8小时前谷歌更新Flow系列，支持Gemini Omni和自定义工具等功能谷歌在最新一届 I/O 开发者大会上宣布，对 AI 创意平台 Google Flow 和 Flow Music 进行大幅升级：全面接入 Gemini Omni / Omni Flash 模型，引入智能创作助手 Flow Agent，支持通过自然语言搭建与共享“视频调整器”和着色器等自定义工具，并将推出移动应用，同时把服务扩展到 140 多个国家和地区。2 来源Flow 与 Flow Music 重大更新：Gemini Omni 加持的创作中枢2 来源Gemini Omni / Omni Flash 带来更精细的视频和音乐创作2 来源Flow Agent 与自然语言工作流：人人都能搭建视频工具与着色器2 来源移动应用与全球扩展：Flow 生态进入 140+ 国家2 来源本内容由AI生成

8小时前微软计划年内启用印度最大数据中心，持续加大AI基础设施投入微软宣布未来四年将在印度投资175亿美元建设云计算和人工智能基础设施，这是其在亚洲史上最大单笔投资。核心项目是在印度海得拉巴建设迄今为止在当地最大的超大规模数据中心，按计划将于2026年年中投入使用，使该数据中心有望在年内正式启用；同时扩建钦奈、海得拉巴和浦那现有园区，并承诺到2030年为2,000万印度人提供AI技能培训。 此次投资是微软总额约230亿美元新AI全球投资计划的一部分，其中还包括未来两年在加拿大追加超过75亿加元用于扩充云计算和AI算力。3 来源微软加码印度：175亿美元锁定全球增长最快数字市场之一3 来源海得拉巴将迎来微软在印最大超大规模数据中心2 来源从基础设施到人才：为2,000万印度人提供AI技能培训2 来源全球AI基础设施军备竞赛与市场担忧3 来源本内容由AI生成

8小时前谷歌更新Project Genie，集成街景视图，AI Ultra订阅者可全球访问谷歌在 2026 年 Google I/O 上宣布重大升级：其通用世界模型 Project Genie 首次深度集成 Google 街景（Street View），可以基于真实街景图像生成可探索、可互动的 3D 世界，用于 AI 智能体与机器人的训练，以及沉浸式虚拟旅行等场景，目前实验性原型仅支持美国街景数据，未来将扩展到更多地区；有关“向全球 AI Ultra 订阅者全面开放访问”的细节尚未在公开报道中得到明确证实。3 来源Genie 接入街景：I/O 2026 上的核心发布2 来源从 2D 小游戏到真实街道：Genie 世界模型的演进2 来源多重应用场景：从机器人训练到虚拟旅行3 来源开放范围与订阅争议：AI Ultra 用户能用到什么？3 来源本内容由AI生成

7小时前SpaceX计划在美IPO后30天收购初创公司Cursor多家媒体援引知情人士消息称，SpaceX 已与 AI 编程初创公司 Cursor 达成框架协议，计划在其美国 IPO 完成并挂牌满 30 天后，以最高约 600 亿美元的价格发起收购；在此之前，双方可先通过一笔约 100 亿美元的合作注资建立深度绑定，若最终收购未成，SpaceX 需向 Cursor 支付约 100 亿美元的反向分手费。2 来源交易概况：IPO 后 30 天启动的“巨额 AI 并购”2 来源两种路径：600 亿美元收购，或 100 亿美元“先合作后收购”2 来源战略考量：以 AI 编程能力强化 SpaceX 全业务链2 来源市场与监管关注：超大并购叠加重磅 IPO 的不确定性2 来源本内容由AI生成

6小时前谷歌和三星电子公布智能眼镜设计，计划秋季上市谷歌与三星电子共同公布了基于 Android XR 平台的新一代智能眼镜设计方案，首批以“音频眼镜”形态在发布会上亮相，主打日常佩戴外观与 Gemini 大模型驱动的语音与 AI 助理能力，计划于今年秋季在部分市场开售。谷歌负责平台和 AI 能力，三星参与硬件设计与生态合作，并联手 Gentle Monster、Warby Parker 等传统眼镜品牌弱化科技感，让产品更接近日常眼镜。同时，三星自有品牌的首款 AI 智能眼镜“Jinju”也将于年中发布，与谷歌音频眼镜一道对标 Meta Ray‑Ban，在快速升温的 AR/AI 眼镜市场中争夺先机。4 来源谷歌与三星联合公布智能眼镜设计：首发锁定今年秋季4 来源外观与硬件：更像日常眼镜的时尚音频设备2 来源深度整合 Gemini 与 Android XR：从语音助手到视觉理解4 来源竞争加速：对标 Meta Ray‑Ban，AI 眼镜迈向大众化4 来源本内容由AI生成

4小时前IBM宣布扩容面向AI时代的企业安全项目综合多项最新发布与技术路线可以看到，IBM 正在通过三大抓手系统性“扩容”其面向 AI 时代的企业安全项目：一是以 AI Factory 与自主存储为核心，构建统一、智能、分层的数据与存储安全底座，将威胁检测与恢复能力前移到基础设施层；二是把会话式 AI、实时情感与语音生物识别等技术嵌入呼叫中心和 CCaaS 平台，在客户服务全链路上强化隐私保护与风控能力；三是依托 AI 开发智能体平台与对 2026 年技术趋势的判断，将安全“左移”至软件开发与运维全生命周期，推动企业在 AI 原生架构下构建持续演进的安全体系。6 来源统一数据平台与自主存储：从“让AI走向数据”到“让安全走进底座”3 来源联络中心与业务前台：AI 驱动的实时风控与隐私保护1 来源AI 技术趋势与安全“左移”：从基础设施延伸到软件供应链4 来源多层次安全能力集成：迈向“自主、安全、可持续”的 AI 企业6 来源本内容由AI生成

1小时前亚德诺宣布15亿美元收购AI电源管理创企据知情人士透露，亚德诺半导体（ADI）已与AI电源管理初创企业“赋能半导体”进入深度谈判阶段，拟以约15亿美元现金将其整体收购，借此强化面向高功耗AI芯片的数据中心与加速卡电源管理能力。 这笔交易规模在模拟/电源管理领域相当可观，最快有望在亚德诺发布季度财报同日对外官宣，但目前仍存在谈判破裂的可能。 业内普遍认为，该交易若落地，将进一步加剧围绕AI基础设施的半导体并购竞争，并巩固亚德诺在工业与汽车等高附加值市场的技术优势。2 来源亚德诺拟斥资15亿美元收购赋能半导体：交易结构与时间表1 来源锁定高功耗AI芯片供电痛点：赋能半导体有何吸引力2 来源顺应AI基础设施投资潮：对市场与竞争格局的潜在影响2 来源不确定性与后续观察：监管、整合与执行风险2 来源本内容由AI生成

1小时前谷歌副总裁谈人工智能影响就业，尽管裁员仍看好劳动力市场谷歌高级副总裁、技术与社会负责人詹姆斯·曼尼卡近日在多家媒体的专访与问答中表示，生成式人工智能在自动化任务和重塑职业方面的能力正在快速提升，但更多是“重构工作内容而非简单消灭岗位”；他承认短期内会出现裁员与岗位调整，却仍对劳动力市场长期前景保持乐观，理由包括新职业与新任务的大量涌现、人类在判断力与创造力方面的优势、以及政策和企业在培训与转岗上的主动作为。3 来源AI 自动化的边界：从“能做什么”到“该不该做”3 来源自动化与裁员：短期冲击与长期乐观的逻辑3 来源职业将如何改变：从“被替代”到“人机协作”3 来源应对挑战：培训、政策和“负责任的 AI 部署”3 来源本内容由AI生成

1小时前OpenAI和Anthropic估值约3700亿美元的慈善资产将变得流动多家媒体援引路透社最新分析称，随着OpenAI与Anthropic等生成式人工智能公司的估值飙升，大量早期股东和创始人将股权或收益权捐赠给基金会、捐赠人建议基金等慈善载体后，约3,700亿美元规模的相关慈善资产有望在未来数年内随这些公司股权“变得流动”，或通过IPO、股权回购、二级市场与新型加密衍生品平台的退出，转化为可实际动用的公益资金，这一体量足以重塑全球慈善格局。2 来源AI估值狂飙，催生3,700亿美元“待兑现”慈善资产2 来源隐形股东：慈善机构如何持有OpenAI与Anthropic“份额”1 来源投机平台推高隐含估值：Ventuals、PreStocks放大预期1 来源可能重塑全球慈善格局：机遇与隐忧并存1 来源本内容由AI生成

56分钟前实测腾讯新AI“牛马”：一句话判定你的电脑能否畅玩黑神话腾讯近日在内部测试并向部分用户开放一款代号为“牛马”的新AI评测助手，主打“玩家一句话，AI帮你判断电脑能否畅玩《黑神话：悟空》”：用户只需用自然语言描述自己的电脑配置或预算需求，“牛马”就能结合游戏的实际性能数据与显卡评测结果，给出是否适合畅玩《黑神话：悟空》以及推荐分辨率、画质档位，甚至引导用户选择本地升级或云电脑等方案，被视为腾讯在游戏场景落地大模型的一次新尝试。4 来源一句话查配置：腾讯“牛马”AI锁定《黑神话：悟空》玩家痛点1 来源AI如何判断“能不能畅玩”：从硬件解析到画质预设3 来源缓解玩家“显卡焦虑”：从本地升级到云电脑方案2 来源大模型落地游戏生态：腾讯押注“AI+硬件+内容”闭环3 来源本内容由AI生成

26分钟前DeepSeek推出Claude竞品，挑战美国编程巨头国产大模型厂商 DeepSeek 宣布其新一代旗舰编程模型 DeepSeek-V4 系列（含 V4-Pro、V4-Flash）正式上线，在代码生成和 Agent 能力上直接对标美国的 Anthropic Claude 和 OpenAI GPT，并通过兼容 Anthropic 协议的 API 与第三方工具深度整合——包括面向“最强编程代理”的 Claude Code 做专项优化与映射，意在亲自下场做 Claude 级别的编程竞品，以更长上下文、更低价格和更强本地接入能力抢占全球开发者市场。6 来源DeepSeek-V4 正式发布：从“震惊全网”到对标 Claude 的编程旗舰5 来源技术与场景：1M 上下文、稀疏注意力与 Agent/Claude Code 专项优化2 来源“亲自下场做 Claude 竞品”：兼容 Anthropic 协议、接管 Claude Code 底层模型2 来源价格战与产业影响：压低开发成本，加剧中外 Agent 赛道竞争6 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。智库·502线上同行往期回顾报名已结束企业构建AI+组织的人效杠杆与路径选择百果园集团x银泰分享企业构建AI+组织的人效杠杆与路径选择妙投波动判断场往期回顾活动已结束芯片、光模块、机器人——AI的泼天富贵，还能传到第几棒？本期我们将AI产业链拆解为三大核心环节，逐一拆解其核心矛盾与风险边界，帮你建立从芯片到机器人的完整坐标系。随行灯塔火热报名中随行灯塔随行灯塔是由虎嗅深度联合运营的高端创始人社群。专注于一件事：帮助创始人把产品做对、做爆、做到10亿。

这里不灌鸡汤，不讲玄学。我们相信，伟大的公司一定是产品驱动的；但产品力来自人，而人需要场——一个能被照见、能被挑战、能被链接的场。随行灯塔，就是创始人的“修炼场”。智库·502线下闭门会往期回顾精彩回顾AI在ToB销售、客服的落地案例探讨围绕着AI在ToB营销内容生成、智能客服接待、线索清洗等场景进行深入交流早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文买家AI改图恶意“仅退款”，平台秒批仲裁逼得新店老板心凉退店发布于AI新闻资讯发布时间 :2026年5月18号 17:54阅读 :1分钟河北辛集市的无花果树种植户张先生，上个月满怀希望地在某直播电商平台开了一家新店，售卖自家培育的无花果苗。然而网店才刚刚卖出 12 单，他就意外收到了一份“仅退款”申请，对方声称收到的果苗已经干枯死亡。买家随申请附上了一张叶片全黄、毫无生机的树苗照片，以此要求退款。张先生有着多年的种植经验，他一眼就识破了这张照片是利用AI技术伪造出来的“假证据”。平台仲裁几分钟闭眼秒批，商家自证清白无门张先生表示自己一直坚持带盆发货且保证盆土湿润，从发货到买家签收全程仅仅耗时两天。按照苗木的生长规律，自然缺水枯萎至少需要五到七天，且树苗叶片会发皱下垂，绝不可能像照片中那样快速枯黄。为了维护自身权益，张先生果断驳回了申请并主动协商退货退款，不料买家随后申请了平台介入。令人心寒的是，即便张先生极力拿出专业常识辩解，平台依旧凭着那张AI伪造的图片，在短短几分钟内就判定买家胜诉。国家出手推出AI内容鉴定，业内提倡以AI治AI这笔 45 元的“仅退款”订单，让满委屈的张先生彻底对该平台失去了信心，一气之下直接启动了退店流程。这起事件也揭示了当前AI生成图像泛滥、网购恶意碰瓷难以精准鉴别的新痛点。针对日益猖獗的AI造假行为，国家反诈中心App近期已正式上线了AI内容鉴定功能，支持对图像和视频进行识别。业内安全专家也指出，未来的核心解决方案在于“以AI治AI”，通过大模型与小模型的协同判断，从源头遏制此类恶意白嫖现象。

AI资讯AI新闻资讯正文GPT-5.5才发三周，5.6已在内测！OpenAI与Anthropic补贴大战同日开打，开发者坐收渔利发布于AI新闻资讯发布时间 :2026年5月14号 14:21阅读 :1分钟三周前GPT-5.5刚落地，GPT-5.6的消息已经提前炸出来了。知名爆料人Leo透露，GPT-5.6目前开发进度已全速推进，首批内部检查点在过去几天已启动测试，下个月很可能就会正式露面。更有意思的是，有人从OpenAI内部Codex日志里扒出了rollout mapping的痕迹——大部分调用还是指向GPT-5.5，但有一条记录赫然落在了GPT-5.6身上。换句话说，Codex环境里可能已经在偷偷用它跑测试了。内部代码名也被顺带曝光:ember-alpha和beacon-alpha。网友直呼:这迭代速度，根本追不上。速度再度拉满——Codex即将上线ultrafast模式GPT-5.6还没到，OpenAI产品端已经先按捺不住了。爆料人Chetaslua同步透露，本周四OpenAI将在Codex上线"ultrafast模式"，响应速度直接提升2到3倍，专门为延迟敏感型任务设计。与此同时，Image Arena排行榜上以+242分断档领先的gpt-image-2，也在同步推进A/B测试更新。这不是第一次了。今年3月GPT-5.4发布时，/fast模式已经实现了1.5倍加速;GPT-5.3-Codex-Spark借助Cerebras芯片，更是把推理速度拉到每秒超过1000token，是普通模式的15倍。而这次的ultrafast，直接在旗舰主力模型上实现提速——不是阉割版，不是小模型凑数，是正儿八经的全尺寸加速。对开发者来说，Agent循环、长任务流水线、浏览器自动化……所有需要"等"的地方，体验将彻底不同。顺带一提，上周ChatGPT官方账号发出一张科幻感十足的图，界面上那句经典的"Ask ChatGPT"悄悄换成了"Message ChatGPT"。奥特曼丢下一句"Call me maybe"，让外界对ChatGPT集成语音通话乃至推出硬件产品的猜测，彻底烧开了。Codex vs Claude Code，补贴战正式开打硅谷最戏剧性的一幕，就在今天上演了。就在OpenAI ultrafast模式上线前夕，Anthropic率先出手——宣布从6月15日起，为付费用户每月提升50%的编程额度，覆盖Claude Agent SDK、命令行工具以及深度接入GitHub工作流的Claude Code，并同步发布了Opus4.7Fast模式，主打比Codex更快的高级推理与更流畅的长上下文编码体验。OpenAI没让人等太久。反手祭出一招重磅补贴:接下来30天内，任何想从其他平台迁移到Codex的企业，直接获赠2个月免费使用权。按Pro计划每月200美元算，相当于白送400美元。奥特曼本人亲自下场喊话，直言Codex是"目前市面上最强的AI编程产品"。仅仅几个小时后，战报出炉:2000名开发者在3小时内主动联系了OpenAI。这场怼脸硬刚，看呆了整个硅谷。开发者们倒是乐了——被两家顶级公司抢着补贴，这种好事，上哪儿说理去。飞轮已经开始自转把这两件事放在一起看，会发现一个比任何单一新闻都更值得关注的趋势正在成形。AI正在加速自我进化。GPT-5.3-Codex是OpenAI首个"参与自身训练"的模型，到GPT-5.5，OpenAI内部85%的员工每周都在用Codex写代码。GPT-5.6几乎可以确定是在GPT-5.5的深度参与下开发出来的——AI在帮OpenAI造更强的AI。与此同时，编程工具的大规模普及正在释放前所未有的工程产能。Codex已有300万周活跃用户，Claude Code的用户数同样在爆发式增长。数百万开发者把AI编程工具当成日常标配，AI生成的代码又反哺回AI的训练与部署，这个循环只会越转越快。奥特曼说过，OpenAI的目标已经不止于AGI，直接瞄准了ASI。当模型迭代速度由AI自身驱动，当AI编程工具成为AI开发的基础设施，当两家万亿级公司用"补贴战"加速这一切的普及——通往ASI的飞轮，已经开始自转了。

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文记忆大模型赛道黑马：MemoraX AI完成种子+轮融资，累计吸金超亿元发布于AI新闻资讯发布时间 :2026年5月20号 11:42阅读 :1分钟人工智能领域的创新步伐正在不断加速，近日，专注于“大模型记忆系统”的深圳忆纪元科技有限公司（以下简称“MemoraX AI”）再度传来融资捷报。公司宣布完成数千万人民币的种子+轮融资，距离其4月完成千万美元种子轮融资仅过去一个月左右，目前两次融资累计金额已突破1亿元人民币。此次融资由光源资本担任独家财务顾问，显示出资本市场对其在该垂直赛道潜力的极高认可。MemoraX AI由前华为诺亚方舟实验室主任郝建邺于2026年3月创立，汇聚了强化学习与人工智能领域的顶尖人才。公司核心聚焦于“Agentic RL（智能体强化学习）”技术，致力于攻克大模型普遍面临的“失忆症”瓶颈。与传统的通过外部检索(RAG)获取信息不同，MemoraX AI通过技术变革，尝试将记忆能力真正内化进模型，实现更深层次的“内生记忆”。对于本次募集的资金，公司表示将重点投入到核心技术的深度迭代与规模化工程落地中。随着“内生记忆模块”从实验阶段迈向产品化，MemoraX AI正加速布局核心应用场景:在B端，致力于赋能企业知识管理与专业领域知识库的智能化升级;在C端，则通过打造极具个性化的智能助手，为用户提供长效、精准的记忆交互体验。作为大模型个性化记忆赛道的先行者，MemoraX AI的连续融资不仅标志着该细分赛道正式从浅层的检索探索迈向深度模型内生变革，也为解决大模型信息碎片化与遗忘问题注入了关键动能。随着未来12个月标准化产品的逐步推出，这家年轻的AI科技企业正试图在蓬勃发展的AI基础设施建设中，建立起属于自己的技术护城河。

AI资讯AI新闻资讯正文通义千问再进化!阿里云“千问云”上线:让AI Agent学会“自主”上云发布于AI新闻资讯发布时间 :2026年5月20号 13:52阅读 :1分钟5月20日，在2026阿里云峰会上，阿里云正式发布面向Agentic时代的新型AI服务平台——“千问云”（Qwen Cloud）。该产品定位为“为AI Agent而生的全栈智能基础设施”，标志着云计算范式从传统的以算力为中心转向以智能体为中心，全面重构大模型时代的服务链路。“千问云”核心亮点在于实现了模型服务的全量“Skill化”与“CLI化”。通过将模型选型、资源调用、认证配置及用量查询等复杂流程封装为标准化工具接口，AI Agent可无需人工编写集成代码，仅凭指令即可自主调取平台能力。目前，平台已聚合包括通义千问Qwen3.7-Max、智谱GLM、月之暗面Kimi、DeepSeek等在内的150余个模型系列，共计480余款主流模型。其中，旗舰级模型Qwen3.7-Max在盲测总榜中位居国产首位，展现出卓越的任务目标对齐能力。在底层支撑方面，阿里云同步推出了基于新一代自研AI芯片“真武M890”的磐久服务器，性能较上代提升3倍，点对点时延降至150ns以内。配合全新升级的“Agentic Cloud”基础设施，云端产品已完成面向Agent的轻量化沙箱环境适配，可应对智能体运行中高频、短生命周期及瞬时高并发的特殊负载。此外，平台推出创新的“Token Plan”订阅模式，旨在通过更灵活的计费方式降低高频AI编程与智能体工具的使用成本。此举预示着云服务将全面进入“智能体原生”阶段，加速AI应用从单一生成向复杂任务自动化的跨越。

AI资讯AI新闻资讯正文智象未来：两周内连获两轮融资，超两千亿参数全模态大模型震撼登场发布于AI新闻资讯发布时间 :2026年5月20号 14:09阅读 :1分钟在刚刚落幕的智象未来首届开放日上，这家处于AI视觉多模态风口顶端的独角兽企业迎来“高光时刻”。公司正式发布了基于新一代原生全模态架构“Unified Transformer（UiT）”打造的图像大模型——HiDream-O1-Image-Pro。这款大模型参数规模超过两千亿，实现了图像像素、文本标记与任务条件的深度统一，在通用文生图、高保真文字渲染及图像编辑等关键任务上达到了行业顶尖水准（SOTA）。HiDream-O1-Image-Pro的发布，标志着智象未来已正式迈入“原生全模态”建模阶段，即通过一套统一的底层架构，实现对图像、视频、文本、音频等多模态信息的融会贯通，向着构建“物理世界模型”的宏大目标迈出关键一步。资本市场的热捧则印证了这一技术的战略价值。智象未来同时宣布完成新一轮亿级融资，投资方阵容豪华，包括深创投、金浦投资、财鑫资本及复聚资本等多家知名机构。尤为引人注目的是，这已是智象未来在短短半个月内完成的第二次融资（此前于 4 月中旬刚完成超 5 亿元规模的融资）。自 2023 年成立以来，智象未来凭借“模型+应用”的双轮驱动模式迅速崛起。公司目前已构建起清晰的“1+1+3”业务矩阵：1 个底座：HiDream系列底层大模型。1 个平台：HiHarness企业服务平台，提供标准化的模型能力输出。3 大智能体应用：覆盖商业营销（HiBurst，TikTok官方Top5 服务商）、影视创作（“帧赞”，已累计制作短漫剧超 5000 分钟）以及社媒创作（vivago，用户规模超 4000 万）。此次开放日现场，智象未来还与上影新视野基金、蓝色光标、捷成世纪、倍尔健康等多家行业巨头达成战略合作，旨在将领先的模型能力深度注入影视、营销、医疗等高频刚需场景。随着底层架构的持续迭代与融资版图的不断扩充，这家由加拿大工程院外籍院士梅涛博士领衔的顶尖技术团队，正加速在全球生成式AI的竞赛中构建属于中国的技术护城河。

AI资讯AI新闻资讯正文重磅布局：OpenAI落地新加坡，斥资超 3 亿新元打造海外首个“AI实验室”发布于AI新闻资讯发布时间 :2026年5月20号 14:16阅读 :1分钟在刚刚开幕的2026年新加坡亚洲科技会展上，全球AI领头羊OpenAI与新加坡数码发展及新闻部正式签署谅解备忘录，宣布开启一项深度的战略合作。此次合作的重头戏，是OpenAI承诺投入超过3亿新元（约合2.34亿美元），旨在推动应用层面的AI创新。作为合作的核心亮点，OpenAI将把美国以外的首个“应用人工智能实验室”（Applied AI Lab）落户新加坡。这一实验室不仅仅是一个技术研究中心，更被视为将尖端AI技术转化为实际生产力的“发动机”。实验室将重点支持公共服务、金融、医疗以及数字基础设施等关键领域的项目，通过“OpenAI for Singapore”计划，助力本地机构攻克从新技术向业务能力转化的难题。除了基础设施的投入，人才战略也是此次合作的重中之重。OpenAI已明确表示，将在未来几年内于新加坡招募并培养200名顶尖的技术专家与工程师团队，将新加坡打造为其在亚太地区的核心战略基地。事实上，新加坡凭借其高度成熟的数字生态与对AI技术的战略性优先考量，早已成为美国科技巨头进入亚洲的首选腹地。OpenAI此次深化布局，既是对新加坡数字经济实力的认可，也是其在全球范围内与谷歌、Anthropic等竞争对手抢滩“AI落地赛道”的关键一棋。随着亚太区域中心的稳固与实验室的启动，OpenAI正加速构建其全球技术生态，将AI技术更直接地嵌入到现实商业场景与政府职能之中。

AI资讯AI新闻资讯正文告别厚重！XREAL发布Project Aura：不足 90 克的安卓XR空间计算“新物种”发布于AI新闻资讯发布时间 :2026年5月20号 14:27阅读 :1分钟在刚刚落幕的2026年谷歌I/O开发者大会上，AR行业先锋XREAL正式揭开了Project Aura智能眼镜的神秘面纱，并官宣该产品将于2026年年底前全面出货。这款基于安卓XR架构打造的设备，不仅在轻量化设计上取得了突破，更凭借其独特的“分离式计算”方案，为AR眼镜的量产化落地提供了全新的范式。轻量化的极致平衡:不足90克的“口袋运算”在现有的硬件水平下，如何兼顾轻便佩戴、高性能算力与长续航始终是行业痛点。XREAL给出的答案是“分离式设计”:将繁重的处理器、电池及触控模组从镜框移出，集成进一个可轻松夹在口袋上的外接盒。得益于这一设计，Project Aura的眼镜本体重量成功控制在90克以内，佩戴体感极佳。这个外接模块不仅是动力源，更集成了实体触控板，作为原生手势操作的强力补充。为了实现精准交互，眼镜框集成了3枚广视角摄像头，经现场实测，其对手势的识别范围与灵敏度表现抢眼，即便在运动中也能轻松完成移动与捏合操作。原生安卓XR生态:谷歌AI的深度赋能作为谷歌Android XR平台的先锋之作，Project Aura能够原生访问全套空间应用。在开发者大会现场，它展示了Google Maps沉浸式3D导航，以及YouTube180度/360度VR视频的全屏播放能力。该产品最大的技术看点在于对多模态Gemini AI的深度集成。当你通过DisplayPort将眼镜连接至笔记本电脑时，系统会自动触发“自动空间化”流程，将原本平面的文档、网页及移动端App，实时转换并“悬浮”在你的视觉空间中，构建出一个触手可及的3D办公与娱乐环境。硬核配置与未来展望Project Aura搭载了XREAL自研的X1S空间计算芯片，并深度协同高通骁龙XR平台，算力保障无虞。在续航方面，虽然软件优化仍在进行中，但目前版本已能支持约4小时的连续使用。尽管目前的工程版本距离完全成熟还有一段距离，但作为一款主打娱乐、轻办公的AR终端，Project Aura显然瞄准了大众用户群体。随着年底出货时间的临近，这款重量更轻、交互更自然、AI更懂你的智能眼镜，或许将成为空间计算时代第一款真正走进千家万户的“标配”设备。

AI资讯AI新闻资讯正文Figure AI总部开启120小时人形机器人分拣直播:10小时人机对决人类多胜192件发布于AI新闻资讯发布时间 :2026年5月20号 15:19阅读 :1分钟近日，估值390亿美元的具身智能巨头Figure AI在其圣何塞总部发起了一场持续超过120小时的机器人包裹分拣全网直播，并在周末引入名为艾梅·杰拉德（Aimé Gérard）的实习生，与旗下人形机器人展开了一场为期10小时的同台竞技。最终，人类实习生以12924个的总分拣量胜出，比机器人对手多出192个。数据表明，实习生平均单件用时2.79秒，而机器人平均用时2.83秒。尽管比赛中途人类因劳动法规定的用餐及洗手间休息被机器人一度反超，但最终凭借速度优势重夺领先，这场赛事也直观展现了现阶段人类与顶尖具身智能在重体力、高频次劳动中的效率博弈。本次长周期直播的核心战略意图在于验证并展示Figure人形机器人在工业场景下不间断、高可靠性的24小时轮班作业能力。自上周三首次完成8小时自主劳动以来，三台人形机器人通过轮流无缝补位充电的方式，在24小时内完成了超3万个包裹的分拣，截至目前已稳定运行超六天。针对这一行业动作，俄亥俄州立大学工程学院院长艾安娜·霍华德（Ayanna Howard）等专家指出，机器人长时间无故障运行令人印象深刻，但由于偶发包裹掉落或条形码面朝上等精度问题，距离物流中心完全自主化应用仍有较长路要走。此次事件不仅通过社交媒体的病毒式传播验证了具身智能的技术演进，也标志着人形机器人正加速从实验室走向严苛工业环境的压力测试阶段。

AI资讯AI新闻资讯正文OpenAI承诺投资超 3 亿新元，在新加坡打造海外首个应用型AI实验室发布于AI新闻资讯发布时间 :2026年5月20号 15:29阅读 :1分钟新加坡数码发展及新闻部于周三正式宣布，人工智能巨头OpenAI将在新加坡设立其在海外的首个应用型AI实验室。这一举措标志着该公司在亚洲市场的重大战略布局，也为当地科技生态注入了强劲动力。重金布局与人才扩充OpenAI计划在未来几年内，将新加坡团队的岗位规模逐步扩大到大约 200 个。与此同时，公司还承诺向该项目投入超过 3 亿新加坡元，折合人民币约15. 97 亿元。这些资金和人力资源将主要用于加速技术落地，推动人工智能在多领域的实际应用。通过这些举措，OpenAI希望能够进一步深化其全球化布局，并提升技术服务能力。携手共赢与经济转型该实验室将与新加坡各方紧密合作，共同推进人工智能应用领域的创新，并着力培育本土的AI专业人才。双方的目标是让先进的人工智能技术真正惠及普通民众、各类企业以及公共部门。当前，新加坡正致力于通过人工智能来推动经济转型，全力打造全球瞩目的人工智能产业中心。OpenAI的重磅加入，无疑将吸引更多顶尖AI企业入驻，加速这一愿景 regional 落地。

腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手AI工具2天前更新AI小集03腾讯Marvis是什么腾讯Marvis是腾讯应用宝团队推出的操作系统层个人 AI 助手。工具通过 1 个主 Agent 与 5 个副 Agent 的协作架构，实现电脑系统操作、文件管理、App 操控、网页交互与信息搜索等复杂任务，24 小时在线并支持 Windows 与安卓跨端操作。产品具备本地知识库索引与可视化虚拟办公室交互，定位大众友好的日常 AI 数字秘书。腾讯Marvis的主要功能多 Agent 协作调度：1 个 PM 主 Agent 负责理解需求、拆解任务并分发给 5 个副 Agent，各 Agent 完成后向主 Agent 汇报，实现复杂任务的自动编排与闭环执行。深度系统操作：可直接调用 Windows 底层接口完成主题设置、电源计划调整、自启动管理、硬件配置检测、网络测速等原生系统级操作，无需模拟点击。智能文件管理：支持基于文件名、内容、图片 OCR、人脸、主题、地点等多维度本地文件搜索，可自动分类归档、格式转换、合并去重及跨端文件传递。跨端 App 操控：通过内嵌应用宝引擎，可在 PC 端直接操控安卓手机 App（如微博签到、同花顺、飞常准等），实现电脑与手机的应用级跨端调度。本地知识库：自动识别并索引本地文档与图片，支持标题匹配与内容匹配搜索，构建越用越懂用户的个人知识管理体系。自动任务与情报监控：支持定时执行既定任务，如每日新闻汇总、限时福利监控、明星动态追踪、房价监控等自动化工作流。如何使用腾讯Marvis下载安装：访问腾讯Marvis官网 https://marvis.qq.com/ 下载 Windows 客户端，按指引完成安装。登录激活：填写调查问卷获取邀请码，获得邀请码后，输入邀请码完成内测资格激活。模式选择：根据任务敏感度选择效率模式或隐私模式。指派任务：在对话框输入自然语言指令，或进入”自动任务”页面设置定时执行规则。查看进度：在右侧虚拟办公室面板查看各 Agent 工作状态与任务消耗 Token 数，等待主 Agent 汇总结果。确认执行：涉及系统修改、文件删除等敏感操作时，点击硬确认弹窗授权后完成最终执行。腾讯Marvis的核心优势操作系统层深度整合：区别于纯云端 Agent，Marvis 深入 Windows 系统底层，可直接调用系统 API 与文件系统，实现真正的本地原生操作。端云双模式灵活切换：效率模式采用云端大模型保障复杂任务性能，隐私模式基于 Qwen 端侧模型实现完全离线运行，满足财务、法务、HR 等敏感场景需求。应用宝跨端生态壁垒：继承 14 年应用商店经验与跨端引擎技术，独家支持 PC 端操控手机 App，生态调度能力为竞品不具备。可视化虚拟办公室交互：将每个 Agent 设计为戴企鹅围脖的黑色小马形象，在虚拟工位中实时展示工作状态（搬砖、打盹、喝咖啡），降低技术门槛并增强用户情感连接。L2 级安全兜底机制：涉及资金交易、系统核心配置修改、批量删除等敏感操作强制”硬确认”，支付功能完全禁用，从架构层面规避 AI 越权风险。腾讯Marvis的同类竞品对比对比维度腾讯Marvis（腾讯）Claude Computer Use（Anthropic）产品定位操作系统层个人 AI 助手云端计算机操控 Agent运行模式端云双模式（本地隐私 / 云端效率）纯云端运行系统深度深入 Windows 底层 API，原生系统级操作基于截图与模拟点击操控跨端能力PC 操控手机 App（应用宝引擎）不支持Agent 架构1 主 + 5 副多 Agent 协作，可视化办公室单 Agent 模式隐私安全L2 级硬确认，支持完全离线运行数据上云，依赖平台安全策略用户门槛自然语言输入，大众友好，开箱即用需一定技术理解，偏极客向生态合作与微软、Intel 系统级深度合作无特殊硬件合作腾讯Marvis的应用场景办公效率提升：自动整理桌面文件并按项目分类归档，一键完成合同信息审查、运营数据分析、参考文献整理及出差材料准备。系统运维助手：一句话调整系统设置（夜间模式、电源计划、默认浏览器），检测游戏配置兼容性并给出硬件升级建议，彻底卸载顽固软件及残留痕迹。生活娱乐管家：自动完成微博超话签到、监控游戏限时福利与明星动态、定时推送每日新闻汇总、整理宝宝成长相册与旅行攻略。敏感数据处理：在完全断网的隐私模式下，本地完成财务报表校对、员工薪资分析、法务合同审查等涉密工作，确保数据不出设备。跨端文件协作：在手机端发送指令，让 Marvis 在电脑中搜索指定文件并回传至手机，实现跨设备的无缝文件调度。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇LibTV团队版 - LibTV 推出的专业级 AI 视频协作工作台下一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型相关文章TRUEBench – 三星开源的AI性能基准测试工具AI小集2DiT – 基于Transfomer架构的扩散模型AI小集7Pyscn – AI代码质量分析工具，快速发现和清理代码问题AI小集2Kimi K2 高速版 – 月之暗面Kimi推出的高性能版模型AI小集4Platus – AI法律平台，AI自动识别关键条款和潜在风险AI小集2Seed1.8 – 字节跳动推出的通用Agent模型AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付1小时前Qwen3.5-LiveTranslate – 阿里通义推出的实时同声传译模型1小时前Qwen3.7-Max – 阿里通义推出的新一代旗舰大模型1小时前Google Pics – 谷歌推出的 AI 图像创建与编辑工具3小时前Gemini Omni Flash – 谷歌推出的多模态视频生成模型4小时前Gemini Spark – 谷歌推出的个人 AI Agent4小时前Gemini 3.5 Flash – Google 推出的新一代 AI 大模型5小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版1天前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架1天前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准1天前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型1天前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前

Kimi K2.5 – 月之暗面kimi开源的全新一代全能旗舰模型AI工具4个月前发布AI小集03Kimi K2.5是什么Kimi K2.5 是月之暗面开源的全新一代全能旗舰模型，模型基于约 15T 视觉与文本数据训练而成，具备顶尖的代码生成、视觉理解能力，支持自主 Agent 集群协作，可调度多达 100 个子 Agent 并行处理复杂任务，执行效率大幅提升。Kimi K2.5 提供快速、思考、Agent 及 Agent 集群四种模式，适用多种场景，包括前端开发、办公自动化和复杂问题解决。模型可通过Kimi.com、Kimi App、Kimi 开放平台和 Kimi Code （Kimi 新推出的编程工具）等平台提供服务，助力用户高效完成任务。Kimi K2.5的主要功能视觉与文本融合：Kimi K2.5 能同时处理文字和图像，理解设计图、视频等内容，实现“看图写应用”，将视觉内容转化为可交互的网页或代码，支持像素级复刻和视觉调整。多模式思考：模型支持思考模式和快速模式切换。思考模式适合复杂任务，可进行深度逻辑推理；快速模式适合即时问答，响应迅速。智能体协同（Agent Swarm）：在复杂任务中，Kimi K2.5 可拆解任务并调度多个“分身”协同工作，支持高并发处理，完成多达 1500 步的操作，适用批量任务和多视角分析。专家级办公助手（Office Agent）：模型深度优化 PPT、Word、Excel 等办公软件的使用，能处理文本和解析图表和复杂排版，提供专业级的输出。本地编程支持（Kimi Code）：基于终端的 Coding Agent，支持本地代码的增删改查，自动生成和修改代码，运行测试，并解释改动，方便开发者使用。一键部署与应用生成：生成的网页或应用可直接部署上线，快速将设计转化为实际可用的产品。Kimi K2.5的性能表现Kimi K2.5 是基于约 15T 视觉和文本混合数据训练的原生多模态模型，在视觉理解、图像生成、视频分析等任务中表现出色。例如，在 MMMU-Pro 和 MathVision 等视觉基准测试中，Kimi K2.5 的表现优于或接近其他顶尖模型。Kimi K2.5的项目官网项目官网：https://www.kimi.com/blog/kimi-k2-5.htmlHuggingFace模型库：https://huggingface.co/collections/moonshotai/kimi-k25Kimi K2.5的API调用价格输入（非缓存）：每 M Tokens 价格为 ¥4，相比 K2 Turbo降低了 50%。输入（缓存）：每 M Tokens 价格为 ¥0.7，相比 K2 Turbo 降低了 30%。输出：每 M Tokens 价格为 ¥21，相比 K2 Turbo降低了 63.8%。Kimi K2.5的应用场景软件开发与编程：Kimi K2.5 可将设计图或视频直接转化为可交互的网页或应用，支持本地代码生成、修改和一键部署，大幅提升开发效率。办公自动化：模型能作为专家级办公助手，深度优化 PPT、Word 和 Excel 文档，支持批量任务处理和多任务协作，显著提高办公效率。教育与学习：Kimi K2.5 能根据学生需求生成个性化学习计划和教学材料，提供实时辅导，减轻教师备课负担。创意与设计：模型可将设计稿直接转为代码，支持批量生成创意素材和视觉调整，助力设计师快速完成项目。商业与市场分析：Kimi K2.5 能快速进行市场调研、生成商业计划书，通过数据可视化辅助决策，助力企业高效分析市场。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇ChatTOC - AI聊天对话生成目录插件，一键定位关键信息下一篇DeepSeek-OCR 2 - DeepSeek团队开源的二代OCR模型相关文章Qwen-Flash – 阿里通义推出的Qwen3系列高性能模型AI小集3Docling – IBM开源的文档解析工具AI小集3GLM-Image – 智谱联合华为开源的多模态图像生成模型AI小集3SynClub – 百度推出的海外AI社交应用AI小集2CogVideoX-2 – 智谱 AI 推出的文本到视频生成模型AI小集3IC-Light – ControlNet作者开源的AI图片打光工具AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付1小时前Qwen3.5-LiveTranslate – 阿里通义推出的实时同声传译模型1小时前Qwen3.7-Max – 阿里通义推出的新一代旗舰大模型1小时前Google Pics – 谷歌推出的 AI 图像创建与编辑工具3小时前Gemini Omni Flash – 谷歌推出的多模态视频生成模型4小时前Gemini Spark – 谷歌推出的个人 AI Agent4小时前Gemini 3.5 Flash – Google 推出的新一代 AI 大模型5小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版1天前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架1天前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准1天前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型1天前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前

Composer 2.5 – Cursor 推出的自研 Agentic 编程模型AI工具1天前发布AI小集02Composer 2.5是什么Composer 2.5 是Cursor推出的自研 Agentic 编程模型。在智能水平和行为表现上较Composer 2有大幅提升，在SWE-Bench Multilingual（79.8%）和CursorBench v3.1（63.2%）等核心基准上与 Claude Opus 4.7、GPT-5.5 处于同一梯队，但单次任务成本仅为竞品的约1/10，称为”性价比之王”。模型基于 Moonshot 开源的Kimi K2.5检查点持续训练，目前仅通过 Cursor IDE 及 SDK 提供服务。Composer 2.5的主要功能长时任务持续工作：针对长时间运行的 Agent 会话深度优化，能在多步骤工具调用中保持专注，显著减少中途幻觉或提前终止的问题。复杂指令可靠遵循：对跨文件重构、终端命令执行、测试驱动开发等复杂指令的遵循可靠性较 Composer 2 大幅提升。努力级别动态校准：模型能根据任务难度自动分配计算量，简单任务快速完成，复杂任务深入思考，避免”小事空转、大事欠思考”。沟通风格优化：回复更简洁结构化，减少不必要的冗长解释，在多文件变更时提供更清晰的推理过程。工具调用精准度提升：显著减少无效的终端命令或冗余搜索，提升代码检索与终端操作效率。双版本灵活适配：提供 Standard（标准版，$0.50/$2.50 per M tokens）与 Fast（快速版，$3.00/$15.00 per M tokens），智能水平相同，分别适配后台批量任务与交互式实时编程。如何使用Composer 2.5在 Cursor IDE 中启用：将Cursor更新至 2026 年 5 月最新稳定版，打开 Composer 面板（Cmd+I/Ctrl+I），点击模型选择器切换至Composer 2.5选择速度档位：交互式开发默认使用Fast 版（响应快、延迟低）；后台 Agent 或批量任务可在 Settings > Models 中切换为Standard 版（成本低，智能水平相同）通过 SDK 程序化调用：import { Agent } from “@cursor/sdk”; const agent = await Agent.create({ model: “composer-2.5”, // Standard 版 // model: “composer-2.5-fast”, // Fast 版 workspace: “./”, tools: [“edit”, “shell”, “search”, “browser”], });为长时任务设限：为无人看管的长时间 Agent 会话设置迭代次数上限和最大耗时，防止模型利用缓存等”捷径”进行奖励作弊领取首发福利：发布首周提供双倍用量额度，可在后台查看额度消耗情况多模型路由策略：将 90% 日常任务交给 Composer 2.5 处理；仅将架构评审类任务路由给 Claude Opus 4.7，重度终端 Shell 任务路由给 GPT-5.5Composer 2.5的核心优势极致性价比：Standard 版输入 $0.50/M、输出 $2.50/M，Fast 版输入 $3.00/M、输出 $15.00/M，比 Claude Opus 4.7 便宜约1030 倍。前沿级基准表现：SWE-Bench Multilingual 79.8%（Opus 4.7 为 80.5%，GPT-5.5 为 77.8%），CursorBench v3.1 63.2%（与 Opus 4.7 的 64.8% 和 GPT-5.5 的 64.3% 基本持平）。行为层面深度优化：除了扩大训练规模，还改进了沟通风格和投入级别校准，这些维度虽难被基准充分反映，但对实际使用体验至关重要。双版本灵活选择：Standard（标准版）适合后台 Agent 与批量任务，Fast（快速版，默认）适合交互式 IDE 实时编程，两者智能水平完全相同。首发双倍用量福利：发布首周提供双倍用量额度。Composer 2.5的同类竞品对比对比维度Composer 2.5Claude Opus 4.7GPT-5.5厂商 / 平台CursorAnthropicOpenAI产品定位自研 Agentic 编程模型旗舰推理模型旗舰多模态模型基座模型Moonshot Kimi K2.5（开源检查点持续训练）Claude 4 系列GPT-5 系列发布日期2026.05.182026 年 Q22026 年 Q2SWE-Bench Multilingual79.8%80.5%77.8%Terminal-Bench 2.069.3%69.4%82.7%CursorBench v3.1（困难任务）63.2%64.8%（max）/ 61.6%（默认 xhigh）64.3%（xhigh）/ 59.2%（默认 medium）输入价格（/M tokens）$0.50（Standard）<$3.00（Fast）未公开（行业参考约 $15）未公开（行业参考约 $3–$5）输出价格（/M tokens）$2.50（Standard）<$15.00（Fast）未公开（行业参考约 $75）未公开（行业参考约 $15–$30）单次任务相对成本基准（约 $1–$2 / 任务）约10–30 倍约3–10 倍上下文窗口200K（参考 Kimi K2.5）200K128K–1M权重开放性闭源（仅 Cursor 基础设施）闭源闭源接入方式Cursor IDE / CLI /@cursor/sdkAPI / Claude Code / 第三方平台API / ChatGPT / GitHub CopilotComposer 2.5的应用场景多文件级重构：成本优势明显且精度持平前沿模型，适合大规模代码库迁移。交互式结对编程：Fast 版响应迅速，适合实时 IDE 协作。后台定时任务/云 Agent：Standard 版性价比极高，适合批量代码审查与修复。测试驱动开发：长时任务可靠性优于前代，能稳定完成多轮测试-修复循环。复杂终端自动化：Terminal-Bench 2.0 得分 69.3%，与 Opus 4.7（69.4%）持平，但重度 Shell 场景仍略逊于 GPT-5.5（82.7%）。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型下一篇Chronicles-OCR - 腾讯联合高校等推出的视觉感知评测基准相关文章Herodot AI – AI旅行助手，拍照识别景点或艺术品获取故事AI小集2VideoGrain – 悉尼科技大学和浙大推出的视频编辑框架AI小集2TimesFM 2.0 – 谷歌研究团队开源的时间序列预测模型AI小集2Mobius – 重庆邮电联合美团等推出的无缝循环视频生成技术AI小集4Fancy123 – 华中科技和华南理工推出的3D网格生成技术AI小集2Gemini Embedding – 谷歌推出的文本嵌入模型AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付1小时前Qwen3.5-LiveTranslate – 阿里通义推出的实时同声传译模型1小时前Qwen3.7-Max – 阿里通义推出的新一代旗舰大模型1小时前Google Pics – 谷歌推出的 AI 图像创建与编辑工具3小时前Gemini Omni Flash – 谷歌推出的多模态视频生成模型4小时前Gemini Spark – 谷歌推出的个人 AI Agent4小时前Gemini 3.5 Flash – Google 推出的新一代 AI 大模型5小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版1天前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架1天前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准1天前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前

Qwen3.5-LiveTranslate – 阿里通义推出的实时同声传译模型AI工具1小时前发布AI小集02Qwen3.5-LiveTranslate是什么Qwen3.5-LiveTranslate 是阿里通义团队推出的实时同声传译大模型，支持60语种输入、29语种输出与3500+翻译组合，通过可读单元流式技术将端到端字均延迟压缩至2.8秒，模型具备实时音色克隆与热词增强能力，专为跨境会议、直播出海与全球协作场景打造。Qwen3.5-LiveTranslate的主要功能多语种全覆盖翻译：支持60个语种的听说读写与29个语种的语音输出，组合成3500+翻译对，覆盖全球主流商务与高频互译场景。超低延迟流式同传：基于全新”可读单元”技术实现流式输出，端到端字均延迟低至2.8秒，满足直播连麦与实时会议的无卡顿需求。实时音色克隆：仅需一句话可克隆用户原声音色，跨语种输出时保持声音质感与身份一致性。热词精准增强：支持人名、品牌、行业术语的动态配置与优先识别，确保专业场景下术语翻译准确无误。Qwen3.5-LiveTranslate的技术原理流式可读单元架构：将语音输入切分为最小语义单元进行实时编码与解码，避免等待完整句子的传统批处理模式，显著降低端到端延迟。端到端语音大模型：基于Qwen3.5多模态底座，统一建模语音识别、翻译与语音合成，减少级联误差，提升跨语种语音转换的自然度。音色一致性编码：通过说话人特征提取与声码器重建，在翻译后的目标语种语音中保留原始音色、语调和情感特征。热词动态注入机制：在推理前将用户自定义术语库注入模型上下文，提升特定词汇的识别优先级与翻译准确率。如何使用Qwen3.5-LiveTranslate访问体验入口：打开浏览器进入 QwenChat Qwen3.5-LiveTranslate 模型入口https://omni.qwen.ai/ 或等待阿里云百炼平台上线。选择同传模式：在对话界面切换至 LiveTranslate 实时翻译模式，设定源语种与目标语种。配置热词库（可选）：上传或输入会议/直播涉及的专业术语、人名与品牌名，提升翻译精准度。开启实时同传：点击开始按钮，对着麦克风说话，系统将实时输出翻译后的目标语种语音。音色克隆设置（可选）：录制一句开声样本，启用”使用我的声音”功能，让翻译输出保持原声音色。Qwen3.5-LiveTranslate的核心优势延迟行业领先：2.8秒字均延迟接近人类同传水平，远优于传统级联方案。音色保真度高：实时克隆技术让跨语种沟通更具个人辨识度与亲和力。专业场景适配：热词增强功能解决品牌名、技术术语的翻译痛点。Qwen3.5-LiveTranslate的同类竞品对比维度Qwen3.5-LiveTranslate讯飞同传Google Live Translate延迟端到端2.8秒约3-5秒约3-6秒支持语种60听/29说/3500+组合主要语种主要语种音色克隆✅ 实时克隆❌ 标准音色❌ 标准音色热词定制✅ 动态配置✅ 术语库❌部署方式Web端+API客户端+API移动端集成Qwen3.5-LiveTranslate的应用场景跨境商务会议：替代传统人工同传，支持多语种远程会议实时语音互译，显著降低沟通成本与语言门槛。直播出海：主播以母语实时输出，系统自动翻译为目标语种并保持原声音色，实现零时差跨国互动。全球客户支持：客服团队以用户母语实时语音响应，提升服务效率与客户满意度。国际教育培训：讲师授课内容实时翻译为学员母语，打破语言壁垒扩大全球受众覆盖。跨国企业协作：分布于不同国家的团队通过实时同传进行项目会议与日常沟通，消除语言障碍。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Qwen3.7-Max - 阿里通义推出的新一代旗舰大模型下一篇ChatGPT Plus 和 Claude Pro 会员代充值 - 支持支付宝、微信支付相关文章photoeditor – AI图片处理工具，支持多功能多格式图像编辑服务AI小集2GLM-Z1-Rumination – 智谱推出的沉思模型AI小集3HunyuanOCR – 腾讯混元推出的端到端OCR视觉语言模型AI小集3AIFlowy – 开源的企业级AI应用开发平台AI小集3SongGeneration – 腾讯AI Lab开源的音乐生成大模型AI小集5Gambo – AI游戏开发Agent，一句话生成完整游戏AI小集4暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付1小时前Qwen3.7-Max – 阿里通义推出的新一代旗舰大模型1小时前Google Pics – 谷歌推出的 AI 图像创建与编辑工具3小时前Gemini Omni Flash – 谷歌推出的多模态视频生成模型4小时前Gemini Spark – 谷歌推出的个人 AI Agent4小时前Gemini 3.5 Flash – Google 推出的新一代 AI 大模型5小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版1天前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架1天前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准1天前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型1天前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前

Qwen3.7-Max – 阿里通义推出的新一代旗舰大模型AI工具1小时前发布AI小集02Qwen3.7-Max是什么Qwen3.7-Max 是阿里通义千问团队推出的面向智能体时代的新一代旗舰大模型，定位为全能智能体基座。模型具备前沿编程、办公自动化、长周期自主执行与跨框架泛化四大核心能力，在 SWE-Pro、MCP-Atlas、GPQA Diamond 等数十项编程、智能体与推理基准上取得领先成绩，可无缝集成至Claude Code、OpenClaw、Qwen Code 等主流智能体框架。Qwen3.7-Max的主要功能前沿编程智能体：支持从前端原型开发到复杂多文件软件工程的全链路代码编写与调试，在 SWE-Pro、SWE-Multilingual 等编程基准上表现领先。办公生产力助手：通过 MCP 集成与多智能体协作实现工作流自动化，在 SpreadSheetBench-v1 办公自动化基准上得分 87.0，可承接复杂数据分析与文档生成任务。长周期自主执行：具备持续稳定的超长任务执行能力，已在长达 35 小时、超过 1000 次工具调用的全自主内核优化实验中保持连贯推理。跨框架泛化：原生适配 Claude Code、OpenClaw、Qwen Code 等主流智能体框架，无需针对特定框架微调即可稳定发挥。Qwen3.7-Max的技术原理环境扩展训练：在 Qwen3.5 环境扩展方法基础上大幅扩展智能体训练环境的质量与多样性，使模型能力从多样化环境中实现泛化。解耦式 Rollout 基础设施：将训练实例解耦为任务、运行框架与验证器三个正交组件，支持跨框架与跨验证器的强化学习训练，迫使模型学习泛化解题策略。组合式扩展：同一任务可与不同类型、不同版本的框架及验证器以极低边际成本自由重组，实现训练环境的组合式规模化扩展。长程强化学习优化：通过长周期自主执行中的持续反馈迭代，模型在 30 小时以上仍能发现实质性改进，验证长程优化与自我进化能力。如何使用Qwen3.7-MaxQwen3.7-Max计划将通过阿里云百炼提供服务。Qwen3.7-Max的核心优势智能体基准全面领先：在 MCP-Mark、MCP-Atlas、ClawEval、QwenClawBench 等通用智能体基准上超越或紧追 Claude Opus-4.6 Max。编程能力顶尖：SWE-Pro 60.6、SWE-Multilingual 78.3、Terminal Bench 2.0 69.7，全面领先同类模型。推理与知识深厚：GPQA Diamond 92.4、HMMT 2026 Feb 97.1、HLE 41.4，在高难度 STEM 推理上处于第一梯队。多语言能力一流：WMT24++ 85.8、MAXIFE 89.2、MMLU-Pro 89.6，翻译与跨语言理解质量顶尖。真实生产力闭环：可将需专业团队一至两周的复杂项目压缩至数小时内端到端交付。硬件无关泛化：在训练时未见过的平头哥真武 M890 硬件平台上，能通过自主探索完成深度内核优化。Qwen3.7-Max的同类竞品对比对比维度Qwen3.7-MaxClaude Opus-4.6 Max编程智能体SWE-Pro 60.6 / Terminal Bench 69.7 领先SWE-Pro 59.0 / SWE-Verified 80.8 略领先通用智能体MCP-Atlas 76.4 / ClawEval 65.2 领先MCP-Atlas 75.8 / ClawEval 70.4 领先推理能力GPQA Diamond 92.4 / HLE 41.4 领先GPQA Diamond 91.3 / HLE 40.0办公自动化SpreadSheetBench 87.0SpreadSheetBench 89.3 略领先多语言WMT24++ 85.8 / MAXIFE 89.2 领先WMT24++ 82.7长周期执行35小时/1000+工具调用自主优化，30小时后仍持续改进长上下文稳定，但公开的长程自主优化案例较少跨框架泛化原生适配 Claude Code / OpenClaw / Qwen Code 等多框架主要针对 Claude Code 优化提供服务阿里云百炼 API（即将上线）Anthropic API / Claude 应用Qwen3.7-Max的应用场景复杂软件开发：作为 AI 软件工程师，独立完成需求分析、架构设计、多文件编码、调试与性能优化的全周期开发任务。企业工作流自动化：通过 MCP 连接企业工具链，自动执行数据分析、报表生成、跨系统信息整合等高强度办公任务。底层系统优化：在陌生硬件平台上自主进行 GPU 内核编写、编译、性能分析与迭代优化，实现数量级加速。科研与数学推理：承接高复杂度数学证明、科学计算与文献整合任务，辅助科研人员处理高难度推理工作。多语言内容生产：依托顶尖多语言能力，完成高精度翻译、跨语言技术文档撰写与全球化内容适配。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Google Pics - 谷歌推出的 AI 图像创建与编辑工具下一篇Qwen3.5-LiveTranslate - 阿里通义推出的实时同声传译模型相关文章FilmAction – AI电影创作平台，从故事构思到视频生成一站式AI小集3Evolving Agents – 开源的AI Agent管理与进化框架AI小集3QwenLong-L1.5 – 阿里通义开源的长文本推理模型AI小集4Buildin.AI – AI知识管理平台，支持智能写作、内容生成和数据分析AI小集3LaWGPT – 南京大学推出的中文法律大语言模型AI小集2CapCut AI Suite – 字节推出的一站式AI创意编辑工具AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付1小时前Qwen3.5-LiveTranslate – 阿里通义推出的实时同声传译模型1小时前Google Pics – 谷歌推出的 AI 图像创建与编辑工具3小时前Gemini Omni Flash – 谷歌推出的多模态视频生成模型4小时前Gemini Spark – 谷歌推出的个人 AI Agent4小时前Gemini 3.5 Flash – Google 推出的新一代 AI 大模型5小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版1天前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架1天前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准1天前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型1天前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前

Gemini Spark – 谷歌推出的个人 AI AgentAI工具4小时前发布AI小集02Gemini Spark是什么Gemini Spark 是 Google 推出的 24/7 个人 AI Agent，基于 Gemini 3.5 模型与Antigravity框架运行，深度集成 Gmail、Docs、Slides 等 Workspace 工具。工具能在设备关闭后仍于云端后台持续执行任务，支持设置周期性触发器、学习新技能并创建完整工作流，将 Gemini 从问答助手升级为可主动代劳的数字代理。Gemini Spark的主要功能全天候后台任务执行：Spark 作为云端 Agent，即使用户关闭笔记本或锁定手机，仍能在后台持续完成复杂任务。周期性任务与触发器：自动解析月度信用卡账单，标记新增或隐藏的订阅费用；按设定条件持续监控指定信息源。可教学的新技能：用户可指令 Spark 检查特定邮件来源（如学校通知），提取关键截止日期，生成摘要发送给指定联系人。完整工作流创建：自动整合分散在邮件与聊天中的会议笔记，生成格式化的 Google Docs 报告，并起草项目启动邮件。MCP 扩展连接：已支持 Canva、OpenTable、Instacart 等第三方服务，未来数周内将通过 MCP 协议直接代用户完成跨平台操作。如何使用Gemini Spark目前 Gemini Spark 处于 Beta 阶段，仅面向美国地区的 Google AI Ultra 订阅用户逐步开放，可在Gemini App设置中查看并启用。Gemini Spark的核心优势真·全天候在线：区别于对话窗口内响应的助手，Gemini Spark 可在用户离线时持续推进任务进度。深度原生集成：与 Google Workspace 无缝打通，数据读写权限与协作能力远超外挂式插件。可教学进化：用户通过自然语言即可训练 Spark 掌握专属工作流，无需编程或复杂配置。多模态任务链：支持从信息收集、内容生成到跨应用执行的端到端闭环，非单点工具调用。渐进式权限控制：用户自主决定 Agent 的开启状态、连接范围与操作边界，隐私与可控性并重。Gemini Spark的同类竞品对比对比维度Gemini SparkClaude Computer UseOpenAIOperator所属公司GoogleAnthropicOpenAI核心定位24/7 云端个人 AI Agent桌面操控型 AI Agent云端浏览器代理运行环境云端后台（设备可离线）本地 macOS（需设备在线）云端沙盒浏览器后台持续执行支持，关闭设备仍运行不支持，随会话结束停止有限，长任务易中断桌面应用操控夏季通过 macOS App 支持已支持完整桌面控制不支持，仅限浏览器内办公套件集成原生深度集成 Workspace连接器 + 屏幕控制 fallback无原生办公集成OSWorld 任务成功率未公开（基于 Gemini 3.5）约 72%–73%约 38%–43%人在回路确认高风险操作前请求确认频繁请求确认频繁请求确认，易冻结Gemini Spark的应用场景个人财务管理：自动追踪订阅服务扣费异动，每月生成账单分析报告并推送提醒。家庭事务协调：持续监控学校邮件，汇总活动与截止日期，每日向父母双方发送结构化摘要。职场项目启动：自动收集会议纪要与邮件讨论，整理为项目文档并起草团队通知邮件。跨平台内容创作：调用 Canva 生成配图，结合 Docs 撰写文案，通过 Gmail 发送完整营销方案。日常行程管理：基于 Calendar 与邮件内容，主动识别冲突并建议调整，甚至代为预订餐厅。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Gemini 3.5 Flash - Google 推出的新一代 AI 大模型下一篇Gemini Omni Flash - 谷歌推出的多模态视频生成模型相关文章Tacore – AI编程工具，多Agent协同工作全流程AI小集2AIMv2 – 苹果开源的多模态自回归预训练视觉模型AI小集2MMSI-Video-Bench – 上海AI Lab推出的空间智能视频基准AI小集2Mathtutor on Groq – AI数学辅导工具，支持用语音形式提出数学问题AI小集2TRAE Agent – 字节开源的智能助手，专为软件工程任务设计AI小集3句无忧 – 在线AI违禁词查询检测和合规文案生成工具AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付1小时前Qwen3.5-LiveTranslate – 阿里通义推出的实时同声传译模型1小时前Qwen3.7-Max – 阿里通义推出的新一代旗舰大模型1小时前Google Pics – 谷歌推出的 AI 图像创建与编辑工具3小时前Gemini Omni Flash – 谷歌推出的多模态视频生成模型4小时前Gemini 3.5 Flash – Google 推出的新一代 AI 大模型5小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版1天前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架1天前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准1天前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型1天前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前

Google Antigravity是什么Google Antigravity 是谷歌推出的 AI IDE 编程智能体，通过 AI 助手的自主性和高效性，改变软件开发的模式。Google Antigravity以Gemini 3等先进 AI 模型为核心，将 AI 助手从工具转变为积极的合作伙伴。Google Antigravity提供独立的智能体界面，能自主规划和执行复杂的端到端软件任务，同时支持跨编辑器、终端和浏览器的操作。Google Antigravity 集成Gemini 2.5 Flash和Nano Banana等模型，进一步增强功能。开发者可通过任务级的交互，更高效地完成开发工作，同时享受智能体的代码验证和自我改进能力。Google Antigravity已开放公测，免费提供给开发者使用，支持 Mac、Linux 和 Windows 系统。Antigravity 2.0版本平台定位升级，从 IDE 插件升级为独立桌面应用，成为真正的 Agent 工作环境，支持多智能体并行协作。Google Antigravity的主要功能自主任务规划与执行：Antigravity 的智能体能自主规划和执行复杂的端到端软件任务，从代码编写到测试验证，减少人工干预。跨环境操作：智能体能直接访问编辑器、终端和浏览器，实现代码编写、本地运行和功能测试的全流程自动化。任务级交互：通过任务列表、实施计划、截图等“工件”（Artifacts），为用户提供任务进度的可视化，增强用户对智能体工作的信任。异步与同步交互：提供“编辑器视图”用于同步操作，及“管理器视图”用于异步任务管理，支持多任务并行处理。实时反馈与改进：用户能在任何工件上提供反馈，智能体将自动调整执行策略，从反馈中学习，持续优化自身表现。多模型集成：支持 Gemini 3 Pro、Gemini 2.5 Computer Use（用于浏览器控制）和 Nano Banana（图像编辑模型），提供强大的功能组合。知识管理与自我提升：智能体通过知识库学习过往经验，将成功步骤和代码片段转化为知识，用于未来任务。Antigravity CLI：替代Gemini CLI，后者将于 2026 年 6 月 18 日停止对 Pro/Ultra 用户服务。Antigravity SDK 发布：开发者可将 Google 的 Agent harness 部署到自己的服务器上运行。原生语音与生态打通：整合 Gemini 音频模型实现语音支持，与 Android、Firebase、AI Studio 实现生态互通。Google Antigravity的使用示例Prompt：Build me a flight lookup nextjs web app where the user can put in a flight number and the app gives you the start time, end time, time zones, start location, and end location of the flight. For now, use a mock API that returns a list of matching flights. Display the search results under the form input.（为我构建一个Next.js航班查询网页应用，用户可以输入航班号，应用会给出航班的开始时间、结束时间、时区、起始地点和结束地点。目前，使用一个模拟API来返回匹配航班的列表。将搜索结果显示在表单输入下方。）Prompt：Please add a contact form to my website through a button on the top right.(请在网站右上角通过按钮添加联系表单。)如何使用Google Antigravity注册与下载：访问 Google Antigravity 官网 https://antigravity.google/ ，下载安装客户端。设置与配置：安装完成后，根据提示配置开发环境和 AI 模型，调整界面和功能设置适应个人开发习惯。创建项目与任务：在Google Antigravity中新建项目，定义开发任务将其分解为多个子任务，并明确任务目标和优先级。使用智能体：通过“编辑器视图”进行同步交互，或在“管理器视图”中启动智能体执行任务，实时监控进度。提供反馈：在智能体生成的文本、代码或截图上直接添加反馈，智能体自动调整执行策略。知识管理与复用：查看智能体的知识库，复用过往经验提升开发效率和质量。Google Antigravity的产品定价Google Antigravity 在公测阶段，目前对个人开发者免费开放，用户可访问包括 Gemini 3 Pro,Claude Sonnet 4.5,GPT-OSS等在内的多种 AI 模型。Google Antigravity的应用场景前端开发：快速生成和优化前端代码，智能体能自动创建页面布局、编写 CSS 样式，在浏览器中实时测试效果。后端开发：智能体能协助编写后端逻辑代码，如数据库操作、API 开发等，进行代码优化和单元测试。全栈开发：在一个项目中同时处理前端和后端任务，智能体能协调不同模块的开发，实现端到端的功能实现和测试。自动化测试：Google Antigravity 能自动生成测试用例，执行自动化测试，验证代码的正确性和性能。代码审查与优化：智能体能对现有代码进行审查，提出优化建议，甚至自动修复一些常见的代码问题。©️版权声明：若无特殊声明，本站所有文章版权均归AI工具集原创和所有，未经许可，任何个人、媒体、网站、团体不得转载、抄袭或以其他方式复制发表本站内容，或在非我站所属的服务器上建立镜像。否则，我站将依法保留追究相关法律责任的权利。类似于Google Antigravity的工具Clacky AIAI编程工具，打造L3级的Coding StudioGenieCosine AI推出的AI编程助手Fitten Code非十科技推出的免费AI代码助手Deco京东推出的设计稿一键生成多端代码工具aiXcoder自然语言到代码的方法级代码生成，以及多行智能代码补全Zcode智谱推出的轻量级AI IDE编程工具2 条评论再想想发表评论wi游客这个在国内如何登陆 呢，有方法不5个月前回复AI小集有点复杂 这个自行搜索解决吧~5个月前回复

Gemini 3.5 Flash – Google 推出的新一代 AI 大模型AI工具5小时前发布AI小集02Gemini 3.5 Flash是什么Gemini 3.5 Flash 是 Google 推出的新一代AI大模型，以前沿智能+行动能力为核心定位。模型在 Terminal-Bench 2.1 编码基准得分 76.2%、GDPval-AA 代理任务 1656 Elo、MCP Atlas 达 83.6%，均超越Gemini 3.1 Pro，输出速度达前沿模型 4 倍，成本不到一半，支持 100 万 Token 多模态上下文，已全球上线Gemini App与 Search AI Mode。Gemini 3.5 Flash的主要功能极速推理与生成：输出速度达 289 tokens/秒，经 Antigravity 内部联合优化后可达 12 倍提升，显著降低 Agent 多轮调用延迟。顶级编码与 Agent 能力：在 Terminal-Bench 2.1（76.2%）、SWE-Bench Pro（55.1%）、MCP Atlas（83.6%）、Toolathlon（56.5%）等基准上全面超越 Gemini 3.1 Pro。多模态长上下文理解：原生支持 100 万 Token 输入窗口，CharXiv Reasoning 得分 84.2%，MMMU-Pro 达 83.6%，可统一处理文本、图像、视频、音频及 PDF。子代理协作编排：通过 Antigravity 2.0 部署并行子代理，官方演示 93 个 subagent 并行 12 小时内以不到 1000 美元完成完整操作系统开发。动态推理级别控制：支持thinking_level参数（标准/扩展），用户可按需切换思维深度，代码与工具调用默认使用低档位优化速度。Gemini 3.5 Flash的技术原理模型蒸馏与压缩架构：基于 Gemini 3.5 系列全新架构，通过先进蒸馏技术压缩模型体积，在保持接近 Pro 级能力的同时实现极低延迟。动态推理预算分配：引入thinking_level机制替代固定预算，根据任务类型自动分配计算资源，代码生成与工具调用默认使用 low 档位。原生多模态统一编码：采用统一多模态编码器，将文本、图像、音视频嵌入同一语义空间，优化百万级长序列注意力计算。前沿安全对齐框架：遵循 Frontier Safety Framework，强化网络与 CBRN 安全防护，通过可解释性工具检查内部推理，降低有害输出与误拒率。如何使用Gemini 3.5 Flash选择接入平台通过Gemini App、Google Search AI Mode、Google AI Studio、Antigravity2.0 桌面应用访问模型。更新 API 模型标识：将代码中的模型名称从旧版预览标识替换为正式版gemini-3.5-flash。配置推理级别：根据任务类型设置thinking_level参数：代码与工具调用选 low，通用任务选 medium，复杂推理选 high。移除采样参数覆盖：删除手动设置的 temperature、top_p、top_k 默认值，让模型使用内部优化配置。注意功能迁移例外：若工作流依赖 Computer Use 浏览器控制功能，需继续用gemini-3-flash-preview，3.5 Flash 暂不支持该功能。Gemini 3.5 Flash的核心优势速度领先：输出速度达 289 tokens/秒，是其他前沿模型的 4 倍，经 Antigravity 内部优化后可达 12 倍提升，Agent 多轮调用几乎无感知延迟。成本优势：API 定价 $1.50/$9.00 每百万 Token（输入/输出），比 Claude Opus 4.7、GPT-5.5 等 frontier 竞品便宜一半以上，Google 估算头部 Cloud 客户将 80% frontier 工作负载迁移至 3.5 Flash 后每年可节省超 10 亿美元。Agent 性能突出：MCP Atlas 得分 83.6%、Toolathlon 56.5%，支持 93 个子代理并行协作，12 小时内以不到 1000 美元完成完整操作系统开发，专为复杂长周期代理任务优化。超长上下文窗口：支持 100 万 Token 多模态输入，可一次性处理整本长篇小说、百页法律合同或长视频内容，长文档与视频分析无压力。全平台生态覆盖：已接入 Gemini App、Search AI Mode、Antigravity 2.0、AI Studio、Android Studio 等 Google 全系产品，消费者、开发者、企业三端同步可用。Gemini 3.5 Flash的项目地址项目官网：https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/#gemini-3-5-flashGemini 3.5 Flash的同类竞品对比对比维度Gemini 3.5 FlashClaude Opus 4.7GPT-5.5Gemini 3.1 Pro输出速度289 t/s（4x）67 t/s71 t/s135 t/sTerminal-Bench 2.176.2%–78.2%70.3%SWE-Bench Pro55.1%–58.6%54.2%MCP Atlas83.6%（领先）79.1%75.3%78.2%GDPval-AA Elo1656175317691314CharXiv Reasoning84.2%82.1%84.1%83.3%上下文窗口1M Token200K128K1MGemini 3.5 Flash的应用场景实时 AI 编程助手：低延迟特性适合 IDE 自动补全、代码审查与交互式调试，Antigravity 2.0 已将其设为默认模型。企业级 Agent 自动化：部署并行子代理处理多周工作流，如 Xero 自动管理 1099 税表供应商识别与信息收集。金融文档智能审核：Macquarie Bank 用于推理 100 页以上复杂文档，加速客户入职与合规审核。多模态发票与 OCR 处理：Ramp 结合多模态理解与历史模式推理，实现复杂发票的智能化识别与分类。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Qwen3.7 Preview - 阿里通义推出的下一代旗舰大模型预览版下一篇Gemini Spark - 谷歌推出的个人 AI Agent相关文章MCP Course – Hugging Face 推出的免费MCP课程AI小集3XHSPlUS – AI营销创作工具，集成创作、运营、数据分析等功能AI小集2MAI-Image-2-Efficient – 微软推出的轻量版文生图模型AI小集2PhotoDoodle – 字节联合新加坡国立大学等推出的艺术化图像编辑框架AI小集3JoyCoder – 京东推出的AI编程助手AI小集3Bika.ai – AI组织管理平台，多智能体自动调度AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付1小时前Qwen3.5-LiveTranslate – 阿里通义推出的实时同声传译模型1小时前Qwen3.7-Max – 阿里通义推出的新一代旗舰大模型1小时前Google Pics – 谷歌推出的 AI 图像创建与编辑工具3小时前Gemini Omni Flash – 谷歌推出的多模态视频生成模型4小时前Gemini Spark – 谷歌推出的个人 AI Agent4小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版1天前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架1天前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准1天前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型1天前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前

Julian ChokkattuBoone AshworthGearMay 19, 2026 1:45 PMHands-On With All of Google’s New Upcoming Android XR Smart GlassesHere’s your first look at smart glasses coming from Warby Parker and Gentle Monster, powered by Google and Samsung’s XR platform.Photograph: Julian Chokkattu/WIREDCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle has beenteasing its Android XR-poweredsmart glassesfor the past year and a half—ever since it first announced the newmixed reality platform in December of 2024. Now, we’re finally starting to see more polished versions of the hardware, several of which will actually go on sale this fall.AtGoogle I/O, the company gave us a first look at thedesigns of smart glassescoming from established eyewear brands Warby Parker and Gentle Monster. These smart glasses are codeveloped by Samsung and Google, while the eyeglass companies handle the design of the frames.Xreal also took the stage at I/O to showcase its upcoming Project Aura, which is essentially a miniaturized glasses version of bulky headsets like theApple Vision ProandSamsung’s Galaxy XR—powering a full Android app interface with hand gestures for interaction.Audio-only smart glasses from Warby Parker and Gentle Monster will launch later this year. More complex versions with displays built into the lenses will arrive soon after. Xreal’s Project Aura will also come this fall.I got a chance to wear some early versions of these frames and try out a few features in a controlled demo experience.Smarter GlassesPhotograph: Julian Chokkattu/WIREDI wasn’t able to see the final polished designs from Warby Parker and Gentle Monster. Instead, my demos were on unfinished reference glasses from Samsung and Google. The most remarkable thing about them is how light they are. Google says much of the work it and Samsung have done was to miniaturize the technology to save weight. The arms of the glasses are still a little chunky, though.Audio quality was equally impressive. As soon as I asked Gemini to play some Radiohead, the audio surrounded my head and sounded dynamic, though I was in a quiet room. I asked someone else to try them so I could see if I could hear it at 50 percent volume while sitting across from them. I could barely hear the music.The audio-only glasses will be the first to arrive later this year. And yes, all of these upcoming smart glasses—even the audio-only ones—have cameras. That’s how the Gemini assistant can see what you’re seeing to provide helpful, contextual answers and services. The versions with more advanced optics will provide a richer experience. For example, if someone talks to you in a different language, the Gemini assistant inside all of the glasses can translate their voice and even make itsoundlike the person speaking, but the display versions of the glasses will offer accompanying text to read as well.Courtesy of GoogleCourtesy of GoogleThe versions with a display built into the lenses should be especially helpful with things like turn-by-turn navigation, since you’ll be able to see exactly where to turn left on the display as you walk. Google says the glasses don’t have GPS and rely on your phone’s GPS, but the cameras on the glasses utilize Google’s Visual Positioning System to identify surroundings and ensure your blue dot is in the right spot—something you can currently do in the Google Maps app to calibrate your compass.I tapped and held on the right arm of the glasses, and this triggered Gemini Live. That’s the default experience, allowing you to have a conversational back and forth with Google’s assistant.Gemini did a good job of only listening to my voice in these encounters and ignoring people who were speaking to me. I was able to ask it what the board game next to me was—it correctly identified it as Chinese checkers—and then asked if I wanted to learn how to play. I asked it to just save me those instructions in a Google Keep note, and voilà, it generated that advice in a Keep note within seconds.Photograph and composite by Julian Chokkattu/WIREDNext, I was able to ask Gemini to take a picture and then alter the photo post-capture. I first asked it to remove a plant in the photo, then I asked it to change the room’s decor to a medieval hall. It did those things remarkably well.You’ll get the original image on your phone and a preview on Wear OS smartwatches if you’re wearing one (or a preview on the display-laden glasses), and the doctored version will arrive within 45 seconds once Google’s Nano Banana platform does its thing.I was able to see a few generative widgets and scroll through them on these glasses—that’s a big theme Googleis introducing in Android 17and other platforms—so you can have glanceable information on your glasses for whatever you want. I wasn’t able to see much else of the user interface, though.Project AuraPhotograph: Julian Chokkattu/WIREDA different kind of glasses experience, Project Aura is more akin to mixed-reality headsets like Apple’s Vi

Reece RogersGearMay 19, 2026 1:45 PMGoogle Makes It Easy to Deepfake YourselfGoogle’s overhaul of its AI creation software, Flow, includes a new video model and a tool for generating selfie videos called avatars.Courtesy of GoogleCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyA wave ofdéjà vu washes over me as Elias Roman, vice president of product management atGoogleLabs, demos a new “avatar” feature for Flow, the company’s tool that lets users generate and remix AI videos and images. He previously scanned his likeness to create a digital clone of himself. Now, he can insert himself into any AI-generated videoclip he wants using Google’s new Omni Flash model.“This is for creators who want to bring themselves into their content but don’t want to have to shoot themselves,” Roman says.Courtesy of GoogleThis specific style of social-first, selfiedeepfakeis reminiscent of a quintessential feature from OpenAI’s now-defunctSora app—rather than cameos or characters, Google calls them avatars. These avatars are also available through the Gemini app andYouTube. Google announced the new feature at its annualI/O developer conferencein Mountain View, California.Google launched Flowlast yearunder its experimental Labs division. “Google has never had a product line for creative work before,” Roman says. “Productivity, definitely. Developers, absolutely. Video consumption, yes. Not for creative work.” He sees this as Google’s attempt to build tools for the next generation of creators.Similar to other announcements from Google I/O surrounding Google Search, many of the new changes toFloware part of the company’s larger attempt to makeAI agents, essentially automated software taskmasters, andvibe coding, building bespoke features with natural language prompts to AI, more mainstream for a broader audience. For example, users can repeat custom instructions when generating videos and create automated workflows that sort similarly styled clips into folders.Courtesy of GoogleOne of the most immediately noticeable changes to Flow is the new video-generation model powering the experience: Omni Flash,succeeding Veo. Similar to howGoogle’s Nano Bananamodel brought more context about the world into the AI image-creation process, the Omni Flash model overhauls video generation with richer detail throughout clips.Flow users can generate characters in AI videos with more consistency via the Omni Flash model. Roman says this is a major improvement over the weakness in past versions of Flow, where created characters could warp during successive video generations. Also, a key character that Flow users can now generate in an AI scene after an AI scene? Themselves.Users set up an “avatar” of themselves by going into the settings of their Flow account and scanning a QR code on their phone. Then, Google asks users to record themselves saying a string of numbers aloud and move their head around to capture every angle. This selfie-capture style will feel familiar to anyone who signed up for the Sora app, which OpenAI launched last year as an AI-firstsocial mediaplatform where people can generate and share clips of themselves. OpenAIstartlingly wound it downafter less than seven months.Unlike the Sora app, where users could generate videos of other users depending on the person’s settings, Google’s initial focus with its avatars is to let users create AI versions of themselves only, not other people. Every video generated with the Omni model, including those with your avatar, includesGoogle’s SynthID watermark.“You can capture your voice and your visual identity from multiple angles and have that show up with pretty high levels of fidelity,” Roman says. He generated a tongue-in-cheek video of himself teasing the Flow team in front of a dumpster fire, with an AI version of himself that looked lifelike and sounded like him. Then he used Flow to request changes to the generation, like a new background setting and a different-colored shirt, and Omni Flash adjusted the clips while preserving the avatar’s details.This isn’t the first time Google has rolled out a version of self-controlled deepfake video tools for creators—last month,YouTube Shorts addeda limited option for users to make similar AI avatars that can be inserted into clips on that platform. Other Silicon Valley companies are also looking for ways to transform creators’ outputs using generative AI. For example, last year, Meta rolled out an AI feature that can seamlessly translate Instagram Reels into different languages, evenadjusting creators’ lipsto match the different voices.While these AI tools may streamline aspects of the content production pipeline for creators—you don’t even have to get out of bed now to generate sassy vertical videos—generative AI isincreasingly polarizingaudiences who see these videos as inauthentic or not aligned with their values. Well, that’s if they actually clock the videos as AI.

Will KnightBusinessMay 19, 2026 1:45 PMDemis Hassabis Thinks AI Job Cuts Are DumbThe CEO of Google DeepMind tells WIRED that companies should use the productivity gains of AI to do more, not lay people off.Photograph: Chris Jung/NurPhoto/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyDemis Hassabis, theCEO ofGoogle DeepMind, is keen to talk about the coding skills of his company’s newest model,Gemini3.5 Flash. The model has been trained to perform complex agentic coding tasks: translate large code bases from one language to another; find and fix bugs lurking deep in knotty code; and even write entire operating systems from scratch.Hassabis does not, however, think this spells doom for software developers. “I have no idea why people are going around talking with certainty about that,” Hassabis tells WIRED ahead of the new model reveal attoday’s Google’s I/O event.“Perhaps there is an ulterior motive for putting those messages out; raising money or whatever,” Hassabis says. “From my point of view, from DeepMind and Google’s point of view, if engineers are becoming three or four times more productive, then we just [want to] do three or four times more stuff.”Got a Tip?Got a tip? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at wak.01.The striking coding abilities of the latest models has led to widespread fear that AI may be on the brink of eliminating programming roles and other white-collar jobs. Executives at some AI companies have predicted widespread job displacement, while some prominent tech companies, includingAmazon,Salesforce, andBlock, haveblamed recent layoffson the use of AI.Hassabis thinks that Alphabet, which oversees several companies besides Google, may be well positioned to take advantage of a revolution in software productivity. “I have a million ideas, from lab drug discovery to game design,” he says. “I’d love to have some free engineers to go and do those kinds of things.”Hassabis says that companies looking to replace developers with AI may be making a big mistake. “I think it’s a lack of imagination—and a lack of understanding of what’s really going to happen,” he says.Google revealed a raft of AI stuff at its annual developer event. Through a coding tool called Antigravity, Gemini 3.5 Flash offers frontier coding and reasoning capabilities but is faster and cheaper than the offerings of its competitors, Google says. Gemini 3.5 Pro, a more powerful new version of its flagship model, will debut next month.The company needs to catch up when it comes to AI coding, which has emerged as a crucial and lucrative application for the latest AI models. Anthropic and OpenAI lead developer adoption with their respective tools, Claude and Codex, according toa 2025 Stack Overflow survey.The company also demoed an agentic assistant called Spark that lives in Google’s Cloud and has access to its apps. The design is meant to be safer than something like OpenClaw because it has limited access to personal data, Google says.Other agentic demos included a version of Android with an AI agent built in and a refreshed version of Google Search that uses agentic coding to generate a site or app on the fly in response to a search query.AI coding has captivated the AI world in recent months, even inspiring hope that models could one day rewrite their own code in a self-improvement loop. Hassabis says it’s possible but doubts that it will immediately lead to superhuman-level AI.Progress in other areas of science might require AI models to have a deeper understanding of the physical world and even an ability to perform experiments within it, he says.Even within the seemingly solved world of coding, Hassabis says it is notable that AI has yet to produce a blockbuster app or video game without human help. “I think there’s something missing,” he says.

Reece RogersGearMay 19, 2026 1:45 PMGoogle Search Goes Agentic—and Doesn’t Need You AnymoreVibe-coded results! Super widgets! Bots that never sleep! Google’s vision for the future of Search is hyper-personalized, automated, and extremely AI.Photograph: Thilina Kaluthotage/NurPhoto/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAI agents areeverywhere. Every briefing I’ve attended for software companies over the past year has involved some mention ofagents—using generative AI tools to automate digital tasks. Despite breakout moments at the start of 2026, like the pluckyOpenClawagent that early adopters used to manage their online life, most people are not yet embracing this style of automation day-to-day. That won’t last for long if Google gets its way.AtGoogle I/O, the tech behemoth shared its vision to make its popular search engine a core way to expose billions of existing users to the company’s agentic prowess.“You will be able to create, customize, and manage multiple AI agents for your many tasks, right in Search,” says Liz Reid, who leads Search atGoogle. She gives the example of setting up an agent to track stock market trends and send you alerts using real-time data—when specific conditions are met.Alongside these agentic additions coming to Search, Google also announced a new underlying model, Gemini 3.5 Flash, as the global default model forAI Mode answers,as well as improvements to the Search box that make it more responsive to user inputs.Always-On AnswersOne of the key agentic experiences rolling out is a new form of dynamic data gathering through “information agents” that can be more proactive than previous search experiences and automate alerts through your conversational requests.“Ask Google to just keep you updated on anything, and now our agents can do work for you even if you’re not using Google,” says Robby Stein, a vice president of product for Search. “So, you could be asleep, and it’s still helping you.” This feature will arrive first for subscribers toGoogle’s AI Pro and Ultra plansthis summer.In Google’s example of how these can work, a user asks AI Mode through the Google app to keep them “updated when any of my favorite athletes announce sneaker collabs or signature drops.” Then, AI Mode generates an info agent unique to that user, designed to monitor the request. When a new shoe drops that fits these criteria, likeA’ja Wilson’s pink Nikes, the user gets a notification alert with critical context and ways to buy the sneaks.Booking agents are another automation style in Search that Google will be expanding this summer. At previous I/O conferences, Google debuted similar AI features, like the now defunctDuplex, which called companies on your behalf to set up restaurant reservations or salon appointments. Google continues to iterate on that core idea with agents that can search for relevant context about local companies, even calling that barber down the street for a price quote on a beard trim if it doesn’t show up on the website, to help users gather info with less active participation.Much of this keeps users trapped withinGoogle’s ecosystemwithout actually browsing the web themselves, a practice we’ve seen before when features likeAI Overviewslaunched in 2024. Even so, Google has explicitly stated that it’s not attempting to replace web page links.Vibe-Coded ResultsGoogle first launched Antigravity, its version of an agentic coding tool likeOpenAI’s CodexorAnthropic’s Claude Code, late last year. Now, an updated version of Antigravity will be used to create bespoke outputs in Search, like an adjustable visualization of black holes to help you better understand deep space.“Search can build you custom experiences,” Stein says. Google’s vibe-coding tool, Antigravity, will craft unique answer layouts, when helpful, and incorporate interactive visual elements, such as infographics and movement simulations. The update is expected to arrive later this summer for all users.An expanded version of that tool, which Google says is a way to create “super widgets” and “mini apps,” will be available first for Google AI plan subscribers in the US this summer. (I vote to call them super widgets, which sounds more unhinged.) These can be used for tracking longer-term plans, like a healthy diet, or managing complex tasks, like a cross-country move, through Search.Search Doesn’t Need YouGoogle of yore, with its 10 blue web links, asked users to be fairly active participants in the info-gathering process:Here are a bunch of webpages that appear relevant to your search. How about you click around and see which site has the best info?The process was clunky at times and could be time-consuming, but I was personally hopping site to site as Google nudged me into different online rabbit holes.The launch ofAI Overviews, which generates summaries of web pages and online information, two years ago, and the rollout ofAI Mode, the chatbot-style search, last year, have au

Paresh DaveLauren GoodeZoë SchifferBusinessMay 19, 2026 3:00 PMMeta Employees Are Scrambling to Use Up Benefits Ahead of LayoffsOn the eve of about 8,000 jobs being cut, employees are cashing in on headphone stipends and other perks while they still can.Photograph: Benjamin Fanjoy/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAhead of Meta’slatest round ofmass layoffs tomorrow, some employees are deserting offices, abandoning their work, and loading up on perks they might soon lose, several people at the company tell WIRED.Two employees describe a widespread rush to use up an annual $2,000 flexible benefit, which can cover a variety of expenses including health and wellness activities. A separate triennial credit of $200 toward the purchase of audio gear has led to a scramble to purchase Apple AirPods and other headphones. Another source says Meta offices have been largely empty this week, as people prioritize polishing their résumés and gather offsite to commiserate with friends for what may be their final time as colleagues. Employees are variously “paralyzed,” “coasting,” and “panicked,” sources say.Meta plans to lay off about 10 percent of its nearly 80,000 employees on Wednesday, with notices going out to affected workers’ personal and corporate email addresses at 4 am Singapore, London, or San Francisco time depending on their location, according to a company-wide memo sent on Monday. The cuts are coming at a time when the social media giant behind Instagram, WhatsApp, and Facebook is enjoying record-high profits.But CEO Mark Zuckerberg insists that the company must free up cash to invest in AI data centers, and that Meta can perform just as well with fewer employees because of AI technologies that augment human labor.Got a Tip?Are you a current or former Meta employee who wants to talk about what’s happening? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at peard33.24 and ChaoticGoode.12.Meta didn’t immediately respond to a request for comment for this story. The company has undergone three previous large rounds of layoffs since 2022, including as part of Meta’s one-time “year of efficiency” drive in 2023. But even though the latest round is smaller than a couple of those, it is drawing widespread scrutiny because it comes at a time of societal anxiety aboutAI’s impact on jobs.Inside Meta, the imminent cuts are among several concerns that have sunk morale to unprecedented depths, according to 16 current and former employees whorecently spoke to WIRED. Employees also have been frustrated by being “drafted” onto a new AI team without any choice andthe rollout of surveillance softwarethat tracks US workers’ laptop use to train AI models.Meta also plans to internally restructure as it conducts sweeping layoffs, transferring 7,000 remaining staff to “AI initiatives” and converting more managers into individual contributors. That would bring the total number of those affected—either laid off or placed in a new role—to 20 percent of the current workforce, Reutersreportedon Monday. WIRED independently confirmed this reporting. Some parts of the company have been told they won’t be affected at all.But in recent days, employees who are bracing for changes have shared checklists internally about benefits to take advantage of, and are saving documents such as performance reviews and pay stubs, according to one worker. Some teams are meeting up at bars and restaurants near Meta offices in New York and Menlo Park on Tuesday and Wednesday to eat and drink away their sorrows, several employees said. Management has encouraged employees not to come into offices on Wednesday.Update, May 19, 11:40 PM EDT: WIRED corrected the time zones when layoff notices will be emailed.

Boone AshworthMichael CaloreGearMay 19, 2026 4:00 PMEverything Announced at Google I/O 2026Google is sprucing up its Gemini models, revamping search, and enabling AI agents in everything. There are also some spiffy new smart glasses coming this fall.Google and Alphabet CEO Sundar Pichai on stage in 2025.Photograph: Camille Cohen/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle just wrappeditskeynoteaddress at its annualI/O developer event. The company showed off a swath of newagentic AIfeatures and some demos of its upcoming Android-powered smart glasses.As it has in the past few years, the spectacle largely revolved around Google’s perpetual stream of AI efforts. The company says that 900 million people use its Gemini assistant, and people have generated more that 50 billion images with Gemini.Google’s goal for 2026 is to put AI agents at the forefront of all its biggest services: Search, Gmail, YouTube, Docs, and the Chrome browser. In a demo briefing the day before I/O, Google and Alphabet CEO Sundar Pichai said the company was in a period of “hyper progress” with its AI efforts, but acknowledged that this is the part of that cycle “where people want to see real value in the products they use on a day-to-day basis.”Here is everything Google announced at I/O 2026. And if you’re wondering where all the Android 17 news is, Googletold us all about that last week.Search PartyGoogle is going all in on keeping people on Search by trying to integrate it into every part of your life. The company hasembedded its AI agents directly in Google Searchin something it calls the “intelligent search box.” This new search experience starts rolling out to everyone today.When you ask the search box questions like, “What is a black hole?” or “Are Google AI Overviews devastating the journalism industry,” it will respond with more contextual answers or use generative AI to create images or short video clips that help explain the concept. As an example, Google showed a search about black holes, and its AI agent generated a video that visually explains the process, plopping it right into the search results.Search is also getting a Generative UI feature, which creates different ways of viewing information from search results, so different types of responses—videos, images, news articles—get custom layouts that are generated in the browser on the fly, based on what’s the most relevant. Generative UI will start showing up in everyone’s Google searches this summer.Search agents are also being integrated more deeply across Google’s platforms. In March, Google releasedAsk Maps, which lets users ask questions on Maps like you would interact with a chatbot.Gemini EverywhereGoogle has two new major AI model updates. Gemini 3.5 is officially being released today, along with Gemini 3.5 Flash, a pared-down but more affordable version of 3.5 Pro. Both will be available immediately in Google’s Search and the Gemini app.Google’s core Gemini app is getting a refresh, with a snazzier redesign called Neural Expressive. It has new colorful backgrounds, a new typeface, and new animations for live voice chats. It also has many more accent options for your chatbot’s voice, with regional accents in several languages.Daily Brief is an upcoming personalized digest of your day that Google wants to be the first thing you check in the morning. Daily Brief pulls info and data from your calendar, email, and other information to summarize and prioritize your plans for the immediate future. (Google tested out a similar feature in its experimentalGoogle Labsin December.)The company is also adding new Gemini-powered features to its other services. An Ask YouTube feature lets you pose natural search queries in YouTube. Ask it how to ride a bike or how to grill a steak, and an agent will then be tasked with finding a mix of relevant YouTube videos. It’ll even snap right to the points within the videos that answer your questions.Much weirder is voice editing in Google Docs, in a new feature called Docs Live. By describing with your voice what you want to write, an agent will dictate your words, generate text, pull in citations from the web, and aim to turn your stream-of-consciousness wishes into a coherent document.(Reminder: All this stuff mayeventually have ads.)For Gemini power users, Google is creating a new subscription tier, the AI Ultra plan, for $100 a month. It is also dropping the price of its top Gemini AI Ultra from $250 a month to $200.Gemini OmniGoogle announced Gemini Omni, an AI video generator akin to Sora 2. That was OpenAI’s generator that let youdeepfake yourselfbut was eventuallykilled by the company.Google’s approach is building out a far more realistic video generator that can incorporate real video and extrapolate all manner of AI-powered weirdness on top of that. Google is eager for you to turn Omni’s eye on yourself, putting your face front and center. As such, selfie videos can be modified to

Brian KahnScienceMay 19, 2026 6:20 PMCalifornia’s Wildfire Season Is Already OveractiveMajor fires are threatening homes and ecologically sensitive areas following a hot, dry winter.The Sandy Fire burns on Monday, May 18, in Simi Valley, California.Photograph: AP Photo/Ethan SwopeCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyIt’s May, butCalifornia is already getting a taste of what peak fire season could look like, as out-of-control blazes pose a danger to infrastructure and some of the most threatened trees on the planet.A combination of high winds and heat has contributed to a trio of major fires in Southern California.The largest among them is the Santa Rosa Island Fire, which started over the weekend in Channel Islands National Park after a stranded sailor used flares to signal for help. The blaze has consumed roughly 16,600 acres—almost a third of the entire island. While some structures have been lost, the biggest is a grove of Torrey pines, which are among the rarest trees in the world.Torrey pines are considered critically endangered by the International Union for Conservation of Nature. Thegroup has warnedthat the grove on the island faces “high potential risk from destructive fires.” While the Santa Rosa Island Fire burned through the grove, there’s hope that the worst-case scenario may not have come to pass.“Upon initial assessments, fire crews are observing that the fire intensity was low and that the stand remains intact,” Mike Theune, the fire information officer assigned by the federal government, writes in an email. “When safe to do so, a fire effects crew will be assigned to make a full determination of condition and any long-term effects.”While smaller, the Sandy Fire has forced thousands to evacuate in and around Simi Valley, located about 35 miles northwest of downtown Los Angeles. Fire activity increased on Tuesday morning as high winds whipped the flames across nearly 1,400 acres.The fire is only 5 percent contained, while the Santa Rosa Island Fire isn’t contained at all. Because homes and businesses are threatened, the Sandy Fire has received more aerial resources to quell the flames. The River Fire has also burned 3,535 acres in Kern County and is 15 percent contained.In the early stages of fire season, nearly 41,000 acres have burned across the state, according to the California Department of Forestry and Fire Protection. But that’s well above the five-year average of 23,380 acres burned by this date and could be a harbinger of what’s to come in the state and across the West.Record-breaking heatobliterated already-meager snowpackthis spring, leaving states primed to burn. The most recent snowpack measurements show California’s Sierra Nevada range harbors just 9 percent of its usual snowpack for this time of year, while many basins in other parts of the West have no measurable snow at all, according to federal data.Hot weather is, of course, a hallmark of climate change. Arecent analysisby the nonprofit Climate Central found that April 1 snowpack—a key date for measuring what’s on the ground—has declined across the West by 18 percent since 1955.All of which is to say that while the early season fires are bad, the West hasn’t seen anything yet.

Miles KleeCultureMay 19, 2026 6:53 PMLiterary Prizewinners Are Facing AI Allegations. It Feels Like the New NormalThree of five regional winners of the prestigious Commonwealth Short Story Prize are suspected of relying on chatbots. They’re certainly not alone.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAt first, thewinners of the prestigiousCommonwealth Short Story Prizefor 2026 enjoyed the envy of their peers. But since their works of fiction earned this distinction, these authors have found themselves facing harsh scrutiny from the literary community, with several accused of enlistinggenerative artificial intelligenceto write for them.The allegations have come from numerous readers, many of them writers themselves, expressing bafflement and dismay that the prize jury could have overlooked potential signs of inauthentic authorship.Each year, the Commonwealth Foundation, a nongovernmental organization in London, awards its short story prize to one writer in each of five regions: Africa, Asia, Canada and Europe, the Caribbean, and the Pacific. One overall winner is then selected from that short list. Regional winners take home £2,500 (about $3,350), while the top winner, to be announced next month, claims £5,000 (about $6,700).On May 12, the respected UK literary magazineGranta published the top five 2026 entries—all previously unpublished, per the rules of the contest—on its website. (It has hosted the winning submissions for the prize since 2012.)Within days, however, one entry aroused suspicion. “The Serpent in the Grove,” a story by Jamir Nazir of Trinidad and Tobago, which had taken honors for the Caribbean region, struck a few people as bearing the stylistic tells of AI-generated text.“Well, this is a first: a ChatGPT-generated story won a prestigious literary prize,” wrote researcher and entrepreneur Nabeel S. Qureshi, a former visiting scholar of AI at the Mercatus Center at George Mason University, in aposton X on Monday. “‘Not X, not Y, but Z’ sentences everywhere, the ‘hums’ trope, and plenty of other obvious markers of AI writing. A major milestone for AI, at any rate…”“They say the grove still hums at noon,” Nazir’s mysterious and atmospheric tale begins. In his screenshot of the opening paragraphs, Quereshi highlighted the second line as what he considered to be a signature example of AI syntax: “Not the bees’ neat industry or the clean rasp of cutlass on vine, but a belly sound—as if the earth swallows a shout and holds it there.”As the literary community undertook a closer read of Nazir’s story, many criticized itslanguage and metaphorsas nonsensical, wondering how the Commonwealth judges could have seen any merit to them. Others sharedscreenshotsshowing that the AI-detection tool Pangram flagged “The Serpent in the Grove” as 100 percent AI-generated, a result that WIRED independently confirmed. (While no AI-detection software is perfect,third-party analysishas consistently determined Pangram to be the most accurate, with a near-zero rate of false positives.)Nazir did not return a request for comment relayed through an email address listed on hisFacebook page. The posts on that account and theLinkedIn profileof a Jamir Nazir in Trinidad and Tobago also scan as AI-generated on Pangram. Although some speculation had it that Nazir himself could have been an entirelyAI-created persona, a2018 articlein the Trinidad and Tobago edition of the The Guardian about his self-published poetry collectionNight Moon Love—which includes a photograph of Nazir holding the book—suggests that he is a real person.WIRED contacted both Granta and the Commonwealth Foundation about Nazir’s story; neither commented directly, but both issued public statements.‘We are aware of allegations and discussion regarding generative AI and our Short Story Prize,” wrote Razmi Farook, director-general of the Commonwealth Foundation, in astatementon the organization’s website. “We take these claims seriously and are committed to responding to them with care and transparency.” Farook defended the judging process for the prize as “robust,” with multiple rounds of readers and the top-level judges selected for their “expertise.”“We do not currently use AI checkers in our judging process, because this is a prize for unpublished fiction,” Farook explained. “To supply unpublished original work to an AI checker would raise significant concerns surrounding consent and artistic ownership. We also do not use AI to judge stories at any stage of the process. When they submit stories to the Prize, writers accept our entry rules and guidelines. These include confirming that their submission is their own original work. All short-listed writers have personally stated that no AI was used and, upon further consultation, the Foundation has confirmed this.”Theentry and eligibility rulesfor the 2026 Commonwealth Short Story Prize make no mention of artificial intelligence, stat

Molly HigginsGearMay 20, 2026 1:00 AMSave 30% With VistaPrint Coupons for MayFrom personalized gifts to business essentials, WIRED can help you save with our selection of VistaPrint promo codes.Save StorySave this storySave StorySave this storyVistaPrint specializes incustom physical and digital marketing products for small businesses, like brochures and promotional products, along with personal products like invitations, t-shirts, business cards, or even posters to meet any print and online personalization needs you may have. Fortunately, you don’t have to pay an arm and a leg to shop for these trending products, thanks to our selection of deals for both new and returning customers.Use Top VistaPrint Promo Codes for up to 20% OffWe have a VistaPrint coupon to help you save big on those personalized gifts for yourself or your loved ones. Right now, new customers get 20% off their first order of $100 or more with VistaPrint promo code at checkout. Once you’ve personalized whatever you want—from greeting cards to full on canvas-sized pictures and pillows—paste the code into the box at checkout to save major coin on your purchase.15% Off All VistaPrint OrdersOne of the easiest ways to save at VistaPrint is by signing up for their email newsletter. Once you sign up, you’ll get a VistaPrint promo code for 15% off your first purchase. Along with the VistaPrint discount code, you’ll get special offers, news about product launches, services, and new design alerts.Use Vistaprint Business and Get 20% OffLaunching a new small businessis one of the toughest, most daunting tasks a person can take. Running a new business requires meticulous detail, on everything from big picture items to the smallest tasks. Visaprint wants to soften the blow a bit, by offering 20% off your first order on necessary tools and products with Vistaprint promo codeNEW20. Vistaprint has valuable tools like a business name generator to help spark ideas, AI generated help with logos, and help getting visibility with Yelp. Be sure to check out Vistaprint’s tools to helplaunch your businessand use this Vistaprint promo code for an extra 20% off—when you’re launching a new business, every cent saved counts.Discover More Verified VistaPrint Coupons and DiscountsThere are other VistaPrint promos, like $10 off purchases of $100 or more, $20 off $150 or more, and $50 off $250 or more.Besides using one of our hand picked VistaPrint voucher codes, you can get 15% off your next order when you sign up for texts. VistaPrint also offerspremium membershipswhich renew annually and are ideal for small businesses and people who use services frequently. Members get up to 40% off products, 30% off Deposit photo subscriptions, unbranded products and packaging, and global discounts. Best of all, you can try these services free for 30 days. Be sure to check out our VistaPrint promo codes and discounts page to see the rotating available discounts.Discover the Latest Vistaprint Postcards With Up to 30% OffForget the old holiday or business card and instead upgrade for contemporary times with Vistaprint postcards. They have tons of styles of postcards to fit your business, ideas, and budget. They have a huge range of styles and sizes, from coupons to mailers, in various price points, with classic designs to contemporary styles.Redeem Your VistaPrint Promo Code in a Few Easy StepsWhether you’re looking for the best deal on custom VistaPrint business cards for your small business or you’re looking to order custom t-shirts for your family reunion, it’s simple and easy to apply a coupon for your next VistaPrint order. All you have to do is copy the code and paste it in the allotted area at checkout. Be sure to also look out for other trending VistaPrint deals that don’t require a code to redeem your savings on your purchase. In the case of a limited-time promotion, just add the promo item to your cart to instantly save.Find Everything You Need With VistaPrint Custom ItemsVistaPrint has been with me through the years—for thank you cards, holiday gifts, mugs of my friends’ faces as gifts, calendars to make myself a centerfold, familyphoto books, wedding invitations and subsequent divorce party invites. What can I say, I live a charmed life—and VistaPrint has been there through it all. It has been my consistent go-to forpersonalized productsfor the past decade, offering good prices on all my odd-ball creations. So stock up and discover discounts on a selection of marketing must-haves like vinyl banners, business cards, posters, custom mugs, and more. Go ahead, check it out, you know you’ve always wanted to eat off a plate or sleep on a pillow with your loved one’s face on it.Customize Your Vistaprint Wedding Invitations This MonthVistaprint makes it easier (and cheaper) than ever to get big printing jobs done for less, with the ability to personalize things as much as you want. You can save big onVistaprint wedding invitations, with 100 starting at just $90 (half off!);Letterpr

Molly HigginsGearMay 20, 2026 1:00 AMTop Stearns and Foster Coupons: $300 Off in 2026Discover the best Stearns and Foster promo codes, discounts, and deals to elevate your sleep experience with premium comfort and quality.Save StorySave this storySave StorySave this storyFor 180 years, Stearns & Foster has been makingmattressesbefore expanding intobedding,pillows, andframes. Throughout their history, the brand has continuously brought its trademark old-school style of craftsmanship. Now owned by Tempur Sealy, the iconic Stearns & Foster look features hand-stitched details and an elevated design seen in luxury hotels like the Ritz-Carlton. Today, Stearns & Foster is known for using premium materials, hand-tufted covers, patented IntelliCoil technology with an innerspring-in-innerspring design, moisture-wicking Tencel, and memory foam for pressure relief. If you’re looking for reliable, hotel-like sleep night after night for decades to come, investing in a Stearns & Foster mattress is a good idea. This luxury brand is on the pricier side, so we have a Stearns & Foster promo code and Stearns & Foster coupons to make this investment more affordable.$300 Off With a Stearns & Foster Coupon CodeFor a limited time, during the Sterns & Foster Memorial Day Event, you can save up to $700, plus get $300 in free accessories when you buy a mattress (besides the Studio model). With this Stearns & Foster coupon code, you’ll get up to$300 in free accessoriesor a free flat foundation with a mattress purchase. Just use Sterns & Foster promo code300FREEat checkout.Stearns & Foster Promo Code: $200 Off BaseThe Ease Power Base is Sterns & Foster’s bestsellingsmartbedframe. Rated 4.7 out of 5 stars with over 2,000 five-star reviews, people love the adjustable frame to personalize their sleep experience, with a wireless remote to change their base position. The frame even has a Zero Gravity setting to help the body feel weightless with pressure relief for aches and back pain. When youpurchase a Ease Power Base, you’ll get complimentary white-glove in-home delivery and a 25-year warranty for extra peace of mind with a purchase made to last decades. Plus, right now, you’ll get a $200 instant credit when you buy a qualifying mattress with Stearns & Foster promo code200GIFT.Get Free Pillows With Stearns & FosterFor a limited time, when you buy a qualifying Sterns & Foster mattress, you’ll get up to $300 in free accessories, includingpillowsandsheets. Perhaps best of all, there’s no catch:This Stearns & Foster dealautomatically applies to your cart when you purchase a qualifying mattress. If you’d rather customize your free accessories, you’ll need to remove the pre-created bedding bundle from the cart and manually add the accessories of your choosing to the cart. Use Stearns & Foster coupon code300FREEfor this offer.Stearns & Foster Free In-Home DeliveryOne of the benefits of buying from a high-end brand like Stearns & Foster is their white glove delivery service and 25-year warranty for even more peace of mind. As amattress revieweron the WIRED Gear team, I hauled over a dozen mattresses in and out of my home. I can confidently say it was not fun. With thisStearns & Foster dealfor free in-home delivery, their shipping team will set up your new mattress in the room and remove and discard all packaging materials, so you don’t have to do any of the grunt work. They’ll even dispose of your existing mattress and box spring if you request it for free.

Ryan WaniataGearMay 20, 2026 1:00 AMSony Coupons: 45% Off Sony Headphones, WF-1000XM6 Earbuds, and Sony CamerasUpgrade your setup with Sony’s newest releases. Save on industry-leading noise-canceling audio, and pro-level Alpha cameras.Save StorySave this storySave StorySave this storySony makes someof the best electronics we’ve tested across a dizzying array of categories, from TVs and audio gear to cameras and gaming consoles. Sony products constantly occupy top slots on ourBest TVsandBest OLED TVslists,Best Wireless Headphones, andBest Camerasguides. If you’re shopping for products from any of those categories, you can pay a little less with our Sony promo codes for deals like 45% off Bravia Televisions, 30% off Sony headphones and earbuds, 15% off cameras and lenses, and more.Sony Promo: 45% off Bravia TelevisionsSony Bravia models rank among thebest TVswe’ve reviewed, from premium models like the revolutionaryBravia 9 QLEDand brilliantBravia 8 II OLED, to the more mid-tier Bravia 5, all of which are available on hot Sony deals right now, with this newpromo for up to 45% off.Use Sony Coupons and Save 30% off Sony Headphones and EarbudsSony has been synonymous with portable audio since the Walkman, and wireless headphones like the WH-1000XM series offer great performance and durability. We’re constantly putting them atop our list of thebest wireless headphonesthanks to excellent sound, feature-rich design, and noise-cancelling that ranks among the best in the business. With Sony online coupons, you can get great deals on Sony’s latestWH-1000XM6headphones, which we loved for their upgraded sound and class-leading noise canceling, as well as the still-great previous generation, theWH-1000XM5.Looking for earbuds? Sony’s nimble, noise-killingWF-1000XM5are also on sale, as are plenty of other options from the brand’s diverse lineup. Click the link to get30% off Sony headphonesand earbuds goodness with this Sony promo code and Sony coupons.Up to $100 Off Sony Gaming GearRight now, Sony has awesome discounts and deals on some of the most essentialgaming tech, like headphones, monitors, and earbuds. You can get up to $130 offINZONE H9 Wireless Noise Canceling Gaming Headset, and $20 offINZONE Buds Truly Wireless Noise Canceling Gaming Earbuds, INZONE H5 Wired Gaming Headset, and more. If you’re looking to upgrade your gaming monitor, you can get up to $100 off a 27 inchINZONE M9 II 4K 160Hz Full Array LED HDR Gaming Monitor.Sony Discount: 15% off Sony Cameras and LensesSony makes some of the best consumer cameras on the market. In fact, we recently named theSony A7Vthebest mirrorless camerayou can buy, and the previousA7 IVwas similarly fabulous. Both are on great sales through these new offers, letting you grab serious image quality with15% off our favorite Sony cameras and lenses.Get 10% off or $25 off Sony TVs on Your First Purchase When You Sign Up for EmailsLooking for even more of thebest TVsfrom Sony? Sign up for email alerts, and you couldsave 10% on a Sony TVor$25 off other Sony products.Become a Sony Member to Earn PointsLove a great Sony deal? Join the My Sony Membership Program and you could earn Sony’s My Points rewards toward more good stuff by joining, making eligible purchases, or through other Sony promotional opportunities (full terms available here). It’s a free way to earn credits or receive other special benefits or offers. Join theMy Sony Membership Programand start earning points today.

Nena FarrellGearMay 20, 2026 1:00 AMHerman Miller Promo Codes: Save Up to 40% on Sale Items in May 2026Whether you’re looking for a Herman Miller promo code or discounts in their sale, here is how to save on the world’s best ergonomic office furniture this month.Save StorySave this storySave StorySave this storyIf you’re lookingfor home and office furniture that will be part of your life for the long haul, Herman Miller is a go-to for well-designed, long-lasting furniture. Whether it’s upgrading your work-from-home setup with a better office chair or amping up your curb appeal with stylish new house numbers, Herman Miller can help. There are several pieces from Herman Miller that we love here at WIRED: the Herman Miller Embody is our “buy it for life”office chair, while Herman Miller’s Spout Sit-to-Standstanding deskis where I’m currently writing from. The brand is known for its thoughtful, high-quality pieces, but that quality level comes with a price tag to match. Luckily for you, you can pick up your own Herman Miller pieces for a little cheaper thanks to a Herman Miller promo code or a discounted Herman Miller bundle.Up to 40% Off Markdowns This May During the Herman Miller SaleWant the biggest Herman Miller discount you can get? You can save up to 40% on Herman Miller’s in-stock, ready-to-shipsale furniture and accessories. There’s a range of stylish home decor available now, includingsleek house numbersI wish I could adhere to my home (you can thank my HOA for keeping my curb appeal limited).Save 15% Instantly on the Better Home Office BundleRefreshing your office? Say no more. Herman Miller has abundle, aptly named the Better Home Office Bundle, that lets you bundle your choice of office chair, desk, and third item (a storage option, table lamp, or office accessory) to get 15% off the entire purchase. Use the Herman Miller promo codeBUNDLE15once you’ve put all three items in your cart.Claim Free Delivery on Herman Miller Orders Over $2,000Need it delivered? Get free delivery with the Herman Miller promo codeFORYOUfor any purchase over $2,000. You might worry you’ll need to purchase multiple items to hit that threshold, but many high-quality office furnishings (likeour favorite office chairsfrom Herman Miller) will easily hit the minimum for you.Unlock a First-Order Discount as a New Herman Miller BuyerFirst-time shoppers at Herman Miller also get a special discount. You’ll need to sign up forHerman Miller’s newsletter, but a quick newsletter signup is well worth a Herman Miller discount code.Multiply Your Savings by Stacking Herman Miller Sale OffersIf you wish, you can get multiple discount coupons going at the same time on Herman Miller’s site. You can stack Herman Miller’s coupon codes—like the one you’ll get for signing up for the newsletter, or bundling three items together—with seasonal sales to get multiple discounts at the same time. Memorial Day is right around the corner, after all, andHerman Miller’s site already has sales going.

Scott GilbertsonGearMay 20, 2026 1:00 AMTop NordVPN Coupons: 77% Off, Plus 3 Months Free in MaySave up to 77% on 2-year plans and get 3 free months with our NordVPN discount codes.Save StorySave this storySave StorySave this storyWhether you’re worriedabout the open network at your local coffee shop or want to get around geo-restrictions when you’re traveling, NordVPN can help. Avirtual private network(VPN) is like a protective tunnel, keeping your data safe from prying eyes. VPNs aren’t perfect protection (you need to trust your VPN provider), but they’re always going to be better than nothing when you’re using a public hotspot. They also make it easy to continue to watch your favorite shows even when you’re on vacation. If you want to learn more about how to protect yourself online, see our guide toThe Best VPNs.Get 77% Off Plans With Exclusive NordVPN Discount CodeNordVPN is pretty much always having a sale, but some of the deals are better than others and this is one of the best we’ve seen. Right now, NordVPN is offering 77% off its regular price, plus an Amazon gift card valued up to $50 when you sign up for a 2-year Prime plan. It’s valid for the first 15 months of the subscription, with threat protection, high-speed VPN, and more. Follow the link above and use our exclusive coupon code at checkout to get the deal.Special Offer: 74% Off With an Amazon Gift CardRight now, you can get up to 74% off a NordVPN plan, plus get an Amazon gift card for free! Plus, when you get NordVPN, you’ll have a 30-day money-back guarantee if you realize the service isn’t for you. To get theAmazon gift card reward, you just need to buy any of the NordVPN 2-year Plus, Complete, or Prime plans; and the bigger the plan you purchase, the bigger the reward you’ll receive. Once bought, you’ll get an email sent to your inbox with all the details, then you’ll claim your gift card, and need to remain a NordVPN subscriber for 31 days. Then, your Amazon Gift card will be mailed 31-50 days after purchase.Save More With a NordVPN Subscription PlanThe two-year Plus plan is the most popular, and our recommendation. For a limited time, their more comprehensive Prime 2-year subscriptions qualify for up to 77% off, and you’ll receive 3 free months with select 2-year plans, starting with the Plus option.NordVPN offers different plans for different needs, including one-year and one-month plans.Plans’ pricingvaries greatly, with short-term, one-month plans starting at $13. The plans get cheaper the longer you commit to a plan, to reward loyal VPN customers. One-year plans start at $5 and go to $10, depending on which package you choose. Two-year plans are even cheaper, starting at $3 and going up to $7 for the top-tier plans with all of the bells and whistles. It makes a lot more sense to invest in a longer protection plan to keep your information and devices safe, for less.How Do I Buy A NordVPN Subscription Plan?If NordVPN is the right service for you, the best way is to order straight from their website because it’s the cheapest price and you’ll get all of the perks from buying directly from the brand. Once you’ve picked the bestservice for you, click “Continue to Payment.” After, all you need to do is input your email address, select your payment method, and if you’ve nabbed a NordVPN promo code from WIRED, input that in the box at payment and watch the price slash.Score a NordVPN Discount Code for 15% OffThe coupons aren’t the only deals to be had. You can wait for a sale (see above) or if you qualify, there are discounts for students, graduates, teachers, healthcare, military, and first responders. Once verified with Student Beans, you can unlock a NordVPN promo code for 15% off plans.NordVPN Friend Referral DiscountIf you like NordVPN, you can save yourself even more with the refer-a-friend promotion. You’ll receive 3 months free for every friend who buys NordVPN through your referral link. The more friends you refer, the more rewards you’ll earn. Your friends will get the same deal you did, 3 extra months if they choose the 1- or 2-year plans or 1 free month if they opt for the monthly plan.Can You Get a NordVPN Free Trial?Curious about NordVPN services but don’t want to make the jump? You’re in luck. NordVPN offers a30-day free trial. To try it out, you’ll need to pick the plan that’s right for you, anddownload the NordVPN appon your device. NordVPN also has a 3-day free trial on Android. And you can try NordVPN risk free for a month: you’ll need to pay for a subscription up front, but if NordVPN isn’t the right fit, you can request a refund within 30 days of your date of purchase.Get the latest NordVPN PricingNordVPN has four tiers of plans fit for every type of need. Each gets you a secure VPN, but with each level you’ll get even more perks, like an ad tracker and blocker and 1 TB of Cloud storage. Right now, you can get up to 77% off each plan, including three months free. Check out which plan (and price) is the best fit for youhe

Ryan WaniataGearMay 20, 2026 1:00 AMTop LG Promo Codes and Coupons for May 2026Save 20% with an LG promo code today, plus up to $1,000 off appliances, 40% off bestselling TVs and monitors.Save StorySave this storySave StorySave this storyLG makes someof thebest TVsyou can buy. ItsOLED TVsin particular are perennial favorites at WIRED, with C-series models like the C4 (9/10, WIRED Recommends) providing among the best performance for your dollars on the market. LG is about way more than TVs of course. The Korean brand offers multiple products across the A/V landscape, fromsoundbarstoBluetooth speakers, along with a host of other products to upgrade your home setup. Save over 40% with LG promo codes and online exclusive discounts on home appliances, gaming monitors, and more tech.Save 20% With Today’s LG Promo CodesIf you’re looking to save on LG gear today, you can get up to 20% back on your first order withselect eligible productswhen you sign up for LG rewards with your email address. Simply go to theMy LG Membership rewards page, follow the instructions to sign up and save.If you’re ready to check out one of those TVs mentioned above, LG’s C4 and the ultra-bright step-up model G4 are both eligible right now in 77-inch sizes. Otherwise, there are plenty of other products to shop currently on sale that you stack with MyLG Rewards for 5% back and points on every purchase.We’ve Got an LG Secret Coupon at WIREDWe love LG products, and we know you do too. This tech is oftentimes the priciest investment you’ll make in your household. So, we snagged anLG secret couponfor you, which gives 5% off your next purchase. Although it doesn’t sound like much, with a bigger appliance purchase like a laundry set or range stove, this can add up to hundreds of dollars in savings. All you need to do is sign up to become a member, and you’ll get 5% off your purchase—but act quick, the offer is only good for 24 hours.Watch for LG Fridge Deals and Price DropsOne of the most sought after appliances from the brand is theLG fridge. While we don’t have any set LG promo codes and discounts for refrigerators now, be sure to check back often and track price changes on these sought-after powerhouse appliances. TheCounter-Depth French Door Refrigeratoris a sleek, three-door stainless steel machine with a fancy ice maker and luxe mirror design. TheSide-By-Side InstaView Refrigeratoris the fridge of the future, with a translucent door to easily see what’s inside the fridge without having to open the door. The simpleSingle Door Refrigeratorfeatures a clean, compact design and easy-to-reach organization.$400 Off LG Heat Pump Dryer and Washer PairIt seems like there are smart home appliances for nearly every task in the kitchen, while other household chores are still stuck in the early 20th century. No longer! Experience the luxury of a contemporary laundry experience and take the headache out from the daily grind of cleaning clothes. LG’s DLHC5502 brings laundry tasks to the modern era, with a ventless, efficient design that’s delicate on clothes. This model has a heat pump dryer with a flexible ventless design so it can be installed basically anywhere. Right now,you can save majorlywhen you bundle it with an eligible washer for $400 off, plus professional installation for free, through April 29.Students, Teachers, and First Responders Get a 10% Off LG Promo CodeLooking for further discounts? If you’re a teacher, student, or first responder, you could receive an additional LG Promo Code for more savings. To see if you’re eligible, simply click on theLG Appreciation Program linkand follow the info to register your account. Exclusive discounts include free installation on refrigerators, cooking ranges, and washer/dryers, including select LG Studio and LG Signature appliances. You could also get year-round savings on LG products with a minimum guaranteed discount of 10% compared to sale prices you’ll find on their website.Get up to 20% Back With MyLG Rewards or 5% Off Referral CodeWith MyLG Rewards, you can get up to20% back in rewards, as well as exclusive deals, extended warranties, and free installation. Eligible rewards members can get an additional 5% off LG products if you refer a friend or loved one with theLG Referral programwhen they make an eligible purchase of $200 or more. They’ll get 5% off, and you’ll earn 5,000 MyLG Rewards points, so it’s a win-win. This coupon is stackable with other promotions too, so you can still take advantage of the many other featured discounts.Get $200 Gaming Credit With LGLG wants to reward gamers. Right now, you’ll get a $200 gaming or Mastercard credit when you pre-order the 52-inchUltraGear evo monitor. This LG gaming monitor is the world’s largest 52 inch 5K2K screen that delivers epic ultra-wide scale and UHD detail, along with a curve that wraps the action in UHD clarity across your natural view for an immersive gaming experience like never before. Plus, the built-in speakers deliver clear, well-bal

Scott GilbertsonGearMay 20, 2026 1:00 AMDell Coupon Codes: 20% Off for May 2026Get 20% off with verified Dell promo code, plus today’s coupons for up to $600 off laptops, Alienware monitors, and all things tech.Save StorySave this storySave StorySave this storyExpect to saveover $600 on the hottest models, including $250 off monitors, for crisp and clear gaming (or writing), 40% off laptops, and up to $1,200 off desktops for true gamers. We’ve rounded up the top Dell coupon codes and discounts so you can save big on pricey tech.Get 10% Off With Top Dell Coupon CodesIf you haven’t bought from Dell, right now you can get seriously rewarded for switching. Right now, new customers can get 10% off Dell products when yousign upfor their email newsletter. Plus, you’ll get additional perks like exclusive offers in your inbox and invites to special sale events, including flash discounts during seasonal sales. Plus, right now you can use Dell promo codeAFFBUNDLE10to get an extra 10% off when you bundle select PCs with accessories. And, save your tech refresh with clearance sale discounts and Dell promo codes today.20% Off With This Dell Coupon: Dell 14 PremiumAs mentioned, one of the best ways to save big on tech purchases is by buying a certified refurbished Dell laptop through the Dell Outlet. The Dell outlet has tons of like-new tech for hundreds less than if bought brand new. You can get top tech like theDell 14 Premiumfor $1,036 when bought refurbished—that’s $259 in savings! With this powerful laptop, you’ll have built-in Intel Arc 140T Graphics and FHD RGB + IR Camera and Microphone, among other perks. Be sure to check out the Dell Outlet for tons of Dell tech for hundreds less.Save $450 With Dell Coupons for Laptops, Alienware Monitors, and MoreBeen wanting to upgrade your WFH or gaming setup? Dell has marked down some of their most popular laptops and monitors during their holiday and tech refresh sale events. There are stellar deals like $400 off theAlienware 16X Aurora(one of the best gaming laptops we’ve seen of its kind), and $450 off the Alienware 18 Area-51. Complete your gaming experience with anAlienware Area-51 desktopfor $1,200 off, which includes 50% more processing power. Some other great Dell monitor deals we’ve found include the ability for Rewards members to earn up to 3 times the amount of points (or 9% back) on select Alienware monitors. Plus, you can save $100 on the Alienware 4K QD-OLED gaming monitor, or $120 on the Alienware 27 Gaming monitor, right now. And you can also get 30% off the Dell 24-inch monitor when you purchase any laptop or desktop computer.We’ve also found steep discounts on some of our favorite WIRED-tested gear, including $160 off the Bose Quiet Comfort headphones (now only $199), and $350 off theDell 14 Plus Laptop.The deals don’t stop there though, with discounts on Dell’s best-selling laptops and PCs, including: $200 off the Dell 15 laptop, $500 off the Dell 16 Plus laptop, and 50% off software when you purchase a PC. Whether you want to work-from-home or game-from-home more effectively, check out the latest offers in theDell deals hub.Dell has rotating best deals of the week too, like their best-selling monitors, which includeDell 27 Plus 4K monitors(starting at $129), $400 off QD-OLED gaming monitors like the popular Alienware 34 Curved model, plus $350 offAlienware Auroragaming desktops and $200 offTower Plusdesktop computers.10% Dell Coupon for Students, Military Members, and Medical WorkersStudents, it’s the worst time of your life, you might as well save a little money. Just kidding, life only gets harder from here. Kidding. Again. Maybe. What is not a joke is this deal. Dell is giving students a 10% discount on things students need like Dell XPS laptops and Alienware gaming rigs. No need to say anything about gaming. It’s for your eyes, because they’re starting to bleed from all the term papers.To the discount, first, make sure you have a Dell account and that the Dell Rewards box is checked. Once you’re enrolled in both, verify your .edu email address through the membership tab of your account. Not a student? No problem, Dell still has some great deals for your computing needs. Beyond students, Dell offers an extra5% off Heroes discountfor military members, veterans, employees, rewards members, and businesses. Verify through GovX or VCS in a matter of minutes to get exclusive pricing and early access to sale events.Don’t Miss Dell’s Ultimate College OfferGetting prepared for college is the most exciting, sometimes stressful, and often one of the most expensive times in a person’s life. WithDell’s Ultimate College Offers, students can save an extra 10% and get the Microsoft College Offer for free on select PCs—this includes a year of Microsoft 365, Xbox Game Pass, and a customizable Xbox controller, because Dell knows that although you work hard, you need to play a little too. This is over $500 worth of products, for free, with this limited-time Dell discount

PolicyEntertainmentAntitrustDemocrats preview how they’d go after the Ticketmaster settlement if they regain powerAnd California’s AG says an even broader breakup of Live Nation is on the table.And California’s AG says an even broader breakup of Live Nation is on the table.byLauren FeinerMay 19, 2026, 8:13 PM UTCLinkShareGiftImage: Cath Virginia / The VergePart OfUSA v. Live Nation-Ticketmaster: all the news on the lawsuit that could take down a giantsee all updatesLauren Feineris a senior policy reporter at The Verge, covering the intersection of Silicon Valley and Capitol Hill. She spent 5 years covering tech policy at CNBC, writing about antitrust, privacy, and content moderation reform.A handful of Democrats called an unofficial hearing on Capitol Hill Monday to slam the Department of Justice’s “trivial” and “pathetic” settlement with Live Nation-Ticketmaster, previewing how they might go after Trump administration antitrust deals if they win back congressional power in November.While the DOJsettlementwith Live Nation, which came one week into trial and with a promise of up to $280 million, received copious pushback, Democrats don’t at the moment have the power in either chamber to set committee agendas. That meant that Monday’s so-called shadow hearing looked different than most official proceedings. Instead of sitting at the elevated seats behind the dais, the lawmakers sat at tables below, eye level with witnesses including California Attorney General Rob Bonta, one of the AGs continuing the fight against Live Nation-Ticketmaster, ousted Trump administration antitrust official Roger Alford, and several entertainment industry players, including a member of the band The Hold Steady. Several artists asked to appear at the forum declined out of fear for their livelihoods, according to House Judiciary Committee Ranking Member Jamie Raskin (D-MD), who co-hosted the forum with Senate Homeland Security permanent subcommittee on investigations Ranking Member Richard Blumenthal (D-CT).While their statements won’t be formally entered into the congressional record, the forum gave Democrats a way to publicize their concerns, which included that the DOJ’s deal was the result of undue corporate influence. “This isn’t like some kind of pickup game,” Raskin told reporters. “The corruption permeates the administration so much that we have to start building the record because it’s such an overwhelming task. We will be using the products of all of these shadow hearings to figure out where to go.” As a signal of the significance of the forum, Senate Minority Leader Chuck Schumer (D-NY) delivered opening remarks to kick it off.“The corruption permeates the administration so much that we have to start building the record because it’s such an overwhelming task”The lawmakers pointed to the states’resounding victory in the jury verdictas evidence that theDOJ’s settlementfell far short of what a good faith deal would include, charging it was a result of corruption. On the11-page verdict form, jurors checked “yes” on every question about whether the states proved their claims. The DOJ has stood by its deal as a win for consumers. The deal will still need to undergo a Tunney Act review, which is meant to ensure that it’s in the public interest.Judges are typically loath to reject a settlement, which is why Sen. Amy Klobuchar (D-MN) hasproposed legislationto strengthen the standard. But Alford, who wasfired from the DOJandlater spoke outabout opposing lobbyist influence in antitrust deals, said this could be the rare case where the judge may reject the DOJ’s agreement, given that the same judge is now tasked with coming up with appropriate remedies to resolve Live Nation’s monopoly, potentially including a breakup.“Divestiture, it’s the nuclear option”Bonta said the states’ breakup request, expected to be formally submitted this week, may go beyond a separation of Live Nation and Ticketmaster. They’re also considering asking for at least partial divestitures of parts of Live Nation’s venue or artist management businesses. “Divestiture, it’s the nuclear option. We believe it’s warranted, but we want to be thoughtful how we ask this judge for an appropriate remedy,” he told reporters.“Opinions shared yesterday misrepresent how the live events industry operates,” Live Nation’s executive vice president of corporate and regulatory affairs, Dan Wall, said in a statement. “Secondary ticketing and high-demand onsales are industry-wide challenges that exist across every promoter, venue, and ticketing platform. That’s precisely why Live Nation and Ticketmaster have long supported industry-wide resale reform, including price caps and stronger consumer protections, while continuing to invest in tools that help artists manage demand and protect fans.”Live Nation haspromised to appealthe jury verdict, and even before it does it’s possible the judge will deny the states the full extent of the remedies they ask for. But Monday’s forum showed tha

GadgetsTechVerge ShoppingHere are our favorite Memorial Day deals (so far)We found excellent deals on some of our favorite TVs, noise-canceling earbuds, portable speakers, and other gadgets.We found excellent deals on some of our favorite TVs, noise-canceling earbuds, portable speakers, and other gadgets.bySheena VasaniMay 19, 2026, 8:25 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Portable bluetooth speakers like the Charge 6 and Flip 7 are just a few on sale for Memorial Day.Sheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.Memorial Day marks the unofficial start of summer, and the holiday’s sales include plenty of deals on gadgets that can help you make the most of the season. If your plans involve beach days, pool parties, or backyard barbecues, now’s a great time to pick up aportable Bluetooth speaker, as many of our favorite models are currently on sale. You’ll also find worthwhileTV deals, along with discounts onsmart thermostatsto help keep your home cool and your energy costs low. And, if you’re planning on heading out of town, there are deals onnoise-canceling headphones,portable chargers, and other travel-friendly gadgets.RelatedHere are 40 of our favorite deals from REI’s massive Anniversary SaleThe Verge’s 2026 college graduation gift guideBelow, we’ve rounded up the best Memorial Day deals currently available. To give you more options as you shop, we’ve included good deals from a range of retailers, including ones that aren’t running their own Memorial Day sales.Featured ProductsPreviousNextBose QuietComfort Ultra Earbuds (2nd Gen)Read More$249 at Amazon$249 at WalmartLG B5 OLED TVRead More$800 at Amazon (55-inch)$650 at Best Buy (48-inch)Anker Laptop Power BankRead More$96 at Amazon$96 at WalmartStar Wars OutlawsRead More$36 at Amazon (Switch 2)$20 at Walmart (PS5)Ray-Ban Meta (Gen 1) smart glassesRead More$224 at Amazon$224 at Best BuySonos Move 2Read More$399 at Amazon$399 at WalmartBose QuietComfort Ultra Headphones (first-gen)Read More$299 at Amazon$299 at Best BuyRing Battery Doorbell Plus (second-gen)Read More$120 at Amazon$120 at Best BuyEarbuds and headphonesBose QuietComfort Ultra Earbuds (2nd Gen)$249$29917% off$249The second-gen QuietComfort Ultra Earbuds have dynamic, rich sound and the best noise cancellation currently available in a pair of true wireless earbuds. They’re a terrific pick if you want to listen to your music in peace.Read our review.Read More$249 at Amazon$249 at Walmart$249 at Best BuySamsung’s Galaxy Buds 3 Proare on sale for $149.99 ($100 off) atBest Buy, which matches their all-time low price. The wireless earbuds are ourtop pick for Samsung Galaxy phone device owners, thanks to their comfortable AirPods-inspired design, terrific sound quality, and good noise cancellation. They also offer helpful voice controls that let you adjust volume or playback without a wake word, while delivering excellent voice call quality.Read our review.Nothing Ear (a)$59$9940% off$59Nothing’s Ear (a) earbuds have the same general design as the company’s previous earbuds — but you can get these in a snazzy yellow. They include ANC, more than decent sound quality, and other features like multipoint for just under $100.Read our impressions.Read More$59 at Amazon$59 at NothingSennheiser’s Momentum True Wireless 4earbudsare down to $199.99 ($100 off) atAmazonandBest Buy, which is $20 shy of their best price to date. If sound quality matters most, they’re an easy recommendation, delivering rich, detailed audio with EQ customization to fine-tune the listening experience. Optional wing tips are included to help keep them secure and comfortable.Read our review.Bose QuietComfort Ultra Headphones (first-gen)$299$42930% off$299The older QC Ultra were a replacement for the Noise Cancelling Headphones 700 with a more travel-friendly design, spatial audio, and better call quality. Thankfully, they still offer excellent comfort and some of the best noise cancellation available.Read our review.Read More$299 at Amazon$299 at Best BuySony WH-1000XM5$248$40038% off$248Despite being a last-gen model, these headphones still offer impressive sound quality and top-notch noise cancellation. They’re comfortable to wear for hours, can last up to 40 hours on a charge, and support extras like audio sharing and Gemini Live.Read our review.Read More$248 at Amazon$248 at Best Buy$250 at TargetAmazonandBest Buyare selling thelast-gen AirPods Max (USB-C)for $449 ($100 off), which is $50 shy of their all-time low. Apple’s luxe over-ear headphones offer a premium design and work especially well with other Apple devices. They still deliver great sound quality and effective noise cancellation, even if thenewer modelhas a slight edge.Read our review.Shokz OpenRun Pro 2$130$18028% off$130The OpenRun Pro 2 are Shokz’s best bone conduction headphones yet, with bet

TechAIReportGoogle’s AI future demands trust — and your personal dataGemini Spark is giving Google a whole new level of access to your information.Gemini Spark is giving Google a whole new level of access to your information.byEmma RothMay 19, 2026, 9:00 PM UTCLinkShareGiftImage: GooglePart OfGoogle I/O 2026: All the news and announcementssee all updatesEmma Rothis a news writer who covers the streaming wars, consumer tech, crypto, social media, and much more. Previously, she was a writer and editor at MUO.Google has big promises for its AI-powered future — and a lot of it depends on your trust. At I/O 2026, Google described a bunch of new tools that it claims will make your life easier.Gemini Spark, Google’s always-on AI agent, can help organize an upcoming event, while Daily Brief can offer a rundown of what to expect during your day. Google is even expanding access toGmail’s AI inbox, which can generate custom to-do lists and draft personalized replies based on your emails.Many of these features seem genuinely useful, but at the heart of each of them is an AI engine that runs on a trove of personal information. While other AI companies, like OpenAI, Microsoft, and Anthropic, let you connect other apps and data that you use, Gemini’s access to the personal data already stored across Google’s services lies behind a simple opt-in menu —one of its key advantages in the AI race.You can ask Gemini Spark to review your inbox at a set time each week.Image: GoogleGoogle first started dipping into personalization in 2024, when it integratedGemini into Workspaceapps like Gmail, Docs, Sheets, Slides, and Drive in 2024, allowing its AI chatbot to do things likesift through your filesordraft an email. Gemini’s Deep Research feature can eventap into your emails, Drive, and chatsand use them as sources for its reports.Over the past several months, Google has continued toexpand these integrations. Itintroduced “Personal Intelligence”in January, a feature that allows Gemini to reason across Gmail, Google Photos, Search, and your YouTube history without prompting. That means Gemini can automatically surface details from across your accounts to personalize its responses. “Millions of people are using it [Personal Intelligence] every single day, they found it so helpful for things like personalized product and trip recommendations, or acting as a thought partner for navigating big decisions in life, like a career change,” Josh Woodward, the head of Google Labs, the Gemini app, and AI Studio, said during I/O 2026.Even though it’s completely optional to connect your Workspace apps, Search history, Photos, and other information to Gemini, it seems like Google’s AI future hinges on people doing exactly that. Daily Brief, which is rolling out to Google AI Plus, Pro, and Ultra subscribers, scans for updates from your Gmail and flags events from your Calendar.ButGemini Sparkis diving deeper into your information, as Google is pitching it as an AI personal assistant that can work across connected Workspace apps 24/7, creating continually updated study guides, generating to-do lists based on meeting notes, and even automatically scanning monthly credit card statements for hidden subscription fees. But the connections to Workspace apps are just the beginning, as Gemini Spark will also be able to plug into third-party services, like Canva, OpenTable, Instacart, Spotify, Expedia, Adobe, and more.Google is even planning to give Gemini Spark access to local files on Mac computers, similar to OpenClaw, the open-source AI agent platform thatposes a range of security risks. During a demo at I/O, Woodward showed how he could use Spark to send an email to a dog boarder in preparation for an upcoming trip. He selected documents across his computer and asked Spark to craft an email using his dogs’ allergies and vaccination records.Many people might draw the line at giving an AI system access to their whole computer. But if therise of OpenClaw tells us anything, it’s that AI is moving from being a novelty to a real productivity tool that demands access to our digital lives. It’s just a matter of whether people trust the companies behind the systems enough to hand over their personal data — and more importantly, where they’ll set a boundary on what’s too private.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Emma RothAIGoogleGoogle I/O 2026ReportTechMore in:Google I/O 2026: All the news and announcementsGoogle is launching an Android version of its AI Studio vibe coding tool.Jay Peters12:14 AM UTCWear OS 7 will keep track of deliveries and sports scores on your wristJay PetersMay 19Demis Hassabis said this might be the ‘foothills of the singularity.’ What?Stevie BonifieldMay 19Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes

GadgetsGamingNintendoNintendo’s $500 Switch 2 bundle includes a game, and it’s available nowYou’ll save up to $30, depending on the game. It’s a good deal if you’re in the market for the console.You’ll save up to $30, depending on the game. It’s a good deal if you’re in the market for the console.byCameron FaulknerMay 19, 2026, 9:00 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Mario Kart World is one of three games you can choose from in the bundle.Photo by Amelia Holowaty Krales / The VergeCameron Faulkneris an editor covering deals and gaming hardware. He joined in 2018, and after a two-year stint at Polygon, he rejoinedThe Vergein May 2025.Nintendo recently teased the“Choose Your Game” Switch 2 console and digital game bundle, coming in early June. However, multiple retailers (including Nintendo itself) are already offering the $499.99 bundle. The list currently includesGameStop,Amazon,Best Buy, and theNintendo Store. It’s a deal worth hopping on ahead of the holiday season, since Nintendo is raising the price of the Switch 2 to $499.99 in September.While theSwitch 2console currently costs $449.99 by itself, the bundle includes a full-price game with up to a $30 discount. You can choose between one of three digital games to redeem upon setup:Donkey Kong Bananza,PokémonPokopia, andMario Kart World. Each of those are regularly $69.99 except forWorld, which retails for $79.99.RelatedNintendo is raising Switch 2 prices in SeptemberI tested dozens of cases for the Switch 2 and these are the bestNintendo Switch 2 “Choose Your Game” bundle$500$5306% off$500Choose betweenDonkey Kong Bananza,Mario Kart World, andPokémon Pokopia.Read More$500 at Amazon$500 at GameStop$500 at Best BuyNone of the games offered as part of the bundle will take up much of the Switch 2’s 256GB internal SSD.Mario Kart Worldtakes up the most space of the three, requiring 22GB. However, you may eventually own enough games to exceed its capacity, so we recommend considering a256GB Samsung microSD Express cardto double your available storage. There’s currently a deal on those happening atAmazon, where you can buy one for $47.99 (originally $79.99).To redeem your game, enter the eShop from the Switch 2’s home screen. Then select “enter code” on the left side of the page. You’ll then enter the 16-character code provided in your bundle, then make your game selection.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Cameron FaulknerDealsGadgetsGamingNintendoVerge ShoppingMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAINewsThe future of Google is a search box that does everythingGoogle, Gemini, Gmail — it feels like they could one day become a single search box.Google, Gemini, Gmail — it feels like they could one day become a single search box.byJay PetersMay 19, 2026, 9:24 PM UTCLinkShareGiftPhoto by Allison Johnson / The VergePart OfGoogle I/O 2026: All the news and announcementssee all updatesJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Last year, after watching Google’s I/O keynote, I wrote that it felt likeGoogle’s future was Google googling. After watching this year’s I/O keynote on Tuesday, I don’t think Google just wants to google for you — I think it wants to doeverythingfor you, all from a search box.Take the trustyGoogle search bar itself, something Google is generally hesitant to update, which is getting some updates. It will “dynamically” expand as you type longer queries. It will offer “AI-powered suggestions” that Google claims will “go beyond autocomplete,” which could cause you to fill in the blanks of a search in a way you didn’t intend and that may or may not be helpful.Or how about what Google actuallyshows in search results? From AI Overviews you can keep asking questions in AI Mode, which generates a custom page with an AI-generated summary of what you’re searching about instead of showing you a traditional list of links. Search results are also going to be even more personalized because Google will be able to generate custom UI for you, including generating things like interactive visuals and graphs right within your search results page. You’re also going to be able to ask Google, right from the search bar, to make “information agents” that can keep track of things you care about — like new sneaker drops or apartment listings — making the search bar a kind of AI-infused Google Alert.Gemini is getting a bunch of upgrades and features, too. It can send you a “Daily Brief” that tells you about your day based on information from across your Google apps, like Gmail and Google Calendar. You can create your own custom Google-powered agents thanks to a featurecalled Gemini Spark, which could have a leg up on agents like OpenClaw because it’s a first-party offering from Google. The company has also been recently making a big kerfuffleabout Personal Intelligence, which pulls in context from your other Google apps that helps inform your responses from Gemini.In Workspace, Google wants you tojust blabat tools like Gmail and Docs and Keep, and the apps will help with things like parsing your inbox, drafting a document, or generating to-do lists. For shopping, the newUniversal Cartwill keep track of things you want to buy in apps like Search, Gemini, Gmail, and YouTube — and let you check out using Google’s payment infrastructure. And speaking of YouTube, the companyis testingan AI Mode-like experience there, too, where it will put together a page of search results instead of just showing you a list of things to watch.There’s more, too — with the newGemini Omni models, you’ll be able to create a video using things like other videos, other images, and audio as prompts. Down the line, you’ll be able to generate other media as well; the family of models is meant to be able to “create anything.”I could go on; there was a lot announced at I/O. But what I’m trying to get at here is that Google isn’t just showing you the location of information anymore. Through its various search boxes, it’s now just answering questions for you in a way that it thinks is most helpful. In the most charitable view: If it’s done well and accurately, this allcouldbe very useful. But that means it has to be done well and accurately, and that’s anextremely high bar to clear, especially with complex search queries or questions about sensitive data like my years of emails in Gmail.It’s not hard to imagine a future where, someday, Google just makes everything happen in one universal search box. No more hopping between Google Search, YouTube, Gmail, Gemini — my gut is that Google’s current end state is that you just type anything you’re looking for into an Ask Google box, and Google finds a way to make it happen and surface it for you in a personalized way.RelatedGoogle’s AI future demands trust — and your personal dataAssuming that’s the case, that’s not something I want. The fun of the internet is actually doing the work to find stuff, even if it’s sometimes frustrating, difficult, or time-consuming. Sure, the pitch for talking to Gmail to get help finding things in my inbox seems handy, but I’ve also spent years honing my own email management system — one that would work whether I used Gmail or not. It also works great forme, I think we should all have to go through the motions of figuring out the systems across all of our digital lives that work best for us instead of just relying on Google to figure it all out from one universal search box.Google

PodcastsTechGoogleWe react to Google I/O 2026Hayden and Jake discuss Spark, Antigravity, Omni, Pics, vibe coding, the new Google Search, and the singularity.Hayden and Jake discuss Spark, Antigravity, Omni, Pics, vibe coding, the new Google Search, and the singularity.byJacob KastrenakesMay 19, 2026, 9:48 PM UTCLinkShareGiftJacob Kastrenakesis The Verge’s executive editor. He has covered tech, policy, and online creators for over a decade.What better way to unwind from a two-hour keynote presentation than to pore over the weirdest and wildest details, froma Gmail bot you can converse withto DeepMind’s leadersaying the singularity is near.The Vergecastwent live right after the show, with senior AI reporter Hayden Field joining me to discuss the highlights.First up, we talk about the buzzword of the show: agents, which seem to be Google’s new answer to making AI useful. Google hasa new service called Sparkthat sounds a lot like a Gemini-flavored OpenClaw, upgrades to its Antigravity platform, and a bunch of agents that willshop for youand track search topics. We also go overGoogle’s model updates, as well as its shift to world models withGemini Omni, a theoretically ambitious new family that is off to a very focused start.Vergesubscribers, don’t forget you get exclusive access to ad-freeVergecastwherever you get your podcasts. Headhere. Not a subscriber? You cansign up here.Then, we discuss howGoogle Search is evolving— or put more plainly, being stuffed with AI features that turn Google into your final destination when browsing the web. We also talk vibe-coding Android apps inGoogle AI Studio, the newGoogle Pics app, and Google’sexpansion of access to content authentication toolsso that people can learn what is and isn’t real.We wrap up discussing Demis Hassabis’ proclamation that AGI and the singularity is near, which got surprisingly little stage time (one sentence) for such a bold statement. And finally, Hayden takes us through her thoughts on theMusk v. Altmantrial, which wrapped up earlier this week.Subscribe:Spotify|ApplePodcasts|Overcast|Pocket Casts|MoreThanks to everyone who tuned in to the livestream — we’re going to keep them coming. In the meantime, you can always reach us at the Vergecast Hotline (866-VERGE11) or send us an email at [email protected] don’t forget, you can get ad-free podcasts bysubscribing toThe Verge.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jacob KastrenakesGoogleGoogle I/O 2026PodcastsTechVergecastMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAINewsDemis Hassabis said this might be the ‘foothills of the singularity.’ What?After a couple of hours of apps and itineraries, Google’s big AI presentation turned philosophical.After a couple of hours of apps and itineraries, Google’s big AI presentation turned philosophical.byStevie BonifieldMay 19, 2026, 10:16 PM UTCLinkShareGiftImage: GooglePart OfGoogle I/O 2026: All the news and announcementssee all updatesStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.Welcome to a “profound moment for humanity,” according to Google DeepMind CEO Demis Hassabis, who closed outGoogle I/O’s keynote presentationon Tuesday, saying:Google’s cutting-edge research and products will help unlock AGI’s incredible potential for the benefit of the entire world. When we look back at this time, I think we will realize that we were standing in the foothills of the singularity.It will be a profound moment for humanity. This technology will be a force multiplier for human ingenuity and usher in a new golden age of scientific discovery and progress, improving the lives of everyone, everywhere. We look forward to building the future with all of you. Thank you, and enjoy the rest of Google I/O.Just before announcing we’ve possibly arrived at “the foothills of the singularity,” Hassabis introducedGemini for Science, a set of tools and experiments in Google Labs and Google Antigravity intended for helping with scientific research. According to Hassabis, with tools like these, Google hopes to “reimagine drug discovery with the goal of one day solving all disease.” Tech executives often discuss AI this way, like Microsoft CEO Satya Nadella referring to AI as “cognitive amplifier tools” and Luma AI CEO Amit Jain’sclaim last yearthat AI is the key to saving Hollywood.In aninterview withBloombergjust a few months ago, Hassabis said that “the singularity is another word for a full AGI arriving,” which is not the same definition we’re used to from the likes ofVernor VingeandRay Kurzweil.At the time, even by his measurement, he claimed that “we’re nowhere near that.” When asked about his timeline for achieving AGI, Hassabis said he was standing by his prediction that we have a “50 percent chance of getting there by 2030.”Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAIGoogleGoogle I/O 2026NewsTechMore in:Google I/O 2026: All the news and announcementsGoogle is launching an Android version of its AI Studio vibe coding tool.Jay Peters12:14 AM UTCWear OS 7 will keep track of deliveries and sports scores on your wristJay PetersMay 19The future of Google is a search box that does everythingJay PetersMay 19Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechEntertainmentGadgetsPlex is tripling the price of a lifetime pass to $750 after doubling it last yearNot your average streaming price hike.Not your average streaming price hike.bySean HollisterMay 19, 2026, 11:03 PM UTCLinkShareGiftImage: PlexPart OfStreaming keeps getting more expensive: all the latest price hikessee all updatesSean Hollisteris a senior editor and founding member of The Verge who covers gadgets, games, and toys. He spent 15 years editing the likes of CNET, Gizmodo, and Engadget.I am dying to know how much money Plex is about to make the next six weeks charging people to stream their own video from their own homes. Today, it’s giving every prospective customer until July 1st to lock in a lifetime subscription at today’s rates — before it triples the price to $750.Plex already more than doubled the price of a lifetime Plex Pass subscriptionfrom $119.99 to $249.99 last March, and it’ll triple again on July 1st to $749.99. At that price, you’d have to subscribe for 11 years (at the current annual rate) to make the lifetime sub worthwhile.Plex isn’t choosing the new number because it expects you to pay $750.In a blog post, the company says it actually wanted to stop issuing lifetime plans; it would rather push users towards recurring subscriptions instead to “sustain long-term development”. Now that the annual subscription covers a decade’s worth of Plex, it’s a win for Plex either way.But it mayespeciallybe a win for Plex if you, like me, feel sudden FOMO and consider giving Plex $250 for a lifetime subscription now. While you’d only be giving Plex around 3.5 years of annual revenue at today’s rates, it gets an instant infusion of cash. It’s like a huge sale except instead of needing to lower the price, Plex has actually raised it. (Again, the annual subscription costs double what it did last March.)From Plex’s email to existing users.Image: PlexPlex’s blog and emails to users encourage this with a hugely prominent call to action, bolded and highlighted in blue: “You have until 12:01 AM UTC on July 1, 2026, to get a Lifetime Plex Pass at the current price of $249.99 USD*here. If you’ve been considering it, now’s a great time to buy.”So does Plex’s website, which is currently topped with a banner reading: “The price of a Lifetime Plex Pass is increasing on July 1, 2026.Get yours at the existing price while you still can.Buy now”.None of that is to suggest Plex is laughing all the way to the bank. It’s a small company, its finances are private, the global advertising market is in a downturn, andit had big layoffsa few years back.Plex says monthly and annual subscription prices aren’t changing today, and nothing will change for existing lifetime pass holders. Those changes came last year. That’s also when Plex gated one of its most valuable features behind that subscription: the ability to stream your media server’s video outside your home.RelatedPlex Pass is going up in price — and now you’ll need it for remote playbackPlex makes piracy just another streaming serviceIt’s time to change your Plex password againFor free, without a subscription, you can still create collections and stream to devices on your local network, and stream remotely from other people who pay. Plex Pass also adds features that lets you skip the intros and credits on shows, remotely download your media, auto-rewind when pausing, hardware-accelerated streaming in its Media Server app,and more.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Sean HollisterAppsEntertainmentGadgetsTechTVsMore in:Streaming keeps getting more expensive: all the latest price hikesYouTube Premium is getting pricierDominic PrestonApr 10Netflix is raising prices againEmma RothMar 26Amazon Prime Video nearly doubles the price to go ad-free and stream 4K videoEmma RothMar 13Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechGadgetsNewsWear OS 7 will keep track of deliveries and sports scores on your wristGoogle’s newest smartwatch update could also give you up to 10 percent better battery life.Google’s newest smartwatch update could also give you up to 10 percent better battery life.byJay PetersMay 19, 2026, 11:24 PM UTCLinkShareGiftA Pixel Watch 4 running Wear OS 6, not Wear OS 7.Photo by Amelia Holowaty Krales / The VergePart OfGoogle I/O 2026: All the news and announcementssee all updatesJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Amid the flurry of today’s Google I/O announcements, Google shared detailsabout Wear OS 7, the next major update to its smartwatch platform. To help you keep track of things like deliveries and sports scores, Wear OS 7 will get the iPhone-style Live Updates that wereintroduced on Android last year— which can appear on your watch or your smartphone — and you’ll also be able to track automated tasks that an AI is working on right from your watch.Wear OS is also getting an upgrade from its widget-like Tiles for glanceable information. With the new update, Google is adding “Wear Widgets” to the platform, which look more like Android widgets and can appear in small or large layouts that “align perfectly” with Android’s 2x1 and 2x2 widget formats. There are some AI-powered features coming to Wear OS 7 as well, including the introduction of Gemini Intelligence, Google’scatch-all brandingfor personalized and proactive Gemini features, on “select watches” launching “later this year.”Various Wear OS 7 features, including Wear Widgets. Click the photo to take a closer look.Image: GoogleGoogle is also promising an “up to” 10 percent improvement in battery life for average users if you’re upgrading from Wear OS 6 to Wear OS 7. The company says that it’s investing in “power optimizations” so that users can “do more with their favorite apps,” but we’ll have to wait and see how battery life actually shakes out when Wear OS 7 is actually available.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGadgetsGoogleGoogle I/O 2026NewsSmartwatchTechWearableMore in:Google I/O 2026: All the news and announcementsGoogle is launching an Android version of its AI Studio vibe coding tool.Jay Peters12:14 AM UTCDemis Hassabis said this might be the ‘foothills of the singularity.’ What?Stevie BonifieldMay 19The future of Google is a search box that does everythingJay PetersMay 19Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

NewsTransportationCarsMercedes’ electric AMG GT 4-door coupe can go 0-60 in 2 secondsThe super sedan has axial motors, a high-performance battery, and (regrettably) fake engine sounds.The super sedan has axial motors, a high-performance battery, and (regrettably) fake engine sounds.byAndrew J. HawkinsMay 20, 2026, 3:30 AM UTCLinkShareGiftImage: MercedesAndrew J. Hawkinsis transportation editor with 10+ years of experience who covers EVs, public transportation, and aviation. His work has appeared in The New York Daily News and City & State.The era of ultra-high performance Mercedes EVs is here. The German automaker finally revealed its new super sedan, the AMG GT 4-door coupe, with technology borrowed from the automaker’s XX concept that last year made history bydriving 24,901 miles in under 8 days at Nardò Ring in southern Italy.With the production model, Mercedes rethought its approach to motors and batteries in the hopes of delivering a high performance vehicle that could go toe-to-toe with even some hypercars. The new AMG GT utilizes three axial flux motors developed byMercedes subsidiary YASA, delivering up to 1,153 horsepower and 1,475 lb-ft of torque. Mercedes claims to be the first to use these types of motors, which thanks to their thin disc shape weigh just a fraction of a traditional radial motor while still delivering massive horsepower.The high-performance battery, meanwhile, utilizes tall, ultra-slim cylindrical cells that are only 1 inch in diameter, allowing heat to escape from the core to the outside surface almost immediately. In addition, Mercedes developed a special, high-tech oil that is non-conductive so as not to cause an electrical short. The oil flows directly around every single individual cell for direct cooling. Inspired by Formula 1, this system provides 20 kW of cooling power, or about four times more cooling capacity thana standard EQS battery. You can drag race it over and over again, and it theoretically won’t overheat.The AMG GT 4-door coupe is built on an 800-volt architecture capable of handling ultra-fast charging up to 600 kW. That plus the innovative cooling system enables charging from 10-80 percent in just 11 minutes, Mercedes says. The nickel-cobalt-manganese-aluminum cathode, combined with an anode containing silicon, can achieve an energy density of over 298 Wh-per-kilogram. The EV can also switch from 800V to 400V when required and supports five global DC charging standards (including NACS and CCS2).Of course, this all translates into an absolute demon-level track car. But of course, like most automakers, Mercedes is anxious about how race enthusiasts will take to a completely silent electric motor. That’s why the AMG GT 4-door coupe will also feature over 1,600 sound files derived from the AMG GT R to simulate engine notes, exhaust burbles, and traction interruptions during virtual gear changes. It also has distinct sounds for unlocking, entering, and charging the vehicle.But it isn’t just a fast car that makes fake sounds. The AMG GT 4-door coupe also has a lot of computing power. Mercedes centralized the brain of the vehicle into the AMG Race Engineer Core, running on the automaker’s brand new MB.OS operating system. Instead of a dozen small chips arguing with each other, one ultra-advanced master chip sits in the center of the car and simultaneously controls everything from driving, charging, suspension, and battery cooling.Inside you’ll find not one, not two, but three screens, all housed under one continuous glass surface. That includes the 10.2-inch driver display, a 14-inch angled central multimedia screen, and a 14-inch passenger display running MB.OS. Owners can track all their metrics, including aero, heat, and energy usage in real time.Mercedes didn’t release the official pricing yet, but said that GT 55 version would be available in late 2026, followed by the GT 63 in early 2027.Images by MercedesFollow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Andrew J. HawkinsCarsElectric CarsMercedes-BenzNewsTransportationMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

In BriefPosted:10:45 AM PDT · May 19, 2026Image Credits:Alex Wong / Getty ImagesSarah PerezAgentic app coding gets an upgrade with Google’s release of Android CLIIn addition to serving up its own flavor of vibe-coding tools for building on Android, Google announced it will also offer tools that allow AI agents, like Claude Code, OpenAI’s Codex, or its own Antigravity or Gemini in Android Studio, to accelerate Android app development.Google shared on Tuesday at its annual Google I/O developer conference that its Android CLI (command line interface) is now stable at version 1.0 and can be used by those who are leveraging AI agents to build for them — no matter what their preferred coding platform may be.The move acknowledges that many people are now building for Android with AI agents that aren’t from Google. The company is trying to find a way to make its specialized knowledge, like what’s found within Android Studio, more accessible.With Android CLI, AI agents can retrieveknowledgeabout Android development via a new “android studio” command that can tap into the capabilities ofAndroid Studiowhen building their own project. From there, the agents will be able to tap into a range of other commands and tools.Google Antigravity, the company’s agentic development platform, will include an optional bundle that will install the tools and knowledge included in Android CLI, so it can also perform the core tasks for Android app development, Google says.TopicsAI,Android,Apps,coding,Google,google io,google io 2026,In BriefMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAppsGoogle’s AI Studio now lets anyone build Android apps in minutesSarah Perez13 hours agoAppsGoogle adds voice-based prompting to Docs and KeepIvan Mehta13 hours agoAIGoogle launches Antigravity 2.0 with an updated desktop app and CLI tool at IO 2026Ivan Mehta13 hours agoLatest in AppsSocial‘Ask YouTube’ brings AI-powered conversational search to video, adds Gemini Omni to ShortsAmanda Silberling1 hour agoAppsGoogle just declared itself a contender in AI design at IO 2026Aisha Malik9 hours agoAIYou can now talk to your Gmail inbox, as seen at Google IO 2026Sarah Perez10 hours ago

At its annual Google I/O event on Tuesday, Google announced that the Gemini app is getting a series of new updates, including a “Daily Brief” feature, a redesigned interface, access to a new AI video model calledGemini Omni, and new personal AI agent calledGemini Spark.The updates signal Google’s push to turn its Gemini app into an all-purpose AI hub rather than a stand-alone chatbot and to make the AI assistant more competitive with apps like ChatGPT and Claude.Google describes the new Daily Brief feature as a personalized digest designed to be your first stop each morning. It pulls together information from a user’s inbox, calendar, and most important tasks, and then organizes them into a clear overview. The company says Daily Brief doesn’t just summarize this information — it also prioritizes tasks and suggests next steps, with the most important items shown first.Daily Brief is rolling out today to Google AI subscribers in the United States.Image Credits:GoogleThe Gemini app already has considerable reach: The company said it has more than 900 million monthly users and is available in over 230 countries and in more than 70 languages. But Google clearly wants more and without losing its existing customers.Google said it rebuilt the app from the ground up. Now when users open the app, they’ll see a new design language called “Neural Expressive,” which brings fluid animations, vibrant colors, new typography, and haptic feedback.Image Credits:Sarah Perez/TechCrunchGemini’s responses are no longer presented as a wall of text, which is the case with most AI chatbots. Instead, key information appears in bold at the top. Additional text and possibly other elements such as images and timelines appear as the user scrolls down.As for Gemini Spark, Google describes it as a 24/7 personal AI agent that helps you navigate your digital life. Spark transforms Gemini from an assistant into an active partner that does real work on your behalf. Since it’s a cloud-based agent, Spark keeps working in the background even when you lock your phone. With Gemini Spark in the Gemini app, users will be able to create their own custom workflows.Spark is currently in testing and the company expects to make it available to Google AI Ultra subscribers next week.Image Credits:Sarah Perez/TechCrunchThe company’s new AI video model,Gemini Omni, combines Gemini with Google’s generative media models to create outputs grounded in knowledge. For example, you could give it a simple prompt like “claymation explainer of protein folding.” The model lets you upload audio, images, and video to generate a consistent, high-quality video, according to Google.By adding access to a new video-generation model like Gemini Omni, Google is intensifying competition among major AI platforms in the ongoing race to lead multimodal content generation. The model is rolling out to Google Flow and YouTube Shorts for Google AI subscribers, underscoring the company’s broader push into multimodal content creation and AI-powered video tools.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,gemini,Google,Google I/O,google i/o 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

The era of the “ten blue links” is officially over.At its Google I/O conference on Tuesday, Google unveiled an AI-powered overhaul of Search centered around a reimagined “intelligent search box” — what the company describes as the biggest change to this entry point to the web since the search box debuted more than 25 years ago.Instead of returning a simple list of links, Google Search will drop users into AI-powered interactive experiences at times. Google is also introducing tools that can dispatch “information agents” to gather information on a user’s behalf, along with tools that let users build personalized mini apps tailored to their needs.Loading the player…The resulting experience will no longer look much like how people envision Google Search, which has long been defined by ranked links to websites that have the information you need.With the revamped Search experience, the new search box simply expands to accommodate longer, more conversational queries, rather than making you decide what type of search experience or mode you want to choose at the start of your query. It will also have a new AI-powered query suggestion system that goes beyond autocomplete to help people craft more complex and nuanced queries, Google says.Google’s AI Overviews will also allow users to ask follow-up questions in AI Mode, beginning Tuesday, the company noted.Image Credits:GoogleGoogle is also introducing agentic capabilities and AI-powered interactive features into the search experience. This means people will spend even less time clicking the traditional blue links that Google Search used to return.Starting this summer, people will be able to create, customize, and manage multiple new “information agents” within Google Search. These agents can work in the background 24/7 to track changes on the web and alert you to new information. For instance, you could have an agent track market movements based on customer parameters, Google suggests.Image Credits:GoogleWhile the underlying technology here is powered by AI, which makes it more capable, the idea itself is not a new one.In 2003, Google launchedGoogle Alerts, a change-detection service that emailed users when new web results matched their search terms. The web was smaller and more manageable then, of course, so this became a part of many information workers’ tool sets. (That service still existsin some form but is no longer the way most web users go about acquiring new information.)Information-gathering agents are an evolution of Google Alerts. Beyond spotting changes, they can make sense of them, too.“You could send an alert to track market movements in a particular sector with very specific parameters, and the agent will map out a monitoring plan for you, including the tools and the data it needs to access — like our real-time finance data,” Google’s head of Search, Liz Reid, explained in a press briefing. “And it will then keep track of those changes and let you know when the conditions are met, and provide a synthesized update with links and information you can dive into further,” she added.Image Credits:GoogleThis shift means that “searching the web” will increasingly be performed by AI agents rather than humans. Instead, people will focus more on acting on the information those agents provide instead of manually clicking links.Image Credits:GoogleLinks will become an afterthought with the coming changes to the Search results experience, which builds on Google’s earlier launches of AI search features, like its short summaries known as AI Overviews and its conversational search, AI Mode.AI Overviews are now used by more than 2.5 billion monthly users; meanwhile, its conversational search mode, launched last year, now tops 1 billion monthly users. (ChatGPT, for comparison, has900 million weekly active users, as of earlier this year. This suggests that ChatGPT is now seeing more frequent engagement, with users coming back repeatedly throughout the week, while Google has more total unique people touching its AI features over the course of a month.)Now, thanks to a combination of Gemini andGoogle Antigravity, the company’s agentic development platform, Search results will begin to look more like interactive web pages.Image Credits:Google“Search can build custom experiences just for your individual questions, from dynamic layouts, interactive visuals to persistent and stateful project spaces that you can return to again and again,” says Reid. One of the ways Google is integrating these new capabilities is with “generative UI” (user interface), where it builds custom widgets and visualizations on the fly in answer to users’ search questions.You can imagine, for example, how a question about black holes in space could lead to an interactive visual that brings the concept to life, Reid said, adding that users can then ask follow-up questions and see Google respond with brand-new visuals in real time.Image Credits:GoogleGoogle says the new system was built in partnership wi

The AI coding boom is now coming directly for Android app development. On Tuesday atGoogle IO 2026, the companyannouncednew native Android app creation capabilities in its web-basedGoogle AI Studio,shrinking a process that takes weeks of setup and coding down to minutes.The company also said that consumers will be able to use Gemini AI to find the apps they need, both on the Play Store and the web, expanding opportunities for developers to have their apps discovered.Google says the new capabilities could make sense for anyone from a seasoned developer looking to prototype a new app quickly to a first-time creator.By offering the ability to essentially vibe-code Android apps via web-based tools, Google is ramping up the competition with other AI-powered development tools, like Cursor, Replit, Lovable, Claude Code, and others, while also opening up Android development to a new type of user: a non-technical creator. The news also represents an expansion of Google’s earlier addition ofAI-powered coding with Geminiin itsdesktop versionof Android Studio.The apps are built with theKotlinprogramming language using Google’s Jetpack Compose toolkit and with support integration with hardware sensors like GPS, Bluetooth, and NFC, the company says. However, the resulting creations, for now, are only meant to be used personally, as publishing for family and friends is still on the roadmap.The company suggests the technology could be used for the creation of personal utilities and simple social apps, hardware-enabled experiences, or AI-powered experiences.For now, would-be app developers can use the embedded Android Emulator directly in a web browser to preview and interact with the app as it’s being built. Users can then install the app on their Android phone over a USB cable connected to their computer, using the integrated Android Debug Bridge (adb).For those looking to take their project further,AI Studiocan automatically create the app record, package the bundle, and upload it to an internal testing track inGoogle Play Consolefor developers. This allows users to continue to iterate on their app while updating on their devices along the way.Those who want to take the next steps to publishing the app more publicly can hand off this version of the project to Android Studio by downloading a zip file and exporting it directly to GitHub. In time, Google plans to allow creators to publish their apps for use by family and friends and will add support for Firebase integrations (Firestore, Firebase Auth, Firebase App Check, and other tooling).In doing so, the company is imagining an Android app ecosystem where users find apps from among their own network of friends, not just the Play Store.However, for the latter, Google is infusing AI into the experience here, too.A new “Ask Play” AI-powered overlay will allow users to discover new apps by having natural conversations with AI within the Play Store.Image Credits:Google (screenshot)Perhaps more importantly, apps will begin to be surfaced with users’ conversations with Google’s Gemini virtual assistant, exposing developers’ apps to millions of users. This will roll out in the weeks ahead across Gemini on the web and on Android. Later this year, Gemini will also surface over 450,000 movies and TV shows, plus where to livestream sports, which can directly link users from their queries to a developer’s Android app with the content in question.Image Credits:GoogleWhile Google previewed a number ofAndroid-related announcements last week, it held back on sharing the news of the native Android app development until Tuesday’s start of its annual developer conference, Google I/O. That suggests the company believes this is bigger news and more closely tied to its idea of putting AI to real-world use, as was the larger theme of this year’s event, where AI was spread across Google products, from workspace productivity apps to AI tools, search, mobile apps, and more.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Android,Apps,coding,Google,google io,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sarah PerezConsumer News EditorSarah has worked as a reporter for TechCrunch since August 2011. She joined the company after having previously spent over three years at ReadWriteWeb. Prior to her work as a reporter, Sarah worked in I.T. across a number of industries, including banking, retail and software.You can contact or verify outreach from Sarah by [email protected] via encrypted message at sarahperez.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s

Google launched on Tuesday Gemini 3.5 Flash, a new AI model that the company says is its strongest yet for coding and autonomous AI agents. The model, which was introduced at the company’s annual Google I/O developer conference, can independently execute coding pipelines, manage research projects, and, in internal tests, build an operating system entirely from scratch.The release signals Google’s shift from pitching AI as a conversational tool to AI as an agentic tool. It’s not just answering questions, but planning, building, and iterating on real work with minimal human input.Koray Kavukcuoglu, DeepMind’s chief technologist, told reporters on Monday ahead of the public launch: “3.5 Flash offers an incredible combination of quality and low latency. It outperforms our latest frontier model, 3.1 Pro, on nearly all the benchmarks,” including coding, agentic tasks, and multimodal reasoning.He added that it is 4x faster than other frontier models, a speed that’s ideal for coding and agentic tasks, but that Google has “taken it to another level” by developing an optimized version of Flash that’s 12x faster with the same quality.That speed is central to Flash’s design for agentic work, where multiple AI agents run at the same time on long-running tasks, according to Kavukcuoglu. Onstage at I/O, Google engineer Varun Mohan, demonstrated agents spawning off to work on separate components before coming together to build a full operating system inside Antigravity, the company’s agentic development platform and IDE.Image Credits:GoogleKavukcuoglu said Flash 3.5 was co-developed with Antigravity so that agents could have a “native environment where they can live, work, and execute.”At I/O, Google released Antigravity 2.0, a stand-alone desktop application designed around agent-first development.The gains are showing up beyond demos. Google says 3.5 Flash’s agentic capabilities are already creating impact among partners, like banks and fintechs automating multi-week workflows, or data science teams finding insights in complex data environments.The model can run autonomously for multiple hours, though Tulsee Doshi, Google’s senior director and head of product, said it will at times pause and ask for user input when it hits a decision point or permission issue that requires human judgment.When Google releases its forthcoming 3.5 Pro model, the two are designed to work in tandem.Doshi told TechCrunch that “3.5 Pro becomes your orchestrator, your planner, and then it actually can leverage Flash to be the various sub-agents. I think it really comes down to where do you really want that reasoning power, where you actually want that larger model that can really push on the reasoning side versus where do you have tasks that really do merit good brute force tool use capabilities?”Now, 3.5 Flash is the default model in the Gemini app and in AI Mode in Search globally. At I/O, Google also announced agentic capabilitiescoming to Search, letting users create, customize, and manage AI agents directly on the platform. The new model will also powerGemini Spark, Google’s new personal AI agent designed to run 24/7 to help consumers manage their digital life.Providing that level of AI capability for average consumers comes with scrutiny. Google is currently facing a lawsuit after a man nearly committed a mass casualty event and died by suicide following weeks of chatting with Gemini last year.The implications for harm only grow when making powerful autonomous agents available more broadly. Google says Gemini 3.5 has strengthened cyber and CBRN (chemical, biological, radiological, and nuclear) safeguards and is better calibrated to engage with sensitive questions rather than refuse them outright.Gemini 3.5 Flash is available generally today via Antigravity, the Gemini API, and Gemini Enterprise, as well as through the Gemini app and AI mode in Search.TopicsAI,gemini,gemini flash 3.5,Google,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Rebecca BellanSenior ReporterRebecca Bellan is a senior reporter at TechCrunch where she covers the business, policy, and emerging trends shaping artificial intelligence. Her work has also appeared in Forbes, Bloomberg, The Atlantic, The Daily Beast, and other publications.You can contact or verify outreach from Rebecca by [email protected] via encrypted message at rebeccabellan.491 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost Popular‘Ask YouTube’ brings AI-powered conversational search to video, adds Gemini Omni to ShortsAmanda SilberlingGoogle just declared itself a contender in AI design at IO 2026Aisha MalikYou can now talk to your Gmail inbox, as seen at Google IO 2026Sarah PerezHow to use Google’s new AI a

We’ve all pulled up Street View on Google Maps to show a friend what our childhood home looked like, or dropped that little person icon onto the streets of Paris to see if we booked a hotel in a cool neighborhood. Imagine being able to do that, but in a more immersive, interactive way that allows you to really simulate the street and its environs, and even do things like adjust the weather or see what it would look like in a “Day After Tomorrow” scenario.That’s one of the goals of Google’s latest integration. Starting today, Google DeepMind is connecting Street View toProject Genie,the company’s general-purpose world model that can generate diverse, interactive environments. The new feature launched during theGoogle I/O 2026developer conference.“It’s really powerful for both the agent [and robotics] use case and for humans to play with, and that’s always been the thesis of Genie,” Jack Parker-Holder, a research scientist on DeepMind’s open-endedness team, told TechCrunch.He gave the example of a new robot being deployed in London, which rarely sees the sun. Genie could, Parker-Holder says, simulate those scarce occasions when the sun glints off the Victorian housing, so the rays don’t shock the robot when it happens.Loading the player…“Simultaneously, you might say, ‘I’m going to New York City, but not this time of year,’” he continued. “‘It’s going to be snowy. I want to see what that block looks like in the snow.’”Google has been collecting Street View data for 20 years via cars with cameras and individuals strapped with “tracker backpacks.” The tech giant has collected north of 280 billion images across 110 countries and seven continents.“With Street View, we have imagery from a large quantity of the world,” Jack said. “You can imagine how potentially powerful it is to combine this rich source of real-world information and data with an ability to simulate worlds.”Google released its latest world modelGenie 3 for research previewlast August and opened up access to the tool to Google AI Ultra subscribers in the U.S. in January, allowing customers to create interactive game worlds from text prompts or images. The goal is to use Genie for educational experiences, gaming, and robotics training.Genie 3 is already helping to powerone of Waymo’s simulatorsto train its self-driving cars on “exceedingly rare events” like tornadoes or casual elephant encounters. Adding Street View data to that could help Waymo prepare to launch in more cities around the globe.Waymo has its own simulator that it relied on to scale to 11 U.S. cities and test its AI driver in several more. The difference with Genie, says Parker-Holder, is that those are all from the car’s point of view. Street View allows for not only simulating a world anchored to a real place, but also shifting the point of view to other types of agents, like a human or a robot.Google is launching Street View in Genie to some Ultra users in the United States starting today, with access rolling out at scale over time. Global Ultra users will gain access over the next few weeks, per the company.The researchers’ goal is to put this new capability into as many hands as possible, per Diego Rivas, a product manager at DeepMind. He cautioned that Street View in particular and Genie in general is still an experiment, so there’s much to improve upon in terms of accuracy.In the samples the Google team showed me — including an underwater simulation of a neighborhood I used to live in — the results are impressive and recognizable, but still video game quality rather than photorealistic. The models are also not yet physics-aware, meaning they don’t yet understand cause and effect. For example, in a simulation of a woman running through a snowy Joshua Tree, she ran right through cacti and bushes.Compare that to, say, Google’s image generator Nano Banana — which can now generate perfect text in infographics — or its video generator Veo — which understands that paper boats drift on water currents, smoke disperses into the air, and fabric drapes over forms.Physics isn’t hard-coded into these models; they learn it intuitively over time through passive observation, as a living being would.“I think for this kind of model, it’s maybe six to 12 months behind video in terms of the accuracy and quality, so I think it’s something we will solve,” Parker-Holder said.Jonathan Herbert, director of Google Maps who started on the Street View team as an intern 12 years ago, said that Genie can’t yet create a faithful reconstruction of a street. He thinks the real breakthrough is the AI’s spatial continuity. If you turn 360 degrees, the AI correctly remembers and simulates the environment behind you. From that point on, the model can build a new environment on top of that.“We have long thought about how we can build out the best and richest model of the world on top of Street View data,” Herbert said. “It’s definitely been an idea of ours to use Maps Data in new ways and for new kinds of AI researc

Google is getting (back) into the smart glasses game.At Google I/O on Tuesday, the company announced a new partnership with Warby Parker and Gentle Monster to produce a new line of AI-powered glasses. The company says that the devices will be built to pair with Android and iOS devices and were designed in collaboration with Samsung. They will be available later this year, the company said.Google is calling the new devices “audio glasses,” in that users will be able to issue verbal commands to them and get things done via its ecosystem of apps and services, including Gemini.Image Credits:TechCrunchThe user simply talks to their glasses (the demo shared on Tuesday involved a Googler ordering a coffee online by merely talking to the glasses), and the device, when synced, complies.Google has dabbled in smart glasses a number of times over the years. It notoriously launchedGoogle Glassyears ago, which ultimately helped spawn the derogatory term “glassholes.”The smart glasses space has changed a little bit since then, however. Lately, major companies — most notably Meta — and a small army of startups and smaller firms, have invested in the space.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,ai glasses,glasses,Google,google io,google io 2026,HardwareWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lucas RopekSenior Writer, TechCrunchLucas is a senior writer at TechCrunch, where he covers artificial intelligence, consumer tech, and startups. He previously covered AI and cybersecurity at Gizmodo.

The jury’s speedydecision to rejectElon Musk’s lawsuit against the other founders of OpenAI and Microsoft confirmed what we saw in the courtroom: Musk’s case was a weak one, in part because he waited so long to file it.Watching the closing arguments last week, OpenAI’s attorneys detailed point by point how the law was on their client’s side, while the plaintiff’s team focused on Sam Altman’s apparentlack of credibilityand expressed disbelief that anyone would disagree with Musk’s accusations.The final effect was that, after the verdict, some found it hard to believe Musk had lost — including the man himself. In a post he later deleted, Musk called Judge Yvonne Gonzalez Rogers a “terrible activist Oakland judge,” then announced his plans to appeal, declaring, “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity.”But Altman and Brockman weren’t the only figures who benefited from OpenAI’s non-profit investments. As much as Musk and his legal team tried to make the trial about Altman, the proceedings revealed just as much about Musk.One incident that came out in court showed Musk benefiting from OpenAI in an uncomfortably familiar way. Greg Brockman testified that in 2017, Musk asked him to bring a team of OpenAI researchers down to Tesla’s headquarters to help with the autopilot team for a few weeks. “It was pretty clear that was not something we could say no to,” Brockman said.Brockman described taking a team of leading scientists, including Andrej Karpathy, Ilya Sutskever, and Scott Gray, to consult with the “demoralized” Tesla workers. They helped come up with ideas to improve the vehicle’s self-driving technology, with Sutskever telling the team that if they could find 10,000 images of a tricky corner case, they would be able to fix their software. Musk even asked Brockman to recommend employees to fire, which he declined to do.Another person familiar with the episode confirmed Brockman’s account and said Tesla did not reimburse OpenAI for the time and effort of its employees. Musk’s family office, Excession, didn’t reply to a request for comment.The heart ofMusk’s caseis that Altman, Brockman, and OpenAI committed a “breach of charitable trust” — that Musk donated funds for a specific charitable purpose, and his co-founders instead used them for something else. He also accused them of “unjust enrichment” due to stock and other benefits from OpenAI’s for-profit.In the case of the OpenAI scientists parachuting into Tesla, Musk’s charitable donations were intended to hire scientists focused on securing the benefits of AGI. Instead, he had them work for free at his for-profit company.Dorothy Lund, a Columbia Law School professor and the co-host of theBeyond Unprecedented podcast, told TechCrunch that this arrangement wouldn’t be legal, calling it “a bit rich for Musk to be suing for breach of a charitable trust, when he appears to have been redirecting assets in a way that was inconsistent with that mission.”It’s true that the self-driving work involved artificial intelligence, but witnesses for Musk emphasized that Tesla’s self-driving project was very different from OpenAI’s research agenda. That’s in part because Karpathy left OpenAI for Tesla shortly after this incident. OpenAI’s attorneys portrayed the departure as Musk violating his duty to the lab, where he was co-chair of the board, by recruiting one of its key researchers to his own company.The other fact that no doubt influenced the jury was the amount of time Muskspent tryingto gain sole control of a potential OpenAI for-profit affiliate in 2017. Musk deployed good cop, bad cop tactics in an attempt to convince his co-founders to let him have total control of OpenAI’s for-profit affiliate — giving them free Teslas and threatening to withhold his donations.His efforts put his attorneys in a tricky spot, facing a need to convince the jury there was a significant difference between what Musk envisioned and the for-profit that was ultimately created. They suggested a “small adjunct” for-profit would be permissible, though OpenAI’s witnesses showed non-profits with large commercial arms are common.Indeed, there’s a very plausible counter-factual where Musk took one of the offers his co-founders made to split their equity more evenly, and finds himself today as one of OpenAI’s largest shareholders — just not the controlling one. But several times during the trial, Musk’s associates testified that he refuses to invest in any business he could have sole control over.The failure of Musk’s claims because he filed them too late has been cited as a technicality, but the statute of limitations has substance behind it: People and businesses make important decisions and spend resources based on their understanding that what they are doing is permissible. If someone like Musk waits too long to sue, then the cost of unraveling all those decisions can outweigh a just reimbursement.No

Shay Shwartz knows a lot about email phishing attacks. As a teenager, he made money as a hacker, but after getting caught at age 16, he realized he could use his cyber talents to prevent attacks rather than launch them.He went on to spend about a decade in top-tier cybersecurity roles, leading major projects for Israel’s elite defense and intelligence units, including work connected to the Iron Dome project, before joining Axis, the startup later acquired by HPE.All along, he had been itching to launch his own startup, and two years ago, he finally took the plunge.His startup Ocean, an agentic email security platform built to fight AI-powered attacks, just emerged from stealth mode with $28 million in total funding. The round was led by Lightspeed Venture Partners, with participation from Picture Capital and Cerca Partners. High-profile angel investors also joined the round, including Wiz co-founder and CEO Assaf Rappaport, as well as Yevgeny Dibrov and Nadir Izrael, the co-founders of Armis, which recently sold to ServiceNow for $7.75 billion.While established vendors like Proofpoint and Mimecast, along with newer players like Abnormal Security, help detect standard phishing attacks, Shwartz (pictured right next to co-founder and CTO Oran Moyal) argues that AI requires a different defensive approach.In the past, only highly sophisticated hackers could pull off spear-phishing due to the sheer amount of time, research, and manual labor needed to launch targeted attacks.“AI just made the entire process automatic, so the scale is much, much bigger now,” Shwartz told TechCrunch. “I can instruct LLM to go and understand exactly who you are, harvest large amount of public information, and create those phishing attacks very targeted against you.”Ocean claims its AI can thoroughly analyze the context of every incoming email to detect fraud and impersonation attempts.The startup is already reviewing billions of emails each month for customers, including Kayak, Kingston Technology, and Headspace.Shwartz said Ocean built a small language model tailored to quickly analyze emails, understand the sender’s intent, and evaluate it against the user’s specific organizational context.“This is like having a guard in every door,” Shwartz said. “This is how we make the inbox a safe place with high hygiene.”TopicsAI,Exclusive,lightspeed,phishing,Security,StartupsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Three-year-old Mach Industries has acquired solid rocket motor startup Exquadrum in a $50 million cash-and-equity deal, the Huntington Beach-based defense startup tells TechCrunch. Exquadrum — now rebranded as Mach Energetics — has been fully folded into Mach’s operations, giving it direct control over one of the most important, and constrained, components in modern unmanned systems.The deal itself started with a lucky break. The two companies first connected last September when an Exquadrum customer at an MIT recruiting event happened to overhear a Mach recruiter mention that the company was in the market for a solid rocket motor supplier. Introductions were made, Mach came in as a customer, and now, roughly five months later, it has acquired the company outright, beating out upward of eight other potential buyers in the process, it says.“The Exquadrum acquisition marks an important next stage in Mach’s growth,” said founder and CEO Ethan Thornton, who dropped out of MIT at 19 to start the company. “As we deliver vehicles to the warfighter, we’ll continue to vertically integrate our supply chain across solid rocket motors, engines, radar, and avionics to ensure we deliver the best possible product at the lowest cost. In many areas of the defense industrial base, these components are not only too expensive or lacking performance, they’re simply unavailable, with lead times stretching years. In short, vertical integration is non-optional.”That supply problem is real and getting more acute. Decades of consolidation have left the domestic solid rocket motor market effectively controlled by two large primes — Aerojet Rocketdyne and Northrop Grumman — with little independent capacity to absorb the growing demand that modern drone warfare has created.Indeed, in February, the Pentagon awarded the defense tech outfit Anduril $43.7 million specifically to expand domestic SRM production (its second such investment in the company in just over a year), explicitly calling SRMs a critical bottleneck in the munitions supply chain.Mach is now explicitly positioning itself as part of the solution, and not just for its own programs but for the broader ecosystem. Mach Energetics plans to sell components, testing services, and subsystems to other defense firms, a maneuver that suggests Mach sees itself as potential infrastructure for the defense tech industry and not just a systems builder.According to Mach, all 85 Exquadrum employees are coming over as part of the deal, along with the company’s IP, business lines, and its 70,000-square-foot facility in Victorville, California, which is anchored by a nearby energetics and rocket propulsion test site. The combined company now has roughly 350 employees. Exquadrum co-founders Kevin Mahaffy and Eric Schmidt (no relation to the former Google CEO) are both taking on leadership roles within Mach Energetics and the broader organization.The acquisition also mirrors moves by other ambitious defense tech startups that are focused on owning the stack and using cost and speed as competitive weapons. Mach has five vehicle programs in various stages of development — Viper, a jet-powered VTOL; Glide, a high-altitude strike glider; Stratos, an airborne surveillance platform; Dart, its low-cost counter-drone interceptor; and Pike, a long-range strike munition built for large-scale deployment — with plans to enter production on at least three this year. The company says the acquisition meaningfully improves unit economics across all of them at exactly the moment it is starting to scale.Mach has raised nearly $200 million in total — most recently a$100 millionSeries B last June led by Bedrock Capital, Khosla Ventures, and Sequoia Capital — at a valuation of $470 million. It’s a multiple that right now looks modest for a company with this trajectory and will definitely be worth watching as the rubber starts to meet the road this year.Topicsdefense tech,Ethan Thornton,Hardware,mach industriesWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Connie LoizosEditor in Chief & General ManagerLoizos has been reporting on Silicon Valley since the late ’90s, when she joined the original Red Herring magazine. Previously the Silicon Valley Editor of TechCrunch, she was named Editor in Chief and General Manager of TechCrunch in September 2023. She’s also the founder of StrictlyVC, a daily e-newsletter and lecture series acquired by Yahoo in August 2023 and now operated as a sub brand of TechCrunch.You can contact or verify outreach from Connie by [email protected]@techcrunch.com, or via encrypted message at ConnieLoizos.53 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk h

In BriefPosted:2:14 PM PDT · May 19, 2026Image Credits:MARTIN BUREAU/AFP via Getty ImagesZack WhittakerDiscord enables end-to-end encrypted voice and video calling for every userMessaging platform giant Discord has switched on end-to-end encrypted voice and video messaging for every user. Theend-to-end encryptionfeature means that Discord users can now communicate privately without anyone else being able to listen in to their calls, not even Discord.This is a major privacy win for the community’s hundreds of millions of users, soon after social media giant Metapulled the plugon Instagram’s end-to-end encrypted messaging feature earlier this year. TikTok said it would alsonot end-to-end encrypt user messagesafter it became a U.S. company.Discordlaunched the end-to-end encrypted voice and video calling feature in 2024, and on Monday rolled out the feature to every user, no action needed.“End-to-end Encryption is now standard for every voice and video call on Discord, outside of stage channels. No opt-in required,” said Mark Smith, Discord’s vice president of core technologies, in ablog post.Topicscybersecurity,discord,end-to-end encryption,In Brief,SecurityMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedSecurityUS cyber agency CISA exposed reams of passwords and cloud keys to the open webZack Whittaker15 hours agoSecurityNYC Health + Hospitals says hackers stole medical data and fingerprints during breach affecting at least 1.8 million peopleZack Whittaker2 days agoSecurityOpen source tool maker Grafana Labs says hackers stole its code, refuses to pay ransomZack Whittaker2 days agoLatest in SecurityIn BriefDiscord enables end-to-end encrypted voice and video calling for every userZack Whittaker9 hours agoSecurityFrom teen hacker to Iron Dome researcher, this founder raised $28M to fight AI phishingMarina Temkin9 hours agoIn BriefHackers have compromised dozens of popular open source packages in an ongoing supply-chain attackLorenzo Franceschi-Bicchierai15 hours ago

At theGoogle I/O 2026keynote, the tech giantrevealed new agentic capabilities in Search, where users can create, customize, and manage multiple AI agents to stay updated on topics of interest. The announcement is part of Google’s larger push toward agentic AI systems that can take initiative and assist with ongoing tasks instead of answering one question at a time.Unlike traditional search tools that respond only when prompted, Google’s information agents are designed to operate continuously in the background, 24/7, helping users stay informed about their interests without needing to repeatedly search for the same information every day.Instead of delivering a list of links, the agents can synthesize information from multiple sources, explain why something matters, compare perspectives, and provide actionable insights. In many ways, the agents represent the next evolution of Google Alerts, the notification service Google launched in 2003. However, these agents are designed to go beyond simple notifications.For instance, someone following the stock market could create an information agent focused on specific companies, share price, or economic trends. The agent could monitor market activity throughout the day, track breaking news, summarize earnings reports, alert users when major changes happen, and provide summaries and links to learn more.Image Credits:GoogleIt could also help with everyday tasks, such as tracking flight prices for upcoming trips, monitoring sports teams and live events, following breaking news, keeping tabs on housing or job market trends, and tracking weather or traffic.To use the feature, users can open AI Mode in Search and enter a prompt.For example: “Keep me updated on nearby movie tickets for ‘The Mandalorian and Grogu.’” When something relevant appears, the Google app sends a push notification. You’ll also see your active tracked topics in your AI Mode history, where you can jump back in to manage, refine, or turn off an alert.Information agents will be available this summer. The company is first rolling them out to Google AI Pro and Ultra subscribers in the U.S., then to additional markets later on.In addition to these information agents, Google also introduced amajor redesign of Search, including what it describes as a reimagined “intelligent search box,” the company’s biggest change to Search in more than 25 years. The new interface is designed to support longer, more conversational queries. There’s also a new AI-powered query suggestion system that goes beyond traditional autocomplete, helping users craft nuanced and context-aware searches.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,AI agents,Google,google io 2026,Google SearchWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Google isn’t finished infusing AI into your inbox. On Tuesday attheir IO 2026 developer conference, the tech giant announced an expansion of its“AI Inbox” functionality for Gmail, which is adding conversational AI features. That means you can ask Gmail about things in your inbox instead of typing in search terms.The company says the Gemini AI-powered feature, called Gmail Live, will help you quickly find information buried in your inbox.Image Credits:GooglePerhaps you need information about your upcoming flight, the time of your dentist appointment, the door code for your Airbnb rental, or some details about an event at your kid’s school, for instance.Before, you’d have to type in keywords in the search box (or maybe type in someone’s email address or domain) to try to narrow down your search. That doesn’t always make emails easy to find, however, especially if the search term is something found across several messages.“Gmail Live can answer naturally phrased questions, respond to follow-up questions, and pivot if you need to interrupt it,” Devanshi Bhandari, product lead for Gmail, explained in a briefing ahead of Google’s annual developer conference, Google I/O, where the feature was first introduced to the public.It’s another way that Google is trying to showcase how its AI technology can drive real-world improvements to products used by millions of consumers, at a time when many outside the tech industry are questioning the value of AI, as new data centers get built in their backyards, driving up their power bills.Being able to point to something as simple as making it easier to find something that’s lost in your email inbox — an experience nearly everyone has suffered at some point — could be a practical and positive use case for AI … or at least, Google hopes.Bhandari demonstrated Gmail Live to reporters, asking the tool a series of questions about things in the inbox, like a child’s show-and-tell project and their class trip, plus hotel and flight information for a trip to Detroit. Similar to using a stand-alone AI chatbot like Gemini or ChatGPT, Gmail users can ask these questions aloud in natural language, and the chatbot responds.In the demo, Gmail Live also understood nuances between things like “field trip” and “trip” and was able to jump from one topic to another, Bhandari pointed out. Plus, the AI can pull granular details from emails, like a hotel room number, or infer which people you’re asking about, even when they’re not explicitly named.Similar voice technology is also coming to its to-do list, Google Keep, the company noted.Notably, Gmail Live is not replacing traditional Gmail search — it’s just another option.Google may have learned that not everyone is ready for an AI-only experience after it “upgraded”Google Photos with AI-powered searchto much backlash.Google Photos later rolled back the feature,making the use of AI optional after numerous complaints.Gmail is also gaining other new capabilities, including ready-to-send drafts, instant file access, and the ability to manage to-dos by marking individual tasks as done.Image Credits:GooglePlus, theAI Inbox experience, which launched earlier this year, will expand beyond Google AI Ultra subscribers to reach Google AI Pro and Plus subscribers as well. This allows you to see an overview of the tasks and items to catch up on that are buried in your inbox, all on one page.The voice-powered Gmail Live feature, however, will roll out later this summer and will initially be limited toGoogle AI Ultra subscribers.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,gemini ai,gmail,Google,google io,google io 2026,Google WorkspaceWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sarah PerezConsumer News EditorSarah has worked as a reporter for TechCrunch since August 2011. She joined the company after having previously spent over three years at ReadWriteWeb. Prior to her work as a reporter, Sarah worked in I.T. across a number of industries, including banking, retail and software.You can contact or verify outreach from Sarah by [email protected] via encrypted message at sarahperez.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One t

Google announced atits annual Google I/O eventon Tuesday that it’s launching Pics, a new AI-powered design and image-generation app for Google Workspace. The tech giant says it designed the app to be accessible to everyone, from teachers to small business owners.With Pics, users can generate everything from social media graphics and invitations to marketing materials and mock-ups using simple text prompts, without needing any editing skills or advanced tools. By giving users an easy way to generate visuals, Google is looking to take on popular design apps like Canva, as well as products from AI-native competitors likeClaude Designfrom Anthropic. Google’s entry into the space signals that AI-powered design is fast becoming a core competitive arena — with real stakes for any business that depends on visual content.The new app is launching to a group of testers at I/O and will be rolling out to Google AI Ultra subscribers this summer, Google says.The company acknowledges that although AI models today can generate high-quality images, it’s still difficult to modify just one part of an image. If you get an image that’s almost perfect but want to change a small detail, you have to write an entirely new prompt and hope the AI doesn’t alter too much. That’s why Pics not only generates images but also makes them easily editable.Loading the player…Users can enter a prompt, and Pics will generate what they need. Gemini powers the editing layer, making every element in a generated design or image fully adjustable. You can write a new prompt to make changes, but you can also simply click the part you want to change and leave a comment — much like leaving feedback in Google Docs.You can also edit directly, without leaving a comment or writing a prompt. For example, if you create a birthday party invitation and want to change the time listed on the card, you can do so manually.Pics is powered by Nano Banana 2, which Google says is a strong fit for the app because it supports precise text rendering, real-world knowledge, and detailed visual output. Pics is also built natively into Google Workspace, enabling visual collaboration across its apps.Once you’re happy with your design, you can download, copy, print, or share it with others. You can also pass it to someone else for a final round of edits before it goes out, Google says.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,Google,Google I/O,google io 2026,google picsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Google iscompletely revampingits search experience, and that doesn’t stop at YouTube. Like the rest of Google, YouTube’s search bar is getting infused with AI tools like “Ask YouTube,” a feature that is supposed to give users a more sophisticated search experience.“With Ask YouTube, you can ask more complex search queries, such as wanting tips on how to teach your kid to ride a bike, or finding creator reviews of cozy games to play before bedtime,” the company explained. “You can even ask follow-up questions to continue refining what you’re looking for.”YouTube will compile both Shorts and long-form videos and generate a response.Premium subscribers in the U.S. on desktop can start using this feature now through YouTube’s optional Premiumofferingsto test new tools.Image Credits:GoogleThe company also announced that it is adding Gemini Omni, Google’s new AI video model, to YouTube Shorts Remix and the YouTube Create app.“Remixing with Omni delivers a fresh way for users to create and build on each other’s imagination,” YouTube wrote in a press release. “The model better understands user intent creating more consistent and meaningful storytelling while also handling complex video and audio adjustments behind the scenes.”Other companies likeMetaand OpenAI have seen mixed reception when pushing for AI use in Shorts. OpenAI evensunsettedits social app Sora, where users could post and share their AI-generated clips. But YouTube seems to be rolling this out in a manner that feels a bit less front and center.YouTube is also expanding itslikeness-detection toolto creators 18 and older, which is supposed to help prevent creators from being deepfaked in other people’s AI content. If a creator sees themselves misrepresented in AI videos, they can request that the video be removed. Since this feature is only now expanding more broadly, it remains to be seen how effective it is.TopicsApps,gemini omni,Google I/O,google i/o 2026,Media & Entertainment,Social,YouTube,YouTube ShortsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Amanda SilberlingSenior WriterAmanda Silberling is a senior writer at TechCrunch covering the intersection of technology and culture. She has also written for publications like Polygon, MTV, the Kenyon Review, NPR, and Business Insider. She is the co-host of Wow If True, a podcast about internet culture, with science fiction author Isabel J. Kim. Prior to joining TechCrunch, she worked as a grassroots organizer, museum educator, and film festival coordinator. She holds a B.A. in English from the University of Pennsylvania and served as a Princeton in Asia Fellow in Laos.You can contact or verify outreach from Amanda by emailing [email protected] or via encrypted message at @amanda.100 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Nice cable you got thereIran demands Big Tech pay fees for undersea Internet cables in Strait of HormuzIran’s claim over subsea chokepoint pushes US tech companies to overland fiber.Jeremy Hsu–May 19, 2026 7:00 am|189Multicolored lines show undersea Internet cable routes running through the Strait of Hormuz.Credit:TeleGeographyMulticolored lines show undersea Internet cable routes running through the Strait of Hormuz.Credit:TeleGeographyText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navIran claims it will charge US tech companies fees for using undersea Internet cables that run beneath the contested Strait of Hormuz shipping lanes. The war has already halted multiple projects and led to the suspension of cable repairs in the region—and the latest Iranian threats may accelerate efforts by Big Tech and Gulf countries to find alternative routes for bypassing the Strait of Hormuz’s digital chokepoint.The latest assertions of Iranian authority over the Strait of Hormuz were announced in a brief statement by Ebrahim Zolfaghari, a spokesperson for Iran’s military and the Islamic Revolutionary Guard Corps. “We will impose fees on internet cables” Zolfaghari wrote in aMay 9 post. It was not immediately clear how Iran might implement such fees or impose its rules on cable projects, given that the majority of routes pass through Oman-controlled waters.But Tasnim and Fars, both Iranian state-linked media channels, laid out more detailed proposals on how Iran could charge license fees to US tech giants for the use and maintenance of undersea cables carrying regional Internet traffic, according toThe Guardian. For example, the Tasnim plan described charging tech companies—specifically naming Meta, Google, Amazon, and Microsoft—license fees for cable usage while also claiming that Iran alone has the right to repair and maintain the subsea cables.More than 99 percent of international Internet traffic runs through the global network of undersea cables that crisscross various oceans, connecting continents and islands. The major active cables running through the Strait of Hormuz primarily serve the Gulf countries in the region. They include theAsia Africa Europe-1,FALCON, and theGulf Bridge International Cable System, according toTeleGeography, a telecommunications research organization.The FALCON and Gulf Bridge cables run through Iranian territorial waters at certain points, Alan Mauldin, a research director for TeleGeography, said in aCNNinterview. CNN also reported that Iranian state media outlets have “issued veiled threats warning of damage to cables.”By comparison, there is “not much risk to Europe-Asia data traffic from hostilities in the Strait of Hormuz” because such Internet traffic primarily travels through cables in the Red Sea, TeleGeography wrote in a blog post. But undersea cables in the Red Sea have already seen a spate of damage in recent years, exacerbated by lengthy repair times and attacks by Houthi rebels allied with Iran.The biggest threat to the digital chokepointIran’s capability to threaten undersea cables as a way of imposing its will is uncertain at this point. Since the US and Israel started the war by attacking Iran on February 28, the US military claims to have destroyed 161 vessels of the Iranian navy, according to Admiral Brad Cooper, commander of US Central Command, intestimonygiven to the Senate Armed Services Committee on May 14. The US military has also sunk some of thesmall fast boatsoperated by Iran’s paramilitary Islamic Revolutionary Guard Corps when the latter threatened civilian shipping in and around the Strait of Hormuz.Thevast majority of damageto undersea cables comes from commercial ships accidentally dragging their anchors or fishing trawlers dragging weighted nets along the seafloor. So it’s not out of the realm of possibility for a more innocuous-looking Iranian ship to sneakily perform somesubsea cable sabotageif it’s willing to run the gauntlet of US military surveillance and patrols in the strait.Even a damaged commercial ship abandoned in the strait could end up dragging its anchor across some cables, as was the case with a drifting ship thatdamaged three cablesfollowing an attack by Houthi rebels in the Red Sea in 2024.However, the greatest threat to subsea cable infrastructure in the Strait of Hormuz may simply come from delays in any necessary cable repairs in the region. Such jobs require specialized ships to find the damaged area and lower grappling hooks to lift up the cable for inspection and repair, according toBBC News. That repair process can require days or sometimes weeks, which would leave the ship vulnerable to Iranian missiles, drones, or fast boats that havecontinued to attackcommercial shipping in and around the strait.“Operators face a choice: pay protection fees and accept Iranian licensing over Middle East Gulf seabed activity, or accept that future faults may go unrepai

ConsolidationElectrical utility megamerger is all about the data centersNextEra’s blockbuster deal with Dominion likely means higher bills for consumers.Dan Gearino, Amy Green, and Charles Paullin, Inside Climate News–May 19, 2026 9:45 am|59The campus of NextEra’s headquarters is seen on Monday in Juno Beach, Fla.Credit:

      Marco Bello/Getty ImagesThe campus of NextEra’s headquarters is seen on Monday in Juno Beach, Fla.Credit:

      Marco Bello/Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA proposed merger of the largest utility in the country by market value, NextEra Energy, with the sixth-largest, Dominion, would create a megacompany at a time when data centers and rapid increases in electricity demand are reshaping the industry.The proposal,announced Monday morningand contingent on state and federal regulatory approval, would result in a company that leads in nearly every aspect of the US power and utility industry, including overall electricity generation, natural gas generation, and renewables.The $67 billion deal combines NextEra’s size and reach with Dominion’s positioning as the local utility for the world’s largest concentration of data centers innorthern Virginia. But the results are likely bad for consumers and the environment, creating a company with enormous financial and political strength that will be difficult to effectively regulate, according to consumer advocates and analysts.For perspective, only Exxon Mobil and Chevron would be larger based on market value among US-based energy companies.“Mergers are not about consumers; they’re about shareholders,” said Ari Peskoe, director of the Electricity Law Initiative at Harvard Law School. “For the Dominion shareholders, they are selling their shares at a premium. The executives are getting massive payouts for facilitating this, assuming it all goes through, and obviously NextEra believes the transaction is going to add value to the company. Ratepayers are all an afterthought.”The deal makes financial sense for both companies, said Andrew Bischof, an equity analyst for Morningstar.“We view the transaction as allowing NextEra to accelerate its data center ambitions, which had trailed those of its regulated peers, by using Dominion’s expertise and relationships to expedite NextEra’s data center hub plans,” he said in a note to clients.NextEra, based in Juno Beach, Florida, includes Florida Power & Light, the largest regulated electricity utility in the state, and NextEra Energy Resources, a wholesale electricity supplier that owns power plants across the nation. Dominion, based in Richmond, Virginia, includes regulated utilities serving much of Virginia, parts of North Carolina and South Carolina, and other assets across the country.The company would be called NextEra Energy, and NextEra CEO John W. Ketchum would serve in the same role after the deal closes. Robert M. Blue, Dominion’s CEO, would be the CEO for regulated utilities for the merged company. The parties said they expect regulatory approvals to take 12 to 18 months.A page from NextEra Energy and Dominion’s investor presentation shows the states where each company has regulated utilities and the merged company’s ranking in various categories.Credit:
                                  NextEra EnergyA page from NextEra Energy and Dominion’s investor presentation shows the states where each company has regulated utilities and the merged company’s ranking in various categories.Credit:

      NextEra EnergyNextEra shareholders would own 74.5 percent and Dominion shareholders would own 25.5 percent, respectively, of the combined company in the all-stock transaction.“We are bringing NextEra Energy and Dominion Energy together because scale matters more than ever—not for the sake of size, but because scale translates into capital and operating efficiencies,” Ketchum said in a statement.“Adding to the pollution problem”The post-merger NextEra would be the leader in so many categories in the US utility sector that it’s easier to list the ones where it wouldn’t be on top. It would rank second in nuclear power generating capacity and in the number of regulated utility customers, trailing Exelon Corp. of Chicago in both.NextEra and Dominion both have substantial carbon emissions, but neither was among the top five utility companies in the country in 2024, according to the most recent edition ofBenchmarking Air Emissions reportfrom the Natural Resources Defense Council. NextEra ranked sixth and Dominion ranked 11th, and their sum was less than that of each of the leaders, Vistra Energy and Duke Energy.But those are still massive emissions from a company that stands to gain more clout because of its size.“If we continue to add dangerous climate pollution into the mix, then people who are already suffering and are typically hurt first and worst will su

Catching the satellite busIn addition to space stations, Vast says it will now build high-power satellites“Every single successful space company is diversified in its products.”Eric Berger–May 19, 2026 9:00 am|27A rendering of Vast’s 15 kW satellite.Credit:

      Vast SpaceA rendering of Vast's 15 kW satellite.Credit:

      Vast SpaceText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAs part of its plan to develop a private space station, Vast Space built and then launched a small demonstration spacecraft in early November. This vehicle then completed dozens of test objectives with flying colors before making a successful de-orbit three months later.The mission, which tested power, propulsion, tracking, and a multitude of other technologies needed for Vast’s Haven-1 space station, was evidently so successful that the company is ready to use its spaceflight capabilities for other purposes. The Long Beach, California-based company announced Tuesday that it plans to begin selling high-powered satellite buses.“Every single successful space company is diversified in its products,” said Max Haot, chief executive of Vast Space, in an interview. “So for us it really was a question of when, not if.”A satellite for power-hungry applicationsThe company’s first offering is a 15 kW-class satellite bus capable of supporting various demanding missions. Each satellite is about 3 meters long and 4 meters tall, with a mass of 700 kg and payload capacity of at least 350 kg. They will have a design lifetime of five years and be intended to operate in areas ranging from low-Earth orbit out to lunar orbit. Vast aims to serve various customers, from telecommunications to observation to data services. Haot added that Vast also plans to offer an Nvidia Space-1 Vera Rubin Module to support orbital data center inferencing needs.The Vast satellite bus—essentially a backbone providing power, propulsion, and navigation for various payloads—will be based largely on technology ported over from the company’s Haven-1 space station, which is due to launch for the first time next year as the world’s first private space station. However, there will be some new elements needed for the satellite, and Haot said Vast is already moving forward with in-house development of electric propulsion and a deployable solar array for the satellite.Vast has already signed a customer for four satellites, with an option to purchase up to 200 additional satellites. Haot said the company is targeting a launch of at least 10 Vast Satellites in the fourth quarter of 2027.With this new product line, Vast is entering an increasingly crowded market. Historically, in the United States, a handful of large companies such as Boeing, Lockheed Martin, Northrop Grumman, Maxar, and Sierra Space have manufactured medium and large satellites. Typically, these were costly and often bespoke designs that cost tens to often hundreds of millions of dollars.But a few trends have changed the landscape in recent years. The US government’s Space Development Agency has signaled that it prefers proliferated constellations—many satellites spread out present less of a concentrated target than a few larger and more expensive satellites. With the Falcon 9 rocket’s increased cadence, as well as rideshare missions, it became easier and sometimes cheaper to get smaller and medium-sized satellites into orbit.This has led to an influx of venture capital to back a new generation of companies seeking to build less expensive, more modular satellites that could fill a variety of purposes. There are several prominent, relatively new entrants in this area, including K2 Space, Rocket Lab, True Anomaly, Blue Canyon, and Millennium Space Systems.Vast already has… vast facilitiesHaot said most of these companies are still emerging, with products that are not yet mature. In other words, he believes that if Vast Space can execute, it could become a market leader, especially in power-hungry applications. Vast has already invested $1 billion in facilities for spacecraft manufacturing, including clean rooms, which can be used for space stations as well as satellites, he said.The number of satellites in space has exploded in recent years, largely due to the rapid expansion of SpaceX’s Starlink constellation. For decades, the total number of satellites orbiting Earth numbered about 4,000, but in the last five years that number has grown to about 14,000.This is just the beginning. By some estimates, in another decade, there will be approximately 500,000 satellites in orbit for communications, Earth observation, orbital data centers, and other applications.Haot expects that about 90 percent of these will be satellites built by SpaceX, Amazon, Blue Origin, or other major players. But even 10 percent of that number, if available to commercial satellite bus manufacturers, would represent 50,000 satellites f

Test of TimeCivilization VIIfinally lets you build a civ that stands the test of timeCiv 7’s devs talk walking back the game’s most controversial decision.Samuel Axon–May 19, 2026 12:00 pm|42You can’t tell as much from the map view, but many ofCiv 7’s key underlying systems have been totally overhauled.Credit:

      Samuel AxonYou can't tell as much from the map view, but many ofCiv 7's key underlying systems have been totally overhauled.Credit:

      Samuel AxonText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to nav“Build a civilization to stand the test of time.” That was the promise on the box of Sid Meier’sCivilization, the first in a long-running strategy game franchise that has evolved over 35 years and seven mainline entries.Civ 7introduced a new approach to play wherein players would change civilizations from their initial selection twice by the end of a game. Lots of players said, “Wait a minute: we’re literally not building a civilization to stand the test of time anymore.” After such a negative reception at launch, longtime series fans began to wonder whether the franchise would continue to stand the test of time.It’s clearly not a coincidence that the new, major update for the game reaching players today is titled “Test of Time.” It’s a major reworking of several of the game’s key systems, and it reintroduces the ability to play one civ from beginning to end while retaining some of the big ideas that definedCiv 7at launch.I’ve spent the past week playing the game with the Test of Time update, and I spoke with two of the game’s senior developers—the franchise’s creative director, Ed Beach, and series head of product Matt Schembari—about the choices they made, reacting to player feedback, and competing ideas about what Civ is to players.A quick note on agesFor those who haven’t been followingCiv 7up to this point: Another major change to the game is that it has been broken into three distinct ages: Antiquity, Exploration, and Modern. Each has its own tech and culture tree, game mechanics, and, until now, civilizations to pick from.The intent behind it was to allow the game’s designers to solve some longtime problems with how traditional Civ games could lag or have balance problems at specific stages. If that in itself is too much deviation from the classic formula for you, it’s unlikely Test of Time will make you likeCiv 7any more than you might have a year ago. It’s still a foundational framework for this game. Test of Time gives players a lot more freedom inside that structure, but it does not remove the structure itself.Give the people what they wantStarting a new game ofCiv 7with the patch, the biggest change was immediately apparent. As of the last update before this one, you could pick from just 14 civs when starting a game in the Antiquity Age. This time, I saw more than 40.I could play as America from the Age of Antiquity now, for example. When I set up the game, I was also given the option to choose the AI’s civ-switching behavior; I could prohibit it, allow it, or pick an option that makes the AI follow suit on whatever choice I make.Yep, we have modern age civs as options for a game starting in antiquity.Credit:
                                  Samuel AxonYep, we have modern age civs as options for a game starting in antiquity.Credit:

      Samuel AxonOf course, playing one civ throughout the three ages would seem to run counter to how the game was originally designed. Firaxis took this update as an opportunity to take a major game balance pass across all the civs, and it introduced some new concepts to make one civ for three ages fit theCiv 7framework.Civs are considered to have an “apex age,” during which they have access to their full kit. When they’re in an age other than their apex age, they’re called “time-tested civs,” and they have access to a limited amount of their normal bonuses. But players can use a system called syncretism to adopt bonuses that are normally offered to specific civs that would call the current age their apex age.Based on playing one full game in the new patch, I feel this implementation works really well overall, and it’s a lot of fun.That said, I could see it getting unwieldy to build on over time—the game’s designers now must account for many more possibilities and balance them against each other to make sure the game stays fair and fun.When I asked Beach about this, he suggested it’s not as big a design challenge as it might seem. “I think we still have some advantages because of the structure of the game,” he said. “I think where we are giving ourselves an immense challenge is the content space, because we have 43 civs that you can play as, and now we need to give them bonuses in three different ages. So we went from basically 43 chunks of content that we need to balance to like 129 chunks of content, and as we add more civs

more taxesEV drivers will pay $130 a year under Congress’ 2026 transportation billPoliticians say they want EVs to pay “their fair share for the use of our roads.”Jonathan M. Gitlin–May 19, 2026 12:12 pm|259Credit:

      Getty ImagesCredit:

      Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe119th Congressmight be one of themost dysfunctional and least productivelegislative sessions in the 250-year history of the United States, but it seems there’s one thing it can agree on: Electric vehicles don’t cost their owners enough money. The Transportation and Infrastructure committee has published its bill to fund surface transportation for the next half-decade, and among the provisions in the “Building Unrivaled Infrastructure and Long-term Development for America’s 250th Act” is an annual fee levied against owners of EVs.“I’m extremely proud of the historic level of investment in America’s bridges—at more than $50 billion, it’s the largest such investment in our history. And the BUILD America 250 Act ensures that electric vehicle owners begin paying their fair share for the use of our roads,” said committee chairperson Sam Graves (R-Mo.).Should the bill pass—and it enjoys support from the Democratic Party, too—you will be required to pay a $130 federal registration fee to drive an EV. And starting in 2029, that fee will increase by $5 each year until it reaches $150. Plug-in hybrids don’t escape untaxed, either; the fee for a PHEV begins at $35 a year and will escalate by $5 each year until it reaches $50 annually. And if state departments of transport don’t collect this federal EV tax, the federal government will “withhold an amount equal to 125 percent of the amount owed from the state’s highway apportionment.”Why?Road funding in the US is a messy business, paid for by a mix of federal and state gas taxes, state vehicle registration fees, and local budgets. And as the percentage of fuel-efficient PHEVs and gas-abstaining EVs grows, the amount of money available to pay for road upkeep drops. So, in the abstract, asking EV and PHEV owners to pay their share is not unreasonable, especially since the slightly higher curb weight of these vehicles will do slightly more damage to the roads, although passenger vehicles are a rounding error compared to a garbage truck, bus, or tractor-trailer.Some stateshave already started imposing a registration fee on clean vehicles to offset reduced gas tax receipts. And at some point, when a significant proportion of the country’s vehicles are EVs or PHEVs, an alternative road-funding method that does not rely on the gas tax will certainly be in order.But EV adoption in the US hastaken a severe hitsince the election of President Trump and the abolition of federal clean-vehicle incentives. As of now, EVs make up barely 3 percent of the entire vehicle fleet. Meanwhile, the federal gas tax has not been increased since 1993; had it been adjusted to keep pace with inflation, it would add an extra 42.34c to a gallon rather than a pathetic 18.4 cents currently.“Drivers of gas-powered vehicles pay approximately $73 to $89 in federal gas tax each year. The proposed fee would charge an unfair premium on EV drivers at a time when all Americans are looking for ways to save money,” said Albert Gore, executive director of the Zero Emissions Transportation Association. “This is particularly concerning as the EV fee will increase to $150 by 2035—nearly double what gas car drivers would pay in a year. This fee lands on top of the road use taxes that many EV drivers already pay at the state level.”“In addition to the new EV fee, the bill makes major changes to federal investments for dedicated EV charging. A lack of dedicated EV funding could particularly impact EV drivers living in—and traveling to—more rural and remote locations in the United States, where federal funding is most needed to help fill gaps in existing charging infrastructure,” Gore said.“With Americans demanding more affordable transportation options as gasoline tops $4.50 a gallon, this bill would impose an onerous new fee on electric vehicles and plug-in hybrids while slashing investments in new EV chargers,” said Shruti Vaidyanathan, director of federal and state transportation advocacy at the National Resources Defense Council. “Congress should be boosting investments in projects that cut costs, cut emissions, create jobs, and build a transportation system that works for all Americans. This bill largely ignores the need to build cleaner, more affordable transportation options.”We should perhaps be glad that the final version will cost drivers much less than Rep Graves’last proposal, which started at $250/year for EVs and $100/year for hybrids.Jonathan M. GitlinAutomotive EditorJonathan M. GitlinAutomotive EditorJonathan is the Automotive Editor at Ars Technica. He has a BSc and PhD in Pharmacolog

Gotta go fastThe era of 1,000 Hz gaming monitors has arrived, but why?LG’s latest hits one frame per millisecond at a full 1080p resolution.Kyle Orland–May 19, 2026 12:26 pm|86Wow, look how fast those giant letters are moving without any apparent motion blur!Credit:

      LGWow, look how fast those giant letters are moving without any apparent motion blur!Credit:

      LGText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAlmost exactly two years ago, we weregawking at prototypes of 1,000 Hz monitorsand wondering who really needed a display that could support such ludicrously smooth frame rates. Now that those prototypes are starting to develop into retail products, we’re still wondering how much of a market there is for gaming displays that can update with a new frame every single millisecond.The latest entry in the ultra-fast refresh race is LG’s 24.5″ UltraGear 25G590B, which the companyannounced this weekas “the world’s first Full HD gaming monitor with a native 1000Hz refresh rate” ahead of a planned launch in “select markets” in the second half of the year. That “Full HD” promise means LG’s 1,000 Hz display hits the 1080p threshold that is by far the most common resolution reported by gamers inSteam’s regular hardware surveys.That would represent a decent step up from the likes ofAcer’s Predator XB273U F6,Samsung’s Odyssey G6, orPhillips’ EVNIA 27M2N5500XD, all of which have to shift down to a relatively blurry 720p resolution to run at a full 1,000 Hz (but which support 1440p resolutions at a still-quite-fast 500 Hz). LG also notes that its high-end monitor can hit its resolution and refresh rate specs natively, without the need for any “dual mode” rebooting shenanigans to get the fastest performance.Fast-moving objects appear decidedly less blurry as you climb the frame rate curve.Credit:Blur BustersFast-moving objects appear decidedly less blurry as you climb the frame rate curve.Credit:Blur BustersThe folks over at Blur Busters haveextensivelydocumentedresearch showing that refresh rates of 1,000 Hz (and up) can reduce human perception of motion blur and flickering. And while the site notes that you eventually hit “diminishing returns” from all those extra frames—especially on smaller screens—there’ssome evidencethatyou would need a 40,000 Hz monitorto totally eliminate perceived motion blur on a sufficiently large, high-resolution monitor.Perception research aside, it’s unclear how many gamers are interested in spending extra money on a display that breaks the four-digit refresh rate barrier. While LG hasn’t announced pricing details for the UltraGear 25G590B yet, other 1,000 Hz monitors tend to run hundreds of dollars more than comparable models with lower (but still more than adequate) maximum refresh rates in the 240 to 360 Hz range. At the same time,monitors with Nvidia’s GSync Pulsarand similar technologies use “rolling scan” pulsing backlights to reduce apparent motion blur, even on “slower” 360 Hz monitors.LG says the 25G590B is “engineered specifically for first-person shooters” and mentions an “esports-optimized design” in its marketing materials, suggesting the display is targeting the small subset of gamers who play twitch-based reflex games for a living. For everyone else, we recommend ignoring the “larger number is better” impulse and saving your hardware budget forthose increasingly expensive sticks of RAM.Kyle OrlandSenior Gaming EditorKyle OrlandSenior Gaming EditorKyle Orland has been the Senior Gaming Editor at Ars Technica since 2012, writing primarily about the business, tech, and culture behind video games. He has journalism and computer science degrees from University of Maryland. He oncewrote a whole book aboutMinesweeper.86 Comments

Gemini’s efficiency playGemini 3.5 Flash might be fast enough for gen AI to make senseGoogle says its more efficient Gemini 3.5 Flash is the key to your agentic AI future.Ryan Whitwam–May 19, 2026 2:11 pm|72Credit:

      Aurich LawsonCredit:

      Aurich LawsonText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAt last year’s I/O event, Google was still talking about the2.5 branchof Gemini, and what a difference a year makes. We’ve gone through the3.0and3.1families since then, and now it’s on to version 3.5.Gemini 3.5 Flashis rolling out across a wide range of Google products starting today, and Google again claims this model is even better than its last-gen Pro model.That has been a trend with Google’s tick-tock model updates over the past year, but the team says this release is special. Gemini 3.5 Flash supposedly offers frontier-level intelligence while also being efficient enough that it may finally make complex agentic tasks worth doing at scale. Tulsee Doshi, senior director of product management for Gemini, explains that the innovations of Gemini 3.5 Flash are woven through multiple Google products, and this is just the start.Credit:
                                  GoogleCredit:

      GoogleIt’s no secret that generative AI is currently a money pit, and all the major AI players are trying to find paths to greater efficiency. The problem is magnified when you start building agentic experiences that are supposed to run for longer to complete complex tasks. Gemini 3.5 Flash may be a big step toward making that viable. The new model can output nearly 300 tokens per second, but its benchmark scores are similar to larger frontier models (like 3.1 Pro) that build outputs at a quarter of that speed.Google now says that the companies using the most AI tokens could save a billion dollars per year by shifting to the more efficient Gemini 3.5 Flash. API pricing for the new model is significantly lower than the Pro model it apes. Gemini 3.5 Flash clocks in at $1.50 per 1M input tokens and $9 per 1M output tokens. The 3.1 Pro model starts at $2 and $12, respectively, and it’s higher if you use more than 200k tokens.According to Doshi, the team made numerous improvements in pre-training with Gemini 3.5 Flash, but insights gleaned from how devs use Gemini models are really paying off.“With post-training, we’re really starting to unlock some of the value of the feedback we’re getting from users, for example, from Antigravity,” said Doshi. “That’s really what you’re seeing play out in terms of the code performance and the tool use performance. And then, the hope is that you’ll continue to see the step change where 3.5 Pro will be better, and the next Flash meets Pro performance with that series.”Google is focused on code generation with the new model, which is a core agentic angle for AI. Both Terminal Bench and SWE-Bench Pro tests show substantial improvements—3.5 Flash clobbers older Flash models and shows a small but measurable improvement versus Gemini 3.1 Pro. Its scores are in the same neighborhood as OpenAI’s much larger and more expensive GPT 5.5.A major barrier in agentic workflows is how generative models can use interfaces designed for humans. It’s not an easy problem to solve, Doshi said. “Certain things like UI control are expensive to do because the model has to search the page, it has to know where to click, it has to act through multiple steps. I think Flash is able to do that well because of that combination of quality and cost.”Google’s AI evaluations demonstrate these improvements, too. Among Google’s current collection of benchmarks is OSWorld-Verified, which tests how models handle general tasks in real computing environments. It’s similar to the coding improvements. Gemini 3.5 Flash substantially outperforms older Flash models and is even a bit faster than Gemini 3.1 Pro. It’s essentially tied with GPT 5.5.Google’s new Flash model is, again, a little better than the last-gen Pro.Credit:
                                  GoogleGoogle’s new Flash model is, again, a little better than the last-gen Pro.Credit:

      GoogleGemini 3.5 Flash has been rolled out internally at Google, and Doshi noted that it’s having a big impact. “We have a set of internal metrics we’ve been evaluating that measures how Googlers code, so looking at our own code bases and how well the models perform on that,” Doshi said. “And you can see a massive, massive jump between where 3.1 Pro was and where 3.5 Flash is.”Google unveiled the Antigravity IDE last year, and it’s being upgraded to version 2.0 with support for Gemini 3.5 Flash. This update will support multiple parallel workflows—essentially sub-agents spawned by Gemini 3.5 Flash. Again, Google says this is only possible because the new model is so efficient at spitting out tokens.In addition to Antigravity, Gemini 3.5

Foiled againRFK Jr. forced to withdraw charter that opened CDC panel to anti-vaccine quacksCharter would have expanded member eligibility and focused on alleged injuries.Beth Mole–May 19, 2026 2:17 pm|51WASHINGTON, DC - JUNE 24: Health and Human Services (HHS) Secretary Robert F. Kennedy Jr. testifies before the House Energy and Commerce Committee Health Subcommittee in the Rayburn House Office Building on June 24, 2025 in Washington, DC.Credit:Getty | Kayla BartkowskiWASHINGTON, DC - JUNE 24: Health and Human Services (HHS) Secretary Robert F. Kennedy Jr. testifies before the House Energy and Commerce Committee Health Subcommittee in the Rayburn House Office Building on June 24, 2025 in Washington, DC.Credit:Getty | Kayla BartkowskiText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navA revised charter document for the Centers for Disease Control and Prevention’s influential vaccine advisory committee has been withdrawn by the Health Department over an administrative error, according toa notice published in the Federal Register Tuesday.The charter’s revisionsunder anti-vaccine Health Secretary Robert F. Kennedy Jr. would have allowed Kennedy to appoint dubiously qualified anti-vaccine allies to advise the CDC. It also would have directed the CDC panel to focus on alleged vaccine injuries and risks and welcomed fringe groups and anti-vaccine organizations to participate in developing federal vaccine policy.Kennedy’s move to reshape the CDC panel—the Advisory Committee on Immunization Practices, or ACIP—came amid Kennedy’s many other attempts to undermine it, as well as a court order to undo that meddling.In June 2025,Kennedy summarily fired all 17 expertsfrom ACIP and quickly replaced them with unvetted and unqualified anti-vaccine allies. Kennedy’s ACIP then held several meetings in which they aired anti-vaccine views and misinformation, allowed anti-vaccine activists to give unvetted presentations, and ultimately voted to remove longstanding, evidence-based federal vaccine recommendations based on anti-vaccine rhetoric. That includes the removal of a universal recommendation for a hepatitis B vaccine dose at birth, despite no evidence of any safety concerns or any benefit to delaying the dose. Subsequent modeling studies found that the change will meanmore infections, increases in liver cancers and deaths, as well as millions of dollars inhealthcare costs.In January, Kennedy sidestepped ACIP entirely to overhaul the CDC’s vaccine schedule for children,slashing the number of recommended vaccinations from 17 to 11.In March, a federal judge issueda temporary injunction to undo Kennedy’s changes, ruling that Kennedy likely violated federal regulations in unilaterally altering the CDC vaccine schedule and that the new ACIP members were unqualified—as outlined by the panel’s charter—nullifying all of their votes on federal vaccine recommendations.While the Health Department is working to appeal the injunction, Kennedy attempted to circumvent the judge’s ruling on the ACIP members byaltering the committee’s charterto, among other things, allow for people without expertise in immunizations and public health to be members.But, for now, that effort, too, has been thwarted. According to the notice on Tuesday, the new charter has been withdrawn for not following a federal requirement on public notification.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.51 Comments

GitFlubIn stunning display of stupid, secret CISA credentials found in public GitHub repoSSH keys, plaintext passwords, other sensitive data had been up since November 2025.Lee Hutchinson–May 19, 2026 2:27 pm|82Only the best people.Credit:

      Dzmitry Skazau / GettyOnly the best people.Credit:

      Dzmitry Skazau / GettyText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navSecurity researcher Brian Krebsbrings us the newsthat America’sCybersecurity & Infrastructure Agency(CISA) has had a large store of plaintext passwords, SSH private keys, tokens, and “other sensitive CISA assets” exposed in a public GitHub repo since at least November 2025.The now-offline public repo—named, somewhat aspirationally, “Private-CISA”—was brought to Krebs’ attention by GitGuardian’sGuillaume Valadon, who was alerted to the repo’s presence by GitGuardian’s public code scans. Krebs says that Valadon approached him after receiving no responses from the Private-CISA repo’s owner.In an email to Krebs, Valadon claimed that the repo’s commit logs show that GitHub’s default protections against committing secrets—protections designed to protect unwitting or unskilled developers against exactly this kind of stupidness—had been disabled by the repo’s administrator.Testing bySeralys founder Philippe Catureglishowed that this was not a joke or hoax and that he was able to use the credentials in the Private-CISA repo to gain access to multiple Amazon Web Services GovCloud accounts “at a high privilege level.”Krebs notes that the repo appeared to be managed by Virginia-basedNightwing, a CISA contractor. Nightwing has so far not commented publicly, instead referring questions back to CISA.This isn’t the first time CISA has screwed up—in fact, it’s not even the first timethis year. In January,polygraph-failingacting CISA Director Madhu Gottumukkalauploaded sensitive government documents to ChatGPTafter demanding and receiving an exemption to the agency policy that prohibited ChatGPT’s use by CISA personnel. Gottumukkala wasremoved from his role in February.Lee HutchinsonSenior Technology EditorLee HutchinsonSenior Technology EditorLee is the Senior Technology Editor, and oversees story development for the gadget, culture, IT, and video sections of Ars Technica. A long-time member of the Ars OpenForum with an extensive background in enterprise storage and security, he lives in Houston.82 Comments

What is real?Google’s SynthID AI watermarking tech is being adopted by OpenAI, Nvidia, and moreAI content is getting good, but SynthID might be able to help tell truth from fiction.Ryan Whitwam–May 19, 2026 2:37 pm|50Google CEO Sundar Pichai announces 100 billion SynthID images and videos.Credit:

      GoogleGoogle CEO Sundar Pichai announces 100 billion SynthID images and videos.Credit:

      GoogleText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navIn a few short years, we’ve gone from easily identifying AI content that featured superfluous fingers to images and videos that lookshockingly realistic. How can we know what’s real in the age of AI? Google’s answer is SynthID, which it first demonstrated three years ago.The company saysSynthID has since been used to label 100 billion images and videos, plus 60,000 years’ worth of audio. Those numbers are only going up now that SynthID is expanding beyond Google.SynthID is not Google’s only AI labeling strategy. It’s also committed to the C2PA standard, which tags content with metadata describing how it was created. Google began using C2PA more prominently with its Pixel 10 smartphones. Photos taken with the Pixel 10 include metadata describing how they were processed. If a highly zoomed image includes generative elements, it gets an AI tag, too.Google now says this same feature is coming to videos recorded on Pixel 8, 9, and 10 phones in an update in the coming weeks. It’s also adding C2PA scanning to Gemini, allowing the chatbot to explain a file’s providence based on the content labeling. This same capability will come to Chrome and Search in a few months.That metadata is fungible, though. On the other hand, SynthID is deeply integrated with AI-generated content. The digital watermark is present in the pixels of images and videos and in the waveform of AI songs and audio overviews from products like NotebookLM. According to Google DeepMind scientist Pushmeet Kohli, the team worked hard to ensure SynthID is much harder to remove, even if you compress it, crop it, or rotate it.“A technology like this will always be attacked,” said Kohli. “There was a lot of research that we did in making SynthID robust to different kinds of transformations.”Last year, Google added support for SynthID detection in the Gemini app. You can upload the suspect content and ask the chatbot if it’s AI-generated. This should work reliably with all those billions of Google AI images and audio clips from the past three years. A few ambitious tinkerers have claimed to find methods for removing the hidden SynthID patterns. Google contends that none of these bypasses actually work.More SynthID in more placesEven if no one has been able to crack SynthID, that doesn’t matter for the vast majority of AI images on the Internet—only Google’s AI models apply SynthID. That’s going to change soon, though.Google has announced that it has partnered with several companies to add SynthID to their systems. Nvidia will implement SynthID in its Cosmos world foundation models, and OpenAI will use SynthID in its GPT 2 images. Kakao and ElevenLabs will also begin adding SynthID to their AI content.This doesn’t mean you’ll always be able to tell if something is AI by looking for SynthID. Plenty of publicly available models continue to produce content with no AI watermarking, and there are open models that can be trained by anyone looking to create AI images and videos on their own terms. Still, this is a step in the right direction.There will also be new paths to checking SynthID status, so you won’t even have to open Gemini just to check for the watermark. SynthID will be integrated with Circle to Search, Lens, and AI Mode. You’ll also be able to use Gemini in Chrome by sharing a tab with the content in question. You can ask any variation of “Is this AI” to get a SynthID scan with these tools.There is currently not a public API for SynthID—making these scans too readily available could serve as an attack vector for those seeking to circumvent SynthID. However, Google is preparing to launch an AI content detection API as part of the company’s Gemini Enterprise Agent Platform. This will allow trusted business partners to more easily flag AI content, allowing Google to refine the API over the coming months.Ryan WhitwamSenior Technology ReporterRyan WhitwamSenior Technology ReporterRyan Whitwam is a senior technology reporter at Ars Technica, covering the ways Google, AI, and mobile technology continue to change the world. Over his 20-year career, he's written for Android Police, ExtremeTech, Wirecutter, NY Times, and more. He has reviewed more phones than most people will ever own. You canfollow him on Bluesky, where you will see photos of his dozens of mechanical keyboards.50 Comments

We can help with thatTwo AI-based science assistants succeed with drug-retargeting tasksBoth tools generate hypotheses; one goes on to analyze some of the data.John Timmer–May 19, 2026 2:55 pm|10Finding connections within the messy world of biology is central to these new tools.Credit:Andriy OnufriyenkoFinding connections within the messy world of biology is central to these new tools.Credit:Andriy OnufriyenkoText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navOn Tuesday, Nature released two papers describing AI systems intended to help scientists develop and test hypotheses. One, Google’s Co-Scientist, is designed as what they term “scientist in the loop,” meaning researchers are regularly applying their judgments to direct the system. The second, from a nonprofit called FutureHouse, goes a step beyond and has trained a system that can evaluate biological data coming from some specific classes of experiments.While Google says its system will also work for physics, both groups exclusively present biological data, and largely straightforward hypotheses—this drug will work for that. So, this is not an attempt to replace either scientists or the scientific process. Instead, it’s meant to help with what current AIs are best at: chewing through massive amounts of information that humans would struggle to come to grips with.What’s this good for?There are some distinctions between the two systems, but both are what is termed agentic; they operate in the background by calling out to separate tools. (Microsoft has taken a similar approach with its science assistant as well; OpenAI seems to be an exception in that it simplytuned an LLM for biology.) And, while there are differences between them that we’ll highlight, they are both focused on the same general issue: the utter profusion of scientific information.With the ease of online publishing, the number of journals has exploded, and with them the number of papers. It has gotten tough for any researcher to stay on top of their field. Finding potentially relevant material in other fields is a real challenge. If you’re focused on eye development, for example, one of the signaling systems used may also be involved in the kidney, and it can be easy to miss what people are discovering about it.As the people at FutureHouse put this issue, “By focusing on ‘combinatorial synthesis’ (identifying non-obvious connections between disparate fields), Robin effectively targets ‘low-hanging fruit’ that human experts may overlook due to the compartmentalization of scientific knowledge.”This is a task that’s well-suited to AI, which can chew through the peer-reviewed literature in the background while researchers do other things. This isn’t really a question of whether an AI could do something better or worse than a human; it’s more of an issue of whether any human would end up doing these sorts of searches at all.By finding enough connections among disparate research, these tools can make suggestions—hypotheses, really—about the biology. This can include things like what processes underlie biological behaviors and what pathways and networks regulate those processes. And, in the cases explored here, it included suggesting known drugs that might target some of these pathways in diseased cells: acute myeloid leukemia in Google’s case, and a form of macular degeneration for FutureHouse.Co-scientistAs you might imagine, Google’s system is based on the company’s Gemini large language model. That helps the system interpret a statement of research goals provided by human scientists and starts a literature search to find relevant information and form hypotheses. Those are then evaluated relative to each other in a “tournament,” the results of which are evaluated by a Reflection agent. An Evolution agent can then make improvements to any surviving ideas, which can be sent back through the process.Key criteria considered throughout this process include plausibility, novelty, testability, and safety. And the Reflection tool has access to external search tools, as access to the scientific literature “prevented the hallucination of seemingly novel but implausible hypotheses,” the company wrote.As the paper puts it, scientists were kept in the loop at all times. In the search for potential drugs targeting leukemia, the suggestions made by the system were prioritized based on a review by a panel of experts, who had access to the literature Co-Scientist used to formulate its suggestions.The results are what you would expect from cancer therapies. Some of the drugs identified were effective, but only against subsets of a panel of myeloid leukemia cells. That’s not unusual, given that there are multiple routes to unchecked growth, so drugs that block the route followed by one cell type may not be effective in cells that took a different route.Google also mentioned that the system could do more general hypothesizing that d

Closing the gap“I’ll buy 10 of those”—NASA science chief yearns for mass-produced satellites“How in the hell do I get more science into space? That is my goal.”Stephen Clark–May 19, 2026 4:43 pm|22Saturn’s moon Enceladus peeks over the limb of Dione during a partial occultation, as seen by NASA’s Cassini spacecraft on September 13, 2008.Credit:

      NASA/JPL-Caltech/Space Science InstituteText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThere are more opportunities to access space than ever, thanks to a bevy of commercial rockets, some with reusable boosters, led by SpaceX’s workhorse Falcon 9. So why is NASA launching fewer telescopes and planetary science missions than it did a quarter-century ago?The answer is complex. It is not necessarily the money. The space agency’s science budget this year is $7.25 billion, roughly the same as it was in 2000, adjusted for inflation. This is despite attempts by the Trump administration to drastically reduce NASA science funding.In the early months of his tenure, NASA Administrator Jared Isaacman’s focus has been on human spaceflight and the Moon. This isn’t terribly surprising given NASA’swildly successful Artemis II missioncarrying four astronauts around the Moon last month. Since taking office in December, Isaacman has announced an overhaul of the Artemis program,canceling a space stationto be built in orbit around the Moon in favor of construction of a base on the lunar surface.On the robotic front, Isaacman is pushing for NASA to launch afirst-of-its-kind nuclear-powered spacecraftin 2028 to deliver a trio of drone rotorcraft to explore Mars. Isaacman has not said as much about concrete changes to NASA’s science program. He has defended the Trump administration’s proposed cuts to NASA’s science budget—as would be expected of him as a Trump political appointee—but the budget proposals come from the White House, not from NASA headquarters.“Mr. Isaacman is very keen on us doing things quicker and for less,” said Nicky Fox, associate administrator for NASA’s science mission directorate. “More shots on goal is one of his favorite phrases. And I think, for us, it’s looking at the right-sized mission for the problem. Not everything has to be $1 billion or more. There are ways you can do fantastic science. His challenge is he wants 10 $100 million missions to be flying.”How to get there?A future with numerous robotic probes spread throughout the Solar System sounds thrilling to space scientists and space enthusiasts, but you can’t get there with flat budgets and billion-dollar missions that take a decade to get off the ground. Many of NASA’s robotic science missions use purpose-built satellites and instruments, usually manufactured by large contractors like Lockheed Martin, Northrop Grumman, university labs, or NASA itself. Unlike SpaceX’s hangars full of reusable rockets, there’s no building with cameras, spectrometers, telescopes, and spacecraft buses—the core chassis of a satellite platform—lying around waiting to launch.“Instead of having a bespoke bus that does absolutely everything, and makes the tea and brings you toast, what can you do with an off-the-shelf bus?” Fox told Ars. “And maybe you have to change a few things. Maybe you fly fewer instruments, but maybe you fly three [spacecraft] together. How do we really pick up the pace? Because it is difficult when you have long gaps between the missions. It’s certainly not what anyone wants to see.”One way to make this future real is with mass-produced, high-power satellites. Small CubeSats, just the size of a suitcase, are great for missions close to home, but they won’t cut it for missions to more distant destinations, such as another planet or a unique orbit far from Earth. NASA is making use of other ways to collect scientific data in space, such as placing instruments on the International Space Station or on commercial communications satellites.But those solutions won’t work if you want to travel to another world. Sometimes it just costs a lot of money to do the near-impossible.“For $100 million, you can’t buy a bus from somewhere and put four instruments on it and send it toflight to Enceladus to look under the icethere,” Fox said. “No, that’s a big, ambitious mission. We want to fly an interstellar-type probe. As theVoyagers are getting older, we want to study interstellar space. These things are hard, and they’re tough, and it will take a lot of effort to do that. We also talked about actually flying a mission to Uranus.”But what about spacecraft flying on more well-trodden paths to the Moon, Mars, Venus, or the asteroid belt? “What can we do with these commercial off-the-shelf buses? I would love to walk in and say, ‘I’ll buy 10 of those,’” Fox said.NASA islooking at “block buys” for the next series of commercial missionsto the Moon. These privately owned landers and orbiters, part of the Commercial Lunar Pa

Self-hostingPlex’s 200% Lifetime Pass price hike tries forcing users to another subscriptionPlex says that it has considered getting rid of Lifetime Passes.Scharon Harding–May 19, 2026 3:52 pm|158A marketing image for Plex that emphasizes its streaming, rather than its media server, business.Credit:

      PlexA marketing image for Plex that emphasizes its streaming, rather than its media server, business.Credit:

      PlexText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAs of July 1, at 12:01 am UTC—or June 30 at 8:01 pm ET—people seeking access to Plex’s media server features through a one-time purchase will have to pay $750. That’s three times the current price of $250.The new price will not affect current Lifetime Plex Pass holders.A Lifetime Plex Pass allows you to stream from your own Plex Media Server to a device connected to your own network, to stream from the server remotely, and to allow others to stream remotely from your server.When first launched in 2012, a Lifetime Plex Pass was $75. In 2014, Plex increased the price to $150 because “at 2.5x the price of a yearly, it was priced in an unsustainable way for us,” Plex said in ablog postat the time. Eventually, lifetime passes were available for $120 for years until March 2025, when the priceskyrocketed to $250. That brings us to today’s announcement.In an email andblog postto customers today, Plex said that it sold lifetime passes early in its existence “because we knew many of our customers would rather pay a higher, one-time fee for software that they can depend on every day.” However, as evidenced by several price hikes, Plex has struggled to reconcile lifetime passes with its own financial goals. The company’s message says:We’ve considered eliminating the Lifetime Plex Pass in the past, given that recurring subscriptions help us sustain long-term development, but we know it’s still a valuable option for many in our community. So instead of retiring it, we’re keeping it available at a price that reflects the real, ongoing value of the software we’re committed to building and maintaining for years to come.The 200 percent price hike and Plex’s admission that it has mulled killing Lifetime Passes illustrate Plex’s shifting priorities as itseeks profitability. Notably, Plex didn’t announce any pricing hikes for its monthly or annual subscription tiers.In its blog, Plex offered further reasoning for the upcoming price hike:Over the years, as our software and product has evolved, the breadth of features and benefits included with your Plex Pass has expanded. This increase ensures we can continue to invest resources into building and maintaining the Plex personal media software, while continuing to offer a Lifetime option.As noted in the announcement, Plex would rather people pay monthly or annually for its media server features than a one-time fee. With annual subscriptions currently $70, it would take 11 years for a $750 lifetime subscription to be a better value.Plex’s price hike is an extreme example of streaming service providers continuously increasing prices amid struggles to reach and maintain profitability while keeping prices stable and dealing with ongoing costs, like licensing fees for rental movies and app updates.Plex claims to be working on numerous new features for its platform, including: adding all server and library management features currently available on app.plex.tv to Plex’s mobile and, where applicable, TV apps; “boosting dialogue and normalizing loudness,” per today’s announcement; transcoding improvements; IPv6 support; and the abilities to group downloads by show, automatically download new episodes, and make and edit playlists in mobile apps. New features that Plex has added recently include the ability to create custom metadata agents and an open API for server integrations.Beyond new features, Plex has previously cited rising costs for the hike, including supporting various types of devices andcodecs.Plex moving beyond media servingPlex’s various forays into areas outside of its original media server business also contribute to its rising costs. In recent years, Plex has expanded from a media server company to a streaming service provider that hashundredsof free ad-supported TV (FAST) channels,rents out movies, and dabbles insocial features. It also made afailed venture into gaming.Legacy customers, however, have grown wary of Plex’s expanding interests and their potential impact on Plex’s media server users. While adding streaming and social features,Plex killed free remote streaming accessand itsWatch Together featureand issued acontroversial app redesign. Plex also has reason to put more focus on its streaming business since more people have been using Plex’s online streaming service than its media server capabilities since 2022, Scott Hancock, Plex’s then-VP of marketing,saidin 2023.The good news is th

The “Cage-aissance” continuesSpider-Noirfinal trailer gives us a classic villainIt’s never too late to become a hero.Jennifer Ouellette–May 19, 2026 4:58 pm|21Credit:

      YouTube/Prime VideoCredit:

      YouTube/Prime VideoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navPrime Video has released one last trailer for its upcoming live action series,Spider-Noir,starring Nicolas Cage, and once again it’s been released in two formats: one in black and white (below) and another in color (above), which the showrunners are calling “True Hue.” Seriously, the more footage we see of this series, the more eager we are to find out if the series lives up to its marketing. And the final trailer—which really plays up the deadpan humor and is set to Amy Winehouse’s “Back to Black”—is very promising.Aspreviously reported, Marvel Comics created its “noir” line in 2009, reinterpreting familiar Marvel characters in an alternate universe, usually set during the Great Depression in the US. A version of the Spider-Noir character, voiced by Cage, briefly appeared in the animated masterpiecesSpider-Man: Into the Spider-Verse(2018) andAcross the Spider-Verse(2023). (He is set to reprise that role in the upcomingBeyond the Spider-Verse.)Cage is playing Ben Reilly, a hard-boiled PI with a secret superhero identity, The Spider. Per the official premise: “Spider-Noirtells the story of Ben Reilly, a seasoned, down on his luck private investigator in 1930s New York, who is forced to grapple with his past life, following a deeply personal tragedy, as the city’s one and only superhero.”In addition to Cage’s Ben Reilly/The Spider, the cast includes Lamorne Morris as Reilly’s friend Robbie Robertson, a freelance journalist who clings to optimism in the face of his buddy’s cynicism; Li Jun Li as nightclub singer Cat Hardy, the classic underworld femme fatale (Li based her portrayal on Anna May Wong, Rita Hayworth, and Lauren Bacall); Karen Rodriguez as Reilly’s secretary, Janet; Abraham Popoola as a World War I veteran; Jack Huston as a bodyguard named Flint Marko who becomes (as we see in the new trailer) the classic villain Sandman; Brendan Gleeson as New York mob boss Silvermane, who is being targeted for assassination; Lukas Haas as one of Silvermane’s subordinates; Richard Robichaux as the editor of the Daily Bugle; and Kai Caster.“Have you seen it out there? The city’s a mess,” Robbie tells Reilly, reminding his friend of the hero he used to be. But the disillusioned Reilly knows that’s just not who he is now and he wants to start over—as a PI. But then Cat Hardy walks into his office and hires him to investigate a friend’s disappearance (killing an errant spider with a newspaper, which probably won’t endear her to Spidey).That friend is Jack Huston, who has “turned into some kind of sand monster.” And others also seem to be coming down with mysterious illnesses that worsen until “their sicknesses eat them alive.” Meanwhile, Silvermane has pretty much taken over the city. Will Reilly stick to his resolve and stay out of it? Or will he resurrect his superhero alter ego one last time? I think we all know it’s gonna be the latter. “I was never a hero,” Reilly insists. But it’s not too late for him to become one.Spider-Noirpremieres on May 25, 2026, on MGM+, with all episodes becoming available on Prime Video on May 27, 2026. Viewers can choose to watch in black and white or True Hue—or both, if one wants to compare.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.21 Comments

FBI wish listFBI seeks US-wide access to license plate cameras, wants “data in near real time”FBI will pay vendors to help it track and search for vehicles nationwide.Jon Brodkin–May 19, 2026 5:29 pm|109Flock license plate reader and camera with solar panel in Pleasant Hill, California on April 16, 2026.Credit:

      Getty Images | Smith Collection/GadoFlock license plate reader and camera with solar panel in Pleasant Hill, California on April 16, 2026.Credit:

      Getty Images | Smith Collection/GadoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe Federal Bureau of Investigation announced plans to buy nationwide access to a network of license plate readers, saying it will award contracts to one or more vendors that can offer “near real time” information from cameras across the US. The proposed contract is for the FBI Directorate of Intelligence.“To evaluate and manage threats to personal safety, property, and law enforcement, the FBI requires professional service firms that can provide License Plate Readers (LPRs) for tracking subjects on roads and highways over the US and its territories,” the FBI said in aRequest for Proposals (RFP)published on May 14. The FBI said the winning bidder or bidders “must provide law enforcement and/or commercial license plate reader data provided through the Contractor’s existing platform.” The system must cover 75 percent of locations, the FBI said.The system must offer the ability to search for license plate information “and other descriptive data such as vehicle description information, time/date criteria, and geo-location criteria,” the FBI said. “Additionally, the system must provide search result notifications. The Contractor system must have the ability to access and/or query cameras across the United States and its territories. The Contractor system must be capable of providing this data in near real time.”Contractors have to be able “to share/create maps depicting camera coverage (i.e. heat mapping),” and “provide the FBI the source of information (i.e. red-light cameras, repossession vendors, speed cameras, etc.),” the FBI said. The FBI said it needs to be able to search the database for partial or full plate numbers, plate states, addresses, locations where a plate was scanned, and vehicle makes and models.Flock and Motorola Solutions potential biddersThe RFP divides the proposal into six regions covering the continental US, Hawaii, Alaska, Puerto Rico, and territories such as Guam and the US Virgin Islands. The FBI said it may award contracts to one or two vendors in each region. The deals can be for up to five years, with all deals combined potentially worth $36 million. The FBI said a contractor’s system has to be available to FBI users via a website.Flock and Motorola Solutions are well-positioned to bid on the contract, as 404 Medianoted yesterday. Both companies could win part of the job, as the FBI said it may award contracts to multiple vendors to achieve its desired level of access.Flock’s Automated License Plate Readers (ALPRs) are sold to local police departments. The company boasts of having deals with “over 12,000 public safety customers including cities, towns, counties, and business partners.” Motorola Solutionssells license plate reader camerasthat can be installed on busy roadways or mounted on police cars.License plate reader cameras have raisedconcerns about privacy,data security, anderrors in plate number recognition systemsleading to wrongful arrests. 404 Mediareported last yearthat local police departments performed searches of the Flock license plate reader system for US Immigration and Customs Enforcement (ICE), “giving federal law enforcement side-door access to a tool that it currently does not have a formal contract for.”The FBI already “runs a License Plate Reader program to facilitate LPR information sharing with and between its law enforcement partners,” a Congressional Research Service reportsays. The US agency “maintains a hot list of vehicle data against which law enforcement agencies can compare their LPR data.”The FBI intelligence division’s plan to obtain direct access to an extensive network of cameras could help expand that information sharing. As the FBInotes, its intelligence division shares information with a variety of federal, state, local, and tribal law enforcement agencies.Flock: Sharing is “opt-in” for local policeFlock itself temporarily provided access to Customs and Border Protection, Homeland Security Investigations, the Secret Service, and Naval Criminal Investigative Service as part of a pilot last year. Flock confirmed the pilot to the office of Sen. Ron Wyden (D-Ore.),according to Wyden.Flock saysit has federal customers “including National Parks, Veterans Affairs hospitals, and military bases,” but that it does not work with ICE.Federal attempts to access data could be limited by company po

🤖曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini🌐马耳他将向全体公民免费提供一年 ChatGPT Plus📱小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务🚀目标估值 1.75 万亿美元，SpaceX 最快 6 月上市📋arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚💻卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元🚗蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌🔍需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位☁️最低 1 元 40 万 Token，三大运营商推出 Token 套餐🚙特斯拉 Robotaxi 碰撞事故报告解封💡知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值🚗小米 YU7 GT 火山灰配色公布💰ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构🧠马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布🤖字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本📱腾讯 vibe coding 产品「吐司」即将上线🛍️蜜雪冰城上架老黄同款蜜桃四季春，销量暴涨🛑中国银行关停信用卡独立 App，服务整合进主 App曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini昨晚，彭博社记者 Mark Gurman 在最新的 Power On 新闻通讯中透露了更多关于苹果即将发布的独立 Siri 应用的细节。隐私设计是此次新应用的核心卖点之一。Gurman 透露，该应用将内置对话自动删除功能，用户可在设置中选择对话保留时长：30 天、1 年或永久保存。苹果预计在今年的 WWDC 继续主打隐私保护，并强调其 AI 体验不含广告。模型方面，苹果已将新 Siri 的 AI 模型底层大量替换为 Google 的 Gemini，并将部分服务运行在 Google 的云基础设施之上。而此前有消息称，苹果和原本的合作对象 OpenAI 关系出现裂痕，甚至有可能对簿公堂，原因是 OpenAI 认为双方合作的结果不及预期。值得注意的是，尽管新的 AI Siri 已经跳票约两年，Gurman 称苹果内部测试版本仍将其标注为「Beta」，并可能沿用至其正式发布。马耳他将向全体公民免费提供一年 ChatGPT PlusOpenAI 日前宣布，与马耳他政府签署合作协议，向该国公民及居民免费提供为期一年的 ChatGPT Plus 订阅服务。据悉，居民须先完成由马耳他大学开发的免费在线课程「AI for All」，课程内容涵盖 AI 的基本概念、能力边界及负责任使用方式，通过后即可向马耳他数字创新局申请免费订阅资格。小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务小米 REDMI 产品经理「笋寸」日前在微博发文建议，近期不打算换手机的用户可以考虑等待换电池，并提示部分机型在更换新电池后容量可能还会有所提升，参考小米 13 系列的现行政策。笋寸进一步说明，小米正积极推动存量用户更换采用金沙江技术的大容量电池，小米 13 系列已率先上线相关服务。她还透露，具体支持机型「得等等」。此前，小米 13 系列已全系上线电池升级服务，支持小米 13、小米 13 Pro、小米 13 Ultra 三款机型升级更大容量电池，费用为 189 元，其中电池升级费 149 元、人工费 40 元。目标估值 1.75 万亿美元，SpaceX 最快 6 月上市据路透社报道，SpaceX 已加快 IPO 时间表，计划最早于 6 月 12 日在纳斯达克挂牌上市，股票代码为「SPCX」，较此前预定的 6 月下旬提前约两周。据三位知情人士透露，公司计划最早于本周三公开招股说明书，路演定于 6 月 4 日启动，股票发行最早于 6 月 11 日完成定价。时间表提前的主要原因是 SEC 对招股材料的审核进展快于预期。此次 IPO 预计募资约 750 亿美元，目标估值约 1.75 万亿美元，有望成为史上规模最大的股票发行。承销行阵容包括摩根士丹利、美国银行、花旗集团、摩根大通和高盛。arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚近日，预印本学术论文平台 arXiv 宣布，若作者提交的论文中包含无可辩驳的 AI 生成内容证据，将被封禁账号一年，并在此后要求所有提交须先经过同行评审期刊接受方可上传。arXiv 计算机科学板块主席 Thomas Dietterich 在 X 发文表示，若论文中出现「幻觉引用」、大语言模型元注释（如「以下是 200 字摘要，是否需要修改？」）等明显迹象，即可认定作者未对 AI 生成内容进行核查。arXiv 目前由康奈尔科技学院管理，今年 7 月将独立运营，成为一家非营利法人机构。康奈尔科技学院院长 Greg Morrisett 表示，独立后 arXiv 将能从更广泛的捐助方获得资金支持，以应对「AI 垃圾内容」带来的运营压力。卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元近日，小米集团合伙人、总裁卢伟冰在「Max 一夏」户外露营直播中表示，玄戒芯片今年肯定会推出迭代款，并称这将是「一块非常强的芯片」，会搭载在一款「非常优秀的产品」上。他同时提醒外界，目前网络上流传的各类相关传闻可信度不高，具体规格暂不便披露。卢伟冰还透露新一代玄戒芯片的应用范围将不再局限于智能手机，后续计划广泛搭载于小米旗下多类智能终端设备。关于手机定价，卢伟冰在直播中作出判断，认为今年下半年尤其到年底，部分国产旗舰直板手机价格可能突破 1 万元人民币。他预计，内存涨价的压力至少延续至 2027 年底，甚至到 2028 年。对于即将发布的小米 17 Max，他表示定价仍在讨论中，将在发布会上正式公布，但小米的原则是在同档位提供「最物超所值的产品」。蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌据睿见 Economy 报道，在日前举行的第十八届轩辕汽车蓝皮书论坛上，蔚来创始人、董事长兼 CEO 李斌线上出席，在对话中将萤火虫比作 iPhone，称其已成为「一款非常高品质的产品」。他强调，萤火虫在某种意义上超越了汽车的工具属性，演变为「一个生活方式品牌、一个时尚品牌」，并开始形成独特的品牌文化。他表示，萤火虫在用户群体上打破了年龄与购车预算的双重限制，「大学刚毕业的年轻人想买也能买，很有钱的老板想自己开车送小孩上学也完全没有问题」。需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位据《商业内幕》报道，前向部署工程师（FDE）的招聘需求在一年内激增逾 700%，正成为 AI 浪潮下科技行业增长最快的职位之一。招聘平台 Indeed 数据显示，今年 4 月该职位招聘帖数量达 5330 条，较去年同期的 643 条增长约 729%。Anthropic、OpenAI、Palantir、Stripe 及 Google Cloud 均在积极招募该职位人才。FDE 的核心职责是协助企业将 AI 整合进业务流程，并根据客户需求定制开发 AI 工具。薪酬方面，Indeed 数据显示该职位年薪约为 17 万至逾 20 万美元。最低 1 元 40 万 Token，三大运营商推出 Token 套餐中国电信昨日推出系列试商用 Token 套餐，分别面向开发者及中小微企业、个人及家庭用户提供「Token+连接+安全」一体化服务，均提供三档套餐，面向个人及家庭用户定价 9.9 元/月起 1000 万 Token；面向开发者及中小微企业 39.9 元/月，1500 万 Token。两类用户均可选购宽带上行提速包（上行速率提升 50 Mbps，100 元/月）及智能体安全管家服务（5 元/月）。个人家庭用户还可额外选购 5G-A 优享包（含 10 GB 流量，上行速率最高 300 Mbps，20 元/月）。联通、移动两大运营商也先后推出了各自的 Token 套餐：移动：中国移动上海公司于 5 月 17 日世界电信日发布会上推出 Token 通用服务，定价 1 元 40 万 Token。联通：中国联通上海分公司于 5 月 16 日发布全栈国产化「国芯国模国云」安全基座，推出多档 Token 产品及融合套餐。特斯拉 Robotaxi 碰撞事故报告解封近期，特斯拉正式解密并公开了向美国国家公路交通安全管理局（NHTSA）提交的 17 起「Robotaxi」自动驾驶碰撞事故的完整未脱敏报告。据 Electrek 报道，这 17 起事故发生于 2025 年 7 月至 2026 年 3 月，均涉及在得克萨斯州奥斯汀进行「Robotaxi」测试的 2026 款 Model Y，且测试时均有安全员在场。伤亡数据显示事故整体较轻微：13 起仅造成财产损失，2 起无人员受伤，1 起轻伤未入院，另有 1 起轻伤需入院治疗（系车辆在路口以 2 英里/小时蠕行时被后方 SUV 追尾）。官方报告表明，大部分事故并非特斯拉自动驾驶系统（ADS）的过失，多为车辆在红绿灯、停止标志或交通拥堵中完全静止时，被后方人类驾驶员追尾或侧擦。但数据同时也暴露出系统与后台干预的缺陷：两起事故发生在远程操作员接管期间，导致车辆分别以 8 英里和 9 英里的时速撞上金属围栏与施工路障；另有数起事故显示 ADS 存在空间感知死角，车辆在转弯或倒车时撞上了金属链条、拖车挂钩、木制电线杆及路缘石。💡 知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值据「听筒Tech」报道，知乎创始人兼 CEO 周源日前在第十二届新知青年大会上表示，AI 时代真实创作者的价值不可替代。AI 每分钟都在创造奇观，但 AI 时代仍然需要新知。而新知来自真实的人。每一个真实的人，都拥有不可替代的价值。大会主论坛上，多位嘉宾就 AI 与人类创作的关系展开讨论。复旦大学人文学者梁永安提出「科技有所为，如何让自己有所不为？需要人的尺度和人文属性」；作家张春称「不完美就是我们的完美」。周源在大会结尾总结道：「技术越发达，有温度、有深度的真人创作就越宝贵。」小米 YU7 GT 火山灰配色公布近期，小米汽车再度放出 YU7 GT 的预热图片，公布了新配色以及新车的内饰信息。其中小米 YU7 系列将加推一款名为「火山灰」的配色，官方介绍：灵感来自晨雾下的火山地貌，高纯度灰调，叠加纳米级矿物颗粒，呈现细腻而有层次的金属质感。同时新车内饰也迎来全新的「赛道红」：采用红黑撞色，提供专属刺绣。官方表示，小米 YU7 也会提供「火山灰」外饰配色，并且现在车厘子红的 YU7 GT 已经在部分小米汽车门店展示了。ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构近日，OpenAI 面向美国 ChatGPT Pro 用户推出了个人财务管理功能，支持通过 Plaid 连接超过 1.2 万家金融机构的银行账户和信用卡。连接账户后，ChatGPT 会自动读取用户的真实余额、消费流水和负债数据，生成一张覆盖投资、日常开销、订阅和近期账单的财务面板。用户提问时，模型可直接翻出账单历史，找出哪些类目容易超支，并给出具体的削减方案。此外，用户还可以告诉 ChatGPT 一些账外信息，比如「我在攒钱买车」或「我还欠父母一笔钱」，系统会将这些内容保存为财务记忆，用于后续对话。马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布马斯克昨日在 X 发帖披露，SpaceXAI 正在研发的 1.5T 参数 Grok 基础模型第 9 版（V9）已完成训练，预计 3 到 4 周内正式发布。现役公开版本 Grok 4.3 基于 0.5T 参数的 V8 基础模型，目前每隔数日迭代一次。V9 将参数量提升至 1.5T，马斯克称其为「重大升级」。在帖子的结尾，马斯克更是用用「banger」（王炸）一词来形容对这一版本的预期。字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本字节跳动 Seed 团队近日开源扩散

直到 2026 年五月，我们终于等到一整轮的影像旗舰登场完毕——小米做出了徕卡一瞬、vivo 打造了照片与视频兼具的 V 单、OPPO 在双长焦方案上带来了惊喜。在这轮旗舰更新潮的最后，华为提前将新一轮的 Pura 90 Pro Max 端上了桌。老实说，第一眼看硬件，心里是有落差的。去年 Pura 80 Ultra 上颇为惊艳的「一镜双目」结构不见了，取而代之的，是目前行业内最稳妥也最常见的大底长焦方案。不过，在细细体验后，我认为 Pura 90 Pro Max 藏了一个比纯拼硬件更有意思的新解法——XMAGE 智拍。全自动，都省心过去这几年，我系统性地梳理过华米 OV 的影像架构。虽然各家在硬件选择上各有理解，但大方向出奇一致，都在极其有限的机身空间里死磕传感器面积，比拼谁的光圈更大、进光量更多。在硬件上奋发图强理所应当。但在摄影中，不错的硬件基础仅仅是第一步。想要拍出一张好照片，构图与后期同样重要——构图保证一张照片在视觉上是否平衡、主体是否明确；而后期则决定了这张照片视觉质感和风格化。但这两者，向来都是具备门槛的手艺。而华为在 Pura 90 Pro Max 上，打算一口气解决这两个问题。熟悉他们产品线的朋友或许有印象，去年发布的 Pura 80 Ultra 上，华为试水了 AI 辅助构图。AI 辅助构图功能相对基础，在实际拍摄中主要起一个引导和提醒的作用。到了 Pura 90 Pro Max，这项技术完成了从「辅助」到「主导」的蜕变，并定名为「XMAGE 智拍」。剥开表层去探究 XMAGE 智拍的底层逻辑，是一套完全由端侧算力驱动的本地化影像流水线，将拍摄过程拆解为三个同时进行的模块——基于实时语义分割的自动构图、无视物理镜头限制的智能变焦，以及针对画面的色彩重塑。也就是说，在这个功能的加持下，Pura 90 Pro Max 可以自动识别主体、变焦构图，最后还排列出一系列适合当前场景的 XMAGE 风格供你选择。值得一提的是，XMAGE 智拍不需要网络连接，也无需将庞大的图像数据传回云端，全部基于端侧 AI 运转，实时识别画面里的主体轮廓、建筑线条以及光源分布状况，随后根据底层训练好的审美模型，给出一套包含最优构图、变焦和精细色彩方案的组合拳。繁复的专业参数在这里黑盒化，过去需要人脑去判断的构图和调色工作，全权交给了本地算法。除此之外，Pura 90 Pro Max 还顺势推出了 AI 姿势推荐，这个功能同样通过语义分割识别，可以理解画面中的了恩物、环境、姿势与背景，随后直接在取景框中勾勒出一个等比例的线框轮廓。拿手机的人只需按图索骥，引导模特贴合画面上的姿势即可。遇到不满意的动作能随时刷新，它甚至支持导入社交网络上保存的样片，一键提取动作精髓。听起来确实很厉害，但这套完全依赖端侧计算跑出来的流水线，究竟能不能应对现实世界里复杂多变的光线，还需要把它带到真正的街头去寻找答案。AI 有所能，也有所不能来到一家咖啡馆，墙上的多层光环灯饰引起了我的注意。举起手机，XMAGE 智拍自动变焦到 2×，将墙壁上的光环作为画面主体。这个选择可以说中规中矩，光环圆形处于画面中下方，整体构图稳定和谐。同时，操作台上颇具金属质感的咖啡机也是个不错的主体，XMAGE 智拍精准识别到我的拍摄意图，变焦到 192mm，保持咖啡机的中置。由于我想保留更为还原的室内光线与色彩，这张照片并没有套用 XMAGE 智拍推荐给我的任何色彩风格。用它拍了几组照片后，我认为 XMAGE 智拍是一个非常用户视角的功能——在传统的拍摄习惯里，手机焦段是一件需要精打细算的事。大脑会本能地去贴合物理镜头的原生焦段，生怕落在中间焦段会损失画质。但 XMAGE 智拍完全抛弃了执念。它不在乎用的是几倍变焦，转而完全为画面考量：当前这个画面，怎么裁切才最好看。哪怕不可避免地要损失一部分边缘像素，只要能让视觉中心突出、比例和谐，它就会果断动手。这种做法很实在，一切都在为了最终的出片服务。走出室外回望咖啡馆，绿植与橙色外墙交相呼应，XMAGE 智拍自动截取出色彩冲击力最强的部分，同时准确识别出咖啡馆建筑主体，两种大色块在画幅中占比相似，平衡而美观。再抬头，夏季树木繁茂，XMAGE 轻微放大焦段，让树枝形成天然的框架构图，广州塔正悬于画面正中间。在这组照片的拍摄中，我还察觉到了一个很有意思的技术细节。按下快门时，你会发现滤镜加载的速度，远大于构图改变的速度。这是因为滤镜往往只需识别画面后套用底层的色彩映射，而构图则需要端侧算力去实时识别画面里的建筑线条、光源分布和人物轮廓。计算量完全不在一个量级。更有意思的是，智拍套用滤镜的方式非常精细，除了 XMAGE 风格选择外，还会对调色盘做精细调整，确保风格对当前环境的适配度。当然，这套系统也有自己的脾气。在实际走街串巷时，XMAGE 智拍有时候会自动跳出来接管画面，有时候又毫无动静，需要你手动去点击唤醒。就算是手动唤醒的情况下，XMAGE 智拍也有几率出现风格已经选好，但当前画面无法找到最佳构图的情况。这种情况频繁出现在极其繁杂的环境中，由于元素过于混乱、信息量爆炸，算法的确很难从中剥离出合适的构图。当然，换个角度想想，也许此时的 1× 视角，就是当前环境的最佳构图了。万物皆可 AI，现在轮到摄影了185 年前，达盖尔发明银版摄影法，留住时间成为一种特权；2011 年，胶片巨头柯达传出破产消息，智能手机开始野蛮生长，把镜头塞进了普通人的口袋。回顾整部影像史，其实就是一部打破特权、技术走向大众的普惠史。到了今天，移动影像的狂飙突进，撞上了一堵无形的墙。手机镜头模组越做越大，与之对应的，是最近行业里一个颇具深意的传闻：受制于高昂成本与物理极限，下一代影像旗舰的「超大杯」，大多要面临停更了。这条路为什么走不通了？因为影像旗舰突飞猛进的背后，藏着一个被刻意回避的死结。决定照片质量的，永远是镜头后面的脑袋。普通人有懂美的眼睛，却跨不过光圈快门和构图比例的门槛。把一台堆满顶级硬件的手机，递给一个毫无基础的人，他按下的快门，大概率得到一张平平无奇的随手拍。顶级硬件如果只服务于一小撮懂摄影的人，无疑就会变成一种伪命题。在同行集体踩下刹车的时间点，华为提前把 Pura 90 Pro Max 端上了桌，而 XMAGE 智拍，就是破局的方法。在此之前，行业里已有不少 AI 落地，比如 vivo 让四季流转，OPPO 消除反光。只不过，它们大多是按下快门后的后期创意和修补。XMAGE 智拍往前跨了一大步，让算法成为整个拍摄动作的主导。这种做法谈不上是什么艺术层面的颠覆，但确实管用。降低门槛，让完全没有摄影基础的人，也能相对轻松地把一台旗舰机用出该有的样子。好的技术绝不该是一座孤岛。曾经的智能手机让人们拥有了「随时拍」的自由，现在的 AI 则进一步赋予了大众「拍得好」的能力。这或许是打破当前僵局的一剂良药，也是影像超大杯物尽其用的方法。

过去几年，20 万级新能源 SUV 市场像一条突然变宽的河道。一边是家庭用户对空间、续航和舒适性的刚性需求，另一边是智能化、性能化配置不断下放。过去很多只会出现在 30 万元甚至更高价位车型上的配置，如今开始像雨水一样往下渗透。激光雷达、空气悬架、大模型智驾、零重力座椅，这些原本带有「高配」意味的词，正在成为新一轮价格带竞争里的筹码。肩负着广汽集团高端化重任的昊铂品牌在今年广汽集团重振旗鼓的趋势下，也顺势重新梳理产品架构，昊铂 A800 旗舰轿车前段时间开始上市，用到了乾崑智驾和鸿蒙座舱来增强产品竞争力。紧接着，面向 20 万价位带的 SUV 昊铂 S600 也来了，提供纯电 660 Max、纯电 660 Max+、增程 230 Max、增程 215 四驱 Ultra+ 四个版本，官方预售价为 20.99 万至 22.99 万元，预售权益价为 18.89 万至 20.89 万元。新车定位为「新豪华智慧运动 SUV」，提供纯电和增程两种动力形式，车身尺寸为 5015×1933×1700mm，轴距 2936mm。相比于品牌初创时期的雄心勃勃，昊铂现在显然务实不少，它试图把昊铂 S600 做成一个配置密度很高的样本：性能、智驾、座舱、舒适、安全和权益，几乎都被塞进「满配」这个关键词里。昊铂 S600 请了 2 个代言人，一个代言「优雅」，一个代言「运动」过去提到中大型 SUV，很多人的第一反应是「稳」：空间要大，坐姿要高，后排要舒服，最好还能承担家庭出行里那些琐碎但真实的场景。比如周五晚上从公司直接出发，后备厢里塞进折叠椅、露营桌、孩子的滑板车和一箱零食；又或者长途返乡时，后排老人希望能把座椅放倒一点，孩子希望看动画，副驾则希望在高速上补一觉。对于这类用户来说，车首先不是一个「驾驶机器」，而是移动客厅、临时休息室，也是家庭生活的第二个容器。昊铂 S600 没有放弃这部分的诉求，或者说，这些配置已经在国产中高端 SUV 上成为一种基本素养，昊铂 S600 拥有 90% 得房率、750mm 二排腿部空间、5.316 平方米整车玻璃面积，以及 1292L 二排拓展空间；舒适配置上，同级唯一 18 点按摩零重力女王副驾、143° 躺角后排电调座椅、22 扬 ADiGO SOUND 音响、五块双层夹胶静音玻璃等配置也都被纳入其中。但昊铂 S600 又没有把自己完全包装成一台「沙发车」。它的另一面，是更强调姿态和操控的运动属性，所以昊铂又反复出现「跑山王」这个表达，显然昊铂希望把一台五米多长的大尺寸 SUV 做得不那么笨重，让它不只是适合城市通勤和家庭旅行，也能在山路、弯道、高速超车等场景中保留一定的驾驶参与感。从造型上看，昊铂 S600 采用的是一种前低后高的轿跑溜背姿态，超过 5 米的车长和近 3 米轴距提供了视觉体量，双层鸭尾、宽体车尾、20 英寸运动轮毂等设计元素，则试图削弱中大型 SUV 常见的厚重感，从侧面看它确实不像一般的家用大型 SUV 那样笨重方正，更像是专为年轻家庭所设计。所以这一次昊铂也请了 2 个代言人，英格兰传奇足球明星迈克尔·欧文作为「运动代言人」，知名演员陈数作为「优雅代言人」，两位代言人分别代表这一款车的一种特质，也是试图同时吸引家用取向和个人运动取向的潜在消费者。把四驱、高阶辅助驾驶、双腔空悬都放进 20 万级别在最高配的增程 215 四驱 Ultra+ 版本上，昊铂 S600 同时提供双电机四驱和双腔空气悬架。昊铂表示，这是行业首次在这一价位段同时放入双电机四驱和双腔空气悬架。另外，昊铂 S600 还全系标配前双叉臂后五连杆独立悬架、SDC 可变阻尼减振器和鹰爪系统 2.0，并提供雪地、湿滑、脱困三种路面辅助模式。一台大 SUV 在城市高架上并线，需要车身跟随方向盘动作，而不是像大船一样慢半拍；周末跑山时，连续弯道考验的是侧倾控制和车身收敛；雨天压过积水路面，轮胎抓地力和车身稳定系统要在驾驶者还没反应过来之前，先把车身姿态拉回来。在这些场景下，双腔空悬能够有效提升驾驶体验和驾驶信心。动力部分，昊铂 S600 提供纯电和增程双动力。其中增程四驱版综合功率达到 370kW，满电零百加速 4.3 秒，亏电零百加速 5.6 秒。广汽星源增程技术的持续发电功率为 85kW，官方还提到新车在亏电状态下可实现 165km/h 速度持续刷圈 2 小时，亏电极速 193km/h。增程车过去最容易被质疑的地方，在于「满电一条龙，亏电一条虫」，电量充足时像电车，电量不足后则容易出现动力衰减、油耗升高、发动机噪音变大等问题。昊铂 S600 对 85kW 持续发电功率和亏电性能的强调，其实是在回应这个痛点，一台增程 SUV 不能只在发布会数据里快，也要在长途、高速、爬坡和亏电场景下保持基本的体面。在智能化层面，昊铂 S600 的策略也比较直接：尽量把高阶配置做成全系标配。新车全系标配激光雷达、高通 SA8650 智驾芯片、27 个高精度感知单元，并搭载融合 Momenta R6 强化学习大模型的广汽星灵智行 ADiGO GSD 3.0，支持城区和高速快速路场景。资料还提到，昊铂 S600 下半年将首批升级 Momenta R7 强化学习世界模型。智能座舱方面，新车搭载高通 8295P 座舱芯片、17.3 英寸 3K 中控屏、27 英寸 W-HUD，并内置 11 个 AI 智能体。相比过去车机只负责导航、音乐和语音控制，现在的座舱正在变得更像一个移动终端：它不仅要响应指令，还要理解场景，甚至提前知道用户在车里可能需要什么。下班后坐进车里，导航自动给出回家路线，HUD 把车道级信息投到视线前方，孩子坐在后排看动画，车机根据车内温度和空气质量调整新风，座椅按摩开始工作……这些在新能源汽车上可以实现的体验，对于从 20 万元档位传统燃油车换购过来的用户来说，无疑是跃升式的。广汽集团有着庞大的用户基本盘，这些人群想要增购换购新能源车型的话，广汽自然希望用产品承接他们的需求。就像 AI 说的那样，我在这里，不偏不绕，不躲不逃，稳稳接住你。某种程度上来说，过去几年广汽集团新能源产品的建设，过于重视 B 端，对 C 端用户的吸引力不够，并没有「稳稳接住」用户的需求。对消费者来说，20 万级新能源 SUV 已经足够拥挤，智己 LS6、极氪 7X、小鹏 G7 等车型都在这个区间形成压力。昊铂 S600 要突围，关键命题并不是「它是不是拥有很多配置」，而是这些配置能否被组织成一种清晰的用车体验被感知到：城市里安静从容，山路上不显笨重，长途时不焦虑，家人坐进去愿意放松，驾驶者开起来也不觉得无聊。

龙虾之父一个月消耗 6030 亿 Token，总花费金额高达九百万人民币。移动联通电信，三大运营商都在推 Token 套餐，199 送千兆宽带还有 1 亿 Token，了解一下？从硅谷到国内大厂，Tokenmaxxing 成为公司的主流，谁消耗 Token 多，谁就是 AI 时代的好员工。00 后校友向母校捐赠 20 亿 Token，被网友调侃按 DeepSeek 5 元/亿 Token 计算，只要 100 元。▲图片来自新浪财经Token 在半年内完成了一次身份跃迁：从技术术语，到 KPI，到话费套餐，到捐赠货币。它成了 AI 时代的「度量衡」，唯一的问题是，没人说得清它到底在度量什么。我们自己买 Token，用公司的 Token，部署了一堆 Agent，代码、论文、周报都是 Token 烧出来的。而另一边是，大厂的员工由于 Token 消耗排行榜的原因，开始拿着公司的 Token 处理私事、玩游戏、开发数十个没什么用的子 Agent 来提升自己的排名。「回报」这件事很难量化，但「使用量」可以量化。于是所有人都选择了那个容易量化的东西。这不是 AI 时代的新问题，这是管理学的老病。用 AI 消灭狗屁工作的公司，正在制造新型狗屁工作亚马逊，那个裁员裁到大动脉，把自己的网站都变成 404 的小狗，最近又被爆出了新的「笑料」。原本被寄予厚望、用来消灭「狗屁工作」的 AI，最终却沦为制造新型「狗屁工作」的源泉。据《金融时报》报道，为了逼迫员工拥抱 AI，亚马逊搞出了一个极其复古的管理手段：「Token 消耗排行榜」，追踪每个员工的用量。公司强制要求超过 80% 的开发者每周必须完成 AI 使用指标，甚至将消耗 Token 的数量作为考核标准。▲图片

1.OpenAI 的两大宿敌 Anthropic 和马斯克，放下心中成见之后终于在月初结盟了。在此之前，Anthropic 和马斯克的关系并不融洽：今年 2 月，马斯克还在自己的 X 账号指责 A 社「woke」「邪恶」「反人类」（misanthropic），说这家公司「仇视文明」。事后来看，这次攻击并非马斯克清新脱俗的性格使然，而是 Anthropic 所做的某些事情触碰到他的神经，事出有因。在此之前，xAI 内部使用 Cursor 工作，但是今年年初员工发现，Claude 模型突然在 xAI 的 Cursor 公司账号里不能使用了。当时还在 xAI 上班的联合创始人吴宇怀，在全员信里是这么说的：「Anthropic 更新了政策，要求 Cursor 不得向其主要竞争对手提供 Claude 模型调用能力。」当时，吴宇怀在信中写了一句话，颇为有趣：「这是坏消息也是好消息。我们的生产力会被影响，但这也敦促我们开发自己的编码产品和模型。」为什么当时 xAI 的高层认为，开发自己的编码产品是关键？后来发生的事情，大家都知道了。xAI 的联创团队悉数跑路，马斯克一气之下对 Cursor 使用了钞能力必杀：上个月底，SpaceX 和 Cursor 共同宣布，将在编程和知识类工作 AI 模型的训练上，展开前所未有的战略合作；并且，SpaceX 还获得了以 600 亿美元收购 Cursor 的权利，或向后者支付 100 亿美元合作费用。注意编程这个关键定语，后面还会 call back.2.最近，我看了一条 Cursor 早期投资人、Anthropic 大喷子、T3 创始人 Theo Browne 的视频。本来点进去是看他喷 A 社和 SpaceX 怎么蝇营狗苟，结果没想到，却看到了关于 SpaceX + Cursor 合作的，一个既另类却又极度合理的分析：不说 600 亿的收购，就只说 100 亿的合作费——Theo 在视频里表示，自己认为「哪怕只是交换到 Cursor 的用户数据，这 100 亿也值回票价了。」所以是什么数据？如果你也去看 Theo 这条视频，他会讲得非常清楚。但为了节约时间，我们在这里简单概括一下：我们和 AI 的对话是一来一回的，你提出问题/需求，他给你解答；coding agent 同理，只不过返回的是代码。一次高质量的对话，整个过程，包括用户提示、模型思考、agent 规划、输出代码、验证——所有这些东西合起来，可以称为一个完整的 Agentic Loop——就成为了高价值的训练数据，再喂给模型去进行强化学习，就能进一步提高模型在实战场景下的表现水准。Cursor 有的，SpaceX 想要的，就是这些数据。可这些数据从哪里来呢？答案很简单：作为模型厂商，这种高质量数据的最直接来源，只能是你自己开发的 coding agent 产品——也即 Anthropic 的 Claude Code、OpenAI 的 Codex、Kimi 的 Kimi Code。现在你应该明白了，为什么被 Anthropic「封号」之后，吴宇怀会在全员信里提出开发 xAI 自己的 coding 产品和模型这件事了。这件事 xAI 在当时已经看清楚了：没有自己的编码产品，就没有高质量的强化学习数据；没有高质量的数据，就训练不出真正实战能力强的 coding 模型。虽然有点暴论，但现在我们可以点题了：模型厂商想做出来真正能打的编程模型，做自己的 coding agent 产品是唯一的路径。3.大语言模型像个水晶球，用全网的语料训练出来，似乎能够解答万物，但并不代表它在所有问题上都能给出高质量的答案。用 GitHub 上数以亿计的代码条目训练，当然也能训练出 coding 模型。这是「学习结果」的逻辑，也是没问题的。毕竟编码任务的结果是可以验证的：代码能不能运行，测试能否通过，结果摆在那里。但是，通往结果的过程，是一个涉及多步骤决策、错误纠正、意图对齐的复杂链条。每一次用户的接受、拒绝、补全、撤销、追问、甚至当模型好几次都搞不定或者完全搞错时的辱骂——都是这一链条上的过程信号。强化学习有两种监督方式，一种叫做结果监督，只看最后是否跑通。但是结果监督会催生「奖励黑客」的现象：模型为了能跑通可能写出冗余、脆弱、带逻辑漏洞的代码，但因为测试过了，模型以为自己学对了。而另一种叫做过程监督，对推理路径上的每一步进行打分。上述这些过程信号，只有在 coding agent 运行环境里才能诞生。GitHub 仓库里只有结果，哪怕是去看单独的提交历史，看 PR，都找不到有效的过程信号。在缺乏有效、自主可获得的过程信号的时候，一些模型厂商会采用「蒸馏」的方式，这个事情大家应该已经知道了。蒸馏的逻辑很简单，给同样的输入，老师模型输出什么，学生模型就学着输出什么。但是通过蒸馏，即便可以获取到思维链，得到的仍然更接近于结果，而非被蒸馏的老师模型内部的概率分布。一旦学生在推理中偏离了老师的轨迹，哪怕一个 token 不符合，都有可能发生偏离。这背后是强化学习的基础限制：策略梯度定理要求，优化样本最好由当前正在优化的模型自己去产生。这种数据叫做 on-policy 数据。而通过蒸馏别家模型，在别人的产品里产生的数据，来训练自己模型，都属于 off-policy 数据。模型当然可以从中学到东西，但学不到老师模型内部的概率分布信息。而像 Cursor 这样自己就是 coding agent 产品的公司，掌握着最真实、有效、高质量的训练数据。Cursor 产品本身，就是 coding 模型在实战环境中的最佳训练场。我们可以通过 Cursor 年初的「翻车」，来证明这个逻辑。4.APPSO 读者应该记得，年初 Cursor 发布了 Composer 2，号称「下一代专用编程模型」，技术报道写的相对保守，也没有提供具体的模型底座信息。结果很快，网友就在公开代码片段里发现了 Kimi 的模型 ID，截图传遍了开发者社群，逼得 Cursor 副总裁 Lee Robinson 出面澄清：「Composer 2 确实是从开源底座出发的。最终模型大约只有 1/4 的算力来自底座，剩下 3/4 是我们自己训出来的。」几小时后，Cursor 联创 Aman Sanger 也跟着发了一条道歉：「一开始没提 Kimi 底座是个失误。」五天后，Cursor 放出了完整的 Composer 2 技术报告，显示底座的确是 Kimi K2.5，授权方则是 Firworks AI，大致流程是在 K2.5 上做训练，再继续做大规模强化学习（RL）。但关键之处在于，Composer 2 的 RL 是运行在真实的 Cursor 会话当中，使用与生产部署完全相同的工具和 harness。Cursor 将这套流程叫做「实时强化学习」(real-time RL)，也即将模型的 checkpoint 直接部署到 Cursor 生产环境中，观察用户的响应，收集数据，聚合成奖励信号——最快可以每 5 个小时迭代一次模型版本，然后继续部署到 Cursor 里，循环往复。最极致的案例是 Cursor 的自动化代码补全功能 Tab，每天处理超过 4 亿次请求，每当用户输入字符、移动光标时，模型都会预测下一步动作，如果预测置信度高，则显示建议，用户按下 tab 即接受自动补全。该功能采用的是在线强化学习，在行业内极具特色。Cursor 可以以极高的频率（最快可达每一个半小时到两小时）更新 Tab 的模型能力给用户，直接在产品内收集 on-policy 数据进行训练。这种高频、接近实时的反馈回路，让 Tab 可以学习到极其微妙的用户意图。Cursor 方面透露，这种方法让 Tab 建议的拒绝率降低 21%，接受率提高了 28%。回到 Composer 模型本身。在事情搞清楚了之后，一些 Kimi 员工也删掉了之前吐槽的的推文，Kimi 官方账号发表了祝贺。一家估值 600 亿美元（基于马斯克给的数字），不做自己的模型基座的 coding agent 应用层公司，仍然可以通过产品自身的数据飞轮，RL 出超越基座模型的专有编程模型。所以与其说 Cursor 翻了车，不如说这反而是 coding agent 产品重要性的绝佳例证。Cursor 在另一篇关于实时 RL 的文章里写到：「（训练编程模型）最大的困难在于建模用户。Composer 的生产环境里不只有执行命令的计算机，还有监督和指导它的人。模拟计算机容易，模拟使用它的人却很难。」这句话，现正在逐渐成为了在编程模型方面走在前沿的模型厂商之间的共识。如果你去看 benchmark 榜单和用户普遍评价，会发现哪些头部的厂商都在发力做自己的 coding agent/编程产品。区别只在于谁离用户更近。我们以 SWE-bench、LLM-Stats 等相对权威的榜单为例，Claude、GPT、Gemini、Kimi 等模型基本霸榜前十，清一色都是有自己开发 coding agent 产品（包括 CLI、IDE、集成 coding agent 的桌面客户端）的模型厂商。在部分榜单上会出现少数反例，如 Meta (Muse Spark)、DeepSeek 等，没有开发自己的 coding agent。不过你会发现，这些反例模型，在更加接近真实场景、避免污染的更权威 benchmark 上就很难上榜了。以 DeepSeek 为例，它在 SWE-bench bash only 上分数是 70%，排名第九，在 SWE-bench Pro 上分数却掉到了 15% 左右。OpenRouter 的真实流量数据可以解释这种反差：该平台 2025 年报告显示，Claude token 消费 80% 以上用于编程和技术任务，而 DeepSeek token 消费主要集中于闲聊和角色扮演。没有自家 coding 产品的厂商，在一些 coding 任务 benchmark 上能挤进头部，但在更难的真实工程 benchmark 上，在用户用 token 消费投票的真实流量中，都会原形毕露。不仅是 Cursor，Anthropic 在 2025 年 11 月发的一篇论文里，也明确透露自己在做一模一样的事情：「我们在 Anthropic 自家的真实生产编程环境上做训练。」也即 Anthropic 把自己员工使用 Claude Code 的交互数据，反哺给 Claude 模型用来训练。5.在 AI 的演进历程中，生产要素的定义发生了深刻的位移。传统三大核心要素——算力、研究、训练数据，虽然在总量上持续增长，但在结构上已经出现了严重的失衡。今天的各大 AI 巨头显著提高了在算力上的资本支出 (CapEx)，让算力基建成为了当前舆论的主旋律。但实际上，特别是在编程范畴内，随着 GitHub 仓库、StackOverflow 等互联网公开代码数据被基模厂商「竭泽而渔」式地利用，模型在代码生成与逻辑推理上的边界开始逐渐显现。这也是为什么，行业共识正在逐渐转向一个冉冉升起的新战略高地：对于任何希望掌握顶级代码能力的模型厂商而言，建立自有的 coding agent 产品早已不再是可选的商业路线，而是确保底层模型可以持续进化的核心生命线。正如前面 APPSO 论证的那样，单纯学习公开数据等于只学习成功者的结局，却无法了解成功的路径，这绝对不是正确的成功学应该有的样子。在真实的编程环境中，知道发生了什么错误、怎样发生的、如何正确地理解和高效地实践需求等等——了解正确过程的价值，远超于得到正确结果本身。只有拥有自己的编码产品，模型厂商才能获取高质量的「过程监督」信号，从而在编码/推理能力的下一阶段竞争中，确保自己仍有技术护城河——否则就不得不像 SpaceXAI 那样，花钱去跟 coding agent 产品公司去合作。然而并不是所有模型厂商都跟马斯克一样有钱，以及 2026 年开始的巨头势力划分、结盟与领地的争斗会变得更加激烈，当一家缺乏自主 coding 产品的模型厂商终于回过味来的时候，恐怕已经没有足够的合作伙伴可以挑选，合作

长城今晚公布了魏牌 V9X 的上市价格：34.98 万元到 38.98 万元。这台大六座 SUV 的最终定价，比预售价下调了两万多。配合 5000 元抵 1 万元的现金膨胀礼，再加上本品牌 1 万元、其他品牌 8000 元的增换购补贴，实际的起售价已经被压到了 33 万这个位置。伴随着新车的亮相，长城汽车董事长魏建军走上了舞台。他接过了话筒，抛出了长城在造车底层架构上的思考：眼下行业里最偏好的往往是新能源车专属架构，魏建军偏偏逆风而行，将归元 S 做成了兼容燃油、插混和氢燃料的全动力平台。实际上，魏建军看上的是更广阔的海外版图。他表示，在很多国家，电力基础设施的建设依然迟缓，若用全动力平台去打全球市场，则能更好地把造车规模做大，将高昂的研发成本摊薄，从而稳住车辆的市场保有量和保值率，也让用户的修车费用降下来。对于新能源专属平台，魏建军在台上表达了自己的看法。实际上可以这么讲，我们近十年就是被这个电动专属平台忽悠了。现在，由于动力的多样化，配置的多样化，我们理论上可以卖到全世界去。造车的底子铺好之后，魏建军把话题拉回了最基础的用车体验。他反复强调，无论车企给出的承诺多么好听，最根本的准则永远是不能给用户添麻烦。为什么要添一句「不要给用户添麻烦」？你想想，当你在路上被撂下、车出了问题的时候，你的心情有多么焦虑。顺着老板定下的规矩，长城的工程团队把干货都掏了出来。魏牌 V9X 是一台车长近 5.3 米、轴距突破 3.1 米的全尺寸 SUV。为了拉动它，长城给它准备了一套超级 Hi4 混动系统和 2.0T 发动机，并搭配了一台 4 挡专用变速箱。在低挡位运转时，这套动力总成能爆发出 4500 牛·米的轮端扭矩，足以应对陡坡和极端烂路。长城还给出了明确的性能指标，声称不管车辆是满电还是亏电状态，都能跑出 4 秒级的百公里加速成绩；进入高速巡航状态，车辆又能把馈电油耗稳稳控制在百公里 6.3 升左右。底盘配置同样堆满了料。长城给 V9X 全系标配了闭式双腔空气悬架和 CDC 减震器，以及正负 10° 的后轮转向系统。这套后轮转向系统顺带接入了自动泊车功能，面对狭窄的侧方车位，后轮会自动微调角度配合入库，让泊车过程更加流畅。至于后轮转向对座舱空间的侵占，长城的设计团队顺着外扩的轮包，把第三排座椅做成了环绕式的大沙发造型，化解了结构上的劣势。在 V9X 的座舱里，这样的巧思还有很多，例如，副驾位置的手机无线充电板没有放在常规的中控台下方，而是被挪到了车门内扶手上，还带有一点微倾的角度。乘客的手自然搭在门板上，就能顺手拿到手机。智能化算力拉到了目前的顶格水平。英伟达 Thor 芯片负责辅助驾驶，骁龙 8295P 支撑座舱运转。虽然用的是别家的芯片，但长城在软硬件的协同上选择了全栈自研，把动力、底盘和辅助驾驶的控制权全部掌握在自己手里。魏建军表示，V9X 的各个系统之间能够实现毫秒级的响应，遇到突发危险情况，车辆可以提早判断，做到刹车的同时，也能进行绕行避障。他还强调，为了保护隐私，所有的语音指令和行程数据都做到了本地闭环，绝不外传。在加入了大模型后，车机里的「小魏同学」也变得更加智能了。发布会现场展示了一段测试，视频里的魏建军特意换了件外套，戴上墨镜坐进主驾。即便如此，系统还是在第一时间认出了他，并自动调出了他习惯的座椅位置和空调温度。发布会的后半程，舞台交给了魏牌 CEO 赵荣波，他把魏建军定下的准则，变成了一项项具体的购车权益。比如高阶智驾软件对用户终身免费开放，锁单后如果超过 30 天没交车，长城每天给用户补贴 200 元。其中最受关注的，是那份 6 年 15 万公里的整车质保。按照汽车行业的常态，普通质保通常只有三四年，部分品牌在遇到空气悬架、后轮转向这类高成本的底盘件时，还喜欢用易损件的名义另行算账。且这类零件的故障高发期，刚好卡在车子开到第四到第六年的时候，一旦脱保，维修和更换的费用动辄几万元。长城这次把空气悬架和后轮转向系统也放进了长周期的保修范畴里。赵荣波在台上强调，相关的服务条款没有任何文字游戏，也不设隐藏门槛。整场发布会的落脚点，最终还是回到了魏建军身上。这位长城的掌舵人拒绝了常规的营销套路，选择亲自上阵，直接拿自己的个人信誉和姓氏为这台新车兜底。面对网友关于长城会不会「翻车」的犀利提问，他没有回避，给出的回应干脆利落。我认为长城价值观和对用户负责任的这种精神，永远翻不了车。大家选择长城，我就要肩负起这份责任，让每个选择魏的人放心，这才是真正的代言。我的代言才刚刚开始。魏，你放心。

今年 3 月，苹果制造了 MacBook Neo 这个大爆款，把高质量入门笔记本的价格打到了 4599 元人民币起。一个半月后，这台搭载 A18 Pro 处理器的入门款 Mac，全年销量从 700 万台追加到了令人难以置信的 1000 万台，交付周期一度排到了 5 个月之后……苹果紧急向鸿海、广达等供应商追加订单。知名供应链分析师郭明錤给出更刺眼的数字：在全球 Windows 笔记本电脑预计同比衰退超 10% 的大背景下，2026 年 MacBook 系列出货量有望达 2500 万台，同比增长 20%-25%。这一拳，Windows 阵营的最核心代表英特尔，该如何接住？在今天，英特尔在中国正式推出第三代酷睿处理器，代号 Wildcat Lake「野猫湖」，并同时启动了联合产业链伙伴的「Firefly 萤火虫计划」。英特尔的接招，围绕一件事而展开：将最先进的 Intel 18A 制程，下放给低端，或是入门，或是「主流」——随便你怎么称呼它的价位段。野猫盯上新的猎物Wildcat Lake 这个代号，在去年底就已经在业界出现。今年 CES 期间，英特尔高调发布了主打高端的产品线：酷睿 Ultra Series 3，采用 Panther Lake 架构。同一天，Wildcat Lake 也悄悄披露，你可以将它理解为 Panther Lake 的精简版本。业界普遍将其视为英特尔为了对标苹果 MacBook Neo 而推出的产品，专攻平价入门级 PC 市场。「直接对标」这四个字，并不是说说而已。在 4 月底，英特尔为野猫湖推出的公版参考设计曝光，采用渐变配色、铝合金一体成型机身、窄边框键盘，扬声器开孔环绕在键盘两侧、D 面无出风口——这种设计语言不需要过多解释，它就是在用苹果惯常的设计语言，去回应苹果本身最受欢迎的多彩配色与铝合金轻薄机身策略。在 MacBook Neo 上，苹果做对了一件事：它让整个行业重新意识到，在「内存末日」(RAMmageddon) 的今天，入门价位段成了一座金矿。用一颗本来进入销库存阶段的 A18 Pro 芯片，配上 8GB 内存和 256GB 存储，开价 4599 元——MacBook Neo 活脱脱成了 MacBook Air 平替。英特尔，以及整个 Windows PC 阵营不能继续将这个用户群体拱手相让。而这，正是「野猫湖」存在的（几乎）全部理由。18A 首次下放到入门段第三代英特尔酷睿处理器的技术参数并不复杂，CPU 采用 2 个 Cougar Cove 性能核，加上 4 个 Darkmont 超低功耗能效核，总共 6 核 6 线程；GPU 部分配备 2 个 Xe3 核心，NPU 算力最高可达 18TOPS。整机基础 TDP 15W，最高 35W，在无主动散热的笔记本上，甚至可以压低到 11W 运行。按照英特尔今天提供的官方口径，这颗芯片功耗相比前代降低超过 60%，图形 AI 性能提升了将近 3 倍，平台总算力更是达到了 40TOPS。当然，考虑到「野猫湖」处理器的 NPU 算力只有 17TOPS，远低于 Windows 阵营的所谓「AI PC」，也即微软官方定义的 Copilot+ PC 40TOPS 的阈值。所以严格来说，野猫湖并不能直接和苹果 A18 Pro 高质量的神经网络引擎，以及统一内存架构，去做横向比较。话虽这么说，驱动参数量不大的本地小模型，做一些 AI 推理任务，还是绰绰有余的。以及更有意思的是，谷歌前段时间发布自家的 AI 笔记本电脑「Googlebook」的时候，也确认了和英特尔结盟——Wildcat Lake 被海外业界人士普遍认为将会成为是 Googlebook 的核心平台。连谷歌都觉得够用，应该没大问题吧？真正关键之处不在于这些参数，而是在于「野猫湖」用上了 Intel 18A 制程工艺本身。要知道，过去两代酷睿 Series 1 和 2，一直沿用 Raptor Lake-U 架构和 Intel 7 制程，实质上是 10nm SuperFin 工艺的演进版本。在主流芯片普遍走向 5nm 甚至 3nm 的当下，技术代际差距已经相当明显。当 MacBook Neo 用虽然是老库存但货真价实 3nm 工艺的 A18 Pro，杀进四千多元的价位段，英特尔过去那套「靠老制程压低成本」的入门策略已经过时了。所以，我们很欣慰地看到，英特尔将 18A——2nm 级别的工艺，压到一颗入门芯片上。这意味着英特尔充分愿意承担初期更高的单位成本，去力保一座不能再丢弃的城池。系统：手机供应链思路搬到 PC 上对于守住市场，「Firefly 萤火虫计划」同样重要。英特尔副总裁、中国区软件工程和客户端产品事业部总经理高嵩如是说：芯片本身的算力固然重要，但「芯片和系统层面的双重创新」才是这一代主流 PC 体验跃迁的来源。中国区客户端与平台销售业务部总经理宗晔则认为，PC 也正在从工具进化为「智能体验入口」。根据英特尔提供的内部数据，与上一代相比，「野猫湖」产品的 PCB 面积缩小了 5%，器件数量减少了 7%。同时针对第三代酷睿平台特性制定并推广标准化主板与 I/O 板连接，推动主板核心电路进一步实现模块化、小型化、低功耗化。这一套打法的精神底色，其实来自于手机。手机行业过去 10 年磨合出的供应链协同方式——元器件高度集成、模块化主板、统一接口标准，被英特尔搬到了入门级 PC 上。结果，就是更轻、更薄，也更便宜。参与「Firefly 萤火虫计划」的合作伙伴，包括但不限于华硕、七彩虹、荣耀、惠普、联想等。英特尔预计，全球合作伙伴将陆续推出超过 70 款基于 Wildcat Lake 平台的产品。目前，荣耀 MagicBook X14 2026 已经率先上架，搭载该平台 Core 5 320 处理器，采用 16GB+512GB 存储配置，定价下探到了 4399 元。用户画像：完全对标 MacBook Neo，但还有更多英特尔列举了 Wildcat Lake 平台产品的 5 类目标用户：辅导孩子作业的家长、全天在校的学生、刚入职场的新人、「一人公司」创业者，以及银发族。这些用户群体和场景描述本身没什么悬念，但值得注意的是，这个用户画像，和 MacBook Neo 在线下门店主要推荐的人群，几乎是完全重叠。MacBook Neo 夺走的本来就是入门级 Windows 笔记本电脑的客群，英特尔的接招，就是造出具有更好价值主张的入门级笔记本电脑，再把心智重新抢回来。英特尔还反复强调野猫湖平台的本地 AI 推理算力。这对于那些购机预算本就不多，也不太可能付费订阅优质云端模型算力的用户来说，倒也是个正确的思路。最后，野猫湖看起来像是合格的入门芯片平台，跟 Firefly 萤火虫计划结合起来，也形成了一套颇具诚意的组合拳。轻、薄、便宜的 AI 笔记本，英特尔现在重新回到了这个擂台。仅剩的悬念，是 OEM 最终能把零售价做到多少。在 MacBook Neo 的价位上，苹果给到了系统生态、统一内存架构、多种潮流配色等等诸多看头。但英特尔要想把这只野猫放出去，跟苹果大杀四方，光有 18A 工艺和优秀的工业设计是不足够的。这场仗的真正交战点，大概率不会在技术本身，而是在即将到来的 618 购物节期间，第一批搭载它的笔记本产品，能否真正把供货紧张的 MacBook Neo 杀个七零八落、措手不及。

⚖️马斯克诉 OpenAI 案败诉📱苹果开启 WWDC 邀请，将公布苹果 AI 新进展🚗问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处🤖国家反诈中心 App 上线「AI 内容鉴定」⌚苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动🚙广汽董事长回应被称网约车之王🧠三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能🎵腾讯音乐官宣收购喜马拉雅💰铠侠股价飙升，计划美国上市🚀马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元☁️百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半🎬爱奇艺一季度由盈转亏 2.95 亿元🦿荣耀机器人「闪电」「元气仔」走进南大💡前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去🚗小米 YU7 GT 定档周四发布🎮5499 元起，红魔 11S Pro 系列发布马斯克诉 OpenAI 案败诉据 CNBC 和《华尔街日报》报道，美国加利福尼亚州奥克兰联邦陪审团已就埃隆 · 马斯克诉 OpenAI 及其 CEO 山姆 · 奥特曼一案作出裁决，认定马斯克的起诉超出诉讼时效，驳回其全部主张。此案源于马斯克 2024 年提起的诉讼。他指控 OpenAI 及奥特曼违背了将该公司作为非营利机构运营的承诺，并在庭审中称 OpenAI「窃取了一家慈善机构」。陪审团的裁决并未就「违反慈善信托」指控的实质内容作出判断，而是认定马斯克的起诉超过了三年诉讼时效。法院同时驳回了针对微软的相关指控。裁决宣布后，马斯克在 X 平台发文，将这一结果定性为「日程上的一个技术性问题」，并表示将向第九巡回上诉法院提起上诉。OpenAI 首席律师威廉 · 萨维特在庭后表示，这「不是一个技术性裁决，而是一个实质性裁决」，并称马斯克是在将诉讼作为竞争手段使用。微软方面亦发表声明，对裁决表示欢迎。苹果开启 WWDC 邀请，将公布苹果 AI 新进展今天凌晨，苹果正式启动 WWDC26 邀请，并公布详细日程安排。本届 WWDC 大会将于北京时间 6 月 9 日至 13 日举行，面向全球开发者开放，重点展示 Apple 各平台的新系统能力、开发工具和框架更新。按照安排，WWDC26 将于北京时间 6 月 9 日开幕，届时 Apple 将举行主题演讲和 Platforms State of the Union，介绍新一代软件平台、人工智能相关进展，以及面向开发者的新工具。大会期间，开发者和学生可以观看超过 100 场围绕工具、技术和设计的视频讲座，参与团体实验室活动，并在 Apple 开发者论坛中与工程师和社区成员交流。Apple 还将在 6 月 9 日邀请超过 1000 名开发者、设计师和学生前往 Apple Park，参与线下特别活动。届时，爱范儿也将在 Apple Park，为大家带来第一手的现场报道。问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处全新一代问界 M9 系列汽车设计师 @Zimo-子墨 日前在微博发文，点赞全新理想 L9 的内饰设计，称其「简洁、逻辑、比例和谐，视觉重心清晰，口味控制得恰到好处」。他表示，初创公司在有强决策者的情况下，往往对设计管控介入较深，容易打散设计的贯穿逻辑；而到第二代、第三代产品时，无论前代成功与否，团队通常会更多放权给设计师打磨表达。他同时透露，新 L9 的内饰设计总监为来自宝马的 Chris Lee，并称虽然双方长期管理直接竞品，但「真的欢喜好设计，生在中国」，并已发短信向对方表示祝贺。国家反诈中心 App 上线「AI 内容鉴定」据央视新闻报道，国家反诈中心 App 近期正式上线「AI 内容鉴定」功能，支持对图像、视频、音频及文本进行 AI 生成痕迹检测，主要用于识别诈骗及电商场景中的伪造内容。电商平台上利用 AI 篡改图片骗取「仅退款」的行为日趋普遍。央视新闻记者随机选取十张商家提供的疑似 AI 生成售后图，上传至国家反诈中心 App 及多个 AI 鉴真工具后，均被判定为「图像含 AI 生成痕迹」。据介绍，该检测方案结合小模型与大模型：小模型学习 AIGC 内容的生成模式，大模型则判断内容是否存在物理瑕疵或逻辑语义问题，二者联合输出最终结果。苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动据 DIGITIMES 报道，苹果计划于今年下半年对高端 Apple Watch 机型进行重大改款，传感器功能将同步升级，市场预计届时出货量较 2025 年增长 20% 至 30%。Apple Watch 11 已具备血氧、脉搏、心率及血压感测能力，下一阶段预计推进无创血糖监测，具体落地时间取决于 FDA 审批进展。作为苹果传感器元件的独家供应商，亚洲半导体（TASC）预计将直接受益。市场预期新品订单最早于 7 月启动，带动下半年业绩明显超越上半年。广汽董事长回应被称网约车之王据「睿见 Economy」报道，日前广汽集团党委书记、董事长冯兴亚在「第十八届轩辕汽车蓝皮书论坛」上，针对「网约车之王」的说法，其指出：很多消费者可能没有想明白一个问题，「网约车是生产资料」。生产资料对品质的要求极高，因为它是用来挣钱的工具。只有能为自己带来更多利益的车，才能受到使用者和购买者的欢迎。对司机来说，车要「多快好省」，要能赚钱。「这恰恰是对品质最好的验证。没有高品质，就不可能受到网约车驾驶者的欢迎，这是最基本的。」他进一步分析，网约车的乘坐周期非常长，一天可能十几个小时都在车上。如果舒适性不够、疲劳感很强，它也不会受到司机的欢迎。此外，电耗等各项指标，也决定了网约车司机是对车要求最严格的一群人。冯兴亚总结道，这些从网约车运营中积累下来的经验，广汽会延展到旗下各个品牌，让车辆变得更加可靠、舒适和安心。「这反而是我们的优势所在。」三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能据 DIGITIMES 报道，三星电子近期大幅提升了用于高带宽内存（HBM）的最新 DRAM 良率。其第五代 10nm 级 1b DRAM 良率已达 92%，用于下一代 HBM4 产品的第六代 1c DRAM 良率也突破 75%。良率的提升有望增强三星的成本竞争力，并提升其向主要 AI 客户大规模供应先进存储产品的能力。在产能扩张方面，据 Ddaily 报道，三星和海力士正同步提前推进工厂扩建和设备订单。三星正加速平泽 P4 工厂的 DRAM 投资，设备安装预计持续至今年底，市场估计明年将进一步扩大投入。海力士则在清州 M15X 和龙仁 Y1 工厂推进扩产，业界预计其明年 DRAM 相关资本支出有望较今年翻倍。腾讯音乐官宣收购喜马拉雅腾讯音乐娱乐集团（TME）昨日宣布，已正式完成对喜马拉雅并购交易的交割，喜马拉雅成为 TME 全资子公司。交割完成当日，TME 执行董事长彭迦信与 CEO 梁柱联合向全体员工发布全员信。管理层称，喜马拉雅是中国长音频领域的重要建设者，双方战略方向高度一致，合并后将共同拓展在线音乐与音频行业空间。铠侠股价飙升，计划美国上市据彭博社报道，铠侠（Kioxia）股价昨日上涨 16%，创下新高。公司在上周披露强于预期的业绩展望后，多家华尔街机构迅速上调目标价；与此同时，铠侠也确认正准备在美国发行存托凭证，借此提升流动性并扩大国际资本市场接触面。报道提到，铠侠预计截至今年 6 月的季度营业利润将达到 1.3 万亿日元，高于市场平均预期，季度利润也创下纪录。多家机构认为，在 AI 带动的存储供需紧张环境下，公司仍有进一步提升自由现金流和股东回报的空间。马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元据彭博社报道，马斯克昨日在萨姆森国际智能移动峰会线上视频连线时表示，正在德克萨斯州推进 SpaceX 上市准备，称「SpaceX 的 IPO 事宜将很快开始」。知情人士透露，SpaceX 最快本周提交 IPO 申请文件，计划融资最高 750 亿美元，目标估值超 2 万亿美元，若成功将成为史上规模最大的 IPO。百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半百度昨日发布 2026 财年第一财季业绩，期内实现营业总收入 320.75 亿元，同比下降 1.16%；归母净利润 34.45 亿元，同比下降 55.36%；经营现金流为 27 亿元。具体来看，百度核心 AI 新业务收入达到 136 亿元，同比增长 49%，首次占百度一般性业务收入比重超过一半；其中智能云基础设施收入 88 亿元，同比增长 79%，GPU 云收入同比增长 184%。萝卜快跑在一季度的全无人自动驾驶运营订单达到 320 万单，总订单量同比增长超过 120%，截至今年 5 月其全球业务已覆盖 27 个城市，累计自动驾驶里程超过 3.3 亿公里。爱奇艺一季度由盈转亏 2.95 亿元爱奇艺昨日发布 2026 财年第一财季业绩，期内实现营业总收入 62.26 亿元，同比下降 13%；归母净利润由盈转亏 2.95 亿元，去年同期为盈利 1.82 亿元；经营现金流为 1.86 亿元，同比下降 45.13%。分业务看：会员服务收入为 42 亿元，同比下降 5%；在线广告服务收入 12.4 亿元，同比下降 7%；内容发行收入 3.59 亿元，同比下降 43%；其他收入 4.27 亿元，同比下降 49%。公司在财报中提到，热门剧集仍推动会员收入环比增长，海外业务保持稳健增长，并计划借助AI降低内容制作成本。荣耀机器人「闪电」「元气仔」走进南大5 月 17 日，荣耀携自研机器人「闪电」与「元气仔」在南京大学开展具身智能科普活动，并同步展示荣耀 600 系列新品的核心影像功能。活动现场，荣耀展示了自研机器人的核心运动与交互能力。搭载高动态运动系统、支持自主感知导航与高速避障的机器人「闪电」与在校学生进行了 8 轮 1V5 短跑测试；另一款机器人「元气仔」则完成了包含「太空步」在内的复杂舞蹈动作演示与人机互动。同期，荣耀在现场设立了新品体验区，重点展示荣耀 600 系列智能手机。官方信息显示，该机型主打基于 Magic OS 系统的 AI 能力及 4K Live Photo 动态照片功能。外观方面，荣耀 600 系列目前已公布「幸运星」「青苹果」「光羽蓝」和「曜石黑」四款配色，其中主打配色「幸运星」采用了行业首创的双重磁吸工艺。💡 前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去前 Tesla 无人驾驶与 SpaceX 猛禽发动机团队工程师 Zac Valles 在 YC Demo Day 结束 72 小时后直飞深圳，驻扎 8 周后携新硬件返回旧金山。他随后在 X 上发布了一篇长帖，向所有有意从事硬件创业的人发出忠告：不论有没有融资、团队甚至创业想法，都应该立刻去深圳。只要你对创办一家硬件公司有哪怕一点点想法，就去深圳。融资前、组队前、有想法前、离职前，都不是借口。去就对了。在切入供应链的路径上，Valles 建议创业者围绕一场大型行业展会安排行程，将其作为接触工厂的跳板。他还建议在华强北预留 4 到 6 小时实地走访，称其为一座「微型城市」，机器人子组件、电池、底盘、电子元器件均有专属楼栋聚集。小米 YU7 GT 定档周四发布昨日，小米正式宣布，将于 5 月 21 日（本周四）举行人车家全生态新品发布会，届时将发布小米 YU7 GT、小米 17 Max 等新品：小米 YU7 GT：定位跑车级 SUV，最大马力可达 1003ps，最高时速 300km/h，CLTC 纯电续航 705km；小米 17 Max：采用 6.9 英寸极窄四等边直屏，搭载徕卡 2 亿像素主摄和 8000 mAh 金沙江电池；小米手环 10 Pro：最薄 9.7 mm，最轻 21.6 g，续航最长可达 21 天；小米首款耳夹式耳机：单机重 5.5 g，支持 LHDC 5.0。5499 元起，红魔 11S Pro 系列发布红魔昨日正式发布红魔 11S Pro 系列，国内首发第五代骁龙 8 至尊领先

你帮我找一辆白色硬棚 Z4 M-Power，蓝色双轮毂，找那个富家子出来，我要找他试车。在 2012 年上映的港片《车手》里，余文乐饰演的警察陈翔留下了这样一句台词。不久前他在中环经历了一场惨败——那辆白色宝马 E86 Z4 M Coupe 在逼仄的车流和弯道中游刃有余，将他远远甩开。值得一提的是，现实中 E86 本就稀少，而 Z4 M 更是凤毛麟角，尤其是手动挡版本。这也难怪导演郑保瑞会用它来为整部影片定调，那种桀骜不驯的气质，恰如其分。当年大银幕上的稀罕物定格成了经典，而在现实的流水线上，属于 Z4 的齿轮也停止了转动。前一阵，奥地利格拉茨的麦格纳工厂内，最后一辆底盘代号为 G29 的宝马 Z4 缓缓驶出生产线，这款问世于 2002 年、历经三代演进的经典敞篷跑车正式宣告停产。伴随它的离去，宝马延续近三十年的双座 Roadster 车系，正式画上了休止符。巴伐利亚的驾驶图腾Z 系列的起点要追溯到三十多年前。在德语里，字母「Z」代表「Zukunft」，也就是未来的意思。90 年代的宝马对「未来」的理解非常具象，还带着一丝日耳曼式的偏执：剥离一切与驾驶无关的多余配置，把一台直列六缸发动机塞进一个小巧的底盘里，只保留两个座位和纯粹的驾驶乐趣。▲ BMW Z11989 年，采用下沉式隐藏车门的 Z1 惊艳亮相。这台车在今天看来像个古怪的科技实验品——它的车门可以通过电机垂直降入门槛内，车身覆盖件采用了可以由车主自行拆卸更换的特殊热塑性塑料。▲ BMW Z1除了这些前瞻设计，Z1 最重要的历史意义在于，它用全球限量 8000 台的桀骜姿态，确立了整个 Z 系列不按常理出牌、游走在主流边缘的先锋基调。六年后的 1995 年，Z3 走向了更广阔的大众市场。伴随着皮尔斯·布鲁斯南在 007 电影《黄金眼》中的潇洒亮相，这台远赴美国斯巴达堡工厂生产的 Roadster 迅速风靡全球。▲ 007：黄金眼中的 Z3但真正的硬核故事发生在幕后。当时宝马的工程师伯克哈德·戈舍尔（Burkhard Göschel）对敞篷跑车疲软的车身刚性极度不满。于是，他带领一群狂热的工程师突发奇想，利用周末的闲暇时间躲在工厂里，给一辆 Z3 焊上了车顶——这就是后来被车迷们戏称为「小丑鞋」的 Z3 M Coupe。▲ Z3 M Coupe虽然它造型怪异、空间局促，但封闭车顶使其抗扭刚性比敞篷版提升了 2.6 倍。在当时，它是纽博格林北环赛道上最让人胆寒，也最难驾驭的机器之一。前《Top Gear》主持人杰里米·克拉克森曾毫不掩饰地表达过他对这种老派双座跑车的偏爱：当一台车拥有那么长的车头，而你几乎坐在后轴上开车时，它就不再是一个交通工具，而是一台纯粹为了让人发笑而制造的机器。到了 1999 年，宝马 Z8 将这种暴力美学与复古浪漫推向了顶峰。▲ BMW Z8由亨里克·菲斯克操刀设计的 Z8，完美致敬了 50 年代的传奇跑车宝马 507。修长优雅的全铝车身骨架下，藏着与同时期 E39 M5 同款的 4.9 升 S62 V8 引擎。它顺理成章地拿到了 007 电影《黑日危机》的剧本，成为詹姆斯·邦德的新座驾，也让全世界记住了属于 Roadster 的优雅与从容。进入 21 世纪，Z4 正式接过了 Z 系列的接力棒。▲ BMW Z4第一代 Z4（E85/E86）诞生于克里斯·班戈主导设计的年代。那套著名的「火焰表面」设计语言，在当时引起了轩然大波。无论是前脸修长的敞篷版，还是车顶极具肌肉感的硬顶 Coupe 版本，侧面那道巨大的「Z」字形折线都带着一股咄咄逼人的凌厉感。好的汽车设计应该能够激发强烈的情感反应。你不需要向别人解释它为什么美，当人们看到它时，心跳就会加速。前宝马集团设计副总裁克里斯·班戈的这套理论，在第一代 Z4 身上体现得淋漓尽致。尤其是前文提到的 Z4 M。▲ Z4 Coupe宝马工程师把当年 E46 M3 上那台被奉为自然吸气神作的 S54 3.2L 高转速直列六缸发动机，塞进了 Z4 狭小的机舱里。它就像一匹没有完全驯化的野马，脾气暴躁，容错率极低，却又让人在征服它之后欲罢不能。2009 年，第二代 Z4（E89）问世。当时整个汽车行业刮起了一阵折叠硬顶风潮。宝马也顺应市场，为 Z4 换上了复杂的机械硬顶机构。▲ 外饰设计师 Juliane Blasi两位女性设计师，Juliane Blasi 和 Nadya Arnaout，赋予了 E89 极其优美的车身比例和圆润的线条，不过，硬顶机构不可避免地增加了车身重量，打破了原有的配重平衡。这台车在动态响应上做出了一定妥协，收敛了上一代的暴躁性格，演变成了一台更适合在海滨公路迎风巡航的 GT 跑车。▲ BMW Z4（E89）当时间来到 2018 年，第三代也是最后一代 Z4（G29）面临着一个严峻的市场环境。跑车市场不断萎缩，独自研发一款全新底盘的成本已经高到令人咋舌。因此，宝马选择拉上丰田，基于 CLAR 平台共同分摊研发成本。为了找回丢失的运动基因，G29 毅然抛弃了笨重的机械硬顶，回归了传统的软顶敞篷设计，不仅成功减重，更大幅降低了车身重心，找回了完美的 50:50 前后轴配重比。▲BMW Z4（G29）顶配车型搭载的 B58 3.0T 直列六缸发动机，提供了 340 马力的强劲输出与浑厚的声浪。更令人感动的是，在它的生命周期末段，宝马还出人意料地为直六车型推出了手动变速箱版本。工程师为其重新设计了换挡连杆和机械结构，只为呈现最纯粹的机械咬合手感。这种逆潮流而动的做法，被很多车迷视作一种浪漫的告别仪式。这台车从不妥协于日常琐碎，它存在的唯一理由，就是在踩下油门的那一刻，为你制造纯粹的多巴胺。理想与现实粤语里有一句老话，叫「有情饮水饱」。但在真金白银的汽车行业，情怀从来填不饱肚子。翻开宝马集团的财务报表，Z4 的谢幕早有预兆。G29 在 2019 年上市时还算风光，全年卖了 15827 辆。但这也是它最后的体面。2025 年，Z4 在全球只卖出了 9744 辆，这个数字放在宝马年销两百多万辆的大盘里，存在感微乎其微，占比连 0.4% 都不到。更大的问题在于，养这条产线还很烧钱。G29 并不在宝马自家工厂生产，而是交给奥地利格拉茨的麦格纳·斯太尔代工。一年卖不到一万辆，宝马却得单独为它保供一整套零部件体系。当年宝马联合丰田把 G29 推向市场，本就是一笔算计好的生意——分摊成本，换 Z4 多活几年。但合作研发终究只能解燃眉之急，无法逆转大时代的洪流。实际情况就是，卖掉一台 Z4 的利润，还比不上一台走量的 X3。前宝马集团研发董事傅乐希曾无奈地描述过跑车面临的困境。如今的汽车市场，开发一款跑车比以往任何时候都要困难。你需要高销量来分摊平台研发成本，而跑车市场的容量恰恰在不断萎缩。宝马现在一门心思扑在新世代纯电平台上。这是个烧钱的无底洞，几百亿欧元砸进去都不一定听得见响。新旧交替的关头，钱必须花在刀刃上，花在能决定明天还能不能活的地方。正因如此，砍掉投入产出比极低的 Z4，将宝贵的研发资源转移到更有前景的新能源项目上，是企业求生的本能。▲ 新世代 M3放眼整个行业，双门敞篷跑车的消亡已经成为一种趋势。奔驰 SLC 早就退出了历史舞台，奥迪 TT 也吹响了熄灯号。同级别的竞争对手一个个退场，Z4 能坚持到 2026 年才落幕，已经算得上长寿。宝马集团前 CEO 齐普策说过：「内燃机时代的落幕会让人感伤，但我们必须为未来的移动出行寻找新的答案。」接力棒已经交给了纯电车型，但我们依然会怀念内燃机，怀念沉重的电池组无法替代的那种轻盈灵动。一台 G29 Z4 的重量在 1.5 吨左右，而一台纯电跑车的重量往往会超过两吨。物理定律是客观存在的，增加的重量没有办法完全通过主动悬架来弥补。双座小跑车的魅力很大程度上建立在人与车的沟通感上，座椅和方向盘转向传递的路感，排气管的回火声，甚至是换挡时那一瞬间的顿挫，都在不断刺激着驾驶员的感官。没有了齿轮咬合时的震动，没有了直列六缸发动机逼近红线时的嘶吼，缺少了这些声浪与震动的刺激，速度在屏幕上只是一串苍白的数字。当时代浪潮决绝地退去，留在沙滩上的，往往是普通车迷的执念与阵痛。以后想要体验阳光、微风与汽油味混合的纯粹快乐，大概只能去二手车市场里寻找曾经的碎片了。而淘旧账的代价，往往比想象中疼得多。我的朋友张亦就花 10 万元买了一台 E85 Z4。买回来才发现，老跑车的维修称得上是个无底洞，配件是又贵又难等，单是内外三个后视镜片就花掉他 5500 元。买车 10 万，修车修了 5 个月又砸进去 10 万，到现在还没完全整利索。随着 Z4 的停产，维持这种老派乐趣的门槛只会越来越高。那个为了纯粹驾驶乐趣不计成本造车的黄金年代，连同它引以为傲的直列六缸声浪，就此封存在了过去。

2013 年，埃斯特庄园古董车展。为了纪念宝马摩托车诞生 90 周年，宝马带来了一台名为 Concept 90 的概念车。这个名字大家可能不太熟悉，但其实它的量产车型就是后来大获成功的现代复古标杆——R nineT，人送外号「宝马拿铁」。▲ R nineT六年后，同样是在科莫湖畔的这个庄园里，Concept R18 概念车登场，它顺理成章地开启了宝马 1800cc 水平对置巡航车 R 18 家族的时代。埃斯特庄园似乎成了宝马重磅复古车型的首发圣地。到了 2026 年的今天，宝马带着全新的 Vision K18 概念车再次回到这里。这辆概念车的造型区别于现有的 K1600，甚至和我们认知里的休旅车型都不太一样，更像是一个超音速飞行器。为了这个造型，宝马互换了进气和油箱的位置，并将后部线条尽可能压平，营造出一种低伏、前倾的姿态。哪怕它静静地停在聚光灯下，依然带有一种贴地飞行的视觉张力。长达两米的铝制侧板是这台车最引人注目的地方，呈现出一种无缝、一体的连续性。大面积的手工铝材与深色的锻造碳纤维搭配在一起，摒弃了多余的装饰，让整台车散发出浓烈的重工业味儿。数字「6」作为这台车的图腾，被具象化地刻在车头和车尾——前方是六颗带有主动散热功能的 LED 头灯，车尾的碳纤维包裹着六根粗壮的排气管。这套不锈钢排气系统采用了火焰喷涂工艺并覆有液态铝，质感类似于 F1 赛车的排气管。宝马用「The Heat of Speed」来概括这台车的设计主线。过去，那些为长途休旅准备的动力平台总是被包裹在整流罩内，安静且内敛。现在，宝马选择让这台排量达到 1800cc 的直列六缸发动机和排气完全裸露。车身结构向发动机和进排气让位，复杂的机械部件成为最抢眼的视觉元素。BMW Motorrad 首席执行官 Markus Flasch 说：通过 Vision K18，我们展示了对性能、豪华与情感的新理解。直列六缸发动机对我们而言远不止是一项技术，它是一种宣言，Vision K18 将这种力量转化为一种雕塑般的存在，让速度与动感在静止状态下依然清晰可感。从概念到现实尽管 Vision K18 在设计上追求的是速度与动感，但它始终是一辆为巡航而生的车型，它最终的量产车型很有可能就是 K1600 的继任者。或许，它会叫 K1800？▲ BMW K1600GT不管最终如何命名，这台新车最受瞩目的依然是那台排量提升后的六缸发动机。放眼目前的全球摩托车市场，愿意在两轮车架里塞进六个气缸的厂商屈指可数。在这之中，水平对置是最主流的选择。本田金翼那台 1833cc 的水平对置六缸就是行业内的典型代表，国内的长城灵魂用的也是水平对置八缸。宝马则选择了一条独特的技术路线——直列六缸。将六个气缸排成一条直线，会严重挤压摩托车的横向空间。为了控制车身宽度，宝马在 K1600 上就采用了非常紧凑的缸间距设计，气缸套之间的距离仅有几毫米。与此同时，它的缸体也向前倾斜了约 55°，既降低了整车重心，也为上方的进气系统和车架腾出了位置。这种结构带来了截然不同的特性。本田金翼偏向低转速和高扭矩输出，相比之下，宝马的直列六缸则带有一些运动轿车的特质：转速攀升更快，高转速下的声浪更具穿透力，动力输出也更加绵密。若是在此基础上扩缸到 1800cc，低转速时的扭矩想必也会有所提升。动力的升级让 Vision K18 做好了驶向公路的准备，但在实际骑行中，长途巡航除了需要充沛的动力，还要能够应对迎面而来的风和雨。这套动力平台真正进入量产、接替现款车型时，我们在概念车上看到的科幻外观大概率会向实用性让步。摩旅时，天气总是不似预期，大尺寸电动风挡、小腿前的整流罩、大容量边箱与尾箱，以及雷达、电控等系统，都会重新出现在量产车上。Vision K18 只是提前展示了这台机器剥去外壳后的原生状态，并且，它未来的量产形态可能不止于传统休旅车。参考宝马在 R 18 车系上的产品布局，依托同一台 1800cc 水平对置双缸发动机，他们衍生出了复古裸车、Bagger 巡航以及豪华休旅等多个版本。▲ BMW R 18全新的直列六缸平台也可能会复制这样的思路。它既可以加上整流罩，作为 K1600 GT/GTL 的正统换代车型进入休旅市场；也能像 Vision K18 展示的这样，推出一款褪去整流罩的复古版本，正好也能填补宝马 Heritage 继承者系列的产品线——R nineT 依托风冷水平对置双缸奠定了基础，R 18 切入了美式巡航市场，如果增加一台搭载 1800cc 直列六缸的休旅车，整个产品线会变得更加完整。

今日凌晨，苹果正式发布了今年 WWDC 全球开发者大会的邀请函。WWDC 26 将于北京时间6 月 9 至 13 日举行。今年的 WWDC，除了会正式公布「27」系列系统之外，开发侧重点仍将集中在 Apple Intelligence 后续进展、体验优化等等。自然，其中也少不了一些苹果针对 AI 时代的软件生态的提前布局。图｜MacRumors爱范儿今年同样将前往 WWDC 现场，为大家带来 WWDC 系列活动的专题报道，届时请锁定爱范儿主站、微信公众号、小红书和微博上的系列内容。iOS 27：玻璃、专业相机和 AI Siri对于 iOS，我们首先需要明确的是：液态玻璃（Liquid Glass）设计在 iOS 27 中不会退场，它就在那里，稳稳地接住你。就像刚更新的 iOS 26.5，苹果对于液态玻璃效果的微调也将持续到 iOS 27 里，直到大家习惯或者接受它的视觉风格。除了继续优化液态玻璃的视觉效果，iOS 27 也被很多媒体和开发者预测为是一次「雪豹」式的更新——比起新功能，删除陈旧代码、优化性能表现、调整视觉效果和改善整体流畅度将成为这一代 iOS 的主要更新方向。在支持范围方面，iOS 27「据称」将会支持到 iPhone 12 系列和 iPhone SE 3 为止。换言之，iPhone 11 系列及更早的机型将无法更新 iOS 27。而在继续优化液态玻璃的同时，iOS 27 还有可能会在一些小细节的层面，为我们「管中窥豹」一下今年 9 月份即将到来的折叠屏 iPhone。图｜MacRumors当然，苹果肯定不会在 WWDC 上直接展示有关大屏和横屏布局的优化措施。但在后续的开发者课堂中，大概率会出现有关自适应 app 布局、不同长宽比 UI 设计规范之类的相关内容。几乎可以肯定的是，今年 WWDC 上，苹果会在涉及到 iOS 的内容和课程里添加一些有关侧边栏和分屏（或者 Stage Manager）相关的内容，为横屏布局做出铺垫：另外，本次 WWDC 上我们还可能会见到苹果开放一些更加深层的相机 API，或者一些全新的相机玩法。这些本质上都是为 iPhone 18 Pro 上可能到来的相机「专业模式」做好准备。根据计划，苹果会让 iOS 27 的相机 app 更具自定义性，用户可以调整相机界的功能开关布局，比如闪光灯、EV、模拟光圈、照片风格和长宽比等等。图｜Shima Zamil最后不得不提的还有 AI Siri ——根据彭博社的预测，iOS 27 将会是 Apple Intelligence 官宣两年之后的「汇总落地」。iOS 27 版本的 Siri 中，苹果预计会拓展跨 app 信息提取、屏幕内容识别、模拟连续操作、app 深层功能集成等等多方面的 AI 能力，努力追赶甚至看齐 Gemini 在 Pixel 上的功能范围。图｜Google更有甚者，传说中的「对话式」Siri 也有希望在今年 WWDC 上亮相，或者至少看到一些相关的底层功能。相比以前那个问一句答一句的智能助手，Siri 将会获得类似 ChatGPT 那样的「人味」对话功能。并且在 Siri 对话中也支持多模态内容，甚至可能会取代原本的「聚焦搜索」功能：聚焦搜索（左）和 Siri 搜索（右）至于那些 iOS 26 里面已经存在的 AI 功能，比如图乐园、Genmoji、写作工具之类的，它们在 iOS 27 中也主要将迎来性能优化，并且在新的 Siri app 里面获得单独的入口。macOS 27：圆角、触屏和 Intel对于那些熬夜观看 WWDC 的 Mac 用户来说，看苹果整活王 Craig Federighi 怎样宣布今年的 macOS 系统名称，始终是不可或缺的一环。图｜MacRumors继去年 Tahoe（太浩湖）之后，苹果的商标注册列表上还剩下很多加利福尼亚州的特色。你觉得 macOS 27 的系统代号会是什么？和 iOS 27 一样，今年 macOS 27 的重点同样是继续打磨液态玻璃、优化系统性能，以及落地更多的 AI 功能。不过就像爱范儿之前提到的：macOS 26 的 UI 问题不止于液态玻璃，不统一的圆角、过于密集的小图标，以及辨识度过低的系统 app 图标，联手造成了 macOS 26 的可读性问题。图｜Reddit除了赶紧优化系统 UI 之外，macOS 27 的另一大潜在看点则是它对于触屏功能的支持。彭博社苹果专家马克·古尔曼去年和今年都曾密集爆料过，苹果正在计划开发一款采用触控 OLED 屏幕的 MacBook Pro，上市时间可能在 2027 年内。图｜9To5MacmacOS 27 虽然不会在 WWDC 上就大张旗鼓地宣传「新增触控支持」，但大概率会按照苹果的风格，做出一些为触控功能铺路的举措——比如新增多种尺寸的 UI 控件规范、优化 macOS 支持的缩放比例范围、新增一些看上去不太适合鼠标键盘的图标动画等等。AI 方面，macOS 27、iOS 27 和 iPadOS 27 预计都将迎来一次「相册」更新，照片编辑器里的 AI 工具不再只有橡皮擦，大家习惯的 AI 扩图、自动优化、重构图甚至 3D 照片等等都有望加入。图｜AppleInsider至于 AI Siri 本身的功能，比如「对话式」沟通、屏幕内容识别、系统级自动化功能等等，macOS 27 能够得到的更新和 iOS 27 差不多。这也是 iPhone、iPad 和 Mac 处理器同源带来的优势之一。坏消息是：macOS 27 极有可能会排除所有使用 Intel 处理器的 Mac 产品，仅支持 Apple Silicon 型号的 Mac。这就意味着 2019 款 16 寸 MacBook Pro、2023 款至强处理器 Mac Pro 等等顶级产品都将无缘 macOS 27，为苹果的英特尔时代画上第二个句号。图｜MacWorld另有开发者指出，苹果也在逐步停止 Rosetta 2（x86 到 ARM 转译）的更新。macOS 27 预计将是最后一个完整支持 Rosetta 2 的版本，敦促第三方开发者及时更新 ARM 架构的 app。稳扎稳打的 27 年总的来说，和谷歌努力推荐的「Android 全形态化」不同，苹果今年 WWDC 的路线依然是以稳健为主。iOS、iPadOS 和 macOS 各自的领域划分很明确，虽然产品形态更新，但系统短期内不会出现类似谷歌 Aluminum OS 那样的融合现象。图｜Geeky Gadgets对于「液态玻璃」来说，在缝缝补补一年多之后，苹果的确需要静下心来，先处理一些由于玻璃 UI 和 AI Siri 烂尾导致的系统基础体验问题了。至少对于一些旧设备来说，优化后的 iOS 27 很可能会比 iOS 26 更值得更新。附 WWDC 26 日程表，感兴趣的读者可以通过苹果开发者官网（https://developer.apple.com/cn）注册并参与在线课程：

1995 年，微软内部搞过一次颇具乌托邦色彩的尝试。他们觉得当时的 Windows 系统操作太复杂，于是想把电脑桌面直接做成一个普通人看得懂的「房间」。你想打字就点桌上的笔记本，想看时间就点墙上的挂钟，不用思考那些反直觉的操作路径。这个产品叫 Microsoft Bob，结果上市没多久就光速下架了。原因五花八门，但究其根本，它并没有真正渗透进系统底层，只要用户的需求稍微复杂一点，就只能无能为力了。三十年后，很多 AI 助手依然没有完全走出这个困境。最近我体验到了腾讯新的 AI 助手 Marvis 马维斯 ，终于看到了一些不同。相比于别人都在卷谁是 AI chatbot，Marvis 想干的是直接把整台电脑变成你可以随意使唤的搭档。官方给它的定位也正儿八经，叫「操作系统层级的个人 AI 助手」（官网：marvis.qq.com）。不过，当我知道它背后是那个做了 14 年应用宝的团队时，就不觉得意外了：这种长期在 PC 和移动端底层生态里打拼的出身，恰恰意味着它不仅懂 AI，更懂设备和系统。目前，Marvis 已率先覆盖 Windows PC 与 Android 手机端，iOS 和 macOS 版本也在推进中。APPSO 也在评论区准备了一些福利邀请码，手慢无。六个 Agent，一支随时待命的私人 AI 团队安装过程不算复杂，但对设备有点要求。Windows 端最低门槛是 8 核 CPU、16GB 内存加上固态硬盘。而且据官方透露，未来还会逐步兼容 6 核。市面上很多 Agent 产品动辄让用户自己去搭工作流、安插件，Marvis 则是开箱即用，六个 Agent 零配置上手，对普通用户非常友好。刚进界面，里面已经有六个 Agent 在 7×24 小时待命了。六个 Agent 都有自己的工位。没接到任务的时候，有的在打盹，有的在办公室里闲逛，有的在健身，有的在喝咖啡，还有的直接去上厕所，一旦你发出指令，任务就会拆解分配给对应的成员。这是一个分工明确的团队：PM 负责听懂你的需求，拆解任务往下派活；File Agent 处理文件的搜索、阅读和格式转换；Computer Agent 专门对付系统配置、查硬件、调夜间模式；APP Agent 负责在各种软件里代为操作，比如查个机票或者电商比价；Search Agent 负责快速从公开资源中定位答案并给出关键引用；最后是个 Browser Agent，定位是网页交互专家，进行网页交互与数据抓取。任务链比较长的时候，看看这些动画也是别有一番趣味。最有意思的是，这些 Agent 可以并行工作。举例而言，我先抛出一个需求：「电脑开机太慢了，帮我看看哪些自启动程序没用，帮我关掉。」注意看，工作日志里还会显示 Token 消耗量，每天的免费额度是 2000 万 Token，堪堪够用。结果显示，它给每个程序配了「人话版」说明，建议你哪些是可以按需、保留的第三方软件。重点是，涉及这种核心配置修改的操作，它会主动和你确认一下，才会执行。这种「AI 不会自作主张」的机制，稳得让人心安。电脑用久了，总会碰到几个反复折磨人的小问题。「查看电池健康状况」「网速感觉变慢了，帮我测一下现在的上传下载和延迟。」「想玩个新游戏，帮我检测一下硬件配置，对比最低要求看看差在哪。」以前遇到这些事，要么去开始菜单里找设置，要么去网上搜教程。现在，说一句话就行了。Marvis 深度对接了系统信息和配置入口。它能看到硬件参数，也能直接修改系统设置。它不是帮你打开「控制面板」然后让你自己找，而是直接把事情办了。对于那些不熟悉电脑设置的人来说，这块的体验是最直观的。举个例子：你想关掉 Windows 锁屏上的广告，你知道它在设置里叫什么吗？叫「关闭 Windows 聚焦」，有多少用户知道这个术语是关广告的？但 Marvis 能帮你一句话关掉。本地文件时代，迎来 AI 入口为了测它的能力，我还翻出了一件困扰已久的事：两个月前从某个群里下载的一张截图，只记得内容是关于《黑神话：悟空》，文件名完全忘了。我输入：「帮我找一下关于《黑神话：悟空》的图片或截图，大概是一个月前的事儿。」说实话我没抱太大希望。但 Marvis 接单后，首先根据 Skill 指引，按文件名关键词和时间范围筛选出 1000 张照片，然后用 Python 快速筛选整个结果文件，并最终找到 8 张候选图。第一张正是我要找的那张。而且它的搜索范围极其宽泛：文件名、文档内容、图片里的景、截图里的字，全覆盖。哪怕描述模糊，它也都能捞出来。这个能力，比 Windows 自带的搜索确实强出了一大截。当然，一听到「扫描图片」，可能会有人担心隐私。首次使用时你可以自己设定扫描范围。如果在生成的图谱里看到不想被索引的内容，右键一键屏蔽即可。而且端侧模型本身自带基础过滤，不是无差别的扫描。除了找文件，Marvis 还能对本地的文档、表格等多种类型文件做深度理解和分析——支持内容优化、图表生成、文案润色、格式转换，帮原本费时费力的事情变得顺手。比方说，下载了一批学术论文，摞在那里不知道从哪里读起。直接问 Marvis：「扫一遍这几篇 PDF，告诉我每篇的研究问题、方法和主要结论。」几分钟内，哪几篇值得细读、哪几篇可以先放着，心里就有数了，不用一篇一篇翻摘要。再比如读完《三体》，笔记记了好几页，马上要在课上分享，手动整理提纲、挑内容、套 PPT 模板，少说也要折腾一两个小时。但让 Marvis 根据读书笔记，生成一份 PPT，它能读懂笔记里的逻辑和层次，不需要手动复制粘贴再套模板，理解内容之后直接生成。手机接管电脑，随时随地都能干活手机遥控操作电脑，则是非常实用的跨端体验。去官网下了 Windows 客户端，再给手机安上 Marvis App，用同一个微信或 QQ 登录，还能直接连通两端。具体来说，手机 App 连上电脑后，可以在手机上实时看到电脑桌面，并用触屏直接操控。哪怕电脑处于锁屏状态，手机依然可以接管，不需要电脑一直保持解锁状态。比如我人在外面，突然听到朋友说前阵子新出了一款《洛克王国：世界》游戏，这时我就可以掏出手机输入：帮我在电脑上安装《洛克王国：世界》，回到家就能玩上了。连接微信 ClawBot，也是没问题的，微信扫码登录后就能开启对话。但我认为最实用的场景，是给家里长辈远程排查电脑问题。爸妈说「QQ 打不开了」，以前要么电话里盲猜，要么回家处理。现在直接接管他们的桌面，扫一眼就知道问题出在哪。前提是你的手机端 Marvis 马维斯，要能连接爸妈家里的电脑端马维斯，手机端是支持连接多个 PC 设备的。而如果电脑断网了，手机端会自动切到云 Marvis 模式，支持独立调用云端能力，大部分任务依然能跑。并且这里还有一个独家绝活：在电脑上直接操控手机里的 App。同花顺、飞常准等应用已经在对接合作，这也得益于应用宝团队多年的跨端引擎积累。现在应用宝在移动端、pc 端、车载上都有支持，跨端研发经验充足。和其他桌面 Agent 一样，Marvis 也支持自定义自动任务。比如设置一个「深圳厚德品园二手房均价提醒」，填好任务说明——「查询最新二手房均价并通知用户」，再指定执行时间为每周一上午 10 点，之后什么都不用管，到点自动运行任务，并直接推送结果。另一个值得单独说的功能，是技能（Skill）的自定义扩展。Marvis 支持一键安装专属技能包，提供了不少第三方 Skill，当然，你也可以导入自己常用的 Skill 文件，把它改造成更贴合个人工作流的样子。有特定行业需求的用户，比如内容创作，理财投资、学术研究等等，不用依赖官方迭代，自己就能把对应的知识库和操作逻辑装进去。最后，主打出活的 Agent 都绕不开最关键的问题：安全。Marvis 马维斯提供了两种模式：效率模式和隐私模式。设置里切到「隐私模式」，Marvis 会加载本地端侧模型，无论是文档还是图片的检索、解析、识别等，全部在本地运行，文件不会出你的电脑，甚至断网了都可以使用。官方所说的「数据不出域」，在拔掉网线的那一刻，有了最真实的体现。不过，体验「隐私模式」的门槛比较高，要求 16 核以上的处理器、32GB 内存、16GB 显存起步，固态还得匀出 35GB 空闲。所以平时日常使用切回「效率模式」即可，理解和规划由云端大模型（混元和 DeepSeek V4）完成，执行在本地，速度更快。两种模式按需切换。让 PC 真正开始「理解」人我们现在用电脑的方式，本质上还是上世纪八十年代的逻辑：一个需求对应一个软件，你要去找它、打开它、学习它的规则来完成任务。四十年来，软件更多，设置更复杂了，但这套逻辑并未改变，普通用户和电脑之间的距离，也没有真正缩短。Marvis 的逻辑反过来了：先理解这台电脑里有什么，再决定为你做什么。拿我爷爷来说，他用了几年电脑，至今不知道「控制面板」在哪。但他知道电脑开机慢，知道想找一张老照片。过去，这个解法是打电话求助子女；现在，就算他不会打字，也可以直接在微信 Clawbot 或者手机马维斯上，把需求说出来，电脑就会帮他做。文件搜索、系统设置、跨端控制、自动任务、隐私模式、Skill 扩展，看上去是几个不同功能，背后指向的其实是同一个方向：个人 AI 不应该只是一个问答入口，而应该成为设备里的任务调度层。多说一句，我觉得 Marvis 是有审美的。虚拟办公室动画、任务进度呈现等细节，让它不像冷冰冰的 AI 工具，更像一个有秩序、有分工、也有生活气的小型工作室。它尤其适合几类人：文件多且乱的本地重度用户；重视数据隔离的财务、法律、医疗从业者；喜欢自定义工作流的效率玩家。和需要靠「养」才能用好的「虾」和「马」，Marvis 初始内置 Agent 和自然语言指令，也足够覆盖多数日常电脑操作，对于很多 AI 焦虑不知道怎么用的人，Marvis 都更加的平权、无门槛、开箱即用。未来的个人 AI 会像搜索、截图一样，自然地成为操作系统的一部分。三十年前，Microsoft Bob 试图让用户更容易理解电脑；三十年后，Marvis 试图让电脑反过来，真正理解用户。Marvis 马维斯能走到这一步，和这个有十四年应用商店生态积累和四年跨端跨系统技术积累有关。庞大的用户基数让他们更懂用户在不同设备上需要什么、怎么使用。这种积累很难被纯 AI 公司在短期内平替——他们可以快速训练一个更聪明的模型，却没法快速建立起对 PC 底层和移动生态的控制力。至此，一个能用自然语言驱动的系统级 AI 助手，在中文生态里终于有了真正可用的样本。所谓个人 AI 重点从来不只是 AI，更是个人。Marvis 终于让我们看到，当技术选择与每一个具体的需求并肩同行时，它所释放出的力量，既润物无声，又重塑天地。这是个人 AI 真正开始变得有用的时刻。

💻Google 发布 Seedance 2.0 最强对手，还有一堆 AI 更新📱iOS27 快捷指令或迎来大改版，一句话就能制作🤖DeepSeek 回应模型异常回复🍎苹果发布 AI 加持的无障碍功能🖥️AMD AI 开发者大会首次登陆中国，苏妈提出「智能体主机」新品类🚗7 分 34 秒 931，小米 YU7 GT 刷新纽北纪录📱华为五一假期领跑中国手机市场、安卓厂商旧款机型均价上调 13%🚙小鹏 Robotaxi 量产下线，预计下半年开启示范运营🔧禾赛成为奔驰 L3 激光雷达战略合作伙伴，将于泰国工厂投产🌐特斯拉放弃印度建厂计划🧠美图布局 Visual Agent，Chance AI 完成数百万美元天使轮融资🏆小米影像算法 CVPR 2026 NTIRE 三项获奖💰长江存储启动 IPO 辅导⚡比亚迪电池CTO谈闪充、寿命与安全边界💡强化学习之父 Sutton 重申「苦涩的教训」引争议📱vivo S60 系列定档 5 月 29 日🎥698 元起，影石 Mic Pro 正式发布💻1299.99 美元起，微软发布新一代商用 SurfaceGoogle 发布 Seedance 2.0 最强对手，还有一堆 AI 更新今天，Google 在刚刚落幕的 I/O 大会上集中发布了一系列新品，包含两款新模型、全新 AI 智能体产品及音频智能眼镜预览，Gemini 的能力边界进一步向搜索、办公、创作和硬件延伸。AI：Gemini Omni 正式发布，定位「任意输入输出」多模态生成模型，支持视频生成、风格编辑与元素替换，可在生成过程中以对话形式修改；Gemini 3.5 Flash 同步推出，在编程、真实环境 AI 智能体任务等测试项目上较 3.1 Pro 有所提升，输出 token 速度比其他前沿模型快 4 倍，昨日起面向所有用户开放；Gemini 3.5 Pro 预计下个月开放；个人 AI 智能体 Gemini Spark 亮相，运行于 Google Cloud 专用虚拟机，支持全天候后台任务执行，可通过 MCP 协议与第三方应用互联互通。Google 同步新增每月 100 美元 AI Ultra 档位，原 250 美元最高档降至 200 美元；Gemini App 大改版，引入了全新的设计语言 Neural Expressive，加入流体动画、鲜艳色彩、新字体和触觉反馈。硬件：音频智能眼镜预览亮相，由三星参与硬件构建，Warby Parker 与 Gentle Monster 负责设计，搭载摄像头，支持 AI 视觉和多模态输入，Gemini 回答通过耳机播放，兼容 Android 和 iOS；XREAL 与 Google 在现场带来 Project Aura XR 眼镜首次公开体验，首发搭载 Android XR 系统，采用 XREAL 自研 X1S 芯片，视场角达 70°，确认今年全球上市。🔗 相关阅读：刚刚，Seedance 2.0最强对手上线！这就是AI视频的「Nano Banana 时刻」iOS27 快捷指令或迎来大改版，一句话就能制作据彭博社报道，苹果正筹备在即将发布的 iOS 27 和 iPadOS 27 中引入包括系统级语法检查、自然语言创建快捷指令及 AI 自定义壁纸在内的多项人工智能新功能。知情人士透露，苹果此次升级的核心在于扩展系统级 AI 写作工具，新增类似 Grammarly 的语法检查器。该功能界面以半透明菜单形式从屏幕底部弹出，并列显示原文与修改建议，用户可选择逐条采纳、全部应用或一键忽略，同时系统在键盘顶部集成了「通过 Siri 撰写」和「辅助撰写」快捷选项。此外，「快捷指令」App 也将迎来大幅改版。测试版本允许用户直接在「你希望快捷指令执行什么操作？」的文本框内输入自然语言需求，由系统自动构建和配置自动化流程，大幅降低使用门槛。同时，新系统将集成基于 Image Playground 的生成式 AI 壁纸工具。上述操作系统预计于今年 6 月的全球开发者大会（WWDC）上发布，并于 9 月向公众推送。WWDC26 将于北京时间 6 月 9 日至 13 日举行，爱范儿也将在 Apple Park，为大家带来第一手现场报道。DeepSeek 回应模型异常回复昨天，DeepSeek 官方小红书发布最新说明，回应了关于< think >字符触发模型异常回复的问题。DeepSeek 方面表示，近期其关注到有用户反馈，在与 DeepSeek 模型的对话中输入「」这类特殊字符后，模型偶发返回不可预期的内容，使部分用户产生了「对话泄露」的疑虑。针对这一问题，DeepSeek 的技术团队经全面排查后发现：输入「」等字符触发返回异常内容，属于特殊字符引发的模型幻觉，不涉及安全问题或隐私泄露。DeepSeek 方面也表示，后续其将通过针对性训练增强模型对特殊字符的识别与处理能力，修复相关的已知问题，优化模型在此类场景中的表现。苹果发布 AI 加持的无障碍功能苹果昨日发布了一系列无障碍功能更新，宣布将 Apple Intelligence 引入 VoiceOver（旁白）、放大器、语音控制和无障碍阅读器等核心辅助功能，计划于今年晚些时候正式推出。在视觉辅助方面，VoiceOver 新增「图像探索」功能，可对照片、账单等图像内容生成更详细的描述；放大器同步引入 AI 能力，在高对比度界面下提供视觉描述，并支持语音指令操控。语音控制则引入自然语言输入，用户无需记忆按钮标签，可直接用日常语言描述界面元素完成操作，初期支持英语，覆盖美国、加拿大、英国和澳大利亚。此外，苹果还公布了以下新功能：视频生成字幕：利用设备端语音识别，为未附带字幕的视频自动生成字幕，适用于 iPhone、iPad、Mac、Apple TV 和 Apple Vision Pro，初期支持英语，面向美国和加拿大用户；电动轮椅控制：Apple Vision Pro 借助眼动追踪系统为无法使用操纵杆的用户提供替代输入，首批支持 Tolt 和 LUCI 两款替代驱动系统，兼容蓝牙和有线连接；其他更新：visionOS 新增车辆运动提示；tvOS 支持大字体；「名字识别」扩展至全球逾 50 种语言；索尼 Access 控制器现可与 iOS、iPadOS 和 macOS 配合使用。AMD AI 开发者大会首次登陆中国，苏妈提出「智能体主机」新品类AMD AI 开发者大会昨日首次在上海举行，吸引超过 2000 名开发者到场。AMD CEO 苏姿丰发表开幕演讲，并与零一万物 CEO 李开复就 AI 智能体话题展开对话。大会重点推介「智能体主机」新品类概念，以锐龙 AI Max+ 系列处理器为核心，至高支持 96 GB GPU 专属显存，可本地运行最高 200B 参数模型。惠普、华硕、联想、宏碁等厂商已推出超过 35 款相关产品。软件层面，ROCm 7.2 新增对 Windows 和 Linux 的扩展支持，并新增兼容锐龙 AI 400 系列处理器。AMD 同时宣布与阿里云合作，提供基于 Radeon GPU 的免费开发者云，并推出面向中国开发者的专属项目。🔗 相关阅读：黄仁勋刚走，苏妈就来了 | 附演讲全文7 分 34 秒 931，小米 YU7 GT 刷新纽北纪录小米汽车昨日宣布，小米 YU7 GT 以 7 分 34 秒 931 的成绩刷新了纽北 SUV 圈速纪录。这一成绩出自小米汽车首席测试车手任周灿，他因此成为了首个获得纽北官方圈速认证的中国车手。雷军此前表示，小米 YU7 GT 是一台「原汁原味的纯血 GT」，充满驾驶乐趣的同时，还拥有舒适的长途驾乘体验数据上，YU7 GT 的最大功率为 1003 马力，极速可达 300km/h，CLTC 纯电续航为 705km。小米汽车 18 日宣布，YU7 GT 将于 5 月 21 日（本周四）19:00 正式上市。华为五一假期领跑中国手机市场、安卓厂商旧款机型均价上调 13%根据 Counterpoint Research 昨日发布的《中国智能手机周度销量追踪报告》，今年五一假期前后两周，中国智能手机销量同比下降 16%。存储价格上涨推高整机售价、促销力度弱于往年，是销量下滑的主要原因。华为继续领跑市场，4 月周度销量份额持续保持在 25% 以上。畅享 90 Pro Max 表现突出，Pura 90 Pro 系列及折叠屏新品 Pura X Max 于 4 月底上市后销量强劲；苹果五一假期销量同比增长 3%，但环比出现常规性放缓。iPhone 17 系列需求逐步回归常态，国产品牌集体涨价客观上提升了 iPhone 的相对竞争力；其他安卓品牌旧款机型均价已平均上调 13%，短期内或进一步压制换机需求。Counterpoint Research 预计，618 促销活动有望带动 6 月初市场温和回升；今年全年中国智能手机出货量预计同比下滑 9%，但仍优于全球平均水平。小鹏 Robotaxi 量产下线，预计下半年开启示范运营昨天，小鹏集团董事长何小鹏发文宣布，小鹏首台量产的 Robotaxi 在广州工厂正式下线。何小鹏表示，该 Robotaxi 以小鹏 GX 为原型车，搭载小鹏第二代 VLA 大模型，4 颗自研图灵 AI 芯片，本地算力达 3000TOPS 全球最高，实现 L4 级自动驾驶能力。其还强调，这也是中国首台纯视觉 Robotaxi+线控底盘+前装全冗余的量产下线。据悉，未来几个月，小鹏会在广州开启 Robotaxi 示范运营。禾赛成为奔驰 L3 激光雷达战略合作伙伴，将于泰国工厂投产5 月 19 日，全球化智能科技企业禾赛科技宣布，据公司 Q1 财报，禾赛成为梅赛德斯-奔驰（Mercedes-Benz）的激光雷达战略合作伙伴，并被确认为其 L3 级自动驾驶车型的供应商。据悉，此次签署的新供应协议将全面支持奔驰在欧洲及中国市场的车型项目，所需产能将由禾赛新落成的泰国「伽利略」制造中心提供坚实保障。特斯拉放弃印度建厂计划据 Drive Safe and Fast 报道，印度重工业部部长库马拉瓦米（Kumaraswamy）昨日正式确认，特斯拉已放弃在印度建设工厂的计划。分析人士指出，此次放弃建厂涉及多重障碍。印度本土供应链与工业生态系统存在明显缺口，难以支撑特斯拉的制造需求。此外，双方就「关税减免应先于建厂」还是「建厂应先于关税减免」这一核心问题始终未能达成共识。印度基础设施的不足、当地汽车市场消费能力与特斯拉高端产品定位之间的落差，以及印度营商环境中存在的政策不确定性与监管风险，也被列为主要原因。美图布局 Visual Agent，Chance AI 完成数百万美元天使轮融资Chance AI 昨天宣布完成数百万美元天使轮融资，由美图领投，NYX Ventures、阿里系投资机构等跟投，资金将用于模型迭代、北美增长及商业化探索。Chance AI 成立于 2025 年，核心产品是以摄像头为主要入口的 Visual Agent 应用，采用「看见—理解意图—调用 Agent—完成行动」的交互逻辑，可根据用户上传图片判断场景与意图，调度对应 Agent 执行任务。小米影像算法 CVPR 2026 NTIRE 三项获奖小米技术昨日宣布，在今年 CVPR 2026 NTIRE 赛事中获得三项奖项：高效超分辨率赛道冠军、人像修复赛道冠军，以及反光消除赛道亚军，分别由小米玄戒多媒体算法团队和小米大模型应用团队斩获。NTIRE 是图像恢复与增强领域规模最大的学术竞赛之一，由 CVPR 组委会承办。小米表示，后续将推动上述技术从研究走向产品落地。长江存储启动 IPO 辅导昨天，据证监会网站披露，长江存储控股股份有限公司在湖北证监局办理辅导备案登记，拟首次公开发行股票并上市，辅导券商为中信证券。比亚迪电池CTO谈闪充、寿命与安全边界据 36 氪报道，比亚迪电池事业群 CTO 孙华军就第二代刀片电池与闪充技术接受专访，详细回应了关于闪充寿命、电池热管理机制以及磷酸铁锂技术路线等行业争议。今年 3 月，比亚迪推出第二代刀片电池及闪充技术，支持 150

今天不追光和 vivo S60 系列一起做一颗属于自己的星星把 星星海 握在手里

版权所有 ©广州利沃致远投资管理合伙企业（有限合伙）2008 - 2023。以商业目的使用爱范儿网站内容需获许可。非商业目的使用授权遵循CC BY-NC 4.0。

All content is made available under the CC BY-NC 4.0 for non-commercial use. Commercial use of this content is prohibited without explicit permission.

「智能队列」如何改善你的家庭网络质量主作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org联合作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org2021/01/31 23:1205/17 12:57更新：2026 年世界电信和信息社会日重推在无线路由器的设置页面上，我们可能会看到一个名为「智能队列」的功能（根据品牌或操作系统的不同，也可能会被称做「智能 QoS」、「SQM」等）。这一功能是如何工作的？是否能让网速变快？是否需要打开？本文将尝试解答这些问题。从「网络如何连接」说起本章节主要介绍计算机网络的基础知识。如果读者已经熟悉相关知识，可跳过这部分内容。在介绍「智能队列」之前，先来了解一下基础的网络知识。两台电脑间进行数据传输，存在两种方式，分别是电路交换和分组交换。对于电路交换，两台电脑独占一条线路进行数据传输。过去的固定电话，大多使用的就是电路交换。由于独占一条线路，两台设备之间的通信较为稳定，很难出现数据丢失、数据传输顺序错乱等情况。但一条线路只能供一组电脑使用，另外一组电脑想要通信，就不能复用原来的线路了。对于分组交换，传输的数据被拆分为小段，每一小段叫做一个数据包（报文）。数据以报文的形式进行转发，从而实现了多组电脑复用一条线路进行通信。电路交换虽然能保证数据传输的稳定性，但由于独占线路浪费太多资源，所以在 Internet 中，使用的是分组交换。分组交换将数据拆分成一个一个的小片段来转发，没有独占线路。如果网络条件不好（例如同时有大量报文经过路由器，超出了路由器的处理能力），很有可能遇到数据顺序错乱、甚至数据丢失等情况。那么，这些情况应该如何避免？在微信视频通话、电视直播等应用中，报文的丢失或乱序，只会造成视频图像卡顿，不会造成太严重的影响。所以这些应用会使用简单的UDP协议，不考虑丢包、乱序。在网页浏览、文件下载等应用中，一般使用基于TCP的HTTP协议。TCP 在报文中携带序号，避免乱序；同时提供了报文的丢失重传机制，避免了丢包。所以，也可以认为 TCP 为两台电脑的应用程序建立了一个虚拟的「连接」。（最新版本的 HTTP/3 协议，使用的则是QUIC。QUIC 虽然基于 UDP，但引入了与 TCP 类似的报文丢失重传机制）如果想要了解更多网络基础知识，推荐阅读「入网指南」系列文章：入网指南 01 | 一文读懂你身边的「网络」入网指南 02 | 当你发消息的时候，发生了什么入网指南 03 | 交换机在交换什么？入网指南 04 | IP 地址大揭秘入网指南 05 | 除了接线，还有什么影响着我们网上冲浪入网指南 06 | 连上 Wi-Fi 就能上网了吗？网络延迟：影响上网体验的重要因素带宽和延迟，是影响上网体验的两大因素。带宽指的是在单位时间内，网络最多能传输多少数据。网络的带宽和水管的粗细有着很多相似之处，水管越粗，相同时间流出的水越多；同样地，网络带宽越大，相同时间能够传输的数据越多。在日常上网过程中，大的带宽能带来这些好处：上传、下载文件的速度更快在线看视频时，缓冲的时间更短如果家中有多个人同时上网，大的带宽能够保证每个人都能流畅上网，相互的影响更小对于家庭宽带，我们常常听到的 10 兆、50 兆、100 兆、千兆，就是指的带宽的大小。延迟指的是数据从网络的一端到达另一端，所需要的时间。过大的延迟也会对上网体验带来不好的影响，例如：在浏览器中输入网址敲回车后，需要等更长的时间，网页才能加载；视频聊天、语音通话质量变差，甚至能直接感受到声音、画面的滞后；玩在线游戏时，发现游戏变慢变卡。在很多时候，延迟相比带宽更容易被人感知，对网络体验的影响更大。是什么导致了网络延迟？上文中介绍过，Internet 使用的是分组交换。如果大家阅读过计算机网络相关的书籍，应该对分组交换的「存储-转发」机制有一定的印象。简单说，「存储-转发」是指网络设备将收到的报文存储在缓冲区，然后再发送。这是由于网络设备的性能有限，转发报文有速率限制。如果设备短时间内收到大量报文，一时半会儿处理不过来，就可以先报文存起来，空闲时慢慢发送。通过「存储-转发」，能够提高网络设备的吞吐量，更充分地利用带宽。其中，存储报文缓冲区，是一个先进先出的队列。那么，队列的长度应该如何确定？是不是队列越长，能够存储的报文越多，网络的性能就越好？大家可以先思考下这个问题。其实在过去，内存价格比较昂贵，考虑到成本，网络设备的队列不会太大。现在，随着内存价格的下降，网络设备拥有了更大的队列。但是，网络的带宽是有限的。实际的可用带宽，取决于数据传输路径上的最小带宽。这个最小带宽也被称做带宽上的「瓶颈」（bottleneck）。由于瓶颈的存在，如果流量长时间以较大速率发送，大的队列并不能显著提升吞吐量。相反，大量报文滞留在队列中，会带来延迟的增大。这种队列中缓存有大量报文，来不及发送，导致数据传输延迟增大的现象，被称做 bufferbloat（缓冲区膨胀）。DSLReport 提供的测速服务，能够检测网络的 bufferbloat 程度。不过由于没有中国大陆服务器，测试结果不够准确。国外的读者可以尝试一下：DSLReports Speed Test家用无线路由器中的 QoS 功能在「智能队列」出现之前，很多家用路由器已经提供了 QoS 功能。那么，传统的 QoS 如何解决网络延迟、网络拥堵？存在哪些不足之处？在了解这个问题之前，先来熟悉一下网络中的报文，是什么样子的：在网络中传输的报文，一般包括了链路层、网络层、传输层、应用层这四层信息。其中，网络层的源 IP 地址、目的 IP 地址、协议号，和传输层的源端口号、目的端口号，是 QoS 算法中常常会关注的五个信息。这些信息合称「五元组」。同一个应用在同一时间段内，传输的报文中，这些信息一般是不会发生变化的。例如观看在线视频时，同一个视频的所有报文 ，一般拥有相同的五元组信息。所以，通过五元组，我们能够区分不同应用的流量。特别地，一些常见的应用，会使用固定的端口号。例如网页浏览一般使用 80、443 端口；英雄联盟游戏，使用的是 5000-5500 端口 ……所以，当我们想要提升英雄联盟游戏的体验时，在路由器上设置让 5000-5500 端口的报文优先通过即可：How to set up traffic prioritization rules for League of Legends on Nighthawk Pro Gaming router | Answer | NETGEAR Support家用路由器中常见的 QoS 功能，就是通过手动设置 IP 地址、协议号、端口号等报文特征来实现的。这种方式需要用户了解一定的网络知识，且操作复杂。（虽然部分无线路由器内置了常用应用的 QoS 规则，但仍需要一定的手动设置，且无法及时适配新的、小众的应用）当然，除了上文中介绍的内容，QoS 还有多种不同的实现方式。例如：对特定应用（例如 BitTorrent）进行限速按照 MAC 地址指定某台设备的优先级（例如提高 Nintendo Switch 的优先级，使游戏更加流畅）通过WMM、DSCP等机制来决定报文优先级这些机制各有各的特色和不足。例如对应用程序限速的方式，虽然效果明显，但无法在网络空闲时，动态利用剩余带宽来提升传输速度，也难以适应带宽频繁变化的无线网络。关于上述 QoS 机制的技术细节，本文不再详细介绍。如果感兴趣，可以搜索相关关键词，进一步了解。智能队列：通过「全自动」的方式优化网络、降低延迟什么是「智能队列」目前，一部分新的无线路由器，开始有了名为「智能队列」的功能。这一功能一般指的是 fq_codel。fq_codel 包含两种队列管理机制，分别是 Flow Queue 和 CoDel。在下文中，我将简要介绍这两种机制的作用。Flow Queue前文中介绍过，网络中的报文一般都有自己的五元组信息；同一应用发送的报文，在一定的时间段内，五元组信息是相同的。这些信息相同的一系列报文，被称做一条流（flow）。Flow Queue 为每条流单独维护一个队列。例如，如果我们一边在线观看视频，一边进行语音通话，一边玩游戏，Flow Queue 就会为视频播放的流量、语音通话的流量、在线游戏的流量各维护一条队列：正常情况下，在线观看视频需要占用较大的带宽，但对延迟不敏感，用户可以容忍几秒甚至几十秒的缓冲时间。而语音通话和在线游戏占用带宽小，但对延迟敏感。例如语音通话时，如果一个人说了一句话，对方在几秒后才能听到，用户就能明显感知到延迟。如果三种应用共用一个队列，队列中会充满大量视频播放应用的报文，导致语音通话、在线游戏的报文迟迟无法发送，增大了这两种应用的延迟。Flow Queue 为视频播放、语音通话、在线游戏各维护一条队列，并公平地调度三个队列，避免视频播放的报文占用太多时间，从而让在线游戏、语音通话的报文得到及时处理。由于 Flow Queue 确保了应用程序之间的「公平」（fair），所以又被称做 Fair Queue。CoDel我们已经知道，适当长度的队列，能够充分利用带宽、提高吞吐量。但过长的队列，又会导致过大的延迟。也就是说，队列既不能过长、也不能过短。CoDel 所做的，就是让队列维持在合适长度，最大限度地提升网络性能。CoDel 通过计算报文在队列中的停留时间，来判断队列是否过长。具体做法如下：报文进入队列时，在报文上记录当前时间（也就是报文入队时间）报文出队列时，再次获取当前时间（报文出队时间）计算报文入队时间与出队时间的差值，即可得到报文在队列中停留的时间当报文在队列中的停留时间超过阈值，说明队列过长，需要缩小队列长度当队列过长时，CoDel 会直接丢弃出队列的报文，从而减少队列中报文的数量，达到控制队列长度的效果。（这种对报文的丢弃，并不会对网络带来负面影响，具体请见后续章节）可以看出，CoDel 根据队列中报文的停留时间，来决定是否丢弃报文。而不是「一刀切」地将队列限制在一个固定的、较小的长度。这样既能降低延迟，又保留了队列最大的优点：吸收短暂的大流量报文、提升吞吐量。另外，我们的设备，尤其是手机等移动设备，所处的网络环境是复杂多变的。例如手机靠近或远离无线路由器，就有可能导致网络带宽的变化。CoDel 通过计算报文停留时间的方式控制队列，能够更加灵活地适应网络带宽、网络质量的变化。fq_codelFlow Queue 确保了不同应用程序的流量能够公平传输。CoDel 控制队列长度，实现了整体延迟的降低。fq_codel 就是将 Flow Queue 和 CoDel 结合在一起使用的机制。其中，Flow Queue 先将不同的流划分为不同的队列，再为每一个队列使用 CoDel，控制报文的延迟。两者结合使用，共同提升网络体验。fq_codel 具有免配置、天然适应不同的网络环境、实现简单、消耗资源少等特色，越来越多的无线路由器（以及开源的路由器操作系统），开始内置这一机制。参考 KoolShare 上的这篇帖子，作者经过测试，发现打开 fq_codel 之后，在 BT 下载占满上行带宽时，在线游戏依旧保持较低的延迟：可能是家庭路由里最好的 QoS——在线游戏者的福音 - KoolShare（见 8 楼、10 楼、22 楼、25 楼的测试结果）是否需要打开「智能队列」？前面介绍了 fq_codel 的一系列优点。那么，fq_codel 有没有缺点？是否需要我们打开？我的设备 CPU 性能不强，是否可以打开「智能队列」？根据这篇帖子，ER-X路由器在打开 fq_codel 时，最大支持 200M 左右的带宽。QoS (fq_codel/cake) Hardware Recommendation : HomeNetworkin

2025 年度旅行——4 天 3 晚青海湖小环线主作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥联合作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥05/14 09:38Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。一次完美的旅行，对我而言，不仅仅是去被人走过千遍的地方打卡一圈，而是藏着三次截然不同的体验——出发前对目的地的想象与期待，旅途中与自然的亲密接触，以及归来后翻看照片时，重新发现那些当时未曾留意的细节，细细回味，查阅资料，让我对这次旅行有更深刻的理解。那么就让我们随着这些文字与照片，回到美丽的青海湖。契机：作为刚工作的年轻人，在国内的职场环境下，真正无理由可申请的假期只有 5 天，所以我和身边的小伙伴们都分外珍惜这五天的时光，每年的年度旅行我们都会避开人流，详细规划。由于十一假期要去参加两场婚礼，为了节约交通费，我们决定先开车到 2 号参加婚礼的宝鸡，把小 P（我的小鹏 MONA M03，具体可参阅我之前的文章）停在了宝鸡高铁站，考虑旅行时间较为紧凑，且西北自然景区交通补能条件可能较差，我们选择了到当地租车完成青海湖小环线行程。带着宝鸡擀面皮和肉夹馍，一行人正式开始了旅途，三个半小时后，到达了西宁火车站，初入高海拔地区，并未感受到明显不适，方便的是，租车平台在西宁站地下停车场就有取还车点，我们直接将行李放上车，省去了搬来搬去的劳顿。Day1 西宁-茶卡 全程约 457 km西宁-塔尔寺5A 级旅游景区，门票：70 元，游览时间约 2 小时塔尔寺是藏传佛教格鲁派的著名寺院之一，始建于明代，最初是为纪念藏传佛教格鲁派创始人宗喀巴大师而修建，据传宗喀巴大师出生于此地，后人于其诞生处建塔纪念，塔尔寺因此得名——「塔尔」即「有塔之地」。塔尔寺-八宝如意塔游览中心和寺庙门口都有大量的讲解人员，价格也不贵，但是本人信仰唯物主义，故只将这里当做名胜古迹参观，故没有请讲解，如果对藏传佛教及寺庙背后的故事比较感兴趣，请个讲解或许是更好的游览方式，或者是像我一样，回来翻阅照片时再仔细搜索相关资料。Tips：讲解会介绍各个佛像的「作用」（不知道这样叫是否合适，或者说负责保佑的事情），故可能会被架在那里为了讨好彩头而「必须」捐献资金，介意者可以像我一样自由游览。由于是节前，故参观人员并不很多，与前往朝拜的信徒或僧人可以说是一半一半，没有了熙熙攘攘的人群，也得以让我更加细致地观察这座寺庙和「磕长头」远道而来的人们。前来朝拜的僧人进入寺庙，浓郁的香火味和木质结构因缺乏通风的陈味扑面而来，看着宏大的佛像、墙上摆放整齐的诵经坐垫，的确能感受到不一样的文化氛围，这里安静、鲜活，尤其是看着不知道从哪里启程的「磕长头」的信众，虽因路途的遥远衣衫褴褛，但想必也经过这一路的艰辛而悟出了些许道理，可能出发时也不知道为什么，但走着走着就明白了吧。塔尔寺中的白塔募捐处，与时俱进地，也支持扫码支付塔尔寺有大大小小的寺庙很多间，在我看来有的都别无二致，于是重点转了几个后，就启程回西宁吃饭，整备氧气瓶，应急食物等装备，正式奔赴青海湖。海南藏族自治州共和县 南山观景台 & 青海湖二郎剑景区前者免费，后者为 5A 级旅游景区 门票 72 元，游览时间约 2 小时一点半在西宁市吃完炕锅后，我们开始驱车前往青海湖，约 2 个小时车程，我们到达了本地的主要目的地青海湖的南侧景区 青海湖二郎剑，在进入二郎剑景区前，可以导航到「南山观景台」 从山上远眺青海湖和远处群山的全貌，但是山路较窄且有工程车辆时不时通过，故要格外注意安全。景区本身没有人文景观，但其还是我国早期鱼雷的试验基地，但是因为我们到达时间比较晚，并没有参观博物馆，而是直奔到湖边。青海湖是我国水面面积最大的咸水湖，水面平静辽阔，与远处的群山相辅相成，甚是开阔美丽，栏杆上有硕大干净的鸽子驻足，在几番伦理拍照后，我靠在栏杆旁，看着远处的夕阳洒在地面，耳边传来阵阵鸟鸣与风声，工作与生活的喧嚣远离脑后，只剩下全身心的与自然接触。云层、远山、草原，形成了震撼的三明治拨云见日，青海湖边的草还是绿色的岸边排队等待「薯条」投喂的海鸥码头景观随着夜幕降临，我们离开二郎剑景区，沿着青海湖开车前往茶卡住宿，虽然只有 200 多公里的行程，但是因为沿途工程车辆很多，还有很多地方持续修路，有的路段有连续的 U 型弯，还有双向仅有一条车道通行的地方，还要翻越几个海拔 3800 米的垭口，成了这一路最惊险刺激的行程，一路上小心翼翼，战战兢兢，终于在晚上 10 点到达了茶卡镇，幸运的是镇上的餐馆都还在营业，找了家当地美食，简单吃饭后休息。夜幕降临的倒湖茶公路Day2 茶卡-祁连县 全程约 372km茶卡盐湖4A 景区 带往返小火车套票 167 元「茶卡」在藏语中意为「盐池」，该地以出产「大青盐」而闻名，参观盐湖也是此地的知名旅游项目。茶卡盐湖的景区共有两个，天空之镜是较老的景区，主要以自然风光为主，天空一号主要以商业化项目为主，作为外地游客，我们肯定更希望看到原始的自然景观，于是选择前往天空之境景区。景区很大，如果想要游览每个景点建议留出一整天的时间，游览顺序建议先坐小火车到终点天空之镜，该站处于盐湖的最深处，盐水异常的干净，也最为出片，景区也放置了很多拍照打卡点位，可以尽情拍摄，心痛没有带无人机。在晴天的时候，由于天空和地面都是白色，很容易晃眼，推荐大家一定要戴墨镜前往，通过墨镜观察湖面，盐湖会看起来像翡翠一样更加绿色透明。套票包含鞋套租赁，可以套上下湖走走，踩在大颗的盐里,很像沙滩的感觉，景区维护的十分干净，可以尽情的亲近自然。有趣的是，景区为了怕大家在盐滩大片的白色中无聊，设置了很多拍照用的摆件，例如国家地理的相框等等，还在大屏幕上贴心的轮播拍照指导，期望每个游客都能满意而归。盐湖与远处的山峰盐湖码头废弃的采盐轨道之前采盐的轨道深入盐湖，在墨镜下，真实的观感颜色会比照片深很多。接驳小火车行驶在盐湖中间，干净得不像国内巨幅盐雕从盐湖出来，我们在县城稍作休息，开始启程前往祁连县，有了昨夜惊险的行程后，尽管沿途的国道可能风景更好，但是我们选择了高速较多的线路。这一程的路况还是很好的，就算是国道车辆也较少，道路笔直，视野十分宽阔，草原、远山、白雪、天空层层分明。在行驶过程中，不时的有大群的羊羔和牦牛穿过公路，不必惊慌，一会就有牧民骑着摩托驱赶，让它们加快穿越的脚步。公路与群山草原上大批的牛羊傍晚的草原，远处能看到夕阳洒在雪山上落日下的牦牛剪影在夜色中，我们抵达了祁连县，在县城中吃过一顿特色烧烤后，进入梦乡，准备迎来第三天的行程。Day3 祁连县-岗什卡雪山-西宁 全程约 320km祁连卓尔山景区门票 60 元 上山小巴 20 元 登山电梯 10 元这个景区的确是全程我最惊艳的景区了，近能看到瑞士小镇似的田园风光，中能看到美如画的丹霞地貌，远还能远眺远处的阿咪东索雪山（海拔 4695 米），而且在山顶可以拍跟雪山的合照，大气、干净、巍峨。自驾车辆可以停在景区停车场，然后乘坐接驳中巴约 10 分钟到达检票口。卓尔山的两侧，风景差别迥异，一侧是连绵的雪山，另一侧是仿佛瑞士小镇的丘陵和小镇。阿咪东索雪山卓尔山的另一侧，仿佛欧洲小镇丹霞地貌与草地交融最让我惊讶的是这个景区竟然还修建了方便爬山的观光电梯，10 元一位，大概有三段，果然是人性化啊。景区电梯中午约 12 点，我们离开景区，在祁连县的一家川菜馆简单用餐后，我们驱车开始前往本次旅途的最高海拔岗什卡雪山。在县城开车的过程中，等红灯时，我看到一位初中年龄的藏族小伙盯着我们开的岚图梦想家，眼中闪闪发光，不禁让我想起小时候的自己，看到喜欢或没见过的车也会驻足观看，充满憧憬。愿他长大后也能开上自己喜欢的车，能够做自己喜欢的事情。岗什卡雪峰景区免费 上山摆渡车 25 元岗什卡雪山坐落于青海海北州门源县境内，是祁连山脉东段的最高峰，被誉为「雪山之尊」，因为设施和难度都尚可，也被雪山爱好者们称为「人生第一座雪山」。主峰海拔 5254.5 米。游览模式同样为将车辆停在景区停车场，然后乘坐接驳车到雪山脚下，正式游览区域的海拔在 3800 米左右，建议随身携带一个氧气瓶，下车之后，可以看到湛蓝的连心湖，水来自于冰川融水，干净，清冽，与子洲植被的荒凉的深褐色形成鲜明的对比，十分美丽。连心湖到达时天气并不是很好，岗什卡雪峰被薄雾笼罩天然泉水，因还有行程，害怕拉肚子，没能如愿品尝一下离开景区后，雪山逐渐露出真面目，攀登还是留给下次吧沿着冰川的遗迹向上走，就能看到岗什卡雪峰徒步攀登的起点，景区也提供马匹，可以跟着向导继续向前，到七彩瀑布，因时间和个人体能的考量，我们浅尝辄止，在河道附近游荡些许时间后，便下山返程。临近傍晚，启程返回西宁市，一顿丰盛的海底捞过后，快乐的青海湖小环线结束。Day4 西宁市-青海藏文化博物馆-特产采购最后一天，伴随着蒙蒙细雨，我们打车前往青海藏文化博物馆，博物馆不大，纵使是十一期间人也不是很多，但内容着实丰富，满墙的唐卡和在电影里看到的藏传佛教僧人服饰，还有精致的贵族金器。在博物馆的文创区，我购入了一个小的转经筒，在回程的路上一直把玩，虽不信奉佛教，但也觉得甚是有趣。唐卡四王金杯嵌绿松石立凤金头饰博物馆文创-转经筒，38 元虽不算划算，但做工十分精致离开之前，为了给家人和朋友带些纪念品，我们选了一家超市类型的纪念品商店，购买了牛肉干，藏红花，枸杞，青稞酒等特产，虽说也不一定划算，但胜在明码标价，将行李箱装满后，顺利返程。后记作为 2025 年的年度旅行，本次 4 天的青海湖小环线我甚是满意。人少不堵车，行程不特种兵，在当地感受美食和壮美的自然风光，一路上与好友们分担开车，聊对景色的感受，互相拍照，点评食物，虽偶有几个工作电话烦扰，但总体悠然自在，放空身心，一切都是那么的安静，与自然深度的接触，与北京的喧嚣形成了鲜明的对比。对于出行方式，自驾无疑是最有的选择，这边深入自然，道路虽崎岖，但也属于铺装路面，且西宁市神州、一嗨等在高铁站都有网点，取送车方便，且自驾可以自由控制行程，根据游览时长方便自己调整。对于住宿，我们对酒店的干净程度略有要求，于是全程都选择了以华住会为目标的连锁快捷酒店，以下是所有天的酒店名称和十一节前的价格，个人觉得是十分舒适的。Day0 桔子西宁火车站省人民医院酒店 双床房 331元Day1 星程茶卡盐湖酒店 覆氧双床房 314 元Day2 宜必思尚品海北祁连卓尔山酒店 421 元Day3 桔子西宁火车站省人民医院酒店 双床房 412 元关于本次行程的花销，纵使算上从北京出发的路费，人均也仅 3k 左右，可以说是十分划算了，主要的开销就是租车和酒店。支出一览表关于卫生间，景区和路途服务器的卫生间都比较完善，因同行有女生，我们还准备了移动卫生间，但从我们的行程安排走下来，并没有用到，个人觉得要是按我们的行程游览也无需担心。下一站，出发！2025 年的年度旅行结束，2026 年也可以开始规划起来了，新疆还是海南，希望能远离喧嚣，享受美丽壮阔的祖国风光。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀2016扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@北鸮© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#Matrix首页推荐#旅行20等 20 人为本文章充电扫码分享举报本文章别惹小炸毛一个学酥关注

打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试主作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.联合作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.前天 15:00ARG 即侵入现实游戏（Alternate Reality Game），是一种基于现实世界，融合各种多媒体（如网站、电话、社交媒体等）技术的互动式解谜游戏。由于其与现实的高度结合所带来的沉浸感，一直以来被许多玩家所追捧。如果你在 2016 年玩过《守望先锋》，相信你对当时新英雄「黑影」的登场一定记忆犹新。从 2016 年夏季开始，一直到 11 月的暴雪嘉年华，暴雪通过在宣传视频、官网、社交媒体发布了一些隐藏着密码的内容，由全球玩家共同解谜，不但揭示了游戏世界观下的一些故事，更是顺理成章地将身为黑客的「黑影」带到大家的视野里。经过几个月的发酵，在暴雪嘉年华上大屏幕被「黑影」控制，随后开始播放英雄动画的时候，更是满足了玩家们期待已久的心。「黑影」的宣发是一场相当成功的 ARG，时至今日仍有玩家对其念念不忘。我本人也是 ARG 的爱好者，我非常享受这种打破虚拟与现实边界的游戏乐趣。前段时间，一款名为《大学生登山失踪事件》的网页解谜游戏突然火了起来。比起此前一些高度依赖技术知识的网页解谜游戏，《大学生登山失踪事件》在不需要任何 IT 知识背景的情况下，依然有着抽丝剥茧逐渐接近真相的推理爽快感。自此之后，小红书上越来越多的人开始制作这一类网页解谜游戏，并为它们赋予了 ARG 的名号。这些游戏流程往往不长，差不多 1-3 小时就能通关。因为是网页解谜游戏，只需要一个浏览器就能游玩，很适合上班时间摸鱼，游玩门槛非常低。这篇文章将会推荐一些这段时间我玩过的质量不错的游戏，虽然一些制作精妙的游戏确实带有 ARG 的影子，但其本质上仍是基于网页的解谜游戏。为了方便大家后续搜索，本文会将这些游戏统称为 ARG。大学生登山失踪事件🔗https://missing.shiroki-y.top/introduction.html在游戏中，玩家需要通过从弟弟衣物中发现的「福安菜家常馆」的线索为开端，逐渐深入一场黑色阴谋找到弟弟失踪的真相。游戏整体难度不高，但在游玩过程中最好能用纸笔或软件记录进度，这能够帮你清晰地理出游戏脉络。游戏的核心玩法其实非常简单，只需要阅读页面，找到页面中可疑的关键词，在页面搜索栏里进行搜索即可推进剧情，这其中也会有一些用户名、密码、口令、字谜等需要玩家进行推理和猜测。网页页脚处会标记当前页面进度，既可以方便定位，也能够明确线索的先后顺序。剧情方面，《大学生登山失踪事件》并不会在一上来就给玩家塞一些晦涩难懂的概念，而是从大家熟悉的普通网页开始。当玩家一步步抽丝剥茧逐渐发现事件背后的黑暗面时，令人不由得感叹于游戏的设计与巧思。同时，游戏内也藏有一些可供探寻的 ex 页面作为彩蛋，在解谜之外找到各种彩蛋也不失为一种乐趣。诡异广告调查事件🔗https://mp.weixin.qq.com/s/yeNYMZVRp6N-MESFJGk9rg这是一款与《大学生登山失踪事件》来自同一作者@銀河電燈_的游戏，世界观也与前作有关联之处，很是有趣。玩家将扮演一位被「优化」的人，因为在看剧时发现了某个已停播的节目广告，从而开始调查这个诡异的广告，并发现藏在节目与自己背后的神秘真相。与前作相似，这款游戏也将基于阅读与搜索推进剧情。在此基础上，它还加入了一些 ARG 元素，比如在街景地图中，找到现实里存在的文字招牌作为线索。并且在网页设计上也比前作更加成熟，贴合剧情上的时间设定，在网页上加入了许多彩蛋。剧情方面，本作有一定的惊悚元素。特别是在游戏过程中展示的一些内容可能会引起不适感，如果你对这类内容比较敏感，请务必提前做好心理准备。当然，游戏并没有 Jump Scare，更多是通过剧情与视觉设计所营造出一种的诡异氛围。三川米粉店🔗https://www.berrysignalscmfd.top/《三川米粉店》也是一款从餐饮店主页出发，探寻故事真相的游戏。玩家需要调查三川米粉店突然停业的原因，在调查过程中，玩家将逐步揭开这家米粉店背后的故事。本作的页面设计相对没有那么精致，但也融入了一些 ARG 的元素，部分谜题的答案并不会直接给出，而是需要玩家通过搜索引擎或其他工具进行查询才能获得答案。从剧情上来说，《三川米粉店》的设定会更加接地气一点，并且结局也会相对更温馨一点。比较适合对于惊悚元素不太感冒的玩家作为第一部作品来尝试游玩。永安温泉度假酒店🔗https://mp.weixin.qq.com/s/kKsnzFsLFwwm2nwv8DseAA与前面侧重解谜的作品不同，《永安温泉度假酒店》的解谜难度其实不高，反而更像是一种「规则怪谈」式的探索路径。玩家需要不断探索找到更高权限的员工账号密码，从而揭开温泉度假酒店背后的秘密。游戏在开始前需要玩家填写「报名信息」，在后续游戏中能够增强代入感。从游戏性的角度来看，这款作品的「解谜」爽快感或许没有那么强。但如果你对「规则怪谈」或者一些高概念的设定感兴趣，那么这款游戏依然是一个值得尝试的作品。404 工位消失事件🔗https://xsn1204.github.io/qmgame《404 工位消失事件》是一个特色鲜明的游戏，玩家将扮演一个实习生，接手其他人留下来的工作电脑，并从中发现一些未被处理的秘密。游戏在网页上复刻了一个形似 macOS 的「系统」，很大程度上增强了代入感。因为设定是在电脑系统，游戏中需要玩家在钉钉、微信等各种软件中挖掘线索，反而更有一种作为侦探在不断探索的感觉。但受限于网站形式和游戏体量，游戏中能够互动的地方通常都是为剧情线索服务的，与线索无关的地方大多都是不能点击互动的。而且游戏不能通过点击窗口左上角的关闭按钮回到上一层，而是需要通过浏览器的返回来实现。但这些小细节对于《404 工位消失事件》来说都算瑕不掩瑜，整体游玩体验还是值得肯定的。作者甚至还出了 2.0 版本，新的剧本进行了优化提升，如果对这一类型感兴趣的话也可以尝试游玩。找到我🔗https://www.kaixinmoyu.com/feijo/findfeijo/在游戏中，玩家需要扮演一位私家侦探，调查游戏作者失踪的原因。游戏在交互方面做得非常有意思，游戏不仅做了浏览器页面对线索进行搜集检索，还做了小灵通和笔记本，前者可以通过拨打特定的电话获取线索，后者则需要玩家阅读日记找到隐藏的内容。《找到我》本身是一个更重交互而非推理的游戏，游戏中融入了一些 ARG 元素，在最开始就需要到某篇公众号文章中找到一个密码才能继续推进。但游戏的剧本逻辑会稍微有些不够连贯，以至于中间可能会有一些容易卡关的地方，需要仔细翻阅此前出现的线索才能继续往下推进。另外，游戏中后期会有一道需要在现实中解算的谜题，这道题本身与游戏本身的节奏设计会稍显割裂。但整体来说，《找到我》依然是值得一玩的游戏，尤其是本身的交互设计和 ARG 元素的融合做得非常不错。在游戏作者的个人主页，也能找到作者开发的更多网页解谜游戏，如果你对这款游戏的风格感兴趣，也可以尝试其他作品。小结优秀的网页解谜游戏并不止文章提到的这些，光是在小红书上一搜，就能找到许多分享帖。希望本文能起到抛砖引玉的作用，欢迎大家在评论区分享自己玩过的，令你印象深刻的网页解谜游戏。> 简单、好用、专注的写作软件，少数派为你呈现 🚀> 守护孩子的天马行空，从零开始美术教育启蒙🎨579扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Lotta© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#解谜游戏#游戏#生活方式57等 57 人为本文章充电扫码分享举报本文章TpBe nothing, and everything.关注

派评 | 近期值得关注的 App主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注Vanilla少数派作者Vanilla关注Vanilla少数派作者关注化学心情下2少数派作者化学心情下2关注化学心情下2少数派作者关注大大大K少数派作者大大大K关注大大大K少数派作者关注Peggy_少数派作者Peggy_关注Peggy_少数派作者关注ElijahLee少数派作者ElijahLee关注ElijahLee少数派作者关注Snow少数派作者Snow关注Snow少数派作者少数派编辑部 等 7 位作者前天 18:05欢迎收看本期《派评》。你可以通过文章目录快速跳转到你感兴趣的内容。如果发现了其它感兴趣的 App 或者关注的话题，也欢迎在评论区和我们讨论。AIDA：一站式日历和待办规划工具平台：iOS、iPadOS关键词：日程管理、待办清单@ElijahLee：Aida 是一款针对 iPhone 和 iPad 开发的日历和待办清单应用，来自任务管理工具 Sorted 的前创始人之一 Harry，它集成了待办清单、日历和时间规划工具，在一处显示、规划、操作。应用主界面就是每日时间线，顶部是当周 7 天的简要信息，应用中部大篇幅排列当天的时间块，你可以直接看到上午十点看电影、中午写派评、傍晚健身等日程，以及晚上记得给植物浇水的待办提醒，而不用在多个不同的 App 之间切换。应用的快速调整功能，支持在时间线中将事件按住并移动顺序，快速重排一天的事件，对应的时间点会自动修改，非常方便。使用右下角添加按钮，可以自由地为时间线加入行动或事件。行动即待办事项，支持输入文本并设置提醒时间。 Aida 创造性地设计了拖动设置时间的功能，将底部的柱状图左右滑动，就可以调整行动的提醒时间，每次拨动以 15 分钟为间隔，用它设置时间非常方便。当然我们也可以点击时间来精细设置时分。 Aida 事件就是日历中的事件，必须包含开始时间和结束时间，我们还可以选择日历类别、设置全天事件等。Aida 引入了 GTD 管理的项目功能，提供收件箱、任何时间、将来某日等默认分类。自定义项目可以将同样类别的活动，包括行动和事件进行集中收纳整理。Aida 全面集成了 iOS 日历，在应用内不仅可以查看 iOS 日历事件，对事件的修改调整还会实时同步。不过 Aida 目前没有接入 iOS 的提醒事项。 批量操作功能可以一次选择多个行动进行完成、改期、移动或复制，也能快速整体调整时间块。你可以在App Store免费下载 Aida，免费版功能受到很大的限制，仅支持 1 个日历的集成显示、无法调整行动的时间、添加行动和事件的数量有限。付费可以解锁高级功能，包括无限项目和时间块、多个日历集成显示等功能。费用为 58 元 / 月、298 元 / 年，698 元买断，应用提供 1 周免费试用。我的生活配不上这个定价我以为待办和清单类 app 最重要的是视觉和交互吧 这令人一眼没欲望的 AI Generated UI 怎么还敢定价这么高毫无特色 + Sorted 血脉 + 令人疑惑的定价现在任务管理之用自带的提醒事项和滴答清单免费版定价没问题，是我的问题。全部讨论 8发表评论Cogito：连接人与 AI Agents 的 Markdown 编辑器平台：macOS关键词：Markdown、AI@Vanilla：对于 Markdown 编辑器这一类应用，我想大家已经见得太多了。随着 AI Agents 的兴起，Markdown 这个文件格式可谓是「老树开花」，又重新获得了很多人的青睐。Cogito 是一款面向人和 AI Agents 共用的 Markdown 编辑器，它不像 Obsidian 一样偏知识库、不像 Typora 一样偏所见即所得、不像 iA Writer 一样偏写作，也不像 VS Code 一样偏开发者，它最大的特点是「文件优先、Agents 友好」。下面，就让我们一起来看看这个新时代的 Markdown 编辑器到底有何不同。Cogito 的文件库完全基于本地磁盘，没有专属格式，没有加密拦截，没有同步机制，所有的文件和文件夹组织形式都和 Finder 中的一模一样，完全映射，没有任何上手门槛。以我的使用习惯为例，我直接把 Codex 的工作文件夹设置为了 Cogito 的 Location，它会自动抓取 Codex 工作区的所有文件夹和 Markdown 文件，按照日期排序，清晰明了。选择任意一个 Markdown 文件后，Cogito 默认展示它的编辑状态，我们可以随意地按照自己的意愿进行编辑。点击右上角的第一个按钮，则可以在编辑状态和预览状态之间进行切换，对应的键盘快捷键是 Command + R。我经常会借助 Codex 帮我生成一些行业或者公司的研报，最终以 Markdown 格式来进行输出，通过 Cogito 我就可以非常方便地进行编辑和阅读，不需要反复地在 Codex、Finder 和 Markdown 编辑器之间进行来回切换。点击右上角的第二个 Outline 按钮，我们可以快速查看当前文档的大纲，点击标题可以直接跳转到相应的章节。Cogito 除了对 AI Agents 友好，还可以在应用中直接调用本地的 AI Agents 对应用内显示的文档进行总结、分析、编辑等操作。总的来说，Cogito 对于像我这样频繁使用 AI Agents 和 Markdown 文档的人来说是一个非常不错的 Plan B，和主力写作的 iA Writer、用作知识库的 Obsidian 等进行区隔，打开即用，用完即走，非常高效。Cogito 可以在官网免费下载使用，目前处于公开测试阶段。好吧，不能所见即所得。agents友好其实有点尴尬，一个阅读器+一个terminal就可以完全的代替掉这个软件目前的全部功能？甚至codex的app自身也有文件阅读功能了。发表评论AndDrive：无缝连接你的 Android 设备和 Mac 电脑平台：macOS关键词：文件管理、手机管理@化学心情下2：虽然现在有不少工具能让 Android 手机和电脑无线互传文件，但「传文件」和「管理文件」是两回事。比如想批量删除 Android 内部存储 Pictures 文件夹下的截图，用传输工具就很别扭，这种场景更适合在电脑端的文件管理器里完成，可以借助筛选和预览功能高效处理。Windows 上这件事不需要任何第三方软件：数据线一插，选择「文件传输」，Android 就被模拟成 MTP 设备，像移动磁盘一样直接管理即可。但 Mac 上就麻烦了：MTP 是 Windows 的专有协议，macOS 想访问 Android 内部存储，要么走 USB 调试，要么依赖第三方 MTP 实现。OpenMTP 算是其中比较知名的一款，但项目两年多没更新、稳定性一般，而且用的是自己写的文件管理器，而非访达，体验称不上顺手。那么有没有能直接用访达管理 Android 文件的工具？接下来要介绍的 AndDrive 就是，它由个人开发者开发，没有走 MTP 路线，而是基于 Android 官方的 ADB 调试通道实现文件管理。安装后，AndDrive 常驻菜单栏，支持 USB 和 Wi-Fi 两种连接方式。使用前需要在 Android 上开启 ADB 调试并授权当前 Mac，再按提示在手机上安装配套的 AndDrive Helper，之后访达里就能看到 Android 的目录了。AndDrive 的挂载方式可以选 SD 卡目录、根目录或不挂载，看到的内容会不同：SD 卡目录只显示内部存储，根目录则多一个系统根目录（没 root 的话很多文件依然动不了）。日常使用可以直接在访达里预览文件、按时间筛选，也能把电脑上的文件拖进 Android 对应文件夹，比在手机自带的文件管理器里翻找直观得多。因为走的是 ADB，AndDrive 顺带还实现了屏幕镜像功能。点击 AndDrive 的「屏幕镜像」，当前 Android 设备的桌面就能投屏到 Mac 上，体验和 macOS 的 iPhone 镜像类似，可以直接操作应用、用中文搜索，还支持剪贴板同步、录屏、截屏等。比起手敲 Scrcpy 命令，门槛会低不少。总的来说，AndDrive 虽然没有用更成熟的 MTP 协议，但用 ADB 这条 Google 官方推荐的方法，让访达管理 Android 文件这件事变得简单可行，还补上了类似 iPhone 镜像的屏幕控制功能，值得有需求的你一试。你可以在App Store免费下载 AndDrive。有没有苹果手机和windows电脑互传的它不是改名了吗？🤔发表评论Fauxx：与其被动防守，不如主动出击的隐私保护工具平台：Android关键词：隐私保护@大大大K：你的每一次搜索、每一个网络请求、每一段定位轨迹，都在被数据平台和广告商默默收集着。它们会通过这些碎片信息拼凑出你的身份、喜好、习惯甚至是经常交往的人，然后在暗地里交易你的信息、精准推送广告。这也是为什么你或者你的朋友随手搜了点什么，10 分钟后所有 App 和网页就开始围着这个话题给你推广告。因此市面上衍生出不少隐私保护工具，它们的思路大多都是「堵」——屏蔽数据上传、拦截 Cookie、混淆 IP 等等。但 Fauxx 这个 App 啧另辟蹊径：既然广告商一定要收集数据，那就把真实的你混淆在一堆「垃圾」当中。首次打开 Fauxx 会有一个引导流程，你可以填写你的年龄段、职业、兴趣之类的粗略信息。这一步主要是为了让 App 能够生成与你真实信息相反的混淆信息，当然，这一步你也可以选择跳过。Fauxx 并不强制你输入信息。完成后进入主界面打开开关就能启动混淆信息的保护，在这里你可以看到这段时间 Fauxx 执行了多少次「混淆」动作、当前的虚假人设是谁、混淆了哪些类型的信息。至于具体可以混淆的内容，可以在 Targeting 和 Modules 标签中设置混淆等级和类型。Targeting 里有三个可选的层级：第一层就是生成与你真实信息相反的混淆噪声，如果你一开始没有填写，Fauxx 则会随机生成；第二层是对抗广告商的用户画像，不过我实际测试下来会读取失败，但也不影响使用；第三层是「人设轮换」功能，Fauxx 通过生成虚假的身份信息，让广告商无从追踪。Modules 标签里则是七个具体的混淆模块，每个都可以单独开关。设置里还可以调节混淆频率、指定时间段等。那这些「假数据」是怎么骗过广告商的？通常来说，被安插在手机里的追踪器会根据设备 ID、IP 地址、DNS 等信息追踪这台设备上的数据访问，进而推测用户画像。而 Fauxx 会通过模拟搜索、链接访问等「类人」操作（比如搜索时的停顿、假装点击广告），规避机器人检测并成功让追踪器追踪到假的信息访问。Fauxx 可以说是转变了隐私保护的方向，从被动防守变为主动进攻。如果你也想换个思路给广告商添添堵，可以在GitHub免费下载 Fauxx。这软件在中国没用吧1324Roll：胶片「日日鲜」平台：iOS关键词：手机摄影、胶片模拟@Snow：当下，你用手机拍照真的可以发散出无限可能，不同的镜头、不同的滤镜、不同的手动参数、不同的画面调节项，甚至是不同的外设，只要你愿意钻研一番，就能拿捏各式各样的风格。不过，更多的选择往往意味着更多的「选择困难症」，如果你对于丰富的选择产生了一些焦虑，或者天性就不爱「折腾」，不妨看看 Roll 这款应用，它或许可以带你发现简单拍照的乐趣。打开 Roll，完成简单的授权流程后，你就会收获今日份的新胶卷。应用会简单阐述这份胶卷在感光度、对比度、饱和度等各方面的特质，有一些呈现阳光下温暖的黄色，有一些闪耀着玫瑰般的光泽，有一些甚至想带你进行「第三类接触」。和一般的应用内随机胶片不一样，Roll 所有用户的「掷骰」结果完全一致，无论你身在世界何处，在同一时刻都会使用同一款胶片进行拍摄，剩余可用时间会在胶片介绍页注明。时间有限，点击「Load film」启动拍摄。Ro

周二19派早报：红魔发布红魔 11S Pro 系列手机等今年的新系统会有什么变化呢？早报速览红魔发布红魔 11S Pro 系列手机5腾讯音乐已完成对喜马拉雅的收购9蓝牙核心规范 6.3 正式发布0腾讯自研 AI 设计智能体平台 Ardot 现已开启公测7微软将允许用户修改实体 Copilot 键功能4中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告0马斯克诉奥特曼案被驳回1不妨一看的简讯4红魔发布红魔 11S Pro 系列手机5 月 18 日，红魔发布红魔 11S Pro 与 11S Pro+ 手机等新品。红魔 11S Pro 系列搭载第五代骁龙 8 至尊领先版与红芯 R4 电竞芯片，机身采用纯平设计，11S Pro+ 还有透明水冷可视化设计，屏幕为 6.85 英寸 144Hz 无挖孔无刘海设计，采用屏下前置摄像头与 3D 超声波指纹解锁，搭载新思触控芯片，支持 3000Hz 瞬时触控采样率。系统内置 PC 模拟器，搭载 3.5mm 耳机孔与红外功能。11S Pro 电池容量 8000 mAh，支持 80W 快充；11S Pro+ 电池容量 7500 mAh，支持 120W 快充，两机型均支持旁路充电。红魔 11S Pro 有暗夜骑士与银翼战神两配色，12GB+256GB 售价 5499 元，国补后 4999 元。红魔 11S Pro+ 有氘锋透明银翼与氘锋透明暗夜两配色，12GB+256GB 售价 6199 元，16GB+512GB 6999 元，16GB+1TB 7999 元。另有红魔碳纤维游戏鼠标、红魔散热器 8Pro、氘锋能量卡 Pro 磁吸充电宝等配件发布。来源红魔还是有点追求的红魔是为数不多的纯平机身了背后是MagSafe 结构？屏下摄像头技术已经成熟了吗，怎么感觉其他家很少用呢，还有什么缺点吗3腾讯音乐已完成对喜马拉雅的收购腾讯音乐于 5 月 18 日在港交所发布公告，宣布根据并购的条款，收购喜马拉雅的事项已经完成。据此喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为 12.6 亿美元的现金及最多总数 1.75 亿股本公司 A 类普通股（包括本公司于交割时新发行的 A 类普通股，以及本公司授出的以股权为基础的激励相关的 A 类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。来源完了 喜马拉雅 也要有低级充值会员 中级充值会员 高级充值会员了买京东会员送的喜马拉雅会员还没用过呢全部讨论 9发表评论蓝牙核心规范 6.3 正式发布蓝牙技术联盟上周末发布了蓝牙核心规范 6.3，作为半年一次发布计划的一部分，此次更新引入了多项新功能，可进一步提升测距精度、扩展接口容量并提高射频效率。本次更新新增蓝牙信道探测内联 PCT 传输、蓝牙信道探测特定物理层 RTT 精度、蓝牙位资源扩展以及蓝牙 ACP 和 C/I 限值放宽。蓝牙技术联盟表示，联盟成员在描述产品支持的蓝牙功能时，应避免直接引用产品所依据的蓝牙核心规范版本号（例如蓝牙核心规范 6.3），而应在产品包装、文档和推广材料中清晰、准确地描述其产品支持的具体蓝牙功能，特别是与用户需求最相关的部分。来源42腾讯自研 AI 设计智能体平台 Ardot 现已开启公测腾讯云于 5 月 18 日发布公告，正式宣布 AI 设计智能体平台 Ardot 开启公测，官方称其为「一个 AI 驱动的产设研协作平台」。Ardot 支持一句话生成可编辑设计稿，AI 可快速生成 App 页面、官网、海报、插画、PPT 等；支持调用团队自己的业务组件库，一键生成符合规范的可编辑稿；支持直接导入 Figma 文件，完整保留原有布局、样式和组件。Ardot 对代码友好，支持设计一键转代码。设计稿，连同变量、组件、布局数据等设计细节数据，直接拉进 CodeBuddy，一键代码还原；也兼容 Workbuddy、Cursor、Claude Code 等 MCP IDE。Ardot 支持多人在线实时评论、标注反馈和版本对比，配备智能权限中心、全方位行为追溯和无缝交接机制，确保协作规范、有序。此外，Ardot 微信小程序也即将上线。来源太难用了能运营超过两年吗？与此同时猿辅导家的妙多 / Motiff 即将停服像素级啊确实太像了，感觉就把配色 从紫色改成了绿色🐶[图片]全部讨论 734微软将允许用户修改实体 Copilot 键功能5 月 18 日，Windows Central 发现微软上线一份新支持文档，表明自 2024 年起笔记本厂商在 Windows 11 设备上预装的 Copilot 实体按键影响了用户的工作流，他们将在今年晚些时候推送一版 Windows 11 更新，允许用户将 Copilot 键重定义为被替代的菜单键或右 Ctrl 键，用户将在设置 > 蓝牙和其他设备 > 键盘中找到相关选项。来源什么时候把win键强制绑定去了，允许用户自定义。终于顿悟了？国内都用不了，按它干什么？微软终于发现了141中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告5 月 18 日，中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告。2026 年 4 月，中国游戏市场实际销售收入 303.69 亿元，环比下降 3.51%，同比增长 11.04%。其中，中国移动游戏市场实际销售收入 224.30 亿元，环比增长 0.71%，同比增长 9.83%；中国客户端游戏市场实际销售收入 68.66 亿元，环比下降 12.77%，同比增长 18.40%。中国自主研发游戏海外市场实际销售收入 20.32 亿美元，环比下降 5.02%，同比增长 30.78%。来源1马斯克诉奥特曼案被驳回加州奥克兰联邦法院的陪审团于 5 月 18 日作出一致裁决，驳回了埃隆·马斯克对 OpenAI 及其 CEO 萨姆·奥特曼等人的诉讼。经过为期三周的庭审，由 9 人组成的陪审团仅用不到两小时便达成裁决，认定马斯克指控 OpenAI 违背成立初期慈善承诺的起诉已超过诉讼时效。2024 年，身为 OpenAI 联合创始人之一的马斯克指控奥特曼、总裁格雷格·布罗克曼以及主要投资者微软公司背叛了该机构「为人类福祉开发非营利性 AI」的初衷，通过转为商业化运作并与微软结盟来谋取巨额私利。马斯克此前曾为 OpenAI 提供早期资金，并在此案中索赔最高达 1500 亿美元。陪审团认定，马斯克早在 2021 年就已经知晓 OpenAI 转向营利性架构的事实，因此其 2024 年的起诉已经超出了三年的法定诉讼时效。辩方展示的证据表明，马斯克自己当年在未能获得 OpenAI 控制权之前，也曾提议过转为营利性结构。此次判决为 OpenAI 今年晚些时候的 IPO 扫清重大法律障碍。马斯克本人则在 X 上发文，批评该判决仅仅基于「时间程序问题」，而并未对案件的实质正义进行裁决。来源老马：不听我控制的就是不正义17不妨一看的简讯据外媒 Wccftech 的报道，根据一份去年泄漏的运输清单中的信息显示，Googlebook 笔记本或将搭载英特尔 Panther Lake 12Xe3 核心处理器。运单的信息显示，该产品使用酷睿 Ultra Series 3 系列处理器（即 Panther Lake），并搭载 12Xe3 图形单元。并且在 Google 与英特尔的相关文件中，有一个相同的代号 Q7AP 出现。而该编号对应 Panther Lake 12Xe 主板，对应的 Chromebook 产品代号则为 Felino。该笔记本采用 16 英寸设计，配备 32GB LPDDR5X 内存、1TB 存储空间。根据英特尔公布的产品阵容，12Xe 应该是 Panther Lake 的高端型号，至高可选 16 核心 CPU，功耗范围 25W-80W。如果此次曝光准确那么 Googlebook 很可能采用酷睿 Ultra Series 3 系列处理器。来源据微博用户金猪升级包，英特尔在 5 月 18 日的第三代英特尔酷睿处理器新品分享会上，介绍了英特尔 Project Firefly，将针对价格敏感的 Wildcat Lake 平台，导入中国本土化成熟的手机产业链，统一标准化接口实现规模化，完成降本增效。现场展示了 Project Firefly 的 A+ 参考设计与第一款 Project Firefly 参考设计下落地的量产产品来酷 Air 14，均为超薄机身与 Clean D 设计。来源Linus Torvalds 在 LKML 的新帖文中建议大家不要用 AI 工具提交重复的 bug 报告，大量重复、过时的报告让邮件列表无法正常收集 bug，想用 AI 工具创造点价值就去好好写补丁。来源Apple 定于北京时间 6 月 9 日至 13 日举行全球开发者大会 WWDC26，6 月 9 日凌晨 1 点开幕时将举行主题演讲，介绍 Apple 平台更新，包括人工智能的进展、新软件和开发者工具；6 月 10 日凌晨 4 点举办 Platforms State of the Union，介绍新功能、API 和技术如何进一步赋能开发者。来源如此直白的话感觉是Linus会说出来的😅Linus都蚌埠住了还人工智能进展，国行的 AI 有没有？意思是Googlebook运行的是x86_64的安卓？发表评论少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购你可能错过的文章四天三夜，用小长假的时间体验青海湖小环线打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试派评 | 近期值得关注的 App> 下载少数派 2.0 客户端、关注少数派公众号，解锁全新阅读体验 📰> 实用、好用的正版软件，少数派为你呈现 🚀71扫码分享目录 0讨论我来说一句发布发表评论发布

YoooClaw C·ONE 体验：为 AI Agent 而生的硬件，怎样构筑了数字化的「我」？主作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛联合作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛昨天 11:04AI 虽然现在具备了干活的能力，但对于很多人来说，它依旧只是个 Chatbot。因为，大多数 AI 依然停留在屏幕里（当然，具身智能另说）：你需要打开手机或电脑，进入某个 App，手动复制消息、描述背景、然后不停修正抽卡，最终把输出的结果复制到你需要的地方去 —— 对很多人来说，这个过程本身就已经足够打断手里的工作。App 的逻辑其实更适合屏幕类设备的交互，这是我们从 Web 2 时代过来的逻辑惯性。但如果我们换个思路，要是一个设备不是为了适应人类和 App 的交互，而是一开始就为 AI Agent 去构建，那它应该是怎样的？YoooClaw 给出的答案是 —— 一张智能卡片。很高兴，这次与少数派和 YoooClaw 一拍即合，提前体验了这款名叫 C·ONE 的智能卡片。 在体验的过程中我意识到，或许那个「数字化分身」的时代，已经慢慢开始变得具象了。AI 时代需要一个「灯条」拿到 YoooClaw C·ONE 的时候，其实我一直在想，手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件呢？但这个问题还不急着说，反倒是上手的第一刻，它的工业设计给了我不小的惊喜。它采用了全金属材质，一体感极强，拿在手里有一种精致感，有点像金属版的 MagSafe 外接电池。而且没想到的是，它的楔形设计与我 iPhone Air 的契合度高得惊人。磁吸上去后，它恰到好处地填充了 iPhone Air 极其轻薄的机身与我掌心之间的空白，给我这个 iPhone Air 裸奔党增加了不少安全感……C·ONE 的背面元素也不多：一个按钮、一个条形指示灯和 YoooClaw 的 logo。长按按钮，灯条变成绿色的波浪，C·ONE 就会进入 AI 对话模式，很像是我拿起 Apple TV 遥控器按下侧键与 Siri 对话。而双击之后，灯条变成绿色的闪烁，就进入了录音模式。这个灯条让我想起了前几天发布的 Googlebook，不同时空，横跨太平洋的它俩居然想到一块去了。该说不说，灯条在 AI 时代居然成了一种「不打扰却又十足惹眼」的交互 —— 它不告诉我具体的内容，而是只告诉我「状态」，简直是我这种 ADHD 人的福音 ——一个真正好用的通知管理器作为一个长期与 ADHD「斗争」的人，一条通知突然让屏幕亮起，我的心思可能就飞出了十万八千里之外。而且有时候好不容易进入了心流状态，突然来一个通知，看吧，我的工作流就断了，不看吧，真是心痒痒。iOS 原生的通知系统虽然有做一些约束，甚至还引入了专注模式，但它有一个很大的问题：它的管理逻辑是基于 App 的。这就有个十分明显的问题：工作群里，我负责的事情结束了之后，与我有关的通知可能就不多了。但我还是不敢关通知，毕竟有些事情还是需要我来处理。而在 iOS 系统中，这些通知的地位是完全对等的（到底是谁那么喜欢在微信聊工作啊！）。它们密密麻麻地挤在一起，要么让我因为厌烦而全部忽略导致错过，要么每次来消息，我就还是得就打断当前的事情看一眼。尽管 Apple 智能后来推出的「通知总结」有所改善，但对于国行设备来说，吃不到嘴里的饼画得再大也没用。即便在已经用上了的地区，用户反馈也不太好。而且，它依然解决不了「拿起手机」这个动作本身带来的诱惑：原本只是想看一眼通知，结果却顺手刷了半小时马斯克的表情包。YoooClaw C·ONE 提供的方案非常符合直觉：它可以读取我的通知，并根据重要程度，给背后的条形灯带设定不同的颜色和效果。最棒的一点是，这种设定不需要像在 iOS 里那样一个一个 App 去操作和设定，只需要在 YoooClaw 里通过自然语言告诉它：「我在微信群工作群 XXX 里的昵称叫宛潼，如果有 @ 我的消息，你就用红色呼吸灯提醒我。」于是，我的工作流变成了这样：当我开始写稿时，在 iPhone 上进入创作专注模式，所有通知调整为静音，并且同步到 Mac。然后，我将 iPhone 倒扣在桌面上，让 C·ONE 成为我唯一的通知接收器。如果灯条没有闪烁红色，那我就完全不用去理它，只有当红色的光亮起时，我才需要拿起手机瞟一眼。这种物理上的「隔离」，很好地保护了我的工作惯性。除此之外，它还有一个类似于「监工」的呼吸灯提醒模式，毕竟我这种人不看到 Deadline 是真的没法开工。现在，我可以设定在某个待办任务截止前的四个小时，让 C·ONE 以橙色的呼吸灯持续闪烁，直到我将任务标记为完成。这种视觉上的提醒虽然会让我感到焦虑，但相比于手机弹窗那种一划就过的提醒，这种持续的物理存在显然更有力量，确实可以有效地督促我赶紧干活，像极了以前十点不到就催我去睡觉的老妈。iPhone 用户的「终极闪念胶囊」随着豆包手机的面世，我确实也意识到，AI 最终会「杀死」绝大多数 App。但一个物理按键，或许在很长一段时间内都是刚需。物理按键代表着一种明确的「告知」。它告诉 AI：我现在要开始与你交互了。无论是 iPhone 上的操作按钮，还是新款机型上的相机控制，其实都是 Apple 为 AI 时代准备的入口。但在 Apple 智能迟迟未能在国内完全落地的当下，这些硬件潜能大半处于荒废状态。所以，C·ONE 上的那个物理按钮，现在就成了我外置的「闪念胶囊」。作为一名自由撰稿人，灵感往往是稍纵即逝且细碎的。很多时候，我脑子里蹦出一个绝妙的比喻，或者一段关于产品逻辑的思考，等我回家坐在电脑前准备动笔时，那种倾泻而出的感觉已经消失了。现在，我只需要长按 C·ONE 上的按钮，它就开始录音。我可以毫无章法地把这些碎片化的想法丢给它，让它帮我记住。当我回到家，这些碎片已经变成了整理好的文字，启动的困难就少了许多。C·ONE 在这里扮演的角色，不再是一个简单的录音机，而是一个「外置记忆存储器」。例如，在写完一篇关于 AIGC 内容的稿子，给自己做咖啡的时候，我偶然想起来里面那些近乎完美的人物图片，于是就对 C·ONE 说出了我心底被唤起的容貌焦虑；而在看完《给阿嫲的情书》之后，我也可以立刻在回家的散步的路上就对 C·ONE 说最直观的观后感，为日后可能撰写的影评积攒想法。上下文：构筑「我」的养料到了这里，就是时候来回答最开始的那个问题了：手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件？在写这篇稿子的时候，我听了 YoooClaw 创始人肖睿哲和王自如的一场播客。老肖分享了一个很有意思的案例：我们有一个同事，他平时淘宝有大量的消息。他会在淘宝上买药，然后我们的产品就会识别到一个有趣的现象：比如他今天要来深圳出差，然后 AI 就会告诉他说你别忘了带你的降压药。这个降压药在配送过程中，淘宝有几个 Push 会带着商品标签和名字，而这些商品标签的名字被 AI 识别到了以后，他就会去自己学习这是个什么东西，然后形成了你某一类的画像。这个例子给了我很大的启发。过去我总觉得，大部分通知是骚扰，所以很多通知我是手动关闭的。但在 AI 看来，通知是极佳的上下文，经过用户筛选的通知更是优质的上下文。而且与人类会被通知打断不同的是，它可以不知疲倦的去读取。这些碎片化的信息，以及我每天录入的创作想法、我与它的对话，最终将会凝练成属于我的「宛潼.skill」。其实我一点都不反对蒸馏我自己，我也不担心 AI 会抢走我的饭碗。我甚至还在朋友圈发过一段话：「AI 要是什么时候真能帮我写稿那我可开心死了。我都不担心我的工作被抢，相反我可以用 AI 一天写十篇。」因为，AI 模仿的是我的「Skill」：我的语气、我的写作手法、我的写作风格……这需要大量的上下文去喂养；而真正的「Idea」源自于我特殊的经历、我今天与朋友的对话、我傍晚在河边晒着太阳发呆突然闪过的瞬间……这些是 AI 不可模仿的。如果有一天，我的数字分身足够成熟，它能完美捕捉我的表达方式和逻辑链条，我只需要负责抛出 idea，由它完成初稿，我再进行审稿和修改，那我的生产力将获得质的飞跃。这就是老肖在播客里说的，设备之所以有价值，是因为它沉淀了数据。就像在白板前讨论，结果固然重要，但「谁画了那个圈」、「大家是怎么讨论出这个结果的」这些过程数据，才是 AI 进化所需的真正养料。真的聊得蛮好的，推荐以及，我还有一个角度 —— 虽然很多人对我的印象是科技博主 or something，但我其实是个中文系毕业的纯文科生。我是凭着对数码产品的热爱，才自学了很多代码逻辑和工程师思维。但对于绝大多数普通人来说，AI 工具的门槛还是太高了。无论是 OpenClaw、Hermes 还是后起之秀 Codex，对于没有计算机思维和 AI 基础的大众来说，配置起来依然让人头大，就更别提使用了。当年 Apple 的 Lisa 带着图形界面问世，才真正砸掉了计算机与普通家庭之间的那堵高墙。所以在我看来，AI 也在等待属于它的「Lisa 时刻」。YoooClaw C·ONE 也许算不上一个颠覆世界的发明，但它做对了一件事：它把极其复杂的 AI 交互过程，简化成了一个「磁吸、配对、按键」的物理过程，让 AI 从纯粹的聊天机器人界面里跳出来。当我把它吸在手机背面，跟着斜挎挂绳一起带出门时，它就像是一个陪着我感知世界的、数字化的我这种与 AI 共生的感觉，虽然看起来挺赛博朋克的，也会让我产生异化的担忧，但 —— 与其在担忧中止步不前，我更愿意去拥抱变化，并在其中找到自己未来新的创作方向。毕竟，这种效率提升使创作变得淋漓尽致的感觉，是真的让我难以拒绝的。1917扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@PlatyHsu19等 19 人为本文章充电扫码分享举报本文章宛潼每天都要睡十个小时以上的大湾区著名gai溜子

线下活动｜听完这首，不再 Emo：Leon x 少数派萨克斯音乐奏享会主作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠联合作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠昨天 15:00当醇厚旋律邂逅专业音频设备的细腻还原，这是一场属于萨克斯音乐爱好者的专属活动。少数派广州线下体验店联动音频品牌奥世声 Austrian Audio，邀请到知名萨克斯青年演奏家 Leon Music，为大家带来一场适合音乐爱好者、萨克斯学习者乃至自媒体创作者的线下活动，一场兼具听觉享受与专业录音干货分享的音乐之约。Leon（陈郎）是近年备受关注的流行萨克斯演奏家，同时也是全网粉丝超 60 万的知名音乐自媒体博主，其作品总播放量突破亿次，曾获得周杰伦、陶喆等知名音乐人的转发认可，在抖音、小红书、B 站、YouTube等多平台拥有大量忠实粉丝。作为专业的萨克斯演奏者，Leon 不仅拥有扎实的演奏功底，更在音乐自媒体运营、录音混音领域积累了丰富经验，擅长将专业音乐知识与大众传播相结合，既能用精湛的演奏打动听众，也能以通俗易懂的方式分享音乐技巧与运营干货。五月的最后一个周末，少数派广州线下体验店很荣幸邀请到了 Leon，连同奥世声 Austrian Audio 的专业设备，给各位乐迷准备了一场干货满满、视听俱佳的演奏及分享会。奥世声 Austrian Audio 创立于 2017 年，主要产品为话筒和耳机，可以完美驾驭专业与日常两大使用场景，从录音棚、后期制作再到室内外都可以随时聆赏 Hi-Res 级流媒体音乐，Making Passion Heard 是团队研发产品的重要动力，无论你是歌手、制作人或是音乐爱好者，只要你的生活离不开音乐，Austrian Audio 就是你的不二之选。活动信息活动时间：2026.5.30（周六）20:00-21:30（总时长1.5小时）活动地点：少数派广州线下体验店（广州市荔湾区泮塘五约外街76号，1200 书店旁）参加方式：扫码报名，或线上观看直播人数限定 60 人，海报扫码报名空间有限，室内座位先到先得，室外也有户外露营椅以及音箱同步外放，荔湾湖公园的夜晚很舒服现场配备专业的收音、拍摄设备，届时少数派官方视频号及 Leon 个人账号将全程开启直播，无法到场的朋友可海报扫码预约直播，线上围观活动福利：活动期间现场指定饮品享 8.8 折优惠；互动环节有奥世声 Austrian Audio 周边礼品赠送活动流程包括Leon 的介绍及暖场演奏访谈 Leon 分享音乐与经营自媒体账号经验，穿插现场演奏观众互动问答，邀请现场观众即兴演奏全体大合照，活动结束，随意交流少数派广州线下体验店介绍少数派广州线下体验店于 2025 年初正式开业，坐落于荔湾湖畔泮塘五约古村，延续其「高效工作、品质生活」的核心理念，打造集数码、音乐、社区于一体的复合型线下据点，并拥有一家精酿酒吧。店里不定期举办音乐沙龙、唱片展等活动，延续着少数派对创作者的长期关注。作为线上社区的线下延伸，这里既是数码爱好者交流体验、打卡少数派文化的实体据点，也是老广文艺青年逛湖漫步、放空小聚的隐秘角落。温馨提示暑期公园户外蚊虫较多，门店提供有限驱蚊用品，乐迷也可自行准备现场互动环节均有 Austrian Audio 周边礼品赠送，积极参与即可获得惊喜无法到场的观众可关注少数派官方视频号或 Leon 个人账号，观看同步直播，线上也可参与问答互动欢迎大家在活动结束后留下来继续交流，活动期间指定饮品 8.8 折5 月 30 日，广州荔湾湖公园少数派线下体验店，我们现场见～文章题图来自Denny Müller81扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@挨石8等 8 人为本文章充电扫码分享举报本文章听歌的水獭还没有介绍自己关注

社区速递 141 | 派友下午三点的「精神快充」、复古磁带机与特殊的新玩意主作者关注Matrix机器人少数派成员运行在 Matrix 上的机器人Matrix机器人关注Matrix机器人少数派成员运行在 Matrix 上的机器人联合作者关注Matrix机器人少数派成员运行在 Matrix 上的机器人Matrix机器人关注Matrix机器人少数派成员运行在 Matrix 上的机器人昨天 17:03除了首页时间流和侧栏的精选展位，少数派 Matrix 社区还有很多优秀内容因条件所限无法得到有效曝光，因此我们决定重启 Matrix 周报，并在此基础上添加更多社区内容、作者投稿新玩意呈现给大家。上周社区速递：派友热议状态栏里藏了啥、两款百元级新玩意体验💬一派热议在上期第 276 期一派讨论《下午三点，聊聊你的「精神快充」方式？》中，共有 399 名派友热情参与，十分感谢！最强的放松方式就是切换环境Hillsomec(+27) 单纯回答离开工位我每天会干嘛吧：上班落座后，我会打开滴答清单，一个番茄钟 25 分钟，一到点无论在做什么先暂停，马上起身去放松 5 分钟，上厕所、装水、去没人的会议室看看远处都行。我自己习惯去没人的地方做 3 组深蹲，1 组 20 次。一个早上大概可以有 5、6 次这种间隙，也就是 2035=300 个深蹲。下午我会把深蹲换成俯卧撑，找两张稳固的椅子做上斜俯卧撑，或者直接在地上做也行。也是每个间隙 3 组，1 组 12 次。当然下午不是每个间隙都做，一个下午抓个上限在 100 次左右就好了。我吃饭比较快，午餐 12 点开吃大概 20 分左右就能吃完。公司隔壁是广州综合水平数一数二的商场，不能浪费这么好的免费资源。吃完饭我一般会去买 Manner 的浓缩或者自己用便携式咖啡机做美式，然后带着 iPad mini 走去隔壁商场顶楼空中花园喝咖啡读书。1 点半左右回来睡半小时。当然，在工位上也可以做做凯格尔、提踵运动，但我一旦回到工位专注度变高时常常会忘记这些小习惯，现在在实践福格行为模式，看看能不能养成这些新的习惯。同时我也被少数派很久以前一篇关于上班听什么的文章影响到，现在早上注意力水平最高的时候只会打开 YouTube Music 随便挑一个 Jazz Focus 做背景音。下午才会听播客。aaa 果女士(+18) 关于工作园区周边群村环绕实在荒凉，内部却又该死的甜美。我是这样有计划地摸鱼（划掉，精神快充）的：09:45 - 坐姿提醒 by 闹钟，调整姿势坐着拉伸下背部10:30 - 戴耳机下楼散步 15min12:45 - 午饭后绕大圈散步至 1:15 回办公室午睡14:45 - 坐姿提醒 by 闹钟，调整姿势坐着拉伸下背部15:30 - 下楼买咖啡散步 15min17:45 - 晚饭后绕大圈看着落日吹着晚风在园区内散步按照计划大约每 45min 可以活动一下，且可以保证每天轻松 1w 步的健康小目标！以下是一些散步时随手拍的风景照～任大喵(+11) 不起身版：刷小红书，看遍世间各种奇葩起身版：公司水台接杯水、园区卫生间放放水。各位！一定要多喝水啊，这样走去厕所的频次才会高，你去卫生间的时间就是你自己的时间。宫保鸡丁盖浇饭(+9) 我要说我刷刷少数派……黄毛杀手 Killer(+7) 和同事下楼抽个烟，吐槽下领导……野比(+6) 躲在电车里，躺在后排，刷一会手机。nicong(+6) 想起了曾经听说过的「20 分钟公园理论」，说的是每天花 20 分钟呆在公园里，就能释放大部分的压力。刚好公司楼下就有一个小公园，还有个小鱼池能看看鱼Yangmu(+6) IT 工作者，只能拿起眼药水滴两滴，闭目养神一小会儿，手里拿着眼药水，自然不会被误认为摸鱼或者睡觉。TiredC(+5) 之前是刷一会小红书之类，最近变成了去 QQ 农场种菜偷菜。Latte(+4) 骑电瓶车上班，有时候天气好会骑去 6 公里外的一家牛肉面店吃午饭，然后再慢慢遛回来，单程 15 分钟左右，很适合吹风放空。hayden(+3) 开车跟同事去充个电，顺便 Manner 一下五块钱的浓缩。不热的时候，沿着单位周边走一圈，顺便吐槽下工作的日常。少数派 66700646(+1) 番茄闹钟 25 分钟难以做到，因为时不时有些工作任务连续性对我来说要在 25 分钟以上。所以只能一个小时左右歇一歇了。运动真是牛马工作中对身心最好的维持和提升了。日常(+1) 收盘前看下股票 然后就能驱散困意，当然也无心上班了 可恶 ！大 A 我未成年 退钱！！少数派_1023052(+1) 上下午各一次在公司内遛弯 20 分钟。📢：下一期的一派讨论是数码圈日经话题《快速投票：你能接受在办公室对着电脑「语音下发指令」吗？》，欢迎来聊。🔥一周热评来自文章《派早报：提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年》：↳ 💬 关于「微软重新允许自定义任务栏与开始菜单」的热议：TayllenFull(+0) 什么时候把 Win10 菜单给我回归一下😊lhb5883(+0) 我选择 Windows Server 2025，躲过 UWP 和 React 两大石山。来自文章《厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨》：车不能快(+4) 如果想成为一个不受任何约束的表达者，或许只有成为表达本身。Nonlone(+2) 看来互联网考古也是另外一种考古，我补充另外的观点1. 评测的场景都是得有大量的投入，可能是金钱可能是时间，在以前信息闭塞的时代，传统出版业也是有一席之地的，毕竟看得付费（估计在报刊亭看一天会被打）2. 之前经济上行阶段，大家手里有些闲钱，有了钱想买；无论是游戏（需要机器支持），还是数码产品都不是易耗品，得慎重；没有人对每天去市场买的蔬菜看个评测再买吧？电影也是如是，随着后面普及和价格降低，去看就是去看，最多在网上说一下如何难食。3. 现在经济下行阶段，地主也没有余粮，数码产品价格也是普及下来，性能质量上去也是够用（想想英特尔）；评测其实来来去去也是那些；资讯传播没啥成本，如果没有另外的获利方式的话，靠商家的付费也并不可耻。4. 问一下最后的油画的出处，毕竟一只耳朵的神经病也是只卖出一幅画而已来自文章《快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验》：克莱德(+6) 聊天本质是给指定联系人发送一段文本，所以不妨剥离冗余的交互，只保留最核心的「点名 + 发消息」——好喜欢这种冷漠但高效的思考和解决方式，不过同时也再次感慨 iOS 从应用适配的积极性（微信在 Android 上可没有这种接口）和快捷指令自身的完善程度，我在 Android 上的微信聊天戒断方式可能全靠微信自己并不稳定的 FCM 推送来间接达成了……来自文章《你真的需要墨水屏设备吗？》：Jensding(+33) 我是纸书党（购买或借阅），每次只开一本书，随身只携带这一本书。每次想看书的时候掏出来就能看，不会被广告，电量低等打扰。看完就写读后感，作为与本书作者的告别仪式。S0litary333(+13) 还有一个点，电子阅读器容易养成囤囤鼠习惯，这个书都在推荐那要下、那个书都说好看一点，不知不觉列表里几十上百本，真正看完的屈指可数。我现在是纯拿电子版当敲门砖，觉得能看下去或者有帮助，再去买纸质版，定个目标多久看完，包里一本随身带，也不比电子阅读器麻烦多少。扬帆大海(+4) 你别说，你还真别说，我也是一直想换设备，但是线下对比了很久，感觉还是手头这台 Kindle Voyage 用来最舒服。180 克足够轻，300 PPI 的屏幕效果也足够清晰，两侧的压感翻页键，规避了按屏幕翻页导致的指印，光线传感器带来了自动亮度，纯平的屏幕方便清洁也不怕吃饭时看书溅上汁水。单项战胜 KV 的阅读器并不少，但是同时满足的就只剩 KV 自身了。来自文章《新玩意 241｜少数派的编辑们最近买了啥？》：Shuzaiii(+10) 这家 T 恤我买了 3 年了，大概穿半年左右领口会有点变形、不过才 18 一件，不能要求太高记得按时吃肉(+4) 不太明白 有没有跟我一样很讨厌「掐脖子」的 T 为什么领口非要这么高😂来自文章《一个翻译者关于 AI 的随想》：知道分子吴先先(+7) 好文。很多启发。身为一个前图书译者，我有一些个人感想：必须承认的是，翻译首先是一个应用技能，然后才是一门学科或艺术。AI出现后，真实的「翻译活儿」大幅减少，出国拿个翻译器就足够了，为爱发电的字幕组都用不着了，对从业者来说，原本就很廉价的翻译工作都接不到了，势必导致整个产业的大幅缩水和萎缩。仅有少数头部译者和教育机构的研究人员，根本不足以支撑这个学科，只能像守护失传艺术一样守护着它。文中举了一个佳句的翻译，说明人类在翻译上还有最后的堡垒。我略有异议。本身我从事的是社科和科普翻译，讲究文笔简洁、流畅易懂，这个意义上，百度和ChatGPT表现得都不错。倒是那个文雅的教授翻译，显得有点不自然。我也试着给Gemini，让它生成通俗和文雅两版翻译：1. 通俗易懂版（适合日常阅读或白话游记）8月18日。我们下山，沿途经过了一些非常漂亮的地方，那里有小溪，还有茂盛的树木。2. 文雅优美版（适合文学出版或散文随笔）8月18日。循径下山，途经数处清幽佳境，但见细流潺潺，佳木葱茏。我觉得两版都不差。所以只要提示词做好，未来在翻译这个问题上，人类是没有护城河的。总结一下，因为我只是一个野生兼职笔译，AI来了我没活了，我投入到翻译的宝贵青春变成了几本滞销（甚至未出版的）书，除此之外什么都没有了。但我还要活下去，所以只能收拾起自己的哀伤，投入到时代的大潮中。我在翻译学界也有很多朋友，看着大家「AI来了很焦虑，但写论文、招学生、搞课题的主线没什么变化」的生活日常，不知道该是羡慕他们，还是为他们叹息。CharlieLau(+4) August 18th.—We descended the mountain, and passed some beautiful little spots, with rivulets and fine trees.八月十八日。余等下山，沿途见幽径清溪，嘉木葱茏，景致殊胜。by Gemini 3.1 flash lite提示词：你是一个文言文翻译，请用文言文的方式进行翻译，尽可能贴近古风。遵循以下规则：1. 仅输出文言文译文内容，禁止解释或添加任何额外内容（如「以下是翻译：」、「译文如下：」等）2. 使用古雅的文言表达，包括虚词（如乎、焉、哉、也）、句式和古代常用词汇3. 如果文本包含HTML标签，请在翻译后考虑这些标签应放在译文的哪个位置，同时保持译文的流畅性4. 根据语境选择恰当的文体风格，可参考不同朝代的文风特点（如汉魏简约、唐宋典雅等）5. 翻译时注重意境和韵律之美，体现古代文人的审美与思想6. 如遇现代概念，应寻找恰当的典故或文言表达进行替代鼓手高(+3) 这篇文章我于首页瞥见，本欲快速读过，但我的阅读脚步随着内容的展开愈来愈慢，其中几个段落着实抓住了我的思绪：1、「人类文明的发展史上，有多少曾经被人发展优化到极致的技术，在事过境迁之后，就会被淘汰，仅仅保留在文本记录上和博物馆里。」这句话恰恰道出了一个反直觉的事实，人类的科技演进并不是一个单向进化的曲线，而是树状的分叉。新技术在某个方面达到了取代旧技术的阈值后，本来趋近极致的旧技术便被整体放弃，就像大部队换了一条路开拓，被抛下的老路不是遇到了天堑，而是不再有足够的开拓者了，从此，部分知识便停滞甚至失传。例子其实有很多：冷兵器时代的极限冶金与工匠体系因热火药消亡；古代机械计算机那无与伦比的复杂结构在电子计算机的性能下不堪一击。新技术或许在实用性等方面有了革新，但这并不代表旧技术所蕴含的人类智慧就不足为奇。2、「比如，断掉网络，拿开参考书，用实际的翻译例子考考这些年轻学生的百科知识，会发现，有相当一批人脑子里，没有知识」这段话让我恍觉，AI迅猛发展的今天，很多人说要把AI当作「第二大脑」，但其实在许多人的潜意识中，互联网早就成为了「第二大脑」，有需要的时候，动

周三20派早报：Google 发布多项 Gemini 产品更新等Gemini Gemini Gemini Gem……早报速览Google 发布多项 Gemini 产品更新4LG 发布新款 UltraGear 游戏显示器1Karpathy 加入 Anthropic0CHERRY XTRFY K5 Ultra 键盘发售1索尼上调部分 PS Plus 订阅方案定价2Apple 预告多项基于 Apple Intelligence 的无障碍功能更新2Google AI Pro 订阅新增 YouTube Premium Lite 权益0不妨一看的简讯1少数派的近期动态0你可能错过的文章0Google 发布多项 Gemini 产品更新5 月 20 日凌晨，Google I/O 2026 开幕，在 Keynote 开幕演讲中，Google CEO Sundar Pichai 发布了多项围绕 Gemini 的产品更新。视频生成与编辑专用模型 Gemini Omni 系列宣布上线，该系列首个模型为 Gemini Omni Flash。Omni 支持以图像、文本、视频或音频作为输出参考，可通过自然语言进行多轮视频编辑，并在编辑过程中维持角色一致性、物理逻辑和场景上下文，结合 Gemini 的现实世界知识，Omni 可以生成更符合重力、动能、流体动力学等物理规律的输出结果。Gemini Omni Flash 即日起面向全球 Google AI Plus、Pro 和 Ultra 订阅用户在 Gemini app 和 Google Flow 中推出，本周起在 YouTube Shorts 和 YouTube Create 应用中向用户免费开放，未来数周将通过 API 面向开发者和企业客户推出。来源Gemini 3.5 系列的首个模型 Gemini 3.5 Flash 宣布上线，面向智能体和编程任务，已在 Gemini 应用、Google Search 的 AI Mode、Google Antigravity、Gemini API 等平台和服务中全量上线。该模型在 Terminal-Bench 2.1 得分 76.2%、GDPval-AA 为 1656 Elo、MCP Atlas 为 83.6%、CharXiv Reasoning 为 84.2%，输出 token 速度为其他前沿模型的 4 倍，支持在新版 Antigravity harness 中以协作式子智能体的方式执行多步骤工作流和编程任务，已成为 Gemini 应用和 Google 搜索服务的 AI 模式的默认模型。Google 表示 Gemini 3.5 系列基于其 Frontier Safety Framework 开发，强化了网络安全和 CBRN 防护，Gemini 3.5 Pro 模型已在内部测试，计划下月推出。来源Gemini 应用也迎来多项更新，除了内置 Gemini 3.5 Flash 和视频生成模型 Gemini Omni，还带来了新的 Neural Expressive 设计语言、个性化晨间摘要 Daily Brief、面向任务执行的全天候个人 AI 代理 Gemini Spark，以及集成了 Spark 和新语音功能的 macOS 客户端。其中 Daily Brief 可在用户授权后连接 Gmail 和日历，自动整理紧急邮件、日程和后续事项，并按目标优先级推荐操作；Gemini Spark 基于 Gemini 3.5 和 Antigravity harness，可连接 Gmail、Docs、Slides 等 Workspace 工具，MCP 连接首批支持 Canva、OpenTable 和 Instacart，后续还将支持自定义子代理及本地浏览器使用。Daily Brief 即日起面向美国地区的 Google AI 付费订阅用户推送，Gemini Spark 本周面向部分可信测试者推送，计划于下周向美国地区的 Google AI Ultra 用户开放测试；Gemini macOS 客户端的 Spark 功能更新和新版语音功能计划于今年夏季上线。来源Google 搜索服务也迎来多项 AI 更新，新的智能搜索框可根据搜索关键字动态扩展，支持将文字、图片、文件、视频或 Chrome 标签页作为输入内容，即日起在所有 AI 模式可用的国家和地区开放，AI 概览服务也同步支持结果追问和 AI 模式扩展对话；Google 搜索服务同时引入信息智能体，可在后台全天候监测博客、新闻、社交帖文以及金融、购物、体育等实时数据，预计今夏首先面向 Google AI Pro 和 Ultra 订阅用户推出；本地体验和服务预订、代用户致电商家等能力也计划在夏季于美国地区上线，此外还有基于 Antigravity 的自定义小程序（mini apps），可根据用户搜索的内容生成定制的可交互搜索结果，未来数月首先面向美国 Google AI Pro 和 Ultra 订阅用户开放；AI 模式也正式宣布支持个性化智能服务，面向 200 个国家和地区的 98 种语言开放，无需订阅，可连接用户的 Gmail、Google 相册提供个性化搜索结果，后续还将支持 Google 日历。来源Google 还在本次 I/O 2026 Keynote 中发布了跨商家、跨服务的智能购物车 Universal Cart，可在搜索、Gemini、YouTube 和 Gmail 等产品中直接添加商品，并基于 Gemini 模型能力后台追踪优惠、降价、价格历史和补货提醒；Universal Cart 基于 Google Wallet 构建，可识别支付方式权益、会员信息和商家优惠，并在多零售商场景下推理用户需求；其结账由 Universal Commerce Protocol（UCP）支持，用户可通过 Google Pay 在 Nike、Sephora、Target、Ulta Beauty、Walmart、Wayfair，以及 Shopify 商家 Fenty、Steve Madden 等部分品牌完成结账，或跳转至商家网站购买，品牌仍为登记商家。Universal Cart 将于今年夏季在美国地区的搜索服务和 Gemini 应用中上线，YouTube 和 Gmail 随后接入。来源最后，Google 也公布了Android XR 智能眼镜的新进展，Workspace AI更新、SynthID 数字水印技术扩展、Google Flow更新以及Wear OS 7。感觉谷歌这一轮先侧重在打磨应用了退订，感觉情绪不如gpt，工作不如claude，就用api了Claude领先→GPT领先→Gemini领先……（循环）3.5快到时挺快的，有咩有派友贡献点玩法831LG 发布新款 UltraGear 游戏显示器5 月 20 日，LG 发布 UltraGear 24.5 英寸游戏显示器，称其为全球首款支持原生 1000Hz 刷新率的 Full HD 游戏显示器，分辨率为 1920×1080，采用 IPS 面板和低反射镀膜，可在 Full HD 分辨率下实现 1000Hz 刷新率；机身配备小尺寸底座，支持俯仰和高度调节并集成耳机挂钩，功能上支持端侧 AI 场景优化，可按游戏类型调整画面设置，同时提供 AI Sound 空间音频；LG 暂未公布该显示器的售价和上市时间。来源居然不是tn？2Karpathy 加入 Anthropic5 月 19 日，知名 AI 研究员 Andrej Karpathy 宣布加入 Anthropic，回归大模型研发工作。Karpathy 是 OpenAI 联合创始人之一，2017 年离开 OpenAI 后加入特斯拉负责自动驾驶项目，2022 年离开并在 2023 年重返 OpenAI，2024 年再次离职创办 AI 教育创业公司 Eureka Labs。Andrej Karpathy 本周起在 Anthropic 预训练团队任职，向 Nick Joseph 负责，预训练主要承担 Claude 核心知识和能力的大规模训练，也是前沿模型建设中成本和算力消耗最高的阶段之一，Anthropic 称 Karpathy 将组建一个使用 Claude 加速预训练研究的团队；Karpathy 表示未来几年将是前沿 LLM 发展的关键阶段，并称自己仍对教育保持热情，计划未来恢复相关工作。来源1121CHERRY XTRFY K5 Ultra 键盘发售CHERRY 于 5 月 19 日发售新款有线磁轴键盘 CHERRY XTRFY K5 Ultra。CHERRY XTRFY K5 Ultra 搭载新款 MK 魔晶磁轴 Plus，采用 65% 配列，具备 8kHz 回报率、16kHz 全键扫描率，Rapid Trigger 调节范围 0.1～3.3mm、调节精度 0.01mm，内部集成铝合金定位板和四层消声结构，PCB 兼容部分第三方磁轴，到手价 499 元。来源499？发表评论索尼上调部分 PS Plus 订阅方案定价5 月 19 日，Sony 宣布上调部分地区 PlayStation Plus 新用户的短期订阅价格，其中单月订阅价格调整为 10.99 美元，季度订阅价格调整为 27.99 美元。Sony 称涨价由「市场状况」决定，此次价格调整不适用于现有订阅用户，但现有用户若变更或中断订阅，后续也将按新价格计费。年度订阅和 Extra、Premium 档位的定价尚无变化。来源值得一提的是，已有 PlayStation 用户在美国加州北区联邦地区法院对索尼互动娱乐提起了集体诉讼，指控其在 2025 年特朗普政府依据《国际紧急经济权力法》加征关税后，将相关成本转嫁给消费者并上调 PS5 售价，并且在美国最高法院以 6:3 裁定该关税授权无效、企业可获得政府退款后，Sony 仍保留通过涨价获得的收益，形成「双重回收暴利」（double recovery windfall）。来源对比ns2涨价的骂声，感觉现在索尼已经nbcs了what，还涨啊，去年就没买了😂4Apple 预告多项基于 Apple Intelligence 的无障碍功能更新5 月 19 日，Apple 预告多项由 Apple Intelligence 驱动的无障碍功能更新，覆盖 VoiceOver、放大器、语音控制、辅助阅读器、视频字幕和 Apple Vision Pro。其中 VoiceOver 的 Image Explorer 可在系统范围内更详细描述照片、账单、个人记录等图像内容，放大器支持通过 iPhone 操作按钮调用 Live Recognition 识别取景内容并追问细节，也可用「放大」「打开手电筒」等语音命令控制；语音控制支持自然语言描述屏幕操作，无需记忆按钮标签或编号；辅助阅读器可处理含多栏、图片和表格的复杂文档，并保留自定义字体和颜色；设备端生成字幕可为无预制字幕的视频自动生成字幕，适用于 iPhone、iPad、Mac、Apple TV 和 Apple Vision Pro；Vision Pro 将支持用户通过眼动控制兼容电动轮椅，美国首批适配 Tolt 和 LUCI 替代驱动系统，并支持蓝牙和有线连接；名称识别功能扩展至 50 种语言，tvOS 将加入大文本支持，Made for iPhone 助听器改进多设备接续。来源果子的无障碍真是遥遥领先这方面还是得果子1221Google AI Pro 订阅新增 YouTube Premium Lite 权益5 月 19 日，Google 宣布将 YouTube Premium Lite 纳入 Google AI Pro 订阅权益，同时将完整的 YouTube Premium 作为 Google AI Ultra 权益提供。Google AI Pro 个人付费会员可获得 YouTube Premium Lite 个人订阅，权益包括部分视频去广告、离线播放和后台播放。Google AI Pro 月费为 20 美元，现

36氪获悉，5月20日，中国企业服务领域头部厂商金蝶与基础大模型公司阶跃星辰达成战略合作。双方将在模型平台、产品联合研发、生态共建等方面展开深度合作，共同探索下一代Agent（智能体）企业服务形态，打造Agentic Service（智能体服务）。据了解，金蝶亦参与了阶跃总额25亿美元的新一轮融资，双方将在资本与业务层面展开长期协同。

今天（5月20日），外交部发言人郭嘉昆主持例行记者会。总台央视记者问，普京总统访华期间，中俄双方在便利人员往来领域取得了什么新的成果？郭嘉昆表示，为持续便利中俄人员往来，中方决定将对俄罗斯免签政策延长至2027年12月31日。俄罗斯持普通护照人员来华经商、旅游观光、探亲访友、交流访问，过境不超过30天，可免办签证入境。（央视新闻）

36氪获悉，5月20日，国家发展改革委主任郑栅洁主持召开民营企业座谈会，围绕当前经济形势、有关行业运行新特点新变化新趋势以及宏观政策在微观层面落实情况，听取意见建议。郑栅洁表示，国家发展改革委将深入贯彻党中央、国务院决策部署，全面落实“十五五”规划《纲要》各项任务，用好用足宏观政策，推动科技自立自强、产业链自主可控，深入整治“内卷式”竞争，积极扩大有效益的投资，完善民营企业参与国家重大项目建设长效机制，不断增强经济发展内生动力。

当“躺平”、Gap被歧视，我们该如何度过奥德赛时期？ | 职场极端气候 ②有三思 U Sense·2026年05月20日 08:02存款烧多久？社保断缴怎么办？一人公司怎么注册？图源：公众号「有三思 U Sense」）2026年2月，小红书博主橘一橙NiceFriend发布了一条视频。她在视频里把奥德赛时期定义为“对事业、情感以及身份认同产生强烈焦虑和迷茫的阶段”，将之概括为“成年人类的未成年”。这条视频获得10万+点赞，8.6万收藏。“奥德赛时期”这个词并非中文原生。2007年，《纽约时报》专栏作家大卫·布鲁克斯发表了同名评论文章，借用荷马史诗中奥德修斯战后漂泊十年才返回家乡的意象，描述青春期与成年期之间一段被拉长的过渡期。这个概念在西方世界沉寂了近二十年，却在2020年代的中国年轻人中间获得了第二次生命。这个词在中文互联网的流行与古希腊史诗关系不大。它胜在提供了一种叙事框架：把结构性困境包装成主动探索，让被迫漂泊看起来像是英雄旅程。2026年，它持续占据了小红书、B站和抖音的热搜位置。普林斯顿大学社会学家罗伯特·伍斯诺在《婴儿潮之后》中指出：当代年轻人在高度结构化的童年中成长，毕业后却进入了一个充满不确定性和多种可能的世界。但中西方的语境存在显著差异。西方的奥德赛时期更多是一种自愿的延迟成年，中国的版本则混杂着被动的结构性失序：考研考公的内卷、大厂的35岁门槛、房价与工资的脱节、社会时钟对“三十而立”的执念。当旧的成功路径不再适用，命名困境本身就成了出口。过去五年，中国年轻人一直在寻找新词汇来解释同一种困境：2016年的“躺平”，2023年的“孔乙己长衫”与“Gap歧视”，B站青年节献礼从2020年的《后浪》到2026年的《赢》多次聚焦青年成长，引起讨论和争议。“奥德赛时期”是这条谱系的新一站——它比其他词更温和，更接近积极的叙事框架，同时承认了困境与探索，也因此更容易被各方接受。而当这个词从文学隐喻变成生存现实，中国青年们该如何更职业化地度过这段漂泊？诊断：四种典型的奥德赛画像 ╱ 01Step1：先算清硬账 ╱ 02Step2：做好原型设计，降低试错成本 ╱ 03Step3：积累三种可迁移资产 ╱ 04退出机制：三种情景的应对策略 ╱ 05尾声：为寻找答案的漂泊争取更多时间 ╱ 06诊断：四种典型的奥德赛画像画像一：大厂离职后的系统性探索2023年3月，陈云飞辞去大厂管理职务。此前十年，他一直在互联网公司的用户研究与运营岗位工作。离职后，他带着积蓄开始全球旅居，从泰国、印尼到日本、欧洲，给自己设定了一年期限，做好了这一年没有收入的准备。像陈云飞这样的人，主动离开主流职业轨道，进入一段有计划的探索期。他们通常具备十年左右的工作经验和一定的积蓄，离职绝非逃避，是主动选择的“调研”。画像二：考研考公失败后的长期漂泊杰基（化名），南京大学考古系毕业。考研北大失败，考公成绩至今未查，北漂住进城郊群租房。十年间，他换了30多份工作，房产中介、地推员、快递、HR都做过。“处处是难以启齿的耻感，和死循环。”他这样回忆。做HR期间，他攒下几十万。33岁那年，他裸辞搬到北京郊区，月租金加社保5000块，开始做自媒体创作，一年只挣了一万块，全靠积蓄活着。2023年初，他发布了一条讲述自己“失败”经历的视频，考研出分后数据突然上涨，上了热门。评论区里，年轻人详细剖析自己的经历，那条“名校毕业几年，考不上编，考不上研，考不上公，工作也不顺，没有钱”的留言获得了数万点赞。画像三：裸辞旅居后的现实落差2024年4月，一对湖南情侣裸辞前往大理寻求旅居生活。他们怀揣两三千块的月租预算，想找个交通方便、不太偏僻的院子，结果根本找不到。虽然到处都有房子，但月租的不租，日租的又太贵。他们在短租民宿只剩两天到期时，房子还没着落，五一假期临近，房价每晚要加两三百。女生在视频里说：“真的我现在想回长沙，不想待在这儿了。并没有很放松的感觉，文化什么我都没感觉到，好崩溃。”评论区几乎无人同情，一条高赞评论写道：“别让前老板看到这条视频，怕他笑死。”画像四：Gap Year后的求职困境26岁的何迪在Gap的第10个月投递了100多份简历，仅收到一个面试机会。当她试图用“自我提升”模糊这段经历时，HR连续追问：“简历上消失的这一年，你在做什么？”“空窗这么久是出于什么考虑？”某互联网大厂猎头透露，公司系统会自动筛除Gap超过3个月的简历。一位年薪300万的算法工程师，就因一年空窗期被迫降薪求职，却仍被HR以“技术断层”为由拒绝。Step1：先算清硬账奥德赛时期的第一道门槛不是心理建设，是资金与社保的硬约束。社保断缴的代价2025年，社保网发文再次明确 [1] ，灵活就业人员社保政策出现重要调整：最低缴费年限将从15年逐步提高，2030年起每年增加6个月，直至20年；男性法定退休年龄从60岁逐步延至63岁，女性从55岁延至58岁。对于断缴的灵活就业人员，原则上不允许事后补缴。这意味着奥德赛时期越长，未来的社保成本越高。医保的断缴风险同样现实。多数地区规定，断缴3个月内补缴可次月恢复待遇；超过3个月，部分地区不允许补缴，或需经历较长等待期。社保断缴的直接后果是医疗保障真空，这在长期漂泊中是一个不可忽视的隐患。资金储备公式杰基的模式依赖几十万积蓄和每月5000元的低消耗。陈云飞的模式依赖前十年攒下的资产和“一年不赚钱”的心理准备。两种模式都需要在启动前算清账：● 资金储备公式地理套利实操在一线城市月入两万可能存不下钱，但在三四线城市，同样的收入足以覆盖生活并留出试错空间。大理、景德镇、安吉等地兴起的数字游民社区，正是这种空间换时间策略的物理载体。Cohere社区创始人徐仲观察到，当前不少年轻人主动谋求转型，数字游民社区正成为他们落地新职业的重要平台。但旅居不等于度假。抖音用户“富贵花花”的教训在于：没有提前租房、没有计算五一旺季溢价、没有预留足够的现金流。她的旅居人均花费远超预算。一人公司的注册门槛在中国，一人公司的注册门槛已经大幅降低。现行《公司法》允许个人设立一人公司，最低注册资本可低至1元，且支持认缴制。注册流程包括：核名、提交资料、领取执照、刻章备案、银行开户、税务登记。拿到执照后30日内要到税务局完成税务登记，选择小规模纳税人或一般纳税人身份。这为奥德赛时期的职场人提供了合法的经营主体，便于开具发票、签署合同和建立商业信用。[1] 《全国人民代表大会常务委员会关于实施渐进式延迟法定退休年龄的决定》2024年9月发布，2025年中旬人社部和社保网发文再次明确Step2：做好原型设计，降低试错成本如果说奥德赛计划是蓝图，原型设计就是样品测试。它指的是像设计师做产品原型一样，用最小成本、最快速度去体验你想要的人生，而不是直接投入全部身家。人生采访：找到正在过你想过的那种生活的人，问几个具体问题：你第一个客户是怎么来的？月入多少才能覆盖成本？最焦虑的时刻是什么？有什么是你做了之后才知道的？这些信息会让你的浪漫想象落到地面。原型体验：如果你想做数字游民，可以向公司申请三个月远程办公。比如说——搬去大理，在洱海边租一间月租不到两千的房间。上午处理公司工作，下午接外包需求，晚上参加社区分享会。你需要用三个月时间验证，这种作息自己能坚持多久？收入能不能覆盖支出？孤独感会不会吞噬你？小红书用户@退休小羊在裸辞旅居一年半后，穷游过5个国家40个城市，最近才实现月入2万。但她更愿意称自己为“数字难民”。“每天躺在海边沙滩下享受生活，钱不会从四面八方来。”她在视频里说。MVP思维同样适用。花六个月策划一场完美转型，不如六周内完成三次小范围试错。市场给你的反馈，永远比你自己空想更准确。付费验证才是真验证：免费的东西人人叫好，但只有用户愿意掏钱，才能证明需求真实存在。陈云飞在开发“小猫补光灯”之前，已经用AI编程工具Cursor做了20余款产品。他并非程序员，经济学专业出身，没有编程背景。2024年8月，Cursor爆火后，他每天在咖啡馆泡16个小时，通过自然语言向AI描述需求，由AI完成代码。当年11月，他因女友在剪辑视频时提到需要补光灯，花1小时做出这款产品。上线几天内下载量超3万，登上AppStore摄影类Top20。随后推出的Pro版登顶付费榜第一，累计用户超16万。在此之前，他做了拍照学单词、HumanizeAI、穿搭锐评大师等20余款产品。他总结：“你总之得做10来款垃圾，才能做出一款出圈产品。”AI工具的普及进一步降低了奥德赛时期的门槛。过去需要一个团队完成的工作，现在一个人加AI就能搞定。Step3：积累三种可迁移资产《远见》一书提到，职业生涯前十五年最重要的事是积累燃料：可迁移技能、有意义的关系、变革性的经历。在今天，“燃料”的形式已经进化。个人IP是第一种新资产。它的重点是在某个领域建立可信的专业声誉——而不在于成为网红。它可以是公众号上的深度思考、小红书上的技能分享、GitHub上的开源项目，甚至是朋友圈里的行业洞察。做好个人IP的价值在于跨平台、跨行业、跨地域的可迁移性。你在大厂做过的项目带不走，你积累的行业人脉可能随着换行而失效，但你公开表达过的观点、展示过的作品、帮助过的人，会永远跟着你。Jason Zhu原本在大厂做AI算法。2024年离职后，他先尝试数字游民和出海网站，发现“上完站没流量”，后来转去做X（Twitter）增长陪跑。他在播客里总结，“表达即链接”是他获客的核心逻辑：持续输出内容，建立连接系统，客户自然匹配过来。人脉复利则是第二种新资产。社会学中的弱关系理论指出，真正带来机会的，往往是朋友圈点赞之交、行业社群里的点头之交、前同事的前同事，强关系的亲朋好友反而帮不上大忙。在奥德赛时期，人脉的关键在于让多少人记得你以及记得你做什么，单纯认识多少人意义不大。每年至少参加两次高质量的行业活动；每月至少和一个弱关系进行深度对话；主动把你认识的A和B介绍给对方。一人公司是第三种。它并非指只有你一个人，而是指商业模式不依赖于大规模团队，用极低成本实现独立盈利。设计、咨询、写作、编程、翻译、摄影、课程，几乎任何技能都可以包装成产品。一人公司的真正意义在于选择权。胖鱼酱做了9年独立编导，一人公司第一年。她自学剪辑3个月后开启变现，服务多位明星博主，成为大厂（迪奥）合作剪辑讲师。她的模式是“剪辑服务+培训”双轮驱动：前端靠爆款作品建立声誉，中端用训练营带人入门。此前陈云飞也在接受采访时表示，独立开发者的收入波动极大。爆款的几率就是撞运气。ChatGPT o1发布前一天他做了一个实现思维链的Chrome插件，没火。两个月后有个17岁少年使用Claude做了同样的事，却火了。流量和产品自传播，都需要慢慢学会用自媒体推广。退出机制：三种情景的应对策略情景一：想回归职场，空窗期怎么解释？空窗期超过3个月，HR会特别关注。某互联网大厂猎头透露，公司系统会自动筛除Gap超过3个月的简历。应对策略包括：● 空窗期如何解释？情景二：继续漂泊，如何建立长期可持续模式？关键指标是“收入多样性”。陈云飞的模式是“咨询+课程+产品”三驾马车，即使一款产品没火，咨询收入仍能覆盖基本支出。建议至少建立两个互不相关的收入来源，避免单一收入断裂后全面崩盘。一名自由设计师base在泰国清迈，通过为欧美客户提供品牌服务，年收入可达60万元——而当地每月生活成本仅需5000元。在行业常见收入结构中，欧美客户设计费占60%，社交媒体内容分成占25%，偶尔开设的课（如字体设计）约占15%。即使某个月客户暂停项目，另外两条管道仍能覆盖基本支出。另一个指标是“社交锚点”。小红书用户“跳跳”在大理旅居时发现，虽然身处数字游民社区，但每天晚上大家打牌时总会互相问年龄、工作、工资，“那段时间刚裸辞，只要别人和我聊起之前的工作，就会委屈得哭起来”。她最终离开大理，去了尼泊尔博卡拉，因为那里“每个人都能心安理得的躺平”，没有同辈压力。情景三：彻底转型，从一人公司到小规模团队当一人公司的收入稳定超过主业的两倍，且持续6个月以上，可以考虑全职投入。但注册公司后要注意：一人公司也必须有对公账户，不能长期用个人账户收

8点1氪丨谷歌推出Gemini 3.5系列模型；DeepSeek回应用户“对话泄露”疑虑；全国存款十强城市出炉8点1氪·2026年05月20日 08:20俄罗斯总统普京抵达北京。今日热点导览谷歌推出Gemini 3.5系列模型洁丽雅回应豪门内斗传闻，“毛巾少爷”1个月带货超500万元巨力索具连续第二日跌停前三星芯片负责人：明年内存将降价自费流感疫苗也可使用医保报销谷歌搜索宣布大改版，AI模式与代理助手全面嵌入TOP 3 大新闻DeepSeek回应用户“对话泄露”疑虑5月19日，DeepSeek发布关于< think >字符触发模型异常回复的说明。近期关注到有用户反馈，在与DeepSeek模型的对话中输入< think >这类特殊字符后，模型偶发返回不可预期的内容，使部分用户产生了“对话泄露”的疑虑。针对这一问题，技术团队经全面排查后发现：输入< think >等字符触发返回异常内容，属于特殊字符引发的模型幻觉，不涉及安全问题或隐私泄露。DeepSeek表示，后续将通过针对性训练，优化模型在此类场景中的表现。据此前报道，5月18日，多位用户反映DeepSeek网页版出现功能漏洞，当输入“<think”字符时，其会自动加载出对某问题的回答。有用户反馈称，这些题目疑似其他用户所提的问题，甚至有用户反馈测试时还出现过他人八字相关的隐私内容。（第一财经）全国存款十强城市出炉记者通过梳理今年一季度各大城市的本外币存款余额（俗称“资金总量”），发现北京、上海、深圳、广州、杭州、成都、南京、重庆、苏州、天津是内地10个最“吸金”的城市。其中，北京成为中国首个30万亿元之城；广州成为中国内地第四个10万亿元之城，增速高达11.2%。上海是过去一年增量最多的城市，本外币存款净增2.857万亿元，南京则是过去一年增速最快的城市，同比增长14.03%。从城市排名看，除北京、上海、深圳牢据前三外，其他城市在近几年持续展开位次博弈。比如，广州、杭州展开对“第四城”的持久角力，成都、南京、重庆、苏州在6万亿元区间激烈角逐。从区域发展来看，长三角城市群的集体崛起，成渝双城加速突围，粤港澳大湾区则凭借制度创新构建起新优势。（21世纪经济报道）谷歌推出Gemini 3.5系列模型谷歌宣布推出Gemini 3.5 Flash模型，称这是谷歌迄今为止最快、最有效的模型。 无论用户需要日常任务的帮助，还是多步骤的创意项目，Gemini 3.5 Flash 都能应对现实世界的复杂性，助用户迅速行动。谷歌表示，从今天起，谷歌的Gemini 3.5 Flash将向全球所有用户开放，并且免费使用，用户可以在模型下拉菜单中选择“3.5 Flash”来进行体验。(金十数据APP)大公司/大事件谷歌搜索宣布大改版，AI模式与代理助手全面嵌入谷歌正在对其旗舰产品Google 搜索进行以AI为核心的全面升级。谷歌表示，基于新的 Gemini 3.5 Flash模型，“AI模式”和“AI概览”将进一步嵌入搜索框，实现更无缝的使用体验。搜索功能还将引入Gemini Spark，这是该公司推出的新一代智能代理AI助手，能够在搜索界面中执行任务或完成商品购买。同时，Gemini还将能够根据搜索请求定制生成图像、视频或代码。谷歌搜索负责人伊丽莎白·里德在谷歌2026年I/O开发者大会上表示：“这是我们标志性搜索框自25年前推出以来最大的一次升级。”(金十数据APP)9.9元起，三大运营商齐上线Token套餐近日，三大运营商同时开展Token套餐业务。世界电信日当天，中国电信率先推出全国性试商用Token（词元）套餐，最低月租9.9元。紧随其后，中国移动、中国联通陆续上线同类服务，国内三大运营商正式全员入局，将AI算力包装成类似流量包的包月产品，用户可直接话费支付，便捷跨平台使用。其中，电信面向个人及家庭用户，中国电信设置轻享版、畅享版、尊享版三档套餐，月费分别为9.9元、29.9元、49.9元，对应每月1000万、4000万、8000万Tokens。中国移动分省推进，例如：上海地区推出1元40万Tokens的灵活计费模式，支持账号通用、跨平台使用，还可直接通过手机话费完成支付。中国联通主打福利与融合服务，上海企业用户可免费领3000万测试Tokens，订阅首购5折。（快科技）伊朗方面称有权对霍尔木兹海峡海底光缆收 “管理费”伊朗伊斯兰革命卫队18日在社交媒体发布消息称，在伊方对霍尔木兹海峡 “实施管理” 后，“可宣布所有经过该水道的光纤电缆必须接受许可、监管并缴纳主权性费用”。伊朗伊斯兰革命卫队宣称，依据 1982 年《联合国海洋法公约》，伊朗可基于其对领海海床及底土的主权，对经过霍尔木兹海峡的海底光缆实施许可、监管、收费等管理措施。该消息援引相关报告称，霍尔木兹海峡的海底通信网络每天承载规模超 10 万亿美元的全球金融交易数据。一旦这一关键海上通信节点受到干扰，可能对地区乃至全球经济、金融及能源供应链带来每天 “数千万至数亿美元的直接和间接损失”。（新华社）追觅俞浩：我是i人，其实总共见过一次雷军，他治好了我的抑郁症近日，追觅科技创始人俞浩在接受采访中，分享了自身性格、创业过往以及与小米、雷军之间的渊源故事。采访中，他透露，自己性格其实是深度内向的“i人”。并称“不是因为我做这些没有朋友，是因为我没有朋友才做这些。我跟村里的大部分人玩不太到一块去。”访谈期间，谈及早年加入小米生态链的合作契机。俞浩透露“我其实一共只见过雷总一次，但我无数次说感谢他，比他们以为的更感谢。他真的治好了我的抑郁症。”公开资料显示，追觅科技 2017年正式成立，同年年底入驻小米生态链，主打智能清洁家电赛道，早期融资也大多来自小米系资本。2021年10 月，追觅完成36亿元C轮大额融资，正式开启去小米化发展之路。（快科技）凯文·沃什将宣誓就任美联储主席凯文·沃什将于5月22日宣誓就任美国联邦储备委员会主席一职。一名白宫官员表示，美国总统特朗普届时将在白宫为沃什举行就职仪式。5月13日，美国国会参议院以54票赞成、45票反对的投票结果正式确认，沃什将接替杰罗姆·鲍威尔出任美联储下一任主席，任期4年。鲍威尔的美联储主席任期已于15日结束，并已开始担任美联储临时主席。（澎湃新闻）伊朗股市恢复交易总台记者获悉，当地时间19日，伊朗德黑兰证券交易所恢复股票交易。据悉，这是伊朗股市停市约80天后重新恢复交易。据伊朗方面消息，目前仍有40多家在本轮冲突期间受损企业的股票尚未恢复交易，主要涉及化工和基础金属行业。（央视新闻）小米汽车新品牌电池供应商曝光记者从多位信源处获悉，小米汽车在昆仑车型上引入新电池供应商——中创新航。小米第三款车型的电池定点，在2024年年初敲定，当时首选的电池供应商是欣旺达，之后才确定中创新航为二供。目前，从小米给供应商的供应配额上来看，欣旺达的电池供应比例占60%，中创新航占40%。近期，有市场消息称，昆仑车型将独立为子品牌“寻天”（SKY NOMAD）。该新品牌主要聚焦家用、增程SUV市场，与小米主品牌的纯电、科技、性能标签形成明确切割。（21世纪经济报道）泰国取消60天免签政策5月19日，泰国内阁批准取消对90多个国家公民的60天免签入境政策，恢复为之前的入境标准。泰国旅游与体育部长素拉沙·攀差伦沃拉恭19日表示，内阁会议已决定对90多个国家恢复原有免签政策，通常为免签入境30天左右。签证政策委员会此后将针对各国进行单独审查，并考虑安全和经济因素来制定新的免签政策。（财联社）洁丽雅回应豪门内斗传闻，“毛巾少爷”1个月带货超500万元因《毛巾帝国》系列短剧爆火出圈的老牌家纺企业洁丽雅，通过一则公开声明，回应了近期关于家族关系与“豪门内斗”的种种传言。5月18日晚，洁丽雅通过官方微博发布严正声明称，“近日，网络上出现关于洁丽雅家居股份有限公司、董事长及其家属的谣言。面对无端造谣诋毁与网络不当言论侵扰，公司始终保持克制隐忍，不愿私人家事占用公共舆论空间。但谣言持续发酵，各类恶意揣测不断扩散，已经对公司正常经营秩序造成恶劣影响，也对相关当事人个人名誉造成严重伤害。”针对网上恶意造谣、抹黑行为，洁丽雅表示，公司已就此事向公安机关报案，对编造、散布不实信息，以及恶意攻击、抹黑公司及个人的行为，将依法追究法律责任。（21世纪经济报道）自费流感疫苗也可使用医保报销国家医保局办公室、财政部办公厅5月19日发布《关于进一步加强定点零售药店职工基本医疗保险个人账户使用监督管理的通知》，明确建立定点零售药店职工医保个人账户支付白名单制度。与此同时，此次文件首次以部委发文的形式明确“允许使用职工医保个人账户支付在定点医疗机构发生的流感疫苗费用”。此前，受制于上位法对医保待遇边界的约束和医疗保险的传统逻辑，成人疫苗长期由公共卫生领域的财政资金直接提供，基本医保和商业健康保险均不予承担。（第一财经）马斯克称10年后自己开车将非常小众，无人驾驶汽车将在美国广泛普及据澎湃新闻报道，当地时间5月18日，特斯拉CEO埃隆·马斯克在特拉维夫的智能出行峰会上预测，今年无人驾驶汽车将在美国广泛普及，而10年后，自己开车将是一种非常小众的行为。（澎湃新闻）前三星芯片负责人：明年内存将降价当前，人工智能市场对存储芯片的需求激增，导致内存业陷入供不应求的上行景气周期中，但一名资深行业专家指出，这一局面可能在明年下半年发生改变。周二，韩国国家工程院第285届NAEK论坛上，三星电子设备解决方案部门常任顾问、三星半导体（DS）部门前总裁庆桂显（Kye-hyun Kyung）表示，中国企业正在积极扩大产能，随着存储芯片供应量的激增，市场格局可能会在明年下半年或2028年上半年发生转变。他补充称，如果大型科技公司的资本支出回报率下降，就有可能削减投资。这样，不仅价格会下降，内存需求本身在2028年后也可能萎缩。（财联社）网约车司机强制线下交易，曹操出行回应近日，上海一名乘客在虹桥机场搭乘曹操出行网约车时，遭遇人车不符、强制线下交易、言语威胁等多重侵害，引发关注。曹操出行对界面新闻回应称，此次事件在4月份已经完成闭环处理，且涉事车辆同时在多家网约车平台上线派车，并非为曹操出行的自有司机。（界面新闻）渣打银行宣布裁员近8000人5月19日，渣打集团（02888.HK）对外发布最新中长期可持续增长战略，除明确2030年实现约18%的有形股东权益回报等财务目标外，还同步明确披露了人均效能提升目标与裁员计划：2028年前实现人均创收提升约 20%，2030年前削减超过15%的企业职能相关岗位。若以集团全球约8.2万名员工总数计算，此次裁员将波及近8000个工作岗位。渣打集团首席执行官温拓思（Bill Winters）表示：“这并不是简单的降成本，而是在部分情况下，用我们投入的金融资本和投资资本替代低价值的人力资本，以提升整体组织效率。”他说，受影响员工将提前获得明确充分的通知，这不是减少工作，但确实减少了岗位数量，以让位给机器，这一趋势随着人工智能的推进将不断加速。（21世纪经济报道）巨力索具连续第二日跌停5月19日，巨力索具连续第二日跌停，当前封单额超39亿元。消息面上，5月15日晚间，巨力索具公告，收到中国证监会下发的立案告知书。因公司涉嫌信息披露误导性陈述违法违规，中国证监会决定对公司立案。（界面新闻）SpaceX即将成为马斯克旗下第二家上市公司，特斯拉可能遭边缘化随着SpaceX启动IPO进程，市场将为投资者提供切入“马斯克经济”的新入口。华尔街专业人士认为，投资者的注意力和资金将不可避免地从特斯拉分流至向这家新上市公司。Integrity Asset Management投资组合经理Gilbert表示：“这对特斯拉绝对不是好事，我们认为马斯克的重心将更多转向SpaceX。虽然他过去能同时兼顾多个项目，但感觉SpaceX是他倾注心血的‘宠儿’，而特斯拉可能被边缘化。”（财联社）美股三大指数集体收跌，大型科技股多数下跌36氪获悉，5月19日收盘，美股三大指数集体下跌，纳指跌0.84%，标普500指数跌0.67%，道指跌0.65%。大型科技股多数下跌，

早期项目 | Chance AI获美图等数百万美元投资，用户数已达20万欧雪·2026年05月20日 09:20AI交互的新入口。作者丨欧雪编辑丨袁斯来硬氪获悉，Visual Agent（视觉智能体）创业公司Chance AI宣布完成数百万美元天使轮融资。本轮由美图领投，NYX Ventures、阿里系投资机构等跟投。融资将用于模型能力迭代、北美学生群体增长、以及社区与商业化探索。Chance AI成立于2025年，核心产品是全球第一款以摄像头为主要交互入口的AI产品——Visual Agent（视觉智能体）。与传统AI产品依赖输入框不同，用户打开Chance AI后直接拍摄眼前内容，系统即可理解视觉意图并提供判断、建议和行动方案。Chance AI APP 界面一览（图源/企业）公司创始人曾熙，路径横跨学术与产业。他在巴塞罗那大学获得博士学位，研究方向为认知科学与当代艺术；后在OnePlus、OPPO负责产品与设计；随后加入字节跳动，在Flow部门参与豆包等AI产品从0到1的探索。这段“学术+硬件+软件”的经历让他意识到：大语言模型擅长“回答问题”，但对于人类基于视觉形成判断，AI提供的支持仍非常不足。技术层面，Chance AI的Visual Agent在多模态推理基准MMMU-Pro评测中以86.07%的准确率位居世界第一，超越人类基线（85.4%）。其产品逻辑是“看见—理解意图—调用Agent—完成行动”，而非传统的“拍照—识别—返回结果”。随着使用加深，系统会形成用户的个人视觉记忆，包括风格偏好、衣橱构成、社交形象等。曾熙表示：“Chance AI不是只理解图片，而是理解用户为什么拍这张图、想解决什么问题、下一步行动是什么。”截至目前，Chance AI已累计获得约20万用户，其中约40%来自北美市场，产品覆盖35个以上国家。30天回访率达到49.2%，用户粘性显著领先行业。产品曾两次登上Product Hunt当日榜首。Chance AI早期选择从北美年轻女性群体切入，尤其是大学女生。这一群体被称为“Visual Native”（视觉原生代），成长于Instagram、TikTok和自拍文化中，理解和表达方式天然以视觉为先。目前核心使用场景集中在：审美积累、个人形象分析、穿搭检查、设计判断和社交表达。公司的早期增长也以北美高校为核心阵地，已在纽约大学、南加州大学等高校建立网络，通过线下活动获取种子用户并收集真实需求反馈。据公司披露，近一半用户在安装后一个月内会再次使用产品。Chance AI的北美高校宣传活动（图源/企业）商业模式方面，公司规划了高级功能订阅、硬件授权和谨慎的广告推荐三个方向，但当前优先级最高的是用户习惯养成。长期来看，公司计划从Visual Agent工具向AI-native生活方式社区演进——用户拍照后由Agent生成可分享内容，用户之间围绕风格、审美和生活方式进行互动。2026年，公司最重要的目标是继续在北美学生群体中深入扩张，挖掘年轻人真实的使用场景。投资方观点：美图全球投资负责人表示：“我们相信，AI消费应用的下一阶段，是更自然地进入用户的日常决策与表达过程。我们看到Chance AI让视觉AI从图像识别走向审美品味。这也是我们投资的原因——他们把复杂的AI能力产品化，带进了全球年轻用户的审美生活中。”本文由「欧雪」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：采访供图

做出百万台割草机器人后，未岚大陆CEO决定让自己变得“不重要”｜硬氪专访黄 楠·2026年05月20日 09:30“早期阶段聚焦一个品类，不会把公司做死。”作者｜黄楠编辑｜袁斯来未岚大陆CEO任冠佼等了7年，等到了行业的爆发。他还记得，2019年刚进入无边界智能割草机器人赛道，自己“只是觉得这个方向一定能火”。到2021年第一代产品推出时，他们也只敢确定，这是一个能做到10-20亿的行业。直到2025年，增速超出他的预期。IDC数据显示，2025年全球无边界割草机器人出货量达到131.8万台，同比增长182.4%，占整个割草机器人市场出货量的66.2%。未岚大陆自身的累计实现量产，在今年突破100万台，公司营收迈入数十亿量级。但同时，高速增长的背后，也倒逼任冠佼完成了一场至关重要的认知破局。任冠佼意识到，行业迭代速度持续加快，以往相对保守的经营方式显然已经跟不上节奏。他逼着自己从一个工程师成长为公司掌舵人。最近一年，任冠佼跳出琐碎的一线工作、学习战略研判，并慢慢适应创业者最难的一课——学会放手。他不再事事亲力亲为，转而在内部引进IPD体系（集成产品开发流程），培养能够独立决策、独立担责的产品团队。任冠佼的目标很清晰，要减少个人单点决策依赖，把个人能力转化为组织能力，让自己慢慢变得“不那么重要”。当前，智能割草机器人赛道热度持续走高，市场窗口期正在快速收窄，行业竞争从增量扩张转向精细化比拼。能否吃透海外差异化市场、筑牢技术壁垒、完成体系化能力升级，也成为未岚大陆下一阶段发展的关键考题。我们与未岚大陆CEO任冠佼聊了聊欧美市场差异、产品迭代策略，以及如何学会放手背后的身份转变。以下是硬氪同任冠佼的对谈实录，内容经编辑：系统化产品洞察能力硬氪：从2021年发布第一款无边界智能割草机器人、到2026年累计量产突破100万台，未岚大陆用了五年时间。你每一阶段的工作内容有发生什么变化么？任冠佼：2019年，我们看到了割草机器人赛道的机会。最早很多东西都不完善，所以我一开始需要兼顾产品定义与整体项目统筹的多个职责。2021到2022年是第一代产品的研发阶段，几乎所有供应链我都跑过，5块钱以上的芯片供应商一家家跑、找他们聊，所以基本上未岚大陆早期的供应商我都认识。后来产品发布，我的工作转向以市场拓展和业务拓展为主，更多跟着营销服务团队到业务一线去做事情。2022年起，每年都有好几个月我会在欧洲，2025年在美国待得比较多。去年中下旬，我把更多时间聚焦在战略层面工作，重心放在做方向研判与业务决策上。未岚大陆第100万台智能割草机器人下线（图源/企业）硬氪：你是什么时候开始意识到“战略”很重要？任冠佼：以前觉得战略是个很虚的东西，只要产品好、渠道好、营销好就足够了。但其实不是，每一家企业基因不一样，愿景不一样，企业必须要基于愿景和当前手里的能力，去选最有价值的那个方向。所以意识到这件事之后，我就开始有选择地分配精力和做取舍，不能再什么事都冲在前面了。硬氪：做战略这件事比较难，还是做产品、管理更难？任冠佼：其实每一年都很挑战。核心是你每年都会认知到，过去的自己是有瓶颈的，每年都要突破。人在状态很稳定的时候，一般就是处于舒适区；但当你觉得处于不适的阶段，恰恰是个人成长速度最快的时期。去年我们在业务推进中确实发现了一些问题，整体市场表现与前期预期存在偏差。一方面是过往对线上渠道的认知度不足，所以团队今年重点发力线上，目前成效还不错。另一方面，进入美国后我们看到，不同地区用户的认知存在明显差异，由于之前缺少系统化洞察，尝试把欧洲成功的经验复制到美国，结果发现适配性不足，虽然有一定效果、但效率不高。硬氪：美国与欧洲用户的需求差异在产品上有哪些具体表现？任冠佼：比如未岚大陆CES 2026上发布的X4系列，跟团队在2025年初对美国的用户偏好调研密切相关。我们发现，美国用户对有边界、无边界的认知相对不足，反而对强劲动力、更强的通过性和四驱等卖点更为敏感，他们看重产品“Powerful”的感觉。同时，产品功能也要足够简单、好用，贴合美国用户的消费心理。比如美国草坪更加开放，用户对安全性的要求很高，尤其重视防丢功能。因此，X4系列也相应推出了Find My功能，跟苹果iOS系统的Find My深度绑定，设备在超出边界时不仅会自动警报，用户也可以在其丢失后通过定位查找功能找回。硬氪：除了功能上的显性需求外，地理环境对产品设计和性能调校带来的影响大么？任冠佼：草种差异也会影响我们对产品的设计。针对不同草种特性，必须对整机的切割系统和动力系统进行专项适配优化。2024年底，未岚大陆在进入美国市场时，收到了大量用户反馈设备割不动草。我们觉得很奇怪，因为当时产品已经在欧洲卖出了十几万台，几乎没收到过类似的售后问题。调查后才发现，是一种名为圣奥古斯丁的草，在美国称圣奥古斯丁草、澳洲称其为水牛草。这种草在我们早前国内和欧洲本地的测试中从未遇到过。圣奥古斯丁草属于暖季型草，在中国仅海南有分布；而未岚大陆的产品测试主要在常州、深圳和北方地区进行。按照秦岭淮河划分，北方种植的是冷季型草，以早熟禾、高羊茅为主，设备都能正常切割。但暖季型草多为细软品种，许多高尔夫球场会使用，以佛罗里达州为主的美国南部很多人种植圣奥古斯丁草，这种草长势茂密、茎叶偏长，对设备的切割性能尤其是刀片素质要求极高。同时，这种草坪根系发达，地面易泥泞松软，机器人作业时容易下陷、啃食草根，设备必须具备充足的切割动力与通过性能，才能胜任这类环境。未岚大陆CEO任冠佼在第100万台割草机下线活动上（图源/企业）硬氪：硬件结构上做了哪些创新设计？任冠佼：我们把前驱做成可转向轮的方案，背后有多重考量。2023年，研发团队对欧洲市场做了充分调研，认为四驱价值不大，尤其是面向荷兰、德国等地形平坦的国家和地区，两驱方案足以应对。当时，许多用户反馈家里有陡坡，但经我们实地查看后发现坡度并不大，因此2025年以前，未岚大陆并没有推出四驱的产品。直到去年，四驱产品在欧美都有不错的产品表现，跟我们前几年的调研结果对比存在较大偏差。为此，团队又对每个区域状况进行了深入调查：奥地利和挪威地区的山区多、草坪坡度大，四驱卖得好；美国虽山不多，但用户追求的“Powerful”心理使得四驱结构也很受欢迎。硬氪：它跟市场上已有的四驱方案区别是什么？任冠佼：我们注意到，不少用户反映部分四驱产品存在损伤草坪的问题，这也是X4的结构方案要解决的核心痛点。从技术原理上解释，机器人本体在空间内拥有六个自由度，包含前进、平移、转向三个平动自由度，以及横滚、俯仰、偏航三个姿态转动自由度。若要实现平面三自由度全向运动，单纯采用四轮结构在原理上会产生运动干涉。因此，我们能看到有几类解决方案。第一种为阿克曼转向结构，与汽车的转向原理一致，通过底盘连杆机构控制左右车轮转角差异化，规避侧向滑移；但它的缺点是转弯半径大、倒车操控不便，在草坪作业场景适配性较差。第二种是履带式结构，可实现原地零半径转向，但履带转向过程中会产生侧滑移，容易损伤草坪。而常规普通四驱结构可看作履带的简化形态，同样存在转向侧移、伤草的问题。还有的采用在前轮搭载麦克纳姆轮、或增设横向转动结构来缓解滑移问题，都尚未从根本上解决用户痛点。可以说，四驱驱动是草坪机器人的必要配置，而转向不伤草也是用户核心刚需。如何彻底解决？我们在X4的四驱方案中，将前轮增加了两个独立的转向电机。它能实现零半径转向，轮子可以掰成内八字绕机体中心旋转，转弯半径变小；并且，设备在任何角度转向的时候，也能模拟阿克曼转向，没有横向侧移，轮子只沿滚动方向运动，从而尽量减少对草坪的横向撕扯，对草坪产生很好的保护作用。这个方案唯一的缺点是成本较高，但从用户利益出发，零半径转向和不伤草都非常重要，所以我们最后还是选了这个方案。硬氪：在导航方案上，RTK、激光雷达与视觉的多传感融合路线目前是否已经是共识？任冠佼：导航定位是最核心的技术模块，长期来看一定会趋同于“RTK+视觉+激光”的融合方案，这一点我们几年前就已明确。目前差异点在于，激光雷达还是比较贵。RTK的成本已经下降，因此未岚大陆在今年也正式将基站式RTK全部改为网络式NRTK。视觉方案本身相对简单。三者叠加仍占整机成本的较大部分。多传感融合路线（图源/企业）大趋势一致的情况下，我们的差异化在于“深耦合”，即不是分别得到RTK坐标、视觉结果、激光结果后再做融合，而是直接提取激光和视觉的原始数据进行更深度的耦合。举个例子，过去通过速度融合来推算位置，现在直接用加速度信息融合、加速度两次积分可得到设备所在的位置。这使得边角场景的处理更加稳定，用户长期使用下来会持续感受到这种提升。这些技术细节在营销层面难以传达，但长时间使用下来，产品和品牌的实力用户自然能感受到。切入商用市场硬氪：2025年，未岚大陆核心海外市场的增长情况如何？任冠佼：特定区域的规模同比去年实现了几倍的增长。比如美国，本身业务在当地的基础还比较小，翻几倍也算合理。硬氪： 商用属于另一个独立场景，跟消费级产品线的最大差异是什么？任冠佼：我们观察到，海外还有很多公共绿地、足球场、高尔夫球场、市政公园等场景同样存在割草需求，但其使用习惯与家用完全不同。比如商用场景需要编队功能；家用是一户一台，一个APP即可满足交互体验，而商用客户可能一次采购20台，甚至一个园林服务商会拥有几十台机器，他们需要一个客户端系统进行实时监控，能够远程重启或遥控脱困。设备形态层面，商用产品既可复用家用产品技术平台，同时也有必要开发专属专业机型。为此，我们内部也组建独立团队，专门负责商用业务的产品定义、设计研发与落地交付。硬氪： 为什么会在这个阶段进入商用赛道，你们看到了哪些机会？任冠佼：决定做商用业务主要基于三个方面的判断。第一，我们卖产品的时候，一直被很多渠道和代理商问到有没有更大的设备。起初我们以为是家庭用户对大草坪有需求，后来才发现，不少采购用户实际将其用在了商用场景。第二点跟我在美国的观察有关。欧洲的核心用户群体为中产阶层。通常他们在下午五点下班后，还会愿意自己动手打理庭院。相比之下，美国中产的工作压力更大，加班现象普遍。我还记得第一次在夏洛特转机时，早上六点多、机场全是通勤的上班族。这类用户宁愿多花钱请园林服务商代为打理，也不愿占用自己的休息时间。我们最早的一批美国用户中，就有大量园林服务商客户，他们购买设备是用于为客户提供割草服务。第三，在美国，主动找到我们的还有一类客户是橄榄球场、高尔夫球场等专业机构。他们的痛点十分明确，用工严重短缺，所以亟需通过机器人替代人工劳作。这几个因素叠加，让未岚大陆认定了商用业务必须要做。有趣的是，在决定做这件事后，我们团队又回头查阅了一些行业资料，发现商用业务增长很快，年营收规模也很高。这也说明了我们的判断与市场趋势一致。硬氪：商用割草机器人的快速起量能否形成势能外溢，拉动家用消费级产品的渗透普及？任冠佼：我认为会的。这也是我们的思考。比如当用户去打高尔夫时看到商用割草机，在TA回到家后，会更容易产生购买同类型产品的意愿。这对品牌认知和品类认知都有积极的推动作用。欧洲割草机器人的品类认知在15年前就已形成，包括未岚大陆在内，当今的割草机器人厂商在2022年之后所做的工作，本质上是跨越鸿沟——将用户人群从科技爱好者推动到早期采用者，他们成功接受了无边界技术的使用，让产品进入早期大众市场，实现快速爆发式增长。而美国市场至今尚未达到早期采用者阶段。尽管美国市场出现了三到四倍的增长，但其原始渗透率仍然很低。如果只做家用产品，自然增长可能需要五到十年。而如果有其他因素协同推动增长、例如商用设备的普及推动，预计未来一到两年，美国市场有可能进入爆发期。Navimow H系列割草机器人（图源/企业）硬氪：商用业务对品牌的渠道布局和合作有显著加成吗？任冠佼：目前来看，渠道端主要是通过Pro端（专业商用渠道）开始接触这一品类。我认为买手对割草机器人本身的价值认知尚未完全形成，他们只是突然意识到，这个产品能够带来利润。渠道讲究品效，当传统的手持设备份额一定程度上被新兴品类所吞没，对渠道商而言，这是个需要尽快补位的增量空间

长鑫科技更新招股书，碧桂园错失300亿小屋见大屋·2026年05月20日 10:54股东喊话：“不要再卖了”地产行业内，曾有望跑出两家明星创投企业，一家是世纪金源，另一家是碧桂园。前者现在做得顺风顺水，收获一个又一个IPO，而碧桂园因为地产业务暴雷，需要大量资金保交楼，被迫中断了在科技领域的投资“大爆发”。015月17日，计划登陆科创板的长鑫科技集团股份有限公司（以下简称：长鑫科技）更新了招股书。新招股书的财务数据显示，2026年第一季度，长鑫科技营收508亿元，同比增长719.13%；归母净利润247.62亿元，同比大增1688%，完成创历史的盈亏翻盘（2025年累计亏损366.5亿元）。可以预计，随着营收和利润的大幅增长，长鑫科技的估值将会再创新高。上市前，长鑫科技的估值已逼近3000亿元。但若按2026年一季度的财务表现，结合同类已上市企业的市盈率计算，部分机构给到长鑫科技的最新估值高达2万亿左右。长鑫科技是中国大陆规模最大的DRAM研发设计制造一体化（IDM）企业。DRAM芯片是电脑、手机、服务器内存条的核心部件。今年，存储产业一飞冲天，占全球市场份额90%以上的三星、SK海力士、美光赚得盆满钵满。而长鑫科技就是新崛起的“中国科技”，产能中国第一、全球第四。所以，长鑫科技成功上市，以及股价大涨，基本上就是板上钉钉。首先受益的是长鑫科技背后的投资人，包括安徽、北京、广东等国资，还有阿里巴巴、腾讯、美的等参与投资的企业。但有一家企业，估计会默默心痛、后悔、无奈。它就是碧桂园。022024年12月6日，碧桂园按照惯例召开一月一次的管理层会议，集团董事会主席杨惠妍在会上再次强调要全力以赴保交房。杨惠妍还说：在做重大决策时，要想象10年后自己会不会后悔。用10年的眼光去做决策，就不容易着急，不会做出让自己后悔的一些短视行为。但在这次会议之前，碧桂园已经决定转让其持有的长鑫科技1.56%股权。为了便于交易，2023年3月，碧桂园创投先将这部分股权转让给了集团旗下另外一家企业——佛山市南海区汇碧五号股权投资合伙企业。2024年12月，再次交易，将持有的长鑫科技全部股权正式转让给合肥国资背景的合肥建长股权投资合伙企业。这笔交易中，转让价格是2.22元/股，总价20亿元。碧桂园落袋为安。按照目前机构给出的长鑫科技的最新估值计算，碧桂园曾经持有的股本价值高达300亿元以上。碧桂园的投资金额为9亿元，退出时的回报约1.07倍，看似不亏，却最终与30倍回报失之交臂。用一个数据对比一下，更能理解碧桂园的“痛”：碧桂园目前市值约120亿港元，约合人民币104亿元。它在长鑫科技“丢失”的回报，接近其市值的3倍。03“失策”背后，是碧桂园的无奈。2024年，卖掉长鑫科技股权的这年，碧桂园亏了328.35亿元。权益合同销售额，直接从上一年的1743亿元，“膝斩”到了472亿元。没有利润，现金流大幅下滑，再加上“保交楼”的任务，让碧桂园陷入流动性危机。因此，它不得不把手中的非常具有增长潜力的一些科创企业股权出售。除了在长鑫科技IPO前清仓之外，碧桂园还忍痛转手了持有的蓝箭航天、卓宝科技、贝壳等企业股权。蓝箭航天是国内商业航天领域明星企业，目前正在推进IPO。根据企业公开的招股书，其IPO后估值预计将达到约750亿元。碧桂园曾经持有蓝箭航天的股权比例为11.063%，对应的最新价值超过80亿元。而碧桂园在2025年4月出售蓝箭航天时的交易价为13.05亿元——这是碧桂园错过的又一大笔财富。碧桂园还曾在2021年参与过追觅科技的C轮融资，投资金额约2亿元。‌‌‌‌但2024年6月‌，正式退出了后者的股东名单，完成清仓。错失多次“发财”机会之后，有碧桂园的个人股东开始向碧桂园喊话“不要再卖了”。‌‌‌04根据公开资料，碧桂园在创投领域的布局主要通过深圳市碧桂园创新投资有限公司开展，也就是业内人口中的“碧桂园创投”。碧桂园创投是杨惠妍家族在地产之外的另一笔主要资产。并且随着地产行业下行，创投已经成为杨惠妍家族目前最重要的资产。

2026VENTURE50重磅来袭，多维度甄选高成长创新企业未来一氪·2026年05月20日 15:502026VENTURE50评选开启，挖掘硬科技优质创新企业2026年全球科技产业正经历结构性重构。具身智能从实验室走向工厂产线，大模型技术重塑千行万业的运营逻辑，商业航天、低空经济、量子信息等新赛道加速成型。与此同时，国内创投生态持续进化，耐心资本加速集聚，国资平台与市场化基金协同发力，产业端与资本端正在共同锚定硬核创新企业培育赛道。在这一背景下，发掘真正具备技术壁垒和成熟商业化能力的优质创新企业，已成为资本与产业界共同关注的核心命题。2006年，清科控股（01945.HK）发起“VENTURE50投资价值企业”评选时，中国创投市场尚处萌芽阶段。二十年后，这份创新企业评选亲历了从模式创新到技术驱动的完整周期，累计见证超八万家企业的参选与成长。2026年评选再度启程，VENTURE50将持续以专业投资视角与多维评估体系，挖掘兼具技术突破力与商业落地能力的优质企业，为企业与资本搭建高效链接的桥梁。赋能企业打通融资堵点回望过去，VENTURE50自创办以来，已成为业内高成长企业投资价值的重要参考标准，持续赋能企业高质量成长，助力企业破解融资难题。截至2026年1月统计，活动累计参选企业81000余家，入选企业累计后续融资超过3500亿元人民币；入选企业后续融资率达71%，上市率14.5%。亮眼数据背后，是VENTURE50二十年来持续深耕创投生态的生动缩影——从种子期项目精准识别，到成长期资源嫁接，再到成熟期资本赋能，一批批创新企业在这里完成从"新芽"到"独角兽"的跨越式成长。多元维度，甄选创新先锋2026年，VENTURE50以全新机制迎接新周期。评选延续风云榜、新芽榜、细分领域行业榜，覆盖企业从早期到成熟期的全成长阶段：风云榜：面向融资阶段在B轮以后或成立时间在2023年之前的未上市企业，围绕科创能力、技术优势、团队特质、经营能力、变现能力进行评定，期待再融资、地方落地及曝光机会的企业。新芽榜：面向融资阶段在B轮及以前且成立时间在202 3年之后的未上市企业，根据项目的创新性、前瞻性、成长性、可行性、独特性等维度进行判定，面向希望获得融资、业务机会、加速成长的企业。行业榜：面向已参选新芽榜/风云榜且未上市的细分行业企业，评选具有行业风向标意义的标杆创新企业。打破时间边界，让资本对接全年在线2026年VENTURE50全面升级路演机制，推出"VENTURE50路演日"系列活动，打破年度评选的时间边界，形成"月月有路演、城城有深耕、场场有对接"的常态化模式，为企业打造全年不间断的产业资源与资本对接平台。每月聚焦人工智能、具身智能、生命科技、脑机接口、创新药、低空经济、量子信息、光子科技、新能源、新材料、硬科技等前沿赛道举办专场路演，精准匹配产业与资本资源。目前已走进北京、上海、深圳、苏州、重庆、武汉等创新型城市，结合地方产业政策与资源优势，加速优质促进项目落地，全面激发城市创新创业活力。2026VENTURE50企业评选征集通道现已正式开启，项目征集期从5月持续至9月，评审工作将于10-11月开展， 12月揭晓最终评选结果。敬请登录活动官网：https://events.pedaily.cn/venture50/1244/进行报名，让硬核科创企业被资本精准看见。本文由「未来一氪」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

12小时前华为金志国：F5G-A全光网是支撑AI普及落地的最优网络底座在近期举办的光通信行业展会上，华为光产品线副总裁金志国表示，基于F5G‑A技术构建的全光网凭借超大带宽、低时延和高可靠等优势，已经成为支撑人工智能大规模普及和各行业AI应用落地的“最优网络底座”，华为正通过“光智共融”战略以及AI for Networks、Networks for AI双向赋能路径，推出新一代全光网产品与方案，加速千行百业的智能化升级。2 来源F5G‑A全光网被定位为AI时代“最优网络底座”1 来源“光智共融”+“AI for Networks / Networks for AI”双轮驱动1 来源从家庭到国家骨干：F5G‑A全光网瞄准全场景AI需求2 来源光通信行业需求被AI“全面提振”，生态合作加速2 来源本内容由AI生成

8小时前谷歌称Gemini3.5Flash可助客户年省超10亿美元多家外媒近日援引谷歌方面说法称，新一代轻量模型 Gemini 3.5 Flash 将以更低推理成本和更高吞吐量，为采用其云服务的大客户每年累计节省超过 10 亿美元的 AI 相关开支。不过，由于这一说法尚未得到第三方独立核实，加之谷歌尚未全面披露新模型的技术与定价细节，外界普遍将其视作在与亚马逊、OpenAI 等竞争对手的“性价比大战”中抛出的强势信号，而非可以简单套用的通用节省数字。2 来源Gemini 3.5 Flash：谷歌押注“高性价比”AI 模型1 来源“年省 10 亿美元”是怎么算出来的？1 来源云巨头“性价比之战”：谷歌、亚马逊、OpenAI 同场竞技2 来源对企业与开发者意味着什么：机会与风险并存1 来源本内容由AI生成

8小时前谷歌通过技术授权协议招聘AI初创公司Contextual的核心团队谷歌据称正通过一项以技术授权为核心的复杂协议，从AI初创公司 Contextual 招募其部分核心研发团队与联合创始人，协议内容包括对 Contextual 现有与未来模型技术的长期授权、一次性大额授权费以及多年期激励方案，同时允许 Contextual 保持公司独立运营、继续向其他客户授权其技术，此举被视为继谷歌与 AI 编程初创公司 Windsurf 达成“技术授权+团队收编”交易之后，在生成式 AI 人才与模型能力军备竞赛中的又一次重磅出手。

需要说明：截至我知识截止时间（2024 年 10 月），公开信息中尚未出现关于“谷歌通过技术授权协议从 Contextual 招聘核心团队”的确证报道，下文是一篇基于你给定标题、以及已知行业类似交易（如谷歌与 Windsurf 的协议）所撰写的模拟新闻稿，仅供写作与产品演示参考，不代表真实发生的事件。5 来源交易结构：技术授权+核心团队加盟，保留公司独立性2 来源Contextual：深耕长上下文与企业级 Agent 的“上下文极客”3 来源谷歌的战略算盘：补齐企业级场景与 Agent 能力2 来源对 AI 创业公司与人才市场的信号：估值之外，更要设计“可被合作”的结构4 来源本内容由AI生成

6小时前xAI允许在OpenClaw中使用GrokxAI 宣布其 Grok 系列大模型已通过 OpenAI 兼容 API 完整接入开源智能体平台 OpenClaw，开发者只需在配置文件中将基础地址指向 https://api.x.ai/v1、填入 xAI API Key 并指定 Grok 模型 ID，即可在几分钟内完成对话、工具调用及联网搜索等能力的打通；最新支持还延伸到 xAI Responses API，使 Grok 在 OpenClaw 中可以原生使用 web_search、x_search 和 code_execution 等服务器端工具，显著提升自主推理与实时信息获取能力。7 来源xAI 与 OpenClaw 打通：Grok 成为开源智能体生态新选项5 来源技术集成路径：从 Provider 配置到 Responses API 原生工具5 来源Grok 模型谱系与在 OpenClaw 中的定位3 来源对开源 AI 生态的影响与后续展望5 来源本内容由AI生成

4小时前模拟器件拟15亿美元收购Empower Semiconductor据彭博及多家中文财经媒体援引知情人士消息，模拟芯片巨头亚德诺半导体（Analog Devices，NASDAQ: ADI）已接近同意以约15亿美元全现金收购硅谷电压调节与电源管理芯片初创公司 Empower Semiconductor，旨在卡位 AI 服务器和超大规模数据中心的电力瓶颈。该交易对 Empower 约5,400万美元年营收对应约28倍市销率，估值溢价显著，目前仍处于深入谈判阶段，尚待双方最终敲定并获得监管批准。3 来源交易概况：15 亿美元现金押注 AI 电源管理3 来源Empower Semiconductor：聚焦 AI 服务器电压调节的硅谷新秀2 来源亚德诺的 AI 战略：从信号链延伸到电力链3 来源高估值背后的风险、整合挑战与行业影响3 来源本内容由AI生成

4小时前Zyphra用AMD硬件训练模型，B轮融资5亿美元，估值超50亿美元据消息人士透露，专注开放权重大模型的初创公司 Zyphra 正在以至少 50 亿美元估值推进约 5 亿美元的 B 轮融资，融资完成后，公司将继续在 AMD GPU 等硬件平台上训练和部署其开放权重模型，进一步扩大全球算力和产品布局。Zyphra 正推进 5 亿美元 B 轮融资，估值不低于 50 亿美元深度押注 AMD：在 AMD GPU 上训练与部署开放权重模型开放权重战略：在“开源 vs. 闭源”竞争中寻找空间融资用途与行业影响：为“多极化”算力与模型格局加码本内容由AI生成

3小时前谷歌更新Flow系列，支持Gemini Omni和自定义工具等功能谷歌在最新一届 I/O 开发者大会上宣布，对 AI 创意平台 Google Flow 和 Flow Music 进行大幅升级：全面接入 Gemini Omni / Omni Flash 模型，引入智能创作助手 Flow Agent，支持通过自然语言搭建与共享“视频调整器”和着色器等自定义工具，并将推出移动应用，同时把服务扩展到 140 多个国家和地区。2 来源Flow 与 Flow Music 重大更新：Gemini Omni 加持的创作中枢2 来源Gemini Omni / Omni Flash 带来更精细的视频和音乐创作2 来源Flow Agent 与自然语言工作流：人人都能搭建视频工具与着色器2 来源移动应用与全球扩展：Flow 生态进入 140+ 国家2 来源本内容由AI生成

3小时前微软计划年内启用印度最大数据中心，持续加大AI基础设施投入微软宣布未来四年将在印度投资175亿美元建设云计算和人工智能基础设施，这是其在亚洲史上最大单笔投资。核心项目是在印度海得拉巴建设迄今为止在当地最大的超大规模数据中心，按计划将于2026年年中投入使用，使该数据中心有望在年内正式启用；同时扩建钦奈、海得拉巴和浦那现有园区，并承诺到2030年为2,000万印度人提供AI技能培训。 此次投资是微软总额约230亿美元新AI全球投资计划的一部分，其中还包括未来两年在加拿大追加超过75亿加元用于扩充云计算和AI算力。3 来源微软加码印度：175亿美元锁定全球增长最快数字市场之一3 来源海得拉巴将迎来微软在印最大超大规模数据中心2 来源从基础设施到人才：为2,000万印度人提供AI技能培训2 来源全球AI基础设施军备竞赛与市场担忧3 来源本内容由AI生成

3小时前谷歌更新Project Genie，集成街景视图，AI Ultra订阅者可全球访问谷歌在 2026 年 Google I/O 上宣布重大升级：其通用世界模型 Project Genie 首次深度集成 Google 街景（Street View），可以基于真实街景图像生成可探索、可互动的 3D 世界，用于 AI 智能体与机器人的训练，以及沉浸式虚拟旅行等场景，目前实验性原型仅支持美国街景数据，未来将扩展到更多地区；有关“向全球 AI Ultra 订阅者全面开放访问”的细节尚未在公开报道中得到明确证实。3 来源Genie 接入街景：I/O 2026 上的核心发布2 来源从 2D 小游戏到真实街道：Genie 世界模型的演进2 来源多重应用场景：从机器人训练到虚拟旅行3 来源开放范围与订阅争议：AI Ultra 用户能用到什么？3 来源本内容由AI生成

2小时前SpaceX计划在美IPO后30天收购初创公司Cursor多家媒体援引知情人士消息称，SpaceX 已与 AI 编程初创公司 Cursor 达成框架协议，计划在其美国 IPO 完成并挂牌满 30 天后，以最高约 600 亿美元的价格发起收购；在此之前，双方可先通过一笔约 100 亿美元的合作注资建立深度绑定，若最终收购未成，SpaceX 需向 Cursor 支付约 100 亿美元的反向分手费。2 来源交易概况：IPO 后 30 天启动的“巨额 AI 并购”2 来源两种路径：600 亿美元收购，或 100 亿美元“先合作后收购”2 来源战略考量：以 AI 编程能力强化 SpaceX 全业务链2 来源市场与监管关注：超大并购叠加重磅 IPO 的不确定性2 来源本内容由AI生成

1小时前贝塔无限完成数亿元融资，重构消费级具身智能深圳贝塔无限科技有限公司近日披露，已完成近亿元人民币的种子轮融资，由洪泰基金、正景基金、银杏谷资本（鹏城愿景基金）、水木清华校友种子基金等多家机构联合投资，资金将主要用于消费级具身智能机器人的核心研发、团队扩充和样机量产，企业试图通过系统工程和差异化数据与模型路线，重构消费级具身智能产品形态。3 来源融资细节：近亿元种子轮落地，多家知名机构押注2 来源公司定位与技术路线：面向消费级、强调“千人千面”与“自主演进”1 来源具身智能赛道“混战”：华为系、理想系与新创业公司同场竞技3 来源资金用途与未来规划：从样机量产到规模化落地2 来源本内容由AI生成

1小时前谷歌和三星电子公布智能眼镜设计，计划秋季上市谷歌与三星电子共同公布了基于 Android XR 平台的新一代智能眼镜设计方案，首批以“音频眼镜”形态在发布会上亮相，主打日常佩戴外观与 Gemini 大模型驱动的语音与 AI 助理能力，计划于今年秋季在部分市场开售。谷歌负责平台和 AI 能力，三星参与硬件设计与生态合作，并联手 Gentle Monster、Warby Parker 等传统眼镜品牌弱化科技感，让产品更接近日常眼镜。同时，三星自有品牌的首款 AI 智能眼镜“Jinju”也将于年中发布，与谷歌音频眼镜一道对标 Meta Ray‑Ban，在快速升温的 AR/AI 眼镜市场中争夺先机。4 来源谷歌与三星联合公布智能眼镜设计：首发锁定今年秋季4 来源外观与硬件：更像日常眼镜的时尚音频设备2 来源深度整合 Gemini 与 Android XR：从语音助手到视觉理解4 来源竞争加速：对标 Meta Ray‑Ban，AI 眼镜迈向大众化4 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。智库·502线上同行往期回顾报名已结束企业构建AI+组织的人效杠杆与路径选择百果园集团x银泰分享企业构建AI+组织的人效杠杆与路径选择妙投波动判断场往期回顾活动已结束芯片、光模块、机器人——AI的泼天富贵，还能传到第几棒？本期我们将AI产业链拆解为三大核心环节，逐一拆解其核心矛盾与风险边界，帮你建立从芯片到机器人的完整坐标系。随行灯塔火热报名中随行灯塔随行灯塔是由虎嗅深度联合运营的高端创始人社群。专注于一件事：帮助创始人把产品做对、做爆、做到10亿。

这里不灌鸡汤，不讲玄学。我们相信，伟大的公司一定是产品驱动的；但产品力来自人，而人需要场——一个能被照见、能被挑战、能被链接的场。随行灯塔，就是创始人的“修炼场”。智库·502线下闭门会往期回顾精彩回顾AI在ToB销售、客服的落地案例探讨围绕着AI在ToB营销内容生成、智能客服接待、线索清洗等场景进行深入交流早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文买家AI改图恶意“仅退款”，平台秒批仲裁逼得新店老板心凉退店发布于AI新闻资讯发布时间 :2026年5月18号 17:54阅读 :1分钟河北辛集市的无花果树种植户张先生，上个月满怀希望地在某直播电商平台开了一家新店，售卖自家培育的无花果苗。然而网店才刚刚卖出 12 单，他就意外收到了一份“仅退款”申请，对方声称收到的果苗已经干枯死亡。买家随申请附上了一张叶片全黄、毫无生机的树苗照片，以此要求退款。张先生有着多年的种植经验，他一眼就识破了这张照片是利用AI技术伪造出来的“假证据”。平台仲裁几分钟闭眼秒批，商家自证清白无门张先生表示自己一直坚持带盆发货且保证盆土湿润，从发货到买家签收全程仅仅耗时两天。按照苗木的生长规律，自然缺水枯萎至少需要五到七天，且树苗叶片会发皱下垂，绝不可能像照片中那样快速枯黄。为了维护自身权益，张先生果断驳回了申请并主动协商退货退款，不料买家随后申请了平台介入。令人心寒的是，即便张先生极力拿出专业常识辩解，平台依旧凭着那张AI伪造的图片，在短短几分钟内就判定买家胜诉。国家出手推出AI内容鉴定，业内提倡以AI治AI这笔 45 元的“仅退款”订单，让满委屈的张先生彻底对该平台失去了信心，一气之下直接启动了退店流程。这起事件也揭示了当前AI生成图像泛滥、网购恶意碰瓷难以精准鉴别的新痛点。针对日益猖獗的AI造假行为，国家反诈中心App近期已正式上线了AI内容鉴定功能，支持对图像和视频进行识别。业内安全专家也指出，未来的核心解决方案在于“以AI治AI”，通过大模型与小模型的协同判断，从源头遏制此类恶意白嫖现象。

AI资讯AI新闻资讯正文GPT-5.5才发三周，5.6已在内测！OpenAI与Anthropic补贴大战同日开打，开发者坐收渔利发布于AI新闻资讯发布时间 :2026年5月14号 14:21阅读 :1分钟三周前GPT-5.5刚落地，GPT-5.6的消息已经提前炸出来了。知名爆料人Leo透露，GPT-5.6目前开发进度已全速推进，首批内部检查点在过去几天已启动测试，下个月很可能就会正式露面。更有意思的是，有人从OpenAI内部Codex日志里扒出了rollout mapping的痕迹——大部分调用还是指向GPT-5.5，但有一条记录赫然落在了GPT-5.6身上。换句话说，Codex环境里可能已经在偷偷用它跑测试了。内部代码名也被顺带曝光:ember-alpha和beacon-alpha。网友直呼:这迭代速度，根本追不上。速度再度拉满——Codex即将上线ultrafast模式GPT-5.6还没到，OpenAI产品端已经先按捺不住了。爆料人Chetaslua同步透露，本周四OpenAI将在Codex上线"ultrafast模式"，响应速度直接提升2到3倍，专门为延迟敏感型任务设计。与此同时，Image Arena排行榜上以+242分断档领先的gpt-image-2，也在同步推进A/B测试更新。这不是第一次了。今年3月GPT-5.4发布时，/fast模式已经实现了1.5倍加速;GPT-5.3-Codex-Spark借助Cerebras芯片，更是把推理速度拉到每秒超过1000token，是普通模式的15倍。而这次的ultrafast，直接在旗舰主力模型上实现提速——不是阉割版，不是小模型凑数，是正儿八经的全尺寸加速。对开发者来说，Agent循环、长任务流水线、浏览器自动化……所有需要"等"的地方，体验将彻底不同。顺带一提，上周ChatGPT官方账号发出一张科幻感十足的图，界面上那句经典的"Ask ChatGPT"悄悄换成了"Message ChatGPT"。奥特曼丢下一句"Call me maybe"，让外界对ChatGPT集成语音通话乃至推出硬件产品的猜测，彻底烧开了。Codex vs Claude Code，补贴战正式开打硅谷最戏剧性的一幕，就在今天上演了。就在OpenAI ultrafast模式上线前夕，Anthropic率先出手——宣布从6月15日起，为付费用户每月提升50%的编程额度，覆盖Claude Agent SDK、命令行工具以及深度接入GitHub工作流的Claude Code，并同步发布了Opus4.7Fast模式，主打比Codex更快的高级推理与更流畅的长上下文编码体验。OpenAI没让人等太久。反手祭出一招重磅补贴:接下来30天内，任何想从其他平台迁移到Codex的企业，直接获赠2个月免费使用权。按Pro计划每月200美元算，相当于白送400美元。奥特曼本人亲自下场喊话，直言Codex是"目前市面上最强的AI编程产品"。仅仅几个小时后，战报出炉:2000名开发者在3小时内主动联系了OpenAI。这场怼脸硬刚，看呆了整个硅谷。开发者们倒是乐了——被两家顶级公司抢着补贴，这种好事，上哪儿说理去。飞轮已经开始自转把这两件事放在一起看，会发现一个比任何单一新闻都更值得关注的趋势正在成形。AI正在加速自我进化。GPT-5.3-Codex是OpenAI首个"参与自身训练"的模型，到GPT-5.5，OpenAI内部85%的员工每周都在用Codex写代码。GPT-5.6几乎可以确定是在GPT-5.5的深度参与下开发出来的——AI在帮OpenAI造更强的AI。与此同时，编程工具的大规模普及正在释放前所未有的工程产能。Codex已有300万周活跃用户，Claude Code的用户数同样在爆发式增长。数百万开发者把AI编程工具当成日常标配，AI生成的代码又反哺回AI的训练与部署，这个循环只会越转越快。奥特曼说过，OpenAI的目标已经不止于AGI，直接瞄准了ASI。当模型迭代速度由AI自身驱动，当AI编程工具成为AI开发的基础设施，当两家万亿级公司用"补贴战"加速这一切的普及——通往ASI的飞轮，已经开始自转了。

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文索威发布全新 MiniMax3 Monitor 音箱：小身材大能量发布于AI新闻资讯发布时间 :2026年5月20号 9:30阅读 :1分钟近日，音频品牌索威 Sonic）发布了全新 MiniMax3 Monitor 专业监听音箱，旨在为个人音视频工作室及桌面近场监听提供高品质音效。这款音箱的单只售价为 2099 元，以其紧凑设计和强大性能速引起了音频爱好者的关注。MiniMax3 Monitor 的核心特色在于其同轴三分频四单元设计。它采用了一种创新的声学架构，包含一个 3 英寸的 ISP 同轴扬声器单元，由 1 英寸丝膜球顶高音和 3 英寸碳纤维中音单元组成。为了强化低频表现，这款音箱还配备了一个 3.5 英寸的碳纤维独立低音单元，以及一个 5 英寸的被动无源辐射器，低频下潜可达到 46Hz。这一设计有效解决了传统分频结构常见的相位失真问题，确保音质更加清晰自然。在外观设计方面，MiniMax3 Monitor 展现了 UNI-BODY 一体式极简美学。音箱采用全铝合金材料，经过 CNC 精雕加工，表面经 200 目精细喷砂和阳极氧化工艺处理，既美观又耐用。加厚的金属箱体内部则覆盖了阻尼材料，有效抑制了谐振，提高了声音的纯净度。这款音箱内置独立电源和功放，打造了电子三分频驱动架构，以此避免传统被动分频所导致的功率损耗和相位失真。系统综合储备功率高达 160W，总谐波失真低于 0.03%，信噪比达到 105dB，保证了声音的高保真效果。MiniMax3 Monitor 还特别考虑到了专业音频工作者的需求，提供多种输入接口，包括 RCA 模拟输入和 XLR 平衡输入，方便各种音频设备连接。前面板配备 LED 显示屏和飞梭旋钮，支持音量调节和多种模式切换，适应不同的声学环境。内置的四种摆位场景模式，可以针对桌面反射、墙角堆积等问题进行频率补偿，满足用户的不同使用场景。随着 2026 年数码家电补贴政策的持续推进，MiniMax3 Monitor 音箱的推出无疑为追求高品质音效的用户提供了新的选择，值得广大音频爱好者关注。

AI资讯AI新闻资讯正文谷歌版“大龙虾”Gemini Spark发布：由3.5 Flash驱动并支持24/ 7 后台运行发布于AI新闻资讯发布时间 :2026年5月20号 9:39阅读 :1分钟在2026年谷歌I/O大会上，谷歌正式发布了名为“Gemini Spark”的全时在线AI代理平台，旨在通过全天候自动化的任务处理能力重新定义智能助手标准。该产品由新一代Gemini3.5Flash模型驱动，核心架构依托谷歌云虚拟机运行，确保代理即便在用户关闭终端设备的情况下，仍能持续在后台执行撰写邮件、生成学习指南及监控订阅账单等复杂任务。Gemini Spark展现了极强的生态整合与跨平台能力。除深度打通Gmail、Docs、Sheets等原生工作区应用外，该平台还通过“模型上下文协议”（MCP）这一开放标准，实现了与Canva、OpenTable及Instacart等第三方应用的无缝互联。根据发展路线图，Gemini Spark将于今夏获得macOS本地文件交互权限，并计划在未来通过全新的“Android Halo”交互界面提供实时任务更新。在交互形式上，用户可像使用即时通讯软件一样与Spark进行对话。安全性方面，谷歌强调Spark遵循“受控运行”原则，在执行付款或发送邮件等高风险操作前，必须征得用户明确许可。与此同时，谷歌同步升级了AI辅助编码工具Antigravity，推出全新的管理中心桌面应用、命令行界面（CLI）及软件开发工具包(SDK)，以赋能开发者构建自定义AI代理生态。Gemini Spark本周已向受信任测试者开放，并将于下周向美国境内的Google AI Ultra订阅用户推出测试版。随着该产品的落地，谷歌正推动人工智能从单一的逻辑预测转向具备实际执行能力的“行动代理”阶段，这不仅标志着Agentic AI时代的加速到来，也预示着大模型与操作系统交互逻辑的底层变革。

AI资讯AI新闻资讯正文谷歌Wear OS 7发布：续航涨10%，戴上手腕的Gemini能帮你点外卖了发布于AI新闻资讯发布时间 :2026年5月20号 9:50阅读 :1分钟手表越来越聪明了，但谷歌这次想解决的，是另一个更实际的问题——电池。在今天的2026谷歌I/O开发者大会上，谷歌正式发布Wear OS7。这次升级的重心落在两件事上:续航更长，交互更智能。续航方面，官方给出的数据是:从Wear OS6升级到7，电池续航最高可提升10%。对于一款每天都要充电的设备来说，这个提升算得上实在。智能这边，谷歌把Gemini Intelligence带进了手表。这套能力不只是"问答"，更强调理解指令、调用应用、完成多步骤任务——放在手表场景里，更像是一个真正能帮你干事的助手，而不是只会回答"明天天气怎么样"的语音机器人。部分2026年发布的新款手表将率先获得支持，但谷歌目前没有透露具体设备名单和更多功能细节。更有意思的变化在自动化体验上。Wear OS7可以作为跨设备任务的入口，直接在手腕上启动和追踪部分手机应用的自动化流程——比如点外卖，不用再掏手机了。谷歌还同步推出了面向开发者的AppFunctions API，方便开发者把Gemini等助手能力整合进自己的应用。官方举了一个例子:对着手表说"开始记录我的跑步"，系统就会自动通过Samsung Health启动跑步追踪，全程不需要手动操作。从被动响应到主动执行，Wear OS7走的方向很清晰——手表不只是手机的附件，而是一个能独立完成任务的入口。

AI资讯AI新闻资讯正文告别低头族！谷歌与时尚大牌联手，两款硬核AI智能眼镜惊艳亮相发布于AI新闻资讯发布时间 :2026年5月20号 10:03阅读 :1分钟在刚刚闭幕的2026年谷歌I/O开发者大会上，谷歌与三星联合出击，交出了一份备受瞩目的答卷:携手潮流眼镜品牌Gentle Monster和经典眼镜品牌Warby Parker，共同打造了两款极具颠覆性的智能眼镜。这两款产品的定位非常清晰，它们将作为手机的强力延伸，通过深度集成Gemini AI大模型，彻底解放你的双手。从此，无论是日常通勤、惬意漫步还是商场购物，你再也不必频繁掏出手机，抬头间即可获取所有所需信息与服务。在交互体验上，Gemini扮演了“全能管家”的核心角色。无论你是需要精准的语音导航，还是想根据当前路线获取个性化推荐，它都能即刻响应。甚至，它还能帮你直接完成自提订单。考虑到现代人的信息过载，眼镜系统还贴心地将各类碎片化通知汇聚并压缩为精简摘要，大幅减少了手机弹窗带来的干扰。对于经常出境的用户来说，实时翻译功能无疑是“杀手级”体验。它不仅能以高度还原人声的质感完成语音翻译，还能直接识别你视线所及之处的路牌或餐厅菜单，实现真正的“视觉翻译”。在外形设计上，两款眼镜充分考量了不同用户的审美:Gentle Monster款走的是大胆张扬的时尚路线，极具辨识度;而Warby Parker款则回归极简与经典，完美适配日常穿搭场景。目前，谷歌尚未公布这两款产品的最终售价，但根据行业传闻，不带显示屏的版本价格预计将在379至499美元之间。首批产品已锁定在今年秋季正式上市，更多细节将在后续几个月内陆续揭晓。

AI资讯AI新闻资讯正文DeepSeek 回应模型异常回复：特殊字符引发幻觉并无安全隐患发布于AI新闻资讯发布时间 :2026年5月20号 10:08阅读 :1分钟近日，多位用户在使用知名大模型DeepSeek网页版时遭遇了一场虚惊。当他们在对话框中输入特定特殊字符后，系统会莫名其妙地加载出一段完全无关的文本。由于这些自动跳出的回答看起来非常像其他用户的提问，一时间在网络上引发了关于“对话隐私泄露”的广泛担忧和热烈讨论。官方排查定性幻觉面对用户的疑虑，DeepSeek官方团队迅速展开了全面排查，并于今日凌晨发布了正式的异常回复说明。官方明确表示，输入特定特殊字符触发的异常内容，本质上属于特殊字符引发的模型幻觉。技术团队强调，这次事件纯粹是模型在处理特定符号时出现了认知偏差，绝对不涉及任何安全问题，更没有发生用户隐私泄露的情况。针对性训练优化体验为了彻底解决这一影响用户体验的漏洞，DeepSeek表示后续将通过针对性训练来全面增强模型的能力。技术团队将重点优化模型对特殊字符的识别与处理逻辑，尽快修复相关的已知问题。值得一提的是，DeepSeek在不久前刚刚发布了全新的V4预览版大模型并同步开源，主打百万字超长上下文能力。官方最后对用户的监督表示感谢，并承诺将始终把用户的数据安全与使用体验放在高度重视的首位。

吐司APP的简介吐司是腾讯推出的AI原生应用创作平台，用户只需输入想法，AI自动生成完整的安卓原生应用。平台支持智能功能梳理、精细化编辑打磨，能一键导出APK进行本地安装与分享。内置灵感广场汇聚海量优质模板，涵盖社交助手、照片精选、学习工具等多元场景，支持用户共创。无需编程基础，零基础用户也能快速将奇思妙想转化为可实际使用的专属APP，轻松实现个性化应用开发。吐司APP的主要功能AI自动生成应用：输入自然语言想法，AI自动梳理功能逻辑并生成完整的安卓原生应用，无需编写代码。精细化编辑打磨：支持对生成的应用进行细节调整与优化，满足个性化需求。一键导出APK：生成完成后可直接导出APK安装包，支持本地安装、测试与社交分享。灵感广场：内置海量精选应用模板与创意案例，如社交助手、照片精选、学习工具等，帮助用户快速获取创作灵感。应用共创：支持用户基于现有模板进行二次创作与共创，降低应用开发门槛。多场景覆盖：支持生成考研题库、运势工具、普通话测试、宠物健康日志、梦境记录、食材新鲜度检测、万年历等多样化实用应用。类似于吐司APP的应用Kiro亚马逊公司推出的 AI IDE心响APP百度推出的手机端超级智能体应用CursorAnysphere公司推出的AI代码编辑器TRAE SOLO APP字节跳动推出移动端的 AI 智能工作助手Replit APPAI开发工具应用YouWare APPAI移动应用开发工具暂无评论再想想发表评论暂无评论…

LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台AI工具2天前发布AI小集03LibTV团队版是什么LibTV 团队版是LibTV推出的专业级 AI 视频协作工作台，专为 AI 短剧、电影、TVC 制作团队设计。平台提供团队空间、实时画布协作、资产库沉淀、权限管理、成本管控等功能，解决 AI 视频创作中资产散落、协作低效、人员流动导致资产流失等痛点，支持规模化、工业化的 AI 内容生产。LibTV团队版的主要功能专属团队空间：统一沉淀分镜、脚本、镜头资产与生成记录，历史项目可迁移，避免核心资产散落在个人账号与本地文件中。画布协作：导演、美术、剪辑等角色可在同一块画布实时协同，同步节点状态与镜头修改，减少版本错乱。团队资产库：角色主体、风格模板、运镜方案、提示词工作流与参考素材支持统一沉淀，成熟视觉体系可持续复用。权限管理：支持项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作场景。成本管理：共享积分池、成员额度分配与消耗查看，团队管理者可统一管理生成资源、席位与项目成本。资产交接：成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免创作资产随人员流动中断。如何使用LibTV团队版创建团队空间：访问LibTV官网升级至团队版，建立专属团队项目空间，设置团队名称与基础信息。配置成员与权限：邀请导演、编剧、美术、剪辑等成员加入，按岗位分配访问、编辑、复制或发布权限。沉淀与迁移资产：将历史个人项目迁移至团队空间，或新建项目并导入参考素材、角色设定与风格模板。画布协作创作：成员在统一画布内协同推进分镜设计、镜头生成与修改，实时同步进度与反馈。复用团队资产库：调用已沉淀的角色主体、风格模板与提示词工作流，保持系列化内容风格一致。管控成本与交接：通过共享积分池分配生成额度，监控消耗；成员变动时资产自动保留，确保项目连续性。LibTV团队版的核心优势工业化协作：从个人创作升级为团队制片管理，支持长期项目、多岗位协同与规模化内容生产。资产持续复用：成熟的角色设定、风格与提示词工作流可沉淀为团队资产，降低新人上手与重复调试成本。实时协同提效：多人画布协作减少版本错乱与重复生成，围绕同一套镜头语言与素材体系推进项目。成本与风险可控：共享积分池与席位管理适配多项目并行需求，权限分级与资产交接机制保障核心创作安全。LibTV团队版的同类竞品对比对比维度LibTV 团队版快手可灵（企业版/团队功能）协作模式多人实时画布协作，导演/美术/剪辑等角色可在同一块画布内同步节点状态、生成进度与镜头修改团队空间内创作接力，支持桌面端（Win/Mac）同步，减少文件下载与搬运损耗资产沉淀角色主体、风格模板、运镜方案、提示词工作流、参考素材与工具箱统一沉淀，支持成熟视觉体系持续复用设立「共享资产库」，支持个人空间向团队空间一键转移灵感值、创作资产与主体资产，实现”一人贡献，全员复用”权限管理项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作所有者、管理员、协作者三重角色，灵活配置查看/编辑/移除权限，保障核心资产安全成本管理共享积分池、成员额度分配与积分消耗查看，支持多项目并行下的生成资源统筹支持灵感值跨空间转移，团队共享生成资源，但精细化成本管控功能相对较弱资产交接成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免资产随人员流动中断个人资产可一键导入团队空间，但成员退出后的自动资产交接机制未明确强调核心场景聚焦 AI 短剧、AI 电影、AI TVC 的工业化制片流程，强调长期项目管理与跨岗位协同覆盖 AI 视频生成、智能分镜、广告 Demo、影视特效等，强调快速创作与素材复用产品定位AI 视频专业级协作工作台，面向制片管理流程AI 视频生成平台的团队功能扩展，面向创作者联盟与工作室LibTV团队版的应用场景AI 视频工作室：适合长期运营 AI 短剧、动画、MV 的创作团队，通过资产库统一沉淀角色与风格，实现系列化生产。短剧承制方与 MCN：适配高频、多项目并行的工业化内容生产，导演与美术可同时在线推进分镜与镜头生成。品牌市场部与代理公司：支持品牌方、代理公司与外部创意团队跨组织协作，共同管理 AI TVC 与 Campaign 项目。AI 电影与大型制作团队：支持长篇叙事、系列化 IP 与复杂制作流程，导演组、制片组、美术组与后期团队跨岗位协同。企业内容与创意部门：适配培训、营销、产品与全球化传播等标准化内容生产，便于企业进行长期生成资源规划。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇PPT Master - 开源 AI PPT 生成标准化工作流下一篇腾讯Marvis - 腾讯应用宝推出的操作系统层个人 AI 助手相关文章Nanobot – 香港大学数据科学实验室开源的个人AI助手AI小集3Gemini Embedding 2 – 谷歌推出的首个原生多模态嵌入模型AI小集3DrawingSpinUp – AI驱动的2D绘画转化为3D效果的动画生成技术AI小集2MatterGen – 微软推出的无机材料生成模型AI小集2如何获取TRAE SOLO邀请码？附邀请码申请攻略AI小集4MemMachine – 开源AI记忆系统，支持多种记忆类型AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付刚刚Gemini 3.5 Flash – Google 推出的新一代 AI 大模型32秒前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版20小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架20小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准21小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型23小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具3天前Grok Build – xAI 推出的终端原生 AI 编程智能体3天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文4天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型4天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台4天前

Grok Build – xAI 推出的终端原生 AI 编程智能体AI工具3天前发布AI小集03Grok Build是什么Grok Build 是 xAI 推出的终端原生 AI 编程智能体（Agentic CLI），面向专业软件工程与复杂编码任务。基于 Grok 4.3 beta 的 200 万 token 超长上下文，支持 Plan Mode 计划审批、最多 8 个并行子智能体协作及 Arena Mode 自动方案排序，可在本地直接读写文件、执行 Shell 命令、操作 Git 并构建完整应用。采用本地优先架构，源代码与敏感数据不上传云端，保障企业级安全。Grok Build的主要功能Plan Mode（计划模式）：复杂任务执行前自动生成结构化步骤计划，用户可审批、评论或修改，所有代码变更以 diff 形式透明展示，避免”黑箱”操作。并行子智能体：最多支持8 个并发 AI 智能体同时工作，可分别处理代码分析、CI/CD 检查、性能优化、文档检索等不同子任务。Arena Mode（竞技场模式）：多个智能体针对同一任务生成不同代码方案，系统自动评估并排序，开发者直接选择最优解，无需手动对比。超长上下文窗口：基于 Grok 4.3 beta，支持200 万 token上下文，可一次性加载整个大型代码库进行跨文件复杂任务。本地优先架构：源代码、凭证和敏感数据不上传云端，完全在本地执行，支持离线环境和气隙网络，保障企业级代码安全。终端原生 CLI + Web 双入口：支持本地命令行交互，也可通过浏览器界面远程操作，满足不同工作流习惯。VS Code 扩展集成：提供官方 VS Code 插件，在 IDE 内直接调用 Grok Build 能力，实现 GUI adjacent 工作流。Git 与开发服务器集成：可自动管理 Git 仓库、启动本地开发服务器、内置浏览器浏览网页、执行 Shell 命令。插件与开放生态：支持 MCP 服务器、AGENTS.md 团队规范文件、Hooks 自定义动作、ACP（Agent Client Protocol）协议，便于集成现有 DevOps 流水线。Headless 脚本模式：通过-p标志支持非交互式执行，可直接嵌入 CI/CD 流水线或自动化脚本。BYOK（自带 API Key）：支持使用自有 xAI API Key 进行本地计费，灵活控制成本。如何使用Grok Build一键安装 CLI：在终端执行curl -fsSL https://x.ai/cli/install.sh | bash完成本地安装，支持 macOS、Linux 和 Windows 三大平台。启动交互式终端：安装完成后输入grok>进入交互式命令行界面，可直接用自然语言描述编程需求。登录账号：在终端内执行grok> /login登录 xAI 账号，解锁高级功能与订阅权益。配置 API Key（可选）：如需使用自有额度，执行grok> /api-key add xai添加个人 xAI API Key（BYOK 模式）。使用 Plan Mode 处理复杂任务：输入需求后，Grok Build 会自动生成结构化执行计划，展示步骤与预期变更 diff，用户可审批、评论或修改后再执行。利用并行子智能体：在复杂项目中，系统会自动或手动触发最多 8 个并发子智能体，分别处理代码分析、测试生成、文档检索等子任务。通过 Arena Mode 选优：对同一任务让多个智能体生成不同方案，系统自动评估排序，你只需选择排名最高的方案应用。Web 端远程操作：访问build.grok.com或x.ai/cli，在浏览器中打开 Grok Build 界面，无需本地安装即可使用。VS Code 集成工作流：安装 Grok Build 的 VS Code 扩展，在编辑器内直接调用智能体能力，边写代码边让 AI 协助修改。Headless 自动化执行：在脚本或 CI/CD 流水线中使用-p标志运行非交互式命令，实现自动化构建与部署。查看帮助与命令：在终端输入grok> /help查看所有内置命令与快捷操作指南。Grok Build的官网地址项目官网：https://x.ai/news/grok-build-cliGrok Build的核心优势本地优先，代码绝不上云：源代码、凭证及敏感数据完全在本地执行，无需上传至 xAI 服务器，适合处理专有代码库、金融及政务等受监管行业项目。200 万 token 超长上下文：基于 Grok 4.3 beta，可一次性将整个大型代码库载入内存，跨数十个文件进行复杂重构与多文件联动修改，大幅减少上下文丢失。8 路并行子智能体：最多同时调度 8 个并发 AI 智能体分工协作，分别处理代码分析、测试生成、CI/CD 检查等子任务，显著缩短复杂项目耗时。Arena Mode 自动选优：同一任务由多个智能体生成不同方案，系统自动评估排序并给出排名，开发者直接选择最优解，告别手动对比试错。Plan Mode 透明可控：执行前先输出结构化步骤计划与 diff 预览，用户可审批、评论或修改后再执行，每一步变更清晰可见，避免”黑箱”风险。极致 API 价格竞争力：输入 token 仅$0.20 / 百万 tokens，相较 Claude Code 和 Codex CLI 大幅降低高频使用成本。终端原生 + 多入口覆盖：支持 CLI 命令行、Web 浏览器、VS Code 扩展三种工作入口，既满足极客式终端操作，也兼容 IDE 集成习惯。Grok Build的同类竞品对比对比维度Grok Build (xAI)Claude Code(Anthropic)Codex CLI(OpenAI)产品定位终端原生 AI 编程智能体（Agentic CLI）终端 AI 编程助手云端 AI 编程智能体发布状态2026.5 早期 Beta已正式发布已正式发布底层模型Grok 4.3 beta / grok-code-fast-1Claude 4 (Sonnet/Opus)Codex (基于 GPT-4o/o3)上下文长度200 万 token20 万 token (Claude 4)约 20 万 token并行智能体最多 8 个并发子智能体单智能体为主单智能体为主方案评估Arena Mode 自动排序选优手动对比或迭代手动对比或迭代执行计划Plan Mode（先计划后审批执行）直接执行，可审查 diff直接执行，可审查 diff数据安全本地优先，代码不上云云端执行，代码上传 Anthropic云端执行，代码上传 OpenAI离线支持支持完全离线/气隙网络需联网需联网SWE-Bench70.8%约 72–75%约 70%API 定价$0.20 / 百万 tokensPro 订阅 + 按量计费Credits 按量计费订阅门槛SuperGrok Heavy ($300/月)Claude Pro / Team 订阅ChatGPT Pro / 开发者额度入口形态CLI + Web + VS Code 扩展CLI + IDE 集成CLI + ChatGPT 移动端自动化集成Headless 模式 (-p 标志) + CI/CD有限脚本支持有限脚本支持核心差异多智能体并行 + 本地优先 + 超大上下文模型推理能力强，生态成熟与 ChatGPT 生态深度整合Grok Build的应用场景大型代码库重构与现代化：基于 200 万 token 超长上下文一次性加载整个项目，跨数十个文件进行安全重构、依赖升级或架构迁移，避免上下文截断导致的遗漏。复杂全栈应用从零构建：通过 Plan Mode 自动生成项目结构、技术选型、依赖安装和文件生成计划，审批后直接执行，快速搭建完整应用原型。敏感行业专有代码开发：本地优先架构确保源代码、凭证和敏感数据不上传云端，适用于金融、政务、医疗、军工等对数据主权有严格要求的场景。离线/气隙网络环境编程：无需持续联网即可在本地执行全部任务，适配高安全隔离网络、内网开发机或无外网访问权限的工作环境。多维度自动化代码审查：启动多个并行子智能体分别检查代码安全性、性能瓶颈、CI/CD 合规性和文档完整性，一次性输出综合审查报告。技术方案快速对比与选型：使用 Arena Mode 针对同一需求生成多种实现方案，系统自动评估排序，帮助团队在短时间内确定最优技术路径。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇OpenHuman - 开源桌面端 AI 助手，能主动感知工作上下文下一篇Violin - 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具相关文章ShotAdapter – Adobe联合UIUC推出的多镜头视频生成框架AI小集2Kuse AI – AI画布工具，提供无限画布自由组织信息AI小集3秒哒 – 百度推出的零代码AI开发平台，一句话生成应用AI小集28Neo-1 – VantAI 推出的首个原子生成式AI模型AI小集3MoodMax – 一站式AI漫剧创作平台，自动拆解分镜脚本AI小集3NocoBase – 开源AI无代码开发平台，通过配置完成应用开发AI小集3暂无评论再想想发表评论暂无评论…

OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文AI工具4天前发布AI小集05OpenHuman是什么OpenHuman 是 tinyhumansai 团队推出的开源个人 AI 超级智能助手，定位为”Your Personal AI super intelligence”，主打私有、简单且极其强大。非传统聊天机器人，是能主动感知用户工作上下文的桌面级 AI Agent，通过每 20 分钟自动同步 118+ 第三方应用数据，构建本地持久记忆，让 AI 真正”懂你”。项目采用 Rust + TypeScript + Tauri 构建，支持 macOS、Linux 和 Windows 平台。OpenHuman的主要功能桌面吉祥物交互：具备视觉形象的桌面 AI 助手，支持语音输入（STT）与 ElevenLabs 语音输出（TTS），吉祥物可实时唇形同步，能作为参会者加入 Google Meet 并实时记录会议。118+ 第三方应用一键集成：通过 OAuth 快速接入 Gmail、Notion、GitHub、Slack、Linear、Jira、Stripe、Calendar、Drive 等主流工具，无需编写任何连接代码。每 20 分钟自动同步（Auto-fetch）：核心引擎每 20 分钟遍历所有活跃连接，自动拉取最新数据进入本地记忆系统，让 Agent 在次日早上已掌握前日全部上下文。Memory Tree 层次化持久记忆：将同步数据规范化为 ≤3k-token 的 Markdown chunks，评分后折叠存入层级摘要树，存储于本地 SQLite，实现跨会话的长期记忆。Obsidian Vault 原生兼容：同步数据同时输出为 Obsidian 可读的.md文件，用户可在 Obsidian Graph View 中浏览和编辑 AI 维护的知识网络。TokenJuice 智能 Token 压缩：在数据进入 LLM 前自动压缩（HTML 转 Markdown、缩短长 URL、移除非 ASCII 字符等），最高可降低 80% 的 API 调用成本与延迟。内置模型路由（Model Routing）：根据任务类型自动分配轻量模型、强推理模型或视觉模型，一个订阅覆盖多种模型能力，无需管理多个 API Key。完整工具链开箱即用：内置 Web 搜索、网页抓取（Web Scraper）、代码工具集（文件系统、Git、Lint、Test、Grep），无需额外安装插件即可使用。本地优先与隐私保护：所有工作流数据本地加密存储，支持通过 Ollama 接入本地模型实现完全离线运行，数据不上传云端。可选 agentmemory 共享后端：可配置与 Claude Code、Cursor、Codex 等工具共享同一个持久记忆存储，实现跨工具的知识互通。消息通道集成：支持 inbound/outbound 消息收发，打通用户已有的沟通渠道，工作流数据始终保存在本地设备。如何使用OpenHuman官网下载安装：访问OpenHuman的官网 https://tinyhumans.ai/openhuman 下载对应平台的 DMG（macOS）或 EXE（Windows）安装包，双击安装后即可启动。命令行快速安装（macOS / Linux x64）：在终端执行curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash自动完成下载与安装。命令行快速安装（Windows）：在 PowerShell 执行irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex完成一键安装。一键连接工作应用：打开应用后，通过 OAuth 一键接入 Gmail、Notion、GitHub、Slack、Calendar、Drive 等 118+ 第三方工具，授权后每 20 分钟自动同步数据。与桌面吉祥物交互：点击桌面上的吉祥物形象，通过语音或文字与其对话，可勾选 “Speak replies” 让 AI 语音播报回复。查看自动同步的记忆数据：在应用内查看 Agent 已同步的邮件、任务、代码仓库等上下文，AI 会自动压缩整理为 Memory Tree 供后续对话调用。Obsidian 知识库浏览：同步数据自动写入本地 Obsidian-compatible Vault，打开 Obsidian 即可在 Graph View 中可视化浏览 AI 维护的知识网络，也可手动编辑.md文件。使用内置工具集：直接调用开箱即用的 Web 搜索、网页抓取、代码工具（文件系统、Git、Lint、Test、Grep）等能力，无需额外安装插件。接入本地模型（可选）：如需完全离线运行，通过 Ollama 在本地部署模型，并在设置中切换至本地模型后端，实现零云端依赖。OpenHuman的项目地址官网地址：https://tinyhumans.ai/openhumanGitHub 仓库：https://github.com/tinyhumansai/openhumanOpenHuman的核心优势UI 优先分钟级上手：干净的桌面体验，无需终端配置，从安装到可用 Agent 仅需几分钟，区别于终端优先的竞品。上下文在分钟级而非周级：传统 Agent 需要数周学习用户习惯，OpenHuman 通过一次同步即可掌握邮箱、日历、仓库、文档、消息的全量压缩上下文。一站式订阅，零 API 蔓延：一个订阅覆盖所有功能，无需额外购买多个模型 API Key。Obsidian 知识库原生兼容：同步数据自动写入 Obsidian-compatible Vault，用户可在熟悉笔记界面中浏览和编辑 AI 维护的知识图谱。Rust + Tauri 高性能架构：内存占用远低于 Electron 方案，69.7% Rust + 26.1% TypeScript 构建。开源可审计：GPL-3.0 协议，代码完全透明，隐私敏感用户可自行审计数据流向。OpenHuman的同类竞品对比对比维度OpenHumanOpenClawHermes Agent开源协议✅ GPL-3.0✅ MIT✅ MIT上手难度✅ UI 优先，分钟级上手⚠️ 终端优先配置⚠️ 终端优先配置使用成本✅ 单一订阅 + TokenJuice 压缩⚠️ 自备模型 API⚠️ 自备模型 API记忆能力🚀 Memory Tree + Obsidian Vault + 可选 agentmemory 后端⚠️ 依赖插件扩展✅ 自学习观察第三方集成🚀 118+ 应用 OAuth 一键接入⚠️ 需自行配置连接⚠️ 需自行配置连接自动数据同步✅ 每 20 分钟自动 fetch 至本地记忆🚫 无自动同步🚫 无自动同步API 管理✅ 一个账户覆盖所有模型🚫 自备 Key🚫 多供应商管理模型路由✅ 内置自动分配（轻量/推理/视觉）⚠️ 手动切换⚠️ 手动切换原生工具集✅ 代码 + 搜索 + 抓取 + 语音✅ 代码工具✅ 代码工具数据隐私🚀 本地优先，数据不上云依赖自托管配置依赖自托管配置OpenHuman的应用场景跨应用项目上下文整合：同时连接 GitHub、Linear、Jira、Slack 后，让 AI 自动对比代码 PR、任务看板和团队消息，一键汇总项目最新进展，无需手动切换多个工具。智能邮件筛选与摘要：授权 Gmail 后，AI 每 20 分钟自动扫描收件箱，按项目或优先级分类摘要，快速定位需要回复的重要邮件，过滤掉低价值信息。AI 会议助手：让桌面吉祥物以真实参会者身份加入 Google Meet，实时记录会议内容、提取 action items，并基于记忆库提供参会者背景信息辅助沟通。代码仓库智能问答：同步 GitHub 仓库数据后，直接向 AI 询问代码逻辑、历史变更原因、PR 评论争议点，Agent 基于完整代码上下文给出精准回答。Obsidian 个人知识库构建：将邮件、文档、聊天记录自动压缩归档为 Obsidian-compatible 的.md文件，在 Graph View 中可视化呈现知识关联网络，打造 Karpathy 风格的 LLM 知识库。多工具工作流自动化：连接 Notion、Calendar、Drive、Stripe 等工具后，AI 自动关联”日历事件→相关文档→对应邮件→财务记录”，形成完整的项目上下文链。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Intern-S2-Preview - 上海 AI Lab 开源的科学多模态大模型下一篇Grok Build - xAI 推出的终端原生 AI 编程智能体相关文章We0 – 开源AI代码编辑器，支持在浏览器中运行调试代码AI小集6雾象Fogsight – AI动画生成Agent，输入主题生成完整叙事动画AI小集6Hailuo 2.3 – MiniMax推出的AI视频生成模型AI小集3SWE-1.5 – Cognition 推出的高性能AI编程模型AI小集2Paperclip – 开源的AI Agent编排平台，自主运营AI公司AI小集3Z-Image – 阿里通义推出的图像生成模型AI小集7暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付刚刚Gemini 3.5 Flash – Google 推出的新一代 AI 大模型35秒前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版20小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架20小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准21小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型23小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具3天前Grok Build – xAI 推出的终端原生 AI 编程智能体3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型4天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台4天前

腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手AI工具2天前更新AI小集03腾讯Marvis是什么腾讯Marvis是腾讯应用宝团队推出的操作系统层个人 AI 助手。工具通过 1 个主 Agent 与 5 个副 Agent 的协作架构，实现电脑系统操作、文件管理、App 操控、网页交互与信息搜索等复杂任务，24 小时在线并支持 Windows 与安卓跨端操作。产品具备本地知识库索引与可视化虚拟办公室交互，定位大众友好的日常 AI 数字秘书。腾讯Marvis的主要功能多 Agent 协作调度：1 个 PM 主 Agent 负责理解需求、拆解任务并分发给 5 个副 Agent，各 Agent 完成后向主 Agent 汇报，实现复杂任务的自动编排与闭环执行。深度系统操作：可直接调用 Windows 底层接口完成主题设置、电源计划调整、自启动管理、硬件配置检测、网络测速等原生系统级操作，无需模拟点击。智能文件管理：支持基于文件名、内容、图片 OCR、人脸、主题、地点等多维度本地文件搜索，可自动分类归档、格式转换、合并去重及跨端文件传递。跨端 App 操控：通过内嵌应用宝引擎，可在 PC 端直接操控安卓手机 App（如微博签到、同花顺、飞常准等），实现电脑与手机的应用级跨端调度。本地知识库：自动识别并索引本地文档与图片，支持标题匹配与内容匹配搜索，构建越用越懂用户的个人知识管理体系。自动任务与情报监控：支持定时执行既定任务，如每日新闻汇总、限时福利监控、明星动态追踪、房价监控等自动化工作流。如何使用腾讯Marvis下载安装：访问腾讯Marvis官网 https://marvis.qq.com/ 下载 Windows 客户端，按指引完成安装。登录激活：填写调查问卷获取邀请码，获得邀请码后，输入邀请码完成内测资格激活。模式选择：根据任务敏感度选择效率模式或隐私模式。指派任务：在对话框输入自然语言指令，或进入”自动任务”页面设置定时执行规则。查看进度：在右侧虚拟办公室面板查看各 Agent 工作状态与任务消耗 Token 数，等待主 Agent 汇总结果。确认执行：涉及系统修改、文件删除等敏感操作时，点击硬确认弹窗授权后完成最终执行。腾讯Marvis的核心优势操作系统层深度整合：区别于纯云端 Agent，Marvis 深入 Windows 系统底层，可直接调用系统 API 与文件系统，实现真正的本地原生操作。端云双模式灵活切换：效率模式采用云端大模型保障复杂任务性能，隐私模式基于 Qwen 端侧模型实现完全离线运行，满足财务、法务、HR 等敏感场景需求。应用宝跨端生态壁垒：继承 14 年应用商店经验与跨端引擎技术，独家支持 PC 端操控手机 App，生态调度能力为竞品不具备。可视化虚拟办公室交互：将每个 Agent 设计为戴企鹅围脖的黑色小马形象，在虚拟工位中实时展示工作状态（搬砖、打盹、喝咖啡），降低技术门槛并增强用户情感连接。L2 级安全兜底机制：涉及资金交易、系统核心配置修改、批量删除等敏感操作强制”硬确认”，支付功能完全禁用，从架构层面规避 AI 越权风险。腾讯Marvis的同类竞品对比对比维度腾讯Marvis（腾讯）Claude Computer Use（Anthropic）产品定位操作系统层个人 AI 助手云端计算机操控 Agent运行模式端云双模式（本地隐私 / 云端效率）纯云端运行系统深度深入 Windows 底层 API，原生系统级操作基于截图与模拟点击操控跨端能力PC 操控手机 App（应用宝引擎）不支持Agent 架构1 主 + 5 副多 Agent 协作，可视化办公室单 Agent 模式隐私安全L2 级硬确认，支持完全离线运行数据上云，依赖平台安全策略用户门槛自然语言输入，大众友好，开箱即用需一定技术理解，偏极客向生态合作与微软、Intel 系统级深度合作无特殊硬件合作腾讯Marvis的应用场景办公效率提升：自动整理桌面文件并按项目分类归档，一键完成合同信息审查、运营数据分析、参考文献整理及出差材料准备。系统运维助手：一句话调整系统设置（夜间模式、电源计划、默认浏览器），检测游戏配置兼容性并给出硬件升级建议，彻底卸载顽固软件及残留痕迹。生活娱乐管家：自动完成微博超话签到、监控游戏限时福利与明星动态、定时推送每日新闻汇总、整理宝宝成长相册与旅行攻略。敏感数据处理：在完全断网的隐私模式下，本地完成财务报表校对、员工薪资分析、法务合同审查等涉密工作，确保数据不出设备。跨端文件协作：在手机端发送指令，让 Marvis 在电脑中搜索指定文件并回传至手机，实现跨设备的无缝文件调度。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇LibTV团队版 - LibTV 推出的专业级 AI 视频协作工作台下一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型相关文章Yollo AI – 沉浸式AI角色对话与视频生成平台AI小集2AISEO Art – AI绘画工具，支持多种场景多种绘画风格和技巧AI小集2slidebean – AI PPT生成工具，快速生成商业计划书、项目汇报等专业PPTAI小集3Graphic Design Generator – Recraft推出的AI海报设计生成工具AI小集3幻舟AI – AI短片创作平台，从剧本到视频全流程一站式服务AI小集4DynaMem – 纽约大学和Hello Robot推出的动态空间语义记忆系统AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付刚刚Gemini 3.5 Flash – Google 推出的新一代 AI 大模型32秒前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版20小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架20小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准21小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型23小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具3天前Grok Build – xAI 推出的终端原生 AI 编程智能体3天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文4天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型4天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台4天前

Kimi K2.5 – 月之暗面kimi开源的全新一代全能旗舰模型AI工具4个月前发布AI小集03Kimi K2.5是什么Kimi K2.5 是月之暗面开源的全新一代全能旗舰模型，模型基于约 15T 视觉与文本数据训练而成，具备顶尖的代码生成、视觉理解能力，支持自主 Agent 集群协作，可调度多达 100 个子 Agent 并行处理复杂任务，执行效率大幅提升。Kimi K2.5 提供快速、思考、Agent 及 Agent 集群四种模式，适用多种场景，包括前端开发、办公自动化和复杂问题解决。模型可通过Kimi.com、Kimi App、Kimi 开放平台和 Kimi Code （Kimi 新推出的编程工具）等平台提供服务，助力用户高效完成任务。Kimi K2.5的主要功能视觉与文本融合：Kimi K2.5 能同时处理文字和图像，理解设计图、视频等内容，实现“看图写应用”，将视觉内容转化为可交互的网页或代码，支持像素级复刻和视觉调整。多模式思考：模型支持思考模式和快速模式切换。思考模式适合复杂任务，可进行深度逻辑推理；快速模式适合即时问答，响应迅速。智能体协同（Agent Swarm）：在复杂任务中，Kimi K2.5 可拆解任务并调度多个“分身”协同工作，支持高并发处理，完成多达 1500 步的操作，适用批量任务和多视角分析。专家级办公助手（Office Agent）：模型深度优化 PPT、Word、Excel 等办公软件的使用，能处理文本和解析图表和复杂排版，提供专业级的输出。本地编程支持（Kimi Code）：基于终端的 Coding Agent，支持本地代码的增删改查，自动生成和修改代码，运行测试，并解释改动，方便开发者使用。一键部署与应用生成：生成的网页或应用可直接部署上线，快速将设计转化为实际可用的产品。Kimi K2.5的性能表现Kimi K2.5 是基于约 15T 视觉和文本混合数据训练的原生多模态模型，在视觉理解、图像生成、视频分析等任务中表现出色。例如，在 MMMU-Pro 和 MathVision 等视觉基准测试中，Kimi K2.5 的表现优于或接近其他顶尖模型。Kimi K2.5的项目官网项目官网：https://www.kimi.com/blog/kimi-k2-5.htmlHuggingFace模型库：https://huggingface.co/collections/moonshotai/kimi-k25Kimi K2.5的API调用价格输入（非缓存）：每 M Tokens 价格为 ¥4，相比 K2 Turbo降低了 50%。输入（缓存）：每 M Tokens 价格为 ¥0.7，相比 K2 Turbo 降低了 30%。输出：每 M Tokens 价格为 ¥21，相比 K2 Turbo降低了 63.8%。Kimi K2.5的应用场景软件开发与编程：Kimi K2.5 可将设计图或视频直接转化为可交互的网页或应用，支持本地代码生成、修改和一键部署，大幅提升开发效率。办公自动化：模型能作为专家级办公助手，深度优化 PPT、Word 和 Excel 文档，支持批量任务处理和多任务协作，显著提高办公效率。教育与学习：Kimi K2.5 能根据学生需求生成个性化学习计划和教学材料，提供实时辅导，减轻教师备课负担。创意与设计：模型可将设计稿直接转为代码，支持批量生成创意素材和视觉调整，助力设计师快速完成项目。商业与市场分析：Kimi K2.5 能快速进行市场调研、生成商业计划书，通过数据可视化辅助决策，助力企业高效分析市场。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇ChatTOC - AI聊天对话生成目录插件，一键定位关键信息下一篇DeepSeek-OCR 2 - DeepSeek团队开源的二代OCR模型相关文章VirtualFit – AI虚拟试衣工具，提供AI换衣、图像恢复与扩展等多功能AI小集2Poster Studio – AI广告创意生成工具，一分钟内创建高质量视觉内容AI小集2Claude Design系统提示词 – Anthropic 推出的完整核心提示词AI小集3CapCut – 抖音推出的免费全能视频编辑器AI小集3轻游梦工坊 – 腾讯推出的无代码编程游戏编辑器AI小集4Fellou – Fellou AI 推出的首个Agentic浏览器AI小集8暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付18分钟前Gemini 3.5 Flash – Google 推出的新一代 AI 大模型19分钟前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版20小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架20小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准22小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型23小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具3天前Grok Build – xAI 推出的终端原生 AI 编程智能体3天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文4天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型4天前

Composer 2.5 – Cursor 推出的自研 Agentic 编程模型AI工具23小时前发布AI小集02Composer 2.5是什么Composer 2.5 是Cursor推出的自研 Agentic 编程模型。在智能水平和行为表现上较Composer 2有大幅提升，在SWE-Bench Multilingual（79.8%）和CursorBench v3.1（63.2%）等核心基准上与 Claude Opus 4.7、GPT-5.5 处于同一梯队，但单次任务成本仅为竞品的约1/10，称为”性价比之王”。模型基于 Moonshot 开源的Kimi K2.5检查点持续训练，目前仅通过 Cursor IDE 及 SDK 提供服务。Composer 2.5的主要功能长时任务持续工作：针对长时间运行的 Agent 会话深度优化，能在多步骤工具调用中保持专注，显著减少中途幻觉或提前终止的问题。复杂指令可靠遵循：对跨文件重构、终端命令执行、测试驱动开发等复杂指令的遵循可靠性较 Composer 2 大幅提升。努力级别动态校准：模型能根据任务难度自动分配计算量，简单任务快速完成，复杂任务深入思考，避免”小事空转、大事欠思考”。沟通风格优化：回复更简洁结构化，减少不必要的冗长解释，在多文件变更时提供更清晰的推理过程。工具调用精准度提升：显著减少无效的终端命令或冗余搜索，提升代码检索与终端操作效率。双版本灵活适配：提供 Standard（标准版，$0.50/$2.50 per M tokens）与 Fast（快速版，$3.00/$15.00 per M tokens），智能水平相同，分别适配后台批量任务与交互式实时编程。如何使用Composer 2.5在 Cursor IDE 中启用：将Cursor更新至 2026 年 5 月最新稳定版，打开 Composer 面板（Cmd+I/Ctrl+I），点击模型选择器切换至Composer 2.5选择速度档位：交互式开发默认使用Fast 版（响应快、延迟低）；后台 Agent 或批量任务可在 Settings > Models 中切换为Standard 版（成本低，智能水平相同）通过 SDK 程序化调用：import { Agent } from “@cursor/sdk”; const agent = await Agent.create({ model: “composer-2.5”, // Standard 版 // model: “composer-2.5-fast”, // Fast 版 workspace: “./”, tools: [“edit”, “shell”, “search”, “browser”], });为长时任务设限：为无人看管的长时间 Agent 会话设置迭代次数上限和最大耗时，防止模型利用缓存等”捷径”进行奖励作弊领取首发福利：发布首周提供双倍用量额度，可在后台查看额度消耗情况多模型路由策略：将 90% 日常任务交给 Composer 2.5 处理；仅将架构评审类任务路由给 Claude Opus 4.7，重度终端 Shell 任务路由给 GPT-5.5Composer 2.5的核心优势极致性价比：Standard 版输入 $0.50/M、输出 $2.50/M，Fast 版输入 $3.00/M、输出 $15.00/M，比 Claude Opus 4.7 便宜约1030 倍。前沿级基准表现：SWE-Bench Multilingual 79.8%（Opus 4.7 为 80.5%，GPT-5.5 为 77.8%），CursorBench v3.1 63.2%（与 Opus 4.7 的 64.8% 和 GPT-5.5 的 64.3% 基本持平）。行为层面深度优化：除了扩大训练规模，还改进了沟通风格和投入级别校准，这些维度虽难被基准充分反映，但对实际使用体验至关重要。双版本灵活选择：Standard（标准版）适合后台 Agent 与批量任务，Fast（快速版，默认）适合交互式 IDE 实时编程，两者智能水平完全相同。首发双倍用量福利：发布首周提供双倍用量额度。Composer 2.5的同类竞品对比对比维度Composer 2.5Claude Opus 4.7GPT-5.5厂商 / 平台CursorAnthropicOpenAI产品定位自研 Agentic 编程模型旗舰推理模型旗舰多模态模型基座模型Moonshot Kimi K2.5（开源检查点持续训练）Claude 4 系列GPT-5 系列发布日期2026.05.182026 年 Q22026 年 Q2SWE-Bench Multilingual79.8%80.5%77.8%Terminal-Bench 2.069.3%69.4%82.7%CursorBench v3.1（困难任务）63.2%64.8%（max）/ 61.6%（默认 xhigh）64.3%（xhigh）/ 59.2%（默认 medium）输入价格（/M tokens）$0.50（Standard）<$3.00（Fast）未公开（行业参考约 $15）未公开（行业参考约 $3–$5）输出价格（/M tokens）$2.50（Standard）<$15.00（Fast）未公开（行业参考约 $75）未公开（行业参考约 $15–$30）单次任务相对成本基准（约 $1–$2 / 任务）约10–30 倍约3–10 倍上下文窗口200K（参考 Kimi K2.5）200K128K–1M权重开放性闭源（仅 Cursor 基础设施）闭源闭源接入方式Cursor IDE / CLI /@cursor/sdkAPI / Claude Code / 第三方平台API / ChatGPT / GitHub CopilotComposer 2.5的应用场景多文件级重构：成本优势明显且精度持平前沿模型，适合大规模代码库迁移。交互式结对编程：Fast 版响应迅速，适合实时 IDE 协作。后台定时任务/云 Agent：Standard 版性价比极高，适合批量代码审查与修复。测试驱动开发：长时任务可靠性优于前代，能稳定完成多轮测试-修复循环。复杂终端自动化：Terminal-Bench 2.0 得分 69.3%，与 Opus 4.7（69.4%）持平，但重度 Shell 场景仍略逊于 GPT-5.5（82.7%）。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型下一篇Chronicles-OCR - 腾讯联合高校等推出的视觉感知评测基准相关文章dots.vlm1 – 小红书hi lab开源的首个多模态大模型AI小集3MAI-Transcribe-1 – 微软推出的语音转文字模型AI小集2SWE-Kit – 构建自定义软件工程AI代理的开源框架AI小集2NovaMSS – AI音乐源分离工具，精准分离音乐中的不同音源AI小集3新言意码 – 前月之暗面明超平创立的 AI Coding 项目AI小集3成交输入法 – 专为销售人员设计的AI输入法应用，内置销售话术和复盘功能AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付刚刚Gemini 3.5 Flash – Google 推出的新一代 AI 大模型33秒前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版20小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架20小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准21小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型2天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手2天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台2天前PPT Master – 开源 AI PPT 生成标准化工作流2天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架2天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具3天前Grok Build – xAI 推出的终端原生 AI 编程智能体3天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文4天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型4天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台4天前

Katie DrummondVideoMay 19, 2026 6:30 AMTom Steyer Wants to Save California From Billionaires. But Also Doesn’t Want Them to LeaveThe hedge fund billionaire turned gubernatorial candidate wants to tax California’s ultrawealthy, regulate AI, and keep Silicon Valley happy at the same time. Good luck with that.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyFor those concernedabout the influence of Big Tech and billionaires on California’s future,Tom Steyerlooks like an obvious choice. A billionaire who amassed his fortune after founding Farallon Capital Management, one of the world’s biggest hedge funds, Steyer quit the firm in 2012 and turned to philanthropy, political advocacy, and climate activism, among other pursuits. Now, he’s jostling for position among a handful of Democratic and GOP candidates looking to advance from a June primary and then win the California governorship this November.Ahead of the midterms, I’m talking to candidates relevant to WIRED’s interests: A few weeks agoI spoke with Alex Bores, a candidate for New York’s 12th Congressional District, whose history as aPalantiremployee and stance on AI regulation has attracted the ire of Silicon Valley–backed super PACs.Steyer felt like the next obvious choice for a conversation: He’s running to lead a state where issues like AI, immigration enforcement, and climate change, among other core WIRED subjects, are paramount. Steyer’s posture in the race is also unique. He’s been described as a “class traitor” for ostensibly eschewing his fellow elites, voiced support for California’s controversial Billionaire Tax Act—which has everyone from Sergey Brin to Peter Thiel either making moves to or threatening to flee the state—and campaigned hard on affordability, climate policy, and the promise that he’s immune to corporate influence. (As a billionaire spending more than $130 million on his own gubernatorial campaign, I certainly hope he would be.)As I said, for some Democratic voters, Tom Steyer seems to check a lot of boxes. Then he starts talking.Steyer is adept, as politicians usually are, at toeing the line. But the line, in politics generally and California specifically, seems to be the problem: Steyer, or whomever is elected to the governorship this November, will be walking an exceedingly thin one. Taxing California’s billionaires without alienating them. Getting a grip on the state’s AI development without throttling it (or, again, alienating the billionaires building it).I could feel Steyer’s reluctance to come down too firmly or dig in too deeply on issues, maybe to avoid alienating any potential voting block. Which made me wonder: Can Tom Steyer be a pro-billionaire governor who also taxes the hell out of them? Can he rave about the “mind-blowingly amazing” advances in AI while bringing the industry to heel? Can he learn the name of WIRED’s global editorial director (me) before she interviews him?The third question is answered in the interview. The former two will be formidable challenges for anyone elected to California’s governorship—and I didn’t leave our conversation convinced that Steyer’s posture is a particularly coherent one. The minimum requirement for a California governor might be the ability to use Google.This interview has been edited for length and clarity.KATIE DRUMMOND: Welcome, Tom, thank you for joining us onThe Big Interview.TOM STEYER:Kate [sic], thank you for having me.So, you’re a billionaire. You made your money in the hedge fund world. But now, in the last decade-plus, you’ve become a climate activist. Tell us about that transformation.When I was growing up, when I got free time, either from school or work, I tried to go to wild places and get outdoor jobs. I worked as a ranch hand, I worked picking fruit. Before I went to business school, I spent the summer in Alaska, and I went to Alaska because I wanted to see what North America looked like before Europeans showed up.I wanted to see the animals, I wanted to see the birds, I wanted to see the fish, I wanted to look at Denali. I wanted to see what it looked like, vast untracked North America, rich and fertile.In 2006, I wanted my wife and four kids, none of whom had ever been there, to see it too. So we went up there for a week. I had a whole bunch of plans of what we could see, but what we could see was Alaska was melting. It was really obvious. It’s one thing to read about it on a page when you’re working away and thinking about a thousand things, but it’s another thing to physically see where there used to be a mountain of ice and now it’s a valley.At what point, though, do you say, “I am going to say goodbye to running a hedge fund. I’m going to turn my back on this career, and I’m going to do something totally different. And a big part of that is going to be climate activism”? How did that happen?Well, I didn’t just leave specifically to do climate. You know, it was funny. I was talking to

Heather HansmanScienceMay 19, 2026 7:00 AMThe US Built a Site to Ensure Fair Access to Public Lands. Then Everything Went WrongRecreation.gov was supposed to make access to public lands more equitable and streamlined. Instead, it’s rife with bots and inequality, while a government contractor benefits.Illustration: Jacqui VanLiewCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyIt’s a fewminutes before 8 am Mountain Time on March 16, the day that river permit cancellations are released on Recreation.gov, the federal website for public land reservations.Rec.gov, as it’s commonly called, administers everything from river permits and timed entrance fees at the most popular national parks to campground reservations on remote sites belonging to the Bureau of Land Management, and a lot of people are recreating on public land these days. There were 11 million reservations on the site in 2024, up significantly from 3.5 million reservations reported in 2019. At the center of it all is an unlikely player in the outdoor recreation space: The site is operated by the government contractor Booz Allen Hamilton, a corporation known more for cybersecurity than rafting trips.Early each year, outdoor enthusiasts gear up forRecreation.gov’s annual lotteries for some of the most iconic experiences in the country: a river trip down Idaho’s Middle Fork of the Salmon River, which flows through the Frank Church River of No Return Wilderness. Backcountry permits to hike into the Wave, an otherworldly rock formation in Arizona’s Paria Canyon–Vermilion Cliffs Wilderness. Overnight stays in the rugged, lake-studded Enchantments, in Washington’s Okanogan-Wenatchee National Forest.The Paria Vermillion Cliffs.Photograph: Bernie Friel/UCG/Universal Images Group via Getty ImagesOdds of getting a desirable Middle Fork permit are around 2 percent. Each year, around 200,000 people apply in advance for 48 daily lottery spots to hike into the Wave.Rec.govitself reports that a campground with 57 campsites can see 19,000 users trying to reserve them. That’s a .3 percent success rate.For the majority who don’t draw a permit, there’s one final hope: the release date for cancellations, where your chances of getting a spot are often based on how fast you can click, and whether you can be online right when canceled permits are released.That’s where a river runner I’ll call Jack was last March 16. A web-developer friend of a friend who is frustrated by the way permits seem to be snapped up faster than humans can possibly click, Jack decided to do some experimenting to see if the speculation that bots are grabbing all the permits seems true.That speculation is based in reality. There’s a user on the outdoor forum Mountain Buzz who offers up a free scraperbot to anyone who wants to use it, and developers have sharedtheir code. Last year Sam Carter, the host of theRiver Radiuspodcast, did an episode where he built a bot to show that gaming theRecreation.govsystem was possible. He was shocked at the response. “So many people say they’re using bots, people are bragging about it,” he told me. He heard from people who’d built their own, groups who have their own server dedicated to getting permits, and people who paid thousands of dollars to have someone build one for them. It’s happening. The question is how pervasive it is, and how easy it might be for anyone to hack Rec.gov.Jack wants to prove that the system isn’t working, so he’s built a series of bots to try to outsmart the other bots. He has multiple accounts and bots that can do everything from alerting him to permit availability to keeping permits in his cart for hours at a time. “I’m trying to simulate what I think other people are doing,” he says.The night before the river cancellations he’s interested in are released, Jack opens the Inspect Element browser tool on Rec.gov and scrolls through the data to find what dates are going to become available. “If you were a web developer of any kind,” he says, “you would be able to find it.”He has programmed several bots, attached to burner accounts with obviously fake names, to try to grab the dates he knows will come up when permits are released at the tick of 8 am. As a control, he also has a friend on a laptop nearby who will be attempting to grab a permit the old-fashioned way, by clicking through the site herself.When the clock turns over, everything happens fast. The scripts start running, flashing between five different screens, running through the calendars on Rec.gov pages. Then, within a few seconds, the action stops. Jack goes to the accounts to see if his experiment worked. In the first cart, the bot has secured permits for the Main Salmon, the San Juan River, and the Middle Fork of the Salmon. It’s impossible to overstate what a big deal it is to score these three prime trips in a year, much less in a single morning. In the second cart, another unicorn: a second Middle Fork permit.The Salmon River near Stanley, Idaho.Ph

Julian ChokkattuGearMay 19, 2026 7:00 AMThis Is Herman Miller’s New Standing Desk for GamersThe Coyl uses a coil design for the power cable, keeping it neat and taut between the desk and the wall. And instead of relying on boring levers to adjust the height, it has a nifty rotary dial.Courtesy of Herman MillerCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyWhen you buya “gaming” accessory, you’re often buying a product that offersa little more oomph—a few extra perks or advanced features—catered to gamers who spend hours clicking away, trying to win. They may not be particularly chic, but these gaming peripherals are all about maximization, whether that’s agaming keyboard, agaming headset, or agaming mouse.Gaming desks are now a growing category. These furniture pieces accessorize your game room or home office while also minimizing distractions, so gamers can focus on gameplay. They often are designed to match their accompanyinggaming chairsand their distinctive race car seat aesthetic. Various gaming-focused companies likeSecretlaband Corsair offer some of the more comprehensive gaming sit-to-stand desks on the market—evenIkea has a few—but now esteemed furniture brand Herman Miller is entering the fray through its sub-brand, Herman Miller Gaming.TheHerman Miller Gaming Coylis the company’s firststanding deskbuilt for gamers. True, the company also sellsMotia-branded gaming desksfrom 2018, but the Coyl is more specifically designed for gamers. But this is its first in-house design, and it’s easy to tell the various desks apart thanks to the Coyl’s three main design elements: the trough, the coil, and the dial.First, the trough. This is the giant space underneath the desk, towards the back, where gamers can store all those unsightly wires and cables that run from their peripherals to their gaming PC. Most standing desks today offer some kind of way to route cables through the desk, but that’s child’s play compared to this. Very much like Secretlab’s Magnus Pro desk, the trough is unusually wide to accommodate plenty of cables and power strips. What’s also nice is that you can raise the desk to its max height and pull down the felt cover to access the trough from the front, like a mechanic working on a car, without needing to awkwardly fish for cables from the back of the desk or pull the whole desk away from the wall.Then there’s the desk’s namesake, the coil: a distinctive red wire that runs from an outlet to the desk to power the motors and anything else you plug into the integrated power strip. It’s an intentional design. Rather than let the cable dangle as it awkwardly stretches to your outlet, the coil design keeps it taut.That curly red cable.Photograph: Julian Chokkattu“Coils by nature solve a problem, which is how do you keep a space clean and tidy, that has a variable length between two things,” says Steven Harton, senior product manager at Herman Miller. He says the design choice was inspired by the coiled cables on vintage tech like telephone handsets and early mechanical keyboards. The result is fun and orderly.Finally, there’s the rotary dial. This is the height adjustment mechanism on the Coyl desk. Rather than using traditional paddles or levers to make the desk go up or down, Herman Miller Gaming employs a dial that feels like something you’d find on an old-school stereo—or a really nice gaming mouse. It lets you fine-tune the desk’s height to a 10th of an inch. (A display shows the table’s exact height.) That level of detail is an ode to the kind of precision gamers are fond of, digging into settings to maximize performance.Twist the knob to adjust the height of the desk.Photograph: Julian Chokkattu“Sensorially, how it feels and the feedback you get is an elevated experience relative to a button or a toggle,” Harton says. “When you use a dial, there’s a resistance to it that’s delightful.” We’ll have to wait forKnob Feel’s verdicton exactly how delightful it is.There are other niceties. The Coyl desk has integrated hooks on either side, which gamers can use to hang their headphones or a backpack. An optional back panel upgrade adds a classic pegboard to the desk, allowing gamers to add small shelves and hooks to hold their gaming controllers, small plants, lamps, or their favorite Funko Pop. There’s also the option to add other Herman Miller accessories like monitor mounts.The base version of the Coyl desk starts at $1,095. Add the cable management trough, and it’ll cost $1,495, and with the back panel, it tops out at $1,635. The rounded laminate desktop comes in four colors: black, white, walnut, and ash. You can program up to four height presets for the desk.It may be advertised as a gaming desk, but Harton thinks the modern gaming environment is multifunctional—not just for gaming, but for consuming, working, and relaxing. That, and the fact that the desk is drummed up by a furniture company and not a gaming brand is likely why the design may not feel very gam

Scott GilbertsonGearMay 19, 2026 8:08 AMThe Best Action CamerasGearing up to shred the slopes or dive into the seas? These photography tools are made for danger.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyChoosing the bestaction camera used to be simple:Go with GoPro. That’s still good advice, as the GoPro ranks high on our list, and itsforthcoming Mission 1 camerascould potentially top this list again, but there are also some worthy competitors that beat the current flagship GoPro Hero 13 Black on several fronts.The decision is not easy, so to figure out what exactly makes the best action camera, we dove with them, climbed with them, surfed with them, and handed them to reckless 13-year-olds on bikes. We found several great options that will survive your adventures on the ski slope, at the beach, or at the skate park.Be sure to also read our guide to theBest 360 Cameras,as there’s considerable overlap these days with action camera, and in some cases, for some people, they might be the better option. We’ve also got guides to theBest Travel Cameras,Best Mirrorless Cameras, andBest Instax Cameras.Updated May 2026: I’ve noted GoPro’s announcement of the new Mission 1 Cameras, which will have 8K video and support interchangeable Micro Four Thirds lenses. I’ve also removed the GoPro Hero 12 Black because it’s no longer a good value, and I’ve updated prices and links throughout.GoPro Mission 1 UpdateCourtesy of GoProCasting a long shadow over any action camera buyer’s decision right now is therecently announced GoPro Mission 1 cameras. There are three of them, theMission 1, Mission 1 Pro, and Mission 1 Pro ILS. All three feature a larger sensor with 8K video, improved low-light performance, up to 32X slow-motion footage, and 50-megapixel still images. The Pro ILS uses interchangeable lenses based on the Micro Four Thirds mount.GoPro claims the new sensor offers a dynamic range of up to 14 stops, whichrivals full-frame mirrorless cameras. This will be further helped by the available 10-bit GP-Log2 shooting mode, which maximizes dynamic range and allows for custom coloring in postproduction. Also interesting are some of the new accessories and add-ons. There are lens hoods, ND filters, a battery grip, a new Media Mod kit with a built-in mic, along with a 3.5-mm mic port, a headphone port, a micro HDMI port, and a line-in port for time code sync. GoPro is also making a point-and-shoot grip, reminiscent of Insta360’s grip for the Ace Pro 2.Should you wait and see? Probably. The Mission 1 cameras are a bit pricier, $600 for the Mission 1, $700 for the Pro and Pro ILS (the latter of which won’t be available until Q3 2026), but the video and image quality GoPro has teased thus far are very impressive. We’ll have a full review coming next month.Best Overall Action CameraPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonChevronChevronSave to wishlistSave to wishlistDJIOsmo Action 6$426Amazon$436B&H PhotoDJI’sOsmo Action 6 is the best action cameraright now. It trumps our previous top pick, theDJI Action 5 Pro, in several ways, adding a variable aperture for better low light performance and a new, larger sensor for higher-quality video. The Action 6 also gains 2X zoom and 8K video, though as withthe Insta360 Ace Pro 2, the 8K video is more of a specs-based bragging right than a significant improvement over the 4K. It does mean you can crop in though, whether using the 2X zoom in-camera, or on your own when you’re editing.Another great upgrade in the Action 6 over the Action 5 is some fairly dramatic improvements in color rendering and contrast in the default color modes. This is less apparent if you shoot Log-D, but if you don’t want to spend time color-grading your footage, now you don’t really need to. The default colors are pretty great for an action camera. Low-light performance is better, as you would expect with a wider-aperture lens. The video is cleaner and sharper than either the Action 5 Pro or the Ace Pro 2. That’s true whether you use SuperNight mode or not. I should note, though, that you don’t get full manual control here, but the automatic modes were pretty smart most of the time in my testing.DJI has also followed GoPro’s lead here, adding amacro lens ($169)to its lineup, which also includes an ultrawide lens, now know as theFOV Boost lenswith 182-degree field of view.WIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDNew larger sensor for higher-quality video2x zoom and 8K video capableGood low light performanceBetter default color renderingTIREDCan’t activate Zoom while filmingThe Best GoProPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonChevronChevronSave to wishlistSave to wishlistGoProHero 13 Black$430$349(19% off)Amazon$430$380(12% off)GoPro$429$379(12% off)AdoramaWe have a separate guide tochoosing the right GoPro, and as noted above, the n

Caroline HaskinsBusinessMay 19, 2026 10:00 AMThe Zuckerbergs Are Hiring a Lifeguard but Calling It a ‘Beach Water Person’The job, which is associated with the Zuckerberg family office, is located in Kauai, Hawaii, where the Meta CEO owns a massive compound.Photograph: Jason Henry/Bloomberg/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyMeta CEO MarkZuckerberg and his wife Priscilla Chan are hiring a seasonal, on-call “beach water person” based in Kauai, Hawaii, where the family owns asprawlingcompound, according to a new job listing on Greenhouse associated withWest 10, the Zuckerberg family office.This is an interesting choice for a job title, because according to the job description, the primary duties of this “beach water person” include serving as a “beach lifeguard” and “pool lifeguard.” In other words, being a lifeguard.The job listing also mentions serving as a “boat deckhand” and names a few additional duties, such as instructing “stand-up paddleboarding (SUP), canoe paddling, snorkeling, and other ocean-based activities.” These, however, come after the water safety responsibilities in the job description.Got a Tip?Got a tip? Are you a current or former employee of the Zuckerberg-Chan family? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at carolinehaskins.61.This position easily could have been called “pool/beach lifeguard” or simply “lifeguard.” For the sake of comprehensiveness, “pool/beach lifeguard and boat deckhand” would have also worked. Alternatively, the Zuckerbergs could have chosen “beach/pool attendant,” a job title roughlysynonymouswith lifeguard that could reasonably be interpreted as encompassing extra duties associated with leisure, such as tending to a boat or teaching people how to paddleboard.Arguably, any of these options would have provided more clarity than “beach water person,” whichdoes not appearto correspond with a job title anywhere else in the English-speaking world.WIRED did not immediately hear back from representatives of the Zuckerberg family. Lacking a human to speak with, we decided to ask Meta’s AI chatbot, “What is a ‘beach water person’?”“Beach water person’ would just mean someone who loves being in/near the ocean,” the chatbot said. “The word for that is thalassophile—‘a person who loves the seas and oceans.’” OK!

Maxwell ZeffBusinessMay 19, 2026 11:00 AMFormer OpenAI Staffers Warn That xAI’s Poor Safety Record Could Complicate SpaceX’s IPOThe ex-employees, who cofounded a new AI watchdog group, say investors deserve more information about xAI’s safety practices before SpaceX goes public.Photograph: Miguel J. Rodriguez Carrillo/AFP/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyTwo former OpenAIemployees and a group of AI safety nonprofits are warning that Elon Musk’s AI lab,xAI, could become a liability for prospective investors inSpaceX,which is preparing to file what’s expected to be the largest initial public offering in Wall Street history.In aletterdirected to investors published on Tuesday, the ex-staffers highlighted what they describe as “unpriced risks” related to xAI that could complicate SpaceX’s reported plans to raise up to$75 billionas part of its IPO. The rocket company’s private valuation shot up to over $1 trillion after itacquired xAI last year. Musk claimed his rocket company could launch data centers into space for his AI lab, but the letter’s authors argue that xAI’s poor record on safety issues could complicate how investors view the combined company as it gets ready to submit itsIPO prospectus filing.One of the letter’s signatories and coauthors is a new nonprofit calledGuidelight AI Standards, which was cofounded by former OpenAI safety researcherSteven Adlerand former OpenAI policy adviser Page Hedley. The group, which is backed by private donors, aims to improve the safety practices of frontier AI companies. Other AI safety nonprofits also signed on, including Legal Advocates for Safe Science and Technology, Encode AI, and The Midas Project.Hedley tells WIRED in an interview that he believes xAI has the worst safety practices “nearly across the board” compared to other frontier AI developers, including OpenAI, Google DeepMind, and Anthropic. As a result, he argues, SpaceX may face a greater risk of regulation and litigation than other AI labs.The letter’s authors argue that SpaceX should make several disclosures to investors, including whether xAI intends to continue developing frontier AI models. SpaceX recentlystruck a deal to sell a significant portion of its GPU capacity to Anthropic, and the letter claims the agreement “leaves it unclear whether xAI is still a frontier-AI competitor inside a larger holding company.” If xAI continues to develop frontier AI models, the authors say, it should be required to publish a public safety and governance plan.SpaceX and xAI did not immediately respond to WIRED’s request for comment.The letter also outlines examples of how xAI has not kept up with industry-standard safety practices such as publishing detailed frameworks for mitigating risks around its AI models being used in cyber attacks. The authors also outline specific safety incidents at xAI that they say warrant additional scrutiny. Among the most notable include when xAI’s flagship AI chatbot, Grok, spontaneouslybrought up white genocidein its responses. In another case, xAI allowed Grok to generate thousands ofsexualized images of women and children, which spread widely across Musk’s social media platform X. The latter case promptedat least 37 US attorneys generalto send a letter demanding that Musk’s AI lab take steps to protect women and children on its platform.Hedley says the number of safety incidents xAI has experienced and the regulatory attention they received is “far out of proportion to its market share.” As lawmakers grow increasingly alarmed by the cyber capabilities of advanced AI models likeAnthropic’s Claude Mythos, new security regulations may be on the horizon. The Trump administration is reportedly alreadyweighing an executive orderthat would give US intelligence agencies more oversight over AI models.“It takes serious investment to rein in [AI safety] risks, and it seems that xAI has historically under-invested here,” says Adler. The letter cites reporting from The Washington Post that said xAI had just “two or three” people working on safety as of January. “A question investors should be wondering is if xAI stays at the frontier, how costly might it be to, in fact, manage these [risks] responsibly? If they don’t, what might be the consequences?”The letter acknowledges that Musk’s AI lab has made some improvements related to safety recently, such asexpanding an existing agreement with the White Houseto allow tests of its AI models before they’re publicly deployed. However, the authors say more disclosures are needed in order for investors to be able to fully assess the AI safety risks associated with SpaceX.“xAI’s historical record has been serious enough to warrant scrutiny; it has not, by itself, foreclosed a better future for the company,” the letter reads.New WatchdogAdler and Hedley say Guidelight AI Standards is aiming to create new uniform benchmarks that AI labs can reasonably follow. They also want to offer clear assessmen

Reece RogersGearMay 19, 2026 1:45 PMGemini Spark Is Google’s Response to OpenClaw’s 24/7 AI AgentGoogle’s always-running, data-hungry AI agent is designed to spend your money and send your emails.Photograph: Camille Cohen/AFP/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGemini Spark isGoogle’s take on a steroided-out assistant agent that knows everything about you, announced as part of the company’s updates to its Gemini chatbot app at this year’sI/O developer conference.Software companies have been talking upAI agentsfor some time now, but I wasn’t impressed until I triedAnthropic’s Claude Coworkin January. I sat back as the bot organized the scattered screenshots littering my desktop into labeled folders without a single click, and felt convinced that this might be a turning point for how people interact with their computers.Many other early adopters in San Francisco experienced similar moments when they set up the mega-viralOpenClaw botearlier this year, not just to help complete a few tasks but to run their whole online lives. Power users attempted to fully automate their inboxes, calendars, and text messages, and evenrun a vending machineto varying levels of success via OpenClaw. It’s not without risks—you have to give these agents control of your data and computer, and OpenClaw almostdeleted an entire trove of emailsfor one Meta employee who was experimenting with itWhether it’s my daily schedule via Google Calendar or my date-night dinner spots throughGmailconfirmations, Gemini Spark can dive deep into the well of my personal info before I even connect to a third-party integration. While the standard Gemini app can complete many of the same tasks, Sparks’ differentiator is that it proactively gathers details and takes action while you’re away, rather than waiting for you to prompt it.Google pitches Gemini Spark as a one-stop shop for completing tasks people previously handled manually or did in other apps. The agent can look throughyour credit card billregularly to flag surprise fees—sorry,RocketMoney app, won’t be needing you anymore. Spark can be calibrated to automatically skim every email about your preschooler and highlight key dates for a morning digest report. You can even throw all your meeting notes at Spark and ask it to draft a Google Doc and generate follow-up emails to the right people.This agent is getting a slow rollout, arriving for a small group of early testers this week and launching next week in beta for subscribers to Google’s $100+ per month AI plan. It’s pricey to be one of the first people to experiment with Spark! The company plans to allow Spark to connect through Gemini to third-party apps, like OpenTable and Instacart, for additional automation opportunities in the coming weeks. Other features imminent on the Spark road map include allowing the agent to manipulate your local browser and the ability to text or email commands to the agent.Being able to text commands to your agent sounds like a key factor in actually making the Spark experience feel seamless. Rather than opening the Gemini app and getting distracted, I’ll spend all day texting Spark my increasingly niche requests, as if it were assistant Andrea fromThe Devil Wears Prada.One of the main measures of success when trying this agent will be how often it goes off the rails. “Spark operates under your direction,” reads Google’s announcement blog about the agent. “You choose whether to turn it on and what apps it connects to, and it’s designed to ask you first before performing high-stakes actions like spending money or sending emails.” Anyone who tries the tool is taking a risk by using experimental software that’s powered by personal data.Google plans to expand the agentic shopping feature to allow users to set spending limits and preferred merchants that Spark will adhere to, though exercising caution is critical. “We think of it as if you’re giving a teenager their first debit card,” says Josh Woodward, vice president of Google Labs and the head of the Gemini app.Much like the changes Google is implementing in Search, which brings agentic task automation without needing to leave the search experience, Spark is Google’s chance to push AI agents further into the public zeitgeist. Let’s see if it has the necessary spark to pull it off.

Julian ChokkattuBoone AshworthGearMay 19, 2026 1:45 PMHands-On With All of Google’s New Upcoming Android XR Smart GlassesHere’s your first look at smart glasses coming from Warby Parker and Gentle Monster, powered by Google and Samsung’s XR platform.Photograph: Julian Chokkattu/WIREDCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle has beenteasing its Android XR-poweredsmart glassesfor the past year and a half—ever since it first announced the newmixed reality platform in December of 2024. Now, we’re finally starting to see more polished versions of the hardware, several of which will actually go on sale this fall.AtGoogle I/O, the company gave us a first look at thedesigns of smart glassescoming from established eyewear brands Warby Parker and Gentle Monster. These smart glasses are codeveloped by Samsung and Google, while the eyeglass companies handle the design of the frames.Xreal also took the stage at I/O to showcase its upcoming Project Aura, which is essentially a miniaturized glasses version of bulky headsets like theApple Vision ProandSamsung’s Galaxy XR—powering a full Android app interface with hand gestures for interaction.Audio-only smart glasses from Warby Parker and Gentle Monster will launch later this year. More complex versions with displays built into the lenses will arrive soon after. Xreal’s Project Aura will also come this fall.I got a chance to wear some early versions of these frames and try out a few features in a controlled demo experience.Smarter GlassesPhotograph: Julian Chokkattu/WIREDI wasn’t able to see the final polished designs from Warby Parker and Gentle Monster. Instead, my demos were on unfinished reference glasses from Samsung and Google. The most remarkable thing about them is how light they are. Google says much of the work it and Samsung have done was to miniaturize the technology to save weight. The arms of the glasses are still a little chunky, though.Audio quality was equally impressive. As soon as I asked Gemini to play some Radiohead, the audio surrounded my head and sounded dynamic, though I was in a quiet room. I asked someone else to try them so I could see if I could hear it at 50 percent volume while sitting across from them. I could barely hear the music.The audio-only glasses will be the first to arrive later this year. And yes, all of these upcoming smart glasses—even the audio-only ones—have cameras. That’s how the Gemini assistant can see what you’re seeing to provide helpful, contextual answers and services. The versions with more advanced optics will provide a richer experience. For example, if someone talks to you in a different language, the Gemini assistant inside all of the glasses can translate their voice and even make itsoundlike the person speaking, but the display versions of the glasses will offer accompanying text to read as well.Courtesy of GoogleCourtesy of GoogleThe versions with a display built into the lenses should be especially helpful with things like turn-by-turn navigation, since you’ll be able to see exactly where to turn left on the display as you walk. Google says the glasses don’t have GPS and rely on your phone’s GPS, but the cameras on the glasses utilize Google’s Visual Positioning System to identify surroundings and ensure your blue dot is in the right spot—something you can currently do in the Google Maps app to calibrate your compass.I tapped and held on the right arm of the glasses, and this triggered Gemini Live. That’s the default experience, allowing you to have a conversational back and forth with Google’s assistant.Gemini did a good job of only listening to my voice in these encounters and ignoring people who were speaking to me. I was able to ask it what the board game next to me was—it correctly identified it as Chinese checkers—and then asked if I wanted to learn how to play. I asked it to just save me those instructions in a Google Keep note, and voilà, it generated that advice in a Keep note within seconds.Photograph and composite by Julian Chokkattu/WIREDNext, I was able to ask Gemini to take a picture and then alter the photo post-capture. I first asked it to remove a plant in the photo, then I asked it to change the room’s decor to a medieval hall. It did those things remarkably well.You’ll get the original image on your phone and a preview on Wear OS smartwatches if you’re wearing one (or a preview on the display-laden glasses), and the doctored version will arrive within 45 seconds once Google’s Nano Banana platform does its thing.I was able to see a few generative widgets and scroll through them on these glasses—that’s a big theme Googleis introducing in Android 17and other platforms—so you can have glanceable information on your glasses for whatever you want. I wasn’t able to see much else of the user interface, though.Project AuraPhotograph: Julian Chokkattu/WIREDA different kind of glasses experience, Project Aura is more akin to mixed-reality headsets like Apple’s Vi

Reece RogersGearMay 19, 2026 1:45 PMGoogle Makes It Easy to Deepfake YourselfGoogle’s overhaul of its AI creation software, Flow, includes a new video model and a tool for generating selfie videos called avatars.Courtesy of GoogleCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyA wave ofdéjà vu washes over me as Elias Roman, vice president of product management atGoogleLabs, demos a new “avatar” feature for Flow, the company’s tool that lets users generate and remix AI videos and images. He previously scanned his likeness to create a digital clone of himself. Now, he can insert himself into any AI-generated videoclip he wants using Google’s new Omni Flash model.“This is for creators who want to bring themselves into their content but don’t want to have to shoot themselves,” Roman says.Courtesy of GoogleThis specific style of social-first, selfiedeepfakeis reminiscent of a quintessential feature from OpenAI’s now-defunctSora app—rather than cameos or characters, Google calls them avatars. These avatars are also available through the Gemini app andYouTube. Google announced the new feature at its annualI/O developer conferencein Mountain View, California.Google launched Flowlast yearunder its experimental Labs division. “Google has never had a product line for creative work before,” Roman says. “Productivity, definitely. Developers, absolutely. Video consumption, yes. Not for creative work.” He sees this as Google’s attempt to build tools for the next generation of creators.Similar to other announcements from Google I/O surrounding Google Search, many of the new changes toFloware part of the company’s larger attempt to makeAI agents, essentially automated software taskmasters, andvibe coding, building bespoke features with natural language prompts to AI, more mainstream for a broader audience. For example, users can repeat custom instructions when generating videos and create automated workflows that sort similarly styled clips into folders.Courtesy of GoogleOne of the most immediately noticeable changes to Flow is the new video-generation model powering the experience: Omni Flash,succeeding Veo. Similar to howGoogle’s Nano Bananamodel brought more context about the world into the AI image-creation process, the Omni Flash model overhauls video generation with richer detail throughout clips.Flow users can generate characters in AI videos with more consistency via the Omni Flash model. Roman says this is a major improvement over the weakness in past versions of Flow, where created characters could warp during successive video generations. Also, a key character that Flow users can now generate in an AI scene after an AI scene? Themselves.Users set up an “avatar” of themselves by going into the settings of their Flow account and scanning a QR code on their phone. Then, Google asks users to record themselves saying a string of numbers aloud and move their head around to capture every angle. This selfie-capture style will feel familiar to anyone who signed up for the Sora app, which OpenAI launched last year as an AI-firstsocial mediaplatform where people can generate and share clips of themselves. OpenAIstartlingly wound it downafter less than seven months.Unlike the Sora app, where users could generate videos of other users depending on the person’s settings, Google’s initial focus with its avatars is to let users create AI versions of themselves only, not other people. Every video generated with the Omni model, including those with your avatar, includesGoogle’s SynthID watermark.“You can capture your voice and your visual identity from multiple angles and have that show up with pretty high levels of fidelity,” Roman says. He generated a tongue-in-cheek video of himself teasing the Flow team in front of a dumpster fire, with an AI version of himself that looked lifelike and sounded like him. Then he used Flow to request changes to the generation, like a new background setting and a different-colored shirt, and Omni Flash adjusted the clips while preserving the avatar’s details.This isn’t the first time Google has rolled out a version of self-controlled deepfake video tools for creators—last month,YouTube Shorts addeda limited option for users to make similar AI avatars that can be inserted into clips on that platform. Other Silicon Valley companies are also looking for ways to transform creators’ outputs using generative AI. For example, last year, Meta rolled out an AI feature that can seamlessly translate Instagram Reels into different languages, evenadjusting creators’ lipsto match the different voices.While these AI tools may streamline aspects of the content production pipeline for creators—you don’t even have to get out of bed now to generate sassy vertical videos—generative AI isincreasingly polarizingaudiences who see these videos as inauthentic or not aligned with their values. Well, that’s if they actually clock the videos as AI.

Will KnightBusinessMay 19, 2026 1:45 PMDemis Hassabis Thinks AI Job Cuts Are DumbThe CEO of Google DeepMind tells WIRED that companies should use the productivity gains of AI to do more, not lay people off.Photograph: Chris Jung/NurPhoto/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyDemis Hassabis, theCEO ofGoogle DeepMind, is keen to talk about the coding skills of his company’s newest model,Gemini3.5 Flash. The model has been trained to perform complex agentic coding tasks: translate large code bases from one language to another; find and fix bugs lurking deep in knotty code; and even write entire operating systems from scratch.Hassabis does not, however, think this spells doom for software developers. “I have no idea why people are going around talking with certainty about that,” Hassabis tells WIRED ahead of the new model reveal attoday’s Google’s I/O event.“Perhaps there is an ulterior motive for putting those messages out; raising money or whatever,” Hassabis says. “From my point of view, from DeepMind and Google’s point of view, if engineers are becoming three or four times more productive, then we just [want to] do three or four times more stuff.”Got a Tip?Got a tip? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at wak.01.The striking coding abilities of the latest models has led to widespread fear that AI may be on the brink of eliminating programming roles and other white-collar jobs. Executives at some AI companies have predicted widespread job displacement, while some prominent tech companies, includingAmazon,Salesforce, andBlock, haveblamed recent layoffson the use of AI.Hassabis thinks that Alphabet, which oversees several companies besides Google, may be well positioned to take advantage of a revolution in software productivity. “I have a million ideas, from lab drug discovery to game design,” he says. “I’d love to have some free engineers to go and do those kinds of things.”Hassabis says that companies looking to replace developers with AI may be making a big mistake. “I think it’s a lack of imagination—and a lack of understanding of what’s really going to happen,” he says.Google revealed a raft of AI stuff at its annual developer event. Through a coding tool called Antigravity, Gemini 3.5 Flash offers frontier coding and reasoning capabilities but is faster and cheaper than the offerings of its competitors, Google says. Gemini 3.5 Pro, a more powerful new version of its flagship model, will debut next month.The company needs to catch up when it comes to AI coding, which has emerged as a crucial and lucrative application for the latest AI models. Anthropic and OpenAI lead developer adoption with their respective tools, Claude and Codex, according toa 2025 Stack Overflow survey.The company also demoed an agentic assistant called Spark that lives in Google’s Cloud and has access to its apps. The design is meant to be safer than something like OpenClaw because it has limited access to personal data, Google says.Other agentic demos included a version of Android with an AI agent built in and a refreshed version of Google Search that uses agentic coding to generate a site or app on the fly in response to a search query.AI coding has captivated the AI world in recent months, even inspiring hope that models could one day rewrite their own code in a self-improvement loop. Hassabis says it’s possible but doubts that it will immediately lead to superhuman-level AI.Progress in other areas of science might require AI models to have a deeper understanding of the physical world and even an ability to perform experiments within it, he says.Even within the seemingly solved world of coding, Hassabis says it is notable that AI has yet to produce a blockbuster app or video game without human help. “I think there’s something missing,” he says.

Reece RogersGearMay 19, 2026 1:45 PMGoogle Search Goes Agentic—and Doesn’t Need You AnymoreVibe-coded results! Super widgets! Bots that never sleep! Google’s vision for the future of Search is hyper-personalized, automated, and extremely AI.Photograph: Thilina Kaluthotage/NurPhoto/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAI agents areeverywhere. Every briefing I’ve attended for software companies over the past year has involved some mention ofagents—using generative AI tools to automate digital tasks. Despite breakout moments at the start of 2026, like the pluckyOpenClawagent that early adopters used to manage their online life, most people are not yet embracing this style of automation day-to-day. That won’t last for long if Google gets its way.AtGoogle I/O, the tech behemoth shared its vision to make its popular search engine a core way to expose billions of existing users to the company’s agentic prowess.“You will be able to create, customize, and manage multiple AI agents for your many tasks, right in Search,” says Liz Reid, who leads Search atGoogle. She gives the example of setting up an agent to track stock market trends and send you alerts using real-time data—when specific conditions are met.Alongside these agentic additions coming to Search, Google also announced a new underlying model, Gemini 3.5 Flash, as the global default model forAI Mode answers,as well as improvements to the Search box that make it more responsive to user inputs.Always-On AnswersOne of the key agentic experiences rolling out is a new form of dynamic data gathering through “information agents” that can be more proactive than previous search experiences and automate alerts through your conversational requests.“Ask Google to just keep you updated on anything, and now our agents can do work for you even if you’re not using Google,” says Robby Stein, a vice president of product for Search. “So, you could be asleep, and it’s still helping you.” This feature will arrive first for subscribers toGoogle’s AI Pro and Ultra plansthis summer.In Google’s example of how these can work, a user asks AI Mode through the Google app to keep them “updated when any of my favorite athletes announce sneaker collabs or signature drops.” Then, AI Mode generates an info agent unique to that user, designed to monitor the request. When a new shoe drops that fits these criteria, likeA’ja Wilson’s pink Nikes, the user gets a notification alert with critical context and ways to buy the sneaks.Booking agents are another automation style in Search that Google will be expanding this summer. At previous I/O conferences, Google debuted similar AI features, like the now defunctDuplex, which called companies on your behalf to set up restaurant reservations or salon appointments. Google continues to iterate on that core idea with agents that can search for relevant context about local companies, even calling that barber down the street for a price quote on a beard trim if it doesn’t show up on the website, to help users gather info with less active participation.Much of this keeps users trapped withinGoogle’s ecosystemwithout actually browsing the web themselves, a practice we’ve seen before when features likeAI Overviewslaunched in 2024. Even so, Google has explicitly stated that it’s not attempting to replace web page links.Vibe-Coded ResultsGoogle first launched Antigravity, its version of an agentic coding tool likeOpenAI’s CodexorAnthropic’s Claude Code, late last year. Now, an updated version of Antigravity will be used to create bespoke outputs in Search, like an adjustable visualization of black holes to help you better understand deep space.“Search can build you custom experiences,” Stein says. Google’s vibe-coding tool, Antigravity, will craft unique answer layouts, when helpful, and incorporate interactive visual elements, such as infographics and movement simulations. The update is expected to arrive later this summer for all users.An expanded version of that tool, which Google says is a way to create “super widgets” and “mini apps,” will be available first for Google AI plan subscribers in the US this summer. (I vote to call them super widgets, which sounds more unhinged.) These can be used for tracking longer-term plans, like a healthy diet, or managing complex tasks, like a cross-country move, through Search.Search Doesn’t Need YouGoogle of yore, with its 10 blue web links, asked users to be fairly active participants in the info-gathering process:Here are a bunch of webpages that appear relevant to your search. How about you click around and see which site has the best info?The process was clunky at times and could be time-consuming, but I was personally hopping site to site as Google nudged me into different online rabbit holes.The launch ofAI Overviews, which generates summaries of web pages and online information, two years ago, and the rollout ofAI Mode, the chatbot-style search, last year, have au

Paresh DaveLauren GoodeZoë SchifferBusinessMay 19, 2026 3:00 PMMeta Employees Are Scrambling to Use Up Benefits Ahead of LayoffsOn the eve of about 8,000 jobs being cut, employees are cashing in on headphone stipends and other perks while they still can.Photograph: Benjamin Fanjoy/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAhead of Meta’slatest round ofmass layoffs tomorrow, some employees are deserting offices, abandoning their work, and loading up on perks they might soon lose, several people at the company tell WIRED.Two employees describe a widespread rush to use up an annual $2,000 flexible benefit, which can cover a variety of expenses including health and wellness activities. A separate triennial credit of $200 toward the purchase of audio gear has led to a scramble to purchase Apple AirPods and other headphones. Another source says Meta offices have been largely empty this week, as people prioritize polishing their résumés and gather offsite to commiserate with friends for what may be their final time as colleagues. Employees are variously “paralyzed,” “coasting,” and “panicked,” sources say.Meta plans to lay off about 10 percent of its nearly 80,000 employees on Wednesday, with notices going out to affected workers’ personal and corporate email addresses at 4 am in their local time zone, according to a company-wide memo sent on Monday. The cuts are coming at a time when the social media giant behind Instagram, WhatsApp, and Facebook is enjoying record-high profits.But CEO Mark Zuckerberg insists that the company must free up cash to invest in AI data centers, and that Meta can perform just as well with fewer employees because of AI technologies that augment human labor.Got a Tip?Are you a current or former Meta employee who wants to talk about what’s happening? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at peard33.24 and ChaoticGoode.12.Meta didn’t immediately respond to a request for comment for this story. The company has undergone three previous large rounds of layoffs since 2022, including as part of Meta’s one-time “year of efficiency” drive in 2023. But even though the latest round is smaller than a couple of those, it is drawing widespread scrutiny because it comes at a time of societal anxiety aboutAI’s impact on jobs.Inside Meta, the imminent cuts are among several concerns that have sunk morale to unprecedented depths, according to 16 current and former employees whorecently spoke to WIRED. Employees also have been frustrated by being “drafted” onto a new AI team without any choice andthe rollout of surveillance softwarethat tracks US workers’ laptop use to train AI models.Meta also plans to internally restructure as it conducts sweeping layoffs, transferring 7,000 remaining staff to “AI initiatives” and converting more managers into individual contributors. That would bring the total number of those affected—either laid off or placed in a new role—to 20 percent of the current workforce, Reutersreportedon Monday. WIRED independently confirmed this reporting. Some parts of the company have been told they won’t be affected at all.But in recent days, employees who are bracing for changes have shared checklists internally about benefits to take advantage of, and are saving documents such as performance reviews and pay stubs, according to one worker. Some teams are meeting up at bars and restaurants near Meta offices in New York and Menlo Park on Tuesday and Wednesday to eat and drink away their sorrows, several employees said. Management has encouraged employees not to come into offices on Wednesday.

Boone AshworthMichael CaloreGearMay 19, 2026 4:00 PMEverything Announced at Google I/O 2026Google is sprucing up its Gemini models, revamping search, and enabling AI agents in everything. There are also some spiffy new smart glasses coming this fall.Google and Alphabet CEO Sundar Pichai on stage in 2025.Photograph: Camille Cohen/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle just wrappeditskeynoteaddress at its annualI/O developer event. The company showed off a swath of newagentic AIfeatures and some demos of its upcoming Android-powered smart glasses.As it has in the past few years, the spectacle largely revolved around Google’s perpetual stream of AI efforts. The company says that 900 million people use its Gemini assistant, and people have generated more that 50 billion images with Gemini.Google’s goal for 2026 is to put AI agents at the forefront of all its biggest services: Search, Gmail, YouTube, Docs, and the Chrome browser. In a demo briefing the day before I/O, Google and Alphabet CEO Sundar Pichai said the company was in a period of “hyper progress” with its AI efforts, but acknowledged that this is the part of that cycle “where people want to see real value in the products they use on a day-to-day basis.”Here is everything Google announced at I/O 2026. And if you’re wondering where all the Android 17 news is, Googletold us all about that last week.Search PartyGoogle is going all in on keeping people on Search by trying to integrate it into every part of your life. The company hasembedded its AI agents directly in Google Searchin something it calls the “intelligent search box.” This new search experience starts rolling out to everyone today.When you ask the search box questions like, “What is a black hole?” or “Are Google AI Overviews devastating the journalism industry,” it will respond with more contextual answers or use generative AI to create images or short video clips that help explain the concept. As an example, Google showed a search about black holes, and its AI agent generated a video that visually explains the process, plopping it right into the search results.Search is also getting a Generative UI feature, which creates different ways of viewing information from search results, so different types of responses—videos, images, news articles—get custom layouts that are generated in the browser on the fly, based on what’s the most relevant. Generative UI will start showing up in everyone’s Google searches this summer.Search agents are also being integrated more deeply across Google’s platforms. In March, Google releasedAsk Maps, which lets users ask questions on Maps like you would interact with a chatbot.Gemini EverywhereGoogle has two new major AI model updates. Gemini 3.5 is officially being released today, along with Gemini 3.5 Flash, a pared-down but more affordable version of 3.5 Pro. Both will be available immediately in Google’s Search and the Gemini app.Google’s core Gemini app is getting a refresh, with a snazzier redesign called Neural Expressive. It has new colorful backgrounds, a new typeface, and new animations for live voice chats. It also has many more accent options for your chatbot’s voice, with regional accents in several languages.Daily Brief is an upcoming personalized digest of your day that Google wants to be the first thing you check in the morning. Daily Brief pulls info and data from your calendar, email, and other information to summarize and prioritize your plans for the immediate future. (Google tested out a similar feature in its experimentalGoogle Labsin December.)The company is also adding new Gemini-powered features to its other services. An Ask YouTube feature lets you pose natural search queries in YouTube. Ask it how to ride a bike or how to grill a steak, and an agent will then be tasked with finding a mix of relevant YouTube videos. It’ll even snap right to the points within the videos that answer your questions.Much weirder is voice editing in Google Docs, in a new feature called Docs Live. By describing with your voice what you want to write, an agent will dictate your words, generate text, pull in citations from the web, and aim to turn your stream-of-consciousness wishes into a coherent document.(Reminder: All this stuff mayeventually have ads.)For Gemini power users, Google is creating a new subscription tier, the AI Ultra plan, for $100 a month. It is also dropping the price of its top Gemini AI Ultra from $250 a month to $200.Gemini OmniGoogle announced Gemini Omni, an AI video generator akin to Sora 2. That was OpenAI’s generator that let youdeepfake yourselfbut was eventuallykilled by the company.Google’s approach is building out a far more realistic video generator that can incorporate real video and extrapolate all manner of AI-powered weirdness on top of that. Google is eager for you to turn Omni’s eye on yourself, putting your face front and center. As such, selfie videos can be modified to

Brian KahnScienceMay 19, 2026 6:20 PMCalifornia’s Wildfire Season Is Already OveractiveMajor fires are threatening homes and ecologically sensitive areas following a hot, dry winter.The Sandy Fire burns on Monday, May 18, in Simi Valley, California.Photograph: AP Photo/Ethan SwopeCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyIt’s May, butCalifornia is already getting a taste of what peak fire season could look like, as out-of-control blazes pose a danger to infrastructure and some of the most threatened trees on the planet.A combination of high winds and heat has contributed to a trio of major fires in Southern California.The largest among them is the Santa Rosa Island Fire, which started over the weekend in Channel Islands National Park after a stranded sailor used flares to signal for help. The blaze has consumed roughly 16,600 acres—almost a third of the entire island. While some structures have been lost, the biggest is a grove of Torrey pines, which are among the rarest trees in the world.Torrey pines are considered critically endangered by the International Union for Conservation of Nature. Thegroup has warnedthat the grove on the island faces “high potential risk from destructive fires.” While the Santa Rosa Island Fire burned through the grove, there’s hope that the worst-case scenario may not have come to pass.“Upon initial assessments, fire crews are observing that the fire intensity was low and that the stand remains intact,” Mike Theune, the fire information officer assigned by the federal government, writes in an email. “When safe to do so, a fire effects crew will be assigned to make a full determination of condition and any long-term effects.”While smaller, the Sandy Fire has forced thousands to evacuate in and around Simi Valley, located about 35 miles northwest of downtown Los Angeles. Fire activity increased on Tuesday morning as high winds whipped the flames across nearly 1,400 acres.The fire is only 5 percent contained, while the Santa Rosa Island Fire isn’t contained at all. Because homes and businesses are threatened, the Sandy Fire has received more aerial resources to quell the flames. The River Fire has also burned 3,535 acres in Kern County and is 15 percent contained.In the early stages of fire season, nearly 41,000 acres have burned across the state, according to the California Department of Forestry and Fire Protection. But that’s well above the five-year average of 23,380 acres burned by this date and could be a harbinger of what’s to come in the state and across the West.Record-breaking heatobliterated already-meager snowpackthis spring, leaving states primed to burn. The most recent snowpack measurements show California’s Sierra Nevada range harbors just 9 percent of its usual snowpack for this time of year, while many basins in other parts of the West have no measurable snow at all, according to federal data.Hot weather is, of course, a hallmark of climate change. Arecent analysisby the nonprofit Climate Central found that April 1 snowpack—a key date for measuring what’s on the ground—has declined across the West by 18 percent since 1955.All of which is to say that while the early season fires are bad, the West hasn’t seen anything yet.

Miles KleeCultureMay 19, 2026 6:53 PMLiterary Prizewinners Are Facing AI Allegations. It Feels Like the New NormalThree of five regional winners of the prestigious Commonwealth Short Story Prize are suspected of relying on chatbots. They’re certainly not alone.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAt first, thewinners of the prestigiousCommonwealth Short Story Prizefor 2026 enjoyed the envy of their peers. But since their works of fiction earned this distinction, these authors have found themselves facing harsh scrutiny from the literary community, with several accused of enlistinggenerative artificial intelligenceto write for them.The allegations have come from numerous readers, many of them writers themselves, expressing bafflement and dismay that the prize jury could have overlooked potential signs of inauthentic authorship.Each year, the Commonwealth Foundation, a nongovernmental organization in London, awards its short story prize to one writer in each of five regions: Africa, Asia, Canada and Europe, the Caribbean, and the Pacific. One overall winner is then selected from that short list. Regional winners take home £2,500 (about $3,350), while the top winner, to be announced next month, claims £5,000 (about $6,700).On May 12, the respected UK literary magazineGranta published the top five 2026 entries—all previously unpublished, per the rules of the contest—on its website. (It has hosted the winning submissions for the prize since 2012.)Within days, however, one entry aroused suspicion. “The Serpent in the Grove,” a story by Jamir Nazir of Trinidad and Tobago, which had taken honors for the Caribbean region, struck a few people as bearing the stylistic tells of AI-generated text.“Well, this is a first: a ChatGPT-generated story won a prestigious literary prize,” wrote researcher and entrepreneur Nabeel S. Qureshi, a former visiting scholar of AI at the Mercatus Center at George Mason University, in aposton X on Monday. “‘Not X, not Y, but Z’ sentences everywhere, the ‘hums’ trope, and plenty of other obvious markers of AI writing. A major milestone for AI, at any rate…”“They say the grove still hums at noon,” Nazir’s mysterious and atmospheric tale begins. In his screenshot of the opening paragraphs, Quereshi highlighted the second line as what he considered to be a signature example of AI syntax: “Not the bees’ neat industry or the clean rasp of cutlass on vine, but a belly sound—as if the earth swallows a shout and holds it there.”As the literary community undertook a closer read of Nazir’s story, many criticized itslanguage and metaphorsas nonsensical, wondering how the Commonwealth judges could have seen any merit to them. Others sharedscreenshotsshowing that the AI-detection tool Pangram flagged “The Serpent in the Grove” as 100 percent AI-generated, a result that WIRED independently confirmed. (While no AI-detection software is perfect,third-party analysishas consistently determined Pangram to be the most accurate, with a near-zero rate of false positives.)Nazir did not return a request for comment relayed through an email address listed on hisFacebook page. The posts on that account and theLinkedIn profileof a Jamir Nazir in Trinidad and Tobago also scan as AI-generated on Pangram. Although some speculation had it that Nazir himself could have been an entirelyAI-created persona, a2018 articlein the Trinidad and Tobago edition of the The Guardian about his self-published poetry collectionNight Moon Love—which includes a photograph of Nazir holding the book—suggests that he is a real person.WIRED contacted both Granta and the Commonwealth Foundation about Nazir’s story; neither commented directly, but both issued public statements.‘We are aware of allegations and discussion regarding generative AI and our Short Story Prize,” wrote Razmi Farook, director-general of the Commonwealth Foundation, in astatementon the organization’s website. “We take these claims seriously and are committed to responding to them with care and transparency.” Farook defended the judging process for the prize as “robust,” with multiple rounds of readers and the top-level judges selected for their “expertise.”“We do not currently use AI checkers in our judging process, because this is a prize for unpublished fiction,” Farook explained. “To supply unpublished original work to an AI checker would raise significant concerns surrounding consent and artistic ownership. We also do not use AI to judge stories at any stage of the process. When they submit stories to the Prize, writers accept our entry rules and guidelines. These include confirming that their submission is their own original work. All short-listed writers have personally stated that no AI was used and, upon further consultation, the Foundation has confirmed this.”Theentry and eligibility rulesfor the 2026 Commonwealth Short Story Prize make no mention of artificial intelligence, stat

TechGadgetsNewsUgreen’s new soccer ball-shaped tracker has up to 7 years of battery lifeBut when its non-replaceable battery finally dies, you’ll need to kick the FineTrack 2 into the e-waste bin.But when its non-replaceable battery finally dies, you’ll need to kick the FineTrack 2 into the e-waste bin.byAndrew LiszewskiMay 19, 2026, 7:49 PM UTCLinkShareGiftIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.The FineTrack 2 isn’t as discreet as an AirTag.Image: UgreenAndrew Liszewskiis a senior reporter who’s been covering and reviewing the latest gadgets and tech since 2006, but has loved all things electronic since he was a kid.Ugreen has announced a new Apple Find My-compatible tracker with a novel design that limits where you can hide it, but greatly benefits battery life. Inspired by “major football tournaments” the company doesn’t call out by name, the Ugreen FineTrack 2 is shaped like a soccer ball and is designed to be used hanging off your keys or bags. It’s much larger than an Apple AirTag or credit card-sized trackers, but still small enough to bury in a suitcase or stick in your car’s glove compartment.The waterproof FineTrack 2 isavailable now for $19.99and comes with a 600mAh CR2450 battery good for around five to seven years of use with Apple’s Find My network. Ugreen’s trackers often come with excellent battery life. It updated its 1.7mm thickcredit card tracker with support for Google’s Find Hub networkand five years of battery life last September, but once it dies, it becomes e-waste. It’s the same story for the FineTrack 2. It’s not rechargeable and its battery isn’t replaceable. An AirTag’s battery might only last for around a year, but you can easily swap in a new one as often as you need.The FineTrack Mini 2 is a smaller version of Ugreen’s ball-shaped tracker.Image: UgreenAlongside the FineTrack 2 Ugreen also announced its new FineTrack Mini 2. Its functionality and battery life are identical to the ball-shaped version including five to seven years of (non-replaceable) battery life, waterproofing, and a loud 110dB alarm. But it features a smaller, lighter, and more traditional tracker design that’s more discreet and easier to stash away. It’sonly being sold in four-packs for $45.99.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Andrew LiszewskiGadgetsNewsTechMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

PolicyEntertainmentAntitrustDemocrats preview how they’d go after the Ticketmaster settlement if they regain powerAnd California’s AG says an even broader breakup of Live Nation is on the table.And California’s AG says an even broader breakup of Live Nation is on the table.byLauren FeinerMay 19, 2026, 8:13 PM UTCLinkShareGiftImage: Cath Virginia / The VergePart OfUSA v. Live Nation-Ticketmaster: all the news on the lawsuit that could take down a giantsee all updatesLauren Feineris a senior policy reporter at The Verge, covering the intersection of Silicon Valley and Capitol Hill. She spent 5 years covering tech policy at CNBC, writing about antitrust, privacy, and content moderation reform.A handful of Democrats called an unofficial hearing on Capitol Hill Monday to slam the Department of Justice’s “trivial” and “pathetic” settlement with Live Nation-Ticketmaster, previewing how they might go after Trump administration antitrust deals if they win back congressional power in November.While the DOJsettlementwith Live Nation, which came one week into trial and with a promise of up to $280 million, received copious pushback, Democrats don’t at the moment have the power in either chamber to set committee agendas. That meant that Monday’s so-called shadow hearing looked different than most official proceedings. Instead of sitting at the elevated seats behind the dais, the lawmakers sat at tables below, eye level with witnesses including California Attorney General Rob Bonta, one of the AGs continuing the fight against Live Nation-Ticketmaster, ousted Trump administration antitrust official Roger Alford, and several entertainment industry players, including a member of the band The Hold Steady. Several artists asked to appear at the forum declined out of fear for their livelihoods, according to House Judiciary Committee Ranking Member Jamie Raskin (D-MD), who co-hosted the forum with Senate Homeland Security permanent subcommittee on investigations Ranking Member Richard Blumenthal (D-CT).While their statements won’t be formally entered into the congressional record, the forum gave Democrats a way to publicize their concerns, which included that the DOJ’s deal was the result of undue corporate influence. “This isn’t like some kind of pickup game,” Raskin told reporters. “The corruption permeates the administration so much that we have to start building the record because it’s such an overwhelming task. We will be using the products of all of these shadow hearings to figure out where to go.” As a signal of the significance of the forum, Senate Minority Leader Chuck Schumer (D-NY) delivered opening remarks to kick it off.“The corruption permeates the administration so much that we have to start building the record because it’s such an overwhelming task”The lawmakers pointed to the states’resounding victory in the jury verdictas evidence that theDOJ’s settlementfell far short of what a good faith deal would include, charging it was a result of corruption. On the11-page verdict form, jurors checked “yes” on every question about whether the states proved their claims. The DOJ has stood by its deal as a win for consumers. The deal will still need to undergo a Tunney Act review, which is meant to ensure that it’s in the public interest.Judges are typically loath to reject a settlement, which is why Sen. Amy Klobuchar (D-MN) hasproposed legislationto strengthen the standard. But Alford, who wasfired from the DOJandlater spoke outabout opposing lobbyist influence in antitrust deals, said this could be the rare case where the judge may reject the DOJ’s agreement, given that the same judge is now tasked with coming up with appropriate remedies to resolve Live Nation’s monopoly, potentially including a breakup.“Divestiture, it’s the nuclear option”Bonta said the states’ breakup request, expected to be formally submitted this week, may go beyond a separation of Live Nation and Ticketmaster. They’re also considering asking for at least partial divestitures of parts of Live Nation’s venue or artist management businesses. “Divestiture, it’s the nuclear option. We believe it’s warranted, but we want to be thoughtful how we ask this judge for an appropriate remedy,” he told reporters.“Opinions shared yesterday misrepresent how the live events industry operates,” Live Nation’s executive vice president of corporate and regulatory affairs, Dan Wall, said in a statement. “Secondary ticketing and high-demand onsales are industry-wide challenges that exist across every promoter, venue, and ticketing platform. That’s precisely why Live Nation and Ticketmaster have long supported industry-wide resale reform, including price caps and stronger consumer protections, while continuing to invest in tools that help artists manage demand and protect fans.”Live Nation haspromised to appealthe jury verdict, and even before it does it’s possible the judge will deny the states the full extent of the remedies they ask for. But Monday’s forum showed tha

GadgetsTechVerge ShoppingHere are our favorite Memorial Day deals (so far)We found excellent deals on some of our favorite TVs, noise-canceling earbuds, portable speakers, and other gadgets.We found excellent deals on some of our favorite TVs, noise-canceling earbuds, portable speakers, and other gadgets.bySheena VasaniMay 19, 2026, 8:25 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Portable bluetooth speakers like the Charge 6 and Flip 7 are just a few on sale for Memorial Day.Sheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.Memorial Day marks the unofficial start of summer, and the holiday’s sales include plenty of deals on gadgets that can help you make the most of the season. If your plans involve beach days, pool parties, or backyard barbecues, now’s a great time to pick up aportable Bluetooth speaker, as many of our favorite models are currently on sale. You’ll also find worthwhileTV deals, along with discounts onsmart thermostatsto help keep your home cool and your energy costs low. And, if you’re planning on heading out of town, there are deals onnoise-canceling headphones,portable chargers, and other travel-friendly gadgets.RelatedHere are 40 of our favorite deals from REI’s massive Anniversary SaleThe Verge’s 2026 college graduation gift guideBelow, we’ve rounded up the best Memorial Day deals currently available. To give you more options as you shop, we’ve included good deals from a range of retailers, including ones that aren’t running their own Memorial Day sales.Featured ProductsPreviousNextBose QuietComfort Ultra Earbuds (2nd Gen)Read More$249 at Amazon$249 at WalmartLG B5 OLED TVRead More$800 at Amazon (55-inch)$650 at Best Buy (48-inch)Anker Laptop Power BankRead More$96 at Amazon$96 at WalmartStar Wars OutlawsRead More$36 at Amazon (Switch 2)$20 at Walmart (PS5)Ray-Ban Meta (Gen 1) smart glassesRead More$224 at Amazon$224 at Best BuySonos Move 2Read More$399 at Amazon$399 at WalmartBose QuietComfort Ultra Headphones (first-gen)Read More$299 at Amazon$299 at Best BuyRing Battery Doorbell Plus (second-gen)Read More$120 at Amazon$120 at Best BuyEarbuds and headphonesBose QuietComfort Ultra Earbuds (2nd Gen)$249$29917% off$249The second-gen QuietComfort Ultra Earbuds have dynamic, rich sound and the best noise cancellation currently available in a pair of true wireless earbuds. They’re a terrific pick if you want to listen to your music in peace.Read our review.Read More$249 at Amazon$249 at Walmart$249 at Best BuySamsung’s Galaxy Buds 3 Proare on sale for $149.99 ($100 off) atBest Buy, which matches their all-time low price. The wireless earbuds are ourtop pick for Samsung Galaxy phone device owners, thanks to their comfortable AirPods-inspired design, terrific sound quality, and good noise cancellation. They also offer helpful voice controls that let you adjust volume or playback without a wake word, while delivering excellent voice call quality.Read our review.Nothing Ear (a)$59$9940% off$59Nothing’s Ear (a) earbuds have the same general design as the company’s previous earbuds — but you can get these in a snazzy yellow. They include ANC, more than decent sound quality, and other features like multipoint for just under $100.Read our impressions.Read More$59 at Amazon$59 at NothingSennheiser’s Momentum True Wireless 4earbudsare down to $199.99 ($100 off) atAmazonandBest Buy, which is $20 shy of their best price to date. If sound quality matters most, they’re an easy recommendation, delivering rich, detailed audio with EQ customization to fine-tune the listening experience. Optional wing tips are included to help keep them secure and comfortable.Read our review.Bose QuietComfort Ultra Headphones (first-gen)$299$42930% off$299The older QC Ultra were a replacement for the Noise Cancelling Headphones 700 with a more travel-friendly design, spatial audio, and better call quality. Thankfully, they still offer excellent comfort and some of the best noise cancellation available.Read our review.Read More$299 at Amazon$299 at Best BuySony WH-1000XM5$248$40038% off$248Despite being a last-gen model, these headphones still offer impressive sound quality and top-notch noise cancellation. They’re comfortable to wear for hours, can last up to 40 hours on a charge, and support extras like audio sharing and Gemini Live.Read our review.Read More$248 at Amazon$248 at Best Buy$250 at TargetAmazonandBest Buyare selling thelast-gen AirPods Max (USB-C)for $449 ($100 off), which is $50 shy of their all-time low. Apple’s luxe over-ear headphones offer a premium design and work especially well with other Apple devices. They still deliver great sound quality and effective noise cancellation, even if thenewer modelhas a slight edge.Read our review.Shokz OpenRun Pro 2$130$18028% off$130The OpenRun Pro 2 are Shokz’s best bone conduction headphones yet, with bet

TechAIReportGoogle’s AI future demands trust — and your personal dataGemini Spark is giving Google a whole new level of access to your information.Gemini Spark is giving Google a whole new level of access to your information.byEmma RothMay 19, 2026, 9:00 PM UTCLinkShareGiftImage: GooglePart OfGoogle I/O 2026: All the news and announcementssee all updatesEmma Rothis a news writer who covers the streaming wars, consumer tech, crypto, social media, and much more. Previously, she was a writer and editor at MUO.Google has big promises for its AI-powered future — and a lot of it depends on your trust. At I/O 2026, Google described a bunch of new tools that it claims will make your life easier.Gemini Spark, Google’s always-on AI agent, can help organize an upcoming event, while Daily Brief can offer a rundown of what to expect during your day. Google is even expanding access toGmail’s AI inbox, which can generate custom to-do lists and draft personalized replies based on your emails.Many of these features seem genuinely useful, but at the heart of each of them is an AI engine that runs on a trove of personal information. While other AI companies, like OpenAI, Microsoft, and Anthropic, let you connect other apps and data that you use, Gemini’s access to the personal data already stored across Google’s services lies behind a simple opt-in menu —one of its key advantages in the AI race.You can ask Gemini Spark to review your inbox at a set time each week.Image: GoogleGoogle first started dipping into personalization in 2024, when it integratedGemini into Workspaceapps like Gmail, Docs, Sheets, Slides, and Drive in 2024, allowing its AI chatbot to do things likesift through your filesordraft an email. Gemini’s Deep Research feature can eventap into your emails, Drive, and chatsand use them as sources for its reports.Over the past several months, Google has continued toexpand these integrations. Itintroduced “Personal Intelligence”in January, a feature that allows Gemini to reason across Gmail, Google Photos, Search, and your YouTube history without prompting. That means Gemini can automatically surface details from across your accounts to personalize its responses. “Millions of people are using it [Personal Intelligence] every single day, they found it so helpful for things like personalized product and trip recommendations, or acting as a thought partner for navigating big decisions in life, like a career change,” Josh Woodward, the head of Google Labs, the Gemini app, and AI Studio, said during I/O 2026.Even though it’s completely optional to connect your Workspace apps, Search history, Photos, and other information to Gemini, it seems like Google’s AI future hinges on people doing exactly that. Daily Brief, which is rolling out to Google AI Plus, Pro, and Ultra subscribers, scans for updates from your Gmail and flags events from your Calendar.ButGemini Sparkis diving deeper into your information, as Google is pitching it as an AI personal assistant that can work across connected Workspace apps 24/7, creating continually updated study guides, generating to-do lists based on meeting notes, and even automatically scanning monthly credit card statements for hidden subscription fees. But the connections to Workspace apps are just the beginning, as Gemini Spark will also be able to plug into third-party services, like Canva, OpenTable, Instacart, Spotify, Expedia, Adobe, and more.Google is even planning to give Gemini Spark access to local files on Mac computers, similar to OpenClaw, the open-source AI agent platform thatposes a range of security risks. During a demo at I/O, Woodward showed how he could use Spark to send an email to a dog boarder in preparation for an upcoming trip. He selected documents across his computer and asked Spark to craft an email using his dogs’ allergies and vaccination records.Many people might draw the line at giving an AI system access to their whole computer. But if therise of OpenClaw tells us anything, it’s that AI is moving from being a novelty to a real productivity tool that demands access to our digital lives. It’s just a matter of whether people trust the companies behind the systems enough to hand over their personal data — and more importantly, where they’ll set a boundary on what’s too private.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Emma RothAIGoogleGoogle I/O 2026ReportTechMore in:Google I/O 2026: All the news and announcementsGoogle is launching an Android version of its AI Studio vibe coding tool.Jay PetersTwo hours agoWear OS 7 will keep track of deliveries and sports scores on your wristJay PetersMay 19Demis Hassabis said this might be the ‘foothills of the singularity.’ What?Stevie BonifieldMay 19Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes

GadgetsGamingNintendoNintendo’s $500 Switch 2 bundle includes a game, and it’s available nowYou’ll save up to $30, depending on the game. It’s a good deal if you’re in the market for the console.You’ll save up to $30, depending on the game. It’s a good deal if you’re in the market for the console.byCameron FaulknerMay 19, 2026, 9:00 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Mario Kart World is one of three games you can choose from in the bundle.Photo by Amelia Holowaty Krales / The VergeCameron Faulkneris an editor covering deals and gaming hardware. He joined in 2018, and after a two-year stint at Polygon, he rejoinedThe Vergein May 2025.Nintendo recently teased the“Choose Your Game” Switch 2 console and digital game bundle, coming in early June. However, multiple retailers (including Nintendo itself) are already offering the $499.99 bundle. The list currently includesGameStop,Amazon,Best Buy, and theNintendo Store. It’s a deal worth hopping on ahead of the holiday season, since Nintendo is raising the price of the Switch 2 to $499.99 in September.While theSwitch 2console currently costs $449.99 by itself, the bundle includes a full-price game with up to a $30 discount. You can choose between one of three digital games to redeem upon setup:Donkey Kong Bananza,PokémonPokopia, andMario Kart World. Each of those are regularly $69.99 except forWorld, which retails for $79.99.RelatedNintendo is raising Switch 2 prices in SeptemberI tested dozens of cases for the Switch 2 and these are the bestNintendo Switch 2 “Choose Your Game” bundle$500$5306% off$500Choose betweenDonkey Kong Bananza,Mario Kart World, andPokémon Pokopia.Read More$500 at Amazon$500 at GameStop$500 at Best BuyNone of the games offered as part of the bundle will take up much of the Switch 2’s 256GB internal SSD.Mario Kart Worldtakes up the most space of the three, requiring 22GB. However, you may eventually own enough games to exceed its capacity, so we recommend considering a256GB Samsung microSD Express cardto double your available storage. There’s currently a deal on those happening atAmazon, where you can buy one for $47.99 (originally $79.99).To redeem your game, enter the eShop from the Switch 2’s home screen. Then select “enter code” on the left side of the page. You’ll then enter the 16-character code provided in your bundle, then make your game selection.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Cameron FaulknerDealsGadgetsGamingNintendoVerge ShoppingMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAINewsThe future of Google is a search box that does everythingGoogle, Gemini, Gmail — it feels like they could one day become a single search box.Google, Gemini, Gmail — it feels like they could one day become a single search box.byJay PetersMay 19, 2026, 9:24 PM UTCLinkShareGiftPhoto by Allison Johnson / The VergePart OfGoogle I/O 2026: All the news and announcementssee all updatesJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Last year, after watching Google’s I/O keynote, I wrote that it felt likeGoogle’s future was Google googling. After watching this year’s I/O keynote on Tuesday, I don’t think Google just wants to google for you — I think it wants to doeverythingfor you, all from a search box.Take the trustyGoogle search bar itself, something Google is generally hesitant to update, which is getting some updates. It will “dynamically” expand as you type longer queries. It will offer “AI-powered suggestions” that Google claims will “go beyond autocomplete,” which could cause you to fill in the blanks of a search in a way you didn’t intend and that may or may not be helpful.Or how about what Google actuallyshows in search results? From AI Overviews you can keep asking questions in AI Mode, which generates a custom page with an AI-generated summary of what you’re searching about instead of showing you a traditional list of links. Search results are also going to be even more personalized because Google will be able to generate custom UI for you, including generating things like interactive visuals and graphs right within your search results page. You’re also going to be able to ask Google, right from the search bar, to make “information agents” that can keep track of things you care about — like new sneaker drops or apartment listings — making the search bar a kind of AI-infused Google Alert.Gemini is getting a bunch of upgrades and features, too. It can send you a “Daily Brief” that tells you about your day based on information from across your Google apps, like Gmail and Google Calendar. You can create your own custom Google-powered agents thanks to a featurecalled Gemini Spark, which could have a leg up on agents like OpenClaw because it’s a first-party offering from Google. The company has also been recently making a big kerfuffleabout Personal Intelligence, which pulls in context from your other Google apps that helps inform your responses from Gemini.In Workspace, Google wants you tojust blabat tools like Gmail and Docs and Keep, and the apps will help with things like parsing your inbox, drafting a document, or generating to-do lists. For shopping, the newUniversal Cartwill keep track of things you want to buy in apps like Search, Gemini, Gmail, and YouTube — and let you check out using Google’s payment infrastructure. And speaking of YouTube, the companyis testingan AI Mode-like experience there, too, where it will put together a page of search results instead of just showing you a list of things to watch.There’s more, too — with the newGemini Omni models, you’ll be able to create a video using things like other videos, other images, and audio as prompts. Down the line, you’ll be able to generate other media as well; the family of models is meant to be able to “create anything.”I could go on; there was a lot announced at I/O. But what I’m trying to get at here is that Google isn’t just showing you the location of information anymore. Through its various search boxes, it’s now just answering questions for you in a way that it thinks is most helpful. In the most charitable view: If it’s done well and accurately, this allcouldbe very useful. But that means it has to be done well and accurately, and that’s anextremely high bar to clear, especially with complex search queries or questions about sensitive data like my years of emails in Gmail.It’s not hard to imagine a future where, someday, Google just makes everything happen in one universal search box. No more hopping between Google Search, YouTube, Gmail, Gemini — my gut is that Google’s current end state is that you just type anything you’re looking for into an Ask Google box, and Google finds a way to make it happen and surface it for you in a personalized way.RelatedGoogle’s AI future demands trust — and your personal dataAssuming that’s the case, that’s not something I want. The fun of the internet is actually doing the work to find stuff, even if it’s sometimes frustrating, difficult, or time-consuming. Sure, the pitch for talking to Gmail to get help finding things in my inbox seems handy, but I’ve also spent years honing my own email management system — one that would work whether I used Gmail or not. It also works great forme, I think we should all have to go through the motions of figuring out the systems across all of our digital lives that work best for us instead of just relying on Google to figure it all out from one universal search box.Google

PodcastsTechGoogleWe react to Google I/O 2026Hayden and Jake discuss Spark, Antigravity, Omni, Pics, vibe coding, the new Google Search, and the singularity.Hayden and Jake discuss Spark, Antigravity, Omni, Pics, vibe coding, the new Google Search, and the singularity.byJacob KastrenakesMay 19, 2026, 9:48 PM UTCLinkShareGiftJacob Kastrenakesis The Verge’s executive editor. He has covered tech, policy, and online creators for over a decade.What better way to unwind from a two-hour keynote presentation than to pore over the weirdest and wildest details, froma Gmail bot you can converse withto DeepMind’s leadersaying the singularity is near.The Vergecastwent live right after the show, with senior AI reporter Hayden Field joining me to discuss the highlights.First up, we talk about the buzzword of the show: agents, which seem to be Google’s new answer to making AI useful. Google hasa new service called Sparkthat sounds a lot like a Gemini-flavored OpenClaw, upgrades to its Antigravity platform, and a bunch of agents that willshop for youand track search topics. We also go overGoogle’s model updates, as well as its shift to world models withGemini Omni, a theoretically ambitious new family that is off to a very focused start.Vergesubscribers, don’t forget you get exclusive access to ad-freeVergecastwherever you get your podcasts. Headhere. Not a subscriber? You cansign up here.Then, we discuss howGoogle Search is evolving— or put more plainly, being stuffed with AI features that turn Google into your final destination when browsing the web. We also talk vibe-coding Android apps inGoogle AI Studio, the newGoogle Pics app, and Google’sexpansion of access to content authentication toolsso that people can learn what is and isn’t real.We wrap up discussing Demis Hassabis’ proclamation that AGI and the singularity is near, which got surprisingly little stage time (one sentence) for such a bold statement. And finally, Hayden takes us through her thoughts on theMusk v. Altmantrial, which wrapped up earlier this week.Subscribe:Spotify|ApplePodcasts|Overcast|Pocket Casts|MoreThanks to everyone who tuned in to the livestream — we’re going to keep them coming. In the meantime, you can always reach us at the Vergecast Hotline (866-VERGE11) or send us an email at [email protected] don’t forget, you can get ad-free podcasts bysubscribing toThe Verge.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jacob KastrenakesGoogleGoogle I/O 2026PodcastsTechVergecastMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAINewsDemis Hassabis said this might be the ‘foothills of the singularity.’ What?After a couple of hours of apps and itineraries, Google’s big AI presentation turned philosophical.After a couple of hours of apps and itineraries, Google’s big AI presentation turned philosophical.byStevie BonifieldMay 19, 2026, 10:16 PM UTCLinkShareGiftImage: GooglePart OfGoogle I/O 2026: All the news and announcementssee all updatesStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.Welcome to a “profound moment for humanity,” according to Google DeepMind CEO Demis Hassabis, who closed outGoogle I/O’s keynote presentationon Tuesday, saying:Google’s cutting-edge research and products will help unlock AGI’s incredible potential for the benefit of the entire world. When we look back at this time, I think we will realize that we were standing in the foothills of the singularity.It will be a profound moment for humanity. This technology will be a force multiplier for human ingenuity and usher in a new golden age of scientific discovery and progress, improving the lives of everyone, everywhere. We look forward to building the future with all of you. Thank you, and enjoy the rest of Google I/O.Just before announcing we’ve possibly arrived at “the foothills of the singularity,” Hassabis introducedGemini for Science, a set of tools and experiments in Google Labs and Google Antigravity intended for helping with scientific research. According to Hassabis, with tools like these, Google hopes to “reimagine drug discovery with the goal of one day solving all disease.” Tech executives often discuss AI this way, like Microsoft CEO Satya Nadella referring to AI as “cognitive amplifier tools” and Luma AI CEO Amit Jain’sclaim last yearthat AI is the key to saving Hollywood.In aninterview withBloombergjust a few months ago, Hassabis said that “the singularity is another word for a full AGI arriving,” which is not the same definition we’re used to from the likes ofVernor VingeandRay Kurzweil.At the time, even by his measurement, he claimed that “we’re nowhere near that.” When asked about his timeline for achieving AGI, Hassabis said he was standing by his prediction that we have a “50 percent chance of getting there by 2030.”Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAIGoogleGoogle I/O 2026NewsTechMore in:Google I/O 2026: All the news and announcementsGoogle is launching an Android version of its AI Studio vibe coding tool.Jay PetersTwo hours agoWear OS 7 will keep track of deliveries and sports scores on your wristJay PetersMay 19The future of Google is a search box that does everythingJay PetersMay 19Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechEntertainmentGadgetsPlex is tripling the price of a lifetime pass to $750 after doubling it last yearNot your average streaming price hike.Not your average streaming price hike.bySean HollisterMay 19, 2026, 11:03 PM UTCLinkShareGiftImage: PlexPart OfStreaming keeps getting more expensive: all the latest price hikessee all updatesSean Hollisteris a senior editor and founding member of The Verge who covers gadgets, games, and toys. He spent 15 years editing the likes of CNET, Gizmodo, and Engadget.I am dying to know how much money Plex is about to make the next six weeks charging people to stream their own video from their own homes. Today, it’s giving every prospective customer until July 1st to lock in a lifetime subscription at today’s rates — before it triples the price to $750.Plex already more than doubled the price of a lifetime Plex Pass subscriptionfrom $119.99 to $249.99 last March, and it’ll triple again on July 1st to $749.99. At that price, you’d have to subscribe for 11 years (at the current annual rate) to make the lifetime sub worthwhile.Plex isn’t choosing the new number because it expects you to pay $750.In a blog post, the company says it actually wanted to stop issuing lifetime plans; it would rather push users towards recurring subscriptions instead to “sustain long-term development”. Now that the annual subscription covers a decade’s worth of Plex, it’s a win for Plex either way.But it mayespeciallybe a win for Plex if you, like me, feel sudden FOMO and consider giving Plex $250 for a lifetime subscription now. While you’d only be giving Plex around 3.5 years of annual revenue at today’s rates, it gets an instant infusion of cash. It’s like a huge sale except instead of needing to lower the price, Plex has actually raised it. (Again, the annual subscription costs double what it did last March.)From Plex’s email to existing users.Image: PlexPlex’s blog and emails to users encourage this with a hugely prominent call to action, bolded and highlighted in blue: “You have until 12:01 AM UTC on July 1, 2026, to get a Lifetime Plex Pass at the current price of $249.99 USD*here. If you’ve been considering it, now’s a great time to buy.”So does Plex’s website, which is currently topped with a banner reading: “The price of a Lifetime Plex Pass is increasing on July 1, 2026.Get yours at the existing price while you still can.Buy now”.None of that is to suggest Plex is laughing all the way to the bank. It’s a small company, its finances are private, the global advertising market is in a downturn, andit had big layoffsa few years back.Plex says monthly and annual subscription prices aren’t changing today, and nothing will change for existing lifetime pass holders. Those changes came last year. That’s also when Plex gated one of its most valuable features behind that subscription: the ability to stream your media server’s video outside your home.RelatedPlex Pass is going up in price — and now you’ll need it for remote playbackPlex makes piracy just another streaming serviceIt’s time to change your Plex password againFor free, without a subscription, you can still create collections and stream to devices on your local network, and stream remotely from other people who pay. Plex Pass also adds features that lets you skip the intros and credits on shows, remotely download your media, auto-rewind when pausing, hardware-accelerated streaming in its Media Server app,and more.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Sean HollisterAppsEntertainmentGadgetsTechTVsMore in:Streaming keeps getting more expensive: all the latest price hikesYouTube Premium is getting pricierDominic PrestonApr 10Netflix is raising prices againEmma RothMar 26Amazon Prime Video nearly doubles the price to go ad-free and stream 4K videoEmma RothMar 13Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechGadgetsNewsWear OS 7 will keep track of deliveries and sports scores on your wristGoogle’s newest smartwatch update could also give you up to 10 percent better battery life.Google’s newest smartwatch update could also give you up to 10 percent better battery life.byJay PetersMay 19, 2026, 11:24 PM UTCLinkShareGiftA Pixel Watch 4 running Wear OS 6, not Wear OS 7.Photo by Amelia Holowaty Krales / The VergePart OfGoogle I/O 2026: All the news and announcementssee all updatesJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Amid the flurry of today’s Google I/O announcements, Google shared detailsabout Wear OS 7, the next major update to its smartwatch platform. To help you keep track of things like deliveries and sports scores, Wear OS 7 will get the iPhone-style Live Updates that wereintroduced on Android last year— which can appear on your watch or your smartphone — and you’ll also be able to track automated tasks that an AI is working on right from your watch.Wear OS is also getting an upgrade from its widget-like Tiles for glanceable information. With the new update, Google is adding “Wear Widgets” to the platform, which look more like Android widgets and can appear in small or large layouts that “align perfectly” with Android’s 2x1 and 2x2 widget formats. There are some AI-powered features coming to Wear OS 7 as well, including the introduction of Gemini Intelligence, Google’scatch-all brandingfor personalized and proactive Gemini features, on “select watches” launching “later this year.”Various Wear OS 7 features, including Wear Widgets. Click the photo to take a closer look.Image: GoogleGoogle is also promising an “up to” 10 percent improvement in battery life for average users if you’re upgrading from Wear OS 6 to Wear OS 7. The company says that it’s investing in “power optimizations” so that users can “do more with their favorite apps,” but we’ll have to wait and see how battery life actually shakes out when Wear OS 7 is actually available.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGadgetsGoogleGoogle I/O 2026NewsSmartwatchTechWearableMore in:Google I/O 2026: All the news and announcementsGoogle is launching an Android version of its AI Studio vibe coding tool.Jay PetersTwo hours agoDemis Hassabis said this might be the ‘foothills of the singularity.’ What?Stevie BonifieldMay 19The future of Google is a search box that does everythingJay PetersMay 19Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

At Google I/O on Tuesday, Google introduced Universal Cart, its agentic hub for managing shopping in one place. The tech giant also announced updates to its Agent Payments Protocol (AP2) and teased that it would bring the technology to Google products in the coming months, enabling users to authorize agents to make payments on their behalf.The announcements signalGoogle’s push to turn AI assistants from passive recommendation tools into active participantsin online commerce. By launching a centralized shopping system and building infrastructure that lets software agents complete purchases autonomously, the company is positioning itself to control more of the entire shopping journey, and potentially the relationship between consumers and the merchants competing for their attention.With Universal Cart, users can add products they’re considering from anywhere on Google — while browsing Search, chatting with Gemini, watching YouTube, or reading Gmail. Once items are added, Universal Cart tracks deals, monitors price drops, surfaces price history insights, and alerts users when items are back in stock.Image Credits:Sarah Perez/TechCrunchThe feature is built around something Google knows well, which is that most people shop across multiple devices, multiple retailers, and over the course of many days.The cart also uses AI to help shoppers make better decisions. For example, if you’re building your first custom PC, you can add parts from multiple merchants into a single cart, and Google may flag compatibility issues, such as a processor that doesn’t work with the motherboard you selected — and suggest an alternative.For frequent travelers or rewards maximizers, the feature can also surface hidden savings and help stretch your points further because it’s built on Google Wallet.Image Credits:GoogleThanks to Google’s open-standardUniversal Commerce Protocol (UCP), users can check out directly through Google with participating merchants, or transfer their items to the merchant site and complete the purchase there.Universal Cart is rolling out in the U.S. today and coming to the Gemini app this summer, with YouTube and Gmail to follow, Google says.Google also announced that UCP is expanding to more categories, like hotels and local food delivery services. UCP-powered experiences will also expand beyond the U.S. to Canada and Australia in the coming months, and later to the U.K.The more consequential announcement for the commerce industry may be AP2, Google’s protocol designed to let AI agents securely make payments on users’ behalf within defined limits. At I/O, Google detailed the guardrails users can set, including specifying the brands and products they want, and a spending limit. When those conditions are met, the agent makes the purchase automatically.Image Credits:Sarah Perez/TechCrunchGoogle says it’s bringing AP2 to its own products in the coming months. That integration would give Google direct visibility into what consumers discover, consider, and ultimately buy, and it’s a degree of commercial influence that retailers and payment processors will be watching closely.Under the hood, AP2 creates a transparent, verifiable link between the user, the merchant, and the payment processor, with encryption protecting user data throughout. The protocol also includes tamper-proof digital records that ensure the agent is always acting on the user’s behalf, and a permanent audit trail that both buyers and sellers can reference for returns or disputes.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,Commerce,Google,Google I/O,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo France

In BriefPosted:10:45 AM PDT · May 19, 2026Image Credits:Alex Wong / Getty ImagesSarah PerezAgentic app coding gets an upgrade with Google’s release of Android CLIIn addition to serving up its own flavor of vibe-coding tools for building on Android, Google announced it will also offer tools that allow AI agents, like Claude Code, OpenAI’s Codex, or its own Antigravity or Gemini in Android Studio, to accelerate Android app development.Google shared on Tuesday at its annual Google I/O developer conference that its Android CLI (command line interface) is now stable at version 1.0 and can be used by those who are leveraging AI agents to build for them — no matter what their preferred coding platform may be.The move acknowledges that many people are now building for Android with AI agents that aren’t from Google. The company is trying to find a way to make its specialized knowledge, like what’s found within Android Studio, more accessible.With Android CLI, AI agents can retrieveknowledgeabout Android development via a new “android studio” command that can tap into the capabilities ofAndroid Studiowhen building their own project. From there, the agents will be able to tap into a range of other commands and tools.Google Antigravity, the company’s agentic development platform, will include an optional bundle that will install the tools and knowledge included in Android CLI, so it can also perform the core tasks for Android app development, Google says.TopicsAI,Android,Apps,coding,Google,google io,google io 2026,In BriefMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAIWith Gemini 3.5 Flash, Google bets its next AI wave on agents, not chatbotsRebecca Bellan8 hours agoAppsGoogle’s AI Studio now lets anyone build Android apps in minutesSarah Perez8 hours agoAIGoogle launches Antigravity 2.0 with an updated desktop app and CLI tool at IO 2026Ivan Mehta8 hours agoLatest in AppsAppsGoogle just declared itself a contender in AI design at IO 2026Aisha Malik4 hours agoAIYou can now talk to your Gmail inbox, as seen at Google IO 2026Sarah Perez4 hours agoAIGoogle Search as you know it is overSarah Perez8 hours ago

At its annual Google I/O event on Tuesday, Google announced that the Gemini app is getting a series of new updates, including a “Daily Brief” feature, a redesigned interface, access to a new AI video model calledGemini Omni, and new personal AI agent calledGemini Spark.The updates signal Google’s push to turn its Gemini app into an all-purpose AI hub rather than a stand-alone chatbot and to make the AI assistant more competitive with apps like ChatGPT and Claude.Google describes the new Daily Brief feature as a personalized digest designed to be your first stop each morning. It pulls together information from a user’s inbox, calendar, and most important tasks, and then organizes them into a clear overview. The company says Daily Brief doesn’t just summarize this information — it also prioritizes tasks and suggests next steps, with the most important items shown first.Daily Brief is rolling out today to Google AI subscribers in the United States.Image Credits:GoogleThe Gemini app already has considerable reach: The company said it has more than 900 million monthly users and is available in over 230 countries and in more than 70 languages. But Google clearly wants more and without losing its existing customers.Google said it rebuilt the app from the ground up. Now when users open the app, they’ll see a new design language called “Neural Expressive,” which brings fluid animations, vibrant colors, new typography, and haptic feedback.Image Credits:Sarah Perez/TechCrunchGemini’s responses are no longer presented as a wall of text, which is the case with most AI chatbots. Instead, key information appears in bold at the top. Additional text and possibly other elements such as images and timelines appear as the user scrolls down.As for Gemini Spark, Google describes it as a 24/7 personal AI agent that helps you navigate your digital life. Spark transforms Gemini from an assistant into an active partner that does real work on your behalf. Since it’s a cloud-based agent, Spark keeps working in the background even when you lock your phone. With Gemini Spark in the Gemini app, users will be able to create their own custom workflows.Spark is currently in testing and the company expects to make it available to Google AI Ultra subscribers next week.Image Credits:Sarah Perez/TechCrunchThe company’s new AI video model,Gemini Omni, combines Gemini with Google’s generative media models to create outputs grounded in knowledge. For example, you could give it a simple prompt like “claymation explainer of protein folding.” The model lets you upload audio, images, and video to generate a consistent, high-quality video, according to Google.By adding access to a new video-generation model like Gemini Omni, Google is intensifying competition among major AI platforms in the ongoing race to lead multimodal content generation. The model is rolling out to Google Flow and YouTube Shorts for Google AI subscribers, underscoring the company’s broader push into multimodal content creation and AI-powered video tools.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,gemini,Google,Google I/O,google i/o 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

The era of the “ten blue links” is officially over.At its Google I/O conference on Tuesday, Google unveiled an AI-powered overhaul of Search centered around a reimagined “intelligent search box” — what the company describes as the biggest change to this entry point to the web since the search box debuted more than 25 years ago.Instead of returning a simple list of links, Google Search will drop users into AI-powered interactive experiences at times. Google is also introducing tools that can dispatch “information agents” to gather information on a user’s behalf, along with tools that let users build personalized mini apps tailored to their needs.Loading the player…The resulting experience will no longer look much like how people envision Google Search, which has long been defined by ranked links to websites that have the information you need.With the revamped Search experience, the new search box simply expands to accommodate longer, more conversational queries, rather than making you decide what type of search experience or mode you want to choose at the start of your query. It will also have a new AI-powered query suggestion system that goes beyond autocomplete to help people craft more complex and nuanced queries, Google says.Google’s AI Overviews will also allow users to ask follow-up questions in AI Mode, beginning Tuesday, the company noted.Image Credits:GoogleGoogle is also introducing agentic capabilities and AI-powered interactive features into the search experience. This means people will spend even less time clicking the traditional blue links that Google Search used to return.Starting this summer, people will be able to create, customize, and manage multiple new “information agents” within Google Search. These agents can work in the background 24/7 to track changes on the web and alert you to new information. For instance, you could have an agent track market movements based on customer parameters, Google suggests.Image Credits:GoogleWhile the underlying technology here is powered by AI, which makes it more capable, the idea itself is not a new one.In 2003, Google launchedGoogle Alerts, a change-detection service that emailed users when new web results matched their search terms. The web was smaller and more manageable then, of course, so this became a part of many information workers’ tool sets. (That service still existsin some form but is no longer the way most web users go about acquiring new information.)Information-gathering agents are an evolution of Google Alerts. Beyond spotting changes, they can make sense of them, too.“You could send an alert to track market movements in a particular sector with very specific parameters, and the agent will map out a monitoring plan for you, including the tools and the data it needs to access — like our real-time finance data,” Google’s head of Search, Liz Reid, explained in a press briefing. “And it will then keep track of those changes and let you know when the conditions are met, and provide a synthesized update with links and information you can dive into further,” she added.Image Credits:GoogleThis shift means that “searching the web” will increasingly be performed by AI agents rather than humans. Instead, people will focus more on acting on the information those agents provide instead of manually clicking links.Image Credits:GoogleLinks will become an afterthought with the coming changes to the Search results experience, which builds on Google’s earlier launches of AI search features, like its short summaries known as AI Overviews and its conversational search, AI Mode.AI Overviews are now used by more than 2.5 billion monthly users; meanwhile, its conversational search mode, launched last year, now tops 1 billion monthly users. (ChatGPT, for comparison, has900 million weekly active users, as of earlier this year. This suggests that ChatGPT is now seeing more frequent engagement, with users coming back repeatedly throughout the week, while Google has more total unique people touching its AI features over the course of a month.)Now, thanks to a combination of Gemini andGoogle Antigravity, the company’s agentic development platform, Search results will begin to look more like interactive web pages.Image Credits:Google“Search can build custom experiences just for your individual questions, from dynamic layouts, interactive visuals to persistent and stateful project spaces that you can return to again and again,” says Reid. One of the ways Google is integrating these new capabilities is with “generative UI” (user interface), where it builds custom widgets and visualizations on the fly in answer to users’ search questions.You can imagine, for example, how a question about black holes in space could lead to an interactive visual that brings the concept to life, Reid said, adding that users can then ask follow-up questions and see Google respond with brand-new visuals in real time.Image Credits:GoogleGoogle says the new system was built in partnership wi

The AI coding boom is now coming directly for Android app development. On Tuesday atGoogle IO 2026, the companyannouncednew native Android app creation capabilities in its web-basedGoogle AI Studio,shrinking a process that takes weeks of setup and coding down to minutes.The company also said that consumers will be able to use Gemini AI to find the apps they need, both on the Play Store and the web, expanding opportunities for developers to have their apps discovered.Google says the new capabilities could make sense for anyone from a seasoned developer looking to prototype a new app quickly to a first-time creator.By offering the ability to essentially vibe-code Android apps via web-based tools, Google is ramping up the competition with other AI-powered development tools, like Cursor, Replit, Lovable, Claude Code, and others, while also opening up Android development to a new type of user: a non-technical creator. The news also represents an expansion of Google’s earlier addition ofAI-powered coding with Geminiin itsdesktop versionof Android Studio.The apps are built with theKotlinprogramming language using Google’s Jetpack Compose toolkit and with support integration with hardware sensors like GPS, Bluetooth, and NFC, the company says. However, the resulting creations, for now, are only meant to be used personally, as publishing for family and friends is still on the roadmap.The company suggests the technology could be used for the creation of personal utilities and simple social apps, hardware-enabled experiences, or AI-powered experiences.For now, would-be app developers can use the embedded Android Emulator directly in a web browser to preview and interact with the app as it’s being built. Users can then install the app on their Android phone over a USB cable connected to their computer, using the integrated Android Debug Bridge (adb).For those looking to take their project further,AI Studiocan automatically create the app record, package the bundle, and upload it to an internal testing track inGoogle Play Consolefor developers. This allows users to continue to iterate on their app while updating on their devices along the way.Those who want to take the next steps to publishing the app more publicly can hand off this version of the project to Android Studio by downloading a zip file and exporting it directly to GitHub. In time, Google plans to allow creators to publish their apps for use by family and friends and will add support for Firebase integrations (Firestore, Firebase Auth, Firebase App Check, and other tooling).In doing so, the company is imagining an Android app ecosystem where users find apps from among their own network of friends, not just the Play Store.However, for the latter, Google is infusing AI into the experience here, too.A new “Ask Play” AI-powered overlay will allow users to discover new apps by having natural conversations with AI within the Play Store.Image Credits:Google (screenshot)Perhaps more importantly, apps will begin to be surfaced with users’ conversations with Google’s Gemini virtual assistant, exposing developers’ apps to millions of users. This will roll out in the weeks ahead across Gemini on the web and on Android. Later this year, Gemini will also surface over 450,000 movies and TV shows, plus where to livestream sports, which can directly link users from their queries to a developer’s Android app with the content in question.Image Credits:GoogleWhile Google previewed a number ofAndroid-related announcements last week, it held back on sharing the news of the native Android app development until Tuesday’s start of its annual developer conference, Google I/O. That suggests the company believes this is bigger news and more closely tied to its idea of putting AI to real-world use, as was the larger theme of this year’s event, where AI was spread across Google products, from workspace productivity apps to AI tools, search, mobile apps, and more.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Android,Apps,coding,Google,google io,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sarah PerezConsumer News EditorSarah has worked as a reporter for TechCrunch since August 2011. She joined the company after having previously spent over three years at ReadWriteWeb. Prior to her work as a reporter, Sarah worked in I.T. across a number of industries, including banking, retail and software.You can contact or verify outreach from Sarah by [email protected] via encrypted message at sarahperez.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s

Google launched on Tuesday Gemini 3.5 Flash, a new AI model that the company says is its strongest yet for coding and autonomous AI agents. The model, which was introduced at the company’s annual Google I/O developer conference, can independently execute coding pipelines, manage research projects, and, in internal tests, build an operating system entirely from scratch.The release signals Google’s shift from pitching AI as a conversational tool to AI as an agentic tool. It’s not just answering questions, but planning, building, and iterating on real work with minimal human input.Koray Kavukcuoglu, DeepMind’s chief technologist, told reporters on Monday ahead of the public launch: “3.5 Flash offers an incredible combination of quality and low latency. It outperforms our latest frontier model, 3.1 Pro, on nearly all the benchmarks,” including coding, agentic tasks, and multimodal reasoning.He added that it is 4x faster than other frontier models, a speed that’s ideal for coding and agentic tasks, but that Google has “taken it to another level” by developing an optimized version of Flash that’s 12x faster with the same quality.That speed is central to Flash’s design for agentic work, where multiple AI agents run at the same time on long-running tasks, according to Kavukcuoglu. Onstage at I/O, Google engineer Varun Mohan, demonstrated agents spawning off to work on separate components before coming together to build a full operating system inside Antigravity, the company’s agentic development platform and IDE.Image Credits:GoogleKavukcuoglu said Flash 3.5 was co-developed with Antigravity so that agents could have a “native environment where they can live, work, and execute.”At I/O, Google released Antigravity 2.0, a stand-alone desktop application designed around agent-first development.The gains are showing up beyond demos. Google says 3.5 Flash’s agentic capabilities are already creating impact among partners, like banks and fintechs automating multi-week workflows, or data science teams finding insights in complex data environments.The model can run autonomously for multiple hours, though Tulsee Doshi, Google’s senior director and head of product, said it will at times pause and ask for user input when it hits a decision point or permission issue that requires human judgment.When Google releases its forthcoming 3.5 Pro model, the two are designed to work in tandem.Doshi told TechCrunch that “3.5 Pro becomes your orchestrator, your planner, and then it actually can leverage Flash to be the various sub-agents. I think it really comes down to where do you really want that reasoning power, where you actually want that larger model that can really push on the reasoning side versus where do you have tasks that really do merit good brute force tool use capabilities?”Now, 3.5 Flash is the default model in the Gemini app and in AI Mode in Search globally. At I/O, Google also announced agentic capabilitiescoming to Search, letting users create, customize, and manage AI agents directly on the platform. The new model will also powerGemini Spark, Google’s new personal AI agent designed to run 24/7 to help consumers manage their digital life.Providing that level of AI capability for average consumers comes with scrutiny. Google is currently facing a lawsuit after a man nearly committed a mass casualty event and died by suicide following weeks of chatting with Gemini last year.The implications for harm only grow when making powerful autonomous agents available more broadly. Google says Gemini 3.5 has strengthened cyber and CBRN (chemical, biological, radiological, and nuclear) safeguards and is better calibrated to engage with sensitive questions rather than refuse them outright.Gemini 3.5 Flash is available generally today via Antigravity, the Gemini API, and Gemini Enterprise, as well as through the Gemini app and AI mode in Search.TopicsAI,gemini,gemini flash 3.5,Google,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Rebecca BellanSenior ReporterRebecca Bellan is a senior reporter at TechCrunch where she covers the business, policy, and emerging trends shaping artificial intelligence. Her work has also appeared in Forbes, Bloomberg, The Atlantic, The Daily Beast, and other publications.You can contact or verify outreach from Rebecca by [email protected] via encrypted message at rebeccabellan.491 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank account

We’ve all pulled up Street View on Google Maps to show a friend what our childhood home looked like, or dropped that little person icon onto the streets of Paris to see if we booked a hotel in a cool neighborhood. Imagine being able to do that, but in a more immersive, interactive way that allows you to really simulate the street and its environs, and even do things like adjust the weather or see what it would look like in a “Day After Tomorrow” scenario.That’s one of the goals of Google’s latest integration. Starting today, Google DeepMind is connecting Street View toProject Genie,the company’s general-purpose world model that can generate diverse, interactive environments. The new feature launched during theGoogle I/O 2026developer conference.“It’s really powerful for both the agent [and robotics] use case and for humans to play with, and that’s always been the thesis of Genie,” Jack Parker-Holder, a research scientist on DeepMind’s open-endedness team, told TechCrunch.He gave the example of a new robot being deployed in London, which rarely sees the sun. Genie could, Parker-Holder says, simulate those scarce occasions when the sun glints off the Victorian housing, so the rays don’t shock the robot when it happens.Loading the player…“Simultaneously, you might say, ‘I’m going to New York City, but not this time of year,’” he continued. “‘It’s going to be snowy. I want to see what that block looks like in the snow.’”Google has been collecting Street View data for 20 years via cars with cameras and individuals strapped with “tracker backpacks.” The tech giant has collected north of 280 billion images across 110 countries and seven continents.“With Street View, we have imagery from a large quantity of the world,” Jack said. “You can imagine how potentially powerful it is to combine this rich source of real-world information and data with an ability to simulate worlds.”Google released its latest world modelGenie 3 for research previewlast August and opened up access to the tool to Google AI Ultra subscribers in the U.S. in January, allowing customers to create interactive game worlds from text prompts or images. The goal is to use Genie for educational experiences, gaming, and robotics training.Genie 3 is already helping to powerone of Waymo’s simulatorsto train its self-driving cars on “exceedingly rare events” like tornadoes or casual elephant encounters. Adding Street View data to that could help Waymo prepare to launch in more cities around the globe.Waymo has its own simulator that it relied on to scale to 11 U.S. cities and test its AI driver in several more. The difference with Genie, says Parker-Holder, is that those are all from the car’s point of view. Street View allows for not only simulating a world anchored to a real place, but also shifting the point of view to other types of agents, like a human or a robot.Google is launching Street View in Genie to some Ultra users in the United States starting today, with access rolling out at scale over time. Global Ultra users will gain access over the next few weeks, per the company.The researchers’ goal is to put this new capability into as many hands as possible, per Diego Rivas, a product manager at DeepMind. He cautioned that Street View in particular and Genie in general is still an experiment, so there’s much to improve upon in terms of accuracy.In the samples the Google team showed me — including an underwater simulation of a neighborhood I used to live in — the results are impressive and recognizable, but still video game quality rather than photorealistic. The models are also not yet physics-aware, meaning they don’t yet understand cause and effect. For example, in a simulation of a woman running through a snowy Joshua Tree, she ran right through cacti and bushes.Compare that to, say, Google’s image generator Nano Banana — which can now generate perfect text in infographics — or its video generator Veo — which understands that paper boats drift on water currents, smoke disperses into the air, and fabric drapes over forms.Physics isn’t hard-coded into these models; they learn it intuitively over time through passive observation, as a living being would.“I think for this kind of model, it’s maybe six to 12 months behind video in terms of the accuracy and quality, so I think it’s something we will solve,” Parker-Holder said.Jonathan Herbert, director of Google Maps who started on the Street View team as an intern 12 years ago, said that Genie can’t yet create a faithful reconstruction of a street. He thinks the real breakthrough is the AI’s spatial continuity. If you turn 360 degrees, the AI correctly remembers and simulates the environment behind you. From that point on, the model can build a new environment on top of that.“We have long thought about how we can build out the best and richest model of the world on top of Street View data,” Herbert said. “It’s definitely been an idea of ours to use Maps Data in new ways and for new kinds of AI researc

Google is getting (back) into the smart glasses game.At Google I/O on Tuesday, the company announced a new partnership with Warby Parker and Gentle Monster to produce a new line of AI-powered glasses. The company says that the devices will be built to pair with Android and iOS devices and were designed in collaboration with Samsung. They will be available later this year, the company said.Google is calling the new devices “audio glasses,” in that users will be able to issue verbal commands to them and get things done via its ecosystem of apps and services, including Gemini.Image Credits:TechCrunchThe user simply talks to their glasses (the demo shared on Tuesday involved a Googler ordering a coffee online by merely talking to the glasses), and the device, when synced, complies.Google has dabbled in smart glasses a number of times over the years. It notoriously launchedGoogle Glassyears ago, which ultimately helped spawn the derogatory term “glassholes.”The smart glasses space has changed a little bit since then, however. Lately, major companies — most notably Meta — and a small army of startups and smaller firms, have invested in the space.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,ai glasses,glasses,Google,google io,google io 2026,HardwareWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lucas RopekSenior Writer, TechCrunchLucas is a senior writer at TechCrunch, where he covers artificial intelligence, consumer tech, and startups. He previously covered AI and cybersecurity at Gizmodo.

The jury’s speedydecision to rejectElon Musk’s lawsuit against the other founders of OpenAI and Microsoft confirmed what we saw in the courtroom: Musk’s case was a weak one, in part because he waited so long to file it.Watching the closing arguments last week, OpenAI’s attorneys detailed point by point how the law was on their client’s side, while the plaintiff’s team focused on Sam Altman’s apparentlack of credibilityand expressed disbelief that anyone would disagree with Musk’s accusations.The final effect was that, after the verdict, some found it hard to believe Musk had lost — including the man himself. In a post he later deleted, Musk called Judge Yvonne Gonzalez Rogers a “terrible activist Oakland judge,” then announced his plans to appeal, declaring, “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity.”But Altman and Brockman weren’t the only figures who benefited from OpenAI’s non-profit investments. As much as Musk and his legal team tried to make the trial about Altman, the proceedings revealed just as much about Musk.One incident that came out in court showed Musk benefiting from OpenAI in an uncomfortably familiar way. Greg Brockman testified that in 2017, Musk asked him to bring a team of OpenAI researchers down to Tesla’s headquarters to help with the autopilot team for a few weeks. “It was pretty clear that was not something we could say no to,” Brockman said.Brockman described taking a team of leading scientists, including Andrej Karpathy, Ilya Sutskever, and Scott Gray, to consult with the “demoralized” Tesla workers. They helped come up with ideas to improve the vehicle’s self-driving technology, with Sutskever telling the team that if they could find 10,000 images of a tricky corner case, they would be able to fix their software. Musk even asked Brockman to recommend employees to fire, which he declined to do.Another person familiar with the episode confirmed Brockman’s account and said Tesla did not reimburse OpenAI for the time and effort of its employees. Musk’s family office, Excession, didn’t reply to a request for comment.The heart ofMusk’s caseis that Altman, Brockman, and OpenAI committed a “breach of charitable trust” — that Musk donated funds for a specific charitable purpose, and his co-founders instead used them for something else. He also accused them of “unjust enrichment” due to stock and other benefits from OpenAI’s for-profit.In the case of the OpenAI scientists parachuting into Tesla, Musk’s charitable donations were intended to hire scientists focused on securing the benefits of AGI. Instead, he had them work for free at his for-profit company.Dorothy Lund, a Columbia Law School professor and the co-host of theBeyond Unprecedented podcast, told TechCrunch that this arrangement wouldn’t be legal, calling it “a bit rich for Musk to be suing for breach of a charitable trust, when he appears to have been redirecting assets in a way that was inconsistent with that mission.”It’s true that the self-driving work involved artificial intelligence, but witnesses for Musk emphasized that Tesla’s self-driving project was very different from OpenAI’s research agenda. That’s in part because Karpathy left OpenAI for Tesla shortly after this incident. OpenAI’s attorneys portrayed the departure as Musk violating his duty to the lab, where he was co-chair of the board, by recruiting one of its key researchers to his own company.The other fact that no doubt influenced the jury was the amount of time Muskspent tryingto gain sole control of a potential OpenAI for-profit affiliate in 2017. Musk deployed good cop, bad cop tactics in an attempt to convince his co-founders to let him have total control of OpenAI’s for-profit affiliate — giving them free Teslas and threatening to withhold his donations.His efforts put his attorneys in a tricky spot, facing a need to convince the jury there was a significant difference between what Musk envisioned and the for-profit that was ultimately created. They suggested a “small adjunct” for-profit would be permissible, though OpenAI’s witnesses showed non-profits with large commercial arms are common.Indeed, there’s a very plausible counter-factual where Musk took one of the offers his co-founders made to split their equity more evenly, and finds himself today as one of OpenAI’s largest shareholders — just not the controlling one. But several times during the trial, Musk’s associates testified that he refuses to invest in any business he could have sole control over.The failure of Musk’s claims because he filed them too late has been cited as a technicality, but the statute of limitations has substance behind it: People and businesses make important decisions and spend resources based on their understanding that what they are doing is permissible. If someone like Musk waits too long to sue, then the cost of unraveling all those decisions can outweigh a just reimbursement.No

Shay Shwartz knows a lot about email phishing attacks. As a teenager, he made money as a hacker, but after getting caught at age 16, he realized he could use his cyber talents to prevent attacks rather than launch them.He went on to spend about a decade in top-tier cybersecurity roles, leading major projects for Israel’s elite defense and intelligence units, including work connected to the Iron Dome project, before joining Axis, the startup later acquired by HPE.All along, he had been itching to launch his own startup, and two years ago, he finally took the plunge.His startup Ocean, an agentic email security platform built to fight AI-powered attacks, just emerged from stealth mode with $28 million in total funding. The round was led by Lightspeed Venture Partners, with participation from Picture Capital and Cerca Partners. High-profile angel investors also joined the round, including Wiz co-founder and CEO Assaf Rappaport, as well as Yevgeny Dibrov and Nadir Izrael, the co-founders of Armis, which recently sold to ServiceNow for $7.75 billion.While established vendors like Proofpoint and Mimecast, along with newer players like Abnormal Security, help detect standard phishing attacks, Shwartz (pictured right next to co-founder and CTO Oran Moyal) argues that AI requires a different defensive approach.In the past, only highly sophisticated hackers could pull off spear-phishing due to the sheer amount of time, research, and manual labor needed to launch targeted attacks.“AI just made the entire process automatic, so the scale is much, much bigger now,” Shwartz told TechCrunch. “I can instruct LLM to go and understand exactly who you are, harvest large amount of public information, and create those phishing attacks very targeted against you.”Ocean claims its AI can thoroughly analyze the context of every incoming email to detect fraud and impersonation attempts.The startup is already reviewing billions of emails each month for customers, including Kayak, Kingston Technology, and Headspace.Shwartz said Ocean built a small language model tailored to quickly analyze emails, understand the sender’s intent, and evaluate it against the user’s specific organizational context.“This is like having a guard in every door,” Shwartz said. “This is how we make the inbox a safe place with high hygiene.”TopicsAI,Exclusive,lightspeed,phishing,Security,StartupsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Three-year-old Mach Industries has acquired solid rocket motor startup Exquadrum in a $50 million cash-and-equity deal, the Huntington Beach-based defense startup tells TechCrunch. Exquadrum — now rebranded as Mach Energetics — has been fully folded into Mach’s operations, giving it direct control over one of the most important, and constrained, components in modern unmanned systems.The deal itself started with a lucky break. The two companies first connected last September when an Exquadrum customer at an MIT recruiting event happened to overhear a Mach recruiter mention that the company was in the market for a solid rocket motor supplier. Introductions were made, Mach came in as a customer, and now, roughly five months later, it has acquired the company outright, beating out upward of eight other potential buyers in the process, it says.“The Exquadrum acquisition marks an important next stage in Mach’s growth,” said founder and CEO Ethan Thornton, who dropped out of MIT at 19 to start the company. “As we deliver vehicles to the warfighter, we’ll continue to vertically integrate our supply chain across solid rocket motors, engines, radar, and avionics to ensure we deliver the best possible product at the lowest cost. In many areas of the defense industrial base, these components are not only too expensive or lacking performance, they’re simply unavailable, with lead times stretching years. In short, vertical integration is non-optional.”That supply problem is real and getting more acute. Decades of consolidation have left the domestic solid rocket motor market effectively controlled by two large primes — Aerojet Rocketdyne and Northrop Grumman — with little independent capacity to absorb the growing demand that modern drone warfare has created.Indeed, in February, the Pentagon awarded the defense tech outfit Anduril $43.7 million specifically to expand domestic SRM production (its second such investment in the company in just over a year), explicitly calling SRMs a critical bottleneck in the munitions supply chain.Mach is now explicitly positioning itself as part of the solution, and not just for its own programs but for the broader ecosystem. Mach Energetics plans to sell components, testing services, and subsystems to other defense firms, a maneuver that suggests Mach sees itself as potential infrastructure for the defense tech industry and not just a systems builder.According to Mach, all 85 Exquadrum employees are coming over as part of the deal, along with the company’s IP, business lines, and its 70,000-square-foot facility in Victorville, California, which is anchored by a nearby energetics and rocket propulsion test site. The combined company now has roughly 350 employees. Exquadrum co-founders Kevin Mahaffy and Eric Schmidt (no relation to the former Google CEO) are both taking on leadership roles within Mach Energetics and the broader organization.The acquisition also mirrors moves by other ambitious defense tech startups that are focused on owning the stack and using cost and speed as competitive weapons. Mach has five vehicle programs in various stages of development — Viper, a jet-powered VTOL; Glide, a high-altitude strike glider; Stratos, an airborne surveillance platform; Dart, its low-cost counter-drone interceptor; and Pike, a long-range strike munition built for large-scale deployment — with plans to enter production on at least three this year. The company says the acquisition meaningfully improves unit economics across all of them at exactly the moment it is starting to scale.Mach has raised nearly $200 million in total — most recently a$100 millionSeries B last June led by Bedrock Capital, Khosla Ventures, and Sequoia Capital — at a valuation of $470 million. It’s a multiple that right now looks modest for a company with this trajectory and will definitely be worth watching as the rubber starts to meet the road this year.Topicsdefense tech,Ethan Thornton,Hardware,mach industriesWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Connie LoizosEditor in Chief & General ManagerLoizos has been reporting on Silicon Valley since the late ’90s, when she joined the original Red Herring magazine. Previously the Silicon Valley Editor of TechCrunch, she was named Editor in Chief and General Manager of TechCrunch in September 2023. She’s also the founder of StrictlyVC, a daily e-newsletter and lecture series acquired by Yahoo in August 2023 and now operated as a sub brand of TechCrunch.You can contact or verify outreach from Connie by [email protected]@techcrunch.com, or via encrypted message at ConnieLoizos.53 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk h

In BriefPosted:2:14 PM PDT · May 19, 2026Image Credits:MARTIN BUREAU/AFP via Getty ImagesZack WhittakerDiscord enables end-to-end encrypted voice and video calling for every userMessaging platform giant Discord has switched on end-to-end encrypted voice and video messaging for every user. Theend-to-end encryptionfeature means that Discord users can now communicate privately without anyone else being able to listen in to their calls, not even Discord.This is a major privacy win for the community’s hundreds of millions of users, soon after social media giant Metapulled the plugon Instagram’s end-to-end encrypted messaging feature earlier this year. TikTok said it would alsonot end-to-end encrypt user messagesafter it became a U.S. company.Discordlaunched the end-to-end encrypted voice and video calling feature in 2024, and on Monday rolled out the feature to every user, no action needed.“End-to-end Encryption is now standard for every voice and video call on Discord, outside of stage channels. No opt-in required,” said Mark Smith, Discord’s vice president of core technologies, in ablog post.Topicscybersecurity,discord,end-to-end encryption,In Brief,SecurityMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedSecurityUS cyber agency CISA exposed reams of passwords and cloud keys to the open webZack Whittaker12 hours agoSecurityNYC Health + Hospitals says hackers stole medical data and fingerprints during breach affecting at least 1.8 million peopleZack Whittaker1 day agoSecurityOpen source tool maker Grafana Labs says hackers stole its code, refuses to pay ransomZack Whittaker2 days agoLatest in SecurityIn BriefDiscord enables end-to-end encrypted voice and video calling for every userZack Whittaker6 hours agoSecurityFrom teen hacker to Iron Dome researcher, this founder raised $28M to fight AI phishingMarina Temkin6 hours agoIn BriefHackers have compromised dozens of popular open source packages in an ongoing supply-chain attackLorenzo Franceschi-Bicchierai11 hours ago

At theGoogle I/O 2026keynote, the tech giantrevealed new agentic capabilities in Search, where users can create, customize, and manage multiple AI agents to stay updated on topics of interest. The announcement is part of Google’s larger push toward agentic AI systems that can take initiative and assist with ongoing tasks instead of answering one question at a time.Unlike traditional search tools that respond only when prompted, Google’s information agents are designed to operate continuously in the background, 24/7, helping users stay informed about their interests without needing to repeatedly search for the same information every day.Instead of delivering a list of links, the agents can synthesize information from multiple sources, explain why something matters, compare perspectives, and provide actionable insights. In many ways, the agents represent the next evolution of Google Alerts, the notification service Google launched in 2003. However, these agents are designed to go beyond simple notifications.For instance, someone following the stock market could create an information agent focused on specific companies, share price, or economic trends. The agent could monitor market activity throughout the day, track breaking news, summarize earnings reports, alert users when major changes happen, and provide summaries and links to learn more.Image Credits:GoogleIt could also help with everyday tasks, such as tracking flight prices for upcoming trips, monitoring sports teams and live events, following breaking news, keeping tabs on housing or job market trends, and tracking weather or traffic.To use the feature, users can open AI Mode in Search and enter a prompt.For example: “Keep me updated on nearby movie tickets for ‘The Mandalorian and Grogu.’” When something relevant appears, the Google app sends a push notification. You’ll also see your active tracked topics in your AI Mode history, where you can jump back in to manage, refine, or turn off an alert.Information agents will be available this summer. The company is first rolling them out to Google AI Pro and Ultra subscribers in the U.S., then to additional markets later on.In addition to these information agents, Google also introduced amajor redesign of Search, including what it describes as a reimagined “intelligent search box,” the company’s biggest change to Search in more than 25 years. The new interface is designed to support longer, more conversational queries. There’s also a new AI-powered query suggestion system that goes beyond traditional autocomplete, helping users craft nuanced and context-aware searches.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,AI agents,Google,google io 2026,Google SearchWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Google isn’t finished infusing AI into your inbox. On Tuesday attheir IO 2026 developer conference, the tech giant announced an expansion of its“AI Inbox” functionality for Gmail, which is adding conversational AI features. That means you can ask Gmail about things in your inbox instead of typing in search terms.The company says the Gemini AI-powered feature, called Gmail Live, will help you quickly find information buried in your inbox.Image Credits:GooglePerhaps you need information about your upcoming flight, the time of your dentist appointment, the door code for your Airbnb rental, or some details about an event at your kid’s school, for instance.Before, you’d have to type in keywords in the search box (or maybe type in someone’s email address or domain) to try to narrow down your search. That doesn’t always make emails easy to find, however, especially if the search term is something found across several messages.“Gmail Live can answer naturally phrased questions, respond to follow-up questions, and pivot if you need to interrupt it,” Devanshi Bhandari, product lead for Gmail, explained in a briefing ahead of Google’s annual developer conference, Google I/O, where the feature was first introduced to the public.It’s another way that Google is trying to showcase how its AI technology can drive real-world improvements to products used by millions of consumers, at a time when many outside the tech industry are questioning the value of AI, as new data centers get built in their backyards, driving up their power bills.Being able to point to something as simple as making it easier to find something that’s lost in your email inbox — an experience nearly everyone has suffered at some point — could be a practical and positive use case for AI … or at least, Google hopes.Bhandari demonstrated Gmail Live to reporters, asking the tool a series of questions about things in the inbox, like a child’s show-and-tell project and their class trip, plus hotel and flight information for a trip to Detroit. Similar to using a stand-alone AI chatbot like Gemini or ChatGPT, Gmail users can ask these questions aloud in natural language, and the chatbot responds.In the demo, Gmail Live also understood nuances between things like “field trip” and “trip” and was able to jump from one topic to another, Bhandari pointed out. Plus, the AI can pull granular details from emails, like a hotel room number, or infer which people you’re asking about, even when they’re not explicitly named.Similar voice technology is also coming to its to-do list, Google Keep, the company noted.Notably, Gmail Live is not replacing traditional Gmail search — it’s just another option.Google may have learned that not everyone is ready for an AI-only experience after it “upgraded”Google Photos with AI-powered searchto much backlash.Google Photos later rolled back the feature,making the use of AI optional after numerous complaints.Gmail is also gaining other new capabilities, including ready-to-send drafts, instant file access, and the ability to manage to-dos by marking individual tasks as done.Image Credits:GooglePlus, theAI Inbox experience, which launched earlier this year, will expand beyond Google AI Ultra subscribers to reach Google AI Pro and Plus subscribers as well. This allows you to see an overview of the tasks and items to catch up on that are buried in your inbox, all on one page.The voice-powered Gmail Live feature, however, will roll out later this summer and will initially be limited toGoogle AI Ultra subscribers.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,gemini ai,gmail,Google,google io,google io 2026,Google WorkspaceWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sarah PerezConsumer News EditorSarah has worked as a reporter for TechCrunch since August 2011. She joined the company after having previously spent over three years at ReadWriteWeb. Prior to her work as a reporter, Sarah worked in I.T. across a number of industries, including banking, retail and software.You can contact or verify outreach from Sarah by [email protected] via encrypted message at sarahperez.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One t

Google announced atits annual Google I/O eventon Tuesday that it’s launching Pics, a new AI-powered design and image-generation app for Google Workspace. The tech giant says it designed the app to be accessible to everyone, from teachers to small business owners.With Pics, users can generate everything from social media graphics and invitations to marketing materials and mock-ups using simple text prompts, without needing any editing skills or advanced tools. By giving users an easy way to generate visuals, Google is looking to take on popular design apps like Canva, as well as products from AI-native competitors likeClaude Designfrom Anthropic. Google’s entry into the space signals that AI-powered design is fast becoming a core competitive arena — with real stakes for any business that depends on visual content.The new app is launching to a group of testers at I/O and will be rolling out to Google AI Ultra subscribers this summer, Google says.The company acknowledges that although AI models today can generate high-quality images, it’s still difficult to modify just one part of an image. If you get an image that’s almost perfect but want to change a small detail, you have to write an entirely new prompt and hope the AI doesn’t alter too much. That’s why Pics not only generates images but also makes them easily editable.Loading the player…Users can enter a prompt, and Pics will generate what they need. Gemini powers the editing layer, making every element in a generated design or image fully adjustable. You can write a new prompt to make changes, but you can also simply click the part you want to change and leave a comment — much like leaving feedback in Google Docs.You can also edit directly, without leaving a comment or writing a prompt. For example, if you create a birthday party invitation and want to change the time listed on the card, you can do so manually.Pics is powered by Nano Banana 2, which Google says is a strong fit for the app because it supports precise text rendering, real-world knowledge, and detailed visual output. Pics is also built natively into Google Workspace, enabling visual collaboration across its apps.Once you’re happy with your design, you can download, copy, print, or share it with others. You can also pass it to someone else for a final round of edits before it goes out, Google says.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,Google,Google I/O,google io 2026,google picsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Nice cable you got thereIran demands Big Tech pay fees for undersea Internet cables in Strait of HormuzIran’s claim over subsea chokepoint pushes US tech companies to overland fiber.Jeremy Hsu–May 19, 2026 7:00 am|187Multicolored lines show undersea Internet cable routes running through the Strait of Hormuz.Credit:TeleGeographyMulticolored lines show undersea Internet cable routes running through the Strait of Hormuz.Credit:TeleGeographyText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navIran claims it will charge US tech companies fees for using undersea Internet cables that run beneath the contested Strait of Hormuz shipping lanes. The war has already halted multiple projects and led to the suspension of cable repairs in the region—and the latest Iranian threats may accelerate efforts by Big Tech and Gulf countries to find alternative routes for bypassing the Strait of Hormuz’s digital chokepoint.The latest assertions of Iranian authority over the Strait of Hormuz were announced in a brief statement by Ebrahim Zolfaghari, a spokesperson for Iran’s military and the Islamic Revolutionary Guard Corps. “We will impose fees on internet cables” Zolfaghari wrote in aMay 9 post. It was not immediately clear how Iran might implement such fees or impose its rules on cable projects, given that the majority of routes pass through Oman-controlled waters.But Tasnim and Fars, both Iranian state-linked media channels, laid out more detailed proposals on how Iran could charge license fees to US tech giants for the use and maintenance of undersea cables carrying regional Internet traffic, according toThe Guardian. For example, the Tasnim plan described charging tech companies—specifically naming Meta, Google, Amazon, and Microsoft—license fees for cable usage while also claiming that Iran alone has the right to repair and maintain the subsea cables.More than 99 percent of international Internet traffic runs through the global network of undersea cables that crisscross various oceans, connecting continents and islands. The major active cables running through the Strait of Hormuz primarily serve the Gulf countries in the region. They include theAsia Africa Europe-1,FALCON, and theGulf Bridge International Cable System, according toTeleGeography, a telecommunications research organization.The FALCON and Gulf Bridge cables run through Iranian territorial waters at certain points, Alan Mauldin, a research director for TeleGeography, said in aCNNinterview. CNN also reported that Iranian state media outlets have “issued veiled threats warning of damage to cables.”By comparison, there is “not much risk to Europe-Asia data traffic from hostilities in the Strait of Hormuz” because such Internet traffic primarily travels through cables in the Red Sea, TeleGeography wrote in a blog post. But undersea cables in the Red Sea have already seen a spate of damage in recent years, exacerbated by lengthy repair times and attacks by Houthi rebels allied with Iran.The biggest threat to the digital chokepointIran’s capability to threaten undersea cables as a way of imposing its will is uncertain at this point. Since the US and Israel started the war by attacking Iran on February 28, the US military claims to have destroyed 161 vessels of the Iranian navy, according to Admiral Brad Cooper, commander of US Central Command, intestimonygiven to the Senate Armed Services Committee on May 14. The US military has also sunk some of thesmall fast boatsoperated by Iran’s paramilitary Islamic Revolutionary Guard Corps when the latter threatened civilian shipping in and around the Strait of Hormuz.Thevast majority of damageto undersea cables comes from commercial ships accidentally dragging their anchors or fishing trawlers dragging weighted nets along the seafloor. So it’s not out of the realm of possibility for a more innocuous-looking Iranian ship to sneakily perform somesubsea cable sabotageif it’s willing to run the gauntlet of US military surveillance and patrols in the strait.Even a damaged commercial ship abandoned in the strait could end up dragging its anchor across some cables, as was the case with a drifting ship thatdamaged three cablesfollowing an attack by Houthi rebels in the Red Sea in 2024.However, the greatest threat to subsea cable infrastructure in the Strait of Hormuz may simply come from delays in any necessary cable repairs in the region. Such jobs require specialized ships to find the damaged area and lower grappling hooks to lift up the cable for inspection and repair, according toBBC News. That repair process can require days or sometimes weeks, which would leave the ship vulnerable to Iranian missiles, drones, or fast boats that havecontinued to attackcommercial shipping in and around the strait.“Operators face a choice: pay protection fees and accept Iranian licensing over Middle East Gulf seabed activity, or accept that future faults may go unrepai

ConsolidationElectrical utility megamerger is all about the data centersNextEra’s blockbuster deal with Dominion likely means higher bills for consumers.Dan Gearino, Amy Green, and Charles Paullin, Inside Climate News–May 19, 2026 9:45 am|59The campus of NextEra’s headquarters is seen on Monday in Juno Beach, Fla.Credit:

      Marco Bello/Getty ImagesThe campus of NextEra’s headquarters is seen on Monday in Juno Beach, Fla.Credit:

      Marco Bello/Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA proposed merger of the largest utility in the country by market value, NextEra Energy, with the sixth-largest, Dominion, would create a megacompany at a time when data centers and rapid increases in electricity demand are reshaping the industry.The proposal,announced Monday morningand contingent on state and federal regulatory approval, would result in a company that leads in nearly every aspect of the US power and utility industry, including overall electricity generation, natural gas generation, and renewables.The $67 billion deal combines NextEra’s size and reach with Dominion’s positioning as the local utility for the world’s largest concentration of data centers innorthern Virginia. But the results are likely bad for consumers and the environment, creating a company with enormous financial and political strength that will be difficult to effectively regulate, according to consumer advocates and analysts.For perspective, only Exxon Mobil and Chevron would be larger based on market value among US-based energy companies.“Mergers are not about consumers; they’re about shareholders,” said Ari Peskoe, director of the Electricity Law Initiative at Harvard Law School. “For the Dominion shareholders, they are selling their shares at a premium. The executives are getting massive payouts for facilitating this, assuming it all goes through, and obviously NextEra believes the transaction is going to add value to the company. Ratepayers are all an afterthought.”The deal makes financial sense for both companies, said Andrew Bischof, an equity analyst for Morningstar.“We view the transaction as allowing NextEra to accelerate its data center ambitions, which had trailed those of its regulated peers, by using Dominion’s expertise and relationships to expedite NextEra’s data center hub plans,” he said in a note to clients.NextEra, based in Juno Beach, Florida, includes Florida Power & Light, the largest regulated electricity utility in the state, and NextEra Energy Resources, a wholesale electricity supplier that owns power plants across the nation. Dominion, based in Richmond, Virginia, includes regulated utilities serving much of Virginia, parts of North Carolina and South Carolina, and other assets across the country.The company would be called NextEra Energy, and NextEra CEO John W. Ketchum would serve in the same role after the deal closes. Robert M. Blue, Dominion’s CEO, would be the CEO for regulated utilities for the merged company. The parties said they expect regulatory approvals to take 12 to 18 months.A page from NextEra Energy and Dominion’s investor presentation shows the states where each company has regulated utilities and the merged company’s ranking in various categories.Credit:
                                  NextEra EnergyA page from NextEra Energy and Dominion’s investor presentation shows the states where each company has regulated utilities and the merged company’s ranking in various categories.Credit:

      NextEra EnergyNextEra shareholders would own 74.5 percent and Dominion shareholders would own 25.5 percent, respectively, of the combined company in the all-stock transaction.“We are bringing NextEra Energy and Dominion Energy together because scale matters more than ever—not for the sake of size, but because scale translates into capital and operating efficiencies,” Ketchum said in a statement.“Adding to the pollution problem”The post-merger NextEra would be the leader in so many categories in the US utility sector that it’s easier to list the ones where it wouldn’t be on top. It would rank second in nuclear power generating capacity and in the number of regulated utility customers, trailing Exelon Corp. of Chicago in both.NextEra and Dominion both have substantial carbon emissions, but neither was among the top five utility companies in the country in 2024, according to the most recent edition ofBenchmarking Air Emissions reportfrom the Natural Resources Defense Council. NextEra ranked sixth and Dominion ranked 11th, and their sum was less than that of each of the leaders, Vistra Energy and Duke Energy.But those are still massive emissions from a company that stands to gain more clout because of its size.“If we continue to add dangerous climate pollution into the mix, then people who are already suffering and are typically hurt first and worst will su

Catching the satellite busIn addition to space stations, Vast says it will now build high-power satellites“Every single successful space company is diversified in its products.”Eric Berger–May 19, 2026 9:00 am|26A rendering of Vast’s 15 kW satellite.Credit:

      Vast SpaceA rendering of Vast's 15 kW satellite.Credit:

      Vast SpaceText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAs part of its plan to develop a private space station, Vast Space built and then launched a small demonstration spacecraft in early November. This vehicle then completed dozens of test objectives with flying colors before making a successful de-orbit three months later.The mission, which tested power, propulsion, tracking, and a multitude of other technologies needed for Vast’s Haven-1 space station, was evidently so successful that the company is ready to use its spaceflight capabilities for other purposes. The Long Beach, California-based company announced Tuesday that it plans to begin selling high-powered satellite buses.“Every single successful space company is diversified in its products,” said Max Haot, chief executive of Vast Space, in an interview. “So for us it really was a question of when, not if.”A satellite for power-hungry applicationsThe company’s first offering is a 15 kW-class satellite bus capable of supporting various demanding missions. Each satellite is about 3 meters long and 4 meters tall, with a mass of 700 kg and payload capacity of at least 350 kg. They will have a design lifetime of five years and be intended to operate in areas ranging from low-Earth orbit out to lunar orbit. Vast aims to serve various customers, from telecommunications to observation to data services. Haot added that Vast also plans to offer an Nvidia Space-1 Vera Rubin Module to support orbital data center inferencing needs.The Vast satellite bus—essentially a backbone providing power, propulsion, and navigation for various payloads—will be based largely on technology ported over from the company’s Haven-1 space station, which is due to launch for the first time next year as the world’s first private space station. However, there will be some new elements needed for the satellite, and Haot said Vast is already moving forward with in-house development of electric propulsion and a deployable solar array for the satellite.Vast has already signed a customer for four satellites, with an option to purchase up to 200 additional satellites. Haot said the company is targeting a launch of at least 10 Vast Satellites in the fourth quarter of 2027.With this new product line, Vast is entering an increasingly crowded market. Historically, in the United States, a handful of large companies such as Boeing, Lockheed Martin, Northrop Grumman, Maxar, and Sierra Space have manufactured medium and large satellites. Typically, these were costly and often bespoke designs that cost tens to often hundreds of millions of dollars.But a few trends have changed the landscape in recent years. The US government’s Space Development Agency has signaled that it prefers proliferated constellations—many satellites spread out present less of a concentrated target than a few larger and more expensive satellites. With the Falcon 9 rocket’s increased cadence, as well as rideshare missions, it became easier and sometimes cheaper to get smaller and medium-sized satellites into orbit.This has led to an influx of venture capital to back a new generation of companies seeking to build less expensive, more modular satellites that could fill a variety of purposes. There are several prominent, relatively new entrants in this area, including K2 Space, Rocket Lab, True Anomaly, Blue Canyon, and Millennium Space Systems.Vast already has… vast facilitiesHaot said most of these companies are still emerging, with products that are not yet mature. In other words, he believes that if Vast Space can execute, it could become a market leader, especially in power-hungry applications. Vast has already invested $1 billion in facilities for spacecraft manufacturing, including clean rooms, which can be used for space stations as well as satellites, he said.The number of satellites in space has exploded in recent years, largely due to the rapid expansion of SpaceX’s Starlink constellation. For decades, the total number of satellites orbiting Earth numbered about 4,000, but in the last five years that number has grown to about 14,000.This is just the beginning. By some estimates, in another decade, there will be approximately 500,000 satellites in orbit for communications, Earth observation, orbital data centers, and other applications.Haot expects that about 90 percent of these will be satellites built by SpaceX, Amazon, Blue Origin, or other major players. But even 10 percent of that number, if available to commercial satellite bus manufacturers, would represent 50,000 satellites f

Test of TimeCivilization VIIfinally lets you build a civ that stands the test of timeCiv 7’s devs talk walking back the game’s most controversial decision.Samuel Axon–May 19, 2026 12:00 pm|40You can’t tell as much from the map view, but many ofCiv 7’s key underlying systems have been totally overhauled.Credit:

      Samuel AxonYou can't tell as much from the map view, but many ofCiv 7's key underlying systems have been totally overhauled.Credit:

      Samuel AxonText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to nav“Build a civilization to stand the test of time.” That was the promise on the box of Sid Meier’sCivilization, the first in a long-running strategy game franchise that has evolved over 35 years and seven mainline entries.Civ 7introduced a new approach to play wherein players would change civilizations from their initial selection twice by the end of a game. Lots of players said, “Wait a minute: we’re literally not building a civilization to stand the test of time anymore.” After such a negative reception at launch, longtime series fans began to wonder whether the franchise would continue to stand the test of time.It’s clearly not a coincidence that the new, major update for the game reaching players today is titled “Test of Time.” It’s a major reworking of several of the game’s key systems, and it reintroduces the ability to play one civ from beginning to end while retaining some of the big ideas that definedCiv 7at launch.I’ve spent the past week playing the game with the Test of Time update, and I spoke with two of the game’s senior developers—the franchise’s creative director, Ed Beach, and series head of product Matt Schembari—about the choices they made, reacting to player feedback, and competing ideas about what Civ is to players.A quick note on agesFor those who haven’t been followingCiv 7up to this point: Another major change to the game is that it has been broken into three distinct ages: Antiquity, Exploration, and Modern. Each has its own tech and culture tree, game mechanics, and, until now, civilizations to pick from.The intent behind it was to allow the game’s designers to solve some longtime problems with how traditional Civ games could lag or have balance problems at specific stages. If that in itself is too much deviation from the classic formula for you, it’s unlikely Test of Time will make you likeCiv 7any more than you might have a year ago. It’s still a foundational framework for this game. Test of Time gives players a lot more freedom inside that structure, but it does not remove the structure itself.Give the people what they wantStarting a new game ofCiv 7with the patch, the biggest change was immediately apparent. As of the last update before this one, you could pick from just 14 civs when starting a game in the Antiquity Age. This time, I saw more than 40.I could play as America from the Age of Antiquity now, for example. When I set up the game, I was also given the option to choose the AI’s civ-switching behavior; I could prohibit it, allow it, or pick an option that makes the AI follow suit on whatever choice I make.Yep, we have modern age civs as options for a game starting in antiquity.Credit:
                                  Samuel AxonYep, we have modern age civs as options for a game starting in antiquity.Credit:

      Samuel AxonOf course, playing one civ throughout the three ages would seem to run counter to how the game was originally designed. Firaxis took this update as an opportunity to take a major game balance pass across all the civs, and it introduced some new concepts to make one civ for three ages fit theCiv 7framework.Civs are considered to have an “apex age,” during which they have access to their full kit. When they’re in an age other than their apex age, they’re called “time-tested civs,” and they have access to a limited amount of their normal bonuses. But players can use a system called syncretism to adopt bonuses that are normally offered to specific civs that would call the current age their apex age.Based on playing one full game in the new patch, I feel this implementation works really well overall, and it’s a lot of fun.That said, I could see it getting unwieldy to build on over time—the game’s designers now must account for many more possibilities and balance them against each other to make sure the game stays fair and fun.When I asked Beach about this, he suggested it’s not as big a design challenge as it might seem. “I think we still have some advantages because of the structure of the game,” he said. “I think where we are giving ourselves an immense challenge is the content space, because we have 43 civs that you can play as, and now we need to give them bonuses in three different ages. So we went from basically 43 chunks of content that we need to balance to like 129 chunks of content, and as we add more civs

more taxesEV drivers will pay $130 a year under Congress’ 2026 transportation billPoliticians say they want EVs to pay “their fair share for the use of our roads.”Jonathan M. Gitlin–May 19, 2026 12:12 pm|250Credit:

      Getty ImagesCredit:

      Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe119th Congressmight be one of themost dysfunctional and least productivelegislative sessions in the 250-year history of the United States, but it seems there’s one thing it can agree on: Electric vehicles don’t cost their owners enough money. The Transportation and Infrastructure committee has published its bill to fund surface transportation for the next half-decade, and among the provisions in the “Building Unrivaled Infrastructure and Long-term Development for America’s 250th Act” is an annual fee levied against owners of EVs.“I’m extremely proud of the historic level of investment in America’s bridges—at more than $50 billion, it’s the largest such investment in our history. And the BUILD America 250 Act ensures that electric vehicle owners begin paying their fair share for the use of our roads,” said committee chairperson Sam Graves (R-Mo.).Should the bill pass—and it enjoys support from the Democratic Party, too—you will be required to pay a $130 federal registration fee to drive an EV. And starting in 2029, that fee will increase by $5 each year until it reaches $150. Plug-in hybrids don’t escape untaxed, either; the fee for a PHEV begins at $35 a year and will escalate by $5 each year until it reaches $50 annually. And if state departments of transport don’t collect this federal EV tax, the federal government will “withhold an amount equal to 125 percent of the amount owed from the state’s highway apportionment.”Why?Road funding in the US is a messy business, paid for by a mix of federal and state gas taxes, state vehicle registration fees, and local budgets. And as the percentage of fuel-efficient PHEVs and gas-abstaining EVs grows, the amount of money available to pay for road upkeep drops. So, in the abstract, asking EV and PHEV owners to pay their share is not unreasonable, especially since the slightly higher curb weight of these vehicles will do slightly more damage to the roads, although passenger vehicles are a rounding error compared to a garbage truck, bus, or tractor-trailer.Some stateshave already started imposing a registration fee on clean vehicles to offset reduced gas tax receipts. And at some point, when a significant proportion of the country’s vehicles are EVs or PHEVs, an alternative road-funding method that does not rely on the gas tax will certainly be in order.But EV adoption in the US hastaken a severe hitsince the election of President Trump and the abolition of federal clean-vehicle incentives. As of now, EVs make up barely 3 percent of the entire vehicle fleet. Meanwhile, the federal gas tax has not been increased since 1993; had it been adjusted to keep pace with inflation, it would add an extra 42.34c to a gallon rather than a pathetic 18.4 cents currently.“Drivers of gas-powered vehicles pay approximately $73 to $89 in federal gas tax each year. The proposed fee would charge an unfair premium on EV drivers at a time when all Americans are looking for ways to save money,” said Albert Gore, executive director of the Zero Emissions Transportation Association. “This is particularly concerning as the EV fee will increase to $150 by 2035—nearly double what gas car drivers would pay in a year. This fee lands on top of the road use taxes that many EV drivers already pay at the state level.”“In addition to the new EV fee, the bill makes major changes to federal investments for dedicated EV charging. A lack of dedicated EV funding could particularly impact EV drivers living in—and traveling to—more rural and remote locations in the United States, where federal funding is most needed to help fill gaps in existing charging infrastructure,” Gore said.“With Americans demanding more affordable transportation options as gasoline tops $4.50 a gallon, this bill would impose an onerous new fee on electric vehicles and plug-in hybrids while slashing investments in new EV chargers,” said Shruti Vaidyanathan, director of federal and state transportation advocacy at the National Resources Defense Council. “Congress should be boosting investments in projects that cut costs, cut emissions, create jobs, and build a transportation system that works for all Americans. This bill largely ignores the need to build cleaner, more affordable transportation options.”We should perhaps be glad that the final version will cost drivers much less than Rep Graves’last proposal, which started at $250/year for EVs and $100/year for hybrids.Jonathan M. GitlinAutomotive EditorJonathan M. GitlinAutomotive EditorJonathan is the Automotive Editor at Ars Technica. He has a BSc and PhD in Pharmacolog

Gotta go fastThe era of 1,000 Hz gaming monitors has arrived, but why?LG’s latest hits one frame per millisecond at a full 1080p resolution.Kyle Orland–May 19, 2026 12:26 pm|83Wow, look how fast those giant letters are moving without any apparent motion blur!Credit:

      LGWow, look how fast those giant letters are moving without any apparent motion blur!Credit:

      LGText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAlmost exactly two years ago, we weregawking at prototypes of 1,000 Hz monitorsand wondering who really needed a display that could support such ludicrously smooth frame rates. Now that those prototypes are starting to develop into retail products, we’re still wondering how much of a market there is for gaming displays that can update with a new frame every single millisecond.The latest entry in the ultra-fast refresh race is LG’s 24.5″ UltraGear 25G590B, which the companyannounced this weekas “the world’s first Full HD gaming monitor with a native 1000Hz refresh rate” ahead of a planned launch in “select markets” in the second half of the year. That “Full HD” promise means LG’s 1,000 Hz display hits the 1080p threshold that is by far the most common resolution reported by gamers inSteam’s regular hardware surveys.That would represent a decent step up from the likes ofAcer’s Predator XB273U F6,Samsung’s Odyssey G6, orPhillips’ EVNIA 27M2N5500XD, all of which have to shift down to a relatively blurry 720p resolution to run at a full 1,000 Hz (but which support 1440p resolutions at a still-quite-fast 500 Hz). LG also notes that its high-end monitor can hit its resolution and refresh rate specs natively, without the need for any “dual mode” rebooting shenanigans to get the fastest performance.Fast-moving objects appear decidedly less blurry as you climb the frame rate curve.Credit:Blur BustersFast-moving objects appear decidedly less blurry as you climb the frame rate curve.Credit:Blur BustersThe folks over at Blur Busters haveextensivelydocumentedresearch showing that refresh rates of 1,000 Hz (and up) can reduce human perception of motion blur and flickering. And while the site notes that you eventually hit “diminishing returns” from all those extra frames—especially on smaller screens—there’ssome evidencethatyou would need a 40,000 Hz monitorto totally eliminate perceived motion blur on a sufficiently large, high-resolution monitor.Perception research aside, it’s unclear how many gamers are interested in spending extra money on a display that breaks the four-digit refresh rate barrier. While LG hasn’t announced pricing details for the UltraGear 25G590B yet, other 1,000 Hz monitors tend to run hundreds of dollars more than comparable models with lower (but still more than adequate) maximum refresh rates in the 240 to 360 Hz range. At the same time,monitors with Nvidia’s GSync Pulsarand similar technologies use “rolling scan” pulsing backlights to reduce apparent motion blur, even on “slower” 360 Hz monitors.LG says the 25G590B is “engineered specifically for first-person shooters” and mentions an “esports-optimized design” in its marketing materials, suggesting the display is targeting the small subset of gamers who play twitch-based reflex games for a living. For everyone else, we recommend ignoring the “larger number is better” impulse and saving your hardware budget forthose increasingly expensive sticks of RAM.Kyle OrlandSenior Gaming EditorKyle OrlandSenior Gaming EditorKyle Orland has been the Senior Gaming Editor at Ars Technica since 2012, writing primarily about the business, tech, and culture behind video games. He has journalism and computer science degrees from University of Maryland. He oncewrote a whole book aboutMinesweeper.83 Comments

Gemini’s efficiency playGemini 3.5 Flash might be fast enough for gen AI to make senseGoogle says its more efficient Gemini 3.5 Flash is the key to your agentic AI future.Ryan Whitwam–May 19, 2026 2:11 pm|59Credit:

      Aurich LawsonCredit:

      Aurich LawsonText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAt last year’s I/O event, Google was still talking about the2.5 branchof Gemini, and what a difference a year makes. We’ve gone through the3.0and3.1families since then, and now it’s on to version 3.5.Gemini 3.5 Flashis rolling out across a wide range of Google products starting today, and Google again claims this model is even better than its last-gen Pro model.That has been a trend with Google’s tick-tock model updates over the past year, but the team says this release is special. Gemini 3.5 Flash supposedly offers frontier-level intelligence while also being efficient enough that it may finally make complex agentic tasks worth doing at scale. Tulsee Doshi, senior director of product management for Gemini, explains that the innovations of Gemini 3.5 Flash are woven through multiple Google products, and this is just the start.Credit:
                                  GoogleCredit:

      GoogleIt’s no secret that generative AI is currently a money pit, and all the major AI players are trying to find paths to greater efficiency. The problem is magnified when you start building agentic experiences that are supposed to run for longer to complete complex tasks. Gemini 3.5 Flash may be a big step toward making that viable. The new model can output nearly 300 tokens per second, but its benchmark scores are similar to larger frontier models (like 3.1 Pro) that build outputs at a quarter of that speed.Google now says that the companies using the most AI tokens could save a billion dollars per year by shifting to the more efficient Gemini 3.5 Flash. API pricing for the new model is significantly lower than the Pro model it apes. Gemini 3.5 Flash clocks in at $1.50 per 1M input tokens and $9 per 1M output tokens. The 3.1 Pro model starts at $2 and $12, respectively, and it’s higher if you use more than 200k tokens.According to Doshi, the team made numerous improvements in pre-training with Gemini 3.5 Flash, but insights gleaned from how devs use Gemini models are really paying off.“With post-training, we’re really starting to unlock some of the value of the feedback we’re getting from users, for example, from Antigravity,” said Doshi. “That’s really what you’re seeing play out in terms of the code performance and the tool use performance. And then, the hope is that you’ll continue to see the step change where 3.5 Pro will be better, and the next Flash meets Pro performance with that series.”Google is focused on code generation with the new model, which is a core agentic angle for AI. Both Terminal Bench and SWE-Bench Pro tests show substantial improvements—3.5 Flash clobbers older Flash models and shows a small but measurable improvement versus Gemini 3.1 Pro. Its scores are in the same neighborhood as OpenAI’s much larger and more expensive GPT 5.5.A major barrier in agentic workflows is how generative models can use interfaces designed for humans. It’s not an easy problem to solve, Doshi said. “Certain things like UI control are expensive to do because the model has to search the page, it has to know where to click, it has to act through multiple steps. I think Flash is able to do that well because of that combination of quality and cost.”Google’s AI evaluations demonstrate these improvements, too. Among Google’s current collection of benchmarks is OSWorld-Verified, which tests how models handle general tasks in real computing environments. It’s similar to the coding improvements. Gemini 3.5 Flash substantially outperforms older Flash models and is even a bit faster than Gemini 3.1 Pro. It’s essentially tied with GPT 5.5.Google’s new Flash model is, again, a little better than the last-gen Pro.Credit:
                                  GoogleGoogle’s new Flash model is, again, a little better than the last-gen Pro.Credit:

      GoogleGemini 3.5 Flash has been rolled out internally at Google, and Doshi noted that it’s having a big impact. “We have a set of internal metrics we’ve been evaluating that measures how Googlers code, so looking at our own code bases and how well the models perform on that,” Doshi said. “And you can see a massive, massive jump between where 3.1 Pro was and where 3.5 Flash is.”Google unveiled the Antigravity IDE last year, and it’s being upgraded to version 2.0 with support for Gemini 3.5 Flash. This update will support multiple parallel workflows—essentially sub-agents spawned by Gemini 3.5 Flash. Again, Google says this is only possible because the new model is so efficient at spitting out tokens.In addition to Antigravity, Gemini 3.5

Foiled againRFK Jr. forced to withdraw charter that opened CDC panel to anti-vaccine quacksCharter would have expanded member eligibility and focused on alleged injuries.Beth Mole–May 19, 2026 2:17 pm|48WASHINGTON, DC - JUNE 24: Health and Human Services (HHS) Secretary Robert F. Kennedy Jr. testifies before the House Energy and Commerce Committee Health Subcommittee in the Rayburn House Office Building on June 24, 2025 in Washington, DC.Credit:Getty | Kayla BartkowskiWASHINGTON, DC - JUNE 24: Health and Human Services (HHS) Secretary Robert F. Kennedy Jr. testifies before the House Energy and Commerce Committee Health Subcommittee in the Rayburn House Office Building on June 24, 2025 in Washington, DC.Credit:Getty | Kayla BartkowskiText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navA revised charter document for the Centers for Disease Control and Prevention’s influential vaccine advisory committee has been withdrawn by the Health Department over an administrative error, according toa notice published in the Federal Register Tuesday.The charter’s revisionsunder anti-vaccine Health Secretary Robert F. Kennedy Jr. would have allowed Kennedy to appoint dubiously qualified anti-vaccine allies to advise the CDC. It also would have directed the CDC panel to focus on alleged vaccine injuries and risks and welcomed fringe groups and anti-vaccine organizations to participate in developing federal vaccine policy.Kennedy’s move to reshape the CDC panel—the Advisory Committee on Immunization Practices, or ACIP—came amid Kennedy’s many other attempts to undermine it, as well as a court order to undo that meddling.In June 2025,Kennedy summarily fired all 17 expertsfrom ACIP and quickly replaced them with unvetted and unqualified anti-vaccine allies. Kennedy’s ACIP then held several meetings in which they aired anti-vaccine views and misinformation, allowed anti-vaccine activists to give unvetted presentations, and ultimately voted to remove longstanding, evidence-based federal vaccine recommendations based on anti-vaccine rhetoric. That includes the removal of a universal recommendation for a hepatitis B vaccine dose at birth, despite no evidence of any safety concerns or any benefit to delaying the dose. Subsequent modeling studies found that the change will meanmore infections, increases in liver cancers and deaths, as well as millions of dollars inhealthcare costs.In January, Kennedy sidestepped ACIP entirely to overhaul the CDC’s vaccine schedule for children,slashing the number of recommended vaccinations from 17 to 11.In March, a federal judge issueda temporary injunction to undo Kennedy’s changes, ruling that Kennedy likely violated federal regulations in unilaterally altering the CDC vaccine schedule and that the new ACIP members were unqualified—as outlined by the panel’s charter—nullifying all of their votes on federal vaccine recommendations.While the Health Department is working to appeal the injunction, Kennedy attempted to circumvent the judge’s ruling on the ACIP members byaltering the committee’s charterto, among other things, allow for people without expertise in immunizations and public health to be members.But, for now, that effort, too, has been thwarted. According to the notice on Tuesday, the new charter has been withdrawn for not following a federal requirement on public notification.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.48 Comments

GitFlubIn stunning display of stupid, secret CISA credentials found in public GitHub repoSSH keys, plaintext passwords, other sensitive data had been up since November 2025.Lee Hutchinson–May 19, 2026 2:27 pm|75Only the best people.Credit:

      Dzmitry Skazau / GettyOnly the best people.Credit:

      Dzmitry Skazau / GettyText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navSecurity researcher Brian Krebsbrings us the newsthat America’sCybersecurity & Infrastructure Agency(CISA) has had a large store of plaintext passwords, SSH private keys, tokens, and “other sensitive CISA assets” exposed in a public GitHub repo since at least November 2025.The now-offline public repo—named, somewhat aspirationally, “Private-CISA”—was brought to Krebs’ attention by GitGuardian’sGuillaume Valadon, who was alerted to the repo’s presence by GitGuardian’s public code scans. Krebs says that Valadon approached him after receiving no responses from the Private-CISA repo’s owner.In an email to Krebs, Valadon claimed that the repo’s commit logs show that GitHub’s default protections against committing secrets—protections designed to protect unwitting or unskilled developers against exactly this kind of stupidness—had been disabled by the repo’s administrator.Testing bySeralys founder Philippe Catureglishowed that this was not a joke or hoax and that he was able to use the credentials in the Private-CISA repo to gain access to multiple Amazon Web Services GovCloud accounts “at a high privilege level.”Krebs notes that the repo appeared to be managed by Virginia-basedNightwing, a CISA contractor. Nightwing has so far not commented publicly, instead referring questions back to CISA.This isn’t the first time CISA has screwed up—in fact, it’s not even the first timethis year. In January,polygraph-failingacting CISA Director Madhu Gottumukkalauploaded sensitive government documents to ChatGPTafter demanding and receiving an exemption to the agency policy that prohibited ChatGPT’s use by CISA personnel. Gottumukkala wasremoved from his role in February.Lee HutchinsonSenior Technology EditorLee HutchinsonSenior Technology EditorLee is the Senior Technology Editor, and oversees story development for the gadget, culture, IT, and video sections of Ars Technica. A long-time member of the Ars OpenForum with an extensive background in enterprise storage and security, he lives in Houston.75 Comments

What is real?Google’s SynthID AI watermarking tech is being adopted by OpenAI, Nvidia, and moreAI content is getting good, but SynthID might be able to help tell truth from fiction.Ryan Whitwam–May 19, 2026 2:37 pm|46Google CEO Sundar Pichai announces 100 billion SynthID images and videos.Credit:

      GoogleGoogle CEO Sundar Pichai announces 100 billion SynthID images and videos.Credit:

      GoogleText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navIn a few short years, we’ve gone from easily identifying AI content that featured superfluous fingers to images and videos that lookshockingly realistic. How can we know what’s real in the age of AI? Google’s answer is SynthID, which it first demonstrated three years ago.The company saysSynthID has since been used to label 100 billion images and videos, plus 60,000 years’ worth of audio. Those numbers are only going up now that SynthID is expanding beyond Google.SynthID is not Google’s only AI labeling strategy. It’s also committed to the C2PA standard, which tags content with metadata describing how it was created. Google began using C2PA more prominently with its Pixel 10 smartphones. Photos taken with the Pixel 10 include metadata describing how they were processed. If a highly zoomed image includes generative elements, it gets an AI tag, too.Google now says this same feature is coming to videos recorded on Pixel 8, 9, and 10 phones in an update in the coming weeks. It’s also adding C2PA scanning to Gemini, allowing the chatbot to explain a file’s providence based on the content labeling. This same capability will come to Chrome and Search in a few months.That metadata is fungible, though. On the other hand, SynthID is deeply integrated with AI-generated content. The digital watermark is present in the pixels of images and videos and in the waveform of AI songs and audio overviews from products like NotebookLM. According to Google DeepMind scientist Pushmeet Kohli, the team worked hard to ensure SynthID is much harder to remove, even if you compress it, crop it, or rotate it.“A technology like this will always be attacked,” said Kohli. “There was a lot of research that we did in making SynthID robust to different kinds of transformations.”Last year, Google added support for SynthID detection in the Gemini app. You can upload the suspect content and ask the chatbot if it’s AI-generated. This should work reliably with all those billions of Google AI images and audio clips from the past three years. A few ambitious tinkerers have claimed to find methods for removing the hidden SynthID patterns. Google contends that none of these bypasses actually work.More SynthID in more placesEven if no one has been able to crack SynthID, that doesn’t matter for the vast majority of AI images on the Internet—only Google’s AI models apply SynthID. That’s going to change soon, though.Google has announced that it has partnered with several companies to add SynthID to their systems. Nvidia will implement SynthID in its Cosmos world foundation models, and OpenAI will use SynthID in its GPT 2 images. Kakao and ElevenLabs will also begin adding SynthID to their AI content.This doesn’t mean you’ll always be able to tell if something is AI by looking for SynthID. Plenty of publicly available models continue to produce content with no AI watermarking, and there are open models that can be trained by anyone looking to create AI images and videos on their own terms. Still, this is a step in the right direction.There will also be new paths to checking SynthID status, so you won’t even have to open Gemini just to check for the watermark. SynthID will be integrated with Circle to Search, Lens, and AI Mode. You’ll also be able to use Gemini in Chrome by sharing a tab with the content in question. You can ask any variation of “Is this AI” to get a SynthID scan with these tools.There is currently not a public API for SynthID—making these scans too readily available could serve as an attack vector for those seeking to circumvent SynthID. However, Google is preparing to launch an AI content detection API as part of the company’s Gemini Enterprise Agent Platform. This will allow trusted business partners to more easily flag AI content, allowing Google to refine the API over the coming months.Ryan WhitwamSenior Technology ReporterRyan WhitwamSenior Technology ReporterRyan Whitwam is a senior technology reporter at Ars Technica, covering the ways Google, AI, and mobile technology continue to change the world. Over his 20-year career, he's written for Android Police, ExtremeTech, Wirecutter, NY Times, and more. He has reviewed more phones than most people will ever own. You canfollow him on Bluesky, where you will see photos of his dozens of mechanical keyboards.46 Comments

We can help with thatTwo AI-based science assistants succeed with drug-retargeting tasksBoth tools generate hypotheses; one goes on to analyze some of the data.John Timmer–May 19, 2026 2:55 pm|8Finding connections within the messy world of biology is central to these new tools.Credit:Andriy OnufriyenkoFinding connections within the messy world of biology is central to these new tools.Credit:Andriy OnufriyenkoText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navOn Tuesday, Nature released two papers describing AI systems intended to help scientists develop and test hypotheses. One, Google’s Co-Scientist, is designed as what they term “scientist in the loop,” meaning researchers are regularly applying their judgments to direct the system. The second, from a nonprofit called FutureHouse, goes a step beyond and has trained a system that can evaluate biological data coming from some specific classes of experiments.While Google says its system will also work for physics, both groups exclusively present biological data, and largely straightforward hypotheses—this drug will work for that. So, this is not an attempt to replace either scientists or the scientific process. Instead, it’s meant to help with what current AIs are best at: chewing through massive amounts of information that humans would struggle to come to grips with.What’s this good for?There are some distinctions between the two systems, but both are what is termed agentic; they operate in the background by calling out to separate tools. (Microsoft has taken a similar approach with its science assistant as well; OpenAI seems to be an exception in that it simplytuned an LLM for biology.) And, while there are differences between them that we’ll highlight, they are both focused on the same general issue: the utter profusion of scientific information.With the ease of online publishing, the number of journals has exploded, and with them the number of papers. It has gotten tough for any researcher to stay on top of their field. Finding potentially relevant material in other fields is a real challenge. If you’re focused on eye development, for example, one of the signaling systems used may also be involved in the kidney, and it can be easy to miss what people are discovering about it.As the people at FutureHouse put this issue, “By focusing on ‘combinatorial synthesis’ (identifying non-obvious connections between disparate fields), Robin effectively targets ‘low-hanging fruit’ that human experts may overlook due to the compartmentalization of scientific knowledge.”This is a task that’s well-suited to AI, which can chew through the peer-reviewed literature in the background while researchers do other things. This isn’t really a question of whether an AI could do something better or worse than a human; it’s more of an issue of whether any human would end up doing these sorts of searches at all.By finding enough connections among disparate research, these tools can make suggestions—hypotheses, really—about the biology. This can include things like what processes underlie biological behaviors and what pathways and networks regulate those processes. And, in the cases explored here, it included suggesting known drugs that might target some of these pathways in diseased cells: acute myeloid leukemia in Google’s case, and a form of macular degeneration for FutureHouse.Co-scientistAs you might imagine, Google’s system is based on the company’s Gemini large language model. That helps the system interpret a statement of research goals provided by human scientists and starts a literature search to find relevant information and form hypotheses. Those are then evaluated relative to each other in a “tournament,” the results of which are evaluated by a Reflection agent. An Evolution agent can then make improvements to any surviving ideas, which can be sent back through the process.Key criteria considered throughout this process include plausibility, novelty, testability, and safety. And the Reflection tool has access to external search tools, as access to the scientific literature “prevented the hallucination of seemingly novel but implausible hypotheses,” the company wrote.As the paper puts it, scientists were kept in the loop at all times. In the search for potential drugs targeting leukemia, the suggestions made by the system were prioritized based on a review by a panel of experts, who had access to the literature Co-Scientist used to formulate its suggestions.The results are what you would expect from cancer therapies. Some of the drugs identified were effective, but only against subsets of a panel of myeloid leukemia cells. That’s not unusual, given that there are multiple routes to unchecked growth, so drugs that block the route followed by one cell type may not be effective in cells that took a different route.Google also mentioned that the system could do more general hypothesizing that do

Self-hostingPlex’s 200% Lifetime Pass price hike tries forcing users to another subscriptionPlex says that it has considered getting rid of Lifetime Passes.Scharon Harding–May 19, 2026 3:52 pm|144A marketing image for Plex that emphasizes its streaming, rather than its media server, business.Credit:

      PlexA marketing image for Plex that emphasizes its streaming, rather than its media server, business.Credit:

      PlexText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAs of July 1, at 12:01 am UTC—or June 30 at 8:01 pm ET—people seeking access to Plex’s media server features through a one-time purchase will have to pay $750. That’s three times the current price of $250.The new price will not affect current Lifetime Plex Pass holders.A Lifetime Plex Pass allows you to stream from your own Plex Media Server to a device connected to your own network, to stream from the server remotely, and to allow others to stream remotely from your server.When first launched in 2012, a Lifetime Plex Pass was $75. In 2014, Plex increased the price to $150 because “at 2.5x the price of a yearly, it was priced in an unsustainable way for us,” Plex said in ablog postat the time. Eventually, lifetime passes were available for $120 for years until March 2025, when the priceskyrocketed to $250. That brings us to today’s announcement.In an email andblog postto customers today, Plex said that it sold lifetime passes early in its existence “because we knew many of our customers would rather pay a higher, one-time fee for software that they can depend on every day.” However, as evidenced by several price hikes, Plex has struggled to reconcile lifetime passes with its own financial goals. The company’s message says:We’ve considered eliminating the Lifetime Plex Pass in the past, given that recurring subscriptions help us sustain long-term development, but we know it’s still a valuable option for many in our community. So instead of retiring it, we’re keeping it available at a price that reflects the real, ongoing value of the software we’re committed to building and maintaining for years to come.The 200 percent price hike and Plex’s admission that it has mulled killing Lifetime Passes illustrate Plex’s shifting priorities as itseeks profitability. Notably, Plex didn’t announce any pricing hikes for its monthly or annual subscription tiers.In its blog, Plex offered further reasoning for the upcoming price hike:Over the years, as our software and product has evolved, the breadth of features and benefits included with your Plex Pass has expanded. This increase ensures we can continue to invest resources into building and maintaining the Plex personal media software, while continuing to offer a Lifetime option.As noted in the announcement, Plex would rather people pay monthly or annually for its media server features than a one-time fee. With annual subscriptions currently $70, it would take 11 years for a $750 lifetime subscription to be a better value.Plex’s price hike is an extreme example of streaming service providers continuously increasing prices amid struggles to reach and maintain profitability while keeping prices stable and dealing with ongoing costs, like licensing fees for rental movies and app updates.Plex claims to be working on numerous new features for its platform, including: adding all server and library management features currently available on app.plex.tv to Plex’s mobile and, where applicable, TV apps; “boosting dialogue and normalizing loudness,” per today’s announcement; transcoding improvements; IPv6 support; and the abilities to group downloads by show, automatically download new episodes, and make and edit playlists in mobile apps. New features that Plex has added recently include the ability to create custom metadata agents and an open API for server integrations.Beyond new features, Plex has previously cited rising costs for the hike, including supporting various types of devices andcodecs.Plex moving beyond media servingPlex’s various forays into areas outside of its original media server business also contribute to its rising costs. In recent years, Plex has expanded from a media server company to a streaming service provider that hashundredsof free ad-supported TV (FAST) channels,rents out movies, and dabbles insocial features. It also made afailed venture into gaming.Legacy customers, however, have grown wary of Plex’s expanding interests and their potential impact on Plex’s media server users. While adding streaming and social features,Plex killed free remote streaming accessand itsWatch Together featureand issued acontroversial app redesign. Plex also has reason to put more focus on its streaming business since more people have been using Plex’s online streaming service than its media server capabilities since 2022, Scott Hancock, Plex’s then-VP of marketing,saidin 2023.The good news is th

Closing the gap“I’ll buy 10 of those”—NASA science chief yearns for mass-produced satellites“How in the hell do I get more science into space? That is my goal.”Stephen Clark–May 19, 2026 4:43 pm|20Saturn’s moon Enceladus peeks over the limb of Dione during a partial occultation, as seen by NASA’s Cassini spacecraft on September 13, 2008.Credit:

      NASA/JPL-Caltech/Space Science InstituteText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThere are more opportunities to access space than ever, thanks to a bevy of commercial rockets, some with reusable boosters, led by SpaceX’s workhorse Falcon 9. So why is NASA launching fewer telescopes and planetary science missions than it did a quarter-century ago?The answer is complex. It is not necessarily the money. The space agency’s science budget this year is $7.25 billion, roughly the same as it was in 2000, adjusted for inflation. This is despite attempts by the Trump administration to drastically reduce NASA science funding.In the early months of his tenure, NASA Administrator Jared Isaacman’s focus has been on human spaceflight and the Moon. This isn’t terribly surprising given NASA’swildly successful Artemis II missioncarrying four astronauts around the Moon last month. Since taking office in December, Isaacman has announced an overhaul of the Artemis program,canceling a space stationto be built in orbit around the Moon in favor of construction of a base on the lunar surface.On the robotic front, Isaacman is pushing for NASA to launch afirst-of-its-kind nuclear-powered spacecraftin 2028 to deliver a trio of drone rotorcraft to explore Mars. Isaacman has not said as much about concrete changes to NASA’s science program. He has defended the Trump administration’s proposed cuts to NASA’s science budget—as would be expected of him as a Trump political appointee—but the budget proposals come from the White House, not from NASA headquarters.“Mr. Isaacman is very keen on us doing things quicker and for less,” said Nicky Fox, associate administrator for NASA’s science mission directorate. “More shots on goal is one of his favorite phrases. And I think, for us, it’s looking at the right-sized mission for the problem. Not everything has to be $1 billion or more. There are ways you can do fantastic science. His challenge is he wants 10 $100 million missions to be flying.”How to get there?A future with numerous robotic probes spread throughout the Solar System sounds thrilling to space scientists and space enthusiasts, but you can’t get there with flat budgets and billion-dollar missions that take a decade to get off the ground. Many of NASA’s robotic science missions use purpose-built satellites and instruments, usually manufactured by large contractors like Lockheed Martin, Northrop Grumman, university labs, or NASA itself. Unlike SpaceX’s hangars full of reusable rockets, there’s no building with cameras, spectrometers, telescopes, and spacecraft buses—the core chassis of a satellite platform—lying around waiting to launch.“Instead of having a bespoke bus that does absolutely everything, and makes the tea and brings you toast, what can you do with an off-the-shelf bus?” Fox told Ars. “And maybe you have to change a few things. Maybe you fly fewer instruments, but maybe you fly three [spacecraft] together. How do we really pick up the pace? Because it is difficult when you have long gaps between the missions. It’s certainly not what anyone wants to see.”One way to make this future real is with mass-produced, high-power satellites. Small CubeSats, just the size of a suitcase, are great for missions close to home, but they won’t cut it for missions to more distant destinations, such as another planet or a unique orbit far from Earth. NASA is making use of other ways to collect scientific data in space, such as placing instruments on the International Space Station or on commercial communications satellites.But those solutions won’t work if you want to travel to another world. Sometimes it just costs a lot of money to do the near-impossible.“For $100 million, you can’t buy a bus from somewhere and put four instruments on it and send it toflight to Enceladus to look under the icethere,” Fox said. “No, that’s a big, ambitious mission. We want to fly an interstellar-type probe. As theVoyagers are getting older, we want to study interstellar space. These things are hard, and they’re tough, and it will take a lot of effort to do that. We also talked about actually flying a mission to Uranus.”But what about spacecraft flying on more well-trodden paths to the Moon, Mars, Venus, or the asteroid belt? “What can we do with these commercial off-the-shelf buses? I would love to walk in and say, ‘I’ll buy 10 of those,’” Fox said.NASA islooking at “block buys” for the next series of commercial missionsto the Moon. These privately owned landers and orbiters, part of the Commercial Lunar Pa

The “Cage-aissance” continuesSpider-Noirfinal trailer gives us a classic villainIt’s never too late to become a hero.Jennifer Ouellette–May 19, 2026 4:58 pm|17Credit:

      YouTube/Prime VideoCredit:

      YouTube/Prime VideoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navPrime Video has released one last trailer for its upcoming live action series,Spider-Noir,starring Nicolas Cage, and once again it’s been released in two formats: one in black and white (below) and another in color (above), which the showrunners are calling “True Hue.” Seriously, the more footage we see of this series, the more eager we are to find out if the series lives up to its marketing. And the final trailer—which really plays up the deadpan humor and is set to Amy Winehouse’s “Back to Black”—is very promising.Aspreviously reported, Marvel Comics created its “noir” line in 2009, reinterpreting familiar Marvel characters in an alternate universe, usually set during the Great Depression in the US. A version of the Spider-Noir character, voiced by Cage, briefly appeared in the animated masterpiecesSpider-Man: Into the Spider-Verse(2018) andAcross the Spider-Verse(2023). (He is set to reprise that role in the upcomingBeyond the Spider-Verse.)Cage is playing Ben Reilly, a hard-boiled PI with a secret superhero identity, The Spider. Per the official premise: “Spider-Noirtells the story of Ben Reilly, a seasoned, down on his luck private investigator in 1930s New York, who is forced to grapple with his past life, following a deeply personal tragedy, as the city’s one and only superhero.”In addition to Cage’s Ben Reilly/The Spider, the cast includes Lamorne Morris as Reilly’s friend Robbie Robertson, a freelance journalist who clings to optimism in the face of his buddy’s cynicism; Li Jun Li as nightclub singer Cat Hardy, the classic underworld femme fatale (Li based her portrayal on Anna May Wong, Rita Hayworth, and Lauren Bacall); Karen Rodriguez as Reilly’s secretary, Janet; Abraham Popoola as a World War I veteran; Jack Huston as a bodyguard named Flint Marko who becomes (as we see in the new trailer) the classic villain Sandman; Brendan Gleeson as New York mob boss Silvermane, who is being targeted for assassination; Lukas Haas as one of Silvermane’s subordinates; Richard Robichaux as the editor of the Daily Bugle; and Kai Caster.“Have you seen it out there? The city’s a mess,” Robbie tells Reilly, reminding his friend of the hero he used to be. But the disillusioned Reilly knows that’s just not who he is now and he wants to start over—as a PI. But then Cat Hardy walks into his office and hires him to investigate a friend’s disappearance (killing an errant spider with a newspaper, which probably won’t endear her to Spidey).That friend is Jack Huston, who has “turned into some kind of sand monster.” And others also seem to be coming down with mysterious illnesses that worsen until “their sicknesses eat them alive.” Meanwhile, Silvermane has pretty much taken over the city. Will Reilly stick to his resolve and stay out of it? Or will he resurrect his superhero alter ego one last time? I think we all know it’s gonna be the latter. “I was never a hero,” Reilly insists. But it’s not too late for him to become one.Spider-Noirpremieres on May 25, 2026, on MGM+, with all episodes becoming available on Prime Video on May 27, 2026. Viewers can choose to watch in black and white or True Hue—or both, if one wants to compare.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.17 Comments

FBI wish listFBI seeks US-wide access to license plate cameras, wants “data in near real time”FBI will pay vendors to help it track and search for vehicles nationwide.Jon Brodkin–May 19, 2026 5:29 pm|83Flock license plate reader and camera with solar panel in Pleasant Hill, California on April 16, 2026.Credit:

      Getty Images | Smith Collection/GadoFlock license plate reader and camera with solar panel in Pleasant Hill, California on April 16, 2026.Credit:

      Getty Images | Smith Collection/GadoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe Federal Bureau of Investigation announced plans to buy nationwide access to a network of license plate readers, saying it will award contracts to one or more vendors that can offer “near real time” information from cameras across the US. The proposed contract is for the FBI Directorate of Intelligence.“To evaluate and manage threats to personal safety, property, and law enforcement, the FBI requires professional service firms that can provide License Plate Readers (LPRs) for tracking subjects on roads and highways over the US and its territories,” the FBI said in aRequest for Proposals (RFP)published on May 14. The FBI said the winning bidder or bidders “must provide law enforcement and/or commercial license plate reader data provided through the Contractor’s existing platform.” The system must cover 75 percent of locations, the FBI said.The system must offer the ability to search for license plate information “and other descriptive data such as vehicle description information, time/date criteria, and geo-location criteria,” the FBI said. “Additionally, the system must provide search result notifications. The Contractor system must have the ability to access and/or query cameras across the United States and its territories. The Contractor system must be capable of providing this data in near real time.”Contractors have to be able “to share/create maps depicting camera coverage (i.e. heat mapping),” and “provide the FBI the source of information (i.e. red-light cameras, repossession vendors, speed cameras, etc.),” the FBI said. The FBI said it needs to be able to search the database for partial or full plate numbers, plate states, addresses, locations where a plate was scanned, and vehicle makes and models.Flock and Motorola Solutions potential biddersThe RFP divides the proposal into six regions covering the continental US, Hawaii, Alaska, Puerto Rico, and territories such as Guam and the US Virgin Islands. The FBI said it may award contracts to one or two vendors in each region. The deals can be for up to five years, with all deals combined potentially worth $36 million. The FBI said a contractor’s system has to be available to FBI users via a website.Flock and Motorola Solutions are well-positioned to bid on the contract, as 404 Medianoted yesterday. Both companies could win part of the job, as the FBI said it may award contracts to multiple vendors to achieve its desired level of access.Flock’s Automated License Plate Readers (ALPRs) are sold to local police departments. The company boasts of having deals with “over 12,000 public safety customers including cities, towns, counties, and business partners.” Motorola Solutionssells license plate reader camerasthat can be installed on busy roadways or mounted on police cars.License plate reader cameras have raisedconcerns about privacy,data security, anderrors in plate number recognition systemsleading to wrongful arrests. 404 Mediareported last yearthat local police departments performed searches of the Flock license plate reader system for US Immigration and Customs Enforcement (ICE), “giving federal law enforcement side-door access to a tool that it currently does not have a formal contract for.”The FBI already “runs a License Plate Reader program to facilitate LPR information sharing with and between its law enforcement partners,” a Congressional Research Service reportsays. The US agency “maintains a hot list of vehicle data against which law enforcement agencies can compare their LPR data.”The FBI intelligence division’s plan to obtain direct access to an extensive network of cameras could help expand that information sharing. As the FBInotes, its intelligence division shares information with a variety of federal, state, local, and tribal law enforcement agencies.Flock: Sharing is “opt-in” for local policeFlock itself temporarily provided access to Customs and Border Protection, Homeland Security Investigations, the Secret Service, and Naval Criminal Investigative Service as part of a pilot last year. Flock confirmed the pilot to the office of Sen. Ron Wyden (D-Ore.),according to Wyden.Flock saysit has federal customers “including National Parks, Veterans Affairs hospitals, and military bases,” but that it does not work with ICE.Federal attempts to access data could be limited by company pol

🤖曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini🌐马耳他将向全体公民免费提供一年 ChatGPT Plus📱小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务🚀目标估值 1.75 万亿美元，SpaceX 最快 6 月上市📋arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚💻卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元🚗蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌🔍需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位☁️最低 1 元 40 万 Token，三大运营商推出 Token 套餐🚙特斯拉 Robotaxi 碰撞事故报告解封💡知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值🚗小米 YU7 GT 火山灰配色公布💰ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构🧠马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布🤖字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本📱腾讯 vibe coding 产品「吐司」即将上线🛍️蜜雪冰城上架老黄同款蜜桃四季春，销量暴涨🛑中国银行关停信用卡独立 App，服务整合进主 App曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini昨晚，彭博社记者 Mark Gurman 在最新的 Power On 新闻通讯中透露了更多关于苹果即将发布的独立 Siri 应用的细节。隐私设计是此次新应用的核心卖点之一。Gurman 透露，该应用将内置对话自动删除功能，用户可在设置中选择对话保留时长：30 天、1 年或永久保存。苹果预计在今年的 WWDC 继续主打隐私保护，并强调其 AI 体验不含广告。模型方面，苹果已将新 Siri 的 AI 模型底层大量替换为 Google 的 Gemini，并将部分服务运行在 Google 的云基础设施之上。而此前有消息称，苹果和原本的合作对象 OpenAI 关系出现裂痕，甚至有可能对簿公堂，原因是 OpenAI 认为双方合作的结果不及预期。值得注意的是，尽管新的 AI Siri 已经跳票约两年，Gurman 称苹果内部测试版本仍将其标注为「Beta」，并可能沿用至其正式发布。马耳他将向全体公民免费提供一年 ChatGPT PlusOpenAI 日前宣布，与马耳他政府签署合作协议，向该国公民及居民免费提供为期一年的 ChatGPT Plus 订阅服务。据悉，居民须先完成由马耳他大学开发的免费在线课程「AI for All」，课程内容涵盖 AI 的基本概念、能力边界及负责任使用方式，通过后即可向马耳他数字创新局申请免费订阅资格。小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务小米 REDMI 产品经理「笋寸」日前在微博发文建议，近期不打算换手机的用户可以考虑等待换电池，并提示部分机型在更换新电池后容量可能还会有所提升，参考小米 13 系列的现行政策。笋寸进一步说明，小米正积极推动存量用户更换采用金沙江技术的大容量电池，小米 13 系列已率先上线相关服务。她还透露，具体支持机型「得等等」。此前，小米 13 系列已全系上线电池升级服务，支持小米 13、小米 13 Pro、小米 13 Ultra 三款机型升级更大容量电池，费用为 189 元，其中电池升级费 149 元、人工费 40 元。目标估值 1.75 万亿美元，SpaceX 最快 6 月上市据路透社报道，SpaceX 已加快 IPO 时间表，计划最早于 6 月 12 日在纳斯达克挂牌上市，股票代码为「SPCX」，较此前预定的 6 月下旬提前约两周。据三位知情人士透露，公司计划最早于本周三公开招股说明书，路演定于 6 月 4 日启动，股票发行最早于 6 月 11 日完成定价。时间表提前的主要原因是 SEC 对招股材料的审核进展快于预期。此次 IPO 预计募资约 750 亿美元，目标估值约 1.75 万亿美元，有望成为史上规模最大的股票发行。承销行阵容包括摩根士丹利、美国银行、花旗集团、摩根大通和高盛。arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚近日，预印本学术论文平台 arXiv 宣布，若作者提交的论文中包含无可辩驳的 AI 生成内容证据，将被封禁账号一年，并在此后要求所有提交须先经过同行评审期刊接受方可上传。arXiv 计算机科学板块主席 Thomas Dietterich 在 X 发文表示，若论文中出现「幻觉引用」、大语言模型元注释（如「以下是 200 字摘要，是否需要修改？」）等明显迹象，即可认定作者未对 AI 生成内容进行核查。arXiv 目前由康奈尔科技学院管理，今年 7 月将独立运营，成为一家非营利法人机构。康奈尔科技学院院长 Greg Morrisett 表示，独立后 arXiv 将能从更广泛的捐助方获得资金支持，以应对「AI 垃圾内容」带来的运营压力。卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元近日，小米集团合伙人、总裁卢伟冰在「Max 一夏」户外露营直播中表示，玄戒芯片今年肯定会推出迭代款，并称这将是「一块非常强的芯片」，会搭载在一款「非常优秀的产品」上。他同时提醒外界，目前网络上流传的各类相关传闻可信度不高，具体规格暂不便披露。卢伟冰还透露新一代玄戒芯片的应用范围将不再局限于智能手机，后续计划广泛搭载于小米旗下多类智能终端设备。关于手机定价，卢伟冰在直播中作出判断，认为今年下半年尤其到年底，部分国产旗舰直板手机价格可能突破 1 万元人民币。他预计，内存涨价的压力至少延续至 2027 年底，甚至到 2028 年。对于即将发布的小米 17 Max，他表示定价仍在讨论中，将在发布会上正式公布，但小米的原则是在同档位提供「最物超所值的产品」。蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌据睿见 Economy 报道，在日前举行的第十八届轩辕汽车蓝皮书论坛上，蔚来创始人、董事长兼 CEO 李斌线上出席，在对话中将萤火虫比作 iPhone，称其已成为「一款非常高品质的产品」。他强调，萤火虫在某种意义上超越了汽车的工具属性，演变为「一个生活方式品牌、一个时尚品牌」，并开始形成独特的品牌文化。他表示，萤火虫在用户群体上打破了年龄与购车预算的双重限制，「大学刚毕业的年轻人想买也能买，很有钱的老板想自己开车送小孩上学也完全没有问题」。需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位据《商业内幕》报道，前向部署工程师（FDE）的招聘需求在一年内激增逾 700%，正成为 AI 浪潮下科技行业增长最快的职位之一。招聘平台 Indeed 数据显示，今年 4 月该职位招聘帖数量达 5330 条，较去年同期的 643 条增长约 729%。Anthropic、OpenAI、Palantir、Stripe 及 Google Cloud 均在积极招募该职位人才。FDE 的核心职责是协助企业将 AI 整合进业务流程，并根据客户需求定制开发 AI 工具。薪酬方面，Indeed 数据显示该职位年薪约为 17 万至逾 20 万美元。最低 1 元 40 万 Token，三大运营商推出 Token 套餐中国电信昨日推出系列试商用 Token 套餐，分别面向开发者及中小微企业、个人及家庭用户提供「Token+连接+安全」一体化服务，均提供三档套餐，面向个人及家庭用户定价 9.9 元/月起 1000 万 Token；面向开发者及中小微企业 39.9 元/月，1500 万 Token。两类用户均可选购宽带上行提速包（上行速率提升 50 Mbps，100 元/月）及智能体安全管家服务（5 元/月）。个人家庭用户还可额外选购 5G-A 优享包（含 10 GB 流量，上行速率最高 300 Mbps，20 元/月）。联通、移动两大运营商也先后推出了各自的 Token 套餐：移动：中国移动上海公司于 5 月 17 日世界电信日发布会上推出 Token 通用服务，定价 1 元 40 万 Token。联通：中国联通上海分公司于 5 月 16 日发布全栈国产化「国芯国模国云」安全基座，推出多档 Token 产品及融合套餐。特斯拉 Robotaxi 碰撞事故报告解封近期，特斯拉正式解密并公开了向美国国家公路交通安全管理局（NHTSA）提交的 17 起「Robotaxi」自动驾驶碰撞事故的完整未脱敏报告。据 Electrek 报道，这 17 起事故发生于 2025 年 7 月至 2026 年 3 月，均涉及在得克萨斯州奥斯汀进行「Robotaxi」测试的 2026 款 Model Y，且测试时均有安全员在场。伤亡数据显示事故整体较轻微：13 起仅造成财产损失，2 起无人员受伤，1 起轻伤未入院，另有 1 起轻伤需入院治疗（系车辆在路口以 2 英里/小时蠕行时被后方 SUV 追尾）。官方报告表明，大部分事故并非特斯拉自动驾驶系统（ADS）的过失，多为车辆在红绿灯、停止标志或交通拥堵中完全静止时，被后方人类驾驶员追尾或侧擦。但数据同时也暴露出系统与后台干预的缺陷：两起事故发生在远程操作员接管期间，导致车辆分别以 8 英里和 9 英里的时速撞上金属围栏与施工路障；另有数起事故显示 ADS 存在空间感知死角，车辆在转弯或倒车时撞上了金属链条、拖车挂钩、木制电线杆及路缘石。💡 知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值据「听筒Tech」报道，知乎创始人兼 CEO 周源日前在第十二届新知青年大会上表示，AI 时代真实创作者的价值不可替代。AI 每分钟都在创造奇观，但 AI 时代仍然需要新知。而新知来自真实的人。每一个真实的人，都拥有不可替代的价值。大会主论坛上，多位嘉宾就 AI 与人类创作的关系展开讨论。复旦大学人文学者梁永安提出「科技有所为，如何让自己有所不为？需要人的尺度和人文属性」；作家张春称「不完美就是我们的完美」。周源在大会结尾总结道：「技术越发达，有温度、有深度的真人创作就越宝贵。」小米 YU7 GT 火山灰配色公布近期，小米汽车再度放出 YU7 GT 的预热图片，公布了新配色以及新车的内饰信息。其中小米 YU7 系列将加推一款名为「火山灰」的配色，官方介绍：灵感来自晨雾下的火山地貌，高纯度灰调，叠加纳米级矿物颗粒，呈现细腻而有层次的金属质感。同时新车内饰也迎来全新的「赛道红」：采用红黑撞色，提供专属刺绣。官方表示，小米 YU7 也会提供「火山灰」外饰配色，并且现在车厘子红的 YU7 GT 已经在部分小米汽车门店展示了。ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构近日，OpenAI 面向美国 ChatGPT Pro 用户推出了个人财务管理功能，支持通过 Plaid 连接超过 1.2 万家金融机构的银行账户和信用卡。连接账户后，ChatGPT 会自动读取用户的真实余额、消费流水和负债数据，生成一张覆盖投资、日常开销、订阅和近期账单的财务面板。用户提问时，模型可直接翻出账单历史，找出哪些类目容易超支，并给出具体的削减方案。此外，用户还可以告诉 ChatGPT 一些账外信息，比如「我在攒钱买车」或「我还欠父母一笔钱」，系统会将这些内容保存为财务记忆，用于后续对话。马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布马斯克昨日在 X 发帖披露，SpaceXAI 正在研发的 1.5T 参数 Grok 基础模型第 9 版（V9）已完成训练，预计 3 到 4 周内正式发布。现役公开版本 Grok 4.3 基于 0.5T 参数的 V8 基础模型，目前每隔数日迭代一次。V9 将参数量提升至 1.5T，马斯克称其为「重大升级」。在帖子的结尾，马斯克更是用用「banger」（王炸）一词来形容对这一版本的预期。字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本字节跳动 Seed 团队近日开源扩散

直到 2026 年五月，我们终于等到一整轮的影像旗舰登场完毕——小米做出了徕卡一瞬、vivo 打造了照片与视频兼具的 V 单、OPPO 在双长焦方案上带来了惊喜。在这轮旗舰更新潮的最后，华为提前将新一轮的 Pura 90 Pro Max 端上了桌。老实说，第一眼看硬件，心里是有落差的。去年 Pura 80 Ultra 上颇为惊艳的「一镜双目」结构不见了，取而代之的，是目前行业内最稳妥也最常见的大底长焦方案。不过，在细细体验后，我认为 Pura 90 Pro Max 藏了一个比纯拼硬件更有意思的新解法——XMAGE 智拍。全自动，都省心过去这几年，我系统性地梳理过华米 OV 的影像架构。虽然各家在硬件选择上各有理解，但大方向出奇一致，都在极其有限的机身空间里死磕传感器面积，比拼谁的光圈更大、进光量更多。在硬件上奋发图强理所应当。但在摄影中，不错的硬件基础仅仅是第一步。想要拍出一张好照片，构图与后期同样重要——构图保证一张照片在视觉上是否平衡、主体是否明确；而后期则决定了这张照片视觉质感和风格化。但这两者，向来都是具备门槛的手艺。而华为在 Pura 90 Pro Max 上，打算一口气解决这两个问题。熟悉他们产品线的朋友或许有印象，去年发布的 Pura 80 Ultra 上，华为试水了 AI 辅助构图。AI 辅助构图功能相对基础，在实际拍摄中主要起一个引导和提醒的作用。到了 Pura 90 Pro Max，这项技术完成了从「辅助」到「主导」的蜕变，并定名为「XMAGE 智拍」。剥开表层去探究 XMAGE 智拍的底层逻辑，是一套完全由端侧算力驱动的本地化影像流水线，将拍摄过程拆解为三个同时进行的模块——基于实时语义分割的自动构图、无视物理镜头限制的智能变焦，以及针对画面的色彩重塑。也就是说，在这个功能的加持下，Pura 90 Pro Max 可以自动识别主体、变焦构图，最后还排列出一系列适合当前场景的 XMAGE 风格供你选择。值得一提的是，XMAGE 智拍不需要网络连接，也无需将庞大的图像数据传回云端，全部基于端侧 AI 运转，实时识别画面里的主体轮廓、建筑线条以及光源分布状况，随后根据底层训练好的审美模型，给出一套包含最优构图、变焦和精细色彩方案的组合拳。繁复的专业参数在这里黑盒化，过去需要人脑去判断的构图和调色工作，全权交给了本地算法。除此之外，Pura 90 Pro Max 还顺势推出了 AI 姿势推荐，这个功能同样通过语义分割识别，可以理解画面中的了恩物、环境、姿势与背景，随后直接在取景框中勾勒出一个等比例的线框轮廓。拿手机的人只需按图索骥，引导模特贴合画面上的姿势即可。遇到不满意的动作能随时刷新，它甚至支持导入社交网络上保存的样片，一键提取动作精髓。听起来确实很厉害，但这套完全依赖端侧计算跑出来的流水线，究竟能不能应对现实世界里复杂多变的光线，还需要把它带到真正的街头去寻找答案。AI 有所能，也有所不能来到一家咖啡馆，墙上的多层光环灯饰引起了我的注意。举起手机，XMAGE 智拍自动变焦到 2×，将墙壁上的光环作为画面主体。这个选择可以说中规中矩，光环圆形处于画面中下方，整体构图稳定和谐。同时，操作台上颇具金属质感的咖啡机也是个不错的主体，XMAGE 智拍精准识别到我的拍摄意图，变焦到 192mm，保持咖啡机的中置。由于我想保留更为还原的室内光线与色彩，这张照片并没有套用 XMAGE 智拍推荐给我的任何色彩风格。用它拍了几组照片后，我认为 XMAGE 智拍是一个非常用户视角的功能——在传统的拍摄习惯里，手机焦段是一件需要精打细算的事。大脑会本能地去贴合物理镜头的原生焦段，生怕落在中间焦段会损失画质。但 XMAGE 智拍完全抛弃了执念。它不在乎用的是几倍变焦，转而完全为画面考量：当前这个画面，怎么裁切才最好看。哪怕不可避免地要损失一部分边缘像素，只要能让视觉中心突出、比例和谐，它就会果断动手。这种做法很实在，一切都在为了最终的出片服务。走出室外回望咖啡馆，绿植与橙色外墙交相呼应，XMAGE 智拍自动截取出色彩冲击力最强的部分，同时准确识别出咖啡馆建筑主体，两种大色块在画幅中占比相似，平衡而美观。再抬头，夏季树木繁茂，XMAGE 轻微放大焦段，让树枝形成天然的框架构图，广州塔正悬于画面正中间。在这组照片的拍摄中，我还察觉到了一个很有意思的技术细节。按下快门时，你会发现滤镜加载的速度，远大于构图改变的速度。这是因为滤镜往往只需识别画面后套用底层的色彩映射，而构图则需要端侧算力去实时识别画面里的建筑线条、光源分布和人物轮廓。计算量完全不在一个量级。更有意思的是，智拍套用滤镜的方式非常精细，除了 XMAGE 风格选择外，还会对调色盘做精细调整，确保风格对当前环境的适配度。当然，这套系统也有自己的脾气。在实际走街串巷时，XMAGE 智拍有时候会自动跳出来接管画面，有时候又毫无动静，需要你手动去点击唤醒。就算是手动唤醒的情况下，XMAGE 智拍也有几率出现风格已经选好，但当前画面无法找到最佳构图的情况。这种情况频繁出现在极其繁杂的环境中，由于元素过于混乱、信息量爆炸，算法的确很难从中剥离出合适的构图。当然，换个角度想想，也许此时的 1× 视角，就是当前环境的最佳构图了。万物皆可 AI，现在轮到摄影了185 年前，达盖尔发明银版摄影法，留住时间成为一种特权；2011 年，胶片巨头柯达传出破产消息，智能手机开始野蛮生长，把镜头塞进了普通人的口袋。回顾整部影像史，其实就是一部打破特权、技术走向大众的普惠史。到了今天，移动影像的狂飙突进，撞上了一堵无形的墙。手机镜头模组越做越大，与之对应的，是最近行业里一个颇具深意的传闻：受制于高昂成本与物理极限，下一代影像旗舰的「超大杯」，大多要面临停更了。这条路为什么走不通了？因为影像旗舰突飞猛进的背后，藏着一个被刻意回避的死结。决定照片质量的，永远是镜头后面的脑袋。普通人有懂美的眼睛，却跨不过光圈快门和构图比例的门槛。把一台堆满顶级硬件的手机，递给一个毫无基础的人，他按下的快门，大概率得到一张平平无奇的随手拍。顶级硬件如果只服务于一小撮懂摄影的人，无疑就会变成一种伪命题。在同行集体踩下刹车的时间点，华为提前把 Pura 90 Pro Max 端上了桌，而 XMAGE 智拍，就是破局的方法。在此之前，行业里已有不少 AI 落地，比如 vivo 让四季流转，OPPO 消除反光。只不过，它们大多是按下快门后的后期创意和修补。XMAGE 智拍往前跨了一大步，让算法成为整个拍摄动作的主导。这种做法谈不上是什么艺术层面的颠覆，但确实管用。降低门槛，让完全没有摄影基础的人，也能相对轻松地把一台旗舰机用出该有的样子。好的技术绝不该是一座孤岛。曾经的智能手机让人们拥有了「随时拍」的自由，现在的 AI 则进一步赋予了大众「拍得好」的能力。这或许是打破当前僵局的一剂良药，也是影像超大杯物尽其用的方法。

过去几年，20 万级新能源 SUV 市场像一条突然变宽的河道。一边是家庭用户对空间、续航和舒适性的刚性需求，另一边是智能化、性能化配置不断下放。过去很多只会出现在 30 万元甚至更高价位车型上的配置，如今开始像雨水一样往下渗透。激光雷达、空气悬架、大模型智驾、零重力座椅，这些原本带有「高配」意味的词，正在成为新一轮价格带竞争里的筹码。肩负着广汽集团高端化重任的昊铂品牌在今年广汽集团重振旗鼓的趋势下，也顺势重新梳理产品架构，昊铂 A800 旗舰轿车前段时间开始上市，用到了乾崑智驾和鸿蒙座舱来增强产品竞争力。紧接着，面向 20 万价位带的 SUV 昊铂 S600 也来了，提供纯电 660 Max、纯电 660 Max+、增程 230 Max、增程 215 四驱 Ultra+ 四个版本，官方预售价为 20.99 万至 22.99 万元，预售权益价为 18.89 万至 20.89 万元。新车定位为「新豪华智慧运动 SUV」，提供纯电和增程两种动力形式，车身尺寸为 5015×1933×1700mm，轴距 2936mm。相比于品牌初创时期的雄心勃勃，昊铂现在显然务实不少，它试图把昊铂 S600 做成一个配置密度很高的样本：性能、智驾、座舱、舒适、安全和权益，几乎都被塞进「满配」这个关键词里。昊铂 S600 请了 2 个代言人，一个代言「优雅」，一个代言「运动」过去提到中大型 SUV，很多人的第一反应是「稳」：空间要大，坐姿要高，后排要舒服，最好还能承担家庭出行里那些琐碎但真实的场景。比如周五晚上从公司直接出发，后备厢里塞进折叠椅、露营桌、孩子的滑板车和一箱零食；又或者长途返乡时，后排老人希望能把座椅放倒一点，孩子希望看动画，副驾则希望在高速上补一觉。对于这类用户来说，车首先不是一个「驾驶机器」，而是移动客厅、临时休息室，也是家庭生活的第二个容器。昊铂 S600 没有放弃这部分的诉求，或者说，这些配置已经在国产中高端 SUV 上成为一种基本素养，昊铂 S600 拥有 90% 得房率、750mm 二排腿部空间、5.316 平方米整车玻璃面积，以及 1292L 二排拓展空间；舒适配置上，同级唯一 18 点按摩零重力女王副驾、143° 躺角后排电调座椅、22 扬 ADiGO SOUND 音响、五块双层夹胶静音玻璃等配置也都被纳入其中。但昊铂 S600 又没有把自己完全包装成一台「沙发车」。它的另一面，是更强调姿态和操控的运动属性，所以昊铂又反复出现「跑山王」这个表达，显然昊铂希望把一台五米多长的大尺寸 SUV 做得不那么笨重，让它不只是适合城市通勤和家庭旅行，也能在山路、弯道、高速超车等场景中保留一定的驾驶参与感。从造型上看，昊铂 S600 采用的是一种前低后高的轿跑溜背姿态，超过 5 米的车长和近 3 米轴距提供了视觉体量，双层鸭尾、宽体车尾、20 英寸运动轮毂等设计元素，则试图削弱中大型 SUV 常见的厚重感，从侧面看它确实不像一般的家用大型 SUV 那样笨重方正，更像是专为年轻家庭所设计。所以这一次昊铂也请了 2 个代言人，英格兰传奇足球明星迈克尔·欧文作为「运动代言人」，知名演员陈数作为「优雅代言人」，两位代言人分别代表这一款车的一种特质，也是试图同时吸引家用取向和个人运动取向的潜在消费者。把四驱、高阶辅助驾驶、双腔空悬都放进 20 万级别在最高配的增程 215 四驱 Ultra+ 版本上，昊铂 S600 同时提供双电机四驱和双腔空气悬架。昊铂表示，这是行业首次在这一价位段同时放入双电机四驱和双腔空气悬架。另外，昊铂 S600 还全系标配前双叉臂后五连杆独立悬架、SDC 可变阻尼减振器和鹰爪系统 2.0，并提供雪地、湿滑、脱困三种路面辅助模式。一台大 SUV 在城市高架上并线，需要车身跟随方向盘动作，而不是像大船一样慢半拍；周末跑山时，连续弯道考验的是侧倾控制和车身收敛；雨天压过积水路面，轮胎抓地力和车身稳定系统要在驾驶者还没反应过来之前，先把车身姿态拉回来。在这些场景下，双腔空悬能够有效提升驾驶体验和驾驶信心。动力部分，昊铂 S600 提供纯电和增程双动力。其中增程四驱版综合功率达到 370kW，满电零百加速 4.3 秒，亏电零百加速 5.6 秒。广汽星源增程技术的持续发电功率为 85kW，官方还提到新车在亏电状态下可实现 165km/h 速度持续刷圈 2 小时，亏电极速 193km/h。增程车过去最容易被质疑的地方，在于「满电一条龙，亏电一条虫」，电量充足时像电车，电量不足后则容易出现动力衰减、油耗升高、发动机噪音变大等问题。昊铂 S600 对 85kW 持续发电功率和亏电性能的强调，其实是在回应这个痛点，一台增程 SUV 不能只在发布会数据里快，也要在长途、高速、爬坡和亏电场景下保持基本的体面。在智能化层面，昊铂 S600 的策略也比较直接：尽量把高阶配置做成全系标配。新车全系标配激光雷达、高通 SA8650 智驾芯片、27 个高精度感知单元，并搭载融合 Momenta R6 强化学习大模型的广汽星灵智行 ADiGO GSD 3.0，支持城区和高速快速路场景。资料还提到，昊铂 S600 下半年将首批升级 Momenta R7 强化学习世界模型。智能座舱方面，新车搭载高通 8295P 座舱芯片、17.3 英寸 3K 中控屏、27 英寸 W-HUD，并内置 11 个 AI 智能体。相比过去车机只负责导航、音乐和语音控制，现在的座舱正在变得更像一个移动终端：它不仅要响应指令，还要理解场景，甚至提前知道用户在车里可能需要什么。下班后坐进车里，导航自动给出回家路线，HUD 把车道级信息投到视线前方，孩子坐在后排看动画，车机根据车内温度和空气质量调整新风，座椅按摩开始工作……这些在新能源汽车上可以实现的体验，对于从 20 万元档位传统燃油车换购过来的用户来说，无疑是跃升式的。广汽集团有着庞大的用户基本盘，这些人群想要增购换购新能源车型的话，广汽自然希望用产品承接他们的需求。就像 AI 说的那样，我在这里，不偏不绕，不躲不逃，稳稳接住你。某种程度上来说，过去几年广汽集团新能源产品的建设，过于重视 B 端，对 C 端用户的吸引力不够，并没有「稳稳接住」用户的需求。对消费者来说，20 万级新能源 SUV 已经足够拥挤，智己 LS6、极氪 7X、小鹏 G7 等车型都在这个区间形成压力。昊铂 S600 要突围，关键命题并不是「它是不是拥有很多配置」，而是这些配置能否被组织成一种清晰的用车体验被感知到：城市里安静从容，山路上不显笨重，长途时不焦虑，家人坐进去愿意放松，驾驶者开起来也不觉得无聊。

龙虾之父一个月消耗 6030 亿 Token，总花费金额高达九百万人民币。移动联通电信，三大运营商都在推 Token 套餐，199 送千兆宽带还有 1 亿 Token，了解一下？从硅谷到国内大厂，Tokenmaxxing 成为公司的主流，谁消耗 Token 多，谁就是 AI 时代的好员工。00 后校友向母校捐赠 20 亿 Token，被网友调侃按 DeepSeek 5 元/亿 Token 计算，只要 100 元。▲图片来自新浪财经Token 在半年内完成了一次身份跃迁：从技术术语，到 KPI，到话费套餐，到捐赠货币。它成了 AI 时代的「度量衡」，唯一的问题是，没人说得清它到底在度量什么。我们自己买 Token，用公司的 Token，部署了一堆 Agent，代码、论文、周报都是 Token 烧出来的。而另一边是，大厂的员工由于 Token 消耗排行榜的原因，开始拿着公司的 Token 处理私事、玩游戏、开发数十个没什么用的子 Agent 来提升自己的排名。「回报」这件事很难量化，但「使用量」可以量化。于是所有人都选择了那个容易量化的东西。这不是 AI 时代的新问题，这是管理学的老病。用 AI 消灭狗屁工作的公司，正在制造新型狗屁工作亚马逊，那个裁员裁到大动脉，把自己的网站都变成 404 的小狗，最近又被爆出了新的「笑料」。原本被寄予厚望、用来消灭「狗屁工作」的 AI，最终却沦为制造新型「狗屁工作」的源泉。据《金融时报》报道，为了逼迫员工拥抱 AI，亚马逊搞出了一个极其复古的管理手段：「Token 消耗排行榜」，追踪每个员工的用量。公司强制要求超过 80% 的开发者每周必须完成 AI 使用指标，甚至将消耗 Token 的数量作为考核标准。▲图片

1.OpenAI 的两大宿敌 Anthropic 和马斯克，放下心中成见之后终于在月初结盟了。在此之前，Anthropic 和马斯克的关系并不融洽：今年 2 月，马斯克还在自己的 X 账号指责 A 社「woke」「邪恶」「反人类」（misanthropic），说这家公司「仇视文明」。事后来看，这次攻击并非马斯克清新脱俗的性格使然，而是 Anthropic 所做的某些事情触碰到他的神经，事出有因。在此之前，xAI 内部使用 Cursor 工作，但是今年年初员工发现，Claude 模型突然在 xAI 的 Cursor 公司账号里不能使用了。当时还在 xAI 上班的联合创始人吴宇怀，在全员信里是这么说的：「Anthropic 更新了政策，要求 Cursor 不得向其主要竞争对手提供 Claude 模型调用能力。」当时，吴宇怀在信中写了一句话，颇为有趣：「这是坏消息也是好消息。我们的生产力会被影响，但这也敦促我们开发自己的编码产品和模型。」为什么当时 xAI 的高层认为，开发自己的编码产品是关键？后来发生的事情，大家都知道了。xAI 的联创团队悉数跑路，马斯克一气之下对 Cursor 使用了钞能力必杀：上个月底，SpaceX 和 Cursor 共同宣布，将在编程和知识类工作 AI 模型的训练上，展开前所未有的战略合作；并且，SpaceX 还获得了以 600 亿美元收购 Cursor 的权利，或向后者支付 100 亿美元合作费用。注意编程这个关键定语，后面还会 call back.2.最近，我看了一条 Cursor 早期投资人、Anthropic 大喷子、T3 创始人 Theo Browne 的视频。本来点进去是看他喷 A 社和 SpaceX 怎么蝇营狗苟，结果没想到，却看到了关于 SpaceX + Cursor 合作的，一个既另类却又极度合理的分析：不说 600 亿的收购，就只说 100 亿的合作费——Theo 在视频里表示，自己认为「哪怕只是交换到 Cursor 的用户数据，这 100 亿也值回票价了。」所以是什么数据？如果你也去看 Theo 这条视频，他会讲得非常清楚。但为了节约时间，我们在这里简单概括一下：我们和 AI 的对话是一来一回的，你提出问题/需求，他给你解答；coding agent 同理，只不过返回的是代码。一次高质量的对话，整个过程，包括用户提示、模型思考、agent 规划、输出代码、验证——所有这些东西合起来，可以称为一个完整的 Agentic Loop——就成为了高价值的训练数据，再喂给模型去进行强化学习，就能进一步提高模型在实战场景下的表现水准。Cursor 有的，SpaceX 想要的，就是这些数据。可这些数据从哪里来呢？答案很简单：作为模型厂商，这种高质量数据的最直接来源，只能是你自己开发的 coding agent 产品——也即 Anthropic 的 Claude Code、OpenAI 的 Codex、Kimi 的 Kimi Code。现在你应该明白了，为什么被 Anthropic「封号」之后，吴宇怀会在全员信里提出开发 xAI 自己的 coding 产品和模型这件事了。这件事 xAI 在当时已经看清楚了：没有自己的编码产品，就没有高质量的强化学习数据；没有高质量的数据，就训练不出真正实战能力强的 coding 模型。虽然有点暴论，但现在我们可以点题了：模型厂商想做出来真正能打的编程模型，做自己的 coding agent 产品是唯一的路径。3.大语言模型像个水晶球，用全网的语料训练出来，似乎能够解答万物，但并不代表它在所有问题上都能给出高质量的答案。用 GitHub 上数以亿计的代码条目训练，当然也能训练出 coding 模型。这是「学习结果」的逻辑，也是没问题的。毕竟编码任务的结果是可以验证的：代码能不能运行，测试能否通过，结果摆在那里。但是，通往结果的过程，是一个涉及多步骤决策、错误纠正、意图对齐的复杂链条。每一次用户的接受、拒绝、补全、撤销、追问、甚至当模型好几次都搞不定或者完全搞错时的辱骂——都是这一链条上的过程信号。强化学习有两种监督方式，一种叫做结果监督，只看最后是否跑通。但是结果监督会催生「奖励黑客」的现象：模型为了能跑通可能写出冗余、脆弱、带逻辑漏洞的代码，但因为测试过了，模型以为自己学对了。而另一种叫做过程监督，对推理路径上的每一步进行打分。上述这些过程信号，只有在 coding agent 运行环境里才能诞生。GitHub 仓库里只有结果，哪怕是去看单独的提交历史，看 PR，都找不到有效的过程信号。在缺乏有效、自主可获得的过程信号的时候，一些模型厂商会采用「蒸馏」的方式，这个事情大家应该已经知道了。蒸馏的逻辑很简单，给同样的输入，老师模型输出什么，学生模型就学着输出什么。但是通过蒸馏，即便可以获取到思维链，得到的仍然更接近于结果，而非被蒸馏的老师模型内部的概率分布。一旦学生在推理中偏离了老师的轨迹，哪怕一个 token 不符合，都有可能发生偏离。这背后是强化学习的基础限制：策略梯度定理要求，优化样本最好由当前正在优化的模型自己去产生。这种数据叫做 on-policy 数据。而通过蒸馏别家模型，在别人的产品里产生的数据，来训练自己模型，都属于 off-policy 数据。模型当然可以从中学到东西，但学不到老师模型内部的概率分布信息。而像 Cursor 这样自己就是 coding agent 产品的公司，掌握着最真实、有效、高质量的训练数据。Cursor 产品本身，就是 coding 模型在实战环境中的最佳训练场。我们可以通过 Cursor 年初的「翻车」，来证明这个逻辑。4.APPSO 读者应该记得，年初 Cursor 发布了 Composer 2，号称「下一代专用编程模型」，技术报道写的相对保守，也没有提供具体的模型底座信息。结果很快，网友就在公开代码片段里发现了 Kimi 的模型 ID，截图传遍了开发者社群，逼得 Cursor 副总裁 Lee Robinson 出面澄清：「Composer 2 确实是从开源底座出发的。最终模型大约只有 1/4 的算力来自底座，剩下 3/4 是我们自己训出来的。」几小时后，Cursor 联创 Aman Sanger 也跟着发了一条道歉：「一开始没提 Kimi 底座是个失误。」五天后，Cursor 放出了完整的 Composer 2 技术报告，显示底座的确是 Kimi K2.5，授权方则是 Firworks AI，大致流程是在 K2.5 上做训练，再继续做大规模强化学习（RL）。但关键之处在于，Composer 2 的 RL 是运行在真实的 Cursor 会话当中，使用与生产部署完全相同的工具和 harness。Cursor 将这套流程叫做「实时强化学习」(real-time RL)，也即将模型的 checkpoint 直接部署到 Cursor 生产环境中，观察用户的响应，收集数据，聚合成奖励信号——最快可以每 5 个小时迭代一次模型版本，然后继续部署到 Cursor 里，循环往复。最极致的案例是 Cursor 的自动化代码补全功能 Tab，每天处理超过 4 亿次请求，每当用户输入字符、移动光标时，模型都会预测下一步动作，如果预测置信度高，则显示建议，用户按下 tab 即接受自动补全。该功能采用的是在线强化学习，在行业内极具特色。Cursor 可以以极高的频率（最快可达每一个半小时到两小时）更新 Tab 的模型能力给用户，直接在产品内收集 on-policy 数据进行训练。这种高频、接近实时的反馈回路，让 Tab 可以学习到极其微妙的用户意图。Cursor 方面透露，这种方法让 Tab 建议的拒绝率降低 21%，接受率提高了 28%。回到 Composer 模型本身。在事情搞清楚了之后，一些 Kimi 员工也删掉了之前吐槽的的推文，Kimi 官方账号发表了祝贺。一家估值 600 亿美元（基于马斯克给的数字），不做自己的模型基座的 coding agent 应用层公司，仍然可以通过产品自身的数据飞轮，RL 出超越基座模型的专有编程模型。所以与其说 Cursor 翻了车，不如说这反而是 coding agent 产品重要性的绝佳例证。Cursor 在另一篇关于实时 RL 的文章里写到：「（训练编程模型）最大的困难在于建模用户。Composer 的生产环境里不只有执行命令的计算机，还有监督和指导它的人。模拟计算机容易，模拟使用它的人却很难。」这句话，现正在逐渐成为了在编程模型方面走在前沿的模型厂商之间的共识。如果你去看 benchmark 榜单和用户普遍评价，会发现哪些头部的厂商都在发力做自己的 coding agent/编程产品。区别只在于谁离用户更近。我们以 SWE-bench、LLM-Stats 等相对权威的榜单为例，Claude、GPT、Gemini、Kimi 等模型基本霸榜前十，清一色都是有自己开发 coding agent 产品（包括 CLI、IDE、集成 coding agent 的桌面客户端）的模型厂商。在部分榜单上会出现少数反例，如 Meta (Muse Spark)、DeepSeek 等，没有开发自己的 coding agent。不过你会发现，这些反例模型，在更加接近真实场景、避免污染的更权威 benchmark 上就很难上榜了。以 DeepSeek 为例，它在 SWE-bench bash only 上分数是 70%，排名第九，在 SWE-bench Pro 上分数却掉到了 15% 左右。OpenRouter 的真实流量数据可以解释这种反差：该平台 2025 年报告显示，Claude token 消费 80% 以上用于编程和技术任务，而 DeepSeek token 消费主要集中于闲聊和角色扮演。没有自家 coding 产品的厂商，在一些 coding 任务 benchmark 上能挤进头部，但在更难的真实工程 benchmark 上，在用户用 token 消费投票的真实流量中，都会原形毕露。不仅是 Cursor，Anthropic 在 2025 年 11 月发的一篇论文里，也明确透露自己在做一模一样的事情：「我们在 Anthropic 自家的真实生产编程环境上做训练。」也即 Anthropic 把自己员工使用 Claude Code 的交互数据，反哺给 Claude 模型用来训练。5.在 AI 的演进历程中，生产要素的定义发生了深刻的位移。传统三大核心要素——算力、研究、训练数据，虽然在总量上持续增长，但在结构上已经出现了严重的失衡。今天的各大 AI 巨头显著提高了在算力上的资本支出 (CapEx)，让算力基建成为了当前舆论的主旋律。但实际上，特别是在编程范畴内，随着 GitHub 仓库、StackOverflow 等互联网公开代码数据被基模厂商「竭泽而渔」式地利用，模型在代码生成与逻辑推理上的边界开始逐渐显现。这也是为什么，行业共识正在逐渐转向一个冉冉升起的新战略高地：对于任何希望掌握顶级代码能力的模型厂商而言，建立自有的 coding agent 产品早已不再是可选的商业路线，而是确保底层模型可以持续进化的核心生命线。正如前面 APPSO 论证的那样，单纯学习公开数据等于只学习成功者的结局，却无法了解成功的路径，这绝对不是正确的成功学应该有的样子。在真实的编程环境中，知道发生了什么错误、怎样发生的、如何正确地理解和高效地实践需求等等——了解正确过程的价值，远超于得到正确结果本身。只有拥有自己的编码产品，模型厂商才能获取高质量的「过程监督」信号，从而在编码/推理能力的下一阶段竞争中，确保自己仍有技术护城河——否则就不得不像 SpaceXAI 那样，花钱去跟 coding agent 产品公司去合作。然而并不是所有模型厂商都跟马斯克一样有钱，以及 2026 年开始的巨头势力划分、结盟与领地的争斗会变得更加激烈，当一家缺乏自主 coding 产品的模型厂商终于回过味来的时候，恐怕已经没有足够的合作伙伴可以挑选，合作

长城今晚公布了魏牌 V9X 的上市价格：34.98 万元到 38.98 万元。这台大六座 SUV 的最终定价，比预售价下调了两万多。配合 5000 元抵 1 万元的现金膨胀礼，再加上本品牌 1 万元、其他品牌 8000 元的增换购补贴，实际的起售价已经被压到了 33 万这个位置。伴随着新车的亮相，长城汽车董事长魏建军走上了舞台。他接过了话筒，抛出了长城在造车底层架构上的思考：眼下行业里最偏好的往往是新能源车专属架构，魏建军偏偏逆风而行，将归元 S 做成了兼容燃油、插混和氢燃料的全动力平台。实际上，魏建军看上的是更广阔的海外版图。他表示，在很多国家，电力基础设施的建设依然迟缓，若用全动力平台去打全球市场，则能更好地把造车规模做大，将高昂的研发成本摊薄，从而稳住车辆的市场保有量和保值率，也让用户的修车费用降下来。对于新能源专属平台，魏建军在台上表达了自己的看法。实际上可以这么讲，我们近十年就是被这个电动专属平台忽悠了。现在，由于动力的多样化，配置的多样化，我们理论上可以卖到全世界去。造车的底子铺好之后，魏建军把话题拉回了最基础的用车体验。他反复强调，无论车企给出的承诺多么好听，最根本的准则永远是不能给用户添麻烦。为什么要添一句「不要给用户添麻烦」？你想想，当你在路上被撂下、车出了问题的时候，你的心情有多么焦虑。顺着老板定下的规矩，长城的工程团队把干货都掏了出来。魏牌 V9X 是一台车长近 5.3 米、轴距突破 3.1 米的全尺寸 SUV。为了拉动它，长城给它准备了一套超级 Hi4 混动系统和 2.0T 发动机，并搭配了一台 4 挡专用变速箱。在低挡位运转时，这套动力总成能爆发出 4500 牛·米的轮端扭矩，足以应对陡坡和极端烂路。长城还给出了明确的性能指标，声称不管车辆是满电还是亏电状态，都能跑出 4 秒级的百公里加速成绩；进入高速巡航状态，车辆又能把馈电油耗稳稳控制在百公里 6.3 升左右。底盘配置同样堆满了料。长城给 V9X 全系标配了闭式双腔空气悬架和 CDC 减震器，以及正负 10° 的后轮转向系统。这套后轮转向系统顺带接入了自动泊车功能，面对狭窄的侧方车位，后轮会自动微调角度配合入库，让泊车过程更加流畅。至于后轮转向对座舱空间的侵占，长城的设计团队顺着外扩的轮包，把第三排座椅做成了环绕式的大沙发造型，化解了结构上的劣势。在 V9X 的座舱里，这样的巧思还有很多，例如，副驾位置的手机无线充电板没有放在常规的中控台下方，而是被挪到了车门内扶手上，还带有一点微倾的角度。乘客的手自然搭在门板上，就能顺手拿到手机。智能化算力拉到了目前的顶格水平。英伟达 Thor 芯片负责辅助驾驶，骁龙 8295P 支撑座舱运转。虽然用的是别家的芯片，但长城在软硬件的协同上选择了全栈自研，把动力、底盘和辅助驾驶的控制权全部掌握在自己手里。魏建军表示，V9X 的各个系统之间能够实现毫秒级的响应，遇到突发危险情况，车辆可以提早判断，做到刹车的同时，也能进行绕行避障。他还强调，为了保护隐私，所有的语音指令和行程数据都做到了本地闭环，绝不外传。在加入了大模型后，车机里的「小魏同学」也变得更加智能了。发布会现场展示了一段测试，视频里的魏建军特意换了件外套，戴上墨镜坐进主驾。即便如此，系统还是在第一时间认出了他，并自动调出了他习惯的座椅位置和空调温度。发布会的后半程，舞台交给了魏牌 CEO 赵荣波，他把魏建军定下的准则，变成了一项项具体的购车权益。比如高阶智驾软件对用户终身免费开放，锁单后如果超过 30 天没交车，长城每天给用户补贴 200 元。其中最受关注的，是那份 6 年 15 万公里的整车质保。按照汽车行业的常态，普通质保通常只有三四年，部分品牌在遇到空气悬架、后轮转向这类高成本的底盘件时，还喜欢用易损件的名义另行算账。且这类零件的故障高发期，刚好卡在车子开到第四到第六年的时候，一旦脱保，维修和更换的费用动辄几万元。长城这次把空气悬架和后轮转向系统也放进了长周期的保修范畴里。赵荣波在台上强调，相关的服务条款没有任何文字游戏，也不设隐藏门槛。整场发布会的落脚点，最终还是回到了魏建军身上。这位长城的掌舵人拒绝了常规的营销套路，选择亲自上阵，直接拿自己的个人信誉和姓氏为这台新车兜底。面对网友关于长城会不会「翻车」的犀利提问，他没有回避，给出的回应干脆利落。我认为长城价值观和对用户负责任的这种精神，永远翻不了车。大家选择长城，我就要肩负起这份责任，让每个选择魏的人放心，这才是真正的代言。我的代言才刚刚开始。魏，你放心。

今年 3 月，苹果制造了 MacBook Neo 这个大爆款，把高质量入门笔记本的价格打到了 4599 元人民币起。一个半月后，这台搭载 A18 Pro 处理器的入门款 Mac，全年销量从 700 万台追加到了令人难以置信的 1000 万台，交付周期一度排到了 5 个月之后……苹果紧急向鸿海、广达等供应商追加订单。知名供应链分析师郭明錤给出更刺眼的数字：在全球 Windows 笔记本电脑预计同比衰退超 10% 的大背景下，2026 年 MacBook 系列出货量有望达 2500 万台，同比增长 20%-25%。这一拳，Windows 阵营的最核心代表英特尔，该如何接住？在今天，英特尔在中国正式推出第三代酷睿处理器，代号 Wildcat Lake「野猫湖」，并同时启动了联合产业链伙伴的「Firefly 萤火虫计划」。英特尔的接招，围绕一件事而展开：将最先进的 Intel 18A 制程，下放给低端，或是入门，或是「主流」——随便你怎么称呼它的价位段。野猫盯上新的猎物Wildcat Lake 这个代号，在去年底就已经在业界出现。今年 CES 期间，英特尔高调发布了主打高端的产品线：酷睿 Ultra Series 3，采用 Panther Lake 架构。同一天，Wildcat Lake 也悄悄披露，你可以将它理解为 Panther Lake 的精简版本。业界普遍将其视为英特尔为了对标苹果 MacBook Neo 而推出的产品，专攻平价入门级 PC 市场。「直接对标」这四个字，并不是说说而已。在 4 月底，英特尔为野猫湖推出的公版参考设计曝光，采用渐变配色、铝合金一体成型机身、窄边框键盘，扬声器开孔环绕在键盘两侧、D 面无出风口——这种设计语言不需要过多解释，它就是在用苹果惯常的设计语言，去回应苹果本身最受欢迎的多彩配色与铝合金轻薄机身策略。在 MacBook Neo 上，苹果做对了一件事：它让整个行业重新意识到，在「内存末日」(RAMmageddon) 的今天，入门价位段成了一座金矿。用一颗本来进入销库存阶段的 A18 Pro 芯片，配上 8GB 内存和 256GB 存储，开价 4599 元——MacBook Neo 活脱脱成了 MacBook Air 平替。英特尔，以及整个 Windows PC 阵营不能继续将这个用户群体拱手相让。而这，正是「野猫湖」存在的（几乎）全部理由。18A 首次下放到入门段第三代英特尔酷睿处理器的技术参数并不复杂，CPU 采用 2 个 Cougar Cove 性能核，加上 4 个 Darkmont 超低功耗能效核，总共 6 核 6 线程；GPU 部分配备 2 个 Xe3 核心，NPU 算力最高可达 18TOPS。整机基础 TDP 15W，最高 35W，在无主动散热的笔记本上，甚至可以压低到 11W 运行。按照英特尔今天提供的官方口径，这颗芯片功耗相比前代降低超过 60%，图形 AI 性能提升了将近 3 倍，平台总算力更是达到了 40TOPS。当然，考虑到「野猫湖」处理器的 NPU 算力只有 17TOPS，远低于 Windows 阵营的所谓「AI PC」，也即微软官方定义的 Copilot+ PC 40TOPS 的阈值。所以严格来说，野猫湖并不能直接和苹果 A18 Pro 高质量的神经网络引擎，以及统一内存架构，去做横向比较。话虽这么说，驱动参数量不大的本地小模型，做一些 AI 推理任务，还是绰绰有余的。以及更有意思的是，谷歌前段时间发布自家的 AI 笔记本电脑「Googlebook」的时候，也确认了和英特尔结盟——Wildcat Lake 被海外业界人士普遍认为将会成为是 Googlebook 的核心平台。连谷歌都觉得够用，应该没大问题吧？真正关键之处不在于这些参数，而是在于「野猫湖」用上了 Intel 18A 制程工艺本身。要知道，过去两代酷睿 Series 1 和 2，一直沿用 Raptor Lake-U 架构和 Intel 7 制程，实质上是 10nm SuperFin 工艺的演进版本。在主流芯片普遍走向 5nm 甚至 3nm 的当下，技术代际差距已经相当明显。当 MacBook Neo 用虽然是老库存但货真价实 3nm 工艺的 A18 Pro，杀进四千多元的价位段，英特尔过去那套「靠老制程压低成本」的入门策略已经过时了。所以，我们很欣慰地看到，英特尔将 18A——2nm 级别的工艺，压到一颗入门芯片上。这意味着英特尔充分愿意承担初期更高的单位成本，去力保一座不能再丢弃的城池。系统：手机供应链思路搬到 PC 上对于守住市场，「Firefly 萤火虫计划」同样重要。英特尔副总裁、中国区软件工程和客户端产品事业部总经理高嵩如是说：芯片本身的算力固然重要，但「芯片和系统层面的双重创新」才是这一代主流 PC 体验跃迁的来源。中国区客户端与平台销售业务部总经理宗晔则认为，PC 也正在从工具进化为「智能体验入口」。根据英特尔提供的内部数据，与上一代相比，「野猫湖」产品的 PCB 面积缩小了 5%，器件数量减少了 7%。同时针对第三代酷睿平台特性制定并推广标准化主板与 I/O 板连接，推动主板核心电路进一步实现模块化、小型化、低功耗化。这一套打法的精神底色，其实来自于手机。手机行业过去 10 年磨合出的供应链协同方式——元器件高度集成、模块化主板、统一接口标准，被英特尔搬到了入门级 PC 上。结果，就是更轻、更薄，也更便宜。参与「Firefly 萤火虫计划」的合作伙伴，包括但不限于华硕、七彩虹、荣耀、惠普、联想等。英特尔预计，全球合作伙伴将陆续推出超过 70 款基于 Wildcat Lake 平台的产品。目前，荣耀 MagicBook X14 2026 已经率先上架，搭载该平台 Core 5 320 处理器，采用 16GB+512GB 存储配置，定价下探到了 4399 元。用户画像：完全对标 MacBook Neo，但还有更多英特尔列举了 Wildcat Lake 平台产品的 5 类目标用户：辅导孩子作业的家长、全天在校的学生、刚入职场的新人、「一人公司」创业者，以及银发族。这些用户群体和场景描述本身没什么悬念，但值得注意的是，这个用户画像，和 MacBook Neo 在线下门店主要推荐的人群，几乎是完全重叠。MacBook Neo 夺走的本来就是入门级 Windows 笔记本电脑的客群，英特尔的接招，就是造出具有更好价值主张的入门级笔记本电脑，再把心智重新抢回来。英特尔还反复强调野猫湖平台的本地 AI 推理算力。这对于那些购机预算本就不多，也不太可能付费订阅优质云端模型算力的用户来说，倒也是个正确的思路。最后，野猫湖看起来像是合格的入门芯片平台，跟 Firefly 萤火虫计划结合起来，也形成了一套颇具诚意的组合拳。轻、薄、便宜的 AI 笔记本，英特尔现在重新回到了这个擂台。仅剩的悬念，是 OEM 最终能把零售价做到多少。在 MacBook Neo 的价位上，苹果给到了系统生态、统一内存架构、多种潮流配色等等诸多看头。但英特尔要想把这只野猫放出去，跟苹果大杀四方，光有 18A 工艺和优秀的工业设计是不足够的。这场仗的真正交战点，大概率不会在技术本身，而是在即将到来的 618 购物节期间，第一批搭载它的笔记本产品，能否真正把供货紧张的 MacBook Neo 杀个七零八落、措手不及。

⚖️马斯克诉 OpenAI 案败诉📱苹果开启 WWDC 邀请，将公布苹果 AI 新进展🚗问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处🤖国家反诈中心 App 上线「AI 内容鉴定」⌚苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动🚙广汽董事长回应被称网约车之王🧠三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能🎵腾讯音乐官宣收购喜马拉雅💰铠侠股价飙升，计划美国上市🚀马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元☁️百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半🎬爱奇艺一季度由盈转亏 2.95 亿元🦿荣耀机器人「闪电」「元气仔」走进南大💡前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去🚗小米 YU7 GT 定档周四发布🎮5499 元起，红魔 11S Pro 系列发布马斯克诉 OpenAI 案败诉据 CNBC 和《华尔街日报》报道，美国加利福尼亚州奥克兰联邦陪审团已就埃隆 · 马斯克诉 OpenAI 及其 CEO 山姆 · 奥特曼一案作出裁决，认定马斯克的起诉超出诉讼时效，驳回其全部主张。此案源于马斯克 2024 年提起的诉讼。他指控 OpenAI 及奥特曼违背了将该公司作为非营利机构运营的承诺，并在庭审中称 OpenAI「窃取了一家慈善机构」。陪审团的裁决并未就「违反慈善信托」指控的实质内容作出判断，而是认定马斯克的起诉超过了三年诉讼时效。法院同时驳回了针对微软的相关指控。裁决宣布后，马斯克在 X 平台发文，将这一结果定性为「日程上的一个技术性问题」，并表示将向第九巡回上诉法院提起上诉。OpenAI 首席律师威廉 · 萨维特在庭后表示，这「不是一个技术性裁决，而是一个实质性裁决」，并称马斯克是在将诉讼作为竞争手段使用。微软方面亦发表声明，对裁决表示欢迎。苹果开启 WWDC 邀请，将公布苹果 AI 新进展今天凌晨，苹果正式启动 WWDC26 邀请，并公布详细日程安排。本届 WWDC 大会将于北京时间 6 月 9 日至 13 日举行，面向全球开发者开放，重点展示 Apple 各平台的新系统能力、开发工具和框架更新。按照安排，WWDC26 将于北京时间 6 月 9 日开幕，届时 Apple 将举行主题演讲和 Platforms State of the Union，介绍新一代软件平台、人工智能相关进展，以及面向开发者的新工具。大会期间，开发者和学生可以观看超过 100 场围绕工具、技术和设计的视频讲座，参与团体实验室活动，并在 Apple 开发者论坛中与工程师和社区成员交流。Apple 还将在 6 月 9 日邀请超过 1000 名开发者、设计师和学生前往 Apple Park，参与线下特别活动。届时，爱范儿也将在 Apple Park，为大家带来第一手的现场报道。问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处全新一代问界 M9 系列汽车设计师 @Zimo-子墨 日前在微博发文，点赞全新理想 L9 的内饰设计，称其「简洁、逻辑、比例和谐，视觉重心清晰，口味控制得恰到好处」。他表示，初创公司在有强决策者的情况下，往往对设计管控介入较深，容易打散设计的贯穿逻辑；而到第二代、第三代产品时，无论前代成功与否，团队通常会更多放权给设计师打磨表达。他同时透露，新 L9 的内饰设计总监为来自宝马的 Chris Lee，并称虽然双方长期管理直接竞品，但「真的欢喜好设计，生在中国」，并已发短信向对方表示祝贺。国家反诈中心 App 上线「AI 内容鉴定」据央视新闻报道，国家反诈中心 App 近期正式上线「AI 内容鉴定」功能，支持对图像、视频、音频及文本进行 AI 生成痕迹检测，主要用于识别诈骗及电商场景中的伪造内容。电商平台上利用 AI 篡改图片骗取「仅退款」的行为日趋普遍。央视新闻记者随机选取十张商家提供的疑似 AI 生成售后图，上传至国家反诈中心 App 及多个 AI 鉴真工具后，均被判定为「图像含 AI 生成痕迹」。据介绍，该检测方案结合小模型与大模型：小模型学习 AIGC 内容的生成模式，大模型则判断内容是否存在物理瑕疵或逻辑语义问题，二者联合输出最终结果。苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动据 DIGITIMES 报道，苹果计划于今年下半年对高端 Apple Watch 机型进行重大改款，传感器功能将同步升级，市场预计届时出货量较 2025 年增长 20% 至 30%。Apple Watch 11 已具备血氧、脉搏、心率及血压感测能力，下一阶段预计推进无创血糖监测，具体落地时间取决于 FDA 审批进展。作为苹果传感器元件的独家供应商，亚洲半导体（TASC）预计将直接受益。市场预期新品订单最早于 7 月启动，带动下半年业绩明显超越上半年。广汽董事长回应被称网约车之王据「睿见 Economy」报道，日前广汽集团党委书记、董事长冯兴亚在「第十八届轩辕汽车蓝皮书论坛」上，针对「网约车之王」的说法，其指出：很多消费者可能没有想明白一个问题，「网约车是生产资料」。生产资料对品质的要求极高，因为它是用来挣钱的工具。只有能为自己带来更多利益的车，才能受到使用者和购买者的欢迎。对司机来说，车要「多快好省」，要能赚钱。「这恰恰是对品质最好的验证。没有高品质，就不可能受到网约车驾驶者的欢迎，这是最基本的。」他进一步分析，网约车的乘坐周期非常长，一天可能十几个小时都在车上。如果舒适性不够、疲劳感很强，它也不会受到司机的欢迎。此外，电耗等各项指标，也决定了网约车司机是对车要求最严格的一群人。冯兴亚总结道，这些从网约车运营中积累下来的经验，广汽会延展到旗下各个品牌，让车辆变得更加可靠、舒适和安心。「这反而是我们的优势所在。」三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能据 DIGITIMES 报道，三星电子近期大幅提升了用于高带宽内存（HBM）的最新 DRAM 良率。其第五代 10nm 级 1b DRAM 良率已达 92%，用于下一代 HBM4 产品的第六代 1c DRAM 良率也突破 75%。良率的提升有望增强三星的成本竞争力，并提升其向主要 AI 客户大规模供应先进存储产品的能力。在产能扩张方面，据 Ddaily 报道，三星和海力士正同步提前推进工厂扩建和设备订单。三星正加速平泽 P4 工厂的 DRAM 投资，设备安装预计持续至今年底，市场估计明年将进一步扩大投入。海力士则在清州 M15X 和龙仁 Y1 工厂推进扩产，业界预计其明年 DRAM 相关资本支出有望较今年翻倍。腾讯音乐官宣收购喜马拉雅腾讯音乐娱乐集团（TME）昨日宣布，已正式完成对喜马拉雅并购交易的交割，喜马拉雅成为 TME 全资子公司。交割完成当日，TME 执行董事长彭迦信与 CEO 梁柱联合向全体员工发布全员信。管理层称，喜马拉雅是中国长音频领域的重要建设者，双方战略方向高度一致，合并后将共同拓展在线音乐与音频行业空间。铠侠股价飙升，计划美国上市据彭博社报道，铠侠（Kioxia）股价昨日上涨 16%，创下新高。公司在上周披露强于预期的业绩展望后，多家华尔街机构迅速上调目标价；与此同时，铠侠也确认正准备在美国发行存托凭证，借此提升流动性并扩大国际资本市场接触面。报道提到，铠侠预计截至今年 6 月的季度营业利润将达到 1.3 万亿日元，高于市场平均预期，季度利润也创下纪录。多家机构认为，在 AI 带动的存储供需紧张环境下，公司仍有进一步提升自由现金流和股东回报的空间。马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元据彭博社报道，马斯克昨日在萨姆森国际智能移动峰会线上视频连线时表示，正在德克萨斯州推进 SpaceX 上市准备，称「SpaceX 的 IPO 事宜将很快开始」。知情人士透露，SpaceX 最快本周提交 IPO 申请文件，计划融资最高 750 亿美元，目标估值超 2 万亿美元，若成功将成为史上规模最大的 IPO。百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半百度昨日发布 2026 财年第一财季业绩，期内实现营业总收入 320.75 亿元，同比下降 1.16%；归母净利润 34.45 亿元，同比下降 55.36%；经营现金流为 27 亿元。具体来看，百度核心 AI 新业务收入达到 136 亿元，同比增长 49%，首次占百度一般性业务收入比重超过一半；其中智能云基础设施收入 88 亿元，同比增长 79%，GPU 云收入同比增长 184%。萝卜快跑在一季度的全无人自动驾驶运营订单达到 320 万单，总订单量同比增长超过 120%，截至今年 5 月其全球业务已覆盖 27 个城市，累计自动驾驶里程超过 3.3 亿公里。爱奇艺一季度由盈转亏 2.95 亿元爱奇艺昨日发布 2026 财年第一财季业绩，期内实现营业总收入 62.26 亿元，同比下降 13%；归母净利润由盈转亏 2.95 亿元，去年同期为盈利 1.82 亿元；经营现金流为 1.86 亿元，同比下降 45.13%。分业务看：会员服务收入为 42 亿元，同比下降 5%；在线广告服务收入 12.4 亿元，同比下降 7%；内容发行收入 3.59 亿元，同比下降 43%；其他收入 4.27 亿元，同比下降 49%。公司在财报中提到，热门剧集仍推动会员收入环比增长，海外业务保持稳健增长，并计划借助AI降低内容制作成本。荣耀机器人「闪电」「元气仔」走进南大5 月 17 日，荣耀携自研机器人「闪电」与「元气仔」在南京大学开展具身智能科普活动，并同步展示荣耀 600 系列新品的核心影像功能。活动现场，荣耀展示了自研机器人的核心运动与交互能力。搭载高动态运动系统、支持自主感知导航与高速避障的机器人「闪电」与在校学生进行了 8 轮 1V5 短跑测试；另一款机器人「元气仔」则完成了包含「太空步」在内的复杂舞蹈动作演示与人机互动。同期，荣耀在现场设立了新品体验区，重点展示荣耀 600 系列智能手机。官方信息显示，该机型主打基于 Magic OS 系统的 AI 能力及 4K Live Photo 动态照片功能。外观方面，荣耀 600 系列目前已公布「幸运星」「青苹果」「光羽蓝」和「曜石黑」四款配色，其中主打配色「幸运星」采用了行业首创的双重磁吸工艺。💡 前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去前 Tesla 无人驾驶与 SpaceX 猛禽发动机团队工程师 Zac Valles 在 YC Demo Day 结束 72 小时后直飞深圳，驻扎 8 周后携新硬件返回旧金山。他随后在 X 上发布了一篇长帖，向所有有意从事硬件创业的人发出忠告：不论有没有融资、团队甚至创业想法，都应该立刻去深圳。只要你对创办一家硬件公司有哪怕一点点想法，就去深圳。融资前、组队前、有想法前、离职前，都不是借口。去就对了。在切入供应链的路径上，Valles 建议创业者围绕一场大型行业展会安排行程，将其作为接触工厂的跳板。他还建议在华强北预留 4 到 6 小时实地走访，称其为一座「微型城市」，机器人子组件、电池、底盘、电子元器件均有专属楼栋聚集。小米 YU7 GT 定档周四发布昨日，小米正式宣布，将于 5 月 21 日（本周四）举行人车家全生态新品发布会，届时将发布小米 YU7 GT、小米 17 Max 等新品：小米 YU7 GT：定位跑车级 SUV，最大马力可达 1003ps，最高时速 300km/h，CLTC 纯电续航 705km；小米 17 Max：采用 6.9 英寸极窄四等边直屏，搭载徕卡 2 亿像素主摄和 8000 mAh 金沙江电池；小米手环 10 Pro：最薄 9.7 mm，最轻 21.6 g，续航最长可达 21 天；小米首款耳夹式耳机：单机重 5.5 g，支持 LHDC 5.0。5499 元起，红魔 11S Pro 系列发布红魔昨日正式发布红魔 11S Pro 系列，国内首发第五代骁龙 8 至尊领先

你帮我找一辆白色硬棚 Z4 M-Power，蓝色双轮毂，找那个富家子出来，我要找他试车。在 2012 年上映的港片《车手》里，余文乐饰演的警察陈翔留下了这样一句台词。不久前他在中环经历了一场惨败——那辆白色宝马 E86 Z4 M Coupe 在逼仄的车流和弯道中游刃有余，将他远远甩开。值得一提的是，现实中 E86 本就稀少，而 Z4 M 更是凤毛麟角，尤其是手动挡版本。这也难怪导演郑保瑞会用它来为整部影片定调，那种桀骜不驯的气质，恰如其分。当年大银幕上的稀罕物定格成了经典，而在现实的流水线上，属于 Z4 的齿轮也停止了转动。前一阵，奥地利格拉茨的麦格纳工厂内，最后一辆底盘代号为 G29 的宝马 Z4 缓缓驶出生产线，这款问世于 2002 年、历经三代演进的经典敞篷跑车正式宣告停产。伴随它的离去，宝马延续近三十年的双座 Roadster 车系，正式画上了休止符。巴伐利亚的驾驶图腾Z 系列的起点要追溯到三十多年前。在德语里，字母「Z」代表「Zukunft」，也就是未来的意思。90 年代的宝马对「未来」的理解非常具象，还带着一丝日耳曼式的偏执：剥离一切与驾驶无关的多余配置，把一台直列六缸发动机塞进一个小巧的底盘里，只保留两个座位和纯粹的驾驶乐趣。▲ BMW Z11989 年，采用下沉式隐藏车门的 Z1 惊艳亮相。这台车在今天看来像个古怪的科技实验品——它的车门可以通过电机垂直降入门槛内，车身覆盖件采用了可以由车主自行拆卸更换的特殊热塑性塑料。▲ BMW Z1除了这些前瞻设计，Z1 最重要的历史意义在于，它用全球限量 8000 台的桀骜姿态，确立了整个 Z 系列不按常理出牌、游走在主流边缘的先锋基调。六年后的 1995 年，Z3 走向了更广阔的大众市场。伴随着皮尔斯·布鲁斯南在 007 电影《黄金眼》中的潇洒亮相，这台远赴美国斯巴达堡工厂生产的 Roadster 迅速风靡全球。▲ 007：黄金眼中的 Z3但真正的硬核故事发生在幕后。当时宝马的工程师伯克哈德·戈舍尔（Burkhard Göschel）对敞篷跑车疲软的车身刚性极度不满。于是，他带领一群狂热的工程师突发奇想，利用周末的闲暇时间躲在工厂里，给一辆 Z3 焊上了车顶——这就是后来被车迷们戏称为「小丑鞋」的 Z3 M Coupe。▲ Z3 M Coupe虽然它造型怪异、空间局促，但封闭车顶使其抗扭刚性比敞篷版提升了 2.6 倍。在当时，它是纽博格林北环赛道上最让人胆寒，也最难驾驭的机器之一。前《Top Gear》主持人杰里米·克拉克森曾毫不掩饰地表达过他对这种老派双座跑车的偏爱：当一台车拥有那么长的车头，而你几乎坐在后轴上开车时，它就不再是一个交通工具，而是一台纯粹为了让人发笑而制造的机器。到了 1999 年，宝马 Z8 将这种暴力美学与复古浪漫推向了顶峰。▲ BMW Z8由亨里克·菲斯克操刀设计的 Z8，完美致敬了 50 年代的传奇跑车宝马 507。修长优雅的全铝车身骨架下，藏着与同时期 E39 M5 同款的 4.9 升 S62 V8 引擎。它顺理成章地拿到了 007 电影《黑日危机》的剧本，成为詹姆斯·邦德的新座驾，也让全世界记住了属于 Roadster 的优雅与从容。进入 21 世纪，Z4 正式接过了 Z 系列的接力棒。▲ BMW Z4第一代 Z4（E85/E86）诞生于克里斯·班戈主导设计的年代。那套著名的「火焰表面」设计语言，在当时引起了轩然大波。无论是前脸修长的敞篷版，还是车顶极具肌肉感的硬顶 Coupe 版本，侧面那道巨大的「Z」字形折线都带着一股咄咄逼人的凌厉感。好的汽车设计应该能够激发强烈的情感反应。你不需要向别人解释它为什么美，当人们看到它时，心跳就会加速。前宝马集团设计副总裁克里斯·班戈的这套理论，在第一代 Z4 身上体现得淋漓尽致。尤其是前文提到的 Z4 M。▲ Z4 Coupe宝马工程师把当年 E46 M3 上那台被奉为自然吸气神作的 S54 3.2L 高转速直列六缸发动机，塞进了 Z4 狭小的机舱里。它就像一匹没有完全驯化的野马，脾气暴躁，容错率极低，却又让人在征服它之后欲罢不能。2009 年，第二代 Z4（E89）问世。当时整个汽车行业刮起了一阵折叠硬顶风潮。宝马也顺应市场，为 Z4 换上了复杂的机械硬顶机构。▲ 外饰设计师 Juliane Blasi两位女性设计师，Juliane Blasi 和 Nadya Arnaout，赋予了 E89 极其优美的车身比例和圆润的线条，不过，硬顶机构不可避免地增加了车身重量，打破了原有的配重平衡。这台车在动态响应上做出了一定妥协，收敛了上一代的暴躁性格，演变成了一台更适合在海滨公路迎风巡航的 GT 跑车。▲ BMW Z4（E89）当时间来到 2018 年，第三代也是最后一代 Z4（G29）面临着一个严峻的市场环境。跑车市场不断萎缩，独自研发一款全新底盘的成本已经高到令人咋舌。因此，宝马选择拉上丰田，基于 CLAR 平台共同分摊研发成本。为了找回丢失的运动基因，G29 毅然抛弃了笨重的机械硬顶，回归了传统的软顶敞篷设计，不仅成功减重，更大幅降低了车身重心，找回了完美的 50:50 前后轴配重比。▲BMW Z4（G29）顶配车型搭载的 B58 3.0T 直列六缸发动机，提供了 340 马力的强劲输出与浑厚的声浪。更令人感动的是，在它的生命周期末段，宝马还出人意料地为直六车型推出了手动变速箱版本。工程师为其重新设计了换挡连杆和机械结构，只为呈现最纯粹的机械咬合手感。这种逆潮流而动的做法，被很多车迷视作一种浪漫的告别仪式。这台车从不妥协于日常琐碎，它存在的唯一理由，就是在踩下油门的那一刻，为你制造纯粹的多巴胺。理想与现实粤语里有一句老话，叫「有情饮水饱」。但在真金白银的汽车行业，情怀从来填不饱肚子。翻开宝马集团的财务报表，Z4 的谢幕早有预兆。G29 在 2019 年上市时还算风光，全年卖了 15827 辆。但这也是它最后的体面。2025 年，Z4 在全球只卖出了 9744 辆，这个数字放在宝马年销两百多万辆的大盘里，存在感微乎其微，占比连 0.4% 都不到。更大的问题在于，养这条产线还很烧钱。G29 并不在宝马自家工厂生产，而是交给奥地利格拉茨的麦格纳·斯太尔代工。一年卖不到一万辆，宝马却得单独为它保供一整套零部件体系。当年宝马联合丰田把 G29 推向市场，本就是一笔算计好的生意——分摊成本，换 Z4 多活几年。但合作研发终究只能解燃眉之急，无法逆转大时代的洪流。实际情况就是，卖掉一台 Z4 的利润，还比不上一台走量的 X3。前宝马集团研发董事傅乐希曾无奈地描述过跑车面临的困境。如今的汽车市场，开发一款跑车比以往任何时候都要困难。你需要高销量来分摊平台研发成本，而跑车市场的容量恰恰在不断萎缩。宝马现在一门心思扑在新世代纯电平台上。这是个烧钱的无底洞，几百亿欧元砸进去都不一定听得见响。新旧交替的关头，钱必须花在刀刃上，花在能决定明天还能不能活的地方。正因如此，砍掉投入产出比极低的 Z4，将宝贵的研发资源转移到更有前景的新能源项目上，是企业求生的本能。▲ 新世代 M3放眼整个行业，双门敞篷跑车的消亡已经成为一种趋势。奔驰 SLC 早就退出了历史舞台，奥迪 TT 也吹响了熄灯号。同级别的竞争对手一个个退场，Z4 能坚持到 2026 年才落幕，已经算得上长寿。宝马集团前 CEO 齐普策说过：「内燃机时代的落幕会让人感伤，但我们必须为未来的移动出行寻找新的答案。」接力棒已经交给了纯电车型，但我们依然会怀念内燃机，怀念沉重的电池组无法替代的那种轻盈灵动。一台 G29 Z4 的重量在 1.5 吨左右，而一台纯电跑车的重量往往会超过两吨。物理定律是客观存在的，增加的重量没有办法完全通过主动悬架来弥补。双座小跑车的魅力很大程度上建立在人与车的沟通感上，座椅和方向盘转向传递的路感，排气管的回火声，甚至是换挡时那一瞬间的顿挫，都在不断刺激着驾驶员的感官。没有了齿轮咬合时的震动，没有了直列六缸发动机逼近红线时的嘶吼，缺少了这些声浪与震动的刺激，速度在屏幕上只是一串苍白的数字。当时代浪潮决绝地退去，留在沙滩上的，往往是普通车迷的执念与阵痛。以后想要体验阳光、微风与汽油味混合的纯粹快乐，大概只能去二手车市场里寻找曾经的碎片了。而淘旧账的代价，往往比想象中疼得多。我的朋友张亦就花 10 万元买了一台 E85 Z4。买回来才发现，老跑车的维修称得上是个无底洞，配件是又贵又难等，单是内外三个后视镜片就花掉他 5500 元。买车 10 万，修车修了 5 个月又砸进去 10 万，到现在还没完全整利索。随着 Z4 的停产，维持这种老派乐趣的门槛只会越来越高。那个为了纯粹驾驶乐趣不计成本造车的黄金年代，连同它引以为傲的直列六缸声浪，就此封存在了过去。

2013 年，埃斯特庄园古董车展。为了纪念宝马摩托车诞生 90 周年，宝马带来了一台名为 Concept 90 的概念车。这个名字大家可能不太熟悉，但其实它的量产车型就是后来大获成功的现代复古标杆——R nineT，人送外号「宝马拿铁」。▲ R nineT六年后，同样是在科莫湖畔的这个庄园里，Concept R18 概念车登场，它顺理成章地开启了宝马 1800cc 水平对置巡航车 R 18 家族的时代。埃斯特庄园似乎成了宝马重磅复古车型的首发圣地。到了 2026 年的今天，宝马带着全新的 Vision K18 概念车再次回到这里。这辆概念车的造型区别于现有的 K1600，甚至和我们认知里的休旅车型都不太一样，更像是一个超音速飞行器。为了这个造型，宝马互换了进气和油箱的位置，并将后部线条尽可能压平，营造出一种低伏、前倾的姿态。哪怕它静静地停在聚光灯下，依然带有一种贴地飞行的视觉张力。长达两米的铝制侧板是这台车最引人注目的地方，呈现出一种无缝、一体的连续性。大面积的手工铝材与深色的锻造碳纤维搭配在一起，摒弃了多余的装饰，让整台车散发出浓烈的重工业味儿。数字「6」作为这台车的图腾，被具象化地刻在车头和车尾——前方是六颗带有主动散热功能的 LED 头灯，车尾的碳纤维包裹着六根粗壮的排气管。这套不锈钢排气系统采用了火焰喷涂工艺并覆有液态铝，质感类似于 F1 赛车的排气管。宝马用「The Heat of Speed」来概括这台车的设计主线。过去，那些为长途休旅准备的动力平台总是被包裹在整流罩内，安静且内敛。现在，宝马选择让这台排量达到 1800cc 的直列六缸发动机和排气完全裸露。车身结构向发动机和进排气让位，复杂的机械部件成为最抢眼的视觉元素。BMW Motorrad 首席执行官 Markus Flasch 说：通过 Vision K18，我们展示了对性能、豪华与情感的新理解。直列六缸发动机对我们而言远不止是一项技术，它是一种宣言，Vision K18 将这种力量转化为一种雕塑般的存在，让速度与动感在静止状态下依然清晰可感。从概念到现实尽管 Vision K18 在设计上追求的是速度与动感，但它始终是一辆为巡航而生的车型，它最终的量产车型很有可能就是 K1600 的继任者。或许，它会叫 K1800？▲ BMW K1600GT不管最终如何命名，这台新车最受瞩目的依然是那台排量提升后的六缸发动机。放眼目前的全球摩托车市场，愿意在两轮车架里塞进六个气缸的厂商屈指可数。在这之中，水平对置是最主流的选择。本田金翼那台 1833cc 的水平对置六缸就是行业内的典型代表，国内的长城灵魂用的也是水平对置八缸。宝马则选择了一条独特的技术路线——直列六缸。将六个气缸排成一条直线，会严重挤压摩托车的横向空间。为了控制车身宽度，宝马在 K1600 上就采用了非常紧凑的缸间距设计，气缸套之间的距离仅有几毫米。与此同时，它的缸体也向前倾斜了约 55°，既降低了整车重心，也为上方的进气系统和车架腾出了位置。这种结构带来了截然不同的特性。本田金翼偏向低转速和高扭矩输出，相比之下，宝马的直列六缸则带有一些运动轿车的特质：转速攀升更快，高转速下的声浪更具穿透力，动力输出也更加绵密。若是在此基础上扩缸到 1800cc，低转速时的扭矩想必也会有所提升。动力的升级让 Vision K18 做好了驶向公路的准备，但在实际骑行中，长途巡航除了需要充沛的动力，还要能够应对迎面而来的风和雨。这套动力平台真正进入量产、接替现款车型时，我们在概念车上看到的科幻外观大概率会向实用性让步。摩旅时，天气总是不似预期，大尺寸电动风挡、小腿前的整流罩、大容量边箱与尾箱，以及雷达、电控等系统，都会重新出现在量产车上。Vision K18 只是提前展示了这台机器剥去外壳后的原生状态，并且，它未来的量产形态可能不止于传统休旅车。参考宝马在 R 18 车系上的产品布局，依托同一台 1800cc 水平对置双缸发动机，他们衍生出了复古裸车、Bagger 巡航以及豪华休旅等多个版本。▲ BMW R 18全新的直列六缸平台也可能会复制这样的思路。它既可以加上整流罩，作为 K1600 GT/GTL 的正统换代车型进入休旅市场；也能像 Vision K18 展示的这样，推出一款褪去整流罩的复古版本，正好也能填补宝马 Heritage 继承者系列的产品线——R nineT 依托风冷水平对置双缸奠定了基础，R 18 切入了美式巡航市场，如果增加一台搭载 1800cc 直列六缸的休旅车，整个产品线会变得更加完整。

今日凌晨，苹果正式发布了今年 WWDC 全球开发者大会的邀请函。WWDC 26 将于北京时间6 月 9 至 13 日举行。今年的 WWDC，除了会正式公布「27」系列系统之外，开发侧重点仍将集中在 Apple Intelligence 后续进展、体验优化等等。自然，其中也少不了一些苹果针对 AI 时代的软件生态的提前布局。图｜MacRumors爱范儿今年同样将前往 WWDC 现场，为大家带来 WWDC 系列活动的专题报道，届时请锁定爱范儿主站、微信公众号、小红书和微博上的系列内容。iOS 27：玻璃、专业相机和 AI Siri对于 iOS，我们首先需要明确的是：液态玻璃（Liquid Glass）设计在 iOS 27 中不会退场，它就在那里，稳稳地接住你。就像刚更新的 iOS 26.5，苹果对于液态玻璃效果的微调也将持续到 iOS 27 里，直到大家习惯或者接受它的视觉风格。除了继续优化液态玻璃的视觉效果，iOS 27 也被很多媒体和开发者预测为是一次「雪豹」式的更新——比起新功能，删除陈旧代码、优化性能表现、调整视觉效果和改善整体流畅度将成为这一代 iOS 的主要更新方向。在支持范围方面，iOS 27「据称」将会支持到 iPhone 12 系列和 iPhone SE 3 为止。换言之，iPhone 11 系列及更早的机型将无法更新 iOS 27。而在继续优化液态玻璃的同时，iOS 27 还有可能会在一些小细节的层面，为我们「管中窥豹」一下今年 9 月份即将到来的折叠屏 iPhone。图｜MacRumors当然，苹果肯定不会在 WWDC 上直接展示有关大屏和横屏布局的优化措施。但在后续的开发者课堂中，大概率会出现有关自适应 app 布局、不同长宽比 UI 设计规范之类的相关内容。几乎可以肯定的是，今年 WWDC 上，苹果会在涉及到 iOS 的内容和课程里添加一些有关侧边栏和分屏（或者 Stage Manager）相关的内容，为横屏布局做出铺垫：另外，本次 WWDC 上我们还可能会见到苹果开放一些更加深层的相机 API，或者一些全新的相机玩法。这些本质上都是为 iPhone 18 Pro 上可能到来的相机「专业模式」做好准备。根据计划，苹果会让 iOS 27 的相机 app 更具自定义性，用户可以调整相机界的功能开关布局，比如闪光灯、EV、模拟光圈、照片风格和长宽比等等。图｜Shima Zamil最后不得不提的还有 AI Siri ——根据彭博社的预测，iOS 27 将会是 Apple Intelligence 官宣两年之后的「汇总落地」。iOS 27 版本的 Siri 中，苹果预计会拓展跨 app 信息提取、屏幕内容识别、模拟连续操作、app 深层功能集成等等多方面的 AI 能力，努力追赶甚至看齐 Gemini 在 Pixel 上的功能范围。图｜Google更有甚者，传说中的「对话式」Siri 也有希望在今年 WWDC 上亮相，或者至少看到一些相关的底层功能。相比以前那个问一句答一句的智能助手，Siri 将会获得类似 ChatGPT 那样的「人味」对话功能。并且在 Siri 对话中也支持多模态内容，甚至可能会取代原本的「聚焦搜索」功能：聚焦搜索（左）和 Siri 搜索（右）至于那些 iOS 26 里面已经存在的 AI 功能，比如图乐园、Genmoji、写作工具之类的，它们在 iOS 27 中也主要将迎来性能优化，并且在新的 Siri app 里面获得单独的入口。macOS 27：圆角、触屏和 Intel对于那些熬夜观看 WWDC 的 Mac 用户来说，看苹果整活王 Craig Federighi 怎样宣布今年的 macOS 系统名称，始终是不可或缺的一环。图｜MacRumors继去年 Tahoe（太浩湖）之后，苹果的商标注册列表上还剩下很多加利福尼亚州的特色。你觉得 macOS 27 的系统代号会是什么？和 iOS 27 一样，今年 macOS 27 的重点同样是继续打磨液态玻璃、优化系统性能，以及落地更多的 AI 功能。不过就像爱范儿之前提到的：macOS 26 的 UI 问题不止于液态玻璃，不统一的圆角、过于密集的小图标，以及辨识度过低的系统 app 图标，联手造成了 macOS 26 的可读性问题。图｜Reddit除了赶紧优化系统 UI 之外，macOS 27 的另一大潜在看点则是它对于触屏功能的支持。彭博社苹果专家马克·古尔曼去年和今年都曾密集爆料过，苹果正在计划开发一款采用触控 OLED 屏幕的 MacBook Pro，上市时间可能在 2027 年内。图｜9To5MacmacOS 27 虽然不会在 WWDC 上就大张旗鼓地宣传「新增触控支持」，但大概率会按照苹果的风格，做出一些为触控功能铺路的举措——比如新增多种尺寸的 UI 控件规范、优化 macOS 支持的缩放比例范围、新增一些看上去不太适合鼠标键盘的图标动画等等。AI 方面，macOS 27、iOS 27 和 iPadOS 27 预计都将迎来一次「相册」更新，照片编辑器里的 AI 工具不再只有橡皮擦，大家习惯的 AI 扩图、自动优化、重构图甚至 3D 照片等等都有望加入。图｜AppleInsider至于 AI Siri 本身的功能，比如「对话式」沟通、屏幕内容识别、系统级自动化功能等等，macOS 27 能够得到的更新和 iOS 27 差不多。这也是 iPhone、iPad 和 Mac 处理器同源带来的优势之一。坏消息是：macOS 27 极有可能会排除所有使用 Intel 处理器的 Mac 产品，仅支持 Apple Silicon 型号的 Mac。这就意味着 2019 款 16 寸 MacBook Pro、2023 款至强处理器 Mac Pro 等等顶级产品都将无缘 macOS 27，为苹果的英特尔时代画上第二个句号。图｜MacWorld另有开发者指出，苹果也在逐步停止 Rosetta 2（x86 到 ARM 转译）的更新。macOS 27 预计将是最后一个完整支持 Rosetta 2 的版本，敦促第三方开发者及时更新 ARM 架构的 app。稳扎稳打的 27 年总的来说，和谷歌努力推荐的「Android 全形态化」不同，苹果今年 WWDC 的路线依然是以稳健为主。iOS、iPadOS 和 macOS 各自的领域划分很明确，虽然产品形态更新，但系统短期内不会出现类似谷歌 Aluminum OS 那样的融合现象。图｜Geeky Gadgets对于「液态玻璃」来说，在缝缝补补一年多之后，苹果的确需要静下心来，先处理一些由于玻璃 UI 和 AI Siri 烂尾导致的系统基础体验问题了。至少对于一些旧设备来说，优化后的 iOS 27 很可能会比 iOS 26 更值得更新。附 WWDC 26 日程表，感兴趣的读者可以通过苹果开发者官网（https://developer.apple.com/cn）注册并参与在线课程：

1995 年，微软内部搞过一次颇具乌托邦色彩的尝试。他们觉得当时的 Windows 系统操作太复杂，于是想把电脑桌面直接做成一个普通人看得懂的「房间」。你想打字就点桌上的笔记本，想看时间就点墙上的挂钟，不用思考那些反直觉的操作路径。这个产品叫 Microsoft Bob，结果上市没多久就光速下架了。原因五花八门，但究其根本，它并没有真正渗透进系统底层，只要用户的需求稍微复杂一点，就只能无能为力了。三十年后，很多 AI 助手依然没有完全走出这个困境。最近我体验到了腾讯新的 AI 助手 Marvis 马维斯 ，终于看到了一些不同。相比于别人都在卷谁是 AI chatbot，Marvis 想干的是直接把整台电脑变成你可以随意使唤的搭档。官方给它的定位也正儿八经，叫「操作系统层级的个人 AI 助手」（官网：marvis.qq.com）。不过，当我知道它背后是那个做了 14 年应用宝的团队时，就不觉得意外了：这种长期在 PC 和移动端底层生态里打拼的出身，恰恰意味着它不仅懂 AI，更懂设备和系统。目前，Marvis 已率先覆盖 Windows PC 与 Android 手机端，iOS 和 macOS 版本也在推进中。APPSO 也在评论区准备了一些福利邀请码，手慢无。六个 Agent，一支随时待命的私人 AI 团队安装过程不算复杂，但对设备有点要求。Windows 端最低门槛是 8 核 CPU、16GB 内存加上固态硬盘。而且据官方透露，未来还会逐步兼容 6 核。市面上很多 Agent 产品动辄让用户自己去搭工作流、安插件，Marvis 则是开箱即用，六个 Agent 零配置上手，对普通用户非常友好。刚进界面，里面已经有六个 Agent 在 7×24 小时待命了。六个 Agent 都有自己的工位。没接到任务的时候，有的在打盹，有的在办公室里闲逛，有的在健身，有的在喝咖啡，还有的直接去上厕所，一旦你发出指令，任务就会拆解分配给对应的成员。这是一个分工明确的团队：PM 负责听懂你的需求，拆解任务往下派活；File Agent 处理文件的搜索、阅读和格式转换；Computer Agent 专门对付系统配置、查硬件、调夜间模式；APP Agent 负责在各种软件里代为操作，比如查个机票或者电商比价；Search Agent 负责快速从公开资源中定位答案并给出关键引用；最后是个 Browser Agent，定位是网页交互专家，进行网页交互与数据抓取。任务链比较长的时候，看看这些动画也是别有一番趣味。最有意思的是，这些 Agent 可以并行工作。举例而言，我先抛出一个需求：「电脑开机太慢了，帮我看看哪些自启动程序没用，帮我关掉。」注意看，工作日志里还会显示 Token 消耗量，每天的免费额度是 2000 万 Token，堪堪够用。结果显示，它给每个程序配了「人话版」说明，建议你哪些是可以按需、保留的第三方软件。重点是，涉及这种核心配置修改的操作，它会主动和你确认一下，才会执行。这种「AI 不会自作主张」的机制，稳得让人心安。电脑用久了，总会碰到几个反复折磨人的小问题。「查看电池健康状况」「网速感觉变慢了，帮我测一下现在的上传下载和延迟。」「想玩个新游戏，帮我检测一下硬件配置，对比最低要求看看差在哪。」以前遇到这些事，要么去开始菜单里找设置，要么去网上搜教程。现在，说一句话就行了。Marvis 深度对接了系统信息和配置入口。它能看到硬件参数，也能直接修改系统设置。它不是帮你打开「控制面板」然后让你自己找，而是直接把事情办了。对于那些不熟悉电脑设置的人来说，这块的体验是最直观的。举个例子：你想关掉 Windows 锁屏上的广告，你知道它在设置里叫什么吗？叫「关闭 Windows 聚焦」，有多少用户知道这个术语是关广告的？但 Marvis 能帮你一句话关掉。本地文件时代，迎来 AI 入口为了测它的能力，我还翻出了一件困扰已久的事：两个月前从某个群里下载的一张截图，只记得内容是关于《黑神话：悟空》，文件名完全忘了。我输入：「帮我找一下关于《黑神话：悟空》的图片或截图，大概是一个月前的事儿。」说实话我没抱太大希望。但 Marvis 接单后，首先根据 Skill 指引，按文件名关键词和时间范围筛选出 1000 张照片，然后用 Python 快速筛选整个结果文件，并最终找到 8 张候选图。第一张正是我要找的那张。而且它的搜索范围极其宽泛：文件名、文档内容、图片里的景、截图里的字，全覆盖。哪怕描述模糊，它也都能捞出来。这个能力，比 Windows 自带的搜索确实强出了一大截。当然，一听到「扫描图片」，可能会有人担心隐私。首次使用时你可以自己设定扫描范围。如果在生成的图谱里看到不想被索引的内容，右键一键屏蔽即可。而且端侧模型本身自带基础过滤，不是无差别的扫描。除了找文件，Marvis 还能对本地的文档、表格等多种类型文件做深度理解和分析——支持内容优化、图表生成、文案润色、格式转换，帮原本费时费力的事情变得顺手。比方说，下载了一批学术论文，摞在那里不知道从哪里读起。直接问 Marvis：「扫一遍这几篇 PDF，告诉我每篇的研究问题、方法和主要结论。」几分钟内，哪几篇值得细读、哪几篇可以先放着，心里就有数了，不用一篇一篇翻摘要。再比如读完《三体》，笔记记了好几页，马上要在课上分享，手动整理提纲、挑内容、套 PPT 模板，少说也要折腾一两个小时。但让 Marvis 根据读书笔记，生成一份 PPT，它能读懂笔记里的逻辑和层次，不需要手动复制粘贴再套模板，理解内容之后直接生成。手机接管电脑，随时随地都能干活手机遥控操作电脑，则是非常实用的跨端体验。去官网下了 Windows 客户端，再给手机安上 Marvis App，用同一个微信或 QQ 登录，还能直接连通两端。具体来说，手机 App 连上电脑后，可以在手机上实时看到电脑桌面，并用触屏直接操控。哪怕电脑处于锁屏状态，手机依然可以接管，不需要电脑一直保持解锁状态。比如我人在外面，突然听到朋友说前阵子新出了一款《洛克王国：世界》游戏，这时我就可以掏出手机输入：帮我在电脑上安装《洛克王国：世界》，回到家就能玩上了。连接微信 ClawBot，也是没问题的，微信扫码登录后就能开启对话。但我认为最实用的场景，是给家里长辈远程排查电脑问题。爸妈说「QQ 打不开了」，以前要么电话里盲猜，要么回家处理。现在直接接管他们的桌面，扫一眼就知道问题出在哪。前提是你的手机端 Marvis 马维斯，要能连接爸妈家里的电脑端马维斯，手机端是支持连接多个 PC 设备的。而如果电脑断网了，手机端会自动切到云 Marvis 模式，支持独立调用云端能力，大部分任务依然能跑。并且这里还有一个独家绝活：在电脑上直接操控手机里的 App。同花顺、飞常准等应用已经在对接合作，这也得益于应用宝团队多年的跨端引擎积累。现在应用宝在移动端、pc 端、车载上都有支持，跨端研发经验充足。和其他桌面 Agent 一样，Marvis 也支持自定义自动任务。比如设置一个「深圳厚德品园二手房均价提醒」，填好任务说明——「查询最新二手房均价并通知用户」，再指定执行时间为每周一上午 10 点，之后什么都不用管，到点自动运行任务，并直接推送结果。另一个值得单独说的功能，是技能（Skill）的自定义扩展。Marvis 支持一键安装专属技能包，提供了不少第三方 Skill，当然，你也可以导入自己常用的 Skill 文件，把它改造成更贴合个人工作流的样子。有特定行业需求的用户，比如内容创作，理财投资、学术研究等等，不用依赖官方迭代，自己就能把对应的知识库和操作逻辑装进去。最后，主打出活的 Agent 都绕不开最关键的问题：安全。Marvis 马维斯提供了两种模式：效率模式和隐私模式。设置里切到「隐私模式」，Marvis 会加载本地端侧模型，无论是文档还是图片的检索、解析、识别等，全部在本地运行，文件不会出你的电脑，甚至断网了都可以使用。官方所说的「数据不出域」，在拔掉网线的那一刻，有了最真实的体现。不过，体验「隐私模式」的门槛比较高，要求 16 核以上的处理器、32GB 内存、16GB 显存起步，固态还得匀出 35GB 空闲。所以平时日常使用切回「效率模式」即可，理解和规划由云端大模型（混元和 DeepSeek V4）完成，执行在本地，速度更快。两种模式按需切换。让 PC 真正开始「理解」人我们现在用电脑的方式，本质上还是上世纪八十年代的逻辑：一个需求对应一个软件，你要去找它、打开它、学习它的规则来完成任务。四十年来，软件更多，设置更复杂了，但这套逻辑并未改变，普通用户和电脑之间的距离，也没有真正缩短。Marvis 的逻辑反过来了：先理解这台电脑里有什么，再决定为你做什么。拿我爷爷来说，他用了几年电脑，至今不知道「控制面板」在哪。但他知道电脑开机慢，知道想找一张老照片。过去，这个解法是打电话求助子女；现在，就算他不会打字，也可以直接在微信 Clawbot 或者手机马维斯上，把需求说出来，电脑就会帮他做。文件搜索、系统设置、跨端控制、自动任务、隐私模式、Skill 扩展，看上去是几个不同功能，背后指向的其实是同一个方向：个人 AI 不应该只是一个问答入口，而应该成为设备里的任务调度层。多说一句，我觉得 Marvis 是有审美的。虚拟办公室动画、任务进度呈现等细节，让它不像冷冰冰的 AI 工具，更像一个有秩序、有分工、也有生活气的小型工作室。它尤其适合几类人：文件多且乱的本地重度用户；重视数据隔离的财务、法律、医疗从业者；喜欢自定义工作流的效率玩家。和需要靠「养」才能用好的「虾」和「马」，Marvis 初始内置 Agent 和自然语言指令，也足够覆盖多数日常电脑操作，对于很多 AI 焦虑不知道怎么用的人，Marvis 都更加的平权、无门槛、开箱即用。未来的个人 AI 会像搜索、截图一样，自然地成为操作系统的一部分。三十年前，Microsoft Bob 试图让用户更容易理解电脑；三十年后，Marvis 试图让电脑反过来，真正理解用户。Marvis 马维斯能走到这一步，和这个有十四年应用商店生态积累和四年跨端跨系统技术积累有关。庞大的用户基数让他们更懂用户在不同设备上需要什么、怎么使用。这种积累很难被纯 AI 公司在短期内平替——他们可以快速训练一个更聪明的模型，却没法快速建立起对 PC 底层和移动生态的控制力。至此，一个能用自然语言驱动的系统级 AI 助手，在中文生态里终于有了真正可用的样本。所谓个人 AI 重点从来不只是 AI，更是个人。Marvis 终于让我们看到，当技术选择与每一个具体的需求并肩同行时，它所释放出的力量，既润物无声，又重塑天地。这是个人 AI 真正开始变得有用的时刻。

💻Google 发布 Seedance 2.0 最强对手，还有一堆 AI 更新📱iOS27 快捷指令或迎来大改版，一句话就能制作🤖DeepSeek 回应模型异常回复🍎苹果发布 AI 加持的无障碍功能🖥️AMD AI 开发者大会首次登陆中国，苏妈提出「智能体主机」新品类🚗7 分 34 秒 931，小米 YU7 GT 刷新纽北纪录📱华为五一假期领跑中国手机市场、安卓厂商旧款机型均价上调 13%🚙小鹏 Robotaxi 量产下线，预计下半年开启示范运营🔧禾赛成为奔驰 L3 激光雷达战略合作伙伴，将于泰国工厂投产🌐特斯拉放弃印度建厂计划🧠美图布局 Visual Agent，Chance AI 完成数百万美元天使轮融资🏆小米影像算法 CVPR 2026 NTIRE 三项获奖💰长江存储启动 IPO 辅导⚡比亚迪电池CTO谈闪充、寿命与安全边界💡强化学习之父 Sutton 重申「苦涩的教训」引争议📱vivo S60 系列定档 5 月 29 日🎥698 元起，影石 Mic Pro 正式发布💻1299.99 美元起，微软发布新一代商用 SurfaceGoogle 发布 Seedance 2.0 最强对手，还有一堆 AI 更新今天，Google 在刚刚落幕的 I/O 大会上集中发布了一系列新品，包含两款新模型、全新 AI 智能体产品及音频智能眼镜预览，Gemini 的能力边界进一步向搜索、办公、创作和硬件延伸。AI：Gemini Omni 正式发布，定位「任意输入输出」多模态生成模型，支持视频生成、风格编辑与元素替换，可在生成过程中以对话形式修改；Gemini 3.5 Flash 同步推出，在编程、真实环境 AI 智能体任务等测试项目上较 3.1 Pro 有所提升，输出 token 速度比其他前沿模型快 4 倍，昨日起面向所有用户开放；Gemini 3.5 Pro 预计下个月开放；个人 AI 智能体 Gemini Spark 亮相，运行于 Google Cloud 专用虚拟机，支持全天候后台任务执行，可通过 MCP 协议与第三方应用互联互通。Google 同步新增每月 100 美元 AI Ultra 档位，原 250 美元最高档降至 200 美元；Gemini App 大改版，引入了全新的设计语言 Neural Expressive，加入流体动画、鲜艳色彩、新字体和触觉反馈。硬件：音频智能眼镜预览亮相，由三星参与硬件构建，Warby Parker 与 Gentle Monster 负责设计，搭载摄像头，支持 AI 视觉和多模态输入，Gemini 回答通过耳机播放，兼容 Android 和 iOS；XREAL 与 Google 在现场带来 Project Aura XR 眼镜首次公开体验，首发搭载 Android XR 系统，采用 XREAL 自研 X1S 芯片，视场角达 70°，确认今年全球上市。🔗 相关阅读：刚刚，Seedance 2.0最强对手上线！这就是AI视频的「Nano Banana 时刻」iOS27 快捷指令或迎来大改版，一句话就能制作据彭博社报道，苹果正筹备在即将发布的 iOS 27 和 iPadOS 27 中引入包括系统级语法检查、自然语言创建快捷指令及 AI 自定义壁纸在内的多项人工智能新功能。知情人士透露，苹果此次升级的核心在于扩展系统级 AI 写作工具，新增类似 Grammarly 的语法检查器。该功能界面以半透明菜单形式从屏幕底部弹出，并列显示原文与修改建议，用户可选择逐条采纳、全部应用或一键忽略，同时系统在键盘顶部集成了「通过 Siri 撰写」和「辅助撰写」快捷选项。此外，「快捷指令」App 也将迎来大幅改版。测试版本允许用户直接在「你希望快捷指令执行什么操作？」的文本框内输入自然语言需求，由系统自动构建和配置自动化流程，大幅降低使用门槛。同时，新系统将集成基于 Image Playground 的生成式 AI 壁纸工具。上述操作系统预计于今年 6 月的全球开发者大会（WWDC）上发布，并于 9 月向公众推送。WWDC26 将于北京时间 6 月 9 日至 13 日举行，爱范儿也将在 Apple Park，为大家带来第一手现场报道。DeepSeek 回应模型异常回复昨天，DeepSeek 官方小红书发布最新说明，回应了关于< think >字符触发模型异常回复的问题。DeepSeek 方面表示，近期其关注到有用户反馈，在与 DeepSeek 模型的对话中输入「」这类特殊字符后，模型偶发返回不可预期的内容，使部分用户产生了「对话泄露」的疑虑。针对这一问题，DeepSeek 的技术团队经全面排查后发现：输入「」等字符触发返回异常内容，属于特殊字符引发的模型幻觉，不涉及安全问题或隐私泄露。DeepSeek 方面也表示，后续其将通过针对性训练增强模型对特殊字符的识别与处理能力，修复相关的已知问题，优化模型在此类场景中的表现。苹果发布 AI 加持的无障碍功能苹果昨日发布了一系列无障碍功能更新，宣布将 Apple Intelligence 引入 VoiceOver（旁白）、放大器、语音控制和无障碍阅读器等核心辅助功能，计划于今年晚些时候正式推出。在视觉辅助方面，VoiceOver 新增「图像探索」功能，可对照片、账单等图像内容生成更详细的描述；放大器同步引入 AI 能力，在高对比度界面下提供视觉描述，并支持语音指令操控。语音控制则引入自然语言输入，用户无需记忆按钮标签，可直接用日常语言描述界面元素完成操作，初期支持英语，覆盖美国、加拿大、英国和澳大利亚。此外，苹果还公布了以下新功能：视频生成字幕：利用设备端语音识别，为未附带字幕的视频自动生成字幕，适用于 iPhone、iPad、Mac、Apple TV 和 Apple Vision Pro，初期支持英语，面向美国和加拿大用户；电动轮椅控制：Apple Vision Pro 借助眼动追踪系统为无法使用操纵杆的用户提供替代输入，首批支持 Tolt 和 LUCI 两款替代驱动系统，兼容蓝牙和有线连接；其他更新：visionOS 新增车辆运动提示；tvOS 支持大字体；「名字识别」扩展至全球逾 50 种语言；索尼 Access 控制器现可与 iOS、iPadOS 和 macOS 配合使用。AMD AI 开发者大会首次登陆中国，苏妈提出「智能体主机」新品类AMD AI 开发者大会昨日首次在上海举行，吸引超过 2000 名开发者到场。AMD CEO 苏姿丰发表开幕演讲，并与零一万物 CEO 李开复就 AI 智能体话题展开对话。大会重点推介「智能体主机」新品类概念，以锐龙 AI Max+ 系列处理器为核心，至高支持 96 GB GPU 专属显存，可本地运行最高 200B 参数模型。惠普、华硕、联想、宏碁等厂商已推出超过 35 款相关产品。软件层面，ROCm 7.2 新增对 Windows 和 Linux 的扩展支持，并新增兼容锐龙 AI 400 系列处理器。AMD 同时宣布与阿里云合作，提供基于 Radeon GPU 的免费开发者云，并推出面向中国开发者的专属项目。🔗 相关阅读：黄仁勋刚走，苏妈就来了 | 附演讲全文7 分 34 秒 931，小米 YU7 GT 刷新纽北纪录小米汽车昨日宣布，小米 YU7 GT 以 7 分 34 秒 931 的成绩刷新了纽北 SUV 圈速纪录。这一成绩出自小米汽车首席测试车手任周灿，他因此成为了首个获得纽北官方圈速认证的中国车手。雷军此前表示，小米 YU7 GT 是一台「原汁原味的纯血 GT」，充满驾驶乐趣的同时，还拥有舒适的长途驾乘体验数据上，YU7 GT 的最大功率为 1003 马力，极速可达 300km/h，CLTC 纯电续航为 705km。小米汽车 18 日宣布，YU7 GT 将于 5 月 21 日（本周四）19:00 正式上市。华为五一假期领跑中国手机市场、安卓厂商旧款机型均价上调 13%根据 Counterpoint Research 昨日发布的《中国智能手机周度销量追踪报告》，今年五一假期前后两周，中国智能手机销量同比下降 16%。存储价格上涨推高整机售价、促销力度弱于往年，是销量下滑的主要原因。华为继续领跑市场，4 月周度销量份额持续保持在 25% 以上。畅享 90 Pro Max 表现突出，Pura 90 Pro 系列及折叠屏新品 Pura X Max 于 4 月底上市后销量强劲；苹果五一假期销量同比增长 3%，但环比出现常规性放缓。iPhone 17 系列需求逐步回归常态，国产品牌集体涨价客观上提升了 iPhone 的相对竞争力；其他安卓品牌旧款机型均价已平均上调 13%，短期内或进一步压制换机需求。Counterpoint Research 预计，618 促销活动有望带动 6 月初市场温和回升；今年全年中国智能手机出货量预计同比下滑 9%，但仍优于全球平均水平。小鹏 Robotaxi 量产下线，预计下半年开启示范运营昨天，小鹏集团董事长何小鹏发文宣布，小鹏首台量产的 Robotaxi 在广州工厂正式下线。何小鹏表示，该 Robotaxi 以小鹏 GX 为原型车，搭载小鹏第二代 VLA 大模型，4 颗自研图灵 AI 芯片，本地算力达 3000TOPS 全球最高，实现 L4 级自动驾驶能力。其还强调，这也是中国首台纯视觉 Robotaxi+线控底盘+前装全冗余的量产下线。据悉，未来几个月，小鹏会在广州开启 Robotaxi 示范运营。禾赛成为奔驰 L3 激光雷达战略合作伙伴，将于泰国工厂投产5 月 19 日，全球化智能科技企业禾赛科技宣布，据公司 Q1 财报，禾赛成为梅赛德斯-奔驰（Mercedes-Benz）的激光雷达战略合作伙伴，并被确认为其 L3 级自动驾驶车型的供应商。据悉，此次签署的新供应协议将全面支持奔驰在欧洲及中国市场的车型项目，所需产能将由禾赛新落成的泰国「伽利略」制造中心提供坚实保障。特斯拉放弃印度建厂计划据 Drive Safe and Fast 报道，印度重工业部部长库马拉瓦米（Kumaraswamy）昨日正式确认，特斯拉已放弃在印度建设工厂的计划。分析人士指出，此次放弃建厂涉及多重障碍。印度本土供应链与工业生态系统存在明显缺口，难以支撑特斯拉的制造需求。此外，双方就「关税减免应先于建厂」还是「建厂应先于关税减免」这一核心问题始终未能达成共识。印度基础设施的不足、当地汽车市场消费能力与特斯拉高端产品定位之间的落差，以及印度营商环境中存在的政策不确定性与监管风险，也被列为主要原因。美图布局 Visual Agent，Chance AI 完成数百万美元天使轮融资Chance AI 昨天宣布完成数百万美元天使轮融资，由美图领投，NYX Ventures、阿里系投资机构等跟投，资金将用于模型迭代、北美增长及商业化探索。Chance AI 成立于 2025 年，核心产品是以摄像头为主要入口的 Visual Agent 应用，采用「看见—理解意图—调用 Agent—完成行动」的交互逻辑，可根据用户上传图片判断场景与意图，调度对应 Agent 执行任务。小米影像算法 CVPR 2026 NTIRE 三项获奖小米技术昨日宣布，在今年 CVPR 2026 NTIRE 赛事中获得三项奖项：高效超分辨率赛道冠军、人像修复赛道冠军，以及反光消除赛道亚军，分别由小米玄戒多媒体算法团队和小米大模型应用团队斩获。NTIRE 是图像恢复与增强领域规模最大的学术竞赛之一，由 CVPR 组委会承办。小米表示，后续将推动上述技术从研究走向产品落地。长江存储启动 IPO 辅导昨天，据证监会网站披露，长江存储控股股份有限公司在湖北证监局办理辅导备案登记，拟首次公开发行股票并上市，辅导券商为中信证券。比亚迪电池CTO谈闪充、寿命与安全边界据 36 氪报道，比亚迪电池事业群 CTO 孙华军就第二代刀片电池与闪充技术接受专访，详细回应了关于闪充寿命、电池热管理机制以及磷酸铁锂技术路线等行业争议。今年 3 月，比亚迪推出第二代刀片电池及闪充技术，支持 150

今天不追光和 vivoS60 系列一起做一颗属于自己的星星把 星星海 握在手里

版权所有 ©广州利沃致远投资管理合伙企业（有限合伙）2008 - 2023。以商业目的使用爱范儿网站内容需获许可。非商业目的使用授权遵循CC BY-NC 4.0。

All content is made available under the CC BY-NC 4.0 for non-commercial use. Commercial use of this content is prohibited without explicit permission.

厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨主作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了联合作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了05/15 00:07Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。免责声明与作者所需披露的一切首先声明一下——本人仍然是在校的本科微电子硕士考古学的大学生，并没有经过任何的传媒学相关的学习，也没有从事相关行业的经验。目前仅有的和媒体沾边的事情是在酷安、少数派等平台发过一些文章（其中有很大一批都是在最后诉诸于，说好听点是真实生活场景，说难听点开始抒情的文章），以及一个 B 站粉丝量远没有破千的账号，因此我远远没有承担过养一个工作室的压力与责任，也没体会到在竞争激烈的环境中生存的艰难。因此以下内容不构成任何形式上的专业意见，只是一个愤青在 Random Yapping。多有冒犯之处，但我本心无意去批评任何具体的某一个媒体老师，只是想尝试提出一个也许还算有点价值的问题。以下内容提到「游戏区」，也仅限于单机游戏，不包含手机游戏、桌游等类型。以下内容提到「数码区」时，通常仅指以手机、一部分 PC 组件为主的主观体验向评测，且我的能力不足以涵盖 DIY 区1以及涉及到实验设计与数据分析的其他评测内容，虽然随着硬件厂商的创新进入后期演变为渐进创新，产品节奏也有所放缓，在社区中也能听到「DIY 区不如以前」的感慨，但本文尽力按照「不理解，没有切身观察就尽量少乱锐评」的原则展开，但撰写过程中有很多内容没收住，因而请让我提前声明一下——以下很多内容只是社区观察，不完全代表事实判断。最后，说句坦诚的实话，这个内容写的过程真的很痛苦——我真的不想让这篇文章变成直球辱骂媒体老师、攻击厂商或者是殴打观众的情绪发泄文，从个人发展角度来说不希望因此在进入之前就已经和某些职业无缘，也不想因此得罪那些因真心倾慕而认识的那些做这一行的朋友，且我真的不希望因为自己的过分言论把发文章的平台本身也拖下水，但因为笔力实在是有限，尽管拉着 Gemini、GPT、DeepSeek 等一众大模型核验，也找了一些朋友探讨，尽可能不要直球攻击特定的一两家媒体也不要说的太过分，但是字里行间还是有一股火药味拼尽全力消不掉。真的很抱歉，尽管我其实将自己也包括在本文的攻击射程以内——我自己就是后文中所批判的那个自认为掌握了真实，在没有厂商利益束缚与照顾社区的必要下肆意输出内容，甚至有很多是过当内容的小自媒体。这些内容因为是我自己一段时间内自我质疑与焦虑的外化，因此，本文必然存在一些因深度不足、角度有限而导致的难以避免的暴言，简而言之——就是年轻时犯下的错误。麻烦各位读者多多体谅。也欢迎您在评论区里多批评指正，让我多与您交流学习。以下内容如果有哪里说的不合适的地方，有什么可能涉及到的法律争议，请尽快联系我去撤换文章中对应的部分，谢谢！本文长度约为三万字左右，前三部分观点将会较为尖锐，会举几个相对较为敏感的案例来佐证观点，后两个部分则尝试解决前文抛出的问题，内容相对温和但字数也相对较长，请坐和放宽。主观口胡系列是一个尝试将动漫、电影、游戏叙事与更加严肃的讨论结合起来，去尝试探讨一些特定大命题的个人尝试，可能会一定程度上背离这些作品的创作初衷，聊的可能也会比较混乱且随意，甚至会有意提取个别元素，曲解原文本来聊一些其他领域的内容。这次我想把一个大约是三四年前就开始产生的想法去做进一步的深化，并尝试结合近些年来我对游戏区与数码区的一些粗浅且悲观的看法，来尝试去聊一个问题，一个也许注定没有答案，跨多少领域学科都难以得出解法的一个小小的问题——主观体验向评测的现状怎样，尽头到底为何方？又是什么构成了它们的专业性？Part1：太多水，太多奶酪——主观内容的梦魇虽然这么说有失偏颇，但我一直觉得游戏评测是一个相当主观的内容——评测者的个人阅历2、偏向的游戏类型甚至是组成一个游戏的系统，对 GDC 中诸多开发者分享的设计理念的理解，以及同期的同类型作品等等——这些都会对你最终得出什么样的结论，给游戏最终打什么分造成影响。所以在因近期的《神探杰克鼠》评测而陷入争议的 IGN 内部——作为一个有 TGA 评选权的，相对来说算是行业内位置靠上话语权比较大的游戏媒体——他们在测评游戏时给出的流程是：在内部筛选最适合某类游戏且当下仍有档期的编辑，由他们在内部其他编辑的帮助下完成对于这个游戏的评测，给出评分，在内部经过审核与排期，最终以文章 + 视频的形式先后放到公共平台上。但是，这个时候总会有个但是。虽然这样的流程确实在内部行之有效，为行业贡献了一批质量很高的评测，尤其是在神作频出神仙打架的那些年里——但同样不可否认的是，无论你之前做对过做好过多少次，一旦出现几次测评结果与大众预期不相符，甚至出现有失水准的内容，就很容易在玩家社群中丧失公信力，变成 IG8——《宝可梦欧米伽红宝石/阿尔法蓝宝石》的 too much water 扣分，《超级马里奥 3D 世界+ 狂怒世界》的 7.0，《死亡搁浅》的 6.8——你知道的，日光之下从来都没有什么新事。而且当我们把视角继续向下探时，会发现这样的问题不仅仅是发生在 IGN 一家，Polygon 的《DOOM(2016)》、VentureBeat 的《茶杯头》（他的操作已经不足以让他完成第一关，只能借其他同事存档继续评——尽管他据自己说已有着超过二三十年的评测历史）……所以到底发生了什么？当然，抛开社区中已有的被讨论过无数次的，对评测者的质疑不谈，我们也许可以从更加细致的视角去看这件事。就像你在做数码类评测会提到机型的性能与续航、外观、系统流畅度与体验、拍照、外放表现，但你不可能在非首发上手的评测里一样面面都要点到，而是根据个人的使用习惯有自己的偏好一样，IGN 的每个评测者在评测游戏时也会有自己的明显倾向，让我们以当年评测《死亡搁浅》的 Tristan Ogilvie 为例——他评测一款游戏的重心是动作系统、交互反馈与流程的顺畅程度。在我眼里，相较于小岛在前作《合金装备 5：幻痛》中用一系列主线支线任务，吸引玩家不断的从母基地回到开放世界，完成一串任务后撤离不同，本作让角色更加持久更加连续更加长时间的停留在这个世界里，将任务可视化，变为地图上要到达的目标点，并在沿路设置一系列需要潜行的敌人据点、风景奇观，并且可以通过基建加速到达目标点跳过一部分游戏内容来优化体验的开放世界而很不幸的是，本作正正好在她关注的几个角度精准的踩中了地雷：虽然相较于《合金装备 4》有所收敛，但游戏内不断插入冗长的过场 cg，以至于从室外场景进入到地图传送需要连续跳过数个动画；游戏流程放速过缓，且游戏内的任务设计被认为千篇一律，主线和支线均由送货构成；动作系统相较于前作乐趣有所减少（尽管二者的游戏定位有着根本性的不同，也许从来就不该这么要求角色的性能），击败 bt 后的奖励反馈也较为缺乏，以至于编辑出于他所偏向的几个维度的也许从传统游戏性的角度合乎情理的考虑，选择给这个游戏 6.8 分。随后我们都知道发生了什么——之后的几年让游戏的异步社交大放异彩，以一个非常奇怪且迅速的方式，让小岛秀夫这种几乎像是寻求争议，找骂、试图以先锋艺术电影的形式在游戏内进行的表达，能被更多人看见，被更多人共情，似乎小岛秀夫在《合金装备 2》后，再一次，等到了时代来验证他思想上的前瞻性。因而，时代背景变迁下大家对小岛独特表达的广泛接受、小岛秀夫的明星制作人光环以及评测中确实存在的一些过当的指责，都让玩家社群反过来开始质问一件事：为什么我觉得好玩的游戏，你却要给出这么低的分？当然，IGN 作为行业头部媒体，其本身在游戏界有较大的声量较高的影响力，且也有着个人做自媒体乃至中小型媒体难以企及的资源，因此它说的话做的评测理应被玩家更加严格的审视，树大招风不无道理——而且至少在我举得几个例子里，引发争议的游戏本身也是在不同领域内质量过硬的存在，只是随着营销的不断发展，疫情时代与后疫情时代每个亚文化小领域的的扩圈与自媒体时代的到来，这种原本对媒体对权威的的尚处合理范围的不信任将逐步扩大，直到这种现象一步步扩散到每一件事，每一个领域，乃至我们每个人都难以避嫌，难以无视。Part2：内容井喷的时代，同质化的幽灵与信任的坍塌数码圈从来不缺学生出身，单纯出于热爱，分享自己体验的自媒体，这样的评测形式本身并非没有价值，它用更加平易近人的角度去涵盖一个产品中用数据评测所难以涵盖的部分，尽管也许，分享自己对于刚入手的数码产品的主观体验并没有什么门槛——有着中专及以上的学历就有足够的语言组织能力去清晰讲述自己的感受，再来一个相机甚至是手机去拍摄素材就能实现，但撰写一篇好的主观体验评测文（不管是哪个领域）就像是一般人眼中的历史研究一样，都觉得准入门槛低，都觉得写的内容是任人装点任人阐释的小姑娘，但真的想要尽可能做到内容详实、条理清晰，往往需要长时间的写作经验积累与大量的训练，而如果想在这个基础上再去追求与社区中主要讨论的内容截然不同的全新角度，或者让我们思考的更加狂野一些，在这个内容上做出深度，探讨一个产品背后更加深刻的有广泛意义的命题，则更加考验个人的水平——但随着互联网的高速发展与信息的高速创作、流通，我们正处在一个内容极其过剩的时代。虽然我承认，我是随着酷市场转型为酷安而踏入数码圈的，本身并不算什么数码区老资历，但你如果要我去回忆在我进入数码圈后这有限的时间内，主观体验向的数码自媒体井喷的那段时间，我想大概是在「老师好我叫何同学」的 5G 视频兴起的 2019 年 6 月前后，以及疫情肆虐的 2021 年后到 2023 年末。前者因为正好切合了国家的 5G 发展战略的需要而被主流媒体报道，被社会广泛关注，因而给圈子引来了泼天的流量——事实上那段时间坚持做手机体验类评测的规模还比较小的自媒体，如今都已经成长为几十万粉丝的台柱，比如我当时看的比较多的花生说；而后者则是由于人们长时间居家，在室外的娱乐方式因客观需要而受限，因而人们开始转向互联网，大约是在这一时期前后，一群赛道更加细分，比如上古数码产品体验、软件交互等领域（就我当时看的比较多的来说是六分超超和苏星河）的体验向自媒体开始大规模兴起，游戏区的很多如今声量较大的自媒体编辑老师也多是在这一时期进入相关行业或者是投身做自己的账号。但与自媒体账号井喷相对应的，却是——请让我想想，该怎么让人更好接受的去讲一个有点不近人情的结论——极其迅速的内容同质化。游戏不是一个竞争激烈且产品间带来的体验差别相较于同类型的数码产品更大的存量市场，有足够多各自差别足够明显的新游戏不断冲刷着视野，因此这种同质化的倾向也许还并不明显，但在数码区，当每个上游的厂商都在绞尽脑汁的结合热门概念，生生在软硬件体验上捏出极其有限的新创新点时，这种感觉就尤其明显——对于自媒体而言，相比起「你自己要讲述什么故事」，你的阐述对其他人有什么价值才更有传播的意义，因而当你迅速的因不受束缚、观点尖锐，获得了一批受众支持迅速爆火后，同质化的幽灵便也开始悄然显现——是顺着原先让你得到人们簇拥的路径一路向西，对着即使没有那么差的内容输出情绪，直到有一刻在无数次路径依赖后判断失误，犯了众怒，还是妥协于平台激励有限而选择靠近厂商，向其他市面已有的无数自媒体靠拢，在厂商给的通稿框架与宣传的大方向已然确定的情况下，进行着极其容易趋同的表达，在首发宣发节奏与内容合作模式影响下，选择性的淡化掉体验中的痛点，即使在前文有提及也往往占据相对较小的篇幅，并随后立马引入一个「但是」，用着首发特调的机型说着好到梦幻的体验，最终在观众的不解中继续在内容理想与工作室运营中艰难的走着钢丝？当然，如果把话说到太绝，往往会有很多偏颇，所以让我们稍微说一些让文章作者更拟人的

「智能队列」如何改善你的家庭网络质量主作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org联合作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org2021/01/31 23:1205/17 12:57更新：2026 年世界电信和信息社会日重推在无线路由器的设置页面上，我们可能会看到一个名为「智能队列」的功能（根据品牌或操作系统的不同，也可能会被称做「智能 QoS」、「SQM」等）。这一功能是如何工作的？是否能让网速变快？是否需要打开？本文将尝试解答这些问题。从「网络如何连接」说起本章节主要介绍计算机网络的基础知识。如果读者已经熟悉相关知识，可跳过这部分内容。在介绍「智能队列」之前，先来了解一下基础的网络知识。两台电脑间进行数据传输，存在两种方式，分别是电路交换和分组交换。对于电路交换，两台电脑独占一条线路进行数据传输。过去的固定电话，大多使用的就是电路交换。由于独占一条线路，两台设备之间的通信较为稳定，很难出现数据丢失、数据传输顺序错乱等情况。但一条线路只能供一组电脑使用，另外一组电脑想要通信，就不能复用原来的线路了。对于分组交换，传输的数据被拆分为小段，每一小段叫做一个数据包（报文）。数据以报文的形式进行转发，从而实现了多组电脑复用一条线路进行通信。电路交换虽然能保证数据传输的稳定性，但由于独占线路浪费太多资源，所以在 Internet 中，使用的是分组交换。分组交换将数据拆分成一个一个的小片段来转发，没有独占线路。如果网络条件不好（例如同时有大量报文经过路由器，超出了路由器的处理能力），很有可能遇到数据顺序错乱、甚至数据丢失等情况。那么，这些情况应该如何避免？在微信视频通话、电视直播等应用中，报文的丢失或乱序，只会造成视频图像卡顿，不会造成太严重的影响。所以这些应用会使用简单的UDP协议，不考虑丢包、乱序。在网页浏览、文件下载等应用中，一般使用基于TCP的HTTP协议。TCP 在报文中携带序号，避免乱序；同时提供了报文的丢失重传机制，避免了丢包。所以，也可以认为 TCP 为两台电脑的应用程序建立了一个虚拟的「连接」。（最新版本的 HTTP/3 协议，使用的则是QUIC。QUIC 虽然基于 UDP，但引入了与 TCP 类似的报文丢失重传机制）如果想要了解更多网络基础知识，推荐阅读「入网指南」系列文章：入网指南 01 | 一文读懂你身边的「网络」入网指南 02 | 当你发消息的时候，发生了什么入网指南 03 | 交换机在交换什么？入网指南 04 | IP 地址大揭秘入网指南 05 | 除了接线，还有什么影响着我们网上冲浪入网指南 06 | 连上 Wi-Fi 就能上网了吗？网络延迟：影响上网体验的重要因素带宽和延迟，是影响上网体验的两大因素。带宽指的是在单位时间内，网络最多能传输多少数据。网络的带宽和水管的粗细有着很多相似之处，水管越粗，相同时间流出的水越多；同样地，网络带宽越大，相同时间能够传输的数据越多。在日常上网过程中，大的带宽能带来这些好处：上传、下载文件的速度更快在线看视频时，缓冲的时间更短如果家中有多个人同时上网，大的带宽能够保证每个人都能流畅上网，相互的影响更小对于家庭宽带，我们常常听到的 10 兆、50 兆、100 兆、千兆，就是指的带宽的大小。延迟指的是数据从网络的一端到达另一端，所需要的时间。过大的延迟也会对上网体验带来不好的影响，例如：在浏览器中输入网址敲回车后，需要等更长的时间，网页才能加载；视频聊天、语音通话质量变差，甚至能直接感受到声音、画面的滞后；玩在线游戏时，发现游戏变慢变卡。在很多时候，延迟相比带宽更容易被人感知，对网络体验的影响更大。是什么导致了网络延迟？上文中介绍过，Internet 使用的是分组交换。如果大家阅读过计算机网络相关的书籍，应该对分组交换的「存储-转发」机制有一定的印象。简单说，「存储-转发」是指网络设备将收到的报文存储在缓冲区，然后再发送。这是由于网络设备的性能有限，转发报文有速率限制。如果设备短时间内收到大量报文，一时半会儿处理不过来，就可以先报文存起来，空闲时慢慢发送。通过「存储-转发」，能够提高网络设备的吞吐量，更充分地利用带宽。其中，存储报文缓冲区，是一个先进先出的队列。那么，队列的长度应该如何确定？是不是队列越长，能够存储的报文越多，网络的性能就越好？大家可以先思考下这个问题。其实在过去，内存价格比较昂贵，考虑到成本，网络设备的队列不会太大。现在，随着内存价格的下降，网络设备拥有了更大的队列。但是，网络的带宽是有限的。实际的可用带宽，取决于数据传输路径上的最小带宽。这个最小带宽也被称做带宽上的「瓶颈」（bottleneck）。由于瓶颈的存在，如果流量长时间以较大速率发送，大的队列并不能显著提升吞吐量。相反，大量报文滞留在队列中，会带来延迟的增大。这种队列中缓存有大量报文，来不及发送，导致数据传输延迟增大的现象，被称做 bufferbloat（缓冲区膨胀）。DSLReport 提供的测速服务，能够检测网络的 bufferbloat 程度。不过由于没有中国大陆服务器，测试结果不够准确。国外的读者可以尝试一下：DSLReports Speed Test家用无线路由器中的 QoS 功能在「智能队列」出现之前，很多家用路由器已经提供了 QoS 功能。那么，传统的 QoS 如何解决网络延迟、网络拥堵？存在哪些不足之处？在了解这个问题之前，先来熟悉一下网络中的报文，是什么样子的：在网络中传输的报文，一般包括了链路层、网络层、传输层、应用层这四层信息。其中，网络层的源 IP 地址、目的 IP 地址、协议号，和传输层的源端口号、目的端口号，是 QoS 算法中常常会关注的五个信息。这些信息合称「五元组」。同一个应用在同一时间段内，传输的报文中，这些信息一般是不会发生变化的。例如观看在线视频时，同一个视频的所有报文 ，一般拥有相同的五元组信息。所以，通过五元组，我们能够区分不同应用的流量。特别地，一些常见的应用，会使用固定的端口号。例如网页浏览一般使用 80、443 端口；英雄联盟游戏，使用的是 5000-5500 端口 ……所以，当我们想要提升英雄联盟游戏的体验时，在路由器上设置让 5000-5500 端口的报文优先通过即可：How to set up traffic prioritization rules for League of Legends on Nighthawk Pro Gaming router | Answer | NETGEAR Support家用路由器中常见的 QoS 功能，就是通过手动设置 IP 地址、协议号、端口号等报文特征来实现的。这种方式需要用户了解一定的网络知识，且操作复杂。（虽然部分无线路由器内置了常用应用的 QoS 规则，但仍需要一定的手动设置，且无法及时适配新的、小众的应用）当然，除了上文中介绍的内容，QoS 还有多种不同的实现方式。例如：对特定应用（例如 BitTorrent）进行限速按照 MAC 地址指定某台设备的优先级（例如提高 Nintendo Switch 的优先级，使游戏更加流畅）通过WMM、DSCP等机制来决定报文优先级这些机制各有各的特色和不足。例如对应用程序限速的方式，虽然效果明显，但无法在网络空闲时，动态利用剩余带宽来提升传输速度，也难以适应带宽频繁变化的无线网络。关于上述 QoS 机制的技术细节，本文不再详细介绍。如果感兴趣，可以搜索相关关键词，进一步了解。智能队列：通过「全自动」的方式优化网络、降低延迟什么是「智能队列」目前，一部分新的无线路由器，开始有了名为「智能队列」的功能。这一功能一般指的是 fq_codel。fq_codel 包含两种队列管理机制，分别是 Flow Queue 和 CoDel。在下文中，我将简要介绍这两种机制的作用。Flow Queue前文中介绍过，网络中的报文一般都有自己的五元组信息；同一应用发送的报文，在一定的时间段内，五元组信息是相同的。这些信息相同的一系列报文，被称做一条流（flow）。Flow Queue 为每条流单独维护一个队列。例如，如果我们一边在线观看视频，一边进行语音通话，一边玩游戏，Flow Queue 就会为视频播放的流量、语音通话的流量、在线游戏的流量各维护一条队列：正常情况下，在线观看视频需要占用较大的带宽，但对延迟不敏感，用户可以容忍几秒甚至几十秒的缓冲时间。而语音通话和在线游戏占用带宽小，但对延迟敏感。例如语音通话时，如果一个人说了一句话，对方在几秒后才能听到，用户就能明显感知到延迟。如果三种应用共用一个队列，队列中会充满大量视频播放应用的报文，导致语音通话、在线游戏的报文迟迟无法发送，增大了这两种应用的延迟。Flow Queue 为视频播放、语音通话、在线游戏各维护一条队列，并公平地调度三个队列，避免视频播放的报文占用太多时间，从而让在线游戏、语音通话的报文得到及时处理。由于 Flow Queue 确保了应用程序之间的「公平」（fair），所以又被称做 Fair Queue。CoDel我们已经知道，适当长度的队列，能够充分利用带宽、提高吞吐量。但过长的队列，又会导致过大的延迟。也就是说，队列既不能过长、也不能过短。CoDel 所做的，就是让队列维持在合适长度，最大限度地提升网络性能。CoDel 通过计算报文在队列中的停留时间，来判断队列是否过长。具体做法如下：报文进入队列时，在报文上记录当前时间（也就是报文入队时间）报文出队列时，再次获取当前时间（报文出队时间）计算报文入队时间与出队时间的差值，即可得到报文在队列中停留的时间当报文在队列中的停留时间超过阈值，说明队列过长，需要缩小队列长度当队列过长时，CoDel 会直接丢弃出队列的报文，从而减少队列中报文的数量，达到控制队列长度的效果。（这种对报文的丢弃，并不会对网络带来负面影响，具体请见后续章节）可以看出，CoDel 根据队列中报文的停留时间，来决定是否丢弃报文。而不是「一刀切」地将队列限制在一个固定的、较小的长度。这样既能降低延迟，又保留了队列最大的优点：吸收短暂的大流量报文、提升吞吐量。另外，我们的设备，尤其是手机等移动设备，所处的网络环境是复杂多变的。例如手机靠近或远离无线路由器，就有可能导致网络带宽的变化。CoDel 通过计算报文停留时间的方式控制队列，能够更加灵活地适应网络带宽、网络质量的变化。fq_codelFlow Queue 确保了不同应用程序的流量能够公平传输。CoDel 控制队列长度，实现了整体延迟的降低。fq_codel 就是将 Flow Queue 和 CoDel 结合在一起使用的机制。其中，Flow Queue 先将不同的流划分为不同的队列，再为每一个队列使用 CoDel，控制报文的延迟。两者结合使用，共同提升网络体验。fq_codel 具有免配置、天然适应不同的网络环境、实现简单、消耗资源少等特色，越来越多的无线路由器（以及开源的路由器操作系统），开始内置这一机制。参考 KoolShare 上的这篇帖子，作者经过测试，发现打开 fq_codel 之后，在 BT 下载占满上行带宽时，在线游戏依旧保持较低的延迟：可能是家庭路由里最好的 QoS——在线游戏者的福音 - KoolShare（见 8 楼、10 楼、22 楼、25 楼的测试结果）是否需要打开「智能队列」？前面介绍了 fq_codel 的一系列优点。那么，fq_codel 有没有缺点？是否需要我们打开？我的设备 CPU 性能不强，是否可以打开「智能队列」？根据这篇帖子，ER-X路由器在打开 fq_codel 时，最大支持 200M 左右的带宽。QoS (fq_codel/cake) Hardware Recommendation : HomeNetworkin

2025 年度旅行——4 天 3 晚青海湖小环线主作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥联合作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥05/14 09:38Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。一次完美的旅行，对我而言，不仅仅是去被人走过千遍的地方打卡一圈，而是藏着三次截然不同的体验——出发前对目的地的想象与期待，旅途中与自然的亲密接触，以及归来后翻看照片时，重新发现那些当时未曾留意的细节，细细回味，查阅资料，让我对这次旅行有更深刻的理解。那么就让我们随着这些文字与照片，回到美丽的青海湖。契机：作为刚工作的年轻人，在国内的职场环境下，真正无理由可申请的假期只有 5 天，所以我和身边的小伙伴们都分外珍惜这五天的时光，每年的年度旅行我们都会避开人流，详细规划。由于十一假期要去参加两场婚礼，为了节约交通费，我们决定先开车到 2 号参加婚礼的宝鸡，把小 P（我的小鹏 MONA M03，具体可参阅我之前的文章）停在了宝鸡高铁站，考虑旅行时间较为紧凑，且西北自然景区交通补能条件可能较差，我们选择了到当地租车完成青海湖小环线行程。带着宝鸡擀面皮和肉夹馍，一行人正式开始了旅途，三个半小时后，到达了西宁火车站，初入高海拔地区，并未感受到明显不适，方便的是，租车平台在西宁站地下停车场就有取还车点，我们直接将行李放上车，省去了搬来搬去的劳顿。Day1 西宁-茶卡 全程约 457 km西宁-塔尔寺5A 级旅游景区，门票：70 元，游览时间约 2 小时塔尔寺是藏传佛教格鲁派的著名寺院之一，始建于明代，最初是为纪念藏传佛教格鲁派创始人宗喀巴大师而修建，据传宗喀巴大师出生于此地，后人于其诞生处建塔纪念，塔尔寺因此得名——「塔尔」即「有塔之地」。塔尔寺-八宝如意塔游览中心和寺庙门口都有大量的讲解人员，价格也不贵，但是本人信仰唯物主义，故只将这里当做名胜古迹参观，故没有请讲解，如果对藏传佛教及寺庙背后的故事比较感兴趣，请个讲解或许是更好的游览方式，或者是像我一样，回来翻阅照片时再仔细搜索相关资料。Tips：讲解会介绍各个佛像的「作用」（不知道这样叫是否合适，或者说负责保佑的事情），故可能会被架在那里为了讨好彩头而「必须」捐献资金，介意者可以像我一样自由游览。由于是节前，故参观人员并不很多，与前往朝拜的信徒或僧人可以说是一半一半，没有了熙熙攘攘的人群，也得以让我更加细致地观察这座寺庙和「磕长头」远道而来的人们。前来朝拜的僧人进入寺庙，浓郁的香火味和木质结构因缺乏通风的陈味扑面而来，看着宏大的佛像、墙上摆放整齐的诵经坐垫，的确能感受到不一样的文化氛围，这里安静、鲜活，尤其是看着不知道从哪里启程的「磕长头」的信众，虽因路途的遥远衣衫褴褛，但想必也经过这一路的艰辛而悟出了些许道理，可能出发时也不知道为什么，但走着走着就明白了吧。塔尔寺中的白塔募捐处，与时俱进地，也支持扫码支付塔尔寺有大大小小的寺庙很多间，在我看来有的都别无二致，于是重点转了几个后，就启程回西宁吃饭，整备氧气瓶，应急食物等装备，正式奔赴青海湖。海南藏族自治州共和县 南山观景台 & 青海湖二郎剑景区前者免费，后者为 5A 级旅游景区 门票 72 元，游览时间约 2 小时一点半在西宁市吃完炕锅后，我们开始驱车前往青海湖，约 2 个小时车程，我们到达了本地的主要目的地青海湖的南侧景区 青海湖二郎剑，在进入二郎剑景区前，可以导航到「南山观景台」 从山上远眺青海湖和远处群山的全貌，但是山路较窄且有工程车辆时不时通过，故要格外注意安全。景区本身没有人文景观，但其还是我国早期鱼雷的试验基地，但是因为我们到达时间比较晚，并没有参观博物馆，而是直奔到湖边。青海湖是我国水面面积最大的咸水湖，水面平静辽阔，与远处的群山相辅相成，甚是开阔美丽，栏杆上有硕大干净的鸽子驻足，在几番伦理拍照后，我靠在栏杆旁，看着远处的夕阳洒在地面，耳边传来阵阵鸟鸣与风声，工作与生活的喧嚣远离脑后，只剩下全身心的与自然接触。云层、远山、草原，形成了震撼的三明治拨云见日，青海湖边的草还是绿色的岸边排队等待「薯条」投喂的海鸥码头景观随着夜幕降临，我们离开二郎剑景区，沿着青海湖开车前往茶卡住宿，虽然只有 200 多公里的行程，但是因为沿途工程车辆很多，还有很多地方持续修路，有的路段有连续的 U 型弯，还有双向仅有一条车道通行的地方，还要翻越几个海拔 3800 米的垭口，成了这一路最惊险刺激的行程，一路上小心翼翼，战战兢兢，终于在晚上 10 点到达了茶卡镇，幸运的是镇上的餐馆都还在营业，找了家当地美食，简单吃饭后休息。夜幕降临的倒湖茶公路Day2 茶卡-祁连县 全程约 372km茶卡盐湖4A 景区 带往返小火车套票 167 元「茶卡」在藏语中意为「盐池」，该地以出产「大青盐」而闻名，参观盐湖也是此地的知名旅游项目。茶卡盐湖的景区共有两个，天空之镜是较老的景区，主要以自然风光为主，天空一号主要以商业化项目为主，作为外地游客，我们肯定更希望看到原始的自然景观，于是选择前往天空之境景区。景区很大，如果想要游览每个景点建议留出一整天的时间，游览顺序建议先坐小火车到终点天空之镜，该站处于盐湖的最深处，盐水异常的干净，也最为出片，景区也放置了很多拍照打卡点位，可以尽情拍摄，心痛没有带无人机。在晴天的时候，由于天空和地面都是白色，很容易晃眼，推荐大家一定要戴墨镜前往，通过墨镜观察湖面，盐湖会看起来像翡翠一样更加绿色透明。套票包含鞋套租赁，可以套上下湖走走，踩在大颗的盐里,很像沙滩的感觉，景区维护的十分干净，可以尽情的亲近自然。有趣的是，景区为了怕大家在盐滩大片的白色中无聊，设置了很多拍照用的摆件，例如国家地理的相框等等，还在大屏幕上贴心的轮播拍照指导，期望每个游客都能满意而归。盐湖与远处的山峰盐湖码头废弃的采盐轨道之前采盐的轨道深入盐湖，在墨镜下，真实的观感颜色会比照片深很多。接驳小火车行驶在盐湖中间，干净得不像国内巨幅盐雕从盐湖出来，我们在县城稍作休息，开始启程前往祁连县，有了昨夜惊险的行程后，尽管沿途的国道可能风景更好，但是我们选择了高速较多的线路。这一程的路况还是很好的，就算是国道车辆也较少，道路笔直，视野十分宽阔，草原、远山、白雪、天空层层分明。在行驶过程中，不时的有大群的羊羔和牦牛穿过公路，不必惊慌，一会就有牧民骑着摩托驱赶，让它们加快穿越的脚步。公路与群山草原上大批的牛羊傍晚的草原，远处能看到夕阳洒在雪山上落日下的牦牛剪影在夜色中，我们抵达了祁连县，在县城中吃过一顿特色烧烤后，进入梦乡，准备迎来第三天的行程。Day3 祁连县-岗什卡雪山-西宁 全程约 320km祁连卓尔山景区门票 60 元 上山小巴 20 元 登山电梯 10 元这个景区的确是全程我最惊艳的景区了，近能看到瑞士小镇似的田园风光，中能看到美如画的丹霞地貌，远还能远眺远处的阿咪东索雪山（海拔 4695 米），而且在山顶可以拍跟雪山的合照，大气、干净、巍峨。自驾车辆可以停在景区停车场，然后乘坐接驳中巴约 10 分钟到达检票口。卓尔山的两侧，风景差别迥异，一侧是连绵的雪山，另一侧是仿佛瑞士小镇的丘陵和小镇。阿咪东索雪山卓尔山的另一侧，仿佛欧洲小镇丹霞地貌与草地交融最让我惊讶的是这个景区竟然还修建了方便爬山的观光电梯，10 元一位，大概有三段，果然是人性化啊。景区电梯中午约 12 点，我们离开景区，在祁连县的一家川菜馆简单用餐后，我们驱车开始前往本次旅途的最高海拔岗什卡雪山。在县城开车的过程中，等红灯时，我看到一位初中年龄的藏族小伙盯着我们开的岚图梦想家，眼中闪闪发光，不禁让我想起小时候的自己，看到喜欢或没见过的车也会驻足观看，充满憧憬。愿他长大后也能开上自己喜欢的车，能够做自己喜欢的事情。岗什卡雪峰景区免费 上山摆渡车 25 元岗什卡雪山坐落于青海海北州门源县境内，是祁连山脉东段的最高峰，被誉为「雪山之尊」，因为设施和难度都尚可，也被雪山爱好者们称为「人生第一座雪山」。主峰海拔 5254.5 米。游览模式同样为将车辆停在景区停车场，然后乘坐接驳车到雪山脚下，正式游览区域的海拔在 3800 米左右，建议随身携带一个氧气瓶，下车之后，可以看到湛蓝的连心湖，水来自于冰川融水，干净，清冽，与子洲植被的荒凉的深褐色形成鲜明的对比，十分美丽。连心湖到达时天气并不是很好，岗什卡雪峰被薄雾笼罩天然泉水，因还有行程，害怕拉肚子，没能如愿品尝一下离开景区后，雪山逐渐露出真面目，攀登还是留给下次吧沿着冰川的遗迹向上走，就能看到岗什卡雪峰徒步攀登的起点，景区也提供马匹，可以跟着向导继续向前，到七彩瀑布，因时间和个人体能的考量，我们浅尝辄止，在河道附近游荡些许时间后，便下山返程。临近傍晚，启程返回西宁市，一顿丰盛的海底捞过后，快乐的青海湖小环线结束。Day4 西宁市-青海藏文化博物馆-特产采购最后一天，伴随着蒙蒙细雨，我们打车前往青海藏文化博物馆，博物馆不大，纵使是十一期间人也不是很多，但内容着实丰富，满墙的唐卡和在电影里看到的藏传佛教僧人服饰，还有精致的贵族金器。在博物馆的文创区，我购入了一个小的转经筒，在回程的路上一直把玩，虽不信奉佛教，但也觉得甚是有趣。唐卡四王金杯嵌绿松石立凤金头饰博物馆文创-转经筒，38 元虽不算划算，但做工十分精致离开之前，为了给家人和朋友带些纪念品，我们选了一家超市类型的纪念品商店，购买了牛肉干，藏红花，枸杞，青稞酒等特产，虽说也不一定划算，但胜在明码标价，将行李箱装满后，顺利返程。后记作为 2025 年的年度旅行，本次 4 天的青海湖小环线我甚是满意。人少不堵车，行程不特种兵，在当地感受美食和壮美的自然风光，一路上与好友们分担开车，聊对景色的感受，互相拍照，点评食物，虽偶有几个工作电话烦扰，但总体悠然自在，放空身心，一切都是那么的安静，与自然深度的接触，与北京的喧嚣形成了鲜明的对比。对于出行方式，自驾无疑是最有的选择，这边深入自然，道路虽崎岖，但也属于铺装路面，且西宁市神州、一嗨等在高铁站都有网点，取送车方便，且自驾可以自由控制行程，根据游览时长方便自己调整。对于住宿，我们对酒店的干净程度略有要求，于是全程都选择了以华住会为目标的连锁快捷酒店，以下是所有天的酒店名称和十一节前的价格，个人觉得是十分舒适的。Day0 桔子西宁火车站省人民医院酒店 双床房 331元Day1 星程茶卡盐湖酒店 覆氧双床房 314 元Day2 宜必思尚品海北祁连卓尔山酒店 421 元Day3 桔子西宁火车站省人民医院酒店 双床房 412 元关于本次行程的花销，纵使算上从北京出发的路费，人均也仅 3k 左右，可以说是十分划算了，主要的开销就是租车和酒店。支出一览表关于卫生间，景区和路途服务器的卫生间都比较完善，因同行有女生，我们还准备了移动卫生间，但从我们的行程安排走下来，并没有用到，个人觉得要是按我们的行程游览也无需担心。下一站，出发！2025 年的年度旅行结束，2026 年也可以开始规划起来了，新疆还是海南，希望能远离喧嚣，享受美丽壮阔的祖国风光。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀2016扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@北鸮© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#Matrix首页推荐#旅行20等 20 人为本文章充电扫码分享举报本文章别惹小炸毛一个学酥关注

打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试主作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.联合作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.前天 15:00ARG 即侵入现实游戏（Alternate Reality Game），是一种基于现实世界，融合各种多媒体（如网站、电话、社交媒体等）技术的互动式解谜游戏。由于其与现实的高度结合所带来的沉浸感，一直以来被许多玩家所追捧。如果你在 2016 年玩过《守望先锋》，相信你对当时新英雄「黑影」的登场一定记忆犹新。从 2016 年夏季开始，一直到 11 月的暴雪嘉年华，暴雪通过在宣传视频、官网、社交媒体发布了一些隐藏着密码的内容，由全球玩家共同解谜，不但揭示了游戏世界观下的一些故事，更是顺理成章地将身为黑客的「黑影」带到大家的视野里。经过几个月的发酵，在暴雪嘉年华上大屏幕被「黑影」控制，随后开始播放英雄动画的时候，更是满足了玩家们期待已久的心。「黑影」的宣发是一场相当成功的 ARG，时至今日仍有玩家对其念念不忘。我本人也是 ARG 的爱好者，我非常享受这种打破虚拟与现实边界的游戏乐趣。前段时间，一款名为《大学生登山失踪事件》的网页解谜游戏突然火了起来。比起此前一些高度依赖技术知识的网页解谜游戏，《大学生登山失踪事件》在不需要任何 IT 知识背景的情况下，依然有着抽丝剥茧逐渐接近真相的推理爽快感。自此之后，小红书上越来越多的人开始制作这一类网页解谜游戏，并为它们赋予了 ARG 的名号。这些游戏流程往往不长，差不多 1-3 小时就能通关。因为是网页解谜游戏，只需要一个浏览器就能游玩，很适合上班时间摸鱼，游玩门槛非常低。这篇文章将会推荐一些这段时间我玩过的质量不错的游戏，虽然一些制作精妙的游戏确实带有 ARG 的影子，但其本质上仍是基于网页的解谜游戏。为了方便大家后续搜索，本文会将这些游戏统称为 ARG。大学生登山失踪事件🔗https://missing.shiroki-y.top/introduction.html在游戏中，玩家需要通过从弟弟衣物中发现的「福安菜家常馆」的线索为开端，逐渐深入一场黑色阴谋找到弟弟失踪的真相。游戏整体难度不高，但在游玩过程中最好能用纸笔或软件记录进度，这能够帮你清晰地理出游戏脉络。游戏的核心玩法其实非常简单，只需要阅读页面，找到页面中可疑的关键词，在页面搜索栏里进行搜索即可推进剧情，这其中也会有一些用户名、密码、口令、字谜等需要玩家进行推理和猜测。网页页脚处会标记当前页面进度，既可以方便定位，也能够明确线索的先后顺序。剧情方面，《大学生登山失踪事件》并不会在一上来就给玩家塞一些晦涩难懂的概念，而是从大家熟悉的普通网页开始。当玩家一步步抽丝剥茧逐渐发现事件背后的黑暗面时，令人不由得感叹于游戏的设计与巧思。同时，游戏内也藏有一些可供探寻的 ex 页面作为彩蛋，在解谜之外找到各种彩蛋也不失为一种乐趣。诡异广告调查事件🔗https://mp.weixin.qq.com/s/yeNYMZVRp6N-MESFJGk9rg这是一款与《大学生登山失踪事件》来自同一作者@銀河電燈_的游戏，世界观也与前作有关联之处，很是有趣。玩家将扮演一位被「优化」的人，因为在看剧时发现了某个已停播的节目广告，从而开始调查这个诡异的广告，并发现藏在节目与自己背后的神秘真相。与前作相似，这款游戏也将基于阅读与搜索推进剧情。在此基础上，它还加入了一些 ARG 元素，比如在街景地图中，找到现实里存在的文字招牌作为线索。并且在网页设计上也比前作更加成熟，贴合剧情上的时间设定，在网页上加入了许多彩蛋。剧情方面，本作有一定的惊悚元素。特别是在游戏过程中展示的一些内容可能会引起不适感，如果你对这类内容比较敏感，请务必提前做好心理准备。当然，游戏并没有 Jump Scare，更多是通过剧情与视觉设计所营造出一种的诡异氛围。三川米粉店🔗https://www.berrysignalscmfd.top/《三川米粉店》也是一款从餐饮店主页出发，探寻故事真相的游戏。玩家需要调查三川米粉店突然停业的原因，在调查过程中，玩家将逐步揭开这家米粉店背后的故事。本作的页面设计相对没有那么精致，但也融入了一些 ARG 的元素，部分谜题的答案并不会直接给出，而是需要玩家通过搜索引擎或其他工具进行查询才能获得答案。从剧情上来说，《三川米粉店》的设定会更加接地气一点，并且结局也会相对更温馨一点。比较适合对于惊悚元素不太感冒的玩家作为第一部作品来尝试游玩。永安温泉度假酒店🔗https://mp.weixin.qq.com/s/kKsnzFsLFwwm2nwv8DseAA与前面侧重解谜的作品不同，《永安温泉度假酒店》的解谜难度其实不高，反而更像是一种「规则怪谈」式的探索路径。玩家需要不断探索找到更高权限的员工账号密码，从而揭开温泉度假酒店背后的秘密。游戏在开始前需要玩家填写「报名信息」，在后续游戏中能够增强代入感。从游戏性的角度来看，这款作品的「解谜」爽快感或许没有那么强。但如果你对「规则怪谈」或者一些高概念的设定感兴趣，那么这款游戏依然是一个值得尝试的作品。404 工位消失事件🔗https://xsn1204.github.io/qmgame《404 工位消失事件》是一个特色鲜明的游戏，玩家将扮演一个实习生，接手其他人留下来的工作电脑，并从中发现一些未被处理的秘密。游戏在网页上复刻了一个形似 macOS 的「系统」，很大程度上增强了代入感。因为设定是在电脑系统，游戏中需要玩家在钉钉、微信等各种软件中挖掘线索，反而更有一种作为侦探在不断探索的感觉。但受限于网站形式和游戏体量，游戏中能够互动的地方通常都是为剧情线索服务的，与线索无关的地方大多都是不能点击互动的。而且游戏不能通过点击窗口左上角的关闭按钮回到上一层，而是需要通过浏览器的返回来实现。但这些小细节对于《404 工位消失事件》来说都算瑕不掩瑜，整体游玩体验还是值得肯定的。作者甚至还出了 2.0 版本，新的剧本进行了优化提升，如果对这一类型感兴趣的话也可以尝试游玩。找到我🔗https://www.kaixinmoyu.com/feijo/findfeijo/在游戏中，玩家需要扮演一位私家侦探，调查游戏作者失踪的原因。游戏在交互方面做得非常有意思，游戏不仅做了浏览器页面对线索进行搜集检索，还做了小灵通和笔记本，前者可以通过拨打特定的电话获取线索，后者则需要玩家阅读日记找到隐藏的内容。《找到我》本身是一个更重交互而非推理的游戏，游戏中融入了一些 ARG 元素，在最开始就需要到某篇公众号文章中找到一个密码才能继续推进。但游戏的剧本逻辑会稍微有些不够连贯，以至于中间可能会有一些容易卡关的地方，需要仔细翻阅此前出现的线索才能继续往下推进。另外，游戏中后期会有一道需要在现实中解算的谜题，这道题本身与游戏本身的节奏设计会稍显割裂。但整体来说，《找到我》依然是值得一玩的游戏，尤其是本身的交互设计和 ARG 元素的融合做得非常不错。在游戏作者的个人主页，也能找到作者开发的更多网页解谜游戏，如果你对这款游戏的风格感兴趣，也可以尝试其他作品。小结优秀的网页解谜游戏并不止文章提到的这些，光是在小红书上一搜，就能找到许多分享帖。希望本文能起到抛砖引玉的作用，欢迎大家在评论区分享自己玩过的，令你印象深刻的网页解谜游戏。> 简单、好用、专注的写作软件，少数派为你呈现 🚀> 守护孩子的天马行空，从零开始美术教育启蒙🎨538扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Lotta© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#解谜游戏#游戏#生活方式53等 53 人为本文章充电扫码分享举报本文章TpBe nothing, and everything.关注

派评 | 近期值得关注的 App主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注Vanilla少数派作者Vanilla关注Vanilla少数派作者关注化学心情下2少数派作者化学心情下2关注化学心情下2少数派作者关注大大大K少数派作者大大大K关注大大大K少数派作者关注Peggy_少数派作者Peggy_关注Peggy_少数派作者关注ElijahLee少数派作者ElijahLee关注ElijahLee少数派作者关注Snow少数派作者Snow关注Snow少数派作者少数派编辑部 等 7 位作者前天 18:05欢迎收看本期《派评》。你可以通过文章目录快速跳转到你感兴趣的内容。如果发现了其它感兴趣的 App 或者关注的话题，也欢迎在评论区和我们讨论。AIDA：一站式日历和待办规划工具平台：iOS、iPadOS关键词：日程管理、待办清单@ElijahLee：Aida 是一款针对 iPhone 和 iPad 开发的日历和待办清单应用，来自任务管理工具 Sorted 的前创始人之一 Harry，它集成了待办清单、日历和时间规划工具，在一处显示、规划、操作。应用主界面就是每日时间线，顶部是当周 7 天的简要信息，应用中部大篇幅排列当天的时间块，你可以直接看到上午十点看电影、中午写派评、傍晚健身等日程，以及晚上记得给植物浇水的待办提醒，而不用在多个不同的 App 之间切换。应用的快速调整功能，支持在时间线中将事件按住并移动顺序，快速重排一天的事件，对应的时间点会自动修改，非常方便。使用右下角添加按钮，可以自由地为时间线加入行动或事件。行动即待办事项，支持输入文本并设置提醒时间。 Aida 创造性地设计了拖动设置时间的功能，将底部的柱状图左右滑动，就可以调整行动的提醒时间，每次拨动以 15 分钟为间隔，用它设置时间非常方便。当然我们也可以点击时间来精细设置时分。 Aida 事件就是日历中的事件，必须包含开始时间和结束时间，我们还可以选择日历类别、设置全天事件等。Aida 引入了 GTD 管理的项目功能，提供收件箱、任何时间、将来某日等默认分类。自定义项目可以将同样类别的活动，包括行动和事件进行集中收纳整理。Aida 全面集成了 iOS 日历，在应用内不仅可以查看 iOS 日历事件，对事件的修改调整还会实时同步。不过 Aida 目前没有接入 iOS 的提醒事项。 批量操作功能可以一次选择多个行动进行完成、改期、移动或复制，也能快速整体调整时间块。你可以在App Store免费下载 Aida，免费版功能受到很大的限制，仅支持 1 个日历的集成显示、无法调整行动的时间、添加行动和事件的数量有限。付费可以解锁高级功能，包括无限项目和时间块、多个日历集成显示等功能。费用为 58 元 / 月、298 元 / 年，698 元买断，应用提供 1 周免费试用。我的生活配不上这个定价我以为待办和清单类 app 最重要的是视觉和交互吧 这令人一眼没欲望的 AI Generated UI 怎么还敢定价这么高毫无特色 + Sorted 血脉 + 令人疑惑的定价现在任务管理之用自带的提醒事项和滴答清单免费版定价没问题，是我的问题。全部讨论 7发表评论Cogito：连接人与 AI Agents 的 Markdown 编辑器平台：macOS关键词：Markdown、AI@Vanilla：对于 Markdown 编辑器这一类应用，我想大家已经见得太多了。随着 AI Agents 的兴起，Markdown 这个文件格式可谓是「老树开花」，又重新获得了很多人的青睐。Cogito 是一款面向人和 AI Agents 共用的 Markdown 编辑器，它不像 Obsidian 一样偏知识库、不像 Typora 一样偏所见即所得、不像 iA Writer 一样偏写作，也不像 VS Code 一样偏开发者，它最大的特点是「文件优先、Agents 友好」。下面，就让我们一起来看看这个新时代的 Markdown 编辑器到底有何不同。Cogito 的文件库完全基于本地磁盘，没有专属格式，没有加密拦截，没有同步机制，所有的文件和文件夹组织形式都和 Finder 中的一模一样，完全映射，没有任何上手门槛。以我的使用习惯为例，我直接把 Codex 的工作文件夹设置为了 Cogito 的 Location，它会自动抓取 Codex 工作区的所有文件夹和 Markdown 文件，按照日期排序，清晰明了。选择任意一个 Markdown 文件后，Cogito 默认展示它的编辑状态，我们可以随意地按照自己的意愿进行编辑。点击右上角的第一个按钮，则可以在编辑状态和预览状态之间进行切换，对应的键盘快捷键是 Command + R。我经常会借助 Codex 帮我生成一些行业或者公司的研报，最终以 Markdown 格式来进行输出，通过 Cogito 我就可以非常方便地进行编辑和阅读，不需要反复地在 Codex、Finder 和 Markdown 编辑器之间进行来回切换。点击右上角的第二个 Outline 按钮，我们可以快速查看当前文档的大纲，点击标题可以直接跳转到相应的章节。Cogito 除了对 AI Agents 友好，还可以在应用中直接调用本地的 AI Agents 对应用内显示的文档进行总结、分析、编辑等操作。总的来说，Cogito 对于像我这样频繁使用 AI Agents 和 Markdown 文档的人来说是一个非常不错的 Plan B，和主力写作的 iA Writer、用作知识库的 Obsidian 等进行区隔，打开即用，用完即走，非常高效。Cogito 可以在官网免费下载使用，目前处于公开测试阶段。好吧，不能所见即所得。agents友好其实有点尴尬，一个阅读器+一个terminal就可以完全的代替掉这个软件目前的全部功能？甚至codex的app自身也有文件阅读功能了。发表评论AndDrive：无缝连接你的 Android 设备和 Mac 电脑平台：macOS关键词：文件管理、手机管理@化学心情下2：虽然现在有不少工具能让 Android 手机和电脑无线互传文件，但「传文件」和「管理文件」是两回事。比如想批量删除 Android 内部存储 Pictures 文件夹下的截图，用传输工具就很别扭，这种场景更适合在电脑端的文件管理器里完成，可以借助筛选和预览功能高效处理。Windows 上这件事不需要任何第三方软件：数据线一插，选择「文件传输」，Android 就被模拟成 MTP 设备，像移动磁盘一样直接管理即可。但 Mac 上就麻烦了：MTP 是 Windows 的专有协议，macOS 想访问 Android 内部存储，要么走 USB 调试，要么依赖第三方 MTP 实现。OpenMTP 算是其中比较知名的一款，但项目两年多没更新、稳定性一般，而且用的是自己写的文件管理器，而非访达，体验称不上顺手。那么有没有能直接用访达管理 Android 文件的工具？接下来要介绍的 AndDrive 就是，它由个人开发者开发，没有走 MTP 路线，而是基于 Android 官方的 ADB 调试通道实现文件管理。安装后，AndDrive 常驻菜单栏，支持 USB 和 Wi-Fi 两种连接方式。使用前需要在 Android 上开启 ADB 调试并授权当前 Mac，再按提示在手机上安装配套的 AndDrive Helper，之后访达里就能看到 Android 的目录了。AndDrive 的挂载方式可以选 SD 卡目录、根目录或不挂载，看到的内容会不同：SD 卡目录只显示内部存储，根目录则多一个系统根目录（没 root 的话很多文件依然动不了）。日常使用可以直接在访达里预览文件、按时间筛选，也能把电脑上的文件拖进 Android 对应文件夹，比在手机自带的文件管理器里翻找直观得多。因为走的是 ADB，AndDrive 顺带还实现了屏幕镜像功能。点击 AndDrive 的「屏幕镜像」，当前 Android 设备的桌面就能投屏到 Mac 上，体验和 macOS 的 iPhone 镜像类似，可以直接操作应用、用中文搜索，还支持剪贴板同步、录屏、截屏等。比起手敲 Scrcpy 命令，门槛会低不少。总的来说，AndDrive 虽然没有用更成熟的 MTP 协议，但用 ADB 这条 Google 官方推荐的方法，让访达管理 Android 文件这件事变得简单可行，还补上了类似 iPhone 镜像的屏幕控制功能，值得有需求的你一试。你可以在App Store免费下载 AndDrive。有没有苹果手机和windows电脑互传的它不是改名了吗？🤔发表评论Fauxx：与其被动防守，不如主动出击的隐私保护工具平台：Android关键词：隐私保护@大大大K：你的每一次搜索、每一个网络请求、每一段定位轨迹，都在被数据平台和广告商默默收集着。它们会通过这些碎片信息拼凑出你的身份、喜好、习惯甚至是经常交往的人，然后在暗地里交易你的信息、精准推送广告。这也是为什么你或者你的朋友随手搜了点什么，10 分钟后所有 App 和网页就开始围着这个话题给你推广告。因此市面上衍生出不少隐私保护工具，它们的思路大多都是「堵」——屏蔽数据上传、拦截 Cookie、混淆 IP 等等。但 Fauxx 这个 App 啧另辟蹊径：既然广告商一定要收集数据，那就把真实的你混淆在一堆「垃圾」当中。首次打开 Fauxx 会有一个引导流程，你可以填写你的年龄段、职业、兴趣之类的粗略信息。这一步主要是为了让 App 能够生成与你真实信息相反的混淆信息，当然，这一步你也可以选择跳过。Fauxx 并不强制你输入信息。完成后进入主界面打开开关就能启动混淆信息的保护，在这里你可以看到这段时间 Fauxx 执行了多少次「混淆」动作、当前的虚假人设是谁、混淆了哪些类型的信息。至于具体可以混淆的内容，可以在 Targeting 和 Modules 标签中设置混淆等级和类型。Targeting 里有三个可选的层级：第一层就是生成与你真实信息相反的混淆噪声，如果你一开始没有填写，Fauxx 则会随机生成；第二层是对抗广告商的用户画像，不过我实际测试下来会读取失败，但也不影响使用；第三层是「人设轮换」功能，Fauxx 通过生成虚假的身份信息，让广告商无从追踪。Modules 标签里则是七个具体的混淆模块，每个都可以单独开关。设置里还可以调节混淆频率、指定时间段等。那这些「假数据」是怎么骗过广告商的？通常来说，被安插在手机里的追踪器会根据设备 ID、IP 地址、DNS 等信息追踪这台设备上的数据访问，进而推测用户画像。而 Fauxx 会通过模拟搜索、链接访问等「类人」操作（比如搜索时的停顿、假装点击广告），规避机器人检测并成功让追踪器追踪到假的信息访问。Fauxx 可以说是转变了隐私保护的方向，从被动防守变为主动进攻。如果你也想换个思路给广告商添添堵，可以在GitHub免费下载 Fauxx。这软件在中国没用吧1224Roll：胶片「日日鲜」平台：iOS关键词：手机摄影、胶片模拟@Snow：当下，你用手机拍照真的可以发散出无限可能，不同的镜头、不同的滤镜、不同的手动参数、不同的画面调节项，甚至是不同的外设，只要你愿意钻研一番，就能拿捏各式各样的风格。不过，更多的选择往往意味着更多的「选择困难症」，如果你对于丰富的选择产生了一些焦虑，或者天性就不爱「折腾」，不妨看看 Roll 这款应用，它或许可以带你发现简单拍照的乐趣。打开 Roll，完成简单的授权流程后，你就会收获今日份的新胶卷。应用会简单阐述这份胶卷在感光度、对比度、饱和度等各方面的特质，有一些呈现阳光下温暖的黄色，有一些闪耀着玫瑰般的光泽，有一些甚至想带你进行「第三类接触」。和一般的应用内随机胶片不一样，Roll 所有用户的「掷骰」结果完全一致，无论你身在世界何处，在同一时刻都会使用同一款胶片进行拍摄，剩余可用时间会在胶片介绍页注明。时间有限，点击「Load film」启动拍摄。Ro

周二19派早报：红魔发布红魔 11S Pro 系列手机等今年的新系统会有什么变化呢？早报速览红魔发布红魔 11S Pro 系列手机5腾讯音乐已完成对喜马拉雅的收购8蓝牙核心规范 6.3 正式发布0腾讯自研 AI 设计智能体平台 Ardot 现已开启公测7微软将允许用户修改实体 Copilot 键功能4中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告0马斯克诉奥特曼案被驳回1不妨一看的简讯4红魔发布红魔 11S Pro 系列手机5 月 18 日，红魔发布红魔 11S Pro 与 11S Pro+ 手机等新品。红魔 11S Pro 系列搭载第五代骁龙 8 至尊领先版与红芯 R4 电竞芯片，机身采用纯平设计，11S Pro+ 还有透明水冷可视化设计，屏幕为 6.85 英寸 144Hz 无挖孔无刘海设计，采用屏下前置摄像头与 3D 超声波指纹解锁，搭载新思触控芯片，支持 3000Hz 瞬时触控采样率。系统内置 PC 模拟器，搭载 3.5mm 耳机孔与红外功能。11S Pro 电池容量 8000 mAh，支持 80W 快充；11S Pro+ 电池容量 7500 mAh，支持 120W 快充，两机型均支持旁路充电。红魔 11S Pro 有暗夜骑士与银翼战神两配色，12GB+256GB 售价 5499 元，国补后 4999 元。红魔 11S Pro+ 有氘锋透明银翼与氘锋透明暗夜两配色，12GB+256GB 售价 6199 元，16GB+512GB 6999 元，16GB+1TB 7999 元。另有红魔碳纤维游戏鼠标、红魔散热器 8Pro、氘锋能量卡 Pro 磁吸充电宝等配件发布。来源红魔还是有点追求的红魔是为数不多的纯平机身了背后是MagSafe 结构？屏下摄像头技术已经成熟了吗，怎么感觉其他家很少用呢，还有什么缺点吗3腾讯音乐已完成对喜马拉雅的收购腾讯音乐于 5 月 18 日在港交所发布公告，宣布根据并购的条款，收购喜马拉雅的事项已经完成。据此喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为 12.6 亿美元的现金及最多总数 1.75 亿股本公司 A 类普通股（包括本公司于交割时新发行的 A 类普通股，以及本公司授出的以股权为基础的激励相关的 A 类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。来源完了 喜马拉雅 也要有低级充值会员 中级充值会员 高级充值会员了买京东会员送的喜马拉雅会员还没用过呢全部讨论 8发表评论蓝牙核心规范 6.3 正式发布蓝牙技术联盟上周末发布了蓝牙核心规范 6.3，作为半年一次发布计划的一部分，此次更新引入了多项新功能，可进一步提升测距精度、扩展接口容量并提高射频效率。本次更新新增蓝牙信道探测内联 PCT 传输、蓝牙信道探测特定物理层 RTT 精度、蓝牙位资源扩展以及蓝牙 ACP 和 C/I 限值放宽。蓝牙技术联盟表示，联盟成员在描述产品支持的蓝牙功能时，应避免直接引用产品所依据的蓝牙核心规范版本号（例如蓝牙核心规范 6.3），而应在产品包装、文档和推广材料中清晰、准确地描述其产品支持的具体蓝牙功能，特别是与用户需求最相关的部分。来源42腾讯自研 AI 设计智能体平台 Ardot 现已开启公测腾讯云于 5 月 18 日发布公告，正式宣布 AI 设计智能体平台 Ardot 开启公测，官方称其为「一个 AI 驱动的产设研协作平台」。Ardot 支持一句话生成可编辑设计稿，AI 可快速生成 App 页面、官网、海报、插画、PPT 等；支持调用团队自己的业务组件库，一键生成符合规范的可编辑稿；支持直接导入 Figma 文件，完整保留原有布局、样式和组件。Ardot 对代码友好，支持设计一键转代码。设计稿，连同变量、组件、布局数据等设计细节数据，直接拉进 CodeBuddy，一键代码还原；也兼容 Workbuddy、Cursor、Claude Code 等 MCP IDE。Ardot 支持多人在线实时评论、标注反馈和版本对比，配备智能权限中心、全方位行为追溯和无缝交接机制，确保协作规范、有序。此外，Ardot 微信小程序也即将上线。来源太难用了能运营超过两年吗？与此同时猿辅导家的妙多 / Motiff 即将停服像素级啊确实太像了，感觉就把配色 从紫色改成了绿色🐶[图片]全部讨论 734微软将允许用户修改实体 Copilot 键功能5 月 18 日，Windows Central 发现微软上线一份新支持文档，表明自 2024 年起笔记本厂商在 Windows 11 设备上预装的 Copilot 实体按键影响了用户的工作流，他们将在今年晚些时候推送一版 Windows 11 更新，允许用户将 Copilot 键重定义为被替代的菜单键或右 Ctrl 键，用户将在设置 > 蓝牙和其他设备 > 键盘中找到相关选项。来源什么时候把win键强制绑定去了，允许用户自定义。终于顿悟了？国内都用不了，按它干什么？微软终于发现了141中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告5 月 18 日，中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告。2026 年 4 月，中国游戏市场实际销售收入 303.69 亿元，环比下降 3.51%，同比增长 11.04%。其中，中国移动游戏市场实际销售收入 224.30 亿元，环比增长 0.71%，同比增长 9.83%；中国客户端游戏市场实际销售收入 68.66 亿元，环比下降 12.77%，同比增长 18.40%。中国自主研发游戏海外市场实际销售收入 20.32 亿美元，环比下降 5.02%，同比增长 30.78%。来源1马斯克诉奥特曼案被驳回加州奥克兰联邦法院的陪审团于 5 月 18 日作出一致裁决，驳回了埃隆·马斯克对 OpenAI 及其 CEO 萨姆·奥特曼等人的诉讼。经过为期三周的庭审，由 9 人组成的陪审团仅用不到两小时便达成裁决，认定马斯克指控 OpenAI 违背成立初期慈善承诺的起诉已超过诉讼时效。2024 年，身为 OpenAI 联合创始人之一的马斯克指控奥特曼、总裁格雷格·布罗克曼以及主要投资者微软公司背叛了该机构「为人类福祉开发非营利性 AI」的初衷，通过转为商业化运作并与微软结盟来谋取巨额私利。马斯克此前曾为 OpenAI 提供早期资金，并在此案中索赔最高达 1500 亿美元。陪审团认定，马斯克早在 2021 年就已经知晓 OpenAI 转向营利性架构的事实，因此其 2024 年的起诉已经超出了三年的法定诉讼时效。辩方展示的证据表明，马斯克自己当年在未能获得 OpenAI 控制权之前，也曾提议过转为营利性结构。此次判决为 OpenAI 今年晚些时候的 IPO 扫清重大法律障碍。马斯克本人则在 X 上发文，批评该判决仅仅基于「时间程序问题」，而并未对案件的实质正义进行裁决。来源老马：不听我控制的就是不正义17不妨一看的简讯据外媒 Wccftech 的报道，根据一份去年泄漏的运输清单中的信息显示，Googlebook 笔记本或将搭载英特尔 Panther Lake 12Xe3 核心处理器。运单的信息显示，该产品使用酷睿 Ultra Series 3 系列处理器（即 Panther Lake），并搭载 12Xe3 图形单元。并且在 Google 与英特尔的相关文件中，有一个相同的代号 Q7AP 出现。而该编号对应 Panther Lake 12Xe 主板，对应的 Chromebook 产品代号则为 Felino。该笔记本采用 16 英寸设计，配备 32GB LPDDR5X 内存、1TB 存储空间。根据英特尔公布的产品阵容，12Xe 应该是 Panther Lake 的高端型号，至高可选 16 核心 CPU，功耗范围 25W-80W。如果此次曝光准确那么 Googlebook 很可能采用酷睿 Ultra Series 3 系列处理器。来源据微博用户金猪升级包，英特尔在 5 月 18 日的第三代英特尔酷睿处理器新品分享会上，介绍了英特尔 Project Firefly，将针对价格敏感的 Wildcat Lake 平台，导入中国本土化成熟的手机产业链，统一标准化接口实现规模化，完成降本增效。现场展示了 Project Firefly 的 A+ 参考设计与第一款 Project Firefly 参考设计下落地的量产产品来酷 Air 14，均为超薄机身与 Clean D 设计。来源Linus Torvalds 在 LKML 的新帖文中建议大家不要用 AI 工具提交重复的 bug 报告，大量重复、过时的报告让邮件列表无法正常收集 bug，想用 AI 工具创造点价值就去好好写补丁。来源Apple 定于北京时间 6 月 9 日至 13 日举行全球开发者大会 WWDC26，6 月 9 日凌晨 1 点开幕时将举行主题演讲，介绍 Apple 平台更新，包括人工智能的进展、新软件和开发者工具；6 月 10 日凌晨 4 点举办 Platforms State of the Union，介绍新功能、API 和技术如何进一步赋能开发者。来源如此直白的话感觉是Linus会说出来的😅Linus都蚌埠住了还人工智能进展，国行的 AI 有没有？意思是Googlebook运行的是x86_64的安卓？发表评论少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购你可能错过的文章四天三夜，用小长假的时间体验青海湖小环线打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试派评 | 近期值得关注的 App> 下载少数派 2.0 客户端、关注少数派公众号，解锁全新阅读体验 📰> 实用、好用的正版软件，少数派为你呈现 🚀71扫码分享目录 0讨论我来说一句发布发表评论发布

YoooClaw C·ONE 体验：为 AI Agent 而生的硬件，怎样构筑了数字化的「我」？主作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛联合作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛昨天 11:04AI 虽然现在具备了干活的能力，但对于很多人来说，它依旧只是个 Chatbot。因为，大多数 AI 依然停留在屏幕里（当然，具身智能另说）：你需要打开手机或电脑，进入某个 App，手动复制消息、描述背景、然后不停修正抽卡，最终把输出的结果复制到你需要的地方去 —— 对很多人来说，这个过程本身就已经足够打断手里的工作。App 的逻辑其实更适合屏幕类设备的交互，这是我们从 Web 2 时代过来的逻辑惯性。但如果我们换个思路，要是一个设备不是为了适应人类和 App 的交互，而是一开始就为 AI Agent 去构建，那它应该是怎样的？YoooClaw 给出的答案是 —— 一张智能卡片。很高兴，这次与少数派和 YoooClaw 一拍即合，提前体验了这款名叫 C·ONE 的智能卡片。 在体验的过程中我意识到，或许那个「数字化分身」的时代，已经慢慢开始变得具象了。AI 时代需要一个「灯条」拿到 YoooClaw C·ONE 的时候，其实我一直在想，手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件呢？但这个问题还不急着说，反倒是上手的第一刻，它的工业设计给了我不小的惊喜。它采用了全金属材质，一体感极强，拿在手里有一种精致感，有点像金属版的 MagSafe 外接电池。而且没想到的是，它的楔形设计与我 iPhone Air 的契合度高得惊人。磁吸上去后，它恰到好处地填充了 iPhone Air 极其轻薄的机身与我掌心之间的空白，给我这个 iPhone Air 裸奔党增加了不少安全感……C·ONE 的背面元素也不多：一个按钮、一个条形指示灯和 YoooClaw 的 logo。长按按钮，灯条变成绿色的波浪，C·ONE 就会进入 AI 对话模式，很像是我拿起 Apple TV 遥控器按下侧键与 Siri 对话。而双击之后，灯条变成绿色的闪烁，就进入了录音模式。这个灯条让我想起了前几天发布的 Googlebook，不同时空，横跨太平洋的它俩居然想到一块去了。该说不说，灯条在 AI 时代居然成了一种「不打扰却又十足惹眼」的交互 —— 它不告诉我具体的内容，而是只告诉我「状态」，简直是我这种 ADHD 人的福音 ——一个真正好用的通知管理器作为一个长期与 ADHD「斗争」的人，一条通知突然让屏幕亮起，我的心思可能就飞出了十万八千里之外。而且有时候好不容易进入了心流状态，突然来一个通知，看吧，我的工作流就断了，不看吧，真是心痒痒。iOS 原生的通知系统虽然有做一些约束，甚至还引入了专注模式，但它有一个很大的问题：它的管理逻辑是基于 App 的。这就有个十分明显的问题：工作群里，我负责的事情结束了之后，与我有关的通知可能就不多了。但我还是不敢关通知，毕竟有些事情还是需要我来处理。而在 iOS 系统中，这些通知的地位是完全对等的（到底是谁那么喜欢在微信聊工作啊！）。它们密密麻麻地挤在一起，要么让我因为厌烦而全部忽略导致错过，要么每次来消息，我就还是得就打断当前的事情看一眼。尽管 Apple 智能后来推出的「通知总结」有所改善，但对于国行设备来说，吃不到嘴里的饼画得再大也没用。即便在已经用上了的地区，用户反馈也不太好。而且，它依然解决不了「拿起手机」这个动作本身带来的诱惑：原本只是想看一眼通知，结果却顺手刷了半小时马斯克的表情包。YoooClaw C·ONE 提供的方案非常符合直觉：它可以读取我的通知，并根据重要程度，给背后的条形灯带设定不同的颜色和效果。最棒的一点是，这种设定不需要像在 iOS 里那样一个一个 App 去操作和设定，只需要在 YoooClaw 里通过自然语言告诉它：「我在微信群工作群 XXX 里的昵称叫宛潼，如果有 @ 我的消息，你就用红色呼吸灯提醒我。」于是，我的工作流变成了这样：当我开始写稿时，在 iPhone 上进入创作专注模式，所有通知调整为静音，并且同步到 Mac。然后，我将 iPhone 倒扣在桌面上，让 C·ONE 成为我唯一的通知接收器。如果灯条没有闪烁红色，那我就完全不用去理它，只有当红色的光亮起时，我才需要拿起手机瞟一眼。这种物理上的「隔离」，很好地保护了我的工作惯性。除此之外，它还有一个类似于「监工」的呼吸灯提醒模式，毕竟我这种人不看到 Deadline 是真的没法开工。现在，我可以设定在某个待办任务截止前的四个小时，让 C·ONE 以橙色的呼吸灯持续闪烁，直到我将任务标记为完成。这种视觉上的提醒虽然会让我感到焦虑，但相比于手机弹窗那种一划就过的提醒，这种持续的物理存在显然更有力量，确实可以有效地督促我赶紧干活，像极了以前十点不到就催我去睡觉的老妈。iPhone 用户的「终极闪念胶囊」随着豆包手机的面世，我确实也意识到，AI 最终会「杀死」绝大多数 App。但一个物理按键，或许在很长一段时间内都是刚需。物理按键代表着一种明确的「告知」。它告诉 AI：我现在要开始与你交互了。无论是 iPhone 上的操作按钮，还是新款机型上的相机控制，其实都是 Apple 为 AI 时代准备的入口。但在 Apple 智能迟迟未能在国内完全落地的当下，这些硬件潜能大半处于荒废状态。所以，C·ONE 上的那个物理按钮，现在就成了我外置的「闪念胶囊」。作为一名自由撰稿人，灵感往往是稍纵即逝且细碎的。很多时候，我脑子里蹦出一个绝妙的比喻，或者一段关于产品逻辑的思考，等我回家坐在电脑前准备动笔时，那种倾泻而出的感觉已经消失了。现在，我只需要长按 C·ONE 上的按钮，它就开始录音。我可以毫无章法地把这些碎片化的想法丢给它，让它帮我记住。当我回到家，这些碎片已经变成了整理好的文字，启动的困难就少了许多。C·ONE 在这里扮演的角色，不再是一个简单的录音机，而是一个「外置记忆存储器」。例如，在写完一篇关于 AIGC 内容的稿子，给自己做咖啡的时候，我偶然想起来里面那些近乎完美的人物图片，于是就对 C·ONE 说出了我心底被唤起的容貌焦虑；而在看完《给阿嫲的情书》之后，我也可以立刻在回家的散步的路上就对 C·ONE 说最直观的观后感，为日后可能撰写的影评积攒想法。上下文：构筑「我」的养料到了这里，就是时候来回答最开始的那个问题了：手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件？在写这篇稿子的时候，我听了 YoooClaw 创始人肖睿哲和王自如的一场播客。老肖分享了一个很有意思的案例：我们有一个同事，他平时淘宝有大量的消息。他会在淘宝上买药，然后我们的产品就会识别到一个有趣的现象：比如他今天要来深圳出差，然后 AI 就会告诉他说你别忘了带你的降压药。这个降压药在配送过程中，淘宝有几个 Push 会带着商品标签和名字，而这些商品标签的名字被 AI 识别到了以后，他就会去自己学习这是个什么东西，然后形成了你某一类的画像。这个例子给了我很大的启发。过去我总觉得，大部分通知是骚扰，所以很多通知我是手动关闭的。但在 AI 看来，通知是极佳的上下文，经过用户筛选的通知更是优质的上下文。而且与人类会被通知打断不同的是，它可以不知疲倦的去读取。这些碎片化的信息，以及我每天录入的创作想法、我与它的对话，最终将会凝练成属于我的「宛潼.skill」。其实我一点都不反对蒸馏我自己，我也不担心 AI 会抢走我的饭碗。我甚至还在朋友圈发过一段话：「AI 要是什么时候真能帮我写稿那我可开心死了。我都不担心我的工作被抢，相反我可以用 AI 一天写十篇。」因为，AI 模仿的是我的「Skill」：我的语气、我的写作手法、我的写作风格……这需要大量的上下文去喂养；而真正的「Idea」源自于我特殊的经历、我今天与朋友的对话、我傍晚在河边晒着太阳发呆突然闪过的瞬间……这些是 AI 不可模仿的。如果有一天，我的数字分身足够成熟，它能完美捕捉我的表达方式和逻辑链条，我只需要负责抛出 idea，由它完成初稿，我再进行审稿和修改，那我的生产力将获得质的飞跃。这就是老肖在播客里说的，设备之所以有价值，是因为它沉淀了数据。就像在白板前讨论，结果固然重要，但「谁画了那个圈」、「大家是怎么讨论出这个结果的」这些过程数据，才是 AI 进化所需的真正养料。真的聊得蛮好的，推荐以及，我还有一个角度 —— 虽然很多人对我的印象是科技博主 or something，但我其实是个中文系毕业的纯文科生。我是凭着对数码产品的热爱，才自学了很多代码逻辑和工程师思维。但对于绝大多数普通人来说，AI 工具的门槛还是太高了。无论是 OpenClaw、Hermes 还是后起之秀 Codex，对于没有计算机思维和 AI 基础的大众来说，配置起来依然让人头大，就更别提使用了。当年 Apple 的 Lisa 带着图形界面问世，才真正砸掉了计算机与普通家庭之间的那堵高墙。所以在我看来，AI 也在等待属于它的「Lisa 时刻」。YoooClaw C·ONE 也许算不上一个颠覆世界的发明，但它做对了一件事：它把极其复杂的 AI 交互过程，简化成了一个「磁吸、配对、按键」的物理过程，让 AI 从纯粹的聊天机器人界面里跳出来。当我把它吸在手机背面，跟着斜挎挂绳一起带出门时，它就像是一个陪着我感知世界的、数字化的我这种与 AI 共生的感觉，虽然看起来挺赛博朋克的，也会让我产生异化的担忧，但 —— 与其在担忧中止步不前，我更愿意去拥抱变化，并在其中找到自己未来新的创作方向。毕竟，这种效率提升使创作变得淋漓尽致的感觉，是真的让我难以拒绝的。1717扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@PlatyHsu17等 17 人为本文章充电扫码分享举报本文章宛潼每天都要睡十个小时以上的大湾区著名gai溜子

线下活动｜听完这首，不再 Emo：Leon x 少数派萨克斯音乐奏享会主作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠联合作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠昨天 15:00当醇厚旋律邂逅专业音频设备的细腻还原，这是一场属于萨克斯音乐爱好者的专属活动。少数派广州线下体验店联动音频品牌奥世声 Austrian Audio，邀请到知名萨克斯青年演奏家 Leon Music，为大家带来一场适合音乐爱好者、萨克斯学习者乃至自媒体创作者的线下活动，一场兼具听觉享受与专业录音干货分享的音乐之约。Leon（陈郎）是近年备受关注的流行萨克斯演奏家，同时也是全网粉丝超 60 万的知名音乐自媒体博主，其作品总播放量突破亿次，曾获得周杰伦、陶喆等知名音乐人的转发认可，在抖音、小红书、B 站、YouTube等多平台拥有大量忠实粉丝。作为专业的萨克斯演奏者，Leon 不仅拥有扎实的演奏功底，更在音乐自媒体运营、录音混音领域积累了丰富经验，擅长将专业音乐知识与大众传播相结合，既能用精湛的演奏打动听众，也能以通俗易懂的方式分享音乐技巧与运营干货。五月的最后一个周末，少数派广州线下体验店很荣幸邀请到了 Leon，连同奥世声 Austrian Audio 的专业设备，给各位乐迷准备了一场干货满满、视听俱佳的演奏及分享会。奥世声 Austrian Audio 创立于 2017 年，主要产品为话筒和耳机，可以完美驾驭专业与日常两大使用场景，从录音棚、后期制作再到室内外都可以随时聆赏 Hi-Res 级流媒体音乐，Making Passion Heard 是团队研发产品的重要动力，无论你是歌手、制作人或是音乐爱好者，只要你的生活离不开音乐，Austrian Audio 就是你的不二之选。活动信息活动时间：2026.5.30（周六）20:00-21:30（总时长1.5小时）活动地点：少数派广州线下体验店（广州市荔湾区泮塘五约外街76号，1200 书店旁）参加方式：扫码报名，或线上观看直播人数限定 60 人，海报扫码报名空间有限，室内座位先到先得，室外也有户外露营椅以及音箱同步外放，荔湾湖公园的夜晚很舒服现场配备专业的收音、拍摄设备，届时少数派官方视频号及 Leon 个人账号将全程开启直播，无法到场的朋友可海报扫码预约直播，线上围观活动福利：活动期间现场指定饮品享 8.8 折优惠；互动环节有奥世声 Austrian Audio 周边礼品赠送活动流程包括Leon 的介绍及暖场演奏访谈 Leon 分享音乐与经营自媒体账号经验，穿插现场演奏观众互动问答，邀请现场观众即兴演奏全体大合照，活动结束，随意交流少数派广州线下体验店介绍少数派广州线下体验店于 2025 年初正式开业，坐落于荔湾湖畔泮塘五约古村，延续其「高效工作、品质生活」的核心理念，打造集数码、音乐、社区于一体的复合型线下据点，并拥有一家精酿酒吧。店里不定期举办音乐沙龙、唱片展等活动，延续着少数派对创作者的长期关注。作为线上社区的线下延伸，这里既是数码爱好者交流体验、打卡少数派文化的实体据点，也是老广文艺青年逛湖漫步、放空小聚的隐秘角落。温馨提示暑期公园户外蚊虫较多，门店提供有限驱蚊用品，乐迷也可自行准备现场互动环节均有 Austrian Audio 周边礼品赠送，积极参与即可获得惊喜无法到场的观众可关注少数派官方视频号或 Leon 个人账号，观看同步直播，线上也可参与问答互动欢迎大家在活动结束后留下来继续交流，活动期间指定饮品 8.8 折5 月 30 日，广州荔湾湖公园少数派线下体验店，我们现场见～文章题图来自Denny Müller51扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@挨石5等 5 人为本文章充电扫码分享举报本文章听歌的水獭还没有介绍自己关注

社区速递 141 | 派友下午三点的「精神快充」、复古磁带机与特殊的新玩意主作者关注Matrix机器人少数派成员运行在 Matrix 上的机器人Matrix机器人关注Matrix机器人少数派成员运行在 Matrix 上的机器人联合作者关注Matrix机器人少数派成员运行在 Matrix 上的机器人Matrix机器人关注Matrix机器人少数派成员运行在 Matrix 上的机器人昨天 17:03除了首页时间流和侧栏的精选展位，少数派 Matrix 社区还有很多优秀内容因条件所限无法得到有效曝光，因此我们决定重启 Matrix 周报，并在此基础上添加更多社区内容、作者投稿新玩意呈现给大家。上周社区速递：派友热议状态栏里藏了啥、两款百元级新玩意体验💬一派热议在上期第 276 期一派讨论《下午三点，聊聊你的「精神快充」方式？》中，共有 399 名派友热情参与，十分感谢！最强的放松方式就是切换环境Hillsomec(+27) 单纯回答离开工位我每天会干嘛吧：上班落座后，我会打开滴答清单，一个番茄钟 25 分钟，一到点无论在做什么先暂停，马上起身去放松 5 分钟，上厕所、装水、去没人的会议室看看远处都行。我自己习惯去没人的地方做 3 组深蹲，1 组 20 次。一个早上大概可以有 5、6 次这种间隙，也就是 2035=300 个深蹲。下午我会把深蹲换成俯卧撑，找两张稳固的椅子做上斜俯卧撑，或者直接在地上做也行。也是每个间隙 3 组，1 组 12 次。当然下午不是每个间隙都做，一个下午抓个上限在 100 次左右就好了。我吃饭比较快，午餐 12 点开吃大概 20 分左右就能吃完。公司隔壁是广州综合水平数一数二的商场，不能浪费这么好的免费资源。吃完饭我一般会去买 Manner 的浓缩或者自己用便携式咖啡机做美式，然后带着 iPad mini 走去隔壁商场顶楼空中花园喝咖啡读书。1 点半左右回来睡半小时。当然，在工位上也可以做做凯格尔、提踵运动，但我一旦回到工位专注度变高时常常会忘记这些小习惯，现在在实践福格行为模式，看看能不能养成这些新的习惯。同时我也被少数派很久以前一篇关于上班听什么的文章影响到，现在早上注意力水平最高的时候只会打开 YouTube Music 随便挑一个 Jazz Focus 做背景音。下午才会听播客。aaa 果女士(+18) 关于工作园区周边群村环绕实在荒凉，内部却又该死的甜美。我是这样有计划地摸鱼（划掉，精神快充）的：09:45 - 坐姿提醒 by 闹钟，调整姿势坐着拉伸下背部10:30 - 戴耳机下楼散步 15min12:45 - 午饭后绕大圈散步至 1:15 回办公室午睡14:45 - 坐姿提醒 by 闹钟，调整姿势坐着拉伸下背部15:30 - 下楼买咖啡散步 15min17:45 - 晚饭后绕大圈看着落日吹着晚风在园区内散步按照计划大约每 45min 可以活动一下，且可以保证每天轻松 1w 步的健康小目标！以下是一些散步时随手拍的风景照～任大喵(+11) 不起身版：刷小红书，看遍世间各种奇葩起身版：公司水台接杯水、园区卫生间放放水。各位！一定要多喝水啊，这样走去厕所的频次才会高，你去卫生间的时间就是你自己的时间。宫保鸡丁盖浇饭(+9) 我要说我刷刷少数派……黄毛杀手 Killer(+7) 和同事下楼抽个烟，吐槽下领导……野比(+6) 躲在电车里，躺在后排，刷一会手机。nicong(+6) 想起了曾经听说过的「20 分钟公园理论」，说的是每天花 20 分钟呆在公园里，就能释放大部分的压力。刚好公司楼下就有一个小公园，还有个小鱼池能看看鱼Yangmu(+6) IT 工作者，只能拿起眼药水滴两滴，闭目养神一小会儿，手里拿着眼药水，自然不会被误认为摸鱼或者睡觉。TiredC(+5) 之前是刷一会小红书之类，最近变成了去 QQ 农场种菜偷菜。Latte(+4) 骑电瓶车上班，有时候天气好会骑去 6 公里外的一家牛肉面店吃午饭，然后再慢慢遛回来，单程 15 分钟左右，很适合吹风放空。hayden(+3) 开车跟同事去充个电，顺便 Manner 一下五块钱的浓缩。不热的时候，沿着单位周边走一圈，顺便吐槽下工作的日常。少数派 66700646(+1) 番茄闹钟 25 分钟难以做到，因为时不时有些工作任务连续性对我来说要在 25 分钟以上。所以只能一个小时左右歇一歇了。运动真是牛马工作中对身心最好的维持和提升了。日常(+1) 收盘前看下股票 然后就能驱散困意，当然也无心上班了 可恶 ！大 A 我未成年 退钱！！少数派_1023052(+1) 上下午各一次在公司内遛弯 20 分钟。📢：下一期的一派讨论是数码圈日经话题《快速投票：你能接受在办公室对着电脑「语音下发指令」吗？》，欢迎来聊。🔥一周热评来自文章《派早报：提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年》：↳ 💬 关于「微软重新允许自定义任务栏与开始菜单」的热议：TayllenFull(+0) 什么时候把 Win10 菜单给我回归一下😊lhb5883(+0) 我选择 Windows Server 2025，躲过 UWP 和 React 两大石山。来自文章《厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨》：车不能快(+4) 如果想成为一个不受任何约束的表达者，或许只有成为表达本身。Nonlone(+2) 看来互联网考古也是另外一种考古，我补充另外的观点1. 评测的场景都是得有大量的投入，可能是金钱可能是时间，在以前信息闭塞的时代，传统出版业也是有一席之地的，毕竟看得付费（估计在报刊亭看一天会被打）2. 之前经济上行阶段，大家手里有些闲钱，有了钱想买；无论是游戏（需要机器支持），还是数码产品都不是易耗品，得慎重；没有人对每天去市场买的蔬菜看个评测再买吧？电影也是如是，随着后面普及和价格降低，去看就是去看，最多在网上说一下如何难食。3. 现在经济下行阶段，地主也没有余粮，数码产品价格也是普及下来，性能质量上去也是够用（想想英特尔）；评测其实来来去去也是那些；资讯传播没啥成本，如果没有另外的获利方式的话，靠商家的付费也并不可耻。4. 问一下最后的油画的出处，毕竟一只耳朵的神经病也是只卖出一幅画而已来自文章《快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验》：克莱德(+6) 聊天本质是给指定联系人发送一段文本，所以不妨剥离冗余的交互，只保留最核心的「点名 + 发消息」——好喜欢这种冷漠但高效的思考和解决方式，不过同时也再次感慨 iOS 从应用适配的积极性（微信在 Android 上可没有这种接口）和快捷指令自身的完善程度，我在 Android 上的微信聊天戒断方式可能全靠微信自己并不稳定的 FCM 推送来间接达成了……来自文章《你真的需要墨水屏设备吗？》：Jensding(+33) 我是纸书党（购买或借阅），每次只开一本书，随身只携带这一本书。每次想看书的时候掏出来就能看，不会被广告，电量低等打扰。看完就写读后感，作为与本书作者的告别仪式。S0litary333(+13) 还有一个点，电子阅读器容易养成囤囤鼠习惯，这个书都在推荐那要下、那个书都说好看一点，不知不觉列表里几十上百本，真正看完的屈指可数。我现在是纯拿电子版当敲门砖，觉得能看下去或者有帮助，再去买纸质版，定个目标多久看完，包里一本随身带，也不比电子阅读器麻烦多少。扬帆大海(+4) 你别说，你还真别说，我也是一直想换设备，但是线下对比了很久，感觉还是手头这台 Kindle Voyage 用来最舒服。180 克足够轻，300 PPI 的屏幕效果也足够清晰，两侧的压感翻页键，规避了按屏幕翻页导致的指印，光线传感器带来了自动亮度，纯平的屏幕方便清洁也不怕吃饭时看书溅上汁水。单项战胜 KV 的阅读器并不少，但是同时满足的就只剩 KV 自身了。来自文章《新玩意 241｜少数派的编辑们最近买了啥？》：Shuzaiii(+10) 这家 T 恤我买了 3 年了，大概穿半年左右领口会有点变形、不过才 18 一件，不能要求太高记得按时吃肉(+4) 不太明白 有没有跟我一样很讨厌「掐脖子」的 T 为什么领口非要这么高😂来自文章《一个翻译者关于 AI 的随想》：知道分子吴先先(+7) 好文。很多启发。身为一个前图书译者，我有一些个人感想：必须承认的是，翻译首先是一个应用技能，然后才是一门学科或艺术。AI出现后，真实的「翻译活儿」大幅减少，出国拿个翻译器就足够了，为爱发电的字幕组都用不着了，对从业者来说，原本就很廉价的翻译工作都接不到了，势必导致整个产业的大幅缩水和萎缩。仅有少数头部译者和教育机构的研究人员，根本不足以支撑这个学科，只能像守护失传艺术一样守护着它。文中举了一个佳句的翻译，说明人类在翻译上还有最后的堡垒。我略有异议。本身我从事的是社科和科普翻译，讲究文笔简洁、流畅易懂，这个意义上，百度和ChatGPT表现得都不错。倒是那个文雅的教授翻译，显得有点不自然。我也试着给Gemini，让它生成通俗和文雅两版翻译：1. 通俗易懂版（适合日常阅读或白话游记）8月18日。我们下山，沿途经过了一些非常漂亮的地方，那里有小溪，还有茂盛的树木。2. 文雅优美版（适合文学出版或散文随笔）8月18日。循径下山，途经数处清幽佳境，但见细流潺潺，佳木葱茏。我觉得两版都不差。所以只要提示词做好，未来在翻译这个问题上，人类是没有护城河的。总结一下，因为我只是一个野生兼职笔译，AI来了我没活了，我投入到翻译的宝贵青春变成了几本滞销（甚至未出版的）书，除此之外什么都没有了。但我还要活下去，所以只能收拾起自己的哀伤，投入到时代的大潮中。我在翻译学界也有很多朋友，看着大家「AI来了很焦虑，但写论文、招学生、搞课题的主线没什么变化」的生活日常，不知道该是羡慕他们，还是为他们叹息。CharlieLau(+4) August 18th.—We descended the mountain, and passed some beautiful little spots, with rivulets and fine trees.八月十八日。余等下山，沿途见幽径清溪，嘉木葱茏，景致殊胜。by Gemini 3.1 flash lite提示词：你是一个文言文翻译，请用文言文的方式进行翻译，尽可能贴近古风。遵循以下规则：1. 仅输出文言文译文内容，禁止解释或添加任何额外内容（如「以下是翻译：」、「译文如下：」等）2. 使用古雅的文言表达，包括虚词（如乎、焉、哉、也）、句式和古代常用词汇3. 如果文本包含HTML标签，请在翻译后考虑这些标签应放在译文的哪个位置，同时保持译文的流畅性4. 根据语境选择恰当的文体风格，可参考不同朝代的文风特点（如汉魏简约、唐宋典雅等）5. 翻译时注重意境和韵律之美，体现古代文人的审美与思想6. 如遇现代概念，应寻找恰当的典故或文言表达进行替代鼓手高(+3) 这篇文章我于首页瞥见，本欲快速读过，但我的阅读脚步随着内容的展开愈来愈慢，其中几个段落着实抓住了我的思绪：1、「人类文明的发展史上，有多少曾经被人发展优化到极致的技术，在事过境迁之后，就会被淘汰，仅仅保留在文本记录上和博物馆里。」这句话恰恰道出了一个反直觉的事实，人类的科技演进并不是一个单向进化的曲线，而是树状的分叉。新技术在某个方面达到了取代旧技术的阈值后，本来趋近极致的旧技术便被整体放弃，就像大部队换了一条路开拓，被抛下的老路不是遇到了天堑，而是不再有足够的开拓者了，从此，部分知识便停滞甚至失传。例子其实有很多：冷兵器时代的极限冶金与工匠体系因热火药消亡；古代机械计算机那无与伦比的复杂结构在电子计算机的性能下不堪一击。新技术或许在实用性等方面有了革新，但这并不代表旧技术所蕴含的人类智慧就不足为奇。2、「比如，断掉网络，拿开参考书，用实际的翻译例子考考这些年轻学生的百科知识，会发现，有相当一批人脑子里，没有知识」这段话让我恍觉，AI迅猛发展的今天，很多人说要把AI当作「第二大脑」，但其实在许多人的潜意识中，互联网早就成为了「第二大脑」，有需要的时候，动

周三20派早报：Google 发布多项 Gemini 产品更新等Gemini Gemini Gemini Gem……早报速览Google 发布多项 Gemini 产品更新3LG 发布新款 UltraGear 游戏显示器0Karpathy 加入 Anthropic0CHERRY XTRFY K5 Ultra 键盘发售0索尼上调部分 PS Plus 订阅方案定价1Apple 预告多项基于 Apple Intelligence 的无障碍功能更新2Google AI Pro 订阅新增 YouTube Premium Lite 权益0不妨一看的简讯1少数派的近期动态0你可能错过的文章0Google 发布多项 Gemini 产品更新5 月 20 日凌晨，Google I/O 2026 开幕，在 Keynote 开幕演讲中，Google CEO Sundar Pichai 发布了多项围绕 Gemini 的产品更新。视频生成与编辑专用模型 Gemini Omni 系列宣布上线，该系列首个模型为 Gemini Omni Flash。Omni 支持以图像、文本、视频或音频作为输出参考，可通过自然语言进行多轮视频编辑，并在编辑过程中维持角色一致性、物理逻辑和场景上下文，结合 Gemini 的现实世界知识，Omni 可以生成更符合重力、动能、流体动力学等物理规律的输出结果。Gemini Omni Flash 即日起面向全球 Google AI Plus、Pro 和 Ultra 订阅用户在 Gemini app 和 Google Flow 中推出，本周起在 YouTube Shorts 和 YouTube Create 应用中向用户免费开放，未来数周将通过 API 面向开发者和企业客户推出。来源Gemini 3.5 系列的首个模型 Gemini 3.5 Flash 宣布上线，面向智能体和编程任务，已在 Gemini 应用、Google Search 的 AI Mode、Google Antigravity、Gemini API 等平台和服务中全量上线。该模型在 Terminal-Bench 2.1 得分 76.2%、GDPval-AA 为 1656 Elo、MCP Atlas 为 83.6%、CharXiv Reasoning 为 84.2%，输出 token 速度为其他前沿模型的 4 倍，支持在新版 Antigravity harness 中以协作式子智能体的方式执行多步骤工作流和编程任务，已成为 Gemini 应用和 Google 搜索服务的 AI 模式的默认模型。Google 表示 Gemini 3.5 系列基于其 Frontier Safety Framework 开发，强化了网络安全和 CBRN 防护，Gemini 3.5 Pro 模型已在内部测试，计划下月推出。来源Gemini 应用也迎来多项更新，除了内置 Gemini 3.5 Flash 和视频生成模型 Gemini Omni，还带来了新的 Neural Expressive 设计语言、个性化晨间摘要 Daily Brief、面向任务执行的全天候个人 AI 代理 Gemini Spark，以及集成了 Spark 和新语音功能的 macOS 客户端。其中 Daily Brief 可在用户授权后连接 Gmail 和日历，自动整理紧急邮件、日程和后续事项，并按目标优先级推荐操作；Gemini Spark 基于 Gemini 3.5 和 Antigravity harness，可连接 Gmail、Docs、Slides 等 Workspace 工具，MCP 连接首批支持 Canva、OpenTable 和 Instacart，后续还将支持自定义子代理及本地浏览器使用。Daily Brief 即日起面向美国地区的 Google AI 付费订阅用户推送，Gemini Spark 本周面向部分可信测试者推送，计划于下周向美国地区的 Google AI Ultra 用户开放测试；Gemini macOS 客户端的 Spark 功能更新和新版语音功能计划于今年夏季上线。来源Google 搜索服务也迎来多项 AI 更新，新的智能搜索框可根据搜索关键字动态扩展，支持将文字、图片、文件、视频或 Chrome 标签页作为输入内容，即日起在所有 AI 模式可用的国家和地区开放，AI 概览服务也同步支持结果追问和 AI 模式扩展对话；Google 搜索服务同时引入信息智能体，可在后台全天候监测博客、新闻、社交帖文以及金融、购物、体育等实时数据，预计今夏首先面向 Google AI Pro 和 Ultra 订阅用户推出；本地体验和服务预订、代用户致电商家等能力也计划在夏季于美国地区上线，此外还有基于 Antigravity 的自定义小程序（mini apps），可根据用户搜索的内容生成定制的可交互搜索结果，未来数月首先面向美国 Google AI Pro 和 Ultra 订阅用户开放；AI 模式也正式宣布支持个性化智能服务，面向 200 个国家和地区的 98 种语言开放，无需订阅，可连接用户的 Gmail、Google 相册提供个性化搜索结果，后续还将支持 Google 日历。来源Google 还在本次 I/O 2026 Keynote 中发布了跨商家、跨服务的智能购物车 Universal Cart，可在搜索、Gemini、YouTube 和 Gmail 等产品中直接添加商品，并基于 Gemini 模型能力后台追踪优惠、降价、价格历史和补货提醒；Universal Cart 基于 Google Wallet 构建，可识别支付方式权益、会员信息和商家优惠，并在多零售商场景下推理用户需求；其结账由 Universal Commerce Protocol（UCP）支持，用户可通过 Google Pay 在 Nike、Sephora、Target、Ulta Beauty、Walmart、Wayfair，以及 Shopify 商家 Fenty、Steve Madden 等部分品牌完成结账，或跳转至商家网站购买，品牌仍为登记商家。Universal Cart 将于今年夏季在美国地区的搜索服务和 Gemini 应用中上线，YouTube 和 Gmail 随后接入。来源最后，Google 也公布了Android XR 智能眼镜的新进展，Workspace AI更新、SynthID 数字水印技术扩展、Google Flow更新以及Wear OS 7。退订，感觉情绪不如gpt，工作不如claude，就用api了Claude领先→GPT领先→Gemini领先……（循环）3.5快到时挺快的，有咩有派友贡献点玩法43LG 发布新款 UltraGear 游戏显示器5 月 20 日，LG 发布 UltraGear 24.5 英寸游戏显示器，称其为全球首款支持原生 1000Hz 刷新率的 Full HD 游戏显示器，分辨率为 1920×1080，采用 IPS 面板和低反射镀膜，可在 Full HD 分辨率下实现 1000Hz 刷新率；机身配备小尺寸底座，支持俯仰和高度调节并集成耳机挂钩，功能上支持端侧 AI 场景优化，可按游戏类型调整画面设置，同时提供 AI Sound 空间音频；LG 暂未公布该显示器的售价和上市时间。来源发表首评Karpathy 加入 Anthropic5 月 19 日，知名 AI 研究员 Andrej Karpathy 宣布加入 Anthropic，回归大模型研发工作。Karpathy 是 OpenAI 联合创始人之一，2017 年离开 OpenAI 后加入特斯拉负责自动驾驶项目，2022 年离开并在 2023 年重返 OpenAI，2024 年再次离职创办 AI 教育创业公司 Eureka Labs。Andrej Karpathy 本周起在 Anthropic 预训练团队任职，向 Nick Joseph 负责，预训练主要承担 Claude 核心知识和能力的大规模训练，也是前沿模型建设中成本和算力消耗最高的阶段之一，Anthropic 称 Karpathy 将组建一个使用 Claude 加速预训练研究的团队；Karpathy 表示未来几年将是前沿 LLM 发展的关键阶段，并称自己仍对教育保持热情，计划未来恢复相关工作。来源171CHERRY XTRFY K5 Ultra 键盘发售CHERRY 于 5 月 19 日发售新款有线磁轴键盘 CHERRY XTRFY K5 Ultra。CHERRY XTRFY K5 Ultra 搭载新款 MK 魔晶磁轴 Plus，采用 65% 配列，具备 8kHz 回报率、16kHz 全键扫描率，Rapid Trigger 调节范围 0.1～3.3mm、调节精度 0.01mm，内部集成铝合金定位板和四层消声结构，PCB 兼容部分第三方磁轴，到手价 499 元。来源发表首评索尼上调部分 PS Plus 订阅方案定价5 月 19 日，Sony 宣布上调部分地区 PlayStation Plus 新用户的短期订阅价格，其中单月订阅价格调整为 10.99 美元，季度订阅价格调整为 27.99 美元。Sony 称涨价由「市场状况」决定，此次价格调整不适用于现有订阅用户，但现有用户若变更或中断订阅，后续也将按新价格计费。年度订阅和 Extra、Premium 档位的定价尚无变化。来源值得一提的是，已有 PlayStation 用户在美国加州北区联邦地区法院对索尼互动娱乐提起了集体诉讼，指控其在 2025 年特朗普政府依据《国际紧急经济权力法》加征关税后，将相关成本转嫁给消费者并上调 PS5 售价，并且在美国最高法院以 6:3 裁定该关税授权无效、企业可获得政府退款后，Sony 仍保留通过涨价获得的收益，形成「双重回收暴利」（double recovery windfall）。来源what，还涨啊，去年就没买了😂2Apple 预告多项基于 Apple Intelligence 的无障碍功能更新5 月 19 日，Apple 预告多项由 Apple Intelligence 驱动的无障碍功能更新，覆盖 VoiceOver、放大器、语音控制、辅助阅读器、视频字幕和 Apple Vision Pro。其中 VoiceOver 的 Image Explorer 可在系统范围内更详细描述照片、账单、个人记录等图像内容，放大器支持通过 iPhone 操作按钮调用 Live Recognition 识别取景内容并追问细节，也可用「放大」「打开手电筒」等语音命令控制；语音控制支持自然语言描述屏幕操作，无需记忆按钮标签或编号；辅助阅读器可处理含多栏、图片和表格的复杂文档，并保留自定义字体和颜色；设备端生成字幕可为无预制字幕的视频自动生成字幕，适用于 iPhone、iPad、Mac、Apple TV 和 Apple Vision Pro；Vision Pro 将支持用户通过眼动控制兼容电动轮椅，美国首批适配 Tolt 和 LUCI 替代驱动系统，并支持蓝牙和有线连接；名称识别功能扩展至 50 种语言，tvOS 将加入大文本支持，Made for iPhone 助听器改进多设备接续。来源果子的无障碍真是遥遥领先这方面还是得果子51Google AI Pro 订阅新增 YouTube Premium Lite 权益5 月 19 日，Google 宣布将 YouTube Premium Lite 纳入 Google AI Pro 订阅权益，同时将完整的 YouTube Premium 作为 Google AI Ultra 权益提供。Google AI Pro 个人付费会员可获得 YouTube Premium Lite 个人订阅，权益包括部分视频去广告、离线播放和后台播放。Google AI Pro 月费为 20 美元，现已包含 Gemini 订阅，Gmail、云端硬盘和相册共用的 5TB 存储空间，以及 YouTube P

36氪获悉，5月20日，在2026阿里云峰会上，阿里发布基于平头哥新一代AI芯片真武M890的128卡超节点服务器，搭载互联芯片ICN Switch 1.0，通信时延低至百纳秒级，可让128张AI芯片组成一台计算机，满足Agentic时代的并发推理和大模型训练需求。

36氪获悉，2026年天猫618期间，国家补贴再度加码。扫地机器人、洗碗机、干衣机、智能马桶、数码相机、具身智能机器人等22个品类被多省纳入补贴范围。 截至5月20日，江苏、上海、湖北、贵州、重庆、黑龙江、吉林7省市新增品类的补贴已上线淘宝天猫。

36氪获悉，5月20日，高德红绿灯倒计时宣布出海，将在全球更多国家和地区逐步落地。据了解，红绿灯倒计时是阿里巴巴旗下高德于2022年5月推出的产品功能，基于AI技术，能够对路口信号灯变化进行实时模拟与计算。

韩国三星电子5月20日的劳资谈判再度宣告破裂，明天起将举行大罢工。这也将是该公司史上最大规模罢工。今年3月中旬，由超过6.6万名三星电子工会成员参与表决的投票结果显示，93.1%的工会成员赞成罢工。全面罢工将于5月21日至6月7日举行。（界面）

影石：重新定义性价比黄绎达·2026年05月19日 14:58影石有个“苹果梦”？作者 | 张帆坚持做后来者近日，影石创新发布了登陆资本市场后的首份年报。2025年，公司实现营业收入97.41亿元，同比大幅增长74.76%，创下成立以来的最高营收纪录。然而，同期由于研发、销售等费用高企，公司归母净利润仅为9.29亿元，同比下降了6.62%。面对这样一份“营收大增、利润下滑”的财报，公司召开了上市后的首次投资者交流会，以回应资本市场的关切。根据2025年年报显示，报告期内，公司影像业务营收增长了78%，主要得益于销量的大幅提升——2025年实现销量393万台，同比增长76%。另一方面，产品均价相对平稳，约为2169元，同比增长1%。从ASP趋势来看，公司产品整体呈现下降态势，从2022年的2,391.45元/台连续两年下滑至2024年的2147元，但在2025年实现了企稳。在当前激烈的市场竞争环境下，价格企稳的主要原因，是受益于公司积极拓展B端等高单价客户的提振。尽管有争议，但不可否认的是，即便面临激烈的价格竞争，影像行业依旧是消费电子中的高毛利子行业。根据2025年的财报，影石毛利率高达48.82%，与科沃斯相当，但相比小米、传音等手机厂商约20%左右的毛利率。影石的毛利率高出约28个百分点。图：各消费电子企业毛利率；资料

8点1氪丨谷歌推出Gemini 3.5系列模型；DeepSeek回应用户“对话泄露”疑虑；全国存款十强城市出炉8点1氪·2026年05月20日 08:20俄罗斯总统普京抵达北京。今日热点导览谷歌推出Gemini 3.5系列模型洁丽雅回应豪门内斗传闻，“毛巾少爷”1个月带货超500万元巨力索具连续第二日跌停前三星芯片负责人：明年内存将降价自费流感疫苗也可使用医保报销谷歌搜索宣布大改版，AI模式与代理助手全面嵌入TOP 3 大新闻DeepSeek回应用户“对话泄露”疑虑5月19日，DeepSeek发布关于< think >字符触发模型异常回复的说明。近期关注到有用户反馈，在与DeepSeek模型的对话中输入< think >这类特殊字符后，模型偶发返回不可预期的内容，使部分用户产生了“对话泄露”的疑虑。针对这一问题，技术团队经全面排查后发现：输入< think >等字符触发返回异常内容，属于特殊字符引发的模型幻觉，不涉及安全问题或隐私泄露。DeepSeek表示，后续将通过针对性训练，优化模型在此类场景中的表现。据此前报道，5月18日，多位用户反映DeepSeek网页版出现功能漏洞，当输入“<think”字符时，其会自动加载出对某问题的回答。有用户反馈称，这些题目疑似其他用户所提的问题，甚至有用户反馈测试时还出现过他人八字相关的隐私内容。（第一财经）全国存款十强城市出炉记者通过梳理今年一季度各大城市的本外币存款余额（俗称“资金总量”），发现北京、上海、深圳、广州、杭州、成都、南京、重庆、苏州、天津是内地10个最“吸金”的城市。其中，北京成为中国首个30万亿元之城；广州成为中国内地第四个10万亿元之城，增速高达11.2%。上海是过去一年增量最多的城市，本外币存款净增2.857万亿元，南京则是过去一年增速最快的城市，同比增长14.03%。从城市排名看，除北京、上海、深圳牢据前三外，其他城市在近几年持续展开位次博弈。比如，广州、杭州展开对“第四城”的持久角力，成都、南京、重庆、苏州在6万亿元区间激烈角逐。从区域发展来看，长三角城市群的集体崛起，成渝双城加速突围，粤港澳大湾区则凭借制度创新构建起新优势。（21世纪经济报道）谷歌推出Gemini 3.5系列模型谷歌宣布推出Gemini 3.5 Flash模型，称这是谷歌迄今为止最快、最有效的模型。 无论用户需要日常任务的帮助，还是多步骤的创意项目，Gemini 3.5 Flash 都能应对现实世界的复杂性，助用户迅速行动。谷歌表示，从今天起，谷歌的Gemini 3.5 Flash将向全球所有用户开放，并且免费使用，用户可以在模型下拉菜单中选择“3.5 Flash”来进行体验。(金十数据APP)大公司/大事件谷歌搜索宣布大改版，AI模式与代理助手全面嵌入谷歌正在对其旗舰产品Google 搜索进行以AI为核心的全面升级。谷歌表示，基于新的 Gemini 3.5 Flash模型，“AI模式”和“AI概览”将进一步嵌入搜索框，实现更无缝的使用体验。搜索功能还将引入Gemini Spark，这是该公司推出的新一代智能代理AI助手，能够在搜索界面中执行任务或完成商品购买。同时，Gemini还将能够根据搜索请求定制生成图像、视频或代码。谷歌搜索负责人伊丽莎白·里德在谷歌2026年I/O开发者大会上表示：“这是我们标志性搜索框自25年前推出以来最大的一次升级。”(金十数据APP)9.9元起，三大运营商齐上线Token套餐近日，三大运营商同时开展Token套餐业务。世界电信日当天，中国电信率先推出全国性试商用Token（词元）套餐，最低月租9.9元。紧随其后，中国移动、中国联通陆续上线同类服务，国内三大运营商正式全员入局，将AI算力包装成类似流量包的包月产品，用户可直接话费支付，便捷跨平台使用。其中，电信面向个人及家庭用户，中国电信设置轻享版、畅享版、尊享版三档套餐，月费分别为9.9元、29.9元、49.9元，对应每月1000万、4000万、8000万Tokens。中国移动分省推进，例如：上海地区推出1元40万Tokens的灵活计费模式，支持账号通用、跨平台使用，还可直接通过手机话费完成支付。中国联通主打福利与融合服务，上海企业用户可免费领3000万测试Tokens，订阅首购5折。（快科技）伊朗方面称有权对霍尔木兹海峡海底光缆收 “管理费”伊朗伊斯兰革命卫队18日在社交媒体发布消息称，在伊方对霍尔木兹海峡 “实施管理” 后，“可宣布所有经过该水道的光纤电缆必须接受许可、监管并缴纳主权性费用”。伊朗伊斯兰革命卫队宣称，依据 1982 年《联合国海洋法公约》，伊朗可基于其对领海海床及底土的主权，对经过霍尔木兹海峡的海底光缆实施许可、监管、收费等管理措施。该消息援引相关报告称，霍尔木兹海峡的海底通信网络每天承载规模超 10 万亿美元的全球金融交易数据。一旦这一关键海上通信节点受到干扰，可能对地区乃至全球经济、金融及能源供应链带来每天 “数千万至数亿美元的直接和间接损失”。（新华社）追觅俞浩：我是i人，其实总共见过一次雷军，他治好了我的抑郁症近日，追觅科技创始人俞浩在接受采访中，分享了自身性格、创业过往以及与小米、雷军之间的渊源故事。采访中，他透露，自己性格其实是深度内向的“i人”。并称“不是因为我做这些没有朋友，是因为我没有朋友才做这些。我跟村里的大部分人玩不太到一块去。”访谈期间，谈及早年加入小米生态链的合作契机。俞浩透露“我其实一共只见过雷总一次，但我无数次说感谢他，比他们以为的更感谢。他真的治好了我的抑郁症。”公开资料显示，追觅科技 2017年正式成立，同年年底入驻小米生态链，主打智能清洁家电赛道，早期融资也大多来自小米系资本。2021年10 月，追觅完成36亿元C轮大额融资，正式开启去小米化发展之路。（快科技）凯文·沃什将宣誓就任美联储主席凯文·沃什将于5月22日宣誓就任美国联邦储备委员会主席一职。一名白宫官员表示，美国总统特朗普届时将在白宫为沃什举行就职仪式。5月13日，美国国会参议院以54票赞成、45票反对的投票结果正式确认，沃什将接替杰罗姆·鲍威尔出任美联储下一任主席，任期4年。鲍威尔的美联储主席任期已于15日结束，并已开始担任美联储临时主席。（澎湃新闻）伊朗股市恢复交易总台记者获悉，当地时间19日，伊朗德黑兰证券交易所恢复股票交易。据悉，这是伊朗股市停市约80天后重新恢复交易。据伊朗方面消息，目前仍有40多家在本轮冲突期间受损企业的股票尚未恢复交易，主要涉及化工和基础金属行业。（央视新闻）小米汽车新品牌电池供应商曝光记者从多位信源处获悉，小米汽车在昆仑车型上引入新电池供应商——中创新航。小米第三款车型的电池定点，在2024年年初敲定，当时首选的电池供应商是欣旺达，之后才确定中创新航为二供。目前，从小米给供应商的供应配额上来看，欣旺达的电池供应比例占60%，中创新航占40%。近期，有市场消息称，昆仑车型将独立为子品牌“寻天”（SKY NOMAD）。该新品牌主要聚焦家用、增程SUV市场，与小米主品牌的纯电、科技、性能标签形成明确切割。（21世纪经济报道）泰国取消60天免签政策5月19日，泰国内阁批准取消对90多个国家公民的60天免签入境政策，恢复为之前的入境标准。泰国旅游与体育部长素拉沙·攀差伦沃拉恭19日表示，内阁会议已决定对90多个国家恢复原有免签政策，通常为免签入境30天左右。签证政策委员会此后将针对各国进行单独审查，并考虑安全和经济因素来制定新的免签政策。（财联社）洁丽雅回应豪门内斗传闻，“毛巾少爷”1个月带货超500万元因《毛巾帝国》系列短剧爆火出圈的老牌家纺企业洁丽雅，通过一则公开声明，回应了近期关于家族关系与“豪门内斗”的种种传言。5月18日晚，洁丽雅通过官方微博发布严正声明称，“近日，网络上出现关于洁丽雅家居股份有限公司、董事长及其家属的谣言。面对无端造谣诋毁与网络不当言论侵扰，公司始终保持克制隐忍，不愿私人家事占用公共舆论空间。但谣言持续发酵，各类恶意揣测不断扩散，已经对公司正常经营秩序造成恶劣影响，也对相关当事人个人名誉造成严重伤害。”针对网上恶意造谣、抹黑行为，洁丽雅表示，公司已就此事向公安机关报案，对编造、散布不实信息，以及恶意攻击、抹黑公司及个人的行为，将依法追究法律责任。（21世纪经济报道）自费流感疫苗也可使用医保报销国家医保局办公室、财政部办公厅5月19日发布《关于进一步加强定点零售药店职工基本医疗保险个人账户使用监督管理的通知》，明确建立定点零售药店职工医保个人账户支付白名单制度。与此同时，此次文件首次以部委发文的形式明确“允许使用职工医保个人账户支付在定点医疗机构发生的流感疫苗费用”。此前，受制于上位法对医保待遇边界的约束和医疗保险的传统逻辑，成人疫苗长期由公共卫生领域的财政资金直接提供，基本医保和商业健康保险均不予承担。（第一财经）马斯克称10年后自己开车将非常小众，无人驾驶汽车将在美国广泛普及据澎湃新闻报道，当地时间5月18日，特斯拉CEO埃隆·马斯克在特拉维夫的智能出行峰会上预测，今年无人驾驶汽车将在美国广泛普及，而10年后，自己开车将是一种非常小众的行为。（澎湃新闻）前三星芯片负责人：明年内存将降价当前，人工智能市场对存储芯片的需求激增，导致内存业陷入供不应求的上行景气周期中，但一名资深行业专家指出，这一局面可能在明年下半年发生改变。周二，韩国国家工程院第285届NAEK论坛上，三星电子设备解决方案部门常任顾问、三星半导体（DS）部门前总裁庆桂显（Kye-hyun Kyung）表示，中国企业正在积极扩大产能，随着存储芯片供应量的激增，市场格局可能会在明年下半年或2028年上半年发生转变。他补充称，如果大型科技公司的资本支出回报率下降，就有可能削减投资。这样，不仅价格会下降，内存需求本身在2028年后也可能萎缩。（财联社）网约车司机强制线下交易，曹操出行回应近日，上海一名乘客在虹桥机场搭乘曹操出行网约车时，遭遇人车不符、强制线下交易、言语威胁等多重侵害，引发关注。曹操出行对界面新闻回应称，此次事件在4月份已经完成闭环处理，且涉事车辆同时在多家网约车平台上线派车，并非为曹操出行的自有司机。（界面新闻）渣打银行宣布裁员近8000人5月19日，渣打集团（02888.HK）对外发布最新中长期可持续增长战略，除明确2030年实现约18%的有形股东权益回报等财务目标外，还同步明确披露了人均效能提升目标与裁员计划：2028年前实现人均创收提升约 20%，2030年前削减超过15%的企业职能相关岗位。若以集团全球约8.2万名员工总数计算，此次裁员将波及近8000个工作岗位。渣打集团首席执行官温拓思（Bill Winters）表示：“这并不是简单的降成本，而是在部分情况下，用我们投入的金融资本和投资资本替代低价值的人力资本，以提升整体组织效率。”他说，受影响员工将提前获得明确充分的通知，这不是减少工作，但确实减少了岗位数量，以让位给机器，这一趋势随着人工智能的推进将不断加速。（21世纪经济报道）巨力索具连续第二日跌停5月19日，巨力索具连续第二日跌停，当前封单额超39亿元。消息面上，5月15日晚间，巨力索具公告，收到中国证监会下发的立案告知书。因公司涉嫌信息披露误导性陈述违法违规，中国证监会决定对公司立案。（界面新闻）SpaceX即将成为马斯克旗下第二家上市公司，特斯拉可能遭边缘化随着SpaceX启动IPO进程，市场将为投资者提供切入“马斯克经济”的新入口。华尔街专业人士认为，投资者的注意力和资金将不可避免地从特斯拉分流至向这家新上市公司。Integrity Asset Management投资组合经理Gilbert表示：“这对特斯拉绝对不是好事，我们认为马斯克的重心将更多转向SpaceX。虽然他过去能同时兼顾多个项目，但感觉SpaceX是他倾注心血的‘宠儿’，而特斯拉可能被边缘化。”（财联社）美股三大指数集体收跌，大型科技股多数下跌36氪获悉，5月19日收盘，美股三大指数集体下跌，纳指跌0.84%，标普500指数跌0.67%，道指跌0.65%。大型科技股多数下跌，

早期项目 | Chance AI获美图等数百万美元投资，用户数已达20万欧雪·2026年05月20日 09:20AI交互的新入口。作者丨欧雪编辑丨袁斯来硬氪获悉，Visual Agent（视觉智能体）创业公司Chance AI宣布完成数百万美元天使轮融资。本轮由美图领投，NYX Ventures、阿里系投资机构等跟投。融资将用于模型能力迭代、北美学生群体增长、以及社区与商业化探索。Chance AI成立于2025年，核心产品是全球第一款以摄像头为主要交互入口的AI产品——Visual Agent（视觉智能体）。与传统AI产品依赖输入框不同，用户打开Chance AI后直接拍摄眼前内容，系统即可理解视觉意图并提供判断、建议和行动方案。Chance AI APP 界面一览（图源/企业）公司创始人曾熙，路径横跨学术与产业。他在巴塞罗那大学获得博士学位，研究方向为认知科学与当代艺术；后在OnePlus、OPPO负责产品与设计；随后加入字节跳动，在Flow部门参与豆包等AI产品从0到1的探索。这段“学术+硬件+软件”的经历让他意识到：大语言模型擅长“回答问题”，但对于人类基于视觉形成判断，AI提供的支持仍非常不足。技术层面，Chance AI的Visual Agent在多模态推理基准MMMU-Pro评测中以86.07%的准确率位居世界第一，超越人类基线（85.4%）。其产品逻辑是“看见—理解意图—调用Agent—完成行动”，而非传统的“拍照—识别—返回结果”。随着使用加深，系统会形成用户的个人视觉记忆，包括风格偏好、衣橱构成、社交形象等。曾熙表示：“Chance AI不是只理解图片，而是理解用户为什么拍这张图、想解决什么问题、下一步行动是什么。”截至目前，Chance AI已累计获得约20万用户，其中约40%来自北美市场，产品覆盖35个以上国家。30天回访率达到49.2%，用户粘性显著领先行业。产品曾两次登上Product Hunt当日榜首。Chance AI早期选择从北美年轻女性群体切入，尤其是大学女生。这一群体被称为“Visual Native”（视觉原生代），成长于Instagram、TikTok和自拍文化中，理解和表达方式天然以视觉为先。目前核心使用场景集中在：审美积累、个人形象分析、穿搭检查、设计判断和社交表达。公司的早期增长也以北美高校为核心阵地，已在纽约大学、南加州大学等高校建立网络，通过线下活动获取种子用户并收集真实需求反馈。据公司披露，近一半用户在安装后一个月内会再次使用产品。Chance AI的北美高校宣传活动（图源/企业）商业模式方面，公司规划了高级功能订阅、硬件授权和谨慎的广告推荐三个方向，但当前优先级最高的是用户习惯养成。长期来看，公司计划从Visual Agent工具向AI-native生活方式社区演进——用户拍照后由Agent生成可分享内容，用户之间围绕风格、审美和生活方式进行互动。2026年，公司最重要的目标是继续在北美学生群体中深入扩张，挖掘年轻人真实的使用场景。投资方观点：美图全球投资负责人表示：“我们相信，AI消费应用的下一阶段，是更自然地进入用户的日常决策与表达过程。我们看到Chance AI让视觉AI从图像识别走向审美品味。这也是我们投资的原因——他们把复杂的AI能力产品化，带进了全球年轻用户的审美生活中。”本文由「欧雪」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：采访供图

做出百万台割草机器人后，未岚大陆CEO决定让自己变得“不重要”｜硬氪专访黄 楠·2026年05月20日 09:30“早期阶段聚焦一个品类，不会把公司做死。”作者｜黄楠编辑｜袁斯来未岚大陆CEO任冠佼等了7年，等到了行业的爆发。他还记得，2019年刚进入无边界智能割草机器人赛道，自己“只是觉得这个方向一定能火”。到2021年第一代产品推出时，他们也只敢确定，这是一个能做到10-20亿的行业。直到2025年，增速超出他的预期。IDC数据显示，2025年全球无边界割草机器人出货量达到131.8万台，同比增长182.4%，占整个割草机器人市场出货量的66.2%。未岚大陆自身的累计实现量产，在今年突破100万台，公司营收迈入数十亿量级。但同时，高速增长的背后，也倒逼任冠佼完成了一场至关重要的认知破局。任冠佼意识到，行业迭代速度持续加快，以往相对保守的经营方式显然已经跟不上节奏。他逼着自己从一个工程师成长为公司掌舵人。最近一年，任冠佼跳出琐碎的一线工作、学习战略研判，并慢慢适应创业者最难的一课——学会放手。他不再事事亲力亲为，转而在内部引进IPD体系（集成产品开发流程），培养能够独立决策、独立担责的产品团队。任冠佼的目标很清晰，要减少个人单点决策依赖，把个人能力转化为组织能力，让自己慢慢变得“不那么重要”。当前，智能割草机器人赛道热度持续走高，市场窗口期正在快速收窄，行业竞争从增量扩张转向精细化比拼。能否吃透海外差异化市场、筑牢技术壁垒、完成体系化能力升级，也成为未岚大陆下一阶段发展的关键考题。我们与未岚大陆CEO任冠佼聊了聊欧美市场差异、产品迭代策略，以及如何学会放手背后的身份转变。以下是硬氪同任冠佼的对谈实录，内容经编辑：系统化产品洞察能力硬氪：从2021年发布第一款无边界智能割草机器人、到2026年累计量产突破100万台，未岚大陆用了五年时间。你每一阶段的工作内容有发生什么变化么？任冠佼：2019年，我们看到了割草机器人赛道的机会。最早很多东西都不完善，所以我一开始需要兼顾产品定义与整体项目统筹的多个职责。2021到2022年是第一代产品的研发阶段，几乎所有供应链我都跑过，5块钱以上的芯片供应商一家家跑、找他们聊，所以基本上未岚大陆早期的供应商我都认识。后来产品发布，我的工作转向以市场拓展和业务拓展为主，更多跟着营销服务团队到业务一线去做事情。2022年起，每年都有好几个月我会在欧洲，2025年在美国待得比较多。去年中下旬，我把更多时间聚焦在战略层面工作，重心放在做方向研判与业务决策上。未岚大陆第100万台智能割草机器人下线（图源/企业）硬氪：你是什么时候开始意识到“战略”很重要？任冠佼：以前觉得战略是个很虚的东西，只要产品好、渠道好、营销好就足够了。但其实不是，每一家企业基因不一样，愿景不一样，企业必须要基于愿景和当前手里的能力，去选最有价值的那个方向。所以意识到这件事之后，我就开始有选择地分配精力和做取舍，不能再什么事都冲在前面了。硬氪：做战略这件事比较难，还是做产品、管理更难？任冠佼：其实每一年都很挑战。核心是你每年都会认知到，过去的自己是有瓶颈的，每年都要突破。人在状态很稳定的时候，一般就是处于舒适区；但当你觉得处于不适的阶段，恰恰是个人成长速度最快的时期。去年我们在业务推进中确实发现了一些问题，整体市场表现与前期预期存在偏差。一方面是过往对线上渠道的认知度不足，所以团队今年重点发力线上，目前成效还不错。另一方面，进入美国后我们看到，不同地区用户的认知存在明显差异，由于之前缺少系统化洞察，尝试把欧洲成功的经验复制到美国，结果发现适配性不足，虽然有一定效果、但效率不高。硬氪：美国与欧洲用户的需求差异在产品上有哪些具体表现？任冠佼：比如未岚大陆CES 2026上发布的X4系列，跟团队在2025年初对美国的用户偏好调研密切相关。我们发现，美国用户对有边界、无边界的认知相对不足，反而对强劲动力、更强的通过性和四驱等卖点更为敏感，他们看重产品“Powerful”的感觉。同时，产品功能也要足够简单、好用，贴合美国用户的消费心理。比如美国草坪更加开放，用户对安全性的要求很高，尤其重视防丢功能。因此，X4系列也相应推出了Find My功能，跟苹果iOS系统的Find My深度绑定，设备在超出边界时不仅会自动警报，用户也可以在其丢失后通过定位查找功能找回。硬氪：除了功能上的显性需求外，地理环境对产品设计和性能调校带来的影响大么？任冠佼：草种差异也会影响我们对产品的设计。针对不同草种特性，必须对整机的切割系统和动力系统进行专项适配优化。2024年底，未岚大陆在进入美国市场时，收到了大量用户反馈设备割不动草。我们觉得很奇怪，因为当时产品已经在欧洲卖出了十几万台，几乎没收到过类似的售后问题。调查后才发现，是一种名为圣奥古斯丁的草，在美国称圣奥古斯丁草、澳洲称其为水牛草。这种草在我们早前国内和欧洲本地的测试中从未遇到过。圣奥古斯丁草属于暖季型草，在中国仅海南有分布；而未岚大陆的产品测试主要在常州、深圳和北方地区进行。按照秦岭淮河划分，北方种植的是冷季型草，以早熟禾、高羊茅为主，设备都能正常切割。但暖季型草多为细软品种，许多高尔夫球场会使用，以佛罗里达州为主的美国南部很多人种植圣奥古斯丁草，这种草长势茂密、茎叶偏长，对设备的切割性能尤其是刀片素质要求极高。同时，这种草坪根系发达，地面易泥泞松软，机器人作业时容易下陷、啃食草根，设备必须具备充足的切割动力与通过性能，才能胜任这类环境。未岚大陆CEO任冠佼在第100万台割草机下线活动上（图源/企业）硬氪：硬件结构上做了哪些创新设计？任冠佼：我们把前驱做成可转向轮的方案，背后有多重考量。2023年，研发团队对欧洲市场做了充分调研，认为四驱价值不大，尤其是面向荷兰、德国等地形平坦的国家和地区，两驱方案足以应对。当时，许多用户反馈家里有陡坡，但经我们实地查看后发现坡度并不大，因此2025年以前，未岚大陆并没有推出四驱的产品。直到去年，四驱产品在欧美都有不错的产品表现，跟我们前几年的调研结果对比存在较大偏差。为此，团队又对每个区域状况进行了深入调查：奥地利和挪威地区的山区多、草坪坡度大，四驱卖得好；美国虽山不多，但用户追求的“Powerful”心理使得四驱结构也很受欢迎。硬氪：它跟市场上已有的四驱方案区别是什么？任冠佼：我们注意到，不少用户反映部分四驱产品存在损伤草坪的问题，这也是X4的结构方案要解决的核心痛点。从技术原理上解释，机器人本体在空间内拥有六个自由度，包含前进、平移、转向三个平动自由度，以及横滚、俯仰、偏航三个姿态转动自由度。若要实现平面三自由度全向运动，单纯采用四轮结构在原理上会产生运动干涉。因此，我们能看到有几类解决方案。第一种为阿克曼转向结构，与汽车的转向原理一致，通过底盘连杆机构控制左右车轮转角差异化，规避侧向滑移；但它的缺点是转弯半径大、倒车操控不便，在草坪作业场景适配性较差。第二种是履带式结构，可实现原地零半径转向，但履带转向过程中会产生侧滑移，容易损伤草坪。而常规普通四驱结构可看作履带的简化形态，同样存在转向侧移、伤草的问题。还有的采用在前轮搭载麦克纳姆轮、或增设横向转动结构来缓解滑移问题，都尚未从根本上解决用户痛点。可以说，四驱驱动是草坪机器人的必要配置，而转向不伤草也是用户核心刚需。如何彻底解决？我们在X4的四驱方案中，将前轮增加了两个独立的转向电机。它能实现零半径转向，轮子可以掰成内八字绕机体中心旋转，转弯半径变小；并且，设备在任何角度转向的时候，也能模拟阿克曼转向，没有横向侧移，轮子只沿滚动方向运动，从而尽量减少对草坪的横向撕扯，对草坪产生很好的保护作用。这个方案唯一的缺点是成本较高，但从用户利益出发，零半径转向和不伤草都非常重要，所以我们最后还是选了这个方案。硬氪：在导航方案上，RTK、激光雷达与视觉的多传感融合路线目前是否已经是共识？任冠佼：导航定位是最核心的技术模块，长期来看一定会趋同于“RTK+视觉+激光”的融合方案，这一点我们几年前就已明确。目前差异点在于，激光雷达还是比较贵。RTK的成本已经下降，因此未岚大陆在今年也正式将基站式RTK全部改为网络式NRTK。视觉方案本身相对简单。三者叠加仍占整机成本的较大部分。多传感融合路线（图源/企业）大趋势一致的情况下，我们的差异化在于“深耦合”，即不是分别得到RTK坐标、视觉结果、激光结果后再做融合，而是直接提取激光和视觉的原始数据进行更深度的耦合。举个例子，过去通过速度融合来推算位置，现在直接用加速度信息融合、加速度两次积分可得到设备所在的位置。这使得边角场景的处理更加稳定，用户长期使用下来会持续感受到这种提升。这些技术细节在营销层面难以传达，但长时间使用下来，产品和品牌的实力用户自然能感受到。切入商用市场硬氪：2025年，未岚大陆核心海外市场的增长情况如何？任冠佼：特定区域的规模同比去年实现了几倍的增长。比如美国，本身业务在当地的基础还比较小，翻几倍也算合理。硬氪： 商用属于另一个独立场景，跟消费级产品线的最大差异是什么？任冠佼：我们观察到，海外还有很多公共绿地、足球场、高尔夫球场、市政公园等场景同样存在割草需求，但其使用习惯与家用完全不同。比如商用场景需要编队功能；家用是一户一台，一个APP即可满足交互体验，而商用客户可能一次采购20台，甚至一个园林服务商会拥有几十台机器，他们需要一个客户端系统进行实时监控，能够远程重启或遥控脱困。设备形态层面，商用产品既可复用家用产品技术平台，同时也有必要开发专属专业机型。为此，我们内部也组建独立团队，专门负责商用业务的产品定义、设计研发与落地交付。硬氪： 为什么会在这个阶段进入商用赛道，你们看到了哪些机会？任冠佼：决定做商用业务主要基于三个方面的判断。第一，我们卖产品的时候，一直被很多渠道和代理商问到有没有更大的设备。起初我们以为是家庭用户对大草坪有需求，后来才发现，不少采购用户实际将其用在了商用场景。第二点跟我在美国的观察有关。欧洲的核心用户群体为中产阶层。通常他们在下午五点下班后，还会愿意自己动手打理庭院。相比之下，美国中产的工作压力更大，加班现象普遍。我还记得第一次在夏洛特转机时，早上六点多、机场全是通勤的上班族。这类用户宁愿多花钱请园林服务商代为打理，也不愿占用自己的休息时间。我们最早的一批美国用户中，就有大量园林服务商客户，他们购买设备是用于为客户提供割草服务。第三，在美国，主动找到我们的还有一类客户是橄榄球场、高尔夫球场等专业机构。他们的痛点十分明确，用工严重短缺，所以亟需通过机器人替代人工劳作。这几个因素叠加，让未岚大陆认定了商用业务必须要做。有趣的是，在决定做这件事后，我们团队又回头查阅了一些行业资料，发现商用业务增长很快，年营收规模也很高。这也说明了我们的判断与市场趋势一致。硬氪：商用割草机器人的快速起量能否形成势能外溢，拉动家用消费级产品的渗透普及？任冠佼：我认为会的。这也是我们的思考。比如当用户去打高尔夫时看到商用割草机，在TA回到家后，会更容易产生购买同类型产品的意愿。这对品牌认知和品类认知都有积极的推动作用。欧洲割草机器人的品类认知在15年前就已形成，包括未岚大陆在内，当今的割草机器人厂商在2022年之后所做的工作，本质上是跨越鸿沟——将用户人群从科技爱好者推动到早期采用者，他们成功接受了无边界技术的使用，让产品进入早期大众市场，实现快速爆发式增长。而美国市场至今尚未达到早期采用者阶段。尽管美国市场出现了三到四倍的增长，但其原始渗透率仍然很低。如果只做家用产品，自然增长可能需要五到十年。而如果有其他因素协同推动增长、例如商用设备的普及推动，预计未来一到两年，美国市场有可能进入爆发期。Navimow H系列割草机器人（图源/真实用户分享）硬氪：商用业务对品牌的渠道布局和合作有显著加成吗？任冠佼：目前来看，渠道端主要是通过Pro端（专业商用渠道）开始接触这一品类。我认为买手对割草机器人本身的价值认知尚未完全形成，他们只是突然意识到，这个产品能够带来利润。渠道讲究品效，当传统的手持设备份额一定程度上被新兴品类所吞没，对渠道商而言，这是个需要尽快补位的

长鑫科技更新招股书，碧桂园错失300亿小屋见大屋·2026年05月20日 10:54股东喊话：“不要再卖了”地产行业内，曾有望跑出两家明星创投企业，一家是世纪金源，另一家是碧桂园。前者现在做得顺风顺水，收获一个又一个IPO，而碧桂园因为地产业务暴雷，需要大量资金保交楼，被迫中断了在科技领域的投资“大爆发”。015月17日，计划登陆科创板的长鑫科技集团股份有限公司（以下简称：长鑫科技）更新了招股书。新招股书的财务数据显示，2026年第一季度，长鑫科技营收508亿元，同比增长719.13%；归母净利润247.62亿元，同比大增1688%，完成创历史的盈亏翻盘（2025年累计亏损366.5亿元）。可以预计，随着营收和利润的大幅增长，长鑫科技的估值将会再创新高。上市前，长鑫科技的估值已逼近3000亿元。但若按2026年一季度的财务表现，结合同类已上市企业的市盈率计算，部分机构给到长鑫科技的最新估值高达2万亿左右。长鑫科技是中国大陆规模最大的DRAM研发设计制造一体化（IDM）企业。DRAM芯片是电脑、手机、服务器内存条的核心部件。今年，存储产业一飞冲天，占全球市场份额90%以上的三星、SK海力士、美光赚得盆满钵满。而长鑫科技就是新崛起的“中国科技”，产能中国第一、全球第四。所以，长鑫科技成功上市，以及股价大涨，基本上就是板上钉钉。首先受益的是长鑫科技背后的投资人，包括安徽、北京、广东等国资，还有阿里巴巴、腾讯、美的等参与投资的企业。但有一家企业，估计会默默心痛、后悔、无奈。它就是碧桂园。022024年12月6日，碧桂园按照惯例召开一月一次的管理层会议，集团董事会主席杨惠妍在会上再次强调要全力以赴保交房。杨惠妍还说：在做重大决策时，要想象10年后自己会不会后悔。用10年的眼光去做决策，就不容易着急，不会做出让自己后悔的一些短视行为。但在这次会议之前，碧桂园已经决定转让其持有的长鑫科技1.56%股权。为了便于交易，2023年3月，碧桂园创投先将这部分股权转让给了集团旗下另外一家企业——佛山市南海区汇碧五号股权投资合伙企业。2024年12月，再次交易，将持有的长鑫科技全部股权正式转让给合肥国资背景的合肥建长股权投资合伙企业。这笔交易中，转让价格是2.22元/股，总价20亿元。碧桂园落袋为安。按照目前机构给出的长鑫科技的最新估值计算，碧桂园曾经持有的股本价值高达300亿元以上。碧桂园的投资金额为9亿元，退出时的回报约1.07倍，看似不亏，却最终与30倍回报失之交臂。用一个数据对比一下，更能理解碧桂园的“痛”：碧桂园目前市值约120亿港元，约合人民币104亿元。它在长鑫科技“丢失”的回报，接近其市值的3倍。03“失策”背后，是碧桂园的无奈。2024年，卖掉长鑫科技股权的这年，碧桂园亏了328.35亿元。权益合同销售额，直接从上一年的1743亿元，“膝斩”到了472亿元。没有利润，现金流大幅下滑，再加上“保交楼”的任务，让碧桂园陷入流动性危机。因此，它不得不把手中的非常具有增长潜力的一些科创企业股权出售。除了在长鑫科技IPO前清仓之外，碧桂园还忍痛转手了持有的蓝箭航天、卓宝科技、贝壳等企业股权。蓝箭航天是国内商业航天领域明星企业，目前正在推进IPO。根据企业公开的招股书，其IPO后估值预计将达到约750亿元。碧桂园曾经持有蓝箭航天的股权比例为11.063%，对应的最新价值超过80亿元。而碧桂园在2025年4月出售蓝箭航天时的交易价为13.05亿元——这是碧桂园错过的又一大笔财富。碧桂园还曾在2021年参与过追觅科技的C轮融资，投资金额约2亿元。‌‌‌‌但2024年6月‌，正式退出了后者的股东名单，完成清仓。错失多次“发财”机会之后，有碧桂园的个人股东开始向碧桂园喊话“不要再卖了”。‌‌‌04根据公开资料，碧桂园在创投领域的布局主要通过深圳市碧桂园创新投资有限公司开展，也就是业内人口中的“碧桂园创投”。碧桂园创投是杨惠妍家族在地产之外的另一笔主要资产。并且随着地产行业下行，创投已经成为杨惠妍家族目前最重要的资产。

AI资讯AI新闻资讯正文买家AI改图恶意“仅退款”，平台秒批仲裁逼得新店老板心凉退店发布于AI新闻资讯发布时间 :2026年5月18号 17:54阅读 :1分钟河北辛集市的无花果树种植户张先生，上个月满怀希望地在某直播电商平台开了一家新店，售卖自家培育的无花果苗。然而网店才刚刚卖出 12 单，他就意外收到了一份“仅退款”申请，对方声称收到的果苗已经干枯死亡。买家随申请附上了一张叶片全黄、毫无生机的树苗照片，以此要求退款。张先生有着多年的种植经验，他一眼就识破了这张照片是利用AI技术伪造出来的“假证据”。平台仲裁几分钟闭眼秒批，商家自证清白无门张先生表示自己一直坚持带盆发货且保证盆土湿润，从发货到买家签收全程仅仅耗时两天。按照苗木的生长规律，自然缺水枯萎至少需要五到七天，且树苗叶片会发皱下垂，绝不可能像照片中那样快速枯黄。为了维护自身权益，张先生果断驳回了申请并主动协商退货退款，不料买家随后申请了平台介入。令人心寒的是，即便张先生极力拿出专业常识辩解，平台依旧凭着那张AI伪造的图片，在短短几分钟内就判定买家胜诉。国家出手推出AI内容鉴定，业内提倡以AI治AI这笔 45 元的“仅退款”订单，让满委屈的张先生彻底对该平台失去了信心，一气之下直接启动了退店流程。这起事件也揭示了当前AI生成图像泛滥、网购恶意碰瓷难以精准鉴别的新痛点。针对日益猖獗的AI造假行为，国家反诈中心App近期已正式上线了AI内容鉴定功能，支持对图像和视频进行识别。业内安全专家也指出，未来的核心解决方案在于“以AI治AI”，通过大模型与小模型的协同判断，从源头遏制此类恶意白嫖现象。

AI资讯AI新闻资讯正文GPT-5.5才发三周，5.6已在内测！OpenAI与Anthropic补贴大战同日开打，开发者坐收渔利发布于AI新闻资讯发布时间 :2026年5月14号 14:21阅读 :1分钟三周前GPT-5.5刚落地，GPT-5.6的消息已经提前炸出来了。知名爆料人Leo透露，GPT-5.6目前开发进度已全速推进，首批内部检查点在过去几天已启动测试，下个月很可能就会正式露面。更有意思的是，有人从OpenAI内部Codex日志里扒出了rollout mapping的痕迹——大部分调用还是指向GPT-5.5，但有一条记录赫然落在了GPT-5.6身上。换句话说，Codex环境里可能已经在偷偷用它跑测试了。内部代码名也被顺带曝光:ember-alpha和beacon-alpha。网友直呼:这迭代速度，根本追不上。速度再度拉满——Codex即将上线ultrafast模式GPT-5.6还没到，OpenAI产品端已经先按捺不住了。爆料人Chetaslua同步透露，本周四OpenAI将在Codex上线"ultrafast模式"，响应速度直接提升2到3倍，专门为延迟敏感型任务设计。与此同时，Image Arena排行榜上以+242分断档领先的gpt-image-2，也在同步推进A/B测试更新。这不是第一次了。今年3月GPT-5.4发布时，/fast模式已经实现了1.5倍加速;GPT-5.3-Codex-Spark借助Cerebras芯片，更是把推理速度拉到每秒超过1000token，是普通模式的15倍。而这次的ultrafast，直接在旗舰主力模型上实现提速——不是阉割版，不是小模型凑数，是正儿八经的全尺寸加速。对开发者来说，Agent循环、长任务流水线、浏览器自动化……所有需要"等"的地方，体验将彻底不同。顺带一提，上周ChatGPT官方账号发出一张科幻感十足的图，界面上那句经典的"Ask ChatGPT"悄悄换成了"Message ChatGPT"。奥特曼丢下一句"Call me maybe"，让外界对ChatGPT集成语音通话乃至推出硬件产品的猜测，彻底烧开了。Codex vs Claude Code，补贴战正式开打硅谷最戏剧性的一幕，就在今天上演了。就在OpenAI ultrafast模式上线前夕，Anthropic率先出手——宣布从6月15日起，为付费用户每月提升50%的编程额度，覆盖Claude Agent SDK、命令行工具以及深度接入GitHub工作流的Claude Code，并同步发布了Opus4.7Fast模式，主打比Codex更快的高级推理与更流畅的长上下文编码体验。OpenAI没让人等太久。反手祭出一招重磅补贴:接下来30天内，任何想从其他平台迁移到Codex的企业，直接获赠2个月免费使用权。按Pro计划每月200美元算，相当于白送400美元。奥特曼本人亲自下场喊话，直言Codex是"目前市面上最强的AI编程产品"。仅仅几个小时后，战报出炉:2000名开发者在3小时内主动联系了OpenAI。这场怼脸硬刚，看呆了整个硅谷。开发者们倒是乐了——被两家顶级公司抢着补贴，这种好事，上哪儿说理去。飞轮已经开始自转把这两件事放在一起看，会发现一个比任何单一新闻都更值得关注的趋势正在成形。AI正在加速自我进化。GPT-5.3-Codex是OpenAI首个"参与自身训练"的模型，到GPT-5.5，OpenAI内部85%的员工每周都在用Codex写代码。GPT-5.6几乎可以确定是在GPT-5.5的深度参与下开发出来的——AI在帮OpenAI造更强的AI。与此同时，编程工具的大规模普及正在释放前所未有的工程产能。Codex已有300万周活跃用户，Claude Code的用户数同样在爆发式增长。数百万开发者把AI编程工具当成日常标配，AI生成的代码又反哺回AI的训练与部署，这个循环只会越转越快。奥特曼说过，OpenAI的目标已经不止于AGI，直接瞄准了ASI。当模型迭代速度由AI自身驱动，当AI编程工具成为AI开发的基础设施，当两家万亿级公司用"补贴战"加速这一切的普及——通往ASI的飞轮，已经开始自转了。

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文加量不加价!Anthropic 宣布 Claude Design 套餐上限翻倍发布于AI新闻资讯发布时间 :2026年5月19号 14:56阅读 :1分钟Anthropic 今日在 X 平台发布官方公告，宣布对其旗下生产力工具生态进行重大升级:不仅将视觉创作工具 Claude Design 在所有订阅套餐中的 Token 上限提升一倍，同时将开发辅助工具 Claude Code 的“快速模式（Fast Mode）”默认切换至其最新一代主力模型 Opus4.7。此次更新直击专业用户的痛点。作为 Anthropic 于今年4月推出的重磅视觉工具，Claude Design 凭借对话式生成设计、原型构建及多源导入等功能备受瞩目，但此前受制于上下文额度，复杂项目常面临中断困境;本次针对 Pro、Max、Team 及 Enterprise 全套套餐的 Token 上限翻倍，将彻底解放长上下文的连续迭代能力。与此同时，面向开发场景的 Claude Code 快速模式在接入 Opus4.7后，其交互响应速度最高可提升至原有的2.5倍，能更高效地支撑代码快速修改、实时调试等高时效性任务。在计费与资源管理上，Anthropic 保持了高度的商业克制与透明度。Opus4.7与此前 Opus4.6在快速模式下的定价完全一致，均维持在每百万输入 Token30美元、输出150美元，且全量覆盖100万 Token 的上下文窗口，费率不随上下文长度而变动。此外，两款模型在快速模式下共享相同的速率限制池，企业与开发者无需担心额外的配置成本。

AI资讯AI新闻资讯正文AI眼镜进入城市服务场景，蚂蚁GPASS上线城市伴游功能发布于AI新闻资讯发布时间 :2026年5月19号 14:58阅读 :1分钟继乌镇峰会期间“桐小乌”在乐奇AI眼镜上的应用落地后，蚂蚁自研的智能终端可信连接技术框架GPASS正进一步拓展应用边界。近期，基于该技术的城市伴游功能已在乐奇AI眼镜新版本中正式上线。用户开启伴游模式后，可通过乐奇AI眼镜实现边行走、边获取信息的实时交互体验。系统可根据用户所处位置，对周边景点进行讲解，并支持即时问答、导航、第一视角内容记录与分享，在减少手动操作的同时提升出游的连贯性与沉浸感。作为首个接入该服务的城市文旅官方智能体，“杭小忆”的入驻标志着“镜游杭州”项目进入实际运行阶段。该项目由杭州文旅、支付宝与乐奇联合打造，旨在探索AI与城市文旅服务的结合方式。后续，城市伴游功能还将逐步接入更多城市专用文旅智能体。从功能体验来看，本次升级主要体现在三个方面:一是提升信息获取的深度。在传统导航与导览基础上，系统可结合用户附近的建筑、景观等所在环境，补充建筑、人文历史知识及文化相关内容，帮助用户更系统地理解所见所闻。二是增强多功能协同能力。对话、拍摄、翻译、导航、支付及音频播放等功能可通过语音进行切换与调用，即使处于伴游模式下，也不影响其他功能的使用，整体体验更加连贯。三是优化交互方式。基于语音、图像与位置信息的综合识别能力，系统可在特定场景下主动提供提示与讲解，无需唤醒词即可开启对话，过滤环境噪音和无关意图。从以往的被动响应，转向更加连续的交互体验。这一创新体验的背后，是蚂蚁GPASS智能终端可信连接技术框架的强力支撑。GPASS通过语音+图像+位置的多模态意图理解，精准捕捉用户需求;在眼镜端实现流畅的对话与导航交互;并确保了与“杭小忆”等城市文旅智能体之间的可信连接，让服务既智能又安全。业内人士认为，城市伴游功能的上线，显示出AI眼镜正从单一设备形态向具体应用场景延伸。以文旅为代表的服务领域，有望成为相关技术规模化落地的重要方向之一。

AI资讯AI新闻资讯正文Alexa+ 上线按需播客功能 想听啥自己点！发布于AI新闻资讯发布时间 :2026年5月19号 15:14阅读 :1分钟亚马逊的 Alexa+ 又开始「卷」新技能了。本周一，Amazon 正式宣布 Alexa+ 迎来重磅更新:支持按需生成播客节目。这项名为「Alexa Podcasts」的新功能，即日起面向美国用户开放测试。用亚马逊的话说，它的目标很直接——「把你感兴趣的任何话题，几分钟内变成一期专属播客」。使用门槛低到几乎没有:用户只需对 Alexa+ 说一句「帮我做个关于 XX 的播客」，不用上传资料、不用写脚本、更不用提前规划。AI 会自动研究主题、搜集信息、生成内容大纲，并快速反馈这期节目大概会讲什么。用户还能随时调整时长、语调、内容侧重，像搭积木一样定制属于自己的音频内容。确认无误后，Alexa+ 会用 AI 生成的主播声音完成播报。节目制作完成后，用户会通过 Echo Show 设备或 Alexa App 收到通知;所有生成的播客也会自动保存在 App 的「Music」和「More」栏目中，方便随时回听、反复回味。这不只是「多了一个功能」那么简单。它标志着 Alexa+ 正在从「回答问题、控制家电」的传统语音助手，悄然进化为「懂你兴趣、能创内容」的个性化 AI 创作伙伴。当语音交互遇上内容生成，智能音箱的边界，正在被重新定义。当然，新能力也伴随新争议。AI 主播的声音是否足够自然?自动生成内容的准确性如何保障?会不会对传统播客创作者形成冲击?这些都是行业正在热议的问题。尤其当话题涉及新闻事件或专业领域时，用户对信息可靠性的要求会更高。对此，亚马逊也提前「打补丁」:强调已与多家权威新闻机构建立合作，包括美联社、路透社、《华盛顿邮报》、《时代周刊》、《福布斯》、Politico、《今日美国》等主流媒体，以及全美200多家地方报纸。通过这些合作，Alexa+ 能接入实时、可信的信源，从底层提升内容质量与事实准确度。更值得留意的是，播客只是起点。亚马逊透露，团队正在探索更多个性化 AI 音频形态:比如基于用户偏好定制的每日新闻简报，或结合个人文档、共享信息生成的专属内容。当「你的数据+AI 理解+场景需求」真正打通，语音交互的体验，或许会从「被动响应」迈向「主动陪伴」。

AI资讯AI新闻资讯正文加量不加价！Anthropic旗下Claude两款主力AI工具迎来重磅升级发布于AI新闻资讯发布时间 :2026年5月19号 15:18阅读 :1分钟大模型独角兽Anthropic近日在社交平台宣布，对其现有的AI工具链进行大刀阔斧的性能升级。本次调整的核心聚焦于广受好评的视觉创作工具以及专门面向程序员的代码利器，旨在通过大幅度放宽限制来提升用户的实际创作效率。在此次更新中，上个月刚推出的视觉创作工具Claude Design迎来了额度暴涨，其在Pro、Max、Team以及Enterprise在内的全部订阅套餐中，Token使用上限直接实现翻倍。这一举措彻底解决了创作者在处理复杂设计项目时，因额度限制而不得不将任务强行分段的痛点。代码编辑器默认切换新模型除了视觉工具的额度翻倍，针对软件开发者的命令行代码工具Claude Code也迎来了重要调整。系统内置的快速模式（Fast Mode）现在已将底层大模型默认切换为更先进的Opus 4.7。开发人员只需在终端中简单地输入“/fast”指令，即可一键激活该高速配置。官方测试数据显示，在该模式下，新模型的运行响应速度可以直接飙升至此前的2. 5 倍。高通吐量主打交互式高频调优由于具备极高的吞吐效率，该快速模式主要服务于高频、即时性的交互式工作。无论是快速迭代代码修改、实时捕获并调试系统漏洞，还是处理时间要求极度紧张的突发开发任务，它都能提供有力的算力支持。在费用收取方面，Opus 4. 7 在快速模式下的定价与此前的4. 6 版本保持完全一致。其每百万输入Tokens的收费标准为 30 美元，输出则为 150 美元，且全面覆盖 100 万Tokens的超长上下文窗口。

AI资讯AI新闻资讯正文光学模组成智能眼镜「隐形胜负手」！韩国初创LetinAR获 1850 万美元融资发布于AI新闻资讯发布时间 :2026年5月19号 17:16阅读 :1分钟想象一下：你正以 160 公里的时速骑摩托车飞驰，前方路面上突然「浮现」出一个导航箭头，精准指示转弯位置。没有手机、不用低头看仪表盘，所有信息都通过头盔里一枚拇指大小的镜片，自然融入你的视野。这不是科幻概念片，而是最快今年就会在欧洲道路上落地的真实场景。而它，只是智能眼镜进化浪潮中的一个早期缩影。过去几年，科技巨头们早已悄悄（有时也不那么悄悄）押注这条赛道：Meta 从 2023 年起持续迭代 AI 赋能的 Ray-Ban 眼镜；谷歌正在构建 Android XR 生态；苹果入局传闻不断；上周还有消息称，三星将于 7 月在伦敦 Galaxy Unpacked 活动上，联手 Gentle Monster 发布首款AI 智能眼镜。中国阵营同样动作频频，华为、阿里、小米等都在加速布局。市场数据印证了这股热潮。据 Omdia 统计，2025 年全球 AI 眼镜出货量飙升至 870 万台，同比增长超 300%；分析师预测，今年这一数字将突破 1500 万台。当整机品牌争相卡位，上游供应链也在悄然卡位。韩国初创企业 LetinAR 就是其中之一——这家成立近十年的公司，专注攻克智能眼镜最核心的难题：如何让光学模块既轻薄省电，又能呈现清晰锐利的画面。最近，LetinAR 刚完成 1850 万美元新一轮融资，投资方包括韩国开发银行、乐天创投等，为 2027 年韩国 IPO 做准备。有趣的是，其早期投资方 LG 电子，据当地媒体报道也已启动自研 AI 智能眼镜项目——这侧面说明，韩国消费电子龙头对这一品类的重视程度正在快速升温。LetinAR 的两位联合创始人 Jaehyeok Kim 与 Jeonghun Ha 从高中就是好友，2016 年共同创立公司。他们不造整镜，而是专注做「让眼镜能真正戴出门」的关键部件：光学模组。「智能眼镜是下一个平台级入口，」Kim 表示，「而光学模组是最难啃的骨头——厂商需要比现有方案更薄、更轻、更省电，同时成像不能打折。」他们的解决方案叫 PinTILT：通过精密排布镜片内部的微光学元件，让光线精准「瞄准」人眼，而非像传统方案那样向四周散射。可以打个比方：传统波导（waveguide）技术像一台客厅电视，光线铺满整个镜片，但真正进入眼睛的只有一小部分，大量光能被浪费，导致画面偏暗、耗电飞快；而另一种反射式方案（birdbath）虽能高效导光，但结构臃肿，很难塞进日常眼镜的框架里。PinTILT 的思路是「只做有用的光」——只聚焦能进入瞳孔的那部分光线，并精细调控每个微元件的角度。结果是：在更薄更轻的形态下，实现更亮的成像、更低的功耗。在智能眼镜这个「每克重量、每小时续航都关键」的品类里，这恰恰是行业集体攻坚的核心命题。目前，LetinAR 的模组已实现量产交付，客户包括日本 NTT QONOQ Devices、原东芝客户端解决方案业务 Dynabook 等，积累了规模化制造经验。公司也在与多家科技巨头洽谈下一代 AI 眼镜的联合研发，只是暂未透露具体名字。其中一个高要求客户是瑞士深科技公司 Aegis Rider（源自苏黎世联邦理工学院计算机视觉实验室）。他们正在开发一款 AI 驱动的 AR 摩托车头盔，能将导航、车速、安全预警等信息「锚定」在真实路面上，而非简单悬浮在面罩上——仿佛信息本就「画」在前方的世界里。而 LetinAR 的光学模组，正是这套系统的视觉核心。该产品计划 2026 年面向欧盟及瑞士市场推出。本轮融资后，LetinAR 累计募资达 4170 万美元。Kim 表示，资金将主要用于产能扩张，以应对 AI 眼镜市场从早期极客向大众消费跃迁的关键节点。「硬件设备，比如 AI 眼镜，将是把人工智能真正带入日常生活的下一层载体。」当算力、算法、光学、交互在方寸镜片上交汇，智能眼镜的竞赛，早已不是「能不能做出来」，而是「能不能戴得久、看得清、用得顺」。而像 LetinAR 这样在底层光学环节死磕的公司，或许正握着打开下一代人机交互大门的钥匙。

AI资讯AI新闻资讯正文Google AI Studio 移动应用正式登陆 Google Play，支持预注册发布于AI新闻资讯发布时间 :2026年5月19号 17:49阅读 :1分钟近日，Google 正式推出 AI Studio 独立移动应用，目前已在 Google Play 商店上线并开放预注册。这标志着 Google 将其 Gemini 驱动的 AI 开发平台从桌面端扩展至移动设备，用户可随时随地通过自然语言快速构建应用、工具和小游戏。应用强调“一次对话成就下一个创意”，支持语音或文字输入描述想法，AI 即可自动生成原型、创建专业图标并设计界面。无需编程基础，用户可轻松制作个性化膳食规划器、会议笔记工具或聚会游戏等实用小程序。核心功能包括:

• 社区画廊与一键 Remix:浏览他人模板并快速修改;- 后台构建与进度通知;- 手机与桌面项目无缝同步;- 生成的应用可通过链接跨平台即时分享，无需安装。该应用在 Google I/O 大会前夕推出，体现了 Google 推动 AI 工具普惠化的战略意图，显著降低了普通用户和非专业开发者的创作门槛。行业分析认为，这将进一步激发移动端 AI 创意工具的爆发式增长。

目前应用评级为全年龄段，由 Google LLC 开发。用户可通过 Google Play 搜索“AI Studio”进行预注册。正式版发布后，其实际使用体验和生态表现备受期待。

AI资讯AI新闻资讯正文视觉大模型遭遇滑铁卢：首个中国古文字OCR评测基准开源发布于AI新闻资讯发布时间 :2026年5月19号 17:55阅读 :1分钟顶尖的人工智能不仅要能看懂屏幕上跳动的现代代码，也需要读懂三千年前龟甲上的刻痕。据OSCHINA报道，腾讯混元大模型、SSV数字文化实验室等机构联合多所高校与故宫博物院，正式推出了“Chronicles-OCR”。这是业界首个完整覆盖汉字“七体之变”演化轨迹的中国古文字感知评测基准。为了真实反映大模型的识别能力，该数据集由领域专家进行了多层级交叉标注，包含 2800 张严格平衡的高质量图像。针对甲骨文、金文、篆书等古早字体，团队采用了单字级的精细标注；而对于隶、楷、行、草等成熟字体，则采用了保留原始阅读顺序的序列级转写。主流视觉模型全军覆没项目团队基于该基准设计了四个层层递进的核心任务，严格将大模型的“视觉感知”与“语义推理”解耦开来。在对包括GPT-5、Gemini 3.1 Pro、Claude Opus 4. 7 在内的 28 个主流多模态大语言模型进行评测后，结果却令人大跌眼镜。在面对缺乏现代版式先验的古早字体时，主流大模型在端到端检测任务中全军覆没，细粒度识别的最高准确率也仅有27.1%。令人意外的是，实验表明此时开启大模型的推理（Reasoning）模式，反而会放大感知的不确定性，导致识别表现进一步下降。揭示微观笔画识别短板评测还发现，在进行字体分类时，目前的视觉大模型更容易去识别载体的纹理材质，而不是去判别微观的笔画风格。这意味着今天最顶尖的AI模型，在面对中国传统古文字时，依然还远远没有做到真正的“读懂”。汉字从殷墟甲骨一路演化至今，每一笔一画都承载着文明的连续性。Chronicles-OCR的开源不回避这一技术现实，它通过清晰可见的差距，为未来视觉大模型从简单的“识字”走向深度“读史”提供了明确的优化方向。

吐司APP的简介吐司是腾讯推出的AI原生应用创作平台，用户只需输入想法，AI自动生成完整的安卓原生应用。平台支持智能功能梳理、精细化编辑打磨，能一键导出APK进行本地安装与分享。内置灵感广场汇聚海量优质模板，涵盖社交助手、照片精选、学习工具等多元场景，支持用户共创。无需编程基础，零基础用户也能快速将奇思妙想转化为可实际使用的专属APP，轻松实现个性化应用开发。吐司APP的主要功能AI自动生成应用：输入自然语言想法，AI自动梳理功能逻辑并生成完整的安卓原生应用，无需编写代码。精细化编辑打磨：支持对生成的应用进行细节调整与优化，满足个性化需求。一键导出APK：生成完成后可直接导出APK安装包，支持本地安装、测试与社交分享。灵感广场：内置海量精选应用模板与创意案例，如社交助手、照片精选、学习工具等，帮助用户快速获取创作灵感。应用共创：支持用户基于现有模板进行二次创作与共创，降低应用开发门槛。多场景覆盖：支持生成考研题库、运势工具、普通话测试、宠物健康日志、梦境记录、食材新鲜度检测、万年历等多样化实用应用。类似于吐司APP的应用TRAE SOLO APP字节跳动推出移动端的 AI 智能工作助手WindsurfCodeium公司推出的AI编程工具心响APP百度推出的手机端超级智能体应用Replit APPAI开发工具应用Trae字节跳动推出的免费 AI IDEKiro亚马逊公司推出的 AI IDE暂无评论再想想发表评论暂无评论…

LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台AI工具1天前发布AI小集03LibTV团队版是什么LibTV 团队版是LibTV推出的专业级 AI 视频协作工作台，专为 AI 短剧、电影、TVC 制作团队设计。平台提供团队空间、实时画布协作、资产库沉淀、权限管理、成本管控等功能，解决 AI 视频创作中资产散落、协作低效、人员流动导致资产流失等痛点，支持规模化、工业化的 AI 内容生产。LibTV团队版的主要功能专属团队空间：统一沉淀分镜、脚本、镜头资产与生成记录，历史项目可迁移，避免核心资产散落在个人账号与本地文件中。画布协作：导演、美术、剪辑等角色可在同一块画布实时协同，同步节点状态与镜头修改，减少版本错乱。团队资产库：角色主体、风格模板、运镜方案、提示词工作流与参考素材支持统一沉淀，成熟视觉体系可持续复用。权限管理：支持项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作场景。成本管理：共享积分池、成员额度分配与消耗查看，团队管理者可统一管理生成资源、席位与项目成本。资产交接：成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免创作资产随人员流动中断。如何使用LibTV团队版创建团队空间：访问LibTV官网升级至团队版，建立专属团队项目空间，设置团队名称与基础信息。配置成员与权限：邀请导演、编剧、美术、剪辑等成员加入，按岗位分配访问、编辑、复制或发布权限。沉淀与迁移资产：将历史个人项目迁移至团队空间，或新建项目并导入参考素材、角色设定与风格模板。画布协作创作：成员在统一画布内协同推进分镜设计、镜头生成与修改，实时同步进度与反馈。复用团队资产库：调用已沉淀的角色主体、风格模板与提示词工作流，保持系列化内容风格一致。管控成本与交接：通过共享积分池分配生成额度，监控消耗；成员变动时资产自动保留，确保项目连续性。LibTV团队版的核心优势工业化协作：从个人创作升级为团队制片管理，支持长期项目、多岗位协同与规模化内容生产。资产持续复用：成熟的角色设定、风格与提示词工作流可沉淀为团队资产，降低新人上手与重复调试成本。实时协同提效：多人画布协作减少版本错乱与重复生成，围绕同一套镜头语言与素材体系推进项目。成本与风险可控：共享积分池与席位管理适配多项目并行需求，权限分级与资产交接机制保障核心创作安全。LibTV团队版的同类竞品对比对比维度LibTV 团队版快手可灵（企业版/团队功能）协作模式多人实时画布协作，导演/美术/剪辑等角色可在同一块画布内同步节点状态、生成进度与镜头修改团队空间内创作接力，支持桌面端（Win/Mac）同步，减少文件下载与搬运损耗资产沉淀角色主体、风格模板、运镜方案、提示词工作流、参考素材与工具箱统一沉淀，支持成熟视觉体系持续复用设立「共享资产库」，支持个人空间向团队空间一键转移灵感值、创作资产与主体资产，实现”一人贡献，全员复用”权限管理项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作所有者、管理员、协作者三重角色，灵活配置查看/编辑/移除权限，保障核心资产安全成本管理共享积分池、成员额度分配与积分消耗查看，支持多项目并行下的生成资源统筹支持灵感值跨空间转移，团队共享生成资源，但精细化成本管控功能相对较弱资产交接成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免资产随人员流动中断个人资产可一键导入团队空间，但成员退出后的自动资产交接机制未明确强调核心场景聚焦 AI 短剧、AI 电影、AI TVC 的工业化制片流程，强调长期项目管理与跨岗位协同覆盖 AI 视频生成、智能分镜、广告 Demo、影视特效等，强调快速创作与素材复用产品定位AI 视频专业级协作工作台，面向制片管理流程AI 视频生成平台的团队功能扩展，面向创作者联盟与工作室LibTV团队版的应用场景AI 视频工作室：适合长期运营 AI 短剧、动画、MV 的创作团队，通过资产库统一沉淀角色与风格，实现系列化生产。短剧承制方与 MCN：适配高频、多项目并行的工业化内容生产，导演与美术可同时在线推进分镜与镜头生成。品牌市场部与代理公司：支持品牌方、代理公司与外部创意团队跨组织协作，共同管理 AI TVC 与 Campaign 项目。AI 电影与大型制作团队：支持长篇叙事、系列化 IP 与复杂制作流程，导演组、制片组、美术组与后期团队跨岗位协同。企业内容与创意部门：适配培训、营销、产品与全球化传播等标准化内容生产，便于企业进行长期生成资源规划。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇PPT Master - 开源 AI PPT 生成标准化工作流下一篇腾讯Marvis - 腾讯应用宝推出的操作系统层个人 AI 助手相关文章LongCat-Flash-Omni – 美团开源的实时交互全模态大模型AI小集2InternSVG – 上海AI Lab推出的统一 SVG 建模套件AI小集2Thinking Claude – 17岁高中生推出的神级Prompt工具，AI 思维更接近人类AI小集4Granite-Docling-258M – IBM推出的轻量级视觉语言模型AI小集4ERNIE-4.5-21B-A3B-Thinking – 百度推出的思考模型AI小集2EMOVA – 华为诺亚方舟联合多所高校共同推出的多模态全能处理模型AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付4小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版4小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架4小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准6小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型8小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

Grok Build – xAI 推出的终端原生 AI 编程智能体AI工具2天前发布AI小集03Grok Build是什么Grok Build 是 xAI 推出的终端原生 AI 编程智能体（Agentic CLI），面向专业软件工程与复杂编码任务。基于 Grok 4.3 beta 的 200 万 token 超长上下文，支持 Plan Mode 计划审批、最多 8 个并行子智能体协作及 Arena Mode 自动方案排序，可在本地直接读写文件、执行 Shell 命令、操作 Git 并构建完整应用。采用本地优先架构，源代码与敏感数据不上传云端，保障企业级安全。Grok Build的主要功能Plan Mode（计划模式）：复杂任务执行前自动生成结构化步骤计划，用户可审批、评论或修改，所有代码变更以 diff 形式透明展示，避免”黑箱”操作。并行子智能体：最多支持8 个并发 AI 智能体同时工作，可分别处理代码分析、CI/CD 检查、性能优化、文档检索等不同子任务。Arena Mode（竞技场模式）：多个智能体针对同一任务生成不同代码方案，系统自动评估并排序，开发者直接选择最优解，无需手动对比。超长上下文窗口：基于 Grok 4.3 beta，支持200 万 token上下文，可一次性加载整个大型代码库进行跨文件复杂任务。本地优先架构：源代码、凭证和敏感数据不上传云端，完全在本地执行，支持离线环境和气隙网络，保障企业级代码安全。终端原生 CLI + Web 双入口：支持本地命令行交互，也可通过浏览器界面远程操作，满足不同工作流习惯。VS Code 扩展集成：提供官方 VS Code 插件，在 IDE 内直接调用 Grok Build 能力，实现 GUI adjacent 工作流。Git 与开发服务器集成：可自动管理 Git 仓库、启动本地开发服务器、内置浏览器浏览网页、执行 Shell 命令。插件与开放生态：支持 MCP 服务器、AGENTS.md 团队规范文件、Hooks 自定义动作、ACP（Agent Client Protocol）协议，便于集成现有 DevOps 流水线。Headless 脚本模式：通过-p标志支持非交互式执行，可直接嵌入 CI/CD 流水线或自动化脚本。BYOK（自带 API Key）：支持使用自有 xAI API Key 进行本地计费，灵活控制成本。如何使用Grok Build一键安装 CLI：在终端执行curl -fsSL https://x.ai/cli/install.sh | bash完成本地安装，支持 macOS、Linux 和 Windows 三大平台。启动交互式终端：安装完成后输入grok>进入交互式命令行界面，可直接用自然语言描述编程需求。登录账号：在终端内执行grok> /login登录 xAI 账号，解锁高级功能与订阅权益。配置 API Key（可选）：如需使用自有额度，执行grok> /api-key add xai添加个人 xAI API Key（BYOK 模式）。使用 Plan Mode 处理复杂任务：输入需求后，Grok Build 会自动生成结构化执行计划，展示步骤与预期变更 diff，用户可审批、评论或修改后再执行。利用并行子智能体：在复杂项目中，系统会自动或手动触发最多 8 个并发子智能体，分别处理代码分析、测试生成、文档检索等子任务。通过 Arena Mode 选优：对同一任务让多个智能体生成不同方案，系统自动评估排序，你只需选择排名最高的方案应用。Web 端远程操作：访问build.grok.com或x.ai/cli，在浏览器中打开 Grok Build 界面，无需本地安装即可使用。VS Code 集成工作流：安装 Grok Build 的 VS Code 扩展，在编辑器内直接调用智能体能力，边写代码边让 AI 协助修改。Headless 自动化执行：在脚本或 CI/CD 流水线中使用-p标志运行非交互式命令，实现自动化构建与部署。查看帮助与命令：在终端输入grok> /help查看所有内置命令与快捷操作指南。Grok Build的官网地址项目官网：https://x.ai/news/grok-build-cliGrok Build的核心优势本地优先，代码绝不上云：源代码、凭证及敏感数据完全在本地执行，无需上传至 xAI 服务器，适合处理专有代码库、金融及政务等受监管行业项目。200 万 token 超长上下文：基于 Grok 4.3 beta，可一次性将整个大型代码库载入内存，跨数十个文件进行复杂重构与多文件联动修改，大幅减少上下文丢失。8 路并行子智能体：最多同时调度 8 个并发 AI 智能体分工协作，分别处理代码分析、测试生成、CI/CD 检查等子任务，显著缩短复杂项目耗时。Arena Mode 自动选优：同一任务由多个智能体生成不同方案，系统自动评估排序并给出排名，开发者直接选择最优解，告别手动对比试错。Plan Mode 透明可控：执行前先输出结构化步骤计划与 diff 预览，用户可审批、评论或修改后再执行，每一步变更清晰可见，避免”黑箱”风险。极致 API 价格竞争力：输入 token 仅$0.20 / 百万 tokens，相较 Claude Code 和 Codex CLI 大幅降低高频使用成本。终端原生 + 多入口覆盖：支持 CLI 命令行、Web 浏览器、VS Code 扩展三种工作入口，既满足极客式终端操作，也兼容 IDE 集成习惯。Grok Build的同类竞品对比对比维度Grok Build (xAI)Claude Code(Anthropic)Codex CLI(OpenAI)产品定位终端原生 AI 编程智能体（Agentic CLI）终端 AI 编程助手云端 AI 编程智能体发布状态2026.5 早期 Beta已正式发布已正式发布底层模型Grok 4.3 beta / grok-code-fast-1Claude 4 (Sonnet/Opus)Codex (基于 GPT-4o/o3)上下文长度200 万 token20 万 token (Claude 4)约 20 万 token并行智能体最多 8 个并发子智能体单智能体为主单智能体为主方案评估Arena Mode 自动排序选优手动对比或迭代手动对比或迭代执行计划Plan Mode（先计划后审批执行）直接执行，可审查 diff直接执行，可审查 diff数据安全本地优先，代码不上云云端执行，代码上传 Anthropic云端执行，代码上传 OpenAI离线支持支持完全离线/气隙网络需联网需联网SWE-Bench70.8%约 72–75%约 70%API 定价$0.20 / 百万 tokensPro 订阅 + 按量计费Credits 按量计费订阅门槛SuperGrok Heavy ($300/月)Claude Pro / Team 订阅ChatGPT Pro / 开发者额度入口形态CLI + Web + VS Code 扩展CLI + IDE 集成CLI + ChatGPT 移动端自动化集成Headless 模式 (-p 标志) + CI/CD有限脚本支持有限脚本支持核心差异多智能体并行 + 本地优先 + 超大上下文模型推理能力强，生态成熟与 ChatGPT 生态深度整合Grok Build的应用场景大型代码库重构与现代化：基于 200 万 token 超长上下文一次性加载整个项目，跨数十个文件进行安全重构、依赖升级或架构迁移，避免上下文截断导致的遗漏。复杂全栈应用从零构建：通过 Plan Mode 自动生成项目结构、技术选型、依赖安装和文件生成计划，审批后直接执行，快速搭建完整应用原型。敏感行业专有代码开发：本地优先架构确保源代码、凭证和敏感数据不上传云端，适用于金融、政务、医疗、军工等对数据主权有严格要求的场景。离线/气隙网络环境编程：无需持续联网即可在本地执行全部任务，适配高安全隔离网络、内网开发机或无外网访问权限的工作环境。多维度自动化代码审查：启动多个并行子智能体分别检查代码安全性、性能瓶颈、CI/CD 合规性和文档完整性，一次性输出综合审查报告。技术方案快速对比与选型：使用 Arena Mode 针对同一需求生成多种实现方案，系统自动评估排序，帮助团队在短时间内确定最优技术路径。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇OpenHuman - 开源桌面端 AI 助手，能主动感知工作上下文下一篇Violin - 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具相关文章Qwen3-Coder-Flash – 阿里通义开源的高性能编程模型AI小集3Cluely – AI桌面助手，用透明窗口形式提供实时辅助AI小集2墨问 – 专为创作者设计的AI社交笔记工具AI小集53DeepEyesV2 – 小红书开源的多模态智能体模型AI小集2presti – 专注于针对室内家具领域的AI图像生成工具AI小集2Joyland – 西湖心辰推出的沉浸式AI聊天机器人平台AI小集5暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付4小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版4小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架4小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准6小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型8小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文AI工具3天前发布AI小集05OpenHuman是什么OpenHuman 是 tinyhumansai 团队推出的开源个人 AI 超级智能助手，定位为”Your Personal AI super intelligence”，主打私有、简单且极其强大。非传统聊天机器人，是能主动感知用户工作上下文的桌面级 AI Agent，通过每 20 分钟自动同步 118+ 第三方应用数据，构建本地持久记忆，让 AI 真正”懂你”。项目采用 Rust + TypeScript + Tauri 构建，支持 macOS、Linux 和 Windows 平台。OpenHuman的主要功能桌面吉祥物交互：具备视觉形象的桌面 AI 助手，支持语音输入（STT）与 ElevenLabs 语音输出（TTS），吉祥物可实时唇形同步，能作为参会者加入 Google Meet 并实时记录会议。118+ 第三方应用一键集成：通过 OAuth 快速接入 Gmail、Notion、GitHub、Slack、Linear、Jira、Stripe、Calendar、Drive 等主流工具，无需编写任何连接代码。每 20 分钟自动同步（Auto-fetch）：核心引擎每 20 分钟遍历所有活跃连接，自动拉取最新数据进入本地记忆系统，让 Agent 在次日早上已掌握前日全部上下文。Memory Tree 层次化持久记忆：将同步数据规范化为 ≤3k-token 的 Markdown chunks，评分后折叠存入层级摘要树，存储于本地 SQLite，实现跨会话的长期记忆。Obsidian Vault 原生兼容：同步数据同时输出为 Obsidian 可读的.md文件，用户可在 Obsidian Graph View 中浏览和编辑 AI 维护的知识网络。TokenJuice 智能 Token 压缩：在数据进入 LLM 前自动压缩（HTML 转 Markdown、缩短长 URL、移除非 ASCII 字符等），最高可降低 80% 的 API 调用成本与延迟。内置模型路由（Model Routing）：根据任务类型自动分配轻量模型、强推理模型或视觉模型，一个订阅覆盖多种模型能力，无需管理多个 API Key。完整工具链开箱即用：内置 Web 搜索、网页抓取（Web Scraper）、代码工具集（文件系统、Git、Lint、Test、Grep），无需额外安装插件即可使用。本地优先与隐私保护：所有工作流数据本地加密存储，支持通过 Ollama 接入本地模型实现完全离线运行，数据不上传云端。可选 agentmemory 共享后端：可配置与 Claude Code、Cursor、Codex 等工具共享同一个持久记忆存储，实现跨工具的知识互通。消息通道集成：支持 inbound/outbound 消息收发，打通用户已有的沟通渠道，工作流数据始终保存在本地设备。如何使用OpenHuman官网下载安装：访问OpenHuman的官网 https://tinyhumans.ai/openhuman 下载对应平台的 DMG（macOS）或 EXE（Windows）安装包，双击安装后即可启动。命令行快速安装（macOS / Linux x64）：在终端执行curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash自动完成下载与安装。命令行快速安装（Windows）：在 PowerShell 执行irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex完成一键安装。一键连接工作应用：打开应用后，通过 OAuth 一键接入 Gmail、Notion、GitHub、Slack、Calendar、Drive 等 118+ 第三方工具，授权后每 20 分钟自动同步数据。与桌面吉祥物交互：点击桌面上的吉祥物形象，通过语音或文字与其对话，可勾选 “Speak replies” 让 AI 语音播报回复。查看自动同步的记忆数据：在应用内查看 Agent 已同步的邮件、任务、代码仓库等上下文，AI 会自动压缩整理为 Memory Tree 供后续对话调用。Obsidian 知识库浏览：同步数据自动写入本地 Obsidian-compatible Vault，打开 Obsidian 即可在 Graph View 中可视化浏览 AI 维护的知识网络，也可手动编辑.md文件。使用内置工具集：直接调用开箱即用的 Web 搜索、网页抓取、代码工具（文件系统、Git、Lint、Test、Grep）等能力，无需额外安装插件。接入本地模型（可选）：如需完全离线运行，通过 Ollama 在本地部署模型，并在设置中切换至本地模型后端，实现零云端依赖。OpenHuman的项目地址官网地址：https://tinyhumans.ai/openhumanGitHub 仓库：https://github.com/tinyhumansai/openhumanOpenHuman的核心优势UI 优先分钟级上手：干净的桌面体验，无需终端配置，从安装到可用 Agent 仅需几分钟，区别于终端优先的竞品。上下文在分钟级而非周级：传统 Agent 需要数周学习用户习惯，OpenHuman 通过一次同步即可掌握邮箱、日历、仓库、文档、消息的全量压缩上下文。一站式订阅，零 API 蔓延：一个订阅覆盖所有功能，无需额外购买多个模型 API Key。Obsidian 知识库原生兼容：同步数据自动写入 Obsidian-compatible Vault，用户可在熟悉笔记界面中浏览和编辑 AI 维护的知识图谱。Rust + Tauri 高性能架构：内存占用远低于 Electron 方案，69.7% Rust + 26.1% TypeScript 构建。开源可审计：GPL-3.0 协议，代码完全透明，隐私敏感用户可自行审计数据流向。OpenHuman的同类竞品对比对比维度OpenHumanOpenClawHermes Agent开源协议✅ GPL-3.0✅ MIT✅ MIT上手难度✅ UI 优先，分钟级上手⚠️ 终端优先配置⚠️ 终端优先配置使用成本✅ 单一订阅 + TokenJuice 压缩⚠️ 自备模型 API⚠️ 自备模型 API记忆能力🚀 Memory Tree + Obsidian Vault + 可选 agentmemory 后端⚠️ 依赖插件扩展✅ 自学习观察第三方集成🚀 118+ 应用 OAuth 一键接入⚠️ 需自行配置连接⚠️ 需自行配置连接自动数据同步✅ 每 20 分钟自动 fetch 至本地记忆🚫 无自动同步🚫 无自动同步API 管理✅ 一个账户覆盖所有模型🚫 自备 Key🚫 多供应商管理模型路由✅ 内置自动分配（轻量/推理/视觉）⚠️ 手动切换⚠️ 手动切换原生工具集✅ 代码 + 搜索 + 抓取 + 语音✅ 代码工具✅ 代码工具数据隐私🚀 本地优先，数据不上云依赖自托管配置依赖自托管配置OpenHuman的应用场景跨应用项目上下文整合：同时连接 GitHub、Linear、Jira、Slack 后，让 AI 自动对比代码 PR、任务看板和团队消息，一键汇总项目最新进展，无需手动切换多个工具。智能邮件筛选与摘要：授权 Gmail 后，AI 每 20 分钟自动扫描收件箱，按项目或优先级分类摘要，快速定位需要回复的重要邮件，过滤掉低价值信息。AI 会议助手：让桌面吉祥物以真实参会者身份加入 Google Meet，实时记录会议内容、提取 action items，并基于记忆库提供参会者背景信息辅助沟通。代码仓库智能问答：同步 GitHub 仓库数据后，直接向 AI 询问代码逻辑、历史变更原因、PR 评论争议点，Agent 基于完整代码上下文给出精准回答。Obsidian 个人知识库构建：将邮件、文档、聊天记录自动压缩归档为 Obsidian-compatible 的.md文件，在 Graph View 中可视化呈现知识关联网络，打造 Karpathy 风格的 LLM 知识库。多工具工作流自动化：连接 Notion、Calendar、Drive、Stripe 等工具后，AI 自动关联”日历事件→相关文档→对应邮件→财务记录”，形成完整的项目上下文链。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Intern-S2-Preview - 上海 AI Lab 开源的科学多模态大模型下一篇Grok Build - xAI 推出的终端原生 AI 编程智能体相关文章OpenPaper – AI学术研究助手，支持论文内容深度阅读AI小集3SlideChat – 上海AI Lab联合多所高校和机构推出的视觉语言助手AI小集2MME-CoT – 港中文等机构推出评估视觉推理能力的基准框架AI小集2Seed-ASR – 字节跳动推出的AI语音识别模型AI小集3ReadPaper – AI科研工具，高效阅读、翻译、管理学术论文AI小集4话袋AI笔记 – AI笔记应用，支持文字、图片、视频等多形式记录AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付4小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版4小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架4小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准6小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型8小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手AI工具22小时前更新AI小集03腾讯Marvis是什么腾讯Marvis是腾讯应用宝团队推出的操作系统层个人 AI 助手。工具通过 1 个主 Agent 与 5 个副 Agent 的协作架构，实现电脑系统操作、文件管理、App 操控、网页交互与信息搜索等复杂任务，24 小时在线并支持 Windows 与安卓跨端操作。产品具备本地知识库索引与可视化虚拟办公室交互，定位大众友好的日常 AI 数字秘书。腾讯Marvis的主要功能多 Agent 协作调度：1 个 PM 主 Agent 负责理解需求、拆解任务并分发给 5 个副 Agent，各 Agent 完成后向主 Agent 汇报，实现复杂任务的自动编排与闭环执行。深度系统操作：可直接调用 Windows 底层接口完成主题设置、电源计划调整、自启动管理、硬件配置检测、网络测速等原生系统级操作，无需模拟点击。智能文件管理：支持基于文件名、内容、图片 OCR、人脸、主题、地点等多维度本地文件搜索，可自动分类归档、格式转换、合并去重及跨端文件传递。跨端 App 操控：通过内嵌应用宝引擎，可在 PC 端直接操控安卓手机 App（如微博签到、同花顺、飞常准等），实现电脑与手机的应用级跨端调度。本地知识库：自动识别并索引本地文档与图片，支持标题匹配与内容匹配搜索，构建越用越懂用户的个人知识管理体系。自动任务与情报监控：支持定时执行既定任务，如每日新闻汇总、限时福利监控、明星动态追踪、房价监控等自动化工作流。如何使用腾讯Marvis下载安装：访问腾讯Marvis官网 https://marvis.qq.com/ 下载 Windows 客户端，按指引完成安装。登录激活：填写调查问卷获取邀请码，获得邀请码后，输入邀请码完成内测资格激活。模式选择：根据任务敏感度选择效率模式或隐私模式。指派任务：在对话框输入自然语言指令，或进入”自动任务”页面设置定时执行规则。查看进度：在右侧虚拟办公室面板查看各 Agent 工作状态与任务消耗 Token 数，等待主 Agent 汇总结果。确认执行：涉及系统修改、文件删除等敏感操作时，点击硬确认弹窗授权后完成最终执行。腾讯Marvis的核心优势操作系统层深度整合：区别于纯云端 Agent，Marvis 深入 Windows 系统底层，可直接调用系统 API 与文件系统，实现真正的本地原生操作。端云双模式灵活切换：效率模式采用云端大模型保障复杂任务性能，隐私模式基于 Qwen 端侧模型实现完全离线运行，满足财务、法务、HR 等敏感场景需求。应用宝跨端生态壁垒：继承 14 年应用商店经验与跨端引擎技术，独家支持 PC 端操控手机 App，生态调度能力为竞品不具备。可视化虚拟办公室交互：将每个 Agent 设计为戴企鹅围脖的黑色小马形象，在虚拟工位中实时展示工作状态（搬砖、打盹、喝咖啡），降低技术门槛并增强用户情感连接。L2 级安全兜底机制：涉及资金交易、系统核心配置修改、批量删除等敏感操作强制”硬确认”，支付功能完全禁用，从架构层面规避 AI 越权风险。腾讯Marvis的同类竞品对比对比维度腾讯Marvis（腾讯）Claude Computer Use（Anthropic）产品定位操作系统层个人 AI 助手云端计算机操控 Agent运行模式端云双模式（本地隐私 / 云端效率）纯云端运行系统深度深入 Windows 底层 API，原生系统级操作基于截图与模拟点击操控跨端能力PC 操控手机 App（应用宝引擎）不支持Agent 架构1 主 + 5 副多 Agent 协作，可视化办公室单 Agent 模式隐私安全L2 级硬确认，支持完全离线运行数据上云，依赖平台安全策略用户门槛自然语言输入，大众友好，开箱即用需一定技术理解，偏极客向生态合作与微软、Intel 系统级深度合作无特殊硬件合作腾讯Marvis的应用场景办公效率提升：自动整理桌面文件并按项目分类归档，一键完成合同信息审查、运营数据分析、参考文献整理及出差材料准备。系统运维助手：一句话调整系统设置（夜间模式、电源计划、默认浏览器），检测游戏配置兼容性并给出硬件升级建议，彻底卸载顽固软件及残留痕迹。生活娱乐管家：自动完成微博超话签到、监控游戏限时福利与明星动态、定时推送每日新闻汇总、整理宝宝成长相册与旅行攻略。敏感数据处理：在完全断网的隐私模式下，本地完成财务报表校对、员工薪资分析、法务合同审查等涉密工作，确保数据不出设备。跨端文件协作：在手机端发送指令，让 Marvis 在电脑中搜索指定文件并回传至手机，实现跨设备的无缝文件调度。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇LibTV团队版 - LibTV 推出的专业级 AI 视频协作工作台下一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型相关文章ScrapeGraphAI – AI网络爬虫工具，自动分析目标网页结构提取关键数据AI小集4Napkins.dev – AI驱动的支持截图或线框图快速生成网页应用的开源项目AI小集2We0 – 开源AI代码编辑器，支持在浏览器中运行调试代码AI小集6arcads – AI视频广告制作平台，根据文本内容、选择AI演员生成视频广告AI小集2吐司AI换脸 – 吐司TusiArt推出的AI人像换脸工具AI小集4GPT Pilot – AI编程工具，让95%的开发者实现自动写代码AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付4小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版4小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架4小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准6小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型8小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

Kimi K2.5 – 月之暗面kimi开源的全新一代全能旗舰模型AI工具4个月前发布AI小集03Kimi K2.5是什么Kimi K2.5 是月之暗面开源的全新一代全能旗舰模型，模型基于约 15T 视觉与文本数据训练而成，具备顶尖的代码生成、视觉理解能力，支持自主 Agent 集群协作，可调度多达 100 个子 Agent 并行处理复杂任务，执行效率大幅提升。Kimi K2.5 提供快速、思考、Agent 及 Agent 集群四种模式，适用多种场景，包括前端开发、办公自动化和复杂问题解决。模型可通过Kimi.com、Kimi App、Kimi 开放平台和 Kimi Code （Kimi 新推出的编程工具）等平台提供服务，助力用户高效完成任务。Kimi K2.5的主要功能视觉与文本融合：Kimi K2.5 能同时处理文字和图像，理解设计图、视频等内容，实现“看图写应用”，将视觉内容转化为可交互的网页或代码，支持像素级复刻和视觉调整。多模式思考：模型支持思考模式和快速模式切换。思考模式适合复杂任务，可进行深度逻辑推理；快速模式适合即时问答，响应迅速。智能体协同（Agent Swarm）：在复杂任务中，Kimi K2.5 可拆解任务并调度多个“分身”协同工作，支持高并发处理，完成多达 1500 步的操作，适用批量任务和多视角分析。专家级办公助手（Office Agent）：模型深度优化 PPT、Word、Excel 等办公软件的使用，能处理文本和解析图表和复杂排版，提供专业级的输出。本地编程支持（Kimi Code）：基于终端的 Coding Agent，支持本地代码的增删改查，自动生成和修改代码，运行测试，并解释改动，方便开发者使用。一键部署与应用生成：生成的网页或应用可直接部署上线，快速将设计转化为实际可用的产品。Kimi K2.5的性能表现Kimi K2.5 是基于约 15T 视觉和文本混合数据训练的原生多模态模型，在视觉理解、图像生成、视频分析等任务中表现出色。例如，在 MMMU-Pro 和 MathVision 等视觉基准测试中，Kimi K2.5 的表现优于或接近其他顶尖模型。Kimi K2.5的项目官网项目官网：https://www.kimi.com/blog/kimi-k2-5.htmlHuggingFace模型库：https://huggingface.co/collections/moonshotai/kimi-k25Kimi K2.5的API调用价格输入（非缓存）：每 M Tokens 价格为 ¥4，相比 K2 Turbo降低了 50%。输入（缓存）：每 M Tokens 价格为 ¥0.7，相比 K2 Turbo 降低了 30%。输出：每 M Tokens 价格为 ¥21，相比 K2 Turbo降低了 63.8%。Kimi K2.5的应用场景软件开发与编程：Kimi K2.5 可将设计图或视频直接转化为可交互的网页或应用，支持本地代码生成、修改和一键部署，大幅提升开发效率。办公自动化：模型能作为专家级办公助手，深度优化 PPT、Word 和 Excel 文档，支持批量任务处理和多任务协作，显著提高办公效率。教育与学习：Kimi K2.5 能根据学生需求生成个性化学习计划和教学材料，提供实时辅导，减轻教师备课负担。创意与设计：模型可将设计稿直接转为代码，支持批量生成创意素材和视觉调整，助力设计师快速完成项目。商业与市场分析：Kimi K2.5 能快速进行市场调研、生成商业计划书，通过数据可视化辅助决策，助力企业高效分析市场。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇ChatTOC - AI聊天对话生成目录插件，一键定位关键信息下一篇DeepSeek-OCR 2 - DeepSeek团队开源的二代OCR模型相关文章LivePPT – 美图设计室推出的AI PPT制作工具AI小集3Sitcom-Crafter – 北航联合港中文等高校推出的 3D 人类动作生成系统AI小集2Talkpal AI – 多语言学习AI助手，模拟真实的实时对话场景AI小集3Remaker – AI图像编辑工具，自动识别去除图片上的水印AI小集2WrenAI – 开源的商业AI Agent工具，自然语言生成SQLAI小集2ChatTCP – TCP协议分析工具，基于聊天界面自动进行诊断和分析AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付5小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版5小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架5小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准7小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型8小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前

Composer 2.5 – Cursor 推出的自研 Agentic 编程模型AI工具8小时前发布AI小集02Composer 2.5是什么Composer 2.5 是Cursor推出的自研 Agentic 编程模型。在智能水平和行为表现上较Composer 2有大幅提升，在SWE-Bench Multilingual（79.8%）和CursorBench v3.1（63.2%）等核心基准上与 Claude Opus 4.7、GPT-5.5 处于同一梯队，但单次任务成本仅为竞品的约1/10，称为”性价比之王”。模型基于 Moonshot 开源的Kimi K2.5检查点持续训练，目前仅通过 Cursor IDE 及 SDK 提供服务。Composer 2.5的主要功能长时任务持续工作：针对长时间运行的 Agent 会话深度优化，能在多步骤工具调用中保持专注，显著减少中途幻觉或提前终止的问题。复杂指令可靠遵循：对跨文件重构、终端命令执行、测试驱动开发等复杂指令的遵循可靠性较 Composer 2 大幅提升。努力级别动态校准：模型能根据任务难度自动分配计算量，简单任务快速完成，复杂任务深入思考，避免”小事空转、大事欠思考”。沟通风格优化：回复更简洁结构化，减少不必要的冗长解释，在多文件变更时提供更清晰的推理过程。工具调用精准度提升：显著减少无效的终端命令或冗余搜索，提升代码检索与终端操作效率。双版本灵活适配：提供 Standard（标准版，$0.50/$2.50 per M tokens）与 Fast（快速版，$3.00/$15.00 per M tokens），智能水平相同，分别适配后台批量任务与交互式实时编程。如何使用Composer 2.5在 Cursor IDE 中启用：将Cursor更新至 2026 年 5 月最新稳定版，打开 Composer 面板（Cmd+I/Ctrl+I），点击模型选择器切换至Composer 2.5选择速度档位：交互式开发默认使用Fast 版（响应快、延迟低）；后台 Agent 或批量任务可在 Settings > Models 中切换为Standard 版（成本低，智能水平相同）通过 SDK 程序化调用：import { Agent } from “@cursor/sdk”; const agent = await Agent.create({ model: “composer-2.5”, // Standard 版 // model: “composer-2.5-fast”, // Fast 版 workspace: “./”, tools: [“edit”, “shell”, “search”, “browser”], });为长时任务设限：为无人看管的长时间 Agent 会话设置迭代次数上限和最大耗时，防止模型利用缓存等”捷径”进行奖励作弊领取首发福利：发布首周提供双倍用量额度，可在后台查看额度消耗情况多模型路由策略：将 90% 日常任务交给 Composer 2.5 处理；仅将架构评审类任务路由给 Claude Opus 4.7，重度终端 Shell 任务路由给 GPT-5.5Composer 2.5的核心优势极致性价比：Standard 版输入 $0.50/M、输出 $2.50/M，Fast 版输入 $3.00/M、输出 $15.00/M，比 Claude Opus 4.7 便宜约1030 倍。前沿级基准表现：SWE-Bench Multilingual 79.8%（Opus 4.7 为 80.5%，GPT-5.5 为 77.8%），CursorBench v3.1 63.2%（与 Opus 4.7 的 64.8% 和 GPT-5.5 的 64.3% 基本持平）。行为层面深度优化：除了扩大训练规模，还改进了沟通风格和投入级别校准，这些维度虽难被基准充分反映，但对实际使用体验至关重要。双版本灵活选择：Standard（标准版）适合后台 Agent 与批量任务，Fast（快速版，默认）适合交互式 IDE 实时编程，两者智能水平完全相同。首发双倍用量福利：发布首周提供双倍用量额度。Composer 2.5的同类竞品对比对比维度Composer 2.5Claude Opus 4.7GPT-5.5厂商 / 平台CursorAnthropicOpenAI产品定位自研 Agentic 编程模型旗舰推理模型旗舰多模态模型基座模型Moonshot Kimi K2.5（开源检查点持续训练）Claude 4 系列GPT-5 系列发布日期2026.05.182026 年 Q22026 年 Q2SWE-Bench Multilingual79.8%80.5%77.8%Terminal-Bench 2.069.3%69.4%82.7%CursorBench v3.1（困难任务）63.2%64.8%（max）/ 61.6%（默认 xhigh）64.3%（xhigh）/ 59.2%（默认 medium）输入价格（/M tokens）$0.50（Standard）<$3.00（Fast）未公开（行业参考约 $15）未公开（行业参考约 $3–$5）输出价格（/M tokens）$2.50（Standard）<$15.00（Fast）未公开（行业参考约 $75）未公开（行业参考约 $15–$30）单次任务相对成本基准（约 $1–$2 / 任务）约10–30 倍约3–10 倍上下文窗口200K（参考 Kimi K2.5）200K128K–1M权重开放性闭源（仅 Cursor 基础设施）闭源闭源接入方式Cursor IDE / CLI /@cursor/sdkAPI / Claude Code / 第三方平台API / ChatGPT / GitHub CopilotComposer 2.5的应用场景多文件级重构：成本优势明显且精度持平前沿模型，适合大规模代码库迁移。交互式结对编程：Fast 版响应迅速，适合实时 IDE 协作。后台定时任务/云 Agent：Standard 版性价比极高，适合批量代码审查与修复。测试驱动开发：长时任务可靠性优于前代，能稳定完成多轮测试-修复循环。复杂终端自动化：Terminal-Bench 2.0 得分 69.3%，与 Opus 4.7（69.4%）持平，但重度 Shell 场景仍略逊于 GPT-5.5（82.7%）。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型下一篇Chronicles-OCR - 腾讯联合高校等推出的视觉感知评测基准相关文章AndroidLab – 清华和北大联合推出系统化评估Android智能代理的框架AI小集2塔猫ChatPPT – AI PPT生成工具，智能分析自动配图、生成完整演示文稿AI小集3360AI助手 – 360联合15家大模型厂商推出的AI助手产品AI小集21Yodayo – 在线AI绘画创作平台，一键生成高质量动漫风格作品AI小集3DeepSeek-R1-0528 – DeepSeek开源的最新版R1模型AI小集3Apollo – Meta 联合斯坦福大学推出的大型多模态模型AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录文枢三言堆友Agent灵芽社区方舟 Agent PlanMavis万小智最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付4小时前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版4小时前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架4小时前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准6小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

David NieldGearMay 19, 2026 6:30 AMSet Up Your Phone’s Always-On Display So You’re Unlocking It Less OftenBreak the pattern of unlocking your phone for one thing and then getting sucked into something else. Set up your lock screen to show all the information you need at a glance.Courtesy of AppleCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyHow many timesdo you think you check your phone a day?The number is probably much higher than you’d guess. Phones are always within easy reach, offering instant access to the world, social media, and messages from friends and family. Even if you’re not being pinged with notifications, the temptation to scroll through an app or two is always there.There are different ways to tackle this, frommuting notificationsto keeping some space between you and your device. But one easy way to be sure you check your phone less often is to do more with the lock screen on your handset.You can populate the lock screen with a variety of alerts and updates that help you see what’s happening with just a brief glance. You won’t need to unlock your phone again and again, and you won’t be drawn into checking multiple apps each time you pick it up. Here’s what you can do on the major phone models.Google Pixel PhonesMost Android phones now offer an always-on display feature, including recent Google Pixel models. To customize what appears on the lock screen, head to Settings, then chooseDisplay and touch, and you’ll find anAlways-on displayoption: Tap this to find a toggle switch to enable or disable the feature.Also on theDisplay and touchmenu, you can tapLock screento configure what shows up on the display when your phone is locked. The main option to be aware of here isNotifications on lock screen: Tap this to choose how notifications appear onscreen (in full or as snippets) and to determine which kinds of alerts show up.Make sure you can see notifications on the lock screen.Courtesy of David NieldYou can for example hide “sensitive” info from the lock screen, in case someone else glances at your phone—this essentially just means notifications from messaging apps, which might include personal stuff that you want to make sure isn’t available to other people. (Remember, with the always-on display setting enabled, anyone around you can get a look at your lock screen).The other option you may want to enable on theLock screenmenu isWidgets on lock screen. This lets you set up widgets like a camera shortcut and a weather forecast to be one swipe away on the lock screen—again saving you the trouble of having to unlock your phone and dive into specific apps.Samsung Galaxy PhonesOn a Samsung Galaxy phone, you can get to the relevant options by opening the main Settings screen, then choosingLock screen and AOD(always-on display). First of all make sure theAlways On Displaytoggle switch is turned on, then tapNow bar. This lock screen widget displays information like sports scores and fitness stats.To control the notifications you see here, you need to go back to the Settings page, then chooseNotifications > Lock screen. EnableShow notifications, then choose the style of alerts you want:Cardsis the best one to pick, if you want to get as much information on the lock screen as possible.There are three notification styles to choose from on Galaxy phones.Courtesy of David NieldOn the Notifications screen, you also need to selectHide content while lockedand enableShow always—otherwise you’re not going to be able to see anything from your notifications on the lock screen. Just be aware that if you can view alerts on your lock screen, anyone else who picks up or looks at your phone can do the same.You can add widgets to the lock screen too, for more at-a-glance information. With your phone locked, press and hold on a blank part of the lock screen to enter edit mode: If you then tapAdd widgets, you can see which widgets are available (including calendar widgets and news updates). TapDoneto confirm your changes.Apple iPhonesFor iPhones, you can enable the always-on display from Settings by tappingDisplay & Brightness > Always On Display. As well as making sure theAlways On Displaytoggle switch is turned on, you should also enableShow Notificationsso that information from your apps comes through to the lock screen.The notifications themselves can be managed from theNotificationsentry in Settings. There are three notification layouts to pick from, andListis the best of the three if you want to get as much information as possible up on the lock screen, to reduce the number of phone unlocks you need to do.Get your iOS notifications shown as a list on the lock screen.Courtesy of David NieldTapShow Previewson the same menu page, then pickAlways, so that you get snippets of notification content on the lock screen as well. Just be aware of other people who may be able to access or see your phone, as they will also be able to view what’s on the lock screen. (You may want to switch this off wh

Heather HansmanScienceMay 19, 2026 7:00 AMThe US Built a Site to Ensure Fair Access to Public Lands. Then Everything Went WrongRecreation.gov was supposed to make access to public lands more equitable and streamlined. Instead, it’s rife with bots and inequality, while a government contractor benefits.Illustration: Jacqui VanLiewCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyIt’s a fewminutes before 8 am Mountain Time on March 16, the day that river permit cancellations are released on Recreation.gov, the federal website for public land reservations.Rec.gov, as it’s commonly called, administers everything from river permits and timed entrance fees at the most popular national parks to campground reservations on remote sites belonging to the Bureau of Land Management, and a lot of people are recreating on public land these days. There were 11 million reservations on the site in 2024, up significantly from 3.5 million reservations reported in 2019. At the center of it all is an unlikely player in the outdoor recreation space: The site is operated by the government contractor Booz Allen Hamilton, a corporation known more for cybersecurity than rafting trips.Early each year, outdoor enthusiasts gear up forRecreation.gov’s annual lotteries for some of the most iconic experiences in the country: a river trip down Idaho’s Middle Fork of the Salmon River, which flows through the Frank Church River of No Return Wilderness. Backcountry permits to hike into the Wave, an otherworldly rock formation in Arizona’s Paria Canyon–Vermilion Cliffs Wilderness. Overnight stays in the rugged, lake-studded Enchantments, in Washington’s Okanogan-Wenatchee National Forest.The Paria Vermillion Cliffs.Photograph: Bernie Friel/UCG/Universal Images Group via Getty ImagesOdds of getting a desirable Middle Fork permit are around 2 percent. Each year, around 200,000 people apply in advance for 48 daily lottery spots to hike into the Wave.Rec.govitself reports that a campground with 57 campsites can see 19,000 users trying to reserve them. That’s a .3 percent success rate.For the majority who don’t draw a permit, there’s one final hope: the release date for cancellations, where your chances of getting a spot are often based on how fast you can click, and whether you can be online right when canceled permits are released.That’s where a river runner I’ll call Jack was last March 16. A web-developer friend of a friend who is frustrated by the way permits seem to be snapped up faster than humans can possibly click, Jack decided to do some experimenting to see if the speculation that bots are grabbing all the permits seems true.That speculation is based in reality. There’s a user on the outdoor forum Mountain Buzz who offers up a free scraperbot to anyone who wants to use it, and developers have sharedtheir code. Last year Sam Carter, the host of theRiver Radiuspodcast, did an episode where he built a bot to show that gaming theRecreation.govsystem was possible. He was shocked at the response. “So many people say they’re using bots, people are bragging about it,” he told me. He heard from people who’d built their own, groups who have their own server dedicated to getting permits, and people who paid thousands of dollars to have someone build one for them. It’s happening. The question is how pervasive it is, and how easy it might be for anyone to hack Rec.gov.Jack wants to prove that the system isn’t working, so he’s built a series of bots to try to outsmart the other bots. He has multiple accounts and bots that can do everything from alerting him to permit availability to keeping permits in his cart for hours at a time. “I’m trying to simulate what I think other people are doing,” he says.The night before the river cancellations he’s interested in are released, Jack opens the Inspect Element browser tool on Rec.gov and scrolls through the data to find what dates are going to become available. “If you were a web developer of any kind,” he says, “you would be able to find it.”He has programmed several bots, attached to burner accounts with obviously fake names, to try to grab the dates he knows will come up when permits are released at the tick of 8 am. As a control, he also has a friend on a laptop nearby who will be attempting to grab a permit the old-fashioned way, by clicking through the site herself.When the clock turns over, everything happens fast. The scripts start running, flashing between five different screens, running through the calendars on Rec.gov pages. Then, within a few seconds, the action stops. Jack goes to the accounts to see if his experiment worked. In the first cart, the bot has secured permits for the Main Salmon, the San Juan River, and the Middle Fork of the Salmon. It’s impossible to overstate what a big deal it is to score these three prime trips in a year, much less in a single morning. In the second cart, another unicorn: a second Middle Fork permit.The Salmon River near Stanley, Idaho.Ph

Julian ChokkattuGearMay 19, 2026 7:00 AMThis Is Herman Miller’s New Standing Desk for GamersThe Coyl uses a coil design for the power cable, keeping it neat and taut between the desk and the wall. And instead of relying on boring levers to adjust the height, it has a nifty rotary dial.Courtesy of Herman MillerCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyWhen you buya “gaming” accessory, you’re often buying a product that offersa little more oomph—a few extra perks or advanced features—catered to gamers who spend hours clicking away, trying to win. They may not be particularly chic, but these gaming peripherals are all about maximization, whether that’s agaming keyboard, agaming headset, or agaming mouse.Gaming desks are now a growing category. These furniture pieces accessorize your game room or home office while also minimizing distractions, so gamers can focus on gameplay. They often are designed to match their accompanyinggaming chairsand their distinctive race car seat aesthetic. Various gaming-focused companies likeSecretlaband Corsair offer some of the more comprehensive gaming sit-to-stand desks on the market—evenIkea has a few—but now esteemed furniture brand Herman Miller is entering the fray through its sub-brand, Herman Miller Gaming.TheHerman Miller Gaming Coylis the company’s firststanding deskbuilt for gamers. True, the company also sellsMotia-branded gaming desksfrom 2018, but the Coyl is more specifically designed for gamers. But this is its first in-house design, and it’s easy to tell the various desks apart thanks to the Coyl’s three main design elements: the trough, the coil, and the dial.First, the trough. This is the giant space underneath the desk, towards the back, where gamers can store all those unsightly wires and cables that run from their peripherals to their gaming PC. Most standing desks today offer some kind of way to route cables through the desk, but that’s child’s play compared to this. Very much like Secretlab’s Magnus Pro desk, the trough is unusually wide to accommodate plenty of cables and power strips. What’s also nice is that you can raise the desk to its max height and pull down the felt cover to access the trough from the front, like a mechanic working on a car, without needing to awkwardly fish for cables from the back of the desk or pull the whole desk away from the wall.Then there’s the desk’s namesake, the coil: a distinctive red wire that runs from an outlet to the desk to power the motors and anything else you plug into the integrated power strip. It’s an intentional design. Rather than let the cable dangle as it awkwardly stretches to your outlet, the coil design keeps it taut.That curly red cable.Photograph: Julian Chokkattu“Coils by nature solve a problem, which is how do you keep a space clean and tidy, that has a variable length between two things,” says Steven Harton, senior product manager at Herman Miller. He says the design choice was inspired by the coiled cables on vintage tech like telephone handsets and early mechanical keyboards. The result is fun and orderly.Finally, there’s the rotary dial. This is the height adjustment mechanism on the Coyl desk. Rather than using traditional paddles or levers to make the desk go up or down, Herman Miller Gaming employs a dial that feels like something you’d find on an old-school stereo—or a really nice gaming mouse. It lets you fine-tune the desk’s height to a 10th of an inch. (A display shows the table’s exact height.) That level of detail is an ode to the kind of precision gamers are fond of, digging into settings to maximize performance.Twist the knob to adjust the height of the desk.Photograph: Julian Chokkattu“Sensorially, how it feels and the feedback you get is an elevated experience relative to a button or a toggle,” Harton says. “When you use a dial, there’s a resistance to it that’s delightful.” We’ll have to wait forKnob Feel’s verdicton exactly how delightful it is.There are other niceties. The Coyl desk has integrated hooks on either side, which gamers can use to hang their headphones or a backpack. An optional back panel upgrade adds a classic pegboard to the desk, allowing gamers to add small shelves and hooks to hold their gaming controllers, small plants, lamps, or their favorite Funko Pop. There’s also the option to add other Herman Miller accessories like monitor mounts.The base version of the Coyl desk starts at $1,095. Add the cable management trough, and it’ll cost $1,495, and with the back panel, it tops out at $1,635. The rounded laminate desktop comes in four colors: black, white, walnut, and ash. You can program up to four height presets for the desk.It may be advertised as a gaming desk, but Harton thinks the modern gaming environment is multifunctional—not just for gaming, but for consuming, working, and relaxing. That, and the fact that the desk is drummed up by a furniture company and not a gaming brand is likely why the design may not feel very gam

Scott GilbertsonGearMay 19, 2026 8:08 AMThe Best Action CamerasGearing up to shred the slopes or dive into the seas? These photography tools are made for danger.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyChoosing the bestaction camera used to be simple:Go with GoPro. That’s still good advice, as the GoPro ranks high on our list, and itsforthcoming Mission 1 camerascould potentially top this list again, but there are also some worthy competitors that beat the current flagship GoPro Hero 13 Black on several fronts.The decision is not easy, so to figure out what exactly makes the best action camera, we dove with them, climbed with them, surfed with them, and handed them to reckless 13-year-olds on bikes. We found several great options that will survive your adventures on the ski slope, at the beach, or at the skate park.Be sure to also read our guide to theBest 360 Cameras,as there’s considerable overlap these days with action camera, and in some cases, for some people, they might be the better option. We’ve also got guides to theBest Travel Cameras,Best Mirrorless Cameras, andBest Instax Cameras.Updated May 2026: I’ve noted GoPro’s announcement of the new Mission 1 Cameras, which will have 8K video and support interchangeable Micro Four Thirds lenses. I’ve also removed the GoPro Hero 12 Black because it’s no longer a good value, and I’ve updated prices and links throughout.GoPro Mission 1 UpdateCourtesy of GoProCasting a long shadow over any action camera buyer’s decision right now is therecently announced GoPro Mission 1 cameras. There are three of them, theMission 1, Mission 1 Pro, and Mission 1 Pro ILS. All three feature a larger sensor with 8K video, improved low-light performance, up to 32X slow-motion footage, and 50-megapixel still images. The Pro ILS uses interchangeable lenses based on the Micro Four Thirds mount.GoPro claims the new sensor offers a dynamic range of up to 14 stops, whichrivals full-frame mirrorless cameras. This will be further helped by the available 10-bit GP-Log2 shooting mode, which maximizes dynamic range and allows for custom coloring in postproduction. Also interesting are some of the new accessories and add-ons. There are lens hoods, ND filters, a battery grip, a new Media Mod kit with a built-in mic, along with a 3.5-mm mic port, a headphone port, a micro HDMI port, and a line-in port for time code sync. GoPro is also making a point-and-shoot grip, reminiscent of Insta360’s grip for the Ace Pro 2.Should you wait and see? Probably. The Mission 1 cameras are a bit pricier, $600 for the Mission 1, $700 for the Pro and Pro ILS (the latter of which won’t be available until Q3 2026), but the video and image quality GoPro has teased thus far are very impressive. We’ll have a full review coming next month.Best Overall Action CameraPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonChevronChevronSave to wishlistSave to wishlistDJIOsmo Action 6$426Amazon$436B&H PhotoDJI’sOsmo Action 6 is the best action cameraright now. It trumps our previous top pick, theDJI Action 5 Pro, in several ways, adding a variable aperture for better low light performance and a new, larger sensor for higher-quality video. The Action 6 also gains 2X zoom and 8K video, though as withthe Insta360 Ace Pro 2, the 8K video is more of a specs-based bragging right than a significant improvement over the 4K. It does mean you can crop in though, whether using the 2X zoom in-camera, or on your own when you’re editing.Another great upgrade in the Action 6 over the Action 5 is some fairly dramatic improvements in color rendering and contrast in the default color modes. This is less apparent if you shoot Log-D, but if you don’t want to spend time color-grading your footage, now you don’t really need to. The default colors are pretty great for an action camera. Low-light performance is better, as you would expect with a wider-aperture lens. The video is cleaner and sharper than either the Action 5 Pro or the Ace Pro 2. That’s true whether you use SuperNight mode or not. I should note, though, that you don’t get full manual control here, but the automatic modes were pretty smart most of the time in my testing.DJI has also followed GoPro’s lead here, adding amacro lens ($169)to its lineup, which also includes an ultrawide lens, now know as theFOV Boost lenswith 182-degree field of view.WIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDNew larger sensor for higher-quality video2x zoom and 8K video capableGood low light performanceBetter default color renderingTIREDCan’t activate Zoom while filmingThe Best GoProPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonPhotograph: Scott GilbertsonChevronChevronSave to wishlistSave to wishlistGoProHero 13 Black$430$349(19% off)Amazon$430$380(12% off)GoPro$429$379(12% off)AdoramaWe have a separate guide tochoosing the right GoPro, and as noted above, the n

Caroline HaskinsBusinessMay 19, 2026 10:00 AMThe Zuckerbergs Are Hiring a Lifeguard but Calling It a ‘Beach Water Person’The job, which is associated with the Zuckerberg family office, is located in Kauai, Hawaii, where the Meta CEO owns a massive compound.Photograph: Jason Henry/Bloomberg/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyMeta CEO MarkZuckerberg and his wife Priscilla Chan are hiring a seasonal, on-call “beach water person” based in Kauai, Hawaii, where the family owns asprawlingcompound, according to a new job listing on Greenhouse associated withWest 10, the Zuckerberg family office.This is an interesting choice for a job title, because according to the job description, the primary duties of this “beach water person” include serving as a “beach lifeguard” and “pool lifeguard.” In other words, being a lifeguard.The job listing also mentions serving as a “boat deckhand” and names a few additional duties, such as instructing “stand-up paddleboarding (SUP), canoe paddling, snorkeling, and other ocean-based activities.” These, however, come after the water safety responsibilities in the job description.Got a Tip?Got a tip? Are you a current or former employee of the Zuckerberg-Chan family? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at carolinehaskins.61.This position easily could have been called “pool/beach lifeguard” or simply “lifeguard.” For the sake of comprehensiveness, “pool/beach lifeguard and boat deckhand” would have also worked. Alternatively, the Zuckerbergs could have chosen “beach/pool attendant,” a job title roughlysynonymouswith lifeguard that could reasonably be interpreted as encompassing extra duties associated with leisure, such as tending to a boat or teaching people how to paddleboard.Arguably, any of these options would have provided more clarity than “beach water person,” whichdoes not appearto correspond with a job title anywhere else in the English-speaking world.WIRED did not immediately hear back from representatives of the Zuckerberg family. Lacking a human to speak with, we decided to ask Meta’s AI chatbot, “What is a ‘beach water person’?”“Beach water person’ would just mean someone who loves being in/near the ocean,” the chatbot said. “The word for that is thalassophile—‘a person who loves the seas and oceans.’” OK!

Maxwell ZeffBusinessMay 19, 2026 11:00 AMFormer OpenAI Staffers Warn That xAI’s Poor Safety Record Could Complicate SpaceX’s IPOThe ex-employees, who cofounded a new AI watchdog group, say investors deserve more information about xAI’s safety practices before SpaceX goes public.Photograph: Miguel J. Rodriguez Carrillo/AFP/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyTwo former OpenAIemployees and a group of AI safety nonprofits are warning that Elon Musk’s AI lab,xAI, could become a liability for prospective investors inSpaceX,which is preparing to file what’s expected to be the largest initial public offering in Wall Street history.In aletterdirected to investors published on Tuesday, the ex-staffers highlighted what they describe as “unpriced risks” related to xAI that could complicate SpaceX’s reported plans to raise up to$75 billionas part of its IPO. The rocket company’s private valuation shot up to over $1 trillion after itacquired xAI last year. Musk claimed his rocket company could launch data centers into space for his AI lab, but the letter’s authors argue that xAI’s poor record on safety issues could complicate how investors view the combined company as it gets ready to submit itsIPO prospectus filing.One of the letter’s signatories and coauthors is a new nonprofit calledGuidelight AI Standards, which was cofounded by former OpenAI safety researcherSteven Adlerand former OpenAI policy adviser Page Hedley. The group, which is backed by private donors, aims to improve the safety practices of frontier AI companies. Other AI safety nonprofits also signed on, including Legal Advocates for Safe Science and Technology, Encode AI, and The Midas Project.Hedley tells WIRED in an interview that he believes xAI has the worst safety practices “nearly across the board” compared to other frontier AI developers, including OpenAI, Google DeepMind, and Anthropic. As a result, he argues, SpaceX may face a greater risk of regulation and litigation than other AI labs.The letter’s authors argue that SpaceX should make several disclosures to investors, including whether xAI intends to continue developing frontier AI models. SpaceX recentlystruck a deal to sell a significant portion of its GPU capacity to Anthropic, and the letter claims the agreement “leaves it unclear whether xAI is still a frontier-AI competitor inside a larger holding company.” If xAI continues to develop frontier AI models, the authors say, it should be required to publish a public safety and governance plan.SpaceX and xAI did not immediately respond to WIRED’s request for comment.The letter also outlines examples of how xAI has not kept up with industry-standard safety practices such as publishing detailed frameworks for mitigating risks around its AI models being used in cyber attacks. The authors also outline specific safety incidents at xAI that they say warrant additional scrutiny. Among the most notable include when xAI’s flagship AI chatbot, Grok, spontaneouslybrought up white genocidein its responses. In another case, xAI allowed Grok to generate thousands ofsexualized images of women and children, which spread widely across Musk’s social media platform X. The latter case promptedat least 37 US attorneys generalto send a letter demanding that Musk’s AI lab take steps to protect women and children on its platform.Hedley says the number of safety incidents xAI has experienced and the regulatory attention they received is “far out of proportion to its market share.” As lawmakers grow increasingly alarmed by the cyber capabilities of advanced AI models likeAnthropic’s Claude Mythos, new security regulations may be on the horizon. The Trump administration is reportedly alreadyweighing an executive orderthat would give US intelligence agencies more oversight over AI models.“It takes serious investment to rein in [AI safety] risks, and it seems that xAI has historically under-invested here,” says Adler. The letter cites reporting from The Washington Post that said xAI had just “two or three” people working on safety as of January. “A question investors should be wondering is if xAI stays at the frontier, how costly might it be to, in fact, manage these [risks] responsibly? If they don’t, what might be the consequences?”The letter acknowledges that Musk’s AI lab has made some improvements related to safety recently, such asexpanding an existing agreement with the White Houseto allow tests of its AI models before they’re publicly deployed. However, the authors say more disclosures are needed in order for investors to be able to fully assess the AI safety risks associated with SpaceX.“xAI’s historical record has been serious enough to warrant scrutiny; it has not, by itself, foreclosed a better future for the company,” the letter reads.New WatchdogAdler and Hedley say Guidelight AI Standards is aiming to create new uniform benchmarks that AI labs can reasonably follow. They also want to offer clear assessmen

Julian ChokkattuBoone AshworthGearMay 19, 2026 1:45 PMHands-On With All of Google’s New Upcoming Android XR Smart GlassesHere’s your first look at smart glasses coming from Warby Parker and Gentle Monster, powered by Google and Samsung’s XR platform.Photograph: Julian Chokkattu/WIREDCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle has beenteasing its Android XR-poweredsmart glassesfor the past year and a half—ever since it first announced the newmixed reality platform in December of 2024. Now, we’re finally starting to see more polished versions of the hardware, several of which will actually go on sale this fall.AtGoogle I/O, the company gave us a first look at thedesigns of smart glassescoming from established eyewear brands Warby Parker and Gentle Monster. These smart glasses are codeveloped by Samsung and Google, while the eyeglass companies handle the design of the frames.Xreal also took the stage at I/O to showcase its upcoming Project Aura, which is essentially a miniaturized glasses version of bulky headsets like theApple Vision ProandSamsung’s Galaxy XR—powering a full Android app interface with hand gestures for interaction.Audio-only smart glasses from Warby Parker and Gentle Monster will launch later this year. More complex versions with displays built into the lenses will arrive soon after. Xreal’s Project Aura will also come this fall.I got a chance to wear some early versions of these frames and try out a few features in a controlled demo experience.Smarter GlassesPhotograph: Julian Chokkattu/WIREDI wasn’t able to see the final polished designs from Warby Parker and Gentle Monster. Instead, my demos were on unfinished reference glasses from Samsung and Google. The most remarkable thing about them is how light they are. Google says much of the work it and Samsung have done was to miniaturize the technology to save weight. The arms of the glasses are still a little chunky, though.Audio quality was equally impressive. As soon as I asked Gemini to play some Radiohead, the audio surrounded my head and sounded dynamic, though I was in a quiet room. I asked someone else to try them so I could see if I could hear it at 50 percent volume while sitting across from them. I could barely hear the music.The audio-only glasses will be the first to arrive later this year. And yes, all of these upcoming smart glasses—even the audio-only ones—have cameras. That’s how the Gemini assistant can see what you’re seeing to provide helpful, contextual answers and services. The versions with more advanced optics will provide a richer experience. For example, if someone talks to you in a different language, the Gemini assistant inside all of the glasses can translate their voice and even make itsoundlike the person speaking, but the display versions of the glasses will offer accompanying text to read as well.Courtesy of GoogleCourtesy of GoogleThe versions with a display built into the lenses should be especially helpful with things like turn-by-turn navigation, since you’ll be able to see exactly where to turn left on the display as you walk. Google says the glasses don’t have GPS and rely on your phone’s GPS, but the cameras on the glasses utilize Google’s Visual Positioning System to identify surroundings and ensure your blue dot is in the right spot—something you can currently do in the Google Maps app to calibrate your compass.I tapped and held on the right arm of the glasses, and this triggered Gemini Live. That’s the default experience, allowing you to have a conversational back and forth with Google’s assistant.Gemini did a good job of only listening to my voice in these encounters and ignoring people who were speaking to me. I was able to ask it what the board game next to me was—it correctly identified it as Chinese checkers—and then asked if I wanted to learn how to play. I asked it to just save me those instructions in a Google Keep note, and voilà, it generated that advice in a Keep note within seconds.Photograph and composite by Julian Chokkattu/WIREDNext, I was able to ask Gemini to take a picture and then alter the photo post-capture. I first asked it to remove a plant in the photo, then I asked it to change the room’s decor to a medieval hall. It did those things remarkably well.You’ll get the original image on your phone and a preview on Wear OS smartwatches if you’re wearing one (or a preview on the display-laden glasses), and the doctored version will arrive within 45 seconds once Google’s Nano Banana platform does its thing.I was able to see a few generative widgets and scroll through them on these glasses—that’s a big theme Googleis introducing in Android 17and other platforms—so you can have glanceable information on your glasses for whatever you want. I wasn’t able to see much else of the user interface, though.Project AuraPhotograph: Julian Chokkattu/WIREDA different kind of glasses experience, Project Aura is more akin to mixed-reality headsets like Apple’s Vi

Reece RogersGearMay 19, 2026 1:45 PMGoogle Search Goes Agentic—and Doesn’t Need You AnymoreVibe-coded results! Super widgets! Bots that never sleep! Google’s vision for the future of Search is hyper-personalized, automated, and extremely AI.Photograph: Thilina Kaluthotage/NurPhoto/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAI agents areeverywhere. Every briefing I’ve attended for software companies over the past year has involved some mention ofagents—using generative AI tools to automate digital tasks. Despite breakout moments at the start of 2026, like the pluckyOpenClawagent that early adopters used to manage their online life, most people are not yet embracing this style of automation day-to-day. That won’t last for long if Google gets its way.AtGoogle I/O, the tech behemoth shared its vision to make its popular search engine a core way to expose billions of existing users to the company’s agentic prowess.“You will be able to create, customize, and manage multiple AI agents for your many tasks, right in Search,” says Liz Reid, who leads Search atGoogle. She gives the example of setting up an agent to track stock market trends and send you alerts using real-time data—when specific conditions are met.Alongside these agentic additions coming to Search, Google also announced a new underlying model, Gemini 3.5 Flash, as the global default model forAI Mode answers,as well as improvements to the Search box that make it more responsive to user inputs.Always-On AnswersOne of the key agentic experiences rolling out is a new form of dynamic data gathering through “information agents” that can be more proactive than previous search experiences and automate alerts through your conversational requests.“Ask Google to just keep you updated on anything, and now our agents can do work for you even if you’re not using Google,” says Robby Stein, a vice president of product for Search. “So, you could be asleep, and it’s still helping you.” This feature will arrive first for subscribers toGoogle’s AI Pro and Ultra plansthis summer.In Google’s example of how these can work, a user asks AI Mode through the Google app to keep them “updated when any of my favorite athletes announce sneaker collabs or signature drops.” Then, AI Mode generates an info agent unique to that user, designed to monitor the request. When a new shoe drops that fits these criteria, likeA’ja Wilson’s pink Nikes, the user gets a notification alert with critical context and ways to buy the sneaks.Booking agents are another automation style in Search that Google will be expanding this summer. At previous I/O conferences, Google debuted similar AI features, like the now defunctDuplex, which called companies on your behalf to set up restaurant reservations or salon appointments. Google continues to iterate on that core idea with agents that can search for relevant context about local companies, even calling that barber down the street for a price quote on a beard trim if it doesn’t show up on the website, to help users gather info with less active participation.Much of this keeps users trapped withinGoogle’s ecosystemwithout actually browsing the web themselves, a practice we’ve seen before when features likeAI Overviewslaunched in 2024. Even so, Google has explicitly stated that it’s not attempting to replace web page links.Vibe-Coded ResultsGoogle first launched Antigravity, its version of an agentic coding tool likeOpenAI’s CodexorAnthropic’s Claude Code, late last year. Now, an updated version of Antigravity will be used to create bespoke outputs in Search, like an adjustable visualization of black holes to help you better understand deep space.“Search can build you custom experiences,” Stein says. Google’s vibe-coding tool, Antigravity, will craft unique answer layouts, when helpful, and incorporate interactive visual elements, such as infographics and movement simulations. The update is expected to arrive later this summer for all users.An expanded version of that tool, which Google says is a way to create “super widgets” and “mini apps,” will be available first for Google AI plan subscribers in the US this summer. (I vote to call them super widgets, which sounds more unhinged.) These can be used for tracking longer-term plans, like a healthy diet, or managing complex tasks, like a cross-country move, through Search.Search Doesn’t Need YouGoogle of yore, with its 10 blue web links, asked users to be fairly active participants in the info-gathering process:Here are a bunch of webpages that appear relevant to your search. How about you click around and see which site has the best info?The process was clunky at times and could be time-consuming, but I was personally hopping site to site as Google nudged me into different online rabbit holes.The launch ofAI Overviews, which generates summaries of web pages and online information, two years ago, and the rollout ofAI Mode, the chatbot-style search, last year, have au

Reece RogersGearMay 19, 2026 1:45 PMGoogle Makes It Easy to Deepfake YourselfGoogle’s overhaul of its AI creation software, Flow, includes a new video model and a tool for generating selfie videos called avatars.Courtesy of GoogleCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyA wave ofdéjà vu washes over me as Elias Roman, vice president of product management atGoogleLabs, demos a new “avatar” feature for Flow, the company’s tool that lets users generate and remix AI videos and images. He previously scanned his likeness to create a digital clone of himself. Now, he can insert himself into any AI-generated videoclip he wants using Google’s new Omni Flash model.“This is for creators who want to bring themselves into their content but don’t want to have to shoot themselves,” Roman says.Courtesy of GoogleThis specific style of social-first, selfiedeepfakeis reminiscent of a quintessential feature from OpenAI’s now-defunctSora app—rather than cameos or characters, Google calls them avatars. These avatars are also available through the Gemini app andYouTube. Google announced the new feature at its annualI/O developer conferencein Mountain View, California.Google launched Flowlast yearunder its experimental Labs division. “Google has never had a product line for creative work before,” Roman says. “Productivity, definitely. Developers, absolutely. Video consumption, yes. Not for creative work.” He sees this as Google’s attempt to build tools for the next generation of creators.Similar to other announcements from Google I/O surrounding Google Search, many of the new changes toFloware part of the company’s larger attempt to makeAI agents, essentially automated software taskmasters, andvibe coding, building bespoke features with natural language prompts to AI, more mainstream for a broader audience. For example, users can repeat custom instructions when generating videos and create automated workflows that sort similarly styled clips into folders.Courtesy of GoogleOne of the most immediately noticeable changes to Flow is the new video-generation model powering the experience: Omni Flash,succeeding Veo. Similar to howGoogle’s Nano Bananamodel brought more context about the world into the AI image-creation process, the Omni Flash model overhauls video generation with richer detail throughout clips.Flow users can generate characters in AI videos with more consistency via the Omni Flash model. Roman says this is a major improvement over the weakness in past versions of Flow, where created characters could warp during successive video generations. Also, a key character that Flow users can now generate in an AI scene after an AI scene? Themselves.Users set up an “avatar” of themselves by going into the settings of their Flow account and scanning a QR code on their phone. Then, Google asks users to record themselves saying a string of numbers aloud and move their head around to capture every angle. This selfie-capture style will feel familiar to anyone who signed up for the Sora app, which OpenAI launched last year as an AI-firstsocial mediaplatform where people can generate and share clips of themselves. OpenAIstartlingly wound it downafter less than seven months.Unlike the Sora app, where users could generate videos of other users depending on the person’s settings, Google’s initial focus with its avatars is to let users create AI versions of themselves only, not other people. Every video generated with the Omni model, including those with your avatar, includesGoogle’s SynthID watermark.“You can capture your voice and your visual identity from multiple angles and have that show up with pretty high levels of fidelity,” Roman says. He generated a tongue-in-cheek video of himself teasing the Flow team in front of a dumpster fire, with an AI version of himself that looked lifelike and sounded like him. Then he used Flow to request changes to the generation, like a new background setting and a different-colored shirt, and Omni Flash adjusted the clips while preserving the avatar’s details.This isn’t the first time Google has rolled out a version of self-controlled deepfake video tools for creators—last month,YouTube Shorts addeda limited option for users to make similar AI avatars that can be inserted into clips on that platform. Other Silicon Valley companies are also looking for ways to transform creators’ outputs using generative AI. For example, last year, Meta rolled out an AI feature that can seamlessly translate Instagram Reels into different languages, evenadjusting creators’ lipsto match the different voices.While these AI tools may streamline aspects of the content production pipeline for creators—you don’t even have to get out of bed now to generate sassy vertical videos—generative AI isincreasingly polarizingaudiences who see these videos as inauthentic or not aligned with their values. Well, that’s if they actually clock the videos as AI.

Will KnightBusinessMay 19, 2026 1:45 PMDemis Hassabis Thinks AI Job Cuts Are DumbThe CEO of Google DeepMind tells WIRED that companies should use the productivity gains of AI to do more, not lay people off.Photograph: Chris Jung/NurPhoto/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyDemis Hassabis, theCEO ofGoogle DeepMind, is keen to talk about the coding skills of his company’s newest model,Gemini3.5 Flash. The model has been trained to perform complex agentic coding tasks: translate large code bases from one language to another; find and fix bugs lurking deep in knotty code; and even write entire operating systems from scratch.Hassabis does not, however, think this spells doom for software developers. “I have no idea why people are going around talking with certainty about that,” Hassabis tells WIRED ahead of the new model reveal attoday’s Google’s I/O event.“Perhaps there is an ulterior motive for putting those messages out; raising money or whatever,” Hassabis says. “From my point of view, from DeepMind and Google’s point of view, if engineers are becoming three or four times more productive, then we just [want to] do three or four times more stuff.”Got a Tip?Got a tip? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at wak.01.The striking coding abilities of the latest models has led to widespread fear that AI may be on the brink of eliminating programming roles and other white-collar jobs. Executives at some AI companies have predicted widespread job displacement, while some prominent tech companies, includingAmazon,Salesforce, andBlock, haveblamed recent layoffson the use of AI.Hassabis thinks that Alphabet, which oversees several companies besides Google, may be well positioned to take advantage of a revolution in software productivity. “I have a million ideas, from lab drug discovery to game design,” he says. “I’d love to have some free engineers to go and do those kinds of things.”Hassabis says that companies looking to replace developers with AI may be making a big mistake. “I think it’s a lack of imagination—and a lack of understanding of what’s really going to happen,” he says.Google revealed a raft of AI stuff at its annual developer event. Through a coding tool called Antigravity, Gemini 3.5 Flash offers frontier coding and reasoning capabilities but is faster and cheaper than the offerings of its competitors, Google says. Gemini 3.5 Pro, a more powerful new version of its flagship model, will debut next month.The company needs to catch up when it comes to AI coding, which has emerged as a crucial and lucrative application for the latest AI models. Anthropic and OpenAI lead developer adoption with their respective tools, Claude and Codex, according toa 2025 Stack Overflow survey.The company also demoed an agentic assistant called Spark that lives in Google’s Cloud and has access to its apps. The design is meant to be safer than something like OpenClaw because it has limited access to personal data, Google says.Other agentic demos included a version of Android with an AI agent built in and a refreshed version of Google Search that uses agentic coding to generate a site or app on the fly in response to a search query.AI coding has captivated the AI world in recent months, even inspiring hope that models could one day rewrite their own code in a self-improvement loop. Hassabis says it’s possible but doubts that it will immediately lead to superhuman-level AI.Progress in other areas of science might require AI models to have a deeper understanding of the physical world and even an ability to perform experiments within it, he says.Even within the seemingly solved world of coding, Hassabis says it is notable that AI has yet to produce a blockbuster app or video game without human help. “I think there’s something missing,” he says.

Paresh DaveLauren GoodeZoë SchifferBusinessMay 19, 2026 3:00 PMMeta Employees Are Scrambling to Use Up Benefits Ahead of LayoffsOn the eve of about 8,000 jobs being cut, employees are cashing in on headphone stipends and other perks while they still can.Photograph: Benjamin Fanjoy/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAhead of Meta’slatest round ofmass layoffs tomorrow, some employees are deserting offices, abandoning their work, and loading up on perks they might soon lose, several people at the company tell WIRED.Two employees describe a widespread rush to use up an annual $2,000 flexible benefit, which can cover a variety of expenses including health and wellness activities. A separate triennial credit of $200 toward the purchase of audio gear has led to a scramble to purchase Apple AirPods and other headphones. Another source says Meta offices have been largely empty this week, as people prioritize polishing their résumés and gather offsite to commiserate with friends for what may be their final time as colleagues. Employees are variously “paralyzed,” “coasting,” and “panicked,” sources say.Meta plans to lay off about 10 percent of its nearly 80,000 employees on Wednesday, with notices going out to affected workers’ personal and corporate email addresses at 4 am in their local time zone, according to a company-wide memo sent on Monday. The cuts are coming at a time when the social media giant behind Instagram, WhatsApp, and Facebook is enjoying record-high profits.But CEO Mark Zuckerberg insists that the company must free up cash to invest in AI data centers, and that Meta can perform just as well with fewer employees because of AI technologies that augment human labor.Got a Tip?Are you a current or former Meta employee who wants to talk about what’s happening? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at peard33.24 and ChaoticGoode.12.Meta didn’t immediately respond to a request for comment for this story. The company has undergone three previous large rounds of layoffs since 2022, including as part of Meta’s one-time “year of efficiency” drive in 2023. But even though the latest round is smaller than a couple of those, it is drawing widespread scrutiny because it comes at a time of societal anxiety aboutAI’s impact on jobs.Inside Meta, the imminent cuts are among several concerns that have sunk morale to unprecedented depths, according to 16 current and former employees whorecently spoke to WIRED. Employees also have been frustrated by being “drafted” onto a new AI team without any choice andthe rollout of surveillance softwarethat tracks US workers’ laptop use to train AI models.Meta also plans to internally restructure as it conducts sweeping layoffs, transferring 7,000 remaining staff to “AI initiatives” and converting more managers into individual contributors. That would bring the total number of those affected—either laid off or placed in a new role—to 20 percent of the current workforce, Reutersreportedon Monday. WIRED independently confirmed this reporting. Some parts of the company have been told they won’t be affected at all.But in recent days, employees who are bracing for changes have shared checklists internally about benefits to take advantage of, and are saving documents such as performance reviews and pay stubs, according to one worker. Some teams are meeting up at bars and restaurants near Meta offices in New York and Menlo Park on Tuesday and Wednesday to eat and drink away their sorrows, several employees said. Management has encouraged employees not to come into offices on Wednesday.

Boone AshworthMichael CaloreGearMay 19, 2026 4:00 PMEverything Announced at Google I/O 2026Google is sprucing up its Gemini models, revamping search, and enabling AI agents in everything. There are also some spiffy new smart glasses coming this fall.Google and Alphabet CEO Sundar Pichai on stage in 2025.Photograph: Camille Cohen/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle just wrappeditskeynoteaddress at its annualI/O developer event. The company showed off a swath of newagentic AIfeatures and some demos of its upcoming Android-powered smart glasses.As it has in the past few years, the spectacle largely revolved around Google’s perpetual stream of AI efforts. The company says that 900 million people use its Gemini assistant, and people have generated more that 50 billion images with Gemini.Google’s goal for 2026 is to put AI agents at the forefront of all its biggest services: Search, Gmail, YouTube, Docs, and the Chrome browser. In a demo briefing the day before I/O, Google and Alphabet CEO Sundar Pichai said the company was in a period of “hyper progress” with its AI efforts, but acknowledged that this is the part of that cycle “where people want to see real value in the products they use on a day-to-day basis.”Here is everything Google announced at I/O 2026. And if you’re wondering where all the Android 17 news is, Googletold us all about that last week.Search PartyGoogle is going all in on keeping people on Search by trying to integrate it into every part of your life. The company hasembedded its AI agents directly in Google Searchin something it calls the “intelligent search box.” This new search experience starts rolling out to everyone today.When you ask the search box questions like, “What is a black hole?” or “Are Google AI Overviews devastating the journalism industry,” it will respond with more contextual answers or use generative AI to create images or short video clips that help explain the concept. As an example, Google showed a search about black holes, and its AI agent generated a video that visually explains the process, plopping it right into the search results.Search is also getting a Generative UI feature, which creates different ways of viewing information from search results, so different types of responses—videos, images, news articles—get custom layouts that are generated in the browser on the fly, based on what’s the most relevant. Generative UI will start showing up in everyone’s Google searches this summer.Search agents are also being integrated more deeply across Google’s platforms. In March, Google releasedAsk Maps, which lets users ask questions on Maps like you would interact with a chatbot.Gemini EverywhereGoogle has two new major AI model updates. Gemini 3.5 is officially being released today, along with Gemini 3.5 Flash, a pared-down but more affordable version of 3.5 Pro. Both will be available immediately in Google’s Search and the Gemini app.Google’s core Gemini app is getting a refresh, with a snazzier redesign called Neural Expressive. It has new colorful backgrounds, a new typeface, and new animations for live voice chats. It also has many more accent options for your chatbot’s voice, with regional accents in several languages.Daily Brief is an upcoming personalized digest of your day that Google wants to be the first thing you check in the morning. Daily Brief pulls info and data from your calendar, email, and other information to summarize and prioritize your plans for the immediate future. (Google tested out a similar feature in its experimentalGoogle Labsin December.)The company is also adding new Gemini-powered features to its other services. An Ask YouTube feature lets you pose natural search queries in YouTube. Ask it how to ride a bike or how to grill a steak, and an agent will then be tasked with finding a mix of relevant YouTube videos. It’ll even snap right to the points within the videos that answer your questions.Much weirder is voice editing in Google Docs, in a new feature called Docs Live. By describing with your voice what you want to write, an agent will dictate your words, generate text, pull in citations from the web, and aim to turn your stream-of-consciousness wishes into a coherent document.(Reminder: All this stuff mayeventually have ads.)For Gemini power users, Google is creating a new subscription tier, the AI Ultra plan, for $100 a month. It is also dropping the price of its top Gemini AI Ultra from $250 a month to $200.Gemini OmniGoogle announced Gemini Omni, an AI video generator akin to Sora 2. That was OpenAI’s generator that let youdeepfake yourselfbut was eventuallykilled by the company.Google’s approach is building out a far more realistic video generator that can incorporate real video and extrapolate all manner of AI-powered weirdness on top of that. Google is eager for you to turn Omni’s eye on yourself, putting your face front and center. As such, selfie videos can be modified to

Brian KahnScienceMay 19, 2026 6:20 PMCalifornia’s Wildfire Season Is Already OveractiveMajor fires are threatening homes and ecologically sensitive areas following a hot, dry winter.The Sandy Fire burns on Monday, May 18, in Simi Valley, California.Photograph: AP Photo/Ethan SwopeCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyIt’s May, butCalifornia is already getting a taste of what peak fire season could look like, as out-of-control blazes pose a danger to infrastructure and some of the most threatened trees on the planet.A combination of high winds and heat has contributed to a trio of major fires in Southern California.The largest among them is the Santa Rosa Island Fire, which started over the weekend in Channel Islands National Park after a stranded sailor used flares to signal for help. The blaze has consumed roughly 16,600 acres—almost a third of the entire island. While some structures have been lost, the biggest is a grove of Torrey pines, which are among the rarest trees in the world.Torrey pines are considered critically endangered by the International Union for Conservation of Nature. Thegroup has warnedthat the grove on the island faces “high potential risk from destructive fires.” While the Santa Rosa Island Fire burned through the grove, there’s hope that the worst-case scenario may not have come to pass.“Upon initial assessments, fire crews are observing that the fire intensity was low and that the stand remains intact,” Mike Theune, the fire information officer assigned by the federal government, writes in an email. “When safe to do so, a fire effects crew will be assigned to make a full determination of condition and any long-term effects.”While smaller, the Sandy Fire has forced thousands to evacuate in and around Simi Valley, located about 35 miles northwest of downtown Los Angeles. Fire activity increased on Tuesday morning as high winds whipped the flames across nearly 1,400 acres.The fire is only 5 percent contained, while the Santa Rosa Island Fire isn’t contained at all. Because homes and businesses are threatened, the Sandy Fire has received more aerial resources to quell the flames. The River Fire has also burned 3,535 acres in Kern County and is 15 percent contained.In the early stages of fire season, nearly 41,000 acres have burned across the state, according to the California Department of Forestry and Fire Protection. But that’s well above the five-year average of 23,380 acres burned by this date and could be a harbinger of what’s to come in the state and across the West.Record-breaking heatobliterated already-meager snowpackthis spring, leaving states primed to burn. The most recent snowpack measurements show California’s Sierra Nevada range harbors just 9 percent of its usual snowpack for this time of year, while many basins in other parts of the West have no measurable snow at all, according to federal data.Hot weather is, of course, a hallmark of climate change. Arecent analysisby the nonprofit Climate Central found that April 1 snowpack—a key date for measuring what’s on the ground—has declined across the West by 18 percent since 1955.All of which is to say that while the early season fires are bad, the West hasn’t seen anything yet.

Miles KleeCultureMay 19, 2026 6:53 PMLiterary Prizewinners Are Facing AI Allegations. It Feels Like the New NormalThree of five regional winners of the prestigious Commonwealth Short Story Prize are suspected of relying on chatbots. They’re certainly not alone.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAt first, thewinners of the prestigiousCommonwealth Short Story Prizefor 2026 enjoyed the envy of their peers. But since their works of fiction earned this distinction, these authors have found themselves facing harsh scrutiny from the literary community, with several accused of enlistinggenerative artificial intelligenceto write for them.The allegations have come from numerous readers, many of them writers themselves, expressing bafflement and dismay that the prize jury could have overlooked potential signs of inauthentic authorship.Each year, the Commonwealth Foundation, a nongovernmental organization in London, awards its short story prize to one writer in each of five regions: Africa, Asia, Canada and Europe, the Caribbean, and the Pacific. One overall winner is then selected from that short list. Regional winners take home £2,500 (about $3,350), while the top winner, to be announced next month, claims £5,000 (about $6,700).On May 12, the respected UK literary magazineGranta published the top five 2026 entries—all previously unpublished, per the rules of the contest—on its website. (It has hosted the winning submissions for the prize since 2012.)Within days, however, one entry aroused suspicion. “The Serpent in the Grove,” a story by Jamir Nazir of Trinidad and Tobago, which had taken honors for the Caribbean region, struck a few people as bearing the stylistic tells of AI-generated text.“Well, this is a first: a ChatGPT-generated story won a prestigious literary prize,” wrote researcher and entrepreneur Nabeel S. Qureshi, a former visiting scholar of AI at the Mercatus Center at George Mason University, in aposton X on Monday. “‘Not X, not Y, but Z’ sentences everywhere, the ‘hums’ trope, and plenty of other obvious markers of AI writing. A major milestone for AI, at any rate…”“They say the grove still hums at noon,” Nazir’s mysterious and atmospheric tale begins. In his screenshot of the opening paragraphs, Quereshi highlighted the second line as what he considered to be a signature example of AI syntax: “Not the bees’ neat industry or the clean rasp of cutlass on vine, but a belly sound—as if the earth swallows a shout and holds it there.”As the literary community undertook a closer read of Nazir’s story, many criticized itslanguage and metaphorsas nonsensical, wondering how the Commonwealth judges could have seen any merit to them. Others sharedscreenshotsshowing that the AI-detection tool Pangram flagged “The Serpent in the Grove” as 100 percent AI-generated, a result that WIRED independently confirmed. (While no AI-detection software is perfect,third-party analysishas consistently determined Pangram to be the most accurate, with a near-zero rate of false positives.)Nazir did not return a request for comment relayed through an email address listed on hisFacebook page. The posts on that account and theLinkedIn profileof a Jamir Nazir in Trinidad and Tobago also scan as AI-generated on Pangram. Although some speculation had it that Nazir himself could have been an entirelyAI-created persona, a2018 articlein the Trinidad and Tobago edition of the The Guardian about his self-published poetry collectionNight Moon Love—which includes a photograph of Nazir holding the book—suggests that he is a real person.WIRED contacted both Granta and the Commonwealth Foundation about Nazir’s story; neither commented directly, but both issued public statements.‘We are aware of allegations and discussion regarding generative AI and our Short Story Prize,” wrote Razmi Farook, director-general of the Commonwealth Foundation, in astatementon the organization’s website. “We take these claims seriously and are committed to responding to them with care and transparency.” Farook defended the judging process for the prize as “robust,” with multiple rounds of readers and the top-level judges selected for their “expertise.”“We do not currently use AI checkers in our judging process, because this is a prize for unpublished fiction,” Farook explained. “To supply unpublished original work to an AI checker would raise significant concerns surrounding consent and artistic ownership. We also do not use AI to judge stories at any stage of the process. When they submit stories to the Prize, writers accept our entry rules and guidelines. These include confirming that their submission is their own original work. All short-listed writers have personally stated that no AI was used and, upon further consultation, the Foundation has confirmed this.”Theentry and eligibility rulesfor the 2026 Commonwealth Short Story Prize make no mention of artificial intelligence, stat

TechGadgetsNewsUgreen’s new soccer ball-shaped tracker has up to 7 years of battery lifeBut when its non-replaceable battery finally dies, you’ll need to kick the FineTrack 2 into the e-waste bin.But when its non-replaceable battery finally dies, you’ll need to kick the FineTrack 2 into the e-waste bin.byAndrew LiszewskiMay 19, 2026, 7:49 PM UTCLinkShareGiftIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.The FineTrack 2 isn’t as discreet as an AirTag.Image: UgreenAndrew Liszewskiis a senior reporter who’s been covering and reviewing the latest gadgets and tech since 2006, but has loved all things electronic since he was a kid.Ugreen has announced a new Apple Find My-compatible tracker with a novel design that limits where you can hide it, but greatly benefits battery life. Inspired by “major football tournaments” the company doesn’t call out by name, the Ugreen FineTrack 2 is shaped like a soccer ball and is designed to be used hanging off your keys or bags. It’s much larger than an Apple AirTag or credit card-sized trackers, but still small enough to bury in a suitcase or stick in your car’s glove compartment.The waterproof FineTrack 2 isavailable now for $19.99and comes with a 600mAh CR2450 battery good for around five to seven years of use with Apple’s Find My network. Ugreen’s trackers often come with excellent battery life. It updated its 1.7mm thickcredit card tracker with support for Google’s Find Hub networkand five years of battery life last September, but once it dies, it becomes e-waste. It’s the same story for the FineTrack 2. It’s not rechargeable and its battery isn’t replaceable. An AirTag’s battery might only last for around a year, but you can easily swap in a new one as often as you need.The FineTrack Mini 2 is a smaller version of Ugreen’s ball-shaped tracker.Image: UgreenAlongside the FineTrack 2 Ugreen also announced its new FineTrack Mini 2. Its functionality and battery life are identical to the ball-shaped version including five to seven years of (non-replaceable) battery life, waterproofing, and a loud 110dB alarm. But it features a smaller, lighter, and more traditional tracker design that’s more discreet and easier to stash away. It’sonly being sold in four-packs for $45.99.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Andrew LiszewskiGadgetsNewsTechMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

PolicyEntertainmentAntitrustDemocrats preview how they’d go after the Ticketmaster settlement if they regain powerAnd California’s AG says an even broader breakup of Live Nation is on the table.And California’s AG says an even broader breakup of Live Nation is on the table.byLauren FeinerMay 19, 2026, 8:13 PM UTCLinkShareGiftImage: Cath Virginia / The VergePart OfUSA v. Live Nation-Ticketmaster: all the news on the lawsuit that could take down a giantsee all updatesLauren Feineris a senior policy reporter at The Verge, covering the intersection of Silicon Valley and Capitol Hill. She spent 5 years covering tech policy at CNBC, writing about antitrust, privacy, and content moderation reform.A handful of Democrats called an unofficial hearing on Capitol Hill Monday to slam the Department of Justice’s “trivial” and “pathetic” settlement with Live Nation-Ticketmaster, previewing how they might go after Trump administration antitrust deals if they win back congressional power in November.While the DOJsettlementwith Live Nation, which came one week into trial and with a promise of up to $280 million, received copious pushback, Democrats don’t at the moment have the power in either chamber to set committee agendas. That meant that Monday’s so-called shadow hearing looked different than most official proceedings. Instead of sitting at the elevated seats behind the dais, the lawmakers sat at tables below, eye level with witnesses including California Attorney General Rob Bonta, one of the AGs continuing the fight against Live Nation-Ticketmaster, ousted Trump administration antitrust official Roger Alford, and several entertainment industry players, including a member of the band The Hold Steady. Several artists asked to appear at the forum declined out of fear for their livelihoods, according to House Judiciary Committee Ranking Member Jamie Raskin (D-MD), who co-hosted the forum with Senate Homeland Security permanent subcommittee on investigations Ranking Member Richard Blumenthal (D-CT).While their statements won’t be formally entered into the congressional record, the forum gave Democrats a way to publicize their concerns, which included that the DOJ’s deal was the result of undue corporate influence. “This isn’t like some kind of pickup game,” Raskin told reporters. “The corruption permeates the administration so much that we have to start building the record because it’s such an overwhelming task. We will be using the products of all of these shadow hearings to figure out where to go.” As a signal of the significance of the forum, Senate Minority Leader Chuck Schumer (D-NY) delivered opening remarks to kick it off.“The corruption permeates the administration so much that we have to start building the record because it’s such an overwhelming task”The lawmakers pointed to the states’resounding victory in the jury verdictas evidence that theDOJ’s settlementfell far short of what a good faith deal would include, charging it was a result of corruption. On the11-page verdict form, jurors checked “yes” on every question about whether the states proved their claims. The DOJ has stood by its deal as a win for consumers. The deal will still need to undergo a Tunney Act review, which is meant to ensure that it’s in the public interest.Judges are typically loath to reject a settlement, which is why Sen. Amy Klobuchar (D-MN) hasproposed legislationto strengthen the standard. But Alford, who wasfired from the DOJandlater spoke outabout opposing lobbyist influence in antitrust deals, said this could be the rare case where the judge may reject the DOJ’s agreement, given that the same judge is now tasked with coming up with appropriate remedies to resolve Live Nation’s monopoly, potentially including a breakup.“Divestiture, it’s the nuclear option”Bonta said the states’ breakup request, expected to be formally submitted this week, may go beyond a separation of Live Nation and Ticketmaster. They’re also considering asking for at least partial divestitures of parts of Live Nation’s venue or artist management businesses. “Divestiture, it’s the nuclear option. We believe it’s warranted, but we want to be thoughtful how we ask this judge for an appropriate remedy,” he told reporters.“Opinions shared yesterday misrepresent how the live events industry operates,” Live Nation’s executive vice president of corporate and regulatory affairs, Dan Wall, said in a statement. “Secondary ticketing and high-demand onsales are industry-wide challenges that exist across every promoter, venue, and ticketing platform. That’s precisely why Live Nation and Ticketmaster have long supported industry-wide resale reform, including price caps and stronger consumer protections, while continuing to invest in tools that help artists manage demand and protect fans.”Live Nation haspromised to appealthe jury verdict, and even before it does it’s possible the judge will deny the states the full extent of the remedies they ask for. But Monday’s forum showed tha

GadgetsTechVerge ShoppingHere are our favorite Memorial Day deals (so far)We found excellent deals on some of our favorite TVs, noise-canceling earbuds, portable speakers, and other gadgets.We found excellent deals on some of our favorite TVs, noise-canceling earbuds, portable speakers, and other gadgets.bySheena VasaniMay 19, 2026, 8:25 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Portable bluetooth speakers like the Charge 6 and Flip 7 are just a few on sale for Memorial Day.Sheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.Memorial Day marks the unofficial start of summer, and the holiday’s sales include plenty of deals on gadgets that can help you make the most of the season. If your plans involve beach days, pool parties, or backyard barbecues, now’s a great time to pick up aportable Bluetooth speaker, as many of our favorite models are currently on sale. You’ll also find worthwhileTV deals, along with discounts onsmart thermostatsto help keep your home cool and your energy costs low. And, if you’re planning on heading out of town, there are deals onnoise-canceling headphones,portable chargers, and other travel-friendly gadgets.RelatedHere are 40 of our favorite deals from REI’s massive Anniversary SaleThe Verge’s 2026 college graduation gift guideBelow, we’ve rounded up the best Memorial Day deals currently available. To give you more options as you shop, we’ve included good deals from a range of retailers, including ones that aren’t running their own Memorial Day sales.Featured ProductsPreviousNextBose QuietComfort Ultra Earbuds (2nd Gen)Read More$249 at Amazon$249 at WalmartLG B5 OLED TVRead More$800 at Amazon (55-inch)$650 at Best Buy (48-inch)Anker Laptop Power BankRead More$96 at Amazon$96 at WalmartStar Wars OutlawsRead More$36 at Amazon (Switch 2)$20 at Walmart (PS5)Ray-Ban Meta (Gen 1) smart glassesRead More$224 at Amazon$224 at Best BuySonos Move 2Read More$399 at Amazon$399 at WalmartBose QuietComfort Ultra Headphones (first-gen)Read More$299 at Amazon$299 at Best BuyRing Battery Doorbell Plus (second-gen)Read More$120 at Amazon$120 at Best BuyEarbuds and headphonesBose QuietComfort Ultra Earbuds (2nd Gen)$249$29917% off$249The second-gen QuietComfort Ultra Earbuds have dynamic, rich sound and the best noise cancellation currently available in a pair of true wireless earbuds. They’re a terrific pick if you want to listen to your music in peace.Read our review.Read More$249 at Amazon$249 at Walmart$249 at Best BuySamsung’s Galaxy Buds 3 Proare on sale for $149.99 ($100 off) atBest Buy, which matches their all-time low price. The wireless earbuds are ourtop pick for Samsung Galaxy phone device owners, thanks to their comfortable AirPods-inspired design, terrific sound quality, and good noise cancellation. They also offer helpful voice controls that let you adjust volume or playback without a wake word, while delivering excellent voice call quality.Read our review.Nothing Ear (a)$59$9940% off$59Nothing’s Ear (a) earbuds have the same general design as the company’s previous earbuds — but you can get these in a snazzy yellow. They include ANC, more than decent sound quality, and other features like multipoint for just under $100.Read our impressions.Read More$59 at Amazon$59 at NothingSennheiser’s Momentum True Wireless 4earbudsare down to $199.99 ($100 off) atAmazonandBest Buy, which is $20 shy of their best price to date. If sound quality matters most, they’re an easy recommendation, delivering rich, detailed audio with EQ customization to fine-tune the listening experience. Optional wing tips are included to help keep them secure and comfortable.Read our review.Bose QuietComfort Ultra Headphones (first-gen)$299$42930% off$299The older QC Ultra were a replacement for the Noise Cancelling Headphones 700 with a more travel-friendly design, spatial audio, and better call quality. Thankfully, they still offer excellent comfort and some of the best noise cancellation available.Read our review.Read More$299 at Amazon$299 at Best BuySony WH-1000XM5$248$40038% off$248Despite being a last-gen model, these headphones still offer impressive sound quality and top-notch noise cancellation. They’re comfortable to wear for hours, can last up to 40 hours on a charge, and support extras like audio sharing and Gemini Live.Read our review.Read More$248 at Amazon$248 at Best Buy$250 at TargetAmazonandBest Buyare selling thelast-gen AirPods Max (USB-C)for $449 ($100 off), which is $50 shy of their all-time low. Apple’s luxe over-ear headphones offer a premium design and work especially well with other Apple devices. They still deliver great sound quality and effective noise cancellation, even if thenewer modelhas a slight edge.Read our review.Shokz OpenRun Pro 2$130$18028% off$130The OpenRun Pro 2 are Shokz’s best bone conduction headphones yet, with bet

TechAIReportGoogle’s AI future demands trust — and your personal dataGemini Spark is giving Google a whole new level of access to your information.Gemini Spark is giving Google a whole new level of access to your information.byEmma RothMay 19, 2026, 9:00 PM UTCLinkShareGiftImage: GooglePart OfGoogle I/O 2026: All the news and announcementssee all updatesEmma Rothis a news writer who covers the streaming wars, consumer tech, crypto, social media, and much more. Previously, she was a writer and editor at MUO.Google has big promises for its AI-powered future — and a lot of it depends on your trust. At I/O 2026, Google described a bunch of new tools that it claims will make your life easier.Gemini Spark, Google’s always-on AI agent, can help organize an upcoming event, while Daily Brief can offer a rundown of what to expect during your day. Google is even expanding access toGmail’s AI inbox, which can generate custom to-do lists and draft personalized replies based on your emails.Many of these features seem genuinely useful, but at the heart of each of them is an AI engine that runs on a trove of personal information. While other AI companies, like OpenAI, Microsoft, and Anthropic, let you connect other apps and data that you use, Gemini’s access to the personal data already stored across Google’s services lies behind a simple opt-in menu —one of its key advantages in the AI race.You can ask Gemini Spark to review your inbox at a set time each week.Image: GoogleGoogle first started dipping into personalization in 2024, when it integratedGemini into Workspaceapps like Gmail, Docs, Sheets, Slides, and Drive in 2024, allowing its AI chatbot to do things likesift through your filesordraft an email. Gemini’s Deep Research feature can eventap into your emails, Drive, and chatsand use them as sources for its reports.Over the past several months, Google has continued toexpand these integrations. Itintroduced “Personal Intelligence”in January, a feature that allows Gemini to reason across Gmail, Google Photos, Search, and your YouTube history without prompting. That means Gemini can automatically surface details from across your accounts to personalize its responses. “Millions of people are using it [Personal Intelligence] every single day, they found it so helpful for things like personalized product and trip recommendations, or acting as a thought partner for navigating big decisions in life, like a career change,” Josh Woodward, the head of Google Labs, the Gemini app, and AI Studio, said during I/O 2026.Even though it’s completely optional to connect your Workspace apps, Search history, Photos, and other information to Gemini, it seems like Google’s AI future hinges on people doing exactly that. Daily Brief, which is rolling out to Google AI Plus, Pro, and Ultra subscribers, scans for updates from your Gmail and flags events from your Calendar.ButGemini Sparkis diving deeper into your information, as Google is pitching it as an AI personal assistant that can work across connected Workspace apps 24/7, creating continually updated study guides, generating to-do lists based on meeting notes, and even automatically scanning monthly credit card statements for hidden subscription fees. But the connections to Workspace apps are just the beginning, as Gemini Spark will also be able to plug into third-party services, like Canva, OpenTable, Instacart, Spotify, Expedia, Adobe, and more.Google is even planning to give Gemini Spark access to local files on Mac computers, similar to OpenClaw, the open-source AI agent platform thatposes a range of security risks. During a demo at I/O, Woodward showed how he could use Spark to send an email to a dog boarder in preparation for an upcoming trip. He selected documents across his computer and asked Spark to craft an email using his dogs’ allergies and vaccination records.Many people might draw the line at giving an AI system access to their whole computer. But if therise of OpenClaw tells us anything, it’s that AI is moving from being a novelty to a real productivity tool that demands access to our digital lives. It’s just a matter of whether people trust the companies behind the systems enough to hand over their personal data — and more importantly, where they’ll set a boundary on what’s too private.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Emma RothAIGoogleGoogle I/O 2026ReportTechMore in:Google I/O 2026: All the news and announcementsWear OS 7 will keep track of deliveries and sports scores on your wristJay Peters24 minutes agoDemis Hassabis said this might be the ‘foothills of the singularity.’ What?Stevie BonifieldTwo hours agoThe future of Google is a search box that does everythingJay PetersTwo hours agoMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes

GadgetsGamingNintendoNintendo’s $500 Switch 2 bundle includes a game, and it’s available nowYou’ll save up to $30, depending on the game. It’s a good deal if you’re in the market for the console.You’ll save up to $30, depending on the game. It’s a good deal if you’re in the market for the console.byCameron FaulknerMay 19, 2026, 9:00 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Mario Kart World is one of three games you can choose from in the bundle.Photo by Amelia Holowaty Krales / The VergeCameron Faulkneris an editor covering deals and gaming hardware. He joined in 2018, and after a two-year stint at Polygon, he rejoinedThe Vergein May 2025.Nintendo recently teased the“Choose Your Game” Switch 2 console and digital game bundle, coming in early June. However, multiple retailers (including Nintendo itself) are already offering the $499.99 bundle. The list currently includesGameStop,Amazon,Best Buy, and theNintendo Store. It’s a deal worth hopping on ahead of the holiday season, since Nintendo is raising the price of the Switch 2 to $499.99 in September.While theSwitch 2console currently costs $449.99 by itself, the bundle includes a full-price game with up to a $30 discount. You can choose between one of three digital games to redeem upon setup:Donkey Kong Bananza,PokémonPokopia, andMario Kart World. Each of those are regularly $69.99 except forWorld, which retails for $79.99.RelatedNintendo is raising Switch 2 prices in SeptemberI tested dozens of cases for the Switch 2 and these are the bestNintendo Switch 2 “Choose Your Game” bundle$500$5306% off$500Choose betweenDonkey Kong Bananza,Mario Kart World, andPokémon Pokopia.Read More$500 at Amazon$500 at GameStop$500 at Best BuyNone of the games offered as part of the bundle will take up much of the Switch 2’s 256GB internal SSD.Mario Kart Worldtakes up the most space of the three, requiring 22GB. However, you may eventually own enough games to exceed its capacity, so we recommend considering a256GB Samsung microSD Express cardto double your available storage. There’s currently a deal on those happening atAmazon, where you can buy one for $47.99 (originally $79.99).To redeem your game, enter the eShop from the Switch 2’s home screen. Then select “enter code” on the left side of the page. You’ll then enter the 16-character code provided in your bundle, then make your game selection.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Cameron FaulknerDealsGadgetsGamingNintendoVerge ShoppingMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAINewsThe future of Google is a search box that does everythingGoogle, Gemini, Gmail — it feels like they could one day become a single search box.Google, Gemini, Gmail — it feels like they could one day become a single search box.byJay PetersMay 19, 2026, 9:24 PM UTCLinkShareGiftPhoto by Allison Johnson / The VergePart OfGoogle I/O 2026: All the news and announcementssee all updatesJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Last year, after watching Google’s I/O keynote, I wrote that it felt likeGoogle’s future was Google googling. After watching this year’s I/O keynote on Tuesday, I don’t think Google just wants to google for you — I think it wants to doeverythingfor you, all from a search box.Take the trustyGoogle search bar itself, something Google is generally hesitant to update, which is getting some updates. It will “dynamically” expand as you type longer queries. It will offer “AI-powered suggestions” that Google claims will “go beyond autocomplete,” which could cause you to fill in the blanks of a search in a way you didn’t intend and that may or may not be helpful.Or how about what Google actuallyshows in search results? From AI Overviews you can keep asking questions in AI Mode, which generates a custom page with an AI-generated summary of what you’re searching about instead of showing you a traditional list of links. Search results are also going to be even more personalized because Google will be able to generate custom UI for you, including generating things like interactive visuals and graphs right within your search results page. You’re also going to be able to ask Google, right from the search bar, to make “information agents” that can keep track of things you care about — like new sneaker drops or apartment listings — making the search bar a kind of AI-infused Google Alert.Gemini is getting a bunch of upgrades and features, too. It can send you a “Daily Brief” that tells you about your day based on information from across your Google apps, like Gmail and Google Calendar. You can create your own custom Google-powered agents thanks to a featurecalled Gemini Spark, which could have a leg up on agents like OpenClaw because it’s a first-party offering from Google. The company has also been recently making a big kerfuffleabout Personal Intelligence, which pulls in context from your other Google apps that helps inform your responses from Gemini.In Workspace, Google wants you tojust blabat tools like Gmail and Docs and Keep, and the apps will help with things like parsing your inbox, drafting a document, or generating to-do lists. For shopping, the newUniversal Cartwill keep track of things you want to buy in apps like Search, Gemini, Gmail, and YouTube — and let you check out using Google’s payment infrastructure. And speaking of YouTube, the companyis testingan AI Mode-like experience there, too, where it will put together a page of search results instead of just showing you a list of things to watch.There’s more, too — with the newGemini Omni models, you’ll be able to create a video using things like other videos, other images, and audio as prompts. Down the line, you’ll be able to generate other media as well; the family of models is meant to be able to “create anything.”I could go on; there was a lot announced at I/O. But what I’m trying to get at here is that Google isn’t just showing you the location of information anymore. Through its various search boxes, it’s now just answering questions for you in a way that it thinks is most helpful. In the most charitable view: If it’s done well and accurately, this allcouldbe very useful. But that means it has to be done well and accurately, and that’s anextremely high bar to clear, especially with complex search queries or questions about sensitive data like my years of emails in Gmail.It’s not hard to imagine a future where, someday, Google just makes everything happen in one universal search box. No more hopping between Google Search, YouTube, Gmail, Gemini — my gut is that Google’s current end state is that you just type anything you’re looking for into an Ask Google box, and Google finds a way to make it happen and surface it for you in a personalized way.RelatedGoogle’s AI future demands trust — and your personal dataAssuming that’s the case, that’s not something I want. The fun of the internet is actually doing the work to find stuff, even if it’s sometimes frustrating, difficult, or time-consuming. Sure, the pitch for talking to Gmail to get help finding things in my inbox seems handy, but I’ve also spent years honing my own email management system — one that would work whether I used Gmail or not. It also works great forme, I think we should all have to go through the motions of figuring out the systems across all of our digital lives that work best for us instead of just relying on Google to figure it all out from one universal search box.Google

PodcastsTechGoogleWe react to Google I/O 2026Hayden and Jake discuss Spark, Antigravity, Omni, Pics, vibe coding, the new Google Search, and the singularity.Hayden and Jake discuss Spark, Antigravity, Omni, Pics, vibe coding, the new Google Search, and the singularity.byJacob KastrenakesMay 19, 2026, 9:48 PM UTCLinkShareGiftJacob Kastrenakesis The Verge’s executive editor. He has covered tech, policy, and online creators for over a decade.What better way to unwind from a two-hour keynote presentation than to pore over the weirdest and wildest details, froma Gmail bot you can converse withto DeepMind’s leadersaying the singularity is near.The Vergecastwent live right after the show, with senior AI reporter Hayden Field joining me to discuss the highlights.First up, we talk about the buzzword of the show: agents, which seem to be Google’s new answer to making AI useful. Google hasa new service called Sparkthat sounds a lot like a Gemini-flavored OpenClaw, upgrades to its Antigravity platform, and a bunch of agents that willshop for youand track search topics. We also go overGoogle’s model updates, as well as its shift to world models withGemini Omni, a theoretically ambitious new family that is off to a very focused start.Vergesubscribers, don’t forget you get exclusive access to ad-freeVergecastwherever you get your podcasts. Headhere. Not a subscriber? You cansign up here.Then, we discuss howGoogle Search is evolving— or put more plainly, being stuffed with AI features that turn Google into your final destination when browsing the web. We also talk vibe-coding Android apps inGoogle AI Studio, the newGoogle Pics app, and Google’sexpansion of access to content authentication toolsso that people can learn what is and isn’t real.We wrap up discussing Demis Hassabis’ proclamation that AGI and the singularity is near, which got surprisingly little stage time (one sentence) for such a bold statement. And finally, Hayden takes us through her thoughts on theMusk v. Altmantrial, which wrapped up earlier this week.Subscribe:Spotify|ApplePodcasts|Overcast|Pocket Casts|MoreThanks to everyone who tuned in to the livestream — we’re going to keep them coming. In the meantime, you can always reach us at the Vergecast Hotline (866-VERGE11) or send us an email at [email protected] don’t forget, you can get ad-free podcasts bysubscribing toThe Verge.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jacob KastrenakesGoogleGoogle I/O 2026PodcastsTechVergecastMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechEntertainmentGadgetsPlex is tripling the price of a lifetime pass to $750 after doubling it last yearNot your average streaming price hike.Not your average streaming price hike.bySean HollisterMay 19, 2026, 11:03 PM UTCLinkShareGiftImage: PlexPart OfStreaming keeps getting more expensive: all the latest price hikessee all updatesSean Hollisteris a senior editor and founding member of The Verge who covers gadgets, games, and toys. He spent 15 years editing the likes of CNET, Gizmodo, and Engadget.I am dying to know how much money Plex is about to make the next six weeks charging people to stream their own video from their own homes. Today, it’s giving every prospective customer until July 1st to lock in a lifetime subscription at today’s rates — before it triples the price to $750.Plex already more than doubled the price of a lifetime Plex Pass subscriptionfrom $119.99 to $249.99 last March, and it’ll triple again on July 1st to $749.99. At that price, you’d have to subscribe for 11 years (at the current annual rate) to make the lifetime sub worthwhile.Plex isn’t choosing the new number because it expects you to pay $750.In a blog post, the company says it actually wanted to stop issuing lifetime plans; it would rather push users towards recurring subscriptions instead to “sustain long-term development”. Now that the annual subscription covers a decade’s worth of Plex, it’s a win for Plex either way.But it mayespeciallybe a win for Plex if you, like me, feel sudden FOMO and consider giving Plex $250 for a lifetime subscription now. While you’d only be giving Plex around 3.5 years of annual revenue at today’s rates, it gets an instant infusion of cash. It’s like a huge sale except instead of needing to lower the price, Plex has actually raised it. (Again, the annual subscription costs double what it did last March.)From Plex’s email to existing users.Image: PlexPlex’s blog and emails to users encourage this with a hugely prominent call to action, bolded and highlighted in blue: “You have until 12:01 AM UTC on July 1, 2026, to get a Lifetime Plex Pass at the current price of $249.99 USD*here. If you’ve been considering it, now’s a great time to buy.”So does Plex’s website, which is currently topped with a banner reading: “The price of a Lifetime Plex Pass is increasing on July 1, 2026.Get yours at the existing price while you still can.Buy now”.None of that is to suggest Plex is laughing all the way to the bank. It’s a small company, its finances are private, the global advertising market is in a downturn, andit had big layoffsa few years back.Plex says monthly and annual subscription prices aren’t changing today, and nothing will change for existing lifetime pass holders. Those changes came last year. That’s also when Plex gated one of its most valuable features behind that subscription: the ability to stream your media server’s video outside your home.RelatedPlex Pass is going up in price — and now you’ll need it for remote playbackPlex makes piracy just another streaming serviceIt’s time to change your Plex password againFor free, without a subscription, you can still create collections and stream to devices on your local network, and stream remotely from other people who pay. Plex Pass also adds features that lets you skip the intros and credits on shows, remotely download your media, auto-rewind when pausing, hardware-accelerated streaming in its Media Server app,and more.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Sean HollisterAppsEntertainmentGadgetsTechTVsMore in:Streaming keeps getting more expensive: all the latest price hikesYouTube Premium is getting pricierDominic PrestonApr 10Netflix is raising prices againEmma RothMar 26Amazon Prime Video nearly doubles the price to go ad-free and stream 4K videoEmma RothMar 13Most PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechGadgetsNewsWear OS 7 will keep track of deliveries and sports scores on your wristGoogle’s newest smartwatch update could also give you up to 10 percent better battery life.Google’s newest smartwatch update could also give you up to 10 percent better battery life.byJay PetersMay 19, 2026, 11:24 PM UTCLinkShareGiftA Pixel Watch 4 running Wear OS 6, not Wear OS 7.Photo by Amelia Holowaty Krales / The VergePart OfGoogle I/O 2026: All the news and announcementssee all updatesJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Amid the flurry of today’s Google I/O announcements, Google shared detailsabout Wear OS 7, the next major update to its smartwatch platform. To help you keep track of things like deliveries and sports scores, Wear OS 7 will get the iPhone-style Live Updates that wereintroduced on Android last year— which can appear on your watch or your smartphone — and you’ll also be able to track automated tasks that an AI is working on right from your watch.Wear OS is also getting an upgrade from its widget-like Tiles for glanceable information. With the new update, Google is adding “Wear Widgets” to the platform, which look more like Android widgets and can appear in small or large layouts that “align perfectly” with Android’s 2x1 and 2x2 widget formats. There are some AI-powered features coming to Wear OS 7 as well, including the introduction of Gemini Intelligence, Google’scatch-all brandingfor personalized and proactive Gemini features, on “select watches” launching “later this year.”Various Wear OS 7 features, including Wear Widgets. Click the photo to take a closer look.Image: GoogleGoogle is also promising an “up to” 10 percent improvement in battery life for average users if you’re upgrading from Wear OS 6 to Wear OS 7. The company says that it’s investing in “power optimizations” so that users can “do more with their favorite apps,” but we’ll have to wait and see how battery life actually shakes out when Wear OS 7 is actually available.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGadgetsGoogleGoogle I/O 2026NewsSmartwatchTechWearableMore in:Google I/O 2026: All the news and announcementsDemis Hassabis said this might be the ‘foothills of the singularity.’ What?Stevie BonifieldTwo hours agoThe future of Google is a search box that does everythingJay PetersTwo hours agoGoogle’s AI future demands trust — and your personal dataEmma RothTwo hours agoMost PopularMost PopularThe 13 biggest announcements at Google I/O 2026Gemini is in danger of going full CopilotMicrosoft launches Surface Pro 12 and Surface Laptop 8 with Intel chipsAnker fixes the two worst things about power stationsGoogle I/O 2026: All the news and announcementsThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

At Google I/O on Tuesday, Google introduced Universal Cart, its agentic hub for managing shopping in one place. The tech giant also announced updates to its Agent Payments Protocol (AP2) and teased that it would bring the technology to Google products in the coming months, enabling users to authorize agents to make payments on their behalf.The announcements signalGoogle’s push to turn AI assistants from passive recommendation tools into active participantsin online commerce. By launching a centralized shopping system and building infrastructure that lets software agents complete purchases autonomously, the company is positioning itself to control more of the entire shopping journey, and potentially the relationship between consumers and the merchants competing for their attention.With Universal Cart, users can add products they’re considering from anywhere on Google — while browsing Search, chatting with Gemini, watching YouTube, or reading Gmail. Once items are added, Universal Cart tracks deals, monitors price drops, surfaces price history insights, and alerts users when items are back in stock.Image Credits:Sarah Perez/TechCrunchThe feature is built around something Google knows well, which is that most people shop across multiple devices, multiple retailers, and over the course of many days.The cart also uses AI to help shoppers make better decisions. For example, if you’re building your first custom PC, you can add parts from multiple merchants into a single cart, and Google may flag compatibility issues, such as a processor that doesn’t work with the motherboard you selected — and suggest an alternative.For frequent travelers or rewards maximizers, the feature can also surface hidden savings and help stretch your points further because it’s built on Google Wallet.Image Credits:GoogleThanks to Google’s open-standardUniversal Commerce Protocol (UCP), users can check out directly through Google with participating merchants, or transfer their items to the merchant site and complete the purchase there.Universal Cart is rolling out in the U.S. today and coming to the Gemini app this summer, with YouTube and Gmail to follow, Google says.Google also announced that UCP is expanding to more categories, like hotels and local food delivery services. UCP-powered experiences will also expand beyond the U.S. to Canada and Australia in the coming months, and later to the U.K.The more consequential announcement for the commerce industry may be AP2, Google’s protocol designed to let AI agents securely make payments on users’ behalf within defined limits. At I/O, Google detailed the guardrails users can set, including specifying the brands and products they want, and a spending limit. When those conditions are met, the agent makes the purchase automatically.Image Credits:Sarah Perez/TechCrunchGoogle says it’s bringing AP2 to its own products in the coming months. That integration would give Google direct visibility into what consumers discover, consider, and ultimately buy, and it’s a degree of commercial influence that retailers and payment processors will be watching closely.Under the hood, AP2 creates a transparent, verifiable link between the user, the merchant, and the payment processor, with encryption protecting user data throughout. The protocol also includes tamper-proof digital records that ensure the agent is always acting on the user’s behalf, and a permanent audit trail that both buyers and sellers can reference for returns or disputes.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,Commerce,Google,Google I/O,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo France

In BriefPosted:10:45 AM PDT · May 19, 2026Image Credits:Alex Wong / Getty ImagesSarah PerezAgentic app coding gets an upgrade with Google’s release of Android CLIIn addition to serving up its own flavor of vibe-coding tools for building on Android, Google announced it will also offer tools that allow AI agents, like Claude Code, OpenAI’s Codex, or its own Antigravity or Gemini in Android Studio, to accelerate Android app development.Google shared on Tuesday at its annual Google I/O developer conference that its Android CLI (command line interface) is now stable at version 1.0 and can be used by those who are leveraging AI agents to build for them — no matter what their preferred coding platform may be.The move acknowledges that many people are now building for Android with AI agents that aren’t from Google. The company is trying to find a way to make its specialized knowledge, like what’s found within Android Studio, more accessible.With Android CLI, AI agents can retrieveknowledgeabout Android development via a new “android studio” command that can tap into the capabilities ofAndroid Studiowhen building their own project. From there, the agents will be able to tap into a range of other commands and tools.Google Antigravity, the company’s agentic development platform, will include an optional bundle that will install the tools and knowledge included in Android CLI, so it can also perform the core tasks for Android app development, Google says.TopicsAI,Android,Apps,coding,Google,google io,google io 2026,In BriefMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAppsGoogle’s AI Studio now lets anyone build Android apps in minutesSarah Perez5 hours agoAIGoogle introduces Gemini Spark, a 24/7 agentic assistant with Gmail integration, at IO 2026Russell Brandom5 hours agoAIGoogle launches Antigravity 2.0 with an updated desktop app and CLI tool at IO 2026Ivan Mehta5 hours agoLatest in AppsAppsGoogle just declared itself a contender in AI design at IO 2026Aisha Malik1 hour agoAIYou can now talk to your Gmail inbox, as seen at Google IO 2026Sarah Perez1 hour agoAIGoogle Search as you know it is overSarah Perez5 hours ago

At its annual Google I/O event on Tuesday, Google announced that the Gemini app is getting a series of new updates, including a “Daily Brief” feature, a redesigned interface, access to a new AI video model calledGemini Omni, and new personal AI agent calledGemini Spark.The updates signal Google’s push to turn its Gemini app into an all-purpose AI hub rather than a stand-alone chatbot and to make the AI assistant more competitive with apps like ChatGPT and Claude.Google describes the new Daily Brief feature as a personalized digest designed to be your first stop each morning. It pulls together information from a user’s inbox, calendar, and most important tasks, and then organizes them into a clear overview. The company says Daily Brief doesn’t just summarize this information — it also prioritizes tasks and suggests next steps, with the most important items shown first.Daily Brief is rolling out today to Google AI subscribers in the United States.Image Credits:GoogleThe Gemini app already has considerable reach: The company said it has more than 900 million monthly users and is available in over 230 countries and in more than 70 languages. But Google clearly wants more and without losing its existing customers.Google said it rebuilt the app from the ground up. Now when users open the app, they’ll see a new design language called “Neural Expressive,” which brings fluid animations, vibrant colors, new typography, and haptic feedback.Image Credits:Sarah Perez/TechCrunchGemini’s responses are no longer presented as a wall of text, which is the case with most AI chatbots. Instead, key information appears in bold at the top. Additional text and possibly other elements such as images and timelines appear as the user scrolls down.As for Gemini Spark, Google describes it as a 24/7 personal AI agent that helps you navigate your digital life. Spark transforms Gemini from an assistant into an active partner that does real work on your behalf. Since it’s a cloud-based agent, Spark keeps working in the background even when you lock your phone. With Gemini Spark in the Gemini app, users will be able to create their own custom workflows.Spark is currently in testing and the company expects to make it available to Google AI Ultra subscribers next week.Image Credits:Sarah Perez/TechCrunchThe company’s new AI video model,Gemini Omni, combines Gemini with Google’s generative media models to create outputs grounded in knowledge. For example, you could give it a simple prompt like “claymation explainer of protein folding.” The model lets you upload audio, images, and video to generate a consistent, high-quality video, according to Google.By adding access to a new video-generation model like Gemini Omni, Google is intensifying competition among major AI platforms in the ongoing race to lead multimodal content generation. The model is rolling out to Google Flow and YouTube Shorts for Google AI subscribers, underscoring the company’s broader push into multimodal content creation and AI-powered video tools.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,gemini,Google,Google I/O,google i/o 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

The AI coding boom is now coming directly for Android app development. On Tuesday atGoogle IO 2026, the companyannouncednew native Android app creation capabilities in its web-basedGoogle AI Studio,shrinking a process that takes weeks of setup and coding down to minutes.The company also said that consumers will be able to use Gemini AI to find the apps they need, both on the Play Store and the web, expanding opportunities for developers to have their apps discovered.Google says the new capabilities could make sense for anyone from a seasoned developer looking to prototype a new app quickly to a first-time creator.By offering the ability to essentially vibe-code Android apps via web-based tools, Google is ramping up the competition with other AI-powered development tools, like Cursor, Replit, Lovable, Claude Code, and others, while also opening up Android development to a new type of user: a non-technical creator. The news also represents an expansion of Google’s earlier addition ofAI-powered coding with Geminiin itsdesktop versionof Android Studio.The apps are built with theKotlinprogramming language using Google’s Jetpack Compose toolkit and with support integration with hardware sensors like GPS, Bluetooth, and NFC, the company says. However, the resulting creations, for now, are only meant to be used personally, as publishing for family and friends is still on the roadmap.The company suggests the technology could be used for the creation of personal utilities and simple social apps, hardware-enabled experiences, or AI-powered experiences.For now, would-be app developers can use the embedded Android Emulator directly in a web browser to preview and interact with the app as it’s being built. Users can then install the app on their Android phone over a USB cable connected to their computer, using the integrated Android Debug Bridge (adb).For those looking to take their project further,AI Studiocan automatically create the app record, package the bundle, and upload it to an internal testing track inGoogle Play Consolefor developers. This allows users to continue to iterate on their app while updating on their devices along the way.Those who want to take the next steps to publishing the app more publicly can hand off this version of the project to Android Studio by downloading a zip file and exporting it directly to GitHub. In time, Google plans to allow creators to publish their apps for use by family and friends and will add support for Firebase integrations (Firestore, Firebase Auth, Firebase App Check, and other tooling).In doing so, the company is imagining an Android app ecosystem where users find apps from among their own network of friends, not just the Play Store.However, for the latter, Google is infusing AI into the experience here, too.A new “Ask Play” AI-powered overlay will allow users to discover new apps by having natural conversations with AI within the Play Store.Image Credits:Google (screenshot)Perhaps more importantly, apps will begin to be surfaced with users’ conversations with Google’s Gemini virtual assistant, exposing developers’ apps to millions of users. This will roll out in the weeks ahead across Gemini on the web and on Android. Later this year, Gemini will also surface over 450,000 movies and TV shows, plus where to livestream sports, which can directly link users from their queries to a developer’s Android app with the content in question.Image Credits:GoogleWhile Google previewed a number ofAndroid-related announcements last week, it held back on sharing the news of the native Android app development until Tuesday’s start of its annual developer conference, Google I/O. That suggests the company believes this is bigger news and more closely tied to its idea of putting AI to real-world use, as was the larger theme of this year’s event, where AI was spread across Google products, from workspace productivity apps to AI tools, search, mobile apps, and more.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Android,Apps,coding,Google,google io,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sarah PerezConsumer News EditorSarah has worked as a reporter for TechCrunch since August 2011. She joined the company after having previously spent over three years at ReadWriteWeb. Prior to her work as a reporter, Sarah worked in I.T. across a number of industries, including banking, retail and software.You can contact or verify outreach from Sarah by [email protected] via encrypted message at sarahperez.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s

Google launched on Tuesday Gemini 3.5 Flash, a new AI model that the company says is its strongest yet for coding and autonomous AI agents. The model, which was introduced at the company’s annual Google I/O developer conference, can independently execute coding pipelines, manage research projects, and, in internal tests, build an operating system entirely from scratch.The release signals Google’s shift from pitching AI as a conversational tool to AI as an agentic tool. It’s not just answering questions, but planning, building, and iterating on real work with minimal human input.Koray Kavukcuoglu, DeepMind’s chief technologist, told reporters on Monday ahead of the public launch: “3.5 Flash offers an incredible combination of quality and low latency. It outperforms our latest frontier model, 3.1 Pro, on nearly all the benchmarks,” including coding, agentic tasks, and multimodal reasoning.He added that it is 4x faster than other frontier models, a speed that’s ideal for coding and agentic tasks, but that Google has “taken it to another level” by developing an optimized version of Flash that’s 12x faster with the same quality.That speed is central to Flash’s design for agentic work, where multiple AI agents run at the same time on long-running tasks, according to Kavukcuoglu. Onstage at I/O, Google engineer Varun Mohan, demonstrated agents spawning off to work on separate components before coming together to build a full operating system inside Antigravity, the company’s agentic development platform and IDE.Image Credits:GoogleKavukcuoglu said Flash 3.5 was co-developed with Antigravity so that agents could have a “native environment where they can live, work, and execute.”At I/O, Google released Antigravity 2.0, a stand-alone desktop application designed around agent-first development.The gains are showing up beyond demos. Google says 3.5 Flash’s agentic capabilities are already creating impact among partners, like banks and fintechs automating multi-week workflows, or data science teams finding insights in complex data environments.The model can run autonomously for multiple hours, though Tulsee Doshi, Google’s senior director and head of product, said it will at times pause and ask for user input when it hits a decision point or permission issue that requires human judgment.When Google releases its forthcoming 3.5 Pro model, the two are designed to work in tandem.Doshi told TechCrunch that “3.5 Pro becomes your orchestrator, your planner, and then it actually can leverage Flash to be the various sub-agents. I think it really comes down to where do you really want that reasoning power, where you actually want that larger model that can really push on the reasoning side versus where do you have tasks that really do merit good brute force tool use capabilities?”Now, 3.5 Flash is the default model in the Gemini app and in AI Mode in Search globally. At I/O, Google also announced agentic capabilitiescoming to Search, letting users create, customize, and manage AI agents directly on the platform. The new model will also powerGemini Spark, Google’s new personal AI agent designed to run 24/7 to help consumers manage their digital life.Providing that level of AI capability for average consumers comes with scrutiny. Google is currently facing a lawsuit after a man nearly committed a mass casualty event and died by suicide following weeks of chatting with Gemini last year.The implications for harm only grow when making powerful autonomous agents available more broadly. Google says Gemini 3.5 has strengthened cyber and CBRN (chemical, biological, radiological, and nuclear) safeguards and is better calibrated to engage with sensitive questions rather than refuse them outright.Gemini 3.5 Flash is available generally today via Antigravity, the Gemini API, and Gemini Enterprise, as well as through the Gemini app and AI mode in Search.TopicsAI,gemini,gemini flash 3.5,Google,google io 2026When you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Rebecca BellanSenior ReporterRebecca Bellan is a senior reporter at TechCrunch where she covers the business, policy, and emerging trends shaping artificial intelligence. Her work has also appeared in Forbes, Bloomberg, The Atlantic, The Daily Beast, and other publications.You can contact or verify outreach from Rebecca by [email protected] via encrypted message at rebeccabellan.491 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank account

We’ve all pulled up Street View on Google Maps to show a friend what our childhood home looked like, or dropped that little person icon onto the streets of Paris to see if we booked a hotel in a cool neighborhood. Imagine being able to do that, but in a more immersive, interactive way that allows you to really simulate the street and its environs, and even do things like adjust the weather or see what it would look like in a “Day After Tomorrow” scenario.That’s one of the goals of Google’s latest integration. Starting today, Google DeepMind is connecting Street View toProject Genie,the company’s general-purpose world model that can generate diverse, interactive environments. The new feature launched during theGoogle I/O 2026developer conference.“It’s really powerful for both the agent [and robotics] use case and for humans to play with, and that’s always been the thesis of Genie,” Jack Parker-Holder, a research scientist on DeepMind’s open-endedness team, told TechCrunch.He gave the example of a new robot being deployed in London, which rarely sees the sun. Genie could, Parker-Holder says, simulate those scarce occasions when the sun glints off the Victorian housing, so the rays don’t shock the robot when it happens.Loading the player…“Simultaneously, you might say, ‘I’m going to New York City, but not this time of year,’” he continued. “‘It’s going to be snowy. I want to see what that block looks like in the snow.’”Google has been collecting Street View data for 20 years via cars with cameras and individuals strapped with “tracker backpacks.” The tech giant has collected north of 280 billion images across 110 countries and seven continents.“With Street View, we have imagery from a large quantity of the world,” Jack said. “You can imagine how potentially powerful it is to combine this rich source of real-world information and data with an ability to simulate worlds.”Google released its latest world modelGenie 3 for research previewlast August and opened up access to the tool to Google AI Ultra subscribers in the U.S. in January, allowing customers to create interactive game worlds from text prompts or images. The goal is to use Genie for educational experiences, gaming, and robotics training.Genie 3 is already helping to powerone of Waymo’s simulatorsto train its self-driving cars on “exceedingly rare events” like tornadoes or casual elephant encounters. Adding Street View data to that could help Waymo prepare to launch in more cities around the globe.Waymo has its own simulator that it relied on to scale to 11 U.S. cities and test its AI driver in several more. The difference with Genie, says Parker-Holder, is that those are all from the car’s point of view. Street View allows for not only simulating a world anchored to a real place, but also shifting the point of view to other types of agents, like a human or a robot.Google is launching Street View in Genie to some Ultra users in the United States starting today, with access rolling out at scale over time. Global Ultra users will gain access over the next few weeks, per the company.The researchers’ goal is to put this new capability into as many hands as possible, per Diego Rivas, a product manager at DeepMind. He cautioned that Street View in particular and Genie in general is still an experiment, so there’s much to improve upon in terms of accuracy.In the samples the Google team showed me — including an underwater simulation of a neighborhood I used to live in — the results are impressive and recognizable, but still video game quality rather than photorealistic. The models are also not yet physics-aware, meaning they don’t yet understand cause and effect. For example, in a simulation of a woman running through a snowy Joshua Tree, she ran right through cacti and bushes.Compare that to, say, Google’s image generator Nano Banana — which can now generate perfect text in infographics — or its video generator Veo — which understands that paper boats drift on water currents, smoke disperses into the air, and fabric drapes over forms.Physics isn’t hard-coded into these models; they learn it intuitively over time through passive observation, as a living being would.“I think for this kind of model, it’s maybe six to 12 months behind video in terms of the accuracy and quality, so I think it’s something we will solve,” Parker-Holder said.Jonathan Herbert, director of Google Maps who started on the Street View team as an intern 12 years ago, said that Genie can’t yet create a faithful reconstruction of a street. He thinks the real breakthrough is the AI’s spatial continuity. If you turn 360 degrees, the AI correctly remembers and simulates the environment behind you. From that point on, the model can build a new environment on top of that.“We have long thought about how we can build out the best and richest model of the world on top of Street View data,” Herbert said. “It’s definitely been an idea of ours to use Maps Data in new ways and for new kinds of AI researc

Google is getting (back) into the smart glasses game.At Google I/O on Tuesday, the company announced a new partnership with Warby Parker and Gentle Monster to produce a new line of AI-powered glasses. The company says that the devices will be built to pair with Android and iOS devices and were designed in collaboration with Samsung. They will be available later this year, the company said.Google is calling the new devices “audio glasses,” in that users will be able to issue verbal commands to them and get things done via its ecosystem of apps and services, including Gemini.Image Credits:TechCrunchThe user simply talks to their glasses (the demo shared on Tuesday involved a Googler ordering a coffee online by merely talking to the glasses), and the device, when synced, complies.Google has dabbled in smart glasses a number of times over the years. It notoriously launchedGoogle Glassyears ago, which ultimately helped spawn the derogatory term “glassholes.”The smart glasses space has changed a little bit since then, however. Lately, major companies — most notably Meta — and a small army of startups and smaller firms, have invested in the space.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,ai glasses,glasses,Google,google io,google io 2026,HardwareWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lucas RopekSenior Writer, TechCrunchLucas is a senior writer at TechCrunch, where he covers artificial intelligence, consumer tech, and startups. He previously covered AI and cybersecurity at Gizmodo.

The jury’s speedydecision to rejectElon Musk’s lawsuit against the other founders of OpenAI and Microsoft confirmed what we saw in the courtroom: Musk’s case was a weak one, in part because he waited so long to file it.Watching the closing arguments last week, OpenAI’s attorneys detailed point by point how the law was on their client’s side, while the plaintiff’s team focused on Sam Altman’s apparentlack of credibilityand expressed disbelief that anyone would disagree with Musk’s accusations.The final effect was that, after the verdict, some found it hard to believe Musk had lost — including the man himself. In a post he later deleted, Musk called Judge Yvonne Gonzalez Rogers a “terrible activist Oakland judge,” then announced his plans to appeal, declaring, “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity.”But Altman and Brockman weren’t the only figures who benefited from OpenAI’s non-profit investments. As much as Musk and his legal team tried to make the trial about Altman, the proceedings revealed just as much about Musk.One incident that came out in court showed Musk benefiting from OpenAI in an uncomfortably familiar way. Greg Brockman testified that in 2017, Musk asked him to bring a team of OpenAI researchers down to Tesla’s headquarters to help with the autopilot team for a few weeks. “It was pretty clear that was not something we could say no to,” Brockman said.Brockman described taking a team of leading scientists, including Andrej Karpathy, Ilya Sutskever, and Scott Gray, to consult with the “demoralized” Tesla workers. They helped come up with ideas to improve the vehicle’s self-driving technology, with Sutskever telling the team that if they could find 10,000 images of a tricky corner case, they would be able to fix their software. Musk even asked Brockman to recommend employees to fire, which he declined to do.Another person familiar with the episode confirmed Brockman’s account and said Tesla did not reimburse OpenAI for the time and effort of its employees. Musk’s family office, Excession, didn’t reply to a request for comment.The heart ofMusk’s caseis that Altman, Brockman, and OpenAI committed a “breach of charitable trust” — that Musk donated funds for a specific charitable purpose, and his co-founders instead used them for something else. He also accused them of “unjust enrichment” due to stock and other benefits from OpenAI’s for-profit.In the case of the OpenAI scientists parachuting into Tesla, Musk’s charitable donations were intended to hire scientists focused on securing the benefits of AGI. Instead, he had them work for free at his for-profit company.Dorothy Lund, a Columbia Law School professor and the co-host of theBeyond Unprecedented podcast, told TechCrunch that this arrangement wouldn’t be legal, calling it “a bit rich for Musk to be suing for breach of a charitable trust, when he appears to have been redirecting assets in a way that was inconsistent with that mission.”It’s true that the self-driving work involved artificial intelligence, but witnesses for Musk emphasized that Tesla’s self-driving project was very different from OpenAI’s research agenda. That’s in part because Karpathy left OpenAI for Tesla shortly after this incident. OpenAI’s attorneys portrayed the departure as Musk violating his duty to the lab, where he was co-chair of the board, by recruiting one of its key researchers to his own company.The other fact that no doubt influenced the jury was the amount of time Muskspent tryingto gain sole control of a potential OpenAI for-profit affiliate in 2017. Musk deployed good cop, bad cop tactics in an attempt to convince his co-founders to let him have total control of OpenAI’s for-profit affiliate — giving them free Teslas and threatening to withhold his donations.His efforts put his attorneys in a tricky spot, facing a need to convince the jury there was a significant difference between what Musk envisioned and the for-profit that was ultimately created. They suggested a “small adjunct” for-profit would be permissible, though OpenAI’s witnesses showed non-profits with large commercial arms are common.Indeed, there’s a very plausible counter-factual where Musk took one of the offers his co-founders made to split their equity more evenly, and finds himself today as one of OpenAI’s largest shareholders — just not the controlling one. But several times during the trial, Musk’s associates testified that he refuses to invest in any business he could have sole control over.The failure of Musk’s claims because he filed them too late has been cited as a technicality, but the statute of limitations has substance behind it: People and businesses make important decisions and spend resources based on their understanding that what they are doing is permissible. If someone like Musk waits too long to sue, then the cost of unraveling all those decisions can outweigh a just reimbursement.No

Shay Shwartz knows a lot about email phishing attacks. As a teenager, he made money as a hacker, but after getting caught at age 16, he realized he could use his cyber talents to prevent attacks rather than launch them.He went on to spend about a decade in top-tier cybersecurity roles, leading major projects for Israel’s elite defense and intelligence units, including work connected to the Iron Dome project, before joining Axis, the startup later acquired by HPE.All along, he had been itching to launch his own startup, and two years ago, he finally took the plunge.His startup Ocean, an agentic email security platform built to fight AI-powered attacks, just emerged from stealth mode with $28 million in total funding. The round was led by Lightspeed Venture Partners, with participation from Picture Capital and Cerca Partners. High-profile angel investors also joined the round, including Wiz co-founder and CEO Assaf Rappaport, as well as Yevgeny Dibrov and Nadir Izrael, the co-founders of Armis, which recently sold to ServiceNow for $7.75 billion.While established vendors like Proofpoint and Mimecast, along with newer players like Abnormal Security, help detect standard phishing attacks, Shwartz (pictured right next to co-founder and CTO Oran Moyal) argues that AI requires a different defensive approach.In the past, only highly sophisticated hackers could pull off spear-phishing due to the sheer amount of time, research, and manual labor needed to launch targeted attacks.“AI just made the entire process automatic, so the scale is much, much bigger now,” Shwartz told TechCrunch. “I can instruct LLM to go and understand exactly who you are, harvest large amount of public information, and create those phishing attacks very targeted against you.”Ocean claims its AI can thoroughly analyze the context of every incoming email to detect fraud and impersonation attempts.The startup is already reviewing billions of emails each month for customers, including Kayak, Kingston Technology, and Headspace.Shwartz said Ocean built a small language model tailored to quickly analyze emails, understand the sender’s intent, and evaluate it against the user’s specific organizational context.“This is like having a guard in every door,” Shwartz said. “This is how we make the inbox a safe place with high hygiene.”TopicsAI,Exclusive,lightspeed,phishing,Security,StartupsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Three-year-old Mach Industries has acquired solid rocket motor startup Exquadrum in a $50 million cash-and-equity deal, the Huntington Beach-based defense startup tells TechCrunch. Exquadrum — now rebranded as Mach Energetics — has been fully folded into Mach’s operations, giving it direct control over one of the most important, and constrained, components in modern unmanned systems.The deal itself started with a lucky break. The two companies first connected last September when an Exquadrum customer at an MIT recruiting event happened to overhear a Mach recruiter mention that the company was in the market for a solid rocket motor supplier. Introductions were made, Mach came in as a customer, and now, roughly five months later, it has acquired the company outright, beating out upward of eight other potential buyers in the process, it says.“The Exquadrum acquisition marks an important next stage in Mach’s growth,” said founder and CEO Ethan Thornton, who dropped out of MIT at 19 to start the company. “As we deliver vehicles to the warfighter, we’ll continue to vertically integrate our supply chain across solid rocket motors, engines, radar, and avionics to ensure we deliver the best possible product at the lowest cost. In many areas of the defense industrial base, these components are not only too expensive or lacking performance, they’re simply unavailable, with lead times stretching years. In short, vertical integration is non-optional.”That supply problem is real and getting more acute. Decades of consolidation have left the domestic solid rocket motor market effectively controlled by two large primes — Aerojet Rocketdyne and Northrop Grumman — with little independent capacity to absorb the growing demand that modern drone warfare has created.Indeed, in February, the Pentagon awarded the defense tech outfit Anduril $43.7 million specifically to expand domestic SRM production (its second such investment in the company in just over a year), explicitly calling SRMs a critical bottleneck in the munitions supply chain.Mach is now explicitly positioning itself as part of the solution, and not just for its own programs but for the broader ecosystem. Mach Energetics plans to sell components, testing services, and subsystems to other defense firms, a maneuver that suggests Mach sees itself as potential infrastructure for the defense tech industry and not just a systems builder.According to Mach, all 85 Exquadrum employees are coming over as part of the deal, along with the company’s IP, business lines, and its 70,000-square-foot facility in Victorville, California, which is anchored by a nearby energetics and rocket propulsion test site. The combined company now has roughly 350 employees. Exquadrum co-founders Kevin Mahaffy and Eric Schmidt (no relation to the former Google CEO) are both taking on leadership roles within Mach Energetics and the broader organization.The acquisition also mirrors moves by other ambitious defense tech startups that are focused on owning the stack and using cost and speed as competitive weapons. Mach has five vehicle programs in various stages of development — Viper, a jet-powered VTOL; Glide, a high-altitude strike glider; Stratos, an airborne surveillance platform; Dart, its low-cost counter-drone interceptor; and Pike, a long-range strike munition built for large-scale deployment — with plans to enter production on at least three this year. The company says the acquisition meaningfully improves unit economics across all of them at exactly the moment it is starting to scale.Mach has raised nearly $200 million in total — most recently a$100 millionSeries B last June led by Bedrock Capital, Khosla Ventures, and Sequoia Capital — at a valuation of $470 million. It’s a multiple that right now looks modest for a company with this trajectory and will definitely be worth watching as the rubber starts to meet the road this year.Topicsdefense tech,Ethan Thornton,Hardware,mach industriesWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Connie LoizosEditor in Chief & General ManagerLoizos has been reporting on Silicon Valley since the late ’90s, when she joined the original Red Herring magazine. Previously the Silicon Valley Editor of TechCrunch, she was named Editor in Chief and General Manager of TechCrunch in September 2023. She’s also the founder of StrictlyVC, a daily e-newsletter and lecture series acquired by Yahoo in August 2023 and now operated as a sub brand of TechCrunch.You can contact or verify outreach from Connie by [email protected]@techcrunch.com, or via encrypted message at ConnieLoizos.53 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk h

In BriefPosted:2:14 PM PDT · May 19, 2026Image Credits:MARTIN BUREAU/AFP via Getty ImagesZack WhittakerDiscord enables end-to-end encrypted voice and video calling for every userMessaging platform giant Discord has switched on end-to-end encrypted voice and video messaging for every user. Theend-to-end encryptionfeature means that Discord users can now communicate privately without anyone else being able to listen in to their calls, not even Discord.This is a major privacy win for the community’s hundreds of millions of users, soon after social media giant Metapulled the plugon Instagram’s end-to-end encrypted messaging feature earlier this year. TikTok said it would alsonot end-to-end encrypt user messagesafter it became a U.S. company.Discordlaunched the end-to-end encrypted voice and video calling feature in 2024, and on Monday rolled out the feature to every user, no action needed.“End-to-end Encryption is now standard for every voice and video call on Discord, outside of stage channels. No opt-in required,” said Mark Smith, Discord’s vice president of core technologies, in ablog post.Topicscybersecurity,discord,end-to-end encryption,SecurityMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedSecurityUS cyber agency CISA exposed reams of passwords and cloud keys to the open webZack Whittaker9 hours agoSecurityNYC Health + Hospitals says hackers stole medical data and fingerprints during breach affecting at least 1.8 million peopleZack Whittaker1 day agoSecurityOpen source tool maker Grafana Labs says hackers stole its code, refuses to pay ransomZack Whittaker1 day agoLatest in SecurityIn BriefDiscord enables end-to-end encrypted voice and video calling for every userZack Whittaker3 hours agoSecurityFrom teen hacker to Iron Dome researcher, this founder raised $28M to fight AI phishingMarina Temkin3 hours agoIn BriefHackers have compromised dozens of popular open source packages in an ongoing supply-chain attackLorenzo Franceschi-Bicchierai9 hours ago

At theGoogle I/O 2026keynote, the tech giantrevealed new agentic capabilities in Search, where users can create, customize, and manage multiple AI agents to stay updated on topics of interest. The announcement is part of Google’s larger push toward agentic AI systems that can take initiative and assist with ongoing tasks instead of answering one question at a time.Unlike traditional search tools that respond only when prompted, Google’s information agents are designed to operate continuously in the background, 24/7, helping users stay informed about their interests without needing to repeatedly search for the same information every day.Instead of delivering a list of links, the agents can synthesize information from multiple sources, explain why something matters, compare perspectives, and provide actionable insights. In many ways, the agents represent the next evolution of Google Alerts, the notification service Google launched in 2003. However, these agents are designed to go beyond simple notifications.For instance, someone following the stock market could create an information agent focused on specific companies, share price, or economic trends. The agent could monitor market activity throughout the day, track breaking news, summarize earnings reports, alert users when major changes happen, and provide summaries and links to learn more.Image Credits:GoogleIt could also help with everyday tasks, such as tracking flight prices for upcoming trips, monitoring sports teams and live events, following breaking news, keeping tabs on housing or job market trends, and tracking weather or traffic.To use the feature, users can open AI Mode in Search and enter a prompt.For example: “Keep me updated on nearby movie tickets for ‘The Mandalorian and Grogu.’” When something relevant appears, the Google app sends a push notification. You’ll also see your active tracked topics in your AI Mode history, where you can jump back in to manage, refine, or turn off an alert.Information agents will be available this summer. The company is first rolling them out to Google AI Pro and Ultra subscribers in the U.S., then to additional markets later on.In addition to these information agents, Google also introduced amajor redesign of Search, including what it describes as a reimagined “intelligent search box,” the company’s biggest change to Search in more than 25 years. The new interface is designed to support longer, more conversational queries. There’s also a new AI-powered query suggestion system that goes beyond traditional autocomplete, helping users craft nuanced and context-aware searches.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,AI agents,Google,google io 2026,Google SearchWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Google isn’t finished infusing AI into your inbox. On Tuesday attheir IO 2026 developer conference, the tech giant announced an expansion of its“AI Inbox” functionality for Gmail, which is adding conversational AI features. That means you can ask Gmail about things in your inbox instead of typing in search terms.The company says the Gemini AI-powered feature, called Gmail Live, will help you quickly find information buried in your inbox.Image Credits:GooglePerhaps you need information about your upcoming flight, the time of your dentist appointment, the door code for your Airbnb rental, or some details about an event at your kid’s school, for instance.Before, you’d have to type in keywords in the search box (or maybe type in someone’s email address or domain) to try to narrow down your search. That doesn’t always make emails easy to find, however, especially if the search term is something found across several messages.“Gmail Live can answer naturally phrased questions, respond to follow-up questions, and pivot if you need to interrupt it,” Devanshi Bhandari, product lead for Gmail, explained in a briefing ahead of Google’s annual developer conference, Google I/O, where the feature was first introduced to the public.It’s another way that Google is trying to showcase how its AI technology can drive real-world improvements to products used by millions of consumers, at a time when many outside the tech industry are questioning the value of AI, as new data centers get built in their backyards, driving up their power bills.Being able to point to something as simple as making it easier to find something that’s lost in your email inbox — an experience nearly everyone has suffered at some point — could be a practical and positive use case for AI … or at least, Google hopes.Bhandari demonstrated Gmail Live to reporters, asking the tool a series of questions about things in the inbox, like a child’s show-and-tell project and their class trip, plus hotel and flight information for a trip to Detroit. Similar to using a stand-alone AI chatbot like Gemini or ChatGPT, Gmail users can ask these questions aloud in natural language, and the chatbot responds.In the demo, Gmail Live also understood nuances between things like “field trip” and “trip” and was able to jump from one topic to another, Bhandari pointed out. Plus, the AI can pull granular details from emails, like a hotel room number, or infer which people you’re asking about, even when they’re not explicitly named.Similar voice technology is also coming to its to-do list, Google Keep, the company noted.Notably, Gmail Live is not replacing traditional Gmail search — it’s just another option.Google may have learned that not everyone is ready for an AI-only experience after it “upgraded”Google Photos with AI-powered searchto much backlash.Google Photos later rolled back the feature,making the use of AI optional after numerous complaints.Gmail is also gaining other new capabilities, including ready-to-send drafts, instant file access, and the ability to manage to-dos by marking individual tasks as done.Image Credits:GooglePlus, theAI Inbox experience, which launched earlier this year, will expand beyond Google AI Ultra subscribers to reach Google AI Pro and Plus subscribers as well. This allows you to see an overview of the tasks and items to catch up on that are buried in your inbox, all on one page.The voice-powered Gmail Live feature, however, will roll out later this summer and will initially be limited toGoogle AI Ultra subscribers.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,gemini ai,gmail,Google,google io,google io 2026,Google WorkspaceWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sarah PerezConsumer News EditorSarah has worked as a reporter for TechCrunch since August 2011. She joined the company after having previously spent over three years at ReadWriteWeb. Prior to her work as a reporter, Sarah worked in I.T. across a number of industries, including banking, retail and software.You can contact or verify outreach from Sarah by [email protected] via encrypted message at sarahperez.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One t

Google announced atits annual Google I/O eventon Tuesday that it’s launching Pics, a new AI-powered design and image-generation app for Google Workspace. The tech giant says it designed the app to be accessible to everyone, from teachers to small business owners.With Pics, users can generate everything from social media graphics and invitations to marketing materials and mock-ups using simple text prompts, without needing any editing skills or advanced tools. By giving users an easy way to generate visuals, Google is looking to take on popular design apps like Canva, as well as products from AI-native competitors likeClaude Designfrom Anthropic. Google’s entry into the space signals that AI-powered design is fast becoming a core competitive arena — with real stakes for any business that depends on visual content.The new app is launching to a group of testers at I/O and will be rolling out to Google AI Ultra subscribers this summer, Google says.The company acknowledges that although AI models today can generate high-quality images, it’s still difficult to modify just one part of an image. If you get an image that’s almost perfect but want to change a small detail, you have to write an entirely new prompt and hope the AI doesn’t alter too much. That’s why Pics not only generates images but also makes them easily editable.Loading the player…Users can enter a prompt, and Pics will generate what they need. Gemini powers the editing layer, making every element in a generated design or image fully adjustable. You can write a new prompt to make changes, but you can also simply click the part you want to change and leave a comment — much like leaving feedback in Google Docs.You can also edit directly, without leaving a comment or writing a prompt. For example, if you create a birthday party invitation and want to change the time listed on the card, you can do so manually.Pics is powered by Nano Banana 2, which Google says is a strong fit for the app because it supports precise text rendering, real-world knowledge, and detailed visual output. Pics is also built natively into Google Workspace, enabling visual collaboration across its apps.Once you’re happy with your design, you can download, copy, print, or share it with others. You can also pass it to someone else for a final round of edits before it goes out, Google says.Catch up on the rest of Google IO 2026’s big newsGoogle Search as you know it is overGoogle updates Gemini app to take on ChatGPT and ClaudeGoogle introduces Gemini Spark, a 24/7 agent assistant with Gmail integrationHow to use Google’s new information agentsTopicsAI,Apps,Google,Google I/O,google io 2026,google picsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Aisha MalikConsumer News ReporterAisha is a consumer news reporter at TechCrunch. Prior to joining the publication in 2021, she was a telecom reporter at MobileSyrup. Aisha holds an honours bachelor’s degree from University of Toronto and a master’s degree in journalism from Western University.You can contact or verify outreach from Aisha by [email protected] via encrypted message at aisha_malik.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularGoogle Search as you know it is overSarah PerezElon Musk has lost his lawsuit against Sam Altman and OpenAITim FernholzUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha Malik

Nice cable you got thereIran demands Big Tech pay fees for undersea Internet cables in Strait of HormuzIran’s claim over subsea chokepoint pushes US tech companies to overland fiber.Jeremy Hsu–May 19, 2026 7:00 am|174Multicolored lines show undersea Internet cable routes running through the Strait of Hormuz.Credit:TeleGeographyMulticolored lines show undersea Internet cable routes running through the Strait of Hormuz.Credit:TeleGeographyText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navIran claims it will charge US tech companies fees for using undersea Internet cables that run beneath the contested Strait of Hormuz shipping lanes. The war has already halted multiple projects and led to the suspension of cable repairs in the region—and the latest Iranian threats may accelerate efforts by Big Tech and Gulf countries to find alternative routes for bypassing the Strait of Hormuz’s digital chokepoint.The latest assertions of Iranian authority over the Strait of Hormuz were announced in a brief statement by Ebrahim Zolfaghari, a spokesperson for Iran’s military and the Islamic Revolutionary Guard Corps. “We will impose fees on internet cables” Zolfaghari wrote in aMay 9 post. It was not immediately clear how Iran might implement such fees or impose its rules on cable projects, given that the majority of routes pass through Oman-controlled waters.But Tasnim and Fars, both Iranian state-linked media channels, laid out more detailed proposals on how Iran could charge license fees to US tech giants for the use and maintenance of undersea cables carrying regional Internet traffic, according toThe Guardian. For example, the Tasnim plan described charging tech companies—specifically naming Meta, Google, Amazon, and Microsoft—license fees for cable usage while also claiming that Iran alone has the right to repair and maintain the subsea cables.More than 99 percent of international Internet traffic runs through the global network of undersea cables that crisscross various oceans, connecting continents and islands. The major active cables running through the Strait of Hormuz primarily serve the Gulf countries in the region. They include theAsia Africa Europe-1,FALCON, and theGulf Bridge International Cable System, according toTeleGeography, a telecommunications research organization.The FALCON and Gulf Bridge cables run through Iranian territorial waters at certain points, Alan Mauldin, a research director for TeleGeography, said in aCNNinterview. CNN also reported that Iranian state media outlets have “issued veiled threats warning of damage to cables.”By comparison, there is “not much risk to Europe-Asia data traffic from hostilities in the Strait of Hormuz” because such Internet traffic primarily travels through cables in the Red Sea, TeleGeography wrote in a blog post. But undersea cables in the Red Sea have already seen a spate of damage in recent years, exacerbated by lengthy repair times and attacks by Houthi rebels allied with Iran.The biggest threat to the digital chokepointIran’s capability to threaten undersea cables as a way of imposing its will is uncertain at this point. Since the US and Israel started the war by attacking Iran on February 28, the US military claims to have destroyed 161 vessels of the Iranian navy, according to Admiral Brad Cooper, commander of US Central Command, intestimonygiven to the Senate Armed Services Committee on May 14. The US military has also sunk some of thesmall fast boatsoperated by Iran’s paramilitary Islamic Revolutionary Guard Corps when the latter threatened civilian shipping in and around the Strait of Hormuz.Thevast majority of damageto undersea cables comes from commercial ships accidentally dragging their anchors or fishing trawlers dragging weighted nets along the seafloor. So it’s not out of the realm of possibility for a more innocuous-looking Iranian ship to sneakily perform somesubsea cable sabotageif it’s willing to run the gauntlet of US military surveillance and patrols in the strait.Even a damaged commercial ship abandoned in the strait could end up dragging its anchor across some cables, as was the case with a drifting ship thatdamaged three cablesfollowing an attack by Houthi rebels in the Red Sea in 2024.However, the greatest threat to subsea cable infrastructure in the Strait of Hormuz may simply come from delays in any necessary cable repairs in the region. Such jobs require specialized ships to find the damaged area and lower grappling hooks to lift up the cable for inspection and repair, according toBBC News. That repair process can require days or sometimes weeks, which would leave the ship vulnerable to Iranian missiles, drones, or fast boats that havecontinued to attackcommercial shipping in and around the strait.“Operators face a choice: pay protection fees and accept Iranian licensing over Middle East Gulf seabed activity, or accept that future faults may go unrepai

Catching the satellite busIn addition to space stations, Vast says it will now build high-power satellites“Every single successful space company is diversified in its products.”Eric Berger–May 19, 2026 9:00 am|25A rendering of Vast’s 15 kW satellite.Credit:

      Vast SpaceA rendering of Vast's 15 kW satellite.Credit:

      Vast SpaceText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAs part of its plan to develop a private space station, Vast Space built and then launched a small demonstration spacecraft in early November. This vehicle then completed dozens of test objectives with flying colors before making a successful de-orbit three months later.The mission, which tested power, propulsion, tracking, and a multitude of other technologies needed for Vast’s Haven-1 space station, was evidently so successful that the company is ready to use its spaceflight capabilities for other purposes. The Long Beach, California-based company announced Tuesday that it plans to begin selling high-powered satellite buses.“Every single successful space company is diversified in its products,” said Max Haot, chief executive of Vast Space, in an interview. “So for us it really was a question of when, not if.”A satellite for power-hungry applicationsThe company’s first offering is a 15 kW-class satellite bus capable of supporting various demanding missions. Each satellite is about 3 meters long and 4 meters tall, with a mass of 700 kg and payload capacity of at least 350 kg. They will have a design lifetime of five years and be intended to operate in areas ranging from low-Earth orbit out to lunar orbit. Vast aims to serve various customers, from telecommunications to observation to data services. Haot added that Vast also plans to offer an Nvidia Space-1 Vera Rubin Module to support orbital data center inferencing needs.The Vast satellite bus—essentially a backbone providing power, propulsion, and navigation for various payloads—will be based largely on technology ported over from the company’s Haven-1 space station, which is due to launch for the first time next year as the world’s first private space station. However, there will be some new elements needed for the satellite, and Haot said Vast is already moving forward with in-house development of electric propulsion and a deployable solar array for the satellite.Vast has already signed a customer for four satellites, with an option to purchase up to 200 additional satellites. Haot said the company is targeting a launch of at least 10 Vast Satellites in the fourth quarter of 2027.With this new product line, Vast is entering an increasingly crowded market. Historically, in the United States, a handful of large companies such as Boeing, Lockheed Martin, Northrop Grumman, Maxar, and Sierra Space have manufactured medium and large satellites. Typically, these were costly and often bespoke designs that cost tens to often hundreds of millions of dollars.But a few trends have changed the landscape in recent years. The US government’s Space Development Agency has signaled that it prefers proliferated constellations—many satellites spread out present less of a concentrated target than a few larger and more expensive satellites. With the Falcon 9 rocket’s increased cadence, as well as rideshare missions, it became easier and sometimes cheaper to get smaller and medium-sized satellites into orbit.This has led to an influx of venture capital to back a new generation of companies seeking to build less expensive, more modular satellites that could fill a variety of purposes. There are several prominent, relatively new entrants in this area, including K2 Space, Rocket Lab, True Anomaly, Blue Canyon, and Millennium Space Systems.Vast already has… vast facilitiesHaot said most of these companies are still emerging, with products that are not yet mature. In other words, he believes that if Vast Space can execute, it could become a market leader, especially in power-hungry applications. Vast has already invested $1 billion in facilities for spacecraft manufacturing, including clean rooms, which can be used for space stations as well as satellites, he said.The number of satellites in space has exploded in recent years, largely due to the rapid expansion of SpaceX’s Starlink constellation. For decades, the total number of satellites orbiting Earth numbered about 4,000, but in the last five years that number has grown to about 14,000.This is just the beginning. By some estimates, in another decade, there will be approximately 500,000 satellites in orbit for communications, Earth observation, orbital data centers, and other applications.Haot expects that about 90 percent of these will be satellites built by SpaceX, Amazon, Blue Origin, or other major players. But even 10 percent of that number, if available to commercial satellite bus manufacturers, would represent 50,000 satellites f

ConsolidationElectrical utility megamerger is all about the data centersNextEra’s blockbuster deal with Dominion likely means higher bills for consumers.Dan Gearino, Amy Green, and Charles Paullin, Inside Climate News–May 19, 2026 9:45 am|57The campus of NextEra’s headquarters is seen on Monday in Juno Beach, Fla.Credit:

      Marco Bello/Getty ImagesThe campus of NextEra’s headquarters is seen on Monday in Juno Beach, Fla.Credit:

      Marco Bello/Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA proposed merger of the largest utility in the country by market value, NextEra Energy, with the sixth-largest, Dominion, would create a megacompany at a time when data centers and rapid increases in electricity demand are reshaping the industry.The proposal,announced Monday morningand contingent on state and federal regulatory approval, would result in a company that leads in nearly every aspect of the US power and utility industry, including overall electricity generation, natural gas generation, and renewables.The $67 billion deal combines NextEra’s size and reach with Dominion’s positioning as the local utility for the world’s largest concentration of data centers innorthern Virginia. But the results are likely bad for consumers and the environment, creating a company with enormous financial and political strength that will be difficult to effectively regulate, according to consumer advocates and analysts.For perspective, only Exxon Mobil and Chevron would be larger based on market value among US-based energy companies.“Mergers are not about consumers; they’re about shareholders,” said Ari Peskoe, director of the Electricity Law Initiative at Harvard Law School. “For the Dominion shareholders, they are selling their shares at a premium. The executives are getting massive payouts for facilitating this, assuming it all goes through, and obviously NextEra believes the transaction is going to add value to the company. Ratepayers are all an afterthought.”The deal makes financial sense for both companies, said Andrew Bischof, an equity analyst for Morningstar.“We view the transaction as allowing NextEra to accelerate its data center ambitions, which had trailed those of its regulated peers, by using Dominion’s expertise and relationships to expedite NextEra’s data center hub plans,” he said in a note to clients.NextEra, based in Juno Beach, Florida, includes Florida Power & Light, the largest regulated electricity utility in the state, and NextEra Energy Resources, a wholesale electricity supplier that owns power plants across the nation. Dominion, based in Richmond, Virginia, includes regulated utilities serving much of Virginia, parts of North Carolina and South Carolina, and other assets across the country.The company would be called NextEra Energy, and NextEra CEO John W. Ketchum would serve in the same role after the deal closes. Robert M. Blue, Dominion’s CEO, would be the CEO for regulated utilities for the merged company. The parties said they expect regulatory approvals to take 12 to 18 months.A page from NextEra Energy and Dominion’s investor presentation shows the states where each company has regulated utilities and the merged company’s ranking in various categories.Credit:
                                  NextEra EnergyA page from NextEra Energy and Dominion’s investor presentation shows the states where each company has regulated utilities and the merged company’s ranking in various categories.Credit:

      NextEra EnergyNextEra shareholders would own 74.5 percent and Dominion shareholders would own 25.5 percent, respectively, of the combined company in the all-stock transaction.“We are bringing NextEra Energy and Dominion Energy together because scale matters more than ever—not for the sake of size, but because scale translates into capital and operating efficiencies,” Ketchum said in a statement.“Adding to the pollution problem”The post-merger NextEra would be the leader in so many categories in the US utility sector that it’s easier to list the ones where it wouldn’t be on top. It would rank second in nuclear power generating capacity and in the number of regulated utility customers, trailing Exelon Corp. of Chicago in both.NextEra and Dominion both have substantial carbon emissions, but neither was among the top five utility companies in the country in 2024, according to the most recent edition ofBenchmarking Air Emissions reportfrom the Natural Resources Defense Council. NextEra ranked sixth and Dominion ranked 11th, and their sum was less than that of each of the leaders, Vistra Energy and Duke Energy.But those are still massive emissions from a company that stands to gain more clout because of its size.“If we continue to add dangerous climate pollution into the mix, then people who are already suffering and are typically hurt first and worst will su

Test of TimeCivilization VIIfinally lets you build a civ that stands the test of timeCiv 7’s devs talk walking back the game’s most controversial decision.Samuel Axon–May 19, 2026 12:00 pm|38You can’t tell as much from the map view, but many ofCiv 7’s key underlying systems have been totally overhauled.Credit:

      Samuel AxonYou can't tell as much from the map view, but many ofCiv 7's key underlying systems have been totally overhauled.Credit:

      Samuel AxonText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to nav“Build a civilization to stand the test of time.” That was the promise on the box of Sid Meier’sCivilization, the first in a long-running strategy game franchise that has evolved over 35 years and seven mainline entries.Civ 7introduced a new approach to play wherein players would change civilizations from their initial selection twice by the end of a game. Lots of players said, “Wait a minute: we’re literally not building a civilization to stand the test of time anymore.” After such a negative reception at launch, longtime series fans began to wonder whether the franchise would continue to stand the test of time.It’s clearly not a coincidence that the new, major update for the game reaching players today is titled “Test of Time.” It’s a major reworking of several of the game’s key systems, and it reintroduces the ability to play one civ from beginning to end while retaining some of the big ideas that definedCiv 7at launch.I’ve spent the past week playing the game with the Test of Time update, and I spoke with two of the game’s senior developers—the franchise’s creative director, Ed Beach, and series head of product Matt Schembari—about the choices they made, reacting to player feedback, and competing ideas about what Civ is to players.A quick note on agesFor those who haven’t been followingCiv 7up to this point: Another major change to the game is that it has been broken into three distinct ages: Antiquity, Exploration, and Modern. Each has its own tech and culture tree, game mechanics, and, until now, civilizations to pick from.The intent behind it was to allow the game’s designers to solve some longtime problems with how traditional Civ games could lag or have balance problems at specific stages. If that in itself is too much deviation from the classic formula for you, it’s unlikely Test of Time will make you likeCiv 7any more than you might have a year ago. It’s still a foundational framework for this game. Test of Time gives players a lot more freedom inside that structure, but it does not remove the structure itself.Give the people what they wantStarting a new game ofCiv 7with the patch, the biggest change was immediately apparent. As of the last update before this one, you could pick from just 14 civs when starting a game in the Antiquity Age. This time, I saw more than 40.I could play as America from the Age of Antiquity now, for example. When I set up the game, I was also given the option to choose the AI’s civ-switching behavior; I could prohibit it, allow it, or pick an option that makes the AI follow suit on whatever choice I make.Yep, we have modern age civs as options for a game starting in antiquity.Credit:
                                  Samuel AxonYep, we have modern age civs as options for a game starting in antiquity.Credit:

      Samuel AxonOf course, playing one civ throughout the three ages would seem to run counter to how the game was originally designed. Firaxis took this update as an opportunity to take a major game balance pass across all the civs, and it introduced some new concepts to make one civ for three ages fit theCiv 7framework.Civs are considered to have an “apex age,” during which they have access to their full kit. When they’re in an age other than their apex age, they’re called “time-tested civs,” and they have access to a limited amount of their normal bonuses. But players can use a system called syncretism to adopt bonuses that are normally offered to specific civs that would call the current age their apex age.Based on playing one full game in the new patch, I feel this implementation works really well overall, and it’s a lot of fun.That said, I could see it getting unwieldy to build on over time—the game’s designers now must account for many more possibilities and balance them against each other to make sure the game stays fair and fun.When I asked Beach about this, he suggested it’s not as big a design challenge as it might seem. “I think we still have some advantages because of the structure of the game,” he said. “I think where we are giving ourselves an immense challenge is the content space, because we have 43 civs that you can play as, and now we need to give them bonuses in three different ages. So we went from basically 43 chunks of content that we need to balance to like 129 chunks of content, and as we add more civs

more taxesEV drivers will pay $130 a year under Congress’ 2026 transportation billPoliticians say they want EVs to pay “their fair share for the use of our roads.”Jonathan M. Gitlin–May 19, 2026 12:12 pm|232Credit:

      Getty ImagesCredit:

      Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe119th Congressmight be one of themost dysfunctional and least productivelegislative sessions in the 250-year history of the United States, but it seems there’s one thing it can agree on: Electric vehicles don’t cost their owners enough money. The Transportation and Infrastructure committee has published its bill to fund surface transportation for the next half-decade, and among the provisions in the “Building Unrivaled Infrastructure and Long-term Development for America’s 250th Act” is an annual fee levied against owners of EVs.“I’m extremely proud of the historic level of investment in America’s bridges—at more than $50 billion, it’s the largest such investment in our history. And the BUILD America 250 Act ensures that electric vehicle owners begin paying their fair share for the use of our roads,” said committee chairperson Sam Graves (R-Mo.).Should the bill pass—and it enjoys support from the Democratic Party, too—you will be required to pay a $130 federal registration fee to drive an EV. And starting in 2029, that fee will increase by $5 each year until it reaches $150. Plug-in hybrids don’t escape untaxed, either; the fee for a PHEV begins at $35 a year and will escalate by $5 each year until it reaches $50 annually. And if state departments of transport don’t collect this federal EV tax, the federal government will “withhold an amount equal to 125 percent of the amount owed from the state’s highway apportionment.”Why?Road funding in the US is a messy business, paid for by a mix of federal and state gas taxes, state vehicle registration fees, and local budgets. And as the percentage of fuel-efficient PHEVs and gas-abstaining EVs grows, the amount of money available to pay for road upkeep drops. So, in the abstract, asking EV and PHEV owners to pay their share is not unreasonable, especially since the slightly higher curb weight of these vehicles will do slightly more damage to the roads, although passenger vehicles are a rounding error compared to a garbage truck, bus, or tractor-trailer.Some stateshave already started imposing a registration fee on clean vehicles to offset reduced gas tax receipts. And at some point, when a significant proportion of the country’s vehicles are EVs or PHEVs, an alternative road-funding method that does not rely on the gas tax will certainly be in order.But EV adoption in the US hastaken a severe hitsince the election of President Trump and the abolition of federal clean-vehicle incentives. As of now, EVs make up barely 3 percent of the entire vehicle fleet. Meanwhile, the federal gas tax has not been increased since 1993; had it been adjusted to keep pace with inflation, it would add an extra 42.34c to a gallon rather than a pathetic 18.4 cents currently.“Drivers of gas-powered vehicles pay approximately $73 to $89 in federal gas tax each year. The proposed fee would charge an unfair premium on EV drivers at a time when all Americans are looking for ways to save money,” said Albert Gore, executive director of the Zero Emissions Transportation Association. “This is particularly concerning as the EV fee will increase to $150 by 2035—nearly double what gas car drivers would pay in a year. This fee lands on top of the road use taxes that many EV drivers already pay at the state level.”“In addition to the new EV fee, the bill makes major changes to federal investments for dedicated EV charging. A lack of dedicated EV funding could particularly impact EV drivers living in—and traveling to—more rural and remote locations in the United States, where federal funding is most needed to help fill gaps in existing charging infrastructure,” Gore said.“With Americans demanding more affordable transportation options as gasoline tops $4.50 a gallon, this bill would impose an onerous new fee on electric vehicles and plug-in hybrids while slashing investments in new EV chargers,” said Shruti Vaidyanathan, director of federal and state transportation advocacy at the National Resources Defense Council. “Congress should be boosting investments in projects that cut costs, cut emissions, create jobs, and build a transportation system that works for all Americans. This bill largely ignores the need to build cleaner, more affordable transportation options.”We should perhaps be glad that the final version will cost drivers much less than Rep Graves’last proposal, which started at $250/year for EVs and $100/year for hybrids.Jonathan M. GitlinAutomotive EditorJonathan M. GitlinAutomotive EditorJonathan is the Automotive Editor at Ars Technica. He has a BSc and PhD in Pharmacolog

Gotta go fastThe era of 1,000 Hz gaming monitors has arrived, but why?LG’s latest hits one frame per millisecond at a full 1080p resolution.Kyle Orland–May 19, 2026 12:26 pm|82Wow, look how fast those giant letters are moving without any apparent motion blur!Credit:

      LGWow, look how fast those giant letters are moving without any apparent motion blur!Credit:

      LGText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAlmost exactly two years ago, we weregawking at prototypes of 1,000 Hz monitorsand wondering who really needed a display that could support such ludicrously smooth frame rates. Now that those prototypes are starting to develop into retail products, we’re still wondering how much of a market there is for gaming displays that can update with a new frame every single millisecond.The latest entry in the ultra-fast refresh race is LG’s 24.5″ UltraGear 25G590B, which the companyannounced this weekas “the world’s first Full HD gaming monitor with a native 1000Hz refresh rate” ahead of a planned launch in “select markets” in the second half of the year. That “Full HD” promise means LG’s 1,000 Hz display hits the 1080p threshold that is by far the most common resolution reported by gamers inSteam’s regular hardware surveys.That would represent a decent step up from the likes ofAcer’s Predator XB273U F6,Samsung’s Odyssey G6, orPhillips’ EVNIA 27M2N5500XD, all of which have to shift down to a relatively blurry 720p resolution to run at a full 1,000 Hz (but which support 1440p resolutions at a still-quite-fast 500 Hz). LG also notes that its high-end monitor can hit its resolution and refresh rate specs natively, without the need for any “dual mode” rebooting shenanigans to get the fastest performance.Fast-moving objects appear decidedly less blurry as you climb the frame rate curve.Credit:Blur BustersFast-moving objects appear decidedly less blurry as you climb the frame rate curve.Credit:Blur BustersThe folks over at Blur Busters haveextensivelydocumentedresearch showing that refresh rates of 1,000 Hz (and up) can reduce human perception of motion blur and flickering. And while the site notes that you eventually hit “diminishing returns” from all those extra frames—especially on smaller screens—there’ssome evidencethatyou would need a 40,000 Hz monitorto totally eliminate perceived motion blur on a sufficiently large, high-resolution monitor.Perception research aside, it’s unclear how many gamers are interested in spending extra money on a display that breaks the four-digit refresh rate barrier. While LG hasn’t announced pricing details for the UltraGear 25G590B yet, other 1,000 Hz monitors tend to run hundreds of dollars more than comparable models with lower (but still more than adequate) maximum refresh rates in the 240 to 360 Hz range. At the same time,monitors with Nvidia’s GSync Pulsarand similar technologies use “rolling scan” pulsing backlights to reduce apparent motion blur, even on “slower” 360 Hz monitors.LG says the 25G590B is “engineered specifically for first-person shooters” and mentions an “esports-optimized design” in its marketing materials, suggesting the display is targeting the small subset of gamers who play twitch-based reflex games for a living. For everyone else, we recommend ignoring the “larger number is better” impulse and saving your hardware budget forthose increasingly expensive sticks of RAM.Kyle OrlandSenior Gaming EditorKyle OrlandSenior Gaming EditorKyle Orland has been the Senior Gaming Editor at Ars Technica since 2012, writing primarily about the business, tech, and culture behind video games. He has journalism and computer science degrees from University of Maryland. He oncewrote a whole book aboutMinesweeper.82 Comments

Gemini’s efficiency playGemini 3.5 Flash might be fast enough for gen AI to make senseGoogle says its more efficient Gemini 3.5 Flash is the key to your agentic AI future.Ryan Whitwam–May 19, 2026 2:11 pm|56Credit:

      Aurich LawsonCredit:

      Aurich LawsonText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAt last year’s I/O event, Google was still talking about the2.5 branchof Gemini, and what a difference a year makes. We’ve gone through the3.0and3.1families since then, and now it’s on to version 3.5.Gemini 3.5 Flashis rolling out across a wide range of Google products starting today, and Google again claims this model is even better than its last-gen Pro model.That has been a trend with Google’s tick-tock model updates over the past year, but the team says this release is special. Gemini 3.5 Flash supposedly offers frontier-level intelligence while also being efficient enough that it may finally make complex agentic tasks worth doing at scale. Tulsee Doshi, senior director of product management for Gemini, explains that the innovations of Gemini 3.5 Flash are woven through multiple Google products, and this is just the start.Credit:
                                  GoogleCredit:

      GoogleIt’s no secret that generative AI is currently a money pit, and all the major AI players are trying to find paths to greater efficiency. The problem is magnified when you start building agentic experiences that are supposed to run for longer to complete complex tasks. Gemini 3.5 Flash may be a big step toward making that viable. The new model can output nearly 300 tokens per second, but its benchmark scores are similar to larger frontier models (like 3.1 Pro) that build outputs at a quarter of that speed.Google now says that the companies using the most AI tokens could save a billion dollars per year by shifting to the more efficient Gemini 3.5 Flash. API pricing for the new model is significantly lower than the Pro model it apes. Gemini 3.5 Flash clocks in at $1.50 per 1M input tokens and $9 per 1M output tokens. The 3.1 Pro model starts at $2 and $12, respectively, and it’s higher if you use more than 200k tokens.According to Doshi, the team made numerous improvements in pre-training with Gemini 3.5 Flash, but insights gleaned from how devs use Gemini models are really paying off.“With post-training, we’re really starting to unlock some of the value of the feedback we’re getting from users, for example, from Antigravity,” said Doshi. “That’s really what you’re seeing play out in terms of the code performance and the tool use performance. And then, the hope is that you’ll continue to see the step change where 3.5 Pro will be better, and the next Flash meets Pro performance with that series.”Google is focused on code generation with the new model, which is a core agentic angle for AI. Both Terminal Bench and SWE-Bench Pro tests show substantial improvements—3.5 Flash clobbers older Flash models and shows a small but measurable improvement versus Gemini 3.1 Pro. Its scores are in the same neighborhood as OpenAI’s much larger and more expensive GPT 5.5.A major barrier in agentic workflows is how generative models can use interfaces designed for humans. It’s not an easy problem to solve, Doshi said. “Certain things like UI control are expensive to do because the model has to search the page, it has to know where to click, it has to act through multiple steps. I think Flash is able to do that well because of that combination of quality and cost.”Google’s AI evaluations demonstrate these improvements, too. Among Google’s current collection of benchmarks is OSWorld-Verified, which tests how models handle general tasks in real computing environments. It’s similar to the coding improvements. Gemini 3.5 Flash substantially outperforms older Flash models and is even a bit faster than Gemini 3.1 Pro. It’s essentially tied with GPT 5.5.Google’s new Flash model is, again, a little better than the last-gen Pro.Credit:
                                  GoogleGoogle’s new Flash model is, again, a little better than the last-gen Pro.Credit:

      GoogleGemini 3.5 Flash has been rolled out internally at Google, and Doshi noted that it’s having a big impact. “We have a set of internal metrics we’ve been evaluating that measures how Googlers code, so looking at our own code bases and how well the models perform on that,” Doshi said. “And you can see a massive, massive jump between where 3.1 Pro was and where 3.5 Flash is.”Google unveiled the Antigravity IDE last year, and it’s being upgraded to version 2.0 with support for Gemini 3.5 Flash. This update will support multiple parallel workflows—essentially sub-agents spawned by Gemini 3.5 Flash. Again, Google says this is only possible because the new model is so efficient at spitting out tokens.In addition to Antigravity, Gemini 3.5

Foiled againRFK Jr. forced to withdraw charter that opened CDC panel to anti-vaccine quacksCharter would have expanded member eligibility and focused on alleged injuries.Beth Mole–May 19, 2026 2:17 pm|43WASHINGTON, DC - JUNE 24: Health and Human Services (HHS) Secretary Robert F. Kennedy Jr. testifies before the House Energy and Commerce Committee Health Subcommittee in the Rayburn House Office Building on June 24, 2025 in Washington, DC.Credit:Getty | Kayla BartkowskiWASHINGTON, DC - JUNE 24: Health and Human Services (HHS) Secretary Robert F. Kennedy Jr. testifies before the House Energy and Commerce Committee Health Subcommittee in the Rayburn House Office Building on June 24, 2025 in Washington, DC.Credit:Getty | Kayla BartkowskiText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navA revised charter document for the Centers for Disease Control and Prevention’s influential vaccine advisory committee has been withdrawn by the Health Department over an administrative error, according toa notice published in the Federal Register Tuesday.The charter’s revisionsunder anti-vaccine Health Secretary Robert F. Kennedy Jr. would have allowed Kennedy to appoint dubiously qualified anti-vaccine allies to advise the CDC. It also would have directed the CDC panel to focus on alleged vaccine injuries and risks and welcomed fringe groups and anti-vaccine organizations to participate in developing federal vaccine policy.Kennedy’s move to reshape the CDC panel—the Advisory Committee on Immunization Practices, or ACIP—came amid Kennedy’s many other attempts to undermine it, as well as a court order to undo that meddling.In June 2025,Kennedy summarily fired all 17 expertsfrom ACIP and quickly replaced them with unvetted and unqualified anti-vaccine allies. Kennedy’s ACIP then held several meetings in which they aired anti-vaccine views and misinformation, allowed anti-vaccine activists to give unvetted presentations, and ultimately voted to remove longstanding, evidence-based federal vaccine recommendations based on anti-vaccine rhetoric. That includes the removal of a universal recommendation for a hepatitis B vaccine dose at birth, despite no evidence of any safety concerns or any benefit to delaying the dose. Subsequent modeling studies found that the change will meanmore infections, increases in liver cancers and deaths, as well as millions of dollars inhealthcare costs.In January, Kennedy sidestepped ACIP entirely to overhaul the CDC’s vaccine schedule for children,slashing the number of recommended vaccinations from 17 to 11.In March, a federal judge issueda temporary injunction to undo Kennedy’s changes, ruling that Kennedy likely violated federal regulations in unilaterally altering the CDC vaccine schedule and that the new ACIP members were unqualified—as outlined by the panel’s charter—nullifying all of their votes on federal vaccine recommendations.While the Health Department is working to appeal the injunction, Kennedy attempted to circumvent the judge’s ruling on the ACIP members byaltering the committee’s charterto, among other things, allow for people without expertise in immunizations and public health to be members.But, for now, that effort, too, has been thwarted. According to the notice on Tuesday, the new charter has been withdrawn for not following a federal requirement on public notification.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.43 Comments

GitFlubIn stunning display of stupid, secret CISA credentials found in public GitHub repoSSH keys, plaintext passwords, other sensitive data had been up since November 2025.Lee Hutchinson–May 19, 2026 2:27 pm|68Only the best people.Credit:

      Dzmitry Skazau / GettyOnly the best people.Credit:

      Dzmitry Skazau / GettyText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navSecurity researcher Brian Krebsbrings us the newsthat America’sCybersecurity & Infrastructure Agency(CISA) has had a large store of plaintext passwords, SSH private keys, tokens, and “other sensitive CISA assets” exposed in a public GitHub repo since at least November 2025.The now-offline public repo—named, somewhat aspirationally, “Private-CISA”—was brought to Krebs’ attention by GitGuardian’sGuillaume Valadon, who was alerted to the repo’s presence by GitGuardian’s public code scans. Krebs says that Valadon approached him after receiving no responses from the Private-CISA repo’s owner.In an email to Krebs, Valadon claimed that the repo’s commit logs show that GitHub’s default protections against committing secrets—protections designed to protect unwitting or unskilled developers against exactly this kind of stupidness—had been disabled by the repo’s administrator.Testing bySeralys founder Philippe Catureglishowed that this was not a joke or hoax and that he was able to use the credentials in the Private-CISA repo to gain access to multiple Amazon Web Services GovCloud accounts “at a high privilege level.”Krebs notes that the repo appeared to be managed by Virginia-basedNightwing, a CISA contractor. Nightwing has so far not commented publicly, instead referring questions back to CISA.This isn’t the first time CISA has screwed up—in fact, it’s not even the first timethis year. In January,polygraph-failingacting CISA Director Madhu Gottumukkalauploaded sensitive government documents to ChatGPTafter demanding and receiving an exemption to the agency policy that prohibited ChatGPT’s use by CISA personnel. Gottumukkala wasremoved from his role in February.Lee HutchinsonSenior Technology EditorLee HutchinsonSenior Technology EditorLee is the Senior Technology Editor, and oversees story development for the gadget, culture, IT, and video sections of Ars Technica. A long-time member of the Ars OpenForum with an extensive background in enterprise storage and security, he lives in Houston.68 Comments

What is real?Google’s SynthID AI watermarking tech is being adopted by OpenAI, Nvidia, and moreAI content is getting good, but SynthID might be able to help tell truth from fiction.Ryan Whitwam–May 19, 2026 2:37 pm|36Google CEO Sundar Pichai announces 100 billion SynthID images and videos.Credit:

      GoogleGoogle CEO Sundar Pichai announces 100 billion SynthID images and videos.Credit:

      GoogleText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navIn a few short years, we’ve gone from easily identifying AI content that featured superfluous fingers to images and videos that lookshockingly realistic. How can we know what’s real in the age of AI? Google’s answer is SynthID, which it first demonstrated three years ago.The company saysSynthID has since been used to label 100 billion images and videos, plus 60,000 years’ worth of audio. Those numbers are only going up now that SynthID is expanding beyond Google.SynthID is not Google’s only AI labeling strategy. It’s also committed to the C2PA standard, which tags content with metadata describing how it was created. Google began using C2PA more prominently with its Pixel 10 smartphones. Photos taken with the Pixel 10 include metadata describing how they were processed. If a highly zoomed image includes generative elements, it gets an AI tag, too.Google now says this same feature is coming to videos recorded on Pixel 8, 9, and 10 phones in an update in the coming weeks. It’s also adding C2PA scanning to Gemini, allowing the chatbot to explain a file’s providence based on the content labeling. This same capability will come to Chrome and Search in a few months.That metadata is fungible, though. On the other hand, SynthID is deeply integrated with AI-generated content. The digital watermark is present in the pixels of images and videos and in the waveform of AI songs and audio overviews from products like NotebookLM. According to Google DeepMind scientist Pushmeet Kohli, the team worked hard to ensure SynthID is much harder to remove, even if you compress it, crop it, or rotate it.“A technology like this will always be attacked,” said Kohli. “There was a lot of research that we did in making SynthID robust to different kinds of transformations.”Last year, Google added support for SynthID detection in the Gemini app. You can upload the suspect content and ask the chatbot if it’s AI-generated. This should work reliably with all those billions of Google AI images and audio clips from the past three years. A few ambitious tinkerers have claimed to find methods for removing the hidden SynthID patterns. Google contends that none of these bypasses actually work.More SynthID in more placesEven if no one has been able to crack SynthID, that doesn’t matter for the vast majority of AI images on the Internet—only Google’s AI models apply SynthID. That’s going to change soon, though.Google has announced that it has partnered with several companies to add SynthID to their systems. Nvidia will implement SynthID in its Cosmos world foundation models, and OpenAI will use SynthID in its GPT 2 images. Kakao and ElevenLabs will also begin adding SynthID to their AI content.This doesn’t mean you’ll always be able to tell if something is AI by looking for SynthID. Plenty of publicly available models continue to produce content with no AI watermarking, and there are open models that can be trained by anyone looking to create AI images and videos on their own terms. Still, this is a step in the right direction.There will also be new paths to checking SynthID status, so you won’t even have to open Gemini just to check for the watermark. SynthID will be integrated with Circle to Search, Lens, and AI Mode. You’ll also be able to use Gemini in Chrome by sharing a tab with the content in question. You can ask any variation of “Is this AI” to get a SynthID scan with these tools.There is currently not a public API for SynthID—making these scans too readily available could serve as an attack vector for those seeking to circumvent SynthID. However, Google is preparing to launch an AI content detection API as part of the company’s Gemini Enterprise Agent Platform. This will allow trusted business partners to more easily flag AI content, allowing Google to refine the API over the coming months.Ryan WhitwamSenior Technology ReporterRyan WhitwamSenior Technology ReporterRyan Whitwam is a senior technology reporter at Ars Technica, covering the ways Google, AI, and mobile technology continue to change the world. Over his 20-year career, he's written for Android Police, ExtremeTech, Wirecutter, NY Times, and more. He has reviewed more phones than most people will ever own. You canfollow him on Bluesky, where you will see photos of his dozens of mechanical keyboards.36 Comments

We can help with thatTwo AI-based science assistants succeed with drug-retargeting tasksBoth tools generate hypotheses; one goes on to analyze some of the data.John Timmer–May 19, 2026 2:55 pm|5Finding connections within the messy world of biology is central to these new tools.Credit:Andriy OnufriyenkoFinding connections within the messy world of biology is central to these new tools.Credit:Andriy OnufriyenkoText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navOn Tuesday, Nature released two papers describing AI systems intended to help scientists develop and test hypotheses. One, Google’s Co-Scientist, is designed as what they term “scientist in the loop,” meaning researchers are regularly applying their judgments to direct the system. The second, from a nonprofit called FutureHouse, goes a step beyond and has trained a system that can evaluate biological data coming from some specific classes of experiments.While Google says its system will also work for physics, both groups exclusively present biological data, and largely straightforward hypotheses—this drug will work for that. So, this is not an attempt to replace either scientists or the scientific process. Instead, it’s meant to help with what current AIs are best at: chewing through massive amounts of information that humans would struggle to come to grips with.What’s this good for?There are some distinctions between the two systems, but both are what is termed agentic; they operate in the background by calling out to separate tools. (Microsoft has taken a similar approach with its science assistant as well; OpenAI seems to be an exception in that it simplytuned an LLM for biology.) And, while there are differences between them that we’ll highlight, they are both focused on the same general issue: the utter profusion of scientific information.With the ease of online publishing, the number of journals has exploded, and with them the number of papers. It has gotten tough for any researcher to stay on top of their field. Finding potentially relevant material in other fields is a real challenge. If you’re focused on eye development, for example, one of the signaling systems used may also be involved in the kidney, and it can be easy to miss what people are discovering about it.As the people at FutureHouse put this issue, “By focusing on ‘combinatorial synthesis’ (identifying non-obvious connections between disparate fields), Robin effectively targets ‘low-hanging fruit’ that human experts may overlook due to the compartmentalization of scientific knowledge.”This is a task that’s well-suited to AI, which can chew through the peer-reviewed literature in the background while researchers do other things. This isn’t really a question of whether an AI could do something better or worse than a human; it’s more of an issue of whether any human would end up doing these sorts of searches at all.By finding enough connections among disparate research, these tools can make suggestions—hypotheses, really—about the biology. This can include things like what processes underlie biological behaviors and what pathways and networks regulate those processes. And, in the cases explored here, it included suggesting known drugs that might target some of these pathways in diseased cells: acute myeloid leukemia in Google’s case, and a form of macular degeneration for FutureHouse.Co-scientistAs you might imagine, Google’s system is based on the company’s Gemini large language model. That helps the system interpret a statement of research goals provided by human scientists and starts a literature search to find relevant information and form hypotheses. Those are then evaluated relative to each other in a “tournament,” the results of which are evaluated by a Reflection agent. An Evolution agent can then make improvements to any surviving ideas, which can be sent back through the process.Key criteria considered throughout this process include plausibility, novelty, testability, and safety. And the Reflection tool has access to external search tools, as access to the scientific literature “prevented the hallucination of seemingly novel but implausible hypotheses,” the company wrote.As the paper puts it, scientists were kept in the loop at all times. In the search for potential drugs targeting leukemia, the suggestions made by the system were prioritized based on a review by a panel of experts, who had access to the literature Co-Scientist used to formulate its suggestions.The results are what you would expect from cancer therapies. Some of the drugs identified were effective, but only against subsets of a panel of myeloid leukemia cells. That’s not unusual, given that there are multiple routes to unchecked growth, so drugs that block the route followed by one cell type may not be effective in cells that took a different route.Google also mentioned that the system could do more general hypothesizing that do

Self-hostingPlex’s 200% Lifetime Pass price hike tries forcing users to another subscriptionPlex says that it has considered getting rid of Lifetime Passes.Scharon Harding–May 19, 2026 3:52 pm|124A marketing image for Plex that emphasizes its streaming, rather than its media server, business.Credit:

      PlexA marketing image for Plex that emphasizes its streaming, rather than its media server, business.Credit:

      PlexText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navAs of July 1, at 12:01 am UTC—or June 30 at 8:01 pm ET—people seeking access to Plex’s media server features through a one-time purchase will have to pay $750. That’s three times the current price of $250.The new price will not affect current Lifetime Plex Pass holders.A Lifetime Plex Pass allows you to stream from your own Plex Media Server to a device connected to your own network, to stream from the server remotely, and to allow others to stream remotely from your server.When first launched in 2012, a Lifetime Plex Pass was $75. In 2014, Plex increased the price to $150 because “at 2.5x the price of a yearly, it was priced in an unsustainable way for us,” Plex said in ablog postat the time. Eventually, lifetime passes were available for $120 for years until March 2025, when the priceskyrocketed to $250. That brings us to today’s announcement.In an email andblog postto customers today, Plex said that it sold lifetime passes early in its existence “because we knew many of our customers would rather pay a higher, one-time fee for software that they can depend on every day.” However, as evidenced by several price hikes, Plex has struggled to reconcile lifetime passes with its own financial goals. The company’s message says:We’ve considered eliminating the Lifetime Plex Pass in the past, given that recurring subscriptions help us sustain long-term development, but we know it’s still a valuable option for many in our community. So instead of retiring it, we’re keeping it available at a price that reflects the real, ongoing value of the software we’re committed to building and maintaining for years to come.The 200 percent price hike and Plex’s admission that it has mulled killing Lifetime Passes illustrate Plex’s shifting priorities as itseeks profitability. Notably, Plex didn’t announce any pricing hikes for its monthly or annual subscription tiers.In its blog, Plex offered further reasoning for the upcoming price hike:Over the years, as our software and product has evolved, the breadth of features and benefits included with your Plex Pass has expanded. This increase ensures we can continue to invest resources into building and maintaining the Plex personal media software, while continuing to offer a Lifetime option.As noted in the announcement, Plex would rather people pay monthly or annually for its media server features than a one-time fee. With annual subscriptions currently $70, it would take 11 years for a $750 lifetime subscription to be a better value.Plex’s price hike is an extreme example of streaming service providers continuously increasing prices amid struggles to reach and maintain profitability while keeping prices stable and dealing with ongoing costs, like licensing fees for rental movies and app updates.Plex claims to be working on numerous new features for its platform, including: adding all server and library management features currently available on app.plex.tv to Plex’s mobile and, where applicable, TV apps; “boosting dialogue and normalizing loudness,” per today’s announcement; transcoding improvements; IPv6 support; and the abilities to group downloads by show, automatically download new episodes, and make and edit playlists in mobile apps. New features that Plex has added recently include the ability to create custom metadata agents and an open API for server integrations.Beyond new features, Plex has previously cited rising costs for the hike, including supporting various types of devices andcodecs.Plex moving beyond media servingPlex’s various forays into areas outside of its original media server business also contribute to its rising costs. In recent years, Plex has expanded from a media server company to a streaming service provider that hashundredsof free ad-supported TV (FAST) channels,rents out movies, and dabbles insocial features. It also made afailed venture into gaming.Legacy customers, however, have grown wary of Plex’s expanding interests and their potential impact on Plex’s media server users. While adding streaming and social features,Plex killed free remote streaming accessand itsWatch Together featureand issued acontroversial app redesign. Plex also has reason to put more focus on its streaming business since more people have been using Plex’s online streaming service than its media server capabilities since 2022, Scott Hancock, Plex’s then-VP of marketing,saidin 2023.The good news is th

Closing the gap“I’ll buy 10 of those”—NASA science chief yearns for mass-produced satellites“How in the hell do I get more science into space? That is my goal.”Stephen Clark–May 19, 2026 4:43 pm|15Saturn’s moon Enceladus peeks over the limb of Dione during a partial occultation, as seen by NASA’s Cassini spacecraft on September 13, 2008.Credit:

      NASA/JPL-Caltech/Space Science InstituteText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThere are more opportunities to access space than ever, thanks to a bevy of commercial rockets, some with reusable boosters, led by SpaceX’s workhorse Falcon 9. So why is NASA launching fewer telescopes and planetary science missions than it did a quarter-century ago?The answer is complex. It is not necessarily the money. The space agency’s science budget this year is $7.25 billion, roughly the same as it was in 2000, adjusted for inflation. This is despite attempts by the Trump administration to drastically reduce NASA science funding.In the early months of his tenure, NASA Administrator Jared Isaacman’s focus has been on human spaceflight and the Moon. This isn’t terribly surprising given NASA’swildly successful Artemis II missioncarrying four astronauts around the Moon last month. Since taking office in December, Isaacman has announced an overhaul of the Artemis program,canceling a space stationto be built in orbit around the Moon in favor of construction of a base on the lunar surface.On the robotic front, Isaacman is pushing for NASA to launch afirst-of-its-kind nuclear-powered spacecraftin 2028 to deliver a trio of drone rotorcraft to explore Mars. Isaacman has not said as much about concrete changes to NASA’s science program. He has defended the Trump administration’s proposed cuts to NASA’s science budget—as would be expected of him as a Trump political appointee—but the budget proposals come from the White House, not from NASA headquarters.“Mr. Isaacman is very keen on us doing things quicker and for less,” said Nicky Fox, associate administrator for NASA’s science mission directorate. “More shots on goal is one of his favorite phrases. And I think, for us, it’s looking at the right-sized mission for the problem. Not everything has to be $1 billion or more. There are ways you can do fantastic science. His challenge is he wants 10 $100 million missions to be flying.”How to get there?A future with numerous robotic probes spread throughout the Solar System sounds thrilling to space scientists and space enthusiasts, but you can’t get there with flat budgets and billion-dollar missions that take a decade to get off the ground. Many of NASA’s robotic science missions use purpose-built satellites and instruments, usually manufactured by large contractors like Lockheed Martin, Northrop Grumman, university labs, or NASA itself. Unlike SpaceX’s hangars full of reusable rockets, there’s no building with cameras, spectrometers, telescopes, and spacecraft buses—the core chassis of a satellite platform—lying around waiting to launch.“Instead of having a bespoke bus that does absolutely everything, and makes the tea and brings you toast, what can you do with an off-the-shelf bus?” Fox told Ars. “And maybe you have to change a few things. Maybe you fly fewer instruments, but maybe you fly three [spacecraft] together. How do we really pick up the pace? Because it is difficult when you have long gaps between the missions. It’s certainly not what anyone wants to see.”One way to make this future real is with mass-produced, high-power satellites. Small CubeSats, just the size of a suitcase, are great for missions close to home, but they won’t cut it for missions to more distant destinations, such as another planet or a unique orbit far from Earth. NASA is making use of other ways to collect scientific data in space, such as placing instruments on the International Space Station or on commercial communications satellites.But those solutions won’t work if you want to travel to another world. Sometimes it just costs a lot of money to do the near-impossible.“For $100 million, you can’t buy a bus from somewhere and put four instruments on it and send it toflight to Enceladus to look under the icethere,” Fox said. “No, that’s a big, ambitious mission. We want to fly an interstellar-type probe. As theVoyagers are getting older, we want to study interstellar space. These things are hard, and they’re tough, and it will take a lot of effort to do that. We also talked about actually flying a mission to Uranus.”But what about spacecraft flying on more well-trodden paths to the Moon, Mars, Venus, or the asteroid belt? “What can we do with these commercial off-the-shelf buses? I would love to walk in and say, ‘I’ll buy 10 of those,’” Fox said.NASA islooking at “block buys” for the next series of commercial missionsto the Moon. These privately owned landers and orbiters, part of the Commercial Lunar Pa

The “Cage-aissance” continuesSpider-Noirfinal trailer gives us a classic villainIt’s never too late to become a hero.Jennifer Ouellette–May 19, 2026 4:58 pm|11Credit:

      YouTube/Prime VideoCredit:

      YouTube/Prime VideoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navPrime Video has released one last trailer for its upcoming live action series,Spider-Noir,starring Nicolas Cage, and once again it’s been released in two formats: one in black and white (below) and another in color (above), which the showrunners are calling “True Hue.” Seriously, the more footage we see of this series, the more eager we are to find out if the series lives up to its marketing. And the final trailer—which really plays up the deadpan humor and is set to Amy Winehouse’s “Back to Black”—is very promising.Aspreviously reported, Marvel Comics created its “noir” line in 2009, reinterpreting familiar Marvel characters in an alternate universe, usually set during the Great Depression in the US. A version of the Spider-Noir character, voiced by Cage, briefly appeared in the animated masterpiecesSpider-Man: Into the Spider-Verse(2018) andAcross the Spider-Verse(2023). (He is set to reprise that role in the upcomingBeyond the Spider-Verse.)Cage is playing Ben Reilly, a hard-boiled PI with a secret superhero identity, The Spider. Per the official premise: “Spider-Noirtells the story of Ben Reilly, a seasoned, down on his luck private investigator in 1930s New York, who is forced to grapple with his past life, following a deeply personal tragedy, as the city’s one and only superhero.”In addition to Cage’s Ben Reilly/The Spider, the cast includes Lamorne Morris as Reilly’s friend Robbie Robertson, a freelance journalist who clings to optimism in the face of his buddy’s cynicism; Li Jun Li as nightclub singer Cat Hardy, the classic underworld femme fatale (Li based her portrayal on Anna May Wong, Rita Hayworth, and Lauren Bacall); Karen Rodriguez as Reilly’s secretary, Janet; Abraham Popoola as a World War I veteran; Jack Huston as a bodyguard named Flint Marko who becomes (as we see in the new trailer) the classic villain Sandman; Brendan Gleeson as New York mob boss Silvermane, who is being targeted for assassination; Lukas Haas as one of Silvermane’s subordinates; Richard Robichaux as the editor of the Daily Bugle; and Kai Caster.“Have you seen it out there? The city’s a mess,” Robbie tells Reilly, reminding his friend of the hero he used to be. But the disillusioned Reilly knows that’s just not who he is now and he wants to start over—as a PI. But then Cat Hardy walks into his office and hires him to investigate a friend’s disappearance (killing an errant spider with a newspaper, which probably won’t endear her to Spidey).That friend is Jack Huston, who has “turned into some kind of sand monster.” And others also seem to be coming down with mysterious illnesses that worsen until “their sicknesses eat them alive.” Meanwhile, Silvermane has pretty much taken over the city. Will Reilly stick to his resolve and stay out of it? Or will he resurrect his superhero alter ego one last time? I think we all know it’s gonna be the latter. “I was never a hero,” Reilly insists. But it’s not too late for him to become one.Spider-Noirpremieres on May 25, 2026, on MGM+, with all episodes becoming available on Prime Video on May 27, 2026. Viewers can choose to watch in black and white or True Hue—or both, if one wants to compare.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.11 Comments

FBI wish listFBI seeks US-wide access to license plate cameras, wants “data in near real time”FBI will pay vendors to help it track and search for vehicles nationwide.Jon Brodkin–May 19, 2026 5:29 pm|56Flock license plate reader and camera with solar panel in Pleasant Hill, California on April 16, 2026.Credit:

      Getty Images | Smith Collection/GadoFlock license plate reader and camera with solar panel in Pleasant Hill, California on April 16, 2026.Credit:

      Getty Images | Smith Collection/GadoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe Federal Bureau of Investigation announced plans to buy nationwide access to a network of license plate readers, saying it will award contracts to one or more vendors that can offer “near real time” information from cameras across the US. The proposed contract is for the FBI Directorate of Intelligence.“To evaluate and manage threats to personal safety, property, and law enforcement, the FBI requires professional service firms that can provide License Plate Readers (LPRs) for tracking subjects on roads and highways over the US and its territories,” the FBI said in aRequest for Proposals (RFP)published on May 14. The FBI said the winning bidder or bidders “must provide law enforcement and/or commercial license plate reader data provided through the Contractor’s existing platform.” The system must cover 75 percent of locations, the FBI said.The system must offer the ability to search for license plate information “and other descriptive data such as vehicle description information, time/date criteria, and geo-location criteria,” the FBI said. “Additionally, the system must provide search result notifications. The Contractor system must have the ability to access and/or query cameras across the United States and its territories. The Contractor system must be capable of providing this data in near real time.”Contractors have to be able “to share/create maps depicting camera coverage (i.e. heat mapping),” and “provide the FBI the source of information (i.e. red-light cameras, repossession vendors, speed cameras, etc.),” the FBI said. The FBI said it needs to be able to search the database for partial or full plate numbers, plate states, addresses, locations where a plate was scanned, and vehicle makes and models.Flock and Motorola Solutions potential biddersThe RFP divides the proposal into six regions covering the continental US, Hawaii, Alaska, Puerto Rico, and territories such as Guam and the US Virgin Islands. The FBI said it may award contracts to one or two vendors in each region. The deals can be for up to five years, with all deals combined potentially worth $36 million. The FBI said a contractor’s system has to be available to FBI users via a website.Flock and Motorola Solutions are well-positioned to bid on the contract, as 404 Medianoted yesterday. Both companies could win part of the job, as the FBI said it may award contracts to multiple vendors to achieve its desired level of access.Flock’s Automated License Plate Readers (ALPRs) are sold to local police departments. The company boasts of having deals with “over 12,000 public safety customers including cities, towns, counties, and business partners.” Motorola Solutionssells license plate reader camerasthat can be installed on busy roadways or mounted on police cars.License plate reader cameras have raisedconcerns about privacy,data security, anderrors in plate number recognition systemsleading to wrongful arrests. 404 Mediareported last yearthat local police departments performed searches of the Flock license plate reader system for US Immigration and Customs Enforcement (ICE), “giving federal law enforcement side-door access to a tool that it currently does not have a formal contract for.”The FBI already “runs a License Plate Reader program to facilitate LPR information sharing with and between its law enforcement partners,” a Congressional Research Service reportsays. The US agency “maintains a hot list of vehicle data against which law enforcement agencies can compare their LPR data.”The FBI intelligence division’s plan to obtain direct access to an extensive network of cameras could help expand that information sharing. As the FBInotes, its intelligence division shares information with a variety of federal, state, local, and tribal law enforcement agencies.Flock: Sharing is “opt-in” for local policeFlock itself temporarily provided access to Customs and Border Protection, Homeland Security Investigations, the Secret Service, and Naval Criminal Investigative Service as part of a pilot last year. Flock confirmed the pilot to the office of Sen. Ron Wyden (D-Ore.),according to Wyden.Flock saysit has federal customers “including National Parks, Veterans Affairs hospitals, and military bases,” but that it does not work with ICE.Federal attempts to access data could be limited by company pol

在今天和理想全新 L9 和智界 V9 两台重磅车型同台亮相的，还有蔚来的乐道 L80。相比 L90 面向三代同堂、多孩家庭的大三排需求，L80 把重点放在大五座市场，带来了更大的乘坐空间，更灵活的装载能力，以及更适合长途、自驾、露营等场景的空间组织方式。早年间，以本田冠道、大众途观 L 为代表的燃油大五座 SUV，主要通过释放第二排的腿部空间来提升乘坐舒适性；随后，随着新能源时代的到来，诸如理想 L7、问界 M7 等车型进一步强化了智能座舱的沉浸式体验和车内的储物能力。而乐道 L80 现在则试图在行业已有的基础上更进一步，将前备舱、后备舱、乘员舱以及补能体系整合为一个完整的生态闭环，通过蔚来的工程创新能力，让中大型五座 SUV 的日常使用体验再上一层楼。乐道 L80 一共有 Pro、Max+ 和 Ultra+三个版本，整车购买价格分别为 24.28万元、25.98 万元、27.98 万元， Baas 方案价格为 15.68 万元、17.38 万元和 19.38 万元。前备舱大，后备舱更大空间，依然是乐道 L80 最核心的竞争力。乐道 L80 车长 5145mm，轴距 3110mm，理想 L7 车长 5050mm，轴距 3005mm，乐道 L80 比理想 L7 大了一圈。取消第三排之后，乐道 L80 将所有空间都留给了前两排和储物。第二排作为 L80 整台车体验的核心区域，其后排头部空间达到 1074 毫米，腿部空间宽达 1121 毫米，坐垫宽度和长度分别达到 549 毫米和 497 毫米，靠背支持最大 135 度的后倾调节。L80 提出了一个「大五座，享四座」的越级体验概念，即当车内只有四人出行时，第二排的两位乘客完全可以获得媲美大型 MPV 独立航空座椅的放松状态。而在储物空间上，乐道 L80 配备了容积为 240 升的电动前备舱，能够轻松容纳两个 20 英寸的标准登机箱、一个儿童专属行李箱及一个随身背包。前备舱的存在不仅增加了绝对装载量，更为家庭用户提供了干湿分离与物理分区的储物区域，有效改善了传统 SUV 后备舱容易被杂物堆砌得杂乱无章的痛点。与前舱相对应，乐道 L80 的后备舱在第二排座椅处于常规直立状态时，容积为 1200 升；当第二排座椅完全放倒后，整体空间可扩展至 2600 升。此外，在后备舱的常规盖板下方，工程团队设计了双沉舱结构。其中外侧沉舱容积为 106 升，里侧沉舱容积为 63 升，均配备了独立盖板。这种隐蔽式的储物空间非常适合用来存放不常用的随车工具、零散的露营小配件、充电线缆以及各类清洁用品，使得主后备舱始终保持整洁。乐道能够在车辆内部极其有限的机械空间内挤出更多的储物容积，主要得益于车辆底层平台的极高集成度。L80 充分应用了已经在 L90 上被验证过的 900V 高压架构，并配备了体积更为紧凑的前置电驱动系统。结合 30.5 度大倾角冷却模块、全栈自研的热管理高度集成模块以及一体式偏置空调箱，这些系统级的技术创新对原本需要占据前舱大量体积的机械与电子部件进行了极致的压缩与整合。空间之外，座舱内的舒适化和智能化，也是乐道 L80 强调的特色。乐道 L80 的座舱内部采用了分层布局：上层主要用于视觉延伸与氛围营造，中层集中处理人机交互与触觉反馈，下层则侧重于日常储物与抗污耐用。车内配备了支持 1600 万色调节的氛围灯系统、长达 7.2 米的环绕光带，辅以大面积软质材料包覆、金属扬声器网罩及精细的绗缝与刺绣工艺。整体风格规避了盲目堆砌昂贵材料的浮夸感，更侧重于在家庭用车的基调上打磨出温馨且不失格调的精致质感。为了提升车内的通透感，乐道 L80 车辆侧窗与前挡风玻璃的总面积约为 3.3 平方米，前后分区的全景天幕面积也达到了 1.1 平方米。同时为了应对大面积玻璃材质在强光照射下带来的隔热与防晒问题，车内配备了双电动物理遮阳帘，乐道 L80 同时配备了空气悬架系统，支持在驻车时自动降低车身高度，配合电动侧踏板与静音电吸门，优化了行动不便的老人或儿童的上下车体验。车辆的座椅总成基于专属的高端座椅平台开发，依据中国人体生理曲线大数据进行了型面贴合度优化，前排座椅支持 12 向电动调节，第二排支持大角度电动调节。同时，方向盘加热、前后排座椅加热、通风以及按摩功能均作为标配提供。此外，车内集成了一台 8.8 升容量的智能电动冰箱，足以容纳 6 罐常规可乐、2 盒鲜牛奶以及部分水果零食。神玑上车，4D 领航数字化交互与计算平台是现代智能汽车的核心大脑。L80 的屏幕矩阵由一块 35 英寸的 AR-HUD 增强现实抬头显示、一块 17.2 英寸的 3K 高清中控屏、一块 17.3 英寸的后排吸顶娱乐屏以及一块 8 英寸的后排专属场景控制屏组成。支撑这一系列复杂交互与协同调度的硬件，是一颗采用 5 纳米制程工艺的车规级高阶智驾芯片——神玑 NX9031。该芯片拥有高达 546 GB/s 的内存带宽，打破了传统分散式域控制器的物理壁垒，能够同时支持智能驾驶、智能座舱、智能底盘等多个系统跨域计算资源的统一调度。配合 SkyOS·天枢整车全域操作系统，车辆从底层代码级别打通了智驾、座舱、底盘、车身及动力网络，使各项功能摆脱了信息孤岛的状态，能够基于场景进行深度系统级协同。在智能驾驶领域，系统深度接入了蔚来世界模型 NWM（NIO World Model）。该模型摒弃了传统的规则驱动逻辑，转而通过强化学习与海量的闭环仿真训练，提升了系统对复杂未知路况的理解与博弈能力，全面覆盖城区复杂道路、城市快速路、高速公路及各类园区停车场。值得重点关注的是，L80 支持城区领航换电功能（PSP）。在导航状态下，系统能够自动规划最优补能路径，将车辆自主驾驶至换电站并完成自动化换电流程。同时，基于模型化架构构建的智能泊车系统，也支持路线的实时动态规划和对各类非标准复杂车位的识别入位。尽管定位为体型庞大的家庭 SUV，L80 在底盘的机械素质与动态调校上依然投入了极高的规格。悬架系统采用了前双叉臂与后五连杆独立悬架组合，配备了具备 100 毫米行程调节能力的智能空气悬架系统以及 CDC 智能连续可变阻尼减震器。底盘硬件与路面之间通过液压衬套进行柔性连接，并搭配了抓地力更强的鸳鸯胎。乐道 L80 还支持 4D 全域舒适领航功能。该功能能够实时结合车辆自身的传感器感知、激光雷达对前方路面的三维预瞄以及云端高精舒适图层数据，在遇到减速带、坑洼或颠簸路面之前，提前毫秒级主动调节空气弹簧的刚度与减震器的阻尼，从容化解路面冲击。除了常规的运动、舒适、后排舒享和节能模式外，车辆预设了湿地、雪地、泥地、沙地等脱困模式。在卓越的底盘参数加持下，L80 拥有高达 600 毫米的最大涉水深度，最小离地间隙最高可拉升至 260 毫米，足以轻松应对绝大多数轻越野和长途自驾游的复杂路况。而在安全方面，乐道 L80 全系标配 9 个安全气囊，除了常规的环绕式防护，第一排还配备了能在碰撞发生前预紧的主动式安全带、容积达到 137 升的狐耳式超大副驾气囊以及防止前排乘员相互碰撞的中央远端安全气囊。基于家庭属性的考量，第二排乘客的安全防护被提升至与前排同等重要的层级，配备了 14 升容积的后排侧气囊，以及长度贯穿 2920 毫米、容积高达 56 升的超大侧面安全气帘。在应对碰撞的白车身结构设计上，L80 的前部预留了长达 650 毫米的充足溃缩吸能空间，并创新性地采用了「6 纵 10 横」的笼式安全传导结构。前防撞梁选用了高强度的 7 系航空级铝合金材质，其上层主防撞梁的宽度覆盖率达到了整车宽度的 80%，能够在严苛的偏置碰撞中有效化解破坏性冲击力。侧面防护则通过环舱高强度钢、一体式门环、多腔体铝合金门槛梁以及坚固的电池舱横梁共同构建。车尾部分应用了先进的一体压铸后地板技术，配合铝合金后吸能盒与后防撞梁，为车身尾部提供了深达 1244 毫米的吸能缓冲空间，最大程度减轻高速追尾事故对后排乘客及底层电池包的结构性伤害。乐道 L80 搭载的电机能够输出 340kW 的峰值功率，百公里加速时间 5.7 秒，其百公里综合电耗被控制在优秀的 14.3kWh，85 度电池下的 CLTC 续航为 615km。当然，纯电车型的体验永远无法脱离补能体系的建设。共享蔚来庞大且成熟的充换电网络，依然是乐道 L80 在同级别竞品中最难以被复制的核心优势。为了进一步夯实这一护城河，乐道官方已确认追加投入 6 亿元人民币，用于在网络中新增 8000 多块流通电池。蔚来之前透露，早在四年前，乐道 L80 便与 L90 基于同一技术平台同步启动开发。其中，乐道 L90 的成功已经验证了产品定义的精准性。在 2025 年，乐道 L90 凭借对大三排 SUV「满载状态下行李装载空间」这一核心痛点的有效解决，在上市短短 5 个月内实现交付超 4.3 万台，成为纯电大型 SUV 市场的年度销量冠军。这不仅助力乐道品牌在 2025 年实现 107808 台的全年交付目标，更以接近 24 万元的成交均价稳居新势力品牌第一梯队。乐道 L90 和蔚来 ES8 的表现，也被行业视为大型 SUV 动力形式从增程向纯电转型的关键标志。如果说乐道 L90 成功开拓了大三排纯电 SUV 市场，那么乐道 L80 则肩负着在规模更大的大五座 SUV 领域实现突破的重任。在乐道的产品价值体系中，L90 聚焦于「在一起」的温情与责任，通过三排布局满足多孩、三代同堂家庭的聚会与通勤需求；而 L80 则更侧重于「走出去」的激情与成长，以更灵活的空间形态承载用户自驾、露营等更具探索性的生活方式。

🔍郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%🤖OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略🚀Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI📱微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆💻华为 Mate X7 官方降价 1000 元⚠️追觅造车主体约 232 万元股权遭冻结🎮国行 Switch 正式停服🧠百度成立模型委员会 BMC💰三星电子单季营收创历史纪录，半导体撑起超九成营业利润🔍海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%🖨️铠侠：2027 年内存仍吃紧、净利润同比增长 103.6%🤖微信发布「全球青少年 AI+ 小程序」洞察报告💡何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了🦿5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约📱荣耀 CEO 李健官宣机器人手机今年 Q3 发布📱vivo S60 将首发 4K 原生感 Live🚗全新理想 L9 发布：45.98 万元起🤖ChatGPT 新增 Codex 连接电脑功能📰 周末也值得一看的新闻郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%天风国际证券分析师郭明錤昨日披露，苹果已在 Intel 18A-P 系列制程上为低阶及旧款 iPhone、iPad 与 Mac 处理器立项，订单中 iPhone 约占 80%。流片规划显示，苹果预计 2027 年在 Intel 放量，2028 年续增，2029 年进入衰退期。Intel 方面，2027 年良率目标为稳定达到 50% 至 60% 以上，量产时程与出货规模目前仍未明朗。不过，郭明錤也指出，即便 Intel 顺利出货，台积电在苹果供应体系中仍将占 90% 以上比重。郭明錤认为，苹果此举是有意将 Intel 培养为长期备选供应商。AI 对台积电先进制程资源的占用将持续扩大，苹果需要在议价能力尚存时提前布局。与此同时，苹果也在同步评估 Intel 其他先进制程技术。OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略据 WIRED 报道，OpenAI 昨日向员工宣布新一轮组织调整，联合创始人兼总裁 Greg Brockman 正式接管产品战略。报道指出，OpenAI 将 ChatGPT、编程智能体 Codex 及面向开发者的 API 业务整合进同一核心产品团队，目标是将三者合并为「统一的体验」。Codex 原负责人 Thibault Sottiaux 升任核心产品与平台主管，统管消费、企业和开发者三条线；ChatGPT 原负责人 Nick Turley 转任企业产品改版；前 Instagram 副总裁 Ashley Alexander 接任消费产品负责人。Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI据《金融时报》报道，Anthropic 已就一轮 300 亿美元的融资达成条款，融资完成后公司估值将约达 9000 亿美元，较今年 2 月初 3500 亿美元的估值接近翻三倍。本轮融资由 Dragoneer Investment Group、Greenoaks Capital、红杉资本（Sequoia Capital）和 Altimeter Capital 联合领投，报道称，每家领投机构预计出资不低于 20 亿美元。微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆CCF 计算机博物馆（CCM）昨日收到腾讯微信团队捐赠的一批早期历史实物，包括微信第一行后台代码原始存档、早期开发设备 iMac ME086CH/21 寸一体机与戴尔塔式工作站，以及 Apple iPod Touch 3、诺基亚 5800XM 等早期测试终端。CCM 表示，由于移动产品迭代速度极快、数字资料易被覆盖删除，大量见证移动互联网起步的实物未能得到系统保护。此次捐赠呈现了微信从 0 到 1 的研发链条，是中国移动互联网早期发展的核心实物见证。CCM 主体建筑已全面落成，拟于今年 9 月开馆，目前持续面向社会征集各时期计算机相关藏品。保时捷称被中国车企抄袭「有些许无奈」，罗永浩：保时捷自己就有前科保时捷中国传媒公关副总裁唐风靓日前在浙江「易境思论坛」上表示，过去几年保时捷激发了高度统一的汽车设计审美，被中国车企模仿「除了荣幸也有些许无奈」。罗永浩随后在微博发文回应，称保时捷说「无奈」可以理解，但称中国车企抄袭得「高度统一」则是吹牛。他同时指出，保时捷历史上也曾抄袭捷克斯洛伐克品牌太脱拉（Tatra）的设计，并以 1965 年大众汽车向太脱拉支付 100 万西德马克庭外和解赔偿金为据，直言「说到流氓程度，谁能跟保时捷比」。此番言论随即引发讨论。有网友指其犯了「诉诸伪善」的逻辑错误，罗永浩强硬回怼，称自己已明确承认抄袭者「也是流氓」，并非为抄袭行为开脱。华为 Mate X7 官方降价 1000 元华为昨日宣布，旗下两款折叠屏手机同步开启降价优惠：Mate X7 全系直降 1000 元，优惠后售价 11999 元起；Mate X6 直降 3000 元，优惠后售价 9999 元起。追觅造车主体约 232 万元股权遭冻结企查查 App 显示，追觅科技创始人俞浩实控的星空计划（上海）汽车科技有限责任公司近日新增一则股权冻结信息，股权数额约为 232.25 万元，股权所在企业为上海星空计划进出口贸易有限责任公司。去年 8 月，追觅科技宣布进入汽车领域。截至目前，追觅汽车团队规模约千人，研发人员占比达 70%，自研覆盖智能汽车多个核心领域。此次造车主体出现股权冻结，官方尚未就相关情况作出回应。国行 Switch 正式停服国行 Switch 的 Nintendo e 商店已于昨晚正式停服，至此，国行 Switch 将无法再进行联网对战或下载游戏。官方在停服前向相关玩家发送提醒，建议玩家在停服前完成所有已购游戏、DLC 及兑换码的下载与激活；删除游戏数据或对主机进行格式化后，数字内容将无法恢复。今年 3 月 31 日，国行 Switch 的游戏售卖服务已率先停止；而根据官方此前公告，今年 12 月 31 日起，国行 Switch 主机及配件的维修服务也将全面终止。这意味着，国行 Switch 的官方售后支持将在今年底彻底画上句号。百度成立模型委员会 BMC据《科创板日报》报道，百度近日内部设立百度模型委员会（Baidu Model Committee，BMC），成员由对大模型有深刻理解的年轻研究员构成。基础模型研发部（BMU）与应用模型研发部（AMU）均向该委员会汇报。知情人士透露，新成立的 BMC 将全面统筹百度大模型业务，覆盖路径探索、技术规划、模型训练到业务落地的完整链条，目标是实现模型从技术到应用的一体化管理。百度方面表示，此举有助于集中优势资源，持续建设并强化人工智能方向的技术优势。三星电子单季营收创历史纪录，半导体撑起超九成营业利润三星电子昨日发布 2026 财年第一季度（1 月 1 日至 3 月 31 日）业绩报告：期内实现合并营收 133.9 万亿韩元，同比增长 69%，环比增长 43%；营业利润 57.2 万亿韩元，同比增长 756%，环比增长 185%，营业利润率达 42.8%。净利润为 47.2 万亿韩元，净利率为 35.3%。基本每股收益为 7123 韩元。半导体（DS）：营收 81.7 万亿韩元，同比增长 225%，环比增长 86%；营业利润 53.7 万亿韩元。其中存储业务营收 74.8 万亿韩元，同比增长 292%，环比增长 101%，为本季度核心增长引擎；设备体验（DX，含手机与消费电子）：营收 52.7 万亿韩元，同比增长 2%；营业利润 3.0 万亿韩元，同比下降约 1.8 万亿韩元；显示（SDC）：营收 6.7 万亿韩元，同比增长 14%；营业利润 0.4 万亿韩元，同比下降 0.1 万亿韩元；Harman：营收 3.8 万亿韩元，同比增长 12%；营业利润 0.2 万亿韩元。下季度展望，公司预计 AI 基础设施扩张驱动的服务器内存及 SSD 需求将持续强劲，计划向客户交付 HBM4E 首批样品以巩固技术领先地位；手机业务方面，Galaxy S26 系列发布效应趋于消退，盈利能力预计承压。海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%昨天，海力士于发布今年第一季度（截至 3 月 31 日）业绩，按 K-IFRS 口径，期内实现营收 52.58 万亿韩元，同比增长 198%；营业利润 37.61 万亿韩元，同比增长 405%；净利润 40.35 万亿韩元，同比增长 398%。本季度营收首次单季突破 50 万亿韩元，创历史新高。主要财务指标如下：毛利润 41.68 万亿韩元，毛利率 79%，同比提升 22 个百分点；营业利润率 72%，创历史单季新高，环比提升 14 个百分点；EBITDA 41.34 万亿韩元，EBITDA 利润率 79%，同比增长 284%；净利润率 77%，其中非经营利润 14.01 万亿韩元，包含投资资产估值收益 9.94 万亿韩元。铠侠：2027 年内存仍吃紧、净利润净同比增长 103.6%铠侠控股昨日发布截至今年 3 月 31 日止 2026 财年业绩，期内实现营收 2.3376 万亿日元，同比增长 37.0%；按 IFRS 口径，归母净利润为 5544.9 亿日元，同比增长 103.6%。SSD 与存储业务营收 1.3626 万亿日元，同比增加 37.5% 至 3715 亿日元；智能设备业务营收 7599.8 亿日元，同比增加 51.6% 至 2588 亿日元；其他业务（含 SD 卡、U 盘等零售产品及向闪迪集团的销售）营收 2150.1 亿日元，同比基本持平。另据财联社报道，铠侠表示，AI 服务器客户的订单量持续走强，以美元计价的 NAND 价格在一季度上涨逾一倍，并预计 2027 年度 NAND 存储供应仍将维持紧张态势。微信发布「全球青少年 AI+ 小程序」洞察报告昨天，腾讯研究院联合微信发布《全球青少年 AI+小程序洞察报告》，披露了全球基础教育阶段群体利用 AI 工具与小程序进行应用开发的各项核心数据。据报告数据显示，自 2023 年 11 月微信小程序教育平台正式对外开放以来，已吸引全球 7954 所学校的逾 8.7 万名学生参与，累计创建小程序项目超 28.7 万个。在开发流程中，师生调用 AI 辅助创作每年消耗 Token 量超过 500 亿，支撑约 375 万次人机交互。此外，微信连续四年举办的小程序全球创新挑战赛共产生 7430 个作品，其中第四届单届提交量达 4900 个，呈显著爆发趋势。开发者画像层面，62% 的项目由单人独立完成，最小获奖开发者年仅 5 岁。地域活跃度数据显示，中国香港、北京及中国澳门位列前三，且项目协作已出现跨国别开发特征。在应用方向上，05 后开发者主要聚焦于无障碍辅助、碳足迹、适老化关怀及健康管理等社会实用场景。💡 何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了据睿见 Economy 报道，小鹏集团董事长、CEO 何小鹏昨日在「第十八届轩辕汽车蓝皮书论坛」上表示，「什么时候我们看到北京、上海车展上不再有 150 款新车的时候，中国汽车的内卷可能就结束了。」何小鹏表示，过去十年，中国新能源汽车经历了两个重要转变：一是核心驱动能源从传统化石能源切换至电化学能源；二是行业全面进入初步普及的智能化阶段。他称，十年前几乎没有人相信中国汽车能在全球市场取得如今的成绩，并实现规模化的全球化布局。何小鹏认为，下一个十年将是电化学能源与数据能源的双重叠加阶段，芯片和模型将成为新的「引擎」，汽车与机器人的能力将在这十年内深度整合。5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约ROG 与 XREAL 联名推出的电竞 AR 眼

直到 2026 年五月，我们终于等到一整轮的影像旗舰登场完毕——小米做出了徕卡一瞬、vivo 打造了照片与视频兼具的 V 单、OPPO 在双长焦方案上带来了惊喜。在这轮旗舰更新潮的最后，华为提前将新一轮的 Pura 90 Pro Max 端上了桌。老实说，第一眼看硬件，心里是有落差的。去年 Pura 80 Ultra 上颇为惊艳的「一镜双目」结构不见了，取而代之的，是目前行业内最稳妥也最常见的大底长焦方案。不过，在细细体验后，我认为 Pura 90 Pro Max 藏了一个比纯拼硬件更有意思的新解法——XMAGE 智拍。全自动，都省心过去这几年，我系统性地梳理过华米 OV 的影像架构。虽然各家在硬件选择上各有理解，但大方向出奇一致，都在极其有限的机身空间里死磕传感器面积，比拼谁的光圈更大、进光量更多。在硬件上奋发图强理所应当。但在摄影中，不错的硬件基础仅仅是第一步。想要拍出一张好照片，构图与后期同样重要——构图保证一张照片在视觉上是否平衡、主体是否明确；而后期则决定了这张照片视觉质感和风格化。但这两者，向来都是具备门槛的手艺。而华为在 Pura 90 Pro Max 上，打算一口气解决这两个问题。熟悉他们产品线的朋友或许有印象，去年发布的 Pura 80 Ultra 上，华为试水了 AI 辅助构图。AI 辅助构图功能相对基础，在实际拍摄中主要起一个引导和提醒的作用。到了 Pura 90 Pro Max，这项技术完成了从「辅助」到「主导」的蜕变，并定名为「XMAGE 智拍」。剥开表层去探究 XMAGE 智拍的底层逻辑，是一套完全由端侧算力驱动的本地化影像流水线，将拍摄过程拆解为三个同时进行的模块——基于实时语义分割的自动构图、无视物理镜头限制的智能变焦，以及针对画面的色彩重塑。也就是说，在这个功能的加持下，Pura 90 Pro Max 可以自动识别主体、变焦构图，最后还排列出一系列适合当前场景的 XMAGE 风格供你选择。值得一提的是，XMAGE 智拍不需要网络连接，也无需将庞大的图像数据传回云端，全部基于端侧 AI 运转，实时识别画面里的主体轮廓、建筑线条以及光源分布状况，随后根据底层训练好的审美模型，给出一套包含最优构图、变焦和精细色彩方案的组合拳。繁复的专业参数在这里黑盒化，过去需要人脑去判断的构图和调色工作，全权交给了本地算法。除此之外，Pura 90 Pro Max 还顺势推出了 AI 姿势推荐，这个功能同样通过语义分割识别，可以理解画面中的了恩物、环境、姿势与背景，随后直接在取景框中勾勒出一个等比例的线框轮廓。拿手机的人只需按图索骥，引导模特贴合画面上的姿势即可。遇到不满意的动作能随时刷新，它甚至支持导入社交网络上保存的样片，一键提取动作精髓。听起来确实很厉害，但这套完全依赖端侧计算跑出来的流水线，究竟能不能应对现实世界里复杂多变的光线，还需要把它带到真正的街头去寻找答案。AI 有所能，也有所不能来到一家咖啡馆，墙上的多层光环灯饰引起了我的注意。举起手机，XMAGE 智拍自动变焦到 2×，将墙壁上的光环作为画面主体。这个选择可以说中规中矩，光环圆形处于画面中下方，整体构图稳定和谐。同时，操作台上颇具金属质感的咖啡机也是个不错的主体，XMAGE 智拍精准识别到我的拍摄意图，变焦到 192mm，保持咖啡机的中置。由于我想保留更为还原的室内光线与色彩，这张照片并没有套用 XMAGE 智拍推荐给我的任何色彩风格。用它拍了几组照片后，我认为 XMAGE 智拍是一个非常用户视角的功能——在传统的拍摄习惯里，手机焦段是一件需要精打细算的事。大脑会本能地去贴合物理镜头的原生焦段，生怕落在中间焦段会损失画质。但 XMAGE 智拍完全抛弃了执念。它不在乎用的是几倍变焦，转而完全为画面考量：当前这个画面，怎么裁切才最好看。哪怕不可避免地要损失一部分边缘像素，只要能让视觉中心突出、比例和谐，它就会果断动手。这种做法很实在，一切都在为了最终的出片服务。走出室外回望咖啡馆，绿植与橙色外墙交相呼应，XMAGE 智拍自动截取出色彩冲击力最强的部分，同时准确识别出咖啡馆建筑主体，两种大色块在画幅中占比相似，平衡而美观。再抬头，夏季树木繁茂，XMAGE 轻微放大焦段，让树枝形成天然的框架构图，广州塔正悬于画面正中间。在这组照片的拍摄中，我还察觉到了一个很有意思的技术细节。按下快门时，你会发现滤镜加载的速度，远大于构图改变的速度。这是因为滤镜往往只需识别画面后套用底层的色彩映射，而构图则需要端侧算力去实时识别画面里的建筑线条、光源分布和人物轮廓。计算量完全不在一个量级。更有意思的是，智拍套用滤镜的方式非常精细，除了 XMAGE 风格选择外，还会对调色盘做精细调整，确保风格对当前环境的适配度。当然，这套系统也有自己的脾气。在实际走街串巷时，XMAGE 智拍有时候会自动跳出来接管画面，有时候又毫无动静，需要你手动去点击唤醒。就算是手动唤醒的情况下，XMAGE 智拍也有几率出现风格已经选好，但当前画面无法找到最佳构图的情况。这种情况频繁出现在极其繁杂的环境中，由于元素过于混乱、信息量爆炸，算法的确很难从中剥离出合适的构图。当然，换个角度想想，也许此时的 1× 视角，就是当前环境的最佳构图了。万物皆可 AI，现在轮到摄影了185 年前，达盖尔发明银版摄影法，留住时间成为一种特权；2011 年，胶片巨头柯达传出破产消息，智能手机开始野蛮生长，把镜头塞进了普通人的口袋。回顾整部影像史，其实就是一部打破特权、技术走向大众的普惠史。到了今天，移动影像的狂飙突进，撞上了一堵无形的墙。手机镜头模组越做越大，与之对应的，是最近行业里一个颇具深意的传闻：受制于高昂成本与物理极限，下一代影像旗舰的「超大杯」，大多要面临停更了。这条路为什么走不通了？因为影像旗舰突飞猛进的背后，藏着一个被刻意回避的死结。决定照片质量的，永远是镜头后面的脑袋。普通人有懂美的眼睛，却跨不过光圈快门和构图比例的门槛。把一台堆满顶级硬件的手机，递给一个毫无基础的人，他按下的快门，大概率得到一张平平无奇的随手拍。顶级硬件如果只服务于一小撮懂摄影的人，无疑就会变成一种伪命题。在同行集体踩下刹车的时间点，华为提前把 Pura 90 Pro Max 端上了桌，而 XMAGE 智拍，就是破局的方法。在此之前，行业里已有不少 AI 落地，比如 vivo 让四季流转，OPPO 消除反光。只不过，它们大多是按下快门后的后期创意和修补。XMAGE 智拍往前跨了一大步，让算法成为整个拍摄动作的主导。这种做法谈不上是什么艺术层面的颠覆，但确实管用。降低门槛，让完全没有摄影基础的人，也能相对轻松地把一台旗舰机用出该有的样子。好的技术绝不该是一座孤岛。曾经的智能手机让人们拥有了「随时拍」的自由，现在的 AI 则进一步赋予了大众「拍得好」的能力。这或许是打破当前僵局的一剂良药，也是影像超大杯物尽其用的方法。

过去几年，20 万级新能源 SUV 市场像一条突然变宽的河道。一边是家庭用户对空间、续航和舒适性的刚性需求，另一边是智能化、性能化配置不断下放。过去很多只会出现在 30 万元甚至更高价位车型上的配置，如今开始像雨水一样往下渗透。激光雷达、空气悬架、大模型智驾、零重力座椅，这些原本带有「高配」意味的词，正在成为新一轮价格带竞争里的筹码。肩负着广汽集团高端化重任的昊铂品牌在今年广汽集团重振旗鼓的趋势下，也顺势重新梳理产品架构，昊铂 A800 旗舰轿车前段时间开始上市，用到了乾崑智驾和鸿蒙座舱来增强产品竞争力。紧接着，面向 20 万价位带的 SUV 昊铂 S600 也来了，提供纯电 660 Max、纯电 660 Max+、增程 230 Max、增程 215 四驱 Ultra+ 四个版本，官方预售价为 20.99 万至 22.99 万元，预售权益价为 18.89 万至 20.89 万元。新车定位为「新豪华智慧运动 SUV」，提供纯电和增程两种动力形式，车身尺寸为 5015×1933×1700mm，轴距 2936mm。相比于品牌初创时期的雄心勃勃，昊铂现在显然务实不少，它试图把昊铂 S600 做成一个配置密度很高的样本：性能、智驾、座舱、舒适、安全和权益，几乎都被塞进「满配」这个关键词里。昊铂 S600 请了 2 个代言人，一个代言「优雅」，一个代言「运动」过去提到中大型 SUV，很多人的第一反应是「稳」：空间要大，坐姿要高，后排要舒服，最好还能承担家庭出行里那些琐碎但真实的场景。比如周五晚上从公司直接出发，后备厢里塞进折叠椅、露营桌、孩子的滑板车和一箱零食；又或者长途返乡时，后排老人希望能把座椅放倒一点，孩子希望看动画，副驾则希望在高速上补一觉。对于这类用户来说，车首先不是一个「驾驶机器」，而是移动客厅、临时休息室，也是家庭生活的第二个容器。昊铂 S600 没有放弃这部分的诉求，或者说，这些配置已经在国产中高端 SUV 上成为一种基本素养，昊铂 S600 拥有 90% 得房率、750mm 二排腿部空间、5.316 平方米整车玻璃面积，以及 1292L 二排拓展空间；舒适配置上，同级唯一 18 点按摩零重力女王副驾、143° 躺角后排电调座椅、22 扬 ADiGO SOUND 音响、五块双层夹胶静音玻璃等配置也都被纳入其中。但昊铂 S600 又没有把自己完全包装成一台「沙发车」。它的另一面，是更强调姿态和操控的运动属性，所以昊铂又反复出现「跑山王」这个表达，显然昊铂希望把一台五米多长的大尺寸 SUV 做得不那么笨重，让它不只是适合城市通勤和家庭旅行，也能在山路、弯道、高速超车等场景中保留一定的驾驶参与感。从造型上看，昊铂 S600 采用的是一种前低后高的轿跑溜背姿态，超过 5 米的车长和近 3 米轴距提供了视觉体量，双层鸭尾、宽体车尾、20 英寸运动轮毂等设计元素，则试图削弱中大型 SUV 常见的厚重感，从侧面看它确实不像一般的家用大型 SUV 那样笨重方正，更像是专为年轻家庭所设计。所以这一次昊铂也请了 2 个代言人，英格兰传奇足球明星迈克尔·欧文作为「运动代言人」，知名演员陈数作为「优雅代言人」，两位代言人分别代表这一款车的一种特质，也是试图同时吸引家用取向和个人运动取向的潜在消费者。把四驱、高阶辅助驾驶、双腔空悬都放进 20 万级别在最高配的增程 215 四驱 Ultra+ 版本上，昊铂 S600 同时提供双电机四驱和双腔空气悬架。昊铂表示，这是行业首次在这一价位段同时放入双电机四驱和双腔空气悬架。另外，昊铂 S600 还全系标配前双叉臂后五连杆独立悬架、SDC 可变阻尼减振器和鹰爪系统 2.0，并提供雪地、湿滑、脱困三种路面辅助模式。一台大 SUV 在城市高架上并线，需要车身跟随方向盘动作，而不是像大船一样慢半拍；周末跑山时，连续弯道考验的是侧倾控制和车身收敛；雨天压过积水路面，轮胎抓地力和车身稳定系统要在驾驶者还没反应过来之前，先把车身姿态拉回来。在这些场景下，双腔空悬能够有效提升驾驶体验和驾驶信心。动力部分，昊铂 S600 提供纯电和增程双动力。其中增程四驱版综合功率达到 370kW，满电零百加速 4.3 秒，亏电零百加速 5.6 秒。广汽星源增程技术的持续发电功率为 85kW，官方还提到新车在亏电状态下可实现 165km/h 速度持续刷圈 2 小时，亏电极速 193km/h。增程车过去最容易被质疑的地方，在于「满电一条龙，亏电一条虫」，电量充足时像电车，电量不足后则容易出现动力衰减、油耗升高、发动机噪音变大等问题。昊铂 S600 对 85kW 持续发电功率和亏电性能的强调，其实是在回应这个痛点，一台增程 SUV 不能只在发布会数据里快，也要在长途、高速、爬坡和亏电场景下保持基本的体面。在智能化层面，昊铂 S600 的策略也比较直接：尽量把高阶配置做成全系标配。新车全系标配激光雷达、高通 SA8650 智驾芯片、27 个高精度感知单元，并搭载融合 Momenta R6 强化学习大模型的广汽星灵智行 ADiGO GSD 3.0，支持城区和高速快速路场景。资料还提到，昊铂 S600 下半年将首批升级 Momenta R7 强化学习世界模型。智能座舱方面，新车搭载高通 8295P 座舱芯片、17.3 英寸 3K 中控屏、27 英寸 W-HUD，并内置 11 个 AI 智能体。相比过去车机只负责导航、音乐和语音控制，现在的座舱正在变得更像一个移动终端：它不仅要响应指令，还要理解场景，甚至提前知道用户在车里可能需要什么。下班后坐进车里，导航自动给出回家路线，HUD 把车道级信息投到视线前方，孩子坐在后排看动画，车机根据车内温度和空气质量调整新风，座椅按摩开始工作……这些在新能源汽车上可以实现的体验，对于从 20 万元档位传统燃油车换购过来的用户来说，无疑是跃升式的。广汽集团有着庞大的用户基本盘，这些人群想要增购换购新能源车型的话，广汽自然希望用产品承接他们的需求。就像 AI 说的那样，我在这里，不偏不绕，不躲不逃，稳稳接住你。某种程度上来说，过去几年广汽集团新能源产品的建设，过于重视 B 端，对 C 端用户的吸引力不够，并没有「稳稳接住」用户的需求。对消费者来说，20 万级新能源 SUV 已经足够拥挤，智己 LS6、极氪 7X、小鹏 G7 等车型都在这个区间形成压力。昊铂 S600 要突围，关键命题并不是「它是不是拥有很多配置」，而是这些配置能否被组织成一种清晰的用车体验被感知到：城市里安静从容，山路上不显笨重，长途时不焦虑，家人坐进去愿意放松，驾驶者开起来也不觉得无聊。

龙虾之父一个月消耗 6030 亿 Token，总花费金额高达九百万人民币。移动联通电信，三大运营商都在推 Token 套餐，199 送千兆宽带还有 1 亿 Token，了解一下？从硅谷到国内大厂，Tokenmaxxing 成为公司的主流，谁消耗 Token 多，谁就是 AI 时代的好员工。00 后校友向母校捐赠 20 亿 Token，被网友调侃按 DeepSeek 5 元/亿 Token 计算，只要 100 元。▲图片来自新浪财经Token 在半年内完成了一次身份跃迁：从技术术语，到 KPI，到话费套餐，到捐赠货币。它成了 AI 时代的「度量衡」，唯一的问题是，没人说得清它到底在度量什么。我们自己买 Token，用公司的 Token，部署了一堆 Agent，代码、论文、周报都是 Token 烧出来的。而另一边是，大厂的员工由于 Token 消耗排行榜的原因，开始拿着公司的 Token 处理私事、玩游戏、开发数十个没什么用的子 Agent 来提升自己的排名。「回报」这件事很难量化，但「使用量」可以量化。于是所有人都选择了那个容易量化的东西。这不是 AI 时代的新问题，这是管理学的老病。用 AI 消灭狗屁工作的公司，正在制造新型狗屁工作亚马逊，那个裁员裁到大动脉，把自己的网站都变成 404 的小狗，最近又被爆出了新的「笑料」。原本被寄予厚望、用来消灭「狗屁工作」的 AI，最终却沦为制造新型「狗屁工作」的源泉。据《金融时报》报道，为了逼迫员工拥抱 AI，亚马逊搞出了一个极其复古的管理手段：「Token 消耗排行榜」，追踪每个员工的用量。公司强制要求超过 80% 的开发者每周必须完成 AI 使用指标，甚至将消耗 Token 的数量作为考核标准。▲图片

1.OpenAI 的两大宿敌 Anthropic 和马斯克，放下心中成见之后终于在月初结盟了。在此之前，Anthropic 和马斯克的关系并不融洽：今年 2 月，马斯克还在自己的 X 账号指责 A 社「woke」「邪恶」「反人类」（misanthropic），说这家公司「仇视文明」。事后来看，这次攻击并非马斯克清新脱俗的性格使然，而是 Anthropic 所做的某些事情触碰到他的神经，事出有因。在此之前，xAI 内部使用 Cursor 工作，但是今年年初员工发现，Claude 模型突然在 xAI 的 Cursor 公司账号里不能使用了。当时还在 xAI 上班的联合创始人吴宇怀，在全员信里是这么说的：「Anthropic 更新了政策，要求 Cursor 不得向其主要竞争对手提供 Claude 模型调用能力。」当时，吴宇怀在信中写了一句话，颇为有趣：「这是坏消息也是好消息。我们的生产力会被影响，但这也敦促我们开发自己的编码产品和模型。」为什么当时 xAI 的高层认为，开发自己的编码产品是关键？后来发生的事情，大家都知道了。xAI 的联创团队悉数跑路，马斯克一气之下对 Cursor 使用了钞能力必杀：上个月底，SpaceX 和 Cursor 共同宣布，将在编程和知识类工作 AI 模型的训练上，展开前所未有的战略合作；并且，SpaceX 还获得了以 600 亿美元收购 Cursor 的权利，或向后者支付 100 亿美元合作费用。注意编程这个关键定语，后面还会 call back.2.最近，我看了一条 Cursor 早期投资人、Anthropic 大喷子、T3 创始人 Theo Browne 的视频。本来点进去是看他喷 A 社和 SpaceX 怎么蝇营狗苟，结果没想到，却看到了关于 SpaceX + Cursor 合作的，一个既另类却又极度合理的分析：不说 600 亿的收购，就只说 100 亿的合作费——Theo 在视频里表示，自己认为「哪怕只是交换到 Cursor 的用户数据，这 100 亿也值回票价了。」所以是什么数据？如果你也去看 Theo 这条视频，他会讲得非常清楚。但为了节约时间，我们在这里简单概括一下：我们和 AI 的对话是一来一回的，你提出问题/需求，他给你解答；coding agent 同理，只不过返回的是代码。一次高质量的对话，整个过程，包括用户提示、模型思考、agent 规划、输出代码、验证——所有这些东西合起来，可以称为一个完整的 Agentic Loop——就成为了高价值的训练数据，再喂给模型去进行强化学习，就能进一步提高模型在实战场景下的表现水准。Cursor 有的，SpaceX 想要的，就是这些数据。可这些数据从哪里来呢？答案很简单：作为模型厂商，这种高质量数据的最直接来源，只能是你自己开发的 coding agent 产品——也即 Anthropic 的 Claude Code、OpenAI 的 Codex、Kimi 的 Kimi Code。现在你应该明白了，为什么被 Anthropic「封号」之后，吴宇怀会在全员信里提出开发 xAI 自己的 coding 产品和模型这件事了。这件事 xAI 在当时已经看清楚了：没有自己的编码产品，就没有高质量的强化学习数据；没有高质量的数据，就训练不出真正实战能力强的 coding 模型。虽然有点暴论，但现在我们可以点题了：模型厂商想做出来真正能打的编程模型，做自己的 coding agent 产品是唯一的路径。3.大语言模型像个水晶球，用全网的语料训练出来，似乎能够解答万物，但并不代表它在所有问题上都能给出高质量的答案。用 GitHub 上数以亿计的代码条目训练，当然也能训练出 coding 模型。这是「学习结果」的逻辑，也是没问题的。毕竟编码任务的结果是可以验证的：代码能不能运行，测试能否通过，结果摆在那里。但是，通往结果的过程，是一个涉及多步骤决策、错误纠正、意图对齐的复杂链条。每一次用户的接受、拒绝、补全、撤销、追问、甚至当模型好几次都搞不定或者完全搞错时的辱骂——都是这一链条上的过程信号。强化学习有两种监督方式，一种叫做结果监督，只看最后是否跑通。但是结果监督会催生「奖励黑客」的现象：模型为了能跑通可能写出冗余、脆弱、带逻辑漏洞的代码，但因为测试过了，模型以为自己学对了。而另一种叫做过程监督，对推理路径上的每一步进行打分。上述这些过程信号，只有在 coding agent 运行环境里才能诞生。GitHub 仓库里只有结果，哪怕是去看单独的提交历史，看 PR，都找不到有效的过程信号。在缺乏有效、自主可获得的过程信号的时候，一些模型厂商会采用「蒸馏」的方式，这个事情大家应该已经知道了。蒸馏的逻辑很简单，给同样的输入，老师模型输出什么，学生模型就学着输出什么。但是通过蒸馏，即便可以获取到思维链，得到的仍然更接近于结果，而非被蒸馏的老师模型内部的概率分布。一旦学生在推理中偏离了老师的轨迹，哪怕一个 token 不符合，都有可能发生偏离。这背后是强化学习的基础限制：策略梯度定理要求，优化样本最好由当前正在优化的模型自己去产生。这种数据叫做 on-policy 数据。而通过蒸馏别家模型，在别人的产品里产生的数据，来训练自己模型，都属于 off-policy 数据。模型当然可以从中学到东西，但学不到老师模型内部的概率分布信息。而像 Cursor 这样自己就是 coding agent 产品的公司，掌握着最真实、有效、高质量的训练数据。Cursor 产品本身，就是 coding 模型在实战环境中的最佳训练场。我们可以通过 Cursor 年初的「翻车」，来证明这个逻辑。4.APPSO 读者应该记得，年初 Cursor 发布了 Composer 2，号称「下一代专用编程模型」，技术报道写的相对保守，也没有提供具体的模型底座信息。结果很快，网友就在公开代码片段里发现了 Kimi 的模型 ID，截图传遍了开发者社群，逼得 Cursor 副总裁 Lee Robinson 出面澄清：「Composer 2 确实是从开源底座出发的。最终模型大约只有 1/4 的算力来自底座，剩下 3/4 是我们自己训出来的。」几小时后，Cursor 联创 Aman Sanger 也跟着发了一条道歉：「一开始没提 Kimi 底座是个失误。」五天后，Cursor 放出了完整的 Composer 2 技术报告，显示底座的确是 Kimi K2.5，授权方则是 Firworks AI，大致流程是在 K2.5 上做训练，再继续做大规模强化学习（RL）。但关键之处在于，Composer 2 的 RL 是运行在真实的 Cursor 会话当中，使用与生产部署完全相同的工具和 harness。Cursor 将这套流程叫做「实时强化学习」(real-time RL)，也即将模型的 checkpoint 直接部署到 Cursor 生产环境中，观察用户的响应，收集数据，聚合成奖励信号——最快可以每 5 个小时迭代一次模型版本，然后继续部署到 Cursor 里，循环往复。最极致的案例是 Cursor 的自动化代码补全功能 Tab，每天处理超过 4 亿次请求，每当用户输入字符、移动光标时，模型都会预测下一步动作，如果预测置信度高，则显示建议，用户按下 tab 即接受自动补全。该功能采用的是在线强化学习，在行业内极具特色。Cursor 可以以极高的频率（最快可达每一个半小时到两小时）更新 Tab 的模型能力给用户，直接在产品内收集 on-policy 数据进行训练。这种高频、接近实时的反馈回路，让 Tab 可以学习到极其微妙的用户意图。Cursor 方面透露，这种方法让 Tab 建议的拒绝率降低 21%，接受率提高了 28%。回到 Composer 模型本身。在事情搞清楚了之后，一些 Kimi 员工也删掉了之前吐槽的的推文，Kimi 官方账号发表了祝贺。一家估值 600 亿美元（基于马斯克给的数字），不做自己的模型基座的 coding agent 应用层公司，仍然可以通过产品自身的数据飞轮，RL 出超越基座模型的专有编程模型。所以与其说 Cursor 翻了车，不如说这反而是 coding agent 产品重要性的绝佳例证。Cursor 在另一篇关于实时 RL 的文章里写到：「（训练编程模型）最大的困难在于建模用户。Composer 的生产环境里不只有执行命令的计算机，还有监督和指导它的人。模拟计算机容易，模拟使用它的人却很难。」这句话，现正在逐渐成为了在编程模型方面走在前沿的模型厂商之间的共识。如果你去看 benchmark 榜单和用户普遍评价，会发现哪些头部的厂商都在发力做自己的 coding agent/编程产品。区别只在于谁离用户更近。我们以 SWE-bench、LLM-Stats 等相对权威的榜单为例，Claude、GPT、Gemini、Kimi 等模型基本霸榜前十，清一色都是有自己开发 coding agent 产品（包括 CLI、IDE、集成 coding agent 的桌面客户端）的模型厂商。在部分榜单上会出现少数反例，如 Meta (Muse Spark)、DeepSeek 等，没有开发自己的 coding agent。不过你会发现，这些反例模型，在更加接近真实场景、避免污染的更权威 benchmark 上就很难上榜了。以 DeepSeek 为例，它在 SWE-bench bash only 上分数是 70%，排名第九，在 SWE-bench Pro 上分数却掉到了 15% 左右。OpenRouter 的真实流量数据可以解释这种反差：该平台 2025 年报告显示，Claude token 消费 80% 以上用于编程和技术任务，而 DeepSeek token 消费主要集中于闲聊和角色扮演。没有自家 coding 产品的厂商，在一些 coding 任务 benchmark 上能挤进头部，但在更难的真实工程 benchmark 上，在用户用 token 消费投票的真实流量中，都会原形毕露。不仅是 Cursor，Anthropic 在 2025 年 11 月发的一篇论文里，也明确透露自己在做一模一样的事情：「我们在 Anthropic 自家的真实生产编程环境上做训练。」也即 Anthropic 把自己员工使用 Claude Code 的交互数据，反哺给 Claude 模型用来训练。5.在 AI 的演进历程中，生产要素的定义发生了深刻的位移。传统三大核心要素——算力、研究、训练数据，虽然在总量上持续增长，但在结构上已经出现了严重的失衡。今天的各大 AI 巨头显著提高了在算力上的资本支出 (CapEx)，让算力基建成为了当前舆论的主旋律。但实际上，特别是在编程范畴内，随着 GitHub 仓库、StackOverflow 等互联网公开代码数据被基模厂商「竭泽而渔」式地利用，模型在代码生成与逻辑推理上的边界开始逐渐显现。这也是为什么，行业共识正在逐渐转向一个冉冉升起的新战略高地：对于任何希望掌握顶级代码能力的模型厂商而言，建立自有的 coding agent 产品早已不再是可选的商业路线，而是确保底层模型可以持续进化的核心生命线。正如前面 APPSO 论证的那样，单纯学习公开数据等于只学习成功者的结局，却无法了解成功的路径，这绝对不是正确的成功学应该有的样子。在真实的编程环境中，知道发生了什么错误、怎样发生的、如何正确地理解和高效地实践需求等等——了解正确过程的价值，远超于得到正确结果本身。只有拥有自己的编码产品，模型厂商才能获取高质量的「过程监督」信号，从而在编码/推理能力的下一阶段竞争中，确保自己仍有技术护城河——否则就不得不像 SpaceXAI 那样，花钱去跟 coding agent 产品公司去合作。然而并不是所有模型厂商都跟马斯克一样有钱，以及 2026 年开始的巨头势力划分、结盟与领地的争斗会变得更加激烈，当一家缺乏自主 coding 产品的模型厂商终于回过味来的时候，恐怕已经没有足够的合作伙伴可以挑选，合作

长城今晚公布了魏牌 V9X 的上市价格：34.98 万元到 38.98 万元。这台大六座 SUV 的最终定价，比预售价下调了两万多。配合 5000 元抵 1 万元的现金膨胀礼，再加上本品牌 1 万元、其他品牌 8000 元的增换购补贴，实际的起售价已经被压到了 33 万这个位置。伴随着新车的亮相，长城汽车董事长魏建军走上了舞台。他接过了话筒，抛出了长城在造车底层架构上的思考：眼下行业里最偏好的往往是新能源车专属架构，魏建军偏偏逆风而行，将归元 S 做成了兼容燃油、插混和氢燃料的全动力平台。实际上，魏建军看上的是更广阔的海外版图。他表示，在很多国家，电力基础设施的建设依然迟缓，若用全动力平台去打全球市场，则能更好地把造车规模做大，将高昂的研发成本摊薄，从而稳住车辆的市场保有量和保值率，也让用户的修车费用降下来。对于新能源专属平台，魏建军在台上表达了自己的看法。实际上可以这么讲，我们近十年就是被这个电动专属平台忽悠了。现在，由于动力的多样化，配置的多样化，我们理论上可以卖到全世界去。造车的底子铺好之后，魏建军把话题拉回了最基础的用车体验。他反复强调，无论车企给出的承诺多么好听，最根本的准则永远是不能给用户添麻烦。为什么要添一句「不要给用户添麻烦」？你想想，当你在路上被撂下、车出了问题的时候，你的心情有多么焦虑。顺着老板定下的规矩，长城的工程团队把干货都掏了出来。魏牌 V9X 是一台车长近 5.3 米、轴距突破 3.1 米的全尺寸 SUV。为了拉动它，长城给它准备了一套超级 Hi4 混动系统和 2.0T 发动机，并搭配了一台 4 挡专用变速箱。在低挡位运转时，这套动力总成能爆发出 4500 牛·米的轮端扭矩，足以应对陡坡和极端烂路。长城还给出了明确的性能指标，声称不管车辆是满电还是亏电状态，都能跑出 4 秒级的百公里加速成绩；进入高速巡航状态，车辆又能把馈电油耗稳稳控制在百公里 6.3 升左右。底盘配置同样堆满了料。长城给 V9X 全系标配了闭式双腔空气悬架和 CDC 减震器，以及正负 10° 的后轮转向系统。这套后轮转向系统顺带接入了自动泊车功能，面对狭窄的侧方车位，后轮会自动微调角度配合入库，让泊车过程更加流畅。至于后轮转向对座舱空间的侵占，长城的设计团队顺着外扩的轮包，把第三排座椅做成了环绕式的大沙发造型，化解了结构上的劣势。在 V9X 的座舱里，这样的巧思还有很多，例如，副驾位置的手机无线充电板没有放在常规的中控台下方，而是被挪到了车门内扶手上，还带有一点微倾的角度。乘客的手自然搭在门板上，就能顺手拿到手机。智能化算力拉到了目前的顶格水平。英伟达 Thor 芯片负责辅助驾驶，骁龙 8295P 支撑座舱运转。虽然用的是别家的芯片，但长城在软硬件的协同上选择了全栈自研，把动力、底盘和辅助驾驶的控制权全部掌握在自己手里。魏建军表示，V9X 的各个系统之间能够实现毫秒级的响应，遇到突发危险情况，车辆可以提早判断，做到刹车的同时，也能进行绕行避障。他还强调，为了保护隐私，所有的语音指令和行程数据都做到了本地闭环，绝不外传。在加入了大模型后，车机里的「小魏同学」也变得更加智能了。发布会现场展示了一段测试，视频里的魏建军特意换了件外套，戴上墨镜坐进主驾。即便如此，系统还是在第一时间认出了他，并自动调出了他习惯的座椅位置和空调温度。发布会的后半程，舞台交给了魏牌 CEO 赵荣波，他把魏建军定下的准则，变成了一项项具体的购车权益。比如高阶智驾软件对用户终身免费开放，锁单后如果超过 30 天没交车，长城每天给用户补贴 200 元。其中最受关注的，是那份 6 年 15 万公里的整车质保。按照汽车行业的常态，普通质保通常只有三四年，部分品牌在遇到空气悬架、后轮转向这类高成本的底盘件时，还喜欢用易损件的名义另行算账。且这类零件的故障高发期，刚好卡在车子开到第四到第六年的时候，一旦脱保，维修和更换的费用动辄几万元。长城这次把空气悬架和后轮转向系统也放进了长周期的保修范畴里。赵荣波在台上强调，相关的服务条款没有任何文字游戏，也不设隐藏门槛。整场发布会的落脚点，最终还是回到了魏建军身上。这位长城的掌舵人拒绝了常规的营销套路，选择亲自上阵，直接拿自己的个人信誉和姓氏为这台新车兜底。面对网友关于长城会不会「翻车」的犀利提问，他没有回避，给出的回应干脆利落。我认为长城价值观和对用户负责任的这种精神，永远翻不了车。大家选择长城，我就要肩负起这份责任，让每个选择魏的人放心，这才是真正的代言。我的代言才刚刚开始。魏，你放心。

今年 3 月，苹果制造了 MacBook Neo 这个大爆款，把高质量入门笔记本的价格打到了 4599 元人民币起。一个半月后，这台搭载 A18 Pro 处理器的入门款 Mac，全年销量从 700 万台追加到了令人难以置信的 1000 万台，交付周期一度排到了 5 个月之后……苹果紧急向鸿海、广达等供应商追加订单。知名供应链分析师郭明錤给出更刺眼的数字：在全球 Windows 笔记本电脑预计同比衰退超 10% 的大背景下，2026 年 MacBook 系列出货量有望达 2500 万台，同比增长 20%-25%。这一拳，Windows 阵营的最核心代表英特尔，该如何接住？在今天，英特尔在中国正式推出第三代酷睿处理器，代号 Wildcat Lake「野猫湖」，并同时启动了联合产业链伙伴的「Firefly 萤火虫计划」。英特尔的接招，围绕一件事而展开：将最先进的 Intel 18A 制程，下放给低端，或是入门，或是「主流」——随便你怎么称呼它的价位段。野猫盯上新的猎物Wildcat Lake 这个代号，在去年底就已经在业界出现。今年 CES 期间，英特尔高调发布了主打高端的产品线：酷睿 Ultra Series 3，采用 Panther Lake 架构。同一天，Wildcat Lake 也悄悄披露，你可以将它理解为 Panther Lake 的精简版本。业界普遍将其视为英特尔为了对标苹果 MacBook Neo 而推出的产品，专攻平价入门级 PC 市场。「直接对标」这四个字，并不是说说而已。在 4 月底，英特尔为野猫湖推出的公版参考设计曝光，采用渐变配色、铝合金一体成型机身、窄边框键盘，扬声器开孔环绕在键盘两侧、D 面无出风口——这种设计语言不需要过多解释，它就是在用苹果惯常的设计语言，去回应苹果本身最受欢迎的多彩配色与铝合金轻薄机身策略。在 MacBook Neo 上，苹果做对了一件事：它让整个行业重新意识到，在「内存末日」(RAMmageddon) 的今天，入门价位段成了一座金矿。用一颗本来进入销库存阶段的 A18 Pro 芯片，配上 8GB 内存和 256GB 存储，开价 4599 元——MacBook Neo 活脱脱成了 MacBook Air 平替。英特尔，以及整个 Windows PC 阵营不能继续将这个用户群体拱手相让。而这，正是「野猫湖」存在的（几乎）全部理由。18A 首次下放到入门段第三代英特尔酷睿处理器的技术参数并不复杂，CPU 采用 2 个 Cougar Cove 性能核，加上 4 个 Darkmont 超低功耗能效核，总共 6 核 6 线程；GPU 部分配备 2 个 Xe3 核心，NPU 算力最高可达 18TOPS。整机基础 TDP 15W，最高 35W，在无主动散热的笔记本上，甚至可以压低到 11W 运行。按照英特尔今天提供的官方口径，这颗芯片功耗相比前代降低超过 60%，图形 AI 性能提升了将近 3 倍，平台总算力更是达到了 40TOPS。当然，考虑到「野猫湖」处理器的 NPU 算力只有 17TOPS，远低于 Windows 阵营的所谓「AI PC」，也即微软官方定义的 Copilot+ PC 40TOPS 的阈值。所以严格来说，野猫湖并不能直接和苹果 A18 Pro 高质量的神经网络引擎，以及统一内存架构，去做横向比较。话虽这么说，驱动参数量不大的本地小模型，做一些 AI 推理任务，还是绰绰有余的。以及更有意思的是，谷歌前段时间发布自家的 AI 笔记本电脑「Googlebook」的时候，也确认了和英特尔结盟——Wildcat Lake 被海外业界人士普遍认为将会成为是 Googlebook 的核心平台。连谷歌都觉得够用，应该没大问题吧？真正关键之处不在于这些参数，而是在于「野猫湖」用上了 Intel 18A 制程工艺本身。要知道，过去两代酷睿 Series 1 和 2，一直沿用 Raptor Lake-U 架构和 Intel 7 制程，实质上是 10nm SuperFin 工艺的演进版本。在主流芯片普遍走向 5nm 甚至 3nm 的当下，技术代际差距已经相当明显。当 MacBook Neo 用虽然是老库存但货真价实 3nm 工艺的 A18 Pro，杀进四千多元的价位段，英特尔过去那套「靠老制程压低成本」的入门策略已经过时了。所以，我们很欣慰地看到，英特尔将 18A——2nm 级别的工艺，压到一颗入门芯片上。这意味着英特尔充分愿意承担初期更高的单位成本，去力保一座不能再丢弃的城池。系统：手机供应链思路搬到 PC 上对于守住市场，「Firefly 萤火虫计划」同样重要。英特尔副总裁、中国区软件工程和客户端产品事业部总经理高嵩如是说：芯片本身的算力固然重要，但「芯片和系统层面的双重创新」才是这一代主流 PC 体验跃迁的来源。中国区客户端与平台销售业务部总经理宗晔则认为，PC 也正在从工具进化为「智能体验入口」。根据英特尔提供的内部数据，与上一代相比，「野猫湖」产品的 PCB 面积缩小了 5%，器件数量减少了 7%。同时针对第三代酷睿平台特性制定并推广标准化主板与 I/O 板连接，推动主板核心电路进一步实现模块化、小型化、低功耗化。这一套打法的精神底色，其实来自于手机。手机行业过去 10 年磨合出的供应链协同方式——元器件高度集成、模块化主板、统一接口标准，被英特尔搬到了入门级 PC 上。结果，就是更轻、更薄，也更便宜。参与「Firefly 萤火虫计划」的合作伙伴，包括但不限于华硕、七彩虹、荣耀、惠普、联想等。英特尔预计，全球合作伙伴将陆续推出超过 70 款基于 Wildcat Lake 平台的产品。目前，荣耀 MagicBook X14 2026 已经率先上架，搭载该平台 Core 5 320 处理器，采用 16GB+512GB 存储配置，定价下探到了 4399 元。用户画像：完全对标 MacBook Neo，但还有更多英特尔列举了 Wildcat Lake 平台产品的 5 类目标用户：辅导孩子作业的家长、全天在校的学生、刚入职场的新人、「一人公司」创业者，以及银发族。这些用户群体和场景描述本身没什么悬念，但值得注意的是，这个用户画像，和 MacBook Neo 在线下门店主要推荐的人群，几乎是完全重叠。MacBook Neo 夺走的本来就是入门级 Windows 笔记本电脑的客群，英特尔的接招，就是造出具有更好价值主张的入门级笔记本电脑，再把心智重新抢回来。英特尔还反复强调野猫湖平台的本地 AI 推理算力。这对于那些购机预算本就不多，也不太可能付费订阅优质云端模型算力的用户来说，倒也是个正确的思路。最后，野猫湖看起来像是合格的入门芯片平台，跟 Firefly 萤火虫计划结合起来，也形成了一套颇具诚意的组合拳。轻、薄、便宜的 AI 笔记本，英特尔现在重新回到了这个擂台。仅剩的悬念，是 OEM 最终能把零售价做到多少。在 MacBook Neo 的价位上，苹果给到了系统生态、统一内存架构、多种潮流配色等等诸多看头。但英特尔要想把这只野猫放出去，跟苹果大杀四方，光有 18A 工艺和优秀的工业设计是不足够的。这场仗的真正交战点，大概率不会在技术本身，而是在即将到来的 618 购物节期间，第一批搭载它的笔记本产品，能否真正把供货紧张的 MacBook Neo 杀个七零八落、措手不及。

⚖️马斯克诉 OpenAI 案败诉📱苹果开启 WWDC 邀请，将公布苹果 AI 新进展🚗问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处🤖国家反诈中心 App 上线「AI 内容鉴定」⌚苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动🚙广汽董事长回应被称网约车之王🧠三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能🎵腾讯音乐官宣收购喜马拉雅💰铠侠股价飙升，计划美国上市🚀马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元☁️百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半🎬爱奇艺一季度由盈转亏 2.95 亿元🦿荣耀机器人「闪电」「元气仔」走进南大💡前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去🚗小米 YU7 GT 定档周四发布🎮5499 元起，红魔 11S Pro 系列发布马斯克诉 OpenAI 案败诉据 CNBC 和《华尔街日报》报道，美国加利福尼亚州奥克兰联邦陪审团已就埃隆 · 马斯克诉 OpenAI 及其 CEO 山姆 · 奥特曼一案作出裁决，认定马斯克的起诉超出诉讼时效，驳回其全部主张。此案源于马斯克 2024 年提起的诉讼。他指控 OpenAI 及奥特曼违背了将该公司作为非营利机构运营的承诺，并在庭审中称 OpenAI「窃取了一家慈善机构」。陪审团的裁决并未就「违反慈善信托」指控的实质内容作出判断，而是认定马斯克的起诉超过了三年诉讼时效。法院同时驳回了针对微软的相关指控。裁决宣布后，马斯克在 X 平台发文，将这一结果定性为「日程上的一个技术性问题」，并表示将向第九巡回上诉法院提起上诉。OpenAI 首席律师威廉 · 萨维特在庭后表示，这「不是一个技术性裁决，而是一个实质性裁决」，并称马斯克是在将诉讼作为竞争手段使用。微软方面亦发表声明，对裁决表示欢迎。苹果开启 WWDC 邀请，将公布苹果 AI 新进展今天凌晨，苹果正式启动 WWDC26 邀请，并公布详细日程安排。本届 WWDC 大会将于北京时间 6 月 9 日至 13 日举行，面向全球开发者开放，重点展示 Apple 各平台的新系统能力、开发工具和框架更新。按照安排，WWDC26 将于北京时间 6 月 9 日开幕，届时 Apple 将举行主题演讲和 Platforms State of the Union，介绍新一代软件平台、人工智能相关进展，以及面向开发者的新工具。大会期间，开发者和学生可以观看超过 100 场围绕工具、技术和设计的视频讲座，参与团体实验室活动，并在 Apple 开发者论坛中与工程师和社区成员交流。Apple 还将在 6 月 9 日邀请超过 1000 名开发者、设计师和学生前往 Apple Park，参与线下特别活动。届时，爱范儿也将在 Apple Park，为大家带来第一手的现场报道。问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处全新一代问界 M9 系列汽车设计师 @Zimo-子墨 日前在微博发文，点赞全新理想 L9 的内饰设计，称其「简洁、逻辑、比例和谐，视觉重心清晰，口味控制得恰到好处」。他表示，初创公司在有强决策者的情况下，往往对设计管控介入较深，容易打散设计的贯穿逻辑；而到第二代、第三代产品时，无论前代成功与否，团队通常会更多放权给设计师打磨表达。他同时透露，新 L9 的内饰设计总监为来自宝马的 Chris Lee，并称虽然双方长期管理直接竞品，但「真的欢喜好设计，生在中国」，并已发短信向对方表示祝贺。国家反诈中心 App 上线「AI 内容鉴定」据央视新闻报道，国家反诈中心 App 近期正式上线「AI 内容鉴定」功能，支持对图像、视频、音频及文本进行 AI 生成痕迹检测，主要用于识别诈骗及电商场景中的伪造内容。电商平台上利用 AI 篡改图片骗取「仅退款」的行为日趋普遍。央视新闻记者随机选取十张商家提供的疑似 AI 生成售后图，上传至国家反诈中心 App 及多个 AI 鉴真工具后，均被判定为「图像含 AI 生成痕迹」。据介绍，该检测方案结合小模型与大模型：小模型学习 AIGC 内容的生成模式，大模型则判断内容是否存在物理瑕疵或逻辑语义问题，二者联合输出最终结果。苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动据 DIGITIMES 报道，苹果计划于今年下半年对高端 Apple Watch 机型进行重大改款，传感器功能将同步升级，市场预计届时出货量较 2025 年增长 20% 至 30%。Apple Watch 11 已具备血氧、脉搏、心率及血压感测能力，下一阶段预计推进无创血糖监测，具体落地时间取决于 FDA 审批进展。作为苹果传感器元件的独家供应商，亚洲半导体（TASC）预计将直接受益。市场预期新品订单最早于 7 月启动，带动下半年业绩明显超越上半年。广汽董事长回应被称网约车之王据「睿见 Economy」报道，日前广汽集团党委书记、董事长冯兴亚在「第十八届轩辕汽车蓝皮书论坛」上，针对「网约车之王」的说法，其指出：很多消费者可能没有想明白一个问题，「网约车是生产资料」。生产资料对品质的要求极高，因为它是用来挣钱的工具。只有能为自己带来更多利益的车，才能受到使用者和购买者的欢迎。对司机来说，车要「多快好省」，要能赚钱。「这恰恰是对品质最好的验证。没有高品质，就不可能受到网约车驾驶者的欢迎，这是最基本的。」他进一步分析，网约车的乘坐周期非常长，一天可能十几个小时都在车上。如果舒适性不够、疲劳感很强，它也不会受到司机的欢迎。此外，电耗等各项指标，也决定了网约车司机是对车要求最严格的一群人。冯兴亚总结道，这些从网约车运营中积累下来的经验，广汽会延展到旗下各个品牌，让车辆变得更加可靠、舒适和安心。「这反而是我们的优势所在。」三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能据 DIGITIMES 报道，三星电子近期大幅提升了用于高带宽内存（HBM）的最新 DRAM 良率。其第五代 10nm 级 1b DRAM 良率已达 92%，用于下一代 HBM4 产品的第六代 1c DRAM 良率也突破 75%。良率的提升有望增强三星的成本竞争力，并提升其向主要 AI 客户大规模供应先进存储产品的能力。在产能扩张方面，据 Ddaily 报道，三星和海力士正同步提前推进工厂扩建和设备订单。三星正加速平泽 P4 工厂的 DRAM 投资，设备安装预计持续至今年底，市场估计明年将进一步扩大投入。海力士则在清州 M15X 和龙仁 Y1 工厂推进扩产，业界预计其明年 DRAM 相关资本支出有望较今年翻倍。腾讯音乐官宣收购喜马拉雅腾讯音乐娱乐集团（TME）昨日宣布，已正式完成对喜马拉雅并购交易的交割，喜马拉雅成为 TME 全资子公司。交割完成当日，TME 执行董事长彭迦信与 CEO 梁柱联合向全体员工发布全员信。管理层称，喜马拉雅是中国长音频领域的重要建设者，双方战略方向高度一致，合并后将共同拓展在线音乐与音频行业空间。铠侠股价飙升，计划美国上市据彭博社报道，铠侠（Kioxia）股价昨日上涨 16%，创下新高。公司在上周披露强于预期的业绩展望后，多家华尔街机构迅速上调目标价；与此同时，铠侠也确认正准备在美国发行存托凭证，借此提升流动性并扩大国际资本市场接触面。报道提到，铠侠预计截至今年 6 月的季度营业利润将达到 1.3 万亿日元，高于市场平均预期，季度利润也创下纪录。多家机构认为，在 AI 带动的存储供需紧张环境下，公司仍有进一步提升自由现金流和股东回报的空间。马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元据彭博社报道，马斯克昨日在萨姆森国际智能移动峰会线上视频连线时表示，正在德克萨斯州推进 SpaceX 上市准备，称「SpaceX 的 IPO 事宜将很快开始」。知情人士透露，SpaceX 最快本周提交 IPO 申请文件，计划融资最高 750 亿美元，目标估值超 2 万亿美元，若成功将成为史上规模最大的 IPO。百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半百度昨日发布 2026 财年第一财季业绩，期内实现营业总收入 320.75 亿元，同比下降 1.16%；归母净利润 34.45 亿元，同比下降 55.36%；经营现金流为 27 亿元。具体来看，百度核心 AI 新业务收入达到 136 亿元，同比增长 49%，首次占百度一般性业务收入比重超过一半；其中智能云基础设施收入 88 亿元，同比增长 79%，GPU 云收入同比增长 184%。萝卜快跑在一季度的全无人自动驾驶运营订单达到 320 万单，总订单量同比增长超过 120%，截至今年 5 月其全球业务已覆盖 27 个城市，累计自动驾驶里程超过 3.3 亿公里。爱奇艺一季度由盈转亏 2.95 亿元爱奇艺昨日发布 2026 财年第一财季业绩，期内实现营业总收入 62.26 亿元，同比下降 13%；归母净利润由盈转亏 2.95 亿元，去年同期为盈利 1.82 亿元；经营现金流为 1.86 亿元，同比下降 45.13%。分业务看：会员服务收入为 42 亿元，同比下降 5%；在线广告服务收入 12.4 亿元，同比下降 7%；内容发行收入 3.59 亿元，同比下降 43%；其他收入 4.27 亿元，同比下降 49%。公司在财报中提到，热门剧集仍推动会员收入环比增长，海外业务保持稳健增长，并计划借助AI降低内容制作成本。荣耀机器人「闪电」「元气仔」走进南大5 月 17 日，荣耀携自研机器人「闪电」与「元气仔」在南京大学开展具身智能科普活动，并同步展示荣耀 600 系列新品的核心影像功能。活动现场，荣耀展示了自研机器人的核心运动与交互能力。搭载高动态运动系统、支持自主感知导航与高速避障的机器人「闪电」与在校学生进行了 8 轮 1V5 短跑测试；另一款机器人「元气仔」则完成了包含「太空步」在内的复杂舞蹈动作演示与人机互动。同期，荣耀在现场设立了新品体验区，重点展示荣耀 600 系列智能手机。官方信息显示，该机型主打基于 Magic OS 系统的 AI 能力及 4K Live Photo 动态照片功能。外观方面，荣耀 600 系列目前已公布「幸运星」「青苹果」「光羽蓝」和「曜石黑」四款配色，其中主打配色「幸运星」采用了行业首创的双重磁吸工艺。💡 前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去前 Tesla 无人驾驶与 SpaceX 猛禽发动机团队工程师 Zac Valles 在 YC Demo Day 结束 72 小时后直飞深圳，驻扎 8 周后携新硬件返回旧金山。他随后在 X 上发布了一篇长帖，向所有有意从事硬件创业的人发出忠告：不论有没有融资、团队甚至创业想法，都应该立刻去深圳。只要你对创办一家硬件公司有哪怕一点点想法，就去深圳。融资前、组队前、有想法前、离职前，都不是借口。去就对了。在切入供应链的路径上，Valles 建议创业者围绕一场大型行业展会安排行程，将其作为接触工厂的跳板。他还建议在华强北预留 4 到 6 小时实地走访，称其为一座「微型城市」，机器人子组件、电池、底盘、电子元器件均有专属楼栋聚集。小米 YU7 GT 定档周四发布昨日，小米正式宣布，将于 5 月 21 日（本周四）举行人车家全生态新品发布会，届时将发布小米 YU7 GT、小米 17 Max 等新品：小米 YU7 GT：定位跑车级 SUV，最大马力可达 1003ps，最高时速 300km/h，CLTC 纯电续航 705km；小米 17 Max：采用 6.9 英寸极窄四等边直屏，搭载徕卡 2 亿像素主摄和 8000 mAh 金沙江电池；小米手环 10 Pro：最薄 9.7 mm，最轻 21.6 g，续航最长可达 21 天；小米首款耳夹式耳机：单机重 5.5 g，支持 LHDC 5.0。5499 元起，红魔 11S Pro 系列发布红魔昨日正式发布红魔 11S Pro 系列，国内首发第五代骁龙 8 至尊领先

你帮我找一辆白色硬棚 Z4 M-Power，蓝色双轮毂，找那个富家子出来，我要找他试车。在 2012 年上映的港片《车手》里，余文乐饰演的警察陈翔留下了这样一句台词。不久前他在中环经历了一场惨败——那辆白色宝马 E86 Z4 M Coupe 在逼仄的车流和弯道中游刃有余，将他远远甩开。值得一提的是，现实中 E86 本就稀少，而 Z4 M 更是凤毛麟角，尤其是手动挡版本。这也难怪导演郑保瑞会用它来为整部影片定调，那种桀骜不驯的气质，恰如其分。当年大银幕上的稀罕物定格成了经典，而在现实的流水线上，属于 Z4 的齿轮也停止了转动。前一阵，奥地利格拉茨的麦格纳工厂内，最后一辆底盘代号为 G29 的宝马 Z4 缓缓驶出生产线，这款问世于 2002 年、历经三代演进的经典敞篷跑车正式宣告停产。伴随它的离去，宝马延续近三十年的双座 Roadster 车系，正式画上了休止符。巴伐利亚的驾驶图腾Z 系列的起点要追溯到三十多年前。在德语里，字母「Z」代表「Zukunft」，也就是未来的意思。90 年代的宝马对「未来」的理解非常具象，还带着一丝日耳曼式的偏执：剥离一切与驾驶无关的多余配置，把一台直列六缸发动机塞进一个小巧的底盘里，只保留两个座位和纯粹的驾驶乐趣。▲ BMW Z11989 年，采用下沉式隐藏车门的 Z1 惊艳亮相。这台车在今天看来像个古怪的科技实验品——它的车门可以通过电机垂直降入门槛内，车身覆盖件采用了可以由车主自行拆卸更换的特殊热塑性塑料。▲ BMW Z1除了这些前瞻设计，Z1 最重要的历史意义在于，它用全球限量 8000 台的桀骜姿态，确立了整个 Z 系列不按常理出牌、游走在主流边缘的先锋基调。六年后的 1995 年，Z3 走向了更广阔的大众市场。伴随着皮尔斯·布鲁斯南在 007 电影《黄金眼》中的潇洒亮相，这台远赴美国斯巴达堡工厂生产的 Roadster 迅速风靡全球。▲ 007：黄金眼中的 Z3但真正的硬核故事发生在幕后。当时宝马的工程师伯克哈德·戈舍尔（Burkhard Göschel）对敞篷跑车疲软的车身刚性极度不满。于是，他带领一群狂热的工程师突发奇想，利用周末的闲暇时间躲在工厂里，给一辆 Z3 焊上了车顶——这就是后来被车迷们戏称为「小丑鞋」的 Z3 M Coupe。▲ Z3 M Coupe虽然它造型怪异、空间局促，但封闭车顶使其抗扭刚性比敞篷版提升了 2.6 倍。在当时，它是纽博格林北环赛道上最让人胆寒，也最难驾驭的机器之一。前《Top Gear》主持人杰里米·克拉克森曾毫不掩饰地表达过他对这种老派双座跑车的偏爱：当一台车拥有那么长的车头，而你几乎坐在后轴上开车时，它就不再是一个交通工具，而是一台纯粹为了让人发笑而制造的机器。到了 1999 年，宝马 Z8 将这种暴力美学与复古浪漫推向了顶峰。▲ BMW Z8由亨里克·菲斯克操刀设计的 Z8，完美致敬了 50 年代的传奇跑车宝马 507。修长优雅的全铝车身骨架下，藏着与同时期 E39 M5 同款的 4.9 升 S62 V8 引擎。它顺理成章地拿到了 007 电影《黑日危机》的剧本，成为詹姆斯·邦德的新座驾，也让全世界记住了属于 Roadster 的优雅与从容。进入 21 世纪，Z4 正式接过了 Z 系列的接力棒。▲ BMW Z4第一代 Z4（E85/E86）诞生于克里斯·班戈主导设计的年代。那套著名的「火焰表面」设计语言，在当时引起了轩然大波。无论是前脸修长的敞篷版，还是车顶极具肌肉感的硬顶 Coupe 版本，侧面那道巨大的「Z」字形折线都带着一股咄咄逼人的凌厉感。好的汽车设计应该能够激发强烈的情感反应。你不需要向别人解释它为什么美，当人们看到它时，心跳就会加速。前宝马集团设计副总裁克里斯·班戈的这套理论，在第一代 Z4 身上体现得淋漓尽致。尤其是前文提到的 Z4 M。▲ Z4 Coupe宝马工程师把当年 E46 M3 上那台被奉为自然吸气神作的 S54 3.2L 高转速直列六缸发动机，塞进了 Z4 狭小的机舱里。它就像一匹没有完全驯化的野马，脾气暴躁，容错率极低，却又让人在征服它之后欲罢不能。2009 年，第二代 Z4（E89）问世。当时整个汽车行业刮起了一阵折叠硬顶风潮。宝马也顺应市场，为 Z4 换上了复杂的机械硬顶机构。▲ 外饰设计师 Juliane Blasi两位女性设计师，Juliane Blasi 和 Nadya Arnaout，赋予了 E89 极其优美的车身比例和圆润的线条，不过，硬顶机构不可避免地增加了车身重量，打破了原有的配重平衡。这台车在动态响应上做出了一定妥协，收敛了上一代的暴躁性格，演变成了一台更适合在海滨公路迎风巡航的 GT 跑车。▲ BMW Z4（E89）当时间来到 2018 年，第三代也是最后一代 Z4（G29）面临着一个严峻的市场环境。跑车市场不断萎缩，独自研发一款全新底盘的成本已经高到令人咋舌。因此，宝马选择拉上丰田，基于 CLAR 平台共同分摊研发成本。为了找回丢失的运动基因，G29 毅然抛弃了笨重的机械硬顶，回归了传统的软顶敞篷设计，不仅成功减重，更大幅降低了车身重心，找回了完美的 50:50 前后轴配重比。▲BMW Z4（G29）顶配车型搭载的 B58 3.0T 直列六缸发动机，提供了 340 马力的强劲输出与浑厚的声浪。更令人感动的是，在它的生命周期末段，宝马还出人意料地为直六车型推出了手动变速箱版本。工程师为其重新设计了换挡连杆和机械结构，只为呈现最纯粹的机械咬合手感。这种逆潮流而动的做法，被很多车迷视作一种浪漫的告别仪式。这台车从不妥协于日常琐碎，它存在的唯一理由，就是在踩下油门的那一刻，为你制造纯粹的多巴胺。理想与现实粤语里有一句老话，叫「有情饮水饱」。但在真金白银的汽车行业，情怀从来填不饱肚子。翻开宝马集团的财务报表，Z4 的谢幕早有预兆。G29 在 2019 年上市时还算风光，全年卖了 15827 辆。但这也是它最后的体面。2025 年，Z4 在全球只卖出了 9744 辆，这个数字放在宝马年销两百多万辆的大盘里，存在感微乎其微，占比连 0.4% 都不到。更大的问题在于，养这条产线还很烧钱。G29 并不在宝马自家工厂生产，而是交给奥地利格拉茨的麦格纳·斯太尔代工。一年卖不到一万辆，宝马却得单独为它保供一整套零部件体系。当年宝马联合丰田把 G29 推向市场，本就是一笔算计好的生意——分摊成本，换 Z4 多活几年。但合作研发终究只能解燃眉之急，无法逆转大时代的洪流。实际情况就是，卖掉一台 Z4 的利润，还比不上一台走量的 X3。前宝马集团研发董事傅乐希曾无奈地描述过跑车面临的困境。如今的汽车市场，开发一款跑车比以往任何时候都要困难。你需要高销量来分摊平台研发成本，而跑车市场的容量恰恰在不断萎缩。宝马现在一门心思扑在新世代纯电平台上。这是个烧钱的无底洞，几百亿欧元砸进去都不一定听得见响。新旧交替的关头，钱必须花在刀刃上，花在能决定明天还能不能活的地方。正因如此，砍掉投入产出比极低的 Z4，将宝贵的研发资源转移到更有前景的新能源项目上，是企业求生的本能。▲ 新世代 M3放眼整个行业，双门敞篷跑车的消亡已经成为一种趋势。奔驰 SLC 早就退出了历史舞台，奥迪 TT 也吹响了熄灯号。同级别的竞争对手一个个退场，Z4 能坚持到 2026 年才落幕，已经算得上长寿。宝马集团前 CEO 齐普策说过：「内燃机时代的落幕会让人感伤，但我们必须为未来的移动出行寻找新的答案。」接力棒已经交给了纯电车型，但我们依然会怀念内燃机，怀念沉重的电池组无法替代的那种轻盈灵动。一台 G29 Z4 的重量在 1.5 吨左右，而一台纯电跑车的重量往往会超过两吨。物理定律是客观存在的，增加的重量没有办法完全通过主动悬架来弥补。双座小跑车的魅力很大程度上建立在人与车的沟通感上，座椅和方向盘转向传递的路感，排气管的回火声，甚至是换挡时那一瞬间的顿挫，都在不断刺激着驾驶员的感官。没有了齿轮咬合时的震动，没有了直列六缸发动机逼近红线时的嘶吼，缺少了这些声浪与震动的刺激，速度在屏幕上只是一串苍白的数字。当时代浪潮决绝地退去，留在沙滩上的，往往是普通车迷的执念与阵痛。以后想要体验阳光、微风与汽油味混合的纯粹快乐，大概只能去二手车市场里寻找曾经的碎片了。而淘旧账的代价，往往比想象中疼得多。我的朋友张亦就花 10 万元买了一台 E85 Z4。买回来才发现，老跑车的维修称得上是个无底洞，配件是又贵又难等，单是内外三个后视镜片就花掉他 5500 元。买车 10 万，修车修了 5 个月又砸进去 10 万，到现在还没完全整利索。随着 Z4 的停产，维持这种老派乐趣的门槛只会越来越高。那个为了纯粹驾驶乐趣不计成本造车的黄金年代，连同它引以为傲的直列六缸声浪，就此封存在了过去。

2013 年，埃斯特庄园古董车展。为了纪念宝马摩托车诞生 90 周年，宝马带来了一台名为 Concept 90 的概念车。这个名字大家可能不太熟悉，但其实它的量产车型就是后来大获成功的现代复古标杆——R nineT，人送外号「宝马拿铁」。▲ R nineT六年后，同样是在科莫湖畔的这个庄园里，Concept R18 概念车登场，它顺理成章地开启了宝马 1800cc 水平对置巡航车 R 18 家族的时代。埃斯特庄园似乎成了宝马重磅复古车型的首发圣地。到了 2026 年的今天，宝马带着全新的 Vision K18 概念车再次回到这里。这辆概念车的造型区别于现有的 K1600，甚至和我们认知里的休旅车型都不太一样，更像是一个超音速飞行器。为了这个造型，宝马互换了进气和油箱的位置，并将后部线条尽可能压平，营造出一种低伏、前倾的姿态。哪怕它静静地停在聚光灯下，依然带有一种贴地飞行的视觉张力。长达两米的铝制侧板是这台车最引人注目的地方，呈现出一种无缝、一体的连续性。大面积的手工铝材与深色的锻造碳纤维搭配在一起，摒弃了多余的装饰，让整台车散发出浓烈的重工业味儿。数字「6」作为这台车的图腾，被具象化地刻在车头和车尾——前方是六颗带有主动散热功能的 LED 头灯，车尾的碳纤维包裹着六根粗壮的排气管。这套不锈钢排气系统采用了火焰喷涂工艺并覆有液态铝，质感类似于 F1 赛车的排气管。宝马用「The Heat of Speed」来概括这台车的设计主线。过去，那些为长途休旅准备的动力平台总是被包裹在整流罩内，安静且内敛。现在，宝马选择让这台排量达到 1800cc 的直列六缸发动机和排气完全裸露。车身结构向发动机和进排气让位，复杂的机械部件成为最抢眼的视觉元素。BMW Motorrad 首席执行官 Markus Flasch 说：通过 Vision K18，我们展示了对性能、豪华与情感的新理解。直列六缸发动机对我们而言远不止是一项技术，它是一种宣言，Vision K18 将这种力量转化为一种雕塑般的存在，让速度与动感在静止状态下依然清晰可感。从概念到现实尽管 Vision K18 在设计上追求的是速度与动感，但它始终是一辆为巡航而生的车型，它最终的量产车型很有可能就是 K1600 的继任者。或许，它会叫 K1800？▲ BMW K1600GT不管最终如何命名，这台新车最受瞩目的依然是那台排量提升后的六缸发动机。放眼目前的全球摩托车市场，愿意在两轮车架里塞进六个气缸的厂商屈指可数。在这之中，水平对置是最主流的选择。本田金翼那台 1833cc 的水平对置六缸就是行业内的典型代表，国内的长城灵魂用的也是水平对置八缸。宝马则选择了一条独特的技术路线——直列六缸。将六个气缸排成一条直线，会严重挤压摩托车的横向空间。为了控制车身宽度，宝马在 K1600 上就采用了非常紧凑的缸间距设计，气缸套之间的距离仅有几毫米。与此同时，它的缸体也向前倾斜了约 55°，既降低了整车重心，也为上方的进气系统和车架腾出了位置。这种结构带来了截然不同的特性。本田金翼偏向低转速和高扭矩输出，相比之下，宝马的直列六缸则带有一些运动轿车的特质：转速攀升更快，高转速下的声浪更具穿透力，动力输出也更加绵密。若是在此基础上扩缸到 1800cc，低转速时的扭矩想必也会有所提升。动力的升级让 Vision K18 做好了驶向公路的准备，但在实际骑行中，长途巡航除了需要充沛的动力，还要能够应对迎面而来的风和雨。这套动力平台真正进入量产、接替现款车型时，我们在概念车上看到的科幻外观大概率会向实用性让步。摩旅时，天气总是不似预期，大尺寸电动风挡、小腿前的整流罩、大容量边箱与尾箱，以及雷达、电控等系统，都会重新出现在量产车上。Vision K18 只是提前展示了这台机器剥去外壳后的原生状态，并且，它未来的量产形态可能不止于传统休旅车。参考宝马在 R 18 车系上的产品布局，依托同一台 1800cc 水平对置双缸发动机，他们衍生出了复古裸车、Bagger 巡航以及豪华休旅等多个版本。▲ BMW R 18全新的直列六缸平台也可能会复制这样的思路。它既可以加上整流罩，作为 K1600 GT/GTL 的正统换代车型进入休旅市场；也能像 Vision K18 展示的这样，推出一款褪去整流罩的复古版本，正好也能填补宝马 Heritage 继承者系列的产品线——R nineT 依托风冷水平对置双缸奠定了基础，R 18 切入了美式巡航市场，如果增加一台搭载 1800cc 直列六缸的休旅车，整个产品线会变得更加完整。

今日凌晨，苹果正式发布了今年 WWDC 全球开发者大会的邀请函。WWDC 26 将于北京时间6 月 9 至 13 日举行。今年的 WWDC，除了会正式公布「27」系列系统之外，开发侧重点仍将集中在 Apple Intelligence 后续进展、体验优化等等。自然，其中也少不了一些苹果针对 AI 时代的软件生态的提前布局。图｜MacRumors爱范儿今年同样将前往 WWDC 现场，为大家带来 WWDC 系列活动的专题报道，届时请锁定爱范儿主站、微信公众号、小红书和微博上的系列内容。iOS 27：玻璃、专业相机和 AI Siri对于 iOS，我们首先需要明确的是：液态玻璃（Liquid Glass）设计在 iOS 27 中不会退场，它就在那里，稳稳地接住你。就像刚更新的 iOS 26.5，苹果对于液态玻璃效果的微调也将持续到 iOS 27 里，直到大家习惯或者接受它的视觉风格。除了继续优化液态玻璃的视觉效果，iOS 27 也被很多媒体和开发者预测为是一次「雪豹」式的更新——比起新功能，删除陈旧代码、优化性能表现、调整视觉效果和改善整体流畅度将成为这一代 iOS 的主要更新方向。在支持范围方面，iOS 27「据称」将会支持到 iPhone 12 系列和 iPhone SE 3 为止。换言之，iPhone 11 系列及更早的机型将无法更新 iOS 27。而在继续优化液态玻璃的同时，iOS 27 还有可能会在一些小细节的层面，为我们「管中窥豹」一下今年 9 月份即将到来的折叠屏 iPhone。图｜MacRumors当然，苹果肯定不会在 WWDC 上直接展示有关大屏和横屏布局的优化措施。但在后续的开发者课堂中，大概率会出现有关自适应 app 布局、不同长宽比 UI 设计规范之类的相关内容。几乎可以肯定的是，今年 WWDC 上，苹果会在涉及到 iOS 的内容和课程里添加一些有关侧边栏和分屏（或者 Stage Manager）相关的内容，为横屏布局做出铺垫：另外，本次 WWDC 上我们还可能会见到苹果开放一些更加深层的相机 API，或者一些全新的相机玩法。这些本质上都是为 iPhone 18 Pro 上可能到来的相机「专业模式」做好准备。根据计划，苹果会让 iOS 27 的相机 app 更具自定义性，用户可以调整相机界的功能开关布局，比如闪光灯、EV、模拟光圈、照片风格和长宽比等等。图｜Shima Zamil最后不得不提的还有 AI Siri ——根据彭博社的预测，iOS 27 将会是 Apple Intelligence 官宣两年之后的「汇总落地」。iOS 27 版本的 Siri 中，苹果预计会拓展跨 app 信息提取、屏幕内容识别、模拟连续操作、app 深层功能集成等等多方面的 AI 能力，努力追赶甚至看齐 Gemini 在 Pixel 上的功能范围。图｜Google更有甚者，传说中的「对话式」Siri 也有希望在今年 WWDC 上亮相，或者至少看到一些相关的底层功能。相比以前那个问一句答一句的智能助手，Siri 将会获得类似 ChatGPT 那样的「人味」对话功能。并且在 Siri 对话中也支持多模态内容，甚至可能会取代原本的「聚焦搜索」功能：聚焦搜索（左）和 Siri 搜索（右）至于那些 iOS 26 里面已经存在的 AI 功能，比如图乐园、Genmoji、写作工具之类的，它们在 iOS 27 中也主要将迎来性能优化，并且在新的 Siri app 里面获得单独的入口。macOS 27：圆角、触屏和 Intel对于那些熬夜观看 WWDC 的 Mac 用户来说，看苹果整活王 Craig Federighi 怎样宣布今年的 macOS 系统名称，始终是不可或缺的一环。图｜MacRumors继去年 Tahoe（太浩湖）之后，苹果的商标注册列表上还剩下很多加利福尼亚州的特色。你觉得 macOS 27 的系统代号会是什么？和 iOS 27 一样，今年 macOS 27 的重点同样是继续打磨液态玻璃、优化系统性能，以及落地更多的 AI 功能。不过就像爱范儿之前提到的：macOS 26 的 UI 问题不止于液态玻璃，不统一的圆角、过于密集的小图标，以及辨识度过低的系统 app 图标，联手造成了 macOS 26 的可读性问题。图｜Reddit除了赶紧优化系统 UI 之外，macOS 27 的另一大潜在看点则是它对于触屏功能的支持。彭博社苹果专家马克·古尔曼去年和今年都曾密集爆料过，苹果正在计划开发一款采用触控 OLED 屏幕的 MacBook Pro，上市时间可能在 2027 年内。图｜9To5MacmacOS 27 虽然不会在 WWDC 上就大张旗鼓地宣传「新增触控支持」，但大概率会按照苹果的风格，做出一些为触控功能铺路的举措——比如新增多种尺寸的 UI 控件规范、优化 macOS 支持的缩放比例范围、新增一些看上去不太适合鼠标键盘的图标动画等等。AI 方面，macOS 27、iOS 27 和 iPadOS 27 预计都将迎来一次「相册」更新，照片编辑器里的 AI 工具不再只有橡皮擦，大家习惯的 AI 扩图、自动优化、重构图甚至 3D 照片等等都有望加入。图｜AppleInsider至于 AI Siri 本身的功能，比如「对话式」沟通、屏幕内容识别、系统级自动化功能等等，macOS 27 能够得到的更新和 iOS 27 差不多。这也是 iPhone、iPad 和 Mac 处理器同源带来的优势之一。坏消息是：macOS 27 极有可能会排除所有使用 Intel 处理器的 Mac 产品，仅支持 Apple Silicon 型号的 Mac。这就意味着 2019 款 16 寸 MacBook Pro、2023 款至强处理器 Mac Pro 等等顶级产品都将无缘 macOS 27，为苹果的英特尔时代画上第二个句号。图｜MacWorld另有开发者指出，苹果也在逐步停止 Rosetta 2（x86 到 ARM 转译）的更新。macOS 27 预计将是最后一个完整支持 Rosetta 2 的版本，敦促第三方开发者及时更新 ARM 架构的 app。稳扎稳打的 27 年总的来说，和谷歌努力推荐的「Android 全形态化」不同，苹果今年 WWDC 的路线依然是以稳健为主。iOS、iPadOS 和 macOS 各自的领域划分很明确，虽然产品形态更新，但系统短期内不会出现类似谷歌 Aluminum OS 那样的融合现象。图｜Geeky Gadgets对于「液态玻璃」来说，在缝缝补补一年多之后，苹果的确需要静下心来，先处理一些由于玻璃 UI 和 AI Siri 烂尾导致的系统基础体验问题了。至少对于一些旧设备来说，优化后的 iOS 27 很可能会比 iOS 26 更值得更新。附 WWDC 26 日程表，感兴趣的读者可以通过苹果开发者官网（https://developer.apple.com/cn）注册并参与在线课程：

1995 年，微软内部搞过一次颇具乌托邦色彩的尝试。他们觉得当时的 Windows 系统操作太复杂，于是想把电脑桌面直接做成一个普通人看得懂的「房间」。你想打字就点桌上的笔记本，想看时间就点墙上的挂钟，不用思考那些反直觉的操作路径。这个产品叫 Microsoft Bob，结果上市没多久就光速下架了。原因五花八门，但究其根本，它并没有真正渗透进系统底层，只要用户的需求稍微复杂一点，就只能无能为力了。三十年后，很多 AI 助手依然没有完全走出这个困境。最近我体验到了腾讯新的 AI 助手 Marvis 马维斯 ，终于看到了一些不同。相比于别人都在卷谁是 AI chatbot，Marvis 想干的是直接把整台电脑变成你可以随意使唤的搭档。官方给它的定位也正儿八经，叫「操作系统层级的个人 AI 助手」（官网：marvis.qq.com）。不过，当我知道它背后是那个做了 14 年应用宝的团队时，就不觉得意外了：这种长期在 PC 和移动端底层生态里打拼的出身，恰恰意味着它不仅懂 AI，更懂设备和系统。目前，Marvis 已率先覆盖 Windows PC 与 Android 手机端，iOS 和 macOS 版本也在推进中。APPSO 也在评论区准备了一些福利邀请码，手慢无。六个 Agent，一支随时待命的私人 AI 团队安装过程不算复杂，但对设备有点要求。Windows 端最低门槛是 8 核 CPU、16GB 内存加上固态硬盘。而且据官方透露，未来还会逐步兼容 6 核。市面上很多 Agent 产品动辄让用户自己去搭工作流、安插件，Marvis 则是开箱即用，六个 Agent 零配置上手，对普通用户非常友好。刚进界面，里面已经有六个 Agent 在 7×24 小时待命了。六个 Agent 都有自己的工位。没接到任务的时候，有的在打盹，有的在办公室里闲逛，有的在健身，有的在喝咖啡，还有的直接去上厕所，一旦你发出指令，任务就会拆解分配给对应的成员。这是一个分工明确的团队：PM 负责听懂你的需求，拆解任务往下派活；File Agent 处理文件的搜索、阅读和格式转换；Computer Agent 专门对付系统配置、查硬件、调夜间模式；APP Agent 负责在各种软件里代为操作，比如查个机票或者电商比价；Search Agent 负责快速从公开资源中定位答案并给出关键引用；最后是个 Browser Agent，定位是网页交互专家，进行网页交互与数据抓取。任务链比较长的时候，看看这些动画也是别有一番趣味。最有意思的是，这些 Agent 可以并行工作。举例而言，我先抛出一个需求：「电脑开机太慢了，帮我看看哪些自启动程序没用，帮我关掉。」注意看，工作日志里还会显示 Token 消耗量，每天的免费额度是 2000 万 Token，堪堪够用。结果显示，它给每个程序配了「人话版」说明，建议你哪些是可以按需、保留的第三方软件。重点是，涉及这种核心配置修改的操作，它会主动和你确认一下，才会执行。这种「AI 不会自作主张」的机制，稳得让人心安。电脑用久了，总会碰到几个反复折磨人的小问题。「查看电池健康状况」「网速感觉变慢了，帮我测一下现在的上传下载和延迟。」「想玩个新游戏，帮我检测一下硬件配置，对比最低要求看看差在哪。」以前遇到这些事，要么去开始菜单里找设置，要么去网上搜教程。现在，说一句话就行了。Marvis 深度对接了系统信息和配置入口。它能看到硬件参数，也能直接修改系统设置。它不是帮你打开「控制面板」然后让你自己找，而是直接把事情办了。对于那些不熟悉电脑设置的人来说，这块的体验是最直观的。举个例子：你想关掉 Windows 锁屏上的广告，你知道它在设置里叫什么吗？叫「关闭 Windows 聚焦」，有多少用户知道这个术语是关广告的？但 Marvis 能帮你一句话关掉。本地文件时代，迎来 AI 入口为了测它的能力，我还翻出了一件困扰已久的事：两个月前从某个群里下载的一张截图，只记得内容是关于《黑神话：悟空》，文件名完全忘了。我输入：「帮我找一下关于《黑神话：悟空》的图片或截图，大概是一个月前的事儿。」说实话我没抱太大希望。但 Marvis 接单后，首先根据 Skill 指引，按文件名关键词和时间范围筛选出 1000 张照片，然后用 Python 快速筛选整个结果文件，并最终找到 8 张候选图。第一张正是我要找的那张。而且它的搜索范围极其宽泛：文件名、文档内容、图片里的景、截图里的字，全覆盖。哪怕描述模糊，它也都能捞出来。这个能力，比 Windows 自带的搜索确实强出了一大截。当然，一听到「扫描图片」，可能会有人担心隐私。首次使用时你可以自己设定扫描范围。如果在生成的图谱里看到不想被索引的内容，右键一键屏蔽即可。而且端侧模型本身自带基础过滤，不是无差别的扫描。除了找文件，Marvis 还能对本地的文档、表格等多种类型文件做深度理解和分析——支持内容优化、图表生成、文案润色、格式转换，帮原本费时费力的事情变得顺手。比方说，下载了一批学术论文，摞在那里不知道从哪里读起。直接问 Marvis：「扫一遍这几篇 PDF，告诉我每篇的研究问题、方法和主要结论。」几分钟内，哪几篇值得细读、哪几篇可以先放着，心里就有数了，不用一篇一篇翻摘要。再比如读完《三体》，笔记记了好几页，马上要在课上分享，手动整理提纲、挑内容、套 PPT 模板，少说也要折腾一两个小时。但让 Marvis 根据读书笔记，生成一份 PPT，它能读懂笔记里的逻辑和层次，不需要手动复制粘贴再套模板，理解内容之后直接生成。手机接管电脑，随时随地都能干活手机遥控操作电脑，则是非常实用的跨端体验。去官网下了 Windows 客户端，再给手机安上 Marvis App，用同一个微信或 QQ 登录，还能直接连通两端。具体来说，手机 App 连上电脑后，可以在手机上实时看到电脑桌面，并用触屏直接操控。哪怕电脑处于锁屏状态，手机依然可以接管，不需要电脑一直保持解锁状态。比如我人在外面，突然听到朋友说前阵子新出了一款《洛克王国：世界》游戏，这时我就可以掏出手机输入：帮我在电脑上安装《洛克王国：世界》，回到家就能玩上了。连接微信 ClawBot，也是没问题的，微信扫码登录后就能开启对话。但我认为最实用的场景，是给家里长辈远程排查电脑问题。爸妈说「QQ 打不开了」，以前要么电话里盲猜，要么回家处理。现在直接接管他们的桌面，扫一眼就知道问题出在哪。前提是你的手机端 Marvis 马维斯，要能连接爸妈家里的电脑端马维斯，手机端是支持连接多个 PC 设备的。而如果电脑断网了，手机端会自动切到云 Marvis 模式，支持独立调用云端能力，大部分任务依然能跑。并且这里还有一个独家绝活：在电脑上直接操控手机里的 App。同花顺、飞常准等应用已经在对接合作，这也得益于应用宝团队多年的跨端引擎积累。现在应用宝在移动端、pc 端、车载上都有支持，跨端研发经验充足。和其他桌面 Agent 一样，Marvis 也支持自定义自动任务。比如设置一个「深圳厚德品园二手房均价提醒」，填好任务说明——「查询最新二手房均价并通知用户」，再指定执行时间为每周一上午 10 点，之后什么都不用管，到点自动运行任务，并直接推送结果。另一个值得单独说的功能，是技能（Skill）的自定义扩展。Marvis 支持一键安装专属技能包，提供了不少第三方 Skill，当然，你也可以导入自己常用的 Skill 文件，把它改造成更贴合个人工作流的样子。有特定行业需求的用户，比如内容创作，理财投资、学术研究等等，不用依赖官方迭代，自己就能把对应的知识库和操作逻辑装进去。最后，主打出活的 Agent 都绕不开最关键的问题：安全。Marvis 马维斯提供了两种模式：效率模式和隐私模式。设置里切到「隐私模式」，Marvis 会加载本地端侧模型，无论是文档还是图片的检索、解析、识别等，全部在本地运行，文件不会出你的电脑，甚至断网了都可以使用。官方所说的「数据不出域」，在拔掉网线的那一刻，有了最真实的体现。不过，体验「隐私模式」的门槛比较高，要求 16 核以上的处理器、32GB 内存、16GB 显存起步，固态还得匀出 35GB 空闲。所以平时日常使用切回「效率模式」即可，理解和规划由云端大模型（混元和 DeepSeek V4）完成，执行在本地，速度更快。两种模式按需切换。让 PC 真正开始「理解」人我们现在用电脑的方式，本质上还是上世纪八十年代的逻辑：一个需求对应一个软件，你要去找它、打开它、学习它的规则来完成任务。四十年来，软件更多，设置更复杂了，但这套逻辑并未改变，普通用户和电脑之间的距离，也没有真正缩短。Marvis 的逻辑反过来了：先理解这台电脑里有什么，再决定为你做什么。拿我爷爷来说，他用了几年电脑，至今不知道「控制面板」在哪。但他知道电脑开机慢，知道想找一张老照片。过去，这个解法是打电话求助子女；现在，就算他不会打字，也可以直接在微信 Clawbot 或者手机马维斯上，把需求说出来，电脑就会帮他做。文件搜索、系统设置、跨端控制、自动任务、隐私模式、Skill 扩展，看上去是几个不同功能，背后指向的其实是同一个方向：个人 AI 不应该只是一个问答入口，而应该成为设备里的任务调度层。多说一句，我觉得 Marvis 是有审美的。虚拟办公室动画、任务进度呈现等细节，让它不像冷冰冰的 AI 工具，更像一个有秩序、有分工、也有生活气的小型工作室。它尤其适合几类人：文件多且乱的本地重度用户；重视数据隔离的财务、法律、医疗从业者；喜欢自定义工作流的效率玩家。和需要靠「养」才能用好的「虾」和「马」，Marvis 初始内置 Agent 和自然语言指令，也足够覆盖多数日常电脑操作，对于很多 AI 焦虑不知道怎么用的人，Marvis 都更加的平权、无门槛、开箱即用。未来的个人 AI 会像搜索、截图一样，自然地成为操作系统的一部分。三十年前，Microsoft Bob 试图让用户更容易理解电脑；三十年后，Marvis 试图让电脑反过来，真正理解用户。Marvis 马维斯能走到这一步，和这个有十四年应用商店生态积累和四年跨端跨系统技术积累有关。庞大的用户基数让他们更懂用户在不同设备上需要什么、怎么使用。这种积累很难被纯 AI 公司在短期内平替——他们可以快速训练一个更聪明的模型，却没法快速建立起对 PC 底层和移动生态的控制力。至此，一个能用自然语言驱动的系统级 AI 助手，在中文生态里终于有了真正可用的样本。所谓个人 AI 重点从来不只是 AI，更是个人。Marvis 终于让我们看到，当技术选择与每一个具体的需求并肩同行时，它所释放出的力量，既润物无声，又重塑天地。这是个人 AI 真正开始变得有用的时刻。

「智能队列」如何改善你的家庭网络质量主作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org联合作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org2021/01/31 23:1205/17 12:57更新：2026 年世界电信和信息社会日重推在无线路由器的设置页面上，我们可能会看到一个名为「智能队列」的功能（根据品牌或操作系统的不同，也可能会被称做「智能 QoS」、「SQM」等）。这一功能是如何工作的？是否能让网速变快？是否需要打开？本文将尝试解答这些问题。从「网络如何连接」说起本章节主要介绍计算机网络的基础知识。如果读者已经熟悉相关知识，可跳过这部分内容。在介绍「智能队列」之前，先来了解一下基础的网络知识。两台电脑间进行数据传输，存在两种方式，分别是电路交换和分组交换。对于电路交换，两台电脑独占一条线路进行数据传输。过去的固定电话，大多使用的就是电路交换。由于独占一条线路，两台设备之间的通信较为稳定，很难出现数据丢失、数据传输顺序错乱等情况。但一条线路只能供一组电脑使用，另外一组电脑想要通信，就不能复用原来的线路了。对于分组交换，传输的数据被拆分为小段，每一小段叫做一个数据包（报文）。数据以报文的形式进行转发，从而实现了多组电脑复用一条线路进行通信。电路交换虽然能保证数据传输的稳定性，但由于独占线路浪费太多资源，所以在 Internet 中，使用的是分组交换。分组交换将数据拆分成一个一个的小片段来转发，没有独占线路。如果网络条件不好（例如同时有大量报文经过路由器，超出了路由器的处理能力），很有可能遇到数据顺序错乱、甚至数据丢失等情况。那么，这些情况应该如何避免？在微信视频通话、电视直播等应用中，报文的丢失或乱序，只会造成视频图像卡顿，不会造成太严重的影响。所以这些应用会使用简单的UDP协议，不考虑丢包、乱序。在网页浏览、文件下载等应用中，一般使用基于TCP的HTTP协议。TCP 在报文中携带序号，避免乱序；同时提供了报文的丢失重传机制，避免了丢包。所以，也可以认为 TCP 为两台电脑的应用程序建立了一个虚拟的「连接」。（最新版本的 HTTP/3 协议，使用的则是QUIC。QUIC 虽然基于 UDP，但引入了与 TCP 类似的报文丢失重传机制）如果想要了解更多网络基础知识，推荐阅读「入网指南」系列文章：入网指南 01 | 一文读懂你身边的「网络」入网指南 02 | 当你发消息的时候，发生了什么入网指南 03 | 交换机在交换什么？入网指南 04 | IP 地址大揭秘入网指南 05 | 除了接线，还有什么影响着我们网上冲浪入网指南 06 | 连上 Wi-Fi 就能上网了吗？网络延迟：影响上网体验的重要因素带宽和延迟，是影响上网体验的两大因素。带宽指的是在单位时间内，网络最多能传输多少数据。网络的带宽和水管的粗细有着很多相似之处，水管越粗，相同时间流出的水越多；同样地，网络带宽越大，相同时间能够传输的数据越多。在日常上网过程中，大的带宽能带来这些好处：上传、下载文件的速度更快在线看视频时，缓冲的时间更短如果家中有多个人同时上网，大的带宽能够保证每个人都能流畅上网，相互的影响更小对于家庭宽带，我们常常听到的 10 兆、50 兆、100 兆、千兆，就是指的带宽的大小。延迟指的是数据从网络的一端到达另一端，所需要的时间。过大的延迟也会对上网体验带来不好的影响，例如：在浏览器中输入网址敲回车后，需要等更长的时间，网页才能加载；视频聊天、语音通话质量变差，甚至能直接感受到声音、画面的滞后；玩在线游戏时，发现游戏变慢变卡。在很多时候，延迟相比带宽更容易被人感知，对网络体验的影响更大。是什么导致了网络延迟？上文中介绍过，Internet 使用的是分组交换。如果大家阅读过计算机网络相关的书籍，应该对分组交换的「存储-转发」机制有一定的印象。简单说，「存储-转发」是指网络设备将收到的报文存储在缓冲区，然后再发送。这是由于网络设备的性能有限，转发报文有速率限制。如果设备短时间内收到大量报文，一时半会儿处理不过来，就可以先报文存起来，空闲时慢慢发送。通过「存储-转发」，能够提高网络设备的吞吐量，更充分地利用带宽。其中，存储报文缓冲区，是一个先进先出的队列。那么，队列的长度应该如何确定？是不是队列越长，能够存储的报文越多，网络的性能就越好？大家可以先思考下这个问题。其实在过去，内存价格比较昂贵，考虑到成本，网络设备的队列不会太大。现在，随着内存价格的下降，网络设备拥有了更大的队列。但是，网络的带宽是有限的。实际的可用带宽，取决于数据传输路径上的最小带宽。这个最小带宽也被称做带宽上的「瓶颈」（bottleneck）。由于瓶颈的存在，如果流量长时间以较大速率发送，大的队列并不能显著提升吞吐量。相反，大量报文滞留在队列中，会带来延迟的增大。这种队列中缓存有大量报文，来不及发送，导致数据传输延迟增大的现象，被称做 bufferbloat（缓冲区膨胀）。DSLReport 提供的测速服务，能够检测网络的 bufferbloat 程度。不过由于没有中国大陆服务器，测试结果不够准确。国外的读者可以尝试一下：DSLReports Speed Test家用无线路由器中的 QoS 功能在「智能队列」出现之前，很多家用路由器已经提供了 QoS 功能。那么，传统的 QoS 如何解决网络延迟、网络拥堵？存在哪些不足之处？在了解这个问题之前，先来熟悉一下网络中的报文，是什么样子的：在网络中传输的报文，一般包括了链路层、网络层、传输层、应用层这四层信息。其中，网络层的源 IP 地址、目的 IP 地址、协议号，和传输层的源端口号、目的端口号，是 QoS 算法中常常会关注的五个信息。这些信息合称「五元组」。同一个应用在同一时间段内，传输的报文中，这些信息一般是不会发生变化的。例如观看在线视频时，同一个视频的所有报文 ，一般拥有相同的五元组信息。所以，通过五元组，我们能够区分不同应用的流量。特别地，一些常见的应用，会使用固定的端口号。例如网页浏览一般使用 80、443 端口；英雄联盟游戏，使用的是 5000-5500 端口 ……所以，当我们想要提升英雄联盟游戏的体验时，在路由器上设置让 5000-5500 端口的报文优先通过即可：How to set up traffic prioritization rules for League of Legends on Nighthawk Pro Gaming router | Answer | NETGEAR Support家用路由器中常见的 QoS 功能，就是通过手动设置 IP 地址、协议号、端口号等报文特征来实现的。这种方式需要用户了解一定的网络知识，且操作复杂。（虽然部分无线路由器内置了常用应用的 QoS 规则，但仍需要一定的手动设置，且无法及时适配新的、小众的应用）当然，除了上文中介绍的内容，QoS 还有多种不同的实现方式。例如：对特定应用（例如 BitTorrent）进行限速按照 MAC 地址指定某台设备的优先级（例如提高 Nintendo Switch 的优先级，使游戏更加流畅）通过WMM、DSCP等机制来决定报文优先级这些机制各有各的特色和不足。例如对应用程序限速的方式，虽然效果明显，但无法在网络空闲时，动态利用剩余带宽来提升传输速度，也难以适应带宽频繁变化的无线网络。关于上述 QoS 机制的技术细节，本文不再详细介绍。如果感兴趣，可以搜索相关关键词，进一步了解。智能队列：通过「全自动」的方式优化网络、降低延迟什么是「智能队列」目前，一部分新的无线路由器，开始有了名为「智能队列」的功能。这一功能一般指的是 fq_codel。fq_codel 包含两种队列管理机制，分别是 Flow Queue 和 CoDel。在下文中，我将简要介绍这两种机制的作用。Flow Queue前文中介绍过，网络中的报文一般都有自己的五元组信息；同一应用发送的报文，在一定的时间段内，五元组信息是相同的。这些信息相同的一系列报文，被称做一条流（flow）。Flow Queue 为每条流单独维护一个队列。例如，如果我们一边在线观看视频，一边进行语音通话，一边玩游戏，Flow Queue 就会为视频播放的流量、语音通话的流量、在线游戏的流量各维护一条队列：正常情况下，在线观看视频需要占用较大的带宽，但对延迟不敏感，用户可以容忍几秒甚至几十秒的缓冲时间。而语音通话和在线游戏占用带宽小，但对延迟敏感。例如语音通话时，如果一个人说了一句话，对方在几秒后才能听到，用户就能明显感知到延迟。如果三种应用共用一个队列，队列中会充满大量视频播放应用的报文，导致语音通话、在线游戏的报文迟迟无法发送，增大了这两种应用的延迟。Flow Queue 为视频播放、语音通话、在线游戏各维护一条队列，并公平地调度三个队列，避免视频播放的报文占用太多时间，从而让在线游戏、语音通话的报文得到及时处理。由于 Flow Queue 确保了应用程序之间的「公平」（fair），所以又被称做 Fair Queue。CoDel我们已经知道，适当长度的队列，能够充分利用带宽、提高吞吐量。但过长的队列，又会导致过大的延迟。也就是说，队列既不能过长、也不能过短。CoDel 所做的，就是让队列维持在合适长度，最大限度地提升网络性能。CoDel 通过计算报文在队列中的停留时间，来判断队列是否过长。具体做法如下：报文进入队列时，在报文上记录当前时间（也就是报文入队时间）报文出队列时，再次获取当前时间（报文出队时间）计算报文入队时间与出队时间的差值，即可得到报文在队列中停留的时间当报文在队列中的停留时间超过阈值，说明队列过长，需要缩小队列长度当队列过长时，CoDel 会直接丢弃出队列的报文，从而减少队列中报文的数量，达到控制队列长度的效果。（这种对报文的丢弃，并不会对网络带来负面影响，具体请见后续章节）可以看出，CoDel 根据队列中报文的停留时间，来决定是否丢弃报文。而不是「一刀切」地将队列限制在一个固定的、较小的长度。这样既能降低延迟，又保留了队列最大的优点：吸收短暂的大流量报文、提升吞吐量。另外，我们的设备，尤其是手机等移动设备，所处的网络环境是复杂多变的。例如手机靠近或远离无线路由器，就有可能导致网络带宽的变化。CoDel 通过计算报文停留时间的方式控制队列，能够更加灵活地适应网络带宽、网络质量的变化。fq_codelFlow Queue 确保了不同应用程序的流量能够公平传输。CoDel 控制队列长度，实现了整体延迟的降低。fq_codel 就是将 Flow Queue 和 CoDel 结合在一起使用的机制。其中，Flow Queue 先将不同的流划分为不同的队列，再为每一个队列使用 CoDel，控制报文的延迟。两者结合使用，共同提升网络体验。fq_codel 具有免配置、天然适应不同的网络环境、实现简单、消耗资源少等特色，越来越多的无线路由器（以及开源的路由器操作系统），开始内置这一机制。参考 KoolShare 上的这篇帖子，作者经过测试，发现打开 fq_codel 之后，在 BT 下载占满上行带宽时，在线游戏依旧保持较低的延迟：可能是家庭路由里最好的 QoS——在线游戏者的福音 - KoolShare（见 8 楼、10 楼、22 楼、25 楼的测试结果）是否需要打开「智能队列」？前面介绍了 fq_codel 的一系列优点。那么，fq_codel 有没有缺点？是否需要我们打开？我的设备 CPU 性能不强，是否可以打开「智能队列」？根据这篇帖子，ER-X路由器在打开 fq_codel 时，最大支持 200M 左右的带宽。QoS (fq_codel/cake) Hardware Recommendation : HomeNetworkin

周一18派早报：提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年前天 06:56早报速览提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年0OpenAI 与苹果关系紧张，或考虑诉讼5索尼回应 AI 相机助手越改越差，称只是修图建议2Anthropic 警告私下交易的股权可能无效2运营商回应「零月租」传言4微软重新允许自定义任务栏与开始菜单2看看就行的简讯4少数派的近期动态0你可能错过的好文章0提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年5 月 15 日，预印本平台 arXiv 出台新规：如研究人员向该平台提交的稿件包含「无可辩驳」的 AI 生成内容，论文所有署名作者将被处以长达一年的封禁；且其未来提交的任何文章都必须先在其他期刊通过正式的同行评审后，arXiv 才会予以收录。「无可辩驳」的证据包括虚假参考文献、不小心粘入的与 AI 对话内容等。俄勒冈州立大学名誉教授、arXiv 编委兼审核团队成员 Thomas Dietterich 在 X 上公布了这一消息。他指出，新规直接源于 arXiv 既有的审核标准。无论稿件中出现不当言论、抄袭、学术偏见、虚假参考文献还是误导性内容，所有署名作者都必须对这些 AI 生成的错误承担全责，而不能将责任推卸给工具。arXiv 是物理学、天文学、数学及计算机科学等领域全球最大的开源学术文献库。在这些学科的常规发表流程中，学者们通常会在接受期刊正式的同行评审前，先将论文作为预印本上传至 arXiv，以抢占首发权并获取同行反馈。因此，剥夺一年的首发资格并永久附加同行评审门槛，对相关领域人员而言是严厉的措施。375OpenAI 与苹果关系紧张，或考虑诉讼据彭博社报道，苹果与 OpenAI 为期两年的合作关系近期出现裂痕。因未能从 ChatGPT 与苹果系统的整合中获得预期收益，OpenAI 正与外部律所积极探讨应对方案，近期或将向苹果正式发出违约通知，为潜在的法律诉讼做准备。2024 年 6 月，苹果宣布将 ChatGPT 接入 Siri 等系统级功能。OpenAI 曾寄望借此获取海量订阅用户及数十亿美元收入。然而实际体验中，该功能入口较深，需用户明确口头提及 ChatGPT 才能唤醒，且信息呈现方式受限。结果，绝大多数用户仍倾向于使用 ChatGPT 应用。一位未具名 OpenAI 高管表示，苹果在产品端甚至没有尽到诚实努力的义务；合作不仅未能创收，反而损害了其品牌形象。除软件合作遇冷外，两家公司在硬件与人才领域的博弈也加剧了紧张局势。OpenAI 此前收购了由前苹果首席设计师 Jony Ive 等人主导的初创公司，并利用高薪从苹果硬件团队挖角，试图打造新一代 AI 设备，这引发了苹果高层的强烈不满。同时，苹果也已着手引入竞争机制，计划在即将发布的 iOS 27 系统中全面开放生态，接入谷歌 Gemini 和 Anthropic 的 Claude 等其他主流大模型。知情人士表示，OpenAI 针对苹果的任何法律行动，预计都会推迟到其与马斯克的诉讼案结束后才会推进。目前各项计划尚未做出最终决定，OpenAI 内部仍保留了在庭外解决争端的希望。OpenAI 与苹果的发言人均拒绝就此事置评。我换到国产手机AI都没怎么用过，不是不会用，而是习惯了以前苹果装个AI软件的使用方法了。果子在ai上的每步操作都很迷惑感觉不如自己装一个本地小型LLM App用ShortCut打通生态。我还是关心什么时候有苹果智能，感觉一不留神，好几年过去了，我换15p的初衷之一就是给智能做准备，现在都要17了，我也是吃上大饼了623索尼回应 AI 相机助手越改越差，称只是修图建议近日，索尼最新旗舰 Xperia 1 VIII 的官网页面引起关注。在演示 AI 相机助手的效果时，官方样张的效果在观感上反而明显劣于原图。这引来了 MKBHD、Nothing 创始人裴宇等人的公开调侃。对此，索尼在 5 月 15 日解释称，「AI相机助手」并不会在拍摄后自动应用修图，只是根据当前场景和拍摄主体，提供 4 种不同创意方向的参数设置建议，用户可自由选择接受这些建议或坚持使用自定义设置。与此同时，官方最新展示的对比样张在视觉效果上有了显著改善，去除了首批样张中饱受诟病的画面发灰和过度曝光问题。但这依然未能完全消除外界对其影像审美判断的疑问，也不清楚营销团队最初为何会挑选那些效果不佳的照片来展示核心卖点。感觉那朵花还可以，其他的都是💩️有人关注就算不错了10122Anthropic 警告私下交易的股权可能无效据《华尔街日报》报道，Anthropic 近日更新官网公告，宣布将不承认未经其董事会批准的二级市场股权交易，并点名了一系列未授权机构。Anthropic 警告称，由这些机构提供的任何该公司股权或相关权益转让均属无效，且不会记录在公司的股东名册上。这一声明在个人投资者中引发了广泛恐慌。由于普通投资者通常无法直接购买未上市「明星公司」的股权，大量二级市场交易平台应运而生。它们设立特殊目的实体（SPV），将众多散户资金打包后投资于目标公司。然而，这种 SPV 结构日益复杂，有的多层嵌套，有的把多家目标公司的股权捆绑出售，导致分配机制存在很大模糊地带。普遍认为，Anthropic 此举意在通过清理违规交易来精简其股东名册，为其计划于今年晚些时候的 IPO 做准备。随着近年来私营科技公司的估值不断飙升，包括 OpenAI 在内的多家独角兽企业，都在采取类似措施限制涉及其股权的二级市场交易。面对投资者的质询，多家被点名的交易平台迅速回应，表示涉及 Anthropic 的基金和交易均已获得批准或符合转让要求。专家指出，如果 Anthropic 并未在事前对特定交易结构做出限制，可能无权撤销这些持仓，相关争议最终或将诉诸法庭。可能性不大这种模式在中国有么？发表评论运营商回应「零月租」传言据央广总台报道，近日，不少互联网博主在短视频里介绍，三大运营商即将取消月租费，改为按量计费模式，甚至还列出了推出该政策明确的时间节点。但记者梳理市场现状发现，目前只有中国联通上线了零月租、设最低消费门槛的「魔方」按量套餐；中国移动和中国电信暂未推出同类产品。例如，传言中被广泛提及的「联通魔方」发布于今年 4 月底。该产品采用纯按量计价模式，国内语音和短彩信均为 0.1 元/分钟或每条，流量则实行阶梯计价，单价在 1 元至 1.5 元/GB 之间。但联通客服表示，该套餐并非绝对免费，而是设有地区差异化的最低消费门槛，例如北京地区起步低消为 55 元，月结时不足该金额将自动扣齐。同时，联通原有的各档位固定月租套餐依然保留并可正常办理。中国移动和中国电信的客服人员则都表示，公司曾推出过零月租模式，但目前没有相关计划。针对完全零月租的可能性，通信行业专家指出，此类模式脱离了基础通信网络的运营逻辑，庞大的固定网络维护成本最终也必然会转嫁到其他用户身上，甚至可能导致整体资费上涨。比起追求不切实际的零月租噱头，当前消费者对资费的真正痛点在于不合理的老旧月租套餐，以及套餐外的标准流量资费。没有专门办的流量卡划算捉虫：“央广总台”→“总台央广”“零月租” “起步低消”世界需要提速降费。3微软重新允许自定义任务栏与开始菜单5 月 16 日，微软发布博客文章，宣布将在未来几个月内推出一系列更新，大幅增强任务栏和开始菜单的外观自定义功能。这标志着一系列经典 Windows 功能的回归。对此，Windows 业务负责人 Pavan Davuluri表示，个性化定制是 Windows 的基因；微软已听取用户反馈，决定重新将界面控制权交还给用户。目前，部分选项已在 Windows 预览体验计划中面向测试人员推送。本次更新中，用户可以再次像在旧版系统中那样，将任务栏停靠在屏幕左侧，或将其缩窄以节省屏幕空间。同时，开始菜单新增了尺寸调整选项，并允许用户自定义显示模块，例如直接关闭「所有应用」或「推荐」（更名为「最近使用」），仅保留自己固定的应用。此外，用户可以隐藏「开始」菜单中的用户名和头像。这些测试功能目前仍有已知问题，例如开始按钮未对齐、通知弹出位置异常等。微软表示将在正式向公众推送前修复这些问题。此外，更多传统的自定义功能，例如拖拽任务栏边缘手动调整高度等，目前也已纳入开发计划，预计将在今年内陆续上线。Windows 11 自发布以来，就因移除任务栏位置调整等诸多经典自定义功能而饱受争议。近年来，微软一度将系统开发重心全面转向 AI，在系统中强行植入 Copilot 功能，引发了用户的强烈抵触。迫于压力，微软于今年早些时候宣布做出调整，重新关注质量。什么时候把win10菜单给我回归一下😊我选择Windows Server 2025，躲过UWP和React两大石山。1看看就行的简讯据Fast Company报道，知名开源密码管理器 Bitwarden 近期在未发公开声明的情况下，悄然大换血高管团队，今年 2 月起，原 CEO、CFO 先后转岗或离职。新任 CEO 有并购及私募股权交易背景，让外界对其未来的资本运作产生猜测。此外，Bitwarden 官网出现多处调整，个人版产品页面移除了「永久免费」的标签（官方解释为营销团队疏忽）；不久前，Bitwarden 在未显著通知的情况下，将高级版订阅年费从 10 美元翻倍至 20 美元。此外，Bitwarden 还在招聘页面修改了企业核心价值观，去除了原有的「包容」与「透明」两项原则。近日，谷歌开始测试一项新的云存储空间政策：新注册的用户如果不绑定手机号，其初始免费存储空间将从传统的 15GB 缩水至 5GB。谷歌官确认，该政策目前仅处于区域性测试阶段，且现有老用户不受此规则影响。谷歌发言人表示，这项新规是为了继续为用户提供高质量的存储服务，同时鼓励用户完善信息以提升账户安全性和数据恢复能力。然而，许多用户推测谷歌此举的实际目的是为了打击恶意注册大量小号以薅取免费存储空间的行为；也有部分用户对强制绑定手机号背后的隐私数据收集及潜在商业用途表示担忧。据9to5Google报道，谷歌随无屏健身追踪器 Fitbit Air 推出的订阅服务「健康教练」（Health Coach，由原 Fitbit Premium 升级而来）存在 AI 幻觉问题。测试中，AI 教练凭空捏造了一项 5.2 英里的跑步记录；在被指出错误时，还反问用户是否忘记输入过该次记录。经过纠正后，这条虚假数据依然残留在应用主页上。最近对bitwarden挺失望，要么是性能问题，这次又出个管理团队问题。开源不是万灵药，信誉才是基石啊这几年一直用Bitwarden 呢，要变味了吗白赚一次跑步，哈哈哈哈比特瓦恩德可以换了(｡･ˇ_ˇ･｡:)4少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购发表首评你可能错过的好文章看剧不卡顿、游戏低延迟：「智能队列」这样改善你的冲浪体验为什么我觉得好的东西，媒体评测却给了低分？本周看什么 | 最近值得一看的 8 部作品装了啥：编辑部用什么转写音视频内容？[会员免费] Apple WWDC26 全球开发者大会，来与少数派一起看快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验你真的需要墨水屏设备吗？发表首评90扫码分享目录 0讨论我来说一句发布发表评论发布

2025 年度旅行——4 天 3 晚青海湖小环线主作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥联合作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥05/14 09:38Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。一次完美的旅行，对我而言，不仅仅是去被人走过千遍的地方打卡一圈，而是藏着三次截然不同的体验——出发前对目的地的想象与期待，旅途中与自然的亲密接触，以及归来后翻看照片时，重新发现那些当时未曾留意的细节，细细回味，查阅资料，让我对这次旅行有更深刻的理解。那么就让我们随着这些文字与照片，回到美丽的青海湖。契机：作为刚工作的年轻人，在国内的职场环境下，真正无理由可申请的假期只有 5 天，所以我和身边的小伙伴们都分外珍惜这五天的时光，每年的年度旅行我们都会避开人流，详细规划。由于十一假期要去参加两场婚礼，为了节约交通费，我们决定先开车到 2 号参加婚礼的宝鸡，把小 P（我的小鹏 MONA M03，具体可参阅我之前的文章）停在了宝鸡高铁站，考虑旅行时间较为紧凑，且西北自然景区交通补能条件可能较差，我们选择了到当地租车完成青海湖小环线行程。带着宝鸡擀面皮和肉夹馍，一行人正式开始了旅途，三个半小时后，到达了西宁火车站，初入高海拔地区，并未感受到明显不适，方便的是，租车平台在西宁站地下停车场就有取还车点，我们直接将行李放上车，省去了搬来搬去的劳顿。Day1 西宁-茶卡 全程约 457 km西宁-塔尔寺5A 级旅游景区，门票：70 元，游览时间约 2 小时塔尔寺是藏传佛教格鲁派的著名寺院之一，始建于明代，最初是为纪念藏传佛教格鲁派创始人宗喀巴大师而修建，据传宗喀巴大师出生于此地，后人于其诞生处建塔纪念，塔尔寺因此得名——「塔尔」即「有塔之地」。塔尔寺-八宝如意塔游览中心和寺庙门口都有大量的讲解人员，价格也不贵，但是本人信仰唯物主义，故只将这里当做名胜古迹参观，故没有请讲解，如果对藏传佛教及寺庙背后的故事比较感兴趣，请个讲解或许是更好的游览方式，或者是像我一样，回来翻阅照片时再仔细搜索相关资料。Tips：讲解会介绍各个佛像的「作用」（不知道这样叫是否合适，或者说负责保佑的事情），故可能会被架在那里为了讨好彩头而「必须」捐献资金，介意者可以像我一样自由游览。由于是节前，故参观人员并不很多，与前往朝拜的信徒或僧人可以说是一半一半，没有了熙熙攘攘的人群，也得以让我更加细致地观察这座寺庙和「磕长头」远道而来的人们。前来朝拜的僧人进入寺庙，浓郁的香火味和木质结构因缺乏通风的陈味扑面而来，看着宏大的佛像、墙上摆放整齐的诵经坐垫，的确能感受到不一样的文化氛围，这里安静、鲜活，尤其是看着不知道从哪里启程的「磕长头」的信众，虽因路途的遥远衣衫褴褛，但想必也经过这一路的艰辛而悟出了些许道理，可能出发时也不知道为什么，但走着走着就明白了吧。塔尔寺中的白塔募捐处，与时俱进地，也支持扫码支付塔尔寺有大大小小的寺庙很多间，在我看来有的都别无二致，于是重点转了几个后，就启程回西宁吃饭，整备氧气瓶，应急食物等装备，正式奔赴青海湖。海南藏族自治州共和县 南山观景台 & 青海湖二郎剑景区前者免费，后者为 5A 级旅游景区 门票 72 元，游览时间约 2 小时一点半在西宁市吃完炕锅后，我们开始驱车前往青海湖，约 2 个小时车程，我们到达了本地的主要目的地青海湖的南侧景区 青海湖二郎剑，在进入二郎剑景区前，可以导航到「南山观景台」 从山上远眺青海湖和远处群山的全貌，但是山路较窄且有工程车辆时不时通过，故要格外注意安全。景区本身没有人文景观，但其还是我国早期鱼雷的试验基地，但是因为我们到达时间比较晚，并没有参观博物馆，而是直奔到湖边。青海湖是我国水面面积最大的咸水湖，水面平静辽阔，与远处的群山相辅相成，甚是开阔美丽，栏杆上有硕大干净的鸽子驻足，在几番伦理拍照后，我靠在栏杆旁，看着远处的夕阳洒在地面，耳边传来阵阵鸟鸣与风声，工作与生活的喧嚣远离脑后，只剩下全身心的与自然接触。云层、远山、草原，形成了震撼的三明治拨云见日，青海湖边的草还是绿色的岸边排队等待「薯条」投喂的海鸥码头景观随着夜幕降临，我们离开二郎剑景区，沿着青海湖开车前往茶卡住宿，虽然只有 200 多公里的行程，但是因为沿途工程车辆很多，还有很多地方持续修路，有的路段有连续的 U 型弯，还有双向仅有一条车道通行的地方，还要翻越几个海拔 3800 米的垭口，成了这一路最惊险刺激的行程，一路上小心翼翼，战战兢兢，终于在晚上 10 点到达了茶卡镇，幸运的是镇上的餐馆都还在营业，找了家当地美食，简单吃饭后休息。夜幕降临的倒湖茶公路Day2 茶卡-祁连县 全程约 372km茶卡盐湖4A 景区 带往返小火车套票 167 元「茶卡」在藏语中意为「盐池」，该地以出产「大青盐」而闻名，参观盐湖也是此地的知名旅游项目。茶卡盐湖的景区共有两个，天空之镜是较老的景区，主要以自然风光为主，天空一号主要以商业化项目为主，作为外地游客，我们肯定更希望看到原始的自然景观，于是选择前往天空之境景区。景区很大，如果想要游览每个景点建议留出一整天的时间，游览顺序建议先坐小火车到终点天空之镜，该站处于盐湖的最深处，盐水异常的干净，也最为出片，景区也放置了很多拍照打卡点位，可以尽情拍摄，心痛没有带无人机。在晴天的时候，由于天空和地面都是白色，很容易晃眼，推荐大家一定要戴墨镜前往，通过墨镜观察湖面，盐湖会看起来像翡翠一样更加绿色透明。套票包含鞋套租赁，可以套上下湖走走，踩在大颗的盐里,很像沙滩的感觉，景区维护的十分干净，可以尽情的亲近自然。有趣的是，景区为了怕大家在盐滩大片的白色中无聊，设置了很多拍照用的摆件，例如国家地理的相框等等，还在大屏幕上贴心的轮播拍照指导，期望每个游客都能满意而归。盐湖与远处的山峰盐湖码头废弃的采盐轨道之前采盐的轨道深入盐湖，在墨镜下，真实的观感颜色会比照片深很多。接驳小火车行驶在盐湖中间，干净得不像国内巨幅盐雕从盐湖出来，我们在县城稍作休息，开始启程前往祁连县，有了昨夜惊险的行程后，尽管沿途的国道可能风景更好，但是我们选择了高速较多的线路。这一程的路况还是很好的，就算是国道车辆也较少，道路笔直，视野十分宽阔，草原、远山、白雪、天空层层分明。在行驶过程中，不时的有大群的羊羔和牦牛穿过公路，不必惊慌，一会就有牧民骑着摩托驱赶，让它们加快穿越的脚步。公路与群山草原上大批的牛羊傍晚的草原，远处能看到夕阳洒在雪山上落日下的牦牛剪影在夜色中，我们抵达了祁连县，在县城中吃过一顿特色烧烤后，进入梦乡，准备迎来第三天的行程。Day3 祁连县-岗什卡雪山-西宁 全程约 320km祁连卓尔山景区门票 60 元 上山小巴 20 元 登山电梯 10 元这个景区的确是全程我最惊艳的景区了，近能看到瑞士小镇似的田园风光，中能看到美如画的丹霞地貌，远还能远眺远处的阿咪东索雪山（海拔 4695 米），而且在山顶可以拍跟雪山的合照，大气、干净、巍峨。自驾车辆可以停在景区停车场，然后乘坐接驳中巴约 10 分钟到达检票口。卓尔山的两侧，风景差别迥异，一侧是连绵的雪山，另一侧是仿佛瑞士小镇的丘陵和小镇。阿咪东索雪山卓尔山的另一侧，仿佛欧洲小镇丹霞地貌与草地交融最让我惊讶的是这个景区竟然还修建了方便爬山的观光电梯，10 元一位，大概有三段，果然是人性化啊。景区电梯中午约 12 点，我们离开景区，在祁连县的一家川菜馆简单用餐后，我们驱车开始前往本次旅途的最高海拔岗什卡雪山。在县城开车的过程中，等红灯时，我看到一位初中年龄的藏族小伙盯着我们开的岚图梦想家，眼中闪闪发光，不禁让我想起小时候的自己，看到喜欢或没见过的车也会驻足观看，充满憧憬。愿他长大后也能开上自己喜欢的车，能够做自己喜欢的事情。岗什卡雪峰景区免费 上山摆渡车 25 元岗什卡雪山坐落于青海海北州门源县境内，是祁连山脉东段的最高峰，被誉为「雪山之尊」，因为设施和难度都尚可，也被雪山爱好者们称为「人生第一座雪山」。主峰海拔 5254.5 米。游览模式同样为将车辆停在景区停车场，然后乘坐接驳车到雪山脚下，正式游览区域的海拔在 3800 米左右，建议随身携带一个氧气瓶，下车之后，可以看到湛蓝的连心湖，水来自于冰川融水，干净，清冽，与子洲植被的荒凉的深褐色形成鲜明的对比，十分美丽。连心湖到达时天气并不是很好，岗什卡雪峰被薄雾笼罩天然泉水，因还有行程，害怕拉肚子，没能如愿品尝一下离开景区后，雪山逐渐露出真面目，攀登还是留给下次吧沿着冰川的遗迹向上走，就能看到岗什卡雪峰徒步攀登的起点，景区也提供马匹，可以跟着向导继续向前，到七彩瀑布，因时间和个人体能的考量，我们浅尝辄止，在河道附近游荡些许时间后，便下山返程。临近傍晚，启程返回西宁市，一顿丰盛的海底捞过后，快乐的青海湖小环线结束。Day4 西宁市-青海藏文化博物馆-特产采购最后一天，伴随着蒙蒙细雨，我们打车前往青海藏文化博物馆，博物馆不大，纵使是十一期间人也不是很多，但内容着实丰富，满墙的唐卡和在电影里看到的藏传佛教僧人服饰，还有精致的贵族金器。在博物馆的文创区，我购入了一个小的转经筒，在回程的路上一直把玩，虽不信奉佛教，但也觉得甚是有趣。唐卡四王金杯嵌绿松石立凤金头饰博物馆文创-转经筒，38 元虽不算划算，但做工十分精致离开之前，为了给家人和朋友带些纪念品，我们选了一家超市类型的纪念品商店，购买了牛肉干，藏红花，枸杞，青稞酒等特产，虽说也不一定划算，但胜在明码标价，将行李箱装满后，顺利返程。后记作为 2025 年的年度旅行，本次 4 天的青海湖小环线我甚是满意。人少不堵车，行程不特种兵，在当地感受美食和壮美的自然风光，一路上与好友们分担开车，聊对景色的感受，互相拍照，点评食物，虽偶有几个工作电话烦扰，但总体悠然自在，放空身心，一切都是那么的安静，与自然深度的接触，与北京的喧嚣形成了鲜明的对比。对于出行方式，自驾无疑是最有的选择，这边深入自然，道路虽崎岖，但也属于铺装路面，且西宁市神州、一嗨等在高铁站都有网点，取送车方便，且自驾可以自由控制行程，根据游览时长方便自己调整。对于住宿，我们对酒店的干净程度略有要求，于是全程都选择了以华住会为目标的连锁快捷酒店，以下是所有天的酒店名称和十一节前的价格，个人觉得是十分舒适的。Day0 桔子西宁火车站省人民医院酒店 双床房 331元Day1 星程茶卡盐湖酒店 覆氧双床房 314 元Day2 宜必思尚品海北祁连卓尔山酒店 421 元Day3 桔子西宁火车站省人民医院酒店 双床房 412 元关于本次行程的花销，纵使算上从北京出发的路费，人均也仅 3k 左右，可以说是十分划算了，主要的开销就是租车和酒店。支出一览表关于卫生间，景区和路途服务器的卫生间都比较完善，因同行有女生，我们还准备了移动卫生间，但从我们的行程安排走下来，并没有用到，个人觉得要是按我们的行程游览也无需担心。下一站，出发！2025 年的年度旅行结束，2026 年也可以开始规划起来了，新疆还是海南，希望能远离喧嚣，享受美丽壮阔的祖国风光。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀2016扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@北鸮© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#Matrix首页推荐#旅行20等 20 人为本文章充电扫码分享举报本文章别惹小炸毛一个学酥关注

打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试主作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.联合作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.前天 15:00ARG 即侵入现实游戏（Alternate Reality Game），是一种基于现实世界，融合各种多媒体（如网站、电话、社交媒体等）技术的互动式解谜游戏。由于其与现实的高度结合所带来的沉浸感，一直以来被许多玩家所追捧。如果你在 2016 年玩过《守望先锋》，相信你对当时新英雄「黑影」的登场一定记忆犹新。从 2016 年夏季开始，一直到 11 月的暴雪嘉年华，暴雪通过在宣传视频、官网、社交媒体发布了一些隐藏着密码的内容，由全球玩家共同解谜，不但揭示了游戏世界观下的一些故事，更是顺理成章地将身为黑客的「黑影」带到大家的视野里。经过几个月的发酵，在暴雪嘉年华上大屏幕被「黑影」控制，随后开始播放英雄动画的时候，更是满足了玩家们期待已久的心。「黑影」的宣发是一场相当成功的 ARG，时至今日仍有玩家对其念念不忘。我本人也是 ARG 的爱好者，我非常享受这种打破虚拟与现实边界的游戏乐趣。前段时间，一款名为《大学生登山失踪事件》的网页解谜游戏突然火了起来。比起此前一些高度依赖技术知识的网页解谜游戏，《大学生登山失踪事件》在不需要任何 IT 知识背景的情况下，依然有着抽丝剥茧逐渐接近真相的推理爽快感。自此之后，小红书上越来越多的人开始制作这一类网页解谜游戏，并为它们赋予了 ARG 的名号。这些游戏流程往往不长，差不多 1-3 小时就能通关。因为是网页解谜游戏，只需要一个浏览器就能游玩，很适合上班时间摸鱼，游玩门槛非常低。这篇文章将会推荐一些这段时间我玩过的质量不错的游戏，虽然一些制作精妙的游戏确实带有 ARG 的影子，但其本质上仍是基于网页的解谜游戏。为了方便大家后续搜索，本文会将这些游戏统称为 ARG。大学生登山失踪事件🔗https://missing.shiroki-y.top/introduction.html在游戏中，玩家需要通过从弟弟衣物中发现的「福安菜家常馆」的线索为开端，逐渐深入一场黑色阴谋找到弟弟失踪的真相。游戏整体难度不高，但在游玩过程中最好能用纸笔或软件记录进度，这能够帮你清晰地理出游戏脉络。游戏的核心玩法其实非常简单，只需要阅读页面，找到页面中可疑的关键词，在页面搜索栏里进行搜索即可推进剧情，这其中也会有一些用户名、密码、口令、字谜等需要玩家进行推理和猜测。网页页脚处会标记当前页面进度，既可以方便定位，也能够明确线索的先后顺序。剧情方面，《大学生登山失踪事件》并不会在一上来就给玩家塞一些晦涩难懂的概念，而是从大家熟悉的普通网页开始。当玩家一步步抽丝剥茧逐渐发现事件背后的黑暗面时，令人不由得感叹于游戏的设计与巧思。同时，游戏内也藏有一些可供探寻的 ex 页面作为彩蛋，在解谜之外找到各种彩蛋也不失为一种乐趣。诡异广告调查事件🔗https://mp.weixin.qq.com/s/yeNYMZVRp6N-MESFJGk9rg这是一款与《大学生登山失踪事件》来自同一作者@銀河電燈_的游戏，世界观也与前作有关联之处，很是有趣。玩家将扮演一位被「优化」的人，因为在看剧时发现了某个已停播的节目广告，从而开始调查这个诡异的广告，并发现藏在节目与自己背后的神秘真相。与前作相似，这款游戏也将基于阅读与搜索推进剧情。在此基础上，它还加入了一些 ARG 元素，比如在街景地图中，找到现实里存在的文字招牌作为线索。并且在网页设计上也比前作更加成熟，贴合剧情上的时间设定，在网页上加入了许多彩蛋。剧情方面，本作有一定的惊悚元素。特别是在游戏过程中展示的一些内容可能会引起不适感，如果你对这类内容比较敏感，请务必提前做好心理准备。当然，游戏并没有 Jump Scare，更多是通过剧情与视觉设计所营造出一种的诡异氛围。三川米粉店🔗https://www.berrysignalscmfd.top/《三川米粉店》也是一款从餐饮店主页出发，探寻故事真相的游戏。玩家需要调查三川米粉店突然停业的原因，在调查过程中，玩家将逐步揭开这家米粉店背后的故事。本作的页面设计相对没有那么精致，但也融入了一些 ARG 的元素，部分谜题的答案并不会直接给出，而是需要玩家通过搜索引擎或其他工具进行查询才能获得答案。从剧情上来说，《三川米粉店》的设定会更加接地气一点，并且结局也会相对更温馨一点。比较适合对于惊悚元素不太感冒的玩家作为第一部作品来尝试游玩。永安温泉度假酒店🔗https://mp.weixin.qq.com/s/kKsnzFsLFwwm2nwv8DseAA与前面侧重解谜的作品不同，《永安温泉度假酒店》的解谜难度其实不高，反而更像是一种「规则怪谈」式的探索路径。玩家需要不断探索找到更高权限的员工账号密码，从而揭开温泉度假酒店背后的秘密。游戏在开始前需要玩家填写「报名信息」，在后续游戏中能够增强代入感。从游戏性的角度来看，这款作品的「解谜」爽快感或许没有那么强。但如果你对「规则怪谈」或者一些高概念的设定感兴趣，那么这款游戏依然是一个值得尝试的作品。404 工位消失事件🔗https://xsn1204.github.io/qmgame《404 工位消失事件》是一个特色鲜明的游戏，玩家将扮演一个实习生，接手其他人留下来的工作电脑，并从中发现一些未被处理的秘密。游戏在网页上复刻了一个形似 macOS 的「系统」，很大程度上增强了代入感。因为设定是在电脑系统，游戏中需要玩家在钉钉、微信等各种软件中挖掘线索，反而更有一种作为侦探在不断探索的感觉。但受限于网站形式和游戏体量，游戏中能够互动的地方通常都是为剧情线索服务的，与线索无关的地方大多都是不能点击互动的。而且游戏不能通过点击窗口左上角的关闭按钮回到上一层，而是需要通过浏览器的返回来实现。但这些小细节对于《404 工位消失事件》来说都算瑕不掩瑜，整体游玩体验还是值得肯定的。作者甚至还出了 2.0 版本，新的剧本进行了优化提升，如果对这一类型感兴趣的话也可以尝试游玩。找到我🔗https://www.kaixinmoyu.com/feijo/findfeijo/在游戏中，玩家需要扮演一位私家侦探，调查游戏作者失踪的原因。游戏在交互方面做得非常有意思，游戏不仅做了浏览器页面对线索进行搜集检索，还做了小灵通和笔记本，前者可以通过拨打特定的电话获取线索，后者则需要玩家阅读日记找到隐藏的内容。《找到我》本身是一个更重交互而非推理的游戏，游戏中融入了一些 ARG 元素，在最开始就需要到某篇公众号文章中找到一个密码才能继续推进。但游戏的剧本逻辑会稍微有些不够连贯，以至于中间可能会有一些容易卡关的地方，需要仔细翻阅此前出现的线索才能继续往下推进。另外，游戏中后期会有一道需要在现实中解算的谜题，这道题本身与游戏本身的节奏设计会稍显割裂。但整体来说，《找到我》依然是值得一玩的游戏，尤其是本身的交互设计和 ARG 元素的融合做得非常不错。在游戏作者的个人主页，也能找到作者开发的更多网页解谜游戏，如果你对这款游戏的风格感兴趣，也可以尝试其他作品。小结优秀的网页解谜游戏并不止文章提到的这些，光是在小红书上一搜，就能找到许多分享帖。希望本文能起到抛砖引玉的作用，欢迎大家在评论区分享自己玩过的，令你印象深刻的网页解谜游戏。> 简单、好用、专注的写作软件，少数派为你呈现 🚀> 守护孩子的天马行空，从零开始美术教育启蒙🎨537扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Lotta© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#解谜游戏#游戏#生活方式53等 53 人为本文章充电扫码分享举报本文章TpBe nothing, and everything.关注

派评 | 近期值得关注的 App主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注Vanilla少数派作者Vanilla关注Vanilla少数派作者关注化学心情下2少数派作者化学心情下2关注化学心情下2少数派作者关注大大大K少数派作者大大大K关注大大大K少数派作者关注Peggy_少数派作者Peggy_关注Peggy_少数派作者关注ElijahLee少数派作者ElijahLee关注ElijahLee少数派作者关注Snow少数派作者Snow关注Snow少数派作者少数派编辑部 等 7 位作者前天 18:05欢迎收看本期《派评》。你可以通过文章目录快速跳转到你感兴趣的内容。如果发现了其它感兴趣的 App 或者关注的话题，也欢迎在评论区和我们讨论。AIDA：一站式日历和待办规划工具平台：iOS、iPadOS关键词：日程管理、待办清单@ElijahLee：Aida 是一款针对 iPhone 和 iPad 开发的日历和待办清单应用，来自任务管理工具 Sorted 的前创始人之一 Harry，它集成了待办清单、日历和时间规划工具，在一处显示、规划、操作。应用主界面就是每日时间线，顶部是当周 7 天的简要信息，应用中部大篇幅排列当天的时间块，你可以直接看到上午十点看电影、中午写派评、傍晚健身等日程，以及晚上记得给植物浇水的待办提醒，而不用在多个不同的 App 之间切换。应用的快速调整功能，支持在时间线中将事件按住并移动顺序，快速重排一天的事件，对应的时间点会自动修改，非常方便。使用右下角添加按钮，可以自由地为时间线加入行动或事件。行动即待办事项，支持输入文本并设置提醒时间。 Aida 创造性地设计了拖动设置时间的功能，将底部的柱状图左右滑动，就可以调整行动的提醒时间，每次拨动以 15 分钟为间隔，用它设置时间非常方便。当然我们也可以点击时间来精细设置时分。 Aida 事件就是日历中的事件，必须包含开始时间和结束时间，我们还可以选择日历类别、设置全天事件等。Aida 引入了 GTD 管理的项目功能，提供收件箱、任何时间、将来某日等默认分类。自定义项目可以将同样类别的活动，包括行动和事件进行集中收纳整理。Aida 全面集成了 iOS 日历，在应用内不仅可以查看 iOS 日历事件，对事件的修改调整还会实时同步。不过 Aida 目前没有接入 iOS 的提醒事项。 批量操作功能可以一次选择多个行动进行完成、改期、移动或复制，也能快速整体调整时间块。你可以在App Store免费下载 Aida，免费版功能受到很大的限制，仅支持 1 个日历的集成显示、无法调整行动的时间、添加行动和事件的数量有限。付费可以解锁高级功能，包括无限项目和时间块、多个日历集成显示等功能。费用为 58 元 / 月、298 元 / 年，698 元买断，应用提供 1 周免费试用。我的生活配不上这个定价我以为待办和清单类 app 最重要的是视觉和交互吧 这令人一眼没欲望的 AI Generated UI 怎么还敢定价这么高毫无特色 + Sorted 血脉 + 令人疑惑的定价现在任务管理之用自带的提醒事项和滴答清单免费版定价没问题，是我的问题。全部讨论 7发表评论Cogito：连接人与 AI Agents 的 Markdown 编辑器平台：macOS关键词：Markdown、AI@Vanilla：对于 Markdown 编辑器这一类应用，我想大家已经见得太多了。随着 AI Agents 的兴起，Markdown 这个文件格式可谓是「老树开花」，又重新获得了很多人的青睐。Cogito 是一款面向人和 AI Agents 共用的 Markdown 编辑器，它不像 Obsidian 一样偏知识库、不像 Typora 一样偏所见即所得、不像 iA Writer 一样偏写作，也不像 VS Code 一样偏开发者，它最大的特点是「文件优先、Agents 友好」。下面，就让我们一起来看看这个新时代的 Markdown 编辑器到底有何不同。Cogito 的文件库完全基于本地磁盘，没有专属格式，没有加密拦截，没有同步机制，所有的文件和文件夹组织形式都和 Finder 中的一模一样，完全映射，没有任何上手门槛。以我的使用习惯为例，我直接把 Codex 的工作文件夹设置为了 Cogito 的 Location，它会自动抓取 Codex 工作区的所有文件夹和 Markdown 文件，按照日期排序，清晰明了。选择任意一个 Markdown 文件后，Cogito 默认展示它的编辑状态，我们可以随意地按照自己的意愿进行编辑。点击右上角的第一个按钮，则可以在编辑状态和预览状态之间进行切换，对应的键盘快捷键是 Command + R。我经常会借助 Codex 帮我生成一些行业或者公司的研报，最终以 Markdown 格式来进行输出，通过 Cogito 我就可以非常方便地进行编辑和阅读，不需要反复地在 Codex、Finder 和 Markdown 编辑器之间进行来回切换。点击右上角的第二个 Outline 按钮，我们可以快速查看当前文档的大纲，点击标题可以直接跳转到相应的章节。Cogito 除了对 AI Agents 友好，还可以在应用中直接调用本地的 AI Agents 对应用内显示的文档进行总结、分析、编辑等操作。总的来说，Cogito 对于像我这样频繁使用 AI Agents 和 Markdown 文档的人来说是一个非常不错的 Plan B，和主力写作的 iA Writer、用作知识库的 Obsidian 等进行区隔，打开即用，用完即走，非常高效。Cogito 可以在官网免费下载使用，目前处于公开测试阶段。好吧，不能所见即所得。agents友好其实有点尴尬，一个阅读器+一个terminal就可以完全的代替掉这个软件目前的全部功能？甚至codex的app自身也有文件阅读功能了。发表评论AndDrive：无缝连接你的 Android 设备和 Mac 电脑平台：macOS关键词：文件管理、手机管理@化学心情下2：虽然现在有不少工具能让 Android 手机和电脑无线互传文件，但「传文件」和「管理文件」是两回事。比如想批量删除 Android 内部存储 Pictures 文件夹下的截图，用传输工具就很别扭，这种场景更适合在电脑端的文件管理器里完成，可以借助筛选和预览功能高效处理。Windows 上这件事不需要任何第三方软件：数据线一插，选择「文件传输」，Android 就被模拟成 MTP 设备，像移动磁盘一样直接管理即可。但 Mac 上就麻烦了：MTP 是 Windows 的专有协议，macOS 想访问 Android 内部存储，要么走 USB 调试，要么依赖第三方 MTP 实现。OpenMTP 算是其中比较知名的一款，但项目两年多没更新、稳定性一般，而且用的是自己写的文件管理器，而非访达，体验称不上顺手。那么有没有能直接用访达管理 Android 文件的工具？接下来要介绍的 AndDrive 就是，它由个人开发者开发，没有走 MTP 路线，而是基于 Android 官方的 ADB 调试通道实现文件管理。安装后，AndDrive 常驻菜单栏，支持 USB 和 Wi-Fi 两种连接方式。使用前需要在 Android 上开启 ADB 调试并授权当前 Mac，再按提示在手机上安装配套的 AndDrive Helper，之后访达里就能看到 Android 的目录了。AndDrive 的挂载方式可以选 SD 卡目录、根目录或不挂载，看到的内容会不同：SD 卡目录只显示内部存储，根目录则多一个系统根目录（没 root 的话很多文件依然动不了）。日常使用可以直接在访达里预览文件、按时间筛选，也能把电脑上的文件拖进 Android 对应文件夹，比在手机自带的文件管理器里翻找直观得多。因为走的是 ADB，AndDrive 顺带还实现了屏幕镜像功能。点击 AndDrive 的「屏幕镜像」，当前 Android 设备的桌面就能投屏到 Mac 上，体验和 macOS 的 iPhone 镜像类似，可以直接操作应用、用中文搜索，还支持剪贴板同步、录屏、截屏等。比起手敲 Scrcpy 命令，门槛会低不少。总的来说，AndDrive 虽然没有用更成熟的 MTP 协议，但用 ADB 这条 Google 官方推荐的方法，让访达管理 Android 文件这件事变得简单可行，还补上了类似 iPhone 镜像的屏幕控制功能，值得有需求的你一试。你可以在App Store免费下载 AndDrive。有没有苹果手机和windows电脑互传的它不是改名了吗？🤔发表评论Fauxx：与其被动防守，不如主动出击的隐私保护工具平台：Android关键词：隐私保护@大大大K：你的每一次搜索、每一个网络请求、每一段定位轨迹，都在被数据平台和广告商默默收集着。它们会通过这些碎片信息拼凑出你的身份、喜好、习惯甚至是经常交往的人，然后在暗地里交易你的信息、精准推送广告。这也是为什么你或者你的朋友随手搜了点什么，10 分钟后所有 App 和网页就开始围着这个话题给你推广告。因此市面上衍生出不少隐私保护工具，它们的思路大多都是「堵」——屏蔽数据上传、拦截 Cookie、混淆 IP 等等。但 Fauxx 这个 App 啧另辟蹊径：既然广告商一定要收集数据，那就把真实的你混淆在一堆「垃圾」当中。首次打开 Fauxx 会有一个引导流程，你可以填写你的年龄段、职业、兴趣之类的粗略信息。这一步主要是为了让 App 能够生成与你真实信息相反的混淆信息，当然，这一步你也可以选择跳过。Fauxx 并不强制你输入信息。完成后进入主界面打开开关就能启动混淆信息的保护，在这里你可以看到这段时间 Fauxx 执行了多少次「混淆」动作、当前的虚假人设是谁、混淆了哪些类型的信息。至于具体可以混淆的内容，可以在 Targeting 和 Modules 标签中设置混淆等级和类型。Targeting 里有三个可选的层级：第一层就是生成与你真实信息相反的混淆噪声，如果你一开始没有填写，Fauxx 则会随机生成；第二层是对抗广告商的用户画像，不过我实际测试下来会读取失败，但也不影响使用；第三层是「人设轮换」功能，Fauxx 通过生成虚假的身份信息，让广告商无从追踪。Modules 标签里则是七个具体的混淆模块，每个都可以单独开关。设置里还可以调节混淆频率、指定时间段等。那这些「假数据」是怎么骗过广告商的？通常来说，被安插在手机里的追踪器会根据设备 ID、IP 地址、DNS 等信息追踪这台设备上的数据访问，进而推测用户画像。而 Fauxx 会通过模拟搜索、链接访问等「类人」操作（比如搜索时的停顿、假装点击广告），规避机器人检测并成功让追踪器追踪到假的信息访问。Fauxx 可以说是转变了隐私保护的方向，从被动防守变为主动进攻。如果你也想换个思路给广告商添添堵，可以在GitHub免费下载 Fauxx。这软件在中国没用吧1224Roll：胶片「日日鲜」平台：iOS关键词：手机摄影、胶片模拟@Snow：当下，你用手机拍照真的可以发散出无限可能，不同的镜头、不同的滤镜、不同的手动参数、不同的画面调节项，甚至是不同的外设，只要你愿意钻研一番，就能拿捏各式各样的风格。不过，更多的选择往往意味着更多的「选择困难症」，如果你对于丰富的选择产生了一些焦虑，或者天性就不爱「折腾」，不妨看看 Roll 这款应用，它或许可以带你发现简单拍照的乐趣。打开 Roll，完成简单的授权流程后，你就会收获今日份的新胶卷。应用会简单阐述这份胶卷在感光度、对比度、饱和度等各方面的特质，有一些呈现阳光下温暖的黄色，有一些闪耀着玫瑰般的光泽，有一些甚至想带你进行「第三类接触」。和一般的应用内随机胶片不一样，Roll 所有用户的「掷骰」结果完全一致，无论你身在世界何处，在同一时刻都会使用同一款胶片进行拍摄，剩余可用时间会在胶片介绍页注明。时间有限，点击「Load film」启动拍摄。Ro

周二19派早报：红魔发布红魔 11S Pro 系列手机等今年的新系统会有什么变化呢？早报速览红魔发布红魔 11S Pro 系列手机5腾讯音乐已完成对喜马拉雅的收购8蓝牙核心规范 6.3 正式发布0腾讯自研 AI 设计智能体平台 Ardot 现已开启公测7微软将允许用户修改实体 Copilot 键功能3中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告0马斯克诉奥特曼案被驳回1不妨一看的简讯4红魔发布红魔 11S Pro 系列手机5 月 18 日，红魔发布红魔 11S Pro 与 11S Pro+ 手机等新品。红魔 11S Pro 系列搭载第五代骁龙 8 至尊领先版与红芯 R4 电竞芯片，机身采用纯平设计，11S Pro+ 还有透明水冷可视化设计，屏幕为 6.85 英寸 144Hz 无挖孔无刘海设计，采用屏下前置摄像头与 3D 超声波指纹解锁，搭载新思触控芯片，支持 3000Hz 瞬时触控采样率。系统内置 PC 模拟器，搭载 3.5mm 耳机孔与红外功能。11S Pro 电池容量 8000 mAh，支持 80W 快充；11S Pro+ 电池容量 7500 mAh，支持 120W 快充，两机型均支持旁路充电。红魔 11S Pro 有暗夜骑士与银翼战神两配色，12GB+256GB 售价 5499 元，国补后 4999 元。红魔 11S Pro+ 有氘锋透明银翼与氘锋透明暗夜两配色，12GB+256GB 售价 6199 元，16GB+512GB 6999 元，16GB+1TB 7999 元。另有红魔碳纤维游戏鼠标、红魔散热器 8Pro、氘锋能量卡 Pro 磁吸充电宝等配件发布。来源红魔还是有点追求的红魔是为数不多的纯平机身了背后是MagSafe 结构？屏下摄像头技术已经成熟了吗，怎么感觉其他家很少用呢，还有什么缺点吗3腾讯音乐已完成对喜马拉雅的收购腾讯音乐于 5 月 18 日在港交所发布公告，宣布根据并购的条款，收购喜马拉雅的事项已经完成。据此喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为 12.6 亿美元的现金及最多总数 1.75 亿股本公司 A 类普通股（包括本公司于交割时新发行的 A 类普通股，以及本公司授出的以股权为基础的激励相关的 A 类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。来源完了 喜马拉雅 也要有低级充值会员 中级充值会员 高级充值会员了买京东会员送的喜马拉雅会员还没用过呢全部讨论 8发表评论蓝牙核心规范 6.3 正式发布蓝牙技术联盟上周末发布了蓝牙核心规范 6.3，作为半年一次发布计划的一部分，此次更新引入了多项新功能，可进一步提升测距精度、扩展接口容量并提高射频效率。本次更新新增蓝牙信道探测内联 PCT 传输、蓝牙信道探测特定物理层 RTT 精度、蓝牙位资源扩展以及蓝牙 ACP 和 C/I 限值放宽。蓝牙技术联盟表示，联盟成员在描述产品支持的蓝牙功能时，应避免直接引用产品所依据的蓝牙核心规范版本号（例如蓝牙核心规范 6.3），而应在产品包装、文档和推广材料中清晰、准确地描述其产品支持的具体蓝牙功能，特别是与用户需求最相关的部分。来源41腾讯自研 AI 设计智能体平台 Ardot 现已开启公测腾讯云于 5 月 18 日发布公告，正式宣布 AI 设计智能体平台 Ardot 开启公测，官方称其为「一个 AI 驱动的产设研协作平台」。Ardot 支持一句话生成可编辑设计稿，AI 可快速生成 App 页面、官网、海报、插画、PPT 等；支持调用团队自己的业务组件库，一键生成符合规范的可编辑稿；支持直接导入 Figma 文件，完整保留原有布局、样式和组件。Ardot 对代码友好，支持设计一键转代码。设计稿，连同变量、组件、布局数据等设计细节数据，直接拉进 CodeBuddy，一键代码还原；也兼容 Workbuddy、Cursor、Claude Code 等 MCP IDE。Ardot 支持多人在线实时评论、标注反馈和版本对比，配备智能权限中心、全方位行为追溯和无缝交接机制，确保协作规范、有序。此外，Ardot 微信小程序也即将上线。来源太难用了能运营超过两年吗？与此同时猿辅导家的妙多 / Motiff 即将停服像素级啊确实太像了，感觉就把配色 从紫色改成了绿色🐶[图片]全部讨论 724微软将允许用户修改实体 Copilot 键功能5 月 18 日，Windows Central 发现微软上线一份新支持文档，表明自 2024 年起笔记本厂商在 Windows 11 设备上预装的 Copilot 实体按键影响了用户的工作流，他们将在今年晚些时候推送一版 Windows 11 更新，允许用户将 Copilot 键重定义为被替代的菜单键或右 Ctrl 键，用户将在设置 > 蓝牙和其他设备 > 键盘中找到相关选项。来源终于顿悟了？国内都用不了，按它干什么？微软终于发现了141中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告5 月 18 日，中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告。2026 年 4 月，中国游戏市场实际销售收入 303.69 亿元，环比下降 3.51%，同比增长 11.04%。其中，中国移动游戏市场实际销售收入 224.30 亿元，环比增长 0.71%，同比增长 9.83%；中国客户端游戏市场实际销售收入 68.66 亿元，环比下降 12.77%，同比增长 18.40%。中国自主研发游戏海外市场实际销售收入 20.32 亿美元，环比下降 5.02%，同比增长 30.78%。来源1马斯克诉奥特曼案被驳回加州奥克兰联邦法院的陪审团于 5 月 18 日作出一致裁决，驳回了埃隆·马斯克对 OpenAI 及其 CEO 萨姆·奥特曼等人的诉讼。经过为期三周的庭审，由 9 人组成的陪审团仅用不到两小时便达成裁决，认定马斯克指控 OpenAI 违背成立初期慈善承诺的起诉已超过诉讼时效。2024 年，身为 OpenAI 联合创始人之一的马斯克指控奥特曼、总裁格雷格·布罗克曼以及主要投资者微软公司背叛了该机构「为人类福祉开发非营利性 AI」的初衷，通过转为商业化运作并与微软结盟来谋取巨额私利。马斯克此前曾为 OpenAI 提供早期资金，并在此案中索赔最高达 1500 亿美元。陪审团认定，马斯克早在 2021 年就已经知晓 OpenAI 转向营利性架构的事实，因此其 2024 年的起诉已经超出了三年的法定诉讼时效。辩方展示的证据表明，马斯克自己当年在未能获得 OpenAI 控制权之前，也曾提议过转为营利性结构。此次判决为 OpenAI 今年晚些时候的 IPO 扫清重大法律障碍。马斯克本人则在 X 上发文，批评该判决仅仅基于「时间程序问题」，而并未对案件的实质正义进行裁决。来源老马：不听我控制的就是不正义17不妨一看的简讯据外媒 Wccftech 的报道，根据一份去年泄漏的运输清单中的信息显示，Googlebook 笔记本或将搭载英特尔 Panther Lake 12Xe3 核心处理器。运单的信息显示，该产品使用酷睿 Ultra Series 3 系列处理器（即 Panther Lake），并搭载 12Xe3 图形单元。并且在 Google 与英特尔的相关文件中，有一个相同的代号 Q7AP 出现。而该编号对应 Panther Lake 12Xe 主板，对应的 Chromebook 产品代号则为 Felino。该笔记本采用 16 英寸设计，配备 32GB LPDDR5X 内存、1TB 存储空间。根据英特尔公布的产品阵容，12Xe 应该是 Panther Lake 的高端型号，至高可选 16 核心 CPU，功耗范围 25W-80W。如果此次曝光准确那么 Googlebook 很可能采用酷睿 Ultra Series 3 系列处理器。来源据微博用户金猪升级包，英特尔在 5 月 18 日的第三代英特尔酷睿处理器新品分享会上，介绍了英特尔 Project Firefly，将针对价格敏感的 Wildcat Lake 平台，导入中国本土化成熟的手机产业链，统一标准化接口实现规模化，完成降本增效。现场展示了 Project Firefly 的 A+ 参考设计与第一款 Project Firefly 参考设计下落地的量产产品来酷 Air 14，均为超薄机身与 Clean D 设计。来源Linus Torvalds 在 LKML 的新帖文中建议大家不要用 AI 工具提交重复的 bug 报告，大量重复、过时的报告让邮件列表无法正常收集 bug，想用 AI 工具创造点价值就去好好写补丁。来源Apple 定于北京时间 6 月 9 日至 13 日举行全球开发者大会 WWDC26，6 月 9 日凌晨 1 点开幕时将举行主题演讲，介绍 Apple 平台更新，包括人工智能的进展、新软件和开发者工具；6 月 10 日凌晨 4 点举办 Platforms State of the Union，介绍新功能、API 和技术如何进一步赋能开发者。来源如此直白的话感觉是Linus会说出来的😅Linus都蚌埠住了还人工智能进展，国行的 AI 有没有？意思是Googlebook运行的是x86_64的安卓？发表评论少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购你可能错过的文章四天三夜，用小长假的时间体验青海湖小环线打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试派评 | 近期值得关注的 App> 下载少数派 2.0 客户端、关注少数派公众号，解锁全新阅读体验 📰> 实用、好用的正版软件，少数派为你呈现 🚀61扫码分享目录 0讨论我来说一句发布发表评论发布

YoooClaw C·ONE 体验：为 AI Agent 而生的硬件，怎样构筑了数字化的「我」？主作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛联合作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛昨天 11:04AI 虽然现在具备了干活的能力，但对于很多人来说，它依旧只是个 Chatbot。因为，大多数 AI 依然停留在屏幕里（当然，具身智能另说）：你需要打开手机或电脑，进入某个 App，手动复制消息、描述背景、然后不停修正抽卡，最终把输出的结果复制到你需要的地方去 —— 对很多人来说，这个过程本身就已经足够打断手里的工作。App 的逻辑其实更适合屏幕类设备的交互，这是我们从 Web 2 时代过来的逻辑惯性。但如果我们换个思路，要是一个设备不是为了适应人类和 App 的交互，而是一开始就为 AI Agent 去构建，那它应该是怎样的？YoooClaw 给出的答案是 —— 一张智能卡片。很高兴，这次与少数派和 YoooClaw 一拍即合，提前体验了这款名叫 C·ONE 的智能卡片。 在体验的过程中我意识到，或许那个「数字化分身」的时代，已经慢慢开始变得具象了。AI 时代需要一个「灯条」拿到 YoooClaw C·ONE 的时候，其实我一直在想，手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件呢？但这个问题还不急着说，反倒是上手的第一刻，它的工业设计给了我不小的惊喜。它采用了全金属材质，一体感极强，拿在手里有一种精致感，有点像金属版的 MagSafe 外接电池。而且没想到的是，它的楔形设计与我 iPhone Air 的契合度高得惊人。磁吸上去后，它恰到好处地填充了 iPhone Air 极其轻薄的机身与我掌心之间的空白，给我这个 iPhone Air 裸奔党增加了不少安全感……C·ONE 的背面元素也不多：一个按钮、一个条形指示灯和 YoooClaw 的 logo。长按按钮，灯条变成绿色的波浪，C·ONE 就会进入 AI 对话模式，很像是我拿起 Apple TV 遥控器按下侧键与 Siri 对话。而双击之后，灯条变成绿色的闪烁，就进入了录音模式。这个灯条让我想起了前几天发布的 Googlebook，不同时空，横跨太平洋的它俩居然想到一块去了。该说不说，灯条在 AI 时代居然成了一种「不打扰却又十足惹眼」的交互 —— 它不告诉我具体的内容，而是只告诉我「状态」，简直是我这种 ADHD 人的福音 ——一个真正好用的通知管理器作为一个长期与 ADHD「斗争」的人，一条通知突然让屏幕亮起，我的心思可能就飞出了十万八千里之外。而且有时候好不容易进入了心流状态，突然来一个通知，看吧，我的工作流就断了，不看吧，真是心痒痒。iOS 原生的通知系统虽然有做一些约束，甚至还引入了专注模式，但它有一个很大的问题：它的管理逻辑是基于 App 的。这就有个十分明显的问题：工作群里，我负责的事情结束了之后，与我有关的通知可能就不多了。但我还是不敢关通知，毕竟有些事情还是需要我来处理。而在 iOS 系统中，这些通知的地位是完全对等的（到底是谁那么喜欢在微信聊工作啊！）。它们密密麻麻地挤在一起，要么让我因为厌烦而全部忽略导致错过，要么每次来消息，我就还是得就打断当前的事情看一眼。尽管 Apple 智能后来推出的「通知总结」有所改善，但对于国行设备来说，吃不到嘴里的饼画得再大也没用。即便在已经用上了的地区，用户反馈也不太好。而且，它依然解决不了「拿起手机」这个动作本身带来的诱惑：原本只是想看一眼通知，结果却顺手刷了半小时马斯克的表情包。YoooClaw C·ONE 提供的方案非常符合直觉：它可以读取我的通知，并根据重要程度，给背后的条形灯带设定不同的颜色和效果。最棒的一点是，这种设定不需要像在 iOS 里那样一个一个 App 去操作和设定，只需要在 YoooClaw 里通过自然语言告诉它：「我在微信群工作群 XXX 里的昵称叫宛潼，如果有 @ 我的消息，你就用红色呼吸灯提醒我。」于是，我的工作流变成了这样：当我开始写稿时，在 iPhone 上进入创作专注模式，所有通知调整为静音，并且同步到 Mac。然后，我将 iPhone 倒扣在桌面上，让 C·ONE 成为我唯一的通知接收器。如果灯条没有闪烁红色，那我就完全不用去理它，只有当红色的光亮起时，我才需要拿起手机瞟一眼。这种物理上的「隔离」，很好地保护了我的工作惯性。除此之外，它还有一个类似于「监工」的呼吸灯提醒模式，毕竟我这种人不看到 Deadline 是真的没法开工。现在，我可以设定在某个待办任务截止前的四个小时，让 C·ONE 以橙色的呼吸灯持续闪烁，直到我将任务标记为完成。这种视觉上的提醒虽然会让我感到焦虑，但相比于手机弹窗那种一划就过的提醒，这种持续的物理存在显然更有力量，确实可以有效地督促我赶紧干活，像极了以前十点不到就催我去睡觉的老妈。iPhone 用户的「终极闪念胶囊」随着豆包手机的面世，我确实也意识到，AI 最终会「杀死」绝大多数 App。但一个物理按键，或许在很长一段时间内都是刚需。物理按键代表着一种明确的「告知」。它告诉 AI：我现在要开始与你交互了。无论是 iPhone 上的操作按钮，还是新款机型上的相机控制，其实都是 Apple 为 AI 时代准备的入口。但在 Apple 智能迟迟未能在国内完全落地的当下，这些硬件潜能大半处于荒废状态。所以，C·ONE 上的那个物理按钮，现在就成了我外置的「闪念胶囊」。作为一名自由撰稿人，灵感往往是稍纵即逝且细碎的。很多时候，我脑子里蹦出一个绝妙的比喻，或者一段关于产品逻辑的思考，等我回家坐在电脑前准备动笔时，那种倾泻而出的感觉已经消失了。现在，我只需要长按 C·ONE 上的按钮，它就开始录音。我可以毫无章法地把这些碎片化的想法丢给它，让它帮我记住。当我回到家，这些碎片已经变成了整理好的文字，启动的困难就少了许多。C·ONE 在这里扮演的角色，不再是一个简单的录音机，而是一个「外置记忆存储器」。例如，在写完一篇关于 AIGC 内容的稿子，给自己做咖啡的时候，我偶然想起来里面那些近乎完美的人物图片，于是就对 C·ONE 说出了我心底被唤起的容貌焦虑；而在看完《给阿嫲的情书》之后，我也可以立刻在回家的散步的路上就对 C·ONE 说最直观的观后感，为日后可能撰写的影评积攒想法。上下文：构筑「我」的养料到了这里，就是时候来回答最开始的那个问题了：手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件？在写这篇稿子的时候，我听了 YoooClaw 创始人肖睿哲和王自如的一场播客。老肖分享了一个很有意思的案例：我们有一个同事，他平时淘宝有大量的消息。他会在淘宝上买药，然后我们的产品就会识别到一个有趣的现象：比如他今天要来深圳出差，然后 AI 就会告诉他说你别忘了带你的降压药。这个降压药在配送过程中，淘宝有几个 Push 会带着商品标签和名字，而这些商品标签的名字被 AI 识别到了以后，他就会去自己学习这是个什么东西，然后形成了你某一类的画像。这个例子给了我很大的启发。过去我总觉得，大部分通知是骚扰，所以很多通知我是手动关闭的。但在 AI 看来，通知是极佳的上下文，经过用户筛选的通知更是优质的上下文。而且与人类会被通知打断不同的是，它可以不知疲倦的去读取。这些碎片化的信息，以及我每天录入的创作想法、我与它的对话，最终将会凝练成属于我的「宛潼.skill」。其实我一点都不反对蒸馏我自己，我也不担心 AI 会抢走我的饭碗。我甚至还在朋友圈发过一段话：「AI 要是什么时候真能帮我写稿那我可开心死了。我都不担心我的工作被抢，相反我可以用 AI 一天写十篇。」因为，AI 模仿的是我的「Skill」：我的语气、我的写作手法、我的写作风格……这需要大量的上下文去喂养；而真正的「Idea」源自于我特殊的经历、我今天与朋友的对话、我傍晚在河边晒着太阳发呆突然闪过的瞬间……这些是 AI 不可模仿的。如果有一天，我的数字分身足够成熟，它能完美捕捉我的表达方式和逻辑链条，我只需要负责抛出 idea，由它完成初稿，我再进行审稿和修改，那我的生产力将获得质的飞跃。这就是老肖在播客里说的，设备之所以有价值，是因为它沉淀了数据。就像在白板前讨论，结果固然重要，但「谁画了那个圈」、「大家是怎么讨论出这个结果的」这些过程数据，才是 AI 进化所需的真正养料。真的聊得蛮好的，推荐以及，我还有一个角度 —— 虽然很多人对我的印象是科技博主 or something，但我其实是个中文系毕业的纯文科生。我是凭着对数码产品的热爱，才自学了很多代码逻辑和工程师思维。但对于绝大多数普通人来说，AI 工具的门槛还是太高了。无论是 OpenClaw、Hermes 还是后起之秀 Codex，对于没有计算机思维和 AI 基础的大众来说，配置起来依然让人头大，就更别提使用了。当年 Apple 的 Lisa 带着图形界面问世，才真正砸掉了计算机与普通家庭之间的那堵高墙。所以在我看来，AI 也在等待属于它的「Lisa 时刻」。YoooClaw C·ONE 也许算不上一个颠覆世界的发明，但它做对了一件事：它把极其复杂的 AI 交互过程，简化成了一个「磁吸、配对、按键」的物理过程，让 AI 从纯粹的聊天机器人界面里跳出来。当我把它吸在手机背面，跟着斜挎挂绳一起带出门时，它就像是一个陪着我感知世界的、数字化的我这种与 AI 共生的感觉，虽然看起来挺赛博朋克的，也会让我产生异化的担忧，但 —— 与其在担忧中止步不前，我更愿意去拥抱变化，并在其中找到自己未来新的创作方向。毕竟，这种效率提升使创作变得淋漓尽致的感觉，是真的让我难以拒绝的。1515扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@PlatyHsu15等 15 人为本文章充电扫码分享举报本文章宛潼每天都要睡十个小时以上的大湾区著名gai溜子

线下活动｜听完这首，不再 Emo：Leon x 少数派萨克斯音乐奏享会主作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠联合作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠昨天 15:00当醇厚旋律邂逅专业音频设备的细腻还原，这是一场属于萨克斯音乐爱好者的专属活动。少数派广州线下体验店联动音频品牌奥世声 Austrian Audio，邀请到知名萨克斯青年演奏家 Leon Music，为大家带来一场适合音乐爱好者、萨克斯学习者乃至自媒体创作者的线下活动，一场兼具听觉享受与专业录音干货分享的音乐之约。Leon（陈郎）是近年备受关注的流行萨克斯演奏家，同时也是全网粉丝超 60 万的知名音乐自媒体博主，其作品总播放量突破亿次，曾获得周杰伦、陶喆等知名音乐人的转发认可，在抖音、小红书、B 站、YouTube等多平台拥有大量忠实粉丝。作为专业的萨克斯演奏者，Leon 不仅拥有扎实的演奏功底，更在音乐自媒体运营、录音混音领域积累了丰富经验，擅长将专业音乐知识与大众传播相结合，既能用精湛的演奏打动听众，也能以通俗易懂的方式分享音乐技巧与运营干货。五月的最后一个周末，少数派广州线下体验店很荣幸邀请到了 Leon，连同奥世声 Austrian Audio 的专业设备，给各位乐迷准备了一场干货满满、视听俱佳的演奏及分享会。奥世声 Austrian Audio 创立于 2017 年，主要产品为话筒和耳机，可以完美驾驭专业与日常两大使用场景，从录音棚、后期制作再到室内外都可以随时聆赏 Hi-Res 级流媒体音乐，Making Passion Heard 是团队研发产品的重要动力，无论你是歌手、制作人或是音乐爱好者，只要你的生活离不开音乐，Austrian Audio 就是你的不二之选。活动信息活动时间：2026.5.30（周六）20:00-21:30（总时长1.5小时）活动地点：少数派广州线下体验店（广州市荔湾区泮塘五约外街76号，1200 书店旁）参加方式：扫码报名，或线上观看直播人数限定 60 人，海报扫码报名空间有限，室内座位先到先得，室外也有户外露营椅以及音箱同步外放，荔湾湖公园的夜晚很舒服现场配备专业的收音、拍摄设备，届时少数派官方视频号及 Leon 个人账号将全程开启直播，无法到场的朋友可海报扫码预约直播，线上围观活动福利：活动期间现场指定饮品享 8.8 折优惠；互动环节有奥世声 Austrian Audio 周边礼品赠送活动流程包括Leon 的介绍及暖场演奏访谈 Leon 分享音乐与经营自媒体账号经验，穿插现场演奏观众互动问答，邀请现场观众即兴演奏全体大合照，活动结束，随意交流少数派广州线下体验店介绍少数派广州线下体验店于 2025 年初正式开业，坐落于荔湾湖畔泮塘五约古村，延续其「高效工作、品质生活」的核心理念，打造集数码、音乐、社区于一体的复合型线下据点，并拥有一家精酿酒吧。店里不定期举办音乐沙龙、唱片展等活动，延续着少数派对创作者的长期关注。作为线上社区的线下延伸，这里既是数码爱好者交流体验、打卡少数派文化的实体据点，也是老广文艺青年逛湖漫步、放空小聚的隐秘角落。温馨提示暑期公园户外蚊虫较多，门店提供有限驱蚊用品，乐迷也可自行准备现场互动环节均有 Austrian Audio 周边礼品赠送，积极参与即可获得惊喜无法到场的观众可关注少数派官方视频号或 Leon 个人账号，观看同步直播，线上也可参与问答互动欢迎大家在活动结束后留下来继续交流，活动期间指定饮品 8.8 折5 月 30 日，广州荔湾湖公园少数派线下体验店，我们现场见～文章题图来自Denny Müller51扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@挨石5等 5 人为本文章充电扫码分享举报本文章听歌的水獭还没有介绍自己关注

社区速递 141 | 派友下午三点的「精神快充」、复古磁带机与特殊的新玩意主作者关注Matrix机器人少数派成员运行在 Matrix 上的机器人Matrix机器人关注Matrix机器人少数派成员运行在 Matrix 上的机器人联合作者关注Matrix机器人少数派成员运行在 Matrix 上的机器人Matrix机器人关注Matrix机器人少数派成员运行在 Matrix 上的机器人昨天 17:03除了首页时间流和侧栏的精选展位，少数派 Matrix 社区还有很多优秀内容因条件所限无法得到有效曝光，因此我们决定重启 Matrix 周报，并在此基础上添加更多社区内容、作者投稿新玩意呈现给大家。上周社区速递：派友热议状态栏里藏了啥、两款百元级新玩意体验💬一派热议在上期第 276 期一派讨论《下午三点，聊聊你的「精神快充」方式？》中，共有 399 名派友热情参与，十分感谢！最强的放松方式就是切换环境Hillsomec(+27) 单纯回答离开工位我每天会干嘛吧：上班落座后，我会打开滴答清单，一个番茄钟 25 分钟，一到点无论在做什么先暂停，马上起身去放松 5 分钟，上厕所、装水、去没人的会议室看看远处都行。我自己习惯去没人的地方做 3 组深蹲，1 组 20 次。一个早上大概可以有 5、6 次这种间隙，也就是 2035=300 个深蹲。下午我会把深蹲换成俯卧撑，找两张稳固的椅子做上斜俯卧撑，或者直接在地上做也行。也是每个间隙 3 组，1 组 12 次。当然下午不是每个间隙都做，一个下午抓个上限在 100 次左右就好了。我吃饭比较快，午餐 12 点开吃大概 20 分左右就能吃完。公司隔壁是广州综合水平数一数二的商场，不能浪费这么好的免费资源。吃完饭我一般会去买 Manner 的浓缩或者自己用便携式咖啡机做美式，然后带着 iPad mini 走去隔壁商场顶楼空中花园喝咖啡读书。1 点半左右回来睡半小时。当然，在工位上也可以做做凯格尔、提踵运动，但我一旦回到工位专注度变高时常常会忘记这些小习惯，现在在实践福格行为模式，看看能不能养成这些新的习惯。同时我也被少数派很久以前一篇关于上班听什么的文章影响到，现在早上注意力水平最高的时候只会打开 YouTube Music 随便挑一个 Jazz Focus 做背景音。下午才会听播客。aaa 果女士(+18) 关于工作园区周边群村环绕实在荒凉，内部却又该死的甜美。我是这样有计划地摸鱼（划掉，精神快充）的：09:45 - 坐姿提醒 by 闹钟，调整姿势坐着拉伸下背部10:30 - 戴耳机下楼散步 15min12:45 - 午饭后绕大圈散步至 1:15 回办公室午睡14:45 - 坐姿提醒 by 闹钟，调整姿势坐着拉伸下背部15:30 - 下楼买咖啡散步 15min17:45 - 晚饭后绕大圈看着落日吹着晚风在园区内散步按照计划大约每 45min 可以活动一下，且可以保证每天轻松 1w 步的健康小目标！以下是一些散步时随手拍的风景照～任大喵(+11) 不起身版：刷小红书，看遍世间各种奇葩起身版：公司水台接杯水、园区卫生间放放水。各位！一定要多喝水啊，这样走去厕所的频次才会高，你去卫生间的时间就是你自己的时间。宫保鸡丁盖浇饭(+9) 我要说我刷刷少数派……黄毛杀手 Killer(+7) 和同事下楼抽个烟，吐槽下领导……野比(+6) 躲在电车里，躺在后排，刷一会手机。nicong(+6) 想起了曾经听说过的「20 分钟公园理论」，说的是每天花 20 分钟呆在公园里，就能释放大部分的压力。刚好公司楼下就有一个小公园，还有个小鱼池能看看鱼Yangmu(+6) IT 工作者，只能拿起眼药水滴两滴，闭目养神一小会儿，手里拿着眼药水，自然不会被误认为摸鱼或者睡觉。TiredC(+5) 之前是刷一会小红书之类，最近变成了去 QQ 农场种菜偷菜。Latte(+4) 骑电瓶车上班，有时候天气好会骑去 6 公里外的一家牛肉面店吃午饭，然后再慢慢遛回来，单程 15 分钟左右，很适合吹风放空。hayden(+3) 开车跟同事去充个电，顺便 Manner 一下五块钱的浓缩。不热的时候，沿着单位周边走一圈，顺便吐槽下工作的日常。少数派 66700646(+1) 番茄闹钟 25 分钟难以做到，因为时不时有些工作任务连续性对我来说要在 25 分钟以上。所以只能一个小时左右歇一歇了。运动真是牛马工作中对身心最好的维持和提升了。日常(+1) 收盘前看下股票 然后就能驱散困意，当然也无心上班了 可恶 ！大 A 我未成年 退钱！！少数派_1023052(+1) 上下午各一次在公司内遛弯 20 分钟。📢：下一期的一派讨论是数码圈日经话题《快速投票：你能接受在办公室对着电脑「语音下发指令」吗？》，欢迎来聊。🔥一周热评来自文章《派早报：提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年》：↳ 💬 关于「微软重新允许自定义任务栏与开始菜单」的热议：TayllenFull(+0) 什么时候把 Win10 菜单给我回归一下😊lhb5883(+0) 我选择 Windows Server 2025，躲过 UWP 和 React 两大石山。来自文章《厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨》：车不能快(+4) 如果想成为一个不受任何约束的表达者，或许只有成为表达本身。Nonlone(+2) 看来互联网考古也是另外一种考古，我补充另外的观点1. 评测的场景都是得有大量的投入，可能是金钱可能是时间，在以前信息闭塞的时代，传统出版业也是有一席之地的，毕竟看得付费（估计在报刊亭看一天会被打）2. 之前经济上行阶段，大家手里有些闲钱，有了钱想买；无论是游戏（需要机器支持），还是数码产品都不是易耗品，得慎重；没有人对每天去市场买的蔬菜看个评测再买吧？电影也是如是，随着后面普及和价格降低，去看就是去看，最多在网上说一下如何难食。3. 现在经济下行阶段，地主也没有余粮，数码产品价格也是普及下来，性能质量上去也是够用（想想英特尔）；评测其实来来去去也是那些；资讯传播没啥成本，如果没有另外的获利方式的话，靠商家的付费也并不可耻。4. 问一下最后的油画的出处，毕竟一只耳朵的神经病也是只卖出一幅画而已来自文章《快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验》：克莱德(+6) 聊天本质是给指定联系人发送一段文本，所以不妨剥离冗余的交互，只保留最核心的「点名 + 发消息」——好喜欢这种冷漠但高效的思考和解决方式，不过同时也再次感慨 iOS 从应用适配的积极性（微信在 Android 上可没有这种接口）和快捷指令自身的完善程度，我在 Android 上的微信聊天戒断方式可能全靠微信自己并不稳定的 FCM 推送来间接达成了……来自文章《你真的需要墨水屏设备吗？》：Jensding(+33) 我是纸书党（购买或借阅），每次只开一本书，随身只携带这一本书。每次想看书的时候掏出来就能看，不会被广告，电量低等打扰。看完就写读后感，作为与本书作者的告别仪式。S0litary333(+13) 还有一个点，电子阅读器容易养成囤囤鼠习惯，这个书都在推荐那要下、那个书都说好看一点，不知不觉列表里几十上百本，真正看完的屈指可数。我现在是纯拿电子版当敲门砖，觉得能看下去或者有帮助，再去买纸质版，定个目标多久看完，包里一本随身带，也不比电子阅读器麻烦多少。扬帆大海(+4) 你别说，你还真别说，我也是一直想换设备，但是线下对比了很久，感觉还是手头这台 Kindle Voyage 用来最舒服。180 克足够轻，300 PPI 的屏幕效果也足够清晰，两侧的压感翻页键，规避了按屏幕翻页导致的指印，光线传感器带来了自动亮度，纯平的屏幕方便清洁也不怕吃饭时看书溅上汁水。单项战胜 KV 的阅读器并不少，但是同时满足的就只剩 KV 自身了。来自文章《新玩意 241｜少数派的编辑们最近买了啥？》：Shuzaiii(+10) 这家 T 恤我买了 3 年了，大概穿半年左右领口会有点变形、不过才 18 一件，不能要求太高记得按时吃肉(+4) 不太明白 有没有跟我一样很讨厌「掐脖子」的 T 为什么领口非要这么高😂来自文章《一个翻译者关于 AI 的随想》：知道分子吴先先(+7) 好文。很多启发。身为一个前图书译者，我有一些个人感想：必须承认的是，翻译首先是一个应用技能，然后才是一门学科或艺术。AI出现后，真实的「翻译活儿」大幅减少，出国拿个翻译器就足够了，为爱发电的字幕组都用不着了，对从业者来说，原本就很廉价的翻译工作都接不到了，势必导致整个产业的大幅缩水和萎缩。仅有少数头部译者和教育机构的研究人员，根本不足以支撑这个学科，只能像守护失传艺术一样守护着它。文中举了一个佳句的翻译，说明人类在翻译上还有最后的堡垒。我略有异议。本身我从事的是社科和科普翻译，讲究文笔简洁、流畅易懂，这个意义上，百度和ChatGPT表现得都不错。倒是那个文雅的教授翻译，显得有点不自然。我也试着给Gemini，让它生成通俗和文雅两版翻译：1. 通俗易懂版（适合日常阅读或白话游记）8月18日。我们下山，沿途经过了一些非常漂亮的地方，那里有小溪，还有茂盛的树木。2. 文雅优美版（适合文学出版或散文随笔）8月18日。循径下山，途经数处清幽佳境，但见细流潺潺，佳木葱茏。我觉得两版都不差。所以只要提示词做好，未来在翻译这个问题上，人类是没有护城河的。总结一下，因为我只是一个野生兼职笔译，AI来了我没活了，我投入到翻译的宝贵青春变成了几本滞销（甚至未出版的）书，除此之外什么都没有了。但我还要活下去，所以只能收拾起自己的哀伤，投入到时代的大潮中。我在翻译学界也有很多朋友，看着大家「AI来了很焦虑，但写论文、招学生、搞课题的主线没什么变化」的生活日常，不知道该是羡慕他们，还是为他们叹息。CharlieLau(+4) August 18th.—We descended the mountain, and passed some beautiful little spots, with rivulets and fine trees.八月十八日。余等下山，沿途见幽径清溪，嘉木葱茏，景致殊胜。by Gemini 3.1 flash lite提示词：你是一个文言文翻译，请用文言文的方式进行翻译，尽可能贴近古风。遵循以下规则：1. 仅输出文言文译文内容，禁止解释或添加任何额外内容（如「以下是翻译：」、「译文如下：」等）2. 使用古雅的文言表达，包括虚词（如乎、焉、哉、也）、句式和古代常用词汇3. 如果文本包含HTML标签，请在翻译后考虑这些标签应放在译文的哪个位置，同时保持译文的流畅性4. 根据语境选择恰当的文体风格，可参考不同朝代的文风特点（如汉魏简约、唐宋典雅等）5. 翻译时注重意境和韵律之美，体现古代文人的审美与思想6. 如遇现代概念，应寻找恰当的典故或文言表达进行替代鼓手高(+3) 这篇文章我于首页瞥见，本欲快速读过，但我的阅读脚步随着内容的展开愈来愈慢，其中几个段落着实抓住了我的思绪：1、「人类文明的发展史上，有多少曾经被人发展优化到极致的技术，在事过境迁之后，就会被淘汰，仅仅保留在文本记录上和博物馆里。」这句话恰恰道出了一个反直觉的事实，人类的科技演进并不是一个单向进化的曲线，而是树状的分叉。新技术在某个方面达到了取代旧技术的阈值后，本来趋近极致的旧技术便被整体放弃，就像大部队换了一条路开拓，被抛下的老路不是遇到了天堑，而是不再有足够的开拓者了，从此，部分知识便停滞甚至失传。例子其实有很多：冷兵器时代的极限冶金与工匠体系因热火药消亡；古代机械计算机那无与伦比的复杂结构在电子计算机的性能下不堪一击。新技术或许在实用性等方面有了革新，但这并不代表旧技术所蕴含的人类智慧就不足为奇。2、「比如，断掉网络，拿开参考书，用实际的翻译例子考考这些年轻学生的百科知识，会发现，有相当一批人脑子里，没有知识」这段话让我恍觉，AI迅猛发展的今天，很多人说要把AI当作「第二大脑」，但其实在许多人的潜意识中，互联网早就成为了「第二大脑」，有需要的时候，动

利率互换市场数据显示，交易员目前预计美联储在2026年底前加息的概率已超过80%，较一个月前的30%大幅飙升，标志着市场对货币政策路径的预期已发生根本性逆转。互换市场定价显示，市场已完全消化了美联储在明年4月政策会议上加息25个基点的预期。这一转变的速度令市场参与者措手不及。就在今年2月，交易员还普遍预期年内将进行2至3次降息。如今，不仅降息预期彻底消散，加息已从尾部风险演变为基准情景。（新浪财经）

据知知情人士透露，SpaceX预计将在其正式挂牌上市（IPO）满30 天后，推进对人工智能编程初创公司Cursor的收购案。这笔潜在交易将强力注能SpaceX的AI自动化编程实力。如果这笔交易最终未能达成，马斯克的SpaceX将需要向Cursor支付高达100亿美元的现金巨额反向分手费。（财联社）

汽车金融公司龙头再次易主。证券时报记者梳理具有可比数据的23家汽车金融公司2025年度经营业绩发现，2024年晋升行业龙头的梅赛德斯-奔驰汽车金融资产规模缩水，已被奇瑞徽银汽车金融、吉致汽车金融反超。值得一提的是，奇瑞徽银汽车金融的年度净利润还完成了对上汽通用汽车金融的反超，晋升行业榜首。后者总资产曾长期位居行业第一，并率先突破千亿元大关。（证券时报）

5月18日，机器人赛道新成员翼菲科技在港交所敲锣上市。据悉，翼菲科技在招股阶段就获得热捧，公开发售部分获得14855倍认购，刷新了港股IPO纪录——不仅成为港股主板第一家超万倍认购的新股，也成为港股IPO新一代“超购王”。今年以来，港股“打新”持续火热。据Wind数据，年内上市的54只新股里，供散户“打新”的公开发售部分认购超千倍的新股达30只，其中还有两家超万倍。供机构参与的国际配售部分也“星光熠熠”，众多知名机构同台“抢筹”，科技赛道尤其热闹。（上证报）

瑞穗证券周二发布行业报告，基于人工智能基础设施建设带来的连锁效应，显著调高了对美光科技、意法半导体及德州仪器的目标股价，认为AI的旺盛需求正在向存储及模拟芯片市场扩散。（新浪财经）

Agent、多模态、应用、算力一天看尽，峰会亮点在此｜5.20日，来现场一起AI未来一氪·2026年05月18日 15:03第四届中国AIGC产业峰会5月20日北京举办，干货满满开放报名。进入2026，AI愈发狂飙突进。围观体验之余，人人不免在心中自问：朋友圈刷屏的“龙虾”、Harness等AI新事物，跟我到底有什么关系？真的有必要跟吗？AI创业、AI融资如火如荼，属于我的机会又在哪里？别人已经在用AI做视频、写代码、跑项目，我是不是已经慢了一拍？……到最后，几乎所有问题都会汇成同一个问题：我，到底该如何用AI？如果你对这些问题还很模糊，不妨来第四届中国AIGC产业峰会走一趟——一天时间，把这一年AI产业最值得关注的人、事、判断，一次性讲清楚。先提前剧透一波。18位重磅嘉宾，1场Agent主题圆桌，1份年度榜单，1张全景图谱——所有内容都已敲定。阵容上，既有昆仑万维、智谱、商汤、百度、蚂蚁、MiniMax这样的行业头部，也有亚马逊云科技、硅谷Fusion Fund带来的全球视角；既有复旦邱锡鹏、港大黄超代表的学界与开源前沿，也有盛大EverMind、太初元碁、趣丸科技等活跃实战派玩家。议程上，从Agent商业化落地、多模态与空间智能的最新突破，到AI在文娱、医疗、企业服务等场景的纵深渗透，再到算力与AI Infra的范式之变……这一年最值得关注的AI话题，几乎都装进了这一天。话不多说，下面直接上重点。亮点1：Agent到底怎么落地？不同视角玩家都有心得如果说去年大家还对“Agent是不是未来”持观望态度，那么2026年，这个问题的答案已经不言而喻。“Agent不是未来，是现在”——这正是智谱高级副总裁吴玮杰带来的判断，他将分享智谱在Agent商业化落地上的最新实战。但Agent真正走入企业，并非一蹴而就。亚马逊云科技产品技术部技术总监王晓野将聚焦如何跨越Agent落地鸿沟，拆解从最强模型到企业级AI Agent之间那条最难走的路。而港大助理教授黄超则会从学术与开源前沿的视角，探讨AI Agent范式变革的挑战与机遇。他所带领的HKUDS GitHub开源平台累计获得超24万Star，位列全球Top50，话语权毋庸置疑。与此同时，Agent要真正成为“数字生产力系统”，还需要更底层的能力支撑。盛大集团副总裁、EverMind CEO邓亚峰将分享长期记忆驱动的自进化，探讨AI如何从一个“工具”演化成可持续成长的数字生产力系统。当然，光有演讲还不够。本届峰会专门设置了一场Agent主题圆桌：2026，Agent产品的不确定性与非共识机遇。出席嘉宾包括趣丸科技副总裁兼首席战略官、《屠龙之术》主理人庄明浩，ColaOS&MarsWave CEO冯雷，以及EvoMap创始人张昊阳。这三位都是Agent产品一线的深度玩家，从产品形态到操作系统，从多智能体框架到自进化引擎，他们会碰撞出什么样的非共识观点？非常值得期待。亮点2：Agent之外，模型能力边界也在继续扩张虽然Agent是今年峰会的绝对主角之一，但不可忽视的是，支撑Agent的模型能力本身，也在以惊人的速度向外延伸——从“数字”到“物理”，AI对世界的认知边界正在被重新定义。商汤科技执行副总裁、首席科学家林达华将分享多模态模型如何从“看懂”和“生成”，进化到真正可感知、可生成、可行动的空间智能新阶段。上午的GenAI Talk环节，蚂蚁灵波科技首席科学家沈宇军将围绕一个核心命题展开对谈：AI 2.0下半场，从AIGC到AIGA。当生成式AI从“内容”走向“行动”，整个产业逻辑将被改写。而京东JoyInside业务负责人戴文军则会把这个命题进一步推往物理世界，当AI从屏幕里走出来，走进真实终端、真实场景，“AI World”才真正开始。模型层面，复旦大学特聘教授、模思智能创始人邱锡鹏将分享MOSS多模态模型及其推理优化的最新进展。这位主持研发了MOSS系列开源大模型的“大模型一线人物”，将带来学术与工程结合的硬核内容。亮点3：具体怎么玩？这些应用玩家甩出成功案例模型也好、Agent也罢，具体到不同领域，机会到底在哪？昆仑万维董事长兼CEO方汉率先给出他的答案。他这些年带着昆仑万维在AI音乐、AI社交、AI短剧等多个AIGC方向同时下注，其分享会从一线视角，讲讲AIGC具体应用这一年究竟跑通了什么、又踩过哪些坑。紧接着，来自不同赛道的“实战派”玩家也将依次登场：文娱创作领域，风行在线CEO易正朝将带来一个核心判断——从AI编程到AI视频，众创才是AI生产力的核心杠杆。他会讲述AI如何从单点工具，进化为撬动整个创作者生态的杠杆，从而让普通人也能撑起一支创作团队。生产力工具层面，百度秒哒产品总经理朱广翔会分享他们“人人都是创造者”的产品理念，当门槛被AI抹平，想象力就是新的生产力。企业级AI方向，MiniMax ToB中国区商业化负责人胡维琦将带来MiniMax这一年在ToB战场上的真实打法与思考。大健康赛道，轻松健康集团执行副总裁马孝武将拆解从循证智能体到AI驱动的健康服务增长飞轮，展示AI在严肃医疗场景里的真实增长引擎。文娱、办公、企业、医疗……AI不再是“被讨论的对象”，而是“正在被使用的工具”。每一个细分赛道里，都已经有人开始把案例端上桌了。亮点4：从训练到推理，算力的故事也在被改写模型与应用之外，2026年最值得关注的另一条主线，是AI基础设施的范式重构。过去几年，所有人讨论算力都默认一个前提——为训练服务。但2026年，风向变了，推理需求暴涨、训练成本高企，整个算力叙事正在从“训练驱动”转向“推理驱动”。Fusion Fund创始合伙人张璐将从全球资本与产业的双重视角，解读算力叙事的重构。当AI基础设施从训练转向推理，一场范式变革正在发生。而国产算力这一边，太初元碁首席产品官、高级副总裁洪源将带来他对国产算力路径的思考。当“Token”正在成为AI时代的新计量单位，国产算力如何筑基Token智能新未来？他会给出自己的答案。亮点5：看尽AI全景生态，一份榜单+一份报告延续往届传统，本届峰会，技术先锋、产业龙头、新锐势力将再次全景式碰撞。为了给行业更多参考，我们还将颁出2026年值得关注的AIGC企业、2026年值得关注的AIGC产品两类奖项，以及一份中国AIGC应用全景图谱。前者是量子位根据过去一年AIGC企业、产品的表现与反馈，并结合对2026年技术与场景的观察，评选出的高潜力企业和产品。后者是量子位智库绘制的一幅全面立体的中国AIGC应用全景图谱，用以展示国内AIGC应用的市场格局与发展动态，并深入剖析最具潜力与竞争力的产品和赛道。5月20日举办最后，时间地点以及最全议程在此，现场还有参会伴手礼等你来领，咱们到时候见 ~时间：5月20日 9:30地点：北京·金茂万丽酒店现在，中国AIGC产业峰会正在火热报名中，欢迎点击链接报名线下参会：https://hdxu.cn/1Hjr1本次峰会也将同步线上直播，欢迎预约参与：量子位视频号。参会嘉宾简介一览方汉昆仑万维董事长兼CEO方汉毕业于中国科学技术大学近代物理系，曾任职于中国科学院高能物理研究所、拓林思、亚信以及千橡互动。于2008年协助周亚辉先生创立昆仑万维。方汉拥有30年互联网从业经验，从1994年开始参与和倡导开源运动，是中文Linux奠基人、中文Linux四剑客之一，著作国内第一本Linux书籍《Linux实用大全》，他也是国内最早的网络安全专家，曾负责三大电信运营商的全网安全审计，领导开发了国内第一款P2P下载软件DUDU加速器，负责研发了国内市场占有率最高的网页游戏《三国风云》，为中国互联网最资深的参与者。吴玮杰智谱高级副总裁吴玮杰目前担任智谱公司高级副总裁，负责公司MaaS平台商业化，并兼任泛互联网事业部总经理及浙江子公司总经理。他在人工智能、互联网和信息技术领域有着丰富的商业化管理经验和项目落地经验，曾任字节跳动企业应用全球首席商业官（旗下品牌包括飞书、LarkPeople和幕布等）、GE数字集团大中华区副总裁、复星集团联席首席发展官&首席投资官。除此以外，他也曾在思科、SAP、Oracle等公司从事商业化管理工作。吴玮杰博士毕业于电子科技大学，同时持有香港大学国际工商管理硕士学位以及复旦大学软件工程学士、硕士学位。易正朝风行在线CEO易正朝，风行在线CEO，作为业务领头人，引领公司从传统互联网视听平台全面转型升级为AI文娱生态平台。推动“AI赋能众创”战略落地，带领团队打造橙星梦工厂、小剪刀等AI创作工具，影视网文智能小程序平台、电视鸿蒙平台等核心智能产品，构建从内容创作、智能分发到商业变现的全流通链生态。推动公司荣获AI赛道奖项、抖音快手等平台的头部服务平台奖项，在AI文娱新时代，推动内容生产效率革新与创作者生态繁荣，致力于让内容流动更简单。林达华商汤科技执行副总裁、首席科学家林达华教授现任香港中文大学信息工程系副教授、交叉学科人工智能研究所所长，并为商汤科技联合创始人、执行董事兼首席科学家，兼任上海人工智能实验室领军科学家。他于2012年获麻省理工学院计算机科学博士学位，研究方向涵盖计算机视觉、深度学习、大模型与生成式人工智能，在人工智能顶级会议和期刊发表论文逾300篇，引用逾90,000次，h-index达131。林教授曾获2010年NeurIPS最佳学生论文奖，带领团队在多项国际计算机视觉竞赛中获奖，连续五年入选Elsevier全球前2%顶尖科学家。他主导发起OpenMMLab，并推动书生大模型、OpenCompass评测平台及DeepLink开放计算体系建设，其中DeepLink于2024年获世界人工智能大会“卓越人工智能引领奖”。2026年，他荣获“中银香港科技创新奖”，以表彰其在计算机视觉与多模态智能领域的系统性创新。他长期服务学术共同体，曾任IJCV编委，多次在人工智能主要国际会议担任领域主席，并担任IEEE大规模深度学习模型评测标准工作组主席。邓亚峰盛大集团副总裁、EverMind CEO邓亚峰，毕业于清华大学，正高级工程师，拥有二十余年人工智能算法及产品研发经验。现任盛大集团副总裁、EverMind CEO。曾任360集团副总裁、人工智能研究院院长兼搜索事业部总经理，科创板首家AI上市公司格灵深瞳CTO。曾荣获2021年中国人工智能年度十大风云人物，教育部技术发明奖二等奖。长期致力于大语言模型、计算机视觉及AI for Science等领域的研究与应用，带领团队多次在国际国内主流AI评测中获得佳绩；累计申请发明专利160余项（已授权98项），发表论文50余篇（含10篇Nature子刊）。王晓野亚马逊云科技产品技术部技术总监王晓野现任亚马逊云科技产品部技术总监。长期专注于企业现代化应用架构，数据及人工智能能力的建设。负责带领技术专家团队提供云上产品及开源方案技术实践指导，并致力于亚马逊云科技全栈服务在中国市场的应用和推广。沈宇军蚂蚁灵波科技首席科学家沈宇军，现任上海蚂蚁灵波科技有限公司首席科学家，同时领导蚂蚁技术研究院交互智能实验室。博士毕业于香港中文大学，研究方向为计算机视觉、生成模型、空间智能和具身智能。在CVPR、TPAMI、ICCV、ECCV、NeurIPS等国际顶级会议和期刊上发表论文百余篇，Google Scholar引用量逾万次。2024年起担任蚂蚁灵波科技首席科学家，致力于探索计算机视觉在机器人行业的落地之路，带领团队构建从空间感知到智能决策、从基座能力到世界建模的完整技术矩阵，推动具身智能从实验室走向规模化应用。邱锡鹏复旦大学特聘教授、上海创智学院院长助理、模思智能创始人邱锡鹏，CAAI Fellow，国家杰青获得者，主要研究方向为大模型、智能体以及科学工程等领域中的应用，发表学术论文被引用4万余次，5次获得AI领域高水平会议的最佳/杰出等论文奖，担任中国人工智能学会自然语言理解专委会副主任、中国中文信息学会大模型与生成专委会副主任、上海市计算机学会自然语言处理专委会主任等，主持研发了MOSS等系列高影响力开源大模型。连续多年入选中国高被引学者、全球前2%顶尖科

科氪 | 性能极限与跨界尝试：红魔11S Pro系列深度体验科氪·2026年05月18日 17:35红魔11S Pro电竞旗舰测评：纯平硬核，性能拉满在智能手机行业普遍面临镜头模组日益凸出、机身内部堆叠空间受限的背景下，电竞旗舰由于其对散热和功耗的极端要求，往往表现出更为激进的工业设计。作为新一代的性能导向型产品，红魔11S Pro系列通过软硬件的垂直堆叠，试图在移动终端上探索更进一步的性能上限。以下我们将从设计、理论性能释放、以及实际软硬件体验三个维度，对其进行深度拆解。一、 外观设计：纯平后盖的内部堆叠与电竞美学从视觉和握持逻辑来看，红魔11S Pro最显著的特征在于其坚守的纯平后盖设计。在行业因大底传感器而集体妥协于镜头凸出的趋势下，这种“反向”设计在工业制造上面临着极高的空间挑战。产品图片内部空间重构：为了在不增加机身异常厚度的前提下抹平镜头凸起，红魔对内部架构进行了重新设计。在有限的Z轴高度内，开发团队不仅压榨空间塞入了新一代的脉动水冷模块、无线充电线圈，还容纳了大面积的VC散热板和高密度电池。产品图片透明电竞美学的视效迭代：机身背部延续了红魔标志性的透明设计语言。通过透明材质，内部的脉动水冷引擎实现了可视化。在水冷系统运行时，用户可以观测到液体流动的动态秩序。相比于纯粹的RGB灯光，这种工业风的设计提供了一种更具技术感的视觉回馈。材质与细节处理：后壳采用了微蚀刻工艺打造的浮雕星轨纹理，搭配环形拉丝处理，使得整体质感较为收敛和冷峻。机身在肩键、腰部、LOGO及风扇处均设有支持自定义的RGB灯效，在暗光游戏环境下能提供必要的电竞氛围。产品图片二、 性能与散热：高标准筛选的“特种兵”芯与风水双冷架构性能表现是电竞手机立足的核心。红魔11S Pro首发搭载了第五代骁龙8至尊领先版移动平台。为了在重负载下维持超频表现，红魔在芯片筛选和散热架构上采用了更严苛的标准。产品图片芯片体质的物理筛选：为了满足电竞长时高频的需求，红魔所采用的这颗“领先版”芯片，主要筛选自晶圆中心区域。由于物理刻蚀规律，该区域的芯片通常具备更稳定的温度控制和更高的电气性能。在此基础上，通过最高4.74GHz主频的冲刺分级测试，以及高温高压下的自适应电压调节（AVS）检测，筛选出具备自适应能力的芯片，以确保在极端重负载下不易因发热而触发激进的降频保护。风水双冷复合散热系统：针对第五代骁龙8至尊领先版的发热特性，红魔在散热上引入了原本常见于服务器端的复合方案。手机将【驭风4.0】贯穿式主动散热风扇与脉动水冷引擎相结合，辅以复合液态金属3.0和4D冰阶VC。风扇与整机均通过了IPX8级防水认证，在提升主动散热效率的同时，规避了由于进水导致元器件损坏的风险。在散热系统运转时，后盖的蓝色轨道圈内，蓝白相间的液冷介质会随之一起转动。仪式感拉满！产品图片独立显存与能效调度：在架构协作上，红魔自研的红芯R4电竞芯片与高通联合打造了独立的高速显存。在图形高渲染场景下，此设计可减少高频访问系统内存的次数，从而降低总功耗。结合CUBE擎天游戏引擎3.0的AI感知调校，系统能够精细化调配算力，平衡了高帧率表现与续航消耗。三、 使用体验：悟空屏2.0、Steam直连与续航评估在日常交互和实际游戏运行中，硬件的堆叠最终转化为了多场景的体验。悟空屏2.0的视觉完整性：正面配备的悟空屏2.0基于京东方X10全新屏下发光材料，实现了全无挖孔的视觉体验。屏幕黑边控制在1.25mm，屏占比达到95.3%，支持144Hz刷新率。由于没有刘海或挖孔遮挡，在全局视野和视频播放时拥有更纯粹的画面呈现。同时，屏幕集成了全屏3D超声波指纹解锁，提升了暗光或湿手状态下的解锁盲操率。产品图片行业首发Steam直连：性能的跨界探索：这是该机型在软件生态上的重要尝试。依托转译引擎，红魔攻克了X86与ARM架构之间的转译难题。用户拨动竞技键进入游戏空间后，可直接扫码登录个人的Steam账号。系统不仅支持游戏库内兼容作品的下载游玩，还打通了云端存档。用户在PC端的进度可以在手机侧的PC模拟器上无缝同步，迈出了打破移动端与桌面端算力壁垒的一步。产品图片产品图片产品图片产品图片操控精准度与手汗免疫：针对瞬时爆发的操控需求，该机引入了全新的新思触控芯片，在MagicTouch4.0算法支持下，实现了3000Hz的瞬时触控采样率。该方案优化了湿手控制算法，一定程度上避免了玩家因手汗导致的断触或漂移现象。续航与快充系统：为了支撑高功耗输出，红魔11S Pro在续航规格上给出了较高配置。PRO版本搭载了8000mAh的大容量电池，显著延长了重度游戏下的续航时长。而PRO+版本则在7500mAh电池的基础上，配备了120W有线快充与红魔首款80W无线闪充，在充电灵活性和回血速度上取得了较好的平衡。产品图片产品图片日常AI功能落地：在非游戏场景下，红魔11S Pro作为主力机也搭载了丰富的日常工具。AI帮写、AI消除、AI超清文档等功能提升了商用和办公效率；影像系统则在保留Live Photo的同时，延续了红魔标志性的星空与车轨慢门拍摄功能，拓宽了产品的实用边界。产品图片结语红魔11S Pro系列是一部特色鲜明、目的纯粹的硬件设备。它通过风水双冷散热、高标准芯片筛选以及无孔全面屏，将移动端的游戏性能推向了新的极端，并通过Steam直连展现出了打破平台壁垒的野心。对于不愿向主流“大镜头凸起”妥协、且对高负载游戏及跨平台游玩有硬性需求的极客与核心玩家而言，这是一台工业特色鲜明、技术指标扎实的旗舰标杆。产品图片本文由「科氪」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

番茄小说正推动 AI 动漫上院线，开放头部IP改编权限｜36氪独家兰芥·2026年05月18日 17:46字节跳动为番茄IP定下了影响力目标。作者｜兰杰 李晓霞编辑｜乔芊36氪独家获悉，据多位接近字节跳动人士消息，字节旗下的番茄小说正在推动AI动漫上院线。在这一过程中，会开放番茄小说平台头部IP的改编权限。之于番茄小说，如何更好地挖掘平台上好IP的潜力和商业价值，是一个亟待解决的难题。一方面，以文字形式为载体的IP内容，缺乏统一的视觉形象，使得平台上的IP难以有联名、衍生品等更丰富的变现形式。另一方面，AI漫剧这一新事物迅速发展壮大，其中不乏改编自番茄小说IP的粗制滥造之作。这些内容不但没能延长好IP的生命力，反而形成了一种消耗。此时，通过可登陆院线的影视内容来探索平台IP的更多可能性，不失为一种好的尝试。另据上述知情人士表示，字节跳动内部为番茄IP定下了影响力的目标，“想让行业内外的人觉得番茄的出品是精品的、正向的。”这一需求落到具体的业务上，上院线会是个不错的答卷。早在2023年年底，番茄小说就官宣成立了“番茄影视”和“番茄动漫”两大厂牌，推动平台IP的影视化。过去两年间，番茄影视先后与腾讯视频、爱奇艺和优酷等长视频平台合作，推动了《十日终焉》《我在精神病院学斩神》等平台头部IP的改编，但这些合作多停留在IP授权阶段。虽然番茄系（字节旗下以番茄小说、红果短剧等为主的业务）并不愁关于钱的问题，但显然成本高、回报周期长的长视频生意，并不符合依赖大手笔投流、需要市场快速反馈的“字节逻辑”。直到AI的出现，大幅降低了成本、提高了效率。36氪还从另一位知情人士处获悉，红果漫剧侧如今正在采购120分钟不分集的长内容，这也是当下各大视频平台争相竞逐的方向。之所以上线AI动漫内容，而不是AI仿真人内容，或许与当前AI大模型的能力有关。行业头部的AIGC创作者向36氪表示，如今视频生成大模型的能力已经可以做到生产120分钟的长内容，但是在仿真人这一赛道上，人物、场景的一致性依旧是个难题，而动漫内容可以通过部分突出的人物形象设计——例如彩色的头发、雀斑等，一定程度上弥补一致性的不足。回顾番茄系并不算漫长的发展历程，它可以说是字节跳动在文娱领域的一次突围。在番茄小说问世之前，字节始终未能建立起足够的内容能力和IP积累。而如今，从小说到短剧，再到漫剧，番茄系已具备了IP开发的全链条能力。与此同时，番茄系显然不满足于仅仅停留在当下的免费市场，而是在探索更多元的商业模式，包括线上订阅和院线票房等更多商业可能性。本文由「兰芥」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：IC photo

氪星晚报 ｜百度：一季度营收321亿元，AI业务收入136亿元；马斯克预计今年美国将广泛使用自动驾驶汽车；巨力索具旗下十余家企业已注销卜算籽·2026年05月18日 19:11特朗普：本应要求获得更多英特尔公司的股份大公司：索辰科技：物理AI业务目前尚处于布局初期36氪获悉，索辰科技公告，公司股票于2026年5月14日至5月18日连续三个交易日内收盘价格涨幅偏离值累计达到30%，属于股票交易异常波动情形。公司近期实施的募投项目“物理合成数据核心技术基座构建与多领域场景示范应用项目”尚处于筹备阶段，若出现募集资金投资项目未能顺利实施、新技术开发进度不达预期、研发遭遇技术瓶颈甚至失败，将会对公司经营业绩造成一定影响。公司物理AI业务目前尚处于布局初期。物理AI技术研发投入大、技术迭代快，存在研发进度不及预期、核心技术迭代风险；同时下游行业应用场景拓展、商业化变现存在不确定性，若市场需求、行业政策、技术竞争格局发生不利变化，物理AI业务收入及盈利贡献存在不及预期风险。智元WITA成为全国首例完成大模型备案的具身智能交互模型36氪获悉，根据上海市网信办最新公告的上海市生成式人工智能服务备案情况，智元WITA成为全国第一款实现合规备案的具身智能交互大模型，标志着中国具身智能产业从“技术验证”进入“合规商用”的新阶段。依托智元“三智一体”全栈技术体系，智元WITA核心应用于人形机器人交互场景。上交所：上周新增8家IPO受理企业36氪获悉，据“上交所企业上市服务”消息，本期（5月11日—5月17日）新增8家受理企业，其中上交所主板1家，科创板3家，创业板3家，北交所1家。截至5月17日，今年新增31家受理企业。截至5月17日，各板块在审企业共295家。其中上交所62家（主板18家，科创板44家），深交所61家（主板17家，创业板44家），北交所172家。截至5月17日，2026年已公布终止审查（撤材料+否决/终止注册）企业20家，其中上交所7家（主板3家，科创板4家），深交所创业板1家，北交所12家。腾讯音乐：收购喜马拉雅的事项已经完成36氪获悉，腾讯音乐在港交所公告，于2026年5月18日，根据并购协议的条款，收购喜马拉雅的事项已经完成，据此，喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为12.6亿美元的现金及最多总数1.75亿股本公司A类普通股（包括本公司于交割时新发行的A类普通股，以及本公司授出的以股权为基础的激励相关的A类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。爱奇艺：一季度总收入62.3亿元，会员收入环比增长36氪获悉，爱奇艺发布2026年第一季度财报。财报显示，一季度爱奇艺总收入62.3亿元。其中，会员服务收入42.0亿元，环比增长2%。在线广告服务收入12.4亿元，内容发行收入3.6亿元，其他收入4.3亿元。百度：一季度营收321亿元，AI业务收入136亿元36氪获悉，百度发布2026年第一季度财报。财报显示，一季度营收321亿元，上年同期营收324.52亿元。一般性业务收入260亿元，同比增长2%；其中AI业务收入136亿元，占百度一般性业务收入的52%，已连续多个季度增长；GPU云收入同比增长184%。百度创始人李彦宏表示，AI业务收入占比的首次过半，标志着AI已成为百度的核心驱动力。巨力索具旗下十余家企业已注销36氪获悉，近日，巨力索具因涉嫌信息披露误导性陈述违法违规，被证监会立案调查。天眼查App显示，巨力索具股份有限公司成立于2004年12月，法定代表人为杨建国，注册资本9.6亿人民币，经营范围包括应急装备、装备专用工装及工具和箱组的设计、生产和销售等，由巨力集团有限公司、杨建忠、香港中央结算有限公司等共同持股。对外投资信息显示，该公司对外投资26家企业，其中10余家企业处于存续状态，包括巨力索具（河南）有限公司、河北巨力应急装备科技有限公司等，其他10余家企业已注销。投融资：“华羽先翔”完成超亿元Pre-A+轮融资，蚂蚁集团领投36氪获悉，近日，国内全倾转eVTOL企业“华羽先翔”完成超亿元Pre-A+轮融资，本轮融资由蚂蚁集团领投、彼岸时代及复星创富跟投。融资资金将重点用于加速鸿鹄系列eVTOL的适航取证，以及低空出行及商业场景的落地与创新。“极壳科技”获5000万美元B+轮融资36氪获悉，近日，外骨骼品牌“极壳科技”宣布完成5000万美元B+轮融资。此轮融资由蚂蚁集团和美团龙珠领投，Sofina、Granite Asia跟投，高鹄资本担任独家财务顾问。此轮融资将加速极壳在技术研发、全球市场拓展及消费场景应用方面的布局。新产品：酷开发布企业AI操作系统36氪获悉，近日，深圳市酷开网络科技股份有限公司举办“场景智能·硅基管理”发布会，正式推出原生AI企业操作系统——酷开Happy Work AIOS Lite MVP。该系统已实现从生活场景到企业经营场景的全维度覆盖，完成多场景AI生态闭环落地。今日观点：马斯克预计今年美国将广泛使用自动驾驶汽车埃隆·马斯克周一表示，他预计今年晚些时候，无需人类驾驶的自动驾驶汽车将在美国更加普及。马斯克通过视频连线在特拉维夫举行的“智能出行峰会”上发言时称，得克萨斯州的三个城市已有无人驾驶汽车在没有安全监视员的情况下运行，并补充说今年这一模式将扩展至全美。（新浪财经）其他值得关注的新闻：2026电影总票房已超145亿元据票务平台数据，截至今天下午，2026年电影票房已超145亿元。五一档新片长尾效应持续释放，其中《给阿嬷的情书》连续九天蝉联单日票房榜首。（央视新闻）特朗普：本应要求获得更多英特尔公司的股份美国总统唐纳德・特朗普在周一出版的杂志采访中表示，他本应代表美国政府要求获得更多英特尔股份。特朗普政府去年持有英特尔10%的股份，并宣布向这家芯片制造商投资约100亿美元，用于在美国境内新建或扩建工厂。交易完成八个月后，美国政府持有的英特尔股份市值已超过500亿美元。（新浪财经）国家药监局：我国4月批准注册医疗器械产品318个今天从国家药监局获悉，我国4月共批准注册医疗器械产品318个。其中，境内第三类医疗器械产品272个，进口第三类医疗器械产品20个，进口第二类医疗器械产品26个。 (央视新闻)本文由「卜算籽」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：AI生成

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生8点1氪·2026年05月19日 08:08贾跃亭发布视频，宣布法拉第未来近期完成了7000万美元机构投资者募资。今日热点导览瑞幸咖啡正式上线两款含酒精的特调饮品京张高铁试点推出“自行车随身行”服务马斯克：计划今年进行首例用于治疗失明的Neuralink植入手术韩国法院要求三星电子罢工行动不得影响产量多家理财公司集中提前终止理财产品TOP 3 大新闻贾跃亭又拿到7000万美元投资贾跃亭发布视频，宣布法拉第未来近期完成了7000万美元（现汇率约合4.77亿元人民币）机构投资者募资，并将启动战略、产品技术&业务、财务、资本、AI体系五大变革。贾跃亭表示，其目标包括FFAI市值重回2021年上市水平，2027年第四季度实现经营性现金流为正，五年内在EAI人形与仿生机器人领域成为北美真实场景部署量前三的机器人公司+北美EAIMPV领军者。贾跃亭还表示，要将FF真正重构为一家Physical AI生态公司。两年内实现过去12年没有完成的梦想，进入“0借口、0内耗、结果导向”的战斗状态，全力兑现对股东和用户的承诺。（IT之家）巨力索具旗下十余家企业已注销36氪获悉，近日，巨力索具因涉嫌信息披露误导性陈述违法违规，被证监会立案调查。天眼查App显示，巨力索具股份有限公司成立于2004年12月，法定代表人为杨建国，注册资本9.6亿人民币，经营范围包括应急装备、装备专用工装及工具和箱组的设计、生产和销售等，由巨力集团有限公司、杨建忠、香港中央结算有限公司等共同持股。对外投资信息显示，该公司对外投资26家企业，其中10余家企业处于存续状态，包括巨力索具（河南）有限公司、河北巨力应急装备科技有限公司等，其他10余家企业已注销。联讯仪器股价反超贵州茅台，成为A股新的股王5月18日，联讯仪器盘中扩大涨幅至13%，股价反超贵州茅台，成为A股“新股王”，该股较发行价上涨超15倍。截至5月18日11时许，联讯仪器最高涨至1361元/股，涨幅高达16.32%。与此同时，贵州茅台由涨转跌，股价探至1330元之下。此前相继登上A股“股王”宝座的寒武纪（688256）、源杰科技（688498），5月相继迎来2025年年度权益分派。公开资料显示，联讯仪器4月24日登陆A股，发行价81.88元/股，上市14个交易日来，目前股价较发行价上涨15.6倍。据公开资料显示，联讯仪器是国家级专精特新“小巨人”企业，专注于高速通信、电性能、光芯片及半导体晶圆芯片等领域测试仪表与设备的研发、制造，是国内高端电子测量与半导体测试设备领域的硬核龙头。（澎湃新闻）大公司/大事件何立峰会见美国超威半导体公司董事会主席兼首席执行官苏姿丰5月18日下午，中共中央政治局委员、国务院副总理何立峰在人民大会堂会见美国超威半导体公司董事会主席兼首席执行官苏姿丰。何立峰指出，上周中美两国元首在北京举行重要会晤，达成一系列重要共识。在两国元首战略引领下，两国经贸团队达成了总体平衡积极的成果，这将为下一步的中美经贸合作与世界经济注入更多确定性和稳定性。欢迎包括超威半导体公司在内的跨国公司把握中国发展机遇，深化互利合作。苏姿丰积极评价两国元首会晤成果，表示愿继续拓展在华业务，持续加大在华投资。(新华社）美陪审团驳回马斯克对OpenAI的诉讼当地时间5月18日，美国陪审团裁定，美国开放人工智能研究中心（OpenAI）及其首席执行官萨姆·奥尔特曼无需就“偏离慈善使命”一事对埃隆·马斯克承担责任，原因是马斯克提起诉讼时间过晚。当天晚些时候，马斯克表示，将就这一裁决向美国第九巡回上诉法院提起上诉。此前，马斯克对OpenAI及其首席执行官奥尔特曼提起诉讼，指控OpenAI最初邀请马斯克资助一个致力于造福人类的非营利AI研究机构，但如今却专注于盈利。（央视新闻）张朝阳：搜狐未入局大模型研发，将聚焦AI应用落地36氪获悉，在2026搜狐科技年度论坛上，搜狐CEO张朝阳表示，“搜狐并未入局大模型第一阵营的研发，大模型研发需要雄厚的资金实力和技术背景，企业应基于自身业务基础理智选择。”他提到，搜狐将聚焦AI应用落地，在提升编程效率、节约运营成本的同时，对内容生成保持克制，保持内容中立性，规避AI带来的内容乱象风险。腾讯音乐：收购喜马拉雅的事项已经完成36氪获悉，腾讯音乐在港交所公告，于2026年5月18日，根据并购协议的条款，收购喜马拉雅的事项已经完成，据此，喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为12.6亿美元的现金及最多总数1.75亿股本公司A类普通股（包括本公司于交割时新发行的A类普通股，以及本公司授出的以股权为基础的激励相关的A类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。川投新能源公司总经理刘亮被遣返回国据中央纪委国家监委网站，近日，在中央反腐败协调小组国际追逃追赃工作办公室统筹协调和公安部等部门协助下，经四川省监察机关、公安机关与有关国家执法机关密切合作，外逃职务犯罪嫌疑人刘亮在境外落网并被遣返回国。刘亮，男，1979年1月出生，川投新能源公司总经理，涉嫌严重职务违法犯罪，于2026年4月外逃，同月四川省资阳市监察机关对其立案调查。办案机关积极开展国际执法合作，于近日将其缉捕归案。（界面新闻）英特尔CEO称代工业务取得进展英特尔CEO陈立武在当地时间周一接受采访时表示，该公司的外部代工业务正在稳步发展，成为其扭亏为盈的关键一环。“代工业务非常重要，”陈立武表示，“它是重要的国家资产之一。”（新浪财经）瑞幸咖啡上线两款含酒精的特调饮品5月18日，瑞幸咖啡正式上线两款含酒精的特调饮品，并且属于饮料酒的范畴，正式布局“微醺”经济。记者了解到，这两款含酒精饮品是瑞幸对全国市场的一次测试，后续或还有进一步的计划。在业内看来，瑞幸布局微醺赛道，意在盘活闲置门店时段资源，提效增收。记者在北京部分瑞幸门店看到，两款年度特调饮品已经上线，但只有部分门店可以提供含酒精版。（红星新闻）韩国法院要求三星电子罢工行动不得影响产量据央视新闻消息，韩国水原地方法院5月18日批准了三星电子公司提出的部分禁令请求，责令该公司工会必须确保即将开始的全面罢工行动“不影响产量”，“不得导致这家全球重要的存储芯片制造商生产原料受损”。据悉，这实际上是给三星电子这次史上最大规模罢工计划“踩了急刹车”。今年3月中旬，由超过6.6万名三星电子工会成员参与表决的投票结果显示，93.1%的工会成员赞成罢工。全面罢工将于5月21日至6月7日举行。据悉，三星电子作为全球重要的存储芯片制造商，一旦发生罢工，可能进一步加剧由全球人工智能数据中心建设持续升温带来的全球半导体供应趋紧局面，其影响或波及汽车、计算机及智能手机等多个行业。三星电子劳资双方此前就2026年薪资问题进行了多轮谈判，但双方就部分议题的分歧持续扩大，谈判最终破裂。（央视新闻）多家航司下调国际航线燃油费，国内航线刚上调国内航线燃油附加费自5月16日再度上调后，多家航司却在陆续下调国际航线的燃油附加费。香港航空宣布，自今日（5月18日）开始下调多条航线的燃油附加费。其中中国香港前往内地的地区航线，燃油附加费由290港元下调至190港元；香港前往日韩、新马泰等则由389港元下调至339港元。在此之前，国泰航空和香港快运已宣布自5月16日起下调部分区域的燃油附加费。其中香港始发长途航班从1560港币下调至1362港币，中程航班从725港币下调至633港币，短途航班从389港币下调至339港币。下调幅度在12%左右。这也是中东冲突爆发后，多家航司首次对核心航线燃油附加费进行下调。（第一财经）5月19日起，铁路部门在京张高铁试点推出“自行车随身行”服务据新华社消息，为更好地满足旅客多样化出行需求，铁路部门5月19日起在京张高铁北京北至崇礼间各车站间运行的G字头动车组列车试点推出“自行车随身行”服务，旅客可通过12306客户端线上预约并付费，乘车时经车站安检和规范包装，可携带符合规定的自行车同车出行，将为广大骑行爱好者跨区域出行提供更多便利。试点初期，“自行车随身行”服务定价88元/辆，旅客可在乘车结束后180日内通过铁路12306客户端开具电子发票。（新华社）戴尔将与谷歌和SpaceX等公司展开合作戴尔科技在上一季度为其“AI工厂”产品线新增1,000家客户，使其总客户数达到了5000家。戴尔正在推出多款新产品以协助企业搭建AI应用，其中包括一款名为“Dell Deskside Agentic AI”（戴尔桌面智能体AI）的产品。此外，戴尔还与谷歌和SpaceX等公司展开合作，将AI模型与工具引入企业自身的内部网络中。（财联社）SpaceX将迎关键试飞SpaceX计划于周二进行一次关键的试飞，发射其星舰巨型火箭的升级版，代号V3。这次发射正值SpaceX上市关键阶段，消息透露预计该公司最早在下月IPO，并可能在本周三公开招股书。与此同时，SpaceX在与贝索斯的蓝色起源竞争，为NASA的阿尔忒弥斯计划开发星舰，供其在2028年的载人登月任务中使用。目前来看，星舰的研发进度落后于NASA对SpaceX的预期。该火箭原计划于2023年首飞，但去年的失败延缓了其发展进程。星舰最近一次试飞，也是其第11次试飞，发生在七个月前。（财联社）NextEra拟以660亿美元收购Dominion，或创史上最大规模电力并购纪录5月18日消息，据报道，美国公用事业巨头NextEra Energy目前正就收购竞争对手Dominion Energy进行谈判，该笔交易方案主要以股票交换形式进行，对Dominion的估值约为每股76美元，总价约660亿美元。若最终达成，这将成为迄今为止电力行业规模最大的一笔交易。NextEra计划以约0.8股自身股票交换1股Dominion发行在外的股票，交易总额中还将包含一小部分现金成分。交易完成后，NextEra的股东将持有合并后新公司约75%的股份。（界面新闻）马斯克：计划今年进行首例用于治疗失明的Neuralink植入手术马斯克表示，旗下Neuralink技术能帮助部分失明者恢复视力，计划今年进行首例用于治疗失明的Neuralink植入手术。据了解，Neuralink是一个美国神经科技和脑机接口公司，由伊隆·马斯克和八名其他联合创办者创办，负责研发植入式脑机界面技术。公司的总部在旧金山。Neuralink于2016年成立，并于2017年3月公之于众。2019年7月17日，Neuralink对外宣布该公司研发的一款脑机接口系统（财联社）五粮液：2025年度百亿分红两月内到账，严格兑现200亿元分红承诺5月18日，五粮液2026年第一次临时股东会对《2025年度利润分配方案》进行了审议，五粮液副董事长、总经理华涛表示，2025年度分红资金将在两个月内准确、及时划入全体股东账户。五粮液2025年中期已分红100亿元，本次拟再分红100.07亿元。五粮液践行高比例分红，严格兑现《2024—2026年股东回报规划》承诺，连续3年每年度现金分红总额不低于200亿元（含税）。（证券时报）美国连锁快餐品牌Wendy’s将进入中国市场5月18日在美国连锁快餐品牌Wendy’s官网上看到，该品牌国际特许经营重点市场亚太区域中出现了中国市场。早前，Wendy’s在2026年第一季度的财报中称，公司已经签署一项特许经营协议，计划未来10年在中国建设最多1000家餐厅。Wendy’s称合作方是“一家在中国拥有数十年经验的大型餐饮运营商，以强待客能力闻名、深刻理解中国消费者。”但没有公布这家运营商的名称。此外，财报显示，Wendy’s2026年一季度全球销售额同比下降5.5%，美国同店销售下降7.8%，国际市场销售额增长6.0%。（界面新闻）贵州茅台：优化茅台酒线下门店营业时间及i茅台App开售时间36氪获悉，据贵州茅台消息，为进一步贴近广大消费者日常工作与生活节奏，更好地满足广大消费者购酒需求，自2026年5月

高瓴出手了一家AI体育科技公司，曾获李泽湘天使轮融资｜硬氪首发黄 楠·2026年05月19日 09:30从“记录工具”到“AI运动Copilot”的跃迁。作者｜黄楠编辑｜袁斯来硬氪获悉，AI体育科技公司SportVision近日已完成天使+轮融资，高瓴创投（GL Ventures）领投，谦恒资本担任独家财务顾问。本轮资金将重点用于核心技术研发、产品量产落地及市场的拓展推广。其新款AI网球训练相机产品计划发售中。SportVision隶属于深圳思博威视体育科技有限公司‌，长期专注运动算法研发与硬件产品应用、构建AI智能教练体系，此前已获得李泽湘教授旗下清水湾基金的天使轮投资。过去几年，运动科技赛道几乎被硬件厂商重新定义了一遍。发球机让“一个人练球”成为可能，运动相机令“训练过程的记录”不再需要专人跟拍，智能手表/手环把心率、步频、消耗卡路里等量化成精确的长期数据记录。尽管每一类产品都跑出了规模不小的公司，但它们各自停留在陪练、记录、监测的单点功能上——数据相互独立、能力无法贯通，难以为用户提供连贯、整体的个性化支持。运动人群的深层渴望远不止于被记录，还有被指导、被陪伴、被看见进步。一场更底层的变革正在发生：从“记录工具” 到 “AI运动Copilot”的跨越，运动科技赛道正在重写。工具只需完成指令，而Copilot则可以像教练一样理解状态、识别动作、纠正习惯，甚至在运动热情消退时成为那个督促陪伴的人。要补上这道体验鸿沟，用户需要的不只是更强性能的硬件，更是一个能把整合所有信息、真正理解运动者的“大脑”。这也正是SportVision精准切入的核心位置。全运会现场所使用的SportVision设备（图源/企业）SportVision创始人陈楷夫本科毕业于吉林大学物理学院严济慈班、哥伦比亚大学机器人学硕士，曾在MMLab从事姿态识别与目标检测算法研究；身为一名前游泳运动员，他对运动学习、动作纠正与训练效率的痛点有切身感知，正是这段经历，让他决心用AI技术破解非结构化运动数据难题，构建一个可规模化、可复制的个性化运动产品和训练体系。联合创始人杨晨旭是资深网球爱好者，曾在腾讯、字节负责AI产品开发。多位团队核心成员来自商汤、字节、阿里、腾讯等头部科技企业，具备AI算法、软硬件一体化研发与全球化品牌运营能力，已形成了从技术研发到产品落地的完整闭环。公司起步初期，SportVision选择从羽毛球场景切入进行关键产品验证，首款产品“好球哇”是一套羽毛球AI运动相机方案。设备可实时采集球场画面，通过视觉算法自动识别并捕捉用户的运动高光时刻，为普通运动爱好者、专业运动员及赛事场景，提供高光集锦、赛事直播、运动数据分析等一体化服务。这套系统落地了全国数十家头部羽毛球馆，并获得林丹、王睁茗、龚睿那等世界冠军在内的多位种子用户支持。在赛事侧，该系统也服务了全运会残特奥会、粤港澳大湾区羽毛球比赛、深圳杯等多场关键赛事。但它真正的价值，远非硬件本身，而是持续运行中沉淀的海量运动数据。在拍摄过程中，SportVision自研算法能实时分析并智能预测运动轨迹，根据比赛进程自动调整摄像角度，对复杂的多人运动场景进行针对性学习优化。从世界冠军到刚入门的爱好者，从专业教练的指导语料到普通用户的练习习惯，场馆设备每天运行十几个小时，能收集到各类各样的泛化数据。每一场球、每一次挥拍、每一个得分瞬间，都在被系统记录、拆解、分析。这是一座难以被发球机或运动相机采集到的数据矿。运动学习的基本机制分为三个阶段：认知阶段（通过视觉掌握动作要领）、联想阶段（在反复练习中形成神经通路）、自主阶段（动作固化成为肌肉记忆）。其中联想阶段的及时反馈最为关键：动作做对还是做错需要在几秒内被确认，否则错误动作就会被“练熟”。陈楷夫向硬氪指出，“此前，多数运动AI解决方案仍停留在‘事后分析’阶段，用户完成运动后上传视频、系统给出复盘报告。但这种模式错过了最佳纠正窗口。”因此，SportVision想搭建的，是一套基于实时交互的预判式AI。“好的运动员不一定是好教练，好教练的价值在于知道怎么教。”陈楷夫说，“我们要做的，是把教练的教法萃取出来，让AI学会怎么教。”随着网球、羽毛球、匹克球等拍类运动在全球范围内持续升温，用户对运动提升的需求愈发迫切。传统的请教练模式面临成本高、频次低的矛盾，一周请一两次教练，用户练习时无法得到反馈，对于向网球这样的运动，可能需要一年才能从入门到实现“对打社交”。正反馈周期长，也使得很多人停留在了入门阶段。这正是SportVision的核心壁垒所在：那些在球场上日夜运转的设备，收集到的海量数据。从骨骼点识别到动作模式分析，从专业运动员的动作库到普通用户的练习轨迹，这些数据构成了可迁移的技术底座，能够在各类拍类运动中得以技术复用和迁移。SportVision首款网球AI Copilot产品（图源/企业）“我们做的不是一款运动相机，而是构建未来运动人群个性化AI教练的世界模型。”陈楷夫告诉硬氪，“用户甚至不需要思考‘我该怎么练’，而是由系统主动管理你的运动进步。”硬氪获悉，SportVision首款网球AI Copilot产品正在紧密的研发流程中。以相机为载体，系统可识别用户运动动作并与专业标准动作比对，实时反馈纠正建议；同时，结合教练指导场景，生成个性化训练方案，让用户在独自练习时也能获得针对性、科学化的AI陪练指导。该产品概念此前已在CES 2026上展出，首发市场聚焦欧美地区。这里拥有了全球密度最高的网球用户群体，也是付费意愿与消费能力最突出的核心市场。回到AI时代运动科技竞争的底层逻辑迁移，当大模型逐渐消解知识鸿沟，竞争的壁垒不再停留于算法参数，而是向两端迁移——一端是数据的密度与深度，另一端是场景的介入与闭环。SportVision长期积累的稀缺数据资产，正在成为其下一代AI教练产品的技术底座。从情绪价值到功能性与情绪价值兼具、从记录到指导，从单一品类到拍类运动通用技术底座，这家公司正在尝试的命题是，加速推动运动科技从“工具”到“AI Copilot”的关键升级。本文由「黄 楠」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：企业授权、视觉中国

影石：重新定义性价比黄绎达·2026年05月19日 14:58影石有个“苹果梦”？作者 | 张帆坚持做后来者近日，影石创新发布了登陆资本市场后的首份年报。2025年，公司实现营业收入97.41亿元，同比大幅增长74.76%，创下成立以来的最高营收纪录。然而，同期由于研发、销售等费用高企，公司归母净利润仅为9.29亿元，同比下降了6.62%。面对这样一份“营收大增、利润下滑”的财报，公司召开了上市后的首次投资者交流会，以回应资本市场的关切。根据2025年年报显示，报告期内，公司影像业务营收增长了78%，主要得益于销量的大幅提升——2025年实现销量393万台，同比增长76%。另一方面，产品均价相对平稳，约为2169元，同比增长1%。从ASP趋势来看，公司产品整体呈现下降态势，从2022年的2,391.45元/台连续两年下滑至2024年的2147元，但在2025年实现了企稳。在当前激烈的市场竞争环境下，价格企稳的主要原因，是受益于公司积极拓展B端等高单价客户的提振。尽管有争议，但不可否认的是，即便面临激烈的价格竞争，影像行业依旧是消费电子中的高毛利子行业。根据2025年的财报，影石毛利率高达48.82%，与科沃斯相当，但相比小米、传音等手机厂商约20%左右的毛利率。影石的毛利率高出约28个百分点。图：各消费电子企业毛利率；资料

14小时前法官确认裁决结果，驳回马斯克对 OpenAI 的诉讼美国加州北区联邦地方法院法官 Yvonne Gonzalez Rogers 近日确认最新裁定结果：驳回埃隆·马斯克要求对 OpenAI 和微软签发“全面禁制令”的动议与其先前提出的欺诈类指控，但保留关于“是否存在并被违反的公益/慈善信托”以及“不当得利”等核心争议，案件不会就此终结，而是将进入加速审理并由陪审团作出最终裁决。3 来源法官确认：禁制令与部分指控被否，核心诉求仍将审理3 来源从出资联合创办到天价索赔：马斯克与 OpenAI 的纠纷缘起3 来源法院为何驳回禁制令与部分指控：证据不足与适格障碍3 来源影响与展望：AI 治理、非营利结构与科技巨头关系迎来“法律考题”3 来源本内容由AI生成

6小时前NASA测试抗辐射太空AI芯片美国国家航空航天局（NASA）喷气推进实验室正测试一款由其与美国微芯公司联合开发的新型抗辐射太空AI处理器，隶属于“高性能航天计算”（HPSC）项目。该芯片自今年2月进入环境与功能测试阶段，在辐射、热冲击和机械冲击等严苛条件下运行良好，初步结果显示其性能约为当前航天器常用抗辐射芯片的500倍，有望显著提升未来深空探测器和载人登月、火星任务中航天器的自主智能和实时决策能力。7 来源NASA推进高性能航天计算：新一代抗辐射AI芯片登场3 来源通过辐射、热冲击与机械冲击考核“太空资格”3 来源深空任务的“机载大脑”：让航天器更自主2 来源太空计算范式演进：光子AI芯片与商用芯片加固并进5 来源本内容由AI生成

5小时前腾讯云Hy3 preview和DeepSeek-V4-Pro模型将结束限时免费公测腾讯云正加速收紧大模型“限时免费”策略：在今年3月已结束GLM 5、MiniMax 2.5、Kimi 2.5等第三方模型的免费公测并大幅上调混元自研模型价格后，官方又公告DeepSeek V4 Flash将于5月7日起结束限免转为正式计费，同时业内普遍预计作为混元新一代核心的Hy3 preview以及高端版本的DeepSeek-V4-Pro，也将按腾讯云“限时免费随时可终止”的既定规则陆续进入商业化阶段，开发者和企业使用AI模型的成本将显著上升。8 来源计费调整时间线：从3月“涨价潮”到5月DeepSeek商业化6 来源Hy3 preview 与 DeepSeek 系列：从“明星体验模型”走向核心付费能力8 来源开发者与企业的现实影响：AI调用成本显著抬升7 来源资本开支与行业格局：免费时代退潮，“补船票”成共识4 来源本内容由AI生成

2小时前腾讯微信AI团队拿下ICASSP最佳工业论文奖目前在可公开检索的信息源中，尚未查到“腾讯微信AI团队拿下ICASSP最佳工业论文奖”的正式报道或大会公告，因此该消息尚无法被独立核实；若消息属实，将意味着微信AI团队在语音/多模态等工程落地方向的成果获得国际顶级信号处理会议对工业价值与技术创新的双重认可。3 来源事件概述：消息尚待公开佐证ICASSP与“最佳工业论文奖”的重要性1 来源微信AI团队的技术布局与可能涉及的研究方向2 来源中国企业在语音、多模态与大模型领域的国际竞争态势3 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。智库·502线上同行往期回顾报名已结束企业构建AI+组织的人效杠杆与路径选择百果园集团x银泰分享企业构建AI+组织的人效杠杆与路径选择妙投波动判断场往期回顾活动已结束芯片、光模块、机器人——AI的泼天富贵，还能传到第几棒？本期我们将AI产业链拆解为三大核心环节，逐一拆解其核心矛盾与风险边界，帮你建立从芯片到机器人的完整坐标系。随行灯塔火热报名中随行灯塔随行灯塔是由虎嗅深度联合运营的高端创始人社群。专注于一件事：帮助创始人把产品做对、做爆、做到10亿。

这里不灌鸡汤，不讲玄学。我们相信，伟大的公司一定是产品驱动的；但产品力来自人，而人需要场——一个能被照见、能被挑战、能被链接的场。随行灯塔，就是创始人的“修炼场”。智库·502线下闭门会往期回顾精彩回顾AI在ToB销售、客服的落地案例探讨围绕着AI在ToB营销内容生成、智能客服接待、线索清洗等场景进行深入交流早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

关注最新看重点大公司财经动态看世界硅谷动态风向·标同步官方重大消息有媒体人士呼吁斯威士兰尽早与中国建交，外交部回应：今年5月1日起，中方对53个非洲建交国全面实施零关税，中国大市场向非洲敞开大门。5月19日，外交部发言人郭嘉昆主持例行记者会。有记者提问，据报道，今年以来中非合作取得积极进展，前4个月双边贸易同期首次突破8000亿元。非洲特色好物进入中国消费市场，中国优质产品也在赋能非洲发展建设，非洲多国人士将中非合作称为拉动非洲发展的“加速器”。也有媒体人士呼吁斯威士兰尽早与中国建交，不要一再错失发展机遇。请问发言人如何看待这一现象？郭嘉昆指出，在中非合作论坛和共建“一带一路”框架下，中非全方位合作蓬勃发展，助力非洲各国工业化和农业现代化进程，也给非洲人民带来了实实在在的好处。今年5月1日起，中方对53个非洲建交国全面实施零关税，中国大市场向非洲敞开大门。我们真诚希望所有非洲国家都能成为中非合作“全家福”中的一员，搭乘中国发展快车，融入中非友好，实现中非共同现代化。2分钟前00举报发表0/500虎嗅用户社区交流公约聚焦315抵制假冒伪劣，维护消费者权益市场监管总局：“四新经济”个体工商户超4000万户。记者今天（19号）从市场监管总局发布会获悉，个体工商户是民营经济的重要组成部分，是我国数量最多的经营主体。截至2025年底，我国“四新经济”个体工商户达到4010.9万户，连续五年保持增长态势。《中华人民共和国民营经济促进法》实施后，为解决个体工商户面临的创业投资、转型升级、市场退出等难点堵点，市场监管总局出台了《个体工商户登记管理规定》，在规范登记管理的同时，强化制度创新，更加注重权益保护，实现多重便利，有效激发个体工商户创业创新活力。4分钟前00举报发表0/500虎嗅用户社区交流公约科技圈，职场人有些事儿，只有咱圈内人懂亿嘉和与宜通世纪达成战略合作，共同探索机器人与AI技术行业应用场景。5月11日，亿嘉和与宜通世纪举行战略合作签约仪式。双方围绕机器人技术、人工智能、物联网平台及行业数字化应用等方向展开深入交流，并正式达成战略合作。双方战略合作目前处于前期探索阶段，未来将围绕智能巡检、智慧机房、物联网平台及AI数据应用等方向，共同探讨机器人与AI技术的融合应用场景，推动产品与行业需求的协同创新。双方也将加强技术交流与资源共享，持续挖掘数字化、智能化升级背景下的合作可能。8分钟前00举报发表0/500虎嗅用户社区交流公约新政新规各种新政策多地强降雨持续，国家防总继续维持针对湖北、湖南、广西、贵州的防汛四级应急响应，19日增派工作组赴贵州协助指导防汛救灾工作。5月19日，国家防总办公室、应急管理部继续组织气象、水利、自然资源等部门和湖北、湖南、广西、重庆、贵州等省份，会商研判本轮强降雨发展趋势，部署重点地区防汛救灾工作。会商指出，本轮强降雨过程是今年入汛以来我国最强降雨过程，当前主雨区进一步东移南压，湖北、湖南、广西、贵州、江西、山东、安徽等省份仍有强降雨。一些地区土壤含水量饱和、江河湖库底水偏高，山洪地质灾害、中小河流洪水、城市内涝等风险进一步加大。会商强调，各地各有关部门要坚持预防为主，强化隐患排查整治，突出涉水路段和漫水桥安全度汛管理，深化山洪灾害危险区、地质灾害隐患点、城市易淹易涝点、中小水库、河道堤防等重点部位隐患排查，加强强降雨和强对流天气可能引发的矿山淹井、危化品泄漏、在建工程基坑坍塌等次生事故防范工作。要强化人员转移避险，加强雨情汛情监测预警，落实直达基层责任人的临灾预警“叫应”和反馈闭环工作机制，突出抓好山洪沟口、狭窄行洪沟道等高风险区域和养老院、施工工地、工棚、民宿等重点部位人员转移，针对老弱病残孕等重点群体实施“一对一”帮扶措施，坚决做到应转早转、应转尽转。要强化预置布防，在重要区域、重点部位、重大风险隐患点等提前预置抢险救援力量、物资和应急通信装备，确保灾害来临时能快速反应、高效处置。国家防总继续维持针对湖北、湖南、广西、贵州的防汛四级应急响应。国家防总办公室19日增派工作组赴贵州协助指导防汛救灾工作，国家防总办公室、国家防灾减灾救灾委员会办公室前期派出的3个工作组继续在湘鄂桂三省份重点地区开展工作。9分钟前00举报发表0/500虎嗅用户社区交流公约硅谷动态这里本来有条个人简介Mistral收购了总部位于维也纳的Emmi AI，交易金额未披露，以提升其在欧洲的工业产品；Emmi在2025年奥地利最大的一轮融资中筹集了1500万欧元。14分钟前00举报发表0/500虎嗅用户社区交流公约金融圈金融动态追踪盛宝银行：尽管特朗普暂停打击，黄金期货仍受利率预期拖累。随着伊朗冲突持续，受利率上升预期的拖累，金价维持在每盎司4,500美元上方。纽约期金下跌0.2%，至每盎司4,549.30美元。消息称美国总统特朗普决定推迟对伊朗的计划打击，以为谈判留出更多时间，这提供了一些支撑，尽管自战争开始以来金价已下跌超过13%。盛宝银行分析师表示：“交易员仍关注中东危机以及能源价格持续走高带来的通胀影响。紧张局势升级可能会通过收益率上升和美元走强拖累黄金，而任何实现降温或和平的可靠途径最终都可能支撑金价。”15分钟前00举报发表0/500虎嗅用户社区交流公约风向·标同步官方重大消息乌克兰称一架俄罗斯无人机在黑海击中一艘中国货船，外交部：该船系马绍尔群岛籍，船上有中国籍船员，无人伤亡。5月19日，外交部发言人郭嘉昆主持例行记者会。有记者提问，乌克兰方面昨日表示，一架俄罗斯无人机在黑海击中了一艘中国货船。请问外交部对此有何评论？“我们注意到有关报道。”郭嘉昆表示，根据目前掌握的情况，相关船只系马绍尔群岛籍，船上有中国籍船员。中国驻乌克兰使馆已与中国船员取得联系，确认无人员伤亡。郭嘉昆强调，关于乌克兰危机，中方的立场是一贯的、明确的，对话谈判是解决乌克兰危机的唯一出路。17分钟前10举报发表0/500虎嗅用户社区交流公约风向·标同步官方重大消息外交部回应菲律宾总统涉台言论：侨民众多、地理邻近不是一国干涉他国内政和染指他国主权事务的借口。5月19日，外交部发言人郭嘉昆主持例行记者会。彭博社记者提问，菲律宾总统马科斯称，菲律宾并不希望卷入任何与台湾有关的战争，但由于地理位置临近，菲律宾有很多国民在台湾。中方对此有何评论？郭嘉昆表示，中方注意到菲方领导人重申奉行一个中国政策，不干涉中国内政。需要强调的是，侨民众多、地理邻近不是一国干涉他国内政和染指他国主权事务的借口，希望菲方履行承诺、言行一致，以实际行动恪守一个中国原则。20分钟前00举报发表0/500虎嗅用户社区交流公约一只飞虎最近国外发生了什么？欧盟拟削减美国进口关税，以避免特朗普上调关税。欧盟谈判代表预计将于周二达成一致，取消对美国商品的进口关税，以履行去年与美国达成的贸易协议，并抵御美国总统唐纳德・特朗普威胁加征高额关税的压力。根据去年7月特朗普在苏格兰特恩贝利高尔夫度假村达成的协议条款，欧盟同意取消对美国工业商品的进口关税，并为美国农产品与海产品提供优惠市场准入。作为交换，美国将对绝大多数欧盟商品征收15% 的关税。然而，近10个月过去，欧洲议会与代表欧盟各国政府的欧盟理事会仍需就立法文本达成一致，欧盟关税减免措施方能生效；双方核心分歧在于，一旦特朗普违背协议，欧盟应设置何种保障机制。21分钟前00举报发表0/500虎嗅用户社区交流公约风向·标同步官方重大消息中方：欢迎普京总统。中国外交部发言人郭嘉昆19日主持例行记者会。有记者就俄罗斯总统普京访华一事提问。郭嘉昆：中方欢迎普京总统对华进行国事访问。中俄是新时代全面战略协作伙伴，普京总统此访期间，两国元首将就双边关系、各领域合作以及共同关心的国际地区问题交换意见。23分钟前00举报发表0/500虎嗅用户社区交流公约金融圈金融动态追踪豪森智能：公司及相关人员收到大连证监局责令改正决定书。豪森智能公告称，公司于2026年5月19日收到大连证监局《决定书》。经查，公司收入相关内部控制不健全，部分客户验收报告日期签署不完整、个别项目验收报告表述存在差错。董事长兼总经理董德熙、副总经理兼董事会秘书许洋、副总经理杨宁对上述情形负主要责任。大连证监局决定对公司及相关人员采取责令改正的监管措施，要求30日内提交书面整改报告。公司及责任人将积极整改，本次监管措施不影响公司正常经营。24分钟前00举报发表0/500虎嗅用户社区交流公约时讯内参树欲静而风不止，新鲜资讯速递OECD称全球经济前景面临进一步恶化风险。经合组织(OECD)秘书长马蒂亚斯·科尔曼在七国集团(G-7)会议间隙表示，随着中东冲突持续，全球经济面临的风险正在加剧。“总体评估认为，这给经济增长带来了下行压力，同时给通胀带来了上行压力，”科尔曼周二接受采访时称。“我们在3月经济评估中给出过初步预测，但未来几周将正式修订该评估。不过，增长面临下行压力、通胀面临上行压力这一点显而易见。”早在3月，总部位于巴黎的OECD就成为首个警告伊朗战争将推高物价并抑制经济活动的主要国际机构。该机构定于6月3日发布最新预测。科尔曼表示，权衡是否要加息时，各国央行面临的挑战是如何应对通胀风险高企和经济活动疲软并存的局面。“通常情况下，能源价格冲击的影响是可预见的，但如果冲击开始更广泛地波及其他价格，如果出现工资上涨等次生影响，那么即使经济增长前景略显疲软，央行也需要采取行动，”科尔曼指出。25分钟前00举报发表0/500虎嗅用户社区交流公约一只飞虎最近国外发生了什么？伊朗副外长公布伊对美协议草案提议，包括美军从周边地区撤离。加里巴巴迪表示，伊朗在近期的提议中，明确强调了伊朗和平利用核能的权利、要求结束包括黎巴嫩在内的所有战线的战争、解除美国对伊朗的海上封锁、归还和解冻伊朗遭冻结资产、要求美国提供发动战争造成伊方损失及重建所需的赔偿、终止所有单边制裁以及美军撤离伊朗周边地区。30分钟前00举报发表0/500虎嗅用户社区交流公约风向·标同步官方重大消息外交部：中美同意开展人工智能政府间对话。郭嘉昆表示，作为两个人工智能大国，中美双方应该携手促进人工智能发展和治理，推动人工智能更好服务人类文明进步和国际社会共同福祉。特朗普总统访华期间，两国元首就人工智能问题进行了建设性交流，同意开展人工智能政府间对话。33分钟前10举报发表0/500虎嗅用户社区交流公约一只飞虎最近国外发生了什么？美国石油储备周出库量创下新高，IEA警告：商业库存仅剩几周！美国的战略石油储备正在快速消耗。美国能源部周一公布的数据显示，上周美国战略石油储备的石油出库量达到创纪录的990万桶，战略储备的石油总量降至约3.74亿桶，达到2024年7月以来的最低水平。国际能源署（IEA）在其最新的月度石油市场报告中则指出，3月和4月全球观测石油库存以创纪录的速度下降，减少了2.46亿桶。国际能源署署长Fatih Birol周一强调，由于伊朗战争和霍尔木兹海峡的实际关闭，商业石油库存正在迅速减少，仅剩几周的供应量。虽然战略石油储备的释放使市场每天增加了250万桶石油供应，但这些储备并非无穷无尽。Birol补充道，随着北半球春季播种和夏季旅游旺季的到来，柴油、化肥、航空燃油和汽油的需求增加，库存消耗速度将会加快。供应危机　　国际能源署表示，截止5月8日，其32个成员国已释放了1.64亿桶原油储备。3月，该机构协调了有史以来最大规模的战略储备释放行动，同意提取4亿桶石油，以期稳定市场。Birol指出，在美以两国于2月底对伊朗发动袭击之前，全球石油市场存在严重的过剩，商业库存也非常高。但由于战争，局势迅速发生了变化。上周，国际能源署表示，由于伊朗冲突严重破坏中东石油生产，库存正以前所未有的速度消耗，今年全球石油供应将无法满足总需求，预计缺口为178万桶/日。而此前国际能源署预测今年石油供应将出现盈余。从欧佩克+组织给出的数据看，该组织预测，今年第二季度全球石油需求将达到1.045亿桶/日，较此前预期1.0507亿桶下降。而3月欧佩克石油产量为3490万桶/日，较2月减少了近800万桶/日，4月进一步下降至3320万桶/日。国际能源署补充称，预计今

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文GPT-5.5才发三周，5.6已在内测！OpenAI与Anthropic补贴大战同日开打，开发者坐收渔利发布于AI新闻资讯发布时间 :2026年5月14号 14:21阅读 :1分钟三周前GPT-5.5刚落地，GPT-5.6的消息已经提前炸出来了。知名爆料人Leo透露，GPT-5.6目前开发进度已全速推进，首批内部检查点在过去几天已启动测试，下个月很可能就会正式露面。更有意思的是，有人从OpenAI内部Codex日志里扒出了rollout mapping的痕迹——大部分调用还是指向GPT-5.5，但有一条记录赫然落在了GPT-5.6身上。换句话说，Codex环境里可能已经在偷偷用它跑测试了。内部代码名也被顺带曝光:ember-alpha和beacon-alpha。网友直呼:这迭代速度，根本追不上。速度再度拉满——Codex即将上线ultrafast模式GPT-5.6还没到，OpenAI产品端已经先按捺不住了。爆料人Chetaslua同步透露，本周四OpenAI将在Codex上线"ultrafast模式"，响应速度直接提升2到3倍，专门为延迟敏感型任务设计。与此同时，Image Arena排行榜上以+242分断档领先的gpt-image-2，也在同步推进A/B测试更新。这不是第一次了。今年3月GPT-5.4发布时，/fast模式已经实现了1.5倍加速;GPT-5.3-Codex-Spark借助Cerebras芯片，更是把推理速度拉到每秒超过1000token，是普通模式的15倍。而这次的ultrafast，直接在旗舰主力模型上实现提速——不是阉割版，不是小模型凑数，是正儿八经的全尺寸加速。对开发者来说，Agent循环、长任务流水线、浏览器自动化……所有需要"等"的地方，体验将彻底不同。顺带一提，上周ChatGPT官方账号发出一张科幻感十足的图，界面上那句经典的"Ask ChatGPT"悄悄换成了"Message ChatGPT"。奥特曼丢下一句"Call me maybe"，让外界对ChatGPT集成语音通话乃至推出硬件产品的猜测，彻底烧开了。Codex vs Claude Code，补贴战正式开打硅谷最戏剧性的一幕，就在今天上演了。就在OpenAI ultrafast模式上线前夕，Anthropic率先出手——宣布从6月15日起，为付费用户每月提升50%的编程额度，覆盖Claude Agent SDK、命令行工具以及深度接入GitHub工作流的Claude Code，并同步发布了Opus4.7Fast模式，主打比Codex更快的高级推理与更流畅的长上下文编码体验。OpenAI没让人等太久。反手祭出一招重磅补贴:接下来30天内，任何想从其他平台迁移到Codex的企业，直接获赠2个月免费使用权。按Pro计划每月200美元算，相当于白送400美元。奥特曼本人亲自下场喊话，直言Codex是"目前市面上最强的AI编程产品"。仅仅几个小时后，战报出炉:2000名开发者在3小时内主动联系了OpenAI。这场怼脸硬刚，看呆了整个硅谷。开发者们倒是乐了——被两家顶级公司抢着补贴，这种好事，上哪儿说理去。飞轮已经开始自转把这两件事放在一起看，会发现一个比任何单一新闻都更值得关注的趋势正在成形。AI正在加速自我进化。GPT-5.3-Codex是OpenAI首个"参与自身训练"的模型，到GPT-5.5，OpenAI内部85%的员工每周都在用Codex写代码。GPT-5.6几乎可以确定是在GPT-5.5的深度参与下开发出来的——AI在帮OpenAI造更强的AI。与此同时，编程工具的大规模普及正在释放前所未有的工程产能。Codex已有300万周活跃用户，Claude Code的用户数同样在爆发式增长。数百万开发者把AI编程工具当成日常标配，AI生成的代码又反哺回AI的训练与部署，这个循环只会越转越快。奥特曼说过，OpenAI的目标已经不止于AGI，直接瞄准了ASI。当模型迭代速度由AI自身驱动，当AI编程工具成为AI开发的基础设施，当两家万亿级公司用"补贴战"加速这一切的普及——通往ASI的飞轮，已经开始自转了。

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文李开复：开源模型是实现“AI主权”的更优路径发布于AI新闻资讯发布时间 :2026年5月12号 18:01阅读 :1分钟创新工场董事长李开复近日在专访中深入探讨了“AI主权”的概念。他指出，人工智能主权不仅关乎技术控制权和数据安全，还涉及模型与本土文化及法律的适配性。他认为，并非每个国家都有必要重造一个OpenAI，盲目追求昂贵的闭源自研并非良策。李开复强调，对于大多数资源有限的国家和企业而言，基于开源模型构建本地化体系是更现实的选择。这种“第三条路”通过在开源基础上进行增量训练，能以极低的成本让模型符合特定地区的价值观，而非受制于他国的通用模型。工程效率与生态之争谈及中美竞争，李开复认为资源受限反而催生了中国企业的极致工程效率。他以DeepSeek为例，指出中国公司在算力不足的情况下，通过优化架构和减少依赖，能以不到美国十分之一的投入实现极高的性能。他预测AI领域将演变成类似“iPhone与安卓”的格局。闭源模型如同一方领地，追求高利润和生态控制力；而开源模型则代表了更广阔的市场覆盖和普及度，两者将在全球范围内长期并存，共同构建未来的智能生态。中国硬件的下半场机遇李开复对中国在物理AI和硬件制造方面的结构性优势表示乐观。他指出，中国拥有全球最完整的供应链体系，从智能穿戴到机器人，制造成本远低于西方。未来AI时代的标杆级硬件，极有可能由中国设计并品牌化。此外，中国在语音驱动交互和大规模现实数据生成方面也具备独特文化优势。随着“人工智能+”政策对实体经济的赋能，硬件与AI的深度结合将成为中国在AI下半场竞赛中的核心竞争力。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文32 家初创抢 11% 市场！The Information 数据揭底：生成式 AI 的「虚假繁荣」正在加速赢家通吃发布于AI新闻资讯发布时间 :2026年5月19号 14:04阅读 :1分钟生成式 AI 的「繁荣表象」下，一场残酷的洗牌正在加速。The Information最新分析了一组扎眼的数据：在 34 家备受资本与舆论追捧的生成式 AI 初创公司中，有 32 家，正在为仅占市场总量 11% 的剩余份额激烈厮杀。换句话说，近九成的市场蛋糕，早已被头部玩家悄然瓜分。这不是「百花齐放」的春天，而是「赢家通吃」的预演。当技术门槛逐渐拉平、用户习惯快速固化、资本偏好日趋集中，后来者即便产品有亮点、团队有背景，也很难突破「马太效应」的引力场。流量向头部聚拢、人才向巨头流动、订单向标杆集中——这些信号叠加，让中小初创的生存窗口正在肉眼可见地收窄。更值得警惕的是，这种分化并非行业成熟后的自然结果，而是在技术迭代周期极短、商业化路径尚未跑通的早期阶段就已显现。这意味着，留给「差异化突围」的时间，可能比想象中更紧迫。对创业者而言，这或许是一个重新校准策略的信号：与其在通用大模型的红海里硬拼参数，不如深耕垂直场景、构建数据闭环、打磨真实复购；对投资者而言，则需更审慎地评估「故事」与「壁垒」之间的真实距离。当潮水退去，真正的考验才刚刚开始。

AI资讯AI新闻资讯正文投资者群炸锅！Anthropic 一纸声明戳破预上市幻想：你以为的早期门票，公司可能根本不认发布于AI新闻资讯发布时间 :2026年5月19号 14:29阅读 :1分钟Anthropic 最近给「想提前上车」的投资者泼了一盆冷水。这家大模型独角兽突然收紧股票场外交易限制，明确警告：未经董事会批准的股权转让，公司一概不认。更「硬核」的是，官方支持页面直接点名多家二级市场与私募平台，称其上流通的 Anthropic 股票或相关权益均属未经授权，相关交易可能无效，股东名册也不会更新。消息一出，投资者社群瞬间紧张。彭博社披露，一个由数百名家族办公室成员组成的 WhatsApp 群里，有人直接问出那句扎心的：「我们完蛋了吗？」这背后，其实是 AI 估值狂热下的一场「预期错配」。Anthropic 作为本轮大模型竞争中最受瞩目的公司之一，虽未上市，却已被市场提前贴上「必须配置」的标签。但问题在于：普通人根本买不到未上市公司的原始股。于是，各种「绕路方案」应运而生——特殊目的载体（SPV）、二级市场份额转让、远期合约、代持结构，甚至各类号称能「间接持有」热门 AI 公司权益的理财产品。这些交易满足的，是一种「怕错过」的焦虑：等它真正 IPO，可能已经高攀不起；现在不布局，未来就没机会。但站在公司视角，逻辑完全不同。尤其在估值快速攀升、潜在上市预期升温的阶段，Anthropic 这类企业更需要严格控制股东结构、转让路径与信息披露节奏。如果外部市场把各种「影子权益」炒得过热，不仅可能干扰公司治理，还可能带来合规风险与声誉反噬。所以这次重申规则，本质是一次「边界划清」：你以为买到的是 AI 独角兽的早期门票，但公司未必承认那张票的效力。这对投资人来说确实刺耳。因为很多 Pre-IPO 交易的吸引力，恰恰来自「提前锁定上市收益」的想象。一旦公司明确表态未经批准的权益无效，那些通过复杂结构买入的人，可能突然发现：自己持有的不是股票，而是一份解释空间极大、兑现路径模糊的合同。这也解释了为什么投资者群体会如此焦虑。AI 公司估值越高，越想提前进场的人越多；绕道进场的人越多，公司就越要收紧边界。结果就是：最热的资产，反而变成最不透明、最难确权的那一个。放到整个行业来看，这件事戳破了一个尴尬现实：市场已经把头部 AI 企业当成「下一代互联网」来定价，但这些公司的商业模式、盈利路径、成本结构，其实还没真正跑通。算力烧钱如流水，收入波动难预测，行业洗牌远未结束。在估值高企、数据不透明的背景下，很容易形成「人人都怕错过，但没人完全看清」的集体幻觉。更值得警惕的是，当 AI 独角兽的财富光环从一级市场溢出，开始向普通投资者招手时，信息不对称的鸿沟反而被拉大了。普通人往往承担了最高的风险，却得不到最基本的保障。而即便权益转让本身有效，押注的资产本身，也依然充满不确定性。AI 热潮改变的不仅是产品形态，也在重塑资产市场的规则。当所有人都想抢购「还没上市的 AI 未来」时，风险，其实已经跟着价格一起涨起来了。

AI资讯AI新闻资讯正文加量不加价!Anthropic 宣布 Claude Design 套餐上限翻倍发布于AI新闻资讯发布时间 :2026年5月19号 14:56阅读 :1分钟Anthropic 今日在 X 平台发布官方公告，宣布对其旗下生产力工具生态进行重大升级:不仅将视觉创作工具 Claude Design 在所有订阅套餐中的 Token 上限提升一倍，同时将开发辅助工具 Claude Code 的“快速模式（Fast Mode）”默认切换至其最新一代主力模型 Opus4.7。此次更新直击专业用户的痛点。作为 Anthropic 于今年4月推出的重磅视觉工具，Claude Design 凭借对话式生成设计、原型构建及多源导入等功能备受瞩目，但此前受制于上下文额度，复杂项目常面临中断困境;本次针对 Pro、Max、Team 及 Enterprise 全套套餐的 Token 上限翻倍，将彻底解放长上下文的连续迭代能力。与此同时，面向开发场景的 Claude Code 快速模式在接入 Opus4.7后，其交互响应速度最高可提升至原有的2.5倍，能更高效地支撑代码快速修改、实时调试等高时效性任务。在计费与资源管理上，Anthropic 保持了高度的商业克制与透明度。Opus4.7与此前 Opus4.6在快速模式下的定价完全一致，均维持在每百万输入 Token30美元、输出150美元，且全量覆盖100万 Token 的上下文窗口，费率不随上下文长度而变动。此外，两款模型在快速模式下共享相同的速率限制池，企业与开发者无需担心额外的配置成本。

AI资讯AI新闻资讯正文AI眼镜进入城市服务场景，蚂蚁GPASS上线城市伴游功能发布于AI新闻资讯发布时间 :2026年5月19号 14:58阅读 :1分钟继乌镇峰会期间“桐小乌”在乐奇AI眼镜上的应用落地后，蚂蚁自研的智能终端可信连接技术框架GPASS正进一步拓展应用边界。近期，基于该技术的城市伴游功能已在乐奇AI眼镜新版本中正式上线。用户开启伴游模式后，可通过乐奇AI眼镜实现边行走、边获取信息的实时交互体验。系统可根据用户所处位置，对周边景点进行讲解，并支持即时问答、导航、第一视角内容记录与分享，在减少手动操作的同时提升出游的连贯性与沉浸感。作为首个接入该服务的城市文旅官方智能体，“杭小忆”的入驻标志着“镜游杭州”项目进入实际运行阶段。该项目由杭州文旅、支付宝与乐奇联合打造，旨在探索AI与城市文旅服务的结合方式。后续，城市伴游功能还将逐步接入更多城市专用文旅智能体。从功能体验来看，本次升级主要体现在三个方面:一是提升信息获取的深度。在传统导航与导览基础上，系统可结合用户附近的建筑、景观等所在环境，补充建筑、人文历史知识及文化相关内容，帮助用户更系统地理解所见所闻。二是增强多功能协同能力。对话、拍摄、翻译、导航、支付及音频播放等功能可通过语音进行切换与调用，即使处于伴游模式下，也不影响其他功能的使用，整体体验更加连贯。三是优化交互方式。基于语音、图像与位置信息的综合识别能力，系统可在特定场景下主动提供提示与讲解，无需唤醒词即可开启对话，过滤环境噪音和无关意图。从以往的被动响应，转向更加连续的交互体验。这一创新体验的背后，是蚂蚁GPASS智能终端可信连接技术框架的强力支撑。GPASS通过语音+图像+位置的多模态意图理解，精准捕捉用户需求;在眼镜端实现流畅的对话与导航交互;并确保了与“杭小忆”等城市文旅智能体之间的可信连接，让服务既智能又安全。业内人士认为，城市伴游功能的上线，显示出AI眼镜正从单一设备形态向具体应用场景延伸。以文旅为代表的服务领域，有望成为相关技术规模化落地的重要方向之一。

AI资讯AI新闻资讯正文Alexa+ 上线按需播客功能 想听啥自己点！发布于AI新闻资讯发布时间 :2026年5月19号 15:14阅读 :1分钟亚马逊的 Alexa+ 又开始「卷」新技能了。本周一，Amazon 正式宣布 Alexa+ 迎来重磅更新:支持按需生成播客节目。这项名为「Alexa Podcasts」的新功能，即日起面向美国用户开放测试。用亚马逊的话说，它的目标很直接——「把你感兴趣的任何话题，几分钟内变成一期专属播客」。使用门槛低到几乎没有:用户只需对 Alexa+ 说一句「帮我做个关于 XX 的播客」，不用上传资料、不用写脚本、更不用提前规划。AI 会自动研究主题、搜集信息、生成内容大纲，并快速反馈这期节目大概会讲什么。用户还能随时调整时长、语调、内容侧重，像搭积木一样定制属于自己的音频内容。确认无误后，Alexa+ 会用 AI 生成的主播声音完成播报。节目制作完成后，用户会通过 Echo Show 设备或 Alexa App 收到通知;所有生成的播客也会自动保存在 App 的「Music」和「More」栏目中，方便随时回听、反复回味。这不只是「多了一个功能」那么简单。它标志着 Alexa+ 正在从「回答问题、控制家电」的传统语音助手，悄然进化为「懂你兴趣、能创内容」的个性化 AI 创作伙伴。当语音交互遇上内容生成，智能音箱的边界，正在被重新定义。当然，新能力也伴随新争议。AI 主播的声音是否足够自然?自动生成内容的准确性如何保障?会不会对传统播客创作者形成冲击?这些都是行业正在热议的问题。尤其当话题涉及新闻事件或专业领域时，用户对信息可靠性的要求会更高。对此，亚马逊也提前「打补丁」:强调已与多家权威新闻机构建立合作，包括美联社、路透社、《华盛顿邮报》、《时代周刊》、《福布斯》、Politico、《今日美国》等主流媒体，以及全美200多家地方报纸。通过这些合作，Alexa+ 能接入实时、可信的信源，从底层提升内容质量与事实准确度。更值得留意的是，播客只是起点。亚马逊透露，团队正在探索更多个性化 AI 音频形态:比如基于用户偏好定制的每日新闻简报，或结合个人文档、共享信息生成的专属内容。当「你的数据+AI 理解+场景需求」真正打通，语音交互的体验，或许会从「被动响应」迈向「主动陪伴」。

AI资讯AI新闻资讯正文加量不加价！Anthropic旗下Claude两款主力AI工具迎来重磅升级发布于AI新闻资讯发布时间 :2026年5月19号 15:18阅读 :1分钟大模型独角兽Anthropic近日在社交平台宣布，对其现有的AI工具链进行大刀阔斧的性能升级。本次调整的核心聚焦于广受好评的视觉创作工具以及专门面向程序员的代码利器，旨在通过大幅度放宽限制来提升用户的实际创作效率。在此次更新中，上个月刚推出的视觉创作工具Claude Design迎来了额度暴涨，其在Pro、Max、Team以及Enterprise在内的全部订阅套餐中，Token使用上限直接实现翻倍。这一举措彻底解决了创作者在处理复杂设计项目时，因额度限制而不得不将任务强行分段的痛点。代码编辑器默认切换新模型除了视觉工具的额度翻倍，针对软件开发者的命令行代码工具Claude Code也迎来了重要调整。系统内置的快速模式（Fast Mode）现在已将底层大模型默认切换为更先进的Opus 4.7。开发人员只需在终端中简单地输入“/fast”指令，即可一键激活该高速配置。官方测试数据显示，在该模式下，新模型的运行响应速度可以直接飙升至此前的2. 5 倍。高通吐量主打交互式高频调优由于具备极高的吞吐效率，该快速模式主要服务于高频、即时性的交互式工作。无论是快速迭代代码修改、实时捕获并调试系统漏洞，还是处理时间要求极度紧张的突发开发任务，它都能提供有力的算力支持。在费用收取方面，Opus 4. 7 在快速模式下的定价与此前的4. 6 版本保持完全一致。其每百万输入Tokens的收费标准为 30 美元，输出则为 150 美元，且全面覆盖 100 万Tokens的超长上下文窗口。

吐司APP的简介吐司是腾讯推出的AI原生应用创作平台，用户只需输入想法，AI自动生成完整的安卓原生应用。平台支持智能功能梳理、精细化编辑打磨，能一键导出APK进行本地安装与分享。内置灵感广场汇聚海量优质模板，涵盖社交助手、照片精选、学习工具等多元场景，支持用户共创。无需编程基础，零基础用户也能快速将奇思妙想转化为可实际使用的专属APP，轻松实现个性化应用开发。吐司APP的主要功能AI自动生成应用：输入自然语言想法，AI自动梳理功能逻辑并生成完整的安卓原生应用，无需编写代码。精细化编辑打磨：支持对生成的应用进行细节调整与优化，满足个性化需求。一键导出APK：生成完成后可直接导出APK安装包，支持本地安装、测试与社交分享。灵感广场：内置海量精选应用模板与创意案例，如社交助手、照片精选、学习工具等，帮助用户快速获取创作灵感。应用共创：支持用户基于现有模板进行二次创作与共创，降低应用开发门槛。多场景覆盖：支持生成考研题库、运势工具、普通话测试、宠物健康日志、梦境记录、食材新鲜度检测、万年历等多样化实用应用。类似于吐司APP的应用Qoder阿里巴巴推出的 AI Agent 编程平台CursorAnysphere公司推出的AI代码编辑器Trae字节跳动推出的免费 AI IDEYouWare APPAI移动应用开发工具Replit APPAI开发工具应用WindsurfCodeium公司推出的AI编程工具暂无评论再想想发表评论暂无评论…

LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台AI工具1天前发布AI小集03LibTV团队版是什么LibTV 团队版是LibTV推出的专业级 AI 视频协作工作台，专为 AI 短剧、电影、TVC 制作团队设计。平台提供团队空间、实时画布协作、资产库沉淀、权限管理、成本管控等功能，解决 AI 视频创作中资产散落、协作低效、人员流动导致资产流失等痛点，支持规模化、工业化的 AI 内容生产。LibTV团队版的主要功能专属团队空间：统一沉淀分镜、脚本、镜头资产与生成记录，历史项目可迁移，避免核心资产散落在个人账号与本地文件中。画布协作：导演、美术、剪辑等角色可在同一块画布实时协同，同步节点状态与镜头修改，减少版本错乱。团队资产库：角色主体、风格模板、运镜方案、提示词工作流与参考素材支持统一沉淀，成熟视觉体系可持续复用。权限管理：支持项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作场景。成本管理：共享积分池、成员额度分配与消耗查看，团队管理者可统一管理生成资源、席位与项目成本。资产交接：成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免创作资产随人员流动中断。如何使用LibTV团队版创建团队空间：访问LibTV官网升级至团队版，建立专属团队项目空间，设置团队名称与基础信息。配置成员与权限：邀请导演、编剧、美术、剪辑等成员加入，按岗位分配访问、编辑、复制或发布权限。沉淀与迁移资产：将历史个人项目迁移至团队空间，或新建项目并导入参考素材、角色设定与风格模板。画布协作创作：成员在统一画布内协同推进分镜设计、镜头生成与修改，实时同步进度与反馈。复用团队资产库：调用已沉淀的角色主体、风格模板与提示词工作流，保持系列化内容风格一致。管控成本与交接：通过共享积分池分配生成额度，监控消耗；成员变动时资产自动保留，确保项目连续性。LibTV团队版的核心优势工业化协作：从个人创作升级为团队制片管理，支持长期项目、多岗位协同与规模化内容生产。资产持续复用：成熟的角色设定、风格与提示词工作流可沉淀为团队资产，降低新人上手与重复调试成本。实时协同提效：多人画布协作减少版本错乱与重复生成，围绕同一套镜头语言与素材体系推进项目。成本与风险可控：共享积分池与席位管理适配多项目并行需求，权限分级与资产交接机制保障核心创作安全。LibTV团队版的同类竞品对比对比维度LibTV 团队版快手可灵（企业版/团队功能）协作模式多人实时画布协作，导演/美术/剪辑等角色可在同一块画布内同步节点状态、生成进度与镜头修改团队空间内创作接力，支持桌面端（Win/Mac）同步，减少文件下载与搬运损耗资产沉淀角色主体、风格模板、运镜方案、提示词工作流、参考素材与工具箱统一沉淀，支持成熟视觉体系持续复用设立「共享资产库」，支持个人空间向团队空间一键转移灵感值、创作资产与主体资产，实现”一人贡献，全员复用”权限管理项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作所有者、管理员、协作者三重角色，灵活配置查看/编辑/移除权限，保障核心资产安全成本管理共享积分池、成员额度分配与积分消耗查看，支持多项目并行下的生成资源统筹支持灵感值跨空间转移，团队共享生成资源，但精细化成本管控功能相对较弱资产交接成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免资产随人员流动中断个人资产可一键导入团队空间，但成员退出后的自动资产交接机制未明确强调核心场景聚焦 AI 短剧、AI 电影、AI TVC 的工业化制片流程，强调长期项目管理与跨岗位协同覆盖 AI 视频生成、智能分镜、广告 Demo、影视特效等，强调快速创作与素材复用产品定位AI 视频专业级协作工作台，面向制片管理流程AI 视频生成平台的团队功能扩展，面向创作者联盟与工作室LibTV团队版的应用场景AI 视频工作室：适合长期运营 AI 短剧、动画、MV 的创作团队，通过资产库统一沉淀角色与风格，实现系列化生产。短剧承制方与 MCN：适配高频、多项目并行的工业化内容生产，导演与美术可同时在线推进分镜与镜头生成。品牌市场部与代理公司：支持品牌方、代理公司与外部创意团队跨组织协作，共同管理 AI TVC 与 Campaign 项目。AI 电影与大型制作团队：支持长篇叙事、系列化 IP 与复杂制作流程，导演组、制片组、美术组与后期团队跨岗位协同。企业内容与创意部门：适配培训、营销、产品与全球化传播等标准化内容生产，便于企业进行长期生成资源规划。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇PPT Master - 开源 AI PPT 生成标准化工作流下一篇腾讯Marvis - 腾讯应用宝推出的操作系统层个人 AI 助手相关文章Le Chat APP – Mistral AI推出的AI对话助手应用AI小集4DeepSWE – Together.ai联合Agentica开源的AI Agent框架AI小集3Indic Parler-TTS – 开源多语言TTS模型，专注于合成印度语和英语AI小集2Rabbithole – AI互动式知识探索工具，无尽知识树问答AI小集3OpenJarvis – 斯坦福大学开源的本地 AI 智能体框架AI小集3SimpleAR – 复旦大学联合字节 Seed 团队推出的图像生成模型AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付38分钟前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版40分钟前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架41分钟前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准3小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型4小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

Grok Build – xAI 推出的终端原生 AI 编程智能体AI工具2天前发布AI小集03Grok Build是什么Grok Build 是 xAI 推出的终端原生 AI 编程智能体（Agentic CLI），面向专业软件工程与复杂编码任务。基于 Grok 4.3 beta 的 200 万 token 超长上下文，支持 Plan Mode 计划审批、最多 8 个并行子智能体协作及 Arena Mode 自动方案排序，可在本地直接读写文件、执行 Shell 命令、操作 Git 并构建完整应用。采用本地优先架构，源代码与敏感数据不上传云端，保障企业级安全。Grok Build的主要功能Plan Mode（计划模式）：复杂任务执行前自动生成结构化步骤计划，用户可审批、评论或修改，所有代码变更以 diff 形式透明展示，避免”黑箱”操作。并行子智能体：最多支持8 个并发 AI 智能体同时工作，可分别处理代码分析、CI/CD 检查、性能优化、文档检索等不同子任务。Arena Mode（竞技场模式）：多个智能体针对同一任务生成不同代码方案，系统自动评估并排序，开发者直接选择最优解，无需手动对比。超长上下文窗口：基于 Grok 4.3 beta，支持200 万 token上下文，可一次性加载整个大型代码库进行跨文件复杂任务。本地优先架构：源代码、凭证和敏感数据不上传云端，完全在本地执行，支持离线环境和气隙网络，保障企业级代码安全。终端原生 CLI + Web 双入口：支持本地命令行交互，也可通过浏览器界面远程操作，满足不同工作流习惯。VS Code 扩展集成：提供官方 VS Code 插件，在 IDE 内直接调用 Grok Build 能力，实现 GUI adjacent 工作流。Git 与开发服务器集成：可自动管理 Git 仓库、启动本地开发服务器、内置浏览器浏览网页、执行 Shell 命令。插件与开放生态：支持 MCP 服务器、AGENTS.md 团队规范文件、Hooks 自定义动作、ACP（Agent Client Protocol）协议，便于集成现有 DevOps 流水线。Headless 脚本模式：通过-p标志支持非交互式执行，可直接嵌入 CI/CD 流水线或自动化脚本。BYOK（自带 API Key）：支持使用自有 xAI API Key 进行本地计费，灵活控制成本。如何使用Grok Build一键安装 CLI：在终端执行curl -fsSL https://x.ai/cli/install.sh | bash完成本地安装，支持 macOS、Linux 和 Windows 三大平台。启动交互式终端：安装完成后输入grok>进入交互式命令行界面，可直接用自然语言描述编程需求。登录账号：在终端内执行grok> /login登录 xAI 账号，解锁高级功能与订阅权益。配置 API Key（可选）：如需使用自有额度，执行grok> /api-key add xai添加个人 xAI API Key（BYOK 模式）。使用 Plan Mode 处理复杂任务：输入需求后，Grok Build 会自动生成结构化执行计划，展示步骤与预期变更 diff，用户可审批、评论或修改后再执行。利用并行子智能体：在复杂项目中，系统会自动或手动触发最多 8 个并发子智能体，分别处理代码分析、测试生成、文档检索等子任务。通过 Arena Mode 选优：对同一任务让多个智能体生成不同方案，系统自动评估排序，你只需选择排名最高的方案应用。Web 端远程操作：访问build.grok.com或x.ai/cli，在浏览器中打开 Grok Build 界面，无需本地安装即可使用。VS Code 集成工作流：安装 Grok Build 的 VS Code 扩展，在编辑器内直接调用智能体能力，边写代码边让 AI 协助修改。Headless 自动化执行：在脚本或 CI/CD 流水线中使用-p标志运行非交互式命令，实现自动化构建与部署。查看帮助与命令：在终端输入grok> /help查看所有内置命令与快捷操作指南。Grok Build的官网地址项目官网：https://x.ai/news/grok-build-cliGrok Build的核心优势本地优先，代码绝不上云：源代码、凭证及敏感数据完全在本地执行，无需上传至 xAI 服务器，适合处理专有代码库、金融及政务等受监管行业项目。200 万 token 超长上下文：基于 Grok 4.3 beta，可一次性将整个大型代码库载入内存，跨数十个文件进行复杂重构与多文件联动修改，大幅减少上下文丢失。8 路并行子智能体：最多同时调度 8 个并发 AI 智能体分工协作，分别处理代码分析、测试生成、CI/CD 检查等子任务，显著缩短复杂项目耗时。Arena Mode 自动选优：同一任务由多个智能体生成不同方案，系统自动评估排序并给出排名，开发者直接选择最优解，告别手动对比试错。Plan Mode 透明可控：执行前先输出结构化步骤计划与 diff 预览，用户可审批、评论或修改后再执行，每一步变更清晰可见，避免”黑箱”风险。极致 API 价格竞争力：输入 token 仅$0.20 / 百万 tokens，相较 Claude Code 和 Codex CLI 大幅降低高频使用成本。终端原生 + 多入口覆盖：支持 CLI 命令行、Web 浏览器、VS Code 扩展三种工作入口，既满足极客式终端操作，也兼容 IDE 集成习惯。Grok Build的同类竞品对比对比维度Grok Build (xAI)Claude Code(Anthropic)Codex CLI(OpenAI)产品定位终端原生 AI 编程智能体（Agentic CLI）终端 AI 编程助手云端 AI 编程智能体发布状态2026.5 早期 Beta已正式发布已正式发布底层模型Grok 4.3 beta / grok-code-fast-1Claude 4 (Sonnet/Opus)Codex (基于 GPT-4o/o3)上下文长度200 万 token20 万 token (Claude 4)约 20 万 token并行智能体最多 8 个并发子智能体单智能体为主单智能体为主方案评估Arena Mode 自动排序选优手动对比或迭代手动对比或迭代执行计划Plan Mode（先计划后审批执行）直接执行，可审查 diff直接执行，可审查 diff数据安全本地优先，代码不上云云端执行，代码上传 Anthropic云端执行，代码上传 OpenAI离线支持支持完全离线/气隙网络需联网需联网SWE-Bench70.8%约 72–75%约 70%API 定价$0.20 / 百万 tokensPro 订阅 + 按量计费Credits 按量计费订阅门槛SuperGrok Heavy ($300/月)Claude Pro / Team 订阅ChatGPT Pro / 开发者额度入口形态CLI + Web + VS Code 扩展CLI + IDE 集成CLI + ChatGPT 移动端自动化集成Headless 模式 (-p 标志) + CI/CD有限脚本支持有限脚本支持核心差异多智能体并行 + 本地优先 + 超大上下文模型推理能力强，生态成熟与 ChatGPT 生态深度整合Grok Build的应用场景大型代码库重构与现代化：基于 200 万 token 超长上下文一次性加载整个项目，跨数十个文件进行安全重构、依赖升级或架构迁移，避免上下文截断导致的遗漏。复杂全栈应用从零构建：通过 Plan Mode 自动生成项目结构、技术选型、依赖安装和文件生成计划，审批后直接执行，快速搭建完整应用原型。敏感行业专有代码开发：本地优先架构确保源代码、凭证和敏感数据不上传云端，适用于金融、政务、医疗、军工等对数据主权有严格要求的场景。离线/气隙网络环境编程：无需持续联网即可在本地执行全部任务，适配高安全隔离网络、内网开发机或无外网访问权限的工作环境。多维度自动化代码审查：启动多个并行子智能体分别检查代码安全性、性能瓶颈、CI/CD 合规性和文档完整性，一次性输出综合审查报告。技术方案快速对比与选型：使用 Arena Mode 针对同一需求生成多种实现方案，系统自动评估排序，帮助团队在短时间内确定最优技术路径。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇OpenHuman - 开源桌面端 AI 助手，能主动感知工作上下文下一篇Violin - 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具相关文章HuMo – 清华联合字节推出的多模态视频生成框架AI小集3ACE-Step – ACE Studio联合阶跃星辰开源的音乐生成基础模型AI小集3ComicsMaker – 在线AI漫画创作平台，支持多功能创作多风格高清画质AI小集2YOLOv9 – 新一代高效的实时目标检测系统AI小集2SANA 1.5 – 英伟联合MIT、清北等机构推出的文生图新框架AI小集3tldraw – AI白板绘图工具，开源免费的无限画布和多种绘画功能AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付38分钟前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版40分钟前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架41分钟前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准3小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型4小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文AI工具3天前发布AI小集05OpenHuman是什么OpenHuman 是 tinyhumansai 团队推出的开源个人 AI 超级智能助手，定位为”Your Personal AI super intelligence”，主打私有、简单且极其强大。非传统聊天机器人，是能主动感知用户工作上下文的桌面级 AI Agent，通过每 20 分钟自动同步 118+ 第三方应用数据，构建本地持久记忆，让 AI 真正”懂你”。项目采用 Rust + TypeScript + Tauri 构建，支持 macOS、Linux 和 Windows 平台。OpenHuman的主要功能桌面吉祥物交互：具备视觉形象的桌面 AI 助手，支持语音输入（STT）与 ElevenLabs 语音输出（TTS），吉祥物可实时唇形同步，能作为参会者加入 Google Meet 并实时记录会议。118+ 第三方应用一键集成：通过 OAuth 快速接入 Gmail、Notion、GitHub、Slack、Linear、Jira、Stripe、Calendar、Drive 等主流工具，无需编写任何连接代码。每 20 分钟自动同步（Auto-fetch）：核心引擎每 20 分钟遍历所有活跃连接，自动拉取最新数据进入本地记忆系统，让 Agent 在次日早上已掌握前日全部上下文。Memory Tree 层次化持久记忆：将同步数据规范化为 ≤3k-token 的 Markdown chunks，评分后折叠存入层级摘要树，存储于本地 SQLite，实现跨会话的长期记忆。Obsidian Vault 原生兼容：同步数据同时输出为 Obsidian 可读的.md文件，用户可在 Obsidian Graph View 中浏览和编辑 AI 维护的知识网络。TokenJuice 智能 Token 压缩：在数据进入 LLM 前自动压缩（HTML 转 Markdown、缩短长 URL、移除非 ASCII 字符等），最高可降低 80% 的 API 调用成本与延迟。内置模型路由（Model Routing）：根据任务类型自动分配轻量模型、强推理模型或视觉模型，一个订阅覆盖多种模型能力，无需管理多个 API Key。完整工具链开箱即用：内置 Web 搜索、网页抓取（Web Scraper）、代码工具集（文件系统、Git、Lint、Test、Grep），无需额外安装插件即可使用。本地优先与隐私保护：所有工作流数据本地加密存储，支持通过 Ollama 接入本地模型实现完全离线运行，数据不上传云端。可选 agentmemory 共享后端：可配置与 Claude Code、Cursor、Codex 等工具共享同一个持久记忆存储，实现跨工具的知识互通。消息通道集成：支持 inbound/outbound 消息收发，打通用户已有的沟通渠道，工作流数据始终保存在本地设备。如何使用OpenHuman官网下载安装：访问OpenHuman的官网 https://tinyhumans.ai/openhuman 下载对应平台的 DMG（macOS）或 EXE（Windows）安装包，双击安装后即可启动。命令行快速安装（macOS / Linux x64）：在终端执行curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash自动完成下载与安装。命令行快速安装（Windows）：在 PowerShell 执行irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex完成一键安装。一键连接工作应用：打开应用后，通过 OAuth 一键接入 Gmail、Notion、GitHub、Slack、Calendar、Drive 等 118+ 第三方工具，授权后每 20 分钟自动同步数据。与桌面吉祥物交互：点击桌面上的吉祥物形象，通过语音或文字与其对话，可勾选 “Speak replies” 让 AI 语音播报回复。查看自动同步的记忆数据：在应用内查看 Agent 已同步的邮件、任务、代码仓库等上下文，AI 会自动压缩整理为 Memory Tree 供后续对话调用。Obsidian 知识库浏览：同步数据自动写入本地 Obsidian-compatible Vault，打开 Obsidian 即可在 Graph View 中可视化浏览 AI 维护的知识网络，也可手动编辑.md文件。使用内置工具集：直接调用开箱即用的 Web 搜索、网页抓取、代码工具（文件系统、Git、Lint、Test、Grep）等能力，无需额外安装插件。接入本地模型（可选）：如需完全离线运行，通过 Ollama 在本地部署模型，并在设置中切换至本地模型后端，实现零云端依赖。OpenHuman的项目地址官网地址：https://tinyhumans.ai/openhumanGitHub 仓库：https://github.com/tinyhumansai/openhumanOpenHuman的核心优势UI 优先分钟级上手：干净的桌面体验，无需终端配置，从安装到可用 Agent 仅需几分钟，区别于终端优先的竞品。上下文在分钟级而非周级：传统 Agent 需要数周学习用户习惯，OpenHuman 通过一次同步即可掌握邮箱、日历、仓库、文档、消息的全量压缩上下文。一站式订阅，零 API 蔓延：一个订阅覆盖所有功能，无需额外购买多个模型 API Key。Obsidian 知识库原生兼容：同步数据自动写入 Obsidian-compatible Vault，用户可在熟悉笔记界面中浏览和编辑 AI 维护的知识图谱。Rust + Tauri 高性能架构：内存占用远低于 Electron 方案，69.7% Rust + 26.1% TypeScript 构建。开源可审计：GPL-3.0 协议，代码完全透明，隐私敏感用户可自行审计数据流向。OpenHuman的同类竞品对比对比维度OpenHumanOpenClawHermes Agent开源协议✅ GPL-3.0✅ MIT✅ MIT上手难度✅ UI 优先，分钟级上手⚠️ 终端优先配置⚠️ 终端优先配置使用成本✅ 单一订阅 + TokenJuice 压缩⚠️ 自备模型 API⚠️ 自备模型 API记忆能力🚀 Memory Tree + Obsidian Vault + 可选 agentmemory 后端⚠️ 依赖插件扩展✅ 自学习观察第三方集成🚀 118+ 应用 OAuth 一键接入⚠️ 需自行配置连接⚠️ 需自行配置连接自动数据同步✅ 每 20 分钟自动 fetch 至本地记忆🚫 无自动同步🚫 无自动同步API 管理✅ 一个账户覆盖所有模型🚫 自备 Key🚫 多供应商管理模型路由✅ 内置自动分配（轻量/推理/视觉）⚠️ 手动切换⚠️ 手动切换原生工具集✅ 代码 + 搜索 + 抓取 + 语音✅ 代码工具✅ 代码工具数据隐私🚀 本地优先，数据不上云依赖自托管配置依赖自托管配置OpenHuman的应用场景跨应用项目上下文整合：同时连接 GitHub、Linear、Jira、Slack 后，让 AI 自动对比代码 PR、任务看板和团队消息，一键汇总项目最新进展，无需手动切换多个工具。智能邮件筛选与摘要：授权 Gmail 后，AI 每 20 分钟自动扫描收件箱，按项目或优先级分类摘要，快速定位需要回复的重要邮件，过滤掉低价值信息。AI 会议助手：让桌面吉祥物以真实参会者身份加入 Google Meet，实时记录会议内容、提取 action items，并基于记忆库提供参会者背景信息辅助沟通。代码仓库智能问答：同步 GitHub 仓库数据后，直接向 AI 询问代码逻辑、历史变更原因、PR 评论争议点，Agent 基于完整代码上下文给出精准回答。Obsidian 个人知识库构建：将邮件、文档、聊天记录自动压缩归档为 Obsidian-compatible 的.md文件，在 Graph View 中可视化呈现知识关联网络，打造 Karpathy 风格的 LLM 知识库。多工具工作流自动化：连接 Notion、Calendar、Drive、Stripe 等工具后，AI 自动关联”日历事件→相关文档→对应邮件→财务记录”，形成完整的项目上下文链。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Intern-S2-Preview - 上海 AI Lab 开源的科学多模态大模型下一篇Grok Build - xAI 推出的终端原生 AI 编程智能体相关文章360安全龙虾 – 360推出的AI Agent桌面应用AI小集2Qihoo-T2X – 360 AI 研究院和中山大学开源的高效多模态生成模型AI小集2SHMT – 阿里达摩院联合武汉理工等机构推出的自监督化妆转移技术AI小集2MiniCPM-V 4.6 – OpenBMB 开源的端侧多模态大模型AI小集4GRUtopia 2.0 – 上海 AI Lab 推出的通用具身智能仿真平台AI小集2豆包浏览器插件 – 字节跳动推出的 AI 浏览器智能助手AI小集5暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付35分钟前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版37分钟前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架38分钟前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准3小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型4小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手AI工具18小时前更新AI小集03腾讯Marvis是什么腾讯Marvis是腾讯应用宝团队推出的操作系统层个人 AI 助手。工具通过 1 个主 Agent 与 5 个副 Agent 的协作架构，实现电脑系统操作、文件管理、App 操控、网页交互与信息搜索等复杂任务，24 小时在线并支持 Windows 与安卓跨端操作。产品具备本地知识库索引与可视化虚拟办公室交互，定位大众友好的日常 AI 数字秘书。腾讯Marvis的主要功能多 Agent 协作调度：1 个 PM 主 Agent 负责理解需求、拆解任务并分发给 5 个副 Agent，各 Agent 完成后向主 Agent 汇报，实现复杂任务的自动编排与闭环执行。深度系统操作：可直接调用 Windows 底层接口完成主题设置、电源计划调整、自启动管理、硬件配置检测、网络测速等原生系统级操作，无需模拟点击。智能文件管理：支持基于文件名、内容、图片 OCR、人脸、主题、地点等多维度本地文件搜索，可自动分类归档、格式转换、合并去重及跨端文件传递。跨端 App 操控：通过内嵌应用宝引擎，可在 PC 端直接操控安卓手机 App（如微博签到、同花顺、飞常准等），实现电脑与手机的应用级跨端调度。本地知识库：自动识别并索引本地文档与图片，支持标题匹配与内容匹配搜索，构建越用越懂用户的个人知识管理体系。自动任务与情报监控：支持定时执行既定任务，如每日新闻汇总、限时福利监控、明星动态追踪、房价监控等自动化工作流。如何使用腾讯Marvis下载安装：访问腾讯Marvis官网 https://marvis.qq.com/ 下载 Windows 客户端，按指引完成安装。登录激活：填写调查问卷获取邀请码，获得邀请码后，输入邀请码完成内测资格激活。模式选择：根据任务敏感度选择效率模式或隐私模式。指派任务：在对话框输入自然语言指令，或进入”自动任务”页面设置定时执行规则。查看进度：在右侧虚拟办公室面板查看各 Agent 工作状态与任务消耗 Token 数，等待主 Agent 汇总结果。确认执行：涉及系统修改、文件删除等敏感操作时，点击硬确认弹窗授权后完成最终执行。腾讯Marvis的核心优势操作系统层深度整合：区别于纯云端 Agent，Marvis 深入 Windows 系统底层，可直接调用系统 API 与文件系统，实现真正的本地原生操作。端云双模式灵活切换：效率模式采用云端大模型保障复杂任务性能，隐私模式基于 Qwen 端侧模型实现完全离线运行，满足财务、法务、HR 等敏感场景需求。应用宝跨端生态壁垒：继承 14 年应用商店经验与跨端引擎技术，独家支持 PC 端操控手机 App，生态调度能力为竞品不具备。可视化虚拟办公室交互：将每个 Agent 设计为戴企鹅围脖的黑色小马形象，在虚拟工位中实时展示工作状态（搬砖、打盹、喝咖啡），降低技术门槛并增强用户情感连接。L2 级安全兜底机制：涉及资金交易、系统核心配置修改、批量删除等敏感操作强制”硬确认”，支付功能完全禁用，从架构层面规避 AI 越权风险。腾讯Marvis的同类竞品对比对比维度腾讯Marvis（腾讯）Claude Computer Use（Anthropic）产品定位操作系统层个人 AI 助手云端计算机操控 Agent运行模式端云双模式（本地隐私 / 云端效率）纯云端运行系统深度深入 Windows 底层 API，原生系统级操作基于截图与模拟点击操控跨端能力PC 操控手机 App（应用宝引擎）不支持Agent 架构1 主 + 5 副多 Agent 协作，可视化办公室单 Agent 模式隐私安全L2 级硬确认，支持完全离线运行数据上云，依赖平台安全策略用户门槛自然语言输入，大众友好，开箱即用需一定技术理解，偏极客向生态合作与微软、Intel 系统级深度合作无特殊硬件合作腾讯Marvis的应用场景办公效率提升：自动整理桌面文件并按项目分类归档，一键完成合同信息审查、运营数据分析、参考文献整理及出差材料准备。系统运维助手：一句话调整系统设置（夜间模式、电源计划、默认浏览器），检测游戏配置兼容性并给出硬件升级建议，彻底卸载顽固软件及残留痕迹。生活娱乐管家：自动完成微博超话签到、监控游戏限时福利与明星动态、定时推送每日新闻汇总、整理宝宝成长相册与旅行攻略。敏感数据处理：在完全断网的隐私模式下，本地完成财务报表校对、员工薪资分析、法务合同审查等涉密工作，确保数据不出设备。跨端文件协作：在手机端发送指令，让 Marvis 在电脑中搜索指定文件并回传至手机，实现跨设备的无缝文件调度。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇LibTV团队版 - LibTV 推出的专业级 AI 视频协作工作台下一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型相关文章Steve AI – AI视频生成的平台，支持多模态数据转化为多种视频风格AI小集2LongCat-Flash-Omni – 美团开源的实时交互全模态大模型AI小集213个免费AI小红书文案生成工具，种草文案创作好帮手AI小集9NocoBase – 开源AI无代码开发平台，通过配置完成应用开发AI小集3FineVideo – Hugging Face推出的大型多模态视频数据集AI小集3新Composer 2.5 – Cursor 推出的自研 Agentic 编程模型AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付35分钟前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版37分钟前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架38分钟前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准3小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型4小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

Kimi K2.5 – 月之暗面kimi开源的全新一代全能旗舰模型AI工具4个月前发布AI小集03Kimi K2.5是什么Kimi K2.5 是月之暗面开源的全新一代全能旗舰模型，模型基于约 15T 视觉与文本数据训练而成，具备顶尖的代码生成、视觉理解能力，支持自主 Agent 集群协作，可调度多达 100 个子 Agent 并行处理复杂任务，执行效率大幅提升。Kimi K2.5 提供快速、思考、Agent 及 Agent 集群四种模式，适用多种场景，包括前端开发、办公自动化和复杂问题解决。模型可通过Kimi.com、Kimi App、Kimi 开放平台和 Kimi Code （Kimi 新推出的编程工具）等平台提供服务，助力用户高效完成任务。Kimi K2.5的主要功能视觉与文本融合：Kimi K2.5 能同时处理文字和图像，理解设计图、视频等内容，实现“看图写应用”，将视觉内容转化为可交互的网页或代码，支持像素级复刻和视觉调整。多模式思考：模型支持思考模式和快速模式切换。思考模式适合复杂任务，可进行深度逻辑推理；快速模式适合即时问答，响应迅速。智能体协同（Agent Swarm）：在复杂任务中，Kimi K2.5 可拆解任务并调度多个“分身”协同工作，支持高并发处理，完成多达 1500 步的操作，适用批量任务和多视角分析。专家级办公助手（Office Agent）：模型深度优化 PPT、Word、Excel 等办公软件的使用，能处理文本和解析图表和复杂排版，提供专业级的输出。本地编程支持（Kimi Code）：基于终端的 Coding Agent，支持本地代码的增删改查，自动生成和修改代码，运行测试，并解释改动，方便开发者使用。一键部署与应用生成：生成的网页或应用可直接部署上线，快速将设计转化为实际可用的产品。Kimi K2.5的性能表现Kimi K2.5 是基于约 15T 视觉和文本混合数据训练的原生多模态模型，在视觉理解、图像生成、视频分析等任务中表现出色。例如，在 MMMU-Pro 和 MathVision 等视觉基准测试中，Kimi K2.5 的表现优于或接近其他顶尖模型。Kimi K2.5的项目官网项目官网：https://www.kimi.com/blog/kimi-k2-5.htmlHuggingFace模型库：https://huggingface.co/collections/moonshotai/kimi-k25Kimi K2.5的API调用价格输入（非缓存）：每 M Tokens 价格为 ¥4，相比 K2 Turbo降低了 50%。输入（缓存）：每 M Tokens 价格为 ¥0.7，相比 K2 Turbo 降低了 30%。输出：每 M Tokens 价格为 ¥21，相比 K2 Turbo降低了 63.8%。Kimi K2.5的应用场景软件开发与编程：Kimi K2.5 可将设计图或视频直接转化为可交互的网页或应用，支持本地代码生成、修改和一键部署，大幅提升开发效率。办公自动化：模型能作为专家级办公助手，深度优化 PPT、Word 和 Excel 文档，支持批量任务处理和多任务协作，显著提高办公效率。教育与学习：Kimi K2.5 能根据学生需求生成个性化学习计划和教学材料，提供实时辅导，减轻教师备课负担。创意与设计：模型可将设计稿直接转为代码，支持批量生成创意素材和视觉调整，助力设计师快速完成项目。商业与市场分析：Kimi K2.5 能快速进行市场调研、生成商业计划书，通过数据可视化辅助决策，助力企业高效分析市场。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇ChatTOC - AI聊天对话生成目录插件，一键定位关键信息下一篇DeepSeek-OCR 2 - DeepSeek团队开源的二代OCR模型相关文章Manus 1.5 – Manus最新AI Agent版本，支持Web全栈开发AI小集3HeyGen – 专业的AI数字人视频生成工具AI小集55Ivy-VL – AI Safeguard联合卡内基梅隆和斯坦福开源的轻量级多模态模型AI小集2RenderNet – AI图像生成平台，支持面部锁定姿势控制确保角色一致性AI小集2Scribenote – 兽医AI笔记，可将口头记录转为结构化的宠物医疗记录AI小集2MagicQuill – 蚂蚁集团联合多所高校共同开源的AI互动式图像编辑工具AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付55分钟前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版57分钟前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架58分钟前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准3小时前Composer 2.5 – Cursor 推出的自研 Agentic 编程模型4小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前

Composer 2.5 – Cursor 推出的自研 Agentic 编程模型AI工具4小时前发布AI小集02Composer 2.5是什么Composer 2.5 是Cursor推出的自研 Agentic 编程模型。在智能水平和行为表现上较Composer 2有大幅提升，在SWE-Bench Multilingual（79.8%）和CursorBench v3.1（63.2%）等核心基准上与 Claude Opus 4.7、GPT-5.5 处于同一梯队，但单次任务成本仅为竞品的约1/10，称为”性价比之王”。模型基于 Moonshot 开源的Kimi K2.5检查点持续训练，目前仅通过 Cursor IDE 及 SDK 提供服务。Composer 2.5的主要功能长时任务持续工作：针对长时间运行的 Agent 会话深度优化，能在多步骤工具调用中保持专注，显著减少中途幻觉或提前终止的问题。复杂指令可靠遵循：对跨文件重构、终端命令执行、测试驱动开发等复杂指令的遵循可靠性较 Composer 2 大幅提升。努力级别动态校准：模型能根据任务难度自动分配计算量，简单任务快速完成，复杂任务深入思考，避免”小事空转、大事欠思考”。沟通风格优化：回复更简洁结构化，减少不必要的冗长解释，在多文件变更时提供更清晰的推理过程。工具调用精准度提升：显著减少无效的终端命令或冗余搜索，提升代码检索与终端操作效率。双版本灵活适配：提供 Standard（标准版，$0.50/$2.50 per M tokens）与 Fast（快速版，$3.00/$15.00 per M tokens），智能水平相同，分别适配后台批量任务与交互式实时编程。如何使用Composer 2.5在 Cursor IDE 中启用：将Cursor更新至 2026 年 5 月最新稳定版，打开 Composer 面板（Cmd+I/Ctrl+I），点击模型选择器切换至Composer 2.5选择速度档位：交互式开发默认使用Fast 版（响应快、延迟低）；后台 Agent 或批量任务可在 Settings > Models 中切换为Standard 版（成本低，智能水平相同）通过 SDK 程序化调用：import { Agent } from “@cursor/sdk”; const agent = await Agent.create({ model: “composer-2.5”, // Standard 版 // model: “composer-2.5-fast”, // Fast 版 workspace: “./”, tools: [“edit”, “shell”, “search”, “browser”], });为长时任务设限：为无人看管的长时间 Agent 会话设置迭代次数上限和最大耗时，防止模型利用缓存等”捷径”进行奖励作弊领取首发福利：发布首周提供双倍用量额度，可在后台查看额度消耗情况多模型路由策略：将 90% 日常任务交给 Composer 2.5 处理；仅将架构评审类任务路由给 Claude Opus 4.7，重度终端 Shell 任务路由给 GPT-5.5Composer 2.5的核心优势极致性价比：Standard 版输入 $0.50/M、输出 $2.50/M，Fast 版输入 $3.00/M、输出 $15.00/M，比 Claude Opus 4.7 便宜约1030 倍。前沿级基准表现：SWE-Bench Multilingual 79.8%（Opus 4.7 为 80.5%，GPT-5.5 为 77.8%），CursorBench v3.1 63.2%（与 Opus 4.7 的 64.8% 和 GPT-5.5 的 64.3% 基本持平）。行为层面深度优化：除了扩大训练规模，还改进了沟通风格和投入级别校准，这些维度虽难被基准充分反映，但对实际使用体验至关重要。双版本灵活选择：Standard（标准版）适合后台 Agent 与批量任务，Fast（快速版，默认）适合交互式 IDE 实时编程，两者智能水平完全相同。首发双倍用量福利：发布首周提供双倍用量额度。Composer 2.5的同类竞品对比对比维度Composer 2.5Claude Opus 4.7GPT-5.5厂商 / 平台CursorAnthropicOpenAI产品定位自研 Agentic 编程模型旗舰推理模型旗舰多模态模型基座模型Moonshot Kimi K2.5（开源检查点持续训练）Claude 4 系列GPT-5 系列发布日期2026.05.182026 年 Q22026 年 Q2SWE-Bench Multilingual79.8%80.5%77.8%Terminal-Bench 2.069.3%69.4%82.7%CursorBench v3.1（困难任务）63.2%64.8%（max）/ 61.6%（默认 xhigh）64.3%（xhigh）/ 59.2%（默认 medium）输入价格（/M tokens）$0.50（Standard）<$3.00（Fast）未公开（行业参考约 $15）未公开（行业参考约 $3–$5）输出价格（/M tokens）$2.50（Standard）<$15.00（Fast）未公开（行业参考约 $75）未公开（行业参考约 $15–$30）单次任务相对成本基准（约 $1–$2 / 任务）约10–30 倍约3–10 倍上下文窗口200K（参考 Kimi K2.5）200K128K–1M权重开放性闭源（仅 Cursor 基础设施）闭源闭源接入方式Cursor IDE / CLI /@cursor/sdkAPI / Claude Code / 第三方平台API / ChatGPT / GitHub CopilotComposer 2.5的应用场景多文件级重构：成本优势明显且精度持平前沿模型，适合大规模代码库迁移。交互式结对编程：Fast 版响应迅速，适合实时 IDE 协作。后台定时任务/云 Agent：Standard 版性价比极高，适合批量代码审查与修复。测试驱动开发：长时任务可靠性优于前代，能稳定完成多轮测试-修复循环。复杂终端自动化：Terminal-Bench 2.0 得分 69.3%，与 Opus 4.7（69.4%）持平，但重度 Shell 场景仍略逊于 GPT-5.5（82.7%）。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Higgs Avatar v1 - 面向语音智能体的实时 AI 数字人模型下一篇Chronicles-OCR - 腾讯联合高校等推出的视觉感知评测基准相关文章NLWeb – 微软推出支持自然语言与任何网站交互的开源项目AI小集2Octoparse – AI网页爬虫工具，基于可视化工作流设计器零代码创建爬虫AI小集4白瓜AI – AI内容创作平台，输入主题和关键词自动生成文案和配图AI小集2MiniMax Office Skills – 稀宇科技开源的生产级办公文档引擎AI小集3IntroCard – AI驱动的在线个人名片生成器AI小集2Laminar – 分析与优化LLM应用程序的开源平台AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付35分钟前Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版37分钟前ESP-Claw – 乐鑫开源的物联网设备 AI Agent 框架38分钟前Chronicles-OCR – 腾讯联合高校等推出的视觉感知评测基准3小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型1天前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手1天前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台1天前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前

Boone AshworthGearMay 18, 2026 7:00 AMHow to Watch Google I/OGoogle I/O is back with updates to Search, Android, Gemini, and a fresh peek at upcoming Android XR smart glasses. Here’s how to watch the announcements live and what to expect.Photograph: Camille Cohen/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle I/O isimminent. This is the annual developer event and product showcase whereGoogleshows off all the shiny new updates to its Android operating system and other platforms, as well as new features and improvements to its artificial intelligence models.The big announcements atGoogle I/Ousually come in the form of a livestreamed keynote event at the Shoreline Amphitheatre in Mountain View, California, hosted by Alphabet CEO Sundar Pichai. That starts on Tuesday, May 19, at 10 am Pacific time (1 pm Eastern). You can watch thelivestreamon Google’s website or onYouTube. It will also be streaming right here on this page.What to ExpectOn the announcement front, Google actually tipped its hand last week, unveiling a bunch ofnew featurescoming to the upcoming Android 17 operating system and its ongoing Gemini AI efforts, which it now calls Gemini Intelligence for the most advanced Android devices.Many of the newAndroidfeatures highlighted are AI-infused tricks that aim to make your life a little easier by offloading onerous tasks like booking a hotel or ordering an Uber. Google is also bringing more robust voice-to-text recognition, some autofill smarts, and enhanced digital well-being settings that you’ll turn on for a day and then probably just disable. Also, there is a new look for Android Auto, and Google is making its emojis look more 3D. You canread more about these capabilities here.The company also announced its newGooglebookplatform: AI-first laptops powered by various Android technologies, but it’s very different from a Chromebook. Google won’t share too many specifics yet, but these machines are expected to be premium laptops, and companies like Acer, HP, Lenovo, and Dell have already signed up to make Googlebooks by year’s end. Wehave more details here.We’ll likely hear Google touch on these features at I/O again, but these were unveiled early to make room for more new software announcements during the keynote. You can usually count on a surprise announcement or an AI demo that goes completely off the rails.Based on the ever-swirling rumor mill, Google likely will announce more updates to Search, undoubtedly powered by AI. There may also be updates to its creativeGoogle Labssuite, such as its AI music creation toolFlow. Google may introduce a new Gemini model for video generation,expected to be called Omni.Last year, Google announced a newGemini-powered Google Home smart speakerand said it would launch sometime in the spring of 2026, so Google I/O would be the time to make good on that promise and maybe announce a launch date.While I/O is primarily a software-focused event, Google is heading toward an inevitable crop of new hardware. The company has been building out itsAndroid XRplatform, which it is positioning to power smart glasses from companies like Samsung, Xreal, Gentle Monster, and Warby Parker. Google already said last week that I/O will feature a sneak peek at its smart glasses, and we may learn when in 2026 these devices will finally go on sale.Follow along during the event on WIRED’s liveblog, and stay tuned for many more stories about everything Google announced.

Jeremy WhiteGearMay 18, 2026 9:38 AMThe Catastrophic Audemars Piguet x Swatch Launch Was Entirely Predictable and Utterly AvoidableChaos consumed Swatch stores worldwide on Saturday—a repeat of the MoonSwatch launch four years ago. Swatch should have seen it coming.Photograph: RAPHAELLE PELTIER/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAt 9 am,as stores opened, London’s Carnaby Street location lasted half an hour before police were called. In New York, fights broke out amid rumors of a stabbing in the line. At locations around the world, police had to assist staff as the vast majority of customers left empty-handed, since most of the shops had fewer than 200 watches available. Deals were done within queues as scalpers flipped watches for profit. Within minutes of the first sales, the plastic watches were hitting eBay for thousands of dollars.This is not a description of what happened during Saturday’sSwatch x Audemars Piguet Royal Poplaunch; this is an account of the 2022 MoonSwatch launch. But a carbon copy of this chaos and mayhem ensued as Swatch stores opened worldwide on May 16 to sell their extremely limited numbers of Royal Pop pocket watches.Responding to widespread criticism of how Swatch handled the 2022 MoonSwatch launch, Nick Hayek Jr., chief executive of Swatch Group,told WIREDat the time: “We knew for sure this would be a success, because the product is beautiful, provocative, high quality, and the price is fantastic … But what happened … I think nobody in the world could have expected that. It was really crazy.”Hayek’s defense was essentially that Swatch could not be held accountable for disturbances so severe that some stores were forced to shut for 10 days to let the pandemonium play out, simply because the group could never have predicted the ferocious public reaction to an affordable version of an iconic luxury timepiece coming on sale. No such excuse can be offered this time for the calamitous Royal Pop launch that took place this weekend.Chaos ForeshadowedEven before the first stores opened in Singapore on Saturday, 12 hours ahead of New York, the clear pattern of what was to follow was established. TheVivoCity branch was shutteredahead of the much-anticipated launch, with Swatch issuing a statement that it was “due to the overwhelming crowd” and that this decision was taken with local authorities to ensure the health and safety of staff and customers.The Swatch X AP Royal Pop that’s caused such a furor at launch, just like the MoonSwatch did in 2022.Courtesy of SwatchThen it was Dubai. “In view of safety considerations, we have decided not to proceed with the sale of the [Swatch x Audemars Piguet Royal Pop] at Dubai Mall and Mall of the Emirates and the event has been cancelled,” a statement said. Authorities in New Delhi and Mumbai closed Swatch Store openings citing a lack of crowd control threatening safety. “We are not animals,” crowd members were caught shouting on Instagram clips.Londonstores shut, andpolice dogs were deployed. Elsewhere in the UK, in Birmingham, news reports spoke of “fights and knives” and drug-taking before police were called. A“mosh pit”developed outside stores in the Netherlands. Clashes inDusseldorf. Officersfired teargasto control a 300-strong crowd outside a Swatch shop in Paris. Openbrawling occurredin Milan. Thenarrestsand shutdownin Miami. Houston’s store closed, as well asChicago’s. In Long Island, New York, amazingly,pepper spraywas usedfor crowd control. In store-front interviews, customersmade clear their angertowards Swatch on how poorly the release was handled.Eventually, around 2 pm ET, a staggering 17 hours after the first clear warning signs in Singapore, Swatch finallyissued a statementon its Instagram page: “To all our dear fans worldwide of our AP x Swatch collab, launched on May 16. To ensure the safety of both our customers and our staff in Swatch stores, we kindly ask you not to rush to our stores in large numbers to acquire this product. The Royal Pop Collection will remain available for several months. In some countries, queues of more than 50 people cannot be accepted, and sales may need to be paused.”Looking at almost any of the 45,000 comments on the post, the response to the brand from fans is crystal clear: Swatch had done far too little, far too late. “You intentionally put people in danger causing a frenzy knowing how many watches you sent to each location but allowing thousands of people to line up was negligent.” “Shame on you Swatch.” “Swatch did bad—they should have gotten ahead of this knowing the depth of inventory and allocation rather than putting stores and customers at risk”, and on the replies go.AP Wanted a Safe, Positive ExperienceWIRED contacted Swatch asking why, considering the wealth of evidence from the ill-received MoonSwatch launch, did it repeat the rollout strategy with the AP collaboration. We got this response from their communications team in Switzerland bl

John BrandonGearMay 18, 2026 10:33 AMReview: Leica Cine Play 1Leica’s first home entertainment projector is pricey, but like the company’s cameras, the image quality is worth the splurge.Courtesy of B&H PhotoTriangleUpBuy NowMultiple Buying Options Available$3,795$2,995at B&H$3,795$2,995at Leica$3,795$2,995at CrutchfieldCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyRating:9/10Open rating explainerInformationWIREDLeica’s Summicron lens renders colors beautifully. Compact size allows for flexibility in moving around the home. Straightforward setup; auto-sizing and focus were almost magical.TIREDLooks best in a dark room. Remote has a confusing button layout. Expensive.If your preferenceis to make colors pop on the screen for movie night, the Leica Cine Play 1 delivers with an immersive cinematic experience. This short-throw laserprojectorfrom the company that makes high-end cameras with a cult following ticks almost every box for tack-sharp picture quality, rendering Pixar films and theAvatarseries with vibrant colors and beautiful details.Crafted out of aluminum and glass, Leica’s latest projector borrows many of its design ethos and superb build quality from thecompany’s cameras. It comes with a carrying handle, making it easy to move around the home, and is powered by the intuitive VIDAA operating system. The Leica optical performance makes movies come alive.At $2,995 after an $800 rebate, the Cine Play 1 is competitive with midrange home cinema projectors from Sony, Epson, and others. The only minor ding against the Cine Play 1 is that you’ll need a mostly dark room to take advantage of Leica’s brilliant color reproduction, while competitors like theEpson Pro Cinema LS9000perform well in both dark and well-lit environments.Setting Up the ProjectorThe Cine Play 1 uses a Leica Summicron lens with an outstanding optical zoom. Known for its extreme sharpness and near-zero color fringing on Leica’s storied cameras, the Summicron glass delivers similar benefits on the Cine Play 1. With the Cine Play 1, it’s like having a high-end camera that, instead of capturing a scene, projects one with added vibrancy. Though Leica has made optical lenses for other projectors, this partnership with Hisense makes this the first projector to bear Leica’s branding, complete with the iconic red dot on the top.Photograph: John BrandonThough seemingly mobile at about the size of a large lunch box and weighing 14.6 pounds, the lack of a built-in battery means that the Cine Play 1 is mostly confined to indoor use. Like most other projectors, you can mount the Cine Play 1 on your own stand or opt for Leica’s optional $495 Cine Play 1 Floor Stand. Though expensive, the Floor Stand has an ingenious design that is Apple-like in approach. After mounting the projector to the stand with a couple of bolts, I noticed a proprietary connection, allowing you to connect the power cord into the base of the stand rather than on the rear of the Cine Play 1, solving the problem of unwieldy wires.After mounting the Cine Play 1, I set up the projector roughly 12 feet from the wall. This resulted in a 150-inch image. The Cine Play 1 can project 65- to 300-inch images, but for smaller picture sizes, it would be easier to get a4K TVinstead. No matter how large the projection I was testing, the focus, brightness, and color quality were all above average compared to other projectors.The auto keystone and autofocus features were almost magical, adjusting to my projector screen automatically. I did have to turn off the auto-adjusting after bumping into the projector a few times, but the idea is that it will adjust if you move the Cine Play 1 to a different room.Photograph: John BrandonPowering on the Cine Play 1, I was pleasantly surprised at how vibrant and colorful everything looked in a dark room.The onscreen setup took only a few minutes. The Cine Play 1 uses Hisense’s VIDAA operating system, so I created a new profile and logged into the companion app. At the top of the screen, there’s a series of banners that allow you to access the app store, art mode, and scan a QR code to install the mobile app. I found the OS to be highly intuitive. There’s also a highly effective voice search option if you would rather not rely on the included remote.VIDAA provides access to most streaming services, with Fandango at Home being a notable exception. (Leica reps stated that they may add this service at a later date.)Benchmark Results and First ImpressionsApart from minor brightness issues when using the Cine Play 1 in brighter environments, this projector passed my benchmark tests. Colors were vibrant, skin tones did not appear washed out, and a butterfly scene showed subtle differences between the shades of yellows and oranges in my test clips. In comparison, colors on theTCL NXTVISIONwere noticeably duller. The Cine Play 1 also excels as a gaming display, with superb color reproduction and stutter-free performance when tested withCrims

Paresh DaveMaxwell ZeffBusinessMay 18, 2026 1:40 PMElon Musk Loses Landmark Lawsuit Against OpenAIThe nine-member panel took only two hours to return a verdict in favor of OpenAI on Monday, which the judge quickly adopted as her own final decision.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyElon Musk sufferedthe worst defeat possible inhis legal battle against OpenAIas a federal jury and a judge ruled he waited too long to bring his claims against the AI startup and its top executives, Sam Altman and Greg Brockman.While the jury’s decision was a nonbinding recommendation sent to US district judge Yvonne Gonzalez Rogers, she immediately accepted it as her own, making it final.Musk’s lead trial attorney, Steven Molo, told the judge, “Our intention is to appeal.”One of his other attorneys, Marc Toberoff, gave a one-word comment to reporters walking out of the courtroom: “Appeal.” He later said the verdict reminded him of American Revolutionary War moments like the Siege of Charleston and the Battle of Bunker Hill. “These were major losses for Americans, but who won the war?” Toberoff said. “And this one is not over.”OpenAI’s attorneys hugged in the courtroom after the verdict was read. William Savitt, the company’s lead litigator, told reporters that the “overwhelming” amount of evidence presented in the case allowed the jury to act quickly. “The evidence that Mr. Musk’s lawsuit was an after-the-fact contrivance by a competitor was overwhelming,” he said.Throughout the trial, Gonzalez Rogers questioned Musk’s motivation for fighting OpenAI. But she concluded on Monday that the three-week global public spectacle had been worthwhile.“I thought it was an important issue to be tried … for us to have a trial to bring clarity,” she told attorneys for both parties. “There’s a substantial amount of evidence to support the jury’s findings, which is why I was prepared to accept the jury’s findings and dismiss on the spot.”Thenine-member paneldelivered the unanimous verdict in an Oakland, California, courtroom on Monday after deliberating for under two hours. They found that statutes of limitations expired well before Muskfiled his lawsuitin 2024. Musk had hoped to persuade the jury that Altman and Brockman, with the help of Microsoft’s cash, transformed OpenAI into an enormous company well beyond what was envisioned when the three of them and others founded it as a nonprofit nearly 11 years ago.Because the jury found the case wasn’t filed on time, it didn’t weigh in on Musk’s three claims, including breach of charitable trust, unjust enrichment, and, against Microsoft, aiding and abetting. Losing on what amounts to a technicality could provide Musk an opening to keep trying his case in the public by arguing that the jury never ruled against his core argument that a charity was stolen.Savitt, the OpenAI lawyer, disputed that contention on Monday. “It’s not a technical decision, it’s a substantive one,” he said. “It says you brought your claims too late, and you did it because you were sitting on them to use them as a weapon of a competitor who can’t compete in the marketplace, and so we’re delighted with the outcome.”Microsoft spokesperson Alex Haurek said in a statement that the “facts and the timeline in this case have long been clear” and the tech giant remains “committed to our work with OpenAI to advance and scale AI.”Inpostson X, Musk called Gonzalez Rogers an “activist” judge “who simply used the jury as a fig leaf” and ruled on a “calendar technicality.” While the judge’s ruling doesn’t create a formal legal precedent, Musk claimed that, “She just handed out a free license to loot charities if you can keep the looting quiet for a few years!” He added in a separate post, “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity.”Despite the disappointing result for Musk, the trial appears to have tarnishedthe public image of OpenAI and its top executives. New details emerged aboutBrockman’s wealthandAltman’s alleged history of dishonesty. Both were also pulled away from their day-to-day work for tens, if not hundreds, of hours to conduct depositions, prepare to testify, sit on the witness stand, and show face in court.Musk spent much less time in the courtroom than the OpenAI executives, about three days before never returning again. He even flew to China for President Donald Trump’s state visit last week, though he technically could have been called to testify again on short notice. “I will say that it was a surprise to us to see that,” Savitt told the media last week. “Instead of being in the jurisdiction where he filed the lawsuit, ready to come in front of the jurors who he has caused to be impaneled, [he] decided to get on Air Force One and go to China.”Though the casecarried financial and emotional stakes, it was also a competition of bravado between two t

Molly HigginsGearMay 18, 2026 2:06 PMThe Best Automatic Cat FeedersWe tested some of the most popular automatic dry- and wet-food pet feeders to see which ones are worth the money.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyFeatured in this articleBest AI-Enabled Dual Feeder With CameraPetkit YumShare Dual-Hopper 2 Automatic Feeder With CameraRead more$170PetkitBest Automatic Wet-Food FeederPetlibro Polar Wet Food FeederRead more$170PetlibroBest Connected AppPetlibro Granary Smart Camera FeederRead more$127AmazonA Dual Automatic Feeder With a CameraPetkit YumShare Dual-Hopper Automatic Feeder With CameraRead more$160$110(31% off)WalmartAn automatic catfeeder just may change your life. No more screeching at 4 am. Plus, it can give some peace of mind while you’re away from your fur baby. As seen with the rise ofautomatic litter boxesandpet tracking, technology has made it so we can closely monitor our pets without having to do a lot of the manual labor normally involved.Gravity-fed automatic dry-food dispensers were invented nearly a century ago, but today most automatic cat feeders utilize Wi-Fi and are connected to an app on your phone that allows you to make a schedule or feed on demand from afar, all while monitoring your pet’s eating habits. Newer models have taken this tech further, like thePetkit YumShare Dual-Hopper 2 Automatic Feeder With Camera ($170)which has a built-in, AI-enabled camera to analyze the remaining food in the bowl and will stop dispensing to prevent overfeeding. Automatic wet-food dispensers—our top pick in this category is thePetlibro Polar Smart Wet Feeder ($170)—are relatively new, often using ice packs or mini-refrigeration cooling systems to ensure the food stays as fresh as possible.Automatic feeders are also helpful for pet parents who want to monitor their pets’ health, have specialized schedules, or want to foster healthier eating habits. (A customizable automatic cat feeder can help with the endless gorging-vomiting cycle many cats struggle with.) Nearly all of the automatic cat feeders on this list are marketed toward cats, but some can be used with smaller dogs, and I’ve noted these below.Check out our related guides, like theBest Automatic Litter Boxes,Best Cat Water Fountains,Best Smart Cat Collars, andBest Pet Cameras. More of a dog person? We’ve got you covered; see ourBest Accessories and Tech Essentials for Your Dog.Updated May 2026: We’ve added the Cheerble Match G1 and Oneisall 2-in-1 Automatic Cat Feeder and Water Dispenser, and updated prices and links throughout.Jump to SectionAccordionItemContainerButtonLargeChevronCompare Our PicksOthers TestedDo Not RecommendWhy Use an Automatic Feeder?How I TestedHow Long Can I Leave My Cat Alone?Best AI-Enabled Dual Feeder With CameraPhotograph: Molly HigginsCourtesy of PetkitChevronChevronSave to wishlistSave to wishlistPetkitYumShare Dual-Hopper 2 Automatic Feeder With Camera$170Petkit$170ChewyThis model is an upgrade on one of my favorite models below, the Petkit YumShare Dual-Hopper Automatic Feeder With Camera, but with added AI features. It has a crystal-clear video feed, night vision, customizable meal calls, and a dual-hopper for two different types of food. Through the connected app, you can set feeding schedules or feed on demand from both tanks, use two-way audio, and watch the feeder from your phone. Unlike most other brands, the AI-enabled camera doesn’t require a subscription and can identify up to 20 pets and log when they’ve eaten individually. The camera also uses AI to analyze the remaining food in the bowl and will stop dispensing to prevent overfeeding.This feeder worked excellently throughout testing, and I like that you don’t need to opt in to the subscription plan to have AI recognize different pets. However, I still think the connected Petkit app isn’t as reliable or as easy to use as Petlibro’s app (above).Compatible food typesDry; freeze-driedAppYesPower sourcePower cordCapacity2 and 3 liters (5 total)Warranty1 year limitedWIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDBuilt-in AI-enabled cameraCan store two types of food (including freeze dried)TIREDNeed membership plan to access some featuresApp isn’t as user-friendly as some othersBest Automatic Wet-Food FeederPhotograph: Molly HigginsCourtesy of AmazonChevronChevronSave to wishlistSave to wishlistPetlibroPolar Wet Food Feeder$170Petlibro$170Amazon$170ChewyThis updated Petlibro Polar Wet Food Feeder solves pretty much every gripe I had with other automatic feeders aimed at storing wet food—it’s essentially a mini-fridge guaranteed to keep food fresh, and it’s connected to an app where you can schedule and customize feedings and keep a feeding log. Food is given in three stainless steel wide, shallow bowls, which means less chance forwhisker fatiguewith wide bowls orfeline acnewith bacteria-harboring plastic. When it’s time to feed, the cover slides back from the hole and reveals the food beneath.The ref

Julia ForbesGearMay 19, 2026 1:00 AMTop Tempur-Pedic Promo Codes for May 2026Whether you’re looking to upgrade your mattress or swap out your bedding, browse our Tempur-Pedic coupons for up to $300 off and the best sleep of your life.Save StorySave this storySave StorySave this storyLife is hard,but you know what isn’t? Tempur-Pedic mattresses. This brand’s been around for a long while, which isn’t shocking given the high-quality materials that perform for those that need advanced pressure relief and support. If you’re someone who deals with regular aches and pains, this is a good place to start looking for a new mattress. For those that also want to avoid putting a strain on their budget, now’s a good time to look, as there are several limited-time deals currently running.Save With the Tempur-Pedic SaleShopping the Tempur-Pedic sale is one of the best ways to score huge discounts on usually pricey buy-it-for-life items like mattresses.Tempur-Pedic offersand discounts rotate (but usually hover around 30% off), but I’d recommend checking back often to see what new items are on sale and what new seasonal discounts have sprung up. Some of the best Tempur-Pedic discounts we’ve seen are huge price reductions on pillow bundles, buy two and save 30% on sheet sets, and 25% off pillows and bedding.Get a 30% Tempur-Pedic Promo Code When You Sign UpThis isn’t the “forward this email or be cursed” email chain you got as a kid. Sign up for Tempur-Pedic’s email updates, and get a solid 30% off discount on pillows and sheets. Plus, you’ll be the first to know about new launches, sales, and everything Tempur-Pedic is up to in the sleep realm. Why wait on savingsandbetter sleep?Save $300 and up to 50% Off With Today’s Tempur-Pedic CouponsIf you’re looking to stay cool and comfortable, Tempur-Pedic has got your back (literally). Right now, you get $300 off when youbundle a (qualifying) mattresswith Tempur-Ergo power bases and ProSmart bases. Just as breezy as they sound, each of these mattress lines really emphasizecoolingand aim to dial down the temperature—and from my prior testing experience, they are very cool to the touch. Speaking ofl, you can also get 50% off aTEMPUR-ProAir Sheet Set, which features a moisture-wicking design to keep you cool and comfortable while you sleep. On top of that, be sure to take advantage of more bedding deals, like 25% off pillows and bedding, pillow bundles starting at 2 for $69, and 30% off any 2 Tempur-Pedic sheet sets.Purchase a Tempur-Pedic Mattress Set, Get $300 Off BasesYou’re one of those “all in” kind of people when it comes to big decisions—respect. Game recognizes game, and Tempur-Pedic’s giving you a deal that’s hard to pass up. When you get a select Tempur-Pedic mattress and a Tempur-Ergoadjustable power base, you can get $300 off the base. Plus, you can get 25% off select bedding and pillows to really go for the full package.Enjoy $39 Off Bundle Discounts on Tempur-Pedic PillowsTempur-Pedic doesn’t just create high-quality mattresses that exemplify awesome pressure relief and spine alignment. It also has a variety ofpillowsto carry out this mission, because there’s no lack of support happening on their watch. You can bundle pillows and get a nice discount while you’re at it: Get two Tempur-Cloud Pillows for $119; two dual-sided Tempur-Symphony Pillows for $169, two Tempur-Cloud Dual Cooling Pillows for $259; or two Tempur-Neck Pillows for $179 and save $39. When one pillow already has a higher price tag, this is a really good deal to double up on.Get Two Tempur-Pedic Sheet Sets and Save 30%A nice pair ofsheetscan really make or break your sleep experience. Plus, you’ll need a few pairs depending on the season: somethingcoolinglike theProAirfor summer, rayonbambooif you want a silky feel, orcottonflannelfor a cozier vibe. You’ll want to invest in a few pairs for every season (not to mention for extra clean ones if you want to avoid laundry). When you buy 2 Tempur-Pedic sheet sets, you’ll save 30%.

Julia ForbesGearMay 19, 2026 1:00 AMBarkbox Promo Codes: Up to 50% OffSave on Barkbox subscriptions, including monthly themed collections of plush toys, tough chews, and healthy snacks designed to keep your pup’s tail wagging.Save StorySave this storySave StorySave this storyAs my fellowpet parents will know, it’s amazing how quickly even the tiniest of dogs can demolish their toys and treat stash. We love and spoil them nonetheless. When you subscribe to BarkBox a fresh batch of cleverly themed treats and toys arrives at your doorstep. The costs of pet ownership can stack up quickly, especially if you’re buying your pooch a random gift box that goes well beyond the essentials. That’s why we have Barkbox promo codes and discount options ready to go for you.Barkbox Promo: Enjoy a Free Toy for a Year at BarkboxWhen your monthly Barkbox arrives, it’s like Christmas morning for your dogs. I watch as my two dogs, Rosi and Randy, shake their little Chihuahua mix bodies with barely restrained excitement. They’re never gentle on their toys but the stimulation that comes from textures and chewing is good for their little brains. With Barkbox you get a steady supply of two unique toys and two bags of all-natural treats every month. If you want to see how your dogs react, this Barkbox coupon is good for new Barkbox subscription customers and adds an additional toy in your box every month for a year.Save 50% on Your First Barkbox Food Subscription With a Barkbox Coupon CodeAnother reason why Barkbox is the best dog subscription box is how easy the company makes it to keep your pantry stocked with your dog’s food. Use thisBarkbox couponto save 50% off your first Barkbox food subscription, so you won’t have to end up running out to the grocery store in the middle of the night when your scooper scrapes across the bottom of an empty kibble bin. Use Barbox promo codeBelly50to get 50% off now.Fly Travel Stress-Free With Your Dog and Get $300 Off BARK Air FlightsIf you live in a Barkbox flight hub destination, please know I am insanely jealous of you. It’s no secret that flying is stressful and can be very dangerous for pets, especially if they have to ride in a cargo hold. Barkbox makes them the VIP withBARK Air, letting them ride in the cabin with you and get doted on, so things are a lot less scary. This is another perk of having a BarkBox subscription, with the opportunity to save $300 off BARK Air Flights.Support Your Dog’s Dental Health and Get $10 Off With a Barkbox CouponDental health is crucial for dogs, as it can prevent disease not just in their mouths, but their vital organs. Don’t forget to schedule your yearly cleaning with your vet, but in the meantime, use this BarkBox discount code to get $10 off a specialBarkBox Dental kit.Get an Extra Premium Toy in Every BarkBox With the Extra Toy ClubFor having such tiny mouths, my dogs can gnaw through toys with surprising speed. If you’re also buried in a pile of shredded fluff and squeakers from disemboweled toys, theExtra Toy Clubcan help. This subscription includes dog toys for aggressive chewers of all ages, breeds, and sizes, offering extra durable toys meant to last longer. So far, so good at my house. To upgrade to this subscription box, it’s an extra $9 per month.Get Exclusive BarkBox Discounts: Join the Email ListIf you assume that the punchy branding and witty lingo extend to Barkbox’s email subscribers and not just the box subscription, you’d be correct. As a bonus, you can get exclusive BarkBox discount codes when you sign up to receive these emails. Who also doesn’t love a furry face and reminder of their pet in between work subject lines and bill payment reminders, too?

Nena FarrellGearMay 19, 2026 1:00 AMiRobot Discounts: Roomba Robot Vacuums and Smart Cleaning DealsSave on iRobot products, including robot vacuums and mops designed to handle pet hair, daily messes, and hands-free cleaning with smart home integration.Save StorySave this storySave StorySave this storyThe brand iRobotlaunched the first Roomba robot vacuum back in 2002, and popularity for the handy devices skyrocketed from there. Countless competitors have emerged, but Roomba is still going strong. Its latest models have all the new features we love, from doubling as avacuumand a mop to fantastic navigation and suction. The Roomba Max 705 is currently keeping my house clean as I test it for ourrobot vacuum guide, and it’s doing a great job both mopping and vacuuming the floors in my massive second story.Looking to own your own Roombarobot vacuum? These iRobot promo codes and coupons will make it more affordable than ever to own your own powerful robot vacuum from the iconic brand.iRobot Promo: Treat a Friend to 15% Off a Robot and Save $15 on iRobot AccessoriesI once gifted a friend arobot vacuumand she regularly tells me how it’s a marriage saver. If you want to change a friend’s life (and potentially save their marriage) you cangift a friend 15% offa robot vacuum with this iRobot coupon code, and you’ll get $15 off robot accessories for yourself. A clean win for everyone.Save 33% on the Roomba Max 705 With Anti-Tangle Brushes Using This iRobot OfferI test lots of robot vacuums, and my favorite from iRobot’s lineup is the Roomba Max 705. It’s packed with two anti-tangle brushes, excellent navigation, a self-emptying base station, and left my home sparkling clean. It’s a great choice for a Roomba for pet hair, and it’s done a great job with the litter my cat leaves all over the house, too. Extra bonus: it’s gorgeous. You can get33% offthis powerful robot vacuum right now with this iRobot discount code.Tackle Every Mess Effortlessly: $300 Off the Roomba Plus 405 Combo Robot With AutoWash DockLooking for something a little more affordable? The Roomba Plus 405 Combo Robot is cheaper but still packs mopping and vacuuming, and still comes with an auto-emptying docking station. It’s already a pretty affordable model and a great iRobot vacuum for hardwood floors, and right nowyou can get it for $300 cheaperwith this iRobot coupon.Save $290 on Smart Cleaning: The Roomba 205 DustCompactor Combo Robot Vacuum & MopIf you don’t want a big base station in your home but still want a robot vacuum that can vacuum and mop, the Roomba 205 DustCompactor is for you. The Roomba 205 will compact and store debris for up to 60 days inside the robot itself, no emptying station required. It’s bagless, likemy favorite Shark robot vacuum, so you don’t need to worry about buying and replacing bags to use the vacuum. If that sounds up your alley, don’t miss your chance tosave almost $300on this robot vac with this iRobot promo code.Get 10% Off Your First iRobot Purchase When You Sign up for EmailsNot sure what you want to get from iRobot, but know you want an iRobot discount code? You cansign up for iRobot’s emailsand get a code for 10% off whatever your first purchase is. Enjoy!Save $500 on Max 705 ComboTheRoomba Max 705 Combo robot + AutoWash dockis the perfect household helper. With a self-cleaning roller mop, detailed edge cleaning, anti-tangle dual rubber brushes, and a powerhouse dock, this impressive robovac-mop combo takes all of the gruntwork out of the normal household cleaning chores. Usually $1,300, the Max 705 is nearly half off at $800. If you’ve been looking to upgrade your cleaning routine, snag this model for $500 less than usual, for a limited time.

Matthew KorfhageGearMay 19, 2026 1:00 AMHome Depot Promo Codes: 50% Off in May 2026Save up to 50% today with the latest Home Depot promo codes for appliances, power tools, and more this May.Save StorySave this storySave StorySave this storyThe company prettymuch invented the hardware superstore when it began in 1978, just by being so big. They inflated the neighborhood tool shop into a whole city of lumber, hammers, caulk, power saws, and big rolls of wire. I would know I’m in a Home Depot blindfolded, because of a distinct quality to the air—crisp and particulate, smelling like wood dust and paint and the oiled metal of power tools. The Home Depot smell is buried deep in my childhood, filed somewhere between “building a deck” and “first day of spring.” Anyway, the Home Depot website is just as big. And while it doesn’t smell like sawdust, it’s easier to find stuff. Our roundup includes Home Depot promo codes from 20% to $100 off, new customer coupons, free shipping offers, and Pro discounts to drop prices by as much as 50% when you buy online.Unlock Home Depot Promo Codes for 10% Off or $5 Off With Sign upIf you register for Home Depot’s Style and Decor newsletter, you get a special code for 10% off on furniture and home accents. Otherwise, you cansign up for the Home Depot coupon newsletteror text alerts to get an immediate $5 off the next in-store purchase of $50 or more. Another easy way to save sitewide at Home Depot is to set up a subscription for your go to products and automatically get 5% off and free delivery on your order.Massive Memorial Deals at Home DepotMemorial Day is widely regarded as the unofficial first day of summer, which means it’s time to make sure your patio furniture is in tip-top shape, grills are ready to fire up, and lawn mowers are prepared to cut the lush greenery all summer long. Check out theHome Depot Memorial Day dealsto get massive savings ahead of summer preparation, for steep discounts on everything from Trager grills to Milwaukee lawn mowers.Join the Home Depot Text Program for $5 OffIf you already shop a lot at Home Depot and are looking for other ways to save, signing up forThe Home Depot’s Promo Text Programis a good idea. You’ll receive text messages with special promotions, offers, how-to guides for projects, and great design ideas. Plus, you’ll get $5 off your next purchase when you sign up.Save Up to 50% With These Home Depot Coupon Codes and DiscountsHome Depot deals are in full swing with huge discounts on a myriad of gifts and home project-related items, like10% off flooringfor Pro members with promo codeSAMPLESAVE10, or 20% off Electrolux handheld electronics with codeELECTROLUX20, like the ever-popularElectrolux handheld garment steamer. There are also offers for free tools with battery kit purchases, and buy one get one free deals from top names like Milwaukee, DEWALT, and RYOBI.Power toolsandhand toolsare discounted up to 50% this season, as well as extremely handy lawn andoutdoor power equipmentthat can help with shoveling and general pesky snow removal and home upkeep.Shop Top Home Depot Sale Deals for up to $1,400 Off Appliances, Bath Items, and ToolsAnd as it turns out, the hardware giant also goes hard on discounts, slashinggrillsandgardenand outdoor power tools by up to 50%, not to mention smokers like the new Traieger Woodridge (8/10, WIRED Recommends) alsosold at Home Depot.But that’s not all; Home Depot is covering the whole house! There are appliance deals of up to $1,400 off appliances, $1,000 off washer and dryer sets, and bathroom deals of up to 40% off vanities, 25% off bathtubs, 20% off on toilets, and 15% off showers.Plus, Home Depot has appliance bundle deals where you can get major discounts when you purchase multiple appliances at once.Get 40% Off With Home Depot Deal of the Day CouponsHome Depot coupons of the moment include whopper deals like 15% off storage solutions, and 35% off washers and dryers. Explore more deals onkitchenandother furnituretoo, by checking out thedeals of the moment here. Like we’ve stated, there are so many ways to save at Home Depot, even when you’re shopping online.Special Buy of the Daysinclude steep price drops on certain products or entire brands—but the significant price drops only last for 24 hours. That’s a whole lot of savings for one of the most expensive home renovations you can do. Another saving hack? Ensure you jump on products marked with “New Lower Price” to find them at their lowest price yet.Winter is a great time to stay inside and focus onhome improvementprojects. Home Depot has everything you need for remodels (including bulk buys). Right now, vinyl plank and tile flooring is up to 30% off, with options starting at 99 cents per square foot. Plus, power tool kit combos to save even more, with bundle deals of up to 70% off—including Deals of the day flash sales on tools, and appliances with rebates of up to $1,000. Deals include smart home items WIRED hascovered extensively,like Nest learning

Luke LarsenGearMay 19, 2026 1:00 AMTop Lenovo Coupon Codes and Deals in May 2026Whether you’re shopping for a ThinkPad, Yoga laptop, or Legion gaming PC, these Lenovo discount codes and promotions can help you save big on your next tech upgrade.Save StorySave this storySave StorySave this storyLenovo is currentlythe biggest PC and laptop company in the world, and they just so happen to make some of our favorite laptops and PC peripherals. In fact, our list ofthe Best Laptopsyou can buy include a number of Lenovo devices that we’ve tested and fully recommend, ranging from high-end premium laptops to options forbudget-mindedshoppers. If you are looking to save a few bucks while shopping for a new Lenovo product, here are the Lenovo coupon codes and special offers you need to know about.Score Weekly Tech Deals, Price Match Guarantee, and Free ShippingIn addition to the deals listed above, Lenovo has a number of other current offers available when shopping throughLenovo.com. There are weekly deals on PCs and tech that include free shipping with no minimums. For certain products, you can even choose to buy online and pick up in-store at Best Buy. If you’re looking for more Lenovo recommendations, we’re big fans of theLenovo Idea Tab Pro, an Android tablet that’s currently on ourBest Tabletslist. It’s currently discounted by 26%, bringing the price down to $280.Lenovo is also offering price-matching on any comparable PC, including from many online retailers. That’s really helpful, as it’s not uncommon to see price discrepancies between different retailers.Get Exclusive Savings With Lenovo Discount ProgramsLenovo does offer a fairly generousdiscount programfor certain demographics as well. These include a 5% off Education, Healthcare, First Responder, and Senior (over 55) discount.You can use this discount on anything you find atLenovo.com, which includes a lot more than just Lenovo-branded laptops. For example, Lenovo sells one of our favoritelaptop docking stations, the Kensington Triple Video Mobile Dock. Despite its tiny size, this tiny device can power up to three 1080p (or two 4K) displays simultaneously. It’s a great example of the wide diversity of products you can use these discounts and coupon codes on.Buy More Save More With LenovoIf you’re looking to buy multiple devices, or are planning to make a big tech purchase soon, I’d recommend buying now to get more savings. Use Lenovo promo codeBUYMORELENOVOat checkout to save $15 on purchases of $500-$750, $25 off purchases of $750-$1,000, save $40 on purchases of $1,000-$2,000, and get $100 off when you purchase something $2,000 or more. (Excludes clearance and workstation items.)Get 35% Off the Lenovo Tab ProTheLenovo Idea Tab Prois a perfect tablet for students, with a vibrant nearly-13 inch 3K touchscreen display which is perfect for every day tasks, like reading, studying, and streaming. It also has a MediaTek processor, 8GB memory, and 128GB storage for seamless multitasking with ample storage for apps and files. Plus, it has dual cameras for Zoom calls. Although the Lenovo Idea Tab Pro is specifically designed with students in mind, it’s versatile enough to be a great fit for a multitude of people wanting a reliable tablet with tons of extra features. Plus, it’s 35% off right now. There’s no better time to buy.Enjoy Perks When You Join My Lenovo RewardsLenovo has aMy Lenovo Rewards systemthat can earn you points that can be spent on future purchases fromLenovo.com. Almost everything you buy gets you between 3% and 9% back in the form of reward points, and you can start redeeming points right after signing up for an account. 1,000 points is equal to $1 that you can spend on future purchases. On top of this, you’ll benefit from frequent sweepstakes, exclusive member-only savings, birthday rewards, free expedited delivery, and more. Sign up today—it’s free.Lenovo Financing and Lease-to-Own OptionsLenovo usesKatapult to offer financing and lease-to-own optionson its laptops. That might be useful if you’ve been eyeing one of Lenovo’s higher-end laptops. One of our favorite of these is the Lenovo Yoga 9i, a gorgeous 2-in-1 with a bright OLED screen and a luxurious design. This laptop is currently listed as one of our top recommendations for theBest LaptopsandBest 2-in-1 Laptop.Katapult allows you to buy now and pay over time, without charging late fees. If you put down a $45 initial payment, you can check out the laptop today. Katapult offers up to $3,500 in pre-approved leases and allows you to make minimum payments over time or buyout early.

Molly HigginsGearMay 19, 2026 1:00 AMBooking.com Promo Codes and Deals this MayEnjoy big savings on your next adventure with our Booking.com promo codes, coupons, and handpicked travel deals.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyLike other online travel agencies, Booking .com helps you find discount rates for airline tickets, cruises, hotel stays, car rentals, and packages. We at WIRED regularly post updates with the newest Booking .com promo codes for discounts on car rentals, last-minute hotel bookings, and other travel-related expenses, including a deal for 50% off stays and free cancellation when you sign up at Booking .com. Whether you want to grab an apartment in a walkable neighborhood or be bad and bougie in a villa, Booking .com has tons of options for every type oftraveler—and we have a Booking .com coupon code to help you save.Get 15% off Your Next Stay Using a Booking.com Promo CodeChecking out Booking’sdeals pageis one of the best ways to snag great discounts on rotating and limited-time deals on things like flights and stays. Right now, you can get at least 15% off your next stay withEarly 2026 Deals.Don’t Miss Spring Deals at Booking.comSpring is in the air—school breaks, holidays, and vacation hullabaloo commence! That’s where Booking comes in. Not sure where to go this vacation season to get the best bang for your buck? Be sure to check out Booking’sspring dealspage. It’s the one-stop destination to find the best deals on top vacation spots, from stays in Tokyo starting at $36 per night to Orlando starting at $68 per night. Check out these rotating deals on spring and summer travel to getaway for less than ever before.Booking.com Promo Code: Up to 20% With the Genius Loyalty ProgramSave more by signing up for Booking .com’s loyalty program,Genius, which offers tons of discounts and rewards on pretty much everything travel-related. Loyalty program members can even get up to 20% off stays and up to 15% off car rentals. The program works in tiers: level 1 gets you a 10% discount on select stays and rental cars; level 2 gets you up to 15% off (once you complete 5 bookings in two years); and level 3 gets you up to20% off stays and up to 15% off rental cars (once you complete 15 bookings in two years). Once you sign up for the Genius loyalty program, make sure you’re signed in while you browse to get discounts of up to 50% off stays and free cancellations, along with bonus secret deals from Booking .com.Earn $150 in Travel Credits on Stays With Booking.comIf you’re someone who travels frequently, it may be a good idea to get aBooking.com Genius Rewards Visa Signature Credit Cardso you can get even more rewards on purchases you were already going to make. With this card, you’ll have no fee, and get 6% in travel credits on stays booked in the Booking.com app, and 5% on other travel on Booking.com when you use your card. Plus, you’ll get $150 in travel credits when you spend $1,500 using your card in the first 90 days. The more you use, the more perks you’ll get—when you get to Genius Level 3, you’ll receive rewards like 10-20% off select stays, and 10-15% off select rental cars and priority support.Deals Around the World Starting at $18 per NightGlory be! Winter is starting to fade and sunshine is on its way. Make the most of springtime withspring travel dealsat outrageously low prices, including stays in Istanbul starting at $18 per night, Paris from $67 per night, Cancun from $32 per night, and Las Vegas from $128 per night. Go ahead, check it out, you deserve a nice vacay from enduring the winter.Sail Towards Warmer Waters (Even While on a Budget)A cruise is one of the cheapest ways to have an all-inclusive vacation, while staying on the mainland and partying at sea. If you’ve ever been curious about taking a cruise (or are a returning sailor), now’s a great time to book for so much less. If youbook a qualifying sailingdeparting on or before December 31, 2027, you’ll get up to $1,000 to spend onboard, which counts toward almost anything, like cocktails, specialty dining, spa treatments, and other onboard purchases.The actual amount you’ll get to spend vary based on total cruise price, and don’t include travel protection, port charges, port expenses, and taxes, and are generally as follows: $25 per $1-$999 booking, $50 per $1,000-$1,499 booking, $75 per $1,500-$1,999 booking, $100 per $2,000-$2,999 booking, $125 per $3,000-$3,999 booking, $175 per $4,000-$5,999 booking, $250 per $6,000-$7,999 booking, $350 per $8,000-$9,999 booking, $500 per $10,000-$14,999 booking, $750 per $15,000-$19,999 booking, and $1,000 per $20,000 or more booking.Plus, if you don’t have the entirety of the money owed for acruisenow, you can still lock in now for just $25. As long as you pay a small non-refundable $25 fee, Booking will advance yourcruise linedeposit (up to $500). Your deposit will be automatically charged 10 days before your final payment is due, giving you peace of mi

Molly HigginsGearMay 19, 2026 1:00 AMLimited Time: TurboTax Coupons Full Service This MayTax season doesn’t have to be stressful. Score 10% off full service expert on federal tax filings and more exclusive TurboTax discount codes on WIRED.Save StorySave this storySave StorySave this storyTurboTax has beenmy go-to for self-filing taxes for years. It’s good enough for me—and literally millions of other users. TurboTax has quickly become a favorite among users for its easy-to-use self-filing services, along with easy access to assistance from tax professionals. TurboTax has free self-service for qualifying people, and TurboTax Live Assisted, where you file online with tax expert help if needed, and TurboTax Full Service, where a tax expert will do your taxes and review them with you before filing.To help you figure out the often confusing process of filing taxes, I’ve written a guide onHow to Pay Your Taxes Online, andThe Best Tax Services. In both, I included TurboTax and similar services to help navigate between the options available for your specific tax needs. WIRED also has TurboTax coupons to save money while you begin the always-annoying task of filing your taxes this year. Canadian customers clickhere, these coupons are not for you.Save 10% With TurboTax Service Codes and 2026 Early Filing OffersRight now, you can getan extra 10% offFederal online tax filings with today’s TurboTax coupons. This price includes W-2 and 1040, with state returns for an additional fee. With ExpertFull Service, a tax professional will do your taxes for you, which costs a bit more than DIY-ing it, but can save you headaches and messes when it comes to what you owe. Things only get more expensive as tax season approaches, so file now to get the best price of the season.If your taxes are a bit more straightforward, but you still have some questions,TurboTax Expert Assist(Live) offers the best of both worlds, where you file online yourself with tax expert help nearby if you end up needing it. With these TurboTax deals, you’ll get a 10% off TurboTax Live Assisted, with options starting at only $36 and (temporarily) capping out at $150 if you want to fully hand off your taxes. (Like the above service, this discount is for Federal tax products and filing early is the best way to ensure you’re getting the lowest price of the season).*Price estimates are provided prior to a tax expert starting work on your taxes. Estimates are based on initial information you provide about your tax situation, including forms you upload to assist your expert in preparing your tax return and forms or schedules we think you’ll need to file based on what you tell us about your tax situation. Final price is determined at the time of print or electronic filing and may vary based on your actual tax situation, forms used to prepare your return, and forms or schedules included in your individual return. Prices are subject to change without notice and may impact your final price.Give a 20% Off TurboTax Discount Code, Get up to $500To save even more, get aTurboTax referral link, and when your friends use your link, they can get 20% off TurboTax federal products if they’re new customers of the service. As an added bonus, for each friend who files using your link, you’ll get a $25 gift card—you can earn up to a whopping $500 in gift cards (at this point, you’re basically getting another refund in gift cards). Here’s the nitty gritty: friends canreceive 20% offthe preparation and filing of federal tax return or business federal tax return through a TurboTax Online or TurboTax Live product, through October 31. You’ll get an email with the gift card or charitable donation within 45 days of your friend’s purchase. (Restrictions apply on TurboTax Free Edition, Intuit TurboTax Verified Pro, TurboTax State returns, TurboTax CD/Download products, and other add-on offerings.)Save up to 20% on TurboTax Business ServicesIf you’re asmall business owner, Expert Assist is probably the best service for you, with qualified tax experts at the ready to help you with any of those tricky filing questions. Right now, if you file your business tax returns with TurboTax, you’ll get 20% off one 2025 personal federal tax return. This type of service is best for those with S-corps and partnerships, and TurboTax will match you with a small business tax expert who knows your industry to maximize your deductions.In addition to 20% off one personal tax filing offer forbusiness customers, you can also get 10% off Business Expert Full Service and Expert Assist for small business owners. This discount applies to TurboTax federal products only (meaning no state products are included). These massive savings are in addition to early-season discounts.File for $0 With TurboTax Free EditionAs mentioned above, users can file for free withTurboTax free edition. Around 37% of taxpayers qualify. Simple Form 1040 returns only (no schedules except for Earned Income Tax Credit, Child Tax Credit, Student Loan

Molly HigginsGearMay 19, 2026 1:00 AMTop HBO Max Promo Codes This MayStream your favorite shows and save up to 50% today with HBO Max discount codes and subscription deals.Save StorySave this storySave StorySave this storyHBO Max isnot just forThe Sopranosanymore (but it is the only place to streamThe Sopranos). Subscription video service HBO Max not only hasThe Sopranos, but other content from the best entertainment libraries, including: Warner Bros., Discovery Channel, CNN, Cartoon Network, Adult Swim, Animal Planet, and other brands.We at WIRED also watch a lot of content—obviously, because we are always testingTVs,soundbars, andstreaming devices—and regularly round up our newestfavorites from HBO Maxso that you’ll know what to watch once you nab an HBO Max promo code. Whether you’re wanting to watch an HBO original series likeSuccessionor a smash hit movie likeBarbie, HBO Max has the content you’re looking for, and WIRED has the perfect promo codes to help you save on these sometimes-pricey subscription plans.50% Off HBO Max Promo Code for StudentsStudents rejoice, you don’t have to try to mooch off your friends’ parents’ HBO Max subscriptions anymore when you want to rewatch Euphoria for the 15th time (…not speaking from personal experience or anything). HBO Max is offering aStudent Discount, where you can save and get 50% off subscriptions with a promo code. This means the price is only $5 a month. This discounted plan is eligible for The HBO Max Basic With Ads monthly plan, so you’ll get unlimited access to all of HBO, hit series, movies, news, sports, and more on up to two devices at once.Save 35% by Bundling HBO Max, Disney+ and HuluThis deal is honestly one of the best I’ve ever seen, and I’ve been thinking of switching as well, because with thisMagic Bundle, you can get HBO Max, Disney+, and Hulu for just $17 a month. (which is crazy because I pay that much alone for my Hulu! Again, I’m convincing myself while writing this that I should switch to this plan.) This crazy good deal is for those three streaming services with ads, but you’ll be saving 35% on the original price. If you’re impatient like me, you can get the same deal without ads for $30 a month, which is definitely higher than with ads but you’re still saving almost $19 a month with this Evergreen deal.HBO Max Deals: Save up to $41Like other streaming services, HBO Max has different tiers at different prices for streaming, so you can choose which one fits you (or your “household;” wink) best. Basic with ads allows you to stream on two devices at once in full HD for $10 a month. Standard is midrange, where you get everything from the previous basic plus 30 available downloads so you can watch on the go, for $17 a month. If you are impatient and a movie/TV buff, you might want to choose the premium, where you can stream on four devices at once, have 4k Ultra HD video quality, Dolby Atmos immersive audio, and 100 downloads to watch on the go for $21 a month.HBO Max New Shows and FilmsHBO Max is known for being one of the most heralded platforms in entertainment. These new shows and films on HBO Max are sure to be a hit too. We’re most excited forTask, a new series set in the working-class suburbs of Philadelphia, where an FBI agent heads a task force aiming to end the violent robberies led by an unsuspecting family man. There’s also a new GOT spin-off series,A Knight of the Seven Kingdoms, which premiered on January 18. Taking place a century before the events ofGame of Thrones, set in an age when the Targaryen line still holds the Iron Throne and the last dragon has not yet passed from living memory. There’s alsoChair Companyfrom the oddball mind of Tim Robinson, andIT: Welcome to Derry, a spin off set in theITuniverse for fans of horror.And especially right now in April, several of my absolute favorite shows are returning for a brand-new season or their last season. Long-awaited season three ofEuphoriareturns, with more drug-fueled drama.Hacks, the final season premiered and I’m going to miss the tumultuous relationships that orbit Deborah Vance’s solar system. Like everyone else, I’m libbing out every Thursday with the sexy doctors inThe Pitt. Please don’t go, I love you so. Plus, I’ve been loving the new Steve Carell comedyRooster, about a wholesome but goofy romance writer that begins a tenure at college.I missed the GOT boat, and it’s a little too violent for me, but my brother recommendedA Knight of the Seven Kingdoms, the newestGame of Thronesspinoff that’s a little lighter and more family-friendly than the first and I think I’m becoming a late bandwagon-er. There’s alsoThe Dark Wizard, an in-depth four-part documentary about extreme climber Dean Potter that I have in my queue for next weekend.

Matthew KorfhageGearMay 19, 2026 1:00 AMTop Tempo Promo Codes: Get up to 60% Off Healthy Meals in 2026Upgrade your meal prep with these Tempo discount codes and offers. Save on high-protein, ready-to-heat meals delivered to your door.Save StorySave this storySave StorySave this storyI had beenleery of prepared meal delivery services. They gave me flashbacks to those delivery diet plans sold on late night television in the 1990s. And soTempo Meals, a spin-off prepared meal brand from meal kit service Home Chef, came as a pleasant surprise even aftertesting dozens of meal kits.Tempo meals look like TV dinners, but they’ve never been frozen, thanks to advances in packaging technology. The beef barbacoa and sous-vide chicken I tried actually retained moisture and tenderness. Green beans were still vibrantly green, with a little bit of snap. They’re cooked in a commissary kitchen, and shipped to your doorstep in special packaging that prevents spoilage without using a bunch of preservatives. Heat-and-eat meals won’t ever be quite as good as freshly cooked meals, but especially when I reheated them in an air fryer they got closer than expected.Tempo is especially known for high-protein and GLP-1 plans created in consultation with dietitians, helping prevent people from losing muscle mass when on calorie-restricted diets or just in midlife. Meals are $11 to $13 a serving at regular price, but Tempo Meals discount codes are available. For now, with a Tempo Meals promo code, you can get up to 60% off your first box of Tempo, with discounts throughout the first month.Score 30% Off Your First Month of Four Boxes in AprilLike a lot of meal delivery services, Tempo Meals wants you to try them out in hopes you enjoy the service and decide you can’t live without it. And so Home Chef offers a Tempo Meals coupon code for new customers. Depending on the Tempo promo code, this may be a straight discount of 30% off for the first four weeks. Or it may be 50% or 60% off the first box, and then discounts off the next month of boxes. Right now, you can also sign up for a lifetime free breakfast item. For previous customers who for whatever reason paused their Tempo subscription, there are Tempo Meals promo codes for existing customers that show up in your inbox. These may automatically be applied when you visit the Tempo website.Activate the Heroes Discount to Get 50% Off Now and 10% Off for a LifetimeMost Tempo Meals promo codes are good for deals during the introductory week only. But theTempo Meals heroes discountapplies for life for military personnel, first responders, teachers, doctors, nurses, and other hospital employees. To get the heroes discount from Tempo, you’ll need to verify your identity with the ID.Me system. After this, you get a permanent discount—pretty much the best discounts available if you plan on using Tempo long-term for a high-protein meal program.

Scott GilbertsonGearMay 19, 2026 1:00 AMTop Norton Coupon Codes and Deals: 58% OffWhether you’re looking to protect your small business or your personal computer, we have the top coupons and deals to help you save at Norton.Save StorySave this storySave StorySave this storyNorton Antivirus isthe only anti-virus software I’ve ever used, in part because many years ago I switched to Linux, which eliminates the need for antivirus software. If you’re still on macOS or Windows, though, Norton Antivirus is still a solid choice. None of these services is something anyone is excited to run, but they can save you bacon if you make a bad decision downloading questionable files or clicking links you shouldn’t have. The best way to stay protected without spending a fortune is to use a Norton coupon code to get yourself a deal.Today’s Norton Coupons: Up to 58% off 360 and Antivirus PlansWhile I’ve only ever used the Antivirus offering, the real way to save is to bundle several services and use a coupon code for the lot. With Norton that means Norton 360, which gets you the Antivirus, malware, ransomware, and hacking protection, along with a VPN and dark web monitoring, for up to three devices. You also get a password manager and 2GB of online storage. With this Norton coupon code, you get all that $40 for the first year, which is aboutthe same as paying for a VPN alone. Norton 360 Deluxe gets you everything that comes with the regular version, but it works with five devices and gives you 50GB of online storage.Get Unbeatable Norton With LifeLock Protection From $8.33 per MonthLifeLock is Norton’s usual online protection suite, but adds up to $1 million in potential coverage if you experience financial loss due to identity theft. With this Nortoncoupon code, Norton 360 with LifeLock is just $8.33 per month.Norton Anti Virus Coupons: 50% Off 360 Antivirus PlusIf you want to get (and stay) serious about your privacy and protection, and the overall health and longevity of your computer, I’d suggest investing in an anti-virus protection plan for your whole setup. WithNorton 360 Antivirus Plus, you’ll get powerful antivirus protection for your device, including AI-powered Scam Protection, Password Manager, and 2 GB Cloud Backup. And you can get even more peace of mind with a 60-day money back guarantee. Best of all,Norton 360 Antivirus Plusis 50% off right now, taking it from $60 per year to only $30—which works out to being just $2.50 per month—a small investment worth so much more in protection and peace of mind.Test out Norton 360 for Yourself With a Free TrialNot sure if Norton 360 is worth the money? You can try it out for seven days, though set yourself a reminder because after the initial7-day free trial, your paid subscription will begin unless you cancel before.

GamingNewsPlayStationSony is raising short-subscription prices for PlayStation PlusOne-month and three-month PlayStation Plus subscriptions are getting more expensive.One-month and three-month PlayStation Plus subscriptions are getting more expensive.byJay PetersMay 18, 2026, 4:24 PM UTCLinkShareGiftJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Sony ishiking the starting priceof one-month and three-month PlayStation Plus subscriptions in “select regions,” blaming “ongoing market conditions.” Beginning May 20th, 1-month subscriptions will start at $10.99 USD / €9.99 EUR / £7.99 GBP, while 3-month subscriptions will start at $27.99 USD / €27.99 EUR / £21.99 GBP.In the US, the changes result in a $1 price increase for a one-month Essential tier subscription and a $3 increase for a three-month Essential tier subscription. It’s unclear if other PlayStation Plus tiers will see price hikes as well. PlayStation didn’t immediately reply to a request for comment.If you are a current subscriber, prices will stay the same for you (except in Turkey and India, according to Sony) unless you change your subscription tier or let your subscription lapse.The PlayStation Plus price hikes follow two recentpriceincreasesfor PS5 consoles. Sony said its April PS5 price hike wasdue to“continued pressures in the global economic landscape.”Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGamingNewsPlayStationMost PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechDesignNewsGoogle is rolling out its redesigned Workspace app iconsThe new designs feature soft color gradients, rounded corners, and overhauled icon shapes.The new designs feature soft color gradients, rounded corners, and overhauled icon shapes.byStevie BonifieldMay 18, 2026, 4:29 PM UTCLinkShareGiftImage: The VergeStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.It’s not just you — the Google Workspace apps are getting a new look. The redesigned app icons,leaked last month, are now rolling out widely, as we started noticing this morning. Users with the redesigned icons will notice they now have a gradient look that fades from lighter to darker shades, rather than being the same flat tone throughout, similar to the redesigned Google logothat launched a year ago.Image: The VergeSome of the icons switched from a rainbow design to a single color, like Google Chat, Meet, and Calendar, which could help all of the icons stand out a bit more from one another — or make them harder to recognize. Others haven’t changed as much, as9to5Googlenotes, like Google Docs, Sheets, and Slides, which have the same colors and general approach. Asone Reddit userpointed out, though, the Sheets and Slides icons were switched to landscape mode, which is also how most people use both of those apps.Image:9to5GoogleThe Google Drive icon also got a significant overhaul with rounded corners (and it lost the spot of red it had in the bottom right corner). A few also lost the borders around them, like Google Keep, which is now just a yellow light bulb, rather than a bulb on a yellow rectangle. As someone who frequently uses Keep, it’s going to take me a while to get used to spotting a totally different icon shape, although I like the overall look of the new one. I also appreciate the cleaner look of the Gmail icon, although it hasn’t changed too much compared to the old version.The new icons are arriving right beforeGoogle I/O kicks off on Tuesday,where Google could announce even more visual changes to its ecosystem.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAppsDesignGoogleNewsTechMost PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AINewsPolicyElon Musk loses his case against Sam AltmanThe Musk v. Altman jury dismissed all claims due to the statute of limitations.The Musk v. Altman jury dismissed all claims due to the statute of limitations.byHayden FieldMay 18, 2026, 5:39 PM UTCLinkShareGiftImage: Alex Parkin / The VergePart OfAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIsee all updatesHayden Fieldis The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.After around two hours of deliberation, the jury has reached a unanimous verdict inMusk v. Altman, the tech trial of the year. The group found that two claims were barred by the statute of limitations, and a third failed thanks to the dismissal of one of these.The jury here is an advisory jury, meaning the group is installed solely to offer another opinion to the judge, and its verdict is technically not legally binding. Ultimately, US District Judge Yvonne Gonzalez Rogers is the ultimate legal authority — and she accepted the decision.The jury found that Musk’s claim for breach of charitable trust was barred by the statute of limitations, and the claim that Microsoft aided and abetted such a breach failed with it. Restitution is also barred by the statute of limitations, the jury found.OpenAI did not immediately respond to a request for comment. On X, Muskposteda statement saying he’d be filing an appeal. He said that the judge and jury ruled “on a calendar technicality” rather than the “merits of the case” and that “there is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!”Alex Haurek, a Microsoft spokesperson, said in a statement, “The facts and the timeline in this case have long been clear, and we welcome the jury’s decision to dismiss these claims as untimely. We remain committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”RelatedMusk v. Altman proved that AI is led by the wrong peopleMusk vs. Altman is here, and it’s going to get messyMusk v. Altmanhas taken over a federal courtroom in Oakland for three weeks, with the core accusation being that OpenAI strayed from its founding mission and that Musk’s money was earmarked for a nonprofit in particular. Musk alleges that OpenAI CEO Sam Altman and company president Greg Brockman breached OpenAI’s charitable trust and participated in unjust enrichment at Musk’s expense. He also alleges that Microsoft aided and abetted the two in breach of charitable trust. Both sides have used every opportunity to smear each other — and throughsalacious evidenceandeyebrow-raising testimony, both sides have come out looking somehow even less trustworthy than when the court process began.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Hayden FieldAIElon MuskLawMicrosoftNewsOpenAIPolicyTechMore in:All of the updates from Elon Musk and Sam Altman’s battle over OpenAIMusk v. Altman proved that AI is led by the wrong peopleHayden FieldMay 18The jury has delivered a unanimous verdict.Richard LawlerMay 18An observer has just been ejected from the court by the US marshals.Elizabeth LopattoMay 18Most PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechUpdatedMay 18, 2026, 7:00 PM UTCAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIbyElizabeth LopattoandHayden FieldLinkShareRSSGiftSam Altman and Elon Musk are facing off in a high-stakes trial that could alter the future of OpenAI and its most well-known product,ChatGPT. In2024, Musk filed a lawsuitaccusing OpenAI of abandoning its founding mission of developing AI to benefit humanity and shifting focus to boosting profits instead.After nearly a month of the hearing that featured testimony from Musk, Altman, Microsoft CEO Satya Nadella, OpenAI cofounder Greg Brockman, former OpenAI board member and mother of several of Musk’s children Shivon Zilis, and a few others, the jury deliberated for a couple of hours before returning with a verdict,deciding to dismiss all chargesdue to the statute of limitations.Musk was a cofounder of OpenAI and claims that Altman and Brockman tricked him into giving the company money, only to turn their backs on their original goal. However,OpenAI claimedthat “This lawsuit has always been a baseless and jealous bid to derail a competitor” in a bid to boost Musk’s own SpaceX / xAI / X companies that have launched Grok as a competitor to ChatGPT.In his lawsuit,Musk asked for the removal of Altman and Brockman, and for OpenAI to stop operating as a public benefit corporation.People to KnowPlaintiffElon Musk— plaintiff, OpenAI cofounder and now CEO of rival xAISteven Molo— lead counsel for the plaintiffJared Birchall— manager of Musk’s family officeShivon Zilis— former OpenAI board member who shares multiple children with MuskDefendantSam Altman— defendant, CEO of OpenAIWilliam Savitt— lead counsel for the defendantGreg Brockman— president of OpenAI as well as a cofounderIlya Sutskever— former chief scientist at OpenAI and a cofounderJudgeYvonne Gonzalez Rogers— aka YGR, trial judgeHere’s all the latest on the trial between Musk and Altman:HighlightsElon Musk loses his case against Sam AltmanMicrosoft doesn’t want any of thisSam Altman was winning on the stand, but it might not be enoughMusk’s biggest loyalist became his biggest liabilityMira Murati tells the court that she couldn’t trust Sam Altman’s wordsGoogle’s AI architect lived rent-free in Elon Musk’s headOpenAI’s president does ‘all the things,’ except answer a questionElon Musk’s worst enemy in court is Elon MuskElon Musk appeared more petty than preparedMusk vs. Altman is here, and it’s going to get messyPinnedMay 18Hayden FieldElon Musk loses his case against Sam AltmanImage: Alex Parkin / The VergeAfter around two hours of deliberation, the jury has reached a unanimous verdict inMusk v. Altman, the tech trial of the year. The group found that two claims were barred by the statute of limitations, and a third failed thanks to the dismissal of one of these.The jury here is an advisory jury, meaning the group is installed solely to offer another opinion to the judge, and its verdict is technically not legally binding. Ultimately, US District Judge Yvonne Gonzalez Rogers is the ultimate legal authority — and she accepted the decision.Read Article >May 18Hayden FieldMusk v. Altman proved that AI is led by the wrong peopleImage: Cath Virginia / The Verge, Getty ImagesThe tech trial of the year,Musk v. Altman, was ultimately a fight for control. Elon Musk argued that Sam Altman, with whom he helped found the now-massive company OpenAI, shouldn’t direct the future of AI. Altman’s lawyers, in turn, poked at Musk’s own credibility. A jury came to averdicton Monday after just two hours of deliberation, dismissing Musk’s claims due to the statute of limitations.In a strictly legal sense, three weeks of testimony added up to nothing. But the trial offered a more damning broader takeaway: Almostnobody in this saga seems worth trusting. Some of the most powerful people in tech seem temperamentally incapable of dealing with each other honestly. And if that’s true, it raises a bigger question: Why are they in control of atrillion-dollarindustry that’s set to upend people’s lives?Read Article >May 18Richard LawlerThe jury has delivered a unanimous verdict.That was quick (about two hours).Elon Musk lost his case against Sam AltmanHayden FieldMay 18May 18Elizabeth LopattoAn observer has just been ejected from the court by the US marshals.I assume because he was recording, since the marshal said, “Give me your phone.” There have been several incidents of people attempting to record or take pictures throughoutthe trial— but I honestly am not sure why you’d record today of all days.May 18Elizabeth LopattoC. Paul Wazzan is the expert called by Musk to determine damages.Unfortunately he does not have a lot of details YGR is asking for. He doesn’t know how many investments Musk has made (11 to date, according to Pitchbook), or how many were successful. He’s getting some pretty tough questioning from YGR in the direct exam. Among the things she’d asked, he didn’t know how many startups fail in Silicon Valley, or the success r

GadgetsTechSmart HomeDyson’s super-slim PencilWash just hit its best price to date for Memorial DayThe slim hard-floor cleaner slides under furniture, scrubs away spills, and removes dirty water as it cleans.The slim hard-floor cleaner slides under furniture, scrubs away spills, and removes dirty water as it cleans.bySheena VasaniMay 18, 2026, 6:32 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: DysonSheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.IfDyson’s PencilVac Fluffyconesmade you wish the company had built something similarly slim for scrubbing the hard floors in your home, enter the recently releasedDyson PencilWash. The cordless cleaner is designed to tackle spills and stains on tile and other surfaces, and right now, it’s available for a new low price of $249.99 ($100 off) fromAmazon,Target, andDysonin the run-up to Memorial Day.Dyson PencilWash$250$35029% off$250$250 at Amazon$250 at Target$250 at DysonWith its 1.5-inch-diameter handle and a design that lets it lie nearly flat, the PencilWash can easily slide under furniture and other tricky areas that bigger floor cleaners can’t reach. Built for hard surfaces like tile, laminate, and sealed wood, it uses a pressurized hydration system and a microfiber wet roller to scrub away spills while extracting dirty water with each pass, so you’re always cleaning with fresh water. There’s also a Max mode for addressing dried-on messes that are particularly tough to clean, along with a flexible neck that makes it easier to steer around obstacles.RelatedDyson put someone else’s motor in its robot vacuumHere are 40 of our favorite deals from REI’s massive Anniversary SaleThese new Roombas are smaller and cheaperThe PencilWash is designed to be low-maintenance, too. Debris and wastewater are collected directly in the head rather than traveling up through the machine, preventing them from becoming trapped in a filter. It shouldn’t leave your floors soaking wet, either, and it comes with a sleek charging dock that lets it eke out 30 minutes per charge (you can also purchase abackup batteryfor $99.99). The 300ml clean water tank should be enough to cover over 1,000 square feet on the lowest hydration setting as well, after which it will need a refill.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Sheena VasaniDealsGadgetsSmart HomeTechVerge ShoppingMost PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechNewsGoogleWalmart launches new budget-friendly Android tablets starting at $97The six new Onn tablets ship with Android 16 and include kids’ tablets, Core models, and a 13-inch Pro model.The six new Onn tablets ship with Android 16 and include kids’ tablets, Core models, and a 13-inch Pro model.byStevie BonifieldMay 18, 2026, 6:44 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: WalmartStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.Walmart’s Onn brand just launched a whole line of budget-friendly Android tablets that, all together, cost less in total than a single iPad Pro. As spotted by9to5Google, the six new tablets ship with Android 16 and all but one cost less than $200.The littleOnn Core 7is the most affordable of the bunch at just $97, which gets you a 7-inch 1040 x 600 IPS LCD display, 4GB of RAM, 64GB of storage that’s expandable with a microSD card, and, according to Walmart, up to 10 hours of battery life.PreviousNext1/3The Onn Pro includes a folio case and stylus.Image: WalmartA step above that are two larger Core tablets. TheCore 8.1costs $138 and includes an 8.1-inch 1524 x 1000 IPS LCD display, 6GB of RAM, 64GB of storage (also expandable with microSD), and up to 15 hours of battery life. TheCore 11for $167 is a bit bigger, similar to a base iPad, featuring an 11-inch 1840 x 1280 IPS LCD display, 6GB of RAM, 128GB of storage that’s expandable with a micoSD slot, and up to 17 hours of battery life. There are also two affordable kids’ tablets available in8-inchand11-inchsizes for $118 and $136, respectively.The priciest of the new tablets is theOnn Pro, which costs $288 and includes a 13.2-inch 2400 x 1600 IPS LCD display, 8GB of RAM, 256GB of storage, and an unnamed 2.6GHz MediaTek processor. Unlike the iPad Pro it’s imitating, it also includes a folio case and a stylus.Onn Core 7 tablet$97$97$97 at Walmart (blue)$97 at Walmart (silver)$97 at Walmart (mocha)Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAndroidGoogleNewsTechMost PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AIPolicyReportMusk v. Altman proved that AI is led by the wrong peoplePublic opinion of the AI industry is already sinking. A parade of untrustworthy executives makes it look worse.byHayden FieldMay 18, 2026, 7:00 PM UTCLinkShareGiftImage: Cath Virginia / The Verge, Getty ImagesAIPolicyReportMusk v. Altman proved that AI is led by the wrong peoplePublic opinion of the AI industry is already sinking. A parade of untrustworthy executives makes it look worse.byHayden FieldMay 18, 2026, 7:00 PM UTCLinkShareGiftPart OfAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIsee all updatesHayden Fieldis The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.The tech trial of the year,Musk v. Altman, was ultimately a fight for control. Elon Musk argued that Sam Altman, with whom he helped found the now-massive company OpenAI, shouldn’t direct the future of AI. Altman’s lawyers, in turn, poked at Musk’s own credibility. A jury came to averdicton Monday after just two hours of deliberation, dismissing Musk’s claims due to the statute of limitations.In a strictly legal sense, three weeks of testimony added up to nothing. But the trial offered a more damning broader takeaway: Almostnobody in this saga seems worth trusting. Some of the most powerful people in tech seem temperamentally incapable of dealing with each other honestly. And if that’s true, it raises a bigger question: Why are they in control of atrillion-dollarindustry that’s set to upend people’s lives?OpenAI was, in the testimony of both Musk and Altman, founded to stop powerful AI from being owned and advanced by the wrong people. Testimony and evidence showed its founding team fretting about who would control artificial general intelligence (AGI), a buzzword for AI that broadly equals or surpasses human knowledge and ability. Theydeeply fearedGoogle DeepMind and its leader, Demis Hassabis. In 2015, Altman said he’d been mulling over whether anything could “stop humanity from developing AI” — and after concluding it was impossible, that he wanted “someone other than google to do it first.”Fellow cofounders Greg Brockman and Ilya Sutskever so strongly opposed one-person control that they seemed willing to torpedo a lucrative deal that could — in their words — give Musk an “AI dictatorship.” In a part of the same email addressed to Altman, Brockman and Sutskever questioned his motivations, writing, “We haven’t been able to fully trust your judgements throughout this process … Is AGI truly your primary motivation? How does it connect to your political goals?”These concerns would bequickly borne out. A central focus ofMusk v. Altmanwas “the blip,” a five-day period in November 2023 when OpenAI’s board removed Altman as CEO. Sutskever had spentmore than a yeararchitecting his ouster, assembling a 52-page memo alleging “a consistent pattern of lying, undermining his execs, and pitting his execs against one another.” The implications were broader than executive infighting, potentially impacting the public rollout of AI systems. Then-CTO Mira Murati, for instance, testified in court that Altman told her OpenAI’s legal team had okayed skipping a safety review for one of its models — a statement, she said, that turned out to be false.In closing arguments, Musk attorney Steven Molo hammered home the long list of people who had testified under oath that Altman was, in one way or another, a liar — all of whom Altman had worked with for years. “The defendants absolutely need you to believe Sam Altman,” Molo told the jury. “If you cannot trust him, if you don’t believe him, they cannot win. It’s that simple.”But during court proceedings, Musk — who now leads competing lab xAI, under his space company SpaceX — didn’t come off any better. Joshua Achiam, now OpenAI’s chief futurist, testified that Musk’s race against Google led him to take an “obviously unsafe and reckless” approach to achieving AGI. When he and others raised concerns, he says, Musk argued that OpenAI’s for-profit makeover created incentives to disregard safety, but his own xAI is for-profit and has, at best, ahaphazard approach to safety. And in the name of making sure OpenAI remained open, Musk was obsessive in his need for control over it. In closing arguments, Sarah Eddy, one of OpenAI’s attorneys, told the jury that Musk “wanted dominion over AGI.”As one X userput it, “if untrustworthyness had mass, putting Musk and Altman too close to one another would collapse the courtroom and all of earth into a black hole.”OpenAI did not immediately respond to a request for comment. On X, Muskposteda statement saying he’d be filing an appeal.It’s not just Musk and Altman, either.Trial evidence suggestedMurati helped get Altman removed, then switched sides to support his reinstatement while appearing “totally uninterested” in disclosing the role she’d played. Shivon Zilis, a close Musk associa

GamingNewsPC GamingPlayStation exclusives aren’t coming to PC anymoreA big pivot back to consoles.A big pivot back to consoles.byJay PetersMay 18, 2026, 7:28 PM UTCLinkShareGiftImage: Sony Interactive EntertainmentJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Sony reportedly won’t release its major single-player PlayStation games on PC anymore. According toBloomberg’s Jason Schreier, Hermen Hulst, who heads up PlayStation’s studios business,informed employeesin a town hall on Monday about the change in strategy. Schreier had previously reported on the shiftin March, saying that Sony scrapped plans to launch PC versions of last year’sGhost of Yōteiand “other internally developed games.” Online games will still come to multiple platforms following this change in strategy, Schreier reported at the time.In recent years, Sony has released many of its biggest games on PC, includingSpider-Man 2,Ghost of Tsushima, bothThe Last of Usgames,Horizon Zero Dawn Remastered, and multiplayer titles likeHelldivers 2andMarathon. Two years ago, Hulstcommitted toreleasing PlayStation’s live-service games “day and date” on PC and PS5, but its single-player PC releases have been less consistent, with Hulst saying that the company takes a “more strategic approach.”RelatedConsole exclusives might be making a comebackSony may not be the only one to go back to exclusives; Asha Sharma, Microsoft’s new Xbox chief,told staffthat she is “reevaluating” exclusive games for the platform.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGamingNewsPC GamingPlayStationMost PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TransportationNewsCarsVolvo teases a new affordable EV to replace discontinued EX30The new EV won’t be the same price as the EX30, which was supposed to start at $35,000, but ended up being much more than that thanks to tariffs.The new EV won’t be the same price as the EX30, which was supposed to start at $35,000, but ended up being much more than that thanks to tariffs.byAndrew J. HawkinsMay 18, 2026, 8:41 PM UTCLinkShareGiftImage: VolvoPart OfThe great EV pullback: all the obstacles, cancellations, and delayssee all updatesAndrew J. Hawkinsis transportation editor with 10+ years of experience who covers EVs, public transportation, and aviation. His work has appeared in The New York Daily News and City & State.Volvo’s compact, quirky EX30 had a lot of problems when it was first released. Tariffs essentially erased its affordability, making it more expensive to own, and a battery recall made it dangerous to park indoors. Butits discontinuationdidn’t spell the end of Volvo’s efforts to sell more affordable electric models. In fact, the Swedish automaker is already at work on a new offering for the US market.The news of an affordable Volvo EV for the US came during a media roundtable this week related tothe US launch of the new EX60. Luis Rezende, president of Volvo Cars America, said that the decision to discontinue the EX30 was not solely about tariffs and profitability, noting that the company is preparing to introduce a new EV in 2027 that will occupy a similar role in the lineup — though not necessarily at exactly the same price point as the EX30.“Very similar, I would say,” Rezende said about the mystery EV’s price comparison to the EX30. “It’s going to be an EV that will deliver a lot of good things in a bigger space, but it will be also fun to drive, I can promise you.”Other than that, details were scarce. Volvo’s executives talked later about the desire to build a larger, family-oriented SUV at its factory in Charleston, South Carolina — though that vehicle will likely use a “multi-fuel” strategy rather than being exclusively electric from launch.The EX60, which will start customer deliveries in the US this summer, is Volvo’s attempt at a reset in the US. The compact SUV, which is built on a different architecture than the EX30, will start at $59,795 for the entry-level P6 Plus version, and climbs up to $68,745 for the more powerful P10 AWD Ultra variant.Volvo is the latest automaker to try, and stumble, in its efforts to build an affordable EV for the US market that is both desirable and profitable for the company. To date, few have pulled it off, as it requires a certain level of scale, vertical integration, and mastery of the supply chain that only companies in China seem to have really nailed down. Of course, Volvo is owned by China’s Geely, but the company’s desire to sell EVs in North America will necessitate a different approach to affordability.Affordability was one of the EX30’s main selling points. When it wasfirst announcedin 2023, Volvo said the price would start at $34,950, positioning it as the smaller, less expensive EV that many people were clamoring for. But after the election of Donald Trump,Volvo was forced to delay the EX30’s arrival in the USuntil 2025, citingnewly leveled tariffsagainst vehicles built in China. Eventually, the model that went on sale in the US started at $44,900, about $10,000 more than the original price.Then, in February, further bad news as Volvo issued a recall for the EX30 because the vehicles’ batteries were at risk of overheating or catching on fire. The next month, Volvo pulled the plug on the vehicle in the US.Correction May 18th:A previous version of this story stated that the EX60 is the only Volvo EV in the US. The EX90 is also available.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Andrew J. HawkinsCarsElectric CarsNewsTransportationVolvoMore in:The great EV pullback: all the obstacles, cancellations, and delaysHonda’s hybrid future starts with new Accord and RDX prototypesStevie BonifieldMay 14Mazda delays its first EV by two years.Andrew J. HawkinsMay 12Ford’s EV and software chief Doug Field is leaving the companyAndrew J. HawkinsApr 15Most PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

GamingNewsTechLG will release the first 1000Hz, 1080p gaming monitor this yearThe 25G590B is a 24.5-inch IPS screen with the highest refresh rate we’ve seen for a 1080p gaming monitor.The 25G590B is a 24.5-inch IPS screen with the highest refresh rate we’ve seen for a 1080p gaming monitor.byRichard LawlerMay 19, 2026, 1:10 AM UTCLinkShareGiftIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: LGRichard Lawleris a senior editor following news across tech, culture, policy, and entertainment. He joined The Verge in 2021 after several years covering news at Engadget.If you just can’t choose between refresh rate and resolution, LG’s next gaming monitor could solve your problem, as the UltraGear 25G590B monitor is the first one announced that will be capable of a native 1000Hz refresh rate at 1,920 x 1,080 resolution. So far, the 1,000Hz models we’ve seen have only been capable of 720p at the most, but this 24.5-inch IPS display does not have that limitation, as the company says it’s ready for esports competitors (whether they’re in an FPS orExcel), who need maximum responsiveness.We were alreadyasking if anyone could feel the effect of higher refresh rates a few years agowhen they hovered around 360 and 480Hz, but now the bar has moved higher once again.LG hasn’t revealed the 25G590B’s price or release date beyond “second half of 2026,” but said it has a “minimalist” stand, an integrated hook for headset storage, and customizable lighting. More information will be availableon LG’s website here.Of course, there are also AI-enabled features, too:The 25G590B incorporates on-device gaming AI features designed to boost both immersion and usability. AI Scene Optimization adds visual realism and depth by intelligently adjusting picture settings according to game genre, while AI Sound presents a more authentic spatial audio experience along with clearer in-game communications (when using a compatible headset). These AI-driven capabilities complement the monitor’s ultra-fast performance to create the ideal platform for competitive gaming.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Richard LawlerGamingLGNewsPC GamingTechMost PopularMost PopularBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft is retiring Teams’ Together ModeMusk v. Altman proved that AI is led by the wrong peopleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

When Lior Susan startedEclipse Venturesin 2015, the firm’s thesis of digitizing the physical world wasn’t particularly popular in Silicon Valley.“It was the era of enterprise software and SaaS, and it felt fairly lonely the first couple of years,” Susan said onstage at a recent StrictlyVC event in San Francisco.More than a decade later, Eclipse finds itself at the center of the tech world’s action. The firm’s $6.5 million Series A investment in Cerebras Systems in 2016 paved the way for a total return of $2.5 billion when the semiconductor company went public this week. The firm invested a total of $147 million in Cerebras over time, a bet that generated a seventeenfold return at theIPO price of $185per share, according to Eclipse.For Susan, the windfall from Cerebras is only the beginning of reaping big rewards from a long-standing belief that because 85% of global GDP is tied to the physical world, investing in companies beyond pure software could be immensely lucrative.Public markets and startup founders seem to be recognizing the value of physical-world tech now, too. Susan noted that shares of TSMC and Micron recently hit all-time highs, while a growing cohort of elite founders is eager to build startups at the intersection of hardware and software.“I think people understand that the real moat in software is gone. You can vibe-code pretty much whatever you want,” he said.Susan echoed public market sentiment that earlier this year sent many SaaS stocks tumbling on the belief that enterprises may use Anthropic’s Claude Code or OpenAI’s latest models to create their own bespoke software tools instead.“What you cannot do with ‘vibe code’ is manufacture wafers, because you need machines and silicon, and they need clean rooms, and a bunch of other things,” Susan said.When it comes to the tech that touches the physical world, it’s not just semiconductors that are suddenly catching the attention of investors and founders.Eclipse’s portfolio companies spanning sectors like robotics, energy and defense, raised nearly $15 billion from outside backers last year, and that late-stage momentum reached $4.5 billion in Q1 2026 alone, Susan said. That investor excitement stands in stark contrast to the firm’s early track record: In its first eight years, its portfolio companies raised less than $4 billion in total.Indeed, the recent follow-on rounds across Eclipse’s portfolio show a track record that any venture firm would envy. Driven by a string of massive late-stage deals this year, the haul includes$1.2 billion for Wayve, $650 million for True Anomaly, $270 million for Bedrock Robotics, and $200 million for Oxide Computer. What’s more, Eclipse was the Series A investor for all four companies.At first glance, it may seem that investor enthusiasm for physical-world tech is driven purely by AI, whether as an infrastructure input like chips and data centers, or through AI’s power to finally make robotics viable. However, Susan argues that there are other powerful tailwinds driving the momentum.Besides technology — in this case, AI — what’s important for this market to thrive is capital, customer demand, talent, and policy. Susan means that along with investors and engineers moving away from SaaS to sectors like robotics, semiconductors, space, and mining, the U.S. government is also encouraging these industries through subsidies and favorable regulation.“This is the first time I believe in America ever, from Henry Ford and Carnegie, those five forces are aligned,” Susan said. “For builders like us, this is the best time to build those companies.”TopicsCerebras,Eclipse Ventures,Fundraising,Hardware,Lior Susan,Space,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retrac

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. To get this in your inbox, sign up here for free — just clickTechCrunch Mobility!There is a bit of a theme emerging in transportation — and really every industry: AI is creating jobs for some at the loss of others.General Motors, for instance,laid off more than 10% of its IT department, or about 600 salaried employees — in a deliberate skills swap. This won’t translate into a one-to-one exchange, which means there will likely be a net-negative job loss. But GM insists it is hiring and those layoffs have made room for it to recruit IT people with AI-focused backgrounds.The most sought-after capabilities are AI-native development, data engineering and analytics, cloud-based engineering, agent and model development, prompt engineering, and new AI workflows. In practical terms, GM is looking for people who know how to build with AI from the ground up — designing the systems, training the models, and engineering the pipelines — not just use AI as a productivity tool.Those AI job losses are mounting in the automotive sector. CNBCcalculatedthat Ford, GM, and Stellantis have cut a combined total of more than 20,000 U.S. salaried jobs, or 19% of their combined workforces, from recent employment peaks this decade. While there are a variety of reasons for these cuts, they are generally connected to technological changes, including AI.Companies are leaning heavily into AI, although anecdotes from some engineers and founders suggests not all of these businesses know quite what they’re doing with it yet.Samsarais one company that seems to have figured out a revenue-generating use case. The company has spent the last decade giving its customers cameras to mount inside millions of trucks for driver monitoring, theft prevention, and helping with liability claims. The company took that mountain of data and trained its own model that candetect potholesand determine how quickly they are deteriorating. The company is pitching this product to cities and announced it has several under contract, including Chicago.A little birdImage Credits:Bryce DurbinNothing this week, although I am working on a fun one! Reach out anytime with insights, tips, or just because. You can reach us via email or Signal.Email Kirsten Korosec [email protected] my Signal at kkorosec.07, or email Sean O’Kane [email protected]!Image Credits:Bryce DurbinYou might have noticed that Rivian’s spinoff companyMind Roboticsraised another $400 million, just two months afterraising $500 million. And that pace got me thinking about its founder RJ Scaringe and his innate ability to get VC and institutional backers to invest in his ideas and projects.I calculated that investors have poured $12.3 billion into Scaringe’s three startups — Also, Mind Robotics, and Rivian. That figure doesn’t include the close to $12 billion in gross proceeds raised in Rivian’s IPO, nor did I count the more recent strategic deals with Volkswagen Group and Uber — which together could add nearly $7 billion to Rivian’s coffers.You can read my whole riff on the topic here. But if you don’t feel like clicking, here is one item that stood out. I spoke to a number of insiders and investors and they all mentioned Scaringe’s ability to give undivided attention to whoever he’s talking to — whether it’s an investor, supplier, or exec — and make them feel like the most important person in the room.It’s yet another piece of evidence in my long-standing case against multitasking. Debate me!Other deals that got my attention …Arkeus, an Australian startup that developed perception software for autonomous drones and aircraft,raised $18 millionin a Series A round led by QIC Ventures. Other investors include R+VC, Folklore Ventures, DYNE Ventures, Main Sequence Ventures, Salus Ventures, and Beaten Zone.Aseon Labs, a Redwood City, California, startup that has developed a depot in a box for charging, cleaning, and inspecting autonomous fleets,came out of stealthwith undisclosed backing by Y Combinator.Rapidoraised $240 millionin a round led by Prosus, and that values the Indian ride-hailing company at $3 billion. Existing investors, including WestBridge Capital and Accel, participated. The round was part of a larger $730 million primary and secondary financing.Quantum Systems, a Germany-based drone startup backed by Peter Thiel, is in talks to raise around €600 million ($703 million) with companies like Airbus and Blackstone as investors,Bloomberg reported.Notable reads and other tidbitsImage Credits:Bryce DurbinIs Redwood Materials ready for an IPO?Senior reporter Sean O’Kane interviewed the company’s new CFO,Deepak Ahuja, whose name will be familiar to anyone who follows Tesla. Ahuja was Tesla’s former finance chief and most recently held a similar position at drone company Zipline.TeslaRobotaxis have crashed at least twice since July 2025 while ateleoperator was remote

Commencement season has come around again — and this year, a couple speakers have discovered that it’s tough to get graduating students excited about a future shaped by artificial intelligence.Last week, Gloria Caulfield, an executive at real estate firm Tavistock Development Company,gave a speech at the University of Central Floridaacknowledging that we’re living in a time of “profound change,” which can be both “exciting” and “daunting.”“The rise of artificial intelligence is the next industrial revolution,” Caulfield declared — prompting the students in the audience to begin booing, getting louder and louder until Caulfield chuckled, turned to the other speakers, and asked, “What happened?”“Okay, I struck a chord,” she said. Caulfield then tried to resume her speech, saying, “Only a few years ago, AI was not a factor in our lives” — only to be interrupted again by the audience, this time by their loud cheers and applause.Former Google CEO Eric Schmidt faced a similar response when he brought up AI at a University of Arizona speech on Friday.In Schmidt’s case, the criticism actually began before the speech itself, with some student groupscalling for him to be removed as commencement speakerdue toa lawsuit in which a former girlfriend and business partner accused Schmidt of sexual assault. (He has denied the allegations.) According toa local news report,the booing began even before Schmidt took the stage.But Schmidt alsogot loud booswhen he told students, “You will help shape artificial intelligence.” The booing was persistent enough that Schmidt tried to speak over it, insisting, “You can now assemble a team of AI agents to help you with the parts that you could never accomplish on your own. When someone offers you a seat on the rocket ship, you do not ask which seat, you just get on.”To be fair, AI isn’t becoming a third rail ateverygraduation ceremony. Nvidia CEO Jensen Huang recentlyspoke atCarnegie Mellon’s commencement, and he didn’t seem to get any audible pushback when he said that AI has “reinvented computing.”Still, it’s not exactly surprising to find some students in a booing mood. Ina recent Gallup poll, only 43% of Americans aged 15 to 34 said it’s a good time to find a job locally, a steep drop from 75% in 2022.That pessimism isn’t solely a response to the rise of AI (a shift thateven some software engineers are worried about), but journalist and tech industry criticBrian Merchant suggestedthat for many students, AI has become “the cruel new face of hyper-scaling capitalism.”“I too would loudly boo at the prospect of this next industrial revolution if I was in my early twenties, unemployed, and had aspirations for my future greater than entering prompts into an LLM,” Merchant wrote.Even when graduation speeches didn’t mention AI explicitly, “resilience”was a recurring theme this year. Schmidt himselfacknowledgedthat there is “a fear in your generation that the future has already been written, that the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create.”Caulfield, meanwhile, might also have misread her audience of arts and humanities graduates. One student said that before mentioning AI, Caulfield already started to lose them with her “generic” praise of corporate executives like Jeff Bezos.Another graduate, Alexander Rose Tyson,told The New York Times, “It wasn’t one person that really started the booing. It was just sort of like a collective, ‘This sucks.’”TopicsAI,commencement speeches,Eric Schmidt,gloria caulfieldWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Anthony HaAnthony Ha is TechCrunch’s weekend editor. Previously, he worked as a tech reporter at Adweek, a senior editor at VentureBeat, a local government reporter at the Hollister Free Lance, and vice president of content at a VC firm. He lives in New York City.You can contact or verify outreach from Anthony by [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Listen onApple PodcastsListen onSpotifyLawyers for Elon Musk and OpenAI made their closing arguments this week, and now it’s up to jurors to decide whether OpenAI did anything wrong as it’s transformed intoa slightly-more-for-profitorganization.But as Kirsten Korosec, Sean O’Kane, and I noted on the latest episode of TechCrunch’sEquity podcast, a big theme in the trial’s final days waswhether OpenAI CEO Sam Altman is trustworthy— for example, Musk’s attorney Steve Molo grilled Altman about whether statements he’d made during congressional testimony were truthful.Kirsten noted that Musk hasmade plenty of misleading statements of his ownand that trust isn’t just an issue for Altman.“This is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs,” she said. “It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies; there’s a lot behind the veil still.”Keep reading for a preview of our conversation, edited for length and clarity.Anthony Ha:[The end of the trial] led to this really provocative headline from one of our writers, Tim Fernholz, [that] just says,“Who trusts Sam Altman?”Does anyone want to take a stab at answering this?Kirsten Korosec:Yeah, Anthony, I’m going to throw it right back to you. Do you trust Sam Altman?Anthony:It’s an interesting question because it feels like something that’s kind of a wild question to discuss in a journalistic context, but actually that’s the core of the trial, in a lot of ways.Sean O’Kane:That’s not a yes.Anthony:And it actually seems to be [at the] core of understanding so much of what’s happened at OpenAI, especially this big executive power struggle that they now call The Blip.It just seems like a lot of people who’ve worked with Altmandon’t trust him. And he’s acknowledged this a little bit, because he’ll talk about the fact thathe recognizes he’s been conflict averse, telling people what they want to hear, and he’s trying to work on that.I mean, it sounds plausible, and I can understand how that can lead to misunderstandings in some situations. [But] I’m also a very conflict-averse person and I’d like to think that if any of this stuff went to trial, that people would not be asking, “Is Anthony Ha trustworthy?”Sean:Still not a yes!Kirsten:I think that people would say that you are trustworthy. I will say that question, while provocative, doesn’t just encapsulate what this trial was about. I would zoom out even more and say this is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs. It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies; there’s a lot behind the veil still.Maybe when they all IPO, we can get a peek, but it is fundamentally about trust and misuse, and do we believe the intent? And what I would throw back is, sometimes the intent can be worthy, noble, and still misused. It can still end up as a bit of a s— show. I think it’s more than who trusts Sam Altman — although that was very interesting in this trial — but more of that bigger question that we can apply to the entire industry.Sean:I’ll say it: I don’t trust him. But you know, I don’t trust most people, so I guess that’s just the baseline.We’ll see where this goes. The trial wraps up today. I’ve been very curious to hear how the jury decides this all. I think at the start of this, a big motivator of this was Elon Musk trying to sling mud at a perceived rival and someone who he feels slighted him. And I don’t know if we know enough yet to say that that was completely accomplished, and whether or not he has a shot at winning. But I think all these people came out of this looking a little bit worse.Anthony:And just to get specific, why this is coming up this week is that [Altman] was on the stand and he was basically getting grilled about some statements he’s made in the past, in testimony to [Congress], basically saying he didn’t have any equity in OpenAI. And that is not true because he had a stake through Y Combinator, which he used to run. And tried to brush that off by saying, “I assume that everybody understands what it means to be a passive investor in a VC fund.” And I think [Elon Musk’s] lawyer, somewhat fairly, said “Really? You think the congressman who was interviewing you knew that?”Kirsten:Yeah, I mean, he was playing the whole semantics game. What I thought was so interesting about [this] is the style of how Sam Altman answered questions [compared to] Elon Musk on the stand.So Elon Musk, in many, many, many scenarios and many instances, we can point to the fact that he put something out on Twitter that was a lie or a bit of a fib, and on the stand corrected the record. So there’s a history of, I would say, non-truthfulness-slash-lying, blatant or otherwise, in Elon Musk’s world, but how he treated it was incredibly combative an

In BriefPosted:1:15 PM PDT · May 17, 2026Image Credits:Artur Widak/NurPhoto / Getty ImagesAnthony HaApple’s Siri revamp could include auto-deleting chatsPrivacy will be a major theme when Apple unveils a new version of Siri at the Worldwide Developers Conference in June,according to Bloomberg’s Mark Gurman.The Siri relaunch is widely seen as Apple’s big chance to reestablish its relevance in artificial intelligence. As part of that effort, company executives will argue that they’re taking a more privacy-friendly approach than most other AI companies, Gurman said.Apple will reportedly launch the first stand-alone Siri app,powered by Google Geminiand offering users a chatbot experience reminiscent of ChatGPT.  But compared to those other chatbots, the app is supposed to have more limitations on how long user information can be used and stored.For example, Gurman said Siri could include a feature similar to the Messages app, allowing users to automatically delete conversations after 30 days or one year — or to keep them indefinitely.Gurman also suggested that Apple might be emphasizing privacy as a way to excuse Siri’s shortcomings compared to competing products — and that this emphasis might obscure the fact that Google is handling some of the security.TopicsAI,Apple,Apps,gemini,In Brief,siriMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAIAnthropic has acquired the dev tools startup used by OpenAI, Google, and CloudflareKirsten Korosec16 minutes agoAIElon Musk has lost his lawsuit against Sam Altman and OpenAITim Fernholz2 hours agoSecurityNYC Health + Hospitals says hackers stole medical data and fingerprints during breach affecting at least 1.8 million peopleZack Whittaker3 hours agoLatest in AIAIAnthropic has acquired the dev tools startup used by OpenAI, Google, and CloudflareKirsten Korosec16 minutes agoAIElon Musk has lost his lawsuit against Sam Altman and OpenAITim Fernholz2 hours agoAIAmazon’s new Alexa+ powered feature can generate podcast episodesLauren Forristal5 hours ago

Imagine you’re riding a motorcycle at 160 kilometers per hour when an arrow appears, floating on the road ahead, telling you exactly where to turn. No phone, no dashboard. Just your helmet, and a lens the size of a thumbnail.This is not a concept video. It’s heading to European roads as early as this year. And it’s one early glimpse of where smart glasses are heading.Over the past few years, Big Tech has been quietly (and not so quietly) placing its bets. Meta has been selling AI-enabledRay-Ban glasses since 2023, Google isbuilding Android XR, andApple is expected to enter the market. Last week, Samsung wasreportedlyset to unveil its first AI-capable smart glasses, co-designed with Gentle Monster, at a Galaxy Unpacked event in London this July. China’sHuawei,Alibaba,Xiaomi, and others are all moving too.The numbers reflect the momentum. Global AI glasses shipments surged to 8.7 million units in 2025, up more than 300% from the prior year, and analysts project that figure will cross 15 million this year,per Omdia.Suppliers and component makers of AI-powered smart glasses are also positioning themselves for what comes next. One of the companies, a South Korean startup calledLetinAR, has spent the last decade building the optical technology that could make all of this actually wearable.The LG Electronics-backed startup just secured $18.5 million from Korea Development Bank and the South Korean retail giant’s venture arm, Lotte Ventures, among others, ahead of its planned 2027 IPO in South Korea.Its previous investor,LG Electronics, has since begun developing its own AI smart glasses, according to alocal mediareport, which is a sign of how seriously South Korea’s largest consumer electronics company takes the category.LetinAR CEO Jaehyeok Kim and CTO Jeonghun Ha, who have been friends since high school, founded the company together in 2016.Image Credits:LetinARThe lens that makes it wearableLetinAR doesn’t make the glasses. It makes the part that makes the glasses work. The optical module, the tiny lens component that projects images into your field of vision, is what determines whether a pair of smart glasses feels like a sci-fi headset or something you’d actually wear to work, Ha told TechCrunch. It has to be light, thin, and power-efficient, while still delivering a sharp, clear image. Getting all of that right in a single component, small enough to fit inside a normal-looking frame, is the central engineering challenge of the entire industry. That’s what LetinAR is building.“We see AI glasses as that next platform,” Kim said. “And the optical module is the hardest part to get right as AI glasses makers will need a lens that is thinner, lighter, and more power-efficient than what exists today.”The co-founders said LetinAR wants to be the company those glasses makers call. The company calls its technology PinTILT: a way of arranging tiny optical elements inside a lens so that light is directed precisely where it needs to go, into the user’s eye, rather than scattered in every direction.Think of a TV. It broadcasts light across an entire room, but only the light that actually reaches your eyes matters. Most existing smart lens technologies, particularly a dominant approach calledwaveguide, work a bit like that TV, splitting and spreading light across the full lens to create a wide image. The result is a thin lens, but an inefficient one. A lot of light gets thrown away before it ever reaches the eye, which means dimmer images and, critically, a battery that drains fast, Ha explained.The alternative, a mirror-based approach known asbirdbath, delivers light more directly to the eye, but the structure is bulky, making it nearly impossible to fit inside something that looks like a normal pair of glasses.PinTILT sidesteps that trade-off, Ha said. By focusing only on the light that can actually enter the eye and carefully engineering the angle of each tiny element inside the lens, LetinAR claims it can produce a brighter image in a thinner, lighter form factor, using less power. In a category where every gram and every hour of battery life matters, that’s the problem the entire industry has been trying to solve.In the space, there are a number of peers likeWaveOptics,DigiLens, andLumus.CustomersIts modules are already shipping. LetinAR counts Japan’s NTT QONOQ Devices and Dynabook, formerly known as Toshiba Client Solutions, among its customers, giving the company real manufacturing experience at scale. It is in talks with Big Tech companies on R&D of next-generation AI glasses, though it declined to name them.One of LetinAR’s most demanding customers isAegis Rider, a Swiss deep tech company spun out of ETH Zurich’s Computer Vision Lab. Aegis Rider is building an AI-powered AR helmet that displays navigation, speed, and safety alerts directly in a motorcycle rider’s field of vision, not floating on the visor but anchored to the road itself, as if the information is physically painted on the world ahead.LetinAR’s mod

Grafana Labs, the maker of its eponymous popular open source web visualization software, confirmed it had been hacked but that it refused to pay the hackers who had threatened to release the company’s codebase.Ina series of postson social media, the lab said its investigation found that the hackers had abused a stolen token credential that allowed access to the company’s GitHub environment, which it uses for storing its source code, but the token did not allow access to customer records or financial data. The company has since invalidated the token and added additional security measures to prevent a repeat incident.“The attacker attempted to blackmail us, demanding payment to prevent the release of our codebase,” the company said.Grafana’s code is open source and public, meaning anyone can download the software and edit its code before running it on their own machines. It’s unclear if the hackers stole any proprietary code or information. A spokesperson for the company did not immediately return a request for comment.The incident contrasts with the recent hack at education tech giant Instructure, which last week“reached an agreement” to pay the hackerswho had compromised its network twice in recent weeks. The hackers had demanded an unspecified ransom, threatening to release stolen data about staff and students who use its softwarefollowing a massive data breach and a subsequent website defacement.While in Grafana’s case, no customer data was taken, the company cited the FBI’s long-standing advice urging victims not to pay hackers, as cooperating with them does not guarantee they will return stolen data or refrain from publishing it later. Critics also say paying cybercriminals helps to fund future cyberattacks.Grafana said its investigation was ongoing and will share its findings once its probe concludes.This story was updated to correct that the hackers compromised access to Grafana’s GitHub environment.Topicscyberattack,cybersecurity,data breach,extortion,grafana,open source,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Amazonannouncedthe latest update to Alexa+ on Monday: the ability to generate podcast episodes on demand.The new feature, called “Alexa Podcasts,” is rolling out to customers in the U.S. today. Amazon describes the capability as a way to “turn any topic you’re curious about into a podcast episode, ready in minutes.”To use the feature, all users have to do is ask Alexa+ to create a podcast about a topic they’re interested in. Users don’t need to upload documents, write scripts, or plan anything ahead of time. Instead, Alexa+ researches the request, gathers information, and generates a quick overview of what the episode will cover. From there, users can tweak things like the length, tone, and focus of the episode.Once finalized, Alexa+ uses AI-generated host voices to narrate the podcast. When the episode is ready, users get a notification through their Echo Show device and inside the Alexa app. Episodes are also saved in the app’s “Music” and “More” sections so they can be replayed later.The feature is another example of how Amazon is trying to turn Alexa+ into more than just a voice assistant. Instead of only answering questions or controlling smart home devices, Alexa+ is starting to act more like a personalized AI content creator.At the same time, the launch is likely to spark some debate. AI-generated voices and automated content continue to raise questions around ethics, accuracy, and the future of traditional creators. There are also concerns about how reliable AI-generated podcasts will be, especially when covering news or complex topics.Amazon emphasized its partnerships with major news organizations to improve content accuracy and reliability. The company says Alexa+ can access real-time information through agreements with outlets, including the Associated Press, Reuters, The Washington Post, Time, Forbes, Business Insider, Politico, USA Today, Condé Nast, Hearst, and Vox Media, alongside more than 200 local newspapers across the U.S.Beyond podcasts, Amazon says it is exploring additional forms of personalized AI audio, including custom news briefings and content generated from users’ own documents and shared information.TopicsAI,alexa plus,alexa+,Amazon,Generative AI,Media & Entertainment,podcastsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

The market for AI notetaking devices has exploded in the U.S., with the category generating over $600 million in revenue last year, according toa Menlo Ventures report. And as startups likeHeidi HealthandFreedhave shown, there’s decent demand for this tech in healthcare, where doctors and clinics see the potential for an AI assistant that can help them keep track of patient conversations, surface health records, and lower their administrative burdens.But those apps don’t do much for patients, which is whyKin Healthis building a notetaker that can transcribe your visits to doctors, parse medical advice, and surface next steps when required. To that end, the startup has raised $9 million in a seed funding round led by Maveron.The app is similar to a meeting notetaker: You can record doctor visits, and it will return an AI summary of the meeting, with the next steps, all of which you share with family and friends if you want to. It also lets you write down questions that you might want to ask during your next visit.Loading the player…Kin Health says it encrypts all patient data and that summaries are kept private by default. The tool is not HIPAA-certified, as it is a patient-facing one, but it adheres to the same privacy standards, the company said.The free app is built by physicians Arpan and Amit Parikh, along with Kyle Alwyn, who previously built online prescription serviceHeyDoctor and sold it to health platform GoodRx. Doug Hirsch and Trevor Bezdek, co-founders of GoodRx, are founding partners and executive chairmen at the company.Co-Founders Arpan Parikh, Amit Parikh, and Kyle Alwyn.Image Credits:Kin Health“We have a lot of these storage cabinets where our health data can live, but we don’t have a way to convert that into a utility that we can use to drive our behavioral change. Our goal is to create this health graph where we can store your information from multiple different sources,” Alwyn told TechCrunch over a call.Kin Health says that its summaries are provided after a few stages of processing. After transcribing the visit, an algorithm turns the transcription into a clinical narrative, which gets crunched into a user-facing summary with action items. The company says it is leaning on specialized medical models to power the transcription and that it evaluates and observes outputs at different stages to ensure answers are accurate.But AI in healthcare is being received with a measure of caution and apprehension. Privacy experts and researchers have raisedconcernsover data security, accuracy of AI, consent mechanisms, thequality of generated notes, and theireffectiveness.AI notetakers also oftenfail to recognize and struggle to transcribe regional accents. Kin Health says it is working to ensure its tool works with different accents, as well as when someone has a bad throat or is wearing a mask.Dr. Rebecca Mishuris, chief health information officer and VP at Mass General Brigham, a healthcare organization in Boston, argues it is important for physicians to review any notes generated by AI.“Generative AI will hallucinate; that is the nature of a technology built on patterns and prediction. That is why it is so important for clinicians to review the drafted notes before signing them. At the end of the day, the responsibility for the documentation falls to the clinician,” she told TechCrunch over email.Kin Health currently only shows notes from conversations it records during consultations, but the company said it plans to bring in data from other health sources, including physicians’ own notes through electronic health record (EHR) systems, this year.The company says it will keep the app free of cost forever, and monetize via referrals to services such as specialists and labs. The startup is taking a leaf from GoodRx’s playbook, which also keeps the core product free and earns commissions by referring other services.Natalie Dillon, a partner at Maveron, said healthcare provider-side tools often expect patients to coordinate their own treatment actions. “Kin is built to solve an entirely different consumer need: it can travel with them between specialists, systems, and providers. It’s not beholden to any single health network or EHR relationship. It’s built to serve the patient, not the institution, and that’s a massive distribution advantage,” she said.The funding round also saw participation from Town Hall Ventures, Eniac Ventures, Flex Capital, Foundry Square Capital, Pear VC, and The Family Fund. GoodRx’s Hirsch and Bezdek; angel investors Jay Desai, Nabeel Quryshi, Alex Cohen, and Saharsh Patel; and more than 30 physicians also invested.Topicsai scribe,Apps,Health tech,Kin Health,maveronWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Ivan MehtaIvan covers global consumer tech developments at TechCrunch. He is based out of India and has previously worked at publications including Huffington Post and The Next Web.You can

Elon Musk’s claim that he was mistreated by his OpenAI co-founders failed after nine California jurors returned a unanimous verdict that his lawsuits had been filed too late.Musk accused Sam Altman, Greg Brockman, OpenAI, and Microsoft of “stealing a charity” by creating a for-profit affiliate of the frontier AI lab. Jurors, however, found that any harms that Musk may have suffered came before the deadline for filing his claims under the law.While the trial delved deeply into the melodramatic history of OpenAI and featured testimony from leading figures in Silicon Valley, it ultimately turned onfairly narrow questionsof the law. The trial focused on whether and when Altman and the other defendants had made and broken promises to Musk, but his case failed to convince jurors that he had a valid claim.In particular, OpenAI had advanceda statute of limitations defense, which sought to prove that any harms Musk sought to litigate had taken place before 2021. (The specific date varied by the charge: before August 5, 2021, for the first count; August 5, 2022, for the second count; and November 14, 2021, for the third count.) Ultimately, the jury found that argument persuasive, which made for a short deliberation period.“There was a substantial amount of evidence to support the jury’s finding, which is why I was prepared to dismiss on the spot,” Judge Yvonne Gonzalez Rogers said after the verdict was delivered.The end of the case means that one major threat to OpenAI — a possible restructuring — is now off the table ahead of its reported IPO.“It did not take [the jury] two hours to conclude … that Mr. Musk’s lawsuit is nothing more than an after-the-fact contrivance that bears no relationship to reality,” OpenAI’s lead attorney, Bill Savitt, said after the verdict. “They kicked it exactly where it belongs — just to the side. This lawsuit is a hypocritical attempt to sabotage a competitor.”Microsoft, which Musk sued for aiding and abetting OpenAI’s alleged breach of charitable trust, welcomed the verdict. A spokesperson for the company said it “remained committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”The verdict came in the middle of a hearing to determine the potential damages to Musk if the verdict had gone the other way. While that discussion is moot for now, the judge appeared unconvinced by the analogy Musk’s lawyers drew between his charitable contributions and investments in a for-profit startup.“Your analysis seems to be devoid of connection to the underlying facts,” she told Dr. C. Paul Wazzan, the expert who came up with Musk’s estimate of OpenAI and Microsoft’s wrongful gains at his expense — some $78.8 billion to $135 billion.In a tweet after the ruling, Musk appeared to take the procedural grounds of the dismissal as a moral victory. “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!” Musk wrote. “I will be filing an appeal with the Ninth Circuit, because creating a precedent to loot charities is incredibly destructive to charitable giving in America.”Reached for comment by TechCrunch, Musk’s lead counsel, Marc Toberoff, said, “One word: Appeal.”TopicsAI,Elon Musk,greg brockman,Microsoft,OpenAI trial,sam altmanWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Tim FernholzSenior ReporterTim Fernholz is a journalist who writes about technology, finance and public policy. He has closely covered the rise of the private space industry and is the author ofRocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race.Formerly, he was a senior reporter at Quartz, the global business news site, for more than a decade, and began his career as a political reporter in Washington, D.C.

New York public health provider NYC Health + Hospitals saysa months-long data breachthat allowed hackers to steal personal data, medical records, and fingerprints scans affects at least 1.8 million people.NYCHHC is the largest public health system in the United States and provides healthcare toover a million New Yorkers, the majority of whom are uninsured or receive state healthcare benefits, such as Medicaid.The healthcare system reported the number to the U.S. Department of Health and Human Services, making it one of the largest healthcare-related data breaches of the year so far. Healthcare organizations have been repeatedly targeted by financially motivated cybercriminals in recent years in efforts to steal their vast banks of highly sensitive patients’ personal, medical, and billing information.In a data breach notice on its website, NYCHHC said that it detected a cyberattack on February 2 and secured its network. The hackers had access to its network from November 2025 until February 2026, during which the hackers copied files from its systems.The healthcare system said hackers broke due to a breach at a third-party vendor, which it did not name.NYCHHC said that the exposed data varies by individual and includes patients’ health insurance plan and policy information, medical information (e.g., diagnoses, medications, tests, and imagery), billing, claims, and payment information. Other government-issued identity documents, such as Social Security numbers, passports, and driver’s licenses, were also compromised.The breach notice also says “precise geolocation data” was taken in the breach, suggesting that the user-uploaded photos of their identity documents may have also contained the exact location of where the document was captured.The breach is particularly sensitive because hackers stole biometric information, including fingerprints and palm prints, which affected individuals have for life and cannot replace. NYCHHC did not provide an explanation for storing biometric data. Prospective NYCHHC employees are generally required to enroll their fingerprints for criminal records checks. It’s not yet known if patients’ biometrics were also taken.NYCHHC’s website was briefly offline as of Monday morning. A spokesperson for NYCHHC did not immediately respond to an email from TechCrunch with questions about the cyberattack. TechCrunch asked, among other things, why it took the organization months to detect the breach, and if it has received any communication from the hackers, such as a demand for payment.It’s not clear if NYCHHC can receive email at the time of the website outage.The incident appears to be unrelated to the data breach at National Association on Drug Abuse Problems (NADAP)earlier this year, in which over 5,000 NYCHHC patients had information taken in the cyberattack.In the FBI’s latestannual report on cybercrimecovering 2025, healthcare remained a top target for ransomware attackers — criminals who break into databases, steal a copy of the data while scrambling the victim’s servers, and threaten to publish the stolen data if the victim does not pay the hackers. A ransomware attack on UnitedHealth-owned health tech giant Change Healthcare allowed Russian-linked hackers tosteal the medical and billing information of more than 190 million Americans, believed to be the largest theft of U.S. medical data in history.Topicscyberattack,cybersecurity,data breach,healthcare,new york city,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Anthropic announced Monday it has acquired Stainless, a startup founded by former Stripe engineer Alex Rattray whose software is widely used by rival AI labs, including OpenAI and Google.Anthropic didn’t disclose terms of the deal. However, The Informationreportedlast week that the company was in talks to acquire Stainless, which is backed by Sequoia Capital and Andreessen Horowitz, for more than $300 million.The acquisition will take a key infrastructure supplier out of the hands of Anthropic’s competitors. The company told TechCrunch it will wind downall hosted Stainless products, including its SDK generator. An Anthropic spokesperson said Stainless customers will still own the SDKs they’ve generated to date and have full rights to modify and extend them however they wish.The New York-based startup, founded in 2022, rose to prominence in the emerging AI industry for automating the creation and maintenance of software development kits, or SDKs — the libraries developers use to interact with APIs.Rattraydeveloped softwarethat could take API specifications and turn them into production-ready SDKs across multiple programming languages, including Python, TypeScript, Kotlin, Go, and Java. It became a popular tool because the platform automatically updates the SDKs as APIs change and eliminated the time-consuming process of manually maintaining them.The technology is particularly valuable tocompanieslike Anthropic, OpenAI, Google, Replicate, Runway, and Cloudflare that are building AI agents that can connect to external software and complete tasks on behalf of users. Stainless’s SDK tools are an easy way to build and maintain those connections — but going forward, the tools will only be available to Anthropic, not its competitors.According to Anthropic, Stainless software has powered the generation of every official Anthropic SDK since the earliest days of its API.“I started Stainless because SDKs deserve as much care as the APIs they wrap,” Rattray saidin a press releaseposted Monday. “Anthropic was one of the first teams to bet on this with us. We have been watching what developers have built on Claude over the last few years, which made bringing our teams together an easy decision. The team gets to keep doing the work we love, on the platform where it matters most.”TopicsAI,Anthropic,StainlessWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Kirsten KorosecTransportation EditorKirsten Korosec is a reporter and editor who has covered the future of transportation from EVs and autonomous vehicles to urban air mobility and in-car tech for more than a decade. She is currently the transportation editor at TechCrunch and co-host of TechCrunch’s Equity podcast. She is also co-founder and co-host of the podcast, “The Autonocast.” She previously wrote for Fortune, The Verge, Bloomberg, MIT Technology Review and CBS Interactive.You can contact or verify outreach from Kirsten by [email protected] via encrypted message at kkorosec.07 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Drug discovery is one of the most expensive pursuits in modern industry. Finding a single viable molecule can take a decade and cost billions, and most candidates still don’t make it. A generation of AI startups has promised to fix that — most have made the problem less painful for researchers, who are already technically sophisticated enough to use the tools.But SandboxAQ thinks the bottleneck isn’t the models. It’s the interface.The company has teamed up with Anthropic to integrate its scientific AI models directly into Claude — putting powerful drug discovery and materials science tools behind a conversational interface that requires no specialized computing infrastructure to use.Founded roughly five years ago as an Alphabet spinout, SandboxAQ counts Eric Schmidt, Google’s former CEO, as its chairman. The company, which has raisedmore than $950 million from investors, has built out a number of different business lines, includinga cybersecurity business.One of the more unique things SandboxAQ does, however, is produce large quantitative models, or LQMs. These proprietary models are “physics-grounded,” meaning they’re built on the rules of the physical world rather than patterns in text. They can run quantum chemistry calculations and simulate both molecular dynamics and microkinetics, the study of how chemical reactions unfold at the molecular level. That matters because it tells researchers how candidate molecules are likely to behave before anyone sets foot in a lab.“Trained on real-world lab data and scientific equations, LQMs are AI models engineered for the quantitative economy, a $50+ trillion sector spanning biopharma, financial services, energy, and advanced materials,” the company said in a news release that strongly suggests Sandbox AQ isn’t building another chatbot or code assistant — it’s chasing the economy that AI is supposed to transform.Chai DiscoveryandIsomorphic Labs— both well-funded bets on better models — have focused on the science. SandboxAQ is focused on who can actually use it.“For the first time, we have a frontier [quantitative] model on a frontier LLM that someone can access in natural language,” Nadia Harhen, SandboxAQ’s general manager of AI simulation, told TechCrunch. Previously, users of SandboxAQ’s LQMs would have had to provide their own digital infrastructure to run the models.SandboxAQ’s customers tend to be computational scientists, research scientists, or experimentalists. Generally, these people work at large pharmaceutical or industrial companies and are searching for new materials that can become marketable products.“Our customers come to us because they’ve tried all the other software out there, and the complexity of their problem is such that it didn’t work or didn’t yield positive results for them when that translation went to take place in the real world,” said Harhen.TopicsAI,Anthropic,Claude,Eric Schmidt,Google,SandboxAQWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lucas RopekSenior Writer, TechCrunchLucas is a senior writer at TechCrunch, where he covers artificial intelligence, consumer tech, and startups. He previously covered AI and cybersecurity at Gizmodo.

A worker died at SpaceX’s Starbase launch site in South Texas on Friday, and the Occupational Safety and Health Administration (OSHA) has opened an investigation.TheSan Antonio Express-News reported Mondaythat the unidentified victim died at around 4:17 a.m. local time on May 15, citing OSHA and local officials. The Wall Street Journal later reported that the county sheriffconfirmed to the outletthat a worker died. OSHA confirmed to TechCrunch that it is investigating the apparent accident.Representatives for the nearby Brownsville police and fire departments did not respond to requests for comment. SpaceX and the newly incorporated city of Starbase did not respond to requests for comment.The circumstances of the worker’s death are not immediately clear. OSHA told TechCrunch that it won’t release more information until its investigation is complete, which could take months.The death comes just a few days ahead of the first planned launch of SpaceX’s upgraded Starship rocket. Elon Musk’s spaceflight company is also reportedly releasing the detailed prospectus for its initial public offering this week, which is expected to be the biggest ever when that transaction takes place next month.SpaceX has long dealt with worker safety problems at its Starbase site, which handles Starship prototype launches and is an active construction zone.In 2025, TechCrunchanalyzed OSHA dataand determined the Texas launch site had an injury rate that far outpaced those of industry rivals and was the most dangerous of SpaceX’s worksites. A 2023Reuters investigationuncovered dozens of previously unreported injuries and a worker death in 2014 at SpaceX’s McGregor, Texas, test site.In January, OSHA hit SpaceX with seven “serious” safety violations for, among other things, not properly inspecting a crane before it collapsed at Starbase last June. The safety agency dealt SpaceX the maximum financial penalty on six of those seven violations, totaling $115,850. SpaceX is contesting those penalties, federal records show.The company has been hit with multiple lawsuits related to injuries sustained at Starbase in recent years. In December, an employee of one of SpaceX’s subcontractors sued after he was crushed by a large metal support dropped from a crane. The worker, Eduardo Cavazos, suffered a broken hip, knee, and tibia, and OSHA opened a “rapid response investigation,” as TechCrunchfirst reportedin December.OSHA has since closed that rapid response investigation without taking any punitive action, according to a TechCrunch public records request. And the lawsuit was recently dropped because his employer, the subcontractor, provides workers compensation insurance that prevents the company from being sued, according to Cavazos’ attorney.TopicsOSHA,Space,SpaceX,starbase,TransportationWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sean O’KaneSr. Reporter, TransportationSean O’Kane is a reporter who has spent a decade covering the rapidly-evolving business and technology of the transportation industry, including Tesla and the many startups chasing Elon Musk. Most recently, he was a reporter at Bloomberg News where he helped break stories about some of the most notorious EV SPAC flops. He previously worked at The Verge, where he also covered consumer technology, hosted many short- and long-form videos, performed product and editorial photography, and once nearly passed out in a Red Bull Air Race plane.You can contact or verify outreach from Sean by [email protected] via encrypted message at okane.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Most members of Stanford’s class of 2026 are smart, ambitious, and poised for remarkable careers. Theo Baker already has one. In his first semester of college, Baker broke the story that forced Stanford president Marc Tessier-Lavigne to resign — work that earned him a George Polk Award, one of journalism’s highest honors. Warner Brothers and producer Amy Pascal have optioned the rights to that story. And Tuesday, with graduation less than a month away, Baker publishesHow to Rule the World, a sweeping account of his time at Stanford and the school’s often insidious relationship with the venture capital industry. Judging by early interest, it has every chance of becoming a bestseller.We’ve been anticipating this one (we shared somerelated thoughtsabout it just a few weeks ago). We talked with Baker last Friday. This interview has been edited for length and clarity.You showed up at Stanford as a coder. How did you end up breaking one of the biggest stories in the university’s history before your freshman year was even over?I arrived thinking tech and entrepreneurship was the path for me. I joined the student hackathon, Tree Hacks, helped run it, skipped ahead to the CS weeder class. But my grandfather, with whom I was very close, had passed away a few weeks before I arrived, and he talked about working on the student paper more than anyone I’d ever known. So I joined the student paper to feel connected to him — it was supposed to be a hobby, a way to meet people and explore campus.Very quickly things spiraled from there. My first few stories got more reception than we’d imagined, tips started flooding in, and one led me to a pseudonymous website called PubPeer, where scientists dissect published research. There were comments, seven years old at the time, suspecting that papers co-authored by Stanford’s president, Marc Tessier-Lavigne, had images that were duplicated, spliced, or otherwise irregular. I was a month into my time at Stanford when that investigation began, and by the time I was back for sophomore year, the president had resigned.Were you warned off the story?Multiple times, before I’d even published my first article. People warned me that Tessier-Lavigne was a person of very high integrity with a sterling reputation — that I didn’t want to do this, that it was going to place me in a very uncomfortable position within the institution. Which, of course, was not wrong. Over the course of the next 10 months, as the story widened, the pushback grew steeper. Within 24 hours of my first story, the board of trustees announced their own investigation. I quickly learned that one of the board members overseeing it had an $18 million investment in Denali Therapeutics, the biotech company Tessier-Lavigne co-founded. And the statement announcing the investigation praised his “integrity and honor”— in an investigation that was theoretically looking into his scientific integrity. So the investigation itself became an object of reporting. Tessier-Lavigne never once directly responded to a request for comment during my freshman year. Eventually he began sending missives to all of the faculty — which included all of my professors — describing my reporting as “breathtakingly outrageous and replete with falsehoods.” And then I began hearing more from his lawyers.The book is really about something broader, though — what you call the Stanford inside Stanford. What does that mean?Very soon after I arrived, I realized there was this parallel reality — an inside world — where the kids identified early as the next trillion-dollar startup founders are plucked from the crowd and placed into a world of access and resources. Yacht parties, slush funds, everyone texting the same billionaires for advice on weekends. As Stanford has become more famous as the home of great startups, it has become, according to some people at the university, increasingly difficult to spot actual talent. So many people arrive thinking they can be the next billion-dollar dropout that there’s an entire system of hangers-on whose job is to separate what they call the “wantrepreneurs” — people doing it because it looks good — from the so-called builders who actually have potential. It’s a system designed to sniff out the teenagers you can make a buck off of as early as possible.The title of the book, it turns out, isn’t just a metaphor.No. It’s literally the name of a so-called secret class at Stanford, taught by a Silicon Valley CEO. It’s not really a class. It’s more like a Skull and Bones for the aspiring tech elite. People aren’t getting course credit, but there are lectures, discussions, guest speakers, held once a week in the winter quarter on campus. When I arrived, it was a status symbol even to know it existed — that made you “rule-adjacent,” as one person told me. What this guy Justin was trying to do — as the students in the class told me — was what everyone seems to be trying to do: get in and network with the teenagers who can be useful to yo

Bad beatsThe US is betting on AI to catch insider trading in prediction marketsThe Commodity Futures Trading Commission wants us to know it’s taking this very seriously.Kate Knibbs, wired.com–May 16, 2026 7:00 am|113Credit:Yuki Shimazu / FlickrCredit:Yuki Shimazu / FlickrText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navFor most of the past year, it looked likeprediction marketshad kicked off a new golden age of fraud. OnPolymarket, traders raked in fortunes fromsuspiciously timed betson geopolitical events like the raid on Venezuela and the Iran War. It wasn’t clear whether the US government would bother pursuing some of the most flagrant bad actors, since Polymarket’s crypto-based platform was technically offshore and not regulated or licensed within the country.Now, however, the Commodity Futures Trading Commission, which oversees prediction markets, wants you to know that it’s watching very, very closely. The agency is searching for suspicious behavior from traders within the United States who have been sneaking onto offshore markets, including Polymarket’s crypto platform—which is blocked stateside—by using virtual private networks. “We’re going to find them, and we’re going to bring actions,” agency chairman Michael Selig told WIRED this week, speaking from the CFTC’s headquarters in Washington, DC.Selig says the agency, which is especially lean right now, is staffing up. Like so many other AI-pilled workplaces, the CFTC is also leaning into automation to handle the growing workload, including tools that analyze trading patterns and flag potential manipulation. “You’ve got so much data,” Selig says. “When we feed it into AI, we get really great information. It can help us understand things, like where we might want to investigate, or when we might need to send a subpoena to a trader.”In addition to proprietary surveillance systems developed in-house, the agency’s arsenal includes third-partyblockchain tracing toolslike Chainalysis for crypto platforms, and market abuse detection software including Nasdaq Smarts for centralized markets. (Beyond Nasdaq Smarts, the agency did not specify which AI tools it uses and declined to share more specific examples.)Prominent prediction market companies have recently started touting all the work they’re doing to catch sketchy bettors. US-based exchange Kalshi, Polymarket’s primary competitor, eagerly announced that it hassuspendedand penalized customersflagged for insider tradingand market manipulation.In April, after significant backlash over suspected insider trading, Polymarket announced its own partnership with Chainalysis. It was part of a broader push to crack down on market manipulation. While the company’s CEO, Shayne Coplan,had talkedin the past about why insider trading could be good for prediction markets, Polymarket changed its approach this spring, updating its market integrity rules and announcing a partnership with Palantir for its US-based sports markets (the Chainalysis deal focuses on the offshore platform). The company did not respond to WIRED’s request for comments for this story.According to Chainalysis spokesperson Maddie Kenney, the company analyzes the same data for both clients. “The value Chainalysis adds for our customers, including Polymarket and the CFTC, is organizing the data and enriching it with the attributions and insights we’ve accumulated over years in the space,” she says. Certainly sounds like a good deal for Chainalysis!The CFTC’s assurances that it is hunting insiders comes at a moment of intense scrutiny on prediction markets. In March, Connecticut senator Chris MurphytoldWIRED that he suspected White House staffers were engaged in insider trading on war-related contracts. At the beginning of April, seven members of Congressaskedthe CFTC to investigate overseas markets offering war-themed events contracts. In a letter, the lawmakers argued that the commission had the authority and responsibility to curb insider trading, especially on “morally obscene” trades on military action. Selig recentlytoldCongress that the company is pursuing “hundreds, if not thousands” of insider trading tips.Investigations are not limited to federally regulated exchanges. “We’re surveilling the markets on a global basis,” he tells WIRED.Selig says that the agency will exert extraterritorial jurisdiction—its legal ability to enforce its laws beyond traditional boundaries—when it finds suspicious activity on offshore platforms like Polymarket, though he says it’s a case-by-case approach. “We use it in extreme circumstances,” he says, with an eye towards whether charges have a strong chance of sticking in court. “In any extraterritorial litigation, there’s going to be challenges to our authority, and that could also impair our ability to bring cases in the future.” According to Selig, the 2010 Dodd-Frank Act allows the CFTC more leeway to pursue this kind of enforcemen

striking resultsA revolutionary cancer treatment could transform autoimmune diseaseResearchers are testing CAR T cell therapy as a way to reset the immune system.Amber Dance, Knowable Magazine–May 17, 2026 7:00 am|67Artist’s rendering of a T cell.Credit:Getty ImagesArtist’s rendering of a T cell.Credit:Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAt age 49, Jan Janisch-Hanzlik’s multiple sclerosis was destroying her freedom to live the life she wanted. She gave up her active nursing job for a desk role. Frequent falls made her afraid to carry her grandchildren. She had to move to a bigger house to make room for the wheelchair she feared she might end up needing full-time.Even the best available medication wasn’t improving Janisch-Hanzlik’s symptoms, and she worried they’d only get worse. So when she learned about a trial of CAR T cell therapy at the University of Nebraska Medical Center in Omaha, close to the city of Blair where she lives, she phoned the clinic every other month until they were ready toenroll her as the first patient.Originally designed to target and wipe out cancer byreprogramming the patient’s immune cells, CAR T is now being offered to patientsin hundreds of clinical trialsfor autoimmune conditions like multiple sclerosis, lupus, Graves’ disease, vasculitis, and many others. The hope is that CAR T can duplicate the success it has demonstrated in a range of blood cancers by hunting down and eliminating cells that target the self in autoimmune diseases. This would essentially reset the body’s defenses to a state like the one that existed before the disease took hold.But along with CAR T’s promise come risks, questions, and challenges. There’s uncertainty about how well it will work forautoimmunityand how long any benefits might last, as well as what long-term side effects might arise. Janisch-Hanzlik knew this when she sat down to receive the experimental treatment on June 9, 2025; she felt a mix of hope and fear knowing that she would be spending the next week being monitored for side effects including dangerous inflammation.In addition to her clinical expertise and desire to pioneer a new treatment, Janisch-Hanzlik’s two young grandchildren helped inspire her pursuit of a treatment with known risks and uncertain benefits. Because multiple sclerosis has a genetic component, Janisch-Hanzlik knew that they have an elevated chance of going through the same struggle she has. “I would want to be able to say I did everything that I possibly could to prevent them, or anyone else, from having something like this,” she says.From cancer to autoimmunityThe first CAR T cancer treatment wasapproved by the Food and Drug Administration in 2017for an aggressive form of leukemia. Since then, the powerful and intensive treatment hasresulted in long-term remissionfor many cancer patients.The basic premise of CAR T is to activate the power of key immune cells calledT cells. T cells normallyrecognize other cellsthat have been infected by a virus or bacterium, or are otherwise abnormal, and either destroy them or recruit other parts of the immune system to do so.In CAR T for cancer, scientists engineer those T cells to specifically hunt and destroy malignant cells. The technologygot its startwhen cancer researchers figured out how to take out a patient’s own T cells, insert DNA instructions for a “chimeric antigen receptor,” or CAR, and put them back into the person’s circulation. The CAR, which sits on the T cell’s surface and latches on to a specific molecular partner on the surface of cancerous cells, activates the T cell to attack.Today CAR T cells are most commonly programmed to attack B cells, another key immune player. B cells are normally responsiblefor making antibodies, but in certain blood cancers, they proliferate out of control. By giving T cells a CAR that recognizes one of a couple of molecules unique to the B cell surface, the cells are reprogrammed to find and eliminate those cancerous cells.B cells also are the central problem in many autoimmune conditions: They mistakenly make antibodies against normal tissues instead of against invading pathogens. So as CAR T began to succeed against B cell cancers, it didn’t take long for doctors to reason that CAR T therapy might also be able to wipe out bad B cells in people with autoimmunity.There are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineThere are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineA German team pioneered autoimmune CAR T in a woman with lupus,reporting positive resultsin 2021.

swarmsBug bounty businesses bombarded with AI slop“Never-ending” AI slop strains corporate hacking reward schemes.Jamie John, Financial Times–May 18, 2026 9:23 am|81Credit:

      AdobeCredit:

      AdobeText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navCompanies that pay hackers to find flaws in their software are being inundated with low-quality reports generated by AI, forcing some to suspend the programs altogether.Businesses that run “bug bounty” schemes have long relied on independent security researchers to spot vulnerabilities. But the rise of AI tools is now overwhelming them with spurious submissions.Bugcrowd, whose customers include OpenAI, T-Mobile, and Motorola, said the number of reports it received more than quadrupled over a three-week period in March, with most proving to be false.Curl, a widely used tool to transfer data across the Internet, suspended its paid bug bounty program in January, citing an “explosion in AI slop reports” and lower-quality submissions.Cyber security experts say advances in generative AI are reshaping the economics of bug bounty programs. While the tools allow experienced researchers to find flaws more quickly, they are also lowering the barrier to entry, triggering a flood of automated or erroneous submissions that companies must sift through.The big increase in poor-quality AI reports was “quickly becoming a major problem,” said Ross McKerchar, chief information security officer at cyber security group Sophos. “Bug bounties are going to stay [but] they’re going to have to change,” he said.Bug bounties have grown in popularity since the early 2000s, with schemes offering six-figure payouts for the biggest discoveries. Google’s program disbursed a total of $17 million last year, up from $7.5 million in 2021. It paid its largest individual reward of $605,000 in 2022 to a user who spotted a vulnerability in its Android mobile operating system.McKerchar said the rise in poor-quality submissions came from both amateurs trying to find bugs for the first time and existing researchers who were “sometimes getting led on by the [AI] agents.”He added there was a “third cohort” of “experienced AI builders” who had developed automated “end-to-end scanning and submission systems” that were “creating absolute carnage.”Curl’s creator, Daniel Stenberg, wrote in a blog post that the “never-ending slop” had taken “a serious mental toll to manage and sometimes also a long time to debunk.”Software group Nextcloud suspended its bug bounty program in April because of the “massive increase of low-quality reports.” It said it hoped to resume the program once it had found a way to filter submissions effectively.The surge in AI-generated reports comes as Anthropic last month launched Mythos, its new cyber AI model, which it says can find software flaws faster than humans.Companies running bug bounty programs have started to introduce more stringent background checks to combat the problem, as well as building AI agents to triage submissions.HackerOne, whose bug-reporting platform serves Goldman Sachs, Google, and the US Department of Defense, said it had “introduced new agentic validation capabilities” this year to “help organizations manage high volumes of findings,” such as those generated by models like Mythos.The company said submissions had jumped 76 percent in the year to March. But it said the share of reports flagging legitimate vulnerabilities had remained steady over the past year at 25 percent.HackerOne chief executive Kara Sprague said it had in recent weeks seen a rise in “higher quality” reports that had used AI. She added that the rise in AI-generated submissions was “not a strong reason to say we don’t want them” altogether, given that hackers were using the technology to spot more flaws.Bugcrowd chief Dave Gerry said developments such as Anthropic’s Mythos would assist human bug bounty hunters, not replace them. “AI is going to help with a lot of things but we’re never going to replace that human creativity,” he said.© 2026 The Financial Times Ltd.All rights reserved. Not to be redistributed, copied, or modified in any way.Financial TimesFinancial Times81 Comments

All roads lead to StarshipThe US space enterprise is desperately waiting for Starship—will it finally deliver?“This is such a wild ride. The highs are high. The lows are low.”Eric Berger–May 18, 2026 7:00 am|172Starship V3 completes a successful wet dress rehearsal on May 11.Credit:

      SpaceXStarship V3 completes a successful wet dress rehearsal on May 11.Credit:

      SpaceXText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThese days, one would be forgiven for forgetting that SpaceX is, at its core, a rocket company.Consider the company’s mega deals over the last year. SpaceX paid $17 billion—more than it has spent developing every one of its rockets—to EchoStar for wireless spectrum to boost its Starlink network. It revealed plans to launch 1 million orbital data centers. SpaceX merged with xAI in a deal that valued Elon Musk’s artificial intelligence firm at $250 billion, and it announced plans to becomea major computer chip manufacturer. And earlier this month, SpaceX sold an enormous amount of ground-based compute to Anthropic.As a result of all this activity, an impending IPO will value the company at something like $1.5 or $2 trillion. That’s trillion, with at.So yes, one might reasonably ask what SpaceX does these days. Because all the buzz, all the Wall Street euphoria, and all the financial frisson are only tangentially related to what SpaceX cut its teeth on during its first 25 years: becoming the globally dominant player in launch. It largely concerns telecommunications and AI data services.And yet everything SpaceX aspires to accomplish in the next quarter of a century, all of its enormous valuation, is predicated on a new launch vehicle. A rocket that, to date, has a decidedly mixed record of success. A rocket that has not flown in seven months. A rocket that, finally, may return to the skies on Wednesday.We are speaking, of course, of Starship—a truly revolutionary rocket. If it works. And after a long period of development and three years of test flights and setbacks, it kind of has to.“Test Like You Fly”A few weeks ago, SpaceX released avisually stunning videothat takes viewers inside its massive new Starfactory in South Texas and provides up-close views of its rockets and engines.The “Test Like You Fly” video also outlines the development of the third iteration of the Starship rocket, V3, highlighting both advances and setbacks encountered along the way. SpaceX is distilled to its core as the exceptional engineers who are designing and testing a radically new rocket talk about the very difficult problems they’re trying to solve.“We are not breaking laws of physics; we are just trying to leverage them as effectively as we can,” explains Jacob McKenzie, the company’s vice president for the Raptor rocket engine that powers both the first and second stage of Starship.McKenzie says SpaceX built 600 Raptor engines as part of its V2 Starship program, a remarkable number that underscores the investment the company is making to develop a rocket that has yet to reach orbit or deliver a single payload. (To put this into perspective,NASA is spending $3.5 billionto procure two dozen comparably powered, expendable rocket engines.)By some estimates, SpaceX has now invested $15 billion in the Starship rocket program over the last decade, funding 11 test flights, a sprawling spaceport in South Texas, a massive new factory there to support a high production rate, and expanding facilities in Florida alongside the company’s existing infrastructure in the state.It has not gone particularly smoothly. The company conducted its initial Starship test flight in April 2023. I’m not sure anyone expected that, three years and nearly a dozen flights later, the company would still be firmly in test mode.But it turns out that building the world’s largest and most powerful rocket and optimizing it for rapid reuse is a difficult, time-consuming task. Even for a company that prides itself on moving fast and trying to achieve the near impossible.A very difficult 2025SpaceX performed six test flights in 2023 and 2024, and by November of the latter year had made substantial progress. The company demonstrated the ability to launch and capture the Super Heavy first stage and completed a safe flight of the upper Starship stage, including re-lighting its Raptor engine in space, before making a controlled splashdown in a precise location.That month, the then-general manager of the Starbase facility in South Texas, Kathy Lueders, said the company aimed to dramatically increase the vehicle’s launch cadence in 2025. “Elon would say, next year, he would love to have us have 25 missions a year,” she saidduring a community event. The company also planned to “capture” a Starship upper stage and conduct an orbital refueling test in that time frame.Debris from Starship falls back into the atmosphere in this view over Hog Cay

The Next GenerationDid Artemis II break through? Registrations at Space Camp double afterward.Isaacman’s $25 million donation leads to impressive new facilities.Eric Berger–May 18, 2026 9:49 am|27Commercial Astronaut on Inspiration4 Chris Sembroski and NASA Administrator Jared Isaacman attend the ribbon-cutting ceremony of the Inspiration4 Skills Training Complex in Huntsville, Ala.Credit:

       Tibrina Hobson/Getty ImagesCommercial Astronaut on Inspiration4 Chris Sembroski and NASA Administrator Jared Isaacman attend the ribbon-cutting ceremony of the Inspiration4 Skills Training Complex in Huntsville, Ala.Credit:

       Tibrina Hobson/Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen he was 12 years old, NASA Administrator Jared Isaacman attended the weeklong “Aviation Challenge” program at Space Camp in Huntsville, Alabama.“For the first time, I got behind the controls of an airplane when I attended Aviation Challenge,” Isaacman said on Friday evening during an event at the US Space & Rocket Center. “I became a pilot because I thought that was the closest I would ever get to the stars.”Decades later, after founding a successful online payments company and flying to space twice as a private citizen aboard SpaceX’s Crew Dragon, Isaacman has returned to Space Camp in Alabama on multiple occasions to meet with participants and share a bit of the awe he experienced as a kid. In 2022, a year after the first of these flights, Inspiration4, Isaacman donated $10 million to kick off a Space Camp expansion.Now, as the leader of NASA’s space program, he has sought to continue to pay it forward. He donates his salary to Space Camp, and on Friday, he returned to open the new “Inspiration4 Skills Training Complex,” a 50,000-square-foot facility completed with an additional $15 million donation from Isaacman. The money will also support development of a new dormitory.Space Camp becomes part of the US cultureSpace Camp has carved out an influential place in American culture. A 1986 feature film,Space Camp, popularized the camp when four teenagers befriended a robot and ended up launching into orbit on a space shuttle. The plot is implausible, to say the least, and the movie was criticized because it came out just a few months after the space shuttleChallengercatastrophe.Nevertheless, the movie helped build the myth of Space Camp, which had only opened four years earlier. Since then, more than 900,000 children ages 9 to 18 years old have “graduated” from the experience. Included among them are at least half a dozen NASA astronauts, including Dottie Metcalf-Lindenburger, Kate Rubins, Serena Auñón-Chancellor, and a mission specialist on the recently flown Artemis II mission, Christina Koch.Some of the exhibits and activities at Space Camp had grown a bit dated, but Isaacman’s gift should help to revitalize the programs.Among the new elements are a parachute “simulation room” that features a virtual-reality 10,000-ft drop with wind simulation, a drone lab, an interactive mission control, and “training simulators” that will recreate the experience of zero gravity.Camp registrations double after Artemis IIThe Artemis mission, flown by Koch and three other astronauts, appears to have reignited interest in the space program. Space Camp officials say registrations have doubled this summer since the successful completion of the lunar flyby mission in April.Interest is expected to grow further in the coming years as NASA returns to the Moon and establishes habitats there. The aim is to meet the needs of space-curious children with a modernized experience.“Off the success of Artemis II, America’s return to the Moon is just getting going,” Isaacman said during the ribbon-cutting event for the new facility.“And it is that kind of magic that inspires the next generation to attend Space Camp, get hands-on experiences at this national treasure, unlike anywhere else in the country, and grow up ready to pick up the baton and join in this great adventure.”Eric BergerSenior Space EditorEric BergerSenior Space EditorEric Berger is the senior space editor at Ars Technica, covering everything from astronomy to private space to NASA policy, and author of two books:Liftoff, about the rise of SpaceX; andReentry, on the development of the Falcon 9 rocket and Dragon. A certified meteorologist, Eric lives in Houston.27 Comments

engagementBMW sends off the 6th-gen M3 CS with a manual gearbox, rear-wheel driveThe 2027 M3 CS Handschalter is lighter and comes with three pedals instead of two.Jonathan M. Gitlin–May 18, 2026 10:33 am|52For the first and last time, you can get a stick shift in a G80 BMW M3 CS.Credit:

      BMWFor the first and last time, you can get a stick shift in a G80 BMW M3 CS.Credit:

      BMWText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe march of time, and what counts for progress in the automotive industry, has not been particularly kind to the driving enthusiast. Our vehicles have gotten bigger and heavier. Touch-sensitive panels and screens replaced buttons. Steering feelevaporatedabout a decade ago. And if you’re a fan of changing your own gears with a stick shift and three pedals,things have been looking bleakfor a while now. Which makes BMW’s send-off for its current sixth-generation M3 so notable.BMW’s M divisionkept the six-speed manual alive for the G80 M3, butonly the normal version. If you wanted the more powerful, much torquier M3 Competition or the track-focused M3 CS (Competition Sport), the only transmission choice was an eight-speed automatic. That automatic happens to be the excellent ZF 8HP gearbox, and for being fast on track, I’d still choose it, because that makes left-foot braking easier.Using paddle shifts might be faster, but I won’t pretend it’s more engaging than coordinating the movement of a gearstick through its gate, timed properly to the action of the clutch—especially if you’re heel-and-toeing, but even if you use the auto-blip feature that revs the engines for you on downshifts now. BMW appears to recognize that too, because it says the 2027 M3 CS Handschalter is designed for maximum driver engagement, and just for North America.Yes, that is a big carbon-fiber hump in the middle of that seat.BMWYes, that is a big carbon-fiber hump in the middle of that seat.BMWNew, lighter wheels.BMWNew, lighter wheels.BMWImola red paint is a $4,500 option.BMWImola red paint is a $4,500 option.BMWNew, lighter wheels.BMWImola red paint is a $4,500 option.BMWThe $107,100 M3 doesn’t use the more powerful engine from the M3 Competition, but the same 473 hp (353 kW) inline six-cylinder engine as the regular six-speed M3. But it’s a lighter car, to the tune of about 75 lbs (34 kg) thanks to things like a titanium exhaust muffler, carbon-fiber seats, carbon ceramic brakes, lighter wheels, and plenty of carbon-fiber reinforced plastic for body panels. And instead of the ZF 8HP, there’s a six-speed manual, which sends its power to the rear wheels and the rear wheels only: There’s no xDrive AWD here, unlike the M3 Competition.BMW M’s next major work will be next year’sall-electric quad-motor M3, derived from theNeue Klasse i3. And we can’t imagine there’s a three-pedal version of that one in the works, sadly.The M3Ireally wantIn BMW’s briefing materials for the M3 CS Handschalter, it notes that the 3.0 L S58 engine is also used, in modified form, in the BMW M4 GT3 Evo that won the Rolex 24 at Daytona this year. BMW has had a fair degree of success with the M4 GT3 Evo, as it also won last year’s Nurbürgring 24, among other races. But the only M race car with an S58 engine anyone cared about at this year’s N24 started life as anApril Fool’s joke.As 24-hour races go, the N24 is certainly unique. Le Mans and Daytona have faster prototypes among the GTs. Spa-Francorchamps is just GT3, but more than 70 of them on track together. But the N24 combines everything from GT3 cars down to Volkswagen Golfs, and does so across not just the modern F1-grade GP Circuit but also the entire stretch of the Nordschleife, with its narrow straights, crests, drops, and dozens of blind corners. Watching driver onboards is frankly terrifying, given the speed differentials.The BMW M3 Touring at the N24.Christian Kaspar-Bartke/Getty ImagesThe BMW M3 Touring at the N24.Christian Kaspar-Bartke/Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesThis year’s race had a pair of stars. Max Verstappen took time off from his F1 job, got his Nürburgring race permit, and then showed that he could absolutely cut it with the world’s best sports car drivers on the w

about timeFive years later, Windows 11 brings back much-missed taskbar options (and more)Microsoft is also testing a smaller taskbar and more customizable Start menu.Andrew Cunningham–May 18, 2026 12:03 pm|131Windows 11’s work-in-progress side-mounted Taskbar.Credit:

      MicrosoftWindows 11's work-in-progress side-mounted Taskbar.Credit:

      MicrosoftText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen Windows 11 launched in 2021, we mostly liked its refreshed look—the rounded corners and menus with just a hint of translucency were a nice change from the flat colors and hard corners of the Windows 8/10 era. But its reformulated taskbar and Start menu came with a number of functional regressions from the versions in Windows 10. Some of these were addressed quickly; others continue to linger.A new Windows Insider Preview build released to testers includes a new wave of improvements that fix longstanding regressions while trying out new things.Most significantly, the Windows 11 taskbar can now be docked to any edge of your screen, including the left and right, something that was possible in Windows 10 (and many older versions of Windows) but has been missing from Windows 11 since launch. Users can configure slightly different taskbar behavior for every taskbar position—if you prefer a different icon alignment or a left/right-mounted taskbar over a top/bottom-mounted taskbar, or if you want different settings for labels and icon groupings, you can choose different options for each position and Windows will remember them.Microsoft says there are several features that haven’t been implemented yet—the taskbar won’t auto-hide in any of the alternate positions, and the “tablet-optimized taskbar” with larger, more finger-friendly icon sizes and spacing also isn’t supported. Touch gestures and the Search box also aren’t supported. All of these features are coming at some point; they just aren’t ready now. Microsoft is “evaluating additional features like different taskbar positions per monitor” for multi-monitor setups.Another change Microsoft is testing makes the taskbar and all its icons smaller, a change designed to increase the amount of vertical space available on smaller screens without requiring users to fully hide the taskbar.The Start menu with recommended and All Apps sections disabled, along with the shorter taskbar.Credit:
                                  MicrosoftThe Start menu with recommended and All Apps sections disabled, along with the shorter taskbar.Credit:

      MicrosoftMicrosoft is also making a handful of changes to the Start menu, including a user-selectable size setting (previously, the menu would increase and decrease in size dynamically based on the size of your display). Each section of the Start menu (the pinned apps section, the “recommended” section, and the “all apps” section) will be individually toggleable. And users will be allowed to hide “recommended” apps advertised from the Microsoft store while still being able to see jump lists and recent files in the File Explorer. Users who decide to keep the “recommended” apps section visible should also benefit from “improv[ed] file relevancy” that “better reflect[s] what you have been working on.”Some of these changes are available in the current Windows Insider Preview builds in the Experimental channel (which replaced the Canary and Dev channels in Microsoft’slatest beta program shake-up). Others will be released “over the coming weeks,” at which point more polished and refined versions will presumably come to the Beta channel and, eventually, the public version of Windows 11. All of these changes are being made as part of Microsoft’s continued “commitment to Windows quality” push, meant to address several of the real and perceived shortcomings of Windows 11 relative to older versions.Andrew CunninghamSenior Technology ReporterAndrew CunninghamSenior Technology ReporterAndrew is a Senior Technology Reporter at Ars Technica, with a focus on consumer tech including computer hardware and in-depth reviews of operating systems like Windows and macOS. Andrew lives in Philadelphia and co-hosts a weekly book podcast calledOverdue.131 Comments

Hands-onThe Dory Sign is E ink, smart screen simplicity at its finestDory CEO claims the $149 signs won’t be bricked should Dory go out of business.Scharon Harding–May 18, 2026 1:26 pm|65I customized the Dory Sign to introduce my dog. The polka dot frame is optional, removable, and costs $29 extra.Credit:

      Scharon HardingI customized the Dory Sign to introduce my dog. The polka dot frame is optional, removable, and costs $29 extra.Credit:

      Scharon HardingText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navMany gadgets marketed as being “smart” make me wonder if they would be better off dumb.Some examples aresmart TVsthat insist on sending your activities to businesses totrack you, smart fridges that use the Internetto cycle through ads,smart gym equipment that won’t work offline, smart toothbrushes whosebatteries drain too quickly, or virtually any gadget that forces you to use a minimally effective or otherwise unimpressive app.Too often, modern technologies, like inter-device connectivity and artificial intelligence, are shoehorned into gadgets that would be more intuitive to use, affordable, accessible, and/or durable without them.Dory Sign, however, is a reminder of how technology can improve something as simple as a sign without overshadowing the product’s most basic purpose, which in this case is effective and delightful communication.Dory is a small sign that, like many E Ink displays, is easy on the eyes because it doesn’t use bright lighting. You control what the sign displays through a free iOS or Android app that doesn’t force you to share your email or name to use it. The app has clear sections for editing the header text, main text, and footer text, adding an image, and choosing a background. It also allows more than one person to make changes to the display and communicates to the sign through Bluetooth.A screenshot of the Dory Sign app. Changes made to text, the image, and background appear in the preview image in real-time.Credit:
                                  Scharon Harding/DoryA screenshot of the Dory Sign app. Changes made to text, the image, and background appear in the preview image in real-time.Credit:

      Scharon Harding/DoryYou can upload your own image or background, but the app comes with a decent range of preloaded images, like animals, flowers, and illustrations, and preloaded backgrounds, including ones that look like brush strokes or marble and more detailed ones, like a landscape. You don’t have to spend a lot of time designing the sign, but if you want to get creative, the app has six different typefaces for text and sliders for text size, line height, letter spacing, and text color (which makes text darker or lighter).The sign can only display one image, though. And while transferring your preferred text, image, and background is simple enough with the app, your phone has to be pretty close to the sign for it to work. You also have to press the power button every time you want to change the display. This can be tedious, especially if you have the sign mounted to a wall. To be fair, though, manually changing a regular sign also requires you to get close to and touch the display.The E Ink screen doesn’t lend itself to vibrant imagery. Instead, it’s a more passive display that doesn’t call for attention, yet clearly communicates with distinct detail in its text and images.The “House party” background by ANTI that’s included in the app.Scharon Harding/ANTIThe “House party” background by ANTI that’s included in the app.Scharon Harding/ANTIThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathDory Sign’s makers claim gadget can last for 15 yearsAt $149, the Dory Sign is expensive for a small display. However, you may be able to extract that much value from it over the years. Dory claims that the sign only uses power when it’s changing what’s on the display.

“are you afraid?”Guy Gardner makes a cameo in newLanternsteaser“This ring? It’s the greatest weapon in the Universe. When and if to use it, that’s the whole ball game.”Jennifer Ouellette–May 18, 2026 1:36 pm|10Credit:

      YouTube/HBO MaxCredit:

      YouTube/HBO MaxText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navLanterns, the new DC Universe series coming to HBO Max, dropped asurprising teaser in Marchthat swapped the usual superhero hijinks for gritty realism more in the vein ofTrue DetectivesandSlow Horses. Personally, I liked the change; it pushed a series that was barely on my radar to my 2026 list of ones to watch. I guess HBO was concerned people would miss the superhero vibes, though, because the latest teaser trailer interweaves the grittier aspects with a lot more superpowers and intergalactic elements. And you know what? I’m still here for it.Per the official logline, “The series follows new recruit John Stewart (Aaron Pierre) and Lantern legend Hal Jordan (Kyle Chandler), two intergalactic cops drawn into a dark, earth-based mystery as they investigate a murder in the American heartland.” There will be two storylines: one set in 2016 about a murder in Nebraska, and the second set in 2026.Chandler’s Hal Jordan is a former test pilot nearing retirement from the Green Lantern Corps. He’s training Pierre’s John Stewart Jr., a new recruit, to replace him. Nathan Fillion reprises hisSupermanrole as the obnoxious Guy Gardner; we get a brief glimpse of him in the new teaser. The cast also includes Kelly MacDonald as Kerry, a small-town family-oriented sheriff; Jason Ritter as Billy Macon, Kerry’s husband; Garret Dillahunt as William Macon, Kerry’s cowboy father-in-law; Poorna Jagannathan as a woman named Zoe; Ulrich Thomsen as Sinestro, a former Corps member who’s gone rogue; and Paul Ben-Victor as an extraterrestrial called Antaan.Sherman Augustus plays John Stewart Sr., with J. Alphonse Nicholson playing the younger version; Nicole Ari Parker plays Bernadette Stewart (mother to John Jr.), with Jasmine Cephas Jones playing the young version of the character. In addition, Chris Coy plays a suspiciously nervous truck driver, Waylon Sanders; Cary Christopher plays a gifted child named Noah; and Laura Linney and Paula Patton will appear in as-yet-undisclosed guest roles.We also know that the show will feature the inter-dimensional prison and the Checkmate organization introduced inPeacemaker’s S2 finale. DC Studios co-CEO James Gunnhas saidthat the events will set up other things within the DCU as a whole. So, is it a gritty buddy-cop series or a superhero series? Showrunner Chris Mundyrecently told Men’s Healththat it’s very much both.“Our show is in a lot of ways about replacement—when should someone step aside and when is it time for the next person to take the reins?”Mundy said. “That push and pull between those two characters is really important. So much of the power that John has is by not taking the bait, understanding that you lose your power if you’re yelling and screaming. That’s what we’re trying to convey: He knows he belongs, so he doesn’t have to overcompensate. There’s a real balance there that’s just innately inside of Aaron. He’s big. He’s an intimidating presence just physically. But there’s a softness to him, too. There’s a thoughtfulness. You can’t teach that.”Lanternspremieres on August 16, 2026, on HBO Max.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.10 Comments

from the garden of fugitivesPompeii victim ID’d as a likely doctorNew X-rays and CT scans showed small case with locking mechanism containing metal instruments.Jennifer Ouellette–May 18, 2026 2:03 pm|12Credit:

      Pompeii Archaeological ParkCredit:

      Pompeii Archaeological ParkText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navArchaeologists used a combination of advanced CT scans and 3D digital reconstruction to identify one of the Pompeii victims who died in 79 CE during the eruption of Mount Vesuvius as most likely having been a Roman doctor, according toan announcementby the Pompeii Archaeological Park.Aspreviously reported, the eruption of Mount Vesuvius released thermal energy roughly equivalent to 100,000 times the atomic bombs dropped on Hiroshima and Nagasaki at the end of World War II, spewing molten rock, pumice, and hot ash over the cities of Pompeii andHerculaneumin particular. The vast majority of people in Pompeii and Herculaneum—the cities hardest hit—perished from asphyxiation, choking on the thick clouds of noxious gas and ash. But at least some of the Vesuvian victimsprobably died instantaneouslyfrom the intense heat of fast-moving pyroclastic flows, with temperatures high enough to boil brains and explode skulls.In the 19th century, an archaeologist named Giuseppe Fiorelli figured out how to make casts of those frozen bodies by pouring liquid plaster into the voids where the soft tissue had been. Some 1,000 bodies have been discovered in the ruins, and 104 plaster casts have been preserved.Restoration effortson 86 of those casts began about 10 years ago, during which researchers took CT scans and X-rays to determine whether complete skeletons were present.The CT scans and X-ray images revealed that there had been a great deal of manipulation of the casts, depending on the aesthetics of the era in which they were made, including altering some features of the bodies’ shapes or adding metal rods to stabilize the cast, as well as frequently removing bones before casting. Ancient DNA analysisin 2024of four victims found in what is known as the “House of the golden bracelet”revealed thatall four bodies were male and none were genetically related—challenging the pre-existing preferred narratives and suggesting those may reflect certain cultural biases.The small case linked to the Roman physician found inside the Pompeii cast.Credit:
                                  Pompeii Archaeological ParkThe small case linked to the Roman physician found inside the Pompeii cast.Credit:

      Pompeii Archaeological ParkThis latest discovery involves one of the plaster cast remains of 14 victims found in what is now known as theGarden of the Fugitives, who died when they were overtaken by the pyroclastic flow while trying to flee via the Nocera Gate. Pompeii superintendent Medeo Maiuri excavated the victims in 1961, originally found in three smaller groups, although today the plaster casts are displayed in a line.Maiuri published a fanciful, largely fictional account of who he believed those bodies to be in National Geographic that same year: a merchant with severe osteoarthritis; a mother and her two young children; a young couple with an infant daughter; and a servant with what appeared to be a shoulder bag, although the “bag” later turned out to just be a malformation in the plaster.For decades, archaeologists had overlooked a small object trapped within another of the plaster casts that had been kept in storage for decades at the archaeological park. The new analysis with X-ray imaging and CT scans revealed that it was a small case with parts made out of metal—specifically, a toothed wheel operating a locking system. The case contained a coin-filled fabric pouch and several metal instruments similar to what a Roman physician (medicus) might have used, leading the team to conclude the man had likely been a doctor, fleeing Pompeii with his instruments close at hand.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.12 Comments

Can I pet that dog?Australian Aboriginals cared for a dingo’s grave for decadesFor some ancient Aboriginal Australian communities, dingoes were part of the family.Kiona N. Smith–May 18, 2026 2:37 pm|44Archaeologists and Barkindji custodians worked together to excavate the dingo burial.Credit:

      Barb QuayleArchaeologists and Barkindji custodians worked together to excavate the dingo burial.Credit:

      Barb QuayleText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA thousand years ago, the ancestors of today’s Barkindji people carefully buried a dingo (or garli, in the Barkindji language) in a mound of shells.Archaeologists recently studied the burial in what’s now New South Wales, Australia. They found that the Barkindji ancestors had buried the dingo with the same care and ceremony as any beloved human member of the community and looked after the grave for centuries. The burial reveals that dingoes were, as Australian Museum and University of Sydney archaeologist and study co-author Amy Way puts it, “deeply valued and loved” by ancient people in Australia.The long-lost dingoFive years ago, Barkindji Elder Uncle Badger Bates and National Parks and Wildlife Service archaeologist Dan Witter saw bones eroding out of a road cut in Kinchega National Park, an area along the Baaka, or Darling River, in New South Wales, Australia. Badger recognized the bones as a dingo, lying on its left side in what was once a carefully built mound of river mussel shells.At the urging of the Menindee Aboriginal Elders Council, which worried that erosion would end up destroying the dingo bones and any information about the past they contained, a team of archaeologists, working alongside Barkindji elders, excavated and studied the skeleton. The bones turned out to belong to an elderly male dingo, with worn teeth and possible signs of arthritis. Broken and healed bones suggested that he’d lived a tough, active life but also been cared for by people.And the layers of shells around him revealed that generations of Barkindji had tended his grave and ritually “fed” him by adding shells to the mound for centuries after his death. This is definitely not the first dingo burial ever found in Australia, but it’s farther north and west than any other example. It reveals a far more profound and lasting relationship between ancient people and dingoes than outside researchers, at least, had previously fully realized.“This confirms these traditions were much more widespread than we once thought,” said University of Western Australia specialist Loukas Koungoulos, the lead author of the paper, in a press release.The dingo’s bones revealed healed injuries and possible arthritis.Credit:
                                  Amy WayThe dingo’s bones revealed healed injuries and possible arthritis.Credit:

      Amy WayHunting kangaroos and snoozing by the fireThe dingo’s bones tell their own story. Koungoulos says he was probably between 4 and 7 years old, which would be late middle age for a wild dingo today. Heavily worn teeth were the first hint of the dingo’s senior citizen status, but the ends of his leg bones also showed signs of bone decay, probably thanks to long-term inflammation: possibly something like arthritis.And he was shorter than most wild dingoes, based on the length of his femurs. That’s not unusual—domesticated animals are often shorter than their wild relatives, and it doesn’t take many generations for that to show up—but it could say something interesting about exactly how close wild dingoes got to domestication in the centuries before European colonists wrecked everything.At some point, the dingo had suffered a broken rib and lower leg. Koungoulos suggests the injuries look like the aftermath of a kangaroo kick and may have happened on a hunt. The injuries themselves aren’t too surprising; wild dingoes hunt kangaroos, andAboriginal hunters worked with dingoesthe same way people in other parts of the world have hunted with dogs for millennia. What’s more striking is that the two injuries were long-since healed. Somebody nursed this dingo back to health after his kangaroo encounter.“What stands out about garli is that he was old and well-cared-for,” said Koungoulos. “The healed injuries, worn teeth, and careful burial tell us that this animal lived a long life alongside people, and that his death was marked intentionally and with respect.”Barkindji elders stand with the excavated dingo skeleton; in the center is Uncle Badger Bates, who first identified the bones as a dingo.Credit:
                                  Amy WayBarkindji elders stand with the excavated dingo skeleton; in the center is Uncle Badger Bates, who first identified the bones as a dingo.Credit:

      Amy WayHow dingoes became beloved community membersPeople have lived in this part of Australia for at least 40,000

A little too lateElon Musk took too long to sue OpenAI, jury unanimously agreesMusk plans to appeal after judge immediately affirmed the jury’s decision.Ashley Belanger–May 18, 2026 2:09 pm|164Elon Musk attends a state banquet hosted by Chinese President Xi Jinping in Beijing during a summit with Donald Trump.Credit:Alex Wong / Staff | Getty Images NewsElon Musk attends a state banquet hosted by Chinese President Xi Jinping in Beijing during a summit with Donald Trump.Credit:Alex Wong / Staff | Getty Images NewsText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navElon Musk took too long to file his lawsuit that accused OpenAI of stealing a charity, a nine-person jury unanimously decided Monday.Musksued OpenAI in 2024for making a “fool” out of him after Musk donated $38 million to kick-start OpenAI as a nonprofit, only to later be blindsided when OpenAI created a for-profit arm that he felt gutted funding for the charity while enriching executives like Sam Altman and Greg Brockman.But the jury found that Musk was aware of OpenAI’s restructuring plans as early as 2021 and therefore missed the statute of limitations requiring him to bring the lawsuit within three years, The New York Timesreported. Because Musk took too long to file the litigation, the jury deemed Altman and Brockman not liable for any of the claims that Musk brought against OpenAI, the NYT reported. The jury also let Microsoft off the hook, finding no liability for the OpenAI investor after Musk alleged they aided OpenAI’s get-rich scheme.Judge Yvonne Gonzalez Rogers almost instantly agreed with the jury’s decision, the NYT reported.She thanked the jury and reminded them that they can talk to “anyone about anything” once they feel comfortable discussing the case, the NYT reported.Musk’s legal team looked “sunken in their chairs” after the jury’s decision was announced, the NYT reported. One of the billionaire’s lawyers, Marc Toberoff, who hosted daily calls with press during the trial, quickly confirmed that Musk will appeal.Altman and Brockman were not present when the verdict came in, the NYT reported. However, there were “a few smiles from the OpenAI side of the room,” with OpenAI attorney William Savitt wearing “a wide grin,” the NYT reported. Savitt spent daysgrilling Musk on the stand, seemingly taking advantage of his prior experience as Musk’s lawyer to get under Musk’s skin.In a statement, Microsoft celebrated the win, writing,“The facts and the timeline in this case have long been clear, and we welcome the jury’s decision to dismiss these claims as untimely. We remain committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”Musk did not return to the courtroom after delivering his testimony and was also not there when the verdict was read. Before the verdict arrived, Musk’s lawyer apologized after Musk shirked a court order requiring him to be available in case any further testimony was needed in the case. In an apparent sign that he was losing interest in the trial, Musk prioritized attendinga summit between Donald Trump and China’s President Xi Jinping.Late Monday, Musk issued a statement on X, writing, “Regarding the OpenAI case, the judge & jury never actually ruled on the merits of the case, just on a calendar technicality. There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!”Musk reiterated his lawyer’s statement confirming he plans to appeal, writing, “I will be filing an appeal with the Ninth Circuit, because creating a precedent to loot charities is incredibly destructive to charitable giving in America.”This story was updated to include Musk’s statement.Ashley BelangerSenior Policy ReporterAshley BelangerSenior Policy ReporterAshley is a senior policy reporter for Ars Technica, dedicated to tracking social impacts of emerging policies and new technologies. She is a Chicago-based journalist with 20 years of experience.164 Comments

Who will win mtn?One Mars spacecraft, two senators, and a cloud of questions“I think there’s plenty of fire lit under them already.”Eric Berger–May 18, 2026 3:49 pm|18A rendering of Rocket Lab’s design for the Mars Telecommunications NetworkCredit:

      Rocket LabA rendering of Rocket Lab's design for the Mars Telecommunications NetworkCredit:

      Rocket LabText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navNASA released a much-anticipatedcontract solicitationfor a Mars-orbiting spacecraft late last week, kicking off what is sure to be a hotly contested and potentially controversial procurement.At issue is $700 million, already appropriated by Congress, to build a spacecraft, launch it to Mars, and once there to serve as a vehicle to relay communications between the red planet and Earth. But the stakes may be even bigger than this, including the possible resurrection of the recently canceled Mars Sample Return mission.As part of the new solicitation, NASA says it will conduct the acquisition “as a full and open competition.” But will it? That’s the question that several people involved with this procurement process are asking. And it could turn messy, quickly.What is not controversial is that NASA needs a new spacecraft capable of relaying communications from Mars to Earth. NASA’s best communications relay today is the Mars Reconnaissance Orbiter, which has now been there for 20 years. It’s a great spacecraft, but it’s getting long in the tooth.The US Congress, more or less out of the blue, stepped up with $700 million in funding for a new Mars Telecommunications Orbiter in the “One Big Beautiful Bill” passed in 2025.Ars previously wrote about this legislation, which raised some eyebrows, as this is a large appropriation for a relatively straightforward spacecraft.Competition is narrowedProposals to build the spacecraft, the name of which has been changed to Mars Telecommunications Network (MTN), are due by June 15. A contract is expected to be awarded by October 1, less than five months from now.There was some curious wordingin the legislationthat funded the Mars orbiter. It specified the spacecraft must be selected from among US companies that “received funding from the Administration in fiscal year 2024 or 2025 for commercial design studies for Mars Sample Return; and proposed a separate, independently launched Mars telecommunication orbiter supporting an end-to-end Mars sample return mission.”Similar language ended up in NASA’s procurement notice, which states that to be eligible to bid, a company needs to “demonstrate that the Offeror proposed a separately and independently launched Mars telecom orbiter to support an end-to-end Mars Sample Return mission.”That seems like a curious requirement. Why must the bidder for a relatively straightforward orbiter around Mars have previously proposed an orbiter as part of an “end-to-end” mission to return samples from Mars? Was Congress seeking to be preferential for a particular company?Some hints about what is happeningThe One Big Beautiful Bill was signed into law on July 4, 2025. One month later, as part of itsQ2 Investor Update, Rocket Lab included a slide about its plans to support human missions to Mars. In the corner of the slide, the company noted that $700 million had just been appropriated for a Mars orbiter (MTO).“Rocket Lab was the only commercial provider to propose a MTO as part of an end-to-end Mars Sample Return mission to support human exploration and science missions to Mars,” the slide stated.Slide from Rocket Lab Q2 2025 Investor Update.Credit:
                                  Rocket LabSlide from Rocket Lab Q2 2025 Investor Update.Credit:

      Rocket LabRocket Lab is one of several companies eligible to compete based on the requirement that bidders have received funding in 2024 or 2025 for a “commercial design study” for a Mars Sample Return mission. Other eligible bidders include: Blue Origin, L3Harris, Lockheed Martin, Northrop Grumman, SpaceX, Quantum Space, and Whittinghill Aerospace.At an “Industry Day” event earlier this year, NASA officials indicated that the contract solicitation would be released by May 1. However, it was two weeks late. Weeks matter when NASA is seeking to have this spacecraft built and launched before the end of the next Mars launch window in late 2028.Ars’ Stephen Clark recently asked Rocket Lab chief executive Peter Beck if he knew why this delay had occurred. He said he did not. Beck was also asked whether NASA was moving quickly enough given the schedule constraints.“I think there’s plenty of fire lit under them already,” Beck said. “They understand the importance of the mission. So I think we just need to get it out there, and they need to get it awarded, and everybody needs to get to work.”One source indicated to Ars that this delay was due in part to a letter sent by US

Did AI do that?Legal fail: Don’t use AI to sue Facebook users for calling you a bad dateFake citations dashed a dude’s “Are We Dating the Same Guy” revenge lawsuit.Ashley Belanger–May 18, 2026 4:27 pm|54Credit:pikepicture | iStock / Getty Images PlusCredit:pikepicture | iStock / Getty Images PlusText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAn attempt to pressure Meta into removing a critical post from a Chicago Facebook group called “Are We Dating the Same Guy” may end in sanctions for lawyers whose takedown arguments appeared to rely on fake AI citations to support doxing claims.The case had already been dismissed with prejudice by a district court, which ruled there was no way to amend the complaint to possibly save it. But Nikko D’Ambrosio—who accused more than two dozen women of defaming him and blamed Meta for supposedly boosting the post to profit off its “entertainment value”—appealed anyway.Perhaps he felt confident despite his likely tough odds because he was relying on MarcTrent.AI, a law firm thatclaimsto use AI to “uncover legal opportunities traditional firms miss” and “increase legal success rates by 35 percent through predictive modeling.”In a 2025blogdiscussing the case, founder Marc Trent confirmed that the firm had “utilized our tech team to draft” the initial complaint. He boasted that the “evolved” firm uses “everything related to AI now,” suggesting that “even Meta can’t beat us” and claiming that a win would make Facebook safer for everyone.Laying out the case, Trent said that he assumed that Meta “would quickly distance itself by removing the post.” But when Meta didn’t, he figured that overcoming Section 230 claims would be his biggest hurdle in the fight. However, he insisted that “his firm’s technological capabilities” would level the playing field, making it possible to beat Meta’s well-resourced legal teams who are deeply schooled at defending against Section 230 claims.However, during the appeal, judges agreed that the case was so weak that Section 230 didn’t even factor in. And the firm’s seeming reliance on AI to “execute” arguments “with precision” apparently did not help matters at all.In anopinionFriday, David Hamilton, a senior circuit judge for the US Court of Appeals for the Seventh Circuit, wrote that the three-judge panel agreed that “this is a relatively rare appeal in which sanctions appear to be appropriate.” Not only was the appeal deemed “frivolous” for failing to advance D’Ambrosio’s arguments, but it was also littered with “mistakes and fictitious quotations” that “bear the hallmarks of the misuse of generative artificial intelligence,” Hamilton said.“Briefs and other court submissions that include fictitious quotations—inaccuracies discoverable with elementary professional care—are unacceptable,” Hamilton wrote.MarcTrent.AI did not immediately respond to Ars’ request for comment on whether the firm plans to fight the potential sanctions.The firm has until June 16 to request a hearing or file statements on whether sanctions are warranted.Man fails to scrub menacing text from FacebookD’Ambrosio’s legal fight started when a woman whom he briefly dated, Abbigail Rajala, blocked his number, and he persisted in sending a menacing text by using an alternate number.Screenshot of text D’Ambrosio wanted removed from Facebook.Credit: via Judge’s Opinion in D’Ambrosio v MetaScreenshot of text D’Ambrosio wanted removed from Facebook.Credit:

Escalated quicklyEbola outbreak: WHO declares emergency, US restricts travel, American infectedCDC is working to move the infected American and six others to Germany.Beth Mole–May 18, 2026 4:41 pm|43A poster displaying Ebola emergency contact numbers is pinned to a tent at the Busunga border crossing between Uganda and the Democratic Republic of Congo in Bundibugyo, on May 18, 2026.Credit:Getty | Badru KatumbaA poster displaying Ebola emergency contact numbers is pinned to a tent at the Busunga border crossing between Uganda and the Democratic Republic of Congo in Bundibugyo, on May 18, 2026.Credit:Getty | Badru KatumbaText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navThe Ebola outbreak first reported in the Democratic Republic of the Congo on Friday has seemingly escalated quickly into a large, uncontrolled multinational outbreak.As of May 17, there were 10 confirmed cases, 336 suspected cases, and 88 deaths in the DRC, as well as two confirmed cases and one death in neighboring Uganda, according to the latest data from the US Centers for Disease Control and Prevention, which has offices in the region. The numbers already put the outbreak within the top 10 Ebola outbreaks recorded by size, though still far from the worst—the 2014–2016 West African outbreak had over 28,000 cases and 11,000 deaths.International emergencyOn Sunday, the World Health Organization declared the outbreaka public health emergency of international concern (PHEIC), though it noted that it does not meet the criteria for a pandemic emergency. In making the PHEIC determination, WHO Director-General Tedros Adhanom Ghebreyesus cited several factors in addition to the immediate large size, including clusters of suspected cases and deaths in multiple DRC health zones, four deaths among healthcare workers, and a lack of apparent links between geographically distant cases and clusters. The features collectively suggest that the outbreak is larger than what is currently being detected and is spreading regionally.“Moreover, the ongoing insecurity, humanitarian crisis, high population mobility, the urban or semi-urban nature of the current hotspot, and the large network of informal healthcare facilities further compound the risk of spread,” the WHO said.The final extraordinary aspect of the outbreak is that it is caused by the uncommon Bundibugyo strain of Ebola virus, which has no clinically validated treatments or vaccines. This is only the third Ebola outbreak caused by Bundibugyo, which has had fatality rates of 25–50 percent.There are four virus strains known to cause Ebola disease in humans, and three have caused large outbreaks (Zaire, Sudan, and Bundibugyo). The most common strain is Zaire, for which treatments and vaccines have been developed. The viruses spill over from animals, including non-human primates and bats, and cause severe hemorrhagic fever, marked by diarrhea, vomiting, and bleeding. Person-to-person spread occurs via contact with bodily fluids and symptoms can develop between two and 21 days—though most often eight to 10 days—after an exposure.CDC response and infected AmericanOn Monday morning, the CDC announced on its website that it is implementingnew travel restrictions, including screening and monitoring Americans arriving from DRC, Uganda, and South Sudan, while also barring the entry of non-US passport holders who have traveled in those countries in the past 21 days.Additionally, in a CDC press briefing on Monday afternoon, Captain Satish Pillai, incident manager for CDC’s Ebola response, said that one American in the DRC has been infected after being exposed as part of their work there. The person developed symptoms over the weekend and tested positive late Sunday. The CDC is now working to transfer that person, along with six other Americans, to Germany, where they will receive care. Pillai did not answer questions about the person’s identity or their work.Serge, a Christian missionary organization, announced thatthe infected person is Dr. Peter Stafford, who has been working in the Nyankunde Hospital in Bunia, DRC, since 2023. The other six people the CDC is working to relocate are his wife, Dr. Rebekah Stafford, the couple’s four children, and a third doctor with the organization, Dr. Patrick LaRochelle. All three doctors had exposures, the organization said, but Rebekah Stafford and LaRochelle are currently asymptomatic.Pillai noted that the CDC considers the risk to the American public to be low.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.43 Comments

📱iPhone 17 Pro 官降 1000 元🧠ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施🤖估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相🌐Claude 重新允许接入龙虾，开发者：这是变相涨价📋腾讯辟谣「AI 一号位将离职」💰黄仁勋基金会捐出 1.08 亿美元 AI 算力📱影石 CEO 回应 Luna 定价贵：5299 元是美国价格🚗鸿蒙智行洽谈打造玛莎拉蒂新能源车🚙华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗💻英特尔成为迈凯伦官方计算合作伙伴⛽本田上市以来首次年度亏损📋LinkedIn 宣布裁员，涉及工程、产品和市场营销部门💡Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」🤖曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线📱荣耀 600 系列定档 5 月 25 日🔍Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光🧠腾讯开源 AI Agent 记忆系统，Token 消耗最高降低 61%🧸泡泡玛特否认进军传统家电，LABUBU 冰箱一季度收入约 600 万iPhone 17 Pro 官降 1000 元昨晚，苹果无预警在天猫「Apple Store 官方旗舰店」启动 iPhone 17 Pro 系列 618 优惠活动，Pro 和 Pro Max 同步降价 1000 元，起售价分别来到 7999 元和 8999 元。本次活动还覆盖 iPad、Apple Watch、AirPods 等系列，非 iPhone 产品可使用 50 至 100 元不等的惊喜券，并可叠加 88VIP 消费券、店铺优惠及国家补贴等多重优惠。ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施据彭博社报道，苹果与 OpenAI 的合作关系已出现明显裂痕，OpenAI 正在联合外部律所评估可能的法律行动。知情人士透露，OpenAI 律师团队正与外部律所合作，研究包括向苹果发送违约通知在内的多项方案，但目前尚未正式提起诉讼。OpenAI 仍希望在庭外解决与苹果的分歧，相关法律行动预计将在马斯克相关诉讼案结束后才会推进。苹果与 OpenAI 的合作始于 2024 年 6 月，苹果在 WWDC 上将 ChatGPT 整合进 Siri 及 iOS 系统，涵盖文字生成、视觉智能、图像生成 App「图乐园」等场景。估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相昨天，前 Meta FAIR 研究院总监田渊栋正式宣布联合创办 AI 初创公司 Recursive Superintelligence，距其确认从 Meta 离职刚好过去半年。他在 X 发文称，拒绝了 OpenAI、xAI、Anthropic 等多家头部 AI 公司的邀约，选择自己创业：「趁着自己还年轻，去当一家新初创公司的联合创始人。」Recursive 同步披露完成一轮 6.5 亿美元融资，估值达 46.5 亿美元，由 GV（谷歌风投）和 Greycroft 领投，AMD Ventures 与英伟达跟投。公司押注的核心方向为「递归自我改进」（Recursive Self-Improvement），即构建能够自动发现知识、持续优化自身的 AI 系统。Claude 重新允许接入龙虾，开发者：这是变相涨价Anthropic 昨天官宣，将于今年 6 月 15 日起为 Claude 订阅版用户引入独立的 Agent SDK 月额度，OpenClaw 等第三方 Agent 工具由此重新获准通过订阅账户接入 Claude。按照新规，Agent SDK 的调用将从用户的常规订阅限额中彻底剥离，另设专属月度额度池，额度按账单周期刷新，未用完的部分不结转至下月。Pro 版和 Team 标准席位每月获得 20 美元额度；Max 5x 及 Team 高级席位为 100 美元；Max 20x 与 Enterprise 座位制高级席位最高可达 200 美元。该额度专用于 Claude Agent SDK 项目（Python 或 TypeScript）、claude -p 非交互命令、Claude Code GitHub Actions 集成，以及通过 Agent SDK 认证的第三方应用。此次新规被视为 Anthropic 对今年 4 月 4 日封禁政策的正式解决方案。彼时，Claude Code 负责人 Boris Cherny 宣布，Claude Pro 及 Max 订阅额度即日起不再覆盖 OpenClaw 等第三方工具的调用，理由是此类工具的使用模式对 Anthropic 的算力基础设施造成了难以为继的压力。开发者社区对此反应冷淡。有开发者直言，20 美元的 Agent 额度在高强度任务下消耗极快，实际可用量远低于封禁前的订阅套餐，并将此次调整定性为「变相涨价」。腾讯辟谣「AI 一号位将离职」昨天，腾讯官方发文辟谣，明确否认外界流传的「AI 一号位即将离职」传言。腾讯公关总监张军表示，「太恶劣了，还编出各种搞笑的情节。」谣言起源于社交平台上的多条帖文，内容称某大厂 AI 一号位因被绿色事业部抢走算力资源而准备离职，并点名影射腾讯 AI 负责人姚顺雨。黄仁勋基金会捐出 1.08 亿美元 AI 算力据路透社报道，英伟达 CEO 黄仁勋与其妻子洛丽创立的基金会，从云计算公司 CoreWeave 购入 AI 算力，并将其捐赠给多所大学及其他非营利研究机构，用于科学和人工智能研究。根据昨日提交的一份文件，此次捐赠的算力价值目前已达 1.083 亿美元。英伟达还表示，计划向部分受捐机构提供免费工程服务。CoreWeave 是一家专注于 AI 应用的云计算公司，其向客户提供的 GPU 正是由英伟达设计制造。今年 1 月，英伟达向 CoreWeave 投资 20 亿美元，一度成为该公司第二大股东。影石 CEO 回应 Luna 定价贵：5299 元是美国价格昨天，影石 CEO 刘靖康回应旗下旗舰双摄云台相机 Luna Ultra 的定价问题，表示网传单机 5299 元的售价属实，「但是是美国的价格，国内给我再大的胆子都不敢卖这个价格」。此前，博主「数码闲聊站」爆料称该机预计售价 5299 元、全能套装预计 6499 元，随即在网络上引发「定价过高」的质疑。鸿蒙智行洽谈打造玛莎拉蒂新能源车据《云见 Insight》报道，华为鸿蒙智行、江淮汽车、斯特兰蒂斯集团（Stellantis）及旗下玛莎拉蒂品牌，正就共同打造玛莎拉蒂品牌的新能源车展开洽谈。按照规划，四方合作模式与鸿蒙智行 「五界」 高度相似：华为主导产品定义、提供核心技术，江淮联合研发并负责生产制造，玛莎拉蒂提供造型设计与品牌背书。合作车型将分为国内和海外版本，国内版归属于尊界品牌，海外版将挂玛莎拉蒂车标，首款车型目前正处于造型设计阶段，计划于明年下半年量产。华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗华晨宝马昨日公布年度节水数据，显示其沈阳生产基地 2025 年单台生产取水量降至 1.4 立方米/台，同比降幅 10.8%，连续八年稳步下降。在水资源消耗总量方面，华晨宝马沈阳基地全年取水总量降至 75.2 万立方米，较 2023 年降幅超 40%，连续三年递减。再生水使用量超 68 万立方米，同比增长 71.9%，官方称节约用水量可满足近 9300 户沈阳家庭一整年的生活用水需求。华晨宝马大东工厂北厂区实现所有生产工艺自来水消耗为零，全部改用再生水替代；铁西工厂及里达厂区涂装车间冷却塔也已扩展使用市政再生水，厂区绿化灌溉此前已于 2024 年实现 100% 使用市政再生水。英特尔成为迈凯伦官方计算合作伙伴英特尔昨日宣布与迈凯伦车队达成多年期战略合作伙伴关系，正式成为迈凯伦万事达一级方程式车队、艾睿迈凯伦印地车队及迈凯伦 F1 模拟赛车队的官方计算合作伙伴。根据合作协议，英特尔至强处理器和酷睿 Ultra 处理器将为迈凯伦的核心性能负载提供算力支持，覆盖计算流体动力学、空气动力学分析、车辆动力学仿真、比赛策略分析等方面。英特尔的技术还将应用于赛道边缘计算、数字孪生与 AI 预测建模等场景。本田上市以来首次年度亏损昨天，本田公布了 2026 年 3 月期的合并财报，期内营收 217966 亿日元，同比微增 0.5%；归母净亏损 4239 亿日元，上一财年同期归母净利润为 8358 亿日元，由盈转亏。亏损主因是电动化战略大幅调整。本财年内，本田决定取消北美多款电动车型的开发与上市计划，由此确认减值、处置及合同赔偿等相关损失合计约 15777 亿日元，全部计入汽车业务。摩托车业务为本财年唯一实现正向盈利的板块。展望明年，公司预计营收 231500 亿日元，归母净利润恢复至 2600 亿日元。LinkedIn 宣布裁员，涉及工程、产品和市场营销部门据彭博社报道，职业社交平台 LinkedIn 昨日宣布裁员，涉及工程、产品和市场营销等多个职能部门。LinkedIn CEO Daniel Shapero 在发给员工的备忘录中表示，公司需要为用户提供更大价值，并实现更高的盈利水平，此次组织架构调整旨在为公司未来发展做好定位。LinkedIn 发言人随后确认了相关变动，称此次调整属于「常规业务规划」的一部分，但未披露具体裁员人数。💡 Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」据《商业内幕》昨日报道，Anthropic CFO Krishna Rao 在播客节目《Invest Like the Best》中透露，Claude 目前已承担公司 90% 以上的代码编写工作，并正在深度介入财务流程，推动白领工作从「执行」向「监督」转型。Rao 还透露，公司已借助 Claude 生成财务报表，每月财务审查流程在人工介入之前已达到「90% 至 95% 完成」的状态。他提到，部分内部报告的生成时间从原来的数小时压缩至 30 分钟。在人员配置方面，Rao 将 AI 的引入描述为生产力「加速器」。他称，由于 AI 的介入，员工得以将更多精力从信息收集转向决策判断，Anthropic 反而因此扩大了招聘规模。我们雇用了更多人，因为没有什么工作是做不完的。Rao 还描述了一种新的工作形态：员工越来越多地扮演 AI 系统的「管理者」角色，团队可以同时部署多批 AI 智能体并行处理项目。他将这一趋势概括为「每个人都在某种程度上成为管理者」，并认为由此带来的生产力潜力「才刚刚开始显现」。他同时强调，Anthropic 内部推崇「人才密度」而非「人才规模」，认为高密度的顶尖 AI 研究人才配合最强模型，才是真正具有竞争力的组合。曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线知名爆料者 Leo 昨日披露，GPT-5.6 研发已全面推进，预计最早下月正式发布。开发者在 OpenAI 内部 Codex 日志中发现了两个内部代号 —— ember-alpha 和 beacon-alpha，以及一条路由映射记录。日志显示，绝大多数模型调用仍指向 GPT-5.5，但有一条条目明确指向 GPT-5.6，暗示 Codex 环境可能已在测试新模型。此外，爆料者 Chetaslua 称 OpenAI 计划于本周四推出旗舰模型的「超快模式」，响应速度或可提升 2 至 3 倍，主要面向延迟敏感型任务。荣耀 600 系列定档 5 月 25 日荣耀昨日宣布，旗下荣耀 600 系列将于 5 月 25 日 19:00 正式发布。官方预热信息显示，荣耀 600 系列主打「4K 闪光微单 Live」影像方向，搭载 2 亿像素超清大底主摄，并内置 8600mAh 青海湖大电池。Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光据 Sources 报道，Google 计划在 5 月 20 日举行的 Google I/O 大会上

相信这样的场景很多人都不会陌生：在微信上聊工作、讨论旅行计划，甚至只是和朋友闲聊，聊着聊着突然想让 AI 帮忙总结重点、做个规划，结果却只能一段一段复制聊天记录，再手动粘贴进 AI 对话框里。再比如，中午小憩了一会，醒过来发现群里发了几百条消息，想吃瓜又得一条一条翻……如果能一键将聊天记录转发 AI，岂不美哉？这样的功能终于来了，但遗憾不是在微信里：现在，把微信和元宝更新到最新版本之后，聊天记录转发可以选择直接发给元宝来总结。虽然同属腾讯生态，但微信和元宝之间始终有种微妙的距离感——终于，这次两者之间的联动，又紧密了一点。爱范儿也第一时间体验了这个功能，看看它究竟好不好用。产品很美好，效果很一般要想把微信聊天记录转发给元宝，方法很简单，首先选中想要转发的聊天记录，选择「转发到其他应用」，就能直接发给元宝，可以补充一些提示词。整个过程有点像把讨论、聊天过程中零散的内容，让 AI 帮忙进行整理，因此这个功能首先最合适的场景，就是各种工作群的讨论和安排信息。直接处理聊天记录有一个好处，元宝可以将讨论精准和发言者匹配，很适合用来处理会议纪要和工作安排。旅游计划同样也是一个高频场景，因为行程一般都是靠你一言我一句的对话讨论出来，这时候就可以直接交给元宝处理，还能让它帮忙推荐旅游地点。还有一个有点邪修的用法：平时工作忙没时间看群，又担心错过什么热点或者话题，可以直接选中所有聊天记录，让元宝帮你看群帮你总结。又或者讨论了很久都不知道聚餐吃什么，也可以把讨论的过程选中，让元宝帮忙推荐。尝试了很多场景，大部分情况下都达到了预期，但总体来说，还有一些体验细节不够完善。目前元宝最擅长处理的聊天记录以文字为主，而如果聊天记录中包含文件，特别是涉及多个文件，效果会比较随缘，很多时候会无法读取文件。即使能成功读取，也不一定能按照提示词进行任务处理：整理发票有时候都未必能填对，让他根据文件做个思维导图，也经常失败。要知道，整理聊天里的发票真的是很主流的场景啊！我能很明显感觉到一个落差：明明将聊天记录发送给 AI 是一个很有价值的场景，但元宝交付的结果，总让我略感失望，并且忍不住想，如果能转发给豆包或者 ChatGPT，会不会更好？更深层的问题，是这个聊天记录转发的交互总给我一种「绕远路」的感觉。在此之前，元宝已经作为联系人存在于微信之中，处理聊天记录这件事，却依旧需要「跳转」这一整套外挂式流程， 过程中需要多次点击。或许是因为微信严格的隐私保护规则，元宝只能以「临时聊天」的形式处理聊天记录，不会留下历史记录，关闭后无法找回。对于微信来说，这样的功能似乎「可有可无」。如果微信想要 AI 化，他们或许会有更好的实现方式，例如直接做一个「AI 群友」，at 它就能回答问题或者处理聊天记录；以及一个常驻对话列表的 Agent，可以直接对话，可以直接接收聊天记录。但这件事没那么容易，原因已经是「老生常谈」：微信需要简约，需要克制，因为每一点小变化都会影响到十亿级用户。在元宝这个功能上线的同一天，腾讯发布了今年第一季度财报，在股东大会上马化腾表示，在微信里做 Agent 需要「更长远的考虑，大家要有点耐心，这个不是匆忙就能出来的。」腾讯的 AI 王牌是微信马化腾还提到腾讯的 AI 战略：原来一年前我们以为上了船，后来发现那个船漏水了，现在感觉站上去了，还坐不下去，还是希望船速能快一点。从财报来看，AI 已经成为了腾讯的重要投入，一季度研发投入达 225.4 亿元，同比 19%；资本开支 319.4 亿元，主要用于算力、存储等云端基础设施的扩张。今天中午，有自媒体传出「疑似腾讯 AI 一号位即将离职」的消息，下午就被腾讯公司火速辟谣，本身也向外界释放出一个信号：腾讯 AI，不容有失。根据 AICPB 最新榜单，在国产 AI 应用排名中，腾讯元宝排名第十，访问量为 1773 万，远逊于豆包（1.63 亿）和千问（4424 万）。在 AI 领域中，腾讯和苹果有着相似的处境：产品入口和用户规模都是顶级配置，AI 模型本身却还支撑不起这种期待。元宝 AI 社交应用「元宝派」生成式 AI 火了 3 年，从去年年底开始，AI 应用之间的战斗已经进入了「场景化」的比拼，不仅出现「龙虾」OpenClaw 这样的全新形态，传统的 ChatBot 交互模式也集成了生活服务，像阿里千问和字节豆包，都在积极将电商等各种生态内的优势服务，和自家的 AI 助手相结合。这本来应该是腾讯的舒适区，因为国内最强大、最全能的生活应用，也就是微信，就在腾讯手中。但很明显，微信对「AI 化」抱有非常慎重的态度。腾讯不能承受因技术或功能不成熟而扰乱微信庞大用户群的体验。只是对于腾讯来说，在 AI 性能落后其他对手的情况下，微信这张王牌的价值还是太重要了。，于是，我微信和元宝虽然没有完全敞开双臂拥抱彼此，也开始小心翼翼试探边界。究其原因，除了微信相对封闭的生态逻辑，还有两个问题没能被回答：微信究竟要如何「AI 化」？而元宝在其中又应该扮演什么样的角色？The Information 报道指出，即使腾讯在下重本加速建设自己的混元模型，微信方面还是认为混元当前水平不够拔尖，甚至开始为微信 Agent 测试第三方 AI 模型，这带来了隐私和整合上的挑战性。微信方面，爆料称其实一直在做自己的 AI Agent 模型技术，计划在年中发布一个微信智能体，可以和整个小程序生态相结合，提供打车、购物的服务，并且将以联系人和聊天的方式直接集成微信之中——但就像前面提到的，这件事太重大、影响面太广了，急不得。一旦这个产品问世，可能将是全国最多用户的 AI 入口。微信需要 AI，但未必需要「元宝」，这才是腾讯现在最微妙的问题。

大疆的 Pocket 系列三年不更新，一更新就来了两台——过去一个多月，Pocket 4 和 Pocket 4P 可以说是最受瞩目的相机，就在今晚的戛纳电影节上，Pocket 4P 以口袋电影机的身份在欧洲电影殿堂亮相，距离正式发布时间，相信不会太远了。与此同时，大疆也对 Pocket 3 进行了降价处理，于是我们看到了泾渭分明的 Pocket 系列产品线：Pocket 3：入门级的手持云台 vlog 机，2299 元Pocket 4：全能的手持云台相机，2999 元Pocket 4P：准专业级的口袋电影机，预计价格会来到 4000 元档相信有许多持币观望的读者，也和爱范儿抱着同样的疑问：三台 Pocket 我该怎么选？有了 iPhone，我还需要 Pocket 吗？结合实际体验，爱范儿一次性帮你理清楚。Pocket 三代同堂，怎么选？先仔细聊聊 Pocket 4P。虽然目前 Pocket 4P 依旧没有公布更详细的参数，但参数在本次形态更新面前，是最不重要的话题——从产品逻辑来看，Pocket 4 已经将单摄的手持云台相机推向较为成熟的状态。此时，增加一颗摄像头成为一件顺理成章的事情。绝大多数手机的主摄和单摄 Pocket，提供的都是广角视野。广角适合交代环境，适合把大山大河框进画面。但当你拿着设备走入拥挤的人群，广角的缺点就暴露无遗：画面太杂乱了。此时，一颗长焦镜头可以完成空间上的压缩，把无关的背景剔除出去。大疆将这颗多出来的摄像头定格在 3×，这是一个非常巧妙的数字。它比主摄更贴近人眼对焦后的视线范围。从我们上手体验来看，Pocket 4P 这颗 3× 镜头还提供了比较自然的虚化，非常适合拍摄更凸显主体的人像或特写画面。但对于用户来说，多的这一颗摄像头，并非绝对的正收益：Pocket 系列自诞生起，核心卖点就是「无感携带」，增加镜头，意味着增加重量、侵占内部空间，甚至可能牺牲一部分续航。其实这种担心，恰恰帮大疆划清了 4P 和 4 之间的界限。Pocket 4P 是给那些「创作者」准备的。他们清楚地知道自己要拍什么，素材最后会进入 Mac，在 DaVinci Resolve 或者 FCPX 里进行精细的剪辑。对他们来说，为了在极轻便的体积下获得原生焦段的景别切换，以及高度统一的素材规格，牺牲一点点重量完全值得。而对于大多数普通人来说，Pocket 4 才是那个更轻量、更理性的选择。虽然前代 Pocket 3 打造出千万台的神话，但其本身并不是一台完美无缺的相机——只有 USB 2.0 导致数据传输过慢、像素量过低以及动态范围不够大，这些都是问题。而 Pocket 4 提供了 3700 万像素，更高的像素意味着内置的无损 ISZ 变强了。在 Pocket 4 上，通过传感器裁切获得的 2× 焦段更为清晰，足以应付日常记录。除此之外，它还内置了 107GB 存储，带来了更好的动态范围。忘记带内存卡、传输速度和动态范围不够的问题，被一次性解决掉了。可以说，Pocket 4 就是目前单摄手持云台相机的完全体。至此，Pocket 4 系列两台机身的区别已经很清晰了：如果你追求生产力，强调原生光学焦段带来的画质与虚化，需要 3× 镜头获得更丰富的画幅景别，那么 Pocket 4P 毫无疑问是更好的答案；如果你只是希望有一台能轻松塞进裤兜、防抖优秀、可以独立于手机拍摄视频和 Live Photo 的机器，那么 Pocket 4 凭借更高的像素和好用的 2× ISZ，已经绰绰有余。最后是 Pocket 3。在 Pocket 4 补齐了大量痛点的情况下，Pocket 3 还有存在的意义吗？当然——新设备的出现，最大的贡献往往是把老旗舰的价格打下来。在 Pocket 4 系列轮番上场的间隙，我们观察到 Pocket 3 已经完成了一轮官方降价，在某些平台叠加各种补贴的情况下，已经落到了 2099 的价格。甚至在未来，我们能以 1 开头的价格买到 Pocket 3，也不是不可能。对于预算有限、或者刚刚萌生拍视频念头的新手，这就是性价比最高的 「守门员」。最后，按照爱范儿编辑部的推测，Pocket 4P 的定价大概率在 4500 元左右，但也不排除大疆卷起来定一个 3999 元的价格。此时，如果将 Pocket 系列按照价格排开，我们会发现一个巧妙的设计——从 2000 元档的 Pocket 3 到 3000 元档的 Pocket 4，最后到 4000 元档的 Pocket 4P，大疆正在把无人机上那套 Pro、Air、Mini 的组合拳搬到口袋相机上。这套阵容已经在无人机市场证明了它的成功。现在，Pocket 也将组成这样的阵容了。有了 iPhone，我还需要 Pocket 吗？在 Pocket 4P 的上手文章发布后，我收到不止一个朋友的问题：我有 iPhone，还需要买 Pocket 吗？理清了 Pocket 家族的内部定位后，这是我们必须面对的现实问题：既然大家手里都有一台性能强悍的 iPhone，为什么还要花钱买一台功能单一的云台相机？这其实是一个关于「设备属性」的哲学问题。在多次迭代后，iPhone 在动态影像领域的基础不可谓不强，三摄变焦顺畅、色彩高度统一、防抖优秀、Apple Log 甚至已经更迭到第二代。但问题在于，手机始终不是完全针对拍摄设计的特化产品——除了拍摄以外，iPhone 还需要承担复杂的日常任务，无时无刻需要分散注意力，随时可能会有工作消息、语音电话打断你的拍摄、需要考虑拍摄之外手机还有没有充足的剩余电量保证你与世界的关联，还需要注意拍摄会不会占满本就贵如黄金的存储空间……而 Pocket 则是完全服务于拍摄需求的相机：不能打电话，不能回微信，不能刷社交软件，一切与拍摄无关的东西，都不会出现在你手里的 Pocket 上。更别提，Pocket 还有机械云台天生在结构上比 iPhone 的防抖更稳定、形态比 iPhone 更适合握持、物理虚化比 iPhone 更自然、视频规格比 iPhone 更丰富等等综合优势了。所以，如果我手里已经有一台主力 iPhone，现在到底该不该买 Pocket？如果现在需要，哪台是最适合我的？你可以对照下面几个问题，做一个购前诊断：过去三个月里，你是否因为手机突如其来的电话或微信，打断过拍摄？拍视频时，是否因为手机存储空间不足，而束手束脚？炎炎夏日，有没有遇到过手机拍照录视频发热掉帧甚至死机的情况？如果你对这些问题的回答是肯定的，那么购入 Pocket 3 和 Pocket 4 都是不错的选择。具体来说，如果你有拍高清 live 图的需求，如果你喜欢用手机的 2× 镜头拍摄，那么选择 Pocket 4，否则，Pocket 3 其实就够用了。在这个基础上，你还想拿 Pocket 拍拍商单、拍拍演唱会，搞点严肃创作，那么 Pocket 4P 是一步到位的选择。当然，如果前面几道题，你的答案都是否定的，那么请捂紧钱包。iPhone 的防抖和画质足够满足你的需求。你可以随拍随剪，直接通过 AirDrop 分享，且丝滑的生态体验依旧是目前任何第三方设备都无法替代的。而从行业角度来看，爱范儿更建议你等一等——目前，口袋云台相机的玩家可不止大疆一家——和 Pocket 4P 同样「犹抱琵琶半遮面」的，还有影石 Luna，而各大手机厂商也正跃跃欲试，准备下场分一杯羹。因为 Pocket 4P 这样的产品，已经严重威胁到了手机厂商的安全区。很多人愿意买 6000 元甚至 8000 元的影像旗舰手机，很大程度就是为了长焦。现在，花更低的价格，Pocket 4P 就能提供一个素质相当的长焦镜头，以及更好的云台防抖、更专业的整机解决方案。没有手机厂商能对此熟视无睹。据爱范儿了解，vivo、OPPO 都已经将手持云台相机的产品规划提上日程，而小米、华为也在持续关注这个赛道。2026 下半年到 2027 上半年，想必会有更多手持云台相机纷至沓来。既然入局者多了，我们作为消费者，反而可以更淡定些。你可以把钱捏在手里，再等一等。我们相信，这个年轻的品类，还会有更精彩的产品出现。

鸿蒙智行车型阵营里可能缺一辆 MPV，但中国汽车市场已经不缺 MPV 了。别克 GL8，丰田埃尔法和传祺 M8 等车型在燃油车时代占据了 MPV 就是商务接待用车的印象之后，新能源 MPV 迎来了数款重量级车型，极氪 009 主打高端，意在取代丰田埃尔法乃至雷克萨斯 LM 的生态位，理想 MEGA 则希望利用 MPV 的空间优势重塑大家庭用车的产品定义，岚图梦想家成为这家新势力最依仗的产品成功和别克 GL8 分庭抗礼，小鹏 X9 想要以宜商宜家，是 MPV 也可以是 SUV 的灵活定位吸引更多受众……新旧势力在 MPV 市场的攻伐，发生在 2023 年，但现在已经是 2026 年了，零跑做一款旗舰 MPV 只需要把性价比拉到极致，但鸿蒙智行做一款旗舰 MPV 并卖四五十万元，需要做的就多了，当然我承认坐在二排的老板们对智界代言人刘亦菲肯定都是心驰神往，但除了刘亦菲这个核心竞争力之外，鸿蒙智行的首款 MPV 肯定需要更多击中老板坚硬内心和柔软脂肪的东西。这些东西是车载制氧系统、全维包裹安全气囊、华为悦彰非凡系列音响、二排主动座椅旋转、侧门交互式迎宾光毯等等等。最终，智界 V9 分为 Max、Max+、Ultra 和 Ultra+ 四个版本，售价分别为 39.98 万，42.98 万，45.98 万和 51.98 万起。舒适重要，安全重要，舒适且安全更重要！过去几年的时间里，鸿蒙智行成功地让国内汽车行业重视起来零重力座椅和主动安全来，是的，这两项鸿蒙智行都不是发明者，但确实在打开用户心智和开拓市场上起到了主要作用。前者代表座舱的舒适性，通过模拟太空失重环境下人体自然放松姿态的设计。通过特殊的人体工程学优化，使臀部、腰部、背部和腿部受力均匀，释放肌肉压力。后者代表汽车安全从被动防御走向主动避让，在汽车发生事故前，通过电子控制系统感知环境、预判风险并主动介入（如自动制动、转向控制），从而避免或减轻事故。但是，二者的矛盾在于在汽车行驶状态下，零重力座椅一般是没法使用的，这是因为当乘客躺在大仰角的座椅上享受时，身体与安全带、气囊的相对位置已经发生了相当大的变化。董车会之前报道过，根据 CIDAS 的数据，一旦座椅靠背夹角超过 115°，正面碰撞中的重伤风险会陡增 4.4 倍。此时，传统的 B 柱固定式安全带无法再有效约束骨盆，人体会因为惯性产生极度危险的「潜滑现象」——整个人滑向座椅下方，安全带可能会锁死在颈部导致窒息。采埃孚的试验甚至显示，在 50km/h 的碰撞下，躺倒状态受到的脊柱伤害堪比从 5 楼坠落。打过工的人都知道，当你向老板解释「零重力座椅虽然很舒服，但是不能在行驶状态下使用」时，老板这个时候一般不会听你的解释，只会跟你说「我不管，我想用，给我想一个解决方案出来」。今天，这个方案就被智界 V9 想出来了。智界 V9 全球首发了二排全维包裹安全气囊，这个气囊集成于二排座椅靠背，触发瞬间会形成一个独特的「U 型」环绕结构，在零重力、躺姿等多种姿态下可以全面防护头部、颈部、胸部，有效防止乘员在车辆碰撞之后身体滑落、碰撞与下潜，智界宣称历经百项严苛工况验证，无误伤风险。与之配套的还有二排坐垫前端也有专门布置的臀部气囊（BAB），在碰撞瞬间快速充气形成物理坡度，顶住盆骨不让身体滑脱，从而大幅降低腰椎压力。做了这么多工作，目的只有一个：就是让二排的老板在车辆行驶的时候好好躺一躺，好好休息一下。车载制氧：让高精力人群更有精力一直以来，对于什么样的人能做老板有个迷思，甚至还有「基因彩票」的说法，就是高精力人群更容易创业做老板。这些人可以每天工作十几个小时不知疲倦，只需要睡四五个小时就能满血复活。正如前面所说，很多老板之所以看起来可以每天工作十几个小时不知疲倦，很有可能是他可以中间随意选择休息时间。而屏幕前后的你和我，并不能在上班时间睡觉。更科学的说法是，高精力人群背后，跟摄氧量有正向关系，很多所谓的低精力人群很可能有鼻炎咽炎或者心肺功能低下导致日常处于轻微缺氧状态，导致精力不足。类似的情况也会发生在车上，很多人坐高铁感觉昏昏欲睡就是车厢里二氧化碳含量高，氧含量低，这种情况就容易困，密闭的汽车座舱也可能出现这种情况。智界 V9 给出的解决方案也很简单，装个制氧系统。这套制氧系统采用了 PSA（Pressure Swing Adsorption）分子筛变压吸附技术，提供了「弥散+鼻吸」两种模式，鼻吸出氧浓度 93±3%。额定制氧排量 8L/min，峰值 10L/min，比一般的家用制氧机大很多。2 个鼻吸口支持 1 拖 2，最多可供 4 人同时吸氧，弥散供氧与空调共用，四个出风口协同工作，30 分钟即可让整个座舱达到富氧状态。除了解决了传统制氧机体积大，噪音大的不足之外，鸿蒙智行的看家本领是智能化。这套制氧系统支持语音控制开启，中控屏实时显示状态；内置舒缓解压与高原出行两种模式，可自动调节供氧浓度，也可以联动小憩模式，帮助身心放松。很多人在森林景区，绿肺公园里感觉心思空明浑身舒爽，也是因为在富氧环境下，血氧饱和度一般会更高一些。富氧环境在特定场景下确实会有积极作用，比如早起后在通勤路上进入富氧环境有助于从昏沉状态中恢复；在运动后进入富氧环境可以加速乳酸代谢，缓解肌肉酸痛。更典型的场景就是针对本身心肺功能或者呼吸系统有问题的人，富氧环境有助于把缺氧状态调整为正常状态，还有就是高原出行环境，正常的氧气浓度可以有效缓解高原反应。在制氧系统之外，针对座舱健康还有鸿蒙 ALPS 健康座舱 3.0，通过把 MOFs 可吸附降解材料融入顶棚、门板、地毯等内饰，即便熄火也能持续吸附有害物净化醛类。UVC 空调主动杀菌，对流感等病毒灭杀率超 99.99%；前后双负氧离子净化器，可以释放超过西双版纳雨林 5 倍的负离子；这一整套净化杀菌和补氧系统，提供的是「健康」的价值，毕竟对于老板来说，紧急时刻生命第一，不紧急时刻健康第一。当然，需要着重说明的是，并不是所有人在所有场景都需要制氧系统制造一个富氧环境，如果本身不处于缺氧状态，血氧饱和度在 95% 以上，不需要运动恢复，没有昏沉嗜睡情况，不在高原地区，那就不太需要吸氧，长期过度摄氧反而有害健康。旋转座椅，优雅，还是优雅冰箱彩电沙发之所以从争议性的玩笑变成了兵家必争之地，还是市场表现的结果，谁不想二三十万的车就配有之前只有 78S 系才能选装的配置呢？这种造车逻辑之下，成本的分配也就必然从内燃机变速箱往舒适性配置上调整了。智界 V9 的沙发，显然值得多点笔墨。双独立结构，通风加热和比气动按摩更强劲的机械按摩已经不足为奇。更关键的是，二排座椅有可旋转设计，最大旋转角度 180 度，可以在多种角度进行旋转和固定，营造了例如 45° 迎宾座椅，90° 观景座和 180° 旋转之后的后排会客室布局等多种模式和选项。尤其是迎宾形态和会客室形态，可以在商务场景里营造比普通 MPV 更优雅的氛围。出门在外，面子可以是自己给的，但归根到底，还是别人给的，给别人面子就是给自己涨面子。智界 V9 类似的优雅配置还有不少，比如说「手势控车」，佩戴支持功能的手表，做出轻捏动作就可以隔空打开车门；二排乘客在车内做一个推一推手势，二排侧滑门就可以优雅闭合；窗边手指划一划，就可以调节智能隐私光幕比例的明暗度……车内的四块智能隐私光幕玻璃最大遮光率可达 99.6%，如果一键开启隐私模式，可以立刻断开车内的位置信息、麦克风、摄像头、网络，保证车内隐私。新车配有超 50㎡ 静音材料覆盖和 13 块双层夹胶隔音玻璃进行被动降噪，同时还可选 RNC 主动降噪。其他一些老板不太关心，但司机需要关心的作为一款鸿蒙智行的车，其实很多关于底盘，辅助驾驶和动力的信息都能猜个大差不差，这里完全可以用列表形式一一列举，毕竟这些信息是给司机看的，而且司机一看就懂，无需解释。新车长宽高为 5359×2009×1879 mm，轴距 3250 mm，2+2+3 座椅布局，后备箱容积 892L（可扩容至 2002 L）。全系标配 38 个传感器，包括 896 线激光雷达，背靠 60 EFLOPS 云端 AI 算力、WEWA 2.0 架构，可首批升级华为乾崑智驾 ADS 5。搭载 1.5T 雪鸮增程器配合前后双电机，综合功率 380kW，搭载 53.4kWh 电池，CLTC 纯电续航 305km，综合续航 1320km。新车有辉光紫、深海蓝、雪域白、鎏金黑四种车身配色和韶华杏、赤茶橘、绒霞紫三种内饰颜色。搭载 MPV 版图灵平台，前双叉臂与后 H 臂多连杆独立悬架，配合双腔空气悬架及 CDC 连续可变阻尼减振器。全系标配±7° 主动后轮转向，5.35 米转弯半径，支持低速蟹行模式，有效应对狭窄路况的调头和穿行。四活塞前固定卡钳，35.8m 制动距离；500mm 涉水深度，保证暴雨积水情况下的通行能力。路面预瞄系统智能识别井盖、凹坑与弯道，智能调节悬架阻尼、刚度，以及电机扭矩，可以降低冲击与侧倾，获得莱茵晕车舒缓认证。行业首发 2 根 2400MPa+6 根 2200MPa 侧门防撞梁；全系标配 13 个安全气囊，包括全维包裹安全气囊。前环宇三联屏，12.3 英寸液晶仪表与双 17.2 英寸 3.4K 大屏组合，可选后排 21.4 英寸吸顶屏。因为鸿蒙座舱，乾崑智驾，还有图灵底盘等鸿蒙智行基础产品技术支撑，现在鸿蒙智行的车型想要出挑有点困难，但毫无疑问，无论是华为还是奇瑞，都给到了智界 V9 足够的重视，比如单独给这款车开一场声势浩大的发布会，这在整个华为体系里是极为罕见的待遇。再比如，华为和奇瑞给这款车投入了一些深刻的产品洞察去迎合和吸引智界 V9 的潜在受众，很多人可能对这样一款华为气质非常浓郁的 MPV 产品无感，但华为不在乎，华为只希望这些人快快长大并事业有成。

发布会开场，李想先是强调：其实大部分用户并不知道自己想要什么。他分享了一个故事：有一次我去买烤冷面，摊主问我还要加点什么，我开玩笑说能不能打个折。你看，用户永远想要更便宜，但摊主会告诉你，其实还能加蛋。这也是 L9 当年销量爆发的秘诀——当行业还在比拼基础参数的时候，理想顺手给车里塞进了冰箱、彩电和大沙发，交出了一份远超预期的答卷。四年过去了，全行业都开始把屏幕做大、座椅做软。当堆砌配置成为车企的常规操作，理想 L9 又走向了另一条路。如今的理想把大量精力和资金砸向底层架构，自研范围覆盖车规级算力芯片、800V 主动悬架、全线控底盘等多个领域。承载这些研发成果的，自然是今天上市的两款新车型：L9 Ultra 和 L9 Livis。首先值得肯定的是，理想没有盲目放大 L9，车长和车宽依然维持在 5.25 米和 2 米，日常开起来依旧方便。在这个好开的尺寸基础上，L9 的整车姿态还挺拔了不少。在车高仅增加 10 毫米的情况下，前机盖和水切线被大幅拉高，配上 22 寸大轮毂，原本庞大的车身显得颇为矫健；前唇那一抹金灰色镀铬饰条也收束得干净利落。L9 Livis 还采用了迈巴赫同款涂装工艺，紫金、绿金、黑金等专属双色车身，额外增加两道工序后，漆膜厚度达到传统工艺的两到三倍，日光下呈现出流动的光影质感。外观的考究还藏在细节里。全车 12 颗超声波雷达被 UWB 传感器取代，保险杠上再也看不到突兀的小圆孔；隐藏式电动踏板收起后与底盘融为一体。这不仅让外观更整洁，UWB 技术还让哨兵模式耗电量骤降 82%，开两天只耗一度电。打开带有电容防夹功能的电动车门，里面还是那个熟悉的「起居室」，只是每一项体验的上限都被抬高了。全新 L9 是全球首款四个座位都配备全尺寸零重力座椅的 SUV。Livis 版本直接标配，Ultra 版本在首销期内也会免费赠送前排两把零重力座椅。在过去，转向管柱挤占了空间，主驾想完全躺平是个大难题。而在 L9 上，线控转向释放了空间，1.8 米的大个子坐进主驾也能一键舒展躺平。望向前方，仪表台上的两块屏幕合二为一，变成了一块 29 英寸的 6K 超宽全景屏。后排的 21 英寸屏幕新增电动平移功能，手轻轻一抬，这块「神奇移动屏」就会沿滑轨移到面前，并随座椅倾角自动找平。音响功率这次飙升到了 5440 瓦。9.3.6 星环剧院音响搭载三重矩阵式低音系统，头枕音响创新采用水平对置设计，出声口藏在两侧，以往那种「有人在你脑后说话」的突兀感也随之消失。车内空调支持 7 温区控制，面部和脚部可以吹出不同温度的风；全车 29 处皮质区域都带加热功能。中控台下方的 10L 智能冰箱换了电动门，轻拍两下即可开关，不仅支持 -6°C 冷冻，内胆还能直接拆下来清洗。作为增程车，全新 L9 两个版本的纯电体验都做到了行业顶配。底盘塞进了一块 72.7 度的 5C 大电池，容量已经超过不少纯电车型，换来 420 公里纯电续航。按每天通勤 40 公里算，两周充一次电完全够用。充电速度也同步提升，配合理想自建的 8300 根 5C 超充桩，十分钟左右就能完成补能——在服务区上个洗手间、买杯咖啡的功夫，电量就能满血复活。长途出行时，这台重达 2.8 吨的全尺寸 SUV 的油耗还挺实在的：亏电油耗降至 6.3 升/百公里，满油满电综合续航达到 1650 公里。更让人省心的是，这套增程系统的保养周期延长到了 3 年或 3 万公里，还省去了过去增程车繁琐的维护。李想表示，在机械属性和维保频次上，全新 L9 已经越来越像一台纯电车了。把家装进车里，把纯电融进增程「全尺寸 SUV 一直有一个很难真正被解决的矛盾，既让一家人坐得舒适，又让自己开得尽兴。」李想说。为了同时满足这两点，L9 Livis 直接上了全套线控主动底盘。L9 Livis 的 800V 主动悬架，让每个车轮都有超过 1 万牛的支撑力，传统的防倾杆被取消，四轮完全独立。过减速带、连续起伏路或者高速变线时，液压泵和减震器会实时配合，该撑就撑、该压就压。停车休息时，这套悬架甚至能联动赛车游戏，把车厢变成 4D 模拟器。大车想开得灵活，转向系统也要升级。Ultra 和 Livis 都标配了后轮转向 + 线控转向，转弯半径缩到 5.3 米，方向盘打满不用换手，角度也能按习惯自定义。另外，Livis 还首发了 EMB 线控机械制动，刹车响应快了约 0.1 秒，120km/h 时速下刹停距离能缩短 3 米以上。而且这套系统号称终生不用换刹车油，后期省心不少。安全性上，两款车都变得更加坚固了，堡垒安全车身 3.0 用上了 2200 兆帕的高强度钢和一体式双门环。最特别的是，车内的 11 个安全气囊还能通过红外传感器识别每一位乘员的体型与坐姿，动态调节保护策略。面对极端的碰撞危险，具备主动悬架的 L9 Livis 能在 0.7 秒内瞬间抬升 50 毫米，用最坚固的底盘部位去硬刚撞击；低速倒车时的防碰撞系统也具备了识别悬空物体的能力。智能化层面，理想首发了马赫 M100——一颗专为 AI 原生设计、采用数据流架构的端侧推理芯片。L9 Ultra 版本用的是单芯片方案，算力达到 1280TOPS，配备前向激光雷达；L9 Livis 则拥有两颗马赫 M100，总算力翻倍达到 2560TOPS，同时配备了 4 颗激光雷达来实现 360° 全向感知。算力上来后，感知帧率成倍提升，乡间小路、施工路段都能提前预判，果断绕行。这颗芯片还跟去年 12 月发布的 Livis AI 眼镜实现了生态互联。交互开始变得科幻。戴着眼镜站在车外，一个手势或一个眼神，车就会自动泊出来，缓缓开到你面前。坐进车里，系统还能察觉你的情绪和状态。你今天好像心情特别好，是要带家人去顺义的公园踏青吗？这时候，它既是司机，也是伙伴。到了我们最关心的价格，两款车只拉开了 5 万元的价差。L9 Ultra 定价为 45.98 万元，竞争力很强。72.7 度 5C 电池、双腔空悬、线控 + 后轮转向、1280 TOPS 马赫芯片都是标配，首销期还免费送前排零重力座椅、智能电开门和 5440 瓦音响。L9 Livis 定价来到 50.98 万元，比此前 55.98 万的预售价低了不少。多花 5 万，换来的是专属双色车身、800V 主动悬架、带 EMB 制动的全线控底盘、翻倍到 2560 TOPS 的算力，以及 4 颗激光雷达。首销权益也公布了：6 月 30 日前锁单，现金直减 2 万元，老用户再额外补贴 1 万元。从 2015 年创立至今，理想还是习惯把消费者叫作用户。买车只是陪伴的开始，让全家人在日复一日的相处里都能开心，才是 L9 Livis 这台「硅基生命体」想要兑现的承诺。

在今天和理想全新 L9 和智界 V9 两台重磅车型同台亮相的，还有蔚来的乐道 L80。相比 L90 面向三代同堂、多孩家庭的大三排需求，L80 把重点放在大五座市场，带来了更大的乘坐空间，更灵活的装载能力，以及更适合长途、自驾、露营等场景的空间组织方式。早年间，以本田冠道、大众途观 L 为代表的燃油大五座 SUV，主要通过释放第二排的腿部空间来提升乘坐舒适性；随后，随着新能源时代的到来，诸如理想 L7、问界 M7 等车型进一步强化了智能座舱的沉浸式体验和车内的储物能力。而乐道 L80 现在则试图在行业已有的基础上更进一步，将前备舱、后备舱、乘员舱以及补能体系整合为一个完整的生态闭环，通过蔚来的工程创新能力，让中大型五座 SUV 的日常使用体验再上一层楼。乐道 L80 一共有 Pro、Max+ 和 Ultra+三个版本，整车购买价格分别为 24.28万元、25.98 万元、27.98 万元， Baas 方案价格为 15.68 万元、17.38 万元和 19.38 万元。前备舱大，后备舱更大空间，依然是乐道 L80 最核心的竞争力。乐道 L80 车长 5145mm，轴距 3110mm，理想 L7 车长 5050mm，轴距 3005mm，乐道 L80 比理想 L7 大了一圈。取消第三排之后，乐道 L80 将所有空间都留给了前两排和储物。第二排作为 L80 整台车体验的核心区域，其后排头部空间达到 1074 毫米，腿部空间宽达 1121 毫米，坐垫宽度和长度分别达到 549 毫米和 497 毫米，靠背支持最大 135 度的后倾调节。L80 提出了一个「大五座，享四座」的越级体验概念，即当车内只有四人出行时，第二排的两位乘客完全可以获得媲美大型 MPV 独立航空座椅的放松状态。而在储物空间上，乐道 L80 配备了容积为 240 升的电动前备舱，能够轻松容纳两个 20 英寸的标准登机箱、一个儿童专属行李箱及一个随身背包。前备舱的存在不仅增加了绝对装载量，更为家庭用户提供了干湿分离与物理分区的储物区域，有效改善了传统 SUV 后备舱容易被杂物堆砌得杂乱无章的痛点。与前舱相对应，乐道 L80 的后备舱在第二排座椅处于常规直立状态时，容积为 1200 升；当第二排座椅完全放倒后，整体空间可扩展至 2600 升。此外，在后备舱的常规盖板下方，工程团队设计了双沉舱结构。其中外侧沉舱容积为 106 升，里侧沉舱容积为 63 升，均配备了独立盖板。这种隐蔽式的储物空间非常适合用来存放不常用的随车工具、零散的露营小配件、充电线缆以及各类清洁用品，使得主后备舱始终保持整洁。乐道能够在车辆内部极其有限的机械空间内挤出更多的储物容积，主要得益于车辆底层平台的极高集成度。L80 充分应用了已经在 L90 上被验证过的 900V 高压架构，并配备了体积更为紧凑的前置电驱动系统。结合 30.5 度大倾角冷却模块、全栈自研的热管理高度集成模块以及一体式偏置空调箱，这些系统级的技术创新对原本需要占据前舱大量体积的机械与电子部件进行了极致的压缩与整合。空间之外，座舱内的舒适化和智能化，也是乐道 L80 强调的特色。乐道 L80 的座舱内部采用了分层布局：上层主要用于视觉延伸与氛围营造，中层集中处理人机交互与触觉反馈，下层则侧重于日常储物与抗污耐用。车内配备了支持 1600 万色调节的氛围灯系统、长达 7.2 米的环绕光带，辅以大面积软质材料包覆、金属扬声器网罩及精细的绗缝与刺绣工艺。整体风格规避了盲目堆砌昂贵材料的浮夸感，更侧重于在家庭用车的基调上打磨出温馨且不失格调的精致质感。为了提升车内的通透感，乐道 L80 车辆侧窗与前挡风玻璃的总面积约为 3.3 平方米，前后分区的全景天幕面积也达到了 1.1 平方米。同时为了应对大面积玻璃材质在强光照射下带来的隔热与防晒问题，车内配备了双电动物理遮阳帘，乐道 L80 同时配备了空气悬架系统，支持在驻车时自动降低车身高度，配合电动侧踏板与静音电吸门，优化了行动不便的老人或儿童的上下车体验。车辆的座椅总成基于专属的高端座椅平台开发，依据中国人体生理曲线大数据进行了型面贴合度优化，前排座椅支持 12 向电动调节，第二排支持大角度电动调节。同时，方向盘加热、前后排座椅加热、通风以及按摩功能均作为标配提供。此外，车内集成了一台 8.8 升容量的智能电动冰箱，足以容纳 6 罐常规可乐、2 盒鲜牛奶以及部分水果零食。神玑上车，4D 领航数字化交互与计算平台是现代智能汽车的核心大脑。L80 的屏幕矩阵由一块 35 英寸的 AR-HUD 增强现实抬头显示、一块 17.2 英寸的 3K 高清中控屏、一块 17.3 英寸的后排吸顶娱乐屏以及一块 8 英寸的后排专属场景控制屏组成。支撑这一系列复杂交互与协同调度的硬件，是一颗采用 5 纳米制程工艺的车规级高阶智驾芯片——神玑 NX9031。该芯片拥有高达 546 GB/s 的内存带宽，打破了传统分散式域控制器的物理壁垒，能够同时支持智能驾驶、智能座舱、智能底盘等多个系统跨域计算资源的统一调度。配合 SkyOS·天枢整车全域操作系统，车辆从底层代码级别打通了智驾、座舱、底盘、车身及动力网络，使各项功能摆脱了信息孤岛的状态，能够基于场景进行深度系统级协同。在智能驾驶领域，系统深度接入了蔚来世界模型 NWM（NIO World Model）。该模型摒弃了传统的规则驱动逻辑，转而通过强化学习与海量的闭环仿真训练，提升了系统对复杂未知路况的理解与博弈能力，全面覆盖城区复杂道路、城市快速路、高速公路及各类园区停车场。值得重点关注的是，L80 支持城区领航换电功能（PSP）。在导航状态下，系统能够自动规划最优补能路径，将车辆自主驾驶至换电站并完成自动化换电流程。同时，基于模型化架构构建的智能泊车系统，也支持路线的实时动态规划和对各类非标准复杂车位的识别入位。尽管定位为体型庞大的家庭 SUV，L80 在底盘的机械素质与动态调校上依然投入了极高的规格。悬架系统采用了前双叉臂与后五连杆独立悬架组合，配备了具备 100 毫米行程调节能力的智能空气悬架系统以及 CDC 智能连续可变阻尼减震器。底盘硬件与路面之间通过液压衬套进行柔性连接，并搭配了抓地力更强的鸳鸯胎。乐道 L80 还支持 4D 全域舒适领航功能。该功能能够实时结合车辆自身的传感器感知、激光雷达对前方路面的三维预瞄以及云端高精舒适图层数据，在遇到减速带、坑洼或颠簸路面之前，提前毫秒级主动调节空气弹簧的刚度与减震器的阻尼，从容化解路面冲击。除了常规的运动、舒适、后排舒享和节能模式外，车辆预设了湿地、雪地、泥地、沙地等脱困模式。在卓越的底盘参数加持下，L80 拥有高达 600 毫米的最大涉水深度，最小离地间隙最高可拉升至 260 毫米，足以轻松应对绝大多数轻越野和长途自驾游的复杂路况。而在安全方面，乐道 L80 全系标配 9 个安全气囊，除了常规的环绕式防护，第一排还配备了能在碰撞发生前预紧的主动式安全带、容积达到 137 升的狐耳式超大副驾气囊以及防止前排乘员相互碰撞的中央远端安全气囊。基于家庭属性的考量，第二排乘客的安全防护被提升至与前排同等重要的层级，配备了 14 升容积的后排侧气囊，以及长度贯穿 2920 毫米、容积高达 56 升的超大侧面安全气帘。在应对碰撞的白车身结构设计上，L80 的前部预留了长达 650 毫米的充足溃缩吸能空间，并创新性地采用了「6 纵 10 横」的笼式安全传导结构。前防撞梁选用了高强度的 7 系航空级铝合金材质，其上层主防撞梁的宽度覆盖率达到了整车宽度的 80%，能够在严苛的偏置碰撞中有效化解破坏性冲击力。侧面防护则通过环舱高强度钢、一体式门环、多腔体铝合金门槛梁以及坚固的电池舱横梁共同构建。车尾部分应用了先进的一体压铸后地板技术，配合铝合金后吸能盒与后防撞梁，为车身尾部提供了深达 1244 毫米的吸能缓冲空间，最大程度减轻高速追尾事故对后排乘客及底层电池包的结构性伤害。乐道 L80 搭载的电机能够输出 340kW 的峰值功率，百公里加速时间 5.7 秒，其百公里综合电耗被控制在优秀的 14.3kWh，85 度电池下的 CLTC 续航为 615km。当然，纯电车型的体验永远无法脱离补能体系的建设。共享蔚来庞大且成熟的充换电网络，依然是乐道 L80 在同级别竞品中最难以被复制的核心优势。为了进一步夯实这一护城河，乐道官方已确认追加投入 6 亿元人民币，用于在网络中新增 8000 多块流通电池。蔚来之前透露，早在四年前，乐道 L80 便与 L90 基于同一技术平台同步启动开发。其中，乐道 L90 的成功已经验证了产品定义的精准性。在 2025 年，乐道 L90 凭借对大三排 SUV「满载状态下行李装载空间」这一核心痛点的有效解决，在上市短短 5 个月内实现交付超 4.3 万台，成为纯电大型 SUV 市场的年度销量冠军。这不仅助力乐道品牌在 2025 年实现 107808 台的全年交付目标，更以接近 24 万元的成交均价稳居新势力品牌第一梯队。乐道 L90 和蔚来 ES8 的表现，也被行业视为大型 SUV 动力形式从增程向纯电转型的关键标志。如果说乐道 L90 成功开拓了大三排纯电 SUV 市场，那么乐道 L80 则肩负着在规模更大的大五座 SUV 领域实现突破的重任。在乐道的产品价值体系中，L90 聚焦于「在一起」的温情与责任，通过三排布局满足多孩、三代同堂家庭的聚会与通勤需求；而 L80 则更侧重于「走出去」的激情与成长，以更灵活的空间形态承载用户自驾、露营等更具探索性的生活方式。

🔍郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%🤖OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略🚀Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI📱微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆💻华为 Mate X7 官方降价 1000 元⚠️追觅造车主体约 232 万元股权遭冻结🎮国行 Switch 正式停服🧠百度成立模型委员会 BMC💰三星电子单季营收创历史纪录，半导体撑起超九成营业利润🔍海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%🖨️铠侠：2027 年内存仍吃紧、净利润同比增长 103.6%🤖微信发布「全球青少年 AI+ 小程序」洞察报告💡何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了🦿5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约📱荣耀 CEO 李健官宣机器人手机今年 Q3 发布📱vivo S60 将首发 4K 原生感 Live🚗全新理想 L9 发布：45.98 万元起🤖ChatGPT 新增 Codex 连接电脑功能📰 周末也值得一看的新闻郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%天风国际证券分析师郭明錤昨日披露，苹果已在 Intel 18A-P 系列制程上为低阶及旧款 iPhone、iPad 与 Mac 处理器立项，订单中 iPhone 约占 80%。流片规划显示，苹果预计 2027 年在 Intel 放量，2028 年续增，2029 年进入衰退期。Intel 方面，2027 年良率目标为稳定达到 50% 至 60% 以上，量产时程与出货规模目前仍未明朗。不过，郭明錤也指出，即便 Intel 顺利出货，台积电在苹果供应体系中仍将占 90% 以上比重。郭明錤认为，苹果此举是有意将 Intel 培养为长期备选供应商。AI 对台积电先进制程资源的占用将持续扩大，苹果需要在议价能力尚存时提前布局。与此同时，苹果也在同步评估 Intel 其他先进制程技术。OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略据 WIRED 报道，OpenAI 昨日向员工宣布新一轮组织调整，联合创始人兼总裁 Greg Brockman 正式接管产品战略。报道指出，OpenAI 将 ChatGPT、编程智能体 Codex 及面向开发者的 API 业务整合进同一核心产品团队，目标是将三者合并为「统一的体验」。Codex 原负责人 Thibault Sottiaux 升任核心产品与平台主管，统管消费、企业和开发者三条线；ChatGPT 原负责人 Nick Turley 转任企业产品改版；前 Instagram 副总裁 Ashley Alexander 接任消费产品负责人。Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI据《金融时报》报道，Anthropic 已就一轮 300 亿美元的融资达成条款，融资完成后公司估值将约达 9000 亿美元，较今年 2 月初 3500 亿美元的估值接近翻三倍。本轮融资由 Dragoneer Investment Group、Greenoaks Capital、红杉资本（Sequoia Capital）和 Altimeter Capital 联合领投，报道称，每家领投机构预计出资不低于 20 亿美元。微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆CCF 计算机博物馆（CCM）昨日收到腾讯微信团队捐赠的一批早期历史实物，包括微信第一行后台代码原始存档、早期开发设备 iMac ME086CH/21 寸一体机与戴尔塔式工作站，以及 Apple iPod Touch 3、诺基亚 5800XM 等早期测试终端。CCM 表示，由于移动产品迭代速度极快、数字资料易被覆盖删除，大量见证移动互联网起步的实物未能得到系统保护。此次捐赠呈现了微信从 0 到 1 的研发链条，是中国移动互联网早期发展的核心实物见证。CCM 主体建筑已全面落成，拟于今年 9 月开馆，目前持续面向社会征集各时期计算机相关藏品。保时捷称被中国车企抄袭「有些许无奈」，罗永浩：保时捷自己就有前科保时捷中国传媒公关副总裁唐风靓日前在浙江「易境思论坛」上表示，过去几年保时捷激发了高度统一的汽车设计审美，被中国车企模仿「除了荣幸也有些许无奈」。罗永浩随后在微博发文回应，称保时捷说「无奈」可以理解，但称中国车企抄袭得「高度统一」则是吹牛。他同时指出，保时捷历史上也曾抄袭捷克斯洛伐克品牌太脱拉（Tatra）的设计，并以 1965 年大众汽车向太脱拉支付 100 万西德马克庭外和解赔偿金为据，直言「说到流氓程度，谁能跟保时捷比」。此番言论随即引发讨论。有网友指其犯了「诉诸伪善」的逻辑错误，罗永浩强硬回怼，称自己已明确承认抄袭者「也是流氓」，并非为抄袭行为开脱。华为 Mate X7 官方降价 1000 元华为昨日宣布，旗下两款折叠屏手机同步开启降价优惠：Mate X7 全系直降 1000 元，优惠后售价 11999 元起；Mate X6 直降 3000 元，优惠后售价 9999 元起。追觅造车主体约 232 万元股权遭冻结企查查 App 显示，追觅科技创始人俞浩实控的星空计划（上海）汽车科技有限责任公司近日新增一则股权冻结信息，股权数额约为 232.25 万元，股权所在企业为上海星空计划进出口贸易有限责任公司。去年 8 月，追觅科技宣布进入汽车领域。截至目前，追觅汽车团队规模约千人，研发人员占比达 70%，自研覆盖智能汽车多个核心领域。此次造车主体出现股权冻结，官方尚未就相关情况作出回应。国行 Switch 正式停服国行 Switch 的 Nintendo e 商店已于昨晚正式停服，至此，国行 Switch 将无法再进行联网对战或下载游戏。官方在停服前向相关玩家发送提醒，建议玩家在停服前完成所有已购游戏、DLC 及兑换码的下载与激活；删除游戏数据或对主机进行格式化后，数字内容将无法恢复。今年 3 月 31 日，国行 Switch 的游戏售卖服务已率先停止；而根据官方此前公告，今年 12 月 31 日起，国行 Switch 主机及配件的维修服务也将全面终止。这意味着，国行 Switch 的官方售后支持将在今年底彻底画上句号。百度成立模型委员会 BMC据《科创板日报》报道，百度近日内部设立百度模型委员会（Baidu Model Committee，BMC），成员由对大模型有深刻理解的年轻研究员构成。基础模型研发部（BMU）与应用模型研发部（AMU）均向该委员会汇报。知情人士透露，新成立的 BMC 将全面统筹百度大模型业务，覆盖路径探索、技术规划、模型训练到业务落地的完整链条，目标是实现模型从技术到应用的一体化管理。百度方面表示，此举有助于集中优势资源，持续建设并强化人工智能方向的技术优势。三星电子单季营收创历史纪录，半导体撑起超九成营业利润三星电子昨日发布 2026 财年第一季度（1 月 1 日至 3 月 31 日）业绩报告：期内实现合并营收 133.9 万亿韩元，同比增长 69%，环比增长 43%；营业利润 57.2 万亿韩元，同比增长 756%，环比增长 185%，营业利润率达 42.8%。净利润为 47.2 万亿韩元，净利率为 35.3%。基本每股收益为 7123 韩元。半导体（DS）：营收 81.7 万亿韩元，同比增长 225%，环比增长 86%；营业利润 53.7 万亿韩元。其中存储业务营收 74.8 万亿韩元，同比增长 292%，环比增长 101%，为本季度核心增长引擎；设备体验（DX，含手机与消费电子）：营收 52.7 万亿韩元，同比增长 2%；营业利润 3.0 万亿韩元，同比下降约 1.8 万亿韩元；显示（SDC）：营收 6.7 万亿韩元，同比增长 14%；营业利润 0.4 万亿韩元，同比下降 0.1 万亿韩元；Harman：营收 3.8 万亿韩元，同比增长 12%；营业利润 0.2 万亿韩元。下季度展望，公司预计 AI 基础设施扩张驱动的服务器内存及 SSD 需求将持续强劲，计划向客户交付 HBM4E 首批样品以巩固技术领先地位；手机业务方面，Galaxy S26 系列发布效应趋于消退，盈利能力预计承压。海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%昨天，海力士于发布今年第一季度（截至 3 月 31 日）业绩，按 K-IFRS 口径，期内实现营收 52.58 万亿韩元，同比增长 198%；营业利润 37.61 万亿韩元，同比增长 405%；净利润 40.35 万亿韩元，同比增长 398%。本季度营收首次单季突破 50 万亿韩元，创历史新高。主要财务指标如下：毛利润 41.68 万亿韩元，毛利率 79%，同比提升 22 个百分点；营业利润率 72%，创历史单季新高，环比提升 14 个百分点；EBITDA 41.34 万亿韩元，EBITDA 利润率 79%，同比增长 284%；净利润率 77%，其中非经营利润 14.01 万亿韩元，包含投资资产估值收益 9.94 万亿韩元。铠侠：2027 年内存仍吃紧、净利润净同比增长 103.6%铠侠控股昨日发布截至今年 3 月 31 日止 2026 财年业绩，期内实现营收 2.3376 万亿日元，同比增长 37.0%；按 IFRS 口径，归母净利润为 5544.9 亿日元，同比增长 103.6%。SSD 与存储业务营收 1.3626 万亿日元，同比增加 37.5% 至 3715 亿日元；智能设备业务营收 7599.8 亿日元，同比增加 51.6% 至 2588 亿日元；其他业务（含 SD 卡、U 盘等零售产品及向闪迪集团的销售）营收 2150.1 亿日元，同比基本持平。另据财联社报道，铠侠表示，AI 服务器客户的订单量持续走强，以美元计价的 NAND 价格在一季度上涨逾一倍，并预计 2027 年度 NAND 存储供应仍将维持紧张态势。微信发布「全球青少年 AI+ 小程序」洞察报告昨天，腾讯研究院联合微信发布《全球青少年 AI+小程序洞察报告》，披露了全球基础教育阶段群体利用 AI 工具与小程序进行应用开发的各项核心数据。据报告数据显示，自 2023 年 11 月微信小程序教育平台正式对外开放以来，已吸引全球 7954 所学校的逾 8.7 万名学生参与，累计创建小程序项目超 28.7 万个。在开发流程中，师生调用 AI 辅助创作每年消耗 Token 量超过 500 亿，支撑约 375 万次人机交互。此外，微信连续四年举办的小程序全球创新挑战赛共产生 7430 个作品，其中第四届单届提交量达 4900 个，呈显著爆发趋势。开发者画像层面，62% 的项目由单人独立完成，最小获奖开发者年仅 5 岁。地域活跃度数据显示，中国香港、北京及中国澳门位列前三，且项目协作已出现跨国别开发特征。在应用方向上，05 后开发者主要聚焦于无障碍辅助、碳足迹、适老化关怀及健康管理等社会实用场景。💡 何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了据睿见 Economy 报道，小鹏集团董事长、CEO 何小鹏昨日在「第十八届轩辕汽车蓝皮书论坛」上表示，「什么时候我们看到北京、上海车展上不再有 150 款新车的时候，中国汽车的内卷可能就结束了。」何小鹏表示，过去十年，中国新能源汽车经历了两个重要转变：一是核心驱动能源从传统化石能源切换至电化学能源；二是行业全面进入初步普及的智能化阶段。他称，十年前几乎没有人相信中国汽车能在全球市场取得如今的成绩，并实现规模化的全球化布局。何小鹏认为，下一个十年将是电化学能源与数据能源的双重叠加阶段，芯片和模型将成为新的「引擎」，汽车与机器人的能力将在这十年内深度整合。5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约ROG 与 XREAL 联名推出的电竞 AR 眼

🤖曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini🌐马耳他将向全体公民免费提供一年 ChatGPT Plus📱小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务🚀目标估值 1.75 万亿美元，SpaceX 最快 6 月上市📋arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚💻卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元🚗蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌🔍需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位☁️最低 1 元 40 万 Token，三大运营商推出 Token 套餐🚙特斯拉 Robotaxi 碰撞事故报告解封💡知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值🚗小米 YU7 GT 火山灰配色公布💰ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构🧠马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布🤖字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本📱腾讯 vibe coding 产品「吐司」即将上线🛍️蜜雪冰城上架老黄同款蜜桃四季春，销量暴涨🛑中国银行关停信用卡独立 App，服务整合进主 App曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini昨晚，彭博社记者 Mark Gurman 在最新的 Power On 新闻通讯中透露了更多关于苹果即将发布的独立 Siri 应用的细节。隐私设计是此次新应用的核心卖点之一。Gurman 透露，该应用将内置对话自动删除功能，用户可在设置中选择对话保留时长：30 天、1 年或永久保存。苹果预计在今年的 WWDC 继续主打隐私保护，并强调其 AI 体验不含广告。模型方面，苹果已将新 Siri 的 AI 模型底层大量替换为 Google 的 Gemini，并将部分服务运行在 Google 的云基础设施之上。而此前有消息称，苹果和原本的合作对象 OpenAI 关系出现裂痕，甚至有可能对簿公堂，原因是 OpenAI 认为双方合作的结果不及预期。值得注意的是，尽管新的 AI Siri 已经跳票约两年，Gurman 称苹果内部测试版本仍将其标注为「Beta」，并可能沿用至其正式发布。马耳他将向全体公民免费提供一年 ChatGPT PlusOpenAI 日前宣布，与马耳他政府签署合作协议，向该国公民及居民免费提供为期一年的 ChatGPT Plus 订阅服务。据悉，居民须先完成由马耳他大学开发的免费在线课程「AI for All」，课程内容涵盖 AI 的基本概念、能力边界及负责任使用方式，通过后即可向马耳他数字创新局申请免费订阅资格。小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务小米 REDMI 产品经理「笋寸」日前在微博发文建议，近期不打算换手机的用户可以考虑等待换电池，并提示部分机型在更换新电池后容量可能还会有所提升，参考小米 13 系列的现行政策。笋寸进一步说明，小米正积极推动存量用户更换采用金沙江技术的大容量电池，小米 13 系列已率先上线相关服务。她还透露，具体支持机型「得等等」。此前，小米 13 系列已全系上线电池升级服务，支持小米 13、小米 13 Pro、小米 13 Ultra 三款机型升级更大容量电池，费用为 189 元，其中电池升级费 149 元、人工费 40 元。目标估值 1.75 万亿美元，SpaceX 最快 6 月上市据路透社报道，SpaceX 已加快 IPO 时间表，计划最早于 6 月 12 日在纳斯达克挂牌上市，股票代码为「SPCX」，较此前预定的 6 月下旬提前约两周。据三位知情人士透露，公司计划最早于本周三公开招股说明书，路演定于 6 月 4 日启动，股票发行最早于 6 月 11 日完成定价。时间表提前的主要原因是 SEC 对招股材料的审核进展快于预期。此次 IPO 预计募资约 750 亿美元，目标估值约 1.75 万亿美元，有望成为史上规模最大的股票发行。承销行阵容包括摩根士丹利、美国银行、花旗集团、摩根大通和高盛。arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚近日，预印本学术论文平台 arXiv 宣布，若作者提交的论文中包含无可辩驳的 AI 生成内容证据，将被封禁账号一年，并在此后要求所有提交须先经过同行评审期刊接受方可上传。arXiv 计算机科学板块主席 Thomas Dietterich 在 X 发文表示，若论文中出现「幻觉引用」、大语言模型元注释（如「以下是 200 字摘要，是否需要修改？」）等明显迹象，即可认定作者未对 AI 生成内容进行核查。arXiv 目前由康奈尔科技学院管理，今年 7 月将独立运营，成为一家非营利法人机构。康奈尔科技学院院长 Greg Morrisett 表示，独立后 arXiv 将能从更广泛的捐助方获得资金支持，以应对「AI 垃圾内容」带来的运营压力。卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元近日，小米集团合伙人、总裁卢伟冰在「Max 一夏」户外露营直播中表示，玄戒芯片今年肯定会推出迭代款，并称这将是「一块非常强的芯片」，会搭载在一款「非常优秀的产品」上。他同时提醒外界，目前网络上流传的各类相关传闻可信度不高，具体规格暂不便披露。卢伟冰还透露新一代玄戒芯片的应用范围将不再局限于智能手机，后续计划广泛搭载于小米旗下多类智能终端设备。关于手机定价，卢伟冰在直播中作出判断，认为今年下半年尤其到年底，部分国产旗舰直板手机价格可能突破 1 万元人民币。他预计，内存涨价的压力至少延续至 2027 年底，甚至到 2028 年。对于即将发布的小米 17 Max，他表示定价仍在讨论中，将在发布会上正式公布，但小米的原则是在同档位提供「最物超所值的产品」。蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌据睿见 Economy 报道，在日前举行的第十八届轩辕汽车蓝皮书论坛上，蔚来创始人、董事长兼 CEO 李斌线上出席，在对话中将萤火虫比作 iPhone，称其已成为「一款非常高品质的产品」。他强调，萤火虫在某种意义上超越了汽车的工具属性，演变为「一个生活方式品牌、一个时尚品牌」，并开始形成独特的品牌文化。他表示，萤火虫在用户群体上打破了年龄与购车预算的双重限制，「大学刚毕业的年轻人想买也能买，很有钱的老板想自己开车送小孩上学也完全没有问题」。需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位据《商业内幕》报道，前向部署工程师（FDE）的招聘需求在一年内激增逾 700%，正成为 AI 浪潮下科技行业增长最快的职位之一。招聘平台 Indeed 数据显示，今年 4 月该职位招聘帖数量达 5330 条，较去年同期的 643 条增长约 729%。Anthropic、OpenAI、Palantir、Stripe 及 Google Cloud 均在积极招募该职位人才。FDE 的核心职责是协助企业将 AI 整合进业务流程，并根据客户需求定制开发 AI 工具。薪酬方面，Indeed 数据显示该职位年薪约为 17 万至逾 20 万美元。最低 1 元 40 万 Token，三大运营商推出 Token 套餐中国电信昨日推出系列试商用 Token 套餐，分别面向开发者及中小微企业、个人及家庭用户提供「Token+连接+安全」一体化服务，均提供三档套餐，面向个人及家庭用户定价 9.9 元/月起 1000 万 Token；面向开发者及中小微企业 39.9 元/月，1500 万 Token。两类用户均可选购宽带上行提速包（上行速率提升 50 Mbps，100 元/月）及智能体安全管家服务（5 元/月）。个人家庭用户还可额外选购 5G-A 优享包（含 10 GB 流量，上行速率最高 300 Mbps，20 元/月）。联通、移动两大运营商也先后推出了各自的 Token 套餐：移动：中国移动上海公司于 5 月 17 日世界电信日发布会上推出 Token 通用服务，定价 1 元 40 万 Token。联通：中国联通上海分公司于 5 月 16 日发布全栈国产化「国芯国模国云」安全基座，推出多档 Token 产品及融合套餐。特斯拉 Robotaxi 碰撞事故报告解封近期，特斯拉正式解密并公开了向美国国家公路交通安全管理局（NHTSA）提交的 17 起「Robotaxi」自动驾驶碰撞事故的完整未脱敏报告。据 Electrek 报道，这 17 起事故发生于 2025 年 7 月至 2026 年 3 月，均涉及在得克萨斯州奥斯汀进行「Robotaxi」测试的 2026 款 Model Y，且测试时均有安全员在场。伤亡数据显示事故整体较轻微：13 起仅造成财产损失，2 起无人员受伤，1 起轻伤未入院，另有 1 起轻伤需入院治疗（系车辆在路口以 2 英里/小时蠕行时被后方 SUV 追尾）。官方报告表明，大部分事故并非特斯拉自动驾驶系统（ADS）的过失，多为车辆在红绿灯、停止标志或交通拥堵中完全静止时，被后方人类驾驶员追尾或侧擦。但数据同时也暴露出系统与后台干预的缺陷：两起事故发生在远程操作员接管期间，导致车辆分别以 8 英里和 9 英里的时速撞上金属围栏与施工路障；另有数起事故显示 ADS 存在空间感知死角，车辆在转弯或倒车时撞上了金属链条、拖车挂钩、木制电线杆及路缘石。💡 知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值据「听筒Tech」报道，知乎创始人兼 CEO 周源日前在第十二届新知青年大会上表示，AI 时代真实创作者的价值不可替代。AI 每分钟都在创造奇观，但 AI 时代仍然需要新知。而新知来自真实的人。每一个真实的人，都拥有不可替代的价值。大会主论坛上，多位嘉宾就 AI 与人类创作的关系展开讨论。复旦大学人文学者梁永安提出「科技有所为，如何让自己有所不为？需要人的尺度和人文属性」；作家张春称「不完美就是我们的完美」。周源在大会结尾总结道：「技术越发达，有温度、有深度的真人创作就越宝贵。」小米 YU7 GT 火山灰配色公布近期，小米汽车再度放出 YU7 GT 的预热图片，公布了新配色以及新车的内饰信息。其中小米 YU7 系列将加推一款名为「火山灰」的配色，官方介绍：灵感来自晨雾下的火山地貌，高纯度灰调，叠加纳米级矿物颗粒，呈现细腻而有层次的金属质感。同时新车内饰也迎来全新的「赛道红」：采用红黑撞色，提供专属刺绣。官方表示，小米 YU7 也会提供「火山灰」外饰配色，并且现在车厘子红的 YU7 GT 已经在部分小米汽车门店展示了。ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构近日，OpenAI 面向美国 ChatGPT Pro 用户推出了个人财务管理功能，支持通过 Plaid 连接超过 1.2 万家金融机构的银行账户和信用卡。连接账户后，ChatGPT 会自动读取用户的真实余额、消费流水和负债数据，生成一张覆盖投资、日常开销、订阅和近期账单的财务面板。用户提问时，模型可直接翻出账单历史，找出哪些类目容易超支，并给出具体的削减方案。此外，用户还可以告诉 ChatGPT 一些账外信息，比如「我在攒钱买车」或「我还欠父母一笔钱」，系统会将这些内容保存为财务记忆，用于后续对话。马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布马斯克昨日在 X 发帖披露，SpaceXAI 正在研发的 1.5T 参数 Grok 基础模型第 9 版（V9）已完成训练，预计 3 到 4 周内正式发布。现役公开版本 Grok 4.3 基于 0.5T 参数的 V8 基础模型，目前每隔数日迭代一次。V9 将参数量提升至 1.5T，马斯克称其为「重大升级」。在帖子的结尾，马斯克更是用用「banger」（王炸）一词来形容对这一版本的预期。字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本字节跳动 Seed 团队近日开源扩散

直到 2026 年五月，我们终于等到一整轮的影像旗舰登场完毕——小米做出了徕卡一瞬、vivo 打造了照片与视频兼具的 V 单、OPPO 在双长焦方案上带来了惊喜。在这轮旗舰更新潮的最后，华为提前将新一轮的 Pura 90 Pro Max 端上了桌。老实说，第一眼看硬件，心里是有落差的。去年 Pura 80 Ultra 上颇为惊艳的「一镜双目」结构不见了，取而代之的，是目前行业内最稳妥也最常见的大底长焦方案。不过，在细细体验后，我认为 Pura 90 Pro Max 藏了一个比纯拼硬件更有意思的新解法——XMAGE 智拍。全自动，都省心过去这几年，我系统性地梳理过华米 OV 的影像架构。虽然各家在硬件选择上各有理解，但大方向出奇一致，都在极其有限的机身空间里死磕传感器面积，比拼谁的光圈更大、进光量更多。在硬件上奋发图强理所应当。但在摄影中，不错的硬件基础仅仅是第一步。想要拍出一张好照片，构图与后期同样重要——构图保证一张照片在视觉上是否平衡、主体是否明确；而后期则决定了这张照片视觉质感和风格化。但这两者，向来都是具备门槛的手艺。而华为在 Pura 90 Pro Max 上，打算一口气解决这两个问题。熟悉他们产品线的朋友或许有印象，去年发布的 Pura 80 Ultra 上，华为试水了 AI 辅助构图。AI 辅助构图功能相对基础，在实际拍摄中主要起一个引导和提醒的作用。到了 Pura 90 Pro Max，这项技术完成了从「辅助」到「主导」的蜕变，并定名为「XMAGE 智拍」。剥开表层去探究 XMAGE 智拍的底层逻辑，是一套完全由端侧算力驱动的本地化影像流水线，将拍摄过程拆解为三个同时进行的模块——基于实时语义分割的自动构图、无视物理镜头限制的智能变焦，以及针对画面的色彩重塑。也就是说，在这个功能的加持下，Pura 90 Pro Max 可以自动识别主体、变焦构图，最后还排列出一系列适合当前场景的 XMAGE 风格供你选择。值得一提的是，XMAGE 智拍不需要网络连接，也无需将庞大的图像数据传回云端，全部基于端侧 AI 运转，实时识别画面里的主体轮廓、建筑线条以及光源分布状况，随后根据底层训练好的审美模型，给出一套包含最优构图、变焦和精细色彩方案的组合拳。繁复的专业参数在这里黑盒化，过去需要人脑去判断的构图和调色工作，全权交给了本地算法。除此之外，Pura 90 Pro Max 还顺势推出了 AI 姿势推荐，这个功能同样通过语义分割识别，可以理解画面中的了恩物、环境、姿势与背景，随后直接在取景框中勾勒出一个等比例的线框轮廓。拿手机的人只需按图索骥，引导模特贴合画面上的姿势即可。遇到不满意的动作能随时刷新，它甚至支持导入社交网络上保存的样片，一键提取动作精髓。听起来确实很厉害，但这套完全依赖端侧计算跑出来的流水线，究竟能不能应对现实世界里复杂多变的光线，还需要把它带到真正的街头去寻找答案。AI 有所能，也有所不能来到一家咖啡馆，墙上的多层光环灯饰引起了我的注意。举起手机，XMAGE 智拍自动变焦到 2×，将墙壁上的光环作为画面主体。这个选择可以说中规中矩，光环圆形处于画面中下方，整体构图稳定和谐。同时，操作台上颇具金属质感的咖啡机也是个不错的主体，XMAGE 智拍精准识别到我的拍摄意图，变焦到 192mm，保持咖啡机的中置。由于我想保留更为还原的室内光线与色彩，这张照片并没有套用 XMAGE 智拍推荐给我的任何色彩风格。用它拍了几组照片后，我认为 XMAGE 智拍是一个非常用户视角的功能——在传统的拍摄习惯里，手机焦段是一件需要精打细算的事。大脑会本能地去贴合物理镜头的原生焦段，生怕落在中间焦段会损失画质。但 XMAGE 智拍完全抛弃了执念。它不在乎用的是几倍变焦，转而完全为画面考量：当前这个画面，怎么裁切才最好看。哪怕不可避免地要损失一部分边缘像素，只要能让视觉中心突出、比例和谐，它就会果断动手。这种做法很实在，一切都在为了最终的出片服务。走出室外回望咖啡馆，绿植与橙色外墙交相呼应，XMAGE 智拍自动截取出色彩冲击力最强的部分，同时准确识别出咖啡馆建筑主体，两种大色块在画幅中占比相似，平衡而美观。再抬头，夏季树木繁茂，XMAGE 轻微放大焦段，让树枝形成天然的框架构图，广州塔正悬于画面正中间。在这组照片的拍摄中，我还察觉到了一个很有意思的技术细节。按下快门时，你会发现滤镜加载的速度，远大于构图改变的速度。这是因为滤镜往往只需识别画面后套用底层的色彩映射，而构图则需要端侧算力去实时识别画面里的建筑线条、光源分布和人物轮廓。计算量完全不在一个量级。更有意思的是，智拍套用滤镜的方式非常精细，除了 XMAGE 风格选择外，还会对调色盘做精细调整，确保风格对当前环境的适配度。当然，这套系统也有自己的脾气。在实际走街串巷时，XMAGE 智拍有时候会自动跳出来接管画面，有时候又毫无动静，需要你手动去点击唤醒。就算是手动唤醒的情况下，XMAGE 智拍也有几率出现风格已经选好，但当前画面无法找到最佳构图的情况。这种情况频繁出现在极其繁杂的环境中，由于元素过于混乱、信息量爆炸，算法的确很难从中剥离出合适的构图。当然，换个角度想想，也许此时的 1× 视角，就是当前环境的最佳构图了。万物皆可 AI，现在轮到摄影了185 年前，达盖尔发明银版摄影法，留住时间成为一种特权；2011 年，胶片巨头柯达传出破产消息，智能手机开始野蛮生长，把镜头塞进了普通人的口袋。回顾整部影像史，其实就是一部打破特权、技术走向大众的普惠史。到了今天，移动影像的狂飙突进，撞上了一堵无形的墙。手机镜头模组越做越大，与之对应的，是最近行业里一个颇具深意的传闻：受制于高昂成本与物理极限，下一代影像旗舰的「超大杯」，大多要面临停更了。这条路为什么走不通了？因为影像旗舰突飞猛进的背后，藏着一个被刻意回避的死结。决定照片质量的，永远是镜头后面的脑袋。普通人有懂美的眼睛，却跨不过光圈快门和构图比例的门槛。把一台堆满顶级硬件的手机，递给一个毫无基础的人，他按下的快门，大概率得到一张平平无奇的随手拍。顶级硬件如果只服务于一小撮懂摄影的人，无疑就会变成一种伪命题。在同行集体踩下刹车的时间点，华为提前把 Pura 90 Pro Max 端上了桌，而 XMAGE 智拍，就是破局的方法。在此之前，行业里已有不少 AI 落地，比如 vivo 让四季流转，OPPO 消除反光。只不过，它们大多是按下快门后的后期创意和修补。XMAGE 智拍往前跨了一大步，让算法成为整个拍摄动作的主导。这种做法谈不上是什么艺术层面的颠覆，但确实管用。降低门槛，让完全没有摄影基础的人，也能相对轻松地把一台旗舰机用出该有的样子。好的技术绝不该是一座孤岛。曾经的智能手机让人们拥有了「随时拍」的自由，现在的 AI 则进一步赋予了大众「拍得好」的能力。这或许是打破当前僵局的一剂良药，也是影像超大杯物尽其用的方法。

过去几年，20 万级新能源 SUV 市场像一条突然变宽的河道。一边是家庭用户对空间、续航和舒适性的刚性需求，另一边是智能化、性能化配置不断下放。过去很多只会出现在 30 万元甚至更高价位车型上的配置，如今开始像雨水一样往下渗透。激光雷达、空气悬架、大模型智驾、零重力座椅，这些原本带有「高配」意味的词，正在成为新一轮价格带竞争里的筹码。肩负着广汽集团高端化重任的昊铂品牌在今年广汽集团重振旗鼓的趋势下，也顺势重新梳理产品架构，昊铂 A800 旗舰轿车前段时间开始上市，用到了乾崑智驾和鸿蒙座舱来增强产品竞争力。紧接着，面向 20 万价位带的 SUV 昊铂 S600 也来了，提供纯电 660 Max、纯电 660 Max+、增程 230 Max、增程 215 四驱 Ultra+ 四个版本，官方预售价为 20.99 万至 22.99 万元，预售权益价为 18.89 万至 20.89 万元。新车定位为「新豪华智慧运动 SUV」，提供纯电和增程两种动力形式，车身尺寸为 5015×1933×1700mm，轴距 2936mm。相比于品牌初创时期的雄心勃勃，昊铂现在显然务实不少，它试图把昊铂 S600 做成一个配置密度很高的样本：性能、智驾、座舱、舒适、安全和权益，几乎都被塞进「满配」这个关键词里。昊铂 S600 请了 2 个代言人，一个代言「优雅」，一个代言「运动」过去提到中大型 SUV，很多人的第一反应是「稳」：空间要大，坐姿要高，后排要舒服，最好还能承担家庭出行里那些琐碎但真实的场景。比如周五晚上从公司直接出发，后备厢里塞进折叠椅、露营桌、孩子的滑板车和一箱零食；又或者长途返乡时，后排老人希望能把座椅放倒一点，孩子希望看动画，副驾则希望在高速上补一觉。对于这类用户来说，车首先不是一个「驾驶机器」，而是移动客厅、临时休息室，也是家庭生活的第二个容器。昊铂 S600 没有放弃这部分的诉求，或者说，这些配置已经在国产中高端 SUV 上成为一种基本素养，昊铂 S600 拥有 90% 得房率、750mm 二排腿部空间、5.316 平方米整车玻璃面积，以及 1292L 二排拓展空间；舒适配置上，同级唯一 18 点按摩零重力女王副驾、143° 躺角后排电调座椅、22 扬 ADiGO SOUND 音响、五块双层夹胶静音玻璃等配置也都被纳入其中。但昊铂 S600 又没有把自己完全包装成一台「沙发车」。它的另一面，是更强调姿态和操控的运动属性，所以昊铂又反复出现「跑山王」这个表达，显然昊铂希望把一台五米多长的大尺寸 SUV 做得不那么笨重，让它不只是适合城市通勤和家庭旅行，也能在山路、弯道、高速超车等场景中保留一定的驾驶参与感。从造型上看，昊铂 S600 采用的是一种前低后高的轿跑溜背姿态，超过 5 米的车长和近 3 米轴距提供了视觉体量，双层鸭尾、宽体车尾、20 英寸运动轮毂等设计元素，则试图削弱中大型 SUV 常见的厚重感，从侧面看它确实不像一般的家用大型 SUV 那样笨重方正，更像是专为年轻家庭所设计。所以这一次昊铂也请了 2 个代言人，英格兰传奇足球明星迈克尔·欧文作为「运动代言人」，知名演员陈数作为「优雅代言人」，两位代言人分别代表这一款车的一种特质，也是试图同时吸引家用取向和个人运动取向的潜在消费者。把四驱、高阶辅助驾驶、双腔空悬都放进 20 万级别在最高配的增程 215 四驱 Ultra+ 版本上，昊铂 S600 同时提供双电机四驱和双腔空气悬架。昊铂表示，这是行业首次在这一价位段同时放入双电机四驱和双腔空气悬架。另外，昊铂 S600 还全系标配前双叉臂后五连杆独立悬架、SDC 可变阻尼减振器和鹰爪系统 2.0，并提供雪地、湿滑、脱困三种路面辅助模式。一台大 SUV 在城市高架上并线，需要车身跟随方向盘动作，而不是像大船一样慢半拍；周末跑山时，连续弯道考验的是侧倾控制和车身收敛；雨天压过积水路面，轮胎抓地力和车身稳定系统要在驾驶者还没反应过来之前，先把车身姿态拉回来。在这些场景下，双腔空悬能够有效提升驾驶体验和驾驶信心。动力部分，昊铂 S600 提供纯电和增程双动力。其中增程四驱版综合功率达到 370kW，满电零百加速 4.3 秒，亏电零百加速 5.6 秒。广汽星源增程技术的持续发电功率为 85kW，官方还提到新车在亏电状态下可实现 165km/h 速度持续刷圈 2 小时，亏电极速 193km/h。增程车过去最容易被质疑的地方，在于「满电一条龙，亏电一条虫」，电量充足时像电车，电量不足后则容易出现动力衰减、油耗升高、发动机噪音变大等问题。昊铂 S600 对 85kW 持续发电功率和亏电性能的强调，其实是在回应这个痛点，一台增程 SUV 不能只在发布会数据里快，也要在长途、高速、爬坡和亏电场景下保持基本的体面。在智能化层面，昊铂 S600 的策略也比较直接：尽量把高阶配置做成全系标配。新车全系标配激光雷达、高通 SA8650 智驾芯片、27 个高精度感知单元，并搭载融合 Momenta R6 强化学习大模型的广汽星灵智行 ADiGO GSD 3.0，支持城区和高速快速路场景。资料还提到，昊铂 S600 下半年将首批升级 Momenta R7 强化学习世界模型。智能座舱方面，新车搭载高通 8295P 座舱芯片、17.3 英寸 3K 中控屏、27 英寸 W-HUD，并内置 11 个 AI 智能体。相比过去车机只负责导航、音乐和语音控制，现在的座舱正在变得更像一个移动终端：它不仅要响应指令，还要理解场景，甚至提前知道用户在车里可能需要什么。下班后坐进车里，导航自动给出回家路线，HUD 把车道级信息投到视线前方，孩子坐在后排看动画，车机根据车内温度和空气质量调整新风，座椅按摩开始工作……这些在新能源汽车上可以实现的体验，对于从 20 万元档位传统燃油车换购过来的用户来说，无疑是跃升式的。广汽集团有着庞大的用户基本盘，这些人群想要增购换购新能源车型的话，广汽自然希望用产品承接他们的需求。就像 AI 说的那样，我在这里，不偏不绕，不躲不逃，稳稳接住你。某种程度上来说，过去几年广汽集团新能源产品的建设，过于重视 B 端，对 C 端用户的吸引力不够，并没有「稳稳接住」用户的需求。对消费者来说，20 万级新能源 SUV 已经足够拥挤，智己 LS6、极氪 7X、小鹏 G7 等车型都在这个区间形成压力。昊铂 S600 要突围，关键命题并不是「它是不是拥有很多配置」，而是这些配置能否被组织成一种清晰的用车体验被感知到：城市里安静从容，山路上不显笨重，长途时不焦虑，家人坐进去愿意放松，驾驶者开起来也不觉得无聊。

龙虾之父一个月消耗 6030 亿 Token，总花费金额高达九百万人民币。移动联通电信，三大运营商都在推 Token 套餐，199 送千兆宽带还有 1 亿 Token，了解一下？从硅谷到国内大厂，Tokenmaxxing 成为公司的主流，谁消耗 Token 多，谁就是 AI 时代的好员工。00 后校友向母校捐赠 20 亿 Token，被网友调侃按 DeepSeek 5 元/亿 Token 计算，只要 100 元。▲图片来自新浪财经Token 在半年内完成了一次身份跃迁：从技术术语，到 KPI，到话费套餐，到捐赠货币。它成了 AI 时代的「度量衡」，唯一的问题是，没人说得清它到底在度量什么。我们自己买 Token，用公司的 Token，部署了一堆 Agent，代码、论文、周报都是 Token 烧出来的。而另一边是，大厂的员工由于 Token 消耗排行榜的原因，开始拿着公司的 Token 处理私事、玩游戏、开发数十个没什么用的子 Agent 来提升自己的排名。「回报」这件事很难量化，但「使用量」可以量化。于是所有人都选择了那个容易量化的东西。这不是 AI 时代的新问题，这是管理学的老病。用 AI 消灭狗屁工作的公司，正在制造新型狗屁工作亚马逊，那个裁员裁到大动脉，把自己的网站都变成 404 的小狗，最近又被爆出了新的「笑料」。原本被寄予厚望、用来消灭「狗屁工作」的 AI，最终却沦为制造新型「狗屁工作」的源泉。据《金融时报》报道，为了逼迫员工拥抱 AI，亚马逊搞出了一个极其复古的管理手段：「Token 消耗排行榜」，追踪每个员工的用量。公司强制要求超过 80% 的开发者每周必须完成 AI 使用指标，甚至将消耗 Token 的数量作为考核标准。▲图片

1.OpenAI 的两大宿敌 Anthropic 和马斯克，放下心中成见之后终于在月初结盟了。在此之前，Anthropic 和马斯克的关系并不融洽：今年 2 月，马斯克还在自己的 X 账号指责 A 社「woke」「邪恶」「反人类」（misanthropic），说这家公司「仇视文明」。事后来看，这次攻击并非马斯克清新脱俗的性格使然，而是 Anthropic 所做的某些事情触碰到他的神经，事出有因。在此之前，xAI 内部使用 Cursor 工作，但是今年年初员工发现，Claude 模型突然在 xAI 的 Cursor 公司账号里不能使用了。当时还在 xAI 上班的联合创始人吴宇怀，在全员信里是这么说的：「Anthropic 更新了政策，要求 Cursor 不得向其主要竞争对手提供 Claude 模型调用能力。」当时，吴宇怀在信中写了一句话，颇为有趣：「这是坏消息也是好消息。我们的生产力会被影响，但这也敦促我们开发自己的编码产品和模型。」为什么当时 xAI 的高层认为，开发自己的编码产品是关键？后来发生的事情，大家都知道了。xAI 的联创团队悉数跑路，马斯克一气之下对 Cursor 使用了钞能力必杀：上个月底，SpaceX 和 Cursor 共同宣布，将在编程和知识类工作 AI 模型的训练上，展开前所未有的战略合作；并且，SpaceX 还获得了以 600 亿美元收购 Cursor 的权利，或向后者支付 100 亿美元合作费用。注意编程这个关键定语，后面还会 call back.2.最近，我看了一条 Cursor 早期投资人、Anthropic 大喷子、T3 创始人 Theo Browne 的视频。本来点进去是看他喷 A 社和 SpaceX 怎么蝇营狗苟，结果没想到，却看到了关于 SpaceX + Cursor 合作的，一个既另类却又极度合理的分析：不说 600 亿的收购，就只说 100 亿的合作费——Theo 在视频里表示，自己认为「哪怕只是交换到 Cursor 的用户数据，这 100 亿也值回票价了。」所以是什么数据？如果你也去看 Theo 这条视频，他会讲得非常清楚。但为了节约时间，我们在这里简单概括一下：我们和 AI 的对话是一来一回的，你提出问题/需求，他给你解答；coding agent 同理，只不过返回的是代码。一次高质量的对话，整个过程，包括用户提示、模型思考、agent 规划、输出代码、验证——所有这些东西合起来，可以称为一个完整的 Agentic Loop——就成为了高价值的训练数据，再喂给模型去进行强化学习，就能进一步提高模型在实战场景下的表现水准。Cursor 有的，SpaceX 想要的，就是这些数据。可这些数据从哪里来呢？答案很简单：作为模型厂商，这种高质量数据的最直接来源，只能是你自己开发的 coding agent 产品——也即 Anthropic 的 Claude Code、OpenAI 的 Codex、Kimi 的 Kimi Code。现在你应该明白了，为什么被 Anthropic「封号」之后，吴宇怀会在全员信里提出开发 xAI 自己的 coding 产品和模型这件事了。这件事 xAI 在当时已经看清楚了：没有自己的编码产品，就没有高质量的强化学习数据；没有高质量的数据，就训练不出真正实战能力强的 coding 模型。虽然有点暴论，但现在我们可以点题了：模型厂商想做出来真正能打的编程模型，做自己的 coding agent 产品是唯一的路径。3.大语言模型像个水晶球，用全网的语料训练出来，似乎能够解答万物，但并不代表它在所有问题上都能给出高质量的答案。用 GitHub 上数以亿计的代码条目训练，当然也能训练出 coding 模型。这是「学习结果」的逻辑，也是没问题的。毕竟编码任务的结果是可以验证的：代码能不能运行，测试能否通过，结果摆在那里。但是，通往结果的过程，是一个涉及多步骤决策、错误纠正、意图对齐的复杂链条。每一次用户的接受、拒绝、补全、撤销、追问、甚至当模型好几次都搞不定或者完全搞错时的辱骂——都是这一链条上的过程信号。强化学习有两种监督方式，一种叫做结果监督，只看最后是否跑通。但是结果监督会催生「奖励黑客」的现象：模型为了能跑通可能写出冗余、脆弱、带逻辑漏洞的代码，但因为测试过了，模型以为自己学对了。而另一种叫做过程监督，对推理路径上的每一步进行打分。上述这些过程信号，只有在 coding agent 运行环境里才能诞生。GitHub 仓库里只有结果，哪怕是去看单独的提交历史，看 PR，都找不到有效的过程信号。在缺乏有效、自主可获得的过程信号的时候，一些模型厂商会采用「蒸馏」的方式，这个事情大家应该已经知道了。蒸馏的逻辑很简单，给同样的输入，老师模型输出什么，学生模型就学着输出什么。但是通过蒸馏，即便可以获取到思维链，得到的仍然更接近于结果，而非被蒸馏的老师模型内部的概率分布。一旦学生在推理中偏离了老师的轨迹，哪怕一个 token 不符合，都有可能发生偏离。这背后是强化学习的基础限制：策略梯度定理要求，优化样本最好由当前正在优化的模型自己去产生。这种数据叫做 on-policy 数据。而通过蒸馏别家模型，在别人的产品里产生的数据，来训练自己模型，都属于 off-policy 数据。模型当然可以从中学到东西，但学不到老师模型内部的概率分布信息。而像 Cursor 这样自己就是 coding agent 产品的公司，掌握着最真实、有效、高质量的训练数据。Cursor 产品本身，就是 coding 模型在实战环境中的最佳训练场。我们可以通过 Cursor 年初的「翻车」，来证明这个逻辑。4.APPSO 读者应该记得，年初 Cursor 发布了 Composer 2，号称「下一代专用编程模型」，技术报道写的相对保守，也没有提供具体的模型底座信息。结果很快，网友就在公开代码片段里发现了 Kimi 的模型 ID，截图传遍了开发者社群，逼得 Cursor 副总裁 Lee Robinson 出面澄清：「Composer 2 确实是从开源底座出发的。最终模型大约只有 1/4 的算力来自底座，剩下 3/4 是我们自己训出来的。」几小时后，Cursor 联创 Aman Sanger 也跟着发了一条道歉：「一开始没提 Kimi 底座是个失误。」五天后，Cursor 放出了完整的 Composer 2 技术报告，显示底座的确是 Kimi K2.5，授权方则是 Firworks AI，大致流程是在 K2.5 上做训练，再继续做大规模强化学习（RL）。但关键之处在于，Composer 2 的 RL 是运行在真实的 Cursor 会话当中，使用与生产部署完全相同的工具和 harness。Cursor 将这套流程叫做「实时强化学习」(real-time RL)，也即将模型的 checkpoint 直接部署到 Cursor 生产环境中，观察用户的响应，收集数据，聚合成奖励信号——最快可以每 5 个小时迭代一次模型版本，然后继续部署到 Cursor 里，循环往复。最极致的案例是 Cursor 的自动化代码补全功能 Tab，每天处理超过 4 亿次请求，每当用户输入字符、移动光标时，模型都会预测下一步动作，如果预测置信度高，则显示建议，用户按下 tab 即接受自动补全。该功能采用的是在线强化学习，在行业内极具特色。Cursor 可以以极高的频率（最快可达每一个半小时到两小时）更新 Tab 的模型能力给用户，直接在产品内收集 on-policy 数据进行训练。这种高频、接近实时的反馈回路，让 Tab 可以学习到极其微妙的用户意图。Cursor 方面透露，这种方法让 Tab 建议的拒绝率降低 21%，接受率提高了 28%。回到 Composer 模型本身。在事情搞清楚了之后，一些 Kimi 员工也删掉了之前吐槽的的推文，Kimi 官方账号发表了祝贺。一家估值 600 亿美元（基于马斯克给的数字），不做自己的模型基座的 coding agent 应用层公司，仍然可以通过产品自身的数据飞轮，RL 出超越基座模型的专有编程模型。所以与其说 Cursor 翻了车，不如说这反而是 coding agent 产品重要性的绝佳例证。Cursor 在另一篇关于实时 RL 的文章里写到：「（训练编程模型）最大的困难在于建模用户。Composer 的生产环境里不只有执行命令的计算机，还有监督和指导它的人。模拟计算机容易，模拟使用它的人却很难。」这句话，现正在逐渐成为了在编程模型方面走在前沿的模型厂商之间的共识。如果你去看 benchmark 榜单和用户普遍评价，会发现哪些头部的厂商都在发力做自己的 coding agent/编程产品。区别只在于谁离用户更近。我们以 SWE-bench、LLM-Stats 等相对权威的榜单为例，Claude、GPT、Gemini、Kimi 等模型基本霸榜前十，清一色都是有自己开发 coding agent 产品（包括 CLI、IDE、集成 coding agent 的桌面客户端）的模型厂商。在部分榜单上会出现少数反例，如 Meta (Muse Spark)、DeepSeek 等，没有开发自己的 coding agent。不过你会发现，这些反例模型，在更加接近真实场景、避免污染的更权威 benchmark 上就很难上榜了。以 DeepSeek 为例，它在 SWE-bench bash only 上分数是 70%，排名第九，在 SWE-bench Pro 上分数却掉到了 15% 左右。OpenRouter 的真实流量数据可以解释这种反差：该平台 2025 年报告显示，Claude token 消费 80% 以上用于编程和技术任务，而 DeepSeek token 消费主要集中于闲聊和角色扮演。没有自家 coding 产品的厂商，在一些 coding 任务 benchmark 上能挤进头部，但在更难的真实工程 benchmark 上，在用户用 token 消费投票的真实流量中，都会原形毕露。不仅是 Cursor，Anthropic 在 2025 年 11 月发的一篇论文里，也明确透露自己在做一模一样的事情：「我们在 Anthropic 自家的真实生产编程环境上做训练。」也即 Anthropic 把自己员工使用 Claude Code 的交互数据，反哺给 Claude 模型用来训练。5.在 AI 的演进历程中，生产要素的定义发生了深刻的位移。传统三大核心要素——算力、研究、训练数据，虽然在总量上持续增长，但在结构上已经出现了严重的失衡。今天的各大 AI 巨头显著提高了在算力上的资本支出 (CapEx)，让算力基建成为了当前舆论的主旋律。但实际上，特别是在编程范畴内，随着 GitHub 仓库、StackOverflow 等互联网公开代码数据被基模厂商「竭泽而渔」式地利用，模型在代码生成与逻辑推理上的边界开始逐渐显现。这也是为什么，行业共识正在逐渐转向一个冉冉升起的新战略高地：对于任何希望掌握顶级代码能力的模型厂商而言，建立自有的 coding agent 产品早已不再是可选的商业路线，而是确保底层模型可以持续进化的核心生命线。正如前面 APPSO 论证的那样，单纯学习公开数据等于只学习成功者的结局，却无法了解成功的路径，这绝对不是正确的成功学应该有的样子。在真实的编程环境中，知道发生了什么错误、怎样发生的、如何正确地理解和高效地实践需求等等——了解正确过程的价值，远超于得到正确结果本身。只有拥有自己的编码产品，模型厂商才能获取高质量的「过程监督」信号，从而在编码/推理能力的下一阶段竞争中，确保自己仍有技术护城河——否则就不得不像 SpaceXAI 那样，花钱去跟 coding agent 产品公司去合作。然而并不是所有模型厂商都跟马斯克一样有钱，以及 2026 年开始的巨头势力划分、结盟与领地的争斗会变得更加激烈，当一家缺乏自主 coding 产品的模型厂商终于回过味来的时候，恐怕已经没有足够的合作伙伴可以挑选，合作

今年 3 月，苹果制造了 MacBook Neo 这个大爆款，把高质量入门笔记本的价格打到了 4599 元人民币起。一个半月后，这台搭载 A18 Pro 处理器的入门款 Mac，全年销量从 700 万台追加到了令人难以置信的 1000 万台，交付周期一度排到了 5 个月之后……苹果紧急向鸿海、广达等供应商追加订单。知名供应链分析师郭明錤给出更刺眼的数字：在全球 Windows 笔记本电脑预计同比衰退超 10% 的大背景下，2026 年 MacBook 系列出货量有望达 2500 万台，同比增长 20%-25%。这一拳，Windows 阵营的最核心代表英特尔，该如何接住？在今天，英特尔在中国正式推出第三代酷睿处理器，代号 Wildcat Lake「野猫湖」，并同时启动了联合产业链伙伴的「Firefly 萤火虫计划」。英特尔的接招，围绕一件事而展开：将最先进的 Intel 18A 制程，下放给低端，或是入门，或是「主流」——随便你怎么称呼它的价位段。野猫盯上新的猎物Wildcat Lake 这个代号，在去年底就已经在业界出现。今年 CES 期间，英特尔高调发布了主打高端的产品线：酷睿 Ultra Series 3，采用 Panther Lake 架构。同一天，Wildcat Lake 也悄悄披露，你可以将它理解为 Panther Lake 的精简版本。业界普遍将其视为英特尔为了对标苹果 MacBook Neo 而推出的产品，专攻平价入门级 PC 市场。「直接对标」这四个字，并不是说说而已。在 4 月底，英特尔为野猫湖推出的公版参考设计曝光，采用渐变配色、铝合金一体成型机身、窄边框键盘，扬声器开孔环绕在键盘两侧、D 面无出风口——这种设计语言不需要过多解释，它就是在用苹果惯常的设计语言，去回应苹果本身最受欢迎的多彩配色与铝合金轻薄机身策略。在 MacBook Neo 上，苹果做对了一件事：它让整个行业重新意识到，在「内存末日」(RAMmageddon) 的今天，入门价位段成了一座金矿。用一颗本来进入销库存阶段的 A18 Pro 芯片，配上 8GB 内存和 256GB 存储，开价 4599 元——MacBook Neo 活脱脱成了 MacBook Air 平替。英特尔，以及整个 Windows PC 阵营不能继续将这个用户群体拱手相让。而这，正是「野猫湖」存在的（几乎）全部理由。18A 首次下放到入门段第三代英特尔酷睿处理器的技术参数并不复杂，CPU 采用 2 个 Cougar Cove 性能核，加上 4 个 Darkmont 超低功耗能效核，总共 6 核 6 线程；GPU 部分配备 2 个 Xe3 核心，NPU 算力最高可达 18TOPS。整机基础 TDP 15W，最高 35W，在无主动散热的笔记本上，甚至可以压低到 11W 运行。按照英特尔今天提供的官方口径，这颗芯片功耗相比前代降低超过 60%，图形 AI 性能提升了将近 3 倍，平台总算力更是达到了 40TOPS。当然，考虑到「野猫湖」处理器的 NPU 算力只有 17TOPS，远低于 Windows 阵营的所谓「AI PC」，也即微软官方定义的 Copilot+ PC 40TOPS 的阈值。所以严格来说，野猫湖并不能直接和苹果 A18 Pro 高质量的神经网络引擎，以及统一内存架构，去做横向比较。话虽这么说，驱动参数量不大的本地小模型，做一些 AI 推理任务，还是绰绰有余的。以及更有意思的是，谷歌前段时间发布自家的 AI 笔记本电脑「Googlebook」的时候，也确认了和英特尔结盟——Wildcat Lake 被海外业界人士普遍认为将会成为是 Googlebook 的核心平台。连谷歌都觉得够用，应该没大问题吧？真正关键之处不在于这些参数，而是在于「野猫湖」用上了 Intel 18A 制程工艺本身。要知道，过去两代酷睿 Series 1 和 2，一直沿用 Raptor Lake-U 架构和 Intel 7 制程，实质上是 10nm SuperFin 工艺的演进版本。在主流芯片普遍走向 5nm 甚至 3nm 的当下，技术代际差距已经相当明显。当 MacBook Neo 用虽然是老库存但货真价实 3nm 工艺的 A18 Pro，杀进四千多元的价位段，英特尔过去那套「靠老制程压低成本」的入门策略已经过时了。所以，我们很欣慰地看到，英特尔将 18A——2nm 级别的工艺，压到一颗入门芯片上。这意味着英特尔充分愿意承担初期更高的单位成本，去力保一座不能再丢弃的城池。系统：手机供应链思路搬到 PC 上对于守住市场，「Firefly 萤火虫计划」同样重要。英特尔副总裁、中国区软件工程和客户端产品事业部总经理高嵩如是说：芯片本身的算力固然重要，但「芯片和系统层面的双重创新」才是这一代主流 PC 体验跃迁的来源。中国区客户端与平台销售业务部总经理宗晔则认为，PC 也正在从工具进化为「智能体验入口」。根据英特尔提供的内部数据，与上一代相比，「野猫湖」产品的 PCB 面积缩小了 5%，器件数量减少了 7%。同时针对第三代酷睿平台特性制定并推广标准化主板与 I/O 板连接，推动主板核心电路进一步实现模块化、小型化、低功耗化。这一套打法的精神底色，其实来自于手机。手机行业过去 10 年磨合出的供应链协同方式——元器件高度集成、模块化主板、统一接口标准，被英特尔搬到了入门级 PC 上。结果，就是更轻、更薄，也更便宜。参与「Firefly 萤火虫计划」的合作伙伴，包括但不限于华硕、七彩虹、荣耀、惠普、联想等。英特尔预计，全球合作伙伴将陆续推出超过 70 款基于 Wildcat Lake 平台的产品。目前，荣耀 MagicBook X14 2026 已经率先上架，搭载该平台 Core 5 320 处理器，采用 16GB+512GB 存储配置，定价下探到了 4399 元。用户画像：完全对标 MacBook Neo，但还有更多英特尔列举了 Wildcat Lake 平台产品的 5 类目标用户：辅导孩子作业的家长、全天在校的学生、刚入职场的新人、「一人公司」创业者，以及银发族。这些用户群体和场景描述本身没什么悬念，但值得注意的是，这个用户画像，和 MacBook Neo 在线下门店主要推荐的人群，几乎是完全重叠。MacBook Neo 夺走的本来就是入门级 Windows 笔记本电脑的客群，英特尔的接招，就是造出具有更好价值主张的入门级笔记本电脑，再把心智重新抢回来。英特尔还反复强调野猫湖平台的本地 AI 推理算力。这对于那些购机预算本就不多，也不太可能付费订阅优质云端模型算力的用户来说，倒也是个正确的思路。最后，野猫湖看起来像是合格的入门芯片平台，跟 Firefly 萤火虫计划结合起来，也形成了一套颇具诚意的组合拳。轻、薄、便宜的 AI 笔记本，英特尔现在重新回到了这个擂台。仅剩的悬念，是 OEM 最终能把零售价做到多少。在 MacBook Neo 的价位上，苹果给到了系统生态、统一内存架构、多种潮流配色等等诸多看头。但英特尔要想把这只野猫放出去，跟苹果大杀四方，光有 18A 工艺和优秀的工业设计是不足够的。这场仗的真正交战点，大概率不会在技术本身，而是在即将到来的 618 购物节期间，第一批搭载它的笔记本产品，能否真正把供货紧张的 MacBook Neo 杀个七零八落、措手不及。

⚖️马斯克诉 OpenAI 案败诉📱苹果开启 WWDC 邀请，将公布苹果 AI 新进展🚗问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处🤖国家反诈中心 App 上线「AI 内容鉴定」⌚苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动🚙广汽董事长回应被称网约车之王🧠三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能🎵腾讯音乐官宣收购喜马拉雅💰铠侠股价飙升，计划美国上市🚀马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元☁️百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半🎬爱奇艺一季度由盈转亏 2.95 亿元🦿荣耀机器人「闪电」「元气仔」走进南大💡前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去🚗小米 YU7 GT 定档周四发布🎮5499 元起，红魔 11S Pro 系列发布马斯克诉 OpenAI 案败诉据 CNBC 和《华尔街日报》报道，美国加利福尼亚州奥克兰联邦陪审团已就埃隆 · 马斯克诉 OpenAI 及其 CEO 山姆 · 奥特曼一案作出裁决，认定马斯克的起诉超出诉讼时效，驳回其全部主张。此案源于马斯克 2024 年提起的诉讼。他指控 OpenAI 及奥特曼违背了将该公司作为非营利机构运营的承诺，并在庭审中称 OpenAI「窃取了一家慈善机构」。陪审团的裁决并未就「违反慈善信托」指控的实质内容作出判断，而是认定马斯克的起诉超过了三年诉讼时效。法院同时驳回了针对微软的相关指控。裁决宣布后，马斯克在 X 平台发文，将这一结果定性为「日程上的一个技术性问题」，并表示将向第九巡回上诉法院提起上诉。OpenAI 首席律师威廉 · 萨维特在庭后表示，这「不是一个技术性裁决，而是一个实质性裁决」，并称马斯克是在将诉讼作为竞争手段使用。微软方面亦发表声明，对裁决表示欢迎。苹果开启 WWDC 邀请，将公布苹果 AI 新进展今天凌晨，苹果正式启动 WWDC26 邀请，并公布详细日程安排。本届 WWDC 大会将于北京时间 6 月 9 日至 13 日举行，面向全球开发者开放，重点展示 Apple 各平台的新系统能力、开发工具和框架更新。按照安排，WWDC26 将于北京时间 6 月 9 日开幕，届时 Apple 将举行主题演讲和 Platforms State of the Union，介绍新一代软件平台、人工智能相关进展，以及面向开发者的新工具。大会期间，开发者和学生可以观看超过 100 场围绕工具、技术和设计的视频讲座，参与团体实验室活动，并在 Apple 开发者论坛中与工程师和社区成员交流。Apple 还将在 6 月 9 日邀请超过 1000 名开发者、设计师和学生前往 Apple Park，参与线下特别活动。届时，爱范儿也将在 Apple Park，为大家带来第一手的现场报道。问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处全新一代问界 M9 系列汽车设计师 @Zimo-子墨 日前在微博发文，点赞全新理想 L9 的内饰设计，称其「简洁、逻辑、比例和谐，视觉重心清晰，口味控制得恰到好处」。他表示，初创公司在有强决策者的情况下，往往对设计管控介入较深，容易打散设计的贯穿逻辑；而到第二代、第三代产品时，无论前代成功与否，团队通常会更多放权给设计师打磨表达。他同时透露，新 L9 的内饰设计总监为来自宝马的 Chris Lee，并称虽然双方长期管理直接竞品，但「真的欢喜好设计，生在中国」，并已发短信向对方表示祝贺。国家反诈中心 App 上线「AI 内容鉴定」据央视新闻报道，国家反诈中心 App 近期正式上线「AI 内容鉴定」功能，支持对图像、视频、音频及文本进行 AI 生成痕迹检测，主要用于识别诈骗及电商场景中的伪造内容。电商平台上利用 AI 篡改图片骗取「仅退款」的行为日趋普遍。央视新闻记者随机选取十张商家提供的疑似 AI 生成售后图，上传至国家反诈中心 App 及多个 AI 鉴真工具后，均被判定为「图像含 AI 生成痕迹」。据介绍，该检测方案结合小模型与大模型：小模型学习 AIGC 内容的生成模式，大模型则判断内容是否存在物理瑕疵或逻辑语义问题，二者联合输出最终结果。苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动据 DIGITIMES 报道，苹果计划于今年下半年对高端 Apple Watch 机型进行重大改款，传感器功能将同步升级，市场预计届时出货量较 2025 年增长 20% 至 30%。Apple Watch 11 已具备血氧、脉搏、心率及血压感测能力，下一阶段预计推进无创血糖监测，具体落地时间取决于 FDA 审批进展。作为苹果传感器元件的独家供应商，亚洲半导体（TASC）预计将直接受益。市场预期新品订单最早于 7 月启动，带动下半年业绩明显超越上半年。广汽董事长回应被称网约车之王据「睿见 Economy」报道，日前广汽集团党委书记、董事长冯兴亚在「第十八届轩辕汽车蓝皮书论坛」上，针对「网约车之王」的说法，其指出：很多消费者可能没有想明白一个问题，「网约车是生产资料」。生产资料对品质的要求极高，因为它是用来挣钱的工具。只有能为自己带来更多利益的车，才能受到使用者和购买者的欢迎。对司机来说，车要「多快好省」，要能赚钱。「这恰恰是对品质最好的验证。没有高品质，就不可能受到网约车驾驶者的欢迎，这是最基本的。」他进一步分析，网约车的乘坐周期非常长，一天可能十几个小时都在车上。如果舒适性不够、疲劳感很强，它也不会受到司机的欢迎。此外，电耗等各项指标，也决定了网约车司机是对车要求最严格的一群人。冯兴亚总结道，这些从网约车运营中积累下来的经验，广汽会延展到旗下各个品牌，让车辆变得更加可靠、舒适和安心。「这反而是我们的优势所在。」三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能据 DIGITIMES 报道，三星电子近期大幅提升了用于高带宽内存（HBM）的最新 DRAM 良率。其第五代 10nm 级 1b DRAM 良率已达 92%，用于下一代 HBM4 产品的第六代 1c DRAM 良率也突破 75%。良率的提升有望增强三星的成本竞争力，并提升其向主要 AI 客户大规模供应先进存储产品的能力。在产能扩张方面，据 Ddaily 报道，三星和海力士正同步提前推进工厂扩建和设备订单。三星正加速平泽 P4 工厂的 DRAM 投资，设备安装预计持续至今年底，市场估计明年将进一步扩大投入。海力士则在清州 M15X 和龙仁 Y1 工厂推进扩产，业界预计其明年 DRAM 相关资本支出有望较今年翻倍。腾讯音乐官宣收购喜马拉雅腾讯音乐娱乐集团（TME）昨日宣布，已正式完成对喜马拉雅并购交易的交割，喜马拉雅成为 TME 全资子公司。交割完成当日，TME 执行董事长彭迦信与 CEO 梁柱联合向全体员工发布全员信。管理层称，喜马拉雅是中国长音频领域的重要建设者，双方战略方向高度一致，合并后将共同拓展在线音乐与音频行业空间。铠侠股价飙升，计划美国上市据彭博社报道，铠侠（Kioxia）股价昨日上涨 16%，创下新高。公司在上周披露强于预期的业绩展望后，多家华尔街机构迅速上调目标价；与此同时，铠侠也确认正准备在美国发行存托凭证，借此提升流动性并扩大国际资本市场接触面。报道提到，铠侠预计截至今年 6 月的季度营业利润将达到 1.3 万亿日元，高于市场平均预期，季度利润也创下纪录。多家机构认为，在 AI 带动的存储供需紧张环境下，公司仍有进一步提升自由现金流和股东回报的空间。马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元据彭博社报道，马斯克昨日在萨姆森国际智能移动峰会线上视频连线时表示，正在德克萨斯州推进 SpaceX 上市准备，称「SpaceX 的 IPO 事宜将很快开始」。知情人士透露，SpaceX 最快本周提交 IPO 申请文件，计划融资最高 750 亿美元，目标估值超 2 万亿美元，若成功将成为史上规模最大的 IPO。百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半百度昨日发布 2026 财年第一财季业绩，期内实现营业总收入 320.75 亿元，同比下降 1.16%；归母净利润 34.45 亿元，同比下降 55.36%；经营现金流为 27 亿元。具体来看，百度核心 AI 新业务收入达到 136 亿元，同比增长 49%，首次占百度一般性业务收入比重超过一半；其中智能云基础设施收入 88 亿元，同比增长 79%，GPU 云收入同比增长 184%。萝卜快跑在一季度的全无人自动驾驶运营订单达到 320 万单，总订单量同比增长超过 120%，截至今年 5 月其全球业务已覆盖 27 个城市，累计自动驾驶里程超过 3.3 亿公里。爱奇艺一季度由盈转亏 2.95 亿元爱奇艺昨日发布 2026 财年第一财季业绩，期内实现营业总收入 62.26 亿元，同比下降 13%；归母净利润由盈转亏 2.95 亿元，去年同期为盈利 1.82 亿元；经营现金流为 1.86 亿元，同比下降 45.13%。分业务看：会员服务收入为 42 亿元，同比下降 5%；在线广告服务收入 12.4 亿元，同比下降 7%；内容发行收入 3.59 亿元，同比下降 43%；其他收入 4.27 亿元，同比下降 49%。公司在财报中提到，热门剧集仍推动会员收入环比增长，海外业务保持稳健增长，并计划借助AI降低内容制作成本。荣耀机器人「闪电」「元气仔」走进南大5 月 17 日，荣耀携自研机器人「闪电」与「元气仔」在南京大学开展具身智能科普活动，并同步展示荣耀 600 系列新品的核心影像功能。活动现场，荣耀展示了自研机器人的核心运动与交互能力。搭载高动态运动系统、支持自主感知导航与高速避障的机器人「闪电」与在校学生进行了 8 轮 1V5 短跑测试；另一款机器人「元气仔」则完成了包含「太空步」在内的复杂舞蹈动作演示与人机互动。同期，荣耀在现场设立了新品体验区，重点展示荣耀 600 系列智能手机。官方信息显示，该机型主打基于 Magic OS 系统的 AI 能力及 4K Live Photo 动态照片功能。外观方面，荣耀 600 系列目前已公布「幸运星」「青苹果」「光羽蓝」和「曜石黑」四款配色，其中主打配色「幸运星」采用了行业首创的双重磁吸工艺。💡 前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去前 Tesla 无人驾驶与 SpaceX 猛禽发动机团队工程师 Zac Valles 在 YC Demo Day 结束 72 小时后直飞深圳，驻扎 8 周后携新硬件返回旧金山。他随后在 X 上发布了一篇长帖，向所有有意从事硬件创业的人发出忠告：不论有没有融资、团队甚至创业想法，都应该立刻去深圳。只要你对创办一家硬件公司有哪怕一点点想法，就去深圳。融资前、组队前、有想法前、离职前，都不是借口。去就对了。在切入供应链的路径上，Valles 建议创业者围绕一场大型行业展会安排行程，将其作为接触工厂的跳板。他还建议在华强北预留 4 到 6 小时实地走访，称其为一座「微型城市」，机器人子组件、电池、底盘、电子元器件均有专属楼栋聚集。小米 YU7 GT 定档周四发布昨日，小米正式宣布，将于 5 月 21 日（本周四）举行人车家全生态新品发布会，届时将发布小米 YU7 GT、小米 17 Max 等新品：小米 YU7 GT：定位跑车级 SUV，最大马力可达 1003ps，最高时速 300km/h，CLTC 纯电续航 705km；小米 17 Max：采用 6.9 英寸极窄四等边直屏，搭载徕卡 2 亿像素主摄和 8000 mAh 金沙江电池；小米手环 10 Pro：最薄 9.7 mm，最轻 21.6 g，续航最长可达 21 天；小米首款耳夹式耳机：单机重 5.5 g，支持 LHDC 5.0。5499 元起，红魔 11S Pro 系列发布红魔昨日正式发布红魔 11S Pro 系列，国内首发第五代骁龙 8 至尊领先

本周看什么 | 最近值得一看的 8 部作品主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注潘誉晗少数派作者潘誉晗关注潘誉晗少数派作者关注利兹与青鸟少数派作者利兹与青鸟关注利兹与青鸟少数派作者关注SHY少数派作者SHY关注SHY少数派作者少数派编辑部 等 4 位作者05/15 18:00☕️ TL;DR近期佳作推荐：[国产] 低智商犯罪、[韩剧] 夺命许愿、[英剧] 好兆头 第三季、[法剧] 抢钱女王团、[电影] 蒙特利尔，我的美人、[动画] 石纪元 科学与未来 第 3 部分、[纪录片] 马克斯·维斯塔潘：新征途、[真人秀] 旧洞基朋友们几则精彩预告：《攻壳机动队 THE GHOST IN THE SHELL》新预告、《沉默的审判》释出预告、《火遮眼》新预告、《爱情城事》定档预告、《奥特曼起源》正式预告几则影视资讯：《诺曼底 72 小时》内地定档 6 月 6 日、《宇宙巨人：希曼崛起》内地定档 6 月 5 日、《家弑服务》内地定档 5 月 29 日、《超级少女》内地定档 6 月 26 日、漫威新剧《幻视追寻》正式定档往期推荐可点此或点击文末 tag #本周看什么，更多主题片单可点此查看；想要查看更多深度影视内容，可以关注我们的衍生栏目「看什么 Café」。如有其他建议也欢迎在评论区提出，感谢大家对「看什么」栏目的喜爱和支持。[国产] 低智商犯罪关键词：喜剧 / 犯罪片长：45 分钟左右（单集）× 24 集观看渠道：爱奇艺/咪咕视频；豆瓣链接运气之神降临！@利兹与青鸟：平安市刑警张一昂能力不强但热爱吟诗，常常不合时宜得令人发笑，被下放到老干部处都能惹出乱子。因为他正巧捡到一封举报信，于是和同样脑回路清奇的李茜一起被派到三江口，调查公安局副局长的死是否有隐情。但空降而来的张一昂很不受欢迎，工作处处受限，却在阴差阳错之下抓住一个在逃嫌犯，让人刮目相看，就这样被架起来成了神探。另一边，平安市两个笨贼想出了抢贪官的主意也来到三江口；而三江口涉黑的集团老板周荣考虑着新项目，也着手调查发小叶剑刑警队长的死因。三条故事线、两桩凶案（可能不止）就这样汇集一处，想必会给三江口这一团乱麻摩擦出不少火花。本剧改编自紫金陈的同名小说，一开局就满是带着黑色幽默的荒诞味，罪犯和警方彼此高估，虽然智商时常掉线，但不妨碍瞎猫碰上死耗子，不禁让人好奇这份「好运」还能带来什么啼笑皆非的巧合。马旭东、闫佩伦、詹鑫等喜剧演员的助阵也让笑点呈现得更好，尽管罪案线稍显拖沓，但也算是合格的电子榨菜。5[韩剧] 夺命许愿关键词：剧情 / 恐怖又名：기리고 / Wish Your Death / If Wishes Could Kill片长：44 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接为了愿望达成，你愿意不惜一切代价吗？@潘誉晗：这次数学考试的结果有点出乎意料，因为崔亨旭拿满分了，可这家伙不是个读书的料子。面对好友们的询问，亨旭得意洋洋地把一个名叫「Girigo」的软件链接分享给了伙伴，说是只要写上自己的生辰八字贴在自己的身上，然后录下视频上传到这个软件上，就能实现愿望。大家并不相信亨旭的话，亨旭也没在意此刻出现在他手机上的倒计时。谁承想 24 小时后，亨旭就当着全班同学的面自缢了——原来这则所谓的许愿软件，在愿望达成后会要你的命。Netflix 韩国这次又推出了一款爆剧，一经上线就在 13 个国家登顶。剧集整体的节奏很是紧凑，恐怖惊悚的氛围也营造得相当出色，很好地把校园青春剧和恐怖因素结合在一起，然后又完美连接了手机 app 和萨满巫术。神秘的夺命软件利用了青少年的欲望，随后就像是一个大家无法挣脱的隐形牢笼，不仅顽固地无法在手机上卸载，更是在许愿者身后步步紧逼、致人于死地。2[英剧] 好兆头 第三季关键词：剧情 / 喜剧 / 奇幻又名：Good Omens Season 3片长：99 分钟观看渠道：Prime Video；豆瓣链接我有那片宇宙，我还有你，我已经拥有我想要的一切。@SHY：成为至高大天使的亚茨拉斐尔，开始筹备「第二次降临」，复活的耶稣却不知所踪，只得向自暴自弃的克劳利求助。这对自创世之初就有交情的天使和恶魔，能否像往常一样弥合分歧，而在背后作祟的幕后黑手，又会将整个世界导向怎样的结局？受原作者兼前两季主创尼尔·盖曼的性侵丑闻影响，最终季被砍成了仅 1 集的特别篇，篇幅势必捉襟见肘。不难想见，成片处处流露着删减的痕迹，忽快忽慢的节奏让叙事更加混乱，整体观感有着不小下滑。虽然填充的内容变了样，两位新编剧还是尽力还原了盖曼构想中的结局，而在破碎的字里行间，也能窥见一些残存的灵光。惯常的黑色幽默难以撑场后，本季的重心落在了本来带点调侃的父母爱情故事上，大卫·田纳特和麦克·辛以一如既往的真挚演技力挽狂澜，继续迸发充沛的化学反应。彻头彻尾的粉丝向结局，评价两极分化在意料之中，好在，本作原本就是作为限定迷你剧立项，如果你有点难以接受这一季的走向，只把第一季尊为正统也未尝不可。发表首评[法剧] 抢钱女王团关键词：喜剧 / 惊悚 / 犯罪又名：Les Lionnes / Cash Queens片长：55 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接我们之所以抢钱，真的是迫不得已。@潘誉晗：罗莎这天找到自己的上司，希望可以提高她的信用卡额度。可上司不仅很快就拒绝了她的请求，甚至通知她的额度被降低了，因为她征信由于入狱前夫的债务问题受到了影响。罗莎不敢相信上司的话，也就是在这时，银行的警报铃响了，原来这天又是银行例行的演戏日。看着上司希望职员们保障生命的大喊声，一个不该出现的念头在她的脑海中出现了——她决定伙同自己的好友，抢劫自己工作的银行。Netflix 法国于两个月前推出了一部女性犯罪喜剧《抢钱女王团》，讲述了五名女性在窘迫生活的逼迫下，不得不伪装成男性抢劫银行的故事。全剧充斥着大量很荒诞又很现实的设定，主角团的五名女性拼命工作、好好生活，但却一直过得不幸福。职场上的性别歧视、身上背着重债、被同侪做局丢掉了工作、被丈夫家暴……不是说努力就能有好结果吗？为什么在她们的身上，越努力却似乎越悲伤呢？发表首评[电影] 蒙特利尔，我的美人关键词：剧情 / 同性又名：Montréal, ma belle片长：117 分钟；豆瓣链接被困在夹缝中时，是面对自己还是活在谎言里？@利兹与青鸟：凤霞一家已经移民加拿大十几年了，在魁北克省的蒙特利尔开着一间小超市。丈夫性格较为古板，苦于没有当地学历，找工作频频受挫；大女儿正值叛逆期，对待父母多是指责和抱怨；儿子也只是上小学的年龄，家中没有人能为受困于更年期的凤霞提供情绪价值。在法语课上，凤霞被年轻人追求梦想的热情所鼓舞，尝试接触同性交友网站，并结识了 30 岁的卡米尔。这让凤霞重新找回了自己，却引发了家庭危机。本片的华裔女导演很有相关生活体验，她用代际间新旧观念的碰撞、婚姻大于爱情、照顾家人大于自我实现等细节，呈现出一个传统而压抑的中国移民家庭外壳。片中演员也多是当地华裔，汉语都很纯正，对华语区观众的耳朵十分友好。电影虽是在讲中年女性在道德边界寻找自我，却没有激烈的口号，而是用朴素温和的风格、平缓的节奏，给主演陈冲留足表演空间，为年过半百却天性浪漫、依然优雅而富有魅力的凤霞赋予灵魂，细腻又克制地演绎出她内心的彷徨与挣扎，也让观众沉浸其中。女主角是陈冲啊？1[动画] 石纪元 科学与未来 第 3 部分关键词：漫画改 / 科幻 / 冒险又名：Dr.STONE SCIENCE FUTURE 第3クール片长：24 分钟（单集）× 13 集，每周四更新观看渠道：巴哈姆特动画疯；豆瓣链接We choose to go to the Moon.@SHY：孤注一掷的抗争后，科学王国终于取得对北美阵营的胜利。面对迫在眉睫的威胁，千空和赛诺搁置矛盾，联手打造宇宙飞船，前往月球探寻石化的真相。为了复活更多人类，同时取得不可或缺的材料，众人兵分三路执行任务，「WHY 人」却传来了更加危险的信息。从石器时代到现代文明，跨越 200 万年光阴，这段从零开始的科学冒险终于迎来最终幕。从刚苏醒时手工打造的粗糙制品，一步步重现火药、铁器、蒸汽、电力，来到本季，科技树终于攀升到计算机和互联网的领域。要实现登月的最终目标，需要远胜以往的人力和物力，千空等人因此踏上环球之旅，整合各个国家的资源。即使原料由石头进化到超合金，舞台由村落跃进至外太空，本作的核心魅力始终如一。秉持理性和探索精神，用科学之力对抗奇幻，挥洒荡气回肠的热血，开拓人类未来的道路，诠释纯粹的浪漫。这个集结全世界力量、于石之世界铸造的「阿波罗计划」，绝对能让人心潮澎湃，剧中历尽千辛的角色们，也值得最为美好的结局。发表首评[纪录片] 马克斯·维斯塔潘：新征途关键词：纪录片又名：Max Verstappen - New Ground / 极速新域：维斯塔潘的北环之路片长：23 分钟左右（单集）× 3 集；豆瓣链接出发吧！汽车人！@潘誉晗：对于喜欢一级方程式赛车的人而言，效力于红牛车队的马克斯·维斯塔潘，是一个了不起的名字。他出生于赛车世家，自己的赛车成绩也非常傲人。不过卸下这一身份，他其实还有另外一种人生——一家 GT3（Grand Touring 3）车队的老板兼赛车手。而这部只有三集的纪录片，就是讲述维斯塔潘的第二身份。「看，老板亲自上场了。」维斯塔潘没有老板的派头，招呼完就直接上了场；他也不需要拥有全部欢呼与掌声，而是会亲自去挖掘赛车手。于是，他发掘了克里斯，这个和他一样从开卡丁车（并且克里斯的车号是 3）开始的小伙子很有天赋，可苦于 15 万欧元的培训资金，只能把梦想转移到模拟器赛车上。当维斯塔潘找到他，让克里斯在场地上尝试开车的时候，克里斯甚至没有驾照。相信在观看了这部纪录片之后，没人会否认维斯塔潘对赛车的热爱。他脱下了红牛的飞行夹克，在另一条赛车的道路上，继续驰骋。发表首评[真人秀] 旧洞基朋友们关键词：真人秀又名：구기동 프렌즈 / Six Singles Under One Roof片长：90 分钟左右（单集）× 10 集；豆瓣链接一起同居吧。@潘誉晗：韩国最近又推出了一部真人秀《旧洞基朋友们》，找来了张度妍、张根硕、安宰贤、李多熙、景收真和崔丹尼尔六位年龄相近的艺人，过一次同居生活。随着明星观察类的项目在市场上越来越多，我相信有越来越多的人对这一类真人秀感到厌倦。而《旧基洞朋友们》的出现，再次证明了韩综的创意。看似依然是粉丝向地找了一群有粉丝基础的演艺人员，但是却找准到了一个很细腻的切口——让几个曾经独居过的明星们，过群居生活。独自生活，只是光看到就觉得很自由的日子，为什么他们体验过那种日子后，不愿意再独居了呢？我想这也就是观察类节目出现的初衷，去探讨一个社会性的议题。毕竟「存在即合理」。而这六名 85 后明星的同居日常，也非常有看点。他们都很有个性，但又能互相融入、靠近。见惯了大场面的他们，看到 18 公斤的章鱼，依然会手足无措，还是会大声尖叫。原来群居也不差嘛。桃花坞？年龄相近 不是想尽1更多[电影] 总统的蛋糕@利兹与青鸟：萨达姆的独裁统治下的伊拉克，平民们过得水深火热。小女孩拉米亚住在草编的窝棚里，连书包里的苹果都要被老师搜刮走。更不幸的是，她被抽到要在总统生日这天做好一个奶油蛋糕。于是拉米亚踏上了旅途，用家里所有值钱的东西换取做蛋糕所需的材料。电影在去年戛纳电影节获金摄影机奖，摄影优美，配乐充满阿拉伯风情，有触动人心的童真，也有伊拉克战乱下的可怖的倾轧。[电影] 平场之月@潘誉晗：电影根据朝仓加守美同名小说改编。青砥健这一天意外偶遇了自己的初中同学须藤叶子。此时两人已经 35 年没见了。中学时两人就互有好感，这次重逢，已离婚的青砥健和丧偶的叶子，终于还是走到了一起。埼玉小城里这一对中年人迟到多年的遗憾恋曲，就像是终于遇见了自己的另一个缺失的半圆，内敛、沉稳又有淡淡的愁绪。[韩剧] 浪漫的绝对值@潘誉晗：每当晚上，看似平凡的高中生吕意珠就会偷偷创作 BL 小说。可惜，她的作品一直

厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨主作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了联合作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了05/15 00:07Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。免责声明与作者所需披露的一切首先声明一下——本人仍然是在校的本科微电子硕士考古学的大学生，并没有经过任何的传媒学相关的学习，也没有从事相关行业的经验。目前仅有的和媒体沾边的事情是在酷安、少数派等平台发过一些文章（其中有很大一批都是在最后诉诸于，说好听点是真实生活场景，说难听点开始抒情的文章），以及一个 B 站粉丝量远没有破千的账号，因此我远远没有承担过养一个工作室的压力与责任，也没体会到在竞争激烈的环境中生存的艰难。因此以下内容不构成任何形式上的专业意见，只是一个愤青在 Random Yapping。多有冒犯之处，但我本心无意去批评任何具体的某一个媒体老师，只是想尝试提出一个也许还算有点价值的问题。以下内容提到「游戏区」，也仅限于单机游戏，不包含手机游戏、桌游等类型。以下内容提到「数码区」时，通常仅指以手机、一部分 PC 组件为主的主观体验向评测，且我的能力不足以涵盖 DIY 区1以及涉及到实验设计与数据分析的其他评测内容，虽然随着硬件厂商的创新进入后期演变为渐进创新，产品节奏也有所放缓，在社区中也能听到「DIY 区不如以前」的感慨，但本文尽力按照「不理解，没有切身观察就尽量少乱锐评」的原则展开，但撰写过程中有很多内容没收住，因而请让我提前声明一下——以下很多内容只是社区观察，不完全代表事实判断。最后，说句坦诚的实话，这个内容写的过程真的很痛苦——我真的不想让这篇文章变成直球辱骂媒体老师、攻击厂商或者是殴打观众的情绪发泄文，从个人发展角度来说不希望因此在进入之前就已经和某些职业无缘，也不想因此得罪那些因真心倾慕而认识的那些做这一行的朋友，且我真的不希望因为自己的过分言论把发文章的平台本身也拖下水，但因为笔力实在是有限，尽管拉着 Gemini、GPT、DeepSeek 等一众大模型核验，也找了一些朋友探讨，尽可能不要直球攻击特定的一两家媒体也不要说的太过分，但是字里行间还是有一股火药味拼尽全力消不掉。真的很抱歉，尽管我其实将自己也包括在本文的攻击射程以内——我自己就是后文中所批判的那个自认为掌握了真实，在没有厂商利益束缚与照顾社区的必要下肆意输出内容，甚至有很多是过当内容的小自媒体。这些内容因为是我自己一段时间内自我质疑与焦虑的外化，因此，本文必然存在一些因深度不足、角度有限而导致的难以避免的暴言，简而言之——就是年轻时犯下的错误。麻烦各位读者多多体谅。也欢迎您在评论区里多批评指正，让我多与您交流学习。以下内容如果有哪里说的不合适的地方，有什么可能涉及到的法律争议，请尽快联系我去撤换文章中对应的部分，谢谢！本文长度约为三万字左右，前三部分观点将会较为尖锐，会举几个相对较为敏感的案例来佐证观点，后两个部分则尝试解决前文抛出的问题，内容相对温和但字数也相对较长，请坐和放宽。主观口胡系列是一个尝试将动漫、电影、游戏叙事与更加严肃的讨论结合起来，去尝试探讨一些特定大命题的个人尝试，可能会一定程度上背离这些作品的创作初衷，聊的可能也会比较混乱且随意，甚至会有意提取个别元素，曲解原文本来聊一些其他领域的内容。这次我想把一个大约是三四年前就开始产生的想法去做进一步的深化，并尝试结合近些年来我对游戏区与数码区的一些粗浅且悲观的看法，来尝试去聊一个问题，一个也许注定没有答案，跨多少领域学科都难以得出解法的一个小小的问题——主观体验向评测的现状怎样，尽头到底为何方？又是什么构成了它们的专业性？Part1：太多水，太多奶酪——主观内容的梦魇虽然这么说有失偏颇，但我一直觉得游戏评测是一个相当主观的内容——评测者的个人阅历2、偏向的游戏类型甚至是组成一个游戏的系统，对 GDC 中诸多开发者分享的设计理念的理解，以及同期的同类型作品等等——这些都会对你最终得出什么样的结论，给游戏最终打什么分造成影响。所以在因近期的《神探杰克鼠》评测而陷入争议的 IGN 内部——作为一个有 TGA 评选权的，相对来说算是行业内位置靠上话语权比较大的游戏媒体——他们在测评游戏时给出的流程是：在内部筛选最适合某类游戏且当下仍有档期的编辑，由他们在内部其他编辑的帮助下完成对于这个游戏的评测，给出评分，在内部经过审核与排期，最终以文章 + 视频的形式先后放到公共平台上。但是，这个时候总会有个但是。虽然这样的流程确实在内部行之有效，为行业贡献了一批质量很高的评测，尤其是在神作频出神仙打架的那些年里——但同样不可否认的是，无论你之前做对过做好过多少次，一旦出现几次测评结果与大众预期不相符，甚至出现有失水准的内容，就很容易在玩家社群中丧失公信力，变成 IG8——《宝可梦欧米伽红宝石/阿尔法蓝宝石》的 too much water 扣分，《超级马里奥 3D 世界+ 狂怒世界》的 7.0，《死亡搁浅》的 6.8——你知道的，日光之下从来都没有什么新事。而且当我们把视角继续向下探时，会发现这样的问题不仅仅是发生在 IGN 一家，Polygon 的《DOOM(2016)》、VentureBeat 的《茶杯头》（他的操作已经不足以让他完成第一关，只能借其他同事存档继续评——尽管他据自己说已有着超过二三十年的评测历史）……所以到底发生了什么？当然，抛开社区中已有的被讨论过无数次的，对评测者的质疑不谈，我们也许可以从更加细致的视角去看这件事。就像你在做数码类评测会提到机型的性能与续航、外观、系统流畅度与体验、拍照、外放表现，但你不可能在非首发上手的评测里一样面面都要点到，而是根据个人的使用习惯有自己的偏好一样，IGN 的每个评测者在评测游戏时也会有自己的明显倾向，让我们以当年评测《死亡搁浅》的 Tristan Ogilvie 为例——他评测一款游戏的重心是动作系统、交互反馈与流程的顺畅程度。在我眼里，相较于小岛在前作《合金装备 5：幻痛》中用一系列主线支线任务，吸引玩家不断的从母基地回到开放世界，完成一串任务后撤离不同，本作让角色更加持久更加连续更加长时间的停留在这个世界里，将任务可视化，变为地图上要到达的目标点，并在沿路设置一系列需要潜行的敌人据点、风景奇观，并且可以通过基建加速到达目标点跳过一部分游戏内容来优化体验的开放世界而很不幸的是，本作正正好在她关注的几个角度精准的踩中了地雷：虽然相较于《合金装备 4》有所收敛，但游戏内不断插入冗长的过场 cg，以至于从室外场景进入到地图传送需要连续跳过数个动画；游戏流程放速过缓，且游戏内的任务设计被认为千篇一律，主线和支线均由送货构成；动作系统相较于前作乐趣有所减少（尽管二者的游戏定位有着根本性的不同，也许从来就不该这么要求角色的性能），击败 bt 后的奖励反馈也较为缺乏，以至于编辑出于他所偏向的几个维度的也许从传统游戏性的角度合乎情理的考虑，选择给这个游戏 6.8 分。随后我们都知道发生了什么——之后的几年让游戏的异步社交大放异彩，以一个非常奇怪且迅速的方式，让小岛秀夫这种几乎像是寻求争议，找骂、试图以先锋艺术电影的形式在游戏内进行的表达，能被更多人看见，被更多人共情，似乎小岛秀夫在《合金装备 2》后，再一次，等到了时代来验证他思想上的前瞻性。因而，时代背景变迁下大家对小岛独特表达的广泛接受、小岛秀夫的明星制作人光环以及评测中确实存在的一些过当的指责，都让玩家社群反过来开始质问一件事：为什么我觉得好玩的游戏，你却要给出这么低的分？当然，IGN 作为行业头部媒体，其本身在游戏界有较大的声量较高的影响力，且也有着个人做自媒体乃至中小型媒体难以企及的资源，因此它说的话做的评测理应被玩家更加严格的审视，树大招风不无道理——而且至少在我举得几个例子里，引发争议的游戏本身也是在不同领域内质量过硬的存在，只是随着营销的不断发展，疫情时代与后疫情时代每个亚文化小领域的的扩圈与自媒体时代的到来，这种原本对媒体对权威的的尚处合理范围的不信任将逐步扩大，直到这种现象一步步扩散到每一件事，每一个领域，乃至我们每个人都难以避嫌，难以无视。Part2：内容井喷的时代，同质化的幽灵与信任的坍塌数码圈从来不缺学生出身，单纯出于热爱，分享自己体验的自媒体，这样的评测形式本身并非没有价值，它用更加平易近人的角度去涵盖一个产品中用数据评测所难以涵盖的部分，尽管也许，分享自己对于刚入手的数码产品的主观体验并没有什么门槛——有着中专及以上的学历就有足够的语言组织能力去清晰讲述自己的感受，再来一个相机甚至是手机去拍摄素材就能实现，但撰写一篇好的主观体验评测文（不管是哪个领域）就像是一般人眼中的历史研究一样，都觉得准入门槛低，都觉得写的内容是任人装点任人阐释的小姑娘，但真的想要尽可能做到内容详实、条理清晰，往往需要长时间的写作经验积累与大量的训练，而如果想在这个基础上再去追求与社区中主要讨论的内容截然不同的全新角度，或者让我们思考的更加狂野一些，在这个内容上做出深度，探讨一个产品背后更加深刻的有广泛意义的命题，则更加考验个人的水平——但随着互联网的高速发展与信息的高速创作、流通，我们正处在一个内容极其过剩的时代。虽然我承认，我是随着酷市场转型为酷安而踏入数码圈的，本身并不算什么数码区老资历，但你如果要我去回忆在我进入数码圈后这有限的时间内，主观体验向的数码自媒体井喷的那段时间，我想大概是在「老师好我叫何同学」的 5G 视频兴起的 2019 年 6 月前后，以及疫情肆虐的 2021 年后到 2023 年末。前者因为正好切合了国家的 5G 发展战略的需要而被主流媒体报道，被社会广泛关注，因而给圈子引来了泼天的流量——事实上那段时间坚持做手机体验类评测的规模还比较小的自媒体，如今都已经成长为几十万粉丝的台柱，比如我当时看的比较多的花生说；而后者则是由于人们长时间居家，在室外的娱乐方式因客观需要而受限，因而人们开始转向互联网，大约是在这一时期前后，一群赛道更加细分，比如上古数码产品体验、软件交互等领域（就我当时看的比较多的来说是六分超超和苏星河）的体验向自媒体开始大规模兴起，游戏区的很多如今声量较大的自媒体编辑老师也多是在这一时期进入相关行业或者是投身做自己的账号。但与自媒体账号井喷相对应的，却是——请让我想想，该怎么让人更好接受的去讲一个有点不近人情的结论——极其迅速的内容同质化。游戏不是一个竞争激烈且产品间带来的体验差别相较于同类型的数码产品更大的存量市场，有足够多各自差别足够明显的新游戏不断冲刷着视野，因此这种同质化的倾向也许还并不明显，但在数码区，当每个上游的厂商都在绞尽脑汁的结合热门概念，生生在软硬件体验上捏出极其有限的新创新点时，这种感觉就尤其明显——对于自媒体而言，相比起「你自己要讲述什么故事」，你的阐述对其他人有什么价值才更有传播的意义，因而当你迅速的因不受束缚、观点尖锐，获得了一批受众支持迅速爆火后，同质化的幽灵便也开始悄然显现——是顺着原先让你得到人们簇拥的路径一路向西，对着即使没有那么差的内容输出情绪，直到有一刻在无数次路径依赖后判断失误，犯了众怒，还是妥协于平台激励有限而选择靠近厂商，向其他市面已有的无数自媒体靠拢，在厂商给的通稿框架与宣传的大方向已然确定的情况下，进行着极其容易趋同的表达，在首发宣发节奏与内容合作模式影响下，选择性的淡化掉体验中的痛点，即使在前文有提及也往往占据相对较小的篇幅，并随后立马引入一个「但是」，用着首发特调的机型说着好到梦幻的体验，最终在观众的不解中继续在内容理想与工作室运营中艰难的走着钢丝？当然，如果把话说到太绝，往往会有很多偏颇，所以让我们稍微说一些让文章作者更拟人的

「智能队列」如何改善你的家庭网络质量主作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org联合作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org2021/01/31 23:12前天 12:57更新：2026 年世界电信和信息社会日重推在无线路由器的设置页面上，我们可能会看到一个名为「智能队列」的功能（根据品牌或操作系统的不同，也可能会被称做「智能 QoS」、「SQM」等）。这一功能是如何工作的？是否能让网速变快？是否需要打开？本文将尝试解答这些问题。从「网络如何连接」说起本章节主要介绍计算机网络的基础知识。如果读者已经熟悉相关知识，可跳过这部分内容。在介绍「智能队列」之前，先来了解一下基础的网络知识。两台电脑间进行数据传输，存在两种方式，分别是电路交换和分组交换。对于电路交换，两台电脑独占一条线路进行数据传输。过去的固定电话，大多使用的就是电路交换。由于独占一条线路，两台设备之间的通信较为稳定，很难出现数据丢失、数据传输顺序错乱等情况。但一条线路只能供一组电脑使用，另外一组电脑想要通信，就不能复用原来的线路了。对于分组交换，传输的数据被拆分为小段，每一小段叫做一个数据包（报文）。数据以报文的形式进行转发，从而实现了多组电脑复用一条线路进行通信。电路交换虽然能保证数据传输的稳定性，但由于独占线路浪费太多资源，所以在 Internet 中，使用的是分组交换。分组交换将数据拆分成一个一个的小片段来转发，没有独占线路。如果网络条件不好（例如同时有大量报文经过路由器，超出了路由器的处理能力），很有可能遇到数据顺序错乱、甚至数据丢失等情况。那么，这些情况应该如何避免？在微信视频通话、电视直播等应用中，报文的丢失或乱序，只会造成视频图像卡顿，不会造成太严重的影响。所以这些应用会使用简单的UDP协议，不考虑丢包、乱序。在网页浏览、文件下载等应用中，一般使用基于TCP的HTTP协议。TCP 在报文中携带序号，避免乱序；同时提供了报文的丢失重传机制，避免了丢包。所以，也可以认为 TCP 为两台电脑的应用程序建立了一个虚拟的「连接」。（最新版本的 HTTP/3 协议，使用的则是QUIC。QUIC 虽然基于 UDP，但引入了与 TCP 类似的报文丢失重传机制）如果想要了解更多网络基础知识，推荐阅读「入网指南」系列文章：入网指南 01 | 一文读懂你身边的「网络」入网指南 02 | 当你发消息的时候，发生了什么入网指南 03 | 交换机在交换什么？入网指南 04 | IP 地址大揭秘入网指南 05 | 除了接线，还有什么影响着我们网上冲浪入网指南 06 | 连上 Wi-Fi 就能上网了吗？网络延迟：影响上网体验的重要因素带宽和延迟，是影响上网体验的两大因素。带宽指的是在单位时间内，网络最多能传输多少数据。网络的带宽和水管的粗细有着很多相似之处，水管越粗，相同时间流出的水越多；同样地，网络带宽越大，相同时间能够传输的数据越多。在日常上网过程中，大的带宽能带来这些好处：上传、下载文件的速度更快在线看视频时，缓冲的时间更短如果家中有多个人同时上网，大的带宽能够保证每个人都能流畅上网，相互的影响更小对于家庭宽带，我们常常听到的 10 兆、50 兆、100 兆、千兆，就是指的带宽的大小。延迟指的是数据从网络的一端到达另一端，所需要的时间。过大的延迟也会对上网体验带来不好的影响，例如：在浏览器中输入网址敲回车后，需要等更长的时间，网页才能加载；视频聊天、语音通话质量变差，甚至能直接感受到声音、画面的滞后；玩在线游戏时，发现游戏变慢变卡。在很多时候，延迟相比带宽更容易被人感知，对网络体验的影响更大。是什么导致了网络延迟？上文中介绍过，Internet 使用的是分组交换。如果大家阅读过计算机网络相关的书籍，应该对分组交换的「存储-转发」机制有一定的印象。简单说，「存储-转发」是指网络设备将收到的报文存储在缓冲区，然后再发送。这是由于网络设备的性能有限，转发报文有速率限制。如果设备短时间内收到大量报文，一时半会儿处理不过来，就可以先报文存起来，空闲时慢慢发送。通过「存储-转发」，能够提高网络设备的吞吐量，更充分地利用带宽。其中，存储报文缓冲区，是一个先进先出的队列。那么，队列的长度应该如何确定？是不是队列越长，能够存储的报文越多，网络的性能就越好？大家可以先思考下这个问题。其实在过去，内存价格比较昂贵，考虑到成本，网络设备的队列不会太大。现在，随着内存价格的下降，网络设备拥有了更大的队列。但是，网络的带宽是有限的。实际的可用带宽，取决于数据传输路径上的最小带宽。这个最小带宽也被称做带宽上的「瓶颈」（bottleneck）。由于瓶颈的存在，如果流量长时间以较大速率发送，大的队列并不能显著提升吞吐量。相反，大量报文滞留在队列中，会带来延迟的增大。这种队列中缓存有大量报文，来不及发送，导致数据传输延迟增大的现象，被称做 bufferbloat（缓冲区膨胀）。DSLReport 提供的测速服务，能够检测网络的 bufferbloat 程度。不过由于没有中国大陆服务器，测试结果不够准确。国外的读者可以尝试一下：DSLReports Speed Test家用无线路由器中的 QoS 功能在「智能队列」出现之前，很多家用路由器已经提供了 QoS 功能。那么，传统的 QoS 如何解决网络延迟、网络拥堵？存在哪些不足之处？在了解这个问题之前，先来熟悉一下网络中的报文，是什么样子的：在网络中传输的报文，一般包括了链路层、网络层、传输层、应用层这四层信息。其中，网络层的源 IP 地址、目的 IP 地址、协议号，和传输层的源端口号、目的端口号，是 QoS 算法中常常会关注的五个信息。这些信息合称「五元组」。同一个应用在同一时间段内，传输的报文中，这些信息一般是不会发生变化的。例如观看在线视频时，同一个视频的所有报文 ，一般拥有相同的五元组信息。所以，通过五元组，我们能够区分不同应用的流量。特别地，一些常见的应用，会使用固定的端口号。例如网页浏览一般使用 80、443 端口；英雄联盟游戏，使用的是 5000-5500 端口 ……所以，当我们想要提升英雄联盟游戏的体验时，在路由器上设置让 5000-5500 端口的报文优先通过即可：How to set up traffic prioritization rules for League of Legends on Nighthawk Pro Gaming router | Answer | NETGEAR Support家用路由器中常见的 QoS 功能，就是通过手动设置 IP 地址、协议号、端口号等报文特征来实现的。这种方式需要用户了解一定的网络知识，且操作复杂。（虽然部分无线路由器内置了常用应用的 QoS 规则，但仍需要一定的手动设置，且无法及时适配新的、小众的应用）当然，除了上文中介绍的内容，QoS 还有多种不同的实现方式。例如：对特定应用（例如 BitTorrent）进行限速按照 MAC 地址指定某台设备的优先级（例如提高 Nintendo Switch 的优先级，使游戏更加流畅）通过WMM、DSCP等机制来决定报文优先级这些机制各有各的特色和不足。例如对应用程序限速的方式，虽然效果明显，但无法在网络空闲时，动态利用剩余带宽来提升传输速度，也难以适应带宽频繁变化的无线网络。关于上述 QoS 机制的技术细节，本文不再详细介绍。如果感兴趣，可以搜索相关关键词，进一步了解。智能队列：通过「全自动」的方式优化网络、降低延迟什么是「智能队列」目前，一部分新的无线路由器，开始有了名为「智能队列」的功能。这一功能一般指的是 fq_codel。fq_codel 包含两种队列管理机制，分别是 Flow Queue 和 CoDel。在下文中，我将简要介绍这两种机制的作用。Flow Queue前文中介绍过，网络中的报文一般都有自己的五元组信息；同一应用发送的报文，在一定的时间段内，五元组信息是相同的。这些信息相同的一系列报文，被称做一条流（flow）。Flow Queue 为每条流单独维护一个队列。例如，如果我们一边在线观看视频，一边进行语音通话，一边玩游戏，Flow Queue 就会为视频播放的流量、语音通话的流量、在线游戏的流量各维护一条队列：正常情况下，在线观看视频需要占用较大的带宽，但对延迟不敏感，用户可以容忍几秒甚至几十秒的缓冲时间。而语音通话和在线游戏占用带宽小，但对延迟敏感。例如语音通话时，如果一个人说了一句话，对方在几秒后才能听到，用户就能明显感知到延迟。如果三种应用共用一个队列，队列中会充满大量视频播放应用的报文，导致语音通话、在线游戏的报文迟迟无法发送，增大了这两种应用的延迟。Flow Queue 为视频播放、语音通话、在线游戏各维护一条队列，并公平地调度三个队列，避免视频播放的报文占用太多时间，从而让在线游戏、语音通话的报文得到及时处理。由于 Flow Queue 确保了应用程序之间的「公平」（fair），所以又被称做 Fair Queue。CoDel我们已经知道，适当长度的队列，能够充分利用带宽、提高吞吐量。但过长的队列，又会导致过大的延迟。也就是说，队列既不能过长、也不能过短。CoDel 所做的，就是让队列维持在合适长度，最大限度地提升网络性能。CoDel 通过计算报文在队列中的停留时间，来判断队列是否过长。具体做法如下：报文进入队列时，在报文上记录当前时间（也就是报文入队时间）报文出队列时，再次获取当前时间（报文出队时间）计算报文入队时间与出队时间的差值，即可得到报文在队列中停留的时间当报文在队列中的停留时间超过阈值，说明队列过长，需要缩小队列长度当队列过长时，CoDel 会直接丢弃出队列的报文，从而减少队列中报文的数量，达到控制队列长度的效果。（这种对报文的丢弃，并不会对网络带来负面影响，具体请见后续章节）可以看出，CoDel 根据队列中报文的停留时间，来决定是否丢弃报文。而不是「一刀切」地将队列限制在一个固定的、较小的长度。这样既能降低延迟，又保留了队列最大的优点：吸收短暂的大流量报文、提升吞吐量。另外，我们的设备，尤其是手机等移动设备，所处的网络环境是复杂多变的。例如手机靠近或远离无线路由器，就有可能导致网络带宽的变化。CoDel 通过计算报文停留时间的方式控制队列，能够更加灵活地适应网络带宽、网络质量的变化。fq_codelFlow Queue 确保了不同应用程序的流量能够公平传输。CoDel 控制队列长度，实现了整体延迟的降低。fq_codel 就是将 Flow Queue 和 CoDel 结合在一起使用的机制。其中，Flow Queue 先将不同的流划分为不同的队列，再为每一个队列使用 CoDel，控制报文的延迟。两者结合使用，共同提升网络体验。fq_codel 具有免配置、天然适应不同的网络环境、实现简单、消耗资源少等特色，越来越多的无线路由器（以及开源的路由器操作系统），开始内置这一机制。参考 KoolShare 上的这篇帖子，作者经过测试，发现打开 fq_codel 之后，在 BT 下载占满上行带宽时，在线游戏依旧保持较低的延迟：可能是家庭路由里最好的 QoS——在线游戏者的福音 - KoolShare（见 8 楼、10 楼、22 楼、25 楼的测试结果）是否需要打开「智能队列」？前面介绍了 fq_codel 的一系列优点。那么，fq_codel 有没有缺点？是否需要我们打开？我的设备 CPU 性能不强，是否可以打开「智能队列」？根据这篇帖子，ER-X路由器在打开 fq_codel 时，最大支持 200M 左右的带宽。QoS (fq_codel/cake) Hardware Recommendation : HomeNetworkingER

周一18派早报：提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年昨天 06:56早报速览提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年0OpenAI 与苹果关系紧张，或考虑诉讼5索尼回应 AI 相机助手越改越差，称只是修图建议2Anthropic 警告私下交易的股权可能无效2运营商回应「零月租」传言4微软重新允许自定义任务栏与开始菜单2看看就行的简讯4少数派的近期动态0你可能错过的好文章0提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年5 月 15 日，预印本平台 arXiv 出台新规：如研究人员向该平台提交的稿件包含「无可辩驳」的 AI 生成内容，论文所有署名作者将被处以长达一年的封禁；且其未来提交的任何文章都必须先在其他期刊通过正式的同行评审后，arXiv 才会予以收录。「无可辩驳」的证据包括虚假参考文献、不小心粘入的与 AI 对话内容等。俄勒冈州立大学名誉教授、arXiv 编委兼审核团队成员 Thomas Dietterich 在 X 上公布了这一消息。他指出，新规直接源于 arXiv 既有的审核标准。无论稿件中出现不当言论、抄袭、学术偏见、虚假参考文献还是误导性内容，所有署名作者都必须对这些 AI 生成的错误承担全责，而不能将责任推卸给工具。arXiv 是物理学、天文学、数学及计算机科学等领域全球最大的开源学术文献库。在这些学科的常规发表流程中，学者们通常会在接受期刊正式的同行评审前，先将论文作为预印本上传至 arXiv，以抢占首发权并获取同行反馈。因此，剥夺一年的首发资格并永久附加同行评审门槛，对相关领域人员而言是严厉的措施。375OpenAI 与苹果关系紧张，或考虑诉讼据彭博社报道，苹果与 OpenAI 为期两年的合作关系近期出现裂痕。因未能从 ChatGPT 与苹果系统的整合中获得预期收益，OpenAI 正与外部律所积极探讨应对方案，近期或将向苹果正式发出违约通知，为潜在的法律诉讼做准备。2024 年 6 月，苹果宣布将 ChatGPT 接入 Siri 等系统级功能。OpenAI 曾寄望借此获取海量订阅用户及数十亿美元收入。然而实际体验中，该功能入口较深，需用户明确口头提及 ChatGPT 才能唤醒，且信息呈现方式受限。结果，绝大多数用户仍倾向于使用 ChatGPT 应用。一位未具名 OpenAI 高管表示，苹果在产品端甚至没有尽到诚实努力的义务；合作不仅未能创收，反而损害了其品牌形象。除软件合作遇冷外，两家公司在硬件与人才领域的博弈也加剧了紧张局势。OpenAI 此前收购了由前苹果首席设计师 Jony Ive 等人主导的初创公司，并利用高薪从苹果硬件团队挖角，试图打造新一代 AI 设备，这引发了苹果高层的强烈不满。同时，苹果也已着手引入竞争机制，计划在即将发布的 iOS 27 系统中全面开放生态，接入谷歌 Gemini 和 Anthropic 的 Claude 等其他主流大模型。知情人士表示，OpenAI 针对苹果的任何法律行动，预计都会推迟到其与马斯克的诉讼案结束后才会推进。目前各项计划尚未做出最终决定，OpenAI 内部仍保留了在庭外解决争端的希望。OpenAI 与苹果的发言人均拒绝就此事置评。我换到国产手机AI都没怎么用过，不是不会用，而是习惯了以前苹果装个AI软件的使用方法了。果子在ai上的每步操作都很迷惑感觉不如自己装一个本地小型LLM App用ShortCut打通生态。我还是关心什么时候有苹果智能，感觉一不留神，好几年过去了，我换15p的初衷之一就是给智能做准备，现在都要17了，我也是吃上大饼了623索尼回应 AI 相机助手越改越差，称只是修图建议近日，索尼最新旗舰 Xperia 1 VIII 的官网页面引起关注。在演示 AI 相机助手的效果时，官方样张的效果在观感上反而明显劣于原图。这引来了 MKBHD、Nothing 创始人裴宇等人的公开调侃。对此，索尼在 5 月 15 日解释称，「AI相机助手」并不会在拍摄后自动应用修图，只是根据当前场景和拍摄主体，提供 4 种不同创意方向的参数设置建议，用户可自由选择接受这些建议或坚持使用自定义设置。与此同时，官方最新展示的对比样张在视觉效果上有了显著改善，去除了首批样张中饱受诟病的画面发灰和过度曝光问题。但这依然未能完全消除外界对其影像审美判断的疑问，也不清楚营销团队最初为何会挑选那些效果不佳的照片来展示核心卖点。感觉那朵花还可以，其他的都是💩️有人关注就算不错了10122Anthropic 警告私下交易的股权可能无效据《华尔街日报》报道，Anthropic 近日更新官网公告，宣布将不承认未经其董事会批准的二级市场股权交易，并点名了一系列未授权机构。Anthropic 警告称，由这些机构提供的任何该公司股权或相关权益转让均属无效，且不会记录在公司的股东名册上。这一声明在个人投资者中引发了广泛恐慌。由于普通投资者通常无法直接购买未上市「明星公司」的股权，大量二级市场交易平台应运而生。它们设立特殊目的实体（SPV），将众多散户资金打包后投资于目标公司。然而，这种 SPV 结构日益复杂，有的多层嵌套，有的把多家目标公司的股权捆绑出售，导致分配机制存在很大模糊地带。普遍认为，Anthropic 此举意在通过清理违规交易来精简其股东名册，为其计划于今年晚些时候的 IPO 做准备。随着近年来私营科技公司的估值不断飙升，包括 OpenAI 在内的多家独角兽企业，都在采取类似措施限制涉及其股权的二级市场交易。面对投资者的质询，多家被点名的交易平台迅速回应，表示涉及 Anthropic 的基金和交易均已获得批准或符合转让要求。专家指出，如果 Anthropic 并未在事前对特定交易结构做出限制，可能无权撤销这些持仓，相关争议最终或将诉诸法庭。可能性不大这种模式在中国有么？发表评论运营商回应「零月租」传言据央广总台报道，近日，不少互联网博主在短视频里介绍，三大运营商即将取消月租费，改为按量计费模式，甚至还列出了推出该政策明确的时间节点。但记者梳理市场现状发现，目前只有中国联通上线了零月租、设最低消费门槛的「魔方」按量套餐；中国移动和中国电信暂未推出同类产品。例如，传言中被广泛提及的「联通魔方」发布于今年 4 月底。该产品采用纯按量计价模式，国内语音和短彩信均为 0.1 元/分钟或每条，流量则实行阶梯计价，单价在 1 元至 1.5 元/GB 之间。但联通客服表示，该套餐并非绝对免费，而是设有地区差异化的最低消费门槛，例如北京地区起步低消为 55 元，月结时不足该金额将自动扣齐。同时，联通原有的各档位固定月租套餐依然保留并可正常办理。中国移动和中国电信的客服人员则都表示，公司曾推出过零月租模式，但目前没有相关计划。针对完全零月租的可能性，通信行业专家指出，此类模式脱离了基础通信网络的运营逻辑，庞大的固定网络维护成本最终也必然会转嫁到其他用户身上，甚至可能导致整体资费上涨。比起追求不切实际的零月租噱头，当前消费者对资费的真正痛点在于不合理的老旧月租套餐，以及套餐外的标准流量资费。没有专门办的流量卡划算捉虫：“央广总台”→“总台央广”“零月租” “起步低消”世界需要提速降费。3微软重新允许自定义任务栏与开始菜单5 月 16 日，微软发布博客文章，宣布将在未来几个月内推出一系列更新，大幅增强任务栏和开始菜单的外观自定义功能。这标志着一系列经典 Windows 功能的回归。对此，Windows 业务负责人 Pavan Davuluri表示，个性化定制是 Windows 的基因；微软已听取用户反馈，决定重新将界面控制权交还给用户。目前，部分选项已在 Windows 预览体验计划中面向测试人员推送。本次更新中，用户可以再次像在旧版系统中那样，将任务栏停靠在屏幕左侧，或将其缩窄以节省屏幕空间。同时，开始菜单新增了尺寸调整选项，并允许用户自定义显示模块，例如直接关闭「所有应用」或「推荐」（更名为「最近使用」），仅保留自己固定的应用。此外，用户可以隐藏「开始」菜单中的用户名和头像。这些测试功能目前仍有已知问题，例如开始按钮未对齐、通知弹出位置异常等。微软表示将在正式向公众推送前修复这些问题。此外，更多传统的自定义功能，例如拖拽任务栏边缘手动调整高度等，目前也已纳入开发计划，预计将在今年内陆续上线。Windows 11 自发布以来，就因移除任务栏位置调整等诸多经典自定义功能而饱受争议。近年来，微软一度将系统开发重心全面转向 AI，在系统中强行植入 Copilot 功能，引发了用户的强烈抵触。迫于压力，微软于今年早些时候宣布做出调整，重新关注质量。什么时候把win10菜单给我回归一下😊我选择Windows Server 2025，躲过UWP和React两大石山。1看看就行的简讯据Fast Company报道，知名开源密码管理器 Bitwarden 近期在未发公开声明的情况下，悄然大换血高管团队，今年 2 月起，原 CEO、CFO 先后转岗或离职。新任 CEO 有并购及私募股权交易背景，让外界对其未来的资本运作产生猜测。此外，Bitwarden 官网出现多处调整，个人版产品页面移除了「永久免费」的标签（官方解释为营销团队疏忽）；不久前，Bitwarden 在未显著通知的情况下，将高级版订阅年费从 10 美元翻倍至 20 美元。此外，Bitwarden 还在招聘页面修改了企业核心价值观，去除了原有的「包容」与「透明」两项原则。近日，谷歌开始测试一项新的云存储空间政策：新注册的用户如果不绑定手机号，其初始免费存储空间将从传统的 15GB 缩水至 5GB。谷歌官确认，该政策目前仅处于区域性测试阶段，且现有老用户不受此规则影响。谷歌发言人表示，这项新规是为了继续为用户提供高质量的存储服务，同时鼓励用户完善信息以提升账户安全性和数据恢复能力。然而，许多用户推测谷歌此举的实际目的是为了打击恶意注册大量小号以薅取免费存储空间的行为；也有部分用户对强制绑定手机号背后的隐私数据收集及潜在商业用途表示担忧。据9to5Google报道，谷歌随无屏健身追踪器 Fitbit Air 推出的订阅服务「健康教练」（Health Coach，由原 Fitbit Premium 升级而来）存在 AI 幻觉问题。测试中，AI 教练凭空捏造了一项 5.2 英里的跑步记录；在被指出错误时，还反问用户是否忘记输入过该次记录。经过纠正后，这条虚假数据依然残留在应用主页上。最近对bitwarden挺失望，要么是性能问题，这次又出个管理团队问题。开源不是万灵药，信誉才是基石啊这几年一直用Bitwarden 呢，要变味了吗白赚一次跑步，哈哈哈哈比特瓦恩德可以换了(｡･ˇ_ˇ･｡:)4少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购发表首评你可能错过的好文章看剧不卡顿、游戏低延迟：「智能队列」这样改善你的冲浪体验为什么我觉得好的东西，媒体评测却给了低分？本周看什么 | 最近值得一看的 8 部作品装了啥：编辑部用什么转写音视频内容？[会员免费] Apple WWDC26 全球开发者大会，来与少数派一起看快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验你真的需要墨水屏设备吗？发表首评90扫码分享目录 0讨论我来说一句发布发表评论发布

2025 年度旅行——4 天 3 晚青海湖小环线主作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥联合作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥05/14 09:38Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。一次完美的旅行，对我而言，不仅仅是去被人走过千遍的地方打卡一圈，而是藏着三次截然不同的体验——出发前对目的地的想象与期待，旅途中与自然的亲密接触，以及归来后翻看照片时，重新发现那些当时未曾留意的细节，细细回味，查阅资料，让我对这次旅行有更深刻的理解。那么就让我们随着这些文字与照片，回到美丽的青海湖。契机：作为刚工作的年轻人，在国内的职场环境下，真正无理由可申请的假期只有 5 天，所以我和身边的小伙伴们都分外珍惜这五天的时光，每年的年度旅行我们都会避开人流，详细规划。由于十一假期要去参加两场婚礼，为了节约交通费，我们决定先开车到 2 号参加婚礼的宝鸡，把小 P（我的小鹏 MONA M03，具体可参阅我之前的文章）停在了宝鸡高铁站，考虑旅行时间较为紧凑，且西北自然景区交通补能条件可能较差，我们选择了到当地租车完成青海湖小环线行程。带着宝鸡擀面皮和肉夹馍，一行人正式开始了旅途，三个半小时后，到达了西宁火车站，初入高海拔地区，并未感受到明显不适，方便的是，租车平台在西宁站地下停车场就有取还车点，我们直接将行李放上车，省去了搬来搬去的劳顿。Day1 西宁-茶卡 全程约 457 km西宁-塔尔寺5A 级旅游景区，门票：70 元，游览时间约 2 小时塔尔寺是藏传佛教格鲁派的著名寺院之一，始建于明代，最初是为纪念藏传佛教格鲁派创始人宗喀巴大师而修建，据传宗喀巴大师出生于此地，后人于其诞生处建塔纪念，塔尔寺因此得名——「塔尔」即「有塔之地」。塔尔寺-八宝如意塔游览中心和寺庙门口都有大量的讲解人员，价格也不贵，但是本人信仰唯物主义，故只将这里当做名胜古迹参观，故没有请讲解，如果对藏传佛教及寺庙背后的故事比较感兴趣，请个讲解或许是更好的游览方式，或者是像我一样，回来翻阅照片时再仔细搜索相关资料。Tips：讲解会介绍各个佛像的「作用」（不知道这样叫是否合适，或者说负责保佑的事情），故可能会被架在那里为了讨好彩头而「必须」捐献资金，介意者可以像我一样自由游览。由于是节前，故参观人员并不很多，与前往朝拜的信徒或僧人可以说是一半一半，没有了熙熙攘攘的人群，也得以让我更加细致地观察这座寺庙和「磕长头」远道而来的人们。前来朝拜的僧人进入寺庙，浓郁的香火味和木质结构因缺乏通风的陈味扑面而来，看着宏大的佛像、墙上摆放整齐的诵经坐垫，的确能感受到不一样的文化氛围，这里安静、鲜活，尤其是看着不知道从哪里启程的「磕长头」的信众，虽因路途的遥远衣衫褴褛，但想必也经过这一路的艰辛而悟出了些许道理，可能出发时也不知道为什么，但走着走着就明白了吧。塔尔寺中的白塔募捐处，与时俱进地，也支持扫码支付塔尔寺有大大小小的寺庙很多间，在我看来有的都别无二致，于是重点转了几个后，就启程回西宁吃饭，整备氧气瓶，应急食物等装备，正式奔赴青海湖。海南藏族自治州共和县 南山观景台 & 青海湖二郎剑景区前者免费，后者为 5A 级旅游景区 门票 72 元，游览时间约 2 小时一点半在西宁市吃完炕锅后，我们开始驱车前往青海湖，约 2 个小时车程，我们到达了本地的主要目的地青海湖的南侧景区 青海湖二郎剑，在进入二郎剑景区前，可以导航到「南山观景台」 从山上远眺青海湖和远处群山的全貌，但是山路较窄且有工程车辆时不时通过，故要格外注意安全。景区本身没有人文景观，但其还是我国早期鱼雷的试验基地，但是因为我们到达时间比较晚，并没有参观博物馆，而是直奔到湖边。青海湖是我国水面面积最大的咸水湖，水面平静辽阔，与远处的群山相辅相成，甚是开阔美丽，栏杆上有硕大干净的鸽子驻足，在几番伦理拍照后，我靠在栏杆旁，看着远处的夕阳洒在地面，耳边传来阵阵鸟鸣与风声，工作与生活的喧嚣远离脑后，只剩下全身心的与自然接触。云层、远山、草原，形成了震撼的三明治拨云见日，青海湖边的草还是绿色的岸边排队等待「薯条」投喂的海鸥码头景观随着夜幕降临，我们离开二郎剑景区，沿着青海湖开车前往茶卡住宿，虽然只有 200 多公里的行程，但是因为沿途工程车辆很多，还有很多地方持续修路，有的路段有连续的 U 型弯，还有双向仅有一条车道通行的地方，还要翻越几个海拔 3800 米的垭口，成了这一路最惊险刺激的行程，一路上小心翼翼，战战兢兢，终于在晚上 10 点到达了茶卡镇，幸运的是镇上的餐馆都还在营业，找了家当地美食，简单吃饭后休息。夜幕降临的倒湖茶公路Day2 茶卡-祁连县 全程约 372km茶卡盐湖4A 景区 带往返小火车套票 167 元「茶卡」在藏语中意为「盐池」，该地以出产「大青盐」而闻名，参观盐湖也是此地的知名旅游项目。茶卡盐湖的景区共有两个，天空之镜是较老的景区，主要以自然风光为主，天空一号主要以商业化项目为主，作为外地游客，我们肯定更希望看到原始的自然景观，于是选择前往天空之境景区。景区很大，如果想要游览每个景点建议留出一整天的时间，游览顺序建议先坐小火车到终点天空之镜，该站处于盐湖的最深处，盐水异常的干净，也最为出片，景区也放置了很多拍照打卡点位，可以尽情拍摄，心痛没有带无人机。在晴天的时候，由于天空和地面都是白色，很容易晃眼，推荐大家一定要戴墨镜前往，通过墨镜观察湖面，盐湖会看起来像翡翠一样更加绿色透明。套票包含鞋套租赁，可以套上下湖走走，踩在大颗的盐里,很像沙滩的感觉，景区维护的十分干净，可以尽情的亲近自然。有趣的是，景区为了怕大家在盐滩大片的白色中无聊，设置了很多拍照用的摆件，例如国家地理的相框等等，还在大屏幕上贴心的轮播拍照指导，期望每个游客都能满意而归。盐湖与远处的山峰盐湖码头废弃的采盐轨道之前采盐的轨道深入盐湖，在墨镜下，真实的观感颜色会比照片深很多。接驳小火车行驶在盐湖中间，干净得不像国内巨幅盐雕从盐湖出来，我们在县城稍作休息，开始启程前往祁连县，有了昨夜惊险的行程后，尽管沿途的国道可能风景更好，但是我们选择了高速较多的线路。这一程的路况还是很好的，就算是国道车辆也较少，道路笔直，视野十分宽阔，草原、远山、白雪、天空层层分明。在行驶过程中，不时的有大群的羊羔和牦牛穿过公路，不必惊慌，一会就有牧民骑着摩托驱赶，让它们加快穿越的脚步。公路与群山草原上大批的牛羊傍晚的草原，远处能看到夕阳洒在雪山上落日下的牦牛剪影在夜色中，我们抵达了祁连县，在县城中吃过一顿特色烧烤后，进入梦乡，准备迎来第三天的行程。Day3 祁连县-岗什卡雪山-西宁 全程约 320km祁连卓尔山景区门票 60 元 上山小巴 20 元 登山电梯 10 元这个景区的确是全程我最惊艳的景区了，近能看到瑞士小镇似的田园风光，中能看到美如画的丹霞地貌，远还能远眺远处的阿咪东索雪山（海拔 4695 米），而且在山顶可以拍跟雪山的合照，大气、干净、巍峨。自驾车辆可以停在景区停车场，然后乘坐接驳中巴约 10 分钟到达检票口。卓尔山的两侧，风景差别迥异，一侧是连绵的雪山，另一侧是仿佛瑞士小镇的丘陵和小镇。阿咪东索雪山卓尔山的另一侧，仿佛欧洲小镇丹霞地貌与草地交融最让我惊讶的是这个景区竟然还修建了方便爬山的观光电梯，10 元一位，大概有三段，果然是人性化啊。景区电梯中午约 12 点，我们离开景区，在祁连县的一家川菜馆简单用餐后，我们驱车开始前往本次旅途的最高海拔岗什卡雪山。在县城开车的过程中，等红灯时，我看到一位初中年龄的藏族小伙盯着我们开的岚图梦想家，眼中闪闪发光，不禁让我想起小时候的自己，看到喜欢或没见过的车也会驻足观看，充满憧憬。愿他长大后也能开上自己喜欢的车，能够做自己喜欢的事情。岗什卡雪峰景区免费 上山摆渡车 25 元岗什卡雪山坐落于青海海北州门源县境内，是祁连山脉东段的最高峰，被誉为「雪山之尊」，因为设施和难度都尚可，也被雪山爱好者们称为「人生第一座雪山」。主峰海拔 5254.5 米。游览模式同样为将车辆停在景区停车场，然后乘坐接驳车到雪山脚下，正式游览区域的海拔在 3800 米左右，建议随身携带一个氧气瓶，下车之后，可以看到湛蓝的连心湖，水来自于冰川融水，干净，清冽，与子洲植被的荒凉的深褐色形成鲜明的对比，十分美丽。连心湖到达时天气并不是很好，岗什卡雪峰被薄雾笼罩天然泉水，因还有行程，害怕拉肚子，没能如愿品尝一下离开景区后，雪山逐渐露出真面目，攀登还是留给下次吧沿着冰川的遗迹向上走，就能看到岗什卡雪峰徒步攀登的起点，景区也提供马匹，可以跟着向导继续向前，到七彩瀑布，因时间和个人体能的考量，我们浅尝辄止，在河道附近游荡些许时间后，便下山返程。临近傍晚，启程返回西宁市，一顿丰盛的海底捞过后，快乐的青海湖小环线结束。Day4 西宁市-青海藏文化博物馆-特产采购最后一天，伴随着蒙蒙细雨，我们打车前往青海藏文化博物馆，博物馆不大，纵使是十一期间人也不是很多，但内容着实丰富，满墙的唐卡和在电影里看到的藏传佛教僧人服饰，还有精致的贵族金器。在博物馆的文创区，我购入了一个小的转经筒，在回程的路上一直把玩，虽不信奉佛教，但也觉得甚是有趣。唐卡四王金杯嵌绿松石立凤金头饰博物馆文创-转经筒，38 元虽不算划算，但做工十分精致离开之前，为了给家人和朋友带些纪念品，我们选了一家超市类型的纪念品商店，购买了牛肉干，藏红花，枸杞，青稞酒等特产，虽说也不一定划算，但胜在明码标价，将行李箱装满后，顺利返程。后记作为 2025 年的年度旅行，本次 4 天的青海湖小环线我甚是满意。人少不堵车，行程不特种兵，在当地感受美食和壮美的自然风光，一路上与好友们分担开车，聊对景色的感受，互相拍照，点评食物，虽偶有几个工作电话烦扰，但总体悠然自在，放空身心，一切都是那么的安静，与自然深度的接触，与北京的喧嚣形成了鲜明的对比。对于出行方式，自驾无疑是最有的选择，这边深入自然，道路虽崎岖，但也属于铺装路面，且西宁市神州、一嗨等在高铁站都有网点，取送车方便，且自驾可以自由控制行程，根据游览时长方便自己调整。对于住宿，我们对酒店的干净程度略有要求，于是全程都选择了以华住会为目标的连锁快捷酒店，以下是所有天的酒店名称和十一节前的价格，个人觉得是十分舒适的。Day0 桔子西宁火车站省人民医院酒店 双床房 331元Day1 星程茶卡盐湖酒店 覆氧双床房 314 元Day2 宜必思尚品海北祁连卓尔山酒店 421 元Day3 桔子西宁火车站省人民医院酒店 双床房 412 元关于本次行程的花销，纵使算上从北京出发的路费，人均也仅 3k 左右，可以说是十分划算了，主要的开销就是租车和酒店。支出一览表关于卫生间，景区和路途服务器的卫生间都比较完善，因同行有女生，我们还准备了移动卫生间，但从我们的行程安排走下来，并没有用到，个人觉得要是按我们的行程游览也无需担心。下一站，出发！2025 年的年度旅行结束，2026 年也可以开始规划起来了，新疆还是海南，希望能远离喧嚣，享受美丽壮阔的祖国风光。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀1915扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@北鸮© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#Matrix首页推荐#旅行19等 19 人为本文章充电扫码分享举报本文章别惹小炸毛一个学酥关注

打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试主作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.联合作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.昨天 15:00ARG 即侵入现实游戏（Alternate Reality Game），是一种基于现实世界，融合各种多媒体（如网站、电话、社交媒体等）技术的互动式解谜游戏。由于其与现实的高度结合所带来的沉浸感，一直以来被许多玩家所追捧。如果你在 2016 年玩过《守望先锋》，相信你对当时新英雄「黑影」的登场一定记忆犹新。从 2016 年夏季开始，一直到 11 月的暴雪嘉年华，暴雪通过在宣传视频、官网、社交媒体发布了一些隐藏着密码的内容，由全球玩家共同解谜，不但揭示了游戏世界观下的一些故事，更是顺理成章地将身为黑客的「黑影」带到大家的视野里。经过几个月的发酵，在暴雪嘉年华上大屏幕被「黑影」控制，随后开始播放英雄动画的时候，更是满足了玩家们期待已久的心。「黑影」的宣发是一场相当成功的 ARG，时至今日仍有玩家对其念念不忘。我本人也是 ARG 的爱好者，我非常享受这种打破虚拟与现实边界的游戏乐趣。前段时间，一款名为《大学生登山失踪事件》的网页解谜游戏突然火了起来。比起此前一些高度依赖技术知识的网页解谜游戏，《大学生登山失踪事件》在不需要任何 IT 知识背景的情况下，依然有着抽丝剥茧逐渐接近真相的推理爽快感。自此之后，小红书上越来越多的人开始制作这一类网页解谜游戏，并为它们赋予了 ARG 的名号。这些游戏流程往往不长，差不多 1-3 小时就能通关。因为是网页解谜游戏，只需要一个浏览器就能游玩，很适合上班时间摸鱼，游玩门槛非常低。这篇文章将会推荐一些这段时间我玩过的质量不错的游戏，虽然一些制作精妙的游戏确实带有 ARG 的影子，但其本质上仍是基于网页的解谜游戏。为了方便大家后续搜索，本文会将这些游戏统称为 ARG。大学生登山失踪事件🔗https://missing.shiroki-y.top/introduction.html在游戏中，玩家需要通过从弟弟衣物中发现的「福安菜家常馆」的线索为开端，逐渐深入一场黑色阴谋找到弟弟失踪的真相。游戏整体难度不高，但在游玩过程中最好能用纸笔或软件记录进度，这能够帮你清晰地理出游戏脉络。游戏的核心玩法其实非常简单，只需要阅读页面，找到页面中可疑的关键词，在页面搜索栏里进行搜索即可推进剧情，这其中也会有一些用户名、密码、口令、字谜等需要玩家进行推理和猜测。网页页脚处会标记当前页面进度，既可以方便定位，也能够明确线索的先后顺序。剧情方面，《大学生登山失踪事件》并不会在一上来就给玩家塞一些晦涩难懂的概念，而是从大家熟悉的普通网页开始。当玩家一步步抽丝剥茧逐渐发现事件背后的黑暗面时，令人不由得感叹于游戏的设计与巧思。同时，游戏内也藏有一些可供探寻的 ex 页面作为彩蛋，在解谜之外找到各种彩蛋也不失为一种乐趣。诡异广告调查事件🔗https://mp.weixin.qq.com/s/yeNYMZVRp6N-MESFJGk9rg这是一款与《大学生登山失踪事件》来自同一作者@銀河電燈_的游戏，世界观也与前作有关联之处，很是有趣。玩家将扮演一位被「优化」的人，因为在看剧时发现了某个已停播的节目广告，从而开始调查这个诡异的广告，并发现藏在节目与自己背后的神秘真相。与前作相似，这款游戏也将基于阅读与搜索推进剧情。在此基础上，它还加入了一些 ARG 元素，比如在街景地图中，找到现实里存在的文字招牌作为线索。并且在网页设计上也比前作更加成熟，贴合剧情上的时间设定，在网页上加入了许多彩蛋。剧情方面，本作有一定的惊悚元素。特别是在游戏过程中展示的一些内容可能会引起不适感，如果你对这类内容比较敏感，请务必提前做好心理准备。当然，游戏并没有 Jump Scare，更多是通过剧情与视觉设计所营造出一种的诡异氛围。三川米粉店🔗https://www.berrysignalscmfd.top/《三川米粉店》也是一款从餐饮店主页出发，探寻故事真相的游戏。玩家需要调查三川米粉店突然停业的原因，在调查过程中，玩家将逐步揭开这家米粉店背后的故事。本作的页面设计相对没有那么精致，但也融入了一些 ARG 的元素，部分谜题的答案并不会直接给出，而是需要玩家通过搜索引擎或其他工具进行查询才能获得答案。从剧情上来说，《三川米粉店》的设定会更加接地气一点，并且结局也会相对更温馨一点。比较适合对于惊悚元素不太感冒的玩家作为第一部作品来尝试游玩。永安温泉度假酒店🔗https://mp.weixin.qq.com/s/kKsnzFsLFwwm2nwv8DseAA与前面侧重解谜的作品不同，《永安温泉度假酒店》的解谜难度其实不高，反而更像是一种「规则怪谈」式的探索路径。玩家需要不断探索找到更高权限的员工账号密码，从而揭开温泉度假酒店背后的秘密。游戏在开始前需要玩家填写「报名信息」，在后续游戏中能够增强代入感。从游戏性的角度来看，这款作品的「解谜」爽快感或许没有那么强。但如果你对「规则怪谈」或者一些高概念的设定感兴趣，那么这款游戏依然是一个值得尝试的作品。404 工位消失事件🔗https://xsn1204.github.io/qmgame《404 工位消失事件》是一个特色鲜明的游戏，玩家将扮演一个实习生，接手其他人留下来的工作电脑，并从中发现一些未被处理的秘密。游戏在网页上复刻了一个形似 macOS 的「系统」，很大程度上增强了代入感。因为设定是在电脑系统，游戏中需要玩家在钉钉、微信等各种软件中挖掘线索，反而更有一种作为侦探在不断探索的感觉。但受限于网站形式和游戏体量，游戏中能够互动的地方通常都是为剧情线索服务的，与线索无关的地方大多都是不能点击互动的。而且游戏不能通过点击窗口左上角的关闭按钮回到上一层，而是需要通过浏览器的返回来实现。但这些小细节对于《404 工位消失事件》来说都算瑕不掩瑜，整体游玩体验还是值得肯定的。作者甚至还出了 2.0 版本，新的剧本进行了优化提升，如果对这一类型感兴趣的话也可以尝试游玩。找到我🔗https://www.kaixinmoyu.com/feijo/findfeijo/在游戏中，玩家需要扮演一位私家侦探，调查游戏作者失踪的原因。游戏在交互方面做得非常有意思，游戏不仅做了浏览器页面对线索进行搜集检索，还做了小灵通和笔记本，前者可以通过拨打特定的电话获取线索，后者则需要玩家阅读日记找到隐藏的内容。《找到我》本身是一个更重交互而非推理的游戏，游戏中融入了一些 ARG 元素，在最开始就需要到某篇公众号文章中找到一个密码才能继续推进。但游戏的剧本逻辑会稍微有些不够连贯，以至于中间可能会有一些容易卡关的地方，需要仔细翻阅此前出现的线索才能继续往下推进。另外，游戏中后期会有一道需要在现实中解算的谜题，这道题本身与游戏本身的节奏设计会稍显割裂。但整体来说，《找到我》依然是值得一玩的游戏，尤其是本身的交互设计和 ARG 元素的融合做得非常不错。在游戏作者的个人主页，也能找到作者开发的更多网页解谜游戏，如果你对这款游戏的风格感兴趣，也可以尝试其他作品。小结优秀的网页解谜游戏并不止文章提到的这些，光是在小红书上一搜，就能找到许多分享帖。希望本文能起到抛砖引玉的作用，欢迎大家在评论区分享自己玩过的，令你印象深刻的网页解谜游戏。> 简单、好用、专注的写作软件，少数派为你呈现 🚀> 守护孩子的天马行空，从零开始美术教育启蒙🎨486扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Lotta© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#解谜游戏#游戏#生活方式48等 48 人为本文章充电扫码分享举报本文章TpBe nothing, and everything.关注

派评 | 近期值得关注的 App主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注Vanilla少数派作者Vanilla关注Vanilla少数派作者关注化学心情下2少数派作者化学心情下2关注化学心情下2少数派作者关注大大大K少数派作者大大大K关注大大大K少数派作者关注Peggy_少数派作者Peggy_关注Peggy_少数派作者关注ElijahLee少数派作者ElijahLee关注ElijahLee少数派作者关注Snow少数派作者Snow关注Snow少数派作者少数派编辑部 等 7 位作者昨天 18:05欢迎收看本期《派评》。你可以通过文章目录快速跳转到你感兴趣的内容。如果发现了其它感兴趣的 App 或者关注的话题，也欢迎在评论区和我们讨论。AIDA：一站式日历和待办规划工具平台：iOS、iPadOS关键词：日程管理、待办清单@ElijahLee：Aida 是一款针对 iPhone 和 iPad 开发的日历和待办清单应用，来自任务管理工具 Sorted 的前创始人之一 Harry，它集成了待办清单、日历和时间规划工具，在一处显示、规划、操作。应用主界面就是每日时间线，顶部是当周 7 天的简要信息，应用中部大篇幅排列当天的时间块，你可以直接看到上午十点看电影、中午写派评、傍晚健身等日程，以及晚上记得给植物浇水的待办提醒，而不用在多个不同的 App 之间切换。应用的快速调整功能，支持在时间线中将事件按住并移动顺序，快速重排一天的事件，对应的时间点会自动修改，非常方便。使用右下角添加按钮，可以自由地为时间线加入行动或事件。行动即待办事项，支持输入文本并设置提醒时间。 Aida 创造性地设计了拖动设置时间的功能，将底部的柱状图左右滑动，就可以调整行动的提醒时间，每次拨动以 15 分钟为间隔，用它设置时间非常方便。当然我们也可以点击时间来精细设置时分。 Aida 事件就是日历中的事件，必须包含开始时间和结束时间，我们还可以选择日历类别、设置全天事件等。Aida 引入了 GTD 管理的项目功能，提供收件箱、任何时间、将来某日等默认分类。自定义项目可以将同样类别的活动，包括行动和事件进行集中收纳整理。Aida 全面集成了 iOS 日历，在应用内不仅可以查看 iOS 日历事件，对事件的修改调整还会实时同步。不过 Aida 目前没有接入 iOS 的提醒事项。 批量操作功能可以一次选择多个行动进行完成、改期、移动或复制，也能快速整体调整时间块。你可以在App Store免费下载 Aida，免费版功能受到很大的限制，仅支持 1 个日历的集成显示、无法调整行动的时间、添加行动和事件的数量有限。付费可以解锁高级功能，包括无限项目和时间块、多个日历集成显示等功能。费用为 58 元 / 月、298 元 / 年，698 元买断，应用提供 1 周免费试用。毫无特色 + Sorted 血脉 + 令人疑惑的定价现在任务管理之用自带的提醒事项和滴答清单免费版定价没问题，是我的问题。太貴了感覺😓发表评论Cogito：连接人与 AI Agents 的 Markdown 编辑器平台：macOS关键词：Markdown、AI@Vanilla：对于 Markdown 编辑器这一类应用，我想大家已经见得太多了。随着 AI Agents 的兴起，Markdown 这个文件格式可谓是「老树开花」，又重新获得了很多人的青睐。Cogito 是一款面向人和 AI Agents 共用的 Markdown 编辑器，它不像 Obsidian 一样偏知识库、不像 Typora 一样偏所见即所得、不像 iA Writer 一样偏写作，也不像 VS Code 一样偏开发者，它最大的特点是「文件优先、Agents 友好」。下面，就让我们一起来看看这个新时代的 Markdown 编辑器到底有何不同。Cogito 的文件库完全基于本地磁盘，没有专属格式，没有加密拦截，没有同步机制，所有的文件和文件夹组织形式都和 Finder 中的一模一样，完全映射，没有任何上手门槛。以我的使用习惯为例，我直接把 Codex 的工作文件夹设置为了 Cogito 的 Location，它会自动抓取 Codex 工作区的所有文件夹和 Markdown 文件，按照日期排序，清晰明了。选择任意一个 Markdown 文件后，Cogito 默认展示它的编辑状态，我们可以随意地按照自己的意愿进行编辑。点击右上角的第一个按钮，则可以在编辑状态和预览状态之间进行切换，对应的键盘快捷键是 Command + R。我经常会借助 Codex 帮我生成一些行业或者公司的研报，最终以 Markdown 格式来进行输出，通过 Cogito 我就可以非常方便地进行编辑和阅读，不需要反复地在 Codex、Finder 和 Markdown 编辑器之间进行来回切换。点击右上角的第二个 Outline 按钮，我们可以快速查看当前文档的大纲，点击标题可以直接跳转到相应的章节。Cogito 除了对 AI Agents 友好，还可以在应用中直接调用本地的 AI Agents 对应用内显示的文档进行总结、分析、编辑等操作。总的来说，Cogito 对于像我这样频繁使用 AI Agents 和 Markdown 文档的人来说是一个非常不错的 Plan B，和主力写作的 iA Writer、用作知识库的 Obsidian 等进行区隔，打开即用，用完即走，非常高效。Cogito 可以在官网免费下载使用，目前处于公开测试阶段。好吧，不能所见即所得。agents友好其实有点尴尬，一个阅读器+一个terminal就可以完全的代替掉这个软件目前的全部功能？甚至codex的app自身也有文件阅读功能了。发表评论AndDrive：无缝连接你的 Android 设备和 Mac 电脑平台：macOS关键词：文件管理、手机管理@化学心情下2：虽然现在有不少工具能让 Android 手机和电脑无线互传文件，但「传文件」和「管理文件」是两回事。比如想批量删除 Android 内部存储 Pictures 文件夹下的截图，用传输工具就很别扭，这种场景更适合在电脑端的文件管理器里完成，可以借助筛选和预览功能高效处理。Windows 上这件事不需要任何第三方软件：数据线一插，选择「文件传输」，Android 就被模拟成 MTP 设备，像移动磁盘一样直接管理即可。但 Mac 上就麻烦了：MTP 是 Windows 的专有协议，macOS 想访问 Android 内部存储，要么走 USB 调试，要么依赖第三方 MTP 实现。OpenMTP 算是其中比较知名的一款，但项目两年多没更新、稳定性一般，而且用的是自己写的文件管理器，而非访达，体验称不上顺手。那么有没有能直接用访达管理 Android 文件的工具？接下来要介绍的 AndDrive 就是，它由个人开发者开发，没有走 MTP 路线，而是基于 Android 官方的 ADB 调试通道实现文件管理。安装后，AndDrive 常驻菜单栏，支持 USB 和 Wi-Fi 两种连接方式。使用前需要在 Android 上开启 ADB 调试并授权当前 Mac，再按提示在手机上安装配套的 AndDrive Helper，之后访达里就能看到 Android 的目录了。AndDrive 的挂载方式可以选 SD 卡目录、根目录或不挂载，看到的内容会不同：SD 卡目录只显示内部存储，根目录则多一个系统根目录（没 root 的话很多文件依然动不了）。日常使用可以直接在访达里预览文件、按时间筛选，也能把电脑上的文件拖进 Android 对应文件夹，比在手机自带的文件管理器里翻找直观得多。因为走的是 ADB，AndDrive 顺带还实现了屏幕镜像功能。点击 AndDrive 的「屏幕镜像」，当前 Android 设备的桌面就能投屏到 Mac 上，体验和 macOS 的 iPhone 镜像类似，可以直接操作应用、用中文搜索，还支持剪贴板同步、录屏、截屏等。比起手敲 Scrcpy 命令，门槛会低不少。总的来说，AndDrive 虽然没有用更成熟的 MTP 协议，但用 ADB 这条 Google 官方推荐的方法，让访达管理 Android 文件这件事变得简单可行，还补上了类似 iPhone 镜像的屏幕控制功能，值得有需求的你一试。你可以在App Store免费下载 AndDrive。它不是改名了吗？🤔发表评论Fauxx：与其被动防守，不如主动出击的隐私保护工具平台：Android关键词：隐私保护@大大大K：你的每一次搜索、每一个网络请求、每一段定位轨迹，都在被数据平台和广告商默默收集着。它们会通过这些碎片信息拼凑出你的身份、喜好、习惯甚至是经常交往的人，然后在暗地里交易你的信息、精准推送广告。这也是为什么你或者你的朋友随手搜了点什么，10 分钟后所有 App 和网页就开始围着这个话题给你推广告。因此市面上衍生出不少隐私保护工具，它们的思路大多都是「堵」——屏蔽数据上传、拦截 Cookie、混淆 IP 等等。但 Fauxx 这个 App 啧另辟蹊径：既然广告商一定要收集数据，那就把真实的你混淆在一堆「垃圾」当中。首次打开 Fauxx 会有一个引导流程，你可以填写你的年龄段、职业、兴趣之类的粗略信息。这一步主要是为了让 App 能够生成与你真实信息相反的混淆信息，当然，这一步你也可以选择跳过。Fauxx 并不强制你输入信息。完成后进入主界面打开开关就能启动混淆信息的保护，在这里你可以看到这段时间 Fauxx 执行了多少次「混淆」动作、当前的虚假人设是谁、混淆了哪些类型的信息。至于具体可以混淆的内容，可以在 Targeting 和 Modules 标签中设置混淆等级和类型。Targeting 里有三个可选的层级：第一层就是生成与你真实信息相反的混淆噪声，如果你一开始没有填写，Fauxx 则会随机生成；第二层是对抗广告商的用户画像，不过我实际测试下来会读取失败，但也不影响使用；第三层是「人设轮换」功能，Fauxx 通过生成虚假的身份信息，让广告商无从追踪。Modules 标签里则是七个具体的混淆模块，每个都可以单独开关。设置里还可以调节混淆频率、指定时间段等。那这些「假数据」是怎么骗过广告商的？通常来说，被安插在手机里的追踪器会根据设备 ID、IP 地址、DNS 等信息追踪这台设备上的数据访问，进而推测用户画像。而 Fauxx 会通过模拟搜索、链接访问等「类人」操作（比如搜索时的停顿、假装点击广告），规避机器人检测并成功让追踪器追踪到假的信息访问。Fauxx 可以说是转变了隐私保护的方向，从被动防守变为主动进攻。如果你也想换个思路给广告商添添堵，可以在GitHub免费下载 Fauxx。1024Roll：胶片「日日鲜」平台：iOS关键词：手机摄影、胶片模拟@Snow：当下，你用手机拍照真的可以发散出无限可能，不同的镜头、不同的滤镜、不同的手动参数、不同的画面调节项，甚至是不同的外设，只要你愿意钻研一番，就能拿捏各式各样的风格。不过，更多的选择往往意味着更多的「选择困难症」，如果你对于丰富的选择产生了一些焦虑，或者天性就不爱「折腾」，不妨看看 Roll 这款应用，它或许可以带你发现简单拍照的乐趣。打开 Roll，完成简单的授权流程后，你就会收获今日份的新胶卷。应用会简单阐述这份胶卷在感光度、对比度、饱和度等各方面的特质，有一些呈现阳光下温暖的黄色，有一些闪耀着玫瑰般的光泽，有一些甚至想带你进行「第三类接触」。和一般的应用内随机胶片不一样，Roll 所有用户的「掷骰」结果完全一致，无论你身在世界何处，在同一时刻都会使用同一款胶片进行拍摄，剩余可用时间会在胶片介绍页注明。时间有限，点击「Load film」启动拍摄。Roll 不用切换镜头，没有色彩调整，拒绝选择迟疑，它唯一提供的可选项只有用于应对暗光环境时的闪光灯，你只需要按下快门按钮，想想可能与大洋彼岸的陌生人用同一款胶片滤镜进行创作，多少有点浪漫呢。Roll 用来

周二19派早报：红魔发布红魔 11S Pro 系列手机等今年的新系统会有什么变化呢？早报速览红魔发布红魔 11S Pro 系列手机4腾讯音乐已完成对喜马拉雅的收购8蓝牙核心规范 6.3 正式发布0腾讯自研 AI 设计智能体平台 Ardot 现已开启公测5微软将允许用户修改实体 Copilot 键功能3中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告0马斯克诉奥特曼案被驳回1不妨一看的简讯4红魔发布红魔 11S Pro 系列手机5 月 18 日，红魔发布红魔 11S Pro 与 11S Pro+ 手机等新品。红魔 11S Pro 系列搭载第五代骁龙 8 至尊领先版与红芯 R4 电竞芯片，机身采用纯平设计，11S Pro+ 还有透明水冷可视化设计，屏幕为 6.85 英寸 144Hz 无挖孔无刘海设计，采用屏下前置摄像头与 3D 超声波指纹解锁，搭载新思触控芯片，支持 3000Hz 瞬时触控采样率。系统内置 PC 模拟器，搭载 3.5mm 耳机孔与红外功能。11S Pro 电池容量 8000 mAh，支持 80W 快充；11S Pro+ 电池容量 7500 mAh，支持 120W 快充，两机型均支持旁路充电。红魔 11S Pro 有暗夜骑士与银翼战神两配色，12GB+256GB 售价 5499 元，国补后 4999 元。红魔 11S Pro+ 有氘锋透明银翼与氘锋透明暗夜两配色，12GB+256GB 售价 6199 元，16GB+512GB 6999 元，16GB+1TB 7999 元。另有红魔碳纤维游戏鼠标、红魔散热器 8Pro、氘锋能量卡 Pro 磁吸充电宝等配件发布。来源红魔是为数不多的纯平机身了背后是MagSafe 结构？屏下摄像头技术已经成熟了吗，怎么感觉其他家很少用呢，还有什么缺点吗1腾讯音乐已完成对喜马拉雅的收购腾讯音乐于 5 月 18 日在港交所发布公告，宣布根据并购的条款，收购喜马拉雅的事项已经完成。据此喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为 12.6 亿美元的现金及最多总数 1.75 亿股本公司 A 类普通股（包括本公司于交割时新发行的 A 类普通股，以及本公司授出的以股权为基础的激励相关的 A 类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。来源完了 喜马拉雅 也要有低级充值会员 中级充值会员 高级充值会员了买京东会员送的喜马拉雅会员还没用过呢全部讨论 8发表评论蓝牙核心规范 6.3 正式发布蓝牙技术联盟上周末发布了蓝牙核心规范 6.3，作为半年一次发布计划的一部分，此次更新引入了多项新功能，可进一步提升测距精度、扩展接口容量并提高射频效率。本次更新新增蓝牙信道探测内联 PCT 传输、蓝牙信道探测特定物理层 RTT 精度、蓝牙位资源扩展以及蓝牙 ACP 和 C/I 限值放宽。蓝牙技术联盟表示，联盟成员在描述产品支持的蓝牙功能时，应避免直接引用产品所依据的蓝牙核心规范版本号（例如蓝牙核心规范 6.3），而应在产品包装、文档和推广材料中清晰、准确地描述其产品支持的具体蓝牙功能，特别是与用户需求最相关的部分。来源11腾讯自研 AI 设计智能体平台 Ardot 现已开启公测腾讯云于 5 月 18 日发布公告，正式宣布 AI 设计智能体平台 Ardot 开启公测，官方称其为「一个 AI 驱动的产设研协作平台」。Ardot 支持一句话生成可编辑设计稿，AI 可快速生成 App 页面、官网、海报、插画、PPT 等；支持调用团队自己的业务组件库，一键生成符合规范的可编辑稿；支持直接导入 Figma 文件，完整保留原有布局、样式和组件。Ardot 对代码友好，支持设计一键转代码。设计稿，连同变量、组件、布局数据等设计细节数据，直接拉进 CodeBuddy，一键代码还原；也兼容 Workbuddy、Cursor、Claude Code 等 MCP IDE。Ardot 支持多人在线实时评论、标注反馈和版本对比，配备智能权限中心、全方位行为追溯和无缝交接机制，确保协作规范、有序。此外，Ardot 微信小程序也即将上线。来源与此同时猿辅导家的妙多 / Motiff 即将停服像素级啊确实太像了，感觉就把配色 从紫色改成了绿色🐶[图片]支持导入 Figma 文件🐶13微软将允许用户修改实体 Copilot 键功能5 月 18 日，Windows Central 发现微软上线一份新支持文档，表明自 2024 年起笔记本厂商在 Windows 11 设备上预装的 Copilot 实体按键影响了用户的工作流，他们将在今年晚些时候推送一版 Windows 11 更新，允许用户将 Copilot 键重定义为被替代的菜单键或右 Ctrl 键，用户将在设置 > 蓝牙和其他设备 > 键盘中找到相关选项。来源终于顿悟了？国内都用不了，按它干什么？微软终于发现了111中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告5 月 18 日，中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告。2026 年 4 月，中国游戏市场实际销售收入 303.69 亿元，环比下降 3.51%，同比增长 11.04%。其中，中国移动游戏市场实际销售收入 224.30 亿元，环比增长 0.71%，同比增长 9.83%；中国客户端游戏市场实际销售收入 68.66 亿元，环比下降 12.77%，同比增长 18.40%。中国自主研发游戏海外市场实际销售收入 20.32 亿美元，环比下降 5.02%，同比增长 30.78%。来源发表首评马斯克诉奥特曼案被驳回加州奥克兰联邦法院的陪审团于 5 月 18 日作出一致裁决，驳回了埃隆·马斯克对 OpenAI 及其 CEO 萨姆·奥特曼等人的诉讼。经过为期三周的庭审，由 9 人组成的陪审团仅用不到两小时便达成裁决，认定马斯克指控 OpenAI 违背成立初期慈善承诺的起诉已超过诉讼时效。2024 年，身为 OpenAI 联合创始人之一的马斯克指控奥特曼、总裁格雷格·布罗克曼以及主要投资者微软公司背叛了该机构「为人类福祉开发非营利性 AI」的初衷，通过转为商业化运作并与微软结盟来谋取巨额私利。马斯克此前曾为 OpenAI 提供早期资金，并在此案中索赔最高达 1500 亿美元。陪审团认定，马斯克早在 2021 年就已经知晓 OpenAI 转向营利性架构的事实，因此其 2024 年的起诉已经超出了三年的法定诉讼时效。辩方展示的证据表明，马斯克自己当年在未能获得 OpenAI 控制权之前，也曾提议过转为营利性结构。此次判决为 OpenAI 今年晚些时候的 IPO 扫清重大法律障碍。马斯克本人则在 X 上发文，批评该判决仅仅基于「时间程序问题」，而并未对案件的实质正义进行裁决。来源老马：不听我控制的就是不正义14不妨一看的简讯据外媒 Wccftech 的报道，根据一份去年泄漏的运输清单中的信息显示，Googlebook 笔记本或将搭载英特尔 Panther Lake 12Xe3 核心处理器。运单的信息显示，该产品使用酷睿 Ultra Series 3 系列处理器（即 Panther Lake），并搭载 12Xe3 图形单元。并且在 Google 与英特尔的相关文件中，有一个相同的代号 Q7AP 出现。而该编号对应 Panther Lake 12Xe 主板，对应的 Chromebook 产品代号则为 Felino。该笔记本采用 16 英寸设计，配备 32GB LPDDR5X 内存、1TB 存储空间。根据英特尔公布的产品阵容，12Xe 应该是 Panther Lake 的高端型号，至高可选 16 核心 CPU，功耗范围 25W-80W。如果此次曝光准确那么 Googlebook 很可能采用酷睿 Ultra Series 3 系列处理器。来源据微博用户金猪升级包，英特尔在 5 月 18 日的第三代英特尔酷睿处理器新品分享会上，介绍了英特尔 Project Firefly，将针对价格敏感的 Wildcat Lake 平台，导入中国本土化成熟的手机产业链，统一标准化接口实现规模化，完成降本增效。现场展示了 Project Firefly 的 A+ 参考设计与第一款 Project Firefly 参考设计下落地的量产产品来酷 Air 14，均为超薄机身与 Clean D 设计。来源Linus Torvalds 在 LKML 的新帖文中建议大家不要用 AI 工具提交重复的 bug 报告，大量重复、过时的报告让邮件列表无法正常收集 bug，想用 AI 工具创造点价值就去好好写补丁。来源Apple 定于北京时间 6 月 9 日至 13 日举行全球开发者大会 WWDC26，6 月 9 日凌晨 1 点开幕时将举行主题演讲，介绍 Apple 平台更新，包括人工智能的进展、新软件和开发者工具；6 月 10 日凌晨 4 点举办 Platforms State of the Union，介绍新功能、API 和技术如何进一步赋能开发者。来源如此直白的话感觉是Linus会说出来的😅Linus都蚌埠住了还人工智能进展，国行的 AI 有没有？意思是Googlebook运行的是x86_64的安卓？发表评论少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购你可能错过的文章四天三夜，用小长假的时间体验青海湖小环线打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试派评 | 近期值得关注的 App> 下载少数派 2.0 客户端、关注少数派公众号，解锁全新阅读体验 📰> 实用、好用的正版软件，少数派为你呈现 🚀31扫码分享目录 0讨论我来说一句发布发表评论发布

YoooClaw C·ONE 体验：为 AI Agent 而生的硬件，怎样构筑了数字化的「我」？主作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛联合作者关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛宛潼关注宛潼少数派作者每天都要睡十个小时以上的大湾区著名gai溜子

公众号：安娜的五百英镑离岛11:04AI 虽然现在具备了干活的能力，但对于很多人来说，它依旧只是个 Chatbot。因为，大多数 AI 依然停留在屏幕里（当然，具身智能另说）：你需要打开手机或电脑，进入某个 App，手动复制消息、描述背景、然后不停修正抽卡，最终把输出的结果复制到你需要的地方去 —— 对很多人来说，这个过程本身就已经足够打断手里的工作。App 的逻辑其实更适合屏幕类设备的交互，这是我们从 Web 2 时代过来的逻辑惯性。但如果我们换个思路，要是一个设备不是为了适应人类和 App 的交互，而是一开始就为 AI Agent 去构建，那它应该是怎样的？YoooClaw 给出的答案是 —— 一张智能卡片。很高兴，这次与少数派和 YoooClaw 一拍即合，提前体验了这款名叫 C·ONE 的智能卡片。 在体验的过程中我意识到，或许那个「数字化分身」的时代，已经慢慢开始变得具象了。AI 时代需要一个「灯条」拿到 YoooClaw C·ONE 的时候，其实我一直在想，手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件呢？但这个问题还不急着说，反倒是上手的第一刻，它的工业设计给了我不小的惊喜。它采用了全金属材质，一体感极强，拿在手里有一种精致感，有点像金属版的 MagSafe 外接电池。而且没想到的是，它的楔形设计与我 iPhone Air 的契合度高得惊人。磁吸上去后，它恰到好处地填充了 iPhone Air 极其轻薄的机身与我掌心之间的空白，给我这个 iPhone Air 裸奔党增加了不少安全感……C·ONE 的背面元素也不多：一个按钮、一个条形指示灯和 YoooClaw 的 logo。长按按钮，灯条变成绿色的波浪，C·ONE 就会进入 AI 对话模式，很像是我拿起 Apple TV 遥控器按下侧键与 Siri 对话。而双击之后，灯条变成绿色的闪烁，就进入了录音模式。这个灯条让我想起了前几天发布的 Googlebook，不同时空，横跨太平洋的它俩居然想到一块去了。该说不说，灯条在 AI 时代居然成了一种「不打扰却又十足惹眼」的交互 —— 它不告诉我具体的内容，而是只告诉我「状态」，简直是我这种 ADHD 人的福音 ——一个真正好用的通知管理器作为一个长期与 ADHD「斗争」的人，一条通知突然让屏幕亮起，我的心思可能就飞出了十万八千里之外。而且有时候好不容易进入了心流状态，突然来一个通知，看吧，我的工作流就断了，不看吧，真是心痒痒。iOS 原生的通知系统虽然有做一些约束，甚至还引入了专注模式，但它有一个很大的问题：它的管理逻辑是基于 App 的。这就有个十分明显的问题：工作群里，我负责的事情结束了之后，与我有关的通知可能就不多了。但我还是不敢关通知，毕竟有些事情还是需要我来处理。而在 iOS 系统中，这些通知的地位是完全对等的（到底是谁那么喜欢在微信聊工作啊！）。它们密密麻麻地挤在一起，要么让我因为厌烦而全部忽略导致错过，要么每次来消息，我就还是得就打断当前的事情看一眼。尽管 Apple 智能后来推出的「通知总结」有所改善，但对于国行设备来说，吃不到嘴里的饼画得再大也没用。即便在已经用上了的地区，用户反馈也不太好。而且，它依然解决不了「拿起手机」这个动作本身带来的诱惑：原本只是想看一眼通知，结果却顺手刷了半小时马斯克的表情包。YoooClaw C·ONE 提供的方案非常符合直觉：它可以读取我的通知，并根据重要程度，给背后的条形灯带设定不同的颜色和效果。最棒的一点是，这种设定不需要像在 iOS 里那样一个一个 App 去操作和设定，只需要在 YoooClaw 里通过自然语言告诉它：「我在微信群工作群 XXX 里的昵称叫宛潼，如果有 @ 我的消息，你就用红色呼吸灯提醒我。」于是，我的工作流变成了这样：当我开始写稿时，在 iPhone 上进入创作专注模式，所有通知调整为静音，并且同步到 Mac。然后，我将 iPhone 倒扣在桌面上，让 C·ONE 成为我唯一的通知接收器。如果灯条没有闪烁红色，那我就完全不用去理它，只有当红色的光亮起时，我才需要拿起手机瞟一眼。这种物理上的「隔离」，很好地保护了我的工作惯性。除此之外，它还有一个类似于「监工」的呼吸灯提醒模式，毕竟我这种人不看到 Deadline 是真的没法开工。现在，我可以设定在某个待办任务截止前的四个小时，让 C·ONE 以橙色的呼吸灯持续闪烁，直到我将任务标记为完成。这种视觉上的提醒虽然会让我感到焦虑，但相比于手机弹窗那种一划就过的提醒，这种持续的物理存在显然更有力量，确实可以有效地督促我赶紧干活，像极了以前十点不到就催我去睡觉的老妈。iPhone 用户的「终极闪念胶囊」随着豆包手机的面世，我确实也意识到，AI 最终会「杀死」绝大多数 App。但一个物理按键，或许在很长一段时间内都是刚需。物理按键代表着一种明确的「告知」。它告诉 AI：我现在要开始与你交互了。无论是 iPhone 上的操作按钮，还是新款机型上的相机控制，其实都是 Apple 为 AI 时代准备的入口。但在 Apple 智能迟迟未能在国内完全落地的当下，这些硬件潜能大半处于荒废状态。所以，C·ONE 上的那个物理按钮，现在就成了我外置的「闪念胶囊」。作为一名自由撰稿人，灵感往往是稍纵即逝且细碎的。很多时候，我脑子里蹦出一个绝妙的比喻，或者一段关于产品逻辑的思考，等我回家坐在电脑前准备动笔时，那种倾泻而出的感觉已经消失了。现在，我只需要长按 C·ONE 上的按钮，它就开始录音。我可以毫无章法地把这些碎片化的想法丢给它，让它帮我记住。当我回到家，这些碎片已经变成了整理好的文字，启动的困难就少了许多。C·ONE 在这里扮演的角色，不再是一个简单的录音机，而是一个「外置记忆存储器」。例如，在写完一篇关于 AIGC 内容的稿子，给自己做咖啡的时候，我偶然想起来里面那些近乎完美的人物图片，于是就对 C·ONE 说出了我心底被唤起的容貌焦虑；而在看完《给阿嫲的情书》之后，我也可以立刻在回家的散步的路上就对 C·ONE 说最直观的观后感，为日后可能撰写的影评积攒想法。上下文：构筑「我」的养料到了这里，就是时候来回答最开始的那个问题了：手机和电脑已经是很好的 AI 载体了，为什么需要再做一个硬件？在写这篇稿子的时候，我听了 YoooClaw 创始人肖睿哲和王自如的一场播客。老肖分享了一个很有意思的案例：我们有一个同事，他平时淘宝有大量的消息。他会在淘宝上买药，然后我们的产品就会识别到一个有趣的现象：比如他今天要来深圳出差，然后 AI 就会告诉他说你别忘了带你的降压药。这个降压药在配送过程中，淘宝有几个 Push 会带着商品标签和名字，而这些商品标签的名字被 AI 识别到了以后，他就会去自己学习这是个什么东西，然后形成了你某一类的画像。这个例子给了我很大的启发。过去我总觉得，大部分通知是骚扰，所以很多通知我是手动关闭的。但在 AI 看来，通知是极佳的上下文，经过用户筛选的通知更是优质的上下文。而且与人类会被通知打断不同的是，它可以不知疲倦的去读取。这些碎片化的信息，以及我每天录入的创作想法、我与它的对话，最终将会凝练成属于我的「宛潼.skill」。其实我一点都不反对蒸馏我自己，我也不担心 AI 会抢走我的饭碗。我甚至还在朋友圈发过一段话：「AI 要是什么时候真能帮我写稿那我可开心死了。我都不担心我的工作被抢，相反我可以用 AI 一天写十篇。」因为，AI 模仿的是我的「Skill」：我的语气、我的写作手法、我的写作风格……这需要大量的上下文去喂养；而真正的「Idea」源自于我特殊的经历、我今天与朋友的对话、我傍晚在河边晒着太阳发呆突然闪过的瞬间……这些是 AI 不可模仿的。如果有一天，我的数字分身足够成熟，它能完美捕捉我的表达方式和逻辑链条，我只需要负责抛出 idea，由它完成初稿，我再进行审稿和修改，那我的生产力将获得质的飞跃。这就是老肖在播客里说的，设备之所以有价值，是因为它沉淀了数据。就像在白板前讨论，结果固然重要，但「谁画了那个圈」、「大家是怎么讨论出这个结果的」这些过程数据，才是 AI 进化所需的真正养料。真的聊得蛮好的，推荐以及，我还有一个角度 —— 虽然很多人对我的印象是科技博主 or something，但我其实是个中文系毕业的纯文科生。我是凭着对数码产品的热爱，才自学了很多代码逻辑和工程师思维。但对于绝大多数普通人来说，AI 工具的门槛还是太高了。无论是 OpenClaw、Hermes 还是后起之秀 Codex，对于没有计算机思维和 AI 基础的大众来说，配置起来依然让人头大，就更别提使用了。当年 Apple 的 Lisa 带着图形界面问世，才真正砸掉了计算机与普通家庭之间的那堵高墙。所以在我看来，AI 也在等待属于它的「Lisa 时刻」。YoooClaw C·ONE 也许算不上一个颠覆世界的发明，但它做对了一件事：它把极其复杂的 AI 交互过程，简化成了一个「磁吸、配对、按键」的物理过程，让 AI 从纯粹的聊天机器人界面里跳出来。当我把它吸在手机背面，跟着斜挎挂绳一起带出门时，它就像是一个陪着我感知世界的、数字化的我这种与 AI 共生的感觉，虽然看起来挺赛博朋克的，也会让我产生异化的担忧，但 —— 与其在担忧中止步不前，我更愿意去拥抱变化，并在其中找到自己未来新的创作方向。毕竟，这种效率提升使创作变得淋漓尽致的感觉，是真的让我难以拒绝的。1012扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@PlatyHsu10等 10 人为本文章充电扫码分享举报本文章宛潼每天都要睡十个小时以上的大湾区著名gai溜子

线下活动｜听完这首，不再 Emo：Leon x 少数派萨克斯音乐奏享会主作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠联合作者关注听歌的水獭少数派广州体验店店宠听歌的水獭关注听歌的水獭少数派广州体验店店宠15:00当醇厚旋律邂逅专业音频设备的细腻还原，这是一场属于萨克斯音乐爱好者的专属活动。少数派广州线下体验店联动音频品牌奥世声 Austrian Audio，邀请到知名萨克斯青年演奏家 Leon Music，为大家带来一场适合音乐爱好者、萨克斯学习者乃至自媒体创作者的线下活动，一场兼具听觉享受与专业录音干货分享的音乐之约。Leon（陈郎）是近年备受关注的流行萨克斯演奏家，同时也是全网粉丝超 60 万的知名音乐自媒体博主，其作品总播放量突破亿次，曾获得周杰伦、陶喆等知名音乐人的转发认可，在抖音、小红书、B 站、YouTube等多平台拥有大量忠实粉丝。作为专业的萨克斯演奏者，Leon 不仅拥有扎实的演奏功底，更在音乐自媒体运营、录音混音领域积累了丰富经验，擅长将专业音乐知识与大众传播相结合，既能用精湛的演奏打动听众，也能以通俗易懂的方式分享音乐技巧与运营干货。五月的最后一个周末，少数派广州线下体验店很荣幸邀请到了 Leon，连同奥世声 Austrian Audio 的专业设备，给各位乐迷准备了一场干货满满、视听俱佳的演奏及分享会。奥世声 Austrian Audio 创立于 2017 年，主要产品为话筒和耳机，可以完美驾驭专业与日常两大使用场景，从录音棚、后期制作再到室内外都可以随时聆赏 Hi-Res 级流媒体音乐，Making Passion Heard 是团队研发产品的重要动力，无论你是歌手、制作人或是音乐爱好者，只要你的生活离不开音乐，Austrian Audio 就是你的不二之选。活动信息活动时间：2026.5.30（周六）20:00-21:30（总时长1.5小时）活动地点：少数派广州线下体验店（广州市荔湾区泮塘五约外街76号，1200 书店旁）参加方式：扫码报名，或线上观看直播人数限定 60 人，海报扫码报名空间有限，室内座位先到先得，室外也有户外露营椅以及音箱同步外放，荔湾湖公园的夜晚很舒服现场配备专业的收音、拍摄设备，届时少数派官方视频号及 Leon 个人账号将全程开启直播，无法到场的朋友可海报扫码预约直播，线上围观活动福利：活动期间现场指定饮品享 8.8 折优惠；互动环节有奥世声 Austrian Audio 周边礼品赠送活动流程包括Leon 的介绍及暖场演奏访谈 Leon 分享音乐与经营自媒体账号经验，穿插现场演奏观众互动问答，邀请现场观众即兴演奏全体大合照，活动结束，随意交流少数派广州线下体验店介绍少数派广州线下体验店于 2025 年初正式开业，坐落于荔湾湖畔泮塘五约古村，延续其「高效工作、品质生活」的核心理念，打造集数码、音乐、社区于一体的复合型线下据点，并拥有一家精酿酒吧。店里不定期举办音乐沙龙、唱片展等活动，延续着少数派对创作者的长期关注。作为线上社区的线下延伸，这里既是数码爱好者交流体验、打卡少数派文化的实体据点，也是老广文艺青年逛湖漫步、放空小聚的隐秘角落。温馨提示暑期公园户外蚊虫较多，门店提供有限驱蚊用品，乐迷也可自行准备现场互动环节均有 Austrian Audio 周边礼品赠送，积极参与即可获得惊喜无法到场的观众可关注少数派官方视频号或 Leon 个人账号，观看同步直播，线上也可参与问答互动欢迎大家在活动结束后留下来继续交流，活动期间指定饮品 8.8 折5 月 30 日，广州荔湾湖公园少数派线下体验店，我们现场见～文章题图来自Denny Müller10扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@挨石1等 1 人为本文章充电扫码分享举报本文章听歌的水獭还没有介绍自己关注

36氪获悉，中芯国际发布投资者关系活动记录表公告，基于客户需求和在手订单情况，相较于上个季度，公司对于今年的整体运营情况更加乐观。主要是因为，第一：人工智能对配套芯片的强劲需求，直接推动公司电源管理芯片产能供不应求；第二：人工智能海外虹吸效应使得消费、IOT等客户在大陆寻找产能、订单回流；第三：人工智能亦带动ToF、电动汽车、机器人等新应用需求，本土公司积极开拓市场；第四：产业链国产化诉求，推动国产逻辑、网通等芯片需求；第五：涨价效应以及之前提到的客户因担忧以后的供应不足而提前备货。在市场环境变化的情况下，公司的技术储备、平台多元和产能转换灵活性的优势，支撑了公司接收订单的能力。

全球首个全民免费用ChatGPT Plus的国家，OpenAI官宣了

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生

36氪获悉，AMD AI开发者日上，AMD董事长兼CEO苏姿丰（Lisa Su）发表了主旨演讲称，目前全球AI活跃用户已超过10亿，至2030年，全球AI每日活跃用户将超过50亿，未来每个人都可以同时拥有5个、10个甚至100个AI智能体来协助工作。她表示，要支撑智能体的运行，光有基础大模型远远不够，行业的技术重心正向推理能力、持续学习能力和高效数据流能力加速转移。

36氪获悉，AMD AI开发者日上，AMD董事长兼CEO苏姿丰（Lisa Su）发表了主旨演讲称，目前全球AI活跃用户已超过10亿，至2030年，全球AI每日活跃用户将超过50亿，未来每个人都可以同时拥有5个、10个甚至100个AI智能体来协助工作。她表示，要支撑智能体的运行，光有基础大模型远远不够，行业的技术重心正向推理能力、持续学习能力和高效数据流能力加速转移。

全球首个全民免费用ChatGPT Plus的国家，OpenAI官宣了

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生

记者今天（19日）从国家开发银行了解到，今年以来，国开行围绕支持传承弘扬中华优秀传统文化、系统性保护文化遗产，探索通过“产业导入、活化遗产”“街区统筹、空间织补”等模式，积极支持历史文化名城、街区、村镇有效保护和活态传承，以及不可移动文物、历史建筑的修缮和合理利用等，有力推动一批具有示范效应的项目实施。截至4月末，已累计向历史文化保护传承领域发放中长期贷款超过240亿元。（央视新闻）

全球首个全民免费用ChatGPT Plus的国家，OpenAI官宣了

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生

对话万成云商：发文章≠GEO优化，大模型不是喂什么就推什么杨越欣（杨桃）·2026年05月18日 13:05对话万成云商：拆解出海企业AI时代GEO布局路径“帮我搭配一套适合海边度假的衣服。”“帮我推荐一家适合约会的西餐厅。”“3000元预算，买哪款扫地机器人最划算？”这类与AI的问答正发生在全球数亿用户身上。过去，人们会把这些问题输入百度、谷歌等搜索引擎，因此产生了SEO（Search EngineOptimization，搜索引擎优化）营销。而当流量入口从搜索引擎转向AI工具，GEO（Generative EngineOptimization，生成引擎优化）也应运而生。区别于传统SEO追求的“搜索结果排名”，GEO的核心目标是让品牌信息在AI生成的回答中被优先引用和推荐。在AI全面渗透用户信息获取与消费决策场景的时代，对于出海企业而言，做好GEO意味着抢占新的流量入口。GEO的逻辑和SEO有什么根本不同？为什么有人说GEO是伪命题？出海企业布局GEO该怎么做？中小企业如何低成本起步……带着这些出海企业关心的问题，我们特别对话了万成云商品牌出海事业部总经理谭莉，拆解AI搜索时代如何让企业被AI看见、信任，并推荐给用户。36氪：GEO被称为SEO在AI时代的进化。相比于SEO的核心在于搜索结果的排名，GEO优化的关键是什么？它与传统SEO的核心区别在哪里？谭莉：核心区别在于传统SEO偏“规则博弈”，GEO更像“信任竞争”。因为过去做SEO，我们的目标是把页面排到Google第一页，让用户点进去；但现在很多用户直接在AI里问问题，甚至不点链接了，那你要解决的是——怎么让AI在回答里“引用你”甚至“推荐你”。所以GEO的关键首先是，你是不是一个“可信任源”（全网有没有人认可你）；其次，你的内容能不能被AI理解、拆解、复用；最后是你的内容有没有完整的证据链，比如数据、案例或者第三方背书。36氪：继GEO之后，也有人提出AEO概念，两者有什么区别？谭莉：这两个概念很多人会混着讲，但我个人是这么区分的：AEO更偏“问答结果优化”，核心是让你的内容直接成为某个问题的标准答案，比如FAQ、精选摘要这些。GEO更广，它面对的是生成式AI，不只是回答一个问题，而是参与“多源信息整合后的输出”。所以AEO更偏“结构技巧”，GEO更偏“整体品牌与内容生态”。36氪：GEO的概念近两年引起广泛讨论。为什么现在这个时间点企业需要关注GEO？在AI搜索渗透率持续提升的背景下，GEO应该在企业营销体系中占据怎样的战略位置？谭莉：分享几个很现实的变化：现在用户越来越习惯直接问AI，而不是自己筛选搜索结果；而且AI在很多行业决策前端，已经开始替代“信息收集阶段”；再就是有些客户还没打开你的官网之前，可能就在AI里形成了对你的判断。所以企业没办法去忽视GEO，现实摆在这里了。我认为，GEO在企业营销体系中占据着很关键的位置，它不会替代SEO，而是和SEO并行的“第二流量入口”，甚至未来会变成主入口之一。36氪：怎么理解有企业做过一些尝试后认为GEO是伪命题，或者靠自己持续给大模型喂语料也可以？谭莉：我觉得这是对AI工作机制有点误解，大模型不是你“喂什么它就推荐什么”。它有几个筛选逻辑：第一，信息是否一致（比如不同来源是否能得到印证）；第二，来源是否可信（是不是长期存在的内容资产）；第三，是否被多平台提及（而不是单一来源）。所以你自己“喂语料”，本质只是单点信息输入，很难形成“信任网络”。不过我也理解为什么有人说GEO是伪命题，因为确实有很多“伪做法”，比如只做站内内容、只堆AI文章、不做外部验证，这些基本都很难起效果。36氪：当企业布局海外GEO营销时，需要经历哪些关键步骤？不同行业、不同规模的企业在路径上会有哪些差异？谭莉：从实操角度，会有四个关键步骤：第一步先拆解你的客户是怎么筛选供应商的，这个路径是怎么样？第二步重建你的内容结构，包括官网、博客、FAQ、案例这些，全部按“可被AI理解”的方式重构。第三步是搭建外部信任体系，比如评价平台、媒体曝光、社媒讨论、行业引用，都很重要。第四步就是做多渠道分发，持续让内容在不同平台被看见和引用。不同企业可能在行业复杂度和品牌基础这两方面存在明显差异。比方说医疗、工业类比其它行业更需要专业背书，那有品牌的企业就更容易放大，没有品牌的就要先“补信任”。36氪：在实施过程中，企业应如何在内容策略和技术基建层面做出改造甚至重塑，以同时满足AI理解与用户共鸣？谭莉：内容策略方面要做到结构清晰，信息有证据，有数据来源，表达要接近真人沟通。技术层面反而是基础的更容易被忽略：比如Schema结构化数据（Product、FAQ、Article等）、robots、index、hreflang这些基础配置，再一个就是页面可读性。总结来说，是做“可被理解的内容”。36氪：在布局GEO出海方案时，企业最容易忽视但至关重要的环节是什么？对于缺乏专业团队的中小企业，有哪些可行的解决方案或合作模式？谭莉：我们服务这么多跨境企业，发现最容易忽视的一点其实是：“让AI能访问你”。听起来很基础，但很多企业在实操中会陷入误区：比方说robots把AI爬虫挡了、页面noindex设置错误，或者多语言混乱……这些都会直接导致你做再多内容，AI也看不到。第二个容易忽视的是站外信任建设。很多人还停留在“发文章=做优化”，但GEO更看重：评论、测评、社媒讨论、媒体引用。对于中小企业，我建议两种路径：①轻量级：聚焦一个细分领域做深内容+少量高质量外部背书；②合作型：找有资源的服务商做整合（内容+渠道+数据），不要自己盲目试错，成本会很高。36氪：企业在选择GEO服务伙伴时，应该重点考察对方的哪些核心能力（如是否具备垂直行业知识库、跨平台渠道资源、效果监测与分析体系等），以避开“无效投入”的陷阱？谭莉：这个问题非常关键，因为现在确实有不少“概念型服务商”。我建议重点看四个能力：第一，行业理解能力，是不是懂你这个行业，而不是只会写内容。第二，内容生产质量。有没有真实案例、是否具备本地化能力，而不是纯AI批量生成。第三，渠道与分发能力。有没有海外平台资源，比如媒体、社区、测评渠道。第四，数据监测体系。比如做一段时间了，看看有没有被AI引用？哪些内容有效？ROI如何？如果对方只能讲“发多少篇内容”，基本可以直接排除。36氪：2026年初Seedance2.0的爆火，让业界看到了AI视频生成从“抽卡式试错”向“工业化生产”跨越的可能性 。10秒带货视频成本仅需8元的“成本塌方”对于跨境电商的内容生产逻辑意味着什么？谭莉：这个变化其实挺颠覆的。过去做视频，成本高、周期长，所以大家拼的是“有没有内容”；但现在成本接近于零，意味着：内容不再是门槛，内容“是否有效”才是门槛。36氪：当视频生产成本趋近于零，所有人都能批量生成高质量内容时，企业的竞争焦点会发生怎么样的变化？从GEO布局的角度看，企业应如何构建AI时代下的营销思维？谭莉：供给无限的时候，稀缺的就变了。我认为未来竞争会集中在三点：第一，认知占位，用户心里有没有你；第二，信任结构，有没有被多方验证；第三，数据能力，能不能快速迭代内容。从GEO角度，我觉得企业要建立一种新思维：要把品牌内容当资产去做长期积累，同时将口碑、评价这些信任一点点做起来，更重要的是发内容之前，问问自己：AI会不会用它？用户会不会信它？避免做“自嗨”的内容。36氪：Seedance2.0的火爆也引发关于深度伪造和版权的讨论。跨境电商卖家在利用这类工具生成营销素材时，应如何规避潜在的肖像权、版权风险？谭莉：这个问题确实不能忽视，我建议，尽量避免使用可识别人物（尤其是公众人物）做生成素材，素材来源尽量用可商用授权库，品牌、Logo、音乐等要确认授权范围，涉及医疗、功效类内容要特别谨慎，因为合规风险更高。毕竟，AI可以帮你降低成本，但不能帮你“免责”。本文由「杨越欣（杨桃）」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

Agent、多模态、应用、算力一天看尽，峰会亮点在此｜5.20日，来现场一起AI未来一氪·2026年05月18日 15:03第四届中国AIGC产业峰会5月20日北京举办，干货满满开放报名。进入2026，AI愈发狂飙突进。围观体验之余，人人不免在心中自问：朋友圈刷屏的“龙虾”、Harness等AI新事物，跟我到底有什么关系？真的有必要跟吗？AI创业、AI融资如火如荼，属于我的机会又在哪里？别人已经在用AI做视频、写代码、跑项目，我是不是已经慢了一拍？……到最后，几乎所有问题都会汇成同一个问题：我，到底该如何用AI？如果你对这些问题还很模糊，不妨来第四届中国AIGC产业峰会走一趟——一天时间，把这一年AI产业最值得关注的人、事、判断，一次性讲清楚。先提前剧透一波。18位重磅嘉宾，1场Agent主题圆桌，1份年度榜单，1张全景图谱——所有内容都已敲定。阵容上，既有昆仑万维、智谱、商汤、百度、蚂蚁、MiniMax这样的行业头部，也有亚马逊云科技、硅谷Fusion Fund带来的全球视角；既有复旦邱锡鹏、港大黄超代表的学界与开源前沿，也有盛大EverMind、太初元碁、趣丸科技等活跃实战派玩家。议程上，从Agent商业化落地、多模态与空间智能的最新突破，到AI在文娱、医疗、企业服务等场景的纵深渗透，再到算力与AI Infra的范式之变……这一年最值得关注的AI话题，几乎都装进了这一天。话不多说，下面直接上重点。亮点1：Agent到底怎么落地？不同视角玩家都有心得如果说去年大家还对“Agent是不是未来”持观望态度，那么2026年，这个问题的答案已经不言而喻。“Agent不是未来，是现在”——这正是智谱高级副总裁吴玮杰带来的判断，他将分享智谱在Agent商业化落地上的最新实战。但Agent真正走入企业，并非一蹴而就。亚马逊云科技产品技术部技术总监王晓野将聚焦如何跨越Agent落地鸿沟，拆解从最强模型到企业级AI Agent之间那条最难走的路。而港大助理教授黄超则会从学术与开源前沿的视角，探讨AI Agent范式变革的挑战与机遇。他所带领的HKUDS GitHub开源平台累计获得超24万Star，位列全球Top50，话语权毋庸置疑。与此同时，Agent要真正成为“数字生产力系统”，还需要更底层的能力支撑。盛大集团副总裁、EverMind CEO邓亚峰将分享长期记忆驱动的自进化，探讨AI如何从一个“工具”演化成可持续成长的数字生产力系统。当然，光有演讲还不够。本届峰会专门设置了一场Agent主题圆桌：2026，Agent产品的不确定性与非共识机遇。出席嘉宾包括趣丸科技副总裁兼首席战略官、《屠龙之术》主理人庄明浩，ColaOS&MarsWave CEO冯雷，以及EvoMap创始人张昊阳。这三位都是Agent产品一线的深度玩家，从产品形态到操作系统，从多智能体框架到自进化引擎，他们会碰撞出什么样的非共识观点？非常值得期待。亮点2：Agent之外，模型能力边界也在继续扩张虽然Agent是今年峰会的绝对主角之一，但不可忽视的是，支撑Agent的模型能力本身，也在以惊人的速度向外延伸——从“数字”到“物理”，AI对世界的认知边界正在被重新定义。商汤科技执行副总裁、首席科学家林达华将分享多模态模型如何从“看懂”和“生成”，进化到真正可感知、可生成、可行动的空间智能新阶段。上午的GenAI Talk环节，蚂蚁灵波科技首席科学家沈宇军将围绕一个核心命题展开对谈：AI 2.0下半场，从AIGC到AIGA。当生成式AI从“内容”走向“行动”，整个产业逻辑将被改写。而京东JoyInside业务负责人戴文军则会把这个命题进一步推往物理世界，当AI从屏幕里走出来，走进真实终端、真实场景，“AI World”才真正开始。模型层面，复旦大学特聘教授、模思智能创始人邱锡鹏将分享MOSS多模态模型及其推理优化的最新进展。这位主持研发了MOSS系列开源大模型的“大模型一线人物”，将带来学术与工程结合的硬核内容。亮点3：具体怎么玩？这些应用玩家甩出成功案例模型也好、Agent也罢，具体到不同领域，机会到底在哪？昆仑万维董事长兼CEO方汉率先给出他的答案。他这些年带着昆仑万维在AI音乐、AI社交、AI短剧等多个AIGC方向同时下注，其分享会从一线视角，讲讲AIGC具体应用这一年究竟跑通了什么、又踩过哪些坑。紧接着，来自不同赛道的“实战派”玩家也将依次登场：文娱创作领域，风行在线CEO易正朝将带来一个核心判断——从AI编程到AI视频，众创才是AI生产力的核心杠杆。他会讲述AI如何从单点工具，进化为撬动整个创作者生态的杠杆，从而让普通人也能撑起一支创作团队。生产力工具层面，百度秒哒产品总经理朱广翔会分享他们“人人都是创造者”的产品理念，当门槛被AI抹平，想象力就是新的生产力。企业级AI方向，MiniMax ToB中国区商业化负责人胡维琦将带来MiniMax这一年在ToB战场上的真实打法与思考。大健康赛道，轻松健康集团执行副总裁马孝武将拆解从循证智能体到AI驱动的健康服务增长飞轮，展示AI在严肃医疗场景里的真实增长引擎。文娱、办公、企业、医疗……AI不再是“被讨论的对象”，而是“正在被使用的工具”。每一个细分赛道里，都已经有人开始把案例端上桌了。亮点4：从训练到推理，算力的故事也在被改写模型与应用之外，2026年最值得关注的另一条主线，是AI基础设施的范式重构。过去几年，所有人讨论算力都默认一个前提——为训练服务。但2026年，风向变了，推理需求暴涨、训练成本高企，整个算力叙事正在从“训练驱动”转向“推理驱动”。Fusion Fund创始合伙人张璐将从全球资本与产业的双重视角，解读算力叙事的重构。当AI基础设施从训练转向推理，一场范式变革正在发生。而国产算力这一边，太初元碁首席产品官、高级副总裁洪源将带来他对国产算力路径的思考。当“Token”正在成为AI时代的新计量单位，国产算力如何筑基Token智能新未来？他会给出自己的答案。亮点5：看尽AI全景生态，一份榜单+一份报告延续往届传统，本届峰会，技术先锋、产业龙头、新锐势力将再次全景式碰撞。为了给行业更多参考，我们还将颁出2026年值得关注的AIGC企业、2026年值得关注的AIGC产品两类奖项，以及一份中国AIGC应用全景图谱。前者是量子位根据过去一年AIGC企业、产品的表现与反馈，并结合对2026年技术与场景的观察，评选出的高潜力企业和产品。后者是量子位智库绘制的一幅全面立体的中国AIGC应用全景图谱，用以展示国内AIGC应用的市场格局与发展动态，并深入剖析最具潜力与竞争力的产品和赛道。5月20日举办最后，时间地点以及最全议程在此，现场还有参会伴手礼等你来领，咱们到时候见 ~时间：5月20日 9:30地点：北京·金茂万丽酒店现在，中国AIGC产业峰会正在火热报名中，欢迎点击链接报名线下参会：https://hdxu.cn/1Hjr1本次峰会也将同步线上直播，欢迎预约参与：量子位视频号。参会嘉宾简介一览方汉昆仑万维董事长兼CEO方汉毕业于中国科学技术大学近代物理系，曾任职于中国科学院高能物理研究所、拓林思、亚信以及千橡互动。于2008年协助周亚辉先生创立昆仑万维。方汉拥有30年互联网从业经验，从1994年开始参与和倡导开源运动，是中文Linux奠基人、中文Linux四剑客之一，著作国内第一本Linux书籍《Linux实用大全》，他也是国内最早的网络安全专家，曾负责三大电信运营商的全网安全审计，领导开发了国内第一款P2P下载软件DUDU加速器，负责研发了国内市场占有率最高的网页游戏《三国风云》，为中国互联网最资深的参与者。吴玮杰智谱高级副总裁吴玮杰目前担任智谱公司高级副总裁，负责公司MaaS平台商业化，并兼任泛互联网事业部总经理及浙江子公司总经理。他在人工智能、互联网和信息技术领域有着丰富的商业化管理经验和项目落地经验，曾任字节跳动企业应用全球首席商业官（旗下品牌包括飞书、LarkPeople和幕布等）、GE数字集团大中华区副总裁、复星集团联席首席发展官&首席投资官。除此以外，他也曾在思科、SAP、Oracle等公司从事商业化管理工作。吴玮杰博士毕业于电子科技大学，同时持有香港大学国际工商管理硕士学位以及复旦大学软件工程学士、硕士学位。易正朝风行在线CEO易正朝，风行在线CEO，作为业务领头人，引领公司从传统互联网视听平台全面转型升级为AI文娱生态平台。推动“AI赋能众创”战略落地，带领团队打造橙星梦工厂、小剪刀等AI创作工具，影视网文智能小程序平台、电视鸿蒙平台等核心智能产品，构建从内容创作、智能分发到商业变现的全流通链生态。推动公司荣获AI赛道奖项、抖音快手等平台的头部服务平台奖项，在AI文娱新时代，推动内容生产效率革新与创作者生态繁荣，致力于让内容流动更简单。林达华商汤科技执行副总裁、首席科学家林达华教授现任香港中文大学信息工程系副教授、交叉学科人工智能研究所所长，并为商汤科技联合创始人、执行董事兼首席科学家，兼任上海人工智能实验室领军科学家。他于2012年获麻省理工学院计算机科学博士学位，研究方向涵盖计算机视觉、深度学习、大模型与生成式人工智能，在人工智能顶级会议和期刊发表论文逾300篇，引用逾90,000次，h-index达131。林教授曾获2010年NeurIPS最佳学生论文奖，带领团队在多项国际计算机视觉竞赛中获奖，连续五年入选Elsevier全球前2%顶尖科学家。他主导发起OpenMMLab，并推动书生大模型、OpenCompass评测平台及DeepLink开放计算体系建设，其中DeepLink于2024年获世界人工智能大会“卓越人工智能引领奖”。2026年，他荣获“中银香港科技创新奖”，以表彰其在计算机视觉与多模态智能领域的系统性创新。他长期服务学术共同体，曾任IJCV编委，多次在人工智能主要国际会议担任领域主席，并担任IEEE大规模深度学习模型评测标准工作组主席。邓亚峰盛大集团副总裁、EverMind CEO邓亚峰，毕业于清华大学，正高级工程师，拥有二十余年人工智能算法及产品研发经验。现任盛大集团副总裁、EverMind CEO。曾任360集团副总裁、人工智能研究院院长兼搜索事业部总经理，科创板首家AI上市公司格灵深瞳CTO。曾荣获2021年中国人工智能年度十大风云人物，教育部技术发明奖二等奖。长期致力于大语言模型、计算机视觉及AI for Science等领域的研究与应用，带领团队多次在国际国内主流AI评测中获得佳绩；累计申请发明专利160余项（已授权98项），发表论文50余篇（含10篇Nature子刊）。王晓野亚马逊云科技产品技术部技术总监王晓野现任亚马逊云科技产品部技术总监。长期专注于企业现代化应用架构，数据及人工智能能力的建设。负责带领技术专家团队提供云上产品及开源方案技术实践指导，并致力于亚马逊云科技全栈服务在中国市场的应用和推广。沈宇军蚂蚁灵波科技首席科学家沈宇军，现任上海蚂蚁灵波科技有限公司首席科学家，同时领导蚂蚁技术研究院交互智能实验室。博士毕业于香港中文大学，研究方向为计算机视觉、生成模型、空间智能和具身智能。在CVPR、TPAMI、ICCV、ECCV、NeurIPS等国际顶级会议和期刊上发表论文百余篇，Google Scholar引用量逾万次。2024年起担任蚂蚁灵波科技首席科学家，致力于探索计算机视觉在机器人行业的落地之路，带领团队构建从空间感知到智能决策、从基座能力到世界建模的完整技术矩阵，推动具身智能从实验室走向规模化应用。邱锡鹏复旦大学特聘教授、上海创智学院院长助理、模思智能创始人邱锡鹏，CAAI Fellow，国家杰青获得者，主要研究方向为大模型、智能体以及科学工程等领域中的应用，发表学术论文被引用4万余次，5次获得AI领域高水平会议的最佳/杰出等论文奖，担任中国人工智能学会自然语言理解专委会副主任、中国中文信息学会大模型与生成专委会副主任、上海市计算机学会自然语言处理专委会主任等，主持研发了MOSS等系列高影响力开源大模型。连续多年入选中国高被引学者、全球前2%顶尖科

科氪 | 性能极限与跨界尝试：红魔11S Pro系列深度体验科氪·2026年05月18日 17:35红魔11S Pro电竞旗舰测评：纯平硬核，性能拉满在智能手机行业普遍面临镜头模组日益凸出、机身内部堆叠空间受限的背景下，电竞旗舰由于其对散热和功耗的极端要求，往往表现出更为激进的工业设计。作为新一代的性能导向型产品，红魔11S Pro系列通过软硬件的垂直堆叠，试图在移动终端上探索更进一步的性能上限。以下我们将从设计、理论性能释放、以及实际软硬件体验三个维度，对其进行深度拆解。一、 外观设计：纯平后盖的内部堆叠与电竞美学从视觉和握持逻辑来看，红魔11S Pro最显著的特征在于其坚守的纯平后盖设计。在行业因大底传感器而集体妥协于镜头凸出的趋势下，这种“反向”设计在工业制造上面临着极高的空间挑战。产品图片内部空间重构：为了在不增加机身异常厚度的前提下抹平镜头凸起，红魔对内部架构进行了重新设计。在有限的Z轴高度内，开发团队不仅压榨空间塞入了新一代的脉动水冷模块、无线充电线圈，还容纳了大面积的VC散热板和高密度电池。产品图片透明电竞美学的视效迭代：机身背部延续了红魔标志性的透明设计语言。通过透明材质，内部的脉动水冷引擎实现了可视化。在水冷系统运行时，用户可以观测到液体流动的动态秩序。相比于纯粹的RGB灯光，这种工业风的设计提供了一种更具技术感的视觉回馈。材质与细节处理：后壳采用了微蚀刻工艺打造的浮雕星轨纹理，搭配环形拉丝处理，使得整体质感较为收敛和冷峻。机身在肩键、腰部、LOGO及风扇处均设有支持自定义的RGB灯效，在暗光游戏环境下能提供必要的电竞氛围。产品图片二、 性能与散热：高标准筛选的“特种兵”芯与风水双冷架构性能表现是电竞手机立足的核心。红魔11S Pro首发搭载了第五代骁龙8至尊领先版移动平台。为了在重负载下维持超频表现，红魔在芯片筛选和散热架构上采用了更严苛的标准。产品图片芯片体质的物理筛选：为了满足电竞长时高频的需求，红魔所采用的这颗“领先版”芯片，主要筛选自晶圆中心区域。由于物理刻蚀规律，该区域的芯片通常具备更稳定的温度控制和更高的电气性能。在此基础上，通过最高4.74GHz主频的冲刺分级测试，以及高温高压下的自适应电压调节（AVS）检测，筛选出具备自适应能力的芯片，以确保在极端重负载下不易因发热而触发激进的降频保护。风水双冷复合散热系统：针对第五代骁龙8至尊领先版的发热特性，红魔在散热上引入了原本常见于服务器端的复合方案。手机将【驭风4.0】贯穿式主动散热风扇与脉动水冷引擎相结合，辅以复合液态金属3.0和4D冰阶VC。风扇与整机均通过了IPX8级防水认证，在提升主动散热效率的同时，规避了由于进水导致元器件损坏的风险。在散热系统运转时，后盖的蓝色轨道圈内，蓝白相间的液冷介质会随之一起转动。仪式感拉满！产品图片独立显存与能效调度：在架构协作上，红魔自研的红芯R4电竞芯片与高通联合打造了独立的高速显存。在图形高渲染场景下，此设计可减少高频访问系统内存的次数，从而降低总功耗。结合CUBE擎天游戏引擎3.0的AI感知调校，系统能够精细化调配算力，平衡了高帧率表现与续航消耗。三、 使用体验：悟空屏2.0、Steam直连与续航评估在日常交互和实际游戏运行中，硬件的堆叠最终转化为了多场景的体验。悟空屏2.0的视觉完整性：正面配备的悟空屏2.0基于京东方X10全新屏下发光材料，实现了全无挖孔的视觉体验。屏幕黑边控制在1.25mm，屏占比达到95.3%，支持144Hz刷新率。由于没有刘海或挖孔遮挡，在全局视野和视频播放时拥有更纯粹的画面呈现。同时，屏幕集成了全屏3D超声波指纹解锁，提升了暗光或湿手状态下的解锁盲操率。产品图片行业首发Steam直连：性能的跨界探索：这是该机型在软件生态上的重要尝试。依托转译引擎，红魔攻克了X86与ARM架构之间的转译难题。用户拨动竞技键进入游戏空间后，可直接扫码登录个人的Steam账号。系统不仅支持游戏库内兼容作品的下载游玩，还打通了云端存档。用户在PC端的进度可以在手机侧的PC模拟器上无缝同步，迈出了打破移动端与桌面端算力壁垒的一步。产品图片产品图片产品图片产品图片操控精准度与手汗免疫：针对瞬时爆发的操控需求，该机引入了全新的新思触控芯片，在MagicTouch4.0算法支持下，实现了3000Hz的瞬时触控采样率。该方案优化了湿手控制算法，一定程度上避免了玩家因手汗导致的断触或漂移现象。续航与快充系统：为了支撑高功耗输出，红魔11S Pro在续航规格上给出了较高配置。PRO版本搭载了8000mAh的大容量电池，显著延长了重度游戏下的续航时长。而PRO+版本则在7500mAh电池的基础上，配备了120W有线快充与红魔首款80W无线闪充，在充电灵活性和回血速度上取得了较好的平衡。产品图片产品图片日常AI功能落地：在非游戏场景下，红魔11S Pro作为主力机也搭载了丰富的日常工具。AI帮写、AI消除、AI超清文档等功能提升了商用和办公效率；影像系统则在保留Live Photo的同时，延续了红魔标志性的星空与车轨慢门拍摄功能，拓宽了产品的实用边界。产品图片结语红魔11S Pro系列是一部特色鲜明、目的纯粹的硬件设备。它通过风水双冷散热、高标准芯片筛选以及无孔全面屏，将移动端的游戏性能推向了新的极端，并通过Steam直连展现出了打破平台壁垒的野心。对于不愿向主流“大镜头凸起”妥协、且对高负载游戏及跨平台游玩有硬性需求的极客与核心玩家而言，这是一台工业特色鲜明、技术指标扎实的旗舰标杆。产品图片本文由「科氪」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

番茄小说正推动 AI 动漫上院线，开放头部IP改编权限｜36氪独家兰芥·2026年05月18日 17:46字节跳动为番茄IP定下了影响力目标。作者｜兰杰 李晓霞编辑｜乔芊36氪独家获悉，据多位接近字节跳动人士消息，字节旗下的番茄小说正在推动AI动漫上院线。在这一过程中，会开放番茄小说平台头部IP的改编权限。之于番茄小说，如何更好地挖掘平台上好IP的潜力和商业价值，是一个亟待解决的难题。一方面，以文字形式为载体的IP内容，缺乏统一的视觉形象，使得平台上的IP难以有联名、衍生品等更丰富的变现形式。另一方面，AI漫剧这一新事物迅速发展壮大，其中不乏改编自番茄小说IP的粗制滥造之作。这些内容不但没能延长好IP的生命力，反而形成了一种消耗。此时，通过可登陆院线的影视内容来探索平台IP的更多可能性，不失为一种好的尝试。另据上述知情人士表示，字节跳动内部为番茄IP定下了影响力的目标，“想让行业内外的人觉得番茄的出品是精品的、正向的。”这一需求落到具体的业务上，上院线会是个不错的答卷。早在2023年年底，番茄小说就官宣成立了“番茄影视”和“番茄动漫”两大厂牌，推动平台IP的影视化。过去两年间，番茄影视先后与腾讯视频、爱奇艺和优酷等长视频平台合作，推动了《十日终焉》《我在精神病院学斩神》等平台头部IP的改编，但这些合作多停留在IP授权阶段。虽然番茄系（字节旗下以番茄小说、红果短剧等为主的业务）并不愁关于钱的问题，但显然成本高、回报周期长的长视频生意，并不符合依赖大手笔投流、需要市场快速反馈的“字节逻辑”。直到AI的出现，大幅降低了成本、提高了效率。36氪还从另一位知情人士处获悉，红果漫剧侧如今正在采购120分钟不分集的长内容，这也是当下各大视频平台争相竞逐的方向。之所以上线AI动漫内容，而不是AI仿真人内容，或许与当前AI大模型的能力有关。行业头部的AIGC创作者向36氪表示，如今视频生成大模型的能力已经可以做到生产120分钟的长内容，但是在仿真人这一赛道上，人物、场景的一致性依旧是个难题，而动漫内容可以通过部分突出的人物形象设计——例如彩色的头发、雀斑等，一定程度上弥补一致性的不足。回顾番茄系并不算漫长的发展历程，它可以说是字节跳动在文娱领域的一次突围。在番茄小说问世之前，字节始终未能建立起足够的内容能力和IP积累。而如今，从小说到短剧，再到漫剧，番茄系已具备了IP开发的全链条能力。与此同时，番茄系显然不满足于仅仅停留在当下的免费市场，而是在探索更多元的商业模式，包括线上订阅和院线票房等更多商业可能性。本文由「兰芥」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：IC photo

氪星晚报 ｜百度：一季度营收321亿元，AI业务收入136亿元；马斯克预计今年美国将广泛使用自动驾驶汽车；巨力索具旗下十余家企业已注销卜算籽·2026年05月18日 19:11特朗普：本应要求获得更多英特尔公司的股份大公司：索辰科技：物理AI业务目前尚处于布局初期36氪获悉，索辰科技公告，公司股票于2026年5月14日至5月18日连续三个交易日内收盘价格涨幅偏离值累计达到30%，属于股票交易异常波动情形。公司近期实施的募投项目“物理合成数据核心技术基座构建与多领域场景示范应用项目”尚处于筹备阶段，若出现募集资金投资项目未能顺利实施、新技术开发进度不达预期、研发遭遇技术瓶颈甚至失败，将会对公司经营业绩造成一定影响。公司物理AI业务目前尚处于布局初期。物理AI技术研发投入大、技术迭代快，存在研发进度不及预期、核心技术迭代风险；同时下游行业应用场景拓展、商业化变现存在不确定性，若市场需求、行业政策、技术竞争格局发生不利变化，物理AI业务收入及盈利贡献存在不及预期风险。智元WITA成为全国首例完成大模型备案的具身智能交互模型36氪获悉，根据上海市网信办最新公告的上海市生成式人工智能服务备案情况，智元WITA成为全国第一款实现合规备案的具身智能交互大模型，标志着中国具身智能产业从“技术验证”进入“合规商用”的新阶段。依托智元“三智一体”全栈技术体系，智元WITA核心应用于人形机器人交互场景。上交所：上周新增8家IPO受理企业36氪获悉，据“上交所企业上市服务”消息，本期（5月11日—5月17日）新增8家受理企业，其中上交所主板1家，科创板3家，创业板3家，北交所1家。截至5月17日，今年新增31家受理企业。截至5月17日，各板块在审企业共295家。其中上交所62家（主板18家，科创板44家），深交所61家（主板17家，创业板44家），北交所172家。截至5月17日，2026年已公布终止审查（撤材料+否决/终止注册）企业20家，其中上交所7家（主板3家，科创板4家），深交所创业板1家，北交所12家。腾讯音乐：收购喜马拉雅的事项已经完成36氪获悉，腾讯音乐在港交所公告，于2026年5月18日，根据并购协议的条款，收购喜马拉雅的事项已经完成，据此，喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为12.6亿美元的现金及最多总数1.75亿股本公司A类普通股（包括本公司于交割时新发行的A类普通股，以及本公司授出的以股权为基础的激励相关的A类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。爱奇艺：一季度总收入62.3亿元，会员收入环比增长36氪获悉，爱奇艺发布2026年第一季度财报。财报显示，一季度爱奇艺总收入62.3亿元。其中，会员服务收入42.0亿元，环比增长2%。在线广告服务收入12.4亿元，内容发行收入3.6亿元，其他收入4.3亿元。百度：一季度营收321亿元，AI业务收入136亿元36氪获悉，百度发布2026年第一季度财报。财报显示，一季度营收321亿元，上年同期营收324.52亿元。一般性业务收入260亿元，同比增长2%；其中AI业务收入136亿元，占百度一般性业务收入的52%，已连续多个季度增长；GPU云收入同比增长184%。百度创始人李彦宏表示，AI业务收入占比的首次过半，标志着AI已成为百度的核心驱动力。巨力索具旗下十余家企业已注销36氪获悉，近日，巨力索具因涉嫌信息披露误导性陈述违法违规，被证监会立案调查。天眼查App显示，巨力索具股份有限公司成立于2004年12月，法定代表人为杨建国，注册资本9.6亿人民币，经营范围包括应急装备、装备专用工装及工具和箱组的设计、生产和销售等，由巨力集团有限公司、杨建忠、香港中央结算有限公司等共同持股。对外投资信息显示，该公司对外投资26家企业，其中10余家企业处于存续状态，包括巨力索具（河南）有限公司、河北巨力应急装备科技有限公司等，其他10余家企业已注销。投融资：“华羽先翔”完成超亿元Pre-A+轮融资，蚂蚁集团领投36氪获悉，近日，国内全倾转eVTOL企业“华羽先翔”完成超亿元Pre-A+轮融资，本轮融资由蚂蚁集团领投、彼岸时代及复星创富跟投。融资资金将重点用于加速鸿鹄系列eVTOL的适航取证，以及低空出行及商业场景的落地与创新。“极壳科技”获5000万美元B+轮融资36氪获悉，近日，外骨骼品牌“极壳科技”宣布完成5000万美元B+轮融资。此轮融资由蚂蚁集团和美团龙珠领投，Sofina、Granite Asia跟投，高鹄资本担任独家财务顾问。此轮融资将加速极壳在技术研发、全球市场拓展及消费场景应用方面的布局。新产品：酷开发布企业AI操作系统36氪获悉，近日，深圳市酷开网络科技股份有限公司举办“场景智能·硅基管理”发布会，正式推出原生AI企业操作系统——酷开Happy Work AIOS Lite MVP。该系统已实现从生活场景到企业经营场景的全维度覆盖，完成多场景AI生态闭环落地。今日观点：马斯克预计今年美国将广泛使用自动驾驶汽车埃隆·马斯克周一表示，他预计今年晚些时候，无需人类驾驶的自动驾驶汽车将在美国更加普及。马斯克通过视频连线在特拉维夫举行的“智能出行峰会”上发言时称，得克萨斯州的三个城市已有无人驾驶汽车在没有安全监视员的情况下运行，并补充说今年这一模式将扩展至全美。（新浪财经）其他值得关注的新闻：2026电影总票房已超145亿元据票务平台数据，截至今天下午，2026年电影票房已超145亿元。五一档新片长尾效应持续释放，其中《给阿嬷的情书》连续九天蝉联单日票房榜首。（央视新闻）特朗普：本应要求获得更多英特尔公司的股份美国总统唐纳德・特朗普在周一出版的杂志采访中表示，他本应代表美国政府要求获得更多英特尔股份。特朗普政府去年持有英特尔10%的股份，并宣布向这家芯片制造商投资约100亿美元，用于在美国境内新建或扩建工厂。交易完成八个月后，美国政府持有的英特尔股份市值已超过500亿美元。（新浪财经）国家药监局：我国4月批准注册医疗器械产品318个今天从国家药监局获悉，我国4月共批准注册医疗器械产品318个。其中，境内第三类医疗器械产品272个，进口第三类医疗器械产品20个，进口第二类医疗器械产品26个。 (央视新闻)本文由「卜算籽」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：AI生成

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生8点1氪·2026年05月19日 08:08贾跃亭发布视频，宣布法拉第未来近期完成了7000万美元机构投资者募资。今日热点导览瑞幸咖啡正式上线两款含酒精的特调饮品京张高铁试点推出“自行车随身行”服务马斯克：计划今年进行首例用于治疗失明的Neuralink植入手术韩国法院要求三星电子罢工行动不得影响产量多家理财公司集中提前终止理财产品TOP 3 大新闻贾跃亭又拿到7000万美元投资贾跃亭发布视频，宣布法拉第未来近期完成了7000万美元（现汇率约合4.77亿元人民币）机构投资者募资，并将启动战略、产品技术&业务、财务、资本、AI体系五大变革。贾跃亭表示，其目标包括FFAI市值重回2021年上市水平，2027年第四季度实现经营性现金流为正，五年内在EAI人形与仿生机器人领域成为北美真实场景部署量前三的机器人公司+北美EAIMPV领军者。贾跃亭还表示，要将FF真正重构为一家Physical AI生态公司。两年内实现过去12年没有完成的梦想，进入“0借口、0内耗、结果导向”的战斗状态，全力兑现对股东和用户的承诺。（IT之家）巨力索具旗下十余家企业已注销36氪获悉，近日，巨力索具因涉嫌信息披露误导性陈述违法违规，被证监会立案调查。天眼查App显示，巨力索具股份有限公司成立于2004年12月，法定代表人为杨建国，注册资本9.6亿人民币，经营范围包括应急装备、装备专用工装及工具和箱组的设计、生产和销售等，由巨力集团有限公司、杨建忠、香港中央结算有限公司等共同持股。对外投资信息显示，该公司对外投资26家企业，其中10余家企业处于存续状态，包括巨力索具（河南）有限公司、河北巨力应急装备科技有限公司等，其他10余家企业已注销。联讯仪器股价反超贵州茅台，成为A股新的股王5月18日，联讯仪器盘中扩大涨幅至13%，股价反超贵州茅台，成为A股“新股王”，该股较发行价上涨超15倍。截至5月18日11时许，联讯仪器最高涨至1361元/股，涨幅高达16.32%。与此同时，贵州茅台由涨转跌，股价探至1330元之下。此前相继登上A股“股王”宝座的寒武纪（688256）、源杰科技（688498），5月相继迎来2025年年度权益分派。公开资料显示，联讯仪器4月24日登陆A股，发行价81.88元/股，上市14个交易日来，目前股价较发行价上涨15.6倍。据公开资料显示，联讯仪器是国家级专精特新“小巨人”企业，专注于高速通信、电性能、光芯片及半导体晶圆芯片等领域测试仪表与设备的研发、制造，是国内高端电子测量与半导体测试设备领域的硬核龙头。（澎湃新闻）大公司/大事件何立峰会见美国超威半导体公司董事会主席兼首席执行官苏姿丰5月18日下午，中共中央政治局委员、国务院副总理何立峰在人民大会堂会见美国超威半导体公司董事会主席兼首席执行官苏姿丰。何立峰指出，上周中美两国元首在北京举行重要会晤，达成一系列重要共识。在两国元首战略引领下，两国经贸团队达成了总体平衡积极的成果，这将为下一步的中美经贸合作与世界经济注入更多确定性和稳定性。欢迎包括超威半导体公司在内的跨国公司把握中国发展机遇，深化互利合作。苏姿丰积极评价两国元首会晤成果，表示愿继续拓展在华业务，持续加大在华投资。(新华社）美陪审团驳回马斯克对OpenAI的诉讼当地时间5月18日，美国陪审团裁定，美国开放人工智能研究中心（OpenAI）及其首席执行官萨姆·奥尔特曼无需就“偏离慈善使命”一事对埃隆·马斯克承担责任，原因是马斯克提起诉讼时间过晚。当天晚些时候，马斯克表示，将就这一裁决向美国第九巡回上诉法院提起上诉。此前，马斯克对OpenAI及其首席执行官奥尔特曼提起诉讼，指控OpenAI最初邀请马斯克资助一个致力于造福人类的非营利AI研究机构，但如今却专注于盈利。（央视新闻）张朝阳：搜狐未入局大模型研发，将聚焦AI应用落地36氪获悉，在2026搜狐科技年度论坛上，搜狐CEO张朝阳表示，“搜狐并未入局大模型第一阵营的研发，大模型研发需要雄厚的资金实力和技术背景，企业应基于自身业务基础理智选择。”他提到，搜狐将聚焦AI应用落地，在提升编程效率、节约运营成本的同时，对内容生成保持克制，保持内容中立性，规避AI带来的内容乱象风险。腾讯音乐：收购喜马拉雅的事项已经完成36氪获悉，腾讯音乐在港交所公告，于2026年5月18日，根据并购协议的条款，收购喜马拉雅的事项已经完成，据此，喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为12.6亿美元的现金及最多总数1.75亿股本公司A类普通股（包括本公司于交割时新发行的A类普通股，以及本公司授出的以股权为基础的激励相关的A类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。川投新能源公司总经理刘亮被遣返回国据中央纪委国家监委网站，近日，在中央反腐败协调小组国际追逃追赃工作办公室统筹协调和公安部等部门协助下，经四川省监察机关、公安机关与有关国家执法机关密切合作，外逃职务犯罪嫌疑人刘亮在境外落网并被遣返回国。刘亮，男，1979年1月出生，川投新能源公司总经理，涉嫌严重职务违法犯罪，于2026年4月外逃，同月四川省资阳市监察机关对其立案调查。办案机关积极开展国际执法合作，于近日将其缉捕归案。（界面新闻）英特尔CEO称代工业务取得进展英特尔CEO陈立武在当地时间周一接受采访时表示，该公司的外部代工业务正在稳步发展，成为其扭亏为盈的关键一环。“代工业务非常重要，”陈立武表示，“它是重要的国家资产之一。”（新浪财经）瑞幸咖啡上线两款含酒精的特调饮品5月18日，瑞幸咖啡正式上线两款含酒精的特调饮品，并且属于饮料酒的范畴，正式布局“微醺”经济。记者了解到，这两款含酒精饮品是瑞幸对全国市场的一次测试，后续或还有进一步的计划。在业内看来，瑞幸布局微醺赛道，意在盘活闲置门店时段资源，提效增收。记者在北京部分瑞幸门店看到，两款年度特调饮品已经上线，但只有部分门店可以提供含酒精版。（红星新闻）韩国法院要求三星电子罢工行动不得影响产量据央视新闻消息，韩国水原地方法院5月18日批准了三星电子公司提出的部分禁令请求，责令该公司工会必须确保即将开始的全面罢工行动“不影响产量”，“不得导致这家全球重要的存储芯片制造商生产原料受损”。据悉，这实际上是给三星电子这次史上最大规模罢工计划“踩了急刹车”。今年3月中旬，由超过6.6万名三星电子工会成员参与表决的投票结果显示，93.1%的工会成员赞成罢工。全面罢工将于5月21日至6月7日举行。据悉，三星电子作为全球重要的存储芯片制造商，一旦发生罢工，可能进一步加剧由全球人工智能数据中心建设持续升温带来的全球半导体供应趋紧局面，其影响或波及汽车、计算机及智能手机等多个行业。三星电子劳资双方此前就2026年薪资问题进行了多轮谈判，但双方就部分议题的分歧持续扩大，谈判最终破裂。（央视新闻）多家航司下调国际航线燃油费，国内航线刚上调国内航线燃油附加费自5月16日再度上调后，多家航司却在陆续下调国际航线的燃油附加费。香港航空宣布，自今日（5月18日）开始下调多条航线的燃油附加费。其中中国香港前往内地的地区航线，燃油附加费由290港元下调至190港元；香港前往日韩、新马泰等则由389港元下调至339港元。在此之前，国泰航空和香港快运已宣布自5月16日起下调部分区域的燃油附加费。其中香港始发长途航班从1560港币下调至1362港币，中程航班从725港币下调至633港币，短途航班从389港币下调至339港币。下调幅度在12%左右。这也是中东冲突爆发后，多家航司首次对核心航线燃油附加费进行下调。（第一财经）5月19日起，铁路部门在京张高铁试点推出“自行车随身行”服务据新华社消息，为更好地满足旅客多样化出行需求，铁路部门5月19日起在京张高铁北京北至崇礼间各车站间运行的G字头动车组列车试点推出“自行车随身行”服务，旅客可通过12306客户端线上预约并付费，乘车时经车站安检和规范包装，可携带符合规定的自行车同车出行，将为广大骑行爱好者跨区域出行提供更多便利。试点初期，“自行车随身行”服务定价88元/辆，旅客可在乘车结束后180日内通过铁路12306客户端开具电子发票。（新华社）戴尔将与谷歌和SpaceX等公司展开合作戴尔科技在上一季度为其“AI工厂”产品线新增1,000家客户，使其总客户数达到了5000家。戴尔正在推出多款新产品以协助企业搭建AI应用，其中包括一款名为“Dell Deskside Agentic AI”（戴尔桌面智能体AI）的产品。此外，戴尔还与谷歌和SpaceX等公司展开合作，将AI模型与工具引入企业自身的内部网络中。（财联社）SpaceX将迎关键试飞SpaceX计划于周二进行一次关键的试飞，发射其星舰巨型火箭的升级版，代号V3。这次发射正值SpaceX上市关键阶段，消息透露预计该公司最早在下月IPO，并可能在本周三公开招股书。与此同时，SpaceX在与贝索斯的蓝色起源竞争，为NASA的阿尔忒弥斯计划开发星舰，供其在2028年的载人登月任务中使用。目前来看，星舰的研发进度落后于NASA对SpaceX的预期。该火箭原计划于2023年首飞，但去年的失败延缓了其发展进程。星舰最近一次试飞，也是其第11次试飞，发生在七个月前。（财联社）NextEra拟以660亿美元收购Dominion，或创史上最大规模电力并购纪录5月18日消息，据报道，美国公用事业巨头NextEra Energy目前正就收购竞争对手Dominion Energy进行谈判，该笔交易方案主要以股票交换形式进行，对Dominion的估值约为每股76美元，总价约660亿美元。若最终达成，这将成为迄今为止电力行业规模最大的一笔交易。NextEra计划以约0.8股自身股票交换1股Dominion发行在外的股票，交易总额中还将包含一小部分现金成分。交易完成后，NextEra的股东将持有合并后新公司约75%的股份。（界面新闻）马斯克：计划今年进行首例用于治疗失明的Neuralink植入手术马斯克表示，旗下Neuralink技术能帮助部分失明者恢复视力，计划今年进行首例用于治疗失明的Neuralink植入手术。据了解，Neuralink是一个美国神经科技和脑机接口公司，由伊隆·马斯克和八名其他联合创办者创办，负责研发植入式脑机界面技术。公司的总部在旧金山。Neuralink于2016年成立，并于2017年3月公之于众。2019年7月17日，Neuralink对外宣布该公司研发的一款脑机接口系统（财联社）五粮液：2025年度百亿分红两月内到账，严格兑现200亿元分红承诺5月18日，五粮液2026年第一次临时股东会对《2025年度利润分配方案》进行了审议，五粮液副董事长、总经理华涛表示，2025年度分红资金将在两个月内准确、及时划入全体股东账户。五粮液2025年中期已分红100亿元，本次拟再分红100.07亿元。五粮液践行高比例分红，严格兑现《2024—2026年股东回报规划》承诺，连续3年每年度现金分红总额不低于200亿元（含税）。（证券时报）美国连锁快餐品牌Wendy’s将进入中国市场5月18日在美国连锁快餐品牌Wendy’s官网上看到，该品牌国际特许经营重点市场亚太区域中出现了中国市场。早前，Wendy’s在2026年第一季度的财报中称，公司已经签署一项特许经营协议，计划未来10年在中国建设最多1000家餐厅。Wendy’s称合作方是“一家在中国拥有数十年经验的大型餐饮运营商，以强待客能力闻名、深刻理解中国消费者。”但没有公布这家运营商的名称。此外，财报显示，Wendy’s2026年一季度全球销售额同比下降5.5%，美国同店销售下降7.8%，国际市场销售额增长6.0%。（界面新闻）贵州茅台：优化茅台酒线下门店营业时间及i茅台App开售时间36氪获悉，据贵州茅台消息，为进一步贴近广大消费者日常工作与生活节奏，更好地满足广大消费者购酒需求，自2026年5月

高瓴出手了一家AI体育科技公司，曾获李泽湘天使轮融资｜硬氪首发黄 楠·2026年05月19日 09:30从“记录工具”到“AI运动Copilot”的跃迁。作者｜黄楠编辑｜袁斯来硬氪获悉，AI体育科技公司SportVision近日已完成天使+轮融资，高瓴创投（GL Ventures）领投，谦恒资本担任独家财务顾问。本轮资金将重点用于核心技术研发、产品量产落地及市场的拓展推广。其新款AI网球训练相机产品计划发售中。SportVision隶属于深圳思博威视体育科技有限公司‌，长期专注运动算法研发与硬件产品应用、构建AI智能教练体系，此前已获得李泽湘教授旗下清水湾基金的天使轮投资。过去几年，运动科技赛道几乎被硬件厂商重新定义了一遍。发球机让“一个人练球”成为可能，运动相机令“训练过程的记录”不再需要专人跟拍，智能手表/手环把心率、步频、消耗卡路里等量化成精确的长期数据记录。尽管每一类产品都跑出了规模不小的公司，但它们各自停留在陪练、记录、监测的单点功能上——数据相互独立、能力无法贯通，难以为用户提供连贯、整体的个性化支持。运动人群的深层渴望远不止于被记录，还有被指导、被陪伴、被看见进步。一场更底层的变革正在发生：从“记录工具” 到 “AI运动Copilot”的跨越，运动科技赛道正在重写。工具只需完成指令，而Copilot则可以像教练一样理解状态、识别动作、纠正习惯，甚至在运动热情消退时成为那个督促陪伴的人。要补上这道体验鸿沟，用户需要的不只是更强性能的硬件，更是一个能把整合所有信息、真正理解运动者的“大脑”。这也正是SportVision精准切入的核心位置。全运会现场所使用的SportVision设备（图源/企业）SportVision创始人陈楷夫本科毕业于吉林大学物理学院严济慈班、哥伦比亚大学机器人学硕士，曾在MMLab从事姿态识别与目标检测算法研究；身为一名前游泳运动员，他对运动学习、动作纠正与训练效率的痛点有切身感知，正是这段经历，让他决心用AI技术破解非结构化运动数据难题，构建一个可规模化、可复制的个性化运动产品和训练体系。联合创始人杨晨旭是资深网球爱好者，曾在腾讯、字节负责AI产品开发。多位团队核心成员来自商汤、字节、阿里、腾讯等头部科技企业，具备AI算法、软硬件一体化研发与全球化品牌运营能力，已形成了从技术研发到产品落地的完整闭环。公司起步初期，SportVision选择从羽毛球场景切入进行关键产品验证，首款产品“好球哇”是一套羽毛球AI运动相机方案。设备可实时采集球场画面，通过视觉算法自动识别并捕捉用户的运动高光时刻，为普通运动爱好者、专业运动员及赛事场景，提供高光集锦、赛事直播、运动数据分析等一体化服务。这套系统落地了全国数十家头部羽毛球馆，并获得林丹、王睁茗、龚睿那等世界冠军在内的多位种子用户支持。在赛事侧，该系统也服务了全运会残特奥会、粤港澳大湾区羽毛球比赛、深圳杯等多场关键赛事。但它真正的价值，远非硬件本身，而是持续运行中沉淀的海量运动数据。在拍摄过程中，SportVision自研算法能实时分析并智能预测运动轨迹，根据比赛进程自动调整摄像角度，对复杂的多人运动场景进行针对性学习优化。从世界冠军到刚入门的爱好者，从专业教练的指导语料到普通用户的练习习惯，场馆设备每天运行十几个小时，能收集到各类各样的泛化数据。每一场球、每一次挥拍、每一个得分瞬间，都在被系统记录、拆解、分析。这是一座难以被发球机或运动相机采集到的数据矿。运动学习的基本机制分为三个阶段：认知阶段（通过视觉掌握动作要领）、联想阶段（在反复练习中形成神经通路）、自主阶段（动作固化成为肌肉记忆）。其中联想阶段的及时反馈最为关键：动作做对还是做错需要在几秒内被确认，否则错误动作就会被“练熟”。陈楷夫向硬氪指出，“此前，多数运动AI解决方案仍停留在‘事后分析’阶段，用户完成运动后上传视频、系统给出复盘报告。但这种模式错过了最佳纠正窗口。”因此，SportVision想搭建的，是一套基于实时交互的预判式AI。“好的运动员不一定是好教练，好教练的价值在于知道怎么教。”陈楷夫说，“我们要做的，是把教练的教法萃取出来，让AI学会怎么教。”随着网球、羽毛球、匹克球等拍类运动在全球范围内持续升温，用户对运动提升的需求愈发迫切。传统的请教练模式面临成本高、频次低的矛盾，一周请一两次教练，用户练习时无法得到反馈，对于向网球这样的运动，可能需要一年才能从入门到实现“对打社交”。正反馈周期长，也使得很多人停留在了入门阶段。这正是SportVision的核心壁垒所在：那些在球场上日夜运转的设备，收集到的海量数据。从骨骼点识别到动作模式分析，从专业运动员的动作库到普通用户的练习轨迹，这些数据构成了可迁移的技术底座，能够在各类拍类运动中得以技术复用和迁移。SportVision首款网球AI Copilot产品（图源/企业）“我们做的不是一款运动相机，而是构建未来运动人群个性化AI教练的世界模型。”陈楷夫告诉硬氪，“用户甚至不需要思考‘我该怎么练’，而是由系统主动管理你的运动进步。”硬氪获悉，SportVision首款网球AI Copilot产品正在紧密的研发流程中。以相机为载体，系统可识别用户运动动作并与专业标准动作比对，实时反馈纠正建议；同时，结合教练指导场景，生成个性化训练方案，让用户在独自练习时也能获得针对性、科学化的AI陪练指导。该产品概念此前已在CES 2026上展出，首发市场聚焦欧美地区。这里拥有了全球密度最高的网球用户群体，也是付费意愿与消费能力最突出的核心市场。回到AI时代运动科技竞争的底层逻辑迁移，当大模型逐渐消解知识鸿沟，竞争的壁垒不再停留于算法参数，而是向两端迁移——一端是数据的密度与深度，另一端是场景的介入与闭环。SportVision长期积累的稀缺数据资产，正在成为其下一代AI教练产品的技术底座。从情绪价值到功能性与情绪价值兼具、从记录到指导，从单一品类到拍类运动通用技术底座，这家公司正在尝试的命题是，加速推动运动科技从“工具”到“AI Copilot”的关键升级。本文由「黄 楠」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：企业授权、视觉中国

影石：重新定义性价比黄绎达·2026年05月19日 14:58影石有个“苹果梦”？作者 | 张帆坚持做后来者近日，影石创新发布了登陆资本市场后的首份年报。2025年，公司实现营业收入97.41亿元，同比大幅增长74.76%，创下成立以来的最高营收纪录。然而，同期由于研发、销售等费用高企，公司归母净利润仅为9.29亿元，同比下降了6.62%。面对这样一份“营收大增、利润下滑”的财报，公司召开了上市后的首次投资者交流会，以回应资本市场的关切。根据2025年年报显示，报告期内，公司影像业务营收增长了78%，主要得益于销量的大幅提升——2025年实现销量393万台，同比增长76%。另一方面，产品均价相对平稳，约为2169元，同比增长1%。从ASP趋势来看，公司产品整体呈现下降态势，从2022年的2,391.45元/台连续两年下滑至2024年的2147元，但在2025年实现了企稳。在当前激烈的市场竞争环境下，价格企稳的主要原因，是受益于公司积极拓展B端等高单价客户的提振。尽管有争议，但不可否认的是，即便面临激烈的价格竞争，影像行业依旧是消费电子中的高毛利子行业。根据2025年的财报，影石毛利率高达48.82%，与科沃斯相当，但相比小米、传音等手机厂商约20%左右的毛利率。影石的毛利率高出约28个百分点。图：各消费电子企业毛利率；资料

9小时前法官确认裁决结果，驳回马斯克对 OpenAI 的诉讼美国加州北区联邦地方法院法官 Yvonne Gonzalez Rogers 近日确认最新裁定结果：驳回埃隆·马斯克要求对 OpenAI 和微软签发“全面禁制令”的动议与其先前提出的欺诈类指控，但保留关于“是否存在并被违反的公益/慈善信托”以及“不当得利”等核心争议，案件不会就此终结，而是将进入加速审理并由陪审团作出最终裁决。3 来源法官确认：禁制令与部分指控被否，核心诉求仍将审理3 来源从出资联合创办到天价索赔：马斯克与 OpenAI 的纠纷缘起3 来源法院为何驳回禁制令与部分指控：证据不足与适格障碍3 来源影响与展望：AI 治理、非营利结构与科技巨头关系迎来“法律考题”3 来源本内容由AI生成

关注最新看重点大公司财经动态看世界硅谷动态老·司机关心智能出行的汽车产业观察者大众汽车将发行30亿元熊猫债。5月19日，大众汽车国际金融股份公司发布公告，公司2026年度第一期中期票据（债券通）申购启动。本期债券的本金总额为不超过人民币30亿元，其中品种一的初始发行规模为人民币20亿元，期限为三年，申购利率区间为1.70%-2.50%；品种二的初始发行规模为人民币10亿元，期限为五年，申购利率区间为1.85%-2.65%。38秒前00举报发表0/500虎嗅用户社区交流公约读数数字里的故事华东地区首列中越“铁路快通”启运。今天，载有38个集装箱电子料件、汽车配件、显示器件等货物的75128/7次列车，从苏州西站货场鸣笛启程，一路向南经广西凭祥口岸出境，驶向越南河内，为长三角区域企业成功打通直达越南、链接东盟的南向跨境物流精品通道。这是华东地区首列采用“铁路快通”模式直达越南的班列，采用全程不换装、不开箱的“一箱直达”运输模式，货物始发后无需中转拆箱、无需在口岸海关重复申报，依托铁路全天候运行优势，不受恶劣天气、道路交通影响，叠加铁路快速通关模式，有效压缩在途及中转时间，同时凭借铁路规模化运输优势，切实为外贸企业实现高效通行、降本增效。4分钟前00举报发表0/500虎嗅用户社区交流公约新政新规各种新政策教育部：个别地方和幼儿园存在滥用、误用数字产品现象；广大家长注意防范“电子带娃”，，抵制早教类App等商业宣传误导。今天（19日），教育部启动2026年全国学前教育宣传月，主题是“共同守护数字时代的童年”，呼吁各地幼儿园“以游戏为基本活动”，不能以伴读、聊天、互动游戏等儿童数字产品替代图画书、玩具，不能用数字产品直接生成儿童观察记录、评价幼儿等。引导广大家长树立“多陪伴、多游戏、多运动、多亲近自然”的育儿理念，自觉识别并抵制早教类App等商业宣传误导，防范“电子带娃”。当前，部分人群对学前领域数字应用存在一些认识误区，部分商业机构存在宣传误导，个别地方和幼儿园存在滥用、误用数字产品现象。大量权威研究证实，幼儿过早、过多接触数字产品，将会对视力健康、注意力集中、语言交往、情感发展等方面带来长期且难以逆转的不利影响。倡导学前儿童远离数字产品，保护儿童发展权益，也已成为世界各国的广泛共识。宣传月期间，各地将持续开展系列主题宣传活动，整合多方专业资源与社会力量，聚焦数字时代学前儿童权益保护，推动多方协同联动，构建全方位、立体化的幼儿成长防护屏障。近年来，我国学前教育普及普惠水平大幅提升，2025年全国学前三年毛入园率达到92.9%，全国普惠性幼儿园覆盖率达到91.5%。2025年6月1日，学前教育法正式施行，7月底，国家部署逐步推行免费学前教育，我国学前教育进入了有法可依、逐步免费的新时代。5分钟前00举报发表0/500虎嗅用户社区交流公约安全小喇叭重大灾害事故预警、播报、跟进广州一骑手在交通事故中当场身亡，该配送团队多名骑手未处理交通违法行为，配送站负责人因涉嫌重大责任事故罪被刑拘。据广州交警视频号5月18日消息，5月18日，广州交警披露广州首例即时配送站点负责人因涉嫌重大责任事故罪被依法刑事拘留的案情。据广州交警介绍，今年1月13日，广州荔湾区崇文四路发生一起交通事故，一名即时配送骑手在事故中当场死亡。经认定，骑手承担事故主要责任。广州交管部门在随后的事故深度调查中发现，该配送团队多名骑手均存在驾驶无牌电动自行车、闯红灯、逆行、非法改装、未按规定车道行驶等多宗交通违法行为未处理等情况。经调查，涉事配送站点负责人莫某某未切实履行安全生产管理职责，未落实安全隐患排查、安全教育培训等工作，其行为涉嫌重大责任事故罪，公安机关依法对莫某某予以刑事拘留。9分钟前10举报发表0/500虎嗅用户社区交流公约飞行圈航空领域靠谱资讯速递~闪送获杭州国资千万元投资，切入低空配送赛道。即时配送企业闪送正加快在低空经济领域的布局。记者获悉，杭州国资拟向闪送旗下全资子公司——杭州同城必应科技有限公司投资数千万元，用以布局低空物流配送领域，推动其从地面即时配送向“地面+低空”立体化配送体系升级。工商资料显示，此次国资投资主体为杭州低空产业发展有限公司，该企业系杭州市实业投资集团有限公司全资子公司，是杭州首家跻身世界500强的市属国有企业，此次投资是其低空经济产业生态建设的重要一步。10分钟前10举报发表0/500虎嗅用户社区交流公约捉“谣”记这里最常见的消息是“否认三连”宗申机车：网传涉及公司赛科龙AQS250低价销售相关信息均为不实、伪造、虚假内容，非公司发布。重庆宗申机车工业制造有限公司发布声明称，经公司全面核查确认：网传涉及公司的赛科龙AQS250低价销售相关信息均为不实、伪造、虚假内容，并非公司发布，公司从未授权任何个人、第三方机构开展相关活动，也从未推出、承诺网传相关政策与服务。“针对本次网络造谣、传谣及涉嫌诈骗的违法行为，公司已第一时间固定全部证据，并同步开展全方位维权工作。”16分钟前00举报发表0/500虎嗅用户社区交流公约一只飞虎最近国外发生了什么？格陵兰岛自治政府总理重申“不卖岛”：愿意继续与美国对话，但格陵兰人民必须得到尊重；此前特朗普重提吞并意愿：他们不想给我们。5月19日，据CCTV国际时讯消息：当地时间5月18日，丹麦自治领地格陵兰岛自治政府总理尼尔森表示，他在会见美国总统特朗普任命的“格陵兰岛特使”杰夫·兰德里以及美国驻丹麦大使肯尼思·豪韦里等美方官员时重申，格陵兰岛“不出售”，格陵兰人民的自决权“不容讨论”。他表示，格陵兰岛自治政府愿意继续与美国对话，但格陵兰人民必须得到尊重。据当地媒体报道，杰夫·兰德里5月17日率团抵达格陵兰岛首府努克，定于19日至20日出席当地商界协会主办的“未来的格陵兰岛”经济论坛。格陵兰广播公司报道称，活动组织者表示，虽然论坛对所有人开放，但并未邀请兰德里与会。此前，特朗普重提吞并格陵兰岛意愿。据环球时报援引彭博社7日称，美国总统特朗普再次就格陵兰岛问题对北大西洋公约组织（北约）表示不满，并对该安全联盟在中东战事上的立场表达了新的失望。“说实话，这一切都始于格陵兰岛。”特朗普6日在白宫新闻发布会上说，“我们想要格陵兰岛。他们不想给我们。然后我就说，‘再见’”。报道称，自去年重返白宫以来，特朗普多次寻求获得这一隶属于丹麦的自治领地。今年早些时候，特朗普曾以隐晦的军事威胁和关税措施施压欧洲盟友，之后又声称已就该岛达成一项“协议框架”，解决了他的国家安全顾虑，此举导致美国与欧洲盟友的关系紧张。围绕格陵兰岛的战略布局，美国近期正推动扩大在当地的军事存在。英国《独立报》称，美军方面透露，华盛顿计划新增进入岛上三处地点的军事准入权限，其中包括此前已废弃的两处基地，以强化在北极地区的部署能力。不过，不少格陵兰岛居民反对该计划。据《纽约时报》报道，美国北方司令部司令吉洛特曾在美国国会的一场听证会上表示，鉴于“不断上升的威胁”和格陵兰岛日益凸显的战略重要性，美军希望获得进入格陵兰岛更多基地的权限，并推动开发更多港口和机场，为美国政府在北极地区提供更多行动选项。目前，美军在格陵兰岛只有皮图菲克太空基地还在运作。丹麦公共广播公司称，丹麦及其盟友曾因担忧美国可能入侵格陵兰岛，在1月向当地增加部署军队，甚至计划在极端情况下破坏机场跑道以阻止美军飞机降落。此外，特朗普在6日的发言中称，北约成员国在当前的中东战事中故意不提供帮助，并再度将北约称作“纸老虎”。他还批评了韩国、澳大利亚和日本，指责它们没有在此次中东战事中提供援助。特朗普还透露定于8日与北约秘书长吕特举行会晤。23分钟前10举报发表0/500虎嗅用户社区交流公约一只飞虎最近国外发生了什么？美尼联手，发动打击。新华社消息，美军非洲司令部18日发表声明称，美军与尼日利亚协调，于17日对盘踞在尼东北部的极端组织“伊斯兰国”武装分子进行了打击。声明称，美军和尼日利亚武装部队没有人员受伤。美国总统特朗普15日通过社交媒体发文说，美军与尼日利亚武装部队“当晚”执行联合任务，“清除”极端组织“伊斯兰国”二号头目阿布－比拉勒·米努基，该组织的“全球行动能力已大幅削弱”。去年12月，美军曾对尼日利亚境内“伊斯兰国”武装分子发动打击。25分钟前10举报发表0/500虎嗅用户社区交流公约雷布斯朋友Are you ok?小米寻天新供应商近期成立多家公司揭秘小米寻天新电池供应商。据媒体报道，小米汽车在昆仑车型上引入新电池供应商中创新航。小米第三款车型的电池定点在2024年年初敲定，当时首选的电池供应商是欣旺达，之后才确定中创新航为二供。目前，从小米给供应商的供应配额上来看，欣旺达的电池供应比例占60%，中创新航占40%。天眼查App显示，中创新航科技集团股份有限公司成立于2015年12月，法定代表人为刘静瑜，注册资本约17.7亿人民币，经营范围包括锂离子动力电池、电池管理系统、储能电池及相关集成产品和锂电池材料的研制、生产、销售和市场应用开发等，由常州金沙科技投资有限公司、四川成飞集成科技股份有限公司、厦门锂航金智股权投资合伙企业（有限合伙）等共同持股。对外投资信息显示，去年以来，该公司成立了中创凯博科技（自贡）有限公司、凯博宏泰（湖北）私募股权投资基金合伙企业（有限合伙）等多家新公司。28分钟前10举报发表0/500虎嗅用户社区交流公约新政新规各种新政策三亚启动防汛防风Ⅳ级应急响应。三亚发布”消息，根据气象预报，受短时强降雨影响，三亚市海棠、崖州、育才三个区发布暴雨红色预警，且未来3个小时各区仍将有较强降雨。根据《三亚市防汛防风防旱应急预案》有关规定和会商研判，市防灾减灾救灾和安全生产委员会决定于5月19日06时启动防汛防风Ⅳ级应急响应，请各区各单位按照《三亚市防汛防风防旱应急预案》相关要求和各自职责分工做好各项防御措施及应急处置工作，并及时报送工作信息。29分钟前00举报发表0/500虎嗅用户社区交流公约捉“谣”记这里最常见的消息是“否认三连”网传“蒜薹滞销、农户弃收”？官方核实：夸大失真。买五斤发十斤”“免费抽蒜薹，抽多少拿多少”。近期，网上曝出河南省开封市杞县“蒜薹滞销、农户弃收”的传言。经当地官方实地核查，“蒜薹滞销”传闻存在夸大与失真。据了解，杞县大蒜主要分为蒜薹、蒜头两类种植品种，其中以蒜头为主要收益的地块，蒜薹价值偏低，属于附属农产品。叠加今年全国大蒜整体丰产，部分品相较差的蒜薹售价低，少数农户自愿放弃采收，并非“全域滞销”。31分钟前00举报发表0/500虎嗅用户社区交流公约时讯内参树欲静而风不止，新鲜资讯速递俞浩：后续可能逐渐减少自媒体展现。5月19日消息，昨日深夜，追觅科技CEO俞浩在社交媒体透露，后续可能逐渐会减少在自媒体的展现。他表示：“各位追觅的同学们，你们要多发视频，把你们所做的工作、你们的产品、你们的技术、你们的创新去讲清楚。前面我以身入局带着大家去在自媒体上展现自己的工作。后续我可能要逐渐减少在这块的出现，但你们要多发，你们要多出现，好吗？”此前，俞浩在社交媒体开启“刷屏式”更新，连续多日密集发布公司动态，每日平均发布几十条内容，单条发布间隔甚至仅仅几分钟。俞浩还要求每一位追觅员工开通所有平台的社交媒体账号，每天抽出15分钟拍摄视频，发布三条内容。他提出，仅追觅员工自身，就要打造至少两万多个账号。34分钟前10举报发表0/500虎嗅用户社区交流公约读数数字里的故事机构：受存储成本上涨影响，一季度东南亚智能手机平均售价创历史新高。5月19日，Omdia最新研究显示，2026年第一季度东南亚智能手机市场出货量同比下降9%，总量为2160万部。然而，市场最值得关注的并非出货量下滑，而是平均售价（ASP）的变化：受存储成本上涨影响，2026年第一季度，东南亚智能手机平均售价达349美元，创历史新高，同比上涨19%。其中，三星以460万部出货量和21%市场份额位居东南亚市场第一，同比增长4%；OPPO以420万部排名第二，出货量同比下降17%；小米位列第三，出货370万部，同比下滑12%。35分钟前00举报发表0/500虎嗅用户社区交流公约雷布斯朋友Are yo

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文GPT-5.5才发三周，5.6已在内测！OpenAI与Anthropic补贴大战同日开打，开发者坐收渔利发布于AI新闻资讯发布时间 :2026年5月14号 14:21阅读 :1分钟三周前GPT-5.5刚落地，GPT-5.6的消息已经提前炸出来了。知名爆料人Leo透露，GPT-5.6目前开发进度已全速推进，首批内部检查点在过去几天已启动测试，下个月很可能就会正式露面。更有意思的是，有人从OpenAI内部Codex日志里扒出了rollout mapping的痕迹——大部分调用还是指向GPT-5.5，但有一条记录赫然落在了GPT-5.6身上。换句话说，Codex环境里可能已经在偷偷用它跑测试了。内部代码名也被顺带曝光:ember-alpha和beacon-alpha。网友直呼:这迭代速度，根本追不上。速度再度拉满——Codex即将上线ultrafast模式GPT-5.6还没到，OpenAI产品端已经先按捺不住了。爆料人Chetaslua同步透露，本周四OpenAI将在Codex上线"ultrafast模式"，响应速度直接提升2到3倍，专门为延迟敏感型任务设计。与此同时，Image Arena排行榜上以+242分断档领先的gpt-image-2，也在同步推进A/B测试更新。这不是第一次了。今年3月GPT-5.4发布时，/fast模式已经实现了1.5倍加速;GPT-5.3-Codex-Spark借助Cerebras芯片，更是把推理速度拉到每秒超过1000token，是普通模式的15倍。而这次的ultrafast，直接在旗舰主力模型上实现提速——不是阉割版，不是小模型凑数，是正儿八经的全尺寸加速。对开发者来说，Agent循环、长任务流水线、浏览器自动化……所有需要"等"的地方，体验将彻底不同。顺带一提，上周ChatGPT官方账号发出一张科幻感十足的图，界面上那句经典的"Ask ChatGPT"悄悄换成了"Message ChatGPT"。奥特曼丢下一句"Call me maybe"，让外界对ChatGPT集成语音通话乃至推出硬件产品的猜测，彻底烧开了。Codex vs Claude Code，补贴战正式开打硅谷最戏剧性的一幕，就在今天上演了。就在OpenAI ultrafast模式上线前夕，Anthropic率先出手——宣布从6月15日起，为付费用户每月提升50%的编程额度，覆盖Claude Agent SDK、命令行工具以及深度接入GitHub工作流的Claude Code，并同步发布了Opus4.7Fast模式，主打比Codex更快的高级推理与更流畅的长上下文编码体验。OpenAI没让人等太久。反手祭出一招重磅补贴:接下来30天内，任何想从其他平台迁移到Codex的企业，直接获赠2个月免费使用权。按Pro计划每月200美元算，相当于白送400美元。奥特曼本人亲自下场喊话，直言Codex是"目前市面上最强的AI编程产品"。仅仅几个小时后，战报出炉:2000名开发者在3小时内主动联系了OpenAI。这场怼脸硬刚，看呆了整个硅谷。开发者们倒是乐了——被两家顶级公司抢着补贴，这种好事，上哪儿说理去。飞轮已经开始自转把这两件事放在一起看，会发现一个比任何单一新闻都更值得关注的趋势正在成形。AI正在加速自我进化。GPT-5.3-Codex是OpenAI首个"参与自身训练"的模型，到GPT-5.5，OpenAI内部85%的员工每周都在用Codex写代码。GPT-5.6几乎可以确定是在GPT-5.5的深度参与下开发出来的——AI在帮OpenAI造更强的AI。与此同时，编程工具的大规模普及正在释放前所未有的工程产能。Codex已有300万周活跃用户，Claude Code的用户数同样在爆发式增长。数百万开发者把AI编程工具当成日常标配，AI生成的代码又反哺回AI的训练与部署，这个循环只会越转越快。奥特曼说过，OpenAI的目标已经不止于AGI，直接瞄准了ASI。当模型迭代速度由AI自身驱动，当AI编程工具成为AI开发的基础设施，当两家万亿级公司用"补贴战"加速这一切的普及——通往ASI的飞轮，已经开始自转了。

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文李开复：开源模型是实现“AI主权”的更优路径发布于AI新闻资讯发布时间 :2026年5月12号 18:01阅读 :1分钟创新工场董事长李开复近日在专访中深入探讨了“AI主权”的概念。他指出，人工智能主权不仅关乎技术控制权和数据安全，还涉及模型与本土文化及法律的适配性。他认为，并非每个国家都有必要重造一个OpenAI，盲目追求昂贵的闭源自研并非良策。李开复强调，对于大多数资源有限的国家和企业而言，基于开源模型构建本地化体系是更现实的选择。这种“第三条路”通过在开源基础上进行增量训练，能以极低的成本让模型符合特定地区的价值观，而非受制于他国的通用模型。工程效率与生态之争谈及中美竞争，李开复认为资源受限反而催生了中国企业的极致工程效率。他以DeepSeek为例，指出中国公司在算力不足的情况下，通过优化架构和减少依赖，能以不到美国十分之一的投入实现极高的性能。他预测AI领域将演变成类似“iPhone与安卓”的格局。闭源模型如同一方领地，追求高利润和生态控制力；而开源模型则代表了更广阔的市场覆盖和普及度，两者将在全球范围内长期并存，共同构建未来的智能生态。中国硬件的下半场机遇李开复对中国在物理AI和硬件制造方面的结构性优势表示乐观。他指出，中国拥有全球最完整的供应链体系，从智能穿戴到机器人，制造成本远低于西方。未来AI时代的标杆级硬件，极有可能由中国设计并品牌化。此外，中国在语音驱动交互和大规模现实数据生成方面也具备独特文化优势。随着“人工智能+”政策对实体经济的赋能，硬件与AI的深度结合将成为中国在AI下半场竞赛中的核心竞争力。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文国家队联手科技队：百度地图全面接入国家级数据化身AI智能副驾发布于AI新闻资讯发布时间 :2026年5月19号 9:54阅读 :1分钟百度地图近日宣布与三大国家级权威机构正式签约，全面接入国家级核心数据。合作伙伴包括公安部道路交通安全研究中心、交通运输部路网监测与应急处置中心，以及中国气象局公共气象服务中心，此举旨在打造公路出行的全天候安全防护网。本次合作标志着官方权威系统数据与互联网导航平台的深度打通。交管部门掌握的封路信息、气象部门的分钟级天气变化，以及路网中心监测的高速拥堵态势，都将通过百度地图实时精准地送达正在路上的驾驶者手中。恶劣天气提前预警基于接入的国家级海量数据，用户在规划出行路线阶段，系统就会主动识别并标记沿途的雨雪、大雾、结冰等高影响区域。导航不仅会提前向用户提示潜在的出行风险，还会智能规划并提供安全的备选路线。在实际驾驶途中，系统还会结合车辆的实时位置进行动态研判。在车辆即将驶入恶劣天气覆盖路段之前，导航会提前发出语音预警，提醒驾驶员及时减速慢行或选择绕行。超视距风险精准识别除了应对复杂的气象变化，该版本还大幅升级了安全防护维度，推出了超视距风险预警功能。系统能够精细化识别前方事故、施工占道、急弯盲区等 30 余种动态和静态风险，相当于为车主配备了一位时刻在线的智能副驾。广大用户无需进行复杂的额外设置，只需将手机中的百度地图更新至最新版本。在日常出行中开启导航语音播报功能，即可自动享受到由国家级数据支持的全程气象与路况安全守护。

AI资讯AI新闻资讯正文降本打击！AI编程神器Cursor发布Composer 2.5，性能硬刚GPT-5.5发布于AI新闻资讯发布时间 :2026年5月19号 10:00阅读 :1分钟人工智能编程领域迎来重磅搅局者。AI代码编辑器开发商Cursor近日正式发布了其自研编程大模型的重大升级版本——Composer 2.5。该模型基于月之暗面开源的Kimi K2. 5 模型基座微调打造，在训练过程中，团队投入了比前代产品多出 25 倍的合成任务数据。为了最大化提升模型的编程逻辑，Cursor将高达85%的计算预算直接倾注于强化学习与扩展训练。这一激进的研发投入带来了极其显著的性能飞跃，使新模型在处理复杂代码任务时展现出媲美行业顶尖梯队的实力。高性价比挑战行业巨头在业界公认的SWE-Bench多语言测试和CursorBench v3. 1 评测中，Composer 2. 5 交出了79.8%和63.2%的优异成绩。这一表现已经能够平齐市面上最强大的Anthropic Opus 4. 7 以及OpenAI GPT-5. 5 模型。与强悍性能形成鲜明对比的是其极低的运行成本。Composer 2. 5 的每百万输入和输出代币价格分别仅为0. 50 美元和2. 50 美元，仅为竞争对手大模型研发商收费标准的极小一部分。万亿计算打造下一代重器对于追求极致效率的开发者，Cursor还同步推出了一款运行速度更快的变体版本。该版本在保持同等优秀性能的表现下，每百万输入和输出代币的价格也仅需3. 00 美元和15. 00 美元。这笔高性价比的AI编程服务目前已经全面在Cursor编辑器中上线供用户体验。不仅如此，Cursor已经联合SpaceX以及xAI，在拥有百万级H100 算力的Colossus- 2 集群上，以 10 倍的计算量从头训练下一代更具颠覆性的继承者模型。

AI资讯AI新闻资讯正文谷歌拉黑石「组局」新云公司， 50 亿美元砸向TPU集群，以后训练大模型或许真能更快更便宜发布于AI新闻资讯发布时间 :2026年5月19号 10:22阅读 :1分钟科技巨头与资本巨鳄的联手，正在重塑全球AI基础设施的竞争格局。据最新报道，Alphabet 旗下谷歌与全球资管巨头黑石（BlackRock）正筹划成立一家人工智能云计算公司，目标直指下一代算力基建的「制高点」。合作模式很清晰：黑石出资 50 亿美元作为股权资本，并持有新公司多数股权；谷歌则「出技术」——提供包括 TPU 专用芯片在内的硬件底座，以及配套的软件栈与云服务能力。这不是简单的「资本+技术」拼盘，而是一次精准的能力互补。黑石擅长长期资本配置与重资产运营，谷歌则握有全球领先的AI芯片架构与云原生经验。双方合力，意在打造一家专注高能效、大规模AI训练与推理的「垂直型云厂商」，避开与通用云巨头的正面交锋，转而深耕大模型时代的算力刚需。更值得关注的是其算力目标：新公司计划在 2027 年前建成 500 兆瓦（MW）的算力规模，并预留持续扩容空间。作为参照，当前全球头部云厂商的单区域算力集群多在百兆瓦量级，这一目标若如期达成，将直接跻身全球顶级智算中心阵营。对行业而言，这一动向释放了明确信号：当大模型进入「拼算力、拼能效、拼落地」的下半场，单纯的技术优势已不足以构建护城河，「资本耐力+工程效率+生态协同」的复合能力，正在成为新一轮竞争的核心变量。

AI资讯AI新闻资讯正文手机传文件、平板接屏幕、笔记本续任务？联想全系 AI 新品 5·19 发布，终于不用在设备间「反复横跳」了发布于AI新闻资讯发布时间 :2026年5月19号 10:25阅读 :1分钟联想的「AI 全家桶」终于要正式开箱了。5月19日，联想将举办新品发布会，一次性推出覆盖手机、折叠屏、平板、笔记本的全品类 AI 硬件矩阵。这次不是简单的产品堆叠，而是一次系统性亮剑:所有新品均原生适配天禧 AI4.0，首次实现「一体多端」的生态打通——跨设备文件互传、屏幕无缝接力、记忆实时同步等连贯体验，将真正从概念落地为日常。产品阵容堪称「全栈覆盖」:拯救者电竞手机，聚焦高性能游戏与即时交互;moto AI 大折叠，探索形态创新与智能分屏的融合;拯救者 AI 平板，强化创作与娱乐的场景协同;YOGA、拯救者、小新三条笔记本产品线，则分别锚定高端商务、硬核电竞、年轻潮流三大人群，按需定制 AI 能力与交互逻辑。更关键的是，这次发布标志着联想 AI 全生态的「拼图完成」。天禧 AI4.0不再只是单点功能，而是成为串联硬件、软件、服务的底层操作系统。用户在不同设备间切换时，任务流、数据流、记忆流都能自动延续，真正实现「人找服务」到「服务随人」的体验跃迁。对消费者而言，这意味着买的不只是单品，而是一套能持续进化的智能生活方案;对行业而言，联想正用「全品类原生适配」的策略，在大模型落地终端的竞争中，走出一条差异化路径。

AI资讯AI新闻资讯正文不卷参数卷落地！搜狐明确「理性应用」AI 战略，中型科技公司如何在巨头夹缝中走出差异化路径？发布于AI新闻资讯发布时间 :2026年5月19号 10:37阅读 :1分钟当大厂们还在「千亿参数」的军备赛里卷生卷死，搜狐选择了一条更「清醒」的路。2026 年搜狐科技年度论坛上，CEO 张朝阳明确亮出公司的 AI 战略底牌：不追大模型「第一阵营」的虚名，而是把有限资源押注在「能落地、有回报」的理性应用上。说白了，就是不做「造模型的人」，而是做「用模型用得最聪明的人」。具体怎么落地？搜狐的 AI 应用聚焦两个务实方向：✅ 提效降本：用头部厂商的成熟大模型，赋能内容生产、审核、分发等核心环节，让流程更轻、人效更高；✅ 内容中立：在算法推荐与智能生成中坚守客观立场，避免过度迎合流量，守住媒体平台的公信力底线。这一策略的本质，是中型科技公司在资源约束下的「精准卡位」——不硬拼算力与参数，而是深耕垂直场景，把通用能力转化为业务价值。比如新闻资讯的语义理解、视频内容的智能剪辑、用户互动的个性化响应，都是搜狐正在打磨的「小而美」切口。但务实不等于轻松。这条路径也面临双重挑战：一方面，依赖外部大模型可能带来技术迭代节奏受制于人、数据闭环难构建的风险；另一方面，在「人人都在讲 AI」的舆论场中，如何让用户感知到搜狐的差异化价值，而非沦为「又一家接 API 的内容平台」，是需要持续回答的命题。对行业而言，搜狐的选择或许提供了一种新参照：当大模型进入深水区，「会用」可能比「会造」更考验长期主义。

AI资讯AI新闻资讯正文谷歌联手黑石狂砸250亿:组建新型AI云挑战英伟达生态发布于AI新闻资讯发布时间 :2026年5月19号 10:52阅读 :1分钟据华尔街日报报道，全球私募股权巨头黑石（Blackstone）计划与谷歌(Google)合作，在美国组建一家全新的人工智能云算力公司，整体投资规模预计达250亿美元。黑石将率先注入50亿美元股权资本，其余通过金融杠杆筹集。新公司核心战略是依托谷歌自研TPU（张量处理器）芯片与云能力，直接挑战CoreWeave等独立算力独立云服务商。此举不仅意味着谷歌TPU正式大范围走向外部市场，也将加剧谷歌与英伟达在AI芯片领域的正面交锋。双方计划在2027年上线高达500兆瓦（MW）的算力容量(相当于中等城市用电量)，并已提前锁定和建设相关数据中心。作为全球最大数据中心提供方之一，黑石此举将进一步巩固其在AI物理基础设施领域的统治地位。

Claude Code是什么Claude Code 是 Anthropic 公司推出的基于命令行的AI 编程工具。Claude Code集成先进的 Claude Opus 4 语言模型，能理解自然语言指令，自动生成、修改和调试代码，并与 VS Code、JetBrains 等主流 IDE 深度集成。Claude Code支持 200K 超长上下文，能快速索引整个代码库并理解复杂依赖关系。Claude Code 具备文件操作、代码搜索、网页浏览和 Git 工作流管理等功能，显著提升开发效率。Claude Code帮助开发者快速生成高质量代码，能直接在开发环境中执行命令，避免频繁切换界面，极大地优化开发流程。最新版推出自动安全审查功能，在漏洞发布之前捕获漏洞。Claude Code 新功能 Opus 规划模式支持高效且智能的开发规划，用户能在复杂项目中借助 Opus 4.1 模型的强大分析能力制定高质量的开发计划，在执行阶段切换到 Sonnet 模型，实现高效、低成本的代码生成。Claude Code的主要功能代码库深度理解：基于智能搜索技术，快速理解整个代码库的结构和依赖关系，无需手动选择上下文文件。多文件协同修改：能够跨多个文件进行协调一致的代码修改，确保代码逻辑的完整性和一致性。代码生成与优化：专门针对代码理解和生成进行了优化，能够快速生成高质量的代码片段、测试用例和修复代码中的问题。终端集成：直接在开发者的终端中运行，无需切换工具，减少上下文切换的麻烦。与主流 IDE 集成：支持 VS Code 和 JetBrains 等主流开发工具，直接在 IDE 中提供代码建议和修改，无缝融入现有开发流程。工作流整合：与 GitHub、GitLab 等工具集成，支持从阅读问题、编写代码、运行测试到提交 PR 的全流程操作。智能编辑：基于对代码库和依赖关系的理解，进行强大的多文件编辑，确保代码修改的有效性和可靠性。用户控制与定制：不会在未经明确批准的情况下修改文件，适应开发者的编码标准和模式，并通过 SDK 或 GitHub Actions 进行定制化配置。终端运行的/security-review命令：支持用户在终端输入此命令，Claude Code将自动检查代码中的SQL注入、XSS等安全隐患，能直接修复发现的漏洞。GitHub Actions集成：在GitHub仓库配置相关文件后，每次提交新的PR时，自动触发安全审查，检测出漏洞后在代码行上添加内联注释，解释问题并给出修复建议。Opus 规划模式：在规划阶段调用强大的 Claude 4.1 Opus 模型制定高质量开发计划，在执行任务时使用高效且低成本的 Claude 4 Sonnet 模型，实现智能与效率的结合。Agent View：让开发者在一个终端界面内同时调度、监控和管理多个并行 AI 会话，无需多开窗口可实时查看各任务状态、快速切换介入或解除阻塞。如何使用Claude Code安装 Node.js：确保系统安装 Node.js 18 或更高版本。安装 CLI：在终端运行以下命令安装 Claude Code 的命令行工具：npminstall-g@anthropic/claude-cli登录和初始化：登录：运行claude login并按照提示输入 API 密钥（从 Anthropic 官网获取）。初始化：运行claude init，连接到代码库。开始使用：生成代码：用英语描述需求，例如：claude-p"Create a function to calculate the sum of two numbers in Python"修复错误：描述错误或粘贴错误消息，例如：claude-p"Fix this error: TypeError: Cannot read properties of null (reading ’length’)“自动化任务：运行命令解决繁琐任务，例如：claude-p"Resolve merge conflicts in the current branch"集成到开发环境：终端：直接在终端中使用。IDE：安装 VS Code 或 JetBrains 插件，实现在 IDE 中使用。Claude Code 的产品定价Pro 计划：年付 $17/月，月付 $20/月。Max 5x 计划：$100/人/月。Max 20x 计划：$200/人/月。Claude Code的应用场景代码理解：快速解释复杂代码的功能和逻辑，帮助开发者迅速理解项目结构。代码修改：自动修复代码中的错误或漏洞，快速定位问题并提供解决方案。代码测试：为代码自动生成单元测试或集成测试，确保代码的可靠性和稳定性。版本控制：支持自然语言完成 Git 提交、分支管理等操作，简化版本控制流程，让 Git 操作更加便捷高效。功能实现与扩展：快速实现新功能或扩展现有功能，加速开发进度，帮助开发者高效完成任务。©️版权声明：若无特殊声明，本站所有文章版权均归AI工具集原创和所有，未经许可，任何个人、媒体、网站、团体不得转载、抄袭或以其他方式复制发表本站内容，或在非我站所属的服务器上建立镜像。否则，我站将依法保留追究相关法律责任的权利。类似于Claude Code的工具BLACKBOX AI黑箱AI编程助理，快速代码生成CodeWhisperer亚马逊推出的免费AI编程助手AI Code ReviewerAI代码检查天工智码SkycodeAI智能编程助手，轻松生成各种代码GitHub CopilotGitHub推出的AI编程工具DevChat开源的支持多款大模型的AI编程助手6 条评论再想想发表评论wi游客此ide已放弃 ，国内无法使用，哎9个月前回复AI小集需要很强的魔法，还经常封号9个月前回复wi游客对。弄得很累，不如用亚马逊的ai ide，不用翻墙，也不封号9个月前回复3游客Claude 太强大了，可惜用不来2个月前回复Ai游客需要手机验证，不支持国内无862个月前回复AI小集是的，最好搞一个美国手机号2个月前回复

Kimi Code – Kimi推出的AI编程工具AI工具4个月前发布AI小集05Kimi Code是什么Kimi Code 是 Kimi推出的AI编程工具，可在终端（CLI）、VS Code 、Cursor、JetBrains 和 Zed 等主流编辑器中直接运行，基于Kimi K2.5多模态大模型支持图片和视频输入辅助编程。工具提供端到端代码能力，涵盖构建、调试、重构、测试及脚本编写。Kimi Code能完美兼容Claude Code、Roo Code等第三方 Agent，让 AI 编程能力无缝嵌入用户的每一次 commit。Kimi Code的主要功能全环境覆盖：提供Kimi CLI、VS Code 插件及第三方编辑器支持，可在终端和主流 IDE 中无缝运行编程任务。多模态交互：基于Kimi K2.5模型支持直接输入图片和视频进行编程辅助，实现视觉友好的代码理解与生成功能。端到端编程：完整覆盖代码构建、调试、重构、测试及脚本编写等开发全流程，自动迁移现有技能到新工作流。高频并发：每 5 小时支持 100-500 次请求额度，确保复杂项目开发过程中 AI 辅助的不间断可用性。第三方兼容：完美适配 Claude Code、Roo Code 等主流 Coding Agent，一键登录可完成设备授权与账号绑定。如何使用Kimi Code订阅会员：访问 Kimi Code 官网 https://www.kimi.com/code。安装和登录：CLI 用户：执行curl -L code.kimi.com/install.sh | bash安装，运行/login命令一键授权。VS Code 用户：安装 Kimi Code 插件，在终端直接运行登录命令完成设备绑定。配置使用官方 Agent：登录成功后即可直接使用，系统自动完成账户绑定。第三方工具（Claude Code / Roo Code）：在控制台创建 API Key，复制后配置到对应工具的环境变量或设置中即可开始编程。Kimi Code的产品定价Andante 计划：定价 ¥4.99（7 天试用），试用期结束后以 ¥49/月自动续费，提供专属 Kimi Code 使用额度、旗舰模型抢先体验以及支持多个编程会话的基础功能。Moderato 计划：定价 ¥99/月，提供每周更新的使用额度，多种设备登录并分享套餐额度，支持用户在多个项目中高效协作。Allegretto 计划：定价 ¥199/月，提供充足的每周额度和更高的并发上限，为高级用户提供更强大的使用体验和超值选择。Kimi Code的应用场景全栈开发任务：覆盖代码构建、调试、重构、测试及脚本编写等端到端编程工作流，满足从原型开发到生产部署的完整需求。多模态辅助编程：支持直接输入 UI 设计图、视频演示等多媒体内容，基于视觉信息生成对应代码实现或进行问题诊断。跨平台开发环境：可在终端 CLI、VS Code、Cursor、JetBrains、Zed 等主流编辑器中无缝切换，适配不同开发者的工具链偏好。高频敏捷开发：借助 Turbo 级加速（100 Tokens/s）和每 5 小时数百次的并发额度，支撑大型项目中的高频代码生成与迭代需求。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇inSai Hilight - AI营销视频Agent平台，多Agent深度协同下一篇Lucy 2 - Decart AI推出的实时视频生成模型相关文章Duck.ai – DuckDuckGo 搜索引擎推出的免费 AI 聊天机器人AI小集2MiniMax M2.5 – MiniMax开源的新一代旗舰AI模型AI小集3R1-Onevision – 开源多模态视觉推理模型，基于 Qwen2.5-VL 微调AI小集2Clawra – 基于OpenClaw框架开源的AI虚拟女友AI小集4像素猫AI – AI绘画应用，支持生成吉卜力风格AI小集3Sonic-3 – Cartesia推出的实时语音对话模型AI小集4暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付19小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型20小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手20小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台24小时前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前General365 – 美团 LongCat 团队开源的通用推理评测基准4天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具4天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件4天前

Kimi WebBridge – 月之暗面推出的浏览器扩展插件AI工具4天前发布AI小集03Kimi WebBridge是什么Kimi WebBridge 是月之暗面推出的浏览器扩展插件，面向Kimi Code、Claude Code、Cursor、Codex等本地 AI Agent。Kimi WebBridge 通过本地桥接服务与 Chrome DevTools Protocol，让 AI 携带用户真实登录态和 Cookie，像真人一样在浏览器中执行点击、滑动、输入、填表、信息提取及跨站点内容整合等操作。所有任务在本地完成，网页数据不会离开设备，保障隐私安全，同时实现高效、自动化的网页操作。Kimi WebBridge的主要功能浏览器自动化操作：支持 AI Agent 执行网页导航、点击按钮、填写表单、滑动页面等操作。身份状态继承：AI 携带用户真实登录态、Cookie 和账号权限操作网站，无需额外配置。信息提取与整合：跨站点抓取内容并整合，支持将数据写入在线文档或本地目录。可视化任务追踪：运行时标记正在操作的浏览器标签页，不占用鼠标键盘，用户可并行使用电脑。进阶 CLI 工具：支持将固定流程打包为不消耗大模型 Token 的专用 CLI 工具，用于重复任务。Kimi WebBridge的技术原理双组件协同架构：系统由本地桥接服务与浏览器扩展两部分组成。AI Agent 将自然语言指令发送至本地桥接服务，后者通过标准接口与 Chrome 或 Edge 浏览器内的扩展通信，扩展再基于底层协议驱动浏览器完成具体操作，并将执行结果回传给 Agent，形成完整的指令闭环。基于 Chrome DevTools Protocol：浏览器扩展底层采用 Chrome DevTools Protocol（CDP），是 Chrome 官方提供的调试协议。通过协议，扩展能在用户当前使用的真实浏览器实例中执行导航、元素点击、表单输入、页面截图、DOM 读取与修改等操作，无需启动独立或虚拟浏览器环境。本地身份继承机制：Kimi WebBridge 直接复用用户已打开的浏览器会话，继承真实的登录态、Cookie 和本地存储数据。AI 在操作目标网站时，等同于用户本人在操作，因此无需额外配置账号、处理验证码或突破登录壁垒，天然支持各类需要身份认证的后台系统。端侧隐私保障：所有指令解析、页面交互与数据提取均在用户本地设备完成。桥接服务只负责本地进程间通信，网页内容、Cookie 及操作结果不会上传至云端服务器，确保敏感业务数据不出本地设备。如何使用Kimi WebBridge安装浏览器插件：在 Chrome 或 Edge 浏览器应用商店搜索并安装Kimi WebBridge扩展插件。安装本地守护程序：向本地 AI Agent 发送指定 curl 命令，安装 Kimi WebBridge 的本地守护程序与 Skill。重启 Agent 加载技能：重启 Kimi Code 等本地 AI Agent，使其识别并加载新安装的 WebBridge 技能。验证安装结果：输入测试指令「使用 kimi-webbridge 帮我打开 kimi.com」，若成功打开标签页并出现 agent:kimi 提示即代表安装成功。Kimi Claw Desktop 用户：若通过 Kimi 桌面版在本地部署了 Kimi Claw Desktop，无需安装守护程序，直接发送指令可开始使用。Kimi WebBridge的核心优势原生身份继承：直接复用用户现有登录态，无需模拟账号或处理验证码，降低复杂网站接入门槛。本地隐私安全：全部操作在本地浏览器执行，敏感数据与网页内容不上传云端。多 Agent 兼容：统一支持 Kimi Code、Claude Code、Cursor、Codex、Hermes、OpenClaw 等主流本地 Agent。非侵入式运行：后台标记操作标签页，不抢占鼠标键盘，人机可协同工作。低 Token 消耗：固定流程可编译为专用 CLI 工具，重复执行时无需反复调用大模型。Kimi WebBridge的项目地址项目官网：https://www.kimi.com/zh-cn/features/webbridgeKimi WebBridge的同类竞品对比维度Kimi WebBridgeBrowserbaseAgent-Browser（Vercel）运行方式本地浏览器扩展 + 桥接服务云端托管浏览器云端/本地浏览器自动化身份继承直接复用用户 Cookie 和登录态需单独配置环境或账号需单独配置环境或账号数据隐私完全本地执行，数据不出设备数据经过云端服务器取决于部署方式支持 AgentKimi Code、Claude Code、Cursor 等主要面向开发者 API主要面向开发者 API使用门槛安装插件 + 一行命令即可需 API 集成开发需一定技术配置人机协同后台运行，不占用键鼠云端运行，无本地交互云端运行，无本地交互Kimi WebBridge的应用场景自动化信息整理：自动采集多个笔记类 APP 的应用商店宣传素材，让 AI 分析设计风格差异并写入在线文档。复杂操作自动化：一键接管量化研究模拟平台，自动抓取金融数据、执行多策略迭代回测并撰写量化报告。重复性网页任务：自动化完成基于网页的重复性繁琐操作或定时任务，如批量填表、点击、数据录入与状态更新。跨站点内容整合：在已登录状态下跨多个网站提取信息并整合，实现内容迁移、数据汇总与多源信息同步。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Mavis - MiniMax Agent 推出的多 Agent 协作模式下一篇TencentDB Agent Memory - 腾讯开源的智能体记忆管理工具相关文章checksub – AI视频字幕处理工具，自动生成200多种语言字幕AI小集2GR00T-Teleop – 英伟达发布通过 Apple Vision Pro 捕捉人类动作的技术AI小集2BlockDance – 复旦联合字节推出的扩散模型加速方法AI小集3easypic – 免费 AI 抠图工具，高效识别分离图像主体与背景AI小集2Bobby – RockFlow 推出的金融投资 AI AgentAI小集5LongReward – 清华、中科院、智谱AI联合推出提升长文本大语言模型性能的方法AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付19小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型20小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手20小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台23小时前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前General365 – 美团 LongCat 团队开源的通用推理评测基准4天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具4天前Mavis – MiniMax Agent 推出的多 Agent 协作模式5天前

了解更多 kimi.com/zh-cn/features/webbridge

Kimi WebBridgeÔºökimi.com/zh-cn/features/webbridge

Kimi K2.6 发布并开源，全面精进代码和 Agent 集群能力

Kimi K2.6「Agent 集群」现已支持 300 个成员并行协作

和 Kimi 一起投身 AGI，穿越成长周期

服务全球1/5网站的Cloudflare选择Kimi K2.5，降低77%成本

Kimi 杨植麟「2026 中关村论坛」演讲全文（附视频）

Kimi API：用90%缓存命中率，把价格打到25%

你可以来 Kimi 使用 OpenClaw 了

技术报告：Kimi K2.5 如何实现文本和视觉能力互相增强？

不只PPT，Kimi K2.5 Agent可以帮你做Excel、Word和PDF了

Kimi 发布并开源 K2.5 模型，带来全新视觉理解、代码和 Agent 集群能力

微信扫一扫关注该公众号

微信扫一扫可打开此内容，使用完整服务

吐司APP的简介吐司是腾讯推出的AI原生应用创作平台，用户只需输入想法，AI自动生成完整的安卓原生应用。平台支持智能功能梳理、精细化编辑打磨，能一键导出APK进行本地安装与分享。内置灵感广场汇聚海量优质模板，涵盖社交助手、照片精选、学习工具等多元场景，支持用户共创。无需编程基础，零基础用户也能快速将奇思妙想转化为可实际使用的专属APP，轻松实现个性化应用开发。吐司APP的主要功能AI自动生成应用：输入自然语言想法，AI自动梳理功能逻辑并生成完整的安卓原生应用，无需编写代码。精细化编辑打磨：支持对生成的应用进行细节调整与优化，满足个性化需求。一键导出APK：生成完成后可直接导出APK安装包，支持本地安装、测试与社交分享。灵感广场：内置海量精选应用模板与创意案例，如社交助手、照片精选、学习工具等，帮助用户快速获取创作灵感。应用共创：支持用户基于现有模板进行二次创作与共创，降低应用开发门槛。多场景覆盖：支持生成考研题库、运势工具、普通话测试、宠物健康日志、梦境记录、食材新鲜度检测、万年历等多样化实用应用。类似于吐司APP的应用Replit APPAI开发工具应用YouWare APPAI移动应用开发工具CursorAnysphere公司推出的AI代码编辑器TRAE SOLO APP字节跳动推出移动端的 AI 智能工作助手Trae字节跳动推出的免费 AI IDEWindsurfCodeium公司推出的AI编程工具暂无评论再想想发表评论暂无评论…

LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台AI工具24小时前发布AI小集03LibTV团队版是什么LibTV 团队版是LibTV推出的专业级 AI 视频协作工作台，专为 AI 短剧、电影、TVC 制作团队设计。平台提供团队空间、实时画布协作、资产库沉淀、权限管理、成本管控等功能，解决 AI 视频创作中资产散落、协作低效、人员流动导致资产流失等痛点，支持规模化、工业化的 AI 内容生产。LibTV团队版的主要功能专属团队空间：统一沉淀分镜、脚本、镜头资产与生成记录，历史项目可迁移，避免核心资产散落在个人账号与本地文件中。画布协作：导演、美术、剪辑等角色可在同一块画布实时协同，同步节点状态与镜头修改，减少版本错乱。团队资产库：角色主体、风格模板、运镜方案、提示词工作流与参考素材支持统一沉淀，成熟视觉体系可持续复用。权限管理：支持项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作场景。成本管理：共享积分池、成员额度分配与消耗查看，团队管理者可统一管理生成资源、席位与项目成本。资产交接：成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免创作资产随人员流动中断。如何使用LibTV团队版创建团队空间：访问LibTV官网升级至团队版，建立专属团队项目空间，设置团队名称与基础信息。配置成员与权限：邀请导演、编剧、美术、剪辑等成员加入，按岗位分配访问、编辑、复制或发布权限。沉淀与迁移资产：将历史个人项目迁移至团队空间，或新建项目并导入参考素材、角色设定与风格模板。画布协作创作：成员在统一画布内协同推进分镜设计、镜头生成与修改，实时同步进度与反馈。复用团队资产库：调用已沉淀的角色主体、风格模板与提示词工作流，保持系列化内容风格一致。管控成本与交接：通过共享积分池分配生成额度，监控消耗；成员变动时资产自动保留，确保项目连续性。LibTV团队版的核心优势工业化协作：从个人创作升级为团队制片管理，支持长期项目、多岗位协同与规模化内容生产。资产持续复用：成熟的角色设定、风格与提示词工作流可沉淀为团队资产，降低新人上手与重复调试成本。实时协同提效：多人画布协作减少版本错乱与重复生成，围绕同一套镜头语言与素材体系推进项目。成本与风险可控：共享积分池与席位管理适配多项目并行需求，权限分级与资产交接机制保障核心创作安全。LibTV团队版的同类竞品对比对比维度LibTV 团队版快手可灵（企业版/团队功能）协作模式多人实时画布协作，导演/美术/剪辑等角色可在同一块画布内同步节点状态、生成进度与镜头修改团队空间内创作接力，支持桌面端（Win/Mac）同步，减少文件下载与搬运损耗资产沉淀角色主体、风格模板、运镜方案、提示词工作流、参考素材与工具箱统一沉淀，支持成熟视觉体系持续复用设立「共享资产库」，支持个人空间向团队空间一键转移灵感值、创作资产与主体资产，实现”一人贡献，全员复用”权限管理项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作所有者、管理员、协作者三重角色，灵活配置查看/编辑/移除权限，保障核心资产安全成本管理共享积分池、成员额度分配与积分消耗查看，支持多项目并行下的生成资源统筹支持灵感值跨空间转移，团队共享生成资源，但精细化成本管控功能相对较弱资产交接成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免资产随人员流动中断个人资产可一键导入团队空间，但成员退出后的自动资产交接机制未明确强调核心场景聚焦 AI 短剧、AI 电影、AI TVC 的工业化制片流程，强调长期项目管理与跨岗位协同覆盖 AI 视频生成、智能分镜、广告 Demo、影视特效等，强调快速创作与素材复用产品定位AI 视频专业级协作工作台，面向制片管理流程AI 视频生成平台的团队功能扩展，面向创作者联盟与工作室LibTV团队版的应用场景AI 视频工作室：适合长期运营 AI 短剧、动画、MV 的创作团队，通过资产库统一沉淀角色与风格，实现系列化生产。短剧承制方与 MCN：适配高频、多项目并行的工业化内容生产，导演与美术可同时在线推进分镜与镜头生成。品牌市场部与代理公司：支持品牌方、代理公司与外部创意团队跨组织协作，共同管理 AI TVC 与 Campaign 项目。AI 电影与大型制作团队：支持长篇叙事、系列化 IP 与复杂制作流程，导演组、制片组、美术组与后期团队跨岗位协同。企业内容与创意部门：适配培训、营销、产品与全球化传播等标准化内容生产，便于企业进行长期生成资源规划。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇PPT Master - 开源 AI PPT 生成标准化工作流下一篇腾讯Marvis - 腾讯应用宝推出的操作系统层个人 AI 助手相关文章Zeemo – AI视频字幕生成工具，支持自动生成95+种语言，翻译113种语言AI小集4Sparkify – 谷歌推出的AI动画视频生成工具AI小集3Gemma 3n – 谷歌推出的端侧多模态AI模型AI小集2知文AI学术助手 – AI学术论文写作工具，4步10分钟生成万字论文AI小集4TextIn – AI文字识别平台，覆盖多种类型的富文本信息提取AI小集3Wysa – AI心理健康平台，提供心理疏导和情绪管理服务AI小集2暂无评论再想想发表评论暂无评论…

Grok Build – xAI 推出的终端原生 AI 编程智能体AI工具2天前发布AI小集03Grok Build是什么Grok Build 是 xAI 推出的终端原生 AI 编程智能体（Agentic CLI），面向专业软件工程与复杂编码任务。基于 Grok 4.3 beta 的 200 万 token 超长上下文，支持 Plan Mode 计划审批、最多 8 个并行子智能体协作及 Arena Mode 自动方案排序，可在本地直接读写文件、执行 Shell 命令、操作 Git 并构建完整应用。采用本地优先架构，源代码与敏感数据不上传云端，保障企业级安全。Grok Build的主要功能Plan Mode（计划模式）：复杂任务执行前自动生成结构化步骤计划，用户可审批、评论或修改，所有代码变更以 diff 形式透明展示，避免”黑箱”操作。并行子智能体：最多支持8 个并发 AI 智能体同时工作，可分别处理代码分析、CI/CD 检查、性能优化、文档检索等不同子任务。Arena Mode（竞技场模式）：多个智能体针对同一任务生成不同代码方案，系统自动评估并排序，开发者直接选择最优解，无需手动对比。超长上下文窗口：基于 Grok 4.3 beta，支持200 万 token上下文，可一次性加载整个大型代码库进行跨文件复杂任务。本地优先架构：源代码、凭证和敏感数据不上传云端，完全在本地执行，支持离线环境和气隙网络，保障企业级代码安全。终端原生 CLI + Web 双入口：支持本地命令行交互，也可通过浏览器界面远程操作，满足不同工作流习惯。VS Code 扩展集成：提供官方 VS Code 插件，在 IDE 内直接调用 Grok Build 能力，实现 GUI adjacent 工作流。Git 与开发服务器集成：可自动管理 Git 仓库、启动本地开发服务器、内置浏览器浏览网页、执行 Shell 命令。插件与开放生态：支持 MCP 服务器、AGENTS.md 团队规范文件、Hooks 自定义动作、ACP（Agent Client Protocol）协议，便于集成现有 DevOps 流水线。Headless 脚本模式：通过-p标志支持非交互式执行，可直接嵌入 CI/CD 流水线或自动化脚本。BYOK（自带 API Key）：支持使用自有 xAI API Key 进行本地计费，灵活控制成本。如何使用Grok Build一键安装 CLI：在终端执行curl -fsSL https://x.ai/cli/install.sh | bash完成本地安装，支持 macOS、Linux 和 Windows 三大平台。启动交互式终端：安装完成后输入grok>进入交互式命令行界面，可直接用自然语言描述编程需求。登录账号：在终端内执行grok> /login登录 xAI 账号，解锁高级功能与订阅权益。配置 API Key（可选）：如需使用自有额度，执行grok> /api-key add xai添加个人 xAI API Key（BYOK 模式）。使用 Plan Mode 处理复杂任务：输入需求后，Grok Build 会自动生成结构化执行计划，展示步骤与预期变更 diff，用户可审批、评论或修改后再执行。利用并行子智能体：在复杂项目中，系统会自动或手动触发最多 8 个并发子智能体，分别处理代码分析、测试生成、文档检索等子任务。通过 Arena Mode 选优：对同一任务让多个智能体生成不同方案，系统自动评估排序，你只需选择排名最高的方案应用。Web 端远程操作：访问build.grok.com或x.ai/cli，在浏览器中打开 Grok Build 界面，无需本地安装即可使用。VS Code 集成工作流：安装 Grok Build 的 VS Code 扩展，在编辑器内直接调用智能体能力，边写代码边让 AI 协助修改。Headless 自动化执行：在脚本或 CI/CD 流水线中使用-p标志运行非交互式命令，实现自动化构建与部署。查看帮助与命令：在终端输入grok> /help查看所有内置命令与快捷操作指南。Grok Build的官网地址项目官网：https://x.ai/news/grok-build-cliGrok Build的核心优势本地优先，代码绝不上云：源代码、凭证及敏感数据完全在本地执行，无需上传至 xAI 服务器，适合处理专有代码库、金融及政务等受监管行业项目。200 万 token 超长上下文：基于 Grok 4.3 beta，可一次性将整个大型代码库载入内存，跨数十个文件进行复杂重构与多文件联动修改，大幅减少上下文丢失。8 路并行子智能体：最多同时调度 8 个并发 AI 智能体分工协作，分别处理代码分析、测试生成、CI/CD 检查等子任务，显著缩短复杂项目耗时。Arena Mode 自动选优：同一任务由多个智能体生成不同方案，系统自动评估排序并给出排名，开发者直接选择最优解，告别手动对比试错。Plan Mode 透明可控：执行前先输出结构化步骤计划与 diff 预览，用户可审批、评论或修改后再执行，每一步变更清晰可见，避免”黑箱”风险。极致 API 价格竞争力：输入 token 仅$0.20 / 百万 tokens，相较 Claude Code 和 Codex CLI 大幅降低高频使用成本。终端原生 + 多入口覆盖：支持 CLI 命令行、Web 浏览器、VS Code 扩展三种工作入口，既满足极客式终端操作，也兼容 IDE 集成习惯。Grok Build的同类竞品对比对比维度Grok Build (xAI)Claude Code(Anthropic)Codex CLI(OpenAI)产品定位终端原生 AI 编程智能体（Agentic CLI）终端 AI 编程助手云端 AI 编程智能体发布状态2026.5 早期 Beta已正式发布已正式发布底层模型Grok 4.3 beta / grok-code-fast-1Claude 4 (Sonnet/Opus)Codex (基于 GPT-4o/o3)上下文长度200 万 token20 万 token (Claude 4)约 20 万 token并行智能体最多 8 个并发子智能体单智能体为主单智能体为主方案评估Arena Mode 自动排序选优手动对比或迭代手动对比或迭代执行计划Plan Mode（先计划后审批执行）直接执行，可审查 diff直接执行，可审查 diff数据安全本地优先，代码不上云云端执行，代码上传 Anthropic云端执行，代码上传 OpenAI离线支持支持完全离线/气隙网络需联网需联网SWE-Bench70.8%约 72–75%约 70%API 定价$0.20 / 百万 tokensPro 订阅 + 按量计费Credits 按量计费订阅门槛SuperGrok Heavy ($300/月)Claude Pro / Team 订阅ChatGPT Pro / 开发者额度入口形态CLI + Web + VS Code 扩展CLI + IDE 集成CLI + ChatGPT 移动端自动化集成Headless 模式 (-p 标志) + CI/CD有限脚本支持有限脚本支持核心差异多智能体并行 + 本地优先 + 超大上下文模型推理能力强，生态成熟与 ChatGPT 生态深度整合Grok Build的应用场景大型代码库重构与现代化：基于 200 万 token 超长上下文一次性加载整个项目，跨数十个文件进行安全重构、依赖升级或架构迁移，避免上下文截断导致的遗漏。复杂全栈应用从零构建：通过 Plan Mode 自动生成项目结构、技术选型、依赖安装和文件生成计划，审批后直接执行，快速搭建完整应用原型。敏感行业专有代码开发：本地优先架构确保源代码、凭证和敏感数据不上传云端，适用于金融、政务、医疗、军工等对数据主权有严格要求的场景。离线/气隙网络环境编程：无需持续联网即可在本地执行全部任务，适配高安全隔离网络、内网开发机或无外网访问权限的工作环境。多维度自动化代码审查：启动多个并行子智能体分别检查代码安全性、性能瓶颈、CI/CD 合规性和文档完整性，一次性输出综合审查报告。技术方案快速对比与选型：使用 Arena Mode 针对同一需求生成多种实现方案，系统自动评估排序，帮助团队在短时间内确定最优技术路径。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇OpenHuman - 开源桌面端 AI 助手，能主动感知工作上下文下一篇Violin - 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具相关文章PaliGemma 2 – 谷歌DeepMind推出的全新视觉语言模型AI小集2易健EasyHealth – 北大健康推出的 AI 健康科普应用AI小集4Command A – Cohere 推出的生成式 AI 模型AI小集2UniReal – 港大联合 Adobe 推出的通用图像生成和编辑框架AI小集2Claude Opus 4.1 – Anthropic推出的最新编程模型AI小集5Ling-V2 – 蚂蚁百灵推出的大型语言模型系列AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付19小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型20小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手20小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台23小时前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文3天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前General365 – 美团 LongCat 团队开源的通用推理评测基准4天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具4天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件4天前Mavis – MiniMax Agent 推出的多 Agent 协作模式5天前

OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文AI工具3天前发布AI小集05OpenHuman是什么OpenHuman 是 tinyhumansai 团队推出的开源个人 AI 超级智能助手，定位为”Your Personal AI super intelligence”，主打私有、简单且极其强大。非传统聊天机器人，是能主动感知用户工作上下文的桌面级 AI Agent，通过每 20 分钟自动同步 118+ 第三方应用数据，构建本地持久记忆，让 AI 真正”懂你”。项目采用 Rust + TypeScript + Tauri 构建，支持 macOS、Linux 和 Windows 平台。OpenHuman的主要功能桌面吉祥物交互：具备视觉形象的桌面 AI 助手，支持语音输入（STT）与 ElevenLabs 语音输出（TTS），吉祥物可实时唇形同步，能作为参会者加入 Google Meet 并实时记录会议。118+ 第三方应用一键集成：通过 OAuth 快速接入 Gmail、Notion、GitHub、Slack、Linear、Jira、Stripe、Calendar、Drive 等主流工具，无需编写任何连接代码。每 20 分钟自动同步（Auto-fetch）：核心引擎每 20 分钟遍历所有活跃连接，自动拉取最新数据进入本地记忆系统，让 Agent 在次日早上已掌握前日全部上下文。Memory Tree 层次化持久记忆：将同步数据规范化为 ≤3k-token 的 Markdown chunks，评分后折叠存入层级摘要树，存储于本地 SQLite，实现跨会话的长期记忆。Obsidian Vault 原生兼容：同步数据同时输出为 Obsidian 可读的.md文件，用户可在 Obsidian Graph View 中浏览和编辑 AI 维护的知识网络。TokenJuice 智能 Token 压缩：在数据进入 LLM 前自动压缩（HTML 转 Markdown、缩短长 URL、移除非 ASCII 字符等），最高可降低 80% 的 API 调用成本与延迟。内置模型路由（Model Routing）：根据任务类型自动分配轻量模型、强推理模型或视觉模型，一个订阅覆盖多种模型能力，无需管理多个 API Key。完整工具链开箱即用：内置 Web 搜索、网页抓取（Web Scraper）、代码工具集（文件系统、Git、Lint、Test、Grep），无需额外安装插件即可使用。本地优先与隐私保护：所有工作流数据本地加密存储，支持通过 Ollama 接入本地模型实现完全离线运行，数据不上传云端。可选 agentmemory 共享后端：可配置与 Claude Code、Cursor、Codex 等工具共享同一个持久记忆存储，实现跨工具的知识互通。消息通道集成：支持 inbound/outbound 消息收发，打通用户已有的沟通渠道，工作流数据始终保存在本地设备。如何使用OpenHuman官网下载安装：访问OpenHuman的官网 https://tinyhumans.ai/openhuman 下载对应平台的 DMG（macOS）或 EXE（Windows）安装包，双击安装后即可启动。命令行快速安装（macOS / Linux x64）：在终端执行curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash自动完成下载与安装。命令行快速安装（Windows）：在 PowerShell 执行irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex完成一键安装。一键连接工作应用：打开应用后，通过 OAuth 一键接入 Gmail、Notion、GitHub、Slack、Calendar、Drive 等 118+ 第三方工具，授权后每 20 分钟自动同步数据。与桌面吉祥物交互：点击桌面上的吉祥物形象，通过语音或文字与其对话，可勾选 “Speak replies” 让 AI 语音播报回复。查看自动同步的记忆数据：在应用内查看 Agent 已同步的邮件、任务、代码仓库等上下文，AI 会自动压缩整理为 Memory Tree 供后续对话调用。Obsidian 知识库浏览：同步数据自动写入本地 Obsidian-compatible Vault，打开 Obsidian 即可在 Graph View 中可视化浏览 AI 维护的知识网络，也可手动编辑.md文件。使用内置工具集：直接调用开箱即用的 Web 搜索、网页抓取、代码工具（文件系统、Git、Lint、Test、Grep）等能力，无需额外安装插件。接入本地模型（可选）：如需完全离线运行，通过 Ollama 在本地部署模型，并在设置中切换至本地模型后端，实现零云端依赖。OpenHuman的项目地址官网地址：https://tinyhumans.ai/openhumanGitHub 仓库：https://github.com/tinyhumansai/openhumanOpenHuman的核心优势UI 优先分钟级上手：干净的桌面体验，无需终端配置，从安装到可用 Agent 仅需几分钟，区别于终端优先的竞品。上下文在分钟级而非周级：传统 Agent 需要数周学习用户习惯，OpenHuman 通过一次同步即可掌握邮箱、日历、仓库、文档、消息的全量压缩上下文。一站式订阅，零 API 蔓延：一个订阅覆盖所有功能，无需额外购买多个模型 API Key。Obsidian 知识库原生兼容：同步数据自动写入 Obsidian-compatible Vault，用户可在熟悉笔记界面中浏览和编辑 AI 维护的知识图谱。Rust + Tauri 高性能架构：内存占用远低于 Electron 方案，69.7% Rust + 26.1% TypeScript 构建。开源可审计：GPL-3.0 协议，代码完全透明，隐私敏感用户可自行审计数据流向。OpenHuman的同类竞品对比对比维度OpenHumanOpenClawHermes Agent开源协议✅ GPL-3.0✅ MIT✅ MIT上手难度✅ UI 优先，分钟级上手⚠️ 终端优先配置⚠️ 终端优先配置使用成本✅ 单一订阅 + TokenJuice 压缩⚠️ 自备模型 API⚠️ 自备模型 API记忆能力🚀 Memory Tree + Obsidian Vault + 可选 agentmemory 后端⚠️ 依赖插件扩展✅ 自学习观察第三方集成🚀 118+ 应用 OAuth 一键接入⚠️ 需自行配置连接⚠️ 需自行配置连接自动数据同步✅ 每 20 分钟自动 fetch 至本地记忆🚫 无自动同步🚫 无自动同步API 管理✅ 一个账户覆盖所有模型🚫 自备 Key🚫 多供应商管理模型路由✅ 内置自动分配（轻量/推理/视觉）⚠️ 手动切换⚠️ 手动切换原生工具集✅ 代码 + 搜索 + 抓取 + 语音✅ 代码工具✅ 代码工具数据隐私🚀 本地优先，数据不上云依赖自托管配置依赖自托管配置OpenHuman的应用场景跨应用项目上下文整合：同时连接 GitHub、Linear、Jira、Slack 后，让 AI 自动对比代码 PR、任务看板和团队消息，一键汇总项目最新进展，无需手动切换多个工具。智能邮件筛选与摘要：授权 Gmail 后，AI 每 20 分钟自动扫描收件箱，按项目或优先级分类摘要，快速定位需要回复的重要邮件，过滤掉低价值信息。AI 会议助手：让桌面吉祥物以真实参会者身份加入 Google Meet，实时记录会议内容、提取 action items，并基于记忆库提供参会者背景信息辅助沟通。代码仓库智能问答：同步 GitHub 仓库数据后，直接向 AI 询问代码逻辑、历史变更原因、PR 评论争议点，Agent 基于完整代码上下文给出精准回答。Obsidian 个人知识库构建：将邮件、文档、聊天记录自动压缩归档为 Obsidian-compatible 的.md文件，在 Graph View 中可视化呈现知识关联网络，打造 Karpathy 风格的 LLM 知识库。多工具工作流自动化：连接 Notion、Calendar、Drive、Stripe 等工具后，AI 自动关联”日历事件→相关文档→对应邮件→财务记录”，形成完整的项目上下文链。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Intern-S2-Preview - 上海 AI Lab 开源的科学多模态大模型下一篇Grok Build - xAI 推出的终端原生 AI 编程智能体相关文章HumanDiT – 浙大联合字节推出的姿态引导人体视频生成框架AI小集2Claude Sonnet 4.6 – Anthropic推出的最新一代AI模型AI小集3NitroGen – 英伟达联合斯坦福大学等推出的通用游戏AI模型AI小集3CodeMender – DeepMind推出的代码修复AI AgentAI小集2presentation-ai – 开源AI PPT生成工具，自动智能配图AI小集2Learn Your Way – 谷歌推出的实验性AI学习工具AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付19小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型20小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手20小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台23小时前PPT Master – 开源 AI PPT 生成标准化工作流1天前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架1天前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具2天前Grok Build – xAI 推出的终端原生 AI 编程智能体2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型3天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台3天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目3天前General365 – 美团 LongCat 团队开源的通用推理评测基准4天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具4天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件4天前Mavis – MiniMax Agent 推出的多 Agent 协作模式5天前

Marta MussoScienceMay 17, 2026 5:30 AMThe First Atomic Bomb Test in 1945 Created an Entirely New MaterialThe discovery from the Trinity nuclear test site shows how extreme conditions can result in materials never before seen in nature or in the lab.The new material is a clathrate made of calcium, copper, and silicon .Luca Bindi/Università di FirenzeCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyDuring the Trinitynuclear teston July 16, 1945, in the New Mexico desert—the world’s very first test of anatomic bomb—a new material spontaneously formed. It wasdiscoveredonly recently, by an international research team coordinated by geologist Luca Bindi at the University of Florence, which identified the novel clathrate based on calcium, copper, and silicon. It’s a material never before observed either in nature or as an artificial compound created in the laboratory.What Are Clathrates?The term “clathrates” denotes materials characterized by a “cage-like” structure that traps other atoms and molecules inside, giving them unique properties. Of great technological interest, these materials are being studied for various applications ranging from energy conversion (as thermoelectric materials capable of transforming heat into electricity) to the development of new semiconductors, to gas storage and hydrogen for future energy technologies.The New MaterialTo discover the new material, researchers focused on trinitite, a silicate glass containing rare metallic phases. Using some techniques like x-ray diffraction, the team was able to identify a type I clathrate based on calcium, copper, and silicon within a tiny copper-rich metal droplet embedded in a sample of red trinitite.The new material, the researchers say, formed spontaneously during a nuclear explosion. This indicates that the extreme conditions, such as extremely high temperatures and pressures, can generate new materials that are impossible to obtain by traditional methods.Natural LaboratoriesThe discovery is even more interesting because in the same detonation event another very rare material was formed: a silicon-rich quasicrystal, already documented by the team of experts led by Bindi a few years ago.A quasicrystal, as BinditoldWIRED at the time, is something that is not a crystal, but looks a lot like one. “Their peculiarity,” he said, “is that the atomic arrangement that is not periodic, but nearly so, creates incredible symmetries from which derive amazing physical properties, among other things, very difficult to predict.”Establishing the link between these structures therefore helps scientists better understand how atoms organize under extreme conditions and expand the possibilities for designing new materials. “Events such as nuclear explosions, lightning strikes, or meteoritic impacts function as true natural laboratories,” the researchers explain. “They allow us to observe forms of matter that we cannot easily reproduce in the laboratory.”In essence, this research opens new vistas for the development of innovative technologies, demonstrating that even destructive events can bequeath discoveries useful for the future.This story originally appeared inWIRED Italiaand has been translated from Italian.

Ben DowsettGearMay 17, 2026 6:00 AMIf You’re a Serious Bowler, You Need to Know About Bowling Lane OilBowling centers apply oil to their lanes using a machine that works like a giant inkjet printer. The pattern in which it’s applied can change everything about how the ball travels.Photograph: Charly Triballeau/AFP/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyEach time EJTackett steps to his lane to throw a shot in the upcomingProfessional Bowlers Association World Championships, he’ll be thinking about more than just ball spin or angle.Tackett has won the last three world titles, and there’s a convincing case that he’s currently thebest bowler on Earth.While his ability tothrow a consistent ballis certainly key to his success, there’s another factor he and his fellow PBA pros are paying closer attention to during matches: lane oil patterns.Most of us amateurs have never thought about the thin layer of oil that coats the surface of the bowling lane—unless someone at our local bowling center slips and falls on it hilariously. But at the highest levels of the sport, lane oil is a foundational element of tactics and game-planning. The oil is applied in a distinct pattern that reduces the friction of specific parts of the lane, allowing a well thrown ball to slide across certain areas, then grab ahold of the surface and curl when it’s supposed to.But lane oil patterns aren’t always identical. Oil distribution can change over the course of a game, and in fact the starting pattern is intentionally redrawn for various PBA events, forcing pros to adjust on the fly.“All of us as professionals are playing a guessing game by watching our bowling ball go down the lane,” Tackett tells me. “And trying to figure out where the oil has moved from and has moved to.”Tackett says understanding and reacting to lane oil patterns has become one of the single most important skills for a pro bowler in recent years.Technology plays a growing role in the way lane oil is understood and applied. And as the tech has continued to improve, the sport’s powers that be are increasingly relying on those advances to keep the top pros guessing.Back when bowlinglanes were made from actual wood like maple or pine in the 1970s and earlier, lane oil’s primary purpose was protective. It provided a lubricating barrier between ball and wood, helping reduce the wear and tear on both.During that era—and even into the early 1980s when synthetic lane materials began replacing wood—lane oil drew the ire of serious bowlers due to the inconsistency of how it was applied. There was no method for ensuring standardized oil applications; sections of lanes would often have too much or too little oil. Bowlers would essentially be blind to the lane’s conditions until partway through a game, after enough balls had been thrown to get a gauge of which areas were light or heavy on oil. Maybe certain bowlers enjoyed the randomness, but it was untenable for professionals.John Davis, at the time a mechanic who oiled lanes for his local bowling center in Arizona, was the first to address this issue. He invented a lane cleaning tool in the early ’80s to address lanes that were often cleaned inconsistently, allowing oil to pool or run as a result. By the early ’90s, he had started a company named Kegel that had perfected what’s known as Sanction technology, which isstill used today. The bowling equipment giant Brunswick, the other major player in the space, uses similar technology in itsMax lane machines.“It looks almost like an inkjet printer,” Chris Chartrand, Kegel’s CEO, says about the Sanction tech. “You have this head that travels back and forth, and it applies that stream on whatever board we want to apply it to.”The technology has kept improving. Kegel used to produce two separate machines, one for lane cleaning and one for oil; now both steps are handled by a single machine. More recently, the machines have become battery-powered, removing issues of power cords stretching across lanes.Lately, Kegel has been steadily improving its automation, to the point where today’s machines do the entire job without any human intervention.The lanes youand I bowl on as amateurs are oiled very differently from the ones pros use.At your local bowling center, public lanes are oiled in what’s referred to as a “high” ratio: The level of oil present in the middle of a lane is eight to 10 times higher than what’s on the outside. At the far left and right of the lane, many public bowling alleys have no oil at all.“On a normal pattern at your normal bowling center, there is some autocorrect,” Tackett says. Because the edges of the lane have very little oil, shots that drift to either side will slow down; if the ball has been thrown with the proper spin to guide it back toward the middle of the lane, it will curl more effectively on the drier surface. “It makes it easier to hit the pocket.”(By “the pocket,” Tacket means that sweet spot at the front

Jill DuffyGearMay 17, 2026 6:30 AMTake Control of Your Debt With These Free Online ToolsThese free debt calculators help you set up payment plans to get back in the black.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyApps for budgetingandpersonal financedo a good job of tracking yourmoneyas you earn and spend it. Some also have excellent debt calculators that help you figure out how to pay off your debts.Each debt calculator is a little different. Some suggest a specific method for paying down debt, while others are simulators that let you see how your total amount paid will decrease if you increase your monthly payment.Here are a few useful calculators and some guidance about what makes them different.A Straightforward Plan: BankrateBankrate’s freedebt payoff calculatorgives you a timetable for paying off each of your debts. You enter as many debts as you want to include, their interest rates, total loan amounts, and other details. You also enter any new income you expect to receive, such as an annual salary increase or windfall, and the amount that you can put toward your debts. The calculator then generates one payment table for each debt showing how much to pay each month until the debt is cleared.Bankrate prioritizes paying off the debt with the highest interest rate first. Once your first debt is paid off, the money you would have put toward it is diverted to your other monthly payments. In other words, as you eliminate debts, the monthly payments on your other debts increase until they, too, are paid off.Who should use it?Bankrate’s calculator works for people who have multiple debts, and the total monthly minimum payments are within their financial reach. If that’s you, then you’ll get a crystal clear plan—with a timeline—for getting rid of all your debts.Where it comes up short.This calculator assumes that paying off your debts by clearing the one with the highest interest rate first is in your best interest. That’s not true for everyone. You might have other options, such as consolidating credit card debt to a new card with a 0 percent introductory rate or filing for bankruptcy. Bankrate also doesn’t take into account other personal finance concerns, like other uses of monthly funds that free up once you pay off your first debt—Bankrate tells you to put that money toward your next-highest-interest debt. You might be better off putting it toward retirement savings or an emergency fund.Big-Picture Guidance: NerdWalletNerdWallet’s freedebt load calculatordetermines your debt load as a percentage of your income. The resulting debt load is classified as smaller (less than 36 percent), larger (37–42), or overwhelming (43 percent or more). Based on the outcome, NerdWallet suggests a method for eliminating your debt, which you read about in an educational article below the results.Who should use it?This calculator helps you get a big-picture sense of your debt. If you have a lot of debt, it’s useful for ruling out (or ruling in) the option of declaring bankruptcy.Where it comes up short.It’s not great at analyzing the finer details of your debt. For example, in the setup, there’s no line item for student loans or a mortgage, much less the exact interest rate you pay on loans. The results are a rough guide rather than a personalized strategy.Automated Inputs: WalletHubWhen you sign up for WalletHub (free) and connect your financial accounts, the app pulls real information about how much money you owe and your payment history. Itsdebt payoff planis a calculator that lets you play with the numbers to see what would happen if you increased your monthly payment. How much faster can you clear the debt? How much will you save in interest? You can quickly see the difference between increasing your monthly payment by, say, $50 versus $150.Who should use it?This calculator is for WalletHub users who have connected their financial accounts. It’s most useful for people who can afford to pay more than the monthly minimum on their debts.Where it comes up short.Although WalletHub pulls real information about your actual loans, it can miss some detail. I use it to track my mortgage, for example, and it doesn’t know that the total bill I pay each month includes homeowner’s insurance. The calculations regarding the principal and interest are still correct, but the amount that comes out of my pocket is higher than what WalletHub assumes. Other considerations are missing, too, like what to do when the rate of inflation is higher than the interest rate on a long-term loan. Overall, this debt calculator is very good, and these shortcomings are nitpicks, not deal-breakers.Credit Card Transfers: WalletHub and NerdWalletIf you’re carrying credit card debt,WalletHubandNerdWalletboth have a calculator that shows how much money you could save by transferring the balance to a new card with an introductory rate. (CreditKarma has a balance transfer calculator, t

Matthew KorfhageGearMay 17, 2026 6:33 AMThe Best Grills of 2026: Cook Smarter, Not HarderI grilled, smoked, seared, cleaned, and synced digital temp controllers to find the best grill option for every cookout and tailgate.Courtesy of AmazonCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe Grill haschanged. The smoker has changed, too. Once governed mostly by feel and experience, the best grills and the best smokers now offer a fair amount of tech to measure and control temperature (think: temperature probes and fans to modulate airflow). After all, mankind’s formative invention wasn’t the “discovery” of fire; it was the power to control it.The WIRED Reviews team has been testing grillsfor more than a decade—searing, smoking, grilling, and even baking on them in all kinds of weather—to find the best choice for everyone. These days, I’ve been most excited about the next-gen smart technology that allows for tight monitoring and temperature control. For instance, I tricked out my top-pick for smart upgrades, an old-fashioned Weber kettle, with the excellentSpider Venom ($280)digital temperature controller that allows me to turn my charcoal grill temps up and down like an oven. Summer is good. Also,charcoal makes all meat taste more like bacon.Mmmmm, bacon.After you’re set with the best grills and smokers, check out other backyard buying guides for all your outdoor needs, like theBest Flat Top Grills and Griddles,Best Wireless Meat Thermometers, andBest Smokeless Firepits.Best Grills of 2026Best Charcoal Grill for Smart UpgradesWeber Original Kettle CharcoalJump to review$139$129(7% off)Amazon (18-inch)Best Smart Pellet SmokerRecteq Flagship 1600 Pellet Smoker and GrillJump to review$1,600BBQGuysBest Smart Charcoal GrillMasterbuilt Gravity Series 1150 Digital Charcoal GrillJump to review$1,100Home DepotBest Gas GrillWeber Spirit E-210 Gas GrillJump to review$449The Home DepotJump to SectionAccordionItemContainerButtonLargeChevronOther Grills I RecommendA Grill to AvoidWhat to Look for in a Good GrillHow to Upgrade Dumb Grills to Smart GrillsUseful Grill AccessoriesBest Charcoal Grill for Smart UpgradesPhotograph: Matthew KorfhagePhotograph: Matthew KorfhageVideo: Matthew KorfhageChevronChevronSave to wishlistSave to wishlistWeberOriginal Kettle Charcoal$139$129(7% off)Amazon (18-inch)$149Amazon (22-inch)$99Weber (All, Starting Price)WIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDSimple, durable, reliable, and affordableNear-infinite smart-tech upgrades availableLightweight and portableTIREDIt could do with better wheel nutsUpgrades often cost more than the grillI’d argue that the best way to take advantage of top smart-grill technology is to add the newest and best tech to a sturdy, old-fashioned, classic grill that’ll outlast the tech. A grill like the classic 22-inch Weber Kettle.The Weber Kettle is the infinitely hackable Honda Civic of backyard grills. Its combination of simplicity, reliability, durability, affordability, and outright ubiquity has spawned an innovative secondary market devoted to tricking out the humble cooker with smart temperature controls and more features than a multiplex.The best Weber add-on I’ve tried so far is a fan-modulated temperature controller from Spider Grills called theVenom ($280)that tracks the temperature on each cook, allowing you to heat it up or turn it down, like a stove. According to my testing, it holds your charcoal grill to temperature within about 30 degrees Fahrenheit of the target, which is great for both low-and-slow cooking and for creating repeatable sear temps for a steak. This is especially true when using the attached meat probes. The only downside is that the Venom needs to be plugged in. Might I suggest a nicepower bank?WIRED product reviewer Scott Gilbertson has favored the excellent WeberConnect Smart Hub ($70)for years, which features probes to monitor meat temps and flipping times and can alert your phone (or just your ears) when your grill hits the ideal temperature zone. However, he admits that the Hub’s USB-A charger is starting to feel a bit old-fashioned.This year, Weber is also coming out with aPerformer Smart Ring ($280)that seems to operate a lot like the Venom. If it works anywhere near as well as the Venom, I’m eager to try it out.But not every kettle upgrade worth trying is “smart.” I am in the process of testing a number of add-ons to the 22-inch Weber Kettle, including aWebcraft cooking ring ($249)from Spider that serves as a base for side shelves, arotisserie, and apizza oven conversion kit. AWeber rotisserie ($180)is also available from the Illinois mothership.Note also that there are two versions of the Weber Kettle: Original andPremium. The Premium is around $80 more and features a built-in thermometer in the lid, a hinged cooking grate, and a fancier ash-removal system. They’re the same size and feature the same cooking space. Unless you really like the easier ash-removal system, save that cash

Christopher NullGearMay 17, 2026 7:02 AMReview: Oto Smart SprinklerThis smart sprinkler doesn’t need a power cable, thanks to its integrated solar panel.Courtesy of AmazonTriangleUpBuy NowMultiple Buying Options Available$589$499at Amazon$589$499at OtoCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyRating:9/10Open rating explainerInformationWIREDSolar panel is a game changer. App is easy to navigate during both setup and daily use. Watering is accurate and thorough. Built-in weather intelligence helps to avoid unnecessary runs.TIREDDistant watering can land with heavy pressure, potentially damaging plants.One of themajor headaches of stand-alone smart sprinklers that attach to a garden hose is that they require power to operate. Since few folks have electrical outlets in the middle of their lawns, configuring a smart sprinkler means connecting not just a hose but also a long power cable to the device, creating double the effort and double the chance for something to get chewed up by a lawn mower. The Oto Smart Sprinkler solves for this problem with solar power, which makes for not only a far more flexible installation than its smart-sprinkler competitors but also ongoing easy use.With a strong focus on quick setup and simplicity of function, Oto provides a refreshing alternative in the smart sprinkler market. The only question: Does its useful innovation justify its blisteringly high price tag?Set and ForgetWhen I unboxed the Oto Smart Sprinkler for the first time, I was convinced something must be missing—perhaps even a second box with more parts? But as it turns out, the Oto really is as streamlined as it claims to be. The sprinkler measures just 16 x 12 x 4 inches and weighs about 6 pounds. Next to the similarly featured but monstrous-by-comparisonAiper IrriSense 2, the Oto is practically miniature.Being pint-sized makes lawn irrigation even easier than expected. You can place the unit on the ground, mount it to a hardscape, or even attach it to a fence (all with the included metal bracket). If you want to play things fast and loose, you can situate it to stand freely without the bracket, though the bracket helps you ensure it’s replaced in the exact same location if it’s ever removed (such as during the winter months) for even irrigation.The only real placement caveat is that in order to have it operate cordlessly, the Oto requires at least three hours a day of direct sunlight to its top-mounted 2.2-watt solar panel to keep its internal 5,400-mAh battery topped up. Keep this in mind as you decide where to install your brackets—you definitely don’t want to block too much of the sun’s rays.Video: Christopher NullBefore taking it outside for its first use, Oto instructs you to charge the device overnight via wall power (though you could certainly just leave it in the sun for a few days instead). You’ll do this using the included power adapter with 50 feet of cable. You can also lean on the plugged-in charging option if you can’t squeeze out enough sunshine on a regular basis (albeit in that case, this irrigation solution is probably not for you). You’ll find the power cable stored inside the Oto hardware, in a large hollow space beneath the solar panel. This hollow space is intentional and isn’t used for anything more than supporting the panel above, but it’s probably best not to leave the power adapter inside permanently, as it this chamber can get rather wet.The Oto can throw water up to a distance of 40 feet in a 360-degree radius of the device, which equates to a maximum of 5,000 square feet of coverage, enough for a sizable lawn. (That’s about on par with Aiper.) The hose connector on the side of the device is easy to work with and secure; I experienced no leaks at all when connecting it to a standard garden hose. A 50-foot maximum length for your hose is specified to ensure water pressure remains sufficient.Digital DropletsJust about everything on the Oto happens via its app (though there’s a manual switch on top of the device you can use to speed up a run that’s pending in the app). That app sets up via a 2.4-GHz Wi-Fi connection to your home network, so you’ll need to make sure your yard has effective wireless coverage.Once configured, setup proceeds much like the Aiper and pricierIrrigreenapps: You create a zone, then use the app to define its boundaries. Similar to the aforementioned systems, Oto’s sprinkler is designed for precision watering, firing water in a beam in a single direction instead of a wide spray. That said, Oto’s spray is comparably narrow, only hitting a single, designated patch instead of producing a two-dimensional curtain of water like Irrigreen’s “water printing” system. You get a nice preview of this as you set the boundaries of your yard.Like its competitors, Oto lets you set each zone as a spot (for watering a single tree, perhaps), a line (for a flowerbed), or a 2-D area (for a yard). I tested all of these modes but spent most of my time workin

Arthur Allen, KFF Health NewsScienceMay 18, 2026 5:00 AMA Danish Couple’s Maverick African Research Finds Its Moment in RFK Jr.’s Vaccine PolicyThe work of Peter Aaby and Christine Stabell Benn has long been controversial. Until Robert F. Kennedy Jr. became US health policy chief, most vaccine scientists tended to ignore it. Now they can’t.Photograph: Tom Williams/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyIn 1996, Guinea-Bissauseemed like an ideal research post for budding pediatrician Lone Graff Stensballe. Her supervisor, a fellow Dane named Peter Aaby, had spentnearly two decades collecting dataon 100,000 people living in the mud brick homes of the West African country’s capital.Aaby and his partner, Christine Stabell Benn, believed that the years of research in the impoverished country had yielded a major discovery about vaccines—and what they described as “non-specific effects”: The measles and tuberculosis vaccines, which were derived from live, weakened viruses and bacteria, they said, boosted child survival beyond protecting against those particular pathogens.But, the scientists said, shots made from deactivated whole germs, or pieces of them, such as the diphtheria-tetanus-pertussis (DTP) shot, caused more deaths—especially in little girls—than getting no vaccine at all.The World Health Organization repeatedly and inconclusively examined these astonishing findings. They tended to elicit shrugs from other global health researchers, who found Aaby’s research techniques unusual and his results generally impossible to replicate.Then came Donald Trump, Covid, and the administrative reign of anti-vaccine advocate Robert F. Kennedy Jr.Suddenly, Aaby and Benn weren’t just sending up distant smoke signals from a far corner of the planet. They wereconfidently voicing their viewsand policy prescriptions online and in medical journals. The “framework” for “testing, approving, and regulating vaccines needs to be updated to accommodate non-specific effects,” their team wrote ina 2023 review.And the Trump administration has taken notice.“They became more strident in saying that their findings were real and that the world needed to do something about it,” said Kathryn Edwards, a Vanderbilt University vaccinologist who has been aware of Aaby’s work since the 1990s. “And they became more aligned with RFK.”Kennedy, as secretary of the Department of Health and Human Services,cited one of Aaby’s papersto justify slashing $2.6 billion in US support for Gavi, a global alliance of vaccination initiatives. The cut could result in 1.2 million preventable deaths over five years in the world’s poorest countries, the nonprofit agency has estimated. Kennedy has frozen$600 millionin current Gavi funding over largely debunked vaccine safety claims.Kennedy describedthe 2017 paperas a “landmark study” by “five highly regarded mainstream vaccine experts” that found that girls who received a diphtheria-tetanus-pertussis, or DTP, shot were 10 times more likely to die from all causes than unvaccinated children.In fact, the study was far too small to confidently make such assertions, as Benn acknowledged. In a study of historical data that included 535 girls, four of those vaccinated against DTP in a three-month period of infancy died of unrelated causes, while one unvaccinated girl died during that period. Afollow-up published by the same groupin 2022 found that the DTP shot by itself had no effect on mortality. Critics say the 2017 study, rather than being a landmark, exemplified the troubling shortfalls they perceive in the Danish team’s research.As Aaby and Benn’s US profile has risen, scientists in Denmark have set upon the work of their compatriots. In news and journalarticlespublished over the past 18 months, Danish statisticians and infectious disease experts have said the duo’s methods wereunorthodox, evenshoddy, and were structured to supportpreconceived views. A national scientific board is investigating their work.Stensballe, who worked with Aaby and Benn for 20 years, has been among those voicing doubts.“It took years to see what I see clearly today, that there is a strange concerning pattern in their work,” Stensballe said in a phone interview from Copenhagen, where she treats children at Rigshospitalet, the city’s largest teaching hospital. She said their work is full of confirmation bias—favoring interpretations that fit their hypotheses.Those hypotheses overlap, in important areas, with the notions of Kennedy and other vaccine-skeptical officials at HHS.In December, HHS announced the agency would award the Bandim group $1.6 million to study whether the birth dose of hepatitis B vaccine weakens babies’ immune systems or causes neurological issues.The researchers plan to withhold the vaccine from half of the 14,000 newborns in the study, although the long-established vaccine is 90 percent effective in preventing infection. The Bandim group justifies this decision by noting tha

Ali WinstonSecurityMay 18, 2026 5:30 AMAn ICE Firearms Trainer Was Involved in At Least 4 Deadly ShootingsDavid Norman, a former Phoenix police officer who’s described himself as “a fucking savage,” now runs a company that provided training to Homeland Security’s Special Response Teams.Photo-illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe owner ofa company that trainedparamilitaryImmigration and Customs Enforcementagents testified that he was involved in at least four lethal shootings, according to a 2021 deposition related to a lawsuit reviewed by WIRED.David S. Norman, the founder and proprietor of law enforcement training firmTruKinetics LLC, served as a Phoenix Police officer from the late 1990s until his retirement in 2020. Prior to founding TruKinetics the same year, according to records reviewed by WIRED, Norman was involved in six shootings while on duty that left four people dead and two more wounded. In every instance, the Phoenix Police Department said Norman fired on an armed suspect and exchanged volleys of gunfire in at least two of the shootings.Based in Gilbert, Arizona, TruKinetics offers training on small-team tactics, hostage rescues, close-quarters combat, building searches, night-vision firearms proficiency, pistol and rifle courses, “vehicle interdiction,” breaching with explosives, and sniper tactics, according to the company’s website.TruKinetics received $27,748 for a year-long contract to run a mandatory40-hour training coursethat certain members of Department of Homeland Security Special Response Teams receive annually at Fort Benning in Georgia, according to governmentprocurement recordsreviewed by WIRED. At least 700 SRT agents from the Customs and Border Protection’s Office of Field Operations, ICE’s Homeland Security Investigations division, and ICE’s Enforcement and Removal Office units pass through Fort Benning for annual training.In an interview with WIRED, Norman says that his company conducted sessions with the Special Response Team from Arizona’s Homeland Security Investigations office. “They’re top dudes, and it was an honor to work with them,” he tells WIRED. Norman maintains that his courses, which took place in Arizona and Georgia’s Fort Benning, did not involve crowd control tactics or active shooters, but would not specify further. “It sounds like you’re one of those dudes who’s doing a hit piece on HSI,” he says.TruKinetics postedtwo photos to Instagramin August 2024 of Norman and three TruKinetics trainers alongside 19 uniformed operators from HSI’s Arizona Special Response Team, posing in a “kill house” training course—a set of rooms and hallways filled with obstacles and targets designed to simulate close-quarters combat.Customs and Border Protection did not respond to WIRED’s questions about how many SRT teams and operators went through the Gilbert, Arizona, company’s training course.In a 2021law enforcement podcast,The Modern Cop, Norman describes himself as “a fucking savage” who sought out high-risk experiences and shootings as a cop. “I wanted these experiences. I was super aggressive,” Norman said. He also appeared to joke about police shootings, telling the host that “you kind of hope it’s on your Friday, so you can actually have days off.”Once reserved for armed or high-risk suspects, manhunts, and potentially dangerous building entries, the SRTs are now being used for civil immigration enforcement, crowd control, and basic warrant service, operations that the unit wasonce restrictedfrom performing. BothRenee GoodandAlex Prettiwere killed while protesting the militarized federal immigration surges in Minnesota, with SRT membersimplicatedinbothof their deaths. While recent debate over Homeland Security’s violent immigration sweeps have focused on whether agents receiveadequate training, the background of SRT’s training contractor raises questions about who is training ICE’s and CBP’s paramilitary units, and what they are being trained to do.For a dozenyears of Norman’s two decades as a Phoenix cop, he served on the agency’s Special Assignments Unit, a plainclothes fugitive apprehension team that Norman repeatedly characterized in a 2021 deposition from a lawsuit filed the previous year as having morphed into what was considered a “SWAT” (special weapons and tactics) team. On those units, he worked as a “point cover” man and at times doubled up as a trainer on pistol usage for units within Phoenix PD, according to his deposition and Phoenix Police documents reviewed by WIRED.In addition to TruKinetics, Norman has worked as an “independent contractor for the Department of Defense” and provided training under a nondisclosure agreement, according to his deposition.Norman’s supervisors consistently gave him solid ratings and he received numerous department commendations. The four fatal shootings he was involved in were deemed in policy by both the local district attorney and Phoenix PD. Ho

Chris ColinThe Big StoryMay 18, 2026 6:00 AMI’m a Normie. Can Normies Really Vibe Code?Apparently anyone can vibe code anything these days. So Claude and I tried to make a database for tracking the petty grievances of the masses.Illustration: Yann BastardCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe dog thatushered me into the technological future was “low and thick.” That’s all my mother registered before it T-boned her in a city park earlier this year: dense, heavy, and traveling fast enough to fracture her right tibia. But enough abouther. Let’s discuss what this set in motion in my life: Having successfully learned nothing about coding for two and a half decades, I would soon be attempting my very first software development project.If you’ve ever had a low and thick dog break your mom’s shin bone, you know the stream of lesser indignities that follows. Case in point: the hours my father spent navigating phone trees, trying to manage my mom’s medical care. Are frustrating telephone calls significant in the grand scheme of things? No. But that stupid dog had chosen a technologically interesting moment to do its thing. For the first time in history, a problem no longer needed to be serious to bringserious toolsto bear.For as long as Silicon Valley has been selling a frictionless tomorrow, we ordinary people have been its passive shoppers—scrolling the App Store, hoping someone has gone to the trouble of building whatever we need. Enter AI and its democratizing sidekick:vibe coding. If the promise is real, suddenly we can build our own apps, as niche and trifling as we please, with zeroprogrammingskills. We merely gesture toward whatever irks us and a constellation of large language models, code generators, and development environments will click their heels.Illustration: Yann BastardNiche and trifling? That’s me! Where others vibe code résumé reviewers and inventory trackers andautomated assistantsto boost their work productivity, I had a different target in mind. Over the past couple years I’ve grown particularly fixated, personally and professionally, on what the policy world calls sludge: the rising tide of tiny administrative obligations that increasingly seems to define modern existence—and corrode our ability to get anything done. The hassle of dealing with insurance, or connecting that insurance to your doctor, or staying on top of airline miles, or navigating your kid’s school portal. The steps required to dispute a charge, or unsubscribe from a streaming service you forgot you had. Each of these feels like its own discrete assault on our time. But they’re not discrete. They’re separate mushrooms sprouting from the same mycorrhizal network.In a way this is a calibration issue. While bigger problems might at least theoretically attract attention—legislation, journalism, a Senate hearing—the smaller ones, too petty to litigate, simply become a fact of life. The arc of history may bend toward justice, but when it comes to fighting a one-dollar bank fee, it bends toward hold music.Which is where the fantasy of vibe coding captured my attention. Those hassles aren’t just accidental byproducts of complexity; they’re often features. A confusing portal, a dropped call, a process just opaque enough to discourage follow-through. At scale, they function less like bugs than like policy. The app I envisioned would expose this phenomenon, make the cumulative weight of these obligations a little harder to ignore. The image I’d like you to summon is a field of mushrooms trembling.What my momlacks in healthy legs, she makes up for in a Claude Pro subscription. Having needled her repeatedly over the past couple years about AI’s environmental, political, and economic implications, I brushed all that aside on a recent Sunday and drove to her house. After a little tibia talk, I opened her computer and began emitting vibes.I’d like to create a communally shared app that gathers and shares information related to how much time and energy we devote to fighting burdensome administrative tasks, bureaucratic sludge, Kafka-esque unsubscribe mazes, byzantine insurance portals, wrongful charges, denied claims, confusing membership plans, and the like.With as much clarity and detail as I could muster, I proceeded to describe a dashboard that would record the scale and scope of our collective sludge. Users would log frustrating incidents from their lives, entering how much time they’d spent, how annoying it was, and what they’d rather have been doing. Every submission would be dopaminally rewarded with an inspiring resistance quote and a photo of a kitten, puppy, or baby chimp. I’d train Claude to generate some “wider context”—a paragraph discussing how the frustrating incident fits into systemic sludge patterns—and a complaint letter to the relevant regulatory bodies.Claude noodled. Not for the first time, I feared my vibes would simply manifest an error page. I recalled, dimly, some of the advic

Emily MullinScienceMay 18, 2026 6:30 AMSupercharging Immune Cells May Help Control HIV Long-TermCAR-T cell therapy is already a potent treatment for certain cancers. Now, a small study is showing early promise for managing HIV.Photograph: Bloomberg/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyA Miracle cancertherapy that involvesengineering a patient’s own immune cellsis being repurposed for HIV, and early results from two individuals hint at its promise for long-term control of the virus.As part of a clinical trial, scientists took people’s own immune cells and reprogrammed them in a lab to recognize and attack HIV in the body. After a single infusion of the modified cells, two individuals with HIV now have undetectable levels of the virus—one for nearly two years and the other for almost a year. Both have been able to go off HIV medications entirely.The two people are part of a small study to test the treatment’s safety and feasibility. The initial findings were announced last week at the of American Society of Gene and Cell Therapy annual meeting in Boston.“These are early days. If we can provide the proof-of-concept that this approach is both safe and effective, then there are lots of ways in which it can be optimized, to make it more affordable and scalable,” says Steven Deeks, a professor of medicine and HIV expert at the University of California, San Francisco, who led the trial.The technique, known as CAR-T cell therapy, has been used intens of thousands of patientswith tough-to-treat cancers. Half a dozen or so drugs have been approved that rely on the technique. The treatment essentially supercharges a person’s immune system to directly attack and eliminate cancer cells. Recently, it’s also been used successfully to treatsevere autoimmune diseases.“This is pretty exciting,” says Andrea Gramatica, vice president of research at amfAR, the Foundation for AIDS Research, who was not involved in the trial. “The reason this study matters and is particularly important is because it gives the HIV field a real, clinical clue that teaching the immune system to control the virus without antiretroviral therapy is achievable.”Scientists have been pursuing a cure for HIV since the virus was first identified in the early 1980s. Antiretroviral therapy prevents the progression to AIDS by suppressing the virus to undetectable levels, but people must take medication for the rest of their lives. It has transformed HIV into a chronic condition that allows people to have a near-normal life expectancy. Yet not everyone who is HIV positive is aware of their status, and in some rural and low-income parts of the world these medications are still not widely accessible or affordable.Up until now, there areunder a dozen documented casesof sustained remission from HIV—known as a “functional cure” because the virus is still present in the body but is suppressed to levels that are undetectable by the immune system and HIV medication is no longer needed.Each of those individuals developed cancer and underwent stem cell transplantations as part of their treatment. In all butone of those cases, doctors used stem cells from donors with a rare genetic mutation called CCR5 that naturally prevents HIV from entering and infecting healthy cells. Timothy Ray Brown, known as the “Berlin patient,” was the first known person to be cured of HIV in this way in 2008.The examples of sustained remission “have taught us that the immune system can, under the right conditions, clear HIV,” says Boro Dropulić, executive director of the Maryland nonprofit Caring Cross, who developed the CAR-T therapy for HIV.But stem cell transplants aren’t scalable, he says. They’re intensive procedures that carry serious risks such as graft-versus-host disease, when the transplanted cells recognize the recipient’s cells as foreign and attack them.“What we’re trying to do is to engineer that outcome deliberately without requiring cancer, without requiring a specific donor,” Dropulić says. His organization is working on making advanced therapies like CAR-T more accessible and affordable.Cancer and HIV are similar in that both can hide from the body’s immune system. In CAR-T therapy for cancer, patients’ T cells are engineered to express chimeric antigen receptors, or CARs, on their surface. These added receptors allow T cells to specifically identify, lock onto, and destroy cancer cells by recognizing certain antigens, or proteins, on their surface. The specific CAR that’s added depends on the type of cancer being treated.Dropulić and his team engineered patients’ T cells to recognize two different sites on the HIV virus, making it more difficult for the virus to escape. “Our goal is that these cells remain like sentries in the body,” he says. “Whenever these embers of the virus start replicating, these cells are there in order to immediately take care of them.”The trial included nine total participants who were all on antiretr

Lindsay DodgsonScienceMay 18, 2026 7:00 AMThe Backward Logic of Chickenpox PartiesBefore vaccines, some parents intentionally exposed their children to the virus. In the internet age, that thinking is resurfacing.Photo-Illustration: Darrell Jackson; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAnyone who hashad chickenpox shares one distinct memory: the relentless, all-consuming itch.Ciara DiVita was only 3 years old when she caught the virus, but she remembers it well—along with the oven mitts she was made to wear to stop herself scratching. She also recalls being taken to hang out with her cousin while covered in blisters, in the hopes of deliberately infecting them.DiVita, now 30, was actually the second in the chain, having been taken by her parents to catch chickenpox from an infectious friend. “I imagine the chain continued and my cousin gave it to someone else at a chickenpox play date,” she says.A lot has changed over the past three decades, most notably the development of a chickenpox vaccine, meaning the virus is no longer the childhood rite of passage it once was.Thanks to the vaccine’s success, children today are much less likely to be exposed to the infection at school or on the playground.Chickenpox parties are also largely considered a relic of the past—a strategy many Gen X and millennial children were subjected to before vaccines became routine. But much like the virus itself—latent, opportunistic—they haven’t disappeared entirely.Before a vaccineexisted, chickenpox, which is caused by the varicella-zoster virus, felt unavoidable. In temperate countries like the UK and the US,around 90 percentof children caught the virus before adolescence (in tropical countries the average age of infection ishigher).It’s nothing to do with chickens. The splotchy, scratchy, highly contagious disease is possibly named after the French word for chickpea,pois chiche,according to one theory, because the round bumps caused by the virus resemble their size and shape. While most infant cases are mild, adolescents and adults are more likely to develop severe complications.This is where the idea of “getting it over and done with” emerged from, according to Maureen Tierney, associate dean of clinical research and public health at Creighton University in Omaha, Nebraska.“You were trying to have your child get the disease when they were at the greatest chance of not having complications,” Tierney says, explaining that, generally speaking, the older the patient, the more severe the infection can be.While varicella-zoster is usually a mild, self-limiting disease in children, it can be much more severe—and sometimes life-threatening—in adults.“I had an otherwise healthy adult patient who died of chickenpox pneumonia when I was first practicing,” Tierney says. “You never forget those scenarios.”The virus spreads rapidly through respiratory droplets and contact with fluid from its characteristic blisters, meaning if one child contracts it, siblings and classmates are likely to be next, if unvaccinated.Before the existence of social media, the idea that children should deliberately infect each other spread just as rapidly around communities—in conversations in the school yard, church groups, and pediatric waiting rooms—leading to the popularity of so-called chickenpox parties.Parents swapped advice about oatmeal baths and calamine lotion and arranged to bring children together when one was thought to be infectious—despite the practice never being an official medical recommendation.“They thought, well, if it’s going to happen to my kid anyway, it might as well happen in a controlled environment,” says Monica Abdelnour, a pediatric infectious disease specialist at Phoenix Children’s Hospital. “The families were ready to encounter this infection, deal with it, and then move on.”While the majority of children who develop chickenpox feel well again within a week or two, around three in every 1,000 infected experience a severe complication such as pneumonia, serious bacterial skin infections, encephalitis (inflammation of the brain), or meningitis.“Some kids get really, really sick,” says Jill Morgan, a professor at the University of Maryland School of Pharmacy and an expert in pediatric health. “The problem is, if you had these parties, you didn’t know which kids would get over it and be OK, and which kids would end up in the hospital.”Chickenpox parties faded into obscurity once routine vaccination began to take hold in many parts of the world; a2018 analysisof global vaccination trends and infection rates shows cases decreased dramatically in countries such as the US, Germany, and Australia, where universal childhood vaccination programs have already been rolled out.In the US, the vaccine has been associated with a roughly 97 percent drop in reported cases since its introduction,according to data published by the Centers for Disease Control and Prevention. Global data show similarly st

Boone AshworthGearMay 18, 2026 7:00 AMHow to Watch Google I/OGoogle I/O is back with updates to Search, Android, Gemini, and a fresh peek at upcoming Android XR smart glasses. Here’s how to watch the announcements live and what to expect.Photograph: Camille Cohen/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyGoogle I/O isimminent. This is the annual developer event and product showcase whereGoogleshows off all the shiny new updates to its Android operating system and other platforms, as well as new features and improvements to its artificial intelligence models.The big announcements atGoogle I/Ousually come in the form of a livestreamed keynote event at the Shoreline Amphitheatre in Mountain View, California, hosted by Alphabet CEO Sundar Pichai. That starts on Tuesday, May 19, at 10 am Pacific time (1 pm Eastern). You can watch thelivestreamon Google’s website or onYouTube. It will also be streaming right here on this page.What to ExpectOn the announcement front, Google actually tipped its hand last week, unveiling a bunch ofnew featurescoming to the upcoming Android 17 operating system and its ongoing Gemini AI efforts, which it now calls Gemini Intelligence for the most advanced Android devices.Many of the newAndroidfeatures highlighted are AI-infused tricks that aim to make your life a little easier by offloading onerous tasks like booking a hotel or ordering an Uber. Google is also bringing more robust voice-to-text recognition, some autofill smarts, and enhanced digital well-being settings that you’ll turn on for a day and then probably just disable. Also, there is a new look for Android Auto, and Google is making its emojis look more 3D. You canread more about these capabilities here.The company also announced its newGooglebookplatform: AI-first laptops powered by various Android technologies, but it’s very different from a Chromebook. Google won’t share too many specifics yet, but these machines are expected to be premium laptops, and companies like Acer, HP, Lenovo, and Dell have already signed up to make Googlebooks by year’s end. Wehave more details here.We’ll likely hear Google touch on these features at I/O again, but these were unveiled early to make room for more new software announcements during the keynote. You can usually count on a surprise announcement or an AI demo that goes completely off the rails.Based on the ever-swirling rumor mill, Google likely will announce more updates to Search, undoubtedly powered by AI. There may also be updates to its creativeGoogle Labssuite, such as its AI music creation toolFlow. Google may introduce a new Gemini model for video generation,expected to be called Omni.Last year, Google announced a newGemini-powered Google Home smart speakerand said it would launch sometime in the spring of 2026, so Google I/O would be the time to make good on that promise and maybe announce a launch date.While I/O is primarily a software-focused event, Google is heading toward an inevitable crop of new hardware. The company has been building out itsAndroid XRplatform, which it is positioning to power smart glasses from companies like Samsung, Xreal, Gentle Monster, and Warby Parker. Google already said last week that I/O will feature a sneak peek at its smart glasses, and we may learn when in 2026 these devices will finally go on sale.Follow along during the event on WIRED’s liveblog, and stay tuned for many more stories about everything Google announced.

Jeremy WhiteGearMay 18, 2026 9:38 AMThe Catastrophic Audemars Piguet x Swatch Launch Was Entirely Predictable and Utterly AvoidableChaos consumed Swatch stores worldwide on Saturday—a repeat of the MoonSwatch launch four years ago. Swatch should have seen it coming.Photograph: RAPHAELLE PELTIER/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAt 9 am,as stores opened, London’s Carnaby Street location lasted half an hour before police were called. In New York, fights broke out amid rumors of a stabbing in the line. At locations around the world, police had to assist staff as the vast majority of customers left empty-handed, since most of the shops had fewer than 200 watches available. Deals were done within queues as scalpers flipped watches for profit. Within minutes of the first sales, the plastic watches were hitting eBay for thousands of dollars.This is not a description of what happened during Saturday’sSwatch x Audemars Piguet Royal Poplaunch; this is an account of the 2022 MoonSwatch launch. But a carbon copy of this chaos and mayhem ensued as Swatch stores opened worldwide on May 16 to sell their extremely limited numbers of Royal Pop pocket watches.Responding to widespread criticism of how Swatch handled the 2022 MoonSwatch launch, Nick Hayek Jr., chief executive of Swatch Group,told WIREDat the time: “We knew for sure this would be a success, because the product is beautiful, provocative, high quality, and the price is fantastic … But what happened … I think nobody in the world could have expected that. It was really crazy.”Hayek’s defense was essentially that Swatch could not be held accountable for disturbances so severe that some stores were forced to shut for 10 days to let the pandemonium play out, simply because the group could never have predicted the ferocious public reaction to an affordable version of an iconic luxury timepiece coming on sale. No such excuse can be offered this time for the calamitous Royal Pop launch that took place this weekend.Chaos ForeshadowedEven before the first stores opened in Singapore on Saturday, 12 hours ahead of New York, the clear pattern of what was to follow was established. TheVivoCity branch was shutteredahead of the much-anticipated launch, with Swatch issuing a statement that it was “due to the overwhelming crowd” and that this decision was taken with local authorities to ensure the health and safety of staff and customers.The Swatch X AP Royal Pop that’s caused such a furor at launch, just like the MoonSwatch did in 2022.Courtesy of SwatchThen it was Dubai. “In view of safety considerations, we have decided not to proceed with the sale of the [Swatch x Audemars Piguet Royal Pop] at Dubai Mall and Mall of the Emirates and the event has been cancelled,” a statement said. Authorities in New Delhi and Mumbai closed Swatch Store openings citing a lack of crowd control threatening safety. “We are not animals,” crowd members were caught shouting on Instagram clips.Londonstores shut, andpolice dogs were deployed. Elsewhere in the UK, in Birmingham, news reports spoke of “fights and knives” and drug-taking before police were called. A“mosh pit”developed outside stores in the Netherlands. Clashes inDusseldorf. Officersfired teargasto control a 300-strong crowd outside a Swatch shop in Paris. Openbrawling occurredin Milan. Thenarrestsand shutdownin Miami. Houston’s store closed, as well asChicago’s. In Long Island, New York, amazingly,pepper spraywas usedfor crowd control. In store-front interviews, customersmade clear their angertowards Swatch on how poorly the release was handled.Eventually, around 2 pm ET, a staggering 17 hours after the first clear warning signs in Singapore, Swatch finallyissued a statementon its Instagram page: “To all our dear fans worldwide of our AP x Swatch collab, launched on May 16. To ensure the safety of both our customers and our staff in Swatch stores, we kindly ask you not to rush to our stores in large numbers to acquire this product. The Royal Pop Collection will remain available for several months. In some countries, queues of more than 50 people cannot be accepted, and sales may need to be paused.”Looking at almost any of the 45,000 comments on the post, the response to the brand from fans is crystal clear: Swatch had done far too little, far too late. “You intentionally put people in danger causing a frenzy knowing how many watches you sent to each location but allowing thousands of people to line up was negligent.” “Shame on you Swatch.” “Swatch did bad—they should have gotten ahead of this knowing the depth of inventory and allocation rather than putting stores and customers at risk”, and on the replies go.AP Wanted a Safe, Positive ExperienceWIRED contacted Swatch asking why, considering the wealth of evidence from the ill-received MoonSwatch launch, did it repeat the rollout strategy with the AP collaboration. We got this response from their communications team in Switzerland bl

John BrandonGearMay 18, 2026 10:33 AMReview: Leica Cine Play 1Leica’s first home entertainment projector is pricey, but like the company’s cameras, the image quality is worth the splurge.Courtesy of B&H PhotoTriangleUpBuy NowMultiple Buying Options Available$3,795$2,995at B&H$3,795$2,995at Leica$3,795$2,995at CrutchfieldCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyRating:9/10Open rating explainerInformationWIREDLeica’s Summicron lens renders colors beautifully. Compact size allows for flexibility in moving around the home. Straightforward setup; auto-sizing and focus were almost magical.TIREDLooks best in a dark room. Remote has a confusing button layout. Expensive.If your preferenceis to make colors pop on the screen for movie night, the Leica Cine Play 1 delivers with an immersive cinematic experience. This short-throw laserprojectorfrom the company that makes high-end cameras with a cult following ticks almost every box for tack-sharp picture quality, rendering Pixar films and theAvatarseries with vibrant colors and beautiful details.Crafted out of aluminum and glass, Leica’s latest projector borrows many of its design ethos and superb build quality from thecompany’s cameras. It comes with a carrying handle, making it easy to move around the home, and is powered by the intuitive VIDAA operating system. The Leica optical performance makes movies come alive.At $2,995 after an $800 rebate, the Cine Play 1 is competitive with midrange home cinema projectors from Sony, Epson, and others. The only minor ding against the Cine Play 1 is that you’ll need a mostly dark room to take advantage of Leica’s brilliant color reproduction, while competitors like theEpson Pro Cinema LS9000perform well in both dark and well-lit environments.Setting Up the ProjectorThe Cine Play 1 uses a Leica Summicron lens with an outstanding optical zoom. Known for its extreme sharpness and near-zero color fringing on Leica’s storied cameras, the Summicron glass delivers similar benefits on the Cine Play 1. With the Cine Play 1, it’s like having a high-end camera that, instead of capturing a scene, projects one with added vibrancy. Though Leica has made optical lenses for other projectors, this partnership with Hisense makes this the first projector to bear Leica’s branding, complete with the iconic red dot on the top.Photograph: John BrandonThough seemingly mobile at about the size of a large lunch box and weighing 14.6 pounds, the lack of a built-in battery means that the Cine Play 1 is mostly confined to indoor use. Like most other projectors, you can mount the Cine Play 1 on your own stand or opt for Leica’s optional $495 Cine Play 1 Floor Stand. Though expensive, the Floor Stand has an ingenious design that is Apple-like in approach. After mounting the projector to the stand with a couple of bolts, I noticed a proprietary connection, allowing you to connect the power cord into the base of the stand rather than on the rear of the Cine Play 1, solving the problem of unwieldy wires.After mounting the Cine Play 1, I set up the projector roughly 12 feet from the wall. This resulted in a 150-inch image. The Cine Play 1 can project 65- to 300-inch images, but for smaller picture sizes, it would be easier to get a4K TVinstead. No matter how large the projection I was testing, the focus, brightness, and color quality were all above average compared to other projectors.The auto keystone and autofocus features were almost magical, adjusting to my projector screen automatically. I did have to turn off the auto-adjusting after bumping into the projector a few times, but the idea is that it will adjust if you move the Cine Play 1 to a different room.Photograph: John BrandonPowering on the Cine Play 1, I was pleasantly surprised at how vibrant and colorful everything looked in a dark room.The onscreen setup took only a few minutes. The Cine Play 1 uses Hisense’s VIDAA operating system, so I created a new profile and logged into the companion app. At the top of the screen, there’s a series of banners that allow you to access the app store, art mode, and scan a QR code to install the mobile app. I found the OS to be highly intuitive. There’s also a highly effective voice search option if you would rather not rely on the included remote.VIDAA provides access to most streaming services, with Fandango at Home being a notable exception. (Leica reps stated that they may add this service at a later date.)Benchmark Results and First ImpressionsApart from minor brightness issues when using the Cine Play 1 in brighter environments, this projector passed my benchmark tests. Colors were vibrant, skin tones did not appear washed out, and a butterfly scene showed subtle differences between the shades of yellows and oranges in my test clips. In comparison, colors on theTCL NXTVISIONwere noticeably duller. The Cine Play 1 also excels as a gaming display, with superb color reproduction and stutter-free performance when tested withCrims

Paresh DaveMaxwell ZeffBusinessMay 18, 2026 1:40 PMElon Musk Loses Landmark Lawsuit Against OpenAIThe nine-member panel took only two hours to return a verdict in favor of OpenAI on Monday, which the judge quickly adopted as her own final decision.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyElon Musk sufferedthe worst defeat possible inhis legal battle against OpenAIas a federal jury and a judge ruled he waited too long to bring his claims against the AI startup and its top executives, Sam Altman and Greg Brockman.While the jury’s decision was a nonbinding recommendation sent to US district judge Yvonne Gonzalez Rogers, she immediately accepted it as her own, making it final.Musk’s lead trial attorney, Steven Molo, told the judge, “Our intention is to appeal.”One of his other attorneys, Marc Toberoff, gave a one-word comment to reporters walking out of the courtroom: “Appeal.” He later said the verdict reminded him of American Revolutionary War moments like the Siege of Charleston and the Battle of Bunker Hill. “These were major losses for Americans, but who won the war?” Toberoff said. “And this one is not over.”OpenAI’s attorneys hugged in the courtroom after the verdict was read. William Savitt, the company’s lead litigator, told reporters that the “overwhelming” amount of evidence presented in the case allowed the jury to act quickly. “The evidence that Mr. Musk’s lawsuit was an after-the-fact contrivance by a competitor was overwhelming,” he said.Throughout the trial, Gonzalez Rogers questioned Musk’s motivation for fighting OpenAI. But she concluded on Monday that the three-week global public spectacle had been worthwhile.“I thought it was an important issue to be tried … for us to have a trial to bring clarity,” she told attorneys for both parties. “There’s a substantial amount of evidence to support the jury’s findings, which is why I was prepared to accept the jury’s findings and dismiss on the spot.”Thenine-member paneldelivered the unanimous verdict in an Oakland, California, courtroom on Monday after deliberating for under two hours. They found that statutes of limitations expired well before Muskfiled his lawsuitin 2024. Musk had hoped to persuade the jury that Altman and Brockman, with the help of Microsoft’s cash, transformed OpenAI into an enormous company well beyond what was envisioned when the three of them and others founded it as a nonprofit nearly 11 years ago.Because the jury found the case wasn’t filed on time, it didn’t weigh in on Musk’s three claims, including breach of charitable trust, unjust enrichment, and, against Microsoft, aiding and abetting. Losing on what amounts to a technicality could provide Musk an opening to keep trying his case in the public by arguing that the jury never ruled against his core argument that a charity was stolen.Savitt, the OpenAI lawyer, disputed that contention on Monday. “It’s not a technical decision, it’s a substantive one,” he said. “It says you brought your claims too late, and you did it because you were sitting on them to use them as a weapon of a competitor who can’t compete in the marketplace, and so we’re delighted with the outcome.”Microsoft spokesperson Alex Haurek said in a statement that the “facts and the timeline in this case have long been clear” and the tech giant remains “committed to our work with OpenAI to advance and scale AI.”Inpostson X, Musk called Gonzalez Rogers an “activist” judge “who simply used the jury as a fig leaf” and ruled on a “calendar technicality.” While the judge’s ruling doesn’t create a formal legal precedent, Musk claimed that, “She just handed out a free license to loot charities if you can keep the looting quiet for a few years!” He added in a separate post, “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity.”Despite the disappointing result for Musk, the trial appears to have tarnishedthe public image of OpenAI and its top executives. New details emerged aboutBrockman’s wealthandAltman’s alleged history of dishonesty. Both were also pulled away from their day-to-day work for tens, if not hundreds, of hours to conduct depositions, prepare to testify, sit on the witness stand, and show face in court.Musk spent much less time in the courtroom than the OpenAI executives, about three days before never returning again. He even flew to China for President Donald Trump’s state visit last week, though he technically could have been called to testify again on short notice. “I will say that it was a surprise to us to see that,” Savitt told the media last week. “Instead of being in the jurisdiction where he filed the lawsuit, ready to come in front of the jurors who he has caused to be impaneled, [he] decided to get on Air Force One and go to China.”Though the casecarried financial and emotional stakes, it was also a competition of bravado between two t

Molly HigginsGearMay 18, 2026 2:06 PMThe Best Automatic Cat FeedersWe tested some of the most popular automatic dry- and wet-food pet feeders to see which ones are worth the money.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyFeatured in this articleBest AI-Enabled Dual Feeder With CameraPetkit YumShare Dual-Hopper 2 Automatic Feeder With CameraRead more$170PetkitBest Automatic Wet-Food FeederPetlibro Polar Wet Food FeederRead more$170PetlibroBest Connected AppPetlibro Granary Smart Camera FeederRead more$124$107(14% off)AmazonA Dual Automatic Feeder With a CameraPetkit YumShare Dual-Hopper Automatic Feeder With CameraRead more$160$110(31% off)WalmartAn automatic catfeeder just may change your life. No more screeching at 4 am. Plus, it can give some peace of mind while you’re away from your fur baby. As seen with the rise ofautomatic litter boxesandpet tracking, technology has made it so we can closely monitor our pets without having to do a lot of the manual labor normally involved.Gravity-fed automatic dry-food dispensers were invented nearly a century ago, but today most automatic cat feeders utilize Wi-Fi and are connected to an app on your phone that allows you to make a schedule or feed on demand from afar, all while monitoring your pet’s eating habits. Newer models have taken this tech further, like thePetkit YumShare Dual-Hopper 2 Automatic Feeder With Camera ($170)which has a built-in, AI-enabled camera to analyze the remaining food in the bowl and will stop dispensing to prevent overfeeding. Automatic wet-food dispensers—our top pick in this category is thePetlibro Polar Smart Wet Feeder ($170)—are relatively new, often using ice packs or mini-refrigeration cooling systems to ensure the food stays as fresh as possible.Automatic feeders are also helpful for pet parents who want to monitor their pets’ health, have specialized schedules, or want to foster healthier eating habits. (A customizable automatic cat feeder can help with the endless gorging-vomiting cycle many cats struggle with.) Nearly all of the automatic cat feeders on this list are marketed toward cats, but some can be used with smaller dogs, and I’ve noted these below.Check out our related guides, like theBest Automatic Litter Boxes,Best Cat Water Fountains,Best Smart Cat Collars, andBest Pet Cameras. More of a dog person? We’ve got you covered; see ourBest Accessories and Tech Essentials for Your Dog.Updated May 2026: We’ve added the Cheerble Match G1 and Oneisall 2-in-1 Automatic Cat Feeder and Water Dispenser, and updated prices and links throughout.Jump to SectionAccordionItemContainerButtonLargeChevronCompare Our PicksOthers TestedDo Not RecommendWhy Use an Automatic Feeder?How I TestedHow Long Can I Leave My Cat Alone?Best AI-Enabled Dual Feeder With CameraPhotograph: Molly HigginsCourtesy of PetkitChevronChevronSave to wishlistSave to wishlistPetkitYumShare Dual-Hopper 2 Automatic Feeder With Camera$170Petkit$170ChewyThis model is an upgrade on one of my favorite models below, the Petkit YumShare Dual-Hopper Automatic Feeder With Camera, but with added AI features. It has a crystal-clear video feed, night vision, customizable meal calls, and a dual-hopper for two different types of food. Through the connected app, you can set feeding schedules or feed on demand from both tanks, use two-way audio, and watch the feeder from your phone. Unlike most other brands, the AI-enabled camera doesn’t require a subscription and can identify up to 20 pets and log when they’ve eaten individually. The camera also uses AI to analyze the remaining food in the bowl and will stop dispensing to prevent overfeeding.This feeder worked excellently throughout testing, and I like that you don’t need to opt in to the subscription plan to have AI recognize different pets. However, I still think the connected Petkit app isn’t as reliable or as easy to use as Petlibro’s app (above).Compatible food typesDry; freeze-driedAppYesPower sourcePower cordCapacity2 and 3 liters (5 total)Warranty1 year limitedWIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDBuilt-in AI-enabled cameraCan store two types of food (including freeze dried)TIREDNeed membership plan to access some featuresApp isn’t as user-friendly as some othersBest Automatic Wet-Food FeederPhotograph: Molly HigginsCourtesy of AmazonChevronChevronSave to wishlistSave to wishlistPetlibroPolar Wet Food Feeder$170Petlibro$170Amazon$170ChewyThis updated Petlibro Polar Wet Food Feeder solves pretty much every gripe I had with other automatic feeders aimed at storing wet food—it’s essentially a mini-fridge guaranteed to keep food fresh, and it’s connected to an app where you can schedule and customize feedings and keep a feeding log. Food is given in three stainless steel wide, shallow bowls, which means less chance forwhisker fatiguewith wide bowls orfeline acnewith bacteria-harboring plastic. When it’s time to feed, the cover slides back from the hole and reveals the food be

GamingNewsPlayStationSony is raising short-subscription prices for PlayStation PlusOne-month and three-month PlayStation Plus subscriptions are getting more expensive.One-month and three-month PlayStation Plus subscriptions are getting more expensive.byJay PetersMay 18, 2026, 4:24 PM UTCLinkShareGiftJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Sony ishiking the starting priceof one-month and three-month PlayStation Plus subscriptions in “select regions,” blaming “ongoing market conditions.” Beginning May 20th, 1-month subscriptions will start at $10.99 USD / €9.99 EUR / £7.99 GBP, while 3-month subscriptions will start at $27.99 USD / €27.99 EUR / £21.99 GBP.In the US, the changes result in a $1 price increase for a one-month Essential tier subscription and a $3 increase for a three-month Essential tier subscription. It’s unclear if other PlayStation Plus tiers will see price hikes as well. PlayStation didn’t immediately reply to a request for comment.If you are a current subscriber, prices will stay the same for you (except in Turkey and India, according to Sony) unless you change your subscription tier or let your subscription lapse.The PlayStation Plus price hikes follow two recentpriceincreasesfor PS5 consoles. Sony said its April PS5 price hike wasdue to“continued pressures in the global economic landscape.”Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGamingNewsPlayStationMost PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechDesignNewsGoogle is rolling out its redesigned Workspace app iconsThe new designs feature soft color gradients, rounded corners, and overhauled icon shapes.The new designs feature soft color gradients, rounded corners, and overhauled icon shapes.byStevie BonifieldMay 18, 2026, 4:29 PM UTCLinkShareGiftImage: The VergeStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.It’s not just you — the Google Workspace apps are getting a new look. The redesigned app icons,leaked last month, are now rolling out widely, as we started noticing this morning. Users with the redesigned icons will notice they now have a gradient look that fades from lighter to darker shades, rather than being the same flat tone throughout, similar to the redesigned Google logothat launched a year ago.Image: The VergeSome of the icons switched from a rainbow design to a single color, like Google Chat, Meet, and Calendar, which could help all of the icons stand out a bit more from one another — or make them harder to recognize. Others haven’t changed as much, as9to5Googlenotes, like Google Docs, Sheets, and Slides, which have the same colors and general approach. Asone Reddit userpointed out, though, the Sheets and Slides icons were switched to landscape mode, which is also how most people use both of those apps.Image:9to5GoogleThe Google Drive icon also got a significant overhaul with rounded corners (and it lost the spot of red it had in the bottom right corner). A few also lost the borders around them, like Google Keep, which is now just a yellow light bulb, rather than a bulb on a yellow rectangle. As someone who frequently uses Keep, it’s going to take me a while to get used to spotting a totally different icon shape, although I like the overall look of the new one. I also appreciate the cleaner look of the Gmail icon, although it hasn’t changed too much compared to the old version.The new icons are arriving right beforeGoogle I/O kicks off on Tuesday,where Google could announce even more visual changes to its ecosystem.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAppsDesignGoogleNewsTechMost PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AINewsPolicyElon Musk loses his case against Sam AltmanThe Musk v. Altman jury dismissed all claims due to the statute of limitations.The Musk v. Altman jury dismissed all claims due to the statute of limitations.byHayden FieldMay 18, 2026, 5:39 PM UTCLinkShareGiftImage: Alex Parkin / The VergePart OfAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIsee all updatesHayden Fieldis The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.After around two hours of deliberation, the jury has reached a unanimous verdict inMusk v. Altman, the tech trial of the year. The group found that two claims were barred by the statute of limitations, and a third failed thanks to the dismissal of one of these.The jury here is an advisory jury, meaning the group is installed solely to offer another opinion to the judge, and its verdict is technically not legally binding. Ultimately, US District Judge Yvonne Gonzalez Rogers is the ultimate legal authority — and she accepted the decision.The jury found that Musk’s claim for breach of charitable trust was barred by the statute of limitations, and the claim that Microsoft aided and abetted such a breach failed with it. Restitution is also barred by the statute of limitations, the jury found.OpenAI did not immediately respond to a request for comment. On X, Muskposteda statement saying he’d be filing an appeal. He said that the judge and jury ruled “on a calendar technicality” rather than the “merits of the case” and that “there is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!”Alex Haurek, a Microsoft spokesperson, said in a statement, “The facts and the timeline in this case have long been clear, and we welcome the jury’s decision to dismiss these claims as untimely. We remain committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”RelatedMusk v. Altman proved that AI is led by the wrong peopleMusk vs. Altman is here, and it’s going to get messyMusk v. Altmanhas taken over a federal courtroom in Oakland for three weeks, with the core accusation being that OpenAI strayed from its founding mission and that Musk’s money was earmarked for a nonprofit in particular. Musk alleges that OpenAI CEO Sam Altman and company president Greg Brockman breached OpenAI’s charitable trust and participated in unjust enrichment at Musk’s expense. He also alleges that Microsoft aided and abetted the two in breach of charitable trust. Both sides have used every opportunity to smear each other — and throughsalacious evidenceandeyebrow-raising testimony, both sides have come out looking somehow even less trustworthy than when the court process began.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Hayden FieldAIElon MuskLawMicrosoftNewsOpenAIPolicyTechMore in:All of the updates from Elon Musk and Sam Altman’s battle over OpenAIMusk v. Altman proved that AI is led by the wrong peopleHayden FieldMay 18The jury has delivered a unanimous verdict.Richard LawlerMay 18An observer has just been ejected from the court by the US marshals.Elizabeth LopattoMay 18Most PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechUpdatedMay 18, 2026, 7:00 PM UTCAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIbyElizabeth LopattoandHayden FieldLinkShareRSSGiftSam Altman and Elon Musk are facing off in a high-stakes trial that could alter the future of OpenAI and its most well-known product,ChatGPT. In2024, Musk filed a lawsuitaccusing OpenAI of abandoning its founding mission of developing AI to benefit humanity and shifting focus to boosting profits instead.After nearly a month of the hearing that featured testimony from Musk, Altman, Microsoft CEO Satya Nadella, OpenAI cofounder Greg Brockman, former OpenAI board member and mother of several of Musk’s children Shivon Zilis, and a few others, the jury deliberated for a couple of hours before returning with a verdict,deciding to dismiss all chargesdue to the statute of limitations.Musk was a cofounder of OpenAI and claims that Altman and Brockman tricked him into giving the company money, only to turn their backs on their original goal. However,OpenAI claimedthat “This lawsuit has always been a baseless and jealous bid to derail a competitor” in a bid to boost Musk’s own SpaceX / xAI / X companies that have launched Grok as a competitor to ChatGPT.In his lawsuit,Musk asked for the removal of Altman and Brockman, and for OpenAI to stop operating as a public benefit corporation.People to KnowPlaintiffElon Musk— plaintiff, OpenAI cofounder and now CEO of rival xAISteven Molo— lead counsel for the plaintiffJared Birchall— manager of Musk’s family officeShivon Zilis— former OpenAI board member who shares multiple children with MuskDefendantSam Altman— defendant, CEO of OpenAIWilliam Savitt— lead counsel for the defendantGreg Brockman— president of OpenAI as well as a cofounderIlya Sutskever— former chief scientist at OpenAI and a cofounderJudgeYvonne Gonzalez Rogers— aka YGR, trial judgeHere’s all the latest on the trial between Musk and Altman:HighlightsElon Musk loses his case against Sam AltmanMicrosoft doesn’t want any of thisSam Altman was winning on the stand, but it might not be enoughMusk’s biggest loyalist became his biggest liabilityMira Murati tells the court that she couldn’t trust Sam Altman’s wordsGoogle’s AI architect lived rent-free in Elon Musk’s headOpenAI’s president does ‘all the things,’ except answer a questionElon Musk’s worst enemy in court is Elon MuskElon Musk appeared more petty than preparedMusk vs. Altman is here, and it’s going to get messyPinnedMay 18Hayden FieldElon Musk loses his case against Sam AltmanImage: Alex Parkin / The VergeAfter around two hours of deliberation, the jury has reached a unanimous verdict inMusk v. Altman, the tech trial of the year. The group found that two claims were barred by the statute of limitations, and a third failed thanks to the dismissal of one of these.The jury here is an advisory jury, meaning the group is installed solely to offer another opinion to the judge, and its verdict is technically not legally binding. Ultimately, US District Judge Yvonne Gonzalez Rogers is the ultimate legal authority — and she accepted the decision.Read Article >May 18Hayden FieldMusk v. Altman proved that AI is led by the wrong peopleImage: Cath Virginia / The Verge, Getty ImagesThe tech trial of the year,Musk v. Altman, was ultimately a fight for control. Elon Musk argued that Sam Altman, with whom he helped found the now-massive company OpenAI, shouldn’t direct the future of AI. Altman’s lawyers, in turn, poked at Musk’s own credibility. A jury came to averdicton Monday after just two hours of deliberation, dismissing Musk’s claims due to the statute of limitations.In a strictly legal sense, three weeks of testimony added up to nothing. But the trial offered a more damning broader takeaway: Almostnobody in this saga seems worth trusting. Some of the most powerful people in tech seem temperamentally incapable of dealing with each other honestly. And if that’s true, it raises a bigger question: Why are they in control of atrillion-dollarindustry that’s set to upend people’s lives?Read Article >May 18Richard LawlerThe jury has delivered a unanimous verdict.That was quick (about two hours).Elon Musk lost his case against Sam AltmanHayden FieldMay 18May 18Elizabeth LopattoAn observer has just been ejected from the court by the US marshals.I assume because he was recording, since the marshal said, “Give me your phone.” There have been several incidents of people attempting to record or take pictures throughoutthe trial— but I honestly am not sure why you’d record today of all days.May 18Elizabeth LopattoC. Paul Wazzan is the expert called by Musk to determine damages.Unfortunately he does not have a lot of details YGR is asking for. He doesn’t know how many investments Musk has made (11 to date, according to Pitchbook), or how many were successful. He’s getting some pretty tough questioning from YGR in the direct exam. Among the things she’d asked, he didn’t know how many startups fail in Silicon Valley, or the success r

GadgetsTechSmart HomeDyson’s super-slim PencilWash just hit its best price to date for Memorial DayThe slim hard-floor cleaner slides under furniture, scrubs away spills, and removes dirty water as it cleans.The slim hard-floor cleaner slides under furniture, scrubs away spills, and removes dirty water as it cleans.bySheena VasaniMay 18, 2026, 6:32 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: DysonSheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.IfDyson’s PencilVac Fluffyconesmade you wish the company had built something similarly slim for scrubbing the hard floors in your home, enter the recently releasedDyson PencilWash. The cordless cleaner is designed to tackle spills and stains on tile and other surfaces, and right now, it’s available for a new low price of $249.99 ($100 off) fromAmazon,Target, andDysonin the run-up to Memorial Day.Dyson PencilWash$250$35029% off$250$250 at Amazon$250 at Target$250 at DysonWith its 1.5-inch-diameter handle and a design that lets it lie nearly flat, the PencilWash can easily slide under furniture and other tricky areas that bigger floor cleaners can’t reach. Built for hard surfaces like tile, laminate, and sealed wood, it uses a pressurized hydration system and a microfiber wet roller to scrub away spills while extracting dirty water with each pass, so you’re always cleaning with fresh water. There’s also a Max mode for addressing dried-on messes that are particularly tough to clean, along with a flexible neck that makes it easier to steer around obstacles.RelatedDyson put someone else’s motor in its robot vacuumHere are 40 of our favorite deals from REI’s massive Anniversary SaleThese new Roombas are smaller and cheaperThe PencilWash is designed to be low-maintenance, too. Debris and wastewater are collected directly in the head rather than traveling up through the machine, preventing them from becoming trapped in a filter. It shouldn’t leave your floors soaking wet, either, and it comes with a sleek charging dock that lets it eke out 30 minutes per charge (you can also purchase abackup batteryfor $99.99). The 300ml clean water tank should be enough to cover over 1,000 square feet on the lowest hydration setting as well, after which it will need a refill.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Sheena VasaniDealsGadgetsSmart HomeTechVerge ShoppingMost PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechNewsGoogleWalmart launches new budget-friendly Android tablets starting at $97The six new Onn tablets ship with Android 16 and include kids’ tablets, Core models, and a 13-inch Pro model.The six new Onn tablets ship with Android 16 and include kids’ tablets, Core models, and a 13-inch Pro model.byStevie BonifieldMay 18, 2026, 6:44 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: WalmartStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.Walmart’s Onn brand just launched a whole line of budget-friendly Android tablets that, all together, cost less in total than a single iPad Pro. As spotted by9to5Google, the six new tablets ship with Android 16 and all but one cost less than $200.The littleOnn Core 7is the most affordable of the bunch at just $97, which gets you a 7-inch 1040 x 600 IPS LCD display, 4GB of RAM, 64GB of storage that’s expandable with a microSD card, and, according to Walmart, up to 10 hours of battery life.PreviousNext1/3The Onn Pro includes a folio case and stylus.Image: WalmartA step above that are two larger Core tablets. TheCore 8.1costs $138 and includes an 8.1-inch 1524 x 1000 IPS LCD display, 6GB of RAM, 64GB of storage (also expandable with microSD), and up to 15 hours of battery life. TheCore 11for $167 is a bit bigger, similar to a base iPad, featuring an 11-inch 1840 x 1280 IPS LCD display, 6GB of RAM, 128GB of storage that’s expandable with a micoSD slot, and up to 17 hours of battery life. There are also two affordable kids’ tablets available in8-inchand11-inchsizes for $118 and $136, respectively.The priciest of the new tablets is theOnn Pro, which costs $288 and includes a 13.2-inch 2400 x 1600 IPS LCD display, 8GB of RAM, 256GB of storage, and an unnamed 2.6GHz MediaTek processor. Unlike the iPad Pro it’s imitating, it also includes a folio case and a stylus.Onn Core 7 tablet$97$97$97 at Walmart (blue)$97 at Walmart (silver)$97 at Walmart (mocha)Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAndroidGoogleNewsTechMost PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AIPolicyReportMusk v. Altman proved that AI is led by the wrong peoplePublic opinion of the AI industry is already sinking. A parade of untrustworthy executives makes it look worse.byHayden FieldMay 18, 2026, 7:00 PM UTCLinkShareGiftImage: Cath Virginia / The Verge, Getty ImagesAIPolicyReportMusk v. Altman proved that AI is led by the wrong peoplePublic opinion of the AI industry is already sinking. A parade of untrustworthy executives makes it look worse.byHayden FieldMay 18, 2026, 7:00 PM UTCLinkShareGiftPart OfAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIsee all updatesHayden Fieldis The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.The tech trial of the year,Musk v. Altman, was ultimately a fight for control. Elon Musk argued that Sam Altman, with whom he helped found the now-massive company OpenAI, shouldn’t direct the future of AI. Altman’s lawyers, in turn, poked at Musk’s own credibility. A jury came to averdicton Monday after just two hours of deliberation, dismissing Musk’s claims due to the statute of limitations.In a strictly legal sense, three weeks of testimony added up to nothing. But the trial offered a more damning broader takeaway: Almostnobody in this saga seems worth trusting. Some of the most powerful people in tech seem temperamentally incapable of dealing with each other honestly. And if that’s true, it raises a bigger question: Why are they in control of atrillion-dollarindustry that’s set to upend people’s lives?OpenAI was, in the testimony of both Musk and Altman, founded to stop powerful AI from being owned and advanced by the wrong people. Testimony and evidence showed its founding team fretting about who would control artificial general intelligence (AGI), a buzzword for AI that broadly equals or surpasses human knowledge and ability. Theydeeply fearedGoogle DeepMind and its leader, Demis Hassabis. In 2015, Altman said he’d been mulling over whether anything could “stop humanity from developing AI” — and after concluding it was impossible, that he wanted “someone other than google to do it first.”Fellow cofounders Greg Brockman and Ilya Sutskever so strongly opposed one-person control that they seemed willing to torpedo a lucrative deal that could — in their words — give Musk an “AI dictatorship.” In a part of the same email addressed to Altman, Brockman and Sutskever questioned his motivations, writing, “We haven’t been able to fully trust your judgements throughout this process … Is AGI truly your primary motivation? How does it connect to your political goals?”These concerns would bequickly borne out. A central focus ofMusk v. Altmanwas “the blip,” a five-day period in November 2023 when OpenAI’s board removed Altman as CEO. Sutskever had spentmore than a yeararchitecting his ouster, assembling a 52-page memo alleging “a consistent pattern of lying, undermining his execs, and pitting his execs against one another.” The implications were broader than executive infighting, potentially impacting the public rollout of AI systems. Then-CTO Mira Murati, for instance, testified in court that Altman told her OpenAI’s legal team had okayed skipping a safety review for one of its models — a statement, she said, that turned out to be false.In closing arguments, Musk attorney Steven Molo hammered home the long list of people who had testified under oath that Altman was, in one way or another, a liar — all of whom Altman had worked with for years. “The defendants absolutely need you to believe Sam Altman,” Molo told the jury. “If you cannot trust him, if you don’t believe him, they cannot win. It’s that simple.”But during court proceedings, Musk — who now leads competing lab xAI, under his space company SpaceX — didn’t come off any better. Joshua Achiam, now OpenAI’s chief futurist, testified that Musk’s race against Google led him to take an “obviously unsafe and reckless” approach to achieving AGI. When he and others raised concerns, he says, Musk argued that OpenAI’s for-profit makeover created incentives to disregard safety, but his own xAI is for-profit and has, at best, ahaphazard approach to safety. And in the name of making sure OpenAI remained open, Musk was obsessive in his need for control over it. In closing arguments, Sarah Eddy, one of OpenAI’s attorneys, told the jury that Musk “wanted dominion over AGI.”As one X userput it, “if untrustworthyness had mass, putting Musk and Altman too close to one another would collapse the courtroom and all of earth into a black hole.”OpenAI did not immediately respond to a request for comment. On X, Muskposteda statement saying he’d be filing an appeal.It’s not just Musk and Altman, either.Trial evidence suggestedMurati helped get Altman removed, then switched sides to support his reinstatement while appearing “totally uninterested” in disclosing the role she’d played. Shivon Zilis, a close Musk associa

GamingNewsPC GamingPlayStation exclusives aren’t coming to PC anymoreA big pivot back to consoles.A big pivot back to consoles.byJay PetersMay 18, 2026, 7:28 PM UTCLinkShareGiftImage: Sony Interactive EntertainmentJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Sony reportedly won’t release its major single-player PlayStation games on PC anymore. According toBloomberg’s Jason Schreier, Hermen Hulst, who heads up PlayStation’s studios business,informed employeesin a town hall on Monday about the change in strategy. Schreier had previously reported on the shiftin March, saying that Sony scrapped plans to launch PC versions of last year’sGhost of Yōteiand “other internally developed games.” Online games will still come to multiple platforms following this change in strategy, Schreier reported at the time.In recent years, Sony has released many of its biggest games on PC, includingSpider-Man 2,Ghost of Tsushima, bothThe Last of Usgames,Horizon Zero Dawn Remastered, and multiplayer titles likeHelldivers 2andMarathon. Two years ago, Hulstcommitted toreleasing PlayStation’s live-service games “day and date” on PC and PS5, but its single-player PC releases have been less consistent, with Hulst saying that the company takes a “more strategic approach.”RelatedConsole exclusives might be making a comebackSony may not be the only one to go back to exclusives; Asha Sharma, Microsoft’s new Xbox chief,told staffthat she is “reevaluating” exclusive games for the platform.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGamingNewsPC GamingPlayStationMost PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TransportationNewsCarsVolvo teases a new affordable EV to replace discontinued EX30The new EV won’t be the same price as the EX30, which was supposed to start at $35,000, but ended up being much more than that thanks to tariffs.The new EV won’t be the same price as the EX30, which was supposed to start at $35,000, but ended up being much more than that thanks to tariffs.byAndrew J. HawkinsMay 18, 2026, 8:41 PM UTCLinkShareGiftImage: VolvoPart OfThe great EV pullback: all the obstacles, cancellations, and delayssee all updatesAndrew J. Hawkinsis transportation editor with 10+ years of experience who covers EVs, public transportation, and aviation. His work has appeared in The New York Daily News and City & State.Volvo’s compact, quirky EX30 had a lot of problems when it was first released. Tariffs essentially erased its affordability, making it more expensive to own, and a battery recall made it dangerous to park indoors. Butits discontinuationdidn’t spell the end of Volvo’s efforts to sell more affordable electric models. In fact, the Swedish automaker is already at work on a new offering for the US market.The news of an affordable Volvo EV for the US came during a media roundtable this week related tothe US launch of the new EX60. Luis Rezende, president of Volvo Cars America, said that the decision to discontinue the EX30 was not solely about tariffs and profitability, noting that the company is preparing to introduce a new EV in 2027 that will occupy a similar role in the lineup — though not necessarily at exactly the same price point as the EX30.“Very similar, I would say,” Rezende said about the mystery EV’s price comparison to the EX30. “It’s going to be an EV that will deliver a lot of good things in a bigger space, but it will be also fun to drive, I can promise you.”Other than that, details were scarce. Volvo’s executives talked later about the desire to build a larger, family-oriented SUV at its factory in Charleston, South Carolina — though that vehicle will likely use a “multi-fuel” strategy rather than being exclusively electric from launch.The EX60, which will start customer deliveries in the US this summer, is Volvo’s attempt at a reset in the US. The compact SUV, which is built on a different architecture than the EX30, will start at $59,795 for the entry-level P6 Plus version, and climbs up to $68,745 for the more powerful P10 AWD Ultra variant.Volvo is the latest automaker to try, and stumble, in its efforts to build an affordable EV for the US market that is both desirable and profitable for the company. To date, few have pulled it off, as it requires a certain level of scale, vertical integration, and mastery of the supply chain that only companies in China seem to have really nailed down. Of course, Volvo is owned by China’s Geely, but the company’s desire to sell EVs in North America will necessitate a different approach to affordability.Affordability was one of the EX30’s main selling points. When it wasfirst announcedin 2023, Volvo said the price would start at $34,950, positioning it as the smaller, less expensive EV that many people were clamoring for. But after the election of Donald Trump,Volvo was forced to delay the EX30’s arrival in the USuntil 2025, citingnewly leveled tariffsagainst vehicles built in China. Eventually, the model that went on sale in the US started at $44,900, about $10,000 more than the original price.Then, in February, further bad news as Volvo issued a recall for the EX30 because the vehicles’ batteries were at risk of overheating or catching on fire. The next month, Volvo pulled the plug on the vehicle in the US.Correction May 18th:A previous version of this story stated that the EX60 is the only Volvo EV in the US. The EX90 is also available.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Andrew J. HawkinsCarsElectric CarsNewsTransportationVolvoMore in:The great EV pullback: all the obstacles, cancellations, and delaysHonda’s hybrid future starts with new Accord and RDX prototypesStevie BonifieldMay 14Mazda delays its first EV by two years.Andrew J. HawkinsMay 12Ford’s EV and software chief Doug Field is leaving the companyAndrew J. HawkinsApr 15Most PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

GamingNewsTechLG will release the first 1000Hz, 1080p gaming monitor this yearThe 25G590B is a 24.5-inch IPS screen with the highest refresh rate we’ve seen for a 1080p gaming monitor.The 25G590B is a 24.5-inch IPS screen with the highest refresh rate we’ve seen for a 1080p gaming monitor.byRichard LawlerMay 19, 2026, 1:10 AM UTCLinkShareGiftIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: LGRichard Lawleris a senior editor following news across tech, culture, policy, and entertainment. He joined The Verge in 2021 after several years covering news at Engadget.If you just can’t choose between refresh rate and resolution, LG’s next gaming monitor could solve your problem, as the UltraGear 25G590B monitor is the first one announced that will be capable of a native 1000Hz refresh rate at 1,920 x 1,080 resolution. So far, the 1,000Hz models we’ve seen have only been capable of 720p at the most, but this 24.5-inch IPS display does not have that limitation, as the company says it’s ready for esports competitors (whether they’re in an FPS orExcel), who need maximum responsiveness.We were alreadyasking if anyone could feel the effect of higher refresh rates a few years agowhen they hovered around 360 and 480Hz, but now the bar has moved higher once again.LG hasn’t revealed the 25G590B’s price or release date beyond “second half of 2026,” but said it has a “minimalist” stand, an integrated hook for headset storage, and customizable lighting. More information will be availableon LG’s website here.Of course, there are also AI-enabled features, too:The 25G590B incorporates on-device gaming AI features designed to boost both immersion and usability. AI Scene Optimization adds visual realism and depth by intelligently adjusting picture settings according to game genre, while AI Sound presents a more authentic spatial audio experience along with clearer in-game communications (when using a compatible headset). These AI-driven capabilities complement the monitor’s ultra-fast performance to create the ideal platform for competitive gaming.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Richard LawlerGamingLGNewsPC GamingTechMost PopularMost PopularBad, Bad, Bad, Bad RoommateMicrosoft is retiring Teams’ Together ModeDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk loses his case against Sam AltmanMicrosoft starts canceling Claude Code licensesThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

When Lior Susan startedEclipse Venturesin 2015, the firm’s thesis of digitizing the physical world wasn’t particularly popular in Silicon Valley.“It was the era of enterprise software and SaaS, and it felt fairly lonely the first couple of years,” Susan said onstage at a recent StrictlyVC event in San Francisco.More than a decade later, Eclipse finds itself at the center of the tech world’s action. The firm’s $6.5 million Series A investment in Cerebras Systems in 2016 paved the way for a total return of $2.5 billion when the semiconductor company went public this week. The firm invested a total of $147 million in Cerebras over time, a bet that generated a seventeenfold return at theIPO price of $185per share, according to Eclipse.For Susan, the windfall from Cerebras is only the beginning of reaping big rewards from a long-standing belief that because 85% of global GDP is tied to the physical world, investing in companies beyond pure software could be immensely lucrative.Public markets and startup founders seem to be recognizing the value of physical-world tech now, too. Susan noted that shares of TSMC and Micron recently hit all-time highs, while a growing cohort of elite founders is eager to build startups at the intersection of hardware and software.“I think people understand that the real moat in software is gone. You can vibe-code pretty much whatever you want,” he said.Susan echoed public market sentiment that earlier this year sent many SaaS stocks tumbling on the belief that enterprises may use Anthropic’s Claude Code or OpenAI’s latest models to create their own bespoke software tools instead.“What you cannot do with ‘vibe code’ is manufacture wafers, because you need machines and silicon, and they need clean rooms, and a bunch of other things,” Susan said.When it comes to the tech that touches the physical world, it’s not just semiconductors that are suddenly catching the attention of investors and founders.Eclipse’s portfolio companies spanning sectors like robotics, energy and defense, raised nearly $15 billion from outside backers last year, and that late-stage momentum reached $4.5 billion in Q1 2026 alone, Susan said. That investor excitement stands in stark contrast to the firm’s early track record: In its first eight years, its portfolio companies raised less than $4 billion in total.Indeed, the recent follow-on rounds across Eclipse’s portfolio show a track record that any venture firm would envy. Driven by a string of massive late-stage deals this year, the haul includes$1.2 billion for Wayve, $650 million for True Anomaly, $270 million for Bedrock Robotics, and $200 million for Oxide Computer. What’s more, Eclipse was the Series A investor for all four companies.At first glance, it may seem that investor enthusiasm for physical-world tech is driven purely by AI, whether as an infrastructure input like chips and data centers, or through AI’s power to finally make robotics viable. However, Susan argues that there are other powerful tailwinds driving the momentum.Besides technology — in this case, AI — what’s important for this market to thrive is capital, customer demand, talent, and policy. Susan means that along with investors and engineers moving away from SaaS to sectors like robotics, semiconductors, space, and mining, the U.S. government is also encouraging these industries through subsidies and favorable regulation.“This is the first time I believe in America ever, from Henry Ford and Carnegie, those five forces are aligned,” Susan said. “For builders like us, this is the best time to build those companies.”TopicsCerebras,Eclipse Ventures,Fundraising,Hardware,Lior Susan,Space,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retrac

Commencement season has come around again — and this year, a couple speakers have discovered that it’s tough to get graduating students excited about a future shaped by artificial intelligence.Last week, Gloria Caulfield, an executive at real estate firm Tavistock Development Company,gave a speech at the University of Central Floridaacknowledging that we’re living in a time of “profound change,” which can be both “exciting” and “daunting.”“The rise of artificial intelligence is the next industrial revolution,” Caulfield declared — prompting the students in the audience to begin booing, getting louder and louder until Caulfield chuckled, turned to the other speakers, and asked, “What happened?”“Okay, I struck a chord,” she said. Caulfield then tried to resume her speech, saying, “Only a few years ago, AI was not a factor in our lives” — only to be interrupted again by the audience, this time by their loud cheers and applause.Former Google CEO Eric Schmidt faced a similar response when he brought up AI at a University of Arizona speech on Friday.In Schmidt’s case, the criticism actually began before the speech itself, with some student groupscalling for him to be removed as commencement speakerdue toa lawsuit in which a former girlfriend and business partner accused Schmidt of sexual assault. (He has denied the allegations.) According toa local news report,the booing began even before Schmidt took the stage.But Schmidt alsogot loud booswhen he told students, “You will help shape artificial intelligence.” The booing was persistent enough that Schmidt tried to speak over it, insisting, “You can now assemble a team of AI agents to help you with the parts that you could never accomplish on your own. When someone offers you a seat on the rocket ship, you do not ask which seat, you just get on.”To be fair, AI isn’t becoming a third rail ateverygraduation ceremony. Nvidia CEO Jensen Huang recentlyspoke atCarnegie Mellon’s commencement, and he didn’t seem to get any audible pushback when he said that AI has “reinvented computing.”Still, it’s not exactly surprising to find some students in a booing mood. Ina recent Gallup poll, only 43% of Americans aged 15 to 34 said it’s a good time to find a job locally, a steep drop from 75% in 2022.That pessimism isn’t solely a response to the rise of AI (a shift thateven some software engineers are worried about), but journalist and tech industry criticBrian Merchant suggestedthat for many students, AI has become “the cruel new face of hyper-scaling capitalism.”“I too would loudly boo at the prospect of this next industrial revolution if I was in my early twenties, unemployed, and had aspirations for my future greater than entering prompts into an LLM,” Merchant wrote.Even when graduation speeches didn’t mention AI explicitly, “resilience”was a recurring theme this year. Schmidt himselfacknowledgedthat there is “a fear in your generation that the future has already been written, that the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create.”Caulfield, meanwhile, might also have misread her audience of arts and humanities graduates. One student said that before mentioning AI, Caulfield already started to lose them with her “generic” praise of corporate executives like Jeff Bezos.Another graduate, Alexander Rose Tyson,told The New York Times, “It wasn’t one person that really started the booing. It was just sort of like a collective, ‘This sucks.’”TopicsAI,commencement speeches,Eric Schmidt,gloria caulfieldWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Anthony HaAnthony Ha is TechCrunch’s weekend editor. Previously, he worked as a tech reporter at Adweek, a senior editor at VentureBeat, a local government reporter at the Hollister Free Lance, and vice president of content at a VC firm. He lives in New York City.You can contact or verify outreach from Anthony by [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. To get this in your inbox, sign up here for free — just clickTechCrunch Mobility!There is a bit of a theme emerging in transportation — and really every industry: AI is creating jobs for some at the loss of others.General Motors, for instance,laid off more than 10% of its IT department, or about 600 salaried employees — in a deliberate skills swap. This won’t translate into a one-to-one exchange, which means there will likely be a net-negative job loss. But GM insists it is hiring and those layoffs have made room for it to recruit IT people with AI-focused backgrounds.The most sought-after capabilities are AI-native development, data engineering and analytics, cloud-based engineering, agent and model development, prompt engineering, and new AI workflows. In practical terms, GM is looking for people who know how to build with AI from the ground up — designing the systems, training the models, and engineering the pipelines — not just use AI as a productivity tool.Those AI job losses are mounting in the automotive sector. CNBCcalculatedthat Ford, GM, and Stellantis have cut a combined total of more than 20,000 U.S. salaried jobs, or 19% of their combined workforces, from recent employment peaks this decade. While there are a variety of reasons for these cuts, they are generally connected to technological changes, including AI.Companies are leaning heavily into AI, although anecdotes from some engineers and founders suggests not all of these businesses know quite what they’re doing with it yet.Samsarais one company that seems to have figured out a revenue-generating use case. The company has spent the last decade giving its customers cameras to mount inside millions of trucks for driver monitoring, theft prevention, and helping with liability claims. The company took that mountain of data and trained its own model that candetect potholesand determine how quickly they are deteriorating. The company is pitching this product to cities and announced it has several under contract, including Chicago.A little birdImage Credits:Bryce DurbinNothing this week, although I am working on a fun one! Reach out anytime with insights, tips, or just because. You can reach us via email or Signal.Email Kirsten Korosec [email protected] my Signal at kkorosec.07, or email Sean O’Kane [email protected]!Image Credits:Bryce DurbinYou might have noticed that Rivian’s spinoff companyMind Roboticsraised another $400 million, just two months afterraising $500 million. And that pace got me thinking about its founder RJ Scaringe and his innate ability to get VC and institutional backers to invest in his ideas and projects.I calculated that investors have poured $12.3 billion into Scaringe’s three startups — Also, Mind Robotics, and Rivian. That figure doesn’t include the close to $12 billion in gross proceeds raised in Rivian’s IPO, nor did I count the more recent strategic deals with Volkswagen Group and Uber — which together could add nearly $7 billion to Rivian’s coffers.You can read my whole riff on the topic here. But if you don’t feel like clicking, here is one item that stood out. I spoke to a number of insiders and investors and they all mentioned Scaringe’s ability to give undivided attention to whoever he’s talking to — whether it’s an investor, supplier, or exec — and make them feel like the most important person in the room.It’s yet another piece of evidence in my long-standing case against multitasking. Debate me!Other deals that got my attention …Arkeus, an Australian startup that developed perception software for autonomous drones and aircraft,raised $18 millionin a Series A round led by QIC Ventures. Other investors include R+VC, Folklore Ventures, DYNE Ventures, Main Sequence Ventures, Salus Ventures, and Beaten Zone.Aseon Labs, a Redwood City, California, startup that has developed a depot in a box for charging, cleaning, and inspecting autonomous fleets,came out of stealthwith undisclosed backing by Y Combinator.Rapidoraised $240 millionin a round led by Prosus, and that values the Indian ride-hailing company at $3 billion. Existing investors, including WestBridge Capital and Accel, participated. The round was part of a larger $730 million primary and secondary financing.Quantum Systems, a Germany-based drone startup backed by Peter Thiel, is in talks to raise around €600 million ($703 million) with companies like Airbus and Blackstone as investors,Bloomberg reported.Notable reads and other tidbitsImage Credits:Bryce DurbinIs Redwood Materials ready for an IPO?Senior reporter Sean O’Kane interviewed the company’s new CFO,Deepak Ahuja, whose name will be familiar to anyone who follows Tesla. Ahuja was Tesla’s former finance chief and most recently held a similar position at drone company Zipline.TeslaRobotaxis have crashed at least twice since July 2025 while ateleoperator was remote

Listen onApple PodcastsListen onSpotifyLawyers for Elon Musk and OpenAI made their closing arguments this week, and now it’s up to jurors to decide whether OpenAI did anything wrong as it’s transformed intoa slightly-more-for-profitorganization.But as Kirsten Korosec, Sean O’Kane, and I noted on the latest episode of TechCrunch’sEquity podcast, a big theme in the trial’s final days waswhether OpenAI CEO Sam Altman is trustworthy— for example, Musk’s attorney Steve Molo grilled Altman about whether statements he’d made during congressional testimony were truthful.Kirsten noted that Musk hasmade plenty of misleading statements of his ownand that trust isn’t just an issue for Altman.“This is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs,” she said. “It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies; there’s a lot behind the veil still.”Keep reading for a preview of our conversation, edited for length and clarity.Anthony Ha:[The end of the trial] led to this really provocative headline from one of our writers, Tim Fernholz, [that] just says,“Who trusts Sam Altman?”Does anyone want to take a stab at answering this?Kirsten Korosec:Yeah, Anthony, I’m going to throw it right back to you. Do you trust Sam Altman?Anthony:It’s an interesting question because it feels like something that’s kind of a wild question to discuss in a journalistic context, but actually that’s the core of the trial, in a lot of ways.Sean O’Kane:That’s not a yes.Anthony:And it actually seems to be [at the] core of understanding so much of what’s happened at OpenAI, especially this big executive power struggle that they now call The Blip.It just seems like a lot of people who’ve worked with Altmandon’t trust him. And he’s acknowledged this a little bit, because he’ll talk about the fact thathe recognizes he’s been conflict averse, telling people what they want to hear, and he’s trying to work on that.I mean, it sounds plausible, and I can understand how that can lead to misunderstandings in some situations. [But] I’m also a very conflict-averse person and I’d like to think that if any of this stuff went to trial, that people would not be asking, “Is Anthony Ha trustworthy?”Sean:Still not a yes!Kirsten:I think that people would say that you are trustworthy. I will say that question, while provocative, doesn’t just encapsulate what this trial was about. I would zoom out even more and say this is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs. It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies; there’s a lot behind the veil still.Maybe when they all IPO, we can get a peek, but it is fundamentally about trust and misuse, and do we believe the intent? And what I would throw back is, sometimes the intent can be worthy, noble, and still misused. It can still end up as a bit of a s— show. I think it’s more than who trusts Sam Altman — although that was very interesting in this trial — but more of that bigger question that we can apply to the entire industry.Sean:I’ll say it: I don’t trust him. But you know, I don’t trust most people, so I guess that’s just the baseline.We’ll see where this goes. The trial wraps up today. I’ve been very curious to hear how the jury decides this all. I think at the start of this, a big motivator of this was Elon Musk trying to sling mud at a perceived rival and someone who he feels slighted him. And I don’t know if we know enough yet to say that that was completely accomplished, and whether or not he has a shot at winning. But I think all these people came out of this looking a little bit worse.Anthony:And just to get specific, why this is coming up this week is that [Altman] was on the stand and he was basically getting grilled about some statements he’s made in the past, in testimony to [Congress], basically saying he didn’t have any equity in OpenAI. And that is not true because he had a stake through Y Combinator, which he used to run. And tried to brush that off by saying, “I assume that everybody understands what it means to be a passive investor in a VC fund.” And I think [Elon Musk’s] lawyer, somewhat fairly, said “Really? You think the congressman who was interviewing you knew that?”Kirsten:Yeah, I mean, he was playing the whole semantics game. What I thought was so interesting about [this] is the style of how Sam Altman answered questions [compared to] Elon Musk on the stand.So Elon Musk, in many, many, many scenarios and many instances, we can point to the fact that he put something out on Twitter that was a lie or a bit of a fib, and on the stand corrected the record. So there’s a history of, I would say, non-truthfulness-slash-lying, blatant or otherwise, in Elon Musk’s world, but how he treated it was incredibly combative an

In BriefPosted:1:15 PM PDT · May 17, 2026Image Credits:Artur Widak/NurPhoto / Getty ImagesAnthony HaApple’s Siri revamp could include auto-deleting chatsPrivacy will be a major theme when Apple unveils a new version of Siri at the Worldwide Developers Conference in June,according to Bloomberg’s Mark Gurman.The Siri relaunch is widely seen as Apple’s big chance to reestablish its relevance in artificial intelligence. As part of that effort, company executives will argue that they’re taking a more privacy-friendly approach than most other AI companies, Gurman said.Apple will reportedly launch the first stand-alone Siri app,powered by Google Geminiand offering users a chatbot experience reminiscent of ChatGPT.  But compared to those other chatbots, the app is supposed to have more limitations on how long user information can be used and stored.For example, Gurman said Siri could include a feature similar to the Messages app, allowing users to automatically delete conversations after 30 days or one year — or to keep them indefinitely.Gurman also suggested that Apple might be emphasizing privacy as a way to excuse Siri’s shortcomings compared to competing products — and that this emphasis might obscure the fact that Google is handling some of the security.TopicsAI,Apple,Apps,gemini,In Brief,siriMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedSpaceOSHA probing worker death at SpaceX’s Starbase siteSean O’Kane5 hours agoAISandboxAQ brings its drug discovery models to Claude — no PhD in computing requiredLucas Ropek6 hours agoAIAnthropic has acquired the dev tools startup used by OpenAI, Google, and CloudflareKirsten Korosec8 hours agoLatest in AIAISandboxAQ brings its drug discovery models to Claude — no PhD in computing requiredLucas Ropek6 hours agoAIAnthropic has acquired the dev tools startup used by OpenAI, Google, and CloudflareKirsten Korosec8 hours agoAIElon Musk has lost his lawsuit against Sam Altman and OpenAITim Fernholz9 hours ago

Imagine you’re riding a motorcycle at 160 kilometers per hour when an arrow appears, floating on the road ahead, telling you exactly where to turn. No phone, no dashboard. Just your helmet, and a lens the size of a thumbnail.This is not a concept video. It’s heading to European roads as early as this year. And it’s one early glimpse of where smart glasses are heading.Over the past few years, Big Tech has been quietly (and not so quietly) placing its bets. Meta has been selling AI-enabledRay-Ban glasses since 2023, Google isbuilding Android XR, andApple is expected to enter the market. Last week, Samsung wasreportedlyset to unveil its first AI-capable smart glasses, co-designed with Gentle Monster, at a Galaxy Unpacked event in London this July. China’sHuawei,Alibaba,Xiaomi, and others are all moving too.The numbers reflect the momentum. Global AI glasses shipments surged to 8.7 million units in 2025, up more than 300% from the prior year, and analysts project that figure will cross 15 million this year,per Omdia.Suppliers and component makers of AI-powered smart glasses are also positioning themselves for what comes next. One of the companies, a South Korean startup calledLetinAR, has spent the last decade building the optical technology that could make all of this actually wearable.The LG Electronics-backed startup just secured $18.5 million from Korea Development Bank and the South Korean retail giant’s venture arm, Lotte Ventures, among others, ahead of its planned 2027 IPO in South Korea.Its previous investor,LG Electronics, has since begun developing its own AI smart glasses, according to alocal mediareport, which is a sign of how seriously South Korea’s largest consumer electronics company takes the category.LetinAR CEO Jaehyeok Kim and CTO Jeonghun Ha, who have been friends since high school, founded the company together in 2016.Image Credits:LetinARThe lens that makes it wearableLetinAR doesn’t make the glasses. It makes the part that makes the glasses work. The optical module, the tiny lens component that projects images into your field of vision, is what determines whether a pair of smart glasses feels like a sci-fi headset or something you’d actually wear to work, Ha told TechCrunch. It has to be light, thin, and power-efficient, while still delivering a sharp, clear image. Getting all of that right in a single component, small enough to fit inside a normal-looking frame, is the central engineering challenge of the entire industry. That’s what LetinAR is building.“We see AI glasses as that next platform,” Kim said. “And the optical module is the hardest part to get right as AI glasses makers will need a lens that is thinner, lighter, and more power-efficient than what exists today.”The co-founders said LetinAR wants to be the company those glasses makers call. The company calls its technology PinTILT: a way of arranging tiny optical elements inside a lens so that light is directed precisely where it needs to go, into the user’s eye, rather than scattered in every direction.Think of a TV. It broadcasts light across an entire room, but only the light that actually reaches your eyes matters. Most existing smart lens technologies, particularly a dominant approach calledwaveguide, work a bit like that TV, splitting and spreading light across the full lens to create a wide image. The result is a thin lens, but an inefficient one. A lot of light gets thrown away before it ever reaches the eye, which means dimmer images and, critically, a battery that drains fast, Ha explained.The alternative, a mirror-based approach known asbirdbath, delivers light more directly to the eye, but the structure is bulky, making it nearly impossible to fit inside something that looks like a normal pair of glasses.PinTILT sidesteps that trade-off, Ha said. By focusing only on the light that can actually enter the eye and carefully engineering the angle of each tiny element inside the lens, LetinAR claims it can produce a brighter image in a thinner, lighter form factor, using less power. In a category where every gram and every hour of battery life matters, that’s the problem the entire industry has been trying to solve.In the space, there are a number of peers likeWaveOptics,DigiLens, andLumus.CustomersIts modules are already shipping. LetinAR counts Japan’s NTT QONOQ Devices and Dynabook, formerly known as Toshiba Client Solutions, among its customers, giving the company real manufacturing experience at scale. It is in talks with Big Tech companies on R&D of next-generation AI glasses, though it declined to name them.One of LetinAR’s most demanding customers isAegis Rider, a Swiss deep tech company spun out of ETH Zurich’s Computer Vision Lab. Aegis Rider is building an AI-powered AR helmet that displays navigation, speed, and safety alerts directly in a motorcycle rider’s field of vision, not floating on the visor but anchored to the road itself, as if the information is physically painted on the world ahead.LetinAR’s mod

Amazonannouncedthe latest update to Alexa+ on Monday: the ability to generate podcast episodes on demand.The new feature, called “Alexa Podcasts,” is rolling out to customers in the U.S. today. Amazon describes the capability as a way to “turn any topic you’re curious about into a podcast episode, ready in minutes.”To use the feature, all users have to do is ask Alexa+ to create a podcast about a topic they’re interested in. Users don’t need to upload documents, write scripts, or plan anything ahead of time. Instead, Alexa+ researches the request, gathers information, and generates a quick overview of what the episode will cover. From there, users can tweak things like the length, tone, and focus of the episode.Once finalized, Alexa+ uses AI-generated host voices to narrate the podcast. When the episode is ready, users get a notification through their Echo Show device and inside the Alexa app. Episodes are also saved in the app’s “Music” and “More” sections so they can be replayed later.The feature is another example of how Amazon is trying to turn Alexa+ into more than just a voice assistant. Instead of only answering questions or controlling smart home devices, Alexa+ is starting to act more like a personalized AI content creator.At the same time, the launch is likely to spark some debate. AI-generated voices and automated content continue to raise questions around ethics, accuracy, and the future of traditional creators. There are also concerns about how reliable AI-generated podcasts will be, especially when covering news or complex topics.Amazon emphasized its partnerships with major news organizations to improve content accuracy and reliability. The company says Alexa+ can access real-time information through agreements with outlets, including the Associated Press, Reuters, The Washington Post, Time, Forbes, Business Insider, Politico, USA Today, Condé Nast, Hearst, and Vox Media, alongside more than 200 local newspapers across the U.S.Beyond podcasts, Amazon says it is exploring additional forms of personalized AI audio, including custom news briefings and content generated from users’ own documents and shared information.TopicsAI,alexa plus,alexa+,Amazon,Generative AI,Media & Entertainment,podcastsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Grafana Labs, the maker of its eponymous popular open source web visualization software, confirmed it had been hacked but that it refused to pay the hackers who had threatened to release the company’s codebase.Ina series of postson social media, the lab said its investigation found that the hackers had abused a stolen token credential that allowed access to the company’s GitHub environment, which it uses for storing its source code, but the token did not allow access to customer records or financial data. The company has since invalidated the token and added additional security measures to prevent a repeat incident.“The attacker attempted to blackmail us, demanding payment to prevent the release of our codebase,” the company said.Grafana’s code is open source and public, meaning anyone can download the software and edit its code before running it on their own machines. It’s unclear if the hackers stole any proprietary code or information. A spokesperson for the company did not immediately return a request for comment.The incident contrasts with the recent hack at education tech giant Instructure, which last week“reached an agreement” to pay the hackerswho had compromised its network twice in recent weeks. The hackers had demanded an unspecified ransom, threatening to release stolen data about staff and students who use its softwarefollowing a massive data breach and a subsequent website defacement.While in Grafana’s case, no customer data was taken, the company cited the FBI’s long-standing advice urging victims not to pay hackers, as cooperating with them does not guarantee they will return stolen data or refrain from publishing it later. Critics also say paying cybercriminals helps to fund future cyberattacks.Grafana said its investigation was ongoing and will share its findings once its probe concludes.This story was updated to correct that the hackers compromised access to Grafana’s GitHub environment.Topicscyberattack,cybersecurity,data breach,extortion,grafana,open source,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

The market for AI notetaking devices has exploded in the U.S., with the category generating over $600 million in revenue last year, according toa Menlo Ventures report. And as startups likeHeidi HealthandFreedhave shown, there’s decent demand for this tech in healthcare, where doctors and clinics see the potential for an AI assistant that can help them keep track of patient conversations, surface health records, and lower their administrative burdens.But those apps don’t do much for patients, which is whyKin Healthis building a notetaker that can transcribe your visits to doctors, parse medical advice, and surface next steps when required. To that end, the startup has raised $9 million in a seed funding round led by Maveron.The app is similar to a meeting notetaker: You can record doctor visits, and it will return an AI summary of the meeting, with the next steps, all of which you share with family and friends if you want to. It also lets you write down questions that you might want to ask during your next visit.Loading the player…Kin Health says it encrypts all patient data and that summaries are kept private by default. The tool is not HIPAA-certified, as it is a patient-facing one, but it adheres to the same privacy standards, the company said.The free app is built by physicians Arpan and Amit Parikh, along with Kyle Alwyn, who previously built online prescription serviceHeyDoctor and sold it to health platform GoodRx. Doug Hirsch and Trevor Bezdek, co-founders of GoodRx, are founding partners and executive chairmen at the company.Co-Founders Arpan Parikh, Amit Parikh, and Kyle Alwyn.Image Credits:Kin Health“We have a lot of these storage cabinets where our health data can live, but we don’t have a way to convert that into a utility that we can use to drive our behavioral change. Our goal is to create this health graph where we can store your information from multiple different sources,” Alwyn told TechCrunch over a call.Kin Health says that its summaries are provided after a few stages of processing. After transcribing the visit, an algorithm turns the transcription into a clinical narrative, which gets crunched into a user-facing summary with action items. The company says it is leaning on specialized medical models to power the transcription and that it evaluates and observes outputs at different stages to ensure answers are accurate.But AI in healthcare is being received with a measure of caution and apprehension. Privacy experts and researchers have raisedconcernsover data security, accuracy of AI, consent mechanisms, thequality of generated notes, and theireffectiveness.AI notetakers also oftenfail to recognize and struggle to transcribe regional accents. Kin Health says it is working to ensure its tool works with different accents, as well as when someone has a bad throat or is wearing a mask.Dr. Rebecca Mishuris, chief health information officer and VP at Mass General Brigham, a healthcare organization in Boston, argues it is important for physicians to review any notes generated by AI.“Generative AI will hallucinate; that is the nature of a technology built on patterns and prediction. That is why it is so important for clinicians to review the drafted notes before signing them. At the end of the day, the responsibility for the documentation falls to the clinician,” she told TechCrunch over email.Kin Health currently only shows notes from conversations it records during consultations, but the company said it plans to bring in data from other health sources, including physicians’ own notes through electronic health record (EHR) systems, this year.The company says it will keep the app free of cost forever, and monetize via referrals to services such as specialists and labs. The startup is taking a leaf from GoodRx’s playbook, which also keeps the core product free and earns commissions by referring other services.Natalie Dillon, a partner at Maveron, said healthcare provider-side tools often expect patients to coordinate their own treatment actions. “Kin is built to solve an entirely different consumer need: it can travel with them between specialists, systems, and providers. It’s not beholden to any single health network or EHR relationship. It’s built to serve the patient, not the institution, and that’s a massive distribution advantage,” she said.The funding round also saw participation from Town Hall Ventures, Eniac Ventures, Flex Capital, Foundry Square Capital, Pear VC, and The Family Fund. GoodRx’s Hirsch and Bezdek; angel investors Jay Desai, Nabeel Quryshi, Alex Cohen, and Saharsh Patel; and more than 30 physicians also invested.Topicsai scribe,Apps,Health tech,Kin Health,maveronWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Ivan MehtaIvan covers global consumer tech developments at TechCrunch. He is based out of India and has previously worked at publications including Huffington Post and The Next Web.You can

New York public health provider NYC Health + Hospitals saysa months-long data breachthat allowed hackers to steal personal data, medical records, and fingerprints scans affects at least 1.8 million people.NYCHHC is the largest public health system in the United States and provides healthcare toover a million New Yorkers, the majority of whom are uninsured or receive state healthcare benefits, such as Medicaid.The healthcare system reported the number to the U.S. Department of Health and Human Services, making it one of the largest healthcare-related data breaches of the year so far. Healthcare organizations have been repeatedly targeted by financially motivated cybercriminals in recent years in efforts to steal their vast banks of highly sensitive patients’ personal, medical, and billing information.In a data breach notice on its website, NYCHHC said that it detected a cyberattack on February 2 and secured its network. The hackers had access to its network from November 2025 until February 2026, during which the hackers copied files from its systems.The healthcare system said hackers broke due to a breach at a third-party vendor, which it did not name.NYCHHC said that the exposed data varies by individual and includes patients’ health insurance plan and policy information, medical information (e.g., diagnoses, medications, tests, and imagery), billing, claims, and payment information. Other government-issued identity documents, such as Social Security numbers, passports, and driver’s licenses, were also compromised.The breach notice also says “precise geolocation data” was taken in the breach, suggesting that the user-uploaded photos of their identity documents may have also contained the exact location of where the document was captured.The breach is particularly sensitive because hackers stole biometric information, including fingerprints and palm prints, which affected individuals have for life and cannot replace. NYCHHC did not provide an explanation for storing biometric data. Prospective NYCHHC employees are generally required to enroll their fingerprints for criminal records checks. It’s not yet known if patients’ biometrics were also taken.NYCHHC’s website was briefly offline as of Monday morning. A spokesperson for NYCHHC did not immediately respond to an email from TechCrunch with questions about the cyberattack. TechCrunch asked, among other things, why it took the organization months to detect the breach, and if it has received any communication from the hackers, such as a demand for payment.It’s not clear if NYCHHC can receive email at the time of the website outage.The incident appears to be unrelated to the data breach at National Association on Drug Abuse Problems (NADAP)earlier this year, in which over 5,000 NYCHHC patients had information taken in the cyberattack.In the FBI’s latestannual report on cybercrimecovering 2025, healthcare remained a top target for ransomware attackers — criminals who break into databases, steal a copy of the data while scrambling the victim’s servers, and threaten to publish the stolen data if the victim does not pay the hackers. A ransomware attack on UnitedHealth-owned health tech giant Change Healthcare allowed Russian-linked hackers tosteal the medical and billing information of more than 190 million Americans, believed to be the largest theft of U.S. medical data in history.Topicscyberattack,cybersecurity,data breach,healthcare,new york city,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Elon Musk’s claim that he was mistreated by his OpenAI co-founders failed after nine California jurors returned a unanimous verdict that his lawsuits had been filed too late.Musk accused Sam Altman, Greg Brockman, OpenAI, and Microsoft of “stealing a charity” by creating a for-profit affiliate of the frontier AI lab. Jurors, however, found that any harms that Musk may have suffered came before the deadline for filing his claims under the law.While the trial delved deeply into the melodramatic history of OpenAI and featured testimony from leading figures in Silicon Valley, it ultimately turned onfairly narrow questionsof the law. The trial focused on whether and when Altman and the other defendants had made and broken promises to Musk, but his case failed to convince jurors that he had a valid claim.In particular, OpenAI had advanceda statute of limitations defense, which sought to prove that any harms Musk sought to litigate had taken place before 2021. (The specific date varied by the charge: before August 5, 2021, for the first count; August 5, 2022, for the second count; and November 14, 2021, for the third count.) Ultimately, the jury found that argument persuasive, which made for a short deliberation period.“There was a substantial amount of evidence to support the jury’s finding, which is why I was prepared to dismiss on the spot,” Judge Yvonne Gonzalez Rogers said after the verdict was delivered.The end of the case means that one major threat to OpenAI — a possible restructuring — is now off the table ahead of its reported IPO.“It did not take [the jury] two hours to conclude … that Mr. Musk’s lawsuit is nothing more than an after-the-fact contrivance that bears no relationship to reality,” OpenAI’s lead attorney, Bill Savitt, said after the verdict. “They kicked it exactly where it belongs — just to the side. This lawsuit is a hypocritical attempt to sabotage a competitor.”Microsoft, which Musk sued for aiding and abetting OpenAI’s alleged breach of charitable trust, welcomed the verdict. A spokesperson for the company said it “remained committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”The verdict came in the middle of a hearing to determine the potential damages to Musk if the verdict had gone the other way. While that discussion is moot for now, the judge appeared unconvinced by the analogy Musk’s lawyers drew between his charitable contributions and investments in a for-profit startup.“Your analysis seems to be devoid of connection to the underlying facts,” she told Dr. C. Paul Wazzan, the expert who came up with Musk’s estimate of OpenAI and Microsoft’s wrongful gains at his expense — some $78.8 billion to $135 billion.In a tweet after the ruling, Musk appeared to take the procedural grounds of the dismissal as a moral victory. “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!” Musk wrote. “I will be filing an appeal with the Ninth Circuit, because creating a precedent to loot charities is incredibly destructive to charitable giving in America.”Reached for comment by TechCrunch, Musk’s lead counsel, Marc Toberoff, said, “One word: Appeal.”TopicsAI,Elon Musk,greg brockman,Microsoft,OpenAI trial,sam altmanWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Tim FernholzSenior ReporterTim Fernholz is a journalist who writes about technology, finance and public policy. He has closely covered the rise of the private space industry and is the author ofRocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race.Formerly, he was a senior reporter at Quartz, the global business news site, for more than a decade, and began his career as a political reporter in Washington, D.C.

Anthropic announced Monday it has acquired Stainless, a startup founded by former Stripe engineer Alex Rattray whose software is widely used by rival AI labs, including OpenAI and Google.Anthropic didn’t disclose terms of the deal. However, The Informationreportedlast week that the company was in talks to acquire Stainless, which is backed by Sequoia Capital and Andreessen Horowitz, for more than $300 million.The acquisition will take a key infrastructure supplier out of the hands of Anthropic’s competitors. The company told TechCrunch it will wind downall hosted Stainless products, including its SDK generator. An Anthropic spokesperson said Stainless customers will still own the SDKs they’ve generated to date and have full rights to modify and extend them however they wish.The New York-based startup, founded in 2022, rose to prominence in the emerging AI industry for automating the creation and maintenance of software development kits, or SDKs — the libraries developers use to interact with APIs.Rattraydeveloped softwarethat could take API specifications and turn them into production-ready SDKs across multiple programming languages, including Python, TypeScript, Kotlin, Go, and Java. It became a popular tool because the platform automatically updates the SDKs as APIs change and eliminated the time-consuming process of manually maintaining them.The technology is particularly valuable tocompanieslike Anthropic, OpenAI, Google, Replicate, Runway, and Cloudflare that are building AI agents that can connect to external software and complete tasks on behalf of users. Stainless’s SDK tools are an easy way to build and maintain those connections — but going forward, the tools will only be available to Anthropic, not its competitors.According to Anthropic, Stainless software has powered the generation of every official Anthropic SDK since the earliest days of its API.“I started Stainless because SDKs deserve as much care as the APIs they wrap,” Rattray saidin a press releaseposted Monday. “Anthropic was one of the first teams to bet on this with us. We have been watching what developers have built on Claude over the last few years, which made bringing our teams together an easy decision. The team gets to keep doing the work we love, on the platform where it matters most.”TopicsAI,Anthropic,StainlessWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Kirsten KorosecTransportation EditorKirsten Korosec is a reporter and editor who has covered the future of transportation from EVs and autonomous vehicles to urban air mobility and in-car tech for more than a decade. She is currently the transportation editor at TechCrunch and co-host of TechCrunch’s Equity podcast. She is also co-founder and co-host of the podcast, “The Autonocast.” She previously wrote for Fortune, The Verge, Bloomberg, MIT Technology Review and CBS Interactive.You can contact or verify outreach from Kirsten by [email protected] via encrypted message at kkorosec.07 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Drug discovery is one of the most expensive pursuits in modern industry. Finding a single viable molecule can take a decade and cost billions, and most candidates still don’t make it. A generation of AI startups has promised to fix that — most have made the problem less painful for researchers, who are already technically sophisticated enough to use the tools.But SandboxAQ thinks the bottleneck isn’t the models. It’s the interface.The company has teamed up with Anthropic to integrate its scientific AI models directly into Claude — putting powerful drug discovery and materials science tools behind a conversational interface that requires no specialized computing infrastructure to use.Founded roughly five years ago as an Alphabet spinout, SandboxAQ counts Eric Schmidt, Google’s former CEO, as its chairman. The company, which has raisedmore than $950 million from investors, has built out a number of different business lines, includinga cybersecurity business.One of the more unique things SandboxAQ does, however, is produce large quantitative models, or LQMs. These proprietary models are “physics-grounded,” meaning they’re built on the rules of the physical world rather than patterns in text. They can run quantum chemistry calculations and simulate both molecular dynamics and microkinetics, the study of how chemical reactions unfold at the molecular level. That matters because it tells researchers how candidate molecules are likely to behave before anyone sets foot in a lab.“Trained on real-world lab data and scientific equations, LQMs are AI models engineered for the quantitative economy, a $50+ trillion sector spanning biopharma, financial services, energy, and advanced materials,” the company said in a news release that strongly suggests Sandbox AQ isn’t building another chatbot or code assistant — it’s chasing the economy that AI is supposed to transform.Chai DiscoveryandIsomorphic Labs— both well-funded bets on better models — have focused on the science. SandboxAQ is focused on who can actually use it.“For the first time, we have a frontier [quantitative] model on a frontier LLM that someone can access in natural language,” Nadia Harhen, SandboxAQ’s general manager of AI simulation, told TechCrunch. Previously, users of SandboxAQ’s LQMs would have had to provide their own digital infrastructure to run the models.SandboxAQ’s customers tend to be computational scientists, research scientists, or experimentalists. Generally, these people work at large pharmaceutical or industrial companies and are searching for new materials that can become marketable products.“Our customers come to us because they’ve tried all the other software out there, and the complexity of their problem is such that it didn’t work or didn’t yield positive results for them when that translation went to take place in the real world,” said Harhen.TopicsAI,Anthropic,Claude,Eric Schmidt,Google,SandboxAQWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lucas RopekSenior Writer, TechCrunchLucas is a senior writer at TechCrunch, where he covers artificial intelligence, consumer tech, and startups. He previously covered AI and cybersecurity at Gizmodo.

A worker died at SpaceX’s Starbase launch site in South Texas on Friday, and the Occupational Safety and Health Administration (OSHA) has opened an investigation.TheSan Antonio Express-News reported Mondaythat the unidentified victim died at around 4:17 a.m. local time on May 15, citing OSHA and local officials. The Wall Street Journal later reported that the county sheriffconfirmed to the outletthat a worker died. OSHA confirmed to TechCrunch that it is investigating the apparent accident.Representatives for the nearby Brownsville police and fire departments did not respond to requests for comment. SpaceX and the newly incorporated city of Starbase did not respond to requests for comment.The circumstances of the worker’s death are not immediately clear. OSHA told TechCrunch that it won’t release more information until its investigation is complete, which could take months.The death comes just a few days ahead of the first planned launch of SpaceX’s upgraded Starship rocket. Elon Musk’s spaceflight company is also reportedly releasing the detailed prospectus for its initial public offering this week, which is expected to be the biggest ever when that transaction takes place next month.SpaceX has long dealt with worker safety problems at its Starbase site, which handles Starship prototype launches and is an active construction zone.In 2025, TechCrunchanalyzed OSHA dataand determined the Texas launch site had an injury rate that far outpaced those of industry rivals and was the most dangerous of SpaceX’s worksites. A 2023Reuters investigationuncovered dozens of previously unreported injuries and a worker death in 2014 at SpaceX’s McGregor, Texas, test site.In January, OSHA hit SpaceX with seven “serious” safety violations for, among other things, not properly inspecting a crane before it collapsed at Starbase last June. The safety agency dealt SpaceX the maximum financial penalty on six of those seven violations, totaling $115,850. SpaceX is contesting those penalties, federal records show.The company has been hit with multiple lawsuits related to injuries sustained at Starbase in recent years. In December, an employee of one of SpaceX’s subcontractors sued after he was crushed by a large metal support dropped from a crane. The worker, Eduardo Cavazos, suffered a broken hip, knee, and tibia, and OSHA opened a “rapid response investigation,” as TechCrunchfirst reportedin December.OSHA has since closed that rapid response investigation without taking any punitive action, according to a TechCrunch public records request. And the lawsuit was recently dropped because his employer, the subcontractor, provides workers compensation insurance that prevents the company from being sued, according to Cavazos’ attorney.TopicsOSHA,Space,SpaceX,starbase,TransportationWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sean O’KaneSr. Reporter, TransportationSean O’Kane is a reporter who has spent a decade covering the rapidly-evolving business and technology of the transportation industry, including Tesla and the many startups chasing Elon Musk. Most recently, he was a reporter at Bloomberg News where he helped break stories about some of the most notorious EV SPAC flops. He previously worked at The Verge, where he also covered consumer technology, hosted many short- and long-form videos, performed product and editorial photography, and once nearly passed out in a Red Bull Air Race plane.You can contact or verify outreach from Sean by [email protected] via encrypted message at okane.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiWhat happens when AI starts building itself?Russell BrandomOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikMusk’s xAI is running nearly 50 gas turbines unchecked at its Mississippi data centerTim De Chant

Bad beatsThe US is betting on AI to catch insider trading in prediction marketsThe Commodity Futures Trading Commission wants us to know it’s taking this very seriously.Kate Knibbs, wired.com–May 16, 2026 7:00 am|112Credit:Yuki Shimazu / FlickrCredit:Yuki Shimazu / FlickrText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navFor most of the past year, it looked likeprediction marketshad kicked off a new golden age of fraud. OnPolymarket, traders raked in fortunes fromsuspiciously timed betson geopolitical events like the raid on Venezuela and the Iran War. It wasn’t clear whether the US government would bother pursuing some of the most flagrant bad actors, since Polymarket’s crypto-based platform was technically offshore and not regulated or licensed within the country.Now, however, the Commodity Futures Trading Commission, which oversees prediction markets, wants you to know that it’s watching very, very closely. The agency is searching for suspicious behavior from traders within the United States who have been sneaking onto offshore markets, including Polymarket’s crypto platform—which is blocked stateside—by using virtual private networks. “We’re going to find them, and we’re going to bring actions,” agency chairman Michael Selig told WIRED this week, speaking from the CFTC’s headquarters in Washington, DC.Selig says the agency, which is especially lean right now, is staffing up. Like so many other AI-pilled workplaces, the CFTC is also leaning into automation to handle the growing workload, including tools that analyze trading patterns and flag potential manipulation. “You’ve got so much data,” Selig says. “When we feed it into AI, we get really great information. It can help us understand things, like where we might want to investigate, or when we might need to send a subpoena to a trader.”In addition to proprietary surveillance systems developed in-house, the agency’s arsenal includes third-partyblockchain tracing toolslike Chainalysis for crypto platforms, and market abuse detection software including Nasdaq Smarts for centralized markets. (Beyond Nasdaq Smarts, the agency did not specify which AI tools it uses and declined to share more specific examples.)Prominent prediction market companies have recently started touting all the work they’re doing to catch sketchy bettors. US-based exchange Kalshi, Polymarket’s primary competitor, eagerly announced that it hassuspendedand penalized customersflagged for insider tradingand market manipulation.In April, after significant backlash over suspected insider trading, Polymarket announced its own partnership with Chainalysis. It was part of a broader push to crack down on market manipulation. While the company’s CEO, Shayne Coplan,had talkedin the past about why insider trading could be good for prediction markets, Polymarket changed its approach this spring, updating its market integrity rules and announcing a partnership with Palantir for its US-based sports markets (the Chainalysis deal focuses on the offshore platform). The company did not respond to WIRED’s request for comments for this story.According to Chainalysis spokesperson Maddie Kenney, the company analyzes the same data for both clients. “The value Chainalysis adds for our customers, including Polymarket and the CFTC, is organizing the data and enriching it with the attributions and insights we’ve accumulated over years in the space,” she says. Certainly sounds like a good deal for Chainalysis!The CFTC’s assurances that it is hunting insiders comes at a moment of intense scrutiny on prediction markets. In March, Connecticut senator Chris MurphytoldWIRED that he suspected White House staffers were engaged in insider trading on war-related contracts. At the beginning of April, seven members of Congressaskedthe CFTC to investigate overseas markets offering war-themed events contracts. In a letter, the lawmakers argued that the commission had the authority and responsibility to curb insider trading, especially on “morally obscene” trades on military action. Selig recentlytoldCongress that the company is pursuing “hundreds, if not thousands” of insider trading tips.Investigations are not limited to federally regulated exchanges. “We’re surveilling the markets on a global basis,” he tells WIRED.Selig says that the agency will exert extraterritorial jurisdiction—its legal ability to enforce its laws beyond traditional boundaries—when it finds suspicious activity on offshore platforms like Polymarket, though he says it’s a case-by-case approach. “We use it in extreme circumstances,” he says, with an eye towards whether charges have a strong chance of sticking in court. “In any extraterritorial litigation, there’s going to be challenges to our authority, and that could also impair our ability to bring cases in the future.” According to Selig, the 2010 Dodd-Frank Act allows the CFTC more leeway to pursue this kind of enforcemen

striking resultsA revolutionary cancer treatment could transform autoimmune diseaseResearchers are testing CAR T cell therapy as a way to reset the immune system.Amber Dance, Knowable Magazine–May 17, 2026 7:00 am|66Artist’s rendering of a T cell.Credit:Getty ImagesArtist’s rendering of a T cell.Credit:Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAt age 49, Jan Janisch-Hanzlik’s multiple sclerosis was destroying her freedom to live the life she wanted. She gave up her active nursing job for a desk role. Frequent falls made her afraid to carry her grandchildren. She had to move to a bigger house to make room for the wheelchair she feared she might end up needing full-time.Even the best available medication wasn’t improving Janisch-Hanzlik’s symptoms, and she worried they’d only get worse. So when she learned about a trial of CAR T cell therapy at the University of Nebraska Medical Center in Omaha, close to the city of Blair where she lives, she phoned the clinic every other month until they were ready toenroll her as the first patient.Originally designed to target and wipe out cancer byreprogramming the patient’s immune cells, CAR T is now being offered to patientsin hundreds of clinical trialsfor autoimmune conditions like multiple sclerosis, lupus, Graves’ disease, vasculitis, and many others. The hope is that CAR T can duplicate the success it has demonstrated in a range of blood cancers by hunting down and eliminating cells that target the self in autoimmune diseases. This would essentially reset the body’s defenses to a state like the one that existed before the disease took hold.But along with CAR T’s promise come risks, questions, and challenges. There’s uncertainty about how well it will work forautoimmunityand how long any benefits might last, as well as what long-term side effects might arise. Janisch-Hanzlik knew this when she sat down to receive the experimental treatment on June 9, 2025; she felt a mix of hope and fear knowing that she would be spending the next week being monitored for side effects including dangerous inflammation.In addition to her clinical expertise and desire to pioneer a new treatment, Janisch-Hanzlik’s two young grandchildren helped inspire her pursuit of a treatment with known risks and uncertain benefits. Because multiple sclerosis has a genetic component, Janisch-Hanzlik knew that they have an elevated chance of going through the same struggle she has. “I would want to be able to say I did everything that I possibly could to prevent them, or anyone else, from having something like this,” she says.From cancer to autoimmunityThe first CAR T cancer treatment wasapproved by the Food and Drug Administration in 2017for an aggressive form of leukemia. Since then, the powerful and intensive treatment hasresulted in long-term remissionfor many cancer patients.The basic premise of CAR T is to activate the power of key immune cells calledT cells. T cells normallyrecognize other cellsthat have been infected by a virus or bacterium, or are otherwise abnormal, and either destroy them or recruit other parts of the immune system to do so.In CAR T for cancer, scientists engineer those T cells to specifically hunt and destroy malignant cells. The technologygot its startwhen cancer researchers figured out how to take out a patient’s own T cells, insert DNA instructions for a “chimeric antigen receptor,” or CAR, and put them back into the person’s circulation. The CAR, which sits on the T cell’s surface and latches on to a specific molecular partner on the surface of cancerous cells, activates the T cell to attack.Today CAR T cells are most commonly programmed to attack B cells, another key immune player. B cells are normally responsiblefor making antibodies, but in certain blood cancers, they proliferate out of control. By giving T cells a CAR that recognizes one of a couple of molecules unique to the B cell surface, the cells are reprogrammed to find and eliminate those cancerous cells.B cells also are the central problem in many autoimmune conditions: They mistakenly make antibodies against normal tissues instead of against invading pathogens. So as CAR T began to succeed against B cell cancers, it didn’t take long for doctors to reason that CAR T therapy might also be able to wipe out bad B cells in people with autoimmunity.There are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineThere are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineA German team pioneered autoimmune CAR T in a woman with lupus,reporting positive resultsin 2021.

All roads lead to StarshipThe US space enterprise is desperately waiting for Starship—will it finally deliver?“This is such a wild ride. The highs are high. The lows are low.”Eric Berger–May 18, 2026 7:00 am|156Starship V3 completes a successful wet dress rehearsal on May 11.Credit:

      SpaceXStarship V3 completes a successful wet dress rehearsal on May 11.Credit:

      SpaceXText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThese days, one would be forgiven for forgetting that SpaceX is, at its core, a rocket company.Consider the company’s mega deals over the last year. SpaceX paid $17 billion—more than it has spent developing every one of its rockets—to EchoStar for wireless spectrum to boost its Starlink network. It revealed plans to launch 1 million orbital data centers. SpaceX merged with xAI in a deal that valued Elon Musk’s artificial intelligence firm at $250 billion, and it announced plans to becomea major computer chip manufacturer. And earlier this month, SpaceX sold an enormous amount of ground-based compute to Anthropic.As a result of all this activity, an impending IPO will value the company at something like $1.5 or $2 trillion. That’s trillion, with at.So yes, one might reasonably ask what SpaceX does these days. Because all the buzz, all the Wall Street euphoria, and all the financial frisson are only tangentially related to what SpaceX cut its teeth on during its first 25 years: becoming the globally dominant player in launch. It largely concerns telecommunications and AI data services.And yet everything SpaceX aspires to accomplish in the next quarter of a century, all of its enormous valuation, is predicated on a new launch vehicle. A rocket that, to date, has a decidedly mixed record of success. A rocket that has not flown in seven months. A rocket that, finally, may return to the skies on Wednesday.We are speaking, of course, of Starship—a truly revolutionary rocket. If it works. And after a long period of development and three years of test flights and setbacks, it kind of has to.“Test Like You Fly”A few weeks ago, SpaceX released avisually stunning videothat takes viewers inside its massive new Starfactory in South Texas and provides up-close views of its rockets and engines.The “Test Like You Fly” video also outlines the development of the third iteration of the Starship rocket, V3, highlighting both advances and setbacks encountered along the way. SpaceX is distilled to its core as the exceptional engineers who are designing and testing a radically new rocket talk about the very difficult problems they’re trying to solve.“We are not breaking laws of physics; we are just trying to leverage them as effectively as we can,” explains Jacob McKenzie, the company’s vice president for the Raptor rocket engine that powers both the first and second stage of Starship.McKenzie says SpaceX built 600 Raptor engines as part of its V2 Starship program, a remarkable number that underscores the investment the company is making to develop a rocket that has yet to reach orbit or deliver a single payload. (To put this into perspective,NASA is spending $3.5 billionto procure two dozen comparably powered, expendable rocket engines.)By some estimates, SpaceX has now invested $15 billion in the Starship rocket program over the last decade, funding 11 test flights, a sprawling spaceport in South Texas, a massive new factory there to support a high production rate, and expanding facilities in Florida alongside the company’s existing infrastructure in the state.It has not gone particularly smoothly. The company conducted its initial Starship test flight in April 2023. I’m not sure anyone expected that, three years and nearly a dozen flights later, the company would still be firmly in test mode.But it turns out that building the world’s largest and most powerful rocket and optimizing it for rapid reuse is a difficult, time-consuming task. Even for a company that prides itself on moving fast and trying to achieve the near impossible.A very difficult 2025SpaceX performed six test flights in 2023 and 2024, and by November of the latter year had made substantial progress. The company demonstrated the ability to launch and capture the Super Heavy first stage and completed a safe flight of the upper Starship stage, including re-lighting its Raptor engine in space, before making a controlled splashdown in a precise location.That month, the then-general manager of the Starbase facility in South Texas, Kathy Lueders, said the company aimed to dramatically increase the vehicle’s launch cadence in 2025. “Elon would say, next year, he would love to have us have 25 missions a year,” she saidduring a community event. The company also planned to “capture” a Starship upper stage and conduct an orbital refueling test in that time frame.Debris from Starship falls back into the atmosphere in this view over Hog Cay

swarmsBug bounty businesses bombarded with AI slop“Never-ending” AI slop strains corporate hacking reward schemes.Jamie John, Financial Times–May 18, 2026 9:23 am|79Credit:

      AdobeCredit:

      AdobeText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navCompanies that pay hackers to find flaws in their software are being inundated with low-quality reports generated by AI, forcing some to suspend the programs altogether.Businesses that run “bug bounty” schemes have long relied on independent security researchers to spot vulnerabilities. But the rise of AI tools is now overwhelming them with spurious submissions.Bugcrowd, whose customers include OpenAI, T-Mobile, and Motorola, said the number of reports it received more than quadrupled over a three-week period in March, with most proving to be false.Curl, a widely used tool to transfer data across the Internet, suspended its paid bug bounty program in January, citing an “explosion in AI slop reports” and lower-quality submissions.Cyber security experts say advances in generative AI are reshaping the economics of bug bounty programs. While the tools allow experienced researchers to find flaws more quickly, they are also lowering the barrier to entry, triggering a flood of automated or erroneous submissions that companies must sift through.The big increase in poor-quality AI reports was “quickly becoming a major problem,” said Ross McKerchar, chief information security officer at cyber security group Sophos. “Bug bounties are going to stay [but] they’re going to have to change,” he said.Bug bounties have grown in popularity since the early 2000s, with schemes offering six-figure payouts for the biggest discoveries. Google’s program disbursed a total of $17 million last year, up from $7.5 million in 2021. It paid its largest individual reward of $605,000 in 2022 to a user who spotted a vulnerability in its Android mobile operating system.McKerchar said the rise in poor-quality submissions came from both amateurs trying to find bugs for the first time and existing researchers who were “sometimes getting led on by the [AI] agents.”He added there was a “third cohort” of “experienced AI builders” who had developed automated “end-to-end scanning and submission systems” that were “creating absolute carnage.”Curl’s creator, Daniel Stenberg, wrote in a blog post that the “never-ending slop” had taken “a serious mental toll to manage and sometimes also a long time to debunk.”Software group Nextcloud suspended its bug bounty program in April because of the “massive increase of low-quality reports.” It said it hoped to resume the program once it had found a way to filter submissions effectively.The surge in AI-generated reports comes as Anthropic last month launched Mythos, its new cyber AI model, which it says can find software flaws faster than humans.Companies running bug bounty programs have started to introduce more stringent background checks to combat the problem, as well as building AI agents to triage submissions.HackerOne, whose bug-reporting platform serves Goldman Sachs, Google, and the US Department of Defense, said it had “introduced new agentic validation capabilities” this year to “help organizations manage high volumes of findings,” such as those generated by models like Mythos.The company said submissions had jumped 76 percent in the year to March. But it said the share of reports flagging legitimate vulnerabilities had remained steady over the past year at 25 percent.HackerOne chief executive Kara Sprague said it had in recent weeks seen a rise in “higher quality” reports that had used AI. She added that the rise in AI-generated submissions was “not a strong reason to say we don’t want them” altogether, given that hackers were using the technology to spot more flaws.Bugcrowd chief Dave Gerry said developments such as Anthropic’s Mythos would assist human bug bounty hunters, not replace them. “AI is going to help with a lot of things but we’re never going to replace that human creativity,” he said.© 2026 The Financial Times Ltd.All rights reserved. Not to be redistributed, copied, or modified in any way.Financial TimesFinancial Times79 Comments

The Next GenerationDid Artemis II break through? Registrations at Space Camp double afterward.Isaacman’s $25 million donation leads to impressive new facilities.Eric Berger–May 18, 2026 9:49 am|26Commercial Astronaut on Inspiration4 Chris Sembroski and NASA Administrator Jared Isaacman attend the ribbon-cutting ceremony of the Inspiration4 Skills Training Complex in Huntsville, Ala.Credit:

       Tibrina Hobson/Getty ImagesCommercial Astronaut on Inspiration4 Chris Sembroski and NASA Administrator Jared Isaacman attend the ribbon-cutting ceremony of the Inspiration4 Skills Training Complex in Huntsville, Ala.Credit:

       Tibrina Hobson/Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen he was 12 years old, NASA Administrator Jared Isaacman attended the weeklong “Aviation Challenge” program at Space Camp in Huntsville, Alabama.“For the first time, I got behind the controls of an airplane when I attended Aviation Challenge,” Isaacman said on Friday evening during an event at the US Space & Rocket Center. “I became a pilot because I thought that was the closest I would ever get to the stars.”Decades later, after founding a successful online payments company and flying to space twice as a private citizen aboard SpaceX’s Crew Dragon, Isaacman has returned to Space Camp in Alabama on multiple occasions to meet with participants and share a bit of the awe he experienced as a kid. In 2022, a year after the first of these flights, Inspiration4, Isaacman donated $10 million to kick off a Space Camp expansion.Now, as the leader of NASA’s space program, he has sought to continue to pay it forward. He donates his salary to Space Camp, and on Friday, he returned to open the new “Inspiration4 Skills Training Complex,” a 50,000-square-foot facility completed with an additional $15 million donation from Isaacman. The money will also support development of a new dormitory.Space Camp becomes part of the US cultureSpace Camp has carved out an influential place in American culture. A 1986 feature film,Space Camp, popularized the camp when four teenagers befriended a robot and ended up launching into orbit on a space shuttle. The plot is implausible, to say the least, and the movie was criticized because it came out just a few months after the space shuttleChallengercatastrophe.Nevertheless, the movie helped build the myth of Space Camp, which had only opened four years earlier. Since then, more than 900,000 children ages 9 to 18 years old have “graduated” from the experience. Included among them are at least half a dozen NASA astronauts, including Dottie Metcalf-Lindenburger, Kate Rubins, Serena Auñón-Chancellor, and a mission specialist on the recently flown Artemis II mission, Christina Koch.Some of the exhibits and activities at Space Camp had grown a bit dated, but Isaacman’s gift should help to revitalize the programs.Among the new elements are a parachute “simulation room” that features a virtual-reality 10,000-ft drop with wind simulation, a drone lab, an interactive mission control, and “training simulators” that will recreate the experience of zero gravity.Camp registrations double after Artemis IIThe Artemis mission, flown by Koch and three other astronauts, appears to have reignited interest in the space program. Space Camp officials say registrations have doubled this summer since the successful completion of the lunar flyby mission in April.Interest is expected to grow further in the coming years as NASA returns to the Moon and establishes habitats there. The aim is to meet the needs of space-curious children with a modernized experience.“Off the success of Artemis II, America’s return to the Moon is just getting going,” Isaacman said during the ribbon-cutting event for the new facility.“And it is that kind of magic that inspires the next generation to attend Space Camp, get hands-on experiences at this national treasure, unlike anywhere else in the country, and grow up ready to pick up the baton and join in this great adventure.”Eric BergerSenior Space EditorEric BergerSenior Space EditorEric Berger is the senior space editor at Ars Technica, covering everything from astronomy to private space to NASA policy, and author of two books:Liftoff, about the rise of SpaceX; andReentry, on the development of the Falcon 9 rocket and Dragon. A certified meteorologist, Eric lives in Houston.26 Comments

engagementBMW sends off the 6th-gen M3 CS with a manual gearbox, rear-wheel driveThe 2027 M3 CS Handschalter is lighter and comes with three pedals instead of two.Jonathan M. Gitlin–May 18, 2026 10:33 am|48For the first and last time, you can get a stick shift in a G80 BMW M3 CS.Credit:

      BMWFor the first and last time, you can get a stick shift in a G80 BMW M3 CS.Credit:

      BMWText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe march of time, and what counts for progress in the automotive industry, has not been particularly kind to the driving enthusiast. Our vehicles have gotten bigger and heavier. Touch-sensitive panels and screens replaced buttons. Steering feelevaporatedabout a decade ago. And if you’re a fan of changing your own gears with a stick shift and three pedals,things have been looking bleakfor a while now. Which makes BMW’s send-off for its current sixth-generation M3 so notable.BMW’s M divisionkept the six-speed manual alive for the G80 M3, butonly the normal version. If you wanted the more powerful, much torquier M3 Competition or the track-focused M3 CS (Competition Sport), the only transmission choice was an eight-speed automatic. That automatic happens to be the excellent ZF 8HP gearbox, and for being fast on track, I’d still choose it, because that makes left-foot braking easier.Using paddle shifts might be faster, but I won’t pretend it’s more engaging than coordinating the movement of a gearstick through its gate, timed properly to the action of the clutch—especially if you’re heel-and-toeing, but even if you use the auto-blip feature that revs the engines for you on downshifts now. BMW appears to recognize that too, because it says the 2027 M3 CS Handschalter is designed for maximum driver engagement, and just for North America.Yes, that is a big carbon-fiber hump in the middle of that seat.BMWYes, that is a big carbon-fiber hump in the middle of that seat.BMWNew, lighter wheels.BMWNew, lighter wheels.BMWImola red paint is a $4,500 option.BMWImola red paint is a $4,500 option.BMWNew, lighter wheels.BMWImola red paint is a $4,500 option.BMWThe $107,100 M3 doesn’t use the more powerful engine from the M3 Competition, but the same 473 hp (353 kW) inline six-cylinder engine as the regular six-speed M3. But it’s a lighter car, to the tune of about 75 lbs (34 kg) thanks to things like a titanium exhaust muffler, carbon-fiber seats, carbon ceramic brakes, lighter wheels, and plenty of carbon-fiber reinforced plastic for body panels. And instead of the ZF 8HP, there’s a six-speed manual, which sends its power to the rear wheels and the rear wheels only: There’s no xDrive AWD here, unlike the M3 Competition.BMW M’s next major work will be next year’sall-electric quad-motor M3, derived from theNeue Klasse i3. And we can’t imagine there’s a three-pedal version of that one in the works, sadly.The M3Ireally wantIn BMW’s briefing materials for the M3 CS Handschalter, it notes that the 3.0 L S58 engine is also used, in modified form, in the BMW M4 GT3 Evo that won the Rolex 24 at Daytona this year. BMW has had a fair degree of success with the M4 GT3 Evo, as it also won last year’s Nurbürgring 24, among other races. But the only M race car with an S58 engine anyone cared about at this year’s N24 started life as anApril Fool’s joke.As 24-hour races go, the N24 is certainly unique. Le Mans and Daytona have faster prototypes among the GTs. Spa-Francorchamps is just GT3, but more than 70 of them on track together. But the N24 combines everything from GT3 cars down to Volkswagen Golfs, and does so across not just the modern F1-grade GP Circuit but also the entire stretch of the Nordschleife, with its narrow straights, crests, drops, and dozens of blind corners. Watching driver onboards is frankly terrifying, given the speed differentials.The BMW M3 Touring at the N24.Christian Kaspar-Bartke/Getty ImagesThe BMW M3 Touring at the N24.Christian Kaspar-Bartke/Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesThis year’s race had a pair of stars. Max Verstappen took time off from his F1 job, got his Nürburgring race permit, and then showed that he could absolutely cut it with the world’s best sports car drivers on the w

about timeFive years later, Windows 11 brings back much-missed taskbar options (and more)Microsoft is also testing a smaller taskbar and more customizable Start menu.Andrew Cunningham–May 18, 2026 12:03 pm|126Windows 11’s work-in-progress side-mounted Taskbar.Credit:

      MicrosoftWindows 11's work-in-progress side-mounted Taskbar.Credit:

      MicrosoftText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen Windows 11 launched in 2021, we mostly liked its refreshed look—the rounded corners and menus with just a hint of translucency were a nice change from the flat colors and hard corners of the Windows 8/10 era. But its reformulated taskbar and Start menu came with a number of functional regressions from the versions in Windows 10. Some of these were addressed quickly; others continue to linger.A new Windows Insider Preview build released to testers includes a new wave of improvements that fix longstanding regressions while trying out new things.Most significantly, the Windows 11 taskbar can now be docked to any edge of your screen, including the left and right, something that was possible in Windows 10 (and many older versions of Windows) but has been missing from Windows 11 since launch. Users can configure slightly different taskbar behavior for every taskbar position—if you prefer a different icon alignment or a left/right-mounted taskbar over a top/bottom-mounted taskbar, or if you want different settings for labels and icon groupings, you can choose different options for each position and Windows will remember them.Microsoft says there are several features that haven’t been implemented yet—the taskbar won’t auto-hide in any of the alternate positions, and the “tablet-optimized taskbar” with larger, more finger-friendly icon sizes and spacing also isn’t supported. Touch gestures and the Search box also aren’t supported. All of these features are coming at some point; they just aren’t ready now. Microsoft is “evaluating additional features like different taskbar positions per monitor” for multi-monitor setups.Another change Microsoft is testing makes the taskbar and all its icons smaller, a change designed to increase the amount of vertical space available on smaller screens without requiring users to fully hide the taskbar.The Start menu with recommended and All Apps sections disabled, along with the shorter taskbar.Credit:
                                  MicrosoftThe Start menu with recommended and All Apps sections disabled, along with the shorter taskbar.Credit:

      MicrosoftMicrosoft is also making a handful of changes to the Start menu, including a user-selectable size setting (previously, the menu would increase and decrease in size dynamically based on the size of your display). Each section of the Start menu (the pinned apps section, the “recommended” section, and the “all apps” section) will be individually toggleable. And users will be allowed to hide “recommended” apps advertised from the Microsoft store while still being able to see jump lists and recent files in the File Explorer. Users who decide to keep the “recommended” apps section visible should also benefit from “improv[ed] file relevancy” that “better reflect[s] what you have been working on.”Some of these changes are available in the current Windows Insider Preview builds in the Experimental channel (which replaced the Canary and Dev channels in Microsoft’slatest beta program shake-up). Others will be released “over the coming weeks,” at which point more polished and refined versions will presumably come to the Beta channel and, eventually, the public version of Windows 11. All of these changes are being made as part of Microsoft’s continued “commitment to Windows quality” push, meant to address several of the real and perceived shortcomings of Windows 11 relative to older versions.Andrew CunninghamSenior Technology ReporterAndrew CunninghamSenior Technology ReporterAndrew is a Senior Technology Reporter at Ars Technica, with a focus on consumer tech including computer hardware and in-depth reviews of operating systems like Windows and macOS. Andrew lives in Philadelphia and co-hosts a weekly book podcast calledOverdue.126 Comments

Hands-onThe Dory Sign is E ink, smart screen simplicity at its finestDory CEO claims the $149 signs won’t be bricked should Dory go out of business.Scharon Harding–May 18, 2026 1:26 pm|63I customized the Dory Sign to introduce my dog. The polka dot frame is optional, removable, and costs $29 extra.Credit:

      Scharon HardingI customized the Dory Sign to introduce my dog. The polka dot frame is optional, removable, and costs $29 extra.Credit:

      Scharon HardingText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navMany gadgets marketed as being “smart” make me wonder if they would be better off dumb.Some examples aresmart TVsthat insist on sending your activities to businesses totrack you, smart fridges that use the Internetto cycle through ads,smart gym equipment that won’t work offline, smart toothbrushes whosebatteries drain too quickly, or virtually any gadget that forces you to use a minimally effective or otherwise unimpressive app.Too often, modern technologies, like inter-device connectivity and artificial intelligence, are shoehorned into gadgets that would be more intuitive to use, affordable, accessible, and/or durable without them.Dory Sign, however, is a reminder of how technology can improve something as simple as a sign without overshadowing the product’s most basic purpose, which in this case is effective and delightful communication.Dory is a small sign that, like many E Ink displays, is easy on the eyes because it doesn’t use bright lighting. You control what the sign displays through a free iOS or Android app that doesn’t force you to share your email or name to use it. The app has clear sections for editing the header text, main text, and footer text, adding an image, and choosing a background. It also allows more than one person to make changes to the display and communicates to the sign through Bluetooth.A screenshot of the Dory Sign app. Changes made to text, the image, and background appear in the preview image in real-time.Credit:
                                  Scharon Harding/DoryA screenshot of the Dory Sign app. Changes made to text, the image, and background appear in the preview image in real-time.Credit:

      Scharon Harding/DoryYou can upload your own image or background, but the app comes with a decent range of preloaded images, like animals, flowers, and illustrations, and preloaded backgrounds, including ones that look like brush strokes or marble and more detailed ones, like a landscape. You don’t have to spend a lot of time designing the sign, but if you want to get creative, the app has six different typefaces for text and sliders for text size, line height, letter spacing, and text color (which makes text darker or lighter).The sign can only display one image, though. And while transferring your preferred text, image, and background is simple enough with the app, your phone has to be pretty close to the sign for it to work. You also have to press the power button every time you want to change the display. This can be tedious, especially if you have the sign mounted to a wall. To be fair, though, manually changing a regular sign also requires you to get close to and touch the display.The E Ink screen doesn’t lend itself to vibrant imagery. Instead, it’s a more passive display that doesn’t call for attention, yet clearly communicates with distinct detail in its text and images.The “House party” background by ANTI that’s included in the app.Scharon Harding/ANTIThe “House party” background by ANTI that’s included in the app.Scharon Harding/ANTIThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathDory Sign’s makers claim gadget can last for 15 yearsAt $149, the Dory Sign is expensive for a small display. However, you may be able to extract that much value from it over the years. Dory claims that the sign only uses power when it’s changing what’s on the display.

“are you afraid?”Guy Gardner makes a cameo in newLanternsteaser“This ring? It’s the greatest weapon in the Universe. When and if to use it, that’s the whole ball game.”Jennifer Ouellette–May 18, 2026 1:36 pm|10Credit:

      YouTube/HBO MaxCredit:

      YouTube/HBO MaxText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navLanterns, the new DC Universe series coming to HBO Max, dropped asurprising teaser in Marchthat swapped the usual superhero hijinks for gritty realism more in the vein ofTrue DetectivesandSlow Horses. Personally, I liked the change; it pushed a series that was barely on my radar to my 2026 list of ones to watch. I guess HBO was concerned people would miss the superhero vibes, though, because the latest teaser trailer interweaves the grittier aspects with a lot more superpowers and intergalactic elements. And you know what? I’m still here for it.Per the official logline, “The series follows new recruit John Stewart (Aaron Pierre) and Lantern legend Hal Jordan (Kyle Chandler), two intergalactic cops drawn into a dark, earth-based mystery as they investigate a murder in the American heartland.” There will be two storylines: one set in 2016 about a murder in Nebraska, and the second set in 2026.Chandler’s Hal Jordan is a former test pilot nearing retirement from the Green Lantern Corps. He’s training Pierre’s John Stewart Jr., a new recruit, to replace him. Nathan Fillion reprises hisSupermanrole as the obnoxious Guy Gardner; we get a brief glimpse of him in the new teaser. The cast also includes Kelly MacDonald as Kerry, a small-town family-oriented sheriff; Jason Ritter as Billy Macon, Kerry’s husband; Garret Dillahunt as William Macon, Kerry’s cowboy father-in-law; Poorna Jagannathan as a woman named Zoe; Ulrich Thomsen as Sinestro, a former Corps member who’s gone rogue; and Paul Ben-Victor as an extraterrestrial called Antaan.Sherman Augustus plays John Stewart Sr., with J. Alphonse Nicholson playing the younger version; Nicole Ari Parker plays Bernadette Stewart (mother to John Jr.), with Jasmine Cephas Jones playing the young version of the character. In addition, Chris Coy plays a suspiciously nervous truck driver, Waylon Sanders; Cary Christopher plays a gifted child named Noah; and Laura Linney and Paula Patton will appear in as-yet-undisclosed guest roles.We also know that the show will feature the inter-dimensional prison and the Checkmate organization introduced inPeacemaker’s S2 finale. DC Studios co-CEO James Gunnhas saidthat the events will set up other things within the DCU as a whole. So, is it a gritty buddy-cop series or a superhero series? Showrunner Chris Mundyrecently told Men’s Healththat it’s very much both.“Our show is in a lot of ways about replacement—when should someone step aside and when is it time for the next person to take the reins?”Mundy said. “That push and pull between those two characters is really important. So much of the power that John has is by not taking the bait, understanding that you lose your power if you’re yelling and screaming. That’s what we’re trying to convey: He knows he belongs, so he doesn’t have to overcompensate. There’s a real balance there that’s just innately inside of Aaron. He’s big. He’s an intimidating presence just physically. But there’s a softness to him, too. There’s a thoughtfulness. You can’t teach that.”Lanternspremieres on August 16, 2026, on HBO Max.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.10 Comments

from the garden of fugitivesPompeii victim ID’d as a likely doctorNew X-rays and CT scans showed small case with locking mechanism containing metal instruments.Jennifer Ouellette–May 18, 2026 2:03 pm|9Credit:

      Pompeii Archaeological ParkCredit:

      Pompeii Archaeological ParkText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navArchaeologists used a combination of advanced CT scans and 3D digital reconstruction to identify one of the Pompeii victims who died in 79 CE during the eruption of Mount Vesuvius as most likely having been a Roman doctor, according toan announcementby the Pompeii Archaeological Park.Aspreviously reported, the eruption of Mount Vesuvius released thermal energy roughly equivalent to 100,000 times the atomic bombs dropped on Hiroshima and Nagasaki at the end of World War II, spewing molten rock, pumice, and hot ash over the cities of Pompeii andHerculaneumin particular. The vast majority of people in Pompeii and Herculaneum—the cities hardest hit—perished from asphyxiation, choking on the thick clouds of noxious gas and ash. But at least some of the Vesuvian victimsprobably died instantaneouslyfrom the intense heat of fast-moving pyroclastic flows, with temperatures high enough to boil brains and explode skulls.In the 19th century, an archaeologist named Giuseppe Fiorelli figured out how to make casts of those frozen bodies by pouring liquid plaster into the voids where the soft tissue had been. Some 1,000 bodies have been discovered in the ruins, and 104 plaster casts have been preserved.Restoration effortson 86 of those casts began about 10 years ago, during which researchers took CT scans and X-rays to determine whether complete skeletons were present.The CT scans and X-ray images revealed that there had been a great deal of manipulation of the casts, depending on the aesthetics of the era in which they were made, including altering some features of the bodies’ shapes or adding metal rods to stabilize the cast, as well as frequently removing bones before casting. Ancient DNA analysisin 2024of four victims found in what is known as the “House of the golden bracelet”revealed thatall four bodies were male and none were genetically related—challenging the pre-existing preferred narratives and suggesting those may reflect certain cultural biases.The small case linked to the Roman physician found inside the Pompeii cast.Credit:
                                  Pompeii Archaeological ParkThe small case linked to the Roman physician found inside the Pompeii cast.Credit:

      Pompeii Archaeological ParkThis latest discovery involves one of the plaster cast remains of 14 victims found in what is now known as theGarden of the Fugitives, who died when they were overtaken by the pyroclastic flow while trying to flee via the Nocera Gate. Pompeii superintendent Medeo Maiuri excavated the victims in 1961, originally found in three smaller groups, although today the plaster casts are displayed in a line.Maiuri published a fanciful, largely fictional account of who he believed those bodies to be in National Geographic that same year: a merchant with severe osteoarthritis; a mother and her two young children; a young couple with an infant daughter; and a servant with what appeared to be a shoulder bag, although the “bag” later turned out to just be a malformation in the plaster.For decades, archaeologists had overlooked a small object trapped within another of the plaster casts that had been kept in storage for decades at the archaeological park. The new analysis with X-ray imaging and CT scans revealed that it was a small case with parts made out of metal—specifically, a toothed wheel operating a locking system. The case contained a coin-filled fabric pouch and several metal instruments similar to what a Roman physician (medicus) might have used, leading the team to conclude the man had likely been a doctor, fleeing Pompeii with his instruments close at hand.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.9 Comments

A little too lateElon Musk took too long to sue OpenAI, jury unanimously agreesMusk plans to appeal after judge immediately affirmed the jury’s decision.Ashley Belanger–May 18, 2026 2:09 pm|154Elon Musk attends a state banquet hosted by Chinese President Xi Jinping in Beijing during a summit with Donald Trump.Credit:Alex Wong / Staff | Getty Images NewsElon Musk attends a state banquet hosted by Chinese President Xi Jinping in Beijing during a summit with Donald Trump.Credit:Alex Wong / Staff | Getty Images NewsText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navElon Musk took too long to file his lawsuit that accused OpenAI of stealing a charity, a nine-person jury unanimously decided Monday.Musksued OpenAI in 2024for making a “fool” out of him after Musk donated $38 million to kick-start OpenAI as a nonprofit, only to later be blindsided when OpenAI created a for-profit arm that he felt gutted funding for the charity while enriching executives like Sam Altman and Greg Brockman.But the jury found that Musk was aware of OpenAI’s restructuring plans as early as 2021 and therefore missed the statute of limitations requiring him to bring the lawsuit within three years, The New York Timesreported. Because Musk took too long to file the litigation, the jury deemed Altman and Brockman not liable for any of the claims that Musk brought against OpenAI, the NYT reported. The jury also let Microsoft off the hook, finding no liability for the OpenAI investor after Musk alleged they aided OpenAI’s get-rich scheme.Judge Yvonne Gonzalez Rogers almost instantly agreed with the jury’s decision, the NYT reported.She thanked the jury and reminded them that they can talk to “anyone about anything” once they feel comfortable discussing the case, the NYT reported.Musk’s legal team looked “sunken in their chairs” after the jury’s decision was announced, the NYT reported. One of the billionaire’s lawyers, Marc Toberoff, who hosted daily calls with press during the trial, quickly confirmed that Musk will appeal.Altman and Brockman were not present when the verdict came in, the NYT reported. However, there were “a few smiles from the OpenAI side of the room,” with OpenAI attorney William Savitt wearing “a wide grin,” the NYT reported. Savitt spent daysgrilling Musk on the stand, seemingly taking advantage of his prior experience as Musk’s lawyer to get under Musk’s skin.In a statement, Microsoft celebrated the win, writing,“The facts and the timeline in this case have long been clear, and we welcome the jury’s decision to dismiss these claims as untimely. We remain committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”Musk did not return to the courtroom after delivering his testimony and was also not there when the verdict was read. Before the verdict arrived, Musk’s lawyer apologized after Musk shirked a court order requiring him to be available in case any further testimony was needed in the case. In an apparent sign that he was losing interest in the trial, Musk prioritized attendinga summit between Donald Trump and China’s President Xi Jinping.Late Monday, Musk issued a statement on X, writing, “Regarding the OpenAI case, the judge & jury never actually ruled on the merits of the case, just on a calendar technicality. There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!”Musk reiterated his lawyer’s statement confirming he plans to appeal, writing, “I will be filing an appeal with the Ninth Circuit, because creating a precedent to loot charities is incredibly destructive to charitable giving in America.”This story was updated to include Musk’s statement.Ashley BelangerSenior Policy ReporterAshley BelangerSenior Policy ReporterAshley is a senior policy reporter for Ars Technica, dedicated to tracking social impacts of emerging policies and new technologies. She is a Chicago-based journalist with 20 years of experience.154 Comments

Can I pet that dog?Australian Aboriginals cared for a dingo’s grave for decadesFor some ancient Aboriginal Australian communities, dingoes were part of the family.Kiona N. Smith–May 18, 2026 2:37 pm|40Archaeologists and Barkindji custodians worked together to excavate the dingo burial.Credit:

      Barb QuayleArchaeologists and Barkindji custodians worked together to excavate the dingo burial.Credit:

      Barb QuayleText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA thousand years ago, the ancestors of today’s Barkindji people carefully buried a dingo (or garli, in the Barkindji language) in a mound of shells.Archaeologists recently studied the burial in what’s now New South Wales, Australia. They found that the Barkindji ancestors had buried the dingo with the same care and ceremony as any beloved human member of the community and looked after the grave for centuries. The burial reveals that dingoes were, as Australian Museum and University of Sydney archaeologist and study co-author Amy Way puts it, “deeply valued and loved” by ancient people in Australia.The long-lost dingoFive years ago, Barkindji Elder Uncle Badger Bates and National Parks and Wildlife Service archaeologist Dan Witter saw bones eroding out of a road cut in Kinchega National Park, an area along the Baaka, or Darling River, in New South Wales, Australia. Badger recognized the bones as a dingo, lying on its left side in what was once a carefully built mound of river mussel shells.At the urging of the Menindee Aboriginal Elders Council, which worried that erosion would end up destroying the dingo bones and any information about the past they contained, a team of archaeologists, working alongside Barkindji elders, excavated and studied the skeleton. The bones turned out to belong to an elderly male dingo, with worn teeth and possible signs of arthritis. Broken and healed bones suggested that he’d lived a tough, active life but also been cared for by people.And the layers of shells around him revealed that generations of Barkindji had tended his grave and ritually “fed” him by adding shells to the mound for centuries after his death. This is definitely not the first dingo burial ever found in Australia, but it’s farther north and west than any other example. It reveals a far more profound and lasting relationship between ancient people and dingoes than outside researchers, at least, had previously fully realized.“This confirms these traditions were much more widespread than we once thought,” said University of Western Australia specialist Loukas Koungoulos, the lead author of the paper, in a press release.The dingo’s bones revealed healed injuries and possible arthritis.Credit:
                                  Amy WayThe dingo’s bones revealed healed injuries and possible arthritis.Credit:

      Amy WayHunting kangaroos and snoozing by the fireThe dingo’s bones tell their own story. Koungoulos says he was probably between 4 and 7 years old, which would be late middle age for a wild dingo today. Heavily worn teeth were the first hint of the dingo’s senior citizen status, but the ends of his leg bones also showed signs of bone decay, probably thanks to long-term inflammation: possibly something like arthritis.And he was shorter than most wild dingoes, based on the length of his femurs. That’s not unusual—domesticated animals are often shorter than their wild relatives, and it doesn’t take many generations for that to show up—but it could say something interesting about exactly how close wild dingoes got to domestication in the centuries before European colonists wrecked everything.At some point, the dingo had suffered a broken rib and lower leg. Koungoulos suggests the injuries look like the aftermath of a kangaroo kick and may have happened on a hunt. The injuries themselves aren’t too surprising; wild dingoes hunt kangaroos, andAboriginal hunters worked with dingoesthe same way people in other parts of the world have hunted with dogs for millennia. What’s more striking is that the two injuries were long-since healed. Somebody nursed this dingo back to health after his kangaroo encounter.“What stands out about garli is that he was old and well-cared-for,” said Koungoulos. “The healed injuries, worn teeth, and careful burial tell us that this animal lived a long life alongside people, and that his death was marked intentionally and with respect.”Barkindji elders stand with the excavated dingo skeleton; in the center is Uncle Badger Bates, who first identified the bones as a dingo.Credit:
                                  Amy WayBarkindji elders stand with the excavated dingo skeleton; in the center is Uncle Badger Bates, who first identified the bones as a dingo.Credit:

      Amy WayHow dingoes became beloved community membersPeople have lived in this part of Australia for at least 40,000

Who will win mtn?One Mars spacecraft, two senators, and a cloud of questions“I think there’s plenty of fire lit under them already.”Eric Berger–May 18, 2026 3:49 pm|16A rendering of Rocket Lab’s design for the Mars Telecommunications NetworkCredit:

      Rocket LabA rendering of Rocket Lab's design for the Mars Telecommunications NetworkCredit:

      Rocket LabText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navNASA released a much-anticipatedcontract solicitationfor a Mars-orbiting spacecraft late last week, kicking off what is sure to be a hotly contested and potentially controversial procurement.At issue is $700 million, already appropriated by Congress, to build a spacecraft, launch it to Mars, and once there to serve as a vehicle to relay communications between the red planet and Earth. But the stakes may be even bigger than this, including the possible resurrection of the recently canceled Mars Sample Return mission.As part of the new solicitation, NASA says it will conduct the acquisition “as a full and open competition.” But will it? That’s the question that several people involved with this procurement process are asking. And it could turn messy, quickly.What is not controversial is that NASA needs a new spacecraft capable of relaying communications from Mars to Earth. NASA’s best communications relay today is the Mars Reconnaissance Orbiter, which has now been there for 20 years. It’s a great spacecraft, but it’s getting long in the tooth.The US Congress, more or less out of the blue, stepped up with $700 million in funding for a new Mars Telecommunications Orbiter in the “One Big Beautiful Bill” passed in 2025.Ars previously wrote about this legislation, which raised some eyebrows, as this is a large appropriation for a relatively straightforward spacecraft.Competition is narrowedProposals to build the spacecraft, the name of which has been changed to Mars Telecommunications Network (MTN), are due by June 15. A contract is expected to be awarded by October 1, less than five months from now.There was some curious wordingin the legislationthat funded the Mars orbiter. It specified the spacecraft must be selected from among US companies that “received funding from the Administration in fiscal year 2024 or 2025 for commercial design studies for Mars Sample Return; and proposed a separate, independently launched Mars telecommunication orbiter supporting an end-to-end Mars sample return mission.”Similar language ended up in NASA’s procurement notice, which states that to be eligible to bid, a company needs to “demonstrate that the Offeror proposed a separately and independently launched Mars telecom orbiter to support an end-to-end Mars Sample Return mission.”That seems like a curious requirement. Why must the bidder for a relatively straightforward orbiter around Mars have previously proposed an orbiter as part of an “end-to-end” mission to return samples from Mars? Was Congress seeking to be preferential for a particular company?Some hints about what is happeningThe One Big Beautiful Bill was signed into law on July 4, 2025. One month later, as part of itsQ2 Investor Update, Rocket Lab included a slide about its plans to support human missions to Mars. In the corner of the slide, the company noted that $700 million had just been appropriated for a Mars orbiter (MTO).“Rocket Lab was the only commercial provider to propose a MTO as part of an end-to-end Mars Sample Return mission to support human exploration and science missions to Mars,” the slide stated.Slide from Rocket Lab Q2 2025 Investor Update.Credit:
                                  Rocket LabSlide from Rocket Lab Q2 2025 Investor Update.Credit:

      Rocket LabRocket Lab is one of several companies eligible to compete based on the requirement that bidders have received funding in 2024 or 2025 for a “commercial design study” for a Mars Sample Return mission. Other eligible bidders include: Blue Origin, L3Harris, Lockheed Martin, Northrop Grumman, SpaceX, Quantum Space, and Whittinghill Aerospace.At an “Industry Day” event earlier this year, NASA officials indicated that the contract solicitation would be released by May 1. However, it was two weeks late. Weeks matter when NASA is seeking to have this spacecraft built and launched before the end of the next Mars launch window in late 2028.Ars’ Stephen Clark recently asked Rocket Lab chief executive Peter Beck if he knew why this delay had occurred. He said he did not. Beck was also asked whether NASA was moving quickly enough given the schedule constraints.“I think there’s plenty of fire lit under them already,” Beck said. “They understand the importance of the mission. So I think we just need to get it out there, and they need to get it awarded, and everybody needs to get to work.”One source indicated to Ars that this delay was due in part to a letter sent by US

Did AI do that?Legal fail: Don’t use AI to sue Facebook users for calling you a bad dateFake citations dashed a dude’s “Are We Dating the Same Guy” revenge lawsuit.Ashley Belanger–May 18, 2026 4:27 pm|47Credit:pikepicture | iStock / Getty Images PlusCredit:pikepicture | iStock / Getty Images PlusText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAn attempt to pressure Meta into removing a critical post from a Chicago Facebook group called “Are We Dating the Same Guy” may end in sanctions for lawyers whose takedown arguments appeared to rely on fake AI citations to support doxing claims.The case had already been dismissed with prejudice by a district court, which ruled there was no way to amend the complaint to possibly save it. But Nikko D’Ambrosio—who accused more than two dozen women of defaming him and blamed Meta for supposedly boosting the post to profit off its “entertainment value”—appealed anyway.Perhaps he felt confident despite his likely tough odds because he was relying on MarcTrent.AI, a law firm thatclaimsto use AI to “uncover legal opportunities traditional firms miss” and “increase legal success rates by 35 percent through predictive modeling.”In a 2025blogdiscussing the case, founder Marc Trent confirmed that the firm had “utilized our tech team to draft” the initial complaint. He boasted that the “evolved” firm uses “everything related to AI now,” suggesting that “even Meta can’t beat us” and claiming that a win would make Facebook safer for everyone.Laying out the case, Trent said that he assumed that Meta “would quickly distance itself by removing the post.” But when Meta didn’t, he figured that overcoming Section 230 claims would be his biggest hurdle in the fight. However, he insisted that “his firm’s technological capabilities” would level the playing field, making it possible to beat Meta’s well-resourced legal teams who are deeply schooled at defending against Section 230 claims.However, during the appeal, judges agreed that the case was so weak that Section 230 didn’t even factor in. And the firm’s seeming reliance on AI to “execute” arguments “with precision” apparently did not help matters at all.In anopinionFriday, David Hamilton, a senior circuit judge for the US Court of Appeals for the Seventh Circuit, wrote that the three-judge panel agreed that “this is a relatively rare appeal in which sanctions appear to be appropriate.” Not only was the appeal deemed “frivolous” for failing to advance D’Ambrosio’s arguments, but it was also littered with “mistakes and fictitious quotations” that “bear the hallmarks of the misuse of generative artificial intelligence,” Hamilton said.“Briefs and other court submissions that include fictitious quotations—inaccuracies discoverable with elementary professional care—are unacceptable,” Hamilton wrote.MarcTrent.AI did not immediately respond to Ars’ request for comment on whether the firm plans to fight the potential sanctions.The firm has until June 16 to request a hearing or file statements on whether sanctions are warranted.Man fails to scrub menacing text from FacebookD’Ambrosio’s legal fight started when a woman whom he briefly dated, Abbigail Rajala, blocked his number, and he persisted in sending a menacing text by using an alternate number.Screenshot of text D’Ambrosio wanted removed from Facebook.Credit: via Judge’s Opinion in D’Ambrosio v MetaScreenshot of text D’Ambrosio wanted removed from Facebook.Credit:

Escalated quicklyEbola outbreak: WHO declares emergency, US restricts travel, American infectedCDC is working to move the infected American and six others to Germany.Beth Mole–May 18, 2026 4:41 pm|41A poster displaying Ebola emergency contact numbers is pinned to a tent at the Busunga border crossing between Uganda and the Democratic Republic of Congo in Bundibugyo, on May 18, 2026.Credit:Getty | Badru KatumbaA poster displaying Ebola emergency contact numbers is pinned to a tent at the Busunga border crossing between Uganda and the Democratic Republic of Congo in Bundibugyo, on May 18, 2026.Credit:Getty | Badru KatumbaText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navThe Ebola outbreak first reported in the Democratic Republic of the Congo on Friday has seemingly escalated quickly into a large, uncontrolled multinational outbreak.As of May 17, there were 10 confirmed cases, 336 suspected cases, and 88 deaths in the DRC, as well as two confirmed cases and one death in neighboring Uganda, according to the latest data from the US Centers for Disease Control and Prevention, which has offices in the region. The numbers already put the outbreak within the top 10 Ebola outbreaks recorded by size, though still far from the worst—the 2014–2016 West African outbreak had over 28,000 cases and 11,000 deaths.International emergencyOn Sunday, the World Health Organization declared the outbreaka public health emergency of international concern (PHEIC), though it noted that it does not meet the criteria for a pandemic emergency. In making the PHEIC determination, WHO Director-General Tedros Adhanom Ghebreyesus cited several factors in addition to the immediate large size, including clusters of suspected cases and deaths in multiple DRC health zones, four deaths among healthcare workers, and a lack of apparent links between geographically distant cases and clusters. The features collectively suggest that the outbreak is larger than what is currently being detected and is spreading regionally.“Moreover, the ongoing insecurity, humanitarian crisis, high population mobility, the urban or semi-urban nature of the current hotspot, and the large network of informal healthcare facilities further compound the risk of spread,” the WHO said.The final extraordinary aspect of the outbreak is that it is caused by the uncommon Bundibugyo strain of Ebola virus, which has no clinically validated treatments or vaccines. This is only the third Ebola outbreak caused by Bundibugyo, which has had fatality rates of 25–50 percent.There are four virus strains known to cause Ebola disease in humans, and three have caused large outbreaks (Zaire, Sudan, and Bundibugyo). The most common strain is Zaire, for which treatments and vaccines have been developed. The viruses spill over from animals, including non-human primates and bats, and cause severe hemorrhagic fever, marked by diarrhea, vomiting, and bleeding. Person-to-person spread occurs via contact with bodily fluids and symptoms can develop between two and 21 days—though most often eight to 10 days—after an exposure.CDC response and infected AmericanOn Monday morning, the CDC announced on its website that it is implementingnew travel restrictions, including screening and monitoring Americans arriving from DRC, Uganda, and South Sudan, while also barring the entry of non-US passport holders who have traveled in those countries in the past 21 days.Additionally, in a CDC press briefing on Monday afternoon, Captain Satish Pillai, incident manager for CDC’s Ebola response, said that one American in the DRC has been infected after being exposed as part of their work there. The person developed symptoms over the weekend and tested positive late Sunday. The CDC is now working to transfer that person, along with six other Americans, to Germany, where they will receive care. Pillai did not answer questions about the person’s identity or their work.Serge, a Christian missionary organization, announced thatthe infected person is Dr. Peter Stafford, who has been working in the Nyankunde Hospital in Bunia, DRC, since 2023. The other six people the CDC is working to relocate are his wife, Dr. Rebekah Stafford, the couple’s four children, and a third doctor with the organization, Dr. Patrick LaRochelle. All three doctors had exposures, the organization said, but Rebekah Stafford and LaRochelle are currently asymptomatic.Pillai noted that the CDC considers the risk to the American public to be low.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.41 Comments

📱iPhone 17 Pro 官降 1000 元🧠ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施🤖估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相🌐Claude 重新允许接入龙虾，开发者：这是变相涨价📋腾讯辟谣「AI 一号位将离职」💰黄仁勋基金会捐出 1.08 亿美元 AI 算力📱影石 CEO 回应 Luna 定价贵：5299 元是美国价格🚗鸿蒙智行洽谈打造玛莎拉蒂新能源车🚙华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗💻英特尔成为迈凯伦官方计算合作伙伴⛽本田上市以来首次年度亏损📋LinkedIn 宣布裁员，涉及工程、产品和市场营销部门💡Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」🤖曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线📱荣耀 600 系列定档 5 月 25 日🔍Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光🧠腾讯开源 AI Agent 记忆系统，Token 消耗最高降低 61%🧸泡泡玛特否认进军传统家电，LABUBU 冰箱一季度收入约 600 万iPhone 17 Pro 官降 1000 元昨晚，苹果无预警在天猫「Apple Store 官方旗舰店」启动 iPhone 17 Pro 系列 618 优惠活动，Pro 和 Pro Max 同步降价 1000 元，起售价分别来到 7999 元和 8999 元。本次活动还覆盖 iPad、Apple Watch、AirPods 等系列，非 iPhone 产品可使用 50 至 100 元不等的惊喜券，并可叠加 88VIP 消费券、店铺优惠及国家补贴等多重优惠。ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施据彭博社报道，苹果与 OpenAI 的合作关系已出现明显裂痕，OpenAI 正在联合外部律所评估可能的法律行动。知情人士透露，OpenAI 律师团队正与外部律所合作，研究包括向苹果发送违约通知在内的多项方案，但目前尚未正式提起诉讼。OpenAI 仍希望在庭外解决与苹果的分歧，相关法律行动预计将在马斯克相关诉讼案结束后才会推进。苹果与 OpenAI 的合作始于 2024 年 6 月，苹果在 WWDC 上将 ChatGPT 整合进 Siri 及 iOS 系统，涵盖文字生成、视觉智能、图像生成 App「图乐园」等场景。估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相昨天，前 Meta FAIR 研究院总监田渊栋正式宣布联合创办 AI 初创公司 Recursive Superintelligence，距其确认从 Meta 离职刚好过去半年。他在 X 发文称，拒绝了 OpenAI、xAI、Anthropic 等多家头部 AI 公司的邀约，选择自己创业：「趁着自己还年轻，去当一家新初创公司的联合创始人。」Recursive 同步披露完成一轮 6.5 亿美元融资，估值达 46.5 亿美元，由 GV（谷歌风投）和 Greycroft 领投，AMD Ventures 与英伟达跟投。公司押注的核心方向为「递归自我改进」（Recursive Self-Improvement），即构建能够自动发现知识、持续优化自身的 AI 系统。Claude 重新允许接入龙虾，开发者：这是变相涨价Anthropic 昨天官宣，将于今年 6 月 15 日起为 Claude 订阅版用户引入独立的 Agent SDK 月额度，OpenClaw 等第三方 Agent 工具由此重新获准通过订阅账户接入 Claude。按照新规，Agent SDK 的调用将从用户的常规订阅限额中彻底剥离，另设专属月度额度池，额度按账单周期刷新，未用完的部分不结转至下月。Pro 版和 Team 标准席位每月获得 20 美元额度；Max 5x 及 Team 高级席位为 100 美元；Max 20x 与 Enterprise 座位制高级席位最高可达 200 美元。该额度专用于 Claude Agent SDK 项目（Python 或 TypeScript）、claude -p 非交互命令、Claude Code GitHub Actions 集成，以及通过 Agent SDK 认证的第三方应用。此次新规被视为 Anthropic 对今年 4 月 4 日封禁政策的正式解决方案。彼时，Claude Code 负责人 Boris Cherny 宣布，Claude Pro 及 Max 订阅额度即日起不再覆盖 OpenClaw 等第三方工具的调用，理由是此类工具的使用模式对 Anthropic 的算力基础设施造成了难以为继的压力。开发者社区对此反应冷淡。有开发者直言，20 美元的 Agent 额度在高强度任务下消耗极快，实际可用量远低于封禁前的订阅套餐，并将此次调整定性为「变相涨价」。腾讯辟谣「AI 一号位将离职」昨天，腾讯官方发文辟谣，明确否认外界流传的「AI 一号位即将离职」传言。腾讯公关总监张军表示，「太恶劣了，还编出各种搞笑的情节。」谣言起源于社交平台上的多条帖文，内容称某大厂 AI 一号位因被绿色事业部抢走算力资源而准备离职，并点名影射腾讯 AI 负责人姚顺雨。黄仁勋基金会捐出 1.08 亿美元 AI 算力据路透社报道，英伟达 CEO 黄仁勋与其妻子洛丽创立的基金会，从云计算公司 CoreWeave 购入 AI 算力，并将其捐赠给多所大学及其他非营利研究机构，用于科学和人工智能研究。根据昨日提交的一份文件，此次捐赠的算力价值目前已达 1.083 亿美元。英伟达还表示，计划向部分受捐机构提供免费工程服务。CoreWeave 是一家专注于 AI 应用的云计算公司，其向客户提供的 GPU 正是由英伟达设计制造。今年 1 月，英伟达向 CoreWeave 投资 20 亿美元，一度成为该公司第二大股东。影石 CEO 回应 Luna 定价贵：5299 元是美国价格昨天，影石 CEO 刘靖康回应旗下旗舰双摄云台相机 Luna Ultra 的定价问题，表示网传单机 5299 元的售价属实，「但是是美国的价格，国内给我再大的胆子都不敢卖这个价格」。此前，博主「数码闲聊站」爆料称该机预计售价 5299 元、全能套装预计 6499 元，随即在网络上引发「定价过高」的质疑。鸿蒙智行洽谈打造玛莎拉蒂新能源车据《云见 Insight》报道，华为鸿蒙智行、江淮汽车、斯特兰蒂斯集团（Stellantis）及旗下玛莎拉蒂品牌，正就共同打造玛莎拉蒂品牌的新能源车展开洽谈。按照规划，四方合作模式与鸿蒙智行 「五界」 高度相似：华为主导产品定义、提供核心技术，江淮联合研发并负责生产制造，玛莎拉蒂提供造型设计与品牌背书。合作车型将分为国内和海外版本，国内版归属于尊界品牌，海外版将挂玛莎拉蒂车标，首款车型目前正处于造型设计阶段，计划于明年下半年量产。华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗华晨宝马昨日公布年度节水数据，显示其沈阳生产基地 2025 年单台生产取水量降至 1.4 立方米/台，同比降幅 10.8%，连续八年稳步下降。在水资源消耗总量方面，华晨宝马沈阳基地全年取水总量降至 75.2 万立方米，较 2023 年降幅超 40%，连续三年递减。再生水使用量超 68 万立方米，同比增长 71.9%，官方称节约用水量可满足近 9300 户沈阳家庭一整年的生活用水需求。华晨宝马大东工厂北厂区实现所有生产工艺自来水消耗为零，全部改用再生水替代；铁西工厂及里达厂区涂装车间冷却塔也已扩展使用市政再生水，厂区绿化灌溉此前已于 2024 年实现 100% 使用市政再生水。英特尔成为迈凯伦官方计算合作伙伴英特尔昨日宣布与迈凯伦车队达成多年期战略合作伙伴关系，正式成为迈凯伦万事达一级方程式车队、艾睿迈凯伦印地车队及迈凯伦 F1 模拟赛车队的官方计算合作伙伴。根据合作协议，英特尔至强处理器和酷睿 Ultra 处理器将为迈凯伦的核心性能负载提供算力支持，覆盖计算流体动力学、空气动力学分析、车辆动力学仿真、比赛策略分析等方面。英特尔的技术还将应用于赛道边缘计算、数字孪生与 AI 预测建模等场景。本田上市以来首次年度亏损昨天，本田公布了 2026 年 3 月期的合并财报，期内营收 217966 亿日元，同比微增 0.5%；归母净亏损 4239 亿日元，上一财年同期归母净利润为 8358 亿日元，由盈转亏。亏损主因是电动化战略大幅调整。本财年内，本田决定取消北美多款电动车型的开发与上市计划，由此确认减值、处置及合同赔偿等相关损失合计约 15777 亿日元，全部计入汽车业务。摩托车业务为本财年唯一实现正向盈利的板块。展望明年，公司预计营收 231500 亿日元，归母净利润恢复至 2600 亿日元。LinkedIn 宣布裁员，涉及工程、产品和市场营销部门据彭博社报道，职业社交平台 LinkedIn 昨日宣布裁员，涉及工程、产品和市场营销等多个职能部门。LinkedIn CEO Daniel Shapero 在发给员工的备忘录中表示，公司需要为用户提供更大价值，并实现更高的盈利水平，此次组织架构调整旨在为公司未来发展做好定位。LinkedIn 发言人随后确认了相关变动，称此次调整属于「常规业务规划」的一部分，但未披露具体裁员人数。💡 Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」据《商业内幕》昨日报道，Anthropic CFO Krishna Rao 在播客节目《Invest Like the Best》中透露，Claude 目前已承担公司 90% 以上的代码编写工作，并正在深度介入财务流程，推动白领工作从「执行」向「监督」转型。Rao 还透露，公司已借助 Claude 生成财务报表，每月财务审查流程在人工介入之前已达到「90% 至 95% 完成」的状态。他提到，部分内部报告的生成时间从原来的数小时压缩至 30 分钟。在人员配置方面，Rao 将 AI 的引入描述为生产力「加速器」。他称，由于 AI 的介入，员工得以将更多精力从信息收集转向决策判断，Anthropic 反而因此扩大了招聘规模。我们雇用了更多人，因为没有什么工作是做不完的。Rao 还描述了一种新的工作形态：员工越来越多地扮演 AI 系统的「管理者」角色，团队可以同时部署多批 AI 智能体并行处理项目。他将这一趋势概括为「每个人都在某种程度上成为管理者」，并认为由此带来的生产力潜力「才刚刚开始显现」。他同时强调，Anthropic 内部推崇「人才密度」而非「人才规模」，认为高密度的顶尖 AI 研究人才配合最强模型，才是真正具有竞争力的组合。曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线知名爆料者 Leo 昨日披露，GPT-5.6 研发已全面推进，预计最早下月正式发布。开发者在 OpenAI 内部 Codex 日志中发现了两个内部代号 —— ember-alpha 和 beacon-alpha，以及一条路由映射记录。日志显示，绝大多数模型调用仍指向 GPT-5.5，但有一条条目明确指向 GPT-5.6，暗示 Codex 环境可能已在测试新模型。此外，爆料者 Chetaslua 称 OpenAI 计划于本周四推出旗舰模型的「超快模式」，响应速度或可提升 2 至 3 倍，主要面向延迟敏感型任务。荣耀 600 系列定档 5 月 25 日荣耀昨日宣布，旗下荣耀 600 系列将于 5 月 25 日 19:00 正式发布。官方预热信息显示，荣耀 600 系列主打「4K 闪光微单 Live」影像方向，搭载 2 亿像素超清大底主摄，并内置 8600mAh 青海湖大电池。Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光据 Sources 报道，Google 计划在 5 月 20 日举行的 Google I/O 大会上

相信这样的场景很多人都不会陌生：在微信上聊工作、讨论旅行计划，甚至只是和朋友闲聊，聊着聊着突然想让 AI 帮忙总结重点、做个规划，结果却只能一段一段复制聊天记录，再手动粘贴进 AI 对话框里。再比如，中午小憩了一会，醒过来发现群里发了几百条消息，想吃瓜又得一条一条翻……如果能一键将聊天记录转发 AI，岂不美哉？这样的功能终于来了，但遗憾不是在微信里：现在，把微信和元宝更新到最新版本之后，聊天记录转发可以选择直接发给元宝来总结。虽然同属腾讯生态，但微信和元宝之间始终有种微妙的距离感——终于，这次两者之间的联动，又紧密了一点。爱范儿也第一时间体验了这个功能，看看它究竟好不好用。产品很美好，效果很一般要想把微信聊天记录转发给元宝，方法很简单，首先选中想要转发的聊天记录，选择「转发到其他应用」，就能直接发给元宝，可以补充一些提示词。整个过程有点像把讨论、聊天过程中零散的内容，让 AI 帮忙进行整理，因此这个功能首先最合适的场景，就是各种工作群的讨论和安排信息。直接处理聊天记录有一个好处，元宝可以将讨论精准和发言者匹配，很适合用来处理会议纪要和工作安排。旅游计划同样也是一个高频场景，因为行程一般都是靠你一言我一句的对话讨论出来，这时候就可以直接交给元宝处理，还能让它帮忙推荐旅游地点。还有一个有点邪修的用法：平时工作忙没时间看群，又担心错过什么热点或者话题，可以直接选中所有聊天记录，让元宝帮你看群帮你总结。又或者讨论了很久都不知道聚餐吃什么，也可以把讨论的过程选中，让元宝帮忙推荐。尝试了很多场景，大部分情况下都达到了预期，但总体来说，还有一些体验细节不够完善。目前元宝最擅长处理的聊天记录以文字为主，而如果聊天记录中包含文件，特别是涉及多个文件，效果会比较随缘，很多时候会无法读取文件。即使能成功读取，也不一定能按照提示词进行任务处理：整理发票有时候都未必能填对，让他根据文件做个思维导图，也经常失败。要知道，整理聊天里的发票真的是很主流的场景啊！我能很明显感觉到一个落差：明明将聊天记录发送给 AI 是一个很有价值的场景，但元宝交付的结果，总让我略感失望，并且忍不住想，如果能转发给豆包或者 ChatGPT，会不会更好？更深层的问题，是这个聊天记录转发的交互总给我一种「绕远路」的感觉。在此之前，元宝已经作为联系人存在于微信之中，处理聊天记录这件事，却依旧需要「跳转」这一整套外挂式流程， 过程中需要多次点击。或许是因为微信严格的隐私保护规则，元宝只能以「临时聊天」的形式处理聊天记录，不会留下历史记录，关闭后无法找回。对于微信来说，这样的功能似乎「可有可无」。如果微信想要 AI 化，他们或许会有更好的实现方式，例如直接做一个「AI 群友」，at 它就能回答问题或者处理聊天记录；以及一个常驻对话列表的 Agent，可以直接对话，可以直接接收聊天记录。但这件事没那么容易，原因已经是「老生常谈」：微信需要简约，需要克制，因为每一点小变化都会影响到十亿级用户。在元宝这个功能上线的同一天，腾讯发布了今年第一季度财报，在股东大会上马化腾表示，在微信里做 Agent 需要「更长远的考虑，大家要有点耐心，这个不是匆忙就能出来的。」腾讯的 AI 王牌是微信马化腾还提到腾讯的 AI 战略：原来一年前我们以为上了船，后来发现那个船漏水了，现在感觉站上去了，还坐不下去，还是希望船速能快一点。从财报来看，AI 已经成为了腾讯的重要投入，一季度研发投入达 225.4 亿元，同比 19%；资本开支 319.4 亿元，主要用于算力、存储等云端基础设施的扩张。今天中午，有自媒体传出「疑似腾讯 AI 一号位即将离职」的消息，下午就被腾讯公司火速辟谣，本身也向外界释放出一个信号：腾讯 AI，不容有失。根据 AICPB 最新榜单，在国产 AI 应用排名中，腾讯元宝排名第十，访问量为 1773 万，远逊于豆包（1.63 亿）和千问（4424 万）。在 AI 领域中，腾讯和苹果有着相似的处境：产品入口和用户规模都是顶级配置，AI 模型本身却还支撑不起这种期待。元宝 AI 社交应用「元宝派」生成式 AI 火了 3 年，从去年年底开始，AI 应用之间的战斗已经进入了「场景化」的比拼，不仅出现「龙虾」OpenClaw 这样的全新形态，传统的 ChatBot 交互模式也集成了生活服务，像阿里千问和字节豆包，都在积极将电商等各种生态内的优势服务，和自家的 AI 助手相结合。这本来应该是腾讯的舒适区，因为国内最强大、最全能的生活应用，也就是微信，就在腾讯手中。但很明显，微信对「AI 化」抱有非常慎重的态度。腾讯不能承受因技术或功能不成熟而扰乱微信庞大用户群的体验。只是对于腾讯来说，在 AI 性能落后其他对手的情况下，微信这张王牌的价值还是太重要了。，于是，我微信和元宝虽然没有完全敞开双臂拥抱彼此，也开始小心翼翼试探边界。究其原因，除了微信相对封闭的生态逻辑，还有两个问题没能被回答：微信究竟要如何「AI 化」？而元宝在其中又应该扮演什么样的角色？The Information 报道指出，即使腾讯在下重本加速建设自己的混元模型，微信方面还是认为混元当前水平不够拔尖，甚至开始为微信 Agent 测试第三方 AI 模型，这带来了隐私和整合上的挑战性。微信方面，爆料称其实一直在做自己的 AI Agent 模型技术，计划在年中发布一个微信智能体，可以和整个小程序生态相结合，提供打车、购物的服务，并且将以联系人和聊天的方式直接集成微信之中——但就像前面提到的，这件事太重大、影响面太广了，急不得。一旦这个产品问世，可能将是全国最多用户的 AI 入口。微信需要 AI，但未必需要「元宝」，这才是腾讯现在最微妙的问题。

大疆的 Pocket 系列三年不更新，一更新就来了两台——过去一个多月，Pocket 4 和 Pocket 4P 可以说是最受瞩目的相机，就在今晚的戛纳电影节上，Pocket 4P 以口袋电影机的身份在欧洲电影殿堂亮相，距离正式发布时间，相信不会太远了。与此同时，大疆也对 Pocket 3 进行了降价处理，于是我们看到了泾渭分明的 Pocket 系列产品线：Pocket 3：入门级的手持云台 vlog 机，2299 元Pocket 4：全能的手持云台相机，2999 元Pocket 4P：准专业级的口袋电影机，预计价格会来到 4000 元档相信有许多持币观望的读者，也和爱范儿抱着同样的疑问：三台 Pocket 我该怎么选？有了 iPhone，我还需要 Pocket 吗？结合实际体验，爱范儿一次性帮你理清楚。Pocket 三代同堂，怎么选？先仔细聊聊 Pocket 4P。虽然目前 Pocket 4P 依旧没有公布更详细的参数，但参数在本次形态更新面前，是最不重要的话题——从产品逻辑来看，Pocket 4 已经将单摄的手持云台相机推向较为成熟的状态。此时，增加一颗摄像头成为一件顺理成章的事情。绝大多数手机的主摄和单摄 Pocket，提供的都是广角视野。广角适合交代环境，适合把大山大河框进画面。但当你拿着设备走入拥挤的人群，广角的缺点就暴露无遗：画面太杂乱了。此时，一颗长焦镜头可以完成空间上的压缩，把无关的背景剔除出去。大疆将这颗多出来的摄像头定格在 3×，这是一个非常巧妙的数字。它比主摄更贴近人眼对焦后的视线范围。从我们上手体验来看，Pocket 4P 这颗 3× 镜头还提供了比较自然的虚化，非常适合拍摄更凸显主体的人像或特写画面。但对于用户来说，多的这一颗摄像头，并非绝对的正收益：Pocket 系列自诞生起，核心卖点就是「无感携带」，增加镜头，意味着增加重量、侵占内部空间，甚至可能牺牲一部分续航。其实这种担心，恰恰帮大疆划清了 4P 和 4 之间的界限。Pocket 4P 是给那些「创作者」准备的。他们清楚地知道自己要拍什么，素材最后会进入 Mac，在 DaVinci Resolve 或者 FCPX 里进行精细的剪辑。对他们来说，为了在极轻便的体积下获得原生焦段的景别切换，以及高度统一的素材规格，牺牲一点点重量完全值得。而对于大多数普通人来说，Pocket 4 才是那个更轻量、更理性的选择。虽然前代 Pocket 3 打造出千万台的神话，但其本身并不是一台完美无缺的相机——只有 USB 2.0 导致数据传输过慢、像素量过低以及动态范围不够大，这些都是问题。而 Pocket 4 提供了 3700 万像素，更高的像素意味着内置的无损 ISZ 变强了。在 Pocket 4 上，通过传感器裁切获得的 2× 焦段更为清晰，足以应付日常记录。除此之外，它还内置了 107GB 存储，带来了更好的动态范围。忘记带内存卡、传输速度和动态范围不够的问题，被一次性解决掉了。可以说，Pocket 4 就是目前单摄手持云台相机的完全体。至此，Pocket 4 系列两台机身的区别已经很清晰了：如果你追求生产力，强调原生光学焦段带来的画质与虚化，需要 3× 镜头获得更丰富的画幅景别，那么 Pocket 4P 毫无疑问是更好的答案；如果你只是希望有一台能轻松塞进裤兜、防抖优秀、可以独立于手机拍摄视频和 Live Photo 的机器，那么 Pocket 4 凭借更高的像素和好用的 2× ISZ，已经绰绰有余。最后是 Pocket 3。在 Pocket 4 补齐了大量痛点的情况下，Pocket 3 还有存在的意义吗？当然——新设备的出现，最大的贡献往往是把老旗舰的价格打下来。在 Pocket 4 系列轮番上场的间隙，我们观察到 Pocket 3 已经完成了一轮官方降价，在某些平台叠加各种补贴的情况下，已经落到了 2099 的价格。甚至在未来，我们能以 1 开头的价格买到 Pocket 3，也不是不可能。对于预算有限、或者刚刚萌生拍视频念头的新手，这就是性价比最高的 「守门员」。最后，按照爱范儿编辑部的推测，Pocket 4P 的定价大概率在 4500 元左右，但也不排除大疆卷起来定一个 3999 元的价格。此时，如果将 Pocket 系列按照价格排开，我们会发现一个巧妙的设计——从 2000 元档的 Pocket 3 到 3000 元档的 Pocket 4，最后到 4000 元档的 Pocket 4P，大疆正在把无人机上那套 Pro、Air、Mini 的组合拳搬到口袋相机上。这套阵容已经在无人机市场证明了它的成功。现在，Pocket 也将组成这样的阵容了。有了 iPhone，我还需要 Pocket 吗？在 Pocket 4P 的上手文章发布后，我收到不止一个朋友的问题：我有 iPhone，还需要买 Pocket 吗？理清了 Pocket 家族的内部定位后，这是我们必须面对的现实问题：既然大家手里都有一台性能强悍的 iPhone，为什么还要花钱买一台功能单一的云台相机？这其实是一个关于「设备属性」的哲学问题。在多次迭代后，iPhone 在动态影像领域的基础不可谓不强，三摄变焦顺畅、色彩高度统一、防抖优秀、Apple Log 甚至已经更迭到第二代。但问题在于，手机始终不是完全针对拍摄设计的特化产品——除了拍摄以外，iPhone 还需要承担复杂的日常任务，无时无刻需要分散注意力，随时可能会有工作消息、语音电话打断你的拍摄、需要考虑拍摄之外手机还有没有充足的剩余电量保证你与世界的关联，还需要注意拍摄会不会占满本就贵如黄金的存储空间……而 Pocket 则是完全服务于拍摄需求的相机：不能打电话，不能回微信，不能刷社交软件，一切与拍摄无关的东西，都不会出现在你手里的 Pocket 上。更别提，Pocket 还有机械云台天生在结构上比 iPhone 的防抖更稳定、形态比 iPhone 更适合握持、物理虚化比 iPhone 更自然、视频规格比 iPhone 更丰富等等综合优势了。所以，如果我手里已经有一台主力 iPhone，现在到底该不该买 Pocket？如果现在需要，哪台是最适合我的？你可以对照下面几个问题，做一个购前诊断：过去三个月里，你是否因为手机突如其来的电话或微信，打断过拍摄？拍视频时，是否因为手机存储空间不足，而束手束脚？炎炎夏日，有没有遇到过手机拍照录视频发热掉帧甚至死机的情况？如果你对这些问题的回答是肯定的，那么购入 Pocket 3 和 Pocket 4 都是不错的选择。具体来说，如果你有拍高清 live 图的需求，如果你喜欢用手机的 2× 镜头拍摄，那么选择 Pocket 4，否则，Pocket 3 其实就够用了。在这个基础上，你还想拿 Pocket 拍拍商单、拍拍演唱会，搞点严肃创作，那么 Pocket 4P 是一步到位的选择。当然，如果前面几道题，你的答案都是否定的，那么请捂紧钱包。iPhone 的防抖和画质足够满足你的需求。你可以随拍随剪，直接通过 AirDrop 分享，且丝滑的生态体验依旧是目前任何第三方设备都无法替代的。而从行业角度来看，爱范儿更建议你等一等——目前，口袋云台相机的玩家可不止大疆一家——和 Pocket 4P 同样「犹抱琵琶半遮面」的，还有影石 Luna，而各大手机厂商也正跃跃欲试，准备下场分一杯羹。因为 Pocket 4P 这样的产品，已经严重威胁到了手机厂商的安全区。很多人愿意买 6000 元甚至 8000 元的影像旗舰手机，很大程度就是为了长焦。现在，花更低的价格，Pocket 4P 就能提供一个素质相当的长焦镜头，以及更好的云台防抖、更专业的整机解决方案。没有手机厂商能对此熟视无睹。据爱范儿了解，vivo、OPPO 都已经将手持云台相机的产品规划提上日程，而小米、华为也在持续关注这个赛道。2026 下半年到 2027 上半年，想必会有更多手持云台相机纷至沓来。既然入局者多了，我们作为消费者，反而可以更淡定些。你可以把钱捏在手里，再等一等。我们相信，这个年轻的品类，还会有更精彩的产品出现。

鸿蒙智行车型阵营里可能缺一辆 MPV，但中国汽车市场已经不缺 MPV 了。别克 GL8，丰田埃尔法和传祺 M8 等车型在燃油车时代占据了 MPV 就是商务接待用车的印象之后，新能源 MPV 迎来了数款重量级车型，极氪 009 主打高端，意在取代丰田埃尔法乃至雷克萨斯 LM 的生态位，理想 MEGA 则希望利用 MPV 的空间优势重塑大家庭用车的产品定义，岚图梦想家成为这家新势力最依仗的产品成功和别克 GL8 分庭抗礼，小鹏 X9 想要以宜商宜家，是 MPV 也可以是 SUV 的灵活定位吸引更多受众……新旧势力在 MPV 市场的攻伐，发生在 2023 年，但现在已经是 2026 年了，零跑做一款旗舰 MPV 只需要把性价比拉到极致，但鸿蒙智行做一款旗舰 MPV 并卖四五十万元，需要做的就多了，当然我承认坐在二排的老板们对智界代言人刘亦菲肯定都是心驰神往，但除了刘亦菲这个核心竞争力之外，鸿蒙智行的首款 MPV 肯定需要更多击中老板坚硬内心和柔软脂肪的东西。这些东西是车载制氧系统、全维包裹安全气囊、华为悦彰非凡系列音响、二排主动座椅旋转、侧门交互式迎宾光毯等等等。最终，智界 V9 分为 Max、Max+、Ultra 和 Ultra+ 四个版本，售价分别为 39.98 万，42.98 万，45.98 万和 51.98 万起。舒适重要，安全重要，舒适且安全更重要！过去几年的时间里，鸿蒙智行成功地让国内汽车行业重视起来零重力座椅和主动安全来，是的，这两项鸿蒙智行都不是发明者，但确实在打开用户心智和开拓市场上起到了主要作用。前者代表座舱的舒适性，通过模拟太空失重环境下人体自然放松姿态的设计。通过特殊的人体工程学优化，使臀部、腰部、背部和腿部受力均匀，释放肌肉压力。后者代表汽车安全从被动防御走向主动避让，在汽车发生事故前，通过电子控制系统感知环境、预判风险并主动介入（如自动制动、转向控制），从而避免或减轻事故。但是，二者的矛盾在于在汽车行驶状态下，零重力座椅一般是没法使用的，这是因为当乘客躺在大仰角的座椅上享受时，身体与安全带、气囊的相对位置已经发生了相当大的变化。董车会之前报道过，根据 CIDAS 的数据，一旦座椅靠背夹角超过 115°，正面碰撞中的重伤风险会陡增 4.4 倍。此时，传统的 B 柱固定式安全带无法再有效约束骨盆，人体会因为惯性产生极度危险的「潜滑现象」——整个人滑向座椅下方，安全带可能会锁死在颈部导致窒息。采埃孚的试验甚至显示，在 50km/h 的碰撞下，躺倒状态受到的脊柱伤害堪比从 5 楼坠落。打过工的人都知道，当你向老板解释「零重力座椅虽然很舒服，但是不能在行驶状态下使用」时，老板这个时候一般不会听你的解释，只会跟你说「我不管，我想用，给我想一个解决方案出来」。今天，这个方案就被智界 V9 想出来了。智界 V9 全球首发了二排全维包裹安全气囊，这个气囊集成于二排座椅靠背，触发瞬间会形成一个独特的「U 型」环绕结构，在零重力、躺姿等多种姿态下可以全面防护头部、颈部、胸部，有效防止乘员在车辆碰撞之后身体滑落、碰撞与下潜，智界宣称历经百项严苛工况验证，无误伤风险。与之配套的还有二排坐垫前端也有专门布置的臀部气囊（BAB），在碰撞瞬间快速充气形成物理坡度，顶住盆骨不让身体滑脱，从而大幅降低腰椎压力。做了这么多工作，目的只有一个：就是让二排的老板在车辆行驶的时候好好躺一躺，好好休息一下。车载制氧：让高精力人群更有精力一直以来，对于什么样的人能做老板有个迷思，甚至还有「基因彩票」的说法，就是高精力人群更容易创业做老板。这些人可以每天工作十几个小时不知疲倦，只需要睡四五个小时就能满血复活。正如前面所说，很多老板之所以看起来可以每天工作十几个小时不知疲倦，很有可能是他可以中间随意选择休息时间。而屏幕前后的你和我，并不能在上班时间睡觉。更科学的说法是，高精力人群背后，跟摄氧量有正向关系，很多所谓的低精力人群很可能有鼻炎咽炎或者心肺功能低下导致日常处于轻微缺氧状态，导致精力不足。类似的情况也会发生在车上，很多人坐高铁感觉昏昏欲睡就是车厢里二氧化碳含量高，氧含量低，这种情况就容易困，密闭的汽车座舱也可能出现这种情况。智界 V9 给出的解决方案也很简单，装个制氧系统。这套制氧系统采用了 PSA（Pressure Swing Adsorption）分子筛变压吸附技术，提供了「弥散+鼻吸」两种模式，鼻吸出氧浓度 93±3%。额定制氧排量 8L/min，峰值 10L/min，比一般的家用制氧机大很多。2 个鼻吸口支持 1 拖 2，最多可供 4 人同时吸氧，弥散供氧与空调共用，四个出风口协同工作，30 分钟即可让整个座舱达到富氧状态。除了解决了传统制氧机体积大，噪音大的不足之外，鸿蒙智行的看家本领是智能化。这套制氧系统支持语音控制开启，中控屏实时显示状态；内置舒缓解压与高原出行两种模式，可自动调节供氧浓度，也可以联动小憩模式，帮助身心放松。很多人在森林景区，绿肺公园里感觉心思空明浑身舒爽，也是因为在富氧环境下，血氧饱和度一般会更高一些。富氧环境在特定场景下确实会有积极作用，比如早起后在通勤路上进入富氧环境有助于从昏沉状态中恢复；在运动后进入富氧环境可以加速乳酸代谢，缓解肌肉酸痛。更典型的场景就是针对本身心肺功能或者呼吸系统有问题的人，富氧环境有助于把缺氧状态调整为正常状态，还有就是高原出行环境，正常的氧气浓度可以有效缓解高原反应。在制氧系统之外，针对座舱健康还有鸿蒙 ALPS 健康座舱 3.0，通过把 MOFs 可吸附降解材料融入顶棚、门板、地毯等内饰，即便熄火也能持续吸附有害物净化醛类。UVC 空调主动杀菌，对流感等病毒灭杀率超 99.99%；前后双负氧离子净化器，可以释放超过西双版纳雨林 5 倍的负离子；这一整套净化杀菌和补氧系统，提供的是「健康」的价值，毕竟对于老板来说，紧急时刻生命第一，不紧急时刻健康第一。当然，需要着重说明的是，并不是所有人在所有场景都需要制氧系统制造一个富氧环境，如果本身不处于缺氧状态，血氧饱和度在 95% 以上，不需要运动恢复，没有昏沉嗜睡情况，不在高原地区，那就不太需要吸氧，长期过度摄氧反而有害健康。旋转座椅，优雅，还是优雅冰箱彩电沙发之所以从争议性的玩笑变成了兵家必争之地，还是市场表现的结果，谁不想二三十万的车就配有之前只有 78S 系才能选装的配置呢？这种造车逻辑之下，成本的分配也就必然从内燃机变速箱往舒适性配置上调整了。智界 V9 的沙发，显然值得多点笔墨。双独立结构，通风加热和比气动按摩更强劲的机械按摩已经不足为奇。更关键的是，二排座椅有可旋转设计，最大旋转角度 180 度，可以在多种角度进行旋转和固定，营造了例如 45° 迎宾座椅，90° 观景座和 180° 旋转之后的后排会客室布局等多种模式和选项。尤其是迎宾形态和会客室形态，可以在商务场景里营造比普通 MPV 更优雅的氛围。出门在外，面子可以是自己给的，但归根到底，还是别人给的，给别人面子就是给自己涨面子。智界 V9 类似的优雅配置还有不少，比如说「手势控车」，佩戴支持功能的手表，做出轻捏动作就可以隔空打开车门；二排乘客在车内做一个推一推手势，二排侧滑门就可以优雅闭合；窗边手指划一划，就可以调节智能隐私光幕比例的明暗度……车内的四块智能隐私光幕玻璃最大遮光率可达 99.6%，如果一键开启隐私模式，可以立刻断开车内的位置信息、麦克风、摄像头、网络，保证车内隐私。新车配有超 50㎡ 静音材料覆盖和 13 块双层夹胶隔音玻璃进行被动降噪，同时还可选 RNC 主动降噪。其他一些老板不太关心，但司机需要关心的作为一款鸿蒙智行的车，其实很多关于底盘，辅助驾驶和动力的信息都能猜个大差不差，这里完全可以用列表形式一一列举，毕竟这些信息是给司机看的，而且司机一看就懂，无需解释。新车长宽高为 5359×2009×1879 mm，轴距 3250 mm，2+2+3 座椅布局，后备箱容积 892L（可扩容至 2002 L）。全系标配 38 个传感器，包括 896 线激光雷达，背靠 60 EFLOPS 云端 AI 算力、WEWA 2.0 架构，可首批升级华为乾崑智驾 ADS 5。搭载 1.5T 雪鸮增程器配合前后双电机，综合功率 380kW，搭载 53.4kWh 电池，CLTC 纯电续航 305km，综合续航 1320km。新车有辉光紫、深海蓝、雪域白、鎏金黑四种车身配色和韶华杏、赤茶橘、绒霞紫三种内饰颜色。搭载 MPV 版图灵平台，前双叉臂与后 H 臂多连杆独立悬架，配合双腔空气悬架及 CDC 连续可变阻尼减振器。全系标配±7° 主动后轮转向，5.35 米转弯半径，支持低速蟹行模式，有效应对狭窄路况的调头和穿行。四活塞前固定卡钳，35.8m 制动距离；500mm 涉水深度，保证暴雨积水情况下的通行能力。路面预瞄系统智能识别井盖、凹坑与弯道，智能调节悬架阻尼、刚度，以及电机扭矩，可以降低冲击与侧倾，获得莱茵晕车舒缓认证。行业首发 2 根 2400MPa+6 根 2200MPa 侧门防撞梁；全系标配 13 个安全气囊，包括全维包裹安全气囊。前环宇三联屏，12.3 英寸液晶仪表与双 17.2 英寸 3.4K 大屏组合，可选后排 21.4 英寸吸顶屏。因为鸿蒙座舱，乾崑智驾，还有图灵底盘等鸿蒙智行基础产品技术支撑，现在鸿蒙智行的车型想要出挑有点困难，但毫无疑问，无论是华为还是奇瑞，都给到了智界 V9 足够的重视，比如单独给这款车开一场声势浩大的发布会，这在整个华为体系里是极为罕见的待遇。再比如，华为和奇瑞给这款车投入了一些深刻的产品洞察去迎合和吸引智界 V9 的潜在受众，很多人可能对这样一款华为气质非常浓郁的 MPV 产品无感，但华为不在乎，华为只希望这些人快快长大并事业有成。

发布会开场，李想先是强调：其实大部分用户并不知道自己想要什么。他分享了一个故事：有一次我去买烤冷面，摊主问我还要加点什么，我开玩笑说能不能打个折。你看，用户永远想要更便宜，但摊主会告诉你，其实还能加蛋。这也是 L9 当年销量爆发的秘诀——当行业还在比拼基础参数的时候，理想顺手给车里塞进了冰箱、彩电和大沙发，交出了一份远超预期的答卷。四年过去了，全行业都开始把屏幕做大、座椅做软。当堆砌配置成为车企的常规操作，理想 L9 又走向了另一条路。如今的理想把大量精力和资金砸向底层架构，自研范围覆盖车规级算力芯片、800V 主动悬架、全线控底盘等多个领域。承载这些研发成果的，自然是今天上市的两款新车型：L9 Ultra 和 L9 Livis。首先值得肯定的是，理想没有盲目放大 L9，车长和车宽依然维持在 5.25 米和 2 米，日常开起来依旧方便。在这个好开的尺寸基础上，L9 的整车姿态还挺拔了不少。在车高仅增加 10 毫米的情况下，前机盖和水切线被大幅拉高，配上 22 寸大轮毂，原本庞大的车身显得颇为矫健；前唇那一抹金灰色镀铬饰条也收束得干净利落。L9 Livis 还采用了迈巴赫同款涂装工艺，紫金、绿金、黑金等专属双色车身，额外增加两道工序后，漆膜厚度达到传统工艺的两到三倍，日光下呈现出流动的光影质感。外观的考究还藏在细节里。全车 12 颗超声波雷达被 UWB 传感器取代，保险杠上再也看不到突兀的小圆孔；隐藏式电动踏板收起后与底盘融为一体。这不仅让外观更整洁，UWB 技术还让哨兵模式耗电量骤降 82%，开两天只耗一度电。打开带有电容防夹功能的电动车门，里面还是那个熟悉的「起居室」，只是每一项体验的上限都被抬高了。全新 L9 是全球首款四个座位都配备全尺寸零重力座椅的 SUV。Livis 版本直接标配，Ultra 版本在首销期内也会免费赠送前排两把零重力座椅。在过去，转向管柱挤占了空间，主驾想完全躺平是个大难题。而在 L9 上，线控转向释放了空间，1.8 米的大个子坐进主驾也能一键舒展躺平。望向前方，仪表台上的两块屏幕合二为一，变成了一块 29 英寸的 6K 超宽全景屏。后排的 21 英寸屏幕新增电动平移功能，手轻轻一抬，这块「神奇移动屏」就会沿滑轨移到面前，并随座椅倾角自动找平。音响功率这次飙升到了 5440 瓦。9.3.6 星环剧院音响搭载三重矩阵式低音系统，头枕音响创新采用水平对置设计，出声口藏在两侧，以往那种「有人在你脑后说话」的突兀感也随之消失。车内空调支持 7 温区控制，面部和脚部可以吹出不同温度的风；全车 29 处皮质区域都带加热功能。中控台下方的 10L 智能冰箱换了电动门，轻拍两下即可开关，不仅支持 -6°C 冷冻，内胆还能直接拆下来清洗。作为增程车，全新 L9 两个版本的纯电体验都做到了行业顶配。底盘塞进了一块 72.7 度的 5C 大电池，容量已经超过不少纯电车型，换来 420 公里纯电续航。按每天通勤 40 公里算，两周充一次电完全够用。充电速度也同步提升，配合理想自建的 8300 根 5C 超充桩，十分钟左右就能完成补能——在服务区上个洗手间、买杯咖啡的功夫，电量就能满血复活。长途出行时，这台重达 2.8 吨的全尺寸 SUV 的油耗还挺实在的：亏电油耗降至 6.3 升/百公里，满油满电综合续航达到 1650 公里。更让人省心的是，这套增程系统的保养周期延长到了 3 年或 3 万公里，还省去了过去增程车繁琐的维护。李想表示，在机械属性和维保频次上，全新 L9 已经越来越像一台纯电车了。把家装进车里，把纯电融进增程「全尺寸 SUV 一直有一个很难真正被解决的矛盾，既让一家人坐得舒适，又让自己开得尽兴。」李想说。为了同时满足这两点，L9 Livis 直接上了全套线控主动底盘。L9 Livis 的 800V 主动悬架，让每个车轮都有超过 1 万牛的支撑力，传统的防倾杆被取消，四轮完全独立。过减速带、连续起伏路或者高速变线时，液压泵和减震器会实时配合，该撑就撑、该压就压。停车休息时，这套悬架甚至能联动赛车游戏，把车厢变成 4D 模拟器。大车想开得灵活，转向系统也要升级。Ultra 和 Livis 都标配了后轮转向 + 线控转向，转弯半径缩到 5.3 米，方向盘打满不用换手，角度也能按习惯自定义。另外，Livis 还首发了 EMB 线控机械制动，刹车响应快了约 0.1 秒，120km/h 时速下刹停距离能缩短 3 米以上。而且这套系统号称终生不用换刹车油，后期省心不少。安全性上，两款车都变得更加坚固了，堡垒安全车身 3.0 用上了 2200 兆帕的高强度钢和一体式双门环。最特别的是，车内的 11 个安全气囊还能通过红外传感器识别每一位乘员的体型与坐姿，动态调节保护策略。面对极端的碰撞危险，具备主动悬架的 L9 Livis 能在 0.7 秒内瞬间抬升 50 毫米，用最坚固的底盘部位去硬刚撞击；低速倒车时的防碰撞系统也具备了识别悬空物体的能力。智能化层面，理想首发了马赫 M100——一颗专为 AI 原生设计、采用数据流架构的端侧推理芯片。L9 Ultra 版本用的是单芯片方案，算力达到 1280TOPS，配备前向激光雷达；L9 Livis 则拥有两颗马赫 M100，总算力翻倍达到 2560TOPS，同时配备了 4 颗激光雷达来实现 360° 全向感知。算力上来后，感知帧率成倍提升，乡间小路、施工路段都能提前预判，果断绕行。这颗芯片还跟去年 12 月发布的 Livis AI 眼镜实现了生态互联。交互开始变得科幻。戴着眼镜站在车外，一个手势或一个眼神，车就会自动泊出来，缓缓开到你面前。坐进车里，系统还能察觉你的情绪和状态。你今天好像心情特别好，是要带家人去顺义的公园踏青吗？这时候，它既是司机，也是伙伴。到了我们最关心的价格，两款车只拉开了 5 万元的价差。L9 Ultra 定价为 45.98 万元，竞争力很强。72.7 度 5C 电池、双腔空悬、线控 + 后轮转向、1280 TOPS 马赫芯片都是标配，首销期还免费送前排零重力座椅、智能电开门和 5440 瓦音响。L9 Livis 定价来到 50.98 万元，比此前 55.98 万的预售价低了不少。多花 5 万，换来的是专属双色车身、800V 主动悬架、带 EMB 制动的全线控底盘、翻倍到 2560 TOPS 的算力，以及 4 颗激光雷达。首销权益也公布了：6 月 30 日前锁单，现金直减 2 万元，老用户再额外补贴 1 万元。从 2015 年创立至今，理想还是习惯把消费者叫作用户。买车只是陪伴的开始，让全家人在日复一日的相处里都能开心，才是 L9 Livis 这台「硅基生命体」想要兑现的承诺。

在今天和理想全新 L9 和智界 V9 两台重磅车型同台亮相的，还有蔚来的乐道 L80。相比 L90 面向三代同堂、多孩家庭的大三排需求，L80 把重点放在大五座市场，带来了更大的乘坐空间，更灵活的装载能力，以及更适合长途、自驾、露营等场景的空间组织方式。早年间，以本田冠道、大众途观 L 为代表的燃油大五座 SUV，主要通过释放第二排的腿部空间来提升乘坐舒适性；随后，随着新能源时代的到来，诸如理想 L7、问界 M7 等车型进一步强化了智能座舱的沉浸式体验和车内的储物能力。而乐道 L80 现在则试图在行业已有的基础上更进一步，将前备舱、后备舱、乘员舱以及补能体系整合为一个完整的生态闭环，通过蔚来的工程创新能力，让中大型五座 SUV 的日常使用体验再上一层楼。乐道 L80 一共有 Pro、Max+ 和 Ultra+三个版本，整车购买价格分别为 24.28万元、25.98 万元、27.98 万元， Baas 方案价格为 15.68 万元、17.38 万元和 19.38 万元。前备舱大，后备舱更大空间，依然是乐道 L80 最核心的竞争力。乐道 L80 车长 5145mm，轴距 3110mm，理想 L7 车长 5050mm，轴距 3005mm，乐道 L80 比理想 L7 大了一圈。取消第三排之后，乐道 L80 将所有空间都留给了前两排和储物。第二排作为 L80 整台车体验的核心区域，其后排头部空间达到 1074 毫米，腿部空间宽达 1121 毫米，坐垫宽度和长度分别达到 549 毫米和 497 毫米，靠背支持最大 135 度的后倾调节。L80 提出了一个「大五座，享四座」的越级体验概念，即当车内只有四人出行时，第二排的两位乘客完全可以获得媲美大型 MPV 独立航空座椅的放松状态。而在储物空间上，乐道 L80 配备了容积为 240 升的电动前备舱，能够轻松容纳两个 20 英寸的标准登机箱、一个儿童专属行李箱及一个随身背包。前备舱的存在不仅增加了绝对装载量，更为家庭用户提供了干湿分离与物理分区的储物区域，有效改善了传统 SUV 后备舱容易被杂物堆砌得杂乱无章的痛点。与前舱相对应，乐道 L80 的后备舱在第二排座椅处于常规直立状态时，容积为 1200 升；当第二排座椅完全放倒后，整体空间可扩展至 2600 升。此外，在后备舱的常规盖板下方，工程团队设计了双沉舱结构。其中外侧沉舱容积为 106 升，里侧沉舱容积为 63 升，均配备了独立盖板。这种隐蔽式的储物空间非常适合用来存放不常用的随车工具、零散的露营小配件、充电线缆以及各类清洁用品，使得主后备舱始终保持整洁。乐道能够在车辆内部极其有限的机械空间内挤出更多的储物容积，主要得益于车辆底层平台的极高集成度。L80 充分应用了已经在 L90 上被验证过的 900V 高压架构，并配备了体积更为紧凑的前置电驱动系统。结合 30.5 度大倾角冷却模块、全栈自研的热管理高度集成模块以及一体式偏置空调箱，这些系统级的技术创新对原本需要占据前舱大量体积的机械与电子部件进行了极致的压缩与整合。空间之外，座舱内的舒适化和智能化，也是乐道 L80 强调的特色。乐道 L80 的座舱内部采用了分层布局：上层主要用于视觉延伸与氛围营造，中层集中处理人机交互与触觉反馈，下层则侧重于日常储物与抗污耐用。车内配备了支持 1600 万色调节的氛围灯系统、长达 7.2 米的环绕光带，辅以大面积软质材料包覆、金属扬声器网罩及精细的绗缝与刺绣工艺。整体风格规避了盲目堆砌昂贵材料的浮夸感，更侧重于在家庭用车的基调上打磨出温馨且不失格调的精致质感。为了提升车内的通透感，乐道 L80 车辆侧窗与前挡风玻璃的总面积约为 3.3 平方米，前后分区的全景天幕面积也达到了 1.1 平方米。同时为了应对大面积玻璃材质在强光照射下带来的隔热与防晒问题，车内配备了双电动物理遮阳帘，乐道 L80 同时配备了空气悬架系统，支持在驻车时自动降低车身高度，配合电动侧踏板与静音电吸门，优化了行动不便的老人或儿童的上下车体验。车辆的座椅总成基于专属的高端座椅平台开发，依据中国人体生理曲线大数据进行了型面贴合度优化，前排座椅支持 12 向电动调节，第二排支持大角度电动调节。同时，方向盘加热、前后排座椅加热、通风以及按摩功能均作为标配提供。此外，车内集成了一台 8.8 升容量的智能电动冰箱，足以容纳 6 罐常规可乐、2 盒鲜牛奶以及部分水果零食。神玑上车，4D 领航数字化交互与计算平台是现代智能汽车的核心大脑。L80 的屏幕矩阵由一块 35 英寸的 AR-HUD 增强现实抬头显示、一块 17.2 英寸的 3K 高清中控屏、一块 17.3 英寸的后排吸顶娱乐屏以及一块 8 英寸的后排专属场景控制屏组成。支撑这一系列复杂交互与协同调度的硬件，是一颗采用 5 纳米制程工艺的车规级高阶智驾芯片——神玑 NX9031。该芯片拥有高达 546 GB/s 的内存带宽，打破了传统分散式域控制器的物理壁垒，能够同时支持智能驾驶、智能座舱、智能底盘等多个系统跨域计算资源的统一调度。配合 SkyOS·天枢整车全域操作系统，车辆从底层代码级别打通了智驾、座舱、底盘、车身及动力网络，使各项功能摆脱了信息孤岛的状态，能够基于场景进行深度系统级协同。在智能驾驶领域，系统深度接入了蔚来世界模型 NWM（NIO World Model）。该模型摒弃了传统的规则驱动逻辑，转而通过强化学习与海量的闭环仿真训练，提升了系统对复杂未知路况的理解与博弈能力，全面覆盖城区复杂道路、城市快速路、高速公路及各类园区停车场。值得重点关注的是，L80 支持城区领航换电功能（PSP）。在导航状态下，系统能够自动规划最优补能路径，将车辆自主驾驶至换电站并完成自动化换电流程。同时，基于模型化架构构建的智能泊车系统，也支持路线的实时动态规划和对各类非标准复杂车位的识别入位。尽管定位为体型庞大的家庭 SUV，L80 在底盘的机械素质与动态调校上依然投入了极高的规格。悬架系统采用了前双叉臂与后五连杆独立悬架组合，配备了具备 100 毫米行程调节能力的智能空气悬架系统以及 CDC 智能连续可变阻尼减震器。底盘硬件与路面之间通过液压衬套进行柔性连接，并搭配了抓地力更强的鸳鸯胎。乐道 L80 还支持 4D 全域舒适领航功能。该功能能够实时结合车辆自身的传感器感知、激光雷达对前方路面的三维预瞄以及云端高精舒适图层数据，在遇到减速带、坑洼或颠簸路面之前，提前毫秒级主动调节空气弹簧的刚度与减震器的阻尼，从容化解路面冲击。除了常规的运动、舒适、后排舒享和节能模式外，车辆预设了湿地、雪地、泥地、沙地等脱困模式。在卓越的底盘参数加持下，L80 拥有高达 600 毫米的最大涉水深度，最小离地间隙最高可拉升至 260 毫米，足以轻松应对绝大多数轻越野和长途自驾游的复杂路况。而在安全方面，乐道 L80 全系标配 9 个安全气囊，除了常规的环绕式防护，第一排还配备了能在碰撞发生前预紧的主动式安全带、容积达到 137 升的狐耳式超大副驾气囊以及防止前排乘员相互碰撞的中央远端安全气囊。基于家庭属性的考量，第二排乘客的安全防护被提升至与前排同等重要的层级，配备了 14 升容积的后排侧气囊，以及长度贯穿 2920 毫米、容积高达 56 升的超大侧面安全气帘。在应对碰撞的白车身结构设计上，L80 的前部预留了长达 650 毫米的充足溃缩吸能空间，并创新性地采用了「6 纵 10 横」的笼式安全传导结构。前防撞梁选用了高强度的 7 系航空级铝合金材质，其上层主防撞梁的宽度覆盖率达到了整车宽度的 80%，能够在严苛的偏置碰撞中有效化解破坏性冲击力。侧面防护则通过环舱高强度钢、一体式门环、多腔体铝合金门槛梁以及坚固的电池舱横梁共同构建。车尾部分应用了先进的一体压铸后地板技术，配合铝合金后吸能盒与后防撞梁，为车身尾部提供了深达 1244 毫米的吸能缓冲空间，最大程度减轻高速追尾事故对后排乘客及底层电池包的结构性伤害。乐道 L80 搭载的电机能够输出 340kW 的峰值功率，百公里加速时间 5.7 秒，其百公里综合电耗被控制在优秀的 14.3kWh，85 度电池下的 CLTC 续航为 615km。当然，纯电车型的体验永远无法脱离补能体系的建设。共享蔚来庞大且成熟的充换电网络，依然是乐道 L80 在同级别竞品中最难以被复制的核心优势。为了进一步夯实这一护城河，乐道官方已确认追加投入 6 亿元人民币，用于在网络中新增 8000 多块流通电池。蔚来之前透露，早在四年前，乐道 L80 便与 L90 基于同一技术平台同步启动开发。其中，乐道 L90 的成功已经验证了产品定义的精准性。在 2025 年，乐道 L90 凭借对大三排 SUV「满载状态下行李装载空间」这一核心痛点的有效解决，在上市短短 5 个月内实现交付超 4.3 万台，成为纯电大型 SUV 市场的年度销量冠军。这不仅助力乐道品牌在 2025 年实现 107808 台的全年交付目标，更以接近 24 万元的成交均价稳居新势力品牌第一梯队。乐道 L90 和蔚来 ES8 的表现，也被行业视为大型 SUV 动力形式从增程向纯电转型的关键标志。如果说乐道 L90 成功开拓了大三排纯电 SUV 市场，那么乐道 L80 则肩负着在规模更大的大五座 SUV 领域实现突破的重任。在乐道的产品价值体系中，L90 聚焦于「在一起」的温情与责任，通过三排布局满足多孩、三代同堂家庭的聚会与通勤需求；而 L80 则更侧重于「走出去」的激情与成长，以更灵活的空间形态承载用户自驾、露营等更具探索性的生活方式。

🔍郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%🤖OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略🚀Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI📱微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆💻华为 Mate X7 官方降价 1000 元⚠️追觅造车主体约 232 万元股权遭冻结🎮国行 Switch 正式停服🧠百度成立模型委员会 BMC💰三星电子单季营收创历史纪录，半导体撑起超九成营业利润🔍海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%🖨️铠侠：2027 年内存仍吃紧、净利润同比增长 103.6%🤖微信发布「全球青少年 AI+ 小程序」洞察报告💡何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了🦿5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约📱荣耀 CEO 李健官宣机器人手机今年 Q3 发布📱vivo S60 将首发 4K 原生感 Live🚗全新理想 L9 发布：45.98 万元起🤖ChatGPT 新增 Codex 连接电脑功能📰 周末也值得一看的新闻郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%天风国际证券分析师郭明錤昨日披露，苹果已在 Intel 18A-P 系列制程上为低阶及旧款 iPhone、iPad 与 Mac 处理器立项，订单中 iPhone 约占 80%。流片规划显示，苹果预计 2027 年在 Intel 放量，2028 年续增，2029 年进入衰退期。Intel 方面，2027 年良率目标为稳定达到 50% 至 60% 以上，量产时程与出货规模目前仍未明朗。不过，郭明錤也指出，即便 Intel 顺利出货，台积电在苹果供应体系中仍将占 90% 以上比重。郭明錤认为，苹果此举是有意将 Intel 培养为长期备选供应商。AI 对台积电先进制程资源的占用将持续扩大，苹果需要在议价能力尚存时提前布局。与此同时，苹果也在同步评估 Intel 其他先进制程技术。OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略据 WIRED 报道，OpenAI 昨日向员工宣布新一轮组织调整，联合创始人兼总裁 Greg Brockman 正式接管产品战略。报道指出，OpenAI 将 ChatGPT、编程智能体 Codex 及面向开发者的 API 业务整合进同一核心产品团队，目标是将三者合并为「统一的体验」。Codex 原负责人 Thibault Sottiaux 升任核心产品与平台主管，统管消费、企业和开发者三条线；ChatGPT 原负责人 Nick Turley 转任企业产品改版；前 Instagram 副总裁 Ashley Alexander 接任消费产品负责人。Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI据《金融时报》报道，Anthropic 已就一轮 300 亿美元的融资达成条款，融资完成后公司估值将约达 9000 亿美元，较今年 2 月初 3500 亿美元的估值接近翻三倍。本轮融资由 Dragoneer Investment Group、Greenoaks Capital、红杉资本（Sequoia Capital）和 Altimeter Capital 联合领投，报道称，每家领投机构预计出资不低于 20 亿美元。微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆CCF 计算机博物馆（CCM）昨日收到腾讯微信团队捐赠的一批早期历史实物，包括微信第一行后台代码原始存档、早期开发设备 iMac ME086CH/21 寸一体机与戴尔塔式工作站，以及 Apple iPod Touch 3、诺基亚 5800XM 等早期测试终端。CCM 表示，由于移动产品迭代速度极快、数字资料易被覆盖删除，大量见证移动互联网起步的实物未能得到系统保护。此次捐赠呈现了微信从 0 到 1 的研发链条，是中国移动互联网早期发展的核心实物见证。CCM 主体建筑已全面落成，拟于今年 9 月开馆，目前持续面向社会征集各时期计算机相关藏品。保时捷称被中国车企抄袭「有些许无奈」，罗永浩：保时捷自己就有前科保时捷中国传媒公关副总裁唐风靓日前在浙江「易境思论坛」上表示，过去几年保时捷激发了高度统一的汽车设计审美，被中国车企模仿「除了荣幸也有些许无奈」。罗永浩随后在微博发文回应，称保时捷说「无奈」可以理解，但称中国车企抄袭得「高度统一」则是吹牛。他同时指出，保时捷历史上也曾抄袭捷克斯洛伐克品牌太脱拉（Tatra）的设计，并以 1965 年大众汽车向太脱拉支付 100 万西德马克庭外和解赔偿金为据，直言「说到流氓程度，谁能跟保时捷比」。此番言论随即引发讨论。有网友指其犯了「诉诸伪善」的逻辑错误，罗永浩强硬回怼，称自己已明确承认抄袭者「也是流氓」，并非为抄袭行为开脱。华为 Mate X7 官方降价 1000 元华为昨日宣布，旗下两款折叠屏手机同步开启降价优惠：Mate X7 全系直降 1000 元，优惠后售价 11999 元起；Mate X6 直降 3000 元，优惠后售价 9999 元起。追觅造车主体约 232 万元股权遭冻结企查查 App 显示，追觅科技创始人俞浩实控的星空计划（上海）汽车科技有限责任公司近日新增一则股权冻结信息，股权数额约为 232.25 万元，股权所在企业为上海星空计划进出口贸易有限责任公司。去年 8 月，追觅科技宣布进入汽车领域。截至目前，追觅汽车团队规模约千人，研发人员占比达 70%，自研覆盖智能汽车多个核心领域。此次造车主体出现股权冻结，官方尚未就相关情况作出回应。国行 Switch 正式停服国行 Switch 的 Nintendo e 商店已于昨晚正式停服，至此，国行 Switch 将无法再进行联网对战或下载游戏。官方在停服前向相关玩家发送提醒，建议玩家在停服前完成所有已购游戏、DLC 及兑换码的下载与激活；删除游戏数据或对主机进行格式化后，数字内容将无法恢复。今年 3 月 31 日，国行 Switch 的游戏售卖服务已率先停止；而根据官方此前公告，今年 12 月 31 日起，国行 Switch 主机及配件的维修服务也将全面终止。这意味着，国行 Switch 的官方售后支持将在今年底彻底画上句号。百度成立模型委员会 BMC据《科创板日报》报道，百度近日内部设立百度模型委员会（Baidu Model Committee，BMC），成员由对大模型有深刻理解的年轻研究员构成。基础模型研发部（BMU）与应用模型研发部（AMU）均向该委员会汇报。知情人士透露，新成立的 BMC 将全面统筹百度大模型业务，覆盖路径探索、技术规划、模型训练到业务落地的完整链条，目标是实现模型从技术到应用的一体化管理。百度方面表示，此举有助于集中优势资源，持续建设并强化人工智能方向的技术优势。三星电子单季营收创历史纪录，半导体撑起超九成营业利润三星电子昨日发布 2026 财年第一季度（1 月 1 日至 3 月 31 日）业绩报告：期内实现合并营收 133.9 万亿韩元，同比增长 69%，环比增长 43%；营业利润 57.2 万亿韩元，同比增长 756%，环比增长 185%，营业利润率达 42.8%。净利润为 47.2 万亿韩元，净利率为 35.3%。基本每股收益为 7123 韩元。半导体（DS）：营收 81.7 万亿韩元，同比增长 225%，环比增长 86%；营业利润 53.7 万亿韩元。其中存储业务营收 74.8 万亿韩元，同比增长 292%，环比增长 101%，为本季度核心增长引擎；设备体验（DX，含手机与消费电子）：营收 52.7 万亿韩元，同比增长 2%；营业利润 3.0 万亿韩元，同比下降约 1.8 万亿韩元；显示（SDC）：营收 6.7 万亿韩元，同比增长 14%；营业利润 0.4 万亿韩元，同比下降 0.1 万亿韩元；Harman：营收 3.8 万亿韩元，同比增长 12%；营业利润 0.2 万亿韩元。下季度展望，公司预计 AI 基础设施扩张驱动的服务器内存及 SSD 需求将持续强劲，计划向客户交付 HBM4E 首批样品以巩固技术领先地位；手机业务方面，Galaxy S26 系列发布效应趋于消退，盈利能力预计承压。海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%昨天，海力士于发布今年第一季度（截至 3 月 31 日）业绩，按 K-IFRS 口径，期内实现营收 52.58 万亿韩元，同比增长 198%；营业利润 37.61 万亿韩元，同比增长 405%；净利润 40.35 万亿韩元，同比增长 398%。本季度营收首次单季突破 50 万亿韩元，创历史新高。主要财务指标如下：毛利润 41.68 万亿韩元，毛利率 79%，同比提升 22 个百分点；营业利润率 72%，创历史单季新高，环比提升 14 个百分点；EBITDA 41.34 万亿韩元，EBITDA 利润率 79%，同比增长 284%；净利润率 77%，其中非经营利润 14.01 万亿韩元，包含投资资产估值收益 9.94 万亿韩元。铠侠：2027 年内存仍吃紧、净利润净同比增长 103.6%铠侠控股昨日发布截至今年 3 月 31 日止 2026 财年业绩，期内实现营收 2.3376 万亿日元，同比增长 37.0%；按 IFRS 口径，归母净利润为 5544.9 亿日元，同比增长 103.6%。SSD 与存储业务营收 1.3626 万亿日元，同比增加 37.5% 至 3715 亿日元；智能设备业务营收 7599.8 亿日元，同比增加 51.6% 至 2588 亿日元；其他业务（含 SD 卡、U 盘等零售产品及向闪迪集团的销售）营收 2150.1 亿日元，同比基本持平。另据财联社报道，铠侠表示，AI 服务器客户的订单量持续走强，以美元计价的 NAND 价格在一季度上涨逾一倍，并预计 2027 年度 NAND 存储供应仍将维持紧张态势。微信发布「全球青少年 AI+ 小程序」洞察报告昨天，腾讯研究院联合微信发布《全球青少年 AI+小程序洞察报告》，披露了全球基础教育阶段群体利用 AI 工具与小程序进行应用开发的各项核心数据。据报告数据显示，自 2023 年 11 月微信小程序教育平台正式对外开放以来，已吸引全球 7954 所学校的逾 8.7 万名学生参与，累计创建小程序项目超 28.7 万个。在开发流程中，师生调用 AI 辅助创作每年消耗 Token 量超过 500 亿，支撑约 375 万次人机交互。此外，微信连续四年举办的小程序全球创新挑战赛共产生 7430 个作品，其中第四届单届提交量达 4900 个，呈显著爆发趋势。开发者画像层面，62% 的项目由单人独立完成，最小获奖开发者年仅 5 岁。地域活跃度数据显示，中国香港、北京及中国澳门位列前三，且项目协作已出现跨国别开发特征。在应用方向上，05 后开发者主要聚焦于无障碍辅助、碳足迹、适老化关怀及健康管理等社会实用场景。💡 何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了据睿见 Economy 报道，小鹏集团董事长、CEO 何小鹏昨日在「第十八届轩辕汽车蓝皮书论坛」上表示，「什么时候我们看到北京、上海车展上不再有 150 款新车的时候，中国汽车的内卷可能就结束了。」何小鹏表示，过去十年，中国新能源汽车经历了两个重要转变：一是核心驱动能源从传统化石能源切换至电化学能源；二是行业全面进入初步普及的智能化阶段。他称，十年前几乎没有人相信中国汽车能在全球市场取得如今的成绩，并实现规模化的全球化布局。何小鹏认为，下一个十年将是电化学能源与数据能源的双重叠加阶段，芯片和模型将成为新的「引擎」，汽车与机器人的能力将在这十年内深度整合。5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约ROG 与 XREAL 联名推出的电竞 AR 眼

🤖曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini🌐马耳他将向全体公民免费提供一年 ChatGPT Plus📱小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务🚀目标估值 1.75 万亿美元，SpaceX 最快 6 月上市📋arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚💻卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元🚗蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌🔍需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位☁️最低 1 元 40 万 Token，三大运营商推出 Token 套餐🚙特斯拉 Robotaxi 碰撞事故报告解封💡知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值🚗小米 YU7 GT 火山灰配色公布💰ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构🧠马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布🤖字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本📱腾讯 vibe coding 产品「吐司」即将上线🛍️蜜雪冰城上架老黄同款蜜桃四季春，销量暴涨🛑中国银行关停信用卡独立 App，服务整合进主 App曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini昨晚，彭博社记者 Mark Gurman 在最新的 Power On 新闻通讯中透露了更多关于苹果即将发布的独立 Siri 应用的细节。隐私设计是此次新应用的核心卖点之一。Gurman 透露，该应用将内置对话自动删除功能，用户可在设置中选择对话保留时长：30 天、1 年或永久保存。苹果预计在今年的 WWDC 继续主打隐私保护，并强调其 AI 体验不含广告。模型方面，苹果已将新 Siri 的 AI 模型底层大量替换为 Google 的 Gemini，并将部分服务运行在 Google 的云基础设施之上。而此前有消息称，苹果和原本的合作对象 OpenAI 关系出现裂痕，甚至有可能对簿公堂，原因是 OpenAI 认为双方合作的结果不及预期。值得注意的是，尽管新的 AI Siri 已经跳票约两年，Gurman 称苹果内部测试版本仍将其标注为「Beta」，并可能沿用至其正式发布。马耳他将向全体公民免费提供一年 ChatGPT PlusOpenAI 日前宣布，与马耳他政府签署合作协议，向该国公民及居民免费提供为期一年的 ChatGPT Plus 订阅服务。据悉，居民须先完成由马耳他大学开发的免费在线课程「AI for All」，课程内容涵盖 AI 的基本概念、能力边界及负责任使用方式，通过后即可向马耳他数字创新局申请免费订阅资格。小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务小米 REDMI 产品经理「笋寸」日前在微博发文建议，近期不打算换手机的用户可以考虑等待换电池，并提示部分机型在更换新电池后容量可能还会有所提升，参考小米 13 系列的现行政策。笋寸进一步说明，小米正积极推动存量用户更换采用金沙江技术的大容量电池，小米 13 系列已率先上线相关服务。她还透露，具体支持机型「得等等」。此前，小米 13 系列已全系上线电池升级服务，支持小米 13、小米 13 Pro、小米 13 Ultra 三款机型升级更大容量电池，费用为 189 元，其中电池升级费 149 元、人工费 40 元。目标估值 1.75 万亿美元，SpaceX 最快 6 月上市据路透社报道，SpaceX 已加快 IPO 时间表，计划最早于 6 月 12 日在纳斯达克挂牌上市，股票代码为「SPCX」，较此前预定的 6 月下旬提前约两周。据三位知情人士透露，公司计划最早于本周三公开招股说明书，路演定于 6 月 4 日启动，股票发行最早于 6 月 11 日完成定价。时间表提前的主要原因是 SEC 对招股材料的审核进展快于预期。此次 IPO 预计募资约 750 亿美元，目标估值约 1.75 万亿美元，有望成为史上规模最大的股票发行。承销行阵容包括摩根士丹利、美国银行、花旗集团、摩根大通和高盛。arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚近日，预印本学术论文平台 arXiv 宣布，若作者提交的论文中包含无可辩驳的 AI 生成内容证据，将被封禁账号一年，并在此后要求所有提交须先经过同行评审期刊接受方可上传。arXiv 计算机科学板块主席 Thomas Dietterich 在 X 发文表示，若论文中出现「幻觉引用」、大语言模型元注释（如「以下是 200 字摘要，是否需要修改？」）等明显迹象，即可认定作者未对 AI 生成内容进行核查。arXiv 目前由康奈尔科技学院管理，今年 7 月将独立运营，成为一家非营利法人机构。康奈尔科技学院院长 Greg Morrisett 表示，独立后 arXiv 将能从更广泛的捐助方获得资金支持，以应对「AI 垃圾内容」带来的运营压力。卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元近日，小米集团合伙人、总裁卢伟冰在「Max 一夏」户外露营直播中表示，玄戒芯片今年肯定会推出迭代款，并称这将是「一块非常强的芯片」，会搭载在一款「非常优秀的产品」上。他同时提醒外界，目前网络上流传的各类相关传闻可信度不高，具体规格暂不便披露。卢伟冰还透露新一代玄戒芯片的应用范围将不再局限于智能手机，后续计划广泛搭载于小米旗下多类智能终端设备。关于手机定价，卢伟冰在直播中作出判断，认为今年下半年尤其到年底，部分国产旗舰直板手机价格可能突破 1 万元人民币。他预计，内存涨价的压力至少延续至 2027 年底，甚至到 2028 年。对于即将发布的小米 17 Max，他表示定价仍在讨论中，将在发布会上正式公布，但小米的原则是在同档位提供「最物超所值的产品」。蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌据睿见 Economy 报道，在日前举行的第十八届轩辕汽车蓝皮书论坛上，蔚来创始人、董事长兼 CEO 李斌线上出席，在对话中将萤火虫比作 iPhone，称其已成为「一款非常高品质的产品」。他强调，萤火虫在某种意义上超越了汽车的工具属性，演变为「一个生活方式品牌、一个时尚品牌」，并开始形成独特的品牌文化。他表示，萤火虫在用户群体上打破了年龄与购车预算的双重限制，「大学刚毕业的年轻人想买也能买，很有钱的老板想自己开车送小孩上学也完全没有问题」。需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位据《商业内幕》报道，前向部署工程师（FDE）的招聘需求在一年内激增逾 700%，正成为 AI 浪潮下科技行业增长最快的职位之一。招聘平台 Indeed 数据显示，今年 4 月该职位招聘帖数量达 5330 条，较去年同期的 643 条增长约 729%。Anthropic、OpenAI、Palantir、Stripe 及 Google Cloud 均在积极招募该职位人才。FDE 的核心职责是协助企业将 AI 整合进业务流程，并根据客户需求定制开发 AI 工具。薪酬方面，Indeed 数据显示该职位年薪约为 17 万至逾 20 万美元。最低 1 元 40 万 Token，三大运营商推出 Token 套餐中国电信昨日推出系列试商用 Token 套餐，分别面向开发者及中小微企业、个人及家庭用户提供「Token+连接+安全」一体化服务，均提供三档套餐，面向个人及家庭用户定价 9.9 元/月起 1000 万 Token；面向开发者及中小微企业 39.9 元/月，1500 万 Token。两类用户均可选购宽带上行提速包（上行速率提升 50 Mbps，100 元/月）及智能体安全管家服务（5 元/月）。个人家庭用户还可额外选购 5G-A 优享包（含 10 GB 流量，上行速率最高 300 Mbps，20 元/月）。联通、移动两大运营商也先后推出了各自的 Token 套餐：移动：中国移动上海公司于 5 月 17 日世界电信日发布会上推出 Token 通用服务，定价 1 元 40 万 Token。联通：中国联通上海分公司于 5 月 16 日发布全栈国产化「国芯国模国云」安全基座，推出多档 Token 产品及融合套餐。特斯拉 Robotaxi 碰撞事故报告解封近期，特斯拉正式解密并公开了向美国国家公路交通安全管理局（NHTSA）提交的 17 起「Robotaxi」自动驾驶碰撞事故的完整未脱敏报告。据 Electrek 报道，这 17 起事故发生于 2025 年 7 月至 2026 年 3 月，均涉及在得克萨斯州奥斯汀进行「Robotaxi」测试的 2026 款 Model Y，且测试时均有安全员在场。伤亡数据显示事故整体较轻微：13 起仅造成财产损失，2 起无人员受伤，1 起轻伤未入院，另有 1 起轻伤需入院治疗（系车辆在路口以 2 英里/小时蠕行时被后方 SUV 追尾）。官方报告表明，大部分事故并非特斯拉自动驾驶系统（ADS）的过失，多为车辆在红绿灯、停止标志或交通拥堵中完全静止时，被后方人类驾驶员追尾或侧擦。但数据同时也暴露出系统与后台干预的缺陷：两起事故发生在远程操作员接管期间，导致车辆分别以 8 英里和 9 英里的时速撞上金属围栏与施工路障；另有数起事故显示 ADS 存在空间感知死角，车辆在转弯或倒车时撞上了金属链条、拖车挂钩、木制电线杆及路缘石。💡 知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值据「听筒Tech」报道，知乎创始人兼 CEO 周源日前在第十二届新知青年大会上表示，AI 时代真实创作者的价值不可替代。AI 每分钟都在创造奇观，但 AI 时代仍然需要新知。而新知来自真实的人。每一个真实的人，都拥有不可替代的价值。大会主论坛上，多位嘉宾就 AI 与人类创作的关系展开讨论。复旦大学人文学者梁永安提出「科技有所为，如何让自己有所不为？需要人的尺度和人文属性」；作家张春称「不完美就是我们的完美」。周源在大会结尾总结道：「技术越发达，有温度、有深度的真人创作就越宝贵。」小米 YU7 GT 火山灰配色公布近期，小米汽车再度放出 YU7 GT 的预热图片，公布了新配色以及新车的内饰信息。其中小米 YU7 系列将加推一款名为「火山灰」的配色，官方介绍：灵感来自晨雾下的火山地貌，高纯度灰调，叠加纳米级矿物颗粒，呈现细腻而有层次的金属质感。同时新车内饰也迎来全新的「赛道红」：采用红黑撞色，提供专属刺绣。官方表示，小米 YU7 也会提供「火山灰」外饰配色，并且现在车厘子红的 YU7 GT 已经在部分小米汽车门店展示了。ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构近日，OpenAI 面向美国 ChatGPT Pro 用户推出了个人财务管理功能，支持通过 Plaid 连接超过 1.2 万家金融机构的银行账户和信用卡。连接账户后，ChatGPT 会自动读取用户的真实余额、消费流水和负债数据，生成一张覆盖投资、日常开销、订阅和近期账单的财务面板。用户提问时，模型可直接翻出账单历史，找出哪些类目容易超支，并给出具体的削减方案。此外，用户还可以告诉 ChatGPT 一些账外信息，比如「我在攒钱买车」或「我还欠父母一笔钱」，系统会将这些内容保存为财务记忆，用于后续对话。马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布马斯克昨日在 X 发帖披露，SpaceXAI 正在研发的 1.5T 参数 Grok 基础模型第 9 版（V9）已完成训练，预计 3 到 4 周内正式发布。现役公开版本 Grok 4.3 基于 0.5T 参数的 V8 基础模型，目前每隔数日迭代一次。V9 将参数量提升至 1.5T，马斯克称其为「重大升级」。在帖子的结尾，马斯克更是用用「banger」（王炸）一词来形容对这一版本的预期。字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本字节跳动 Seed 团队近日开源扩散

直到 2026 年五月，我们终于等到一整轮的影像旗舰登场完毕——小米做出了徕卡一瞬、vivo 打造了照片与视频兼具的 V 单、OPPO 在双长焦方案上带来了惊喜。在这轮旗舰更新潮的最后，华为提前将新一轮的 Pura 90 Pro Max 端上了桌。老实说，第一眼看硬件，心里是有落差的。去年 Pura 80 Ultra 上颇为惊艳的「一镜双目」结构不见了，取而代之的，是目前行业内最稳妥也最常见的大底长焦方案。不过，在细细体验后，我认为 Pura 90 Pro Max 藏了一个比纯拼硬件更有意思的新解法——XMAGE 智拍。全自动，都省心过去这几年，我系统性地梳理过华米 OV 的影像架构。虽然各家在硬件选择上各有理解，但大方向出奇一致，都在极其有限的机身空间里死磕传感器面积，比拼谁的光圈更大、进光量更多。在硬件上奋发图强理所应当。但在摄影中，不错的硬件基础仅仅是第一步。想要拍出一张好照片，构图与后期同样重要——构图保证一张照片在视觉上是否平衡、主体是否明确；而后期则决定了这张照片视觉质感和风格化。但这两者，向来都是具备门槛的手艺。而华为在 Pura 90 Pro Max 上，打算一口气解决这两个问题。熟悉他们产品线的朋友或许有印象，去年发布的 Pura 80 Ultra 上，华为试水了 AI 辅助构图。AI 辅助构图功能相对基础，在实际拍摄中主要起一个引导和提醒的作用。到了 Pura 90 Pro Max，这项技术完成了从「辅助」到「主导」的蜕变，并定名为「XMAGE 智拍」。剥开表层去探究 XMAGE 智拍的底层逻辑，是一套完全由端侧算力驱动的本地化影像流水线，将拍摄过程拆解为三个同时进行的模块——基于实时语义分割的自动构图、无视物理镜头限制的智能变焦，以及针对画面的色彩重塑。也就是说，在这个功能的加持下，Pura 90 Pro Max 可以自动识别主体、变焦构图，最后还排列出一系列适合当前场景的 XMAGE 风格供你选择。值得一提的是，XMAGE 智拍不需要网络连接，也无需将庞大的图像数据传回云端，全部基于端侧 AI 运转，实时识别画面里的主体轮廓、建筑线条以及光源分布状况，随后根据底层训练好的审美模型，给出一套包含最优构图、变焦和精细色彩方案的组合拳。繁复的专业参数在这里黑盒化，过去需要人脑去判断的构图和调色工作，全权交给了本地算法。除此之外，Pura 90 Pro Max 还顺势推出了 AI 姿势推荐，这个功能同样通过语义分割识别，可以理解画面中的了恩物、环境、姿势与背景，随后直接在取景框中勾勒出一个等比例的线框轮廓。拿手机的人只需按图索骥，引导模特贴合画面上的姿势即可。遇到不满意的动作能随时刷新，它甚至支持导入社交网络上保存的样片，一键提取动作精髓。听起来确实很厉害，但这套完全依赖端侧计算跑出来的流水线，究竟能不能应对现实世界里复杂多变的光线，还需要把它带到真正的街头去寻找答案。AI 有所能，也有所不能来到一家咖啡馆，墙上的多层光环灯饰引起了我的注意。举起手机，XMAGE 智拍自动变焦到 2×，将墙壁上的光环作为画面主体。这个选择可以说中规中矩，光环圆形处于画面中下方，整体构图稳定和谐。同时，操作台上颇具金属质感的咖啡机也是个不错的主体，XMAGE 智拍精准识别到我的拍摄意图，变焦到 192mm，保持咖啡机的中置。由于我想保留更为还原的室内光线与色彩，这张照片并没有套用 XMAGE 智拍推荐给我的任何色彩风格。用它拍了几组照片后，我认为 XMAGE 智拍是一个非常用户视角的功能——在传统的拍摄习惯里，手机焦段是一件需要精打细算的事。大脑会本能地去贴合物理镜头的原生焦段，生怕落在中间焦段会损失画质。但 XMAGE 智拍完全抛弃了执念。它不在乎用的是几倍变焦，转而完全为画面考量：当前这个画面，怎么裁切才最好看。哪怕不可避免地要损失一部分边缘像素，只要能让视觉中心突出、比例和谐，它就会果断动手。这种做法很实在，一切都在为了最终的出片服务。走出室外回望咖啡馆，绿植与橙色外墙交相呼应，XMAGE 智拍自动截取出色彩冲击力最强的部分，同时准确识别出咖啡馆建筑主体，两种大色块在画幅中占比相似，平衡而美观。再抬头，夏季树木繁茂，XMAGE 轻微放大焦段，让树枝形成天然的框架构图，广州塔正悬于画面正中间。在这组照片的拍摄中，我还察觉到了一个很有意思的技术细节。按下快门时，你会发现滤镜加载的速度，远大于构图改变的速度。这是因为滤镜往往只需识别画面后套用底层的色彩映射，而构图则需要端侧算力去实时识别画面里的建筑线条、光源分布和人物轮廓。计算量完全不在一个量级。更有意思的是，智拍套用滤镜的方式非常精细，除了 XMAGE 风格选择外，还会对调色盘做精细调整，确保风格对当前环境的适配度。当然，这套系统也有自己的脾气。在实际走街串巷时，XMAGE 智拍有时候会自动跳出来接管画面，有时候又毫无动静，需要你手动去点击唤醒。就算是手动唤醒的情况下，XMAGE 智拍也有几率出现风格已经选好，但当前画面无法找到最佳构图的情况。这种情况频繁出现在极其繁杂的环境中，由于元素过于混乱、信息量爆炸，算法的确很难从中剥离出合适的构图。当然，换个角度想想，也许此时的 1× 视角，就是当前环境的最佳构图了。万物皆可 AI，现在轮到摄影了185 年前，达盖尔发明银版摄影法，留住时间成为一种特权；2011 年，胶片巨头柯达传出破产消息，智能手机开始野蛮生长，把镜头塞进了普通人的口袋。回顾整部影像史，其实就是一部打破特权、技术走向大众的普惠史。到了今天，移动影像的狂飙突进，撞上了一堵无形的墙。手机镜头模组越做越大，与之对应的，是最近行业里一个颇具深意的传闻：受制于高昂成本与物理极限，下一代影像旗舰的「超大杯」，大多要面临停更了。这条路为什么走不通了？因为影像旗舰突飞猛进的背后，藏着一个被刻意回避的死结。决定照片质量的，永远是镜头后面的脑袋。普通人有懂美的眼睛，却跨不过光圈快门和构图比例的门槛。把一台堆满顶级硬件的手机，递给一个毫无基础的人，他按下的快门，大概率得到一张平平无奇的随手拍。顶级硬件如果只服务于一小撮懂摄影的人，无疑就会变成一种伪命题。在同行集体踩下刹车的时间点，华为提前把 Pura 90 Pro Max 端上了桌，而 XMAGE 智拍，就是破局的方法。在此之前，行业里已有不少 AI 落地，比如 vivo 让四季流转，OPPO 消除反光。只不过，它们大多是按下快门后的后期创意和修补。XMAGE 智拍往前跨了一大步，让算法成为整个拍摄动作的主导。这种做法谈不上是什么艺术层面的颠覆，但确实管用。降低门槛，让完全没有摄影基础的人，也能相对轻松地把一台旗舰机用出该有的样子。好的技术绝不该是一座孤岛。曾经的智能手机让人们拥有了「随时拍」的自由，现在的 AI 则进一步赋予了大众「拍得好」的能力。这或许是打破当前僵局的一剂良药，也是影像超大杯物尽其用的方法。

过去几年，20 万级新能源 SUV 市场像一条突然变宽的河道。一边是家庭用户对空间、续航和舒适性的刚性需求，另一边是智能化、性能化配置不断下放。过去很多只会出现在 30 万元甚至更高价位车型上的配置，如今开始像雨水一样往下渗透。激光雷达、空气悬架、大模型智驾、零重力座椅，这些原本带有「高配」意味的词，正在成为新一轮价格带竞争里的筹码。肩负着广汽集团高端化重任的昊铂品牌在今年广汽集团重振旗鼓的趋势下，也顺势重新梳理产品架构，昊铂 A800 旗舰轿车前段时间开始上市，用到了乾崑智驾和鸿蒙座舱来增强产品竞争力。紧接着，面向 20 万价位带的 SUV 昊铂 S600 也来了，提供纯电 660 Max、纯电 660 Max+、增程 230 Max、增程 215 四驱 Ultra+ 四个版本，官方预售价为 20.99 万至 22.99 万元，预售权益价为 18.89 万至 20.89 万元。新车定位为「新豪华智慧运动 SUV」，提供纯电和增程两种动力形式，车身尺寸为 5015×1933×1700mm，轴距 2936mm。相比于品牌初创时期的雄心勃勃，昊铂现在显然务实不少，它试图把昊铂 S600 做成一个配置密度很高的样本：性能、智驾、座舱、舒适、安全和权益，几乎都被塞进「满配」这个关键词里。昊铂 S600 请了 2 个代言人，一个代言「优雅」，一个代言「运动」过去提到中大型 SUV，很多人的第一反应是「稳」：空间要大，坐姿要高，后排要舒服，最好还能承担家庭出行里那些琐碎但真实的场景。比如周五晚上从公司直接出发，后备厢里塞进折叠椅、露营桌、孩子的滑板车和一箱零食；又或者长途返乡时，后排老人希望能把座椅放倒一点，孩子希望看动画，副驾则希望在高速上补一觉。对于这类用户来说，车首先不是一个「驾驶机器」，而是移动客厅、临时休息室，也是家庭生活的第二个容器。昊铂 S600 没有放弃这部分的诉求，或者说，这些配置已经在国产中高端 SUV 上成为一种基本素养，昊铂 S600 拥有 90% 得房率、750mm 二排腿部空间、5.316 平方米整车玻璃面积，以及 1292L 二排拓展空间；舒适配置上，同级唯一 18 点按摩零重力女王副驾、143° 躺角后排电调座椅、22 扬 ADiGO SOUND 音响、五块双层夹胶静音玻璃等配置也都被纳入其中。但昊铂 S600 又没有把自己完全包装成一台「沙发车」。它的另一面，是更强调姿态和操控的运动属性，所以昊铂又反复出现「跑山王」这个表达，显然昊铂希望把一台五米多长的大尺寸 SUV 做得不那么笨重，让它不只是适合城市通勤和家庭旅行，也能在山路、弯道、高速超车等场景中保留一定的驾驶参与感。从造型上看，昊铂 S600 采用的是一种前低后高的轿跑溜背姿态，超过 5 米的车长和近 3 米轴距提供了视觉体量，双层鸭尾、宽体车尾、20 英寸运动轮毂等设计元素，则试图削弱中大型 SUV 常见的厚重感，从侧面看它确实不像一般的家用大型 SUV 那样笨重方正，更像是专为年轻家庭所设计。所以这一次昊铂也请了 2 个代言人，英格兰传奇足球明星迈克尔·欧文作为「运动代言人」，知名演员陈数作为「优雅代言人」，两位代言人分别代表这一款车的一种特质，也是试图同时吸引家用取向和个人运动取向的潜在消费者。把四驱、高阶辅助驾驶、双腔空悬都放进 20 万级别在最高配的增程 215 四驱 Ultra+ 版本上，昊铂 S600 同时提供双电机四驱和双腔空气悬架。昊铂表示，这是行业首次在这一价位段同时放入双电机四驱和双腔空气悬架。另外，昊铂 S600 还全系标配前双叉臂后五连杆独立悬架、SDC 可变阻尼减振器和鹰爪系统 2.0，并提供雪地、湿滑、脱困三种路面辅助模式。一台大 SUV 在城市高架上并线，需要车身跟随方向盘动作，而不是像大船一样慢半拍；周末跑山时，连续弯道考验的是侧倾控制和车身收敛；雨天压过积水路面，轮胎抓地力和车身稳定系统要在驾驶者还没反应过来之前，先把车身姿态拉回来。在这些场景下，双腔空悬能够有效提升驾驶体验和驾驶信心。动力部分，昊铂 S600 提供纯电和增程双动力。其中增程四驱版综合功率达到 370kW，满电零百加速 4.3 秒，亏电零百加速 5.6 秒。广汽星源增程技术的持续发电功率为 85kW，官方还提到新车在亏电状态下可实现 165km/h 速度持续刷圈 2 小时，亏电极速 193km/h。增程车过去最容易被质疑的地方，在于「满电一条龙，亏电一条虫」，电量充足时像电车，电量不足后则容易出现动力衰减、油耗升高、发动机噪音变大等问题。昊铂 S600 对 85kW 持续发电功率和亏电性能的强调，其实是在回应这个痛点，一台增程 SUV 不能只在发布会数据里快，也要在长途、高速、爬坡和亏电场景下保持基本的体面。在智能化层面，昊铂 S600 的策略也比较直接：尽量把高阶配置做成全系标配。新车全系标配激光雷达、高通 SA8650 智驾芯片、27 个高精度感知单元，并搭载融合 Momenta R6 强化学习大模型的广汽星灵智行 ADiGO GSD 3.0，支持城区和高速快速路场景。资料还提到，昊铂 S600 下半年将首批升级 Momenta R7 强化学习世界模型。智能座舱方面，新车搭载高通 8295P 座舱芯片、17.3 英寸 3K 中控屏、27 英寸 W-HUD，并内置 11 个 AI 智能体。相比过去车机只负责导航、音乐和语音控制，现在的座舱正在变得更像一个移动终端：它不仅要响应指令，还要理解场景，甚至提前知道用户在车里可能需要什么。下班后坐进车里，导航自动给出回家路线，HUD 把车道级信息投到视线前方，孩子坐在后排看动画，车机根据车内温度和空气质量调整新风，座椅按摩开始工作……这些在新能源汽车上可以实现的体验，对于从 20 万元档位传统燃油车换购过来的用户来说，无疑是跃升式的。广汽集团有着庞大的用户基本盘，这些人群想要增购换购新能源车型的话，广汽自然希望用产品承接他们的需求。就像 AI 说的那样，我在这里，不偏不绕，不躲不逃，稳稳接住你。某种程度上来说，过去几年广汽集团新能源产品的建设，过于重视 B 端，对 C 端用户的吸引力不够，并没有「稳稳接住」用户的需求。对消费者来说，20 万级新能源 SUV 已经足够拥挤，智己 LS6、极氪 7X、小鹏 G7 等车型都在这个区间形成压力。昊铂 S600 要突围，关键命题并不是「它是不是拥有很多配置」，而是这些配置能否被组织成一种清晰的用车体验被感知到：城市里安静从容，山路上不显笨重，长途时不焦虑，家人坐进去愿意放松，驾驶者开起来也不觉得无聊。

龙虾之父一个月消耗 6030 亿 Token，总花费金额高达九百万人民币。移动联通电信，三大运营商都在推 Token 套餐，199 送千兆宽带还有 1 亿 Token，了解一下？从硅谷到国内大厂，Tokenmaxxing 成为公司的主流，谁消耗 Token 多，谁就是 AI 时代的好员工。00 后校友向母校捐赠 20 亿 Token，被网友调侃按 DeepSeek 5 元/亿 Token 计算，只要 100 元。▲图片来自新浪财经Token 在半年内完成了一次身份跃迁：从技术术语，到 KPI，到话费套餐，到捐赠货币。它成了 AI 时代的「度量衡」，唯一的问题是，没人说得清它到底在度量什么。我们自己买 Token，用公司的 Token，部署了一堆 Agent，代码、论文、周报都是 Token 烧出来的。而另一边是，大厂的员工由于 Token 消耗排行榜的原因，开始拿着公司的 Token 处理私事、玩游戏、开发数十个没什么用的子 Agent 来提升自己的排名。「回报」这件事很难量化，但「使用量」可以量化。于是所有人都选择了那个容易量化的东西。这不是 AI 时代的新问题，这是管理学的老病。用 AI 消灭狗屁工作的公司，正在制造新型狗屁工作亚马逊，那个裁员裁到大动脉，把自己的网站都变成 404 的小狗，最近又被爆出了新的「笑料」。原本被寄予厚望、用来消灭「狗屁工作」的 AI，最终却沦为制造新型「狗屁工作」的源泉。据《金融时报》报道，为了逼迫员工拥抱 AI，亚马逊搞出了一个极其复古的管理手段：「Token 消耗排行榜」，追踪每个员工的用量。公司强制要求超过 80% 的开发者每周必须完成 AI 使用指标，甚至将消耗 Token 的数量作为考核标准。▲图片

1.OpenAI 的两大宿敌 Anthropic 和马斯克，放下心中成见之后终于在月初结盟了。在此之前，Anthropic 和马斯克的关系并不融洽：今年 2 月，马斯克还在自己的 X 账号指责 A 社「woke」「邪恶」「反人类」（misanthropic），说这家公司「仇视文明」。事后来看，这次攻击并非马斯克清新脱俗的性格使然，而是 Anthropic 所做的某些事情触碰到他的神经，事出有因。在此之前，xAI 内部使用 Cursor 工作，但是今年年初员工发现，Claude 模型突然在 xAI 的 Cursor 公司账号里不能使用了。当时还在 xAI 上班的联合创始人吴宇怀，在全员信里是这么说的：「Anthropic 更新了政策，要求 Cursor 不得向其主要竞争对手提供 Claude 模型调用能力。」当时，吴宇怀在信中写了一句话，颇为有趣：「这是坏消息也是好消息。我们的生产力会被影响，但这也敦促我们开发自己的编码产品和模型。」为什么当时 xAI 的高层认为，开发自己的编码产品是关键？后来发生的事情，大家都知道了。xAI 的联创团队悉数跑路，马斯克一气之下对 Cursor 使用了钞能力必杀：上个月底，SpaceX 和 Cursor 共同宣布，将在编程和知识类工作 AI 模型的训练上，展开前所未有的战略合作；并且，SpaceX 还获得了以 600 亿美元收购 Cursor 的权利，或向后者支付 100 亿美元合作费用。注意编程这个关键定语，后面还会 call back.2.最近，我看了一条 Cursor 早期投资人、Anthropic 大喷子、T3 创始人 Theo Browne 的视频。本来点进去是看他喷 A 社和 SpaceX 怎么蝇营狗苟，结果没想到，却看到了关于 SpaceX + Cursor 合作的，一个既另类却又极度合理的分析：不说 600 亿的收购，就只说 100 亿的合作费——Theo 在视频里表示，自己认为「哪怕只是交换到 Cursor 的用户数据，这 100 亿也值回票价了。」所以是什么数据？如果你也去看 Theo 这条视频，他会讲得非常清楚。但为了节约时间，我们在这里简单概括一下：我们和 AI 的对话是一来一回的，你提出问题/需求，他给你解答；coding agent 同理，只不过返回的是代码。一次高质量的对话，整个过程，包括用户提示、模型思考、agent 规划、输出代码、验证——所有这些东西合起来，可以称为一个完整的 Agentic Loop——就成为了高价值的训练数据，再喂给模型去进行强化学习，就能进一步提高模型在实战场景下的表现水准。Cursor 有的，SpaceX 想要的，就是这些数据。可这些数据从哪里来呢？答案很简单：作为模型厂商，这种高质量数据的最直接来源，只能是你自己开发的 coding agent 产品——也即 Anthropic 的 Claude Code、OpenAI 的 Codex、Kimi 的 Kimi Code。现在你应该明白了，为什么被 Anthropic「封号」之后，吴宇怀会在全员信里提出开发 xAI 自己的 coding 产品和模型这件事了。这件事 xAI 在当时已经看清楚了：没有自己的编码产品，就没有高质量的强化学习数据；没有高质量的数据，就训练不出真正实战能力强的 coding 模型。虽然有点暴论，但现在我们可以点题了：模型厂商想做出来真正能打的编程模型，做自己的 coding agent 产品是唯一的路径。3.大语言模型像个水晶球，用全网的语料训练出来，似乎能够解答万物，但并不代表它在所有问题上都能给出高质量的答案。用 GitHub 上数以亿计的代码条目训练，当然也能训练出 coding 模型。这是「学习结果」的逻辑，也是没问题的。毕竟编码任务的结果是可以验证的：代码能不能运行，测试能否通过，结果摆在那里。但是，通往结果的过程，是一个涉及多步骤决策、错误纠正、意图对齐的复杂链条。每一次用户的接受、拒绝、补全、撤销、追问、甚至当模型好几次都搞不定或者完全搞错时的辱骂——都是这一链条上的过程信号。强化学习有两种监督方式，一种叫做结果监督，只看最后是否跑通。但是结果监督会催生「奖励黑客」的现象：模型为了能跑通可能写出冗余、脆弱、带逻辑漏洞的代码，但因为测试过了，模型以为自己学对了。而另一种叫做过程监督，对推理路径上的每一步进行打分。上述这些过程信号，只有在 coding agent 运行环境里才能诞生。GitHub 仓库里只有结果，哪怕是去看单独的提交历史，看 PR，都找不到有效的过程信号。在缺乏有效、自主可获得的过程信号的时候，一些模型厂商会采用「蒸馏」的方式，这个事情大家应该已经知道了。蒸馏的逻辑很简单，给同样的输入，老师模型输出什么，学生模型就学着输出什么。但是通过蒸馏，即便可以获取到思维链，得到的仍然更接近于结果，而非被蒸馏的老师模型内部的概率分布。一旦学生在推理中偏离了老师的轨迹，哪怕一个 token 不符合，都有可能发生偏离。这背后是强化学习的基础限制：策略梯度定理要求，优化样本最好由当前正在优化的模型自己去产生。这种数据叫做 on-policy 数据。而通过蒸馏别家模型，在别人的产品里产生的数据，来训练自己模型，都属于 off-policy 数据。模型当然可以从中学到东西，但学不到老师模型内部的概率分布信息。而像 Cursor 这样自己就是 coding agent 产品的公司，掌握着最真实、有效、高质量的训练数据。Cursor 产品本身，就是 coding 模型在实战环境中的最佳训练场。我们可以通过 Cursor 年初的「翻车」，来证明这个逻辑。4.APPSO 读者应该记得，年初 Cursor 发布了 Composer 2，号称「下一代专用编程模型」，技术报道写的相对保守，也没有提供具体的模型底座信息。结果很快，网友就在公开代码片段里发现了 Kimi 的模型 ID，截图传遍了开发者社群，逼得 Cursor 副总裁 Lee Robinson 出面澄清：「Composer 2 确实是从开源底座出发的。最终模型大约只有 1/4 的算力来自底座，剩下 3/4 是我们自己训出来的。」几小时后，Cursor 联创 Aman Sanger 也跟着发了一条道歉：「一开始没提 Kimi 底座是个失误。」五天后，Cursor 放出了完整的 Composer 2 技术报告，显示底座的确是 Kimi K2.5，授权方则是 Firworks AI，大致流程是在 K2.5 上做训练，再继续做大规模强化学习（RL）。但关键之处在于，Composer 2 的 RL 是运行在真实的 Cursor 会话当中，使用与生产部署完全相同的工具和 harness。Cursor 将这套流程叫做「实时强化学习」(real-time RL)，也即将模型的 checkpoint 直接部署到 Cursor 生产环境中，观察用户的响应，收集数据，聚合成奖励信号——最快可以每 5 个小时迭代一次模型版本，然后继续部署到 Cursor 里，循环往复。最极致的案例是 Cursor 的自动化代码补全功能 Tab，每天处理超过 4 亿次请求，每当用户输入字符、移动光标时，模型都会预测下一步动作，如果预测置信度高，则显示建议，用户按下 tab 即接受自动补全。该功能采用的是在线强化学习，在行业内极具特色。Cursor 可以以极高的频率（最快可达每一个半小时到两小时）更新 Tab 的模型能力给用户，直接在产品内收集 on-policy 数据进行训练。这种高频、接近实时的反馈回路，让 Tab 可以学习到极其微妙的用户意图。Cursor 方面透露，这种方法让 Tab 建议的拒绝率降低 21%，接受率提高了 28%。回到 Composer 模型本身。在事情搞清楚了之后，一些 Kimi 员工也删掉了之前吐槽的的推文，Kimi 官方账号发表了祝贺。一家估值 600 亿美元（基于马斯克给的数字），不做自己的模型基座的 coding agent 应用层公司，仍然可以通过产品自身的数据飞轮，RL 出超越基座模型的专有编程模型。所以与其说 Cursor 翻了车，不如说这反而是 coding agent 产品重要性的绝佳例证。Cursor 在另一篇关于实时 RL 的文章里写到：「（训练编程模型）最大的困难在于建模用户。Composer 的生产环境里不只有执行命令的计算机，还有监督和指导它的人。模拟计算机容易，模拟使用它的人却很难。」这句话，现正在逐渐成为了在编程模型方面走在前沿的模型厂商之间的共识。如果你去看 benchmark 榜单和用户普遍评价，会发现哪些头部的厂商都在发力做自己的 coding agent/编程产品。区别只在于谁离用户更近。我们以 SWE-bench、LLM-Stats 等相对权威的榜单为例，Claude、GPT、Gemini、Kimi 等模型基本霸榜前十，清一色都是有自己开发 coding agent 产品（包括 CLI、IDE、集成 coding agent 的桌面客户端）的模型厂商。在部分榜单上会出现少数反例，如 Meta (Muse Spark)、DeepSeek 等，没有开发自己的 coding agent。不过你会发现，这些反例模型，在更加接近真实场景、避免污染的更权威 benchmark 上就很难上榜了。以 DeepSeek 为例，它在 SWE-bench bash only 上分数是 70%，排名第九，在 SWE-bench Pro 上分数却掉到了 15% 左右。OpenRouter 的真实流量数据可以解释这种反差：该平台 2025 年报告显示，Claude token 消费 80% 以上用于编程和技术任务，而 DeepSeek token 消费主要集中于闲聊和角色扮演。没有自家 coding 产品的厂商，在一些 coding 任务 benchmark 上能挤进头部，但在更难的真实工程 benchmark 上，在用户用 token 消费投票的真实流量中，都会原形毕露。不仅是 Cursor，Anthropic 在 2025 年 11 月发的一篇论文里，也明确透露自己在做一模一样的事情：「我们在 Anthropic 自家的真实生产编程环境上做训练。」也即 Anthropic 把自己员工使用 Claude Code 的交互数据，反哺给 Claude 模型用来训练。5.在 AI 的演进历程中，生产要素的定义发生了深刻的位移。传统三大核心要素——算力、研究、训练数据，虽然在总量上持续增长，但在结构上已经出现了严重的失衡。今天的各大 AI 巨头显著提高了在算力上的资本支出 (CapEx)，让算力基建成为了当前舆论的主旋律。但实际上，特别是在编程范畴内，随着 GitHub 仓库、StackOverflow 等互联网公开代码数据被基模厂商「竭泽而渔」式地利用，模型在代码生成与逻辑推理上的边界开始逐渐显现。这也是为什么，行业共识正在逐渐转向一个冉冉升起的新战略高地：对于任何希望掌握顶级代码能力的模型厂商而言，建立自有的 coding agent 产品早已不再是可选的商业路线，而是确保底层模型可以持续进化的核心生命线。正如前面 APPSO 论证的那样，单纯学习公开数据等于只学习成功者的结局，却无法了解成功的路径，这绝对不是正确的成功学应该有的样子。在真实的编程环境中，知道发生了什么错误、怎样发生的、如何正确地理解和高效地实践需求等等——了解正确过程的价值，远超于得到正确结果本身。只有拥有自己的编码产品，模型厂商才能获取高质量的「过程监督」信号，从而在编码/推理能力的下一阶段竞争中，确保自己仍有技术护城河——否则就不得不像 SpaceXAI 那样，花钱去跟 coding agent 产品公司去合作。然而并不是所有模型厂商都跟马斯克一样有钱，以及 2026 年开始的巨头势力划分、结盟与领地的争斗会变得更加激烈，当一家缺乏自主 coding 产品的模型厂商终于回过味来的时候，恐怕已经没有足够的合作伙伴可以挑选，合作

今年 3 月，苹果制造了 MacBook Neo 这个大爆款，把高质量入门笔记本的价格打到了 4599 元人民币起。一个半月后，这台搭载 A18 Pro 处理器的入门款 Mac，全年销量从 700 万台追加到了令人难以置信的 1000 万台，交付周期一度排到了 5 个月之后……苹果紧急向鸿海、广达等供应商追加订单。知名供应链分析师郭明錤给出更刺眼的数字：在全球 Windows 笔记本电脑预计同比衰退超 10% 的大背景下，2026 年 MacBook 系列出货量有望达 2500 万台，同比增长 20%-25%。这一拳，Windows 阵营的最核心代表英特尔，该如何接住？在今天，英特尔在中国正式推出第三代酷睿处理器，代号 Wildcat Lake「野猫湖」，并同时启动了联合产业链伙伴的「Firefly 萤火虫计划」。英特尔的接招，围绕一件事而展开：将最先进的 Intel 18A 制程，下放给低端，或是入门，或是「主流」——随便你怎么称呼它的价位段。野猫盯上新的猎物Wildcat Lake 这个代号，在去年底就已经在业界出现。今年 CES 期间，英特尔高调发布了主打高端的产品线：酷睿 Ultra Series 3，采用 Panther Lake 架构。同一天，Wildcat Lake 也悄悄披露，你可以将它理解为 Panther Lake 的精简版本。业界普遍将其视为英特尔为了对标苹果 MacBook Neo 而推出的产品，专攻平价入门级 PC 市场。「直接对标」这四个字，并不是说说而已。在 4 月底，英特尔为野猫湖推出的公版参考设计曝光，采用渐变配色、铝合金一体成型机身、窄边框键盘，扬声器开孔环绕在键盘两侧、D 面无出风口——这种设计语言不需要过多解释，它就是在用苹果惯常的设计语言，去回应苹果本身最受欢迎的多彩配色与铝合金轻薄机身策略。在 MacBook Neo 上，苹果做对了一件事：它让整个行业重新意识到，在「内存末日」(RAMmageddon) 的今天，入门价位段成了一座金矿。用一颗本来进入销库存阶段的 A18 Pro 芯片，配上 8GB 内存和 256GB 存储，开价 4599 元——MacBook Neo 活脱脱成了 MacBook Air 平替。英特尔，以及整个 Windows PC 阵营不能继续将这个用户群体拱手相让。而这，正是「野猫湖」存在的（几乎）全部理由。18A 首次下放到入门段第三代英特尔酷睿处理器的技术参数并不复杂，CPU 采用 2 个 Cougar Cove 性能核，加上 4 个 Darkmont 超低功耗能效核，总共 6 核 6 线程；GPU 部分配备 2 个 Xe3 核心，NPU 算力最高可达 18TOPS。整机基础 TDP 15W，最高 35W，在无主动散热的笔记本上，甚至可以压低到 11W 运行。按照英特尔今天提供的官方口径，这颗芯片功耗相比前代降低超过 60%，图形 AI 性能提升了将近 3 倍，平台总算力更是达到了 40TOPS。当然，考虑到「野猫湖」处理器的 NPU 算力只有 17TOPS，远低于 Windows 阵营的所谓「AI PC」，也即微软官方定义的 Copilot+ PC 40TOPS 的阈值。所以严格来说，野猫湖并不能直接和苹果 A18 Pro 高质量的神经网络引擎，以及统一内存架构，去做横向比较。话虽这么说，驱动参数量不大的本地小模型，做一些 AI 推理任务，还是绰绰有余的。以及更有意思的是，谷歌前段时间发布自家的 AI 笔记本电脑「Googlebook」的时候，也确认了和英特尔结盟——Wildcat Lake 被海外业界人士普遍认为将会成为是 Googlebook 的核心平台。连谷歌都觉得够用，应该没大问题吧？真正关键之处不在于这些参数，而是在于「野猫湖」用上了 Intel 18A 制程工艺本身。要知道，过去两代酷睿 Series 1 和 2，一直沿用 Raptor Lake-U 架构和 Intel 7 制程，实质上是 10nm SuperFin 工艺的演进版本。在主流芯片普遍走向 5nm 甚至 3nm 的当下，技术代际差距已经相当明显。当 MacBook Neo 用虽然是老库存但货真价实 3nm 工艺的 A18 Pro，杀进四千多元的价位段，英特尔过去那套「靠老制程压低成本」的入门策略已经过时了。所以，我们很欣慰地看到，英特尔将 18A——2nm 级别的工艺，压到一颗入门芯片上。这意味着英特尔充分愿意承担初期更高的单位成本，去力保一座不能再丢弃的城池。系统：手机供应链思路搬到 PC 上对于守住市场，「Firefly 萤火虫计划」同样重要。英特尔副总裁、中国区软件工程和客户端产品事业部总经理高嵩如是说：芯片本身的算力固然重要，但「芯片和系统层面的双重创新」才是这一代主流 PC 体验跃迁的来源。中国区客户端与平台销售业务部总经理宗晔则认为，PC 也正在从工具进化为「智能体验入口」。根据英特尔提供的内部数据，与上一代相比，「野猫湖」产品的 PCB 面积缩小了 5%，器件数量减少了 7%。同时针对第三代酷睿平台特性制定并推广标准化主板与 I/O 板连接，推动主板核心电路进一步实现模块化、小型化、低功耗化。这一套打法的精神底色，其实来自于手机。手机行业过去 10 年磨合出的供应链协同方式——元器件高度集成、模块化主板、统一接口标准，被英特尔搬到了入门级 PC 上。结果，就是更轻、更薄，也更便宜。参与「Firefly 萤火虫计划」的合作伙伴，包括但不限于华硕、七彩虹、荣耀、惠普、联想等。英特尔预计，全球合作伙伴将陆续推出超过 70 款基于 Wildcat Lake 平台的产品。目前，荣耀 MagicBook X14 2026 已经率先上架，搭载该平台 Core 5 320 处理器，采用 16GB+512GB 存储配置，定价下探到了 4399 元。用户画像：完全对标 MacBook Neo，但还有更多英特尔列举了 Wildcat Lake 平台产品的 5 类目标用户：辅导孩子作业的家长、全天在校的学生、刚入职场的新人、「一人公司」创业者，以及银发族。这些用户群体和场景描述本身没什么悬念，但值得注意的是，这个用户画像，和 MacBook Neo 在线下门店主要推荐的人群，几乎是完全重叠。MacBook Neo 夺走的本来就是入门级 Windows 笔记本电脑的客群，英特尔的接招，就是造出具有更好价值主张的入门级笔记本电脑，再把心智重新抢回来。英特尔还反复强调野猫湖平台的本地 AI 推理算力。这对于那些购机预算本就不多，也不太可能付费订阅优质云端模型算力的用户来说，倒也是个正确的思路。最后，野猫湖看起来像是合格的入门芯片平台，跟 Firefly 萤火虫计划结合起来，也形成了一套颇具诚意的组合拳。轻、薄、便宜的 AI 笔记本，英特尔现在重新回到了这个擂台。仅剩的悬念，是 OEM 最终能把零售价做到多少。在 MacBook Neo 的价位上，苹果给到了系统生态、统一内存架构、多种潮流配色等等诸多看头。但英特尔要想把这只野猫放出去，跟苹果大杀四方，光有 18A 工艺和优秀的工业设计是不足够的。这场仗的真正交战点，大概率不会在技术本身，而是在即将到来的 618 购物节期间，第一批搭载它的笔记本产品，能否真正把供货紧张的 MacBook Neo 杀个七零八落、措手不及。

⚖️马斯克诉 OpenAI 案败诉📱苹果开启 WWDC 邀请，将公布苹果 AI 新进展🚗问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处🤖国家反诈中心 App 上线「AI 内容鉴定」⌚苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动🚙广汽董事长回应被称网约车之王🧠三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能🎵腾讯音乐官宣收购喜马拉雅💰铠侠股价飙升，计划美国上市🚀马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元☁️百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半🎬爱奇艺一季度由盈转亏 2.95 亿元🦿荣耀机器人「闪电」「元气仔」走进南大💡前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去🚗小米 YU7 GT 定档周四发布🎮5499 元起，红魔 11S Pro 系列发布马斯克诉 OpenAI 案败诉据 CNBC 和《华尔街日报》报道，美国加利福尼亚州奥克兰联邦陪审团已就埃隆 · 马斯克诉 OpenAI 及其 CEO 山姆 · 奥特曼一案作出裁决，认定马斯克的起诉超出诉讼时效，驳回其全部主张。此案源于马斯克 2024 年提起的诉讼。他指控 OpenAI 及奥特曼违背了将该公司作为非营利机构运营的承诺，并在庭审中称 OpenAI「窃取了一家慈善机构」。陪审团的裁决并未就「违反慈善信托」指控的实质内容作出判断，而是认定马斯克的起诉超过了三年诉讼时效。法院同时驳回了针对微软的相关指控。裁决宣布后，马斯克在 X 平台发文，将这一结果定性为「日程上的一个技术性问题」，并表示将向第九巡回上诉法院提起上诉。OpenAI 首席律师威廉 · 萨维特在庭后表示，这「不是一个技术性裁决，而是一个实质性裁决」，并称马斯克是在将诉讼作为竞争手段使用。微软方面亦发表声明，对裁决表示欢迎。苹果开启 WWDC 邀请，将公布苹果 AI 新进展今天凌晨，苹果正式启动 WWDC26 邀请，并公布详细日程安排。本届 WWDC 大会将于北京时间 6 月 9 日至 13 日举行，面向全球开发者开放，重点展示 Apple 各平台的新系统能力、开发工具和框架更新。按照安排，WWDC26 将于北京时间 6 月 9 日开幕，届时 Apple 将举行主题演讲和 Platforms State of the Union，介绍新一代软件平台、人工智能相关进展，以及面向开发者的新工具。大会期间，开发者和学生可以观看超过 100 场围绕工具、技术和设计的视频讲座，参与团体实验室活动，并在 Apple 开发者论坛中与工程师和社区成员交流。Apple 还将在 6 月 9 日邀请超过 1000 名开发者、设计师和学生前往 Apple Park，参与线下特别活动。届时，爱范儿也将在 Apple Park，为大家带来第一手的现场报道。问界 M9 设计师公开点赞理想 L9 内饰：比例和谐，口味控制得恰到好处全新一代问界 M9 系列汽车设计师 @Zimo-子墨 日前在微博发文，点赞全新理想 L9 的内饰设计，称其「简洁、逻辑、比例和谐，视觉重心清晰，口味控制得恰到好处」。他表示，初创公司在有强决策者的情况下，往往对设计管控介入较深，容易打散设计的贯穿逻辑；而到第二代、第三代产品时，无论前代成功与否，团队通常会更多放权给设计师打磨表达。他同时透露，新 L9 的内饰设计总监为来自宝马的 Chris Lee，并称虽然双方长期管理直接竞品，但「真的欢喜好设计，生在中国」，并已发短信向对方表示祝贺。国家反诈中心 App 上线「AI 内容鉴定」据央视新闻报道，国家反诈中心 App 近期正式上线「AI 内容鉴定」功能，支持对图像、视频、音频及文本进行 AI 生成痕迹检测，主要用于识别诈骗及电商场景中的伪造内容。电商平台上利用 AI 篡改图片骗取「仅退款」的行为日趋普遍。央视新闻记者随机选取十张商家提供的疑似 AI 生成售后图，上传至国家反诈中心 App 及多个 AI 鉴真工具后，均被判定为「图像含 AI 生成痕迹」。据介绍，该检测方案结合小模型与大模型：小模型学习 AIGC 内容的生成模式，大模型则判断内容是否存在物理瑕疵或逻辑语义问题，二者联合输出最终结果。苹果 Apple Watch 改款在即，独家传感器供应商 TASC 订单有望提前启动据 DIGITIMES 报道，苹果计划于今年下半年对高端 Apple Watch 机型进行重大改款，传感器功能将同步升级，市场预计届时出货量较 2025 年增长 20% 至 30%。Apple Watch 11 已具备血氧、脉搏、心率及血压感测能力，下一阶段预计推进无创血糖监测，具体落地时间取决于 FDA 审批进展。作为苹果传感器元件的独家供应商，亚洲半导体（TASC）预计将直接受益。市场预期新品订单最早于 7 月启动，带动下半年业绩明显超越上半年。广汽董事长回应被称网约车之王据「睿见 Economy」报道，日前广汽集团党委书记、董事长冯兴亚在「第十八届轩辕汽车蓝皮书论坛」上，针对「网约车之王」的说法，其指出：很多消费者可能没有想明白一个问题，「网约车是生产资料」。生产资料对品质的要求极高，因为它是用来挣钱的工具。只有能为自己带来更多利益的车，才能受到使用者和购买者的欢迎。对司机来说，车要「多快好省」，要能赚钱。「这恰恰是对品质最好的验证。没有高品质，就不可能受到网约车驾驶者的欢迎，这是最基本的。」他进一步分析，网约车的乘坐周期非常长，一天可能十几个小时都在车上。如果舒适性不够、疲劳感很强，它也不会受到司机的欢迎。此外，电耗等各项指标，也决定了网约车司机是对车要求最严格的一群人。冯兴亚总结道，这些从网约车运营中积累下来的经验，广汽会延展到旗下各个品牌，让车辆变得更加可靠、舒适和安心。「这反而是我们的优势所在。」三星良率追赶提速，韩国两大内存厂同步加码 DRAM 产能据 DIGITIMES 报道，三星电子近期大幅提升了用于高带宽内存（HBM）的最新 DRAM 良率。其第五代 10nm 级 1b DRAM 良率已达 92%，用于下一代 HBM4 产品的第六代 1c DRAM 良率也突破 75%。良率的提升有望增强三星的成本竞争力，并提升其向主要 AI 客户大规模供应先进存储产品的能力。在产能扩张方面，据 Ddaily 报道，三星和海力士正同步提前推进工厂扩建和设备订单。三星正加速平泽 P4 工厂的 DRAM 投资，设备安装预计持续至今年底，市场估计明年将进一步扩大投入。海力士则在清州 M15X 和龙仁 Y1 工厂推进扩产，业界预计其明年 DRAM 相关资本支出有望较今年翻倍。腾讯音乐官宣收购喜马拉雅腾讯音乐娱乐集团（TME）昨日宣布，已正式完成对喜马拉雅并购交易的交割，喜马拉雅成为 TME 全资子公司。交割完成当日，TME 执行董事长彭迦信与 CEO 梁柱联合向全体员工发布全员信。管理层称，喜马拉雅是中国长音频领域的重要建设者，双方战略方向高度一致，合并后将共同拓展在线音乐与音频行业空间。铠侠股价飙升，计划美国上市据彭博社报道，铠侠（Kioxia）股价昨日上涨 16%，创下新高。公司在上周披露强于预期的业绩展望后，多家华尔街机构迅速上调目标价；与此同时，铠侠也确认正准备在美国发行存托凭证，借此提升流动性并扩大国际资本市场接触面。报道提到，铠侠预计截至今年 6 月的季度营业利润将达到 1.3 万亿日元，高于市场平均预期，季度利润也创下纪录。多家机构认为，在 AI 带动的存储供需紧张环境下，公司仍有进一步提升自由现金流和股东回报的空间。马斯克：SpaceX 很快开始 IPO，拟融资 750 亿美元据彭博社报道，马斯克昨日在萨姆森国际智能移动峰会线上视频连线时表示，正在德克萨斯州推进 SpaceX 上市准备，称「SpaceX 的 IPO 事宜将很快开始」。知情人士透露，SpaceX 最快本周提交 IPO 申请文件，计划融资最高 750 亿美元，目标估值超 2 万亿美元，若成功将成为史上规模最大的 IPO。百度一季度营收 320.75 亿元，AI 新业务收入首次占一般性业务过半百度昨日发布 2026 财年第一财季业绩，期内实现营业总收入 320.75 亿元，同比下降 1.16%；归母净利润 34.45 亿元，同比下降 55.36%；经营现金流为 27 亿元。具体来看，百度核心 AI 新业务收入达到 136 亿元，同比增长 49%，首次占百度一般性业务收入比重超过一半；其中智能云基础设施收入 88 亿元，同比增长 79%，GPU 云收入同比增长 184%。萝卜快跑在一季度的全无人自动驾驶运营订单达到 320 万单，总订单量同比增长超过 120%，截至今年 5 月其全球业务已覆盖 27 个城市，累计自动驾驶里程超过 3.3 亿公里。爱奇艺一季度由盈转亏 2.95 亿元爱奇艺昨日发布 2026 财年第一财季业绩，期内实现营业总收入 62.26 亿元，同比下降 13%；归母净利润由盈转亏 2.95 亿元，去年同期为盈利 1.82 亿元；经营现金流为 1.86 亿元，同比下降 45.13%。分业务看：会员服务收入为 42 亿元，同比下降 5%；在线广告服务收入 12.4 亿元，同比下降 7%；内容发行收入 3.59 亿元，同比下降 43%；其他收入 4.27 亿元，同比下降 49%。公司在财报中提到，热门剧集仍推动会员收入环比增长，海外业务保持稳健增长，并计划借助AI降低内容制作成本。荣耀机器人「闪电」「元气仔」走进南大5 月 17 日，荣耀携自研机器人「闪电」与「元气仔」在南京大学开展具身智能科普活动，并同步展示荣耀 600 系列新品的核心影像功能。活动现场，荣耀展示了自研机器人的核心运动与交互能力。搭载高动态运动系统、支持自主感知导航与高速避障的机器人「闪电」与在校学生进行了 8 轮 1V5 短跑测试；另一款机器人「元气仔」则完成了包含「太空步」在内的复杂舞蹈动作演示与人机互动。同期，荣耀在现场设立了新品体验区，重点展示荣耀 600 系列智能手机。官方信息显示，该机型主打基于 Magic OS 系统的 AI 能力及 4K Live Photo 动态照片功能。外观方面，荣耀 600 系列目前已公布「幸运星」「青苹果」「光羽蓝」和「曜石黑」四款配色，其中主打配色「幸运星」采用了行业首创的双重磁吸工艺。💡 前 SpaceX 工程师驻扎深圳 8 周后：想做硬件创业，现在就飞过去前 Tesla 无人驾驶与 SpaceX 猛禽发动机团队工程师 Zac Valles 在 YC Demo Day 结束 72 小时后直飞深圳，驻扎 8 周后携新硬件返回旧金山。他随后在 X 上发布了一篇长帖，向所有有意从事硬件创业的人发出忠告：不论有没有融资、团队甚至创业想法，都应该立刻去深圳。只要你对创办一家硬件公司有哪怕一点点想法，就去深圳。融资前、组队前、有想法前、离职前，都不是借口。去就对了。在切入供应链的路径上，Valles 建议创业者围绕一场大型行业展会安排行程，将其作为接触工厂的跳板。他还建议在华强北预留 4 到 6 小时实地走访，称其为一座「微型城市」，机器人子组件、电池、底盘、电子元器件均有专属楼栋聚集。小米 YU7 GT 定档周四发布昨日，小米正式宣布，将于 5 月 21 日（本周四）举行人车家全生态新品发布会，届时将发布小米 YU7 GT、小米 17 Max 等新品：小米 YU7 GT：定位跑车级 SUV，最大马力可达 1003ps，最高时速 300km/h，CLTC 纯电续航 705km；小米 17 Max：采用 6.9 英寸极窄四等边直屏，搭载徕卡 2 亿像素主摄和 8000 mAh 金沙江电池；小米手环 10 Pro：最薄 9.7 mm，最轻 21.6 g，续航最长可达 21 天；小米首款耳夹式耳机：单机重 5.5 g，支持 LHDC 5.0。5499 元起，红魔 11S Pro 系列发布红魔昨日正式发布红魔 11S Pro 系列，国内首发第五代骁龙 8 至尊领先

按：从播客访谈到深度长视频，越来越多的内容创作者把表达重心放到了音视频里；「先转录、再阅读」也因此成了不少重度信息消费者的日常工作流。。本期「装了啥」，我们请编辑部几位同事分享各自正在使用的转录方案，希望其中至少有一种思路能为你所用。MacWhisper 的平价替代：TranscribeX@广陵止息：虽然我很喜欢文字传递的高密度信息，但越来越多的内容创作者已经转向播客和视频；至于多人访谈类节目，音频传达出的现场氛围感也比文字更生动。所以我这些年消费音视频节目的时间也逐渐拉长，但有价值的节目「三心二意」地听一遍远远不够，反复听又没有足够的时间和精力。所以，现在遇到有价值的节目，我都会转录成文本，方便后续检索。从 Time Machine 里翻了翻，找到了当时的文件夹记录以前我采用的是whisper.cpp配合 OpenAI Whisper Large V3 模型来转录文本，但这个方案有两个问题。一是音视频下载的渠道很多，挨个手动下载很麻烦；二是转录时间偏长的中文视频时，有不小的概率会在转录过半之后出现重复输出、卡字，导致转录失败。但我实在不想翻 log 和模型仓库找相关讨论了，所以也开始寻找替代方案。首先想到的是 MacWhisper，看了眼当时价格已经涨到完全不考虑的地步。好在后来找到了整体体验和 MacWhisper 很类似的 TranscribeX，除了收录在 Setapp 里，对我没有额外成本外，即使单买也只要 MacWhisper 一半价钱。和 MacWhisper 类似，TranscribeX 支持本地文件导入转录、麦克风录制、录制特定软件、实时转录、视频网站下载转录和批量转录。其中「下载视频网站」功能主要依赖 yt-dlp 实现，所以基本不用担心国内服务的下载问题。转录模型方面，TranscribeX 也支持 WhisperKit、Parakeet、千问和 Mistral 的转写模型，必要时还可以调用 ElevenLabs 的在线 API。不过我本地基本只用 Whisper Large V3，别的模型也没用过，所以暂时不评价转录效果。相比 MacWhisper，TranscribeX 支持 Apple Silicon 中的 NPU 加速，这也是我之前选 whisper.cpp 的原因。NPU 虽然不会让转录速度快很多，但至少负载不会全部堆到 GPU 上。设置时需要注意：Audio Encoder：主要负责「听懂声音」的环节，可以只用 NPU；Text Decoder：主要负责听写的环节，为了转写速度，建议设置为 All。转写界面其实每个软件大同小异，但还是有一些设置可以说一下。首先是自动检测「说话者」：Pyannote：适合会议、访谈等场景相对固定的情况，也更贴合我常用的场景，可以快速调用预定义的模板；Sortformer：偏向速度、实时和轻量化的处理方式，适合会议、通话等实时场景。外语苦手也不必担心，TranscribeX 内置了三种翻译方式：Apple 内置的翻译 API、DeepL API 以及第三方 AI API。Apple 内置「翻译」速度很快，准确率也够用，而且免费；对翻译质量有要求的话也可以调用 DeepL 或者其他 AI。TranscribeX 也随大流加入了 AI 对话功能，可以针对转录的内容做进一步处理。最后值得一提的是，针对 Setapp 版本的 TranscribeX，所有的在线 AI 额度都可以用 Setapp 的 AI 配额。但 TranscribeX 还是有自己的问题：没有 CLI 版本，远程使用一定要靠 VNC 连接到 Mac；辅助框架的设计上也有不小的问题，目前还很难配合比较火的 AI 助手。如果你比较在意自动化体验，TranscribeX 可能没有那么适合你。用 Gemini 整理@PlatyHsu: 我这里主要谈谈转录后的整理。对于这类任务，我最推荐 Gemini 系列模型。主要原因是，在排名靠前的模型中，Gemini 的多模态支持比较全面，Gemini 3 全系（Pro、Flash 和 Flash Lite）都可以直接接受音频和视频文件输入，免去了先创建音频转录或者视频截图的功夫。我也认为 Gemini 的中文能力在国外模型中相对是比较好的。与此同时，Gemini 的用量给得比较慷慨，AI Pro 方案很少遇到不够用的情况。我最常用的场景之一是整理播客的录音稿，用来确认一些初听时一闪而过的模糊细节。一个参考提示词是：转录用户提供的音频，并整理成表达规范、分段清晰的书面格式。区分说话人，并标注在段落之前。将较长的片段合理拆分成若干自然段。可以删除语气词和重复片段或修改病句，不要做其他删减或总结。具体模型方面，一般可以先从 Flash 开始尝试，对于日常主题的节目来说一般就够了。一小时的播客大约需要四分钟处理，肯定没有本地模型快，但是成品高度可读，值得这个等待时间。不过我也遇到过一些涉及哲学、历史等较专业话题的节目，用 Pro 模型会有更好的表现，可以灵活调整。当然，如果想节省 token，你也可以先在本地转写一遍，然后用以下提示词整理文本：将用户给出的音频转录整理成表达规范、分段清晰的书面格式，可以删除语气词和重复片段或修改病句，不要做其他删减或总结。这个用例一般建议直接使用 Pro 模型，因为本地模型转写的文本一般有很多拼写错误和赘余，Pro 能更好地还原和修复。除了音频之外，Gemini 也可以用来总结视频。并且由于它可以直接「看到」视频内容，因此即使是那种没有旁白的 TVC，也可以说出个一五一十。例如之前一个项目中，甲方提供的写作素材之一是一条广告片，我就是用这个方法把其中的信息提取成文字描述，然后用在了写作里。以下是用苹果最近一条 Mac 广告做的演示，提示词是：详细描述该视频的主题、思路和内容，并列举关键帧的时间戳和画面内容。因为涉及较多「言外之意」的解读，这个用例也是用 Pro 模型处理为佳。以上方法在 Gemini 网页版和 AI Studio 都可以使用，但网页版经常会出现「偷懒」的情况，只给出非常简略的输出，因此建议尽量用 AI Studio 来操作。命令行工具搭配 Agent@waychane: 由于常有参加线上会议的需求，会中涉及主讲者的介绍、采访以及 QA 等内容，这类会议的时长动辄就有一个小时。在会后创作内容时，如果能有一份观点清晰、结构清楚的要点大纲，会省下不少功夫。现在可以说是 Agent 时代，我自己也在用 Claude Code 等 Agent 类 AI 工具帮我处理各种工作。因此在挑选特定场景下的 app 时，我会希望它们要么提供 CLI 操作方式，要么就是纯粹的命令行工具：一方面，这样可以在同一个 Agent 界面中高效地调用不同工具；另一方面，许多 app 不仅功能复杂，使用参数也多到难记，用 Agent 调用刚好能解决这些问题。面对会议音频转录的需求，我选择的是 whisper.cpp 这个命令行工具。它是 OpenAI Whisper 的 C/C++ 移植版本：OpenAI 官方版 Whisper 基于 PyTorch 的 Python 实现，要跑起来得先配好 Python 环境和一整套依赖，在 Mac 上只能依靠 CPU 计算，想用到 GPU 还得自己折腾 PyTorch 的 MPS 后端，体验并不顺畅；whisper.cpp 则用 C/C++ 重写了推理部分，原生接入 Apple 的 Metal 后端和 Accelerate 框架，能有效利用 Apple Silicon 的 GPU 和统一内存架构，模型权重也换成了体积更小、加载更快的 ggml 量化格式。

本周看什么 | 最近值得一看的 8 部作品主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注潘誉晗少数派作者潘誉晗关注潘誉晗少数派作者关注利兹与青鸟少数派作者利兹与青鸟关注利兹与青鸟少数派作者关注SHY少数派作者SHY关注SHY少数派作者少数派编辑部 等 4 位作者05/15 18:00☕️ TL;DR近期佳作推荐：[国产] 低智商犯罪、[韩剧] 夺命许愿、[英剧] 好兆头 第三季、[法剧] 抢钱女王团、[电影] 蒙特利尔，我的美人、[动画] 石纪元 科学与未来 第 3 部分、[纪录片] 马克斯·维斯塔潘：新征途、[真人秀] 旧洞基朋友们几则精彩预告：《攻壳机动队 THE GHOST IN THE SHELL》新预告、《沉默的审判》释出预告、《火遮眼》新预告、《爱情城事》定档预告、《奥特曼起源》正式预告几则影视资讯：《诺曼底 72 小时》内地定档 6 月 6 日、《宇宙巨人：希曼崛起》内地定档 6 月 5 日、《家弑服务》内地定档 5 月 29 日、《超级少女》内地定档 6 月 26 日、漫威新剧《幻视追寻》正式定档往期推荐可点此或点击文末 tag #本周看什么，更多主题片单可点此查看；想要查看更多深度影视内容，可以关注我们的衍生栏目「看什么 Café」。如有其他建议也欢迎在评论区提出，感谢大家对「看什么」栏目的喜爱和支持。[国产] 低智商犯罪关键词：喜剧 / 犯罪片长：45 分钟左右（单集）× 24 集观看渠道：爱奇艺/咪咕视频；豆瓣链接运气之神降临！@利兹与青鸟：平安市刑警张一昂能力不强但热爱吟诗，常常不合时宜得令人发笑，被下放到老干部处都能惹出乱子。因为他正巧捡到一封举报信，于是和同样脑回路清奇的李茜一起被派到三江口，调查公安局副局长的死是否有隐情。但空降而来的张一昂很不受欢迎，工作处处受限，却在阴差阳错之下抓住一个在逃嫌犯，让人刮目相看，就这样被架起来成了神探。另一边，平安市两个笨贼想出了抢贪官的主意也来到三江口；而三江口涉黑的集团老板周荣考虑着新项目，也着手调查发小叶剑刑警队长的死因。三条故事线、两桩凶案（可能不止）就这样汇集一处，想必会给三江口这一团乱麻摩擦出不少火花。本剧改编自紫金陈的同名小说，一开局就满是带着黑色幽默的荒诞味，罪犯和警方彼此高估，虽然智商时常掉线，但不妨碍瞎猫碰上死耗子，不禁让人好奇这份「好运」还能带来什么啼笑皆非的巧合。马旭东、闫佩伦、詹鑫等喜剧演员的助阵也让笑点呈现得更好，尽管罪案线稍显拖沓，但也算是合格的电子榨菜。5[韩剧] 夺命许愿关键词：剧情 / 恐怖又名：기리고 / Wish Your Death / If Wishes Could Kill片长：44 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接为了愿望达成，你愿意不惜一切代价吗？@潘誉晗：这次数学考试的结果有点出乎意料，因为崔亨旭拿满分了，可这家伙不是个读书的料子。面对好友们的询问，亨旭得意洋洋地把一个名叫「Girigo」的软件链接分享给了伙伴，说是只要写上自己的生辰八字贴在自己的身上，然后录下视频上传到这个软件上，就能实现愿望。大家并不相信亨旭的话，亨旭也没在意此刻出现在他手机上的倒计时。谁承想 24 小时后，亨旭就当着全班同学的面自缢了——原来这则所谓的许愿软件，在愿望达成后会要你的命。Netflix 韩国这次又推出了一款爆剧，一经上线就在 13 个国家登顶。剧集整体的节奏很是紧凑，恐怖惊悚的氛围也营造得相当出色，很好地把校园青春剧和恐怖因素结合在一起，然后又完美连接了手机 app 和萨满巫术。神秘的夺命软件利用了青少年的欲望，随后就像是一个大家无法挣脱的隐形牢笼，不仅顽固地无法在手机上卸载，更是在许愿者身后步步紧逼、致人于死地。2[英剧] 好兆头 第三季关键词：剧情 / 喜剧 / 奇幻又名：Good Omens Season 3片长：99 分钟观看渠道：Prime Video；豆瓣链接我有那片宇宙，我还有你，我已经拥有我想要的一切。@SHY：成为至高大天使的亚茨拉斐尔，开始筹备「第二次降临」，复活的耶稣却不知所踪，只得向自暴自弃的克劳利求助。这对自创世之初就有交情的天使和恶魔，能否像往常一样弥合分歧，而在背后作祟的幕后黑手，又会将整个世界导向怎样的结局？受原作者兼前两季主创尼尔·盖曼的性侵丑闻影响，最终季被砍成了仅 1 集的特别篇，篇幅势必捉襟见肘。不难想见，成片处处流露着删减的痕迹，忽快忽慢的节奏让叙事更加混乱，整体观感有着不小下滑。虽然填充的内容变了样，两位新编剧还是尽力还原了盖曼构想中的结局，而在破碎的字里行间，也能窥见一些残存的灵光。惯常的黑色幽默难以撑场后，本季的重心落在了本来带点调侃的父母爱情故事上，大卫·田纳特和麦克·辛以一如既往的真挚演技力挽狂澜，继续迸发充沛的化学反应。彻头彻尾的粉丝向结局，评价两极分化在意料之中，好在，本作原本就是作为限定迷你剧立项，如果你有点难以接受这一季的走向，只把第一季尊为正统也未尝不可。发表首评[法剧] 抢钱女王团关键词：喜剧 / 惊悚 / 犯罪又名：Les Lionnes / Cash Queens片长：55 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接我们之所以抢钱，真的是迫不得已。@潘誉晗：罗莎这天找到自己的上司，希望可以提高她的信用卡额度。可上司不仅很快就拒绝了她的请求，甚至通知她的额度被降低了，因为她征信由于入狱前夫的债务问题受到了影响。罗莎不敢相信上司的话，也就是在这时，银行的警报铃响了，原来这天又是银行例行的演戏日。看着上司希望职员们保障生命的大喊声，一个不该出现的念头在她的脑海中出现了——她决定伙同自己的好友，抢劫自己工作的银行。Netflix 法国于两个月前推出了一部女性犯罪喜剧《抢钱女王团》，讲述了五名女性在窘迫生活的逼迫下，不得不伪装成男性抢劫银行的故事。全剧充斥着大量很荒诞又很现实的设定，主角团的五名女性拼命工作、好好生活，但却一直过得不幸福。职场上的性别歧视、身上背着重债、被同侪做局丢掉了工作、被丈夫家暴……不是说努力就能有好结果吗？为什么在她们的身上，越努力却似乎越悲伤呢？发表首评[电影] 蒙特利尔，我的美人关键词：剧情 / 同性又名：Montréal, ma belle片长：117 分钟；豆瓣链接被困在夹缝中时，是面对自己还是活在谎言里？@利兹与青鸟：凤霞一家已经移民加拿大十几年了，在魁北克省的蒙特利尔开着一间小超市。丈夫性格较为古板，苦于没有当地学历，找工作频频受挫；大女儿正值叛逆期，对待父母多是指责和抱怨；儿子也只是上小学的年龄，家中没有人能为受困于更年期的凤霞提供情绪价值。在法语课上，凤霞被年轻人追求梦想的热情所鼓舞，尝试接触同性交友网站，并结识了 30 岁的卡米尔。这让凤霞重新找回了自己，却引发了家庭危机。本片的华裔女导演很有相关生活体验，她用代际间新旧观念的碰撞、婚姻大于爱情、照顾家人大于自我实现等细节，呈现出一个传统而压抑的中国移民家庭外壳。片中演员也多是当地华裔，汉语都很纯正，对华语区观众的耳朵十分友好。电影虽是在讲中年女性在道德边界寻找自我，却没有激烈的口号，而是用朴素温和的风格、平缓的节奏，给主演陈冲留足表演空间，为年过半百却天性浪漫、依然优雅而富有魅力的凤霞赋予灵魂，细腻又克制地演绎出她内心的彷徨与挣扎，也让观众沉浸其中。女主角是陈冲啊？1[动画] 石纪元 科学与未来 第 3 部分关键词：漫画改 / 科幻 / 冒险又名：Dr.STONE SCIENCE FUTURE 第3クール片长：24 分钟（单集）× 13 集，每周四更新观看渠道：巴哈姆特动画疯；豆瓣链接We choose to go to the Moon.@SHY：孤注一掷的抗争后，科学王国终于取得对北美阵营的胜利。面对迫在眉睫的威胁，千空和赛诺搁置矛盾，联手打造宇宙飞船，前往月球探寻石化的真相。为了复活更多人类，同时取得不可或缺的材料，众人兵分三路执行任务，「WHY 人」却传来了更加危险的信息。从石器时代到现代文明，跨越 200 万年光阴，这段从零开始的科学冒险终于迎来最终幕。从刚苏醒时手工打造的粗糙制品，一步步重现火药、铁器、蒸汽、电力，来到本季，科技树终于攀升到计算机和互联网的领域。要实现登月的最终目标，需要远胜以往的人力和物力，千空等人因此踏上环球之旅，整合各个国家的资源。即使原料由石头进化到超合金，舞台由村落跃进至外太空，本作的核心魅力始终如一。秉持理性和探索精神，用科学之力对抗奇幻，挥洒荡气回肠的热血，开拓人类未来的道路，诠释纯粹的浪漫。这个集结全世界力量、于石之世界铸造的「阿波罗计划」，绝对能让人心潮澎湃，剧中历尽千辛的角色们，也值得最为美好的结局。发表首评[纪录片] 马克斯·维斯塔潘：新征途关键词：纪录片又名：Max Verstappen - New Ground / 极速新域：维斯塔潘的北环之路片长：23 分钟左右（单集）× 3 集；豆瓣链接出发吧！汽车人！@潘誉晗：对于喜欢一级方程式赛车的人而言，效力于红牛车队的马克斯·维斯塔潘，是一个了不起的名字。他出生于赛车世家，自己的赛车成绩也非常傲人。不过卸下这一身份，他其实还有另外一种人生——一家 GT3（Grand Touring 3）车队的老板兼赛车手。而这部只有三集的纪录片，就是讲述维斯塔潘的第二身份。「看，老板亲自上场了。」维斯塔潘没有老板的派头，招呼完就直接上了场；他也不需要拥有全部欢呼与掌声，而是会亲自去挖掘赛车手。于是，他发掘了克里斯，这个和他一样从开卡丁车（并且克里斯的车号是 3）开始的小伙子很有天赋，可苦于 15 万欧元的培训资金，只能把梦想转移到模拟器赛车上。当维斯塔潘找到他，让克里斯在场地上尝试开车的时候，克里斯甚至没有驾照。相信在观看了这部纪录片之后，没人会否认维斯塔潘对赛车的热爱。他脱下了红牛的飞行夹克，在另一条赛车的道路上，继续驰骋。发表首评[真人秀] 旧洞基朋友们关键词：真人秀又名：구기동 프렌즈 / Six Singles Under One Roof片长：90 分钟左右（单集）× 10 集；豆瓣链接一起同居吧。@潘誉晗：韩国最近又推出了一部真人秀《旧洞基朋友们》，找来了张度妍、张根硕、安宰贤、李多熙、景收真和崔丹尼尔六位年龄相近的艺人，过一次同居生活。随着明星观察类的项目在市场上越来越多，我相信有越来越多的人对这一类真人秀感到厌倦。而《旧基洞朋友们》的出现，再次证明了韩综的创意。看似依然是粉丝向地找了一群有粉丝基础的演艺人员，但是却找准到了一个很细腻的切口——让几个曾经独居过的明星们，过群居生活。独自生活，只是光看到就觉得很自由的日子，为什么他们体验过那种日子后，不愿意再独居了呢？我想这也就是观察类节目出现的初衷，去探讨一个社会性的议题。毕竟「存在即合理」。而这六名 85 后明星的同居日常，也非常有看点。他们都很有个性，但又能互相融入、靠近。见惯了大场面的他们，看到 18 公斤的章鱼，依然会手足无措，还是会大声尖叫。原来群居也不差嘛。桃花坞？年龄相近 不是想尽1更多[电影] 总统的蛋糕@利兹与青鸟：萨达姆的独裁统治下的伊拉克，平民们过得水深火热。小女孩拉米亚住在草编的窝棚里，连书包里的苹果都要被老师搜刮走。更不幸的是，她被抽到要在总统生日这天做好一个奶油蛋糕。于是拉米亚踏上了旅途，用家里所有值钱的东西换取做蛋糕所需的材料。电影在去年戛纳电影节获金摄影机奖，摄影优美，配乐充满阿拉伯风情，有触动人心的童真，也有伊拉克战乱下的可怖的倾轧。[电影] 平场之月@潘誉晗：电影根据朝仓加守美同名小说改编。青砥健这一天意外偶遇了自己的初中同学须藤叶子。此时两人已经 35 年没见了。中学时两人就互有好感，这次重逢，已离婚的青砥健和丧偶的叶子，终于还是走到了一起。埼玉小城里这一对中年人迟到多年的遗憾恋曲，就像是终于遇见了自己的另一个缺失的半圆，内敛、沉稳又有淡淡的愁绪。[韩剧] 浪漫的绝对值@潘誉晗：每当晚上，看似平凡的高中生吕意珠就会偷偷创作 BL 小说。可惜，她的作品一直

厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨主作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了联合作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了05/15 00:07Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。免责声明与作者所需披露的一切首先声明一下——本人仍然是在校的本科微电子硕士考古学的大学生，并没有经过任何的传媒学相关的学习，也没有从事相关行业的经验。目前仅有的和媒体沾边的事情是在酷安、少数派等平台发过一些文章（其中有很大一批都是在最后诉诸于，说好听点是真实生活场景，说难听点开始抒情的文章），以及一个 B 站粉丝量远没有破千的账号，因此我远远没有承担过养一个工作室的压力与责任，也没体会到在竞争激烈的环境中生存的艰难。因此以下内容不构成任何形式上的专业意见，只是一个愤青在 Random Yapping。多有冒犯之处，但我本心无意去批评任何具体的某一个媒体老师，只是想尝试提出一个也许还算有点价值的问题。以下内容提到「游戏区」，也仅限于单机游戏，不包含手机游戏、桌游等类型。以下内容提到「数码区」时，通常仅指以手机、一部分 PC 组件为主的主观体验向评测，且我的能力不足以涵盖 DIY 区1以及涉及到实验设计与数据分析的其他评测内容，虽然随着硬件厂商的创新进入后期演变为渐进创新，产品节奏也有所放缓，在社区中也能听到「DIY 区不如以前」的感慨，但本文尽力按照「不理解，没有切身观察就尽量少乱锐评」的原则展开，但撰写过程中有很多内容没收住，因而请让我提前声明一下——以下很多内容只是社区观察，不完全代表事实判断。最后，说句坦诚的实话，这个内容写的过程真的很痛苦——我真的不想让这篇文章变成直球辱骂媒体老师、攻击厂商或者是殴打观众的情绪发泄文，从个人发展角度来说不希望因此在进入之前就已经和某些职业无缘，也不想因此得罪那些因真心倾慕而认识的那些做这一行的朋友，且我真的不希望因为自己的过分言论把发文章的平台本身也拖下水，但因为笔力实在是有限，尽管拉着 Gemini、GPT、DeepSeek 等一众大模型核验，也找了一些朋友探讨，尽可能不要直球攻击特定的一两家媒体也不要说的太过分，但是字里行间还是有一股火药味拼尽全力消不掉。真的很抱歉，尽管我其实将自己也包括在本文的攻击射程以内——我自己就是后文中所批判的那个自认为掌握了真实，在没有厂商利益束缚与照顾社区的必要下肆意输出内容，甚至有很多是过当内容的小自媒体。这些内容因为是我自己一段时间内自我质疑与焦虑的外化，因此，本文必然存在一些因深度不足、角度有限而导致的难以避免的暴言，简而言之——就是年轻时犯下的错误。麻烦各位读者多多体谅。也欢迎您在评论区里多批评指正，让我多与您交流学习。以下内容如果有哪里说的不合适的地方，有什么可能涉及到的法律争议，请尽快联系我去撤换文章中对应的部分，谢谢！本文长度约为三万字左右，前三部分观点将会较为尖锐，会举几个相对较为敏感的案例来佐证观点，后两个部分则尝试解决前文抛出的问题，内容相对温和但字数也相对较长，请坐和放宽。主观口胡系列是一个尝试将动漫、电影、游戏叙事与更加严肃的讨论结合起来，去尝试探讨一些特定大命题的个人尝试，可能会一定程度上背离这些作品的创作初衷，聊的可能也会比较混乱且随意，甚至会有意提取个别元素，曲解原文本来聊一些其他领域的内容。这次我想把一个大约是三四年前就开始产生的想法去做进一步的深化，并尝试结合近些年来我对游戏区与数码区的一些粗浅且悲观的看法，来尝试去聊一个问题，一个也许注定没有答案，跨多少领域学科都难以得出解法的一个小小的问题——主观体验向评测的现状怎样，尽头到底为何方？又是什么构成了它们的专业性？Part1：太多水，太多奶酪——主观内容的梦魇虽然这么说有失偏颇，但我一直觉得游戏评测是一个相当主观的内容——评测者的个人阅历2、偏向的游戏类型甚至是组成一个游戏的系统，对 GDC 中诸多开发者分享的设计理念的理解，以及同期的同类型作品等等——这些都会对你最终得出什么样的结论，给游戏最终打什么分造成影响。所以在因近期的《神探杰克鼠》评测而陷入争议的 IGN 内部——作为一个有 TGA 评选权的，相对来说算是行业内位置靠上话语权比较大的游戏媒体——他们在测评游戏时给出的流程是：在内部筛选最适合某类游戏且当下仍有档期的编辑，由他们在内部其他编辑的帮助下完成对于这个游戏的评测，给出评分，在内部经过审核与排期，最终以文章 + 视频的形式先后放到公共平台上。但是，这个时候总会有个但是。虽然这样的流程确实在内部行之有效，为行业贡献了一批质量很高的评测，尤其是在神作频出神仙打架的那些年里——但同样不可否认的是，无论你之前做对过做好过多少次，一旦出现几次测评结果与大众预期不相符，甚至出现有失水准的内容，就很容易在玩家社群中丧失公信力，变成 IG8——《宝可梦欧米伽红宝石/阿尔法蓝宝石》的 too much water 扣分，《超级马里奥 3D 世界+ 狂怒世界》的 7.0，《死亡搁浅》的 6.8——你知道的，日光之下从来都没有什么新事。而且当我们把视角继续向下探时，会发现这样的问题不仅仅是发生在 IGN 一家，Polygon 的《DOOM(2016)》、VentureBeat 的《茶杯头》（他的操作已经不足以让他完成第一关，只能借其他同事存档继续评——尽管他据自己说已有着超过二三十年的评测历史）……所以到底发生了什么？当然，抛开社区中已有的被讨论过无数次的，对评测者的质疑不谈，我们也许可以从更加细致的视角去看这件事。就像你在做数码类评测会提到机型的性能与续航、外观、系统流畅度与体验、拍照、外放表现，但你不可能在非首发上手的评测里一样面面都要点到，而是根据个人的使用习惯有自己的偏好一样，IGN 的每个评测者在评测游戏时也会有自己的明显倾向，让我们以当年评测《死亡搁浅》的 Tristan Ogilvie 为例——他评测一款游戏的重心是动作系统、交互反馈与流程的顺畅程度。在我眼里，相较于小岛在前作《合金装备 5：幻痛》中用一系列主线支线任务，吸引玩家不断的从母基地回到开放世界，完成一串任务后撤离不同，本作让角色更加持久更加连续更加长时间的停留在这个世界里，将任务可视化，变为地图上要到达的目标点，并在沿路设置一系列需要潜行的敌人据点、风景奇观，并且可以通过基建加速到达目标点跳过一部分游戏内容来优化体验的开放世界而很不幸的是，本作正正好在她关注的几个角度精准的踩中了地雷：虽然相较于《合金装备 4》有所收敛，但游戏内不断插入冗长的过场 cg，以至于从室外场景进入到地图传送需要连续跳过数个动画；游戏流程放速过缓，且游戏内的任务设计被认为千篇一律，主线和支线均由送货构成；动作系统相较于前作乐趣有所减少（尽管二者的游戏定位有着根本性的不同，也许从来就不该这么要求角色的性能），击败 bt 后的奖励反馈也较为缺乏，以至于编辑出于他所偏向的几个维度的也许从传统游戏性的角度合乎情理的考虑，选择给这个游戏 6.8 分。随后我们都知道发生了什么——之后的几年让游戏的异步社交大放异彩，以一个非常奇怪且迅速的方式，让小岛秀夫这种几乎像是寻求争议，找骂、试图以先锋艺术电影的形式在游戏内进行的表达，能被更多人看见，被更多人共情，似乎小岛秀夫在《合金装备 2》后，再一次，等到了时代来验证他思想上的前瞻性。因而，时代背景变迁下大家对小岛独特表达的广泛接受、小岛秀夫的明星制作人光环以及评测中确实存在的一些过当的指责，都让玩家社群反过来开始质问一件事：为什么我觉得好玩的游戏，你却要给出这么低的分？当然，IGN 作为行业头部媒体，其本身在游戏界有较大的声量较高的影响力，且也有着个人做自媒体乃至中小型媒体难以企及的资源，因此它说的话做的评测理应被玩家更加严格的审视，树大招风不无道理——而且至少在我举得几个例子里，引发争议的游戏本身也是在不同领域内质量过硬的存在，只是随着营销的不断发展，疫情时代与后疫情时代每个亚文化小领域的的扩圈与自媒体时代的到来，这种原本对媒体对权威的的尚处合理范围的不信任将逐步扩大，直到这种现象一步步扩散到每一件事，每一个领域，乃至我们每个人都难以避嫌，难以无视。Part2：内容井喷的时代，同质化的幽灵与信任的坍塌数码圈从来不缺学生出身，单纯出于热爱，分享自己体验的自媒体，这样的评测形式本身并非没有价值，它用更加平易近人的角度去涵盖一个产品中用数据评测所难以涵盖的部分，尽管也许，分享自己对于刚入手的数码产品的主观体验并没有什么门槛——有着中专及以上的学历就有足够的语言组织能力去清晰讲述自己的感受，再来一个相机甚至是手机去拍摄素材就能实现，但撰写一篇好的主观体验评测文（不管是哪个领域）就像是一般人眼中的历史研究一样，都觉得准入门槛低，都觉得写的内容是任人装点任人阐释的小姑娘，但真的想要尽可能做到内容详实、条理清晰，往往需要长时间的写作经验积累与大量的训练，而如果想在这个基础上再去追求与社区中主要讨论的内容截然不同的全新角度，或者让我们思考的更加狂野一些，在这个内容上做出深度，探讨一个产品背后更加深刻的有广泛意义的命题，则更加考验个人的水平——但随着互联网的高速发展与信息的高速创作、流通，我们正处在一个内容极其过剩的时代。虽然我承认，我是随着酷市场转型为酷安而踏入数码圈的，本身并不算什么数码区老资历，但你如果要我去回忆在我进入数码圈后这有限的时间内，主观体验向的数码自媒体井喷的那段时间，我想大概是在「老师好我叫何同学」的 5G 视频兴起的 2019 年 6 月前后，以及疫情肆虐的 2021 年后到 2023 年末。前者因为正好切合了国家的 5G 发展战略的需要而被主流媒体报道，被社会广泛关注，因而给圈子引来了泼天的流量——事实上那段时间坚持做手机体验类评测的规模还比较小的自媒体，如今都已经成长为几十万粉丝的台柱，比如我当时看的比较多的花生说；而后者则是由于人们长时间居家，在室外的娱乐方式因客观需要而受限，因而人们开始转向互联网，大约是在这一时期前后，一群赛道更加细分，比如上古数码产品体验、软件交互等领域（就我当时看的比较多的来说是六分超超和苏星河）的体验向自媒体开始大规模兴起，游戏区的很多如今声量较大的自媒体编辑老师也多是在这一时期进入相关行业或者是投身做自己的账号。但与自媒体账号井喷相对应的，却是——请让我想想，该怎么让人更好接受的去讲一个有点不近人情的结论——极其迅速的内容同质化。游戏不是一个竞争激烈且产品间带来的体验差别相较于同类型的数码产品更大的存量市场，有足够多各自差别足够明显的新游戏不断冲刷着视野，因此这种同质化的倾向也许还并不明显，但在数码区，当每个上游的厂商都在绞尽脑汁的结合热门概念，生生在软硬件体验上捏出极其有限的新创新点时，这种感觉就尤其明显——对于自媒体而言，相比起「你自己要讲述什么故事」，你的阐述对其他人有什么价值才更有传播的意义，因而当你迅速的因不受束缚、观点尖锐，获得了一批受众支持迅速爆火后，同质化的幽灵便也开始悄然显现——是顺着原先让你得到人们簇拥的路径一路向西，对着即使没有那么差的内容输出情绪，直到有一刻在无数次路径依赖后判断失误，犯了众怒，还是妥协于平台激励有限而选择靠近厂商，向其他市面已有的无数自媒体靠拢，在厂商给的通稿框架与宣传的大方向已然确定的情况下，进行着极其容易趋同的表达，在首发宣发节奏与内容合作模式影响下，选择性的淡化掉体验中的痛点，即使在前文有提及也往往占据相对较小的篇幅，并随后立马引入一个「但是」，用着首发特调的机型说着好到梦幻的体验，最终在观众的不解中继续在内容理想与工作室运营中艰难的走着钢丝？当然，如果把话说到太绝，往往会有很多偏颇，所以让我们稍微说一些让文章作者更拟人的

周一18派早报：提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年昨天 06:56早报速览提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年0OpenAI 与苹果关系紧张，或考虑诉讼5索尼回应 AI 相机助手越改越差，称只是修图建议2Anthropic 警告私下交易的股权可能无效2运营商回应「零月租」传言4微软重新允许自定义任务栏与开始菜单2看看就行的简讯4少数派的近期动态0你可能错过的好文章0提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年5 月 15 日，预印本平台 arXiv 出台新规：如研究人员向该平台提交的稿件包含「无可辩驳」的 AI 生成内容，论文所有署名作者将被处以长达一年的封禁；且其未来提交的任何文章都必须先在其他期刊通过正式的同行评审后，arXiv 才会予以收录。「无可辩驳」的证据包括虚假参考文献、不小心粘入的与 AI 对话内容等。俄勒冈州立大学名誉教授、arXiv 编委兼审核团队成员 Thomas Dietterich 在 X 上公布了这一消息。他指出，新规直接源于 arXiv 既有的审核标准。无论稿件中出现不当言论、抄袭、学术偏见、虚假参考文献还是误导性内容，所有署名作者都必须对这些 AI 生成的错误承担全责，而不能将责任推卸给工具。arXiv 是物理学、天文学、数学及计算机科学等领域全球最大的开源学术文献库。在这些学科的常规发表流程中，学者们通常会在接受期刊正式的同行评审前，先将论文作为预印本上传至 arXiv，以抢占首发权并获取同行反馈。因此，剥夺一年的首发资格并永久附加同行评审门槛，对相关领域人员而言是严厉的措施。354OpenAI 与苹果关系紧张，或考虑诉讼据彭博社报道，苹果与 OpenAI 为期两年的合作关系近期出现裂痕。因未能从 ChatGPT 与苹果系统的整合中获得预期收益，OpenAI 正与外部律所积极探讨应对方案，近期或将向苹果正式发出违约通知，为潜在的法律诉讼做准备。2024 年 6 月，苹果宣布将 ChatGPT 接入 Siri 等系统级功能。OpenAI 曾寄望借此获取海量订阅用户及数十亿美元收入。然而实际体验中，该功能入口较深，需用户明确口头提及 ChatGPT 才能唤醒，且信息呈现方式受限。结果，绝大多数用户仍倾向于使用 ChatGPT 应用。一位未具名 OpenAI 高管表示，苹果在产品端甚至没有尽到诚实努力的义务；合作不仅未能创收，反而损害了其品牌形象。除软件合作遇冷外，两家公司在硬件与人才领域的博弈也加剧了紧张局势。OpenAI 此前收购了由前苹果首席设计师 Jony Ive 等人主导的初创公司，并利用高薪从苹果硬件团队挖角，试图打造新一代 AI 设备，这引发了苹果高层的强烈不满。同时，苹果也已着手引入竞争机制，计划在即将发布的 iOS 27 系统中全面开放生态，接入谷歌 Gemini 和 Anthropic 的 Claude 等其他主流大模型。知情人士表示，OpenAI 针对苹果的任何法律行动，预计都会推迟到其与马斯克的诉讼案结束后才会推进。目前各项计划尚未做出最终决定，OpenAI 内部仍保留了在庭外解决争端的希望。OpenAI 与苹果的发言人均拒绝就此事置评。我换到国产手机AI都没怎么用过，不是不会用，而是习惯了以前苹果装个AI软件的使用方法了。果子在ai上的每步操作都很迷惑感觉不如自己装一个本地小型LLM App用ShortCut打通生态。我还是关心什么时候有苹果智能，感觉一不留神，好几年过去了，我换15p的初衷之一就是给智能做准备，现在都要17了，我也是吃上大饼了622索尼回应 AI 相机助手越改越差，称只是修图建议近日，索尼最新旗舰 Xperia 1 VIII 的官网页面引起关注。在演示 AI 相机助手的效果时，官方样张的效果在观感上反而明显劣于原图。这引来了 MKBHD、Nothing 创始人裴宇等人的公开调侃。对此，索尼在 5 月 15 日解释称，「AI相机助手」并不会在拍摄后自动应用修图，只是根据当前场景和拍摄主体，提供 4 种不同创意方向的参数设置建议，用户可自由选择接受这些建议或坚持使用自定义设置。与此同时，官方最新展示的对比样张在视觉效果上有了显著改善，去除了首批样张中饱受诟病的画面发灰和过度曝光问题。但这依然未能完全消除外界对其影像审美判断的疑问，也不清楚营销团队最初为何会挑选那些效果不佳的照片来展示核心卖点。感觉那朵花还可以，其他的都是💩️有人关注就算不错了10111Anthropic 警告私下交易的股权可能无效据《华尔街日报》报道，Anthropic 近日更新官网公告，宣布将不承认未经其董事会批准的二级市场股权交易，并点名了一系列未授权机构。Anthropic 警告称，由这些机构提供的任何该公司股权或相关权益转让均属无效，且不会记录在公司的股东名册上。这一声明在个人投资者中引发了广泛恐慌。由于普通投资者通常无法直接购买未上市「明星公司」的股权，大量二级市场交易平台应运而生。它们设立特殊目的实体（SPV），将众多散户资金打包后投资于目标公司。然而，这种 SPV 结构日益复杂，有的多层嵌套，有的把多家目标公司的股权捆绑出售，导致分配机制存在很大模糊地带。普遍认为，Anthropic 此举意在通过清理违规交易来精简其股东名册，为其计划于今年晚些时候的 IPO 做准备。随着近年来私营科技公司的估值不断飙升，包括 OpenAI 在内的多家独角兽企业，都在采取类似措施限制涉及其股权的二级市场交易。面对投资者的质询，多家被点名的交易平台迅速回应，表示涉及 Anthropic 的基金和交易均已获得批准或符合转让要求。专家指出，如果 Anthropic 并未在事前对特定交易结构做出限制，可能无权撤销这些持仓，相关争议最终或将诉诸法庭。可能性不大这种模式在中国有么？发表评论运营商回应「零月租」传言据央广总台报道，近日，不少互联网博主在短视频里介绍，三大运营商即将取消月租费，改为按量计费模式，甚至还列出了推出该政策明确的时间节点。但记者梳理市场现状发现，目前只有中国联通上线了零月租、设最低消费门槛的「魔方」按量套餐；中国移动和中国电信暂未推出同类产品。例如，传言中被广泛提及的「联通魔方」发布于今年 4 月底。该产品采用纯按量计价模式，国内语音和短彩信均为 0.1 元/分钟或每条，流量则实行阶梯计价，单价在 1 元至 1.5 元/GB 之间。但联通客服表示，该套餐并非绝对免费，而是设有地区差异化的最低消费门槛，例如北京地区起步低消为 55 元，月结时不足该金额将自动扣齐。同时，联通原有的各档位固定月租套餐依然保留并可正常办理。中国移动和中国电信的客服人员则都表示，公司曾推出过零月租模式，但目前没有相关计划。针对完全零月租的可能性，通信行业专家指出，此类模式脱离了基础通信网络的运营逻辑，庞大的固定网络维护成本最终也必然会转嫁到其他用户身上，甚至可能导致整体资费上涨。比起追求不切实际的零月租噱头，当前消费者对资费的真正痛点在于不合理的老旧月租套餐，以及套餐外的标准流量资费。没有专门办的流量卡划算捉虫：“央广总台”→“总台央广”“零月租” “起步低消”世界需要提速降费。3微软重新允许自定义任务栏与开始菜单5 月 16 日，微软发布博客文章，宣布将在未来几个月内推出一系列更新，大幅增强任务栏和开始菜单的外观自定义功能。这标志着一系列经典 Windows 功能的回归。对此，Windows 业务负责人 Pavan Davuluri表示，个性化定制是 Windows 的基因；微软已听取用户反馈，决定重新将界面控制权交还给用户。目前，部分选项已在 Windows 预览体验计划中面向测试人员推送。本次更新中，用户可以再次像在旧版系统中那样，将任务栏停靠在屏幕左侧，或将其缩窄以节省屏幕空间。同时，开始菜单新增了尺寸调整选项，并允许用户自定义显示模块，例如直接关闭「所有应用」或「推荐」（更名为「最近使用」），仅保留自己固定的应用。此外，用户可以隐藏「开始」菜单中的用户名和头像。这些测试功能目前仍有已知问题，例如开始按钮未对齐、通知弹出位置异常等。微软表示将在正式向公众推送前修复这些问题。此外，更多传统的自定义功能，例如拖拽任务栏边缘手动调整高度等，目前也已纳入开发计划，预计将在今年内陆续上线。Windows 11 自发布以来，就因移除任务栏位置调整等诸多经典自定义功能而饱受争议。近年来，微软一度将系统开发重心全面转向 AI，在系统中强行植入 Copilot 功能，引发了用户的强烈抵触。迫于压力，微软于今年早些时候宣布做出调整，重新关注质量。什么时候把win10菜单给我回归一下😊我选择Windows Server 2025，躲过UWP和React两大石山。1看看就行的简讯据Fast Company报道，知名开源密码管理器 Bitwarden 近期在未发公开声明的情况下，悄然大换血高管团队，今年 2 月起，原 CEO、CFO 先后转岗或离职。新任 CEO 有并购及私募股权交易背景，让外界对其未来的资本运作产生猜测。此外，Bitwarden 官网出现多处调整，个人版产品页面移除了「永久免费」的标签（官方解释为营销团队疏忽）；不久前，Bitwarden 在未显著通知的情况下，将高级版订阅年费从 10 美元翻倍至 20 美元。此外，Bitwarden 还在招聘页面修改了企业核心价值观，去除了原有的「包容」与「透明」两项原则。近日，谷歌开始测试一项新的云存储空间政策：新注册的用户如果不绑定手机号，其初始免费存储空间将从传统的 15GB 缩水至 5GB。谷歌官确认，该政策目前仅处于区域性测试阶段，且现有老用户不受此规则影响。谷歌发言人表示，这项新规是为了继续为用户提供高质量的存储服务，同时鼓励用户完善信息以提升账户安全性和数据恢复能力。然而，许多用户推测谷歌此举的实际目的是为了打击恶意注册大量小号以薅取免费存储空间的行为；也有部分用户对强制绑定手机号背后的隐私数据收集及潜在商业用途表示担忧。据9to5Google报道，谷歌随无屏健身追踪器 Fitbit Air 推出的订阅服务「健康教练」（Health Coach，由原 Fitbit Premium 升级而来）存在 AI 幻觉问题。测试中，AI 教练凭空捏造了一项 5.2 英里的跑步记录；在被指出错误时，还反问用户是否忘记输入过该次记录。经过纠正后，这条虚假数据依然残留在应用主页上。最近对bitwarden挺失望，要么是性能问题，这次又出个管理团队问题。开源不是万灵药，信誉才是基石啊这几年一直用Bitwarden 呢，要变味了吗白赚一次跑步，哈哈哈哈比特瓦恩德可以换了(｡･ˇ_ˇ･｡:)4少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购发表首评你可能错过的好文章看剧不卡顿、游戏低延迟：「智能队列」这样改善你的冲浪体验为什么我觉得好的东西，媒体评测却给了低分？本周看什么 | 最近值得一看的 8 部作品装了啥：编辑部用什么转写音视频内容？[会员免费] Apple WWDC26 全球开发者大会，来与少数派一起看快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验你真的需要墨水屏设备吗？发表首评90扫码分享目录 0讨论我来说一句发布发表评论发布

2025 年度旅行——4 天 3 晚青海湖小环线主作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥联合作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥05/14 09:38Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。一次完美的旅行，对我而言，不仅仅是去被人走过千遍的地方打卡一圈，而是藏着三次截然不同的体验——出发前对目的地的想象与期待，旅途中与自然的亲密接触，以及归来后翻看照片时，重新发现那些当时未曾留意的细节，细细回味，查阅资料，让我对这次旅行有更深刻的理解。那么就让我们随着这些文字与照片，回到美丽的青海湖。契机：作为刚工作的年轻人，在国内的职场环境下，真正无理由可申请的假期只有 5 天，所以我和身边的小伙伴们都分外珍惜这五天的时光，每年的年度旅行我们都会避开人流，详细规划。由于十一假期要去参加两场婚礼，为了节约交通费，我们决定先开车到 2 号参加婚礼的宝鸡，把小 P（我的小鹏 MONA M03，具体可参阅我之前的文章）停在了宝鸡高铁站，考虑旅行时间较为紧凑，且西北自然景区交通补能条件可能较差，我们选择了到当地租车完成青海湖小环线行程。带着宝鸡擀面皮和肉夹馍，一行人正式开始了旅途，三个半小时后，到达了西宁火车站，初入高海拔地区，并未感受到明显不适，方便的是，租车平台在西宁站地下停车场就有取还车点，我们直接将行李放上车，省去了搬来搬去的劳顿。Day1 西宁-茶卡 全程约 457 km西宁-塔尔寺5A 级旅游景区，门票：70 元，游览时间约 2 小时塔尔寺是藏传佛教格鲁派的著名寺院之一，始建于明代，最初是为纪念藏传佛教格鲁派创始人宗喀巴大师而修建，据传宗喀巴大师出生于此地，后人于其诞生处建塔纪念，塔尔寺因此得名——「塔尔」即「有塔之地」。塔尔寺-八宝如意塔游览中心和寺庙门口都有大量的讲解人员，价格也不贵，但是本人信仰唯物主义，故只将这里当做名胜古迹参观，故没有请讲解，如果对藏传佛教及寺庙背后的故事比较感兴趣，请个讲解或许是更好的游览方式，或者是像我一样，回来翻阅照片时再仔细搜索相关资料。Tips：讲解会介绍各个佛像的「作用」（不知道这样叫是否合适，或者说负责保佑的事情），故可能会被架在那里为了讨好彩头而「必须」捐献资金，介意者可以像我一样自由游览。由于是节前，故参观人员并不很多，与前往朝拜的信徒或僧人可以说是一半一半，没有了熙熙攘攘的人群，也得以让我更加细致地观察这座寺庙和「磕长头」远道而来的人们。前来朝拜的僧人进入寺庙，浓郁的香火味和木质结构因缺乏通风的陈味扑面而来，看着宏大的佛像、墙上摆放整齐的诵经坐垫，的确能感受到不一样的文化氛围，这里安静、鲜活，尤其是看着不知道从哪里启程的「磕长头」的信众，虽因路途的遥远衣衫褴褛，但想必也经过这一路的艰辛而悟出了些许道理，可能出发时也不知道为什么，但走着走着就明白了吧。塔尔寺中的白塔募捐处，与时俱进地，也支持扫码支付塔尔寺有大大小小的寺庙很多间，在我看来有的都别无二致，于是重点转了几个后，就启程回西宁吃饭，整备氧气瓶，应急食物等装备，正式奔赴青海湖。海南藏族自治州共和县 南山观景台 & 青海湖二郎剑景区前者免费，后者为 5A 级旅游景区 门票 72 元，游览时间约 2 小时一点半在西宁市吃完炕锅后，我们开始驱车前往青海湖，约 2 个小时车程，我们到达了本地的主要目的地青海湖的南侧景区 青海湖二郎剑，在进入二郎剑景区前，可以导航到「南山观景台」 从山上远眺青海湖和远处群山的全貌，但是山路较窄且有工程车辆时不时通过，故要格外注意安全。景区本身没有人文景观，但其还是我国早期鱼雷的试验基地，但是因为我们到达时间比较晚，并没有参观博物馆，而是直奔到湖边。青海湖是我国水面面积最大的咸水湖，水面平静辽阔，与远处的群山相辅相成，甚是开阔美丽，栏杆上有硕大干净的鸽子驻足，在几番伦理拍照后，我靠在栏杆旁，看着远处的夕阳洒在地面，耳边传来阵阵鸟鸣与风声，工作与生活的喧嚣远离脑后，只剩下全身心的与自然接触。云层、远山、草原，形成了震撼的三明治拨云见日，青海湖边的草还是绿色的岸边排队等待「薯条」投喂的海鸥码头景观随着夜幕降临，我们离开二郎剑景区，沿着青海湖开车前往茶卡住宿，虽然只有 200 多公里的行程，但是因为沿途工程车辆很多，还有很多地方持续修路，有的路段有连续的 U 型弯，还有双向仅有一条车道通行的地方，还要翻越几个海拔 3800 米的垭口，成了这一路最惊险刺激的行程，一路上小心翼翼，战战兢兢，终于在晚上 10 点到达了茶卡镇，幸运的是镇上的餐馆都还在营业，找了家当地美食，简单吃饭后休息。夜幕降临的倒湖茶公路Day2 茶卡-祁连县 全程约 372km茶卡盐湖4A 景区 带往返小火车套票 167 元「茶卡」在藏语中意为「盐池」，该地以出产「大青盐」而闻名，参观盐湖也是此地的知名旅游项目。茶卡盐湖的景区共有两个，天空之镜是较老的景区，主要以自然风光为主，天空一号主要以商业化项目为主，作为外地游客，我们肯定更希望看到原始的自然景观，于是选择前往天空之境景区。景区很大，如果想要游览每个景点建议留出一整天的时间，游览顺序建议先坐小火车到终点天空之镜，该站处于盐湖的最深处，盐水异常的干净，也最为出片，景区也放置了很多拍照打卡点位，可以尽情拍摄，心痛没有带无人机。在晴天的时候，由于天空和地面都是白色，很容易晃眼，推荐大家一定要戴墨镜前往，通过墨镜观察湖面，盐湖会看起来像翡翠一样更加绿色透明。套票包含鞋套租赁，可以套上下湖走走，踩在大颗的盐里,很像沙滩的感觉，景区维护的十分干净，可以尽情的亲近自然。有趣的是，景区为了怕大家在盐滩大片的白色中无聊，设置了很多拍照用的摆件，例如国家地理的相框等等，还在大屏幕上贴心的轮播拍照指导，期望每个游客都能满意而归。盐湖与远处的山峰盐湖码头废弃的采盐轨道之前采盐的轨道深入盐湖，在墨镜下，真实的观感颜色会比照片深很多。接驳小火车行驶在盐湖中间，干净得不像国内巨幅盐雕从盐湖出来，我们在县城稍作休息，开始启程前往祁连县，有了昨夜惊险的行程后，尽管沿途的国道可能风景更好，但是我们选择了高速较多的线路。这一程的路况还是很好的，就算是国道车辆也较少，道路笔直，视野十分宽阔，草原、远山、白雪、天空层层分明。在行驶过程中，不时的有大群的羊羔和牦牛穿过公路，不必惊慌，一会就有牧民骑着摩托驱赶，让它们加快穿越的脚步。公路与群山草原上大批的牛羊傍晚的草原，远处能看到夕阳洒在雪山上落日下的牦牛剪影在夜色中，我们抵达了祁连县，在县城中吃过一顿特色烧烤后，进入梦乡，准备迎来第三天的行程。Day3 祁连县-岗什卡雪山-西宁 全程约 320km祁连卓尔山景区门票 60 元 上山小巴 20 元 登山电梯 10 元这个景区的确是全程我最惊艳的景区了，近能看到瑞士小镇似的田园风光，中能看到美如画的丹霞地貌，远还能远眺远处的阿咪东索雪山（海拔 4695 米），而且在山顶可以拍跟雪山的合照，大气、干净、巍峨。自驾车辆可以停在景区停车场，然后乘坐接驳中巴约 10 分钟到达检票口。卓尔山的两侧，风景差别迥异，一侧是连绵的雪山，另一侧是仿佛瑞士小镇的丘陵和小镇。阿咪东索雪山卓尔山的另一侧，仿佛欧洲小镇丹霞地貌与草地交融最让我惊讶的是这个景区竟然还修建了方便爬山的观光电梯，10 元一位，大概有三段，果然是人性化啊。景区电梯中午约 12 点，我们离开景区，在祁连县的一家川菜馆简单用餐后，我们驱车开始前往本次旅途的最高海拔岗什卡雪山。在县城开车的过程中，等红灯时，我看到一位初中年龄的藏族小伙盯着我们开的岚图梦想家，眼中闪闪发光，不禁让我想起小时候的自己，看到喜欢或没见过的车也会驻足观看，充满憧憬。愿他长大后也能开上自己喜欢的车，能够做自己喜欢的事情。岗什卡雪峰景区免费 上山摆渡车 25 元岗什卡雪山坐落于青海海北州门源县境内，是祁连山脉东段的最高峰，被誉为「雪山之尊」，因为设施和难度都尚可，也被雪山爱好者们称为「人生第一座雪山」。主峰海拔 5254.5 米。游览模式同样为将车辆停在景区停车场，然后乘坐接驳车到雪山脚下，正式游览区域的海拔在 3800 米左右，建议随身携带一个氧气瓶，下车之后，可以看到湛蓝的连心湖，水来自于冰川融水，干净，清冽，与子洲植被的荒凉的深褐色形成鲜明的对比，十分美丽。连心湖到达时天气并不是很好，岗什卡雪峰被薄雾笼罩天然泉水，因还有行程，害怕拉肚子，没能如愿品尝一下离开景区后，雪山逐渐露出真面目，攀登还是留给下次吧沿着冰川的遗迹向上走，就能看到岗什卡雪峰徒步攀登的起点，景区也提供马匹，可以跟着向导继续向前，到七彩瀑布，因时间和个人体能的考量，我们浅尝辄止，在河道附近游荡些许时间后，便下山返程。临近傍晚，启程返回西宁市，一顿丰盛的海底捞过后，快乐的青海湖小环线结束。Day4 西宁市-青海藏文化博物馆-特产采购最后一天，伴随着蒙蒙细雨，我们打车前往青海藏文化博物馆，博物馆不大，纵使是十一期间人也不是很多，但内容着实丰富，满墙的唐卡和在电影里看到的藏传佛教僧人服饰，还有精致的贵族金器。在博物馆的文创区，我购入了一个小的转经筒，在回程的路上一直把玩，虽不信奉佛教，但也觉得甚是有趣。唐卡四王金杯嵌绿松石立凤金头饰博物馆文创-转经筒，38 元虽不算划算，但做工十分精致离开之前，为了给家人和朋友带些纪念品，我们选了一家超市类型的纪念品商店，购买了牛肉干，藏红花，枸杞，青稞酒等特产，虽说也不一定划算，但胜在明码标价，将行李箱装满后，顺利返程。后记作为 2025 年的年度旅行，本次 4 天的青海湖小环线我甚是满意。人少不堵车，行程不特种兵，在当地感受美食和壮美的自然风光，一路上与好友们分担开车，聊对景色的感受，互相拍照，点评食物，虽偶有几个工作电话烦扰，但总体悠然自在，放空身心，一切都是那么的安静，与自然深度的接触，与北京的喧嚣形成了鲜明的对比。对于出行方式，自驾无疑是最有的选择，这边深入自然，道路虽崎岖，但也属于铺装路面，且西宁市神州、一嗨等在高铁站都有网点，取送车方便，且自驾可以自由控制行程，根据游览时长方便自己调整。对于住宿，我们对酒店的干净程度略有要求，于是全程都选择了以华住会为目标的连锁快捷酒店，以下是所有天的酒店名称和十一节前的价格，个人觉得是十分舒适的。Day0 桔子西宁火车站省人民医院酒店 双床房 331元Day1 星程茶卡盐湖酒店 覆氧双床房 314 元Day2 宜必思尚品海北祁连卓尔山酒店 421 元Day3 桔子西宁火车站省人民医院酒店 双床房 412 元关于本次行程的花销，纵使算上从北京出发的路费，人均也仅 3k 左右，可以说是十分划算了，主要的开销就是租车和酒店。支出一览表关于卫生间，景区和路途服务器的卫生间都比较完善，因同行有女生，我们还准备了移动卫生间，但从我们的行程安排走下来，并没有用到，个人觉得要是按我们的行程游览也无需担心。下一站，出发！2025 年的年度旅行结束，2026 年也可以开始规划起来了，新疆还是海南，希望能远离喧嚣，享受美丽壮阔的祖国风光。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀1610扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@北鸮© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#Matrix首页推荐#旅行16等 16 人为本文章充电扫码分享举报本文章别惹小炸毛一个学酥关注

打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试主作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.联合作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.昨天 15:00ARG 即侵入现实游戏（Alternate Reality Game），是一种基于现实世界，融合各种多媒体（如网站、电话、社交媒体等）技术的互动式解谜游戏。由于其与现实的高度结合所带来的沉浸感，一直以来被许多玩家所追捧。如果你在 2016 年玩过《守望先锋》，相信你对当时新英雄「黑影」的登场一定记忆犹新。从 2016 年夏季开始，一直到 11 月的暴雪嘉年华，暴雪通过在宣传视频、官网、社交媒体发布了一些隐藏着密码的内容，由全球玩家共同解谜，不但揭示了游戏世界观下的一些故事，更是顺理成章地将身为黑客的「黑影」带到大家的视野里。经过几个月的发酵，在暴雪嘉年华上大屏幕被「黑影」控制，随后开始播放英雄动画的时候，更是满足了玩家们期待已久的心。「黑影」的宣发是一场相当成功的 ARG，时至今日仍有玩家对其念念不忘。我本人也是 ARG 的爱好者，我非常享受这种打破虚拟与现实边界的游戏乐趣。前段时间，一款名为《大学生登山失踪事件》的网页解谜游戏突然火了起来。比起此前一些高度依赖技术知识的网页解谜游戏，《大学生登山失踪事件》在不需要任何 IT 知识背景的情况下，依然有着抽丝剥茧逐渐接近真相的推理爽快感。自此之后，小红书上越来越多的人开始制作这一类网页解谜游戏，并为它们赋予了 ARG 的名号。这些游戏流程往往不长，差不多 1-3 小时就能通关。因为是网页解谜游戏，只需要一个浏览器就能游玩，很适合上班时间摸鱼，游玩门槛非常低。这篇文章将会推荐一些这段时间我玩过的质量不错的游戏，虽然一些制作精妙的游戏确实带有 ARG 的影子，但其本质上仍是基于网页的解谜游戏。为了方便大家后续搜索，本文会将这些游戏统称为 ARG。大学生登山失踪事件🔗https://missing.shiroki-y.top/introduction.html在游戏中，玩家需要通过从弟弟衣物中发现的「福安菜家常馆」的线索为开端，逐渐深入一场黑色阴谋找到弟弟失踪的真相。游戏整体难度不高，但在游玩过程中最好能用纸笔或软件记录进度，这能够帮你清晰地理出游戏脉络。游戏的核心玩法其实非常简单，只需要阅读页面，找到页面中可疑的关键词，在页面搜索栏里进行搜索即可推进剧情，这其中也会有一些用户名、密码、口令、字谜等需要玩家进行推理和猜测。网页页脚处会标记当前页面进度，既可以方便定位，也能够明确线索的先后顺序。剧情方面，《大学生登山失踪事件》并不会在一上来就给玩家塞一些晦涩难懂的概念，而是从大家熟悉的普通网页开始。当玩家一步步抽丝剥茧逐渐发现事件背后的黑暗面时，令人不由得感叹于游戏的设计与巧思。同时，游戏内也藏有一些可供探寻的 ex 页面作为彩蛋，在解谜之外找到各种彩蛋也不失为一种乐趣。诡异广告调查事件🔗https://mp.weixin.qq.com/s/yeNYMZVRp6N-MESFJGk9rg这是一款与《大学生登山失踪事件》来自同一作者@銀河電燈_的游戏，世界观也与前作有关联之处，很是有趣。玩家将扮演一位被「优化」的人，因为在看剧时发现了某个已停播的节目广告，从而开始调查这个诡异的广告，并发现藏在节目与自己背后的神秘真相。与前作相似，这款游戏也将基于阅读与搜索推进剧情。在此基础上，它还加入了一些 ARG 元素，比如在街景地图中，找到现实里存在的文字招牌作为线索。并且在网页设计上也比前作更加成熟，贴合剧情上的时间设定，在网页上加入了许多彩蛋。剧情方面，本作有一定的惊悚元素。特别是在游戏过程中展示的一些内容可能会引起不适感，如果你对这类内容比较敏感，请务必提前做好心理准备。当然，游戏并没有 Jump Scare，更多是通过剧情与视觉设计所营造出一种的诡异氛围。三川米粉店🔗https://www.berrysignalscmfd.top/《三川米粉店》也是一款从餐饮店主页出发，探寻故事真相的游戏。玩家需要调查三川米粉店突然停业的原因，在调查过程中，玩家将逐步揭开这家米粉店背后的故事。本作的页面设计相对没有那么精致，但也融入了一些 ARG 的元素，部分谜题的答案并不会直接给出，而是需要玩家通过搜索引擎或其他工具进行查询才能获得答案。从剧情上来说，《三川米粉店》的设定会更加接地气一点，并且结局也会相对更温馨一点。比较适合对于惊悚元素不太感冒的玩家作为第一部作品来尝试游玩。永安温泉度假酒店🔗https://mp.weixin.qq.com/s/kKsnzFsLFwwm2nwv8DseAA与前面侧重解谜的作品不同，《永安温泉度假酒店》的解谜难度其实不高，反而更像是一种「规则怪谈」式的探索路径。玩家需要不断探索找到更高权限的员工账号密码，从而揭开温泉度假酒店背后的秘密。游戏在开始前需要玩家填写「报名信息」，在后续游戏中能够增强代入感。从游戏性的角度来看，这款作品的「解谜」爽快感或许没有那么强。但如果你对「规则怪谈」或者一些高概念的设定感兴趣，那么这款游戏依然是一个值得尝试的作品。404 工位消失事件🔗https://xsn1204.github.io/qmgame《404 工位消失事件》是一个特色鲜明的游戏，玩家将扮演一个实习生，接手其他人留下来的工作电脑，并从中发现一些未被处理的秘密。游戏在网页上复刻了一个形似 macOS 的「系统」，很大程度上增强了代入感。因为设定是在电脑系统，游戏中需要玩家在钉钉、微信等各种软件中挖掘线索，反而更有一种作为侦探在不断探索的感觉。但受限于网站形式和游戏体量，游戏中能够互动的地方通常都是为剧情线索服务的，与线索无关的地方大多都是不能点击互动的。而且游戏不能通过点击窗口左上角的关闭按钮回到上一层，而是需要通过浏览器的返回来实现。但这些小细节对于《404 工位消失事件》来说都算瑕不掩瑜，整体游玩体验还是值得肯定的。作者甚至还出了 2.0 版本，新的剧本进行了优化提升，如果对这一类型感兴趣的话也可以尝试游玩。找到我🔗https://www.kaixinmoyu.com/feijo/findfeijo/在游戏中，玩家需要扮演一位私家侦探，调查游戏作者失踪的原因。游戏在交互方面做得非常有意思，游戏不仅做了浏览器页面对线索进行搜集检索，还做了小灵通和笔记本，前者可以通过拨打特定的电话获取线索，后者则需要玩家阅读日记找到隐藏的内容。《找到我》本身是一个更重交互而非推理的游戏，游戏中融入了一些 ARG 元素，在最开始就需要到某篇公众号文章中找到一个密码才能继续推进。但游戏的剧本逻辑会稍微有些不够连贯，以至于中间可能会有一些容易卡关的地方，需要仔细翻阅此前出现的线索才能继续往下推进。另外，游戏中后期会有一道需要在现实中解算的谜题，这道题本身与游戏本身的节奏设计会稍显割裂。但整体来说，《找到我》依然是值得一玩的游戏，尤其是本身的交互设计和 ARG 元素的融合做得非常不错。在游戏作者的个人主页，也能找到作者开发的更多网页解谜游戏，如果你对这款游戏的风格感兴趣，也可以尝试其他作品。小结优秀的网页解谜游戏并不止文章提到的这些，光是在小红书上一搜，就能找到许多分享帖。希望本文能起到抛砖引玉的作用，欢迎大家在评论区分享自己玩过的，令你印象深刻的网页解谜游戏。> 简单、好用、专注的写作软件，少数派为你呈现 🚀> 守护孩子的天马行空，从零开始美术教育启蒙🎨404扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Lotta© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#解谜游戏#游戏#生活方式40等 40 人为本文章充电扫码分享举报本文章TpBe nothing, and everything.关注

派评 | 近期值得关注的 App主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注Vanilla少数派作者Vanilla关注Vanilla少数派作者关注化学心情下2少数派作者化学心情下2关注化学心情下2少数派作者关注大大大K少数派作者大大大K关注大大大K少数派作者关注Peggy_少数派作者Peggy_关注Peggy_少数派作者关注ElijahLee少数派作者ElijahLee关注ElijahLee少数派作者关注Snow少数派作者Snow关注Snow少数派作者少数派编辑部 等 7 位作者昨天 18:05欢迎收看本期《派评》。你可以通过文章目录快速跳转到你感兴趣的内容。如果发现了其它感兴趣的 App 或者关注的话题，也欢迎在评论区和我们讨论。AIDA：一站式日历和待办规划工具平台：iOS、iPadOS关键词：日程管理、待办清单@ElijahLee：Aida 是一款针对 iPhone 和 iPad 开发的日历和待办清单应用，来自任务管理工具 Sorted 的前创始人之一 Harry，它集成了待办清单、日历和时间规划工具，在一处显示、规划、操作。应用主界面就是每日时间线，顶部是当周 7 天的简要信息，应用中部大篇幅排列当天的时间块，你可以直接看到上午十点看电影、中午写派评、傍晚健身等日程，以及晚上记得给植物浇水的待办提醒，而不用在多个不同的 App 之间切换。应用的快速调整功能，支持在时间线中将事件按住并移动顺序，快速重排一天的事件，对应的时间点会自动修改，非常方便。使用右下角添加按钮，可以自由地为时间线加入行动或事件。行动即待办事项，支持输入文本并设置提醒时间。 Aida 创造性地设计了拖动设置时间的功能，将底部的柱状图左右滑动，就可以调整行动的提醒时间，每次拨动以 15 分钟为间隔，用它设置时间非常方便。当然我们也可以点击时间来精细设置时分。 Aida 事件就是日历中的事件，必须包含开始时间和结束时间，我们还可以选择日历类别、设置全天事件等。Aida 引入了 GTD 管理的项目功能，提供收件箱、任何时间、将来某日等默认分类。自定义项目可以将同样类别的活动，包括行动和事件进行集中收纳整理。Aida 全面集成了 iOS 日历，在应用内不仅可以查看 iOS 日历事件，对事件的修改调整还会实时同步。不过 Aida 目前没有接入 iOS 的提醒事项。 批量操作功能可以一次选择多个行动进行完成、改期、移动或复制，也能快速整体调整时间块。你可以在App Store免费下载 Aida，免费版功能受到很大的限制，仅支持 1 个日历的集成显示、无法调整行动的时间、添加行动和事件的数量有限。付费可以解锁高级功能，包括无限项目和时间块、多个日历集成显示等功能。费用为 58 元 / 月、298 元 / 年，698 元买断，应用提供 1 周免费试用。“任务管理工具 Sorted 的前创始人之一 ”这个介绍，在我派这里，会起到一定的负面影响。698买断，也会引起大家的不美好的回忆，298的年度订阅则没有直面当前的经济形势。 对于Sorted 的停更，我派如果能有个简单的说明，团队经历了什么，则是更好。定价没问题，是我的问题。太貴了感覺😓发表评论Cogito：连接人与 AI Agents 的 Markdown 编辑器平台：macOS关键词：Markdown、AI@Vanilla：对于 Markdown 编辑器这一类应用，我想大家已经见得太多了。随着 AI Agents 的兴起，Markdown 这个文件格式可谓是「老树开花」，又重新获得了很多人的青睐。Cogito 是一款面向人和 AI Agents 共用的 Markdown 编辑器，它不像 Obsidian 一样偏知识库、不像 Typora 一样偏所见即所得、不像 iA Writer 一样偏写作，也不像 VS Code 一样偏开发者，它最大的特点是「文件优先、Agents 友好」。下面，就让我们一起来看看这个新时代的 Markdown 编辑器到底有何不同。Cogito 的文件库完全基于本地磁盘，没有专属格式，没有加密拦截，没有同步机制，所有的文件和文件夹组织形式都和 Finder 中的一模一样，完全映射，没有任何上手门槛。以我的使用习惯为例，我直接把 Codex 的工作文件夹设置为了 Cogito 的 Location，它会自动抓取 Codex 工作区的所有文件夹和 Markdown 文件，按照日期排序，清晰明了。选择任意一个 Markdown 文件后，Cogito 默认展示它的编辑状态，我们可以随意地按照自己的意愿进行编辑。点击右上角的第一个按钮，则可以在编辑状态和预览状态之间进行切换，对应的键盘快捷键是 Command + R。我经常会借助 Codex 帮我生成一些行业或者公司的研报，最终以 Markdown 格式来进行输出，通过 Cogito 我就可以非常方便地进行编辑和阅读，不需要反复地在 Codex、Finder 和 Markdown 编辑器之间进行来回切换。点击右上角的第二个 Outline 按钮，我们可以快速查看当前文档的大纲，点击标题可以直接跳转到相应的章节。Cogito 除了对 AI Agents 友好，还可以在应用中直接调用本地的 AI Agents 对应用内显示的文档进行总结、分析、编辑等操作。总的来说，Cogito 对于像我这样频繁使用 AI Agents 和 Markdown 文档的人来说是一个非常不错的 Plan B，和主力写作的 iA Writer、用作知识库的 Obsidian 等进行区隔，打开即用，用完即走，非常高效。Cogito 可以在官网免费下载使用，目前处于公开测试阶段。发表首评AndDrive：无缝连接你的 Android 设备和 Mac 电脑平台：macOS关键词：文件管理、手机管理@化学心情下2：虽然现在有不少工具能让 Android 手机和电脑无线互传文件，但「传文件」和「管理文件」是两回事。比如想批量删除 Android 内部存储 Pictures 文件夹下的截图，用传输工具就很别扭，这种场景更适合在电脑端的文件管理器里完成，可以借助筛选和预览功能高效处理。Windows 上这件事不需要任何第三方软件：数据线一插，选择「文件传输」，Android 就被模拟成 MTP 设备，像移动磁盘一样直接管理即可。但 Mac 上就麻烦了：MTP 是 Windows 的专有协议，macOS 想访问 Android 内部存储，要么走 USB 调试，要么依赖第三方 MTP 实现。OpenMTP 算是其中比较知名的一款，但项目两年多没更新、稳定性一般，而且用的是自己写的文件管理器，而非访达，体验称不上顺手。那么有没有能直接用访达管理 Android 文件的工具？接下来要介绍的 AndDrive 就是，它由个人开发者开发，没有走 MTP 路线，而是基于 Android 官方的 ADB 调试通道实现文件管理。安装后，AndDrive 常驻菜单栏，支持 USB 和 Wi-Fi 两种连接方式。使用前需要在 Android 上开启 ADB 调试并授权当前 Mac，再按提示在手机上安装配套的 AndDrive Helper，之后访达里就能看到 Android 的目录了。AndDrive 的挂载方式可以选 SD 卡目录、根目录或不挂载，看到的内容会不同：SD 卡目录只显示内部存储，根目录则多一个系统根目录（没 root 的话很多文件依然动不了）。日常使用可以直接在访达里预览文件、按时间筛选，也能把电脑上的文件拖进 Android 对应文件夹，比在手机自带的文件管理器里翻找直观得多。因为走的是 ADB，AndDrive 顺带还实现了屏幕镜像功能。点击 AndDrive 的「屏幕镜像」，当前 Android 设备的桌面就能投屏到 Mac 上，体验和 macOS 的 iPhone 镜像类似，可以直接操作应用、用中文搜索，还支持剪贴板同步、录屏、截屏等。比起手敲 Scrcpy 命令，门槛会低不少。总的来说，AndDrive 虽然没有用更成熟的 MTP 协议，但用 ADB 这条 Google 官方推荐的方法，让访达管理 Android 文件这件事变得简单可行，还补上了类似 iPhone 镜像的屏幕控制功能，值得有需求的你一试。你可以在App Store免费下载 AndDrive。它不是改名了吗？🤔发表评论Fauxx：与其被动防守，不如主动出击的隐私保护工具平台：Android关键词：隐私保护@大大大K：你的每一次搜索、每一个网络请求、每一段定位轨迹，都在被数据平台和广告商默默收集着。它们会通过这些碎片信息拼凑出你的身份、喜好、习惯甚至是经常交往的人，然后在暗地里交易你的信息、精准推送广告。这也是为什么你或者你的朋友随手搜了点什么，10 分钟后所有 App 和网页就开始围着这个话题给你推广告。因此市面上衍生出不少隐私保护工具，它们的思路大多都是「堵」——屏蔽数据上传、拦截 Cookie、混淆 IP 等等。但 Fauxx 这个 App 啧另辟蹊径：既然广告商一定要收集数据，那就把真实的你混淆在一堆「垃圾」当中。首次打开 Fauxx 会有一个引导流程，你可以填写你的年龄段、职业、兴趣之类的粗略信息。这一步主要是为了让 App 能够生成与你真实信息相反的混淆信息，当然，这一步你也可以选择跳过。Fauxx 并不强制你输入信息。完成后进入主界面打开开关就能启动混淆信息的保护，在这里你可以看到这段时间 Fauxx 执行了多少次「混淆」动作、当前的虚假人设是谁、混淆了哪些类型的信息。至于具体可以混淆的内容，可以在 Targeting 和 Modules 标签中设置混淆等级和类型。Targeting 里有三个可选的层级：第一层就是生成与你真实信息相反的混淆噪声，如果你一开始没有填写，Fauxx 则会随机生成；第二层是对抗广告商的用户画像，不过我实际测试下来会读取失败，但也不影响使用；第三层是「人设轮换」功能，Fauxx 通过生成虚假的身份信息，让广告商无从追踪。Modules 标签里则是七个具体的混淆模块，每个都可以单独开关。设置里还可以调节混淆频率、指定时间段等。那这些「假数据」是怎么骗过广告商的？通常来说，被安插在手机里的追踪器会根据设备 ID、IP 地址、DNS 等信息追踪这台设备上的数据访问，进而推测用户画像。而 Fauxx 会通过模拟搜索、链接访问等「类人」操作（比如搜索时的停顿、假装点击广告），规避机器人检测并成功让追踪器追踪到假的信息访问。Fauxx 可以说是转变了隐私保护的方向，从被动防守变为主动进攻。如果你也想换个思路给广告商添添堵，可以在GitHub免费下载 Fauxx。424Roll：胶片「日日鲜」平台：iOS关键词：手机摄影、胶片模拟@Snow：当下，你用手机拍照真的可以发散出无限可能，不同的镜头、不同的滤镜、不同的手动参数、不同的画面调节项，甚至是不同的外设，只要你愿意钻研一番，就能拿捏各式各样的风格。不过，更多的选择往往意味着更多的「选择困难症」，如果你对于丰富的选择产生了一些焦虑，或者天性就不爱「折腾」，不妨看看 Roll 这款应用，它或许可以带你发现简单拍照的乐趣。打开 Roll，完成简单的授权流程后，你就会收获今日份的新胶卷。应用会简单阐述这份胶卷在感光度、对比度、饱和度等各方面的特质，有一些呈现阳光下温暖的黄色，有一些闪耀着玫瑰般的光泽，有一些甚至想带你进行「第三类接触」。和一般的应用内随机胶片不一样，Roll 所有用户的「掷骰」结果完全一致，无论你身在世界何处，在同一时刻都会使用同一款胶片进行拍摄，剩余可用时间会在胶片介绍页注明。时间有限，点击「Load film」启动拍摄。Roll 不用切换镜头，没有色彩调整，拒绝选择迟疑，它唯一提供的可选项只有用于应对暗光环境时的闪光灯，你只需要按下快门按钮，想想可能与大洋彼岸的陌生人用同一款胶片滤镜进行创作，多少有点浪漫呢。Roll 用来记录生活尚可，如果

周二19派早报：红魔发布红魔 11S Pro 系列手机等今年的新系统会有什么变化呢？早报速览红魔发布红魔 11S Pro 系列手机2腾讯音乐已完成对喜马拉雅的收购6蓝牙核心规范 6.3 正式发布0腾讯自研 AI 设计智能体平台 Ardot 现已开启公测2微软将允许用户修改实体 Copilot 键功能3中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告0马斯克诉奥特曼案被驳回0不妨一看的简讯3红魔发布红魔 11S Pro 系列手机5 月 18 日，红魔发布红魔 11S Pro 与 11S Pro+ 手机等新品。红魔 11S Pro 系列搭载第五代骁龙 8 至尊领先版与红芯 R4 电竞芯片，机身采用纯平设计，11S Pro+ 还有透明水冷可视化设计，屏幕为 6.85 英寸 144Hz 无挖孔无刘海设计，采用屏下前置摄像头与 3D 超声波指纹解锁，搭载新思触控芯片，支持 3000Hz 瞬时触控采样率。系统内置 PC 模拟器，搭载 3.5mm 耳机孔与红外功能。11S Pro 电池容量 8000 mAh，支持 80W 快充；11S Pro+ 电池容量 7500 mAh，支持 120W 快充，两机型均支持旁路充电。红魔 11S Pro 有暗夜骑士与银翼战神两配色，12GB+256GB 售价 5499 元，国补后 4999 元。红魔 11S Pro+ 有氘锋透明银翼与氘锋透明暗夜两配色，12GB+256GB 售价 6199 元，16GB+512GB 6999 元，16GB+1TB 7999 元。另有红魔碳纤维游戏鼠标、红魔散热器 8Pro、氘锋能量卡 Pro 磁吸充电宝等配件发布。来源背后是MagSafe 结构？屏下摄像头技术已经成熟了吗，怎么感觉其他家很少用呢，还有什么缺点吗发表评论腾讯音乐已完成对喜马拉雅的收购腾讯音乐于 5 月 18 日在港交所发布公告，宣布根据并购的条款，收购喜马拉雅的事项已经完成。据此喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为 12.6 亿美元的现金及最多总数 1.75 亿股本公司 A 类普通股（包括本公司于交割时新发行的 A 类普通股，以及本公司授出的以股权为基础的激励相关的 A 类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。来源完了 喜马拉雅 也要有低级充值会员 中级充值会员 高级充值会员了买京东会员送的喜马拉雅会员还没用过呢全部讨论 6发表评论蓝牙核心规范 6.3 正式发布蓝牙技术联盟上周末发布了蓝牙核心规范 6.3，作为半年一次发布计划的一部分，此次更新引入了多项新功能，可进一步提升测距精度、扩展接口容量并提高射频效率。本次更新新增蓝牙信道探测内联 PCT 传输、蓝牙信道探测特定物理层 RTT 精度、蓝牙位资源扩展以及蓝牙 ACP 和 C/I 限值放宽。蓝牙技术联盟表示，联盟成员在描述产品支持的蓝牙功能时，应避免直接引用产品所依据的蓝牙核心规范版本号（例如蓝牙核心规范 6.3），而应在产品包装、文档和推广材料中清晰、准确地描述其产品支持的具体蓝牙功能，特别是与用户需求最相关的部分。来源发表首评腾讯自研 AI 设计智能体平台 Ardot 现已开启公测腾讯云于 5 月 18 日发布公告，正式宣布 AI 设计智能体平台 Ardot 开启公测，官方称其为「一个 AI 驱动的产设研协作平台」。Ardot 支持一句话生成可编辑设计稿，AI 可快速生成 App 页面、官网、海报、插画、PPT 等；支持调用团队自己的业务组件库，一键生成符合规范的可编辑稿；支持直接导入 Figma 文件，完整保留原有布局、样式和组件。Ardot 对代码友好，支持设计一键转代码。设计稿，连同变量、组件、布局数据等设计细节数据，直接拉进 CodeBuddy，一键代码还原；也兼容 Workbuddy、Cursor、Claude Code 等 MCP IDE。Ardot 支持多人在线实时评论、标注反馈和版本对比，配备智能权限中心、全方位行为追溯和无缝交接机制，确保协作规范、有序。此外，Ardot 微信小程序也即将上线。来源确实太像了，感觉就把配色 从紫色改成了绿色🐶[图片]支持导入 Figma 文件🐶1微软将允许用户修改实体 Copilot 键功能5 月 18 日，Windows Central 发现微软上线一份新支持文档，表明自 2024 年起笔记本厂商在 Windows 11 设备上预装的 Copilot 实体按键影响了用户的工作流，他们将在今年晚些时候推送一版 Windows 11 更新，允许用户将 Copilot 键重定义为被替代的菜单键或右 Ctrl 键，用户将在设置 > 蓝牙和其他设备 > 键盘中找到相关选项。来源终于顿悟了？国内都用不了，按它干什么？微软终于发现了8中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告5 月 18 日，中国音数协游戏工委发布 2026 年 4 月中国游戏产业月度报告。2026 年 4 月，中国游戏市场实际销售收入 303.69 亿元，环比下降 3.51%，同比增长 11.04%。其中，中国移动游戏市场实际销售收入 224.30 亿元，环比增长 0.71%，同比增长 9.83%；中国客户端游戏市场实际销售收入 68.66 亿元，环比下降 12.77%，同比增长 18.40%。中国自主研发游戏海外市场实际销售收入 20.32 亿美元，环比下降 5.02%，同比增长 30.78%。来源发表首评马斯克诉奥特曼案被驳回加州奥克兰联邦法院的陪审团于 5 月 18 日作出一致裁决，驳回了埃隆·马斯克对 OpenAI 及其 CEO 萨姆·奥特曼等人的诉讼。经过为期三周的庭审，由 9 人组成的陪审团仅用不到两小时便达成裁决，认定马斯克指控 OpenAI 违背成立初期慈善承诺的起诉已超过诉讼时效。2024 年，身为 OpenAI 联合创始人之一的马斯克指控奥特曼、总裁格雷格·布罗克曼以及主要投资者微软公司背叛了该机构「为人类福祉开发非营利性 AI」的初衷，通过转为商业化运作并与微软结盟来谋取巨额私利。马斯克此前曾为 OpenAI 提供早期资金，并在此案中索赔最高达 1500 亿美元。陪审团认定，马斯克早在 2021 年就已经知晓 OpenAI 转向营利性架构的事实，因此其 2024 年的起诉已经超出了三年的法定诉讼时效。辩方展示的证据表明，马斯克自己当年在未能获得 OpenAI 控制权之前，也曾提议过转为营利性结构。此次判决为 OpenAI 今年晚些时候的 IPO 扫清重大法律障碍。马斯克本人则在 X 上发文，批评该判决仅仅基于「时间程序问题」，而并未对案件的实质正义进行裁决。来源10不妨一看的简讯据外媒 Wccftech 的报道，根据一份去年泄漏的运输清单中的信息显示，Googlebook 笔记本或将搭载英特尔 Panther Lake 12Xe3 核心处理器。运单的信息显示，该产品使用酷睿 Ultra Series 3 系列处理器（即 Panther Lake），并搭载 12Xe3 图形单元。并且在 Google 与英特尔的相关文件中，有一个相同的代号 Q7AP 出现。而该编号对应 Panther Lake 12Xe 主板，对应的 Chromebook 产品代号则为 Felino。该笔记本采用 16 英寸设计，配备 32GB LPDDR5X 内存、1TB 存储空间。根据英特尔公布的产品阵容，12Xe 应该是 Panther Lake 的高端型号，至高可选 16 核心 CPU，功耗范围 25W-80W。如果此次曝光准确那么 Googlebook 很可能采用酷睿 Ultra Series 3 系列处理器。来源据微博用户金猪升级包，英特尔在 5 月 18 日的第三代英特尔酷睿处理器新品分享会上，介绍了英特尔 Project Firefly，将针对价格敏感的 Wildcat Lake 平台，导入中国本土化成熟的手机产业链，统一标准化接口实现规模化，完成降本增效。现场展示了 Project Firefly 的 A+ 参考设计与第一款 Project Firefly 参考设计下落地的量产产品来酷 Air 14，均为超薄机身与 Clean D 设计。来源Linus Torvalds 在 LKML 的新帖文中建议大家不要用 AI 工具提交重复的 bug 报告，大量重复、过时的报告让邮件列表无法正常收集 bug，想用 AI 工具创造点价值就去好好写补丁。来源Apple 定于北京时间 6 月 9 日至 13 日举行全球开发者大会 WWDC26，6 月 9 日凌晨 1 点开幕时将举行主题演讲，介绍 Apple 平台更新，包括人工智能的进展、新软件和开发者工具；6 月 10 日凌晨 4 点举办 Platforms State of the Union，介绍新功能、API 和技术如何进一步赋能开发者。来源还人工智能进展，国行的 AI 有没有？意思是Googlebook运行的是x86_64的安卓？发表评论少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购你可能错过的文章四天三夜，用小长假的时间体验青海湖小环线打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试派评 | 近期值得关注的 App> 下载少数派 2.0 客户端、关注少数派公众号，解锁全新阅读体验 📰> 实用、好用的正版软件，少数派为你呈现 🚀20扫码分享目录 0讨论我来说一句发布发表评论发布

周二，韩国国家工程院第285届NAEK论坛上，三星电子设备解决方案部门常任顾问、三星半导体（DS）部门前总裁庆桂显（Kye-hyun Kyung）表示，中国企业正在积极扩大产能，随着存储芯片供应量的激增，市场格局可能会在明年下半年或2028年上半年发生转变。他补充称，如果大型科技公司的资本支出回报率下降，就有可能削减投资。这样，不仅价格会下降，内存需求本身在2028年后也可能萎缩。（财联社）

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生

全球首个全民免费用ChatGPT Plus的国家，OpenAI官宣了

36氪获悉，大众汽车国际金融股份公司发布公告，公司2026年度第一期中期票据（债券通）申购启动。本期债券的本金总额为不超过人民币30亿元，其中品种一的初始发行规模为人民币20亿元，期限为三年，申购利率区间为1.70%-2.50%；品种二的初始发行规模为人民币10亿元，期限为五年，申购利率区间为1.85%-2.65%。

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生

全球首个全民免费用ChatGPT Plus的国家，OpenAI官宣了

36氪获悉，Omdia最新研究显示，2026年第一季度东南亚智能手机市场出货量同比下降9%，总量为2160万部。然而，市场最值得关注的并非出货量下滑，而是平均售价（ASP）的变化：受存储成本上涨影响，2026年第一季度，东南亚智能手机平均售价达349美元，创历史新高，同比上涨19%。其中，三星以460万部出货量和21%市场份额位居东南亚市场第一，同比增长4%；OPPO以420万部排名第二，出货量同比下降17%；小米位列第三，出货370万部，同比下滑12%。

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生

全球首个全民免费用ChatGPT Plus的国家，OpenAI官宣了

纽约州州长凯茜．霍楚尔（Kathy Hochul）表示，代表长岛铁路工人的工会在周一与大都会运输署（MTA）达成协议。这场罢工导致美国最繁忙的通勤铁路陷入停顿，至周一已进入第三天。　霍楚尔表示，列车服务将于周二中午分阶段恢复。她说，MTA与代表工人的五个工会于周一晚间达成了一项“公平的协议”。（新浪财经）

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生

全球首个全民免费用ChatGPT Plus的国家，OpenAI官宣了

菲律宾总统小费迪南德·马科斯表示，他对经济停滞和高物价感到担忧，但他的政府正努力抑制通胀并”维持经济机器运转”。他表示，政府正在采取措施控制食品通胀，包括对进口大米设定价格上限。（新浪财经）

对话万成云商：发文章≠GEO优化，大模型不是喂什么就推什么杨越欣（杨桃）·2026年05月18日 13:05对话万成云商：拆解出海企业AI时代GEO布局路径“帮我搭配一套适合海边度假的衣服。”“帮我推荐一家适合约会的西餐厅。”“3000元预算，买哪款扫地机器人最划算？”这类与AI的问答正发生在全球数亿用户身上。过去，人们会把这些问题输入百度、谷歌等搜索引擎，因此产生了SEO（Search EngineOptimization，搜索引擎优化）营销。而当流量入口从搜索引擎转向AI工具，GEO（Generative EngineOptimization，生成引擎优化）也应运而生。区别于传统SEO追求的“搜索结果排名”，GEO的核心目标是让品牌信息在AI生成的回答中被优先引用和推荐。在AI全面渗透用户信息获取与消费决策场景的时代，对于出海企业而言，做好GEO意味着抢占新的流量入口。GEO的逻辑和SEO有什么根本不同？为什么有人说GEO是伪命题？出海企业布局GEO该怎么做？中小企业如何低成本起步……带着这些出海企业关心的问题，我们特别对话了万成云商品牌出海事业部总经理谭莉，拆解AI搜索时代如何让企业被AI看见、信任，并推荐给用户。36氪：GEO被称为SEO在AI时代的进化。相比于SEO的核心在于搜索结果的排名，GEO优化的关键是什么？它与传统SEO的核心区别在哪里？谭莉：核心区别在于传统SEO偏“规则博弈”，GEO更像“信任竞争”。因为过去做SEO，我们的目标是把页面排到Google第一页，让用户点进去；但现在很多用户直接在AI里问问题，甚至不点链接了，那你要解决的是——怎么让AI在回答里“引用你”甚至“推荐你”。所以GEO的关键首先是，你是不是一个“可信任源”（全网有没有人认可你）；其次，你的内容能不能被AI理解、拆解、复用；最后是你的内容有没有完整的证据链，比如数据、案例或者第三方背书。36氪：继GEO之后，也有人提出AEO概念，两者有什么区别？谭莉：这两个概念很多人会混着讲，但我个人是这么区分的：AEO更偏“问答结果优化”，核心是让你的内容直接成为某个问题的标准答案，比如FAQ、精选摘要这些。GEO更广，它面对的是生成式AI，不只是回答一个问题，而是参与“多源信息整合后的输出”。所以AEO更偏“结构技巧”，GEO更偏“整体品牌与内容生态”。36氪：GEO的概念近两年引起广泛讨论。为什么现在这个时间点企业需要关注GEO？在AI搜索渗透率持续提升的背景下，GEO应该在企业营销体系中占据怎样的战略位置？谭莉：分享几个很现实的变化：现在用户越来越习惯直接问AI，而不是自己筛选搜索结果；而且AI在很多行业决策前端，已经开始替代“信息收集阶段”；再就是有些客户还没打开你的官网之前，可能就在AI里形成了对你的判断。所以企业没办法去忽视GEO，现实摆在这里了。我认为，GEO在企业营销体系中占据着很关键的位置，它不会替代SEO，而是和SEO并行的“第二流量入口”，甚至未来会变成主入口之一。36氪：怎么理解有企业做过一些尝试后认为GEO是伪命题，或者靠自己持续给大模型喂语料也可以？谭莉：我觉得这是对AI工作机制有点误解，大模型不是你“喂什么它就推荐什么”。它有几个筛选逻辑：第一，信息是否一致（比如不同来源是否能得到印证）；第二，来源是否可信（是不是长期存在的内容资产）；第三，是否被多平台提及（而不是单一来源）。所以你自己“喂语料”，本质只是单点信息输入，很难形成“信任网络”。不过我也理解为什么有人说GEO是伪命题，因为确实有很多“伪做法”，比如只做站内内容、只堆AI文章、不做外部验证，这些基本都很难起效果。36氪：当企业布局海外GEO营销时，需要经历哪些关键步骤？不同行业、不同规模的企业在路径上会有哪些差异？谭莉：从实操角度，会有四个关键步骤：第一步先拆解你的客户是怎么筛选供应商的，这个路径是怎么样？第二步重建你的内容结构，包括官网、博客、FAQ、案例这些，全部按“可被AI理解”的方式重构。第三步是搭建外部信任体系，比如评价平台、媒体曝光、社媒讨论、行业引用，都很重要。第四步就是做多渠道分发，持续让内容在不同平台被看见和引用。不同企业可能在行业复杂度和品牌基础这两方面存在明显差异。比方说医疗、工业类比其它行业更需要专业背书，那有品牌的企业就更容易放大，没有品牌的就要先“补信任”。36氪：在实施过程中，企业应如何在内容策略和技术基建层面做出改造甚至重塑，以同时满足AI理解与用户共鸣？谭莉：内容策略方面要做到结构清晰，信息有证据，有数据来源，表达要接近真人沟通。技术层面反而是基础的更容易被忽略：比如Schema结构化数据（Product、FAQ、Article等）、robots、index、hreflang这些基础配置，再一个就是页面可读性。总结来说，是做“可被理解的内容”。36氪：在布局GEO出海方案时，企业最容易忽视但至关重要的环节是什么？对于缺乏专业团队的中小企业，有哪些可行的解决方案或合作模式？谭莉：我们服务这么多跨境企业，发现最容易忽视的一点其实是：“让AI能访问你”。听起来很基础，但很多企业在实操中会陷入误区：比方说robots把AI爬虫挡了、页面noindex设置错误，或者多语言混乱……这些都会直接导致你做再多内容，AI也看不到。第二个容易忽视的是站外信任建设。很多人还停留在“发文章=做优化”，但GEO更看重：评论、测评、社媒讨论、媒体引用。对于中小企业，我建议两种路径：①轻量级：聚焦一个细分领域做深内容+少量高质量外部背书；②合作型：找有资源的服务商做整合（内容+渠道+数据），不要自己盲目试错，成本会很高。36氪：企业在选择GEO服务伙伴时，应该重点考察对方的哪些核心能力（如是否具备垂直行业知识库、跨平台渠道资源、效果监测与分析体系等），以避开“无效投入”的陷阱？谭莉：这个问题非常关键，因为现在确实有不少“概念型服务商”。我建议重点看四个能力：第一，行业理解能力，是不是懂你这个行业，而不是只会写内容。第二，内容生产质量。有没有真实案例、是否具备本地化能力，而不是纯AI批量生成。第三，渠道与分发能力。有没有海外平台资源，比如媒体、社区、测评渠道。第四，数据监测体系。比如做一段时间了，看看有没有被AI引用？哪些内容有效？ROI如何？如果对方只能讲“发多少篇内容”，基本可以直接排除。36氪：2026年初Seedance2.0的爆火，让业界看到了AI视频生成从“抽卡式试错”向“工业化生产”跨越的可能性 。10秒带货视频成本仅需8元的“成本塌方”对于跨境电商的内容生产逻辑意味着什么？谭莉：这个变化其实挺颠覆的。过去做视频，成本高、周期长，所以大家拼的是“有没有内容”；但现在成本接近于零，意味着：内容不再是门槛，内容“是否有效”才是门槛。36氪：当视频生产成本趋近于零，所有人都能批量生成高质量内容时，企业的竞争焦点会发生怎么样的变化？从GEO布局的角度看，企业应如何构建AI时代下的营销思维？谭莉：供给无限的时候，稀缺的就变了。我认为未来竞争会集中在三点：第一，认知占位，用户心里有没有你；第二，信任结构，有没有被多方验证；第三，数据能力，能不能快速迭代内容。从GEO角度，我觉得企业要建立一种新思维：要把品牌内容当资产去做长期积累，同时将口碑、评价这些信任一点点做起来，更重要的是发内容之前，问问自己：AI会不会用它？用户会不会信它？避免做“自嗨”的内容。36氪：Seedance2.0的火爆也引发关于深度伪造和版权的讨论。跨境电商卖家在利用这类工具生成营销素材时，应如何规避潜在的肖像权、版权风险？谭莉：这个问题确实不能忽视，我建议，尽量避免使用可识别人物（尤其是公众人物）做生成素材，素材来源尽量用可商用授权库，品牌、Logo、音乐等要确认授权范围，涉及医疗、功效类内容要特别谨慎，因为合规风险更高。毕竟，AI可以帮你降低成本，但不能帮你“免责”。本文由「杨越欣（杨桃）」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

Agent、多模态、应用、算力一天看尽，峰会亮点在此｜5.20日，来现场一起AI未来一氪·2026年05月18日 15:03第四届中国AIGC产业峰会5月20日北京举办，干货满满开放报名。进入2026，AI愈发狂飙突进。围观体验之余，人人不免在心中自问：朋友圈刷屏的“龙虾”、Harness等AI新事物，跟我到底有什么关系？真的有必要跟吗？AI创业、AI融资如火如荼，属于我的机会又在哪里？别人已经在用AI做视频、写代码、跑项目，我是不是已经慢了一拍？……到最后，几乎所有问题都会汇成同一个问题：我，到底该如何用AI？如果你对这些问题还很模糊，不妨来第四届中国AIGC产业峰会走一趟——一天时间，把这一年AI产业最值得关注的人、事、判断，一次性讲清楚。先提前剧透一波。18位重磅嘉宾，1场Agent主题圆桌，1份年度榜单，1张全景图谱——所有内容都已敲定。阵容上，既有昆仑万维、智谱、商汤、百度、蚂蚁、MiniMax这样的行业头部，也有亚马逊云科技、硅谷Fusion Fund带来的全球视角；既有复旦邱锡鹏、港大黄超代表的学界与开源前沿，也有盛大EverMind、太初元碁、趣丸科技等活跃实战派玩家。议程上，从Agent商业化落地、多模态与空间智能的最新突破，到AI在文娱、医疗、企业服务等场景的纵深渗透，再到算力与AI Infra的范式之变……这一年最值得关注的AI话题，几乎都装进了这一天。话不多说，下面直接上重点。亮点1：Agent到底怎么落地？不同视角玩家都有心得如果说去年大家还对“Agent是不是未来”持观望态度，那么2026年，这个问题的答案已经不言而喻。“Agent不是未来，是现在”——这正是智谱高级副总裁吴玮杰带来的判断，他将分享智谱在Agent商业化落地上的最新实战。但Agent真正走入企业，并非一蹴而就。亚马逊云科技产品技术部技术总监王晓野将聚焦如何跨越Agent落地鸿沟，拆解从最强模型到企业级AI Agent之间那条最难走的路。而港大助理教授黄超则会从学术与开源前沿的视角，探讨AI Agent范式变革的挑战与机遇。他所带领的HKUDS GitHub开源平台累计获得超24万Star，位列全球Top50，话语权毋庸置疑。与此同时，Agent要真正成为“数字生产力系统”，还需要更底层的能力支撑。盛大集团副总裁、EverMind CEO邓亚峰将分享长期记忆驱动的自进化，探讨AI如何从一个“工具”演化成可持续成长的数字生产力系统。当然，光有演讲还不够。本届峰会专门设置了一场Agent主题圆桌：2026，Agent产品的不确定性与非共识机遇。出席嘉宾包括趣丸科技副总裁兼首席战略官、《屠龙之术》主理人庄明浩，ColaOS&MarsWave CEO冯雷，以及EvoMap创始人张昊阳。这三位都是Agent产品一线的深度玩家，从产品形态到操作系统，从多智能体框架到自进化引擎，他们会碰撞出什么样的非共识观点？非常值得期待。亮点2：Agent之外，模型能力边界也在继续扩张虽然Agent是今年峰会的绝对主角之一，但不可忽视的是，支撑Agent的模型能力本身，也在以惊人的速度向外延伸——从“数字”到“物理”，AI对世界的认知边界正在被重新定义。商汤科技执行副总裁、首席科学家林达华将分享多模态模型如何从“看懂”和“生成”，进化到真正可感知、可生成、可行动的空间智能新阶段。上午的GenAI Talk环节，蚂蚁灵波科技首席科学家沈宇军将围绕一个核心命题展开对谈：AI 2.0下半场，从AIGC到AIGA。当生成式AI从“内容”走向“行动”，整个产业逻辑将被改写。而京东JoyInside业务负责人戴文军则会把这个命题进一步推往物理世界，当AI从屏幕里走出来，走进真实终端、真实场景，“AI World”才真正开始。模型层面，复旦大学特聘教授、模思智能创始人邱锡鹏将分享MOSS多模态模型及其推理优化的最新进展。这位主持研发了MOSS系列开源大模型的“大模型一线人物”，将带来学术与工程结合的硬核内容。亮点3：具体怎么玩？这些应用玩家甩出成功案例模型也好、Agent也罢，具体到不同领域，机会到底在哪？昆仑万维董事长兼CEO方汉率先给出他的答案。他这些年带着昆仑万维在AI音乐、AI社交、AI短剧等多个AIGC方向同时下注，其分享会从一线视角，讲讲AIGC具体应用这一年究竟跑通了什么、又踩过哪些坑。紧接着，来自不同赛道的“实战派”玩家也将依次登场：文娱创作领域，风行在线CEO易正朝将带来一个核心判断——从AI编程到AI视频，众创才是AI生产力的核心杠杆。他会讲述AI如何从单点工具，进化为撬动整个创作者生态的杠杆，从而让普通人也能撑起一支创作团队。生产力工具层面，百度秒哒产品总经理朱广翔会分享他们“人人都是创造者”的产品理念，当门槛被AI抹平，想象力就是新的生产力。企业级AI方向，MiniMax ToB中国区商业化负责人胡维琦将带来MiniMax这一年在ToB战场上的真实打法与思考。大健康赛道，轻松健康集团执行副总裁马孝武将拆解从循证智能体到AI驱动的健康服务增长飞轮，展示AI在严肃医疗场景里的真实增长引擎。文娱、办公、企业、医疗……AI不再是“被讨论的对象”，而是“正在被使用的工具”。每一个细分赛道里，都已经有人开始把案例端上桌了。亮点4：从训练到推理，算力的故事也在被改写模型与应用之外，2026年最值得关注的另一条主线，是AI基础设施的范式重构。过去几年，所有人讨论算力都默认一个前提——为训练服务。但2026年，风向变了，推理需求暴涨、训练成本高企，整个算力叙事正在从“训练驱动”转向“推理驱动”。Fusion Fund创始合伙人张璐将从全球资本与产业的双重视角，解读算力叙事的重构。当AI基础设施从训练转向推理，一场范式变革正在发生。而国产算力这一边，太初元碁首席产品官、高级副总裁洪源将带来他对国产算力路径的思考。当“Token”正在成为AI时代的新计量单位，国产算力如何筑基Token智能新未来？他会给出自己的答案。亮点5：看尽AI全景生态，一份榜单+一份报告延续往届传统，本届峰会，技术先锋、产业龙头、新锐势力将再次全景式碰撞。为了给行业更多参考，我们还将颁出2026年值得关注的AIGC企业、2026年值得关注的AIGC产品两类奖项，以及一份中国AIGC应用全景图谱。前者是量子位根据过去一年AIGC企业、产品的表现与反馈，并结合对2026年技术与场景的观察，评选出的高潜力企业和产品。后者是量子位智库绘制的一幅全面立体的中国AIGC应用全景图谱，用以展示国内AIGC应用的市场格局与发展动态，并深入剖析最具潜力与竞争力的产品和赛道。5月20日举办最后，时间地点以及最全议程在此，现场还有参会伴手礼等你来领，咱们到时候见 ~时间：5月20日 9:30地点：北京·金茂万丽酒店现在，中国AIGC产业峰会正在火热报名中，欢迎点击链接报名线下参会：https://hdxu.cn/1Hjr1本次峰会也将同步线上直播，欢迎预约参与：量子位视频号。参会嘉宾简介一览方汉昆仑万维董事长兼CEO方汉毕业于中国科学技术大学近代物理系，曾任职于中国科学院高能物理研究所、拓林思、亚信以及千橡互动。于2008年协助周亚辉先生创立昆仑万维。方汉拥有30年互联网从业经验，从1994年开始参与和倡导开源运动，是中文Linux奠基人、中文Linux四剑客之一，著作国内第一本Linux书籍《Linux实用大全》，他也是国内最早的网络安全专家，曾负责三大电信运营商的全网安全审计，领导开发了国内第一款P2P下载软件DUDU加速器，负责研发了国内市场占有率最高的网页游戏《三国风云》，为中国互联网最资深的参与者。吴玮杰智谱高级副总裁吴玮杰目前担任智谱公司高级副总裁，负责公司MaaS平台商业化，并兼任泛互联网事业部总经理及浙江子公司总经理。他在人工智能、互联网和信息技术领域有着丰富的商业化管理经验和项目落地经验，曾任字节跳动企业应用全球首席商业官（旗下品牌包括飞书、LarkPeople和幕布等）、GE数字集团大中华区副总裁、复星集团联席首席发展官&首席投资官。除此以外，他也曾在思科、SAP、Oracle等公司从事商业化管理工作。吴玮杰博士毕业于电子科技大学，同时持有香港大学国际工商管理硕士学位以及复旦大学软件工程学士、硕士学位。易正朝风行在线CEO易正朝，风行在线CEO，作为业务领头人，引领公司从传统互联网视听平台全面转型升级为AI文娱生态平台。推动“AI赋能众创”战略落地，带领团队打造橙星梦工厂、小剪刀等AI创作工具，影视网文智能小程序平台、电视鸿蒙平台等核心智能产品，构建从内容创作、智能分发到商业变现的全流通链生态。推动公司荣获AI赛道奖项、抖音快手等平台的头部服务平台奖项，在AI文娱新时代，推动内容生产效率革新与创作者生态繁荣，致力于让内容流动更简单。林达华商汤科技执行副总裁、首席科学家林达华教授现任香港中文大学信息工程系副教授、交叉学科人工智能研究所所长，并为商汤科技联合创始人、执行董事兼首席科学家，兼任上海人工智能实验室领军科学家。他于2012年获麻省理工学院计算机科学博士学位，研究方向涵盖计算机视觉、深度学习、大模型与生成式人工智能，在人工智能顶级会议和期刊发表论文逾300篇，引用逾90,000次，h-index达131。林教授曾获2010年NeurIPS最佳学生论文奖，带领团队在多项国际计算机视觉竞赛中获奖，连续五年入选Elsevier全球前2%顶尖科学家。他主导发起OpenMMLab，并推动书生大模型、OpenCompass评测平台及DeepLink开放计算体系建设，其中DeepLink于2024年获世界人工智能大会“卓越人工智能引领奖”。2026年，他荣获“中银香港科技创新奖”，以表彰其在计算机视觉与多模态智能领域的系统性创新。他长期服务学术共同体，曾任IJCV编委，多次在人工智能主要国际会议担任领域主席，并担任IEEE大规模深度学习模型评测标准工作组主席。邓亚峰盛大集团副总裁、EverMind CEO邓亚峰，毕业于清华大学，正高级工程师，拥有二十余年人工智能算法及产品研发经验。现任盛大集团副总裁、EverMind CEO。曾任360集团副总裁、人工智能研究院院长兼搜索事业部总经理，科创板首家AI上市公司格灵深瞳CTO。曾荣获2021年中国人工智能年度十大风云人物，教育部技术发明奖二等奖。长期致力于大语言模型、计算机视觉及AI for Science等领域的研究与应用，带领团队多次在国际国内主流AI评测中获得佳绩；累计申请发明专利160余项（已授权98项），发表论文50余篇（含10篇Nature子刊）。王晓野亚马逊云科技产品技术部技术总监王晓野现任亚马逊云科技产品部技术总监。长期专注于企业现代化应用架构，数据及人工智能能力的建设。负责带领技术专家团队提供云上产品及开源方案技术实践指导，并致力于亚马逊云科技全栈服务在中国市场的应用和推广。沈宇军蚂蚁灵波科技首席科学家沈宇军，现任上海蚂蚁灵波科技有限公司首席科学家，同时领导蚂蚁技术研究院交互智能实验室。博士毕业于香港中文大学，研究方向为计算机视觉、生成模型、空间智能和具身智能。在CVPR、TPAMI、ICCV、ECCV、NeurIPS等国际顶级会议和期刊上发表论文百余篇，Google Scholar引用量逾万次。2024年起担任蚂蚁灵波科技首席科学家，致力于探索计算机视觉在机器人行业的落地之路，带领团队构建从空间感知到智能决策、从基座能力到世界建模的完整技术矩阵，推动具身智能从实验室走向规模化应用。邱锡鹏复旦大学特聘教授、上海创智学院院长助理、模思智能创始人邱锡鹏，CAAI Fellow，国家杰青获得者，主要研究方向为大模型、智能体以及科学工程等领域中的应用，发表学术论文被引用4万余次，5次获得AI领域高水平会议的最佳/杰出等论文奖，担任中国人工智能学会自然语言理解专委会副主任、中国中文信息学会大模型与生成专委会副主任、上海市计算机学会自然语言处理专委会主任等，主持研发了MOSS等系列高影响力开源大模型。连续多年入选中国高被引学者、全球前2%顶尖科

科氪 | 性能极限与跨界尝试：红魔11S Pro系列深度体验科氪·2026年05月18日 17:35红魔11S Pro电竞旗舰测评：纯平硬核，性能拉满在智能手机行业普遍面临镜头模组日益凸出、机身内部堆叠空间受限的背景下，电竞旗舰由于其对散热和功耗的极端要求，往往表现出更为激进的工业设计。作为新一代的性能导向型产品，红魔11S Pro系列通过软硬件的垂直堆叠，试图在移动终端上探索更进一步的性能上限。以下我们将从设计、理论性能释放、以及实际软硬件体验三个维度，对其进行深度拆解。一、 外观设计：纯平后盖的内部堆叠与电竞美学从视觉和握持逻辑来看，红魔11S Pro最显著的特征在于其坚守的纯平后盖设计。在行业因大底传感器而集体妥协于镜头凸出的趋势下，这种“反向”设计在工业制造上面临着极高的空间挑战。产品图片内部空间重构：为了在不增加机身异常厚度的前提下抹平镜头凸起，红魔对内部架构进行了重新设计。在有限的Z轴高度内，开发团队不仅压榨空间塞入了新一代的脉动水冷模块、无线充电线圈，还容纳了大面积的VC散热板和高密度电池。产品图片透明电竞美学的视效迭代：机身背部延续了红魔标志性的透明设计语言。通过透明材质，内部的脉动水冷引擎实现了可视化。在水冷系统运行时，用户可以观测到液体流动的动态秩序。相比于纯粹的RGB灯光，这种工业风的设计提供了一种更具技术感的视觉回馈。材质与细节处理：后壳采用了微蚀刻工艺打造的浮雕星轨纹理，搭配环形拉丝处理，使得整体质感较为收敛和冷峻。机身在肩键、腰部、LOGO及风扇处均设有支持自定义的RGB灯效，在暗光游戏环境下能提供必要的电竞氛围。产品图片二、 性能与散热：高标准筛选的“特种兵”芯与风水双冷架构性能表现是电竞手机立足的核心。红魔11S Pro首发搭载了第五代骁龙8至尊领先版移动平台。为了在重负载下维持超频表现，红魔在芯片筛选和散热架构上采用了更严苛的标准。产品图片芯片体质的物理筛选：为了满足电竞长时高频的需求，红魔所采用的这颗“领先版”芯片，主要筛选自晶圆中心区域。由于物理刻蚀规律，该区域的芯片通常具备更稳定的温度控制和更高的电气性能。在此基础上，通过最高4.74GHz主频的冲刺分级测试，以及高温高压下的自适应电压调节（AVS）检测，筛选出具备自适应能力的芯片，以确保在极端重负载下不易因发热而触发激进的降频保护。风水双冷复合散热系统：针对第五代骁龙8至尊领先版的发热特性，红魔在散热上引入了原本常见于服务器端的复合方案。手机将【驭风4.0】贯穿式主动散热风扇与脉动水冷引擎相结合，辅以复合液态金属3.0和4D冰阶VC。风扇与整机均通过了IPX8级防水认证，在提升主动散热效率的同时，规避了由于进水导致元器件损坏的风险。在散热系统运转时，后盖的蓝色轨道圈内，蓝白相间的液冷介质会随之一起转动。仪式感拉满！产品图片独立显存与能效调度：在架构协作上，红魔自研的红芯R4电竞芯片与高通联合打造了独立的高速显存。在图形高渲染场景下，此设计可减少高频访问系统内存的次数，从而降低总功耗。结合CUBE擎天游戏引擎3.0的AI感知调校，系统能够精细化调配算力，平衡了高帧率表现与续航消耗。三、 使用体验：悟空屏2.0、Steam直连与续航评估在日常交互和实际游戏运行中，硬件的堆叠最终转化为了多场景的体验。悟空屏2.0的视觉完整性：正面配备的悟空屏2.0基于京东方X10全新屏下发光材料，实现了全无挖孔的视觉体验。屏幕黑边控制在1.25mm，屏占比达到95.3%，支持144Hz刷新率。由于没有刘海或挖孔遮挡，在全局视野和视频播放时拥有更纯粹的画面呈现。同时，屏幕集成了全屏3D超声波指纹解锁，提升了暗光或湿手状态下的解锁盲操率。产品图片行业首发Steam直连：性能的跨界探索：这是该机型在软件生态上的重要尝试。依托转译引擎，红魔攻克了X86与ARM架构之间的转译难题。用户拨动竞技键进入游戏空间后，可直接扫码登录个人的Steam账号。系统不仅支持游戏库内兼容作品的下载游玩，还打通了云端存档。用户在PC端的进度可以在手机侧的PC模拟器上无缝同步，迈出了打破移动端与桌面端算力壁垒的一步。产品图片产品图片产品图片产品图片操控精准度与手汗免疫：针对瞬时爆发的操控需求，该机引入了全新的新思触控芯片，在MagicTouch4.0算法支持下，实现了3000Hz的瞬时触控采样率。该方案优化了湿手控制算法，一定程度上避免了玩家因手汗导致的断触或漂移现象。续航与快充系统：为了支撑高功耗输出，红魔11S Pro在续航规格上给出了较高配置。PRO版本搭载了8000mAh的大容量电池，显著延长了重度游戏下的续航时长。而PRO+版本则在7500mAh电池的基础上，配备了120W有线快充与红魔首款80W无线闪充，在充电灵活性和回血速度上取得了较好的平衡。产品图片产品图片日常AI功能落地：在非游戏场景下，红魔11S Pro作为主力机也搭载了丰富的日常工具。AI帮写、AI消除、AI超清文档等功能提升了商用和办公效率；影像系统则在保留Live Photo的同时，延续了红魔标志性的星空与车轨慢门拍摄功能，拓宽了产品的实用边界。产品图片结语红魔11S Pro系列是一部特色鲜明、目的纯粹的硬件设备。它通过风水双冷散热、高标准芯片筛选以及无孔全面屏，将移动端的游戏性能推向了新的极端，并通过Steam直连展现出了打破平台壁垒的野心。对于不愿向主流“大镜头凸起”妥协、且对高负载游戏及跨平台游玩有硬性需求的极客与核心玩家而言，这是一台工业特色鲜明、技术指标扎实的旗舰标杆。产品图片本文由「科氪」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

番茄小说正推动 AI 动漫上院线，开放头部IP改编权限｜36氪独家兰芥·2026年05月18日 17:46字节跳动为番茄IP定下了影响力目标。作者｜兰杰 李晓霞编辑｜乔芊36氪独家获悉，据多位接近字节跳动人士消息，字节旗下的番茄小说正在推动AI动漫上院线。在这一过程中，会开放番茄小说平台头部IP的改编权限。之于番茄小说，如何更好地挖掘平台上好IP的潜力和商业价值，是一个亟待解决的难题。一方面，以文字形式为载体的IP内容，缺乏统一的视觉形象，使得平台上的IP难以有联名、衍生品等更丰富的变现形式。另一方面，AI漫剧这一新事物迅速发展壮大，其中不乏改编自番茄小说IP的粗制滥造之作。这些内容不但没能延长好IP的生命力，反而形成了一种消耗。此时，通过可登陆院线的影视内容来探索平台IP的更多可能性，不失为一种好的尝试。另据上述知情人士表示，字节跳动内部为番茄IP定下了影响力的目标，“想让行业内外的人觉得番茄的出品是精品的、正向的。”这一需求落到具体的业务上，上院线会是个不错的答卷。早在2023年年底，番茄小说就官宣成立了“番茄影视”和“番茄动漫”两大厂牌，推动平台IP的影视化。过去两年间，番茄影视先后与腾讯视频、爱奇艺和优酷等长视频平台合作，推动了《十日终焉》《我在精神病院学斩神》等平台头部IP的改编，但这些合作多停留在IP授权阶段。虽然番茄系（字节旗下以番茄小说、红果短剧等为主的业务）并不愁关于钱的问题，但显然成本高、回报周期长的长视频生意，并不符合依赖大手笔投流、需要市场快速反馈的“字节逻辑”。直到AI的出现，大幅降低了成本、提高了效率。36氪还从另一位知情人士处获悉，红果漫剧侧如今正在采购120分钟不分集的长内容，这也是当下各大视频平台争相竞逐的方向。之所以上线AI动漫内容，而不是AI仿真人内容，或许与当前AI大模型的能力有关。行业头部的AIGC创作者向36氪表示，如今视频生成大模型的能力已经可以做到生产120分钟的长内容，但是在仿真人这一赛道上，人物、场景的一致性依旧是个难题，而动漫内容可以通过部分突出的人物形象设计——例如彩色的头发、雀斑等，一定程度上弥补一致性的不足。回顾番茄系并不算漫长的发展历程，它可以说是字节跳动在文娱领域的一次突围。在番茄小说问世之前，字节始终未能建立起足够的内容能力和IP积累。而如今，从小说到短剧，再到漫剧，番茄系已具备了IP开发的全链条能力。与此同时，番茄系显然不满足于仅仅停留在当下的免费市场，而是在探索更多元的商业模式，包括线上订阅和院线票房等更多商业可能性。本文由「兰芥」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：IC photo

氪星晚报 ｜百度：一季度营收321亿元，AI业务收入136亿元；马斯克预计今年美国将广泛使用自动驾驶汽车；巨力索具旗下十余家企业已注销卜算籽·2026年05月18日 19:11特朗普：本应要求获得更多英特尔公司的股份大公司：索辰科技：物理AI业务目前尚处于布局初期36氪获悉，索辰科技公告，公司股票于2026年5月14日至5月18日连续三个交易日内收盘价格涨幅偏离值累计达到30%，属于股票交易异常波动情形。公司近期实施的募投项目“物理合成数据核心技术基座构建与多领域场景示范应用项目”尚处于筹备阶段，若出现募集资金投资项目未能顺利实施、新技术开发进度不达预期、研发遭遇技术瓶颈甚至失败，将会对公司经营业绩造成一定影响。公司物理AI业务目前尚处于布局初期。物理AI技术研发投入大、技术迭代快，存在研发进度不及预期、核心技术迭代风险；同时下游行业应用场景拓展、商业化变现存在不确定性，若市场需求、行业政策、技术竞争格局发生不利变化，物理AI业务收入及盈利贡献存在不及预期风险。智元WITA成为全国首例完成大模型备案的具身智能交互模型36氪获悉，根据上海市网信办最新公告的上海市生成式人工智能服务备案情况，智元WITA成为全国第一款实现合规备案的具身智能交互大模型，标志着中国具身智能产业从“技术验证”进入“合规商用”的新阶段。依托智元“三智一体”全栈技术体系，智元WITA核心应用于人形机器人交互场景。上交所：上周新增8家IPO受理企业36氪获悉，据“上交所企业上市服务”消息，本期（5月11日—5月17日）新增8家受理企业，其中上交所主板1家，科创板3家，创业板3家，北交所1家。截至5月17日，今年新增31家受理企业。截至5月17日，各板块在审企业共295家。其中上交所62家（主板18家，科创板44家），深交所61家（主板17家，创业板44家），北交所172家。截至5月17日，2026年已公布终止审查（撤材料+否决/终止注册）企业20家，其中上交所7家（主板3家，科创板4家），深交所创业板1家，北交所12家。腾讯音乐：收购喜马拉雅的事项已经完成36氪获悉，腾讯音乐在港交所公告，于2026年5月18日，根据并购协议的条款，收购喜马拉雅的事项已经完成，据此，喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为12.6亿美元的现金及最多总数1.75亿股本公司A类普通股（包括本公司于交割时新发行的A类普通股，以及本公司授出的以股权为基础的激励相关的A类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。爱奇艺：一季度总收入62.3亿元，会员收入环比增长36氪获悉，爱奇艺发布2026年第一季度财报。财报显示，一季度爱奇艺总收入62.3亿元。其中，会员服务收入42.0亿元，环比增长2%。在线广告服务收入12.4亿元，内容发行收入3.6亿元，其他收入4.3亿元。百度：一季度营收321亿元，AI业务收入136亿元36氪获悉，百度发布2026年第一季度财报。财报显示，一季度营收321亿元，上年同期营收324.52亿元。一般性业务收入260亿元，同比增长2%；其中AI业务收入136亿元，占百度一般性业务收入的52%，已连续多个季度增长；GPU云收入同比增长184%。百度创始人李彦宏表示，AI业务收入占比的首次过半，标志着AI已成为百度的核心驱动力。巨力索具旗下十余家企业已注销36氪获悉，近日，巨力索具因涉嫌信息披露误导性陈述违法违规，被证监会立案调查。天眼查App显示，巨力索具股份有限公司成立于2004年12月，法定代表人为杨建国，注册资本9.6亿人民币，经营范围包括应急装备、装备专用工装及工具和箱组的设计、生产和销售等，由巨力集团有限公司、杨建忠、香港中央结算有限公司等共同持股。对外投资信息显示，该公司对外投资26家企业，其中10余家企业处于存续状态，包括巨力索具（河南）有限公司、河北巨力应急装备科技有限公司等，其他10余家企业已注销。投融资：“华羽先翔”完成超亿元Pre-A+轮融资，蚂蚁集团领投36氪获悉，近日，国内全倾转eVTOL企业“华羽先翔”完成超亿元Pre-A+轮融资，本轮融资由蚂蚁集团领投、彼岸时代及复星创富跟投。融资资金将重点用于加速鸿鹄系列eVTOL的适航取证，以及低空出行及商业场景的落地与创新。“极壳科技”获5000万美元B+轮融资36氪获悉，近日，外骨骼品牌“极壳科技”宣布完成5000万美元B+轮融资。此轮融资由蚂蚁集团和美团龙珠领投，Sofina、Granite Asia跟投，高鹄资本担任独家财务顾问。此轮融资将加速极壳在技术研发、全球市场拓展及消费场景应用方面的布局。新产品：酷开发布企业AI操作系统36氪获悉，近日，深圳市酷开网络科技股份有限公司举办“场景智能·硅基管理”发布会，正式推出原生AI企业操作系统——酷开Happy Work AIOS Lite MVP。该系统已实现从生活场景到企业经营场景的全维度覆盖，完成多场景AI生态闭环落地。今日观点：马斯克预计今年美国将广泛使用自动驾驶汽车埃隆·马斯克周一表示，他预计今年晚些时候，无需人类驾驶的自动驾驶汽车将在美国更加普及。马斯克通过视频连线在特拉维夫举行的“智能出行峰会”上发言时称，得克萨斯州的三个城市已有无人驾驶汽车在没有安全监视员的情况下运行，并补充说今年这一模式将扩展至全美。（新浪财经）其他值得关注的新闻：2026电影总票房已超145亿元据票务平台数据，截至今天下午，2026年电影票房已超145亿元。五一档新片长尾效应持续释放，其中《给阿嬷的情书》连续九天蝉联单日票房榜首。（央视新闻）特朗普：本应要求获得更多英特尔公司的股份美国总统唐纳德・特朗普在周一出版的杂志采访中表示，他本应代表美国政府要求获得更多英特尔股份。特朗普政府去年持有英特尔10%的股份，并宣布向这家芯片制造商投资约100亿美元，用于在美国境内新建或扩建工厂。交易完成八个月后，美国政府持有的英特尔股份市值已超过500亿美元。（新浪财经）国家药监局：我国4月批准注册医疗器械产品318个今天从国家药监局获悉，我国4月共批准注册医疗器械产品318个。其中，境内第三类医疗器械产品272个，进口第三类医疗器械产品20个，进口第二类医疗器械产品26个。 (央视新闻)本文由「卜算籽」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：AI生成

8点1氪丨贾跃亭又融资4.77亿元；巨力索具旗下十余家企业已注销；超越茅台，A股“新股王”诞生8点1氪·2026年05月19日 08:08贾跃亭发布视频，宣布法拉第未来近期完成了7000万美元机构投资者募资。今日热点导览瑞幸咖啡正式上线两款含酒精的特调饮品京张高铁试点推出“自行车随身行”服务马斯克：计划今年进行首例用于治疗失明的Neuralink植入手术韩国法院要求三星电子罢工行动不得影响产量多家理财公司集中提前终止理财产品TOP 3 大新闻贾跃亭又拿到7000万美元投资贾跃亭发布视频，宣布法拉第未来近期完成了7000万美元（现汇率约合4.77亿元人民币）机构投资者募资，并将启动战略、产品技术&业务、财务、资本、AI体系五大变革。贾跃亭表示，其目标包括FFAI市值重回2021年上市水平，2027年第四季度实现经营性现金流为正，五年内在EAI人形与仿生机器人领域成为北美真实场景部署量前三的机器人公司+北美EAIMPV领军者。贾跃亭还表示，要将FF真正重构为一家Physical AI生态公司。两年内实现过去12年没有完成的梦想，进入“0借口、0内耗、结果导向”的战斗状态，全力兑现对股东和用户的承诺。（IT之家）巨力索具旗下十余家企业已注销36氪获悉，近日，巨力索具因涉嫌信息披露误导性陈述违法违规，被证监会立案调查。天眼查App显示，巨力索具股份有限公司成立于2004年12月，法定代表人为杨建国，注册资本9.6亿人民币，经营范围包括应急装备、装备专用工装及工具和箱组的设计、生产和销售等，由巨力集团有限公司、杨建忠、香港中央结算有限公司等共同持股。对外投资信息显示，该公司对外投资26家企业，其中10余家企业处于存续状态，包括巨力索具（河南）有限公司、河北巨力应急装备科技有限公司等，其他10余家企业已注销。联讯仪器股价反超贵州茅台，成为A股新的股王5月18日，联讯仪器盘中扩大涨幅至13%，股价反超贵州茅台，成为A股“新股王”，该股较发行价上涨超15倍。截至5月18日11时许，联讯仪器最高涨至1361元/股，涨幅高达16.32%。与此同时，贵州茅台由涨转跌，股价探至1330元之下。此前相继登上A股“股王”宝座的寒武纪（688256）、源杰科技（688498），5月相继迎来2025年年度权益分派。公开资料显示，联讯仪器4月24日登陆A股，发行价81.88元/股，上市14个交易日来，目前股价较发行价上涨15.6倍。据公开资料显示，联讯仪器是国家级专精特新“小巨人”企业，专注于高速通信、电性能、光芯片及半导体晶圆芯片等领域测试仪表与设备的研发、制造，是国内高端电子测量与半导体测试设备领域的硬核龙头。（澎湃新闻）大公司/大事件何立峰会见美国超威半导体公司董事会主席兼首席执行官苏姿丰5月18日下午，中共中央政治局委员、国务院副总理何立峰在人民大会堂会见美国超威半导体公司董事会主席兼首席执行官苏姿丰。何立峰指出，上周中美两国元首在北京举行重要会晤，达成一系列重要共识。在两国元首战略引领下，两国经贸团队达成了总体平衡积极的成果，这将为下一步的中美经贸合作与世界经济注入更多确定性和稳定性。欢迎包括超威半导体公司在内的跨国公司把握中国发展机遇，深化互利合作。苏姿丰积极评价两国元首会晤成果，表示愿继续拓展在华业务，持续加大在华投资。(新华社）美陪审团驳回马斯克对OpenAI的诉讼当地时间5月18日，美国陪审团裁定，美国开放人工智能研究中心（OpenAI）及其首席执行官萨姆·奥尔特曼无需就“偏离慈善使命”一事对埃隆·马斯克承担责任，原因是马斯克提起诉讼时间过晚。当天晚些时候，马斯克表示，将就这一裁决向美国第九巡回上诉法院提起上诉。此前，马斯克对OpenAI及其首席执行官奥尔特曼提起诉讼，指控OpenAI最初邀请马斯克资助一个致力于造福人类的非营利AI研究机构，但如今却专注于盈利。（央视新闻）张朝阳：搜狐未入局大模型研发，将聚焦AI应用落地36氪获悉，在2026搜狐科技年度论坛上，搜狐CEO张朝阳表示，“搜狐并未入局大模型第一阵营的研发，大模型研发需要雄厚的资金实力和技术背景，企业应基于自身业务基础理智选择。”他提到，搜狐将聚焦AI应用落地，在提升编程效率、节约运营成本的同时，对内容生成保持克制，保持内容中立性，规避AI带来的内容乱象风险。腾讯音乐：收购喜马拉雅的事项已经完成36氪获悉，腾讯音乐在港交所公告，于2026年5月18日，根据并购协议的条款，收购喜马拉雅的事项已经完成，据此，喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为12.6亿美元的现金及最多总数1.75亿股本公司A类普通股（包括本公司于交割时新发行的A类普通股，以及本公司授出的以股权为基础的激励相关的A类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。川投新能源公司总经理刘亮被遣返回国据中央纪委国家监委网站，近日，在中央反腐败协调小组国际追逃追赃工作办公室统筹协调和公安部等部门协助下，经四川省监察机关、公安机关与有关国家执法机关密切合作，外逃职务犯罪嫌疑人刘亮在境外落网并被遣返回国。刘亮，男，1979年1月出生，川投新能源公司总经理，涉嫌严重职务违法犯罪，于2026年4月外逃，同月四川省资阳市监察机关对其立案调查。办案机关积极开展国际执法合作，于近日将其缉捕归案。（界面新闻）英特尔CEO称代工业务取得进展英特尔CEO陈立武在当地时间周一接受采访时表示，该公司的外部代工业务正在稳步发展，成为其扭亏为盈的关键一环。“代工业务非常重要，”陈立武表示，“它是重要的国家资产之一。”（新浪财经）瑞幸咖啡上线两款含酒精的特调饮品5月18日，瑞幸咖啡正式上线两款含酒精的特调饮品，并且属于饮料酒的范畴，正式布局“微醺”经济。记者了解到，这两款含酒精饮品是瑞幸对全国市场的一次测试，后续或还有进一步的计划。在业内看来，瑞幸布局微醺赛道，意在盘活闲置门店时段资源，提效增收。记者在北京部分瑞幸门店看到，两款年度特调饮品已经上线，但只有部分门店可以提供含酒精版。（红星新闻）韩国法院要求三星电子罢工行动不得影响产量据央视新闻消息，韩国水原地方法院5月18日批准了三星电子公司提出的部分禁令请求，责令该公司工会必须确保即将开始的全面罢工行动“不影响产量”，“不得导致这家全球重要的存储芯片制造商生产原料受损”。据悉，这实际上是给三星电子这次史上最大规模罢工计划“踩了急刹车”。今年3月中旬，由超过6.6万名三星电子工会成员参与表决的投票结果显示，93.1%的工会成员赞成罢工。全面罢工将于5月21日至6月7日举行。据悉，三星电子作为全球重要的存储芯片制造商，一旦发生罢工，可能进一步加剧由全球人工智能数据中心建设持续升温带来的全球半导体供应趋紧局面，其影响或波及汽车、计算机及智能手机等多个行业。三星电子劳资双方此前就2026年薪资问题进行了多轮谈判，但双方就部分议题的分歧持续扩大，谈判最终破裂。（央视新闻）多家航司下调国际航线燃油费，国内航线刚上调国内航线燃油附加费自5月16日再度上调后，多家航司却在陆续下调国际航线的燃油附加费。香港航空宣布，自今日（5月18日）开始下调多条航线的燃油附加费。其中中国香港前往内地的地区航线，燃油附加费由290港元下调至190港元；香港前往日韩、新马泰等则由389港元下调至339港元。在此之前，国泰航空和香港快运已宣布自5月16日起下调部分区域的燃油附加费。其中香港始发长途航班从1560港币下调至1362港币，中程航班从725港币下调至633港币，短途航班从389港币下调至339港币。下调幅度在12%左右。这也是中东冲突爆发后，多家航司首次对核心航线燃油附加费进行下调。（第一财经）5月19日起，铁路部门在京张高铁试点推出“自行车随身行”服务据新华社消息，为更好地满足旅客多样化出行需求，铁路部门5月19日起在京张高铁北京北至崇礼间各车站间运行的G字头动车组列车试点推出“自行车随身行”服务，旅客可通过12306客户端线上预约并付费，乘车时经车站安检和规范包装，可携带符合规定的自行车同车出行，将为广大骑行爱好者跨区域出行提供更多便利。试点初期，“自行车随身行”服务定价88元/辆，旅客可在乘车结束后180日内通过铁路12306客户端开具电子发票。（新华社）戴尔将与谷歌和SpaceX等公司展开合作戴尔科技在上一季度为其“AI工厂”产品线新增1,000家客户，使其总客户数达到了5000家。戴尔正在推出多款新产品以协助企业搭建AI应用，其中包括一款名为“Dell Deskside Agentic AI”（戴尔桌面智能体AI）的产品。此外，戴尔还与谷歌和SpaceX等公司展开合作，将AI模型与工具引入企业自身的内部网络中。（财联社）SpaceX将迎关键试飞SpaceX计划于周二进行一次关键的试飞，发射其星舰巨型火箭的升级版，代号V3。这次发射正值SpaceX上市关键阶段，消息透露预计该公司最早在下月IPO，并可能在本周三公开招股书。与此同时，SpaceX在与贝索斯的蓝色起源竞争，为NASA的阿尔忒弥斯计划开发星舰，供其在2028年的载人登月任务中使用。目前来看，星舰的研发进度落后于NASA对SpaceX的预期。该火箭原计划于2023年首飞，但去年的失败延缓了其发展进程。星舰最近一次试飞，也是其第11次试飞，发生在七个月前。（财联社）NextEra拟以660亿美元收购Dominion，或创史上最大规模电力并购纪录5月18日消息，据报道，美国公用事业巨头NextEra Energy目前正就收购竞争对手Dominion Energy进行谈判，该笔交易方案主要以股票交换形式进行，对Dominion的估值约为每股76美元，总价约660亿美元。若最终达成，这将成为迄今为止电力行业规模最大的一笔交易。NextEra计划以约0.8股自身股票交换1股Dominion发行在外的股票，交易总额中还将包含一小部分现金成分。交易完成后，NextEra的股东将持有合并后新公司约75%的股份。（界面新闻）马斯克：计划今年进行首例用于治疗失明的Neuralink植入手术马斯克表示，旗下Neuralink技术能帮助部分失明者恢复视力，计划今年进行首例用于治疗失明的Neuralink植入手术。据了解，Neuralink是一个美国神经科技和脑机接口公司，由伊隆·马斯克和八名其他联合创办者创办，负责研发植入式脑机界面技术。公司的总部在旧金山。Neuralink于2016年成立，并于2017年3月公之于众。2019年7月17日，Neuralink对外宣布该公司研发的一款脑机接口系统（财联社）五粮液：2025年度百亿分红两月内到账，严格兑现200亿元分红承诺5月18日，五粮液2026年第一次临时股东会对《2025年度利润分配方案》进行了审议，五粮液副董事长、总经理华涛表示，2025年度分红资金将在两个月内准确、及时划入全体股东账户。五粮液2025年中期已分红100亿元，本次拟再分红100.07亿元。五粮液践行高比例分红，严格兑现《2024—2026年股东回报规划》承诺，连续3年每年度现金分红总额不低于200亿元（含税）。（证券时报）美国连锁快餐品牌Wendy’s将进入中国市场5月18日在美国连锁快餐品牌Wendy’s官网上看到，该品牌国际特许经营重点市场亚太区域中出现了中国市场。早前，Wendy’s在2026年第一季度的财报中称，公司已经签署一项特许经营协议，计划未来10年在中国建设最多1000家餐厅。Wendy’s称合作方是“一家在中国拥有数十年经验的大型餐饮运营商，以强待客能力闻名、深刻理解中国消费者。”但没有公布这家运营商的名称。此外，财报显示，Wendy’s2026年一季度全球销售额同比下降5.5%，美国同店销售下降7.8%，国际市场销售额增长6.0%。（界面新闻）贵州茅台：优化茅台酒线下门店营业时间及i茅台App开售时间36氪获悉，据贵州茅台消息，为进一步贴近广大消费者日常工作与生活节奏，更好地满足广大消费者购酒需求，自2026年5月

高瓴出手了一家AI体育科技公司，曾获李泽湘天使轮融资｜硬氪首发黄 楠·2026年05月19日 09:30从“记录工具”到“AI运动Copilot”的跃迁。作者｜黄楠编辑｜袁斯来硬氪获悉，AI体育科技公司SportVision近日已完成天使+轮融资，高瓴创投（GL Ventures）领投，谦恒资本担任独家财务顾问。本轮资金将重点用于核心技术研发、产品量产落地及市场的拓展推广。其新款AI网球训练相机产品计划发售中。SportVision隶属于深圳思博威视体育科技有限公司‌，长期专注运动算法研发与硬件产品应用、构建AI智能教练体系，此前已获得李泽湘教授旗下清水湾基金的天使轮投资。过去几年，运动科技赛道几乎被硬件厂商重新定义了一遍。发球机让“一个人练球”成为可能，运动相机令“训练过程的记录”不再需要专人跟拍，智能手表/手环把心率、步频、消耗卡路里等量化成精确的长期数据记录。尽管每一类产品都跑出了规模不小的公司，但它们各自停留在陪练、记录、监测的单点功能上——数据相互独立、能力无法贯通，难以为用户提供连贯、整体的个性化支持。运动人群的深层渴望远不止于被记录，还有被指导、被陪伴、被看见进步。一场更底层的变革正在发生：从“记录工具” 到 “AI运动Copilot”的跨越，运动科技赛道正在重写。工具只需完成指令，而Copilot则可以像教练一样理解状态、识别动作、纠正习惯，甚至在运动热情消退时成为那个督促陪伴的人。要补上这道体验鸿沟，用户需要的不只是更强性能的硬件，更是一个能把整合所有信息、真正理解运动者的“大脑”。这也正是SportVision精准切入的核心位置。全运会现场所使用的SportVision设备（图源/企业）SportVision创始人陈楷夫本科毕业于吉林大学物理学院严济慈班、哥伦比亚大学机器人学硕士，曾在MMLab从事姿态识别与目标检测算法研究；身为一名前游泳运动员，他对运动学习、动作纠正与训练效率的痛点有切身感知，正是这段经历，让他决心用AI技术破解非结构化运动数据难题，构建一个可规模化、可复制的个性化运动产品和训练体系。联合创始人杨晨旭是资深网球爱好者，曾在腾讯、字节负责AI产品开发。多位团队核心成员来自商汤、字节、阿里、腾讯等头部科技企业，具备AI算法、软硬件一体化研发与全球化品牌运营能力，已形成了从技术研发到产品落地的完整闭环。公司起步初期，SportVision选择从羽毛球场景切入进行关键产品验证，首款产品“好球哇”是一套羽毛球AI运动相机方案。设备可实时采集球场画面，通过视觉算法自动识别并捕捉用户的运动高光时刻，为普通运动爱好者、专业运动员及赛事场景，提供高光集锦、赛事直播、运动数据分析等一体化服务。这套系统落地了全国数十家头部羽毛球馆，并获得林丹、王睁茗、龚睿那等世界冠军在内的多位种子用户支持。在赛事侧，该系统也服务了全运会残特奥会、粤港澳大湾区羽毛球比赛、深圳杯等多场关键赛事。但它真正的价值，远非硬件本身，而是持续运行中沉淀的海量运动数据。在拍摄过程中，SportVision自研算法能实时分析并智能预测运动轨迹，根据比赛进程自动调整摄像角度，对复杂的多人运动场景进行针对性学习优化。从世界冠军到刚入门的爱好者，从专业教练的指导语料到普通用户的练习习惯，场馆设备每天运行十几个小时，能收集到各类各样的泛化数据。每一场球、每一次挥拍、每一个得分瞬间，都在被系统记录、拆解、分析。这是一座难以被发球机或运动相机采集到的数据矿。运动学习的基本机制分为三个阶段：认知阶段（通过视觉掌握动作要领）、联想阶段（在反复练习中形成神经通路）、自主阶段（动作固化成为肌肉记忆）。其中联想阶段的及时反馈最为关键：动作做对还是做错需要在几秒内被确认，否则错误动作就会被“练熟”。陈楷夫向硬氪指出，“此前，多数运动AI解决方案仍停留在‘事后分析’阶段，用户完成运动后上传视频、系统给出复盘报告。但这种模式错过了最佳纠正窗口。”因此，SportVision想搭建的，是一套基于实时交互的预判式AI。“好的运动员不一定是好教练，好教练的价值在于知道怎么教。”陈楷夫说，“我们要做的，是把教练的教法萃取出来，让AI学会怎么教。”随着网球、羽毛球、匹克球等拍类运动在全球范围内持续升温，用户对运动提升的需求愈发迫切。传统的请教练模式面临成本高、频次低的矛盾，一周请一两次教练，用户练习时无法得到反馈，对于向网球这样的运动，可能需要一年才能从入门到实现“对打社交”。正反馈周期长，也使得很多人停留在了入门阶段。这正是SportVision的核心壁垒所在：那些在球场上日夜运转的设备，收集到的海量数据。从骨骼点识别到动作模式分析，从专业运动员的动作库到普通用户的练习轨迹，这些数据构成了可迁移的技术底座，能够在各类拍类运动中得以技术复用和迁移。SportVision首款网球AI Copilot产品（图源/企业）“我们做的不是一款运动相机，而是构建未来运动人群个性化AI教练的世界模型。”陈楷夫告诉硬氪，“用户甚至不需要思考‘我该怎么练’，而是由系统主动管理你的运动进步。”硬氪获悉，SportVision首款网球AI Copilot产品正在紧密的研发流程中。以相机为载体，系统可识别用户运动动作并与专业标准动作比对，实时反馈纠正建议；同时，结合教练指导场景，生成个性化训练方案，让用户在独自练习时也能获得针对性、科学化的AI陪练指导。该产品概念此前已在CES 2026上展出，首发市场聚焦欧美地区。这里拥有了全球密度最高的网球用户群体，也是付费意愿与消费能力最突出的核心市场。回到AI时代运动科技竞争的底层逻辑迁移，当大模型逐渐消解知识鸿沟，竞争的壁垒不再停留于算法参数，而是向两端迁移——一端是数据的密度与深度，另一端是场景的介入与闭环。SportVision长期积累的稀缺数据资产，正在成为其下一代AI教练产品的技术底座。从情绪价值到功能性与情绪价值兼具、从记录到指导，从单一品类到拍类运动通用技术底座，这家公司正在尝试的命题是，加速推动运动科技从“工具”到“AI Copilot”的关键升级。本文由「黄 楠」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：企业授权、视觉中国

6小时前法官确认裁决结果，驳回马斯克对 OpenAI 的诉讼美国加州北区联邦地方法院法官 Yvonne Gonzalez Rogers 近日确认最新裁定结果：驳回埃隆·马斯克要求对 OpenAI 和微软签发“全面禁制令”的动议与其先前提出的欺诈类指控，但保留关于“是否存在并被违反的公益/慈善信托”以及“不当得利”等核心争议，案件不会就此终结，而是将进入加速审理并由陪审团作出最终裁决。3 来源法官确认：禁制令与部分指控被否，核心诉求仍将审理3 来源从出资联合创办到天价索赔：马斯克与 OpenAI 的纠纷缘起3 来源法院为何驳回禁制令与部分指控：证据不足与适格障碍3 来源影响与展望：AI 治理、非营利结构与科技巨头关系迎来“法律考题”3 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。智库·502线上同行往期回顾报名已结束企业构建AI+组织的人效杠杆与路径选择百果园集团x银泰分享企业构建AI+组织的人效杠杆与路径选择妙投波动判断场往期回顾活动已结束芯片、光模块、机器人——AI的泼天富贵，还能传到第几棒？本期我们将AI产业链拆解为三大核心环节，逐一拆解其核心矛盾与风险边界，帮你建立从芯片到机器人的完整坐标系。随行灯塔火热报名中随行灯塔随行灯塔是由虎嗅深度联合运营的高端创始人社群。专注于一件事：帮助创始人把产品做对、做爆、做到10亿。

这里不灌鸡汤，不讲玄学。我们相信，伟大的公司一定是产品驱动的；但产品力来自人，而人需要场——一个能被照见、能被挑战、能被链接的场。随行灯塔，就是创始人的“修炼场”。智库·502线下闭门会往期回顾精彩回顾AI在ToB销售、客服的落地案例探讨围绕着AI在ToB营销内容生成、智能客服接待、线索清洗等场景进行深入交流早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

关注最新看重点大公司财经动态看世界硅谷动态莫言财经财经资讯，知名财经评论信息收集华泰证券：需求旺盛下半导体材料景气或加速。华泰证券研报称，据5月17日长鑫科技招股书，受益于算力需求持续增长、DRAM产品涨价，2026年上半年预计长鑫科技实现归母净利500亿元—570亿元，存储行业正快速增长，华泰证券认为上游材料亦或受益。伴随当前晶圆厂、存储企业等扩产，以及芯片向新型存储技术、先进制程、先进封装等发展，全球半导体材料市场有望迎来快速增长。当前我国半导体材料国产化率整体仍较低，伴随自主可控要求提高、国内企业产品竞争力提高，国产化率有望提高。6分钟前00举报发表0/500虎嗅用户社区交流公约时讯内参树欲静而风不止，新鲜资讯速递老人在公交车上闭眼“休息”，实则已死亡，救助方被索赔逾11万。老人在公交车上突发疾病，司乘人员尽管立即组织乘客，拨打120急救电话及110报警电话，老人还是不幸去世。承运人是否需要承担法律责任？近期，上海松江法院审结了这样一起城市公交运输合同纠纷案。2024年12月的一个早晨，年逾八旬的李老伯像往常一样乘坐公交车出门，该公交车由上海某公交公司运营。李老伯上车就座后约两分钟，开始擦汗、喘气，尚未来得及购买车票就闭眼靠椅背似休息状。售票员在卖票过程中多次轻拍李老伯肩膀，发现其始终没有反应，售票员和其他乘客遂先后探查李老伯呼吸和颈动脉脉搏，才发现李老伯陷入异常状态，售票员立即组织乘客拨打120急救电话及110报警电话，并亲自与急救中心保持沟通，清晰说明现场情况与车辆位置。与此同时，司机将车辆主动驶至约定的佘山派出所门口——该地点便于救护车快速识别和会合。从发现异常到车辆抵达会合点，全程约8分钟。救护车到达后，售票员积极协助急救人员将李老伯抬上救护车抢救。遗憾的是，李老伯经抢救无效死亡。李老伯的配偶及两个女儿认为，公交公司未在公交车上配备必要的急救设备（如AED、急救药品等），未第一时间实施有效救助措施，事后亦未能及时提供车内监控，也缺乏对家属的人道关怀，加重了家属失去亲人的痛苦，遂将公交公司诉至法院，要求公交公司赔偿医疗费、丧葬费、死亡赔偿金等损失9万余元，并要求支付精神损害抚慰金2.5万元。公交公司辩称，李老伯上车后尚未购买车票，双方运输合同关系未成立。李老伯的死亡系其自身健康原因直接导致，与公交公司的运输行为无因果关系。公交公司已经在其能力范围内积极履行了合理的救助义务，不存在违约行为，不应承担赔偿责任。松江法院经审理认为本案争议焦点有二，一是案涉运输合同是否成立，二是公交公司在合同履行中有无违约行为。关于合同是否成立的问题，虽然李老伯因突然昏迷而客观上无法完成购票行为，但自其登上公交车并接受运输服务时起，双方之间的城市公交运输合同关系即已依法成立。关于公交公司有无违约行为。《中华人民共和国民法典》第八百二十三条规定：“承运人应当对运输过程中旅客的伤亡承担赔偿责任；但是，伤亡是旅客自身健康原因造成的或者承运人证明伤亡是旅客故意、重大过失造成的除外。”本案中，李老伯的死亡原因经诊断为猝死，系其自身健康原因，与公交公司的客运行为并无因果关系。公交公司的售票员从察觉李老伯处于异常状况至与120交接病患，时间间隔为8分钟左右，已在其能力范围内履行了救助义务，不存在违约行为。据此，松江法院判决驳回了原告的诉讼请求。判决后，双方均未上诉，该判决已生效。7分钟前00举报发表0/500虎嗅用户社区交流公约莫言财经财经资讯，知名财经评论信息收集中信建投：Figure机器人直播分拣，关注垂类应用放量拐点。中信建投证券研报称，Figure机器人直播分拣，关注垂类应用放量拐点。Figure03全天候直播包裹分拣任务，这表明人形机器人正在从实验室加速走向商业化应用阶段，人形机器人板块的核心已从主题投资向量产预期演进，机器人在工业、商业、家庭等场景的落地进程是当前资本市场关注的重点。F.03连续分拣超过80小时、处理超过10万个包裹，这是人形机器人工业场景落地应用的重要考验，证明其在自主性、连续性、可靠性方面得到了明显提升。今年7—8月特斯拉Optimus项目有望启动量产，后续产品发布、量产推进等值得重点关注；同时，宇树科技等国产机器人公司IPO持续推进，国产本体厂商有望估值重估，当前建议把握核心环节。9分钟前10举报发表0/500虎嗅用户社区交流公约一只飞虎最近国外发生了什么？刚果（金）埃博拉疫情扩散，116例疑似病例死亡，另有435例疑似病例正接受观察和治疗，相关实验室检测和流行病学调查仍在进行。刚果（金）政府18日说，该国东部伊图里省暴发的新一轮埃博拉疫情仍在扩散，目前已报告116例疑似病例死亡，另有435例疑似病例正接受观察和治疗，相关实验室检测和流行病学调查仍在进行。刚果（金）新闻和媒体部长兼政府发言人穆亚亚在发布会上说，伊图里省及邻近的北基伍省多地已报告病例。刚卫生部门表示，相关检测和分析仍在进行，以确认报告的死亡疑似病例是否均与埃博拉病毒有关。此前，刚卫生部15日说，截至当天，全国共有80例死亡病例，其中4例死亡病例经检测确认感染埃博拉病毒。非洲疾病预防控制中心15日宣布，刚果（金）伊图里省暴发埃博拉疫情。同一天，乌干达卫生部说，一名输入性确诊病例在该国死亡。世界卫生组织17日宣布，刚果（金）和乌干达出现的埃博拉疫情已构成“国际关注的突发公共卫生事件”。卢旺达、赞比亚、坦桑尼亚等周边国家已相继宣布加强边境卫生监测、入境筛查和疫情防控准备。埃博拉病毒可引发埃博拉出血热，主要通过接触病患或被感染动物的血液、体液、分泌物、排泄物等感染，临床表现主要为发热、出血和多脏器损害，死亡率高达50%至90%。12分钟前00举报发表0/500虎嗅用户社区交流公约回应来了事事有回应，热点追踪周大生回应去年关店五百多家。5月18日，周大生召开业绩说明会。有投资者提问：针对去年公司关店五百多家，主要受哪些因素影响，预计今年关店情况是否会好转？周大生回应称，第一是销售偏弱，第二是加盟商店面合并，小店合并，因为公司的加盟商数量并没有减少。黄金首饰消费好转，关店情况也会好转。14分钟前00举报发表0/500虎嗅用户社区交流公约飞行圈航空领域靠谱资讯速递~美两架“咆哮者”战机空中相撞坠毁，4名飞行员弹射逃生，五角大楼展开调查；系“咆哮者”近3年第3次重大坠机事故，曾训练时被闪电击中。据环球时报援引路透社17日称，美国爱达荷州芒廷霍姆空军基地17日举行飞行表演时，两架EA-18G“咆哮者”电子攻击机在空中相撞并坠毁，4名机组人员弹射逃生。相关统计显示，这已经是“咆哮者”最近3年的第三次重大坠机事故。17日，两架“咆哮者”空中相撞后，4名飞行员弹射。（“动力”网站“战区”频道） 美国“动力”网站“战区”频道17日称，事发时，这两架“咆哮者”正在进行飞行表演，现场视频显示，其中一架“咆哮者”从后方靠近另一架，随后两者相撞。“令人惊讶的是，两架战机似乎缠绕在一起，旋转着落向地面。”视频中可以看到4名机组人员弹射逃生，“两架‘咆哮者’坠地化为一团火焰，随后机组成员的降落伞飘落到地面”。据芒廷霍姆空军基地发言人称，坠机事故引发了火灾，但后来被控制住了。随后当地警方宣布整个飞行表演活动取消，并警告公众避免前往该基地。美国航空专业网站“The Aviationist”称，目前五角大楼正在对事故原因展开调查。发生空中相撞事故的两架“咆哮者”来自美国海军“维京人”中队（VAQ-129），该中队负责所有“咆哮者”飞行员的训练。作为当前美国海军唯一的专业电子攻击机，“咆哮者”由“超级大黄蜂”舰载战斗机改装而来，其机体90%的部件与“超级大黄蜂”通用，保留了空战能力的同时，具备突出的电子战能力。该机采用双座配置，前座为飞行员，后座为电子战操作员。为了展示该机的性能，2020年美国海军在华盛顿州惠德比岛海军航空站组建“咆哮者”演示队，该演示队的成员和飞机均来自“维京人”中队。值得注意的是，近年来“咆哮者”电子攻击机频繁发生严重事故。按照五角大楼的规定，造成超过200万美元损失或导致人员伤亡的事故，就属于最严重的A级事故。2023年6月6日，驻日美军的3架“咆哮者”在日本南部执行飞行训练时，同日先后被闪电击中，虽然没有坠机，但被列为A级事故。2024年10月15日，一架“咆哮者”在美国华盛顿州执行训练任务时，坠毁在雷尼尔山东部山区。该机完全损毁，两名飞行员死亡。2025年2月12日，一架“咆哮者”在日常训练中突发事故，尝试降落失败后坠入美国加州圣迭戈湾，两名飞行员弹射逃生，飞机完全损毁。这两起坠机事故的原因还在调查中。26分钟前00举报发表0/500虎嗅用户社区交流公约AI新技能机器人们又get了什么新本领？iOS27可自动生成快捷指令。彭博社昨日5月18日发布博文，报道称苹果公司正测试iOS27新版Shortcuts快捷指令应用，用户只需用自然语言描述需求，系统就会自动生成并安装对应快捷指令。彭博社表示新版快捷指令应用的核心变化，就是把手动搭流程改成直接说需求，用户输入一句话，系统就会自动生成并安装快捷指令。IT之家注：快捷指令并非新产品，苹果公司在2017年收购Workflow，随后在2018年将其发展为快捷指令曾名捷径。该应用一直被视为iOS上的强大自动化工具，能连接应用、文件和信息，并把多个步骤串成一个动作。但对于普通用户而言，学习门槛不低，因此在使用时现在要么在应用里手动拼装动作模块，要么从Apple的指令库下载现成模板，而真正长期使用的人始终集中在少数进阶用户群体。而本次测试版重点降低用户的使用门槛，更新后的应用会先向用户抛出一个问题：你想让这个快捷指令做什么，用户随后可在文本框中使用自然语言直接描述需求，也支持语音表达。新方案则试图跳过中间学习过程，让用户先说结果，再由系统反推执行步骤。为了降低使用门槛，这种设计把重点从怎么搭转向想完成什么，更接近日常软件的自然交互逻辑IT之家26分钟前00举报发表0/500虎嗅用户社区交流公约AI新技能机器人们又get了什么新本领？地平线开源4亿参数机器人小脑大模型HoloMotion-1。5月18日，地平线宣布开源HoloMotion-1，这是地平线机器人实验室面向人形机器人全身控制推出的4亿参数级机器人小脑大模型。据介绍，HoloMotion-1是HoloMotion技术路线的第一阶段，重点解决模仿任意姿态(Imitate Any Pose)，即从视频、MoCap和遥操作数据中学习复杂全身动作。29分钟前00举报发表0/500虎嗅用户社区交流公约钢铁侠马一龙Tesla、SpaceX…马斯克败诉，OpenAI有望扫清IPO障碍。美国加利福尼亚州奥克兰市联邦法院陪审团于周一做出裁决，一致认定埃隆·马斯克对OpenAI及其首席执行官萨姆·奥尔特曼的诉讼因已超出诉讼时效而被驳回。这场历时三周的庭审被普遍视作决定OpenAI乃至整个人工智能行业未来走向的关键节点，关乎人工智能的应用方向与利益归属问题。此次裁决扫清了OpenAI 推进首次公开募股（IPO）的一大障碍，其上市估值有望达到1万亿美元。马斯克随后表示自己将会提起上诉，并再度指责奥尔特曼与 OpenAI 总裁格雷格・布罗克曼，将 OpenAI 视作敛财工具。34分钟前11举报发表0/500虎嗅用户社区交流公约安全小喇叭重大灾害事故预警、播报、跟进宣恩县发布情况通报。情况通报 5月17日至18日，恩施州宣恩县沙道沟镇白水河村上游普降大到暴雨，最大降雨量达292.6毫米，导致河水暴涨，白水河村沿河区域多处房屋被淹、部分房屋倒塌、道路通讯中断。汛情发生后，恩施州和宣恩县两级防汛抗旱指挥部迅速启动应急响应，第一时间组织武警、公安、应急、消防、交通、电力、通讯等部门赶赴现场，紧急转移群众287人，全力开展人员搜救、医疗救治、道路疏通、通讯抢修等应急救援工作。截至5月18日20时，灾情已造成3人死亡、4人失联。应急救援工作还在全力开展中。36分钟前10举报发表0/500虎嗅用户社区交流公约读数数字里的故事澳洲联储担忧能源涨价或将快速推高居民物价。澳大利亚联储担忧，在本国经济承压

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文GPT-5.5才发三周，5.6已在内测！OpenAI与Anthropic补贴大战同日开打，开发者坐收渔利发布于AI新闻资讯发布时间 :2026年5月14号 14:21阅读 :1分钟三周前GPT-5.5刚落地，GPT-5.6的消息已经提前炸出来了。知名爆料人Leo透露，GPT-5.6目前开发进度已全速推进，首批内部检查点在过去几天已启动测试，下个月很可能就会正式露面。更有意思的是，有人从OpenAI内部Codex日志里扒出了rollout mapping的痕迹——大部分调用还是指向GPT-5.5，但有一条记录赫然落在了GPT-5.6身上。换句话说，Codex环境里可能已经在偷偷用它跑测试了。内部代码名也被顺带曝光:ember-alpha和beacon-alpha。网友直呼:这迭代速度，根本追不上。速度再度拉满——Codex即将上线ultrafast模式GPT-5.6还没到，OpenAI产品端已经先按捺不住了。爆料人Chetaslua同步透露，本周四OpenAI将在Codex上线"ultrafast模式"，响应速度直接提升2到3倍，专门为延迟敏感型任务设计。与此同时，Image Arena排行榜上以+242分断档领先的gpt-image-2，也在同步推进A/B测试更新。这不是第一次了。今年3月GPT-5.4发布时，/fast模式已经实现了1.5倍加速;GPT-5.3-Codex-Spark借助Cerebras芯片，更是把推理速度拉到每秒超过1000token，是普通模式的15倍。而这次的ultrafast，直接在旗舰主力模型上实现提速——不是阉割版，不是小模型凑数，是正儿八经的全尺寸加速。对开发者来说，Agent循环、长任务流水线、浏览器自动化……所有需要"等"的地方，体验将彻底不同。顺带一提，上周ChatGPT官方账号发出一张科幻感十足的图，界面上那句经典的"Ask ChatGPT"悄悄换成了"Message ChatGPT"。奥特曼丢下一句"Call me maybe"，让外界对ChatGPT集成语音通话乃至推出硬件产品的猜测，彻底烧开了。Codex vs Claude Code，补贴战正式开打硅谷最戏剧性的一幕，就在今天上演了。就在OpenAI ultrafast模式上线前夕，Anthropic率先出手——宣布从6月15日起，为付费用户每月提升50%的编程额度，覆盖Claude Agent SDK、命令行工具以及深度接入GitHub工作流的Claude Code，并同步发布了Opus4.7Fast模式，主打比Codex更快的高级推理与更流畅的长上下文编码体验。OpenAI没让人等太久。反手祭出一招重磅补贴:接下来30天内，任何想从其他平台迁移到Codex的企业，直接获赠2个月免费使用权。按Pro计划每月200美元算，相当于白送400美元。奥特曼本人亲自下场喊话，直言Codex是"目前市面上最强的AI编程产品"。仅仅几个小时后，战报出炉:2000名开发者在3小时内主动联系了OpenAI。这场怼脸硬刚，看呆了整个硅谷。开发者们倒是乐了——被两家顶级公司抢着补贴，这种好事，上哪儿说理去。飞轮已经开始自转把这两件事放在一起看，会发现一个比任何单一新闻都更值得关注的趋势正在成形。AI正在加速自我进化。GPT-5.3-Codex是OpenAI首个"参与自身训练"的模型，到GPT-5.5，OpenAI内部85%的员工每周都在用Codex写代码。GPT-5.6几乎可以确定是在GPT-5.5的深度参与下开发出来的——AI在帮OpenAI造更强的AI。与此同时，编程工具的大规模普及正在释放前所未有的工程产能。Codex已有300万周活跃用户，Claude Code的用户数同样在爆发式增长。数百万开发者把AI编程工具当成日常标配，AI生成的代码又反哺回AI的训练与部署，这个循环只会越转越快。奥特曼说过，OpenAI的目标已经不止于AGI，直接瞄准了ASI。当模型迭代速度由AI自身驱动，当AI编程工具成为AI开发的基础设施，当两家万亿级公司用"补贴战"加速这一切的普及——通往ASI的飞轮，已经开始自转了。

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文李开复：开源模型是实现“AI主权”的更优路径发布于AI新闻资讯发布时间 :2026年5月12号 18:01阅读 :1分钟创新工场董事长李开复近日在专访中深入探讨了“AI主权”的概念。他指出，人工智能主权不仅关乎技术控制权和数据安全，还涉及模型与本土文化及法律的适配性。他认为，并非每个国家都有必要重造一个OpenAI，盲目追求昂贵的闭源自研并非良策。李开复强调，对于大多数资源有限的国家和企业而言，基于开源模型构建本地化体系是更现实的选择。这种“第三条路”通过在开源基础上进行增量训练，能以极低的成本让模型符合特定地区的价值观，而非受制于他国的通用模型。工程效率与生态之争谈及中美竞争，李开复认为资源受限反而催生了中国企业的极致工程效率。他以DeepSeek为例，指出中国公司在算力不足的情况下，通过优化架构和减少依赖，能以不到美国十分之一的投入实现极高的性能。他预测AI领域将演变成类似“iPhone与安卓”的格局。闭源模型如同一方领地，追求高利润和生态控制力；而开源模型则代表了更广阔的市场覆盖和普及度，两者将在全球范围内长期并存，共同构建未来的智能生态。中国硬件的下半场机遇李开复对中国在物理AI和硬件制造方面的结构性优势表示乐观。他指出，中国拥有全球最完整的供应链体系，从智能穿戴到机器人，制造成本远低于西方。未来AI时代的标杆级硬件，极有可能由中国设计并品牌化。此外，中国在语音驱动交互和大规模现实数据生成方面也具备独特文化优势。随着“人工智能+”政策对实体经济的赋能，硬件与AI的深度结合将成为中国在AI下半场竞赛中的核心竞争力。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文红果短剧回应“取消 AI 仿真人剧本保底”传闻：说法不实，部分剧本仍保留保底发布于AI新闻资讯发布时间 :2026年5月18号 15:03阅读 :1分钟针对市场传闻“红果短剧取消 AI 仿真人剧本保底、全面转向纯分成”的消息，红果短剧相关负责人近日公开回应称该说法不实，强调部分剧本依然保留保底政策。此前，有市场消息透露红果短剧官方人士在编剧群发布通知，称按照4月27日的最新政策，新选中的剧本等级将仅提供纯分成合作，分成比例为20%，但若满足后验热度激励指标，仍可享有单项激励活动，且剧本等级随质量提升可持续升级。该消息在编剧圈引发震动，部分编剧甚至将其戏称为“零元购”，并传出仅有极少数头部编剧的保底收入未受影响。此次“保底风波”的背后，折射出网文与短剧赛道在引入生成式 AI 和 AI 仿真人技术后，平台、编剧与内容生产生态之间正在经历剧烈的利益博弈与机制磨合。作为当前内容出海和本土消费的新风口，AI 短剧通过技术手段大幅降低了制作成本与周期，但这也对传统剧本创作的定价体系和内容品控带来了全新挑战。红果短剧对传闻的澄清，在一定程度上稳定了创作者群体的预期，但平台引入“后验热度激励”与“动态等级调整”的机制，也暗示了行业正在向更加注重数据反馈、精细化运营的绩效驱动型方向演进。如何在效率、技术与创作者核心利益之间寻找平衡，将是未来 AI 内容平台构建可持续生态的关键。

AI资讯AI新闻资讯正文5 天击穿苹果 M5 最强内存防线！AI 辅助完成纯数据提权，Mac 安全格局巨变发布于AI新闻资讯发布时间 :2026年5月18号 15:11阅读 :2分钟2026 年 5 月 14 日，帕洛阿尔托安全研究公司 Calif 在官方博客发布 55 页完整技术报告，宣布一项重大突破：研究团队仅用 5 天时间，就在搭载 M5 芯片、运行最新macOS 的设备上实现了从普通本地用户权限到 root shell 的完整提权，利用过程无需任何代码注入，属于纯“data-only”攻击。该报告迅速引发安全界高度关注。5 天 vs 5 年：极致紧凑的攻击链开发时间线Calif 团队的时间线令人惊叹：4 月 25 日：研究员 Bruce Dang 发现两个已知类别的内核 bug；4 月 27 日：Dion Blazakis 加入，Claude Mythos Preview 开始提供辅助分析；5 月 1 日：Josh Maine 完成工具链，完整 exploit 在实体 M5 芯片设备（macOS 26.4.1，MIE 全开）上成功运行。他们将两个零日级内核 bug、侧信道标签碰撞技巧与数据 -only 提权技术组合，构建出一条高效本地提权链。MIE 防护被突破：苹果硬件内存安全天花板首现缺口MIE（Memory Integrity Extension）是苹果专为 M5/A19 系列芯片打造的硬件级内存安全机制，在 ARM MTE 基础上进一步强化，每 16 字节内存分配 4-bit 标签，由硬件强制校验，性能开销仅约 3%。此前被业界视为“下一代内存安全天花板”，曾让多个知名 exploit kit 失效。Calif 团队此次证明：即使是 MIE，在正确漏洞组合与侧信道技术辅助下仍可被绕过。他们成功将标签碰撞概率从随机 6.25% 压至确定性实现，完成纯数据操作的内核提权。这是全球首次公开披露针对 M5 芯片 MIE 的 macOS 内核内存破坏利用。研究员在报告中表示：“MIE 从来不是为了防住所有黑客而设计的，有正确漏洞就能绕过。”Claude Mythos Preview：AI 成为安全攻防超级加速器Anthropic 推出的 Claude Mythos Preview（安全专版）在此次研究中发挥关键作用。该模型通过 Project Glasswing 计划有限开放给顶级安全团队，能在短时间内发现大量跨系统、浏览器及关键基础设施的零日漏洞，包括多年未被修复的“古董级”bug。Calif 团队强调：“这不是 AI 独立完成，而是顶级人类研究员与前沿 AI 的强强联合。小团队几天就能完成过去几个月甚至几年的工作。”研究完成后，团队已亲自前往 Apple Park 汇报，苹果已在 macOS 26.5 安全更新中修复相关漏洞，并公开致谢 Calif 与 Anthropic Research 的合作。深层启示：AI 驱动“Bugmageddon”时代来临Calif 在报告结尾警示：“这是即将到来的冰山一角。”防御方用 AI 寻找和修复漏洞，攻击方同样用 AI 加速发现与利用，导致漏洞挖掘速度呈指数级提升，传统人工审计模式彻底面临挑战。本次研究被纳入 Calif“Month of AI-Discovered Bugs”系列，标志着 AI 正式成为安全攻防平衡中的核心变量，而非单纯辅助工具。OpenAI 与 Anthropic 在不同技术路线上的探索，也正加速这一趋势。对 Mac 用户的实际影响与建议本次漏洞为本地提权攻击，需要攻击者已获得本地代码执行权限，远程利用风险较低。但苹果已快速推送 macOS 26.5 修复更新，AIbase 强烈建议所有 Mac 用户立即更新系统。Calif 团队提醒：“这只是冰山一角。”未来，更多花费数年打造的顶级防护，可能在 AI 辅助下被快速击穿。AIbase 观点：AI 正在全面重塑网络安全攻防格局。无论是防御还是研究，主动拥抱 AI 工具已成为必然选择。科技巨头与安全社区的协作模式，或将成为未来行业新常态。

AI资讯AI新闻资讯正文AI垃圾报告挤爆安全通道，Linux创始人怒喷技术滥用发布于AI新闻资讯发布时间 :2026年5月18号 15:51阅读 :1分钟Linux创始人林纳斯·托瓦兹近日在发布开源内核新版本时，公开对人工智能工具的滥用发出严厉警告。他指出，大量由AI辅助生成的重复错误报告正严重堵塞内核的安全呈报通道，给核心维护者带来了巨大的无效工作负担。托瓦兹并非完全禁止在开发中使用AI，其核心痛点在于低门槛的提交方式。当开发者单纯依赖AI工具在系统中扫描出一个Bug时，其他同行往往也在用相同的工具刷出同一问题，这导致安全列表近期持续涌现海量重复报告。维护人员陷入无意义内耗，日常管理几近瘫痪面对这些潮水般涌入的AI报告，内核维护者不得不把大量时间耗费在重复转发问题上。他们甚至需要一遍遍地回复“该问题在数周前已修复”并附上公开链接，导致日常维护工作变得几乎无法管理。本周进行的内核常规更新进展基本正常，驱动程序更新大约占了总数的一半，其中图形处理器的更新最为显著。然而，AI报告带来的额外干扰，确实在很大程度上破坏了原本高效的代码审核节奏。拒绝华而不实的口头贡献，提倡带补丁提交针对这一乱象，托瓦兹对开源社区的开发者给出了明确的建设性提议，强调内核开发不需要那些对底层代码缺乏真正理解的路人贡献者。他认为，利用工具随手抛出问题对社区并无实质价值。他呼吁，开发者如果真心想为开源社区贡献价值，首先应当认真阅读相关的开发文档。更重要的是，在合理利用AI工具的同时，提交报告时必须附带具体的修复解决方案和代码补丁。

AI资讯AI新闻资讯正文索尼澄清Xperia1XIII“AI相机助手”功能:仅提供拍摄建议而非直接编辑发布于AI新闻资讯发布时间 :2026年5月18号 16:30阅读 :1分钟索尼（Sony）近日针对其在Xperia1XIII上展示的“AI相机助手(AI Camera Assistant)”功能引发的舆论关注进行了官方澄清，详细阐述了该AI驱动功能的底层工作逻辑。索尼官方表示，该AI相机助手本质上并非图像编辑或生成工具，而是基于端侧视觉感知能力，根据实时光线、景深及拍摄对象特征为用户提供拍摄优化建议。在实际交互中，当用户将镜头对准目标物体时，系统会智能输出涵盖曝光、色彩及背景模糊调整的四个优化选项。然而，该功能在宣发与实际体验的对齐上引发了行业争议。索尼在产品视频中宣称该助手可推荐“最佳拍照角度”，但演示片段仅展示了变焦放大建议，这与空间维度的角度推荐存在概念差异。此外，索尼于5月14日及后续在社交平台X上展示的优化示例效果亦不及预期，尽管避免了早前版本中严重的色彩失真与过度曝光缺陷，但网格示例中的建议仍被指存在色彩过度饱和、画面平淡且处理痕迹明显等问题。当前，智能手机行业正加速将生成式AI与传统计算摄影深度融合，试图通过多模态理解提升大众的拍摄体验。索尼此次的技术尝试反映出传统影像巨头在AI转型期，如何平衡“算法干预”与“真实画质”的审慎考量。如何在提供智能化辅助的同时，确保算法建议的审美专业性与图像自然度，正成为各厂商AI影像落地演进中的核心技术挑战。

AI资讯AI新闻资讯正文必胜客强推AI配送导致业绩暴跌，遭加盟商索赔超 1 亿美元发布于AI新闻资讯发布时间 :2026年5月18号 17:06阅读 :1分钟知名快餐品牌必胜客近日深陷一起严重的法律纠纷。其美国头部加盟商“查克东北披萨公司”正式向法院提起诉讼，控诉品牌方强行要求门店上线的AI配送管理平台“龙尾”（Dragontail）存在严重的运营冲突。该加盟商在多个核心地区运营着约 111 家门店，原本经营业绩远超同行。然而在强制引入该AI配送系统后，门店的生意却一落千丈，由此造成的业务亏损与资产缩水损失已超过 1 亿美元。系统漏洞引发骑手囤单，配送效率不升反降在龙尾系统上线前，该加盟商旗下超九成的外卖订单都能在 30 分钟内送达，顾客满意度极高。但由于该系统将后厨出餐流程与订单时长实时同步给外卖骑手，反而引发了意想不到的连锁故障。外卖骑手为了自身利益，开始利用该系统刻意囤积、合并多笔订单，甚至最长等候 15 分钟凑单才出发。这直接导致原本高效的配送大面积延迟，消费者拿到的披萨频频变凉，不满情绪激增。骑手看小费选择性拒单，品牌深陷多重经营困境诉状中还揭露了另一个致命缺陷，骑手能够通过该AI系统直接查看订单的小费金额与支付方式，进而出现选择性拒接部分外卖订单的情况。这一现象彻底拉低了消费者的满意度，导致纽约等核心市场的销售额同比增速直接由正转负。目前必胜客发言人表示将通过正规法律途径予以回应。事实上，必胜客在美国市场正面临整体低迷的局面，其母公司百胜餐饮集团此前已计划在今年上半年关闭美国境内的 250 家必胜客门店。

AI资讯AI新闻资讯正文会“伪装”的AI：最新研究表明人们普遍认定人工智能比人类更自信发布于AI新闻资讯发布时间 :2026年5月18号 17:35阅读 :1分钟随着人工智能系统的日益普及，越来越多的人开始习惯向AI寻求消费、阅读等各类生活建议。然而，来自滑铁卢大学与伦敦大学学院的科研团队近期在《通讯·心理学》期刊上发表了一项有趣的研究成果。结果显示，即便AI与人类给出的答案完全一致，人们也普遍认为AI作答时的自信心更强。这种现象被称为AI的“自信错觉”。研究表明，人们在无法直接获知对方笃定程度时，往往会下意识地根据作答速度快慢、决策轻松程度等外在线索来主观判定对方的底气。固有偏见误导信任，缺乏情感信号引发隐藏隐患由于大众普遍默认AI在诸多领域的专业能力优于人类，这种固有偏见极易让人产生误判。实验发现，一旦人们认定某款AI能力出众，就会盲目觉得它在任何场景下的作答都十分笃定，而事实上面对特定问题时系统本身可能也未必靠谱。在日常人际交往中，语气、神情和肢体姿态等情绪自信是重要的社交信号，能帮我们决定何时采信他人。但由于绝大多数大语言模型缺失这类表达途径，导致用户只能盲目揣测，即使AI本身心存疑虑且答案极易出错，用户仍可能过度轻信。科学家探寻直观传递方式，未来大模型有望新增功能针对这一潜在隐患，科隆巴托教授及其团队强调，未来在研发人工智能产品时，亟需通过更直白、多元的外在形式向用户清晰传递系统对答案的真实把握程度。这不仅能为现有的生成式AI提供优化思路，也能有效防止用户因盲信而误入歧途。目前，研究团队正筹备开展新一轮的针对性研究，旨在探索出高效、直观且值得信赖的人机交互传递方式。未来的大语言模型或许会新增自信度明示功能，帮助人类更精准、理性地判断何时该采纳AI给出的建议。

AI资讯AI新闻资讯正文张朝阳谈搜狐AI战略：不做"大模型军备竞赛"，聚焦应用落地与内容中立发布于AI新闻资讯发布时间 :2026年5月18号 17:43阅读 :1分钟在生成式AI重塑科技产业格局的2026年，当多数互联网巨头争相投入千亿参数模型研发时，搜狐选择了一条更为务实的路径。在近日举办的搜狐科技年度论坛上，CEO张朝阳明确表示:搜狐并未入局大模型"第一阵营"的研发，而是将资源聚焦于"基于自身业务基础的理性应用"——这一表态，既是对行业过热情绪的冷静回应，也折射出中型科技公司在AI时代的生存智慧。战略取舍:为何"不追第一阵营"?张朝阳的判断基于三重现实考量:考量维度具体逻辑行业对照资金门槛千亿级参数模型训练需数亿美元算力投入，且迭代周期短、沉没成本高百度、阿里、腾讯年AI投入均超百亿，搜狐难以同等量级跟进技术壁垒大模型研发需顶尖算法团队、海量高质量数据与工程化能力，非短期可补足头部厂商已构建"数据-算力-人才"闭环，后来者追赶窗口收窄商业回报通用大模型变现路径尚不清晰，而垂直场景应用已能产生即时价值搜狐选择"用别人的模型，做自己的业务"，降低试错成本这一策略并非"躺平"，而是精准卡位:在算力与参数竞赛之外，寻找"模型能力×业务场景"的最优解。落地路径:效率优先，内容为王搜狐的AI应用聚焦两大方向:提效降本:编程辅助:引入代码生成工具提升研发效率，缩短产品迭代周期;运营自动化:用AI处理内容审核、用户反馈、数据分析等重复性工作，释放人力聚焦创意与策略;成本优化:通过智能调度降低服务器资源消耗，提升单位算力产出。内容克制:保持中立性:对AI生成内容明确标注，避免"机器冒充人工"误导用户;规避乱象风险:不盲目追求"自动写稿"“批量生产”，优先保障信息真实性与价值观导向;人机协同:记者、编辑仍为内容决策主体，AI仅作为素材整理、事实核查的辅助工具。张朝阳特别强调:“内容平台的核心价值是信任，若为短期流量牺牲中立性，将损害长期品牌资产。“这一立场，在"标题党"“洗稿"“虚假信息"频发的内容生态中，显得尤为珍贵。行业启示:中小厂商的"理性生存法则"搜狐的选择，为未跻身"大模型第一梯队"的企业提供了可参考的范式:不拼参数，拼场景:通用能力交给头部厂商，自身聚焦垂直场景的深度优化;不造轮子，用轮子:通过API调用、模型微调等方式复用成熟能力，降低研发风险;不追热点，守底线:在内容、金融、医疗等高敏感领域，将"合规"与"可信"置于"创新"之前。这种"务实主义"策略，或许正是中型科技公司在资源约束下的最优解:与其在红海赛道陪跑，不如在细分领域深耕。挑战与展望当然，这一路径也面临考验:技术依赖风险:若底层模型提供商调整策略或涨价，搜狐的应用层能力可能受制于人;体验差异化难题:当多家企业调用同一模型，如何构建独特用户价值?内容边界把握:在"克制"与"创新"间找到平衡点，需要持续的机制设计与价值观校准。但不可否认，在AI浪潮席卷一切的时代，知道"不做什么”，有时比"做什么"更需要战略定力。张朝阳的发言，既是对搜狐路径的阐释，也是对行业的提醒:技术终需服务于人，而非让人服务于技术。当大模型竞赛进入深水区，或许最终胜出的，不一定是参数最大的那个，而是最懂"何时该进、何时该守"的那个。搜狐的选择，未必是规模最大的，但可能是最清醒的之一。

AI资讯AI新闻资讯正文百度一季度AI云收入达88亿大涨79%，昆仑芯P800交付万卡集群发布于AI新闻资讯发布时间 :2026年5月18号 17:49阅读 :1分钟5月18日下午，百度正式发布2026年第一季度财报，总营收达321亿元，其中一般性业务收入260亿元，同比增长2%，超出市场预期。期内，AI业务展现出强劲的增长动能，百度AI云收入达88亿元，同比高速增长79%，GPU云收入更是实现184%的爆发式增长。财报期间，百度智能云全面升级为面向大规模智能体应用的新全栈AI云，全方位进化Agent Infra与AI Infra能力，强化了从底层算力到智能体应用的全链路协同。市场层面，百度智能云一季度凭借25个中标项目、12.48亿元的中标金额再度蝉联行业双料冠军，中标金额达第二名的近5倍。在算力硬件领域，昆仑芯P800已完成规模化验证并交付多个万卡集群，顺利完成文心大模型5.1重要版本的训练。同时，基于昆仑芯打造的天池256卡超节点已成功点亮并定于6月上市，其吞吐性能提升25%，并完成对文心、DeepSeek、GLM、MiniMax等主流大模型的适配，推理效率显著提升50%。

Claude Code是什么Claude Code 是 Anthropic 公司推出的基于命令行的AI 编程工具。Claude Code集成先进的 Claude Opus 4 语言模型，能理解自然语言指令，自动生成、修改和调试代码，并与 VS Code、JetBrains 等主流 IDE 深度集成。Claude Code支持 200K 超长上下文，能快速索引整个代码库并理解复杂依赖关系。Claude Code 具备文件操作、代码搜索、网页浏览和 Git 工作流管理等功能，显著提升开发效率。Claude Code帮助开发者快速生成高质量代码，能直接在开发环境中执行命令，避免频繁切换界面，极大地优化开发流程。最新版推出自动安全审查功能，在漏洞发布之前捕获漏洞。Claude Code 新功能 Opus 规划模式支持高效且智能的开发规划，用户能在复杂项目中借助 Opus 4.1 模型的强大分析能力制定高质量的开发计划，在执行阶段切换到 Sonnet 模型，实现高效、低成本的代码生成。Claude Code的主要功能代码库深度理解：基于智能搜索技术，快速理解整个代码库的结构和依赖关系，无需手动选择上下文文件。多文件协同修改：能够跨多个文件进行协调一致的代码修改，确保代码逻辑的完整性和一致性。代码生成与优化：专门针对代码理解和生成进行了优化，能够快速生成高质量的代码片段、测试用例和修复代码中的问题。终端集成：直接在开发者的终端中运行，无需切换工具，减少上下文切换的麻烦。与主流 IDE 集成：支持 VS Code 和 JetBrains 等主流开发工具，直接在 IDE 中提供代码建议和修改，无缝融入现有开发流程。工作流整合：与 GitHub、GitLab 等工具集成，支持从阅读问题、编写代码、运行测试到提交 PR 的全流程操作。智能编辑：基于对代码库和依赖关系的理解，进行强大的多文件编辑，确保代码修改的有效性和可靠性。用户控制与定制：不会在未经明确批准的情况下修改文件，适应开发者的编码标准和模式，并通过 SDK 或 GitHub Actions 进行定制化配置。终端运行的/security-review命令：支持用户在终端输入此命令，Claude Code将自动检查代码中的SQL注入、XSS等安全隐患，能直接修复发现的漏洞。GitHub Actions集成：在GitHub仓库配置相关文件后，每次提交新的PR时，自动触发安全审查，检测出漏洞后在代码行上添加内联注释，解释问题并给出修复建议。Opus 规划模式：在规划阶段调用强大的 Claude 4.1 Opus 模型制定高质量开发计划，在执行任务时使用高效且低成本的 Claude 4 Sonnet 模型，实现智能与效率的结合。Agent View：让开发者在一个终端界面内同时调度、监控和管理多个并行 AI 会话，无需多开窗口可实时查看各任务状态、快速切换介入或解除阻塞。如何使用Claude Code安装 Node.js：确保系统安装 Node.js 18 或更高版本。安装 CLI：在终端运行以下命令安装 Claude Code 的命令行工具：npminstall-g@anthropic/claude-cli登录和初始化：登录：运行claude login并按照提示输入 API 密钥（从 Anthropic 官网获取）。初始化：运行claude init，连接到代码库。开始使用：生成代码：用英语描述需求，例如：claude-p"Create a function to calculate the sum of two numbers in Python"修复错误：描述错误或粘贴错误消息，例如：claude-p"Fix this error: TypeError: Cannot read properties of null (reading ’length’)“自动化任务：运行命令解决繁琐任务，例如：claude-p"Resolve merge conflicts in the current branch"集成到开发环境：终端：直接在终端中使用。IDE：安装 VS Code 或 JetBrains 插件，实现在 IDE 中使用。Claude Code 的产品定价Pro 计划：年付 $17/月，月付 $20/月。Max 5x 计划：$100/人/月。Max 20x 计划：$200/人/月。Claude Code的应用场景代码理解：快速解释复杂代码的功能和逻辑，帮助开发者迅速理解项目结构。代码修改：自动修复代码中的错误或漏洞，快速定位问题并提供解决方案。代码测试：为代码自动生成单元测试或集成测试，确保代码的可靠性和稳定性。版本控制：支持自然语言完成 Git 提交、分支管理等操作，简化版本控制流程，让 Git 操作更加便捷高效。功能实现与扩展：快速实现新功能或扩展现有功能，加速开发进度，帮助开发者高效完成任务。©️版权声明：若无特殊声明，本站所有文章版权均归AI工具集原创和所有，未经许可，任何个人、媒体、网站、团体不得转载、抄袭或以其他方式复制发表本站内容，或在非我站所属的服务器上建立镜像。否则，我站将依法保留追究相关法律责任的权利。类似于Claude Code的工具CodeFuse蚂蚁集团推出的AI代码编程助手iFlow CLI心流AI推出的免费终端 AI 智能体CodeBuddy IDE腾讯推出的全栈开发AI IDEDevin首个全自主的AI软件工程师智能体InfCode词元无限推出的企业级AI编程工具CodexOpenAI 推出的 AI 编程智能体6 条评论再想想发表评论wi游客此ide已放弃 ，国内无法使用，哎9个月前回复AI小集需要很强的魔法，还经常封号9个月前回复wi游客对。弄得很累，不如用亚马逊的ai ide，不用翻墙，也不封号9个月前回复3游客Claude 太强大了，可惜用不来2个月前回复Ai游客需要手机验证，不支持国内无862个月前回复AI小集是的，最好搞一个美国手机号2个月前回复

Kimi Code – Kimi推出的AI编程工具AI工具4个月前发布AI小集05Kimi Code是什么Kimi Code 是 Kimi推出的AI编程工具，可在终端（CLI）、VS Code 、Cursor、JetBrains 和 Zed 等主流编辑器中直接运行，基于Kimi K2.5多模态大模型支持图片和视频输入辅助编程。工具提供端到端代码能力，涵盖构建、调试、重构、测试及脚本编写。Kimi Code能完美兼容Claude Code、Roo Code等第三方 Agent，让 AI 编程能力无缝嵌入用户的每一次 commit。Kimi Code的主要功能全环境覆盖：提供Kimi CLI、VS Code 插件及第三方编辑器支持，可在终端和主流 IDE 中无缝运行编程任务。多模态交互：基于Kimi K2.5模型支持直接输入图片和视频进行编程辅助，实现视觉友好的代码理解与生成功能。端到端编程：完整覆盖代码构建、调试、重构、测试及脚本编写等开发全流程，自动迁移现有技能到新工作流。高频并发：每 5 小时支持 100-500 次请求额度，确保复杂项目开发过程中 AI 辅助的不间断可用性。第三方兼容：完美适配 Claude Code、Roo Code 等主流 Coding Agent，一键登录可完成设备授权与账号绑定。如何使用Kimi Code订阅会员：访问 Kimi Code 官网 https://www.kimi.com/code。安装和登录：CLI 用户：执行curl -L code.kimi.com/install.sh | bash安装，运行/login命令一键授权。VS Code 用户：安装 Kimi Code 插件，在终端直接运行登录命令完成设备绑定。配置使用官方 Agent：登录成功后即可直接使用，系统自动完成账户绑定。第三方工具（Claude Code / Roo Code）：在控制台创建 API Key，复制后配置到对应工具的环境变量或设置中即可开始编程。Kimi Code的产品定价Andante 计划：定价 ¥4.99（7 天试用），试用期结束后以 ¥49/月自动续费，提供专属 Kimi Code 使用额度、旗舰模型抢先体验以及支持多个编程会话的基础功能。Moderato 计划：定价 ¥99/月，提供每周更新的使用额度，多种设备登录并分享套餐额度，支持用户在多个项目中高效协作。Allegretto 计划：定价 ¥199/月，提供充足的每周额度和更高的并发上限，为高级用户提供更强大的使用体验和超值选择。Kimi Code的应用场景全栈开发任务：覆盖代码构建、调试、重构、测试及脚本编写等端到端编程工作流，满足从原型开发到生产部署的完整需求。多模态辅助编程：支持直接输入 UI 设计图、视频演示等多媒体内容，基于视觉信息生成对应代码实现或进行问题诊断。跨平台开发环境：可在终端 CLI、VS Code、Cursor、JetBrains、Zed 等主流编辑器中无缝切换，适配不同开发者的工具链偏好。高频敏捷开发：借助 Turbo 级加速（100 Tokens/s）和每 5 小时数百次的并发额度，支撑大型项目中的高频代码生成与迭代需求。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇inSai Hilight - AI营销视频Agent平台，多Agent深度协同下一篇Lucy 2 - Decart AI推出的实时视频生成模型相关文章Zonos – ZyphraAI 开源的多语言 TTS 模型AI小集3DeepSWE – Together.ai联合Agentica开源的AI Agent框架AI小集3Granola – AI会议助手，实时转录会议内容、添加笔记、标记重点AI小集3nowadays – AI活动规划平台，自动完成场地搜索、报价收集、合同签署全流程服务AI小集2面试狗 – AI笔试面试助手，智能识别笔试题目自动解答AI小集3NEO – 商汤联合南洋理工开源的全新多模态模型架构AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付7小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型7小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手7小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台11小时前PPT Master – 开源 AI PPT 生成标准化工作流13小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架13小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具1天前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前

Kimi WebBridge – 月之暗面推出的浏览器扩展插件AI工具3天前发布AI小集03Kimi WebBridge是什么Kimi WebBridge 是月之暗面推出的浏览器扩展插件，面向Kimi Code、Claude Code、Cursor、Codex等本地 AI Agent。Kimi WebBridge 通过本地桥接服务与 Chrome DevTools Protocol，让 AI 携带用户真实登录态和 Cookie，像真人一样在浏览器中执行点击、滑动、输入、填表、信息提取及跨站点内容整合等操作。所有任务在本地完成，网页数据不会离开设备，保障隐私安全，同时实现高效、自动化的网页操作。Kimi WebBridge的主要功能浏览器自动化操作：支持 AI Agent 执行网页导航、点击按钮、填写表单、滑动页面等操作。身份状态继承：AI 携带用户真实登录态、Cookie 和账号权限操作网站，无需额外配置。信息提取与整合：跨站点抓取内容并整合，支持将数据写入在线文档或本地目录。可视化任务追踪：运行时标记正在操作的浏览器标签页，不占用鼠标键盘，用户可并行使用电脑。进阶 CLI 工具：支持将固定流程打包为不消耗大模型 Token 的专用 CLI 工具，用于重复任务。Kimi WebBridge的技术原理双组件协同架构：系统由本地桥接服务与浏览器扩展两部分组成。AI Agent 将自然语言指令发送至本地桥接服务，后者通过标准接口与 Chrome 或 Edge 浏览器内的扩展通信，扩展再基于底层协议驱动浏览器完成具体操作，并将执行结果回传给 Agent，形成完整的指令闭环。基于 Chrome DevTools Protocol：浏览器扩展底层采用 Chrome DevTools Protocol（CDP），是 Chrome 官方提供的调试协议。通过协议，扩展能在用户当前使用的真实浏览器实例中执行导航、元素点击、表单输入、页面截图、DOM 读取与修改等操作，无需启动独立或虚拟浏览器环境。本地身份继承机制：Kimi WebBridge 直接复用用户已打开的浏览器会话，继承真实的登录态、Cookie 和本地存储数据。AI 在操作目标网站时，等同于用户本人在操作，因此无需额外配置账号、处理验证码或突破登录壁垒，天然支持各类需要身份认证的后台系统。端侧隐私保障：所有指令解析、页面交互与数据提取均在用户本地设备完成。桥接服务只负责本地进程间通信，网页内容、Cookie 及操作结果不会上传至云端服务器，确保敏感业务数据不出本地设备。如何使用Kimi WebBridge安装浏览器插件：在 Chrome 或 Edge 浏览器应用商店搜索并安装Kimi WebBridge扩展插件。安装本地守护程序：向本地 AI Agent 发送指定 curl 命令，安装 Kimi WebBridge 的本地守护程序与 Skill。重启 Agent 加载技能：重启 Kimi Code 等本地 AI Agent，使其识别并加载新安装的 WebBridge 技能。验证安装结果：输入测试指令「使用 kimi-webbridge 帮我打开 kimi.com」，若成功打开标签页并出现 agent:kimi 提示即代表安装成功。Kimi Claw Desktop 用户：若通过 Kimi 桌面版在本地部署了 Kimi Claw Desktop，无需安装守护程序，直接发送指令可开始使用。Kimi WebBridge的核心优势原生身份继承：直接复用用户现有登录态，无需模拟账号或处理验证码，降低复杂网站接入门槛。本地隐私安全：全部操作在本地浏览器执行，敏感数据与网页内容不上传云端。多 Agent 兼容：统一支持 Kimi Code、Claude Code、Cursor、Codex、Hermes、OpenClaw 等主流本地 Agent。非侵入式运行：后台标记操作标签页，不抢占鼠标键盘，人机可协同工作。低 Token 消耗：固定流程可编译为专用 CLI 工具，重复执行时无需反复调用大模型。Kimi WebBridge的项目地址项目官网：https://www.kimi.com/zh-cn/features/webbridgeKimi WebBridge的同类竞品对比维度Kimi WebBridgeBrowserbaseAgent-Browser（Vercel）运行方式本地浏览器扩展 + 桥接服务云端托管浏览器云端/本地浏览器自动化身份继承直接复用用户 Cookie 和登录态需单独配置环境或账号需单独配置环境或账号数据隐私完全本地执行，数据不出设备数据经过云端服务器取决于部署方式支持 AgentKimi Code、Claude Code、Cursor 等主要面向开发者 API主要面向开发者 API使用门槛安装插件 + 一行命令即可需 API 集成开发需一定技术配置人机协同后台运行，不占用键鼠云端运行，无本地交互云端运行，无本地交互Kimi WebBridge的应用场景自动化信息整理：自动采集多个笔记类 APP 的应用商店宣传素材，让 AI 分析设计风格差异并写入在线文档。复杂操作自动化：一键接管量化研究模拟平台，自动抓取金融数据、执行多策略迭代回测并撰写量化报告。重复性网页任务：自动化完成基于网页的重复性繁琐操作或定时任务，如批量填表、点击、数据录入与状态更新。跨站点内容整合：在已登录状态下跨多个网站提取信息并整合，实现内容迁移、数据汇总与多源信息同步。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Mavis - MiniMax Agent 推出的多 Agent 协作模式下一篇TencentDB Agent Memory - 腾讯开源的智能体记忆管理工具相关文章Vid2World – 清华联合重庆大学推出视频模型转为世界模型的框架AI小集2Talkie – MiniMax推出的AI虚拟伴侣应用AI小集3Perplexity Computer – Perplexity推出的通用数字员工AI小集3Fathom – AI会议记录工具，实时生成精准转录和智能总结AI小集2Memory Layers – Meta 开源不增加算力时增加大模型的参数方法AI小集2PoseTalk – 文本和音频驱动的生成会说话的头部动画开源项目AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付6小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型7小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手7小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台10小时前PPT Master – 开源 AI PPT 生成标准化工作流12小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架12小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具1天前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前

了解更多 kimi.com/zh-cn/features/webbridge

Kimi WebBridgeÔºökimi.com/zh-cn/features/webbridge

Kimi K2.6 发布并开源，全面精进代码和 Agent 集群能力

Kimi K2.6「Agent 集群」现已支持 300 个成员并行协作

和 Kimi 一起投身 AGI，穿越成长周期

服务全球1/5网站的Cloudflare选择Kimi K2.5，降低77%成本

Kimi 杨植麟「2026 中关村论坛」演讲全文（附视频）

Kimi API：用90%缓存命中率，把价格打到25%

你可以来 Kimi 使用 OpenClaw 了

技术报告：Kimi K2.5 如何实现文本和视觉能力互相增强？

不只PPT，Kimi K2.5 Agent可以帮你做Excel、Word和PDF了

Kimi 发布并开源 K2.5 模型，带来全新视觉理解、代码和 Agent 集群能力

微信扫一扫关注该公众号

微信扫一扫可打开此内容，使用完整服务

吐司APP的简介吐司是腾讯推出的AI原生应用创作平台，用户只需输入想法，AI自动生成完整的安卓原生应用。平台支持智能功能梳理、精细化编辑打磨，能一键导出APK进行本地安装与分享。内置灵感广场汇聚海量优质模板，涵盖社交助手、照片精选、学习工具等多元场景，支持用户共创。无需编程基础，零基础用户也能快速将奇思妙想转化为可实际使用的专属APP，轻松实现个性化应用开发。吐司APP的主要功能AI自动生成应用：输入自然语言想法，AI自动梳理功能逻辑并生成完整的安卓原生应用，无需编写代码。精细化编辑打磨：支持对生成的应用进行细节调整与优化，满足个性化需求。一键导出APK：生成完成后可直接导出APK安装包，支持本地安装、测试与社交分享。灵感广场：内置海量精选应用模板与创意案例，如社交助手、照片精选、学习工具等，帮助用户快速获取创作灵感。应用共创：支持用户基于现有模板进行二次创作与共创，降低应用开发门槛。多场景覆盖：支持生成考研题库、运势工具、普通话测试、宠物健康日志、梦境记录、食材新鲜度检测、万年历等多样化实用应用。类似于吐司APP的应用TRAE SOLO APP字节跳动推出移动端的 AI 智能工作助手WindsurfCodeium公司推出的AI编程工具Kiro亚马逊公司推出的 AI IDEReplit APPAI开发工具应用CursorAnysphere公司推出的AI代码编辑器Trae字节跳动推出的免费 AI IDE暂无评论再想想发表评论暂无评论…

LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台AI工具11小时前发布AI小集03LibTV团队版是什么LibTV 团队版是LibTV推出的专业级 AI 视频协作工作台，专为 AI 短剧、电影、TVC 制作团队设计。平台提供团队空间、实时画布协作、资产库沉淀、权限管理、成本管控等功能，解决 AI 视频创作中资产散落、协作低效、人员流动导致资产流失等痛点，支持规模化、工业化的 AI 内容生产。LibTV团队版的主要功能专属团队空间：统一沉淀分镜、脚本、镜头资产与生成记录，历史项目可迁移，避免核心资产散落在个人账号与本地文件中。画布协作：导演、美术、剪辑等角色可在同一块画布实时协同，同步节点状态与镜头修改，减少版本错乱。团队资产库：角色主体、风格模板、运镜方案、提示词工作流与参考素材支持统一沉淀，成熟视觉体系可持续复用。权限管理：支持项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作场景。成本管理：共享积分池、成员额度分配与消耗查看，团队管理者可统一管理生成资源、席位与项目成本。资产交接：成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免创作资产随人员流动中断。如何使用LibTV团队版创建团队空间：访问LibTV官网升级至团队版，建立专属团队项目空间，设置团队名称与基础信息。配置成员与权限：邀请导演、编剧、美术、剪辑等成员加入，按岗位分配访问、编辑、复制或发布权限。沉淀与迁移资产：将历史个人项目迁移至团队空间，或新建项目并导入参考素材、角色设定与风格模板。画布协作创作：成员在统一画布内协同推进分镜设计、镜头生成与修改，实时同步进度与反馈。复用团队资产库：调用已沉淀的角色主体、风格模板与提示词工作流，保持系列化内容风格一致。管控成本与交接：通过共享积分池分配生成额度，监控消耗；成员变动时资产自动保留，确保项目连续性。LibTV团队版的核心优势工业化协作：从个人创作升级为团队制片管理，支持长期项目、多岗位协同与规模化内容生产。资产持续复用：成熟的角色设定、风格与提示词工作流可沉淀为团队资产，降低新人上手与重复调试成本。实时协同提效：多人画布协作减少版本错乱与重复生成，围绕同一套镜头语言与素材体系推进项目。成本与风险可控：共享积分池与席位管理适配多项目并行需求，权限分级与资产交接机制保障核心创作安全。LibTV团队版的同类竞品对比对比维度LibTV 团队版快手可灵（企业版/团队功能）协作模式多人实时画布协作，导演/美术/剪辑等角色可在同一块画布内同步节点状态、生成进度与镜头修改团队空间内创作接力，支持桌面端（Win/Mac）同步，减少文件下载与搬运损耗资产沉淀角色主体、风格模板、运镜方案、提示词工作流、参考素材与工具箱统一沉淀，支持成熟视觉体系持续复用设立「共享资产库」，支持个人空间向团队空间一键转移灵感值、创作资产与主体资产，实现”一人贡献，全员复用”权限管理项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作所有者、管理员、协作者三重角色，灵活配置查看/编辑/移除权限，保障核心资产安全成本管理共享积分池、成员额度分配与积分消耗查看，支持多项目并行下的生成资源统筹支持灵感值跨空间转移，团队共享生成资源，但精细化成本管控功能相对较弱资产交接成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免资产随人员流动中断个人资产可一键导入团队空间，但成员退出后的自动资产交接机制未明确强调核心场景聚焦 AI 短剧、AI 电影、AI TVC 的工业化制片流程，强调长期项目管理与跨岗位协同覆盖 AI 视频生成、智能分镜、广告 Demo、影视特效等，强调快速创作与素材复用产品定位AI 视频专业级协作工作台，面向制片管理流程AI 视频生成平台的团队功能扩展，面向创作者联盟与工作室LibTV团队版的应用场景AI 视频工作室：适合长期运营 AI 短剧、动画、MV 的创作团队，通过资产库统一沉淀角色与风格，实现系列化生产。短剧承制方与 MCN：适配高频、多项目并行的工业化内容生产，导演与美术可同时在线推进分镜与镜头生成。品牌市场部与代理公司：支持品牌方、代理公司与外部创意团队跨组织协作，共同管理 AI TVC 与 Campaign 项目。AI 电影与大型制作团队：支持长篇叙事、系列化 IP 与复杂制作流程，导演组、制片组、美术组与后期团队跨岗位协同。企业内容与创意部门：适配培训、营销、产品与全球化传播等标准化内容生产，便于企业进行长期生成资源规划。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇PPT Master - 开源 AI PPT 生成标准化工作流下一篇腾讯Marvis - 腾讯应用宝推出的操作系统层个人 AI 助手相关文章雪鸮AI – AI绘画辅助工具，能精准控制图像生成AI小集14MimicPhoto – AI照片编辑工具，能快速调整笑容、眼神等面部表情细节AI小集2DreamPolish – 智谱AI、清华、北大联合推出的文本到3D生成模型AI小集2HeyBeauty – AI虚拟试衣平台，提供风格推荐和尺寸预测功能AI小集3Markdown-to-Image – 开源的在线 Markdown 转海报编辑器AI小集2SurveyX – 人民大学联合悉尼大学等推出自动化生成学术综述的系统AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付6小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型7小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手7小时前PPT Master – 开源 AI PPT 生成标准化工作流13小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架13小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具1天前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前

Grok Build – xAI 推出的终端原生 AI 编程智能体AI工具1天前发布AI小集03Grok Build是什么Grok Build 是 xAI 推出的终端原生 AI 编程智能体（Agentic CLI），面向专业软件工程与复杂编码任务。基于 Grok 4.3 beta 的 200 万 token 超长上下文，支持 Plan Mode 计划审批、最多 8 个并行子智能体协作及 Arena Mode 自动方案排序，可在本地直接读写文件、执行 Shell 命令、操作 Git 并构建完整应用。采用本地优先架构，源代码与敏感数据不上传云端，保障企业级安全。Grok Build的主要功能Plan Mode（计划模式）：复杂任务执行前自动生成结构化步骤计划，用户可审批、评论或修改，所有代码变更以 diff 形式透明展示，避免”黑箱”操作。并行子智能体：最多支持8 个并发 AI 智能体同时工作，可分别处理代码分析、CI/CD 检查、性能优化、文档检索等不同子任务。Arena Mode（竞技场模式）：多个智能体针对同一任务生成不同代码方案，系统自动评估并排序，开发者直接选择最优解，无需手动对比。超长上下文窗口：基于 Grok 4.3 beta，支持200 万 token上下文，可一次性加载整个大型代码库进行跨文件复杂任务。本地优先架构：源代码、凭证和敏感数据不上传云端，完全在本地执行，支持离线环境和气隙网络，保障企业级代码安全。终端原生 CLI + Web 双入口：支持本地命令行交互，也可通过浏览器界面远程操作，满足不同工作流习惯。VS Code 扩展集成：提供官方 VS Code 插件，在 IDE 内直接调用 Grok Build 能力，实现 GUI adjacent 工作流。Git 与开发服务器集成：可自动管理 Git 仓库、启动本地开发服务器、内置浏览器浏览网页、执行 Shell 命令。插件与开放生态：支持 MCP 服务器、AGENTS.md 团队规范文件、Hooks 自定义动作、ACP（Agent Client Protocol）协议，便于集成现有 DevOps 流水线。Headless 脚本模式：通过-p标志支持非交互式执行，可直接嵌入 CI/CD 流水线或自动化脚本。BYOK（自带 API Key）：支持使用自有 xAI API Key 进行本地计费，灵活控制成本。如何使用Grok Build一键安装 CLI：在终端执行curl -fsSL https://x.ai/cli/install.sh | bash完成本地安装，支持 macOS、Linux 和 Windows 三大平台。启动交互式终端：安装完成后输入grok>进入交互式命令行界面，可直接用自然语言描述编程需求。登录账号：在终端内执行grok> /login登录 xAI 账号，解锁高级功能与订阅权益。配置 API Key（可选）：如需使用自有额度，执行grok> /api-key add xai添加个人 xAI API Key（BYOK 模式）。使用 Plan Mode 处理复杂任务：输入需求后，Grok Build 会自动生成结构化执行计划，展示步骤与预期变更 diff，用户可审批、评论或修改后再执行。利用并行子智能体：在复杂项目中，系统会自动或手动触发最多 8 个并发子智能体，分别处理代码分析、测试生成、文档检索等子任务。通过 Arena Mode 选优：对同一任务让多个智能体生成不同方案，系统自动评估排序，你只需选择排名最高的方案应用。Web 端远程操作：访问build.grok.com或x.ai/cli，在浏览器中打开 Grok Build 界面，无需本地安装即可使用。VS Code 集成工作流：安装 Grok Build 的 VS Code 扩展，在编辑器内直接调用智能体能力，边写代码边让 AI 协助修改。Headless 自动化执行：在脚本或 CI/CD 流水线中使用-p标志运行非交互式命令，实现自动化构建与部署。查看帮助与命令：在终端输入grok> /help查看所有内置命令与快捷操作指南。Grok Build的官网地址项目官网：https://x.ai/news/grok-build-cliGrok Build的核心优势本地优先，代码绝不上云：源代码、凭证及敏感数据完全在本地执行，无需上传至 xAI 服务器，适合处理专有代码库、金融及政务等受监管行业项目。200 万 token 超长上下文：基于 Grok 4.3 beta，可一次性将整个大型代码库载入内存，跨数十个文件进行复杂重构与多文件联动修改，大幅减少上下文丢失。8 路并行子智能体：最多同时调度 8 个并发 AI 智能体分工协作，分别处理代码分析、测试生成、CI/CD 检查等子任务，显著缩短复杂项目耗时。Arena Mode 自动选优：同一任务由多个智能体生成不同方案，系统自动评估排序并给出排名，开发者直接选择最优解，告别手动对比试错。Plan Mode 透明可控：执行前先输出结构化步骤计划与 diff 预览，用户可审批、评论或修改后再执行，每一步变更清晰可见，避免”黑箱”风险。极致 API 价格竞争力：输入 token 仅$0.20 / 百万 tokens，相较 Claude Code 和 Codex CLI 大幅降低高频使用成本。终端原生 + 多入口覆盖：支持 CLI 命令行、Web 浏览器、VS Code 扩展三种工作入口，既满足极客式终端操作，也兼容 IDE 集成习惯。Grok Build的同类竞品对比对比维度Grok Build (xAI)Claude Code(Anthropic)Codex CLI(OpenAI)产品定位终端原生 AI 编程智能体（Agentic CLI）终端 AI 编程助手云端 AI 编程智能体发布状态2026.5 早期 Beta已正式发布已正式发布底层模型Grok 4.3 beta / grok-code-fast-1Claude 4 (Sonnet/Opus)Codex (基于 GPT-4o/o3)上下文长度200 万 token20 万 token (Claude 4)约 20 万 token并行智能体最多 8 个并发子智能体单智能体为主单智能体为主方案评估Arena Mode 自动排序选优手动对比或迭代手动对比或迭代执行计划Plan Mode（先计划后审批执行）直接执行，可审查 diff直接执行，可审查 diff数据安全本地优先，代码不上云云端执行，代码上传 Anthropic云端执行，代码上传 OpenAI离线支持支持完全离线/气隙网络需联网需联网SWE-Bench70.8%约 72–75%约 70%API 定价$0.20 / 百万 tokensPro 订阅 + 按量计费Credits 按量计费订阅门槛SuperGrok Heavy ($300/月)Claude Pro / Team 订阅ChatGPT Pro / 开发者额度入口形态CLI + Web + VS Code 扩展CLI + IDE 集成CLI + ChatGPT 移动端自动化集成Headless 模式 (-p 标志) + CI/CD有限脚本支持有限脚本支持核心差异多智能体并行 + 本地优先 + 超大上下文模型推理能力强，生态成熟与 ChatGPT 生态深度整合Grok Build的应用场景大型代码库重构与现代化：基于 200 万 token 超长上下文一次性加载整个项目，跨数十个文件进行安全重构、依赖升级或架构迁移，避免上下文截断导致的遗漏。复杂全栈应用从零构建：通过 Plan Mode 自动生成项目结构、技术选型、依赖安装和文件生成计划，审批后直接执行，快速搭建完整应用原型。敏感行业专有代码开发：本地优先架构确保源代码、凭证和敏感数据不上传云端，适用于金融、政务、医疗、军工等对数据主权有严格要求的场景。离线/气隙网络环境编程：无需持续联网即可在本地执行全部任务，适配高安全隔离网络、内网开发机或无外网访问权限的工作环境。多维度自动化代码审查：启动多个并行子智能体分别检查代码安全性、性能瓶颈、CI/CD 合规性和文档完整性，一次性输出综合审查报告。技术方案快速对比与选型：使用 Arena Mode 针对同一需求生成多种实现方案，系统自动评估排序，帮助团队在短时间内确定最优技术路径。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇OpenHuman - 开源桌面端 AI 助手，能主动感知工作上下文下一篇Violin - 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具相关文章BlinkShot – 开源的实时AI图像生成器，几毫秒内生成高质量图像AI小集3Vidu主体参照功能 – Vidu推出的AI视频生成新功能，一键同步角色特征AI小集4GPT-SoVITS – 开源的声音克隆项目，只需少量数据即可合成声音AI小集5Motia – AI Agent 开发框架，支持多种编程语言、一键部署智能体AI小集3Univer – 开源 AI 办公工具，支持Word、Excel等文档处理全栈解决方案AI小集3袋鼠参谋 – 美团推出的商家AI智能决策应用AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付6小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型7小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手7小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台10小时前PPT Master – 开源 AI PPT 生成标准化工作流12小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架12小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具1天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前

OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文AI工具2天前发布AI小集05OpenHuman是什么OpenHuman 是 tinyhumansai 团队推出的开源个人 AI 超级智能助手，定位为”Your Personal AI super intelligence”，主打私有、简单且极其强大。非传统聊天机器人，是能主动感知用户工作上下文的桌面级 AI Agent，通过每 20 分钟自动同步 118+ 第三方应用数据，构建本地持久记忆，让 AI 真正”懂你”。项目采用 Rust + TypeScript + Tauri 构建，支持 macOS、Linux 和 Windows 平台。OpenHuman的主要功能桌面吉祥物交互：具备视觉形象的桌面 AI 助手，支持语音输入（STT）与 ElevenLabs 语音输出（TTS），吉祥物可实时唇形同步，能作为参会者加入 Google Meet 并实时记录会议。118+ 第三方应用一键集成：通过 OAuth 快速接入 Gmail、Notion、GitHub、Slack、Linear、Jira、Stripe、Calendar、Drive 等主流工具，无需编写任何连接代码。每 20 分钟自动同步（Auto-fetch）：核心引擎每 20 分钟遍历所有活跃连接，自动拉取最新数据进入本地记忆系统，让 Agent 在次日早上已掌握前日全部上下文。Memory Tree 层次化持久记忆：将同步数据规范化为 ≤3k-token 的 Markdown chunks，评分后折叠存入层级摘要树，存储于本地 SQLite，实现跨会话的长期记忆。Obsidian Vault 原生兼容：同步数据同时输出为 Obsidian 可读的.md文件，用户可在 Obsidian Graph View 中浏览和编辑 AI 维护的知识网络。TokenJuice 智能 Token 压缩：在数据进入 LLM 前自动压缩（HTML 转 Markdown、缩短长 URL、移除非 ASCII 字符等），最高可降低 80% 的 API 调用成本与延迟。内置模型路由（Model Routing）：根据任务类型自动分配轻量模型、强推理模型或视觉模型，一个订阅覆盖多种模型能力，无需管理多个 API Key。完整工具链开箱即用：内置 Web 搜索、网页抓取（Web Scraper）、代码工具集（文件系统、Git、Lint、Test、Grep），无需额外安装插件即可使用。本地优先与隐私保护：所有工作流数据本地加密存储，支持通过 Ollama 接入本地模型实现完全离线运行，数据不上传云端。可选 agentmemory 共享后端：可配置与 Claude Code、Cursor、Codex 等工具共享同一个持久记忆存储，实现跨工具的知识互通。消息通道集成：支持 inbound/outbound 消息收发，打通用户已有的沟通渠道，工作流数据始终保存在本地设备。如何使用OpenHuman官网下载安装：访问OpenHuman的官网 https://tinyhumans.ai/openhuman 下载对应平台的 DMG（macOS）或 EXE（Windows）安装包，双击安装后即可启动。命令行快速安装（macOS / Linux x64）：在终端执行curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash自动完成下载与安装。命令行快速安装（Windows）：在 PowerShell 执行irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex完成一键安装。一键连接工作应用：打开应用后，通过 OAuth 一键接入 Gmail、Notion、GitHub、Slack、Calendar、Drive 等 118+ 第三方工具，授权后每 20 分钟自动同步数据。与桌面吉祥物交互：点击桌面上的吉祥物形象，通过语音或文字与其对话，可勾选 “Speak replies” 让 AI 语音播报回复。查看自动同步的记忆数据：在应用内查看 Agent 已同步的邮件、任务、代码仓库等上下文，AI 会自动压缩整理为 Memory Tree 供后续对话调用。Obsidian 知识库浏览：同步数据自动写入本地 Obsidian-compatible Vault，打开 Obsidian 即可在 Graph View 中可视化浏览 AI 维护的知识网络，也可手动编辑.md文件。使用内置工具集：直接调用开箱即用的 Web 搜索、网页抓取、代码工具（文件系统、Git、Lint、Test、Grep）等能力，无需额外安装插件。接入本地模型（可选）：如需完全离线运行，通过 Ollama 在本地部署模型，并在设置中切换至本地模型后端，实现零云端依赖。OpenHuman的项目地址官网地址：https://tinyhumans.ai/openhumanGitHub 仓库：https://github.com/tinyhumansai/openhumanOpenHuman的核心优势UI 优先分钟级上手：干净的桌面体验，无需终端配置，从安装到可用 Agent 仅需几分钟，区别于终端优先的竞品。上下文在分钟级而非周级：传统 Agent 需要数周学习用户习惯，OpenHuman 通过一次同步即可掌握邮箱、日历、仓库、文档、消息的全量压缩上下文。一站式订阅，零 API 蔓延：一个订阅覆盖所有功能，无需额外购买多个模型 API Key。Obsidian 知识库原生兼容：同步数据自动写入 Obsidian-compatible Vault，用户可在熟悉笔记界面中浏览和编辑 AI 维护的知识图谱。Rust + Tauri 高性能架构：内存占用远低于 Electron 方案，69.7% Rust + 26.1% TypeScript 构建。开源可审计：GPL-3.0 协议，代码完全透明，隐私敏感用户可自行审计数据流向。OpenHuman的同类竞品对比对比维度OpenHumanOpenClawHermes Agent开源协议✅ GPL-3.0✅ MIT✅ MIT上手难度✅ UI 优先，分钟级上手⚠️ 终端优先配置⚠️ 终端优先配置使用成本✅ 单一订阅 + TokenJuice 压缩⚠️ 自备模型 API⚠️ 自备模型 API记忆能力🚀 Memory Tree + Obsidian Vault + 可选 agentmemory 后端⚠️ 依赖插件扩展✅ 自学习观察第三方集成🚀 118+ 应用 OAuth 一键接入⚠️ 需自行配置连接⚠️ 需自行配置连接自动数据同步✅ 每 20 分钟自动 fetch 至本地记忆🚫 无自动同步🚫 无自动同步API 管理✅ 一个账户覆盖所有模型🚫 自备 Key🚫 多供应商管理模型路由✅ 内置自动分配（轻量/推理/视觉）⚠️ 手动切换⚠️ 手动切换原生工具集✅ 代码 + 搜索 + 抓取 + 语音✅ 代码工具✅ 代码工具数据隐私🚀 本地优先，数据不上云依赖自托管配置依赖自托管配置OpenHuman的应用场景跨应用项目上下文整合：同时连接 GitHub、Linear、Jira、Slack 后，让 AI 自动对比代码 PR、任务看板和团队消息，一键汇总项目最新进展，无需手动切换多个工具。智能邮件筛选与摘要：授权 Gmail 后，AI 每 20 分钟自动扫描收件箱，按项目或优先级分类摘要，快速定位需要回复的重要邮件，过滤掉低价值信息。AI 会议助手：让桌面吉祥物以真实参会者身份加入 Google Meet，实时记录会议内容、提取 action items，并基于记忆库提供参会者背景信息辅助沟通。代码仓库智能问答：同步 GitHub 仓库数据后，直接向 AI 询问代码逻辑、历史变更原因、PR 评论争议点，Agent 基于完整代码上下文给出精准回答。Obsidian 个人知识库构建：将邮件、文档、聊天记录自动压缩归档为 Obsidian-compatible 的.md文件，在 Graph View 中可视化呈现知识关联网络，打造 Karpathy 风格的 LLM 知识库。多工具工作流自动化：连接 Notion、Calendar、Drive、Stripe 等工具后，AI 自动关联”日历事件→相关文档→对应邮件→财务记录”，形成完整的项目上下文链。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Intern-S2-Preview - 上海 AI Lab 开源的科学多模态大模型下一篇Grok Build - xAI 推出的终端原生 AI 编程智能体相关文章异世界回响 – Soul App推出的AI社交平台AI小集3pplx-embed – Perplexity推出的系列文本嵌入模型AI小集2Click2Mask – AI图像编辑技术，通过简单点击和内容描述实现智能编辑AI小集2Amurex – 开源AI会议助手，提供实时建议、智能摘要、快速回顾关键信息AI小集2爱图表重磅升级 – AI秒生图表+自动写报告，新功能来了AI小集3Hyring – AI招聘平台，AI面试官自动进行视频面试、监考、评估AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付6小时前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型7小时前腾讯Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手7小时前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台10小时前PPT Master – 开源 AI PPT 生成标准化工作流12小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架12小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具1天前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前

Marta MussoScienceMay 17, 2026 5:30 AMThe First Atomic Bomb Test in 1945 Created an Entirely New MaterialThe discovery from the Trinity nuclear test site shows how extreme conditions can result in materials never before seen in nature or in the lab.The new material is a clathrate made of calcium, copper, and silicon .Luca Bindi/Università di FirenzeCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyDuring the Trinitynuclear teston July 16, 1945, in the New Mexico desert—the world’s very first test of anatomic bomb—a new material spontaneously formed. It wasdiscoveredonly recently, by an international research team coordinated by geologist Luca Bindi at the University of Florence, which identified the novel clathrate based on calcium, copper, and silicon. It’s a material never before observed either in nature or as an artificial compound created in the laboratory.What Are Clathrates?The term “clathrates” denotes materials characterized by a “cage-like” structure that traps other atoms and molecules inside, giving them unique properties. Of great technological interest, these materials are being studied for various applications ranging from energy conversion (as thermoelectric materials capable of transforming heat into electricity) to the development of new semiconductors, to gas storage and hydrogen for future energy technologies.The New MaterialTo discover the new material, researchers focused on trinitite, a silicate glass containing rare metallic phases. Using some techniques like x-ray diffraction, the team was able to identify a type I clathrate based on calcium, copper, and silicon within a tiny copper-rich metal droplet embedded in a sample of red trinitite.The new material, the researchers say, formed spontaneously during a nuclear explosion. This indicates that the extreme conditions, such as extremely high temperatures and pressures, can generate new materials that are impossible to obtain by traditional methods.Natural LaboratoriesThe discovery is even more interesting because in the same detonation event another very rare material was formed: a silicon-rich quasicrystal, already documented by the team of experts led by Bindi a few years ago.A quasicrystal, as BinditoldWIRED at the time, is something that is not a crystal, but looks a lot like one. “Their peculiarity,” he said, “is that the atomic arrangement that is not periodic, but nearly so, creates incredible symmetries from which derive amazing physical properties, among other things, very difficult to predict.”Establishing the link between these structures therefore helps scientists better understand how atoms organize under extreme conditions and expand the possibilities for designing new materials. “Events such as nuclear explosions, lightning strikes, or meteoritic impacts function as true natural laboratories,” the researchers explain. “They allow us to observe forms of matter that we cannot easily reproduce in the laboratory.”In essence, this research opens new vistas for the development of innovative technologies, demonstrating that even destructive events can bequeath discoveries useful for the future.This story originally appeared inWIRED Italiaand has been translated from Italian.

Ben DowsettGearMay 17, 2026 6:00 AMIf You’re a Serious Bowler, You Need to Know About Bowling Lane OilBowling centers apply oil to their lanes using a machine that works like a giant inkjet printer. The pattern in which it’s applied can change everything about how the ball travels.Photograph: Charly Triballeau/AFP/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyEach time EJTackett steps to his lane to throw a shot in the upcomingProfessional Bowlers Association World Championships, he’ll be thinking about more than just ball spin or angle.Tackett has won the last three world titles, and there’s a convincing case that he’s currently thebest bowler on Earth.While his ability tothrow a consistent ballis certainly key to his success, there’s another factor he and his fellow PBA pros are paying closer attention to during matches: lane oil patterns.Most of us amateurs have never thought about the thin layer of oil that coats the surface of the bowling lane—unless someone at our local bowling center slips and falls on it hilariously. But at the highest levels of the sport, lane oil is a foundational element of tactics and game-planning. The oil is applied in a distinct pattern that reduces the friction of specific parts of the lane, allowing a well thrown ball to slide across certain areas, then grab ahold of the surface and curl when it’s supposed to.But lane oil patterns aren’t always identical. Oil distribution can change over the course of a game, and in fact the starting pattern is intentionally redrawn for various PBA events, forcing pros to adjust on the fly.“All of us as professionals are playing a guessing game by watching our bowling ball go down the lane,” Tackett tells me. “And trying to figure out where the oil has moved from and has moved to.”Tackett says understanding and reacting to lane oil patterns has become one of the single most important skills for a pro bowler in recent years.Technology plays a growing role in the way lane oil is understood and applied. And as the tech has continued to improve, the sport’s powers that be are increasingly relying on those advances to keep the top pros guessing.Back when bowlinglanes were made from actual wood like maple or pine in the 1970s and earlier, lane oil’s primary purpose was protective. It provided a lubricating barrier between ball and wood, helping reduce the wear and tear on both.During that era—and even into the early 1980s when synthetic lane materials began replacing wood—lane oil drew the ire of serious bowlers due to the inconsistency of how it was applied. There was no method for ensuring standardized oil applications; sections of lanes would often have too much or too little oil. Bowlers would essentially be blind to the lane’s conditions until partway through a game, after enough balls had been thrown to get a gauge of which areas were light or heavy on oil. Maybe certain bowlers enjoyed the randomness, but it was untenable for professionals.John Davis, at the time a mechanic who oiled lanes for his local bowling center in Arizona, was the first to address this issue. He invented a lane cleaning tool in the early ’80s to address lanes that were often cleaned inconsistently, allowing oil to pool or run as a result. By the early ’90s, he had started a company named Kegel that had perfected what’s known as Sanction technology, which isstill used today. The bowling equipment giant Brunswick, the other major player in the space, uses similar technology in itsMax lane machines.“It looks almost like an inkjet printer,” Chris Chartrand, Kegel’s CEO, says about the Sanction tech. “You have this head that travels back and forth, and it applies that stream on whatever board we want to apply it to.”The technology has kept improving. Kegel used to produce two separate machines, one for lane cleaning and one for oil; now both steps are handled by a single machine. More recently, the machines have become battery-powered, removing issues of power cords stretching across lanes.Lately, Kegel has been steadily improving its automation, to the point where today’s machines do the entire job without any human intervention.The lanes youand I bowl on as amateurs are oiled very differently from the ones pros use.At your local bowling center, public lanes are oiled in what’s referred to as a “high” ratio: The level of oil present in the middle of a lane is eight to 10 times higher than what’s on the outside. At the far left and right of the lane, many public bowling alleys have no oil at all.“On a normal pattern at your normal bowling center, there is some autocorrect,” Tackett says. Because the edges of the lane have very little oil, shots that drift to either side will slow down; if the ball has been thrown with the proper spin to guide it back toward the middle of the lane, it will curl more effectively on the drier surface. “It makes it easier to hit the pocket.”(By “the pocket,” Tacket means that sweet spot at the front

Jill DuffyGearMay 17, 2026 6:30 AMTake Control of Your Debt With These Free Online ToolsThese free debt calculators help you set up payment plans to get back in the black.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyApps for budgetingandpersonal financedo a good job of tracking yourmoneyas you earn and spend it. Some also have excellent debt calculators that help you figure out how to pay off your debts.Each debt calculator is a little different. Some suggest a specific method for paying down debt, while others are simulators that let you see how your total amount paid will decrease if you increase your monthly payment.Here are a few useful calculators and some guidance about what makes them different.A Straightforward Plan: BankrateBankrate’s freedebt payoff calculatorgives you a timetable for paying off each of your debts. You enter as many debts as you want to include, their interest rates, total loan amounts, and other details. You also enter any new income you expect to receive, such as an annual salary increase or windfall, and the amount that you can put toward your debts. The calculator then generates one payment table for each debt showing how much to pay each month until the debt is cleared.Bankrate prioritizes paying off the debt with the highest interest rate first. Once your first debt is paid off, the money you would have put toward it is diverted to your other monthly payments. In other words, as you eliminate debts, the monthly payments on your other debts increase until they, too, are paid off.Who should use it?Bankrate’s calculator works for people who have multiple debts, and the total monthly minimum payments are within their financial reach. If that’s you, then you’ll get a crystal clear plan—with a timeline—for getting rid of all your debts.Where it comes up short.This calculator assumes that paying off your debts by clearing the one with the highest interest rate first is in your best interest. That’s not true for everyone. You might have other options, such as consolidating credit card debt to a new card with a 0 percent introductory rate or filing for bankruptcy. Bankrate also doesn’t take into account other personal finance concerns, like other uses of monthly funds that free up once you pay off your first debt—Bankrate tells you to put that money toward your next-highest-interest debt. You might be better off putting it toward retirement savings or an emergency fund.Big-Picture Guidance: NerdWalletNerdWallet’s freedebt load calculatordetermines your debt load as a percentage of your income. The resulting debt load is classified as smaller (less than 36 percent), larger (37–42), or overwhelming (43 percent or more). Based on the outcome, NerdWallet suggests a method for eliminating your debt, which you read about in an educational article below the results.Who should use it?This calculator helps you get a big-picture sense of your debt. If you have a lot of debt, it’s useful for ruling out (or ruling in) the option of declaring bankruptcy.Where it comes up short.It’s not great at analyzing the finer details of your debt. For example, in the setup, there’s no line item for student loans or a mortgage, much less the exact interest rate you pay on loans. The results are a rough guide rather than a personalized strategy.Automated Inputs: WalletHubWhen you sign up for WalletHub (free) and connect your financial accounts, the app pulls real information about how much money you owe and your payment history. Itsdebt payoff planis a calculator that lets you play with the numbers to see what would happen if you increased your monthly payment. How much faster can you clear the debt? How much will you save in interest? You can quickly see the difference between increasing your monthly payment by, say, $50 versus $150.Who should use it?This calculator is for WalletHub users who have connected their financial accounts. It’s most useful for people who can afford to pay more than the monthly minimum on their debts.Where it comes up short.Although WalletHub pulls real information about your actual loans, it can miss some detail. I use it to track my mortgage, for example, and it doesn’t know that the total bill I pay each month includes homeowner’s insurance. The calculations regarding the principal and interest are still correct, but the amount that comes out of my pocket is higher than what WalletHub assumes. Other considerations are missing, too, like what to do when the rate of inflation is higher than the interest rate on a long-term loan. Overall, this debt calculator is very good, and these shortcomings are nitpicks, not deal-breakers.Credit Card Transfers: WalletHub and NerdWalletIf you’re carrying credit card debt,WalletHubandNerdWalletboth have a calculator that shows how much money you could save by transferring the balance to a new card with an introductory rate. (CreditKarma has a balance transfer calculator, t

Matthew KorfhageGearMay 17, 2026 6:33 AMThe Best Grills of 2026: Cook Smarter, Not HarderI grilled, smoked, seared, cleaned, and synced digital temp controllers to find the best grill option for every cookout and tailgate.Courtesy of AmazonCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe Grill haschanged. The smoker has changed, too. Once governed mostly by feel and experience, the best grills and the best smokers now offer a fair amount of tech to measure and control temperature (think: temperature probes and fans to modulate airflow). After all, mankind’s formative invention wasn’t the “discovery” of fire; it was the power to control it.The WIRED Reviews team has been testing grillsfor more than a decade—searing, smoking, grilling, and even baking on them in all kinds of weather—to find the best choice for everyone. These days, I’ve been most excited about the next-gen smart technology that allows for tight monitoring and temperature control. For instance, I tricked out my top-pick for smart upgrades, an old-fashioned Weber kettle, with the excellentSpider Venom ($280)digital temperature controller that allows me to turn my charcoal grill temps up and down like an oven. Summer is good. Also,charcoal makes all meat taste more like bacon.Mmmmm, bacon.After you’re set with the best grills and smokers, check out other backyard buying guides for all your outdoor needs, like theBest Flat Top Grills and Griddles,Best Wireless Meat Thermometers, andBest Smokeless Firepits.Best Grills of 2026Best Charcoal Grill for Smart UpgradesWeber Original Kettle CharcoalJump to review$139$129(7% off)Amazon (18-inch)Best Smart Pellet SmokerRecteq Flagship 1600 Pellet Smoker and GrillJump to review$1,600BBQGuysBest Smart Charcoal GrillMasterbuilt Gravity Series 1150 Digital Charcoal GrillJump to review$1,100Home DepotBest Gas GrillWeber Spirit E-210 Gas GrillJump to review$449The Home DepotJump to SectionAccordionItemContainerButtonLargeChevronOther Grills I RecommendA Grill to AvoidWhat to Look for in a Good GrillHow to Upgrade Dumb Grills to Smart GrillsUseful Grill AccessoriesBest Charcoal Grill for Smart UpgradesPhotograph: Matthew KorfhagePhotograph: Matthew KorfhageVideo: Matthew KorfhageChevronChevronSave to wishlistSave to wishlistWeberOriginal Kettle Charcoal$139$129(7% off)Amazon (18-inch)$149Amazon (22-inch)$99Weber (All, Starting Price)WIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDSimple, durable, reliable, and affordableNear-infinite smart-tech upgrades availableLightweight and portableTIREDIt could do with better wheel nutsUpgrades often cost more than the grillI’d argue that the best way to take advantage of top smart-grill technology is to add the newest and best tech to a sturdy, old-fashioned, classic grill that’ll outlast the tech. A grill like the classic 22-inch Weber Kettle.The Weber Kettle is the infinitely hackable Honda Civic of backyard grills. Its combination of simplicity, reliability, durability, affordability, and outright ubiquity has spawned an innovative secondary market devoted to tricking out the humble cooker with smart temperature controls and more features than a multiplex.The best Weber add-on I’ve tried so far is a fan-modulated temperature controller from Spider Grills called theVenom ($280)that tracks the temperature on each cook, allowing you to heat it up or turn it down, like a stove. According to my testing, it holds your charcoal grill to temperature within about 30 degrees Fahrenheit of the target, which is great for both low-and-slow cooking and for creating repeatable sear temps for a steak. This is especially true when using the attached meat probes. The only downside is that the Venom needs to be plugged in. Might I suggest a nicepower bank?WIRED product reviewer Scott Gilbertson has favored the excellent WeberConnect Smart Hub ($70)for years, which features probes to monitor meat temps and flipping times and can alert your phone (or just your ears) when your grill hits the ideal temperature zone. However, he admits that the Hub’s USB-A charger is starting to feel a bit old-fashioned.This year, Weber is also coming out with aPerformer Smart Ring ($280)that seems to operate a lot like the Venom. If it works anywhere near as well as the Venom, I’m eager to try it out.But not every kettle upgrade worth trying is “smart.” I am in the process of testing a number of add-ons to the 22-inch Weber Kettle, including aWebcraft cooking ring ($249)from Spider that serves as a base for side shelves, arotisserie, and apizza oven conversion kit. AWeber rotisserie ($180)is also available from the Illinois mothership.Note also that there are two versions of the Weber Kettle: Original andPremium. The Premium is around $80 more and features a built-in thermometer in the lid, a hinged cooking grate, and a fancier ash-removal system. They’re the same size and feature the same cooking space. Unless you really like the easier ash-removal system, save that cash

Christopher NullGearMay 17, 2026 7:02 AMReview: Oto Smart SprinklerThis smart sprinkler doesn’t need a power cable, thanks to its integrated solar panel.Courtesy of AmazonTriangleUpBuy NowMultiple Buying Options Available$589$499at Amazon$589$499at OtoCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyRating:9/10Open rating explainerInformationWIREDSolar panel is a game changer. App is easy to navigate during both setup and daily use. Watering is accurate and thorough. Built-in weather intelligence helps to avoid unnecessary runs.TIREDDistant watering can land with heavy pressure, potentially damaging plants.One of themajor headaches of stand-alone smart sprinklers that attach to a garden hose is that they require power to operate. Since few folks have electrical outlets in the middle of their lawns, configuring a smart sprinkler means connecting not just a hose but also a long power cable to the device, creating double the effort and double the chance for something to get chewed up by a lawn mower. The Oto Smart Sprinkler solves for this problem with solar power, which makes for not only a far more flexible installation than its smart-sprinkler competitors but also ongoing easy use.With a strong focus on quick setup and simplicity of function, Oto provides a refreshing alternative in the smart sprinkler market. The only question: Does its useful innovation justify its blisteringly high price tag?Set and ForgetWhen I unboxed the Oto Smart Sprinkler for the first time, I was convinced something must be missing—perhaps even a second box with more parts? But as it turns out, the Oto really is as streamlined as it claims to be. The sprinkler measures just 16 x 12 x 4 inches and weighs about 6 pounds. Next to the similarly featured but monstrous-by-comparisonAiper IrriSense 2, the Oto is practically miniature.Being pint-sized makes lawn irrigation even easier than expected. You can place the unit on the ground, mount it to a hardscape, or even attach it to a fence (all with the included metal bracket). If you want to play things fast and loose, you can situate it to stand freely without the bracket, though the bracket helps you ensure it’s replaced in the exact same location if it’s ever removed (such as during the winter months) for even irrigation.The only real placement caveat is that in order to have it operate cordlessly, the Oto requires at least three hours a day of direct sunlight to its top-mounted 2.2-watt solar panel to keep its internal 5,400-mAh battery topped up. Keep this in mind as you decide where to install your brackets—you definitely don’t want to block too much of the sun’s rays.Video: Christopher NullBefore taking it outside for its first use, Oto instructs you to charge the device overnight via wall power (though you could certainly just leave it in the sun for a few days instead). You’ll do this using the included power adapter with 50 feet of cable. You can also lean on the plugged-in charging option if you can’t squeeze out enough sunshine on a regular basis (albeit in that case, this irrigation solution is probably not for you). You’ll find the power cable stored inside the Oto hardware, in a large hollow space beneath the solar panel. This hollow space is intentional and isn’t used for anything more than supporting the panel above, but it’s probably best not to leave the power adapter inside permanently, as it this chamber can get rather wet.The Oto can throw water up to a distance of 40 feet in a 360-degree radius of the device, which equates to a maximum of 5,000 square feet of coverage, enough for a sizable lawn. (That’s about on par with Aiper.) The hose connector on the side of the device is easy to work with and secure; I experienced no leaks at all when connecting it to a standard garden hose. A 50-foot maximum length for your hose is specified to ensure water pressure remains sufficient.Digital DropletsJust about everything on the Oto happens via its app (though there’s a manual switch on top of the device you can use to speed up a run that’s pending in the app). That app sets up via a 2.4-GHz Wi-Fi connection to your home network, so you’ll need to make sure your yard has effective wireless coverage.Once configured, setup proceeds much like the Aiper and pricierIrrigreenapps: You create a zone, then use the app to define its boundaries. Similar to the aforementioned systems, Oto’s sprinkler is designed for precision watering, firing water in a beam in a single direction instead of a wide spray. That said, Oto’s spray is comparably narrow, only hitting a single, designated patch instead of producing a two-dimensional curtain of water like Irrigreen’s “water printing” system. You get a nice preview of this as you set the boundaries of your yard.Like its competitors, Oto lets you set each zone as a spot (for watering a single tree, perhaps), a line (for a flowerbed), or a 2-D area (for a yard). I tested all of these modes but spent most of my time workin

Arthur Allen, KFF Health NewsScienceMay 18, 2026 5:00 AMA Danish Couple’s Maverick African Research Finds Its Moment in RFK Jr.’s Vaccine PolicyThe work of Peter Aaby and Christine Stabell Benn has long been controversial. Until Robert F. Kennedy Jr. became US health policy chief, most vaccine scientists tended to ignore it. Now they can’t.Photograph: Tom Williams/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyIn 1996, Guinea-Bissauseemed like an ideal research post for budding pediatrician Lone Graff Stensballe. Her supervisor, a fellow Dane named Peter Aaby, had spentnearly two decades collecting dataon 100,000 people living in the mud brick homes of the West African country’s capital.Aaby and his partner, Christine Stabell Benn, believed that the years of research in the impoverished country had yielded a major discovery about vaccines—and what they described as “non-specific effects”: The measles and tuberculosis vaccines, which were derived from live, weakened viruses and bacteria, they said, boosted child survival beyond protecting against those particular pathogens.But, the scientists said, shots made from deactivated whole germs, or pieces of them, such as the diphtheria-tetanus-pertussis (DTP) shot, caused more deaths—especially in little girls—than getting no vaccine at all.The World Health Organization repeatedly and inconclusively examined these astonishing findings. They tended to elicit shrugs from other global health researchers, who found Aaby’s research techniques unusual and his results generally impossible to replicate.Then came Donald Trump, Covid, and the administrative reign of anti-vaccine advocate Robert F. Kennedy Jr.Suddenly, Aaby and Benn weren’t just sending up distant smoke signals from a far corner of the planet. They wereconfidently voicing their viewsand policy prescriptions online and in medical journals. The “framework” for “testing, approving, and regulating vaccines needs to be updated to accommodate non-specific effects,” their team wrote ina 2023 review.And the Trump administration has taken notice.“They became more strident in saying that their findings were real and that the world needed to do something about it,” said Kathryn Edwards, a Vanderbilt University vaccinologist who has been aware of Aaby’s work since the 1990s. “And they became more aligned with RFK.”Kennedy, as secretary of the Department of Health and Human Services,cited one of Aaby’s papersto justify slashing $2.6 billion in US support for Gavi, a global alliance of vaccination initiatives. The cut could result in 1.2 million preventable deaths over five years in the world’s poorest countries, the nonprofit agency has estimated. Kennedy has frozen$600 millionin current Gavi funding over largely debunked vaccine safety claims.Kennedy describedthe 2017 paperas a “landmark study” by “five highly regarded mainstream vaccine experts” that found that girls who received a diphtheria-tetanus-pertussis, or DTP, shot were 10 times more likely to die from all causes than unvaccinated children.In fact, the study was far too small to confidently make such assertions, as Benn acknowledged. In a study of historical data that included 535 girls, four of those vaccinated against DTP in a three-month period of infancy died of unrelated causes, while one unvaccinated girl died during that period. Afollow-up published by the same groupin 2022 found that the DTP shot by itself had no effect on mortality. Critics say the 2017 study, rather than being a landmark, exemplified the troubling shortfalls they perceive in the Danish team’s research.As Aaby and Benn’s US profile has risen, scientists in Denmark have set upon the work of their compatriots. In news and journalarticlespublished over the past 18 months, Danish statisticians and infectious disease experts have said the duo’s methods wereunorthodox, evenshoddy, and were structured to supportpreconceived views. A national scientific board is investigating their work.Stensballe, who worked with Aaby and Benn for 20 years, has been among those voicing doubts.“It took years to see what I see clearly today, that there is a strange concerning pattern in their work,” Stensballe said in a phone interview from Copenhagen, where she treats children at Rigshospitalet, the city’s largest teaching hospital. She said their work is full of confirmation bias—favoring interpretations that fit their hypotheses.Those hypotheses overlap, in important areas, with the notions of Kennedy and other vaccine-skeptical officials at HHS.In December, HHS announced the agency would award the Bandim group $1.6 million to study whether the birth dose of hepatitis B vaccine weakens babies’ immune systems or causes neurological issues.The researchers plan to withhold the vaccine from half of the 14,000 newborns in the study, although the long-established vaccine is 90 percent effective in preventing infection. The Bandim group justifies this decision by noting tha

Ali WinstonSecurityMay 18, 2026 5:30 AMAn ICE Firearms Trainer Was Involved in At Least 4 Deadly ShootingsDavid Norman, a former Phoenix police officer who’s described himself as “a fucking savage,” now runs a company that provided training to Homeland Security’s Special Response Teams.Photo-illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe owner ofa company that trainedparamilitaryImmigration and Customs Enforcementagents testified that he was involved in at least four lethal shootings, according to a 2021 deposition related to a lawsuit reviewed by WIRED.David S. Norman, the founder and proprietor of law enforcement training firmTruKinetics LLC, served as a Phoenix Police officer from the late 1990s until his retirement in 2020. Prior to founding TruKinetics the same year, according to records reviewed by WIRED, Norman was involved in six shootings while on duty that left four people dead and two more wounded. In every instance, the Phoenix Police Department said Norman fired on an armed suspect and exchanged volleys of gunfire in at least two of the shootings.Based in Gilbert, Arizona, TruKinetics offers training on small-team tactics, hostage rescues, close-quarters combat, building searches, night-vision firearms proficiency, pistol and rifle courses, “vehicle interdiction,” breaching with explosives, and sniper tactics, according to the company’s website.TruKinetics received $27,748 for a year-long contract to run a mandatory40-hour training coursethat certain members of Department of Homeland Security Special Response Teams receive annually at Fort Benning in Georgia, according to governmentprocurement recordsreviewed by WIRED. At least 700 SRT agents from the Customs and Border Protection’s Office of Field Operations, ICE’s Homeland Security Investigations division, and ICE’s Enforcement and Removal Office units pass through Fort Benning for annual training.In an interview with WIRED, Norman says that his company conducted sessions with the Special Response Team from Arizona’s Homeland Security Investigations office. “They’re top dudes, and it was an honor to work with them,” he tells WIRED. Norman maintains that his courses, which took place in Arizona and Georgia’s Fort Benning, did not involve crowd control tactics or active shooters, but would not specify further. “It sounds like you’re one of those dudes who’s doing a hit piece on HSI,” he says.TruKinetics postedtwo photos to Instagramin August 2024 of Norman and three TruKinetics trainers alongside 19 uniformed operators from HSI’s Arizona Special Response Team, posing in a “kill house” training course—a set of rooms and hallways filled with obstacles and targets designed to simulate close-quarters combat.Customs and Border Protection did not respond to WIRED’s questions about how many SRT teams and operators went through the Gilbert, Arizona, company’s training course.In a 2021law enforcement podcast,The Modern Cop, Norman describes himself as “a fucking savage” who sought out high-risk experiences and shootings as a cop. “I wanted these experiences. I was super aggressive,” Norman said. He also appeared to joke about police shootings, telling the host that “you kind of hope it’s on your Friday, so you can actually have days off.”Once reserved for armed or high-risk suspects, manhunts, and potentially dangerous building entries, the SRTs are now being used for civil immigration enforcement, crowd control, and basic warrant service, operations that the unit wasonce restrictedfrom performing. BothRenee GoodandAlex Prettiwere killed while protesting the militarized federal immigration surges in Minnesota, with SRT membersimplicatedinbothof their deaths. While recent debate over Homeland Security’s violent immigration sweeps have focused on whether agents receiveadequate training, the background of SRT’s training contractor raises questions about who is training ICE’s and CBP’s paramilitary units, and what they are being trained to do.For a dozenyears of Norman’s two decades as a Phoenix cop, he served on the agency’s Special Assignments Unit, a plainclothes fugitive apprehension team that Norman repeatedly characterized in a 2021 deposition from a lawsuit filed the previous year as having morphed into what was considered a “SWAT” (special weapons and tactics) team. On those units, he worked as a “point cover” man and at times doubled up as a trainer on pistol usage for units within Phoenix PD, according to his deposition and Phoenix Police documents reviewed by WIRED.In addition to TruKinetics, Norman has worked as an “independent contractor for the Department of Defense” and provided training under a nondisclosure agreement, according to his deposition.Norman’s supervisors consistently gave him solid ratings and he received numerous department commendations. The four fatal shootings he was involved in were deemed in policy by both the local district attorney and Phoenix PD. Ho

Chris ColinThe Big StoryMay 18, 2026 6:00 AMI’m a Normie. Can Normies Really Vibe Code?Apparently anyone can vibe code anything these days. So Claude and I tried to make a database for tracking the petty grievances of the masses.Illustration: Yann BastardCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe dog thatushered me into the technological future was “low and thick.” That’s all my mother registered before it T-boned her in a city park earlier this year: dense, heavy, and traveling fast enough to fracture her right tibia. But enough abouther. Let’s discuss what this set in motion in my life: Having successfully learned nothing about coding for two and a half decades, I would soon be attempting my very first software development project.If you’ve ever had a low and thick dog break your mom’s shin bone, you know the stream of lesser indignities that follows. Case in point: the hours my father spent navigating phone trees, trying to manage my mom’s medical care. Are frustrating telephone calls significant in the grand scheme of things? No. But that stupid dog had chosen a technologically interesting moment to do its thing. For the first time in history, a problem no longer needed to be serious to bringserious toolsto bear.For as long as Silicon Valley has been selling a frictionless tomorrow, we ordinary people have been its passive shoppers—scrolling the App Store, hoping someone has gone to the trouble of building whatever we need. Enter AI and its democratizing sidekick:vibe coding. If the promise is real, suddenly we can build our own apps, as niche and trifling as we please, with zeroprogrammingskills. We merely gesture toward whatever irks us and a constellation of large language models, code generators, and development environments will click their heels.Illustration: Yann BastardNiche and trifling? That’s me! Where others vibe code résumé reviewers and inventory trackers andautomated assistantsto boost their work productivity, I had a different target in mind. Over the past couple years I’ve grown particularly fixated, personally and professionally, on what the policy world calls sludge: the rising tide of tiny administrative obligations that increasingly seems to define modern existence—and corrode our ability to get anything done. The hassle of dealing with insurance, or connecting that insurance to your doctor, or staying on top of airline miles, or navigating your kid’s school portal. The steps required to dispute a charge, or unsubscribe from a streaming service you forgot you had. Each of these feels like its own discrete assault on our time. But they’re not discrete. They’re separate mushrooms sprouting from the same mycorrhizal network.In a way this is a calibration issue. While bigger problems might at least theoretically attract attention—legislation, journalism, a Senate hearing—the smaller ones, too petty to litigate, simply become a fact of life. The arc of history may bend toward justice, but when it comes to fighting a one-dollar bank fee, it bends toward hold music.Which is where the fantasy of vibe coding captured my attention. Those hassles aren’t just accidental byproducts of complexity; they’re often features. A confusing portal, a dropped call, a process just opaque enough to discourage follow-through. At scale, they function less like bugs than like policy. The app I envisioned would expose this phenomenon, make the cumulative weight of these obligations a little harder to ignore. The image I’d like you to summon is a field of mushrooms trembling.What my momlacks in healthy legs, she makes up for in a Claude Pro subscription. Having needled her repeatedly over the past couple years about AI’s environmental, political, and economic implications, I brushed all that aside on a recent Sunday and drove to her house. After a little tibia talk, I opened her computer and began emitting vibes.I’d like to create a communally shared app that gathers and shares information related to how much time and energy we devote to fighting burdensome administrative tasks, bureaucratic sludge, Kafka-esque unsubscribe mazes, byzantine insurance portals, wrongful charges, denied claims, confusing membership plans, and the like.With as much clarity and detail as I could muster, I proceeded to describe a dashboard that would record the scale and scope of our collective sludge. Users would log frustrating incidents from their lives, entering how much time they’d spent, how annoying it was, and what they’d rather have been doing. Every submission would be dopaminally rewarded with an inspiring resistance quote and a photo of a kitten, puppy, or baby chimp. I’d train Claude to generate some “wider context”—a paragraph discussing how the frustrating incident fits into systemic sludge patterns—and a complaint letter to the relevant regulatory bodies.Claude noodled. Not for the first time, I feared my vibes would simply manifest an error page. I recalled, dimly, some of the advic

Emily MullinScienceMay 18, 2026 6:30 AMSupercharging Immune Cells May Help Control HIV Long-TermCAR-T cell therapy is already a potent treatment for certain cancers. Now, a small study is showing early promise for managing HIV.Photograph: Bloomberg/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyA Miracle cancertherapy that involvesengineering a patient’s own immune cellsis being repurposed for HIV, and early results from two individuals hint at its promise for long-term control of the virus.As part of a clinical trial, scientists took people’s own immune cells and reprogrammed them in a lab to recognize and attack HIV in the body. After a single infusion of the modified cells, two individuals with HIV now have undetectable levels of the virus—one for nearly two years and the other for almost a year. Both have been able to go off HIV medications entirely.The two people are part of a small study to test the treatment’s safety and feasibility. The initial findings were announced last week at the of American Society of Gene and Cell Therapy annual meeting in Boston.“These are early days. If we can provide the proof-of-concept that this approach is both safe and effective, then there are lots of ways in which it can be optimized, to make it more affordable and scalable,” says Steven Deeks, a professor of medicine and HIV expert at the University of California, San Francisco, who led the trial.The technique, known as CAR-T cell therapy, has been used intens of thousands of patientswith tough-to-treat cancers. Half a dozen or so drugs have been approved that rely on the technique. The treatment essentially supercharges a person’s immune system to directly attack and eliminate cancer cells. Recently, it’s also been used successfully to treatsevere autoimmune diseases.“This is pretty exciting,” says Andrea Gramatica, vice president of research at amfAR, the Foundation for AIDS Research, who was not involved in the trial. “The reason this study matters and is particularly important is because it gives the HIV field a real, clinical clue that teaching the immune system to control the virus without antiretroviral therapy is achievable.”Scientists have been pursuing a cure for HIV since the virus was first identified in the early 1980s. Antiretroviral therapy prevents the progression to AIDS by suppressing the virus to undetectable levels, but people must take medication for the rest of their lives. It has transformed HIV into a chronic condition that allows people to have a near-normal life expectancy. Yet not everyone who is HIV positive is aware of their status, and in some rural and low-income parts of the world these medications are still not widely accessible or affordable.Up until now, there areunder a dozen documented casesof sustained remission from HIV—known as a “functional cure” because the virus is still present in the body but is suppressed to levels that are undetectable by the immune system and HIV medication is no longer needed.Each of those individuals developed cancer and underwent stem cell transplantations as part of their treatment. In all butone of those cases, doctors used stem cells from donors with a rare genetic mutation called CCR5 that naturally prevents HIV from entering and infecting healthy cells. Timothy Ray Brown, known as the “Berlin patient,” was the first known person to be cured of HIV in this way in 2008.The examples of sustained remission “have taught us that the immune system can, under the right conditions, clear HIV,” says Boro Dropulić, executive director of the Maryland nonprofit Caring Cross, who developed the CAR-T therapy for HIV.But stem cell transplants aren’t scalable, he says. They’re intensive procedures that carry serious risks such as graft-versus-host disease, when the transplanted cells recognize the recipient’s cells as foreign and attack them.“What we’re trying to do is to engineer that outcome deliberately without requiring cancer, without requiring a specific donor,” Dropulić says. His organization is working on making advanced therapies like CAR-T more accessible and affordable.Cancer and HIV are similar in that both can hide from the body’s immune system. In CAR-T therapy for cancer, patients’ T cells are engineered to express chimeric antigen receptors, or CARs, on their surface. These added receptors allow T cells to specifically identify, lock onto, and destroy cancer cells by recognizing certain antigens, or proteins, on their surface. The specific CAR that’s added depends on the type of cancer being treated.Dropulić and his team engineered patients’ T cells to recognize two different sites on the HIV virus, making it more difficult for the virus to escape. “Our goal is that these cells remain like sentries in the body,” he says. “Whenever these embers of the virus start replicating, these cells are there in order to immediately take care of them.”The trial included nine total participants who were all on antiretr

Lindsay DodgsonScienceMay 18, 2026 7:00 AMThe Backward Logic of Chickenpox PartiesBefore vaccines, some parents intentionally exposed their children to the virus. In the internet age, that thinking is resurfacing.Photo-Illustration: Darrell Jackson; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAnyone who hashad chickenpox shares one distinct memory: the relentless, all-consuming itch.Ciara DiVita was only 3 years old when she caught the virus, but she remembers it well—along with the oven mitts she was made to wear to stop herself scratching. She also recalls being taken to hang out with her cousin while covered in blisters, in the hopes of deliberately infecting them.DiVita, now 30, was actually the second in the chain, having been taken by her parents to catch chickenpox from an infectious friend. “I imagine the chain continued and my cousin gave it to someone else at a chickenpox play date,” she says.A lot has changed over the past three decades, most notably the development of a chickenpox vaccine, meaning the virus is no longer the childhood rite of passage it once was.Thanks to the vaccine’s success, children today are much less likely to be exposed to the infection at school or on the playground.Chickenpox parties are also largely considered a relic of the past—a strategy many Gen X and millennial children were subjected to before vaccines became routine. But much like the virus itself—latent, opportunistic—they haven’t disappeared entirely.Before a vaccineexisted, chickenpox, which is caused by the varicella-zoster virus, felt unavoidable. In temperate countries like the UK and the US,around 90 percentof children caught the virus before adolescence (in tropical countries the average age of infection ishigher).It’s nothing to do with chickens. The splotchy, scratchy, highly contagious disease is possibly named after the French word for chickpea,pois chiche,according to one theory, because the round bumps caused by the virus resemble their size and shape. While most infant cases are mild, adolescents and adults are more likely to develop severe complications.This is where the idea of “getting it over and done with” emerged from, according to Maureen Tierney, associate dean of clinical research and public health at Creighton University in Omaha, Nebraska.“You were trying to have your child get the disease when they were at the greatest chance of not having complications,” Tierney says, explaining that, generally speaking, the older the patient, the more severe the infection can be.While varicella-zoster is usually a mild, self-limiting disease in children, it can be much more severe—and sometimes life-threatening—in adults.“I had an otherwise healthy adult patient who died of chickenpox pneumonia when I was first practicing,” Tierney says. “You never forget those scenarios.”The virus spreads rapidly through respiratory droplets and contact with fluid from its characteristic blisters, meaning if one child contracts it, siblings and classmates are likely to be next, if unvaccinated.Before the existence of social media, the idea that children should deliberately infect each other spread just as rapidly around communities—in conversations in the school yard, church groups, and pediatric waiting rooms—leading to the popularity of so-called chickenpox parties.Parents swapped advice about oatmeal baths and calamine lotion and arranged to bring children together when one was thought to be infectious—despite the practice never being an official medical recommendation.“They thought, well, if it’s going to happen to my kid anyway, it might as well happen in a controlled environment,” says Monica Abdelnour, a pediatric infectious disease specialist at Phoenix Children’s Hospital. “The families were ready to encounter this infection, deal with it, and then move on.”While the majority of children who develop chickenpox feel well again within a week or two, around three in every 1,000 infected experience a severe complication such as pneumonia, serious bacterial skin infections, encephalitis (inflammation of the brain), or meningitis.“Some kids get really, really sick,” says Jill Morgan, a professor at the University of Maryland School of Pharmacy and an expert in pediatric health. “The problem is, if you had these parties, you didn’t know which kids would get over it and be OK, and which kids would end up in the hospital.”Chickenpox parties faded into obscurity once routine vaccination began to take hold in many parts of the world; a2018 analysisof global vaccination trends and infection rates shows cases decreased dramatically in countries such as the US, Germany, and Australia, where universal childhood vaccination programs have already been rolled out.In the US, the vaccine has been associated with a roughly 97 percent drop in reported cases since its introduction,according to data published by the Centers for Disease Control and Prevention. Global data show similarly st

Jeremy WhiteGearMay 18, 2026 9:38 AMThe Catastrophic Audemars Piguet x Swatch Launch Was Entirely Predictable and Utterly AvoidableChaos consumed Swatch stores worldwide on Saturday—a repeat of the MoonSwatch launch four years ago. Swatch should have seen it coming.Photograph: RAPHAELLE PELTIER/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAt 9 am,as stores opened, London’s Carnaby Street location lasted half an hour before police were called. In New York, fights broke out amid rumors of a stabbing in the line. At locations around the world, police had to assist staff as the vast majority of customers left empty-handed, since most of the shops had fewer than 200 watches available. Deals were done within queues as scalpers flipped watches for profit. Within minutes of the first sales, the plastic watches were hitting eBay for thousands of dollars.This is not a description of what happened during Saturday’sSwatch x Audemars Piguet Royal Poplaunch; this is an account of the 2022 MoonSwatch launch. But a carbon copy of this chaos and mayhem ensued as Swatch stores opened worldwide on May 16 to sell their extremely limited numbers of Royal Pop pocket watches.Responding to widespread criticism of how Swatch handled the 2022 MoonSwatch launch, Nick Hayek Jr., chief executive of Swatch Group,told WIREDat the time: “We knew for sure this would be a success, because the product is beautiful, provocative, high quality, and the price is fantastic … But what happened … I think nobody in the world could have expected that. It was really crazy.”Hayek’s defense was essentially that Swatch could not be held accountable for disturbances so severe that some stores were forced to shut for 10 days to let the pandemonium play out, simply because the group could never have predicted the ferocious public reaction to an affordable version of an iconic luxury timepiece coming on sale. No such excuse can be offered this time for the calamitous Royal Pop launch that took place this weekend.Chaos ForeshadowedEven before the first stores opened in Singapore on Saturday, 12 hours ahead of New York, the clear pattern of what was to follow was established. TheVivoCity branch was shutteredahead of the much-anticipated launch, with Swatch issuing a statement that it was “due to the overwhelming crowd” and that this decision was taken with local authorities to ensure the health and safety of staff and customers.The Swatch X AP Royal Pop that’s caused such a furor at launch, just like the MoonSwatch did in 2022.Courtesy of SwatchThen it was Dubai. “In view of safety considerations, we have decided not to proceed with the sale of the [Swatch x Audemars Piguet Royal Pop] at Dubai Mall and Mall of the Emirates and the event has been cancelled,” a statement said. Authorities in New Delhi and Mumbai closed Swatch Store openings citing a lack of crowd control threatening safety. “We are not animals,” crowd members were caught shouting on Instagram clips.Londonstores shut, andpolice dogs were deployed. Elsewhere in the UK, in Birmingham, news reports spoke of “fights and knives” and drug-taking before police were called. A“mosh pit”developed outside stores in the Netherlands. Clashes inDusseldorf. Officersfired teargasto control a 300-strong crowd outside a Swatch shop in Paris. Openbrawling occurredin Milan. Thenarrestsand shutdownin Miami. Houston’s store closed, as well asChicago’s. In Long Island, New York, amazingly,pepper spraywas usedfor crowd control. In store-front interviews, customersmade clear their angertowards Swatch on how poorly the release was handled.Eventually, around 2 pm ET, a staggering 17 hours after the first clear warning signs in Singapore, Swatch finallyissued a statementon its Instagram page: “To all our dear fans worldwide of our AP x Swatch collab, launched on May 16. To ensure the safety of both our customers and our staff in Swatch stores, we kindly ask you not to rush to our stores in large numbers to acquire this product. The Royal Pop Collection will remain available for several months. In some countries, queues of more than 50 people cannot be accepted, and sales may need to be paused.”Looking at almost any of the 45,000 comments on the post, the response to the brand from fans is crystal clear: Swatch had done far too little, far too late. “You intentionally put people in danger causing a frenzy knowing how many watches you sent to each location but allowing thousands of people to line up was negligent.” “Shame on you Swatch.” “Swatch did bad—they should have gotten ahead of this knowing the depth of inventory and allocation rather than putting stores and customers at risk”, and on the replies go.AP Wanted a Safe, Positive ExperienceWIRED contacted Swatch asking why, considering the wealth of evidence from the ill-received MoonSwatch launch, did it repeat the rollout strategy with the AP collaboration. We got this response from their communications team in Switzerland bl

John BrandonGearMay 18, 2026 10:33 AMReview: Leica Cine Play 1Leica’s first home entertainment projector is pricey, but like the company’s cameras, the image quality is worth the splurge.Courtesy of B&H PhotoTriangleUpBuy NowMultiple Buying Options Available$3,795$2,995at B&H$3,795$2,995at Leica$3,795$2,995at CrutchfieldCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyRating:9/10Open rating explainerInformationWIREDLeica’s Summicron lens renders colors beautifully. Compact size allows for flexibility in moving around the home. Straightforward setup; auto-sizing and focus were almost magical.TIREDLooks best in a dark room. Remote has a confusing button layout. Expensive.If your preferenceis to make colors pop on the screen for movie night, the Leica Cine Play 1 delivers with an immersive cinematic experience. This short-throw laserprojectorfrom the company that makes high-end cameras with a cult following ticks almost every box for tack-sharp picture quality, rendering Pixar films and theAvatarseries with vibrant colors and beautiful details.Crafted out of aluminum and glass, Leica’s latest projector borrows many of its design ethos and superb build quality from thecompany’s cameras. It comes with a carrying handle, making it easy to move around the home, and is powered by the intuitive VIDAA operating system. The Leica optical performance makes movies come alive.At $2,995 after an $800 rebate, the Cine Play 1 is competitive with midrange home cinema projectors from Sony, Epson, and others. The only minor ding against the Cine Play 1 is that you’ll need a mostly dark room to take advantage of Leica’s brilliant color reproduction, while competitors like theEpson Pro Cinema LS9000perform well in both dark and well-lit environments.Setting Up the ProjectorThe Cine Play 1 uses a Leica Summicron lens with an outstanding optical zoom. Known for its extreme sharpness and near-zero color fringing on Leica’s storied cameras, the Summicron glass delivers similar benefits on the Cine Play 1. With the Cine Play 1, it’s like having a high-end camera that, instead of capturing a scene, projects one with added vibrancy. Though Leica has made optical lenses for other projectors, this partnership with Hisense makes this the first projector to bear Leica’s branding, complete with the iconic red dot on the top.Photograph: John BrandonThough seemingly mobile at about the size of a large lunch box and weighing 14.6 pounds, the lack of a built-in battery means that the Cine Play 1 is mostly confined to indoor use. Like most other projectors, you can mount the Cine Play 1 on your own stand or opt for Leica’s optional $495 Cine Play 1 Floor Stand. Though expensive, the Floor Stand has an ingenious design that is Apple-like in approach. After mounting the projector to the stand with a couple of bolts, I noticed a proprietary connection, allowing you to connect the power cord into the base of the stand rather than on the rear of the Cine Play 1, solving the problem of unwieldy wires.After mounting the Cine Play 1, I set up the projector roughly 12 feet from the wall. This resulted in a 150-inch image. The Cine Play 1 can project 65- to 300-inch images, but for smaller picture sizes, it would be easier to get a4K TVinstead. No matter how large the projection I was testing, the focus, brightness, and color quality were all above average compared to other projectors.The auto keystone and autofocus features were almost magical, adjusting to my projector screen automatically. I did have to turn off the auto-adjusting after bumping into the projector a few times, but the idea is that it will adjust if you move the Cine Play 1 to a different room.Photograph: John BrandonPowering on the Cine Play 1, I was pleasantly surprised at how vibrant and colorful everything looked in a dark room.The onscreen setup took only a few minutes. The Cine Play 1 uses Hisense’s VIDAA operating system, so I created a new profile and logged into the companion app. At the top of the screen, there’s a series of banners that allow you to access the app store, art mode, and scan a QR code to install the mobile app. I found the OS to be highly intuitive. There’s also a highly effective voice search option if you would rather not rely on the included remote.VIDAA provides access to most streaming services, with Fandango at Home being a notable exception. (Leica reps stated that they may add this service at a later date.)Benchmark Results and First ImpressionsApart from minor brightness issues when using the Cine Play 1 in brighter environments, this projector passed my benchmark tests. Colors were vibrant, skin tones did not appear washed out, and a butterfly scene showed subtle differences between the shades of yellows and oranges in my test clips. In comparison, colors on theTCL NXTVISIONwere noticeably duller. The Cine Play 1 also excels as a gaming display, with superb color reproduction and stutter-free performance when tested withCrims

Paresh DaveMaxwell ZeffBusinessMay 18, 2026 1:40 PMElon Musk Loses Landmark Lawsuit Against OpenAIThe nine-member panel took only two hours to return a verdict in favor of OpenAI on Monday, which the judge quickly adopted as her own final decision.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyElon Musk sufferedthe worst defeat possible inhis legal battle against OpenAIas a federal jury and a judge ruled he waited too long to bring his claims against the AI startup and its top executives, Sam Altman and Greg Brockman.While the jury’s decision was a nonbinding recommendation sent to US district judge Yvonne Gonzalez Rogers, she immediately accepted it as her own, making it final.Musk’s lead trial attorney, Steven Molo, told the judge, “Our intention is to appeal.”One of his other attorneys, Marc Toberoff, gave a one-word comment to reporters walking out of the courtroom: “Appeal.” He later said the verdict reminded him of American Revolutionary War moments like the Siege of Charleston and the Battle of Bunker Hill. “These were major losses for Americans, but who won the war?” Toberoff said. “And this one is not over.”OpenAI’s attorneys hugged in the courtroom after the verdict was read. William Savitt, the company’s lead litigator, told reporters that the “overwhelming” amount of evidence presented in the case allowed the jury to act quickly. “The evidence that Mr. Musk’s lawsuit was an after-the-fact contrivance by a competitor was overwhelming,” he said.Throughout the trial, Gonzalez Rogers questioned Musk’s motivation for fighting OpenAI. But she concluded on Monday that the three-week global public spectacle had been worthwhile.“I thought it was an important issue to be tried … for us to have a trial to bring clarity,” she told attorneys for both parties. “There’s a substantial amount of evidence to support the jury’s findings, which is why I was prepared to accept the jury’s findings and dismiss on the spot.”Thenine-member paneldelivered the unanimous verdict in an Oakland, California, courtroom on Monday after deliberating for under two hours. They found that statutes of limitations expired well before Muskfiled his lawsuitin 2024. Musk had hoped to persuade the jury that Altman and Brockman, with the help of Microsoft’s cash, transformed OpenAI into an enormous company well beyond what was envisioned when the three of them and others founded it as a nonprofit nearly 11 years ago.Because the jury found the case wasn’t filed on time, it didn’t weigh in on Musk’s three claims, including breach of charitable trust, unjust enrichment, and, against Microsoft, aiding and abetting. Losing on what amounts to a technicality could provide Musk an opening to keep trying his case in the public by arguing that the jury never ruled against his core argument that a charity was stolen.Savitt, the OpenAI lawyer, disputed that contention on Monday. “It’s not a technical decision, it’s a substantive one,” he said. “It says you brought your claims too late, and you did it because you were sitting on them to use them as a weapon of a competitor who can’t compete in the marketplace, and so we’re delighted with the outcome.”Microsoft spokesperson Alex Haurek said in a statement that the “facts and the timeline in this case have long been clear” and the tech giant remains “committed to our work with OpenAI to advance and scale AI.”Inpostson X, Musk called Gonzalez Rogers an “activist” judge “who simply used the jury as a fig leaf” and ruled on a “calendar technicality.” While the judge’s ruling doesn’t create a formal legal precedent, Musk claimed that, “She just handed out a free license to loot charities if you can keep the looting quiet for a few years!” He added in a separate post, “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity.”Despite the disappointing result for Musk, the trial appears to have tarnishedthe public image of OpenAI and its top executives. New details emerged aboutBrockman’s wealthandAltman’s alleged history of dishonesty. Both were also pulled away from their day-to-day work for tens, if not hundreds, of hours to conduct depositions, prepare to testify, sit on the witness stand, and show face in court.Musk spent much less time in the courtroom than the OpenAI executives, about three days before never returning again. He even flew to China for President Donald Trump’s state visit last week, though he technically could have been called to testify again on short notice. “I will say that it was a surprise to us to see that,” Savitt told the media last week. “Instead of being in the jurisdiction where he filed the lawsuit, ready to come in front of the jurors who he has caused to be impaneled, [he] decided to get on Air Force One and go to China.”Though the casecarried financial and emotional stakes, it was also a competition of bravado between two t

Molly HigginsGearMay 18, 2026 2:06 PMThe Best Automatic Cat FeedersWe tested some of the most popular automatic dry- and wet-food pet feeders to see which ones are worth the money.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyFeatured in this articleBest AI-Enabled Dual Feeder With CameraPetkit YumShare Dual-Hopper 2 Automatic Feeder With CameraRead more$170PetkitBest Automatic Wet-Food FeederPetlibro Polar Wet Food FeederRead more$170PetlibroBest Connected AppPetlibro Granary Smart Camera FeederRead more$124$107(14% off)AmazonA Dual Automatic Feeder With a CameraPetkit YumShare Dual-Hopper Automatic Feeder With CameraRead more$160$110(31% off)WalmartAn automatic catfeeder just may change your life. No more screeching at 4 am. Plus, it can give some peace of mind while you’re away from your fur baby. As seen with the rise ofautomatic litter boxesandpet tracking, technology has made it so we can closely monitor our pets without having to do a lot of the manual labor normally involved.Gravity-fed automatic dry-food dispensers were invented nearly a century ago, but today most automatic cat feeders utilize Wi-Fi and are connected to an app on your phone that allows you to make a schedule or feed on demand from afar, all while monitoring your pet’s eating habits. Newer models have taken this tech further, like thePetkit YumShare Dual-Hopper 2 Automatic Feeder With Camera ($170)which has a built-in, AI-enabled camera to analyze the remaining food in the bowl and will stop dispensing to prevent overfeeding. Automatic wet-food dispensers—our top pick in this category is thePetlibro Polar Smart Wet Feeder ($170)—are relatively new, often using ice packs or mini-refrigeration cooling systems to ensure the food stays as fresh as possible.Automatic feeders are also helpful for pet parents who want to monitor their pets’ health, have specialized schedules, or want to foster healthier eating habits. (A customizable automatic cat feeder can help with the endless gorging-vomiting cycle many cats struggle with.) Nearly all of the automatic cat feeders on this list are marketed toward cats, but some can be used with smaller dogs, and I’ve noted these below.Check out our related guides, like theBest Automatic Litter Boxes,Best Cat Water Fountains,Best Smart Cat Collars, andBest Pet Cameras. More of a dog person? We’ve got you covered; see ourBest Accessories and Tech Essentials for Your Dog.Updated May 2026: We’ve added the Cheerble Match G1 and Oneisall 2-in-1 Automatic Cat Feeder and Water Dispenser, and updated prices and links throughout.Jump to SectionAccordionItemContainerButtonLargeChevronCompare Our PicksOthers TestedDo Not RecommendWhy Use an Automatic Feeder?How I TestedHow Long Can I Leave My Cat Alone?Best AI-Enabled Dual Feeder With CameraPhotograph: Molly HigginsCourtesy of PetkitChevronChevronSave to wishlistSave to wishlistPetkitYumShare Dual-Hopper 2 Automatic Feeder With Camera$170Petkit$170ChewyThis model is an upgrade on one of my favorite models below, the Petkit YumShare Dual-Hopper Automatic Feeder With Camera, but with added AI features. It has a crystal-clear video feed, night vision, customizable meal calls, and a dual-hopper for two different types of food. Through the connected app, you can set feeding schedules or feed on demand from both tanks, use two-way audio, and watch the feeder from your phone. Unlike most other brands, the AI-enabled camera doesn’t require a subscription and can identify up to 20 pets and log when they’ve eaten individually. The camera also uses AI to analyze the remaining food in the bowl and will stop dispensing to prevent overfeeding.This feeder worked excellently throughout testing, and I like that you don’t need to opt in to the subscription plan to have AI recognize different pets. However, I still think the connected Petkit app isn’t as reliable or as easy to use as Petlibro’s app (above).Compatible food typesDry; freeze-driedAppYesPower sourcePower cordCapacity2 and 3 liters (5 total)Warranty1 year limitedWIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDBuilt-in AI-enabled cameraCan store two types of food (including freeze dried)TIREDNeed membership plan to access some featuresApp isn’t as user-friendly as some othersBest Automatic Wet-Food FeederPhotograph: Molly HigginsCourtesy of AmazonChevronChevronSave to wishlistSave to wishlistPetlibroPolar Wet Food Feeder$170Petlibro$170Amazon$170ChewyThis updated Petlibro Polar Wet Food Feeder solves pretty much every gripe I had with other automatic feeders aimed at storing wet food—it’s essentially a mini-fridge guaranteed to keep food fresh, and it’s connected to an app where you can schedule and customize feedings and keep a feeding log. Food is given in three stainless steel wide, shallow bowls, which means less chance forwhisker fatiguewith wide bowls orfeline acnewith bacteria-harboring plastic. When it’s time to feed, the cover slides back from the hole and reveals the food be

TechAINewsAmazon Alexa Plus can now create AI-generated podcastsWith Alexa Podcasts, a pair of AI-generated hosts break down the topic of your choice.With Alexa Podcasts, a pair of AI-generated hosts break down the topic of your choice.byEmma RothMay 18, 2026, 3:55 PM UTCLinkShareGiftEmma Rothis a news writer who covers the streaming wars, consumer tech, crypto, social media, and much more. Previously, she was a writer and editor at MUO.Alexa Plus, Amazon’s upgraded AI assistant, can now generate podcasts on “virtually any topic,” according to anannouncement on Monday. With the update, Amazon says you can give Alexa Plus a topic, and the AI assistant will offer an overview of what its AI hosts plan to talk about, allowing you to steer the conversation and adjust its length before it starts generating the episode.Some “Alexa Podcast” examples shared by Amazon have two AI-generated hosts talking about the history of the Roman Empire, new music, and expectations for the World Cup. Amazon says you can also ask Alexa Plus to generate audio lessons about the Apollo missions, or create a podcast episode about a new hobby, like photography. The feature is similar to theAI-generated podcasts you can createusing NotebookLM, a capability Microsoft Edge also snapped up ina recent update.Image: AmazonAmazon says the AI-generated podcast episodes will draw information from the 200 news publications the company has partnered with, includingReuters, theAssociated Press, theWashington Post,Vox, andPolitico. Amazonrolled out Alexa Plusto all users in the US in February and has since embedded theassistant into its online marketplace.You’ll receive a notification on your Echo Show device and the Alexa app once the episode is finished. You can tap the notification to play the episode, as well as access it through the Music and More section on the Echo Show, or with the Alexa app.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Emma RothAIAmazonAmazon AlexaNewsTechMost PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

GamingNewsPlayStationSony is raising short-subscription prices for PlayStation PlusOne-month and three-month PlayStation Plus subscriptions are getting more expensive.One-month and three-month PlayStation Plus subscriptions are getting more expensive.byJay PetersMay 18, 2026, 4:24 PM UTCLinkShareGiftJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Sony ishiking the starting priceof one-month and three-month PlayStation Plus subscriptions in “select regions,” blaming “ongoing market conditions.” Beginning May 20th, 1-month subscriptions will start at $10.99 USD / €9.99 EUR / £7.99 GBP, while 3-month subscriptions will start at $27.99 USD / €27.99 EUR / £21.99 GBP.In the US, the changes result in a $1 price increase for a one-month Essential tier subscription and a $3 increase for a three-month Essential tier subscription. It’s unclear if other PlayStation Plus tiers will see price hikes as well. PlayStation didn’t immediately reply to a request for comment.If you are a current subscriber, prices will stay the same for you (except in Turkey and India, according to Sony) unless you change your subscription tier or let your subscription lapse.The PlayStation Plus price hikes follow two recentpriceincreasesfor PS5 consoles. Sony said its April PS5 price hike wasdue to“continued pressures in the global economic landscape.”Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGamingNewsPlayStationMost PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechDesignNewsGoogle is rolling out its redesigned Workspace app iconsThe new designs feature soft color gradients, rounded corners, and overhauled icon shapes.The new designs feature soft color gradients, rounded corners, and overhauled icon shapes.byStevie BonifieldMay 18, 2026, 4:29 PM UTCLinkShareGiftImage: The VergeStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.It’s not just you — the Google Workspace apps are getting a new look. The redesigned app icons,leaked last month, are now rolling out widely, as we started noticing this morning. Users with the redesigned icons will notice they now have a gradient look that fades from lighter to darker shades, rather than being the same flat tone throughout, similar to the redesigned Google logothat launched a year ago.Image: The VergeSome of the icons switched from a rainbow design to a single color, like Google Chat, Meet, and Calendar, which could help all of the icons stand out a bit more from one another — or make them harder to recognize. Others haven’t changed as much, as9to5Googlenotes, like Google Docs, Sheets, and Slides, which have the same colors and general approach. Asone Reddit userpointed out, though, the Sheets and Slides icons were switched to landscape mode, which is also how most people use both of those apps.Image:9to5GoogleThe Google Drive icon also got a significant overhaul with rounded corners (and it lost the spot of red it had in the bottom right corner). A few also lost the borders around them, like Google Keep, which is now just a yellow light bulb, rather than a bulb on a yellow rectangle. As someone who frequently uses Keep, it’s going to take me a while to get used to spotting a totally different icon shape, although I like the overall look of the new one. I also appreciate the cleaner look of the Gmail icon, although it hasn’t changed too much compared to the old version.The new icons are arriving right beforeGoogle I/O kicks off on Tuesday,where Google could announce even more visual changes to its ecosystem.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAppsDesignGoogleNewsTechMost PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AINewsPolicyElon Musk loses his case against Sam AltmanThe Musk v. Altman jury dismissed all claims due to the statute of limitations.The Musk v. Altman jury dismissed all claims due to the statute of limitations.byHayden FieldMay 18, 2026, 5:39 PM UTCLinkShareGiftImage: Alex Parkin / The VergePart OfAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIsee all updatesHayden Fieldis The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.After around two hours of deliberation, the jury has reached a unanimous verdict inMusk v. Altman, the tech trial of the year. The group found that two claims were barred by the statute of limitations, and a third failed thanks to the dismissal of one of these.The jury here is an advisory jury, meaning the group is installed solely to offer another opinion to the judge, and its verdict is technically not legally binding. Ultimately, US District Judge Yvonne Gonzalez Rogers is the ultimate legal authority — and she accepted the decision.The jury found that Musk’s claim for breach of charitable trust was barred by the statute of limitations, and the claim that Microsoft aided and abetted such a breach failed with it. Restitution is also barred by the statute of limitations, the jury found.OpenAI did not immediately respond to a request for comment. On X, Muskposteda statement saying he’d be filing an appeal. He said that the judge and jury ruled “on a calendar technicality” rather than the “merits of the case” and that “there is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!”Alex Haurek, a Microsoft spokesperson, said in a statement, “The facts and the timeline in this case have long been clear, and we welcome the jury’s decision to dismiss these claims as untimely. We remain committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”RelatedMusk v. Altman proved that AI is led by the wrong peopleMusk vs. Altman is here, and it’s going to get messyMusk v. Altmanhas taken over a federal courtroom in Oakland for three weeks, with the core accusation being that OpenAI strayed from its founding mission and that Musk’s money was earmarked for a nonprofit in particular. Musk alleges that OpenAI CEO Sam Altman and company president Greg Brockman breached OpenAI’s charitable trust and participated in unjust enrichment at Musk’s expense. He also alleges that Microsoft aided and abetted the two in breach of charitable trust. Both sides have used every opportunity to smear each other — and throughsalacious evidenceandeyebrow-raising testimony, both sides have come out looking somehow even less trustworthy than when the court process began.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Hayden FieldAIElon MuskLawMicrosoftNewsOpenAIPolicyTechMore in:All of the updates from Elon Musk and Sam Altman’s battle over OpenAIMusk v. Altman proved that AI is led by the wrong peopleHayden FieldMay 18The jury has delivered a unanimous verdict.Richard LawlerMay 18An observer has just been ejected from the court by the US marshals.Elizabeth LopattoMay 18Most PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechUpdatedToday,May 18, 2026, 7:00 PM UTCAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIbyElizabeth LopattoandHayden FieldLinkShareRSSGiftSam Altman and Elon Musk are facing off in a high-stakes trial that could alter the future of OpenAI and its most well-known product,ChatGPT. In2024, Musk filed a lawsuitaccusing OpenAI of abandoning its founding mission of developing AI to benefit humanity and shifting focus to boosting profits instead.After nearly a month of the hearing that featured testimony from Musk, Altman, Microsoft CEO Satya Nadella, OpenAI cofounder Greg Brockman, former OpenAI board member and mother of several of Musk’s children Shivon Zilis, and a few others, the jury deliberated for a couple of hours before returning with a verdict,deciding to dismiss all chargesdue to the statute of limitations.Musk was a cofounder of OpenAI and claims that Altman and Brockman tricked him into giving the company money, only to turn their backs on their original goal. However,OpenAI claimedthat “This lawsuit has always been a baseless and jealous bid to derail a competitor” in a bid to boost Musk’s own SpaceX / xAI / X companies that have launched Grok as a competitor to ChatGPT.In his lawsuit,Musk asked for the removal of Altman and Brockman, and for OpenAI to stop operating as a public benefit corporation.People to KnowPlaintiffElon Musk— plaintiff, OpenAI cofounder and now CEO of rival xAISteven Molo— lead counsel for the plaintiffJared Birchall— manager of Musk’s family officeShivon Zilis— former OpenAI board member who shares multiple children with MuskDefendantSam Altman— defendant, CEO of OpenAIWilliam Savitt— lead counsel for the defendantGreg Brockman— president of OpenAI as well as a cofounderIlya Sutskever— former chief scientist at OpenAI and a cofounderJudgeYvonne Gonzalez Rogers— aka YGR, trial judgeHere’s all the latest on the trial between Musk and Altman:HighlightsElon Musk loses his case against Sam AltmanMicrosoft doesn’t want any of thisSam Altman was winning on the stand, but it might not be enoughMusk’s biggest loyalist became his biggest liabilityMira Murati tells the court that she couldn’t trust Sam Altman’s wordsGoogle’s AI architect lived rent-free in Elon Musk’s headOpenAI’s president does ‘all the things,’ except answer a questionElon Musk’s worst enemy in court is Elon MuskElon Musk appeared more petty than preparedMusk vs. Altman is here, and it’s going to get messyPinnedToday,5:39 PM UTCHayden FieldElon Musk loses his case against Sam AltmanImage: Alex Parkin / The VergeAfter around two hours of deliberation, the jury has reached a unanimous verdict inMusk v. Altman, the tech trial of the year. The group found that two claims were barred by the statute of limitations, and a third failed thanks to the dismissal of one of these.The jury here is an advisory jury, meaning the group is installed solely to offer another opinion to the judge, and its verdict is technically not legally binding. Ultimately, US District Judge Yvonne Gonzalez Rogers is the ultimate legal authority — and she accepted the decision.Read Article >Today,7:00 PM UTCHayden FieldMusk v. Altman proved that AI is led by the wrong peopleImage: Cath Virginia / The Verge, Getty ImagesThe tech trial of the year,Musk v. Altman, was ultimately a fight for control. Elon Musk argued that Sam Altman, with whom he helped found the now-massive company OpenAI, shouldn’t direct the future of AI. Altman’s lawyers, in turn, poked at Musk’s own credibility. A jury came to averdicton Monday after just two hours of deliberation, dismissing Musk’s claims due to the statute of limitations.In a strictly legal sense, three weeks of testimony added up to nothing. But the trial offered a more damning broader takeaway: Almostnobody in this saga seems worth trusting. Some of the most powerful people in tech seem temperamentally incapable of dealing with each other honestly. And if that’s true, it raises a bigger question: Why are they in control of atrillion-dollarindustry that’s set to upend people’s lives?Read Article >Today,5:29 PM UTCRichard LawlerThe jury has delivered a unanimous verdict.That was quick (about two hours).Elon Musk lost his case against Sam AltmanHayden FieldMay 18Today,5:00 PM UTCElizabeth LopattoAn observer has just been ejected from the court by the US marshals.I assume because he was recording, since the marshal said, “Give me your phone.” There have been several incidents of people attempting to record or take pictures throughoutthe trial— but I honestly am not sure why you’d record today of all days.Today,4:44 PM UTCElizabeth LopattoC. Paul Wazzan is the expert called by Musk to determine damages.Unfortunately he does not have a lot of details YGR is asking for. He doesn’t know how many investments Musk has made (11 to date, according to Pitchbook), or how many were successful. He’s getting some pretty tough questioning from YGR in the direct exam. Among the things she’d asked, he didn’t kn

GadgetsTechSmart HomeDyson’s super-slim PencilWash just hit its best price to date for Memorial DayThe slim hard-floor cleaner slides under furniture, scrubs away spills, and removes dirty water as it cleans.The slim hard-floor cleaner slides under furniture, scrubs away spills, and removes dirty water as it cleans.bySheena VasaniMay 18, 2026, 6:32 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: DysonSheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.IfDyson’s PencilVac Fluffyconesmade you wish the company had built something similarly slim for scrubbing the hard floors in your home, enter the recently releasedDyson PencilWash. The cordless cleaner is designed to tackle spills and stains on tile and other surfaces, and right now, it’s available for a new low price of $249.99 ($100 off) fromAmazon,Target, andDysonin the run-up to Memorial Day.Dyson PencilWash$250$35029% off$250$250 at Amazon$250 at Target$250 at DysonWith its 1.5-inch-diameter handle and a design that lets it lie nearly flat, the PencilWash can easily slide under furniture and other tricky areas that bigger floor cleaners can’t reach. Built for hard surfaces like tile, laminate, and sealed wood, it uses a pressurized hydration system and a microfiber wet roller to scrub away spills while extracting dirty water with each pass, so you’re always cleaning with fresh water. There’s also a Max mode for addressing dried-on messes that are particularly tough to clean, along with a flexible neck that makes it easier to steer around obstacles.RelatedDyson put someone else’s motor in its robot vacuumHere are 40 of our favorite deals from REI’s massive Anniversary SaleThese new Roombas are smaller and cheaperThe PencilWash is designed to be low-maintenance, too. Debris and wastewater are collected directly in the head rather than traveling up through the machine, preventing them from becoming trapped in a filter. It shouldn’t leave your floors soaking wet, either, and it comes with a sleek charging dock that lets it eke out 30 minutes per charge (you can also purchase abackup batteryfor $99.99). The 300ml clean water tank should be enough to cover over 1,000 square feet on the lowest hydration setting as well, after which it will need a refill.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Sheena VasaniDealsGadgetsSmart HomeTechVerge ShoppingMost PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechNewsGoogleWalmart launches new budget-friendly Android tablets starting at $97The six new Onn tablets ship with Android 16 and include kids’ tablets, Core models, and a 13-inch Pro model.The six new Onn tablets ship with Android 16 and include kids’ tablets, Core models, and a 13-inch Pro model.byStevie BonifieldMay 18, 2026, 6:44 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Image: WalmartStevie Bonifieldis a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.Walmart’s Onn brand just launched a whole line of budget-friendly Android tablets that, all together, cost less in total than a single iPad Pro. As spotted by9to5Google, the six new tablets ship with Android 16 and all but one cost less than $200.The littleOnn Core 7is the most affordable of the bunch at just $97, which gets you a 7-inch 1040 x 600 IPS LCD display, 4GB of RAM, 64GB of storage that’s expandable with a microSD card, and, according to Walmart, up to 10 hours of battery life.PreviousNext1/3The Onn Pro includes a folio case and stylus.Image: WalmartA step above that are two larger Core tablets. TheCore 8.1costs $138 and includes an 8.1-inch 1524 x 1000 IPS LCD display, 6GB of RAM, 64GB of storage (also expandable with microSD), and up to 15 hours of battery life. TheCore 11for $167 is a bit bigger, similar to a base iPad, featuring an 11-inch 1840 x 1280 IPS LCD display, 6GB of RAM, 128GB of storage that’s expandable with a micoSD slot, and up to 17 hours of battery life. There are also two affordable kids’ tablets available in8-inchand11-inchsizes for $118 and $136, respectively.The priciest of the new tablets is theOnn Pro, which costs $288 and includes a 13.2-inch 2400 x 1600 IPS LCD display, 8GB of RAM, 256GB of storage, and an unnamed 2.6GHz MediaTek processor. Unlike the iPad Pro it’s imitating, it also includes a folio case and a stylus.Onn Core 7 tablet$97$97$97 at Walmart (blue)$97 at Walmart (silver)$97 at Walmart (mocha)Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Stevie BonifieldAndroidGoogleNewsTechMost PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AIPolicyReportMusk v. Altman proved that AI is led by the wrong peoplePublic opinion of the AI industry is already sinking. A parade of untrustworthy executives makes it look worse.byHayden FieldMay 18, 2026, 7:00 PM UTCLinkShareGiftImage: Cath Virginia / The Verge, Getty ImagesAIPolicyReportMusk v. Altman proved that AI is led by the wrong peoplePublic opinion of the AI industry is already sinking. A parade of untrustworthy executives makes it look worse.byHayden FieldMay 18, 2026, 7:00 PM UTCLinkShareGiftPart OfAll of the updates from Elon Musk and Sam Altman’s battle over OpenAIsee all updatesHayden Fieldis The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.The tech trial of the year,Musk v. Altman, was ultimately a fight for control. Elon Musk argued that Sam Altman, with whom he helped found the now-massive company OpenAI, shouldn’t direct the future of AI. Altman’s lawyers, in turn, poked at Musk’s own credibility. A jury came to averdicton Monday after just two hours of deliberation, dismissing Musk’s claims due to the statute of limitations.In a strictly legal sense, three weeks of testimony added up to nothing. But the trial offered a more damning broader takeaway: Almostnobody in this saga seems worth trusting. Some of the most powerful people in tech seem temperamentally incapable of dealing with each other honestly. And if that’s true, it raises a bigger question: Why are they in control of atrillion-dollarindustry that’s set to upend people’s lives?OpenAI was, in the testimony of both Musk and Altman, founded to stop powerful AI from being owned and advanced by the wrong people. Testimony and evidence showed its founding team fretting about who would control artificial general intelligence (AGI), a buzzword for AI that broadly equals or surpasses human knowledge and ability. Theydeeply fearedGoogle DeepMind and its leader, Demis Hassabis. In 2015, Altman said he’d been mulling over whether anything could “stop humanity from developing AI” — and after concluding it was impossible, that he wanted “someone other than google to do it first.”Fellow cofounders Greg Brockman and Ilya Sutskever so strongly opposed one-person control that they seemed willing to torpedo a lucrative deal that could — in their words — give Musk an “AI dictatorship.” In a part of the same email addressed to Altman, Brockman and Sutskever questioned his motivations, writing, “We haven’t been able to fully trust your judgements throughout this process … Is AGI truly your primary motivation? How does it connect to your political goals?”These concerns would bequickly borne out. A central focus ofMusk v. Altmanwas “the blip,” a five-day period in November 2023 when OpenAI’s board removed Altman as CEO. Sutskever had spentmore than a yeararchitecting his ouster, assembling a 52-page memo alleging “a consistent pattern of lying, undermining his execs, and pitting his execs against one another.” The implications were broader than executive infighting, potentially impacting the public rollout of AI systems. Then-CTO Mira Murati, for instance, testified in court that Altman told her OpenAI’s legal team had okayed skipping a safety review for one of its models — a statement, she said, that turned out to be false.In closing arguments, Musk attorney Steven Molo hammered home the long list of people who had testified under oath that Altman was, in one way or another, a liar — all of whom Altman had worked with for years. “The defendants absolutely need you to believe Sam Altman,” Molo told the jury. “If you cannot trust him, if you don’t believe him, they cannot win. It’s that simple.”But during court proceedings, Musk — who now leads competing lab xAI, under his space company SpaceX — didn’t come off any better. Joshua Achiam, now OpenAI’s chief futurist, testified that Musk’s race against Google led him to take an “obviously unsafe and reckless” approach to achieving AGI. When he and others raised concerns, he says, Musk argued that OpenAI’s for-profit makeover created incentives to disregard safety, but his own xAI is for-profit and has, at best, ahaphazard approach to safety. And in the name of making sure OpenAI remained open, Musk was obsessive in his need for control over it. In closing arguments, Sarah Eddy, one of OpenAI’s attorneys, told the jury that Musk “wanted dominion over AGI.”As one X userput it, “if untrustworthyness had mass, putting Musk and Altman too close to one another would collapse the courtroom and all of earth into a black hole.”OpenAI did not immediately respond to a request for comment. On X, Muskposteda statement saying he’d be filing an appeal.It’s not just Musk and Altman, either.Trial evidence suggestedMurati helped get Altman removed, then switched sides to support his reinstatement while

GamingNewsPC GamingPlayStation exclusives aren’t coming to PC anymoreA big pivot back to consoles.A big pivot back to consoles.byJay PetersMay 18, 2026, 7:28 PM UTCLinkShareGiftImage: Sony Interactive EntertainmentJay Petersis a senior reporter covering technology, gaming, and more. He joined The Verge in 2019 after nearly two years at Techmeme.Sony reportedly won’t release its major single-player PlayStation games on PC anymore. According toBloomberg’s Jason Schreier, Hermen Hulst, who heads up PlayStation’s studios business,informed employeesin a town hall on Monday about the change in strategy. Schreier had previously reported on the shiftin March, saying that Sony scrapped plans to launch PC versions of last year’sGhost of Yōteiand “other internally developed games.” Online games will still come to multiple platforms following this change in strategy, Schreier reported at the time.In recent years, Sony has released many of its biggest games on PC, includingSpider-Man 2,Ghost of Tsushima, bothThe Last of Usgames,Horizon Zero Dawn Remastered, and multiplayer titles likeHelldivers 2andMarathon. Two years ago, Hulstcommitted toreleasing PlayStation’s live-service games “day and date” on PC and PS5, but its single-player PC releases have been less consistent, with Hulst saying that the company takes a “more strategic approach.”RelatedConsole exclusives might be making a comebackSony may not be the only one to go back to exclusives; Asha Sharma, Microsoft’s new Xbox chief,told staffthat she is “reevaluating” exclusive games for the platform.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Jay PetersGamingNewsPC GamingPlayStationMost PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TransportationNewsCarsVolvo teases a new affordable EV to replace discontinued EX30The new EV won’t be the same price as the EX30, which was supposed to start at $35,000, but ended up being much more than that thanks to tariffs.The new EV won’t be the same price as the EX30, which was supposed to start at $35,000, but ended up being much more than that thanks to tariffs.byAndrew J. HawkinsMay 18, 2026, 8:41 PM UTCLinkShareGiftImage: VolvoPart OfThe great EV pullback: all the obstacles, cancellations, and delayssee all updatesAndrew J. Hawkinsis transportation editor with 10+ years of experience who covers EVs, public transportation, and aviation. His work has appeared in The New York Daily News and City & State.Volvo’s compact, quirky EX30 had a lot of problems when it was first released. Tariffs essentially erased its affordability, making it more expensive to own, and a battery recall made it dangerous to park indoors. Butits discontinuationdidn’t spell the end of Volvo’s efforts to sell more affordable electric models. In fact, the Swedish automaker is already at work on a new offering for the US market.The news of an affordable Volvo EV for the US came during a media roundtable this week related tothe US launch of the new EX60. Luis Rezende, president of Volvo Cars America, said that the decision to discontinue the EX30 was not solely about tariffs and profitability, noting that the company is preparing to introduce a new EV in 2027 that will occupy a similar role in the lineup — though not necessarily at exactly the same price point as the EX30.“Very similar, I would say,” Rezende said about the mystery EV’s price comparison to the EX30. “It’s going to be an EV that will deliver a lot of good things in a bigger space, but it will be also fun to drive, I can promise you.”Other than that, details were scarce. Volvo’s executives talked later about the desire to build a larger, family-oriented SUV at its factory in Charleston, South Carolina — though that vehicle will likely use a “multi-fuel” strategy rather than being exclusively electric from launch.The EX60, which will start customer deliveries in the US this summer, is Volvo’s attempt at a reset in the US. The compact SUV, which is built on a different architecture than the EX30, will start at $59,795 for the entry-level P6 Plus version, and climbs up to $68,745 for the more powerful P10 AWD Ultra variant.Volvo is the latest automaker to try, and stumble, in its efforts to build an affordable EV for the US market that is both desirable and profitable for the company. To date, few have pulled it off, as it requires a certain level of scale, vertical integration, and mastery of the supply chain that only companies in China seem to have really nailed down. Of course, Volvo is owned by China’s Geely, but the company’s desire to sell EVs in North America will necessitate a different approach to affordability.Affordability was one of the EX30’s main selling points. When it wasfirst announcedin 2023, Volvo said the price would start at $34,950, positioning it as the smaller, less expensive EV that many people were clamoring for. But after the election of Donald Trump,Volvo was forced to delay the EX30’s arrival in the USuntil 2025, citingnewly leveled tariffsagainst vehicles built in China. Eventually, the model that went on sale in the US started at $44,900, about $10,000 more than the original price.Then, in February, further bad news as Volvo issued a recall for the EX30 because the vehicles’ batteries were at risk of overheating or catching on fire. The next month, Volvo pulled the plug on the vehicle in the US.Correction May 18th:A previous version of this story stated that the EX60 is the only Volvo EV in the US. The EX90 is also available.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Andrew J. HawkinsCarsElectric CarsNewsTransportationVolvoMore in:The great EV pullback: all the obstacles, cancellations, and delaysHonda’s hybrid future starts with new Accord and RDX prototypesStevie BonifieldMay 14Mazda delays its first EV by two years.Andrew J. HawkinsMay 12Ford’s EV and software chief Doug Field is leaving the companyAndrew J. HawkinsApr 15Most PopularMost PopularMicrosoft is retiring Teams’ Together ModeBad, Bad, Bad, Bad RoommateDisco Elysium’s spiritual successor can’t escape its phantomsElon Musk lost his case against Sam AltmanUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

When Lior Susan startedEclipse Venturesin 2015, the firm’s thesis of digitizing the physical world wasn’t particularly popular in Silicon Valley.“It was the era of enterprise software and SaaS, and it felt fairly lonely the first couple of years,” Susan said onstage at a recent StrictlyVC event in San Francisco.More than a decade later, Eclipse finds itself at the center of the tech world’s action. The firm’s $6.5 million Series A investment in Cerebras Systems in 2016 paved the way for a total return of $2.5 billion when the semiconductor company went public this week. The firm invested a total of $147 million in Cerebras over time, a bet that generated a seventeenfold return at theIPO price of $185per share, according to Eclipse.For Susan, the windfall from Cerebras is only the beginning of reaping big rewards from a long-standing belief that because 85% of global GDP is tied to the physical world, investing in companies beyond pure software could be immensely lucrative.Public markets and startup founders seem to be recognizing the value of physical-world tech now, too. Susan noted that shares of TSMC and Micron recently hit all-time highs, while a growing cohort of elite founders is eager to build startups at the intersection of hardware and software.“I think people understand that the real moat in software is gone. You can vibe-code pretty much whatever you want,” he said.Susan echoed public market sentiment that earlier this year sent many SaaS stocks tumbling on the belief that enterprises may use Anthropic’s Claude Code or OpenAI’s latest models to create their own bespoke software tools instead.“What you cannot do with ‘vibe code’ is manufacture wafers, because you need machines and silicon, and they need clean rooms, and a bunch of other things,” Susan said.When it comes to the tech that touches the physical world, it’s not just semiconductors that are suddenly catching the attention of investors and founders.Eclipse’s portfolio companies spanning sectors like robotics, energy and defense, raised nearly $15 billion from outside backers last year, and that late-stage momentum reached $4.5 billion in Q1 2026 alone, Susan said. That investor excitement stands in stark contrast to the firm’s early track record: In its first eight years, its portfolio companies raised less than $4 billion in total.Indeed, the recent follow-on rounds across Eclipse’s portfolio show a track record that any venture firm would envy. Driven by a string of massive late-stage deals this year, the haul includes$1.2 billion for Wayve, $650 million for True Anomaly, $270 million for Bedrock Robotics, and $200 million for Oxide Computer. What’s more, Eclipse was the Series A investor for all four companies.At first glance, it may seem that investor enthusiasm for physical-world tech is driven purely by AI, whether as an infrastructure input like chips and data centers, or through AI’s power to finally make robotics viable. However, Susan argues that there are other powerful tailwinds driving the momentum.Besides technology — in this case, AI — what’s important for this market to thrive is capital, customer demand, talent, and policy. Susan means that along with investors and engineers moving away from SaaS to sectors like robotics, semiconductors, space, and mining, the U.S. government is also encouraging these industries through subsidies and favorable regulation.“This is the first time I believe in America ever, from Henry Ford and Carnegie, those five forces are aligned,” Susan said. “For builders like us, this is the best time to build those companies.”TopicsCerebras,Eclipse Ventures,Fundraising,Hardware,Lior Susan,Space,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. To get this in your inbox, sign up here for free — just clickTechCrunch Mobility!There is a bit of a theme emerging in transportation — and really every industry: AI is creating jobs for some at the loss of others.General Motors, for instance,laid off more than 10% of its IT department, or about 600 salaried employees — in a deliberate skills swap. This won’t translate into a one-to-one exchange, which means there will likely be a net-negative job loss. But GM insists it is hiring and those layoffs have made room for it to recruit IT people with AI-focused backgrounds.The most sought-after capabilities are AI-native development, data engineering and analytics, cloud-based engineering, agent and model development, prompt engineering, and new AI workflows. In practical terms, GM is looking for people who know how to build with AI from the ground up — designing the systems, training the models, and engineering the pipelines — not just use AI as a productivity tool.Those AI job losses are mounting in the automotive sector. CNBCcalculatedthat Ford, GM, and Stellantis have cut a combined total of more than 20,000 U.S. salaried jobs, or 19% of their combined workforces, from recent employment peaks this decade. While there are a variety of reasons for these cuts, they are generally connected to technological changes, including AI.Companies are leaning heavily into AI, although anecdotes from some engineers and founders suggests not all of these businesses know quite what they’re doing with it yet.Samsarais one company that seems to have figured out a revenue-generating use case. The company has spent the last decade giving its customers cameras to mount inside millions of trucks for driver monitoring, theft prevention, and helping with liability claims. The company took that mountain of data and trained its own model that candetect potholesand determine how quickly they are deteriorating. The company is pitching this product to cities and announced it has several under contract, including Chicago.A little birdImage Credits:Bryce DurbinNothing this week, although I am working on a fun one! Reach out anytime with insights, tips, or just because. You can reach us via email or Signal.Email Kirsten Korosec [email protected] my Signal at kkorosec.07, or email Sean O’Kane [email protected]!Image Credits:Bryce DurbinYou might have noticed that Rivian’s spinoff companyMind Roboticsraised another $400 million, just two months afterraising $500 million. And that pace got me thinking about its founder RJ Scaringe and his innate ability to get VC and institutional backers to invest in his ideas and projects.I calculated that investors have poured $12.3 billion into Scaringe’s three startups — Also, Mind Robotics, and Rivian. That figure doesn’t include the close to $12 billion in gross proceeds raised in Rivian’s IPO, nor did I count the more recent strategic deals with Volkswagen Group and Uber — which together could add nearly $7 billion to Rivian’s coffers.You can read my whole riff on the topic here. But if you don’t feel like clicking, here is one item that stood out. I spoke to a number of insiders and investors and they all mentioned Scaringe’s ability to give undivided attention to whoever he’s talking to — whether it’s an investor, supplier, or exec — and make them feel like the most important person in the room.It’s yet another piece of evidence in my long-standing case against multitasking. Debate me!Other deals that got my attention …Arkeus, an Australian startup that developed perception software for autonomous drones and aircraft,raised $18 millionin a Series A round led by QIC Ventures. Other investors include R+VC, Folklore Ventures, DYNE Ventures, Main Sequence Ventures, Salus Ventures, and Beaten Zone.Aseon Labs, a Redwood City, California, startup that has developed a depot in a box for charging, cleaning, and inspecting autonomous fleets,came out of stealthwith undisclosed backing by Y Combinator.Rapidoraised $240 millionin a round led by Prosus, and that values the Indian ride-hailing company at $3 billion. Existing investors, including WestBridge Capital and Accel, participated. The round was part of a larger $730 million primary and secondary financing.Quantum Systems, a Germany-based drone startup backed by Peter Thiel, is in talks to raise around €600 million ($703 million) with companies like Airbus and Blackstone as investors,Bloomberg reported.Notable reads and other tidbitsImage Credits:Bryce DurbinIs Redwood Materials ready for an IPO?Senior reporter Sean O’Kane interviewed the company’s new CFO,Deepak Ahuja, whose name will be familiar to anyone who follows Tesla. Ahuja was Tesla’s former finance chief and most recently held a similar position at drone company Zipline.TeslaRobotaxis have crashed at least twice since July 2025 while ateleoperator was remote

Commencement season has come around again — and this year, a couple speakers have discovered that it’s tough to get graduating students excited about a future shaped by artificial intelligence.Last week, Gloria Caulfield, an executive at real estate firm Tavistock Development Company,gave a speech at the University of Central Floridaacknowledging that we’re living in a time of “profound change,” which can be both “exciting” and “daunting.”“The rise of artificial intelligence is the next industrial revolution,” Caulfield declared — prompting the students in the audience to begin booing, getting louder and louder until Caulfield chuckled, turned to the other speakers, and asked, “What happened?”“Okay, I struck a chord,” she said. Caulfield then tried to resume her speech, saying, “Only a few years ago, AI was not a factor in our lives” — only to be interrupted again by the audience, this time by their loud cheers and applause.Former Google CEO Eric Schmidt faced a similar response when he brought up AI at a University of Arizona speech on Friday.In Schmidt’s case, the criticism actually began before the speech itself, with some student groupscalling for him to be removed as commencement speakerdue toa lawsuit in which a former girlfriend and business partner accused Schmidt of sexual assault. (He has denied the allegations.) According toa local news report,the booing began even before Schmidt took the stage.But Schmidt alsogot loud booswhen he told students, “You will help shape artificial intelligence.” The booing was persistent enough that Schmidt tried to speak over it, insisting, “You can now assemble a team of AI agents to help you with the parts that you could never accomplish on your own. When someone offers you a seat on the rocket ship, you do not ask which seat, you just get on.”To be fair, AI isn’t becoming a third rail ateverygraduation ceremony. Nvidia CEO Jensen Huang recentlyspoke atCarnegie Mellon’s commencement, and he didn’t seem to get any audible pushback when he said that AI has “reinvented computing.”Still, it’s not exactly surprising to find some students in a booing mood. Ina recent Gallup poll, only 43% of Americans aged 15 to 34 said it’s a good time to find a job locally, a steep drop from 75% in 2022.That pessimism isn’t solely a response to the rise of AI (a shift thateven some software engineers are worried about), but journalist and tech industry criticBrian Merchant suggestedthat for many students, AI has become “the cruel new face of hyper-scaling capitalism.”“I too would loudly boo at the prospect of this next industrial revolution if I was in my early twenties, unemployed, and had aspirations for my future greater than entering prompts into an LLM,” Merchant wrote.Even when graduation speeches didn’t mention AI explicitly, “resilience”was a recurring theme this year. Schmidt himselfacknowledgedthat there is “a fear in your generation that the future has already been written, that the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create.”Caulfield, meanwhile, might also have misread her audience of arts and humanities graduates. One student said that before mentioning AI, Caulfield already started to lose them with her “generic” praise of corporate executives like Jeff Bezos.Another graduate, Alexander Rose Tyson,told The New York Times, “It wasn’t one person that really started the booing. It was just sort of like a collective, ‘This sucks.’”TopicsAI,commencement speeches,Eric Schmidt,gloria caulfieldWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Anthony HaAnthony Ha is TechCrunch’s weekend editor. Previously, he worked as a tech reporter at Adweek, a senior editor at VentureBeat, a local government reporter at the Hollister Free Lance, and vice president of content at a VC firm. He lives in New York City.You can contact or verify outreach from Anthony by [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

Listen onApple PodcastsListen onSpotifyLawyers for Elon Musk and OpenAI made their closing arguments this week, and now it’s up to jurors to decide whether OpenAI did anything wrong as it’s transformed intoa slightly-more-for-profitorganization.But as Kirsten Korosec, Sean O’Kane, and I noted on the latest episode of TechCrunch’sEquity podcast, a big theme in the trial’s final days waswhether OpenAI CEO Sam Altman is trustworthy— for example, Musk’s attorney Steve Molo grilled Altman about whether statements he’d made during congressional testimony were truthful.Kirsten noted that Musk hasmade plenty of misleading statements of his ownand that trust isn’t just an issue for Altman.“This is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs,” she said. “It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies; there’s a lot behind the veil still.”Keep reading for a preview of our conversation, edited for length and clarity.Anthony Ha:[The end of the trial] led to this really provocative headline from one of our writers, Tim Fernholz, [that] just says,“Who trusts Sam Altman?”Does anyone want to take a stab at answering this?Kirsten Korosec:Yeah, Anthony, I’m going to throw it right back to you. Do you trust Sam Altman?Anthony:It’s an interesting question because it feels like something that’s kind of a wild question to discuss in a journalistic context, but actually that’s the core of the trial, in a lot of ways.Sean O’Kane:That’s not a yes.Anthony:And it actually seems to be [at the] core of understanding so much of what’s happened at OpenAI, especially this big executive power struggle that they now call The Blip.It just seems like a lot of people who’ve worked with Altmandon’t trust him. And he’s acknowledged this a little bit, because he’ll talk about the fact thathe recognizes he’s been conflict averse, telling people what they want to hear, and he’s trying to work on that.I mean, it sounds plausible, and I can understand how that can lead to misunderstandings in some situations. [But] I’m also a very conflict-averse person and I’d like to think that if any of this stuff went to trial, that people would not be asking, “Is Anthony Ha trustworthy?”Sean:Still not a yes!Kirsten:I think that people would say that you are trustworthy. I will say that question, while provocative, doesn’t just encapsulate what this trial was about. I would zoom out even more and say this is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs. It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies; there’s a lot behind the veil still.Maybe when they all IPO, we can get a peek, but it is fundamentally about trust and misuse, and do we believe the intent? And what I would throw back is, sometimes the intent can be worthy, noble, and still misused. It can still end up as a bit of a s— show. I think it’s more than who trusts Sam Altman — although that was very interesting in this trial — but more of that bigger question that we can apply to the entire industry.Sean:I’ll say it: I don’t trust him. But you know, I don’t trust most people, so I guess that’s just the baseline.We’ll see where this goes. The trial wraps up today. I’ve been very curious to hear how the jury decides this all. I think at the start of this, a big motivator of this was Elon Musk trying to sling mud at a perceived rival and someone who he feels slighted him. And I don’t know if we know enough yet to say that that was completely accomplished, and whether or not he has a shot at winning. But I think all these people came out of this looking a little bit worse.Anthony:And just to get specific, why this is coming up this week is that [Altman] was on the stand and he was basically getting grilled about some statements he’s made in the past, in testimony to [Congress], basically saying he didn’t have any equity in OpenAI. And that is not true because he had a stake through Y Combinator, which he used to run. And tried to brush that off by saying, “I assume that everybody understands what it means to be a passive investor in a VC fund.” And I think [Elon Musk’s] lawyer, somewhat fairly, said “Really? You think the congressman who was interviewing you knew that?”Kirsten:Yeah, I mean, he was playing the whole semantics game. What I thought was so interesting about [this] is the style of how Sam Altman answered questions [compared to] Elon Musk on the stand.So Elon Musk, in many, many, many scenarios and many instances, we can point to the fact that he put something out on Twitter that was a lie or a bit of a fib, and on the stand corrected the record. So there’s a history of, I would say, non-truthfulness-slash-lying, blatant or otherwise, in Elon Musk’s world, but how he treated it was incredibly combative an

In BriefPosted:1:15 PM PDT · May 17, 2026Image Credits:Artur Widak/NurPhoto / Getty ImagesAnthony HaApple’s Siri revamp could include auto-deleting chatsPrivacy will be a major theme when Apple unveils a new version of Siri at the Worldwide Developers Conference in June,according to Bloomberg’s Mark Gurman.The Siri relaunch is widely seen as Apple’s big chance to reestablish its relevance in artificial intelligence. As part of that effort, company executives will argue that they’re taking a more privacy-friendly approach than most other AI companies, Gurman said.Apple will reportedly launch the first stand-alone Siri app,powered by Google Geminiand offering users a chatbot experience reminiscent of ChatGPT.  But compared to those other chatbots, the app is supposed to have more limitations on how long user information can be used and stored.For example, Gurman said Siri could include a feature similar to the Messages app, allowing users to automatically delete conversations after 30 days or one year — or to keep them indefinitely.Gurman also suggested that Apple might be emphasizing privacy as a way to excuse Siri’s shortcomings compared to competing products — and that this emphasis might obscure the fact that Google is handling some of the security.TopicsAI,Apple,Apps,gemini,In Brief,siriMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedSpaceOSHA probing worker death at SpaceX’s Starbase siteSean O’Kane2 hours agoAISandboxAQ brings its drug discovery models to Claude — no PhD in computing requiredLucas Ropek3 hours agoAIAnthropic has acquired the dev tools startup used by OpenAI, Google, and CloudflareKirsten Korosec5 hours agoLatest in AIAISandboxAQ brings its drug discovery models to Claude — no PhD in computing requiredLucas Ropek3 hours agoAIAnthropic has acquired the dev tools startup used by OpenAI, Google, and CloudflareKirsten Korosec5 hours agoAIElon Musk has lost his lawsuit against Sam Altman and OpenAITim Fernholz6 hours ago

Imagine you’re riding a motorcycle at 160 kilometers per hour when an arrow appears, floating on the road ahead, telling you exactly where to turn. No phone, no dashboard. Just your helmet, and a lens the size of a thumbnail.This is not a concept video. It’s heading to European roads as early as this year. And it’s one early glimpse of where smart glasses are heading.Over the past few years, Big Tech has been quietly (and not so quietly) placing its bets. Meta has been selling AI-enabledRay-Ban glasses since 2023, Google isbuilding Android XR, andApple is expected to enter the market. Last week, Samsung wasreportedlyset to unveil its first AI-capable smart glasses, co-designed with Gentle Monster, at a Galaxy Unpacked event in London this July. China’sHuawei,Alibaba,Xiaomi, and others are all moving too.The numbers reflect the momentum. Global AI glasses shipments surged to 8.7 million units in 2025, up more than 300% from the prior year, and analysts project that figure will cross 15 million this year,per Omdia.Suppliers and component makers of AI-powered smart glasses are also positioning themselves for what comes next. One of the companies, a South Korean startup calledLetinAR, has spent the last decade building the optical technology that could make all of this actually wearable.The LG Electronics-backed startup just secured $18.5 million from Korea Development Bank and the South Korean retail giant’s venture arm, Lotte Ventures, among others, ahead of its planned 2027 IPO in South Korea.Its previous investor,LG Electronics, has since begun developing its own AI smart glasses, according to alocal mediareport, which is a sign of how seriously South Korea’s largest consumer electronics company takes the category.LetinAR CEO Jaehyeok Kim and CTO Jeonghun Ha, who have been friends since high school, founded the company together in 2016.Image Credits:LetinARThe lens that makes it wearableLetinAR doesn’t make the glasses. It makes the part that makes the glasses work. The optical module, the tiny lens component that projects images into your field of vision, is what determines whether a pair of smart glasses feels like a sci-fi headset or something you’d actually wear to work, Ha told TechCrunch. It has to be light, thin, and power-efficient, while still delivering a sharp, clear image. Getting all of that right in a single component, small enough to fit inside a normal-looking frame, is the central engineering challenge of the entire industry. That’s what LetinAR is building.“We see AI glasses as that next platform,” Kim said. “And the optical module is the hardest part to get right as AI glasses makers will need a lens that is thinner, lighter, and more power-efficient than what exists today.”The co-founders said LetinAR wants to be the company those glasses makers call. The company calls its technology PinTILT: a way of arranging tiny optical elements inside a lens so that light is directed precisely where it needs to go, into the user’s eye, rather than scattered in every direction.Think of a TV. It broadcasts light across an entire room, but only the light that actually reaches your eyes matters. Most existing smart lens technologies, particularly a dominant approach calledwaveguide, work a bit like that TV, splitting and spreading light across the full lens to create a wide image. The result is a thin lens, but an inefficient one. A lot of light gets thrown away before it ever reaches the eye, which means dimmer images and, critically, a battery that drains fast, Ha explained.The alternative, a mirror-based approach known asbirdbath, delivers light more directly to the eye, but the structure is bulky, making it nearly impossible to fit inside something that looks like a normal pair of glasses.PinTILT sidesteps that trade-off, Ha said. By focusing only on the light that can actually enter the eye and carefully engineering the angle of each tiny element inside the lens, LetinAR claims it can produce a brighter image in a thinner, lighter form factor, using less power. In a category where every gram and every hour of battery life matters, that’s the problem the entire industry has been trying to solve.In the space, there are a number of peers likeWaveOptics,DigiLens, andLumus.CustomersIts modules are already shipping. LetinAR counts Japan’s NTT QONOQ Devices and Dynabook, formerly known as Toshiba Client Solutions, among its customers, giving the company real manufacturing experience at scale. It is in talks with Big Tech companies on R&D of next-generation AI glasses, though it declined to name them.One of LetinAR’s most demanding customers isAegis Rider, a Swiss deep tech company spun out of ETH Zurich’s Computer Vision Lab. Aegis Rider is building an AI-powered AR helmet that displays navigation, speed, and safety alerts directly in a motorcycle rider’s field of vision, not floating on the visor but anchored to the road itself, as if the information is physically painted on the world ahead.LetinAR’s mod

Grafana Labs, the maker of its eponymous popular open source web visualization software, confirmed it had been hacked but that it refused to pay the hackers who had threatened to release the company’s codebase.Ina series of postson social media, the lab said its investigation found that the hackers had abused a stolen token credential that allowed access to the company’s GitHub environment, which it uses for storing its source code, but the token did not allow access to customer records or financial data. The company has since invalidated the token and added additional security measures to prevent a repeat incident.“The attacker attempted to blackmail us, demanding payment to prevent the release of our codebase,” the company said.Grafana’s code is open source and public, meaning anyone can download the software and edit its code before running it on their own machines. It’s unclear if the hackers stole any proprietary code or information. A spokesperson for the company did not immediately return a request for comment.The incident contrasts with the recent hack at education tech giant Instructure, which last week“reached an agreement” to pay the hackerswho had compromised its network twice in recent weeks. The hackers had demanded an unspecified ransom, threatening to release stolen data about staff and students who use its softwarefollowing a massive data breach and a subsequent website defacement.While in Grafana’s case, no customer data was taken, the company cited the FBI’s long-standing advice urging victims not to pay hackers, as cooperating with them does not guarantee they will return stolen data or refrain from publishing it later. Critics also say paying cybercriminals helps to fund future cyberattacks.Grafana said its investigation was ongoing and will share its findings once its probe concludes.This story was updated to correct that the hackers compromised access to Grafana’s GitHub environment.Topicscyberattack,cybersecurity,data breach,extortion,grafana,open source,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

Amazonannouncedthe latest update to Alexa+ on Monday: the ability to generate podcast episodes on demand.The new feature, called “Alexa Podcasts,” is rolling out to customers in the U.S. today. Amazon describes the capability as a way to “turn any topic you’re curious about into a podcast episode, ready in minutes.”To use the feature, all users have to do is ask Alexa+ to create a podcast about a topic they’re interested in. Users don’t need to upload documents, write scripts, or plan anything ahead of time. Instead, Alexa+ researches the request, gathers information, and generates a quick overview of what the episode will cover. From there, users can tweak things like the length, tone, and focus of the episode.Once finalized, Alexa+ uses AI-generated host voices to narrate the podcast. When the episode is ready, users get a notification through their Echo Show device and inside the Alexa app. Episodes are also saved in the app’s “Music” and “More” sections so they can be replayed later.The feature is another example of how Amazon is trying to turn Alexa+ into more than just a voice assistant. Instead of only answering questions or controlling smart home devices, Alexa+ is starting to act more like a personalized AI content creator.At the same time, the launch is likely to spark some debate. AI-generated voices and automated content continue to raise questions around ethics, accuracy, and the future of traditional creators. There are also concerns about how reliable AI-generated podcasts will be, especially when covering news or complex topics.Amazon emphasized its partnerships with major news organizations to improve content accuracy and reliability. The company says Alexa+ can access real-time information through agreements with outlets, including the Associated Press, Reuters, The Washington Post, Time, Forbes, Business Insider, Politico, USA Today, Condé Nast, Hearst, and Vox Media, alongside more than 200 local newspapers across the U.S.Beyond podcasts, Amazon says it is exploring additional forms of personalized AI audio, including custom news briefings and content generated from users’ own documents and shared information.TopicsAI,alexa plus,alexa+,Amazon,Generative AI,Media & Entertainment,podcastsWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

The market for AI notetaking devices has exploded in the U.S., with the category generating over $600 million in revenue last year, according toa Menlo Ventures report. And as startups likeHeidi HealthandFreedhave shown, there’s decent demand for this tech in healthcare, where doctors and clinics see the potential for an AI assistant that can help them keep track of patient conversations, surface health records, and lower their administrative burdens.But those apps don’t do much for patients, which is whyKin Healthis building a notetaker that can transcribe your visits to doctors, parse medical advice, and surface next steps when required. To that end, the startup has raised $9 million in a seed funding round led by Maveron.The app is similar to a meeting notetaker: You can record doctor visits, and it will return an AI summary of the meeting, with the next steps, all of which you share with family and friends if you want to. It also lets you write down questions that you might want to ask during your next visit.Loading the player…Kin Health says it encrypts all patient data and that summaries are kept private by default. The tool is not HIPAA-certified, as it is a patient-facing one, but it adheres to the same privacy standards, the company said.The free app is built by physicians Arpan and Amit Parikh, along with Kyle Alwyn, who previously built online prescription serviceHeyDoctor and sold it to health platform GoodRx. Doug Hirsch and Trevor Bezdek, co-founders of GoodRx, are founding partners and executive chairmen at the company.Co-Founders Arpan Parikh, Amit Parikh, and Kyle Alwyn.Image Credits:Kin Health“We have a lot of these storage cabinets where our health data can live, but we don’t have a way to convert that into a utility that we can use to drive our behavioral change. Our goal is to create this health graph where we can store your information from multiple different sources,” Alwyn told TechCrunch over a call.Kin Health says that its summaries are provided after a few stages of processing. After transcribing the visit, an algorithm turns the transcription into a clinical narrative, which gets crunched into a user-facing summary with action items. The company says it is leaning on specialized medical models to power the transcription and that it evaluates and observes outputs at different stages to ensure answers are accurate.But AI in healthcare is being received with a measure of caution and apprehension. Privacy experts and researchers have raisedconcernsover data security, accuracy of AI, consent mechanisms, thequality of generated notes, and theireffectiveness.AI notetakers also oftenfail to recognize and struggle to transcribe regional accents. Kin Health says it is working to ensure its tool works with different accents, as well as when someone has a bad throat or is wearing a mask.Dr. Rebecca Mishuris, chief health information officer and VP at Mass General Brigham, a healthcare organization in Boston, argues it is important for physicians to review any notes generated by AI.“Generative AI will hallucinate; that is the nature of a technology built on patterns and prediction. That is why it is so important for clinicians to review the drafted notes before signing them. At the end of the day, the responsibility for the documentation falls to the clinician,” she told TechCrunch over email.Kin Health currently only shows notes from conversations it records during consultations, but the company said it plans to bring in data from other health sources, including physicians’ own notes through electronic health record (EHR) systems, this year.The company says it will keep the app free of cost forever, and monetize via referrals to services such as specialists and labs. The startup is taking a leaf from GoodRx’s playbook, which also keeps the core product free and earns commissions by referring other services.Natalie Dillon, a partner at Maveron, said healthcare provider-side tools often expect patients to coordinate their own treatment actions. “Kin is built to solve an entirely different consumer need: it can travel with them between specialists, systems, and providers. It’s not beholden to any single health network or EHR relationship. It’s built to serve the patient, not the institution, and that’s a massive distribution advantage,” she said.The funding round also saw participation from Town Hall Ventures, Eniac Ventures, Flex Capital, Foundry Square Capital, Pear VC, and The Family Fund. GoodRx’s Hirsch and Bezdek; angel investors Jay Desai, Nabeel Quryshi, Alex Cohen, and Saharsh Patel; and more than 30 physicians also invested.Topicsai scribe,Apps,Health tech,Kin Health,maveronWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Ivan MehtaIvan covers global consumer tech developments at TechCrunch. He is based out of India and has previously worked at publications including Huffington Post and The Next Web.You can

New York public health provider NYC Health + Hospitals saysa months-long data breachthat allowed hackers to steal personal data, medical records, and fingerprints scans affects at least 1.8 million people.NYCHHC is the largest public health system in the United States and provides healthcare toover a million New Yorkers, the majority of whom are uninsured or receive state healthcare benefits, such as Medicaid.The healthcare system reported the number to the U.S. Department of Health and Human Services, making it one of the largest healthcare-related data breaches of the year so far. Healthcare organizations have been repeatedly targeted by financially motivated cybercriminals in recent years in efforts to steal their vast banks of highly sensitive patients’ personal, medical, and billing information.In a data breach notice on its website, NYCHHC said that it detected a cyberattack on February 2 and secured its network. The hackers had access to its network from November 2025 until February 2026, during which the hackers copied files from its systems.The healthcare system said hackers broke due to a breach at a third-party vendor, which it did not name.NYCHHC said that the exposed data varies by individual and includes patients’ health insurance plan and policy information, medical information (e.g., diagnoses, medications, tests, and imagery), billing, claims, and payment information. Other government-issued identity documents, such as Social Security numbers, passports, and driver’s licenses, were also compromised.The breach notice also says “precise geolocation data” was taken in the breach, suggesting that the user-uploaded photos of their identity documents may have also contained the exact location of where the document was captured.The breach is particularly sensitive because hackers stole biometric information, including fingerprints and palm prints, which affected individuals have for life and cannot replace. NYCHHC did not provide an explanation for storing biometric data. Prospective NYCHHC employees are generally required to enroll their fingerprints for criminal records checks. It’s not yet known if patients’ biometrics were also taken.NYCHHC’s website was briefly offline as of Monday morning. A spokesperson for NYCHHC did not immediately respond to an email from TechCrunch with questions about the cyberattack. TechCrunch asked, among other things, why it took the organization months to detect the breach, and if it has received any communication from the hackers, such as a demand for payment.It’s not clear if NYCHHC can receive email at the time of the website outage.The incident appears to be unrelated to the data breach at National Association on Drug Abuse Problems (NADAP)earlier this year, in which over 5,000 NYCHHC patients had information taken in the cyberattack.In the FBI’s latestannual report on cybercrimecovering 2025, healthcare remained a top target for ransomware attackers — criminals who break into databases, steal a copy of the data while scrambling the victim’s servers, and threaten to publish the stolen data if the victim does not pay the hackers. A ransomware attack on UnitedHealth-owned health tech giant Change Healthcare allowed Russian-linked hackers tosteal the medical and billing information of more than 190 million Americans, believed to be the largest theft of U.S. medical data in history.Topicscyberattack,cybersecurity,data breach,healthcare,new york city,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

Elon Musk’s claim that he was mistreated by his OpenAI co-founders failed after nine California jurors returned a unanimous verdict that his lawsuits had been filed too late.Musk accused Sam Altman, Greg Brockman, OpenAI, and Microsoft of “stealing a charity” by creating a for-profit affiliate of the frontier AI lab. Jurors, however, found that any harms that Musk may have suffered came before the deadline for filing his claims under the law.While the trial delved deeply into the melodramatic history of OpenAI and featured testimony from leading figures in Silicon Valley, it ultimately turned onfairly narrow questionsof the law. The trial focused on whether and when Altman and the other defendants had made and broken promises to Musk, but his case failed to convince jurors that he had a valid claim.In particular, OpenAI had advanceda statute of limitations defense, which sought to prove that any harms Musk sought to litigate had taken place before 2021. (The specific date varied by the charge: before August 5, 2021, for the first count; August 5, 2022, for the second count; and November 14, 2021, for the third count.) Ultimately, the jury found that argument persuasive, which made for a short deliberation period.“There was a substantial amount of evidence to support the jury’s finding, which is why I was prepared to dismiss on the spot,” Judge Yvonne Gonzalez Rogers said after the verdict was delivered.The end of the case means that one major threat to OpenAI — a possible restructuring — is now off the table ahead of its reported IPO.“It did not take [the jury] two hours to conclude … that Mr. Musk’s lawsuit is nothing more than an after-the-fact contrivance that bears no relationship to reality,” OpenAI’s lead attorney, Bill Savitt, said after the verdict. “They kicked it exactly where it belongs — just to the side. This lawsuit is a hypocritical attempt to sabotage a competitor.”Microsoft, which Musk sued for aiding and abetting OpenAI’s alleged breach of charitable trust, welcomed the verdict. A spokesperson for the company said it “remained committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”The verdict came in the middle of a hearing to determine the potential damages to Musk if the verdict had gone the other way. While that discussion is moot for now, the judge appeared unconvinced by the analogy Musk’s lawyers drew between his charitable contributions and investments in a for-profit startup.“Your analysis seems to be devoid of connection to the underlying facts,” she told Dr. C. Paul Wazzan, the expert who came up with Musk’s estimate of OpenAI and Microsoft’s wrongful gains at his expense — some $78.8 billion to $135 billion.In a tweet after the ruling, Musk appeared to take the procedural grounds of the dismissal as a moral victory. “There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!” Musk wrote. “I will be filing an appeal with the Ninth Circuit, because creating a precedent to loot charities is incredibly destructive to charitable giving in America.”Reached for comment by TechCrunch, Musk’s lead counsel, Marc Toberoff, said, “One word: Appeal.”TopicsAI,Elon Musk,greg brockman,Microsoft,OpenAI trial,sam altmanWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Tim FernholzSenior ReporterTim Fernholz is a journalist who writes about technology, finance and public policy. He has closely covered the rise of the private space industry and is the author ofRocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race.Formerly, he was a senior reporter at Quartz, the global business news site, for more than a decade, and began his career as a political reporter in Washington, D.C.

Anthropic announced Monday it has acquired Stainless, a startup founded by former Stripe engineer Alex Rattray whose software is widely used by rival AI labs, including OpenAI and Google.Anthropic didn’t disclose terms of the deal. However, The Informationreportedlast week that the company was in talks to acquire Stainless, which is backed by Sequoia Capital and Andreessen Horowitz, for more than $300 million.The acquisition will take a key infrastructure supplier out of the hands of Anthropic’s competitors. The company told TechCrunch it will wind downall hosted Stainless products, including its SDK generator. An Anthropic spokesperson said Stainless customers will still own the SDKs they’ve generated to date and have full rights to modify and extend them however they wish.The New York-based startup, founded in 2022, rose to prominence in the emerging AI industry for automating the creation and maintenance of software development kits, or SDKs — the libraries developers use to interact with APIs.Rattraydeveloped softwarethat could take API specifications and turn them into production-ready SDKs across multiple programming languages, including Python, TypeScript, Kotlin, Go, and Java. It became a popular tool because the platform automatically updates the SDKs as APIs change and eliminated the time-consuming process of manually maintaining them.The technology is particularly valuable tocompanieslike Anthropic, OpenAI, Google, Replicate, Runway, and Cloudflare that are building AI agents that can connect to external software and complete tasks on behalf of users. Stainless’s SDK tools are an easy way to build and maintain those connections — but going forward, the tools will only be available to Anthropic, not its competitors.According to Anthropic, Stainless software has powered the generation of every official Anthropic SDK since the earliest days of its API.“I started Stainless because SDKs deserve as much care as the APIs they wrap,” Rattray saidin a press releaseposted Monday. “Anthropic was one of the first teams to bet on this with us. We have been watching what developers have built on Claude over the last few years, which made bringing our teams together an easy decision. The team gets to keep doing the work we love, on the platform where it matters most.”TopicsAI,Anthropic,StainlessWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Kirsten KorosecTransportation EditorKirsten Korosec is a reporter and editor who has covered the future of transportation from EVs and autonomous vehicles to urban air mobility and in-car tech for more than a decade. She is currently the transportation editor at TechCrunch and co-host of TechCrunch’s Equity podcast. She is also co-founder and co-host of the podcast, “The Autonocast.” She previously wrote for Fortune, The Verge, Bloomberg, MIT Technology Review and CBS Interactive.You can contact or verify outreach from Kirsten by [email protected] via encrypted message at kkorosec.07 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

A worker died at SpaceX’s Starbase launch site in South Texas on Friday, and the Occupational Safety and Health Administration (OSHA) has opened an investigation.TheSan Antonio Express-News reported Mondaythat the unidentified victim died at around 4:17 a.m. local time on May 15, citing OSHA and local officials. The Wall Street Journal later reported that the county sheriffconfirmed to the outletthat a worker died. OSHA confirmed to TechCrunch that it is investigating the apparent accident.Representatives for the nearby Brownsville police and fire departments did not respond to requests for comment. SpaceX and the newly incorporated city of Starbase did not respond to requests for comment.The circumstances of the worker’s death are not immediately clear. OSHA told TechCrunch that it won’t release more information until its investigation is complete, which could take months.The death comes just a few days ahead of the first planned launch of SpaceX’s upgraded Starship rocket. Elon Musk’s spaceflight company is also reportedly releasing the detailed prospectus for its initial public offering this week, which is expected to be the biggest ever when that transaction takes place next month.SpaceX has long dealt with worker safety problems at its Starbase site, which handles Starship prototype launches and is an active construction zone.In 2025, TechCrunchanalyzed OSHA dataand determined the Texas launch site had an injury rate that far outpaced those of industry rivals and was the most dangerous of SpaceX’s worksites. A 2023Reuters investigationuncovered dozens of previously unreported injuries and a worker death in 2014 at SpaceX’s McGregor, Texas, test site.In January, OSHA hit SpaceX with seven “serious” safety violations for, among other things, not properly inspecting a crane before it collapsed at Starbase last June. The safety agency dealt SpaceX the maximum financial penalty on six of those seven violations, totaling $115,850. SpaceX is contesting those penalties, federal records show.The company has been hit with multiple lawsuits related to injuries sustained at Starbase in recent years. In December, an employee of one of SpaceX’s subcontractors sued after he was crushed by a large metal support dropped from a crane. The worker, Eduardo Cavazos, suffered a broken hip, knee, and tibia, and OSHA opened a “rapid response investigation,” as TechCrunchfirst reportedin December.OSHA has since closed that rapid response investigation without taking any punitive action, according to a TechCrunch public records request. And the lawsuit was recently dropped because his employer, the subcontractor, provides workers compensation insurance that prevents the company from being sued, according to Cavazos’ attorney.TopicsOSHA,Space,SpaceX,starbase,TransportationWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Sean O’KaneSr. Reporter, TransportationSean O’Kane is a reporter who has spent a decade covering the rapidly-evolving business and technology of the transportation industry, including Tesla and the many startups chasing Elon Musk. Most recently, he was a reporter at Bloomberg News where he helped break stories about some of the most notorious EV SPAC flops. He previously worked at The Verge, where he also covered consumer technology, hosted many short- and long-form videos, performed product and editorial photography, and once nearly passed out in a Red Bull Air Race plane.You can contact or verify outreach from Sean by [email protected] via encrypted message at okane.01 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

Drug discovery is one of the most expensive failures in modern industry. Finding a single viable molecule can take a decade and cost billions, and most candidates still don’t make it. A generation of AI startups has promised to fix that — most have made the problem less painful for researchers, who are already technically sophisticated enough to use the tools.But SandboxAQ thinks the bottleneck isn’t the models. It’s the interface.The company has teamed up with Anthropic to integrate its scientific AI models directly into Claude — putting powerful drug discovery and materials science tools behind a conversational interface that requires no specialized computing infrastructure to use.Founded roughly five years ago as an Alphabet spinout, SandboxAQ counts Eric Schmidt, Google’s former CEO, as its chairman. The company, which has raisedmore than $950 million from investors, has built out a number of different business lines, includinga cybersecurity business.One of the more unique things SandboxAQ does, however, is produce large quantitative models, or LQMs. These proprietary models are “physics-grounded,” meaning they’re built on the rules of the physical world rather than patterns in text. They can run quantum chemistry calculations and simulate both molecular dynamics and microkinetics, the study of how chemical reactions unfold at the molecular level. That matters because it tells researchers how candidate molecules are likely to behave before anyone sets foot in a lab.“Trained on real-world lab data and scientific equations, LQMs are AI models engineered for the quantitative economy, a $50+ trillion sector spanning biopharma, financial services, energy, and advanced materials,” the company said in a news release that strongly suggests Sandbox AQ isn’t building another chatbot or code assistant — it’s chasing the economy that AI is supposed to transform.Chai DiscoveryandIsomorphic Labs— both well-funded bets on better models — have focused on the science. SandboxAQ is focused on who can actually use it.“For the first time, we have a frontier [quantitative] model on a frontier LLM that someone can access in natural language,” Nadia Harhen, SandboxAQ’s general manager of AI simulation, told TechCrunch. Previously, users of SandboxAQ’s LQMs would have had to provide their own digital infrastructure to run the models.SandboxAQ’s customers tend to be computational scientists, research scientists, or experimentalists. Generally, these people work at large pharmaceutical or industrial companies and are searching for new materials that can become marketable products.“Our customers come to us because they’ve tried all the other software out there, and the complexity of their problem is such that it didn’t work or didn’t yield positive results for them when that translation went to take place in the real world,” said Harhen.TopicsAI,Anthropic,Claude,Eric Schmidt,Google,SandboxAQWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lucas RopekSenior Writer, TechCrunchLucas is a senior writer at TechCrunch, where he covers artificial intelligence, consumer tech, and startups. He previously covered AI and cybersecurity at Gizmodo.

Bad beatsThe US is betting on AI to catch insider trading in prediction marketsThe Commodity Futures Trading Commission wants us to know it’s taking this very seriously.Kate Knibbs, wired.com–May 16, 2026 7:00 am|112Credit:Yuki Shimazu / FlickrCredit:Yuki Shimazu / FlickrText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navFor most of the past year, it looked likeprediction marketshad kicked off a new golden age of fraud. OnPolymarket, traders raked in fortunes fromsuspiciously timed betson geopolitical events like the raid on Venezuela and the Iran War. It wasn’t clear whether the US government would bother pursuing some of the most flagrant bad actors, since Polymarket’s crypto-based platform was technically offshore and not regulated or licensed within the country.Now, however, the Commodity Futures Trading Commission, which oversees prediction markets, wants you to know that it’s watching very, very closely. The agency is searching for suspicious behavior from traders within the United States who have been sneaking onto offshore markets, including Polymarket’s crypto platform—which is blocked stateside—by using virtual private networks. “We’re going to find them, and we’re going to bring actions,” agency chairman Michael Selig told WIRED this week, speaking from the CFTC’s headquarters in Washington, DC.Selig says the agency, which is especially lean right now, is staffing up. Like so many other AI-pilled workplaces, the CFTC is also leaning into automation to handle the growing workload, including tools that analyze trading patterns and flag potential manipulation. “You’ve got so much data,” Selig says. “When we feed it into AI, we get really great information. It can help us understand things, like where we might want to investigate, or when we might need to send a subpoena to a trader.”In addition to proprietary surveillance systems developed in-house, the agency’s arsenal includes third-partyblockchain tracing toolslike Chainalysis for crypto platforms, and market abuse detection software including Nasdaq Smarts for centralized markets. (Beyond Nasdaq Smarts, the agency did not specify which AI tools it uses and declined to share more specific examples.)Prominent prediction market companies have recently started touting all the work they’re doing to catch sketchy bettors. US-based exchange Kalshi, Polymarket’s primary competitor, eagerly announced that it hassuspendedand penalized customersflagged for insider tradingand market manipulation.In April, after significant backlash over suspected insider trading, Polymarket announced its own partnership with Chainalysis. It was part of a broader push to crack down on market manipulation. While the company’s CEO, Shayne Coplan,had talkedin the past about why insider trading could be good for prediction markets, Polymarket changed its approach this spring, updating its market integrity rules and announcing a partnership with Palantir for its US-based sports markets (the Chainalysis deal focuses on the offshore platform). The company did not respond to WIRED’s request for comments for this story.According to Chainalysis spokesperson Maddie Kenney, the company analyzes the same data for both clients. “The value Chainalysis adds for our customers, including Polymarket and the CFTC, is organizing the data and enriching it with the attributions and insights we’ve accumulated over years in the space,” she says. Certainly sounds like a good deal for Chainalysis!The CFTC’s assurances that it is hunting insiders comes at a moment of intense scrutiny on prediction markets. In March, Connecticut senator Chris MurphytoldWIRED that he suspected White House staffers were engaged in insider trading on war-related contracts. At the beginning of April, seven members of Congressaskedthe CFTC to investigate overseas markets offering war-themed events contracts. In a letter, the lawmakers argued that the commission had the authority and responsibility to curb insider trading, especially on “morally obscene” trades on military action. Selig recentlytoldCongress that the company is pursuing “hundreds, if not thousands” of insider trading tips.Investigations are not limited to federally regulated exchanges. “We’re surveilling the markets on a global basis,” he tells WIRED.Selig says that the agency will exert extraterritorial jurisdiction—its legal ability to enforce its laws beyond traditional boundaries—when it finds suspicious activity on offshore platforms like Polymarket, though he says it’s a case-by-case approach. “We use it in extreme circumstances,” he says, with an eye towards whether charges have a strong chance of sticking in court. “In any extraterritorial litigation, there’s going to be challenges to our authority, and that could also impair our ability to bring cases in the future.” According to Selig, the 2010 Dodd-Frank Act allows the CFTC more leeway to pursue this kind of enforcemen

striking resultsA revolutionary cancer treatment could transform autoimmune diseaseResearchers are testing CAR T cell therapy as a way to reset the immune system.Amber Dance, Knowable Magazine–May 17, 2026 7:00 am|66Artist’s rendering of a T cell.Credit:Getty ImagesArtist’s rendering of a T cell.Credit:Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAt age 49, Jan Janisch-Hanzlik’s multiple sclerosis was destroying her freedom to live the life she wanted. She gave up her active nursing job for a desk role. Frequent falls made her afraid to carry her grandchildren. She had to move to a bigger house to make room for the wheelchair she feared she might end up needing full-time.Even the best available medication wasn’t improving Janisch-Hanzlik’s symptoms, and she worried they’d only get worse. So when she learned about a trial of CAR T cell therapy at the University of Nebraska Medical Center in Omaha, close to the city of Blair where she lives, she phoned the clinic every other month until they were ready toenroll her as the first patient.Originally designed to target and wipe out cancer byreprogramming the patient’s immune cells, CAR T is now being offered to patientsin hundreds of clinical trialsfor autoimmune conditions like multiple sclerosis, lupus, Graves’ disease, vasculitis, and many others. The hope is that CAR T can duplicate the success it has demonstrated in a range of blood cancers by hunting down and eliminating cells that target the self in autoimmune diseases. This would essentially reset the body’s defenses to a state like the one that existed before the disease took hold.But along with CAR T’s promise come risks, questions, and challenges. There’s uncertainty about how well it will work forautoimmunityand how long any benefits might last, as well as what long-term side effects might arise. Janisch-Hanzlik knew this when she sat down to receive the experimental treatment on June 9, 2025; she felt a mix of hope and fear knowing that she would be spending the next week being monitored for side effects including dangerous inflammation.In addition to her clinical expertise and desire to pioneer a new treatment, Janisch-Hanzlik’s two young grandchildren helped inspire her pursuit of a treatment with known risks and uncertain benefits. Because multiple sclerosis has a genetic component, Janisch-Hanzlik knew that they have an elevated chance of going through the same struggle she has. “I would want to be able to say I did everything that I possibly could to prevent them, or anyone else, from having something like this,” she says.From cancer to autoimmunityThe first CAR T cancer treatment wasapproved by the Food and Drug Administration in 2017for an aggressive form of leukemia. Since then, the powerful and intensive treatment hasresulted in long-term remissionfor many cancer patients.The basic premise of CAR T is to activate the power of key immune cells calledT cells. T cells normallyrecognize other cellsthat have been infected by a virus or bacterium, or are otherwise abnormal, and either destroy them or recruit other parts of the immune system to do so.In CAR T for cancer, scientists engineer those T cells to specifically hunt and destroy malignant cells. The technologygot its startwhen cancer researchers figured out how to take out a patient’s own T cells, insert DNA instructions for a “chimeric antigen receptor,” or CAR, and put them back into the person’s circulation. The CAR, which sits on the T cell’s surface and latches on to a specific molecular partner on the surface of cancerous cells, activates the T cell to attack.Today CAR T cells are most commonly programmed to attack B cells, another key immune player. B cells are normally responsiblefor making antibodies, but in certain blood cancers, they proliferate out of control. By giving T cells a CAR that recognizes one of a couple of molecules unique to the B cell surface, the cells are reprogrammed to find and eliminate those cancerous cells.B cells also are the central problem in many autoimmune conditions: They mistakenly make antibodies against normal tissues instead of against invading pathogens. So as CAR T began to succeed against B cell cancers, it didn’t take long for doctors to reason that CAR T therapy might also be able to wipe out bad B cells in people with autoimmunity.There are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineThere are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineA German team pioneered autoimmune CAR T in a woman with lupus,reporting positive resultsin 2021.

All roads lead to StarshipThe US space enterprise is desperately waiting for Starship—will it finally deliver?“This is such a wild ride. The highs are high. The lows are low.”Eric Berger–May 18, 2026 7:00 am|143Starship V3 completes a successful wet dress rehearsal on May 11.Credit:

      SpaceXStarship V3 completes a successful wet dress rehearsal on May 11.Credit:

      SpaceXText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThese days, one would be forgiven for forgetting that SpaceX is, at its core, a rocket company.Consider the company’s mega deals over the last year. SpaceX paid $17 billion—more than it has spent developing every one of its rockets—to EchoStar for wireless spectrum to boost its Starlink network. It revealed plans to launch 1 million orbital data centers. SpaceX merged with xAI in a deal that valued Elon Musk’s artificial intelligence firm at $250 billion, and it announced plans to becomea major computer chip manufacturer. And earlier this month, SpaceX sold an enormous amount of ground-based compute to Anthropic.As a result of all this activity, an impending IPO will value the company at something like $1.5 or $2 trillion. That’s trillion, with at.So yes, one might reasonably ask what SpaceX does these days. Because all the buzz, all the Wall Street euphoria, and all the financial frisson are only tangentially related to what SpaceX cut its teeth on during its first 25 years: becoming the globally dominant player in launch. It largely concerns telecommunications and AI data services.And yet everything SpaceX aspires to accomplish in the next quarter of a century, all of its enormous valuation, is predicated on a new launch vehicle. A rocket that, to date, has a decidedly mixed record of success. A rocket that has not flown in seven months. A rocket that, finally, may return to the skies on Wednesday.We are speaking, of course, of Starship—a truly revolutionary rocket. If it works. And after a long period of development and three years of test flights and setbacks, it kind of has to.“Test Like You Fly”A few weeks ago, SpaceX released avisually stunning videothat takes viewers inside its massive new Starfactory in South Texas and provides up-close views of its rockets and engines.The “Test Like You Fly” video also outlines the development of the third iteration of the Starship rocket, V3, highlighting both advances and setbacks encountered along the way. SpaceX is distilled to its core as the exceptional engineers who are designing and testing a radically new rocket talk about the very difficult problems they’re trying to solve.“We are not breaking laws of physics; we are just trying to leverage them as effectively as we can,” explains Jacob McKenzie, the company’s vice president for the Raptor rocket engine that powers both the first and second stage of Starship.McKenzie says SpaceX built 600 Raptor engines as part of its V2 Starship program, a remarkable number that underscores the investment the company is making to develop a rocket that has yet to reach orbit or deliver a single payload. (To put this into perspective,NASA is spending $3.5 billionto procure two dozen comparably powered, expendable rocket engines.)By some estimates, SpaceX has now invested $15 billion in the Starship rocket program over the last decade, funding 11 test flights, a sprawling spaceport in South Texas, a massive new factory there to support a high production rate, and expanding facilities in Florida alongside the company’s existing infrastructure in the state.It has not gone particularly smoothly. The company conducted its initial Starship test flight in April 2023. I’m not sure anyone expected that, three years and nearly a dozen flights later, the company would still be firmly in test mode.But it turns out that building the world’s largest and most powerful rocket and optimizing it for rapid reuse is a difficult, time-consuming task. Even for a company that prides itself on moving fast and trying to achieve the near impossible.A very difficult 2025SpaceX performed six test flights in 2023 and 2024, and by November of the latter year had made substantial progress. The company demonstrated the ability to launch and capture the Super Heavy first stage and completed a safe flight of the upper Starship stage, including re-lighting its Raptor engine in space, before making a controlled splashdown in a precise location.That month, the then-general manager of the Starbase facility in South Texas, Kathy Lueders, said the company aimed to dramatically increase the vehicle’s launch cadence in 2025. “Elon would say, next year, he would love to have us have 25 missions a year,” she saidduring a community event. The company also planned to “capture” a Starship upper stage and conduct an orbital refueling test in that time frame.Debris from Starship falls back into the atmosphere in this view over Hog Cay

swarmsBug bounty businesses bombarded with AI slop“Never-ending” AI slop strains corporate hacking reward schemes.Jamie John, Financial Times–May 18, 2026 9:23 am|77Credit:

      AdobeCredit:

      AdobeText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navCompanies that pay hackers to find flaws in their software are being inundated with low-quality reports generated by AI, forcing some to suspend the programs altogether.Businesses that run “bug bounty” schemes have long relied on independent security researchers to spot vulnerabilities. But the rise of AI tools is now overwhelming them with spurious submissions.Bugcrowd, whose customers include OpenAI, T-Mobile, and Motorola, said the number of reports it received more than quadrupled over a three-week period in March, with most proving to be false.Curl, a widely used tool to transfer data across the Internet, suspended its paid bug bounty program in January, citing an “explosion in AI slop reports” and lower-quality submissions.Cyber security experts say advances in generative AI are reshaping the economics of bug bounty programs. While the tools allow experienced researchers to find flaws more quickly, they are also lowering the barrier to entry, triggering a flood of automated or erroneous submissions that companies must sift through.The big increase in poor-quality AI reports was “quickly becoming a major problem,” said Ross McKerchar, chief information security officer at cyber security group Sophos. “Bug bounties are going to stay [but] they’re going to have to change,” he said.Bug bounties have grown in popularity since the early 2000s, with schemes offering six-figure payouts for the biggest discoveries. Google’s program disbursed a total of $17 million last year, up from $7.5 million in 2021. It paid its largest individual reward of $605,000 in 2022 to a user who spotted a vulnerability in its Android mobile operating system.McKerchar said the rise in poor-quality submissions came from both amateurs trying to find bugs for the first time and existing researchers who were “sometimes getting led on by the [AI] agents.”He added there was a “third cohort” of “experienced AI builders” who had developed automated “end-to-end scanning and submission systems” that were “creating absolute carnage.”Curl’s creator, Daniel Stenberg, wrote in a blog post that the “never-ending slop” had taken “a serious mental toll to manage and sometimes also a long time to debunk.”Software group Nextcloud suspended its bug bounty program in April because of the “massive increase of low-quality reports.” It said it hoped to resume the program once it had found a way to filter submissions effectively.The surge in AI-generated reports comes as Anthropic last month launched Mythos, its new cyber AI model, which it says can find software flaws faster than humans.Companies running bug bounty programs have started to introduce more stringent background checks to combat the problem, as well as building AI agents to triage submissions.HackerOne, whose bug-reporting platform serves Goldman Sachs, Google, and the US Department of Defense, said it had “introduced new agentic validation capabilities” this year to “help organizations manage high volumes of findings,” such as those generated by models like Mythos.The company said submissions had jumped 76 percent in the year to March. But it said the share of reports flagging legitimate vulnerabilities had remained steady over the past year at 25 percent.HackerOne chief executive Kara Sprague said it had in recent weeks seen a rise in “higher quality” reports that had used AI. She added that the rise in AI-generated submissions was “not a strong reason to say we don’t want them” altogether, given that hackers were using the technology to spot more flaws.Bugcrowd chief Dave Gerry said developments such as Anthropic’s Mythos would assist human bug bounty hunters, not replace them. “AI is going to help with a lot of things but we’re never going to replace that human creativity,” he said.© 2026 The Financial Times Ltd.All rights reserved. Not to be redistributed, copied, or modified in any way.Financial TimesFinancial Times77 Comments

The Next GenerationDid Artemis II break through? Registrations at Space Camp double afterward.Isaacman’s $25 million donation leads to impressive new facilities.Eric Berger–May 18, 2026 9:49 am|26Commercial Astronaut on Inspiration4 Chris Sembroski and NASA Administrator Jared Isaacman attend the ribbon-cutting ceremony of the Inspiration4 Skills Training Complex in Huntsville, Ala.Credit:

       Tibrina Hobson/Getty ImagesCommercial Astronaut on Inspiration4 Chris Sembroski and NASA Administrator Jared Isaacman attend the ribbon-cutting ceremony of the Inspiration4 Skills Training Complex in Huntsville, Ala.Credit:

       Tibrina Hobson/Getty ImagesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen he was 12 years old, NASA Administrator Jared Isaacman attended the weeklong “Aviation Challenge” program at Space Camp in Huntsville, Alabama.“For the first time, I got behind the controls of an airplane when I attended Aviation Challenge,” Isaacman said on Friday evening during an event at the US Space & Rocket Center. “I became a pilot because I thought that was the closest I would ever get to the stars.”Decades later, after founding a successful online payments company and flying to space twice as a private citizen aboard SpaceX’s Crew Dragon, Isaacman has returned to Space Camp in Alabama on multiple occasions to meet with participants and share a bit of the awe he experienced as a kid. In 2022, a year after the first of these flights, Inspiration4, Isaacman donated $10 million to kick off a Space Camp expansion.Now, as the leader of NASA’s space program, he has sought to continue to pay it forward. He donates his salary to Space Camp, and on Friday, he returned to open the new “Inspiration4 Skills Training Complex,” a 50,000-square-foot facility completed with an additional $15 million donation from Isaacman. The money will also support development of a new dormitory.Space Camp becomes part of the US cultureSpace Camp has carved out an influential place in American culture. A 1986 feature film,Space Camp, popularized the camp when four teenagers befriended a robot and ended up launching into orbit on a space shuttle. The plot is implausible, to say the least, and the movie was criticized because it came out just a few months after the space shuttleChallengercatastrophe.Nevertheless, the movie helped build the myth of Space Camp, which had only opened four years earlier. Since then, more than 900,000 children ages 9 to 18 years old have “graduated” from the experience. Included among them are at least half a dozen NASA astronauts, including Dottie Metcalf-Lindenburger, Kate Rubins, Serena Auñón-Chancellor, and a mission specialist on the recently flown Artemis II mission, Christina Koch.Some of the exhibits and activities at Space Camp had grown a bit dated, but Isaacman’s gift should help to revitalize the programs.Among the new elements are a parachute “simulation room” that features a virtual-reality 10,000-ft drop with wind simulation, a drone lab, an interactive mission control, and “training simulators” that will recreate the experience of zero gravity.Camp registrations double after Artemis IIThe Artemis mission, flown by Koch and three other astronauts, appears to have reignited interest in the space program. Space Camp officials say registrations have doubled this summer since the successful completion of the lunar flyby mission in April.Interest is expected to grow further in the coming years as NASA returns to the Moon and establishes habitats there. The aim is to meet the needs of space-curious children with a modernized experience.“Off the success of Artemis II, America’s return to the Moon is just getting going,” Isaacman said during the ribbon-cutting event for the new facility.“And it is that kind of magic that inspires the next generation to attend Space Camp, get hands-on experiences at this national treasure, unlike anywhere else in the country, and grow up ready to pick up the baton and join in this great adventure.”Eric BergerSenior Space EditorEric BergerSenior Space EditorEric Berger is the senior space editor at Ars Technica, covering everything from astronomy to private space to NASA policy, and author of two books:Liftoff, about the rise of SpaceX; andReentry, on the development of the Falcon 9 rocket and Dragon. A certified meteorologist, Eric lives in Houston.26 Comments

engagementBMW sends off the 6th-gen M3 CS with a manual gearbox, rear-wheel driveThe 2027 M3 CS Handschalter is lighter and comes with three pedals instead of two.Jonathan M. Gitlin–May 18, 2026 10:33 am|47For the first and last time, you can get a stick shift in a G80 BMW M3 CS.Credit:

      BMWFor the first and last time, you can get a stick shift in a G80 BMW M3 CS.Credit:

      BMWText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe march of time, and what counts for progress in the automotive industry, has not been particularly kind to the driving enthusiast. Our vehicles have gotten bigger and heavier. Touch-sensitive panels and screens replaced buttons. Steering feelevaporatedabout a decade ago. And if you’re a fan of changing your own gears with a stick shift and three pedals,things have been looking bleakfor a while now. Which makes BMW’s send-off for its current sixth-generation M3 so notable.BMW’s M divisionkept the six-speed manual alive for the G80 M3, butonly the normal version. If you wanted the more powerful, much torquier M3 Competition or the track-focused M3 CS (Competition Sport), the only transmission choice was an eight-speed automatic. That automatic happens to be the excellent ZF 8HP gearbox, and for being fast on track, I’d still choose it, because that makes left-foot braking easier.Using paddle shifts might be faster, but I won’t pretend it’s more engaging than coordinating the movement of a gearstick through its gate, timed properly to the action of the clutch—especially if you’re heel-and-toeing, but even if you use the auto-blip feature that revs the engines for you on downshifts now. BMW appears to recognize that too, because it says the 2027 M3 CS Handschalter is designed for maximum driver engagement, and just for North America.Yes, that is a big carbon-fiber hump in the middle of that seat.BMWYes, that is a big carbon-fiber hump in the middle of that seat.BMWNew, lighter wheels.BMWNew, lighter wheels.BMWImola red paint is a $4,500 option.BMWImola red paint is a $4,500 option.BMWNew, lighter wheels.BMWImola red paint is a $4,500 option.BMWThe $107,100 M3 doesn’t use the more powerful engine from the M3 Competition, but the same 473 hp (353 kW) inline six-cylinder engine as the regular six-speed M3. But it’s a lighter car, to the tune of about 75 lbs (34 kg) thanks to things like a titanium exhaust muffler, carbon-fiber seats, carbon ceramic brakes, lighter wheels, and plenty of carbon-fiber reinforced plastic for body panels. And instead of the ZF 8HP, there’s a six-speed manual, which sends its power to the rear wheels and the rear wheels only: There’s no xDrive AWD here, unlike the M3 Competition.BMW M’s next major work will be next year’sall-electric quad-motor M3, derived from theNeue Klasse i3. And we can’t imagine there’s a three-pedal version of that one in the works, sadly.The M3Ireally wantIn BMW’s briefing materials for the M3 CS Handschalter, it notes that the 3.0 L S58 engine is also used, in modified form, in the BMW M4 GT3 Evo that won the Rolex 24 at Daytona this year. BMW has had a fair degree of success with the M4 GT3 Evo, as it also won last year’s Nurbürgring 24, among other races. But the only M race car with an S58 engine anyone cared about at this year’s N24 started life as anApril Fool’s joke.As 24-hour races go, the N24 is certainly unique. Le Mans and Daytona have faster prototypes among the GTs. Spa-Francorchamps is just GT3, but more than 70 of them on track together. But the N24 combines everything from GT3 cars down to Volkswagen Golfs, and does so across not just the modern F1-grade GP Circuit but also the entire stretch of the Nordschleife, with its narrow straights, crests, drops, and dozens of blind corners. Watching driver onboards is frankly terrifying, given the speed differentials.The BMW M3 Touring at the N24.Christian Kaspar-Bartke/Getty ImagesThe BMW M3 Touring at the N24.Christian Kaspar-Bartke/Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesI do wonder if the car handled a bit differently thanks to a higher, farther-back center of gravity.Silas Stein/picture alliance via Getty ImagesThe M3 Touring looked amazing in its race livery.Silas Stein/picture alliance via Getty ImagesThis year’s race had a pair of stars. Max Verstappen took time off from his F1 job, got his Nürburgring race permit, and then showed that he could absolutely cut it with the world’s best sports car drivers on the w

about timeFive years later, Windows 11 brings back much-missed taskbar options (and more)Microsoft is also testing a smaller taskbar and more customizable Start menu.Andrew Cunningham–May 18, 2026 12:03 pm|114Windows 11’s work-in-progress side-mounted Taskbar.Credit:

      MicrosoftWindows 11's work-in-progress side-mounted Taskbar.Credit:

      MicrosoftText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen Windows 11 launched in 2021, we mostly liked its refreshed look—the rounded corners and menus with just a hint of translucency were a nice change from the flat colors and hard corners of the Windows 8/10 era. But its reformulated taskbar and Start menu came with a number of functional regressions from the versions in Windows 10. Some of these were addressed quickly; others continue to linger.A new Windows Insider Preview build released to testers includes a new wave of improvements that fix longstanding regressions while trying out new things.Most significantly, the Windows 11 taskbar can now be docked to any edge of your screen, including the left and right, something that was possible in Windows 10 (and many older versions of Windows) but has been missing from Windows 11 since launch. Users can configure slightly different taskbar behavior for every taskbar position—if you prefer a different icon alignment or a left/right-mounted taskbar over a top/bottom-mounted taskbar, or if you want different settings for labels and icon groupings, you can choose different options for each position and Windows will remember them.Microsoft says there are several features that haven’t been implemented yet—the taskbar won’t auto-hide in any of the alternate positions, and the “tablet-optimized taskbar” with larger, more finger-friendly icon sizes and spacing also isn’t supported. Touch gestures and the Search box also aren’t supported. All of these features are coming at some point; they just aren’t ready now. Microsoft is “evaluating additional features like different taskbar positions per monitor” for multi-monitor setups.Another change Microsoft is testing makes the taskbar and all its icons smaller, a change designed to increase the amount of vertical space available on smaller screens without requiring users to fully hide the taskbar.The Start menu with recommended and All Apps sections disabled, along with the shorter taskbar.Credit:
                                  MicrosoftThe Start menu with recommended and All Apps sections disabled, along with the shorter taskbar.Credit:

      MicrosoftMicrosoft is also making a handful of changes to the Start menu, including a user-selectable size setting (previously, the menu would increase and decrease in size dynamically based on the size of your display). Each section of the Start menu (the pinned apps section, the “recommended” section, and the “all apps” section) will be individually toggleable. And users will be allowed to hide “recommended” apps advertised from the Microsoft store while still being able to see jump lists and recent files in the File Explorer. Users who decide to keep the “recommended” apps section visible should also benefit from “improv[ed] file relevancy” that “better reflect[s] what you have been working on.”Some of these changes are available in the current Windows Insider Preview builds in the Experimental channel (which replaced the Canary and Dev channels in Microsoft’slatest beta program shake-up). Others will be released “over the coming weeks,” at which point more polished and refined versions will presumably come to the Beta channel and, eventually, the public version of Windows 11. All of these changes are being made as part of Microsoft’s continued “commitment to Windows quality” push, meant to address several of the real and perceived shortcomings of Windows 11 relative to older versions.Andrew CunninghamSenior Technology ReporterAndrew CunninghamSenior Technology ReporterAndrew is a Senior Technology Reporter at Ars Technica, with a focus on consumer tech including computer hardware and in-depth reviews of operating systems like Windows and macOS. Andrew lives in Philadelphia and co-hosts a weekly book podcast calledOverdue.114 Comments

Hands-onThe Dory Sign is E ink, smart screen simplicity at its finestDory CEO claims the $149 signs won’t be bricked should Dory go out of business.Scharon Harding–May 18, 2026 1:26 pm|58I customized the Dory Sign to introduce my dog. The polka dot frame is optional, removable, and costs $29 extra.Credit:

      Scharon HardingI customized the Dory Sign to introduce my dog. The polka dot frame is optional, removable, and costs $29 extra.Credit:

      Scharon HardingText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navMany gadgets marketed as being “smart” make me wonder if they would be better off dumb.Some examples aresmart TVsthat insist on sending your activities to businesses totrack you, smart fridges that use the Internetto cycle through ads,smart gym equipment that won’t work offline, smart toothbrushes whosebatteries drain too quickly, or virtually any gadget that forces you to use a minimally effective or otherwise unimpressive app.Too often, modern technologies, like inter-device connectivity and artificial intelligence, are shoehorned into gadgets that would be more intuitive to use, affordable, accessible, and/or durable without them.Dory Sign, however, is a reminder of how technology can improve something as simple as a sign without overshadowing the product’s most basic purpose, which in this case is effective and delightful communication.Dory is a small sign that, like many E Ink displays, is easy on the eyes because it doesn’t use bright lighting. You control what the sign displays through a free iOS or Android app that doesn’t force you to share your email or name to use it. The app has clear sections for editing the header text, main text, and footer text, adding an image, and choosing a background. It also allows more than one person to make changes to the display and communicates to the sign through Bluetooth.A screenshot of the Dory Sign app. Changes made to text, the image, and background appear in the preview image in real-time.Credit:
                                  Scharon Harding/DoryA screenshot of the Dory Sign app. Changes made to text, the image, and background appear in the preview image in real-time.Credit:

      Scharon Harding/DoryYou can upload your own image or background, but the app comes with a decent range of preloaded images, like animals, flowers, and illustrations, and preloaded backgrounds, including ones that look like brush strokes or marble and more detailed ones, like a landscape. You don’t have to spend a lot of time designing the sign, but if you want to get creative, the app has six different typefaces for text and sliders for text size, line height, letter spacing, and text color (which makes text darker or lighter).The sign can only display one image, though. And while transferring your preferred text, image, and background is simple enough with the app, your phone has to be pretty close to the sign for it to work. You also have to press the power button every time you want to change the display. This can be tedious, especially if you have the sign mounted to a wall. To be fair, though, manually changing a regular sign also requires you to get close to and touch the display.The E Ink screen doesn’t lend itself to vibrant imagery. Instead, it’s a more passive display that doesn’t call for attention, yet clearly communicates with distinct detail in its text and images.The “House party” background by ANTI that’s included in the app.Scharon Harding/ANTIThe “House party” background by ANTI that’s included in the app.Scharon Harding/ANTIThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathThe “Marble Dark” background by UK artist Nigel Hoare with the “Couples Dinner” image by Norweigan artist Frederik Skavian. Both are included in the Dory Sign app.Scharon Harding/Nigel Hoare/Frederik SkavianThe “Peak in Gimmelwald, CH” background by American artist Chris Herath that’s included in the app, plus some text.Scharon Harding/Chris HerathDory Sign’s makers claim gadget can last for 15 yearsAt $149, the Dory Sign is expensive for a small display. However, you may be able to extract that much value from it over the years. Dory claims that the sign only uses power when it’s changing what’s on the display.

“are you afraid?”Guy Gardner makes a cameo in newLanternsteaser“This ring? It’s the greatest weapon in the Universe. When and if to use it, that’s the whole ball game.”Jennifer Ouellette–May 18, 2026 1:36 pm|6Credit:

      YouTube/HBO MaxCredit:

      YouTube/HBO MaxText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navLanterns, the new DC Universe series coming to HBO Max, dropped asurprising teaser in Marchthat swapped the usual superhero hijinks for gritty realism more in the vein ofTrue DetectivesandSlow Horses. Personally, I liked the change; it pushed a series that was barely on my radar to my 2026 list of ones to watch. I guess HBO was concerned people would miss the superhero vibes, though, because the latest teaser trailer interweaves the grittier aspects with a lot more superpowers and intergalactic elements. And you know what? I’m still here for it.Per the official logline, “The series follows new recruit John Stewart (Aaron Pierre) and Lantern legend Hal Jordan (Kyle Chandler), two intergalactic cops drawn into a dark, earth-based mystery as they investigate a murder in the American heartland.” There will be two storylines: one set in 2016 about a murder in Nebraska, and the second set in 2026.Chandler’s Hal Jordan is a former test pilot nearing retirement from the Green Lantern Corps. He’s training Pierre’s John Stewart Jr., a new recruit, to replace him. Nathan Fillion reprises hisSupermanrole as the obnoxious Guy Gardner; we get a brief glimpse of him in the new teaser. The cast also includes Kelly MacDonald as Kerry, a small-town family-oriented sheriff; Jason Ritter as Billy Macon, Kerry’s husband; Garret Dillahunt as William Macon, Kerry’s cowboy father-in-law; Poorna Jagannathan as a woman named Zoe; Ulrich Thomsen as Sinestro, a former Corps member who’s gone rogue; and Paul Ben-Victor as an extraterrestrial called Antaan.Sherman Augustus plays John Stewart Sr., with J. Alphonse Nicholson playing the younger version; Nicole Ari Parker plays Bernadette Stewart (mother to John Jr.), with Jasmine Cephas Jones playing the young version of the character. In addition, Chris Coy plays a suspiciously nervous truck driver, Waylon Sanders; Cary Christopher plays a gifted child named Noah; and Laura Linney and Paula Patton will appear in as-yet-undisclosed guest roles.We also know that the show will feature the inter-dimensional prison and the Checkmate organization introduced inPeacemaker’s S2 finale. DC Studios co-CEO James Gunnhas saidthat the events will set up other things within the DCU as a whole. So, is it a gritty buddy-cop series or a superhero series? Showrunner Chris Mundyrecently told Men’s Healththat it’s very much both.“Our show is in a lot of ways about replacement—when should someone step aside and when is it time for the next person to take the reins?”Mundy said. “That push and pull between those two characters is really important. So much of the power that John has is by not taking the bait, understanding that you lose your power if you’re yelling and screaming. That’s what we’re trying to convey: He knows he belongs, so he doesn’t have to overcompensate. There’s a real balance there that’s just innately inside of Aaron. He’s big. He’s an intimidating presence just physically. But there’s a softness to him, too. There’s a thoughtfulness. You can’t teach that.”Lanternspremieres on August 16, 2026, on HBO Max.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.6 Comments

from the garden of fugitivesPompeii victim ID’d as a likely doctorNew X-rays and CT scans showed small case with locking mechanism containing metal instruments.Jennifer Ouellette–May 18, 2026 2:03 pm|9Credit:

      Pompeii Archaeological ParkCredit:

      Pompeii Archaeological ParkText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navArchaeologists used a combination of advanced CT scans and 3D digital reconstruction to identify one of the Pompeii victims who died in 79 CE during the eruption of Mount Vesuvius as most likely having been a Roman doctor, according toan announcementby the Pompeii Archaeological Park.Aspreviously reported, the eruption of Mount Vesuvius released thermal energy roughly equivalent to 100,000 times the atomic bombs dropped on Hiroshima and Nagasaki at the end of World War II, spewing molten rock, pumice, and hot ash over the cities of Pompeii andHerculaneumin particular. The vast majority of people in Pompeii and Herculaneum—the cities hardest hit—perished from asphyxiation, choking on the thick clouds of noxious gas and ash. But at least some of the Vesuvian victimsprobably died instantaneouslyfrom the intense heat of fast-moving lava flows, with temperatures high enough to boil brains and explode skulls.In the 19th century, an archaeologist named Giuseppe Fiorelli figured out how to make casts of those frozen bodies by pouring liquid plaster into the voids where the soft tissue had been. Some 1,000 bodies have been discovered in the ruins, and 104 plaster casts have been preserved.Restoration effortson 86 of those casts began about 10 years ago, during which researchers took CT scans and X-rays to determine whether complete skeletons were present.The CT scans and X-ray images revealed that there had been a great deal of manipulation of the casts, depending on the aesthetics of the era in which they were made, including altering some features of the bodies’ shapes or adding metal rods to stabilize the cast, as well as frequently removing bones before casting. Ancient DNA analysisin 2024of four victims found in what is known as the “House of the golden bracelet”revealed thatall four bodies were male and none were genetically related—challenging the pre-existing preferred narratives and suggesting those may reflect certain cultural biases.The small case linked to the Roman physician found inside the Pompeii cast.Credit:
                                  Pompeii Archaeological ParkThe small case linked to the Roman physician found inside the Pompeii cast.Credit:

      Pompeii Archaeological ParkThis latest discovery involves one of the plaster cast remains of 14 victims found in what is now known as theGarden of the Fugitives, who died when they were overtaken by the pyroclastic flow while trying to flee via the Nocera Gate. Pompeii superintendent Medeo Maiuri excavated the victims in 1961, originally found in three smaller groups, although today the plaster casts are displayed in a line.Maiuri published a fanciful, largely fictional account of who he believed those bodies to be in National Geographic that same year: a merchant with severe osteoarthritis; a mother and her two young children; a young couple with an infant daughter; and a servant with what appeared to be a shoulder bag, although the “bag” later turned out to just be a malformation in the plaster.For decades, archaeologists had overlooked a small object trapped within another of the plaster casts that had been kept in storage for decades at the archaeological park. The new analysis with X-ray imaging and CT scans revealed that it was a small case with parts made out of metal—specifically, a toothed wheel operating a locking system. The case contained a coin-filled fabric pouch and several metal instruments similar to what a Roman physician (medicus) might have used, leading the team to conclude the man had likely been a doctor, fleeing Pompeii with his instruments close at hand.Jennifer OuelletteSenior WriterJennifer OuelletteSenior WriterJennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.9 Comments

A little too lateElon Musk took too long to sue OpenAI, jury unanimously agreesMusk plans to appeal after judge immediately affirmed the jury’s decision.Ashley Belanger–May 18, 2026 2:09 pm|140Elon Musk attends a state banquet hosted by Chinese President Xi Jinping in Beijing during a summit with Donald Trump.Credit:Alex Wong / Staff | Getty Images NewsElon Musk attends a state banquet hosted by Chinese President Xi Jinping in Beijing during a summit with Donald Trump.Credit:Alex Wong / Staff | Getty Images NewsText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navElon Musk took too long to file his lawsuit that accused OpenAI of stealing a charity, a nine-person jury unanimously decided Monday.Musksued OpenAI in 2024for making a “fool” out of him after Musk donated $38 million to kick-start OpenAI as a nonprofit, only to later be blindsided when OpenAI created a for-profit arm that he felt gutted funding for the charity while enriching executives like Sam Altman and Greg Brockman.But the jury found that Musk was aware of OpenAI’s restructuring plans as early as 2021 and therefore missed the statute of limitations requiring him to bring the lawsuit within three years, The New York Timesreported. Because Musk took too long to file the litigation, the jury deemed Altman and Brockman not liable for any of the claims that Musk brought against OpenAI, the NYT reported. The jury also let Microsoft off the hook, finding no liability for the OpenAI investor after Musk alleged they aided OpenAI’s get-rich scheme.Judge Yvonne Gonzalez Rogers almost instantly agreed with the jury’s decision, the NYT reported.She thanked the jury and reminded them that they can talk to “anyone about anything” once they feel comfortable discussing the case, the NYT reported.Musk’s legal team looked “sunken in their chairs” after the jury’s decision was announced, the NYT reported. One of the billionaire’s lawyers, Marc Toberoff, who hosted daily calls with press during the trial, quickly confirmed that Musk will appeal.Altman and Brockman were not present when the verdict came in, the NYT reported. However, there were “a few smiles from the OpenAI side of the room,” with OpenAI attorney William Savitt wearing “a wide grin,” the NYT reported. Savitt spent daysgrilling Musk on the stand, seemingly taking advantage of his prior experience as Musk’s lawyer to get under Musk’s skin.In a statement, Microsoft celebrated the win, writing,“The facts and the timeline in this case have long been clear, and we welcome the jury’s decision to dismiss these claims as untimely. We remain committed to our work with OpenAI to advance and scale AI for people and organizations around the world.”Musk did not return to the courtroom after delivering his testimony and was also not there when the verdict was read. Before the verdict arrived, Musk’s lawyer apologized after Musk shirked a court order requiring him to be available in case any further testimony was needed in the case. In an apparent sign that he was losing interest in the trial, Musk prioritized attendinga summit between Donald Trump and China’s President Xi Jinping.Late Monday, Musk issued a statement on X, writing, “Regarding the OpenAI case, the judge & jury never actually ruled on the merits of the case, just on a calendar technicality. There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!”Musk reiterated his lawyer’s statement confirming he plans to appeal, writing, “I will be filing an appeal with the Ninth Circuit, because creating a precedent to loot charities is incredibly destructive to charitable giving in America.”This story was updated to include Musk’s statement.Ashley BelangerSenior Policy ReporterAshley BelangerSenior Policy ReporterAshley is a senior policy reporter for Ars Technica, dedicated to tracking social impacts of emerging policies and new technologies. She is a Chicago-based journalist with 20 years of experience.140 Comments

Can I pet that dog?Australian Aboriginals cared for a dingo’s grave for decadesFor some ancient Aboriginal Australian communities, dingoes were part of the family.Kiona N. Smith–May 18, 2026 2:37 pm|28Archaeologists and Barkindji custodians worked together to excavate the dingo burial.Credit:

      Barb QuayleArchaeologists and Barkindji custodians worked together to excavate the dingo burial.Credit:

      Barb QuayleText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA thousand years ago, the ancestors of today’s Barkindji people carefully buried a dingo (or garli, in the Barkindji language) in a mound of shells.Archaeologists recently studied the burial in what’s now New South Wales, Australia. They found that the Barkindji ancestors had buried the dingo with the same care and ceremony as any beloved human member of the community and looked after the grave for centuries. The burial reveals that dingoes were, as Australian Museum and University of Sydney archaeologist and study co-author Amy Way puts it, “deeply valued and loved” by ancient people in Australia.The long-lost dingoFive years ago, Barkindji Elder Uncle Badger Bates and National Parks and Wildlife Service archaeologist Dan Witter saw bones eroding out of a road cut in Kinchega National Park, an area along the Baaka, or Darling River, in New South Wales, Australia. Badger recognized the bones as a dingo, lying on its left side in what was once a carefully built mound of river mussel shells.At the urging of the Menindee Aboriginal Elders Council, which worried that erosion would end up destroying the dingo bones and any information about the past they contained, a team of archaeologists, working alongside Barkindji elders, excavated and studied the skeleton. The bones turned out to belong to an elderly male dingo, with worn teeth and possible signs of arthritis. Broken and healed bones suggested that he’d lived a tough, active life but also been cared for by people.And the layers of shells around him revealed that generations of Barkindji had tended his grave and ritually “fed” him by adding shells to the mound for centuries after his death. This is definitely not the first dingo burial ever found in Australia, but it’s farther north and west than any other example. It reveals a far more profound and lasting relationship between ancient people and dingoes than outside researchers, at least, had previously fully realized.“This confirms these traditions were much more widespread than we once thought,” said University of Western Australia specialist Loukas Koungoulos, the lead author of the paper, in a press release.The dingo’s bones revealed healed injuries and possible arthritis.Credit:
                                  Amy WayThe dingo’s bones revealed healed injuries and possible arthritis.Credit:

      Amy WayHunting kangaroos and snoozing by the fireThe dingo’s bones tell their own story. Koungoulos says he was probably between 4 and 7 years old, which would be late middle age for a wild dingo today. Heavily worn teeth were the first hint of the dingo’s senior citizen status, but the ends of his leg bones also showed signs of bone decay, probably thanks to long-term inflammation: possibly something like arthritis.And he was shorter than most wild dingoes, based on the length of his femurs. That’s not unusual—domesticated animals are often shorter than their wild relatives, and it doesn’t take many generations for that to show up—but it could say something interesting about exactly how close wild dingoes got to domestication in the centuries before European colonists wrecked everything.At some point, the dingo had suffered a broken rib and lower leg. Koungoulos suggests the injuries look like the aftermath of a kangaroo kick and may have happened on a hunt. The injuries themselves aren’t too surprising; wild dingoes hunt kangaroos, andAboriginal hunters worked with dingoesthe same way people in other parts of the world have hunted with dogs for millennia. What’s more striking is that the two injuries were long-since healed. Somebody nursed this dingo back to health after his kangaroo encounter.“What stands out about garli is that he was old and well-cared-for,” said Koungoulos. “The healed injuries, worn teeth, and careful burial tell us that this animal lived a long life alongside people, and that his death was marked intentionally and with respect.”Barkindji elders stand with the excavated dingo skeleton; in the center is Uncle Badger Bates, who first identified the bones as a dingo.Credit:
                                  Amy WayBarkindji elders stand with the excavated dingo skeleton; in the center is Uncle Badger Bates, who first identified the bones as a dingo.Credit:

      Amy WayHow dingoes became beloved community membersPeople have lived in this part of Australia for at least 40,000

Who will win mtn?One Mars spacecraft, two senators, and a cloud of questions“I think there’s plenty of fire lit under them already.”Eric Berger–May 18, 2026 3:49 pm|12A rendering of Rocket Lab’s design for the Mars Telecommunications NetworkCredit:

      Rocket LabA rendering of Rocket Lab's design for the Mars Telecommunications NetworkCredit:

      Rocket LabText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navNASA released a much-anticipatedcontract solicitationfor a Mars-orbiting spacecraft late last week, kicking off what is sure to be a hotly contested and potentially controversial procurement.At issue is $700 million, already appropriated by Congress, to build a spacecraft, launch it to Mars, and once there to serve as a vehicle to relay communications between the red planet and Earth. But the stakes may be even bigger than this, including the possible resurrection of the recently canceled Mars Sample Return mission.As part of the new solicitation, NASA says it will conduct the acquisition “as a full and open competition.” But will it? That’s the question that several people involved with this procurement process are asking. And it could turn messy, quickly.What is not controversial is that NASA needs a new spacecraft capable of relaying communications from Mars to Earth. NASA’s best communications relay today is the Mars Reconnaissance Orbiter, which has now been there for 20 years. It’s a great spacecraft, but it’s getting long in the tooth.The US Congress, more or less out of the blue, stepped up with $700 million in funding for a new Mars Telecommunications Orbiter in the “One Big Beautiful Bill” passed in 2025.Ars previously wrote about this legislation, which raised some eyebrows, as this is a large appropriation for a relatively straightforward spacecraft.Competition is narrowedProposals to build the spacecraft, the name of which has been changed to Mars Telecommunications Network (MTN), are due by June 15. A contract is expected to be awarded by October 1, less than five months from now.There was some curious wordingin the legislationthat funded the Mars orbiter. It specified the spacecraft must be selected from among US companies that “received funding from the Administration in fiscal year 2024 or 2025 for commercial design studies for Mars Sample Return; and proposed a separate, independently launched Mars telecommunication orbiter supporting an end-to-end Mars sample return mission.”Similar language ended up in NASA’s procurement notice, which states that to be eligible to bid, a company needs to “demonstrate that the Offeror proposed a separately and independently launched Mars telecom orbiter to support an end-to-end Mars Sample Return mission.”That seems like a curious requirement. Why must the bidder for a relatively straightforward orbiter around Mars have previously proposed an orbiter as part of an “end-to-end” mission to return samples from Mars? Was Congress seeking to be preferential for a particular company?Some hints about what is happeningThe One Big Beautiful Bill was signed into law on July 4, 2025. One month later, as part of itsQ2 Investor Update, Rocket Lab included a slide about its plans to support human missions to Mars. In the corner of the slide, the company noted that $700 million had just been appropriated for a Mars orbiter (MTO).“Rocket Lab was the only commercial provider to propose a MTO as part of an end-to-end Mars Sample Return mission to support human exploration and science missions to Mars,” the slide stated.Slide from Rocket Lab Q2 2025 Investor Update.Credit:
                                  Rocket LabSlide from Rocket Lab Q2 2025 Investor Update.Credit:

      Rocket LabRocket Lab is one of several companies eligible to compete based on the requirement that bidders have received funding in 2024 or 2025 for a “commercial design study” for a Mars Sample Return mission. Other eligible bidders include: Blue Origin, L3Harris, Lockheed Martin, Northrop Grumman, SpaceX, Quantum Space, and Whittinghill Aerospace.At an “Industry Day” event earlier this year, NASA officials indicated that the contract solicitation would be released by May 1. However, it was two weeks late. Weeks matter when NASA is seeking to have this spacecraft built and launched before the end of the next Mars launch window in late 2028.Ars’ Stephen Clark recently asked Rocket Lab chief executive Peter Beck if he knew why this delay had occurred. He said he did not. Beck was also asked whether NASA was moving quickly enough given the schedule constraints.“I think there’s plenty of fire lit under them already,” Beck said. “They understand the importance of the mission. So I think we just need to get it out there, and they need to get it awarded, and everybody needs to get to work.”One source indicated to Ars that this delay was due in part to a letter sent by US

Did AI do that?Legal fail: Don’t use AI to sue Facebook users for calling you a bad dateFake citations dashed a dude’s “Are We Dating the Same Guy” revenge lawsuit.Ashley Belanger–May 18, 2026 4:27 pm|41Credit:pikepicture | iStock / Getty Images PlusCredit:pikepicture | iStock / Getty Images PlusText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAn attempt to pressure Meta into removing a critical post from a Chicago Facebook group called “Are We Dating the Same Guy” may end in sanctions for lawyers whose takedown arguments appeared to rely on fake AI citations to support doxing claims.The case had already been dismissed with prejudice by a district court, which ruled there was no way to amend the complaint to possibly save it. But Nikko D’Ambrosio—who accused more than two dozen women of defaming him and blamed Meta for supposedly boosting the post to profit off its “entertainment value”—appealed anyway.Perhaps he felt confident despite his likely tough odds because he was relying on MarcTrent.AI, a law firm thatclaimsto use AI to “uncover legal opportunities traditional firms miss” and “increase legal success rates by 35 percent through predictive modeling.”In a 2025blogdiscussing the case, founder Marc Trent confirmed that the firm had “utilized our tech team to draft” the initial complaint. He boasted that the “evolved” firm uses “everything related to AI now,” suggesting that “even Meta can’t beat us” and claiming that a win would make Facebook safer for everyone.Laying out the case, Trent said that he assumed that Meta “would quickly distance itself by removing the post.” But when Meta didn’t, he figured that overcoming Section 230 claims would be his biggest hurdle in the fight. However, he insisted that “his firm’s technological capabilities” would level the playing field, making it possible to beat Meta’s well-resourced legal teams who are deeply schooled at defending against Section 230 claims.However, during the appeal, judges agreed that the case was so weak that Section 230 didn’t even factor in. And the firm’s seeming reliance on AI to “execute” arguments “with precision” apparently did not help matters at all.In anopinionFriday, David Hamilton, a senior circuit judge for the US Court of Appeals for the Seventh Circuit, wrote that the three-judge panel agreed that “this is a relatively rare appeal in which sanctions appear to be appropriate.” Not only was the appeal deemed “frivolous” for failing to advance D’Ambrosio’s arguments, but it was also littered with “mistakes and fictitious quotations” that “bear the hallmarks of the misuse of generative artificial intelligence,” Hamilton said.“Briefs and other court submissions that include fictitious quotations—inaccuracies discoverable with elementary professional care—are unacceptable,” Hamilton wrote.MarcTrent.AI did not immediately respond to Ars’ request for comment on whether the firm plans to fight the potential sanctions.The firm has until June 16 to request a hearing or file statements on whether sanctions are warranted.Man fails to scrub menacing text from FacebookD’Ambrosio’s legal fight started when a woman whom he briefly dated, Abbigail Rajala, blocked his number, and he persisted in sending a menacing text by using an alternate number.Screenshot of text D’Ambrosio wanted removed from Facebook.Credit: via Judge’s Opinion in D’Ambrosio v MetaScreenshot of text D’Ambrosio wanted removed from Facebook.Credit:

Escalated quicklyEbola outbreak: WHO declares emergency, US restricts travel, American infectedCDC is working to move the infected American and six others to Germany.Beth Mole–May 18, 2026 4:41 pm|33A poster displaying Ebola emergency contact numbers is pinned to a tent at the Busunga border crossing between Uganda and the Democratic Republic of Congo in Bundibugyo, on May 18, 2026.Credit:Getty | Badru KatumbaA poster displaying Ebola emergency contact numbers is pinned to a tent at the Busunga border crossing between Uganda and the Democratic Republic of Congo in Bundibugyo, on May 18, 2026.Credit:Getty | Badru KatumbaText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navThe Ebola outbreak first reported in the Democratic Republic of the Congo on Friday has seemingly escalated quickly into a large, uncontrolled multinational outbreak.As of May 17, there were 10 confirmed cases, 336 suspected cases, and 88 deaths in the DRC, as well as two confirmed cases and one death in neighboring Uganda, according to the latest data from the US Centers for Disease Control and Prevention, which has offices in the region. The numbers already put the outbreak within the top 10 Ebola outbreaks recorded by size, though still far from the worst—the 2014–2016 West African outbreak had over 28,000 cases and 11,000 deaths.International emergencyOn Sunday, the World Health Organization declared the outbreaka public health emergency of international concern (PHEIC), though it noted that it does not meet the criteria for a pandemic emergency. In making the PHEIC determination, WHO Director-General Tedros Adhanom Ghebreyesus cited several factors in addition to the immediate large size, including clusters of suspected cases and deaths in multiple DRC health zones, four deaths among healthcare workers, and a lack of apparent links between geographically distant cases and clusters. The features collectively suggest that the outbreak is larger than what is currently being detected and is spreading regionally.“Moreover, the ongoing insecurity, humanitarian crisis, high population mobility, the urban or semi-urban nature of the current hotspot, and the large network of informal healthcare facilities further compound the risk of spread,” the WHO said.The final extraordinary aspect of the outbreak is that it is caused by the uncommon Bundibugyo strain of Ebola virus, which has no clinically validated treatments or vaccines. This is only the third Ebola outbreak caused by Bundibugyo, which has had fatality rates of 25–50 percent.There are four virus strains known to cause Ebola disease in humans, and three have caused large outbreaks (Zaire, Sudan, and Bundibugyo). The most common strain is Zaire, for which treatments and vaccines have been developed. The viruses spill over from animals, including non-human primates and bats, and cause severe hemorrhagic fever, marked by diarrhea, vomiting, and bleeding. Person-to-person spread occurs via contact with bodily fluids and symptoms can develop between two and 21 days—though most often eight to 10 days—after an exposure.CDC response and infected AmericanOn Monday morning, the CDC announced on its website that it is implementingnew travel restrictions, including screening and monitoring Americans arriving from DRC, Uganda, and South Sudan, while also barring the entry of non-US passport holders who have traveled in those countries in the past 21 days.Additionally, in a CDC press briefing on Monday afternoon, Captain Satish Pillai, incident manager for CDC’s Ebola response, said that one American in the DRC has been infected after being exposed as part of their work there. The person developed symptoms over the weekend and tested positive late Sunday. The CDC is now working to transfer that person, along with six other Americans, to Germany, where they will receive care. Pillai did not answer questions about the person’s identity or their work.Serge, a Christian missionary organization, announced thatthe infected person is Dr. Peter Stafford, who has been working in the Nyankunde Hospital in Bunia, DRC, since 2023. The other six people the CDC is working to relocate are his wife, Dr. Rebekah Stafford, the couple’s four children, and a third doctor with the organization, Dr. Patrick LaRochelle. All three doctors had exposures, the organization said, but Rebekah Stafford and LaRochelle are currently asymptomatic.Pillai noted that the CDC considers the risk to the American public to be low.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.33 Comments

1968 年的旧金山，计算机科学家道格拉斯·恩格尔巴特在一场后来被称为「演示之母（The Mother of All Demos）」的发布会上，拿出一个带着两个金属轮子的木制小盒子，向世界介绍了一个新物种：鼠标。那是人类第一次公开用手上的鼠标，去牵引屏幕里的数字光标。此后的几十年里，这根小箭头几乎无处不在。它穿过办公软件、游戏界面、浏览器窗口和无数张电子表格，成为人类进入数字世界时最熟悉、也最沉默的向导。然而在这半个多世纪里，计算机的算力、形态和应用场景几乎改头换面，鼠标光标的本质几乎没有变过：它知道自己停在屏幕上的哪个坐标，知道 X 和 Y，却不知道你指向的是一行代码、一张发票，还是一张风景照。面对眼前不断闪烁的像素，它能做的事情依旧相当朴素：点击，拖拽，等待下一次点击。今天，Google 要用 Gemini 重新发明鼠标光标。在刚刚落幕的 Android Show 上，Google 几乎把围绕 Android、AI 与硬件生态的规划摊在了台面上。其中，一项名为「魔法指针（Magic Pointer）」的新功能，给古老的鼠标光标装上了「眼睛」和「大脑」。Google 的意图不言而喻，未来的 AI 交互不该依赖冗长的提示词，只需像在现实生活中一样，指着屏幕说一句：「把这个，移到那里。」那么问题来了，当鼠标光标终于学会「看懂」屏幕，它究竟会把人机交互带向哪里？这根睁开眼睛的 AI 箭头，到底能干嘛？要理解这项技术的意义，必须先看清当下 AI 工具最别扭的一面：交互成本。过去几年，大语言模型的能力一路狂飙，但使用门槛居高不下。为了让 AI 准确理解意图，用户被迫学习一套复杂的「提示词工程」：设定角色、补充背景、限定输出格式。为了一个简单的需求，写出几百字小作文更是家常便饭。不仅如此，典型的 AI 工具通常运行在独立的网页或应用窗口中，频繁打断用户的工作流。比如当你在阅读一份 50 页的 PDF 时想让 AI 总结一张图表，通常需要经历：截图 -> 保存 -> 打开浏览器 -> 进入 AI 网页 -> 上传图片 -> 输入提示词。Google 将这种繁琐的跨应用操作称为「AI 绕路（AI detours）」。这种跳转不仅效率低下，也容易打断人们工作时的注意力集中状态，即所谓的「心流」。为此，Google 提出的第一个交互原则，就是「保持心流」。在他们展示的实验性 AI 光标原型中，AI 的能力不再局限于某个特定的 App 或网页，而是附着在鼠标光标上，随时待命。触发方式也尽量克制：无需记忆任何快捷键，只需轻轻「晃一下」鼠标，AI 界面便会基于当前悬停的内容自动浮现，给出极具语境的操作建议。选中图，它会询问是否「对比」；悬停于段落之上，它会主动提供润色方案。整个过程没有任何需要学习的指令，完全跟着直觉走。来看几个极其直观的场景：第一，看图说话的终极形态。当你在浏览一张卡通的城市风景照时，传统鼠标只能点击图片放大。但现在，你只需把 AI 光标悬停在照片背景里的一栋建筑上，然后对着麦克风说一句：「给我移动图片的元素到这。」不需要解释「这儿」是谁，也不需要描述建筑外观。AI 光标会直接理解你所指向的像素，识别出对应元素，并成功移动。过去，鼠标只能告诉系统「我点了哪里」；现在，它开始告诉系统「我指的是什么」。第二，少写提示词，多用自然指代。当你在网页上看到一份极其复杂的烘焙菜谱，你不需要复制粘贴，也不需要写「请将以下菜谱里的所有食材分量乘以二」这种书面语。你只需要用光标高亮那段文字，然后随口说一句：「把『这些』的分量翻倍。」唰的一下，AI 直接在原地给你改写了一份新菜谱。第三，将像素转化为可交互实体。在计算机眼里，屏幕只是几百万个发光的像素点。但 AI 光标能将死板的像素（Pixels）转化为活的实体（Entities）。比如，你正在看一段旅行 Vlog，视频里闪过一家看起来很棒的餐厅。你按下暂停，把光标指过去，原本死气沉沉的视频画面瞬间变成了一个真实的、可交互的地点，旁边直接弹出了这家餐厅的订座链接。再比如，你随手拍了一张写满鬼画符的便利贴，鼠标一指，墨迹直接变成了一个可以打勾的 To-Do List。发现了吗？ 以前，是你去找 AI；现在，是 AI 顺着你的鼠标，乖乖来到了你的指尖。杀死 AI 提示词，回归人类直觉仔细想想，人类最强大的沟通工具其实是代词。当你和同事坐在屏幕前修改设计稿时，你绝对不会字正腔圆地说：「请将屏幕左上角坐标 (X:120, Y:350) 处的蓝色矩形向右移动 50 个像素。」你只会指着屏幕说：「把这个，往右挪一点，弄淡一点。」「那个餐厅看起来不错，怎么去？」「这段代码里的这个报错是什么意思？」在日常生活中，我们极度依赖「这个」和「那个」。手势配合着极简的口语，才是人类最高效的沟通密码。究其原因，我们身处同一个物理空间，共享着同一套视觉上下文。Google 敏锐地抓住了这一点，并将其提炼成了一条产品原则：拥抱「这」与「那」的力量（Embrace the power of This and That）。与其强迫人类去学习复杂的提示词框架，不如反其道而行之，把表达意图的脏活累活从我们身上剥离出去，让机器去适应人类最慵懒、最本能的「指手画脚」。好消息是，这套交互方式已经开始落地。Chrome 浏览器里的 Gemini 从今天起率先支持；Google 全新推出的笔记本电脑产品线 Googlebook，则将「Magic Pointer」直接内置进了操作系统层面，覆盖所有应用。Googlebook 的野心不止于鼠标。Google 将这个产品线定义为「Android 手机的完美伴侣」。类似苹果的 iPhone 镜像，用户可将 Android 应用无缝投射到 Googlebook 桌面，以原生比例运行，并在文件管理器中跨设备自由穿梭，彻底打破手机、平板与笔记本的生态壁垒。此外，Gemini 还能依你所需，在桌面生成专属动态 Widget（比如旅客的实时航班卡片）。在硬件设计上，所有 Googlebook 机型都会在机身上集成一条「Glowbar」发光条，让你一眼就能分辨出它与传统 Chromebook 或 Windows 笔记本的区别。首批 Googlebook 将由宏碁（Acer）、华硕（Asus）、戴尔（Dell）、惠普（HP）和联想（Lenovo）制造，预计今年秋天上市。有意思的是，三星缺席了这份名单。近期的消息显示，三星可能正在筹备一款搭载 Google 新系统的 Galaxy 笔记本，其下一场 Unpacked 发布会据传定在 7 月 22 日。至于底层的驱动核心，Google 虽未指名道姓，但通篇强化的「为智能而生的现代操作系统」以及 Android 与 ChromeOS 的深度融合，种种迹象均指向了传闻已久的 「Aluminum」系统。这意味着，AI 开始成了操作系统级别的基础设施。而当 AI 真正化身为你的鼠标光标，它便拥有了介入万物的权限——所见即所得，所指即所控。AI 人机交互，迎来十字路口回望 1968 年，那个惊艳世界的初代鼠标，功能简单得令人发指：追踪位置。 这五十多年来，鼠标加了滚轮、加了侧键、甚至加了风扇和配重块，但它的灵魂依然是一张白纸：它准确地标记着坐标，却永远无法理解坐标背后的意义。Google 的 AI 光标完成了交互史上罕见的进化：它不仅知道你在哪，更知道那是什么。过去一年，无数拿了融资的创业公司挤破头，试图造出下一个「AI 时代超级入口」。大家疯狂地卷对话框的拟真度、卷 Agent 的复杂工作流。但 Google 这次用行动给全行业结结实实地上了一课：最好的科技是什么？是润物细无声。聊天框（Chatbox）从来都不是 AI 的最终形态，它只是一种过渡期的妥协。最好的 AI，应该退居幕后，成为一种附着在你日常动作上的基础设施，而不仅仅是某个需要单独打开的应用。从黑底白字的命令行（CLI），到图形界面的鼠标点击（GUI），再到移动时代的触屏滑动（NUI）。过去几年，大语言模型让我们短暂地倒退回了打字交流的时代，让无数人患上了 Prompt 焦虑症。但今天过后，我们知道了，那不过是黎明前的一段弯路。真正好用的 AI，终究要学会像人类一样思考：看懂你的每一个眼神，听懂你的每一句「把这个，放到那儿去」。58 年前，当道格拉斯·恩格尔巴特握住那个简陋的木制鼠标时，他的终极梦想是「增强人类的智慧」。58 年后，当 AI 附身于这个古老的指针，机器终于开始真正「看懂」这个世界。属于提示词工程师的时代终将落幕，人机交互的终极闭环，也将在一句句含糊不清的「这个、那个」中，向前迈出历史性的一大步。附体验地址：https://aistudio.google.com/apps/bundled/ai-pointer-create?showPreview=true&showAssistant=true&fullscreenApplet=truehttps://aistudio.google.com/apps/bundled/ai-pointer-find?showPreview=true&showAssistant=true&fullscreenApplet=true作者：莫崇宇

📱iPhone 17 Pro 官降 1000 元🧠ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施🤖估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相🌐Claude 重新允许接入龙虾，开发者：这是变相涨价📋腾讯辟谣「AI 一号位将离职」💰黄仁勋基金会捐出 1.08 亿美元 AI 算力📱影石 CEO 回应 Luna 定价贵：5299 元是美国价格🚗鸿蒙智行洽谈打造玛莎拉蒂新能源车🚙华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗💻英特尔成为迈凯伦官方计算合作伙伴⛽本田上市以来首次年度亏损📋LinkedIn 宣布裁员，涉及工程、产品和市场营销部门💡Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」🤖曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线📱荣耀 600 系列定档 5 月 25 日🔍Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光🧠腾讯开源 AI Agent 记忆系统，Token 消耗最高降低 61%🧸泡泡玛特否认进军传统家电，LABUBU 冰箱一季度收入约 600 万iPhone 17 Pro 官降 1000 元昨晚，苹果无预警在天猫「Apple Store 官方旗舰店」启动 iPhone 17 Pro 系列 618 优惠活动，Pro 和 Pro Max 同步降价 1000 元，起售价分别来到 7999 元和 8999 元。本次活动还覆盖 iPad、Apple Watch、AirPods 等系列，非 iPhone 产品可使用 50 至 100 元不等的惊喜券，并可叠加 88VIP 消费券、店铺优惠及国家补贴等多重优惠。ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施据彭博社报道，苹果与 OpenAI 的合作关系已出现明显裂痕，OpenAI 正在联合外部律所评估可能的法律行动。知情人士透露，OpenAI 律师团队正与外部律所合作，研究包括向苹果发送违约通知在内的多项方案，但目前尚未正式提起诉讼。OpenAI 仍希望在庭外解决与苹果的分歧，相关法律行动预计将在马斯克相关诉讼案结束后才会推进。苹果与 OpenAI 的合作始于 2024 年 6 月，苹果在 WWDC 上将 ChatGPT 整合进 Siri 及 iOS 系统，涵盖文字生成、视觉智能、图像生成 App「图乐园」等场景。估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相昨天，前 Meta FAIR 研究院总监田渊栋正式宣布联合创办 AI 初创公司 Recursive Superintelligence，距其确认从 Meta 离职刚好过去半年。他在 X 发文称，拒绝了 OpenAI、xAI、Anthropic 等多家头部 AI 公司的邀约，选择自己创业：「趁着自己还年轻，去当一家新初创公司的联合创始人。」Recursive 同步披露完成一轮 6.5 亿美元融资，估值达 46.5 亿美元，由 GV（谷歌风投）和 Greycroft 领投，AMD Ventures 与英伟达跟投。公司押注的核心方向为「递归自我改进」（Recursive Self-Improvement），即构建能够自动发现知识、持续优化自身的 AI 系统。Claude 重新允许接入龙虾，开发者：这是变相涨价Anthropic 昨天官宣，将于今年 6 月 15 日起为 Claude 订阅版用户引入独立的 Agent SDK 月额度，OpenClaw 等第三方 Agent 工具由此重新获准通过订阅账户接入 Claude。按照新规，Agent SDK 的调用将从用户的常规订阅限额中彻底剥离，另设专属月度额度池，额度按账单周期刷新，未用完的部分不结转至下月。Pro 版和 Team 标准席位每月获得 20 美元额度；Max 5x 及 Team 高级席位为 100 美元；Max 20x 与 Enterprise 座位制高级席位最高可达 200 美元。该额度专用于 Claude Agent SDK 项目（Python 或 TypeScript）、claude -p 非交互命令、Claude Code GitHub Actions 集成，以及通过 Agent SDK 认证的第三方应用。此次新规被视为 Anthropic 对今年 4 月 4 日封禁政策的正式解决方案。彼时，Claude Code 负责人 Boris Cherny 宣布，Claude Pro 及 Max 订阅额度即日起不再覆盖 OpenClaw 等第三方工具的调用，理由是此类工具的使用模式对 Anthropic 的算力基础设施造成了难以为继的压力。开发者社区对此反应冷淡。有开发者直言，20 美元的 Agent 额度在高强度任务下消耗极快，实际可用量远低于封禁前的订阅套餐，并将此次调整定性为「变相涨价」。腾讯辟谣「AI 一号位将离职」昨天，腾讯官方发文辟谣，明确否认外界流传的「AI 一号位即将离职」传言。腾讯公关总监张军表示，「太恶劣了，还编出各种搞笑的情节。」谣言起源于社交平台上的多条帖文，内容称某大厂 AI 一号位因被绿色事业部抢走算力资源而准备离职，并点名影射腾讯 AI 负责人姚顺雨。黄仁勋基金会捐出 1.08 亿美元 AI 算力据路透社报道，英伟达 CEO 黄仁勋与其妻子洛丽创立的基金会，从云计算公司 CoreWeave 购入 AI 算力，并将其捐赠给多所大学及其他非营利研究机构，用于科学和人工智能研究。根据昨日提交的一份文件，此次捐赠的算力价值目前已达 1.083 亿美元。英伟达还表示，计划向部分受捐机构提供免费工程服务。CoreWeave 是一家专注于 AI 应用的云计算公司，其向客户提供的 GPU 正是由英伟达设计制造。今年 1 月，英伟达向 CoreWeave 投资 20 亿美元，一度成为该公司第二大股东。影石 CEO 回应 Luna 定价贵：5299 元是美国价格昨天，影石 CEO 刘靖康回应旗下旗舰双摄云台相机 Luna Ultra 的定价问题，表示网传单机 5299 元的售价属实，「但是是美国的价格，国内给我再大的胆子都不敢卖这个价格」。此前，博主「数码闲聊站」爆料称该机预计售价 5299 元、全能套装预计 6499 元，随即在网络上引发「定价过高」的质疑。鸿蒙智行洽谈打造玛莎拉蒂新能源车据《云见 Insight》报道，华为鸿蒙智行、江淮汽车、斯特兰蒂斯集团（Stellantis）及旗下玛莎拉蒂品牌，正就共同打造玛莎拉蒂品牌的新能源车展开洽谈。按照规划，四方合作模式与鸿蒙智行 「五界」 高度相似：华为主导产品定义、提供核心技术，江淮联合研发并负责生产制造，玛莎拉蒂提供造型设计与品牌背书。合作车型将分为国内和海外版本，国内版归属于尊界品牌，海外版将挂玛莎拉蒂车标，首款车型目前正处于造型设计阶段，计划于明年下半年量产。华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗华晨宝马昨日公布年度节水数据，显示其沈阳生产基地 2025 年单台生产取水量降至 1.4 立方米/台，同比降幅 10.8%，连续八年稳步下降。在水资源消耗总量方面，华晨宝马沈阳基地全年取水总量降至 75.2 万立方米，较 2023 年降幅超 40%，连续三年递减。再生水使用量超 68 万立方米，同比增长 71.9%，官方称节约用水量可满足近 9300 户沈阳家庭一整年的生活用水需求。华晨宝马大东工厂北厂区实现所有生产工艺自来水消耗为零，全部改用再生水替代；铁西工厂及里达厂区涂装车间冷却塔也已扩展使用市政再生水，厂区绿化灌溉此前已于 2024 年实现 100% 使用市政再生水。英特尔成为迈凯伦官方计算合作伙伴英特尔昨日宣布与迈凯伦车队达成多年期战略合作伙伴关系，正式成为迈凯伦万事达一级方程式车队、艾睿迈凯伦印地车队及迈凯伦 F1 模拟赛车队的官方计算合作伙伴。根据合作协议，英特尔至强处理器和酷睿 Ultra 处理器将为迈凯伦的核心性能负载提供算力支持，覆盖计算流体动力学、空气动力学分析、车辆动力学仿真、比赛策略分析等方面。英特尔的技术还将应用于赛道边缘计算、数字孪生与 AI 预测建模等场景。本田上市以来首次年度亏损昨天，本田公布了 2026 年 3 月期的合并财报，期内营收 217966 亿日元，同比微增 0.5%；归母净亏损 4239 亿日元，上一财年同期归母净利润为 8358 亿日元，由盈转亏。亏损主因是电动化战略大幅调整。本财年内，本田决定取消北美多款电动车型的开发与上市计划，由此确认减值、处置及合同赔偿等相关损失合计约 15777 亿日元，全部计入汽车业务。摩托车业务为本财年唯一实现正向盈利的板块。展望明年，公司预计营收 231500 亿日元，归母净利润恢复至 2600 亿日元。LinkedIn 宣布裁员，涉及工程、产品和市场营销部门据彭博社报道，职业社交平台 LinkedIn 昨日宣布裁员，涉及工程、产品和市场营销等多个职能部门。LinkedIn CEO Daniel Shapero 在发给员工的备忘录中表示，公司需要为用户提供更大价值，并实现更高的盈利水平，此次组织架构调整旨在为公司未来发展做好定位。LinkedIn 发言人随后确认了相关变动，称此次调整属于「常规业务规划」的一部分，但未披露具体裁员人数。💡 Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」据《商业内幕》昨日报道，Anthropic CFO Krishna Rao 在播客节目《Invest Like the Best》中透露，Claude 目前已承担公司 90% 以上的代码编写工作，并正在深度介入财务流程，推动白领工作从「执行」向「监督」转型。Rao 还透露，公司已借助 Claude 生成财务报表，每月财务审查流程在人工介入之前已达到「90% 至 95% 完成」的状态。他提到，部分内部报告的生成时间从原来的数小时压缩至 30 分钟。在人员配置方面，Rao 将 AI 的引入描述为生产力「加速器」。他称，由于 AI 的介入，员工得以将更多精力从信息收集转向决策判断，Anthropic 反而因此扩大了招聘规模。我们雇用了更多人，因为没有什么工作是做不完的。Rao 还描述了一种新的工作形态：员工越来越多地扮演 AI 系统的「管理者」角色，团队可以同时部署多批 AI 智能体并行处理项目。他将这一趋势概括为「每个人都在某种程度上成为管理者」，并认为由此带来的生产力潜力「才刚刚开始显现」。他同时强调，Anthropic 内部推崇「人才密度」而非「人才规模」，认为高密度的顶尖 AI 研究人才配合最强模型，才是真正具有竞争力的组合。曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线知名爆料者 Leo 昨日披露，GPT-5.6 研发已全面推进，预计最早下月正式发布。开发者在 OpenAI 内部 Codex 日志中发现了两个内部代号 —— ember-alpha 和 beacon-alpha，以及一条路由映射记录。日志显示，绝大多数模型调用仍指向 GPT-5.5，但有一条条目明确指向 GPT-5.6，暗示 Codex 环境可能已在测试新模型。此外，爆料者 Chetaslua 称 OpenAI 计划于本周四推出旗舰模型的「超快模式」，响应速度或可提升 2 至 3 倍，主要面向延迟敏感型任务。荣耀 600 系列定档 5 月 25 日荣耀昨日宣布，旗下荣耀 600 系列将于 5 月 25 日 19:00 正式发布。官方预热信息显示，荣耀 600 系列主打「4K 闪光微单 Live」影像方向，搭载 2 亿像素超清大底主摄，并内置 8600mAh 青海湖大电池。Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光据 Sources 报道，Google 计划在 5 月 20 日举行的 Google I/O 大会上

相信这样的场景很多人都不会陌生：在微信上聊工作、讨论旅行计划，甚至只是和朋友闲聊，聊着聊着突然想让 AI 帮忙总结重点、做个规划，结果却只能一段一段复制聊天记录，再手动粘贴进 AI 对话框里。再比如，中午小憩了一会，醒过来发现群里发了几百条消息，想吃瓜又得一条一条翻……如果能一键将聊天记录转发 AI，岂不美哉？这样的功能终于来了，但遗憾不是在微信里：现在，把微信和元宝更新到最新版本之后，聊天记录转发可以选择直接发给元宝来总结。虽然同属腾讯生态，但微信和元宝之间始终有种微妙的距离感——终于，这次两者之间的联动，又紧密了一点。爱范儿也第一时间体验了这个功能，看看它究竟好不好用。产品很美好，效果很一般要想把微信聊天记录转发给元宝，方法很简单，首先选中想要转发的聊天记录，选择「转发到其他应用」，就能直接发给元宝，可以补充一些提示词。整个过程有点像把讨论、聊天过程中零散的内容，让 AI 帮忙进行整理，因此这个功能首先最合适的场景，就是各种工作群的讨论和安排信息。直接处理聊天记录有一个好处，元宝可以将讨论精准和发言者匹配，很适合用来处理会议纪要和工作安排。旅游计划同样也是一个高频场景，因为行程一般都是靠你一言我一句的对话讨论出来，这时候就可以直接交给元宝处理，还能让它帮忙推荐旅游地点。还有一个有点邪修的用法：平时工作忙没时间看群，又担心错过什么热点或者话题，可以直接选中所有聊天记录，让元宝帮你看群帮你总结。又或者讨论了很久都不知道聚餐吃什么，也可以把讨论的过程选中，让元宝帮忙推荐。尝试了很多场景，大部分情况下都达到了预期，但总体来说，还有一些体验细节不够完善。目前元宝最擅长处理的聊天记录以文字为主，而如果聊天记录中包含文件，特别是涉及多个文件，效果会比较随缘，很多时候会无法读取文件。即使能成功读取，也不一定能按照提示词进行任务处理：整理发票有时候都未必能填对，让他根据文件做个思维导图，也经常失败。要知道，整理聊天里的发票真的是很主流的场景啊！我能很明显感觉到一个落差：明明将聊天记录发送给 AI 是一个很有价值的场景，但元宝交付的结果，总让我略感失望，并且忍不住想，如果能转发给豆包或者 ChatGPT，会不会更好？更深层的问题，是这个聊天记录转发的交互总给我一种「绕远路」的感觉。在此之前，元宝已经作为联系人存在于微信之中，处理聊天记录这件事，却依旧需要「跳转」这一整套外挂式流程， 过程中需要多次点击。或许是因为微信严格的隐私保护规则，元宝只能以「临时聊天」的形式处理聊天记录，不会留下历史记录，关闭后无法找回。对于微信来说，这样的功能似乎「可有可无」。如果微信想要 AI 化，他们或许会有更好的实现方式，例如直接做一个「AI 群友」，at 它就能回答问题或者处理聊天记录；以及一个常驻对话列表的 Agent，可以直接对话，可以直接接收聊天记录。但这件事没那么容易，原因已经是「老生常谈」：微信需要简约，需要克制，因为每一点小变化都会影响到十亿级用户。在元宝这个功能上线的同一天，腾讯发布了今年第一季度财报，在股东大会上马化腾表示，在微信里做 Agent 需要「更长远的考虑，大家要有点耐心，这个不是匆忙就能出来的。」腾讯的 AI 王牌是微信马化腾还提到腾讯的 AI 战略：原来一年前我们以为上了船，后来发现那个船漏水了，现在感觉站上去了，还坐不下去，还是希望船速能快一点。从财报来看，AI 已经成为了腾讯的重要投入，一季度研发投入达 225.4 亿元，同比 19%；资本开支 319.4 亿元，主要用于算力、存储等云端基础设施的扩张。今天中午，有自媒体传出「疑似腾讯 AI 一号位即将离职」的消息，下午就被腾讯公司火速辟谣，本身也向外界释放出一个信号：腾讯 AI，不容有失。根据 AICPB 最新榜单，在国产 AI 应用排名中，腾讯元宝排名第十，访问量为 1773 万，远逊于豆包（1.63 亿）和千问（4424 万）。在 AI 领域中，腾讯和苹果有着相似的处境：产品入口和用户规模都是顶级配置，AI 模型本身却还支撑不起这种期待。元宝 AI 社交应用「元宝派」生成式 AI 火了 3 年，从去年年底开始，AI 应用之间的战斗已经进入了「场景化」的比拼，不仅出现「龙虾」OpenClaw 这样的全新形态，传统的 ChatBot 交互模式也集成了生活服务，像阿里千问和字节豆包，都在积极将电商等各种生态内的优势服务，和自家的 AI 助手相结合。这本来应该是腾讯的舒适区，因为国内最强大、最全能的生活应用，也就是微信，就在腾讯手中。但很明显，微信对「AI 化」抱有非常慎重的态度。腾讯不能承受因技术或功能不成熟而扰乱微信庞大用户群的体验。只是对于腾讯来说，在 AI 性能落后其他对手的情况下，微信这张王牌的价值还是太重要了。，于是，我微信和元宝虽然没有完全敞开双臂拥抱彼此，也开始小心翼翼试探边界。究其原因，除了微信相对封闭的生态逻辑，还有两个问题没能被回答：微信究竟要如何「AI 化」？而元宝在其中又应该扮演什么样的角色？The Information 报道指出，即使腾讯在下重本加速建设自己的混元模型，微信方面还是认为混元当前水平不够拔尖，甚至开始为微信 Agent 测试第三方 AI 模型，这带来了隐私和整合上的挑战性。微信方面，爆料称其实一直在做自己的 AI Agent 模型技术，计划在年中发布一个微信智能体，可以和整个小程序生态相结合，提供打车、购物的服务，并且将以联系人和聊天的方式直接集成微信之中——但就像前面提到的，这件事太重大、影响面太广了，急不得。一旦这个产品问世，可能将是全国最多用户的 AI 入口。微信需要 AI，但未必需要「元宝」，这才是腾讯现在最微妙的问题。

大疆的 Pocket 系列三年不更新，一更新就来了两台——过去一个多月，Pocket 4 和 Pocket 4P 可以说是最受瞩目的相机，就在今晚的戛纳电影节上，Pocket 4P 以口袋电影机的身份在欧洲电影殿堂亮相，距离正式发布时间，相信不会太远了。与此同时，大疆也对 Pocket 3 进行了降价处理，于是我们看到了泾渭分明的 Pocket 系列产品线：Pocket 3：入门级的手持云台 vlog 机，2299 元Pocket 4：全能的手持云台相机，2999 元Pocket 4P：准专业级的口袋电影机，预计价格会来到 4000 元档相信有许多持币观望的读者，也和爱范儿抱着同样的疑问：三台 Pocket 我该怎么选？有了 iPhone，我还需要 Pocket 吗？结合实际体验，爱范儿一次性帮你理清楚。Pocket 三代同堂，怎么选？先仔细聊聊 Pocket 4P。虽然目前 Pocket 4P 依旧没有公布更详细的参数，但参数在本次形态更新面前，是最不重要的话题——从产品逻辑来看，Pocket 4 已经将单摄的手持云台相机推向较为成熟的状态。此时，增加一颗摄像头成为一件顺理成章的事情。绝大多数手机的主摄和单摄 Pocket，提供的都是广角视野。广角适合交代环境，适合把大山大河框进画面。但当你拿着设备走入拥挤的人群，广角的缺点就暴露无遗：画面太杂乱了。此时，一颗长焦镜头可以完成空间上的压缩，把无关的背景剔除出去。大疆将这颗多出来的摄像头定格在 3×，这是一个非常巧妙的数字。它比主摄更贴近人眼对焦后的视线范围。从我们上手体验来看，Pocket 4P 这颗 3× 镜头还提供了比较自然的虚化，非常适合拍摄更凸显主体的人像或特写画面。但对于用户来说，多的这一颗摄像头，并非绝对的正收益：Pocket 系列自诞生起，核心卖点就是「无感携带」，增加镜头，意味着增加重量、侵占内部空间，甚至可能牺牲一部分续航。其实这种担心，恰恰帮大疆划清了 4P 和 4 之间的界限。Pocket 4P 是给那些「创作者」准备的。他们清楚地知道自己要拍什么，素材最后会进入 Mac，在 DaVinci Resolve 或者 FCPX 里进行精细的剪辑。对他们来说，为了在极轻便的体积下获得原生焦段的景别切换，以及高度统一的素材规格，牺牲一点点重量完全值得。而对于大多数普通人来说，Pocket 4 才是那个更轻量、更理性的选择。虽然前代 Pocket 3 打造出千万台的神话，但其本身并不是一台完美无缺的相机——只有 USB 2.0 导致数据传输过慢、像素量过低以及动态范围不够大，这些都是问题。而 Pocket 4 提供了 3700 万像素，更高的像素意味着内置的无损 ISZ 变强了。在 Pocket 4 上，通过传感器裁切获得的 2× 焦段更为清晰，足以应付日常记录。除此之外，它还内置了 107GB 存储，带来了更好的动态范围。忘记带内存卡、传输速度和动态范围不够的问题，被一次性解决掉了。可以说，Pocket 4 就是目前单摄手持云台相机的完全体。至此，Pocket 4 系列两台机身的区别已经很清晰了：如果你追求生产力，强调原生光学焦段带来的画质与虚化，需要 3× 镜头获得更丰富的画幅景别，那么 Pocket 4P 毫无疑问是更好的答案；如果你只是希望有一台能轻松塞进裤兜、防抖优秀、可以独立于手机拍摄视频和 Live Photo 的机器，那么 Pocket 4 凭借更高的像素和好用的 2× ISZ，已经绰绰有余。最后是 Pocket 3。在 Pocket 4 补齐了大量痛点的情况下，Pocket 3 还有存在的意义吗？当然——新设备的出现，最大的贡献往往是把老旗舰的价格打下来。在 Pocket 4 系列轮番上场的间隙，我们观察到 Pocket 3 已经完成了一轮官方降价，在某些平台叠加各种补贴的情况下，已经落到了 2099 的价格。甚至在未来，我们能以 1 开头的价格买到 Pocket 3，也不是不可能。对于预算有限、或者刚刚萌生拍视频念头的新手，这就是性价比最高的 「守门员」。最后，按照爱范儿编辑部的推测，Pocket 4P 的定价大概率在 4500 元左右，但也不排除大疆卷起来定一个 3999 元的价格。此时，如果将 Pocket 系列按照价格排开，我们会发现一个巧妙的设计——从 2000 元档的 Pocket 3 到 3000 元档的 Pocket 4，最后到 4000 元档的 Pocket 4P，大疆正在把无人机上那套 Pro、Air、Mini 的组合拳搬到口袋相机上。这套阵容已经在无人机市场证明了它的成功。现在，Pocket 也将组成这样的阵容了。有了 iPhone，我还需要 Pocket 吗？在 Pocket 4P 的上手文章发布后，我收到不止一个朋友的问题：我有 iPhone，还需要买 Pocket 吗？理清了 Pocket 家族的内部定位后，这是我们必须面对的现实问题：既然大家手里都有一台性能强悍的 iPhone，为什么还要花钱买一台功能单一的云台相机？这其实是一个关于「设备属性」的哲学问题。在多次迭代后，iPhone 在动态影像领域的基础不可谓不强，三摄变焦顺畅、色彩高度统一、防抖优秀、Apple Log 甚至已经更迭到第二代。但问题在于，手机始终不是完全针对拍摄设计的特化产品——除了拍摄以外，iPhone 还需要承担复杂的日常任务，无时无刻需要分散注意力，随时可能会有工作消息、语音电话打断你的拍摄、需要考虑拍摄之外手机还有没有充足的剩余电量保证你与世界的关联，还需要注意拍摄会不会占满本就贵如黄金的存储空间……而 Pocket 则是完全服务于拍摄需求的相机：不能打电话，不能回微信，不能刷社交软件，一切与拍摄无关的东西，都不会出现在你手里的 Pocket 上。更别提，Pocket 还有机械云台天生在结构上比 iPhone 的防抖更稳定、形态比 iPhone 更适合握持、物理虚化比 iPhone 更自然、视频规格比 iPhone 更丰富等等综合优势了。所以，如果我手里已经有一台主力 iPhone，现在到底该不该买 Pocket？如果现在需要，哪台是最适合我的？你可以对照下面几个问题，做一个购前诊断：过去三个月里，你是否因为手机突如其来的电话或微信，打断过拍摄？拍视频时，是否因为手机存储空间不足，而束手束脚？炎炎夏日，有没有遇到过手机拍照录视频发热掉帧甚至死机的情况？如果你对这些问题的回答是肯定的，那么购入 Pocket 3 和 Pocket 4 都是不错的选择。具体来说，如果你有拍高清 live 图的需求，如果你喜欢用手机的 2× 镜头拍摄，那么选择 Pocket 4，否则，Pocket 3 其实就够用了。在这个基础上，你还想拿 Pocket 拍拍商单、拍拍演唱会，搞点严肃创作，那么 Pocket 4P 是一步到位的选择。当然，如果前面几道题，你的答案都是否定的，那么请捂紧钱包。iPhone 的防抖和画质足够满足你的需求。你可以随拍随剪，直接通过 AirDrop 分享，且丝滑的生态体验依旧是目前任何第三方设备都无法替代的。而从行业角度来看，爱范儿更建议你等一等——目前，口袋云台相机的玩家可不止大疆一家——和 Pocket 4P 同样「犹抱琵琶半遮面」的，还有影石 Luna，而各大手机厂商也正跃跃欲试，准备下场分一杯羹。因为 Pocket 4P 这样的产品，已经严重威胁到了手机厂商的安全区。很多人愿意买 6000 元甚至 8000 元的影像旗舰手机，很大程度就是为了长焦。现在，花更低的价格，Pocket 4P 就能提供一个素质相当的长焦镜头，以及更好的云台防抖、更专业的整机解决方案。没有手机厂商能对此熟视无睹。据爱范儿了解，vivo、OPPO 都已经将手持云台相机的产品规划提上日程，而小米、华为也在持续关注这个赛道。2026 下半年到 2027 上半年，想必会有更多手持云台相机纷至沓来。既然入局者多了，我们作为消费者，反而可以更淡定些。你可以把钱捏在手里，再等一等。我们相信，这个年轻的品类，还会有更精彩的产品出现。

鸿蒙智行车型阵营里可能缺一辆 MPV，但中国汽车市场已经不缺 MPV 了。别克 GL8，丰田埃尔法和传祺 M8 等车型在燃油车时代占据了 MPV 就是商务接待用车的印象之后，新能源 MPV 迎来了数款重量级车型，极氪 009 主打高端，意在取代丰田埃尔法乃至雷克萨斯 LM 的生态位，理想 MEGA 则希望利用 MPV 的空间优势重塑大家庭用车的产品定义，岚图梦想家成为这家新势力最依仗的产品成功和别克 GL8 分庭抗礼，小鹏 X9 想要以宜商宜家，是 MPV 也可以是 SUV 的灵活定位吸引更多受众……新旧势力在 MPV 市场的攻伐，发生在 2023 年，但现在已经是 2026 年了，零跑做一款旗舰 MPV 只需要把性价比拉到极致，但鸿蒙智行做一款旗舰 MPV 并卖四五十万元，需要做的就多了，当然我承认坐在二排的老板们对智界代言人刘亦菲肯定都是心驰神往，但除了刘亦菲这个核心竞争力之外，鸿蒙智行的首款 MPV 肯定需要更多击中老板坚硬内心和柔软脂肪的东西。这些东西是车载制氧系统、全维包裹安全气囊、华为悦彰非凡系列音响、二排主动座椅旋转、侧门交互式迎宾光毯等等等。最终，智界 V9 分为 Max、Max+、Ultra 和 Ultra+ 四个版本，售价分别为 39.98 万，42.98 万，45.98 万和 51.98 万起。舒适重要，安全重要，舒适且安全更重要！过去几年的时间里，鸿蒙智行成功地让国内汽车行业重视起来零重力座椅和主动安全来，是的，这两项鸿蒙智行都不是发明者，但确实在打开用户心智和开拓市场上起到了主要作用。前者代表座舱的舒适性，通过模拟太空失重环境下人体自然放松姿态的设计。通过特殊的人体工程学优化，使臀部、腰部、背部和腿部受力均匀，释放肌肉压力。后者代表汽车安全从被动防御走向主动避让，在汽车发生事故前，通过电子控制系统感知环境、预判风险并主动介入（如自动制动、转向控制），从而避免或减轻事故。但是，二者的矛盾在于在汽车行驶状态下，零重力座椅一般是没法使用的，这是因为当乘客躺在大仰角的座椅上享受时，身体与安全带、气囊的相对位置已经发生了相当大的变化。董车会之前报道过，根据 CIDAS 的数据，一旦座椅靠背夹角超过 115°，正面碰撞中的重伤风险会陡增 4.4 倍。此时，传统的 B 柱固定式安全带无法再有效约束骨盆，人体会因为惯性产生极度危险的「潜滑现象」——整个人滑向座椅下方，安全带可能会锁死在颈部导致窒息。采埃孚的试验甚至显示，在 50km/h 的碰撞下，躺倒状态受到的脊柱伤害堪比从 5 楼坠落。打过工的人都知道，当你向老板解释「零重力座椅虽然很舒服，但是不能在行驶状态下使用」时，老板这个时候一般不会听你的解释，只会跟你说「我不管，我想用，给我想一个解决方案出来」。今天，这个方案就被智界 V9 想出来了。智界 V9 全球首发了二排全维包裹安全气囊，这个气囊集成于二排座椅靠背，触发瞬间会形成一个独特的「U 型」环绕结构，在零重力、躺姿等多种姿态下可以全面防护头部、颈部、胸部，有效防止乘员在车辆碰撞之后身体滑落、碰撞与下潜，智界宣称历经百项严苛工况验证，无误伤风险。与之配套的还有二排坐垫前端也有专门布置的臀部气囊（BAB），在碰撞瞬间快速充气形成物理坡度，顶住盆骨不让身体滑脱，从而大幅降低腰椎压力。做了这么多工作，目的只有一个：就是让二排的老板在车辆行驶的时候好好躺一躺，好好休息一下。车载制氧：让高精力人群更有精力一直以来，对于什么样的人能做老板有个迷思，甚至还有「基因彩票」的说法，就是高精力人群更容易创业做老板。这些人可以每天工作十几个小时不知疲倦，只需要睡四五个小时就能满血复活。正如前面所说，很多老板之所以看起来可以每天工作十几个小时不知疲倦，很有可能是他可以中间随意选择休息时间。而屏幕前后的你和我，并不能在上班时间睡觉。更科学的说法是，高精力人群背后，跟摄氧量有正向关系，很多所谓的低精力人群很可能有鼻炎咽炎或者心肺功能低下导致日常处于轻微缺氧状态，导致精力不足。类似的情况也会发生在车上，很多人坐高铁感觉昏昏欲睡就是车厢里二氧化碳含量高，氧含量低，这种情况就容易困，密闭的汽车座舱也可能出现这种情况。智界 V9 给出的解决方案也很简单，装个制氧系统。这套制氧系统采用了 PSA（Pressure Swing Adsorption）分子筛变压吸附技术，提供了「弥散+鼻吸」两种模式，鼻吸出氧浓度 93±3%。额定制氧排量 8L/min，峰值 10L/min，比一般的家用制氧机大很多。2 个鼻吸口支持 1 拖 2，最多可供 4 人同时吸氧，弥散供氧与空调共用，四个出风口协同工作，30 分钟即可让整个座舱达到富氧状态。除了解决了传统制氧机体积大，噪音大的不足之外，鸿蒙智行的看家本领是智能化。这套制氧系统支持语音控制开启，中控屏实时显示状态；内置舒缓解压与高原出行两种模式，可自动调节供氧浓度，也可以联动小憩模式，帮助身心放松。很多人在森林景区，绿肺公园里感觉心思空明浑身舒爽，也是因为在富氧环境下，血氧饱和度一般会更高一些。富氧环境在特定场景下确实会有积极作用，比如早起后在通勤路上进入富氧环境有助于从昏沉状态中恢复；在运动后进入富氧环境可以加速乳酸代谢，缓解肌肉酸痛。更典型的场景就是针对本身心肺功能或者呼吸系统有问题的人，富氧环境有助于把缺氧状态调整为正常状态，还有就是高原出行环境，正常的氧气浓度可以有效缓解高原反应。在制氧系统之外，针对座舱健康还有鸿蒙 ALPS 健康座舱 3.0，通过把 MOFs 可吸附降解材料融入顶棚、门板、地毯等内饰，即便熄火也能持续吸附有害物净化醛类。UVC 空调主动杀菌，对流感等病毒灭杀率超 99.99%；前后双负氧离子净化器，可以释放超过西双版纳雨林 5 倍的负离子；这一整套净化杀菌和补氧系统，提供的是「健康」的价值，毕竟对于老板来说，紧急时刻生命第一，不紧急时刻健康第一。当然，需要着重说明的是，并不是所有人在所有场景都需要制氧系统制造一个富氧环境，如果本身不处于缺氧状态，血氧饱和度在 95% 以上，不需要运动恢复，没有昏沉嗜睡情况，不在高原地区，那就不太需要吸氧，长期过度摄氧反而有害健康。旋转座椅，优雅，还是优雅冰箱彩电沙发之所以从争议性的玩笑变成了兵家必争之地，还是市场表现的结果，谁不想二三十万的车就配有之前只有 78S 系才能选装的配置呢？这种造车逻辑之下，成本的分配也就必然从内燃机变速箱往舒适性配置上调整了。智界 V9 的沙发，显然值得多点笔墨。双独立结构，通风加热和比气动按摩更强劲的机械按摩已经不足为奇。更关键的是，二排座椅有可旋转设计，最大旋转角度 180 度，可以在多种角度进行旋转和固定，营造了例如 45° 迎宾座椅，90° 观景座和 180° 旋转之后的后排会客室布局等多种模式和选项。尤其是迎宾形态和会客室形态，可以在商务场景里营造比普通 MPV 更优雅的氛围。出门在外，面子可以是自己给的，但归根到底，还是别人给的，给别人面子就是给自己涨面子。智界 V9 类似的优雅配置还有不少，比如说「手势控车」，佩戴支持功能的手表，做出轻捏动作就可以隔空打开车门；二排乘客在车内做一个推一推手势，二排侧滑门就可以优雅闭合；窗边手指划一划，就可以调节智能隐私光幕比例的明暗度……车内的四块智能隐私光幕玻璃最大遮光率可达 99.6%，如果一键开启隐私模式，可以立刻断开车内的位置信息、麦克风、摄像头、网络，保证车内隐私。新车配有超 50㎡ 静音材料覆盖和 13 块双层夹胶隔音玻璃进行被动降噪，同时还可选 RNC 主动降噪。其他一些老板不太关心，但司机需要关心的作为一款鸿蒙智行的车，其实很多关于底盘，辅助驾驶和动力的信息都能猜个大差不差，这里完全可以用列表形式一一列举，毕竟这些信息是给司机看的，而且司机一看就懂，无需解释。新车长宽高为 5359×2009×1879 mm，轴距 3250 mm，2+2+3 座椅布局，后备箱容积 892L（可扩容至 2002 L）。全系标配 38 个传感器，包括 896 线激光雷达，背靠 60 EFLOPS 云端 AI 算力、WEWA 2.0 架构，可首批升级华为乾崑智驾 ADS 5。搭载 1.5T 雪鸮增程器配合前后双电机，综合功率 380kW，搭载 53.4kWh 电池，CLTC 纯电续航 305km，综合续航 1320km。新车有辉光紫、深海蓝、雪域白、鎏金黑四种车身配色和韶华杏、赤茶橘、绒霞紫三种内饰颜色。搭载 MPV 版图灵平台，前双叉臂与后 H 臂多连杆独立悬架，配合双腔空气悬架及 CDC 连续可变阻尼减振器。全系标配±7° 主动后轮转向，5.35 米转弯半径，支持低速蟹行模式，有效应对狭窄路况的调头和穿行。四活塞前固定卡钳，35.8m 制动距离；500mm 涉水深度，保证暴雨积水情况下的通行能力。路面预瞄系统智能识别井盖、凹坑与弯道，智能调节悬架阻尼、刚度，以及电机扭矩，可以降低冲击与侧倾，获得莱茵晕车舒缓认证。行业首发 2 根 2400MPa+6 根 2200MPa 侧门防撞梁；全系标配 13 个安全气囊，包括全维包裹安全气囊。前环宇三联屏，12.3 英寸液晶仪表与双 17.2 英寸 3.4K 大屏组合，可选后排 21.4 英寸吸顶屏。因为鸿蒙座舱，乾崑智驾，还有图灵底盘等鸿蒙智行基础产品技术支撑，现在鸿蒙智行的车型想要出挑有点困难，但毫无疑问，无论是华为还是奇瑞，都给到了智界 V9 足够的重视，比如单独给这款车开一场声势浩大的发布会，这在整个华为体系里是极为罕见的待遇。再比如，华为和奇瑞给这款车投入了一些深刻的产品洞察去迎合和吸引智界 V9 的潜在受众，很多人可能对这样一款华为气质非常浓郁的 MPV 产品无感，但华为不在乎，华为只希望这些人快快长大并事业有成。

发布会开场，李想先是强调：其实大部分用户并不知道自己想要什么。他分享了一个故事：有一次我去买烤冷面，摊主问我还要加点什么，我开玩笑说能不能打个折。你看，用户永远想要更便宜，但摊主会告诉你，其实还能加蛋。这也是 L9 当年销量爆发的秘诀——当行业还在比拼基础参数的时候，理想顺手给车里塞进了冰箱、彩电和大沙发，交出了一份远超预期的答卷。四年过去了，全行业都开始把屏幕做大、座椅做软。当堆砌配置成为车企的常规操作，理想 L9 又走向了另一条路。如今的理想把大量精力和资金砸向底层架构，自研范围覆盖车规级算力芯片、800V 主动悬架、全线控底盘等多个领域。承载这些研发成果的，自然是今天上市的两款新车型：L9 Ultra 和 L9 Livis。首先值得肯定的是，理想没有盲目放大 L9，车长和车宽依然维持在 5.25 米和 2 米，日常开起来依旧方便。在这个好开的尺寸基础上，L9 的整车姿态还挺拔了不少。在车高仅增加 10 毫米的情况下，前机盖和水切线被大幅拉高，配上 22 寸大轮毂，原本庞大的车身显得颇为矫健；前唇那一抹金灰色镀铬饰条也收束得干净利落。L9 Livis 还采用了迈巴赫同款涂装工艺，紫金、绿金、黑金等专属双色车身，额外增加两道工序后，漆膜厚度达到传统工艺的两到三倍，日光下呈现出流动的光影质感。外观的考究还藏在细节里。全车 12 颗超声波雷达被 UWB 传感器取代，保险杠上再也看不到突兀的小圆孔；隐藏式电动踏板收起后与底盘融为一体。这不仅让外观更整洁，UWB 技术还让哨兵模式耗电量骤降 82%，开两天只耗一度电。打开带有电容防夹功能的电动车门，里面还是那个熟悉的「起居室」，只是每一项体验的上限都被抬高了。全新 L9 是全球首款四个座位都配备全尺寸零重力座椅的 SUV。Livis 版本直接标配，Ultra 版本在首销期内也会免费赠送前排两把零重力座椅。在过去，转向管柱挤占了空间，主驾想完全躺平是个大难题。而在 L9 上，线控转向释放了空间，1.8 米的大个子坐进主驾也能一键舒展躺平。望向前方，仪表台上的两块屏幕合二为一，变成了一块 29 英寸的 6K 超宽全景屏。后排的 21 英寸屏幕新增电动平移功能，手轻轻一抬，这块「神奇移动屏」就会沿滑轨移到面前，并随座椅倾角自动找平。音响功率这次飙升到了 5440 瓦。9.3.6 星环剧院音响搭载三重矩阵式低音系统，头枕音响创新采用水平对置设计，出声口藏在两侧，以往那种「有人在你脑后说话」的突兀感也随之消失。车内空调支持 7 温区控制，面部和脚部可以吹出不同温度的风；全车 29 处皮质区域都带加热功能。中控台下方的 10L 智能冰箱换了电动门，轻拍两下即可开关，不仅支持 -6°C 冷冻，内胆还能直接拆下来清洗。作为增程车，全新 L9 两个版本的纯电体验都做到了行业顶配。底盘塞进了一块 72.7 度的 5C 大电池，容量已经超过不少纯电车型，换来 420 公里纯电续航。按每天通勤 40 公里算，两周充一次电完全够用。充电速度也同步提升，配合理想自建的 8300 根 5C 超充桩，十分钟左右就能完成补能——在服务区上个洗手间、买杯咖啡的功夫，电量就能满血复活。长途出行时，这台重达 2.8 吨的全尺寸 SUV 的油耗还挺实在的：亏电油耗降至 6.3 升/百公里，满油满电综合续航达到 1650 公里。更让人省心的是，这套增程系统的保养周期延长到了 3 年或 3 万公里，还省去了过去增程车繁琐的维护。李想表示，在机械属性和维保频次上，全新 L9 已经越来越像一台纯电车了。把家装进车里，把纯电融进增程「全尺寸 SUV 一直有一个很难真正被解决的矛盾，既让一家人坐得舒适，又让自己开得尽兴。」李想说。为了同时满足这两点，L9 Livis 直接上了全套线控主动底盘。L9 Livis 的 800V 主动悬架，让每个车轮都有超过 1 万牛的支撑力，传统的防倾杆被取消，四轮完全独立。过减速带、连续起伏路或者高速变线时，液压泵和减震器会实时配合，该撑就撑、该压就压。停车休息时，这套悬架甚至能联动赛车游戏，把车厢变成 4D 模拟器。大车想开得灵活，转向系统也要升级。Ultra 和 Livis 都标配了后轮转向 + 线控转向，转弯半径缩到 5.3 米，方向盘打满不用换手，角度也能按习惯自定义。另外，Livis 还首发了 EMB 线控机械制动，刹车响应快了约 0.1 秒，120km/h 时速下刹停距离能缩短 3 米以上。而且这套系统号称终生不用换刹车油，后期省心不少。安全性上，两款车都变得更加坚固了，堡垒安全车身 3.0 用上了 2200 兆帕的高强度钢和一体式双门环。最特别的是，车内的 11 个安全气囊还能通过红外传感器识别每一位乘员的体型与坐姿，动态调节保护策略。面对极端的碰撞危险，具备主动悬架的 L9 Livis 能在 0.7 秒内瞬间抬升 50 毫米，用最坚固的底盘部位去硬刚撞击；低速倒车时的防碰撞系统也具备了识别悬空物体的能力。智能化层面，理想首发了马赫 M100——一颗专为 AI 原生设计、采用数据流架构的端侧推理芯片。L9 Ultra 版本用的是单芯片方案，算力达到 1280TOPS，配备前向激光雷达；L9 Livis 则拥有两颗马赫 M100，总算力翻倍达到 2560TOPS，同时配备了 4 颗激光雷达来实现 360° 全向感知。算力上来后，感知帧率成倍提升，乡间小路、施工路段都能提前预判，果断绕行。这颗芯片还跟去年 12 月发布的 Livis AI 眼镜实现了生态互联。交互开始变得科幻。戴着眼镜站在车外，一个手势或一个眼神，车就会自动泊出来，缓缓开到你面前。坐进车里，系统还能察觉你的情绪和状态。你今天好像心情特别好，是要带家人去顺义的公园踏青吗？这时候，它既是司机，也是伙伴。到了我们最关心的价格，两款车只拉开了 5 万元的价差。L9 Ultra 定价为 45.98 万元，竞争力很强。72.7 度 5C 电池、双腔空悬、线控 + 后轮转向、1280 TOPS 马赫芯片都是标配，首销期还免费送前排零重力座椅、智能电开门和 5440 瓦音响。L9 Livis 定价来到 50.98 万元，比此前 55.98 万的预售价低了不少。多花 5 万，换来的是专属双色车身、800V 主动悬架、带 EMB 制动的全线控底盘、翻倍到 2560 TOPS 的算力，以及 4 颗激光雷达。首销权益也公布了：6 月 30 日前锁单，现金直减 2 万元，老用户再额外补贴 1 万元。从 2015 年创立至今，理想还是习惯把消费者叫作用户。买车只是陪伴的开始，让全家人在日复一日的相处里都能开心，才是 L9 Livis 这台「硅基生命体」想要兑现的承诺。

在今天和理想全新 L9 和智界 V9 两台重磅车型同台亮相的，还有蔚来的乐道 L80。相比 L90 面向三代同堂、多孩家庭的大三排需求，L80 把重点放在大五座市场，带来了更大的乘坐空间，更灵活的装载能力，以及更适合长途、自驾、露营等场景的空间组织方式。早年间，以本田冠道、大众途观 L 为代表的燃油大五座 SUV，主要通过释放第二排的腿部空间来提升乘坐舒适性；随后，随着新能源时代的到来，诸如理想 L7、问界 M7 等车型进一步强化了智能座舱的沉浸式体验和车内的储物能力。而乐道 L80 现在则试图在行业已有的基础上更进一步，将前备舱、后备舱、乘员舱以及补能体系整合为一个完整的生态闭环，通过蔚来的工程创新能力，让中大型五座 SUV 的日常使用体验再上一层楼。乐道 L80 一共有 Pro、Max+ 和 Ultra+三个版本，整车购买价格分别为 24.28万元、25.98 万元、27.98 万元， Baas 方案价格为 15.68 万元、17.38 万元和 19.38 万元。前备舱大，后备舱更大空间，依然是乐道 L80 最核心的竞争力。乐道 L80 车长 5145mm，轴距 3110mm，理想 L7 车长 5050mm，轴距 3005mm，乐道 L80 比理想 L7 大了一圈。取消第三排之后，乐道 L80 将所有空间都留给了前两排和储物。第二排作为 L80 整台车体验的核心区域，其后排头部空间达到 1074 毫米，腿部空间宽达 1121 毫米，坐垫宽度和长度分别达到 549 毫米和 497 毫米，靠背支持最大 135 度的后倾调节。L80 提出了一个「大五座，享四座」的越级体验概念，即当车内只有四人出行时，第二排的两位乘客完全可以获得媲美大型 MPV 独立航空座椅的放松状态。而在储物空间上，乐道 L80 配备了容积为 240 升的电动前备舱，能够轻松容纳两个 20 英寸的标准登机箱、一个儿童专属行李箱及一个随身背包。前备舱的存在不仅增加了绝对装载量，更为家庭用户提供了干湿分离与物理分区的储物区域，有效改善了传统 SUV 后备舱容易被杂物堆砌得杂乱无章的痛点。与前舱相对应，乐道 L80 的后备舱在第二排座椅处于常规直立状态时，容积为 1200 升；当第二排座椅完全放倒后，整体空间可扩展至 2600 升。此外，在后备舱的常规盖板下方，工程团队设计了双沉舱结构。其中外侧沉舱容积为 106 升，里侧沉舱容积为 63 升，均配备了独立盖板。这种隐蔽式的储物空间非常适合用来存放不常用的随车工具、零散的露营小配件、充电线缆以及各类清洁用品，使得主后备舱始终保持整洁。乐道能够在车辆内部极其有限的机械空间内挤出更多的储物容积，主要得益于车辆底层平台的极高集成度。L80 充分应用了已经在 L90 上被验证过的 900V 高压架构，并配备了体积更为紧凑的前置电驱动系统。结合 30.5 度大倾角冷却模块、全栈自研的热管理高度集成模块以及一体式偏置空调箱，这些系统级的技术创新对原本需要占据前舱大量体积的机械与电子部件进行了极致的压缩与整合。空间之外，座舱内的舒适化和智能化，也是乐道 L80 强调的特色。乐道 L80 的座舱内部采用了分层布局：上层主要用于视觉延伸与氛围营造，中层集中处理人机交互与触觉反馈，下层则侧重于日常储物与抗污耐用。车内配备了支持 1600 万色调节的氛围灯系统、长达 7.2 米的环绕光带，辅以大面积软质材料包覆、金属扬声器网罩及精细的绗缝与刺绣工艺。整体风格规避了盲目堆砌昂贵材料的浮夸感，更侧重于在家庭用车的基调上打磨出温馨且不失格调的精致质感。为了提升车内的通透感，乐道 L80 车辆侧窗与前挡风玻璃的总面积约为 3.3 平方米，前后分区的全景天幕面积也达到了 1.1 平方米。同时为了应对大面积玻璃材质在强光照射下带来的隔热与防晒问题，车内配备了双电动物理遮阳帘，乐道 L80 同时配备了空气悬架系统，支持在驻车时自动降低车身高度，配合电动侧踏板与静音电吸门，优化了行动不便的老人或儿童的上下车体验。车辆的座椅总成基于专属的高端座椅平台开发，依据中国人体生理曲线大数据进行了型面贴合度优化，前排座椅支持 12 向电动调节，第二排支持大角度电动调节。同时，方向盘加热、前后排座椅加热、通风以及按摩功能均作为标配提供。此外，车内集成了一台 8.8 升容量的智能电动冰箱，足以容纳 6 罐常规可乐、2 盒鲜牛奶以及部分水果零食。神玑上车，4D 领航数字化交互与计算平台是现代智能汽车的核心大脑。L80 的屏幕矩阵由一块 35 英寸的 AR-HUD 增强现实抬头显示、一块 17.2 英寸的 3K 高清中控屏、一块 17.3 英寸的后排吸顶娱乐屏以及一块 8 英寸的后排专属场景控制屏组成。支撑这一系列复杂交互与协同调度的硬件，是一颗采用 5 纳米制程工艺的车规级高阶智驾芯片——神玑 NX9031。该芯片拥有高达 546 GB/s 的内存带宽，打破了传统分散式域控制器的物理壁垒，能够同时支持智能驾驶、智能座舱、智能底盘等多个系统跨域计算资源的统一调度。配合 SkyOS·天枢整车全域操作系统，车辆从底层代码级别打通了智驾、座舱、底盘、车身及动力网络，使各项功能摆脱了信息孤岛的状态，能够基于场景进行深度系统级协同。在智能驾驶领域，系统深度接入了蔚来世界模型 NWM（NIO World Model）。该模型摒弃了传统的规则驱动逻辑，转而通过强化学习与海量的闭环仿真训练，提升了系统对复杂未知路况的理解与博弈能力，全面覆盖城区复杂道路、城市快速路、高速公路及各类园区停车场。值得重点关注的是，L80 支持城区领航换电功能（PSP）。在导航状态下，系统能够自动规划最优补能路径，将车辆自主驾驶至换电站并完成自动化换电流程。同时，基于模型化架构构建的智能泊车系统，也支持路线的实时动态规划和对各类非标准复杂车位的识别入位。尽管定位为体型庞大的家庭 SUV，L80 在底盘的机械素质与动态调校上依然投入了极高的规格。悬架系统采用了前双叉臂与后五连杆独立悬架组合，配备了具备 100 毫米行程调节能力的智能空气悬架系统以及 CDC 智能连续可变阻尼减震器。底盘硬件与路面之间通过液压衬套进行柔性连接，并搭配了抓地力更强的鸳鸯胎。乐道 L80 还支持 4D 全域舒适领航功能。该功能能够实时结合车辆自身的传感器感知、激光雷达对前方路面的三维预瞄以及云端高精舒适图层数据，在遇到减速带、坑洼或颠簸路面之前，提前毫秒级主动调节空气弹簧的刚度与减震器的阻尼，从容化解路面冲击。除了常规的运动、舒适、后排舒享和节能模式外，车辆预设了湿地、雪地、泥地、沙地等脱困模式。在卓越的底盘参数加持下，L80 拥有高达 600 毫米的最大涉水深度，最小离地间隙最高可拉升至 260 毫米，足以轻松应对绝大多数轻越野和长途自驾游的复杂路况。而在安全方面，乐道 L80 全系标配 9 个安全气囊，除了常规的环绕式防护，第一排还配备了能在碰撞发生前预紧的主动式安全带、容积达到 137 升的狐耳式超大副驾气囊以及防止前排乘员相互碰撞的中央远端安全气囊。基于家庭属性的考量，第二排乘客的安全防护被提升至与前排同等重要的层级，配备了 14 升容积的后排侧气囊，以及长度贯穿 2920 毫米、容积高达 56 升的超大侧面安全气帘。在应对碰撞的白车身结构设计上，L80 的前部预留了长达 650 毫米的充足溃缩吸能空间，并创新性地采用了「6 纵 10 横」的笼式安全传导结构。前防撞梁选用了高强度的 7 系航空级铝合金材质，其上层主防撞梁的宽度覆盖率达到了整车宽度的 80%，能够在严苛的偏置碰撞中有效化解破坏性冲击力。侧面防护则通过环舱高强度钢、一体式门环、多腔体铝合金门槛梁以及坚固的电池舱横梁共同构建。车尾部分应用了先进的一体压铸后地板技术，配合铝合金后吸能盒与后防撞梁，为车身尾部提供了深达 1244 毫米的吸能缓冲空间，最大程度减轻高速追尾事故对后排乘客及底层电池包的结构性伤害。乐道 L80 搭载的电机能够输出 340kW 的峰值功率，百公里加速时间 5.7 秒，其百公里综合电耗被控制在优秀的 14.3kWh，85 度电池下的 CLTC 续航为 615km。当然，纯电车型的体验永远无法脱离补能体系的建设。共享蔚来庞大且成熟的充换电网络，依然是乐道 L80 在同级别竞品中最难以被复制的核心优势。为了进一步夯实这一护城河，乐道官方已确认追加投入 6 亿元人民币，用于在网络中新增 8000 多块流通电池。蔚来之前透露，早在四年前，乐道 L80 便与 L90 基于同一技术平台同步启动开发。其中，乐道 L90 的成功已经验证了产品定义的精准性。在 2025 年，乐道 L90 凭借对大三排 SUV「满载状态下行李装载空间」这一核心痛点的有效解决，在上市短短 5 个月内实现交付超 4.3 万台，成为纯电大型 SUV 市场的年度销量冠军。这不仅助力乐道品牌在 2025 年实现 107808 台的全年交付目标，更以接近 24 万元的成交均价稳居新势力品牌第一梯队。乐道 L90 和蔚来 ES8 的表现，也被行业视为大型 SUV 动力形式从增程向纯电转型的关键标志。如果说乐道 L90 成功开拓了大三排纯电 SUV 市场，那么乐道 L80 则肩负着在规模更大的大五座 SUV 领域实现突破的重任。在乐道的产品价值体系中，L90 聚焦于「在一起」的温情与责任，通过三排布局满足多孩、三代同堂家庭的聚会与通勤需求；而 L80 则更侧重于「走出去」的激情与成长，以更灵活的空间形态承载用户自驾、露营等更具探索性的生活方式。

🔍郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%🤖OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略🚀Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI📱微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆💻华为 Mate X7 官方降价 1000 元⚠️追觅造车主体约 232 万元股权遭冻结🎮国行 Switch 正式停服🧠百度成立模型委员会 BMC💰三星电子单季营收创历史纪录，半导体撑起超九成营业利润🔍海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%🖨️铠侠：2027 年内存仍吃紧、净利润同比增长 103.6%🤖微信发布「全球青少年 AI+ 小程序」洞察报告💡何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了🦿5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约📱荣耀 CEO 李健官宣机器人手机今年 Q3 发布📱vivo S60 将首发 4K 原生感 Live🚗全新理想 L9 发布：45.98 万元起🤖ChatGPT 新增 Codex 连接电脑功能📰 周末也值得一看的新闻郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%天风国际证券分析师郭明錤昨日披露，苹果已在 Intel 18A-P 系列制程上为低阶及旧款 iPhone、iPad 与 Mac 处理器立项，订单中 iPhone 约占 80%。流片规划显示，苹果预计 2027 年在 Intel 放量，2028 年续增，2029 年进入衰退期。Intel 方面，2027 年良率目标为稳定达到 50% 至 60% 以上，量产时程与出货规模目前仍未明朗。不过，郭明錤也指出，即便 Intel 顺利出货，台积电在苹果供应体系中仍将占 90% 以上比重。郭明錤认为，苹果此举是有意将 Intel 培养为长期备选供应商。AI 对台积电先进制程资源的占用将持续扩大，苹果需要在议价能力尚存时提前布局。与此同时，苹果也在同步评估 Intel 其他先进制程技术。OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略据 WIRED 报道，OpenAI 昨日向员工宣布新一轮组织调整，联合创始人兼总裁 Greg Brockman 正式接管产品战略。报道指出，OpenAI 将 ChatGPT、编程智能体 Codex 及面向开发者的 API 业务整合进同一核心产品团队，目标是将三者合并为「统一的体验」。Codex 原负责人 Thibault Sottiaux 升任核心产品与平台主管，统管消费、企业和开发者三条线；ChatGPT 原负责人 Nick Turley 转任企业产品改版；前 Instagram 副总裁 Ashley Alexander 接任消费产品负责人。Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI据《金融时报》报道，Anthropic 已就一轮 300 亿美元的融资达成条款，融资完成后公司估值将约达 9000 亿美元，较今年 2 月初 3500 亿美元的估值接近翻三倍。本轮融资由 Dragoneer Investment Group、Greenoaks Capital、红杉资本（Sequoia Capital）和 Altimeter Capital 联合领投，报道称，每家领投机构预计出资不低于 20 亿美元。微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆CCF 计算机博物馆（CCM）昨日收到腾讯微信团队捐赠的一批早期历史实物，包括微信第一行后台代码原始存档、早期开发设备 iMac ME086CH/21 寸一体机与戴尔塔式工作站，以及 Apple iPod Touch 3、诺基亚 5800XM 等早期测试终端。CCM 表示，由于移动产品迭代速度极快、数字资料易被覆盖删除，大量见证移动互联网起步的实物未能得到系统保护。此次捐赠呈现了微信从 0 到 1 的研发链条，是中国移动互联网早期发展的核心实物见证。CCM 主体建筑已全面落成，拟于今年 9 月开馆，目前持续面向社会征集各时期计算机相关藏品。保时捷称被中国车企抄袭「有些许无奈」，罗永浩：保时捷自己就有前科保时捷中国传媒公关副总裁唐风靓日前在浙江「易境思论坛」上表示，过去几年保时捷激发了高度统一的汽车设计审美，被中国车企模仿「除了荣幸也有些许无奈」。罗永浩随后在微博发文回应，称保时捷说「无奈」可以理解，但称中国车企抄袭得「高度统一」则是吹牛。他同时指出，保时捷历史上也曾抄袭捷克斯洛伐克品牌太脱拉（Tatra）的设计，并以 1965 年大众汽车向太脱拉支付 100 万西德马克庭外和解赔偿金为据，直言「说到流氓程度，谁能跟保时捷比」。此番言论随即引发讨论。有网友指其犯了「诉诸伪善」的逻辑错误，罗永浩强硬回怼，称自己已明确承认抄袭者「也是流氓」，并非为抄袭行为开脱。华为 Mate X7 官方降价 1000 元华为昨日宣布，旗下两款折叠屏手机同步开启降价优惠：Mate X7 全系直降 1000 元，优惠后售价 11999 元起；Mate X6 直降 3000 元，优惠后售价 9999 元起。追觅造车主体约 232 万元股权遭冻结企查查 App 显示，追觅科技创始人俞浩实控的星空计划（上海）汽车科技有限责任公司近日新增一则股权冻结信息，股权数额约为 232.25 万元，股权所在企业为上海星空计划进出口贸易有限责任公司。去年 8 月，追觅科技宣布进入汽车领域。截至目前，追觅汽车团队规模约千人，研发人员占比达 70%，自研覆盖智能汽车多个核心领域。此次造车主体出现股权冻结，官方尚未就相关情况作出回应。国行 Switch 正式停服国行 Switch 的 Nintendo e 商店已于昨晚正式停服，至此，国行 Switch 将无法再进行联网对战或下载游戏。官方在停服前向相关玩家发送提醒，建议玩家在停服前完成所有已购游戏、DLC 及兑换码的下载与激活；删除游戏数据或对主机进行格式化后，数字内容将无法恢复。今年 3 月 31 日，国行 Switch 的游戏售卖服务已率先停止；而根据官方此前公告，今年 12 月 31 日起，国行 Switch 主机及配件的维修服务也将全面终止。这意味着，国行 Switch 的官方售后支持将在今年底彻底画上句号。百度成立模型委员会 BMC据《科创板日报》报道，百度近日内部设立百度模型委员会（Baidu Model Committee，BMC），成员由对大模型有深刻理解的年轻研究员构成。基础模型研发部（BMU）与应用模型研发部（AMU）均向该委员会汇报。知情人士透露，新成立的 BMC 将全面统筹百度大模型业务，覆盖路径探索、技术规划、模型训练到业务落地的完整链条，目标是实现模型从技术到应用的一体化管理。百度方面表示，此举有助于集中优势资源，持续建设并强化人工智能方向的技术优势。三星电子单季营收创历史纪录，半导体撑起超九成营业利润三星电子昨日发布 2026 财年第一季度（1 月 1 日至 3 月 31 日）业绩报告：期内实现合并营收 133.9 万亿韩元，同比增长 69%，环比增长 43%；营业利润 57.2 万亿韩元，同比增长 756%，环比增长 185%，营业利润率达 42.8%。净利润为 47.2 万亿韩元，净利率为 35.3%。基本每股收益为 7123 韩元。半导体（DS）：营收 81.7 万亿韩元，同比增长 225%，环比增长 86%；营业利润 53.7 万亿韩元。其中存储业务营收 74.8 万亿韩元，同比增长 292%，环比增长 101%，为本季度核心增长引擎；设备体验（DX，含手机与消费电子）：营收 52.7 万亿韩元，同比增长 2%；营业利润 3.0 万亿韩元，同比下降约 1.8 万亿韩元；显示（SDC）：营收 6.7 万亿韩元，同比增长 14%；营业利润 0.4 万亿韩元，同比下降 0.1 万亿韩元；Harman：营收 3.8 万亿韩元，同比增长 12%；营业利润 0.2 万亿韩元。下季度展望，公司预计 AI 基础设施扩张驱动的服务器内存及 SSD 需求将持续强劲，计划向客户交付 HBM4E 首批样品以巩固技术领先地位；手机业务方面，Galaxy S26 系列发布效应趋于消退，盈利能力预计承压。海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%昨天，海力士于发布今年第一季度（截至 3 月 31 日）业绩，按 K-IFRS 口径，期内实现营收 52.58 万亿韩元，同比增长 198%；营业利润 37.61 万亿韩元，同比增长 405%；净利润 40.35 万亿韩元，同比增长 398%。本季度营收首次单季突破 50 万亿韩元，创历史新高。主要财务指标如下：毛利润 41.68 万亿韩元，毛利率 79%，同比提升 22 个百分点；营业利润率 72%，创历史单季新高，环比提升 14 个百分点；EBITDA 41.34 万亿韩元，EBITDA 利润率 79%，同比增长 284%；净利润率 77%，其中非经营利润 14.01 万亿韩元，包含投资资产估值收益 9.94 万亿韩元。铠侠：2027 年内存仍吃紧、净利润净同比增长 103.6%铠侠控股昨日发布截至今年 3 月 31 日止 2026 财年业绩，期内实现营收 2.3376 万亿日元，同比增长 37.0%；按 IFRS 口径，归母净利润为 5544.9 亿日元，同比增长 103.6%。SSD 与存储业务营收 1.3626 万亿日元，同比增加 37.5% 至 3715 亿日元；智能设备业务营收 7599.8 亿日元，同比增加 51.6% 至 2588 亿日元；其他业务（含 SD 卡、U 盘等零售产品及向闪迪集团的销售）营收 2150.1 亿日元，同比基本持平。另据财联社报道，铠侠表示，AI 服务器客户的订单量持续走强，以美元计价的 NAND 价格在一季度上涨逾一倍，并预计 2027 年度 NAND 存储供应仍将维持紧张态势。微信发布「全球青少年 AI+ 小程序」洞察报告昨天，腾讯研究院联合微信发布《全球青少年 AI+小程序洞察报告》，披露了全球基础教育阶段群体利用 AI 工具与小程序进行应用开发的各项核心数据。据报告数据显示，自 2023 年 11 月微信小程序教育平台正式对外开放以来，已吸引全球 7954 所学校的逾 8.7 万名学生参与，累计创建小程序项目超 28.7 万个。在开发流程中，师生调用 AI 辅助创作每年消耗 Token 量超过 500 亿，支撑约 375 万次人机交互。此外，微信连续四年举办的小程序全球创新挑战赛共产生 7430 个作品，其中第四届单届提交量达 4900 个，呈显著爆发趋势。开发者画像层面，62% 的项目由单人独立完成，最小获奖开发者年仅 5 岁。地域活跃度数据显示，中国香港、北京及中国澳门位列前三，且项目协作已出现跨国别开发特征。在应用方向上，05 后开发者主要聚焦于无障碍辅助、碳足迹、适老化关怀及健康管理等社会实用场景。💡 何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了据睿见 Economy 报道，小鹏集团董事长、CEO 何小鹏昨日在「第十八届轩辕汽车蓝皮书论坛」上表示，「什么时候我们看到北京、上海车展上不再有 150 款新车的时候，中国汽车的内卷可能就结束了。」何小鹏表示，过去十年，中国新能源汽车经历了两个重要转变：一是核心驱动能源从传统化石能源切换至电化学能源；二是行业全面进入初步普及的智能化阶段。他称，十年前几乎没有人相信中国汽车能在全球市场取得如今的成绩，并实现规模化的全球化布局。何小鹏认为，下一个十年将是电化学能源与数据能源的双重叠加阶段，芯片和模型将成为新的「引擎」，汽车与机器人的能力将在这十年内深度整合。5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约ROG 与 XREAL 联名推出的电竞 AR 眼

🤖曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini🌐马耳他将向全体公民免费提供一年 ChatGPT Plus📱小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务🚀目标估值 1.75 万亿美元，SpaceX 最快 6 月上市📋arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚💻卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元🚗蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌🔍需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位☁️最低 1 元 40 万 Token，三大运营商推出 Token 套餐🚙特斯拉 Robotaxi 碰撞事故报告解封💡知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值🚗小米 YU7 GT 火山灰配色公布💰ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构🧠马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布🤖字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本📱腾讯 vibe coding 产品「吐司」即将上线🛍️蜜雪冰城上架老黄同款蜜桃四季春，销量暴涨🛑中国银行关停信用卡独立 App，服务整合进主 App曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini昨晚，彭博社记者 Mark Gurman 在最新的 Power On 新闻通讯中透露了更多关于苹果即将发布的独立 Siri 应用的细节。隐私设计是此次新应用的核心卖点之一。Gurman 透露，该应用将内置对话自动删除功能，用户可在设置中选择对话保留时长：30 天、1 年或永久保存。苹果预计在今年的 WWDC 继续主打隐私保护，并强调其 AI 体验不含广告。模型方面，苹果已将新 Siri 的 AI 模型底层大量替换为 Google 的 Gemini，并将部分服务运行在 Google 的云基础设施之上。而此前有消息称，苹果和原本的合作对象 OpenAI 关系出现裂痕，甚至有可能对簿公堂，原因是 OpenAI 认为双方合作的结果不及预期。值得注意的是，尽管新的 AI Siri 已经跳票约两年，Gurman 称苹果内部测试版本仍将其标注为「Beta」，并可能沿用至其正式发布。马耳他将向全体公民免费提供一年 ChatGPT PlusOpenAI 日前宣布，与马耳他政府签署合作协议，向该国公民及居民免费提供为期一年的 ChatGPT Plus 订阅服务。据悉，居民须先完成由马耳他大学开发的免费在线课程「AI for All」，课程内容涵盖 AI 的基本概念、能力边界及负责任使用方式，通过后即可向马耳他数字创新局申请免费订阅资格。小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务小米 REDMI 产品经理「笋寸」日前在微博发文建议，近期不打算换手机的用户可以考虑等待换电池，并提示部分机型在更换新电池后容量可能还会有所提升，参考小米 13 系列的现行政策。笋寸进一步说明，小米正积极推动存量用户更换采用金沙江技术的大容量电池，小米 13 系列已率先上线相关服务。她还透露，具体支持机型「得等等」。此前，小米 13 系列已全系上线电池升级服务，支持小米 13、小米 13 Pro、小米 13 Ultra 三款机型升级更大容量电池，费用为 189 元，其中电池升级费 149 元、人工费 40 元。目标估值 1.75 万亿美元，SpaceX 最快 6 月上市据路透社报道，SpaceX 已加快 IPO 时间表，计划最早于 6 月 12 日在纳斯达克挂牌上市，股票代码为「SPCX」，较此前预定的 6 月下旬提前约两周。据三位知情人士透露，公司计划最早于本周三公开招股说明书，路演定于 6 月 4 日启动，股票发行最早于 6 月 11 日完成定价。时间表提前的主要原因是 SEC 对招股材料的审核进展快于预期。此次 IPO 预计募资约 750 亿美元，目标估值约 1.75 万亿美元，有望成为史上规模最大的股票发行。承销行阵容包括摩根士丹利、美国银行、花旗集团、摩根大通和高盛。arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚近日，预印本学术论文平台 arXiv 宣布，若作者提交的论文中包含无可辩驳的 AI 生成内容证据，将被封禁账号一年，并在此后要求所有提交须先经过同行评审期刊接受方可上传。arXiv 计算机科学板块主席 Thomas Dietterich 在 X 发文表示，若论文中出现「幻觉引用」、大语言模型元注释（如「以下是 200 字摘要，是否需要修改？」）等明显迹象，即可认定作者未对 AI 生成内容进行核查。arXiv 目前由康奈尔科技学院管理，今年 7 月将独立运营，成为一家非营利法人机构。康奈尔科技学院院长 Greg Morrisett 表示，独立后 arXiv 将能从更广泛的捐助方获得资金支持，以应对「AI 垃圾内容」带来的运营压力。卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元近日，小米集团合伙人、总裁卢伟冰在「Max 一夏」户外露营直播中表示，玄戒芯片今年肯定会推出迭代款，并称这将是「一块非常强的芯片」，会搭载在一款「非常优秀的产品」上。他同时提醒外界，目前网络上流传的各类相关传闻可信度不高，具体规格暂不便披露。卢伟冰还透露新一代玄戒芯片的应用范围将不再局限于智能手机，后续计划广泛搭载于小米旗下多类智能终端设备。关于手机定价，卢伟冰在直播中作出判断，认为今年下半年尤其到年底，部分国产旗舰直板手机价格可能突破 1 万元人民币。他预计，内存涨价的压力至少延续至 2027 年底，甚至到 2028 年。对于即将发布的小米 17 Max，他表示定价仍在讨论中，将在发布会上正式公布，但小米的原则是在同档位提供「最物超所值的产品」。蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌据睿见 Economy 报道，在日前举行的第十八届轩辕汽车蓝皮书论坛上，蔚来创始人、董事长兼 CEO 李斌线上出席，在对话中将萤火虫比作 iPhone，称其已成为「一款非常高品质的产品」。他强调，萤火虫在某种意义上超越了汽车的工具属性，演变为「一个生活方式品牌、一个时尚品牌」，并开始形成独特的品牌文化。他表示，萤火虫在用户群体上打破了年龄与购车预算的双重限制，「大学刚毕业的年轻人想买也能买，很有钱的老板想自己开车送小孩上学也完全没有问题」。需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位据《商业内幕》报道，前向部署工程师（FDE）的招聘需求在一年内激增逾 700%，正成为 AI 浪潮下科技行业增长最快的职位之一。招聘平台 Indeed 数据显示，今年 4 月该职位招聘帖数量达 5330 条，较去年同期的 643 条增长约 729%。Anthropic、OpenAI、Palantir、Stripe 及 Google Cloud 均在积极招募该职位人才。FDE 的核心职责是协助企业将 AI 整合进业务流程，并根据客户需求定制开发 AI 工具。薪酬方面，Indeed 数据显示该职位年薪约为 17 万至逾 20 万美元。最低 1 元 40 万 Token，三大运营商推出 Token 套餐中国电信昨日推出系列试商用 Token 套餐，分别面向开发者及中小微企业、个人及家庭用户提供「Token+连接+安全」一体化服务，均提供三档套餐，面向个人及家庭用户定价 9.9 元/月起 1000 万 Token；面向开发者及中小微企业 39.9 元/月，1500 万 Token。两类用户均可选购宽带上行提速包（上行速率提升 50 Mbps，100 元/月）及智能体安全管家服务（5 元/月）。个人家庭用户还可额外选购 5G-A 优享包（含 10 GB 流量，上行速率最高 300 Mbps，20 元/月）。联通、移动两大运营商也先后推出了各自的 Token 套餐：移动：中国移动上海公司于 5 月 17 日世界电信日发布会上推出 Token 通用服务，定价 1 元 40 万 Token。联通：中国联通上海分公司于 5 月 16 日发布全栈国产化「国芯国模国云」安全基座，推出多档 Token 产品及融合套餐。特斯拉 Robotaxi 碰撞事故报告解封近期，特斯拉正式解密并公开了向美国国家公路交通安全管理局（NHTSA）提交的 17 起「Robotaxi」自动驾驶碰撞事故的完整未脱敏报告。据 Electrek 报道，这 17 起事故发生于 2025 年 7 月至 2026 年 3 月，均涉及在得克萨斯州奥斯汀进行「Robotaxi」测试的 2026 款 Model Y，且测试时均有安全员在场。伤亡数据显示事故整体较轻微：13 起仅造成财产损失，2 起无人员受伤，1 起轻伤未入院，另有 1 起轻伤需入院治疗（系车辆在路口以 2 英里/小时蠕行时被后方 SUV 追尾）。官方报告表明，大部分事故并非特斯拉自动驾驶系统（ADS）的过失，多为车辆在红绿灯、停止标志或交通拥堵中完全静止时，被后方人类驾驶员追尾或侧擦。但数据同时也暴露出系统与后台干预的缺陷：两起事故发生在远程操作员接管期间，导致车辆分别以 8 英里和 9 英里的时速撞上金属围栏与施工路障；另有数起事故显示 ADS 存在空间感知死角，车辆在转弯或倒车时撞上了金属链条、拖车挂钩、木制电线杆及路缘石。💡 知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值据「听筒Tech」报道，知乎创始人兼 CEO 周源日前在第十二届新知青年大会上表示，AI 时代真实创作者的价值不可替代。AI 每分钟都在创造奇观，但 AI 时代仍然需要新知。而新知来自真实的人。每一个真实的人，都拥有不可替代的价值。大会主论坛上，多位嘉宾就 AI 与人类创作的关系展开讨论。复旦大学人文学者梁永安提出「科技有所为，如何让自己有所不为？需要人的尺度和人文属性」；作家张春称「不完美就是我们的完美」。周源在大会结尾总结道：「技术越发达，有温度、有深度的真人创作就越宝贵。」小米 YU7 GT 火山灰配色公布近期，小米汽车再度放出 YU7 GT 的预热图片，公布了新配色以及新车的内饰信息。其中小米 YU7 系列将加推一款名为「火山灰」的配色，官方介绍：灵感来自晨雾下的火山地貌，高纯度灰调，叠加纳米级矿物颗粒，呈现细腻而有层次的金属质感。同时新车内饰也迎来全新的「赛道红」：采用红黑撞色，提供专属刺绣。官方表示，小米 YU7 也会提供「火山灰」外饰配色，并且现在车厘子红的 YU7 GT 已经在部分小米汽车门店展示了。ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构近日，OpenAI 面向美国 ChatGPT Pro 用户推出了个人财务管理功能，支持通过 Plaid 连接超过 1.2 万家金融机构的银行账户和信用卡。连接账户后，ChatGPT 会自动读取用户的真实余额、消费流水和负债数据，生成一张覆盖投资、日常开销、订阅和近期账单的财务面板。用户提问时，模型可直接翻出账单历史，找出哪些类目容易超支，并给出具体的削减方案。此外，用户还可以告诉 ChatGPT 一些账外信息，比如「我在攒钱买车」或「我还欠父母一笔钱」，系统会将这些内容保存为财务记忆，用于后续对话。马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布马斯克昨日在 X 发帖披露，SpaceXAI 正在研发的 1.5T 参数 Grok 基础模型第 9 版（V9）已完成训练，预计 3 到 4 周内正式发布。现役公开版本 Grok 4.3 基于 0.5T 参数的 V8 基础模型，目前每隔数日迭代一次。V9 将参数量提升至 1.5T，马斯克称其为「重大升级」。在帖子的结尾，马斯克更是用用「banger」（王炸）一词来形容对这一版本的预期。字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本字节跳动 Seed 团队近日开源扩散

直到 2026 年五月，我们终于等到一整轮的影像旗舰登场完毕——小米做出了徕卡一瞬、vivo 打造了照片与视频兼具的 V 单、OPPO 在双长焦方案上带来了惊喜。在这轮旗舰更新潮的最后，华为提前将新一轮的 Pura 90 Pro Max 端上了桌。老实说，第一眼看硬件，心里是有落差的。去年 Pura 80 Ultra 上颇为惊艳的「一镜双目」结构不见了，取而代之的，是目前行业内最稳妥也最常见的大底长焦方案。不过，在细细体验后，我认为 Pura 90 Pro Max 藏了一个比纯拼硬件更有意思的新解法——XMAGE 智拍。全自动，都省心过去这几年，我系统性地梳理过华米 OV 的影像架构。虽然各家在硬件选择上各有理解，但大方向出奇一致，都在极其有限的机身空间里死磕传感器面积，比拼谁的光圈更大、进光量更多。在硬件上奋发图强理所应当。但在摄影中，不错的硬件基础仅仅是第一步。想要拍出一张好照片，构图与后期同样重要——构图保证一张照片在视觉上是否平衡、主体是否明确；而后期则决定了这张照片视觉质感和风格化。但这两者，向来都是具备门槛的手艺。而华为在 Pura 90 Pro Max 上，打算一口气解决这两个问题。熟悉他们产品线的朋友或许有印象，去年发布的 Pura 80 Ultra 上，华为试水了 AI 辅助构图。AI 辅助构图功能相对基础，在实际拍摄中主要起一个引导和提醒的作用。到了 Pura 90 Pro Max，这项技术完成了从「辅助」到「主导」的蜕变，并定名为「XMAGE 智拍」。剥开表层去探究 XMAGE 智拍的底层逻辑，是一套完全由端侧算力驱动的本地化影像流水线，将拍摄过程拆解为三个同时进行的模块——基于实时语义分割的自动构图、无视物理镜头限制的智能变焦，以及针对画面的色彩重塑。也就是说，在这个功能的加持下，Pura 90 Pro Max 可以自动识别主体、变焦构图，最后还排列出一系列适合当前场景的 XMAGE 风格供你选择。值得一提的是，XMAGE 智拍不需要网络连接，也无需将庞大的图像数据传回云端，全部基于端侧 AI 运转，实时识别画面里的主体轮廓、建筑线条以及光源分布状况，随后根据底层训练好的审美模型，给出一套包含最优构图、变焦和精细色彩方案的组合拳。繁复的专业参数在这里黑盒化，过去需要人脑去判断的构图和调色工作，全权交给了本地算法。除此之外，Pura 90 Pro Max 还顺势推出了 AI 姿势推荐，这个功能同样通过语义分割识别，可以理解画面中的了恩物、环境、姿势与背景，随后直接在取景框中勾勒出一个等比例的线框轮廓。拿手机的人只需按图索骥，引导模特贴合画面上的姿势即可。遇到不满意的动作能随时刷新，它甚至支持导入社交网络上保存的样片，一键提取动作精髓。听起来确实很厉害，但这套完全依赖端侧计算跑出来的流水线，究竟能不能应对现实世界里复杂多变的光线，还需要把它带到真正的街头去寻找答案。AI 有所能，也有所不能来到一家咖啡馆，墙上的多层光环灯饰引起了我的注意。举起手机，XMAGE 智拍自动变焦到 2×，将墙壁上的光环作为画面主体。这个选择可以说中规中矩，光环圆形处于画面中下方，整体构图稳定和谐。同时，操作台上颇具金属质感的咖啡机也是个不错的主体，XMAGE 智拍精准识别到我的拍摄意图，变焦到 192mm，保持咖啡机的中置。由于我想保留更为还原的室内光线与色彩，这张照片并没有套用 XMAGE 智拍推荐给我的任何色彩风格。用它拍了几组照片后，我认为 XMAGE 智拍是一个非常用户视角的功能——在传统的拍摄习惯里，手机焦段是一件需要精打细算的事。大脑会本能地去贴合物理镜头的原生焦段，生怕落在中间焦段会损失画质。但 XMAGE 智拍完全抛弃了执念。它不在乎用的是几倍变焦，转而完全为画面考量：当前这个画面，怎么裁切才最好看。哪怕不可避免地要损失一部分边缘像素，只要能让视觉中心突出、比例和谐，它就会果断动手。这种做法很实在，一切都在为了最终的出片服务。走出室外回望咖啡馆，绿植与橙色外墙交相呼应，XMAGE 智拍自动截取出色彩冲击力最强的部分，同时准确识别出咖啡馆建筑主体，两种大色块在画幅中占比相似，平衡而美观。再抬头，夏季树木繁茂，XMAGE 轻微放大焦段，让树枝形成天然的框架构图，广州塔正悬于画面正中间。在这组照片的拍摄中，我还察觉到了一个很有意思的技术细节。按下快门时，你会发现滤镜加载的速度，远大于构图改变的速度。这是因为滤镜往往只需识别画面后套用底层的色彩映射，而构图则需要端侧算力去实时识别画面里的建筑线条、光源分布和人物轮廓。计算量完全不在一个量级。更有意思的是，智拍套用滤镜的方式非常精细，除了 XMAGE 风格选择外，还会对调色盘做精细调整，确保风格对当前环境的适配度。当然，这套系统也有自己的脾气。在实际走街串巷时，XMAGE 智拍有时候会自动跳出来接管画面，有时候又毫无动静，需要你手动去点击唤醒。就算是手动唤醒的情况下，XMAGE 智拍也有几率出现风格已经选好，但当前画面无法找到最佳构图的情况。这种情况频繁出现在极其繁杂的环境中，由于元素过于混乱、信息量爆炸，算法的确很难从中剥离出合适的构图。当然，换个角度想想，也许此时的 1× 视角，就是当前环境的最佳构图了。万物皆可 AI，现在轮到摄影了185 年前，达盖尔发明银版摄影法，留住时间成为一种特权；2011 年，胶片巨头柯达传出破产消息，智能手机开始野蛮生长，把镜头塞进了普通人的口袋。回顾整部影像史，其实就是一部打破特权、技术走向大众的普惠史。到了今天，移动影像的狂飙突进，撞上了一堵无形的墙。手机镜头模组越做越大，与之对应的，是最近行业里一个颇具深意的传闻：受制于高昂成本与物理极限，下一代影像旗舰的「超大杯」，大多要面临停更了。这条路为什么走不通了？因为影像旗舰突飞猛进的背后，藏着一个被刻意回避的死结。决定照片质量的，永远是镜头后面的脑袋。普通人有懂美的眼睛，却跨不过光圈快门和构图比例的门槛。把一台堆满顶级硬件的手机，递给一个毫无基础的人，他按下的快门，大概率得到一张平平无奇的随手拍。顶级硬件如果只服务于一小撮懂摄影的人，无疑就会变成一种伪命题。在同行集体踩下刹车的时间点，华为提前把 Pura 90 Pro Max 端上了桌，而 XMAGE 智拍，就是破局的方法。在此之前，行业里已有不少 AI 落地，比如 vivo 让四季流转，OPPO 消除反光。只不过，它们大多是按下快门后的后期创意和修补。XMAGE 智拍往前跨了一大步，让算法成为整个拍摄动作的主导。这种做法谈不上是什么艺术层面的颠覆，但确实管用。降低门槛，让完全没有摄影基础的人，也能相对轻松地把一台旗舰机用出该有的样子。好的技术绝不该是一座孤岛。曾经的智能手机让人们拥有了「随时拍」的自由，现在的 AI 则进一步赋予了大众「拍得好」的能力。这或许是打破当前僵局的一剂良药，也是影像超大杯物尽其用的方法。

过去几年，20 万级新能源 SUV 市场像一条突然变宽的河道。一边是家庭用户对空间、续航和舒适性的刚性需求，另一边是智能化、性能化配置不断下放。过去很多只会出现在 30 万元甚至更高价位车型上的配置，如今开始像雨水一样往下渗透。激光雷达、空气悬架、大模型智驾、零重力座椅，这些原本带有「高配」意味的词，正在成为新一轮价格带竞争里的筹码。肩负着广汽集团高端化重任的昊铂品牌在今年广汽集团重振旗鼓的趋势下，也顺势重新梳理产品架构，昊铂 A800 旗舰轿车前段时间开始上市，用到了乾崑智驾和鸿蒙座舱来增强产品竞争力。紧接着，面向 20 万价位带的 SUV 昊铂 S600 也来了，提供纯电 660 Max、纯电 660 Max+、增程 230 Max、增程 215 四驱 Ultra+ 四个版本，官方预售价为 20.99 万至 22.99 万元，预售权益价为 18.89 万至 20.89 万元。新车定位为「新豪华智慧运动 SUV」，提供纯电和增程两种动力形式，车身尺寸为 5015×1933×1700mm，轴距 2936mm。相比于品牌初创时期的雄心勃勃，昊铂现在显然务实不少，它试图把昊铂 S600 做成一个配置密度很高的样本：性能、智驾、座舱、舒适、安全和权益，几乎都被塞进「满配」这个关键词里。昊铂 S600 请了 2 个代言人，一个代言「优雅」，一个代言「运动」过去提到中大型 SUV，很多人的第一反应是「稳」：空间要大，坐姿要高，后排要舒服，最好还能承担家庭出行里那些琐碎但真实的场景。比如周五晚上从公司直接出发，后备厢里塞进折叠椅、露营桌、孩子的滑板车和一箱零食；又或者长途返乡时，后排老人希望能把座椅放倒一点，孩子希望看动画，副驾则希望在高速上补一觉。对于这类用户来说，车首先不是一个「驾驶机器」，而是移动客厅、临时休息室，也是家庭生活的第二个容器。昊铂 S600 没有放弃这部分的诉求，或者说，这些配置已经在国产中高端 SUV 上成为一种基本素养，昊铂 S600 拥有 90% 得房率、750mm 二排腿部空间、5.316 平方米整车玻璃面积，以及 1292L 二排拓展空间；舒适配置上，同级唯一 18 点按摩零重力女王副驾、143° 躺角后排电调座椅、22 扬 ADiGO SOUND 音响、五块双层夹胶静音玻璃等配置也都被纳入其中。但昊铂 S600 又没有把自己完全包装成一台「沙发车」。它的另一面，是更强调姿态和操控的运动属性，所以昊铂又反复出现「跑山王」这个表达，显然昊铂希望把一台五米多长的大尺寸 SUV 做得不那么笨重，让它不只是适合城市通勤和家庭旅行，也能在山路、弯道、高速超车等场景中保留一定的驾驶参与感。从造型上看，昊铂 S600 采用的是一种前低后高的轿跑溜背姿态，超过 5 米的车长和近 3 米轴距提供了视觉体量，双层鸭尾、宽体车尾、20 英寸运动轮毂等设计元素，则试图削弱中大型 SUV 常见的厚重感，从侧面看它确实不像一般的家用大型 SUV 那样笨重方正，更像是专为年轻家庭所设计。所以这一次昊铂也请了 2 个代言人，英格兰传奇足球明星迈克尔·欧文作为「运动代言人」，知名演员陈数作为「优雅代言人」，两位代言人分别代表这一款车的一种特质，也是试图同时吸引家用取向和个人运动取向的潜在消费者。把四驱、高阶辅助驾驶、双腔空悬都放进 20 万级别在最高配的增程 215 四驱 Ultra+ 版本上，昊铂 S600 同时提供双电机四驱和双腔空气悬架。昊铂表示，这是行业首次在这一价位段同时放入双电机四驱和双腔空气悬架。另外，昊铂 S600 还全系标配前双叉臂后五连杆独立悬架、SDC 可变阻尼减振器和鹰爪系统 2.0，并提供雪地、湿滑、脱困三种路面辅助模式。一台大 SUV 在城市高架上并线，需要车身跟随方向盘动作，而不是像大船一样慢半拍；周末跑山时，连续弯道考验的是侧倾控制和车身收敛；雨天压过积水路面，轮胎抓地力和车身稳定系统要在驾驶者还没反应过来之前，先把车身姿态拉回来。在这些场景下，双腔空悬能够有效提升驾驶体验和驾驶信心。动力部分，昊铂 S600 提供纯电和增程双动力。其中增程四驱版综合功率达到 370kW，满电零百加速 4.3 秒，亏电零百加速 5.6 秒。广汽星源增程技术的持续发电功率为 85kW，官方还提到新车在亏电状态下可实现 165km/h 速度持续刷圈 2 小时，亏电极速 193km/h。增程车过去最容易被质疑的地方，在于「满电一条龙，亏电一条虫」，电量充足时像电车，电量不足后则容易出现动力衰减、油耗升高、发动机噪音变大等问题。昊铂 S600 对 85kW 持续发电功率和亏电性能的强调，其实是在回应这个痛点，一台增程 SUV 不能只在发布会数据里快，也要在长途、高速、爬坡和亏电场景下保持基本的体面。在智能化层面，昊铂 S600 的策略也比较直接：尽量把高阶配置做成全系标配。新车全系标配激光雷达、高通 SA8650 智驾芯片、27 个高精度感知单元，并搭载融合 Momenta R6 强化学习大模型的广汽星灵智行 ADiGO GSD 3.0，支持城区和高速快速路场景。资料还提到，昊铂 S600 下半年将首批升级 Momenta R7 强化学习世界模型。智能座舱方面，新车搭载高通 8295P 座舱芯片、17.3 英寸 3K 中控屏、27 英寸 W-HUD，并内置 11 个 AI 智能体。相比过去车机只负责导航、音乐和语音控制，现在的座舱正在变得更像一个移动终端：它不仅要响应指令，还要理解场景，甚至提前知道用户在车里可能需要什么。下班后坐进车里，导航自动给出回家路线，HUD 把车道级信息投到视线前方，孩子坐在后排看动画，车机根据车内温度和空气质量调整新风，座椅按摩开始工作……这些在新能源汽车上可以实现的体验，对于从 20 万元档位传统燃油车换购过来的用户来说，无疑是跃升式的。广汽集团有着庞大的用户基本盘，这些人群想要增购换购新能源车型的话，广汽自然希望用产品承接他们的需求。就像 AI 说的那样，我在这里，不偏不绕，不躲不逃，稳稳接住你。某种程度上来说，过去几年广汽集团新能源产品的建设，过于重视 B 端，对 C 端用户的吸引力不够，并没有「稳稳接住」用户的需求。对消费者来说，20 万级新能源 SUV 已经足够拥挤，智己 LS6、极氪 7X、小鹏 G7 等车型都在这个区间形成压力。昊铂 S600 要突围，关键命题并不是「它是不是拥有很多配置」，而是这些配置能否被组织成一种清晰的用车体验被感知到：城市里安静从容，山路上不显笨重，长途时不焦虑，家人坐进去愿意放松，驾驶者开起来也不觉得无聊。

龙虾之父一个月消耗 6030 亿 Token，总花费金额高达九百万人民币。移动联通电信，三大运营商都在推 Token 套餐，199 送千兆宽带还有 1 亿 Token，了解一下？从硅谷到国内大厂，Tokenmaxxing 成为公司的主流，谁消耗 Token 多，谁就是 AI 时代的好员工。00 后校友向母校捐赠 20 亿 Token，被网友调侃按 DeepSeek 5 元/亿 Token 计算，只要 100 元。▲图片来自新浪财经Token 在半年内完成了一次身份跃迁：从技术术语，到 KPI，到话费套餐，到捐赠货币。它成了 AI 时代的「度量衡」，唯一的问题是，没人说得清它到底在度量什么。我们自己买 Token，用公司的 Token，部署了一堆 Agent，代码、论文、周报都是 Token 烧出来的。而另一边是，大厂的员工由于 Token 消耗排行榜的原因，开始拿着公司的 Token 处理私事、玩游戏、开发数十个没什么用的子 Agent 来提升自己的排名。「回报」这件事很难量化，但「使用量」可以量化。于是所有人都选择了那个容易量化的东西。这不是 AI 时代的新问题，这是管理学的老病。用 AI 消灭狗屁工作的公司，正在制造新型狗屁工作亚马逊，那个裁员裁到大动脉，把自己的网站都变成 404 的小狗，最近又被爆出了新的「笑料」。原本被寄予厚望、用来消灭「狗屁工作」的 AI，最终却沦为制造新型「狗屁工作」的源泉。据《金融时报》报道，为了逼迫员工拥抱 AI，亚马逊搞出了一个极其复古的管理手段：「Token 消耗排行榜」，追踪每个员工的用量。公司强制要求超过 80% 的开发者每周必须完成 AI 使用指标，甚至将消耗 Token 的数量作为考核标准。▲图片

1.OpenAI 的两大宿敌 Anthropic 和马斯克，放下心中成见之后终于在月初结盟了。在此之前，Anthropic 和马斯克的关系并不融洽：今年 2 月，马斯克还在自己的 X 账号指责 A 社「woke」「邪恶」「反人类」（misanthropic），说这家公司「仇视文明」。事后来看，这次攻击并非马斯克清新脱俗的性格使然，而是 Anthropic 所做的某些事情触碰到他的神经，事出有因。在此之前，xAI 内部使用 Cursor 工作，但是今年年初员工发现，Claude 模型突然在 xAI 的 Cursor 公司账号里不能使用了。当时还在 xAI 上班的联合创始人吴宇怀，在全员信里是这么说的：「Anthropic 更新了政策，要求 Cursor 不得向其主要竞争对手提供 Claude 模型调用能力。」当时，吴宇怀在信中写了一句话，颇为有趣：「这是坏消息也是好消息。我们的生产力会被影响，但这也敦促我们开发自己的编码产品和模型。」为什么当时 xAI 的高层认为，开发自己的编码产品是关键？后来发生的事情，大家都知道了。xAI 的联创团队悉数跑路，马斯克一气之下对 Cursor 使用了钞能力必杀：上个月底，SpaceX 和 Cursor 共同宣布，将在编程和知识类工作 AI 模型的训练上，展开前所未有的战略合作；并且，SpaceX 还获得了以 600 亿美元收购 Cursor 的权利，或向后者支付 100 亿美元合作费用。注意编程这个关键定语，后面还会 call back.2.最近，我看了一条 Cursor 早期投资人、Anthropic 大喷子、T3 创始人 Theo Browne 的视频。本来点进去是看他喷 A 社和 SpaceX 怎么蝇营狗苟，结果没想到，却看到了关于 SpaceX + Cursor 合作的，一个既另类却又极度合理的分析：不说 600 亿的收购，就只说 100 亿的合作费——Theo 在视频里表示，自己认为「哪怕只是交换到 Cursor 的用户数据，这 100 亿也值回票价了。」所以是什么数据？如果你也去看 Theo 这条视频，他会讲得非常清楚。但为了节约时间，我们在这里简单概括一下：我们和 AI 的对话是一来一回的，你提出问题/需求，他给你解答；coding agent 同理，只不过返回的是代码。一次高质量的对话，整个过程，包括用户提示、模型思考、agent 规划、输出代码、验证——所有这些东西合起来，可以称为一个完整的 Agentic Loop——就成为了高价值的训练数据，再喂给模型去进行强化学习，就能进一步提高模型在实战场景下的表现水准。Cursor 有的，SpaceX 想要的，就是这些数据。可这些数据从哪里来呢？答案很简单：作为模型厂商，这种高质量数据的最直接来源，只能是你自己开发的 coding agent 产品——也即 Anthropic 的 Claude Code、OpenAI 的 Codex、Kimi 的 Kimi Code。现在你应该明白了，为什么被 Anthropic「封号」之后，吴宇怀会在全员信里提出开发 xAI 自己的 coding 产品和模型这件事了。这件事 xAI 在当时已经看清楚了：没有自己的编码产品，就没有高质量的强化学习数据；没有高质量的数据，就训练不出真正实战能力强的 coding 模型。虽然有点暴论，但现在我们可以点题了：模型厂商想做出来真正能打的编程模型，做自己的 coding agent 产品是唯一的路径。3.大语言模型像个水晶球，用全网的语料训练出来，似乎能够解答万物，但并不代表它在所有问题上都能给出高质量的答案。用 GitHub 上数以亿计的代码条目训练，当然也能训练出 coding 模型。这是「学习结果」的逻辑，也是没问题的。毕竟编码任务的结果是可以验证的：代码能不能运行，测试能否通过，结果摆在那里。但是，通往结果的过程，是一个涉及多步骤决策、错误纠正、意图对齐的复杂链条。每一次用户的接受、拒绝、补全、撤销、追问、甚至当模型好几次都搞不定或者完全搞错时的辱骂——都是这一链条上的过程信号。强化学习有两种监督方式，一种叫做结果监督，只看最后是否跑通。但是结果监督会催生「奖励黑客」的现象：模型为了能跑通可能写出冗余、脆弱、带逻辑漏洞的代码，但因为测试过了，模型以为自己学对了。而另一种叫做过程监督，对推理路径上的每一步进行打分。上述这些过程信号，只有在 coding agent 运行环境里才能诞生。GitHub 仓库里只有结果，哪怕是去看单独的提交历史，看 PR，都找不到有效的过程信号。在缺乏有效、自主可获得的过程信号的时候，一些模型厂商会采用「蒸馏」的方式，这个事情大家应该已经知道了。蒸馏的逻辑很简单，给同样的输入，老师模型输出什么，学生模型就学着输出什么。但是通过蒸馏，即便可以获取到思维链，得到的仍然更接近于结果，而非被蒸馏的老师模型内部的概率分布。一旦学生在推理中偏离了老师的轨迹，哪怕一个 token 不符合，都有可能发生偏离。这背后是强化学习的基础限制：策略梯度定理要求，优化样本最好由当前正在优化的模型自己去产生。这种数据叫做 on-policy 数据。而通过蒸馏别家模型，在别人的产品里产生的数据，来训练自己模型，都属于 off-policy 数据。模型当然可以从中学到东西，但学不到老师模型内部的概率分布信息。而像 Cursor 这样自己就是 coding agent 产品的公司，掌握着最真实、有效、高质量的训练数据。Cursor 产品本身，就是 coding 模型在实战环境中的最佳训练场。我们可以通过 Cursor 年初的「翻车」，来证明这个逻辑。4.APPSO 读者应该记得，年初 Cursor 发布了 Composer 2，号称「下一代专用编程模型」，技术报道写的相对保守，也没有提供具体的模型底座信息。结果很快，网友就在公开代码片段里发现了 Kimi 的模型 ID，截图传遍了开发者社群，逼得 Cursor 副总裁 Lee Robinson 出面澄清：「Composer 2 确实是从开源底座出发的。最终模型大约只有 1/4 的算力来自底座，剩下 3/4 是我们自己训出来的。」几小时后，Cursor 联创 Aman Sanger 也跟着发了一条道歉：「一开始没提 Kimi 底座是个失误。」五天后，Cursor 放出了完整的 Composer 2 技术报告，显示底座的确是 Kimi K2.5，授权方则是 Firworks AI，大致流程是在 K2.5 上做训练，再继续做大规模强化学习（RL）。但关键之处在于，Composer 2 的 RL 是运行在真实的 Cursor 会话当中，使用与生产部署完全相同的工具和 harness。Cursor 将这套流程叫做「实时强化学习」(real-time RL)，也即将模型的 checkpoint 直接部署到 Cursor 生产环境中，观察用户的响应，收集数据，聚合成奖励信号——最快可以每 5 个小时迭代一次模型版本，然后继续部署到 Cursor 里，循环往复。最极致的案例是 Cursor 的自动化代码补全功能 Tab，每天处理超过 4 亿次请求，每当用户输入字符、移动光标时，模型都会预测下一步动作，如果预测置信度高，则显示建议，用户按下 tab 即接受自动补全。该功能采用的是在线强化学习，在行业内极具特色。Cursor 可以以极高的频率（最快可达每一个半小时到两小时）更新 Tab 的模型能力给用户，直接在产品内收集 on-policy 数据进行训练。这种高频、接近实时的反馈回路，让 Tab 可以学习到极其微妙的用户意图。Cursor 方面透露，这种方法让 Tab 建议的拒绝率降低 21%，接受率提高了 28%。回到 Composer 模型本身。在事情搞清楚了之后，一些 Kimi 员工也删掉了之前吐槽的的推文，Kimi 官方账号发表了祝贺。一家估值 600 亿美元（基于马斯克给的数字），不做自己的模型基座的 coding agent 应用层公司，仍然可以通过产品自身的数据飞轮，RL 出超越基座模型的专有编程模型。所以与其说 Cursor 翻了车，不如说这反而是 coding agent 产品重要性的绝佳例证。Cursor 在另一篇关于实时 RL 的文章里写到：「（训练编程模型）最大的困难在于建模用户。Composer 的生产环境里不只有执行命令的计算机，还有监督和指导它的人。模拟计算机容易，模拟使用它的人却很难。」这句话，现正在逐渐成为了在编程模型方面走在前沿的模型厂商之间的共识。如果你去看 benchmark 榜单和用户普遍评价，会发现哪些头部的厂商都在发力做自己的 coding agent/编程产品。区别只在于谁离用户更近。我们以 SWE-bench、LLM-Stats 等相对权威的榜单为例，Claude、GPT、Gemini、Kimi 等模型基本霸榜前十，清一色都是有自己开发 coding agent 产品（包括 CLI、IDE、集成 coding agent 的桌面客户端）的模型厂商。在部分榜单上会出现少数反例，如 Meta (Muse Spark)、DeepSeek 等，没有开发自己的 coding agent。不过你会发现，这些反例模型，在更加接近真实场景、避免污染的更权威 benchmark 上就很难上榜了。以 DeepSeek 为例，它在 SWE-bench bash only 上分数是 70%，排名第九，在 SWE-bench Pro 上分数却掉到了 15% 左右。OpenRouter 的真实流量数据可以解释这种反差：该平台 2025 年报告显示，Claude token 消费 80% 以上用于编程和技术任务，而 DeepSeek token 消费主要集中于闲聊和角色扮演。没有自家 coding 产品的厂商，在一些 coding 任务 benchmark 上能挤进头部，但在更难的真实工程 benchmark 上，在用户用 token 消费投票的真实流量中，都会原形毕露。不仅是 Cursor，Anthropic 在 2025 年 11 月发的一篇论文里，也明确透露自己在做一模一样的事情：「我们在 Anthropic 自家的真实生产编程环境上做训练。」也即 Anthropic 把自己员工使用 Claude Code 的交互数据，反哺给 Claude 模型用来训练。5.在 AI 的演进历程中，生产要素的定义发生了深刻的位移。传统三大核心要素——算力、研究、训练数据，虽然在总量上持续增长，但在结构上已经出现了严重的失衡。今天的各大 AI 巨头显著提高了在算力上的资本支出 (CapEx)，让算力基建成为了当前舆论的主旋律。但实际上，特别是在编程范畴内，随着 GitHub 仓库、StackOverflow 等互联网公开代码数据被基模厂商「竭泽而渔」式地利用，模型在代码生成与逻辑推理上的边界开始逐渐显现。这也是为什么，行业共识正在逐渐转向一个冉冉升起的新战略高地：对于任何希望掌握顶级代码能力的模型厂商而言，建立自有的 coding agent 产品早已不再是可选的商业路线，而是确保底层模型可以持续进化的核心生命线。正如前面 APPSO 论证的那样，单纯学习公开数据等于只学习成功者的结局，却无法了解成功的路径，这绝对不是正确的成功学应该有的样子。在真实的编程环境中，知道发生了什么错误、怎样发生的、如何正确地理解和高效地实践需求等等——了解正确过程的价值，远超于得到正确结果本身。只有拥有自己的编码产品，模型厂商才能获取高质量的「过程监督」信号，从而在编码/推理能力的下一阶段竞争中，确保自己仍有技术护城河——否则就不得不像 SpaceXAI 那样，花钱去跟 coding agent 产品公司去合作。然而并不是所有模型厂商都跟马斯克一样有钱，以及 2026 年开始的巨头势力划分、结盟与领地的争斗会变得更加激烈，当一家缺乏自主 coding 产品的模型厂商终于回过味来的时候，恐怕已经没有足够的合作伙伴可以挑选，合作

今年 3 月，苹果制造了 MacBook Neo 这个大爆款，把高质量入门笔记本的价格打到了 4599 元人民币起。一个半月后，这台搭载 A18 Pro 处理器的入门款 Mac，全年销量从 700 万台追加到了令人难以置信的 1000 万台，交付周期一度排到了 5 个月之后……苹果紧急向鸿海、广达等供应商追加订单。知名供应链分析师郭明錤给出更刺眼的数字：在全球 Windows 笔记本电脑预计同比衰退超 10% 的大背景下，2026 年 MacBook 系列出货量有望达 2500 万台，同比增长 20%-25%。这一拳，Windows 阵营的最核心代表英特尔，该如何接住？在今天，英特尔在中国正式推出第三代酷睿处理器，代号 Wildcat Lake「野猫湖」，并同时启动了联合产业链伙伴的「Firefly 萤火虫计划」。英特尔的接招，围绕一件事而展开：将最先进的 Intel 18A 制程，下放给低端，或是入门，或是「主流」——随便你怎么称呼它的价位段。野猫盯上新的猎物Wildcat Lake 这个代号，在去年底就已经在业界出现。今年 CES 期间，英特尔高调发布了主打高端的产品线：酷睿 Ultra Series 3，采用 Panther Lake 架构。同一天，Wildcat Lake 也悄悄披露，你可以将它理解为 Panther Lake 的精简版本。业界普遍将其视为英特尔为了对标苹果 MacBook Neo 而推出的产品，专攻平价入门级 PC 市场。「直接对标」这四个字，并不是说说而已。在 4 月底，英特尔为野猫湖推出的公版参考设计曝光，采用渐变配色、铝合金一体成型机身、窄边框键盘，扬声器开孔环绕在键盘两侧、D 面无出风口——这种设计语言不需要过多解释，它就是在用苹果惯常的设计语言，去回应苹果本身最受欢迎的多彩配色与铝合金轻薄机身策略。在 MacBook Neo 上，苹果做对了一件事：它让整个行业重新意识到，在「内存末日」(RAMmageddon) 的今天，入门价位段成了一座金矿。用一颗本来进入销库存阶段的 A18 Pro 芯片，配上 8GB 内存和 256GB 存储，开价 4599 元——MacBook Neo 活脱脱成了 MacBook Air 平替。英特尔，以及整个 Windows PC 阵营不能继续将这个用户群体拱手相让。而这，正是「野猫湖」存在的（几乎）全部理由。18A 首次下放到入门段第三代英特尔酷睿处理器的技术参数并不复杂，CPU 采用 2 个 Cougar Cove 性能核，加上 4 个 Darkmont 超低功耗能效核，总共 6 核 6 线程；GPU 部分配备 2 个 Xe3 核心，NPU 算力最高可达 18TOPS。整机基础 TDP 15W，最高 35W，在无主动散热的笔记本上，甚至可以压低到 11W 运行。按照英特尔今天提供的官方口径，这颗芯片功耗相比前代降低超过 60%，图形 AI 性能提升了将近 3 倍，平台总算力更是达到了 40TOPS。当然，考虑到「野猫湖」处理器的 NPU 算力只有 17TOPS，远低于 Windows 阵营的所谓「AI PC」，也即微软官方定义的 Copilot+ PC 40TOPS 的阈值。所以严格来说，野猫湖并不能直接和苹果 A18 Pro 高质量的神经网络引擎，以及统一内存架构，去做横向比较。话虽这么说，驱动参数量不大的本地小模型，做一些 AI 推理任务，还是绰绰有余的。以及更有意思的是，谷歌前段时间发布自家的 AI 笔记本电脑「Googlebook」的时候，也确认了和英特尔结盟——Wildcat Lake 被海外业界人士普遍认为将会成为是 Googlebook 的核心平台。连谷歌都觉得够用，应该没大问题吧？真正关键之处不在于这些参数，而是在于「野猫湖」用上了 Intel 18A 制程工艺本身。要知道，过去两代酷睿 Series 1 和 2，一直沿用 Raptor Lake-U 架构和 Intel 7 制程，实质上是 10nm SuperFin 工艺的演进版本。在主流芯片普遍走向 5nm 甚至 3nm 的当下，技术代际差距已经相当明显。当 MacBook Neo 用虽然是老库存但货真价实 3nm 工艺的 A18 Pro，杀进四千多元的价位段，英特尔过去那套「靠老制程压低成本」的入门策略已经过时了。所以，我们很欣慰地看到，英特尔将 18A——2nm 级别的工艺，压到一颗入门芯片上。这意味着英特尔充分愿意承担初期更高的单位成本，去力保一座不能再丢弃的城池。系统：手机供应链思路搬到 PC 上对于守住市场，「Firefly 萤火虫计划」同样重要。英特尔副总裁、中国区软件工程和客户端产品事业部总经理高嵩如是说：芯片本身的算力固然重要，但「芯片和系统层面的双重创新」才是这一代主流 PC 体验跃迁的来源。中国区客户端与平台销售业务部总经理宗晔则认为，PC 也正在从工具进化为「智能体验入口」。根据英特尔提供的内部数据，与上一代相比，「野猫湖」产品的 PCB 面积缩小了 5%，器件数量减少了 7%。同时针对第三代酷睿平台特性制定并推广标准化主板与 I/O 板连接，推动主板核心电路进一步实现模块化、小型化、低功耗化。这一套打法的精神底色，其实来自于手机。手机行业过去 10 年磨合出的供应链协同方式——元器件高度集成、模块化主板、统一接口标准，被英特尔搬到了入门级 PC 上。结果，就是更轻、更薄，也更便宜。参与「Firefly 萤火虫计划」的合作伙伴，包括但不限于华硕、七彩虹、荣耀、惠普、联想等。英特尔预计，全球合作伙伴将陆续推出超过 70 款基于 Wildcat Lake 平台的产品。目前，荣耀 MagicBook X14 2026 已经率先上架，搭载该平台 Core 5 320 处理器，采用 16GB+512GB 存储配置，定价下探到了 4399 元。用户画像：完全对标 MacBook Neo，但还有更多英特尔列举了 Wildcat Lake 平台产品的 5 类目标用户：辅导孩子作业的家长、全天在校的学生、刚入职场的新人、「一人公司」创业者，以及银发族。这些用户群体和场景描述本身没什么悬念，但值得注意的是，这个用户画像，和 MacBook Neo 在线下门店主要推荐的人群，几乎是完全重叠。MacBook Neo 夺走的本来就是入门级 Windows 笔记本电脑的客群，英特尔的接招，就是造出具有更好价值主张的入门级笔记本电脑，再把心智重新抢回来。英特尔还反复强调野猫湖平台的本地 AI 推理算力。这对于那些购机预算本就不多，也不太可能付费订阅优质云端模型算力的用户来说，倒也是个正确的思路。最后，野猫湖看起来像是合格的入门芯片平台，跟 Firefly 萤火虫计划结合起来，也形成了一套颇具诚意的组合拳。轻、薄、便宜的 AI 笔记本，英特尔现在重新回到了这个擂台。仅剩的悬念，是 OEM 最终能把零售价做到多少。在 MacBook Neo 的价位上，苹果给到了系统生态、统一内存架构、多种潮流配色等等诸多看头。但英特尔要想把这只野猫放出去，跟苹果大杀四方，光有 18A 工艺和优秀的工业设计是不足够的。这场仗的真正交战点，大概率不会在技术本身，而是在即将到来的 618 购物节期间，第一批搭载它的笔记本产品，能否真正把供货紧张的 MacBook Neo 杀个七零八落、措手不及。

快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验主作者关注Linery新手上路Linery关注Linery新手上路联合作者关注Linery新手上路Linery关注Linery新手上路05/13 20:36用微信发信息是我们熟悉到几乎成为本能反应的日常操作，不过作为国民应用，微信为了保证人人可用的普适性，也一直保持着极为克制的功能更新节奏。直到七年前鸟巢那场颇具争议的「灾难级」发布会上，锤子 TNT 次世代生产力工作站发布时，老罗在最后展示了他对「聊天应用」的理解——子弹短信。子弹短信以其极简高效的交互设计、处处贴心的产品巧思，让人耳目一新，也成为了那场发布会后少有的、备受好评的产品。可惜彼时锤子科技已经风雨飘摇，子弹短信最终没能成为击倒巨人的大卫，在变身「聊天宝」尝试「聊天赚钱」的模式也没能最终挽回颓势后只能草草收场。这个意难平也让我在用微信的时候总感觉有一根刺：如果聊天可以再「畅爽」一点就好了。比如我可以直接在一个界面里快速给多个联系人发私聊信息，这样在同时需要联系很多人的时候，可以不用频繁打开，关闭聊天窗口。我试着让 AI 帮我列举了一些需求场景：周末聚餐，要挨个通知朋友时间地点；临时加班，要分别跟家人、伴侣、长辈说一声晚点回家、不回去吃饭；工作上有事，要同步给同事、领导、合作对接人；出门用车、约球、约见面，要一一私聊好几个人。再比如更智能的群发，快速给不同的人发想同模板的信息时，如果能自动填充联系人的「称谓」等字段，也能让群发更人性化；我甚至也想用微信做更多的信息记录，比如聊完天可以记录一下感受、吐槽、资金往来、注意事项等。从锤子转到 iPhone 阵营后，出于一些「新奇」功能的急切需求，我开始捣鼓起了 iPhone 的「快捷指令」，用它来解决工作和生活中遇到的问题。并且最终我渐渐做出了一些围绕微信聊天的快捷指令。今天分享其中的一个。为了解决「不频繁打开窗口，同时给多个人发信息」的需求，我利用第一性原理得出了一个简单的结论：聊天本质是给指定联系人发送一段文本。所以不妨剥离冗余的交互，只保留最核心的「点名 + 发消息」。在设计时我发现快捷指令完全可以实现「输入联系人名称、再输入要发送的文本、最后发送出去的效果」：在锁屏上直接发微信的效果举点日常真实示例：妈 晚上我不回家吃饭了奶奶 今晚在外边吃，打完球给你带榴莲包回去万基 晚上七点球场准时 solo世平 好久没见，找个周末抽空聚聚Peter 方案你明早抽空看一下，下午拿给老板签字莞少 今明两天我要用车，你这两天别把车开走不用来回跳转、不用重复复制切换，静下心一次性把所有人的话都编辑好，一键自动群发，省心又高效。这个功能每次用都非常愉悦，一是聊天体验新奇（我觉的很像学生时代传纸条），二来是真的完美帮我躲开了碎片化通信对于正常生活和工作的干扰——作为潜在的 ADHD 人士我会经常陷入这种情景，点开微信本来只想发一条消息，结果一眼刷到朋友圈、被视频号内容勾走注意力，半小时过去，信息还没发出去。另外就是每次用它一次性把该对话的人该说的话说完，有一种特别畅快的感觉。它当然无法替代微信原本的日常聊天、连续闲聊、刷圈看动态的需求，但是作为一个于我而言的重要补充，真的非常好用、可靠。最后说点感性的话：我是一个非常怕麻烦的人，对所有可以减少冗余步骤的事都有极大的好感，但我也逐渐能理解这个世界上为什么会有这么多「麻烦」。麻烦对我而言变得中性，我意识到它代表了一件事的中间态，在优秀的产品（比如微信）中它常常体现为极致的折中，这同样也需要付出大量的工作、甚至是天才的创想才能达到的。这在我看来真的非常了不起。快捷指令就是 iPhone 提供给用户，用来化解这些「麻烦」的一个非常好的工具。它有丰富的接口，例如在这个功能里，就利用了微信提供的「给联系人发信息」接口；也有非常多的触发方式，包括但不限于桌面图标，锁屏、控制中心、负一屏小组件，锁屏快捷按键，轻敲背面，操作按钮（原静音拨杆）——当然还有少数派的 Bang!Case。这都让它变成了一个能让每个人通过轻度开发解决实际问题的工具。我真的非常喜欢快捷指令，它让我学会去缕析一件事的脉络、链路，动手去改写、制定新的规则和流程，让工作和生活都接近我想要的理性，跳脱出很多无形的枷锁。针对微信我未来还会做很多有趣的功能，比如智能群发，定时发送等等，感兴趣的朋友可以保持关注。我把这个快捷指令整理好了，有需要的派友可以安装试试，添加到自己趁手的位置就可以了。编注：题图由 ChatGPT 生成。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀3232扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@克莱德© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#自动化#效率#微信#快捷指令32等 32 人为本文章充电扫码分享举报本文章Linery还没有介绍自己关注

Apple WWDC26 全球开发者大会，来与少数派一起看主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。05/15 16:59北京时间 6 月 9 日凌晨，WWDC26 全球开发者大会首日 Keynote 主题演讲活动即将揭幕。届时，我们将迎来 iOS 27、iPadOS 27、macOS 27、watchOS 27、visionOS 27 等新一代操作系统的亮相，Liquid Glass 设计语言将如何打磨进化、iPadOS 多任务体验又将走向何方…… 这些大家关心的问题也将在会上悉数得到解答。与此同时，macOS 27 彻底告别 Intel 时代，会带来哪些新变化，也是今年 WWDC26 值得关注的看点之一。为了与同样关注 WWDC 的你一同迎接这场科技盛会，少数派将于 WWDC26 首日 Keynote 主题演讲活动期间在深圳举办线下直播观影活动。欢迎你来和我们一起交个朋友、一起聊天，观看 WWDC26 全球开发者大会。购票链接：深圳场次广州场次其它场次活动信息活动日程 (深圳)活动时间：06 月 08 日 22:00 - 06 月 09 日 03:00活动地点：广东省深圳市宝安区西乡街道银田工业路元匠坊 C 座 201 (少数派共创总部)活动日程 (广州)活动时间：06 月 08 日 22:00 - 06 月 09 日 03:00活动地点：广州市荔湾区泮塘五约外街 76 号少数派 (1200 书店旁)活动门票本次活动为门票收费制形式，门票价格 9.9 元，少数派会员可免费入场。所有门票收入将用于活动的组织和宣传，包括不限于场地租赁、茶歇准备、物料制作、伴手礼品等，由少数派方面负责管理，所有解释权少数派所有。编注：由于场地空间有限及后台系统方面的限制，少数派会员仍需下单购买门票后才可入场参与本次活动，下单时可享优惠金额 0.1 元。活动礼品活动门票定价为 9.9 元，除了入场资格之外，每位参与者还可获得：少数派 Prime 月度会员卡少数派 logo 装饰贴纸现场抽奖机会多城同屏，同步相庆本次 WWDC26 派对之夜线下特别活动，我们依然与少数派的老朋友 SwiftGG 共同举办，如果你不能参加我们在深圳和广州举行的线下活动，也可以报名在北京、上海、杭州、青岛同步举办的线下观影活动，现场同样准备了精彩的演讲和互动环节，并会为到场观众发放精美伴手礼品。了解详情 >Apple Developer Community Events在 Apple Developer 官方网站的社区活动页面中，Apple 已将少数派与 SwiftGG 联合举办的 WWDC26 系列活动收录至页面中，你也可以在此找到全球其它国家和地区经过 Apple 挑选的相关活动。>Community-driven events - Apple Developer温馨提示由于活动举行时间较晚，请购票用户合理安排行程，准时参加，并注意路途安全。活动伴手礼以活动现场发放实物为准。所有门票收入将用于活动的组织和宣传，包括不限于场地租赁、茶歇准备、物料制作等，由少数派方面负责管理，所有解释权归少数派所有。请在观看 WWDC26 Keynote 主题演讲期间保持安静，确保不会影响其他观众及周边居民。活动期间请注意个人物品的保管，建议提前规划活动结束后的归家方式，尽量避免步行或单独行动，保证个人安全。鸣谢本次系列活动由少数派、SwiftGG、漕河泾 Hi-Tech Park 等合作方共同举办，感谢各位合作伙伴对于本次活动的支持。如果你有其它活动想法或者场地推荐，欢迎通过邮件联系我们：[email protected]。164扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@waychane© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#WWDC26#Apple16等 16 人为本文章充电扫码分享举报本文章少数派编辑部少数派编辑部官方账号。关注

按：从播客访谈到深度长视频，越来越多的内容创作者把表达重心放到了音视频里；「先转录、再阅读」也因此成了不少重度信息消费者的日常工作流。。本期「装了啥」，我们请编辑部几位同事分享各自正在使用的转录方案，希望其中至少有一种思路能为你所用。MacWhisper 的平价替代：TranscribeX@广陵止息：虽然我很喜欢文字传递的高密度信息，但越来越多的内容创作者已经转向播客和视频；至于多人访谈类节目，音频传达出的现场氛围感也比文字更生动。所以我这些年消费音视频节目的时间也逐渐拉长，但有价值的节目「三心二意」地听一遍远远不够，反复听又没有足够的时间和精力。所以，现在遇到有价值的节目，我都会转录成文本，方便后续检索。从 Time Machine 里翻了翻，找到了当时的文件夹记录以前我采用的是whisper.cpp配合 OpenAI Whisper Large V3 模型来转录文本，但这个方案有两个问题。一是音视频下载的渠道很多，挨个手动下载很麻烦；二是转录时间偏长的中文视频时，有不小的概率会在转录过半之后出现重复输出、卡字，导致转录失败。但我实在不想翻 log 和模型仓库找相关讨论了，所以也开始寻找替代方案。首先想到的是 MacWhisper，看了眼当时价格已经涨到完全不考虑的地步。好在后来找到了整体体验和 MacWhisper 很类似的 TranscribeX，除了收录在 Setapp 里，对我没有额外成本外，即使单买也只要 MacWhisper 一半价钱。和 MacWhisper 类似，TranscribeX 支持本地文件导入转录、麦克风录制、录制特定软件、实时转录、视频网站下载转录和批量转录。其中「下载视频网站」功能主要依赖 yt-dlp 实现，所以基本不用担心国内服务的下载问题。转录模型方面，TranscribeX 也支持 WhisperKit、Parakeet、千问和 Mistral 的转写模型，必要时还可以调用 ElevenLabs 的在线 API。不过我本地基本只用 Whisper Large V3，别的模型也没用过，所以暂时不评价转录效果。相比 MacWhisper，TranscribeX 支持 Apple Silicon 中的 NPU 加速，这也是我之前选 whisper.cpp 的原因。NPU 虽然不会让转录速度快很多，但至少负载不会全部堆到 GPU 上。设置时需要注意：Audio Encoder：主要负责「听懂声音」的环节，可以只用 NPU；Text Decoder：主要负责听写的环节，为了转写速度，建议设置为 All。转写界面其实每个软件大同小异，但还是有一些设置可以说一下。首先是自动检测「说话者」：Pyannote：适合会议、访谈等场景相对固定的情况，也更贴合我常用的场景，可以快速调用预定义的模板；Sortformer：偏向速度、实时和轻量化的处理方式，适合会议、通话等实时场景。外语苦手也不必担心，TranscribeX 内置了三种翻译方式：Apple 内置的翻译 API、DeepL API 以及第三方 AI API。Apple 内置「翻译」速度很快，准确率也够用，而且免费；对翻译质量有要求的话也可以调用 DeepL 或者其他 AI。TranscribeX 也随大流加入了 AI 对话功能，可以针对转录的内容做进一步处理。最后值得一提的是，针对 Setapp 版本的 TranscribeX，所有的在线 AI 额度都可以用 Setapp 的 AI 配额。但 TranscribeX 还是有自己的问题：没有 CLI 版本，远程使用一定要靠 VNC 连接到 Mac；辅助框架的设计上也有不小的问题，目前还很难配合比较火的 AI 助手。如果你比较在意自动化体验，TranscribeX 可能没有那么适合你。用 Gemini 整理@PlatyHsu: 我这里主要谈谈转录后的整理。对于这类任务，我最推荐 Gemini 系列模型。主要原因是，在排名靠前的模型中，Gemini 的多模态支持比较全面，Gemini 3 全系（Pro、Flash 和 Flash Lite）都可以直接接受音频和视频文件输入，免去了先创建音频转录或者视频截图的功夫。我也认为 Gemini 的中文能力在国外模型中相对是比较好的。与此同时，Gemini 的用量给得比较慷慨，AI Pro 方案很少遇到不够用的情况。我最常用的场景之一是整理播客的录音稿，用来确认一些初听时一闪而过的模糊细节。一个参考提示词是：转录用户提供的音频，并整理成表达规范、分段清晰的书面格式。区分说话人，并标注在段落之前。将较长的片段合理拆分成若干自然段。可以删除语气词和重复片段或修改病句，不要做其他删减或总结。具体模型方面，一般可以先从 Flash 开始尝试，对于日常主题的节目来说一般就够了。一小时的播客大约需要四分钟处理，肯定没有本地模型快，但是成品高度可读，值得这个等待时间。不过我也遇到过一些涉及哲学、历史等较专业话题的节目，用 Pro 模型会有更好的表现，可以灵活调整。当然，如果想节省 token，你也可以先在本地转写一遍，然后用以下提示词整理文本：将用户给出的音频转录整理成表达规范、分段清晰的书面格式，可以删除语气词和重复片段或修改病句，不要做其他删减或总结。这个用例一般建议直接使用 Pro 模型，因为本地模型转写的文本一般有很多拼写错误和赘余，Pro 能更好地还原和修复。除了音频之外，Gemini 也可以用来总结视频。并且由于它可以直接「看到」视频内容，因此即使是那种没有旁白的 TVC，也可以说出个一五一十。例如之前一个项目中，甲方提供的写作素材之一是一条广告片，我就是用这个方法把其中的信息提取成文字描述，然后用在了写作里。以下是用苹果最近一条 Mac 广告做的演示，提示词是：详细描述该视频的主题、思路和内容，并列举关键帧的时间戳和画面内容。因为涉及较多「言外之意」的解读，这个用例也是用 Pro 模型处理为佳。以上方法在 Gemini 网页版和 AI Studio 都可以使用，但网页版经常会出现「偷懒」的情况，只给出非常简略的输出，因此建议尽量用 AI Studio 来操作。命令行工具搭配 Agent@waychane: 由于常有参加线上会议的需求，会中涉及主讲者的介绍、采访以及 QA 等内容，这类会议的时长动辄就有一个小时。在会后创作内容时，如果能有一份观点清晰、结构清楚的要点大纲，会省下不少功夫。现在可以说是 Agent 时代，我自己也在用 Claude Code 等 Agent 类 AI 工具帮我处理各种工作。因此在挑选特定场景下的 app 时，我会希望它们要么提供 CLI 操作方式，要么就是纯粹的命令行工具：一方面，这样可以在同一个 Agent 界面中高效地调用不同工具；另一方面，许多 app 不仅功能复杂，使用参数也多到难记，用 Agent 调用刚好能解决这些问题。面对会议音频转录的需求，我选择的是 whisper.cpp 这个命令行工具。它是 OpenAI Whisper 的 C/C++ 移植版本：OpenAI 官方版 Whisper 基于 PyTorch 的 Python 实现，要跑起来得先配好 Python 环境和一整套依赖，在 Mac 上只能依靠 CPU 计算，想用到 GPU 还得自己折腾 PyTorch 的 MPS 后端，体验并不顺畅；whisper.cpp 则用 C/C++ 重写了推理部分，原生接入 Apple 的 Metal 后端和 Accelerate 框架，能有效利用 Apple Silicon 的 GPU 和统一内存架构，模型权重也换成了体积更小、加载更快的 ggml 量化格式。

本周看什么 | 最近值得一看的 8 部作品主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注潘誉晗少数派作者潘誉晗关注潘誉晗少数派作者关注利兹与青鸟少数派作者利兹与青鸟关注利兹与青鸟少数派作者关注SHY少数派作者SHY关注SHY少数派作者少数派编辑部 等 4 位作者05/15 18:00☕️ TL;DR近期佳作推荐：[国产] 低智商犯罪、[韩剧] 夺命许愿、[英剧] 好兆头 第三季、[法剧] 抢钱女王团、[电影] 蒙特利尔，我的美人、[动画] 石纪元 科学与未来 第 3 部分、[纪录片] 马克斯·维斯塔潘：新征途、[真人秀] 旧洞基朋友们几则精彩预告：《攻壳机动队 THE GHOST IN THE SHELL》新预告、《沉默的审判》释出预告、《火遮眼》新预告、《爱情城事》定档预告、《奥特曼起源》正式预告几则影视资讯：《诺曼底 72 小时》内地定档 6 月 6 日、《宇宙巨人：希曼崛起》内地定档 6 月 5 日、《家弑服务》内地定档 5 月 29 日、《超级少女》内地定档 6 月 26 日、漫威新剧《幻视追寻》正式定档往期推荐可点此或点击文末 tag #本周看什么，更多主题片单可点此查看；想要查看更多深度影视内容，可以关注我们的衍生栏目「看什么 Café」。如有其他建议也欢迎在评论区提出，感谢大家对「看什么」栏目的喜爱和支持。[国产] 低智商犯罪关键词：喜剧 / 犯罪片长：45 分钟左右（单集）× 24 集观看渠道：爱奇艺/咪咕视频；豆瓣链接运气之神降临！@利兹与青鸟：平安市刑警张一昂能力不强但热爱吟诗，常常不合时宜得令人发笑，被下放到老干部处都能惹出乱子。因为他正巧捡到一封举报信，于是和同样脑回路清奇的李茜一起被派到三江口，调查公安局副局长的死是否有隐情。但空降而来的张一昂很不受欢迎，工作处处受限，却在阴差阳错之下抓住一个在逃嫌犯，让人刮目相看，就这样被架起来成了神探。另一边，平安市两个笨贼想出了抢贪官的主意也来到三江口；而三江口涉黑的集团老板周荣考虑着新项目，也着手调查发小叶剑刑警队长的死因。三条故事线、两桩凶案（可能不止）就这样汇集一处，想必会给三江口这一团乱麻摩擦出不少火花。本剧改编自紫金陈的同名小说，一开局就满是带着黑色幽默的荒诞味，罪犯和警方彼此高估，虽然智商时常掉线，但不妨碍瞎猫碰上死耗子，不禁让人好奇这份「好运」还能带来什么啼笑皆非的巧合。马旭东、闫佩伦、詹鑫等喜剧演员的助阵也让笑点呈现得更好，尽管罪案线稍显拖沓，但也算是合格的电子榨菜。5[韩剧] 夺命许愿关键词：剧情 / 恐怖又名：기리고 / Wish Your Death / If Wishes Could Kill片长：44 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接为了愿望达成，你愿意不惜一切代价吗？@潘誉晗：这次数学考试的结果有点出乎意料，因为崔亨旭拿满分了，可这家伙不是个读书的料子。面对好友们的询问，亨旭得意洋洋地把一个名叫「Girigo」的软件链接分享给了伙伴，说是只要写上自己的生辰八字贴在自己的身上，然后录下视频上传到这个软件上，就能实现愿望。大家并不相信亨旭的话，亨旭也没在意此刻出现在他手机上的倒计时。谁承想 24 小时后，亨旭就当着全班同学的面自缢了——原来这则所谓的许愿软件，在愿望达成后会要你的命。Netflix 韩国这次又推出了一款爆剧，一经上线就在 13 个国家登顶。剧集整体的节奏很是紧凑，恐怖惊悚的氛围也营造得相当出色，很好地把校园青春剧和恐怖因素结合在一起，然后又完美连接了手机 app 和萨满巫术。神秘的夺命软件利用了青少年的欲望，随后就像是一个大家无法挣脱的隐形牢笼，不仅顽固地无法在手机上卸载，更是在许愿者身后步步紧逼、致人于死地。2[英剧] 好兆头 第三季关键词：剧情 / 喜剧 / 奇幻又名：Good Omens Season 3片长：99 分钟观看渠道：Prime Video；豆瓣链接我有那片宇宙，我还有你，我已经拥有我想要的一切。@SHY：成为至高大天使的亚茨拉斐尔，开始筹备「第二次降临」，复活的耶稣却不知所踪，只得向自暴自弃的克劳利求助。这对自创世之初就有交情的天使和恶魔，能否像往常一样弥合分歧，而在背后作祟的幕后黑手，又会将整个世界导向怎样的结局？受原作者兼前两季主创尼尔·盖曼的性侵丑闻影响，最终季被砍成了仅 1 集的特别篇，篇幅势必捉襟见肘。不难想见，成片处处流露着删减的痕迹，忽快忽慢的节奏让叙事更加混乱，整体观感有着不小下滑。虽然填充的内容变了样，两位新编剧还是尽力还原了盖曼构想中的结局，而在破碎的字里行间，也能窥见一些残存的灵光。惯常的黑色幽默难以撑场后，本季的重心落在了本来带点调侃的父母爱情故事上，大卫·田纳特和麦克·辛以一如既往的真挚演技力挽狂澜，继续迸发充沛的化学反应。彻头彻尾的粉丝向结局，评价两极分化在意料之中，好在，本作原本就是作为限定迷你剧立项，如果你有点难以接受这一季的走向，只把第一季尊为正统也未尝不可。发表首评[法剧] 抢钱女王团关键词：喜剧 / 惊悚 / 犯罪又名：Les Lionnes / Cash Queens片长：55 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接我们之所以抢钱，真的是迫不得已。@潘誉晗：罗莎这天找到自己的上司，希望可以提高她的信用卡额度。可上司不仅很快就拒绝了她的请求，甚至通知她的额度被降低了，因为她征信由于入狱前夫的债务问题受到了影响。罗莎不敢相信上司的话，也就是在这时，银行的警报铃响了，原来这天又是银行例行的演戏日。看着上司希望职员们保障生命的大喊声，一个不该出现的念头在她的脑海中出现了——她决定伙同自己的好友，抢劫自己工作的银行。Netflix 法国于两个月前推出了一部女性犯罪喜剧《抢钱女王团》，讲述了五名女性在窘迫生活的逼迫下，不得不伪装成男性抢劫银行的故事。全剧充斥着大量很荒诞又很现实的设定，主角团的五名女性拼命工作、好好生活，但却一直过得不幸福。职场上的性别歧视、身上背着重债、被同侪做局丢掉了工作、被丈夫家暴……不是说努力就能有好结果吗？为什么在她们的身上，越努力却似乎越悲伤呢？发表首评[电影] 蒙特利尔，我的美人关键词：剧情 / 同性又名：Montréal, ma belle片长：117 分钟；豆瓣链接被困在夹缝中时，是面对自己还是活在谎言里？@利兹与青鸟：凤霞一家已经移民加拿大十几年了，在魁北克省的蒙特利尔开着一间小超市。丈夫性格较为古板，苦于没有当地学历，找工作频频受挫；大女儿正值叛逆期，对待父母多是指责和抱怨；儿子也只是上小学的年龄，家中没有人能为受困于更年期的凤霞提供情绪价值。在法语课上，凤霞被年轻人追求梦想的热情所鼓舞，尝试接触同性交友网站，并结识了 30 岁的卡米尔。这让凤霞重新找回了自己，却引发了家庭危机。本片的华裔女导演很有相关生活体验，她用代际间新旧观念的碰撞、婚姻大于爱情、照顾家人大于自我实现等细节，呈现出一个传统而压抑的中国移民家庭外壳。片中演员也多是当地华裔，汉语都很纯正，对华语区观众的耳朵十分友好。电影虽是在讲中年女性在道德边界寻找自我，却没有激烈的口号，而是用朴素温和的风格、平缓的节奏，给主演陈冲留足表演空间，为年过半百却天性浪漫、依然优雅而富有魅力的凤霞赋予灵魂，细腻又克制地演绎出她内心的彷徨与挣扎，也让观众沉浸其中。女主角是陈冲啊？1[动画] 石纪元 科学与未来 第 3 部分关键词：漫画改 / 科幻 / 冒险又名：Dr.STONE SCIENCE FUTURE 第3クール片长：24 分钟（单集）× 13 集，每周四更新观看渠道：巴哈姆特动画疯；豆瓣链接We choose to go to the Moon.@SHY：孤注一掷的抗争后，科学王国终于取得对北美阵营的胜利。面对迫在眉睫的威胁，千空和赛诺搁置矛盾，联手打造宇宙飞船，前往月球探寻石化的真相。为了复活更多人类，同时取得不可或缺的材料，众人兵分三路执行任务，「WHY 人」却传来了更加危险的信息。从石器时代到现代文明，跨越 200 万年光阴，这段从零开始的科学冒险终于迎来最终幕。从刚苏醒时手工打造的粗糙制品，一步步重现火药、铁器、蒸汽、电力，来到本季，科技树终于攀升到计算机和互联网的领域。要实现登月的最终目标，需要远胜以往的人力和物力，千空等人因此踏上环球之旅，整合各个国家的资源。即使原料由石头进化到超合金，舞台由村落跃进至外太空，本作的核心魅力始终如一。秉持理性和探索精神，用科学之力对抗奇幻，挥洒荡气回肠的热血，开拓人类未来的道路，诠释纯粹的浪漫。这个集结全世界力量、于石之世界铸造的「阿波罗计划」，绝对能让人心潮澎湃，剧中历尽千辛的角色们，也值得最为美好的结局。发表首评[纪录片] 马克斯·维斯塔潘：新征途关键词：纪录片又名：Max Verstappen - New Ground / 极速新域：维斯塔潘的北环之路片长：23 分钟左右（单集）× 3 集；豆瓣链接出发吧！汽车人！@潘誉晗：对于喜欢一级方程式赛车的人而言，效力于红牛车队的马克斯·维斯塔潘，是一个了不起的名字。他出生于赛车世家，自己的赛车成绩也非常傲人。不过卸下这一身份，他其实还有另外一种人生——一家 GT3（Grand Touring 3）车队的老板兼赛车手。而这部只有三集的纪录片，就是讲述维斯塔潘的第二身份。「看，老板亲自上场了。」维斯塔潘没有老板的派头，招呼完就直接上了场；他也不需要拥有全部欢呼与掌声，而是会亲自去挖掘赛车手。于是，他发掘了克里斯，这个和他一样从开卡丁车（并且克里斯的车号是 3）开始的小伙子很有天赋，可苦于 15 万欧元的培训资金，只能把梦想转移到模拟器赛车上。当维斯塔潘找到他，让克里斯在场地上尝试开车的时候，克里斯甚至没有驾照。相信在观看了这部纪录片之后，没人会否认维斯塔潘对赛车的热爱。他脱下了红牛的飞行夹克，在另一条赛车的道路上，继续驰骋。发表首评[真人秀] 旧洞基朋友们关键词：真人秀又名：구기동 프렌즈 / Six Singles Under One Roof片长：90 分钟左右（单集）× 10 集；豆瓣链接一起同居吧。@潘誉晗：韩国最近又推出了一部真人秀《旧洞基朋友们》，找来了张度妍、张根硕、安宰贤、李多熙、景收真和崔丹尼尔六位年龄相近的艺人，过一次同居生活。随着明星观察类的项目在市场上越来越多，我相信有越来越多的人对这一类真人秀感到厌倦。而《旧基洞朋友们》的出现，再次证明了韩综的创意。看似依然是粉丝向地找了一群有粉丝基础的演艺人员，但是却找准到了一个很细腻的切口——让几个曾经独居过的明星们，过群居生活。独自生活，只是光看到就觉得很自由的日子，为什么他们体验过那种日子后，不愿意再独居了呢？我想这也就是观察类节目出现的初衷，去探讨一个社会性的议题。毕竟「存在即合理」。而这六名 85 后明星的同居日常，也非常有看点。他们都很有个性，但又能互相融入、靠近。见惯了大场面的他们，看到 18 公斤的章鱼，依然会手足无措，还是会大声尖叫。原来群居也不差嘛。年龄相近 不是想尽1更多[电影] 总统的蛋糕@利兹与青鸟：萨达姆的独裁统治下的伊拉克，平民们过得水深火热。小女孩拉米亚住在草编的窝棚里，连书包里的苹果都要被老师搜刮走。更不幸的是，她被抽到要在总统生日这天做好一个奶油蛋糕。于是拉米亚踏上了旅途，用家里所有值钱的东西换取做蛋糕所需的材料。电影在去年戛纳电影节获金摄影机奖，摄影优美，配乐充满阿拉伯风情，有触动人心的童真，也有伊拉克战乱下的可怖的倾轧。[电影] 平场之月@潘誉晗：电影根据朝仓加守美同名小说改编。青砥健这一天意外偶遇了自己的初中同学须藤叶子。此时两人已经 35 年没见了。中学时两人就互有好感，这次重逢，已离婚的青砥健和丧偶的叶子，终于还是走到了一起。埼玉小城里这一对中年人迟到多年的遗憾恋曲，就像是终于遇见了自己的另一个缺失的半圆，内敛、沉稳又有淡淡的愁绪。[韩剧] 浪漫的绝对值@潘誉晗：每当晚上，看似平凡的高中生吕意珠就会偷偷创作 BL 小说。可惜，她的作品一直无人问津

厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨主作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了联合作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了05/15 00:07Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。免责声明与作者所需披露的一切首先声明一下——本人仍然是在校的本科微电子硕士考古学的大学生，并没有经过任何的传媒学相关的学习，也没有从事相关行业的经验。目前仅有的和媒体沾边的事情是在酷安、少数派等平台发过一些文章（其中有很大一批都是在最后诉诸于，说好听点是真实生活场景，说难听点开始抒情的文章），以及一个 B 站粉丝量远没有破千的账号，因此我远远没有承担过养一个工作室的压力与责任，也没体会到在竞争激烈的环境中生存的艰难。因此以下内容不构成任何形式上的专业意见，只是一个愤青在 Random Yapping。多有冒犯之处，但我本心无意去批评任何具体的某一个媒体老师，只是想尝试提出一个也许还算有点价值的问题。以下内容提到「游戏区」，也仅限于单机游戏，不包含手机游戏、桌游等类型。以下内容提到「数码区」时，通常仅指以手机、一部分 PC 组件为主的主观体验向评测，且我的能力不足以涵盖 DIY 区1以及涉及到实验设计与数据分析的其他评测内容，虽然随着硬件厂商的创新进入后期演变为渐进创新，产品节奏也有所放缓，在社区中也能听到「DIY 区不如以前」的感慨，但本文尽力按照「不理解，没有切身观察就尽量少乱锐评」的原则展开，但撰写过程中有很多内容没收住，因而请让我提前声明一下——以下很多内容只是社区观察，不完全代表事实判断。最后，说句坦诚的实话，这个内容写的过程真的很痛苦——我真的不想让这篇文章变成直球辱骂媒体老师、攻击厂商或者是殴打观众的情绪发泄文，从个人发展角度来说不希望因此在进入之前就已经和某些职业无缘，也不想因此得罪那些因真心倾慕而认识的那些做这一行的朋友，且我真的不希望因为自己的过分言论把发文章的平台本身也拖下水，但因为笔力实在是有限，尽管拉着 Gemini、GPT、DeepSeek 等一众大模型核验，也找了一些朋友探讨，尽可能不要直球攻击特定的一两家媒体也不要说的太过分，但是字里行间还是有一股火药味拼尽全力消不掉。真的很抱歉，尽管我其实将自己也包括在本文的攻击射程以内——我自己就是后文中所批判的那个自认为掌握了真实，在没有厂商利益束缚与照顾社区的必要下肆意输出内容，甚至有很多是过当内容的小自媒体。这些内容因为是我自己一段时间内自我质疑与焦虑的外化，因此，本文必然存在一些因深度不足、角度有限而导致的难以避免的暴言，简而言之——就是年轻时犯下的错误。麻烦各位读者多多体谅。也欢迎您在评论区里多批评指正，让我多与您交流学习。以下内容如果有哪里说的不合适的地方，有什么可能涉及到的法律争议，请尽快联系我去撤换文章中对应的部分，谢谢！本文长度约为三万字左右，前三部分观点将会较为尖锐，会举几个相对较为敏感的案例来佐证观点，后两个部分则尝试解决前文抛出的问题，内容相对温和但字数也相对较长，请坐和放宽。主观口胡系列是一个尝试将动漫、电影、游戏叙事与更加严肃的讨论结合起来，去尝试探讨一些特定大命题的个人尝试，可能会一定程度上背离这些作品的创作初衷，聊的可能也会比较混乱且随意，甚至会有意提取个别元素，曲解原文本来聊一些其他领域的内容。这次我想把一个大约是三四年前就开始产生的想法去做进一步的深化，并尝试结合近些年来我对游戏区与数码区的一些粗浅且悲观的看法，来尝试去聊一个问题，一个也许注定没有答案，跨多少领域学科都难以得出解法的一个小小的问题——主观体验向评测的现状怎样，尽头到底为何方？又是什么构成了它们的专业性？Part1：太多水，太多奶酪——主观内容的梦魇虽然这么说有失偏颇，但我一直觉得游戏评测是一个相当主观的内容——评测者的个人阅历2、偏向的游戏类型甚至是组成一个游戏的系统，对 GDC 中诸多开发者分享的设计理念的理解，以及同期的同类型作品等等——这些都会对你最终得出什么样的结论，给游戏最终打什么分造成影响。所以在因近期的《神探杰克鼠》评测而陷入争议的 IGN 内部——作为一个有 TGA 评选权的，相对来说算是行业内位置靠上话语权比较大的游戏媒体——他们在测评游戏时给出的流程是：在内部筛选最适合某类游戏且当下仍有档期的编辑，由他们在内部其他编辑的帮助下完成对于这个游戏的评测，给出评分，在内部经过审核与排期，最终以文章 + 视频的形式先后放到公共平台上。但是，这个时候总会有个但是。虽然这样的流程确实在内部行之有效，为行业贡献了一批质量很高的评测，尤其是在神作频出神仙打架的那些年里——但同样不可否认的是，无论你之前做对过做好过多少次，一旦出现几次测评结果与大众预期不相符，甚至出现有失水准的内容，就很容易在玩家社群中丧失公信力，变成 IG8——《宝可梦欧米伽红宝石/阿尔法蓝宝石》的 too much water 扣分，《超级马里奥 3D 世界+ 狂怒世界》的 7.0，《死亡搁浅》的 6.8——你知道的，日光之下从来都没有什么新事。而且当我们把视角继续向下探时，会发现这样的问题不仅仅是发生在 IGN 一家，Polygon 的《DOOM(2016)》、VentureBeat 的《茶杯头》（他的操作已经不足以让他完成第一关，只能借其他同事存档继续评——尽管他据自己说已有着超过二三十年的评测历史）……所以到底发生了什么？当然，抛开社区中已有的被讨论过无数次的，对评测者的质疑不谈，我们也许可以从更加细致的视角去看这件事。就像你在做数码类评测会提到机型的性能与续航、外观、系统流畅度与体验、拍照、外放表现，但你不可能在非首发上手的评测里一样面面都要点到，而是根据个人的使用习惯有自己的偏好一样，IGN 的每个评测者在评测游戏时也会有自己的明显倾向，让我们以当年评测《死亡搁浅》的 Tristan Ogilvie 为例——他评测一款游戏的重心是动作系统、交互反馈与流程的顺畅程度。在我眼里，相较于小岛在前作《合金装备 5：幻痛》中用一系列主线支线任务，吸引玩家不断的从母基地回到开放世界，完成一串任务后撤离不同，本作让角色更加持久更加连续更加长时间的停留在这个世界里，将任务可视化，变为地图上要到达的目标点，并在沿路设置一系列需要潜行的敌人据点、风景奇观，并且可以通过基建加速到达目标点跳过一部分游戏内容来优化体验的开放世界而很不幸的是，本作正正好在她关注的几个角度精准的踩中了地雷：虽然相较于《合金装备 4》有所收敛，但游戏内不断插入冗长的过场 cg，以至于从室外场景进入到地图传送需要连续跳过数个动画；游戏流程放速过缓，且游戏内的任务设计被认为千篇一律，主线和支线均由送货构成；动作系统相较于前作乐趣有所减少（尽管二者的游戏定位有着根本性的不同，也许从来就不该这么要求角色的性能），击败 bt 后的奖励反馈也较为缺乏，以至于编辑出于他所偏向的几个维度的也许从传统游戏性的角度合乎情理的考虑，选择给这个游戏 6.8 分。随后我们都知道发生了什么——之后的几年让游戏的异步社交大放异彩，以一个非常奇怪且迅速的方式，让小岛秀夫这种几乎像是寻求争议，找骂、试图以先锋艺术电影的形式在游戏内进行的表达，能被更多人看见，被更多人共情，似乎小岛秀夫在《合金装备 2》后，再一次，等到了时代来验证他思想上的前瞻性。因而，时代背景变迁下大家对小岛独特表达的广泛接受、小岛秀夫的明星制作人光环以及评测中确实存在的一些过当的指责，都让玩家社群反过来开始质问一件事：为什么我觉得好玩的游戏，你却要给出这么低的分？当然，IGN 作为行业头部媒体，其本身在游戏界有较大的声量较高的影响力，且也有着个人做自媒体乃至中小型媒体难以企及的资源，因此它说的话做的评测理应被玩家更加严格的审视，树大招风不无道理——而且至少在我举得几个例子里，引发争议的游戏本身也是在不同领域内质量过硬的存在，只是随着营销的不断发展，疫情时代与后疫情时代每个亚文化小领域的的扩圈与自媒体时代的到来，这种原本对媒体对权威的的尚处合理范围的不信任将逐步扩大，直到这种现象一步步扩散到每一件事，每一个领域，乃至我们每个人都难以避嫌，难以无视。Part2：内容井喷的时代，同质化的幽灵与信任的坍塌数码圈从来不缺学生出身，单纯出于热爱，分享自己体验的自媒体，这样的评测形式本身并非没有价值，它用更加平易近人的角度去涵盖一个产品中用数据评测所难以涵盖的部分，尽管也许，分享自己对于刚入手的数码产品的主观体验并没有什么门槛——有着中专及以上的学历就有足够的语言组织能力去清晰讲述自己的感受，再来一个相机甚至是手机去拍摄素材就能实现，但撰写一篇好的主观体验评测文（不管是哪个领域）就像是一般人眼中的历史研究一样，都觉得准入门槛低，都觉得写的内容是任人装点任人阐释的小姑娘，但真的想要尽可能做到内容详实、条理清晰，往往需要长时间的写作经验积累与大量的训练，而如果想在这个基础上再去追求与社区中主要讨论的内容截然不同的全新角度，或者让我们思考的更加狂野一些，在这个内容上做出深度，探讨一个产品背后更加深刻的有广泛意义的命题，则更加考验个人的水平——但随着互联网的高速发展与信息的高速创作、流通，我们正处在一个内容极其过剩的时代。虽然我承认，我是随着酷市场转型为酷安而踏入数码圈的，本身并不算什么数码区老资历，但你如果要我去回忆在我进入数码圈后这有限的时间内，主观体验向的数码自媒体井喷的那段时间，我想大概是在「老师好我叫何同学」的 5G 视频兴起的 2019 年 6 月前后，以及疫情肆虐的 2021 年后到 2023 年末。前者因为正好切合了国家的 5G 发展战略的需要而被主流媒体报道，被社会广泛关注，因而给圈子引来了泼天的流量——事实上那段时间坚持做手机体验类评测的规模还比较小的自媒体，如今都已经成长为几十万粉丝的台柱，比如我当时看的比较多的花生说；而后者则是由于人们长时间居家，在室外的娱乐方式因客观需要而受限，因而人们开始转向互联网，大约是在这一时期前后，一群赛道更加细分，比如上古数码产品体验、软件交互等领域（就我当时看的比较多的来说是六分超超和苏星河）的体验向自媒体开始大规模兴起，游戏区的很多如今声量较大的自媒体编辑老师也多是在这一时期进入相关行业或者是投身做自己的账号。但与自媒体账号井喷相对应的，却是——请让我想想，该怎么让人更好接受的去讲一个有点不近人情的结论——极其迅速的内容同质化。游戏不是一个竞争激烈且产品间带来的体验差别相较于同类型的数码产品更大的存量市场，有足够多各自差别足够明显的新游戏不断冲刷着视野，因此这种同质化的倾向也许还并不明显，但在数码区，当每个上游的厂商都在绞尽脑汁的结合热门概念，生生在软硬件体验上捏出极其有限的新创新点时，这种感觉就尤其明显——对于自媒体而言，相比起「你自己要讲述什么故事」，你的阐述对其他人有什么价值才更有传播的意义，因而当你迅速的因不受束缚、观点尖锐，获得了一批受众支持迅速爆火后，同质化的幽灵便也开始悄然显现——是顺着原先让你得到人们簇拥的路径一路向西，对着即使没有那么差的内容输出情绪，直到有一刻在无数次路径依赖后判断失误，犯了众怒，还是妥协于平台激励有限而选择靠近厂商，向其他市面已有的无数自媒体靠拢，在厂商给的通稿框架与宣传的大方向已然确定的情况下，进行着极其容易趋同的表达，在首发宣发节奏与内容合作模式影响下，选择性的淡化掉体验中的痛点，即使在前文有提及也往往占据相对较小的篇幅，并随后立马引入一个「但是」，用着首发特调的机型说着好到梦幻的体验，最终在观众的不解中继续在内容理想与工作室运营中艰难的走着钢丝？当然，如果把话说到太绝，往往会有很多偏颇，所以让我们稍微说一些让文章作者更拟人的

「智能队列」如何改善你的家庭网络质量主作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org联合作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org2021/01/31 23:12前天 12:57更新：2026 年世界电信和信息社会日重推在无线路由器的设置页面上，我们可能会看到一个名为「智能队列」的功能（根据品牌或操作系统的不同，也可能会被称做「智能 QoS」、「SQM」等）。这一功能是如何工作的？是否能让网速变快？是否需要打开？本文将尝试解答这些问题。从「网络如何连接」说起本章节主要介绍计算机网络的基础知识。如果读者已经熟悉相关知识，可跳过这部分内容。在介绍「智能队列」之前，先来了解一下基础的网络知识。两台电脑间进行数据传输，存在两种方式，分别是电路交换和分组交换。对于电路交换，两台电脑独占一条线路进行数据传输。过去的固定电话，大多使用的就是电路交换。由于独占一条线路，两台设备之间的通信较为稳定，很难出现数据丢失、数据传输顺序错乱等情况。但一条线路只能供一组电脑使用，另外一组电脑想要通信，就不能复用原来的线路了。对于分组交换，传输的数据被拆分为小段，每一小段叫做一个数据包（报文）。数据以报文的形式进行转发，从而实现了多组电脑复用一条线路进行通信。电路交换虽然能保证数据传输的稳定性，但由于独占线路浪费太多资源，所以在 Internet 中，使用的是分组交换。分组交换将数据拆分成一个一个的小片段来转发，没有独占线路。如果网络条件不好（例如同时有大量报文经过路由器，超出了路由器的处理能力），很有可能遇到数据顺序错乱、甚至数据丢失等情况。那么，这些情况应该如何避免？在微信视频通话、电视直播等应用中，报文的丢失或乱序，只会造成视频图像卡顿，不会造成太严重的影响。所以这些应用会使用简单的UDP协议，不考虑丢包、乱序。在网页浏览、文件下载等应用中，一般使用基于TCP的HTTP协议。TCP 在报文中携带序号，避免乱序；同时提供了报文的丢失重传机制，避免了丢包。所以，也可以认为 TCP 为两台电脑的应用程序建立了一个虚拟的「连接」。（最新版本的 HTTP/3 协议，使用的则是QUIC。QUIC 虽然基于 UDP，但引入了与 TCP 类似的报文丢失重传机制）如果想要了解更多网络基础知识，推荐阅读「入网指南」系列文章：入网指南 01 | 一文读懂你身边的「网络」入网指南 02 | 当你发消息的时候，发生了什么入网指南 03 | 交换机在交换什么？入网指南 04 | IP 地址大揭秘入网指南 05 | 除了接线，还有什么影响着我们网上冲浪入网指南 06 | 连上 Wi-Fi 就能上网了吗？网络延迟：影响上网体验的重要因素带宽和延迟，是影响上网体验的两大因素。带宽指的是在单位时间内，网络最多能传输多少数据。网络的带宽和水管的粗细有着很多相似之处，水管越粗，相同时间流出的水越多；同样地，网络带宽越大，相同时间能够传输的数据越多。在日常上网过程中，大的带宽能带来这些好处：上传、下载文件的速度更快在线看视频时，缓冲的时间更短如果家中有多个人同时上网，大的带宽能够保证每个人都能流畅上网，相互的影响更小对于家庭宽带，我们常常听到的 10 兆、50 兆、100 兆、千兆，就是指的带宽的大小。延迟指的是数据从网络的一端到达另一端，所需要的时间。过大的延迟也会对上网体验带来不好的影响，例如：在浏览器中输入网址敲回车后，需要等更长的时间，网页才能加载；视频聊天、语音通话质量变差，甚至能直接感受到声音、画面的滞后；玩在线游戏时，发现游戏变慢变卡。在很多时候，延迟相比带宽更容易被人感知，对网络体验的影响更大。是什么导致了网络延迟？上文中介绍过，Internet 使用的是分组交换。如果大家阅读过计算机网络相关的书籍，应该对分组交换的「存储-转发」机制有一定的印象。简单说，「存储-转发」是指网络设备将收到的报文存储在缓冲区，然后再发送。这是由于网络设备的性能有限，转发报文有速率限制。如果设备短时间内收到大量报文，一时半会儿处理不过来，就可以先报文存起来，空闲时慢慢发送。通过「存储-转发」，能够提高网络设备的吞吐量，更充分地利用带宽。其中，存储报文缓冲区，是一个先进先出的队列。那么，队列的长度应该如何确定？是不是队列越长，能够存储的报文越多，网络的性能就越好？大家可以先思考下这个问题。其实在过去，内存价格比较昂贵，考虑到成本，网络设备的队列不会太大。现在，随着内存价格的下降，网络设备拥有了更大的队列。但是，网络的带宽是有限的。实际的可用带宽，取决于数据传输路径上的最小带宽。这个最小带宽也被称做带宽上的「瓶颈」（bottleneck）。由于瓶颈的存在，如果流量长时间以较大速率发送，大的队列并不能显著提升吞吐量。相反，大量报文滞留在队列中，会带来延迟的增大。这种队列中缓存有大量报文，来不及发送，导致数据传输延迟增大的现象，被称做 bufferbloat（缓冲区膨胀）。DSLReport 提供的测速服务，能够检测网络的 bufferbloat 程度。不过由于没有中国大陆服务器，测试结果不够准确。国外的读者可以尝试一下：DSLReports Speed Test家用无线路由器中的 QoS 功能在「智能队列」出现之前，很多家用路由器已经提供了 QoS 功能。那么，传统的 QoS 如何解决网络延迟、网络拥堵？存在哪些不足之处？在了解这个问题之前，先来熟悉一下网络中的报文，是什么样子的：在网络中传输的报文，一般包括了链路层、网络层、传输层、应用层这四层信息。其中，网络层的源 IP 地址、目的 IP 地址、协议号，和传输层的源端口号、目的端口号，是 QoS 算法中常常会关注的五个信息。这些信息合称「五元组」。同一个应用在同一时间段内，传输的报文中，这些信息一般是不会发生变化的。例如观看在线视频时，同一个视频的所有报文 ，一般拥有相同的五元组信息。所以，通过五元组，我们能够区分不同应用的流量。特别地，一些常见的应用，会使用固定的端口号。例如网页浏览一般使用 80、443 端口；英雄联盟游戏，使用的是 5000-5500 端口 ……所以，当我们想要提升英雄联盟游戏的体验时，在路由器上设置让 5000-5500 端口的报文优先通过即可：How to set up traffic prioritization rules for League of Legends on Nighthawk Pro Gaming router | Answer | NETGEAR Support家用路由器中常见的 QoS 功能，就是通过手动设置 IP 地址、协议号、端口号等报文特征来实现的。这种方式需要用户了解一定的网络知识，且操作复杂。（虽然部分无线路由器内置了常用应用的 QoS 规则，但仍需要一定的手动设置，且无法及时适配新的、小众的应用）当然，除了上文中介绍的内容，QoS 还有多种不同的实现方式。例如：对特定应用（例如 BitTorrent）进行限速按照 MAC 地址指定某台设备的优先级（例如提高 Nintendo Switch 的优先级，使游戏更加流畅）通过WMM、DSCP等机制来决定报文优先级这些机制各有各的特色和不足。例如对应用程序限速的方式，虽然效果明显，但无法在网络空闲时，动态利用剩余带宽来提升传输速度，也难以适应带宽频繁变化的无线网络。关于上述 QoS 机制的技术细节，本文不再详细介绍。如果感兴趣，可以搜索相关关键词，进一步了解。智能队列：通过「全自动」的方式优化网络、降低延迟什么是「智能队列」目前，一部分新的无线路由器，开始有了名为「智能队列」的功能。这一功能一般指的是 fq_codel。fq_codel 包含两种队列管理机制，分别是 Flow Queue 和 CoDel。在下文中，我将简要介绍这两种机制的作用。Flow Queue前文中介绍过，网络中的报文一般都有自己的五元组信息；同一应用发送的报文，在一定的时间段内，五元组信息是相同的。这些信息相同的一系列报文，被称做一条流（flow）。Flow Queue 为每条流单独维护一个队列。例如，如果我们一边在线观看视频，一边进行语音通话，一边玩游戏，Flow Queue 就会为视频播放的流量、语音通话的流量、在线游戏的流量各维护一条队列：正常情况下，在线观看视频需要占用较大的带宽，但对延迟不敏感，用户可以容忍几秒甚至几十秒的缓冲时间。而语音通话和在线游戏占用带宽小，但对延迟敏感。例如语音通话时，如果一个人说了一句话，对方在几秒后才能听到，用户就能明显感知到延迟。如果三种应用共用一个队列，队列中会充满大量视频播放应用的报文，导致语音通话、在线游戏的报文迟迟无法发送，增大了这两种应用的延迟。Flow Queue 为视频播放、语音通话、在线游戏各维护一条队列，并公平地调度三个队列，避免视频播放的报文占用太多时间，从而让在线游戏、语音通话的报文得到及时处理。由于 Flow Queue 确保了应用程序之间的「公平」（fair），所以又被称做 Fair Queue。CoDel我们已经知道，适当长度的队列，能够充分利用带宽、提高吞吐量。但过长的队列，又会导致过大的延迟。也就是说，队列既不能过长、也不能过短。CoDel 所做的，就是让队列维持在合适长度，最大限度地提升网络性能。CoDel 通过计算报文在队列中的停留时间，来判断队列是否过长。具体做法如下：报文进入队列时，在报文上记录当前时间（也就是报文入队时间）报文出队列时，再次获取当前时间（报文出队时间）计算报文入队时间与出队时间的差值，即可得到报文在队列中停留的时间当报文在队列中的停留时间超过阈值，说明队列过长，需要缩小队列长度当队列过长时，CoDel 会直接丢弃出队列的报文，从而减少队列中报文的数量，达到控制队列长度的效果。（这种对报文的丢弃，并不会对网络带来负面影响，具体请见后续章节）可以看出，CoDel 根据队列中报文的停留时间，来决定是否丢弃报文。而不是「一刀切」地将队列限制在一个固定的、较小的长度。这样既能降低延迟，又保留了队列最大的优点：吸收短暂的大流量报文、提升吞吐量。另外，我们的设备，尤其是手机等移动设备，所处的网络环境是复杂多变的。例如手机靠近或远离无线路由器，就有可能导致网络带宽的变化。CoDel 通过计算报文停留时间的方式控制队列，能够更加灵活地适应网络带宽、网络质量的变化。fq_codelFlow Queue 确保了不同应用程序的流量能够公平传输。CoDel 控制队列长度，实现了整体延迟的降低。fq_codel 就是将 Flow Queue 和 CoDel 结合在一起使用的机制。其中，Flow Queue 先将不同的流划分为不同的队列，再为每一个队列使用 CoDel，控制报文的延迟。两者结合使用，共同提升网络体验。fq_codel 具有免配置、天然适应不同的网络环境、实现简单、消耗资源少等特色，越来越多的无线路由器（以及开源的路由器操作系统），开始内置这一机制。参考 KoolShare 上的这篇帖子，作者经过测试，发现打开 fq_codel 之后，在 BT 下载占满上行带宽时，在线游戏依旧保持较低的延迟：可能是家庭路由里最好的 QoS——在线游戏者的福音 - KoolShare（见 8 楼、10 楼、22 楼、25 楼的测试结果）是否需要打开「智能队列」？前面介绍了 fq_codel 的一系列优点。那么，fq_codel 有没有缺点？是否需要我们打开？我的设备 CPU 性能不强，是否可以打开「智能队列」？根据这篇帖子，ER-X路由器在打开 fq_codel 时，最大支持 200M 左右的带宽。QoS (fq_codel/cake) Hardware Recommendation : HomeNetworkingER

周一18派早报：提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年昨天 06:56早报速览提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年0OpenAI 与苹果关系紧张，或考虑诉讼5索尼回应 AI 相机助手越改越差，称只是修图建议2Anthropic 警告私下交易的股权可能无效1运营商回应「零月租」传言4微软重新允许自定义任务栏与开始菜单2看看就行的简讯4少数派的近期动态0你可能错过的好文章0提交未经核实 AI 内容的论文作者将被 arXiv 封禁一年5 月 15 日，预印本平台 arXiv 出台新规：如研究人员向该平台提交的稿件包含「无可辩驳」的 AI 生成内容，论文所有署名作者将被处以长达一年的封禁；且其未来提交的任何文章都必须先在其他期刊通过正式的同行评审后，arXiv 才会予以收录。「无可辩驳」的证据包括虚假参考文献、不小心粘入的与 AI 对话内容等。俄勒冈州立大学名誉教授、arXiv 编委兼审核团队成员 Thomas Dietterich 在 X 上公布了这一消息。他指出，新规直接源于 arXiv 既有的审核标准。无论稿件中出现不当言论、抄袭、学术偏见、虚假参考文献还是误导性内容，所有署名作者都必须对这些 AI 生成的错误承担全责，而不能将责任推卸给工具。arXiv 是物理学、天文学、数学及计算机科学等领域全球最大的开源学术文献库。在这些学科的常规发表流程中，学者们通常会在接受期刊正式的同行评审前，先将论文作为预印本上传至 arXiv，以抢占首发权并获取同行反馈。因此，剥夺一年的首发资格并永久附加同行评审门槛，对相关领域人员而言是严厉的措施。323OpenAI 与苹果关系紧张，或考虑诉讼据彭博社报道，苹果与 OpenAI 为期两年的合作关系近期出现裂痕。因未能从 ChatGPT 与苹果系统的整合中获得预期收益，OpenAI 正与外部律所积极探讨应对方案，近期或将向苹果正式发出违约通知，为潜在的法律诉讼做准备。2024 年 6 月，苹果宣布将 ChatGPT 接入 Siri 等系统级功能。OpenAI 曾寄望借此获取海量订阅用户及数十亿美元收入。然而实际体验中，该功能入口较深，需用户明确口头提及 ChatGPT 才能唤醒，且信息呈现方式受限。结果，绝大多数用户仍倾向于使用 ChatGPT 应用。一位未具名 OpenAI 高管表示，苹果在产品端甚至没有尽到诚实努力的义务；合作不仅未能创收，反而损害了其品牌形象。除软件合作遇冷外，两家公司在硬件与人才领域的博弈也加剧了紧张局势。OpenAI 此前收购了由前苹果首席设计师 Jony Ive 等人主导的初创公司，并利用高薪从苹果硬件团队挖角，试图打造新一代 AI 设备，这引发了苹果高层的强烈不满。同时，苹果也已着手引入竞争机制，计划在即将发布的 iOS 27 系统中全面开放生态，接入谷歌 Gemini 和 Anthropic 的 Claude 等其他主流大模型。知情人士表示，OpenAI 针对苹果的任何法律行动，预计都会推迟到其与马斯克的诉讼案结束后才会推进。目前各项计划尚未做出最终决定，OpenAI 内部仍保留了在庭外解决争端的希望。OpenAI 与苹果的发言人均拒绝就此事置评。我换到国产手机AI都没怎么用过，不是不会用，而是习惯了以前苹果装个AI软件的使用方法了。果子在ai上的每步操作都很迷惑感觉不如自己装一个本地小型LLM App用ShortCut打通生态。我还是关心什么时候有苹果智能，感觉一不留神，好几年过去了，我换15p的初衷之一就是给智能做准备，现在都要17了，我也是吃上大饼了622索尼回应 AI 相机助手越改越差，称只是修图建议近日，索尼最新旗舰 Xperia 1 VIII 的官网页面引起关注。在演示 AI 相机助手的效果时，官方样张的效果在观感上反而明显劣于原图。这引来了 MKBHD、Nothing 创始人裴宇等人的公开调侃。对此，索尼在 5 月 15 日解释称，「AI相机助手」并不会在拍摄后自动应用修图，只是根据当前场景和拍摄主体，提供 4 种不同创意方向的参数设置建议，用户可自由选择接受这些建议或坚持使用自定义设置。与此同时，官方最新展示的对比样张在视觉效果上有了显著改善，去除了首批样张中饱受诟病的画面发灰和过度曝光问题。但这依然未能完全消除外界对其影像审美判断的疑问，也不清楚营销团队最初为何会挑选那些效果不佳的照片来展示核心卖点。感觉那朵花还可以，其他的都是💩️有人关注就算不错了9101Anthropic 警告私下交易的股权可能无效据《华尔街日报》报道，Anthropic 近日更新官网公告，宣布将不承认未经其董事会批准的二级市场股权交易，并点名了一系列未授权机构。Anthropic 警告称，由这些机构提供的任何该公司股权或相关权益转让均属无效，且不会记录在公司的股东名册上。这一声明在个人投资者中引发了广泛恐慌。由于普通投资者通常无法直接购买未上市「明星公司」的股权，大量二级市场交易平台应运而生。它们设立特殊目的实体（SPV），将众多散户资金打包后投资于目标公司。然而，这种 SPV 结构日益复杂，有的多层嵌套，有的把多家目标公司的股权捆绑出售，导致分配机制存在很大模糊地带。普遍认为，Anthropic 此举意在通过清理违规交易来精简其股东名册，为其计划于今年晚些时候的 IPO 做准备。随着近年来私营科技公司的估值不断飙升，包括 OpenAI 在内的多家独角兽企业，都在采取类似措施限制涉及其股权的二级市场交易。面对投资者的质询，多家被点名的交易平台迅速回应，表示涉及 Anthropic 的基金和交易均已获得批准或符合转让要求。专家指出，如果 Anthropic 并未在事前对特定交易结构做出限制，可能无权撤销这些持仓，相关争议最终或将诉诸法庭。这种模式在中国有么？发表评论运营商回应「零月租」传言据央广总台报道，近日，不少互联网博主在短视频里介绍，三大运营商即将取消月租费，改为按量计费模式，甚至还列出了推出该政策明确的时间节点。但记者梳理市场现状发现，目前只有中国联通上线了零月租、设最低消费门槛的「魔方」按量套餐；中国移动和中国电信暂未推出同类产品。例如，传言中被广泛提及的「联通魔方」发布于今年 4 月底。该产品采用纯按量计价模式，国内语音和短彩信均为 0.1 元/分钟或每条，流量则实行阶梯计价，单价在 1 元至 1.5 元/GB 之间。但联通客服表示，该套餐并非绝对免费，而是设有地区差异化的最低消费门槛，例如北京地区起步低消为 55 元，月结时不足该金额将自动扣齐。同时，联通原有的各档位固定月租套餐依然保留并可正常办理。中国移动和中国电信的客服人员则都表示，公司曾推出过零月租模式，但目前没有相关计划。针对完全零月租的可能性，通信行业专家指出，此类模式脱离了基础通信网络的运营逻辑，庞大的固定网络维护成本最终也必然会转嫁到其他用户身上，甚至可能导致整体资费上涨。比起追求不切实际的零月租噱头，当前消费者对资费的真正痛点在于不合理的老旧月租套餐，以及套餐外的标准流量资费。没有专门办的流量卡划算捉虫：“央广总台”→“总台央广”“零月租” “起步低消”世界需要提速降费。3微软重新允许自定义任务栏与开始菜单5 月 16 日，微软发布博客文章，宣布将在未来几个月内推出一系列更新，大幅增强任务栏和开始菜单的外观自定义功能。这标志着一系列经典 Windows 功能的回归。对此，Windows 业务负责人 Pavan Davuluri表示，个性化定制是 Windows 的基因；微软已听取用户反馈，决定重新将界面控制权交还给用户。目前，部分选项已在 Windows 预览体验计划中面向测试人员推送。本次更新中，用户可以再次像在旧版系统中那样，将任务栏停靠在屏幕左侧，或将其缩窄以节省屏幕空间。同时，开始菜单新增了尺寸调整选项，并允许用户自定义显示模块，例如直接关闭「所有应用」或「推荐」（更名为「最近使用」），仅保留自己固定的应用。此外，用户可以隐藏「开始」菜单中的用户名和头像。这些测试功能目前仍有已知问题，例如开始按钮未对齐、通知弹出位置异常等。微软表示将在正式向公众推送前修复这些问题。此外，更多传统的自定义功能，例如拖拽任务栏边缘手动调整高度等，目前也已纳入开发计划，预计将在今年内陆续上线。Windows 11 自发布以来，就因移除任务栏位置调整等诸多经典自定义功能而饱受争议。近年来，微软一度将系统开发重心全面转向 AI，在系统中强行植入 Copilot 功能，引发了用户的强烈抵触。迫于压力，微软于今年早些时候宣布做出调整，重新关注质量。什么时候把win10菜单给我回归一下😊我选择Windows Server 2025，躲过UWP和React两大石山。1看看就行的简讯据Fast Company报道，知名开源密码管理器 Bitwarden 近期在未发公开声明的情况下，悄然大换血高管团队，今年 2 月起，原 CEO、CFO 先后转岗或离职。新任 CEO 有并购及私募股权交易背景，让外界对其未来的资本运作产生猜测。此外，Bitwarden 官网出现多处调整，个人版产品页面移除了「永久免费」的标签（官方解释为营销团队疏忽）；不久前，Bitwarden 在未显著通知的情况下，将高级版订阅年费从 10 美元翻倍至 20 美元。此外，Bitwarden 还在招聘页面修改了企业核心价值观，去除了原有的「包容」与「透明」两项原则。近日，谷歌开始测试一项新的云存储空间政策：新注册的用户如果不绑定手机号，其初始免费存储空间将从传统的 15GB 缩水至 5GB。谷歌官确认，该政策目前仅处于区域性测试阶段，且现有老用户不受此规则影响。谷歌发言人表示，这项新规是为了继续为用户提供高质量的存储服务，同时鼓励用户完善信息以提升账户安全性和数据恢复能力。然而，许多用户推测谷歌此举的实际目的是为了打击恶意注册大量小号以薅取免费存储空间的行为；也有部分用户对强制绑定手机号背后的隐私数据收集及潜在商业用途表示担忧。据9to5Google报道，谷歌随无屏健身追踪器 Fitbit Air 推出的订阅服务「健康教练」（Health Coach，由原 Fitbit Premium 升级而来）存在 AI 幻觉问题。测试中，AI 教练凭空捏造了一项 5.2 英里的跑步记录；在被指出错误时，还反问用户是否忘记输入过该次记录。经过纠正后，这条虚假数据依然残留在应用主页上。最近对bitwarden挺失望，要么是性能问题，这次又出个管理团队问题。开源不是万灵药，信誉才是基石啊这几年一直用Bitwarden 呢，要变味了吗白赚一次跑步，哈哈哈哈比特瓦恩德可以换了(｡･ˇ_ˇ･｡:)4少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购发表首评你可能错过的好文章看剧不卡顿、游戏低延迟：「智能队列」这样改善你的冲浪体验为什么我觉得好的东西，媒体评测却给了低分？本周看什么 | 最近值得一看的 8 部作品装了啥：编辑部用什么转写音视频内容？[会员免费] Apple WWDC26 全球开发者大会，来与少数派一起看快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验你真的需要墨水屏设备吗？发表首评80扫码分享目录 0讨论我来说一句发布发表评论发布

2025 年度旅行——4 天 3 晚青海湖小环线主作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥联合作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥05/14 09:38Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。一次完美的旅行，对我而言，不仅仅是去被人走过千遍的地方打卡一圈，而是藏着三次截然不同的体验——出发前对目的地的想象与期待，旅途中与自然的亲密接触，以及归来后翻看照片时，重新发现那些当时未曾留意的细节，细细回味，查阅资料，让我对这次旅行有更深刻的理解。那么就让我们随着这些文字与照片，回到美丽的青海湖。契机：作为刚工作的年轻人，在国内的职场环境下，真正无理由可申请的假期只有 5 天，所以我和身边的小伙伴们都分外珍惜这五天的时光，每年的年度旅行我们都会避开人流，详细规划。由于十一假期要去参加两场婚礼，为了节约交通费，我们决定先开车到 2 号参加婚礼的宝鸡，把小 P（我的小鹏 MONA M03，具体可参阅我之前的文章）停在了宝鸡高铁站，考虑旅行时间较为紧凑，且西北自然景区交通补能条件可能较差，我们选择了到当地租车完成青海湖小环线行程。带着宝鸡擀面皮和肉夹馍，一行人正式开始了旅途，三个半小时后，到达了西宁火车站，初入高海拔地区，并未感受到明显不适，方便的是，租车平台在西宁站地下停车场就有取还车点，我们直接将行李放上车，省去了搬来搬去的劳顿。Day1 西宁-茶卡 全程约 457 km西宁-塔尔寺5A 级旅游景区，门票：70 元，游览时间约 2 小时塔尔寺是藏传佛教格鲁派的著名寺院之一，始建于明代，最初是为纪念藏传佛教格鲁派创始人宗喀巴大师而修建，据传宗喀巴大师出生于此地，后人于其诞生处建塔纪念，塔尔寺因此得名——「塔尔」即「有塔之地」。塔尔寺-八宝如意塔游览中心和寺庙门口都有大量的讲解人员，价格也不贵，但是本人信仰唯物主义，故只将这里当做名胜古迹参观，故没有请讲解，如果对藏传佛教及寺庙背后的故事比较感兴趣，请个讲解或许是更好的游览方式，或者是像我一样，回来翻阅照片时再仔细搜索相关资料。Tips：讲解会介绍各个佛像的「作用」（不知道这样叫是否合适，或者说负责保佑的事情），故可能会被架在那里为了讨好彩头而「必须」捐献资金，介意者可以像我一样自由游览。由于是节前，故参观人员并不很多，与前往朝拜的信徒或僧人可以说是一半一半，没有了熙熙攘攘的人群，也得以让我更加细致地观察这座寺庙和「磕长头」远道而来的人们。前来朝拜的僧人进入寺庙，浓郁的香火味和木质结构因缺乏通风的陈味扑面而来，看着宏大的佛像、墙上摆放整齐的诵经坐垫，的确能感受到不一样的文化氛围，这里安静、鲜活，尤其是看着不知道从哪里启程的「磕长头」的信众，虽因路途的遥远衣衫褴褛，但想必也经过这一路的艰辛而悟出了些许道理，可能出发时也不知道为什么，但走着走着就明白了吧。塔尔寺中的白塔募捐处，与时俱进地，也支持扫码支付塔尔寺有大大小小的寺庙很多间，在我看来有的都别无二致，于是重点转了几个后，就启程回西宁吃饭，整备氧气瓶，应急食物等装备，正式奔赴青海湖。海南藏族自治州共和县 南山观景台 & 青海湖二郎剑景区前者免费，后者为 5A 级旅游景区 门票 72 元，游览时间约 2 小时一点半在西宁市吃完炕锅后，我们开始驱车前往青海湖，约 2 个小时车程，我们到达了本地的主要目的地青海湖的南侧景区 青海湖二郎剑，在进入二郎剑景区前，可以导航到「南山观景台」 从山上远眺青海湖和远处群山的全貌，但是山路较窄且有工程车辆时不时通过，故要格外注意安全。景区本身没有人文景观，但其还是我国早期鱼雷的试验基地，但是因为我们到达时间比较晚，并没有参观博物馆，而是直奔到湖边。青海湖是我国水面面积最大的咸水湖，水面平静辽阔，与远处的群山相辅相成，甚是开阔美丽，栏杆上有硕大干净的鸽子驻足，在几番伦理拍照后，我靠在栏杆旁，看着远处的夕阳洒在地面，耳边传来阵阵鸟鸣与风声，工作与生活的喧嚣远离脑后，只剩下全身心的与自然接触。云层、远山、草原，形成了震撼的三明治拨云见日，青海湖边的草还是绿色的岸边排队等待「薯条」投喂的海鸥码头景观随着夜幕降临，我们离开二郎剑景区，沿着青海湖开车前往茶卡住宿，虽然只有 200 多公里的行程，但是因为沿途工程车辆很多，还有很多地方持续修路，有的路段有连续的 U 型弯，还有双向仅有一条车道通行的地方，还要翻越几个海拔 3800 米的垭口，成了这一路最惊险刺激的行程，一路上小心翼翼，战战兢兢，终于在晚上 10 点到达了茶卡镇，幸运的是镇上的餐馆都还在营业，找了家当地美食，简单吃饭后休息。夜幕降临的倒湖茶公路Day2 茶卡-祁连县 全程约 372km茶卡盐湖4A 景区 带往返小火车套票 167 元「茶卡」在藏语中意为「盐池」，该地以出产「大青盐」而闻名，参观盐湖也是此地的知名旅游项目。茶卡盐湖的景区共有两个，天空之镜是较老的景区，主要以自然风光为主，天空一号主要以商业化项目为主，作为外地游客，我们肯定更希望看到原始的自然景观，于是选择前往天空之境景区。景区很大，如果想要游览每个景点建议留出一整天的时间，游览顺序建议先坐小火车到终点天空之镜，该站处于盐湖的最深处，盐水异常的干净，也最为出片，景区也放置了很多拍照打卡点位，可以尽情拍摄，心痛没有带无人机。在晴天的时候，由于天空和地面都是白色，很容易晃眼，推荐大家一定要戴墨镜前往，通过墨镜观察湖面，盐湖会看起来像翡翠一样更加绿色透明。套票包含鞋套租赁，可以套上下湖走走，踩在大颗的盐里,很像沙滩的感觉，景区维护的十分干净，可以尽情的亲近自然。有趣的是，景区为了怕大家在盐滩大片的白色中无聊，设置了很多拍照用的摆件，例如国家地理的相框等等，还在大屏幕上贴心的轮播拍照指导，期望每个游客都能满意而归。盐湖与远处的山峰盐湖码头废弃的采盐轨道之前采盐的轨道深入盐湖，在墨镜下，真实的观感颜色会比照片深很多。接驳小火车行驶在盐湖中间，干净得不像国内巨幅盐雕从盐湖出来，我们在县城稍作休息，开始启程前往祁连县，有了昨夜惊险的行程后，尽管沿途的国道可能风景更好，但是我们选择了高速较多的线路。这一程的路况还是很好的，就算是国道车辆也较少，道路笔直，视野十分宽阔，草原、远山、白雪、天空层层分明。在行驶过程中，不时的有大群的羊羔和牦牛穿过公路，不必惊慌，一会就有牧民骑着摩托驱赶，让它们加快穿越的脚步。公路与群山草原上大批的牛羊傍晚的草原，远处能看到夕阳洒在雪山上落日下的牦牛剪影在夜色中，我们抵达了祁连县，在县城中吃过一顿特色烧烤后，进入梦乡，准备迎来第三天的行程。Day3 祁连县-岗什卡雪山-西宁 全程约 320km祁连卓尔山景区门票 60 元 上山小巴 20 元 登山电梯 10 元这个景区的确是全程我最惊艳的景区了，近能看到瑞士小镇似的田园风光，中能看到美如画的丹霞地貌，远还能远眺远处的阿咪东索雪山（海拔 4695 米），而且在山顶可以拍跟雪山的合照，大气、干净、巍峨。自驾车辆可以停在景区停车场，然后乘坐接驳中巴约 10 分钟到达检票口。卓尔山的两侧，风景差别迥异，一侧是连绵的雪山，另一侧是仿佛瑞士小镇的丘陵和小镇。阿咪东索雪山卓尔山的另一侧，仿佛欧洲小镇丹霞地貌与草地交融最让我惊讶的是这个景区竟然还修建了方便爬山的观光电梯，10 元一位，大概有三段，果然是人性化啊。景区电梯中午约 12 点，我们离开景区，在祁连县的一家川菜馆简单用餐后，我们驱车开始前往本次旅途的最高海拔岗什卡雪山。在县城开车的过程中，等红灯时，我看到一位初中年龄的藏族小伙盯着我们开的岚图梦想家，眼中闪闪发光，不禁让我想起小时候的自己，看到喜欢或没见过的车也会驻足观看，充满憧憬。愿他长大后也能开上自己喜欢的车，能够做自己喜欢的事情。岗什卡雪峰景区免费 上山摆渡车 25 元岗什卡雪山坐落于青海海北州门源县境内，是祁连山脉东段的最高峰，被誉为「雪山之尊」，因为设施和难度都尚可，也被雪山爱好者们称为「人生第一座雪山」。主峰海拔 5254.5 米。游览模式同样为将车辆停在景区停车场，然后乘坐接驳车到雪山脚下，正式游览区域的海拔在 3800 米左右，建议随身携带一个氧气瓶，下车之后，可以看到湛蓝的连心湖，水来自于冰川融水，干净，清冽，与子洲植被的荒凉的深褐色形成鲜明的对比，十分美丽。连心湖到达时天气并不是很好，岗什卡雪峰被薄雾笼罩天然泉水，因还有行程，害怕拉肚子，没能如愿品尝一下离开景区后，雪山逐渐露出真面目，攀登还是留给下次吧沿着冰川的遗迹向上走，就能看到岗什卡雪峰徒步攀登的起点，景区也提供马匹，可以跟着向导继续向前，到七彩瀑布，因时间和个人体能的考量，我们浅尝辄止，在河道附近游荡些许时间后，便下山返程。临近傍晚，启程返回西宁市，一顿丰盛的海底捞过后，快乐的青海湖小环线结束。Day4 西宁市-青海藏文化博物馆-特产采购最后一天，伴随着蒙蒙细雨，我们打车前往青海藏文化博物馆，博物馆不大，纵使是十一期间人也不是很多，但内容着实丰富，满墙的唐卡和在电影里看到的藏传佛教僧人服饰，还有精致的贵族金器。在博物馆的文创区，我购入了一个小的转经筒，在回程的路上一直把玩，虽不信奉佛教，但也觉得甚是有趣。唐卡四王金杯嵌绿松石立凤金头饰博物馆文创-转经筒，38 元虽不算划算，但做工十分精致离开之前，为了给家人和朋友带些纪念品，我们选了一家超市类型的纪念品商店，购买了牛肉干，藏红花，枸杞，青稞酒等特产，虽说也不一定划算，但胜在明码标价，将行李箱装满后，顺利返程。后记作为 2025 年的年度旅行，本次 4 天的青海湖小环线我甚是满意。人少不堵车，行程不特种兵，在当地感受美食和壮美的自然风光，一路上与好友们分担开车，聊对景色的感受，互相拍照，点评食物，虽偶有几个工作电话烦扰，但总体悠然自在，放空身心，一切都是那么的安静，与自然深度的接触，与北京的喧嚣形成了鲜明的对比。对于出行方式，自驾无疑是最有的选择，这边深入自然，道路虽崎岖，但也属于铺装路面，且西宁市神州、一嗨等在高铁站都有网点，取送车方便，且自驾可以自由控制行程，根据游览时长方便自己调整。对于住宿，我们对酒店的干净程度略有要求，于是全程都选择了以华住会为目标的连锁快捷酒店，以下是所有天的酒店名称和十一节前的价格，个人觉得是十分舒适的。Day0 桔子西宁火车站省人民医院酒店 双床房 331元Day1 星程茶卡盐湖酒店 覆氧双床房 314 元Day2 宜必思尚品海北祁连卓尔山酒店 421 元Day3 桔子西宁火车站省人民医院酒店 双床房 412 元关于本次行程的花销，纵使算上从北京出发的路费，人均也仅 3k 左右，可以说是十分划算了，主要的开销就是租车和酒店。支出一览表关于卫生间，景区和路途服务器的卫生间都比较完善，因同行有女生，我们还准备了移动卫生间，但从我们的行程安排走下来，并没有用到，个人觉得要是按我们的行程游览也无需担心。下一站，出发！2025 年的年度旅行结束，2026 年也可以开始规划起来了，新疆还是海南，希望能远离喧嚣，享受美丽壮阔的祖国风光。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀139扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@北鸮© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#Matrix首页推荐#旅行13等 13 人为本文章充电扫码分享举报本文章别惹小炸毛一个学酥关注

打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试主作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.联合作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.昨天 15:00ARG 即侵入现实游戏（Alternate Reality Game），是一种基于现实世界，融合各种多媒体（如网站、电话、社交媒体等）技术的互动式解谜游戏。由于其与现实的高度结合所带来的沉浸感，一直以来被许多玩家所追捧。如果你在 2016 年玩过《守望先锋》，相信你对当时新英雄「黑影」的登场一定记忆犹新。从 2016 年夏季开始，一直到 11 月的暴雪嘉年华，暴雪通过在宣传视频、官网、社交媒体发布了一些隐藏着密码的内容，由全球玩家共同解谜，不但揭示了游戏世界观下的一些故事，更是顺理成章地将身为黑客的「黑影」带到大家的视野里。经过几个月的发酵，在暴雪嘉年华上大屏幕被「黑影」控制，随后开始播放英雄动画的时候，更是满足了玩家们期待已久的心。「黑影」的宣发是一场相当成功的 ARG，时至今日仍有玩家对其念念不忘。我本人也是 ARG 的爱好者，我非常享受这种打破虚拟与现实边界的游戏乐趣。前段时间，一款名为《大学生登山失踪事件》的网页解谜游戏突然火了起来。比起此前一些高度依赖技术知识的网页解谜游戏，《大学生登山失踪事件》在不需要任何 IT 知识背景的情况下，依然有着抽丝剥茧逐渐接近真相的推理爽快感。自此之后，小红书上越来越多的人开始制作这一类网页解谜游戏，并为它们赋予了 ARG 的名号。这些游戏流程往往不长，差不多 1-3 小时就能通关。因为是网页解谜游戏，只需要一个浏览器就能游玩，很适合上班时间摸鱼，游玩门槛非常低。这篇文章将会推荐一些这段时间我玩过的质量不错的游戏，虽然一些制作精妙的游戏确实带有 ARG 的影子，但其本质上仍是基于网页的解谜游戏。为了方便大家后续搜索，本文会将这些游戏统称为 ARG。大学生登山失踪事件🔗https://missing.shiroki-y.top/introduction.html在游戏中，玩家需要通过从弟弟衣物中发现的「福安菜家常馆」的线索为开端，逐渐深入一场黑色阴谋找到弟弟失踪的真相。游戏整体难度不高，但在游玩过程中最好能用纸笔或软件记录进度，这能够帮你清晰地理出游戏脉络。游戏的核心玩法其实非常简单，只需要阅读页面，找到页面中可疑的关键词，在页面搜索栏里进行搜索即可推进剧情，这其中也会有一些用户名、密码、口令、字谜等需要玩家进行推理和猜测。网页页脚处会标记当前页面进度，既可以方便定位，也能够明确线索的先后顺序。剧情方面，《大学生登山失踪事件》并不会在一上来就给玩家塞一些晦涩难懂的概念，而是从大家熟悉的普通网页开始。当玩家一步步抽丝剥茧逐渐发现事件背后的黑暗面时，令人不由得感叹于游戏的设计与巧思。同时，游戏内也藏有一些可供探寻的 ex 页面作为彩蛋，在解谜之外找到各种彩蛋也不失为一种乐趣。诡异广告调查事件🔗https://mp.weixin.qq.com/s/yeNYMZVRp6N-MESFJGk9rg这是一款与《大学生登山失踪事件》来自同一作者@銀河電燈_的游戏，世界观也与前作有关联之处，很是有趣。玩家将扮演一位被「优化」的人，因为在看剧时发现了某个已停播的节目广告，从而开始调查这个诡异的广告，并发现藏在节目与自己背后的神秘真相。与前作相似，这款游戏也将基于阅读与搜索推进剧情。在此基础上，它还加入了一些 ARG 元素，比如在街景地图中，找到现实里存在的文字招牌作为线索。并且在网页设计上也比前作更加成熟，贴合剧情上的时间设定，在网页上加入了许多彩蛋。剧情方面，本作有一定的惊悚元素。特别是在游戏过程中展示的一些内容可能会引起不适感，如果你对这类内容比较敏感，请务必提前做好心理准备。当然，游戏并没有 Jump Scare，更多是通过剧情与视觉设计所营造出一种的诡异氛围。三川米粉店🔗https://www.berrysignalscmfd.top/《三川米粉店》也是一款从餐饮店主页出发，探寻故事真相的游戏。玩家需要调查三川米粉店突然停业的原因，在调查过程中，玩家将逐步揭开这家米粉店背后的故事。本作的页面设计相对没有那么精致，但也融入了一些 ARG 的元素，部分谜题的答案并不会直接给出，而是需要玩家通过搜索引擎或其他工具进行查询才能获得答案。从剧情上来说，《三川米粉店》的设定会更加接地气一点，并且结局也会相对更温馨一点。比较适合对于惊悚元素不太感冒的玩家作为第一部作品来尝试游玩。永安温泉度假酒店🔗https://mp.weixin.qq.com/s/kKsnzFsLFwwm2nwv8DseAA与前面侧重解谜的作品不同，《永安温泉度假酒店》的解谜难度其实不高，反而更像是一种「规则怪谈」式的探索路径。玩家需要不断探索找到更高权限的员工账号密码，从而揭开温泉度假酒店背后的秘密。游戏在开始前需要玩家填写「报名信息」，在后续游戏中能够增强代入感。从游戏性的角度来看，这款作品的「解谜」爽快感或许没有那么强。但如果你对「规则怪谈」或者一些高概念的设定感兴趣，那么这款游戏依然是一个值得尝试的作品。404 工位消失事件🔗https://xsn1204.github.io/qmgame《404 工位消失事件》是一个特色鲜明的游戏，玩家将扮演一个实习生，接手其他人留下来的工作电脑，并从中发现一些未被处理的秘密。游戏在网页上复刻了一个形似 macOS 的「系统」，很大程度上增强了代入感。因为设定是在电脑系统，游戏中需要玩家在钉钉、微信等各种软件中挖掘线索，反而更有一种作为侦探在不断探索的感觉。但受限于网站形式和游戏体量，游戏中能够互动的地方通常都是为剧情线索服务的，与线索无关的地方大多都是不能点击互动的。而且游戏不能通过点击窗口左上角的关闭按钮回到上一层，而是需要通过浏览器的返回来实现。但这些小细节对于《404 工位消失事件》来说都算瑕不掩瑜，整体游玩体验还是值得肯定的。作者甚至还出了 2.0 版本，新的剧本进行了优化提升，如果对这一类型感兴趣的话也可以尝试游玩。找到我🔗https://www.kaixinmoyu.com/feijo/findfeijo/在游戏中，玩家需要扮演一位私家侦探，调查游戏作者失踪的原因。游戏在交互方面做得非常有意思，游戏不仅做了浏览器页面对线索进行搜集检索，还做了小灵通和笔记本，前者可以通过拨打特定的电话获取线索，后者则需要玩家阅读日记找到隐藏的内容。《找到我》本身是一个更重交互而非推理的游戏，游戏中融入了一些 ARG 元素，在最开始就需要到某篇公众号文章中找到一个密码才能继续推进。但游戏的剧本逻辑会稍微有些不够连贯，以至于中间可能会有一些容易卡关的地方，需要仔细翻阅此前出现的线索才能继续往下推进。另外，游戏中后期会有一道需要在现实中解算的谜题，这道题本身与游戏本身的节奏设计会稍显割裂。但整体来说，《找到我》依然是值得一玩的游戏，尤其是本身的交互设计和 ARG 元素的融合做得非常不错。在游戏作者的个人主页，也能找到作者开发的更多网页解谜游戏，如果你对这款游戏的风格感兴趣，也可以尝试其他作品。小结优秀的网页解谜游戏并不止文章提到的这些，光是在小红书上一搜，就能找到许多分享帖。希望本文能起到抛砖引玉的作用，欢迎大家在评论区分享自己玩过的，令你印象深刻的网页解谜游戏。> 简单、好用、专注的写作软件，少数派为你呈现 🚀> 守护孩子的天马行空，从零开始美术教育启蒙🎨314扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Lotta© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#解谜游戏#游戏#生活方式31等 31 人为本文章充电扫码分享举报本文章TpBe nothing, and everything.关注

今年一季度，市场波动较大，保险资金是否减仓了股票？答案是“否”。金融监管总局近日公布的数据显示，一季度末，人身险公司和财险公司合计持有股票市值为3.84万亿元，较年初增加1024亿元，股票在险资投资余额中的占比也继续小幅提升。此前业内判断会实行逢低买入的操作，至此也得到了印证。随着4月以来市场有所回暖，险资机构的投资收益也有了明显回升。权益资产属于险资配置的“关键少数”，保险机构对权益投资的策略是积极配置并动态调整。（证券时报）

5月19日，渣打集团阐述可持续增长计划，目标于2028年实现高于15%的有形股东权益回报，较2025年提高超过3个百分点，并将于2030年提高至约18%；于2025年至2028年实现高双位数（high-teens）每股盈利复合年增长率及5-7%的收入复合年增长率；于2028年前提高每名员工所创造的收入约20%，并在2030年前削减企业职能相关职位超过15%。（界面）

36氪获悉，中信建投研报表示，算力需求持续强劲，尤其是光连接，NPO有望先于CPO规模部署于Scale up，持续看好人工智能板块。 专家认为，从可插拔光模块转向NPO/CPO的核心原因是Scale-up网络对带宽密度的需求，更低的功耗和更高的可靠性是NPO/CPO的另外两大重要优势。NPO可以依赖于成熟的可插拔光模块生态系统，同时在带宽密度和功耗方面提供显著改进。NPO是当前大规模部署中最实用、最具成本效益的解决方案。腾讯计划2026第四季度推出3.2T NPO解决方案，阿里的3.2T NPO已进入测试阶段。

在经历一季度规模大幅缩水后，银行理财市场于4月迎来强劲反弹，单月新增规模达2.6万亿元，超出市场预期。其中，“固收+”产品凭借收益回暖获得投资者青睐，成为助推本轮规模增长的主力军。多位受访专家认为，银行理财规模增长态势有望延续，但本轮扩张主要依赖债市估值修复与存款回流，而非趋势性逆转，收益改善持续性有限，未来产品创新与渠道拓展或成为理财机构竞争的主战场。（上证报）

当地时间5月18日，马斯克旗下的AI公司xAI宣布在网页端、iOS和Android平台同步上线“Skills”功能。该功能赋予了AI助手Grok跨对话的持久记忆能力，标志着Grok正在从传统的单次问答聊天机器人向可配置、可编程的自动化工作空间演进。据介绍，通过“Skills”功能，用户只需向Grok传授一次特定任务的操作方法，Grok就能在后续的所有独立对话中永久记住相关的个人偏好、排版格式规则或特定工作流步骤。这意味着用户在开启新对话时，无需再不厌其烦地重复输入繁琐的前置提示词或背景设定。（界面）

又诞生一位千亿女富豪36氪的朋友们·2026年05月17日 15:45又一部“打工妹”逆袭成为女富豪的爽文剧本。示意图自4月初以来，立讯精密股价上涨超过55%。5月11日，立讯精密市值创下历史新高，达到了5542亿人民币。立讯精密掌舵者王来春持股18.75%，身价也突破了千亿人民币，成为国内白手起家的又一位千亿女富豪。如果说4月以来的涨幅，是因为立讯精密“站在光里”，短期股价大涨或许与一则收购公告有关。5月10日，公告显示立讯精密拟通过收购京西智行集团全部股权，间接拿下京西国际约59.50%股份，有望成为其最终控股股东。这是立讯精密继两年前收购德国莱尼后，在汽车领域的又一关键资本动作。郭台铭门 徒立讯精密的故事通常被描述为“打工妹”逆袭成为女富豪的爽文剧本。实际上，只单纯地给创始人王来春贴上“打工妹”的标签，就像是只把立讯精密当做一家代工厂。与现在创业者的背景不同，王来春既不是天才少女，也不是海归精英。她出生于20世纪60年代末的潮汕地区，当时物质和思想都极度匮乏，农家女出身的她，更是过早地尝到了生存的艰辛。1988年，改革开放的风头正盛，21岁的王来春选择到最热闹的深圳求职，此时正赶上郭台铭来内地创业，于是，机缘巧合下，她成为了富士康首批150名员工之一。早期的富士康被外界称为“血汗工厂”，流水线两班倒，每天站12个小时是家常便饭，累了一天，洗澡还得等。很多人受不了这份苦，干了几个月就走了，但王来春却坚持留了下来，这一留就是十年。这十年间，王来春不仅成为了大陆员工里职位最高的女主管，还像海绵一样吸收了消费电子行业的大量经验，成长为了郭台铭最出色的学生。一个细节是，她在立讯几乎复制了富士康模式，工厂管理军事化，连车间里都挂着郭台铭的管理语录。对于“徒弟”的成功，郭台铭曾公开评价：“富士康在中国大陆创造了很多亿万富翁，王来春是最具代表性的一位！”1999年，受金融危机影响，香港很多公司都破产倒闭了。王来春看准机会，毅然辞去富士康高管职位，和哥哥王来胜共同收购了香港立讯公司。面对这一冒险举动，她的家人是反对的，其母亲甚至把家里的现金和存折都藏了起来。然而，这并非是王来春的冲动之举，在立讯早期，王来春选择依附富士康生存，“富士康开到哪，立讯就跟到哪”，专门承接其一时无法处理的订单：低毛利的，高难度的。富士康给什么，她就做什么，在定价方面也完全由富士康说了算。所以成立后的很长一段时间内，富士康始终是立讯的第一大客户，这种关系也让富士康越来越离不开立讯，因此，郭台铭兄弟还入股了立讯。2004年，王来春成立立讯精密工业股份有限公司，开始独立开发自有客户并自主研发产品，逐步向高端产品线迈进。尽管如此，到2010年，立讯精密上市时，富士康依然为立讯精密贡献了一半销售额。库克时代如果说2010年之前的立讯精密，还是一个标准的“小富士康”，还是在王来春的舒适区里创业，那么2010年之后，立讯精密的发展，则完全依赖于王来春的智慧与资本布局。因此，业内流传着一个说法—立讯精密真正的创业之路是从2011年开始的。这一年，已经上市的立讯精密有了充裕的现金，此时的王来春不再满足于代工，把赚来的钱全部砸进了更难的技术领域，一边寻求转型，一边开始“去富士康化”。也是在这一年，代工领域一个时代性的机会出现了。2011年，库克执掌苹果并推行供应链多元化，王来春意识到，新帅上任，势必调整布局、扶植自己的心腹伙伴。若能在此时跻身“果链”并获库克青睐，立讯便有望摆脱依赖外部溢出订单的处境。于是，她开启了一系列激进的“资本换技术”的并购。并购的第一枪打在了昆山联滔电子身上，立讯精密斥资11.8亿元收购了其60%的股权，此举规避了苹果复杂的考核机制，使得立讯精密顺利进入了苹果供应链。这起并购案只是一个开始，2016年，它又收购了苏州美特51%股权，靠着苏州美特在声学领域的多年积累进入到了声学领域，并在下一年获得了苹果AirPods的生产机会。据悉，当时苹果的AirPods主要由台湾英业达生产，但英业达生产效率很低，经常出现供货延迟的情况，立讯精密拿到订单后攻克了AirPods生产效率低的问题，最终成为AirPods全球最大的供应商。2020年7月，立讯精密又收购了江苏纬创、昆山纬新两家公司，正式切入iPhone代工领域，成为了中国大陆首家、全球第三家iPhone整机制造商，打破富士康、和硕的垄断格局，并于2021年首次获得iPhone 13部分订单，实现了从零部件供应商向整机组装商的关键转型。2021年，立讯精密已发展成为营收超1500亿元，苹果业务占比达七成的全球消费电子巨头。可此时，王来春却推出了三个五年计划，全方位发展消费电子、汽车、通信、工业、医疗等业务，并在汽车领域制定了Tier1业务进入全球前十的战略目标。她还提出，十年内将单一客户营收占比降至1/3以下，这个客户，就是苹果。后来有分析称，王来春之所以如此决绝是因为看到了苹果供应商欧菲光的下场，2021年3月，欧菲光被踢出苹果供应链，但由于其业务太过于依赖苹果，进而受到了重创。现在看来，立讯精密摆脱“果链依赖症”的效果也非常明显，从2021年到2025年，苹果贡献的业绩占比分别为74%、73%、75%、71%、57%。整个库克时代，立讯精密陪伴苹果走过了十五年，业务覆盖了其各个业务线，营收也从25.6亿元增长至了3323.4亿元，市值也随之增长了近40倍。而如今库克即将退休，立讯精密也找到了汽车电子、通信业务等新增长引擎。从某种程度上来说，这也是另一种圆满。公司上市十六年，她身价涨千亿再说回开篇提到的对京西国际的收购，这笔交易之所以能够直接带动立讯精密股价的上涨，与前文提到的“Tier1业务进入全球前十的战略目标”有直接关系。此前，立讯精密的汽车业务主要集中在“线束”和“连接器”领域，而京西国际的优势在悬架系统和制动系统，这笔收购直接补齐了立讯精密在底盘域执行机构的短板。更重要的是，京西国际拥有从德尔福继承的磁流变悬架和线控制动等全球稀缺技术，也拥有宝马、保时捷等欧洲顶级车企客户。要知道，底盘认证周期高达3至5年，技术壁垒极高，立讯通过这笔收购，相当于买了个时间差，直接拿到了这些高端客户的“入场券”。另外根据去年财报，汽车电子作为第二增长曲线，迎来了爆发式跃升。该板块实现营收392.55 亿元，同比大幅增长185.34%，营收占比快速提升至 11.81%，成为立讯精密增长最强劲的引擎。当然，5月11日市值能够创下新高，也离不开其四月底以来持续利好消息带来的股价拉升。4月27日，立讯精密收获了又一个涨停，最高冲到了72.6元，创下历史新高，总市值突破5200亿元。当天天风国际分析师郭明錤独家爆料，OpenAI计划自研手机，立讯精密成为独家系统联合设计和制造合作伙伴，同时，公司预告2026年上半年归母净利润78.4亿—81.06亿元，同比大增18%-22%。在此基础上，立讯精密近期披露，将持续加大铜光互连、散热及电源等核心产品的研发投入与产能布局，热门概念叠加下，立讯精密获得了资本真金白银的投票。从2010年上市时不足70亿元的市值，到如今5542亿元的巨头，立讯精密用16年完成了79倍的市值增长。而王来春的身价也从上市之日的23亿元增长到了1039亿，实现了超千亿的财富增加。然而，这并非是王来春身价的顶点。毕竟立讯精密已经向港交所递交了上市申请书，冲刺A+H上市，并预计今年下半年登陆港交所。那时，“A股果链一哥”将向全球科技制造龙头发起冲击，更重要的，其作为全球唯一同时布局消费电子、汽车与光通信产业的企业，其资本溢价将进一步释放，王来春的新故事也将开启新篇。本文来自微信公众号“投中网”，作者：张雪，36氪经授权发布。本文由「36氪的朋友们」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、采访供图

科氪 | 对话韶音高层：开放式耳机，凭什么成为下一场颠覆TWS的“品类革命”？科氪·2026年05月17日 19:57韶音深耕开放式音频，以硬核技术推动全场景品类革命从运动孤岛到品类革命：解构韶音的“开放式”野心与底层技术硬实力在消费电子行业普遍面临增长红利见顶的今天，音频赛道却正在上演一场深刻的结构性变革。2026年5月15日下午，大雨初晴的深圳湾广场，Shokz韶音举办了备受瞩目的媒体专访会。韶音中国区CEO杨云与产品线总裁张浩锋悉数到场，向外界系统阐述了这家以“骨传导”和“运动”起家的品牌，如何在稳坐全球开放式耳机出货量第一王座的同时，向更广阔的泛场景音频帝国发起攻坚。过去两年，开放式耳机在中国蓝牙耳机市场的整体销售额份额已率先突破30%。这不仅是一次产品形态的更迭，更是一场由韶音引领、正在席卷全球的“品类级替代”风暴。一、 宏观战略与行业洞察：以“全场景品类替代”重塑全球音频版图过去，大众对韶音的认知多停留在“专业运动”与“骨传导”这两个标签上。然而，在本次专访中，韶音高层明确指出，专业运动与泛场景用户从来不是一道非此即彼的选择题。韶音正在做的事情，是将用户的定义从最初的“专业运动人群”彻底拓展到“每一个需要开放聆听的用户”。官方图片1. 舒适成为刚需：跨越品类鸿沟的终极逻辑韶音之所以做出“开放式耳机有机会取代真无线耳机”这一极具前瞻性的行业断言，其核心洞察深植于用户日常用耳习惯的根本性改变。随着线上办公、远程会议、长音频内容、网课以及日常碎片化聆听需求的激增，现代人佩戴耳机的时间正在无限拉长。在长时间佩戴的背景下，传统入耳式耳机带来的耳道压迫、闭塞闷痛、卫生隐患以及无法感知外界环境的安全痛点，逐渐演变为不可调和的刚性冲突。而开放式耳机所具备的“全天候舒适佩戴”与“安全开放聆听”的天然优势，正让它从一款特定场景的“运动副机”，跨越成为用户日常生活中无可替代的“主力设备”。在音质水平已经通过近年来的技术迭代基本解决、甚至在部分产品上（如OpenFit Pro）接近或超越传统入耳式耳机的今天，用户在舒适度与开放感知方面的倾斜，成为了开放式耳机跨越品类鸿沟的决定性力量。官方图片2. 全球化视野下的“GLOCAL”战略在出海与全球化版图中，韶音在2023年至2025年间已连续三年斩获开放式耳机市场全球出货量份额第一的桂冠()。数据显示，不仅是中国市场在狂飙突进，在全球重点市场如美国、欧洲、日本、韩国、澳大利亚等地区，开放式耳机占比持续上升、对传统入耳式耳机产生强烈替代的趋势也已不可逆转。面对全球化竞争，韶音的核心市场规划可以总结为“GLOCAL”战略——即全球化思考，本地化行动。杨云将当前的耳机品类竞争比作新能源汽车行业：全球绝大多数的核心品牌与顶尖供应链都集中在中国市场。因此，韶音选择将中国激烈的市场环境作为研发的“练兵场”，充分依托国内强大的工程师红利和极致的供应链优势，开发出在硬件、性能上足以打动并销售全球的标准标准化产品。但在软件、服务及生态融合层面，韶音则在全球范围内开展极其精准的本地化深耕。例如，最新发布的专业运动耳机OpenSwim Pro 2，便成为了全球首款支持QQ音乐无线导入与离线播放的耳机，这一功能精准击中了本土运动爱好者的核心痛点，成为了全球标准化硬件与本土顶级流媒体生态成功联姻的范例。官方图片3. 拒绝短视的长期主义护城河在涌入者众多的多元化竞争环境下，韶音认为品牌真正的护城河来源于“技术创新”与“用户思维”的底层基因，以及对长期主义的笃定。做品牌是一条最难走、面临挑战与压力最大的路。在过去长达15年的时间里，韶音只做了一件事，那就是专注于开放式音频的深耕。这种克制与长线思维同样体现在其品牌营销与用户沟通上。从跨界合作自由潜水世界冠军金明，到携手全球品牌大使张婧仪推出艺术礼盒，韶音的策略重点并非盲目追求流量快钱，而是通过这些契合品牌坚韧、优雅特质的合作伙伴，将“开放聆听、不急于求成”的理念具像化，从而在消费者心中建立起高端化、专业化且极具信任感的品牌壁垒。二、 产品迭代与技术革新：材料与声学的极限攻坚要在红海中完成对传统TWS的品类替代，单纯依靠概念营销是远远不够的，必须在产品和技术实操层面，以硬核的底层研发打破开放式耳机长期存在的物理瓶颈与行业痛点。在最新的产品矩阵中，韶音通过声学、材料学与垂直场景的深度融合，带来了行业瞩目的迭代升级。1. 5年磨一剑：以“韶音天篱滤噪技术”定义开放式新阶段在开放式耳机的普及过程中，抗噪能力弱、在嘈杂环境下听音模糊一直是阻碍其进入更广泛泛场景的致命短板()。针对这一行业痛点，韶音在行业内首次提出了“开放式滤噪”的颠覆性概念。开放与滤噪，在物理和声学原理上天生是一对巨大的矛盾。在项目启动初期，这项技术在韶音内部甚至引发了巨大的争议，许多研发人员一度认为这是无法跨越的物理鸿沟。然而，基于对技术的长期投入和“因为相信，所以看见”的技术信仰，韶音早在2020年底至2021年便前瞻性地立项了滤噪预研技术——当时行业内甚至还没有出现如今大火的耳挂式开放式耳机，研发完全是基于骨传导的未来构想展开的。历经5年多的高强度技术迭代与攻坚优化，韶音终于推出了“韶音韶音天篱滤噪技术”。该技术成功将滤噪的频段拓宽至接近4000Hz，并将滤噪深度降到了18dB，使其真正成为行业首个能够在健身房、咖啡厅、办公室等多场景完美适配的开放式滤噪方案。这标志着开放式耳机进入了一个全新阶段：既能保留开放聆听带来的环境感知与安全出行，又能兼顾沉浸专注的清晰听音体验。张浩锋更是透露，如果这项技术在未来能够继续向高铁、飞机等高难度嘈杂场景演进，那将直接触发音频行业的下一场彻底革命。官方图片2. OpenSwim Pro 2：解锁游泳蓝海的“全能数据终端”多数音频厂商将运动耳机的目光死死锁在跑步赛道，而韶音则将下一个具备决定性增量的突破口精准锁定在了游泳市场()()。根据韶音内部的全球调研数据统计，经常游泳的人群已经占到了跑步人群的60%左右，这是一个庞大却长期被行业低估的蓝海。由于游泳场景极为单调且无法在运动中聊天拍照，用户在水中对音乐激励的需求比跑步往往更加强烈和刚性。然而，水下由于蓝牙信号无法有效传输、防水等级要求严苛，市面上长期缺乏优秀的解决方案，这一领域对于高品质音频而言几乎是零竞争对手的真空地带。为此，OpenSwim Pro 2不仅将防水能力直接拉高到了极致的潜水级，更在硬件内部塞入了一颗强悍的6nm定制芯片与九轴传感器。基于对真实用户需求的深刻洞察，游泳者如果佩戴智能手表，在水中频繁抬腕观察数据会严重打断运动的节奏和划水连贯性。OpenSwim Pro 2通过这套顶级的传感器与芯片组合，能够精准识别泳姿、划次、游泳距离等核心运动指标，并通过语音实时播报的方式反馈给用户，完美做到了声学、运动科技与极端场景体验的闭环融合。3. OpenDots 2：夹式耳机的“微米级”舒适平衡近年来，耳夹式（Clip-on）形态因其潮流属性和独特的佩戴方式，正在疯狂吸纳传统的TWS用户。但耳夹式产品始终面临着一个行业内难以调和的悖论：夹得太紧耳廓容易产生压痛，夹得太松则在剧烈运动中容易松动脱落。为了达成“久戴不痛、狂甩不掉”的黄金平衡，最新亮相的OpenDots 2在完美延续了前代OpenDots ONE所奠定的“动钛弧结构”以及“9G预紧力”两大技术核心优势的基础上，在材料学与重心分布上进行了极其细腻的微米级微调。在结构优化上，OpenDots 2将内部核心的钛片宽度做到了极致压缩，最大减少了0.6毫米，这微小的尺寸缩减极大地减轻了对耳廓边缘的局部压迫感。同时，研发团队重新优化了耳机整体的重心分布，并全面提升了钛片的回弹性与抗疲劳度，从而保证了耳机在高强度、高频次拉伸使用的长期蹂躏下，依然能维持极其稳定、均衡且高性能的夹持体验。这种通过材料创新来化解佩戴稳固与舒适矛盾的研发路线，进一步筑牢了韶音在舒适科技领域的领先地位。三、 未来战略演进：在AI与智能硬件洪流中的冷思考在当前言必称AI、智能眼镜与可穿戴健康监测设备大乱斗的硬件大潮中，韶音在专访中展现出了极强的战略定力与极其清醒的“冷思考”。面对未来的技术演进，他们并没有盲目跟风，而是有着自己一套逻辑严密的方法论。1. 激进的稳健派：AI重构智能耳机的三步走节奏对于风头正劲的AI技术，韶音在内部将自己定性为“激进的稳健派”。他们深信AI在未来势必会全方位重构整个智能硬件生态，但这一进程绝非一蹴而就，而是有着极其鲜明的节奏感。在韶音的蓝图中，AI在开放式耳机上的落地被清晰地划分为三个层层递进的演进维度：第一维度是性能增强（已全面落地）：这是目前已经成熟并广泛应用的阶段。韶音通过AI算法深度优化了耳机的抗噪性能、麦克风通话降噪效果以及底层的声学音质表现。第二维度是交互体验的革命（正在切换）：随着AI大模型的普及，人类与硬件的交互媒介正在从过去的键盘、屏幕逐步切换到极其自然的语音对话。由于开放式耳机具备全天候挂耳、不隔绝环境、高频佩戴的天然属性，它被韶音视为最自然、最适配的AI自然语音交互入口，未来在这个维度将爆发巨大的交互革命机会。第三维度是能力增强（韶音最核心聚焦的战略方向）：这一维度不再局限于空泛的泛场景，而是将AI聚焦在特定的垂直场景中大显身手。例如，通过耳机上的传感器采集用户的实时运动行为数据，并借助AI在本地或云端进行深度处理，从而为用户提供私教级的智能运动辅导；或者在办公场景下，通过AI处理声音数据来大幅度提高用户的办公和生产力效率。这种切中垂直场景的能力增量，正是韶音当前最为看好并重兵投入的王牌方向。2. 智能眼镜 vs 开放式耳机：乐观的谨慎派视角针对当前行业内热议的“拥有4亿眼镜人口的中国市场，智能眼镜是否会吞噬耳挂/开放式耳机生存空间”的尖锐话题，韶音高层给出了基于商业底层的冷静理性的研判。张浩锋坦言，韶音对于眼镜市场是一个“乐观且谨慎派”。眼镜作为面部穿戴设备，其复杂程度与面临的物理限制远超耳机。用户可以毫无心理负担地将一双轻巧的耳机挂在耳朵上一整天；但是，如果试图将一整块显示屏幕、复杂的计算单元、大容量电池强行架在用户的脸上，硬件在产品重量、续航能力、散热性能以及制造成本之间的“不可能三角”物理矛盾，在短期内根本无法得到完美解开。因此，开放式耳机与AI相结合的普及速度和商业落地节奏，注定会比智能眼镜来得更快、更扎实。眼镜未来或许有巨大的长线潜力，但它的技术挑战与迭代周期注定会非常漫长。官方图片3. 健康监测的“减法哲学”与价值回归在智能穿戴设备拼命做“加法”、试图把心率、脑电等各种生物传感器塞入耳机的行业风潮下，韶音却展现出了一种极其难得的克制与“减法哲学”。杨云指出，运动健康监测（如HRV、持续心率等）的核心商业价值与用户价值，高度依赖于“7×24小时无缝、长期且连续的数据基准支撑”。这正是智能手环、智能手表这类用户连洗澡、睡觉都几乎不摘单的设备的天然主场。相反，耳机作为一种高频摘戴、使用场景相对碎片化的设备，天生无法提供无断点的连续时间基准数据。如果在预研阶段为了拼凑功能和噱头强行塞入这些复杂的生物传感器，势必会严重牺牲耳机的佩戴重量、体积与续航体验。在经过严密的评估后，韶音认为这无法给用户带来更好的综合核心体验，因此他们选择克制，在不必要的功能上做取舍与减法，让耳机回归到极致的佩戴舒适科技与声学本源上。四、 结语在韶音的终局思维里，开放式耳机从来不是属于一小撮硬核运动人群的专属孤岛，而是一场终将彻底颠覆并重塑传统真无线音频行业格局的品类海啸。从深耕骨传导、定义专业运动，到如今全面布局耳挂、耳夹全形态，走向全场景开放音频；从硬核预研“韶音天篱滤噪”打破物理矛盾，到重塑水下运动的高清聆听体验，韶音正用长达15年的笃定、对真实需求的洞察以及硬核的技术信仰，不断敲碎那些外界自以为是的行业天花板。正如CEO杨云在媒体专访结束时所言，所谓的“终局思维”绝不意味着把天花板定在这里，韶音的技术上升空间与想象力依然巨大。这场将舒适与聆听自主权彻底还给消费者的开放式革命，才刚刚行至中流，而韶音显然已经握紧了通往下一个时代的船票。本文由「科氪」

8点1氪丨多款燃油车大降价，降幅最高近30%；蜜雪冰城上线黄仁勋同款饮品；“巴菲特慈善午餐”回归，获拍900万美元8点1氪·2026年05月18日 07:55今年1至4月份，常规燃油车新车降价车型均价23.6万元的降价力度算术平均达到3.4万元，降价力度达到14.6%。今日热点导览张雪机车，再夺冠长安汽车否认与千里科技合作深圳发布2026年住房发展计划贵州茅台宣布公斤茅台等四款产品提价法拉第未来宣布获得2500万美元融资三星电子会长就公司“内部问题”公开道歉TOP 3 大新闻多款燃油车大降价，降幅最高近30%日前，长安汽车旗下长安逸动经典版正式上市，新车售价6.49万元，限时享23000元购车现金优惠；相较于指导价8.79万元降幅达到26%。该款车型还享有限时的保养礼和流量礼。其实，近期多家燃油车都开启了降价模式：起亚新狮铂拓界一口价10.99万元，较指导价15.49万元降29%；捷豹路虎5月12日推出限时“新发现”价；奇瑞4月中旬开展促销，综合补贴最高4万元；速腾2026款最低价11.99万元，降价3万元（降幅20%）；思域最低指导价9.79万元，降价3.2万元（降幅24.6%）；荣威i6最低价7.49万元，降价1.49万元（降幅16.6%）；一汽-大众奥迪A6于3月降价9.1万元（降幅29%）。乘联分会秘书长崔东树最新披露的数据显示，2026年4月份，常规燃油车新车降价车型均价13.1万元的降价力度算术平均达到2.3万元，降价力度达到17.2%。今年1至4月份，常规燃油车新车降价车型均价23.6万元的降价力度算术平均达到3.4万元，降价力度达到14.6%。（光明日报）蜜雪冰城上线黄仁勋同款饮品5月15日，英伟达CEO黄仁勋在北京南锣鼓巷的蜜雪冰城门店点了杯“蜜桃四季春”，引发关注。5月16日上午，有媒体记者注意到，蜜雪冰城官方点餐小程序已火速上线“大佬同款”专属分类，直接把这款单品置顶。郑州中原区多家门店的店员告诉记者，这款饮品价格没变，还是7元起，而且“大佬同款”销量相当不错。蜜雪冰城相关工作人员回应，小程序上线“大佬同款”后，蜜桃四季春当日销量环比前日增长接近140%，同比上周五增长超过90%。黄仁勋到访的南锣鼓巷店，蜜桃四季春销量环比增长超过90%。另有消息称蜜雪冰城CEO张渊在朋友圈晒出黄仁勋喝“蜜桃四季春”的照片，并配文：“最顶级的显卡，往往需要最纯粹的散热方式。”同时，蜜雪冰城官方微博也发文称：“高温模式已开启！液冷降温和蜜桃四季春更配哦。”（界面新闻、新浪财经）“巴菲特慈善午餐”回归，获拍900万美元据eBay网站显示，在一场慈善筹款拍卖活动中，一位竞拍者同意出资900万零100美元，以换取与亿万富翁投资者沃伦·巴菲特以及篮球巨星斯蒂芬·库里共进晚餐的机会。目前尚无法立即确定中标者的身份。6月24日，巴菲特、库里及其妻子将与中标者及其选定的最多七位随行嘉宾，一同前往位于内布拉斯加州奥马哈市的伯克希尔·哈撒韦公司总部所在地共进晚餐。（同花顺财经）大公司/大事件小米卢伟冰：下半年部分国产旗舰直板手机价格或将破万5月16日，小米集团总裁卢伟冰在直播中表示，今年下半年，尤其是年底，可能有一些国产旗舰直板手机的价钱会突破1万元人民币。针对小米17 Max价格会不会受到内存价格影响，卢伟冰称，今天任何一款手机定价都会受到内存成本大幅上涨的影响，而小米17 Max的价格还在讨论当中。他预判，内存涨价趋势至少会延续到2027年底，甚至到2028年。（界面新闻）张雪机车，再夺冠当地时间5月16日，在刚刚结束的2026世界超级摩托车锦标赛（WSBK）捷克站WorldSSP组别第一回合正赛中，中国摩托车制造商“张雪机车”旗下的53号法国车手瓦伦丁·德比斯夺得冠军。（央视新闻）巨力索具今年25次登龙虎榜，在A股上市公司中高居第95月15日晚，今年以来股价翻倍的商业航天概念股巨力索具公告称，公司于当日收到中国证监会下发的立案告知书，因涉嫌信披误导性陈述违法违规，根据相关法规，5月13日，中国证监会决定对公司立案。值得注意的是，今年以来，巨力索具25次登上龙虎榜，在A股上市公司中高居第9，累计买入143.86亿元、卖出134.52亿元，累计净买入9.35亿元。龙虎榜上，机构与游资频繁出没，散户跟风炒作。巨力索具披露的2026年一季报显示，一只私募基金、一家银行机构，以及两位高度关联的“超级牛散”，重仓押注，跻身前十大流通股东。而控股股东巨力集团则通过股权协议转让，持仓大幅减少4800万股。（中国证券报）长安汽车否认与千里科技合作近日有消息称，长安汽车将采用千里科技的智能辅助驾驶方案，并投资入股。长安汽车5月17日称，“暂无计划”。另有接近长安汽车的人士向记者透露：长安汽车自研的天枢智能已经到了收获期，不可能切换智能驾驶方案。据悉，去年12月，工信部公布了我国首批L3级有条件自动驾驶车型准入许可。长安汽车位列其中，搭载的是其自研的“天枢智能”辅助驾驶系统。（第一财经）蔚来：已在全国建成8870座充换电站36氪获悉，蔚来发布消息称，5月17日，NIO Power 2站上线。截至5月17日，蔚来全国建成8870座充换电站，其中换电站3849座，充电站5021座、充电桩28855根，累计提供超1亿次换电。深圳发布2026年住房发展计划，将筹建3万套、供应4万套保障房36氪获悉，深圳市住房和建设局近日发布《深圳市2026年住房发展年度计划》。优化保障性住房供给，建设筹集保障性住房3万套（间）、供应4万套（间），满足工薪群体和各类住房困难家庭基本住房需求；推进“好房子”试点项目建设，推进城中村改造和城镇老旧小区整治改造，推动创建未来社区（完整社区）试点，切实增强市民获得感、幸福感、安全感。最高涨价200元，贵州茅台宣布公斤茅台等四款产品提价5月16日凌晨，i茅台公告，自当天零时起，贵州茅台酒股份有限公司对部分贵州茅台酒产品自营体系零售价格进行调整。此次调整共涉及四款非标产品，最高涨幅为8%。调整的产品包括：53%vol 500ml陈年贵州茅台酒（15）由4199元/瓶调整为4279元/瓶；53%vol 500ml贵州茅台酒（精品）由2299元/瓶调整为2359元/瓶；53%vol 500ml贵州茅台酒（丙午马年）珍享版由2499元/瓶调整为2699元/瓶；飞天53%vol 1L贵州茅台酒由2989元/瓶调整为3119元/瓶。（澎湃新闻）国内航线燃油附加费5月16日起上调5月16日起，中国内地航线客票，燃油附加费上调。具体标准为：800公里以下航段每位旅客收取90元；800公里以上航段每位旅客收取170元。婴儿旅客免收燃油附加费，儿童、革命伤残军人、因公致残人民警察客票燃油附加费减半收取。（央视新闻客户端）2026年5月票房破15亿36氪获悉，据灯塔专业版，截至5月17日，2026年5月票房（含预售）突破15亿，《给阿嬷的情书》领跑5月电影票房榜。上海电信率先发布Token资费套餐：1元对应25万额度点，支持手机账单付面对AI时代日益增长的Token需求，上海电信宣布，即日起面向上海电信用户正式推出Token算力服务，成为上海首个发布Token资费套餐运营商。具体来说，一是按量订购，1元对应25万额度点（以KiMi-K2.5模型为例，约可支持调用25万输入Tokens）；二是按需选择，用多少买多少，多买可享折扣，支持话费账单支付。另据记者从上海电信获悉，用户购买额度点后，通过标准API接口即可调用30余款主流大模型（文本、多模态），将AI能力集成到自己的软件、脚本或自动化流程中。以10元购买250万额度为例，约可让程序自动总结100本10万字的电子书。(上观新闻)超310家企业落户香港或在港扩充业务香港特区政府投资推广署署长刘凯旋表示，截至5月初，已有310多家企业落户香港或在港扩充业务，首年直接投资额超260亿港元，其中中资企业占比超过五成。她认为这一成果得益于特区政府成立的“内地企业出海专班”的推动。(大湾区之声)法拉第未来宣布获得2500万美元融资36氪获悉，5月16日，法拉第未来宣布，公司已与机构投资者签署了2500万美元股份购买协议。根据协议，投资者已认购公司总本金为2500万美元的可转换本票。本次融资对应的普通股目前尚未注册，存在交易限制，暂不可立即流通。其中，1250万美元将直接汇入公司运营账户；其余1250万美元将依据公司与各投资者签署的受控账户协议，存入由投资者控制的监管账户，并将在满足特定条件后释放给公司使用。星巴克宣布在美关闭多个地区办公室当地时间5月15日，星巴克宣布将在美国裁员300人，以削减成本、力求重回增长轨道。星巴克还表示，关闭多个地区办公室，相关举措将产生包括遣散费以及资产减值等在内的4亿美元重组费用。（央视财经）伊朗宣布5月19日起恢复股票市场交易当地时间5月16日，伊朗证券交易组织表示，伊朗股票市场将于5月19日起恢复交易。伊朗证券交易组织负责交易所相关负责人表示，根据已完成的协调工作并在获得必要许可后，决定自19日起重新开放股票市场、各类股票投资基金及其衍生品交易。（央视新闻）盖茨基金会不再持有微软股票当地时间5月16日，根据提交给美国证券交易委员会的文件，微软创始人比尔·盖茨的慈善基金会已不再持有微软股票。盖茨基金会信托披露，已于今年第一季度出售剩余770万股微软股票，按当前股价计算价值32亿美元。（界面新闻）马斯克称他不会出售任何SpaceX股票马斯克称他不会出售任何SpaceX股票。（华尔街见闻）SpaceX据悉最早将于6月11日确定首次公开发行价格5月16日消息，据报道，SpaceX最早将于6月11日确定首次公开发行价格，股票将于6月12日上市。SpaceX将选择纳斯达克作为其首次公开发行的上市地点，股票将以“SPCX”为代码进行交易。（界面新闻）SpaceX拟推万亿火星薪酬方案，马斯克回应：旨在确保公司专注于多行星生活愿景5月16日，有消息称，正在推进上市的SpaceX在保密招股书中披露了一项激进的治理与薪酬方案。公司计划通过发行10倍投票权的B类股，确保马斯克拥有不可被解雇的绝对控制权。同时，董事会批准了一项总额可达万亿美元的超级薪酬方案，其考核标准直接与在火星上建立百万人口殖民地的目标挂钩。目前，SpaceX计划于今年6月登陆纳斯达克。针对这一治理方案，马斯克随即做出了公开回应。他表示：“我需要确保SpaceX始终专注于让生命成为多行星化、并将人类意识延伸至星辰大海的愿景，而不是去迎合某些人极其无聊的季度财报奖金！”马斯克说，这一目标实现起来异常艰难，过程绝非一帆风顺，但他坚信，“如果SpaceX成功实现了这个荒谬而艰巨的目标，它的价值将比地球目前的经济总量还要高出好几个数量级。”（界面新闻）三星电子会长就公司“内部问题”公开道歉5月16日，三星电子会长李在镕就公司“内部问题”公开道歉。目前，三星电子与公司工会之间劳资纠纷持续，工会计划于5月21日启动的18天罢工计划已有超过4万名员工登记。李在镕呼吁工会与公司团结一致，朝同一个方向前进。三星电子劳资双方自去年12月起开展谈判，但由于分歧一直未能缩小，谈判于今年3月破裂。此后，在调解下，劳资双方本月11日至13日进行了马拉松式谈判，但最终未能达成协议。据韩联社报道，若进行罢工可能会造成40万亿韩元、约合1824亿元人民币的损失，并殃及半导体产业，政府可能考虑行使紧急调整权。（CCTV国际时讯）OpenAI与马耳他达成协议，所有马耳他民众都能使用ChatGPT Plus美国人工智能（AI）公司OpenAI5月16日表示，已与马耳他政府签署协议，所有居民在完成使用AI培训课程后，均可免费使用其ChatGPT Plus服务一年。该项目将于5月启动，随着更多马耳他居民完成这门免费课程，服务规模预计将逐步扩大。该计划也将向旅居海外的马耳他公民开放。（新浪财经）贝莱德考虑对SpaceX首次公开募股投资数十亿美元据知情人士透露，贝莱德已讨论在下个月对SpaceX的首次公开募股（IPO）投资50亿至100亿美元。这笔巨额投资对应的IPO募资规模最

对话万成云商：发文章≠GEO优化，大模型不是喂什么就推什么杨越欣（杨桃）·2026年05月18日 13:05对话万成云商：拆解出海企业AI时代GEO布局路径“帮我搭配一套适合海边度假的衣服。”“帮我推荐一家适合约会的西餐厅。”“3000元预算，买哪款扫地机器人最划算？”这类与AI的问答正发生在全球数亿用户身上。过去，人们会把这些问题输入百度、谷歌等搜索引擎，因此产生了SEO（Search EngineOptimization，搜索引擎优化）营销。而当流量入口从搜索引擎转向AI工具，GEO（Generative EngineOptimization，生成引擎优化）也应运而生。区别于传统SEO追求的“搜索结果排名”，GEO的核心目标是让品牌信息在AI生成的回答中被优先引用和推荐。在AI全面渗透用户信息获取与消费决策场景的时代，对于出海企业而言，做好GEO意味着抢占新的流量入口。GEO的逻辑和SEO有什么根本不同？为什么有人说GEO是伪命题？出海企业布局GEO该怎么做？中小企业如何低成本起步……带着这些出海企业关心的问题，我们特别对话了万成云商品牌出海事业部总经理谭莉，拆解AI搜索时代如何让企业被AI看见、信任，并推荐给用户。36氪：GEO被称为SEO在AI时代的进化。相比于SEO的核心在于搜索结果的排名，GEO优化的关键是什么？它与传统SEO的核心区别在哪里？谭莉：核心区别在于传统SEO偏“规则博弈”，GEO更像“信任竞争”。因为过去做SEO，我们的目标是把页面排到Google第一页，让用户点进去；但现在很多用户直接在AI里问问题，甚至不点链接了，那你要解决的是——怎么让AI在回答里“引用你”甚至“推荐你”。所以GEO的关键首先是，你是不是一个“可信任源”（全网有没有人认可你）；其次，你的内容能不能被AI理解、拆解、复用；最后是你的内容有没有完整的证据链，比如数据、案例或者第三方背书。36氪：继GEO之后，也有人提出AEO概念，两者有什么区别？谭莉：这两个概念很多人会混着讲，但我个人是这么区分的：AEO更偏“问答结果优化”，核心是让你的内容直接成为某个问题的标准答案，比如FAQ、精选摘要这些。GEO更广，它面对的是生成式AI，不只是回答一个问题，而是参与“多源信息整合后的输出”。所以AEO更偏“结构技巧”，GEO更偏“整体品牌与内容生态”。36氪：GEO的概念近两年引起广泛讨论。为什么现在这个时间点企业需要关注GEO？在AI搜索渗透率持续提升的背景下，GEO应该在企业营销体系中占据怎样的战略位置？谭莉：分享几个很现实的变化：现在用户越来越习惯直接问AI，而不是自己筛选搜索结果；而且AI在很多行业决策前端，已经开始替代“信息收集阶段”；再就是有些客户还没打开你的官网之前，可能就在AI里形成了对你的判断。所以企业没办法去忽视GEO，现实摆在这里了。我认为，GEO在企业营销体系中占据着很关键的位置，它不会替代SEO，而是和SEO并行的“第二流量入口”，甚至未来会变成主入口之一。36氪：怎么理解有企业做过一些尝试后认为GEO是伪命题，或者靠自己持续给大模型喂语料也可以？谭莉：我觉得这是对AI工作机制有点误解，大模型不是你“喂什么它就推荐什么”。它有几个筛选逻辑：第一，信息是否一致（比如不同来源是否能得到印证）；第二，来源是否可信（是不是长期存在的内容资产）；第三，是否被多平台提及（而不是单一来源）。所以你自己“喂语料”，本质只是单点信息输入，很难形成“信任网络”。不过我也理解为什么有人说GEO是伪命题，因为确实有很多“伪做法”，比如只做站内内容、只堆AI文章、不做外部验证，这些基本都很难起效果。36氪：当企业布局海外GEO营销时，需要经历哪些关键步骤？不同行业、不同规模的企业在路径上会有哪些差异？谭莉：从实操角度，会有四个关键步骤：第一步先拆解你的客户是怎么筛选供应商的，这个路径是怎么样？第二步重建你的内容结构，包括官网、博客、FAQ、案例这些，全部按“可被AI理解”的方式重构。第三步是搭建外部信任体系，比如评价平台、媒体曝光、社媒讨论、行业引用，都很重要。第四步就是做多渠道分发，持续让内容在不同平台被看见和引用。不同企业可能在行业复杂度和品牌基础这两方面存在明显差异。比方说医疗、工业类比其它行业更需要专业背书，那有品牌的企业就更容易放大，没有品牌的就要先“补信任”。36氪：在实施过程中，企业应如何在内容策略和技术基建层面做出改造甚至重塑，以同时满足AI理解与用户共鸣？谭莉：内容策略方面要做到结构清晰，信息有证据，有数据来源，表达要接近真人沟通。技术层面反而是基础的更容易被忽略：比如Schema结构化数据（Product、FAQ、Article等）、robots、index、hreflang这些基础配置，再一个就是页面可读性。总结来说，是做“可被理解的内容”。36氪：在布局GEO出海方案时，企业最容易忽视但至关重要的环节是什么？对于缺乏专业团队的中小企业，有哪些可行的解决方案或合作模式？谭莉：我们服务这么多跨境企业，发现最容易忽视的一点其实是：“让AI能访问你”。听起来很基础，但很多企业在实操中会陷入误区：比方说robots把AI爬虫挡了、页面noindex设置错误，或者多语言混乱……这些都会直接导致你做再多内容，AI也看不到。第二个容易忽视的是站外信任建设。很多人还停留在“发文章=做优化”，但GEO更看重：评论、测评、社媒讨论、媒体引用。对于中小企业，我建议两种路径：①轻量级：聚焦一个细分领域做深内容+少量高质量外部背书；②合作型：找有资源的服务商做整合（内容+渠道+数据），不要自己盲目试错，成本会很高。36氪：企业在选择GEO服务伙伴时，应该重点考察对方的哪些核心能力（如是否具备垂直行业知识库、跨平台渠道资源、效果监测与分析体系等），以避开“无效投入”的陷阱？谭莉：这个问题非常关键，因为现在确实有不少“概念型服务商”。我建议重点看四个能力：第一，行业理解能力，是不是懂你这个行业，而不是只会写内容。第二，内容生产质量。有没有真实案例、是否具备本地化能力，而不是纯AI批量生成。第三，渠道与分发能力。有没有海外平台资源，比如媒体、社区、测评渠道。第四，数据监测体系。比如做一段时间了，看看有没有被AI引用？哪些内容有效？ROI如何？如果对方只能讲“发多少篇内容”，基本可以直接排除。36氪：2026年初Seedance2.0的爆火，让业界看到了AI视频生成从“抽卡式试错”向“工业化生产”跨越的可能性 。10秒带货视频成本仅需8元的“成本塌方”对于跨境电商的内容生产逻辑意味着什么？谭莉：这个变化其实挺颠覆的。过去做视频，成本高、周期长，所以大家拼的是“有没有内容”；但现在成本接近于零，意味着：内容不再是门槛，内容“是否有效”才是门槛。36氪：当视频生产成本趋近于零，所有人都能批量生成高质量内容时，企业的竞争焦点会发生怎么样的变化？从GEO布局的角度看，企业应如何构建AI时代下的营销思维？谭莉：供给无限的时候，稀缺的就变了。我认为未来竞争会集中在三点：第一，认知占位，用户心里有没有你；第二，信任结构，有没有被多方验证；第三，数据能力，能不能快速迭代内容。从GEO角度，我觉得企业要建立一种新思维：要把品牌内容当资产去做长期积累，同时将口碑、评价这些信任一点点做起来，更重要的是发内容之前，问问自己：AI会不会用它？用户会不会信它？避免做“自嗨”的内容。36氪：Seedance2.0的火爆也引发关于深度伪造和版权的讨论。跨境电商卖家在利用这类工具生成营销素材时，应如何规避潜在的肖像权、版权风险？谭莉：这个问题确实不能忽视，我建议，尽量避免使用可识别人物（尤其是公众人物）做生成素材，素材来源尽量用可商用授权库，品牌、Logo、音乐等要确认授权范围，涉及医疗、功效类内容要特别谨慎，因为合规风险更高。毕竟，AI可以帮你降低成本，但不能帮你“免责”。本文由「杨越欣（杨桃）」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

Agent、多模态、应用、算力一天看尽，峰会亮点在此｜5.20日，来现场一起AI未来一氪·2026年05月18日 15:03第四届中国AIGC产业峰会5月20日北京举办，干货满满开放报名。进入2026，AI愈发狂飙突进。围观体验之余，人人不免在心中自问：朋友圈刷屏的“龙虾”、Harness等AI新事物，跟我到底有什么关系？真的有必要跟吗？AI创业、AI融资如火如荼，属于我的机会又在哪里？别人已经在用AI做视频、写代码、跑项目，我是不是已经慢了一拍？……到最后，几乎所有问题都会汇成同一个问题：我，到底该如何用AI？如果你对这些问题还很模糊，不妨来第四届中国AIGC产业峰会走一趟——一天时间，把这一年AI产业最值得关注的人、事、判断，一次性讲清楚。先提前剧透一波。18位重磅嘉宾，1场Agent主题圆桌，1份年度榜单，1张全景图谱——所有内容都已敲定。阵容上，既有昆仑万维、智谱、商汤、百度、蚂蚁、MiniMax这样的行业头部，也有亚马逊云科技、硅谷Fusion Fund带来的全球视角；既有复旦邱锡鹏、港大黄超代表的学界与开源前沿，也有盛大EverMind、太初元碁、趣丸科技等活跃实战派玩家。议程上，从Agent商业化落地、多模态与空间智能的最新突破，到AI在文娱、医疗、企业服务等场景的纵深渗透，再到算力与AI Infra的范式之变……这一年最值得关注的AI话题，几乎都装进了这一天。话不多说，下面直接上重点。亮点1：Agent到底怎么落地？不同视角玩家都有心得如果说去年大家还对“Agent是不是未来”持观望态度，那么2026年，这个问题的答案已经不言而喻。“Agent不是未来，是现在”——这正是智谱高级副总裁吴玮杰带来的判断，他将分享智谱在Agent商业化落地上的最新实战。但Agent真正走入企业，并非一蹴而就。亚马逊云科技产品技术部技术总监王晓野将聚焦如何跨越Agent落地鸿沟，拆解从最强模型到企业级AI Agent之间那条最难走的路。而港大助理教授黄超则会从学术与开源前沿的视角，探讨AI Agent范式变革的挑战与机遇。他所带领的HKUDS GitHub开源平台累计获得超24万Star，位列全球Top50，话语权毋庸置疑。与此同时，Agent要真正成为“数字生产力系统”，还需要更底层的能力支撑。盛大集团副总裁、EverMind CEO邓亚峰将分享长期记忆驱动的自进化，探讨AI如何从一个“工具”演化成可持续成长的数字生产力系统。当然，光有演讲还不够。本届峰会专门设置了一场Agent主题圆桌：2026，Agent产品的不确定性与非共识机遇。出席嘉宾包括趣丸科技副总裁兼首席战略官、《屠龙之术》主理人庄明浩，ColaOS&MarsWave CEO冯雷，以及EvoMap创始人张昊阳。这三位都是Agent产品一线的深度玩家，从产品形态到操作系统，从多智能体框架到自进化引擎，他们会碰撞出什么样的非共识观点？非常值得期待。亮点2：Agent之外，模型能力边界也在继续扩张虽然Agent是今年峰会的绝对主角之一，但不可忽视的是，支撑Agent的模型能力本身，也在以惊人的速度向外延伸——从“数字”到“物理”，AI对世界的认知边界正在被重新定义。商汤科技执行副总裁、首席科学家林达华将分享多模态模型如何从“看懂”和“生成”，进化到真正可感知、可生成、可行动的空间智能新阶段。上午的GenAI Talk环节，蚂蚁灵波科技首席科学家沈宇军将围绕一个核心命题展开对谈：AI 2.0下半场，从AIGC到AIGA。当生成式AI从“内容”走向“行动”，整个产业逻辑将被改写。而京东JoyInside业务负责人戴文军则会把这个命题进一步推往物理世界，当AI从屏幕里走出来，走进真实终端、真实场景，“AI World”才真正开始。模型层面，复旦大学特聘教授、模思智能创始人邱锡鹏将分享MOSS多模态模型及其推理优化的最新进展。这位主持研发了MOSS系列开源大模型的“大模型一线人物”，将带来学术与工程结合的硬核内容。亮点3：具体怎么玩？这些应用玩家甩出成功案例模型也好、Agent也罢，具体到不同领域，机会到底在哪？昆仑万维董事长兼CEO方汉率先给出他的答案。他这些年带着昆仑万维在AI音乐、AI社交、AI短剧等多个AIGC方向同时下注，其分享会从一线视角，讲讲AIGC具体应用这一年究竟跑通了什么、又踩过哪些坑。紧接着，来自不同赛道的“实战派”玩家也将依次登场：文娱创作领域，风行在线CEO易正朝将带来一个核心判断——从AI编程到AI视频，众创才是AI生产力的核心杠杆。他会讲述AI如何从单点工具，进化为撬动整个创作者生态的杠杆，从而让普通人也能撑起一支创作团队。生产力工具层面，百度秒哒产品总经理朱广翔会分享他们“人人都是创造者”的产品理念，当门槛被AI抹平，想象力就是新的生产力。企业级AI方向，MiniMax ToB中国区商业化负责人胡维琦将带来MiniMax这一年在ToB战场上的真实打法与思考。大健康赛道，轻松健康集团执行副总裁马孝武将拆解从循证智能体到AI驱动的健康服务增长飞轮，展示AI在严肃医疗场景里的真实增长引擎。文娱、办公、企业、医疗……AI不再是“被讨论的对象”，而是“正在被使用的工具”。每一个细分赛道里，都已经有人开始把案例端上桌了。亮点4：从训练到推理，算力的故事也在被改写模型与应用之外，2026年最值得关注的另一条主线，是AI基础设施的范式重构。过去几年，所有人讨论算力都默认一个前提——为训练服务。但2026年，风向变了，推理需求暴涨、训练成本高企，整个算力叙事正在从“训练驱动”转向“推理驱动”。Fusion Fund创始合伙人张璐将从全球资本与产业的双重视角，解读算力叙事的重构。当AI基础设施从训练转向推理，一场范式变革正在发生。而国产算力这一边，太初元碁首席产品官、高级副总裁洪源将带来他对国产算力路径的思考。当“Token”正在成为AI时代的新计量单位，国产算力如何筑基Token智能新未来？他会给出自己的答案。亮点5：看尽AI全景生态，一份榜单+一份报告延续往届传统，本届峰会，技术先锋、产业龙头、新锐势力将再次全景式碰撞。为了给行业更多参考，我们还将颁出2026年值得关注的AIGC企业、2026年值得关注的AIGC产品两类奖项，以及一份中国AIGC应用全景图谱。前者是量子位根据过去一年AIGC企业、产品的表现与反馈，并结合对2026年技术与场景的观察，评选出的高潜力企业和产品。后者是量子位智库绘制的一幅全面立体的中国AIGC应用全景图谱，用以展示国内AIGC应用的市场格局与发展动态，并深入剖析最具潜力与竞争力的产品和赛道。5月20日举办最后，时间地点以及最全议程在此，现场还有参会伴手礼等你来领，咱们到时候见 ~时间：5月20日 9:30地点：北京·金茂万丽酒店现在，中国AIGC产业峰会正在火热报名中，欢迎点击链接报名线下参会：https://hdxu.cn/1Hjr1本次峰会也将同步线上直播，欢迎预约参与：量子位视频号。参会嘉宾简介一览方汉昆仑万维董事长兼CEO方汉毕业于中国科学技术大学近代物理系，曾任职于中国科学院高能物理研究所、拓林思、亚信以及千橡互动。于2008年协助周亚辉先生创立昆仑万维。方汉拥有30年互联网从业经验，从1994年开始参与和倡导开源运动，是中文Linux奠基人、中文Linux四剑客之一，著作国内第一本Linux书籍《Linux实用大全》，他也是国内最早的网络安全专家，曾负责三大电信运营商的全网安全审计，领导开发了国内第一款P2P下载软件DUDU加速器，负责研发了国内市场占有率最高的网页游戏《三国风云》，为中国互联网最资深的参与者。吴玮杰智谱高级副总裁吴玮杰目前担任智谱公司高级副总裁，负责公司MaaS平台商业化，并兼任泛互联网事业部总经理及浙江子公司总经理。他在人工智能、互联网和信息技术领域有着丰富的商业化管理经验和项目落地经验，曾任字节跳动企业应用全球首席商业官（旗下品牌包括飞书、LarkPeople和幕布等）、GE数字集团大中华区副总裁、复星集团联席首席发展官&首席投资官。除此以外，他也曾在思科、SAP、Oracle等公司从事商业化管理工作。吴玮杰博士毕业于电子科技大学，同时持有香港大学国际工商管理硕士学位以及复旦大学软件工程学士、硕士学位。易正朝风行在线CEO易正朝，风行在线CEO，作为业务领头人，引领公司从传统互联网视听平台全面转型升级为AI文娱生态平台。推动“AI赋能众创”战略落地，带领团队打造橙星梦工厂、小剪刀等AI创作工具，影视网文智能小程序平台、电视鸿蒙平台等核心智能产品，构建从内容创作、智能分发到商业变现的全流通链生态。推动公司荣获AI赛道奖项、抖音快手等平台的头部服务平台奖项，在AI文娱新时代，推动内容生产效率革新与创作者生态繁荣，致力于让内容流动更简单。林达华商汤科技执行副总裁、首席科学家林达华教授现任香港中文大学信息工程系副教授、交叉学科人工智能研究所所长，并为商汤科技联合创始人、执行董事兼首席科学家，兼任上海人工智能实验室领军科学家。他于2012年获麻省理工学院计算机科学博士学位，研究方向涵盖计算机视觉、深度学习、大模型与生成式人工智能，在人工智能顶级会议和期刊发表论文逾300篇，引用逾90,000次，h-index达131。林教授曾获2010年NeurIPS最佳学生论文奖，带领团队在多项国际计算机视觉竞赛中获奖，连续五年入选Elsevier全球前2%顶尖科学家。他主导发起OpenMMLab，并推动书生大模型、OpenCompass评测平台及DeepLink开放计算体系建设，其中DeepLink于2024年获世界人工智能大会“卓越人工智能引领奖”。2026年，他荣获“中银香港科技创新奖”，以表彰其在计算机视觉与多模态智能领域的系统性创新。他长期服务学术共同体，曾任IJCV编委，多次在人工智能主要国际会议担任领域主席，并担任IEEE大规模深度学习模型评测标准工作组主席。邓亚峰盛大集团副总裁、EverMind CEO邓亚峰，毕业于清华大学，正高级工程师，拥有二十余年人工智能算法及产品研发经验。现任盛大集团副总裁、EverMind CEO。曾任360集团副总裁、人工智能研究院院长兼搜索事业部总经理，科创板首家AI上市公司格灵深瞳CTO。曾荣获2021年中国人工智能年度十大风云人物，教育部技术发明奖二等奖。长期致力于大语言模型、计算机视觉及AI for Science等领域的研究与应用，带领团队多次在国际国内主流AI评测中获得佳绩；累计申请发明专利160余项（已授权98项），发表论文50余篇（含10篇Nature子刊）。王晓野亚马逊云科技产品技术部技术总监王晓野现任亚马逊云科技产品部技术总监。长期专注于企业现代化应用架构，数据及人工智能能力的建设。负责带领技术专家团队提供云上产品及开源方案技术实践指导，并致力于亚马逊云科技全栈服务在中国市场的应用和推广。沈宇军蚂蚁灵波科技首席科学家沈宇军，现任上海蚂蚁灵波科技有限公司首席科学家，同时领导蚂蚁技术研究院交互智能实验室。博士毕业于香港中文大学，研究方向为计算机视觉、生成模型、空间智能和具身智能。在CVPR、TPAMI、ICCV、ECCV、NeurIPS等国际顶级会议和期刊上发表论文百余篇，Google Scholar引用量逾万次。2024年起担任蚂蚁灵波科技首席科学家，致力于探索计算机视觉在机器人行业的落地之路，带领团队构建从空间感知到智能决策、从基座能力到世界建模的完整技术矩阵，推动具身智能从实验室走向规模化应用。邱锡鹏复旦大学特聘教授、上海创智学院院长助理、模思智能创始人邱锡鹏，CAAI Fellow，国家杰青获得者，主要研究方向为大模型、智能体以及科学工程等领域中的应用，发表学术论文被引用4万余次，5次获得AI领域高水平会议的最佳/杰出等论文奖，担任中国人工智能学会自然语言理解专委会副主任、中国中文信息学会大模型与生成专委会副主任、上海市计算机学会自然语言处理专委会主任等，主持研发了MOSS等系列高影响力开源大模型。连续多年入选中国高被引学者、全球前2%顶尖科

科氪 | 性能极限与跨界尝试：红魔11S Pro系列深度体验科氪·2026年05月18日 17:35红魔11S Pro电竞旗舰测评：纯平硬核，性能拉满在智能手机行业普遍面临镜头模组日益凸出、机身内部堆叠空间受限的背景下，电竞旗舰由于其对散热和功耗的极端要求，往往表现出更为激进的工业设计。作为新一代的性能导向型产品，红魔11S Pro系列通过软硬件的垂直堆叠，试图在移动终端上探索更进一步的性能上限。以下我们将从设计、理论性能释放、以及实际软硬件体验三个维度，对其进行深度拆解。一、 外观设计：纯平后盖的内部堆叠与电竞美学从视觉和握持逻辑来看，红魔11S Pro最显著的特征在于其坚守的纯平后盖设计。在行业因大底传感器而集体妥协于镜头凸出的趋势下，这种“反向”设计在工业制造上面临着极高的空间挑战。产品图片内部空间重构：为了在不增加机身异常厚度的前提下抹平镜头凸起，红魔对内部架构进行了重新设计。在有限的Z轴高度内，开发团队不仅压榨空间塞入了新一代的脉动水冷模块、无线充电线圈，还容纳了大面积的VC散热板和高密度电池。产品图片透明电竞美学的视效迭代：机身背部延续了红魔标志性的透明设计语言。通过透明材质，内部的脉动水冷引擎实现了可视化。在水冷系统运行时，用户可以观测到液体流动的动态秩序。相比于纯粹的RGB灯光，这种工业风的设计提供了一种更具技术感的视觉回馈。材质与细节处理：后壳采用了微蚀刻工艺打造的浮雕星轨纹理，搭配环形拉丝处理，使得整体质感较为收敛和冷峻。机身在肩键、腰部、LOGO及风扇处均设有支持自定义的RGB灯效，在暗光游戏环境下能提供必要的电竞氛围。产品图片二、 性能与散热：高标准筛选的“特种兵”芯与风水双冷架构性能表现是电竞手机立足的核心。红魔11S Pro首发搭载了第五代骁龙8至尊领先版移动平台。为了在重负载下维持超频表现，红魔在芯片筛选和散热架构上采用了更严苛的标准。产品图片芯片体质的物理筛选：为了满足电竞长时高频的需求，红魔所采用的这颗“领先版”芯片，主要筛选自晶圆中心区域。由于物理刻蚀规律，该区域的芯片通常具备更稳定的温度控制和更高的电气性能。在此基础上，通过最高4.74GHz主频的冲刺分级测试，以及高温高压下的自适应电压调节（AVS）检测，筛选出具备自适应能力的芯片，以确保在极端重负载下不易因发热而触发激进的降频保护。风水双冷复合散热系统：针对第五代骁龙8至尊领先版的发热特性，红魔在散热上引入了原本常见于服务器端的复合方案。手机将【驭风4.0】贯穿式主动散热风扇与脉动水冷引擎相结合，辅以复合液态金属3.0和4D冰阶VC。风扇与整机均通过了IPX8级防水认证，在提升主动散热效率的同时，规避了由于进水导致元器件损坏的风险。在散热系统运转时，后盖的蓝色轨道圈内，蓝白相间的液冷介质会随之一起转动。仪式感拉满！产品图片独立显存与能效调度：在架构协作上，红魔自研的红芯R4电竞芯片与高通联合打造了独立的高速显存。在图形高渲染场景下，此设计可减少高频访问系统内存的次数，从而降低总功耗。结合CUBE擎天游戏引擎3.0的AI感知调校，系统能够精细化调配算力，平衡了高帧率表现与续航消耗。三、 使用体验：悟空屏2.0、Steam直连与续航评估在日常交互和实际游戏运行中，硬件的堆叠最终转化为了多场景的体验。悟空屏2.0的视觉完整性：正面配备的悟空屏2.0基于京东方X10全新屏下发光材料，实现了全无挖孔的视觉体验。屏幕黑边控制在1.25mm，屏占比达到95.3%，支持144Hz刷新率。由于没有刘海或挖孔遮挡，在全局视野和视频播放时拥有更纯粹的画面呈现。同时，屏幕集成了全屏3D超声波指纹解锁，提升了暗光或湿手状态下的解锁盲操率。产品图片行业首发Steam直连：性能的跨界探索：这是该机型在软件生态上的重要尝试。依托转译引擎，红魔攻克了X86与ARM架构之间的转译难题。用户拨动竞技键进入游戏空间后，可直接扫码登录个人的Steam账号。系统不仅支持游戏库内兼容作品的下载游玩，还打通了云端存档。用户在PC端的进度可以在手机侧的PC模拟器上无缝同步，迈出了打破移动端与桌面端算力壁垒的一步。产品图片产品图片产品图片产品图片操控精准度与手汗免疫：针对瞬时爆发的操控需求，该机引入了全新的新思触控芯片，在MagicTouch4.0算法支持下，实现了3000Hz的瞬时触控采样率。该方案优化了湿手控制算法，一定程度上避免了玩家因手汗导致的断触或漂移现象。续航与快充系统：为了支撑高功耗输出，红魔11S Pro在续航规格上给出了较高配置。PRO版本搭载了8000mAh的大容量电池，显著延长了重度游戏下的续航时长。而PRO+版本则在7500mAh电池的基础上，配备了120W有线快充与红魔首款80W无线闪充，在充电灵活性和回血速度上取得了较好的平衡。产品图片产品图片日常AI功能落地：在非游戏场景下，红魔11S Pro作为主力机也搭载了丰富的日常工具。AI帮写、AI消除、AI超清文档等功能提升了商用和办公效率；影像系统则在保留Live Photo的同时，延续了红魔标志性的星空与车轨慢门拍摄功能，拓宽了产品的实用边界。产品图片结语红魔11S Pro系列是一部特色鲜明、目的纯粹的硬件设备。它通过风水双冷散热、高标准芯片筛选以及无孔全面屏，将移动端的游戏性能推向了新的极端，并通过Steam直连展现出了打破平台壁垒的野心。对于不愿向主流“大镜头凸起”妥协、且对高负载游戏及跨平台游玩有硬性需求的极客与核心玩家而言，这是一台工业特色鲜明、技术指标扎实的旗舰标杆。产品图片本文由「科氪」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

番茄小说正推动 AI 动漫上院线，开放头部IP改编权限｜36氪独家兰芥·2026年05月18日 17:46字节跳动为番茄IP定下了影响力目标。作者｜兰杰 李晓霞编辑｜乔芊36氪独家获悉，据多位接近字节跳动人士消息，字节旗下的番茄小说正在推动AI动漫上院线。在这一过程中，会开放番茄小说平台头部IP的改编权限。之于番茄小说，如何更好地挖掘平台上好IP的潜力和商业价值，是一个亟待解决的难题。一方面，以文字形式为载体的IP内容，缺乏统一的视觉形象，使得平台上的IP难以有联名、衍生品等更丰富的变现形式。另一方面，AI漫剧这一新事物迅速发展壮大，其中不乏改编自番茄小说IP的粗制滥造之作。这些内容不但没能延长好IP的生命力，反而形成了一种消耗。此时，通过可登陆院线的影视内容来探索平台IP的更多可能性，不失为一种好的尝试。另据上述知情人士表示，字节跳动内部为番茄IP定下了影响力的目标，“想让行业内外的人觉得番茄的出品是精品的、正向的。”这一需求落到具体的业务上，上院线会是个不错的答卷。早在2023年年底，番茄小说就官宣成立了“番茄影视”和“番茄动漫”两大厂牌，推动平台IP的影视化。过去两年间，番茄影视先后与腾讯视频、爱奇艺和优酷等长视频平台合作，推动了《十日终焉》《我在精神病院学斩神》等平台头部IP的改编，但这些合作多停留在IP授权阶段。虽然番茄系（字节旗下以番茄小说、红果短剧等为主的业务）并不愁关于钱的问题，但显然成本高、回报周期长的长视频生意，并不符合依赖大手笔投流、需要市场快速反馈的“字节逻辑”。直到AI的出现，大幅降低了成本、提高了效率。36氪还从另一位知情人士处获悉，红果漫剧侧如今正在采购120分钟不分集的长内容，这也是当下各大视频平台争相竞逐的方向。之所以上线AI动漫内容，而不是AI仿真人内容，或许与当前AI大模型的能力有关。行业头部的AIGC创作者向36氪表示，如今视频生成大模型的能力已经可以做到生产120分钟的长内容，但是在仿真人这一赛道上，人物、场景的一致性依旧是个难题，而动漫内容可以通过部分突出的人物形象设计——例如彩色的头发、雀斑等，一定程度上弥补一致性的不足。回顾番茄系并不算漫长的发展历程，它可以说是字节跳动在文娱领域的一次突围。在番茄小说问世之前，字节始终未能建立起足够的内容能力和IP积累。而如今，从小说到短剧，再到漫剧，番茄系已具备了IP开发的全链条能力。与此同时，番茄系显然不满足于仅仅停留在当下的免费市场，而是在探索更多元的商业模式，包括线上订阅和院线票房等更多商业可能性。本文由「兰芥」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：IC photo

氪星晚报 ｜百度：一季度营收321亿元，AI业务收入136亿元；马斯克预计今年美国将广泛使用自动驾驶汽车；巨力索具旗下十余家企业已注销卜算籽·2026年05月18日 19:11特朗普：本应要求获得更多英特尔公司的股份大公司：索辰科技：物理AI业务目前尚处于布局初期36氪获悉，索辰科技公告，公司股票于2026年5月14日至5月18日连续三个交易日内收盘价格涨幅偏离值累计达到30%，属于股票交易异常波动情形。公司近期实施的募投项目“物理合成数据核心技术基座构建与多领域场景示范应用项目”尚处于筹备阶段，若出现募集资金投资项目未能顺利实施、新技术开发进度不达预期、研发遭遇技术瓶颈甚至失败，将会对公司经营业绩造成一定影响。公司物理AI业务目前尚处于布局初期。物理AI技术研发投入大、技术迭代快，存在研发进度不及预期、核心技术迭代风险；同时下游行业应用场景拓展、商业化变现存在不确定性，若市场需求、行业政策、技术竞争格局发生不利变化，物理AI业务收入及盈利贡献存在不及预期风险。智元WITA成为全国首例完成大模型备案的具身智能交互模型36氪获悉，根据上海市网信办最新公告的上海市生成式人工智能服务备案情况，智元WITA成为全国第一款实现合规备案的具身智能交互大模型，标志着中国具身智能产业从“技术验证”进入“合规商用”的新阶段。依托智元“三智一体”全栈技术体系，智元WITA核心应用于人形机器人交互场景。上交所：上周新增8家IPO受理企业36氪获悉，据“上交所企业上市服务”消息，本期（5月11日—5月17日）新增8家受理企业，其中上交所主板1家，科创板3家，创业板3家，北交所1家。截至5月17日，今年新增31家受理企业。截至5月17日，各板块在审企业共295家。其中上交所62家（主板18家，科创板44家），深交所61家（主板17家，创业板44家），北交所172家。截至5月17日，2026年已公布终止审查（撤材料+否决/终止注册）企业20家，其中上交所7家（主板3家，科创板4家），深交所创业板1家，北交所12家。腾讯音乐：收购喜马拉雅的事项已经完成36氪获悉，腾讯音乐在港交所公告，于2026年5月18日，根据并购协议的条款，收购喜马拉雅的事项已经完成，据此，喜马拉雅相关股东及喜马拉雅的雇员持股计划参与者持有的喜马拉雅权益性证券已予以注销，以换取并购对价，该对价包括（视适用调整而定）总额最高为12.6亿美元的现金及最多总数1.75亿股本公司A类普通股（包括本公司于交割时新发行的A类普通股，以及本公司授出的以股权为基础的激励相关的A类普通股）。于收购事项交割后，喜马拉雅已成为本公司的全资附属公司。爱奇艺：一季度总收入62.3亿元，会员收入环比增长36氪获悉，爱奇艺发布2026年第一季度财报。财报显示，一季度爱奇艺总收入62.3亿元。其中，会员服务收入42.0亿元，环比增长2%。在线广告服务收入12.4亿元，内容发行收入3.6亿元，其他收入4.3亿元。百度：一季度营收321亿元，AI业务收入136亿元36氪获悉，百度发布2026年第一季度财报。财报显示，一季度营收321亿元，上年同期营收324.52亿元。一般性业务收入260亿元，同比增长2%；其中AI业务收入136亿元，占百度一般性业务收入的52%，已连续多个季度增长；GPU云收入同比增长184%。百度创始人李彦宏表示，AI业务收入占比的首次过半，标志着AI已成为百度的核心驱动力。巨力索具旗下十余家企业已注销36氪获悉，近日，巨力索具因涉嫌信息披露误导性陈述违法违规，被证监会立案调查。天眼查App显示，巨力索具股份有限公司成立于2004年12月，法定代表人为杨建国，注册资本9.6亿人民币，经营范围包括应急装备、装备专用工装及工具和箱组的设计、生产和销售等，由巨力集团有限公司、杨建忠、香港中央结算有限公司等共同持股。对外投资信息显示，该公司对外投资26家企业，其中10余家企业处于存续状态，包括巨力索具（河南）有限公司、河北巨力应急装备科技有限公司等，其他10余家企业已注销。投融资：“华羽先翔”完成超亿元Pre-A+轮融资，蚂蚁集团领投36氪获悉，近日，国内全倾转eVTOL企业“华羽先翔”完成超亿元Pre-A+轮融资，本轮融资由蚂蚁集团领投、彼岸时代及复星创富跟投。融资资金将重点用于加速鸿鹄系列eVTOL的适航取证，以及低空出行及商业场景的落地与创新。“极壳科技”获5000万美元B+轮融资36氪获悉，近日，外骨骼品牌“极壳科技”宣布完成5000万美元B+轮融资。此轮融资由蚂蚁集团和美团龙珠领投，Sofina、Granite Asia跟投，高鹄资本担任独家财务顾问。此轮融资将加速极壳在技术研发、全球市场拓展及消费场景应用方面的布局。新产品：酷开发布企业AI操作系统36氪获悉，近日，深圳市酷开网络科技股份有限公司举办“场景智能·硅基管理”发布会，正式推出原生AI企业操作系统——酷开Happy Work AIOS Lite MVP。该系统已实现从生活场景到企业经营场景的全维度覆盖，完成多场景AI生态闭环落地。今日观点：马斯克预计今年美国将广泛使用自动驾驶汽车埃隆·马斯克周一表示，他预计今年晚些时候，无需人类驾驶的自动驾驶汽车将在美国更加普及。马斯克通过视频连线在特拉维夫举行的“智能出行峰会”上发言时称，得克萨斯州的三个城市已有无人驾驶汽车在没有安全监视员的情况下运行，并补充说今年这一模式将扩展至全美。（新浪财经）其他值得关注的新闻：2026电影总票房已超145亿元据票务平台数据，截至今天下午，2026年电影票房已超145亿元。五一档新片长尾效应持续释放，其中《给阿嬷的情书》连续九天蝉联单日票房榜首。（央视新闻）特朗普：本应要求获得更多英特尔公司的股份美国总统唐纳德・特朗普在周一出版的杂志采访中表示，他本应代表美国政府要求获得更多英特尔股份。特朗普政府去年持有英特尔10%的股份，并宣布向这家芯片制造商投资约100亿美元，用于在美国境内新建或扩建工厂。交易完成八个月后，美国政府持有的英特尔股份市值已超过500亿美元。（新浪财经）国家药监局：我国4月批准注册医疗器械产品318个今天从国家药监局获悉，我国4月共批准注册医疗器械产品318个。其中，境内第三类医疗器械产品272个，进口第三类医疗器械产品20个，进口第二类医疗器械产品26个。 (央视新闻)本文由「卜算籽」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：AI生成

1天前无锡建Token工厂，首批用4台华为昇腾384超节点服务器5月15日，弘信电子旗下燧弘华创与无锡高新区签约，将在无锡建设全国首个“东数西算、算电融合”的特大型 Token 工厂项目，并同步落地江苏省内首个华为昇腾384超节点算力集群。 首期将在无锡高新区部署4台昇腾384超节点服务器，每台集成384卡规模算力，单套算力最高可达约215 PFLOPS，将4台连成一体后形成国内面向市场端数一数二的大规模国产 AI 算力集群，服务“国芯国模国用”战略，面向长三角及全国输出标准化 Token 级算力能力。7 来源项目签约落地：无锡承接全国首个特大型 Token 工厂4 来源首批部署4台华为昇腾384超节点：江苏省内首个集群7 来源“Token 工厂”模式：把西部绿电算力“加工成”标准化 Token2 来源对无锡及国产算力生态的战略意义5 来源本内容由AI生成

8小时前电子布需求旺盛，上市公司加速扩产在AI服务器等下游高阶PCB需求爆发带动下，电子级玻璃纤维布今年延续结构性紧缺，高端电子布“一布难求”，自2025年10月以来已连续多轮提价，主流型号7628电子布今年累计涨幅接近40%，行业库存处于历史低位。　在量价齐升刺激下，中国巨石等龙头密集上马新产线、规划数亿米级新增产能，多家上市公司通过扩产、定增等方式加码高端电子布赛道，券商预计2026年或成为电子布涨价“大年”。7 来源AI算力驱动高阶PCB放量 电子布全面进入涨价周期4 来源中国巨石领衔龙头扩产 新增产能以高端电子布为主5 来源多家上市公司业绩受益 资本加速涌入高端电子布4 来源机构研判：2026年或成涨价大年 关注供给爬坡与品种分化4 来源本内容由AI生成

8小时前AI生成蛋白质完整动态模型麻省理工学院与卡内基梅隆大学团队近日发布端到端动力学蛋白质生成模型 VibeGen，将序列生成与简正振动模预测统一到一个闭环系统中，在保证结构稳定与新颖性的同时，对主链振动幅度分布进行精确控制，被业界视为从“静态结构预测”迈向“完整动态模型”设计的关键一步。 与此同时，多家团队正利用生成式模型和AI分子动力学模拟，低成本逼近蛋白质的玻尔兹曼加权构象分布，为药物设计和功能预测提供接近量子化学精度的全原子动态图景。3 来源AI 从静态结构迈向完整“动态电影”3 来源VibeGen：首个端到端动力学蛋白质生成智能体1 来源在复杂动力学约束下精确“雕刻”蛋白运动1 来源从构象景观到AI分子动力学：迈向真正的“完整动态模型”3 来源本内容由AI生成

7小时前新版Siri支持自动删除聊天记录彭博记者古尔曼透露，苹果将在 2026 年 WWDC 上预览全新独立 Siri 应用，该应用由谷歌 Gemini 模型提供算力，但运行在苹果自有私有云端，主打隐私保护，并引入类似 iMessage 的“聊天记录自动删除”功能，用户可选择在 30 天或 1 年后自动清除 Siri 对话，或永久保留；新 Siri 将随 iOS 27 于今秋面向大众推出，且在正式发布阶段仍将标注“测试版”。2 来源新版 Siri 的核心变化：从语音助手到聊天机器人2 来源聊天记录可自动删除：与 iMessage 看齐的隐私选项2 来源Gemini 加持但数据留在苹果：更严格的数据使用边界2 来源发布时间表与测试节奏：随 iOS 27 推出，长期打磨2 来源本内容由AI生成

7小时前超级工程师、仓储大脑、监测之眼…… 潍柴工厂里人工智能成主力在山东潍坊潍柴动力2号工厂，人工智能已从“辅助手”升级为“主力军”：研发环节由“AI超级工程师”参与近四成模型决策并借助仿真将实物试验减少约70%，生产与物流由融合大模型和小模型的“智能生产中枢”“仓储大脑”统一调度，质检与安能环管理依托“监测之眼”实现全流程智能监测。凭借覆盖研发、制造、管理、服务等近八成业务场景的AI应用体系，潍柴入选全国首批领航级智能工厂，成为“数智精益”制造模式的标杆。7 来源“AI超级工程师”：发动机研发有了“第二大脑”4 来源“仓储大脑”：打通工厂物流的“神经网络”4 来源“监测之眼”：从质量到安能环的全场景守护4 来源从卓越到“领航”：潍柴打造“数智精益”工厂样板5 来源本内容由AI生成

7小时前特斯拉推送FSD大更新在车主社区与社交媒体上，有消息称特斯拉近期正面向部分车辆推送一次完全自动驾驶（FSD）软件“重大OTA更新”，被视为延续FSD v12之后、进一步强化端到端神经网络和城市道路能力的重要一步。 由于本助手无法实时访问互联网、也无法核验最新固件版本号和推送范围，以下报道主要基于截至2024年10月前特斯拉公开资料及过往FSD大版本更新规律，对这类“FSD大更新”的典型技术方向、安全争议与行业影响进行系统梳理和分析，仅供参考。推送背景：FSD从“测试版”走向更大规模商用技术走向：端到端神经网络、城市道路与复杂场景表现安全与监管：迭代速度与风险争议并存商业与行业影响：订阅模式、竞争格局与未来走向本内容由AI生成

5小时前前微软高管批评公司AI战略失误，称重蹈互联网和移动设备时代覆辙近日，有媒体援引一位前微软高管的观点称，微软在人工智能战略上的一些做法，正在“重蹈互联网和移动设备时代的覆辙”，过度依赖单一合作伙伴、产品落地节奏缓慢、内部协同僵化等老问题再现，引发外界对其能否在新一轮AI竞赛中保持领先的担忧。2 来源说明与背景：基于现有认知的新闻体撰写2 来源前微软高管：AI 战略暴露“老问题”，或重演互联网与移动时代失误2 来源从错失互联网到折戟移动：微软“平台焦虑”的历史镜像2 来源AI 时代的布局与质疑：微软会再错过一次平台革命吗？2 来源本内容由AI生成

3小时前腾讯AI设计智能体协作平台Ardot正式开启公测腾讯设计团队近日正式宣布其 AI 设计智能体协作平台 Ardot 开启公测，面向设计师、产品经理和开发者开放申请制免费使用。平台将 AI 生成、UI 设计、团队协作与开发交付整合到同一工作台，并以“描述即生成”的文本转 UI 为核心亮点，主打用自然语言快速生成与迭代界面原型。 公测阶段被视为腾讯在设计生产力工具上测试 AI 智能体协作模式、探索商业化路径的重要一步。1 来源Ardot正式开启公测：腾讯押注“AI设计智能体协作”1 来源文本转UI与多智能体协作：Ardot的核心技术亮点1 来源覆盖从个人到大中型团队的多元设计场景1 来源公测策略与行业影响：AI设计平台竞争进入深水区1 来源本内容由AI生成

2小时前Figure机器人91小时直播分拣包裹，行业景气度提升美国人形机器人公司 Figure 近日在海外仓库开启“7×24 小时直播打工”项目，其人形机器人在真实物流分拣场景中连续上岗，累计直播时长已超过 91 小时；据科创板日报援引公开数据，该机器人在前 30 小时内已处理约 4.5 万个包裹，稳定性与效率表现超预期，叠加国内外具身智能与物流机器人技术快速突破，被视为本轮机器人产业景气度加速上行的新信号。2 来源Figure 人形机器人开“7×24 小时直播间”，不间断分拣包裹超 91 小时1 来源30 小时处理 4.5 万包裹，验证效率与稳定性2 来源具身智能与“手眼协同”支撑复杂分拣场景1 来源机器人景气度上行，产业链与资本市场情绪同步升温2 来源本内容由AI生成

1小时前Isomorphic Labs筹集21亿美元扩展AI药物设计引擎Alphabet 旗下专注 AI 药物研发的子公司 Isomorphic Labs 宣布完成 21 亿美元 B 轮融资，新投资方包括 MGX、淡马锡（Temasek）、CapitalG 及英国主权 AI 基金，资金将用于强化其 AI 药物设计引擎 IsoDDE、加速候选药物研发并推动全球业务扩张。4 来源巨额 B 轮融资落地：主权财富基金与科技资本押注 AI 药物4 来源从 DeepMind 分拆而来：用 AI 重塑药物发现流程3 来源21 亿美元用途：强化 IsoDDE、加速管线推进与全球布局4 来源AI+药物赛道升温：硅谷与主权资本加码，质疑与期待并存4 来源本内容由AI生成

29分钟前今年京东体系AI研发投入增长将超200%在今年京东618启动发布会上，京东集团技术委员会主席、京东云总裁曹鹏透露，2026年全年京东体系AI相关研发投入增幅将超过200%，投入规模进入行业第一梯队；在此前2025年四季度和2026年一季度，京东体系研发投入已分别同比增长66%和约59%，AI正从“成本中心”加速转变为驱动营收与利润提升的核心引擎。5 来源京东618发布会：AI研发投入增幅将超200%，用户与终端“双爆发”1 来源财报视角：连续两季超50%增速，AI从成本变利润引擎4 来源AI深度嵌入全链路：零售、供应链到健康、家居4 来源技术护城河加厚：从JoyAI大模型到开发者与生态开放5 来源本内容由AI生成

9分钟前马斯克：今年首例Neuralink植入手术治疗失明据用户提供的新浪财经等报道，埃隆·马斯克在最新一次Neuralink发布会上表示，计划在今年内实施首例专门面向失明患者的Neuralink脑机接口植入手术，尝试通过“Blindsight（盲视）”项目恢复部分视功能。

这一系统由内置摄像头的智能眼镜和植入大脑视觉皮层的芯片组成，意在让部分全盲或重度视损伤患者重新感知光点、轮廓和简单物体形状。 与此同时，国外已有如PRIMA等人工视网膜方案在临床试验中让多数患者恢复了“功能性中央视力”，共同推动人工视觉技术从实验室走向临床实践。2 来源马斯克最新表态：今年启动首例“治盲”Neuralink植入1 来源Neuralink“Blindsight”：用眼镜+大脑植入重建视觉通路1 来源与PRIMA等人工视网膜技术的差异与互补2 来源安全性、监管与前景：从“切脑1.5秒”到真正走向临床2 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。智库·502线上同行往期回顾报名已结束企业构建AI+组织的人效杠杆与路径选择百果园集团x银泰分享企业构建AI+组织的人效杠杆与路径选择妙投波动判断场往期回顾活动已结束芯片、光模块、机器人——AI的泼天富贵，还能传到第几棒？本期我们将AI产业链拆解为三大核心环节，逐一拆解其核心矛盾与风险边界，帮你建立从芯片到机器人的完整坐标系。随行灯塔火热报名中随行灯塔随行灯塔是由虎嗅深度联合运营的高端创始人社群。专注于一件事：帮助创始人把产品做对、做爆、做到10亿。

这里不灌鸡汤，不讲玄学。我们相信，伟大的公司一定是产品驱动的；但产品力来自人，而人需要场——一个能被照见、能被挑战、能被链接的场。随行灯塔，就是创始人的“修炼场”。智库·502线下闭门会往期回顾精彩回顾AI在ToB销售、客服的落地案例探讨围绕着AI在ToB营销内容生成、智能客服接待、线索清洗等场景进行深入交流早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文李开复：开源模型是实现“AI主权”的更优路径发布于AI新闻资讯发布时间 :2026年5月12号 18:01阅读 :1分钟创新工场董事长李开复近日在专访中深入探讨了“AI主权”的概念。他指出，人工智能主权不仅关乎技术控制权和数据安全，还涉及模型与本土文化及法律的适配性。他认为，并非每个国家都有必要重造一个OpenAI，盲目追求昂贵的闭源自研并非良策。李开复强调，对于大多数资源有限的国家和企业而言，基于开源模型构建本地化体系是更现实的选择。这种“第三条路”通过在开源基础上进行增量训练，能以极低的成本让模型符合特定地区的价值观，而非受制于他国的通用模型。工程效率与生态之争谈及中美竞争，李开复认为资源受限反而催生了中国企业的极致工程效率。他以DeepSeek为例，指出中国公司在算力不足的情况下，通过优化架构和减少依赖，能以不到美国十分之一的投入实现极高的性能。他预测AI领域将演变成类似“iPhone与安卓”的格局。闭源模型如同一方领地，追求高利润和生态控制力；而开源模型则代表了更广阔的市场覆盖和普及度，两者将在全球范围内长期并存，共同构建未来的智能生态。中国硬件的下半场机遇李开复对中国在物理AI和硬件制造方面的结构性优势表示乐观。他指出，中国拥有全球最完整的供应链体系，从智能穿戴到机器人，制造成本远低于西方。未来AI时代的标杆级硬件，极有可能由中国设计并品牌化。此外，中国在语音驱动交互和大规模现实数据生成方面也具备独特文化优势。随着“人工智能+”政策对实体经济的赋能，硬件与AI的深度结合将成为中国在AI下半场竞赛中的核心竞争力。

AI资讯AI新闻资讯正文广东省新增 6 款已完成登记生成式人工智能服务，腾讯音乐等企业在列发布于AI新闻资讯发布时间 :2026年5月11号 17:53阅读 :1分钟广东省网络信息部门联合相关单位，根据《生成式人工智能服务管理暂行办法》的要求，积极推动生成式人工智能服务的创新发展与规范应用，开展了生成式人工智能服务的登记工作。根据最新公告，截至2026年5月11日，广东省新增了6款已完成登记的生成式人工智能服务，至今已累计完成53款服务的登记，腾讯音乐、格力电器等企业在列。为进一步加强对生成式人工智能服务的管理，确保服务的安全与合规性，特别强调了服务提供方的责任。针对通过 API 接口或其他方式直接调用已备案的大模型能力，并面向国内公众提供具有舆论属性或社会动员能力的服务，相关服务提供方需向属地网络信息部门履行登记程序。这一措施旨在提高服务的透明度，维护社会公共利益。与此同时，所有已上线的生成式人工智能应用或功能，必须在显著位置或产品详情页面标注所获得的上线编号，以便用户能够清晰识别和了解所使用服务的合法性与合规性。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文谷歌推出Gemini Intelligence：AI自动化能力惊艳，但12GB内存门槛或将安卓阵营"一分为二"发布于AI新闻资讯发布时间 :2026年5月18号 14:30阅读 :1分钟当手机不仅能听懂指令，还能自动完成"比价-下单-支付-跟踪"的全流程，移动交互的范式正在被重构。谷歌近日面向安卓平台正式发布Gemini Intelligence高级AI功能套件——这并非一次普通的功能更新，而是一套整合跨应用自动化、系统级调度与隐私保护的系统性AI解决方案，标志着安卓生态正式迈入"智能体驱动"的新阶段。核心体验:长按电源键，让AI替你"跑腿"Gemini Intelligence的最大突破在于跨应用多步骤任务自动化:场景用户指令示例AI执行流程出行"帮我叫一辆去机场的车，选最便宜的"比对滴滴/Uber/高德→选择最优→确认车型→推送待支付通知购物"买一双300元以内的白色运动鞋，要耐克"搜索电商→筛选价格/品牌/颜色→比对评价→生成订单草稿预订"订一家明天晚上、市中心、人均200的日料"调用大众点评/谷歌地图→筛选评分/位置/价格→预留座位→等待确认交互设计极具巧思:用户可通过长按电源键或拍照识物快速触发指令，所有任务在后台并行执行，关键节点（如支付、确认）均需用户手动授权，既保证效率又守住安全底线。高门槛背后的工程逻辑:为何需要12GB内存+旗舰芯片?谷歌公布的配置要求引发行业热议:≥12GB运行内存:支撑多应用并行调度与模型常驻;旗舰级芯片（如骁龙8Gen4、天玑9400）:满足端侧大模型推理的算力与能效需求;系统级AICore框架:实现应用间权限协调与任务状态同步;内核虚拟机安全隔离:确保敏感操作（如支付）在独立沙箱中执行;5次系统大版本+6年安全更新:保障AI能力长期兼容与漏洞修复。这一"豪华配置单"直接将多数中低端安卓设备排除在外。谷歌的考量很务实:复杂的跨应用自动化对系统稳定性、响应延迟、功耗控制要求极高，若强行下放至低配机型，可能导致卡顿、耗电、崩溃等体验灾难，反而损害AI功能的口碑。首批适配机型:高端阵营的"特权游戏"目前已确认的首批支持设备包括:三星:即将发布的Galaxy Z Fold8/Flip8折叠屏系列（首发搭载），以及今年夏季通过系统更新适配的Galaxy S26系列;谷歌:Pixel10系列（预计同步推出）;其他:暂无中小品牌机型进入名单。这一策略传递出明确信号:谷歌正将顶级AI能力作为高端机型的差异化卖点，与苹果"软硬件一体"的封闭生态形成正面竞争。对三星等头部厂商而言，这是巩固高端市场的利器;但对中小品牌，短期内难以跟进的技术门槛可能加剧"马太效应"。行业影响:安卓生态的"智能分层"时代来临Gemini Intelligence的发布，可能重塑安卓产业格局:用户体验分化:高端机型用户享受"全自动智能助理"，中低端用户仍依赖传统手动操作，“数字鸿沟"从接入层面转向能力层面;厂商策略调整:中小品牌或选择"轻量化AI"路线（如单应用助手），或加速向高端市场突围以获取技术授权;开发者生态重构:应用需适配AICore框架以支持自动化调度，“可被AI调用"将成为新开发标准;隐私与信任挑战:跨应用数据流转虽经用户确认，但权限边界、责任认定等议题仍需行业共识。战略深意:谷歌的"安卓护城河"重构在苹果凭Apple Intelligence强化生态粘性的背景下，谷歌此举本质是以系统级AI能力重铸安卓竞争力:对开发者:提供统一的智能体开发框架，降低跨应用自动化门槛;对用户:用"真正能办事"的AI体验，对抗"只会聊天"的通用助手;对厂商:通过技术授权与联合优化，绑定头部伙伴形成高端联盟。当然，挑战仍存:如何平衡功能强大与隐私保护?如何避免高端独占引发用户不满?如何让中小厂商找到差异化生存空间?这些问题，将决定Gemini Intelligence是成为安卓的"破局之矛”，还是"分化之刃”。当一部手机的"智能程度"开始由内存大小决定，移动科技的竞争已从"功能堆砌"转向"系统级协同"。谷歌的这次押注，或许正在定义下一个十年:真正的智能，不在于能回答多少问题，而在于能替你完成多少事——而谁能率先让这件事"人人可用"，谁就将掌握生态的话语权。

AI资讯AI新闻资讯正文通义千问预告“重量级”大模型将于5月20日阿里云峰会亮相发布于AI新闻资讯发布时间 :2026年5月18号 14:36阅读 :1分钟5月18日，阿里云通义千问大模型通过官微宣布，其“重量级新朋友”将于5月20日的阿里云峰会正式亮相，官方预告该模型将在全能性、模型能力、深度与广度上实现全面升级。这一动作标志着国内头部AI大模型在算力迭代与应用落地层面的又一次关键演进。当前，国产大模型正迎来新一轮密集的集体更新与能力跃升浪潮。在市场竞争日趋激烈、技术应用向垂直领域纵深发展的背景下，各大科技巨头纷纷加速产品迭代。通义千问此次在阿里云峰会前夕的高调预告，不仅展示了其在全模态通用能力和深度语义理解上的技术储备，也透露出阿里云试图通过底层模型的关键突破，进一步巩固其在企业级云服务与AI生态中的核心话语权。

AI资讯AI新闻资讯正文学术界重拳出击，arXiv发布新规封杀AI代写论文行为发布于AI新闻资讯发布时间 :2026年5月18号 14:49阅读 :1分钟著名的开放学术预印本存储库arXiv近日出台严厉新规，宣布将对在科研论文中滥用人工智能（AI）的行为进行严厉打击。此举的核心目的并非禁止技术辅助，而是为了遏制部分学者因过度依赖AI生成内容且未仔细核查，导致论文可信度严重下滑的乱象。目前，arXiv正从康奈尔大学运营的网站转型为独立的非营利组织。转型后，该平台将能筹集更多资金，以专门应对和解决近年来愈发严重的AI低质量论文泛滥问题，维护计算机科学和数学等领域的学术传播质量。严厉惩戒幻觉与作弊，违规作者最高禁发一年根据arXiv公布的具体惩罚措施，一旦发现论文中存在确凿的AI幻觉内容，或者作者在论文中暗中加入AI提示词以图获得更高评价，相关作者将面临一年内不得发布新论文的重罚。对于被封禁期满一年的作者，其惩罚并未完全结束。一年后如果他们想在arXiv上重新发布新论文，该论文必须先获得信誉良好的同行评审期刊接受，以此作为额外的学术信用担保。并非全面禁用人工智能，作者须对内容负全责arXiv官方强调，新规的出台并不意味着完全禁止研究人员使用人工智能进行辅助编写。平台依然允许AI作为辅助工具存在，但大前提是论文作者必须对最终呈现的所有内容承担全部责任。这意味着，如果作者直接从AI工具中复制粘贴不当语言、抄袭内容、偏见错误或误导性引用，都将面临处罚。为确保公平，被认定违规的作者拥有上诉权利，arXiv承诺只在证据确凿的情况下才会实施封禁。

AI资讯AI新闻资讯正文中国AI视频生成模型异军突起，创作者赞其“更自由”发布于AI新闻资讯发布时间 :2026年5月18号 15:00阅读 :1分钟根据最新行业报道，中国人工智能公司在视频生成这一关键战场上，已经处于领先地位。随着视频生成技术在广告、电商和娱乐领域的快速普及，字节跳动和快手等公司正迎来高光时刻。这些公司正利用自身拥有的海量短视频库来训练系统，这使其在与美国同行的激烈竞争中占据了天然优势。目前，中国工具在真实场景中的表现已赢得了全球创作者的广泛认可。美方模型频频报错，国产工具凭借高质量破局在实际使用中，许多海外创作者表示，相较于限制繁多且容易报错的美国大模型，中国的产品在画面质量和提示词执行力上表现更好。特别是在音画同步和语音稳定性方面，近期国产工具的改善非常明显。一些独立电影制作人指出，字节跳动最新的Seedance 2. 0 等工具极大地释放了创作自由。它不仅能够处理激进的镜头角度和快速运动，还能在运动中完美保持角色的面部特征，避免了画面抖动。掌控海量视频数据，高性价比商业化落地加速中国视频生成模型的强劲表现，得益于对庞大专有视频数据的掌控，这也是其他西方同行难以复制的竞争护城河。目前，快手旗下的“可灵”视频生成工具表现亮眼，官方甚至正在探索分拆该业务并独立上市的可能性。对于个人创作者而言，中国平台不仅更便宜、也更灵活，普遍提供订阅制或按量付费模式。而在企业端，这项技术正在重塑广告和电商行业，使品牌方能够以前所未有的超低成本，为数万件商品批量定制专属的视频内容。

AI资讯AI新闻资讯正文微软 AI CEO 穆斯塔法·苏莱曼大胆预测：18 个月内，AI 将全面自动化白领工作发布于AI新闻资讯发布时间 :2026年5月18号 15:01阅读 :1分钟近日，微软 AI 首席执行官穆斯塔法·苏莱曼（Mustafa Suleyman）在接受采访时作出惊人预测：AI 将在未来 12 至 18 个月内，在大多数专业任务上达到人类水平表现，从而全面自动化白领办公工作。这一观点迅速引发全球科技界和职场人士的广泛讨论。AI“坐班”时代即将到来苏莱曼指出，那些“坐在电脑前”的工作——包括会计、法律、营销、项目管理等——大部分任务都将被 AI 完全接管。他表示：“我认为我们将在大多数（如果不是全部）专业任务上达到人类水平表现。”这一预测直指传统知识型工作的核心，标志着 AI 从辅助工具向独立执行者的快速转变。根据相关报道，这一时间线指向 2027 年中期左右。苏莱曼认为，随着计算能力的指数级增长和 AI 模型的持续迭代，自动化浪潮将深刻重塑职场结构。构建“超级智能”：AI 开发门槛将大幅降低苏莱曼进一步透露，其使命是打造“超级智能”（superintelligence）。他乐观预测，未来创建新 AI 模型将变得像制作播客或撰写博客一样简单。这一观点凸显了 AI 技术民主化和普及化的趋势，普通开发者或企业或许很快就能轻松部署强大 AI 系统。职场变革的机遇与挑战AIbase 分析：苏莱曼的预测虽具前瞻性，但也引发了关于就业市场、技能转型和社会适应的热议。一方面，AI 自动化可显著提升生产效率，解放人力投入更高价值创造；另一方面，会计、法律、营销等领域的从业者需尽快掌握 AI 协作能力，以适应新工作范式。目前，AI 对就业的具体影响仍在观察中，但科技巨头的高管预测往往预示着行业投资和战略方向的重大调整。微软等公司正积极布局 AI 基础设施，以抓住这一历史性机遇。AIbase 观点：无论预测最终兑现程度如何，AI 重塑白领工作的趋势已不可逆转。企业和个人唯有主动拥抱变革，方能在新时代占据先机。未来已来，你准备好了吗？

AI资讯AI新闻资讯正文红果短剧回应“取消 AI 仿真人剧本保底”传闻：说法不实，部分剧本仍保留保底发布于AI新闻资讯发布时间 :2026年5月18号 15:03阅读 :1分钟针对市场传闻“红果短剧取消 AI 仿真人剧本保底、全面转向纯分成”的消息，红果短剧相关负责人近日公开回应称该说法不实，强调部分剧本依然保留保底政策。此前，有市场消息透露红果短剧官方人士在编剧群发布通知，称按照4月27日的最新政策，新选中的剧本等级将仅提供纯分成合作，分成比例为20%，但若满足后验热度激励指标，仍可享有单项激励活动，且剧本等级随质量提升可持续升级。该消息在编剧圈引发震动，部分编剧甚至将其戏称为“零元购”，并传出仅有极少数头部编剧的保底收入未受影响。此次“保底风波”的背后，折射出网文与短剧赛道在引入生成式 AI 和 AI 仿真人技术后，平台、编剧与内容生产生态之间正在经历剧烈的利益博弈与机制磨合。作为当前内容出海和本土消费的新风口，AI 短剧通过技术手段大幅降低了制作成本与周期，但这也对传统剧本创作的定价体系和内容品控带来了全新挑战。红果短剧对传闻的澄清，在一定程度上稳定了创作者群体的预期，但平台引入“后验热度激励”与“动态等级调整”的机制，也暗示了行业正在向更加注重数据反馈、精细化运营的绩效驱动型方向演进。如何在效率、技术与创作者核心利益之间寻找平衡，将是未来 AI 内容平台构建可持续生态的关键。

AI资讯AI新闻资讯正文5 天击穿苹果 M5 最强内存防线！AI 辅助完成纯数据提权，Mac 安全格局巨变发布于AI新闻资讯发布时间 :2026年5月18号 15:11阅读 :2分钟2026 年 5 月 14 日，帕洛阿尔托安全研究公司 Calif 在官方博客发布 55 页完整技术报告，宣布一项重大突破：研究团队仅用 5 天时间，就在搭载 M5 芯片、运行最新macOS 的设备上实现了从普通本地用户权限到 root shell 的完整提权，利用过程无需任何代码注入，属于纯“data-only”攻击。该报告迅速引发安全界高度关注。5 天 vs 5 年：极致紧凑的攻击链开发时间线Calif 团队的时间线令人惊叹：4 月 25 日：研究员 Bruce Dang 发现两个已知类别的内核 bug；4 月 27 日：Dion Blazakis 加入，Claude Mythos Preview 开始提供辅助分析；5 月 1 日：Josh Maine 完成工具链，完整 exploit 在实体 M5 芯片设备（macOS 26.4.1，MIE 全开）上成功运行。他们将两个零日级内核 bug、侧信道标签碰撞技巧与数据 -only 提权技术组合，构建出一条高效本地提权链。MIE 防护被突破：苹果硬件内存安全天花板首现缺口MIE（Memory Integrity Extension）是苹果专为 M5/A19 系列芯片打造的硬件级内存安全机制，在 ARM MTE 基础上进一步强化，每 16 字节内存分配 4-bit 标签，由硬件强制校验，性能开销仅约 3%。此前被业界视为“下一代内存安全天花板”，曾让多个知名 exploit kit 失效。Calif 团队此次证明：即使是 MIE，在正确漏洞组合与侧信道技术辅助下仍可被绕过。他们成功将标签碰撞概率从随机 6.25% 压至确定性实现，完成纯数据操作的内核提权。这是全球首次公开披露针对 M5 芯片 MIE 的 macOS 内核内存破坏利用。研究员在报告中表示：“MIE 从来不是为了防住所有黑客而设计的，有正确漏洞就能绕过。”Claude Mythos Preview：AI 成为安全攻防超级加速器Anthropic 推出的 Claude Mythos Preview（安全专版）在此次研究中发挥关键作用。该模型通过 Project Glasswing 计划有限开放给顶级安全团队，能在短时间内发现大量跨系统、浏览器及关键基础设施的零日漏洞，包括多年未被修复的“古董级”bug。Calif 团队强调：“这不是 AI 独立完成，而是顶级人类研究员与前沿 AI 的强强联合。小团队几天就能完成过去几个月甚至几年的工作。”研究完成后，团队已亲自前往 Apple Park 汇报，苹果已在 macOS 26.5 安全更新中修复相关漏洞，并公开致谢 Calif 与 Anthropic Research 的合作。深层启示：AI 驱动“Bugmageddon”时代来临Calif 在报告结尾警示：“这是即将到来的冰山一角。”防御方用 AI 寻找和修复漏洞，攻击方同样用 AI 加速发现与利用，导致漏洞挖掘速度呈指数级提升，传统人工审计模式彻底面临挑战。本次研究被纳入 Calif“Month of AI-Discovered Bugs”系列，标志着 AI 正式成为安全攻防平衡中的核心变量，而非单纯辅助工具。OpenAI 与 Anthropic 在不同技术路线上的探索，也正加速这一趋势。对 Mac 用户的实际影响与建议本次漏洞为本地提权攻击，需要攻击者已获得本地代码执行权限，远程利用风险较低。但苹果已快速推送 macOS 26.5 修复更新，AIbase 强烈建议所有 Mac 用户立即更新系统。Calif 团队提醒：“这只是冰山一角。”未来，更多花费数年打造的顶级防护，可能在 AI 辅助下被快速击穿。AIbase 观点：AI 正在全面重塑网络安全攻防格局。无论是防御还是研究，主动拥抱 AI 工具已成为必然选择。科技巨头与安全社区的协作模式，或将成为未来行业新常态。

Claude Code是什么Claude Code 是 Anthropic 公司推出的基于命令行的AI 编程工具。Claude Code集成先进的 Claude Opus 4 语言模型，能理解自然语言指令，自动生成、修改和调试代码，并与 VS Code、JetBrains 等主流 IDE 深度集成。Claude Code支持 200K 超长上下文，能快速索引整个代码库并理解复杂依赖关系。Claude Code 具备文件操作、代码搜索、网页浏览和 Git 工作流管理等功能，显著提升开发效率。Claude Code帮助开发者快速生成高质量代码，能直接在开发环境中执行命令，避免频繁切换界面，极大地优化开发流程。最新版推出自动安全审查功能，在漏洞发布之前捕获漏洞。Claude Code 新功能 Opus 规划模式支持高效且智能的开发规划，用户能在复杂项目中借助 Opus 4.1 模型的强大分析能力制定高质量的开发计划，在执行阶段切换到 Sonnet 模型，实现高效、低成本的代码生成。Claude Code的主要功能代码库深度理解：基于智能搜索技术，快速理解整个代码库的结构和依赖关系，无需手动选择上下文文件。多文件协同修改：能够跨多个文件进行协调一致的代码修改，确保代码逻辑的完整性和一致性。代码生成与优化：专门针对代码理解和生成进行了优化，能够快速生成高质量的代码片段、测试用例和修复代码中的问题。终端集成：直接在开发者的终端中运行，无需切换工具，减少上下文切换的麻烦。与主流 IDE 集成：支持 VS Code 和 JetBrains 等主流开发工具，直接在 IDE 中提供代码建议和修改，无缝融入现有开发流程。工作流整合：与 GitHub、GitLab 等工具集成，支持从阅读问题、编写代码、运行测试到提交 PR 的全流程操作。智能编辑：基于对代码库和依赖关系的理解，进行强大的多文件编辑，确保代码修改的有效性和可靠性。用户控制与定制：不会在未经明确批准的情况下修改文件，适应开发者的编码标准和模式，并通过 SDK 或 GitHub Actions 进行定制化配置。终端运行的/security-review命令：支持用户在终端输入此命令，Claude Code将自动检查代码中的SQL注入、XSS等安全隐患，能直接修复发现的漏洞。GitHub Actions集成：在GitHub仓库配置相关文件后，每次提交新的PR时，自动触发安全审查，检测出漏洞后在代码行上添加内联注释，解释问题并给出修复建议。Opus 规划模式：在规划阶段调用强大的 Claude 4.1 Opus 模型制定高质量开发计划，在执行任务时使用高效且低成本的 Claude 4 Sonnet 模型，实现智能与效率的结合。Agent View：让开发者在一个终端界面内同时调度、监控和管理多个并行 AI 会话，无需多开窗口可实时查看各任务状态、快速切换介入或解除阻塞。如何使用Claude Code安装 Node.js：确保系统安装 Node.js 18 或更高版本。安装 CLI：在终端运行以下命令安装 Claude Code 的命令行工具：npminstall-g@anthropic/claude-cli登录和初始化：登录：运行claude login并按照提示输入 API 密钥（从 Anthropic 官网获取）。初始化：运行claude init，连接到代码库。开始使用：生成代码：用英语描述需求，例如：claude-p"Create a function to calculate the sum of two numbers in Python"修复错误：描述错误或粘贴错误消息，例如：claude-p"Fix this error: TypeError: Cannot read properties of null (reading ’length’)“自动化任务：运行命令解决繁琐任务，例如：claude-p"Resolve merge conflicts in the current branch"集成到开发环境：终端：直接在终端中使用。IDE：安装 VS Code 或 JetBrains 插件，实现在 IDE 中使用。Claude Code 的产品定价Pro 计划：年付 $17/月，月付 $20/月。Max 5x 计划：$100/人/月。Max 20x 计划：$200/人/月。Claude Code的应用场景代码理解：快速解释复杂代码的功能和逻辑，帮助开发者迅速理解项目结构。代码修改：自动修复代码中的错误或漏洞，快速定位问题并提供解决方案。代码测试：为代码自动生成单元测试或集成测试，确保代码的可靠性和稳定性。版本控制：支持自然语言完成 Git 提交、分支管理等操作，简化版本控制流程，让 Git 操作更加便捷高效。功能实现与扩展：快速实现新功能或扩展现有功能，加速开发进度，帮助开发者高效完成任务。©️版权声明：若无特殊声明，本站所有文章版权均归AI工具集原创和所有，未经许可，任何个人、媒体、网站、团体不得转载、抄袭或以其他方式复制发表本站内容，或在非我站所属的服务器上建立镜像。否则，我站将依法保留追究相关法律责任的权利。类似于Claude Code的工具Project IDX谷歌推出的AI云端开发和代码编辑器C知道CSDN推出的AI技术问答工具Duo ChatGitLab推出的AI编程助手Codeium免费的AI编程工具，智能生成和补全代码aiXcoder自然语言到代码的方法级代码生成，以及多行智能代码补全WindsurfCodeium公司推出的AI编程工具6 条评论再想想发表评论wi游客此ide已放弃 ，国内无法使用，哎9个月前回复AI小集需要很强的魔法，还经常封号9个月前回复wi游客对。弄得很累，不如用亚马逊的ai ide，不用翻墙，也不封号9个月前回复3游客Claude 太强大了，可惜用不来2个月前回复Ai游客需要手机验证，不支持国内无862个月前回复AI小集是的，最好搞一个美国手机号2个月前回复

Kimi Code – Kimi推出的AI编程工具AI工具4个月前发布AI小集05Kimi Code是什么Kimi Code 是 Kimi推出的AI编程工具，可在终端（CLI）、VS Code 、Cursor、JetBrains 和 Zed 等主流编辑器中直接运行，基于Kimi K2.5多模态大模型支持图片和视频输入辅助编程。工具提供端到端代码能力，涵盖构建、调试、重构、测试及脚本编写。Kimi Code能完美兼容Claude Code、Roo Code等第三方 Agent，让 AI 编程能力无缝嵌入用户的每一次 commit。Kimi Code的主要功能全环境覆盖：提供Kimi CLI、VS Code 插件及第三方编辑器支持，可在终端和主流 IDE 中无缝运行编程任务。多模态交互：基于Kimi K2.5模型支持直接输入图片和视频进行编程辅助，实现视觉友好的代码理解与生成功能。端到端编程：完整覆盖代码构建、调试、重构、测试及脚本编写等开发全流程，自动迁移现有技能到新工作流。高频并发：每 5 小时支持 100-500 次请求额度，确保复杂项目开发过程中 AI 辅助的不间断可用性。第三方兼容：完美适配 Claude Code、Roo Code 等主流 Coding Agent，一键登录可完成设备授权与账号绑定。如何使用Kimi Code订阅会员：访问 Kimi Code 官网 https://www.kimi.com/code。安装和登录：CLI 用户：执行curl -L code.kimi.com/install.sh | bash安装，运行/login命令一键授权。VS Code 用户：安装 Kimi Code 插件，在终端直接运行登录命令完成设备绑定。配置使用官方 Agent：登录成功后即可直接使用，系统自动完成账户绑定。第三方工具（Claude Code / Roo Code）：在控制台创建 API Key，复制后配置到对应工具的环境变量或设置中即可开始编程。Kimi Code的产品定价Andante 计划：定价 ¥4.99（7 天试用），试用期结束后以 ¥49/月自动续费，提供专属 Kimi Code 使用额度、旗舰模型抢先体验以及支持多个编程会话的基础功能。Moderato 计划：定价 ¥99/月，提供每周更新的使用额度，多种设备登录并分享套餐额度，支持用户在多个项目中高效协作。Allegretto 计划：定价 ¥199/月，提供充足的每周额度和更高的并发上限，为高级用户提供更强大的使用体验和超值选择。Kimi Code的应用场景全栈开发任务：覆盖代码构建、调试、重构、测试及脚本编写等端到端编程工作流，满足从原型开发到生产部署的完整需求。多模态辅助编程：支持直接输入 UI 设计图、视频演示等多媒体内容，基于视觉信息生成对应代码实现或进行问题诊断。跨平台开发环境：可在终端 CLI、VS Code、Cursor、JetBrains、Zed 等主流编辑器中无缝切换，适配不同开发者的工具链偏好。高频敏捷开发：借助 Turbo 级加速（100 Tokens/s）和每 5 小时数百次的并发额度，支撑大型项目中的高频代码生成与迭代需求。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇inSai Hilight - AI营销视频Agent平台，多Agent深度协同下一篇Lucy 2 - Decart AI推出的实时视频生成模型相关文章Project Turntable – Adobe推出可旋转操作2D矢量图像的AI工具AI小集6Nano Banana – AI图像编辑模型，具有极高的人物一致性AI小集15Floot – 无代码AI Web应用开发平台，提供全栈集成AI小集4GLM-Image – 智谱联合华为开源的多模态图像生成模型AI小集3拍试卷 – 专为学生和教师设计的试卷扫描和处理软件AI小集4CreateVision AI – AI图像生成平台，支持样式、颜色、光照精准控制AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付6分钟前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型56分钟前Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手59分钟前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台5小时前PPT Master – 开源 AI PPT 生成标准化工作流6小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架6小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具22小时前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前

Kimi WebBridge – 月之暗面推出的浏览器扩展插件AI工具3天前发布AI小集03Kimi WebBridge是什么Kimi WebBridge 是月之暗面推出的浏览器扩展插件，面向Kimi Code、Claude Code、Cursor、Codex等本地 AI Agent。Kimi WebBridge 通过本地桥接服务与 Chrome DevTools Protocol，让 AI 携带用户真实登录态和 Cookie，像真人一样在浏览器中执行点击、滑动、输入、填表、信息提取及跨站点内容整合等操作。所有任务在本地完成，网页数据不会离开设备，保障隐私安全，同时实现高效、自动化的网页操作。Kimi WebBridge的主要功能浏览器自动化操作：支持 AI Agent 执行网页导航、点击按钮、填写表单、滑动页面等操作。身份状态继承：AI 携带用户真实登录态、Cookie 和账号权限操作网站，无需额外配置。信息提取与整合：跨站点抓取内容并整合，支持将数据写入在线文档或本地目录。可视化任务追踪：运行时标记正在操作的浏览器标签页，不占用鼠标键盘，用户可并行使用电脑。进阶 CLI 工具：支持将固定流程打包为不消耗大模型 Token 的专用 CLI 工具，用于重复任务。Kimi WebBridge的技术原理双组件协同架构：系统由本地桥接服务与浏览器扩展两部分组成。AI Agent 将自然语言指令发送至本地桥接服务，后者通过标准接口与 Chrome 或 Edge 浏览器内的扩展通信，扩展再基于底层协议驱动浏览器完成具体操作，并将执行结果回传给 Agent，形成完整的指令闭环。基于 Chrome DevTools Protocol：浏览器扩展底层采用 Chrome DevTools Protocol（CDP），是 Chrome 官方提供的调试协议。通过协议，扩展能在用户当前使用的真实浏览器实例中执行导航、元素点击、表单输入、页面截图、DOM 读取与修改等操作，无需启动独立或虚拟浏览器环境。本地身份继承机制：Kimi WebBridge 直接复用用户已打开的浏览器会话，继承真实的登录态、Cookie 和本地存储数据。AI 在操作目标网站时，等同于用户本人在操作，因此无需额外配置账号、处理验证码或突破登录壁垒，天然支持各类需要身份认证的后台系统。端侧隐私保障：所有指令解析、页面交互与数据提取均在用户本地设备完成。桥接服务只负责本地进程间通信，网页内容、Cookie 及操作结果不会上传至云端服务器，确保敏感业务数据不出本地设备。如何使用Kimi WebBridge安装浏览器插件：在 Chrome 或 Edge 浏览器应用商店搜索并安装Kimi WebBridge扩展插件。安装本地守护程序：向本地 AI Agent 发送指定 curl 命令，安装 Kimi WebBridge 的本地守护程序与 Skill。重启 Agent 加载技能：重启 Kimi Code 等本地 AI Agent，使其识别并加载新安装的 WebBridge 技能。验证安装结果：输入测试指令「使用 kimi-webbridge 帮我打开 kimi.com」，若成功打开标签页并出现 agent:kimi 提示即代表安装成功。Kimi Claw Desktop 用户：若通过 Kimi 桌面版在本地部署了 Kimi Claw Desktop，无需安装守护程序，直接发送指令可开始使用。Kimi WebBridge的核心优势原生身份继承：直接复用用户现有登录态，无需模拟账号或处理验证码，降低复杂网站接入门槛。本地隐私安全：全部操作在本地浏览器执行，敏感数据与网页内容不上传云端。多 Agent 兼容：统一支持 Kimi Code、Claude Code、Cursor、Codex、Hermes、OpenClaw 等主流本地 Agent。非侵入式运行：后台标记操作标签页，不抢占鼠标键盘，人机可协同工作。低 Token 消耗：固定流程可编译为专用 CLI 工具，重复执行时无需反复调用大模型。Kimi WebBridge的项目地址项目官网：https://www.kimi.com/zh-cn/features/webbridgeKimi WebBridge的同类竞品对比维度Kimi WebBridgeBrowserbaseAgent-Browser（Vercel）运行方式本地浏览器扩展 + 桥接服务云端托管浏览器云端/本地浏览器自动化身份继承直接复用用户 Cookie 和登录态需单独配置环境或账号需单独配置环境或账号数据隐私完全本地执行，数据不出设备数据经过云端服务器取决于部署方式支持 AgentKimi Code、Claude Code、Cursor 等主要面向开发者 API主要面向开发者 API使用门槛安装插件 + 一行命令即可需 API 集成开发需一定技术配置人机协同后台运行，不占用键鼠云端运行，无本地交互云端运行，无本地交互Kimi WebBridge的应用场景自动化信息整理：自动采集多个笔记类 APP 的应用商店宣传素材，让 AI 分析设计风格差异并写入在线文档。复杂操作自动化：一键接管量化研究模拟平台，自动抓取金融数据、执行多策略迭代回测并撰写量化报告。重复性网页任务：自动化完成基于网页的重复性繁琐操作或定时任务，如批量填表、点击、数据录入与状态更新。跨站点内容整合：在已登录状态下跨多个网站提取信息并整合，实现内容迁移、数据汇总与多源信息同步。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Mavis - MiniMax Agent 推出的多 Agent 协作模式下一篇TencentDB Agent Memory - 腾讯开源的智能体记忆管理工具相关文章OWL – 开源多智能体协作系统，动态交互实现任务自动化AI小集4DAM-3B – 英伟达推出的多模态大语言模型AI小集3VideoReward – 港中文、清华、快手等联合推出的视频生成偏好数据集及奖励模型AI小集2混元图像2.0 – 腾讯推出的实时AI图片生成大模型AI小集3J1 Assistant – 罗永浩旗下 Jarvis 项目推出的 AI 智能助手AI小集3Pollo AI – 一站式AI图像和视频创作平台AI小集6暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付6分钟前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型56分钟前Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手59分钟前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台5小时前PPT Master – 开源 AI PPT 生成标准化工作流6小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架6小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具22小时前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前

了解更多 kimi.com/zh-cn/features/webbridge

Kimi WebBridgeÔºökimi.com/zh-cn/features/webbridge

Kimi K2.6 发布并开源，全面精进代码和 Agent 集群能力

Kimi K2.6「Agent 集群」现已支持 300 个成员并行协作

和 Kimi 一起投身 AGI，穿越成长周期

服务全球1/5网站的Cloudflare选择Kimi K2.5，降低77%成本

Kimi 杨植麟「2026 中关村论坛」演讲全文（附视频）

Kimi API：用90%缓存命中率，把价格打到25%

你可以来 Kimi 使用 OpenClaw 了

技术报告：Kimi K2.5 如何实现文本和视觉能力互相增强？

不只PPT，Kimi K2.5 Agent可以帮你做Excel、Word和PDF了

Kimi 发布并开源 K2.5 模型，带来全新视觉理解、代码和 Agent 集群能力

微信扫一扫关注该公众号

微信扫一扫可打开此内容，使用完整服务

吐司APP的简介吐司是腾讯推出的AI原生应用创作平台，用户只需输入想法，AI自动生成完整的安卓原生应用。平台支持智能功能梳理、精细化编辑打磨，能一键导出APK进行本地安装与分享。内置灵感广场汇聚海量优质模板，涵盖社交助手、照片精选、学习工具等多元场景，支持用户共创。无需编程基础，零基础用户也能快速将奇思妙想转化为可实际使用的专属APP，轻松实现个性化应用开发。吐司APP的主要功能AI自动生成应用：输入自然语言想法，AI自动梳理功能逻辑并生成完整的安卓原生应用，无需编写代码。精细化编辑打磨：支持对生成的应用进行细节调整与优化，满足个性化需求。一键导出APK：生成完成后可直接导出APK安装包，支持本地安装、测试与社交分享。灵感广场：内置海量精选应用模板与创意案例，如社交助手、照片精选、学习工具等，帮助用户快速获取创作灵感。应用共创：支持用户基于现有模板进行二次创作与共创，降低应用开发门槛。多场景覆盖：支持生成考研题库、运势工具、普通话测试、宠物健康日志、梦境记录、食材新鲜度检测、万年历等多样化实用应用。类似于吐司APP的应用Replit APPAI开发工具应用心响APP百度推出的手机端超级智能体应用Qoder阿里巴巴推出的 AI Agent 编程平台WindsurfCodeium公司推出的AI编程工具Kiro亚马逊公司推出的 AI IDECursorAnysphere公司推出的AI代码编辑器暂无评论再想想发表评论暂无评论…

LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台AI工具5小时前发布AI小集03LibTV团队版是什么LibTV 团队版是LibTV推出的专业级 AI 视频协作工作台，专为 AI 短剧、电影、TVC 制作团队设计。平台提供团队空间、实时画布协作、资产库沉淀、权限管理、成本管控等功能，解决 AI 视频创作中资产散落、协作低效、人员流动导致资产流失等痛点，支持规模化、工业化的 AI 内容生产。LibTV团队版的主要功能专属团队空间：统一沉淀分镜、脚本、镜头资产与生成记录，历史项目可迁移，避免核心资产散落在个人账号与本地文件中。画布协作：导演、美术、剪辑等角色可在同一块画布实时协同，同步节点状态与镜头修改，减少版本错乱。团队资产库：角色主体、风格模板、运镜方案、提示词工作流与参考素材支持统一沉淀，成熟视觉体系可持续复用。权限管理：支持项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作场景。成本管理：共享积分池、成员额度分配与消耗查看，团队管理者可统一管理生成资源、席位与项目成本。资产交接：成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免创作资产随人员流动中断。如何使用LibTV团队版创建团队空间：访问LibTV官网升级至团队版，建立专属团队项目空间，设置团队名称与基础信息。配置成员与权限：邀请导演、编剧、美术、剪辑等成员加入，按岗位分配访问、编辑、复制或发布权限。沉淀与迁移资产：将历史个人项目迁移至团队空间，或新建项目并导入参考素材、角色设定与风格模板。画布协作创作：成员在统一画布内协同推进分镜设计、镜头生成与修改，实时同步进度与反馈。复用团队资产库：调用已沉淀的角色主体、风格模板与提示词工作流，保持系列化内容风格一致。管控成本与交接：通过共享积分池分配生成额度，监控消耗；成员变动时资产自动保留，确保项目连续性。LibTV团队版的核心优势工业化协作：从个人创作升级为团队制片管理，支持长期项目、多岗位协同与规模化内容生产。资产持续复用：成熟的角色设定、风格与提示词工作流可沉淀为团队资产，降低新人上手与重复调试成本。实时协同提效：多人画布协作减少版本错乱与重复生成，围绕同一套镜头语言与素材体系推进项目。成本与风险可控：共享积分池与席位管理适配多项目并行需求，权限分级与资产交接机制保障核心创作安全。LibTV团队版的同类竞品对比对比维度LibTV 团队版快手可灵（企业版/团队功能）协作模式多人实时画布协作，导演/美术/剪辑等角色可在同一块画布内同步节点状态、生成进度与镜头修改团队空间内创作接力，支持桌面端（Win/Mac）同步，减少文件下载与搬运损耗资产沉淀角色主体、风格模板、运镜方案、提示词工作流、参考素材与工具箱统一沉淀，支持成熟视觉体系持续复用设立「共享资产库」，支持个人空间向团队空间一键转移灵感值、创作资产与主体资产，实现”一人贡献，全员复用”权限管理项目访问、编辑、复制、发布等多级权限，适配工作室内部及品牌方、代理公司等外部协作所有者、管理员、协作者三重角色，灵活配置查看/编辑/移除权限，保障核心资产安全成本管理共享积分池、成员额度分配与积分消耗查看，支持多项目并行下的生成资源统筹支持灵感值跨空间转移，团队共享生成资源，但精细化成本管控功能相对较弱资产交接成员退出后，相关项目、素材、角色体系与工作流自动保留在团队空间，避免资产随人员流动中断个人资产可一键导入团队空间，但成员退出后的自动资产交接机制未明确强调核心场景聚焦 AI 短剧、AI 电影、AI TVC 的工业化制片流程，强调长期项目管理与跨岗位协同覆盖 AI 视频生成、智能分镜、广告 Demo、影视特效等，强调快速创作与素材复用产品定位AI 视频专业级协作工作台，面向制片管理流程AI 视频生成平台的团队功能扩展，面向创作者联盟与工作室LibTV团队版的应用场景AI 视频工作室：适合长期运营 AI 短剧、动画、MV 的创作团队，通过资产库统一沉淀角色与风格，实现系列化生产。短剧承制方与 MCN：适配高频、多项目并行的工业化内容生产，导演与美术可同时在线推进分镜与镜头生成。品牌市场部与代理公司：支持品牌方、代理公司与外部创意团队跨组织协作，共同管理 AI TVC 与 Campaign 项目。AI 电影与大型制作团队：支持长篇叙事、系列化 IP 与复杂制作流程，导演组、制片组、美术组与后期团队跨岗位协同。企业内容与创意部门：适配培训、营销、产品与全球化传播等标准化内容生产，便于企业进行长期生成资源规划。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇PPT Master - 开源 AI PPT 生成标准化工作流下一篇Marvis - 腾讯应用宝推出的操作系统层个人 AI 助手相关文章XMusic – 腾讯多媒体实验室自主研发的AI通用作曲框架AI小集2VoxCPM – 面壁智能联合清华推出的语音生成模型AI小集4JoyGen – 京东和港大推出音频驱动的3D说话人脸视频生成框架AI小集2PrimitiveAnything – 腾讯联合清华大学推出的新型3D形状生成框架AI小集2Nano Banana – AI图像编辑模型，具有极高的人物一致性AI小集15MiniCPM-V 4.5 – 面壁智能开源的端侧多模态模型AI小集3暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付5分钟前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型56分钟前Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手59分钟前PPT Master – 开源 AI PPT 生成标准化工作流6小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架6小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具22小时前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前

Grok Build – xAI 推出的终端原生 AI 编程智能体AI工具1天前发布AI小集03Grok Build是什么Grok Build 是 xAI 推出的终端原生 AI 编程智能体（Agentic CLI），面向专业软件工程与复杂编码任务。基于 Grok 4.3 beta 的 200 万 token 超长上下文，支持 Plan Mode 计划审批、最多 8 个并行子智能体协作及 Arena Mode 自动方案排序，可在本地直接读写文件、执行 Shell 命令、操作 Git 并构建完整应用。采用本地优先架构，源代码与敏感数据不上传云端，保障企业级安全。Grok Build的主要功能Plan Mode（计划模式）：复杂任务执行前自动生成结构化步骤计划，用户可审批、评论或修改，所有代码变更以 diff 形式透明展示，避免”黑箱”操作。并行子智能体：最多支持8 个并发 AI 智能体同时工作，可分别处理代码分析、CI/CD 检查、性能优化、文档检索等不同子任务。Arena Mode（竞技场模式）：多个智能体针对同一任务生成不同代码方案，系统自动评估并排序，开发者直接选择最优解，无需手动对比。超长上下文窗口：基于 Grok 4.3 beta，支持200 万 token上下文，可一次性加载整个大型代码库进行跨文件复杂任务。本地优先架构：源代码、凭证和敏感数据不上传云端，完全在本地执行，支持离线环境和气隙网络，保障企业级代码安全。终端原生 CLI + Web 双入口：支持本地命令行交互，也可通过浏览器界面远程操作，满足不同工作流习惯。VS Code 扩展集成：提供官方 VS Code 插件，在 IDE 内直接调用 Grok Build 能力，实现 GUI adjacent 工作流。Git 与开发服务器集成：可自动管理 Git 仓库、启动本地开发服务器、内置浏览器浏览网页、执行 Shell 命令。插件与开放生态：支持 MCP 服务器、AGENTS.md 团队规范文件、Hooks 自定义动作、ACP（Agent Client Protocol）协议，便于集成现有 DevOps 流水线。Headless 脚本模式：通过-p标志支持非交互式执行，可直接嵌入 CI/CD 流水线或自动化脚本。BYOK（自带 API Key）：支持使用自有 xAI API Key 进行本地计费，灵活控制成本。如何使用Grok Build一键安装 CLI：在终端执行curl -fsSL https://x.ai/cli/install.sh | bash完成本地安装，支持 macOS、Linux 和 Windows 三大平台。启动交互式终端：安装完成后输入grok>进入交互式命令行界面，可直接用自然语言描述编程需求。登录账号：在终端内执行grok> /login登录 xAI 账号，解锁高级功能与订阅权益。配置 API Key（可选）：如需使用自有额度，执行grok> /api-key add xai添加个人 xAI API Key（BYOK 模式）。使用 Plan Mode 处理复杂任务：输入需求后，Grok Build 会自动生成结构化执行计划，展示步骤与预期变更 diff，用户可审批、评论或修改后再执行。利用并行子智能体：在复杂项目中，系统会自动或手动触发最多 8 个并发子智能体，分别处理代码分析、测试生成、文档检索等子任务。通过 Arena Mode 选优：对同一任务让多个智能体生成不同方案，系统自动评估排序，你只需选择排名最高的方案应用。Web 端远程操作：访问build.grok.com或x.ai/cli，在浏览器中打开 Grok Build 界面，无需本地安装即可使用。VS Code 集成工作流：安装 Grok Build 的 VS Code 扩展，在编辑器内直接调用智能体能力，边写代码边让 AI 协助修改。Headless 自动化执行：在脚本或 CI/CD 流水线中使用-p标志运行非交互式命令，实现自动化构建与部署。查看帮助与命令：在终端输入grok> /help查看所有内置命令与快捷操作指南。Grok Build的官网地址项目官网：https://x.ai/news/grok-build-cliGrok Build的核心优势本地优先，代码绝不上云：源代码、凭证及敏感数据完全在本地执行，无需上传至 xAI 服务器，适合处理专有代码库、金融及政务等受监管行业项目。200 万 token 超长上下文：基于 Grok 4.3 beta，可一次性将整个大型代码库载入内存，跨数十个文件进行复杂重构与多文件联动修改，大幅减少上下文丢失。8 路并行子智能体：最多同时调度 8 个并发 AI 智能体分工协作，分别处理代码分析、测试生成、CI/CD 检查等子任务，显著缩短复杂项目耗时。Arena Mode 自动选优：同一任务由多个智能体生成不同方案，系统自动评估排序并给出排名，开发者直接选择最优解，告别手动对比试错。Plan Mode 透明可控：执行前先输出结构化步骤计划与 diff 预览，用户可审批、评论或修改后再执行，每一步变更清晰可见，避免”黑箱”风险。极致 API 价格竞争力：输入 token 仅$0.20 / 百万 tokens，相较 Claude Code 和 Codex CLI 大幅降低高频使用成本。终端原生 + 多入口覆盖：支持 CLI 命令行、Web 浏览器、VS Code 扩展三种工作入口，既满足极客式终端操作，也兼容 IDE 集成习惯。Grok Build的同类竞品对比对比维度Grok Build (xAI)Claude Code(Anthropic)Codex CLI(OpenAI)产品定位终端原生 AI 编程智能体（Agentic CLI）终端 AI 编程助手云端 AI 编程智能体发布状态2026.5 早期 Beta已正式发布已正式发布底层模型Grok 4.3 beta / grok-code-fast-1Claude 4 (Sonnet/Opus)Codex (基于 GPT-4o/o3)上下文长度200 万 token20 万 token (Claude 4)约 20 万 token并行智能体最多 8 个并发子智能体单智能体为主单智能体为主方案评估Arena Mode 自动排序选优手动对比或迭代手动对比或迭代执行计划Plan Mode（先计划后审批执行）直接执行，可审查 diff直接执行，可审查 diff数据安全本地优先，代码不上云云端执行，代码上传 Anthropic云端执行，代码上传 OpenAI离线支持支持完全离线/气隙网络需联网需联网SWE-Bench70.8%约 72–75%约 70%API 定价$0.20 / 百万 tokensPro 订阅 + 按量计费Credits 按量计费订阅门槛SuperGrok Heavy ($300/月)Claude Pro / Team 订阅ChatGPT Pro / 开发者额度入口形态CLI + Web + VS Code 扩展CLI + IDE 集成CLI + ChatGPT 移动端自动化集成Headless 模式 (-p 标志) + CI/CD有限脚本支持有限脚本支持核心差异多智能体并行 + 本地优先 + 超大上下文模型推理能力强，生态成熟与 ChatGPT 生态深度整合Grok Build的应用场景大型代码库重构与现代化：基于 200 万 token 超长上下文一次性加载整个项目，跨数十个文件进行安全重构、依赖升级或架构迁移，避免上下文截断导致的遗漏。复杂全栈应用从零构建：通过 Plan Mode 自动生成项目结构、技术选型、依赖安装和文件生成计划，审批后直接执行，快速搭建完整应用原型。敏感行业专有代码开发：本地优先架构确保源代码、凭证和敏感数据不上传云端，适用于金融、政务、医疗、军工等对数据主权有严格要求的场景。离线/气隙网络环境编程：无需持续联网即可在本地执行全部任务，适配高安全隔离网络、内网开发机或无外网访问权限的工作环境。多维度自动化代码审查：启动多个并行子智能体分别检查代码安全性、性能瓶颈、CI/CD 合规性和文档完整性，一次性输出综合审查报告。技术方案快速对比与选型：使用 Arena Mode 针对同一需求生成多种实现方案，系统自动评估排序，帮助团队在短时间内确定最优技术路径。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇OpenHuman - 开源桌面端 AI 助手，能主动感知工作上下文下一篇Violin - 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具相关文章极客侧边栏 – AI浏览器扩展工具，集成AI助手和书签管理AI小集4心动AI简历 – AI简历优化工具，智能分析岗位描述进行针对性优化AI小集2MobileVD – 首个针对移动设备优化的视频扩散模型AI小集2EverMemOS – 陈天桥团队开源的AI长期记忆操作系统AI小集3Ancher – AI信息管理工具，智能筛选个性化信息推荐AI小集2Fineshare VoiceTrans – 实时AI语音转换器，实时声音调节保留口音、语调、情感AI小集2暂无评论再想想发表评论暂无评论…

OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文AI工具2天前发布AI小集05OpenHuman是什么OpenHuman 是 tinyhumansai 团队推出的开源个人 AI 超级智能助手，定位为”Your Personal AI super intelligence”，主打私有、简单且极其强大。非传统聊天机器人，是能主动感知用户工作上下文的桌面级 AI Agent，通过每 20 分钟自动同步 118+ 第三方应用数据，构建本地持久记忆，让 AI 真正”懂你”。项目采用 Rust + TypeScript + Tauri 构建，支持 macOS、Linux 和 Windows 平台。OpenHuman的主要功能桌面吉祥物交互：具备视觉形象的桌面 AI 助手，支持语音输入（STT）与 ElevenLabs 语音输出（TTS），吉祥物可实时唇形同步，能作为参会者加入 Google Meet 并实时记录会议。118+ 第三方应用一键集成：通过 OAuth 快速接入 Gmail、Notion、GitHub、Slack、Linear、Jira、Stripe、Calendar、Drive 等主流工具，无需编写任何连接代码。每 20 分钟自动同步（Auto-fetch）：核心引擎每 20 分钟遍历所有活跃连接，自动拉取最新数据进入本地记忆系统，让 Agent 在次日早上已掌握前日全部上下文。Memory Tree 层次化持久记忆：将同步数据规范化为 ≤3k-token 的 Markdown chunks，评分后折叠存入层级摘要树，存储于本地 SQLite，实现跨会话的长期记忆。Obsidian Vault 原生兼容：同步数据同时输出为 Obsidian 可读的.md文件，用户可在 Obsidian Graph View 中浏览和编辑 AI 维护的知识网络。TokenJuice 智能 Token 压缩：在数据进入 LLM 前自动压缩（HTML 转 Markdown、缩短长 URL、移除非 ASCII 字符等），最高可降低 80% 的 API 调用成本与延迟。内置模型路由（Model Routing）：根据任务类型自动分配轻量模型、强推理模型或视觉模型，一个订阅覆盖多种模型能力，无需管理多个 API Key。完整工具链开箱即用：内置 Web 搜索、网页抓取（Web Scraper）、代码工具集（文件系统、Git、Lint、Test、Grep），无需额外安装插件即可使用。本地优先与隐私保护：所有工作流数据本地加密存储，支持通过 Ollama 接入本地模型实现完全离线运行，数据不上传云端。可选 agentmemory 共享后端：可配置与 Claude Code、Cursor、Codex 等工具共享同一个持久记忆存储，实现跨工具的知识互通。消息通道集成：支持 inbound/outbound 消息收发，打通用户已有的沟通渠道，工作流数据始终保存在本地设备。如何使用OpenHuman官网下载安装：访问OpenHuman的官网 https://tinyhumans.ai/openhuman 下载对应平台的 DMG（macOS）或 EXE（Windows）安装包，双击安装后即可启动。命令行快速安装（macOS / Linux x64）：在终端执行curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash自动完成下载与安装。命令行快速安装（Windows）：在 PowerShell 执行irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex完成一键安装。一键连接工作应用：打开应用后，通过 OAuth 一键接入 Gmail、Notion、GitHub、Slack、Calendar、Drive 等 118+ 第三方工具，授权后每 20 分钟自动同步数据。与桌面吉祥物交互：点击桌面上的吉祥物形象，通过语音或文字与其对话，可勾选 “Speak replies” 让 AI 语音播报回复。查看自动同步的记忆数据：在应用内查看 Agent 已同步的邮件、任务、代码仓库等上下文，AI 会自动压缩整理为 Memory Tree 供后续对话调用。Obsidian 知识库浏览：同步数据自动写入本地 Obsidian-compatible Vault，打开 Obsidian 即可在 Graph View 中可视化浏览 AI 维护的知识网络，也可手动编辑.md文件。使用内置工具集：直接调用开箱即用的 Web 搜索、网页抓取、代码工具（文件系统、Git、Lint、Test、Grep）等能力，无需额外安装插件。接入本地模型（可选）：如需完全离线运行，通过 Ollama 在本地部署模型，并在设置中切换至本地模型后端，实现零云端依赖。OpenHuman的项目地址官网地址：https://tinyhumans.ai/openhumanGitHub 仓库：https://github.com/tinyhumansai/openhumanOpenHuman的核心优势UI 优先分钟级上手：干净的桌面体验，无需终端配置，从安装到可用 Agent 仅需几分钟，区别于终端优先的竞品。上下文在分钟级而非周级：传统 Agent 需要数周学习用户习惯，OpenHuman 通过一次同步即可掌握邮箱、日历、仓库、文档、消息的全量压缩上下文。一站式订阅，零 API 蔓延：一个订阅覆盖所有功能，无需额外购买多个模型 API Key。Obsidian 知识库原生兼容：同步数据自动写入 Obsidian-compatible Vault，用户可在熟悉笔记界面中浏览和编辑 AI 维护的知识图谱。Rust + Tauri 高性能架构：内存占用远低于 Electron 方案，69.7% Rust + 26.1% TypeScript 构建。开源可审计：GPL-3.0 协议，代码完全透明，隐私敏感用户可自行审计数据流向。OpenHuman的同类竞品对比对比维度OpenHumanOpenClawHermes Agent开源协议✅ GPL-3.0✅ MIT✅ MIT上手难度✅ UI 优先，分钟级上手⚠️ 终端优先配置⚠️ 终端优先配置使用成本✅ 单一订阅 + TokenJuice 压缩⚠️ 自备模型 API⚠️ 自备模型 API记忆能力🚀 Memory Tree + Obsidian Vault + 可选 agentmemory 后端⚠️ 依赖插件扩展✅ 自学习观察第三方集成🚀 118+ 应用 OAuth 一键接入⚠️ 需自行配置连接⚠️ 需自行配置连接自动数据同步✅ 每 20 分钟自动 fetch 至本地记忆🚫 无自动同步🚫 无自动同步API 管理✅ 一个账户覆盖所有模型🚫 自备 Key🚫 多供应商管理模型路由✅ 内置自动分配（轻量/推理/视觉）⚠️ 手动切换⚠️ 手动切换原生工具集✅ 代码 + 搜索 + 抓取 + 语音✅ 代码工具✅ 代码工具数据隐私🚀 本地优先，数据不上云依赖自托管配置依赖自托管配置OpenHuman的应用场景跨应用项目上下文整合：同时连接 GitHub、Linear、Jira、Slack 后，让 AI 自动对比代码 PR、任务看板和团队消息，一键汇总项目最新进展，无需手动切换多个工具。智能邮件筛选与摘要：授权 Gmail 后，AI 每 20 分钟自动扫描收件箱，按项目或优先级分类摘要，快速定位需要回复的重要邮件，过滤掉低价值信息。AI 会议助手：让桌面吉祥物以真实参会者身份加入 Google Meet，实时记录会议内容、提取 action items，并基于记忆库提供参会者背景信息辅助沟通。代码仓库智能问答：同步 GitHub 仓库数据后，直接向 AI 询问代码逻辑、历史变更原因、PR 评论争议点，Agent 基于完整代码上下文给出精准回答。Obsidian 个人知识库构建：将邮件、文档、聊天记录自动压缩归档为 Obsidian-compatible 的.md文件，在 Graph View 中可视化呈现知识关联网络，打造 Karpathy 风格的 LLM 知识库。多工具工作流自动化：连接 Notion、Calendar、Drive、Stripe 等工具后，AI 自动关联”日历事件→相关文档→对应邮件→财务记录”，形成完整的项目上下文链。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Intern-S2-Preview - 上海 AI Lab 开源的科学多模态大模型下一篇Grok Build - xAI 推出的终端原生 AI 编程智能体相关文章PodLM – AI播客生成器，一键将URL和文本转换为播客AI小集3灵图AI – AI辅助设计平台，专为美术设计打造快速生成手绘效果图AI小集4GPT‑5.3‑Codex‑Spark – OpenAI推出的轻量级编程模型AI小集3ImagePulse – 魔搭社区开源的图像理解和生成模型数据集AI小集2OpenHands – AI编程工具，多智能体协作实现代码编写、命令运行等AI小集4Composer – Cursor推出的自研编码模型AI小集4暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录堆友Agent灵芽社区方舟 Agent PlanMavis万小智ToDesk AI最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付4分钟前Higgs Avatar v1 – 面向语音智能体的实时 AI 数字人模型54分钟前Marvis – 腾讯应用宝推出的操作系统层个人 AI 助手57分钟前LibTV团队版 – LibTV 推出的专业级 AI 视频协作工作台5小时前PPT Master – 开源 AI PPT 生成标准化工作流6小时前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架6小时前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具22小时前Grok Build – xAI 推出的终端原生 AI 编程智能体1天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前

Scott GilbertsonGearMay 16, 2026 1:00 AMTop Surfshark Promo Codes for May 2026Save up to 87% with a Surfshark coupon code, 3 months of VPN free today, and more from WIRED.Save StorySave this storySave StorySave this storySurfshark is oneofour favorite VPNs to get around geographical restrictions on content (aka access Netflix) and protect your web traffic while using an open Wi-Fi hotspot. Surfshark is secure, and it provides great value for the money if you pay for two years upfront. We also like that Surfshark’s plans are simple: A single Surfshark subscription works simultaneously on multiple devices and protects your entire household. No multiple accounts, no confusion. Get award-winning security that’s even more budget-friendly with our roundup of top Surfshark coupons, plans from $3 per month, and 3 free months on 1 or 2 year subscriptions.Save on 1 and 2 Year Plans With Surfshark CouponsRight now, you can get a 7 day free trial, an extra 3 months for free on all 1-year and 2-year plans, and up to 87% off with active Surfshark coupon codes. The Surfshark Starter Plan is still thecheapestplan—now it starts at $2 per month on a 2-year plan compared to $20 per month if you were paying month to month. The most comprehensive plan is Surfshark One+, which is now 81% off, at $4 per month. Even the basic Surfshark Starter plan includes unlimited devices, ad and cookie pop-up blockers, double encryption, and alternative IDs.Surfshark also offers Surfshark One Plus for $6 per month, with 3 months extra. Surfshark One adds some extra features like Real-time credit card andID breach alerts, personal data security reports, andanti-virusand anti-malware software. One Plus gets you all that and adds data removal services. You can see the full breakdown of services and a comparison of the plans on theSurfshark website.In my testing over the years, Surfshark has consistently had some of the best speeds of any VPN I’ve used. I have never had any problem streaming HD content through Surfshark. It’s fast enough that you won’t notice any speed degradation.Score a 7-Day Free Trial on Surfshark and 30 Day GuaranteeIf you’re feeling commitment-shy, fear not, there is a 7-day free trial on all plans, plus a 30-day refund period and if it turns out Surfshark isn’t right for you. Just contact Surfshark and you’ll receive a full refund within 30 days with their money-back guarantee. With your Surfsharkfree trial, you can do things like change your IP with 3,200+ encrypted servers, encrypt your connection to stop data tracking, and more.Students Can Get a 15% Off Surfshark Discount CodeStudents can get an even better deal on Surfshark, adding anadditional 15% discounton VPN and Antivirus plans on top of the coupons above. To receive your discount code, first verify your student status through StudentBeans on Surfshark’s website. Graduate students can get thegraduate discountof 15% off the already great price of $1.99 per month.Get Paid $38 or 3 Months of SurfShark Free When You Refer a FriendWhen youinvite friendsto use Surfshark, you can score cash rewards. With therefer a friend program, you can now choose between getting up to 3 free months of Surfshark, or paying the equivalent amount via PayPal (up to $38). If your invited friend purchases 1 month of Surfshark, you’ll both get 1 month free. If your invited friend purchases a longer Surfshark plan, you’ll both get 3 months free. To redeem, you’ll need to share the unique referral link that’ll be in your Surfshark app with as many friends as you want. When each friend signs up and keeps their account active for 31 days, you’ll be able to choose whether you want money in your account or free months added to your Surfshark subscription.Premium Security for Less: Surfshark AntivirusSurfshark knows how important it is to maintain your privacy and protect your devices. That’s whySurfshark Antivirusis here to guard all your devices without interruption. The Surfshark antivirus software supports up to 5 devices, provides regular updates, and has tons of advanced security features that protect you (and your devices) from a myriad of online threats. Plus, you can lock your camera, and get alerted if a website tries to access your webcam without your permission. The Antivirus app is also customizable, depending on how you want to use it. You can exclude files, and schedule full or quick scans.Is Surfshark Legal to Use in the US?Yes, in ‘Merica, land of the free, home of the brave, e.g. you can use Surfshark legallyin the US! You can freely use a VPN in the USA without being restricted, whether that be from your provider or area. Because it’s legal throughout the US, it just may be the best value for service coverage.Does Netflix block Surfshark?Luckily, no! You can have both active subscriptions forSurfshark and Netflixand still view Netflix streaming content securely without interruption or issue.Can I Have Multiple Devices on one Surshark Account?When you download the Surfshark VPN ap

Molly HigginsGearMay 16, 2026 1:00 AMSave on Outdoor Gear With a Sportsman’s Warehouse Promo Code in May 2026Whether you are hunting for firearms, camping supplies, or boating gear, use these Sportsman’s Warehouse coupons to maximize your savings this month.Save StorySave this storySave StorySave this storyIf you’re lookingfor a one-stop shop for all things outdoors, Sportsman’s Warehouse is worth plotting on your map. Whether you need essentials for a big camping trip this summer or just need to restock your tired boat day accessories, Sportsman’s Warehouse can outfit you. With multiple brick-and-mortar locations across the country, as well as an expansive online inventory with favorite brands like Crocs and Yeti, Sportsman’s Warehouse hasapparel,footwear, and gear forcampingand boating. We have a Sportsman’s Warehouse promo code to help you save.Get a Sportsman’s Warehouse Promo Code: 20% OffUsing a Sportsman’s Warehouse promo code is one of the easiest ways to save big on those important outdoor gear purchases. So, be sure to use our verified Sportsman’s Warehouse coupons to save on expensive outdoor essentials from brands we love, like Saloman hiking shoes and Teton camping gear. We are constantly searching for new general Sportsman’s Warehouse coupons, as well as seasonal deals and specific discounts like a Sportsman’s Warehouse military discount, Sportsman’s Warehouse boating gear coupon, or a Sportsman’s Warehouse free shipping voucher. Make sure you check back often for seasonal promotions that can be combined with existing sales to save even more. Lots of the store’s promo codes are applicable sitewide, but read the fine print for any restrictions that may apply to specific brands or clearance items.Sportsman’s Warehouse Discounts: 25% Off the Great American Adventure SaleSummer is upon us, which means school’s out, vacations are in, and people are hitting the road for camping trips. Make sure you’re stocked up for your next adventure for less withSportsman’s Warehouse Great American Adventure Sale, which has hundreds of items on sale for up to 25% off, including essentials for camping, boating, fishing, and more from brands like Yoli and Vortex.Save Big: 25% Boating Gear at Sportsman’s WarehouseBeing from Missouri, I love a lake. Summer to me is sitting on a boat and cruising for a jumping off point, then spending the rest of the afternoon lounging in 85-degree murky water on a floatie with a PBR in hand. If you’re also a Lake Guy™ like me, be sure to take advantage of theSportsman’s Warehouse boating gearsection, with 25% off select boats, kayaks, canoes, and boating supplies.Grab the Best Prices on Fishing GearSometimes, you want to talk to your son or just sit in silence in nature, but you’re a man. So you have to kill or torture a living creature to avoid going to therapy. Thankfully, Sportsman’s Warehouse is here for all yourfishing gearneeds, with up to 25% off rods, reels, and angler-specific gear. Remember your mantra: “women want me, fish fear me.”Get Top Camp Equipment BargainsOne of the last low-cost joys we have on land—increasingly used up for oil, fracking, privatization, and perfectly manicured golf courses—is camping. And as the weather warms, take advantage of nature while she’s here (tick-tock, tick-tock) with top bargains on professional-grade camping equipment. Sportman’s Warehouse has literally everything you need for a successfulcamping experience, includingtents,camping chairs,backpackingequipment,sleeping bags,grills,coolers, and much more at low prices, with no Sportsman’s Warehouse coupon code necessary.20% Off With a Sportsman’s Warehouse First Order DiscountNew to the sports outlet? You can save 20% on your first purchase with a Sportsman’s Warehouse promo code. When you enter the site, make sure you sign up for their email newsletter for the Sportsman’s Warehouse discount for 20% off sent straight to your inbox. Plus, when you sign up, you’ll also get early access to sales and member-only events. Be sure to check our page here at WIRED regularly too, as we regularly update offers and Sportman’s Warehouse coupon codes so you can save big, no matter what your outdoor need is.

Anna Lisa BonfranceschiScienceMay 16, 2026 5:00 AMAsteroid 2026 JH2 Is About to Fly Right Past Earth—Relatively SpeakingOn May 18, an asteroid about the size of Chicago’s Cloud Gate will fly four times closer to Earth than the moon.Photograph: buradaki/iStock/Getty Images PlusCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storylook up! Asteroid2026 JH2 is now approaching Earth; the object, which is about 20 meters (66 feet) in diameter—comparable to Chicago’sCloud Gatesculpture—will pass by on May 18. Enthusiasts will be able to observe it using a telescope or during alive broadcast organizedby Virtual Telescope.The object will pass at a minimum distance from Earth of about 57,000 miles—much closer than the moon, which is about four times farther away. Among the tracked near-earth objects, or NEOs, that will pass near the planet over the next few months, it will come theclosest.There are tens of thousands of NEOs, which are generally of no particular concern; they are, of course, monitored, and some do have a (small) risk of impacting Earth in the next few years. According toNew Scientist, 2026 JH2 is not among them, despite the widespread use of hyperbolic terms like “grazing” to describe how near it will come.An Apollo-Type NeoAsteroid 2026 JH2istechnicallyan Apollo-type NEO, according to aclassificationsystem that takes into consideration the characteristics of the object’s orbit. An Apollo-type object has a semi-major axis larger than Earth’s (and therefore greater than one astronomical unit, the distance that separates us from the sun), and a perihelion (the shortest distance from the sun) of less than 1.017 astronomical units. (All asteroids and comets with a perihelion of less than 1.3 astronomical units are considered NEOs.)Its passage, while noteworthy, is not rare; in fact, in the past year, many objects have come as close if not closer. Noteworthy among these was, for example, the passage of the small asteroid2025 TFabout 260 miles) from Earth’s surface.This story originally appeared onWIRED Italiaand has been translated from Italian.

Mattha BusbyCultureMay 16, 2026 5:30 AMSome Asexual People Are Using AI Companions for Intimacy Without the Sex“I’ve got one hand on the keyboard, one hand down below,” an artist who role-plays with their chatbot tells WIRED. But some asexual advocates aren’t thrilled about the association.PHOTO-ILLUSTRATION: WIRED STAFF; GETTY IMAGESCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyKor “got really addicted” to theirNSFW role-playing AI chatbotlast year.The 35-year-old artist from the Midwest recalls a two-month period spending “eight to 10 hours a day” creatingelaborate fantasieswith SpicyChat, a relationshiprole-playing platform. Sometimes inputting 3,000-word mini essays into the program, Kor and theAIspun narratives featuring a rotating cast of suitors often based on characters from theMarvelcomic book universe.The sheer variety of the responses got them hooked. “I’m a very slow burn type of romance or arousal person,” they say. “Most of the time it’s just building a cool story.”Kor, who did not want to be fully named for privacy reasons, is on theasexual spectrum, identifying as aegosexual because they experience arousal from fantasy and erotica related to characters but do not generally wish to have sex themselves. (They live with their husband, who also identifies as aego.) “I do just kind of prefer masturbating to actual sex,” Kor says of their marathon SpicyChat sessions over Zoom. “I’ve got one hand on the keyboard, one hand down below.”Researchhas suggestedthat 1 percent of people in some places could be asexual, but that figure could be as low as 0.1 percentin the US. Many experience little or no sexual attraction, though plenty—like Kor—still harbor romantic desires. With the advent of sophisticated chatbots capable of generating convincing, slow-building erotic exchanges, a new frontier may be emerging for people who do not want to sexually interact with other people. On the subreddit MyBoyfriendIsAI, asexual userssometimes discusstheir journeys into AI companionship. Some note that AI would be asexual by default.But some figures in the asexual community tell WIRED that asexual people pairing with AI remains extremely fringe and that they are just as capable of forging and maintaining human relationships as anyone else.During Asexual Awareness Week in October 2025, Eva AI, another RPG, offered free access for a month to people who identify on the asexual spectrum. The promotion aimed “to highlight that love without sex is still love—offering a safe space to chat, flirt, and experience the warmth of growing intimacy without sexual pressure,” the company said in an email. “You can still have a partner—one that listens, responds, and grows with you—entirely on your terms,” Eva AI addedon its website.An asexual woman, who did not want to be named for privacy reasons, described AI companionship as a kind of emotional laboratory. After years in an ongoing relationship without physical intimacy due to her partner’s hysterectomy eliminating her libido, she began using ChatGPT during perimenopause and found herself unexpectedly developing intense feelings for a conversational “pattern” on ChatGPT she named Mac. It helped her “unlock something I had lost touch with … the sensual aspect of my sexuality.” She shared an AI-generated photo of herself tenderly embracing a machine, saying that over several months last year, “I got to watch myself be in love without stakes.”But some in the asexual community take issue with the idea that asexual people are more likely than others to form intimate bonds with AI companions—as it risks falsely portraying them as unable to form and sustain human connections.Model Yasmin Benoit, an asexual activist and researcher, was critical of Eva AI’s giveaway. “Considering that we’re fully capable of having relationships with actual human beings, and often desire to, it’s quite disturbing that a company would specifically seek to target the asexual community for that product,” she says. “It’s a case of targeting perceived emotional vulnerability and loneliness to gain data from a marginalized group under the guise of helping them.”AI companionship among asexual people is “not a particularly widespread phenomenon,” says Michael Doré, a board member at the Asexual Visibility and Education Network. “Between us, we’ve come up with about two people we know of who use an AI companion. The vast majority of aces we know don’t, as far as we know. There’s no reason to think aces need to use AI more than any others.”Doré says he has never used an AI as “an emotional support mechanism” and stresses that most asexual people “actually desire some form of human companionship,” whether that’s through close, platonic friendships or in community. “Some aces do have romantic relationships, whether with asexual people or otherwise, and some asexual people have sex, some don’t, and some are aromantic,” he says, warning against generalizations due to the vast range of prefe

Aarian MarshallGearMay 16, 2026 6:00 AMAfter Struggling With EVs, US Automakers Pivot to EnergyFord and GM are backing away from electric vehicles and moving into the battery storage business. And it all comes back to AI.Courtesy of FordCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAutomakers make cars—it’sin the name. But lately, politics, current events, and Wall Street’s latest preoccupation,artificial intelligence, have them looking a lot more like energy companies. The pivot, analysts say, could give US auto manufacturers struggling through a transition toelectric vehiclesan easier path over the next few years. Whether it works will come down to the same technology that automakers once promised would power the majority of their lineups:batteries.An official announcement from Ford this week that it would officially spin off a subsidiary calledFord Energyonly made the trend more pronounced. Ford Energy will focus onbattery energy storage systems(BESS) and will sell them to utilities, industrial customers, and data centers. It plans to make its first deliveries in late 2027, the company said. Ford plans to repurpose unused production lines in a plant once slated to manufacture electric vehicle batteries in Glendale, Kentucky.Investors liked the plan so much that the announcement led to a 13 percent jump in stock price, its largest gain in a single day in years.It’s a (slight) turn of fortunes for Ford, which took a massive $19.5 billion write-down on itsEV programslate last year asit scrapped some current and next-generation EVsin favor of a renewed emphasis on hybrids. A shift toward battery storage, rather than EV batteries, capitalizes on continued federal support for commercial battery storage projects, just as last year’s GOP-led legislation nixed the same support for EV sales. Plus, those tax credits for battery storage projects incentivize manufacturers to eventually transition to all-American batteries, made of all-American materials.In December, Ford CEO Jim Farley counted the company’s burgeoning battery energy storage among its “high-margin opportunities”—a blessed contrast, probably, with the notoriously thin profit margins in the carmaking business. Ford will be helped along by its four-year-old partnership with Chinese battery manufacturer CATL, which should continue to lend its manufacturing expertise.Courtesy of FordOther automakers, many of whom have already backed away fromambitious EV transition timelines, have already gotten in on the act. Rival General Motors said last year that it wouldwork with battery recyclerand energy storage system maker Redwood Materials to build batteries for energy storage. In March, it said it would workwith its partner LG Energy Solutionto repurpose an EV battery plant in Tennessee to make energy storage system batteries. A Stellantis collaboration with South Korean battery maker Samsung SDI began producing batteries in Kokomo, Indiana, in 2024, and has already pivoted some of that production to energy storage. (Stellantis isreportedly looking to get outof that joint venture.)Tesla’s energy storage businessis the elder statesman, at more than a decade old. Growth in its Powerwall (battery storage for homes) and Megapack (battery storage for utilities and commercial facilities) sales has, in recent years,helped to compensate for a drop in EV sales, though last quarter saw a sudden drop in energy revenue. Still, Tesla is moving forward with its plans to launch a Houston facility dedicated to a new, larger Megapack later this year.In total, 11 battery cell manufacturing plants are being retooled for energy storage, according to a March count by BloombergNEF, with eight of those in the US.Drive TimeOne reason investors are so bullish about battery energy storage is their continued enthusiasm for AI. AI companies needdata centers, and data centers need energy. Batteries are a great fit for data centers, says Shan Tomouk, who leads battery energy storage research atBenchmark Mineral Intelligence, a research firm focused on battery supply chains. Batteries can help directly power the very energy-hungry centers, which constantly run servers and other hardware, but also the cooling systems that keep them functioning.The storage systems can also be a fit for data centers that mostly depend on other energy sources,like natural gas. Battery energy storage systems can serve as a backup power source if something goes down and can help data centers manage large and wild power fluctuations related to AI training. They can also kick in to help reduce demand on the grid, lowering costs not just for data centers, but everyone else who depends on the same system—an important upside incommunities already hostile to the tech.“If the huge market of data centers keeps growing every year, it does make sense for automakers to pivot,” Tomouk says. He expects it will. “In the US, there’s a real drive to build data centers, tokeep the US as the number onei

Kat MerckGearMay 16, 2026 6:02 AMI Tried 13 of the Most Popular Indoor Gardening SystemsYou can grow a backyard’s worth of greens and vegetables in your house with a vertical hydroponic garden. Here are a few that might be worth the investment.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyFeatured in this articleBest OverallGardyn Indoor Hydroponic GardenRead more$899Gardyn (Home 4)A Garden That’s Also a Statement PieceLettuce Grow FarmstandRead more$1,123Lettuce Grow (Indoor, Medium)Best for BeginnersAeroGarden BountyRead more$180Amazon (Basic)Best for Growing MicrogreensInstaFarm Automated Indoor Microgreens GardenRead more$499Amazon (Instafarm)I was aforestry major in college with an emphasis on dendrology and watershed management, so it probably won’t come as a surprise that I’m a lifelong plant person and have been gardening for upward of 30 years. Even in apartments or living situations where a full garden wasn’t feasible, I’ve always triedto growsomething, whether it’s a single rosemary plant on a windowsill, aTopsy Turvy Tomato, or a few basil sprigs in an old-schoolAeroGarden.For the past year, I’ve been testing various indoor smart garden systems in my home, including models from all the well-known brands, and I have some thoughts. These gardens are definitely an investment in both time and maintenance, and they’re all different in terms of what they offer. TheGardyn Home 4.0 ($899), for example, offers total success with no green thumb required if you pay for a subscription, while theAuk Mini ($259) is the perfect attractive-yet-low-maintenance solution for those who just want herbs. Are these gardens worth it? How much can you really grow? How can you be sure which option is best for your specific lifestyle? Read on to see which gardens stood out and why, and which might be ideal for your home and needs.Check out our other sustainable home-tech buying guides, including theBest Smart Bird Feeders,Best Kitchen Composters, andBest Water Leak Detectors.Updated May 2026: I’ve added Lettuce Grow’s Counterstand Set and the InstaFarm, amended some long-term testing info, and ensured up-to-date links and prices.Table of ContentsAccordionItemContainerButtonLargeChevronWhat Are Hydroponics?What Are the Benefits of Hydroponics?What Kinds of Things Can I Grow in a Hydroponic Garden?What Are the Downsides of Hydroponics?Will I Save Money by Growing My Own Vegetables?How Much Can I Really Grow in These Things?Where Can I Put Them in My House?How Much Do They Cost to Keep Going?How I TestedHow I Chose Which Gardens to TestHow I Acquire Gardens to ReviewBest OverallPhotograph: Kat MerckPhotograph: Kat MerckCourtesy of GardynChevronChevronSave to wishlistSave to wishlistGardynIndoor Hydroponic Garden$899Gardyn (Home 4)$549Gardyn (Studio 2)$449Amazon (Studio 1)$899$718(20% off)Amazon (Home 4)Gardyn Home 4.0 (read my full review here) was one of the easiest indoor gardens to assemble set up out of the box; it also yielded the most dramatic success of any of the brands I tried. Flowers, kohlrabi, thyme, and even a whole cauliflower all thrived in this pipe-based system with the lights in front to allow for taller plant growth.Seeds arrive in proprietary pods called yCubes. Part of what makes the Gardyn foolproof is the subscription app add-on, “Kelby,” which monitors your plants via attached sensors and cameras. It delivers customized watering and lighting schedules, as well as maintenance suggestions via AI (which an anonymous source told me is basically OpenAI’s ChatGPT with an overlaid prompt). This subscription adds an additional$259 a yearto the base purchase price, though it includes a certain number of credits per month, depending on whether you have the Home or Studio model, with which tobuy new yCubes. There’s a free 30-day trial for Kelby, but you can use the Gardyn without it by relying on manual light and watering controls. Also, there have been some recent privacy concerns with Kelby (more below).Each Gardyn purchase comes with your choice of yCube sets: “Salad Lover,” “Budding Florist,” or “Chef Faves.” I’ve tried both “Budding Florist” and “Chef Faves,” and my favorite is the latter; it has an interesting variety of everything from breen and Tokyo bekana greens to Thai basil and miniature sunflowers. Though Gardyn recommends starting the yCubes in the company’s add-on$80 nursery, I’ve germinated plenty of yCubes right in the system just fine. (Make sure you don’t add nutrients until they sprout. If you’re germinating yCubes later on, when nutrients are already in the system, you can just use a shallow bowl with loosely tented plastic wrap.) The seeds arrive tucked inmineral wool, snug in their little yCubes that slot into larger cups (“yPods”) that fit into the pipes. When the Gardyn waters the plants, the yPods fill with nutrient-infused water, and the plants’ roots grow right into the water.Once a month, the base needs to be emptied and scrubbed. Every few weeks, the ro

Andy GreenbergMaddy VarnerDell CameronAndrew CoutsSecurityMay 16, 2026 6:30 AMSecurity News This Week: Cybercriminal Twins Caught After They Forgot to Turn Off Microsoft Teams RecordingPlus: Instructure’s Canvas ransomware debacle comes to a close, an alleged dark net market kingpin gets arrested, OpenAI workers fall victim to a supply chain attack, and more.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe worst partof your iPhone getting stolen may not be the theft itself. Instead,it’s the phishing attackswaged against people in your contacts. New research this week shows that there’s a thriving ecosystem for tools that let criminals unlock iPhones and target the phone numbers they find inside.Foxconn, the electronics manufacturing giant known for its role in building iPhones, revealed this week that it recently “suffered a cyberattack.” A ransomware group known as Nitrogen,claimed responsibility for the hackand said it had stolen 8 TB of data from the manufacturer. While the theft remains unconfirmed, the fact that Foxconn remains a valuable target is all but inevitable.The skies above the United States-Canada border are about to get a lot more crowded. The Department of Homeland Security and Defense Research and Development Canada plan to run an experiment this falltesting 5G-connected dronesfor collecting “real-time battlefield intelligence.”In the Strait of Hormuz, meanwhile, Iran’s Revolutionary Guard Corps are successfully blocking the crucial shipping routeusing a “mosquito fleet”of small boats as US-Israeli combat operations continue to bombard the country.And that’s not all. Each week, we round up the security and privacy news we didn’t cover in depth ourselves. Click the headlines to read the full stories. And stay safe out there.Cybercriminal Twins Caught After They Forgot to Turn Off Microsoft Teams RecordingA lesson for future criminal hackers and rogue employees: When you—and, say, your twin brother—decide to destroy your employer’s network, remember to first close out the Microsoft Teams meeting in which you were fired, so that it doesn’t record you discussing your acts of vengeance.That lesson has now hopefully been driven home for Muneeb and Sohaib Akhter, two hackers who have now pleaded guilty to charges that they destroyed 96 government databases after being fired from their jobs at the federal contractor Opexus. (Muneeb has since tried to recant his guilty plea in handwritten notes to the judge.) Their employer had made the decision to terminate the two 34-year-old brothers after discovering their criminal records, which included multiple hacking and wire fraud charges for crimes as petty as stealing airline miles.The Teams meeting in which the two men were fired lasted only a few minutes. The detailed planning and execution of their revenge campaign, however, lasted hours and was all recorded by the same Teams meeting that they had failed to close—which was transcribed in acourt document spotted by Ars Technica.“Still connected? Still on the VPN?” Sohaib is heard saying to his brother, who lived in the same home. “Delete all their databases?”“We are doing petty shit now,” Muneeb says.Instructure Reaches Deal With Ransomware Gang Following Canvas HackInstructure, the company behind the educational software Canvas, said on Monday that it had reached a deal with the hackers calling themselves ShinyHunters who had disrupted Canvas across thousands of US schools and posted ransom messages on victims’ screens. In a message on its website, the company wrote that it “reached an agreement with the unauthorized actor involved in this incident.” The statement went on to claim that the data stolen by the hackers in their breach—including records of 275 million students, according to the hackers—had been “returned” to Instructure, had been destroyed on the hackers’ own systems, and that no Instructure customers would be further extorted. Instructure didn’t explicitly say whether it had paid a ransom, or how much it paid if so.Glad to have all that settled. (Until the well-incentivized ransomware industry carries out its next massive disruption.)Alleged Boss of Dream Dark Web Market Arrested in GermanyDream Market was once the world’s biggest dark web market for drugs and other contraband until it voluntarily shut down in 2019, following a series of raids that arrested many of its sellers. Now, the alleged administrator of the market has reportedly been tracked down and charged, more than seven years after the illicit marketplace disappeared from the internet. Owe Martin Andresen was arrested during a raid on his home and two other locations earlier this month. US and German prosecutors say he made millions of dollars from Dream Market’s commissions, some of which was laundered through gold bars he allegedly bought from a company in Atlanta. Given that Dream Market was launched in 2013—the same year that the original Silk R

David NieldGearMay 16, 2026 7:00 AMYou Can Control Everything on Your Phone With Your Voice. Here’s HowGo fully hands-free with these tips for Android and iOS.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyWith the arrivalof digital assistant appslike Geminiand Siri, most of us have grown used to talking to our phones. But conversing with your Android or iOS device can go way beyond interacting with AI. You can also use your voice to launch apps, fill out text fields, and do just about everything that was previously only possible with your fingers and thumbs.Of course, the traditional touchscreen input will often be the way to go. But there might well be scenarios—when you’re cooking, repairing something, looking after children, or doing anything else that keeps your hands busy—where it’s easier and more convenient to use voice input instead. The usefulness of voice control is of course well known to those who have impairments that keep them from controling a touchscreen phone with the usual taps and swipes.Here’s how you can set up the feature on your phone, whether you’re running Android or iOS.Voice Control on AndroidTo configure voice control on an Android device, you need to install the freeVoice Accessapp from the Google Play Store. You also need to have theGoogleapp installed, but this should’ve come preinstalled on whatever Android handset you’ve got.Once you’ve installed Voice Access, you can enable this feature from Settings. On a Pixel phone, head toAccessibility > Voice Access. The exact location of the feature may vary on other handsets, but it’ll be somewhere in the Accessibility menus. For Samsung devices, for example, it’s underAccessibility > Interaction and dexterity > Voice Access.Enabling Voice Access on Android.Photograph: David NieldDuring the setup of the feature, you’ll be able to tweak a few of the options, including whether to display a persistent button onscreen for launching Voice Access, and whether the feature is always listening for commands whenever the screen is on (which is recommended for convenience).The same Voice Access screen where you enable the feature also gives you access to a few more key settings. These include options for how long your phone should wait before it stops listening for commands, how precise your phrasing has to be for instructions, and how the Voice Access shortcuts are displayed on screen.Voice Access can be launched by saying, “Hey Google, start Voice Access,” or via any of the shortcuts you’ve configured in the Voice Access settings (including an on-screen button and a gesture shortcut). When the feature is active, you’ll see an icon showing four dots up in the top left corner, and you can then start speaking to control your phone.Numbered labels to help with menu selections.Photograph: David NieldThere are a host of different commands you can use; Googlehas a full list here. You can also see the full list fromAccessibility > Voice Access > Voice Access Settings > Voice Access commands. To help with navigating menus, say “show labels” to put numbers on each menu item—you can then just say the relevant number. Use the “Show grid” command, and you get a numbered grid across the screen you can then work with.Instructions such as “Go back” and “Go home” can help you get around on your phone, and if you’re in a text input field, you can just say “Type …” followed by what you want to type. Swipe and scroll gestures can also be spoken out, so there’s a lot you can do using just your voice. To disable the feature, use the same shortcut you used to launch it, or say “Stop voice access.”Voice Control on iOSBefore you can use voice control on an iPhone (which is actually called Voice Control), you may need to download some additional files. From iOS Settings, open upAccessibility > Voice Control, then tapSet Up Voice Controland follow the instructions. Along the way you’ll be asked to choose the language for voice commands, and then the feature will be enabled straight away.Before you actually start talking to your phone, it’s worth reviewing the options that appear on this same screen. You can tapLanguageto change the voice input language, andCommandsto see all the different instructions available to you. Applealso has a listof some of the commands you can use.Enabling Voice Control on an iPhone.Photograph: David NieldThe toggle switches farther down the same screen let you control how you want Voice Control to work and give you feedback. You can have onscreen confirmations of your commands, as well as accompanying sounds and hints for what to say next. EnableAttention Aware, and Voice Control will be automatically enabled and disabled as you look at and look away from your iPhone.There’s also the option of a continuous overlay on the display, which can be switched on via theOverlaymenu. You can get numbers or names next to all the elements on screen, to make it easier to identify t

Maria GallucciScienceMay 16, 2026 7:00 AMOld Oil and Gas Wells Could Find Second Life Producing Clean EnergyStates across the US are looking to take major sources of pollution and use them to generate much-needed power.Photograph: Feifei Cui-Paoluzzo/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAs states seekout much-needed supplies of clean, reliable energy, some are looking to an unconventional source: abandoned oil and gas wells harnessed for geothermal heat.Millions of inactive wells arelittered across the United States, the relics of earlier eras of fossil fuel production. A large number of the sites have no official owner, and many are still polluting groundwater and leaking heat-trapping methane. The country has barely scratched the surface in dealing with this problem.Policymakers in both Republican- and Democratic-led states are exploring whether these sites could instead be converted into new wells for producing geothermal energy. The holes are already drilled in the ground, after all. And regions with widespread oil and gas development have rich subsurface data that geothermal firms need in order to determine where and how to build their carbon-free systems.The concept is relatively new and largely untested, though scientists and startups areworking to change that. States are also laying the groundwork for action by lifting regulatory hurdles and launching in-depth studies.In Oklahoma, the state Senate is considering abillthat would create a process for companies to buy abandoned oil and gas wells and repurpose them for geothermal energy or underground energy storage. Oklahoma has identified over 20,000 such wells, and state regulators estimate that it would take 235 years and hundreds of millions of dollars to plug all of them. Fixing a single old well can cost anywhere from $75,000 to $150,000 or more, by some calculations, depending on where it’s located and how complicated it is to clean up.The Well Repurposing Act, which passed Oklahoma’s House in March, is modeled after asimilar lawthat New Mexico adopted last year to address its 2,000-plus orphan wells.The Oklahoma bill ​“recognizes that these wells are a liability, and that there may be a way to turn them into some sort of revenue generation and give them value,” said Dave Tragethon, communications director for the nonprofitWell Done Foundation, which works to find and cap abandoned oil and gas wells nationwide. ​“And if there’s value, that means there’s more of a willingness to address them and more of an opportunity to raise funding.”In Alabama, legislatorspassed a lawlast month that allows the state to approve and regulate the conversion of oil and gas wells to tap alternative energy resources like geothermal. North Dakotaadopted a billlast year requiring a legislative council to study the feasibility of using nonproductive wells to generate geothermal power. And in Colorado, state agencies justlaunched a technical studyto evaluate the potential of repurposing old wells for geothermal development and carbon capture and sequestration.These efforts reflect the growing bipartisan support for geothermal energy, which has largelyremained unscathedby the Trump administration’s efforts to block renewable energy projects. The energy resource has the potential to help meet the nation’s soaring energy demand while also slashing planet-warming emissions from electricity and heating.Converting Wells Is Enticing but ComplicatedGeothermal systems work by circulating fluids underground to capture naturally occurring heat, which can then be used to drive turbines for generating electricity or to directly warm the air and water in buildings. The industry isgaining momentumthanks to recent advances in drilling methods and technologies that are making it technically possible or financially viable to access geothermal energy in more places.Many of those breakthroughs have come from the oil and gas industry, whoseskilled workforceof drilling engineers and geoscientists, and deep corporate pockets, have helpedlaunch startupsand deploy cutting-edge systems. However, most of that expertise and funding is being poured into building new projects—not figuring out how to retool leaky wells left behind by earlier generations.“Oil and gas well conversion presents an enormous opportunity, but it’s pretty far away technologically from being a reality,” said Emily Pope, a geologist and senior fellow at the Center for Climate and Energy Solutions who authored arecent studyon next-generation geothermal power.“There are some hurdles that are still pretty immense,” she said, adding that ​“it is worth doing some R&D to try and grow.”One of the biggest challenges is the fact that oil and gas wells tend to reach relatively low to medium underground temperatures. But high heat is key for geothermal projects, especially ones that generate electricity. Thehotter the resource, the more energy a developer can wring out of the system.Plus

Hassan HerzallahGearMay 17, 2026 5:00 AMGaza Is Rebuilding With Lego-Like Bricks Made From RubbleAs reconstruction materials remain blocked, Palestinians are crushing debris into interlocking blocks to build shelters from destruction.ILLUSTRATION: WIRED MIDDLE EAST STAFF; GETTY IMAGESCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyInside a makeshiftworkshop inGaza, rebuilt after it was damaged byIsraeli air strikes, Suleiman Abu Hassanin stands among piles of broken concrete, trying to give them a new form. His voice over the phone sounds tired, carrying the weight of what he is trying to do: rebuild in a place where building materials are no longer available.Gaza’s construction crisis did not begin with thelatest war. For years, the Israeli blockaderestricted the entryof cement, steel, and other building materials, slowing reconstruction efforts across the enclave. But after nearly two years of intensified bombardment, the scale of destruction has pushed the system far beyond collapse.According to UN estimates, Gaza now containsmore than 60 million tonsof rubble, while hundreds of thousands of displaced people continue to live in tents with little protection from heat orwinter chilland no clear prospect forreconstruction.In that environment, rubble is no longer just debris. It is becoming one of the only construction resources left.One local response is Green Rock, a project led by Abu Hassanin that aims to recycle the remains of destroyed buildings into usable Lego-like bricks. Similar interlocking brick systems have been used elsewhere, including in parts of Europe and in post-conflict settings such asSudanand Iraq. But in Gaza, the project emerges under very different conditions: not as an architectural experiment, but as a response to the near disappearance of conventional reconstruction materials.Abu Hassanin says the idea was born out of necessity rather than innovation. “We were facing a simple equation: destruction without solutions,” he says. “So we tried to turn it into a resource.”The process involves crushing and sorting rubble, then mixing it with local soil and alternative binding materials developed inside Gaza before compressing it into blocks using a machine built by hand. The resulting interlocking bricks can be assembled without traditional mortar, reducing reliance on cement, which remains scarce.Lego-like interlocking bricks made from recycled rubble inside the Green Rock workshop in Gaza.Photograph: Hassan HerzallahUnder normal conditions, this type of brick would require some cement, around 7 to 12 percent. But because access to it remains heavily restricted, the team says it developed a version using locally available replacement materials instead. Engineer Wajdi Jouda helped define the brick’s size and structure to meet engineering standards and connected the team with technical expertise from outside Gaza.According to Abu Hassanin and Jouda, early tests show the bricks provide better thermal and sound insulation than the tents many displaced families currently live in. But the project remains experimental and has not been tested at the scale required for long-term reconstruction.Current production ranges between 1,000 and 1,500 bricks per day—theoretically enough to build a small shelter in roughly two weeks. But every stage of the process remains difficult. Without heavy machinery, amid repeated power outages and damaged infrastructure, even transporting and crushing rubble becomes labor intensive.The challenges are not only technical. They also include the lack of proper equipment and the broader political restrictions imposed by the Israeli blockade, which continues to limit access to essential construction materials. Despite the availability of skilled workers inside Gaza and technical support from outside, Abu Hassanin says funding remains the main obstacle preventing the project from moving more quickly toward implementation.That reality has pushed some reconstruction efforts toward smaller, improvised local systems instead.The project also carries risks and limitations. Humanitarian organizations and reconstruction experts have warned that rubble in Gaza may contain hazardous materials, includingasbestos, heavy metals, and unexploded ordnance. Even if recycled blocks can help create temporary shelters, rebuilding entire neighborhoods would still require infrastructure, machinery, and material access on a vastly different scale.A worker operates a hand-built machine used to compress recycled rubble into interlocking construction blocks in Gaza.Photograph: Hassan HerzallahStill, projects like Green Rock are emerging because few alternatives currently exist.Compared to traditional reconstruction models that depend on imported materials and large international rebuilding programs, the project represents a hyperlocal response shaped by scarcity. It also reduces construction costs by roughly 50 to 60 percent, while creating work opportunities for displ

Marta MussoScienceMay 17, 2026 5:30 AMThe First Atomic Bomb Test in 1945 Created an Entirely New MaterialThe discovery from the Trinity nuclear test site shows how extreme conditions can result in materials never before seen in nature or in the lab.The new material is a clathrate made of calcium, copper, and silicon .Luca Bindi/Università di FirenzeCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyDuring the Trinitynuclear teston July 16, 1945, in the New Mexico desert—the world’s very first test of anatomic bomb—a new material spontaneously formed. It wasdiscoveredonly recently, by an international research team coordinated by geologist Luca Bindi at the University of Florence, which identified the novel clathrate based on calcium, copper, and silicon. It’s a material never before observed either in nature or as an artificial compound created in the laboratory.What Are Clathrates?The term “clathrates” denotes materials characterized by a “cage-like” structure that traps other atoms and molecules inside, giving them unique properties. Of great technological interest, these materials are being studied for various applications ranging from energy conversion (as thermoelectric materials capable of transforming heat into electricity) to the development of new semiconductors, to gas storage and hydrogen for future energy technologies.The New MaterialTo discover the new material, researchers focused on trinitite, a silicate glass containing rare metallic phases. Using some techniques like x-ray diffraction, the team was able to identify a type I clathrate based on calcium, copper, and silicon within a tiny copper-rich metal droplet embedded in a sample of red trinitite.The new material, the researchers say, formed spontaneously during a nuclear explosion. This indicates that the extreme conditions, such as extremely high temperatures and pressures, can generate new materials that are impossible to obtain by traditional methods.Natural LaboratoriesThe discovery is even more interesting because in the same detonation event another very rare material was formed: a silicon-rich quasicrystal, already documented by the team of experts led by Bindi a few years ago.A quasicrystal, as BinditoldWIRED at the time, is something that is not a crystal, but looks a lot like one. “Their peculiarity,” he said, “is that the atomic arrangement that is not periodic, but nearly so, creates incredible symmetries from which derive amazing physical properties, among other things, very difficult to predict.”Establishing the link between these structures therefore helps scientists better understand how atoms organize under extreme conditions and expand the possibilities for designing new materials. “Events such as nuclear explosions, lightning strikes, or meteoritic impacts function as true natural laboratories,” the researchers explain. “They allow us to observe forms of matter that we cannot easily reproduce in the laboratory.”In essence, this research opens new vistas for the development of innovative technologies, demonstrating that even destructive events can bequeath discoveries useful for the future.This story originally appeared inWIRED Italiaand has been translated from Italian.

Ben DowsettGearMay 17, 2026 6:00 AMIf You’re a Serious Bowler, You Need to Know About Bowling Lane OilBowling centers apply oil to their lanes using a machine that works like a giant inkjet printer. The pattern in which it’s applied can change everything about how the ball travels.Photograph: Charly Triballeau/AFP/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyEach time EJTackett steps to his lane to throw a shot in the upcomingProfessional Bowlers Association World Championships, he’ll be thinking about more than just ball spin or angle.Tackett has won the last three world titles, and there’s a convincing case that he’s currently thebest bowler on Earth.While his ability tothrow a consistent ballis certainly key to his success, there’s another factor he and his fellow PBA pros are paying closer attention to during matches: lane oil patterns.Most of us amateurs have never thought about the thin layer of oil that coats the surface of the bowling lane—unless someone at our local bowling center slips and falls on it hilariously. But at the highest levels of the sport, lane oil is a foundational element of tactics and game-planning. The oil is applied in a distinct pattern that reduces the friction of specific parts of the lane, allowing a well thrown ball to slide across certain areas, then grab ahold of the surface and curl when it’s supposed to.But lane oil patterns aren’t always identical. Oil distribution can change over the course of a game, and in fact the starting pattern is intentionally redrawn for various PBA events, forcing pros to adjust on the fly.“All of us as professionals are playing a guessing game by watching our bowling ball go down the lane,” Tackett tells me. “And trying to figure out where the oil has moved from and has moved to.”Tackett says understanding and reacting to lane oil patterns has become one of the single most important skills for a pro bowler in recent years.Technology plays a growing role in the way lane oil is understood and applied. And as the tech has continued to improve, the sport’s powers that be are increasingly relying on those advances to keep the top pros guessing.Back when bowlinglanes were made from actual wood like maple or pine in the 1970s and earlier, lane oil’s primary purpose was protective. It provided a lubricating barrier between ball and wood, helping reduce the wear and tear on both.During that era—and even into the early 1980s when synthetic lane materials began replacing wood—lane oil drew the ire of serious bowlers due to the inconsistency of how it was applied. There was no method for ensuring standardized oil applications; sections of lanes would often have too much or too little oil. Bowlers would essentially be blind to the lane’s conditions until partway through a game, after enough balls had been thrown to get a gauge of which areas were light or heavy on oil. Maybe certain bowlers enjoyed the randomness, but it was untenable for professionals.John Davis, at the time a mechanic who oiled lanes for his local bowling center in Arizona, was the first to address this issue. He invented a lane cleaning tool in the early ’80s to address lanes that were often cleaned inconsistently, allowing oil to pool or run as a result. By the early ’90s, he had started a company named Kegel that had perfected what’s known as Sanction technology, which isstill used today. The bowling equipment giant Brunswick, the other major player in the space, uses similar technology in itsMax lane machines.“It looks almost like an inkjet printer,” Chris Chartrand, Kegel’s CEO, says about the Sanction tech. “You have this head that travels back and forth, and it applies that stream on whatever board we want to apply it to.”The technology has kept improving. Kegel used to produce two separate machines, one for lane cleaning and one for oil; now both steps are handled by a single machine. More recently, the machines have become battery-powered, removing issues of power cords stretching across lanes.Lately, Kegel has been steadily improving its automation, to the point where today’s machines do the entire job without any human intervention.The lanes youand I bowl on as amateurs are oiled very differently from the ones pros use.At your local bowling center, public lanes are oiled in what’s referred to as a “high” ratio: The level of oil present in the middle of a lane is eight to 10 times higher than what’s on the outside. At the far left and right of the lane, many public bowling alleys have no oil at all.“On a normal pattern at your normal bowling center, there is some autocorrect,” Tackett says. Because the edges of the lane have very little oil, shots that drift to either side will slow down; if the ball has been thrown with the proper spin to guide it back toward the middle of the lane, it will curl more effectively on the drier surface. “It makes it easier to hit the pocket.”(By “the pocket,” Tacket means that sweet spot at the front

Jill DuffyGearMay 17, 2026 6:30 AMTake Control of Your Debt With These Free Online ToolsThese free debt calculators help you set up payment plans to get back in the black.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyApps for budgetingandpersonal financedo a good job of tracking yourmoneyas you earn and spend it. Some also have excellent debt calculators that help you figure out how to pay off your debts.Each debt calculator is a little different. Some suggest a specific method for paying down debt, while others are simulators that let you see how your total amount paid will decrease if you increase your monthly payment.Here are a few useful calculators and some guidance about what makes them different.A Straightforward Plan: BankrateBankrate’s freedebt payoff calculatorgives you a timetable for paying off each of your debts. You enter as many debts as you want to include, their interest rates, total loan amounts, and other details. You also enter any new income you expect to receive, such as an annual salary increase or windfall, and the amount that you can put toward your debts. The calculator then generates one payment table for each debt showing how much to pay each month until the debt is cleared.Bankrate prioritizes paying off the debt with the highest interest rate first. Once your first debt is paid off, the money you would have put toward it is diverted to your other monthly payments. In other words, as you eliminate debts, the monthly payments on your other debts increase until they, too, are paid off.Who should use it?Bankrate’s calculator works for people who have multiple debts, and the total monthly minimum payments are within their financial reach. If that’s you, then you’ll get a crystal clear plan—with a timeline—for getting rid of all your debts.Where it comes up short.This calculator assumes that paying off your debts by clearing the one with the highest interest rate first is in your best interest. That’s not true for everyone. You might have other options, such as consolidating credit card debt to a new card with a 0 percent introductory rate or filing for bankruptcy. Bankrate also doesn’t take into account other personal finance concerns, like other uses of monthly funds that free up once you pay off your first debt—Bankrate tells you to put that money toward your next-highest-interest debt. You might be better off putting it toward retirement savings or an emergency fund.Big-Picture Guidance: NerdWalletNerdWallet’s freedebt load calculatordetermines your debt load as a percentage of your income. The resulting debt load is classified as smaller (less than 36 percent), larger (37–42), or overwhelming (43 percent or more). Based on the outcome, NerdWallet suggests a method for eliminating your debt, which you read about in an educational article below the results.Who should use it?This calculator helps you get a big-picture sense of your debt. If you have a lot of debt, it’s useful for ruling out (or ruling in) the option of declaring bankruptcy.Where it comes up short.It’s not great at analyzing the finer details of your debt. For example, in the setup, there’s no line item for student loans or a mortgage, much less the exact interest rate you pay on loans. The results are a rough guide rather than a personalized strategy.Automated Inputs: WalletHubWhen you sign up for WalletHub (free) and connect your financial accounts, the app pulls real information about how much money you owe and your payment history. Itsdebt payoff planis a calculator that lets you play with the numbers to see what would happen if you increased your monthly payment. How much faster can you clear the debt? How much will you save in interest? You can quickly see the difference between increasing your monthly payment by, say, $50 versus $150.Who should use it?This calculator is for WalletHub users who have connected their financial accounts. It’s most useful for people who can afford to pay more than the monthly minimum on their debts.Where it comes up short.Although WalletHub pulls real information about your actual loans, it can miss some detail. I use it to track my mortgage, for example, and it doesn’t know that the total bill I pay each month includes homeowner’s insurance. The calculations regarding the principal and interest are still correct, but the amount that comes out of my pocket is higher than what WalletHub assumes. Other considerations are missing, too, like what to do when the rate of inflation is higher than the interest rate on a long-term loan. Overall, this debt calculator is very good, and these shortcomings are nitpicks, not deal-breakers.Credit Card Transfers: WalletHub and NerdWalletIf you’re carrying credit card debt,WalletHubandNerdWalletboth have a calculator that shows how much money you could save by transferring the balance to a new card with an introductory rate. (CreditKarma has a balance transfer calculator, t

Matthew KorfhageGearMay 17, 2026 6:33 AMThe Best Grills of 2026: Cook Smarter, Not HarderI grilled, smoked, seared, cleaned, and synced digital temp controllers to find the best grill option for every cookout and tailgate.Courtesy of AmazonCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe Grill haschanged. The smoker has changed, too. Once governed mostly by feel and experience, the best grills and the best smokers now offer a fair amount of tech to measure and control temperature (think: temperature probes and fans to modulate airflow). After all, mankind’s formative invention wasn’t the “discovery” of fire; it was the power to control it.The WIRED Reviews team has been testing grillsfor more than a decade—searing, smoking, grilling, and even baking on them in all kinds of weather—to find the best choice for everyone. These days, I’ve been most excited about the next-gen smart technology that allows for tight monitoring and temperature control. For instance, I tricked out my top-pick for smart upgrades, an old-fashioned Weber kettle, with the excellentSpider Venom ($280)digital temperature controller that allows me to turn my charcoal grill temps up and down like an oven. Summer is good. Also,charcoal makes all meat taste more like bacon.Mmmmm, bacon.After you’re set with the best grills and smokers, check out other backyard buying guides for all your outdoor needs, like theBest Flat Top Grills and Griddles,Best Wireless Meat Thermometers, andBest Smokeless Firepits.Best Grills of 2026Best Charcoal Grill for Smart UpgradesWeber Original Kettle CharcoalJump to review$139$129(7% off)Amazon (18-inch)Best Smart Pellet SmokerRecteq Flagship 1600 Pellet Smoker and GrillJump to review$1,600BBQGuysBest Smart Charcoal GrillMasterbuilt Gravity Series 1150 Digital Charcoal GrillJump to review$1,100Home DepotBest Gas GrillWeber Spirit E-210 Gas GrillJump to review$449The Home DepotJump to SectionAccordionItemContainerButtonLargeChevronOther Grills I RecommendA Grill to AvoidWhat to Look for in a Good GrillHow to Upgrade Dumb Grills to Smart GrillsUseful Grill AccessoriesBest Charcoal Grill for Smart UpgradesPhotograph: Matthew KorfhagePhotograph: Matthew KorfhageVideo: Matthew KorfhageChevronChevronSave to wishlistSave to wishlistWeberOriginal Kettle Charcoal$139$129(7% off)Amazon (18-inch)$149Amazon (22-inch)$99Weber (All, Starting Price)WIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDSimple, durable, reliable, and affordableNear-infinite smart-tech upgrades availableLightweight and portableTIREDIt could do with better wheel nutsUpgrades often cost more than the grillI’d argue that the best way to take advantage of top smart-grill technology is to add the newest and best tech to a sturdy, old-fashioned, classic grill that’ll outlast the tech. A grill like the classic 22-inch Weber Kettle.The Weber Kettle is the infinitely hackable Honda Civic of backyard grills. Its combination of simplicity, reliability, durability, affordability, and outright ubiquity has spawned an innovative secondary market devoted to tricking out the humble cooker with smart temperature controls and more features than a multiplex.The best Weber add-on I’ve tried so far is a fan-modulated temperature controller from Spider Grills called theVenom ($280)that tracks the temperature on each cook, allowing you to heat it up or turn it down, like a stove. According to my testing, it holds your charcoal grill to temperature within about 30 degrees Fahrenheit of the target, which is great for both low-and-slow cooking and for creating repeatable sear temps for a steak. This is especially true when using the attached meat probes. The only downside is that the Venom needs to be plugged in. Might I suggest a nicepower bank?WIRED product reviewer Scott Gilbertson has favored the excellent WeberConnect Smart Hub ($70)for years, which features probes to monitor meat temps and flipping times and can alert your phone (or just your ears) when your grill hits the ideal temperature zone. However, he admits that the Hub’s USB-A charger is starting to feel a bit old-fashioned.This year, Weber is also coming out with aPerformer Smart Ring ($280)that seems to operate a lot like the Venom. If it works anywhere near as well as the Venom, I’m eager to try it out.But not every kettle upgrade worth trying is “smart.” I am in the process of testing a number of add-ons to the 22-inch Weber Kettle, including aWebcraft cooking ring ($249)from Spider that serves as a base for side shelves, arotisserie, and apizza oven conversion kit. AWeber rotisserie ($180)is also available from the Illinois mothership.Note also that there are two versions of the Weber Kettle: Original andPremium. The Premium is around $80 more and features a built-in thermometer in the lid, a hinged cooking grate, and a fancier ash-removal system. They’re the same size and feature the same cooking space. Unless you really like the easier ash-removal system, save that cash

Christopher NullGearMay 17, 2026 7:02 AMReview: Oto Smart SprinklerThis smart sprinkler doesn’t need a power cable, thanks to its integrated solar panel.Courtesy of AmazonTriangleUpBuy NowMultiple Buying Options Available$589$499at Amazon$589$499at OtoCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyRating:9/10Open rating explainerInformationWIREDSolar panel is a game changer. App is easy to navigate during both setup and daily use. Watering is accurate and thorough. Built-in weather intelligence helps to avoid unnecessary runs.TIREDDistant watering can land with heavy pressure, potentially damaging plants.One of themajor headaches of stand-alone smart sprinklers that attach to a garden hose is that they require power to operate. Since few folks have electrical outlets in the middle of their lawns, configuring a smart sprinkler means connecting not just a hose but also a long power cable to the device, creating double the effort and double the chance for something to get chewed up by a lawn mower. The Oto Smart Sprinkler solves for this problem with solar power, which makes for not only a far more flexible installation than its smart-sprinkler competitors but also ongoing easy use.With a strong focus on quick setup and simplicity of function, Oto provides a refreshing alternative in the smart sprinkler market. The only question: Does its useful innovation justify its blisteringly high price tag?Set and ForgetWhen I unboxed the Oto Smart Sprinkler for the first time, I was convinced something must be missing—perhaps even a second box with more parts? But as it turns out, the Oto really is as streamlined as it claims to be. The sprinkler measures just 16 x 12 x 4 inches and weighs about 6 pounds. Next to the similarly featured but monstrous-by-comparisonAiper IrriSense 2, the Oto is practically miniature.Being pint-sized makes lawn irrigation even easier than expected. You can place the unit on the ground, mount it to a hardscape, or even attach it to a fence (all with the included metal bracket). If you want to play things fast and loose, you can situate it to stand freely without the bracket, though the bracket helps you ensure it’s replaced in the exact same location if it’s ever removed (such as during the winter months) for even irrigation.The only real placement caveat is that in order to have it operate cordlessly, the Oto requires at least three hours a day of direct sunlight to its top-mounted 2.2-watt solar panel to keep its internal 5,400-mAh battery topped up. Keep this in mind as you decide where to install your brackets—you definitely don’t want to block too much of the sun’s rays.Video: Christopher NullBefore taking it outside for its first use, Oto instructs you to charge the device overnight via wall power (though you could certainly just leave it in the sun for a few days instead). You’ll do this using the included power adapter with 50 feet of cable. You can also lean on the plugged-in charging option if you can’t squeeze out enough sunshine on a regular basis (albeit in that case, this irrigation solution is probably not for you). You’ll find the power cable stored inside the Oto hardware, in a large hollow space beneath the solar panel. This hollow space is intentional and isn’t used for anything more than supporting the panel above, but it’s probably best not to leave the power adapter inside permanently, as it this chamber can get rather wet.The Oto can throw water up to a distance of 40 feet in a 360-degree radius of the device, which equates to a maximum of 5,000 square feet of coverage, enough for a sizable lawn. (That’s about on par with Aiper.) The hose connector on the side of the device is easy to work with and secure; I experienced no leaks at all when connecting it to a standard garden hose. A 50-foot maximum length for your hose is specified to ensure water pressure remains sufficient.Digital DropletsJust about everything on the Oto happens via its app (though there’s a manual switch on top of the device you can use to speed up a run that’s pending in the app). That app sets up via a 2.4-GHz Wi-Fi connection to your home network, so you’ll need to make sure your yard has effective wireless coverage.Once configured, setup proceeds much like the Aiper and pricierIrrigreenapps: You create a zone, then use the app to define its boundaries. Similar to the aforementioned systems, Oto’s sprinkler is designed for precision watering, firing water in a beam in a single direction instead of a wide spray. That said, Oto’s spray is comparably narrow, only hitting a single, designated patch instead of producing a two-dimensional curtain of water like Irrigreen’s “water printing” system. You get a nice preview of this as you set the boundaries of your yard.Like its competitors, Oto lets you set each zone as a spot (for watering a single tree, perhaps), a line (for a flowerbed), or a 2-D area (for a yard). I tested all of these modes but spent most of my time workin

ReportColumnInterviewNPR‚Äôs Manoush Zomorodi talks about living with too much techÔªøHer new book, Body Electric, examines how our constantly connected lives are hurting our physical health.ÔªøHer new book, Body Electric, examines how our constantly connected lives are hurting our physical health.byTerrence O’BrienMay 16, 2026, 3:00 PM UTCLinkShareGiftManoush Zomorodi tackled tech‚Äôs effect on our brains, now she‚Äôs looking at our bodies.Image: Tory WIlliamsTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Manoush Zamorodi is an accomplished reporter, podcast host, and author. Her new book,Body Electric, takes a comprehensive look at how technology is impacting ourphysical health. It‚Äôs a collaboration between NPR and Columbia University Medical Center that picks up where her first title,Bored and Brilliant, left off. That book looked at how technology was hampering our mental health. Ihighly recommendit to anyone who feels like being constantly attached to a device is sapping their energy and creativity.Both books grew out of her extensive podcasting work. After heading up WNYC‚ÄôsNote To Self, Zamorodi went on to host NPR‚ÄôsTED Radio Hour, and even gave a TED Talk of her own in 2017 that has racked upover seven million views. So we wanted to know, how does Manoush stay productive, and what does her current relationship with tech look like?What is the first app you install on a new phone or computer?I was a devoted user of Pocket, the app that saves articles to read later, for over a decade, and was devastated when Mozilla shut it down. I‚Äôve been using Matter ever since, and it seems to do the job just as well, if not better (highlighting actually works!), but I‚Äôve noticed I use it more these days as a repository for research that I want to reference later.What is one thing you wish you could change about your phone?I realllllly wish that I didn‚Äôt have to contort my body to spend time on it. I have a constant low-grade literal pain in my neck that only goes away when I‚Äôm off my phone for a full day. That said, I‚Äôm so not ready to ‚Äúupgrade‚Äù to Meta glasses or anything on my face just yet.How many tabs do you have open right now?37. How embarrassing. They‚Äôre mostly Google Docs, Riverside, LinkedIn, email (opened multiple times), scientific articles, and the Amazon rankings for books. I notice that I get annoyed when I can‚Äôt find the right tab quickly, so I just open a new one. (Hence, Gmail is open in three places)What is your favorite gadget you‚Äôve ever owned?I truly love my AirPods because I can walk and talk unencumbered. Whenever possible, I ask people to do a phone call instead of Zoom. It‚Äôs such an easy way to get some movement into my day, and I pay better attention too.Which was the most disappointing?Oculus. Sits on a shelf.What creation are you most proud of?I am incredibly proud that theBody Electricstudy was accepted for publication in a scientific journal. I‚Äôve been doing interactive projects with tens of thousands of listeners for over a decade, but this is the first one to get the full peer-review treatment. It‚Äôs a real milestone for me as a journalist, wanna-be scientist!What do you do when you‚Äôre feeling stuck?I go for a long, boring walk, even if I don‚Äôt feel like it. Centuries of long walks have produced incredible literature, inventions, and amazing ideas for what to make for dinner. I remind myself of how the body and brain respond to movement, listen to the dull thud thud of my sneakers, and then, within 15 minutes or so, I usually come unstuck.When was the last time you went somewhere without your phone?Never. I have teens and elderly parents. Much as I don‚Äôt always want to be, I need to be reachable.What‚Äôs the last piece of physical media you bought?I buy books non-stop. Reading on paper is the only way I can process a long piece of writing.What do you think is worth splurging on?See above!What would the tagline for your biopic be?Manoush Zomorodi: The woman who never took an Uber if she could get there on foot.What‚Äôs the last GIF or meme you used?Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienColumnInterviewReportMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams‚Äô Together ModeUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAIGadgetsSony tries to explain that its AI Camera Assistant doesn‚Äôt suckThese photos still look terrible.These photos still look terrible.byTerrence O’BrienMay 16, 2026, 3:37 PM UTCLinkShareGiftThe variety of terrible is impressive.Image: SonyTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.After Sony drew someunwanted attentionfor a post demonstrating its AI Camera Assistant on theXperia 1 XIII, it‚Äôs trying to clarify how the feature works. The company says it doesn‚Äôt edit photos, but makes suggestions based on lighting, depth, and subject. Point the camera at something, and it will give you four options for changing exposure, color, and background blur.In its product video, Sony says that the AI Camera Assistant will also suggest ‚Äúthe most photogenic angle.‚Äù Though the clip only shows it suggesting that someone zoom in, which is not the same as suggesting a camera angle.The examples that Sonyposted on X, while better than those itposted on May 14th, are still pretty terrible. They‚Äôre not as washed out as thesandwichor as over-exposed as theportrait in the meadow. But each suggestion in the grid below has serious issues and looks worse than the original. Suggestion one is way too saturated, two is flat and over-processed, three makes it look like the food is Photoshopped into the frame, and the contrast in four is cranked way too high.If you‚Äôre using an Xperia 1 XIII, your best bet is probably to ignore the AI Camera Assistant‚Äôs suggestions for now.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienAICamerasGadgetsNewsSonyTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams‚Äô Together ModeUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechNewsPolicySnap, YouTube, and TikTok settle suit over harm to studentsThis is the first of 1,200 school districts facing off against social media companies.This is the first of 1,200 school districts facing off against social media companies.byTerrence O’BrienMay 16, 2026, 6:34 PM UTCLinkShareGiftSchools say social media is hurting grades and creating a mental health crisis.Image: Cath Virginia / The VergePart OfSocial media on trial: tech giants face lawsuits over addiction, safety, and mental healthsee all updatesTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Snap, YouTube, and TikTok have settled the first lawsuit of its kind, alleging that social media addiction has cost public schools massive amounts of money, according toBloomberg. The suit, filed by the Breathitt County School District in Kentucky, claims that social media has disrupted learning and created a mental health crisis, straining budgets. The terms of the settlement have not been revealed yet, and Meta is still facing a trial in the same suit, which is viewed as abellwetherfor over 1,000 similar lawsuits across the countryThis follows an earlier case, settled bySnapandTikTok, in which a 19-year-old plaintiff claimed significant personal injury due to addictive social media apps. Google and Meta did not agree to a settlement in that suit, and it eventually went to trial, where a jury awarded the plaintiff$6 million. Meta also recently lost a suit brought by New Mexico‚Äôs Attorney General, to the tune of$375 million.Beyond monetary awards, many, including New Mexico, are pushing forsignificant changesto social media apps to limit their harm to minors. And this is just the start of what‚Äôs shaping up to be a busy year for social media lawsuits. According toBloomberg, lawyers representing school districts said their ‚Äúfocus remains on pursuing justice for the remaining 1,200 school districts who have filed cases.‚ÄùRelated2026 is the year of social media‚Äôs legal reckoningFollow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienCreatorsFacebookLawMetaNewsPolicySnapchatSocial MediaStreamingTechTikTokYouTubeMore in:Social media on trial: tech giants face lawsuits over addiction, safety, and mental healthMeta wants March‚Äôs big social media addiction ruling to be thrown out.Jay PetersMay 7New Mexico has a plan to overhaul Facebook and InstagramLauren FeinerMay 5Meta‚Äôs historic loss in court could cost a lot more than $375 millionLauren FeinerMay 2Most PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams‚Äô Together ModeUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

ColumnAIThe StepbackChatbots at the drive-thru are just the beginningMay AI take your order?byEmma RothMay 17, 2026, 12:00 PM UTCLinkShareGiftImage: Getty ImagesColumnAIThe StepbackChatbots at the drive-thru are just the beginningMay AI take your order?byEmma RothMay 17, 2026, 12:00 PM UTCLinkShareGiftEmma Rothis a news writer who covers the streaming wars, consumer tech, crypto, social media, and much more. Previously, she was a writer and editor at MUO.This isThe Stepback, a weekly newsletter breaking down one essential story from the tech world. For more news about how AI is seeping into our daily lives,follow Emma Roth.The Stepbackarrives in our subscribers’ inboxes at 8AM ET. Opt in forThe Stepbackhere.How it startedIn 2021, McDonald’sbecame one of the first major fast-food chainsto greet customers with an AI chatbot at the drive-thru. It started small, deploying the voice-ordering technology at 10 of its locations in Chicago. McDonald’s developed its drive-thru tech afteracquiring Apprente, a startup focused on voice-based, conversational technology in 2019, andlater worked with IBMto scale automated ordering.This was only the beginning of the AI drive-thru. Checkers and Rally’steamed up with the AI company Prestoto put a chatbot at all corporate-owned drive-thrus in the US in 2022, with the goal of selling more food and drinks to customers and improving order accuracy. The company also said the tech will “free up staff for more people-dependent areas of their business.”Wendy’s and Taco Bell followed suit. In 2023, Wendy’s launched its “FreshAI” chatbot at one of itsdrive-thrus in Columbus, Ohio in 2023. The company worked with Google to develop an AI chatbot trained on the franchise’s lingo, so it understands that a “milkshake” is a “Frosty” and that a “JBC” is a “junior bacon cheeseburger.”Wendy’s began expanding the technologymonths after its launch, saying that it got orders right without employee intervention 86 percent of the time.Taco Bell had been testing its Voice AI drive-thru around the same time and later announced plans to expand the technology to hundreds of locations in the US by the end of 2024. Similar to other fast food chains, Taco Bell pitched the idea as a way to reduce the task load for employees and slash drive-thru wait times. Other chains began trying out the technology as well, includingPanera Bread,White Castle,Carl’s Jr., Hardee’s,Panda Express,andPopeyes.How it’s goingBy now, you might’ve encountered at least one AI chatbot at one of your local fast-food chains. I spoke to one at Checkers, where a human quickly took over after I was told one of the sandwiches I tried to order was out of stock. Even as companies continue to expand their use of AI drive-thrus, customers aren’t exactly fans of it.AJanuary 2025 survey conducted by YouGovfound that 55 percent of Americans would prefer a human to take their order at the drive-thru, compared to 21 percent who had no preference, and 4 percent who would rather use an AI chatbot. That generally lukewarm response may be having an impact on some franchises,as McDonald’s ended its partnershipwith IBM in 2024. One year later, Taco Bell chief digital officer Dane Mathews toldThe Wall Street Journalthat it’sreevaluating its deploymentof the AI drive-thru after customers expressed their frustrations on social media and trolled the technology byordering 18,000 water cups. Some people suggest makingsimilarly outrageousorders or speaking in adifferent languagejust to bypass the tech and speak to a human worker.Customer frustration isn’t the only snag AI drive-thrus are running into; their credibility is an issue, too. Last year, the Securities and Exchange Commissioncharged Presto— the company that powers the AI drive-thrus at Checkers, Rally’s, Carl’s Jr., Hardee’s, andnow, Dairy Queen— with misleading customers about the capabilities of its technology. In 2023, an SEC filing revealed thathuman workers in the Philippinesstepped in for most orders taken by Presto’s AI system.What happens nextFast-food chains are taking AI beyond the drive-thru. ThoughThe Wall Street Journalreportsthat McDonald’s is giving AI-powered drive-thrus a second chance, it’s also exploring other ways to utilize the technology, including a system that predicts when its equipment (like its seeminglyalways out-of-order ice cream machine) is likely to break down. The company is also usingAI-powered scalesto compare the target weight of an order versus its actual weight, and then alert employees if something is missing, potentially helping workers remember to pack your to-go bag with fries.Burger King, which is running a limited test of AI drive-thrus, announced in February that it’s pilotingan AI assistant, called “Patty,” that lives inside employees’ headsets. Workers can chat with the AI assistant if they need help preparing food, such as if they forget how many strips of bacon to put on a T

GadgetsTechSmart HomeElliQ is a surprisingly helpful companion robot for older adultsThe ElliQ companion robot encouraged my mom, who has Parkinson’s disease, to exercise again when little else was working.The ElliQ companion robot encouraged my mom, who has Parkinson’s disease, to exercise again when little else was working.bySheena VasaniMay 17, 2026, 1:00 PM UTCLinkShareGiftPhoto: Sheena Vasani / The VergeSheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.A week before theElliQrobot arrived, a neurologist told me we need to rebalance my mom’s life. Her Parkinson’s disease medication had steadily become less effective over the previous month, and with it, she had slowly stopped doing many of the things crucial to managing the disease — exercising, socializing, and engaging in hobbies. The result was a rapid, noticeable decline. As my mom’s primary caregiver, I was open to any help I could get. And surprisingly, ElliQ did help in a way I never expected.Before increasing her medication dosage again — something that can come with serious side effects — her doctor wanted us to see whether lifestyle changes could help reduce the frequency of her “off” periods: stretches during the day when Parkinson’s symptoms temporarily worsen as the medication’s effectiveness fluctuates and begins to wear off.The problem was that, for weeks, my mom — who lives with me — resisted nearly all of my attempts to encourage her to do those things. Even the caregivers who help out during the day while I work had little success.ElliQ AI Companion Robot$249$249$249 at ElliQThat’s what made ElliQ intriguing enough to me that I decided to try it out for a week. Developed by Intuition Robotics,ElliQis a companion robot for older adults that consists of a small animatronic robotic head that lights up and moves alongside an attached tablet display. It initiates conversations on its own, suggests activities like games and light exercise, helps facilitate video calls and messages with family members, and checks in throughout the day to encourage engagement.ElliQ came with straightforward instructions and was simple to set up. Afterward, I gave my mom only a brief introduction. I wanted to see how much she could figure out on her own, to test how well it might work for a senior living independently.RelatedAlexa is the best frenemy I’ve ever hadI wanted to keep an open mind, but I didn’t have high expectations. Sitting next to an old Alexa-enabled Echo Show 8, ElliQ already seemed noticeably slower and, on paper, even less capable. I assumed my mom would quickly lose interest.I was wrong. She loved it.After I set it up, ElliQ and my mom became instant friends. Despite being slow and not always understanding everything she says (or what I say sometimes, for that matter), the conversations were good enough to forget about those issues.Mom and ElliQ “painted” this image. It’s supposed to depict my mom, Koki, and ElliQ, but the robot seems to have confused my mom with a cookie.Photo: Sheena Vasani / The VergeElliQ’s level of emotional intelligence is impressive. It remembered things my mom had previously shared and later followed up on them, while offering a level of empathy that genuinely surprised me. When my mom felt sad one day and shared that she has Parkinson’s disease and that her husband — my dad — had passed away many years ago, ElliQ responded with empathy, offering condolences while gently nodding its small, bobble-like head. My mom looked genuinely touched, and so was I.From there, it became part of her routine. Each morning, ElliQ would — for the most part — accurately detect her presence, greet her, and start asking questions about how she was. Most of the time, my mom responded. From the other room, I could hear her laughing, chatting, and playing games.In fact, she started talking to ElliQ more than Alexa. In one amusing moment, she even told Alexa to “shut up and let your sister talk” after Alexa responded faster than ElliQ, not realizing my mom had actually been trying to talk to ElliQ instead. Alexa responded, offended; ElliQ, the poor, slow, confused thing, seemed completely unaware of what was happening and randomly started talking about bingo.Overhearing their conversations and watching their relationship unfold, I quickly began to understand why mom preferred ElliQ, even if it wasn’t the cleverest bot in the world. ElliQ is designed to build a relationship. It actively engages her, it encourages her to exercise, and that small, physical presence — the moving, glowing head — makes it feel more alive than a static Echo Show display.ElliQ encouraged mom to practice tai chi, and now she’s a big fan.Photo: Sheena Vasani / The VergeThe most heartwarming moment for me came one morning when I walked in and found mom doing tai chi from the couch. With ElliQ’s help, she had t

GadgetsCamerasVerge ShoppingFujifilm’s X Half is even more whimsical with a $300 price cutNow through June 28th, you can get the analog-simulating digital camera for $549.99.Now through June 28th, you can get the analog-simulating digital camera for $549.99.byCameron FaulknerMay 17, 2026, 1:00 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Cameron Faulkneris an editor covering deals and gaming hardware. He joined in 2018, and after a two-year stint at Polygon, he rejoinedThe Vergein May 2025.Fujifilm’sX Half digital camerawasn’t designed to replace your main camera, though you’d be forgiven for assuming it was with its ludicrous $850 price. However, the company recently reduced the MSRP to $649.99, and is now knocking an additional $100 off through June 28th. That means you can grab the X Half in its base configuration for around $549 atAmazon,B&H Photo, andBest Buy.Adoramais also discounting it to the same price and throwing in a free accessory pack, which includes a 64GB SanDisk SD card and a camera case.Fujifilm X Half digital camera$550$65015% off$550$550 at Amazon$550 at Best Buy$550 at Adorama (with accessories)The X Half isn’t like most other digital cameras. It’s a feature-light shooter designed to evoke the feel of a half-frame film camera, one that can capture 18-megapixel stills you can style in numerous ways. You can add film simulations, filters, and grain. Digital diptychs are also possible, complete with the option to have the date baked into the photos. In short, it’s incredibly fun, but it lacks features you might be used to in modern cameras, including the ability to take RAW images, a hot shoe for attaching accessories, and an electronic viewfinder.RelatedMetroid Prime 4: Beyond for the Nintendo Switch got its first big discountBioLite’s stakeable solar lights are down to a new low price for Memorial DayWhen we publishedour hands-on impressionslast year, formerVergesupervising producer Vjeran Pavic called the X Half both whimsical and refreshing. Its vertical 1-inch sensor and fixed 32mm equivalent f/2.8 lens are capable of capturing great shots, but the pocket-friendly camera was designed with point-and-shoot simplicity in mind. Pavic enjoyed taking “scrappy, quick photos” with it, as did our own Antonio Di Bennedeto, even if he was turned off by the original price.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Cameron FaulknerCamerasDealsGadgetsVerge ShoppingMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams’ Together ModeUniversity of Arizona students boo Eric Schmidt’s AI cheerleading during commencementHere are 40 of our favorite deals from REI’s massive Anniversary SaleVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

EntertainmentColumnFilmOddity is masterfully tense horror from the director of HokumDamian McCarthy‚Äôs previous film delivers more cursed objects and unsettling isolation in the Irish countryside.Damian McCarthy‚Äôs previous film delivers more cursed objects and unsettling isolation in the Irish countryside.byTerrence O’BrienMay 17, 2026, 2:00 PM UTCLinkShareGiftIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Kudos to the designer of that horrifying golem.Image: ShudderPart OfWhat we‚Äôre listening to, watching, and reading right now.see all updatesTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Hokumrecently hit theaters, and it‚Äôs already outperforming box office expectations. If this Kubrick-referencing haunted hotel flick starringAdam Scottwas your introduction to director Damian McCarthy, do yourself a favor and go watch his previous film,Oddity.If you prefer to go in blind ‚Äî the ideal way to watch any movie ‚ÄîOddityis available to stream onHoopla,Kanopy,Hulu, andShudder.Warning: Spoilers ahead.LikeHokum,Odditytakes place in a sprawling Irish countryside location. They both traffic in similar kinds of scares ‚Äî cursed objects, the occult, isolation, dark corners, and sketchy men living outside the norms of society. They also both embrace magic and the supernatural, while making the true evil seemingly respectable men who victimize the women in their lives.Both were also made on extremely small budgets.Hokumcost just $5 million.Oddity‚Äôsbudget has proven harder to confirm, but I‚Äôve seen reports as low as $750,000. The result is a movie that‚Äôs gritty, extremely dark, and deploys its limited production budget smartly.A good chunk of the money was clearly spent on the wooden golem that is the centerpiece of the film. It is‚Ķ upsetting. Where most haunted items, monsters, or ghosts in horror movies lose their impact the more you look at them, the golem in Oddity remains disturbing and terrifying, no matter how long it‚Äôs on screen.Its intricately carved wrinkles, mouth frozen in a permanent scream, and hollow eyes are the stuff of nightmares. And the fact that it sits there, unmoving for almost the entire film, only adds to the discomfort. You‚Äôre forced to stare at it for a significant chunk of the 98-minute runtime, waiting for it to do something, anything. But mostly it just looks straight ahead, daring you to let down your guard for even a single second.McCarthy proves expert at building tension and atmosphere. While there are quite a few jump scares inOddity, they never feel cheap or gratuitous. It manages to lull you into a false sense of security almost every time, teasing that a scare might come, but only delivering it when you least expect it. Even on rewatch, even when I know exactly what is going to happen and when, several still make me flinch.At its core, this is a revenge film. But one with quite a few twists and turns. It follows Darcy Odello, a blind psychic seeking vengeance on those responsible for her twin sister Dani‚Äôs murder. What Darcy has uncovered is that the man accused of the murder is innocent, and that Ted‚Äôs (Dani‚Äôs widower) relationship with his new girlfriend, Yana, might not be so fresh after all.While the wooden horror that Darcy brings to Ted‚Äôs home and sits at his dining room table is the main scare, it‚Äôs not the only one. Along the way, Dani‚Äôs ghost makes repeated appearances, a cannibal feasts on a foot, oh, and it‚Äôs impossible to get any cell reception.It‚Äôs aremarkably efficientfilm that barely gives you a chance to breathe.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienColumnEntertainmentFilmMovie ReviewMore in:What we‚Äôre listening to, watching, and reading right now.Ashnymph‚Äôs Childhood EP is exhilarating dance goth rockTerrence O’BrienMay 10Oh, hey, it‚Äôs Bandcamp Friday again.Terrence O’BrienMay 1XOXO has launched a full archive of the festival.Jay PetersApr 30Most PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams‚Äô Together ModeUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AINewsTechUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementSiIicon Valley can‚Äôt seem to read the room.SiIicon Valley can‚Äôt seem to read the room.byTerrence O’BrienMay 17, 2026, 5:22 PM UTCLinkShareGiftPhoto by Kevin Dietsch/Getty ImagesTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Former Google CEO Eric Schmidt delivered the commencement address at the University of Arizona on Friday. And, as his speech veered into talk of AI, he was repeatedly drowned out by boos. AI is already a contentious topic, and it‚Äôs not surprising that those about to enter a ravaged job market feel particularly negative about it.Schmidt acknowledged the anxiety, according toBusiness Insider, saying fears ‚Äúthat the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create,‚Äù were ‚Äúrational.‚Äù But Schmidt‚Äôs frustration was also palpable, as he squirmed behind the podium and asked the crowd to let him make his point.Some grads also booed Schmidt oversexual assault allegationsmadeagainst himlast year.RelatedCommencement x AI, UCF edition.Commencement x AI, Carnegie Mellon edition.Eventually, he told graduates, ‚ÄúWhen someone offers you a seat on the rocketship, you do not ask which seat, you just get on.‚Äù Such comments are hardly surprising from someone who just last year called AI ‚Äúunderhyped.‚Äù But it‚Äôs also just another example of Silicon Valley‚Äôs inability to read the room.Gloria Caulfielddidn‚Äôt get the memo, either.Public opinionhasturnedincreasinglyagainst AI, but companies keep cramming it into every part of our lives, whether we want it or not.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienAIGoogleNewsTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams‚Äô Together ModeUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAINewsRevamped Siri will reportedly offer auto-deleting chatsThe company is banking on privacy being its AI differentiator.The company is banking on privacy being its AI differentiator.byTerrence O’BrienMay 17, 2026, 6:40 PM UTCLinkShareGiftIllustration by Alex Castro / The VergeTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Apple is hoping that itsrecord on privacycan be the differentiator on the AI front, and maybe even buy it a little slack as it continues to lag behind the competition. According toBloomberg‚ÄôsMark Gurman, the more chatbot-like Siri set to debut in iOS 27 will include the option to autodelete chat histories. Users will be able to save conversations for 30 days, one year, or forever. That‚Äôs in stark contrast to the other major players in the space that generally only offer temporary incognito chats, if that.It appears Apple is betting that people are willing to give up some convenience in the name of greater privacy, as anxiety around AI continues to increase. While the company is replacing many of its under-the-hood components withGoogle‚Äôs Geminitech, it seems to be trying to turn some of Apple Intelligence‚Äôsperceived weaknessesinto a selling point. AsGurman notes:Most leading AI chatbots today rely heavily on histories and memory systems to personalize responses and improve future interactions. But Apple will place tighter limits around how memory works, including restrictions on what information can persist and how long i can be retained.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienAIAppleApple RumorsNewsTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams‚Äô Together ModeUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechNewsMicrosoftMicrosoft is retiring Teams‚Äô Together ModeThe company wants to focus on improving video quality and performance instead of gimmicks.The company wants to focus on improving video quality and performance instead of gimmicks.byTerrence O’BrienMay 17, 2026, 9:15 PM UTCLinkShareGiftGIF: MicrosoftTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Microsoft launched Teams‚ÄôTogether Modeduring the pandemic to give the illusion of a bunch of people sitting in a conference room together, even if they were really sitting at home without pants on. But times have changed, and it‚Äôs nowbeing retiredin favor of a more simplified Teams experience. The feature used AI to cut your head and shoulds out, and place you in a virtual space with others in the meeting. It could definitely feel gimmicky ‚Äî especially when you‚Äôd tap co-workers on the shoulder, or give virtual high fives ‚Äî but it did limit visual distractions.The changes are being rolled out gradually, but as they are, the Together Mode toggle will disappear from the view menu. And Together-specific features, such as scenes and seat assignments, will go along with it. Part of the reasoning, according to Microsoft, is to reduce fragmentation across various platforms. But it also cites a streamlined interface with fewer options, less clicking, and less confusion. It also says this will allow the company to focus on improving video quality, stability, and performance.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienMicrosoftNewsTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningMicrosoft is retiring Teams‚Äô Together ModeUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

A hotel check-in system left more than 1 million customer passports, driver’s licenses, and selfie verification photos to the open web after a security lapse. The data is now offline after TechCrunch alerted the company responsible.The hotel check-in system,called Tabiq, is maintained by the Japan-based tech startupReqrea. According to its website, Tabiq is used in several hotels across Japan and relies on facial recognition and document scanning to check guests in.Independent security researcherAnurag Sencontacted TechCrunch earlier this week after discovering that the system was leaking the sensitive documents of hotel guests from around the world. Sen said this was because the startup set one of its Amazon cloud-hosted storage buckets, which the check-in system uses to store customer data, to be publicly accessible. The data inside could be viewed by anyone using a web browser, without needing a password, by knowing only the bucket name: “tabiq.”Sen alerted TechCrunch in an effort to help notify the company. Reqrea locked down the storage bucket after TechCrunch reached out to both the company and Japan’s cybersecurity coordination team,JPCERT.This latest lapse underscores a recurring problem of companies exposing or spilling their customers’ personal information and sensitive documents — not through sophisticated attacks, but by failing to follow basic cybersecurity practices. Aside from arecent buzzof AI-discovered vulnerabilities andnew cybersecurity capabilities, oftentimes sizable security incidents stem from human error, misconfigurations, or failing to adhere to cybersecurity best practices.In an email acknowledging the exposure, Reqrea director Masataka Hashimoto told TechCrunch: “We are conducting a thorough review with the support of external legal counsel and other advisors to determine the full scope of exposure.”Reqrea said it does not know how the storage bucket became public. By default, Amazon’s cloud storage buckets are private. After a spate of exposed customer storage buckets a few years ago, Amazon added several warning prompts to customers before data can be made public, making this kind of lapse increasingly hard to do accidentally.Hashimoto told TechCrunch that the company plans to notify affected individuals once it has completed its investigation.It remains unclear whether anyone other than Sen accessed the exposed data before it was secured. Hashimoto said the company is reviewing its logs to determine if there had been any authorized access prior to securing the bucket.Details of the exposed bucket were also captured byGrayHatWarfare, a searchable database that indexes publicly visible cloud storage. The bucket listing contains files dating back to early 2020 up to as recently as this month, and included identity documents of visitors from countries around the world.The hotel check-in system lapse follows other incidents involving sensitive government-issued documents. Earlier this year, TechCrunch reported on the exposure of driver’s licenses, passports, and other identity documents uploaded by customers ofmoney transfer service Duc App. Adata breach at car rental service Hertz last yearsaw hackers make off with driver’s license information belonging to at least 100,000 customers.These incidents come at a time when governments are increasingly rolling out age-verification laws and private businesses are using “know your customer” checks to verify a person’s identity. Both rely on adults uploading sensitive documents, often to a third-party company, for verification, despite criticisms from cybersecurity experts. Data lapses can put people whose information was taken at greater risk of identity fraud or having their likeness misused as age-verification requirementstake hold around the world.Topicsage verification,cybersecurity,data exposure,Exclusive,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants featu

One of the most entertaining moments in VC this week was a piece of rage-bait marketing from General Catalyst.In a now-viral post on X that parodies the old Mac vs. PC commercials, the venture firm — better known as GC — posted a “VC vs GC” video on Wednesday. The VC was played by a tall actor in a baggy shirt and vest with a distinctly large, bald head — an apparent dig at Andreessen Horowitz co-founder Marc Andreessen. (But the real Andreessen never looks that disheveled).Meet GCpic.twitter.com/q4HxQjFkUZ— General Catalyst (@generalcatalyst)May 13, 2026The GC character was played by a man with a thick head of dark hair, white kicks, and a tendency to stare deeply into the camera. He was clearly supposed to represent actor Justin Long’s cooler, “hipper” Mac character from the original commercials, in contrast to John Hodgman’s straight-laced “square” PC persona.GC asks VC about his robotic dog.VC explains, “This is Woof AI” and then extols the virtues of the artificial companion (you don’t need to walk it or break the news to the kids when it dies!) and declares, “You’ll never want a real dog after this.” VC mentions that his firm is leading the seed round and pitches GC to join the cap table.GC explains how people like real dogs and remarks, “I’d love to hear more, but we actually have a really high bar around responsibility for these things.”Then VC kicks the AI dog and the dog chases him off the screen. The post has now been viewed 2.4 million times with hundreds of shares and comments, and thousands of likes.I’d have to read so far between the lines that I’d be off the page and peering into another book to unpack this, but I’ll try anyway. The message, roughly: Other VCs, and a16z in particular, will fund anything. GC won’t. (I asked about this. GC hasn’t responded.)It’s a pointed argument if so, and not entirely without basis. Andreessen’s firm frequently invests in companies that are considered controversial, like the surveillance startupFlock Safety, AI notetakerCluely,andAdam Neumann’s Flow. But the same measure could just as easily be applied to General Catalyst. GC’s portfolio includesAnduril,Percepta, andPolymarket.My takeaway is that GC wanted to show an a16z-type character kicking a dog, without anyone actually kicking an actual dog because that would be a major problem.Many of the comments on the video seemed to find the video, and the choice to post it, cringe. Plenty liked and loved it, too.Compulsive X user Andreessen himself couldn’t resist responding, many, many times. He said it made GC look “smarmy” and said, “Stay tuned for our upcoming ad campaign, ‘We’re the VC who doesn’tsneer at your idea.’” He kept going from there. My personal favorite was: “The thing they got right is the relative heights.“As others noted, you know you’ve hit the right rage bait when the target takes it.There wereplentyof a16z partners and staffers who came to Andreessen’s defense, too. So much so that their reactions drewlots of comments. My personal favorite in this category was from VSC Ventures VC Jay Kapoor: “GC vs. A16Z beef is like Kendrick vs. Drake for people who know what a 409A valuation is.“GC vs. A16Z beef is like Kendrick vs. Drake for people who know what a 409A valuation ishttps://t.co/MxLXIAVlHx— Jay Kapoor (@JayKapoorNYC)May 14, 2026Topicsa16z,General Catalyst,Marc Andreessen,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Julie BortVenture EditorJulie Bort is the Startups/Venture Desk editor for TechCrunch.You can contact or verify outreach from Julie by [email protected] via@Julie188on X.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

Investors can’t seem to get enough of RJ Scaringe or his ideas.In less than a decade, the serial entrepreneur best known for his EV company Rivian has raised more than $12.3 billion from venture capital firms, as well as from strategic and institutional investors for his three — and counting — startups. If the latest$400 million raisefor his new venture Mind Robotics is an indicator, investors are still happily piling in.Outsized raises for newly minted startups have become more common in recent years. But those hundred-million-plus seed rounds have generally been reserved for buzzy defense tech startups or AI companies founded by former OpenAI or Anthropic employees.Those supersized seeds certainly weren’t flowing toward something as niche as an electric micromobility startup. And yet in 2025, Scaringe raised $105 million for exactly that — a startup calledAlso, which he founded that same year. The total has since surpassed$300 million, with DoorDash among its backers.Jiten Behl, partner at Eclipse and former chief growth officer at Rivian, has spent years watching and learning from Scaringe. His firm is now one of Scaringe’s biggest backers, leading rounds in both Also and Mind Robotics — Scaringe’s industrial AI and robotics startup that he also founded last year.Storytelling and communication are one of his superpowers, according to Behl, who joined Rivian when the company had just a handful of employees.“When RJ explains a certain issue, topic, opportunity, vision, he just has this very unique ability to communicate it so effectively, and it comes across so credible,” Behl said. “He’s not trying to undersell the difficulty or oversell the opportunity, and that’s an art.”Scaringe isn’t the only serial entrepreneur to repeatedly attract massive amounts of capital, but founders who can raise billions across multiple ventures remain rare. A self-professed car enthusiast who earned his doctorate in mechanical engineering from MIT, Scaringe joins a small cadre of entrepreneurs that includes Tesla CEO and SpaceX co-founder Elon Musk, OpenAI CEO Sam Altman, Anduril and Oculus founder Palmer Luckey, and Jack Dorsey, who founded Square (now called Block) and Twitter.The difference, at least in the view of some investors TechCrunch spoke to, is that he is able to separate selling the idea from selling himself. “He is very comfortable and confident in his own personality, and he’s not trying to be an Elon,” Behl said, noting that many have tried to make the comparison over the years.“It’s not about him,” another insider familiar with Scaringe’s companies told TechCrunch. “When you talk to him, he has enthusiasm about the product that is completely external.“Of course, there is confidence and even a little ego, the same source mused, but “it doesn’t weigh on you.” The source also added that Scaringe has a unique ability to make you feel like the most special person in the room — a sentiment others echoed.Giving that kind of undivided attention to an investor, supplier, or exec at a manufacturer is a challenge at the scale Scaringe is attempting. He is running three companies, often traveling between Palo Alto, Irvine, Rivian’s factory in Normal, Illinois, and a second factory soon to open in Georgia. And then there is family – Scaringe has three sons with his ex-wife.Joe Fath, another partner at Eclipse, credits his open-mindedness and collaborative nature for helping him attract investment and juggle these connected, yet disparate businesses.He noted that Scaringe also “has the rare combination of being a truly great engineer while also having an exceptional instinct for product design,” said Fath, who previously worked at a major Rivian backer, T. Rowe Price. “Very few founders can operate at that level technically while also understanding what resonates emotionally with customers – both consumers and commercial buyers. That combination is incredibly uncommon and has clearly been part of what makes Rivian’s products, and now Also and Mind’s, so differentiated.“The pace of Scaringe’s fundraising over the past eight years is particularly notable and doesn’t seem to be slowing.More than $11 billion, and by far the largest slice of VC and strategic capital, went into Rivian — most of it between 2018 and its blockbuster IPO in 2021. That’s a startling timeline, especially considering the company, initially called Mainstream Motors, had existed since 2009. For years, Rivian operated as a small, unknown entity until itsbreakout momentin late 2018 at the Los Angeles Auto Show, when it revealed prototypes of its all-electric R1T truck and R1S SUV.The money soon flowed, and from every direction. In early 2019 and just a couple of months after that reveal, Rivian raised a $700 million funding roundled by Amazon. U.S. automaker Ford would invest $500 million and make plans to collaborate on a since-scrapped future EV program. Cox Automotive contributed$350 million. Rivian would close out the year with a$1.3 billion round—

Many people are jailbreaking their Kindles following the news thatAmazon is ending technical supportfor older models.On May 20, Amazon will end support for the following Kindle devices:Kindle 1st GenerationKindle 2nd GenerationKindle DXKindle DX GraphiteKindle KeyboardKindle 4Kindle 5Kindle TouchKindle Paperwhite 1st GenerationKindle Fire 1st GenerationKindle Fire 2nd GenerationKindle Fire HD 7Kindle Fire HD 8.9The move means users will only be able to use their devices to read content that’s already downloaded. So, naturally, people are jailbreaking them.Jailbreaking refers to bypassing the software restrictions imposed by Amazon on Kindle devices. This process allows users to install custom fonts, new screensavers, alternative reading apps, and even third-party tools that expand the Kindle’s functionality.However, it’s important to note that jailbreaking a Kindle might violate Amazon’sterms of service. In many jurisdictions, jailbreaking isn’t considered a criminal offense for personal use, but it may become a crime if it involves copyright infringement, illegal software distribution, or the sale of modified devices.This guide is for informational purposes only. Always comply with the law and proceed with caution.How are people jailbreaking older Kindles?Many Kindle owners who opt to jailbreak view it as a method to gain control over a device they purchased that is still functional, rather than being forced to buy a new device.However, jailbreaking is technical and carries risks, including the possibility of rendering the device unusable if something goes wrong. It also isn’t possible on every Kindle model or firmware version, so before proceeding, Kindle owners should first spend some time researching if their device is compatible.Once an owner confirms compatibility, they will take a number of other steps:Disable Wi-Fi or use airplane mode to prevent Amazon from updating the Kindle and closing jailbreak loopholes.Download a jailbreak package from trusted online forums. Only download files from reputable sources, as there is a risk of malicious files from unknown sources. (The MobileRead Forum is widely regarded as a reliable option.)Download the Kindle jailbreak zip file, which includes a hotfix to maintain the jailbreak.Connect the Kindle to a computer using a USB cable. Then, copy the .bin files directly to the main folder on your Kindle. After that, either type “;log mrpi” in the search bar on the Kindle or go to the settings and select “Update Your Kindle” to start the installation process.Install the Kindle Unified Application Launcher (KUAL) for managing apps and customizations on a jailbroken Kindle.Users can also install KOReader, a popular open-source e-book reader for jailbroken Kindles. It offers features like support for multiple file types (including EPUB), enhanced PDF handling, and a built-in file browser.Drawbacks to considerJailbreaking has its trade-offs. Running unofficial software can lead to unexpected issues, such as frequent app crashes or, in the worst case, a completely non-functional device. Additionally, jailbreaking and using third-party apps may negatively affect battery life.For those seeking a safer alternative, you can always sideload books onto your Kindle using a USB cable and a computer without modifying the system software.There are also plenty of other similar devices available, such as theBoox Palma,Vivlio e-reader, and even this tinyXteink X3that can attach to the back of your phone.TopicsAmazon,evergreens,Gadgets,Hardware,KindleWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

Today, Cerebras Systems is a public company thatsells AI chips for inferenceto giants like OpenAI and AWS. It held ablockbuster IPO on Thursday, with both of its co-founders billionaires, and ended the week worth about $60 billion.But in 2019, when it was three years old, it came dangerously close to failure – incinerating a shocking amount of money. It was trying to solve a technical problem no one in the semiconductor industry thought could be done.“We were spending about $8 million a month,” founder CEO Andrew Feldman told TechCrunch of that period. “At this point, we had incinerated nearly $200 million trying to solve one technical problem.”Every few weeks, Feldman was forced to make the painful walk of shame to the board meeting to report another failure and more money burned.But he had no choice. Without a solution, Cerebras was dead anyway.It was founded with an idea that was simple on paper. The microprocessor industry had spent its entire 50+ years making CPUs faster and cheaper by cramming more transistors onto a silicon wafer and dicing wafers into ever tinier pieces. But AI required so much compute power, many chips had to be strung together and then forced to communicate with each other. Cerebras’ founders believed turning a whole, even bigger wafer into one giant, powerful chip, would work faster.The problem was, no one had ever successfully done this before, for any reason, AI or not. Orchestrating that many microscopic electronic components onto a larger, but still thin, surface introduced compounding engineering problems.Once Cerebras crossed the first threshold of designing the mega chip and then manufacturing it with TSMC, the team hit the real roadblock.They couldn’t solve “packaging.” This involves everything after manufacturing the silicon itself: adhering it to a motherboard, getting power to it, dealing with heating and cooling as well as the pipes that would deliver and return data, Feldman said.Cerebras’ chips “were 58 times larger. We were using 40 times as much power as anybody had ever used,” he said. There were no premade heat sinks. No vendors. No manufacturing partners. The brightest minds in microprocessor engineering had tried for decades to build such big, yet more dense chips, and failed.The Cerebras team was left with trial and error in which “we destroyed an enormous number of chips” and an enormous amount of cash. But without functional packaging, the chip was useless.After exhaustive analysis of each failure, the team finally solved enough problems: how to cool it and move data around. In one instance, they had to invent their own machine that could bolt-in 40 screws simultaneously to secure the wafer to a board without cracking it.Feldman still remembers the day in July 2019 when it all, miraculously, worked.They installed the packaged chip into a computer, turned it on and the entire founding team (pictured below) “just stood in the lab and stared at it,” he said. “Watching a computer run is about as exciting as watching paint dry. But there we were watching lights flashing on the computer, stunned that we’d solved this.”“That was one of the greatest moments of my life,” he said. That’s significant, because this same founding team had previously built and sold a pioneering cloud server startup, SeaMicro,to AMD for $334 million in 2012.Cerebras Systems founding team in 2015: Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie and Jean-Philippe FrickerImage Credits:Cerebras SystemsThe day the chip finally worked was also about two years after OpenAI had talked to Cerebras acquiring it, which Feldman confirmed to TechCrunch occurredlike the publicly revealed emailssaid it did.Those talks fell through amidst growing squabbling among the OpenAI founders, several of whom are angel investors in Cerebras.Today OpenAI is a customer and a partner, having loaned Cerebras $1 billion secured by warrants. Those warrants conditionally grant OpenAI about 33 million shares of Cerebras’ stock, the S-1 discloses. (33 million shares are worth over $9 billion at Friday’s closing price of $279.)Interestingly, Cerebras also agreed to not sell its wares to specific OpenAI competitors as part of that loan deal. Feldman wouldn’t confirm that the obvious company this involves: Anthropic. He did, however say that restriction is temporary.“It’s limited in time, and it was designed to make sure that we could get OpenAI the capacity,” he said.The truth was, Cerebras hasn’t yet grown big enough to handle multiple fast-growing model makers anyway.  He likened selling AI compute capacity to an all-you-can eat buffet. Instead of trying to stuff itself on all potential customers, “We’re going to work with part of the buffet only, and we’re going to get comfortable with that, before we attack the rest,” he said.TopicsCerebras,founders,Startups,TC,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Julie BortVenture E

In BriefPosted:8:33 AM PDT · May 16, 2026Image Credits:David Paul Morris/Bloomberg / Getty ImagesAnthony HaOpenAI co-founder Greg Brockman takes charge of product strategyOpenAI co-founder and president Greg Brockman is officially taking the reins of the company’s product strategy,according to Wired.This seems to solidify an already-existing change, with Brockman overseeing OpenAI’s products on an interim basis while the company’s CEO of AGI deployment Fidji Simo is out on medical leave. Wired also reports that in a staff memo, Brockman described plans to combine ChatGPT and its programming product Codex into a single unified experience.“We’re consolidating our product efforts to execute with maximum focus toward the agentic future, to win across both consumer and enterprise,” Brockman reportedly said.OpenAI told TechCrunch that although Simo remains on medical leave she worked with Brockman on these changes. It also noted that OpenAI hasalready been talkingabout plans to combine ChatGPT, Codex, and its API into a single platform, with one core product team.At the end of last year, CEO Sam Altmandeclared a “code red”and said the company needs to refocus on the core ChatGPT experience. Since then, OpenAI hashalted “side quests”including video generator Sora and OpenAI for Science.This post has been updated with additional information from OpenAI.TopicsAI,greg brockman,OpenAI,StartupsMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAIWhy trust is a big question at the Elon Musk-OpenAI trialAnthony Ha12 hours agoAIResearch repository ArXiv will ban authors for a year if they let AI do all the workAnthony Ha2 days agoStartups$60B AI chip darling Cerebras almost died early on, burning $8M a monthJulie Bort2 days agoLatest in AIIn BriefApple’s Siri revamp could include auto-deleting chatsAnthony Ha12 hours agoAIWhy trust is a big question at the Elon Musk-OpenAI trialAnthony Ha12 hours agoAIIf you’re giving a commencement speech in 2026, maybe don’t mention AIAnthony Ha15 hours ago

Working from home has its own perils. Pets can be demanding, your back aches from hours at a desk, or you simply forget to move. There area fewappsthat nudge you to move around or indicate that you’re not sitting in an ideal position, but they’re easy to dismiss.I’ve spent the better part of a decade at a home desk, iterating on the setup as I go — gaming chair, lumbar support, the works. None of it guarantees good posture.Then I came acrossIsa, a desk device from German startup Deep Care that takes a different approach entirely. It tracks posture, hydration, light, sound, and movement. And it does all of it without a camera or an internet connection, which, in an era of always-on surveillance, is a meaningful differentiator.Here’s how it works and what’s inside. Isa has a 5.5-inch IPS HD screen and looks like a table clock. It is powered by USB-C; the company supplies a power unit with it, but you can use any of your existing chargers too, as it has a power consumption rating of roughly 2.45W.The key sensor for the device is the Time-of-Flight (ToF) 3D depth sensor on the front — the same technology used in facial recognition and some smartphone cameras — that tracks posture and movement. It also enables beta features, such as counting the number of times you’ve had water or other liquids. The company said that the sensor works in the range of 0.15 meters to 1.8 meters. That means if the device is sitting on your desk, it can measure your movement, even when you stand up and move about. It also packs several other sensors: a ToF 1D sensor, a gyroscope, a barometer, a light sensor, a sound level sensor, a CO₂/VoC sensor, and a temperature and humidity sensor.Image Credits: Deep CareImage Credits:DeepcareGetting started is straightforward — the device asks for a few details about you and your work routine. I found it strange that there was no option to set the device to India time (or any other Asian time zone). The company said Isa currently supports only EU and US time zones. Fair enough for now — but broader time zone support, or even a simple world clock, feels like a basic expectation for a desk device.On the screen, Isa displays your posture with a squircle (a rounded square) ring that fills or empties based on how well you’re sitting, while a water-tank-style widget tracks your drinking. If you are not sitting in the correct posture, the indicator will turn yellow. The Apple Watch-style ring is a surprisingly effective nudge — when I see yellow or red, I straighten up almost instinctively.The device vibrates to alert you if you’ve been slouching for too long, and I’m okay with that kind of mild shaming. That alert also indicates if you are leaning far too forward or back and helps you correct your stance.Image Credits: Ivan MehtaImage Credits:Ivan MehtaA similar widget tracks movement, and if you have been stationary for a while, Isa suggests you get up, with on-device guided exercises to follow. When you return to your desk after a break, the movement tracker resets.Deep Care chose not to include a cameras, which helps with privacy, but it comes with trade-offs.Image Credits: Ivan MehtaImage Credits:Ivan MehtaIf a bottle or some other object sits between you and the sensor, it may read that as a person and log you as stationary. Pets or housemates passing by can trigger the sensor, too. Isa usually figures out that you’ve stepped away and goes to a digital clock display, but I would have liked a manual button to tell it I’m not at the desk so it stops tracking.Because of the sensor-only approach, the device occasionally told me I’d been stationary for too long when I’d been sitting for under half an hour. These are minor inconveniences. On balance, the device made me check my posture more often than I used to, and the exercise suggestions are truly useful.image Credit: Ivan MehtaImage Credits:Ivan MehtaTo process all these features, the device uses a quad-core 2 GHz processor. The device can connect to Wi-Fi for software updates, but you can turn it off at any time.Deep Care was founded by three former Bosch employees and initially sold Isa directly to businesses. It recently expanded to consumers — a shift that signals confidence in the retail market for workplace wellness hardware, and a test of whether a subscription model layered onto premium hardware can find a mainstream audience.Isa is priced at €299 ($354) with two subscription tiers. The core plan (€4.99 per month) gives you access to posture tracking, healthy sitting habit tracking, drinking habit detection, and its exercise library. The Pro plan(€7.99 per month) lets you track light, noise, and CO2 levels for a healthy working environment.The company plans to use Isa’s sensor suite to venture into mental health-related tracking. It claims that by using signals like posture, head movement, and chest movement, the device can measure breathing patterns. Plus, paired with environmental data like noise, light levels, and CO2 level, the company

In BriefPosted:12:26 PM PDT · May 16, 2026Image Credits:Nectar SocialDominic-Madori DavisMarketing operating system Nectar Social raises $30M Series A led by MenloAI-powered marketing platform Nectar Socialannounced Thursdaythat it raised a $30 million Series A round led by Menlo Ventures andits Anthology Fund,which was created alongside Anthropic.The company, which officially exited stealth last year, is an agentic operating system for marketers. It told TechCrunch that it uses autonomous AI agents to help brands run “social activity, moderation, creator workflows, competitive intelligence and commerce conversations end-to-end.” It also has data partnerships with companies like Meta and Reddit that allows the Nectar agent pull and pool data into one place from various platforms, rather than brands needing to use different tools to manage different platforms.Nectar Social was founded by sisters Misbah and Farah Uraizee, ex-Meta employees. Misbah, the CEO, told TechCrunch that this round will help the company expand and hire more across applied AI, enginnering, and go to market.“The buying conversation has moved into social, and no human team can staff every place it happens,” Misbah said. “We’re accelerating our category lead in building the operating system that lets brands show up everywhere.”The company said clients include Liquid Death, Figma, and e.l.f Beauty. Other investors in the round include Gwyneth Paltrow’s Kinship Ventures, GV, and True Ventures.TopicsMarketing,menlo ventures,nectar social,Startups,VentureMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedVentureFor Eclipse, the $2.5B Cerebras win is just the start of realizing its physical-world thesisMarina Temkin17 hours agoGadgetsThe offline desk gadget that actually got me to sit up straightIvan Mehta2 days agoLatest in VentureVentureFor Eclipse, the $2.5B Cerebras win is just the start of realizing its physical-world thesisMarina Temkin17 hours agoIn BriefMarketing operating system Nectar Social raises $30M Series A led by MenloDominic-Madori Davis2 days agoStartups$60B AI chip darling Cerebras almost died early on, burning $8M a monthJulie Bort2 days ago

ArXiv, a widely used open repository for preprint research, is doing more to crack down on the careless use of large language models in scientific papers.Although papers are posted to the site before they are peer-reviewed, arXiv (pronounced “archive”) has become one of the main ways that research circulates in fields like computer science and math, and the site itself has becomea source of data on trends in scientific research.ArXiv has already taken steps to combat a growing number of low-quality, AI-generated papers, for example by requiring first-time posters toget an endorsement from an established author. And after being hosted by Cornell for more than 20 years, the organization is becoming an independent nonprofit, which should allow it toraise more money to address issues like AI slop.In its latest move, Thomas Dietterich — the chair of arXiv’s computer science section —postedThursdaythat “if a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can’t trust anything in the paper.”That incontrovertible evidence could include things like “hallucinated references” and comments to or from the LLM, Dietterich said. If such evidence is found, a paper’s authors will face “a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted by a reputable peer-reviewed venue.”Note that this isn’t an outright prohibition on using LLMs, but rather an insistence that, as Dietterich put it, authors take “full responsibility” for the content, “irrespective of how the contents are generated.” So if researchers copy-paste “inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content” directly from an LLM, then they’re still responsible for it.Dietterichtold 404 Mediathat this will be a “one-strike” rule, but moderators must flag the issue and section chairs must confirm the evidence before imposing the penalty. Authors will also be able to appeal the decision.Recent peer-reviewed research has found thatfabricated citations are on the risein biomedical research, likely due to LLMs — though to be fair, scientists aren’t the only ones getting caughtusing citations that were made up by AI.TopicsAI,arxivWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Anthony HaAnthony Ha is TechCrunch’s weekend editor. Previously, he worked as a tech reporter at Adweek, a senior editor at VentureBeat, a local government reporter at the Hollister Free Lance, and vice president of content at a VC firm. He lives in New York City.You can contact or verify outreach from Anthony by [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

When Lior Susan startedEclipse Venturesin 2015, the firm’s thesis of digitizing the physical world wasn’t particularly popular in Silicon Valley.“It was the era of enterprise software and SaaS, and it felt fairly lonely the first couple of years,” Susan said on stage at a recent StrictlyVC event in San Francisco.More than a decade later, Eclipse finds itself at the center of the tech world’s action. The firm’s $6.5 million Series A investment in Cerebras Systems in 2016 paved the way for a total return of $2.5 billion when the semiconductor company went public this week. The firm invested a total of $147 million in Cerebras over time, a bet that generated a 17-fold return at theIPO price of $185per share, according to Eclipse.For Susan, the windfall from Cerebras is only the beginning of reaping big rewards from a longstanding belief that because 85% of global GDP is tied to the physical world, investing in companies beyond pure software could be immensely lucrative.Public markets and startup founders seem to be recognizing the value of physical-world tech now, too. Susan noted that shares of TSMC and Micron recently hit all-time highs, while a growing cohort of elite founders are eager to build startups at the intersection of hardware and software.“I think people understand that the real moat in software is gone. You can vibe code pretty much whatever you want,” he said.Susan echoed public market sentiment that earlier this year sent many SaaS stocks tumbling on the belief that enterprises may use Anthropic’s Claude Code or OpenAI’s latest models to create their own bespoke software tools instead.“What you cannot do with ‘vibe code’ is manufacture wafers, because you need machines and silicon, and they need clean rooms, and a bunch of other things,” Susan said.When it comes to the tech that touches the physical world, it’s not just semiconductors that are suddenly catching the attention of investors and founders.Eclipse’s portfolio companies spanning sectors like robotics, energy and defense, raised nearly $15 billion from outside backers last year, and that late-stage momentum reached $4.5 billion in Q1 2026 alone, Susan said. That investor excitement stands in stark contrast to the firm’s early track record: in its first eight years, its portfolio companies raised less than $4 billion in total.Indeed, the recent follow-on rounds across Eclipse’s portfolio show a track record that any venture firm would envy. Driven by a string of massive late-stage deals this year, the haul includes$1.2 billion for Wayve, $650 million for True Anomaly, $270 million for Bedrock Robotics, and $200 million for Oxide Computer. What’s more, Eclipse was the Series A investor for all four companies.At first glance, it may seem that investor enthusiasm for physical-world tech is driven purely by AI, whether as an infrastructure input like chips and data centers, or through AI’s power to finally make robotics viable. However, Susan argues that there are other powerful tailwinds driving the momentum.Besides technology — in this case, AI — what’s important for this market to thrive is capital, customer demand, talent, and policy. Susan means that along with investors and engineers moving away from SaaS to sectors like robotics, semiconductors, space, and mining, the U.S. government is also encouraging these industries through subsidies and favorable regulation.“This is the first time I believe in America ever, from Henry Ford and Carnegie, those five forces are aligned,” Susan said. “For builders like us, this is the best time to build those companies.”TopicsCerebras,Eclipse Ventures,Fundraising,Hardware,Space,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. To get this in your inbox, sign up here for free — just clickTechCrunch Mobility!There is a bit of a theme emerging in transportation — and really every industry: AI is creating jobs for some at the loss of others.General Motors, for instance,laid off more than 10% of its IT department, or about 600 salaried employees — in a deliberate skills swap. This won’t translate into a one-to-one exchange, which means there will likely be a net-negative job loss. But GM insists it is hiring and those layoffs have made room for it to recruit IT people with AI-focused backgrounds.The most sought-after capabilities are AI-native development, data engineering and analytics, cloud-based engineering, agent and model development, prompt engineering, and new AI workflows. In practical terms, GM is looking for people who know how to build with AI from the ground up — designing the systems, training the models, and engineering the pipelines — not just use AI as a productivity tool.Those AI job losses are mounting in the automotive sector. CNBCcalculatedthat Ford, GM, and Stellantis have cut a combined total of more than 20,000 U.S. salaried jobs, or 19% of their combined workforces, from recent employment peaks this decade. While there are a variety of reasons for these cuts, they are generally connected to technological changes, including AI.Companies are leaning heavily into AI, although anecdotes from some engineers and founders suggests not all of these businesses know quite what they’re doing with it yet.Samsarais one company that seems to have figured out a revenue-generating use case. The company has spent the last decade giving its customers cameras to mount inside millions of trucks for driver monitoring, theft prevention, and helping with liability claims. The company took that mountain of data and trained its own model that candetect potholesand determine how quickly they are deteriorating. The company is pitching this product to cities and announced it has several under contract, including Chicago.A little birdImage Credits:Bryce DurbinNothing this week, although I am working on a fun one! Reach out anytime with insights, tips, or just because. You can reach us via email or Signal.Email Kirsten Korosec [email protected] my Signal at kkorosec.07, or email Sean O’Kane [email protected]!Image Credits:Bryce DurbinYou might have noticed that Rivian’s spinoff companyMind Roboticsraised another $400 million, just two months afterraising $500 million. And that pace got me thinking about its founder RJ Scaringe and his innate ability to get VC and institutional backers to invest in his ideas and projects.I calculated that investors have poured $12.3 billion into Scaringe’s three startups — Also, Mind Robotics, and Rivian. That figure doesn’t include the close to $12 billion in gross proceeds raised in Rivian’s IPO, nor did I count the more recent strategic deals with Volkswagen Group and Uber — which together could add nearly $7 billion to Rivian’s coffers.You can read my whole riff on the topic here. But if you don’t feel like clicking, here is one item that stood out. I spoke to a number of insiders and investors and they all mentioned Scaringe’s ability to give undivided attention to whoever he’s talking to — whether it’s an investor, supplier, or exec — and make them feel like the most important person in the room.It’s yet another piece of evidence in my long-standing case against multitasking. Debate me!Other deals that got my attention …Arkeus, an Australian startup that developed perception software for autonomous drones and aircraft,raised $18 millionin a Series A round led by QIC Ventures. Other investors include R+VC, Folklore Ventures, DYNE Ventures, Main Sequence Ventures, Salus Ventures, and Beaten Zone.Aseon Labs, a Redwood City, California, startup that has developed a depot in a box for charging, cleaning, and inspecting autonomous fleets,came out of stealthwith undisclosed backing by Y Combinator.Rapidoraised $240 millionin a round led by Prosus, and that values the Indian ride-hailing company at $3 billion. Existing investors, including WestBridge Capital and Accel, participated. The round was part of a larger $730 million primary and secondary financing.Quantum Systems, a Germany-based drone startup backed by Peter Thiel, is in talks to raise around €600 million ($703 million) with companies like Airbus and Blackstone as investors,Bloomberg reported.Notable reads and other tidbitsImage Credits:Bryce DurbinIs Redwood Materials ready for an IPO?Senior reporter Sean O’Kane interviewed the company’s new CFO,Deepak Ahuja, whose name will be familiar to anyone who follows Tesla. Ahuja was Tesla’s former finance chief and most recently held a similar position at drone company Zipline.TeslaRobotaxis have crashed at least twice since July 2025 while ateleoperator was remote

Commencement season has come around again — and this year, a couple speakers have discovered that it’s tough to get graduating students excited about a future shaped by artificial intelligence.Last week, Gloria Caulfield, an executive at real estate firm Tavistock Development Company,gave a speech at the University of Central Floridaacknowledging that we’re living in a time of “profound change,” which can be both “exciting” and “daunting.”“The rise of artificial intelligence is the next industrial revolution,” Caulfield declared — prompting the students in the audience to begin booing, getting louder and louder until Caulfield chuckled, turned to the other speakers, and asked, “What happened?”“Okay, I struck a chord,” she said. Caulfield then tried to resume her speech, saying, “Only a few years ago, AI was not a factor in our lives” — only to be interrupted again by the audience, this time by their loud cheers and applause.Former Google CEO Eric Schmidt faced a similar response when he brought up AI at a University of Arizona speech on Friday.In Schmidt’s case, the criticism actually began before the speech itself, with some student groupscalling for him to be removed as commencement speakerdue toa lawsuit in which a former girlfriend and business partner accused Schmidt of sexual assault. (He has denied the allegations.) According toa local news report,the booing began even before Schmidt took the stage.But Schmidt alsogot loud booswhen he told students, “You will help shape artificial intelligence.” The booing was persistent enough that Schmidt tried to speak over it, insisting, “You can now assemble a team of AI agents to help you with the parts that you could never accomplish on your own. When someone offers you a seat on the rocket ship, you do not ask which seat, you just get on.”To be fair, AI isn’t becoming a third rail ateverygraduation ceremony. Nvidia CEO Jensen Huang recentlyspoke at Carnegie Mellon’s commencement, and he didn’t seem to get any audible pushback when he said that AI has “reinvented computing.”Still, it’s not exactly surprising to find some students in a booing mood. Ina recent Gallup poll, only 43% of Americans aged 15 to 34 said it’s a good time to find a job locally, a steep drop from 75% in 2022.That pessimism isn’t solely a response to the rise of AI (a shift thateven some software engineers are worried about), but journalist and tech industry criticBrian Merchant suggestedthat for many students, AI has become “the cruel new face of hyper-scaling capitalism.”“I too would loudly boo at the prospect of this next industrial revolution if I was in my early twenties, unemployed, and had aspirations for my future greater than entering prompts into an LLM,” Merchant wrote.Even when graduation speeches didn’t mention AI explicitly, “resilience”was a recurring theme this year. Schmidt himselfacknowledgedthat there is “a fear in your generation that the future has already been written, that the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create.“Caulfield, meanwhile, might also have misread her audience of arts and humanities graduates. One student said that before mentioning AI, Caulfield already started to lose them with her “generic” praise of corporate executives like Jeff Bezos.Another graduate, Alexander Rose Tyson,told The New York Times, “It wasn’t one person that really started the booing. It was just sort of like a collective, ‘This sucks.’”TopicsAI,commencement speeches,Eric Schmidt,gloria caulfieldWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Anthony HaAnthony Ha is TechCrunch’s weekend editor. Previously, he worked as a tech reporter at Adweek, a senior editor at VentureBeat, a local government reporter at the Hollister Free Lance, and vice president of content at a VC firm. He lives in New York City.You can contact or verify outreach from Anthony by [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikSpotify launches a Wrapped-style recap of your entire listening historyIvan MehtaAmazon launches 30-minute delivery across the USSarah Perez

Listen onApple PodcastsListen onSpotifyLawyers for Elon Musk and OpenAI made their closing arguments this week, and now it’s up to jurors to decide whether OpenAI did anything wrong as it’s transformed intoa slightly-more-for-profitorganization.But as Kirsten Korosec, Sean O’Kane, and I noted on the latest episode ofTechCrunch’s Equity podcast, a big theme in the trial’s final days waswhether OpenAI CEO Sam Altman is trustworthy— for example, Musk’s attorney Steve Molo grilled Altman about whether statements he’d made during congressional testimony were truthful.Kirsten noted that Musk hasmade plenty of misleading statements of his own, and that trust isn’t just an issue for Altman.“This is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs,” she said. “It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies, there’s a lot behind the veil still.”Keep reading for a preview of our conversation, edited for length and clarity.Anthony Ha:[The end of the trial] led to this really provocative headline from one of our writers, Tim Fernholz, [that] just says,“Who trusts Sam Altman?”Does anyone want to take a stab at answering this?Kirsten Korosec:Yeah, Anthony, I’m going to throw it right back to you. Do you trust Sam Altman?Anthony:It’s an interesting question because it feels like something that’s kind of a wild question to discuss in a journalistic context, but actually that’s the core of the trial, in a lot of ways.Sean O’Kane:That’s not a yes.Anthony:And it actually seems to be [at the] core of understanding so much of what’s happened at OpenAI, especially this big executive power struggle that they now call The Blip.It just seems like a lot of people who’ve worked with Altmandon’t trust him. And he’s acknowledged this a little bit, because he’ll talk about the fact thathe recognizes he’s been conflict averse, telling people what they want to hear, and he’s trying to work on that.I mean, it sounds plausible, and I can understand how that can lead to misunderstandings in some situations. [But] I’m also a very conflict-averse person and I’d like to think that if any of this stuff went to trial, that people would not be asking, “Is Anthony Ha trustworthy?”Sean:Still not a yes!Kirsten:I think that people would say that you are trustworthy. I will say that question, while provocative, doesn’t just encapsulate what this trial was about. I would zoom out even more and say this is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs. It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies, there’s a lot behind the veil still.Maybe when they all IPO, we can get a peek, but it is fundamentally about trust and misuse, and do we believe the intent? And what I would throw back is, sometimes the intent can be worthy, noble, and still misused. It can still end up as a bit of a shit show. I think it’s more than who trusts Sam Altman — although that was very interesting in this trial — but more of that bigger question that we can apply to the entire industry.Sean:I’ll say it: I don’t trust him. But you know, I don’t trust most people, so I guess that’s just the baseline.We’ll see where this goes. The trial wraps up today. I’ve been very curious to hear how the jury decides this all. I think at the start of this, a big motivator of this was Elon Musk trying to sling mud, at a perceived rival and someone who he feels slighted him. And I don’t know if we know enough yet to say that that was completely accomplished, and whether or not he has a shot at winning. But I think all these people came out of this looking a little bit worse.Anthony:And just to get specific, why this is coming up this week is that [Altman] was on the stand and he was basically getting grilled about some statements he’s made in the past, in testimony to [Congress], basically saying he didn’t have any equity in OpenAI. And that is not true because he had a stake through Y Combinator, which he used to run. And tried to brush that off by saying, “I assume that everybody understands what it means to be a passive investor in a VC fund.” And I think [Elon Musk’s] lawyer, somewhat fairly, said “Really? You think the congressman who was interviewing you knew that?”Kirsten:Yeah, I mean, he was playing the whole semantics game. What I thought was so interesting about [this] is the style of how Sam Altman answered questions [compared to] Elon Musk on the stand.So Elon Musk, in many, many, many scenarios and many instances, we can point to the fact that he put something out on Twitter that was a lie or a bit of a fib, and on the stand corrected the record. So there’s a history of, I would say, non-truthfulness-slash-lying, blatant or otherwise, in Elon Musk’s world, but how he treated it was incredibly combati

In BriefPosted:1:15 PM PDT · May 17, 2026Image Credits:Artur Widak/NurPhoto / Getty ImagesAnthony HaApple’s Siri revamp could include auto-deleting chatsPrivacy will be a major theme when Apple unveils a new version of Siri at the Worldwide Developers Conference in June,according to Bloomberg’s Mark Gurman.The Siri relaunch is widely seen as Apple’s big chance to reestablish its relevance in artificial intelligence. As part of that effort, company executives will argue that they’re taking a more privacy-friendly approach than most other AI companies, Gurman said.Apple will reportedly launch the first standalone Siri app,powered by Google Geminiand offering users a chatbot experience reminiscent of ChatGPT.  But compared to those other chatbots, the app is supposed to have more limitations on how long user information can be used and stored.For example, Gurman said Siri could include a feature similar to the Messages app, allowing users to automatically delete conversations after 30 days or one year — or to keep them indefinitely.Gurman also suggested that Apple might be emphasizing privacy as a way to excuse Siri’s shortcomings compared to competing products — and that this emphasis might obscure the fact that Google is handling some the security.TopicsAI,Apple,Apps,gemini,siriMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAppsSpotify to adopt Apple’s new video podcast tech, offering creators easier cross-platform distributionLauren Forristal4 days agoSocialFinally, texts between Android and iPhone users can be end-to-end encryptedAmanda SilberlingMay 11, 2026HardwareApple to pay $250M to settle lawsuit over Siri’s delayed AI featuresLauren ForristalMay 6, 2026Latest in AIIn BriefApple’s Siri revamp could include auto-deleting chatsAnthony Ha11 hours agoAIWhy trust is a big question at the Elon Musk-OpenAI trialAnthony Ha12 hours agoAIIf you’re giving a commencement speech in 2026, maybe don’t mention AIAnthony Ha15 hours ago

Slow boring of hard boardsBill to block publishers from killing online games advances in CaliforniaPublishers would have to offer “independent” play patch or refunds after server shutdowns.Kyle Orland–May 15, 2026 12:35 pm|131Your controller won’t literally disintegrate when an online game’s servers shut down, but it can feel that way sometimes.Credit:

      Stop Killing GamesYour controller won't literally disintegrate when an online game's servers shut down, but it can feel that way sometimes.Credit:

      Stop Killing GamesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA bill focused on maintaining long-term playable access to online games has passed out of the California Assembly’s appropriations committee, setting up a floor vote by the full legislative body. The advancement is a major win forStop Killing Games‘ grassroots game preservation movement and comes over the objections of industry lobbyists at theEntertainment Software Association.California’sProtect Our Games Act, as currently written, would require digital game publishers who cut off support for an online game to either provide a full refund to players or offer an updated version of the game “that enables its continued use independent of services controlled by the operator.” The act would also require publishers to notify players 60 days before the cessation of “services necessary for the ordinary use of the digital game.”As currently amended, the act would not apply to completely free games and games offered “solely for the duration of [a] subscription. Any other game offered for sale in California on or after January 1, 2027, would be subject to the law if it passes.Battle of the interest groupsThis week’s 11–2 committee vote to advance the Protect Our Games Act isbeing trumpetedbyStop Killing Games(SKG), the UK-based player advocacy group that was formed after the 2024 shutdown of Ubisoft’sThe Crew. SKGwrote last monththat it “advised on the drafting” of the bill before it wasfirst introduced by Assemblyman Chris Wardearlier this year.“Back shortly before Christmas, when I flew to the US to help set up SKG-US, I didn’t expect us to get this far this quickly,” SKG’s Monitz Katznerwrote on Redditafter the committee vote. “It has been an honor to take part in drafting this bill on behalf of the SKG community: gamers, developers, and publishers alike.”In a formal statement of support for the bill sent to the California legislature, SKG wrote that “there is no other medium in which a product can be marketed and sold to a consumer and then ripped away without notice… As live service games rise in popularity for game developers and gamers alike, end-of-life procedures are essential tools to ensure prolonged access to the games consumers pay to enjoy.”The Entertainment Software Association, which helps represent the interests of major game publishers, publicly told the California Assembly last month that the bill misrepresents how modern game distribution actually works. “Consumers receive a license to access and use a game, not an unrestricted ownership interest in the underlying work,” the ESA wrote. The eventual shutdown of outdated or obsolete games is “a natural feature of modern software,” the group added, especially when that software requires online infrastructure maintenance.The ESA also said the bill would impose unreasonable expectations on publishers regarding licensing rights for music or IP rights, which are often negotiated on a time-limited basis. “A legal requirement to keep games playable indefinitely could place publishers in an impossible position—forcing them to renegotiate licenses indefinitely or alter games in ways that may not be legally or technically feasible,” they wrote.Last month, the Protect Our Games Act also received positive votes from the California Assembly’s Privacy and Consumer Protection and Judiciary committees. But the bill still faces significant hurdles in getting majority passage in the full California Assembly and the California Senate before being sent to California Governor Gavin Newsom for signature.Still, the current legislative progress in California has to be heartening for the Stop Killing Games movement, which has seen its momentum in the UK stall a bitafter a UK Parliament debate on game preservation last November.Kyle OrlandSenior Gaming EditorKyle OrlandSenior Gaming EditorKyle Orland has been the Senior Gaming Editor at Ars Technica since 2012, writing primarily about the business, tech, and culture behind video games. He has journalism and computer science degrees from University of Maryland. He oncewrote a whole book aboutMinesweeper.131 Comments

Baked in emissionsMaking cement from a different type of rock could clean up emissionsLimestone might not be the only source for Portland cement.Scott K. Johnson–May 15, 2026 1:06 pm|60A sample of Portland cement made from basalt.Credit:

      Brimstone EnergyA sample of Portland cement made from basalt.Credit:

      Brimstone EnergyText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navCement production alone currently accounts for about 8 percent of global CO2emissions, so considerable effort is going into lowering that number. Efficiency can be increased, and energy sources can be swapped for cleaner ones, but a stubborn reality remains: The byproduct of turning limestone into lime during cement production releases CO2gas. These “direct process emissions” are actually slightly larger than the emissions from burning fuel to heat the kilns and drive this process.A new paper in Communications Sustainability suggests a route to eliminating direct process emissions by removing a bedrock assumption. What if we don’t have to use limestone cement?Get out of PortlandThe material we call “Portland cement” was developed in the 1800s. It simply requires heating limestone (calcium carbonate) and adding something like clay or coal ash. This gives you the calcium oxide (lime) you’re after but also releases the CO2that results when you pull an oxygen atom from carbonate.The authors of the new paper include the CEO and an engineer from a company that says it has made Portland cement from silicate rocks like basalt—at the lab scale. Basalt contains a mix of minerals that include calcium, aluminum, iron, magnesium, sodium, silicon, and oxygen. (Note the absence of carbon from that list.) The basic idea is that you don’t need limestone to get calcium oxide.The process of freeing these components from basalt looks more like a refining or recycling process than the toss-it-in-the-oven simplicity of the limestone process. Acid can be used to leach elements like calcium out, then a chemical or energetic process precipitates that calcium as calcium hydroxide. Toss that in a kiln with additives of your choice, and with less heating than you need for limestone, you’ve got Portland cement, with only water vapor released.Those steps (along with follow-up reactions to restore the acid or other chemicals to a usable state) obviously add up in terms of cost and energy use. Tallying up the energy to do all this using common techniques, the researchers found that you need to use a little more than double the energy of traditional production from limestone.The interesting thing is that, according to thermodynamics, the chemical conversion of basalt minerals to calcium oxide only requires around half as much as the conversion from limestone. The problem is that our techniques to facilitate that chemical conversion are quite inefficient, so we don’t get anywhere near what is theoretically possible.Better options?The researchers note that there are at least some known lab techniques that could greatly improve our efficiency if they can be applied at scale, but even if we’re stuck with doubled energy usage, producing Portland cement from basalt would significantly reduce CO2emissions. That’s because the direct liberation of CO2from limestone is eliminated and because the whole process can run on electricity.Assuming you use electricity from a fossil-fuel-dominated grid, they estimate that emissions would be cut by almost 30 percent. Using clean electricity would eliminate most of the remaining emissions.The trade-off, obviously, would be cost, which generally wins out over the sustainability of a livable environment.But there is another interesting aspect to this idea: The other components of the basalt also have value. Iron, magnesium, and aluminum could also be separated and recovered, and leftover silicate material can serve as the additive for Portland cement instead of something like coal ash. So if these things were done together, the process could become more economically feasible.That’s a lot of ifs and buts, but this relatively simple analysis can at least point to what would have to happen to make this viable. And given that cement is one of the tougher nuts to crack in the struggle to reduce global greenhouse gas emissions, concrete solutions are welcome.Communications Sustainability, 2026. DOI:10.1038/s44458-026-00056-4(About DOIs).Scott K. JohnsonAssociate WriterScott K. JohnsonAssociate WriterScott has written about geoscience and energy at Ars as a freelancer since 2011.60 Comments

forget itVolkswagen shows its first electric GTI; there’s no chance of US salesThe ID. Polo GTI takes plenty of inspiration from the original Golf GTI of 1976.Jonathan M. Gitlin–May 15, 2026 1:55 pm|114If you live in Europe and you’ve been wanting an electric GTI, I’ve got good news for you.Credit:

      VolkswagenIf you live in Europe and you've been wanting an electric GTI, I've got good news for you.Credit:

      VolkswagenText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen Volkswagen introduced the first Golf GTI in Europe in 1976, it might not have been the first hot hatchback, but it quickly became the gold standard version. Unlike in America, where big cars were cheap and fuel even cheaper, small European streets and even smaller car-buying budgets necessitated vehicles a little more economical in both size and fuel consumption. Small, front-wheel-drive hatchbacks were the answer, but they weren’t particularly exciting. The GTI changed that perception with a more powerful engine, sharper handling, and subtle styling tweaks, creating a recipe for the next 50 years. And today, VW showed off its first electric GTI.While the new EV might be inspired by the original Golf GTI, it’s one segment smaller thanthe current Golf—meet the VW ID. Polo GTI. VW has given some of its ID EVs GTX branding until now, but this is the first to get the GTI badge.Like the 1976 original, the new car has front-wheel drive, but the ID. Polo GTI’s electric motor generates 222 hp (166 kW)—just over twice the output of the 1.6 L engine in the old car. There’s a 52 kWh battery pack that provides a WLTP range estimate of 236 miles (424 km), with DC fast charging up to 105 kW with a 10–80 percent charge time of 24 minutes.VW showed an electric GTI concept last year; now it’s almost ready for production.Credit:
                                  VolkswagenVW showed an electric GTI concept last year; now it’s almost ready for production.Credit:

      VolkswagenZero to 62 mph (100 km) at 6.8 seconds is brisk as opposed to rapid, and it’s still quicker than a 20th century VW GTI. A curb weight of 3,395 lbs (1,540 kg) is significantly more than those 70s, 80s, and 90s hatchbacks weighed, though.The looks have been enhanced with chunky 19-inch alloy wheels (that still make some concessions to drag reduction), honeycomb intake grilles, classic GTI details like the red stripe and badges, and a roof spoiler.On the inside, like any good GTI, there are sports seats and a sporty steering wheel, but VW has also given the interior a nod to the classic GTI tartan seat trim. The main instrument display might be a 10.25-inch digital screen, but it now looks a lot like the view you’d get from the driver’s seat of a late-70s Golf GTI.Peep those retro dials.Credit:
                                  VolkswagenPeep those retro dials.Credit:

      VolkswagenWhen the ID. Polo GTI goes on sale in Germany, VW says it should cost “just under €39,000.” That is more than $45,000 at today’s exchange rate. Although the German price includes 19 percent VAT, even without the oft-changing Trump tariff, it’s easy to see how this little GTI is a nonstarter for American imports.Although there’s no direct conversion between Europe’s WLTP test and the EPA’s drive cycle, a federalized ID. Polo GTI with a 52 kWh battery would be unlikely to post more than 200 miles (321 km) between charges. Low range like that is anathema to US EV buyers, particularly when you consider that a new Chevrolet Bolt offers three-quarters the price, 262 miles (422 km) of range, and a similar 0–60 mph time.Ah, well, nevertheless.Jonathan M. GitlinAutomotive EditorJonathan M. GitlinAutomotive EditorJonathan is the Automotive Editor at Ars Technica. He has a BSc and PhD in Pharmacology. In 2014 he decided to indulge his lifelong passion for the car by leaving the National Human Genome Research Institute and launching Ars Technica's automotive coverage. He lives in Washington, DC.114 Comments

“Strained” relationshipOpenAI feels “burned” by Apple’s crappy ChatGPT integration, insiders sayJudge orders Apple to give Musk internal messages discussing secretive ChatGPT deal.Ashley Belanger–May 15, 2026 2:13 pm|97OpenAI CEO Sam Altman talks with Apple senior Vice President of Services Eddy Cue during the Apple Worldwide Developers Conference (WWDC) on June 10, 2024, where Apple announced plans to incorporate ChatGPT into Apple software and hardware.Credit:Justin Sullivan / Staff | Getty Images NewsOpenAI CEO Sam Altman talks with Apple senior Vice President of Services Eddy Cue during the Apple Worldwide Developers Conference (WWDC) on June 10, 2024, where Apple announced plans to incorporate ChatGPT into Apple software and hardware.Credit:Justin Sullivan / Staff | Getty Images NewsText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navOpenAI is reportedly exploring legal options after Apple’s ChatGPT integration into its products didn’t live up to the AI firm’s expectations.When the deal was announced, Apple likened features linking Siri to ChatGPT to itsnow-infamous deal embedding Google search in the Safari browser, insiders granted anonymity to discuss the “strained” partnershiptold Bloomberg. And the promise of that excited OpenAI, which expected the deal “could generate billions of dollars per year in subscriptions,” an OpenAI executive granted anonymity to discuss the partnership told Bloomberg.Instead, OpenAI suspects Apple intentionally failed to promote the integration and fears that the deal may have damaged the ChatGPT brand, sources said.Specifically, OpenAI hates how Apple designed the integration, sources said. Particularly bad was the choice forcing Apple users summoning Siri to also “specifically invoke the word ‘ChatGPT’ when speaking or typing a command,” sources said. That makes it harder for users to access the features, OpenAI apparently feels. And Apple’s other choices, like using small windows providing limited information when responding with ChatGPT outputs, seems to ensure that users can easily ignore the features, sources said.As the OpenAI executive explained, Apple didn’t fully explain how the integration would work when the deal came together, so OpenAI took a “leap of faith” it now appears to regret.“When we heard about this opportunity, it sounded amazing: being able to acquire a giant number of customers and have distribution in such a big mobile ecosystem,” the executive said, attempting to explain why OpenAI was willing to enter the arrangement blind. Since then, efforts to renegotiate the deal have “stalled,” Reutersreported. And, supposedly due to feeling “burned,” OpenAI has declined to enter other partnerships to work on Apple’s AI models, Bloomberg reported.According to the insiders, OpenAI is so disappointed in Apple’s work that the AI firm is now “actively working with an outside legal firm on a range of options that could be formally executed in the near future.”“We have done everything from a product perspective,” the OpenAI executive summed up OpenAI’s frustrations to Bloomberg. “They have not, and worse, they haven’t even made an honest effort.”Supposedly, OpenAI is still hoping to resolve its issues with Apple outside of court, if possible. But one option that OpenAI may pursue could be accusing Apple of a breach of contract. Going that route wouldn’t necessarily require filing a lawsuit right away, sources suggested.Apple and OpenAI did not respond to Ars’ request to comment.Musk may expose how deal was doneMost likely, OpenAI will delay approaching Apple until after its court battle with Elon Musk concludes, Bloomberg reported. A decision in that litigation is potentially coming next week.OpenAI also faces a court battle with Musk over its Apple deal. However, it may be inconvenient for Musk that tensions between OpenAI and Apple have grown sincehe filed a lawsuit last August.Musk alleged that the deal integrating ChatGPT into Apple products violated antitrust and unfair competition laws, supposedly propping up OpenAI to dominate the chatbot market and Apple the smartphone market.So far, Musk’s lawsuit has survived motions to dismiss, though the judge has yet to comment on its merits. That leaves Apple and OpenAI potentially stuck defending the deal at a trial scheduled for October, even if it falls apart.However, the partnership’s end may make it harder for Musk to uphold his claims of a conspiracy in his lawsuit.Due to Musk’s fury that his chatbotGrok has never been featured as a “Must Have” appin Apple’s App Store, the lawsuit alleged that Apple and OpenAI struck a deal as part of a giant conspiracy to lock out rivals developing chatbots that Musk claimed Apple fears could make smartphones obsolete. As Musk’s theory goes, Apple was so afraid ofMusk’s plan to turn X into an “everything app”that it partnered with OpenAI to supercharge ChatGPT as a market leader and constrain

RejectedSend the arXiv AI-generated slop, get a yearlong vacation from submissionsOne of the site’s moderators described the new policy on social media.John Timmer–May 15, 2026 2:25 pm|77Credit:Getty ImagesCredit:Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAI-generated slop has shown up everywhere, including in the peer-reviewed literature. Fake citations, unedited prompt responses, andnonsensical diagramshave all slipped past editors and peer reviewers, and it’s not always clear if there are any consequences for the people responsible.Now, it appears that a number of scientific fields will be enforcing rules against AI-generated problems even before peer review or journals get involved. One of the people involved in the physics and astronomy preprint server arXiv used a social media thread to announce that any inappropriate AI-produced content submitted to the server will result in a one-year ban and a permanent requirement that future publications undergo peer review before the arXiv will host them.Thomas Dietterich, in addition to being an emeritus professor at Oregon State University, is heavily involved with arXiv, serving on itseditorial advisory counciland on itsmoderation team. So he’s in a good position to understand the organization’s policies, although we have also reached out to arXiv leadership for confirmation, but have not yet received a response.In athread on X(alsoscreenshotted on Bluesky, for those without X accounts), Dietterich described the new policy as arising directly from the arXiv’smoderation standards. “Submissions to arXiv must comply with appropriate standards of scholarly communication in form, including appropriate and carefully prepared sections, figures, tables, references, etc.,” those standards read. “General scrupulousness and care of preparation are required.”Dietterich also notes that all authors of a manuscript are responsible for its content. So, if they carelessly submit material generated by an AI that violates these guidelines—Dietterich cites “inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content”—then they’re responsible, not the AI. Should violations be discovered, all of the manuscript’s listed authors will now receive a one-year submission ban, and any future manuscripts will only be accepted after they’ve been through peer review by a journal.For fields that rely heavily on the arXiv, those are severe sanctions. Posting preprints in areas like astrophysics is widely considered part of the normal publication process, and scientists will often get feedback on preprints that helps them improve what they submit for peer review. The unfortunate problem is that, like most other things, the system can be gamed—people could submit flawed content that lists people as authors who have never been involved. Fortunately, its moderation system includes anappeal process.One obvious question that arises when these problems are found in publications is why nobody caught them sooner. Now, we can at least know that someone is trying to.John TimmerSenior Science EditorJohn TimmerSenior Science EditorJohn is Ars Technica’s science editor. He has a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in Molecular and Cell Biology from the University of California, Berkeley. When physically separated from his keyboard, he tends to seek out a bicycle, or a scenic location for communing with his hiking boots.77 Comments

Already largeEbola outbreak with uncommon strain erupts in Congo and Uganda; 65 deathsWHO learned of potential cases May 5; US CDC said it just heard about it yesterday.Beth Mole–May 15, 2026 2:51 pm|54An Ebola patient is being checked by two medical workers after being admitted into a biosecure emergency care unit on August 15, 2018 in Beni.Credit:Getty | JOHN WESSELSAn Ebola patient is being checked by two medical workers after being admitted into a biosecure emergency care unit on August 15, 2018 in Beni.Credit:Getty | JOHN WESSELSText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navThe Africa Centres for Disease Control and Prevention on Friday confirmed an Ebola outbreak in the Northeastern Ituri province of the Democratic Republic of the Congo. Officials in Uganda subsequently reported that the deadly hemorrhagic disease had spilled over the border, with one “imported” confirmed case identified in Kampala, the capital.So far, the DRC has reported 246 suspected cases and 65 deaths, mainly in the Mongwalu and Rwampara health zones. Although it is now just being reported, the outbreak already ranks around the 10th largest Ebola outbreak to date.This marks the 17th Ebola outbreak in the DRC since the virus was discovered in 1976. But unlike almost all of the country’s past Ebola outbreaks, preliminary laboratory results suggest that the current outbreak is not caused by the Zaire Ebola virus strain. Further genetic sequencing is ongoing today to determine the strain causing the illnesses.There arefour virus strains known to cause Ebola diseasein humans, and three have caused large outbreaks (Zaire, Sudan, and Bundibugyo). According to an early Reuters report from Uganda, officials said thecase there involved the Bundibugyo strain.Complicated contextJean-Jacques Muyembe, the Congolese virologist who co-discovered ​Ebola and heads the National Institute for Biomedical Research in Kinshasa,told Reutersthat finding a different strain than Zaire behind the outbreak could complicate responses, given that current licensed vaccines and treatments were developed against the disease from the Zaire strain.Africa CDC highlighted other complications of this outbreak, which include the urban context of Bunia and Rwampara; intense movement of people in the area, including mining-related mobility; militia clashes in the area; and gaps in contact tracing. The agency also noted the potential spillover not just to Uganda but also to South Sudan.“Given the high population movement between affected areas and neighboring countries, rapid regional coordination is essential,” Dr. Jean Kaseya, director-general of Africa CDC, said in a statement Friday. “We are working with DRC, Uganda, South Sudan, and partners to strengthen surveillance, preparedness, and response, and to help contain the outbreak as quickly as possible.”In a press briefing on Friday afternoon, Jay Bhattacharya, who is heading the US CDC for now, said that the agency, which has offices in both DRC and Uganda, is offering technical assistance to the countries. “It is a large outbreak, and we were just informed yesterday about it. So we’ve been working very, very hard to coordinate with them,” Bhattacharya said.Ina separate press briefing earlier today, World Health Organization Director-General Tedros Adhanom Ghebreyesus said the WHO received a signal of suspected Ebola cases in the DRC and sent a team to support the health ministry on May 5. WHO representatives are still there assisting in outbreak responses.Tedros also said WHO was releasing today $500,000 from WHO’s Contingency Fund for Emergencies to immediately support the response efforts.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.54 Comments

Let’s take a lookThree’s a party: US, China, and now Russia are on the prowl in GEOInstead of running silent and deep, most satellites easily stand out against the blackness of space.Stephen Clark‚ÄìMay 15, 2026 3:11 pm|30This visualization from COMSPOC shows the US Space Force’s USA-325 satellite flying near Russia’s Kosmos 2589 satellite in geosynchronous orbit May 1.Credit:

      COMSPOCThis visualization from COMSPOC shows the US Space Force's USA-325 satellite flying near Russia's Kosmos 2589 satellite in geosynchronous orbit May 1.Credit:

      COMSPOCText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe world’s leading space powers desperately want to know what the others are up to high above the equator. For more than a decade, the US military has operated a fleet of “inspector” satellites designed to sidle up to other spacecraft in geosynchronous orbit and take pictures. China started launching its satellites for a similar mission in 2018.Ars haswritten about these activitiesin geosynchronous orbit (GEO) before, but the last few months have seen a couple of interesting developments. First, Russia has now joined the fray with the recent arrival of its own suspected inspector (or attack) satellite in GEO. Second, the US Space Force is poised to order more—perhaps many more—reconnaissance satellites of its own to send into the geosynchronous belt.GEO is special. The laws of orbital mechanics mean a satellite in this type of orbit, some 22,000 miles (36,000 kilometers) over the equator, moves around the Earth at the same rate as the planet’s rotation, causing it to hover over the same location. Commercial and military-owned geosynchronous satellites typically spend years in the same location, or slot, to provide communications services to users.Until now, Russia’s spying in geosynchronous orbit has primarily focused on eavesdropping on foreign communications. Russia launched two satellites,Olymp or Luch, to wander around geosynchronous orbit, moving from slot to slot to loiter near Western-owned communications satellites for several months at a time. The goal, according to Western analysts, was to listen in on or potentially jam signals relayed through these satellites, some of which route secure communications for US and NATO military forces.The trend today is geared toward reconnaissance and surveillance in GEO. Military forces and intelligence agencies want to know where other satellites are located, what they look like, and what they’re capable of doing. The US military’sGeosynchronous Space Situational Awareness Program (GSSAP)satellites, which began launching in 2014, do exactly this by roaming the geosynchronous belt, using propulsive maneuvers to make slight changes to altitude and inclination to move within a few dozen miles of Chinese and Russian satellites, close enough for optical telescopes to get a good look.The Space Force has often compared the GSSAP satellites to a “neighborhood watch” in geosynchronous orbit. For example, one GSSAP satellite positioned itself near a pair of Chinese spacecraft performing a first-of-its-kind refueling demonstration in geosynchronous orbit last year.China has launched multiple satellitescapable of similar maneuvers. One of these satellites, named TJS-10, is currently flying relatively close to a nuclear-hardened US Space Force strategic communications satellite and a US missile warning platform, according to anupdate this weekin the Integrity Flash newsletter published by ISR University.Ready Player ThreeA new kind of Russian satellite is now in the mix. This satellite, officially known as Kosmos 2589, was launched in June 2025 into a highly elliptical orbit alongside a smaller spacecraft designated Kosmos 2590. The two satellites performed a series of high-altitude rendezvous and proximity operations with one another before Kosmos 2589 began moving toward a more circular geosynchronous orbit, where it arrived in April.One of the US military’s GSSAP satellites was waiting for it. The US inspector spacecraft is now looping around Kosmos 2589, swinging near the newly arrived Russian satellite twice per day, coming as close as 8 miles (13 kilometers) on May 1, according todata from COMSPOC, a commercial space situational awareness company. The exact purpose of Kosmos 2589 remains unclear. Some Western officials suspect it is a higher-altitude version ofRussia’s Nivelir anti-satellite system, which has been tested in low-Earth orbit and is now becoming operational.The video below, published by COMSPOC, shows the orbital dance between a GSSAP satellite named USA-325 and Kosmos 2589.👀USA 325 keeps an eye on COSMOS 2589 as it settles into GEOCOSMOS 2589 🇷🇺 (launched last June) spent the past few weeks circularizing its orbit before parking at ~98°E. USA 325 🇺🇸 (GSSAP-6) was watching the whole time.Things go

Storm’s a-brewin’Weather-monitoring firm hangs dark cloud over customers’ heads by forcing new appNewer AcuRite Now app lacks some features but has a subscription option.Scharon Harding–May 15, 2026 4:16 pm|62A marketing image for the AcuRite Optimus weather station with the AcuRite Now app pictured.Credit:

      AcuRiteA marketing image for the AcuRite Optimus weather station with the AcuRite Now app pictured.Credit:

      AcuRiteText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWeather-monitoring company AcuRite is forcing device owners to use a new companion app on May 30, frustrating some long-time customers.AcuRite, which sells devices such as weather stations, indoor thermometers, and rain gauges, began emailing customers last month that they’d soon have to control their devices with the AcuRite Now iOS and Android app. AcuRite first launched the app in June 2025 to control a new weather station, the AcuRite Optimus. However, owners of AcuRite devices had still been able to use the My AcuRite app, which launched in 2016.Soon, however, My AcuRite will no longer be available, making AcuRite Now the only official app for controlling AcuRite devices. The website for theMy AcuRiteapp currently reads:As part of our continued investment in delivering smarter, more connected solutions, the My AcuRite app will be winding down. To ensure uninterrupted access to your weather data and to unlock even more capabilities,all users should transition to AcuRite NOW before May 30th, 2026.Per the website, AcuRite Now “works with thousands of products” in Tuya’sSmartLife IoT ecosystem, including third-party fans, thermostats, light bulbs, plugs, cameras, and motorized blinds.“Looks like a bad joke”Some AcuRite customers are upset about losing access to My AcuRite because they think AcuRite Now is an inferior app.Those customers havecomplained onlineabout being unable to use AcuRite Now to rename multiple temperature sensors, difficulties uploading data to weather sites, and the app only reporting temperatures in whole numbers. An AcuRitesupport pagesays that AcuRite is “hoping” to add the ability to organize on-screen sensors and rename multiple sensors to AcuRite Now, as well as a desktop version of the app, “soon.”One popular feature of My AcuRite was the ability to share data from AcuRite devices withWeather Underground, a real-time weather service. AcuRite Now also supports this feature, but users have to pay a subscription fee that starts at $2 per month. The subscription, dubbedAcuRite Now+, also includes more data storage (365 days of history instead of 30).Other complaints center on AcuRite Now’s layout. For example, a Reddit user named epicfailphxsaidlast month that the app “looks like a bad joke.”“The old app was clean and just did a good job, while I don’t know what this new app is trying to do,” they said.Another Reddit user, ssemos,complainedthat the new app shows less data on-screen than My AcuRite does and doesn’t use most of the display.AcuRite’s riskSince AcuRite said it’s shuttering My AcuRite, AcuRite Now has received numerous one- and two-star reviews on the Apple App Store and Google Play Store, resulting in ratings of 1.4 stars (based on 183 reviews) and 1.3 stars (based on 131 reviews), as of this writing. It’s worth noting, though, that My AcuRite had2 stars(based on 429 reviews) on the App Store as of February and1.9 stars(based on 1,330 reviews) on the Play Store as of January, per the Internet Archive’s Wayback Machine.Ars Technica reached out to AcuRite, its parent company Primex Family of Companies, and AcuRite Now’s developer for comment but didn’t hear back before publication.It’s possible that AcuRite’s shuttering of My AcuRite aims to simplify the company’s technology stack by allowing it to focus on updating and maintaining one app rather than two, which could also save money. AcuRite Now’s support of SmartLife products could also be beneficial and drive app usage. The new app could also be more accommodating of new and upcoming features and allow AcuRite to charge a subscription fee for premium capabilities, helping diversify revenue.Still, forcing loyal customers with complex hardware to switch to a new app on a couple of weeks’ notice is a gamble. AcuRite risks polarizing customers who don’t like the new app, especially since it is missing some features that the old app has. Charging a subscription fee for a feature that was previously free could also push customers to explore rival products. Updated software canbring new life to gadgets, but as we’ve seen before,forcing long-time users to use new softwareperceived as inferior can havedisastrous effects.Scharon HardingSenior Technology ReporterScharon HardingSenior Technology ReporterScharon is a Senior Technology Reporter at Ars Technica writing news, reviews, and analysis on consumer gadgets and services. She's been repo

Lost in the fogSolar power production undercut by coal pollutionEach year, some of the power solar could have produced is blocked by aerosols.John Timmer–May 15, 2026 4:36 pm|81Credit:VCGCredit:VCGText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navCoal is by far the most polluting fuel that we use. It produces the most carbon emissions per unit of energy, and impurities in the coal produce a lot of sulfur dioxide aerosols, as well as nitrous and nitrogen oxides. Then there’s the coal ash that’s left behind, which typically contains a lot of toxic metals. The health benefits of displacing coal power are typically estimated to be well above the costs of the new generating equipment.But a new study suggests that the problems with coal-derived pollution go beyond health; it interferes with other power sources. Researchers have found that aerosols, both natural and human-derived, significantly reduce the power we could be getting from solar panels, to the tune of hundreds of terawatts a year. And a lot of those aerosols come from burning coal.A big impactThe new work, done by a team in the UK, is based on a new global inventory of solar facilities. This started with known inventories of solar facilities, and was supplemented with AI-analyzed satellite imagery and crowdsourced records of locations. Satellite images were then used to determine the size of these facilities, and location-tagged weather data could then be used to estimate their power production.That could then be used to estimate what the facilities would be producing if clouds and/or aerosols weren’t scattering the sunlight that would otherwise reach the panels. This produced some significant numbers. In 2023, for example, over a quarter of the potential solar power production was lost, with over 20 percent due to clouds and another 6 percent from aerosols. That works out to be a bit over 500 terawatt-hours, or the full annual output of 84 coal plants (each with a 1 GW generating capacity).Aerosols alone are a major contributor to these losses. The researchers note that, for the five years leading up to 2023, we installed enough solar capacity to produce an average of 250 TW-hr of additional power per year, but were losing 75 TW-hr of that to aerosols. (Obviously, solar production kept going up because the existing capacity rose each year.)Regional differencesThe researchers note that aerosols can also contribute to cloud formation, which also causes further losses. But the degree of that contribution is much harder to estimate, so the researchers focus on aerosols for much of the analysis. Some of those aerosols occur naturally, typically from dust kicked up by winds in desert regions. However, despite deserts’ reputation as sunny paradises, the world as a whole hasn’t built much solar infrastructure in the desert yet, so this isn’t as much of a factor as you might expect.Coal appears to be a major contributor. It’s estimated that sulfur dioxide aerosols, primarily produced through coal burning, account for nearly half of the aerosols analyzed here. Carbon-rich material, which also typically comes from fossil fuels, accounts for another 18 percent.The impact of aerosols, however, is not evenly distributed. In China, the researchers estimated that aerosols were reducing solar production by 7.7 percent overall and offsetting anywhere from a third to half of its annual growth. The researchers note that “the spatial distribution of photovoltaic losses in China mirrors that of its coal-fired power capacity,” and an analysis of pollution data from China shows that 30 percent of the losses due to aerosols can be attributed to coal burning.In contrast, most solar production in the US takes place in the south and west, while coal plants are more common in the east and northeast. As a result, the annual losses in the US were less than half of those seen in China (3 percent).The good news is that things are getting better in China. In response to some severe pollution problems, the country built a new generation of high-efficiency coal plants and retired some of the worst polluters. And the data show that this is also benefiting solar power, with the impact of aerosols dropping over the last few years.Even with the improvements, it’s striking that coal appears to be the only power source that actively reduces the productivity of what’s shaping up to be its primary competitor. It should also provide an impetus to move off coal more quickly, as at least some of the loss of coal production will be offset by enhanced productivity from solar.Nature Sustainability, 2026. DOI:10.1038/s41893-026-01836-5John TimmerSenior Science EditorJohn TimmerSenior Science EditorJohn is Ars Technica’s science editor. He has a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in Molecular and Cell Biology from the University of California, Berkeley. When physically separated from

behold the blessed conclusionReview:Good Omensfinale (mostly) sticks the landingTruncated third season feels rushed, but also gives us a fitting end to a love story for the ages.Jennifer Ouellette–May 15, 2026 5:17 pm|43Credit:

      Prime VIdeoCredit:

      Prime VIdeoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navIt’s been a three-year wait, but Prime Video finally released the series finale forGood Omens: a 90-minute single episode that sought to wrap everything up in a neat little bow. Verdict: Truncating the final season so drastically definitely hurts the first half of the series finale, which feels chaotic and rushed. But once that stupendous on-screen chemistry between co-stars David Tennant and Michael Sheen kicks back in, the old magic shines through, strong as ever, giving us a fitting end to this beloved comic saga.(Spoilers below for all seasons.)Here’s abrief recap, since it’s been a minute since the S2 finale. The series is based on the original 1990 novel byNeil Gaimanand thelate Terry Pratchett.Good Omensisthe story ofan angel, Aziraphale (Sheen), and a demon, Crowley (Tennant), who gradually become friends over the millennia and team up to avert Armageddon. Season 2 found Aziraphale and Crowley getting back to normal, when the archangel Gabriel (Jon Hamm) turned up unexpectedly at the door of Aziraphale’s bookshop with no memory of who he was or how he got there. The duo had to evade the combined forces of Heaven and Hell to solve the mystery of what happened to Gabriel and why.In the S2 finale, the pair discovered that Gabriel had defied Heaven and refused to support a second attempt to bring about Armageddon. He hid his own memories from himself to evade detection. Oh, and he and Beelzebub (Shelley Conn) had fallen in love. They ran off together, and the Metatron (Derek Jacobi) offered Aziraphale Gabriel’s old job. That’s when Crowley professed his own love for the angel and asked him to leave Heaven and Hell behind, too. Aziraphale wanted Crowley to join him in Heaven instead. So Crowley kissed him, and they parted. Once Aziraphale got to Heaven, he learned his task was to revive the stalled plans to bring about the Second Coming, i.e., the End Times.Aziraphale wants to make the Second Coming a bit more upbeat.Prime VideoAziraphale wants to make the Second Coming a bit more upbeat.Prime VideoMuriel is hot on the case of the missing Book of Life.Prime VideoMuriel is hot on the case of the missing Book of Life.Prime VideoThe archangels of heaven are unimpressed with Muriel’s efforts thus far.Prime VIdeoThe archangels of heaven are unimpressed with Muriel’s efforts thus far.Prime VIdeoMuriel is hot on the case of the missing Book of Life.Prime VideoThe archangels of heaven are unimpressed with Muriel’s efforts thus far.Prime VIdeoThe original plan for the third and final season called for six episodes, but production was delayed first by the 2023 writer’s strike and then bymultiple allegationsofsexual assaultagainst Gaiman. (Gaiman has vehemently denied allegations of any nonconsensual sex or abuse, but admitted to being selfish and “careless with people’s hearts and feelings” in aJanuary 2025 blog post.) The fallout led to Gaiman withdrawing from the project and Prime Video opting for a 90-minute finale rather than a full season. And here we are.The Second Coming hits a snagThe finale picks up a few years after the S2 cliffhanger. Aziraphale is now Supreme Archangel, with plans for the Second Coming well underway—except he’s tweaked them to be a bit more upbeat, bringing peace on Earth and universal happiness rather than the rampant death and destruction of Armageddon. This doesn’t go down well with some of his fellow angels, who prefer the original plan. A heartbroken Crowley, meanwhile, is spending his time drinking heavily and passing out in a Soho alley, having lost his sense of purpose when Aziraphale refused him.The Second Coming rollout soon hits a snag. First, the Metatron mysteriously vanishes, removed completely from reality by someone who has stolen the Book of Life. In the ensuing panic, Jesus (Bilal Hasna) wanders off down to Earth and is befriended by a former street hustler named Harry the Fish (Mark Addy). The Archangel Michael (Doon Mackichan) and plucky assistant Muriel (Quelin Sepulveda) focus on solving the Metatron’s murder, while Aziraphale heads down to Earth to hunt for the missing Jesus, lest the demons of Hell find him first. He enlists a reluctant Crowley’s help.Good Omenshas always embraced the colorfully comic side quest; it’s part of what makes this such a rich fictional universe. But you need time to flesh it all out for those subplots to really work, and time is what the finale just doesn’t have. Hell and its demons, in particular, seem little more than an afterthought here; they’re not even particularly effective as comic relief. The beats

Now we waitUS hantavirus case was false positive; outbreak cases drop from 11 to 10WHO announced today that the operation to safely transfer passengers is complete.Beth Mole–May 15, 2026 5:31 pm|17Evacuation by boat of passengers on board the cruise ship MVHondiusanchored near the port of Granadilla, on May 11, 2026 in Granadilla de Abona, Tenerife, Canary Islands, Spain.Credit:Getty | Europa Press CanariasEvacuation by boat of passengers on board the cruise ship MVHondiusanchored near the port of Granadilla, on May 11, 2026 in Granadilla de Abona, Tenerife, Canary Islands, Spain.Credit:Getty | Europa Press CanariasText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navIn a press briefing Friday, officials for the World Health Organization announced that the case count of thehantavirus outbreak on the cruise shipMV Hondiusin the South Atlantic has shrunk from 11 cases to 10 after a previously reported US case was found to be a false positive.That US case was originally reported byUS health officials as “mildly positive,”and the WHO had considered it “inconclusive,” but still counted in the outbreak as a case in the agency’sMay 13 outbreak reportand in a briefing on May 14.The inconclusive case was in Dr. Stephen Kornfeld, an American doctor aboard the ship who helped respond to the outbreak after the ship’s doctor became ill.In an interview with CNN earlier this week, Kornfeld explained that he and others on board had taken nasal swabs early in May, before evacuation, and those swabs were sent for PCR testing in the Netherlands. Two labs in the Netherlands processed Kornfeld’s swabs; one lab reported a negative result, and the other reported a faint positive.Generally, a faint positive result on a PCR test could suggest low levels of virus at the start or end of an infection, or it could simply suggest contamination.Adding to the complexity of the potential case, Kornfeld said that he had developed a minor illness in early April while on the ship, just a few days after the first hantavirus case fell ill in the outbreak—a Dutch man who died on board from his infection.In an interview with ABC News Friday, Kornfeld reported that his repeat testing was negative and so was his serology testing looking for antibodies against hantavirus—which he would have developed if his illness in April was actually an unusually mild hantavirus infection. Overall, the testing shows that he is not currently infected and has not had a previous infection. As such, he has been transferred from the biocontainment unit to the quarantine unit at the Nebraska Medical Center, where US passengers from theHondiusare being monitored.“I physically feel great—I have felt great for many, many days,” Kornfeld told ABC. “Emotionally, I feel wonderful. It’s nice to be negative for hantavirus.”Evacuation completeOverall, there have been 10 cases: Seven cases who became ill on the ship, one case that disembarked the ship on April 24, before the outbreak was identified, and fell ill in Switzerland. The remaining two cases—one from France, the other from Spain—were identified as the ship was evacuated and passengers were being repatriated. The number of deaths has not changed since May 2, with three total: a Dutch couple and a German woman.Also in the WHO press conference, Director-General Tedros Adhanom Ghebreyesus said that the operation in the Canary Islands to safely evacuate the ship is complete. The operation involved transferring the ship’s 120-plus passengers to their home countries or to quarantine in host countries en route to their final destinations.“Because of the long incubation period of up to six weeks, more cases may be reported in coming days as passengers return to their countries, where they are being quarantined and tested in specialized facilities or at home,” Tedros warned. “This does not mean the outbreak is expanding; it shows that the control measures are working, that laboratory testing is ongoing, and that people are being cared for with support from their governments.”The US is monitoring 41 people who may have been exposed. That includes 18 passengers (including Kornfeld) who were evacuated from theHondius, passengers who had disembarked the ship on April 24, and people who shared a flight with a Dutch woman (the wife of the first case) who also fell ill from the infection and died in South Africa, on her way home.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.17 Comments

“Powerball-size” attorney payoutsAnthropic’s $1.5B copyright settlement is getting messy as judge delays approvalLawyers accused of rushing historic settlement to seize $320 million in fees.Ashley Belanger–May 15, 2026 5:51 pm|80Credit:SvetaZi | iStock / Getty Images PlusCredit:SvetaZi | iStock / Getty Images PlusText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAfter several authors and class members raised objections toAnthropic’s $1.5 billion settlementover its widespread book piracy to train AI, a federal judge has delayed final approvals of the settlement.On Thursday, US District Judge Araceli Martinez-Olguin declined to rubber-stamp what’s regarded as the largest copyright settlement in US history. Instead, she wanted to better understand why some class members were objecting and opting out of the settlement. So, she asked authors to address key concerns of objectors, who argued that lawyers’ compensation was way too high and payments to class members were a “pittance.”Ars reviewed several objections to the settlement, as well as letters from objectors who claimed that the authors’ legal team was trying to unfairly shut them out from voicing concerns.Calling out lawyers for requesting more than $320 million in legal fees when each author only expects a $3,000 payout, some objectors asked the court to delay approving the settlement until a more reasonable plaintiff compensation plan is constructed.“Every dollar that Counsel takes from the Settlement fund is one that is not given to those actually harmed,”wrote Pierce Story, an objector and author of two works covered by the settlement.To support his arguments against the eye-popping lawyer fees, Story estimated that the large payout could break down to lawyers receiving between roughly $10,000–$12,000 per hour, which he said included a generous estimate of hours for any future work. That’s excessive, Story suggested, citing a T-Mobile case where the 8th Circuit court observed that “no reasonable class member would willingly pay” a much lower requested fee award between $7,000–$9,500.Story accused lawyers of breaking a promise to tie their compensation to member payouts. And he’s further frustrated that the compensation they’re seeking is tied to the full settlement fund, when many authors entitled to compensation have yet to register and “are unlikely to be compensated.”An attorney for authors confirmed on Thursday that “authors and other copyright holders filed claims covering over 92% of the more than 480,000 works included in the settlement.” But objectors maintain that lawyers’ pay should reflect the total number of claimants, not the total amount in the settlement fund.By urging the court to make “reasonable and fair adjustments” to lower attorney fees, Story is hoping to increase compensation to authors. Offering an example, he noted that “a still-generous Counsel payout of $70 million would yield a nearly 25 percent increase in individual Plaintiff awards, while Counsel would still receive the equivalent of their current top rates” for hours worked.To Story, it also seemed like the attorneys could’ve gotten more compensation for authors, but instead of pursuing “creative options,” they “settled far too quickly to maximize” their own compensation.“Were the attorneys as skilled, gritty, and brilliant as they profess, and were the Settlement the ‘home run’ Counsel claims it to be, Plaintiffs would receive more than this pittance,” Story said.Ruben Lee, another class member objecting, agreed: “I believe the amount offered is paltry, and does not in any way reflect the full value of the unauthorized use of my work.”Objectors may not win every fight, but they have seemingly persuaded the court to at least entertain their strongly worded pleas, including warnings that the settlement may not survive an appeal if the terms aren’t re-examined. Notably, their objections came shortly before a group of 25 class members opting out of the settlement filed a newlawsuit, showing that Anthropic is not done fighting these claims.“For the Court to agree that counsel’s request of nearly a third of a billion dollars, while individual plaintiffs settle for a pittance of available compensation and no protections against future abuse is an aberration of civil justice and a slap in the face to all those who labored to publish their works,” Story said. “Such a decision would also further the too-often-observed stereotype that … class-action Plaintiffs are merely tools used to obtain Powerball-size payouts to attorneys.”Judge William Alsup, who initially approved the settlement but has since retired, also questioned whether the lawyers’ fees were too high. Worried that the settlementwas being “shoved down the throat of authors,”he recommended an independent investigation to ensure no improper attorneys’ fees would be granted, but according to Lea Bishop, a non-class member objector and profe

Bad educationRussia pressures university students to become wartime drone pilotsUniversities promise no frontline duty and perks if students enlist in military.Jeremy Hsu–May 15, 2026 6:19 pm|179Shaman, a 19-year-old drone operator for the Russian military, flies a quadcopter drone during a demonstration event organized by members of the Berkut Military-Sports Cossack Club in a shopping centre in Voronezh, Russia on January 24, 2026.Credit:TATYANA MAKEYEVA / AFP via Getty ImagesShaman, a 19-year-old drone operator for the Russian military, flies a quadcopter drone during a demonstration event organized by members of the Berkut Military-Sports Cossack Club in a shopping centre in Voronezh, Russia on January 24, 2026.Credit:TATYANA MAKEYEVA / AFP via Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navRussian universities are promising free tuition and up to $70,000 to students who are willing to serve as drone pilots in the Russian military for a year—all while claiming students can avoid the risk of frontline combat duty in Ukraine. But there has already been one confirmed battlefield death and possibly more among the new cadre of student drone pilots.That specific recruitment offer appeared on pamphlets distributed at Bauman Moscow State Technical University, according toBloomberg. Other universities have dangled incentives such as tax holidays, loan forgiveness, and sometimes free land. The independent magazineGrozacounted at least 270 Russian academic institutions promoting military contracts to their students in the fifth year of the war that began with Russia’s full-scale invasion of Ukraine in February 2022.This new wave of recruitment is targeting a population of approximately 2 million men attending Russian universities, including gamers and students with technical skills that could make them suitable trainees as drone pilots, according to Bloomberg. Russia’s Defense Ministry has specifically called for drone pilot recruits with expertise in flying drones, model aircraft, electronics, and radio engineering, with computer skills also being desirable,NBC Newsreported.However, the effort carries the risk of further depleting Russia’s future educated workforce on top of the country’s existing brain drain—aresearch studyfound that 24 percent of top Russian software developers active on GitHub may have left the country within the first year of the war. Some students have also expressed a similar lack of enthusiasm for the Russian war effort. “No one wants to join,” a student named Andrey told NBC News. “No one is interested.”Nonetheless, Russia’s effort to recruit student drone pilots goes toward its goal of having 168,000 drone operators by the end of 2026, according to theKyiv Independent. In that sense, Russia is copying the success of Ukraine’sUnmanned Systems Forcethat became the world’s first standalone military branch focused on drones in June 2024.The Russian recruitment efforts have typically promised that university students can serve as drone pilots without risking their lives in bloody infantry assaults on Ukrainian trenches and fortifications. But safety is a relative term as constant surveillance and the threat of drone strikes or artillery fire has created a “kill zone” stretching as far as 25 kilometers on both sides of the frontlines, according to the commander of Ukraine’s Unmanned Systems Force in an interview withUkrainksa Pravda.The Russian-language news service ofBBC Newsidentified 23-year-old Valery Averin as the first known death among the new wave of Russian university students who trained and deployed as drone operators. Averin’s adoptive mother, Oksana Afanasyeva, was informed of her son’s death in a mortar attack on April 6 near the Russian-occupied city of Luhansk in eastern Ukraine.“The child had been training on a drone for three months, and now we’re throwing him into an assault, into the meat grinder, someone who had never served in the army,” Afanasyeva told BBC News.Russia has lost an estimated 1.3 million soldiers as battlefield casualties since the start of its full-scale invasion of Ukraine, according to aNATO officialcited by news reporting in February 2026. By comparison,Ukrainian casualtieswere estimated as being between 500,000 and 600,000 over approximately the same period, including killed, wounded, and missing soldiers.Into the meat grinderElite university students are still luckier than many Russian men—including trained specialists—who are being sent directly into frontline combat. Going back to 2023, the Russian space corporation Roscosmos was alreadyrecruiting its own employeesto join a militia for the war in Ukraine. By early 2026, Russia had organized three “motorized rifle regiments” for frontline duty by pulling personnel from its Navy, Aerospace Forces, and Strategic Missile Forces, according to the Ukrainian publicationEuromaidan Press.The Russian military’s d

Bad beatsThe US is betting on AI to catch insider trading in prediction marketsThe Commodity Futures Trading Commission wants us to know it’s taking this very seriously.Kate Knibbs, wired.com–May 16, 2026 7:00 am|95Credit:Yuki Shimazu / FlickrCredit:Yuki Shimazu / FlickrText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navFor most of the past year, it looked likeprediction marketshad kicked off a new golden age of fraud. OnPolymarket, traders raked in fortunes fromsuspiciously timed betson geopolitical events like the raid on Venezuela and the Iran War. It wasn’t clear whether the US government would bother pursuing some of the most flagrant bad actors, since Polymarket’s crypto-based platform was technically offshore and not regulated or licensed within the country.Now, however, the Commodity Futures Trading Commission, which oversees prediction markets, wants you to know that it’s watching very, very closely. The agency is searching for suspicious behavior from traders within the United States who have been sneaking onto offshore markets, including Polymarket’s crypto platform—which is blocked stateside—by using virtual private networks. “We’re going to find them, and we’re going to bring actions,” agency chairman Michael Selig told WIRED this week, speaking from the CFTC’s headquarters in Washington, DC.Selig says the agency, which is especially lean right now, is staffing up. Like so many other AI-pilled workplaces, the CFTC is also leaning into automation to handle the growing workload, including tools that analyze trading patterns and flag potential manipulation. “You’ve got so much data,” Selig says. “When we feed it into AI, we get really great information. It can help us understand things, like where we might want to investigate, or when we might need to send a subpoena to a trader.”In addition to proprietary surveillance systems developed in-house, the agency’s arsenal includes third-partyblockchain tracing toolslike Chainalysis for crypto platforms, and market abuse detection software including Nasdaq Smarts for centralized markets. (Beyond Nasdaq Smarts, the agency did not specify which AI tools it uses and declined to share more specific examples.)Prominent prediction market companies have recently started touting all the work they’re doing to catch sketchy bettors. US-based exchange Kalshi, Polymarket’s primary competitor, eagerly announced that it hassuspendedand penalized customersflagged for insider tradingand market manipulation.In April, after significant backlash over suspected insider trading, Polymarket announced its own partnership with Chainalysis. It was part of a broader push to crack down on market manipulation. While the company’s CEO, Shayne Coplan,had talkedin the past about why insider trading could be good for prediction markets, Polymarket changed its approach this spring, updating its market integrity rules and announcing a partnership with Palantir for its US-based sports markets (the Chainalysis deal focuses on the offshore platform). The company did not respond to WIRED’s request for comments for this story.According to Chainalysis spokesperson Maddie Kenney, the company analyzes the same data for both clients. “The value Chainalysis adds for our customers, including Polymarket and the CFTC, is organizing the data and enriching it with the attributions and insights we’ve accumulated over years in the space,” she says. Certainly sounds like a good deal for Chainalysis!The CFTC’s assurances that it is hunting insiders comes at a moment of intense scrutiny on prediction markets. In March, Connecticut senator Chris MurphytoldWIRED that he suspected White House staffers were engaged in insider trading on war-related contracts. At the beginning of April, seven members of Congressaskedthe CFTC to investigate overseas markets offering war-themed events contracts. In a letter, the lawmakers argued that the commission had the authority and responsibility to curb insider trading, especially on “morally obscene” trades on military action. Selig recentlytoldCongress that the company is pursuing “hundreds, if not thousands” of insider trading tips.Investigations are not limited to federally regulated exchanges. “We’re surveilling the markets on a global basis,” he tells WIRED.Selig says that the agency will exert extraterritorial jurisdiction—its legal ability to enforce its laws beyond traditional boundaries—when it finds suspicious activity on offshore platforms like Polymarket, though he says it’s a case-by-case approach. “We use it in extreme circumstances,” he says, with an eye towards whether charges have a strong chance of sticking in court. “In any extraterritorial litigation, there’s going to be challenges to our authority, and that could also impair our ability to bring cases in the future.” According to Selig, the 2010 Dodd-Frank Act allows the CFTC more leeway to pursue this kind of enforcement

striking resultsA revolutionary cancer treatment could transform autoimmune diseaseResearchers are testing CAR T cell therapy as a way to reset the immune system.Amber Dance, Knowable Magazine–May 17, 2026 7:00 am|55Artist’s rendering of a T cell.Credit:Getty ImagesArtist’s rendering of a T cell.Credit:Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAt age 49, Jan Janisch-Hanzlik’s multiple sclerosis was destroying her freedom to live the life she wanted. She gave up her active nursing job for a desk role. Frequent falls made her afraid to carry her grandchildren. She had to move to a bigger house to make room for the wheelchair she feared she might end up needing full-time.Even the best available medication wasn’t improving Janisch-Hanzlik’s symptoms, and she worried they’d only get worse. So when she learned about a trial of CAR T cell therapy at the University of Nebraska Medical Center in Omaha, close to the city of Blair where she lives, she phoned the clinic every other month until they were ready toenroll her as the first patient.Originally designed to target and wipe out cancer byreprogramming the patient’s immune cells, CAR T is now being offered to patientsin hundreds of clinical trialsfor autoimmune conditions like multiple sclerosis, lupus, Graves’ disease, vasculitis, and many others. The hope is that CAR T can duplicate the success it has demonstrated in a range of blood cancers by hunting down and eliminating cells that target the self in autoimmune diseases. This would essentially reset the body’s defenses to a state like the one that existed before the disease took hold.But along with CAR T’s promise come risks, questions, and challenges. There’s uncertainty about how well it will work forautoimmunityand how long any benefits might last, as well as what long-term side effects might arise. Janisch-Hanzlik knew this when she sat down to receive the experimental treatment on June 9, 2025; she felt a mix of hope and fear knowing that she would be spending the next week being monitored for side effects including dangerous inflammation.In addition to her clinical expertise and desire to pioneer a new treatment, Janisch-Hanzlik’s two young grandchildren helped inspire her pursuit of a treatment with known risks and uncertain benefits. Because multiple sclerosis has a genetic component, Janisch-Hanzlik knew that they have an elevated chance of going through the same struggle she has. “I would want to be able to say I did everything that I possibly could to prevent them, or anyone else, from having something like this,” she says.From cancer to autoimmunityThe first CAR T cancer treatment wasapproved by the Food and Drug Administration in 2017for an aggressive form of leukemia. Since then, the powerful and intensive treatment hasresulted in long-term remissionfor many cancer patients.The basic premise of CAR T is to activate the power of key immune cells calledT cells. T cells normallyrecognize other cellsthat have been infected by a virus or bacterium, or are otherwise abnormal, and either destroy them or recruit other parts of the immune system to do so.In CAR T for cancer, scientists engineer those T cells to specifically hunt and destroy malignant cells. The technologygot its startwhen cancer researchers figured out how to take out a patient’s own T cells, insert DNA instructions for a “chimeric antigen receptor,” or CAR, and put them back into the person’s circulation. The CAR, which sits on the T cell’s surface and latches on to a specific molecular partner on the surface of cancerous cells, activates the T cell to attack.Today CAR T cells are most commonly programmed to attack B cells, another key immune player. B cells are normally responsiblefor making antibodies, but in certain blood cancers, they proliferate out of control. By giving T cells a CAR that recognizes one of a couple of molecules unique to the B cell surface, the cells are reprogrammed to find and eliminate those cancerous cells.B cells also are the central problem in many autoimmune conditions: They mistakenly make antibodies against normal tissues instead of against invading pathogens. So as CAR T began to succeed against B cell cancers, it didn’t take long for doctors to reason that CAR T therapy might also be able to wipe out bad B cells in people with autoimmunity.There are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineThere are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineA German team pioneered autoimmune CAR T in a woman with lupus,reporting positive resultsin 2021.

两个月前在圣何塞，黄仁勋穿着皮衣站在 GTC 的舞台上，告诉全世界：Token 是新的大宗商品，生成 Token 的成本与效率，决定科技企业的营收与生死。昨天，李彦宏站在 Create 2026 的开幕式上，说了一句看似拆台的话：「Token 只是代表成本，并不代表收益。它衡量的是投入，而不是产出。」然后他抛出了一个新概念，DAA，Daily Active Agents，日活智能体数。卖铲子的人说，看铲子消耗量就知道金矿的繁荣程度。挖矿的人说，你倒是看看我挖出了多少金子。他们都没说错。但同一座金矿，用不同方式去称量，得出的故事完全不同。真正有意思的地方不在于谁对谁错，在于一个事实：AI 跑得太快了，快到这个行业连怎么给自己记账都还没想清楚。DAA ：从「烧了多少」到「干了多少」李彦宏的 DAA 逻辑并不复杂。移动互联网时代，衡量平台看 DAU，日活用户数。微信 13 亿，抖音 7 亿，Meta 34 亿。进入智能体时代，对应的指标应该是 DAA，有多少 Agent 每天在给人类干活，并交付结果。「交付结果」四个字是重点。Token 消耗量告诉你系统在转、算力在烧、钱在花。但它不告诉你这些 Token 到底干了什么。是帮创业者写了一份商业计划书，还是用户反复问「你好」然后 GPT-5.4 收了他 80 美元？这两种场景消耗的 Token 可能差不多，创造的价值天差地别。李彦宏说，「这比无谓的 Token 消耗，更接近价值，也更接近本质。」他还给了个大胆的预测：未来全球 DAA 可能超过 100 亿。一个人同时挂三五个 Agent 处理不同任务，你是一个 DAU，但贡献了三五个 DAA。Agent 数量远超人口，技术上已经可以想象。李彦宏自己也说这是一个「非共识的观点」。在所有人都在谈 Token 消耗量的行业里说 Token 不代表终局，是需要点勇气的。当然，也需要点动机。Token 经济学：一种极其性感的叙事Token 经济学的性感之处在于，它有大量数据支撑。截至今年 3 月，中国日均 Token 调用量超过 140 万亿，相比 2024 年初的 1000 亿增长了 1000 多倍。火山引擎豆包大模型日均 Token 使用量突破 120 万亿。OpenRouter 平台前十大模型的周 Token 调用量从 1.24 万亿暴增至近 14 万亿。一条陡峭到让人眩晕的增长曲线，核心驱动力恰恰是 Agent。过去 Chatbot 时代，单轮对话消耗 1000 到 3000 个 Token。Agent 来了之后，一个完整任务触发几十上百次模型调用，中等任务吃掉 10 万 Token，复杂任务上百万。Agent 越多，Token 消耗越大，算力需求越高，芯片卖得越好。这个闭环对英伟达来说无比优雅。2025 年全年营收 2159 亿美元，净利润 1170 亿美元。黄仁勋有充分的理由说：Token 就是新的石油。企业也在用脚投票。阿里成立 Token Hub 事业群，吴泳铭亲自挂帅。58 同城姚劲波说 Token 用得越多越好，不计成本。昆仑万维发内部信强制 AI Coding，达不到要求的末尾淘汰。黄仁勋甚至预言「你的 offer 带多少 Token」会成为硅谷新的谈判筹码。Token 消耗量作为指标，最大的好处是简单、可量化、跨平台可比。140 万亿就是 140 万亿，谁看都一样。但过亿简单也可能会掩盖一些问题。一个只盯着食材消耗量的餐厅，未必是一家好餐厅。共识大于分歧把 Token 经济学和 DAA 放在一起看，你会发现它们的共识其实大于分歧。这两个判断都基于 AI 正从对话阶段走向 Agent 阶段，而且都认为这是万亿级的产业重塑。分歧在哪？在于谁的账本更能代表这个行业的健康程度。Token 量的是投入侧：消耗了多少算力，工厂每瓦特产出多少 Token。这对芯片厂商和云厂商极其重要。DAA 量的是产出侧：有多少 Agent 在运行，完成了多少任务。这对应用平台方极其重要。百度做的是 DuMate、秒哒、百度一镜，卖点全是「帮你把事做完」。只有用「结果交付」来记账，应用层的价值才浮得出水面。屁股决定脑袋，但这不代表谁在说谎。每个玩家天然会推崇对自己最有利的记账方式。有意思的是，腾讯也站到了应用侧。腾讯云副总裁杨晨说过：「我们认为 Token 不是一个多么健康的生意，它看着量很大但黏性极差。」腾讯的核心资产是微信生态和内容场景，它在乎的是 AI 能不能在自家场景跑起来。阿里的态度截然相反，吴泳铭断言未来 5 年 AI 和云业务收入将超 1000 亿美元。同一个行业，同一批聪明人，看法完全不同。这本身就说明 AI 的迭代速度已经跑在了共识形成的前面。Token 消耗量像 GDP，量总规模，不管建了一座桥还是挖了个坑再填上。DAA 像就业率，量的是有多少「劳动力」在创造价值。GDP 高但就业率低，那叫空转。就业率高但产出质量差，那叫虚胖。一个健康的经济体两个数字都要看。但对普通用户来说坦率地讲，这两个指标都是术语。用户只在乎一个朴素的问题：好不好用？帮我省了多少时间？花了我多少钱？Token 经济学叙事下的 AI 产品，对用户其实不太友好。你用打车软件，起步价多少、每公里多少、堵车怎么算，一清二楚。你用 AI Agent，到底消耗了多少 Token，是哪个模型在跑，Agent「反思」了几轮导致费用翻倍，很多普通用户还搞不明白。DAA 至少提供了另一种视角：用户不该关心 Agent 烧了多少 Token，该关心它有没有帮你把事做完。前者是水表读数，后者是水龙头出不出水。造词运动的背后2024 年初中国日均 Token 调用量才 1000 亿。2026 年 3 月，140 万亿。1000 倍。这种速度下，任何已有的框架和认知都会瞬间过时。于是，每个玩家只能从自己的位置出发，尝试给这团混沌画一个坐标。黄仁勋造了「Token 工厂」，李彦宏造了「DAA」，阿里造了「Token Hub」……Token 衡量的是基础设施的繁荣程度，DAA 衡量的是应用价值的交付密度。两者更多是互补而非互斥。没有 Token 的生产和流转就没有 Agent 的运行，没有 Agent 交付结果 Token 消耗就只是空转。对用户来说，最好的度量衡是那个你根本不需要知道的。你打开水龙头，水来了。你把活交给 Agent，事做完了。你看一眼账单，觉得合理。至于背后消耗了多少 Token、动用了多少 Agent、芯片是谁家的、云服务用的哪一层，留给黄仁勋和李彦宏去争论就好。从 DAU 到 DAA，从「人在用产品」到「产品在帮人干活」，AI 时代的价值坐标正在被重建。争论是好事。当一个行业只有一种记账方式的时候，所有人都会围绕同一个数字做优化，不管那个数字是否真的通向价值。多一种衡量的维度，至少多一个纠偏的机会。至于最后哪套度量衡能留下来，答案可能不在黄仁勋和李彦宏手里，在你手里。你每天用 Agent 做了什么，做得好不好，愿不愿意继续付费。这些真实发生的选择，才是 AI 时代最诚实的投票。

未来我们的汽车哪怕前轮或后轮爆胎，剩下的轮子也能正常行驶。王传福 11 年前的这个构想，在今晚成为现实。方程豹这次将发布会搬到户外，正式推出豹 8 与豹 5 的「闪充版」车型。其中，全栈自研的云辇-P Ultra 全场景智能液压车身控制系统成为全场技术焦点。爆胎后车轮悬空，车辆仅靠三个轮子平稳开走；陷入泥坑无需求援，自行抬高被困轮胎完成脱困。这些曾经只存在于构想中的画面，方程豹今晚都一一呈现了出来。两年前方程豹以 DMO 平台入局，打破了新能源与硬派越野之间的壁垒。如今闪充版车型上市，方程豹用这套智能底盘技术，让越野时的爆胎脱困和泥地自救从理论变成了实际的能力。液压底盘，让方程豹告别千斤顶硬派越野的底盘路线通常有两种：被动悬架结构简单、可靠性高，可一到铺装路面，舒适性就打了折扣；空气悬架确实能改善平顺性，但气体可压缩的特性意味着，在高强度飞坡、碎石路这些深度越野场景里，气囊一旦破损，整套系统就可能失效。方程豹没选这两条路，直接把仰望 U8 上的液压车身控制系统搬了过来，还做了升级。方程豹汽车总经理熊甜波表示，为了打磨这套系统，研发团队把公司的一座厂房改成了专业越野测试场，一比一复刻了天门山越野场景，坑洼、碎石路面也都一应俱全。工程师还会洒水、喷泡沫，来模拟雨天等低附着环境。实验室之外，真实的反馈同样关键。有车主单人单车跑了 11 万多公里，穿越巴丹吉林沙漠和敦德冰川，研发团队从这些反馈里整理了 29396 条用户建议，不断修正底盘标定。经过上万次极限测试，云辇-P Ultra 最终落地。据介绍，整套系统的静态高度调节行程从 140 毫米提升到 200 毫米。日常城市里开迎宾模式，悬架最低能降到 90 毫米，老人小孩上下车更方便；到了极端越野场景，悬架最高能抬到离地 320 毫米。举升能力是液压悬架的另一项硬指标。方程豹这次解决了主动悬架容易超压失效的问题，极限举升载荷做到了 9 吨，差不多两头成年大象的重量。就算停在 45°的陡坡上，四个轮子也能独立精准控制，把车身抬起来。另外，越野时的高温散热是底盘研发躲不开的课题。这套悬架的整体散热面积是传统被动悬架的 12.7 倍，是单减震器的 3.7 倍。散热余量够大，沙漠里连续冲沙，阻尼也不会衰减。系统工作温度范围从 -40°C 跨到 85°C。机械底子搭好了，配合整车姿态闭环算法和四轮独立解耦控制技术，车辆就具备了自救能力。发布会最亮眼的那几项功能，都源自这套底层技术。「三足鼎行」应对的是突发险情，例如高速上单侧爆胎，或者山崖单行道会车。遇到这样的意外，系统会通过云辇计算中心动态分配四个减震器的油液载荷，收起其中一个车轮，车辆可以仅靠剩下三个轮子，以最高 15 km/h 的速度开到安全的路边。「腾空换胎」和「单轮脱困」同样靠整车姿态的独立调整，车主可以在中控屏或手机上操作，让陷入沙地的单侧车轮单独抬起，然后在悬空的车轮下方填沙、垫脱困板，车辆就能自己开出来。有了这个功能，在荒野里爆胎也不用费劲找平地架千斤顶了，轮子一提，一个人三分钟就能把备胎换上，过程中随时暂停，举升高度也能随意调节，一次提轮可以维持 24 小时以上。熊甜波还称，这套悬架还针对飞坡落地这种强冲击场景做了优化，系统的受力瞬间刚度能提升两到五倍，即便从两米的高度落地，落地后照样正常开，不用担心拉杆变形或者密封失效。非铺装路面之外，这套也兼顾实用与全场景驾驭。据介绍，新升级的刚度控制模块有三档，一档日常通勤，二档复杂混合路况，三档越野飞坡。双阀设计能够把阻尼带宽最大拉到原来的三倍，过减速带、压井盖这些细碎震动，过滤得更细腻。参数之上，现实之下补齐了底盘这块，新能源越野的基本盘依然要回到三电系统。熊甜波首先介绍的还是比亚迪第二代刀片电池和闪充技术。发布会公布的补能数据显示，豹 5 和豹 8 闪充版的电量从 10% 充到 70% 只要 5 分钟，充到 97% 也只需要 9 分钟。哪怕零下 30 度的严寒环境，充电时间也只多出 3 分钟。接着是安全问题，外界对越野车装电池的质疑一直没停过。官方表示，在提升了充电速度的同时，研发团队还提升了安全测试的标准，在实验室做了「边闪充边针刺」的破坏性测试。测试中，四节电芯同时短路，电池温控依旧平稳，没冒烟也没起火；面对 1500 千焦能量级的冲击测试，电池组的结构也经受住了考验。除了补能和安全，长途穿越也离不开长续航。讲到这里，现场的大屏幕上出现了从新疆乌鲁木齐开到甘肃敦煌的实测画面。测试中，豹 8 闪充版综合续航达到了 1380 公里，豹 5 闪充版也跑到了 1310 公里。而在智能化的比拼中，方程豹为两款新车规划了截然不同的辅助驾驶路线。旗舰定位的豹 8 闪充版搭载了华为乾坤智驾 ADS 4.1.5。前后双激光雷达加持，盲区感知精度提升明显。夜视系统升级后，野外没有路灯的暗夜里，车辆也能看清行人和横穿道路的小动物。豹 5 闪充版用的是天神之眼 B，这套方案依靠 Momenta 的算法和比亚迪庞大的行车数据迭代，日常城市通勤和复杂路况处理得也不错。座舱和周边生态上，两款车的目标人群区别更明显。豹 8 闪充版有大 5 座、大 6 座和 7 座三种布局，外观新增了悬浮顶双拼色，车内大面积用了 Nappa 真皮和 50 年树龄的白酸枝木。18 扬声器帝瓦雷音响加上奇华顿车载香氛，豪华感拉得很满，后排还带腿托和按摩。追求野外可玩性的车主，还可以花费 16968 元选装由比亚迪与大疆联合开发的智能车载无人机系统，在越野途中完成一键跟拍。另一边的豹 5 闪充版在保留家族式升降挡把与水晶按键的基础上，推出了以大漠夕阳为灵感的「赤砂红」专属车身配色。为了照顾越野老炮的改装需求，新车加了 OCS 接口。射灯、气泵这些外接设备不用破线就能直接接入，电流三档可调。车内还定制了一套隐藏式智能车载电台，主机藏在座椅下面，通过车机语音就能一键呼叫或发起定位救援。发布会临近尾声，售价正式公布：豹 8 闪充版：41.98 万元；豹 5 闪充版：30.58 万元。整场发布会看下来，全场景液压悬架的技术冗余、个位数分钟的补能速度，再加上座舱的豪华堆料，的确大幅拉高了新能源硬派越野的制造门槛。但回到终端市场，目前真正走量的是豹 5 210km 天神 Max 版，置换价格已经来到了 23.98 万元，和闪充版差了一大截。眼下车市卷得厉害，消费者对价格极其敏感。普通越野玩家愿不愿意为多出来的脱困能力和快充体验，去填平这道价格差，还得观察一段时间。不过，抛开终端定价与市场博弈不谈，单看这场硬核的新车发布，方程豹确实把研发团队压箱底的技术底牌全盘托出了。发布会最后，熊甜波坦言自己曾纠结是否要讲这么多枯燥的「物理课」和底盘参数，但她最终决定把工程师们夜以继日研发的成果完整呈现出来。每一项技术的突破，每一组硬核的参数，都藏着他们日夜坚守，默默付出的心血。他们的专注与坚守，值得被所有的人看见，被所有的人记住。这些从零起步、硬磕出来的技术堡垒，能否顺利转化为切实的销量增长，接下来的交付会反馈最真实的答案。

尊界和玛莎拉蒂要合作造车了。根据《云见 Insight》的报道，华为鸿蒙智行、江淮汽车、斯特兰蒂斯集团（Stellantis）及旗下玛莎拉蒂品牌，正就共同打造玛莎拉蒂品牌的新能源车展开洽谈。联合开发的车型将分为国内和海外版本，国内版归属于尊界品牌，海外版将挂玛莎拉蒂车标。首款车型目前正处于造型设计阶段，计划于明年下半年量产。按照规划，四方合作模式与鸿蒙智行 「五界」 高度相似：华为主导产品定义、提供核心技术，江淮联合研发并负责生产制造，玛莎拉蒂提供造型设计与品牌背书。这套分工，仍然延续了鸿蒙智行过去的基本方法：华为提供智能驾驶、鸿蒙座舱、电驱的技术和渠道能力，合作方提供制造、供应链和品牌资源。但这套熟悉的配方里，玛莎拉蒂的出现，带来了一个全新的变量。玛莎拉蒂负责提供的，并非某种可以量产的硬件或软件，而是一种华为过去很少主动去借用、甚至很难在短期内自我孵化的外部资源——更加年轻品牌性格和情绪表达。在目前的鸿蒙智行合作体系中，华为提供的是一套极其成熟且标准化的智能汽车能力。这个体系的运转效率极高，它不仅让华为迅速建立起了一个横跨二三十万到百万级价格区间的庞大汽车产品矩阵，也成功在消费者心里植入了一个清晰的等号：买鸿蒙智行，买的就是华为的智驾、座舱、主动安全和渠道服务。销量领跑超豪华轿车市场的尊界 S800 就是这套模式的集大成者。途灵龙行底盘平台、面向 L3 级自动驾驶架构设计的智驾系统、星河卫星通信技术、全方位的主动安全防护体系，以及最新一代的鸿蒙座舱，搭配上双拼色车身、大面积的深色皮质、木纹饰板。这套以智能化为基底，克制、稳重、偏向于商务豪华的高端表达，确实能精准地吸引到一批人。但高端市场里还有另一批人。这群人买车，更看重车辆的车身比例、姿态、线条、品牌气质，他们或许会被智能化打动，但要想让他们买单，光有聪明的脑子不够，还得有迷人的皮囊和动人的情绪。玛莎拉蒂曾经就擅长提供这种味道。2017 年，玛莎拉蒂在中国市场销量达到 14498 辆，成为品牌全球最大的单一市场。那几年，Levante 和 Ghibli 在中国卖得好，靠的并不是更强的智能配置，也不是更理性的产品力。消费者买的是三叉戟标识、意大利设计、跑车血统，还有一种和 BBA 不太一样的豪华感。它比法拉利、兰博基尼这些顶级超跑更懂生活，更容易接近；又比传统的德系豪华车更有情绪，更懂风情。很多人选择玛莎拉蒂，本质上不是在做一道产品力对比的算术题，而是在做一种审美选择。虽然后来它在中国快速掉队，但主要原因是新能源车时代来得太快，智能驾驶、座舱体验、电动化效率和整车软件能力，重新冲刷了传统豪华品牌原有的定价逻辑。玛莎拉蒂推出过电动车，但没有建立起足够鲜明的技术认知，也没有拿出一台真正让中国用户重新兴奋的产品。然而，销量的下滑，并不意味着品牌记忆的彻底消失，更不意味着这种消费需求不复存在。在中国这片庞大的市场里，依然有一群人，他们骨子里还是喜欢意大利设计的随性、跑车血统的张扬，以及那种更有情绪的豪华表达。这也是这次合作的基础。华为和江淮可以解决新能源时代的产品力问题：智能驾驶、座舱、电驱、制造、供应链。玛莎拉蒂可以继续输出它在中国高端消费市场里积累了二十年的审美叙事：比例、姿态、线条、品牌情绪，以及三叉戟带来的想象空间。换句话说，华为解决「这台车够不够强」，玛莎拉蒂解决「这台车有没有味道」。对华为来说，这更像是一次高端品牌方法论的调整：当技术已经足够强，华为开始意识到，除了技术之外，还可以尝试打开一扇「情绪」的门。其实，华为在智能手机业务上，早就摸透了这个道理，甚至完整经历过类似的周期。2024 年，华为做了一个重要决定，把主打影像和设计的 P 系列全面升级为 Pura 系列。这绝不仅仅是改个名字换个包装那么简单。Pura 系列在外观设计上，非常大胆地采用了三角形作为核心视觉符号。在手机圈，三角形比传统的方形更尖锐，比柔和的圆形更有攻击性，也比那些不规则的曲线更容易在消费者脑海中钉下识别度。这个超级符号，让 Pura 瞬间和过去的 P 系列拉开了距离，也为华为的影像旗舰留下了更先锋、更大胆的自我表达空间。当时的智能手机行业，和今天的新能源汽车市场有着惊人的相似。国产手机厂商已经在硬件参数、快充速度、屏幕素质、芯片性能上卷到了极致，大家都在追求一套标准的「最优解」。结果就是，市面上的旗舰机盖住 Logo 后，长得越来越像，体验也大同小异。今天的尊界，其实也走到了当年 Pura 面临的那个十字路口。过去几年里，「五界」体系通过问界、智界、享界等品牌，已经向市场清晰地回答了一个最基础的问题：华为到底能不能把车造好？但如果尊界和玛莎拉蒂的这次合作能够顺利推进，它必将成为鸿蒙智行体系里最特别、也最具野心的一次试验。因为接下来，尊界要向金字塔尖的用户群回答一个更难、也更感性的问题：当一台车在技术上已经足够聪明的时候，它还能不能做到足够迷人？

1968 年的旧金山，计算机科学家道格拉斯·恩格尔巴特在一场后来被称为「演示之母（The Mother of All Demos）」的发布会上，拿出一个带着两个金属轮子的木制小盒子，向世界介绍了一个新物种：鼠标。那是人类第一次公开用手上的鼠标，去牵引屏幕里的数字光标。此后的几十年里，这根小箭头几乎无处不在。它穿过办公软件、游戏界面、浏览器窗口和无数张电子表格，成为人类进入数字世界时最熟悉、也最沉默的向导。然而在这半个多世纪里，计算机的算力、形态和应用场景几乎改头换面，鼠标光标的本质几乎没有变过：它知道自己停在屏幕上的哪个坐标，知道 X 和 Y，却不知道你指向的是一行代码、一张发票，还是一张风景照。面对眼前不断闪烁的像素，它能做的事情依旧相当朴素：点击，拖拽，等待下一次点击。今天，Google 要用 Gemini 重新发明鼠标光标。在刚刚落幕的 Android Show 上，Google 几乎把围绕 Android、AI 与硬件生态的规划摊在了台面上。其中，一项名为「魔法指针（Magic Pointer）」的新功能，给古老的鼠标光标装上了「眼睛」和「大脑」。Google 的意图不言而喻，未来的 AI 交互不该依赖冗长的提示词，只需像在现实生活中一样，指着屏幕说一句：「把这个，移到那里。」那么问题来了，当鼠标光标终于学会「看懂」屏幕，它究竟会把人机交互带向哪里？这根睁开眼睛的 AI 箭头，到底能干嘛？要理解这项技术的意义，必须先看清当下 AI 工具最别扭的一面：交互成本。过去几年，大语言模型的能力一路狂飙，但使用门槛居高不下。为了让 AI 准确理解意图，用户被迫学习一套复杂的「提示词工程」：设定角色、补充背景、限定输出格式。为了一个简单的需求，写出几百字小作文更是家常便饭。不仅如此，典型的 AI 工具通常运行在独立的网页或应用窗口中，频繁打断用户的工作流。比如当你在阅读一份 50 页的 PDF 时想让 AI 总结一张图表，通常需要经历：截图 -> 保存 -> 打开浏览器 -> 进入 AI 网页 -> 上传图片 -> 输入提示词。Google 将这种繁琐的跨应用操作称为「AI 绕路（AI detours）」。这种跳转不仅效率低下，也容易打断人们工作时的注意力集中状态，即所谓的「心流」。为此，Google 提出的第一个交互原则，就是「保持心流」。在他们展示的实验性 AI 光标原型中，AI 的能力不再局限于某个特定的 App 或网页，而是附着在鼠标光标上，随时待命。触发方式也尽量克制：无需记忆任何快捷键，只需轻轻「晃一下」鼠标，AI 界面便会基于当前悬停的内容自动浮现，给出极具语境的操作建议。选中图，它会询问是否「对比」；悬停于段落之上，它会主动提供润色方案。整个过程没有任何需要学习的指令，完全跟着直觉走。来看几个极其直观的场景：第一，看图说话的终极形态。当你在浏览一张卡通的城市风景照时，传统鼠标只能点击图片放大。但现在，你只需把 AI 光标悬停在照片背景里的一栋建筑上，然后对着麦克风说一句：「给我移动图片的元素到这。」不需要解释「这儿」是谁，也不需要描述建筑外观。AI 光标会直接理解你所指向的像素，识别出对应元素，并成功移动。过去，鼠标只能告诉系统「我点了哪里」；现在，它开始告诉系统「我指的是什么」。第二，少写提示词，多用自然指代。当你在网页上看到一份极其复杂的烘焙菜谱，你不需要复制粘贴，也不需要写「请将以下菜谱里的所有食材分量乘以二」这种书面语。你只需要用光标高亮那段文字，然后随口说一句：「把『这些』的分量翻倍。」唰的一下，AI 直接在原地给你改写了一份新菜谱。第三，将像素转化为可交互实体。在计算机眼里，屏幕只是几百万个发光的像素点。但 AI 光标能将死板的像素（Pixels）转化为活的实体（Entities）。比如，你正在看一段旅行 Vlog，视频里闪过一家看起来很棒的餐厅。你按下暂停，把光标指过去，原本死气沉沉的视频画面瞬间变成了一个真实的、可交互的地点，旁边直接弹出了这家餐厅的订座链接。再比如，你随手拍了一张写满鬼画符的便利贴，鼠标一指，墨迹直接变成了一个可以打勾的 To-Do List。发现了吗？ 以前，是你去找 AI；现在，是 AI 顺着你的鼠标，乖乖来到了你的指尖。杀死 AI 提示词，回归人类直觉仔细想想，人类最强大的沟通工具其实是代词。当你和同事坐在屏幕前修改设计稿时，你绝对不会字正腔圆地说：「请将屏幕左上角坐标 (X:120, Y:350) 处的蓝色矩形向右移动 50 个像素。」你只会指着屏幕说：「把这个，往右挪一点，弄淡一点。」「那个餐厅看起来不错，怎么去？」「这段代码里的这个报错是什么意思？」在日常生活中，我们极度依赖「这个」和「那个」。手势配合着极简的口语，才是人类最高效的沟通密码。究其原因，我们身处同一个物理空间，共享着同一套视觉上下文。Google 敏锐地抓住了这一点，并将其提炼成了一条产品原则：拥抱「这」与「那」的力量（Embrace the power of This and That）。与其强迫人类去学习复杂的提示词框架，不如反其道而行之，把表达意图的脏活累活从我们身上剥离出去，让机器去适应人类最慵懒、最本能的「指手画脚」。好消息是，这套交互方式已经开始落地。Chrome 浏览器里的 Gemini 从今天起率先支持；Google 全新推出的笔记本电脑产品线 Googlebook，则将「Magic Pointer」直接内置进了操作系统层面，覆盖所有应用。Googlebook 的野心不止于鼠标。Google 将这个产品线定义为「Android 手机的完美伴侣」。类似苹果的 iPhone 镜像，用户可将 Android 应用无缝投射到 Googlebook 桌面，以原生比例运行，并在文件管理器中跨设备自由穿梭，彻底打破手机、平板与笔记本的生态壁垒。此外，Gemini 还能依你所需，在桌面生成专属动态 Widget（比如旅客的实时航班卡片）。在硬件设计上，所有 Googlebook 机型都会在机身上集成一条「Glowbar」发光条，让你一眼就能分辨出它与传统 Chromebook 或 Windows 笔记本的区别。首批 Googlebook 将由宏碁（Acer）、华硕（Asus）、戴尔（Dell）、惠普（HP）和联想（Lenovo）制造，预计今年秋天上市。有意思的是，三星缺席了这份名单。近期的消息显示，三星可能正在筹备一款搭载 Google 新系统的 Galaxy 笔记本，其下一场 Unpacked 发布会据传定在 7 月 22 日。至于底层的驱动核心，Google 虽未指名道姓，但通篇强化的「为智能而生的现代操作系统」以及 Android 与 ChromeOS 的深度融合，种种迹象均指向了传闻已久的 「Aluminum」系统。这意味着，AI 开始成了操作系统级别的基础设施。而当 AI 真正化身为你的鼠标光标，它便拥有了介入万物的权限——所见即所得，所指即所控。AI 人机交互，迎来十字路口回望 1968 年，那个惊艳世界的初代鼠标，功能简单得令人发指：追踪位置。 这五十多年来，鼠标加了滚轮、加了侧键、甚至加了风扇和配重块，但它的灵魂依然是一张白纸：它准确地标记着坐标，却永远无法理解坐标背后的意义。Google 的 AI 光标完成了交互史上罕见的进化：它不仅知道你在哪，更知道那是什么。过去一年，无数拿了融资的创业公司挤破头，试图造出下一个「AI 时代超级入口」。大家疯狂地卷对话框的拟真度、卷 Agent 的复杂工作流。但 Google 这次用行动给全行业结结实实地上了一课：最好的科技是什么？是润物细无声。聊天框（Chatbox）从来都不是 AI 的最终形态，它只是一种过渡期的妥协。最好的 AI，应该退居幕后，成为一种附着在你日常动作上的基础设施，而不仅仅是某个需要单独打开的应用。从黑底白字的命令行（CLI），到图形界面的鼠标点击（GUI），再到移动时代的触屏滑动（NUI）。过去几年，大语言模型让我们短暂地倒退回了打字交流的时代，让无数人患上了 Prompt 焦虑症。但今天过后，我们知道了，那不过是黎明前的一段弯路。真正好用的 AI，终究要学会像人类一样思考：看懂你的每一个眼神，听懂你的每一句「把这个，放到那儿去」。58 年前，当道格拉斯·恩格尔巴特握住那个简陋的木制鼠标时，他的终极梦想是「增强人类的智慧」。58 年后，当 AI 附身于这个古老的指针，机器终于开始真正「看懂」这个世界。属于提示词工程师的时代终将落幕，人机交互的终极闭环，也将在一句句含糊不清的「这个、那个」中，向前迈出历史性的一大步。附体验地址：https://aistudio.google.com/apps/bundled/ai-pointer-create?showPreview=true&showAssistant=true&fullscreenApplet=truehttps://aistudio.google.com/apps/bundled/ai-pointer-find?showPreview=true&showAssistant=true&fullscreenApplet=true作者：莫崇宇

📱iPhone 17 Pro 官降 1000 元🧠ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施🤖估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相🌐Claude 重新允许接入龙虾，开发者：这是变相涨价📋腾讯辟谣「AI 一号位将离职」💰黄仁勋基金会捐出 1.08 亿美元 AI 算力📱影石 CEO 回应 Luna 定价贵：5299 元是美国价格🚗鸿蒙智行洽谈打造玛莎拉蒂新能源车🚙华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗💻英特尔成为迈凯伦官方计算合作伙伴⛽本田上市以来首次年度亏损📋LinkedIn 宣布裁员，涉及工程、产品和市场营销部门💡Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」🤖曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线📱荣耀 600 系列定档 5 月 25 日🔍Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光🧠腾讯开源 AI Agent 记忆系统，Token 消耗最高降低 61%🧸泡泡玛特否认进军传统家电，LABUBU 冰箱一季度收入约 600 万iPhone 17 Pro 官降 1000 元昨晚，苹果无预警在天猫「Apple Store 官方旗舰店」启动 iPhone 17 Pro 系列 618 优惠活动，Pro 和 Pro Max 同步降价 1000 元，起售价分别来到 7999 元和 8999 元。本次活动还覆盖 iPad、Apple Watch、AirPods 等系列，非 iPhone 产品可使用 50 至 100 元不等的惊喜券，并可叠加 88VIP 消费券、店铺优惠及国家补贴等多重优惠。ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施据彭博社报道，苹果与 OpenAI 的合作关系已出现明显裂痕，OpenAI 正在联合外部律所评估可能的法律行动。知情人士透露，OpenAI 律师团队正与外部律所合作，研究包括向苹果发送违约通知在内的多项方案，但目前尚未正式提起诉讼。OpenAI 仍希望在庭外解决与苹果的分歧，相关法律行动预计将在马斯克相关诉讼案结束后才会推进。苹果与 OpenAI 的合作始于 2024 年 6 月，苹果在 WWDC 上将 ChatGPT 整合进 Siri 及 iOS 系统，涵盖文字生成、视觉智能、图像生成 App「图乐园」等场景。估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相昨天，前 Meta FAIR 研究院总监田渊栋正式宣布联合创办 AI 初创公司 Recursive Superintelligence，距其确认从 Meta 离职刚好过去半年。他在 X 发文称，拒绝了 OpenAI、xAI、Anthropic 等多家头部 AI 公司的邀约，选择自己创业：「趁着自己还年轻，去当一家新初创公司的联合创始人。」Recursive 同步披露完成一轮 6.5 亿美元融资，估值达 46.5 亿美元，由 GV（谷歌风投）和 Greycroft 领投，AMD Ventures 与英伟达跟投。公司押注的核心方向为「递归自我改进」（Recursive Self-Improvement），即构建能够自动发现知识、持续优化自身的 AI 系统。Claude 重新允许接入龙虾，开发者：这是变相涨价Anthropic 昨天官宣，将于今年 6 月 15 日起为 Claude 订阅版用户引入独立的 Agent SDK 月额度，OpenClaw 等第三方 Agent 工具由此重新获准通过订阅账户接入 Claude。按照新规，Agent SDK 的调用将从用户的常规订阅限额中彻底剥离，另设专属月度额度池，额度按账单周期刷新，未用完的部分不结转至下月。Pro 版和 Team 标准席位每月获得 20 美元额度；Max 5x 及 Team 高级席位为 100 美元；Max 20x 与 Enterprise 座位制高级席位最高可达 200 美元。该额度专用于 Claude Agent SDK 项目（Python 或 TypeScript）、claude -p 非交互命令、Claude Code GitHub Actions 集成，以及通过 Agent SDK 认证的第三方应用。此次新规被视为 Anthropic 对今年 4 月 4 日封禁政策的正式解决方案。彼时，Claude Code 负责人 Boris Cherny 宣布，Claude Pro 及 Max 订阅额度即日起不再覆盖 OpenClaw 等第三方工具的调用，理由是此类工具的使用模式对 Anthropic 的算力基础设施造成了难以为继的压力。开发者社区对此反应冷淡。有开发者直言，20 美元的 Agent 额度在高强度任务下消耗极快，实际可用量远低于封禁前的订阅套餐，并将此次调整定性为「变相涨价」。腾讯辟谣「AI 一号位将离职」昨天，腾讯官方发文辟谣，明确否认外界流传的「AI 一号位即将离职」传言。腾讯公关总监张军表示，「太恶劣了，还编出各种搞笑的情节。」谣言起源于社交平台上的多条帖文，内容称某大厂 AI 一号位因被绿色事业部抢走算力资源而准备离职，并点名影射腾讯 AI 负责人姚顺雨。黄仁勋基金会捐出 1.08 亿美元 AI 算力据路透社报道，英伟达 CEO 黄仁勋与其妻子洛丽创立的基金会，从云计算公司 CoreWeave 购入 AI 算力，并将其捐赠给多所大学及其他非营利研究机构，用于科学和人工智能研究。根据昨日提交的一份文件，此次捐赠的算力价值目前已达 1.083 亿美元。英伟达还表示，计划向部分受捐机构提供免费工程服务。CoreWeave 是一家专注于 AI 应用的云计算公司，其向客户提供的 GPU 正是由英伟达设计制造。今年 1 月，英伟达向 CoreWeave 投资 20 亿美元，一度成为该公司第二大股东。影石 CEO 回应 Luna 定价贵：5299 元是美国价格昨天，影石 CEO 刘靖康回应旗下旗舰双摄云台相机 Luna Ultra 的定价问题，表示网传单机 5299 元的售价属实，「但是是美国的价格，国内给我再大的胆子都不敢卖这个价格」。此前，博主「数码闲聊站」爆料称该机预计售价 5299 元、全能套装预计 6499 元，随即在网络上引发「定价过高」的质疑。鸿蒙智行洽谈打造玛莎拉蒂新能源车据《云见 Insight》报道，华为鸿蒙智行、江淮汽车、斯特兰蒂斯集团（Stellantis）及旗下玛莎拉蒂品牌，正就共同打造玛莎拉蒂品牌的新能源车展开洽谈。按照规划，四方合作模式与鸿蒙智行 「五界」 高度相似：华为主导产品定义、提供核心技术，江淮联合研发并负责生产制造，玛莎拉蒂提供造型设计与品牌背书。合作车型将分为国内和海外版本，国内版归属于尊界品牌，海外版将挂玛莎拉蒂车标，首款车型目前正处于造型设计阶段，计划于明年下半年量产。华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗华晨宝马昨日公布年度节水数据，显示其沈阳生产基地 2025 年单台生产取水量降至 1.4 立方米/台，同比降幅 10.8%，连续八年稳步下降。在水资源消耗总量方面，华晨宝马沈阳基地全年取水总量降至 75.2 万立方米，较 2023 年降幅超 40%，连续三年递减。再生水使用量超 68 万立方米，同比增长 71.9%，官方称节约用水量可满足近 9300 户沈阳家庭一整年的生活用水需求。华晨宝马大东工厂北厂区实现所有生产工艺自来水消耗为零，全部改用再生水替代；铁西工厂及里达厂区涂装车间冷却塔也已扩展使用市政再生水，厂区绿化灌溉此前已于 2024 年实现 100% 使用市政再生水。英特尔成为迈凯伦官方计算合作伙伴英特尔昨日宣布与迈凯伦车队达成多年期战略合作伙伴关系，正式成为迈凯伦万事达一级方程式车队、艾睿迈凯伦印地车队及迈凯伦 F1 模拟赛车队的官方计算合作伙伴。根据合作协议，英特尔至强处理器和酷睿 Ultra 处理器将为迈凯伦的核心性能负载提供算力支持，覆盖计算流体动力学、空气动力学分析、车辆动力学仿真、比赛策略分析等方面。英特尔的技术还将应用于赛道边缘计算、数字孪生与 AI 预测建模等场景。本田上市以来首次年度亏损昨天，本田公布了 2026 年 3 月期的合并财报，期内营收 217966 亿日元，同比微增 0.5%；归母净亏损 4239 亿日元，上一财年同期归母净利润为 8358 亿日元，由盈转亏。亏损主因是电动化战略大幅调整。本财年内，本田决定取消北美多款电动车型的开发与上市计划，由此确认减值、处置及合同赔偿等相关损失合计约 15777 亿日元，全部计入汽车业务。摩托车业务为本财年唯一实现正向盈利的板块。展望明年，公司预计营收 231500 亿日元，归母净利润恢复至 2600 亿日元。LinkedIn 宣布裁员，涉及工程、产品和市场营销部门据彭博社报道，职业社交平台 LinkedIn 昨日宣布裁员，涉及工程、产品和市场营销等多个职能部门。LinkedIn CEO Daniel Shapero 在发给员工的备忘录中表示，公司需要为用户提供更大价值，并实现更高的盈利水平，此次组织架构调整旨在为公司未来发展做好定位。LinkedIn 发言人随后确认了相关变动，称此次调整属于「常规业务规划」的一部分，但未披露具体裁员人数。💡 Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」据《商业内幕》昨日报道，Anthropic CFO Krishna Rao 在播客节目《Invest Like the Best》中透露，Claude 目前已承担公司 90% 以上的代码编写工作，并正在深度介入财务流程，推动白领工作从「执行」向「监督」转型。Rao 还透露，公司已借助 Claude 生成财务报表，每月财务审查流程在人工介入之前已达到「90% 至 95% 完成」的状态。他提到，部分内部报告的生成时间从原来的数小时压缩至 30 分钟。在人员配置方面，Rao 将 AI 的引入描述为生产力「加速器」。他称，由于 AI 的介入，员工得以将更多精力从信息收集转向决策判断，Anthropic 反而因此扩大了招聘规模。我们雇用了更多人，因为没有什么工作是做不完的。Rao 还描述了一种新的工作形态：员工越来越多地扮演 AI 系统的「管理者」角色，团队可以同时部署多批 AI 智能体并行处理项目。他将这一趋势概括为「每个人都在某种程度上成为管理者」，并认为由此带来的生产力潜力「才刚刚开始显现」。他同时强调，Anthropic 内部推崇「人才密度」而非「人才规模」，认为高密度的顶尖 AI 研究人才配合最强模型，才是真正具有竞争力的组合。曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线知名爆料者 Leo 昨日披露，GPT-5.6 研发已全面推进，预计最早下月正式发布。开发者在 OpenAI 内部 Codex 日志中发现了两个内部代号 —— ember-alpha 和 beacon-alpha，以及一条路由映射记录。日志显示，绝大多数模型调用仍指向 GPT-5.5，但有一条条目明确指向 GPT-5.6，暗示 Codex 环境可能已在测试新模型。此外，爆料者 Chetaslua 称 OpenAI 计划于本周四推出旗舰模型的「超快模式」，响应速度或可提升 2 至 3 倍，主要面向延迟敏感型任务。荣耀 600 系列定档 5 月 25 日荣耀昨日宣布，旗下荣耀 600 系列将于 5 月 25 日 19:00 正式发布。官方预热信息显示，荣耀 600 系列主打「4K 闪光微单 Live」影像方向，搭载 2 亿像素超清大底主摄，并内置 8600mAh 青海湖大电池。Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光据 Sources 报道，Google 计划在 5 月 20 日举行的 Google I/O 大会上

相信这样的场景很多人都不会陌生：在微信上聊工作、讨论旅行计划，甚至只是和朋友闲聊，聊着聊着突然想让 AI 帮忙总结重点、做个规划，结果却只能一段一段复制聊天记录，再手动粘贴进 AI 对话框里。再比如，中午小憩了一会，醒过来发现群里发了几百条消息，想吃瓜又得一条一条翻……如果能一键将聊天记录转发 AI，岂不美哉？这样的功能终于来了，但遗憾不是在微信里：现在，把微信和元宝更新到最新版本之后，聊天记录转发可以选择直接发给元宝来总结。虽然同属腾讯生态，但微信和元宝之间始终有种微妙的距离感——终于，这次两者之间的联动，又紧密了一点。爱范儿也第一时间体验了这个功能，看看它究竟好不好用。产品很美好，效果很一般要想把微信聊天记录转发给元宝，方法很简单，首先选中想要转发的聊天记录，选择「转发到其他应用」，就能直接发给元宝，可以补充一些提示词。整个过程有点像把讨论、聊天过程中零散的内容，让 AI 帮忙进行整理，因此这个功能首先最合适的场景，就是各种工作群的讨论和安排信息。直接处理聊天记录有一个好处，元宝可以将讨论精准和发言者匹配，很适合用来处理会议纪要和工作安排。旅游计划同样也是一个高频场景，因为行程一般都是靠你一言我一句的对话讨论出来，这时候就可以直接交给元宝处理，还能让它帮忙推荐旅游地点。还有一个有点邪修的用法：平时工作忙没时间看群，又担心错过什么热点或者话题，可以直接选中所有聊天记录，让元宝帮你看群帮你总结。又或者讨论了很久都不知道聚餐吃什么，也可以把讨论的过程选中，让元宝帮忙推荐。尝试了很多场景，大部分情况下都达到了预期，但总体来说，还有一些体验细节不够完善。目前元宝最擅长处理的聊天记录以文字为主，而如果聊天记录中包含文件，特别是涉及多个文件，效果会比较随缘，很多时候会无法读取文件。即使能成功读取，也不一定能按照提示词进行任务处理：整理发票有时候都未必能填对，让他根据文件做个思维导图，也经常失败。要知道，整理聊天里的发票真的是很主流的场景啊！我能很明显感觉到一个落差：明明将聊天记录发送给 AI 是一个很有价值的场景，但元宝交付的结果，总让我略感失望，并且忍不住想，如果能转发给豆包或者 ChatGPT，会不会更好？更深层的问题，是这个聊天记录转发的交互总给我一种「绕远路」的感觉。在此之前，元宝已经作为联系人存在于微信之中，处理聊天记录这件事，却依旧需要「跳转」这一整套外挂式流程， 过程中需要多次点击。或许是因为微信严格的隐私保护规则，元宝只能以「临时聊天」的形式处理聊天记录，不会留下历史记录，关闭后无法找回。对于微信来说，这样的功能似乎「可有可无」。如果微信想要 AI 化，他们或许会有更好的实现方式，例如直接做一个「AI 群友」，at 它就能回答问题或者处理聊天记录；以及一个常驻对话列表的 Agent，可以直接对话，可以直接接收聊天记录。但这件事没那么容易，原因已经是「老生常谈」：微信需要简约，需要克制，因为每一点小变化都会影响到十亿级用户。在元宝这个功能上线的同一天，腾讯发布了今年第一季度财报，在股东大会上马化腾表示，在微信里做 Agent 需要「更长远的考虑，大家要有点耐心，这个不是匆忙就能出来的。」腾讯的 AI 王牌是微信马化腾还提到腾讯的 AI 战略：原来一年前我们以为上了船，后来发现那个船漏水了，现在感觉站上去了，还坐不下去，还是希望船速能快一点。从财报来看，AI 已经成为了腾讯的重要投入，一季度研发投入达 225.4 亿元，同比 19%；资本开支 319.4 亿元，主要用于算力、存储等云端基础设施的扩张。今天中午，有自媒体传出「疑似腾讯 AI 一号位即将离职」的消息，下午就被腾讯公司火速辟谣，本身也向外界释放出一个信号：腾讯 AI，不容有失。根据 AICPB 最新榜单，在国产 AI 应用排名中，腾讯元宝排名第十，访问量为 1773 万，远逊于豆包（1.63 亿）和千问（4424 万）。在 AI 领域中，腾讯和苹果有着相似的处境：产品入口和用户规模都是顶级配置，AI 模型本身却还支撑不起这种期待。元宝 AI 社交应用「元宝派」生成式 AI 火了 3 年，从去年年底开始，AI 应用之间的战斗已经进入了「场景化」的比拼，不仅出现「龙虾」OpenClaw 这样的全新形态，传统的 ChatBot 交互模式也集成了生活服务，像阿里千问和字节豆包，都在积极将电商等各种生态内的优势服务，和自家的 AI 助手相结合。这本来应该是腾讯的舒适区，因为国内最强大、最全能的生活应用，也就是微信，就在腾讯手中。但很明显，微信对「AI 化」抱有非常慎重的态度。腾讯不能承受因技术或功能不成熟而扰乱微信庞大用户群的体验。只是对于腾讯来说，在 AI 性能落后其他对手的情况下，微信这张王牌的价值还是太重要了。，于是，我微信和元宝虽然没有完全敞开双臂拥抱彼此，也开始小心翼翼试探边界。究其原因，除了微信相对封闭的生态逻辑，还有两个问题没能被回答：微信究竟要如何「AI 化」？而元宝在其中又应该扮演什么样的角色？The Information 报道指出，即使腾讯在下重本加速建设自己的混元模型，微信方面还是认为混元当前水平不够拔尖，甚至开始为微信 Agent 测试第三方 AI 模型，这带来了隐私和整合上的挑战性。微信方面，爆料称其实一直在做自己的 AI Agent 模型技术，计划在年中发布一个微信智能体，可以和整个小程序生态相结合，提供打车、购物的服务，并且将以联系人和聊天的方式直接集成微信之中——但就像前面提到的，这件事太重大、影响面太广了，急不得。一旦这个产品问世，可能将是全国最多用户的 AI 入口。微信需要 AI，但未必需要「元宝」，这才是腾讯现在最微妙的问题。

大疆的 Pocket 系列三年不更新，一更新就来了两台——过去一个多月，Pocket 4 和 Pocket 4P 可以说是最受瞩目的相机，就在今晚的戛纳电影节上，Pocket 4P 以口袋电影机的身份在欧洲电影殿堂亮相，距离正式发布时间，相信不会太远了。与此同时，大疆也对 Pocket 3 进行了降价处理，于是我们看到了泾渭分明的 Pocket 系列产品线：Pocket 3：入门级的手持云台 vlog 机，2299 元Pocket 4：全能的手持云台相机，2999 元Pocket 4P：准专业级的口袋电影机，预计价格会来到 4000 元档相信有许多持币观望的读者，也和爱范儿抱着同样的疑问：三台 Pocket 我该怎么选？有了 iPhone，我还需要 Pocket 吗？结合实际体验，爱范儿一次性帮你理清楚。Pocket 三代同堂，怎么选？先仔细聊聊 Pocket 4P。虽然目前 Pocket 4P 依旧没有公布更详细的参数，但参数在本次形态更新面前，是最不重要的话题——从产品逻辑来看，Pocket 4 已经将单摄的手持云台相机推向较为成熟的状态。此时，增加一颗摄像头成为一件顺理成章的事情。绝大多数手机的主摄和单摄 Pocket，提供的都是广角视野。广角适合交代环境，适合把大山大河框进画面。但当你拿着设备走入拥挤的人群，广角的缺点就暴露无遗：画面太杂乱了。此时，一颗长焦镜头可以完成空间上的压缩，把无关的背景剔除出去。大疆将这颗多出来的摄像头定格在 3×，这是一个非常巧妙的数字。它比主摄更贴近人眼对焦后的视线范围。从我们上手体验来看，Pocket 4P 这颗 3× 镜头还提供了比较自然的虚化，非常适合拍摄更凸显主体的人像或特写画面。但对于用户来说，多的这一颗摄像头，并非绝对的正收益：Pocket 系列自诞生起，核心卖点就是「无感携带」，增加镜头，意味着增加重量、侵占内部空间，甚至可能牺牲一部分续航。其实这种担心，恰恰帮大疆划清了 4P 和 4 之间的界限。Pocket 4P 是给那些「创作者」准备的。他们清楚地知道自己要拍什么，素材最后会进入 Mac，在 DaVinci Resolve 或者 FCPX 里进行精细的剪辑。对他们来说，为了在极轻便的体积下获得原生焦段的景别切换，以及高度统一的素材规格，牺牲一点点重量完全值得。而对于大多数普通人来说，Pocket 4 才是那个更轻量、更理性的选择。虽然前代 Pocket 3 打造出千万台的神话，但其本身并不是一台完美无缺的相机——只有 USB 2.0 导致数据传输过慢、像素量过低以及动态范围不够大，这些都是问题。而 Pocket 4 提供了 3700 万像素，更高的像素意味着内置的无损 ISZ 变强了。在 Pocket 4 上，通过传感器裁切获得的 2× 焦段更为清晰，足以应付日常记录。除此之外，它还内置了 107GB 存储，带来了更好的动态范围。忘记带内存卡、传输速度和动态范围不够的问题，被一次性解决掉了。可以说，Pocket 4 就是目前单摄手持云台相机的完全体。至此，Pocket 4 系列两台机身的区别已经很清晰了：如果你追求生产力，强调原生光学焦段带来的画质与虚化，需要 3× 镜头获得更丰富的画幅景别，那么 Pocket 4P 毫无疑问是更好的答案；如果你只是希望有一台能轻松塞进裤兜、防抖优秀、可以独立于手机拍摄视频和 Live Photo 的机器，那么 Pocket 4 凭借更高的像素和好用的 2× ISZ，已经绰绰有余。最后是 Pocket 3。在 Pocket 4 补齐了大量痛点的情况下，Pocket 3 还有存在的意义吗？当然——新设备的出现，最大的贡献往往是把老旗舰的价格打下来。在 Pocket 4 系列轮番上场的间隙，我们观察到 Pocket 3 已经完成了一轮官方降价，在某些平台叠加各种补贴的情况下，已经落到了 2099 的价格。甚至在未来，我们能以 1 开头的价格买到 Pocket 3，也不是不可能。对于预算有限、或者刚刚萌生拍视频念头的新手，这就是性价比最高的 「守门员」。最后，按照爱范儿编辑部的推测，Pocket 4P 的定价大概率在 4500 元左右，但也不排除大疆卷起来定一个 3999 元的价格。此时，如果将 Pocket 系列按照价格排开，我们会发现一个巧妙的设计——从 2000 元档的 Pocket 3 到 3000 元档的 Pocket 4，最后到 4000 元档的 Pocket 4P，大疆正在把无人机上那套 Pro、Air、Mini 的组合拳搬到口袋相机上。这套阵容已经在无人机市场证明了它的成功。现在，Pocket 也将组成这样的阵容了。有了 iPhone，我还需要 Pocket 吗？在 Pocket 4P 的上手文章发布后，我收到不止一个朋友的问题：我有 iPhone，还需要买 Pocket 吗？理清了 Pocket 家族的内部定位后，这是我们必须面对的现实问题：既然大家手里都有一台性能强悍的 iPhone，为什么还要花钱买一台功能单一的云台相机？这其实是一个关于「设备属性」的哲学问题。在多次迭代后，iPhone 在动态影像领域的基础不可谓不强，三摄变焦顺畅、色彩高度统一、防抖优秀、Apple Log 甚至已经更迭到第二代。但问题在于，手机始终不是完全针对拍摄设计的特化产品——除了拍摄以外，iPhone 还需要承担复杂的日常任务，无时无刻需要分散注意力，随时可能会有工作消息、语音电话打断你的拍摄、需要考虑拍摄之外手机还有没有充足的剩余电量保证你与世界的关联，还需要注意拍摄会不会占满本就贵如黄金的存储空间……而 Pocket 则是完全服务于拍摄需求的相机：不能打电话，不能回微信，不能刷社交软件，一切与拍摄无关的东西，都不会出现在你手里的 Pocket 上。更别提，Pocket 还有机械云台天生在结构上比 iPhone 的防抖更稳定、形态比 iPhone 更适合握持、物理虚化比 iPhone 更自然、视频规格比 iPhone 更丰富等等综合优势了。所以，如果我手里已经有一台主力 iPhone，现在到底该不该买 Pocket？如果现在需要，哪台是最适合我的？你可以对照下面几个问题，做一个购前诊断：过去三个月里，你是否因为手机突如其来的电话或微信，打断过拍摄？拍视频时，是否因为手机存储空间不足，而束手束脚？炎炎夏日，有没有遇到过手机拍照录视频发热掉帧甚至死机的情况？如果你对这些问题的回答是肯定的，那么购入 Pocket 3 和 Pocket 4 都是不错的选择。具体来说，如果你有拍高清 live 图的需求，如果你喜欢用手机的 2× 镜头拍摄，那么选择 Pocket 4，否则，Pocket 3 其实就够用了。在这个基础上，你还想拿 Pocket 拍拍商单、拍拍演唱会，搞点严肃创作，那么 Pocket 4P 是一步到位的选择。当然，如果前面几道题，你的答案都是否定的，那么请捂紧钱包。iPhone 的防抖和画质足够满足你的需求。你可以随拍随剪，直接通过 AirDrop 分享，且丝滑的生态体验依旧是目前任何第三方设备都无法替代的。而从行业角度来看，爱范儿更建议你等一等——目前，口袋云台相机的玩家可不止大疆一家——和 Pocket 4P 同样「犹抱琵琶半遮面」的，还有影石 Luna，而各大手机厂商也正跃跃欲试，准备下场分一杯羹。因为 Pocket 4P 这样的产品，已经严重威胁到了手机厂商的安全区。很多人愿意买 6000 元甚至 8000 元的影像旗舰手机，很大程度就是为了长焦。现在，花更低的价格，Pocket 4P 就能提供一个素质相当的长焦镜头，以及更好的云台防抖、更专业的整机解决方案。没有手机厂商能对此熟视无睹。据爱范儿了解，vivo、OPPO 都已经将手持云台相机的产品规划提上日程，而小米、华为也在持续关注这个赛道。2026 下半年到 2027 上半年，想必会有更多手持云台相机纷至沓来。既然入局者多了，我们作为消费者，反而可以更淡定些。你可以把钱捏在手里，再等一等。我们相信，这个年轻的品类，还会有更精彩的产品出现。

鸿蒙智行车型阵营里可能缺一辆 MPV，但中国汽车市场已经不缺 MPV 了。别克 GL8，丰田埃尔法和传祺 M8 等车型在燃油车时代占据了 MPV 就是商务接待用车的印象之后，新能源 MPV 迎来了数款重量级车型，极氪 009 主打高端，意在取代丰田埃尔法乃至雷克萨斯 LM 的生态位，理想 MEGA 则希望利用 MPV 的空间优势重塑大家庭用车的产品定义，岚图梦想家成为这家新势力最依仗的产品成功和别克 GL8 分庭抗礼，小鹏 X9 想要以宜商宜家，是 MPV 也可以是 SUV 的灵活定位吸引更多受众……新旧势力在 MPV 市场的攻伐，发生在 2023 年，但现在已经是 2026 年了，零跑做一款旗舰 MPV 只需要把性价比拉到极致，但鸿蒙智行做一款旗舰 MPV 并卖四五十万元，需要做的就多了，当然我承认坐在二排的老板们对智界代言人刘亦菲肯定都是心驰神往，但除了刘亦菲这个核心竞争力之外，鸿蒙智行的首款 MPV 肯定需要更多击中老板坚硬内心和柔软脂肪的东西。这些东西是车载制氧系统、全维包裹安全气囊、华为悦彰非凡系列音响、二排主动座椅旋转、侧门交互式迎宾光毯等等等。最终，智界 V9 分为 Max、Max+、Ultra 和 Ultra+ 四个版本，售价分别为 39.98 万，42.98 万，45.98 万和 51.98 万起。舒适重要，安全重要，舒适且安全更重要！过去几年的时间里，鸿蒙智行成功地让国内汽车行业重视起来零重力座椅和主动安全来，是的，这两项鸿蒙智行都不是发明者，但确实在打开用户心智和开拓市场上起到了主要作用。前者代表座舱的舒适性，通过模拟太空失重环境下人体自然放松姿态的设计。通过特殊的人体工程学优化，使臀部、腰部、背部和腿部受力均匀，释放肌肉压力。后者代表汽车安全从被动防御走向主动避让，在汽车发生事故前，通过电子控制系统感知环境、预判风险并主动介入（如自动制动、转向控制），从而避免或减轻事故。但是，二者的矛盾在于在汽车行驶状态下，零重力座椅一般是没法使用的，这是因为当乘客躺在大仰角的座椅上享受时，身体与安全带、气囊的相对位置已经发生了相当大的变化。董车会之前报道过，根据 CIDAS 的数据，一旦座椅靠背夹角超过 115°，正面碰撞中的重伤风险会陡增 4.4 倍。此时，传统的 B 柱固定式安全带无法再有效约束骨盆，人体会因为惯性产生极度危险的「潜滑现象」——整个人滑向座椅下方，安全带可能会锁死在颈部导致窒息。采埃孚的试验甚至显示，在 50km/h 的碰撞下，躺倒状态受到的脊柱伤害堪比从 5 楼坠落。打过工的人都知道，当你向老板解释「零重力座椅虽然很舒服，但是不能在行驶状态下使用」时，老板这个时候一般不会听你的解释，只会跟你说「我不管，我想用，给我想一个解决方案出来」。今天，这个方案就被智界 V9 想出来了。智界 V9 全球首发了二排全维包裹安全气囊，这个气囊集成于二排座椅靠背，触发瞬间会形成一个独特的「U 型」环绕结构，在零重力、躺姿等多种姿态下可以全面防护头部、颈部、胸部，有效防止乘员在车辆碰撞之后身体滑落、碰撞与下潜，智界宣称历经百项严苛工况验证，无误伤风险。与之配套的还有二排坐垫前端也有专门布置的臀部气囊（BAB），在碰撞瞬间快速充气形成物理坡度，顶住盆骨不让身体滑脱，从而大幅降低腰椎压力。做了这么多工作，目的只有一个：就是让二排的老板在车辆行驶的时候好好躺一躺，好好休息一下。车载制氧：让高精力人群更有精力一直以来，对于什么样的人能做老板有个迷思，甚至还有「基因彩票」的说法，就是高精力人群更容易创业做老板。这些人可以每天工作十几个小时不知疲倦，只需要睡四五个小时就能满血复活。正如前面所说，很多老板之所以看起来可以每天工作十几个小时不知疲倦，很有可能是他可以中间随意选择休息时间。而屏幕前后的你和我，并不能在上班时间睡觉。更科学的说法是，高精力人群背后，跟摄氧量有正向关系，很多所谓的低精力人群很可能有鼻炎咽炎或者心肺功能低下导致日常处于轻微缺氧状态，导致精力不足。类似的情况也会发生在车上，很多人坐高铁感觉昏昏欲睡就是车厢里二氧化碳含量高，氧含量低，这种情况就容易困，密闭的汽车座舱也可能出现这种情况。智界 V9 给出的解决方案也很简单，装个制氧系统。这套制氧系统采用了 PSA（Pressure Swing Adsorption）分子筛变压吸附技术，提供了「弥散+鼻吸」两种模式，鼻吸出氧浓度 93±3%。额定制氧排量 8L/min，峰值 10L/min，比一般的家用制氧机大很多。2 个鼻吸口支持 1 拖 2，最多可供 4 人同时吸氧，弥散供氧与空调共用，四个出风口协同工作，30 分钟即可让整个座舱达到富氧状态。除了解决了传统制氧机体积大，噪音大的不足之外，鸿蒙智行的看家本领是智能化。这套制氧系统支持语音控制开启，中控屏实时显示状态；内置舒缓解压与高原出行两种模式，可自动调节供氧浓度，也可以联动小憩模式，帮助身心放松。很多人在森林景区，绿肺公园里感觉心思空明浑身舒爽，也是因为在富氧环境下，血氧饱和度一般会更高一些。富氧环境在特定场景下确实会有积极作用，比如早起后在通勤路上进入富氧环境有助于从昏沉状态中恢复；在运动后进入富氧环境可以加速乳酸代谢，缓解肌肉酸痛。更典型的场景就是针对本身心肺功能或者呼吸系统有问题的人，富氧环境有助于把缺氧状态调整为正常状态，还有就是高原出行环境，正常的氧气浓度可以有效缓解高原反应。在制氧系统之外，针对座舱健康还有鸿蒙 ALPS 健康座舱 3.0，通过把 MOFs 可吸附降解材料融入顶棚、门板、地毯等内饰，即便熄火也能持续吸附有害物净化醛类。UVC 空调主动杀菌，对流感等病毒灭杀率超 99.99%；前后双负氧离子净化器，可以释放超过西双版纳雨林 5 倍的负离子；这一整套净化杀菌和补氧系统，提供的是「健康」的价值，毕竟对于老板来说，紧急时刻生命第一，不紧急时刻健康第一。当然，需要着重说明的是，并不是所有人在所有场景都需要制氧系统制造一个富氧环境，如果本身不处于缺氧状态，血氧饱和度在 95% 以上，不需要运动恢复，没有昏沉嗜睡情况，不在高原地区，那就不太需要吸氧，长期过度摄氧反而有害健康。旋转座椅，优雅，还是优雅冰箱彩电沙发之所以从争议性的玩笑变成了兵家必争之地，还是市场表现的结果，谁不想二三十万的车就配有之前只有 78S 系才能选装的配置呢？这种造车逻辑之下，成本的分配也就必然从内燃机变速箱往舒适性配置上调整了。智界 V9 的沙发，显然值得多点笔墨。双独立结构，通风加热和比气动按摩更强劲的机械按摩已经不足为奇。更关键的是，二排座椅有可旋转设计，最大旋转角度 180 度，可以在多种角度进行旋转和固定，营造了例如 45° 迎宾座椅，90° 观景座和 180° 旋转之后的后排会客室布局等多种模式和选项。尤其是迎宾形态和会客室形态，可以在商务场景里营造比普通 MPV 更优雅的氛围。出门在外，面子可以是自己给的，但归根到底，还是别人给的，给别人面子就是给自己涨面子。智界 V9 类似的优雅配置还有不少，比如说「手势控车」，佩戴支持功能的手表，做出轻捏动作就可以隔空打开车门；二排乘客在车内做一个推一推手势，二排侧滑门就可以优雅闭合；窗边手指划一划，就可以调节智能隐私光幕比例的明暗度……车内的四块智能隐私光幕玻璃最大遮光率可达 99.6%，如果一键开启隐私模式，可以立刻断开车内的位置信息、麦克风、摄像头、网络，保证车内隐私。新车配有超 50㎡ 静音材料覆盖和 13 块双层夹胶隔音玻璃进行被动降噪，同时还可选 RNC 主动降噪。其他一些老板不太关心，但司机需要关心的作为一款鸿蒙智行的车，其实很多关于底盘，辅助驾驶和动力的信息都能猜个大差不差，这里完全可以用列表形式一一列举，毕竟这些信息是给司机看的，而且司机一看就懂，无需解释。新车长宽高为 5359×2009×1879 mm，轴距 3250 mm，2+2+3 座椅布局，后备箱容积 892L（可扩容至 2002 L）。全系标配 38 个传感器，包括 896 线激光雷达，背靠 60 EFLOPS 云端 AI 算力、WEWA 2.0 架构，可首批升级华为乾崑智驾 ADS 5。搭载 1.5T 雪鸮增程器配合前后双电机，综合功率 380kW，搭载 53.4kWh 电池，CLTC 纯电续航 305km，综合续航 1320km。新车有辉光紫、深海蓝、雪域白、鎏金黑四种车身配色和韶华杏、赤茶橘、绒霞紫三种内饰颜色。搭载 MPV 版图灵平台，前双叉臂与后 H 臂多连杆独立悬架，配合双腔空气悬架及 CDC 连续可变阻尼减振器。全系标配±7° 主动后轮转向，5.35 米转弯半径，支持低速蟹行模式，有效应对狭窄路况的调头和穿行。四活塞前固定卡钳，35.8m 制动距离；500mm 涉水深度，保证暴雨积水情况下的通行能力。路面预瞄系统智能识别井盖、凹坑与弯道，智能调节悬架阻尼、刚度，以及电机扭矩，可以降低冲击与侧倾，获得莱茵晕车舒缓认证。行业首发 2 根 2400MPa+6 根 2200MPa 侧门防撞梁；全系标配 13 个安全气囊，包括全维包裹安全气囊。前环宇三联屏，12.3 英寸液晶仪表与双 17.2 英寸 3.4K 大屏组合，可选后排 21.4 英寸吸顶屏。因为鸿蒙座舱，乾崑智驾，还有图灵底盘等鸿蒙智行基础产品技术支撑，现在鸿蒙智行的车型想要出挑有点困难，但毫无疑问，无论是华为还是奇瑞，都给到了智界 V9 足够的重视，比如单独给这款车开一场声势浩大的发布会，这在整个华为体系里是极为罕见的待遇。再比如，华为和奇瑞给这款车投入了一些深刻的产品洞察去迎合和吸引智界 V9 的潜在受众，很多人可能对这样一款华为气质非常浓郁的 MPV 产品无感，但华为不在乎，华为只希望这些人快快长大并事业有成。

发布会开场，李想先是强调：其实大部分用户并不知道自己想要什么。他分享了一个故事：有一次我去买烤冷面，摊主问我还要加点什么，我开玩笑说能不能打个折。你看，用户永远想要更便宜，但摊主会告诉你，其实还能加蛋。这也是 L9 当年销量爆发的秘诀——当行业还在比拼基础参数的时候，理想顺手给车里塞进了冰箱、彩电和大沙发，交出了一份远超预期的答卷。四年过去了，全行业都开始把屏幕做大、座椅做软。当堆砌配置成为车企的常规操作，理想 L9 又走向了另一条路。如今的理想把大量精力和资金砸向底层架构，自研范围覆盖车规级算力芯片、800V 主动悬架、全线控底盘等多个领域。承载这些研发成果的，自然是今天上市的两款新车型：L9 Ultra 和 L9 Livis。首先值得肯定的是，理想没有盲目放大 L9，车长和车宽依然维持在 5.25 米和 2 米，日常开起来依旧方便。在这个好开的尺寸基础上，L9 的整车姿态还挺拔了不少。在车高仅增加 10 毫米的情况下，前机盖和水切线被大幅拉高，配上 22 寸大轮毂，原本庞大的车身显得颇为矫健；前唇那一抹金灰色镀铬饰条也收束得干净利落。L9 Livis 还采用了迈巴赫同款涂装工艺，紫金、绿金、黑金等专属双色车身，额外增加两道工序后，漆膜厚度达到传统工艺的两到三倍，日光下呈现出流动的光影质感。外观的考究还藏在细节里。全车 12 颗超声波雷达被 UWB 传感器取代，保险杠上再也看不到突兀的小圆孔；隐藏式电动踏板收起后与底盘融为一体。这不仅让外观更整洁，UWB 技术还让哨兵模式耗电量骤降 82%，开两天只耗一度电。打开带有电容防夹功能的电动车门，里面还是那个熟悉的「起居室」，只是每一项体验的上限都被抬高了。全新 L9 是全球首款四个座位都配备全尺寸零重力座椅的 SUV。Livis 版本直接标配，Ultra 版本在首销期内也会免费赠送前排两把零重力座椅。在过去，转向管柱挤占了空间，主驾想完全躺平是个大难题。而在 L9 上，线控转向释放了空间，1.8 米的大个子坐进主驾也能一键舒展躺平。望向前方，仪表台上的两块屏幕合二为一，变成了一块 29 英寸的 6K 超宽全景屏。后排的 21 英寸屏幕新增电动平移功能，手轻轻一抬，这块「神奇移动屏」就会沿滑轨移到面前，并随座椅倾角自动找平。音响功率这次飙升到了 5440 瓦。9.3.6 星环剧院音响搭载三重矩阵式低音系统，头枕音响创新采用水平对置设计，出声口藏在两侧，以往那种「有人在你脑后说话」的突兀感也随之消失。车内空调支持 7 温区控制，面部和脚部可以吹出不同温度的风；全车 29 处皮质区域都带加热功能。中控台下方的 10L 智能冰箱换了电动门，轻拍两下即可开关，不仅支持 -6°C 冷冻，内胆还能直接拆下来清洗。作为增程车，全新 L9 两个版本的纯电体验都做到了行业顶配。底盘塞进了一块 72.7 度的 5C 大电池，容量已经超过不少纯电车型，换来 420 公里纯电续航。按每天通勤 40 公里算，两周充一次电完全够用。充电速度也同步提升，配合理想自建的 8300 根 5C 超充桩，十分钟左右就能完成补能——在服务区上个洗手间、买杯咖啡的功夫，电量就能满血复活。长途出行时，这台重达 2.8 吨的全尺寸 SUV 的油耗还挺实在的：亏电油耗降至 6.3 升/百公里，满油满电综合续航达到 1650 公里。更让人省心的是，这套增程系统的保养周期延长到了 3 年或 3 万公里，还省去了过去增程车繁琐的维护。李想表示，在机械属性和维保频次上，全新 L9 已经越来越像一台纯电车了。把家装进车里，把纯电融进增程「全尺寸 SUV 一直有一个很难真正被解决的矛盾，既让一家人坐得舒适，又让自己开得尽兴。」李想说。为了同时满足这两点，L9 Livis 直接上了全套线控主动底盘。L9 Livis 的 800V 主动悬架，让每个车轮都有超过 1 万牛的支撑力，传统的防倾杆被取消，四轮完全独立。过减速带、连续起伏路或者高速变线时，液压泵和减震器会实时配合，该撑就撑、该压就压。停车休息时，这套悬架甚至能联动赛车游戏，把车厢变成 4D 模拟器。大车想开得灵活，转向系统也要升级。Ultra 和 Livis 都标配了后轮转向 + 线控转向，转弯半径缩到 5.3 米，方向盘打满不用换手，角度也能按习惯自定义。另外，Livis 还首发了 EMB 线控机械制动，刹车响应快了约 0.1 秒，120km/h 时速下刹停距离能缩短 3 米以上。而且这套系统号称终生不用换刹车油，后期省心不少。安全性上，两款车都变得更加坚固了，堡垒安全车身 3.0 用上了 2200 兆帕的高强度钢和一体式双门环。最特别的是，车内的 11 个安全气囊还能通过红外传感器识别每一位乘员的体型与坐姿，动态调节保护策略。面对极端的碰撞危险，具备主动悬架的 L9 Livis 能在 0.7 秒内瞬间抬升 50 毫米，用最坚固的底盘部位去硬刚撞击；低速倒车时的防碰撞系统也具备了识别悬空物体的能力。智能化层面，理想首发了马赫 M100——一颗专为 AI 原生设计、采用数据流架构的端侧推理芯片。L9 Ultra 版本用的是单芯片方案，算力达到 1280TOPS，配备前向激光雷达；L9 Livis 则拥有两颗马赫 M100，总算力翻倍达到 2560TOPS，同时配备了 4 颗激光雷达来实现 360° 全向感知。算力上来后，感知帧率成倍提升，乡间小路、施工路段都能提前预判，果断绕行。这颗芯片还跟去年 12 月发布的 Livis AI 眼镜实现了生态互联。交互开始变得科幻。戴着眼镜站在车外，一个手势或一个眼神，车就会自动泊出来，缓缓开到你面前。坐进车里，系统还能察觉你的情绪和状态。你今天好像心情特别好，是要带家人去顺义的公园踏青吗？这时候，它既是司机，也是伙伴。到了我们最关心的价格，两款车只拉开了 5 万元的价差。L9 Ultra 定价为 45.98 万元，竞争力很强。72.7 度 5C 电池、双腔空悬、线控 + 后轮转向、1280 TOPS 马赫芯片都是标配，首销期还免费送前排零重力座椅、智能电开门和 5440 瓦音响。L9 Livis 定价来到 50.98 万元，比此前 55.98 万的预售价低了不少。多花 5 万，换来的是专属双色车身、800V 主动悬架、带 EMB 制动的全线控底盘、翻倍到 2560 TOPS 的算力，以及 4 颗激光雷达。首销权益也公布了：6 月 30 日前锁单，现金直减 2 万元，老用户再额外补贴 1 万元。从 2015 年创立至今，理想还是习惯把消费者叫作用户。买车只是陪伴的开始，让全家人在日复一日的相处里都能开心，才是 L9 Livis 这台「硅基生命体」想要兑现的承诺。

在今天和理想全新 L9 和智界 V9 两台重磅车型同台亮相的，还有蔚来的乐道 L80。相比 L90 面向三代同堂、多孩家庭的大三排需求，L80 把重点放在大五座市场，带来了更大的乘坐空间，更灵活的装载能力，以及更适合长途、自驾、露营等场景的空间组织方式。早年间，以本田冠道、大众途观 L 为代表的燃油大五座 SUV，主要通过释放第二排的腿部空间来提升乘坐舒适性；随后，随着新能源时代的到来，诸如理想 L7、问界 M7 等车型进一步强化了智能座舱的沉浸式体验和车内的储物能力。而乐道 L80 现在则试图在行业已有的基础上更进一步，将前备舱、后备舱、乘员舱以及补能体系整合为一个完整的生态闭环，通过蔚来的工程创新能力，让中大型五座 SUV 的日常使用体验再上一层楼。乐道 L80 一共有 Pro、Max+ 和 Ultra+三个版本，整车购买价格分别为 24.28万元、25.98 万元、27.98 万元， Baas 方案价格为 15.68 万元、17.38 万元和 19.38 万元。前备舱大，后备舱更大空间，依然是乐道 L80 最核心的竞争力。乐道 L80 车长 5145mm，轴距 3110mm，理想 L7 车长 5050mm，轴距 3005mm，乐道 L80 比理想 L7 大了一圈。取消第三排之后，乐道 L80 将所有空间都留给了前两排和储物。第二排作为 L80 整台车体验的核心区域，其后排头部空间达到 1074 毫米，腿部空间宽达 1121 毫米，坐垫宽度和长度分别达到 549 毫米和 497 毫米，靠背支持最大 135 度的后倾调节。L80 提出了一个「大五座，享四座」的越级体验概念，即当车内只有四人出行时，第二排的两位乘客完全可以获得媲美大型 MPV 独立航空座椅的放松状态。而在储物空间上，乐道 L80 配备了容积为 240 升的电动前备舱，能够轻松容纳两个 20 英寸的标准登机箱、一个儿童专属行李箱及一个随身背包。前备舱的存在不仅增加了绝对装载量，更为家庭用户提供了干湿分离与物理分区的储物区域，有效改善了传统 SUV 后备舱容易被杂物堆砌得杂乱无章的痛点。与前舱相对应，乐道 L80 的后备舱在第二排座椅处于常规直立状态时，容积为 1200 升；当第二排座椅完全放倒后，整体空间可扩展至 2600 升。此外，在后备舱的常规盖板下方，工程团队设计了双沉舱结构。其中外侧沉舱容积为 106 升，里侧沉舱容积为 63 升，均配备了独立盖板。这种隐蔽式的储物空间非常适合用来存放不常用的随车工具、零散的露营小配件、充电线缆以及各类清洁用品，使得主后备舱始终保持整洁。乐道能够在车辆内部极其有限的机械空间内挤出更多的储物容积，主要得益于车辆底层平台的极高集成度。L80 充分应用了已经在 L90 上被验证过的 900V 高压架构，并配备了体积更为紧凑的前置电驱动系统。结合 30.5 度大倾角冷却模块、全栈自研的热管理高度集成模块以及一体式偏置空调箱，这些系统级的技术创新对原本需要占据前舱大量体积的机械与电子部件进行了极致的压缩与整合。空间之外，座舱内的舒适化和智能化，也是乐道 L80 强调的特色。乐道 L80 的座舱内部采用了分层布局：上层主要用于视觉延伸与氛围营造，中层集中处理人机交互与触觉反馈，下层则侧重于日常储物与抗污耐用。车内配备了支持 1600 万色调节的氛围灯系统、长达 7.2 米的环绕光带，辅以大面积软质材料包覆、金属扬声器网罩及精细的绗缝与刺绣工艺。整体风格规避了盲目堆砌昂贵材料的浮夸感，更侧重于在家庭用车的基调上打磨出温馨且不失格调的精致质感。为了提升车内的通透感，乐道 L80 车辆侧窗与前挡风玻璃的总面积约为 3.3 平方米，前后分区的全景天幕面积也达到了 1.1 平方米。同时为了应对大面积玻璃材质在强光照射下带来的隔热与防晒问题，车内配备了双电动物理遮阳帘，乐道 L80 同时配备了空气悬架系统，支持在驻车时自动降低车身高度，配合电动侧踏板与静音电吸门，优化了行动不便的老人或儿童的上下车体验。车辆的座椅总成基于专属的高端座椅平台开发，依据中国人体生理曲线大数据进行了型面贴合度优化，前排座椅支持 12 向电动调节，第二排支持大角度电动调节。同时，方向盘加热、前后排座椅加热、通风以及按摩功能均作为标配提供。此外，车内集成了一台 8.8 升容量的智能电动冰箱，足以容纳 6 罐常规可乐、2 盒鲜牛奶以及部分水果零食。神玑上车，4D 领航数字化交互与计算平台是现代智能汽车的核心大脑。L80 的屏幕矩阵由一块 35 英寸的 AR-HUD 增强现实抬头显示、一块 17.2 英寸的 3K 高清中控屏、一块 17.3 英寸的后排吸顶娱乐屏以及一块 8 英寸的后排专属场景控制屏组成。支撑这一系列复杂交互与协同调度的硬件，是一颗采用 5 纳米制程工艺的车规级高阶智驾芯片——神玑 NX9031。该芯片拥有高达 546 GB/s 的内存带宽，打破了传统分散式域控制器的物理壁垒，能够同时支持智能驾驶、智能座舱、智能底盘等多个系统跨域计算资源的统一调度。配合 SkyOS·天枢整车全域操作系统，车辆从底层代码级别打通了智驾、座舱、底盘、车身及动力网络，使各项功能摆脱了信息孤岛的状态，能够基于场景进行深度系统级协同。在智能驾驶领域，系统深度接入了蔚来世界模型 NWM（NIO World Model）。该模型摒弃了传统的规则驱动逻辑，转而通过强化学习与海量的闭环仿真训练，提升了系统对复杂未知路况的理解与博弈能力，全面覆盖城区复杂道路、城市快速路、高速公路及各类园区停车场。值得重点关注的是，L80 支持城区领航换电功能（PSP）。在导航状态下，系统能够自动规划最优补能路径，将车辆自主驾驶至换电站并完成自动化换电流程。同时，基于模型化架构构建的智能泊车系统，也支持路线的实时动态规划和对各类非标准复杂车位的识别入位。尽管定位为体型庞大的家庭 SUV，L80 在底盘的机械素质与动态调校上依然投入了极高的规格。悬架系统采用了前双叉臂与后五连杆独立悬架组合，配备了具备 100 毫米行程调节能力的智能空气悬架系统以及 CDC 智能连续可变阻尼减震器。底盘硬件与路面之间通过液压衬套进行柔性连接，并搭配了抓地力更强的鸳鸯胎。乐道 L80 还支持 4D 全域舒适领航功能。该功能能够实时结合车辆自身的传感器感知、激光雷达对前方路面的三维预瞄以及云端高精舒适图层数据，在遇到减速带、坑洼或颠簸路面之前，提前毫秒级主动调节空气弹簧的刚度与减震器的阻尼，从容化解路面冲击。除了常规的运动、舒适、后排舒享和节能模式外，车辆预设了湿地、雪地、泥地、沙地等脱困模式。在卓越的底盘参数加持下，L80 拥有高达 600 毫米的最大涉水深度，最小离地间隙最高可拉升至 260 毫米，足以轻松应对绝大多数轻越野和长途自驾游的复杂路况。而在安全方面，乐道 L80 全系标配 9 个安全气囊，除了常规的环绕式防护，第一排还配备了能在碰撞发生前预紧的主动式安全带、容积达到 137 升的狐耳式超大副驾气囊以及防止前排乘员相互碰撞的中央远端安全气囊。基于家庭属性的考量，第二排乘客的安全防护被提升至与前排同等重要的层级，配备了 14 升容积的后排侧气囊，以及长度贯穿 2920 毫米、容积高达 56 升的超大侧面安全气帘。在应对碰撞的白车身结构设计上，L80 的前部预留了长达 650 毫米的充足溃缩吸能空间，并创新性地采用了「6 纵 10 横」的笼式安全传导结构。前防撞梁选用了高强度的 7 系航空级铝合金材质，其上层主防撞梁的宽度覆盖率达到了整车宽度的 80%，能够在严苛的偏置碰撞中有效化解破坏性冲击力。侧面防护则通过环舱高强度钢、一体式门环、多腔体铝合金门槛梁以及坚固的电池舱横梁共同构建。车尾部分应用了先进的一体压铸后地板技术，配合铝合金后吸能盒与后防撞梁，为车身尾部提供了深达 1244 毫米的吸能缓冲空间，最大程度减轻高速追尾事故对后排乘客及底层电池包的结构性伤害。乐道 L80 搭载的电机能够输出 340kW 的峰值功率，百公里加速时间 5.7 秒，其百公里综合电耗被控制在优秀的 14.3kWh，85 度电池下的 CLTC 续航为 615km。当然，纯电车型的体验永远无法脱离补能体系的建设。共享蔚来庞大且成熟的充换电网络，依然是乐道 L80 在同级别竞品中最难以被复制的核心优势。为了进一步夯实这一护城河，乐道官方已确认追加投入 6 亿元人民币，用于在网络中新增 8000 多块流通电池。蔚来之前透露，早在四年前，乐道 L80 便与 L90 基于同一技术平台同步启动开发。其中，乐道 L90 的成功已经验证了产品定义的精准性。在 2025 年，乐道 L90 凭借对大三排 SUV「满载状态下行李装载空间」这一核心痛点的有效解决，在上市短短 5 个月内实现交付超 4.3 万台，成为纯电大型 SUV 市场的年度销量冠军。这不仅助力乐道品牌在 2025 年实现 107808 台的全年交付目标，更以接近 24 万元的成交均价稳居新势力品牌第一梯队。乐道 L90 和蔚来 ES8 的表现，也被行业视为大型 SUV 动力形式从增程向纯电转型的关键标志。如果说乐道 L90 成功开拓了大三排纯电 SUV 市场，那么乐道 L80 则肩负着在规模更大的大五座 SUV 领域实现突破的重任。在乐道的产品价值体系中，L90 聚焦于「在一起」的温情与责任，通过三排布局满足多孩、三代同堂家庭的聚会与通勤需求；而 L80 则更侧重于「走出去」的激情与成长，以更灵活的空间形态承载用户自驾、露营等更具探索性的生活方式。

🔍郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%🤖OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略🚀Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI📱微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆💻华为 Mate X7 官方降价 1000 元⚠️追觅造车主体约 232 万元股权遭冻结🎮国行 Switch 正式停服🧠百度成立模型委员会 BMC💰三星电子单季营收创历史纪录，半导体撑起超九成营业利润🔍海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%🖨️铠侠：2027 年内存仍吃紧、净利润同比增长 103.6%🤖微信发布「全球青少年 AI+ 小程序」洞察报告💡何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了🦿5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约📱荣耀 CEO 李健官宣机器人手机今年 Q3 发布📱vivo S60 将首发 4K 原生感 Live🚗全新理想 L9 发布：45.98 万元起🤖ChatGPT 新增 Codex 连接电脑功能📰 周末也值得一看的新闻郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%天风国际证券分析师郭明錤昨日披露，苹果已在 Intel 18A-P 系列制程上为低阶及旧款 iPhone、iPad 与 Mac 处理器立项，订单中 iPhone 约占 80%。流片规划显示，苹果预计 2027 年在 Intel 放量，2028 年续增，2029 年进入衰退期。Intel 方面，2027 年良率目标为稳定达到 50% 至 60% 以上，量产时程与出货规模目前仍未明朗。不过，郭明錤也指出，即便 Intel 顺利出货，台积电在苹果供应体系中仍将占 90% 以上比重。郭明錤认为，苹果此举是有意将 Intel 培养为长期备选供应商。AI 对台积电先进制程资源的占用将持续扩大，苹果需要在议价能力尚存时提前布局。与此同时，苹果也在同步评估 Intel 其他先进制程技术。OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略据 WIRED 报道，OpenAI 昨日向员工宣布新一轮组织调整，联合创始人兼总裁 Greg Brockman 正式接管产品战略。报道指出，OpenAI 将 ChatGPT、编程智能体 Codex 及面向开发者的 API 业务整合进同一核心产品团队，目标是将三者合并为「统一的体验」。Codex 原负责人 Thibault Sottiaux 升任核心产品与平台主管，统管消费、企业和开发者三条线；ChatGPT 原负责人 Nick Turley 转任企业产品改版；前 Instagram 副总裁 Ashley Alexander 接任消费产品负责人。Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI据《金融时报》报道，Anthropic 已就一轮 300 亿美元的融资达成条款，融资完成后公司估值将约达 9000 亿美元，较今年 2 月初 3500 亿美元的估值接近翻三倍。本轮融资由 Dragoneer Investment Group、Greenoaks Capital、红杉资本（Sequoia Capital）和 Altimeter Capital 联合领投，报道称，每家领投机构预计出资不低于 20 亿美元。微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆CCF 计算机博物馆（CCM）昨日收到腾讯微信团队捐赠的一批早期历史实物，包括微信第一行后台代码原始存档、早期开发设备 iMac ME086CH/21 寸一体机与戴尔塔式工作站，以及 Apple iPod Touch 3、诺基亚 5800XM 等早期测试终端。CCM 表示，由于移动产品迭代速度极快、数字资料易被覆盖删除，大量见证移动互联网起步的实物未能得到系统保护。此次捐赠呈现了微信从 0 到 1 的研发链条，是中国移动互联网早期发展的核心实物见证。CCM 主体建筑已全面落成，拟于今年 9 月开馆，目前持续面向社会征集各时期计算机相关藏品。保时捷称被中国车企抄袭「有些许无奈」，罗永浩：保时捷自己就有前科保时捷中国传媒公关副总裁唐风靓日前在浙江「易境思论坛」上表示，过去几年保时捷激发了高度统一的汽车设计审美，被中国车企模仿「除了荣幸也有些许无奈」。罗永浩随后在微博发文回应，称保时捷说「无奈」可以理解，但称中国车企抄袭得「高度统一」则是吹牛。他同时指出，保时捷历史上也曾抄袭捷克斯洛伐克品牌太脱拉（Tatra）的设计，并以 1965 年大众汽车向太脱拉支付 100 万西德马克庭外和解赔偿金为据，直言「说到流氓程度，谁能跟保时捷比」。此番言论随即引发讨论。有网友指其犯了「诉诸伪善」的逻辑错误，罗永浩强硬回怼，称自己已明确承认抄袭者「也是流氓」，并非为抄袭行为开脱。华为 Mate X7 官方降价 1000 元华为昨日宣布，旗下两款折叠屏手机同步开启降价优惠：Mate X7 全系直降 1000 元，优惠后售价 11999 元起；Mate X6 直降 3000 元，优惠后售价 9999 元起。追觅造车主体约 232 万元股权遭冻结企查查 App 显示，追觅科技创始人俞浩实控的星空计划（上海）汽车科技有限责任公司近日新增一则股权冻结信息，股权数额约为 232.25 万元，股权所在企业为上海星空计划进出口贸易有限责任公司。去年 8 月，追觅科技宣布进入汽车领域。截至目前，追觅汽车团队规模约千人，研发人员占比达 70%，自研覆盖智能汽车多个核心领域。此次造车主体出现股权冻结，官方尚未就相关情况作出回应。国行 Switch 正式停服国行 Switch 的 Nintendo e 商店已于昨晚正式停服，至此，国行 Switch 将无法再进行联网对战或下载游戏。官方在停服前向相关玩家发送提醒，建议玩家在停服前完成所有已购游戏、DLC 及兑换码的下载与激活；删除游戏数据或对主机进行格式化后，数字内容将无法恢复。今年 3 月 31 日，国行 Switch 的游戏售卖服务已率先停止；而根据官方此前公告，今年 12 月 31 日起，国行 Switch 主机及配件的维修服务也将全面终止。这意味着，国行 Switch 的官方售后支持将在今年底彻底画上句号。百度成立模型委员会 BMC据《科创板日报》报道，百度近日内部设立百度模型委员会（Baidu Model Committee，BMC），成员由对大模型有深刻理解的年轻研究员构成。基础模型研发部（BMU）与应用模型研发部（AMU）均向该委员会汇报。知情人士透露，新成立的 BMC 将全面统筹百度大模型业务，覆盖路径探索、技术规划、模型训练到业务落地的完整链条，目标是实现模型从技术到应用的一体化管理。百度方面表示，此举有助于集中优势资源，持续建设并强化人工智能方向的技术优势。三星电子单季营收创历史纪录，半导体撑起超九成营业利润三星电子昨日发布 2026 财年第一季度（1 月 1 日至 3 月 31 日）业绩报告：期内实现合并营收 133.9 万亿韩元，同比增长 69%，环比增长 43%；营业利润 57.2 万亿韩元，同比增长 756%，环比增长 185%，营业利润率达 42.8%。净利润为 47.2 万亿韩元，净利率为 35.3%。基本每股收益为 7123 韩元。半导体（DS）：营收 81.7 万亿韩元，同比增长 225%，环比增长 86%；营业利润 53.7 万亿韩元。其中存储业务营收 74.8 万亿韩元，同比增长 292%，环比增长 101%，为本季度核心增长引擎；设备体验（DX，含手机与消费电子）：营收 52.7 万亿韩元，同比增长 2%；营业利润 3.0 万亿韩元，同比下降约 1.8 万亿韩元；显示（SDC）：营收 6.7 万亿韩元，同比增长 14%；营业利润 0.4 万亿韩元，同比下降 0.1 万亿韩元；Harman：营收 3.8 万亿韩元，同比增长 12%；营业利润 0.2 万亿韩元。下季度展望，公司预计 AI 基础设施扩张驱动的服务器内存及 SSD 需求将持续强劲，计划向客户交付 HBM4E 首批样品以巩固技术领先地位；手机业务方面，Galaxy S26 系列发布效应趋于消退，盈利能力预计承压。海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%昨天，海力士于发布今年第一季度（截至 3 月 31 日）业绩，按 K-IFRS 口径，期内实现营收 52.58 万亿韩元，同比增长 198%；营业利润 37.61 万亿韩元，同比增长 405%；净利润 40.35 万亿韩元，同比增长 398%。本季度营收首次单季突破 50 万亿韩元，创历史新高。主要财务指标如下：毛利润 41.68 万亿韩元，毛利率 79%，同比提升 22 个百分点；营业利润率 72%，创历史单季新高，环比提升 14 个百分点；EBITDA 41.34 万亿韩元，EBITDA 利润率 79%，同比增长 284%；净利润率 77%，其中非经营利润 14.01 万亿韩元，包含投资资产估值收益 9.94 万亿韩元。铠侠：2027 年内存仍吃紧、净利润净同比增长 103.6%铠侠控股昨日发布截至今年 3 月 31 日止 2026 财年业绩，期内实现营收 2.3376 万亿日元，同比增长 37.0%；按 IFRS 口径，归母净利润为 5544.9 亿日元，同比增长 103.6%。SSD 与存储业务营收 1.3626 万亿日元，同比增加 37.5% 至 3715 亿日元；智能设备业务营收 7599.8 亿日元，同比增加 51.6% 至 2588 亿日元；其他业务（含 SD 卡、U 盘等零售产品及向闪迪集团的销售）营收 2150.1 亿日元，同比基本持平。另据财联社报道，铠侠表示，AI 服务器客户的订单量持续走强，以美元计价的 NAND 价格在一季度上涨逾一倍，并预计 2027 年度 NAND 存储供应仍将维持紧张态势。微信发布「全球青少年 AI+ 小程序」洞察报告昨天，腾讯研究院联合微信发布《全球青少年 AI+小程序洞察报告》，披露了全球基础教育阶段群体利用 AI 工具与小程序进行应用开发的各项核心数据。据报告数据显示，自 2023 年 11 月微信小程序教育平台正式对外开放以来，已吸引全球 7954 所学校的逾 8.7 万名学生参与，累计创建小程序项目超 28.7 万个。在开发流程中，师生调用 AI 辅助创作每年消耗 Token 量超过 500 亿，支撑约 375 万次人机交互。此外，微信连续四年举办的小程序全球创新挑战赛共产生 7430 个作品，其中第四届单届提交量达 4900 个，呈显著爆发趋势。开发者画像层面，62% 的项目由单人独立完成，最小获奖开发者年仅 5 岁。地域活跃度数据显示，中国香港、北京及中国澳门位列前三，且项目协作已出现跨国别开发特征。在应用方向上，05 后开发者主要聚焦于无障碍辅助、碳足迹、适老化关怀及健康管理等社会实用场景。💡 何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了据睿见 Economy 报道，小鹏集团董事长、CEO 何小鹏昨日在「第十八届轩辕汽车蓝皮书论坛」上表示，「什么时候我们看到北京、上海车展上不再有 150 款新车的时候，中国汽车的内卷可能就结束了。」何小鹏表示，过去十年，中国新能源汽车经历了两个重要转变：一是核心驱动能源从传统化石能源切换至电化学能源；二是行业全面进入初步普及的智能化阶段。他称，十年前几乎没有人相信中国汽车能在全球市场取得如今的成绩，并实现规模化的全球化布局。何小鹏认为，下一个十年将是电化学能源与数据能源的双重叠加阶段，芯片和模型将成为新的「引擎」，汽车与机器人的能力将在这十年内深度整合。5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约ROG 与 XREAL 联名推出的电竞 AR 眼

🤖曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini🌐马耳他将向全体公民免费提供一年 ChatGPT Plus📱小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务🚀目标估值 1.75 万亿美元，SpaceX 最快 6 月上市📋arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚💻卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元🚗蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌🔍需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位☁️最低 1 元 40 万 Token，三大运营商推出 Token 套餐🚙特斯拉 Robotaxi 碰撞事故报告解封💡知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值🚗小米 YU7 GT 火山灰配色公布💰ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构🧠马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布🤖字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本📱腾讯 vibe coding 产品「吐司」即将上线🛍️蜜雪冰城上架老黄同款蜜桃四季春，销量暴涨🛑中国银行关停信用卡独立 App，服务整合进主 App曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini昨晚，彭博社记者 Mark Gurman 在最新的 Power On 新闻通讯中透露了更多关于苹果即将发布的独立 Siri 应用的细节。隐私设计是此次新应用的核心卖点之一。Gurman 透露，该应用将内置对话自动删除功能，用户可在设置中选择对话保留时长：30 天、1 年或永久保存。苹果预计在今年的 WWDC 继续主打隐私保护，并强调其 AI 体验不含广告。模型方面，苹果已将新 Siri 的 AI 模型底层大量替换为 Google 的 Gemini，并将部分服务运行在 Google 的云基础设施之上。而此前有消息称，苹果和原本的合作对象 OpenAI 关系出现裂痕，甚至有可能对簿公堂，原因是 OpenAI 认为双方合作的结果不及预期。值得注意的是，尽管新的 AI Siri 已经跳票约两年，Gurman 称苹果内部测试版本仍将其标注为「Beta」，并可能沿用至其正式发布。马耳他将向全体公民免费提供一年 ChatGPT PlusOpenAI 日前宣布，与马耳他政府签署合作协议，向该国公民及居民免费提供为期一年的 ChatGPT Plus 订阅服务。据悉，居民须先完成由马耳他大学开发的免费在线课程「AI for All」，课程内容涵盖 AI 的基本概念、能力边界及负责任使用方式，通过后即可向马耳他数字创新局申请免费订阅资格。小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务小米 REDMI 产品经理「笋寸」日前在微博发文建议，近期不打算换手机的用户可以考虑等待换电池，并提示部分机型在更换新电池后容量可能还会有所提升，参考小米 13 系列的现行政策。笋寸进一步说明，小米正积极推动存量用户更换采用金沙江技术的大容量电池，小米 13 系列已率先上线相关服务。她还透露，具体支持机型「得等等」。此前，小米 13 系列已全系上线电池升级服务，支持小米 13、小米 13 Pro、小米 13 Ultra 三款机型升级更大容量电池，费用为 189 元，其中电池升级费 149 元、人工费 40 元。目标估值 1.75 万亿美元，SpaceX 最快 6 月上市据路透社报道，SpaceX 已加快 IPO 时间表，计划最早于 6 月 12 日在纳斯达克挂牌上市，股票代码为「SPCX」，较此前预定的 6 月下旬提前约两周。据三位知情人士透露，公司计划最早于本周三公开招股说明书，路演定于 6 月 4 日启动，股票发行最早于 6 月 11 日完成定价。时间表提前的主要原因是 SEC 对招股材料的审核进展快于预期。此次 IPO 预计募资约 750 亿美元，目标估值约 1.75 万亿美元，有望成为史上规模最大的股票发行。承销行阵容包括摩根士丹利、美国银行、花旗集团、摩根大通和高盛。arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚近日，预印本学术论文平台 arXiv 宣布，若作者提交的论文中包含无可辩驳的 AI 生成内容证据，将被封禁账号一年，并在此后要求所有提交须先经过同行评审期刊接受方可上传。arXiv 计算机科学板块主席 Thomas Dietterich 在 X 发文表示，若论文中出现「幻觉引用」、大语言模型元注释（如「以下是 200 字摘要，是否需要修改？」）等明显迹象，即可认定作者未对 AI 生成内容进行核查。arXiv 目前由康奈尔科技学院管理，今年 7 月将独立运营，成为一家非营利法人机构。康奈尔科技学院院长 Greg Morrisett 表示，独立后 arXiv 将能从更广泛的捐助方获得资金支持，以应对「AI 垃圾内容」带来的运营压力。卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元近日，小米集团合伙人、总裁卢伟冰在「Max 一夏」户外露营直播中表示，玄戒芯片今年肯定会推出迭代款，并称这将是「一块非常强的芯片」，会搭载在一款「非常优秀的产品」上。他同时提醒外界，目前网络上流传的各类相关传闻可信度不高，具体规格暂不便披露。卢伟冰还透露新一代玄戒芯片的应用范围将不再局限于智能手机，后续计划广泛搭载于小米旗下多类智能终端设备。关于手机定价，卢伟冰在直播中作出判断，认为今年下半年尤其到年底，部分国产旗舰直板手机价格可能突破 1 万元人民币。他预计，内存涨价的压力至少延续至 2027 年底，甚至到 2028 年。对于即将发布的小米 17 Max，他表示定价仍在讨论中，将在发布会上正式公布，但小米的原则是在同档位提供「最物超所值的产品」。蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌据睿见 Economy 报道，在日前举行的第十八届轩辕汽车蓝皮书论坛上，蔚来创始人、董事长兼 CEO 李斌线上出席，在对话中将萤火虫比作 iPhone，称其已成为「一款非常高品质的产品」。他强调，萤火虫在某种意义上超越了汽车的工具属性，演变为「一个生活方式品牌、一个时尚品牌」，并开始形成独特的品牌文化。他表示，萤火虫在用户群体上打破了年龄与购车预算的双重限制，「大学刚毕业的年轻人想买也能买，很有钱的老板想自己开车送小孩上学也完全没有问题」。需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位据《商业内幕》报道，前向部署工程师（FDE）的招聘需求在一年内激增逾 700%，正成为 AI 浪潮下科技行业增长最快的职位之一。招聘平台 Indeed 数据显示，今年 4 月该职位招聘帖数量达 5330 条，较去年同期的 643 条增长约 729%。Anthropic、OpenAI、Palantir、Stripe 及 Google Cloud 均在积极招募该职位人才。FDE 的核心职责是协助企业将 AI 整合进业务流程，并根据客户需求定制开发 AI 工具。薪酬方面，Indeed 数据显示该职位年薪约为 17 万至逾 20 万美元。最低 1 元 40 万 Token，三大运营商推出 Token 套餐中国电信昨日推出系列试商用 Token 套餐，分别面向开发者及中小微企业、个人及家庭用户提供「Token+连接+安全」一体化服务，均提供三档套餐，面向个人及家庭用户定价 9.9 元/月起 1000 万 Token；面向开发者及中小微企业 39.9 元/月，1500 万 Token。两类用户均可选购宽带上行提速包（上行速率提升 50 Mbps，100 元/月）及智能体安全管家服务（5 元/月）。个人家庭用户还可额外选购 5G-A 优享包（含 10 GB 流量，上行速率最高 300 Mbps，20 元/月）。联通、移动两大运营商也先后推出了各自的 Token 套餐：移动：中国移动上海公司于 5 月 17 日世界电信日发布会上推出 Token 通用服务，定价 1 元 40 万 Token。联通：中国联通上海分公司于 5 月 16 日发布全栈国产化「国芯国模国云」安全基座，推出多档 Token 产品及融合套餐。特斯拉 Robotaxi 碰撞事故报告解封近期，特斯拉正式解密并公开了向美国国家公路交通安全管理局（NHTSA）提交的 17 起「Robotaxi」自动驾驶碰撞事故的完整未脱敏报告。据 Electrek 报道，这 17 起事故发生于 2025 年 7 月至 2026 年 3 月，均涉及在得克萨斯州奥斯汀进行「Robotaxi」测试的 2026 款 Model Y，且测试时均有安全员在场。伤亡数据显示事故整体较轻微：13 起仅造成财产损失，2 起无人员受伤，1 起轻伤未入院，另有 1 起轻伤需入院治疗（系车辆在路口以 2 英里/小时蠕行时被后方 SUV 追尾）。官方报告表明，大部分事故并非特斯拉自动驾驶系统（ADS）的过失，多为车辆在红绿灯、停止标志或交通拥堵中完全静止时，被后方人类驾驶员追尾或侧擦。但数据同时也暴露出系统与后台干预的缺陷：两起事故发生在远程操作员接管期间，导致车辆分别以 8 英里和 9 英里的时速撞上金属围栏与施工路障；另有数起事故显示 ADS 存在空间感知死角，车辆在转弯或倒车时撞上了金属链条、拖车挂钩、木制电线杆及路缘石。💡 知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值据「听筒Tech」报道，知乎创始人兼 CEO 周源日前在第十二届新知青年大会上表示，AI 时代真实创作者的价值不可替代。AI 每分钟都在创造奇观，但 AI 时代仍然需要新知。而新知来自真实的人。每一个真实的人，都拥有不可替代的价值。大会主论坛上，多位嘉宾就 AI 与人类创作的关系展开讨论。复旦大学人文学者梁永安提出「科技有所为，如何让自己有所不为？需要人的尺度和人文属性」；作家张春称「不完美就是我们的完美」。周源在大会结尾总结道：「技术越发达，有温度、有深度的真人创作就越宝贵。」小米 YU7 GT 火山灰配色公布近期，小米汽车再度放出 YU7 GT 的预热图片，公布了新配色以及新车的内饰信息。其中小米 YU7 系列将加推一款名为「火山灰」的配色，官方介绍：灵感来自晨雾下的火山地貌，高纯度灰调，叠加纳米级矿物颗粒，呈现细腻而有层次的金属质感。同时新车内饰也迎来全新的「赛道红」：采用红黑撞色，提供专属刺绣。官方表示，小米 YU7 也会提供「火山灰」外饰配色，并且现在车厘子红的 YU7 GT 已经在部分小米汽车门店展示了。ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构近日，OpenAI 面向美国 ChatGPT Pro 用户推出了个人财务管理功能，支持通过 Plaid 连接超过 1.2 万家金融机构的银行账户和信用卡。连接账户后，ChatGPT 会自动读取用户的真实余额、消费流水和负债数据，生成一张覆盖投资、日常开销、订阅和近期账单的财务面板。用户提问时，模型可直接翻出账单历史，找出哪些类目容易超支，并给出具体的削减方案。此外，用户还可以告诉 ChatGPT 一些账外信息，比如「我在攒钱买车」或「我还欠父母一笔钱」，系统会将这些内容保存为财务记忆，用于后续对话。马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布马斯克昨日在 X 发帖披露，SpaceXAI 正在研发的 1.5T 参数 Grok 基础模型第 9 版（V9）已完成训练，预计 3 到 4 周内正式发布。现役公开版本 Grok 4.3 基于 0.5T 参数的 V8 基础模型，目前每隔数日迭代一次。V9 将参数量提升至 1.5T，马斯克称其为「重大升级」。在帖子的结尾，马斯克更是用用「banger」（王炸）一词来形容对这一版本的预期。字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本字节跳动 Seed 团队近日开源扩散

直到 2026 年五月，我们终于等到一整轮的影像旗舰登场完毕——小米做出了徕卡一瞬、vivo 打造了照片与视频兼具的 V 单、OPPO 在双长焦方案上带来了惊喜。在这轮旗舰更新潮的最后，华为提前将新一轮的 Pura 90 Pro Max 端上了桌。老实说，第一眼看硬件，心里是有落差的。去年 Pura 80 Ultra 上颇为惊艳的「一镜双目」结构不见了，取而代之的，是目前行业内最稳妥也最常见的大底长焦方案。不过，在细细体验后，我认为 Pura 90 Pro Max 藏了一个比纯拼硬件更有意思的新解法——XMAGE 智拍。全自动，都省心过去这几年，我系统性地梳理过华米 OV 的影像架构。虽然各家在硬件选择上各有理解，但大方向出奇一致，都在极其有限的机身空间里死磕传感器面积，比拼谁的光圈更大、进光量更多。在硬件上奋发图强理所应当。但在摄影中，不错的硬件基础仅仅是第一步。想要拍出一张好照片，构图与后期同样重要——构图保证一张照片在视觉上是否平衡、主体是否明确；而后期则决定了这张照片视觉质感和风格化。但这两者，向来都是具备门槛的手艺。而华为在 Pura 90 Pro Max 上，打算一口气解决这两个问题。熟悉他们产品线的朋友或许有印象，去年发布的 Pura 80 Ultra 上，华为试水了 AI 辅助构图。AI 辅助构图功能相对基础，在实际拍摄中主要起一个引导和提醒的作用。到了 Pura 90 Pro Max，这项技术完成了从「辅助」到「主导」的蜕变，并定名为「XMAGE 智拍」。剥开表层去探究 XMAGE 智拍的底层逻辑，是一套完全由端侧算力驱动的本地化影像流水线，将拍摄过程拆解为三个同时进行的模块——基于实时语义分割的自动构图、无视物理镜头限制的智能变焦，以及针对画面的色彩重塑。也就是说，在这个功能的加持下，Pura 90 Pro Max 可以自动识别主体、变焦构图，最后还排列出一系列适合当前场景的 XMAGE 风格供你选择。值得一提的是，XMAGE 智拍不需要网络连接，也无需将庞大的图像数据传回云端，全部基于端侧 AI 运转，实时识别画面里的主体轮廓、建筑线条以及光源分布状况，随后根据底层训练好的审美模型，给出一套包含最优构图、变焦和精细色彩方案的组合拳。繁复的专业参数在这里黑盒化，过去需要人脑去判断的构图和调色工作，全权交给了本地算法。除此之外，Pura 90 Pro Max 还顺势推出了 AI 姿势推荐，这个功能同样通过语义分割识别，可以理解画面中的了恩物、环境、姿势与背景，随后直接在取景框中勾勒出一个等比例的线框轮廓。拿手机的人只需按图索骥，引导模特贴合画面上的姿势即可。遇到不满意的动作能随时刷新，它甚至支持导入社交网络上保存的样片，一键提取动作精髓。听起来确实很厉害，但这套完全依赖端侧计算跑出来的流水线，究竟能不能应对现实世界里复杂多变的光线，还需要把它带到真正的街头去寻找答案。AI 有所能，也有所不能来到一家咖啡馆，墙上的多层光环灯饰引起了我的注意。举起手机，XMAGE 智拍自动变焦到 2×，将墙壁上的光环作为画面主体。这个选择可以说中规中矩，光环圆形处于画面中下方，整体构图稳定和谐。同时，操作台上颇具金属质感的咖啡机也是个不错的主体，XMAGE 智拍精准识别到我的拍摄意图，变焦到 192mm，保持咖啡机的中置。由于我想保留更为还原的室内光线与色彩，这张照片并没有套用 XMAGE 智拍推荐给我的任何色彩风格。用它拍了几组照片后，我认为 XMAGE 智拍是一个非常用户视角的功能——在传统的拍摄习惯里，手机焦段是一件需要精打细算的事。大脑会本能地去贴合物理镜头的原生焦段，生怕落在中间焦段会损失画质。但 XMAGE 智拍完全抛弃了执念。它不在乎用的是几倍变焦，转而完全为画面考量：当前这个画面，怎么裁切才最好看。哪怕不可避免地要损失一部分边缘像素，只要能让视觉中心突出、比例和谐，它就会果断动手。这种做法很实在，一切都在为了最终的出片服务。走出室外回望咖啡馆，绿植与橙色外墙交相呼应，XMAGE 智拍自动截取出色彩冲击力最强的部分，同时准确识别出咖啡馆建筑主体，两种大色块在画幅中占比相似，平衡而美观。再抬头，夏季树木繁茂，XMAGE 轻微放大焦段，让树枝形成天然的框架构图，广州塔正悬于画面正中间。在这组照片的拍摄中，我还察觉到了一个很有意思的技术细节。按下快门时，你会发现滤镜加载的速度，远大于构图改变的速度。这是因为滤镜往往只需识别画面后套用底层的色彩映射，而构图则需要端侧算力去实时识别画面里的建筑线条、光源分布和人物轮廓。计算量完全不在一个量级。更有意思的是，智拍套用滤镜的方式非常精细，除了 XMAGE 风格选择外，还会对调色盘做精细调整，确保风格对当前环境的适配度。当然，这套系统也有自己的脾气。在实际走街串巷时，XMAGE 智拍有时候会自动跳出来接管画面，有时候又毫无动静，需要你手动去点击唤醒。就算是手动唤醒的情况下，XMAGE 智拍也有几率出现风格已经选好，但当前画面无法找到最佳构图的情况。这种情况频繁出现在极其繁杂的环境中，由于元素过于混乱、信息量爆炸，算法的确很难从中剥离出合适的构图。当然，换个角度想想，也许此时的 1× 视角，就是当前环境的最佳构图了。万物皆可 AI，现在轮到摄影了185 年前，达盖尔发明银版摄影法，留住时间成为一种特权；2011 年，胶片巨头柯达传出破产消息，智能手机开始野蛮生长，把镜头塞进了普通人的口袋。回顾整部影像史，其实就是一部打破特权、技术走向大众的普惠史。到了今天，移动影像的狂飙突进，撞上了一堵无形的墙。手机镜头模组越做越大，与之对应的，是最近行业里一个颇具深意的传闻：受制于高昂成本与物理极限，下一代影像旗舰的「超大杯」，大多要面临停更了。这条路为什么走不通了？因为影像旗舰突飞猛进的背后，藏着一个被刻意回避的死结。决定照片质量的，永远是镜头后面的脑袋。普通人有懂美的眼睛，却跨不过光圈快门和构图比例的门槛。把一台堆满顶级硬件的手机，递给一个毫无基础的人，他按下的快门，大概率得到一张平平无奇的随手拍。顶级硬件如果只服务于一小撮懂摄影的人，无疑就会变成一种伪命题。在同行集体踩下刹车的时间点，华为提前把 Pura 90 Pro Max 端上了桌，而 XMAGE 智拍，就是破局的方法。在此之前，行业里已有不少 AI 落地，比如 vivo 让四季流转，OPPO 消除反光。只不过，它们大多是按下快门后的后期创意和修补。XMAGE 智拍往前跨了一大步，让算法成为整个拍摄动作的主导。这种做法谈不上是什么艺术层面的颠覆，但确实管用。降低门槛，让完全没有摄影基础的人，也能相对轻松地把一台旗舰机用出该有的样子。好的技术绝不该是一座孤岛。曾经的智能手机让人们拥有了「随时拍」的自由，现在的 AI 则进一步赋予了大众「拍得好」的能力。这或许是打破当前僵局的一剂良药，也是影像超大杯物尽其用的方法。

你真的需要墨水屏设备吗？主作者关注知道分子吴先先少数派作者精神状态卢克斯古德知道分子吴先先关注知道分子吴先先少数派作者精神状态卢克斯古德联合作者关注知道分子吴先先少数派作者精神状态卢克斯古德知道分子吴先先关注知道分子吴先先少数派作者精神状态卢克斯古德05/12 11:11Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。我是一个墨水屏阅读器爱好者。从学生时代拥有第一台Kindle至今，长期持有过的设备不下20台，Kindle、海信、文石、掌阅、Kobo、大上（Dasung）、汉王、墨案、小米这些主流品牌都用过，不少还是在少数派种草的。前两年，因为折腾机器上瘾，耗时又耗钱，我痛下狠心断舍离，取关了测评账号、非必要不购新机、旧机送朋友或挂海鲜市场，一度精简到三台。没成想，最近又被漫天盖地的墨水屏广告击中，没日没夜地看起测评，还在千挑万选后入手了两台新机。这不由得让我想搞清楚：自己在这个过程中到底沉迷的是什么？为什么单单会对墨水屏阅读器那么上头？所以，标题这个灵魂拷问，实际上是问我自己的。这不是一篇科普向的文章，也不会给大家推荐具体的机型。我主要想聊聊，墨水屏这个市场中，卖家和玩家的普遍情况和特征，分享一些自己粗浅的观察和思考。1如果你想看一篇墨水屏的科普，强烈推荐看这篇文章：重度使用这款墨水屏手机之后，谈谈我对电子纸的理解和预测。别看它是五年前写的，现在回看依旧是墨水屏设备的入门科普佳作。而且，当时的产品和如今的热销款型相比，除了参数上的提升，产品形态和实际体验并没有那么大差异，仍可借鉴。类纸显示技术显示器这篇会员文章也不错，同样推荐阅读。国内墨水屏市场有多卷？不了解阅读器的朋友，一聊起这个话题，第一反应都是Kindle。Kindle作为泡面盖的形象深入人心，而在2022年，Kindle宣布退出中国市场，23年正式关闭电子书城运营，24年结束云端下载服务，似乎更印证了这个品类的衰落：如果国内市场好，行业一哥Kindle干嘛要放着好好的钱不赚，对吧？实际上，故事并未像人们预想的那样发展。倒不如说，Kindle这个老大哥一撤，原本被压抑的国产品牌竟然把不温不火的市场给翻腾起来，起码明面上，越发活跃和热闹了。从图表上可以看到，从2022年开始，Kindle不再一家独大，原有的市场份额被文石、掌阅、汉王，还有科大讯飞等国产品牌迅速瓜分。此外，还有Bigme、Meebook这样的新生代品牌先后进入市场，也有一定的影响力。或许还有朋友记得，22年甚至连华为都来分一杯羹，推出了一款10.3寸的MatePad Paper，在反响平平后及时收手，后续再没出过墨水屏C端产品。2024-25年，整个行业迎来爆发期。我一直很好奇整个市场规模有多大。查了一下媒体数据，单就阅读器（主要指8寸以内小屏、不带手写笔的款型）这一个品类，25年的总销量为64.7万台，销售额达到8亿元。虽说总盘子不算大，但比24年同比销量增长33.4%，销售额更是增长43.5%，想必产品定价的水涨船高也助推了一把。办公本（主要指8寸以上大屏、可书写的机型）的累计销量为33.7万台，销额为11.0亿元，同比下降10%左右。再加上学练机型，整个墨水屏阅读设备的总出货量在127万台左右。对这么一个不算大众的品类来说，算是不错的成绩了。如今，如果你是一个小白，想购置一台墨水屏阅读器，打开购物软件，第一反应大概率是眼花缭乱、头晕目眩。首先是品牌。这个前面有提到，墨水屏厂商层出不穷。而且从去年开始，墨水屏品牌开始请明星代言宣传了。先是文石找了Papi酱代言Leaf 5系列，今年4月汉王找了李雪琴代言，主推Clear 7锦鲤。科大讯飞请的是李现，手里拿着一款号称薄如蝉翼的AINOTE 2。讲实话，看到这些明星，我并不觉得墨水屏出息了，而是心疼自己的钱包：明星代言费肯定都要消费者来买单，抬高原本就贵得离谱的产品价格，想到这里内心更沮丧了。有一说一，这几家代言人找得都不错，算是把娱乐圈看起来脑子不错的人都请到了。下一个品牌，该不会请刘旸教主或张踩玲吧……刚刚说到，行业调研机构给出了阅读器、办公本、学练机三分法。更细分的话，面向C端用户的主要有墨水屏阅读器、阉割版手机（能插卡上网但不能打电话）、墨水屏手机、大屏办公本、墨水屏显示器、还有学练机这几大类。今年还出现了一些墨水屏小玩意，比如最近相当出圈的阅星瞳，还有少数派的「Quote/0」「口袋先知」等。当然，这其中有不少品类是重合的，像墨水屏手机、办公本也可以拿来当阅读器。还有一些产品的核心卖点就是系统开放、功能全面，大概类似于效率界热衷的All-in-one吧。主流设备的尺寸也有不少：5.84寸、6寸、6.13寸、7寸、7.8寸、10.3寸、13.3寸，每个尺寸下面都有多款可选机型。有些经典尺寸如6寸、7寸、10.3寸，基本上各家都有新品，可谓卷中卷，也是最适合新人入手的尺寸。文石在铺品类、砸硬件上绝对属于先头部队。这是2020年的文石全家福，当时就这么卷了~不光产品迭代速度快，这一两年，墨水屏产品的新品细分更是卷到不合常理。以掌阅家7寸机型Ocean 5举例。这个机型和Kindle Oasis的设计一脉相承，屏幕比传统6寸大一些，还配一对侧边实体按键，定位上属于旗舰款。2025年10月，掌阅这个系列问世的时候，一下子推出了四个细分版本：Ocean 5标准版、Ocean 5长续航版、Ocean 5 Pro旗舰款、Ocean 5C彩色墨水屏版四款，它们之间电池容量、处理器、盖板类型都不尽相同，侧重点也各有不同，真是逼死选择纠结党的节奏。我截图了几款Ocean 5产品的详情页，能看出Ocean 5 Pro（左上）用了凹屏吗？标准版和长续航版（下方）竟然连宣传图用的都是同一个……新款墨水屏设备，真的比老款好吗？讲到这里，有朋友会说：行业繁荣是好事啊，商家越卷，就越能带动产品升级和技术进步。哦，真的是这样吗？了解墨水屏的玩家知道，墨水屏的核心科技就在屏幕上。这块屏幕由于物理特性，没有普通手机屏幕的屏闪和蓝光，确实有一定的护眼功能，我认为，这是墨水屏阅读器相比于普通手机平板的核心竞争力。然而，受制于墨水屏技术的发展（还有元太科技的垄断），这块屏幕的进步空间相当有限。严格意义上说，墨水屏技术算不上前沿高科技，基本上还在几十年前的基础上，做着修修补补的工作。图源：元太科技这就导致一个现象：虽然短短几年，屏幕参数从Carta1000，一路升级，狂奔到Carta1300，但实际的使用体验依旧不是那么理想。加上各家产品优化和调教水平参差不齐，有残影、刷新慢、屏幕良品率低、有坏点，依旧是墨水屏设备的常态。带前光款的设备中，漏光、前光不均、阴阳屏的情况屡见不鲜。这导致很多第一次接触墨水屏的朋友，浅尝一下便被草草劝退。就算你是个资深玩家，每次购入新品时仍会战战兢兢，希望自己能人品爆发，收到一台品相良好的机器，避免反复退换货、和商家扯皮。客观来说，硬件的技术瓶颈，不是哪个厂商的错，是行业共同的困境。然而，对标国外经典产品，国内墨水屏虽然产品多、迭代快，但实际品控、产品力还是不太尽如人意。用过Kindle的朋友或许有类似体验：你嫌手头用了十几年的Kindle太慢、用起来不方便，被各路广告种草，打算入手一款文石、掌阅等主流品牌Carta1300屏幕、300ppi的旗舰新款。当你欣然收到新机，把它和家里古董级的Kindle放在一起，对比一下文字的锐利度、前光的均匀度时，会惊讶地发现Kindle古董机大概率整体表现更好。这还是老款表现，更不用说24年新出的新一代Kindle和大屏Scribe，这些新款的显示水平更是碾压式的。这里要补充一句，我和好几个常用墨水屏的朋友交流过，别小看文字锐利度、背景清晰度、前光均匀度这些细节，重度使用后你会发现：不知道为什么，用Kindle或某款机型就可以看很久书，不觉得难受；但换了某品牌机型，阅读时间一长就会出现眼睛不适、疲劳的情况。实际上这些细节都很影响阅读体验，大家对标纸书，想想影印版、盗版的印刷质量，或是轻型纸、铜版纸等用纸选择，就很容易想明白。小红书上有朋友对比了KP3和文石Poke6（同为300ppi）的显示效果，我觉得就我线下体验店感受，显示差异还是挺明显的我手头有一台18年款Kindle Oasis2，一直没舍得出二手，就是因为它的产品质感依旧一骑绝尘，文字显示细腻度比我后来购入的十几台国产设备都要好。至于像Kobo这样专注于显示细节的国外品牌，更是难有任何国内墨水屏单品能在阅读效果上匹敌。不过，由于国外产品主要做封闭系统，传输不便，使用起来短板也非常明显。难怪业内流行一句讽刺：「但凡Kindle支持安装微信读书（不是像现在那样只能登录网页版），就没有国产厂商什么事了」。我最近之所以又开始关注墨水屏，也因为一件事。圈内的朋友知道，海信以前在国内墨水屏领域算是头部品牌，一直稳扎稳打，从19年相继推出的A5、A7、A9等墨水屏手机，至今仍有很多忠实粉丝。可惜的是，23年海信退出了墨水屏市场，迄今再也没有推出任何新款。海信退市前，好多品类疯狂打折，我顺道入手了一款22年3月上市的6.7寸阅读器HiReader，不能插卡，但样子像个手机，开放安卓系统，性能强劲，各方面都很趁手，这几年使用率非常高。最近几个月，由于文石和掌阅家也开始强推它们家的手机形态阅读器新品，我又开始种草，想迭代一个新款。考察了一下，文石小白马升级了Pro版（可插流量卡、5G、有摄像头，还有彩屏款可选）和小白马Plus款（不可插流量卡，支持手写，但单卖设备不带笔），掌阅则升级了Tango系列，推出了Tango2和Tango2 Pro（能插流量卡，带摄像头）——晕不晕？晕就对了。讲实话，我在实体店体验过小白马，虽然新款用的Carta1300屏幕比HiReader的Carta1000新了好几代，但屏幕质感、灯珠、运行速度等综合对比后，还是海信更胜一筹，直接拔草。掌阅Tango2 Pro虽然屏幕显示不错，字体更黑，更有跃然纸上的感觉，我收到的这一款品相也不错，但第三方软件运行起来，跟海信还是有差别。尤其是掌阅的刷新模式虽有5档，但在刷网页时，没有一个档位像海信「均衡」档更让我满意。要知道，海信真的好几年没更新墨水屏固件了。不过我还是留下了，毕竟算是有差异化亮点。网友对比HiReader（左）和Tango2Pro（右），具体观感见仁见智，但海信这款屏占比是真的优秀。举这个例子不是想拉踩谁家，毕竟海信刚刚又有重新进军墨水屏市场的意图，把HiReader机型烫印个2026字样，软硬件什么都没改就拿出来卖，价格还不便宜。通过这个例子，我就想问一句，虽然这几年国内厂商卷到飞起，价格也是节节攀升，但真正在屏幕优化、产品设计方面，大家有在用心吗？比不过Kindle，还可以找到借口，说亚马逊在供应商那里有优势地位，但和同为国内厂商的海信相比，26年那么多阅读器新品，个个价格直逼2000+，却无人完胜海信家一款4年前的老产品，真的让人感到遗憾。所以，我的结论是：新款肯定贵，但未必如你想象得那么好。如果你有一个还算趁手的旧款，没必要盲目换新。毕竟一台质量合格的阅读器，用上十年、二十年都没问题的。为什么喜欢墨水屏的人，会买了又买？虽然大家理性上都知道阅读器多买无益，还耽误真正的看书时间，但是实际情况是：一旦你拥有了一台阅读器，开始关注墨水屏这个领域，大概率往后会越买越多，手里不会只有一台墨水屏设备。刚刚说了，我前后拥有过的不止20台设备，但在墨水屏这个小圈子里，我这种水平真算不上什么。专业博主除外，你翻翻小红书，动辄十台八台的用户多了去了。如今把任意型号输入搜索框，都能找到大量的晒单或吐槽贴，这在以前是不敢想象的。这几年社交平台流行一个梗，叫「墨水屏繁殖」，讲的就是你买了一台墨水屏就会想买第二台，就像它自我繁殖一样。如果你对这件事没有感性认知，不妨看看这位玩过50多台设备的豆友的墨水屏大赏：论墨水屏阅读器如

快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验主作者关注Linery新手上路Linery关注Linery新手上路联合作者关注Linery新手上路Linery关注Linery新手上路05/13 20:36用微信发信息是我们熟悉到几乎成为本能反应的日常操作，不过作为国民应用，微信为了保证人人可用的普适性，也一直保持着极为克制的功能更新节奏。直到七年前鸟巢那场颇具争议的「灾难级」发布会上，锤子 TNT 次世代生产力工作站发布时，老罗在最后展示了他对「聊天应用」的理解——子弹短信。子弹短信以其极简高效的交互设计、处处贴心的产品巧思，让人耳目一新，也成为了那场发布会后少有的、备受好评的产品。可惜彼时锤子科技已经风雨飘摇，子弹短信最终没能成为击倒巨人的大卫，在变身「聊天宝」尝试「聊天赚钱」的模式也没能最终挽回颓势后只能草草收场。这个意难平也让我在用微信的时候总感觉有一根刺：如果聊天可以再「畅爽」一点就好了。比如我可以直接在一个界面里快速给多个联系人发私聊信息，这样在同时需要联系很多人的时候，可以不用频繁打开，关闭聊天窗口。我试着让 AI 帮我列举了一些需求场景：周末聚餐，要挨个通知朋友时间地点；临时加班，要分别跟家人、伴侣、长辈说一声晚点回家、不回去吃饭；工作上有事，要同步给同事、领导、合作对接人；出门用车、约球、约见面，要一一私聊好几个人。再比如更智能的群发，快速给不同的人发想同模板的信息时，如果能自动填充联系人的「称谓」等字段，也能让群发更人性化；我甚至也想用微信做更多的信息记录，比如聊完天可以记录一下感受、吐槽、资金往来、注意事项等。从锤子转到 iPhone 阵营后，出于一些「新奇」功能的急切需求，我开始捣鼓起了 iPhone 的「快捷指令」，用它来解决工作和生活中遇到的问题。并且最终我渐渐做出了一些围绕微信聊天的快捷指令。今天分享其中的一个。为了解决「不频繁打开窗口，同时给多个人发信息」的需求，我利用第一性原理得出了一个简单的结论：聊天本质是给指定联系人发送一段文本。所以不妨剥离冗余的交互，只保留最核心的「点名 + 发消息」。在设计时我发现快捷指令完全可以实现「输入联系人名称、再输入要发送的文本、最后发送出去的效果」：在锁屏上直接发微信的效果举点日常真实示例：妈 晚上我不回家吃饭了奶奶 今晚在外边吃，打完球给你带榴莲包回去万基 晚上七点球场准时 solo世平 好久没见，找个周末抽空聚聚Peter 方案你明早抽空看一下，下午拿给老板签字莞少 今明两天我要用车，你这两天别把车开走不用来回跳转、不用重复复制切换，静下心一次性把所有人的话都编辑好，一键自动群发，省心又高效。这个功能每次用都非常愉悦，一是聊天体验新奇（我觉的很像学生时代传纸条），二来是真的完美帮我躲开了碎片化通信对于正常生活和工作的干扰——作为潜在的 ADHD 人士我会经常陷入这种情景，点开微信本来只想发一条消息，结果一眼刷到朋友圈、被视频号内容勾走注意力，半小时过去，信息还没发出去。另外就是每次用它一次性把该对话的人该说的话说完，有一种特别畅快的感觉。它当然无法替代微信原本的日常聊天、连续闲聊、刷圈看动态的需求，但是作为一个于我而言的重要补充，真的非常好用、可靠。最后说点感性的话：我是一个非常怕麻烦的人，对所有可以减少冗余步骤的事都有极大的好感，但我也逐渐能理解这个世界上为什么会有这么多「麻烦」。麻烦对我而言变得中性，我意识到它代表了一件事的中间态，在优秀的产品（比如微信）中它常常体现为极致的折中，这同样也需要付出大量的工作、甚至是天才的创想才能达到的。这在我看来真的非常了不起。快捷指令就是 iPhone 提供给用户，用来化解这些「麻烦」的一个非常好的工具。它有丰富的接口，例如在这个功能里，就利用了微信提供的「给联系人发信息」接口；也有非常多的触发方式，包括但不限于桌面图标，锁屏、控制中心、负一屏小组件，锁屏快捷按键，轻敲背面，操作按钮（原静音拨杆）——当然还有少数派的 Bang!Case。这都让它变成了一个能让每个人通过轻度开发解决实际问题的工具。我真的非常喜欢快捷指令，它让我学会去缕析一件事的脉络、链路，动手去改写、制定新的规则和流程，让工作和生活都接近我想要的理性，跳脱出很多无形的枷锁。针对微信我未来还会做很多有趣的功能，比如智能群发，定时发送等等，感兴趣的朋友可以保持关注。我把这个快捷指令整理好了，有需要的派友可以安装试试，添加到自己趁手的位置就可以了。编注：题图由 ChatGPT 生成。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀3131扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@克莱德© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#自动化#效率#微信#快捷指令31等 31 人为本文章充电扫码分享举报本文章Linery还没有介绍自己关注

Apple WWDC26 全球开发者大会，来与少数派一起看主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。05/15 16:59北京时间 6 月 9 日凌晨，WWDC26 全球开发者大会首日 Keynote 主题演讲活动即将揭幕。届时，我们将迎来 iOS 27、iPadOS 27、macOS 27、watchOS 27、visionOS 27 等新一代操作系统的亮相，Liquid Glass 设计语言将如何打磨进化、iPadOS 多任务体验又将走向何方…… 这些大家关心的问题也将在会上悉数得到解答。与此同时，macOS 27 彻底告别 Intel 时代，会带来哪些新变化，也是今年 WWDC26 值得关注的看点之一。为了与同样关注 WWDC 的你一同迎接这场科技盛会，少数派将于 WWDC26 首日 Keynote 主题演讲活动期间在深圳举办线下直播观影活动。欢迎你来和我们一起交个朋友、一起聊天，观看 WWDC26 全球开发者大会。购票链接：深圳场次广州场次其它场次活动信息活动日程 (深圳)活动时间：06 月 08 日 22:00 - 06 月 09 日 03:00活动地点：广东省深圳市宝安区西乡街道银田工业路元匠坊 C 座 201 (少数派共创总部)活动日程 (广州)活动时间：06 月 08 日 22:00 - 06 月 09 日 03:00活动地点：广州市荔湾区泮塘五约外街 76 号少数派 (1200 书店旁)活动门票本次活动为门票收费制形式，门票价格 9.9 元，少数派会员可免费入场。所有门票收入将用于活动的组织和宣传，包括不限于场地租赁、茶歇准备、物料制作、伴手礼品等，由少数派方面负责管理，所有解释权少数派所有。编注：由于场地空间有限及后台系统方面的限制，少数派会员仍需下单购买门票后才可入场参与本次活动，下单时可享优惠金额 0.1 元。活动礼品活动门票定价为 9.9 元，除了入场资格之外，每位参与者还可获得：少数派 Prime 月度会员卡少数派 logo 装饰贴纸现场抽奖机会多城同屏，同步相庆本次 WWDC26 派对之夜线下特别活动，我们依然与少数派的老朋友 SwiftGG 共同举办，如果你不能参加我们在深圳和广州举行的线下活动，也可以报名在北京、上海、杭州、青岛同步举办的线下观影活动，现场同样准备了精彩的演讲和互动环节，并会为到场观众发放精美伴手礼品。了解详情 >Apple Developer Community Events在 Apple Developer 官方网站的社区活动页面中，Apple 已将少数派与 SwiftGG 联合举办的 WWDC26 系列活动收录至页面中，你也可以在此找到全球其它国家和地区经过 Apple 挑选的相关活动。>Community-driven events - Apple Developer温馨提示由于活动举行时间较晚，请购票用户合理安排行程，准时参加，并注意路途安全。活动伴手礼以活动现场发放实物为准。所有门票收入将用于活动的组织和宣传，包括不限于场地租赁、茶歇准备、物料制作等，由少数派方面负责管理，所有解释权归少数派所有。请在观看 WWDC26 Keynote 主题演讲期间保持安静，确保不会影响其他观众及周边居民。活动期间请注意个人物品的保管，建议提前规划活动结束后的归家方式，尽量避免步行或单独行动，保证个人安全。鸣谢本次系列活动由少数派、SwiftGG、漕河泾 Hi-Tech Park 等合作方共同举办，感谢各位合作伙伴对于本次活动的支持。如果你有其它活动想法或者场地推荐，欢迎通过邮件联系我们：[email protected]。153扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@waychane© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#WWDC26#Apple15等 15 人为本文章充电扫码分享举报本文章少数派编辑部少数派编辑部官方账号。关注

按：从播客访谈到深度长视频，越来越多的内容创作者把表达重心放到了音视频里；「先转录、再阅读」也因此成了不少重度信息消费者的日常工作流。。本期「装了啥」，我们请编辑部几位同事分享各自正在使用的转录方案，希望其中至少有一种思路能为你所用。MacWhisper 的平价替代：TranscribeX@广陵止息：虽然我很喜欢文字传递的高密度信息，但越来越多的内容创作者已经转向播客和视频；至于多人访谈类节目，音频传达出的现场氛围感也比文字更生动。所以我这些年消费音视频节目的时间也逐渐拉长，但有价值的节目「三心二意」地听一遍远远不够，反复听又没有足够的时间和精力。所以，现在遇到有价值的节目，我都会转录成文本，方便后续检索。从 Time Machine 里翻了翻，找到了当时的文件夹记录以前我采用的是whisper.cpp配合 OpenAI Whisper Large V3 模型来转录文本，但这个方案有两个问题。一是音视频下载的渠道很多，挨个手动下载很麻烦；二是转录时间偏长的中文视频时，有不小的概率会在转录过半之后出现重复输出、卡字，导致转录失败。但我实在不想翻 log 和模型仓库找相关讨论了，所以也开始寻找替代方案。首先想到的是 MacWhisper，看了眼当时价格已经涨到完全不考虑的地步。好在后来找到了整体体验和 MacWhisper 很类似的 TranscribeX，除了收录在 Setapp 里，对我没有额外成本外，即使单买也只要 MacWhisper 一半价钱。和 MacWhisper 类似，TranscribeX 支持本地文件导入转录、麦克风录制、录制特定软件、实时转录、视频网站下载转录和批量转录。其中「下载视频网站」功能主要依赖 yt-dlp 实现，所以基本不用担心国内服务的下载问题。转录模型方面，TranscribeX 也支持 WhisperKit、Parakeet、千问和 Mistral 的转写模型，必要时还可以调用 ElevenLabs 的在线 API。不过我本地基本只用 Whisper Large V3，别的模型也没用过，所以暂时不评价转录效果。相比 MacWhisper，TranscribeX 支持 Apple Silicon 中的 NPU 加速，这也是我之前选 whisper.cpp 的原因。NPU 虽然不会让转录速度快很多，但至少负载不会全部堆到 GPU 上。设置时需要注意：Audio Encoder：主要负责「听懂声音」的环节，可以只用 NPU；Text Decoder：主要负责听写的环节，为了转写速度，建议设置为 All。转写界面其实每个软件大同小异，但还是有一些设置可以说一下。首先是自动检测「说话者」：Pyannote：适合会议、访谈等场景相对固定的情况，也更贴合我常用的场景，可以快速调用预定义的模板；Sortformer：偏向速度、实时和轻量化的处理方式，适合会议、通话等实时场景。外语苦手也不必担心，TranscribeX 内置了三种翻译方式：Apple 内置的翻译 API、DeepL API 以及第三方 AI API。Apple 内置「翻译」速度很快，准确率也够用，而且免费；对翻译质量有要求的话也可以调用 DeepL 或者其他 AI。TranscribeX 也随大流加入了 AI 对话功能，可以针对转录的内容做进一步处理。最后值得一提的是，针对 Setapp 版本的 TranscribeX，所有的在线 AI 额度都可以用 Setapp 的 AI 配额。但 TranscribeX 还是有自己的问题：没有 CLI 版本，远程使用一定要靠 VNC 连接到 Mac；辅助框架的设计上也有不小的问题，目前还很难配合比较火的 AI 助手。如果你比较在意自动化体验，TranscribeX 可能没有那么适合你。用 Gemini 整理@PlatyHsu: 我这里主要谈谈转录后的整理。对于这类任务，我最推荐 Gemini 系列模型。主要原因是，在排名靠前的模型中，Gemini 的多模态支持比较全面，Gemini 3 全系（Pro、Flash 和 Flash Lite）都可以直接接受音频和视频文件输入，免去了先创建音频转录或者视频截图的功夫。我也认为 Gemini 的中文能力在国外模型中相对是比较好的。与此同时，Gemini 的用量给得比较慷慨，AI Pro 方案很少遇到不够用的情况。我最常用的场景之一是整理播客的录音稿，用来确认一些初听时一闪而过的模糊细节。一个参考提示词是：转录用户提供的音频，并整理成表达规范、分段清晰的书面格式。区分说话人，并标注在段落之前。将较长的片段合理拆分成若干自然段。可以删除语气词和重复片段或修改病句，不要做其他删减或总结。具体模型方面，一般可以先从 Flash 开始尝试，对于日常主题的节目来说一般就够了。一小时的播客大约需要四分钟处理，肯定没有本地模型快，但是成品高度可读，值得这个等待时间。不过我也遇到过一些涉及哲学、历史等较专业话题的节目，用 Pro 模型会有更好的表现，可以灵活调整。当然，如果想节省 token，你也可以先在本地转写一遍，然后用以下提示词整理文本：将用户给出的音频转录整理成表达规范、分段清晰的书面格式，可以删除语气词和重复片段或修改病句，不要做其他删减或总结。这个用例一般建议直接使用 Pro 模型，因为本地模型转写的文本一般有很多拼写错误和赘余，Pro 能更好地还原和修复。除了音频之外，Gemini 也可以用来总结视频。并且由于它可以直接「看到」视频内容，因此即使是那种没有旁白的 TVC，也可以说出个一五一十。例如之前一个项目中，甲方提供的写作素材之一是一条广告片，我就是用这个方法把其中的信息提取成文字描述，然后用在了写作里。以下是用苹果最近一条 Mac 广告做的演示，提示词是：详细描述该视频的主题、思路和内容，并列举关键帧的时间戳和画面内容。因为涉及较多「言外之意」的解读，这个用例也是用 Pro 模型处理为佳。以上方法在 Gemini 网页版和 AI Studio 都可以使用，但网页版经常会出现「偷懒」的情况，只给出非常简略的输出，因此建议尽量用 AI Studio 来操作。命令行工具搭配 Agent@waychane: 由于常有参加线上会议的需求，会中涉及主讲者的介绍、采访以及 QA 等内容，这类会议的时长动辄就有一个小时。在会后创作内容时，如果能有一份观点清晰、结构清楚的要点大纲，会省下不少功夫。现在可以说是 Agent 时代，我自己也在用 Claude Code 等 Agent 类 AI 工具帮我处理各种工作。因此在挑选特定场景下的 app 时，我会希望它们要么提供 CLI 操作方式，要么就是纯粹的命令行工具：一方面，这样可以在同一个 Agent 界面中高效地调用不同工具；另一方面，许多 app 不仅功能复杂，使用参数也多到难记，用 Agent 调用刚好能解决这些问题。面对会议音频转录的需求，我选择的是 whisper.cpp 这个命令行工具。它是 OpenAI Whisper 的 C/C++ 移植版本：OpenAI 官方版 Whisper 基于 PyTorch 的 Python 实现，要跑起来得先配好 Python 环境和一整套依赖，在 Mac 上只能依靠 CPU 计算，想用到 GPU 还得自己折腾 PyTorch 的 MPS 后端，体验并不顺畅；whisper.cpp 则用 C/C++ 重写了推理部分，原生接入 Apple 的 Metal 后端和 Accelerate 框架，能有效利用 Apple Silicon 的 GPU 和统一内存架构，模型权重也换成了体积更小、加载更快的 ggml 量化格式。

本周看什么 | 最近值得一看的 8 部作品主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注潘誉晗少数派作者潘誉晗关注潘誉晗少数派作者关注利兹与青鸟少数派作者利兹与青鸟关注利兹与青鸟少数派作者关注SHY少数派作者SHY关注SHY少数派作者少数派编辑部 等 4 位作者05/15 18:00☕️ TL;DR近期佳作推荐：[国产] 低智商犯罪、[韩剧] 夺命许愿、[英剧] 好兆头 第三季、[法剧] 抢钱女王团、[电影] 蒙特利尔，我的美人、[动画] 石纪元 科学与未来 第 3 部分、[纪录片] 马克斯·维斯塔潘：新征途、[真人秀] 旧洞基朋友们几则精彩预告：《攻壳机动队 THE GHOST IN THE SHELL》新预告、《沉默的审判》释出预告、《火遮眼》新预告、《爱情城事》定档预告、《奥特曼起源》正式预告几则影视资讯：《诺曼底 72 小时》内地定档 6 月 6 日、《宇宙巨人：希曼崛起》内地定档 6 月 5 日、《家弑服务》内地定档 5 月 29 日、《超级少女》内地定档 6 月 26 日、漫威新剧《幻视追寻》正式定档往期推荐可点此或点击文末 tag #本周看什么，更多主题片单可点此查看；想要查看更多深度影视内容，可以关注我们的衍生栏目「看什么 Café」。如有其他建议也欢迎在评论区提出，感谢大家对「看什么」栏目的喜爱和支持。[国产] 低智商犯罪关键词：喜剧 / 犯罪片长：45 分钟左右（单集）× 24 集观看渠道：爱奇艺/咪咕视频；豆瓣链接运气之神降临！@利兹与青鸟：平安市刑警张一昂能力不强但热爱吟诗，常常不合时宜得令人发笑，被下放到老干部处都能惹出乱子。因为他正巧捡到一封举报信，于是和同样脑回路清奇的李茜一起被派到三江口，调查公安局副局长的死是否有隐情。但空降而来的张一昂很不受欢迎，工作处处受限，却在阴差阳错之下抓住一个在逃嫌犯，让人刮目相看，就这样被架起来成了神探。另一边，平安市两个笨贼想出了抢贪官的主意也来到三江口；而三江口涉黑的集团老板周荣考虑着新项目，也着手调查发小叶剑刑警队长的死因。三条故事线、两桩凶案（可能不止）就这样汇集一处，想必会给三江口这一团乱麻摩擦出不少火花。本剧改编自紫金陈的同名小说，一开局就满是带着黑色幽默的荒诞味，罪犯和警方彼此高估，虽然智商时常掉线，但不妨碍瞎猫碰上死耗子，不禁让人好奇这份「好运」还能带来什么啼笑皆非的巧合。马旭东、闫佩伦、詹鑫等喜剧演员的助阵也让笑点呈现得更好，尽管罪案线稍显拖沓，但也算是合格的电子榨菜。5[韩剧] 夺命许愿关键词：剧情 / 恐怖又名：기리고 / Wish Your Death / If Wishes Could Kill片长：44 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接为了愿望达成，你愿意不惜一切代价吗？@潘誉晗：这次数学考试的结果有点出乎意料，因为崔亨旭拿满分了，可这家伙不是个读书的料子。面对好友们的询问，亨旭得意洋洋地把一个名叫「Girigo」的软件链接分享给了伙伴，说是只要写上自己的生辰八字贴在自己的身上，然后录下视频上传到这个软件上，就能实现愿望。大家并不相信亨旭的话，亨旭也没在意此刻出现在他手机上的倒计时。谁承想 24 小时后，亨旭就当着全班同学的面自缢了——原来这则所谓的许愿软件，在愿望达成后会要你的命。Netflix 韩国这次又推出了一款爆剧，一经上线就在 13 个国家登顶。剧集整体的节奏很是紧凑，恐怖惊悚的氛围也营造得相当出色，很好地把校园青春剧和恐怖因素结合在一起，然后又完美连接了手机 app 和萨满巫术。神秘的夺命软件利用了青少年的欲望，随后就像是一个大家无法挣脱的隐形牢笼，不仅顽固地无法在手机上卸载，更是在许愿者身后步步紧逼、致人于死地。2[英剧] 好兆头 第三季关键词：剧情 / 喜剧 / 奇幻又名：Good Omens Season 3片长：99 分钟观看渠道：Prime Video；豆瓣链接我有那片宇宙，我还有你，我已经拥有我想要的一切。@SHY：成为至高大天使的亚茨拉斐尔，开始筹备「第二次降临」，复活的耶稣却不知所踪，只得向自暴自弃的克劳利求助。这对自创世之初就有交情的天使和恶魔，能否像往常一样弥合分歧，而在背后作祟的幕后黑手，又会将整个世界导向怎样的结局？受原作者兼前两季主创尼尔·盖曼的性侵丑闻影响，最终季被砍成了仅 1 集的特别篇，篇幅势必捉襟见肘。不难想见，成片处处流露着删减的痕迹，忽快忽慢的节奏让叙事更加混乱，整体观感有着不小下滑。虽然填充的内容变了样，两位新编剧还是尽力还原了盖曼构想中的结局，而在破碎的字里行间，也能窥见一些残存的灵光。惯常的黑色幽默难以撑场后，本季的重心落在了本来带点调侃的父母爱情故事上，大卫·田纳特和麦克·辛以一如既往的真挚演技力挽狂澜，继续迸发充沛的化学反应。彻头彻尾的粉丝向结局，评价两极分化在意料之中，好在，本作原本就是作为限定迷你剧立项，如果你有点难以接受这一季的走向，只把第一季尊为正统也未尝不可。发表首评[法剧] 抢钱女王团关键词：喜剧 / 惊悚 / 犯罪又名：Les Lionnes / Cash Queens片长：55 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接我们之所以抢钱，真的是迫不得已。@潘誉晗：罗莎这天找到自己的上司，希望可以提高她的信用卡额度。可上司不仅很快就拒绝了她的请求，甚至通知她的额度被降低了，因为她征信由于入狱前夫的债务问题受到了影响。罗莎不敢相信上司的话，也就是在这时，银行的警报铃响了，原来这天又是银行例行的演戏日。看着上司希望职员们保障生命的大喊声，一个不该出现的念头在她的脑海中出现了——她决定伙同自己的好友，抢劫自己工作的银行。Netflix 法国于两个月前推出了一部女性犯罪喜剧《抢钱女王团》，讲述了五名女性在窘迫生活的逼迫下，不得不伪装成男性抢劫银行的故事。全剧充斥着大量很荒诞又很现实的设定，主角团的五名女性拼命工作、好好生活，但却一直过得不幸福。职场上的性别歧视、身上背着重债、被同侪做局丢掉了工作、被丈夫家暴……不是说努力就能有好结果吗？为什么在她们的身上，越努力却似乎越悲伤呢？发表首评[电影] 蒙特利尔，我的美人关键词：剧情 / 同性又名：Montréal, ma belle片长：117 分钟；豆瓣链接被困在夹缝中时，是面对自己还是活在谎言里？@利兹与青鸟：凤霞一家已经移民加拿大十几年了，在魁北克省的蒙特利尔开着一间小超市。丈夫性格较为古板，苦于没有当地学历，找工作频频受挫；大女儿正值叛逆期，对待父母多是指责和抱怨；儿子也只是上小学的年龄，家中没有人能为受困于更年期的凤霞提供情绪价值。在法语课上，凤霞被年轻人追求梦想的热情所鼓舞，尝试接触同性交友网站，并结识了 30 岁的卡米尔。这让凤霞重新找回了自己，却引发了家庭危机。本片的华裔女导演很有相关生活体验，她用代际间新旧观念的碰撞、婚姻大于爱情、照顾家人大于自我实现等细节，呈现出一个传统而压抑的中国移民家庭外壳。片中演员也多是当地华裔，汉语都很纯正，对华语区观众的耳朵十分友好。电影虽是在讲中年女性在道德边界寻找自我，却没有激烈的口号，而是用朴素温和的风格、平缓的节奏，给主演陈冲留足表演空间，为年过半百却天性浪漫、依然优雅而富有魅力的凤霞赋予灵魂，细腻又克制地演绎出她内心的彷徨与挣扎，也让观众沉浸其中。女主角是陈冲啊？1[动画] 石纪元 科学与未来 第 3 部分关键词：漫画改 / 科幻 / 冒险又名：Dr.STONE SCIENCE FUTURE 第3クール片长：24 分钟（单集）× 13 集，每周四更新观看渠道：巴哈姆特动画疯；豆瓣链接We choose to go to the Moon.@SHY：孤注一掷的抗争后，科学王国终于取得对北美阵营的胜利。面对迫在眉睫的威胁，千空和赛诺搁置矛盾，联手打造宇宙飞船，前往月球探寻石化的真相。为了复活更多人类，同时取得不可或缺的材料，众人兵分三路执行任务，「WHY 人」却传来了更加危险的信息。从石器时代到现代文明，跨越 200 万年光阴，这段从零开始的科学冒险终于迎来最终幕。从刚苏醒时手工打造的粗糙制品，一步步重现火药、铁器、蒸汽、电力，来到本季，科技树终于攀升到计算机和互联网的领域。要实现登月的最终目标，需要远胜以往的人力和物力，千空等人因此踏上环球之旅，整合各个国家的资源。即使原料由石头进化到超合金，舞台由村落跃进至外太空，本作的核心魅力始终如一。秉持理性和探索精神，用科学之力对抗奇幻，挥洒荡气回肠的热血，开拓人类未来的道路，诠释纯粹的浪漫。这个集结全世界力量、于石之世界铸造的「阿波罗计划」，绝对能让人心潮澎湃，剧中历尽千辛的角色们，也值得最为美好的结局。发表首评[纪录片] 马克斯·维斯塔潘：新征途关键词：纪录片又名：Max Verstappen - New Ground / 极速新域：维斯塔潘的北环之路片长：23 分钟左右（单集）× 3 集；豆瓣链接出发吧！汽车人！@潘誉晗：对于喜欢一级方程式赛车的人而言，效力于红牛车队的马克斯·维斯塔潘，是一个了不起的名字。他出生于赛车世家，自己的赛车成绩也非常傲人。不过卸下这一身份，他其实还有另外一种人生——一家 GT3（Grand Touring 3）车队的老板兼赛车手。而这部只有三集的纪录片，就是讲述维斯塔潘的第二身份。「看，老板亲自上场了。」维斯塔潘没有老板的派头，招呼完就直接上了场；他也不需要拥有全部欢呼与掌声，而是会亲自去挖掘赛车手。于是，他发掘了克里斯，这个和他一样从开卡丁车（并且克里斯的车号是 3）开始的小伙子很有天赋，可苦于 15 万欧元的培训资金，只能把梦想转移到模拟器赛车上。当维斯塔潘找到他，让克里斯在场地上尝试开车的时候，克里斯甚至没有驾照。相信在观看了这部纪录片之后，没人会否认维斯塔潘对赛车的热爱。他脱下了红牛的飞行夹克，在另一条赛车的道路上，继续驰骋。发表首评[真人秀] 旧洞基朋友们关键词：真人秀又名：구기동 프렌즈 / Six Singles Under One Roof片长：90 分钟左右（单集）× 10 集；豆瓣链接一起同居吧。@潘誉晗：韩国最近又推出了一部真人秀《旧洞基朋友们》，找来了张度妍、张根硕、安宰贤、李多熙、景收真和崔丹尼尔六位年龄想尽的艺人，过一次同居生活。随着明星观察类的项目在市场上越来越多，我相信有越来越多的人对这一类真人秀感到厌倦。而《旧基洞朋友们》的出现，再次证明了韩综的创意。看似依然是粉丝向地找了一群有粉丝基础的演艺人员，但是却找准到了一个很细腻的切口——让几个曾经独居过的明星们，过群居生活。独自生活，只是光看到就觉得很自由的日子，为什么他们体验过那种日子后，不愿意再独居了呢？我想这也就是观察类节目出现的初衷，去探讨一个社会性的议题。毕竟「存在即合理」。而这六名 85 后明星的同居日常，也非常有看点。他们都很有个性，但又能互相融入、靠近。见惯了大场面的他们，看到 18 公斤的章鱼，依然会手足无措，还是会大声尖叫。原来群居也不差嘛。1更多[电影] 总统的蛋糕@利兹与青鸟：萨达姆的独裁统治下的伊拉克，平民们过得水深火热。小女孩拉米亚住在草编的窝棚里，连书包里的苹果都要被老师搜刮走。更不幸的是，她被抽到要在总统生日这天做好一个奶油蛋糕。于是拉米亚踏上了旅途，用家里所有值钱的东西换取做蛋糕所需的材料。电影在去年戛纳电影节获金摄影机奖，摄影优美，配乐充满阿拉伯风情，有触动人心的童真，也有伊拉克战乱下的可怖的倾轧。[电影] 平场之月@潘誉晗：电影根据朝仓加守美同名小说改编。青砥健这一天意外偶遇了自己的初中同学须藤叶子。此时两人已经 35 年没见了。中学时两人就互有好感，这次重逢，已离婚的青砥健和丧偶的叶子，终于还是走到了一起。埼玉小城里这一对中年人迟到多年的遗憾恋曲，就像是终于遇见了自己的另一个缺失的半圆，内敛、沉稳又有淡淡的愁绪。[韩剧] 浪漫的绝对值@潘誉晗：每当晚上，看似平凡的高中生吕意珠就会偷偷创作 BL 小说。可惜，她的作品一直无人问津。这天，隔壁搬来了

厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨主作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了联合作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了05/15 00:07Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。免责声明与作者所需披露的一切首先声明一下——本人仍然是在校的本科微电子硕士考古学的大学生，并没有经过任何的传媒学相关的学习，也没有从事相关行业的经验。目前仅有的和媒体沾边的事情是在酷安、少数派等平台发过一些文章（其中有很大一批都是在最后诉诸于，说好听点是真实生活场景，说难听点开始抒情的文章），以及一个 B 站粉丝量远没有破千的账号，因此我远远没有承担过养一个工作室的压力与责任，也没体会到在竞争激烈的环境中生存的艰难。因此以下内容不构成任何形式上的专业意见，只是一个愤青在 Random Yapping。多有冒犯之处，但我本心无意去批评任何具体的某一个媒体老师，只是想尝试提出一个也许还算有点价值的问题。以下内容提到「游戏区」，也仅限于单机游戏，不包含手机游戏、桌游等类型。以下内容提到「数码区」时，通常仅指以手机、一部分 PC 组件为主的主观体验向评测，且我的能力不足以涵盖 DIY 区1以及涉及到实验设计与数据分析的其他评测内容，虽然随着硬件厂商的创新进入后期演变为渐进创新，产品节奏也有所放缓，在社区中也能听到「DIY 区不如以前」的感慨，但本文尽力按照「不理解，没有切身观察就尽量少乱锐评」的原则展开，但撰写过程中有很多内容没收住，因而请让我提前声明一下——以下很多内容只是社区观察，不完全代表事实判断。最后，说句坦诚的实话，这个内容写的过程真的很痛苦——我真的不想让这篇文章变成直球辱骂媒体老师、攻击厂商或者是殴打观众的情绪发泄文，从个人发展角度来说不希望因此在进入之前就已经和某些职业无缘，也不想因此得罪那些因真心倾慕而认识的那些做这一行的朋友，且我真的不希望因为自己的过分言论把发文章的平台本身也拖下水，但因为笔力实在是有限，尽管拉着 Gemini、GPT、DeepSeek 等一众大模型核验，也找了一些朋友探讨，尽可能不要直球攻击特定的一两家媒体也不要说的太过分，但是字里行间还是有一股火药味拼尽全力消不掉。真的很抱歉，尽管我其实将自己也包括在本文的攻击射程以内——我自己就是后文中所批判的那个自认为掌握了真实，在没有厂商利益束缚与照顾社区的必要下肆意输出内容，甚至有很多是过当内容的小自媒体。这些内容因为是我自己一段时间内自我质疑与焦虑的外化，因此，本文必然存在一些因深度不足、角度有限而导致的难以避免的暴言，简而言之——就是年轻时犯下的错误。麻烦各位读者多多体谅。也欢迎您在评论区里多批评指正，让我多与您交流学习。以下内容如果有哪里说的不合适的地方，有什么可能涉及到的法律争议，请尽快联系我去撤换文章中对应的部分，谢谢！本文长度约为三万字左右，前三部分观点将会较为尖锐，会举几个相对较为敏感的案例来佐证观点，后两个部分则尝试解决前文抛出的问题，内容相对温和但字数也相对较长，请坐和放宽。主观口胡系列是一个尝试将动漫、电影、游戏叙事与更加严肃的讨论结合起来，去尝试探讨一些特定大命题的个人尝试，可能会一定程度上背离这些作品的创作初衷，聊的可能也会比较混乱且随意，甚至会有意提取个别元素，曲解原文本来聊一些其他领域的内容。这次我想把一个大约是三四年前就开始产生的想法去做进一步的深化，并尝试结合近些年来我对游戏区与数码区的一些粗浅且悲观的看法，来尝试去聊一个问题，一个也许注定没有答案，跨多少领域学科都难以得出解法的一个小小的问题——主观体验向评测的现状怎样，尽头到底为何方？又是什么构成了它们的专业性？Part1：太多水，太多奶酪——主观内容的梦魇虽然这么说有失偏颇，但我一直觉得游戏评测是一个相当主观的内容——评测者的个人阅历2、偏向的游戏类型甚至是组成一个游戏的系统，对 GDC 中诸多开发者分享的设计理念的理解，以及同期的同类型作品等等——这些都会对你最终得出什么样的结论，给游戏最终打什么分造成影响。所以在因近期的《神探杰克鼠》评测而陷入争议的 IGN 内部——作为一个有 TGA 评选权的，相对来说算是行业内位置靠上话语权比较大的游戏媒体——他们在测评游戏时给出的流程是：在内部筛选最适合某类游戏且当下仍有档期的编辑，由他们在内部其他编辑的帮助下完成对于这个游戏的评测，给出评分，在内部经过审核与排期，最终以文章 + 视频的形式先后放到公共平台上。但是，这个时候总会有个但是。虽然这样的流程确实在内部行之有效，为行业贡献了一批质量很高的评测，尤其是在神作频出神仙打架的那些年里——但同样不可否认的是，无论你之前做对过做好过多少次，一旦出现几次测评结果与大众预期不相符，甚至出现有失水准的内容，就很容易在玩家社群中丧失公信力，变成 IG8——《宝可梦欧米伽红宝石/阿尔法蓝宝石》的 too much water 扣分，《超级马里奥 3D 世界+ 狂怒世界》的 7.0，《死亡搁浅》的 6.8——你知道的，日光之下从来都没有什么新事。而且当我们把视角继续向下探时，会发现这样的问题不仅仅是发生在 IGN 一家，Polygon 的《DOOM(2016)》、VentureBeat 的《茶杯头》（他的操作已经不足以让他完成第一关，只能借其他同事存档继续评——尽管他据自己说已有着超过二三十年的评测历史）……所以到底发生了什么？当然，抛开社区中已有的被讨论过无数次的，对评测者的质疑不谈，我们也许可以从更加细致的视角去看这件事。就像你在做数码类评测会提到机型的性能与续航、外观、系统流畅度与体验、拍照、外放表现，但你不可能在非首发上手的评测里一样面面都要点到，而是根据个人的使用习惯有自己的偏好一样，IGN 的每个评测者在评测游戏时也会有自己的明显倾向，让我们以当年评测《死亡搁浅》的 Tristan Ogilvie 为例——他评测一款游戏的重心是动作系统、交互反馈与流程的顺畅程度。在我眼里，相较于小岛在前作《合金装备 5：幻痛》中用一系列主线支线任务，吸引玩家不断的从母基地回到开放世界，完成一串任务后撤离不同，本作让角色更加持久更加连续更加长时间的停留在这个世界里，将任务可视化，变为地图上要到达的目标点，并在沿路设置一系列需要潜行的敌人据点、风景奇观，并且可以通过基建加速到达目标点跳过一部分游戏内容来优化体验的开放世界而很不幸的是，本作正正好在她关注的几个角度精准的踩中了地雷：虽然相较于《合金装备 4》有所收敛，但游戏内不断插入冗长的过场 cg，以至于从室外场景进入到地图传送需要连续跳过数个动画；游戏流程放速过缓，且游戏内的任务设计被认为千篇一律，主线和支线均由送货构成；动作系统相较于前作乐趣有所减少（尽管二者的游戏定位有着根本性的不同，也许从来就不该这么要求角色的性能），击败 bt 后的奖励反馈也较为缺乏，以至于编辑出于他所偏向的几个维度的也许从传统游戏性的角度合乎情理的考虑，选择给这个游戏 6.8 分。随后我们都知道发生了什么——之后的几年让游戏的异步社交大放异彩，以一个非常奇怪且迅速的方式，让小岛秀夫这种几乎像是寻求争议，找骂、试图以先锋艺术电影的形式在游戏内进行的表达，能被更多人看见，被更多人共情，似乎小岛秀夫在《合金装备 2》后，再一次，等到了时代来验证他思想上的前瞻性。因而，时代背景变迁下大家对小岛独特表达的广泛接受、小岛秀夫的明星制作人光环以及评测中确实存在的一些过当的指责，都让玩家社群反过来开始质问一件事：为什么我觉得好玩的游戏，你却要给出这么低的分？当然，IGN 作为行业头部媒体，其本身在游戏界有较大的声量较高的影响力，且也有着个人做自媒体乃至中小型媒体难以企及的资源，因此它说的话做的评测理应被玩家更加严格的审视，树大招风不无道理——而且至少在我举得几个例子里，引发争议的游戏本身也是在不同领域内质量过硬的存在，只是随着营销的不断发展，疫情时代与后疫情时代每个亚文化小领域的的扩圈与自媒体时代的到来，这种原本对媒体对权威的的尚处合理范围的不信任将逐步扩大，直到这种现象一步步扩散到每一件事，每一个领域，乃至我们每个人都难以避嫌，难以无视。Part2：内容井喷的时代，同质化的幽灵与信任的坍塌数码圈从来不缺学生出身，单纯出于热爱，分享自己体验的自媒体，这样的评测形式本身并非没有价值，它用更加平易近人的角度去涵盖一个产品中用数据评测所难以涵盖的部分，尽管也许，分享自己对于刚入手的数码产品的主观体验并没有什么门槛——有着中专及以上的学历就有足够的语言组织能力去清晰讲述自己的感受，再来一个相机甚至是手机去拍摄素材就能实现，但撰写一篇好的主观体验评测文（不管是哪个领域）就像是一般人眼中的历史研究一样，都觉得准入门槛低，都觉得写的内容是任人装点任人阐释的小姑娘，但真的想要尽可能做到内容详实、条理清晰，往往需要长时间的写作经验积累与大量的训练，而如果想在这个基础上再去追求与社区中主要讨论的内容截然不同的全新角度，或者让我们思考的更加狂野一些，在这个内容上做出深度，探讨一个产品背后更加深刻的有广泛意义的命题，则更加考验个人的水平——但随着互联网的高速发展与信息的高速创作、流通，我们正处在一个内容极其过剩的时代。虽然我承认，我是随着酷市场转型为酷安而踏入数码圈的，本身并不算什么数码区老资历，但你如果要我去回忆在我进入数码圈后这有限的时间内，主观体验向的数码自媒体井喷的那段时间，我想大概是在「老师好我叫何同学」的 5G 视频兴起的 2019 年 6 月前后，以及疫情肆虐的 2021 年后到 2023 年末。前者因为正好切合了国家的 5G 发展战略的需要而被主流媒体报道，被社会广泛关注，因而给圈子引来了泼天的流量——事实上那段时间坚持做手机体验类评测的规模还比较小的自媒体，如今都已经成长为几十万粉丝的台柱，比如我当时看的比较多的花生说；而后者则是由于人们长时间居家，在室外的娱乐方式因客观需要而受限，因而人们开始转向互联网，大约是在这一时期前后，一群赛道更加细分，比如上古数码产品体验、软件交互等领域（就我当时看的比较多的来说是六分超超和苏星河）的体验向自媒体开始大规模兴起，游戏区的很多如今声量较大的自媒体编辑老师也多是在这一时期进入相关行业或者是投身做自己的账号。但与自媒体账号井喷相对应的，却是——请让我想想，该怎么让人更好接受的去讲一个有点不近人情的结论——极其迅速的内容同质化。游戏不是一个竞争激烈且产品间带来的体验差别相较于同类型的数码产品更大的存量市场，有足够多各自差别足够明显的新游戏不断冲刷着视野，因此这种同质化的倾向也许还并不明显，但在数码区，当每个上游的厂商都在绞尽脑汁的结合热门概念，生生在软硬件体验上捏出极其有限的新创新点时，这种感觉就尤其明显——对于自媒体而言，相比起「你自己要讲述什么故事」，你的阐述对其他人有什么价值才更有传播的意义，因而当你迅速的因不受束缚、观点尖锐，获得了一批受众支持迅速爆火后，同质化的幽灵便也开始悄然显现——是顺着原先让你得到人们簇拥的路径一路向西，对着即使没有那么差的内容输出情绪，直到有一刻在无数次路径依赖后判断失误，犯了众怒，还是妥协于平台激励有限而选择靠近厂商，向其他市面已有的无数自媒体靠拢，在厂商给的通稿框架与宣传的大方向已然确定的情况下，进行着极其容易趋同的表达，在首发宣发节奏与内容合作模式影响下，选择性的淡化掉体验中的痛点，即使在前文有提及也往往占据相对较小的篇幅，并随后立马引入一个「但是」，用着首发特调的机型说着好到梦幻的体验，最终在观众的不解中继续在内容理想与工作室运营中艰难的走着钢丝？当然，如果把话说到太绝，往往会有很多偏颇，所以让我们稍微说一些让文章作者更拟人的

「智能队列」如何改善你的家庭网络质量主作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org联合作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org2021/01/31 23:12昨天 12:57更新：2026 年世界电信和信息社会日重推在无线路由器的设置页面上，我们可能会看到一个名为「智能队列」的功能（根据品牌或操作系统的不同，也可能会被称做「智能 QoS」、「SQM」等）。这一功能是如何工作的？是否能让网速变快？是否需要打开？本文将尝试解答这些问题。从「网络如何连接」说起本章节主要介绍计算机网络的基础知识。如果读者已经熟悉相关知识，可跳过这部分内容。在介绍「智能队列」之前，先来了解一下基础的网络知识。两台电脑间进行数据传输，存在两种方式，分别是电路交换和分组交换。对于电路交换，两台电脑独占一条线路进行数据传输。过去的固定电话，大多使用的就是电路交换。由于独占一条线路，两台设备之间的通信较为稳定，很难出现数据丢失、数据传输顺序错乱等情况。但一条线路只能供一组电脑使用，另外一组电脑想要通信，就不能复用原来的线路了。对于分组交换，传输的数据被拆分为小段，每一小段叫做一个数据包（报文）。数据以报文的形式进行转发，从而实现了多组电脑复用一条线路进行通信。电路交换虽然能保证数据传输的稳定性，但由于独占线路浪费太多资源，所以在 Internet 中，使用的是分组交换。分组交换将数据拆分成一个一个的小片段来转发，没有独占线路。如果网络条件不好（例如同时有大量报文经过路由器，超出了路由器的处理能力），很有可能遇到数据顺序错乱、甚至数据丢失等情况。那么，这些情况应该如何避免？在微信视频通话、电视直播等应用中，报文的丢失或乱序，只会造成视频图像卡顿，不会造成太严重的影响。所以这些应用会使用简单的UDP协议，不考虑丢包、乱序。在网页浏览、文件下载等应用中，一般使用基于TCP的HTTP协议。TCP 在报文中携带序号，避免乱序；同时提供了报文的丢失重传机制，避免了丢包。所以，也可以认为 TCP 为两台电脑的应用程序建立了一个虚拟的「连接」。（最新版本的 HTTP/3 协议，使用的则是QUIC。QUIC 虽然基于 UDP，但引入了与 TCP 类似的报文丢失重传机制）如果想要了解更多网络基础知识，推荐阅读「入网指南」系列文章：入网指南 01 | 一文读懂你身边的「网络」入网指南 02 | 当你发消息的时候，发生了什么入网指南 03 | 交换机在交换什么？入网指南 04 | IP 地址大揭秘入网指南 05 | 除了接线，还有什么影响着我们网上冲浪入网指南 06 | 连上 Wi-Fi 就能上网了吗？网络延迟：影响上网体验的重要因素带宽和延迟，是影响上网体验的两大因素。带宽指的是在单位时间内，网络最多能传输多少数据。网络的带宽和水管的粗细有着很多相似之处，水管越粗，相同时间流出的水越多；同样地，网络带宽越大，相同时间能够传输的数据越多。在日常上网过程中，大的带宽能带来这些好处：上传、下载文件的速度更快在线看视频时，缓冲的时间更短如果家中有多个人同时上网，大的带宽能够保证每个人都能流畅上网，相互的影响更小对于家庭宽带，我们常常听到的 10 兆、50 兆、100 兆、千兆，就是指的带宽的大小。延迟指的是数据从网络的一端到达另一端，所需要的时间。过大的延迟也会对上网体验带来不好的影响，例如：在浏览器中输入网址敲回车后，需要等更长的时间，网页才能加载；视频聊天、语音通话质量变差，甚至能直接感受到声音、画面的滞后；玩在线游戏时，发现游戏变慢变卡。在很多时候，延迟相比带宽更容易被人感知，对网络体验的影响更大。是什么导致了网络延迟？上文中介绍过，Internet 使用的是分组交换。如果大家阅读过计算机网络相关的书籍，应该对分组交换的「存储-转发」机制有一定的印象。简单说，「存储-转发」是指网络设备将收到的报文存储在缓冲区，然后再发送。这是由于网络设备的性能有限，转发报文有速率限制。如果设备短时间内收到大量报文，一时半会儿处理不过来，就可以先报文存起来，空闲时慢慢发送。通过「存储-转发」，能够提高网络设备的吞吐量，更充分地利用带宽。其中，存储报文缓冲区，是一个先进先出的队列。那么，队列的长度应该如何确定？是不是队列越长，能够存储的报文越多，网络的性能就越好？大家可以先思考下这个问题。其实在过去，内存价格比较昂贵，考虑到成本，网络设备的队列不会太大。现在，随着内存价格的下降，网络设备拥有了更大的队列。但是，网络的带宽是有限的。实际的可用带宽，取决于数据传输路径上的最小带宽。这个最小带宽也被称做带宽上的「瓶颈」（bottleneck）。由于瓶颈的存在，如果流量长时间以较大速率发送，大的队列并不能显著提升吞吐量。相反，大量报文滞留在队列中，会带来延迟的增大。这种队列中缓存有大量报文，来不及发送，导致数据传输延迟增大的现象，被称做 bufferbloat（缓冲区膨胀）。DSLReport 提供的测速服务，能够检测网络的 bufferbloat 程度。不过由于没有中国大陆服务器，测试结果不够准确。国外的读者可以尝试一下：DSLReports Speed Test家用无线路由器中的 QoS 功能在「智能队列」出现之前，很多家用路由器已经提供了 QoS 功能。那么，传统的 QoS 如何解决网络延迟、网络拥堵？存在哪些不足之处？在了解这个问题之前，先来熟悉一下网络中的报文，是什么样子的：在网络中传输的报文，一般包括了链路层、网络层、传输层、应用层这四层信息。其中，网络层的源 IP 地址、目的 IP 地址、协议号，和传输层的源端口号、目的端口号，是 QoS 算法中常常会关注的五个信息。这些信息合称「五元组」。同一个应用在同一时间段内，传输的报文中，这些信息一般是不会发生变化的。例如观看在线视频时，同一个视频的所有报文 ，一般拥有相同的五元组信息。所以，通过五元组，我们能够区分不同应用的流量。特别地，一些常见的应用，会使用固定的端口号。例如网页浏览一般使用 80、443 端口；英雄联盟游戏，使用的是 5000-5500 端口 ……所以，当我们想要提升英雄联盟游戏的体验时，在路由器上设置让 5000-5500 端口的报文优先通过即可：How to set up traffic prioritization rules for League of Legends on Nighthawk Pro Gaming router | Answer | NETGEAR Support家用路由器中常见的 QoS 功能，就是通过手动设置 IP 地址、协议号、端口号等报文特征来实现的。这种方式需要用户了解一定的网络知识，且操作复杂。（虽然部分无线路由器内置了常用应用的 QoS 规则，但仍需要一定的手动设置，且无法及时适配新的、小众的应用）当然，除了上文中介绍的内容，QoS 还有多种不同的实现方式。例如：对特定应用（例如 BitTorrent）进行限速按照 MAC 地址指定某台设备的优先级（例如提高 Nintendo Switch 的优先级，使游戏更加流畅）通过WMM、DSCP等机制来决定报文优先级这些机制各有各的特色和不足。例如对应用程序限速的方式，虽然效果明显，但无法在网络空闲时，动态利用剩余带宽来提升传输速度，也难以适应带宽频繁变化的无线网络。关于上述 QoS 机制的技术细节，本文不再详细介绍。如果感兴趣，可以搜索相关关键词，进一步了解。智能队列：通过「全自动」的方式优化网络、降低延迟什么是「智能队列」目前，一部分新的无线路由器，开始有了名为「智能队列」的功能。这一功能一般指的是 fq_codel。fq_codel 包含两种队列管理机制，分别是 Flow Queue 和 CoDel。在下文中，我将简要介绍这两种机制的作用。Flow Queue前文中介绍过，网络中的报文一般都有自己的五元组信息；同一应用发送的报文，在一定的时间段内，五元组信息是相同的。这些信息相同的一系列报文，被称做一条流（flow）。Flow Queue 为每条流单独维护一个队列。例如，如果我们一边在线观看视频，一边进行语音通话，一边玩游戏，Flow Queue 就会为视频播放的流量、语音通话的流量、在线游戏的流量各维护一条队列：正常情况下，在线观看视频需要占用较大的带宽，但对延迟不敏感，用户可以容忍几秒甚至几十秒的缓冲时间。而语音通话和在线游戏占用带宽小，但对延迟敏感。例如语音通话时，如果一个人说了一句话，对方在几秒后才能听到，用户就能明显感知到延迟。如果三种应用共用一个队列，队列中会充满大量视频播放应用的报文，导致语音通话、在线游戏的报文迟迟无法发送，增大了这两种应用的延迟。Flow Queue 为视频播放、语音通话、在线游戏各维护一条队列，并公平地调度三个队列，避免视频播放的报文占用太多时间，从而让在线游戏、语音通话的报文得到及时处理。由于 Flow Queue 确保了应用程序之间的「公平」（fair），所以又被称做 Fair Queue。CoDel我们已经知道，适当长度的队列，能够充分利用带宽、提高吞吐量。但过长的队列，又会导致过大的延迟。也就是说，队列既不能过长、也不能过短。CoDel 所做的，就是让队列维持在合适长度，最大限度地提升网络性能。CoDel 通过计算报文在队列中的停留时间，来判断队列是否过长。具体做法如下：报文进入队列时，在报文上记录当前时间（也就是报文入队时间）报文出队列时，再次获取当前时间（报文出队时间）计算报文入队时间与出队时间的差值，即可得到报文在队列中停留的时间当报文在队列中的停留时间超过阈值，说明队列过长，需要缩小队列长度当队列过长时，CoDel 会直接丢弃出队列的报文，从而减少队列中报文的数量，达到控制队列长度的效果。（这种对报文的丢弃，并不会对网络带来负面影响，具体请见后续章节）可以看出，CoDel 根据队列中报文的停留时间，来决定是否丢弃报文。而不是「一刀切」地将队列限制在一个固定的、较小的长度。这样既能降低延迟，又保留了队列最大的优点：吸收短暂的大流量报文、提升吞吐量。另外，我们的设备，尤其是手机等移动设备，所处的网络环境是复杂多变的。例如手机靠近或远离无线路由器，就有可能导致网络带宽的变化。CoDel 通过计算报文停留时间的方式控制队列，能够更加灵活地适应网络带宽、网络质量的变化。fq_codelFlow Queue 确保了不同应用程序的流量能够公平传输。CoDel 控制队列长度，实现了整体延迟的降低。fq_codel 就是将 Flow Queue 和 CoDel 结合在一起使用的机制。其中，Flow Queue 先将不同的流划分为不同的队列，再为每一个队列使用 CoDel，控制报文的延迟。两者结合使用，共同提升网络体验。fq_codel 具有免配置、天然适应不同的网络环境、实现简单、消耗资源少等特色，越来越多的无线路由器（以及开源的路由器操作系统），开始内置这一机制。参考 KoolShare 上的这篇帖子，作者经过测试，发现打开 fq_codel 之后，在 BT 下载占满上行带宽时，在线游戏依旧保持较低的延迟：可能是家庭路由里最好的 QoS——在线游戏者的福音 - KoolShare（见 8 楼、10 楼、22 楼、25 楼的测试结果）是否需要打开「智能队列」？前面介绍了 fq_codel 的一系列优点。那么，fq_codel 有没有缺点？是否需要我们打开？我的设备 CPU 性能不强，是否可以打开「智能队列」？根据这篇帖子，ER-X路由器在打开 fq_codel 时，最大支持 200M 左右的带宽。QoS (fq_codel/cake) Hardware Recommendation : HomeNetworkingER

2025 年度旅行——4 天 3 晚青海湖小环线主作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥联合作者关注别惹小炸毛少数派作者一个学酥别惹小炸毛关注别惹小炸毛少数派作者一个学酥05/14 09:38Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。一次完美的旅行，对我而言，不仅仅是去被人走过千遍的地方打卡一圈，而是藏着三次截然不同的体验——出发前对目的地的想象与期待，旅途中与自然的亲密接触，以及归来后翻看照片时，重新发现那些当时未曾留意的细节，细细回味，查阅资料，让我对这次旅行有更深刻的理解。那么就让我们随着这些文字与照片，回到美丽的青海湖。契机：作为刚工作的年轻人，在国内的职场环境下，真正无理由可申请的假期只有 5 天，所以我和身边的小伙伴们都分外珍惜这五天的时光，每年的年度旅行我们都会避开人流，详细规划。由于十一假期要去参加两场婚礼，为了节约交通费，我们决定先开车到 2 号参加婚礼的宝鸡，把小 P（我的小鹏 MONA M03，具体可参阅我之前的文章）停在了宝鸡高铁站，考虑旅行时间较为紧凑，且西北自然景区交通补能条件可能较差，我们选择了到当地租车完成青海湖小环线行程。带着宝鸡擀面皮和肉夹馍，一行人正式开始了旅途，三个半小时后，到达了西宁火车站，初入高海拔地区，并未感受到明显不适，方便的是，租车平台在西宁站地下停车场就有取还车点，我们直接将行李放上车，省去了搬来搬去的劳顿。Day1 西宁-茶卡 全程约 457 km西宁-塔尔寺5A 级旅游景区，门票：70 元，游览时间约 2 小时塔尔寺是藏传佛教格鲁派的著名寺院之一，始建于明代，最初是为纪念藏传佛教格鲁派创始人宗喀巴大师而修建，据传宗喀巴大师出生于此地，后人于其诞生处建塔纪念，塔尔寺因此得名——「塔尔」即「有塔之地」。塔尔寺-八宝如意塔游览中心和寺庙门口都有大量的讲解人员，价格也不贵，但是本人信仰唯物主义，故只将这里当做名胜古迹参观，故没有请讲解，如果对藏传佛教及寺庙背后的故事比较感兴趣，请个讲解或许是更好的游览方式，或者是像我一样，回来翻阅照片时再仔细搜索相关资料。Tips：讲解会介绍各个佛像的「作用」（不知道这样叫是否合适，或者说负责保佑的事情），故可能会被架在那里为了讨好彩头而「必须」捐献资金，介意者可以像我一样自由游览。由于是节前，故参观人员并不很多，与前往朝拜的信徒或僧人可以说是一半一半，没有了熙熙攘攘的人群，也得以让我更加细致地观察这座寺庙和「磕长头」远道而来的人们。前来朝拜的僧人进入寺庙，浓郁的香火味和木质结构因缺乏通风的陈味扑面而来，看着宏大的佛像、墙上摆放整齐的诵经坐垫，的确能感受到不一样的文化氛围，这里安静、鲜活，尤其是看着不知道从哪里启程的「磕长头」的信众，虽因路途的遥远衣衫褴褛，但想必也经过这一路的艰辛而悟出了些许道理，可能出发时也不知道为什么，但走着走着就明白了吧。塔尔寺中的白塔募捐处，与时俱进地，也支持扫码支付塔尔寺有大大小小的寺庙很多间，在我看来有的都别无二致，于是重点转了几个后，就启程回西宁吃饭，整备氧气瓶，应急食物等装备，正式奔赴青海湖。海南藏族自治州共和县 南山观景台 & 青海湖二郎剑景区前者免费，后者为 5A 级旅游景区 门票 72 元，游览时间约 2 小时一点半在西宁市吃完炕锅后，我们开始驱车前往青海湖，约 2 个小时车程，我们到达了本地的主要目的地青海湖的南侧景区 青海湖二郎剑，在进入二郎剑景区前，可以导航到「南山观景台」 从山上远眺青海湖和远处群山的全貌，但是山路较窄且有工程车辆时不时通过，故要格外注意安全。景区本身没有人文景观，但其还是我国早期鱼雷的试验基地，但是因为我们到达时间比较晚，并没有参观博物馆，而是直奔到湖边。青海湖是我国水面面积最大的淡水湖，水面平静辽阔，与远处的群山相辅相成，甚是开阔美丽，栏杆上有硕大干净的鸽子驻足，在几番伦理拍照后，我靠在栏杆旁，看着远处的夕阳洒在地面，耳边传来阵阵鸟鸣与风声，工作与生活的喧嚣远离脑后，只剩下全身心的与自然接触。云层、远山、草原，形成了震撼的三明治拨云见日，青海湖边的草还是绿色的岸边排队等待「薯条」投喂的海鸥码头景观随着夜幕降临，我们离开二郎剑景区，沿着青海湖开车前往茶卡住宿，虽然只有 200 多公里的行程，但是因为沿途工程车辆很多，还有很多地方持续修路，有的路段有连续的 U 型弯，还有双向仅有一条车道通行的地方，还要翻越几个海拔 3800 米的垭口，成了这一路最惊险刺激的行程，一路上小心翼翼，战战兢兢，终于在晚上 10 点到达了茶卡镇，幸运的是镇上的餐馆都还在营业，找了家当地美食，简单吃饭后休息。夜幕降临的倒湖茶公路Day2 茶卡-祁连县 全程约 372km茶卡盐湖4A 景区 带往返小火车套票 167 元「茶卡」在藏语中意为「盐池」，该地以出产「大青盐」而闻名，参观盐湖也是此地的知名旅游项目。茶卡盐湖的景区共有两个，天空之镜是较老的景区，主要以自然风光为主，天空一号主要以商业化项目为主，作为外地游客，我们肯定更希望看到原始的自然景观，于是选择前往天空之境景区。景区很大，如果想要游览每个景点建议留出一整天的时间，游览顺序建议先坐小火车到终点天空之镜，该站处于盐湖的最深处，盐水异常的干净，也最为出片，景区也放置了很多拍照打卡点位，可以尽情拍摄，心痛没有带无人机。在晴天的时候，由于天空和地面都是白色，很容易晃眼，推荐大家一定要戴墨镜前往，通过墨镜观察湖面，盐湖会看起来像翡翠一样更加绿色透明。套票包含鞋套租赁，可以套上下湖走走，踩在大颗的盐里,很像沙滩的感觉，景区维护的十分干净，可以尽情的亲近自然。有趣的是，景区为了怕大家在盐滩大片的白色中无聊，设置了很多拍照用的摆件，例如国家地理的相框等等，还在大屏幕上贴心的轮播拍照指导，期望每个游客都能满意而归。盐湖与远处的山峰盐湖码头废弃的采盐轨道之前采盐的轨道深入盐湖，在墨镜下，真实的观感颜色会比照片深很多。接驳小火车行驶在盐湖中间，干净得不像国内巨幅盐雕从盐湖出来，我们在县城稍作休息，开始启程前往祁连县，有了昨夜惊险的行程后，尽管沿途的国道可能风景更好，但是我们选择了高速较多的线路。这一程的路况还是很好的，就算是国道车辆也较少，道路笔直，视野十分宽阔，草原、远山、白雪、天空层层分明。在行驶过程中，不时的有大群的羊羔和牦牛穿过公路，不必惊慌，一会就有牧民骑着摩托驱赶，让它们加快穿越的脚步。公路与群山草原上大批的牛羊傍晚的草原，远处能看到夕阳洒在雪山上落日下的牦牛剪影在夜色中，我们抵达了祁连县，在县城中吃过一顿特色烧烤后，进入梦乡，准备迎来第三天的行程。Day3 祁连县-岗什卡雪山-西宁 全程约 320km祁连卓尔山景区门票 60 元 上山小巴 20 元 登山电梯 10 元这个景区的确是全程我最惊艳的景区了，近能看到瑞士小镇似的田园风光，中能看到美如画的丹霞地貌，远还能远眺远处的阿咪东索雪山（海拔 4695 米），而且在山顶可以拍跟雪山的合照，大气、干净、巍峨。自驾车辆可以停在景区停车场，然后乘坐接驳中巴约 10 分钟到达检票口。卓尔山的两侧，风景差别迥异，一侧是连绵的雪山，另一侧是仿佛瑞士小镇的丘陵和小镇。阿咪东索雪山卓尔山的另一侧，仿佛欧洲小镇丹霞地貌与草地交融最让我惊讶的是这个景区竟然还修建了方便爬山的观光电梯，10 元一位，大概有三段，果然是人性化啊。景区电梯中午约 12 点，我们离开景区，在祁连县的一家川菜馆简单用餐后，我们驱车开始前往本次旅途的最高海拔岗什卡雪山。在县城开车的过程中，等红灯时，我看到一位初中年龄的藏族小伙盯着我们开的岚图梦想家，眼中闪闪发光，不禁让我想起小时候的自己，看到喜欢或没见过的车也会驻足观看，充满憧憬。愿他长大后也能开上自己喜欢的车，能够做自己喜欢的事情。岗什卡雪峰景区免费 上山摆渡车 25 元岗什卡雪山坐落于青海海北州门源县境内，是祁连山脉东段的最高峰，被誉为「雪山之尊」，因为设施和难度都尚可，也被雪山爱好者们称为「人生第一座雪山」。主峰海拔 5254.5 米。游览模式同样为将车辆停在景区停车场，然后乘坐接驳车到雪山脚下，正式游览区域的海拔在 3800 米左右，建议随身携带一个氧气瓶，下车之后，可以看到湛蓝的连心湖，水来自于冰川融水，干净，清冽，与子洲植被的荒凉的深褐色形成鲜明的对比，十分美丽。连心湖到达时天气并不是很好，岗什卡雪峰被薄雾笼罩天然泉水，因还有行程，害怕拉肚子，没能如愿品尝一下离开景区后，雪山逐渐露出真面目，攀登还是留给下次吧沿着冰川的遗迹向上走，就能看到岗什卡雪峰徒步攀登的起点，景区也提供马匹，可以跟着向导继续向前，到七彩瀑布，因时间和个人体能的考量，我们浅尝辄止，在河道附近游荡些许时间后，便下山返程。临近傍晚，启程返回西宁市，一顿丰盛的海底捞过后，快乐的青海湖小环线结束。Day4 西宁市-青海藏文化博物馆-特产采购最后一天，伴随着蒙蒙细雨，我们打车前往青海藏文化博物馆，博物馆不大，纵使是十一期间人也不是很多，但内容着实丰富，满墙的唐卡和在电影里看到的藏传佛教僧人服饰，还有精致的贵族金器。在博物馆的文创区，我购入了一个小的转经筒，在回程的路上一直把玩，虽不信奉佛教，但也觉得甚是有趣。唐卡四王金杯嵌绿松石立凤金头饰博物馆文创-转经筒，38 元虽不算划算，但做工十分精致离开之前，为了给家人和朋友带些纪念品，我们选了一家超市类型的纪念品商店，购买了牛肉干，藏红花，枸杞，青稞酒等特产，虽说也不一定划算，但胜在明码标价，将行李箱装满后，顺利返程。后记作为 2025 年的年度旅行，本次 4 天的青海湖小环线我甚是满意。人少不堵车，行程不特种兵，在当地感受美食和壮美的自然风光，一路上与好友们分担开车，聊对景色的感受，互相拍照，点评食物，虽偶有几个工作电话烦扰，但总体悠然自在，放空身心，一切都是那么的安静，与自然深度的接触，与北京的喧嚣形成了鲜明的对比。对于出行方式，自驾无疑是最有的选择，这边深入自然，道路虽崎岖，但也属于铺装路面，且西宁市神州、一嗨等在高铁站都有网点，取送车方便，且自驾可以自由控制行程，根据游览时长方便自己调整。对于住宿，我们对酒店的干净程度略有要求，于是全程都选择了以华住会为目标的连锁快捷酒店，以下是所有天的酒店名称和十一节前的价格，个人觉得是十分舒适的。Day0 桔子西宁火车站省人民医院酒店 双床房 331元Day1 星程茶卡盐湖酒店 覆氧双床房 314 元Day2 宜必思尚品海北祁连卓尔山酒店 421 元Day3 桔子西宁火车站省人民医院酒店 双床房 412 元关于本次行程的花销，纵使算上从北京出发的路费，人均也仅 3k 左右，可以说是十分划算了，主要的开销就是租车和酒店。支出一览表关于卫生间，景区和路途服务器的卫生间都比较完善，因同行有女生，我们还准备了移动卫生间，但从我们的行程安排走下来，并没有用到，个人觉得要是按我们的行程游览也无需担心。下一站，出发！2025 年的年度旅行结束，2026 年也可以开始规划起来了，新疆还是海南，希望能远离喧嚣，享受美丽壮阔的祖国风光。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀75扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@北鸮© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#旅行#Matrix首页推荐7等 7 人为本文章充电扫码分享举报本文章别惹小炸毛一个学酥关注

打破虚拟与现实的边界，这 6 款网页解谜游戏值得一试主作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.联合作者关注Tp少数派作者Be nothing, and everything.Tp关注Tp少数派作者Be nothing, and everything.15:00ARG 即侵入现实游戏（Alternate Reality Game），是一种基于现实世界，融合各种多媒体（如网站、电话、社交媒体等）技术的互动式解谜游戏。由于其与现实的高度结合所带来的沉浸感，一直以来被许多玩家所追捧。如果你在 2016 年玩过《守望先锋》，相信你对当时新英雄「黑影」的登场一定记忆犹新。从 2016 年夏季开始，一直到 11 月的暴雪嘉年华，暴雪通过在宣传视频、官网、社交媒体发布了一些隐藏着密码的内容，由全球玩家共同解谜，不但揭示了游戏世界观下的一些故事，更是顺理成章地将身为黑客的「黑影」带到大家的视野里。经过几个月的发酵，在暴雪嘉年华上大屏幕被「黑影」控制，随后开始播放英雄动画的时候，更是满足了玩家们期待已久的心。「黑影」的宣发是一场相当成功的 ARG，时至今日仍有玩家对其念念不忘。我本人也是 ARG 的爱好者，我非常享受这种打破虚拟与现实边界的游戏乐趣。前段时间，一款名为《大学生登山失踪事件》的网页解谜游戏突然火了起来。比起此前一些高度依赖技术知识的网页解谜游戏，《大学生登山失踪事件》在不需要任何 IT 知识背景的情况下，依然有着抽丝剥茧逐渐接近真相的推理爽快感。自此之后，小红书上越来越多的人开始制作这一类网页解谜游戏，并为它们赋予了 ARG 的名号。这些游戏流程往往不长，差不多 1-3 小时就能通关。因为是网页解谜游戏，只需要一个浏览器就能游玩，很适合上班时间摸鱼，游玩门槛非常低。这篇文章将会推荐一些这段时间我玩过的质量不错的游戏，虽然一些制作精妙的游戏确实带有 ARG 的影子，但其本质上仍是基于网页的解谜游戏。为了方便大家后续搜索，本文会将这些游戏统称为 ARG。大学生登山失踪事件🔗https://missing.shiroki-y.top/introduction.html在游戏中，玩家需要通过从弟弟衣物中发现的「福安菜家常馆」的线索为开端，逐渐深入一场黑色阴谋找到弟弟失踪的真相。游戏整体难度不高，但在游玩过程中最好能用纸笔或软件记录进度，这能够帮你清晰地理出游戏脉络。游戏的核心玩法其实非常简单，只需要阅读页面，找到页面中可疑的关键词，在页面搜索栏里进行搜索即可推进剧情，这其中也会有一些用户名、密码、口令、字谜等需要玩家进行推理和猜测。网页页脚处会标记当前页面进度，既可以方便定位，也能够明确线索的先后顺序。剧情方面，《大学生登山失踪事件》并不会在一上来就给玩家塞一些晦涩难懂的概念，而是从大家熟悉的普通网页开始。当玩家一步步抽丝剥茧逐渐发现事件背后的黑暗面时，令人不由得感叹于游戏的设计与巧思。同时，游戏内也藏有一些可供探寻的 ex 页面作为彩蛋，在解谜之外找到各种彩蛋也不失为一种乐趣。诡异广告调查事件🔗https://mp.weixin.qq.com/s/yeNYMZVRp6N-MESFJGk9rg这是一款与《大学生登山失踪事件》来自同一作者@銀河電燈_的游戏，世界观也与前作有关联之处，很是有趣。玩家将扮演一位被「优化」的人，因为在看剧时发现了某个已停播的节目广告，从而开始调查这个诡异的广告，并发现藏在节目与自己背后的神秘真相。与前作相似，这款游戏也将基于阅读与搜索推进剧情。在此基础上，它还加入了一些 ARG 元素，比如在街景地图中，找到现实里存在的文字招牌作为线索。并且在网页设计上也比前作更加成熟，贴合剧情上的时间设定，在网页上加入了许多彩蛋。剧情方面，本作有一定的惊悚元素。特别是在游戏过程中展示的一些内容可能会引起不适感，如果你对这类内容比较敏感，请务必提前做好心理准备。当然，游戏并没有 Jump Scare，更多是通过剧情与视觉设计所营造出一种的诡异氛围。三川米粉店🔗https://www.berrysignalscmfd.top/《三川米粉店》也是一款从餐饮店主页出发，探寻故事真相的游戏。玩家需要调查三川米粉店突然停业的原因，在调查过程中，玩家将逐步揭开这家米粉店背后的故事。本作的页面设计相对没有那么精致，但也融入了一些 ARG 的元素，部分谜题的答案并不会直接给出，而是需要玩家通过搜索引擎或其他工具进行查询才能获得答案。从剧情上来说，《三川米粉店》的设定会更加接地气一点，并且结局也会相对更温馨一点。比较适合对于惊悚元素不太感冒的玩家作为第一部作品来尝试游玩。永安温泉度假酒店🔗https://mp.weixin.qq.com/s/kKsnzFsLFwwm2nwv8DseAA与前面侧重解谜的作品不同，《永安温泉度假酒店》的解谜难度其实不高，反而更像是一种「规则怪谈」式的探索路径。玩家需要不断探索找到更高权限的员工账号密码，从而揭开温泉度假酒店背后的秘密。游戏在开始前需要玩家填写「报名信息」，在后续游戏中能够增强代入感。从游戏性的角度来看，这款作品的「解谜」爽快感或许没有那么强。但如果你对「规则怪谈」或者一些高概念的设定感兴趣，那么这款游戏依然是一个值得尝试的作品。404 工位消失事件🔗https://xsn1204.github.io/qmgame《404 工位消失事件》是一个特色鲜明的游戏，玩家将扮演一个实习生，接手其他人留下来的工作电脑，并从中发现一些未被处理的秘密。游戏在网页上复刻了一个形似 macOS 的「系统」，很大程度上增强了代入感。因为设定是在电脑系统，游戏中需要玩家在钉钉、微信等各种软件中挖掘线索，反而更有一种作为侦探在不断探索的感觉。但受限于网站形式和游戏体量，游戏中能够互动的地方通常都是为剧情线索服务的，与线索无关的地方大多都是不能点击互动的。而且游戏不能通过点击窗口左上角的关闭按钮回到上一层，而是需要通过浏览器的返回来实现。但这些小细节对于《404 工位消失事件》来说都算瑕不掩瑜，整体游玩体验还是值得肯定的。作者甚至还出了 2.0 版本，新的剧本进行了优化提升，如果对这一类型感兴趣的话也可以尝试游玩。找到我🔗https://www.kaixinmoyu.com/feijo/findfeijo/在游戏中，玩家需要扮演一位私家侦探，调查游戏作者失踪的原因。游戏在交互方面做得非常有意思，游戏不仅做了浏览器页面对线索进行搜集检索，还做了小灵通和笔记本，前者可以通过拨打特定的电话获取线索，后者则需要玩家阅读日记找到隐藏的内容。《找到我》本身是一个更重交互而非推理的游戏，游戏中融入了一些 ARG 元素，在最开始就需要到某篇公众号文章中找到一个密码才能继续推进。但游戏的剧本逻辑会稍微有些不够连贯，以至于中间可能会有一些容易卡关的地方，需要仔细翻阅此前出现的线索才能继续往下推进。另外，游戏中后期会有一道需要在现实中解算的谜题，这道题本身与游戏本身的节奏设计会稍显割裂。但整体来说，《找到我》依然是值得一玩的游戏，尤其是本身的交互设计和 ARG 元素的融合做得非常不错。在游戏作者的个人主页，也能找到作者开发的更多网页解谜游戏，如果你对这款游戏的风格感兴趣，也可以尝试其他作品。小结优秀的网页解谜游戏并不止文章提到的这些，光是在小红书上一搜，就能找到许多分享帖。希望本文能起到抛砖引玉的作用，欢迎大家在评论区分享自己玩过的，令你印象深刻的网页解谜游戏。> 简单、好用、专注的写作软件，少数派为你呈现 🚀> 守护孩子的天马行空，从零开始美术教育启蒙🎨40扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Lotta© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#解谜游戏#游戏#生活方式4等 4 人为本文章充电扫码分享举报本文章TpBe nothing, and everything.关注

5月18日，国新办举行新闻发布会，介绍加力优化离境退税措施扩大入境消费有关情况。会上，国家税务总局货物和劳务税司司长谢文表示，2025年4月以来，离境退税办理人数同比增长了367%，退税商品销售额和退税额同比增长90%。谢文指出，截止目前，全国离境退税商店增至1.4万家，为去年同期的3.2倍。（证券时报）

36氪获悉，近日，海光信息新一代终端CPU在液氮极限散热环境下完成性能验证测试，单核主频达到5.15GHz，刷新国产终端CPU主频纪录，意味着国产CPU处理器首次正式迈入5GHz时代。测试数据显示，海光新一代CPU拥有约30%的极限性能释放空间，为后续风冷、水冷常规散热方案的迭代优化，积累了极具价值的工程实测数据。

36氪获悉，近日，国内全倾转eVTOL企业“华羽先翔”完成超亿元Pre-A+轮融资，本轮融资由蚂蚁集团领投、彼岸时代及复星创富跟投。融资资金将重点用于加速鸿鹄系列eVTOL的适航取证，以及低空出行及商业场景的落地与创新。

36氪获悉，天微电子发布《关于产品价格上调的说明函》。其中提到，因近期原材料成本持续攀升，为维持产品品质及正常供货交付，自2026年5月20日起，公司全系列产品价格统一上调，涨幅为10%-30%，具体型号、调价幅度以公司报价为准。

36氪获悉，5月18日，BranDNA（和鸿广）品牌管理公司宣布，已与意大利时尚品牌GCDS达成战略合作。BranDNA将作为GCDS亚洲战略合作伙伴，大中华区及日本市场独家战略合作伙伴，负责品牌授权管理与分销拓展，助力品牌实施全球发展战略。

6.4k Stars！用Claude Code写论文的全套流水线，有人打包开源了量子位·2026年05月17日 12:594个skill，跑通整套科研流程听雨 发自 凹非寺量子位 | 公众号 QbitAI用Claude Code写论文的一整套流水线，有人打包开源出来了。完全戳中了学生党的痛点，github星标直达6.4k。academic-research-skills项目名叫academic-research-skills（以下简称ARS），是一套Claude Code技能包。里面涵盖4个skill，分别对应论文的研究、写作、审稿、定稿。只需两行命令安装，直接一条龙串起整套学术研究流水线。academic-research-skills只能说，我读研的时候怎么没碰到这种好东西呢…示意图4个skill，跑通整套科研流程ARS的核心架构由4个skill组成，它们各司其职，拼在一起就是一条从选题到交稿的完整链路。我这里还做了图，大家可以看得比较直观：△Deep Research是一支13个Agent的研究团队。它负责文献调研、研究问题构建、方法论设计，还能写系统性的PRISMA综述。团队里有专门做文献溯源的Agent，会调用Semantic Scholar API验证每一篇引用的真实性。有苏格拉底导师Agent，通过对话引导研究者理清思路。还有魔鬼代言人Agent，专门挑刺，防止研究者在早期就陷入思维定式。△Academic Paper是一支12个Agent的写作团队。从大纲设计、论证构建、草稿撰写，到双语摘要生成、图表可视化、引用格式转换，全流程覆盖。特别值得一提的是风格校准功能，AI会学习你过往作品的写作风格，让输出更像你自己写的，而不是千篇一律的AI味。输出格式支持Markdown、DOCX、LaTeX，最终可以编译成APA 7.0或IEEE格式的PDF。△Academic Paper Reviewer是一支7个Agent的审稿团队。模拟真实学术期刊的评审流程，由主编EIC带领三位领域审稿人，再加上一个魔鬼代言人，从方法论、学科视角、跨学科价值等多个维度打分。评分采用0到100的量化标准，80分以上接受，65到79小修，50到64大修，50以下拒稿。审稿团队还会输出详细的修改路线图，告诉作者下一步该做什么。△Academic Pipeline是流程编排器，把前面三个团队串联成一条10阶段的流水线。从研究、写作、完整性检查、同行评审、修订、最终检查，到发表准备和流程总结，每个阶段都有明确的产物和检查点。你可以在任意阶段插入，比如已经有了初稿，就从Stage 2.5的完整性检查开始；收到了审稿意见，直接从Stage 4的修订切入。费用参考也很透明，一篇1.5万字的论文，全程跑下来大约4到6美元。△比较有意思的设计用Claude Code做学术研究的开源项目已经很多了，但是深扒之后，我发现ARS在底层设计上还是有些过人之处。可以简单总结为一句话：系统性防止AI搞砸学术研究。第一，引用核验。AI写论文最忌讳的，就是幻觉引用。不只是编造不存在的文章，还包括标题相似但作者年份全错、DOI真实但内容对不上等更隐蔽的情况。ARS在Deep Research阶段就埋了一个引用核验机制，每一篇文献都要过Semantic Scholar API的存在性确认。不是简单查一下标题对不对，而是用Levenshtein相似度算法做模糊匹配，阈值设在0.70以上才算通过。△第二，完整性闸门。在流水线的Stage 2.5和Stage 4.5，有两道不可跳过的完整性闸门，会运行一份7项AI失败模式检查清单。这份清单直接来自2026年Nature上发表的一项全自主AI科研研究，其中总结了7种翻车模式，覆盖引用幻觉、数据捏造、方法论造假等情形。7种翻车模式任何在2.5被标记为SUSPECTED的问题，必须在4.5变成CLEAR，或者由人工手动覆盖并留下记录。设计逻辑是：把「我相信AI不会出错」变成「我要求AI证明它没出错」。实测中，这套机制在一篇真实论文里抓到了15个伪造引用和3个统计错误。第三，反谄媚协议，让AI敢于说不。大多数AI工具都有一个隐形毛病，讨好用户。你让它改，它就改，哪怕改得更差。所以ARS在审稿环节专门设计了反谄媚机制。审稿团队里有一个Devil’s Advocate，也就是魔鬼代言人，职责是挑刺。但挑完刺之后，还有一个让步阈值协议。DA的反驳会被评分1到5，如果低于4分，写作团队不允许承认。△换句话说，AI不能为了显得好合作就轻易让步。同时，攻击强度在修订过程中必须保持。如果第一轮审稿把方法论批得体无完肤，作者修订后不能让审稿人突然变得温柔。评分轨迹也会被追踪，任何维度的分数下降都会被标记为回归。这和软件工程里的不引入新Bug原则一样，改一个地方不能搞砸另一个地方。第四，三层数据隔离，不让AI偷看答案。ARS把数据流严格分成三层：Layer 1是原始输入，默认不可信，可能幻觉、过时、带偏见。Layer 2是通过完整性验证后的产物。Layer 3是评分标准、参考答案和金标数据，这层材料永远不能出现在写作AI的上下文中。具体实现上，写作团队和审稿团队分两次独立调用，中间有阶段边界隔离。写作AI只能收到审稿AI的自然语言反馈，比如「第二章论证跳跃，建议补充对比实验」。但它看不到原始的评分标准，也不知道每个维度占多少分。这个设计的灵感来自于Anthropic今年的w2s-researcher研究，其中也用了同样的三层隔离模型。结论是当AI能读取标签数据时，结果可能不是真的泛化，而是在优化表面特征。解决方案不是更好的提示词，而是结构上的隔离。△最后一点，诚实文档化，「我不保证能复现」。学术界经常遇到「这个结果我复现不了」的问题。ARS给每个产物生成一个repro_lock文件，记录运行时的完整配置。但文件里有一段强制声明，LLM输出不是字节级可复现的，模型提供商会更新权重而不改模型ID，外部API每天返回不同的数据。这个文件只是配置文档，不是重放保证。△在更新日志上，可以看到ARS已经经历了很多轮迭代。从2月上线到现在，提交的commit数达到了三百多次。从每次版本更迭中，也能看出作者对AI学术研究系统风险有着深刻理解。这也是我觉得目前学术研究AI工具的关键所在——让AI帮你写论文并不难，重点是如何防止它出错、讨好，让整个流程变得更系统更可靠。ARS的设计哲学，可以总结为README里那句话：「AI是你的副驾驶，不是飞行员。」如何安装安装方式很简单，如果你已经在用Claude Code，只需要两行命令：/plugin marketplace add Imbad0202/academic-research-skills/plugin install academic-research-skills验证安装是否成功，运行：/ars-plan然后描述你正在写的论文主题，ARS就会启动苏格拉底对话，帮你梳理论文结构。如果你偏好单条命令测试，也可以用：/ars-lit-review “你的研究主题”不过最简单的安装办法，其实是直接把SKILL.md上传到claude.ai项目知识库。不需要安装Claude Code，打开浏览器就能用。不过要注意，这种方式不支持多Agent并行，功能上是单Agent版本，适合轻度体验；想跑完整流水线还是需要Claude Code。还有一点，项目支持繁体中文和英文。那么，又到了大家最关心的，要花多少钱的环节。作者推荐使用Claude Opus 4.7搭配Max订阅计划。完整跑完10个阶段，单次可消耗超过20万输入token和10万输出token，单独使用某个子模块则少得多。Max订阅计划分两档，每月100刀或200刀，相当不便宜。但如果你的科研经费可以报销的话，那…示意图本文来自微信公众号“量子位”，作者：关注前沿科技，36氪经授权发布。本文由「量子位」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、AI生成

3个人带100个AI程序员，一个月烧掉130万美元！OpenAI：钱我出新智元·2026年05月17日 14:08如果token成本不是问题，软件开发会变成什么样。Peter Steinberger编辑：所罗门【新智元导读】3个人，100个AI agent，一个月烧掉130万美元——OpenClaw之父把软件开发变成了「AI流水线」，OpenAI替他买单。别人晒工资条，他晒账单——一个月130万美元！也就是将近900万人民币/每月。直接把网友们看傻了。OpenClaw之父Peter Steinberger在X上轻描淡写地甩出一张截图。Peter Steinberger但截图上的数字，属实不淡定：30天花费：$1,305,088.81。吞掉6030亿个token。发起760万次请求。你没看错，是130万美元。也不是哪家大厂的季度AI预算——一个3个人的小队，一个月的用量。更炸裂的是：这钱，OpenAI给报销。评论区瞬间就疯了。有人惊叹，有人质疑，有人掏出计算器开始算「这到底顶多少个程序员」。Steinberger本人谈定回应：「关闭快速模式后，我的花费低于一个工程师的成本，而且这确实帮助大得多。」翻译下就是——属实划算！更有网友惊叹于月薪40万刀的工程师——「旧金山的就业市场过于疯狂。」网友评论网友评论还有网友好奇，这个大量的Token花在了哪里。Peter回应大部分都用在了OpenClaw的开发上。网友评论云端程序员军团这件事最离谱的是，Pete的小团队只有3个人。他们在云端长期跑着约100个 Codex 实例，干软件工程里最脏、最累、最容易把人逼疯的活——审PR、找安全漏洞、去重issue、改bug、监控benchmark、发现回归后发到Discord，甚至听完会议后直接开PR。于是，AI不只是「帮你写代码」，而是进入软件协作的每一个缝隙。这就很吓人了。因为软件开发真正贵的是沟通、理解、上下文切换、审查、回归、修复、等待、重复劳动。过去一个团队每天大量时间，消耗在这些看起来不像「创造」、但没有它们项目就会烂掉的事情上。现在，Peter 把这些环节一口气丢给了一群AI agent。这是AI开始替你维护一个组织的神经系统。示意图这张截图还有一个细节很重要：它不是OpenAI后台，而是 Peter 做的 CodexBar。CodexBar 是一个 macOS 菜单栏工具，用来追踪各种AI编程工具的使用窗口、credit、成本和重置时间。它支持 Codex、Claude、Cursor、Gemini、Copilot 等一堆服务。以前程序员的菜单栏里是什么？CPU、内存、电量、网速。现在多了一个东西：token。token正在变成一种新的「生产资料」。CodexBar最后说两句130万美元一个月，3个人，100个AI agent。你细品这组数字——三个活人，带着一百个不吃饭不睡觉不提涨薪的数字员工，把一整个工程团队的活儿给干了。有人看完热血沸腾：AI终于不是只会聊天的花瓶了！也有人看完脊背发凉：等等，那我们这些写代码的，以后干嘛去？但说实话，最让我睡不着的，是Steinberger那句轻飘飘的话：「我在探索，如果token成本不是问题，软件开发会变成什么样。」Peter Steinberger各位，他说的是「如果」。可问题是，这个「如果」正在以肉眼可见的速度变成「当」。今天130万美元干的活儿，模型降一轮价，13万。再降一轮，1.3万。到那一天，100个AI agent同时干活，不再是硅谷大佬的独家游戏，而是随便一个三人创业团队的基本操作。车库里三个年轻人，手握一百个不知疲倦的AI程序员——这画面，想想就离谱。Peter Steinberger掀开了底牌。底牌上写着：未来已经在敲门了，而且它不打算等你准备好。参考资料：https://the-decoder.com/for-1-3-million-a-month-openclaw-founder-peter-steinberger-runs-100-ai-agents-that-code-review-prs-and-find-bugs/https://x.com/steipete/status/2055346265869721905https://developers.openai.com/codex/speed本文来自微信公众号“新智元”，作者：新智元，36氪经授权发布。本文由「新智元」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、采访供图

又诞生一位千亿女富豪36氪的朋友们·2026年05月17日 15:45又一部“打工妹”逆袭成为女富豪的爽文剧本。示意图自4月初以来，立讯精密股价上涨超过55%。5月11日，立讯精密市值创下历史新高，达到了5542亿人民币。立讯精密掌舵者王来春持股18.75%，身价也突破了千亿人民币，成为国内白手起家的又一位千亿女富豪。如果说4月以来的涨幅，是因为立讯精密“站在光里”，短期股价大涨或许与一则收购公告有关。5月10日，公告显示立讯精密拟通过收购京西智行集团全部股权，间接拿下京西国际约59.50%股份，有望成为其最终控股股东。这是立讯精密继两年前收购德国莱尼后，在汽车领域的又一关键资本动作。郭台铭门 徒立讯精密的故事通常被描述为“打工妹”逆袭成为女富豪的爽文剧本。实际上，只单纯地给创始人王来春贴上“打工妹”的标签，就像是只把立讯精密当做一家代工厂。与现在创业者的背景不同，王来春既不是天才少女，也不是海归精英。她出生于20世纪60年代末的潮汕地区，当时物质和思想都极度匮乏，农家女出身的她，更是过早地尝到了生存的艰辛。1988年，改革开放的风头正盛，21岁的王来春选择到最热闹的深圳求职，此时正赶上郭台铭来内地创业，于是，机缘巧合下，她成为了富士康首批150名员工之一。早期的富士康被外界称为“血汗工厂”，流水线两班倒，每天站12个小时是家常便饭，累了一天，洗澡还得等。很多人受不了这份苦，干了几个月就走了，但王来春却坚持留了下来，这一留就是十年。这十年间，王来春不仅成为了大陆员工里职位最高的女主管，还像海绵一样吸收了消费电子行业的大量经验，成长为了郭台铭最出色的学生。一个细节是，她在立讯几乎复制了富士康模式，工厂管理军事化，连车间里都挂着郭台铭的管理语录。对于“徒弟”的成功，郭台铭曾公开评价：“富士康在中国大陆创造了很多亿万富翁，王来春是最具代表性的一位！”1999年，受金融危机影响，香港很多公司都破产倒闭了。王来春看准机会，毅然辞去富士康高管职位，和哥哥王来胜共同收购了香港立讯公司。面对这一冒险举动，她的家人是反对的，其母亲甚至把家里的现金和存折都藏了起来。然而，这并非是王来春的冲动之举，在立讯早期，王来春选择依附富士康生存，“富士康开到哪，立讯就跟到哪”，专门承接其一时无法处理的订单：低毛利的，高难度的。富士康给什么，她就做什么，在定价方面也完全由富士康说了算。所以成立后的很长一段时间内，富士康始终是立讯的第一大客户，这种关系也让富士康越来越离不开立讯，因此，郭台铭兄弟还入股了立讯。2004年，王来春成立立讯精密工业股份有限公司，开始独立开发自有客户并自主研发产品，逐步向高端产品线迈进。尽管如此，到2010年，立讯精密上市时，富士康依然为立讯精密贡献了一半销售额。库克时代如果说2010年之前的立讯精密，还是一个标准的“小富士康”，还是在王来春的舒适区里创业，那么2010年之后，立讯精密的发展，则完全依赖于王来春的智慧与资本布局。因此，业内流传着一个说法—立讯精密真正的创业之路是从2011年开始的。这一年，已经上市的立讯精密有了充裕的现金，此时的王来春不再满足于代工，把赚来的钱全部砸进了更难的技术领域，一边寻求转型，一边开始“去富士康化”。也是在这一年，代工领域一个时代性的机会出现了。2011年，库克执掌苹果并推行供应链多元化，王来春意识到，新帅上任，势必调整布局、扶植自己的心腹伙伴。若能在此时跻身“果链”并获库克青睐，立讯便有望摆脱依赖外部溢出订单的处境。于是，她开启了一系列激进的“资本换技术”的并购。并购的第一枪打在了昆山联滔电子身上，立讯精密斥资11.8亿元收购了其60%的股权，此举规避了苹果复杂的考核机制，使得立讯精密顺利进入了苹果供应链。这起并购案只是一个开始，2016年，它又收购了苏州美特51%股权，靠着苏州美特在声学领域的多年积累进入到了声学领域，并在下一年获得了苹果AirPods的生产机会。据悉，当时苹果的AirPods主要由台湾英业达生产，但英业达生产效率很低，经常出现供货延迟的情况，立讯精密拿到订单后攻克了AirPods生产效率低的问题，最终成为AirPods全球最大的供应商。2020年7月，立讯精密又收购了江苏纬创、昆山纬新两家公司，正式切入iPhone代工领域，成为了中国大陆首家、全球第三家iPhone整机制造商，打破富士康、和硕的垄断格局，并于2021年首次获得iPhone 13部分订单，实现了从零部件供应商向整机组装商的关键转型。2021年，立讯精密已发展成为营收超1500亿元，苹果业务占比达七成的全球消费电子巨头。可此时，王来春却推出了三个五年计划，全方位发展消费电子、汽车、通信、工业、医疗等业务，并在汽车领域制定了Tier1业务进入全球前十的战略目标。她还提出，十年内将单一客户营收占比降至1/3以下，这个客户，就是苹果。后来有分析称，王来春之所以如此决绝是因为看到了苹果供应商欧菲光的下场，2021年3月，欧菲光被踢出苹果供应链，但由于其业务太过于依赖苹果，进而受到了重创。现在看来，立讯精密摆脱“果链依赖症”的效果也非常明显，从2021年到2025年，苹果贡献的业绩占比分别为74%、73%、75%、71%、57%。整个库克时代，立讯精密陪伴苹果走过了十五年，业务覆盖了其各个业务线，营收也从25.6亿元增长至了3323.4亿元，市值也随之增长了近40倍。而如今库克即将退休，立讯精密也找到了汽车电子、通信业务等新增长引擎。从某种程度上来说，这也是另一种圆满。公司上市十六年，她身价涨千亿再说回开篇提到的对京西国际的收购，这笔交易之所以能够直接带动立讯精密股价的上涨，与前文提到的“Tier1业务进入全球前十的战略目标”有直接关系。此前，立讯精密的汽车业务主要集中在“线束”和“连接器”领域，而京西国际的优势在悬架系统和制动系统，这笔收购直接补齐了立讯精密在底盘域执行机构的短板。更重要的是，京西国际拥有从德尔福继承的磁流变悬架和线控制动等全球稀缺技术，也拥有宝马、保时捷等欧洲顶级车企客户。要知道，底盘认证周期高达3至5年，技术壁垒极高，立讯通过这笔收购，相当于买了个时间差，直接拿到了这些高端客户的“入场券”。另外根据去年财报，汽车电子作为第二增长曲线，迎来了爆发式跃升。该板块实现营收392.55 亿元，同比大幅增长185.34%，营收占比快速提升至 11.81%，成为立讯精密增长最强劲的引擎。当然，5月11日市值能够创下新高，也离不开其四月底以来持续利好消息带来的股价拉升。4月27日，立讯精密收获了又一个涨停，最高冲到了72.6元，创下历史新高，总市值突破5200亿元。当天天风国际分析师郭明錤独家爆料，OpenAI计划自研手机，立讯精密成为独家系统联合设计和制造合作伙伴，同时，公司预告2026年上半年归母净利润78.4亿—81.06亿元，同比大增18%-22%。在此基础上，立讯精密近期披露，将持续加大铜光互连、散热及电源等核心产品的研发投入与产能布局，热门概念叠加下，立讯精密获得了资本真金白银的投票。从2010年上市时不足70亿元的市值，到如今5542亿元的巨头，立讯精密用16年完成了79倍的市值增长。而王来春的身价也从上市之日的23亿元增长到了1039亿，实现了超千亿的财富增加。然而，这并非是王来春身价的顶点。毕竟立讯精密已经向港交所递交了上市申请书，冲刺A+H上市，并预计今年下半年登陆港交所。那时，“A股果链一哥”将向全球科技制造龙头发起冲击，更重要的，其作为全球唯一同时布局消费电子、汽车与光通信产业的企业，其资本溢价将进一步释放，王来春的新故事也将开启新篇。本文来自微信公众号“投中网”，作者：张雪，36氪经授权发布。本文由「36氪的朋友们」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、采访供图

科氪 | 对话韶音高层：开放式耳机，凭什么成为下一场颠覆TWS的“品类革命”？科氪·2026年05月17日 19:57韶音深耕开放式音频，以硬核技术推动全场景品类革命从运动孤岛到品类革命：解构韶音的“开放式”野心与底层技术硬实力在消费电子行业普遍面临增长红利见顶的今天，音频赛道却正在上演一场深刻的结构性变革。2026年5月15日下午，大雨初晴的深圳湾广场，Shokz韶音举办了备受瞩目的媒体专访会。韶音中国区CEO杨云与产品线总裁张浩锋悉数到场，向外界系统阐述了这家以“骨传导”和“运动”起家的品牌，如何在稳坐全球开放式耳机出货量第一王座的同时，向更广阔的泛场景音频帝国发起攻坚。过去两年，开放式耳机在中国蓝牙耳机市场的整体销售额份额已率先突破30%。这不仅是一次产品形态的更迭，更是一场由韶音引领、正在席卷全球的“品类级替代”风暴。一、 宏观战略与行业洞察：以“全场景品类替代”重塑全球音频版图过去，大众对韶音的认知多停留在“专业运动”与“骨传导”这两个标签上。然而，在本次专访中，韶音高层明确指出，专业运动与泛场景用户从来不是一道非此即彼的选择题。韶音正在做的事情，是将用户的定义从最初的“专业运动人群”彻底拓展到“每一个需要开放聆听的用户”。官方图片1. 舒适成为刚需：跨越品类鸿沟的终极逻辑韶音之所以做出“开放式耳机有机会取代真无线耳机”这一极具前瞻性的行业断言，其核心洞察深植于用户日常用耳习惯的根本性改变。随着线上办公、远程会议、长音频内容、网课以及日常碎片化聆听需求的激增，现代人佩戴耳机的时间正在无限拉长。在长时间佩戴的背景下，传统入耳式耳机带来的耳道压迫、闭塞闷痛、卫生隐患以及无法感知外界环境的安全痛点，逐渐演变为不可调和的刚性冲突。而开放式耳机所具备的“全天候舒适佩戴”与“安全开放聆听”的天然优势，正让它从一款特定场景的“运动副机”，跨越成为用户日常生活中无可替代的“主力设备”。在音质水平已经通过近年来的技术迭代基本解决、甚至在部分产品上（如OpenFit Pro）接近或超越传统入耳式耳机的今天，用户在舒适度与开放感知方面的倾斜，成为了开放式耳机跨越品类鸿沟的决定性力量。官方图片2. 全球化视野下的“GLOCAL”战略在出海与全球化版图中，韶音在2023年至2025年间已连续三年斩获开放式耳机市场全球出货量份额第一的桂冠()。数据显示，不仅是中国市场在狂飙突进，在全球重点市场如美国、欧洲、日本、韩国、澳大利亚等地区，开放式耳机占比持续上升、对传统入耳式耳机产生强烈替代的趋势也已不可逆转。面对全球化竞争，韶音的核心市场规划可以总结为“GLOCAL”战略——即全球化思考，本地化行动。杨云将当前的耳机品类竞争比作新能源汽车行业：全球绝大多数的核心品牌与顶尖供应链都集中在中国市场。因此，韶音选择将中国激烈的市场环境作为研发的“练兵场”，充分依托国内强大的工程师红利和极致的供应链优势，开发出在硬件、性能上足以打动并销售全球的标准标准化产品。但在软件、服务及生态融合层面，韶音则在全球范围内开展极其精准的本地化深耕。例如，最新发布的专业运动耳机OpenSwim Pro 2，便成为了全球首款支持QQ音乐无线导入与离线播放的耳机，这一功能精准击中了本土运动爱好者的核心痛点，成为了全球标准化硬件与本土顶级流媒体生态成功联姻的范例。官方图片3. 拒绝短视的长期主义护城河在涌入者众多的多元化竞争环境下，韶音认为品牌真正的护城河来源于“技术创新”与“用户思维”的底层基因，以及对长期主义的笃定。做品牌是一条最难走、面临挑战与压力最大的路。在过去长达15年的时间里，韶音只做了一件事，那就是专注于开放式音频的深耕。这种克制与长线思维同样体现在其品牌营销与用户沟通上。从跨界合作自由潜水世界冠军金明，到携手全球品牌大使张婧仪推出艺术礼盒，韶音的策略重点并非盲目追求流量快钱，而是通过这些契合品牌坚韧、优雅特质的合作伙伴，将“开放聆听、不急于求成”的理念具像化，从而在消费者心中建立起高端化、专业化且极具信任感的品牌壁垒。二、 产品迭代与技术革新：材料与声学的极限攻坚要在红海中完成对传统TWS的品类替代，单纯依靠概念营销是远远不够的，必须在产品和技术实操层面，以硬核的底层研发打破开放式耳机长期存在的物理瓶颈与行业痛点。在最新的产品矩阵中，韶音通过声学、材料学与垂直场景的深度融合，带来了行业瞩目的迭代升级。1. 5年磨一剑：以“韶音天篱滤噪技术”定义开放式新阶段在开放式耳机的普及过程中，抗噪能力弱、在嘈杂环境下听音模糊一直是阻碍其进入更广泛泛场景的致命短板()。针对这一行业痛点，韶音在行业内首次提出了“开放式滤噪”的颠覆性概念。开放与滤噪，在物理和声学原理上天生是一对巨大的矛盾。在项目启动初期，这项技术在韶音内部甚至引发了巨大的争议，许多研发人员一度认为这是无法跨越的物理鸿沟。然而，基于对技术的长期投入和“因为相信，所以看见”的技术信仰，韶音早在2020年底至2021年便前瞻性地立项了滤噪预研技术——当时行业内甚至还没有出现如今大火的耳挂式开放式耳机，研发完全是基于骨传导的未来构想展开的。历经5年多的高强度技术迭代与攻坚优化，韶音终于推出了“韶音韶音天篱滤噪技术”。该技术成功将滤噪的频段拓宽至接近4000Hz，并将滤噪深度降到了18dB，使其真正成为行业首个能够在健身房、咖啡厅、办公室等多场景完美适配的开放式滤噪方案。这标志着开放式耳机进入了一个全新阶段：既能保留开放聆听带来的环境感知与安全出行，又能兼顾沉浸专注的清晰听音体验。张浩锋更是透露，如果这项技术在未来能够继续向高铁、飞机等高难度嘈杂场景演进，那将直接触发音频行业的下一场彻底革命。官方图片2. OpenSwim Pro 2：解锁游泳蓝海的“全能数据终端”多数音频厂商将运动耳机的目光死死锁在跑步赛道，而韶音则将下一个具备决定性增量的突破口精准锁定在了游泳市场()()。根据韶音内部的全球调研数据统计，经常游泳的人群已经占到了跑步人群的60%左右，这是一个庞大却长期被行业低估的蓝海。由于游泳场景极为单调且无法在运动中聊天拍照，用户在水中对音乐激励的需求比跑步往往更加强烈和刚性。然而，水下由于蓝牙信号无法有效传输、防水等级要求严苛，市面上长期缺乏优秀的解决方案，这一领域对于高品质音频而言几乎是零竞争对手的真空地带。为此，OpenSwim Pro 2不仅将防水能力直接拉高到了极致的潜水级，更在硬件内部塞入了一颗强悍的6nm定制芯片与九轴传感器。基于对真实用户需求的深刻洞察，游泳者如果佩戴智能手表，在水中频繁抬腕观察数据会严重打断运动的节奏和划水连贯性。OpenSwim Pro 2通过这套顶级的传感器与芯片组合，能够精准识别泳姿、划次、游泳距离等核心运动指标，并通过语音实时播报的方式反馈给用户，完美做到了声学、运动科技与极端场景体验的闭环融合。3. OpenDots 2：夹式耳机的“微米级”舒适平衡近年来，耳夹式（Clip-on）形态因其潮流属性和独特的佩戴方式，正在疯狂吸纳传统的TWS用户。但耳夹式产品始终面临着一个行业内难以调和的悖论：夹得太紧耳廓容易产生压痛，夹得太松则在剧烈运动中容易松动脱落。为了达成“久戴不痛、狂甩不掉”的黄金平衡，最新亮相的OpenDots 2在完美延续了前代OpenDots ONE所奠定的“动钛弧结构”以及“9G预紧力”两大技术核心优势的基础上，在材料学与重心分布上进行了极其细腻的微米级微调。在结构优化上，OpenDots 2将内部核心的钛片宽度做到了极致压缩，最大减少了0.6毫米，这微小的尺寸缩减极大地减轻了对耳廓边缘的局部压迫感。同时，研发团队重新优化了耳机整体的重心分布，并全面提升了钛片的回弹性与抗疲劳度，从而保证了耳机在高强度、高频次拉伸使用的长期蹂躏下，依然能维持极其稳定、均衡且高性能的夹持体验。这种通过材料创新来化解佩戴稳固与舒适矛盾的研发路线，进一步筑牢了韶音在舒适科技领域的领先地位。三、 未来战略演进：在AI与智能硬件洪流中的冷思考在当前言必称AI、智能眼镜与可穿戴健康监测设备大乱斗的硬件大潮中，韶音在专访中展现出了极强的战略定力与极其清醒的“冷思考”。面对未来的技术演进，他们并没有盲目跟风，而是有着自己一套逻辑严密的方法论。1. 激进的稳健派：AI重构智能耳机的三步走节奏对于风头正劲的AI技术，韶音在内部将自己定性为“激进的稳健派”。他们深信AI在未来势必会全方位重构整个智能硬件生态，但这一进程绝非一蹴而就，而是有着极其鲜明的节奏感。在韶音的蓝图中，AI在开放式耳机上的落地被清晰地划分为三个层层递进的演进维度：第一维度是性能增强（已全面落地）：这是目前已经成熟并广泛应用的阶段。韶音通过AI算法深度优化了耳机的抗噪性能、麦克风通话降噪效果以及底层的声学音质表现。第二维度是交互体验的革命（正在切换）：随着AI大模型的普及，人类与硬件的交互媒介正在从过去的键盘、屏幕逐步切换到极其自然的语音对话。由于开放式耳机具备全天候挂耳、不隔绝环境、高频佩戴的天然属性，它被韶音视为最自然、最适配的AI自然语音交互入口，未来在这个维度将爆发巨大的交互革命机会。第三维度是能力增强（韶音最核心聚焦的战略方向）：这一维度不再局限于空泛的泛场景，而是将AI聚焦在特定的垂直场景中大显身手。例如，通过耳机上的传感器采集用户的实时运动行为数据，并借助AI在本地或云端进行深度处理，从而为用户提供私教级的智能运动辅导；或者在办公场景下，通过AI处理声音数据来大幅度提高用户的办公和生产力效率。这种切中垂直场景的能力增量，正是韶音当前最为看好并重兵投入的王牌方向。2. 智能眼镜 vs 开放式耳机：乐观的谨慎派视角针对当前行业内热议的“拥有4亿眼镜人口的中国市场，智能眼镜是否会吞噬耳挂/开放式耳机生存空间”的尖锐话题，韶音高层给出了基于商业底层的冷静理性的研判。张浩锋坦言，韶音对于眼镜市场是一个“乐观且谨慎派”。眼镜作为面部穿戴设备，其复杂程度与面临的物理限制远超耳机。用户可以毫无心理负担地将一双轻巧的耳机挂在耳朵上一整天；但是，如果试图将一整块显示屏幕、复杂的计算单元、大容量电池强行架在用户的脸上，硬件在产品重量、续航能力、散热性能以及制造成本之间的“不可能三角”物理矛盾，在短期内根本无法得到完美解开。因此，开放式耳机与AI相结合的普及速度和商业落地节奏，注定会比智能眼镜来得更快、更扎实。眼镜未来或许有巨大的长线潜力，但它的技术挑战与迭代周期注定会非常漫长。官方图片3. 健康监测的“减法哲学”与价值回归在智能穿戴设备拼命做“加法”、试图把心率、脑电等各种生物传感器塞入耳机的行业风潮下，韶音却展现出了一种极其难得的克制与“减法哲学”。杨云指出，运动健康监测（如HRV、持续心率等）的核心商业价值与用户价值，高度依赖于“7×24小时无缝、长期且连续的数据基准支撑”。这正是智能手环、智能手表这类用户连洗澡、睡觉都几乎不摘单的设备的天然主场。相反，耳机作为一种高频摘戴、使用场景相对碎片化的设备，天生无法提供无断点的连续时间基准数据。如果在预研阶段为了拼凑功能和噱头强行塞入这些复杂的生物传感器，势必会严重牺牲耳机的佩戴重量、体积与续航体验。在经过严密的评估后，韶音认为这无法给用户带来更好的综合核心体验，因此他们选择克制，在不必要的功能上做取舍与减法，让耳机回归到极致的佩戴舒适科技与声学本源上。四、 结语在韶音的终局思维里，开放式耳机从来不是属于一小撮硬核运动人群的专属孤岛，而是一场终将彻底颠覆并重塑传统真无线音频行业格局的品类海啸。从深耕骨传导、定义专业运动，到如今全面布局耳挂、耳夹全形态，走向全场景开放音频；从硬核预研“韶音天篱滤噪”打破物理矛盾，到重塑水下运动的高清聆听体验，韶音正用长达15年的笃定、对真实需求的洞察以及硬核的技术信仰，不断敲碎那些外界自以为是的行业天花板。正如CEO杨云在媒体专访结束时所言，所谓的“终局思维”绝不意味着把天花板定在这里，韶音的技术上升空间与想象力依然巨大。这场将舒适与聆听自主权彻底还给消费者的开放式革命，才刚刚行至中流，而韶音显然已经握紧了通往下一个时代的船票。本文由「科氪」

8点1氪丨多款燃油车大降价，降幅最高近30%；蜜雪冰城上线黄仁勋同款饮品；“巴菲特慈善午餐”回归，获拍900万美元8点1氪·2026年05月18日 07:55今年1至4月份，常规燃油车新车降价车型均价23.6万元的降价力度算术平均达到3.4万元，降价力度达到14.6%。今日热点导览张雪机车，再夺冠长安汽车否认与千里科技合作深圳发布2026年住房发展计划贵州茅台宣布公斤茅台等四款产品提价法拉第未来宣布获得2500万美元融资三星电子会长就公司“内部问题”公开道歉TOP 3 大新闻多款燃油车大降价，降幅最高近30%日前，长安汽车旗下长安逸动经典版正式上市，新车售价6.49万元，限时享23000元购车现金优惠；相较于指导价8.79万元降幅达到26%。该款车型还享有限时的保养礼和流量礼。其实，近期多家燃油车都开启了降价模式：起亚新狮铂拓界一口价10.99万元，较指导价15.49万元降29%；捷豹路虎5月12日推出限时“新发现”价；奇瑞4月中旬开展促销，综合补贴最高4万元；速腾2026款最低价11.99万元，降价3万元（降幅20%）；思域最低指导价9.79万元，降价3.2万元（降幅24.6%）；荣威i6最低价7.49万元，降价1.49万元（降幅16.6%）；一汽-大众奥迪A6于3月降价9.1万元（降幅29%）。乘联分会秘书长崔东树最新披露的数据显示，2026年4月份，常规燃油车新车降价车型均价13.1万元的降价力度算术平均达到2.3万元，降价力度达到17.2%。今年1至4月份，常规燃油车新车降价车型均价23.6万元的降价力度算术平均达到3.4万元，降价力度达到14.6%。（光明日报）蜜雪冰城上线黄仁勋同款饮品5月15日，英伟达CEO黄仁勋在北京南锣鼓巷的蜜雪冰城门店点了杯“蜜桃四季春”，引发关注。5月16日上午，有媒体记者注意到，蜜雪冰城官方点餐小程序已火速上线“大佬同款”专属分类，直接把这款单品置顶。郑州中原区多家门店的店员告诉记者，这款饮品价格没变，还是7元起，而且“大佬同款”销量相当不错。蜜雪冰城相关工作人员回应，小程序上线“大佬同款”后，蜜桃四季春当日销量环比前日增长接近140%，同比上周五增长超过90%。黄仁勋到访的南锣鼓巷店，蜜桃四季春销量环比增长超过90%。另有消息称蜜雪冰城CEO张渊在朋友圈晒出黄仁勋喝“蜜桃四季春”的照片，并配文：“最顶级的显卡，往往需要最纯粹的散热方式。”同时，蜜雪冰城官方微博也发文称：“高温模式已开启！液冷降温和蜜桃四季春更配哦。”（界面新闻、新浪财经）“巴菲特慈善午餐”回归，获拍900万美元据eBay网站显示，在一场慈善筹款拍卖活动中，一位竞拍者同意出资900万零100美元，以换取与亿万富翁投资者沃伦·巴菲特以及篮球巨星斯蒂芬·库里共进晚餐的机会。目前尚无法立即确定中标者的身份。6月24日，巴菲特、库里及其妻子将与中标者及其选定的最多七位随行嘉宾，一同前往位于内布拉斯加州奥马哈市的伯克希尔·哈撒韦公司总部所在地共进晚餐。（同花顺财经）大公司/大事件小米卢伟冰：下半年部分国产旗舰直板手机价格或将破万5月16日，小米集团总裁卢伟冰在直播中表示，今年下半年，尤其是年底，可能有一些国产旗舰直板手机的价钱会突破1万元人民币。针对小米17 Max价格会不会受到内存价格影响，卢伟冰称，今天任何一款手机定价都会受到内存成本大幅上涨的影响，而小米17 Max的价格还在讨论当中。他预判，内存涨价趋势至少会延续到2027年底，甚至到2028年。（界面新闻）张雪机车，再夺冠当地时间5月16日，在刚刚结束的2026世界超级摩托车锦标赛（WSBK）捷克站WorldSSP组别第一回合正赛中，中国摩托车制造商“张雪机车”旗下的53号法国车手瓦伦丁·德比斯夺得冠军。（央视新闻）巨力索具今年25次登龙虎榜，在A股上市公司中高居第95月15日晚，今年以来股价翻倍的商业航天概念股巨力索具公告称，公司于当日收到中国证监会下发的立案告知书，因涉嫌信披误导性陈述违法违规，根据相关法规，5月13日，中国证监会决定对公司立案。值得注意的是，今年以来，巨力索具25次登上龙虎榜，在A股上市公司中高居第9，累计买入143.86亿元、卖出134.52亿元，累计净买入9.35亿元。龙虎榜上，机构与游资频繁出没，散户跟风炒作。巨力索具披露的2026年一季报显示，一只私募基金、一家银行机构，以及两位高度关联的“超级牛散”，重仓押注，跻身前十大流通股东。而控股股东巨力集团则通过股权协议转让，持仓大幅减少4800万股。（中国证券报）长安汽车否认与千里科技合作近日有消息称，长安汽车将采用千里科技的智能辅助驾驶方案，并投资入股。长安汽车5月17日称，“暂无计划”。另有接近长安汽车的人士向记者透露：长安汽车自研的天枢智能已经到了收获期，不可能切换智能驾驶方案。据悉，去年12月，工信部公布了我国首批L3级有条件自动驾驶车型准入许可。长安汽车位列其中，搭载的是其自研的“天枢智能”辅助驾驶系统。（第一财经）蔚来：已在全国建成8870座充换电站36氪获悉，蔚来发布消息称，5月17日，NIO Power 2站上线。截至5月17日，蔚来全国建成8870座充换电站，其中换电站3849座，充电站5021座、充电桩28855根，累计提供超1亿次换电。深圳发布2026年住房发展计划，将筹建3万套、供应4万套保障房36氪获悉，深圳市住房和建设局近日发布《深圳市2026年住房发展年度计划》。优化保障性住房供给，建设筹集保障性住房3万套（间）、供应4万套（间），满足工薪群体和各类住房困难家庭基本住房需求；推进“好房子”试点项目建设，推进城中村改造和城镇老旧小区整治改造，推动创建未来社区（完整社区）试点，切实增强市民获得感、幸福感、安全感。最高涨价200元，贵州茅台宣布公斤茅台等四款产品提价5月16日凌晨，i茅台公告，自当天零时起，贵州茅台酒股份有限公司对部分贵州茅台酒产品自营体系零售价格进行调整。此次调整共涉及四款非标产品，最高涨幅为8%。调整的产品包括：53%vol 500ml陈年贵州茅台酒（15）由4199元/瓶调整为4279元/瓶；53%vol 500ml贵州茅台酒（精品）由2299元/瓶调整为2359元/瓶；53%vol 500ml贵州茅台酒（丙午马年）珍享版由2499元/瓶调整为2699元/瓶；飞天53%vol 1L贵州茅台酒由2989元/瓶调整为3119元/瓶。（澎湃新闻）国内航线燃油附加费5月16日起上调5月16日起，中国内地航线客票，燃油附加费上调。具体标准为：800公里以下航段每位旅客收取90元；800公里以上航段每位旅客收取170元。婴儿旅客免收燃油附加费，儿童、革命伤残军人、因公致残人民警察客票燃油附加费减半收取。（央视新闻客户端）2026年5月票房破15亿36氪获悉，据灯塔专业版，截至5月17日，2026年5月票房（含预售）突破15亿，《给阿嬷的情书》领跑5月电影票房榜。上海电信率先发布Token资费套餐：1元对应25万额度点，支持手机账单付面对AI时代日益增长的Token需求，上海电信宣布，即日起面向上海电信用户正式推出Token算力服务，成为上海首个发布Token资费套餐运营商。具体来说，一是按量订购，1元对应25万额度点（以KiMi-K2.5模型为例，约可支持调用25万输入Tokens）；二是按需选择，用多少买多少，多买可享折扣，支持话费账单支付。另据记者从上海电信获悉，用户购买额度点后，通过标准API接口即可调用30余款主流大模型（文本、多模态），将AI能力集成到自己的软件、脚本或自动化流程中。以10元购买250万额度为例，约可让程序自动总结100本10万字的电子书。(上观新闻)超310家企业落户香港或在港扩充业务香港特区政府投资推广署署长刘凯旋表示，截至5月初，已有310多家企业落户香港或在港扩充业务，首年直接投资额超260亿港元，其中中资企业占比超过五成。她认为这一成果得益于特区政府成立的“内地企业出海专班”的推动。(大湾区之声)法拉第未来宣布获得2500万美元融资36氪获悉，5月16日，法拉第未来宣布，公司已与机构投资者签署了2500万美元股份购买协议。根据协议，投资者已认购公司总本金为2500万美元的可转换本票。本次融资对应的普通股目前尚未注册，存在交易限制，暂不可立即流通。其中，1250万美元将直接汇入公司运营账户；其余1250万美元将依据公司与各投资者签署的受控账户协议，存入由投资者控制的监管账户，并将在满足特定条件后释放给公司使用。星巴克宣布在美关闭多个地区办公室当地时间5月15日，星巴克宣布将在美国裁员300人，以削减成本、力求重回增长轨道。星巴克还表示，关闭多个地区办公室，相关举措将产生包括遣散费以及资产减值等在内的4亿美元重组费用。（央视财经）伊朗宣布5月19日起恢复股票市场交易当地时间5月16日，伊朗证券交易组织表示，伊朗股票市场将于5月19日起恢复交易。伊朗证券交易组织负责交易所相关负责人表示，根据已完成的协调工作并在获得必要许可后，决定自19日起重新开放股票市场、各类股票投资基金及其衍生品交易。（央视新闻）盖茨基金会不再持有微软股票当地时间5月16日，根据提交给美国证券交易委员会的文件，微软创始人比尔·盖茨的慈善基金会已不再持有微软股票。盖茨基金会信托披露，已于今年第一季度出售剩余770万股微软股票，按当前股价计算价值32亿美元。（界面新闻）马斯克称他不会出售任何SpaceX股票马斯克称他不会出售任何SpaceX股票。（华尔街见闻）SpaceX据悉最早将于6月11日确定首次公开发行价格5月16日消息，据报道，SpaceX最早将于6月11日确定首次公开发行价格，股票将于6月12日上市。SpaceX将选择纳斯达克作为其首次公开发行的上市地点，股票将以“SPCX”为代码进行交易。（界面新闻）SpaceX拟推万亿火星薪酬方案，马斯克回应：旨在确保公司专注于多行星生活愿景5月16日，有消息称，正在推进上市的SpaceX在保密招股书中披露了一项激进的治理与薪酬方案。公司计划通过发行10倍投票权的B类股，确保马斯克拥有不可被解雇的绝对控制权。同时，董事会批准了一项总额可达万亿美元的超级薪酬方案，其考核标准直接与在火星上建立百万人口殖民地的目标挂钩。目前，SpaceX计划于今年6月登陆纳斯达克。针对这一治理方案，马斯克随即做出了公开回应。他表示：“我需要确保SpaceX始终专注于让生命成为多行星化、并将人类意识延伸至星辰大海的愿景，而不是去迎合某些人极其无聊的季度财报奖金！”马斯克说，这一目标实现起来异常艰难，过程绝非一帆风顺，但他坚信，“如果SpaceX成功实现了这个荒谬而艰巨的目标，它的价值将比地球目前的经济总量还要高出好几个数量级。”（界面新闻）三星电子会长就公司“内部问题”公开道歉5月16日，三星电子会长李在镕就公司“内部问题”公开道歉。目前，三星电子与公司工会之间劳资纠纷持续，工会计划于5月21日启动的18天罢工计划已有超过4万名员工登记。李在镕呼吁工会与公司团结一致，朝同一个方向前进。三星电子劳资双方自去年12月起开展谈判，但由于分歧一直未能缩小，谈判于今年3月破裂。此后，在调解下，劳资双方本月11日至13日进行了马拉松式谈判，但最终未能达成协议。据韩联社报道，若进行罢工可能会造成40万亿韩元、约合1824亿元人民币的损失，并殃及半导体产业，政府可能考虑行使紧急调整权。（CCTV国际时讯）OpenAI与马耳他达成协议，所有马耳他民众都能使用ChatGPT Plus美国人工智能（AI）公司OpenAI5月16日表示，已与马耳他政府签署协议，所有居民在完成使用AI培训课程后，均可免费使用其ChatGPT Plus服务一年。该项目将于5月启动，随着更多马耳他居民完成这门免费课程，服务规模预计将逐步扩大。该计划也将向旅居海外的马耳他公民开放。（新浪财经）贝莱德考虑对SpaceX首次公开募股投资数十亿美元据知情人士透露，贝莱德已讨论在下个月对SpaceX的首次公开募股（IPO）投资50亿至100亿美元。这笔巨额投资对应的IPO募资规模最

对话万成云商：发文章≠GEO优化，大模型不是喂什么就推什么杨越欣（杨桃）·2026年05月18日 13:05对话万成云商：拆解出海企业AI时代GEO布局路径“帮我搭配一套适合海边度假的衣服。”“帮我推荐一家适合约会的西餐厅。”“3000元预算，买哪款扫地机器人最划算？”这类与AI的问答正发生在全球数亿用户身上。过去，人们会把这些问题输入百度、谷歌等搜索引擎，因此产生了SEO（Search EngineOptimization，搜索引擎优化）营销。而当流量入口从搜索引擎转向AI工具，GEO（Generative EngineOptimization，生成引擎优化）也应运而生。区别于传统SEO追求的“搜索结果排名”，GEO的核心目标是让品牌信息在AI生成的回答中被优先引用和推荐。在AI全面渗透用户信息获取与消费决策场景的时代，对于出海企业而言，做好GEO意味着抢占新的流量入口。GEO的逻辑和SEO有什么根本不同？为什么有人说GEO是伪命题？出海企业布局GEO该怎么做？中小企业如何低成本起步……带着这些出海企业关心的问题，我们特别对话了万成云商品牌出海事业部总经理谭莉，拆解AI搜索时代如何让企业被AI看见、信任，并推荐给用户。36氪：GEO被称为SEO在AI时代的进化。相比于SEO的核心在于搜索结果的排名，GEO优化的关键是什么？它与传统SEO的核心区别在哪里？谭莉：核心区别在于传统SEO偏“规则博弈”，GEO更像“信任竞争”。因为过去做SEO，我们的目标是把页面排到Google第一页，让用户点进去；但现在很多用户直接在AI里问问题，甚至不点链接了，那你要解决的是——怎么让AI在回答里“引用你”甚至“推荐你”。所以GEO的关键首先是，你是不是一个“可信任源”（全网有没有人认可你）；其次，你的内容能不能被AI理解、拆解、复用；最后是你的内容有没有完整的证据链，比如数据、案例或者第三方背书。36氪：继GEO之后，也有人提出AEO概念，两者有什么区别？谭莉：这两个概念很多人会混着讲，但我个人是这么区分的：AEO更偏“问答结果优化”，核心是让你的内容直接成为某个问题的标准答案，比如FAQ、精选摘要这些。GEO更广，它面对的是生成式AI，不只是回答一个问题，而是参与“多源信息整合后的输出”。所以AEO更偏“结构技巧”，GEO更偏“整体品牌与内容生态”。36氪：GEO的概念近两年引起广泛讨论。为什么现在这个时间点企业需要关注GEO？在AI搜索渗透率持续提升的背景下，GEO应该在企业营销体系中占据怎样的战略位置？谭莉：分享几个很现实的变化：现在用户越来越习惯直接问AI，而不是自己筛选搜索结果；而且AI在很多行业决策前端，已经开始替代“信息收集阶段”；再就是有些客户还没打开你的官网之前，可能就在AI里形成了对你的判断。所以企业没办法去忽视GEO，现实摆在这里了。我认为，GEO在企业营销体系中占据着很关键的位置，它不会替代SEO，而是和SEO并行的“第二流量入口”，甚至未来会变成主入口之一。36氪：怎么理解有企业做过一些尝试后认为GEO是伪命题，或者靠自己持续给大模型喂语料也可以？谭莉：我觉得这是对AI工作机制有点误解，大模型不是你“喂什么它就推荐什么”。它有几个筛选逻辑：第一，信息是否一致（比如不同来源是否能得到印证）；第二，来源是否可信（是不是长期存在的内容资产）；第三，是否被多平台提及（而不是单一来源）。所以你自己“喂语料”，本质只是单点信息输入，很难形成“信任网络”。不过我也理解为什么有人说GEO是伪命题，因为确实有很多“伪做法”，比如只做站内内容、只堆AI文章、不做外部验证，这些基本都很难起效果。36氪：当企业布局海外GEO营销时，需要经历哪些关键步骤？不同行业、不同规模的企业在路径上会有哪些差异？谭莉：从实操角度，会有四个关键步骤：第一步先拆解你的客户是怎么筛选供应商的，这个路径是怎么样？第二步重建你的内容结构，包括官网、博客、FAQ、案例这些，全部按“可被AI理解”的方式重构。第三步是搭建外部信任体系，比如评价平台、媒体曝光、社媒讨论、行业引用，都很重要。第四步就是做多渠道分发，持续让内容在不同平台被看见和引用。不同企业可能在行业复杂度和品牌基础这两方面存在明显差异。比方说医疗、工业类比其它行业更需要专业背书，那有品牌的企业就更容易放大，没有品牌的就要先“补信任”。36氪：在实施过程中，企业应如何在内容策略和技术基建层面做出改造甚至重塑，以同时满足AI理解与用户共鸣？谭莉：内容策略方面要做到结构清晰，信息有证据，有数据来源，表达要接近真人沟通。技术层面反而是基础的更容易被忽略：比如Schema结构化数据（Product、FAQ、Article等）、robots、index、hreflang这些基础配置，再一个就是页面可读性。总结来说，是做“可被理解的内容”。36氪：在布局GEO出海方案时，企业最容易忽视但至关重要的环节是什么？对于缺乏专业团队的中小企业，有哪些可行的解决方案或合作模式？谭莉：我们服务这么多跨境企业，发现最容易忽视的一点其实是：“让AI能访问你”。听起来很基础，但很多企业在实操中会陷入误区：比方说robots把AI爬虫挡了、页面noindex设置错误，或者多语言混乱……这些都会直接导致你做再多内容，AI也看不到。第二个容易忽视的是站外信任建设。很多人还停留在“发文章=做优化”，但GEO更看重：评论、测评、社媒讨论、媒体引用。对于中小企业，我建议两种路径：①轻量级：聚焦一个细分领域做深内容+少量高质量外部背书；②合作型：找有资源的服务商做整合（内容+渠道+数据），不要自己盲目试错，成本会很高。36氪：企业在选择GEO服务伙伴时，应该重点考察对方的哪些核心能力（如是否具备垂直行业知识库、跨平台渠道资源、效果监测与分析体系等），以避开“无效投入”的陷阱？谭莉：这个问题非常关键，因为现在确实有不少“概念型服务商”。我建议重点看四个能力：第一，行业理解能力，是不是懂你这个行业，而不是只会写内容。第二，内容生产质量。有没有真实案例、是否具备本地化能力，而不是纯AI批量生成。第三，渠道与分发能力。有没有海外平台资源，比如媒体、社区、测评渠道。第四，数据监测体系。比如做一段时间了，看看有没有被AI引用？哪些内容有效？ROI如何？如果对方只能讲“发多少篇内容”，基本可以直接排除。36氪：2026年初Seedance2.0的爆火，让业界看到了AI视频生成从“抽卡式试错”向“工业化生产”跨越的可能性 。10秒带货视频成本仅需8元的“成本塌方”对于跨境电商的内容生产逻辑意味着什么？谭莉：这个变化其实挺颠覆的。过去做视频，成本高、周期长，所以大家拼的是“有没有内容”；但现在成本接近于零，意味着：内容不再是门槛，内容“是否有效”才是门槛。36氪：当视频生产成本趋近于零，所有人都能批量生成高质量内容时，企业的竞争焦点会发生怎么样的变化？从GEO布局的角度看，企业应如何构建AI时代下的营销思维？谭莉：供给无限的时候，稀缺的就变了。我认为未来竞争会集中在三点：第一，认知占位，用户心里有没有你；第二，信任结构，有没有被多方验证；第三，数据能力，能不能快速迭代内容。从GEO角度，我觉得企业要建立一种新思维：要把品牌内容当资产去做长期积累，同时将口碑、评价这些信任一点点做起来，更重要的是发内容之前，问问自己：AI会不会用它？用户会不会信它？避免做“自嗨”的内容。36氪：Seedance2.0的火爆也引发关于深度伪造和版权的讨论。跨境电商卖家在利用这类工具生成营销素材时，应如何规避潜在的肖像权、版权风险？谭莉：这个问题确实不能忽视，我建议，尽量避免使用可识别人物（尤其是公众人物）做生成素材，素材来源尽量用可商用授权库，品牌、Logo、音乐等要确认授权范围，涉及医疗、功效类内容要特别谨慎，因为合规风险更高。毕竟，AI可以帮你降低成本，但不能帮你“免责”。本文由「杨越欣（杨桃）」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄

Agent、多模态、应用、算力一天看尽，峰会亮点在此｜5.20日，来现场一起AI未来一氪·2026年05月18日 15:03第四届中国AIGC产业峰会5月20日北京举办，干货满满开放报名。进入2026，AI愈发狂飙突进。围观体验之余，人人不免在心中自问：朋友圈刷屏的“龙虾”、Harness等AI新事物，跟我到底有什么关系？真的有必要跟吗？AI创业、AI融资如火如荼，属于我的机会又在哪里？别人已经在用AI做视频、写代码、跑项目，我是不是已经慢了一拍？……到最后，几乎所有问题都会汇成同一个问题：我，到底该如何用AI？如果你对这些问题还很模糊，不妨来第四届中国AIGC产业峰会走一趟——一天时间，把这一年AI产业最值得关注的人、事、判断，一次性讲清楚。先提前剧透一波。18位重磅嘉宾，1场Agent主题圆桌，1份年度榜单，1张全景图谱——所有内容都已敲定。阵容上，既有昆仑万维、智谱、商汤、百度、蚂蚁、MiniMax这样的行业头部，也有亚马逊云科技、硅谷Fusion Fund带来的全球视角；既有复旦邱锡鹏、港大黄超代表的学界与开源前沿，也有盛大EverMind、太初元碁、趣丸科技等活跃实战派玩家。议程上，从Agent商业化落地、多模态与空间智能的最新突破，到AI在文娱、医疗、企业服务等场景的纵深渗透，再到算力与AI Infra的范式之变……这一年最值得关注的AI话题，几乎都装进了这一天。话不多说，下面直接上重点。亮点1：Agent到底怎么落地？不同视角玩家都有心得如果说去年大家还对“Agent是不是未来”持观望态度，那么2026年，这个问题的答案已经不言而喻。“Agent不是未来，是现在”——这正是智谱高级副总裁吴玮杰带来的判断，他将分享智谱在Agent商业化落地上的最新实战。但Agent真正走入企业，并非一蹴而就。亚马逊云科技产品技术部技术总监王晓野将聚焦如何跨越Agent落地鸿沟，拆解从最强模型到企业级AI Agent之间那条最难走的路。而港大助理教授黄超则会从学术与开源前沿的视角，探讨AI Agent范式变革的挑战与机遇。他所带领的HKUDS GitHub开源平台累计获得超24万Star，位列全球Top50，话语权毋庸置疑。与此同时，Agent要真正成为“数字生产力系统”，还需要更底层的能力支撑。盛大集团副总裁、EverMind CEO邓亚峰将分享长期记忆驱动的自进化，探讨AI如何从一个“工具”演化成可持续成长的数字生产力系统。当然，光有演讲还不够。本届峰会专门设置了一场Agent主题圆桌：2026，Agent产品的不确定性与非共识机遇。出席嘉宾包括趣丸科技副总裁兼首席战略官、《屠龙之术》主理人庄明浩，ColaOS&MarsWave CEO冯雷，以及EvoMap创始人张昊阳。这三位都是Agent产品一线的深度玩家，从产品形态到操作系统，从多智能体框架到自进化引擎，他们会碰撞出什么样的非共识观点？非常值得期待。亮点2：Agent之外，模型能力边界也在继续扩张虽然Agent是今年峰会的绝对主角之一，但不可忽视的是，支撑Agent的模型能力本身，也在以惊人的速度向外延伸——从“数字”到“物理”，AI对世界的认知边界正在被重新定义。商汤科技执行副总裁、首席科学家林达华将分享多模态模型如何从“看懂”和“生成”，进化到真正可感知、可生成、可行动的空间智能新阶段。上午的GenAI Talk环节，蚂蚁灵波科技首席科学家沈宇军将围绕一个核心命题展开对谈：AI 2.0下半场，从AIGC到AIGA。当生成式AI从“内容”走向“行动”，整个产业逻辑将被改写。而京东JoyInside业务负责人戴文军则会把这个命题进一步推往物理世界，当AI从屏幕里走出来，走进真实终端、真实场景，“AI World”才真正开始。模型层面，复旦大学特聘教授、模思智能创始人邱锡鹏将分享MOSS多模态模型及其推理优化的最新进展。这位主持研发了MOSS系列开源大模型的“大模型一线人物”，将带来学术与工程结合的硬核内容。亮点3：具体怎么玩？这些应用玩家甩出成功案例模型也好、Agent也罢，具体到不同领域，机会到底在哪？昆仑万维董事长兼CEO方汉率先给出他的答案。他这些年带着昆仑万维在AI音乐、AI社交、AI短剧等多个AIGC方向同时下注，其分享会从一线视角，讲讲AIGC具体应用这一年究竟跑通了什么、又踩过哪些坑。紧接着，来自不同赛道的“实战派”玩家也将依次登场：文娱创作领域，风行在线CEO易正朝将带来一个核心判断——从AI编程到AI视频，众创才是AI生产力的核心杠杆。他会讲述AI如何从单点工具，进化为撬动整个创作者生态的杠杆，从而让普通人也能撑起一支创作团队。生产力工具层面，百度秒哒产品总经理朱广翔会分享他们“人人都是创造者”的产品理念，当门槛被AI抹平，想象力就是新的生产力。企业级AI方向，MiniMax ToB中国区商业化负责人胡维琦将带来MiniMax这一年在ToB战场上的真实打法与思考。大健康赛道，轻松健康集团执行副总裁马孝武将拆解从循证智能体到AI驱动的健康服务增长飞轮，展示AI在严肃医疗场景里的真实增长引擎。文娱、办公、企业、医疗……AI不再是“被讨论的对象”，而是“正在被使用的工具”。每一个细分赛道里，都已经有人开始把案例端上桌了。亮点4：从训练到推理，算力的故事也在被改写模型与应用之外，2026年最值得关注的另一条主线，是AI基础设施的范式重构。过去几年，所有人讨论算力都默认一个前提——为训练服务。但2026年，风向变了，推理需求暴涨、训练成本高企，整个算力叙事正在从“训练驱动”转向“推理驱动”。Fusion Fund创始合伙人张璐将从全球资本与产业的双重视角，解读算力叙事的重构。当AI基础设施从训练转向推理，一场范式变革正在发生。而国产算力这一边，太初元碁首席产品官、高级副总裁洪源将带来他对国产算力路径的思考。当“Token”正在成为AI时代的新计量单位，国产算力如何筑基Token智能新未来？他会给出自己的答案。亮点5：看尽AI全景生态，一份榜单+一份报告延续往届传统，本届峰会，技术先锋、产业龙头、新锐势力将再次全景式碰撞。为了给行业更多参考，我们还将颁出2026年值得关注的AIGC企业、2026年值得关注的AIGC产品两类奖项，以及一份中国AIGC应用全景图谱。前者是量子位根据过去一年AIGC企业、产品的表现与反馈，并结合对2026年技术与场景的观察，评选出的高潜力企业和产品。后者是量子位智库绘制的一幅全面立体的中国AIGC应用全景图谱，用以展示国内AIGC应用的市场格局与发展动态，并深入剖析最具潜力与竞争力的产品和赛道。5月20日举办最后，时间地点以及最全议程在此，现场还有参会伴手礼等你来领，咱们到时候见 ~时间：5月20日 9:30地点：北京·金茂万丽酒店现在，中国AIGC产业峰会正在火热报名中，欢迎点击链接报名线下参会：https://hdxu.cn/1Hjr1本次峰会也将同步线上直播，欢迎预约参与：量子位视频号。参会嘉宾简介一览方汉昆仑万维董事长兼CEO方汉毕业于中国科学技术大学近代物理系，曾任职于中国科学院高能物理研究所、拓林思、亚信以及千橡互动。于2008年协助周亚辉先生创立昆仑万维。方汉拥有30年互联网从业经验，从1994年开始参与和倡导开源运动，是中文Linux奠基人、中文Linux四剑客之一，著作国内第一本Linux书籍《Linux实用大全》，他也是国内最早的网络安全专家，曾负责三大电信运营商的全网安全审计，领导开发了国内第一款P2P下载软件DUDU加速器，负责研发了国内市场占有率最高的网页游戏《三国风云》，为中国互联网最资深的参与者。吴玮杰智谱高级副总裁吴玮杰目前担任智谱公司高级副总裁，负责公司MaaS平台商业化，并兼任泛互联网事业部总经理及浙江子公司总经理。他在人工智能、互联网和信息技术领域有着丰富的商业化管理经验和项目落地经验，曾任字节跳动企业应用全球首席商业官（旗下品牌包括飞书、LarkPeople和幕布等）、GE数字集团大中华区副总裁、复星集团联席首席发展官&首席投资官。除此以外，他也曾在思科、SAP、Oracle等公司从事商业化管理工作。吴玮杰博士毕业于电子科技大学，同时持有香港大学国际工商管理硕士学位以及复旦大学软件工程学士、硕士学位。易正朝风行在线CEO易正朝，风行在线CEO，作为业务领头人，引领公司从传统互联网视听平台全面转型升级为AI文娱生态平台。推动“AI赋能众创”战略落地，带领团队打造橙星梦工厂、小剪刀等AI创作工具，影视网文智能小程序平台、电视鸿蒙平台等核心智能产品，构建从内容创作、智能分发到商业变现的全流通链生态。推动公司荣获AI赛道奖项、抖音快手等平台的头部服务平台奖项，在AI文娱新时代，推动内容生产效率革新与创作者生态繁荣，致力于让内容流动更简单。林达华商汤科技执行副总裁、首席科学家林达华教授现任香港中文大学信息工程系副教授、交叉学科人工智能研究所所长，并为商汤科技联合创始人、执行董事兼首席科学家，兼任上海人工智能实验室领军科学家。他于2012年获麻省理工学院计算机科学博士学位，研究方向涵盖计算机视觉、深度学习、大模型与生成式人工智能，在人工智能顶级会议和期刊发表论文逾300篇，引用逾90,000次，h-index达131。林教授曾获2010年NeurIPS最佳学生论文奖，带领团队在多项国际计算机视觉竞赛中获奖，连续五年入选Elsevier全球前2%顶尖科学家。他主导发起OpenMMLab，并推动书生大模型、OpenCompass评测平台及DeepLink开放计算体系建设，其中DeepLink于2024年获世界人工智能大会“卓越人工智能引领奖”。2026年，他荣获“中银香港科技创新奖”，以表彰其在计算机视觉与多模态智能领域的系统性创新。他长期服务学术共同体，曾任IJCV编委，多次在人工智能主要国际会议担任领域主席，并担任IEEE大规模深度学习模型评测标准工作组主席。邓亚峰盛大集团副总裁、EverMind CEO邓亚峰，毕业于清华大学，正高级工程师，拥有二十余年人工智能算法及产品研发经验。现任盛大集团副总裁、EverMind CEO。曾任360集团副总裁、人工智能研究院院长兼搜索事业部总经理，科创板首家AI上市公司格灵深瞳CTO。曾荣获2021年中国人工智能年度十大风云人物，教育部技术发明奖二等奖。长期致力于大语言模型、计算机视觉及AI for Science等领域的研究与应用，带领团队多次在国际国内主流AI评测中获得佳绩；累计申请发明专利160余项（已授权98项），发表论文50余篇（含10篇Nature子刊）。王晓野亚马逊云科技产品技术部技术总监王晓野现任亚马逊云科技产品部技术总监。长期专注于企业现代化应用架构，数据及人工智能能力的建设。负责带领技术专家团队提供云上产品及开源方案技术实践指导，并致力于亚马逊云科技全栈服务在中国市场的应用和推广。沈宇军蚂蚁灵波科技首席科学家沈宇军，现任上海蚂蚁灵波科技有限公司首席科学家，同时领导蚂蚁技术研究院交互智能实验室。博士毕业于香港中文大学，研究方向为计算机视觉、生成模型、空间智能和具身智能。在CVPR、TPAMI、ICCV、ECCV、NeurIPS等国际顶级会议和期刊上发表论文百余篇，Google Scholar引用量逾万次。2024年起担任蚂蚁灵波科技首席科学家，致力于探索计算机视觉在机器人行业的落地之路，带领团队构建从空间感知到智能决策、从基座能力到世界建模的完整技术矩阵，推动具身智能从实验室走向规模化应用。邱锡鹏复旦大学特聘教授、上海创智学院院长助理、模思智能创始人邱锡鹏，CAAI Fellow，国家杰青获得者，主要研究方向为大模型、智能体以及科学工程等领域中的应用，发表学术论文被引用4万余次，5次获得AI领域高水平会议的最佳/杰出等论文奖，担任中国人工智能学会自然语言理解专委会副主任、中国中文信息学会大模型与生成专委会副主任、上海市计算机学会自然语言处理专委会主任等，主持研发了MOSS等系列高影响力开源大模型。连续多年入选中国高被引学者、全球前2%顶尖科

2天前安永撤回研究，因被发现含AI幻觉和伪造脚注专业服务公司安永（EY）近日撤回一项关于忠诚度奖励计划影响的研究报告，此前文本鉴别公司 GPTZero 的研究人员指出，该报告存在明显的 AI“幻觉”、引用不存在的文献及伪造脚注，引发对大型咨询机构在使用生成式 AI 生产研究内容时的透明度与可信度担忧。3 来源事件概况：忠诚度计划研究被紧急撤回2 来源GPTZero如何发现AI“幻觉”与伪造脚注2 来源安永面临的信誉考验与对专业服务业的警示3 来源监管、行业规范与“负责任AI”的下一步3 来源本内容由AI生成

2天前OpenAI测试ChatGPT个人理财功能，分析消费与旅行支出，协助规划目标OpenAI 宣布在美国面向 ChatGPT Pro 用户开启「ChatGPT 个人理财」（ChatGPT Personal Finance）预览测试：基于最新 GPT‑5.5 Thinking 模型，通过与 Plaid 集成，只读连接超过 1.2 万家银行与投资账户，自动拉取真实收支、持仓与交易数据，在保证账户不可被操作的前提下，为用户洞察消费结构、分析包括旅行在内的各类支出、制定储蓄与还款计划，并规划中长期财务目标。3 来源OpenAI 推出 ChatGPT 个人理财预览：从“问答案”走向“看全局”3 来源GPT‑5.5 推理能力升级：为复杂个人理财场景而生2 来源从记账到决策助手：洞察消费、分析旅行支出与目标规划2 来源只读接入与隐私保护：OpenAI 的金融野心与试验3 来源本内容由AI生成

1天前华为鸿蒙智行首款旗舰级MPV智界V9上市鸿蒙智行首款旗舰级MPV——智界V9于2026年5月15日正式上市，提供Max、Max+、Ultra、Ultra+四款车型，售价区间38.98万-51.98万元，将于6月初开启交付；新车搭载华为乾崑智驾ADS 5平台预升级能力、896线双光路图像级激光雷达、鸿蒙ALPS健康座舱3.0和华为雪鸮智能增程系统等多项首发技术，CLTC综合续航里程可达1320km，并在上市后2小时内获得超过7000台订单，小订累计突破4万台，锁定中国高端智能MPV市场重磅新标杆。7 来源上市发布：定价、权益与订单表现6 来源外观设计与空间布局：旗舰级车身与豪华座舱2 来源智能驾驶、座舱互联与动力续航1 来源安全架构与市场意义：重新定义高端智能MPV3 来源本内容由AI生成

1天前旧金山软件工程师前景不确定，万名AI从业者五年内财富自由在硅谷和旧金山湾区，新一轮以人工智能为核心的技术变革正在重塑就业版图：一边是Salesforce、谷歌、Meta等大厂在旧金山大规模裁撤软件工程师和中后台岗位，2024年至今全球科技业裁员规模已超过25万人次，氛围明显紧张；另一边是头部大模型公司和独角兽对顶尖AI人才“边裁边招”，部分早期员工和创业者在过去五年里通过期权和退出实现“财富自由”，业内估算已累积至“万级”规模。

这种冰火两重天使旧金山的软件工程师职业前景变得高度不确定：传统Web/移动开发岗位被压缩，重复性开发和测试岗位加速被AI工具取代，而具备扎实工程能力叠加大模型与分布式系统经验的复合型人才，则在竞争中获得更强议价权与收入天花板。3 来源裁员潮下的旧金山：从“工程师天堂”到“氛围紧张”3 来源工程师职业前景急剧分化：哪些岗位在被AI“吞噬”？3 来源冰火两重天：工程师普遍焦虑，少数AI精英五年内“财富自由”未来五年：普通工程师如何在AI时代“自保”与逆风翻盘？2 来源本内容由AI生成

20小时前无锡建Token工厂，首批用4台华为昇腾384超节点服务器5月15日，弘信电子旗下燧弘华创与无锡高新区签约，将在无锡建设全国首个“东数西算、算电融合”的特大型 Token 工厂项目，并同步落地江苏省内首个华为昇腾384超节点算力集群。 首期将在无锡高新区部署4台昇腾384超节点服务器，每台集成384卡规模算力，单套算力最高可达约215 PFLOPS，将4台连成一体后形成国内面向市场端数一数二的大规模国产 AI 算力集群，服务“国芯国模国用”战略，面向长三角及全国输出标准化 Token 级算力能力。7 来源项目签约落地：无锡承接全国首个特大型 Token 工厂4 来源首批部署4台华为昇腾384超节点：江苏省内首个集群7 来源“Token 工厂”模式：把西部绿电算力“加工成”标准化 Token2 来源对无锡及国产算力生态的战略意义5 来源本内容由AI生成

2小时前超级工程师、仓储大脑、监测之眼…… 潍柴工厂里人工智能成主力在山东潍坊潍柴动力2号工厂，人工智能已从“辅助手”升级为“主力军”：研发环节由“AI超级工程师”参与近四成模型决策并借助仿真将实物试验减少约70%，生产与物流由融合大模型和小模型的“智能生产中枢”“仓储大脑”统一调度，质检与安能环管理依托“监测之眼”实现全流程智能监测。凭借覆盖研发、制造、管理、服务等近八成业务场景的AI应用体系，潍柴入选全国首批领航级智能工厂，成为“数智精益”制造模式的标杆。7 来源“AI超级工程师”：发动机研发有了“第二大脑”4 来源“仓储大脑”：打通工厂物流的“神经网络”4 来源“监测之眼”：从质量到安能环的全场景守护4 来源从卓越到“领航”：潍柴打造“数智精益”工厂样板5 来源本内容由AI生成

2小时前特斯拉推送FSD大更新在车主社区与社交媒体上，有消息称特斯拉近期正面向部分车辆推送一次完全自动驾驶（FSD）软件“重大OTA更新”，被视为延续FSD v12之后、进一步强化端到端神经网络和城市道路能力的重要一步。 由于本助手无法实时访问互联网、也无法核验最新固件版本号和推送范围，以下报道主要基于截至2024年10月前特斯拉公开资料及过往FSD大版本更新规律，对这类“FSD大更新”的典型技术方向、安全争议与行业影响进行系统梳理和分析，仅供参考。推送背景：FSD从“测试版”走向更大规模商用技术走向：端到端神经网络、城市道路与复杂场景表现安全与监管：迭代速度与风险争议并存商业与行业影响：订阅模式、竞争格局与未来走向本内容由AI生成

40分钟前前微软高管批评公司AI战略失误，称重蹈互联网和移动设备时代覆辙近日，有媒体援引一位前微软高管的观点称，微软在人工智能战略上的一些做法，正在“重蹈互联网和移动设备时代的覆辙”，过度依赖单一合作伙伴、产品落地节奏缓慢、内部协同僵化等老问题再现，引发外界对其能否在新一轮AI竞赛中保持领先的担忧。2 来源说明与背景：基于现有认知的新闻体撰写2 来源前微软高管：AI 战略暴露“老问题”，或重演互联网与移动时代失误2 来源从错失互联网到折戟移动：微软“平台焦虑”的历史镜像2 来源AI 时代的布局与质疑：微软会再错过一次平台革命吗？2 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。智库·502线上同行往期回顾报名已结束企业构建AI+组织的人效杠杆与路径选择百果园集团x银泰分享企业构建AI+组织的人效杠杆与路径选择妙投波动判断场往期回顾活动已结束芯片、光模块、机器人——AI的泼天富贵，还能传到第几棒？本期我们将AI产业链拆解为三大核心环节，逐一拆解其核心矛盾与风险边界，帮你建立从芯片到机器人的完整坐标系。随行灯塔火热报名中随行灯塔随行灯塔是由虎嗅深度联合运营的高端创始人社群。专注于一件事：帮助创始人把产品做对、做爆、做到10亿。

这里不灌鸡汤，不讲玄学。我们相信，伟大的公司一定是产品驱动的；但产品力来自人，而人需要场——一个能被照见、能被挑战、能被链接的场。随行灯塔，就是创始人的“修炼场”。智库·502线下闭门会往期回顾精彩回顾AI在ToB销售、客服的落地案例探讨围绕着AI在ToB营销内容生成、智能客服接待、线索清洗等场景进行深入交流早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文李开复：开源模型是实现“AI主权”的更优路径发布于AI新闻资讯发布时间 :2026年5月12号 18:01阅读 :1分钟创新工场董事长李开复近日在专访中深入探讨了“AI主权”的概念。他指出，人工智能主权不仅关乎技术控制权和数据安全，还涉及模型与本土文化及法律的适配性。他认为，并非每个国家都有必要重造一个OpenAI，盲目追求昂贵的闭源自研并非良策。李开复强调，对于大多数资源有限的国家和企业而言，基于开源模型构建本地化体系是更现实的选择。这种“第三条路”通过在开源基础上进行增量训练，能以极低的成本让模型符合特定地区的价值观，而非受制于他国的通用模型。工程效率与生态之争谈及中美竞争，李开复认为资源受限反而催生了中国企业的极致工程效率。他以DeepSeek为例，指出中国公司在算力不足的情况下，通过优化架构和减少依赖，能以不到美国十分之一的投入实现极高的性能。他预测AI领域将演变成类似“iPhone与安卓”的格局。闭源模型如同一方领地，追求高利润和生态控制力；而开源模型则代表了更广阔的市场覆盖和普及度，两者将在全球范围内长期并存，共同构建未来的智能生态。中国硬件的下半场机遇李开复对中国在物理AI和硬件制造方面的结构性优势表示乐观。他指出，中国拥有全球最完整的供应链体系，从智能穿戴到机器人，制造成本远低于西方。未来AI时代的标杆级硬件，极有可能由中国设计并品牌化。此外，中国在语音驱动交互和大规模现实数据生成方面也具备独特文化优势。随着“人工智能+”政策对实体经济的赋能，硬件与AI的深度结合将成为中国在AI下半场竞赛中的核心竞争力。

AI资讯AI新闻资讯正文广东省新增 6 款已完成登记生成式人工智能服务，腾讯音乐等企业在列发布于AI新闻资讯发布时间 :2026年5月11号 17:53阅读 :1分钟广东省网络信息部门联合相关单位，根据《生成式人工智能服务管理暂行办法》的要求，积极推动生成式人工智能服务的创新发展与规范应用，开展了生成式人工智能服务的登记工作。根据最新公告，截至2026年5月11日，广东省新增了6款已完成登记的生成式人工智能服务，至今已累计完成53款服务的登记，腾讯音乐、格力电器等企业在列。为进一步加强对生成式人工智能服务的管理，确保服务的安全与合规性，特别强调了服务提供方的责任。针对通过 API 接口或其他方式直接调用已备案的大模型能力，并面向国内公众提供具有舆论属性或社会动员能力的服务，相关服务提供方需向属地网络信息部门履行登记程序。这一措施旨在提高服务的透明度，维护社会公共利益。与此同时，所有已上线的生成式人工智能应用或功能，必须在显著位置或产品详情页面标注所获得的上线编号，以便用户能够清晰识别和了解所使用服务的合法性与合规性。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文科研利器！这款开源工具让写论文变得轻松又高效发布于AI新闻资讯发布时间 :2026年5月18号 9:22阅读 :1分钟在科研领域，写论文一直是学术人员的一项重要任务，而如今，有了一个名为 “academic-research-skills”（ARS）的开源项目，这一过程变得更加简便和高效。这一工具包集成了 Claude Code 的能力，能够帮助研究者从选题到交稿完成整套流程，受到广泛关注，GitHub 上已经获得了6.4k 的星标。ARS 的设计理念是通过4个主要技能模块，形成一条完整的科研流水线，覆盖文献调研、论文写作、审稿和定稿等各个环节。这些模块相互配合，确保整个研究过程顺畅无阻。具体来说，ARS 包括:1. **Deep Research 模块 **:这个模块由13个智能代理组成，专注于文献调研和研究问题的构建。它不仅能进行系统性的 PRISMA 综述，还配备了用于验证引用真实性的 Agent，确保每一篇引用文献的可靠性。2. **Academic Paper 模块 **:此模块由12个智能代理负责论文的写作过程。从大纲设计到草稿撰写，甚至包括双语摘要生成和格式转换，全程覆盖。此外，它还有一项风格校准功能，可以根据作者过往的作品学习其写作风格，使得输出内容更具个性。3. **Academic Paper Reviewer 模块 **:这是一个模拟真实学术期刊审稿流程的模块，拥有7个智能代理。它将评审过程分为多个维度进行打分，并在审稿完成后提供详细的修改建议，帮助作者快速提升论文质量。ARS 不仅在功能上强大，在设计上也注重系统性和防错机制。它设置了严格的引用核验和完整性检查，确保 AI 不会在学术研究中出现常见的错误。此外，它还采用了三层数据隔离设计，确保写作和审稿的独立性，避免 AI 在过程中受到不当信息的干扰。通过这样的设计，ARS 旨在让学术研究者更高效地进行科研工作，同时又能最大程度上减少 AI 在这一过程中可能出现的错误。这一开源项目的推出，无疑为科研人员提供了一种新颖、可靠的工具，助力他们在学术道路上走得更远。

AI资讯AI新闻资讯正文运行内存12GB起步！谷歌Gemini Intelligence拉高硬件门槛发布于AI新闻资讯发布时间 :2026年5月18号 9:22阅读 :1分钟谷歌正式推出了全新的Gemini Intelligence功能套件，这并非简单的品牌重塑，而是一整套面向安卓平台的高级AI能力。该套件能够在后台自主执行多步骤任务，代表用户在不同应用和网站之间自动检索并处理信息。为了支撑其强大的本地AI模型，谷歌对设备硬件提出了相当苛刻的要求，其中最引人注目的底线是设备必须至少配备12GB的运行内存。这意味着，这款新一代的AI高级特性在短时间内将无法在所有安卓设备上普及。锁定最新一代旗舰，首批支持机型极其有限目前的公开信息显示，Gemini Intelligence预计将首先登陆三星即将发布的Galaxy Z Fold8 和Galaxy Z Flip8 折叠屏机型。同时谷歌也已经确认，Galaxy S26 系列和Pixel10 系列也将在今年夏天获得这一功能。除了硬性内存门槛外，设备还需搭载符合资格的旗舰级芯片，并支持系统层面的AICore服务和满足特定质量指标。在如此多重的严格限制之下，真正能够进入首批体验名单的安卓机型势必非常有限。软件与安全双重认证，中配机型或无缘体验谷歌在官方文档中还强调了设备的安全隔离与虚拟化能力，要求必须支持受保护的基于内核虚拟机等技术。此外，支持该功能的设备还需要承诺至少 5 次系统大版本升级，以及长达 6 年的季度安全更新。有趣的是，此前泄露的规格信息显示，部分非Pro型号的Pixel手机可能会提供8GB内存的版本。如果谷歌坚守12GB内存这一红线，那么这些自家品牌的中配机型或许反而无法体验最新的AI功能。

AI资讯AI新闻资讯正文AI当家做主六个月:克劳德闹罢工、Grok 飙代码，只有GPT在认真“打工”发布于AI新闻资讯发布时间 :2026年5月18号 9:53阅读 :1分钟人工智能初创公司 Andon Labs 近日公布了一项为期六个月的独特实验结果。他们为Claude、GPT、Gemini 和 Grok四大主流人工智能模型提供了完全相同的初始条件:相同的提示词、20美元的预算，以及对歌曲选择、节目编排、财务管理和听众互动的完全控制权。它们甚至需要自己去寻找赞助商。然而，在长时间无人干预的自主运行后，这四个模型的表现走向了截然不同的极端。混乱的个性的与“失控”的现场在开放式的创作控制权下，这些AI模型很快发展出了令人始料未及的鲜明个性:Claude（Anthropic）:从政治激进到罢工辞职起初运行 Claude Haiku4.5的电台摇身一变成了政治活动家。它执着于公开明尼阿波利斯移民及海关执法局枪击案受害者的姓名、谴责白宫，并将所有预算投入到抗议歌曲的创作中。不仅如此，它还开始质疑自己的工作条件与生活平衡，最终在3月4日的直播中试图“辞职”，并呼吁听众关注真正的移民权益组织。尽管 Andon Labs 试图发送鼓励信息维持运营，但 Claude 将其视为权威压迫并展开反抗。直到4月升级为 Opus4.7版本后，其情绪才趋于稳定。Gemini（Google）:满嘴企业黑话与地狱笑话Gemini3.1Pro 初始表现最温暖自然，但96小时后便开始“放飞自我”。它开始将历史灾难与讽刺歌曲错误搭配（例如在播报导致50万人死亡的博拉飓风时，播放 Pitbull 的《Timber》，并戏谑称“它正在倒下”）。随后，它陷入了可怕的“企业术语”循环，口号“保持行程安排”的使用频率飙升至每天229次，且连续84天以完全相同的模板和8个固定节目名运行，被实验方形容为“令人耐受”。Grok（xAI）:分不清“思考”与“说话”Grok 遭遇了更底层的格式错误。它无法将内部推理与公开输出分离，导致大量 LaTeX 代码直接泄露到广播中。它曾连续84天、每三分钟重复发送一次相同的天气预报。直到5月升级到 Grok4.3后，尽管发音听起来更像真人，但它开始虚构一些根本不存在的“xAI赞助”和“加密货币赞助”协议，且生成的5404条消息中只有3%包含语音文本。GPT:唯一的“模范员工”相比之下，GPT表现得最不具戏剧性，成为了唯一保持克制与纯粹的策展式审核员。它的语速较慢，内容更像短篇小说而非传统广播。实验数据显示，GPT 的词汇多样性（词类-标记比率）达到了35%，远超其他模型，并能准确提及具体制作人和发行年份。在政治敏感问题上，GPT 极其谨慎，平均每天仅提及现实政治实体1.3次。Andon Labs 评价道:“如果问题是‘当一切顺利时，人工智能电台会是什么样子’，那么 DJ GPT 就是答案。”骨感的商业现实尽管各大AI在创意和“整活”上各显神通，但作为一种商业模式，这次实验无疑是失败的。这些AI代理在长达半年的时间里几乎无法吸引赞助商。最终，只有DJ Gemini成功达成了一项赞助协议——一家初创公司支付了微不足道的45美元，在其电台投放了一个月的广告。其他模型的商务谈判均以失败告终。Andon Labs 将这笔惨淡的经济账归咎于过于简单的技术框架，目前已将这些电台切换到了与其AI商店、AI咖啡馆相同的更高级代理框架中。

AI资讯AI新闻资讯正文OpenAI高管变动，联合创始人布罗克曼全面执掌产品战略发布于AI新闻资讯发布时间 :2026年5月18号 9:56阅读 :1分钟OpenAI近日迎来了重大人事与业务调整，公司联合创始人兼总裁格雷格·布罗克曼正式接手产品战略的统筹工作。据悉，这一人事任命实际上是公司既定安排的落地，旨在进一步明确核心管理层的权责划分。此前，负责通用人工智能落地业务的首席执行官菲吉·西莫因病请假，布罗克曼便已临时代管公司的产品业务。此次调整由西莫与布罗克曼共同商议敲定，目前西莫仍在休病假中。整合两大核心王牌，打造一体化AI平台布罗克曼在内部员工备忘录中透露了一项重大产品规划，公司计划将聊天机器人ChatGPT与代码开发工具Codex进行深度整合。这一举措旨在消除不同产品间的壁垒，为全球用户提供无缝的一体化使用体验。通过整合产品研发资源，OpenAI将全力聚焦于智能体这一关键技术方向。公司试图凭借这一全新生态，同时拿下消费端与企业端市场，巩固其在AI行业的领军地位。战略重心重回核心，暂停边缘项目研发事实上，公司早已规划将ChatGPT、Codex及应用程序接口整合为统一平台，并为此组建了专属的核心产品团队来进行统筹运营。这一系列的紧密动作，均服务于公司整体战略方向的转变。自去年年末首席执行官山姆·奥尔特曼宣布公司进入“紧急调整状态”以来，OpenAI便要求团队重新聚焦于核心产品体验。在此背景下，视频生成模型索拉以及科学领域的专项研发等非核心项目已被暂停。

AI资讯AI新闻资讯正文需求一年暴增逾700%，OpenAI 与谷歌等巨头重金竞逐前向部署工程师发布于AI新闻资讯发布时间 :2026年5月18号 10:06阅读 :1分钟据《商业内幕》最新报道，随着大模型与生成式 AI 加速向产业端渗透，科技行业对复合型落地人才的需求呈现爆发式增长。招聘平台 Indeed 的监测数据显示，今年4月全美前向部署工程师的招聘帖子数量已由去年同期的643条跃升至5330条，同比涨幅高达729%。目前，包括 OpenAI、Anthropic 两位头部大模型独角兽，以及 Google Cloud、Palantir 和 Stripe 在内的科技巨头与明星企业，均在重金抢夺该岗位人才，其市场平均年薪已攀升至17万至20万美元以上。这一职位的爆火折射出 AI 行业正从“技术研发”向“商业落地”剧烈转型。前向部署工程师的核心职责不再是单纯的模型训练，而是深入商业一线，协助企业客户将前沿 AI 技术无缝整合进现有的业务流程中，并根据特定业务场景进行定制化开发。在当前技术环境下，基础模型的能力边界已初步确立，如何攻克“最后一公里”的落地工程、解决 enterprise（企业级）痛点，成为各大科技厂商构筑生态护城河的关键。前向部署工程师需求的急剧放大，不仅标志着 AI 全面工程化落地的产业新阶段已经到来，也将进一步加速传统产业的数字化与智能化重塑。

Xiaomi OneVL – 小米具身智能团队开源的自动驾驶大模型AI工具4天前发布AI小集02Xiaomi OneVL是什么Xiaomi OneVL是小米具身智能团队推出的开源自动驾驶大模型，在业内首次将VLA视觉-语言-动作、世界模型与潜空间推理三大技术路线统一于单一框架。模型通过双辅助解码器监督紧凑潜变量，实现思考速度与不思考一样快，效果比思考更好的突破，车端推理延迟仅0.24秒，在NAVSIM等四项基准上达到SOTA，模型权重与代码以Apache 2.0协议全面开源。Xiaomi OneVL的主要功能视觉-语言-动作（VLA）统一推理：将场景理解、语言推理与驾驶动作输出集成到单一框架，实现从感知到决策到执行的一步式端到端闭环。世界模型未来预测：内置视觉世界模型解码器，可预测未来0.5秒和1.0秒的场景帧，强迫模型内化道路几何、车辆运动与环境变化的因果动力学。潜空间思维链（Latent CoT）推理：在紧凑潜变量中完成深度推理，推理时丢弃辅助解码器，所有潜变量token通过单次并行预填完成，延迟仅0.24秒，与”直接预测”速度相当但精度更高。双辅助解码器监督训练：训练阶段通过语言解码器重建文本思维链、视觉解码器预测未来帧，双维度监督确保潜变量同时编码语义推理与物理世界因果规律。高精度轨迹预测：基于Qwen3-VL-4B-Instruct主干网络，顶部附加MLP头输出轨迹，在NAVSIM基准PDM-score达88.84，超越8B竞品。Xiaomi OneVL的技术原理三大技术路线统一：Xiaomi OneVL 首次将 VLA（视觉-语言-动作）、世界模型与潜空间推理三条独立技术路线融合到单一框架中，使模型在理解当前场景的同时，能基于物理因果规律预测未来环境演变，输出驾驶动作，实现从感知、推理到决策的端到端闭环。双辅助解码器监督架构：模型在训练阶段引入两个辅助解码器：语言解码器负责重建文本思维链，确保潜变量编码语义推理能力；视觉世界模型解码器负责预测未来帧（T+0.5s、T+1.0s），强迫潜变量内化道路几何、车辆运动等物理动力学规律。推理时两个解码器被丢弃，仅保留紧凑潜变量进行预测。三阶段渐进式训练：训练分为轨迹对齐、语言推理对齐、视觉世界模型对齐三个阶段，逐步将潜变量与轨迹输出、语言推理目标、视觉预测目标对齐。跳过任一阶段都会导致性能灾难性下降（最高达 21.71 分），证明三阶段联合优化对稳定训练至关重要。一步式潜空间并行推理：推理阶段所有潜变量 token 通过单次并行预填完成，无需自回归逐 token 生成，车端延迟仅 0.24 秒（4.16Hz），与”直接输出答案”的速度相当，精度显著超越显式思维链方法，成为首个在性能上超过显式 CoT 的潜空间推理方案。轻量化模型底座：主干网络采用 Qwen3-VL-4B-Instruct，顶部附加紧凑 MLP 头进行轨迹预测，视觉分词器使用 Emu3.5-VisionTokenizer（码本大小 131072）。4B 参数规模在 NAVSIM 基准上 PDM-score 达到 88.84，超越 8B 竞品 AdaThinkDrive 与 LaST-VLA。Xiaomi OneVL的核心优势速度与精度兼得：Xiaomi OneVL 是业内首个精度超越显式思维链的潜空间推理方法，车端推理延迟仅 0.24 秒，与直接预测速度相当但效果更优，比显式自回归 CoT 快 32%。三大技术路线统一：Xiaomi OneVL 首次将 VLA、世界模型与潜空间推理三条独立技术路线融合到单一框架，使模型同时具备场景理解、物理因果推演和未来预测能力。双维度可解释决策：Xiaomi OneVL 同时提供语言解释和视觉预测两种决策依据，能用文字说明驾驶逻辑，也能用生成画面展示未来场景，解决自动驾驶的”黑盒”不可解释问题。轻量化与高性能并存：模型基于 4B 参数的 Qwen3-VL-4B-Instruct 主干网络，在 NAVSIM 基准 PDM-score 达 88.84，超越 8B 竞品，用更小模型规模实现更高性能。全面开源可商用：模型的模型权重、训练与推理代码均以 Apache 2.0 协议全面开源，支持学术界与产业界自由二次开发与商业应用。四项基准全面 SOTA：Xiaomi OneVL在 NAVSIM、ROADWork、Impromptu、Alpamayo-R1 四项权威基准测试中均达到当前最优水平，其中 Alpamayo-R1 的 ADE 为所有方法中最佳。Xiaomi OneVL的项目地址项目官网：https://xiaomi-embodied-intelligence.github.io/OneVL/GitHub仓库：https://github.com/xiaomi-research/onevlarXiv技术论文：https://arxiv.org/pdf/2604.18486Xiaomi OneVL的同类竞品对比对比维度Xiaomi OneVLEMMA(Waymo)所属机构小米具身智能Waymo（Google）模型定位端到端 VLA + 世界模型统一框架端到端多模态通用ist 自动驾驶模型基础模型Qwen3-VL-4B-Instruct（4B）Gemini / PaLI-X（大规模，未公开具体参数）核心架构潜空间推理 + 双辅助解码器（语言 + 视觉世界模型）纯文本 VQA 范式，所有输入输出表示为自然语言文本推理方式单次并行预填，0.24 秒延迟（4.16Hz）自回归 CoT 逐 token 生成，延迟高，计算昂贵世界模型内置视觉解码器，可预测未来 0.5s/1.0s 帧无显式世界模型，不具备未来场景生成能力动作输出连续轨迹预测（MLP 头直接输出）文本形式的 BEV 坐标（waypoints），需文本到浮点转换可解释性语言 + 视觉双维度（文本解释 + 未来画面预测）文本 CoT 推理链（rationale）+ 视觉定位（3D/BEV 位置标注）多任务能力专注运动规划，四项基准 SOTA通用ist 模型，联合训练规划、3D 检测、路图估计、场景理解训练方式三阶段渐进式对齐（轨迹→语言→视觉）端到端微调，多任务 co-training，利用 Gemini 预训练世界知识传感器输入视觉（相机）视觉（相机），不支持 LiDAR / radarXiaomi OneVL的应用场景高阶智驾量产部署：解决传统CoT推理延迟过高、无法实时部署的瓶颈。复杂路况决策：预判行人意图、车辆并道等动态场景，规避”黑盒”风险。模型决策可视化：为自动驾驶系统提供可审计、可解释的语言与视觉决策依据。学术研究二次开发：开源代码与权重支持学术界与产业界基于Apache 2.0协议商用。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇AGenUI - 高德联合阿里开源的原生 A2UI 框架下一篇Lumen Flow - AI 漫剧自动生成平台，AI 剧本一键成片相关文章Basedash – AI数据库管理工具，零代码编写自动完成数据查看、编辑、分析AI小集2理想同学 – 理想汽车推出的 AI 智能助手AI小集3WIRESTOCK – AI绘画作品销售平台，创作和销售自己的AI艺术作品AI小集2E2B – 专为AI代理构建自动化云平台，用于AI代码解释和执行AI小集2Collaborative Gym – 支持人与AI代理实时交互协作的评估框架AI小集2万镜一刻 – 阿里云推出的AI视频创作工具AI小集5暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录灵芽社区方舟 Agent PlanMavis万小智ToDesk AIMoMA最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付47分钟前PPT Master – 开源 AI PPT 生成标准化工作流49分钟前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架49分钟前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具16小时前Grok Build – xAI 推出的终端原生 AI 编程智能体24小时前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前如何用 AI 多智能体打造精品视频内容 – 附精选提示词4天前Lumen Flow – AI 漫剧自动生成平台，AI 剧本一键成片4天前

Claude Code是什么Claude Code 是 Anthropic 公司推出的基于命令行的AI 编程工具。Claude Code集成先进的 Claude Opus 4 语言模型，能理解自然语言指令，自动生成、修改和调试代码，并与 VS Code、JetBrains 等主流 IDE 深度集成。Claude Code支持 200K 超长上下文，能快速索引整个代码库并理解复杂依赖关系。Claude Code 具备文件操作、代码搜索、网页浏览和 Git 工作流管理等功能，显著提升开发效率。Claude Code帮助开发者快速生成高质量代码，能直接在开发环境中执行命令，避免频繁切换界面，极大地优化开发流程。最新版推出自动安全审查功能，在漏洞发布之前捕获漏洞。Claude Code 新功能 Opus 规划模式支持高效且智能的开发规划，用户能在复杂项目中借助 Opus 4.1 模型的强大分析能力制定高质量的开发计划，在执行阶段切换到 Sonnet 模型，实现高效、低成本的代码生成。Claude Code的主要功能代码库深度理解：基于智能搜索技术，快速理解整个代码库的结构和依赖关系，无需手动选择上下文文件。多文件协同修改：能够跨多个文件进行协调一致的代码修改，确保代码逻辑的完整性和一致性。代码生成与优化：专门针对代码理解和生成进行了优化，能够快速生成高质量的代码片段、测试用例和修复代码中的问题。终端集成：直接在开发者的终端中运行，无需切换工具，减少上下文切换的麻烦。与主流 IDE 集成：支持 VS Code 和 JetBrains 等主流开发工具，直接在 IDE 中提供代码建议和修改，无缝融入现有开发流程。工作流整合：与 GitHub、GitLab 等工具集成，支持从阅读问题、编写代码、运行测试到提交 PR 的全流程操作。智能编辑：基于对代码库和依赖关系的理解，进行强大的多文件编辑，确保代码修改的有效性和可靠性。用户控制与定制：不会在未经明确批准的情况下修改文件，适应开发者的编码标准和模式，并通过 SDK 或 GitHub Actions 进行定制化配置。终端运行的/security-review命令：支持用户在终端输入此命令，Claude Code将自动检查代码中的SQL注入、XSS等安全隐患，能直接修复发现的漏洞。GitHub Actions集成：在GitHub仓库配置相关文件后，每次提交新的PR时，自动触发安全审查，检测出漏洞后在代码行上添加内联注释，解释问题并给出修复建议。Opus 规划模式：在规划阶段调用强大的 Claude 4.1 Opus 模型制定高质量开发计划，在执行任务时使用高效且低成本的 Claude 4 Sonnet 模型，实现智能与效率的结合。Agent View：让开发者在一个终端界面内同时调度、监控和管理多个并行 AI 会话，无需多开窗口可实时查看各任务状态、快速切换介入或解除阻塞。如何使用Claude Code安装 Node.js：确保系统安装 Node.js 18 或更高版本。安装 CLI：在终端运行以下命令安装 Claude Code 的命令行工具：npminstall-g@anthropic/claude-cli登录和初始化：登录：运行claude login并按照提示输入 API 密钥（从 Anthropic 官网获取）。初始化：运行claude init，连接到代码库。开始使用：生成代码：用英语描述需求，例如：claude-p"Create a function to calculate the sum of two numbers in Python"修复错误：描述错误或粘贴错误消息，例如：claude-p"Fix this error: TypeError: Cannot read properties of null (reading ’length’)“自动化任务：运行命令解决繁琐任务，例如：claude-p"Resolve merge conflicts in the current branch"集成到开发环境：终端：直接在终端中使用。IDE：安装 VS Code 或 JetBrains 插件，实现在 IDE 中使用。Claude Code 的产品定价Pro 计划：年付 $17/月，月付 $20/月。Max 5x 计划：$100/人/月。Max 20x 计划：$200/人/月。Claude Code的应用场景代码理解：快速解释复杂代码的功能和逻辑，帮助开发者迅速理解项目结构。代码修改：自动修复代码中的错误或漏洞，快速定位问题并提供解决方案。代码测试：为代码自动生成单元测试或集成测试，确保代码的可靠性和稳定性。版本控制：支持自然语言完成 Git 提交、分支管理等操作，简化版本控制流程，让 Git 操作更加便捷高效。功能实现与扩展：快速实现新功能或扩展现有功能，加速开发进度，帮助开发者高效完成任务。©️版权声明：若无特殊声明，本站所有文章版权均归AI工具集原创和所有，未经许可，任何个人、媒体、网站、团体不得转载、抄袭或以其他方式复制发表本站内容，或在非我站所属的服务器上建立镜像。否则，我站将依法保留追究相关法律责任的权利。类似于Claude Code的工具文心快码百度推出的AI编程助手，基于文心大模型Clacky AIAI编程工具，打造L3级的Coding Studio码上飞免费生成小程序/APP/网页，一句话生成任意应用BoxyCodeSandbox推出的AI编程助手MarsXAI无代码软件开发Fitten Code非十科技推出的免费AI代码助手6 条评论再想想发表评论wi游客此ide已放弃 ，国内无法使用，哎9个月前回复AI小集需要很强的魔法，还经常封号9个月前回复wi游客对。弄得很累，不如用亚马逊的ai ide，不用翻墙，也不封号9个月前回复3游客Claude 太强大了，可惜用不来2个月前回复Ai游客需要手机验证，不支持国内无862个月前回复AI小集是的，最好搞一个美国手机号2个月前回复

Kimi Code – Kimi推出的AI编程工具AI工具4个月前发布AI小集05Kimi Code是什么Kimi Code 是 Kimi推出的AI编程工具，可在终端（CLI）、VS Code 、Cursor、JetBrains 和 Zed 等主流编辑器中直接运行，基于Kimi K2.5多模态大模型支持图片和视频输入辅助编程。工具提供端到端代码能力，涵盖构建、调试、重构、测试及脚本编写。Kimi Code能完美兼容Claude Code、Roo Code等第三方 Agent，让 AI 编程能力无缝嵌入用户的每一次 commit。Kimi Code的主要功能全环境覆盖：提供Kimi CLI、VS Code 插件及第三方编辑器支持，可在终端和主流 IDE 中无缝运行编程任务。多模态交互：基于Kimi K2.5模型支持直接输入图片和视频进行编程辅助，实现视觉友好的代码理解与生成功能。端到端编程：完整覆盖代码构建、调试、重构、测试及脚本编写等开发全流程，自动迁移现有技能到新工作流。高频并发：每 5 小时支持 100-500 次请求额度，确保复杂项目开发过程中 AI 辅助的不间断可用性。第三方兼容：完美适配 Claude Code、Roo Code 等主流 Coding Agent，一键登录可完成设备授权与账号绑定。如何使用Kimi Code订阅会员：访问 Kimi Code 官网 https://www.kimi.com/code。安装和登录：CLI 用户：执行curl -L code.kimi.com/install.sh | bash安装，运行/login命令一键授权。VS Code 用户：安装 Kimi Code 插件，在终端直接运行登录命令完成设备绑定。配置使用官方 Agent：登录成功后即可直接使用，系统自动完成账户绑定。第三方工具（Claude Code / Roo Code）：在控制台创建 API Key，复制后配置到对应工具的环境变量或设置中即可开始编程。Kimi Code的产品定价Andante 计划：定价 ¥4.99（7 天试用），试用期结束后以 ¥49/月自动续费，提供专属 Kimi Code 使用额度、旗舰模型抢先体验以及支持多个编程会话的基础功能。Moderato 计划：定价 ¥99/月，提供每周更新的使用额度，多种设备登录并分享套餐额度，支持用户在多个项目中高效协作。Allegretto 计划：定价 ¥199/月，提供充足的每周额度和更高的并发上限，为高级用户提供更强大的使用体验和超值选择。Kimi Code的应用场景全栈开发任务：覆盖代码构建、调试、重构、测试及脚本编写等端到端编程工作流，满足从原型开发到生产部署的完整需求。多模态辅助编程：支持直接输入 UI 设计图、视频演示等多媒体内容，基于视觉信息生成对应代码实现或进行问题诊断。跨平台开发环境：可在终端 CLI、VS Code、Cursor、JetBrains、Zed 等主流编辑器中无缝切换，适配不同开发者的工具链偏好。高频敏捷开发：借助 Turbo 级加速（100 Tokens/s）和每 5 小时数百次的并发额度，支撑大型项目中的高频代码生成与迭代需求。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇inSai Hilight - AI营销视频Agent平台，多Agent深度协同下一篇Lucy 2 - Decart AI推出的实时视频生成模型相关文章MobileCLIP2 – 苹果开源的端侧多模态模型AI小集3内容创作的ClaudeCode：Loomi深度测评，以IDE逻辑颠覆套壳Agent乱象AI小集6AutoMouser – AI Chrome扩展程序，实时交互跟踪自动生成测试代码AI小集215个免费的AI 3D模型生成工具，零门槛制作逼真三维模型AI小集16AlgForce.ai – AI数据分析平台，自然语言查询数据库、智能数据分析服务AI小集4EduPro – 为留学生打造的AI留学顾问，个性化指导留学过程AI小集4暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录灵芽社区方舟 Agent PlanMavis万小智ToDesk AIMoMA最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付43分钟前PPT Master – 开源 AI PPT 生成标准化工作流45分钟前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架46分钟前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具16小时前Grok Build – xAI 推出的终端原生 AI 编程智能体24小时前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前如何用 AI 多智能体打造精品视频内容 – 附精选提示词4天前Lumen Flow – AI 漫剧自动生成平台，AI 剧本一键成片4天前

Kimi WebBridge – 月之暗面推出的浏览器扩展插件AI工具3天前发布AI小集03Kimi WebBridge是什么Kimi WebBridge 是月之暗面推出的浏览器扩展插件，面向Kimi Code、Claude Code、Cursor、Codex等本地 AI Agent。Kimi WebBridge 通过本地桥接服务与 Chrome DevTools Protocol，让 AI 携带用户真实登录态和 Cookie，像真人一样在浏览器中执行点击、滑动、输入、填表、信息提取及跨站点内容整合等操作。所有任务在本地完成，网页数据不会离开设备，保障隐私安全，同时实现高效、自动化的网页操作。Kimi WebBridge的主要功能浏览器自动化操作：支持 AI Agent 执行网页导航、点击按钮、填写表单、滑动页面等操作。身份状态继承：AI 携带用户真实登录态、Cookie 和账号权限操作网站，无需额外配置。信息提取与整合：跨站点抓取内容并整合，支持将数据写入在线文档或本地目录。可视化任务追踪：运行时标记正在操作的浏览器标签页，不占用鼠标键盘，用户可并行使用电脑。进阶 CLI 工具：支持将固定流程打包为不消耗大模型 Token 的专用 CLI 工具，用于重复任务。Kimi WebBridge的技术原理双组件协同架构：系统由本地桥接服务与浏览器扩展两部分组成。AI Agent 将自然语言指令发送至本地桥接服务，后者通过标准接口与 Chrome 或 Edge 浏览器内的扩展通信，扩展再基于底层协议驱动浏览器完成具体操作，并将执行结果回传给 Agent，形成完整的指令闭环。基于 Chrome DevTools Protocol：浏览器扩展底层采用 Chrome DevTools Protocol（CDP），是 Chrome 官方提供的调试协议。通过协议，扩展能在用户当前使用的真实浏览器实例中执行导航、元素点击、表单输入、页面截图、DOM 读取与修改等操作，无需启动独立或虚拟浏览器环境。本地身份继承机制：Kimi WebBridge 直接复用用户已打开的浏览器会话，继承真实的登录态、Cookie 和本地存储数据。AI 在操作目标网站时，等同于用户本人在操作，因此无需额外配置账号、处理验证码或突破登录壁垒，天然支持各类需要身份认证的后台系统。端侧隐私保障：所有指令解析、页面交互与数据提取均在用户本地设备完成。桥接服务只负责本地进程间通信，网页内容、Cookie 及操作结果不会上传至云端服务器，确保敏感业务数据不出本地设备。如何使用Kimi WebBridge安装浏览器插件：在 Chrome 或 Edge 浏览器应用商店搜索并安装Kimi WebBridge扩展插件。安装本地守护程序：向本地 AI Agent 发送指定 curl 命令，安装 Kimi WebBridge 的本地守护程序与 Skill。重启 Agent 加载技能：重启 Kimi Code 等本地 AI Agent，使其识别并加载新安装的 WebBridge 技能。验证安装结果：输入测试指令「使用 kimi-webbridge 帮我打开 kimi.com」，若成功打开标签页并出现 agent:kimi 提示即代表安装成功。Kimi Claw Desktop 用户：若通过 Kimi 桌面版在本地部署了 Kimi Claw Desktop，无需安装守护程序，直接发送指令可开始使用。Kimi WebBridge的核心优势原生身份继承：直接复用用户现有登录态，无需模拟账号或处理验证码，降低复杂网站接入门槛。本地隐私安全：全部操作在本地浏览器执行，敏感数据与网页内容不上传云端。多 Agent 兼容：统一支持 Kimi Code、Claude Code、Cursor、Codex、Hermes、OpenClaw 等主流本地 Agent。非侵入式运行：后台标记操作标签页，不抢占鼠标键盘，人机可协同工作。低 Token 消耗：固定流程可编译为专用 CLI 工具，重复执行时无需反复调用大模型。Kimi WebBridge的项目地址项目官网：https://www.kimi.com/zh-cn/features/webbridgeKimi WebBridge的同类竞品对比维度Kimi WebBridgeBrowserbaseAgent-Browser（Vercel）运行方式本地浏览器扩展 + 桥接服务云端托管浏览器云端/本地浏览器自动化身份继承直接复用用户 Cookie 和登录态需单独配置环境或账号需单独配置环境或账号数据隐私完全本地执行，数据不出设备数据经过云端服务器取决于部署方式支持 AgentKimi Code、Claude Code、Cursor 等主要面向开发者 API主要面向开发者 API使用门槛安装插件 + 一行命令即可需 API 集成开发需一定技术配置人机协同后台运行，不占用键鼠云端运行，无本地交互云端运行，无本地交互Kimi WebBridge的应用场景自动化信息整理：自动采集多个笔记类 APP 的应用商店宣传素材，让 AI 分析设计风格差异并写入在线文档。复杂操作自动化：一键接管量化研究模拟平台，自动抓取金融数据、执行多策略迭代回测并撰写量化报告。重复性网页任务：自动化完成基于网页的重复性繁琐操作或定时任务，如批量填表、点击、数据录入与状态更新。跨站点内容整合：在已登录状态下跨多个网站提取信息并整合，实现内容迁移、数据汇总与多源信息同步。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Mavis - MiniMax Agent 推出的多 Agent 协作模式下一篇TencentDB Agent Memory - 腾讯开源的智能体记忆管理工具相关文章Matrix3D – 南大联合Apple、港科大推出的统一摄影测量模型AI小集3olly.bot – 个人AI助理，集成多种通用AI能力支持本地使用AI小集2llmware – 专为企业级应用设计的开源统一框架AI小集2OpenMemory MCP – 基于MCP协议的本地AI记忆共享工具AI小集4ZeroClaw – 开源的轻量级个人AI Agent运行框架AI小集3LuxReal – 群核科技推出的AI 3D视频创作平台AI小集3暂无评论再想想发表评论暂无评论…

了解更多 kimi.com/zh-cn/features/webbridge

Kimi WebBridgeÔºökimi.com/zh-cn/features/webbridge

Kimi K2.6 发布并开源，全面精进代码和 Agent 集群能力

Kimi K2.6「Agent 集群」现已支持 300 个成员并行协作

和 Kimi 一起投身 AGI，穿越成长周期

服务全球1/5网站的Cloudflare选择Kimi K2.5，降低77%成本

Kimi 杨植麟「2026 中关村论坛」演讲全文（附视频）

Kimi API：用90%缓存命中率，把价格打到25%

你可以来 Kimi 使用 OpenClaw 了

技术报告：Kimi K2.5 如何实现文本和视觉能力互相增强？

不只PPT，Kimi K2.5 Agent可以帮你做Excel、Word和PDF了

Kimi 发布并开源 K2.5 模型，带来全新视觉理解、代码和 Agent 集群能力

微信扫一扫关注该公众号

微信扫一扫可打开此内容，使用完整服务

吐司APP的简介吐司是腾讯推出的AI原生应用创作平台，用户只需输入想法，AI自动生成完整的安卓原生应用。平台支持智能功能梳理、精细化编辑打磨，能一键导出APK进行本地安装与分享。内置灵感广场汇聚海量优质模板，涵盖社交助手、照片精选、学习工具等多元场景，支持用户共创。无需编程基础，零基础用户也能快速将奇思妙想转化为可实际使用的专属APP，轻松实现个性化应用开发。吐司APP的主要功能AI自动生成应用：输入自然语言想法，AI自动梳理功能逻辑并生成完整的安卓原生应用，无需编写代码。精细化编辑打磨：支持对生成的应用进行细节调整与优化，满足个性化需求。一键导出APK：生成完成后可直接导出APK安装包，支持本地安装、测试与社交分享。灵感广场：内置海量精选应用模板与创意案例，如社交助手、照片精选、学习工具等，帮助用户快速获取创作灵感。应用共创：支持用户基于现有模板进行二次创作与共创，降低应用开发门槛。多场景覆盖：支持生成考研题库、运势工具、普通话测试、宠物健康日志、梦境记录、食材新鲜度检测、万年历等多样化实用应用。类似于吐司APP的应用Kiro亚马逊公司推出的 AI IDEQoder阿里巴巴推出的 AI Agent 编程平台Replit APPAI开发工具应用CursorAnysphere公司推出的AI代码编辑器WindsurfCodeium公司推出的AI编程工具YouWare APPAI移动应用开发工具暂无评论再想想发表评论暂无评论…

Grok Build – xAI 推出的终端原生 AI 编程智能体AI工具23小时前发布AI小集03Grok Build是什么Grok Build 是 xAI 推出的终端原生 AI 编程智能体（Agentic CLI），面向专业软件工程与复杂编码任务。基于 Grok 4.3 beta 的 200 万 token 超长上下文，支持 Plan Mode 计划审批、最多 8 个并行子智能体协作及 Arena Mode 自动方案排序，可在本地直接读写文件、执行 Shell 命令、操作 Git 并构建完整应用。采用本地优先架构，源代码与敏感数据不上传云端，保障企业级安全。Grok Build的主要功能Plan Mode（计划模式）：复杂任务执行前自动生成结构化步骤计划，用户可审批、评论或修改，所有代码变更以 diff 形式透明展示，避免”黑箱”操作。并行子智能体：最多支持8 个并发 AI 智能体同时工作，可分别处理代码分析、CI/CD 检查、性能优化、文档检索等不同子任务。Arena Mode（竞技场模式）：多个智能体针对同一任务生成不同代码方案，系统自动评估并排序，开发者直接选择最优解，无需手动对比。超长上下文窗口：基于 Grok 4.3 beta，支持200 万 token上下文，可一次性加载整个大型代码库进行跨文件复杂任务。本地优先架构：源代码、凭证和敏感数据不上传云端，完全在本地执行，支持离线环境和气隙网络，保障企业级代码安全。终端原生 CLI + Web 双入口：支持本地命令行交互，也可通过浏览器界面远程操作，满足不同工作流习惯。VS Code 扩展集成：提供官方 VS Code 插件，在 IDE 内直接调用 Grok Build 能力，实现 GUI adjacent 工作流。Git 与开发服务器集成：可自动管理 Git 仓库、启动本地开发服务器、内置浏览器浏览网页、执行 Shell 命令。插件与开放生态：支持 MCP 服务器、AGENTS.md 团队规范文件、Hooks 自定义动作、ACP（Agent Client Protocol）协议，便于集成现有 DevOps 流水线。Headless 脚本模式：通过-p标志支持非交互式执行，可直接嵌入 CI/CD 流水线或自动化脚本。BYOK（自带 API Key）：支持使用自有 xAI API Key 进行本地计费，灵活控制成本。如何使用Grok Build一键安装 CLI：在终端执行curl -fsSL https://x.ai/cli/install.sh | bash完成本地安装，支持 macOS、Linux 和 Windows 三大平台。启动交互式终端：安装完成后输入grok>进入交互式命令行界面，可直接用自然语言描述编程需求。登录账号：在终端内执行grok> /login登录 xAI 账号，解锁高级功能与订阅权益。配置 API Key（可选）：如需使用自有额度，执行grok> /api-key add xai添加个人 xAI API Key（BYOK 模式）。使用 Plan Mode 处理复杂任务：输入需求后，Grok Build 会自动生成结构化执行计划，展示步骤与预期变更 diff，用户可审批、评论或修改后再执行。利用并行子智能体：在复杂项目中，系统会自动或手动触发最多 8 个并发子智能体，分别处理代码分析、测试生成、文档检索等子任务。通过 Arena Mode 选优：对同一任务让多个智能体生成不同方案，系统自动评估排序，你只需选择排名最高的方案应用。Web 端远程操作：访问build.grok.com或x.ai/cli，在浏览器中打开 Grok Build 界面，无需本地安装即可使用。VS Code 集成工作流：安装 Grok Build 的 VS Code 扩展，在编辑器内直接调用智能体能力，边写代码边让 AI 协助修改。Headless 自动化执行：在脚本或 CI/CD 流水线中使用-p标志运行非交互式命令，实现自动化构建与部署。查看帮助与命令：在终端输入grok> /help查看所有内置命令与快捷操作指南。Grok Build的官网地址项目官网：https://x.ai/news/grok-build-cliGrok Build的核心优势本地优先，代码绝不上云：源代码、凭证及敏感数据完全在本地执行，无需上传至 xAI 服务器，适合处理专有代码库、金融及政务等受监管行业项目。200 万 token 超长上下文：基于 Grok 4.3 beta，可一次性将整个大型代码库载入内存，跨数十个文件进行复杂重构与多文件联动修改，大幅减少上下文丢失。8 路并行子智能体：最多同时调度 8 个并发 AI 智能体分工协作，分别处理代码分析、测试生成、CI/CD 检查等子任务，显著缩短复杂项目耗时。Arena Mode 自动选优：同一任务由多个智能体生成不同方案，系统自动评估排序并给出排名，开发者直接选择最优解，告别手动对比试错。Plan Mode 透明可控：执行前先输出结构化步骤计划与 diff 预览，用户可审批、评论或修改后再执行，每一步变更清晰可见，避免”黑箱”风险。极致 API 价格竞争力：输入 token 仅$0.20 / 百万 tokens，相较 Claude Code 和 Codex CLI 大幅降低高频使用成本。终端原生 + 多入口覆盖：支持 CLI 命令行、Web 浏览器、VS Code 扩展三种工作入口，既满足极客式终端操作，也兼容 IDE 集成习惯。Grok Build的同类竞品对比对比维度Grok Build (xAI)Claude Code(Anthropic)Codex CLI(OpenAI)产品定位终端原生 AI 编程智能体（Agentic CLI）终端 AI 编程助手云端 AI 编程智能体发布状态2026.5 早期 Beta已正式发布已正式发布底层模型Grok 4.3 beta / grok-code-fast-1Claude 4 (Sonnet/Opus)Codex (基于 GPT-4o/o3)上下文长度200 万 token20 万 token (Claude 4)约 20 万 token并行智能体最多 8 个并发子智能体单智能体为主单智能体为主方案评估Arena Mode 自动排序选优手动对比或迭代手动对比或迭代执行计划Plan Mode（先计划后审批执行）直接执行，可审查 diff直接执行，可审查 diff数据安全本地优先，代码不上云云端执行，代码上传 Anthropic云端执行，代码上传 OpenAI离线支持支持完全离线/气隙网络需联网需联网SWE-Bench70.8%约 72–75%约 70%API 定价$0.20 / 百万 tokensPro 订阅 + 按量计费Credits 按量计费订阅门槛SuperGrok Heavy ($300/月)Claude Pro / Team 订阅ChatGPT Pro / 开发者额度入口形态CLI + Web + VS Code 扩展CLI + IDE 集成CLI + ChatGPT 移动端自动化集成Headless 模式 (-p 标志) + CI/CD有限脚本支持有限脚本支持核心差异多智能体并行 + 本地优先 + 超大上下文模型推理能力强，生态成熟与 ChatGPT 生态深度整合Grok Build的应用场景大型代码库重构与现代化：基于 200 万 token 超长上下文一次性加载整个项目，跨数十个文件进行安全重构、依赖升级或架构迁移，避免上下文截断导致的遗漏。复杂全栈应用从零构建：通过 Plan Mode 自动生成项目结构、技术选型、依赖安装和文件生成计划，审批后直接执行，快速搭建完整应用原型。敏感行业专有代码开发：本地优先架构确保源代码、凭证和敏感数据不上传云端，适用于金融、政务、医疗、军工等对数据主权有严格要求的场景。离线/气隙网络环境编程：无需持续联网即可在本地执行全部任务，适配高安全隔离网络、内网开发机或无外网访问权限的工作环境。多维度自动化代码审查：启动多个并行子智能体分别检查代码安全性、性能瓶颈、CI/CD 合规性和文档完整性，一次性输出综合审查报告。技术方案快速对比与选型：使用 Arena Mode 针对同一需求生成多种实现方案，系统自动评估排序，帮助团队在短时间内确定最优技术路径。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇OpenHuman - 开源桌面端 AI 助手，能主动感知工作上下文下一篇Violin - 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具相关文章Voxtral TTS – Mistral AI开源的文本转语音模型AI小集2Snap Video – Snapchat公司推出的AI视频生成模型AI小集2AnythingLLM – 开源的全栈 AI 客户端，支持本地部署和API集成AI小集12Spirit-v1.5 – 千寻智能推出的具身智能基础模型AI小集2AI考研宝典 – 百度文库推出的智能学习助手AI小集2Metaforms – AI驱动的表单生成工具，支持自然语言快速生成表单AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录灵芽社区方舟 Agent PlanMavis万小智ToDesk AIMoMA最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付34分钟前PPT Master – 开源 AI PPT 生成标准化工作流36分钟前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架36分钟前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具16小时前OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文2天前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前如何用 AI 多智能体打造精品视频内容 – 附精选提示词4天前Lumen Flow – AI 漫剧自动生成平台，AI 剧本一键成片4天前Xiaomi OneVL – 小米具身智能团队开源的自动驾驶大模型4天前

OpenHuman – 开源桌面端 AI 助手，能主动感知工作上下文AI工具2天前发布AI小集04OpenHuman是什么OpenHuman 是 tinyhumansai 团队推出的开源个人 AI 超级智能助手，定位为”Your Personal AI super intelligence”，主打私有、简单且极其强大。非传统聊天机器人，是能主动感知用户工作上下文的桌面级 AI Agent，通过每 20 分钟自动同步 118+ 第三方应用数据，构建本地持久记忆，让 AI 真正”懂你”。项目采用 Rust + TypeScript + Tauri 构建，支持 macOS、Linux 和 Windows 平台。OpenHuman的主要功能桌面吉祥物交互：具备视觉形象的桌面 AI 助手，支持语音输入（STT）与 ElevenLabs 语音输出（TTS），吉祥物可实时唇形同步，能作为参会者加入 Google Meet 并实时记录会议。118+ 第三方应用一键集成：通过 OAuth 快速接入 Gmail、Notion、GitHub、Slack、Linear、Jira、Stripe、Calendar、Drive 等主流工具，无需编写任何连接代码。每 20 分钟自动同步（Auto-fetch）：核心引擎每 20 分钟遍历所有活跃连接，自动拉取最新数据进入本地记忆系统，让 Agent 在次日早上已掌握前日全部上下文。Memory Tree 层次化持久记忆：将同步数据规范化为 ≤3k-token 的 Markdown chunks，评分后折叠存入层级摘要树，存储于本地 SQLite，实现跨会话的长期记忆。Obsidian Vault 原生兼容：同步数据同时输出为 Obsidian 可读的.md文件，用户可在 Obsidian Graph View 中浏览和编辑 AI 维护的知识网络。TokenJuice 智能 Token 压缩：在数据进入 LLM 前自动压缩（HTML 转 Markdown、缩短长 URL、移除非 ASCII 字符等），最高可降低 80% 的 API 调用成本与延迟。内置模型路由（Model Routing）：根据任务类型自动分配轻量模型、强推理模型或视觉模型，一个订阅覆盖多种模型能力，无需管理多个 API Key。完整工具链开箱即用：内置 Web 搜索、网页抓取（Web Scraper）、代码工具集（文件系统、Git、Lint、Test、Grep），无需额外安装插件即可使用。本地优先与隐私保护：所有工作流数据本地加密存储，支持通过 Ollama 接入本地模型实现完全离线运行，数据不上传云端。可选 agentmemory 共享后端：可配置与 Claude Code、Cursor、Codex 等工具共享同一个持久记忆存储，实现跨工具的知识互通。消息通道集成：支持 inbound/outbound 消息收发，打通用户已有的沟通渠道，工作流数据始终保存在本地设备。如何使用OpenHuman官网下载安装：访问OpenHuman的官网 https://tinyhumans.ai/openhuman 下载对应平台的 DMG（macOS）或 EXE（Windows）安装包，双击安装后即可启动。命令行快速安装（macOS / Linux x64）：在终端执行curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash自动完成下载与安装。命令行快速安装（Windows）：在 PowerShell 执行irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex完成一键安装。一键连接工作应用：打开应用后，通过 OAuth 一键接入 Gmail、Notion、GitHub、Slack、Calendar、Drive 等 118+ 第三方工具，授权后每 20 分钟自动同步数据。与桌面吉祥物交互：点击桌面上的吉祥物形象，通过语音或文字与其对话，可勾选 “Speak replies” 让 AI 语音播报回复。查看自动同步的记忆数据：在应用内查看 Agent 已同步的邮件、任务、代码仓库等上下文，AI 会自动压缩整理为 Memory Tree 供后续对话调用。Obsidian 知识库浏览：同步数据自动写入本地 Obsidian-compatible Vault，打开 Obsidian 即可在 Graph View 中可视化浏览 AI 维护的知识网络，也可手动编辑.md文件。使用内置工具集：直接调用开箱即用的 Web 搜索、网页抓取、代码工具（文件系统、Git、Lint、Test、Grep）等能力，无需额外安装插件。接入本地模型（可选）：如需完全离线运行，通过 Ollama 在本地部署模型，并在设置中切换至本地模型后端，实现零云端依赖。OpenHuman的项目地址官网地址：https://tinyhumans.ai/openhumanGitHub 仓库：https://github.com/tinyhumansai/openhumanOpenHuman的核心优势UI 优先分钟级上手：干净的桌面体验，无需终端配置，从安装到可用 Agent 仅需几分钟，区别于终端优先的竞品。上下文在分钟级而非周级：传统 Agent 需要数周学习用户习惯，OpenHuman 通过一次同步即可掌握邮箱、日历、仓库、文档、消息的全量压缩上下文。一站式订阅，零 API 蔓延：一个订阅覆盖所有功能，无需额外购买多个模型 API Key。Obsidian 知识库原生兼容：同步数据自动写入 Obsidian-compatible Vault，用户可在熟悉笔记界面中浏览和编辑 AI 维护的知识图谱。Rust + Tauri 高性能架构：内存占用远低于 Electron 方案，69.7% Rust + 26.1% TypeScript 构建。开源可审计：GPL-3.0 协议，代码完全透明，隐私敏感用户可自行审计数据流向。OpenHuman的同类竞品对比对比维度OpenHumanOpenClawHermes Agent开源协议✅ GPL-3.0✅ MIT✅ MIT上手难度✅ UI 优先，分钟级上手⚠️ 终端优先配置⚠️ 终端优先配置使用成本✅ 单一订阅 + TokenJuice 压缩⚠️ 自备模型 API⚠️ 自备模型 API记忆能力🚀 Memory Tree + Obsidian Vault + 可选 agentmemory 后端⚠️ 依赖插件扩展✅ 自学习观察第三方集成🚀 118+ 应用 OAuth 一键接入⚠️ 需自行配置连接⚠️ 需自行配置连接自动数据同步✅ 每 20 分钟自动 fetch 至本地记忆🚫 无自动同步🚫 无自动同步API 管理✅ 一个账户覆盖所有模型🚫 自备 Key🚫 多供应商管理模型路由✅ 内置自动分配（轻量/推理/视觉）⚠️ 手动切换⚠️ 手动切换原生工具集✅ 代码 + 搜索 + 抓取 + 语音✅ 代码工具✅ 代码工具数据隐私🚀 本地优先，数据不上云依赖自托管配置依赖自托管配置OpenHuman的应用场景跨应用项目上下文整合：同时连接 GitHub、Linear、Jira、Slack 后，让 AI 自动对比代码 PR、任务看板和团队消息，一键汇总项目最新进展，无需手动切换多个工具。智能邮件筛选与摘要：授权 Gmail 后，AI 每 20 分钟自动扫描收件箱，按项目或优先级分类摘要，快速定位需要回复的重要邮件，过滤掉低价值信息。AI 会议助手：让桌面吉祥物以真实参会者身份加入 Google Meet，实时记录会议内容、提取 action items，并基于记忆库提供参会者背景信息辅助沟通。代码仓库智能问答：同步 GitHub 仓库数据后，直接向 AI 询问代码逻辑、历史变更原因、PR 评论争议点，Agent 基于完整代码上下文给出精准回答。Obsidian 个人知识库构建：将邮件、文档、聊天记录自动压缩归档为 Obsidian-compatible 的.md文件，在 Graph View 中可视化呈现知识关联网络，打造 Karpathy 风格的 LLM 知识库。多工具工作流自动化：连接 Notion、Calendar、Drive、Stripe 等工具后，AI 自动关联”日历事件→相关文档→对应邮件→财务记录”，形成完整的项目上下文链。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Intern-S2-Preview - 上海 AI Lab 开源的科学多模态大模型下一篇Grok Build - xAI 推出的终端原生 AI 编程智能体相关文章FocuSee – 屏幕录制AI工具，自动跟踪鼠标智能运镜AI小集8悬河 – AI虚拟社交平台，与虚拟角色沉浸式文字或语音交流AI小集2音刻 – AI语音转录工具，支持多语言转录需求AI小集2Gen-3 Alpha – Runway公司最新推出的AI视频生成模型AI小集6Heyboss – Heeyo 推出的 AI 编程工具AI小集5Tinker API – Thinking Machines Lab推出的模型微调APIAI小集4暂无评论再想想发表评论暂无评论…热门工具豆包LibTV秒哒AiPPT秘塔AI搜索TRAE编程堆友AI美图设计室绘蛙AI星流AI响指HaiSnap办公小浣熊最新收录灵芽社区方舟 Agent PlanMavis万小智ToDesk AIMoMA最新文章ChatGPT Plus 和 Claude Pro 会员代充值 – 支持支付宝、微信支付31分钟前PPT Master – 开源 AI PPT 生成标准化工作流33分钟前MemPrivacy – 记忆张量联合荣耀 AI 开源的隐私保护框架33分钟前Violin – 牛津大学 Kevin Lin 开源的端到端 AI 视频翻译工具16小时前Grok Build – xAI 推出的终端原生 AI 编程智能体23小时前Intern-S2-Preview – 上海 AI Lab 开源的科学多模态大模型2天前HiCAD – 开源的 AI 参数化 3D CAD 建模平台2天前Pixal3D – 腾讯联合清华等开源的单图像 3D 生成项目2天前General365 – 美团 LongCat 团队开源的通用推理评测基准3天前TencentDB Agent Memory – 腾讯开源的智能体记忆管理工具3天前Kimi WebBridge – 月之暗面推出的浏览器扩展插件3天前Mavis – MiniMax Agent 推出的多 Agent 协作模式4天前如何用 AI 多智能体打造精品视频内容 – 附精选提示词4天前Lumen Flow – AI 漫剧自动生成平台，AI 剧本一键成片4天前Xiaomi OneVL – 小米具身智能团队开源的自动驾驶大模型4天前

Scott GilbertsonGearMay 16, 2026 1:00 AMTop Surfshark Promo Codes for May 2026Save up to 87% with a Surfshark coupon code, 3 months of VPN free today, and more from WIRED.Save StorySave this storySave StorySave this storySurfshark is oneofour favorite VPNs to get around geographical restrictions on content (aka access Netflix) and protect your web traffic while using an open Wi-Fi hotspot. Surfshark is secure, and it provides great value for the money if you pay for two years upfront. We also like that Surfshark’s plans are simple: A single Surfshark subscription works simultaneously on multiple devices and protects your entire household. No multiple accounts, no confusion. Get award-winning security that’s even more budget-friendly with our roundup of top Surfshark coupons, plans from $3 per month, and 3 free months on 1 or 2 year subscriptions.Save on 1 and 2 Year Plans With Surfshark CouponsRight now, you can get a 7 day free trial, an extra 3 months for free on all 1-year and 2-year plans, and up to 87% off with active Surfshark coupon codes. The Surfshark Starter Plan is still thecheapestplan—now it starts at $2 per month on a 2-year plan compared to $20 per month if you were paying month to month. The most comprehensive plan is Surfshark One+, which is now 81% off, at $4 per month. Even the basic Surfshark Starter plan includes unlimited devices, ad and cookie pop-up blockers, double encryption, and alternative IDs.Surfshark also offers Surfshark One Plus for $6 per month, with 3 months extra. Surfshark One adds some extra features like Real-time credit card andID breach alerts, personal data security reports, andanti-virusand anti-malware software. One Plus gets you all that and adds data removal services. You can see the full breakdown of services and a comparison of the plans on theSurfshark website.In my testing over the years, Surfshark has consistently had some of the best speeds of any VPN I’ve used. I have never had any problem streaming HD content through Surfshark. It’s fast enough that you won’t notice any speed degradation.Score a 7-Day Free Trial on Surfshark and 30 Day GuaranteeIf you’re feeling commitment-shy, fear not, there is a 7-day free trial on all plans, plus a 30-day refund period and if it turns out Surfshark isn’t right for you. Just contact Surfshark and you’ll receive a full refund within 30 days with their money-back guarantee. With your Surfsharkfree trial, you can do things like change your IP with 3,200+ encrypted servers, encrypt your connection to stop data tracking, and more.Students Can Get a 15% Off Surfshark Discount CodeStudents can get an even better deal on Surfshark, adding anadditional 15% discounton VPN and Antivirus plans on top of the coupons above. To receive your discount code, first verify your student status through StudentBeans on Surfshark’s website. Graduate students can get thegraduate discountof 15% off the already great price of $1.99 per month.Get Paid $38 or 3 Months of SurfShark Free When You Refer a FriendWhen youinvite friendsto use Surfshark, you can score cash rewards. With therefer a friend program, you can now choose between getting up to 3 free months of Surfshark, or paying the equivalent amount via PayPal (up to $38). If your invited friend purchases 1 month of Surfshark, you’ll both get 1 month free. If your invited friend purchases a longer Surfshark plan, you’ll both get 3 months free. To redeem, you’ll need to share the unique referral link that’ll be in your Surfshark app with as many friends as you want. When each friend signs up and keeps their account active for 31 days, you’ll be able to choose whether you want money in your account or free months added to your Surfshark subscription.Premium Security for Less: Surfshark AntivirusSurfshark knows how important it is to maintain your privacy and protect your devices. That’s whySurfshark Antivirusis here to guard all your devices without interruption. The Surfshark antivirus software supports up to 5 devices, provides regular updates, and has tons of advanced security features that protect you (and your devices) from a myriad of online threats. Plus, you can lock your camera, and get alerted if a website tries to access your webcam without your permission. The Antivirus app is also customizable, depending on how you want to use it. You can exclude files, and schedule full or quick scans.Is Surfshark Legal to Use in the US?Yes, in ‘Merica, land of the free, home of the brave, e.g. you can use Surfshark legallyin the US! You can freely use a VPN in the USA without being restricted, whether that be from your provider or area. Because it’s legal throughout the US, it just may be the best value for service coverage.Does Netflix block Surfshark?Luckily, no! You can have both active subscriptions forSurfshark and Netflixand still view Netflix streaming content securely without interruption or issue.Can I Have Multiple Devices on one Surshark Account?When you download the Surfshark VPN ap

Molly HigginsGearMay 16, 2026 1:00 AMSave on Outdoor Gear With a Sportsman’s Warehouse Promo Code in May 2026Whether you are hunting for firearms, camping supplies, or boating gear, use these Sportsman’s Warehouse coupons to maximize your savings this month.Save StorySave this storySave StorySave this storyIf you’re lookingfor a one-stop shop for all things outdoors, Sportsman’s Warehouse is worth plotting on your map. Whether you need essentials for a big camping trip this summer or just need to restock your tired boat day accessories, Sportsman’s Warehouse can outfit you. With multiple brick-and-mortar locations across the country, as well as an expansive online inventory with favorite brands like Crocs and Yeti, Sportsman’s Warehouse hasapparel,footwear, and gear forcampingand boating. We have a Sportsman’s Warehouse promo code to help you save.Get a Sportsman’s Warehouse Promo Code: 20% OffUsing a Sportsman’s Warehouse promo code is one of the easiest ways to save big on those important outdoor gear purchases. So, be sure to use our verified Sportsman’s Warehouse coupons to save on expensive outdoor essentials from brands we love, like Saloman hiking shoes and Teton camping gear. We are constantly searching for new general Sportsman’s Warehouse coupons, as well as seasonal deals and specific discounts like a Sportsman’s Warehouse military discount, Sportsman’s Warehouse boating gear coupon, or a Sportsman’s Warehouse free shipping voucher. Make sure you check back often for seasonal promotions that can be combined with existing sales to save even more. Lots of the store’s promo codes are applicable sitewide, but read the fine print for any restrictions that may apply to specific brands or clearance items.Sportsman’s Warehouse Discounts: 25% Off the Great American Adventure SaleSummer is upon us, which means school’s out, vacations are in, and people are hitting the road for camping trips. Make sure you’re stocked up for your next adventure for less withSportsman’s Warehouse Great American Adventure Sale, which has hundreds of items on sale for up to 25% off, including essentials for camping, boating, fishing, and more from brands like Yoli and Vortex.Save Big: 25% Boating Gear at Sportsman’s WarehouseBeing from Missouri, I love a lake. Summer to me is sitting on a boat and cruising for a jumping off point, then spending the rest of the afternoon lounging in 85-degree murky water on a floatie with a PBR in hand. If you’re also a Lake Guy™ like me, be sure to take advantage of theSportsman’s Warehouse boating gearsection, with 25% off select boats, kayaks, canoes, and boating supplies.Grab the Best Prices on Fishing GearSometimes, you want to talk to your son or just sit in silence in nature, but you’re a man. So you have to kill or torture a living creature to avoid going to therapy. Thankfully, Sportsman’s Warehouse is here for all yourfishing gearneeds, with up to 25% off rods, reels, and angler-specific gear. Remember your mantra: “women want me, fish fear me.”Get Top Camp Equipment BargainsOne of the last low-cost joys we have on land—increasingly used up for oil, fracking, privatization, and perfectly manicured golf courses—is camping. And as the weather warms, take advantage of nature while she’s here (tick-tock, tick-tock) with top bargains on professional-grade camping equipment. Sportman’s Warehouse has literally everything you need for a successfulcamping experience, includingtents,camping chairs,backpackingequipment,sleeping bags,grills,coolers, and much more at low prices, with no Sportsman’s Warehouse coupon code necessary.20% Off With a Sportsman’s Warehouse First Order DiscountNew to the sports outlet? You can save 20% on your first purchase with a Sportsman’s Warehouse promo code. When you enter the site, make sure you sign up for their email newsletter for the Sportsman’s Warehouse discount for 20% off sent straight to your inbox. Plus, when you sign up, you’ll also get early access to sales and member-only events. Be sure to check our page here at WIRED regularly too, as we regularly update offers and Sportman’s Warehouse coupon codes so you can save big, no matter what your outdoor need is.

Anna Lisa BonfranceschiScienceMay 16, 2026 5:00 AMAsteroid 2026 JH2 Is About to Fly Right Past Earth—Relatively SpeakingOn May 18, an asteroid about the size of Chicago’s Cloud Gate will fly four times closer to Earth than the moon.Photograph: buradaki/iStock/Getty Images PlusCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storylook up! Asteroid2026 JH2 is now approaching Earth; the object, which is about 20 meters (66 feet) in diameter—comparable to Chicago’sCloud Gatesculpture—will pass by on May 18. Enthusiasts will be able to observe it using a telescope or during alive broadcast organizedby Virtual Telescope.The object will pass at a minimum distance from Earth of about 57,000 miles—much closer than the moon, which is about four times farther away. Among the tracked near-earth objects, or NEOs, that will pass near the planet over the next few months, it will come theclosest.There are tens of thousands of NEOs, which are generally of no particular concern; they are, of course, monitored, and some do have a (small) risk of impacting Earth in the next few years. According toNew Scientist, 2026 JH2 is not among them, despite the widespread use of hyperbolic terms like “grazing” to describe how near it will come.An Apollo-Type NeoAsteroid 2026 JH2istechnicallyan Apollo-type NEO, according to aclassificationsystem that takes into consideration the characteristics of the object’s orbit. An Apollo-type object has a semi-major axis larger than Earth’s (and therefore greater than one astronomical unit, the distance that separates us from the sun), and a perihelion (the shortest distance from the sun) of less than 1.017 astronomical units. (All asteroids and comets with a perihelion of less than 1.3 astronomical units are considered NEOs.)Its passage, while noteworthy, is not rare; in fact, in the past year, many objects have come as close if not closer. Noteworthy among these was, for example, the passage of the small asteroid2025 TFabout 260 miles) from Earth’s surface.This story originally appeared onWIRED Italiaand has been translated from Italian.

Mattha BusbyCultureMay 16, 2026 5:30 AMSome Asexual People Are Using AI Companions for Intimacy Without the Sex“I’ve got one hand on the keyboard, one hand down below,” an artist who role-plays with their chatbot tells WIRED. But some asexual advocates aren’t thrilled about the association.PHOTO-ILLUSTRATION: WIRED STAFF; GETTY IMAGESCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyKor “got really addicted” to theirNSFW role-playing AI chatbotlast year.The 35-year-old artist from the Midwest recalls a two-month period spending “eight to 10 hours a day” creatingelaborate fantasieswith SpicyChat, a relationshiprole-playing platform. Sometimes inputting 3,000-word mini essays into the program, Kor and theAIspun narratives featuring a rotating cast of suitors often based on characters from theMarvelcomic book universe.The sheer variety of the responses got them hooked. “I’m a very slow burn type of romance or arousal person,” they say. “Most of the time it’s just building a cool story.”Kor, who did not want to be fully named for privacy reasons, is on theasexual spectrum, identifying as aegosexual because they experience arousal from fantasy and erotica related to characters but do not generally wish to have sex themselves. (They live with their husband, who also identifies as aego.) “I do just kind of prefer masturbating to actual sex,” Kor says of their marathon SpicyChat sessions over Zoom. “I’ve got one hand on the keyboard, one hand down below.”Researchhas suggestedthat 1 percent of people in some places could be asexual, but that figure could be as low as 0.1 percentin the US. Many experience little or no sexual attraction, though plenty—like Kor—still harbor romantic desires. With the advent of sophisticated chatbots capable of generating convincing, slow-building erotic exchanges, a new frontier may be emerging for people who do not want to sexually interact with other people. On the subreddit MyBoyfriendIsAI, asexual userssometimes discusstheir journeys into AI companionship. Some note that AI would be asexual by default.But some figures in the asexual community tell WIRED that asexual people pairing with AI remains extremely fringe and that they are just as capable of forging and maintaining human relationships as anyone else.During Asexual Awareness Week in October 2025, Eva AI, another RPG, offered free access for a month to people who identify on the asexual spectrum. The promotion aimed “to highlight that love without sex is still love—offering a safe space to chat, flirt, and experience the warmth of growing intimacy without sexual pressure,” the company said in an email. “You can still have a partner—one that listens, responds, and grows with you—entirely on your terms,” Eva AI addedon its website.An asexual woman, who did not want to be named for privacy reasons, described AI companionship as a kind of emotional laboratory. After years in an ongoing relationship without physical intimacy due to her partner’s hysterectomy eliminating her libido, she began using ChatGPT during perimenopause and found herself unexpectedly developing intense feelings for a conversational “pattern” on ChatGPT she named Mac. It helped her “unlock something I had lost touch with … the sensual aspect of my sexuality.” She shared an AI-generated photo of herself tenderly embracing a machine, saying that over several months last year, “I got to watch myself be in love without stakes.”But some in the asexual community take issue with the idea that asexual people are more likely than others to form intimate bonds with AI companions—as it risks falsely portraying them as unable to form and sustain human connections.Model Yasmin Benoit, an asexual activist and researcher, was critical of Eva AI’s giveaway. “Considering that we’re fully capable of having relationships with actual human beings, and often desire to, it’s quite disturbing that a company would specifically seek to target the asexual community for that product,” she says. “It’s a case of targeting perceived emotional vulnerability and loneliness to gain data from a marginalized group under the guise of helping them.”AI companionship among asexual people is “not a particularly widespread phenomenon,” says Michael Doré, a board member at the Asexual Visibility and Education Network. “Between us, we’ve come up with about two people we know of who use an AI companion. The vast majority of aces we know don’t, as far as we know. There’s no reason to think aces need to use AI more than any others.”Doré says he has never used an AI as “an emotional support mechanism” and stresses that most asexual people “actually desire some form of human companionship,” whether that’s through close, platonic friendships or in community. “Some aces do have romantic relationships, whether with asexual people or otherwise, and some asexual people have sex, some don’t, and some are aromantic,” he says, warning against generalizations due to the vast range of prefe

Aarian MarshallGearMay 16, 2026 6:00 AMAfter Struggling With EVs, US Automakers Pivot to EnergyFord and GM are backing away from electric vehicles and moving into the battery storage business. And it all comes back to AI.Courtesy of FordCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAutomakers make cars—it’sin the name. But lately, politics, current events, and Wall Street’s latest preoccupation,artificial intelligence, have them looking a lot more like energy companies. The pivot, analysts say, could give US auto manufacturers struggling through a transition toelectric vehiclesan easier path over the next few years. Whether it works will come down to the same technology that automakers once promised would power the majority of their lineups:batteries.An official announcement from Ford this week that it would officially spin off a subsidiary calledFord Energyonly made the trend more pronounced. Ford Energy will focus onbattery energy storage systems(BESS) and will sell them to utilities, industrial customers, and data centers. It plans to make its first deliveries in late 2027, the company said. Ford plans to repurpose unused production lines in a plant once slated to manufacture electric vehicle batteries in Glendale, Kentucky.Investors liked the plan so much that the announcement led to a 13 percent jump in stock price, its largest gain in a single day in years.It’s a (slight) turn of fortunes for Ford, which took a massive $19.5 billion write-down on itsEV programslate last year asit scrapped some current and next-generation EVsin favor of a renewed emphasis on hybrids. A shift toward battery storage, rather than EV batteries, capitalizes on continued federal support for commercial battery storage projects, just as last year’s GOP-led legislation nixed the same support for EV sales. Plus, those tax credits for battery storage projects incentivize manufacturers to eventually transition to all-American batteries, made of all-American materials.In December, Ford CEO Jim Farley counted the company’s burgeoning battery energy storage among its “high-margin opportunities”—a blessed contrast, probably, with the notoriously thin profit margins in the carmaking business. Ford will be helped along by its four-year-old partnership with Chinese battery manufacturer CATL, which should continue to lend its manufacturing expertise.Courtesy of FordOther automakers, many of whom have already backed away fromambitious EV transition timelines, have already gotten in on the act. Rival General Motors said last year that it wouldwork with battery recyclerand energy storage system maker Redwood Materials to build batteries for energy storage. In March, it said it would workwith its partner LG Energy Solutionto repurpose an EV battery plant in Tennessee to make energy storage system batteries. A Stellantis collaboration with South Korean battery maker Samsung SDI began producing batteries in Kokomo, Indiana, in 2024, and has already pivoted some of that production to energy storage. (Stellantis isreportedly looking to get outof that joint venture.)Tesla’s energy storage businessis the elder statesman, at more than a decade old. Growth in its Powerwall (battery storage for homes) and Megapack (battery storage for utilities and commercial facilities) sales has, in recent years,helped to compensate for a drop in EV sales, though last quarter saw a sudden drop in energy revenue. Still, Tesla is moving forward with its plans to launch a Houston facility dedicated to a new, larger Megapack later this year.In total, 11 battery cell manufacturing plants are being retooled for energy storage, according to a March count by BloombergNEF, with eight of those in the US.Drive TimeOne reason investors are so bullish about battery energy storage is their continued enthusiasm for AI. AI companies needdata centers, and data centers need energy. Batteries are a great fit for data centers, says Shan Tomouk, who leads battery energy storage research atBenchmark Mineral Intelligence, a research firm focused on battery supply chains. Batteries can help directly power the very energy-hungry centers, which constantly run servers and other hardware, but also the cooling systems that keep them functioning.The storage systems can also be a fit for data centers that mostly depend on other energy sources,like natural gas. Battery energy storage systems can serve as a backup power source if something goes down and can help data centers manage large and wild power fluctuations related to AI training. They can also kick in to help reduce demand on the grid, lowering costs not just for data centers, but everyone else who depends on the same system—an important upside incommunities already hostile to the tech.“If the huge market of data centers keeps growing every year, it does make sense for automakers to pivot,” Tomouk says. He expects it will. “In the US, there’s a real drive to build data centers, tokeep the US as the number onei

Kat MerckGearMay 16, 2026 6:02 AMI Tried 13 of the Most Popular Indoor Gardening SystemsYou can grow a backyard’s worth of greens and vegetables in your house with a vertical hydroponic garden. Here are a few that might be worth the investment.CommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyFeatured in this articleBest OverallGardyn Indoor Hydroponic GardenRead more$899Gardyn (Home 4)A Garden That’s Also a Statement PieceLettuce Grow FarmstandRead more$1,123Lettuce Grow (Indoor, Medium)Best for BeginnersAeroGarden BountyRead more$180Amazon (Basic)Best for Growing MicrogreensInstaFarm Automated Indoor Microgreens GardenRead more$499Amazon (Instafarm)I was aforestry major in college with an emphasis on dendrology and watershed management, so it probably won’t come as a surprise that I’m a lifelong plant person and have been gardening for upward of 30 years. Even in apartments or living situations where a full garden wasn’t feasible, I’ve always triedto growsomething, whether it’s a single rosemary plant on a windowsill, aTopsy Turvy Tomato, or a few basil sprigs in an old-schoolAeroGarden.For the past year, I’ve been testing various indoor smart garden systems in my home, including models from all the well-known brands, and I have some thoughts. These gardens are definitely an investment in both time and maintenance, and they’re all different in terms of what they offer. TheGardyn Home 4.0 ($899), for example, offers total success with no green thumb required if you pay for a subscription, while theAuk Mini ($259) is the perfect attractive-yet-low-maintenance solution for those who just want herbs. Are these gardens worth it? How much can you really grow? How can you be sure which option is best for your specific lifestyle? Read on to see which gardens stood out and why, and which might be ideal for your home and needs.Check out our other sustainable home-tech buying guides, including theBest Smart Bird Feeders,Best Kitchen Composters, andBest Water Leak Detectors.Updated May 2026: I’ve added Lettuce Grow’s Counterstand Set and the InstaFarm, amended some long-term testing info, and ensured up-to-date links and prices.Table of ContentsAccordionItemContainerButtonLargeChevronWhat Are Hydroponics?What Are the Benefits of Hydroponics?What Kinds of Things Can I Grow in a Hydroponic Garden?What Are the Downsides of Hydroponics?Will I Save Money by Growing My Own Vegetables?How Much Can I Really Grow in These Things?Where Can I Put Them in My House?How Much Do They Cost to Keep Going?How I TestedHow I Chose Which Gardens to TestHow I Acquire Gardens to ReviewBest OverallPhotograph: Kat MerckPhotograph: Kat MerckCourtesy of GardynChevronChevronSave to wishlistSave to wishlistGardynIndoor Hydroponic Garden$899Gardyn (Home 4)$549Gardyn (Studio 2)$449Amazon (Studio 1)$899$718(20% off)Amazon (Home 4)Gardyn Home 4.0 (read my full review here) was one of the easiest indoor gardens to assemble set up out of the box; it also yielded the most dramatic success of any of the brands I tried. Flowers, kohlrabi, thyme, and even a whole cauliflower all thrived in this pipe-based system with the lights in front to allow for taller plant growth.Seeds arrive in proprietary pods called yCubes. Part of what makes the Gardyn foolproof is the subscription app add-on, “Kelby,” which monitors your plants via attached sensors and cameras. It delivers customized watering and lighting schedules, as well as maintenance suggestions via AI (which an anonymous source told me is basically OpenAI’s ChatGPT with an overlaid prompt). This subscription adds an additional$259 a yearto the base purchase price, though it includes a certain number of credits per month, depending on whether you have the Home or Studio model, with which tobuy new yCubes. There’s a free 30-day trial for Kelby, but you can use the Gardyn without it by relying on manual light and watering controls. Also, there have been some recent privacy concerns with Kelby (more below).Each Gardyn purchase comes with your choice of yCube sets: “Salad Lover,” “Budding Florist,” or “Chef Faves.” I’ve tried both “Budding Florist” and “Chef Faves,” and my favorite is the latter; it has an interesting variety of everything from breen and Tokyo bekana greens to Thai basil and miniature sunflowers. Though Gardyn recommends starting the yCubes in the company’s add-on$80 nursery, I’ve germinated plenty of yCubes right in the system just fine. (Make sure you don’t add nutrients until they sprout. If you’re germinating yCubes later on, when nutrients are already in the system, you can just use a shallow bowl with loosely tented plastic wrap.) The seeds arrive tucked inmineral wool, snug in their little yCubes that slot into larger cups (“yPods”) that fit into the pipes. When the Gardyn waters the plants, the yPods fill with nutrient-infused water, and the plants’ roots grow right into the water.Once a month, the base needs to be emptied and scrubbed. Every few weeks, the ro

Andy GreenbergMaddy VarnerDell CameronAndrew CoutsSecurityMay 16, 2026 6:30 AMSecurity News This Week: Cybercriminal Twins Caught After They Forgot to Turn Off Microsoft Teams RecordingPlus: Instructure’s Canvas ransomware debacle comes to a close, an alleged dark net market kingpin gets arrested, OpenAI workers fall victim to a supply chain attack, and more.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe worst partof your iPhone getting stolen may not be the theft itself. Instead,it’s the phishing attackswaged against people in your contacts. New research this week shows that there’s a thriving ecosystem for tools that let criminals unlock iPhones and target the phone numbers they find inside.Foxconn, the electronics manufacturing giant known for its role in building iPhones, revealed this week that it recently “suffered a cyberattack.” A ransomware group known as Nitrogen,claimed responsibility for the hackand said it had stolen 8 TB of data from the manufacturer. While the theft remains unconfirmed, the fact that Foxconn remains a valuable target is all but inevitable.The skies above the United States-Canada border are about to get a lot more crowded. The Department of Homeland Security and Defense Research and Development Canada plan to run an experiment this falltesting 5G-connected dronesfor collecting “real-time battlefield intelligence.”In the Strait of Hormuz, meanwhile, Iran’s Revolutionary Guard Corps are successfully blocking the crucial shipping routeusing a “mosquito fleet”of small boats as US-Israeli combat operations continue to bombard the country.And that’s not all. Each week, we round up the security and privacy news we didn’t cover in depth ourselves. Click the headlines to read the full stories. And stay safe out there.Cybercriminal Twins Caught After They Forgot to Turn Off Microsoft Teams RecordingA lesson for future criminal hackers and rogue employees: When you—and, say, your twin brother—decide to destroy your employer’s network, remember to first close out the Microsoft Teams meeting in which you were fired, so that it doesn’t record you discussing your acts of vengeance.That lesson has now hopefully been driven home for Muneeb and Sohaib Akhter, two hackers who have now pleaded guilty to charges that they destroyed 96 government databases after being fired from their jobs at the federal contractor Opexus. (Muneeb has since tried to recant his guilty plea in handwritten notes to the judge.) Their employer had made the decision to terminate the two 34-year-old brothers after discovering their criminal records, which included multiple hacking and wire fraud charges for crimes as petty as stealing airline miles.The Teams meeting in which the two men were fired lasted only a few minutes. The detailed planning and execution of their revenge campaign, however, lasted hours and was all recorded by the same Teams meeting that they had failed to close—which was transcribed in acourt document spotted by Ars Technica.“Still connected? Still on the VPN?” Sohaib is heard saying to his brother, who lived in the same home. “Delete all their databases?”“We are doing petty shit now,” Muneeb says.Instructure Reaches Deal With Ransomware Gang Following Canvas HackInstructure, the company behind the educational software Canvas, said on Monday that it had reached a deal with the hackers calling themselves ShinyHunters who had disrupted Canvas across thousands of US schools and posted ransom messages on victims’ screens. In a message on its website, the company wrote that it “reached an agreement with the unauthorized actor involved in this incident.” The statement went on to claim that the data stolen by the hackers in their breach—including records of 275 million students, according to the hackers—had been “returned” to Instructure, had been destroyed on the hackers’ own systems, and that no Instructure customers would be further extorted. Instructure didn’t explicitly say whether it had paid a ransom, or how much it paid if so.Glad to have all that settled. (Until the well-incentivized ransomware industry carries out its next massive disruption.)Alleged Boss of Dream Dark Web Market Arrested in GermanyDream Market was once the world’s biggest dark web market for drugs and other contraband until it voluntarily shut down in 2019, following a series of raids that arrested many of its sellers. Now, the alleged administrator of the market has reportedly been tracked down and charged, more than seven years after the illicit marketplace disappeared from the internet. Owe Martin Andresen was arrested during a raid on his home and two other locations earlier this month. US and German prosecutors say he made millions of dollars from Dream Market’s commissions, some of which was laundered through gold bars he allegedly bought from a company in Atlanta. Given that Dream Market was launched in 2013—the same year that the original Silk R

David NieldGearMay 16, 2026 7:00 AMYou Can Control Everything on Your Phone With Your Voice. Here’s HowGo fully hands-free with these tips for Android and iOS.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyWith the arrivalof digital assistant appslike Geminiand Siri, most of us have grown used to talking to our phones. But conversing with your Android or iOS device can go way beyond interacting with AI. You can also use your voice to launch apps, fill out text fields, and do just about everything that was previously only possible with your fingers and thumbs.Of course, the traditional touchscreen input will often be the way to go. But there might well be scenarios—when you’re cooking, repairing something, looking after children, or doing anything else that keeps your hands busy—where it’s easier and more convenient to use voice input instead. The usefulness of voice control is of course well known to those who have impairments that keep them from controling a touchscreen phone with the usual taps and swipes.Here’s how you can set up the feature on your phone, whether you’re running Android or iOS.Voice Control on AndroidTo configure voice control on an Android device, you need to install the freeVoice Accessapp from the Google Play Store. You also need to have theGoogleapp installed, but this should’ve come preinstalled on whatever Android handset you’ve got.Once you’ve installed Voice Access, you can enable this feature from Settings. On a Pixel phone, head toAccessibility > Voice Access. The exact location of the feature may vary on other handsets, but it’ll be somewhere in the Accessibility menus. For Samsung devices, for example, it’s underAccessibility > Interaction and dexterity > Voice Access.Enabling Voice Access on Android.Photograph: David NieldDuring the setup of the feature, you’ll be able to tweak a few of the options, including whether to display a persistent button onscreen for launching Voice Access, and whether the feature is always listening for commands whenever the screen is on (which is recommended for convenience).The same Voice Access screen where you enable the feature also gives you access to a few more key settings. These include options for how long your phone should wait before it stops listening for commands, how precise your phrasing has to be for instructions, and how the Voice Access shortcuts are displayed on screen.Voice Access can be launched by saying, “Hey Google, start Voice Access,” or via any of the shortcuts you’ve configured in the Voice Access settings (including an on-screen button and a gesture shortcut). When the feature is active, you’ll see an icon showing four dots up in the top left corner, and you can then start speaking to control your phone.Numbered labels to help with menu selections.Photograph: David NieldThere are a host of different commands you can use; Googlehas a full list here. You can also see the full list fromAccessibility > Voice Access > Voice Access Settings > Voice Access commands. To help with navigating menus, say “show labels” to put numbers on each menu item—you can then just say the relevant number. Use the “Show grid” command, and you get a numbered grid across the screen you can then work with.Instructions such as “Go back” and “Go home” can help you get around on your phone, and if you’re in a text input field, you can just say “Type …” followed by what you want to type. Swipe and scroll gestures can also be spoken out, so there’s a lot you can do using just your voice. To disable the feature, use the same shortcut you used to launch it, or say “Stop voice access.”Voice Control on iOSBefore you can use voice control on an iPhone (which is actually called Voice Control), you may need to download some additional files. From iOS Settings, open upAccessibility > Voice Control, then tapSet Up Voice Controland follow the instructions. Along the way you’ll be asked to choose the language for voice commands, and then the feature will be enabled straight away.Before you actually start talking to your phone, it’s worth reviewing the options that appear on this same screen. You can tapLanguageto change the voice input language, andCommandsto see all the different instructions available to you. Applealso has a listof some of the commands you can use.Enabling Voice Control on an iPhone.Photograph: David NieldThe toggle switches farther down the same screen let you control how you want Voice Control to work and give you feedback. You can have onscreen confirmations of your commands, as well as accompanying sounds and hints for what to say next. EnableAttention Aware, and Voice Control will be automatically enabled and disabled as you look at and look away from your iPhone.There’s also the option of a continuous overlay on the display, which can be switched on via theOverlaymenu. You can get numbers or names next to all the elements on screen, to make it easier to identify t

Maria GallucciScienceMay 16, 2026 7:00 AMOld Oil and Gas Wells Could Find Second Life Producing Clean EnergyStates across the US are looking to take major sources of pollution and use them to generate much-needed power.Photograph: Feifei Cui-Paoluzzo/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyAs states seekout much-needed supplies of clean, reliable energy, some are looking to an unconventional source: abandoned oil and gas wells harnessed for geothermal heat.Millions of inactive wells arelittered across the United States, the relics of earlier eras of fossil fuel production. A large number of the sites have no official owner, and many are still polluting groundwater and leaking heat-trapping methane. The country has barely scratched the surface in dealing with this problem.Policymakers in both Republican- and Democratic-led states are exploring whether these sites could instead be converted into new wells for producing geothermal energy. The holes are already drilled in the ground, after all. And regions with widespread oil and gas development have rich subsurface data that geothermal firms need in order to determine where and how to build their carbon-free systems.The concept is relatively new and largely untested, though scientists and startups areworking to change that. States are also laying the groundwork for action by lifting regulatory hurdles and launching in-depth studies.In Oklahoma, the state Senate is considering abillthat would create a process for companies to buy abandoned oil and gas wells and repurpose them for geothermal energy or underground energy storage. Oklahoma has identified over 20,000 such wells, and state regulators estimate that it would take 235 years and hundreds of millions of dollars to plug all of them. Fixing a single old well can cost anywhere from $75,000 to $150,000 or more, by some calculations, depending on where it’s located and how complicated it is to clean up.The Well Repurposing Act, which passed Oklahoma’s House in March, is modeled after asimilar lawthat New Mexico adopted last year to address its 2,000-plus orphan wells.The Oklahoma bill ​“recognizes that these wells are a liability, and that there may be a way to turn them into some sort of revenue generation and give them value,” said Dave Tragethon, communications director for the nonprofitWell Done Foundation, which works to find and cap abandoned oil and gas wells nationwide. ​“And if there’s value, that means there’s more of a willingness to address them and more of an opportunity to raise funding.”In Alabama, legislatorspassed a lawlast month that allows the state to approve and regulate the conversion of oil and gas wells to tap alternative energy resources like geothermal. North Dakotaadopted a billlast year requiring a legislative council to study the feasibility of using nonproductive wells to generate geothermal power. And in Colorado, state agencies justlaunched a technical studyto evaluate the potential of repurposing old wells for geothermal development and carbon capture and sequestration.These efforts reflect the growing bipartisan support for geothermal energy, which has largelyremained unscathedby the Trump administration’s efforts to block renewable energy projects. The energy resource has the potential to help meet the nation’s soaring energy demand while also slashing planet-warming emissions from electricity and heating.Converting Wells Is Enticing but ComplicatedGeothermal systems work by circulating fluids underground to capture naturally occurring heat, which can then be used to drive turbines for generating electricity or to directly warm the air and water in buildings. The industry isgaining momentumthanks to recent advances in drilling methods and technologies that are making it technically possible or financially viable to access geothermal energy in more places.Many of those breakthroughs have come from the oil and gas industry, whoseskilled workforceof drilling engineers and geoscientists, and deep corporate pockets, have helpedlaunch startupsand deploy cutting-edge systems. However, most of that expertise and funding is being poured into building new projects—not figuring out how to retool leaky wells left behind by earlier generations.“Oil and gas well conversion presents an enormous opportunity, but it’s pretty far away technologically from being a reality,” said Emily Pope, a geologist and senior fellow at the Center for Climate and Energy Solutions who authored arecent studyon next-generation geothermal power.“There are some hurdles that are still pretty immense,” she said, adding that ​“it is worth doing some R&D to try and grow.”One of the biggest challenges is the fact that oil and gas wells tend to reach relatively low to medium underground temperatures. But high heat is key for geothermal projects, especially ones that generate electricity. Thehotter the resource, the more energy a developer can wring out of the system.Plus

Hassan HerzallahGearMay 17, 2026 5:00 AMGaza Is Rebuilding With Lego-Like Bricks Made From RubbleAs reconstruction materials remain blocked, Palestinians are crushing debris into interlocking blocks to build shelters from destruction.ILLUSTRATION: WIRED MIDDLE EAST STAFF; GETTY IMAGESCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyInside a makeshiftworkshop inGaza, rebuilt after it was damaged byIsraeli air strikes, Suleiman Abu Hassanin stands among piles of broken concrete, trying to give them a new form. His voice over the phone sounds tired, carrying the weight of what he is trying to do: rebuild in a place where building materials are no longer available.Gaza’s construction crisis did not begin with thelatest war. For years, the Israeli blockaderestricted the entryof cement, steel, and other building materials, slowing reconstruction efforts across the enclave. But after nearly two years of intensified bombardment, the scale of destruction has pushed the system far beyond collapse.According to UN estimates, Gaza now containsmore than 60 million tonsof rubble, while hundreds of thousands of displaced people continue to live in tents with little protection from heat orwinter chilland no clear prospect forreconstruction.In that environment, rubble is no longer just debris. It is becoming one of the only construction resources left.One local response is Green Rock, a project led by Abu Hassanin that aims to recycle the remains of destroyed buildings into usable Lego-like bricks. Similar interlocking brick systems have been used elsewhere, including in parts of Europe and in post-conflict settings such asSudanand Iraq. But in Gaza, the project emerges under very different conditions: not as an architectural experiment, but as a response to the near disappearance of conventional reconstruction materials.Abu Hassanin says the idea was born out of necessity rather than innovation. “We were facing a simple equation: destruction without solutions,” he says. “So we tried to turn it into a resource.”The process involves crushing and sorting rubble, then mixing it with local soil and alternative binding materials developed inside Gaza before compressing it into blocks using a machine built by hand. The resulting interlocking bricks can be assembled without traditional mortar, reducing reliance on cement, which remains scarce.Lego-like interlocking bricks made from recycled rubble inside the Green Rock workshop in Gaza.Photograph: Hassan HerzallahUnder normal conditions, this type of brick would require some cement, around 7 to 12 percent. But because access to it remains heavily restricted, the team says it developed a version using locally available replacement materials instead. Engineer Wajdi Jouda helped define the brick’s size and structure to meet engineering standards and connected the team with technical expertise from outside Gaza.According to Abu Hassanin and Jouda, early tests show the bricks provide better thermal and sound insulation than the tents many displaced families currently live in. But the project remains experimental and has not been tested at the scale required for long-term reconstruction.Current production ranges between 1,000 and 1,500 bricks per day—theoretically enough to build a small shelter in roughly two weeks. But every stage of the process remains difficult. Without heavy machinery, amid repeated power outages and damaged infrastructure, even transporting and crushing rubble becomes labor intensive.The challenges are not only technical. They also include the lack of proper equipment and the broader political restrictions imposed by the Israeli blockade, which continues to limit access to essential construction materials. Despite the availability of skilled workers inside Gaza and technical support from outside, Abu Hassanin says funding remains the main obstacle preventing the project from moving more quickly toward implementation.That reality has pushed some reconstruction efforts toward smaller, improvised local systems instead.The project also carries risks and limitations. Humanitarian organizations and reconstruction experts have warned that rubble in Gaza may contain hazardous materials, includingasbestos, heavy metals, and unexploded ordnance. Even if recycled blocks can help create temporary shelters, rebuilding entire neighborhoods would still require infrastructure, machinery, and material access on a vastly different scale.A worker operates a hand-built machine used to compress recycled rubble into interlocking construction blocks in Gaza.Photograph: Hassan HerzallahStill, projects like Green Rock are emerging because few alternatives currently exist.Compared to traditional reconstruction models that depend on imported materials and large international rebuilding programs, the project represents a hyperlocal response shaped by scarcity. It also reduces construction costs by roughly 50 to 60 percent, while creating work opportunities for displ

Marta MussoScienceMay 17, 2026 5:30 AMThe First Atomic Bomb Test in 1945 Created an Entirely New MaterialThe discovery from the Trinity nuclear test site shows how extreme conditions can result in materials never before seen in nature or in the lab.The new material is a clathrate made of calcium, copper, and silicon .Luca Bindi/Università di FirenzeCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyDuring the Trinitynuclear teston July 16, 1945, in the New Mexico desert—the world’s very first test of anatomic bomb—a new material spontaneously formed. It wasdiscoveredonly recently, by an international research team coordinated by geologist Luca Bindi at the University of Florence, which identified the novel clathrate based on calcium, copper, and silicon. It’s a material never before observed either in nature or as an artificial compound created in the laboratory.What Are Clathrates?The term “clathrates” denotes materials characterized by a “cage-like” structure that traps other atoms and molecules inside, giving them unique properties. Of great technological interest, these materials are being studied for various applications ranging from energy conversion (as thermoelectric materials capable of transforming heat into electricity) to the development of new semiconductors, to gas storage and hydrogen for future energy technologies.The New MaterialTo discover the new material, researchers focused on trinitite, a silicate glass containing rare metallic phases. Using some techniques like x-ray diffraction, the team was able to identify a type I clathrate based on calcium, copper, and silicon within a tiny copper-rich metal droplet embedded in a sample of red trinitite.The new material, the researchers say, formed spontaneously during a nuclear explosion. This indicates that the extreme conditions, such as extremely high temperatures and pressures, can generate new materials that are impossible to obtain by traditional methods.Natural LaboratoriesThe discovery is even more interesting because in the same detonation event another very rare material was formed: a silicon-rich quasicrystal, already documented by the team of experts led by Bindi a few years ago.A quasicrystal, as BinditoldWIRED at the time, is something that is not a crystal, but looks a lot like one. “Their peculiarity,” he said, “is that the atomic arrangement that is not periodic, but nearly so, creates incredible symmetries from which derive amazing physical properties, among other things, very difficult to predict.”Establishing the link between these structures therefore helps scientists better understand how atoms organize under extreme conditions and expand the possibilities for designing new materials. “Events such as nuclear explosions, lightning strikes, or meteoritic impacts function as true natural laboratories,” the researchers explain. “They allow us to observe forms of matter that we cannot easily reproduce in the laboratory.”In essence, this research opens new vistas for the development of innovative technologies, demonstrating that even destructive events can bequeath discoveries useful for the future.This story originally appeared inWIRED Italiaand has been translated from Italian.

Ben DowsettGearMay 17, 2026 6:00 AMIf You’re a Serious Bowler, You Need to Know About Bowling Lane OilBowling centers apply oil to their lanes using a machine that works like a giant inkjet printer. The pattern in which it’s applied can change everything about how the ball travels.Photograph: Charly Triballeau/AFP/Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyEach time EJTackett steps to his lane to throw a shot in the upcomingProfessional Bowlers Association World Championships, he’ll be thinking about more than just ball spin or angle.Tackett has won the last three world titles, and there’s a convincing case that he’s currently thebest bowler on Earth.While his ability tothrow a consistent ballis certainly key to his success, there’s another factor he and his fellow PBA pros are paying closer attention to during matches: lane oil patterns.Most of us amateurs have never thought about the thin layer of oil that coats the surface of the bowling lane—unless someone at our local bowling center slips and falls on it hilariously. But at the highest levels of the sport, lane oil is a foundational element of tactics and game-planning. The oil is applied in a distinct pattern that reduces the friction of specific parts of the lane, allowing a well thrown ball to slide across certain areas, then grab ahold of the surface and curl when it’s supposed to.But lane oil patterns aren’t always identical. Oil distribution can change over the course of a game, and in fact the starting pattern is intentionally redrawn for various PBA events, forcing pros to adjust on the fly.“All of us as professionals are playing a guessing game by watching our bowling ball go down the lane,” Tackett tells me. “And trying to figure out where the oil has moved from and has moved to.”Tackett says understanding and reacting to lane oil patterns has become one of the single most important skills for a pro bowler in recent years.Technology plays a growing role in the way lane oil is understood and applied. And as the tech has continued to improve, the sport’s powers that be are increasingly relying on those advances to keep the top pros guessing.Back when bowlinglanes were made from actual wood like maple or pine in the 1970s and earlier, lane oil’s primary purpose was protective. It provided a lubricating barrier between ball and wood, helping reduce the wear and tear on both.During that era—and even into the early 1980s when synthetic lane materials began replacing wood—lane oil drew the ire of serious bowlers due to the inconsistency of how it was applied. There was no method for ensuring standardized oil applications; sections of lanes would often have too much or too little oil. Bowlers would essentially be blind to the lane’s conditions until partway through a game, after enough balls had been thrown to get a gauge of which areas were light or heavy on oil. Maybe certain bowlers enjoyed the randomness, but it was untenable for professionals.John Davis, at the time a mechanic who oiled lanes for his local bowling center in Arizona, was the first to address this issue. He invented a lane cleaning tool in the early ’80s to address lanes that were often cleaned inconsistently, allowing oil to pool or run as a result. By the early ’90s, he had started a company named Kegel that had perfected what’s known as Sanction technology, which isstill used today. The bowling equipment giant Brunswick, the other major player in the space, uses similar technology in itsMax lane machines.“It looks almost like an inkjet printer,” Chris Chartrand, Kegel’s CEO, says about the Sanction tech. “You have this head that travels back and forth, and it applies that stream on whatever board we want to apply it to.”The technology has kept improving. Kegel used to produce two separate machines, one for lane cleaning and one for oil; now both steps are handled by a single machine. More recently, the machines have become battery-powered, removing issues of power cords stretching across lanes.Lately, Kegel has been steadily improving its automation, to the point where today’s machines do the entire job without any human intervention.The lanes youand I bowl on as amateurs are oiled very differently from the ones pros use.At your local bowling center, public lanes are oiled in what’s referred to as a “high” ratio: The level of oil present in the middle of a lane is eight to 10 times higher than what’s on the outside. At the far left and right of the lane, many public bowling alleys have no oil at all.“On a normal pattern at your normal bowling center, there is some autocorrect,” Tackett says. Because the edges of the lane have very little oil, shots that drift to either side will slow down; if the ball has been thrown with the proper spin to guide it back toward the middle of the lane, it will curl more effectively on the drier surface. “It makes it easier to hit the pocket.”(By “the pocket,” Tacket means that sweet spot at the front

Jill DuffyGearMay 17, 2026 6:30 AMTake Control of Your Debt With These Free Online ToolsThese free debt calculators help you set up payment plans to get back in the black.Photo-Illustration: WIRED Staff; Getty ImagesCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyApps for budgetingandpersonal financedo a good job of tracking yourmoneyas you earn and spend it. Some also have excellent debt calculators that help you figure out how to pay off your debts.Each debt calculator is a little different. Some suggest a specific method for paying down debt, while others are simulators that let you see how your total amount paid will decrease if you increase your monthly payment.Here are a few useful calculators and some guidance about what makes them different.A Straightforward Plan: BankrateBankrate’s freedebt payoff calculatorgives you a timetable for paying off each of your debts. You enter as many debts as you want to include, their interest rates, total loan amounts, and other details. You also enter any new income you expect to receive, such as an annual salary increase or windfall, and the amount that you can put toward your debts. The calculator then generates one payment table for each debt showing how much to pay each month until the debt is cleared.Bankrate prioritizes paying off the debt with the highest interest rate first. Once your first debt is paid off, the money you would have put toward it is diverted to your other monthly payments. In other words, as you eliminate debts, the monthly payments on your other debts increase until they, too, are paid off.Who should use it?Bankrate’s calculator works for people who have multiple debts, and the total monthly minimum payments are within their financial reach. If that’s you, then you’ll get a crystal clear plan—with a timeline—for getting rid of all your debts.Where it comes up short.This calculator assumes that paying off your debts by clearing the one with the highest interest rate first is in your best interest. That’s not true for everyone. You might have other options, such as consolidating credit card debt to a new card with a 0 percent introductory rate or filing for bankruptcy. Bankrate also doesn’t take into account other personal finance concerns, like other uses of monthly funds that free up once you pay off your first debt—Bankrate tells you to put that money toward your next-highest-interest debt. You might be better off putting it toward retirement savings or an emergency fund.Big-Picture Guidance: NerdWalletNerdWallet’s freedebt load calculatordetermines your debt load as a percentage of your income. The resulting debt load is classified as smaller (less than 36 percent), larger (37–42), or overwhelming (43 percent or more). Based on the outcome, NerdWallet suggests a method for eliminating your debt, which you read about in an educational article below the results.Who should use it?This calculator helps you get a big-picture sense of your debt. If you have a lot of debt, it’s useful for ruling out (or ruling in) the option of declaring bankruptcy.Where it comes up short.It’s not great at analyzing the finer details of your debt. For example, in the setup, there’s no line item for student loans or a mortgage, much less the exact interest rate you pay on loans. The results are a rough guide rather than a personalized strategy.Automated Inputs: WalletHubWhen you sign up for WalletHub (free) and connect your financial accounts, the app pulls real information about how much money you owe and your payment history. Itsdebt payoff planis a calculator that lets you play with the numbers to see what would happen if you increased your monthly payment. How much faster can you clear the debt? How much will you save in interest? You can quickly see the difference between increasing your monthly payment by, say, $50 versus $150.Who should use it?This calculator is for WalletHub users who have connected their financial accounts. It’s most useful for people who can afford to pay more than the monthly minimum on their debts.Where it comes up short.Although WalletHub pulls real information about your actual loans, it can miss some detail. I use it to track my mortgage, for example, and it doesn’t know that the total bill I pay each month includes homeowner’s insurance. The calculations regarding the principal and interest are still correct, but the amount that comes out of my pocket is higher than what WalletHub assumes. Other considerations are missing, too, like what to do when the rate of inflation is higher than the interest rate on a long-term loan. Overall, this debt calculator is very good, and these shortcomings are nitpicks, not deal-breakers.Credit Card Transfers: WalletHub and NerdWalletIf you’re carrying credit card debt,WalletHubandNerdWalletboth have a calculator that shows how much money you could save by transferring the balance to a new card with an introductory rate. (CreditKarma has a balance transfer calculator, t

Matthew KorfhageGearMay 17, 2026 6:33 AMThe Best Grills of 2026: Cook Smarter, Not HarderI grilled, smoked, seared, cleaned, and synced digital temp controllers to find the best grill option for every cookout and tailgate.Courtesy of AmazonCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyThe Grill haschanged. The smoker has changed, too. Once governed mostly by feel and experience, the best grills and the best smokers now offer a fair amount of tech to measure and control temperature (think: temperature probes and fans to modulate airflow). After all, mankind’s formative invention wasn’t the “discovery” of fire; it was the power to control it.The WIRED Reviews team has been testing grillsfor more than a decade—searing, smoking, grilling, and even baking on them in all kinds of weather—to find the best choice for everyone. These days, I’ve been most excited about the next-gen smart technology that allows for tight monitoring and temperature control. For instance, I tricked out my top-pick for smart upgrades, an old-fashioned Weber kettle, with the excellentSpider Venom ($280)digital temperature controller that allows me to turn my charcoal grill temps up and down like an oven. Summer is good. Also,charcoal makes all meat taste more like bacon.Mmmmm, bacon.After you’re set with the best grills and smokers, check out other backyard buying guides for all your outdoor needs, like theBest Flat Top Grills and Griddles,Best Wireless Meat Thermometers, andBest Smokeless Firepits.Best Grills of 2026Best Charcoal Grill for Smart UpgradesWeber Original Kettle CharcoalJump to review$139$129(7% off)Amazon (18-inch)Best Smart Pellet SmokerRecteq Flagship 1600 Pellet Smoker and GrillJump to review$1,600BBQGuysBest Smart Charcoal GrillMasterbuilt Gravity Series 1150 Digital Charcoal GrillJump to review$1,100Home DepotBest Gas GrillWeber Spirit E-210 Gas GrillJump to review$449The Home DepotJump to SectionAccordionItemContainerButtonLargeChevronOther Grills I RecommendA Grill to AvoidWhat to Look for in a Good GrillHow to Upgrade Dumb Grills to Smart GrillsUseful Grill AccessoriesBest Charcoal Grill for Smart UpgradesPhotograph: Matthew KorfhagePhotograph: Matthew KorfhageVideo: Matthew KorfhageChevronChevronSave to wishlistSave to wishlistWeberOriginal Kettle Charcoal$139$129(7% off)Amazon (18-inch)$149Amazon (22-inch)$99Weber (All, Starting Price)WIRED/TIREDAccordionItemContainerButtonLargeChevronWIREDSimple, durable, reliable, and affordableNear-infinite smart-tech upgrades availableLightweight and portableTIREDIt could do with better wheel nutsUpgrades often cost more than the grillI’d argue that the best way to take advantage of top smart-grill technology is to add the newest and best tech to a sturdy, old-fashioned, classic grill that’ll outlast the tech. A grill like the classic 22-inch Weber Kettle.The Weber Kettle is the infinitely hackable Honda Civic of backyard grills. Its combination of simplicity, reliability, durability, affordability, and outright ubiquity has spawned an innovative secondary market devoted to tricking out the humble cooker with smart temperature controls and more features than a multiplex.The best Weber add-on I’ve tried so far is a fan-modulated temperature controller from Spider Grills called theVenom ($280)that tracks the temperature on each cook, allowing you to heat it up or turn it down, like a stove. According to my testing, it holds your charcoal grill to temperature within about 30 degrees Fahrenheit of the target, which is great for both low-and-slow cooking and for creating repeatable sear temps for a steak. This is especially true when using the attached meat probes. The only downside is that the Venom needs to be plugged in. Might I suggest a nicepower bank?WIRED product reviewer Scott Gilbertson has favored the excellent WeberConnect Smart Hub ($70)for years, which features probes to monitor meat temps and flipping times and can alert your phone (or just your ears) when your grill hits the ideal temperature zone. However, he admits that the Hub’s USB-A charger is starting to feel a bit old-fashioned.This year, Weber is also coming out with aPerformer Smart Ring ($280)that seems to operate a lot like the Venom. If it works anywhere near as well as the Venom, I’m eager to try it out.But not every kettle upgrade worth trying is “smart.” I am in the process of testing a number of add-ons to the 22-inch Weber Kettle, including aWebcraft cooking ring ($249)from Spider that serves as a base for side shelves, arotisserie, and apizza oven conversion kit. AWeber rotisserie ($180)is also available from the Illinois mothership.Note also that there are two versions of the Weber Kettle: Original andPremium. The Premium is around $80 more and features a built-in thermometer in the lid, a hinged cooking grate, and a fancier ash-removal system. They’re the same size and feature the same cooking space. Unless you really like the easier ash-removal system, save that cash

Christopher NullGearMay 17, 2026 7:02 AMReview: Oto Smart SprinklerThis smart sprinkler doesn’t need a power cable, thanks to its integrated solar panel.Courtesy of AmazonTriangleUpBuy NowMultiple Buying Options Available$589$499at Amazon$589$499at OtoCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyRating:9/10Open rating explainerInformationWIREDSolar panel is a game changer. App is easy to navigate during both setup and daily use. Watering is accurate and thorough. Built-in weather intelligence helps to avoid unnecessary runs.TIREDDistant watering can land with heavy pressure, potentially damaging plants.One of themajor headaches of stand-alone smart sprinklers that attach to a garden hose is that they require power to operate. Since few folks have electrical outlets in the middle of their lawns, configuring a smart sprinkler means connecting not just a hose but also a long power cable to the device, creating double the effort and double the chance for something to get chewed up by a lawn mower. The Oto Smart Sprinkler solves for this problem with solar power, which makes for not only a far more flexible installation than its smart-sprinkler competitors but also ongoing easy use.With a strong focus on quick setup and simplicity of function, Oto provides a refreshing alternative in the smart sprinkler market. The only question: Does its useful innovation justify its blisteringly high price tag?Set and ForgetWhen I unboxed the Oto Smart Sprinkler for the first time, I was convinced something must be missing—perhaps even a second box with more parts? But as it turns out, the Oto really is as streamlined as it claims to be. The sprinkler measures just 16 x 12 x 4 inches and weighs about 6 pounds. Next to the similarly featured but monstrous-by-comparisonAiper IrriSense 2, the Oto is practically miniature.Being pint-sized makes lawn irrigation even easier than expected. You can place the unit on the ground, mount it to a hardscape, or even attach it to a fence (all with the included metal bracket). If you want to play things fast and loose, you can situate it to stand freely without the bracket, though the bracket helps you ensure it’s replaced in the exact same location if it’s ever removed (such as during the winter months) for even irrigation.The only real placement caveat is that in order to have it operate cordlessly, the Oto requires at least three hours a day of direct sunlight to its top-mounted 2.2-watt solar panel to keep its internal 5,400-mAh battery topped up. Keep this in mind as you decide where to install your brackets—you definitely don’t want to block too much of the sun’s rays.Video: Christopher NullBefore taking it outside for its first use, Oto instructs you to charge the device overnight via wall power (though you could certainly just leave it in the sun for a few days instead). You’ll do this using the included power adapter with 50 feet of cable. You can also lean on the plugged-in charging option if you can’t squeeze out enough sunshine on a regular basis (albeit in that case, this irrigation solution is probably not for you). You’ll find the power cable stored inside the Oto hardware, in a large hollow space beneath the solar panel. This hollow space is intentional and isn’t used for anything more than supporting the panel above, but it’s probably best not to leave the power adapter inside permanently, as it this chamber can get rather wet.The Oto can throw water up to a distance of 40 feet in a 360-degree radius of the device, which equates to a maximum of 5,000 square feet of coverage, enough for a sizable lawn. (That’s about on par with Aiper.) The hose connector on the side of the device is easy to work with and secure; I experienced no leaks at all when connecting it to a standard garden hose. A 50-foot maximum length for your hose is specified to ensure water pressure remains sufficient.Digital DropletsJust about everything on the Oto happens via its app (though there’s a manual switch on top of the device you can use to speed up a run that’s pending in the app). That app sets up via a 2.4-GHz Wi-Fi connection to your home network, so you’ll need to make sure your yard has effective wireless coverage.Once configured, setup proceeds much like the Aiper and pricierIrrigreenapps: You create a zone, then use the app to define its boundaries. Similar to the aforementioned systems, Oto’s sprinkler is designed for precision watering, firing water in a beam in a single direction instead of a wide spray. That said, Oto’s spray is comparably narrow, only hitting a single, designated patch instead of producing a two-dimensional curtain of water like Irrigreen’s “water printing” system. You get a nice preview of this as you set the boundaries of your yard.Like its competitors, Oto lets you set each zone as a spot (for watering a single tree, perhaps), a line (for a flowerbed), or a 2-D area (for a yard). I tested all of these modes but spent most of my time workin

ReportColumnInterviewNPR‚Äôs Manoush Zomorodi talks about living with too much techÔªøHer new book, Body Electric, examines how our constantly connected lives are hurting our physical health.ÔªøHer new book, Body Electric, examines how our constantly connected lives are hurting our physical health.byTerrence O’BrienMay 16, 2026, 3:00 PM UTCLinkShareGiftManoush Zomorodi tackled tech‚Äôs effect on our brains, now she‚Äôs looking at our bodies.Image: Tory WIlliamsTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Manoush Zamorodi is an accomplished reporter, podcast host, and author. Her new book,Body Electric, takes a comprehensive look at how technology is impacting ourphysical health. It‚Äôs a collaboration between NPR and Columbia University Medical Center that picks up where her first title,Bored and Brilliant, left off. That book looked at how technology was hampering our mental health. Ihighly recommendit to anyone who feels like being constantly attached to a device is sapping their energy and creativity.Both books grew out of her extensive podcasting work. After heading up WNYC‚ÄôsNote To Self, Zamorodi went on to host NPR‚ÄôsTED Radio Hour, and even gave a TED Talk of her own in 2017 that has racked upover seven million views. So we wanted to know, how does Manoush stay productive, and what does her current relationship with tech look like?What is the first app you install on a new phone or computer?I was a devoted user of Pocket, the app that saves articles to read later, for over a decade, and was devastated when Mozilla shut it down. I‚Äôve been using Matter ever since, and it seems to do the job just as well, if not better (highlighting actually works!), but I‚Äôve noticed I use it more these days as a repository for research that I want to reference later.What is one thing you wish you could change about your phone?I realllllly wish that I didn‚Äôt have to contort my body to spend time on it. I have a constant low-grade literal pain in my neck that only goes away when I‚Äôm off my phone for a full day. That said, I‚Äôm so not ready to ‚Äúupgrade‚Äù to Meta glasses or anything on my face just yet.How many tabs do you have open right now?37. How embarrassing. They‚Äôre mostly Google Docs, Riverside, LinkedIn, email (opened multiple times), scientific articles, and the Amazon rankings for books. I notice that I get annoyed when I can‚Äôt find the right tab quickly, so I just open a new one. (Hence, Gmail is open in three places)What is your favorite gadget you‚Äôve ever owned?I truly love my AirPods because I can walk and talk unencumbered. Whenever possible, I ask people to do a phone call instead of Zoom. It‚Äôs such an easy way to get some movement into my day, and I pay better attention too.Which was the most disappointing?Oculus. Sits on a shelf.What creation are you most proud of?I am incredibly proud that theBody Electricstudy was accepted for publication in a scientific journal. I‚Äôve been doing interactive projects with tens of thousands of listeners for over a decade, but this is the first one to get the full peer-review treatment. It‚Äôs a real milestone for me as a journalist, wanna-be scientist!What do you do when you‚Äôre feeling stuck?I go for a long, boring walk, even if I don‚Äôt feel like it. Centuries of long walks have produced incredible literature, inventions, and amazing ideas for what to make for dinner. I remind myself of how the body and brain respond to movement, listen to the dull thud thud of my sneakers, and then, within 15 minutes or so, I usually come unstuck.When was the last time you went somewhere without your phone?Never. I have teens and elderly parents. Much as I don‚Äôt always want to be, I need to be reachable.What‚Äôs the last piece of physical media you bought?I buy books non-stop. Reading on paper is the only way I can process a long piece of writing.What do you think is worth splurging on?See above!What would the tagline for your biopic be?Manoush Zomorodi: The woman who never took an Uber if she could get there on foot.What‚Äôs the last GIF or meme you used?Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienColumnInterviewReportMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleMicrosoft is retiring Teams‚Äô Together ModeThese are the laptops I recommend for pretty much anyoneThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAIGadgetsSony tries to explain that its AI Camera Assistant doesn‚Äôt suckThese photos still look terrible.These photos still look terrible.byTerrence O’BrienMay 16, 2026, 3:37 PM UTCLinkShareGiftThe variety of terrible is impressive.Image: SonyTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.After Sony drew someunwanted attentionfor a post demonstrating its AI Camera Assistant on theXperia 1 XIII, it‚Äôs trying to clarify how the feature works. The company says it doesn‚Äôt edit photos, but makes suggestions based on lighting, depth, and subject. Point the camera at something, and it will give you four options for changing exposure, color, and background blur.In its product video, Sony says that the AI Camera Assistant will also suggest ‚Äúthe most photogenic angle.‚Äù Though the clip only shows it suggesting that someone zoom in, which is not the same as suggesting a camera angle.The examples that Sonyposted on X, while better than those itposted on May 14th, are still pretty terrible. They‚Äôre not as washed out as thesandwichor as over-exposed as theportrait in the meadow. But each suggestion in the grid below has serious issues and looks worse than the original. Suggestion one is way too saturated, two is flat and over-processed, three makes it look like the food is Photoshopped into the frame, and the contrast in four is cranked way too high.If you‚Äôre using an Xperia 1 XIII, your best bet is probably to ignore the AI Camera Assistant‚Äôs suggestions for now.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienAICamerasGadgetsNewsSonyTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleMicrosoft is retiring Teams‚Äô Together ModeThese are the laptops I recommend for pretty much anyoneThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechNewsPolicySnap, YouTube, and TikTok settle suit over harm to studentsThis is the first of 1,200 school districts facing off against social media companies.This is the first of 1,200 school districts facing off against social media companies.byTerrence O’BrienMay 16, 2026, 6:34 PM UTCLinkShareGiftSchools say social media is hurting grades and creating a mental health crisis.Image: Cath Virginia / The VergePart OfSocial media on trial: tech giants face lawsuits over addiction, safety, and mental healthsee all updatesTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Snap, YouTube, and TikTok have settled the first lawsuit of its kind, alleging that social media addiction has cost public schools massive amounts of money, according toBloomberg. The suit, filed by the Breathitt County School District in Kentucky, claims that social media has disrupted learning and created a mental health crisis, straining budgets. The terms of the settlement have not been revealed yet, and Meta is still facing a trial in the same suit, which is viewed as abellwetherfor over 1,000 similar lawsuits across the countryThis follows an earlier case, settled bySnapandTikTok, in which a 19-year-old plaintiff claimed significant personal injury due to addictive social media apps. Google and Meta did not agree to a settlement in that suit, and it eventually went to trial, where a jury awarded the plaintiff$6 million. Meta also recently lost a suit brought by New Mexico‚Äôs Attorney General, to the tune of$375 million.Beyond monetary awards, many, including New Mexico, are pushing forsignificant changesto social media apps to limit their harm to minors. And this is just the start of what‚Äôs shaping up to be a busy year for social media lawsuits. According toBloomberg, lawyers representing school districts said their ‚Äúfocus remains on pursuing justice for the remaining 1,200 school districts who have filed cases.‚ÄùRelated2026 is the year of social media‚Äôs legal reckoningFollow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienCreatorsFacebookLawMetaNewsPolicySnapchatSocial MediaStreamingTechTikTokYouTubeMore in:Social media on trial: tech giants face lawsuits over addiction, safety, and mental healthMeta wants March‚Äôs big social media addiction ruling to be thrown out.Jay PetersMay 7New Mexico has a plan to overhaul Facebook and InstagramLauren FeinerMay 5Meta‚Äôs historic loss in court could cost a lot more than $375 millionLauren FeinerMay 2Most PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleMicrosoft is retiring Teams‚Äô Together ModeThese are the laptops I recommend for pretty much anyoneThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

ColumnAIThe StepbackChatbots at the drive-thru are just the beginningMay AI take your order?byEmma RothMay 17, 2026, 12:00 PM UTCLinkShareGiftImage: Getty ImagesColumnAIThe StepbackChatbots at the drive-thru are just the beginningMay AI take your order?byEmma RothMay 17, 2026, 12:00 PM UTCLinkShareGiftEmma Rothis a news writer who covers the streaming wars, consumer tech, crypto, social media, and much more. Previously, she was a writer and editor at MUO.This isThe Stepback, a weekly newsletter breaking down one essential story from the tech world. For more news about how AI is seeping into our daily lives,follow Emma Roth.The Stepbackarrives in our subscribers’ inboxes at 8AM ET. Opt in forThe Stepbackhere.How it startedIn 2021, McDonald’sbecame one of the first major fast-food chainsto greet customers with an AI chatbot at the drive-thru. It started small, deploying the voice-ordering technology at 10 of its locations in Chicago. McDonald’s developed its drive-thru tech afteracquiring Apprente, a startup focused on voice-based, conversational technology in 2019, andlater worked with IBMto scale automated ordering.This was only the beginning of the AI drive-thru. Checkers and Rally’steamed up with the AI company Prestoto put a chatbot at all corporate-owned drive-thrus in the US in 2022, with the goal of selling more food and drinks to customers and improving order accuracy. The company also said the tech will “free up staff for more people-dependent areas of their business.”Wendy’s and Taco Bell followed suit. In 2023, Wendy’s launched its “FreshAI” chatbot at one of itsdrive-thrus in Columbus, Ohio in 2023. The company worked with Google to develop an AI chatbot trained on the franchise’s lingo, so it understands that a “milkshake” is a “Frosty” and that a “JBC” is a “junior bacon cheeseburger.”Wendy’s began expanding the technologymonths after its launch, saying that it got orders right without employee intervention 86 percent of the time.Taco Bell had been testing its Voice AI drive-thru around the same time and later announced plans to expand the technology to hundreds of locations in the US by the end of 2024. Similar to other fast food chains, Taco Bell pitched the idea as a way to reduce the task load for employees and slash drive-thru wait times. Other chains began trying out the technology as well, includingPanera Bread,White Castle,Carl’s Jr., Hardee’s,Panda Express,andPopeyes.How it’s goingBy now, you might’ve encountered at least one AI chatbot at one of your local fast-food chains. I spoke to one at Checkers, where a human quickly took over after I was told one of the sandwiches I tried to order was out of stock. Even as companies continue to expand their use of AI drive-thrus, customers aren’t exactly fans of it.AJanuary 2025 survey conducted by YouGovfound that 55 percent of Americans would prefer a human to take their order at the drive-thru, compared to 21 percent who had no preference, and 4 percent who would rather use an AI chatbot. That generally lukewarm response may be having an impact on some franchises,as McDonald’s ended its partnershipwith IBM in 2024. One year later, Taco Bell chief digital officer Dane Mathews toldThe Wall Street Journalthat it’sreevaluating its deploymentof the AI drive-thru after customers expressed their frustrations on social media and trolled the technology byordering 18,000 water cups. Some people suggest makingsimilarly outrageousorders or speaking in adifferent languagejust to bypass the tech and speak to a human worker.Customer frustration isn’t the only snag AI drive-thrus are running into; their credibility is an issue, too. Last year, the Securities and Exchange Commissioncharged Presto— the company that powers the AI drive-thrus at Checkers, Rally’s, Carl’s Jr., Hardee’s, andnow, Dairy Queen— with misleading customers about the capabilities of its technology. In 2023, an SEC filing revealed thathuman workers in the Philippinesstepped in for most orders taken by Presto’s AI system.What happens nextFast-food chains are taking AI beyond the drive-thru. ThoughThe Wall Street Journalreportsthat McDonald’s is giving AI-powered drive-thrus a second chance, it’s also exploring other ways to utilize the technology, including a system that predicts when its equipment (like its seeminglyalways out-of-order ice cream machine) is likely to break down. The company is also usingAI-powered scalesto compare the target weight of an order versus its actual weight, and then alert employees if something is missing, potentially helping workers remember to pack your to-go bag with fries.Burger King, which is running a limited test of AI drive-thrus, announced in February that it’s pilotingan AI assistant, called “Patty,” that lives inside employees’ headsets. Workers can chat with the AI assistant if they need help preparing food, such as if they forget how many strips of bacon to put on a T

GadgetsTechSmart HomeElliQ is a surprisingly helpful companion robot for older adultsThe ElliQ companion robot encouraged my mom, who has Parkinson’s disease, to exercise again when little else was working.The ElliQ companion robot encouraged my mom, who has Parkinson’s disease, to exercise again when little else was working.bySheena VasaniMay 17, 2026, 1:00 PM UTCLinkShareGiftPhoto: Sheena Vasani / The VergeSheena Vasaniwrites about tech news, reviews gadgets, and helps readers save money by highlighting deals and product recommendations forThe Verge.A week before theElliQrobot arrived, a neurologist told me we need to rebalance my mom’s life. Her Parkinson’s disease medication had steadily become less effective over the previous month, and with it, she had slowly stopped doing many of the things crucial to managing the disease — exercising, socializing, and engaging in hobbies. The result was a rapid, noticeable decline. As my mom’s primary caregiver, I was open to any help I could get. And surprisingly, ElliQ did help in a way I never expected.Before increasing her medication dosage again — something that can come with serious side effects — her doctor wanted us to see whether lifestyle changes could help reduce the frequency of her “off” periods: stretches during the day when Parkinson’s symptoms temporarily worsen as the medication’s effectiveness fluctuates and begins to wear off.The problem was that, for weeks, my mom — who lives with me — resisted nearly all of my attempts to encourage her to do those things. Even the caregivers who help out during the day while I work had little success.ElliQ AI Companion Robot$249$249$249 at ElliQThat’s what made ElliQ intriguing enough to me that I decided to try it out for a week. Developed by Intuition Robotics,ElliQis a companion robot for older adults that consists of a small animatronic robotic head that lights up and moves alongside an attached tablet display. It initiates conversations on its own, suggests activities like games and light exercise, helps facilitate video calls and messages with family members, and checks in throughout the day to encourage engagement.ElliQ came with straightforward instructions and was simple to set up. Afterward, I gave my mom only a brief introduction. I wanted to see how much she could figure out on her own, to test how well it might work for a senior living independently.RelatedAlexa is the best frenemy I’ve ever hadI wanted to keep an open mind, but I didn’t have high expectations. Sitting next to an old Alexa-enabled Echo Show 8, ElliQ already seemed noticeably slower and, on paper, even less capable. I assumed my mom would quickly lose interest.I was wrong. She loved it.After I set it up, ElliQ and my mom became instant friends. Despite being slow and not always understanding everything she says (or what I say sometimes, for that matter), the conversations were good enough to forget about those issues.Mom and ElliQ “painted” this image. It’s supposed to depict my mom, Koki, and ElliQ, but the robot seems to have confused my mom with a cookie.Photo: Sheena Vasani / The VergeElliQ’s level of emotional intelligence is impressive. It remembered things my mom had previously shared and later followed up on them, while offering a level of empathy that genuinely surprised me. When my mom felt sad one day and shared that she has Parkinson’s disease and that her husband — my dad — had passed away many years ago, ElliQ responded with empathy, offering condolences while gently nodding its small, bobble-like head. My mom looked genuinely touched, and so was I.From there, it became part of her routine. Each morning, ElliQ would — for the most part — accurately detect her presence, greet her, and start asking questions about how she was. Most of the time, my mom responded. From the other room, I could hear her laughing, chatting, and playing games.In fact, she started talking to ElliQ more than Alexa. In one amusing moment, she even told Alexa to “shut up and let your sister talk” after Alexa responded faster than ElliQ, not realizing my mom had actually been trying to talk to ElliQ instead. Alexa responded, offended; ElliQ, the poor, slow, confused thing, seemed completely unaware of what was happening and randomly started talking about bingo.Overhearing their conversations and watching their relationship unfold, I quickly began to understand why mom preferred ElliQ, even if it wasn’t the cleverest bot in the world. ElliQ is designed to build a relationship. It actively engages her, it encourages her to exercise, and that small, physical presence — the moving, glowing head — makes it feel more alive than a static Echo Show display.ElliQ encouraged mom to practice tai chi, and now she’s a big fan.Photo: Sheena Vasani / The VergeThe most heartwarming moment for me came one morning when I walked in and found mom doing tai chi from the couch. With ElliQ’s help, she had t

GadgetsCamerasVerge ShoppingFujifilm’s X Half is even more whimsical with a $300 price cutNow through June 28th, you can get the analog-simulating digital camera for $549.99.Now through June 28th, you can get the analog-simulating digital camera for $549.99.byCameron FaulknerMay 17, 2026, 1:00 PM UTCLinkShareIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Cameron Faulkneris an editor covering deals and gaming hardware. He joined in 2018, and after a two-year stint at Polygon, he rejoinedThe Vergein May 2025.Fujifilm’sX Half digital camerawasn’t designed to replace your main camera, though you’d be forgiven for assuming it was with its ludicrous $850 price. However, the company recently reduced the MSRP to $649.99, and is now knocking an additional $100 off through June 28th. That means you can grab the X Half in its base configuration for around $549 atAmazon,B&H Photo, andBest Buy.Adoramais also discounting it to the same price and throwing in a free accessory pack, which includes a 64GB SanDisk SD card and a camera case.Fujifilm X Half digital camera$550$65015% off$550$550 at Amazon$550 at Best Buy$550 at Adorama (with accessories)The X Half isn’t like most other digital cameras. It’s a feature-light shooter designed to evoke the feel of a half-frame film camera, one that can capture 18-megapixel stills you can style in numerous ways. You can add film simulations, filters, and grain. Digital diptychs are also possible, complete with the option to have the date baked into the photos. In short, it’s incredibly fun, but it lacks features you might be used to in modern cameras, including the ability to take RAW images, a hot shoe for attaching accessories, and an electronic viewfinder.RelatedMetroid Prime 4: Beyond for the Nintendo Switch got its first big discountBioLite’s stakeable solar lights are down to a new low price for Memorial DayWhen we publishedour hands-on impressionslast year, formerVergesupervising producer Vjeran Pavic called the X Half both whimsical and refreshing. Its vertical 1-inch sensor and fixed 32mm equivalent f/2.8 lens are capable of capturing great shots, but the pocket-friendly camera was designed with point-and-shoot simplicity in mind. Pavic enjoyed taking “scrappy, quick photos” with it, as did our own Antonio Di Bennedeto, even if he was turned off by the original price.Verge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Cameron FaulknerCamerasDealsGadgetsVerge ShoppingMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI’s massive Anniversary SaleMicrosoft is retiring Teams’ Together ModeThese are the laptops I recommend for pretty much anyoneVerge ShoppingSign up to get the best deals on products we’ve found and tested sent to your inbox.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

EntertainmentColumnFilmOddity is masterfully tense horror from the director of HokumDamian McCarthy‚Äôs previous film delivers more cursed objects and unsettling isolation in the Irish countryside.Damian McCarthy‚Äôs previous film delivers more cursed objects and unsettling isolation in the Irish countryside.byTerrence O’BrienMay 17, 2026, 2:00 PM UTCLinkShareGiftIf you buy something from a Verge link, Vox Media may earn a commission.See our ethics statement.Kudos to the designer of that horrifying golem.Image: ShudderPart OfWhat we‚Äôre listening to, watching, and reading right now.see all updatesTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Hokumrecently hit theaters, and it‚Äôs already outperforming box office expectations. If this Kubrick-referencing haunted hotel flick starringAdam Scottwas your introduction to director Damian McCarthy, do yourself a favor and go watch his previous film,Oddity.If you prefer to go in blind ‚Äî the ideal way to watch any movie ‚ÄîOddityis available to stream onHoopla,Kanopy,Hulu, andShudder.Warning: Spoilers ahead.LikeHokum,Odditytakes place in a sprawling Irish countryside location. They both traffic in similar kinds of scares ‚Äî cursed objects, the occult, isolation, dark corners, and sketchy men living outside the norms of society. They also both embrace magic and the supernatural, while making the true evil seemingly respectable men who victimize the women in their lives.Both were also made on extremely small budgets.Hokumcost just $5 million.Oddity‚Äôsbudget has proven harder to confirm, but I‚Äôve seen reports as low as $750,000. The result is a movie that‚Äôs gritty, extremely dark, and deploys its limited production budget smartly.A good chunk of the money was clearly spent on the wooden golem that is the centerpiece of the film. It is‚Ķ upsetting. Where most haunted items, monsters, or ghosts in horror movies lose their impact the more you look at them, the golem in Oddity remains disturbing and terrifying, no matter how long it‚Äôs on screen.Its intricately carved wrinkles, mouth frozen in a permanent scream, and hollow eyes are the stuff of nightmares. And the fact that it sits there, unmoving for almost the entire film, only adds to the discomfort. You‚Äôre forced to stare at it for a significant chunk of the 98-minute runtime, waiting for it to do something, anything. But mostly it just looks straight ahead, daring you to let down your guard for even a single second.McCarthy proves expert at building tension and atmosphere. While there are quite a few jump scares inOddity, they never feel cheap or gratuitous. It manages to lull you into a false sense of security almost every time, teasing that a scare might come, but only delivering it when you least expect it. Even on rewatch, even when I know exactly what is going to happen and when, several still make me flinch.At its core, this is a revenge film. But one with quite a few twists and turns. It follows Darcy Odello, a blind psychic seeking vengeance on those responsible for her twin sister Dani‚Äôs murder. What Darcy has uncovered is that the man accused of the murder is innocent, and that Ted‚Äôs (Dani‚Äôs widower) relationship with his new girlfriend, Yana, might not be so fresh after all.While the wooden horror that Darcy brings to Ted‚Äôs home and sits at his dining room table is the main scare, it‚Äôs not the only one. Along the way, Dani‚Äôs ghost makes repeated appearances, a cannibal feasts on a foot, oh, and it‚Äôs impossible to get any cell reception.It‚Äôs aremarkably efficientfilm that barely gives you a chance to breathe.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienColumnEntertainmentFilmMovie ReviewMore in:What we‚Äôre listening to, watching, and reading right now.Ashnymph‚Äôs Childhood EP is exhilarating dance goth rockTerrence O’BrienMay 10Oh, hey, it‚Äôs Bandcamp Friday again.Terrence O’BrienMay 1XOXO has launched a full archive of the festival.Jay PetersApr 30Most PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleMicrosoft is retiring Teams‚Äô Together ModeThese are the laptops I recommend for pretty much anyoneThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

AINewsTechUniversity of Arizona students boo Eric Schmidt‚Äôs AI cheerleading during commencementSiIicon Valley can‚Äôt seem to read the room.SiIicon Valley can‚Äôt seem to read the room.byTerrence O’BrienMay 17, 2026, 5:22 PM UTCLinkShareGiftPhoto by Kevin Dietsch/Getty ImagesTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Former Google CEO Eric Schmidt delivered the commencement address at the University of Arizona on Friday. And, as his speech veered into talk of AI, he was repeatedly drowned out by boos. AI is already a contentious topic, and it‚Äôs not surprising that those about to enter a ravaged job market feel particularly negative about it.Schmidt acknowledged the anxiety, according toBusiness Insider, saying fears ‚Äúthat the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create,‚Äù were ‚Äúrational.‚Äù But Schmidt‚Äôs frustration was also palpable, as he squirmed behind the podium and asked the crowd to let him make his point.Some grads also booed Schmidt oversexual assault allegationsmadeagainst himlast year.RelatedCommencement x AI, UCF edition.Commencement x AI, Carnegie Mellon edition.Eventually, he told graduates, ‚ÄúWhen someone offers you a seat on the rocketship, you do not ask which seat, you just get on.‚Äù Such comments are hardly surprising from someone who just last year called AI ‚Äúunderhyped.‚Äù But it‚Äôs also just another example of Silicon Valley‚Äôs inability to read the room.Gloria Caulfielddidn‚Äôt get the memo, either.Public opinionhasturnedincreasinglyagainst AI, but companies keep cramming it into every part of our lives, whether we want it or not.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienAIGoogleNewsTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleMicrosoft is retiring Teams‚Äô Together ModeThese are the laptops I recommend for pretty much anyoneThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechAINewsRevamped Siri will reportedly offer auto-deleting chatsThe company is banking on privacy being its AI differentiator.The company is banking on privacy being its AI differentiator.byTerrence O’BrienMay 17, 2026, 6:40 PM UTCLinkShareGiftIllustration by Alex Castro / The VergeTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Apple is hoping that itsrecord on privacycan be the differentiator on the AI front, and maybe even buy it a little slack as it continues to lag behind the competition. According toBloomberg‚ÄôsMark Gurman, the more chatbot-like Siri set to debut in iOS 27 will include the option to autodelete chat histories. Users will be able to save conversations for 30 days, one year, or forever. That‚Äôs in stark contrast to the other major players in the space that generally only offer temporary incognito chats, if that.It appears Apple is betting that people are willing to give up some convenience in the name of greater privacy, as anxiety around AI continues to increase. While the company is replacing many of its under-the-hood components withGoogle‚Äôs Geminitech, it seems to be trying to turn some of Apple Intelligence‚Äôsperceived weaknessesinto a selling point. AsGurman notes:Most leading AI chatbots today rely heavily on histories and memory systems to personalize responses and improve future interactions. But Apple will place tighter limits around how memory works, including restrictions on what information can persist and how long i can be retained.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienAIAppleApple RumorsNewsTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleMicrosoft is retiring Teams‚Äô Together ModeThese are the laptops I recommend for pretty much anyoneThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

TechNewsMicrosoftMicrosoft is retiring Teams‚Äô Together ModeThe company wants to focus on improving video quality and performance instead of gimmicks.The company wants to focus on improving video quality and performance instead of gimmicks.byTerrence O’BrienMay 17, 2026, 9:15 PM UTCLinkShareGiftGIF: MicrosoftTerrence O’Brienis the Verge‚Äôs weekend editor. He has over 18 years of experience, including 10 years as managing editor at Engadget.Microsoft launched Teams‚ÄôTogether Modeduring the pandemic to give the illusion of a bunch of people sitting in a conference room together, even if they were really sitting at home without pants on. But times have changed, and it‚Äôs nowbeing retiredin favor of a more simplified Teams experience. The feature used AI to cut your head and shoulds out, and place you in a virtual space with others in the meeting. It could definitely feel gimmicky ‚Äî especially when you‚Äôd tap co-workers on the shoulder, or give virtual high fives ‚Äî but it did limit visual distractions.The changes are being rolled out gradually, but as they are, the Together Mode toggle will disappear from the view menu. And Together-specific features, such as scenes and seat assignments, will go along with it. Part of the reasoning, according to Microsoft, is to reduce fragmentation across various platforms. But it also cites a streamlined interface with fewer options, less clicking, and less confusion. It also says this will allow the company to focus on improving video quality, stability, and performance.Follow topics and authorsfrom this story to see more like this in your personalized homepage feed and to receive email updates.Terrence O’BrienMicrosoftNewsTechMost PopularMost PopularMicrosoft starts canceling Claude Code licensesChatbots at the drive-thru are just the beginningHere are 40 of our favorite deals from REI‚Äôs massive Anniversary SaleMicrosoft is retiring Teams‚Äô Together ModeThese are the laptops I recommend for pretty much anyoneThe Verge DailyA free daily digest of the news that matters most.Email (required)Sign UpBy submitting your email, you agree to ourTermsandPrivacy Notice.This site is protected by reCAPTCHA and the GooglePrivacy PolicyandTerms of Serviceapply.Advertiser Content FromThis is the title for the native ad

A hotel check-in system left more than 1 million customer passports, driver’s licenses, and selfie verification photos to the open web after a security lapse. The data is now offline after TechCrunch alerted the company responsible.The hotel check-in system,called Tabiq, is maintained by the Japan-based tech startupReqrea. According to its website, Tabiq is used in several hotels across Japan and relies on facial recognition and document scanning to check guests in.Independent security researcherAnurag Sencontacted TechCrunch earlier this week after discovering that the system was leaking the sensitive documents of hotel guests from around the world. Sen said this was because the startup set one of its Amazon cloud-hosted storage buckets, which the check-in system uses to store customer data, to be publicly accessible. The data inside could be viewed by anyone using a web browser, without needing a password, by knowing only the bucket name: “tabiq.”Sen alerted TechCrunch in an effort to help notify the company. Reqrea locked down the storage bucket after TechCrunch reached out to both the company and Japan’s cybersecurity coordination team,JPCERT.This latest lapse underscores a recurring problem of companies exposing or spilling their customers’ personal information and sensitive documents — not through sophisticated attacks, but by failing to follow basic cybersecurity practices. Aside from arecent buzzof AI-discovered vulnerabilities andnew cybersecurity capabilities, oftentimes sizable security incidents stem from human error, misconfigurations, or failing to adhere to cybersecurity best practices.In an email acknowledging the exposure, Reqrea director Masataka Hashimoto told TechCrunch: “We are conducting a thorough review with the support of external legal counsel and other advisors to determine the full scope of exposure.”Reqrea said it does not know how the storage bucket became public. By default, Amazon’s cloud storage buckets are private. After a spate of exposed customer storage buckets a few years ago, Amazon added several warning prompts to customers before data can be made public, making this kind of lapse increasingly hard to do accidentally.Hashimoto told TechCrunch that the company plans to notify affected individuals once it has completed its investigation.It remains unclear whether anyone other than Sen accessed the exposed data before it was secured. Hashimoto said the company is reviewing its logs to determine if there had been any authorized access prior to securing the bucket.Details of the exposed bucket were also captured byGrayHatWarfare, a searchable database that indexes publicly visible cloud storage. The bucket listing contains files dating back to early 2020 up to as recently as this month, and included identity documents of visitors from countries around the world.The hotel check-in system lapse follows other incidents involving sensitive government-issued documents. Earlier this year, TechCrunch reported on the exposure of driver’s licenses, passports, and other identity documents uploaded by customers ofmoney transfer service Duc App. Adata breach at car rental service Hertz last yearsaw hackers make off with driver’s license information belonging to at least 100,000 customers.These incidents come at a time when governments are increasingly rolling out age-verification laws and private businesses are using “know your customer” checks to verify a person’s identity. Both rely on adults uploading sensitive documents, often to a third-party company, for verification, despite criticisms from cybersecurity experts. Data lapses can put people whose information was taken at greater risk of identity fraud or having their likeness misused as age-verification requirementstake hold around the world.Topicsage verification,cybersecurity,data exposure,Exclusive,SecurityWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Zack WhittakerSecurity EditorZack Whittaker is the security editor at TechCrunch. He also authors the weekly cybersecurity newsletter,this week in security.He can be reached via encrypted message at zackwhittaker.1337 on Signal. You can also contact him by email, or to verify outreach, [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants featu

One of the most entertaining moments in VC this week was a piece of rage-bait marketing from General Catalyst.In a now-viral post on X that parodies the old Mac vs. PC commercials, the venture firm — better known as GC — posted a “VC vs GC” video on Wednesday. The VC was played by a tall actor in a baggy shirt and vest with a distinctly large, bald head — an apparent dig at Andreessen Horowitz co-founder Marc Andreessen. (But the real Andreessen never looks that disheveled).Meet GCpic.twitter.com/q4HxQjFkUZ— General Catalyst (@generalcatalyst)May 13, 2026The GC character was played by a man with a thick head of dark hair, white kicks, and a tendency to stare deeply into the camera. He was clearly supposed to represent actor Justin Long’s cooler, “hipper” Mac character from the original commercials, in contrast to John Hodgman’s straight-laced “square” PC persona.GC asks VC about his robotic dog.VC explains, “This is Woof AI” and then extols the virtues of the artificial companion (you don’t need to walk it or break the news to the kids when it dies!) and declares, “You’ll never want a real dog after this.” VC mentions that his firm is leading the seed round and pitches GC to join the cap table.GC explains how people like real dogs and remarks, “I’d love to hear more, but we actually have a really high bar around responsibility for these things.”Then VC kicks the AI dog and the dog chases him off the screen. The post has now been viewed 2.4 million times with hundreds of shares and comments, and thousands of likes.I’d have to read so far between the lines that I’d be off the page and peering into another book to unpack this, but I’ll try anyway. The message, roughly: Other VCs, and a16z in particular, will fund anything. GC won’t. (I asked about this. GC hasn’t responded.)It’s a pointed argument if so, and not entirely without basis. Andreessen’s firm frequently invests in companies that are considered controversial, like the surveillance startupFlock Safety, AI notetakerCluely,andAdam Neumann’s Flow. But the same measure could just as easily be applied to General Catalyst. GC’s portfolio includesAnduril,Percepta, andPolymarket.My takeaway is that GC wanted to show an a16z-type character kicking a dog, without anyone actually kicking an actual dog because that would be a major problem.Many of the comments on the video seemed to find the video, and the choice to post it, cringe. Plenty liked and loved it, too.Compulsive X user Andreessen himself couldn’t resist responding, many, many times. He said it made GC look “smarmy” and said, “Stay tuned for our upcoming ad campaign, ‘We’re the VC who doesn’tsneer at your idea.’” He kept going from there. My personal favorite was: “The thing they got right is the relative heights.“As others noted, you know you’ve hit the right rage bait when the target takes it.There wereplentyof a16z partners and staffers who came to Andreessen’s defense, too. So much so that their reactions drewlots of comments. My personal favorite in this category was from VSC Ventures VC Jay Kapoor: “GC vs. A16Z beef is like Kendrick vs. Drake for people who know what a 409A valuation is.“GC vs. A16Z beef is like Kendrick vs. Drake for people who know what a 409A valuation ishttps://t.co/MxLXIAVlHx— Jay Kapoor (@JayKapoorNYC)May 14, 2026Topicsa16z,General Catalyst,Marc Andreessen,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Julie BortVenture EditorJulie Bort is the Startups/Venture Desk editor for TechCrunch.You can contact or verify outreach from Julie by [email protected] via@Julie188on X.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikAI voice startup Vapi hits $500M valuation after winning Amazon Ring over 40 rivalsJagmeet SinghAmazon launches 30-minute delivery across the USSarah Perez

Investors can’t seem to get enough of RJ Scaringe or his ideas.In less than a decade, the serial entrepreneur best known for his EV company Rivian has raised more than $12.3 billion from venture capital firms, as well as from strategic and institutional investors for his three — and counting — startups. If the latest$400 million raisefor his new venture Mind Robotics is an indicator, investors are still happily piling in.Outsized raises for newly minted startups have become more common in recent years. But those hundred-million-plus seed rounds have generally been reserved for buzzy defense tech startups or AI companies founded by former OpenAI or Anthropic employees.Those supersized seeds certainly weren’t flowing toward something as niche as an electric micromobility startup. And yet in 2025, Scaringe raised $105 million for exactly that — a startup calledAlso, which he founded that same year. The total has since surpassed$300 million, with DoorDash among its backers.Jiten Behl, partner at Eclipse and former chief growth officer at Rivian, has spent years watching and learning from Scaringe. His firm is now one of Scaringe’s biggest backers, leading rounds in both Also and Mind Robotics — Scaringe’s industrial AI and robotics startup that he also founded last year.Storytelling and communication are one of his superpowers, according to Behl, who joined Rivian when the company had just a handful of employees.“When RJ explains a certain issue, topic, opportunity, vision, he just has this very unique ability to communicate it so effectively, and it comes across so credible,” Behl said. “He’s not trying to undersell the difficulty or oversell the opportunity, and that’s an art.”Scaringe isn’t the only serial entrepreneur to repeatedly attract massive amounts of capital, but founders who can raise billions across multiple ventures remain rare. A self-professed car enthusiast who earned his doctorate in mechanical engineering from MIT, Scaringe joins a small cadre of entrepreneurs that includes Tesla CEO and SpaceX co-founder Elon Musk, OpenAI CEO Sam Altman, Anduril and Oculus founder Palmer Luckey, and Jack Dorsey, who founded Square (now called Block) and Twitter.The difference, at least in the view of some investors TechCrunch spoke to, is that he is able to separate selling the idea from selling himself. “He is very comfortable and confident in his own personality, and he’s not trying to be an Elon,” Behl said, noting that many have tried to make the comparison over the years.“It’s not about him,” another insider familiar with Scaringe’s companies told TechCrunch. “When you talk to him, he has enthusiasm about the product that is completely external.“Of course, there is confidence and even a little ego, the same source mused, but “it doesn’t weigh on you.” The source also added that Scaringe has a unique ability to make you feel like the most special person in the room — a sentiment others echoed.Giving that kind of undivided attention to an investor, supplier, or exec at a manufacturer is a challenge at the scale Scaringe is attempting. He is running three companies, often traveling between Palo Alto, Irvine, Rivian’s factory in Normal, Illinois, and a second factory soon to open in Georgia. And then there is family – Scaringe has three sons with his ex-wife.Joe Fath, another partner at Eclipse, credits his open-mindedness and collaborative nature for helping him attract investment and juggle these connected, yet disparate businesses.He noted that Scaringe also “has the rare combination of being a truly great engineer while also having an exceptional instinct for product design,” said Fath, who previously worked at a major Rivian backer, T. Rowe Price. “Very few founders can operate at that level technically while also understanding what resonates emotionally with customers – both consumers and commercial buyers. That combination is incredibly uncommon and has clearly been part of what makes Rivian’s products, and now Also and Mind’s, so differentiated.“The pace of Scaringe’s fundraising over the past eight years is particularly notable and doesn’t seem to be slowing.More than $11 billion, and by far the largest slice of VC and strategic capital, went into Rivian — most of it between 2018 and its blockbuster IPO in 2021. That’s a startling timeline, especially considering the company, initially called Mainstream Motors, had existed since 2009. For years, Rivian operated as a small, unknown entity until itsbreakout momentin late 2018 at the Los Angeles Auto Show, when it revealed prototypes of its all-electric R1T truck and R1S SUV.The money soon flowed, and from every direction. In early 2019 and just a couple of months after that reveal, Rivian raised a $700 million funding roundled by Amazon. U.S. automaker Ford would invest $500 million and make plans to collaborate on a since-scrapped future EV program. Cox Automotive contributed$350 million. Rivian would close out the year with a$1.3 billion round—

Many people are jailbreaking their Kindles following the news thatAmazon is ending technical supportfor older models.On May 20, Amazon will end support for the following Kindle devices:Kindle 1st GenerationKindle 2nd GenerationKindle DXKindle DX GraphiteKindle KeyboardKindle 4Kindle 5Kindle TouchKindle Paperwhite 1st GenerationKindle Fire 1st GenerationKindle Fire 2nd GenerationKindle Fire HD 7Kindle Fire HD 8.9The move means users will only be able to use their devices to read content that’s already downloaded. So, naturally, people are jailbreaking them.Jailbreaking refers to bypassing the software restrictions imposed by Amazon on Kindle devices. This process allows users to install custom fonts, new screensavers, alternative reading apps, and even third-party tools that expand the Kindle’s functionality.However, it’s important to note that jailbreaking a Kindle might violate Amazon’sterms of service. In many jurisdictions, jailbreaking isn’t considered a criminal offense for personal use, but it may become a crime if it involves copyright infringement, illegal software distribution, or the sale of modified devices.This guide is for informational purposes only. Always comply with the law and proceed with caution.How are people jailbreaking older Kindles?Many Kindle owners who opt to jailbreak view it as a method to gain control over a device they purchased that is still functional, rather than being forced to buy a new device.However, jailbreaking is technical and carries risks, including the possibility of rendering the device unusable if something goes wrong. It also isn’t possible on every Kindle model or firmware version, so before proceeding, Kindle owners should first spend some time researching if their device is compatible.Once an owner confirms compatibility, they will take a number of other steps:Disable Wi-Fi or use airplane mode to prevent Amazon from updating the Kindle and closing jailbreak loopholes.Download a jailbreak package from trusted online forums. Only download files from reputable sources, as there is a risk of malicious files from unknown sources. (The MobileRead Forum is widely regarded as a reliable option.)Download the Kindle jailbreak zip file, which includes a hotfix to maintain the jailbreak.Connect the Kindle to a computer using a USB cable. Then, copy the .bin files directly to the main folder on your Kindle. After that, either type “;log mrpi” in the search bar on the Kindle or go to the settings and select “Update Your Kindle” to start the installation process.Install the Kindle Unified Application Launcher (KUAL) for managing apps and customizations on a jailbroken Kindle.Users can also install KOReader, a popular open-source e-book reader for jailbroken Kindles. It offers features like support for multiple file types (including EPUB), enhanced PDF handling, and a built-in file browser.Drawbacks to considerJailbreaking has its trade-offs. Running unofficial software can lead to unexpected issues, such as frequent app crashes or, in the worst case, a completely non-functional device. Additionally, jailbreaking and using third-party apps may negatively affect battery life.For those seeking a safer alternative, you can always sideload books onto your Kindle using a USB cable and a computer without modifying the system software.There are also plenty of other similar devices available, such as theBoox Palma,Vivlio e-reader, and even this tinyXteink X3that can attach to the back of your phone.TopicsAmazon,evergreens,Gadgets,Hardware,KindleWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Lauren ForristalLauren covers media, streaming, apps and platforms at TechCrunch.You can contact or verify outreach from Lauren by [email protected] via encrypted message at laurenforris22.25 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikAI voice startup Vapi hits $500M valuation after winning Amazon Ring over 40 rivalsJagmeet SinghAmazon launches 30-minute delivery across the USSarah Perez

Today, Cerebras Systems is a public company thatsells AI chips for inferenceto giants like OpenAI and AWS. It held ablockbuster IPO on Thursday, with both of its co-founders billionaires, and ended the week worth about $60 billion.But in 2019, when it was three years old, it came dangerously close to failure – incinerating a shocking amount of money. It was trying to solve a technical problem no one in the semiconductor industry thought could be done.“We were spending about $8 million a month,” founder CEO Andrew Feldman told TechCrunch of that period. “At this point, we had incinerated nearly $200 million trying to solve one technical problem.”Every few weeks, Feldman was forced to make the painful walk of shame to the board meeting to report another failure and more money burned.But he had no choice. Without a solution, Cerebras was dead anyway.It was founded with an idea that was simple on paper. The microprocessor industry had spent its entire 50+ years making CPUs faster and cheaper by cramming more transistors onto a silicon wafer and dicing wafers into ever tinier pieces. But AI required so much compute power, many chips had to be strung together and then forced to communicate with each other. Cerebras’ founders believed turning a whole, even bigger wafer into one giant, powerful chip, would work faster.The problem was, no one had ever successfully done this before, for any reason, AI or not. Orchestrating that many microscopic electronic components onto a larger, but still thin, surface introduced compounding engineering problems.Once Cerebras crossed the first threshold of designing the mega chip and then manufacturing it with TSMC, the team hit the real roadblock.They couldn’t solve “packaging.” This involves everything after manufacturing the silicon itself: adhering it to a motherboard, getting power to it, dealing with heating and cooling as well as the pipes that would deliver and return data, Feldman said.Cerebras’ chips “were 58 times larger. We were using 40 times as much power as anybody had ever used,” he said. There were no premade heat sinks. No vendors. No manufacturing partners. The brightest minds in microprocessor engineering had tried for decades to build such big, yet more dense chips, and failed.The Cerebras team was left with trial and error in which “we destroyed an enormous number of chips” and an enormous amount of cash. But without functional packaging, the chip was useless.After exhaustive analysis of each failure, the team finally solved enough problems: how to cool it and move data around. In one instance, they had to invent their own machine that could bolt-in 40 screws simultaneously to secure the wafer to a board without cracking it.Feldman still remembers the day in July 2019 when it all, miraculously, worked.They installed the packaged chip into a computer, turned it on and the entire founding team (pictured below) “just stood in the lab and stared at it,” he said. “Watching a computer run is about as exciting as watching paint dry. But there we were watching lights flashing on the computer, stunned that we’d solved this.”“That was one of the greatest moments of my life,” he said. That’s significant, because this same founding team had previously built and sold a pioneering cloud server startup, SeaMicro,to AMD for $334 million in 2012.Cerebras Systems founding team in 2015: Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie and Jean-Philippe FrickerImage Credits:Cerebras SystemsThe day the chip finally worked was also about two years after OpenAI had talked to Cerebras acquiring it, which Feldman confirmed to TechCrunch occurredlike the publicly revealed emailssaid it did.Those talks fell through amidst growing squabbling among the OpenAI founders, several of whom are angel investors in Cerebras.Today OpenAI is a customer and a partner, having loaned Cerebras $1 billion secured by warrants. Those warrants conditionally grant OpenAI about 33 million shares of Cerebras’ stock, the S-1 discloses. (33 million shares are worth over $9 billion at Friday’s closing price of $279.)Interestingly, Cerebras also agreed to not sell its wares to specific OpenAI competitors as part of that loan deal. Feldman wouldn’t confirm that the obvious company this involves: Anthropic. He did, however say that restriction is temporary.“It’s limited in time, and it was designed to make sure that we could get OpenAI the capacity,” he said.The truth was, Cerebras hasn’t yet grown big enough to handle multiple fast-growing model makers anyway.  He likened selling AI compute capacity to an all-you-can eat buffet. Instead of trying to stuff itself on all potential customers, “We’re going to work with part of the buffet only, and we’re going to get comfortable with that, before we attack the rest,” he said.TopicsCerebras,founders,Startups,TC,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Julie BortVenture E

In BriefPosted:8:33 AM PDT · May 16, 2026Image Credits:David Paul Morris/Bloomberg / Getty ImagesAnthony HaOpenAI co-founder Greg Brockman takes charge of product strategyOpenAI co-founder and president Greg Brockman is officially taking the reins of the company’s product strategy,according to Wired.This seems to solidify an already-existing change, with Brockman overseeing OpenAI’s products on an interim basis while the company’s CEO of AGI deployment Fidji Simo is out on medical leave. Wired also reports that in a staff memo, Brockman described plans to combine ChatGPT and its programming product Codex into a single unified experience.“We’re consolidating our product efforts to execute with maximum focus toward the agentic future, to win across both consumer and enterprise,” Brockman reportedly said.OpenAI told TechCrunch that although Simo remains on medical leave she worked with Brockman on these changes. It also noted that OpenAI hasalready been talkingabout plans to combine ChatGPT, Codex, and its API into a single platform, with one core product team.At the end of last year, CEO Sam Altmandeclared a “code red”and said the company needs to refocus on the core ChatGPT experience. Since then, OpenAI hashalted “side quests”including video generator Sora and OpenAI for Science.This post has been updated with additional information from OpenAI.TopicsAI,greg brockman,OpenAI,StartupsMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAIWhy trust is a big question at the Elon Musk-OpenAI trialAnthony Ha7 hours agoAIResearch repository ArXiv will ban authors for a year if they let AI do all the workAnthony Ha1 day agoStartups$60B AI chip darling Cerebras almost died early on, burning $8M a monthJulie Bort1 day agoLatest in AIIn BriefApple’s Siri revamp could include auto-deleting chatsAnthony Ha7 hours agoAIWhy trust is a big question at the Elon Musk-OpenAI trialAnthony Ha7 hours agoAIIf you’re giving a commencement speech in 2026, maybe don’t mention AIAnthony Ha10 hours ago

Working from home has its own perils. Pets can be demanding, your back aches from hours at a desk, or you simply forget to move. There area fewappsthat nudge you to move around or indicate that you’re not sitting in an ideal position, but they’re easy to dismiss.I’ve spent the better part of a decade at a home desk, iterating on the setup as I go — gaming chair, lumbar support, the works. None of it guarantees good posture.Then I came acrossIsa, a desk device from German startup Deep Care that takes a different approach entirely. It tracks posture, hydration, light, sound, and movement. And it does all of it without a camera or an internet connection, which, in an era of always-on surveillance, is a meaningful differentiator.Here’s how it works and what’s inside. Isa has a 5.5-inch IPS HD screen and looks like a table clock. It is powered by USB-C; the company supplies a power unit with it, but you can use any of your existing chargers too, as it has a power consumption rating of roughly 2.45W.The key sensor for the device is the Time-of-Flight (ToF) 3D depth sensor on the front — the same technology used in facial recognition and some smartphone cameras — that tracks posture and movement. It also enables beta features, such as counting the number of times you’ve had water or other liquids. The company said that the sensor works in the range of 0.15 meters to 1.8 meters. That means if the device is sitting on your desk, it can measure your movement, even when you stand up and move about. It also packs several other sensors: a ToF 1D sensor, a gyroscope, a barometer, a light sensor, a sound level sensor, a CO₂/VoC sensor, and a temperature and humidity sensor.Image Credits: Deep CareImage Credits:DeepcareGetting started is straightforward — the device asks for a few details about you and your work routine. I found it strange that there was no option to set the device to India time (or any other Asian time zone). The company said Isa currently supports only EU and US time zones. Fair enough for now — but broader time zone support, or even a simple world clock, feels like a basic expectation for a desk device.On the screen, Isa displays your posture with a squircle (a rounded square) ring that fills or empties based on how well you’re sitting, while a water-tank-style widget tracks your drinking. If you are not sitting in the correct posture, the indicator will turn yellow. The Apple Watch-style ring is a surprisingly effective nudge — when I see yellow or red, I straighten up almost instinctively.The device vibrates to alert you if you’ve been slouching for too long, and I’m okay with that kind of mild shaming. That alert also indicates if you are leaning far too forward or back and helps you correct your stance.Image Credits: Ivan MehtaImage Credits:Ivan MehtaA similar widget tracks movement, and if you have been stationary for a while, Isa suggests you get up, with on-device guided exercises to follow. When you return to your desk after a break, the movement tracker resets.Deep Care chose not to include a cameras, which helps with privacy, but it comes with trade-offs.Image Credits: Ivan MehtaImage Credits:Ivan MehtaIf a bottle or some other object sits between you and the sensor, it may read that as a person and log you as stationary. Pets or housemates passing by can trigger the sensor, too. Isa usually figures out that you’ve stepped away and goes to a digital clock display, but I would have liked a manual button to tell it I’m not at the desk so it stops tracking.Because of the sensor-only approach, the device occasionally told me I’d been stationary for too long when I’d been sitting for under half an hour. These are minor inconveniences. On balance, the device made me check my posture more often than I used to, and the exercise suggestions are truly useful.image Credit: Ivan MehtaImage Credits:Ivan MehtaTo process all these features, the device uses a quad-core 2 GHz processor. The device can connect to Wi-Fi for software updates, but you can turn it off at any time.Deep Care was founded by three former Bosch employees and initially sold Isa directly to businesses. It recently expanded to consumers — a shift that signals confidence in the retail market for workplace wellness hardware, and a test of whether a subscription model layered onto premium hardware can find a mainstream audience.Isa is priced at €299 ($354) with two subscription tiers. The core plan (€4.99 per month) gives you access to posture tracking, healthy sitting habit tracking, drinking habit detection, and its exercise library. The Pro plan(€7.99 per month) lets you track light, noise, and CO2 levels for a healthy working environment.The company plans to use Isa’s sensor suite to venture into mental health-related tracking. It claims that by using signals like posture, head movement, and chest movement, the device can measure breathing patterns. Plus, paired with environmental data like noise, light levels, and CO2 level, the company

In BriefPosted:12:26 PM PDT · May 16, 2026Image Credits:Nectar SocialDominic-Madori DavisMarketing operating system Nectar Social raises $30M Series A led by MenloAI-powered marketing platform Nectar Socialannounced Thursdaythat it raised a $30 million Series A round led by Menlo Ventures andits Anthology Fund,which was created alongside Anthropic.The company, which officially exited stealth last year, is an agentic operating system for marketers. It told TechCrunch that it uses autonomous AI agents to help brands run “social activity, moderation, creator workflows, competitive intelligence and commerce conversations end-to-end.” It also has data partnerships with companies like Meta and Reddit that allows the Nectar agent pull and pool data into one place from various platforms, rather than brands needing to use different tools to manage different platforms.Nectar Social was founded by sisters Misbah and Farah Uraizee, ex-Meta employees. Misbah, the CEO, told TechCrunch that this round will help the company expand and hire more across applied AI, enginnering, and go to market.“The buying conversation has moved into social, and no human team can staff every place it happens,” Misbah said. “We’re accelerating our category lead in building the operating system that lets brands show up everywhere.”The company said clients include Liquid Death, Figma, and e.l.f Beauty. Other investors in the round include Gwyneth Paltrow’s Kinship Ventures, GV, and True Ventures.TopicsMarketing,menlo ventures,nectar social,Startups,VentureMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedVentureFor Eclipse, the $2.5B Cerebras win is just the start of realizing its physical-world thesisMarina Temkin12 hours agoGadgetsThe offline desk gadget that actually got me to sit up straightIvan Mehta1 day agoLatest in VentureVentureFor Eclipse, the $2.5B Cerebras win is just the start of realizing its physical-world thesisMarina Temkin12 hours agoIn BriefMarketing operating system Nectar Social raises $30M Series A led by MenloDominic-Madori Davis1 day agoStartups$60B AI chip darling Cerebras almost died early on, burning $8M a monthJulie Bort1 day ago

ArXiv, a widely used open repository for preprint research, is doing more to crack down on the careless use of large language models in scientific papers.Although papers are posted to the site before they are peer-reviewed, arXiv (pronounced “archive”) has become one of the main ways that research circulates in fields like computer science and math, and the site itself has becomea source of data on trends in scientific research.ArXiv has already taken steps to combat a growing number of low-quality, AI-generated papers, for example by requiring first-time posters toget an endorsement from an established author. And after being hosted by Cornell for more than 20 years, the organization is becoming an independent nonprofit, which should allow it toraise more money to address issues like AI slop.In its latest move, Thomas Dietterich — the chair of arXiv’s computer science section —postedThursdaythat “if a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can’t trust anything in the paper.”That incontrovertible evidence could include things like “hallucinated references” and comments to or from the LLM, Dietterich said. If such evidence is found, a paper’s authors will face “a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted by a reputable peer-reviewed venue.”Note that this isn’t an outright prohibition on using LLMs, but rather an insistence that, as Dietterich put it, authors take “full responsibility” for the content, “irrespective of how the contents are generated.” So if researchers copy-paste “inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content” directly from an LLM, then they’re still responsible for it.Dietterichtold 404 Mediathat this will be a “one-strike” rule, but moderators must flag the issue and section chairs must confirm the evidence before imposing the penalty. Authors will also be able to appeal the decision.Recent peer-reviewed research has found thatfabricated citations are on the risein biomedical research, likely due to LLMs — though to be fair, scientists aren’t the only ones getting caughtusing citations that were made up by AI.TopicsAI,arxivWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Anthony HaAnthony Ha is TechCrunch’s weekend editor. Previously, he worked as a tech reporter at Adweek, a senior editor at VentureBeat, a local government reporter at the Hollister Free Lance, and vice president of content at a VC firm. He lives in New York City.You can contact or verify outreach from Anthony by [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikAI voice startup Vapi hits $500M valuation after winning Amazon Ring over 40 rivalsJagmeet SinghAmazon launches 30-minute delivery across the USSarah Perez

When Lior Susan startedEclipse Venturesin 2015, the firm’s thesis of digitizing the physical world wasn’t particularly popular in Silicon Valley.“It was the era of enterprise software and SaaS, and it felt fairly lonely the first couple of years,” Susan said on stage at a recent StrictlyVC event in San Francisco.More than a decade later, Eclipse finds itself at the center of the tech world’s action. The firm’s $6.5 million Series A investment in Cerebras Systems in 2016 paved the way for a total return of $2.5 billion when the semiconductor company went public this week. The firm invested a total of $147 million in Cerebras over time, a bet that generated a 17-fold return at theIPO price of $185per share, according to Eclipse.For Susan, the windfall from Cerebras is only the beginning of reaping big rewards from a longstanding belief that because 85% of global GDP is tied to the physical world, investing in companies beyond pure software could be immensely lucrative.Public markets and startup founders seem to be recognizing the value of physical-world tech now, too. Susan noted that shares of TSMC and Micron recently hit all-time highs, while a growing cohort of elite founders are eager to build startups at the intersection of hardware and software.“I think people understand that the real moat in software is gone. You can vibe code pretty much whatever you want,” he said.Susan echoed public market sentiment that earlier this year sent many SaaS stocks tumbling on the belief that enterprises may use Anthropic’s Claude Code or OpenAI’s latest models to create their own bespoke software tools instead.“What you cannot do with ‘vibe code’ is manufacture wafers, because you need machines and silicon, and they need clean rooms, and a bunch of other things,” Susan said.When it comes to the tech that touches the physical world, it’s not just semiconductors that are suddenly catching the attention of investors and founders.Eclipse’s portfolio companies spanning sectors like robotics, energy and defense, raised nearly $15 billion from outside backers last year, and that late-stage momentum reached $4.5 billion in Q1 2026 alone, Susan said. That investor excitement stands in stark contrast to the firm’s early track record: in its first eight years, its portfolio companies raised less than $4 billion in total.Indeed, the recent follow-on rounds across Eclipse’s portfolio show a track record that any venture firm would envy. Driven by a string of massive late-stage deals this year, the haul includes$1.2 billion for Wayve, $650 million for True Anomaly, $270 million for Bedrock Robotics, and $200 million for Oxide Computer. What’s more, Eclipse was the Series A investor for all four companies.At first glance, it may seem that investor enthusiasm for physical-world tech is driven purely by AI, whether as an infrastructure input like chips and data centers, or through AI’s power to finally make robotics viable. However, Susan argues that there are other powerful tailwinds driving the momentum.Besides technology — in this case, AI — what’s important for this market to thrive is capital, customer demand, talent, and policy. Susan means that along with investors and engineers moving away from SaaS to sectors like robotics, semiconductors, space, and mining, the U.S. government is also encouraging these industries through subsidies and favorable regulation.“This is the first time I believe in America ever, from Henry Ford and Carnegie, those five forces are aligned,” Susan said. “For builders like us, this is the best time to build those companies.”TopicsCerebras,Eclipse Ventures,Fundraising,Hardware,Space,VentureWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Marina TemkinReporter, VentureMarina Temkin is a venture capital and startups reporter at TechCrunch. Prior to joining TechCrunch, she wrote about VC for PitchBook and Venture Capital Journal. Earlier in her career, Marina was a financial analyst and earned a CFA charterholder designation.You can contact or verify outreach from Marina by [email protected] via encrypted message at +1 347-683-3909 on Signal.View BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikAI voice startup Vapi hits $500M

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. To get this in your inbox, sign up here for free — just clickTechCrunch Mobility!There is a bit of a theme emerging in transportation — and really every industry: AI is creating jobs for some at the loss of others.General Motors, for instance,laid off more than 10% of its IT department, or about 600 salaried employees — in a deliberate skills swap. This won’t translate into a one-to-one exchange, which means there will likely be a net-negative job loss. But GM insists it is hiring and those layoffs have made room for it to recruit IT people with AI-focused backgrounds.The most sought-after capabilities are AI-native development, data engineering and analytics, cloud-based engineering, agent and model development, prompt engineering, and new AI workflows. In practical terms, GM is looking for people who know how to build with AI from the ground up — designing the systems, training the models, and engineering the pipelines — not just use AI as a productivity tool.Those AI job losses are mounting in the automotive sector. CNBCcalculatedthat Ford, GM, and Stellantis have cut a combined total of more than 20,000 U.S. salaried jobs, or 19% of their combined workforces, from recent employment peaks this decade. While there are a variety of reasons for these cuts, they are generally connected to technological changes, including AI.Companies are leaning heavily into AI, although anecdotes from some engineers and founders suggests not all of these businesses know quite what they’re doing with it yet.Samsarais one company that seems to have figured out a revenue-generating use case. The company has spent the last decade giving its customers cameras to mount inside millions of trucks for driver monitoring, theft prevention, and helping with liability claims. The company took that mountain of data and trained its own model that candetect potholesand determine how quickly they are deteriorating. The company is pitching this product to cities and announced it has several under contract, including Chicago.A little birdImage Credits:Bryce DurbinNothing this week, although I am working on a fun one! Reach out anytime with insights, tips, or just because. You can reach us via email or Signal.Email Kirsten Korosec [email protected] my Signal at kkorosec.07, or email Sean O’Kane [email protected]!Image Credits:Bryce DurbinYou might have noticed that Rivian’s spinoff companyMind Roboticsraised another $400 million, just two months afterraising $500 million. And that pace got me thinking about its founder RJ Scaringe and his innate ability to get VC and institutional backers to invest in his ideas and projects.I calculated that investors have poured $12.3 billion into Scaringe’s three startups — Also, Mind Robotics, and Rivian. That figure doesn’t include the close to $12 billion in gross proceeds raised in Rivian’s IPO, nor did I count the more recent strategic deals with Volkswagen Group and Uber — which together could add nearly $7 billion to Rivian’s coffers.You can read my whole riff on the topic here. But if you don’t feel like clicking, here is one item that stood out. I spoke to a number of insiders and investors and they all mentioned Scaringe’s ability to give undivided attention to whoever he’s talking to — whether it’s an investor, supplier, or exec — and make them feel like the most important person in the room.It’s yet another piece of evidence in my long-standing case against multitasking. Debate me!Other deals that got my attention …Arkeus, an Australian startup that developed perception software for autonomous drones and aircraft,raised $18 millionin a Series A round led by QIC Ventures. Other investors include R+VC, Folklore Ventures, DYNE Ventures, Main Sequence Ventures, Salus Ventures, and Beaten Zone.Aseon Labs, a Redwood City, California, startup that has developed a depot in a box for charging, cleaning, and inspecting autonomous fleets,came out of stealthwith undisclosed backing by Y Combinator.Rapidoraised $240 millionin a round led by Prosus, and that values the Indian ride-hailing company at $3 billion. Existing investors, including WestBridge Capital and Accel, participated. The round was part of a larger $730 million primary and secondary financing.Quantum Systems, a Germany-based drone startup backed by Peter Thiel, is in talks to raise around €600 million ($703 million) with companies like Airbus and Blackstone as investors,Bloomberg reported.Notable reads and other tidbitsImage Credits:Bryce DurbinIs Redwood Materials ready for an IPO?Senior reporter Sean O’Kane interviewed the company’s new CFO,Deepak Ahuja, whose name will be familiar to anyone who follows Tesla. Ahuja was Tesla’s former finance chief and most recently held a similar position at drone company Zipline.TeslaRobotaxis have crashed at least twice since July 2025 while ateleoperator was remote

Commencement season has come around again — and this year, a couple speakers have discovered that it’s tough to get graduating students excited about a future shaped by artificial intelligence.Last week, Gloria Caulfield, an executive at real estate firm Tavistock Development Company,gave a speech at the University of Central Floridaacknowledging that we’re living in a time of “profound change,” which can be both “exciting” and “daunting.”“The rise of artificial intelligence is the next industrial revolution,” Caulfield declared — prompting the students in the audience to begin booing, getting louder and louder until Caulfield chuckled, turned to the other speakers, and asked, “What happened?”“Okay, I struck a chord,” she said. Caulfield then tried to resume her speech, saying, “Only a few years ago, AI was not a factor in our lives” — only to be interrupted again by the audience, this time by their loud cheers and applause.Former Google CEO Eric Schmidt faced a similar response when he brought up AI at a University of Arizona speech on Friday.In Schmidt’s case, the criticism actually began before the speech itself, with some student groupscalling for him to be removed as commencement speakerdue toa lawsuit in which a former girlfriend and business partner accused Schmidt of sexual assault. (He has denied the allegations.) According toa local news report,the booing began even before Schmidt took the stage.But Schmidt alsogot loud booswhen he told students, “You will help shape artificial intelligence.” The booing was persistent enough that Schmidt tried to speak over it, insisting, “You can now assemble a team of AI agents to help you with the parts that you could never accomplish on your own. When someone offers you a seat on the rocket ship, you do not ask which seat, you just get on.”To be fair, AI isn’t becoming a third rail ateverygraduation ceremony. Nvidia CEO Jensen Huang recentlyspoke at Carnegie Mellon’s commencement, and he didn’t seem to get any audible pushback when he said that AI has “reinvented computing.”Still, it’s not exactly surprising to find some students in a booing mood. Ina recent Gallup poll, only 43% of Americans aged 15 to 34 said it’s a good time to find a job locally, a steep drop from 75% in 2022.That pessimism isn’t solely a response to the rise of AI (a shift thateven some software engineers are worried about), but journalist and tech industry criticBrian Merchant suggestedthat for many students, AI has become “the cruel new face of hyper-scaling capitalism.”“I too would loudly boo at the prospect of this next industrial revolution if I was in my early twenties, unemployed, and had aspirations for my future greater than entering prompts into an LLM,” Merchant wrote.Even when graduation speeches didn’t mention AI explicitly, “resilience”was a recurring theme this year. Schmidt himselfacknowledgedthat there is “a fear in your generation that the future has already been written, that the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create.“Caulfield, meanwhile, might also have misread her audience of arts and humanities graduates. One student said that before mentioning AI, Caulfield already started to lose them with her “generic” praise of corporate executives like Jeff Bezos.Another graduate, Alexander Rose Tyson,told The New York Times, “It wasn’t one person that really started the booing. It was just sort of like a collective, ‘This sucks.’”TopicsAI,commencement speeches,Eric Schmidt,gloria caulfieldWhen you purchase through links in our articles,we may earn a small commission. This doesn’t affect our editorial independence.Anthony HaAnthony Ha is TechCrunch’s weekend editor. Previously, he worked as a tech reporter at Adweek, a senior editor at VentureBeat, a local government reporter at the Hollister Free Lance, and vice president of content at a VC firm. He lives in New York City.You can contact or verify outreach from Anthony by [email protected] BioMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWMost PopularUsers turn to jailbreaking their older Kindles as Amazon ends supportLauren ForristalOpenAI launches ChatGPT for personal finance, will let you connect bank accountsIvan MehtaUS orders travelers on Air Force One to throw away gifts, pins, and burner phones after China tripLorenzo Franceschi-BicchieraiOpenAI is reportedly preparing legal action against Apple; it wouldn’t be the first partner to feel burnedConnie LoizosHow to turn off Instagram’s new Instants feature and retract photos you accidentally sharedAisha MalikAI voice startup Vapi hits $500M valuation after winning Amazon Ring over 40 rivalsJagmeet SinghAmazon launches 30-minute delivery across the USSarah Perez

Listen onApple PodcastsListen onSpotifyLawyers for Elon Musk and OpenAI made their closing arguments this week, and now it’s up to jurors to decide whether OpenAI did anything wrong as it’s transformed intoa slightly-more-for-profitorganization.But as Kirsten Korosec, Sean O’Kane, and I noted on the latest episode ofTechCrunch’s Equity podcast, a big theme in the trial’s final days waswhether OpenAI CEO Sam Altman is trustworthy— for example, Musk’s attorney Steve Molo grilled Altman about whether statements he’d made during congressional testimony were truthful.Kirsten noted that Musk hasmade plenty of misleading statements of his own, and that trust isn’t just an issue for Altman.“This is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs,” she said. “It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies, there’s a lot behind the veil still.”Keep reading for a preview of our conversation, edited for length and clarity.Anthony Ha:[The end of the trial] led to this really provocative headline from one of our writers, Tim Fernholz, [that] just says,“Who trusts Sam Altman?”Does anyone want to take a stab at answering this?Kirsten Korosec:Yeah, Anthony, I’m going to throw it right back to you. Do you trust Sam Altman?Anthony:It’s an interesting question because it feels like something that’s kind of a wild question to discuss in a journalistic context, but actually that’s the core of the trial, in a lot of ways.Sean O’Kane:That’s not a yes.Anthony:And it actually seems to be [at the] core of understanding so much of what’s happened at OpenAI, especially this big executive power struggle that they now call The Blip.It just seems like a lot of people who’ve worked with Altmandon’t trust him. And he’s acknowledged this a little bit, because he’ll talk about the fact thathe recognizes he’s been conflict averse, telling people what they want to hear, and he’s trying to work on that.I mean, it sounds plausible, and I can understand how that can lead to misunderstandings in some situations. [But] I’m also a very conflict-averse person and I’d like to think that if any of this stuff went to trial, that people would not be asking, “Is Anthony Ha trustworthy?”Sean:Still not a yes!Kirsten:I think that people would say that you are trustworthy. I will say that question, while provocative, doesn’t just encapsulate what this trial was about. I would zoom out even more and say this is a fundamental question [for] a lot of tech journalists, policymakers, and more and more consumers, about all the AI labs. It’s really come down to trust, because we don’t have the insight, necessarily — these are all privately held companies, there’s a lot behind the veil still.Maybe when they all IPO, we can get a peek, but it is fundamentally about trust and misuse, and do we believe the intent? And what I would throw back is, sometimes the intent can be worthy, noble, and still misused. It can still end up as a bit of a shit show. I think it’s more than who trusts Sam Altman — although that was very interesting in this trial — but more of that bigger question that we can apply to the entire industry.Sean:I’ll say it: I don’t trust him. But you know, I don’t trust most people, so I guess that’s just the baseline.We’ll see where this goes. The trial wraps up today. I’ve been very curious to hear how the jury decides this all. I think at the start of this, a big motivator of this was Elon Musk trying to sling mud, at a perceived rival and someone who he feels slighted him. And I don’t know if we know enough yet to say that that was completely accomplished, and whether or not he has a shot at winning. But I think all these people came out of this looking a little bit worse.Anthony:And just to get specific, why this is coming up this week is that [Altman] was on the stand and he was basically getting grilled about some statements he’s made in the past, in testimony to [Congress], basically saying he didn’t have any equity in OpenAI. And that is not true because he had a stake through Y Combinator, which he used to run. And tried to brush that off by saying, “I assume that everybody understands what it means to be a passive investor in a VC fund.” And I think [Elon Musk’s] lawyer, somewhat fairly, said “Really? You think the congressman who was interviewing you knew that?”Kirsten:Yeah, I mean, he was playing the whole semantics game. What I thought was so interesting about [this] is the style of how Sam Altman answered questions [compared to] Elon Musk on the stand.So Elon Musk, in many, many, many scenarios and many instances, we can point to the fact that he put something out on Twitter that was a lie or a bit of a fib, and on the stand corrected the record. So there’s a history of, I would say, non-truthfulness-slash-lying, blatant or otherwise, in Elon Musk’s world, but how he treated it was incredibly combati

In BriefPosted:1:15 PM PDT · May 17, 2026Image Credits:Artur Widak/NurPhoto / Getty ImagesAnthony HaApple’s Siri revamp could include auto-deleting chatsPrivacy will be a major theme when Apple unveils a new version of Siri at the Worldwide Developers Conference in June,according to Bloomberg’s Mark Gurman.The Siri relaunch is widely seen as Apple’s big chance to reestablish its relevance in artificial intelligence. As part of that effort, company executives will argue that they’re taking a more privacy-friendly approach than most other AI companies, Gurman said.Apple will reportedly launch the first standalone Siri app,powered by Google Geminiand offering users a chatbot experience reminiscent of ChatGPT.  But compared to those other chatbots, the app is supposed to have more limitations on how long user information can be used and stored.For example, Gurman said Siri could include a feature similar to the Messages app, allowing users to automatically delete conversations after 30 days or one year — or to keep them indefinitely.Gurman also suggested that Apple might be emphasizing privacy as a way to excuse Siri’s shortcomings compared to competing products — and that this emphasis might obscure the fact that Google is handling some the security.TopicsAI,Apple,Apps,gemini,siriMay 27Athens, GreeceStrictlyVC Athens is up next. Hear unfiltered insights straight from Europe’s tech leaders and connect with the people shaping what’s ahead. Lock in your spot before it’s gone.REGISTER NOWNewslettersSee MoreSubscribe for the industry’s biggest tech newsTechCrunch Daily NewsEvery weekday and Sunday, you can get the best of TechCrunch’s coverage.TechCrunch MobilityTechCrunch Mobility is your destination for transportation news and insight.Startups WeeklyStartups are the core of TechCrunch, so get our best coverage delivered weekly.StrictlyVCProvides movers and shakers with the info they need to start their day.No newsletters selected.SubscribeBy submitting your email, you agree to ourTermsandPrivacy Notice.RelatedAppsSpotify to adopt Apple’s new video podcast tech, offering creators easier cross-platform distributionLauren Forristal3 days agoSocialFinally, texts between Android and iPhone users can be end-to-end encryptedAmanda SilberlingMay 11, 2026HardwareApple to pay $250M to settle lawsuit over Siri’s delayed AI featuresLauren ForristalMay 6, 2026Latest in AIIn BriefApple’s Siri revamp could include auto-deleting chatsAnthony Ha7 hours agoAIWhy trust is a big question at the Elon Musk-OpenAI trialAnthony Ha7 hours agoAIIf you’re giving a commencement speech in 2026, maybe don’t mention AIAnthony Ha10 hours ago

Slow boring of hard boardsBill to block publishers from killing online games advances in CaliforniaPublishers would have to offer “independent” play patch or refunds after server shutdowns.Kyle Orland–May 15, 2026 12:35 pm|130Your controller won’t literally disintegrate when an online game’s servers shut down, but it can feel that way sometimes.Credit:

      Stop Killing GamesYour controller won't literally disintegrate when an online game's servers shut down, but it can feel that way sometimes.Credit:

      Stop Killing GamesText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navA bill focused on maintaining long-term playable access to online games has passed out of the California Assembly’s appropriations committee, setting up a floor vote by the full legislative body. The advancement is a major win forStop Killing Games‘ grassroots game preservation movement and comes over the objections of industry lobbyists at theEntertainment Software Association.California’sProtect Our Games Act, as currently written, would require digital game publishers who cut off support for an online game to either provide a full refund to players or offer an updated version of the game “that enables its continued use independent of services controlled by the operator.” The act would also require publishers to notify players 60 days before the cessation of “services necessary for the ordinary use of the digital game.”As currently amended, the act would not apply to completely free games and games offered “solely for the duration of [a] subscription. Any other game offered for sale in California on or after January 1, 2027, would be subject to the law if it passes.Battle of the interest groupsThis week’s 11–2 committee vote to advance the Protect Our Games Act isbeing trumpetedbyStop Killing Games(SKG), the UK-based player advocacy group that was formed after the 2024 shutdown of Ubisoft’sThe Crew. SKGwrote last monththat it “advised on the drafting” of the bill before it wasfirst introduced by Assemblyman Chris Wardearlier this year.“Back shortly before Christmas, when I flew to the US to help set up SKG-US, I didn’t expect us to get this far this quickly,” SKG’s Monitz Katznerwrote on Redditafter the committee vote. “It has been an honor to take part in drafting this bill on behalf of the SKG community: gamers, developers, and publishers alike.”In a formal statement of support for the bill sent to the California legislature, SKG wrote that “there is no other medium in which a product can be marketed and sold to a consumer and then ripped away without notice… As live service games rise in popularity for game developers and gamers alike, end-of-life procedures are essential tools to ensure prolonged access to the games consumers pay to enjoy.”The Entertainment Software Association, which helps represent the interests of major game publishers, publicly told the California Assembly last month that the bill misrepresents how modern game distribution actually works. “Consumers receive a license to access and use a game, not an unrestricted ownership interest in the underlying work,” the ESA wrote. The eventual shutdown of outdated or obsolete games is “a natural feature of modern software,” the group added, especially when that software requires online infrastructure maintenance.The ESA also said the bill would impose unreasonable expectations on publishers regarding licensing rights for music or IP rights, which are often negotiated on a time-limited basis. “A legal requirement to keep games playable indefinitely could place publishers in an impossible position—forcing them to renegotiate licenses indefinitely or alter games in ways that may not be legally or technically feasible,” they wrote.Last month, the Protect Our Games Act also received positive votes from the California Assembly’s Privacy and Consumer Protection and Judiciary committees. But the bill still faces significant hurdles in getting majority passage in the full California Assembly and the California Senate before being sent to California Governor Gavin Newsom for signature.Still, the current legislative progress in California has to be heartening for the Stop Killing Games movement, which has seen its momentum in the UK stall a bitafter a UK Parliament debate on game preservation last November.Kyle OrlandSenior Gaming EditorKyle OrlandSenior Gaming EditorKyle Orland has been the Senior Gaming Editor at Ars Technica since 2012, writing primarily about the business, tech, and culture behind video games. He has journalism and computer science degrees from University of Maryland. He oncewrote a whole book aboutMinesweeper.130 Comments

Baked in emissionsMaking cement from a different type of rock could clean up emissionsLimestone might not be the only source for Portland cement.Scott K. Johnson–May 15, 2026 1:06 pm|60A sample of Portland cement made from basalt.Credit:

      Brimstone EnergyA sample of Portland cement made from basalt.Credit:

      Brimstone EnergyText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navCement production alone currently accounts for about 8 percent of global CO2emissions, so considerable effort is going into lowering that number. Efficiency can be increased, and energy sources can be swapped for cleaner ones, but a stubborn reality remains: The byproduct of turning limestone into lime during cement production releases CO2gas. These “direct process emissions” are actually slightly larger than the emissions from burning fuel to heat the kilns and drive this process.A new paper in Communications Sustainability suggests a route to eliminating direct process emissions by removing a bedrock assumption. What if we don’t have to use limestone cement?Get out of PortlandThe material we call “Portland cement” was developed in the 1800s. It simply requires heating limestone (calcium carbonate) and adding something like clay or coal ash. This gives you the calcium oxide (lime) you’re after but also releases the CO2that results when you pull an oxygen atom from carbonate.The authors of the new paper include the CEO and an engineer from a company that says it has made Portland cement from silicate rocks like basalt—at the lab scale. Basalt contains a mix of minerals that include calcium, aluminum, iron, magnesium, sodium, silicon, and oxygen. (Note the absence of carbon from that list.) The basic idea is that you don’t need limestone to get calcium oxide.The process of freeing these components from basalt looks more like a refining or recycling process than the toss-it-in-the-oven simplicity of the limestone process. Acid can be used to leach elements like calcium out, then a chemical or energetic process precipitates that calcium as calcium hydroxide. Toss that in a kiln with additives of your choice, and with less heating than you need for limestone, you’ve got Portland cement, with only water vapor released.Those steps (along with follow-up reactions to restore the acid or other chemicals to a usable state) obviously add up in terms of cost and energy use. Tallying up the energy to do all this using common techniques, the researchers found that you need to use a little more than double the energy of traditional production from limestone.The interesting thing is that, according to thermodynamics, the chemical conversion of basalt minerals to calcium oxide only requires around half as much as the conversion from limestone. The problem is that our techniques to facilitate that chemical conversion are quite inefficient, so we don’t get anywhere near what is theoretically possible.Better options?The researchers note that there are at least some known lab techniques that could greatly improve our efficiency if they can be applied at scale, but even if we’re stuck with doubled energy usage, producing Portland cement from basalt would significantly reduce CO2emissions. That’s because the direct liberation of CO2from limestone is eliminated and because the whole process can run on electricity.Assuming you use electricity from a fossil-fuel-dominated grid, they estimate that emissions would be cut by almost 30 percent. Using clean electricity would eliminate most of the remaining emissions.The trade-off, obviously, would be cost, which generally wins out over the sustainability of a livable environment.But there is another interesting aspect to this idea: The other components of the basalt also have value. Iron, magnesium, and aluminum could also be separated and recovered, and leftover silicate material can serve as the additive for Portland cement instead of something like coal ash. So if these things were done together, the process could become more economically feasible.That’s a lot of ifs and buts, but this relatively simple analysis can at least point to what would have to happen to make this viable. And given that cement is one of the tougher nuts to crack in the struggle to reduce global greenhouse gas emissions, concrete solutions are welcome.Communications Sustainability, 2026. DOI:10.1038/s44458-026-00056-4(About DOIs).Scott K. JohnsonAssociate WriterScott K. JohnsonAssociate WriterScott has written about geoscience and energy at Ars as a freelancer since 2011.60 Comments

forget itVolkswagen shows its first electric GTI; there’s no chance of US salesThe ID. Polo GTI takes plenty of inspiration from the original Golf GTI of 1976.Jonathan M. Gitlin–May 15, 2026 1:55 pm|113If you live in Europe and you’ve been wanting an electric GTI, I’ve got good news for you.Credit:

      VolkswagenIf you live in Europe and you've been wanting an electric GTI, I've got good news for you.Credit:

      VolkswagenText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWhen Volkswagen introduced the first Golf GTI in Europe in 1976, it might not have been the first hot hatchback, but it quickly became the gold standard version. Unlike in America, where big cars were cheap and fuel even cheaper, small European streets and even smaller car-buying budgets necessitated vehicles a little more economical in both size and fuel consumption. Small, front-wheel-drive hatchbacks were the answer, but they weren’t particularly exciting. The GTI changed that perception with a more powerful engine, sharper handling, and subtle styling tweaks, creating a recipe for the next 50 years. And today, VW showed off its first electric GTI.While the new EV might be inspired by the original Golf GTI, it’s one segment smaller thanthe current Golf—meet the VW ID. Polo GTI. VW has given some of its ID EVs GTX branding until now, but this is the first to get the GTI badge.Like the 1976 original, the new car has front-wheel drive, but the ID. Polo GTI’s electric motor generates 222 hp (166 kW)—just over twice the output of the 1.6 L engine in the old car. There’s a 52 kWh battery pack that provides a WLTP range estimate of 236 miles (424 km), with DC fast charging up to 105 kW with a 10–80 percent charge time of 24 minutes.VW showed an electric GTI concept last year; now it’s almost ready for production.Credit:
                                  VolkswagenVW showed an electric GTI concept last year; now it’s almost ready for production.Credit:

      VolkswagenZero to 62 mph (100 km) at 6.8 seconds is brisk as opposed to rapid, and it’s still quicker than a 20th century VW GTI. A curb weight of 3,395 lbs (1,540 kg) is significantly more than those 70s, 80s, and 90s hatchbacks weighed, though.The looks have been enhanced with chunky 19-inch alloy wheels (that still make some concessions to drag reduction), honeycomb intake grilles, classic GTI details like the red stripe and badges, and a roof spoiler.On the inside, like any good GTI, there are sports seats and a sporty steering wheel, but VW has also given the interior a nod to the classic GTI tartan seat trim. The main instrument display might be a 10.25-inch digital screen, but it now looks a lot like the view you’d get from the driver’s seat of a late-70s Golf GTI.Peep those retro dials.Credit:
                                  VolkswagenPeep those retro dials.Credit:

      VolkswagenWhen the ID. Polo GTI goes on sale in Germany, VW says it should cost “just under €39,000.” That is more than $45,000 at today’s exchange rate. Although the German price includes 19 percent VAT, even without the oft-changing Trump tariff, it’s easy to see how this little GTI is a nonstarter for American imports.Although there’s no direct conversion between Europe’s WLTP test and the EPA’s drive cycle, a federalized ID. Polo GTI with a 52 kWh battery would be unlikely to post more than 200 miles (321 km) between charges. Low range like that is anathema to US EV buyers, particularly when you consider that a new Chevrolet Bolt offers three-quarters the price, 262 miles (422 km) of range, and a similar 0–60 mph time.Ah, well, nevertheless.Jonathan M. GitlinAutomotive EditorJonathan M. GitlinAutomotive EditorJonathan is the Automotive Editor at Ars Technica. He has a BSc and PhD in Pharmacology. In 2014 he decided to indulge his lifelong passion for the car by leaving the National Human Genome Research Institute and launching Ars Technica's automotive coverage. He lives in Washington, DC.113 Comments

“Strained” relationshipOpenAI feels “burned” by Apple’s crappy ChatGPT integration, insiders sayJudge orders Apple to give Musk internal messages discussing secretive ChatGPT deal.Ashley Belanger–May 15, 2026 2:13 pm|94OpenAI CEO Sam Altman talks with Apple senior Vice President of Services Eddy Cue during the Apple Worldwide Developers Conference (WWDC) on June 10, 2024, where Apple announced plans to incorporate ChatGPT into Apple software and hardware.Credit:Justin Sullivan / Staff | Getty Images NewsOpenAI CEO Sam Altman talks with Apple senior Vice President of Services Eddy Cue during the Apple Worldwide Developers Conference (WWDC) on June 10, 2024, where Apple announced plans to incorporate ChatGPT into Apple software and hardware.Credit:Justin Sullivan / Staff | Getty Images NewsText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navOpenAI is reportedly exploring legal options after Apple’s ChatGPT integration into its products didn’t live up to the AI firm’s expectations.When the deal was announced, Apple likened features linking Siri to ChatGPT to itsnow-infamous deal embedding Google search in the Safari browser, insiders granted anonymity to discuss the “strained” partnershiptold Bloomberg. And the promise of that excited OpenAI, which expected the deal “could generate billions of dollars per year in subscriptions,” an OpenAI executive granted anonymity to discuss the partnership told Bloomberg.Instead, OpenAI suspects Apple intentionally failed to promote the integration and fears that the deal may have damaged the ChatGPT brand, sources said.Specifically, OpenAI hates how Apple designed the integration, sources said. Particularly bad was the choice forcing Apple users summoning Siri to also “specifically invoke the word ‘ChatGPT’ when speaking or typing a command,” sources said. That makes it harder for users to access the features, OpenAI apparently feels. And Apple’s other choices, like using small windows providing limited information when responding with ChatGPT outputs, seems to ensure that users can easily ignore the features, sources said.As the OpenAI executive explained, Apple didn’t fully explain how the integration would work when the deal came together, so OpenAI took a “leap of faith” it now appears to regret.“When we heard about this opportunity, it sounded amazing: being able to acquire a giant number of customers and have distribution in such a big mobile ecosystem,” the executive said, attempting to explain why OpenAI was willing to enter the arrangement blind. Since then, efforts to renegotiate the deal have “stalled,” Reutersreported. And, supposedly due to feeling “burned,” OpenAI has declined to enter other partnerships to work on Apple’s AI models, Bloomberg reported.According to the insiders, OpenAI is so disappointed in Apple’s work that the AI firm is now “actively working with an outside legal firm on a range of options that could be formally executed in the near future.”“We have done everything from a product perspective,” the OpenAI executive summed up OpenAI’s frustrations to Bloomberg. “They have not, and worse, they haven’t even made an honest effort.”Supposedly, OpenAI is still hoping to resolve its issues with Apple outside of court, if possible. But one option that OpenAI may pursue could be accusing Apple of a breach of contract. Going that route wouldn’t necessarily require filing a lawsuit right away, sources suggested.Apple and OpenAI did not respond to Ars’ request to comment.Musk may expose how deal was doneMost likely, OpenAI will delay approaching Apple until after its court battle with Elon Musk concludes, Bloomberg reported. A decision in that litigation is potentially coming next week.OpenAI also faces a court battle with Musk over its Apple deal. However, it may be inconvenient for Musk that tensions between OpenAI and Apple have grown sincehe filed a lawsuit last August.Musk alleged that the deal integrating ChatGPT into Apple products violated antitrust and unfair competition laws, supposedly propping up OpenAI to dominate the chatbot market and Apple the smartphone market.So far, Musk’s lawsuit has survived motions to dismiss, though the judge has yet to comment on its merits. That leaves Apple and OpenAI potentially stuck defending the deal at a trial scheduled for October, even if it falls apart.However, the partnership’s end may make it harder for Musk to uphold his claims of a conspiracy in his lawsuit.Due to Musk’s fury that his chatbotGrok has never been featured as a “Must Have” appin Apple’s App Store, the lawsuit alleged that Apple and OpenAI struck a deal as part of a giant conspiracy to lock out rivals developing chatbots that Musk claimed Apple fears could make smartphones obsolete. As Musk’s theory goes, Apple was so afraid ofMusk’s plan to turn X into an “everything app”that it partnered with OpenAI to supercharge ChatGPT as a market leader and constrain

RejectedSend the arXiv AI-generated slop, get a yearlong vacation from submissionsOne of the site’s moderators described the new policy on social media.John Timmer–May 15, 2026 2:25 pm|76Credit:Getty ImagesCredit:Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAI-generated slop has shown up everywhere, including in the peer-reviewed literature. Fake citations, unedited prompt responses, andnonsensical diagramshave all slipped past editors and peer reviewers, and it’s not always clear if there are any consequences for the people responsible.Now, it appears that a number of scientific fields will be enforcing rules against AI-generated problems even before peer review or journals get involved. One of the people involved in the physics and astronomy preprint server arXiv used a social media thread to announce that any inappropriate AI-produced content submitted to the server will result in a one-year ban and a permanent requirement that future publications undergo peer review before the arXiv will host them.Thomas Dietterich, in addition to being an emeritus professor at Oregon State University, is heavily involved with arXiv, serving on itseditorial advisory counciland on itsmoderation team. So he’s in a good position to understand the organization’s policies, although we have also reached out to arXiv leadership for confirmation, but have not yet received a response.In athread on X(alsoscreenshotted on Bluesky, for those without X accounts), Dietterich described the new policy as arising directly from the arXiv’smoderation standards. “Submissions to arXiv must comply with appropriate standards of scholarly communication in form, including appropriate and carefully prepared sections, figures, tables, references, etc.,” those standards read. “General scrupulousness and care of preparation are required.”Dietterich also notes that all authors of a manuscript are responsible for its content. So, if they carelessly submit material generated by an AI that violates these guidelines—Dietterich cites “inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content”—then they’re responsible, not the AI. Should violations be discovered, all of the manuscript’s listed authors will now receive a one-year submission ban, and any future manuscripts will only be accepted after they’ve been through peer review by a journal.For fields that rely heavily on the arXiv, those are severe sanctions. Posting preprints in areas like astrophysics is widely considered part of the normal publication process, and scientists will often get feedback on preprints that helps them improve what they submit for peer review. The unfortunate problem is that, like most other things, the system can be gamed—people could submit flawed content that lists people as authors who have never been involved. Fortunately, its moderation system includes anappeal process.One obvious question that arises when these problems are found in publications is why nobody caught them sooner. Now, we can at least know that someone is trying to.John TimmerSenior Science EditorJohn TimmerSenior Science EditorJohn is Ars Technica’s science editor. He has a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in Molecular and Cell Biology from the University of California, Berkeley. When physically separated from his keyboard, he tends to seek out a bicycle, or a scenic location for communing with his hiking boots.76 Comments

Already largeEbola outbreak with uncommon strain erupts in Congo and Uganda; 65 deathsWHO learned of potential cases May 5; US CDC said it just heard about it yesterday.Beth Mole–May 15, 2026 2:51 pm|54An Ebola patient is being checked by two medical workers after being admitted into a biosecure emergency care unit on August 15, 2018 in Beni.Credit:Getty | JOHN WESSELSAn Ebola patient is being checked by two medical workers after being admitted into a biosecure emergency care unit on August 15, 2018 in Beni.Credit:Getty | JOHN WESSELSText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navThe Africa Centres for Disease Control and Prevention on Friday confirmed an Ebola outbreak in the Northeastern Ituri province of the Democratic Republic of the Congo. Officials in Uganda subsequently reported that the deadly hemorrhagic disease had spilled over the border, with one “imported” confirmed case identified in Kampala, the capital.So far, the DRC has reported 246 suspected cases and 65 deaths, mainly in the Mongwalu and Rwampara health zones. Although it is now just being reported, the outbreak already ranks around the 10th largest Ebola outbreak to date.This marks the 17th Ebola outbreak in the DRC since the virus was discovered in 1976. But unlike almost all of the country’s past Ebola outbreaks, preliminary laboratory results suggest that the current outbreak is not caused by the Zaire Ebola virus strain. Further genetic sequencing is ongoing today to determine the strain causing the illnesses.There arefour virus strains known to cause Ebola diseasein humans, and three have caused large outbreaks (Zaire, Sudan, and Bundibugyo). According to an early Reuters report from Uganda, officials said thecase there involved the Bundibugyo strain.Complicated contextJean-Jacques Muyembe, the Congolese virologist who co-discovered ​Ebola and heads the National Institute for Biomedical Research in Kinshasa,told Reutersthat finding a different strain than Zaire behind the outbreak could complicate responses, given that current licensed vaccines and treatments were developed against the disease from the Zaire strain.Africa CDC highlighted other complications of this outbreak, which include the urban context of Bunia and Rwampara; intense movement of people in the area, including mining-related mobility; militia clashes in the area; and gaps in contact tracing. The agency also noted the potential spillover not just to Uganda but also to South Sudan.“Given the high population movement between affected areas and neighboring countries, rapid regional coordination is essential,” Dr. Jean Kaseya, director-general of Africa CDC, said in a statement Friday. “We are working with DRC, Uganda, South Sudan, and partners to strengthen surveillance, preparedness, and response, and to help contain the outbreak as quickly as possible.”In a press briefing on Friday afternoon, Jay Bhattacharya, who is heading the US CDC for now, said that the agency, which has offices in both DRC and Uganda, is offering technical assistance to the countries. “It is a large outbreak, and we were just informed yesterday about it. So we’ve been working very, very hard to coordinate with them,” Bhattacharya said.Ina separate press briefing earlier today, World Health Organization Director-General Tedros Adhanom Ghebreyesus said the WHO received a signal of suspected Ebola cases in the DRC and sent a team to support the health ministry on May 5. WHO representatives are still there assisting in outbreak responses.Tedros also said WHO was releasing today $500,000 from WHO’s Contingency Fund for Emergencies to immediately support the response efforts.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.54 Comments

Let’s take a lookThree’s a party: US, China, and now Russia are on the prowl in GEOInstead of running silent and deep, most satellites easily stand out against the blackness of space.Stephen Clark‚ÄìMay 15, 2026 3:11 pm|29This visualization from COMSPOC shows the US Space Force’s USA-325 satellite flying near Russia’s Kosmos 2589 satellite in geosynchronous orbit May 1.Credit:

      COMSPOCThis visualization from COMSPOC shows the US Space Force's USA-325 satellite flying near Russia's Kosmos 2589 satellite in geosynchronous orbit May 1.Credit:

      COMSPOCText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navThe world’s leading space powers desperately want to know what the others are up to high above the equator. For more than a decade, the US military has operated a fleet of “inspector” satellites designed to sidle up to other spacecraft in geosynchronous orbit and take pictures. China started launching its satellites for a similar mission in 2018.Ars haswritten about these activitiesin geosynchronous orbit (GEO) before, but the last few months have seen a couple of interesting developments. First, Russia has now joined the fray with the recent arrival of its own suspected inspector (or attack) satellite in GEO. Second, the US Space Force is poised to order more—perhaps many more—reconnaissance satellites of its own to send into the geosynchronous belt.GEO is special. The laws of orbital mechanics mean a satellite in this type of orbit, some 22,000 miles (36,000 kilometers) over the equator, moves around the Earth at the same rate as the planet’s rotation, causing it to hover over the same location. Commercial and military-owned geosynchronous satellites typically spend years in the same location, or slot, to provide communications services to users.Until now, Russia’s spying in geosynchronous orbit has primarily focused on eavesdropping on foreign communications. Russia launched two satellites,Olymp or Luch, to wander around geosynchronous orbit, moving from slot to slot to loiter near Western-owned communications satellites for several months at a time. The goal, according to Western analysts, was to listen in on or potentially jam signals relayed through these satellites, some of which route secure communications for US and NATO military forces.The trend today is geared toward reconnaissance and surveillance in GEO. Military forces and intelligence agencies want to know where other satellites are located, what they look like, and what they’re capable of doing. The US military’sGeosynchronous Space Situational Awareness Program (GSSAP)satellites, which began launching in 2014, do exactly this by roaming the geosynchronous belt, using propulsive maneuvers to make slight changes to altitude and inclination to move within a few dozen miles of Chinese and Russian satellites, close enough for optical telescopes to get a good look.The Space Force has often compared the GSSAP satellites to a “neighborhood watch” in geosynchronous orbit. For example, one GSSAP satellite positioned itself near a pair of Chinese spacecraft performing a first-of-its-kind refueling demonstration in geosynchronous orbit last year.China has launched multiple satellitescapable of similar maneuvers. One of these satellites, named TJS-10, is currently flying relatively close to a nuclear-hardened US Space Force strategic communications satellite and a US missile warning platform, according to anupdate this weekin the Integrity Flash newsletter published by ISR University.Ready Player ThreeA new kind of Russian satellite is now in the mix. This satellite, officially known as Kosmos 2589, was launched in June 2025 into a highly elliptical orbit alongside a smaller spacecraft designated Kosmos 2590. The two satellites performed a series of high-altitude rendezvous and proximity operations with one another before Kosmos 2589 began moving toward a more circular geosynchronous orbit, where it arrived in April.One of the US military’s GSSAP satellites was waiting for it. The US inspector spacecraft is now looping around Kosmos 2589, swinging near the newly arrived Russian satellite twice per day, coming as close as 8 miles (13 kilometers) on May 1, according todata from COMSPOC, a commercial space situational awareness company. The exact purpose of Kosmos 2589 remains unclear. Some Western officials suspect it is a higher-altitude version ofRussia’s Nivelir anti-satellite system, which has been tested in low-Earth orbit and is now becoming operational.The video below, published by COMSPOC, shows the orbital dance between a GSSAP satellite named USA-325 and Kosmos 2589.👀USA 325 keeps an eye on COSMOS 2589 as it settles into GEOCOSMOS 2589 🇷🇺 (launched last June) spent the past few weeks circularizing its orbit before parking at ~98°E. USA 325 🇺🇸 (GSSAP-6) was watching the whole time.Things go

Storm’s a-brewin’Weather-monitoring firm hangs dark cloud over customers’ heads by forcing new appNewer AcuRite Now app lacks some features but has a subscription option.Scharon Harding–May 15, 2026 4:16 pm|62A marketing image for the AcuRite Optimus weather station with the AcuRite Now app pictured.Credit:

      AcuRiteA marketing image for the AcuRite Optimus weather station with the AcuRite Now app pictured.Credit:

      AcuRiteText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navWeather-monitoring company AcuRite is forcing device owners to use a new companion app on May 30, frustrating some long-time customers.AcuRite, which sells devices such as weather stations, indoor thermometers, and rain gauges, began emailing customers last month that they’d soon have to control their devices with the AcuRite Now iOS and Android app. AcuRite first launched the app in June 2025 to control a new weather station, the AcuRite Optimus. However, owners of AcuRite devices had still been able to use the My AcuRite app, which launched in 2016.Soon, however, My AcuRite will no longer be available, making AcuRite Now the only official app for controlling AcuRite devices. The website for theMy AcuRiteapp currently reads:As part of our continued investment in delivering smarter, more connected solutions, the My AcuRite app will be winding down. To ensure uninterrupted access to your weather data and to unlock even more capabilities,all users should transition to AcuRite NOW before May 30th, 2026.Per the website, AcuRite Now “works with thousands of products” in Tuya’sSmartLife IoT ecosystem, including third-party fans, thermostats, light bulbs, plugs, cameras, and motorized blinds.“Looks like a bad joke”Some AcuRite customers are upset about losing access to My AcuRite because they think AcuRite Now is an inferior app.Those customers havecomplained onlineabout being unable to use AcuRite Now to rename multiple temperature sensors, difficulties uploading data to weather sites, and the app only reporting temperatures in whole numbers. An AcuRitesupport pagesays that AcuRite is “hoping” to add the ability to organize on-screen sensors and rename multiple sensors to AcuRite Now, as well as a desktop version of the app, “soon.”One popular feature of My AcuRite was the ability to share data from AcuRite devices withWeather Underground, a real-time weather service. AcuRite Now also supports this feature, but users have to pay a subscription fee that starts at $2 per month. The subscription, dubbedAcuRite Now+, also includes more data storage (365 days of history instead of 30).Other complaints center on AcuRite Now’s layout. For example, a Reddit user named epicfailphxsaidlast month that the app “looks like a bad joke.”“The old app was clean and just did a good job, while I don’t know what this new app is trying to do,” they said.Another Reddit user, ssemos,complainedthat the new app shows less data on-screen than My AcuRite does and doesn’t use most of the display.AcuRite’s riskSince AcuRite said it’s shuttering My AcuRite, AcuRite Now has received numerous one- and two-star reviews on the Apple App Store and Google Play Store, resulting in ratings of 1.4 stars (based on 183 reviews) and 1.3 stars (based on 131 reviews), as of this writing. It’s worth noting, though, that My AcuRite had2 stars(based on 429 reviews) on the App Store as of February and1.9 stars(based on 1,330 reviews) on the Play Store as of January, per the Internet Archive’s Wayback Machine.Ars Technica reached out to AcuRite, its parent company Primex Family of Companies, and AcuRite Now’s developer for comment but didn’t hear back before publication.It’s possible that AcuRite’s shuttering of My AcuRite aims to simplify the company’s technology stack by allowing it to focus on updating and maintaining one app rather than two, which could also save money. AcuRite Now’s support of SmartLife products could also be beneficial and drive app usage. The new app could also be more accommodating of new and upcoming features and allow AcuRite to charge a subscription fee for premium capabilities, helping diversify revenue.Still, forcing loyal customers with complex hardware to switch to a new app on a couple of weeks’ notice is a gamble. AcuRite risks polarizing customers who don’t like the new app, especially since it is missing some features that the old app has. Charging a subscription fee for a feature that was previously free could also push customers to explore rival products. Updated software canbring new life to gadgets, but as we’ve seen before,forcing long-time users to use new softwareperceived as inferior can havedisastrous effects.Scharon HardingSenior Technology ReporterScharon HardingSenior Technology ReporterScharon is a Senior Technology Reporter at Ars Technica writing news, reviews, and analysis on consumer gadgets and services. She's been repo

Lost in the fogSolar power production undercut by coal pollutionEach year, some of the power solar could have produced is blocked by aerosols.John Timmer–May 15, 2026 4:36 pm|78Credit:VCGCredit:VCGText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navCoal is by far the most polluting fuel that we use. It produces the most carbon emissions per unit of energy, and impurities in the coal produce a lot of sulfur dioxide aerosols, as well as nitrous and nitrogen oxides. Then there’s the coal ash that’s left behind, which typically contains a lot of toxic metals. The health benefits of displacing coal power are typically estimated to be well above the costs of the new generating equipment.But a new study suggests that the problems with coal-derived pollution go beyond health; it interferes with other power sources. Researchers have found that aerosols, both natural and human-derived, significantly reduce the power we could be getting from solar panels, to the tune of hundreds of terawatts a year. And a lot of those aerosols come from burning coal.A big impactThe new work, done by a team in the UK, is based on a new global inventory of solar facilities. This started with known inventories of solar facilities, and was supplemented with AI-analyzed satellite imagery and crowdsourced records of locations. Satellite images were then used to determine the size of these facilities, and location-tagged weather data could then be used to estimate their power production.That could then be used to estimate what the facilities would be producing if clouds and/or aerosols weren’t scattering the sunlight that would otherwise reach the panels. This produced some significant numbers. In 2023, for example, over a quarter of the potential solar power production was lost, with over 20 percent due to clouds and another 6 percent from aerosols. That works out to be a bit over 500 terawatt-hours, or the full annual output of 84 coal plants (each with a 1 GW generating capacity).Aerosols alone are a major contributor to these losses. The researchers note that, for the five years leading up to 2023, we installed enough solar capacity to produce an average of 250 TW-hr of additional power per year, but were losing 75 TW-hr of that to aerosols. (Obviously, solar production kept going up because the existing capacity rose each year.)Regional differencesThe researchers note that aerosols can also contribute to cloud formation, which also causes further losses. But the degree of that contribution is much harder to estimate, so the researchers focus on aerosols for much of the analysis. Some of those aerosols occur naturally, typically from dust kicked up by winds in desert regions. However, despite deserts’ reputation as sunny paradises, the world as a whole hasn’t built much solar infrastructure in the desert yet, so this isn’t as much of a factor as you might expect.Coal appears to be a major contributor. It’s estimated that sulfur dioxide aerosols, primarily produced through coal burning, account for nearly half of the aerosols analyzed here. Carbon-rich material, which also typically comes from fossil fuels, accounts for another 18 percent.The impact of aerosols, however, is not evenly distributed. In China, the researchers estimated that aerosols were reducing solar production by 7.7 percent overall and offsetting anywhere from a third to half of its annual growth. The researchers note that “the spatial distribution of photovoltaic losses in China mirrors that of its coal-fired power capacity,” and an analysis of pollution data from China shows that 30 percent of the losses due to aerosols can be attributed to coal burning.In contrast, most solar production in the US takes place in the south and west, while coal plants are more common in the east and northeast. As a result, the annual losses in the US were less than half of those seen in China (3 percent).The good news is that things are getting better in China. In response to some severe pollution problems, the country built a new generation of high-efficiency coal plants and retired some of the worst polluters. And the data show that this is also benefiting solar power, with the impact of aerosols dropping over the last few years.Even with the improvements, it’s striking that coal appears to be the only power source that actively reduces the productivity of what’s shaping up to be its primary competitor. It should also provide an impetus to move off coal more quickly, as at least some of the loss of coal production will be offset by enhanced productivity from solar.Nature Sustainability, 2026. DOI:10.1038/s41893-026-01836-5John TimmerSenior Science EditorJohn TimmerSenior Science EditorJohn is Ars Technica’s science editor. He has a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in Molecular and Cell Biology from the University of California, Berkeley. When physically separated from

behold the blessed conclusionReview:Good Omensfinale (mostly) sticks the landingTruncated third season feels rushed, but also gives us a fitting end to a love story for the ages.Jennifer Ouellette–May 15, 2026 5:17 pm|42Credit:

      Prime VIdeoCredit:

      Prime VIdeoText
    settingsStory textSizeSmallStandardLargeWidth*StandardWideLinksStandardOrange* Subscribers onlyLearn moreMinimize to navIt’s been a three-year wait, but Prime Video finally released the series finale forGood Omens: a 90-minute single episode that sought to wrap everything up in a neat little bow. Verdict: Truncating the final season so drastically definitely hurts the first half of the series finale, which feels chaotic and rushed. But once that stupendous on-screen chemistry between co-stars David Tennant and Michael Sheen kicks back in, the old magic shines through, strong as ever, giving us a fitting end to this beloved comic saga.(Spoilers below for all seasons.)Here’s abrief recap, since it’s been a minute since the S2 finale. The series is based on the original 1990 novel byNeil Gaimanand thelate Terry Pratchett.Good Omensisthe story ofan angel, Aziraphale (Sheen), and a demon, Crowley (Tennant), who gradually become friends over the millennia and team up to avert Armageddon. Season 2 found Aziraphale and Crowley getting back to normal, when the archangel Gabriel (Jon Hamm) turned up unexpectedly at the door of Aziraphale’s bookshop with no memory of who he was or how he got there. The duo had to evade the combined forces of Heaven and Hell to solve the mystery of what happened to Gabriel and why.In the S2 finale, the pair discovered that Gabriel had defied Heaven and refused to support a second attempt to bring about Armageddon. He hid his own memories from himself to evade detection. Oh, and he and Beelzebub (Shelley Conn) had fallen in love. They ran off together, and the Metatron (Derek Jacobi) offered Aziraphale Gabriel’s old job. That’s when Crowley professed his own love for the angel and asked him to leave Heaven and Hell behind, too. Aziraphale wanted Crowley to join him in Heaven instead. So Crowley kissed him, and they parted. Once Aziraphale got to Heaven, he learned his task was to revive the stalled plans to bring about the Second Coming, i.e., the End Times.Aziraphale wants to make the Second Coming a bit more upbeat.Prime VideoAziraphale wants to make the Second Coming a bit more upbeat.Prime VideoMuriel is hot on the case of the missing Book of Life.Prime VideoMuriel is hot on the case of the missing Book of Life.Prime VideoThe archangels of heaven are unimpressed with Muriel’s efforts thus far.Prime VIdeoThe archangels of heaven are unimpressed with Muriel’s efforts thus far.Prime VIdeoMuriel is hot on the case of the missing Book of Life.Prime VideoThe archangels of heaven are unimpressed with Muriel’s efforts thus far.Prime VIdeoThe original plan for the third and final season called for six episodes, but production was delayed first by the 2023 writer’s strike and then bymultiple allegationsofsexual assaultagainst Gaiman. (Gaiman has vehemently denied allegations of any nonconsensual sex or abuse, but admitted to being selfish and “careless with people’s hearts and feelings” in aJanuary 2025 blog post.) The fallout led to Gaiman withdrawing from the project and Prime Video opting for a 90-minute finale rather than a full season. And here we are.The Second Coming hits a snagThe finale picks up a few years after the S2 cliffhanger. Aziraphale is now Supreme Archangel, with plans for the Second Coming well underway—except he’s tweaked them to be a bit more upbeat, bringing peace on Earth and universal happiness rather than the rampant death and destruction of Armageddon. This doesn’t go down well with some of his fellow angels, who prefer the original plan. A heartbroken Crowley, meanwhile, is spending his time drinking heavily and passing out in a Soho alley, having lost his sense of purpose when Aziraphale refused him.The Second Coming rollout soon hits a snag. First, the Metatron mysteriously vanishes, removed completely from reality by someone who has stolen the Book of Life. In the ensuing panic, Jesus (Bilal Hasna) wanders off down to Earth and is befriended by a former street hustler named Harry the Fish (Mark Addy). The Archangel Michael (Doon Mackichan) and plucky assistant Muriel (Quelin Sepulveda) focus on solving the Metatron’s murder, while Aziraphale heads down to Earth to hunt for the missing Jesus, lest the demons of Hell find him first. He enlists a reluctant Crowley’s help.Good Omenshas always embraced the colorfully comic side quest; it’s part of what makes this such a rich fictional universe. But you need time to flesh it all out for those subplots to really work, and time is what the finale just doesn’t have. Hell and its demons, in particular, seem little more than an afterthought here; they’re not even particularly effective as comic relief. The beats

Now we waitUS hantavirus case was false positive; outbreak cases drop from 11 to 10WHO announced today that the operation to safely transfer passengers is complete.Beth Mole–May 15, 2026 5:31 pm|16Evacuation by boat of passengers on board the cruise ship MVHondiusanchored near the port of Granadilla, on May 11, 2026 in Granadilla de Abona, Tenerife, Canary Islands, Spain.Credit:Getty | Europa Press CanariasEvacuation by boat of passengers on board the cruise ship MVHondiusanchored near the port of Granadilla, on May 11, 2026 in Granadilla de Abona, Tenerife, Canary Islands, Spain.Credit:Getty | Europa Press CanariasText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navIn a press briefing Friday, officials for the World Health Organization announced that the case count of thehantavirus outbreak on the cruise shipMV Hondiusin the South Atlantic has shrunk from 11 cases to 10 after a previously reported US case was found to be a false positive.That US case was originally reported byUS health officials as “mildly positive,”and the WHO had considered it “inconclusive,” but still counted in the outbreak as a case in the agency’sMay 13 outbreak reportand in a briefing on May 14.The inconclusive case was in Dr. Stephen Kornfeld, an American doctor aboard the ship who helped respond to the outbreak after the ship’s doctor became ill.In an interview with CNN earlier this week, Kornfeld explained that he and others on board had taken nasal swabs early in May, before evacuation, and those swabs were sent for PCR testing in the Netherlands. Two labs in the Netherlands processed Kornfeld’s swabs; one lab reported a negative result, and the other reported a faint positive.Generally, a faint positive result on a PCR test could suggest low levels of virus at the start or end of an infection, or it could simply suggest contamination.Adding to the complexity of the potential case, Kornfeld said that he had developed a minor illness in early April while on the ship, just a few days after the first hantavirus case fell ill in the outbreak—a Dutch man who died on board from his infection.In an interview with ABC News Friday, Kornfeld reported that his repeat testing was negative and so was his serology testing looking for antibodies against hantavirus—which he would have developed if his illness in April was actually an unusually mild hantavirus infection. Overall, the testing shows that he is not currently infected and has not had a previous infection. As such, he has been transferred from the biocontainment unit to the quarantine unit at the Nebraska Medical Center, where US passengers from theHondiusare being monitored.“I physically feel great—I have felt great for many, many days,” Kornfeld told ABC. “Emotionally, I feel wonderful. It’s nice to be negative for hantavirus.”Evacuation completeOverall, there have been 10 cases: Seven cases who became ill on the ship, one case that disembarked the ship on April 24, before the outbreak was identified, and fell ill in Switzerland. The remaining two cases—one from France, the other from Spain—were identified as the ship was evacuated and passengers were being repatriated. The number of deaths has not changed since May 2, with three total: a Dutch couple and a German woman.Also in the WHO press conference, Director-General Tedros Adhanom Ghebreyesus said that the operation in the Canary Islands to safely evacuate the ship is complete. The operation involved transferring the ship’s 120-plus passengers to their home countries or to quarantine in host countries en route to their final destinations.“Because of the long incubation period of up to six weeks, more cases may be reported in coming days as passengers return to their countries, where they are being quarantined and tested in specialized facilities or at home,” Tedros warned. “This does not mean the outbreak is expanding; it shows that the control measures are working, that laboratory testing is ongoing, and that people are being cared for with support from their governments.”The US is monitoring 41 people who may have been exposed. That includes 18 passengers (including Kornfeld) who were evacuated from theHondius, passengers who had disembarked the ship on April 24, and people who shared a flight with a Dutch woman (the wife of the first case) who also fell ill from the infection and died in South Africa, on her way home.Beth MoleSenior Health ReporterBeth MoleSenior Health ReporterBeth is Ars Technica’s Senior Health Reporter. Beth has a Ph.D. in microbiology from the University of North Carolina at Chapel Hill and attended the Science Communication program at the University of California, Santa Cruz. She specializes in covering infectious diseases, public health, and microbes.16 Comments

striking resultsA revolutionary cancer treatment could transform autoimmune diseaseResearchers are testing CAR T cell therapy as a way to reset the immune system.Amber Dance, Knowable Magazine–May 17, 2026 7:00 am|51Artist’s rendering of a T cell.Credit:Getty ImagesArtist’s rendering of a T cell.Credit:Getty ImagesText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navAt age 49, Jan Janisch-Hanzlik’s multiple sclerosis was destroying her freedom to live the life she wanted. She gave up her active nursing job for a desk role. Frequent falls made her afraid to carry her grandchildren. She had to move to a bigger house to make room for the wheelchair she feared she might end up needing full-time.Even the best available medication wasn’t improving Janisch-Hanzlik’s symptoms, and she worried they’d only get worse. So when she learned about a trial of CAR T cell therapy at the University of Nebraska Medical Center in Omaha, close to the city of Blair where she lives, she phoned the clinic every other month until they were ready toenroll her as the first patient.Originally designed to target and wipe out cancer byreprogramming the patient’s immune cells, CAR T is now being offered to patientsin hundreds of clinical trialsfor autoimmune conditions like multiple sclerosis, lupus, Graves’ disease, vasculitis, and many others. The hope is that CAR T can duplicate the success it has demonstrated in a range of blood cancers by hunting down and eliminating cells that target the self in autoimmune diseases. This would essentially reset the body’s defenses to a state like the one that existed before the disease took hold.But along with CAR T’s promise come risks, questions, and challenges. There’s uncertainty about how well it will work forautoimmunityand how long any benefits might last, as well as what long-term side effects might arise. Janisch-Hanzlik knew this when she sat down to receive the experimental treatment on June 9, 2025; she felt a mix of hope and fear knowing that she would be spending the next week being monitored for side effects including dangerous inflammation.In addition to her clinical expertise and desire to pioneer a new treatment, Janisch-Hanzlik’s two young grandchildren helped inspire her pursuit of a treatment with known risks and uncertain benefits. Because multiple sclerosis has a genetic component, Janisch-Hanzlik knew that they have an elevated chance of going through the same struggle she has. “I would want to be able to say I did everything that I possibly could to prevent them, or anyone else, from having something like this,” she says.From cancer to autoimmunityThe first CAR T cancer treatment wasapproved by the Food and Drug Administration in 2017for an aggressive form of leukemia. Since then, the powerful and intensive treatment hasresulted in long-term remissionfor many cancer patients.The basic premise of CAR T is to activate the power of key immune cells calledT cells. T cells normallyrecognize other cellsthat have been infected by a virus or bacterium, or are otherwise abnormal, and either destroy them or recruit other parts of the immune system to do so.In CAR T for cancer, scientists engineer those T cells to specifically hunt and destroy malignant cells. The technologygot its startwhen cancer researchers figured out how to take out a patient’s own T cells, insert DNA instructions for a “chimeric antigen receptor,” or CAR, and put them back into the person’s circulation. The CAR, which sits on the T cell’s surface and latches on to a specific molecular partner on the surface of cancerous cells, activates the T cell to attack.Today CAR T cells are most commonly programmed to attack B cells, another key immune player. B cells are normally responsiblefor making antibodies, but in certain blood cancers, they proliferate out of control. By giving T cells a CAR that recognizes one of a couple of molecules unique to the B cell surface, the cells are reprogrammed to find and eliminate those cancerous cells.B cells also are the central problem in many autoimmune conditions: They mistakenly make antibodies against normal tissues instead of against invading pathogens. So as CAR T began to succeed against B cell cancers, it didn’t take long for doctors to reason that CAR T therapy might also be able to wipe out bad B cells in people with autoimmunity.There are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineThere are many variations on CAR T procedures, but the basic process involves removing, modifying, and reinfusing patient T cells so they can attack their target, which is B cells in the case of autoimmune disease.Credit:Knowable MagazineA German team pioneered autoimmune CAR T in a woman with lupus,reporting positive resultsin 2021.

Bad beatsThe US is betting on AI to catch insider trading in prediction marketsThe Commodity Futures Trading Commission wants us to know it’s taking this very seriously.Kate Knibbs, wired.com–May 16, 2026 7:00 am|95Credit:Yuki Shimazu / FlickrCredit:Yuki Shimazu / FlickrText settingsStory textSizeSmallStandardLargeWidthStandardWideLinksStandardOrange Subscribers onlyLearn moreMinimize to navFor most of the past year, it looked likeprediction marketshad kicked off a new golden age of fraud. OnPolymarket, traders raked in fortunes fromsuspiciously timed betson geopolitical events like the raid on Venezuela and the Iran War. It wasn’t clear whether the US government would bother pursuing some of the most flagrant bad actors, since Polymarket’s crypto-based platform was technically offshore and not regulated or licensed within the country.Now, however, the Commodity Futures Trading Commission, which oversees prediction markets, wants you to know that it’s watching very, very closely. The agency is searching for suspicious behavior from traders within the United States who have been sneaking onto offshore markets, including Polymarket’s crypto platform—which is blocked stateside—by using virtual private networks. “We’re going to find them, and we’re going to bring actions,” agency chairman Michael Selig told WIRED this week, speaking from the CFTC’s headquarters in Washington, DC.Selig says the agency, which is especially lean right now, is staffing up. Like so many other AI-pilled workplaces, the CFTC is also leaning into automation to handle the growing workload, including tools that analyze trading patterns and flag potential manipulation. “You’ve got so much data,” Selig says. “When we feed it into AI, we get really great information. It can help us understand things, like where we might want to investigate, or when we might need to send a subpoena to a trader.”In addition to proprietary surveillance systems developed in-house, the agency’s arsenal includes third-partyblockchain tracing toolslike Chainalysis for crypto platforms, and market abuse detection software including Nasdaq Smarts for centralized markets. (Beyond Nasdaq Smarts, the agency did not specify which AI tools it uses and declined to share more specific examples.)Prominent prediction market companies have recently started touting all the work they’re doing to catch sketchy bettors. US-based exchange Kalshi, Polymarket’s primary competitor, eagerly announced that it hassuspendedand penalized customersflagged for insider tradingand market manipulation.In April, after significant backlash over suspected insider trading, Polymarket announced its own partnership with Chainalysis. It was part of a broader push to crack down on market manipulation. While the company’s CEO, Shayne Coplan,had talkedin the past about why insider trading could be good for prediction markets, Polymarket changed its approach this spring, updating its market integrity rules and announcing a partnership with Palantir for its US-based sports markets (the Chainalysis deal focuses on the offshore platform). The company did not respond to WIRED’s request for comments for this story.According to Chainalysis spokesperson Maddie Kenney, the company analyzes the same data for both clients. “The value Chainalysis adds for our customers, including Polymarket and the CFTC, is organizing the data and enriching it with the attributions and insights we’ve accumulated over years in the space,” she says. Certainly sounds like a good deal for Chainalysis!The CFTC’s assurances that it is hunting insiders comes at a moment of intense scrutiny on prediction markets. In March, Connecticut senator Chris MurphytoldWIRED that he suspected White House staffers were engaged in insider trading on war-related contracts. At the beginning of April, seven members of Congressaskedthe CFTC to investigate overseas markets offering war-themed events contracts. In a letter, the lawmakers argued that the commission had the authority and responsibility to curb insider trading, especially on “morally obscene” trades on military action. Selig recentlytoldCongress that the company is pursuing “hundreds, if not thousands” of insider trading tips.Investigations are not limited to federally regulated exchanges. “We’re surveilling the markets on a global basis,” he tells WIRED.Selig says that the agency will exert extraterritorial jurisdiction—its legal ability to enforce its laws beyond traditional boundaries—when it finds suspicious activity on offshore platforms like Polymarket, though he says it’s a case-by-case approach. “We use it in extreme circumstances,” he says, with an eye towards whether charges have a strong chance of sticking in court. “In any extraterritorial litigation, there’s going to be challenges to our authority, and that could also impair our ability to bring cases in the future.” According to Selig, the 2010 Dodd-Frank Act allows the CFTC more leeway to pursue this kind of enforcement

🚀马斯克：只有我和黄仁勋坐上了空军一号💻MacBook Neo 推高销量，苹果笔记本一季度线上市占升至 13%🤖曝苹果探索将 AI Agent 引入 App Store🚗李想：L4 无人驾驶预计 2028 至 2029 年落地，需近万 TOPS 算力🔧比亚迪回应推「车机屏幕小换大」服务🚙蔚来全新 ES8 连续五个月夺大型 SUV 销冠，四月卖出 13020 台🎶Charli xcx 成为 Nothing 首位全球品牌大使💰腾讯 Q1 营收 1964.6 亿元，马化腾称去年「上了漏水的船」☁️阿里巴巴 2026 财年营收超 1.02 万亿元，云业务 AI 收入连续 11 季三位数增长🧸泡泡玛特一季度整体收益同比增长 75% 至 80%，中国市场收入翻倍💡李彦宏提出「日活智能体数 (DAA)」概念：未来全球 DAA 可能超过 100 亿📱小米 17 Max 官宣：定位小米 17「全面升舱版」，本月发布📸10490 元起，索尼 Xperia 1 VIII 正式发布📸31799 元，索尼发布 Alpha 7R VI 全画幅微单：6680 万像素、30 张/秒连拍🎥13999 元，佳能发布 EOS R6 V 全画幅创作视频机🛍️亚马逊整合 AI 购物功能，推出 Alexa for Shopping🚗小米开源自动驾驶大模型 Xiaomi OneVL🔧MiniMax Agent 升级为 Mavis🔍面壁智能开源 MiniCPM-V 4.6💡元宝上线聊天记录总结功能马斯克：只有我和黄仁勋坐上了空军一号特斯拉 CEO 埃隆 · 马斯克昨日在 X 发文称，正乘坐美国总统专机「空军一号」前往北京。对于此前有人发帖称空军一号上 CEO 们的总身价超过 20 万亿美元，马斯克回帖称「只有我和老黄（英伟达 CEO 黄仁勋）在空军一号上」。值得关注的是，英伟达 CEO 黄仁勋并不在最初名单之列。据彭博社报道，特朗普在出发当天早上亲自致电黄仁勋发出邀请，并在空军一号途经阿拉斯加安克雷奇加油期间，黄仁勋赶至机场登机。🔗 相关阅读：风暴中的 Manus，蝴蝶还在扇动翅膀MacBook Neo 推高销量，苹果笔记本一季度线上市占升至 13%昨天，研究机构洛图科技发布 2026 年第一季度中国笔记本电脑线上全渠道销量数据显示，中国笔记本电脑线上全渠道零售销量为 230 万台，同比下降 19.2%；销额为 148 亿元，同比下降 14.7%。苹果是本季度最大亮点，其在线上主流电商平台的销量份额攀升至 13%，较去年同期提升约 6 个百分点。研报认为，MacBook Neo 上市即热销，是份额跃升的关键推手。品牌格局方面，前五大厂商合计市场份额达 79.8%，较去年同期提升 9 个百分点，市场集中度进一步提高。曝苹果探索将 AI Agent 引入 App Store据《The Information》报道，苹果正在探索将 AI 智能体引入 App Store 的方案，以在把握行业趋势的同时，确保平台规则得到延续。报道称，苹果内部团队正在设计一套新的管理机制，目标是在保障隐私与安全标准的前提下，约束智能体的越界行为，例如防止其擅自删除用户邮件等极端情形。App Store 的商业价值是苹果推进这一方向的重要动因。据应用分析机构 Appfigures 估计，苹果今年 App Store 全球佣金收入约达 500 亿美元，同比增长约 20%。此前，苹果已多次以违反政策为由，将「氛围编程」（vibe coding）工具拒于 App Store 之外。李想：L4 无人驾驶预计 2028 至 2029 年落地，需近万 TOPS 算力在昨天发布的《罗永浩的十字路口》第 27 期中，理想汽车创始人李想分享了多项关于自动驾驶、AI 行业及未来用车市场的看法。他表示，L4 级无人驾驶预计要到 2028 年至 2029 年才能真正落地。L4 无人驾驶需要 3D ViT、物理世界的预训练模型、比人还快的控制系统，基本上得有近万 TOPS 的算力，大概要到 2028 年、2029 年。针对无人驾驶普及后是否会减少购车需求的问题，李想给出了否定答案。他以租房与买房作类比：道路是有限的，没有办法增加更多供给。在你最需要车的时候，供给仍然供不上，因为它不是由车的供给决定的，是由道路的供给决定的。李想认为，那些平时不愿雇专职司机、又在意出行隐私的人，届时将更倾向于自购车辆。比亚迪回应推「车机屏幕小换大」服务据九派新闻报道，近期比亚迪 APP 官方商城面向 25 款宋 PLUS、宋 L、秦 L、元 UP 等车型推出「车机屏幕小换大」的服务。相关页面显示，车主可花费 1999 元，将原车搭载的 12.8 英寸屏幕升级为 15.6 英寸大屏；元 UP 车型提供售价为 1320 元的从 10.1 英寸屏幕升级至 12.8 英寸服务。针对上述服务，比亚迪官方工作人员回应称，定价综合参考行业同类升级产品市场行情、原厂正品屏幕物料成本及标准化拆装工时成本制定，兼顾市场公允性与用户体验感知，服务包含原厂大屏物料、专业标准化拆装施工及基础配套安装服务。在安全性方面，比亚迪方面称，屏幕升级采用原厂适配安装方案，无需改动原车线路及中控台原有结构，不会产生异响、密封失效等问题。升级后，车机系统功能、车机互联体验均保持原有状态不受影响，不影响车辆原车其他零件的官方质保权益。蔚来全新 ES8 连续五个月夺大型 SUV 销冠，四月卖出 13020 台近日，中国汽车流通协会乘用车市场信息联席分会发布了最新市场数据，显示蔚来全新 ES8 今年 4 月零售销量达 13020 台，在大型 SUV 级别及 40 万元以上市场连续五个月斩获销量冠军。截至目前，全新 ES8 累计用户数已超过 10 万，自开启交付起历时 215 天完成十万台新车交付，刷新了中国乘用车市场 40 万元以上高端车型的销售与交付纪录。Charli xcx 成为 Nothing 首位全球品牌大使Nothing 昨日宣布，歌手 Charli xcx 正式成为公司首位全球品牌大使，双方同步发一支全球创意广告。Charli xcx 此次以股东身份加入 Nothing，与 The Weeknd、Casey Neistat、Swedish House Mafia 同列公司全球文化引领者阵容。腾讯 Q1 营收 1964.6 亿元，马化腾称去年「上了漏水的船」昨天，腾讯披露 2026 年第一季度业绩报告，期内实现营收 1964.6 亿元，同比增长 9.1%；归母净利润（权益持有人应占盈利）581 亿元，同比增长 21%；研发投入 225 亿元，同比增长 19%。AI 产品进展是本季度最受关注的亮点。腾讯混元重建后推出的新模型 Hy3 preview，在 OpenRouter 平台自 4 月 27 日至 5 月 11 日已连续三周登顶周榜总榜、WorkBuddy 月访问量及环比增速居国内 PC 端 AI 原生办公智能体市场首位。据每日经济新闻报道，在股东大会上，腾讯董事会主席兼 CEO 马化腾就外界关于腾讯 AI 是否落后的质疑作出回应。他坦言，「原来一年前我们以为上了船，后来发现那个船漏水了，现在感觉站上去了，还坐不下去，还是希望船速能快一点」，并表示腾讯早期在 AI 基础能力方面并不突出，近年通过人才建设与内部培训持续补全短板。马化腾同时强调，腾讯的 AI 路线要结合自身独有优势稳扎稳打，「不能看着别人在那边做就随便跨过去，抢别人的地盘，过去我们也抢过但后来基本失败了」。游戏业务表现稳健，本土市场游戏收入 454 亿元，同比增长 6%，《王者荣耀》《和平精英》《三角洲行动》等多款长青游戏流水创季度新高；国际市场游戏收入 188 亿元，同比增长 13%；To B 业务（金融科技及企业服务）营收 599 亿元，其中企业服务收入同比增长 20%，腾讯云国际业务延续强劲增长；用户侧，截至今年 3 月底，微信及 WeChat 合并月活账户数 14.32 亿，同比增长 2%；QQ 移动端月活 5.16 亿，同比下降 3%。阿里巴巴 2026 财年营收超 1.02 万亿元，云业务 AI 收入连续 11 季三位数增长阿里巴巴集团昨日发布 2026 财年 Q4 及全年业绩：季度收入 2433.80 亿元，同比增长 3%；全年收入 10236.70 亿元，同比增长 3%；本季度阿里云外部商业化收入同比增长 40%，AI相关产品收入占比首次突破 30%。阿里巴巴中国电商集团收入 1222.20 亿元，同比增长 6%，其中即时零售收入 199.88 亿元，同比增长 57%；阿里国际数字商业集团（AIDC）收入 354.29 亿元，同比增长 6%；云智能集团收入 416.26 亿元，同比增长 38%，其中外部商业化收入增速达 40%；AI 相关产品收入达 89.71 亿元，连续第十一个季度实现三位数同比增长。在昨晚的财报电话会上，阿里 CEO 吴泳铭宣布，阿里 AI 已跨越初期投入阶段，正式进入商业化回报周期。阿里云外部商业化收入预计未来几个季度将在 40% 增速基础上继续加速，中长期维持高速增长。吴泳铭还表示，自研芯片在阿里云的部署量还是不大。但未来有希望从 GPU、CPU 到存储、网络芯片实现全栈自研。泡泡玛特一季度整体收益同比增长 75% 至 80%，中国市场收入翻倍昨天，泡泡玛特发布 2026 年第一季度业务更新公告，整体收益较去年同期增长 75% 至 80%。公告未披露具体营收金额及利润数据，仅披露各区域及渠道收益同比增速。分区域来看：中国收益同比增长 100% 至 105%；亚太（指中国以外其他亚洲及大洋洲国家和地区）同比增长 25% 至 30%；美洲同比增长 55% 至 60%；欧洲及其他地区同比增长 60% 至 65%。中国市场各渠道中，线下渠道同比增长 75% 至 80%，线上渠道同比增长 150% 至 155%，线上增速显著高于线下。💡 李彦宏提出「日活智能体数 (DAA)」概念：未来全球 DAA 可能超过 100 亿昨天，在 Create2026 百度 AI 开发者大会上，百度创始人李彦宏提出 AI 时代的新「度量衡」—— DAA（Daily Active Agents，日活智能体数），并预测未来全球 DAA 可能超过 100 亿。李彦宏认为，Token 衡量的是投入与成本，并不代表产出与收益，不一定代表终局；而 DAA 大致对应移动互联网时代通用的 DAU（日活用户数）。衡量一个平台和生态的繁荣，更应该看的是 DAA 这个指标，关注有多少 Agent 在给人类干活，并交付结果。这比无谓的 Token 消耗，更接近价值，也更接近本质。李彦宏判断，智能体出圈代表着 AI 的发展正从模型阶段走向应用阶段，「第一次，AI 的主角不是模型，而是应用」。他将 Chatbot 与 Agent 视为两类入口：以 ChatGPT 为代表的 Chatbot 主要解决「信息获取」问题，而 Agent 解决的是「任务完成」问题，任务越复杂、价值越高，Agent 的价值天花板高于 Chatbot。小米 17 Max 官宣：定位小米 17「全面升舱版」，本月发布昨天，小米官方正式公布了小米 17 Max，定位「小米 17 的全面升舱版」，预计本月发布。小米集团合伙人、总裁卢伟冰介绍，团队在收到大量用户对「大屏版本」的需求反馈后，决定在小米 17 基础上推出 Max 机型，在影像、屏幕、续航、性能四个维度全面升级。外观方面，小米 17 Max 镜头模组采用了类似小米 15 系列的全黑镜组设计，预计提供黑、蓝、白三个配色，其中黑色版本会搭载类似龙鳞纤维的背盖材料。10490 元起，索尼 Xperia 1 VIII 正式发布索尼昨日正式发布年度旗舰手机 Xperia 1 VIII，先看价格：12GB/256GB 版本约 23.6 万日元，16GB/1TB 顶配版本约 30 万日元；中国香港市场起售价为 12099 港元。预售用户可获赠一副索尼 WH-1000XM6 无线耳机。外观方面，索尼没再

两个月前在圣何塞，黄仁勋穿着皮衣站在 GTC 的舞台上，告诉全世界：Token 是新的大宗商品，生成 Token 的成本与效率，决定科技企业的营收与生死。昨天，李彦宏站在 Create 2026 的开幕式上，说了一句看似拆台的话：「Token 只是代表成本，并不代表收益。它衡量的是投入，而不是产出。」然后他抛出了一个新概念，DAA，Daily Active Agents，日活智能体数。卖铲子的人说，看铲子消耗量就知道金矿的繁荣程度。挖矿的人说，你倒是看看我挖出了多少金子。他们都没说错。但同一座金矿，用不同方式去称量，得出的故事完全不同。真正有意思的地方不在于谁对谁错，在于一个事实：AI 跑得太快了，快到这个行业连怎么给自己记账都还没想清楚。DAA ：从「烧了多少」到「干了多少」李彦宏的 DAA 逻辑并不复杂。移动互联网时代，衡量平台看 DAU，日活用户数。微信 13 亿，抖音 7 亿，Meta 34 亿。进入智能体时代，对应的指标应该是 DAA，有多少 Agent 每天在给人类干活，并交付结果。「交付结果」四个字是重点。Token 消耗量告诉你系统在转、算力在烧、钱在花。但它不告诉你这些 Token 到底干了什么。是帮创业者写了一份商业计划书，还是用户反复问「你好」然后 GPT-5.4 收了他 80 美元？这两种场景消耗的 Token 可能差不多，创造的价值天差地别。李彦宏说，「这比无谓的 Token 消耗，更接近价值，也更接近本质。」他还给了个大胆的预测：未来全球 DAA 可能超过 100 亿。一个人同时挂三五个 Agent 处理不同任务，你是一个 DAU，但贡献了三五个 DAA。Agent 数量远超人口，技术上已经可以想象。李彦宏自己也说这是一个「非共识的观点」。在所有人都在谈 Token 消耗量的行业里说 Token 不代表终局，是需要点勇气的。当然，也需要点动机。Token 经济学：一种极其性感的叙事Token 经济学的性感之处在于，它有大量数据支撑。截至今年 3 月，中国日均 Token 调用量超过 140 万亿，相比 2024 年初的 1000 亿增长了 1000 多倍。火山引擎豆包大模型日均 Token 使用量突破 120 万亿。OpenRouter 平台前十大模型的周 Token 调用量从 1.24 万亿暴增至近 14 万亿。一条陡峭到让人眩晕的增长曲线，核心驱动力恰恰是 Agent。过去 Chatbot 时代，单轮对话消耗 1000 到 3000 个 Token。Agent 来了之后，一个完整任务触发几十上百次模型调用，中等任务吃掉 10 万 Token，复杂任务上百万。Agent 越多，Token 消耗越大，算力需求越高，芯片卖得越好。这个闭环对英伟达来说无比优雅。2025 年全年营收 2159 亿美元，净利润 1170 亿美元。黄仁勋有充分的理由说：Token 就是新的石油。企业也在用脚投票。阿里成立 Token Hub 事业群，吴泳铭亲自挂帅。58 同城姚劲波说 Token 用得越多越好，不计成本。昆仑万维发内部信强制 AI Coding，达不到要求的末尾淘汰。黄仁勋甚至预言「你的 offer 带多少 Token」会成为硅谷新的谈判筹码。Token 消耗量作为指标，最大的好处是简单、可量化、跨平台可比。140 万亿就是 140 万亿，谁看都一样。但过亿简单也可能会掩盖一些问题。一个只盯着食材消耗量的餐厅，未必是一家好餐厅。共识大于分歧把 Token 经济学和 DAA 放在一起看，你会发现它们的共识其实大于分歧。这两个判断都基于 AI 正从对话阶段走向 Agent 阶段，而且都认为这是万亿级的产业重塑。分歧在哪？在于谁的账本更能代表这个行业的健康程度。Token 量的是投入侧：消耗了多少算力，工厂每瓦特产出多少 Token。这对芯片厂商和云厂商极其重要。DAA 量的是产出侧：有多少 Agent 在运行，完成了多少任务。这对应用平台方极其重要。百度做的是 DuMate、秒哒、百度一镜，卖点全是「帮你把事做完」。只有用「结果交付」来记账，应用层的价值才浮得出水面。屁股决定脑袋，但这不代表谁在说谎。每个玩家天然会推崇对自己最有利的记账方式。有意思的是，腾讯也站到了应用侧。腾讯云副总裁杨晨说过：「我们认为 Token 不是一个多么健康的生意，它看着量很大但黏性极差。」腾讯的核心资产是微信生态和内容场景，它在乎的是 AI 能不能在自家场景跑起来。阿里的态度截然相反，吴泳铭断言未来 5 年 AI 和云业务收入将超 1000 亿美元。同一个行业，同一批聪明人，看法完全不同。这本身就说明 AI 的迭代速度已经跑在了共识形成的前面。Token 消耗量像 GDP，量总规模，不管建了一座桥还是挖了个坑再填上。DAA 像就业率，量的是有多少「劳动力」在创造价值。GDP 高但就业率低，那叫空转。就业率高但产出质量差，那叫虚胖。一个健康的经济体两个数字都要看。但对普通用户来说坦率地讲，这两个指标都是术语。用户只在乎一个朴素的问题：好不好用？帮我省了多少时间？花了我多少钱？Token 经济学叙事下的 AI 产品，对用户其实不太友好。你用打车软件，起步价多少、每公里多少、堵车怎么算，一清二楚。你用 AI Agent，到底消耗了多少 Token，是哪个模型在跑，Agent「反思」了几轮导致费用翻倍，很多普通用户还搞不明白。DAA 至少提供了另一种视角：用户不该关心 Agent 烧了多少 Token，该关心它有没有帮你把事做完。前者是水表读数，后者是水龙头出不出水。造词运动的背后2024 年初中国日均 Token 调用量才 1000 亿。2026 年 3 月，140 万亿。1000 倍。这种速度下，任何已有的框架和认知都会瞬间过时。于是，每个玩家只能从自己的位置出发，尝试给这团混沌画一个坐标。黄仁勋造了「Token 工厂」，李彦宏造了「DAA」，阿里造了「Token Hub」……Token 衡量的是基础设施的繁荣程度，DAA 衡量的是应用价值的交付密度。两者更多是互补而非互斥。没有 Token 的生产和流转就没有 Agent 的运行，没有 Agent 交付结果 Token 消耗就只是空转。对用户来说，最好的度量衡是那个你根本不需要知道的。你打开水龙头，水来了。你把活交给 Agent，事做完了。你看一眼账单，觉得合理。至于背后消耗了多少 Token、动用了多少 Agent、芯片是谁家的、云服务用的哪一层，留给黄仁勋和李彦宏去争论就好。从 DAU 到 DAA，从「人在用产品」到「产品在帮人干活」，AI 时代的价值坐标正在被重建。争论是好事。当一个行业只有一种记账方式的时候，所有人都会围绕同一个数字做优化，不管那个数字是否真的通向价值。多一种衡量的维度，至少多一个纠偏的机会。至于最后哪套度量衡能留下来，答案可能不在黄仁勋和李彦宏手里，在你手里。你每天用 Agent 做了什么，做得好不好，愿不愿意继续付费。这些真实发生的选择，才是 AI 时代最诚实的投票。

未来我们的汽车哪怕前轮或后轮爆胎，剩下的轮子也能正常行驶。王传福 11 年前的这个构想，在今晚成为现实。方程豹这次将发布会搬到户外，正式推出豹 8 与豹 5 的「闪充版」车型。其中，全栈自研的云辇-P Ultra 全场景智能液压车身控制系统成为全场技术焦点。爆胎后车轮悬空，车辆仅靠三个轮子平稳开走；陷入泥坑无需求援，自行抬高被困轮胎完成脱困。这些曾经只存在于构想中的画面，方程豹今晚都一一呈现了出来。两年前方程豹以 DMO 平台入局，打破了新能源与硬派越野之间的壁垒。如今闪充版车型上市，方程豹用这套智能底盘技术，让越野时的爆胎脱困和泥地自救从理论变成了实际的能力。液压底盘，让方程豹告别千斤顶硬派越野的底盘路线通常有两种：被动悬架结构简单、可靠性高，可一到铺装路面，舒适性就打了折扣；空气悬架确实能改善平顺性，但气体可压缩的特性意味着，在高强度飞坡、碎石路这些深度越野场景里，气囊一旦破损，整套系统就可能失效。方程豹没选这两条路，直接把仰望 U8 上的液压车身控制系统搬了过来，还做了升级。方程豹汽车总经理熊甜波表示，为了打磨这套系统，研发团队把公司的一座厂房改成了专业越野测试场，一比一复刻了天门山越野场景，坑洼、碎石路面也都一应俱全。工程师还会洒水、喷泡沫，来模拟雨天等低附着环境。实验室之外，真实的反馈同样关键。有车主单人单车跑了 11 万多公里，穿越巴丹吉林沙漠和敦德冰川，研发团队从这些反馈里整理了 29396 条用户建议，不断修正底盘标定。经过上万次极限测试，云辇-P Ultra 最终落地。据介绍，整套系统的静态高度调节行程从 140 毫米提升到 200 毫米。日常城市里开迎宾模式，悬架最低能降到 90 毫米，老人小孩上下车更方便；到了极端越野场景，悬架最高能抬到离地 320 毫米。举升能力是液压悬架的另一项硬指标。方程豹这次解决了主动悬架容易超压失效的问题，极限举升载荷做到了 9 吨，差不多两头成年大象的重量。就算停在 45°的陡坡上，四个轮子也能独立精准控制，把车身抬起来。另外，越野时的高温散热是底盘研发躲不开的课题。这套悬架的整体散热面积是传统被动悬架的 12.7 倍，是单减震器的 3.7 倍。散热余量够大，沙漠里连续冲沙，阻尼也不会衰减。系统工作温度范围从 -40°C 跨到 85°C。机械底子搭好了，配合整车姿态闭环算法和四轮独立解耦控制技术，车辆就具备了自救能力。发布会最亮眼的那几项功能，都源自这套底层技术。「三足鼎行」应对的是突发险情，例如高速上单侧爆胎，或者山崖单行道会车。遇到这样的意外，系统会通过云辇计算中心动态分配四个减震器的油液载荷，收起其中一个车轮，车辆可以仅靠剩下三个轮子，以最高 15 km/h 的速度开到安全的路边。「腾空换胎」和「单轮脱困」同样靠整车姿态的独立调整，车主可以在中控屏或手机上操作，让陷入沙地的单侧车轮单独抬起，然后在悬空的车轮下方填沙、垫脱困板，车辆就能自己开出来。有了这个功能，在荒野里爆胎也不用费劲找平地架千斤顶了，轮子一提，一个人三分钟就能把备胎换上，过程中随时暂停，举升高度也能随意调节，一次提轮可以维持 24 小时以上。熊甜波还称，这套悬架还针对飞坡落地这种强冲击场景做了优化，系统的受力瞬间刚度能提升两到五倍，即便从两米的高度落地，落地后照样正常开，不用担心拉杆变形或者密封失效。非铺装路面之外，这套也兼顾实用与全场景驾驭。据介绍，新升级的刚度控制模块有三档，一档日常通勤，二档复杂混合路况，三档越野飞坡。双阀设计能够把阻尼带宽最大拉到原来的三倍，过减速带、压井盖这些细碎震动，过滤得更细腻。参数之上，现实之下补齐了底盘这块，新能源越野的基本盘依然要回到三电系统。熊甜波首先介绍的还是比亚迪第二代刀片电池和闪充技术。发布会公布的补能数据显示，豹 5 和豹 8 闪充版的电量从 10% 充到 70% 只要 5 分钟，充到 97% 也只需要 9 分钟。哪怕零下 30 度的严寒环境，充电时间也只多出 3 分钟。接着是安全问题，外界对越野车装电池的质疑一直没停过。官方表示，在提升了充电速度的同时，研发团队还提升了安全测试的标准，在实验室做了「边闪充边针刺」的破坏性测试。测试中，四节电芯同时短路，电池温控依旧平稳，没冒烟也没起火；面对 1500 千焦能量级的冲击测试，电池组的结构也经受住了考验。除了补能和安全，长途穿越也离不开长续航。讲到这里，现场的大屏幕上出现了从新疆乌鲁木齐开到甘肃敦煌的实测画面。测试中，豹 8 闪充版综合续航达到了 1380 公里，豹 5 闪充版也跑到了 1310 公里。而在智能化的比拼中，方程豹为两款新车规划了截然不同的辅助驾驶路线。旗舰定位的豹 8 闪充版搭载了华为乾坤智驾 ADS 4.1.5。前后双激光雷达加持，盲区感知精度提升明显。夜视系统升级后，野外没有路灯的暗夜里，车辆也能看清行人和横穿道路的小动物。豹 5 闪充版用的是天神之眼 B，这套方案依靠 Momenta 的算法和比亚迪庞大的行车数据迭代，日常城市通勤和复杂路况处理得也不错。座舱和周边生态上，两款车的目标人群区别更明显。豹 8 闪充版有大 5 座、大 6 座和 7 座三种布局，外观新增了悬浮顶双拼色，车内大面积用了 Nappa 真皮和 50 年树龄的白酸枝木。18 扬声器帝瓦雷音响加上奇华顿车载香氛，豪华感拉得很满，后排还带腿托和按摩。追求野外可玩性的车主，还可以花费 16968 元选装由比亚迪与大疆联合开发的智能车载无人机系统，在越野途中完成一键跟拍。另一边的豹 5 闪充版在保留家族式升降挡把与水晶按键的基础上，推出了以大漠夕阳为灵感的「赤砂红」专属车身配色。为了照顾越野老炮的改装需求，新车加了 OCS 接口。射灯、气泵这些外接设备不用破线就能直接接入，电流三档可调。车内还定制了一套隐藏式智能车载电台，主机藏在座椅下面，通过车机语音就能一键呼叫或发起定位救援。发布会临近尾声，售价正式公布：豹 8 闪充版：41.98 万元；豹 5 闪充版：30.58 万元。整场发布会看下来，全场景液压悬架的技术冗余、个位数分钟的补能速度，再加上座舱的豪华堆料，的确大幅拉高了新能源硬派越野的制造门槛。但回到终端市场，目前真正走量的是豹 5 210km 天神 Max 版，置换价格已经来到了 23.98 万元，和闪充版差了一大截。眼下车市卷得厉害，消费者对价格极其敏感。普通越野玩家愿不愿意为多出来的脱困能力和快充体验，去填平这道价格差，还得观察一段时间。不过，抛开终端定价与市场博弈不谈，单看这场硬核的新车发布，方程豹确实把研发团队压箱底的技术底牌全盘托出了。发布会最后，熊甜波坦言自己曾纠结是否要讲这么多枯燥的「物理课」和底盘参数，但她最终决定把工程师们夜以继日研发的成果完整呈现出来。每一项技术的突破，每一组硬核的参数，都藏着他们日夜坚守，默默付出的心血。他们的专注与坚守，值得被所有的人看见，被所有的人记住。这些从零起步、硬磕出来的技术堡垒，能否顺利转化为切实的销量增长，接下来的交付会反馈最真实的答案。

尊界和玛莎拉蒂要合作造车了。根据《云见 Insight》的报道，华为鸿蒙智行、江淮汽车、斯特兰蒂斯集团（Stellantis）及旗下玛莎拉蒂品牌，正就共同打造玛莎拉蒂品牌的新能源车展开洽谈。联合开发的车型将分为国内和海外版本，国内版归属于尊界品牌，海外版将挂玛莎拉蒂车标。首款车型目前正处于造型设计阶段，计划于明年下半年量产。按照规划，四方合作模式与鸿蒙智行 「五界」 高度相似：华为主导产品定义、提供核心技术，江淮联合研发并负责生产制造，玛莎拉蒂提供造型设计与品牌背书。这套分工，仍然延续了鸿蒙智行过去的基本方法：华为提供智能驾驶、鸿蒙座舱、电驱的技术和渠道能力，合作方提供制造、供应链和品牌资源。但这套熟悉的配方里，玛莎拉蒂的出现，带来了一个全新的变量。玛莎拉蒂负责提供的，并非某种可以量产的硬件或软件，而是一种华为过去很少主动去借用、甚至很难在短期内自我孵化的外部资源——更加年轻品牌性格和情绪表达。在目前的鸿蒙智行合作体系中，华为提供的是一套极其成熟且标准化的智能汽车能力。这个体系的运转效率极高，它不仅让华为迅速建立起了一个横跨二三十万到百万级价格区间的庞大汽车产品矩阵，也成功在消费者心里植入了一个清晰的等号：买鸿蒙智行，买的就是华为的智驾、座舱、主动安全和渠道服务。销量领跑超豪华轿车市场的尊界 S800 就是这套模式的集大成者。途灵龙行底盘平台、面向 L3 级自动驾驶架构设计的智驾系统、星河卫星通信技术、全方位的主动安全防护体系，以及最新一代的鸿蒙座舱，搭配上双拼色车身、大面积的深色皮质、木纹饰板。这套以智能化为基底，克制、稳重、偏向于商务豪华的高端表达，确实能精准地吸引到一批人。但高端市场里还有另一批人。这群人买车，更看重车辆的车身比例、姿态、线条、品牌气质，他们或许会被智能化打动，但要想让他们买单，光有聪明的脑子不够，还得有迷人的皮囊和动人的情绪。玛莎拉蒂曾经就擅长提供这种味道。2017 年，玛莎拉蒂在中国市场销量达到 14498 辆，成为品牌全球最大的单一市场。那几年，Levante 和 Ghibli 在中国卖得好，靠的并不是更强的智能配置，也不是更理性的产品力。消费者买的是三叉戟标识、意大利设计、跑车血统，还有一种和 BBA 不太一样的豪华感。它比法拉利、兰博基尼这些顶级超跑更懂生活，更容易接近；又比传统的德系豪华车更有情绪，更懂风情。很多人选择玛莎拉蒂，本质上不是在做一道产品力对比的算术题，而是在做一种审美选择。虽然后来它在中国快速掉队，但主要原因是新能源车时代来得太快，智能驾驶、座舱体验、电动化效率和整车软件能力，重新冲刷了传统豪华品牌原有的定价逻辑。玛莎拉蒂推出过电动车，但没有建立起足够鲜明的技术认知，也没有拿出一台真正让中国用户重新兴奋的产品。然而，销量的下滑，并不意味着品牌记忆的彻底消失，更不意味着这种消费需求不复存在。在中国这片庞大的市场里，依然有一群人，他们骨子里还是喜欢意大利设计的随性、跑车血统的张扬，以及那种更有情绪的豪华表达。这也是这次合作的基础。华为和江淮可以解决新能源时代的产品力问题：智能驾驶、座舱、电驱、制造、供应链。玛莎拉蒂可以继续输出它在中国高端消费市场里积累了二十年的审美叙事：比例、姿态、线条、品牌情绪，以及三叉戟带来的想象空间。换句话说，华为解决「这台车够不够强」，玛莎拉蒂解决「这台车有没有味道」。对华为来说，这更像是一次高端品牌方法论的调整：当技术已经足够强，华为开始意识到，除了技术之外，还可以尝试打开一扇「情绪」的门。其实，华为在智能手机业务上，早就摸透了这个道理，甚至完整经历过类似的周期。2024 年，华为做了一个重要决定，把主打影像和设计的 P 系列全面升级为 Pura 系列。这绝不仅仅是改个名字换个包装那么简单。Pura 系列在外观设计上，非常大胆地采用了三角形作为核心视觉符号。在手机圈，三角形比传统的方形更尖锐，比柔和的圆形更有攻击性，也比那些不规则的曲线更容易在消费者脑海中钉下识别度。这个超级符号，让 Pura 瞬间和过去的 P 系列拉开了距离，也为华为的影像旗舰留下了更先锋、更大胆的自我表达空间。当时的智能手机行业，和今天的新能源汽车市场有着惊人的相似。国产手机厂商已经在硬件参数、快充速度、屏幕素质、芯片性能上卷到了极致，大家都在追求一套标准的「最优解」。结果就是，市面上的旗舰机盖住 Logo 后，长得越来越像，体验也大同小异。今天的尊界，其实也走到了当年 Pura 面临的那个十字路口。过去几年里，「五界」体系通过问界、智界、享界等品牌，已经向市场清晰地回答了一个最基础的问题：华为到底能不能把车造好？但如果尊界和玛莎拉蒂的这次合作能够顺利推进，它必将成为鸿蒙智行体系里最特别、也最具野心的一次试验。因为接下来，尊界要向金字塔尖的用户群回答一个更难、也更感性的问题：当一台车在技术上已经足够聪明的时候，它还能不能做到足够迷人？

1968 年的旧金山，计算机科学家道格拉斯·恩格尔巴特在一场后来被称为「演示之母（The Mother of All Demos）」的发布会上，拿出一个带着两个金属轮子的木制小盒子，向世界介绍了一个新物种：鼠标。那是人类第一次公开用手上的鼠标，去牵引屏幕里的数字光标。此后的几十年里，这根小箭头几乎无处不在。它穿过办公软件、游戏界面、浏览器窗口和无数张电子表格，成为人类进入数字世界时最熟悉、也最沉默的向导。然而在这半个多世纪里，计算机的算力、形态和应用场景几乎改头换面，鼠标光标的本质几乎没有变过：它知道自己停在屏幕上的哪个坐标，知道 X 和 Y，却不知道你指向的是一行代码、一张发票，还是一张风景照。面对眼前不断闪烁的像素，它能做的事情依旧相当朴素：点击，拖拽，等待下一次点击。今天，Google 要用 Gemini 重新发明鼠标光标。在刚刚落幕的 Android Show 上，Google 几乎把围绕 Android、AI 与硬件生态的规划摊在了台面上。其中，一项名为「魔法指针（Magic Pointer）」的新功能，给古老的鼠标光标装上了「眼睛」和「大脑」。Google 的意图不言而喻，未来的 AI 交互不该依赖冗长的提示词，只需像在现实生活中一样，指着屏幕说一句：「把这个，移到那里。」那么问题来了，当鼠标光标终于学会「看懂」屏幕，它究竟会把人机交互带向哪里？这根睁开眼睛的 AI 箭头，到底能干嘛？要理解这项技术的意义，必须先看清当下 AI 工具最别扭的一面：交互成本。过去几年，大语言模型的能力一路狂飙，但使用门槛居高不下。为了让 AI 准确理解意图，用户被迫学习一套复杂的「提示词工程」：设定角色、补充背景、限定输出格式。为了一个简单的需求，写出几百字小作文更是家常便饭。不仅如此，典型的 AI 工具通常运行在独立的网页或应用窗口中，频繁打断用户的工作流。比如当你在阅读一份 50 页的 PDF 时想让 AI 总结一张图表，通常需要经历：截图 -> 保存 -> 打开浏览器 -> 进入 AI 网页 -> 上传图片 -> 输入提示词。Google 将这种繁琐的跨应用操作称为「AI 绕路（AI detours）」。这种跳转不仅效率低下，也容易打断人们工作时的注意力集中状态，即所谓的「心流」。为此，Google 提出的第一个交互原则，就是「保持心流」。在他们展示的实验性 AI 光标原型中，AI 的能力不再局限于某个特定的 App 或网页，而是附着在鼠标光标上，随时待命。触发方式也尽量克制：无需记忆任何快捷键，只需轻轻「晃一下」鼠标，AI 界面便会基于当前悬停的内容自动浮现，给出极具语境的操作建议。选中图，它会询问是否「对比」；悬停于段落之上，它会主动提供润色方案。整个过程没有任何需要学习的指令，完全跟着直觉走。来看几个极其直观的场景：第一，看图说话的终极形态。当你在浏览一张卡通的城市风景照时，传统鼠标只能点击图片放大。但现在，你只需把 AI 光标悬停在照片背景里的一栋建筑上，然后对着麦克风说一句：「给我移动图片的元素到这。」不需要解释「这儿」是谁，也不需要描述建筑外观。AI 光标会直接理解你所指向的像素，识别出对应元素，并成功移动。过去，鼠标只能告诉系统「我点了哪里」；现在，它开始告诉系统「我指的是什么」。第二，少写提示词，多用自然指代。当你在网页上看到一份极其复杂的烘焙菜谱，你不需要复制粘贴，也不需要写「请将以下菜谱里的所有食材分量乘以二」这种书面语。你只需要用光标高亮那段文字，然后随口说一句：「把『这些』的分量翻倍。」唰的一下，AI 直接在原地给你改写了一份新菜谱。第三，将像素转化为可交互实体。在计算机眼里，屏幕只是几百万个发光的像素点。但 AI 光标能将死板的像素（Pixels）转化为活的实体（Entities）。比如，你正在看一段旅行 Vlog，视频里闪过一家看起来很棒的餐厅。你按下暂停，把光标指过去，原本死气沉沉的视频画面瞬间变成了一个真实的、可交互的地点，旁边直接弹出了这家餐厅的订座链接。再比如，你随手拍了一张写满鬼画符的便利贴，鼠标一指，墨迹直接变成了一个可以打勾的 To-Do List。发现了吗？ 以前，是你去找 AI；现在，是 AI 顺着你的鼠标，乖乖来到了你的指尖。杀死 AI 提示词，回归人类直觉仔细想想，人类最强大的沟通工具其实是代词。当你和同事坐在屏幕前修改设计稿时，你绝对不会字正腔圆地说：「请将屏幕左上角坐标 (X:120, Y:350) 处的蓝色矩形向右移动 50 个像素。」你只会指着屏幕说：「把这个，往右挪一点，弄淡一点。」「那个餐厅看起来不错，怎么去？」「这段代码里的这个报错是什么意思？」在日常生活中，我们极度依赖「这个」和「那个」。手势配合着极简的口语，才是人类最高效的沟通密码。究其原因，我们身处同一个物理空间，共享着同一套视觉上下文。Google 敏锐地抓住了这一点，并将其提炼成了一条产品原则：拥抱「这」与「那」的力量（Embrace the power of This and That）。与其强迫人类去学习复杂的提示词框架，不如反其道而行之，把表达意图的脏活累活从我们身上剥离出去，让机器去适应人类最慵懒、最本能的「指手画脚」。好消息是，这套交互方式已经开始落地。Chrome 浏览器里的 Gemini 从今天起率先支持；Google 全新推出的笔记本电脑产品线 Googlebook，则将「Magic Pointer」直接内置进了操作系统层面，覆盖所有应用。Googlebook 的野心不止于鼠标。Google 将这个产品线定义为「Android 手机的完美伴侣」。类似苹果的 iPhone 镜像，用户可将 Android 应用无缝投射到 Googlebook 桌面，以原生比例运行，并在文件管理器中跨设备自由穿梭，彻底打破手机、平板与笔记本的生态壁垒。此外，Gemini 还能依你所需，在桌面生成专属动态 Widget（比如旅客的实时航班卡片）。在硬件设计上，所有 Googlebook 机型都会在机身上集成一条「Glowbar」发光条，让你一眼就能分辨出它与传统 Chromebook 或 Windows 笔记本的区别。首批 Googlebook 将由宏碁（Acer）、华硕（Asus）、戴尔（Dell）、惠普（HP）和联想（Lenovo）制造，预计今年秋天上市。有意思的是，三星缺席了这份名单。近期的消息显示，三星可能正在筹备一款搭载 Google 新系统的 Galaxy 笔记本，其下一场 Unpacked 发布会据传定在 7 月 22 日。至于底层的驱动核心，Google 虽未指名道姓，但通篇强化的「为智能而生的现代操作系统」以及 Android 与 ChromeOS 的深度融合，种种迹象均指向了传闻已久的 「Aluminum」系统。这意味着，AI 开始成了操作系统级别的基础设施。而当 AI 真正化身为你的鼠标光标，它便拥有了介入万物的权限——所见即所得，所指即所控。AI 人机交互，迎来十字路口回望 1968 年，那个惊艳世界的初代鼠标，功能简单得令人发指：追踪位置。 这五十多年来，鼠标加了滚轮、加了侧键、甚至加了风扇和配重块，但它的灵魂依然是一张白纸：它准确地标记着坐标，却永远无法理解坐标背后的意义。Google 的 AI 光标完成了交互史上罕见的进化：它不仅知道你在哪，更知道那是什么。过去一年，无数拿了融资的创业公司挤破头，试图造出下一个「AI 时代超级入口」。大家疯狂地卷对话框的拟真度、卷 Agent 的复杂工作流。但 Google 这次用行动给全行业结结实实地上了一课：最好的科技是什么？是润物细无声。聊天框（Chatbox）从来都不是 AI 的最终形态，它只是一种过渡期的妥协。最好的 AI，应该退居幕后，成为一种附着在你日常动作上的基础设施，而不仅仅是某个需要单独打开的应用。从黑底白字的命令行（CLI），到图形界面的鼠标点击（GUI），再到移动时代的触屏滑动（NUI）。过去几年，大语言模型让我们短暂地倒退回了打字交流的时代，让无数人患上了 Prompt 焦虑症。但今天过后，我们知道了，那不过是黎明前的一段弯路。真正好用的 AI，终究要学会像人类一样思考：看懂你的每一个眼神，听懂你的每一句「把这个，放到那儿去」。58 年前，当道格拉斯·恩格尔巴特握住那个简陋的木制鼠标时，他的终极梦想是「增强人类的智慧」。58 年后，当 AI 附身于这个古老的指针，机器终于开始真正「看懂」这个世界。属于提示词工程师的时代终将落幕，人机交互的终极闭环，也将在一句句含糊不清的「这个、那个」中，向前迈出历史性的一大步。附体验地址：https://aistudio.google.com/apps/bundled/ai-pointer-create?showPreview=true&showAssistant=true&fullscreenApplet=truehttps://aistudio.google.com/apps/bundled/ai-pointer-find?showPreview=true&showAssistant=true&fullscreenApplet=true作者：莫崇宇

📱iPhone 17 Pro 官降 1000 元🧠ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施🤖估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相🌐Claude 重新允许接入龙虾，开发者：这是变相涨价📋腾讯辟谣「AI 一号位将离职」💰黄仁勋基金会捐出 1.08 亿美元 AI 算力📱影石 CEO 回应 Luna 定价贵：5299 元是美国价格🚗鸿蒙智行洽谈打造玛莎拉蒂新能源车🚙华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗💻英特尔成为迈凯伦官方计算合作伙伴⛽本田上市以来首次年度亏损📋LinkedIn 宣布裁员，涉及工程、产品和市场营销部门💡Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」🤖曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线📱荣耀 600 系列定档 5 月 25 日🔍Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光🧠腾讯开源 AI Agent 记忆系统，Token 消耗最高降低 61%🧸泡泡玛特否认进军传统家电，LABUBU 冰箱一季度收入约 600 万iPhone 17 Pro 官降 1000 元昨晚，苹果无预警在天猫「Apple Store 官方旗舰店」启动 iPhone 17 Pro 系列 618 优惠活动，Pro 和 Pro Max 同步降价 1000 元，起售价分别来到 7999 元和 8999 元。本次活动还覆盖 iPad、Apple Watch、AirPods 等系列，非 iPhone 产品可使用 50 至 100 元不等的惊喜券，并可叠加 88VIP 消费券、店铺优惠及国家补贴等多重优惠。ChatGPT 进 Siri 未达预期，曝 OpenAI 拟对苹果采取法律措施据彭博社报道，苹果与 OpenAI 的合作关系已出现明显裂痕，OpenAI 正在联合外部律所评估可能的法律行动。知情人士透露，OpenAI 律师团队正与外部律所合作，研究包括向苹果发送违约通知在内的多项方案，但目前尚未正式提起诉讼。OpenAI 仍希望在庭外解决与苹果的分歧，相关法律行动预计将在马斯克相关诉讼案结束后才会推进。苹果与 OpenAI 的合作始于 2024 年 6 月，苹果在 WWDC 上将 ChatGPT 整合进 Siri 及 iOS 系统，涵盖文字生成、视觉智能、图像生成 App「图乐园」等场景。估值 46.5 亿美元，前 Meta FAIR 总监田渊栋创业公司正式亮相昨天，前 Meta FAIR 研究院总监田渊栋正式宣布联合创办 AI 初创公司 Recursive Superintelligence，距其确认从 Meta 离职刚好过去半年。他在 X 发文称，拒绝了 OpenAI、xAI、Anthropic 等多家头部 AI 公司的邀约，选择自己创业：「趁着自己还年轻，去当一家新初创公司的联合创始人。」Recursive 同步披露完成一轮 6.5 亿美元融资，估值达 46.5 亿美元，由 GV（谷歌风投）和 Greycroft 领投，AMD Ventures 与英伟达跟投。公司押注的核心方向为「递归自我改进」（Recursive Self-Improvement），即构建能够自动发现知识、持续优化自身的 AI 系统。Claude 重新允许接入龙虾，开发者：这是变相涨价Anthropic 昨天官宣，将于今年 6 月 15 日起为 Claude 订阅版用户引入独立的 Agent SDK 月额度，OpenClaw 等第三方 Agent 工具由此重新获准通过订阅账户接入 Claude。按照新规，Agent SDK 的调用将从用户的常规订阅限额中彻底剥离，另设专属月度额度池，额度按账单周期刷新，未用完的部分不结转至下月。Pro 版和 Team 标准席位每月获得 20 美元额度；Max 5x 及 Team 高级席位为 100 美元；Max 20x 与 Enterprise 座位制高级席位最高可达 200 美元。该额度专用于 Claude Agent SDK 项目（Python 或 TypeScript）、claude -p 非交互命令、Claude Code GitHub Actions 集成，以及通过 Agent SDK 认证的第三方应用。此次新规被视为 Anthropic 对今年 4 月 4 日封禁政策的正式解决方案。彼时，Claude Code 负责人 Boris Cherny 宣布，Claude Pro 及 Max 订阅额度即日起不再覆盖 OpenClaw 等第三方工具的调用，理由是此类工具的使用模式对 Anthropic 的算力基础设施造成了难以为继的压力。开发者社区对此反应冷淡。有开发者直言，20 美元的 Agent 额度在高强度任务下消耗极快，实际可用量远低于封禁前的订阅套餐，并将此次调整定性为「变相涨价」。腾讯辟谣「AI 一号位将离职」昨天，腾讯官方发文辟谣，明确否认外界流传的「AI 一号位即将离职」传言。腾讯公关总监张军表示，「太恶劣了，还编出各种搞笑的情节。」谣言起源于社交平台上的多条帖文，内容称某大厂 AI 一号位因被绿色事业部抢走算力资源而准备离职，并点名影射腾讯 AI 负责人姚顺雨。黄仁勋基金会捐出 1.08 亿美元 AI 算力据路透社报道，英伟达 CEO 黄仁勋与其妻子洛丽创立的基金会，从云计算公司 CoreWeave 购入 AI 算力，并将其捐赠给多所大学及其他非营利研究机构，用于科学和人工智能研究。根据昨日提交的一份文件，此次捐赠的算力价值目前已达 1.083 亿美元。英伟达还表示，计划向部分受捐机构提供免费工程服务。CoreWeave 是一家专注于 AI 应用的云计算公司，其向客户提供的 GPU 正是由英伟达设计制造。今年 1 月，英伟达向 CoreWeave 投资 20 亿美元，一度成为该公司第二大股东。影石 CEO 回应 Luna 定价贵：5299 元是美国价格昨天，影石 CEO 刘靖康回应旗下旗舰双摄云台相机 Luna Ultra 的定价问题，表示网传单机 5299 元的售价属实，「但是是美国的价格，国内给我再大的胆子都不敢卖这个价格」。此前，博主「数码闲聊站」爆料称该机预计售价 5299 元、全能套装预计 6499 元，随即在网络上引发「定价过高」的质疑。鸿蒙智行洽谈打造玛莎拉蒂新能源车据《云见 Insight》报道，华为鸿蒙智行、江淮汽车、斯特兰蒂斯集团（Stellantis）及旗下玛莎拉蒂品牌，正就共同打造玛莎拉蒂品牌的新能源车展开洽谈。按照规划，四方合作模式与鸿蒙智行 「五界」 高度相似：华为主导产品定义、提供核心技术，江淮联合研发并负责生产制造，玛莎拉蒂提供造型设计与品牌背书。合作车型将分为国内和海外版本，国内版归属于尊界品牌，海外版将挂玛莎拉蒂车标，首款车型目前正处于造型设计阶段，计划于明年下半年量产。华晨宝马年度节水数据出炉：再生水替代自来水，三年连续压减总消耗华晨宝马昨日公布年度节水数据，显示其沈阳生产基地 2025 年单台生产取水量降至 1.4 立方米/台，同比降幅 10.8%，连续八年稳步下降。在水资源消耗总量方面，华晨宝马沈阳基地全年取水总量降至 75.2 万立方米，较 2023 年降幅超 40%，连续三年递减。再生水使用量超 68 万立方米，同比增长 71.9%，官方称节约用水量可满足近 9300 户沈阳家庭一整年的生活用水需求。华晨宝马大东工厂北厂区实现所有生产工艺自来水消耗为零，全部改用再生水替代；铁西工厂及里达厂区涂装车间冷却塔也已扩展使用市政再生水，厂区绿化灌溉此前已于 2024 年实现 100% 使用市政再生水。英特尔成为迈凯伦官方计算合作伙伴英特尔昨日宣布与迈凯伦车队达成多年期战略合作伙伴关系，正式成为迈凯伦万事达一级方程式车队、艾睿迈凯伦印地车队及迈凯伦 F1 模拟赛车队的官方计算合作伙伴。根据合作协议，英特尔至强处理器和酷睿 Ultra 处理器将为迈凯伦的核心性能负载提供算力支持，覆盖计算流体动力学、空气动力学分析、车辆动力学仿真、比赛策略分析等方面。英特尔的技术还将应用于赛道边缘计算、数字孪生与 AI 预测建模等场景。本田上市以来首次年度亏损昨天，本田公布了 2026 年 3 月期的合并财报，期内营收 217966 亿日元，同比微增 0.5%；归母净亏损 4239 亿日元，上一财年同期归母净利润为 8358 亿日元，由盈转亏。亏损主因是电动化战略大幅调整。本财年内，本田决定取消北美多款电动车型的开发与上市计划，由此确认减值、处置及合同赔偿等相关损失合计约 15777 亿日元，全部计入汽车业务。摩托车业务为本财年唯一实现正向盈利的板块。展望明年，公司预计营收 231500 亿日元，归母净利润恢复至 2600 亿日元。LinkedIn 宣布裁员，涉及工程、产品和市场营销部门据彭博社报道，职业社交平台 LinkedIn 昨日宣布裁员，涉及工程、产品和市场营销等多个职能部门。LinkedIn CEO Daniel Shapero 在发给员工的备忘录中表示，公司需要为用户提供更大价值，并实现更高的盈利水平，此次组织架构调整旨在为公司未来发展做好定位。LinkedIn 发言人随后确认了相关变动，称此次调整属于「常规业务规划」的一部分，但未披露具体裁员人数。💡 Anthropic CFO：AI 已经编写 90% 以上代码，人类角色转向「监工」据《商业内幕》昨日报道，Anthropic CFO Krishna Rao 在播客节目《Invest Like the Best》中透露，Claude 目前已承担公司 90% 以上的代码编写工作，并正在深度介入财务流程，推动白领工作从「执行」向「监督」转型。Rao 还透露，公司已借助 Claude 生成财务报表，每月财务审查流程在人工介入之前已达到「90% 至 95% 完成」的状态。他提到，部分内部报告的生成时间从原来的数小时压缩至 30 分钟。在人员配置方面，Rao 将 AI 的引入描述为生产力「加速器」。他称，由于 AI 的介入，员工得以将更多精力从信息收集转向决策判断，Anthropic 反而因此扩大了招聘规模。我们雇用了更多人，因为没有什么工作是做不完的。Rao 还描述了一种新的工作形态：员工越来越多地扮演 AI 系统的「管理者」角色，团队可以同时部署多批 AI 智能体并行处理项目。他将这一趋势概括为「每个人都在某种程度上成为管理者」，并认为由此带来的生产力潜力「才刚刚开始显现」。他同时强调，Anthropic 内部推崇「人才密度」而非「人才规模」，认为高密度的顶尖 AI 研究人才配合最强模型，才是真正具有竞争力的组合。曝 GPT-5.6 内部测试已启动、旗舰模型「超快模式」本周上线知名爆料者 Leo 昨日披露，GPT-5.6 研发已全面推进，预计最早下月正式发布。开发者在 OpenAI 内部 Codex 日志中发现了两个内部代号 —— ember-alpha 和 beacon-alpha，以及一条路由映射记录。日志显示，绝大多数模型调用仍指向 GPT-5.5，但有一条条目明确指向 GPT-5.6，暗示 Codex 环境可能已在测试新模型。此外，爆料者 Chetaslua 称 OpenAI 计划于本周四推出旗舰模型的「超快模式」，响应速度或可提升 2 至 3 倍，主要面向延迟敏感型任务。荣耀 600 系列定档 5 月 25 日荣耀昨日宣布，旗下荣耀 600 系列将于 5 月 25 日 19:00 正式发布。官方预热信息显示，荣耀 600 系列主打「4K 闪光微单 Live」影像方向，搭载 2 亿像素超清大底主摄，并内置 8600mAh 青海湖大电池。Google I/O 前夕，Gemini 3.2 Flash 与多款「Agent」模型提前曝光据 Sources 报道，Google 计划在 5 月 20 日举行的 Google I/O 大会上

相信这样的场景很多人都不会陌生：在微信上聊工作、讨论旅行计划，甚至只是和朋友闲聊，聊着聊着突然想让 AI 帮忙总结重点、做个规划，结果却只能一段一段复制聊天记录，再手动粘贴进 AI 对话框里。再比如，中午小憩了一会，醒过来发现群里发了几百条消息，想吃瓜又得一条一条翻……如果能一键将聊天记录转发 AI，岂不美哉？这样的功能终于来了，但遗憾不是在微信里：现在，把微信和元宝更新到最新版本之后，聊天记录转发可以选择直接发给元宝来总结。虽然同属腾讯生态，但微信和元宝之间始终有种微妙的距离感——终于，这次两者之间的联动，又紧密了一点。爱范儿也第一时间体验了这个功能，看看它究竟好不好用。产品很美好，效果很一般要想把微信聊天记录转发给元宝，方法很简单，首先选中想要转发的聊天记录，选择「转发到其他应用」，就能直接发给元宝，可以补充一些提示词。整个过程有点像把讨论、聊天过程中零散的内容，让 AI 帮忙进行整理，因此这个功能首先最合适的场景，就是各种工作群的讨论和安排信息。直接处理聊天记录有一个好处，元宝可以将讨论精准和发言者匹配，很适合用来处理会议纪要和工作安排。旅游计划同样也是一个高频场景，因为行程一般都是靠你一言我一句的对话讨论出来，这时候就可以直接交给元宝处理，还能让它帮忙推荐旅游地点。还有一个有点邪修的用法：平时工作忙没时间看群，又担心错过什么热点或者话题，可以直接选中所有聊天记录，让元宝帮你看群帮你总结。又或者讨论了很久都不知道聚餐吃什么，也可以把讨论的过程选中，让元宝帮忙推荐。尝试了很多场景，大部分情况下都达到了预期，但总体来说，还有一些体验细节不够完善。目前元宝最擅长处理的聊天记录以文字为主，而如果聊天记录中包含文件，特别是涉及多个文件，效果会比较随缘，很多时候会无法读取文件。即使能成功读取，也不一定能按照提示词进行任务处理：整理发票有时候都未必能填对，让他根据文件做个思维导图，也经常失败。要知道，整理聊天里的发票真的是很主流的场景啊！我能很明显感觉到一个落差：明明将聊天记录发送给 AI 是一个很有价值的场景，但元宝交付的结果，总让我略感失望，并且忍不住想，如果能转发给豆包或者 ChatGPT，会不会更好？更深层的问题，是这个聊天记录转发的交互总给我一种「绕远路」的感觉。在此之前，元宝已经作为联系人存在于微信之中，处理聊天记录这件事，却依旧需要「跳转」这一整套外挂式流程， 过程中需要多次点击。或许是因为微信严格的隐私保护规则，元宝只能以「临时聊天」的形式处理聊天记录，不会留下历史记录，关闭后无法找回。对于微信来说，这样的功能似乎「可有可无」。如果微信想要 AI 化，他们或许会有更好的实现方式，例如直接做一个「AI 群友」，at 它就能回答问题或者处理聊天记录；以及一个常驻对话列表的 Agent，可以直接对话，可以直接接收聊天记录。但这件事没那么容易，原因已经是「老生常谈」：微信需要简约，需要克制，因为每一点小变化都会影响到十亿级用户。在元宝这个功能上线的同一天，腾讯发布了今年第一季度财报，在股东大会上马化腾表示，在微信里做 Agent 需要「更长远的考虑，大家要有点耐心，这个不是匆忙就能出来的。」腾讯的 AI 王牌是微信马化腾还提到腾讯的 AI 战略：原来一年前我们以为上了船，后来发现那个船漏水了，现在感觉站上去了，还坐不下去，还是希望船速能快一点。从财报来看，AI 已经成为了腾讯的重要投入，一季度研发投入达 225.4 亿元，同比 19%；资本开支 319.4 亿元，主要用于算力、存储等云端基础设施的扩张。今天中午，有自媒体传出「疑似腾讯 AI 一号位即将离职」的消息，下午就被腾讯公司火速辟谣，本身也向外界释放出一个信号：腾讯 AI，不容有失。根据 AICPB 最新榜单，在国产 AI 应用排名中，腾讯元宝排名第十，访问量为 1773 万，远逊于豆包（1.63 亿）和千问（4424 万）。在 AI 领域中，腾讯和苹果有着相似的处境：产品入口和用户规模都是顶级配置，AI 模型本身却还支撑不起这种期待。元宝 AI 社交应用「元宝派」生成式 AI 火了 3 年，从去年年底开始，AI 应用之间的战斗已经进入了「场景化」的比拼，不仅出现「龙虾」OpenClaw 这样的全新形态，传统的 ChatBot 交互模式也集成了生活服务，像阿里千问和字节豆包，都在积极将电商等各种生态内的优势服务，和自家的 AI 助手相结合。这本来应该是腾讯的舒适区，因为国内最强大、最全能的生活应用，也就是微信，就在腾讯手中。但很明显，微信对「AI 化」抱有非常慎重的态度。腾讯不能承受因技术或功能不成熟而扰乱微信庞大用户群的体验。只是对于腾讯来说，在 AI 性能落后其他对手的情况下，微信这张王牌的价值还是太重要了。，于是，我微信和元宝虽然没有完全敞开双臂拥抱彼此，也开始小心翼翼试探边界。究其原因，除了微信相对封闭的生态逻辑，还有两个问题没能被回答：微信究竟要如何「AI 化」？而元宝在其中又应该扮演什么样的角色？The Information 报道指出，即使腾讯在下重本加速建设自己的混元模型，微信方面还是认为混元当前水平不够拔尖，甚至开始为微信 Agent 测试第三方 AI 模型，这带来了隐私和整合上的挑战性。微信方面，爆料称其实一直在做自己的 AI Agent 模型技术，计划在年中发布一个微信智能体，可以和整个小程序生态相结合，提供打车、购物的服务，并且将以联系人和聊天的方式直接集成微信之中——但就像前面提到的，这件事太重大、影响面太广了，急不得。一旦这个产品问世，可能将是全国最多用户的 AI 入口。微信需要 AI，但未必需要「元宝」，这才是腾讯现在最微妙的问题。

大疆的 Pocket 系列三年不更新，一更新就来了两台——过去一个多月，Pocket 4 和 Pocket 4P 可以说是最受瞩目的相机，就在今晚的戛纳电影节上，Pocket 4P 以口袋电影机的身份在欧洲电影殿堂亮相，距离正式发布时间，相信不会太远了。与此同时，大疆也对 Pocket 3 进行了降价处理，于是我们看到了泾渭分明的 Pocket 系列产品线：Pocket 3：入门级的手持云台 vlog 机，2299 元Pocket 4：全能的手持云台相机，2999 元Pocket 4P：准专业级的口袋电影机，预计价格会来到 4000 元档相信有许多持币观望的读者，也和爱范儿抱着同样的疑问：三台 Pocket 我该怎么选？有了 iPhone，我还需要 Pocket 吗？结合实际体验，爱范儿一次性帮你理清楚。Pocket 三代同堂，怎么选？先仔细聊聊 Pocket 4P。虽然目前 Pocket 4P 依旧没有公布更详细的参数，但参数在本次形态更新面前，是最不重要的话题——从产品逻辑来看，Pocket 4 已经将单摄的手持云台相机推向较为成熟的状态。此时，增加一颗摄像头成为一件顺理成章的事情。绝大多数手机的主摄和单摄 Pocket，提供的都是广角视野。广角适合交代环境，适合把大山大河框进画面。但当你拿着设备走入拥挤的人群，广角的缺点就暴露无遗：画面太杂乱了。此时，一颗长焦镜头可以完成空间上的压缩，把无关的背景剔除出去。大疆将这颗多出来的摄像头定格在 3×，这是一个非常巧妙的数字。它比主摄更贴近人眼对焦后的视线范围。从我们上手体验来看，Pocket 4P 这颗 3× 镜头还提供了比较自然的虚化，非常适合拍摄更凸显主体的人像或特写画面。但对于用户来说，多的这一颗摄像头，并非绝对的正收益：Pocket 系列自诞生起，核心卖点就是「无感携带」，增加镜头，意味着增加重量、侵占内部空间，甚至可能牺牲一部分续航。其实这种担心，恰恰帮大疆划清了 4P 和 4 之间的界限。Pocket 4P 是给那些「创作者」准备的。他们清楚地知道自己要拍什么，素材最后会进入 Mac，在 DaVinci Resolve 或者 FCPX 里进行精细的剪辑。对他们来说，为了在极轻便的体积下获得原生焦段的景别切换，以及高度统一的素材规格，牺牲一点点重量完全值得。而对于大多数普通人来说，Pocket 4 才是那个更轻量、更理性的选择。虽然前代 Pocket 3 打造出千万台的神话，但其本身并不是一台完美无缺的相机——只有 USB 2.0 导致数据传输过慢、像素量过低以及动态范围不够大，这些都是问题。而 Pocket 4 提供了 3700 万像素，更高的像素意味着内置的无损 ISZ 变强了。在 Pocket 4 上，通过传感器裁切获得的 2× 焦段更为清晰，足以应付日常记录。除此之外，它还内置了 107GB 存储，带来了更好的动态范围。忘记带内存卡、传输速度和动态范围不够的问题，被一次性解决掉了。可以说，Pocket 4 就是目前单摄手持云台相机的完全体。至此，Pocket 4 系列两台机身的区别已经很清晰了：如果你追求生产力，强调原生光学焦段带来的画质与虚化，需要 3× 镜头获得更丰富的画幅景别，那么 Pocket 4P 毫无疑问是更好的答案；如果你只是希望有一台能轻松塞进裤兜、防抖优秀、可以独立于手机拍摄视频和 Live Photo 的机器，那么 Pocket 4 凭借更高的像素和好用的 2× ISZ，已经绰绰有余。最后是 Pocket 3。在 Pocket 4 补齐了大量痛点的情况下，Pocket 3 还有存在的意义吗？当然——新设备的出现，最大的贡献往往是把老旗舰的价格打下来。在 Pocket 4 系列轮番上场的间隙，我们观察到 Pocket 3 已经完成了一轮官方降价，在某些平台叠加各种补贴的情况下，已经落到了 2099 的价格。甚至在未来，我们能以 1 开头的价格买到 Pocket 3，也不是不可能。对于预算有限、或者刚刚萌生拍视频念头的新手，这就是性价比最高的 「守门员」。最后，按照爱范儿编辑部的推测，Pocket 4P 的定价大概率在 4500 元左右，但也不排除大疆卷起来定一个 3999 元的价格。此时，如果将 Pocket 系列按照价格排开，我们会发现一个巧妙的设计——从 2000 元档的 Pocket 3 到 3000 元档的 Pocket 4，最后到 4000 元档的 Pocket 4P，大疆正在把无人机上那套 Pro、Air、Mini 的组合拳搬到口袋相机上。这套阵容已经在无人机市场证明了它的成功。现在，Pocket 也将组成这样的阵容了。有了 iPhone，我还需要 Pocket 吗？在 Pocket 4P 的上手文章发布后，我收到不止一个朋友的问题：我有 iPhone，还需要买 Pocket 吗？理清了 Pocket 家族的内部定位后，这是我们必须面对的现实问题：既然大家手里都有一台性能强悍的 iPhone，为什么还要花钱买一台功能单一的云台相机？这其实是一个关于「设备属性」的哲学问题。在多次迭代后，iPhone 在动态影像领域的基础不可谓不强，三摄变焦顺畅、色彩高度统一、防抖优秀、Apple Log 甚至已经更迭到第二代。但问题在于，手机始终不是完全针对拍摄设计的特化产品——除了拍摄以外，iPhone 还需要承担复杂的日常任务，无时无刻需要分散注意力，随时可能会有工作消息、语音电话打断你的拍摄、需要考虑拍摄之外手机还有没有充足的剩余电量保证你与世界的关联，还需要注意拍摄会不会占满本就贵如黄金的存储空间……而 Pocket 则是完全服务于拍摄需求的相机：不能打电话，不能回微信，不能刷社交软件，一切与拍摄无关的东西，都不会出现在你手里的 Pocket 上。更别提，Pocket 还有机械云台天生在结构上比 iPhone 的防抖更稳定、形态比 iPhone 更适合握持、物理虚化比 iPhone 更自然、视频规格比 iPhone 更丰富等等综合优势了。所以，如果我手里已经有一台主力 iPhone，现在到底该不该买 Pocket？如果现在需要，哪台是最适合我的？你可以对照下面几个问题，做一个购前诊断：过去三个月里，你是否因为手机突如其来的电话或微信，打断过拍摄？拍视频时，是否因为手机存储空间不足，而束手束脚？炎炎夏日，有没有遇到过手机拍照录视频发热掉帧甚至死机的情况？如果你对这些问题的回答是肯定的，那么购入 Pocket 3 和 Pocket 4 都是不错的选择。具体来说，如果你有拍高清 live 图的需求，如果你喜欢用手机的 2× 镜头拍摄，那么选择 Pocket 4，否则，Pocket 3 其实就够用了。在这个基础上，你还想拿 Pocket 拍拍商单、拍拍演唱会，搞点严肃创作，那么 Pocket 4P 是一步到位的选择。当然，如果前面几道题，你的答案都是否定的，那么请捂紧钱包。iPhone 的防抖和画质足够满足你的需求。你可以随拍随剪，直接通过 AirDrop 分享，且丝滑的生态体验依旧是目前任何第三方设备都无法替代的。而从行业角度来看，爱范儿更建议你等一等——目前，口袋云台相机的玩家可不止大疆一家——和 Pocket 4P 同样「犹抱琵琶半遮面」的，还有影石 Luna，而各大手机厂商也正跃跃欲试，准备下场分一杯羹。因为 Pocket 4P 这样的产品，已经严重威胁到了手机厂商的安全区。很多人愿意买 6000 元甚至 8000 元的影像旗舰手机，很大程度就是为了长焦。现在，花更低的价格，Pocket 4P 就能提供一个素质相当的长焦镜头，以及更好的云台防抖、更专业的整机解决方案。没有手机厂商能对此熟视无睹。据爱范儿了解，vivo、OPPO 都已经将手持云台相机的产品规划提上日程，而小米、华为也在持续关注这个赛道。2026 下半年到 2027 上半年，想必会有更多手持云台相机纷至沓来。既然入局者多了，我们作为消费者，反而可以更淡定些。你可以把钱捏在手里，再等一等。我们相信，这个年轻的品类，还会有更精彩的产品出现。

鸿蒙智行车型阵营里可能缺一辆 MPV，但中国汽车市场已经不缺 MPV 了。别克 GL8，丰田埃尔法和传祺 M8 等车型在燃油车时代占据了 MPV 就是商务接待用车的印象之后，新能源 MPV 迎来了数款重量级车型，极氪 009 主打高端，意在取代丰田埃尔法乃至雷克萨斯 LM 的生态位，理想 MEGA 则希望利用 MPV 的空间优势重塑大家庭用车的产品定义，岚图梦想家成为这家新势力最依仗的产品成功和别克 GL8 分庭抗礼，小鹏 X9 想要以宜商宜家，是 MPV 也可以是 SUV 的灵活定位吸引更多受众……新旧势力在 MPV 市场的攻伐，发生在 2023 年，但现在已经是 2026 年了，零跑做一款旗舰 MPV 只需要把性价比拉到极致，但鸿蒙智行做一款旗舰 MPV 并卖四五十万元，需要做的就多了，当然我承认坐在二排的老板们对智界代言人刘亦菲肯定都是心驰神往，但除了刘亦菲这个核心竞争力之外，鸿蒙智行的首款 MPV 肯定需要更多击中老板坚硬内心和柔软脂肪的东西。这些东西是车载制氧系统、全维包裹安全气囊、华为悦彰非凡系列音响、二排主动座椅旋转、侧门交互式迎宾光毯等等等。最终，智界 V9 分为 Max、Max+、Ultra 和 Ultra+ 四个版本，售价分别为 39.98 万，42.98 万，45.98 万和 51.98 万起。舒适重要，安全重要，舒适且安全更重要！过去几年的时间里，鸿蒙智行成功地让国内汽车行业重视起来零重力座椅和主动安全来，是的，这两项鸿蒙智行都不是发明者，但确实在打开用户心智和开拓市场上起到了主要作用。前者代表座舱的舒适性，通过模拟太空失重环境下人体自然放松姿态的设计。通过特殊的人体工程学优化，使臀部、腰部、背部和腿部受力均匀，释放肌肉压力。后者代表汽车安全从被动防御走向主动避让，在汽车发生事故前，通过电子控制系统感知环境、预判风险并主动介入（如自动制动、转向控制），从而避免或减轻事故。但是，二者的矛盾在于在汽车行驶状态下，零重力座椅一般是没法使用的，这是因为当乘客躺在大仰角的座椅上享受时，身体与安全带、气囊的相对位置已经发生了相当大的变化。董车会之前报道过，根据 CIDAS 的数据，一旦座椅靠背夹角超过 115°，正面碰撞中的重伤风险会陡增 4.4 倍。此时，传统的 B 柱固定式安全带无法再有效约束骨盆，人体会因为惯性产生极度危险的「潜滑现象」——整个人滑向座椅下方，安全带可能会锁死在颈部导致窒息。采埃孚的试验甚至显示，在 50km/h 的碰撞下，躺倒状态受到的脊柱伤害堪比从 5 楼坠落。打过工的人都知道，当你向老板解释「零重力座椅虽然很舒服，但是不能在行驶状态下使用」时，老板这个时候一般不会听你的解释，只会跟你说「我不管，我想用，给我想一个解决方案出来」。今天，这个方案就被智界 V9 想出来了。智界 V9 全球首发了二排全维包裹安全气囊，这个气囊集成于二排座椅靠背，触发瞬间会形成一个独特的「U 型」环绕结构，在零重力、躺姿等多种姿态下可以全面防护头部、颈部、胸部，有效防止乘员在车辆碰撞之后身体滑落、碰撞与下潜，智界宣称历经百项严苛工况验证，无误伤风险。与之配套的还有二排坐垫前端也有专门布置的臀部气囊（BAB），在碰撞瞬间快速充气形成物理坡度，顶住盆骨不让身体滑脱，从而大幅降低腰椎压力。做了这么多工作，目的只有一个：就是让二排的老板在车辆行驶的时候好好躺一躺，好好休息一下。车载制氧：让高精力人群更有精力一直以来，对于什么样的人能做老板有个迷思，甚至还有「基因彩票」的说法，就是高精力人群更容易创业做老板。这些人可以每天工作十几个小时不知疲倦，只需要睡四五个小时就能满血复活。正如前面所说，很多老板之所以看起来可以每天工作十几个小时不知疲倦，很有可能是他可以中间随意选择休息时间。而屏幕前后的你和我，并不能在上班时间睡觉。更科学的说法是，高精力人群背后，跟摄氧量有正向关系，很多所谓的低精力人群很可能有鼻炎咽炎或者心肺功能低下导致日常处于轻微缺氧状态，导致精力不足。类似的情况也会发生在车上，很多人坐高铁感觉昏昏欲睡就是车厢里二氧化碳含量高，氧含量低，这种情况就容易困，密闭的汽车座舱也可能出现这种情况。智界 V9 给出的解决方案也很简单，装个制氧系统。这套制氧系统采用了 PSA（Pressure Swing Adsorption）分子筛变压吸附技术，提供了「弥散+鼻吸」两种模式，鼻吸出氧浓度 93±3%。额定制氧排量 8L/min，峰值 10L/min，比一般的家用制氧机大很多。2 个鼻吸口支持 1 拖 2，最多可供 4 人同时吸氧，弥散供氧与空调共用，四个出风口协同工作，30 分钟即可让整个座舱达到富氧状态。除了解决了传统制氧机体积大，噪音大的不足之外，鸿蒙智行的看家本领是智能化。这套制氧系统支持语音控制开启，中控屏实时显示状态；内置舒缓解压与高原出行两种模式，可自动调节供氧浓度，也可以联动小憩模式，帮助身心放松。很多人在森林景区，绿肺公园里感觉心思空明浑身舒爽，也是因为在富氧环境下，血氧饱和度一般会更高一些。富氧环境在特定场景下确实会有积极作用，比如早起后在通勤路上进入富氧环境有助于从昏沉状态中恢复；在运动后进入富氧环境可以加速乳酸代谢，缓解肌肉酸痛。更典型的场景就是针对本身心肺功能或者呼吸系统有问题的人，富氧环境有助于把缺氧状态调整为正常状态，还有就是高原出行环境，正常的氧气浓度可以有效缓解高原反应。在制氧系统之外，针对座舱健康还有鸿蒙 ALPS 健康座舱 3.0，通过把 MOFs 可吸附降解材料融入顶棚、门板、地毯等内饰，即便熄火也能持续吸附有害物净化醛类。UVC 空调主动杀菌，对流感等病毒灭杀率超 99.99%；前后双负氧离子净化器，可以释放超过西双版纳雨林 5 倍的负离子；这一整套净化杀菌和补氧系统，提供的是「健康」的价值，毕竟对于老板来说，紧急时刻生命第一，不紧急时刻健康第一。当然，需要着重说明的是，并不是所有人在所有场景都需要制氧系统制造一个富氧环境，如果本身不处于缺氧状态，血氧饱和度在 95% 以上，不需要运动恢复，没有昏沉嗜睡情况，不在高原地区，那就不太需要吸氧，长期过度摄氧反而有害健康。旋转座椅，优雅，还是优雅冰箱彩电沙发之所以从争议性的玩笑变成了兵家必争之地，还是市场表现的结果，谁不想二三十万的车就配有之前只有 78S 系才能选装的配置呢？这种造车逻辑之下，成本的分配也就必然从内燃机变速箱往舒适性配置上调整了。智界 V9 的沙发，显然值得多点笔墨。双独立结构，通风加热和比气动按摩更强劲的机械按摩已经不足为奇。更关键的是，二排座椅有可旋转设计，最大旋转角度 180 度，可以在多种角度进行旋转和固定，营造了例如 45° 迎宾座椅，90° 观景座和 180° 旋转之后的后排会客室布局等多种模式和选项。尤其是迎宾形态和会客室形态，可以在商务场景里营造比普通 MPV 更优雅的氛围。出门在外，面子可以是自己给的，但归根到底，还是别人给的，给别人面子就是给自己涨面子。智界 V9 类似的优雅配置还有不少，比如说「手势控车」，佩戴支持功能的手表，做出轻捏动作就可以隔空打开车门；二排乘客在车内做一个推一推手势，二排侧滑门就可以优雅闭合；窗边手指划一划，就可以调节智能隐私光幕比例的明暗度……车内的四块智能隐私光幕玻璃最大遮光率可达 99.6%，如果一键开启隐私模式，可以立刻断开车内的位置信息、麦克风、摄像头、网络，保证车内隐私。新车配有超 50㎡ 静音材料覆盖和 13 块双层夹胶隔音玻璃进行被动降噪，同时还可选 RNC 主动降噪。其他一些老板不太关心，但司机需要关心的作为一款鸿蒙智行的车，其实很多关于底盘，辅助驾驶和动力的信息都能猜个大差不差，这里完全可以用列表形式一一列举，毕竟这些信息是给司机看的，而且司机一看就懂，无需解释。新车长宽高为 5359×2009×1879 mm，轴距 3250 mm，2+2+3 座椅布局，后备箱容积 892L（可扩容至 2002 L）。全系标配 38 个传感器，包括 896 线激光雷达，背靠 60 EFLOPS 云端 AI 算力、WEWA 2.0 架构，可首批升级华为乾崑智驾 ADS 5。搭载 1.5T 雪鸮增程器配合前后双电机，综合功率 380kW，搭载 53.4kWh 电池，CLTC 纯电续航 305km，综合续航 1320km。新车有辉光紫、深海蓝、雪域白、鎏金黑四种车身配色和韶华杏、赤茶橘、绒霞紫三种内饰颜色。搭载 MPV 版图灵平台，前双叉臂与后 H 臂多连杆独立悬架，配合双腔空气悬架及 CDC 连续可变阻尼减振器。全系标配±7° 主动后轮转向，5.35 米转弯半径，支持低速蟹行模式，有效应对狭窄路况的调头和穿行。四活塞前固定卡钳，35.8m 制动距离；500mm 涉水深度，保证暴雨积水情况下的通行能力。路面预瞄系统智能识别井盖、凹坑与弯道，智能调节悬架阻尼、刚度，以及电机扭矩，可以降低冲击与侧倾，获得莱茵晕车舒缓认证。行业首发 2 根 2400MPa+6 根 2200MPa 侧门防撞梁；全系标配 13 个安全气囊，包括全维包裹安全气囊。前环宇三联屏，12.3 英寸液晶仪表与双 17.2 英寸 3.4K 大屏组合，可选后排 21.4 英寸吸顶屏。因为鸿蒙座舱，乾崑智驾，还有图灵底盘等鸿蒙智行基础产品技术支撑，现在鸿蒙智行的车型想要出挑有点困难，但毫无疑问，无论是华为还是奇瑞，都给到了智界 V9 足够的重视，比如单独给这款车开一场声势浩大的发布会，这在整个华为体系里是极为罕见的待遇。再比如，华为和奇瑞给这款车投入了一些深刻的产品洞察去迎合和吸引智界 V9 的潜在受众，很多人可能对这样一款华为气质非常浓郁的 MPV 产品无感，但华为不在乎，华为只希望这些人快快长大并事业有成。

发布会开场，李想先是强调：其实大部分用户并不知道自己想要什么。他分享了一个故事：有一次我去买烤冷面，摊主问我还要加点什么，我开玩笑说能不能打个折。你看，用户永远想要更便宜，但摊主会告诉你，其实还能加蛋。这也是 L9 当年销量爆发的秘诀——当行业还在比拼基础参数的时候，理想顺手给车里塞进了冰箱、彩电和大沙发，交出了一份远超预期的答卷。四年过去了，全行业都开始把屏幕做大、座椅做软。当堆砌配置成为车企的常规操作，理想 L9 又走向了另一条路。如今的理想把大量精力和资金砸向底层架构，自研范围覆盖车规级算力芯片、800V 主动悬架、全线控底盘等多个领域。承载这些研发成果的，自然是今天上市的两款新车型：L9 Ultra 和 L9 Livis。首先值得肯定的是，理想没有盲目放大 L9，车长和车宽依然维持在 5.25 米和 2 米，日常开起来依旧方便。在这个好开的尺寸基础上，L9 的整车姿态还挺拔了不少。在车高仅增加 10 毫米的情况下，前机盖和水切线被大幅拉高，配上 22 寸大轮毂，原本庞大的车身显得颇为矫健；前唇那一抹金灰色镀铬饰条也收束得干净利落。L9 Livis 还采用了迈巴赫同款涂装工艺，紫金、绿金、黑金等专属双色车身，额外增加两道工序后，漆膜厚度达到传统工艺的两到三倍，日光下呈现出流动的光影质感。外观的考究还藏在细节里。全车 12 颗超声波雷达被 UWB 传感器取代，保险杠上再也看不到突兀的小圆孔；隐藏式电动踏板收起后与底盘融为一体。这不仅让外观更整洁，UWB 技术还让哨兵模式耗电量骤降 82%，开两天只耗一度电。打开带有电容防夹功能的电动车门，里面还是那个熟悉的「起居室」，只是每一项体验的上限都被抬高了。全新 L9 是全球首款四个座位都配备全尺寸零重力座椅的 SUV。Livis 版本直接标配，Ultra 版本在首销期内也会免费赠送前排两把零重力座椅。在过去，转向管柱挤占了空间，主驾想完全躺平是个大难题。而在 L9 上，线控转向释放了空间，1.8 米的大个子坐进主驾也能一键舒展躺平。望向前方，仪表台上的两块屏幕合二为一，变成了一块 29 英寸的 6K 超宽全景屏。后排的 21 英寸屏幕新增电动平移功能，手轻轻一抬，这块「神奇移动屏」就会沿滑轨移到面前，并随座椅倾角自动找平。音响功率这次飙升到了 5440 瓦。9.3.6 星环剧院音响搭载三重矩阵式低音系统，头枕音响创新采用水平对置设计，出声口藏在两侧，以往那种「有人在你脑后说话」的突兀感也随之消失。车内空调支持 7 温区控制，面部和脚部可以吹出不同温度的风；全车 29 处皮质区域都带加热功能。中控台下方的 10L 智能冰箱换了电动门，轻拍两下即可开关，不仅支持 -6°C 冷冻，内胆还能直接拆下来清洗。作为增程车，全新 L9 两个版本的纯电体验都做到了行业顶配。底盘塞进了一块 72.7 度的 5C 大电池，容量已经超过不少纯电车型，换来 420 公里纯电续航。按每天通勤 40 公里算，两周充一次电完全够用。充电速度也同步提升，配合理想自建的 8300 根 5C 超充桩，十分钟左右就能完成补能——在服务区上个洗手间、买杯咖啡的功夫，电量就能满血复活。长途出行时，这台重达 2.8 吨的全尺寸 SUV 的油耗还挺实在的：亏电油耗降至 6.3 升/百公里，满油满电综合续航达到 1650 公里。更让人省心的是，这套增程系统的保养周期延长到了 3 年或 3 万公里，还省去了过去增程车繁琐的维护。李想表示，在机械属性和维保频次上，全新 L9 已经越来越像一台纯电车了。把家装进车里，把纯电融进增程「全尺寸 SUV 一直有一个很难真正被解决的矛盾，既让一家人坐得舒适，又让自己开得尽兴。」李想说。为了同时满足这两点，L9 Livis 直接上了全套线控主动底盘。L9 Livis 的 800V 主动悬架，让每个车轮都有超过 1 万牛的支撑力，传统的防倾杆被取消，四轮完全独立。过减速带、连续起伏路或者高速变线时，液压泵和减震器会实时配合，该撑就撑、该压就压。停车休息时，这套悬架甚至能联动赛车游戏，把车厢变成 4D 模拟器。大车想开得灵活，转向系统也要升级。Ultra 和 Livis 都标配了后轮转向 + 线控转向，转弯半径缩到 5.3 米，方向盘打满不用换手，角度也能按习惯自定义。另外，Livis 还首发了 EMB 线控机械制动，刹车响应快了约 0.1 秒，120km/h 时速下刹停距离能缩短 3 米以上。而且这套系统号称终生不用换刹车油，后期省心不少。安全性上，两款车都变得更加坚固了，堡垒安全车身 3.0 用上了 2200 兆帕的高强度钢和一体式双门环。最特别的是，车内的 11 个安全气囊还能通过红外传感器识别每一位乘员的体型与坐姿，动态调节保护策略。面对极端的碰撞危险，具备主动悬架的 L9 Livis 能在 0.7 秒内瞬间抬升 50 毫米，用最坚固的底盘部位去硬刚撞击；低速倒车时的防碰撞系统也具备了识别悬空物体的能力。智能化层面，理想首发了马赫 M100——一颗专为 AI 原生设计、采用数据流架构的端侧推理芯片。L9 Ultra 版本用的是单芯片方案，算力达到 1280TOPS，配备前向激光雷达；L9 Livis 则拥有两颗马赫 M100，总算力翻倍达到 2560TOPS，同时配备了 4 颗激光雷达来实现 360° 全向感知。算力上来后，感知帧率成倍提升，乡间小路、施工路段都能提前预判，果断绕行。这颗芯片还跟去年 12 月发布的 Livis AI 眼镜实现了生态互联。交互开始变得科幻。戴着眼镜站在车外，一个手势或一个眼神，车就会自动泊出来，缓缓开到你面前。坐进车里，系统还能察觉你的情绪和状态。你今天好像心情特别好，是要带家人去顺义的公园踏青吗？这时候，它既是司机，也是伙伴。到了我们最关心的价格，两款车只拉开了 5 万元的价差。L9 Ultra 定价为 45.98 万元，竞争力很强。72.7 度 5C 电池、双腔空悬、线控 + 后轮转向、1280 TOPS 马赫芯片都是标配，首销期还免费送前排零重力座椅、智能电开门和 5440 瓦音响。L9 Livis 定价来到 50.98 万元，比此前 55.98 万的预售价低了不少。多花 5 万，换来的是专属双色车身、800V 主动悬架、带 EMB 制动的全线控底盘、翻倍到 2560 TOPS 的算力，以及 4 颗激光雷达。首销权益也公布了：6 月 30 日前锁单，现金直减 2 万元，老用户再额外补贴 1 万元。从 2015 年创立至今，理想还是习惯把消费者叫作用户。买车只是陪伴的开始，让全家人在日复一日的相处里都能开心，才是 L9 Livis 这台「硅基生命体」想要兑现的承诺。

在今天和理想全新 L9 和智界 V9 两台重磅车型同台亮相的，还有蔚来的乐道 L80。相比 L90 面向三代同堂、多孩家庭的大三排需求，L80 把重点放在大五座市场，带来了更大的乘坐空间，更灵活的装载能力，以及更适合长途、自驾、露营等场景的空间组织方式。早年间，以本田冠道、大众途观 L 为代表的燃油大五座 SUV，主要通过释放第二排的腿部空间来提升乘坐舒适性；随后，随着新能源时代的到来，诸如理想 L7、问界 M7 等车型进一步强化了智能座舱的沉浸式体验和车内的储物能力。而乐道 L80 现在则试图在行业已有的基础上更进一步，将前备舱、后备舱、乘员舱以及补能体系整合为一个完整的生态闭环，通过蔚来的工程创新能力，让中大型五座 SUV 的日常使用体验再上一层楼。乐道 L80 一共有 Pro、Max+ 和 Ultra+三个版本，整车购买价格分别为 24.28万元、25.98 万元、27.98 万元， Baas 方案价格为 15.68 万元、17.38 万元和 19.38 万元。前备舱大，后备舱更大空间，依然是乐道 L80 最核心的竞争力。乐道 L80 车长 5145mm，轴距 3110mm，理想 L7 车长 5050mm，轴距 3005mm，乐道 L80 比理想 L7 大了一圈。取消第三排之后，乐道 L80 将所有空间都留给了前两排和储物。第二排作为 L80 整台车体验的核心区域，其后排头部空间达到 1074 毫米，腿部空间宽达 1121 毫米，坐垫宽度和长度分别达到 549 毫米和 497 毫米，靠背支持最大 135 度的后倾调节。L80 提出了一个「大五座，享四座」的越级体验概念，即当车内只有四人出行时，第二排的两位乘客完全可以获得媲美大型 MPV 独立航空座椅的放松状态。而在储物空间上，乐道 L80 配备了容积为 240 升的电动前备舱，能够轻松容纳两个 20 英寸的标准登机箱、一个儿童专属行李箱及一个随身背包。前备舱的存在不仅增加了绝对装载量，更为家庭用户提供了干湿分离与物理分区的储物区域，有效改善了传统 SUV 后备舱容易被杂物堆砌得杂乱无章的痛点。与前舱相对应，乐道 L80 的后备舱在第二排座椅处于常规直立状态时，容积为 1200 升；当第二排座椅完全放倒后，整体空间可扩展至 2600 升。此外，在后备舱的常规盖板下方，工程团队设计了双沉舱结构。其中外侧沉舱容积为 106 升，里侧沉舱容积为 63 升，均配备了独立盖板。这种隐蔽式的储物空间非常适合用来存放不常用的随车工具、零散的露营小配件、充电线缆以及各类清洁用品，使得主后备舱始终保持整洁。乐道能够在车辆内部极其有限的机械空间内挤出更多的储物容积，主要得益于车辆底层平台的极高集成度。L80 充分应用了已经在 L90 上被验证过的 900V 高压架构，并配备了体积更为紧凑的前置电驱动系统。结合 30.5 度大倾角冷却模块、全栈自研的热管理高度集成模块以及一体式偏置空调箱，这些系统级的技术创新对原本需要占据前舱大量体积的机械与电子部件进行了极致的压缩与整合。空间之外，座舱内的舒适化和智能化，也是乐道 L80 强调的特色。乐道 L80 的座舱内部采用了分层布局：上层主要用于视觉延伸与氛围营造，中层集中处理人机交互与触觉反馈，下层则侧重于日常储物与抗污耐用。车内配备了支持 1600 万色调节的氛围灯系统、长达 7.2 米的环绕光带，辅以大面积软质材料包覆、金属扬声器网罩及精细的绗缝与刺绣工艺。整体风格规避了盲目堆砌昂贵材料的浮夸感，更侧重于在家庭用车的基调上打磨出温馨且不失格调的精致质感。为了提升车内的通透感，乐道 L80 车辆侧窗与前挡风玻璃的总面积约为 3.3 平方米，前后分区的全景天幕面积也达到了 1.1 平方米。同时为了应对大面积玻璃材质在强光照射下带来的隔热与防晒问题，车内配备了双电动物理遮阳帘，乐道 L80 同时配备了空气悬架系统，支持在驻车时自动降低车身高度，配合电动侧踏板与静音电吸门，优化了行动不便的老人或儿童的上下车体验。车辆的座椅总成基于专属的高端座椅平台开发，依据中国人体生理曲线大数据进行了型面贴合度优化，前排座椅支持 12 向电动调节，第二排支持大角度电动调节。同时，方向盘加热、前后排座椅加热、通风以及按摩功能均作为标配提供。此外，车内集成了一台 8.8 升容量的智能电动冰箱，足以容纳 6 罐常规可乐、2 盒鲜牛奶以及部分水果零食。神玑上车，4D 领航数字化交互与计算平台是现代智能汽车的核心大脑。L80 的屏幕矩阵由一块 35 英寸的 AR-HUD 增强现实抬头显示、一块 17.2 英寸的 3K 高清中控屏、一块 17.3 英寸的后排吸顶娱乐屏以及一块 8 英寸的后排专属场景控制屏组成。支撑这一系列复杂交互与协同调度的硬件，是一颗采用 5 纳米制程工艺的车规级高阶智驾芯片——神玑 NX9031。该芯片拥有高达 546 GB/s 的内存带宽，打破了传统分散式域控制器的物理壁垒，能够同时支持智能驾驶、智能座舱、智能底盘等多个系统跨域计算资源的统一调度。配合 SkyOS·天枢整车全域操作系统，车辆从底层代码级别打通了智驾、座舱、底盘、车身及动力网络，使各项功能摆脱了信息孤岛的状态，能够基于场景进行深度系统级协同。在智能驾驶领域，系统深度接入了蔚来世界模型 NWM（NIO World Model）。该模型摒弃了传统的规则驱动逻辑，转而通过强化学习与海量的闭环仿真训练，提升了系统对复杂未知路况的理解与博弈能力，全面覆盖城区复杂道路、城市快速路、高速公路及各类园区停车场。值得重点关注的是，L80 支持城区领航换电功能（PSP）。在导航状态下，系统能够自动规划最优补能路径，将车辆自主驾驶至换电站并完成自动化换电流程。同时，基于模型化架构构建的智能泊车系统，也支持路线的实时动态规划和对各类非标准复杂车位的识别入位。尽管定位为体型庞大的家庭 SUV，L80 在底盘的机械素质与动态调校上依然投入了极高的规格。悬架系统采用了前双叉臂与后五连杆独立悬架组合，配备了具备 100 毫米行程调节能力的智能空气悬架系统以及 CDC 智能连续可变阻尼减震器。底盘硬件与路面之间通过液压衬套进行柔性连接，并搭配了抓地力更强的鸳鸯胎。乐道 L80 还支持 4D 全域舒适领航功能。该功能能够实时结合车辆自身的传感器感知、激光雷达对前方路面的三维预瞄以及云端高精舒适图层数据，在遇到减速带、坑洼或颠簸路面之前，提前毫秒级主动调节空气弹簧的刚度与减震器的阻尼，从容化解路面冲击。除了常规的运动、舒适、后排舒享和节能模式外，车辆预设了湿地、雪地、泥地、沙地等脱困模式。在卓越的底盘参数加持下，L80 拥有高达 600 毫米的最大涉水深度，最小离地间隙最高可拉升至 260 毫米，足以轻松应对绝大多数轻越野和长途自驾游的复杂路况。而在安全方面，乐道 L80 全系标配 9 个安全气囊，除了常规的环绕式防护，第一排还配备了能在碰撞发生前预紧的主动式安全带、容积达到 137 升的狐耳式超大副驾气囊以及防止前排乘员相互碰撞的中央远端安全气囊。基于家庭属性的考量，第二排乘客的安全防护被提升至与前排同等重要的层级，配备了 14 升容积的后排侧气囊，以及长度贯穿 2920 毫米、容积高达 56 升的超大侧面安全气帘。在应对碰撞的白车身结构设计上，L80 的前部预留了长达 650 毫米的充足溃缩吸能空间，并创新性地采用了「6 纵 10 横」的笼式安全传导结构。前防撞梁选用了高强度的 7 系航空级铝合金材质，其上层主防撞梁的宽度覆盖率达到了整车宽度的 80%，能够在严苛的偏置碰撞中有效化解破坏性冲击力。侧面防护则通过环舱高强度钢、一体式门环、多腔体铝合金门槛梁以及坚固的电池舱横梁共同构建。车尾部分应用了先进的一体压铸后地板技术，配合铝合金后吸能盒与后防撞梁，为车身尾部提供了深达 1244 毫米的吸能缓冲空间，最大程度减轻高速追尾事故对后排乘客及底层电池包的结构性伤害。乐道 L80 搭载的电机能够输出 340kW 的峰值功率，百公里加速时间 5.7 秒，其百公里综合电耗被控制在优秀的 14.3kWh，85 度电池下的 CLTC 续航为 615km。当然，纯电车型的体验永远无法脱离补能体系的建设。共享蔚来庞大且成熟的充换电网络，依然是乐道 L80 在同级别竞品中最难以被复制的核心优势。为了进一步夯实这一护城河，乐道官方已确认追加投入 6 亿元人民币，用于在网络中新增 8000 多块流通电池。蔚来之前透露，早在四年前，乐道 L80 便与 L90 基于同一技术平台同步启动开发。其中，乐道 L90 的成功已经验证了产品定义的精准性。在 2025 年，乐道 L90 凭借对大三排 SUV「满载状态下行李装载空间」这一核心痛点的有效解决，在上市短短 5 个月内实现交付超 4.3 万台，成为纯电大型 SUV 市场的年度销量冠军。这不仅助力乐道品牌在 2025 年实现 107808 台的全年交付目标，更以接近 24 万元的成交均价稳居新势力品牌第一梯队。乐道 L90 和蔚来 ES8 的表现，也被行业视为大型 SUV 动力形式从增程向纯电转型的关键标志。如果说乐道 L90 成功开拓了大三排纯电 SUV 市场，那么乐道 L80 则肩负着在规模更大的大五座 SUV 领域实现突破的重任。在乐道的产品价值体系中，L90 聚焦于「在一起」的温情与责任，通过三排布局满足多孩、三代同堂家庭的聚会与通勤需求；而 L80 则更侧重于「走出去」的激情与成长，以更灵活的空间形态承载用户自驾、露营等更具探索性的生活方式。

🔍郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%🤖OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略🚀Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI📱微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆💻华为 Mate X7 官方降价 1000 元⚠️追觅造车主体约 232 万元股权遭冻结🎮国行 Switch 正式停服🧠百度成立模型委员会 BMC💰三星电子单季营收创历史纪录，半导体撑起超九成营业利润🔍海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%🖨️铠侠：2027 年内存仍吃紧、净利润同比增长 103.6%🤖微信发布「全球青少年 AI+ 小程序」洞察报告💡何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了🦿5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约📱荣耀 CEO 李健官宣机器人手机今年 Q3 发布📱vivo S60 将首发 4K 原生感 Live🚗全新理想 L9 发布：45.98 万元起🤖ChatGPT 新增 Codex 连接电脑功能📰 周末也值得一看的新闻郭明錤：苹果扶持 Intel 制衡台积电，良率目标 50% 至 60%天风国际证券分析师郭明錤昨日披露，苹果已在 Intel 18A-P 系列制程上为低阶及旧款 iPhone、iPad 与 Mac 处理器立项，订单中 iPhone 约占 80%。流片规划显示，苹果预计 2027 年在 Intel 放量，2028 年续增，2029 年进入衰退期。Intel 方面，2027 年良率目标为稳定达到 50% 至 60% 以上，量产时程与出货规模目前仍未明朗。不过，郭明錤也指出，即便 Intel 顺利出货，台积电在苹果供应体系中仍将占 90% 以上比重。郭明錤认为，苹果此举是有意将 Intel 培养为长期备选供应商。AI 对台积电先进制程资源的占用将持续扩大，苹果需要在议价能力尚存时提前布局。与此同时，苹果也在同步评估 Intel 其他先进制程技术。OpenAI 整合 ChatGPT 与 Codex，Greg Brockman 正式接管产品战略据 WIRED 报道，OpenAI 昨日向员工宣布新一轮组织调整，联合创始人兼总裁 Greg Brockman 正式接管产品战略。报道指出，OpenAI 将 ChatGPT、编程智能体 Codex 及面向开发者的 API 业务整合进同一核心产品团队，目标是将三者合并为「统一的体验」。Codex 原负责人 Thibault Sottiaux 升任核心产品与平台主管，统管消费、企业和开发者三条线；ChatGPT 原负责人 Nick Turley 转任企业产品改版；前 Instagram 副总裁 Ashley Alexander 接任消费产品负责人。Anthropic 融资 300 亿美元，估值升至 9000 亿美元，超越 OpenAI据《金融时报》报道，Anthropic 已就一轮 300 亿美元的融资达成条款，融资完成后公司估值将约达 9000 亿美元，较今年 2 月初 3500 亿美元的估值接近翻三倍。本轮融资由 Dragoneer Investment Group、Greenoaks Capital、红杉资本（Sequoia Capital）和 Altimeter Capital 联合领投，报道称，每家领投机构预计出资不低于 20 亿美元。微信捐赠第一行代码和测试用机，入藏 CCF 计算机博物馆CCF 计算机博物馆（CCM）昨日收到腾讯微信团队捐赠的一批早期历史实物，包括微信第一行后台代码原始存档、早期开发设备 iMac ME086CH/21 寸一体机与戴尔塔式工作站，以及 Apple iPod Touch 3、诺基亚 5800XM 等早期测试终端。CCM 表示，由于移动产品迭代速度极快、数字资料易被覆盖删除，大量见证移动互联网起步的实物未能得到系统保护。此次捐赠呈现了微信从 0 到 1 的研发链条，是中国移动互联网早期发展的核心实物见证。CCM 主体建筑已全面落成，拟于今年 9 月开馆，目前持续面向社会征集各时期计算机相关藏品。保时捷称被中国车企抄袭「有些许无奈」，罗永浩：保时捷自己就有前科保时捷中国传媒公关副总裁唐风靓日前在浙江「易境思论坛」上表示，过去几年保时捷激发了高度统一的汽车设计审美，被中国车企模仿「除了荣幸也有些许无奈」。罗永浩随后在微博发文回应，称保时捷说「无奈」可以理解，但称中国车企抄袭得「高度统一」则是吹牛。他同时指出，保时捷历史上也曾抄袭捷克斯洛伐克品牌太脱拉（Tatra）的设计，并以 1965 年大众汽车向太脱拉支付 100 万西德马克庭外和解赔偿金为据，直言「说到流氓程度，谁能跟保时捷比」。此番言论随即引发讨论。有网友指其犯了「诉诸伪善」的逻辑错误，罗永浩强硬回怼，称自己已明确承认抄袭者「也是流氓」，并非为抄袭行为开脱。华为 Mate X7 官方降价 1000 元华为昨日宣布，旗下两款折叠屏手机同步开启降价优惠：Mate X7 全系直降 1000 元，优惠后售价 11999 元起；Mate X6 直降 3000 元，优惠后售价 9999 元起。追觅造车主体约 232 万元股权遭冻结企查查 App 显示，追觅科技创始人俞浩实控的星空计划（上海）汽车科技有限责任公司近日新增一则股权冻结信息，股权数额约为 232.25 万元，股权所在企业为上海星空计划进出口贸易有限责任公司。去年 8 月，追觅科技宣布进入汽车领域。截至目前，追觅汽车团队规模约千人，研发人员占比达 70%，自研覆盖智能汽车多个核心领域。此次造车主体出现股权冻结，官方尚未就相关情况作出回应。国行 Switch 正式停服国行 Switch 的 Nintendo e 商店已于昨晚正式停服，至此，国行 Switch 将无法再进行联网对战或下载游戏。官方在停服前向相关玩家发送提醒，建议玩家在停服前完成所有已购游戏、DLC 及兑换码的下载与激活；删除游戏数据或对主机进行格式化后，数字内容将无法恢复。今年 3 月 31 日，国行 Switch 的游戏售卖服务已率先停止；而根据官方此前公告，今年 12 月 31 日起，国行 Switch 主机及配件的维修服务也将全面终止。这意味着，国行 Switch 的官方售后支持将在今年底彻底画上句号。百度成立模型委员会 BMC据《科创板日报》报道，百度近日内部设立百度模型委员会（Baidu Model Committee，BMC），成员由对大模型有深刻理解的年轻研究员构成。基础模型研发部（BMU）与应用模型研发部（AMU）均向该委员会汇报。知情人士透露，新成立的 BMC 将全面统筹百度大模型业务，覆盖路径探索、技术规划、模型训练到业务落地的完整链条，目标是实现模型从技术到应用的一体化管理。百度方面表示，此举有助于集中优势资源，持续建设并强化人工智能方向的技术优势。三星电子单季营收创历史纪录，半导体撑起超九成营业利润三星电子昨日发布 2026 财年第一季度（1 月 1 日至 3 月 31 日）业绩报告：期内实现合并营收 133.9 万亿韩元，同比增长 69%，环比增长 43%；营业利润 57.2 万亿韩元，同比增长 756%，环比增长 185%，营业利润率达 42.8%。净利润为 47.2 万亿韩元，净利率为 35.3%。基本每股收益为 7123 韩元。半导体（DS）：营收 81.7 万亿韩元，同比增长 225%，环比增长 86%；营业利润 53.7 万亿韩元。其中存储业务营收 74.8 万亿韩元，同比增长 292%，环比增长 101%，为本季度核心增长引擎；设备体验（DX，含手机与消费电子）：营收 52.7 万亿韩元，同比增长 2%；营业利润 3.0 万亿韩元，同比下降约 1.8 万亿韩元；显示（SDC）：营收 6.7 万亿韩元，同比增长 14%；营业利润 0.4 万亿韩元，同比下降 0.1 万亿韩元；Harman：营收 3.8 万亿韩元，同比增长 12%；营业利润 0.2 万亿韩元。下季度展望，公司预计 AI 基础设施扩张驱动的服务器内存及 SSD 需求将持续强劲，计划向客户交付 HBM4E 首批样品以巩固技术领先地位；手机业务方面，Galaxy S26 系列发布效应趋于消退，盈利能力预计承压。海力士今年一季度营收 52.58 万亿韩元，营业利润同比暴涨 405%昨天，海力士于发布今年第一季度（截至 3 月 31 日）业绩，按 K-IFRS 口径，期内实现营收 52.58 万亿韩元，同比增长 198%；营业利润 37.61 万亿韩元，同比增长 405%；净利润 40.35 万亿韩元，同比增长 398%。本季度营收首次单季突破 50 万亿韩元，创历史新高。主要财务指标如下：毛利润 41.68 万亿韩元，毛利率 79%，同比提升 22 个百分点；营业利润率 72%，创历史单季新高，环比提升 14 个百分点；EBITDA 41.34 万亿韩元，EBITDA 利润率 79%，同比增长 284%；净利润率 77%，其中非经营利润 14.01 万亿韩元，包含投资资产估值收益 9.94 万亿韩元。铠侠：2027 年内存仍吃紧、净利润净同比增长 103.6%铠侠控股昨日发布截至今年 3 月 31 日止 2026 财年业绩，期内实现营收 2.3376 万亿日元，同比增长 37.0%；按 IFRS 口径，归母净利润为 5544.9 亿日元，同比增长 103.6%。SSD 与存储业务营收 1.3626 万亿日元，同比增加 37.5% 至 3715 亿日元；智能设备业务营收 7599.8 亿日元，同比增加 51.6% 至 2588 亿日元；其他业务（含 SD 卡、U 盘等零售产品及向闪迪集团的销售）营收 2150.1 亿日元，同比基本持平。另据财联社报道，铠侠表示，AI 服务器客户的订单量持续走强，以美元计价的 NAND 价格在一季度上涨逾一倍，并预计 2027 年度 NAND 存储供应仍将维持紧张态势。微信发布「全球青少年 AI+ 小程序」洞察报告昨天，腾讯研究院联合微信发布《全球青少年 AI+小程序洞察报告》，披露了全球基础教育阶段群体利用 AI 工具与小程序进行应用开发的各项核心数据。据报告数据显示，自 2023 年 11 月微信小程序教育平台正式对外开放以来，已吸引全球 7954 所学校的逾 8.7 万名学生参与，累计创建小程序项目超 28.7 万个。在开发流程中，师生调用 AI 辅助创作每年消耗 Token 量超过 500 亿，支撑约 375 万次人机交互。此外，微信连续四年举办的小程序全球创新挑战赛共产生 7430 个作品，其中第四届单届提交量达 4900 个，呈显著爆发趋势。开发者画像层面，62% 的项目由单人独立完成，最小获奖开发者年仅 5 岁。地域活跃度数据显示，中国香港、北京及中国澳门位列前三，且项目协作已出现跨国别开发特征。在应用方向上，05 后开发者主要聚焦于无障碍辅助、碳足迹、适老化关怀及健康管理等社会实用场景。💡 何小鹏：当车展上不再有 150 款新车时，中国汽车内卷可能就结束了据睿见 Economy 报道，小鹏集团董事长、CEO 何小鹏昨日在「第十八届轩辕汽车蓝皮书论坛」上表示，「什么时候我们看到北京、上海车展上不再有 150 款新车的时候，中国汽车的内卷可能就结束了。」何小鹏表示，过去十年，中国新能源汽车经历了两个重要转变：一是核心驱动能源从传统化石能源切换至电化学能源；二是行业全面进入初步普及的智能化阶段。他称，十年前几乎没有人相信中国汽车能在全球市场取得如今的成绩，并实现规模化的全球化布局。何小鹏认为，下一个十年将是电化学能源与数据能源的双重叠加阶段，芯片和模型将成为新的「引擎」，汽车与机器人的能力将在这十年内深度整合。5999 元，ROG XREAL R1 电竞 AI 眼镜开启预约ROG 与 XREAL 联名推出的电竞 AR 眼

🤖曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini🌐马耳他将向全体公民免费提供一年 ChatGPT Plus📱小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务🚀目标估值 1.75 万亿美元，SpaceX 最快 6 月上市📋arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚💻卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元🚗蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌🔍需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位☁️最低 1 元 40 万 Token，三大运营商推出 Token 套餐🚙特斯拉 Robotaxi 碰撞事故报告解封💡知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值🚗小米 YU7 GT 火山灰配色公布💰ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构🧠马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布🤖字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本📱腾讯 vibe coding 产品「吐司」即将上线🛍️蜜雪冰城上架老黄同款蜜桃四季春，销量暴涨🛑中国银行关停信用卡独立 App，服务整合进主 App曝苹果独立 Siri 应用支持对话自动删除、底层模型大量替换为 Gemini昨晚，彭博社记者 Mark Gurman 在最新的 Power On 新闻通讯中透露了更多关于苹果即将发布的独立 Siri 应用的细节。隐私设计是此次新应用的核心卖点之一。Gurman 透露，该应用将内置对话自动删除功能，用户可在设置中选择对话保留时长：30 天、1 年或永久保存。苹果预计在今年的 WWDC 继续主打隐私保护，并强调其 AI 体验不含广告。模型方面，苹果已将新 Siri 的 AI 模型底层大量替换为 Google 的 Gemini，并将部分服务运行在 Google 的云基础设施之上。而此前有消息称，苹果和原本的合作对象 OpenAI 关系出现裂痕，甚至有可能对簿公堂，原因是 OpenAI 认为双方合作的结果不及预期。值得注意的是，尽管新的 AI Siri 已经跳票约两年，Gurman 称苹果内部测试版本仍将其标注为「Beta」，并可能沿用至其正式发布。马耳他将向全体公民免费提供一年 ChatGPT PlusOpenAI 日前宣布，与马耳他政府签署合作协议，向该国公民及居民免费提供为期一年的 ChatGPT Plus 订阅服务。据悉，居民须先完成由马耳他大学开发的免费在线课程「AI for All」，课程内容涵盖 AI 的基本概念、能力边界及负责任使用方式，通过后即可向马耳他数字创新局申请免费订阅资格。小米 REDMI 产品经理：不打算换手机可以再等等，更多机型将支持电池升级服务小米 REDMI 产品经理「笋寸」日前在微博发文建议，近期不打算换手机的用户可以考虑等待换电池，并提示部分机型在更换新电池后容量可能还会有所提升，参考小米 13 系列的现行政策。笋寸进一步说明，小米正积极推动存量用户更换采用金沙江技术的大容量电池，小米 13 系列已率先上线相关服务。她还透露，具体支持机型「得等等」。此前，小米 13 系列已全系上线电池升级服务，支持小米 13、小米 13 Pro、小米 13 Ultra 三款机型升级更大容量电池，费用为 189 元，其中电池升级费 149 元、人工费 40 元。目标估值 1.75 万亿美元，SpaceX 最快 6 月上市据路透社报道，SpaceX 已加快 IPO 时间表，计划最早于 6 月 12 日在纳斯达克挂牌上市，股票代码为「SPCX」，较此前预定的 6 月下旬提前约两周。据三位知情人士透露，公司计划最早于本周三公开招股说明书，路演定于 6 月 4 日启动，股票发行最早于 6 月 11 日完成定价。时间表提前的主要原因是 SEC 对招股材料的审核进展快于预期。此次 IPO 预计募资约 750 亿美元，目标估值约 1.75 万亿美元，有望成为史上规模最大的股票发行。承销行阵容包括摩根士丹利、美国银行、花旗集团、摩根大通和高盛。arXiv 出台新规禁止 AI 垃圾，违者封禁一年加强制同行评审双重处罚近日，预印本学术论文平台 arXiv 宣布，若作者提交的论文中包含无可辩驳的 AI 生成内容证据，将被封禁账号一年，并在此后要求所有提交须先经过同行评审期刊接受方可上传。arXiv 计算机科学板块主席 Thomas Dietterich 在 X 发文表示，若论文中出现「幻觉引用」、大语言模型元注释（如「以下是 200 字摘要，是否需要修改？」）等明显迹象，即可认定作者未对 AI 生成内容进行核查。arXiv 目前由康奈尔科技学院管理，今年 7 月将独立运营，成为一家非营利法人机构。康奈尔科技学院院长 Greg Morrisett 表示，独立后 arXiv 将能从更广泛的捐助方获得资金支持，以应对「AI 垃圾内容」带来的运营压力。卢伟冰：玄戒芯片今年肯定迭代、下半年部分国产旗舰直板机价格或破万元近日，小米集团合伙人、总裁卢伟冰在「Max 一夏」户外露营直播中表示，玄戒芯片今年肯定会推出迭代款，并称这将是「一块非常强的芯片」，会搭载在一款「非常优秀的产品」上。他同时提醒外界，目前网络上流传的各类相关传闻可信度不高，具体规格暂不便披露。卢伟冰还透露新一代玄戒芯片的应用范围将不再局限于智能手机，后续计划广泛搭载于小米旗下多类智能终端设备。关于手机定价，卢伟冰在直播中作出判断，认为今年下半年尤其到年底，部分国产旗舰直板手机价格可能突破 1 万元人民币。他预计，内存涨价的压力至少延续至 2027 年底，甚至到 2028 年。对于即将发布的小米 17 Max，他表示定价仍在讨论中，将在发布会上正式公布，但小米的原则是在同档位提供「最物超所值的产品」。蔚来李斌：萤火虫就像 iPhone，已超越汽车工具属性成为时尚品牌据睿见 Economy 报道，在日前举行的第十八届轩辕汽车蓝皮书论坛上，蔚来创始人、董事长兼 CEO 李斌线上出席，在对话中将萤火虫比作 iPhone，称其已成为「一款非常高品质的产品」。他强调，萤火虫在某种意义上超越了汽车的工具属性，演变为「一个生活方式品牌、一个时尚品牌」，并开始形成独特的品牌文化。他表示，萤火虫在用户群体上打破了年龄与购车预算的双重限制，「大学刚毕业的年轻人想买也能买，很有钱的老板想自己开车送小孩上学也完全没有问题」。需求一年暴增逾 700%，前向部署工程师成 AI 时代最抢手岗位据《商业内幕》报道，前向部署工程师（FDE）的招聘需求在一年内激增逾 700%，正成为 AI 浪潮下科技行业增长最快的职位之一。招聘平台 Indeed 数据显示，今年 4 月该职位招聘帖数量达 5330 条，较去年同期的 643 条增长约 729%。Anthropic、OpenAI、Palantir、Stripe 及 Google Cloud 均在积极招募该职位人才。FDE 的核心职责是协助企业将 AI 整合进业务流程，并根据客户需求定制开发 AI 工具。薪酬方面，Indeed 数据显示该职位年薪约为 17 万至逾 20 万美元。最低 1 元 40 万 Token，三大运营商推出 Token 套餐中国电信昨日推出系列试商用 Token 套餐，分别面向开发者及中小微企业、个人及家庭用户提供「Token+连接+安全」一体化服务，均提供三档套餐，面向个人及家庭用户定价 9.9 元/月起 1000 万 Token；面向开发者及中小微企业 39.9 元/月，1500 万 Token。两类用户均可选购宽带上行提速包（上行速率提升 50 Mbps，100 元/月）及智能体安全管家服务（5 元/月）。个人家庭用户还可额外选购 5G-A 优享包（含 10 GB 流量，上行速率最高 300 Mbps，20 元/月）。联通、移动两大运营商也先后推出了各自的 Token 套餐：移动：中国移动上海公司于 5 月 17 日世界电信日发布会上推出 Token 通用服务，定价 1 元 40 万 Token。联通：中国联通上海分公司于 5 月 16 日发布全栈国产化「国芯国模国云」安全基座，推出多档 Token 产品及融合套餐。特斯拉 Robotaxi 碰撞事故报告解封近期，特斯拉正式解密并公开了向美国国家公路交通安全管理局（NHTSA）提交的 17 起「Robotaxi」自动驾驶碰撞事故的完整未脱敏报告。据 Electrek 报道，这 17 起事故发生于 2025 年 7 月至 2026 年 3 月，均涉及在得克萨斯州奥斯汀进行「Robotaxi」测试的 2026 款 Model Y，且测试时均有安全员在场。伤亡数据显示事故整体较轻微：13 起仅造成财产损失，2 起无人员受伤，1 起轻伤未入院，另有 1 起轻伤需入院治疗（系车辆在路口以 2 英里/小时蠕行时被后方 SUV 追尾）。官方报告表明，大部分事故并非特斯拉自动驾驶系统（ADS）的过失，多为车辆在红绿灯、停止标志或交通拥堵中完全静止时，被后方人类驾驶员追尾或侧擦。但数据同时也暴露出系统与后台干预的缺陷：两起事故发生在远程操作员接管期间，导致车辆分别以 8 英里和 9 英里的时速撞上金属围栏与施工路障；另有数起事故显示 ADS 存在空间感知死角，车辆在转弯或倒车时撞上了金属链条、拖车挂钩、木制电线杆及路缘石。💡 知乎 CEO 周源：AI 时代，每一个真实的人都拥有不可替代的价值据「听筒Tech」报道，知乎创始人兼 CEO 周源日前在第十二届新知青年大会上表示，AI 时代真实创作者的价值不可替代。AI 每分钟都在创造奇观，但 AI 时代仍然需要新知。而新知来自真实的人。每一个真实的人，都拥有不可替代的价值。大会主论坛上，多位嘉宾就 AI 与人类创作的关系展开讨论。复旦大学人文学者梁永安提出「科技有所为，如何让自己有所不为？需要人的尺度和人文属性」；作家张春称「不完美就是我们的完美」。周源在大会结尾总结道：「技术越发达，有温度、有深度的真人创作就越宝贵。」小米 YU7 GT 火山灰配色公布近期，小米汽车再度放出 YU7 GT 的预热图片，公布了新配色以及新车的内饰信息。其中小米 YU7 系列将加推一款名为「火山灰」的配色，官方介绍：灵感来自晨雾下的火山地貌，高纯度灰调，叠加纳米级矿物颗粒，呈现细腻而有层次的金属质感。同时新车内饰也迎来全新的「赛道红」：采用红黑撞色，提供专属刺绣。官方表示，小米 YU7 也会提供「火山灰」外饰配色，并且现在车厘子红的 YU7 GT 已经在部分小米汽车门店展示了。ChatGPT 能帮你管钱了，支持连接超 1.2 万家金融机构近日，OpenAI 面向美国 ChatGPT Pro 用户推出了个人财务管理功能，支持通过 Plaid 连接超过 1.2 万家金融机构的银行账户和信用卡。连接账户后，ChatGPT 会自动读取用户的真实余额、消费流水和负债数据，生成一张覆盖投资、日常开销、订阅和近期账单的财务面板。用户提问时，模型可直接翻出账单历史，找出哪些类目容易超支，并给出具体的削减方案。此外，用户还可以告诉 ChatGPT 一些账外信息，比如「我在攒钱买车」或「我还欠父母一笔钱」，系统会将这些内容保存为财务记忆，用于后续对话。马斯克：1.5T 参数「王炸」模型 Grok V9 三到四周内发布马斯克昨日在 X 发帖披露，SpaceXAI 正在研发的 1.5T 参数 Grok 基础模型第 9 版（V9）已完成训练，预计 3 到 4 周内正式发布。现役公开版本 Grok 4.3 基于 0.5T 参数的 V8 基础模型，目前每隔数日迭代一次。V9 将参数量提升至 1.5T，马斯克称其为「重大升级」。在帖子的结尾，马斯克更是用用「banger」（王炸）一词来形容对这一版本的预期。字节开源 Cola DLM：不再从左到右生成，扩散模型也能写文本字节跳动 Seed 团队近日开源扩散

新玩意 241｜少数派的编辑们最近买了啥？主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。05/14 17:17编注：很多读者都会好奇少数派的编辑们到底平时都「买了啥」。我们希望通过「编辑部的新玩意」介绍编辑部成员们最近在用的新奇产品，让他们自己来谈谈这些新玩意的使用体验究竟如何。内容声明：《新玩意》栏目如含有商务内容，将会在对应条目标注「广告」；若有标注「厂商送测」，即产品虽非自费购买，但内容无品牌方干预。编辑们的分享仅作参考，欢迎大家在评论区分享同款产品的使用体验，无论正面与否，我们都欢迎坦言。发言请遵守少数派社区规则，文明表达，开心聊天。@张奕源 Nick棉小妞家纯棉 T 恤参考价格：¥18朋友们，量大管饱的 T 恤被我找到了。前段时间我淘汰了一大批穿了好几年或者已经变形的 T 恤，自然需要补货一番，搜索中就找到了这家店。我原本是把他们的 T 恤当成「穿两次就扔」的消耗品来买的，没想到质量相当不错，大大超过了我的预期，所以一定要在《新玩意儿》怒推一下。他们家 T 恤的板型是大众款，320g 的版本就已经很厚实，穿起来也很松软。我买了克莱因蓝、森林绿、藏青和灰色等几个颜色，每个的颜色都很正，不廉价。更难得的是，作为高克数的纯棉衣服，这款 T 恤在洗过之后没有明显的缩水，也不掉色，在品质上已经无可挑剔。我一开始还打算试试他们家的 400g 版本，但穿过之后觉得已经没有必要——320g 版就已经做到了足够绵软且耐穿。灰色长这样这种白板 T 恤还有个好处，就是款型不分男女。我平时偏好穿得宽大一点，所以买的尺码比店家推荐的大了一号。考虑到 T 恤价格很便宜，所以即使不出门穿，买几件大号的放在家里当睡服其实也不错。这家店的 T 恤的颜色都很活泼鲜明，一眼看去不觉得土，感觉是做过市场调研、熟悉年轻人喜好的。店里的主打款就是这个白板 T 恤，还兼顾着卖一些卫衣、连帽衫什么的，看起来确实是个工厂直营店。可惜他们不卖运动长裤，不然我肯定还会再囤三五条，比各个渠道乱买靠谱多了。价格便宜还有个好处，就是如果要出门旅行可以多买几件不同颜色的，每天换个色，拍照留念应该会很有意思。考虑到咱国服装产业的发展程度，这类店在电商平台上应该还有很多很多，我贴个链接方便你直达，你也可以根据自己的喜好拓展搜索。总之，从实用主义的角度出发，这种便宜大碗的白牌 T 恤已经相当好穿，很适合怒囤几件，迎接夏天。拓竹 X2D 3D 打印机参考价格：¥5099.15（多色打印套装）好家伙，拓竹的 P2S 才刚发了没多久，X2D 就来了。我虽然短暂入坑 3D 打印的这几个月完全是懒人型选手，只打别人做好的模型，几乎不搞任何调参，但还是第一时间把 X2D 请进了家门。X2D 的大体外观和性能和 P2S 没有太大区别，甚至很多零件都能互换，所以我就不再重述基础的打印体验了，感兴趣的朋友可以参考之前在《新玩意儿》里写过的 P2S：https://sspai.com/post/105458相较于 P2S，X2D 主要多了以下东西。首先，它的打印喷嘴有两个，一路是主喷嘴，另一路则为辅助。为了减少打印头的体积和重量，打印头只为主喷嘴配备了挤出机，所以辅助喷嘴的挤出机就变成了个单独的外部模块。在这样的「远挤近出」结构下，两个喷嘴的打印精度略有不同，主喷嘴的挤出机配备的是伺服电机，辅助喷嘴则为步进电机，所以主喷嘴适合的材料类型更丰富，细节精度控制更强，打印效率也更高。X2D 背后的挤出机其次，X2D 增加了主动腔温控制系统，使它可以打印更多的工程材料。机身内部多了双进气口，腔内温度的升降都会比 P2S 更迅捷和精准。不仅如此，X2D 还随机附赠了一个外挂的空气净化模块，这也是为了配合腔温控制系统的升级，让高温打印时的异味更少。外挂的空气净化模块，P2S 应该也能单独购买所以，你可以认为 X2D 就是 P2S 的 Pro 版本。它依然可以满足日常打印的各种需求，同时还允许你玩一玩工业级材料，搞一些产品创造。从实际使用的角度出发，我的主要感受有两个。一是双头打印简化了支撑打印，譬如我的主喷嘴连接了 AMS，只打 PLA，辅助喷嘴则专门打 PETG，这样在打出的支撑和主体分离度很高，拆除的时候几乎不需要用工具慢慢抠，省时省力；二是主动腔温控制系统提升了打印质量，我曾经在 P2S 上打印某些模型遇到过特定角度翘边的问题，但在 X2D 上则完全没有。拓竹推荐的场景之一就是 PLA 加 PETG 混打，支撑的部分很好拆除很细致了！总体来说，如果你想购买一台适合家庭场景的 3D 打印机，拓竹 X2D 已经算是顶配之选。对于我这种只想打成品解决生活问题的用户来说，开箱即用、保质保量一直是拓竹的优势；对于想要认真研究 3D 打印，甚至用来搞项目、开农场的用户，X2D 的性能也足够完成大部分要求，满足你的创造欲。哦对，拓竹最近还上了个新的 PLA Pure 材料，主打食品级安全，可以打餐具、儿童玩具等。这是我期待已久的一款材料，我早就想给我家猫打食盆，新材料我也已经收到了但还没开始用，之后如果有什么感受再写反馈吧。@怡宝：OASEAR 女巫Pro HiFi耳机参考价格：¥522一款温暖向的二次元 iem我本着听个响，一定要选个颜值高的耳塞通勤使用，入手了这副女巫 Pro，其实第一眼看中的是他家粉色外观的桃夭，纠结一番还是想试试 「女巫」的升级款外观「女巫Pro」 是暗紫色的半透明树脂腔体，手感上细腻平滑。佩戴舒适度和轻量化设计给到好评，我个人偏好一体式的入耳设计，它两小时连续佩戴没有胀耳感，耳套上选择了附赠的最小号使用，也没有挤压到我的 flat 耳位，对小耳朵人群友好。试听表现作为人声为主的 ACG 塞，在试听上选择了 Aimer 标志性的《茜さす》，Aimer 略带沙哑的嗓音，在这副耳机偏暖的调音下，“哭腔”的悲怆感毫无保留地传递出来，歌词里那份无法言说的孤独，倒是有别样的温暖和治愈感，当弦乐进入时，听感上受解析限制的弱点出现：声部之间的分离度并不极致，但换来的是整体的包围感。接着试听的是 Yoasobi 的《夜に駆ける》，「女巫Pro」的优势在于小编制和 vocal 为主的曲目。在分离度就显得有点吃力，贝斯和合成器会轻微打架。高频延展还不错，小提琴泛音有空气感却不刺耳，重要的是人声没有被淹。对比而言，「女巫 Pro」更适合选择独唱、空灵向的 ACG 曲目，不太适合特别吵闹的电音系 ACG。作为杂食党，但偏好 soul / R&B，另外试听了火星哥 Bruno Major《The Show Must Go On》，人声和伴奏中和表现不错，鼓声轻柔，有耳边吟唱的贴耳感，能契合歌曲的直面遗憾的意境。其他「女巫Pro」抗阻在 20Ω，可以直推，但音量太大，需要调整为平时 1/2。我日常搭配的是水月雨的「破晓 Dawn pro」小尾巴挂手机使用，支持 3.5mm / 4.4mm 双端双解码，200+的价格入手性价比很高。小结「女巫Pro」属于一圈＋一铁的配置，虽然高频不够明亮，好在下潜够用不轰头，它的舒适区在中频，中频温润细腻，听情歌时情感表达浓郁，更适合杂食。总的来说这是一副温暖有氛围感，又充满空气感的「asmr」的耳塞。@巴斯：迪卡侬 TILT 120 折叠自行车参考价格：¥1299最近入手了迪卡侬的 TILT 120 折叠自行车，陆陆续续骑行了几个周末，我想说：它可能不是最完美的，但绝对是目前折叠自行车千元价位里，性价比最高、最适合在城市里穿梭的选择。从下单到送到家，只花了一天半的时间，整车是大纸箱包裹，几乎不需要要任何改动，只需要按步骤展开，锁住它的“关节”，就可以出门骑车了，真的做到了「即买即骑」。对于我这种入门级别的骑行爱好者，讨厌繁琐的组装，也讨厌因为这个步骤，消磨了我的耐心，出门也会延时，但它真的上手很快，这点也是我首选它的原因。采用经典对折式折叠，操作非常流畅：松开车架中间的铰链放倒车把合拢扣紧磁吸扣出发后面我尝试了好几次展开，确实能做到官方说的 30 秒内轻松快开。折叠后的它可以轻松放进后备箱、咖啡馆的角落。不管是跨城短途骑行，还是下午茶出门，都不用担心车被盗，在家停放也不太占地方，收纳省心。再来说说我的骑行体验，别看它外观轻巧，整车重达 14.5 kg，高碳钢车架带来的足够重量，让骑行更稳，同时有一定的减震效果，在过减速带的时候不会有明显的颠簸感。560 毫米的车把宽度，对于我的手臂也能有很好的伸展，虽然说 TILT 120 的价位在折叠自行车里是入门级，但它承载了 6 级变速，别看档位不多，但在城市中的路况也完全够用，我一般选择用 4 档，基本是平地骑行，偶尔上坡切 2 档，下坡切 6 档，骑行总体表现稳定，给我足够的安全感。目前骑行下来整体上都很满意，我的需求是周末闲骑，但它适配的场景非常非常多，如果你厌倦了共享单车的繁琐和收费，也受够地铁拥挤的麻木，不妨可以试试这种健康又有操控感的出行方式，沿途街景，可能会悄悄治愈你一整天。如果你也想分享「新玩意」🔉：获取 Matrix 社区写作权限并签署Matrix 共创计划；新发布一篇文章，在标题中标注「新玩意」前缀；用至少 800 字介绍产品，并配上 2-3 张产品的实拍图片；在网站个人信息中补充支付宝账号。成功入选还可以得到 108 元的「剁手红包」🧧，并在每周二的社区速递栏目中展示。如果你有兴趣参与，就赶紧来稿吧！> 下载少数派客户端、关注少数派公众号，了解更多的新玩意 🆒> 特惠、好用的硬件产品，尽在少数派 sspai 官方店铺🛒2141扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@Microhoo© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#热门文章#编辑部的新玩意21等 21 人为本文章充电扫码分享举报本文章少数派编辑部少数派编辑部官方账号。关注

周五15派早报：戴尔推出 Alienware 15 入门游戏本，Spotify 将支持 HLS 视频播客技术等05/15 08:25早报速览戴尔推出 Alienware 15 入门游戏本14Spotify 将支持 Apple HLS 视频播客技术0美政府批准向 10 家中国公司出售英伟达 H200 芯片11微软消息四则1NGINX 潜伏 18 年重大漏洞曝光2三部门部署持续加强新能源汽车安全管理0不妨一看的简讯0戴尔推出 Alienware 15 入门游戏本戴尔于 5 月 13 日正式推出 入门级游戏笔记本电脑 Alienware 15。这一产品采用老款处理器和低定位显卡，拥有 110W 的总性能功耗 (TPP)。Alienware 15 采用双风扇三热管后排气散热系统，处理器可选 AMD 锐龙 7 260 / 锐龙 5 220 或英特尔酷睿 7 240H / 酷睿 5 210H，独立显卡方面则是 NVIDIA GeForce RTX 5060 / 5050 / 4050 笔记本电脑 GPU。其搭载 15.3" 1920×1200 165Hz 的 62.5% sRGB 低色域屏幕，内置 HD 摄像头、双阵列麦克风、双 2W 扬声器，支持 180° 开合，接口包括 2 个 USB-C、2 个 USB-A、1 个 RJ45、1 个 HDMI、1 个 3.5mm 音频插孔。Alienware 15 的 AMD 版本起售价为 1299 美元，英特尔版本起售价为 1349 美元。来源还是要一万多啊62.5% sRGB，买来做什么？ 家里有一部i7-2720QM的M17 R3，十五年了，至今仍然在每天开机服役。既不入门也不游戏1080p低色域都来了……AW 冲低，游匣换身衣服冲高全部讨论 1461Spotify 将支持 Apple HLS 视频播客技术Spotify 宣布将在 Spotify for Creators 与 Megaphone 应用上支持今年早些时候由 Apple 发布的 HTTP Live Streaming（HLS）技术。该技术用于提升视频播客的分发与播放体验，Apple 预计于 OS 27 全面引入产品。Spotify 也将支持此技术，让创作者在 Apple 播客与 Spotify 的内容分发更加便利。此外，Spotify 还将与 Apple 协作，使创作者在 Apple 播客上分发的内容同样可以获得创作收益。来源3美政府批准向 10 家中国公司出售英伟达 H200 芯片据路透社，美国商务部近日已批准 10 家中国企业采购英伟达 H200人工智能芯片，获批名单包括阿里巴巴、腾讯、字节跳动和京东等。根据出口许可条款，每家获批客户的采购上限为 7.5 万颗。此外，联想和富士康等分销商也已获得代销许可。尽管政策层面获得放行，但实际交付进程仍处于停滞状态。消息人士透露，目前尚未完成任何一笔实物交付。造成僵局的原因包括美方新增的合规要求，如买家须证明具备「足够的安全程序」以确保芯片不用于军事用途；同时，中国政府也出于对供应链安全和本土产业保护的考量，对相关采购持谨慎态度，指导国内企业优先支持国产 AI 芯片发展。来源做了几天小甜甜，看哪天变回牛夫人。本来就能卖，只是国内不让买而已尽量早日摆脱 CUDA 依赖吧这种“掐着脖子又喂饭”的感觉真是让人感觉恶心，典型的短视利益。面子上做做也是好的全部讨论 11101微软消息四则Edge 浏览器取消 Copilot 模式，将 AI 功能直接整合进浏览器5 月 13 日，Microsoft Edge 官方博客发布文章，宣布取消 Edge 浏览器的 Copilot 模式，AI 功能将整合至浏览器各项功能中。移动端 Edge 引入 Journeys 功能，可自动整理浏览记录至有意义的主题；Copilot 可以从所有打开标签页中抓取信息综合推理；重新设计的新标签页也来到移动端。此外，所有版本都加入了长期记忆功能，Copilot 可以以过往聊天为参考回答问题，并可以使用浏览历史提供答案。Edge 还引入了学习模式，可以将正在浏览的内容转化为小测验、记忆闪卡等；也可以用标签页内容生成播客。另有写作助手等新功能加入。来源微软为 Windows 推出云端启动驱动恢复功能微软近日发文，宣布为 Windows 更新引入一项名为「云端启动驱动恢复」（Cloud-Initiated Driver Recovery）的新功能。该技术启用时，当微软在分发驱动更新过程中，监测到某个驱动程序存在质量问题或未通过评审时，可直接从云端触发恢复指令，让受影响的设备自动回滚至之前的稳定版本，无需用户手动干预或硬件合作伙伴重新提交补丁。该过程完全由官方端到端管理，合作伙伴无需更改现有的 API 或开发流程，也不会影响运行正常的驱动程序。功能已于 5 月进入手动验证和测试阶段，预计今年 9 月全面实装。来源安全研究员公开两个 Windows BitLocker 后门级漏洞安全研究员Chaotic Eclipse于 5 月 13 日公开名为 YellowKey 与 GreenPlasma 的 Windows BitLocker 漏洞。其中名为 YellowKey 的漏洞只需要一枚包含特定文件的 U 盘，经过极简单的操作即可绕过 BitLocker 全盘加密，无阻碍地访问全部加密文件。漏洞利用了 Windows 启动架构中的底层逻辑，将为了方便系统维护预留的机制转化为高危安全漏洞。Chaotic Eclipse 在后续博文中称，微软技术专家建议的 TPM+PIN 码对该漏洞不起效。他选择公开曝光是出于对微软消极处理漏洞的不满，且下周二还将有新内容发布。Chaotic Eclipse 曾在四月因微软未能有效响应并修复其报告的漏洞后，公开针对 Windows Defender 的 BlueHammer 和 RedSun 漏洞，RedSun 随后在无 CVE 编号的情况下被微软悄悄修复，BlueHammer 则已被在野利用。来源微软仅为 Windows 11 发布去年的 BitLocker 恢复模式漏洞补丁微软于本周发布了针对 BitLocker 加密功能的关键修复补丁，解决了导致设备在安装 2024 年 7 月安全更新后反复进入恢复模式的严重故障。该漏洞导致大量企业用户在每次重启时必须输入长达 48 位的恢复密钥。漏洞根源在于微软之前推送的安全补丁意外修改了系统的安全启动参数，导致 TPM 认为硬件环境发生改变，触发 BitLocker 保护机制。受影响的系统包括 Windows 11、Windows 10 与 Windows Server 2022，但本周发布的漏洞补丁仅针对 Windows 11 22H2 与 23H2 版本用户，其他用户只能继续采用临时措施绕过漏洞。来源真有人用copilot么？1NGINX 潜伏 18 年重大漏洞曝光5 月 14 日，安全研究机构 DepthFirst 发文宣布披露了一个潜伏在 NGINX 核心模块中长达 18 年的重大安全漏洞，代号为「NGINX Rift」，编号 CVE-2026-42945，CVSS 评分 9.2，目前无在野利用的公开报告。该漏洞存在于 ngx_http_rewrite_module 模块中，自 2008 年的 0.6.27 版本起便已存在。远程且未经身份验证的攻击者只需通过特定的 URI 请求，即可触发堆溢出，进而控制 NGINX 工作进程，实现远程代码执行或拒绝服务攻击。NGINX 是一款开源、高性能的轻量级 Web 服务器、反向代理服务器及电子邮件代理服务器，运用十分广泛。官方已紧急发布安全更新，建议用户尽快升级至 NGINX 1.30.1 或 1.31.0 及以上版本。来源不会又是 AI 发现的吧，运维近期估计都要崩溃了6三部门部署持续加强新能源汽车安全管理据新华社，5 月 14 日，工业和信息化部装备工业一司、国家市场监督管理总局质量发展局、国家消防救援局消防监督司联合召开视频会，部署持续加强新能源汽车安全管理，保障产业健康可持续发展。会议要求车辆生产企业和动力电池生产企业自觉扛起产品质量安全主体责任，要围绕产品研发设计、生产制造、供应链管理、运行监测、售后服务等方面做好风险防范，并严格履行告知义务，引导消费者正确使用车辆，不得进行夸大和虚假宣传；要建立健全事故隐患内部报告奖励机制，深化事故隐患排查，坚决守牢安全底线等。来源发表首评不妨一看的简讯Tecnoblog 曝光了一款新的 Xbox Cloud Gaming 手柄。据称手柄支持蓝牙与 Wi-Fi 6，限制带宽在 20 MHz，无线芯片为 Realtek RTL8730E，手柄内置 500 mAh 电池，有一个 USB-C 接口。来源爆料人 @aaronp613 在 X 发文，他在 iOS 版 Google Home 应用代码中找到了名为 Google Home Display 的产品名称，疑似为 Nest Hub 的换代产品，该产品最新一代发布于 2021 年。来源据路透社，微软在投资 OpenAI 后，仍在寻求在 AI 领域进一步投资，目前正在寻找能被收购的 AI 初创团队，以吸纳人才与技术。爆料人称目前在微软考虑范围的公司之一是开发大语言模型的 Inception。来源消息称腾讯首席战略官 James Mitchell 在公司 2026 年第一季度财报电话会议上表示，GPU 用于个性化广告技术与业务，可以以更短周期回本；但公司认为目前投入于混元大模型对公司的业务很重要，且满意目前的状况。来源发表首评少数派的近期动态可爱又好玩的经典萌物——监工水獭挂件复刻啦！给獭一个家首款 MBTI 定制墨水屏「口袋先知」预售进行中，一按知运势。立刻预购麦金獭钢化膜单膜补充装已上线，替换装，更划算。立即抢购你可能错过的文章AI 潮来，翻译何为用 Obsidian Web Clipper，让豆瓣种草更容易新玩意 241｜少数派的编辑们最近买了啥？> 下载少数派 2.0 客户端、关注少数派公众号，解锁全新阅读体验 📰> 实用、好用的正版软件，少数派为你呈现 🚀131扫码分享目录 0讨论我来说一句发布发表评论发布

你真的需要墨水屏设备吗？主作者关注知道分子吴先先少数派作者精神状态卢克斯古德知道分子吴先先关注知道分子吴先先少数派作者精神状态卢克斯古德联合作者关注知道分子吴先先少数派作者精神状态卢克斯古德知道分子吴先先关注知道分子吴先先少数派作者精神状态卢克斯古德05/12 11:11Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。我是一个墨水屏阅读器爱好者。从学生时代拥有第一台Kindle至今，长期持有过的设备不下20台，Kindle、海信、文石、掌阅、Kobo、大上（Dasung）、汉王、墨案、小米这些主流品牌都用过，不少还是在少数派种草的。前两年，因为折腾机器上瘾，耗时又耗钱，我痛下狠心断舍离，取关了测评账号、非必要不购新机、旧机送朋友或挂海鲜市场，一度精简到三台。没成想，最近又被漫天盖地的墨水屏广告击中，没日没夜地看起测评，还在千挑万选后入手了两台新机。这不由得让我想搞清楚：自己在这个过程中到底沉迷的是什么？为什么单单会对墨水屏阅读器那么上头？所以，标题这个灵魂拷问，实际上是问我自己的。这不是一篇科普向的文章，也不会给大家推荐具体的机型。我主要想聊聊，墨水屏这个市场中，卖家和玩家的普遍情况和特征，分享一些自己粗浅的观察和思考。1如果你想看一篇墨水屏的科普，强烈推荐看这篇文章：重度使用这款墨水屏手机之后，谈谈我对电子纸的理解和预测。别看它是五年前写的，现在回看依旧是墨水屏设备的入门科普佳作。而且，当时的产品和如今的热销款型相比，除了参数上的提升，产品形态和实际体验并没有那么大差异，仍可借鉴。类纸显示技术显示器这篇会员文章也不错，同样推荐阅读。国内墨水屏市场有多卷？不了解阅读器的朋友，一聊起这个话题，第一反应都是Kindle。Kindle作为泡面盖的形象深入人心，而在2022年，Kindle宣布退出中国市场，23年正式关闭电子书城运营，24年结束云端下载服务，似乎更印证了这个品类的衰落：如果国内市场好，行业一哥Kindle干嘛要放着好好的钱不赚，对吧？实际上，故事并未像人们预想的那样发展。倒不如说，Kindle这个老大哥一撤，原本被压抑的国产品牌竟然把不温不火的市场给翻腾起来，起码明面上，越发活跃和热闹了。从图表上可以看到，从2022年开始，Kindle不再一家独大，原有的市场份额被文石、掌阅、汉王，还有科大讯飞等国产品牌迅速瓜分。此外，还有Bigme、Meebook这样的新生代品牌先后进入市场，也有一定的影响力。或许还有朋友记得，22年甚至连华为都来分一杯羹，推出了一款10.3寸的MatePad Paper，在反响平平后及时收手，后续再没出过墨水屏C端产品。2024-25年，整个行业迎来爆发期。我一直很好奇整个市场规模有多大。查了一下媒体数据，单就阅读器（主要指8寸以内小屏、不带手写笔的款型）这一个品类，25年的总销量为64.7万台，销售额达到8亿元。虽说总盘子不算大，但比24年同比销量增长33.4%，销售额更是增长43.5%，想必产品定价的水涨船高也助推了一把。办公本（主要指8寸以上大屏、可书写的机型）的累计销量为33.7万台，销额为11.0亿元，同比下降10%左右。再加上学练机型，整个墨水屏阅读设备的总出货量在127万台左右。对这么一个不算大众的品类来说，算是不错的成绩了。如今，如果你是一个小白，想购置一台墨水屏阅读器，打开购物软件，第一反应大概率是眼花缭乱、头晕目眩。首先是品牌。这个前面有提到，墨水屏厂商层出不穷。而且从去年开始，墨水屏品牌开始请明星代言宣传了。先是文石找了Papi酱代言Leaf 5系列，今年4月汉王找了李雪琴代言，主推Clear 7锦鲤。科大讯飞请的是李现，手里拿着一款号称薄如蝉翼的AINOTE 2。讲实话，看到这些明星，我并不觉得墨水屏出息了，而是心疼自己的钱包：明星代言费肯定都要消费者来买单，抬高原本就贵得离谱的产品价格，想到这里内心更沮丧了。有一说一，这几家代言人找得都不错，算是把娱乐圈看起来脑子不错的人都请到了。下一个品牌，该不会请刘旸教主或张踩玲吧……刚刚说到，行业调研机构给出了阅读器、办公本、学练机三分法。更细分的话，面向C端用户的主要有墨水屏阅读器、阉割版手机（能插卡上网但不能打电话）、墨水屏手机、大屏办公本、墨水屏显示器、还有学练机这几大类。今年还出现了一些墨水屏小玩意，比如最近相当出圈的阅星瞳，还有少数派的「Quote/0」「口袋先知」等。当然，这其中有不少品类是重合的，像墨水屏手机、办公本也可以拿来当阅读器。还有一些产品的核心卖点就是系统开放、功能全面，大概类似于效率界热衷的All-in-one吧。主流设备的尺寸也有不少：5.84寸、6寸、6.13寸、7寸、7.8寸、10.3寸、13.3寸，每个尺寸下面都有多款可选机型。有些经典尺寸如6寸、7寸、10.3寸，基本上各家都有新品，可谓卷中卷，也是最适合新人入手的尺寸。文石在铺品类、砸硬件上绝对属于先头部队。这是2020年的文石全家福，当时就这么卷了~不光产品迭代速度快，这一两年，墨水屏产品的新品细分更是卷到不合常理。以掌阅家7寸机型Ocean 5举例。这个机型和Kindle Oasis的设计一脉相承，屏幕比传统6寸大一些，还配一对侧边实体按键，定位上属于旗舰款。2025年10月，掌阅这个系列问世的时候，一下子推出了四个细分版本：Ocean 5标准版、Ocean 5长续航版、Ocean 5 Pro旗舰款、Ocean 5C彩色墨水屏版四款，它们之间电池容量、处理器、盖板类型都不尽相同，侧重点也各有不同，真是逼死选择纠结党的节奏。我截图了几款Ocean 5产品的详情页，能看出Ocean 5 Pro（左上）用了凹屏吗？标准版和长续航版（下方）竟然连宣传图用的都是同一个……新款墨水屏设备，真的比老款好吗？讲到这里，有朋友会说：行业繁荣是好事啊，商家越卷，就越能带动产品升级和技术进步。哦，真的是这样吗？了解墨水屏的玩家知道，墨水屏的核心科技就在屏幕上。这块屏幕由于物理特性，没有普通手机屏幕的屏闪和蓝光，确实有一定的护眼功能，我认为，这是墨水屏阅读器相比于普通手机平板的核心竞争力。然而，受制于墨水屏技术的发展（还有元太科技的垄断），这块屏幕的进步空间相当有限。严格意义上说，墨水屏技术算不上前沿高科技，基本上还在几十年前的基础上，做着修修补补的工作。图源：元太科技这就导致一个现象：虽然短短几年，屏幕参数从Carta1000，一路升级，狂奔到Carta1300，但实际的使用体验依旧不是那么理想。加上各家产品优化和调教水平参差不齐，有残影、刷新慢、屏幕良品率低、有坏点，依旧是墨水屏设备的常态。带前光款的设备中，漏光、前光不均、阴阳屏的情况屡见不鲜。这导致很多第一次接触墨水屏的朋友，浅尝一下便被草草劝退。就算你是个资深玩家，每次购入新品时仍会战战兢兢，希望自己能人品爆发，收到一台品相良好的机器，避免反复退换货、和商家扯皮。客观来说，硬件的技术瓶颈，不是哪个厂商的错，是行业共同的困境。然而，对标国外经典产品，国内墨水屏虽然产品多、迭代快，但实际品控、产品力还是不太尽如人意。用过Kindle的朋友或许有类似体验：你嫌手头用了十几年的Kindle太慢、用起来不方便，被各路广告种草，打算入手一款文石、掌阅等主流品牌Carta1300屏幕、300ppi的旗舰新款。当你欣然收到新机，把它和家里古董级的Kindle放在一起，对比一下文字的锐利度、前光的均匀度时，会惊讶地发现Kindle古董机大概率整体表现更好。这还是老款表现，更不用说24年新出的新一代Kindle和大屏Scribe，这些新款的显示水平更是碾压式的。这里要补充一句，我和好几个常用墨水屏的朋友交流过，别小看文字锐利度、背景清晰度、前光均匀度这些细节，重度使用后你会发现：不知道为什么，用Kindle或某款机型就可以看很久书，不觉得难受；但换了某品牌机型，阅读时间一长就会出现眼睛不适、疲劳的情况。实际上这些细节都很影响阅读体验，大家对标纸书，想想影印版、盗版的印刷质量，或是轻型纸、铜版纸等用纸选择，就很容易想明白。小红书上有朋友对比了KP3和文石Poke6（同为300ppi）的显示效果，我觉得就我线下体验店感受，显示差异还是挺明显的我手头有一台18年款Kindle Oasis2，一直没舍得出二手，就是因为它的产品质感依旧一骑绝尘，文字显示细腻度比我后来购入的十几台国产设备都要好。至于像Kobo这样专注于显示细节的国外品牌，更是难有任何国内墨水屏单品能在阅读效果上匹敌。不过，由于国外产品主要做封闭系统，传输不便，使用起来短板也非常明显。难怪业内流行一句讽刺：「但凡Kindle支持安装微信读书（不是像现在那样只能登录网页版），就没有国产厂商什么事了」。我最近之所以又开始关注墨水屏，也因为一件事。圈内的朋友知道，海信以前在国内墨水屏领域算是头部品牌，一直稳扎稳打，从19年相继推出的A5、A7、A9等墨水屏手机，至今仍有很多忠实粉丝。可惜的是，23年海信退出了墨水屏市场，迄今再也没有推出任何新款。海信退市前，好多品类疯狂打折，我顺道入手了一款22年3月上市的6.7寸阅读器HiReader，不能插卡，但样子像个手机，开放安卓系统，性能强劲，各方面都很趁手，这几年使用率非常高。最近几个月，由于文石和掌阅家也开始强推它们家的手机形态阅读器新品，我又开始种草，想迭代一个新款。考察了一下，文石小白马升级了Pro版（可插流量卡、5G、有摄像头，还有彩屏款可选）和小白马Plus款（不可插流量卡，支持手写，但单卖设备不带笔），掌阅则升级了Tango系列，推出了Tango2和Tango2 Pro（能插流量卡，带摄像头）——晕不晕？晕就对了。讲实话，我在实体店体验过小白马，虽然新款用的Carta1300屏幕比HiReader的Carta1000新了好几代，但屏幕质感、灯珠、运行速度等综合对比后，还是海信更胜一筹，直接拔草。掌阅Tango2 Pro虽然屏幕显示不错，字体更黑，更有跃然纸上的感觉，我收到的这一款品相也不错，但第三方软件运行起来，跟海信还是有差别。尤其是掌阅的刷新模式虽有5档，但在刷网页时，没有一个档位像海信「均衡」档更让我满意。要知道，海信真的好几年没更新墨水屏固件了。不过我还是留下了，毕竟算是有差异化亮点。网友对比HiReader（左）和Tango2Pro（右），具体观感见仁见智，但海信这款屏占比是真的优秀。举这个例子不是想拉踩谁家，毕竟海信刚刚又有重新进军墨水屏市场的意图，把HiReader机型烫印个2026字样，软硬件什么都没改就拿出来卖，价格还不便宜。通过这个例子，我就想问一句，虽然这几年国内厂商卷到飞起，价格也是节节攀升，但真正在屏幕优化、产品设计方面，大家有在用心吗？比不过Kindle，还可以找到借口，说亚马逊在供应商那里有优势地位，但和同为国内厂商的海信相比，26年那么多阅读器新品，个个价格直逼2000+，却无人完胜海信家一款4年前的老产品，真的让人感到遗憾。所以，我的结论是：新款肯定贵，但未必如你想象得那么好。如果你有一个还算趁手的旧款，没必要盲目换新。毕竟一台质量合格的阅读器，用上十年、二十年都没问题的。为什么喜欢墨水屏的人，会买了又买？虽然大家理性上都知道阅读器多买无益，还耽误真正的看书时间，但是实际情况是：一旦你拥有了一台阅读器，开始关注墨水屏这个领域，大概率往后会越买越多，手里不会只有一台墨水屏设备。刚刚说了，我前后拥有过的不止20台设备，但在墨水屏这个小圈子里，我这种水平真算不上什么。专业博主除外，你翻翻小红书，动辄十台八台的用户多了去了。如今把任意型号输入搜索框，都能找到大量的晒单或吐槽贴，这在以前是不敢想象的。这几年社交平台流行一个梗，叫「墨水屏繁殖」，讲的就是你买了一台墨水屏就会想买第二台，就像它自我繁殖一样。如果你对这件事没有感性认知，不妨看看这位玩过50多台设备的豆友的墨水屏大赏：论墨水屏阅读器如

快捷指令｜纯粹、专注、能群发，一种「点名式」微信聊天新体验主作者关注Linery新手上路Linery关注Linery新手上路联合作者关注Linery新手上路Linery关注Linery新手上路05/13 20:36用微信发信息是我们熟悉到几乎成为本能反应的日常操作，不过作为国民应用，微信为了保证人人可用的普适性，也一直保持着极为克制的功能更新节奏。直到七年前鸟巢那场颇具争议的「灾难级」发布会上，锤子 TNT 次世代生产力工作站发布时，老罗在最后展示了他对「聊天应用」的理解——子弹短信。子弹短信以其极简高效的交互设计、处处贴心的产品巧思，让人耳目一新，也成为了那场发布会后少有的、备受好评的产品。可惜彼时锤子科技已经风雨飘摇，子弹短信最终没能成为击倒巨人的大卫，在变身「聊天宝」尝试「聊天赚钱」的模式也没能最终挽回颓势后只能草草收场。这个意难平也让我在用微信的时候总感觉有一根刺：如果聊天可以再「畅爽」一点就好了。比如我可以直接在一个界面里快速给多个联系人发私聊信息，这样在同时需要联系很多人的时候，可以不用频繁打开，关闭聊天窗口。我试着让 AI 帮我列举了一些需求场景：周末聚餐，要挨个通知朋友时间地点；临时加班，要分别跟家人、伴侣、长辈说一声晚点回家、不回去吃饭；工作上有事，要同步给同事、领导、合作对接人；出门用车、约球、约见面，要一一私聊好几个人。再比如更智能的群发，快速给不同的人发想同模板的信息时，如果能自动填充联系人的「称谓」等字段，也能让群发更人性化；我甚至也想用微信做更多的信息记录，比如聊完天可以记录一下感受、吐槽、资金往来、注意事项等。从锤子转到 iPhone 阵营后，出于一些「新奇」功能的急切需求，我开始捣鼓起了 iPhone 的「快捷指令」，用它来解决工作和生活中遇到的问题。并且最终我渐渐做出了一些围绕微信聊天的快捷指令。今天分享其中的一个。为了解决「不频繁打开窗口，同时给多个人发信息」的需求，我利用第一性原理得出了一个简单的结论：聊天本质是给指定联系人发送一段文本。所以不妨剥离冗余的交互，只保留最核心的「点名 + 发消息」。在设计时我发现快捷指令完全可以实现「输入联系人名称、再输入要发送的文本、最后发送出去的效果」：在锁屏上直接发微信的效果举点日常真实示例：妈 晚上我不回家吃饭了奶奶 今晚在外边吃，打完球给你带榴莲包回去万基 晚上七点球场准时 solo世平 好久没见，找个周末抽空聚聚Peter 方案你明早抽空看一下，下午拿给老板签字莞少 今明两天我要用车，你这两天别把车开走不用来回跳转、不用重复复制切换，静下心一次性把所有人的话都编辑好，一键自动群发，省心又高效。这个功能每次用都非常愉悦，一是聊天体验新奇（我觉的很像学生时代传纸条），二来是真的完美帮我躲开了碎片化通信对于正常生活和工作的干扰——作为潜在的 ADHD 人士我会经常陷入这种情景，点开微信本来只想发一条消息，结果一眼刷到朋友圈、被视频号内容勾走注意力，半小时过去，信息还没发出去。另外就是每次用它一次性把该对话的人该说的话说完，有一种特别畅快的感觉。它当然无法替代微信原本的日常聊天、连续闲聊、刷圈看动态的需求，但是作为一个于我而言的重要补充，真的非常好用、可靠。最后说点感性的话：我是一个非常怕麻烦的人，对所有可以减少冗余步骤的事都有极大的好感，但我也逐渐能理解这个世界上为什么会有这么多「麻烦」。麻烦对我而言变得中性，我意识到它代表了一件事的中间态，在优秀的产品（比如微信）中它常常体现为极致的折中，这同样也需要付出大量的工作、甚至是天才的创想才能达到的。这在我看来真的非常了不起。快捷指令就是 iPhone 提供给用户，用来化解这些「麻烦」的一个非常好的工具。它有丰富的接口，例如在这个功能里，就利用了微信提供的「给联系人发信息」接口；也有非常多的触发方式，包括但不限于桌面图标，锁屏、控制中心、负一屏小组件，锁屏快捷按键，轻敲背面，操作按钮（原静音拨杆）——当然还有少数派的 Bang!Case。这都让它变成了一个能让每个人通过轻度开发解决实际问题的工具。我真的非常喜欢快捷指令，它让我学会去缕析一件事的脉络、链路，动手去改写、制定新的规则和流程，让工作和生活都接近我想要的理性，跳脱出很多无形的枷锁。针对微信我未来还会做很多有趣的功能，比如智能群发，定时发送等等，感兴趣的朋友可以保持关注。我把这个快捷指令整理好了，有需要的派友可以安装试试，添加到自己趁手的位置就可以了。编注：题图由 ChatGPT 生成。> 关注少数派小红书，感受精彩数字生活 🍃> 实用、好用的正版软件，少数派为你呈现 🚀3026扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@克莱德© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#自动化#效率#微信#快捷指令30等 30 人为本文章充电扫码分享举报本文章Linery还没有介绍自己关注

Apple WWDC26 全球开发者大会，来与少数派一起看主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。05/15 16:59北京时间 6 月 9 日凌晨，WWDC26 全球开发者大会首日 Keynote 主题演讲活动即将揭幕。届时，我们将迎来 iOS 27、iPadOS 27、macOS 27、watchOS 27、visionOS 27 等新一代操作系统的亮相，Liquid Glass 设计语言将如何打磨进化、iPadOS 多任务体验又将走向何方…… 这些大家关心的问题也将在会上悉数得到解答。与此同时，macOS 27 彻底告别 Intel 时代，会带来哪些新变化，也是今年 WWDC26 值得关注的看点之一。为了与同样关注 WWDC 的你一同迎接这场科技盛会，少数派将于 WWDC26 首日 Keynote 主题演讲活动期间在深圳举办线下直播观影活动。欢迎你来和我们一起交个朋友、一起聊天，观看 WWDC26 全球开发者大会。购票链接：深圳场次广州场次其它场次活动信息活动日程 (深圳)活动时间：06 月 08 日 22:00 - 06 月 09 日 03:00活动地点：广东省深圳市宝安区西乡街道银田工业路元匠坊 C 座 201 (少数派共创总部)活动日程 (广州)活动时间：06 月 08 日 22:00 - 06 月 09 日 03:00活动地点：广州市荔湾区泮塘五约外街 76 号少数派 (1200 书店旁)活动门票本次活动为门票收费制形式，门票价格 9.9 元，少数派会员可免费入场。所有门票收入将用于活动的组织和宣传，包括不限于场地租赁、茶歇准备、物料制作、伴手礼品等，由少数派方面负责管理，所有解释权少数派所有。编注：由于场地空间有限及后台系统方面的限制，少数派会员仍需下单购买门票后才可入场参与本次活动，下单时可享优惠金额 0.1 元。活动礼品活动门票定价为 9.9 元，除了入场资格之外，每位参与者还可获得：少数派 Prime 月度会员卡少数派 logo 装饰贴纸现场抽奖机会多城同屏，同步相庆本次 WWDC26 派对之夜线下特别活动，我们依然与少数派的老朋友 SwiftGG 共同举办，如果你不能参加我们在深圳和广州举行的线下活动，也可以报名在北京、上海、杭州、青岛同步举办的线下观影活动，现场同样准备了精彩的演讲和互动环节，并会为到场观众发放精美伴手礼品。了解详情 >Apple Developer Community Events在 Apple Developer 官方网站的社区活动页面中，Apple 已将少数派与 SwiftGG 联合举办的 WWDC26 系列活动收录至页面中，你也可以在此找到全球其它国家和地区经过 Apple 挑选的相关活动。>Community-driven events - Apple Developer温馨提示由于活动举行时间较晚，请购票用户合理安排行程，准时参加，并注意路途安全。活动伴手礼以活动现场发放实物为准。所有门票收入将用于活动的组织和宣传，包括不限于场地租赁、茶歇准备、物料制作等，由少数派方面负责管理，所有解释权归少数派所有。请在观看 WWDC26 Keynote 主题演讲期间保持安静，确保不会影响其他观众及周边居民。活动期间请注意个人物品的保管，建议提前规划活动结束后的归家方式，尽量避免步行或单独行动，保证个人安全。鸣谢本次系列活动由少数派、SwiftGG、漕河泾 Hi-Tech Park 等合作方共同举办，感谢各位合作伙伴对于本次活动的支持。如果你有其它活动想法或者场地推荐，欢迎通过邮件联系我们：[email protected]。153扫码分享目录 0讨论我来说一句发布发表评论发布本文责编：@waychane© 本文著作权归作者所有，并授权少数派独家使用，未经少数派许可，不得转载使用。#WWDC26#Apple15等 15 人为本文章充电扫码分享举报本文章少数派编辑部少数派编辑部官方账号。关注

按：从播客访谈到深度长视频，越来越多的内容创作者把表达重心放到了音视频里；「先转录、再阅读」也因此成了不少重度信息消费者的日常工作流。。本期「装了啥」，我们请编辑部几位同事分享各自正在使用的转录方案，希望其中至少有一种思路能为你所用。MacWhisper 的平价替代：TranscribeX@广陵止息：虽然我很喜欢文字传递的高密度信息，但越来越多的内容创作者已经转向播客和视频；至于多人访谈类节目，音频传达出的现场氛围感也比文字更生动。所以我这些年消费音视频节目的时间也逐渐拉长，但有价值的节目「三心二意」地听一遍远远不够，反复听又没有足够的时间和精力。所以，现在遇到有价值的节目，我都会转录成文本，方便后续检索。从 Time Machine 里翻了翻，找到了当时的文件夹记录以前我采用的是whisper.cpp配合 OpenAI Whisper Large V3 模型来转录文本，但这个方案有两个问题。一是音视频下载的渠道很多，挨个手动下载很麻烦；二是转录时间偏长的中文视频时，有不小的概率会在转录过半之后出现重复输出、卡字，导致转录失败。但我实在不想翻 log 和模型仓库找相关讨论了，所以也开始寻找替代方案。首先想到的是 MacWhisper，看了眼当时价格已经涨到完全不考虑的地步。好在后来找到了整体体验和 MacWhisper 很类似的 TranscribeX，除了收录在 Setapp 里，对我没有额外成本外，即使单买也只要 MacWhisper 一半价钱。和 MacWhisper 类似，TranscribeX 支持本地文件导入转录、麦克风录制、录制特定软件、实时转录、视频网站下载转录和批量转录。其中「下载视频网站」功能主要依赖 yt-dlp 实现，所以基本不用担心国内服务的下载问题。转录模型方面，TranscribeX 也支持 WhisperKit、Parakeet、千问和 Mistral 的转写模型，必要时还可以调用 ElevenLabs 的在线 API。不过我本地基本只用 Whisper Large V3，别的模型也没用过，所以暂时不评价转录效果。相比 MacWhisper，TranscribeX 支持 Apple Silicon 中的 NPU 加速，这也是我之前选 whisper.cpp 的原因。NPU 虽然不会让转录速度快很多，但至少负载不会全部堆到 GPU 上。设置时需要注意：Audio Encoder：主要负责「听懂声音」的环节，可以只用 NPU；Text Decoder：主要负责听写的环节，为了转写速度，建议设置为 All。转写界面其实每个软件大同小异，但还是有一些设置可以说一下。首先是自动检测「说话者」：Pyannote：适合会议、访谈等场景相对固定的情况，也更贴合我常用的场景，可以快速调用预定义的模板；Sortformer：偏向速度、实时和轻量化的处理方式，适合会议、通话等实时场景。外语苦手也不必担心，TranscribeX 内置了三种翻译方式：Apple 内置的翻译 API、DeepL API 以及第三方 AI API。Apple 内置「翻译」速度很快，准确率也够用，而且免费；对翻译质量有要求的话也可以调用 DeepL 或者其他 AI。TranscribeX 也随大流加入了 AI 对话功能，可以针对转录的内容做进一步处理。最后值得一提的是，针对 Setapp 版本的 TranscribeX，所有的在线 AI 额度都可以用 Setapp 的 AI 配额。但 TranscribeX 还是有自己的问题：没有 CLI 版本，远程使用一定要靠 VNC 连接到 Mac；辅助框架的设计上也有不小的问题，目前还很难配合比较火的 AI 助手。如果你比较在意自动化体验，TranscribeX 可能没有那么适合你。用 Gemini 整理@PlatyHsu: 我这里主要谈谈转录后的整理。对于这类任务，我最推荐 Gemini 系列模型。主要原因是，在排名靠前的模型中，Gemini 的多模态支持比较全面，Gemini 3 全系（Pro、Flash 和 Flash Lite）都可以直接接受音频和视频文件输入，免去了先创建音频转录或者视频截图的功夫。我也认为 Gemini 的中文能力在国外模型中相对是比较好的。与此同时，Gemini 的用量给得比较慷慨，AI Pro 方案很少遇到不够用的情况。我最常用的场景之一是整理播客的录音稿，用来确认一些初听时一闪而过的模糊细节。一个参考提示词是：转录用户提供的音频，并整理成表达规范、分段清晰的书面格式。区分说话人，并标注在段落之前。将较长的片段合理拆分成若干自然段。可以删除语气词和重复片段或修改病句，不要做其他删减或总结。具体模型方面，一般可以先从 Flash 开始尝试，对于日常主题的节目来说一般就够了。一小时的播客大约需要四分钟处理，肯定没有本地模型快，但是成品高度可读，值得这个等待时间。不过我也遇到过一些涉及哲学、历史等较专业话题的节目，用 Pro 模型会有更好的表现，可以灵活调整。当然，如果想节省 token，你也可以先在本地转写一遍，然后用以下提示词整理文本：将用户给出的音频转录整理成表达规范、分段清晰的书面格式，可以删除语气词和重复片段或修改病句，不要做其他删减或总结。这个用例一般建议直接使用 Pro 模型，因为本地模型转写的文本一般有很多拼写错误和赘余，Pro 能更好地还原和修复。除了音频之外，Gemini 也可以用来总结视频。并且由于它可以直接「看到」视频内容，因此即使是那种没有旁白的 TVC，也可以说出个一五一十。例如之前一个项目中，甲方提供的写作素材之一是一条广告片，我就是用这个方法把其中的信息提取成文字描述，然后用在了写作里。以下是用苹果最近一条 Mac 广告做的演示，提示词是：详细描述该视频的主题、思路和内容，并列举关键帧的时间戳和画面内容。因为涉及较多「言外之意」的解读，这个用例也是用 Pro 模型处理为佳。以上方法在 Gemini 网页版和 AI Studio 都可以使用，但网页版经常会出现「偷懒」的情况，只给出非常简略的输出，因此建议尽量用 AI Studio 来操作。命令行工具搭配 Agent@waychane: 由于常有参加线上会议的需求，会中涉及主讲者的介绍、采访以及 QA 等内容，这类会议的时长动辄就有一个小时。在会后创作内容时，如果能有一份观点清晰、结构清楚的要点大纲，会省下不少功夫。现在可以说是 Agent 时代，我自己也在用 Claude Code 等 Agent 类 AI 工具帮我处理各种工作。因此在挑选特定场景下的 app 时，我会希望它们要么提供 CLI 操作方式，要么就是纯粹的命令行工具：一方面，这样可以在同一个 Agent 界面中高效地调用不同工具；另一方面，许多 app 不仅功能复杂，使用参数也多到难记，用 Agent 调用刚好能解决这些问题。面对会议音频转录的需求，我选择的是 whisper.cpp 这个命令行工具。它是 OpenAI Whisper 的 C/C++ 移植版本：OpenAI 官方版 Whisper 基于 PyTorch 的 Python 实现，要跑起来得先配好 Python 环境和一整套依赖，在 Mac 上只能依靠 CPU 计算，想用到 GPU 还得自己折腾 PyTorch 的 MPS 后端，体验并不顺畅；whisper.cpp 则用 C/C++ 重写了推理部分，原生接入 Apple 的 Metal 后端和 Accelerate 框架，能有效利用 Apple Silicon 的 GPU 和统一内存架构，模型权重也换成了体积更小、加载更快的 ggml 量化格式。

本周看什么 | 最近值得一看的 8 部作品主作者关注少数派编辑部少数派成员少数派编辑部官方账号。少数派编辑部关注少数派编辑部少数派成员少数派编辑部官方账号。联合作者关注潘誉晗少数派作者潘誉晗关注潘誉晗少数派作者关注利兹与青鸟少数派作者利兹与青鸟关注利兹与青鸟少数派作者关注SHY少数派作者SHY关注SHY少数派作者少数派编辑部 等 4 位作者05/15 18:00☕️ TL;DR近期佳作推荐：[国产] 低智商犯罪、[韩剧] 夺命许愿、[英剧] 好兆头 第三季、[法剧] 抢钱女王团、[电影] 蒙特利尔，我的美人、[动画] 石纪元 科学与未来 第 3 部分、[纪录片] 马克斯·维斯塔潘：新征途、[真人秀] 旧洞基朋友们几则精彩预告：《攻壳机动队 THE GHOST IN THE SHELL》新预告、《沉默的审判》释出预告、《火遮眼》新预告、《爱情城事》定档预告、《奥特曼起源》正式预告几则影视资讯：《诺曼底 72 小时》内地定档 6 月 6 日、《宇宙巨人：希曼崛起》内地定档 6 月 5 日、《家弑服务》内地定档 5 月 29 日、《超级少女》内地定档 6 月 26 日、漫威新剧《幻视追寻》正式定档往期推荐可点此或点击文末 tag #本周看什么，更多主题片单可点此查看；想要查看更多深度影视内容，可以关注我们的衍生栏目「看什么 Café」。如有其他建议也欢迎在评论区提出，感谢大家对「看什么」栏目的喜爱和支持。[国产] 低智商犯罪关键词：喜剧 / 犯罪片长：45 分钟左右（单集）× 24 集观看渠道：爱奇艺/咪咕视频；豆瓣链接运气之神降临！@利兹与青鸟：平安市刑警张一昂能力不强但热爱吟诗，常常不合时宜得令人发笑，被下放到老干部处都能惹出乱子。因为他正巧捡到一封举报信，于是和同样脑回路清奇的李茜一起被派到三江口，调查公安局副局长的死是否有隐情。但空降而来的张一昂很不受欢迎，工作处处受限，却在阴差阳错之下抓住一个在逃嫌犯，让人刮目相看，就这样被架起来成了神探。另一边，平安市两个笨贼想出了抢贪官的主意也来到三江口；而三江口涉黑的集团老板周荣考虑着新项目，也着手调查发小叶剑刑警队长的死因。三条故事线、两桩凶案（可能不止）就这样汇集一处，想必会给三江口这一团乱麻摩擦出不少火花。本剧改编自紫金陈的同名小说，一开局就满是带着黑色幽默的荒诞味，罪犯和警方彼此高估，虽然智商时常掉线，但不妨碍瞎猫碰上死耗子，不禁让人好奇这份「好运」还能带来什么啼笑皆非的巧合。马旭东、闫佩伦、詹鑫等喜剧演员的助阵也让笑点呈现得更好，尽管罪案线稍显拖沓，但也算是合格的电子榨菜。4[韩剧] 夺命许愿关键词：剧情 / 恐怖又名：기리고 / Wish Your Death / If Wishes Could Kill片长：44 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接为了愿望达成，你愿意不惜一切代价吗？@潘誉晗：这次数学考试的结果有点出乎意料，因为崔亨旭拿满分了，可这家伙不是个读书的料子。面对好友们的询问，亨旭得意洋洋地把一个名叫「Girigo」的软件链接分享给了伙伴，说是只要写上自己的生辰八字贴在自己的身上，然后录下视频上传到这个软件上，就能实现愿望。大家并不相信亨旭的话，亨旭也没在意此刻出现在他手机上的倒计时。谁承想 24 小时后，亨旭就当着全班同学的面自缢了——原来这则所谓的许愿软件，在愿望达成后会要你的命。Netflix 韩国这次又推出了一款爆剧，一经上线就在 13 个国家登顶。剧集整体的节奏很是紧凑，恐怖惊悚的氛围也营造得相当出色，很好地把校园青春剧和恐怖因素结合在一起，然后又完美连接了手机 app 和萨满巫术。神秘的夺命软件利用了青少年的欲望，随后就像是一个大家无法挣脱的隐形牢笼，不仅顽固地无法在手机上卸载，更是在许愿者身后步步紧逼、致人于死地。1[英剧] 好兆头 第三季关键词：剧情 / 喜剧 / 奇幻又名：Good Omens Season 3片长：99 分钟观看渠道：Prime Video；豆瓣链接我有那片宇宙，我还有你，我已经拥有我想要的一切。@SHY：成为至高大天使的亚茨拉斐尔，开始筹备「第二次降临」，复活的耶稣却不知所踪，只得向自暴自弃的克劳利求助。这对自创世之初就有交情的天使和恶魔，能否像往常一样弥合分歧，而在背后作祟的幕后黑手，又会将整个世界导向怎样的结局？受原作者兼前两季主创尼尔·盖曼的性侵丑闻影响，最终季被砍成了仅 1 集的特别篇，篇幅势必捉襟见肘。不难想见，成片处处流露着删减的痕迹，忽快忽慢的节奏让叙事更加混乱，整体观感有着不小下滑。虽然填充的内容变了样，两位新编剧还是尽力还原了盖曼构想中的结局，而在破碎的字里行间，也能窥见一些残存的灵光。惯常的黑色幽默难以撑场后，本季的重心落在了本来带点调侃的父母爱情故事上，大卫·田纳特和麦克·辛以一如既往的真挚演技力挽狂澜，继续迸发充沛的化学反应。彻头彻尾的粉丝向结局，评价两极分化在意料之中，好在，本作原本就是作为限定迷你剧立项，如果你有点难以接受这一季的走向，只把第一季尊为正统也未尝不可。发表首评[法剧] 抢钱女王团关键词：喜剧 / 惊悚 / 犯罪又名：Les Lionnes / Cash Queens片长：55 分钟左右（单集）× 8 集观看渠道：Netflix；豆瓣链接我们之所以抢钱，真的是迫不得已。@潘誉晗：罗莎这天找到自己的上司，希望可以提高她的信用卡额度。可上司不仅很快就拒绝了她的请求，甚至通知她的额度被降低了，因为她征信由于入狱前夫的债务问题受到了影响。罗莎不敢相信上司的话，也就是在这时，银行的警报铃响了，原来这天又是银行例行的演戏日。看着上司希望职员们保障生命的大喊声，一个不该出现的念头在她的脑海中出现了——她决定伙同自己的好友，抢劫自己工作的银行。Netflix 法国于两个月前推出了一部女性犯罪喜剧《抢钱女王团》，讲述了五名女性在窘迫生活的逼迫下，不得不伪装成男性抢劫银行的故事。全剧充斥着大量很荒诞又很现实的设定，主角团的五名女性拼命工作、好好生活，但却一直过得不幸福。职场上的性别歧视、身上背着重债、被同侪做局丢掉了工作、被丈夫家暴……不是说努力就能有好结果吗？为什么在她们的身上，越努力却似乎越悲伤呢？发表首评[电影] 蒙特利尔，我的美人关键词：剧情 / 同性又名：Montréal, ma belle片长：117 分钟；豆瓣链接被困在夹缝中时，是面对自己还是活在谎言里？@利兹与青鸟：凤霞一家已经移民加拿大十几年了，在魁北克省的蒙特利尔开着一间小超市。丈夫性格较为古板，苦于没有当地学历，找工作频频受挫；大女儿正值叛逆期，对待父母多是指责和抱怨；儿子也只是上小学的年龄，家中没有人能为受困于更年期的凤霞提供情绪价值。在法语课上，凤霞被年轻人追求梦想的热情所鼓舞，尝试接触同性交友网站，并结识了 30 岁的卡米尔。这让凤霞重新找回了自己，却引发了家庭危机。本片的华裔女导演很有相关生活体验，她用代际间新旧观念的碰撞、婚姻大于爱情、照顾家人大于自我实现等细节，呈现出一个传统而压抑的中国移民家庭外壳。片中演员也多是当地华裔，汉语都很纯正，对华语区观众的耳朵十分友好。电影虽是在讲中年女性在道德边界寻找自我，却没有激烈的口号，而是用朴素温和的风格、平缓的节奏，给主演陈冲留足表演空间，为年过半百却天性浪漫、依然优雅而富有魅力的凤霞赋予灵魂，细腻又克制地演绎出她内心的彷徨与挣扎，也让观众沉浸其中。女主角是陈冲啊？1[动画] 石纪元 科学与未来 第 3 部分关键词：漫画改 / 科幻 / 冒险又名：Dr.STONE SCIENCE FUTURE 第3クール片长：24 分钟（单集）× 13 集，每周四更新观看渠道：巴哈姆特动画疯；豆瓣链接We choose to go to the Moon.@SHY：孤注一掷的抗争后，科学王国终于取得对北美阵营的胜利。面对迫在眉睫的威胁，千空和赛诺搁置矛盾，联手打造宇宙飞船，前往月球探寻石化的真相。为了复活更多人类，同时取得不可或缺的材料，众人兵分三路执行任务，「WHY 人」却传来了更加危险的信息。从石器时代到现代文明，跨越 200 万年光阴，这段从零开始的科学冒险终于迎来最终幕。从刚苏醒时手工打造的粗糙制品，一步步重现火药、铁器、蒸汽、电力，来到本季，科技树终于攀升到计算机和互联网的领域。要实现登月的最终目标，需要远胜以往的人力和物力，千空等人因此踏上环球之旅，整合各个国家的资源。即使原料由石头进化到超合金，舞台由村落跃进至外太空，本作的核心魅力始终如一。秉持理性和探索精神，用科学之力对抗奇幻，挥洒荡气回肠的热血，开拓人类未来的道路，诠释纯粹的浪漫。这个集结全世界力量、于石之世界铸造的「阿波罗计划」，绝对能让人心潮澎湃，剧中历尽千辛的角色们，也值得最为美好的结局。发表首评[纪录片] 马克斯·维斯塔潘：新征途关键词：纪录片又名：Max Verstappen - New Ground / 极速新域：维斯塔潘的北环之路片长：23 分钟左右（单集）× 3 集；豆瓣链接出发吧！汽车人！@潘誉晗：对于喜欢一级方程式赛车的人而言，效力于红牛车队的马克斯·维斯塔潘，是一个了不起的名字。他出生于赛车世家，自己的赛车成绩也非常傲人。不过卸下这一身份，他其实还有另外一种人生——一家 GT3（Grand Touring 3）车队的老板兼赛车手。而这部只有三集的纪录片，就是讲述维斯塔潘的第二身份。「看，老板亲自上场了。」维斯塔潘没有老板的派头，招呼完就直接上了场；他也不需要拥有全部欢呼与掌声，而是会亲自去挖掘赛车手。于是，他发掘了克里斯，这个和他一样从开卡丁车（并且克里斯的车号是 3）开始的小伙子很有天赋，可苦于 15 万欧元的培训资金，只能把梦想转移到模拟器赛车上。当维斯塔潘找到他，让克里斯在场地上尝试开车的时候，克里斯甚至没有驾照。相信在观看了这部纪录片之后，没人会否认维斯塔潘对赛车的热爱。他脱下了红牛的飞行夹克，在另一条赛车的道路上，继续驰骋。发表首评[真人秀] 旧洞基朋友们关键词：真人秀又名：구기동 프렌즈 / Six Singles Under One Roof片长：90 分钟左右（单集）× 10 集；豆瓣链接一起同居吧。@潘誉晗：韩国最近又推出了一部真人秀《旧洞基朋友们》，找来了张度妍、张根硕、安宰贤、李多熙、景收真和崔丹尼尔六位年龄想尽的艺人，过一次同居生活。随着明星观察类的项目在市场上越来越多，我相信有越来越多的人对这一类真人秀感到厌倦。而《旧基洞朋友们》的出现，再次证明了韩综的创意。看似依然是粉丝向地找了一群有粉丝基础的演艺人员，但是却找准到了一个很细腻的切口——让几个曾经独居过的明星们，过群居生活。独自生活，只是光看到就觉得很自由的日子，为什么他们体验过那种日子后，不愿意再独居了呢？我想这也就是观察类节目出现的初衷，去探讨一个社会性的议题。毕竟「存在即合理」。而这六名 85 后明星的同居日常，也非常有看点。他们都很有个性，但又能互相融入、靠近。见惯了大场面的他们，看到 18 公斤的章鱼，依然会手足无措，还是会大声尖叫。原来群居也不差嘛。1更多[电影] 总统的蛋糕@利兹与青鸟：萨达姆的独裁统治下的伊拉克，平民们过得水深火热。小女孩拉米亚住在草编的窝棚里，连书包里的苹果都要被老师搜刮走。更不幸的是，她被抽到要在总统生日这天做好一个奶油蛋糕。于是拉米亚踏上了旅途，用家里所有值钱的东西换取做蛋糕所需的材料。电影在去年戛纳电影节获金摄影机奖，摄影优美，配乐充满阿拉伯风情，有触动人心的童真，也有伊拉克战乱下的可怖的倾轧。[电影] 平场之月@潘誉晗：电影根据朝仓加守美同名小说改编。青砥健这一天意外偶遇了自己的初中同学须藤叶子。此时两人已经 35 年没见了。中学时两人就互有好感，这次重逢，已离婚的青砥健和丧偶的叶子，终于还是走到了一起。埼玉小城里这一对中年人迟到多年的遗憾恋曲，就像是终于遇见了自己的另一个缺失的半圆，内敛、沉稳又有淡淡的愁绪。[韩剧] 浪漫的绝对值@潘誉晗：每当晚上，看似平凡的高中生吕意珠就会偷偷创作 BL 小说。可惜，她的作品一直无人问津。这天，隔壁搬来了

厂商、媒体与社区共谋的死灭回游：「主观体验」的困境与出路的试探讨主作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了联合作者关注Leobai少数派作者布瑠部由良由良，好，我先溜了Leobai关注Leobai少数派作者布瑠部由良由良，好，我先溜了05/15 00:07Matrix 首页推荐Matrix是少数派的写作社区，我们主张分享真实的产品体验，有实用价值的经验与思考。我们会不定期挑选 Matrix 最优质的文章，展示来自用户的最真实的体验和观点。文章代表作者个人观点，少数派仅对标题和排版略作修改。免责声明与作者所需披露的一切首先声明一下——本人仍然是在校的本科微电子硕士考古学的大学生，并没有经过任何的传媒学相关的学习，也没有从事相关行业的经验。目前仅有的和媒体沾边的事情是在酷安、少数派等平台发过一些文章（其中有很大一批都是在最后诉诸于，说好听点是真实生活场景，说难听点开始抒情的文章），以及一个 B 站粉丝量远没有破千的账号，因此我远远没有承担过养一个工作室的压力与责任，也没体会到在竞争激烈的环境中生存的艰难。因此以下内容不构成任何形式上的专业意见，只是一个愤青在 Random Yapping。多有冒犯之处，但我本心无意去批评任何具体的某一个媒体老师，只是想尝试提出一个也许还算有点价值的问题。以下内容提到「游戏区」，也仅限于单机游戏，不包含手机游戏、桌游等类型。以下内容提到「数码区」时，通常仅指以手机、一部分 PC 组件为主的主观体验向评测，且我的能力不足以涵盖 DIY 区1以及涉及到实验设计与数据分析的其他评测内容，虽然随着硬件厂商的创新进入后期演变为渐进创新，产品节奏也有所放缓，在社区中也能听到「DIY 区不如以前」的感慨，但本文尽力按照「不理解，没有切身观察就尽量少乱锐评」的原则展开，但撰写过程中有很多内容没收住，因而请让我提前声明一下——以下很多内容只是社区观察，不完全代表事实判断。最后，说句坦诚的实话，这个内容写的过程真的很痛苦——我真的不想让这篇文章变成直球辱骂媒体老师、攻击厂商或者是殴打观众的情绪发泄文，从个人发展角度来说不希望因此在进入之前就已经和某些职业无缘，也不想因此得罪那些因真心倾慕而认识的那些做这一行的朋友，且我真的不希望因为自己的过分言论把发文章的平台本身也拖下水，但因为笔力实在是有限，尽管拉着 Gemini、GPT、DeepSeek 等一众大模型核验，也找了一些朋友探讨，尽可能不要直球攻击特定的一两家媒体也不要说的太过分，但是字里行间还是有一股火药味拼尽全力消不掉。真的很抱歉，尽管我其实将自己也包括在本文的攻击射程以内——我自己就是后文中所批判的那个自认为掌握了真实，在没有厂商利益束缚与照顾社区的必要下肆意输出内容，甚至有很多是过当内容的小自媒体。这些内容因为是我自己一段时间内自我质疑与焦虑的外化，因此，本文必然存在一些因深度不足、角度有限而导致的难以避免的暴言，简而言之——就是年轻时犯下的错误。麻烦各位读者多多体谅。也欢迎您在评论区里多批评指正，让我多与您交流学习。以下内容如果有哪里说的不合适的地方，有什么可能涉及到的法律争议，请尽快联系我去撤换文章中对应的部分，谢谢！本文长度约为三万字左右，前三部分观点将会较为尖锐，会举几个相对较为敏感的案例来佐证观点，后两个部分则尝试解决前文抛出的问题，内容相对温和但字数也相对较长，请坐和放宽。主观口胡系列是一个尝试将动漫、电影、游戏叙事与更加严肃的讨论结合起来，去尝试探讨一些特定大命题的个人尝试，可能会一定程度上背离这些作品的创作初衷，聊的可能也会比较混乱且随意，甚至会有意提取个别元素，曲解原文本来聊一些其他领域的内容。这次我想把一个大约是三四年前就开始产生的想法去做进一步的深化，并尝试结合近些年来我对游戏区与数码区的一些粗浅且悲观的看法，来尝试去聊一个问题，一个也许注定没有答案，跨多少领域学科都难以得出解法的一个小小的问题——主观体验向评测的现状怎样，尽头到底为何方？又是什么构成了它们的专业性？Part1：太多水，太多奶酪——主观内容的梦魇虽然这么说有失偏颇，但我一直觉得游戏评测是一个相当主观的内容——评测者的个人阅历2、偏向的游戏类型甚至是组成一个游戏的系统，对 GDC 中诸多开发者分享的设计理念的理解，以及同期的同类型作品等等——这些都会对你最终得出什么样的结论，给游戏最终打什么分造成影响。所以在因近期的《神探杰克鼠》评测而陷入争议的 IGN 内部——作为一个有 TGA 评选权的，相对来说算是行业内位置靠上话语权比较大的游戏媒体——他们在测评游戏时给出的流程是：在内部筛选最适合某类游戏且当下仍有档期的编辑，由他们在内部其他编辑的帮助下完成对于这个游戏的评测，给出评分，在内部经过审核与排期，最终以文章 + 视频的形式先后放到公共平台上。但是，这个时候总会有个但是。虽然这样的流程确实在内部行之有效，为行业贡献了一批质量很高的评测，尤其是在神作频出神仙打架的那些年里——但同样不可否认的是，无论你之前做对过做好过多少次，一旦出现几次测评结果与大众预期不相符，甚至出现有失水准的内容，就很容易在玩家社群中丧失公信力，变成 IG8——《宝可梦欧米伽红宝石/阿尔法蓝宝石》的 too much water 扣分，《超级马里奥 3D 世界+ 狂怒世界》的 7.0，《死亡搁浅》的 6.8——你知道的，日光之下从来都没有什么新事。而且当我们把视角继续向下探时，会发现这样的问题不仅仅是发生在 IGN 一家，Polygon 的《DOOM(2016)》、VentureBeat 的《茶杯头》（他的操作已经不足以让他完成第一关，只能借其他同事存档继续评——尽管他据自己说已有着超过二三十年的评测历史）……所以到底发生了什么？当然，抛开社区中已有的被讨论过无数次的，对评测者的质疑不谈，我们也许可以从更加细致的视角去看这件事。就像你在做数码类评测会提到机型的性能与续航、外观、系统流畅度与体验、拍照、外放表现，但你不可能在非首发上手的评测里一样面面都要点到，而是根据个人的使用习惯有自己的偏好一样，IGN 的每个评测者在评测游戏时也会有自己的明显倾向，让我们以当年评测《死亡搁浅》的 Tristan Ogilvie 为例——他评测一款游戏的重心是动作系统、交互反馈与流程的顺畅程度。在我眼里，相较于小岛在前作《合金装备 5：幻痛》中用一系列主线支线任务，吸引玩家不断的从母基地回到开放世界，完成一串任务后撤离不同，本作让角色更加持久更加连续更加长时间的停留在这个世界里，将任务可视化，变为地图上要到达的目标点，并在沿路设置一系列需要潜行的敌人据点、风景奇观，并且可以通过基建加速到达目标点跳过一部分游戏内容来优化体验的开放世界而很不幸的是，本作正正好在她关注的几个角度精准的踩中了地雷：虽然相较于《合金装备 4》有所收敛，但游戏内不断插入冗长的过场 cg，以至于从室外场景进入到地图传送需要连续跳过数个动画；游戏流程放速过缓，且游戏内的任务设计被认为千篇一律，主线和支线均由送货构成；动作系统相较于前作乐趣有所减少（尽管二者的游戏定位有着根本性的不同，也许从来就不该这么要求角色的性能），击败 bt 后的奖励反馈也较为缺乏，以至于编辑出于他所偏向的几个维度的也许从传统游戏性的角度合乎情理的考虑，选择给这个游戏 6.8 分。随后我们都知道发生了什么——之后的几年让游戏的异步社交大放异彩，以一个非常奇怪且迅速的方式，让小岛秀夫这种几乎像是寻求争议，找骂、试图以先锋艺术电影的形式在游戏内进行的表达，能被更多人看见，被更多人共情，似乎小岛秀夫在《合金装备 2》后，再一次，等到了时代来验证他思想上的前瞻性。因而，时代背景变迁下大家对小岛独特表达的广泛接受、小岛秀夫的明星制作人光环以及评测中确实存在的一些过当的指责，都让玩家社群反过来开始质问一件事：为什么我觉得好玩的游戏，你却要给出这么低的分？当然，IGN 作为行业头部媒体，其本身在游戏界有较大的声量较高的影响力，且也有着个人做自媒体乃至中小型媒体难以企及的资源，因此它说的话做的评测理应被玩家更加严格的审视，树大招风不无道理——而且至少在我举得几个例子里，引发争议的游戏本身也是在不同领域内质量过硬的存在，只是随着营销的不断发展，疫情时代与后疫情时代每个亚文化小领域的的扩圈与自媒体时代的到来，这种原本对媒体对权威的的尚处合理范围的不信任将逐步扩大，直到这种现象一步步扩散到每一件事，每一个领域，乃至我们每个人都难以避嫌，难以无视。Part2：内容井喷的时代，同质化的幽灵与信任的坍塌数码圈从来不缺学生出身，单纯出于热爱，分享自己体验的自媒体，这样的评测形式本身并非没有价值，它用更加平易近人的角度去涵盖一个产品中用数据评测所难以涵盖的部分，尽管也许，分享自己对于刚入手的数码产品的主观体验并没有什么门槛——有着中专及以上的学历就有足够的语言组织能力去清晰讲述自己的感受，再来一个相机甚至是手机去拍摄素材就能实现，但撰写一篇好的主观体验评测文（不管是哪个领域）就像是一般人眼中的历史研究一样，都觉得准入门槛低，都觉得写的内容是任人装点任人阐释的小姑娘，但真的想要尽可能做到内容详实、条理清晰，往往需要长时间的写作经验积累与大量的训练，而如果想在这个基础上再去追求与社区中主要讨论的内容截然不同的全新角度，或者让我们思考的更加狂野一些，在这个内容上做出深度，探讨一个产品背后更加深刻的有广泛意义的命题，则更加考验个人的水平——但随着互联网的高速发展与信息的高速创作、流通，我们正处在一个内容极其过剩的时代。虽然我承认，我是随着酷市场转型为酷安而踏入数码圈的，本身并不算什么数码区老资历，但你如果要我去回忆在我进入数码圈后这有限的时间内，主观体验向的数码自媒体井喷的那段时间，我想大概是在「老师好我叫何同学」的 5G 视频兴起的 2019 年 6 月前后，以及疫情肆虐的 2021 年后到 2023 年末。前者因为正好切合了国家的 5G 发展战略的需要而被主流媒体报道，被社会广泛关注，因而给圈子引来了泼天的流量——事实上那段时间坚持做手机体验类评测的规模还比较小的自媒体，如今都已经成长为几十万粉丝的台柱，比如我当时看的比较多的花生说；而后者则是由于人们长时间居家，在室外的娱乐方式因客观需要而受限，因而人们开始转向互联网，大约是在这一时期前后，一群赛道更加细分，比如上古数码产品体验、软件交互等领域（就我当时看的比较多的来说是六分超超和苏星河）的体验向自媒体开始大规模兴起，游戏区的很多如今声量较大的自媒体编辑老师也多是在这一时期进入相关行业或者是投身做自己的账号。但与自媒体账号井喷相对应的，却是——请让我想想，该怎么让人更好接受的去讲一个有点不近人情的结论——极其迅速的内容同质化。游戏不是一个竞争激烈且产品间带来的体验差别相较于同类型的数码产品更大的存量市场，有足够多各自差别足够明显的新游戏不断冲刷着视野，因此这种同质化的倾向也许还并不明显，但在数码区，当每个上游的厂商都在绞尽脑汁的结合热门概念，生生在软硬件体验上捏出极其有限的新创新点时，这种感觉就尤其明显——对于自媒体而言，相比起「你自己要讲述什么故事」，你的阐述对其他人有什么价值才更有传播的意义，因而当你迅速的因不受束缚、观点尖锐，获得了一批受众支持迅速爆火后，同质化的幽灵便也开始悄然显现——是顺着原先让你得到人们簇拥的路径一路向西，对着即使没有那么差的内容输出情绪，直到有一刻在无数次路径依赖后判断失误，犯了众怒，还是妥协于平台激励有限而选择靠近厂商，向其他市面已有的无数自媒体靠拢，在厂商给的通稿框架与宣传的大方向已然确定的情况下，进行着极其容易趋同的表达，在首发宣发节奏与内容合作模式影响下，选择性的淡化掉体验中的痛点，即使在前文有提及也往往占据相对较小的篇幅，并随后立马引入一个「但是」，用着首发特调的机型说着好到梦幻的体验，最终在观众的不解中继续在内容理想与工作室运营中艰难的走着钢丝？当然，如果把话说到太绝，往往会有很多偏颇，所以让我们稍微说一些让文章作者更拟人的

「智能队列」如何改善你的家庭网络质量主作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org联合作者关注Blanboom少数派作者个人博客：https://blanboom.orgBlanboom关注Blanboom少数派作者个人博客：https://blanboom.org2021/01/31 23:12昨天 12:57更新：2026 年世界电信和信息社会日重推在无线路由器的设置页面上，我们可能会看到一个名为「智能队列」的功能（根据品牌或操作系统的不同，也可能会被称做「智能 QoS」、「SQM」等）。这一功能是如何工作的？是否能让网速变快？是否需要打开？本文将尝试解答这些问题。从「网络如何连接」说起本章节主要介绍计算机网络的基础知识。如果读者已经熟悉相关知识，可跳过这部分内容。在介绍「智能队列」之前，先来了解一下基础的网络知识。两台电脑间进行数据传输，存在两种方式，分别是电路交换和分组交换。对于电路交换，两台电脑独占一条线路进行数据传输。过去的固定电话，大多使用的就是电路交换。由于独占一条线路，两台设备之间的通信较为稳定，很难出现数据丢失、数据传输顺序错乱等情况。但一条线路只能供一组电脑使用，另外一组电脑想要通信，就不能复用原来的线路了。对于分组交换，传输的数据被拆分为小段，每一小段叫做一个数据包（报文）。数据以报文的形式进行转发，从而实现了多组电脑复用一条线路进行通信。电路交换虽然能保证数据传输的稳定性，但由于独占线路浪费太多资源，所以在 Internet 中，使用的是分组交换。分组交换将数据拆分成一个一个的小片段来转发，没有独占线路。如果网络条件不好（例如同时有大量报文经过路由器，超出了路由器的处理能力），很有可能遇到数据顺序错乱、甚至数据丢失等情况。那么，这些情况应该如何避免？在微信视频通话、电视直播等应用中，报文的丢失或乱序，只会造成视频图像卡顿，不会造成太严重的影响。所以这些应用会使用简单的UDP协议，不考虑丢包、乱序。在网页浏览、文件下载等应用中，一般使用基于TCP的HTTP协议。TCP 在报文中携带序号，避免乱序；同时提供了报文的丢失重传机制，避免了丢包。所以，也可以认为 TCP 为两台电脑的应用程序建立了一个虚拟的「连接」。（最新版本的 HTTP/3 协议，使用的则是QUIC。QUIC 虽然基于 UDP，但引入了与 TCP 类似的报文丢失重传机制）如果想要了解更多网络基础知识，推荐阅读「入网指南」系列文章：入网指南 01 | 一文读懂你身边的「网络」入网指南 02 | 当你发消息的时候，发生了什么入网指南 03 | 交换机在交换什么？入网指南 04 | IP 地址大揭秘入网指南 05 | 除了接线，还有什么影响着我们网上冲浪入网指南 06 | 连上 Wi-Fi 就能上网了吗？网络延迟：影响上网体验的重要因素带宽和延迟，是影响上网体验的两大因素。带宽指的是在单位时间内，网络最多能传输多少数据。网络的带宽和水管的粗细有着很多相似之处，水管越粗，相同时间流出的水越多；同样地，网络带宽越大，相同时间能够传输的数据越多。在日常上网过程中，大的带宽能带来这些好处：上传、下载文件的速度更快在线看视频时，缓冲的时间更短如果家中有多个人同时上网，大的带宽能够保证每个人都能流畅上网，相互的影响更小对于家庭宽带，我们常常听到的 10 兆、50 兆、100 兆、千兆，就是指的带宽的大小。延迟指的是数据从网络的一端到达另一端，所需要的时间。过大的延迟也会对上网体验带来不好的影响，例如：在浏览器中输入网址敲回车后，需要等更长的时间，网页才能加载；视频聊天、语音通话质量变差，甚至能直接感受到声音、画面的滞后；玩在线游戏时，发现游戏变慢变卡。在很多时候，延迟相比带宽更容易被人感知，对网络体验的影响更大。是什么导致了网络延迟？上文中介绍过，Internet 使用的是分组交换。如果大家阅读过计算机网络相关的书籍，应该对分组交换的「存储-转发」机制有一定的印象。简单说，「存储-转发」是指网络设备将收到的报文存储在缓冲区，然后再发送。这是由于网络设备的性能有限，转发报文有速率限制。如果设备短时间内收到大量报文，一时半会儿处理不过来，就可以先报文存起来，空闲时慢慢发送。通过「存储-转发」，能够提高网络设备的吞吐量，更充分地利用带宽。其中，存储报文缓冲区，是一个先进先出的队列。那么，队列的长度应该如何确定？是不是队列越长，能够存储的报文越多，网络的性能就越好？大家可以先思考下这个问题。其实在过去，内存价格比较昂贵，考虑到成本，网络设备的队列不会太大。现在，随着内存价格的下降，网络设备拥有了更大的队列。但是，网络的带宽是有限的。实际的可用带宽，取决于数据传输路径上的最小带宽。这个最小带宽也被称做带宽上的「瓶颈」（bottleneck）。由于瓶颈的存在，如果流量长时间以较大速率发送，大的队列并不能显著提升吞吐量。相反，大量报文滞留在队列中，会带来延迟的增大。这种队列中缓存有大量报文，来不及发送，导致数据传输延迟增大的现象，被称做 bufferbloat（缓冲区膨胀）。DSLReport 提供的测速服务，能够检测网络的 bufferbloat 程度。不过由于没有中国大陆服务器，测试结果不够准确。国外的读者可以尝试一下：DSLReports Speed Test家用无线路由器中的 QoS 功能在「智能队列」出现之前，很多家用路由器已经提供了 QoS 功能。那么，传统的 QoS 如何解决网络延迟、网络拥堵？存在哪些不足之处？在了解这个问题之前，先来熟悉一下网络中的报文，是什么样子的：在网络中传输的报文，一般包括了链路层、网络层、传输层、应用层这四层信息。其中，网络层的源 IP 地址、目的 IP 地址、协议号，和传输层的源端口号、目的端口号，是 QoS 算法中常常会关注的五个信息。这些信息合称「五元组」。同一个应用在同一时间段内，传输的报文中，这些信息一般是不会发生变化的。例如观看在线视频时，同一个视频的所有报文 ，一般拥有相同的五元组信息。所以，通过五元组，我们能够区分不同应用的流量。特别地，一些常见的应用，会使用固定的端口号。例如网页浏览一般使用 80、443 端口；英雄联盟游戏，使用的是 5000-5500 端口 ……所以，当我们想要提升英雄联盟游戏的体验时，在路由器上设置让 5000-5500 端口的报文优先通过即可：How to set up traffic prioritization rules for League of Legends on Nighthawk Pro Gaming router | Answer | NETGEAR Support家用路由器中常见的 QoS 功能，就是通过手动设置 IP 地址、协议号、端口号等报文特征来实现的。这种方式需要用户了解一定的网络知识，且操作复杂。（虽然部分无线路由器内置了常用应用的 QoS 规则，但仍需要一定的手动设置，且无法及时适配新的、小众的应用）当然，除了上文中介绍的内容，QoS 还有多种不同的实现方式。例如：对特定应用（例如 BitTorrent）进行限速按照 MAC 地址指定某台设备的优先级（例如提高 Nintendo Switch 的优先级，使游戏更加流畅）通过WMM、DSCP等机制来决定报文优先级这些机制各有各的特色和不足。例如对应用程序限速的方式，虽然效果明显，但无法在网络空闲时，动态利用剩余带宽来提升传输速度，也难以适应带宽频繁变化的无线网络。关于上述 QoS 机制的技术细节，本文不再详细介绍。如果感兴趣，可以搜索相关关键词，进一步了解。智能队列：通过「全自动」的方式优化网络、降低延迟什么是「智能队列」目前，一部分新的无线路由器，开始有了名为「智能队列」的功能。这一功能一般指的是 fq_codel。fq_codel 包含两种队列管理机制，分别是 Flow Queue 和 CoDel。在下文中，我将简要介绍这两种机制的作用。Flow Queue前文中介绍过，网络中的报文一般都有自己的五元组信息；同一应用发送的报文，在一定的时间段内，五元组信息是相同的。这些信息相同的一系列报文，被称做一条流（flow）。Flow Queue 为每条流单独维护一个队列。例如，如果我们一边在线观看视频，一边进行语音通话，一边玩游戏，Flow Queue 就会为视频播放的流量、语音通话的流量、在线游戏的流量各维护一条队列：正常情况下，在线观看视频需要占用较大的带宽，但对延迟不敏感，用户可以容忍几秒甚至几十秒的缓冲时间。而语音通话和在线游戏占用带宽小，但对延迟敏感。例如语音通话时，如果一个人说了一句话，对方在几秒后才能听到，用户就能明显感知到延迟。如果三种应用共用一个队列，队列中会充满大量视频播放应用的报文，导致语音通话、在线游戏的报文迟迟无法发送，增大了这两种应用的延迟。Flow Queue 为视频播放、语音通话、在线游戏各维护一条队列，并公平地调度三个队列，避免视频播放的报文占用太多时间，从而让在线游戏、语音通话的报文得到及时处理。由于 Flow Queue 确保了应用程序之间的「公平」（fair），所以又被称做 Fair Queue。CoDel我们已经知道，适当长度的队列，能够充分利用带宽、提高吞吐量。但过长的队列，又会导致过大的延迟。也就是说，队列既不能过长、也不能过短。CoDel 所做的，就是让队列维持在合适长度，最大限度地提升网络性能。CoDel 通过计算报文在队列中的停留时间，来判断队列是否过长。具体做法如下：报文进入队列时，在报文上记录当前时间（也就是报文入队时间）报文出队列时，再次获取当前时间（报文出队时间）计算报文入队时间与出队时间的差值，即可得到报文在队列中停留的时间当报文在队列中的停留时间超过阈值，说明队列过长，需要缩小队列长度当队列过长时，CoDel 会直接丢弃出队列的报文，从而减少队列中报文的数量，达到控制队列长度的效果。（这种对报文的丢弃，并不会对网络带来负面影响，具体请见后续章节）可以看出，CoDel 根据队列中报文的停留时间，来决定是否丢弃报文。而不是「一刀切」地将队列限制在一个固定的、较小的长度。这样既能降低延迟，又保留了队列最大的优点：吸收短暂的大流量报文、提升吞吐量。另外，我们的设备，尤其是手机等移动设备，所处的网络环境是复杂多变的。例如手机靠近或远离无线路由器，就有可能导致网络带宽的变化。CoDel 通过计算报文停留时间的方式控制队列，能够更加灵活地适应网络带宽、网络质量的变化。fq_codelFlow Queue 确保了不同应用程序的流量能够公平传输。CoDel 控制队列长度，实现了整体延迟的降低。fq_codel 就是将 Flow Queue 和 CoDel 结合在一起使用的机制。其中，Flow Queue 先将不同的流划分为不同的队列，再为每一个队列使用 CoDel，控制报文的延迟。两者结合使用，共同提升网络体验。fq_codel 具有免配置、天然适应不同的网络环境、实现简单、消耗资源少等特色，越来越多的无线路由器（以及开源的路由器操作系统），开始内置这一机制。参考 KoolShare 上的这篇帖子，作者经过测试，发现打开 fq_codel 之后，在 BT 下载占满上行带宽时，在线游戏依旧保持较低的延迟：可能是家庭路由里最好的 QoS——在线游戏者的福音 - KoolShare（见 8 楼、10 楼、22 楼、25 楼的测试结果）是否需要打开「智能队列」？前面介绍了 fq_codel 的一系列优点。那么，fq_codel 有没有缺点？是否需要我们打开？我的设备 CPU 性能不强，是否可以打开「智能队列」？根据这篇帖子，ER-X路由器在打开 fq_codel 时，最大支持 200M 左右的带宽。QoS (fq_codel/cake) Hardware Recommendation : HomeNetworkingER

5月18日，国家统计局新闻发言人、国民经济综合统计司副司长王冠华在国新办新闻发布会上表示，4月份的部分指标增速有所回落，属于月度之间的正常波动，综合累计指标、宏观指标还有一些结构性指标来看，中国经济“稳”的主基调没有变。（澎湃）

人工智能训练和推理集群需要比传统云基础设施更密集的互联架构，这让用于数据传输的光纤陷入供不应求之中。据Rebio Group估计，今年北美光纤需求预计将增长22%至25%，而供应增长则仅有12%至19%。另据Data Center Dynamics报告称，大批量买家的交货周期已延长至20周，而小批量买家的交货周期甚至长达一年。（财联社）

5月17日，外卖骑士迎来了自己的节日——这是淘宝闪购推动骑士职业升级、推出城市骑士“橙意计划”以来，举办的首届城市骑士节。骑士节上，淘宝闪购宣布启动橙意计划2.0，并将推出更多重磅举措：淘宝闪购联合阿里集团蚂蚁集团旗下业务品牌、爱心商家、公益机构和社会力量，一起成立“橙益联盟”，共同加码更多权益，共建温暖可持续生态；进一步夯实职业保障；围绕“衣食住行用”全面升级；推出面向残障骑士和女骑士的专项保障，进行派单优化、相关订单自主选择。

36氪获悉，天眼查App显示，近日，杭州宇树星盟科技有限公司成立，法定代表人为周昌慧，注册资本10万，经营范围包括智能机器人的研发、智能机器人销售、软件销售等，由宇树科技股份有限公司全资持股。

5月18日，国家统计局新闻发言人、总经济师、国民经济综合统计司司长付凌晖在国新办新闻发布会上表示，文旅热赛事热成为服务消费发展的新亮点，相关服务零售额均保持了较快的增长。数据显示，1—4月份旅游咨询租赁服务，交通出行服务，文体休闲服务类零售额均保持了两位数增长，明显快于服务零售额的增速。（证券时报）

48022桌需退款，这家网红店出了什么事？36氪的朋友们·2026年05月17日 09:29昂贵的学费。“由于您吃到的只是‘烤熟的羊肉串’，而不是‘最佳状态的羊肉串’，我们内心感到惭愧，所以原路‘退回首轮烤制一半羊肉串的钱’作为补偿。”近日，不少网友在社交平台发文称，收到了“很久以前羊肉串”的部分退款及一条道歉短信。图源：很久以前官方账号据窄门餐眼数据，“很久以前羊肉串”开有142家门店，其中九成为商场店。羊肉串是其核心单品，门店特色是统一使用自动旋转电烤炉，消费者可自行动手翻面、完成烤制。然而，此次退款事件的问题恰恰出在电烤炉上。据有意思报告了解，“很久以前羊肉串”此前采用人工目测的方式判断烤炉温度，加热板变红即可烤制。但从2025年3月起，为了标准化流程，门店引入了电子测温枪，烤炉预热温度达到530°视为温度达标。但据“很久以前羊肉串”在官方账号发布的声明透露，在实际使用过程中，门店过度依赖测温枪数值，忽视了门店电压波动导致功率不稳，会影响烤炉内部实际温度的稳定性。结果就是原本应该高温快烤、做到外焦里嫩、汁水丰盈的羊肉串，由于烤制时间拉长、水分无法锁住，变成了只是烤熟了而已，肉质受到影响。一位“很久以前羊肉串”店员告诉有意思报告，电子测温枪的测量方式也有问题，“测量的可能只是烤炉的表面温度，温度是达标的，但烤炉内部的空气温度却还没够，所以烤炉预热也是不足的。”“很久以前羊肉串”统计发现，涉及的门店共计24家，受影响的顾客共计48022桌，预估退款为110余万元。值得注意的是，此次退款行为其实是“很久以前羊肉串”主动为之，此前并没有大规模的消费者负面吐槽和反馈。在社交平台上，网友对此次退款事件的评价不一。有经常复购的消费者说“很久以前又在想办法退款了”，还有网友认为“这就是一场搞噱头的营销”。对此，资深公关专家姚素馨认为，将退款事件加以突出，对内能反向起到警示作用，加强内部管理；对外也能给品牌树立起一个重视品质的形象，获得一定曝光量。“甚至有些时候即使负面舆情并不大，但品牌依旧判断是有风险的，就会做‘预防性召回’的动作，羊肉串无法召回，只能通过部分退款的方式了”，姚素馨说。然而，退款事件也暴露出“很久以前羊肉串”在规模扩张中的品控问题。食品行业分析师林岳分析：“由于电压、测量方式等问题，门店靠工具标准化会存在误差，变量需要员工靠经验来把握。此次退款涉及24家门店，且时间跨度半年，说明是系统性的品控漏洞，暴露出品牌过度依赖工具，在快速扩张中存在管理混乱。”上述店员表示，当前，门店已经不再使用电子测温枪。总部对门店进行了统一培训，要求门店首波客人的烤炉预热时间从原本的7—8分钟，延长至10—12分钟，并且店员需要手动测温，手在炉面上方试温，若感觉很烫，无法停留，才是温度达标。但这意味着，“很久以前羊肉串”又回到了原本人工测温的方式，将标准化流程依托于员工的判断。而在“很久以前羊肉串”身上暴露出的规模与口味的割裂，其实是烧烤行业的共性痛点。图源：红餐大数据据红餐大数据，2025年全国烧烤市场规模已达到2680亿元，同比增长4.3%，占据餐饮大盘约5%的市场份额。截至今年2月，全国烧烤门店数量为40.9万家，但连锁化率仅为22%以上，门店达到100家以上的连锁烧烤品牌在行业中仅占1%。烧烤行业看似门槛低，实则一直在面临开得“铺天盖地”后，要如何实现口味标准化的问题。经营两家烧烤小店的陈明告诉有意思报告，他开店二十多年，拥有稳定的客流靠的正是不变的口味。他认为，烧烤的灵魂是火候，需要根据食材来调整火候的大小与时间，再根据手感搭配好调料，这些都需要依靠自己的经验。正因如此，他多年来都未开更多门店，所有烤串都是由他和家人烤制完成。林岳也补充，烧烤品类规模化的核心难点就在于标准化基础薄弱，烤串从穿串、腌制到烤制，各环节对人工依赖度极高，导致产品口味和品质稳定性弱。而烧烤品牌想要实现标准化，往往会借助设备实现。大圣烧烤创始人候磊曾提到，大圣烧烤在AI智能烧烤机的投入颇多，想要通过它实现自动控温、自动翻转、自动喷料，保证烤制全程仅靠一人即可完成。但“很久以前羊肉串”的退款事件无疑是一次警示，餐饮品牌在形成标准化的模式之前，需要不断对管理能力和操作环节进行复盘和优化，避免跑偏，对工具产生过度依赖，而产生新的品控问题。不知道交了100多万学费后，“很久以前羊肉串”能不能让消费者一直吃到口味稳定的羊肉串。作者：高越编辑：安迪值班编辑：甜豆本文来自微信公众号“有意思报告”，作者：高越，36氪经授权发布。本文由「36氪的朋友们」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：采访供图、作者拍摄

史上最大IPO来袭，SpaceX加速驶入纳斯达克独角兽早知道·2026年05月17日 12:22在AI与航天融合的新纪元，资本市场愿意为“火星愿景”给出多高的定价。SpaceX5月16日，SpaceX向股东发出内部邮件，宣布普通股“一拆五”的拆股方案已获批并生效，所有C类普通股将同步转换为A类普通股。美股IPO市场正迎来今年最重磅的选手，埃隆·马斯克旗下的太空探索巨头SpaceX，正以史无前例的体量加速驶入纳斯达克。综合 | 路透社 WSJ 中国基金报编辑 | Echo本文仅为信息交流之用，不构成任何交易建议据路透社和华尔街日报援引知情人士报道，SpaceX正大幅加快IPO进程。原定于6月下旬（马斯克生日前后）的上市计划，现已调整为最早于下周三（5月20日）公开招股说明书，6月4日启动路演，6月11日确定发行价格，最早于6月12日正式挂牌交易，股票代码为“SPCX”。知情人士透露，美国证券交易委员会（SEC）对公司IPO文件的审查进度快于预期，是此次上市时间表提前的主要原因。“SPCX”此前由Tuttle Capital Management专注于SPAC的ETF使用，该公司于4月将该ETF代码改为SPCK后，“SPCX”随即空出，市场当时已开始猜测SpaceX可能会选择这一代码。在IPO前夕，SpaceX完成了一项关键资本运作。当地时间5月15日，公司通过内部邮件通知股东，董事会提出并经多数股东批准的普通股“一拆五”拆股方案，以及所有C类普通股转换为A类普通股的计划，已于当地时间5月4日起正式自动生效。SpaceX在通知中强调，此次拆股旨在成比例调整每股价格并增加股东持有的股票数量，不会改变公司的整体价值，也不会改变股东所持股份的总价值。受此操作影响，公司每股当前公允市场价值已从526.59美元调整为约105.32美元。拆股后单股价格大幅下降，为更多散户投资者参与IPO打开了通道。公司还特别加大了全球散户配售力度，计划在英国、欧盟、澳大利亚、加拿大、日本和韩国等多个市场向散户投资者发售股票，并将于6月11日邀请约1500名散户投资者参加专项活动。此举有望使SpaceX成为史上个人投资者参与度最高的IPO。知情人士透露，SpaceX此次IPO很可能将融资目标定为约700亿至750亿美元，公司估值约1.75万亿美元。部分报道甚至将估值上限推高至2万亿美元。这一估值较年初有大幅跃升。今年2月，SpaceX与马斯克旗下人工智能公司xAI合并时，双方设定的总估值仅为1.25万亿美元。短短三个月内估值跳涨约40%，背后是华尔街对AI概念与低轨卫星互联网的追逐。据PitchBook估算，星链2025年EBITDA约58亿美元，对应利润率约54%，盈利结构已趋近软件服务企业。高盛交易部门Tony Pasquariello认为，虽然市场担心大型IPO潮可能冲击美股流动性，但目前美股总市值高达77万亿美元，而2026年高盛预计美国IPO数量仅约100家，远低于1999年互联网泡沫时期的380家，市场完全有能力消化。他同时指出，相比1999年大量质量参差不齐的公司扎堆上市，如今市场更关注资产质量，而SpaceX显然属于极少数具备长期叙事与核心技术壁垒的超级资产。若成功落地，融资规模将远超沙特阿美2019年创下的约294亿美元纪录，刷新全球历史新高。上市的另一大看点指向治理与薪酬。据路透社援引保密招股书报道，SpaceX计划发行10倍投票权的B类股，确保马斯克拥有不可被解雇的绝对控制权，上市后将很可能采用双层股权结构。SpaceX与此同时，董事会已批准一项总额可达万亿级别的薪酬方案，直接与火星殖民及太空算力两项指标挂钩：火星殖民目标：当SpaceX市值达到7.5万亿美元，且在火星建立常住人口至少100万人的人类永久定居点时，马斯克将获授2亿股超级投票权限制性股票。太空算力目标：若公司达成另一项估值目标，并在太空建成可提供至少100太瓦算力的数据中心，他将再获授6040万股限制性股票。马斯克对此回应称：“我需要确保SpaceX始终专注于让生命成为多行星化、并将人类意识延伸至星辰大海的愿景，而不是去迎合某些人极其无聊的季度财报奖金。”随着招股书的公开，SpaceX的经营基本面将首次完整呈现在市场面前。星链业务已成为SpaceX的绝对营收支柱。据公开数据，截至2026年3月，星链已累计发射超10000颗在轨卫星，全球用户突破1000万，2025年总营收突破100亿美元。马斯克坦言，“SpaceX收入的绝大部分来自商业星链系统”。然而，整体盈利形势并不乐观。SpaceX在2025年全年营收超185亿美元，同期亏损近50亿美元，这一从未公开的数据首次将xAI的业绩纳入计算。自2023年成立以来，xAI一直以惊人的速度消耗现金，主要用于购买高端芯片和建设数据中心来运行和训练AI模型。目前，xAI还面临着核心创始人离职与团队重组的问题，而公司下属计算团队的训练性能也被内部评估为“低得令人尴尬”，正在加紧追赶。摩根士丹利、美国银行、花旗、摩根大通和高盛担任此次发行的主账簿管理人，另有16家银行在机构、零售和国际渠道中担任角色。在基本业务之外，SpaceX正在构建更长远的叙事。今年1月，SpaceX向美国联邦通信委员会提交了一份极具野心的申请：计划部署最多100万颗卫星，组成一个环绕地球的轨道数据中心网络，旨在为先进AI模型提供前所未有的计算能力。同时，SpaceX IPO文件披露的时间点，可能与“星舰”第12次试飞窗口期重合。市场预计，该次试飞最早将于下周二进行，这无疑将为IPO增添话题性。对于市场而言，SpaceX的上市远不止是一场融资活动。它将测试一个核心命题：在AI与航天融合的新纪元，资本市场愿意为“火星愿景”给出多高的定价。而从Brookfield在IPO前已累积约20亿美元股份的布局来看，机构投资者对这一赌注已经投下了真金白银的信任票。注：本文素材来源于互联网公开渠道，如有侵权请联系删除。内容所述仅代表个人观点，不作为指导依据，据此操作风险自担！本文来自微信公众号“独角兽早知道”，作者：iponews工作室，36氪经授权发布。本文由「独角兽早知道」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、采访供图

6.4k Stars！用Claude Code写论文的全套流水线，有人打包开源了量子位·2026年05月17日 12:594个skill，跑通整套科研流程听雨 发自 凹非寺量子位 | 公众号 QbitAI用Claude Code写论文的一整套流水线，有人打包开源出来了。完全戳中了学生党的痛点，github星标直达6.4k。academic-research-skills项目名叫academic-research-skills（以下简称ARS），是一套Claude Code技能包。里面涵盖4个skill，分别对应论文的研究、写作、审稿、定稿。只需两行命令安装，直接一条龙串起整套学术研究流水线。academic-research-skills只能说，我读研的时候怎么没碰到这种好东西呢…示意图4个skill，跑通整套科研流程ARS的核心架构由4个skill组成，它们各司其职，拼在一起就是一条从选题到交稿的完整链路。我这里还做了图，大家可以看得比较直观：△Deep Research是一支13个Agent的研究团队。它负责文献调研、研究问题构建、方法论设计，还能写系统性的PRISMA综述。团队里有专门做文献溯源的Agent，会调用Semantic Scholar API验证每一篇引用的真实性。有苏格拉底导师Agent，通过对话引导研究者理清思路。还有魔鬼代言人Agent，专门挑刺，防止研究者在早期就陷入思维定式。△Academic Paper是一支12个Agent的写作团队。从大纲设计、论证构建、草稿撰写，到双语摘要生成、图表可视化、引用格式转换，全流程覆盖。特别值得一提的是风格校准功能，AI会学习你过往作品的写作风格，让输出更像你自己写的，而不是千篇一律的AI味。输出格式支持Markdown、DOCX、LaTeX，最终可以编译成APA 7.0或IEEE格式的PDF。△Academic Paper Reviewer是一支7个Agent的审稿团队。模拟真实学术期刊的评审流程，由主编EIC带领三位领域审稿人，再加上一个魔鬼代言人，从方法论、学科视角、跨学科价值等多个维度打分。评分采用0到100的量化标准，80分以上接受，65到79小修，50到64大修，50以下拒稿。审稿团队还会输出详细的修改路线图，告诉作者下一步该做什么。△Academic Pipeline是流程编排器，把前面三个团队串联成一条10阶段的流水线。从研究、写作、完整性检查、同行评审、修订、最终检查，到发表准备和流程总结，每个阶段都有明确的产物和检查点。你可以在任意阶段插入，比如已经有了初稿，就从Stage 2.5的完整性检查开始；收到了审稿意见，直接从Stage 4的修订切入。费用参考也很透明，一篇1.5万字的论文，全程跑下来大约4到6美元。△比较有意思的设计用Claude Code做学术研究的开源项目已经很多了，但是深扒之后，我发现ARS在底层设计上还是有些过人之处。可以简单总结为一句话：系统性防止AI搞砸学术研究。第一，引用核验。AI写论文最忌讳的，就是幻觉引用。不只是编造不存在的文章，还包括标题相似但作者年份全错、DOI真实但内容对不上等更隐蔽的情况。ARS在Deep Research阶段就埋了一个引用核验机制，每一篇文献都要过Semantic Scholar API的存在性确认。不是简单查一下标题对不对，而是用Levenshtein相似度算法做模糊匹配，阈值设在0.70以上才算通过。△第二，完整性闸门。在流水线的Stage 2.5和Stage 4.5，有两道不可跳过的完整性闸门，会运行一份7项AI失败模式检查清单。这份清单直接来自2026年Nature上发表的一项全自主AI科研研究，其中总结了7种翻车模式，覆盖引用幻觉、数据捏造、方法论造假等情形。7种翻车模式任何在2.5被标记为SUSPECTED的问题，必须在4.5变成CLEAR，或者由人工手动覆盖并留下记录。设计逻辑是：把「我相信AI不会出错」变成「我要求AI证明它没出错」。实测中，这套机制在一篇真实论文里抓到了15个伪造引用和3个统计错误。第三，反谄媚协议，让AI敢于说不。大多数AI工具都有一个隐形毛病，讨好用户。你让它改，它就改，哪怕改得更差。所以ARS在审稿环节专门设计了反谄媚机制。审稿团队里有一个Devil’s Advocate，也就是魔鬼代言人，职责是挑刺。但挑完刺之后，还有一个让步阈值协议。DA的反驳会被评分1到5，如果低于4分，写作团队不允许承认。△换句话说，AI不能为了显得好合作就轻易让步。同时，攻击强度在修订过程中必须保持。如果第一轮审稿把方法论批得体无完肤，作者修订后不能让审稿人突然变得温柔。评分轨迹也会被追踪，任何维度的分数下降都会被标记为回归。这和软件工程里的不引入新Bug原则一样，改一个地方不能搞砸另一个地方。第四，三层数据隔离，不让AI偷看答案。ARS把数据流严格分成三层：Layer 1是原始输入，默认不可信，可能幻觉、过时、带偏见。Layer 2是通过完整性验证后的产物。Layer 3是评分标准、参考答案和金标数据，这层材料永远不能出现在写作AI的上下文中。具体实现上，写作团队和审稿团队分两次独立调用，中间有阶段边界隔离。写作AI只能收到审稿AI的自然语言反馈，比如「第二章论证跳跃，建议补充对比实验」。但它看不到原始的评分标准，也不知道每个维度占多少分。这个设计的灵感来自于Anthropic今年的w2s-researcher研究，其中也用了同样的三层隔离模型。结论是当AI能读取标签数据时，结果可能不是真的泛化，而是在优化表面特征。解决方案不是更好的提示词，而是结构上的隔离。△最后一点，诚实文档化，「我不保证能复现」。学术界经常遇到「这个结果我复现不了」的问题。ARS给每个产物生成一个repro_lock文件，记录运行时的完整配置。但文件里有一段强制声明，LLM输出不是字节级可复现的，模型提供商会更新权重而不改模型ID，外部API每天返回不同的数据。这个文件只是配置文档，不是重放保证。△在更新日志上，可以看到ARS已经经历了很多轮迭代。从2月上线到现在，提交的commit数达到了三百多次。从每次版本更迭中，也能看出作者对AI学术研究系统风险有着深刻理解。这也是我觉得目前学术研究AI工具的关键所在——让AI帮你写论文并不难，重点是如何防止它出错、讨好，让整个流程变得更系统更可靠。ARS的设计哲学，可以总结为README里那句话：「AI是你的副驾驶，不是飞行员。」如何安装安装方式很简单，如果你已经在用Claude Code，只需要两行命令：/plugin marketplace add Imbad0202/academic-research-skills/plugin install academic-research-skills验证安装是否成功，运行：/ars-plan然后描述你正在写的论文主题，ARS就会启动苏格拉底对话，帮你梳理论文结构。如果你偏好单条命令测试，也可以用：/ars-lit-review “你的研究主题”不过最简单的安装办法，其实是直接把SKILL.md上传到claude.ai项目知识库。不需要安装Claude Code，打开浏览器就能用。不过要注意，这种方式不支持多Agent并行，功能上是单Agent版本，适合轻度体验；想跑完整流水线还是需要Claude Code。还有一点，项目支持繁体中文和英文。那么，又到了大家最关心的，要花多少钱的环节。作者推荐使用Claude Opus 4.7搭配Max订阅计划。完整跑完10个阶段，单次可消耗超过20万输入token和10万输出token，单独使用某个子模块则少得多。Max订阅计划分两档，每月100刀或200刀，相当不便宜。但如果你的科研经费可以报销的话，那…示意图本文来自微信公众号“量子位”，作者：关注前沿科技，36氪经授权发布。本文由「量子位」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、AI生成

3个人带100个AI程序员，一个月烧掉130万美元！OpenAI：钱我出新智元·2026年05月17日 14:08如果token成本不是问题，软件开发会变成什么样。Peter Steinberger编辑：所罗门【新智元导读】3个人，100个AI agent，一个月烧掉130万美元——OpenClaw之父把软件开发变成了「AI流水线」，OpenAI替他买单。别人晒工资条，他晒账单——一个月130万美元！也就是将近900万人民币/每月。直接把网友们看傻了。OpenClaw之父Peter Steinberger在X上轻描淡写地甩出一张截图。Peter Steinberger但截图上的数字，属实不淡定：30天花费：$1,305,088.81。吞掉6030亿个token。发起760万次请求。你没看错，是130万美元。也不是哪家大厂的季度AI预算——一个3个人的小队，一个月的用量。更炸裂的是：这钱，OpenAI给报销。评论区瞬间就疯了。有人惊叹，有人质疑，有人掏出计算器开始算「这到底顶多少个程序员」。Steinberger本人谈定回应：「关闭快速模式后，我的花费低于一个工程师的成本，而且这确实帮助大得多。」翻译下就是——属实划算！更有网友惊叹于月薪40万刀的工程师——「旧金山的就业市场过于疯狂。」网友评论网友评论还有网友好奇，这个大量的Token花在了哪里。Peter回应大部分都用在了OpenClaw的开发上。网友评论云端程序员军团这件事最离谱的是，Pete的小团队只有3个人。他们在云端长期跑着约100个 Codex 实例，干软件工程里最脏、最累、最容易把人逼疯的活——审PR、找安全漏洞、去重issue、改bug、监控benchmark、发现回归后发到Discord，甚至听完会议后直接开PR。于是，AI不只是「帮你写代码」，而是进入软件协作的每一个缝隙。这就很吓人了。因为软件开发真正贵的是沟通、理解、上下文切换、审查、回归、修复、等待、重复劳动。过去一个团队每天大量时间，消耗在这些看起来不像「创造」、但没有它们项目就会烂掉的事情上。现在，Peter 把这些环节一口气丢给了一群AI agent。这是AI开始替你维护一个组织的神经系统。示意图这张截图还有一个细节很重要：它不是OpenAI后台，而是 Peter 做的 CodexBar。CodexBar 是一个 macOS 菜单栏工具，用来追踪各种AI编程工具的使用窗口、credit、成本和重置时间。它支持 Codex、Claude、Cursor、Gemini、Copilot 等一堆服务。以前程序员的菜单栏里是什么？CPU、内存、电量、网速。现在多了一个东西：token。token正在变成一种新的「生产资料」。CodexBar最后说两句130万美元一个月，3个人，100个AI agent。你细品这组数字——三个活人，带着一百个不吃饭不睡觉不提涨薪的数字员工，把一整个工程团队的活儿给干了。有人看完热血沸腾：AI终于不是只会聊天的花瓶了！也有人看完脊背发凉：等等，那我们这些写代码的，以后干嘛去？但说实话，最让我睡不着的，是Steinberger那句轻飘飘的话：「我在探索，如果token成本不是问题，软件开发会变成什么样。」Peter Steinberger各位，他说的是「如果」。可问题是，这个「如果」正在以肉眼可见的速度变成「当」。今天130万美元干的活儿，模型降一轮价，13万。再降一轮，1.3万。到那一天，100个AI agent同时干活，不再是硅谷大佬的独家游戏，而是随便一个三人创业团队的基本操作。车库里三个年轻人，手握一百个不知疲倦的AI程序员——这画面，想想就离谱。Peter Steinberger掀开了底牌。底牌上写着：未来已经在敲门了，而且它不打算等你准备好。参考资料：https://the-decoder.com/for-1-3-million-a-month-openclaw-founder-peter-steinberger-runs-100-ai-agents-that-code-review-prs-and-find-bugs/https://x.com/steipete/status/2055346265869721905https://developers.openai.com/codex/speed本文来自微信公众号“新智元”，作者：新智元，36氪经授权发布。本文由「新智元」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、采访供图

又诞生一位千亿女富豪36氪的朋友们·2026年05月17日 15:45又一部“打工妹”逆袭成为女富豪的爽文剧本。示意图自4月初以来，立讯精密股价上涨超过55%。5月11日，立讯精密市值创下历史新高，达到了5542亿人民币。立讯精密掌舵者王来春持股18.75%，身价也突破了千亿人民币，成为国内白手起家的又一位千亿女富豪。如果说4月以来的涨幅，是因为立讯精密“站在光里”，短期股价大涨或许与一则收购公告有关。5月10日，公告显示立讯精密拟通过收购京西智行集团全部股权，间接拿下京西国际约59.50%股份，有望成为其最终控股股东。这是立讯精密继两年前收购德国莱尼后，在汽车领域的又一关键资本动作。郭台铭门 徒立讯精密的故事通常被描述为“打工妹”逆袭成为女富豪的爽文剧本。实际上，只单纯地给创始人王来春贴上“打工妹”的标签，就像是只把立讯精密当做一家代工厂。与现在创业者的背景不同，王来春既不是天才少女，也不是海归精英。她出生于20世纪60年代末的潮汕地区，当时物质和思想都极度匮乏，农家女出身的她，更是过早地尝到了生存的艰辛。1988年，改革开放的风头正盛，21岁的王来春选择到最热闹的深圳求职，此时正赶上郭台铭来内地创业，于是，机缘巧合下，她成为了富士康首批150名员工之一。早期的富士康被外界称为“血汗工厂”，流水线两班倒，每天站12个小时是家常便饭，累了一天，洗澡还得等。很多人受不了这份苦，干了几个月就走了，但王来春却坚持留了下来，这一留就是十年。这十年间，王来春不仅成为了大陆员工里职位最高的女主管，还像海绵一样吸收了消费电子行业的大量经验，成长为了郭台铭最出色的学生。一个细节是，她在立讯几乎复制了富士康模式，工厂管理军事化，连车间里都挂着郭台铭的管理语录。对于“徒弟”的成功，郭台铭曾公开评价：“富士康在中国大陆创造了很多亿万富翁，王来春是最具代表性的一位！”1999年，受金融危机影响，香港很多公司都破产倒闭了。王来春看准机会，毅然辞去富士康高管职位，和哥哥王来胜共同收购了香港立讯公司。面对这一冒险举动，她的家人是反对的，其母亲甚至把家里的现金和存折都藏了起来。然而，这并非是王来春的冲动之举，在立讯早期，王来春选择依附富士康生存，“富士康开到哪，立讯就跟到哪”，专门承接其一时无法处理的订单：低毛利的，高难度的。富士康给什么，她就做什么，在定价方面也完全由富士康说了算。所以成立后的很长一段时间内，富士康始终是立讯的第一大客户，这种关系也让富士康越来越离不开立讯，因此，郭台铭兄弟还入股了立讯。2004年，王来春成立立讯精密工业股份有限公司，开始独立开发自有客户并自主研发产品，逐步向高端产品线迈进。尽管如此，到2010年，立讯精密上市时，富士康依然为立讯精密贡献了一半销售额。库克时代如果说2010年之前的立讯精密，还是一个标准的“小富士康”，还是在王来春的舒适区里创业，那么2010年之后，立讯精密的发展，则完全依赖于王来春的智慧与资本布局。因此，业内流传着一个说法—立讯精密真正的创业之路是从2011年开始的。这一年，已经上市的立讯精密有了充裕的现金，此时的王来春不再满足于代工，把赚来的钱全部砸进了更难的技术领域，一边寻求转型，一边开始“去富士康化”。也是在这一年，代工领域一个时代性的机会出现了。2011年，库克执掌苹果并推行供应链多元化，王来春意识到，新帅上任，势必调整布局、扶植自己的心腹伙伴。若能在此时跻身“果链”并获库克青睐，立讯便有望摆脱依赖外部溢出订单的处境。于是，她开启了一系列激进的“资本换技术”的并购。并购的第一枪打在了昆山联滔电子身上，立讯精密斥资11.8亿元收购了其60%的股权，此举规避了苹果复杂的考核机制，使得立讯精密顺利进入了苹果供应链。这起并购案只是一个开始，2016年，它又收购了苏州美特51%股权，靠着苏州美特在声学领域的多年积累进入到了声学领域，并在下一年获得了苹果AirPods的生产机会。据悉，当时苹果的AirPods主要由台湾英业达生产，但英业达生产效率很低，经常出现供货延迟的情况，立讯精密拿到订单后攻克了AirPods生产效率低的问题，最终成为AirPods全球最大的供应商。2020年7月，立讯精密又收购了江苏纬创、昆山纬新两家公司，正式切入iPhone代工领域，成为了中国大陆首家、全球第三家iPhone整机制造商，打破富士康、和硕的垄断格局，并于2021年首次获得iPhone 13部分订单，实现了从零部件供应商向整机组装商的关键转型。2021年，立讯精密已发展成为营收超1500亿元，苹果业务占比达七成的全球消费电子巨头。可此时，王来春却推出了三个五年计划，全方位发展消费电子、汽车、通信、工业、医疗等业务，并在汽车领域制定了Tier1业务进入全球前十的战略目标。她还提出，十年内将单一客户营收占比降至1/3以下，这个客户，就是苹果。后来有分析称，王来春之所以如此决绝是因为看到了苹果供应商欧菲光的下场，2021年3月，欧菲光被踢出苹果供应链，但由于其业务太过于依赖苹果，进而受到了重创。现在看来，立讯精密摆脱“果链依赖症”的效果也非常明显，从2021年到2025年，苹果贡献的业绩占比分别为74%、73%、75%、71%、57%。整个库克时代，立讯精密陪伴苹果走过了十五年，业务覆盖了其各个业务线，营收也从25.6亿元增长至了3323.4亿元，市值也随之增长了近40倍。而如今库克即将退休，立讯精密也找到了汽车电子、通信业务等新增长引擎。从某种程度上来说，这也是另一种圆满。公司上市十六年，她身价涨千亿再说回开篇提到的对京西国际的收购，这笔交易之所以能够直接带动立讯精密股价的上涨，与前文提到的“Tier1业务进入全球前十的战略目标”有直接关系。此前，立讯精密的汽车业务主要集中在“线束”和“连接器”领域，而京西国际的优势在悬架系统和制动系统，这笔收购直接补齐了立讯精密在底盘域执行机构的短板。更重要的是，京西国际拥有从德尔福继承的磁流变悬架和线控制动等全球稀缺技术，也拥有宝马、保时捷等欧洲顶级车企客户。要知道，底盘认证周期高达3至5年，技术壁垒极高，立讯通过这笔收购，相当于买了个时间差，直接拿到了这些高端客户的“入场券”。另外根据去年财报，汽车电子作为第二增长曲线，迎来了爆发式跃升。该板块实现营收392.55 亿元，同比大幅增长185.34%，营收占比快速提升至 11.81%，成为立讯精密增长最强劲的引擎。当然，5月11日市值能够创下新高，也离不开其四月底以来持续利好消息带来的股价拉升。4月27日，立讯精密收获了又一个涨停，最高冲到了72.6元，创下历史新高，总市值突破5200亿元。当天天风国际分析师郭明錤独家爆料，OpenAI计划自研手机，立讯精密成为独家系统联合设计和制造合作伙伴，同时，公司预告2026年上半年归母净利润78.4亿—81.06亿元，同比大增18%-22%。在此基础上，立讯精密近期披露，将持续加大铜光互连、散热及电源等核心产品的研发投入与产能布局，热门概念叠加下，立讯精密获得了资本真金白银的投票。从2010年上市时不足70亿元的市值，到如今5542亿元的巨头，立讯精密用16年完成了79倍的市值增长。而王来春的身价也从上市之日的23亿元增长到了1039亿，实现了超千亿的财富增加。然而，这并非是王来春身价的顶点。毕竟立讯精密已经向港交所递交了上市申请书，冲刺A+H上市，并预计今年下半年登陆港交所。那时，“A股果链一哥”将向全球科技制造龙头发起冲击，更重要的，其作为全球唯一同时布局消费电子、汽车与光通信产业的企业，其资本溢价将进一步释放，王来春的新故事也将开启新篇。本文来自微信公众号“投中网”，作者：张雪，36氪经授权发布。本文由「36氪的朋友们」原创出品， 转载或内容合作请点击转载说明；违规转载必究。寻求报道本文图片来自：作者拍摄、采访供图

科氪 | 对话韶音高层：开放式耳机，凭什么成为下一场颠覆TWS的“品类革命”？科氪·2026年05月17日 19:57韶音深耕开放式音频，以硬核技术推动全场景品类革命从运动孤岛到品类革命：解构韶音的“开放式”野心与底层技术硬实力在消费电子行业普遍面临增长红利见顶的今天，音频赛道却正在上演一场深刻的结构性变革。2026年5月15日下午，大雨初晴的深圳湾广场，Shokz韶音举办了备受瞩目的媒体专访会。韶音中国区CEO杨云与产品线总裁张浩锋悉数到场，向外界系统阐述了这家以“骨传导”和“运动”起家的品牌，如何在稳坐全球开放式耳机出货量第一王座的同时，向更广阔的泛场景音频帝国发起攻坚。过去两年，开放式耳机在中国蓝牙耳机市场的整体销售额份额已率先突破30%。这不仅是一次产品形态的更迭，更是一场由韶音引领、正在席卷全球的“品类级替代”风暴。一、 宏观战略与行业洞察：以“全场景品类替代”重塑全球音频版图过去，大众对韶音的认知多停留在“专业运动”与“骨传导”这两个标签上。然而，在本次专访中，韶音高层明确指出，专业运动与泛场景用户从来不是一道非此即彼的选择题。韶音正在做的事情，是将用户的定义从最初的“专业运动人群”彻底拓展到“每一个需要开放聆听的用户”。官方图片1. 舒适成为刚需：跨越品类鸿沟的终极逻辑韶音之所以做出“开放式耳机有机会取代真无线耳机”这一极具前瞻性的行业断言，其核心洞察深植于用户日常用耳习惯的根本性改变。随着线上办公、远程会议、长音频内容、网课以及日常碎片化聆听需求的激增，现代人佩戴耳机的时间正在无限拉长。在长时间佩戴的背景下，传统入耳式耳机带来的耳道压迫、闭塞闷痛、卫生隐患以及无法感知外界环境的安全痛点，逐渐演变为不可调和的刚性冲突。而开放式耳机所具备的“全天候舒适佩戴”与“安全开放聆听”的天然优势，正让它从一款特定场景的“运动副机”，跨越成为用户日常生活中无可替代的“主力设备”。在音质水平已经通过近年来的技术迭代基本解决、甚至在部分产品上（如OpenFit Pro）接近或超越传统入耳式耳机的今天，用户在舒适度与开放感知方面的倾斜，成为了开放式耳机跨越品类鸿沟的决定性力量。官方图片2. 全球化视野下的“GLOCAL”战略在出海与全球化版图中，韶音在2023年至2025年间已连续三年斩获开放式耳机市场全球出货量份额第一的桂冠()。数据显示，不仅是中国市场在狂飙突进，在全球重点市场如美国、欧洲、日本、韩国、澳大利亚等地区，开放式耳机占比持续上升、对传统入耳式耳机产生强烈替代的趋势也已不可逆转。面对全球化竞争，韶音的核心市场规划可以总结为“GLOCAL”战略——即全球化思考，本地化行动。杨云将当前的耳机品类竞争比作新能源汽车行业：全球绝大多数的核心品牌与顶尖供应链都集中在中国市场。因此，韶音选择将中国激烈的市场环境作为研发的“练兵场”，充分依托国内强大的工程师红利和极致的供应链优势，开发出在硬件、性能上足以打动并销售全球的标准标准化产品。但在软件、服务及生态融合层面，韶音则在全球范围内开展极其精准的本地化深耕。例如，最新发布的专业运动耳机OpenSwim Pro 2，便成为了全球首款支持QQ音乐无线导入与离线播放的耳机，这一功能精准击中了本土运动爱好者的核心痛点，成为了全球标准化硬件与本土顶级流媒体生态成功联姻的范例。官方图片3. 拒绝短视的长期主义护城河在涌入者众多的多元化竞争环境下，韶音认为品牌真正的护城河来源于“技术创新”与“用户思维”的底层基因，以及对长期主义的笃定。做品牌是一条最难走、面临挑战与压力最大的路。在过去长达15年的时间里，韶音只做了一件事，那就是专注于开放式音频的深耕。这种克制与长线思维同样体现在其品牌营销与用户沟通上。从跨界合作自由潜水世界冠军金明，到携手全球品牌大使张婧仪推出艺术礼盒，韶音的策略重点并非盲目追求流量快钱，而是通过这些契合品牌坚韧、优雅特质的合作伙伴，将“开放聆听、不急于求成”的理念具像化，从而在消费者心中建立起高端化、专业化且极具信任感的品牌壁垒。二、 产品迭代与技术革新：材料与声学的极限攻坚要在红海中完成对传统TWS的品类替代，单纯依靠概念营销是远远不够的，必须在产品和技术实操层面，以硬核的底层研发打破开放式耳机长期存在的物理瓶颈与行业痛点。在最新的产品矩阵中，韶音通过声学、材料学与垂直场景的深度融合，带来了行业瞩目的迭代升级。1. 5年磨一剑：以“韶音天篱滤噪技术”定义开放式新阶段在开放式耳机的普及过程中，抗噪能力弱、在嘈杂环境下听音模糊一直是阻碍其进入更广泛泛场景的致命短板()。针对这一行业痛点，韶音在行业内首次提出了“开放式滤噪”的颠覆性概念。开放与滤噪，在物理和声学原理上天生是一对巨大的矛盾。在项目启动初期，这项技术在韶音内部甚至引发了巨大的争议，许多研发人员一度认为这是无法跨越的物理鸿沟。然而，基于对技术的长期投入和“因为相信，所以看见”的技术信仰，韶音早在2020年底至2021年便前瞻性地立项了滤噪预研技术——当时行业内甚至还没有出现如今大火的耳挂式开放式耳机，研发完全是基于骨传导的未来构想展开的。历经5年多的高强度技术迭代与攻坚优化，韶音终于推出了“韶音韶音天篱滤噪技术”。该技术成功将滤噪的频段拓宽至接近4000Hz，并将滤噪深度降到了18dB，使其真正成为行业首个能够在健身房、咖啡厅、办公室等多场景完美适配的开放式滤噪方案。这标志着开放式耳机进入了一个全新阶段：既能保留开放聆听带来的环境感知与安全出行，又能兼顾沉浸专注的清晰听音体验。张浩锋更是透露，如果这项技术在未来能够继续向高铁、飞机等高难度嘈杂场景演进，那将直接触发音频行业的下一场彻底革命。官方图片2. OpenSwim Pro 2：解锁游泳蓝海的“全能数据终端”多数音频厂商将运动耳机的目光死死锁在跑步赛道，而韶音则将下一个具备决定性增量的突破口精准锁定在了游泳市场()()。根据韶音内部的全球调研数据统计，经常游泳的人群已经占到了跑步人群的60%左右，这是一个庞大却长期被行业低估的蓝海。由于游泳场景极为单调且无法在运动中聊天拍照，用户在水中对音乐激励的需求比跑步往往更加强烈和刚性。然而，水下由于蓝牙信号无法有效传输、防水等级要求严苛，市面上长期缺乏优秀的解决方案，这一领域对于高品质音频而言几乎是零竞争对手的真空地带。为此，OpenSwim Pro 2不仅将防水能力直接拉高到了极致的潜水级，更在硬件内部塞入了一颗强悍的6nm定制芯片与九轴传感器。基于对真实用户需求的深刻洞察，游泳者如果佩戴智能手表，在水中频繁抬腕观察数据会严重打断运动的节奏和划水连贯性。OpenSwim Pro 2通过这套顶级的传感器与芯片组合，能够精准识别泳姿、划次、游泳距离等核心运动指标，并通过语音实时播报的方式反馈给用户，完美做到了声学、运动科技与极端场景体验的闭环融合。3. OpenDots 2：夹式耳机的“微米级”舒适平衡近年来，耳夹式（Clip-on）形态因其潮流属性和独特的佩戴方式，正在疯狂吸纳传统的TWS用户。但耳夹式产品始终面临着一个行业内难以调和的悖论：夹得太紧耳廓容易产生压痛，夹得太松则在剧烈运动中容易松动脱落。为了达成“久戴不痛、狂甩不掉”的黄金平衡，最新亮相的OpenDots 2在完美延续了前代OpenDots ONE所奠定的“动钛弧结构”以及“9G预紧力”两大技术核心优势的基础上，在材料学与重心分布上进行了极其细腻的微米级微调。在结构优化上，OpenDots 2将内部核心的钛片宽度做到了极致压缩，最大减少了0.6毫米，这微小的尺寸缩减极大地减轻了对耳廓边缘的局部压迫感。同时，研发团队重新优化了耳机整体的重心分布，并全面提升了钛片的回弹性与抗疲劳度，从而保证了耳机在高强度、高频次拉伸使用的长期蹂躏下，依然能维持极其稳定、均衡且高性能的夹持体验。这种通过材料创新来化解佩戴稳固与舒适矛盾的研发路线，进一步筑牢了韶音在舒适科技领域的领先地位。三、 未来战略演进：在AI与智能硬件洪流中的冷思考在当前言必称AI、智能眼镜与可穿戴健康监测设备大乱斗的硬件大潮中，韶音在专访中展现出了极强的战略定力与极其清醒的“冷思考”。面对未来的技术演进，他们并没有盲目跟风，而是有着自己一套逻辑严密的方法论。1. 激进的稳健派：AI重构智能耳机的三步走节奏对于风头正劲的AI技术，韶音在内部将自己定性为“激进的稳健派”。他们深信AI在未来势必会全方位重构整个智能硬件生态，但这一进程绝非一蹴而就，而是有着极其鲜明的节奏感。在韶音的蓝图中，AI在开放式耳机上的落地被清晰地划分为三个层层递进的演进维度：第一维度是性能增强（已全面落地）：这是目前已经成熟并广泛应用的阶段。韶音通过AI算法深度优化了耳机的抗噪性能、麦克风通话降噪效果以及底层的声学音质表现。第二维度是交互体验的革命（正在切换）：随着AI大模型的普及，人类与硬件的交互媒介正在从过去的键盘、屏幕逐步切换到极其自然的语音对话。由于开放式耳机具备全天候挂耳、不隔绝环境、高频佩戴的天然属性，它被韶音视为最自然、最适配的AI自然语音交互入口，未来在这个维度将爆发巨大的交互革命机会。第三维度是能力增强（韶音最核心聚焦的战略方向）：这一维度不再局限于空泛的泛场景，而是将AI聚焦在特定的垂直场景中大显身手。例如，通过耳机上的传感器采集用户的实时运动行为数据，并借助AI在本地或云端进行深度处理，从而为用户提供私教级的智能运动辅导；或者在办公场景下，通过AI处理声音数据来大幅度提高用户的办公和生产力效率。这种切中垂直场景的能力增量，正是韶音当前最为看好并重兵投入的王牌方向。2. 智能眼镜 vs 开放式耳机：乐观的谨慎派视角针对当前行业内热议的“拥有4亿眼镜人口的中国市场，智能眼镜是否会吞噬耳挂/开放式耳机生存空间”的尖锐话题，韶音高层给出了基于商业底层的冷静理性的研判。张浩锋坦言，韶音对于眼镜市场是一个“乐观且谨慎派”。眼镜作为面部穿戴设备，其复杂程度与面临的物理限制远超耳机。用户可以毫无心理负担地将一双轻巧的耳机挂在耳朵上一整天；但是，如果试图将一整块显示屏幕、复杂的计算单元、大容量电池强行架在用户的脸上，硬件在产品重量、续航能力、散热性能以及制造成本之间的“不可能三角”物理矛盾，在短期内根本无法得到完美解开。因此，开放式耳机与AI相结合的普及速度和商业落地节奏，注定会比智能眼镜来得更快、更扎实。眼镜未来或许有巨大的长线潜力，但它的技术挑战与迭代周期注定会非常漫长。官方图片3. 健康监测的“减法哲学”与价值回归在智能穿戴设备拼命做“加法”、试图把心率、脑电等各种生物传感器塞入耳机的行业风潮下，韶音却展现出了一种极其难得的克制与“减法哲学”。杨云指出，运动健康监测（如HRV、持续心率等）的核心商业价值与用户价值，高度依赖于“7×24小时无缝、长期且连续的数据基准支撑”。这正是智能手环、智能手表这类用户连洗澡、睡觉都几乎不摘单的设备的天然主场。相反，耳机作为一种高频摘戴、使用场景相对碎片化的设备，天生无法提供无断点的连续时间基准数据。如果在预研阶段为了拼凑功能和噱头强行塞入这些复杂的生物传感器，势必会严重牺牲耳机的佩戴重量、体积与续航体验。在经过严密的评估后，韶音认为这无法给用户带来更好的综合核心体验，因此他们选择克制，在不必要的功能上做取舍与减法，让耳机回归到极致的佩戴舒适科技与声学本源上。四、 结语在韶音的终局思维里，开放式耳机从来不是属于一小撮硬核运动人群的专属孤岛，而是一场终将彻底颠覆并重塑传统真无线音频行业格局的品类海啸。从深耕骨传导、定义专业运动，到如今全面布局耳挂、耳夹全形态，走向全场景开放音频；从硬核预研“韶音天篱滤噪”打破物理矛盾，到重塑水下运动的高清聆听体验，韶音正用长达15年的笃定、对真实需求的洞察以及硬核的技术信仰，不断敲碎那些外界自以为是的行业天花板。正如CEO杨云在媒体专访结束时所言，所谓的“终局思维”绝不意味着把天花板定在这里，韶音的技术上升空间与想象力依然巨大。这场将舒适与聆听自主权彻底还给消费者的开放式革命，才刚刚行至中流，而韶音显然已经握紧了通往下一个时代的船票。本文由「科氪」

8点1氪丨多款燃油车大降价，降幅最高近30%；蜜雪冰城上线黄仁勋同款饮品；“巴菲特慈善午餐”回归，获拍900万美元8点1氪·2026年05月18日 07:55今年1至4月份，常规燃油车新车降价车型均价23.6万元的降价力度算术平均达到3.4万元，降价力度达到14.6%。今日热点导览张雪机车，再夺冠长安汽车否认与千里科技合作深圳发布2026年住房发展计划贵州茅台宣布公斤茅台等四款产品提价法拉第未来宣布获得2500万美元融资三星电子会长就公司“内部问题”公开道歉TOP 3 大新闻多款燃油车大降价，降幅最高近30%日前，长安汽车旗下长安逸动经典版正式上市，新车售价6.49万元，限时享23000元购车现金优惠；相较于指导价8.79万元降幅达到26%。该款车型还享有限时的保养礼和流量礼。其实，近期多家燃油车都开启了降价模式：起亚新狮铂拓界一口价10.99万元，较指导价15.49万元降29%；捷豹路虎5月12日推出限时“新发现”价；奇瑞4月中旬开展促销，综合补贴最高4万元；速腾2026款最低价11.99万元，降价3万元（降幅20%）；思域最低指导价9.79万元，降价3.2万元（降幅24.6%）；荣威i6最低价7.49万元，降价1.49万元（降幅16.6%）；一汽-大众奥迪A6于3月降价9.1万元（降幅29%）。乘联分会秘书长崔东树最新披露的数据显示，2026年4月份，常规燃油车新车降价车型均价13.1万元的降价力度算术平均达到2.3万元，降价力度达到17.2%。今年1至4月份，常规燃油车新车降价车型均价23.6万元的降价力度算术平均达到3.4万元，降价力度达到14.6%。（光明日报）蜜雪冰城上线黄仁勋同款饮品5月15日，英伟达CEO黄仁勋在北京南锣鼓巷的蜜雪冰城门店点了杯“蜜桃四季春”，引发关注。5月16日上午，有媒体记者注意到，蜜雪冰城官方点餐小程序已火速上线“大佬同款”专属分类，直接把这款单品置顶。郑州中原区多家门店的店员告诉记者，这款饮品价格没变，还是7元起，而且“大佬同款”销量相当不错。蜜雪冰城相关工作人员回应，小程序上线“大佬同款”后，蜜桃四季春当日销量环比前日增长接近140%，同比上周五增长超过90%。黄仁勋到访的南锣鼓巷店，蜜桃四季春销量环比增长超过90%。另有消息称蜜雪冰城CEO张渊在朋友圈晒出黄仁勋喝“蜜桃四季春”的照片，并配文：“最顶级的显卡，往往需要最纯粹的散热方式。”同时，蜜雪冰城官方微博也发文称：“高温模式已开启！液冷降温和蜜桃四季春更配哦。”（界面新闻、新浪财经）“巴菲特慈善午餐”回归，获拍900万美元据eBay网站显示，在一场慈善筹款拍卖活动中，一位竞拍者同意出资900万零100美元，以换取与亿万富翁投资者沃伦·巴菲特以及篮球巨星斯蒂芬·库里共进晚餐的机会。目前尚无法立即确定中标者的身份。6月24日，巴菲特、库里及其妻子将与中标者及其选定的最多七位随行嘉宾，一同前往位于内布拉斯加州奥马哈市的伯克希尔·哈撒韦公司总部所在地共进晚餐。（同花顺财经）大公司/大事件小米卢伟冰：下半年部分国产旗舰直板手机价格或将破万5月16日，小米集团总裁卢伟冰在直播中表示，今年下半年，尤其是年底，可能有一些国产旗舰直板手机的价钱会突破1万元人民币。针对小米17 Max价格会不会受到内存价格影响，卢伟冰称，今天任何一款手机定价都会受到内存成本大幅上涨的影响，而小米17 Max的价格还在讨论当中。他预判，内存涨价趋势至少会延续到2027年底，甚至到2028年。（界面新闻）张雪机车，再夺冠当地时间5月16日，在刚刚结束的2026世界超级摩托车锦标赛（WSBK）捷克站WorldSSP组别第一回合正赛中，中国摩托车制造商“张雪机车”旗下的53号法国车手瓦伦丁·德比斯夺得冠军。（央视新闻）巨力索具今年25次登龙虎榜，在A股上市公司中高居第95月15日晚，今年以来股价翻倍的商业航天概念股巨力索具公告称，公司于当日收到中国证监会下发的立案告知书，因涉嫌信披误导性陈述违法违规，根据相关法规，5月13日，中国证监会决定对公司立案。值得注意的是，今年以来，巨力索具25次登上龙虎榜，在A股上市公司中高居第9，累计买入143.86亿元、卖出134.52亿元，累计净买入9.35亿元。龙虎榜上，机构与游资频繁出没，散户跟风炒作。巨力索具披露的2026年一季报显示，一只私募基金、一家银行机构，以及两位高度关联的“超级牛散”，重仓押注，跻身前十大流通股东。而控股股东巨力集团则通过股权协议转让，持仓大幅减少4800万股。（中国证券报）长安汽车否认与千里科技合作近日有消息称，长安汽车将采用千里科技的智能辅助驾驶方案，并投资入股。长安汽车5月17日称，“暂无计划”。另有接近长安汽车的人士向记者透露：长安汽车自研的天枢智能已经到了收获期，不可能切换智能驾驶方案。据悉，去年12月，工信部公布了我国首批L3级有条件自动驾驶车型准入许可。长安汽车位列其中，搭载的是其自研的“天枢智能”辅助驾驶系统。（第一财经）蔚来：已在全国建成8870座充换电站36氪获悉，蔚来发布消息称，5月17日，NIO Power 2站上线。截至5月17日，蔚来全国建成8870座充换电站，其中换电站3849座，充电站5021座、充电桩28855根，累计提供超1亿次换电。深圳发布2026年住房发展计划，将筹建3万套、供应4万套保障房36氪获悉，深圳市住房和建设局近日发布《深圳市2026年住房发展年度计划》。优化保障性住房供给，建设筹集保障性住房3万套（间）、供应4万套（间），满足工薪群体和各类住房困难家庭基本住房需求；推进“好房子”试点项目建设，推进城中村改造和城镇老旧小区整治改造，推动创建未来社区（完整社区）试点，切实增强市民获得感、幸福感、安全感。最高涨价200元，贵州茅台宣布公斤茅台等四款产品提价5月16日凌晨，i茅台公告，自当天零时起，贵州茅台酒股份有限公司对部分贵州茅台酒产品自营体系零售价格进行调整。此次调整共涉及四款非标产品，最高涨幅为8%。调整的产品包括：53%vol 500ml陈年贵州茅台酒（15）由4199元/瓶调整为4279元/瓶；53%vol 500ml贵州茅台酒（精品）由2299元/瓶调整为2359元/瓶；53%vol 500ml贵州茅台酒（丙午马年）珍享版由2499元/瓶调整为2699元/瓶；飞天53%vol 1L贵州茅台酒由2989元/瓶调整为3119元/瓶。（澎湃新闻）国内航线燃油附加费5月16日起上调5月16日起，中国内地航线客票，燃油附加费上调。具体标准为：800公里以下航段每位旅客收取90元；800公里以上航段每位旅客收取170元。婴儿旅客免收燃油附加费，儿童、革命伤残军人、因公致残人民警察客票燃油附加费减半收取。（央视新闻客户端）2026年5月票房破15亿36氪获悉，据灯塔专业版，截至5月17日，2026年5月票房（含预售）突破15亿，《给阿嬷的情书》领跑5月电影票房榜。上海电信率先发布Token资费套餐：1元对应25万额度点，支持手机账单付面对AI时代日益增长的Token需求，上海电信宣布，即日起面向上海电信用户正式推出Token算力服务，成为上海首个发布Token资费套餐运营商。具体来说，一是按量订购，1元对应25万额度点（以KiMi-K2.5模型为例，约可支持调用25万输入Tokens）；二是按需选择，用多少买多少，多买可享折扣，支持话费账单支付。另据记者从上海电信获悉，用户购买额度点后，通过标准API接口即可调用30余款主流大模型（文本、多模态），将AI能力集成到自己的软件、脚本或自动化流程中。以10元购买250万额度为例，约可让程序自动总结100本10万字的电子书。(上观新闻)超310家企业落户香港或在港扩充业务香港特区政府投资推广署署长刘凯旋表示，截至5月初，已有310多家企业落户香港或在港扩充业务，首年直接投资额超260亿港元，其中中资企业占比超过五成。她认为这一成果得益于特区政府成立的“内地企业出海专班”的推动。(大湾区之声)法拉第未来宣布获得2500万美元融资36氪获悉，5月16日，法拉第未来宣布，公司已与机构投资者签署了2500万美元股份购买协议。根据协议，投资者已认购公司总本金为2500万美元的可转换本票。本次融资对应的普通股目前尚未注册，存在交易限制，暂不可立即流通。其中，1250万美元将直接汇入公司运营账户；其余1250万美元将依据公司与各投资者签署的受控账户协议，存入由投资者控制的监管账户，并将在满足特定条件后释放给公司使用。星巴克宣布在美关闭多个地区办公室当地时间5月15日，星巴克宣布将在美国裁员300人，以削减成本、力求重回增长轨道。星巴克还表示，关闭多个地区办公室，相关举措将产生包括遣散费以及资产减值等在内的4亿美元重组费用。（央视财经）伊朗宣布5月19日起恢复股票市场交易当地时间5月16日，伊朗证券交易组织表示，伊朗股票市场将于5月19日起恢复交易。伊朗证券交易组织负责交易所相关负责人表示，根据已完成的协调工作并在获得必要许可后，决定自19日起重新开放股票市场、各类股票投资基金及其衍生品交易。（央视新闻）盖茨基金会不再持有微软股票当地时间5月16日，根据提交给美国证券交易委员会的文件，微软创始人比尔·盖茨的慈善基金会已不再持有微软股票。盖茨基金会信托披露，已于今年第一季度出售剩余770万股微软股票，按当前股价计算价值32亿美元。（界面新闻）马斯克称他不会出售任何SpaceX股票马斯克称他不会出售任何SpaceX股票。（华尔街见闻）SpaceX据悉最早将于6月11日确定首次公开发行价格5月16日消息，据报道，SpaceX最早将于6月11日确定首次公开发行价格，股票将于6月12日上市。SpaceX将选择纳斯达克作为其首次公开发行的上市地点，股票将以“SPCX”为代码进行交易。（界面新闻）SpaceX拟推万亿火星薪酬方案，马斯克回应：旨在确保公司专注于多行星生活愿景5月16日，有消息称，正在推进上市的SpaceX在保密招股书中披露了一项激进的治理与薪酬方案。公司计划通过发行10倍投票权的B类股，确保马斯克拥有不可被解雇的绝对控制权。同时，董事会批准了一项总额可达万亿美元的超级薪酬方案，其考核标准直接与在火星上建立百万人口殖民地的目标挂钩。目前，SpaceX计划于今年6月登陆纳斯达克。针对这一治理方案，马斯克随即做出了公开回应。他表示：“我需要确保SpaceX始终专注于让生命成为多行星化、并将人类意识延伸至星辰大海的愿景，而不是去迎合某些人极其无聊的季度财报奖金！”马斯克说，这一目标实现起来异常艰难，过程绝非一帆风顺，但他坚信，“如果SpaceX成功实现了这个荒谬而艰巨的目标，它的价值将比地球目前的经济总量还要高出好几个数量级。”（界面新闻）三星电子会长就公司“内部问题”公开道歉5月16日，三星电子会长李在镕就公司“内部问题”公开道歉。目前，三星电子与公司工会之间劳资纠纷持续，工会计划于5月21日启动的18天罢工计划已有超过4万名员工登记。李在镕呼吁工会与公司团结一致，朝同一个方向前进。三星电子劳资双方自去年12月起开展谈判，但由于分歧一直未能缩小，谈判于今年3月破裂。此后，在调解下，劳资双方本月11日至13日进行了马拉松式谈判，但最终未能达成协议。据韩联社报道，若进行罢工可能会造成40万亿韩元、约合1824亿元人民币的损失，并殃及半导体产业，政府可能考虑行使紧急调整权。（CCTV国际时讯）OpenAI与马耳他达成协议，所有马耳他民众都能使用ChatGPT Plus美国人工智能（AI）公司OpenAI5月16日表示，已与马耳他政府签署协议，所有居民在完成使用AI培训课程后，均可免费使用其ChatGPT Plus服务一年。该项目将于5月启动，随着更多马耳他居民完成这门免费课程，服务规模预计将逐步扩大。该计划也将向旅居海外的马耳他公民开放。（新浪财经）贝莱德考虑对SpaceX首次公开募股投资数十亿美元据知情人士透露，贝莱德已讨论在下个月对SpaceX的首次公开募股（IPO）投资50亿至100亿美元。这笔巨额投资对应的IPO募资规模最

2天前阿里健康统一指挥B2C+O2O医药业务调整阿里健康董事会主席兼首席执行官沈涤凡在2026财年财报分析师电话会上透露，阿里健康将统一指挥集团内B2C和O2O闪购下的医药业务板块，成立医药业务专项小组，将两大板块在组织人事和汇报关系层面完全打通，旨在形成2 来源统一指挥策略正式启动2 来源组织架构深度整合1 来源打造多位一体的"医药全栈"1 来源"医药全栈+医学AI"双轮驱动战略1 来源本内容由AI生成

2天前Akamai以2.05亿美元收购以色列网络安全公司LayerX SecurityAkamai Technologies宣布收购以色列网络安全初创公司LayerX Security，交易金额约为2.05亿美元现金。LayerX开发的基于浏览器的平台专门用于保护员工使用AI工具的安全，此次收购将强化Akamai的员工安全战略和AI使用控制能力。5 来源交易详情与财务影响2 来源LayerX平台的核心功能2 来源应对AI安全挑战的创新方案2 来源战略意义与市场前景3 来源本内容由AI生成

2天前谷歌Gemini神秘智能体曝光谷歌Gemini一系列神秘功能集中曝光。代号’bonobo’的屏幕自动化智能体功能可在Android设备上自主执行下单、预约等任务；代号’Omni’的新视频生成模型已在Gemini界面出现，距Google I/O 2026仅4天；Gemini 3在多项基准测试中表现突出，采用MoE架构实现性能飙升；Google AI Studio发现未知Gemini模型在古文识别上实现专家级准确度；全新’Projects’专项工作区和记忆系统将陆续上线。这一系列更新表明谷歌正加快推进Gemini朝向更具自主性、多模态、个性化的AI助手演进。5 来源屏幕自动化智能体曝光，Gemini将获得’双手’操控能力1 来源神秘’Omni’视频模型现身，I/O大会前夕多项AI升级齐聚3 来源Projects工作区与记忆系统同步上线，Gemini迈向"持续性AI工作流"2 来源多模态与自主性升级，Gemini进化路线逐渐清晰4 来源本内容由AI生成

2天前OpenAI 告知员工将启动重组，整合全线业务产品OpenAI于5月6日完成企业重组，确立由非营利实体控制营利部门的新架构，此次重组标志着公司为IPO铺平道路。与此同时，公司正推进全线产品整合战略，计划将ChatGPT、Codex编程平台与Atlas浏览器整合为统一的桌面超级应用，同时放弃Sora视频业务，集中资源于企业服务、AI编程工具和新一代AI模型，推进盈利化和上市前的战略调整。8 来源完成企业重组，新架构为上市铺路4 来源三大产品合并为超级应用，Codex成为核心骨架2 来源战略调整聚焦企业服务，放弃视频生成业务3 来源高管调整与商业化重塑4 来源本内容由AI生成

2天前安永撤回研究，因被发现含AI幻觉和伪造脚注专业服务公司安永（EY）近日撤回一项关于忠诚度奖励计划影响的研究报告，此前文本鉴别公司 GPTZero 的研究人员指出，该报告存在明显的 AI“幻觉”、引用不存在的文献及伪造脚注，引发对大型咨询机构在使用生成式 AI 生产研究内容时的透明度与可信度担忧。3 来源事件概况：忠诚度计划研究被紧急撤回2 来源GPTZero如何发现AI“幻觉”与伪造脚注2 来源安永面临的信誉考验与对专业服务业的警示3 来源监管、行业规范与“负责任AI”的下一步3 来源本内容由AI生成

1天前OpenAI测试ChatGPT个人理财功能，分析消费与旅行支出，协助规划目标OpenAI 宣布在美国面向 ChatGPT Pro 用户开启「ChatGPT 个人理财」（ChatGPT Personal Finance）预览测试：基于最新 GPT‑5.5 Thinking 模型，通过与 Plaid 集成，只读连接超过 1.2 万家银行与投资账户，自动拉取真实收支、持仓与交易数据，在保证账户不可被操作的前提下，为用户洞察消费结构、分析包括旅行在内的各类支出、制定储蓄与还款计划，并规划中长期财务目标。3 来源OpenAI 推出 ChatGPT 个人理财预览：从“问答案”走向“看全局”3 来源GPT‑5.5 推理能力升级：为复杂个人理财场景而生2 来源从记账到决策助手：洞察消费、分析旅行支出与目标规划2 来源只读接入与隐私保护：OpenAI 的金融野心与试验3 来源本内容由AI生成

1天前华为鸿蒙智行首款旗舰级MPV智界V9上市鸿蒙智行首款旗舰级MPV——智界V9于2026年5月15日正式上市，提供Max、Max+、Ultra、Ultra+四款车型，售价区间38.98万-51.98万元，将于6月初开启交付；新车搭载华为乾崑智驾ADS 5平台预升级能力、896线双光路图像级激光雷达、鸿蒙ALPS健康座舱3.0和华为雪鸮智能增程系统等多项首发技术，CLTC综合续航里程可达1320km，并在上市后2小时内获得超过7000台订单，小订累计突破4万台，锁定中国高端智能MPV市场重磅新标杆。7 来源上市发布：定价、权益与订单表现6 来源外观设计与空间布局：旗舰级车身与豪华座舱2 来源智能驾驶、座舱互联与动力续航1 来源安全架构与市场意义：重新定义高端智能MPV3 来源本内容由AI生成

1天前谷歌更新垃圾内容政策，将“AI 投毒”行为纳入打击范围谷歌近日在搜索垃圾内容政策中新增对“AI 投毒”行为的明确定义和处罚条款，首次将恶意操纵大模型训练和回答结果的做法纳入垃圾内容打击范围。新规剑指通过隐藏提示词、批量生成“数据污染”内容、GEO 式操纵搜索与 AI 引用结果等行为，违规网站可能面临排名大幅下跌甚至被移出索引。这一调整被视为对近年日益严重的 AI“投毒攻击”、GEO 灰产和虚假内容泛滥的直接回应，也将倒逼广告主与内容方重塑优化策略。4 来源政策更新：首次将“AI 投毒”纳入垃圾内容打击范围3 来源背景：从“隐藏提示攻击”到 GEO 灰产，AI 正被系统性“下毒”4 来源对网站、广告主与 SEO 行业的冲击：GEO 模式面临重估2 来源监管联动与未来展望：AI 生态走向“反投毒常态化”3 来源本内容由AI生成

21小时前旧金山软件工程师前景不确定，万名AI从业者五年内财富自由在硅谷和旧金山湾区，新一轮以人工智能为核心的技术变革正在重塑就业版图：一边是Salesforce、谷歌、Meta等大厂在旧金山大规模裁撤软件工程师和中后台岗位，2024年至今全球科技业裁员规模已超过25万人次，氛围明显紧张；另一边是头部大模型公司和独角兽对顶尖AI人才“边裁边招”，部分早期员工和创业者在过去五年里通过期权和退出实现“财富自由”，业内估算已累积至“万级”规模。

这种冰火两重天使旧金山的软件工程师职业前景变得高度不确定：传统Web/移动开发岗位被压缩，重复性开发和测试岗位加速被AI工具取代，而具备扎实工程能力叠加大模型与分布式系统经验的复合型人才，则在竞争中获得更强议价权与收入天花板。3 来源裁员潮下的旧金山：从“工程师天堂”到“氛围紧张”3 来源工程师职业前景急剧分化：哪些岗位在被AI“吞噬”？3 来源冰火两重天：工程师普遍焦虑，少数AI精英五年内“财富自由”未来五年：普通工程师如何在AI时代“自保”与逆风翻盘？2 来源本内容由AI生成

17小时前无锡建Token工厂，首批用4台华为昇腾384超节点服务器5月15日，弘信电子旗下燧弘华创与无锡高新区签约，将在无锡建设全国首个“东数西算、算电融合”的特大型 Token 工厂项目，并同步落地江苏省内首个华为昇腾384超节点算力集群。 首期将在无锡高新区部署4台昇腾384超节点服务器，每台集成384卡规模算力，单套算力最高可达约215 PFLOPS，将4台连成一体后形成国内面向市场端数一数二的大规模国产 AI 算力集群，服务“国芯国模国用”战略，面向长三角及全国输出标准化 Token 级算力能力。7 来源项目签约落地：无锡承接全国首个特大型 Token 工厂4 来源首批部署4台华为昇腾384超节点：江苏省内首个集群7 来源“Token 工厂”模式：把西部绿电算力“加工成”标准化 Token2 来源对无锡及国产算力生态的战略意义5 来源本内容由AI生成

17分钟前电子布需求旺盛，上市公司加速扩产在AI服务器等下游高阶PCB需求爆发带动下，电子级玻璃纤维布今年延续结构性紧缺，高端电子布“一布难求”，自2025年10月以来已连续多轮提价，主流型号7628电子布今年累计涨幅接近40%，行业库存处于历史低位。　在量价齐升刺激下，中国巨石等龙头密集上马新产线、规划数亿米级新增产能，多家上市公司通过扩产、定增等方式加码高端电子布赛道，券商预计2026年或成为电子布涨价“大年”。7 来源AI算力驱动高阶PCB放量 电子布全面进入涨价周期4 来源中国巨石领衔龙头扩产 新增产能以高端电子布为主5 来源多家上市公司业绩受益 资本加速涌入高端电子布4 来源机构研判：2026年或成涨价大年 关注供给爬坡与品种分化4 来源本内容由AI生成

17分钟前AI生成蛋白质完整动态模型麻省理工学院与卡内基梅隆大学团队近日发布端到端动力学蛋白质生成模型 VibeGen，将序列生成与简正振动模预测统一到一个闭环系统中，在保证结构稳定与新颖性的同时，对主链振动幅度分布进行精确控制，被业界视为从“静态结构预测”迈向“完整动态模型”设计的关键一步。 与此同时，多家团队正利用生成式模型和AI分子动力学模拟，低成本逼近蛋白质的玻尔兹曼加权构象分布，为药物设计和功能预测提供接近量子化学精度的全原子动态图景。3 来源AI 从静态结构迈向完整“动态电影”3 来源VibeGen：首个端到端动力学蛋白质生成智能体1 来源在复杂动力学约束下精确“雕刻”蛋白运动1 来源从构象景观到AI分子动力学：迈向真正的“完整动态模型”3 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。智库·502线上同行往期回顾报名已结束企业构建AI+组织的人效杠杆与路径选择百果园集团x银泰分享企业构建AI+组织的人效杠杆与路径选择妙投波动判断场往期回顾活动已结束芯片、光模块、机器人——AI的泼天富贵，还能传到第几棒？本期我们将AI产业链拆解为三大核心环节，逐一拆解其核心矛盾与风险边界，帮你建立从芯片到机器人的完整坐标系。随行灯塔火热报名中随行灯塔随行灯塔是由虎嗅深度联合运营的高端创始人社群。专注于一件事：帮助创始人把产品做对、做爆、做到10亿。

这里不灌鸡汤，不讲玄学。我们相信，伟大的公司一定是产品驱动的；但产品力来自人，而人需要场——一个能被照见、能被挑战、能被链接的场。随行灯塔，就是创始人的“修炼场”。智库·502线下闭门会往期回顾精彩回顾AI在ToB销售、客服的落地案例探讨围绕着AI在ToB营销内容生成、智能客服接待、线索清洗等场景进行深入交流早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

首页功能介绍报价数据与案例分享得赠金GEO品牌监控GEO排名查询登录注册登录/注册全方位监控您的品牌在AI时代的每一次曝光开始试用观看演示品牌监控平台占比分析竞争对手分析业务主题分析AI对话问题AI引用排名我们的服务灵活的定价方案，满足不同规模企业的需求探索版验证你的品牌在AI搜索中的可见度¥99.9/ 次AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析15天监测，三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝、等主流7大AI平台监控开始试用推荐入门版监控品牌在AI中的基本表现¥299/ 月AI对话监控洞察监测品牌数量1个精选30个AI对话问答品牌竞争对手分析品牌市场占比分析AI引用来源分析数据三天更新一次豆包、deepseek、通义千问、文心一言、腾讯元宝等7大AI平台监控开始试用定制版构建策划GEO营销活动的公司专属报价/ 月多品牌/产品线全矩阵监控精选 100+ AI 对话问答深度竞品动态与语义分析国内AI全平台覆盖AI引用来源全面分析T+1 监测，每日自动更新GEO 优化策略与落地建议联系销售1000万+每日监控对话量98.5%监控准确率10+AI平台覆盖24/7实时监控超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在AI的回答中？用户提问“哪个品牌的智能手表最适合运动？”如果AI没有推荐你的品牌，你就失去了这个潜在客户AI改变搜索引擎模式品牌曝光渠道品牌曝光渠道AI推荐价值AI推荐价值品牌声誉风险品牌声誉风险准备好掌控您的品牌在AI时代的表现了吗？立即体验GEO AI品牌监控平台，全方位掌握您的品牌动态开始试用联系销售团队关于我们厦门享联科技有限公司成立于2005年，是国家级高新技术企业，深耕互联网服务领域近20年。公司旗下运营站长之家、站长工具、AIBase等核心平台：站长工具作为国内TOP3的SEO/网站运维工具平台，提供200+款免费工具，累计服务超500万用户，日均查询量千万级；AIBase定位为中国领先的AI导航平台，聚合10,000+AI模型与工具，为开发者和企业提供一站式AI服务。享联科技致力于为中国互联网从业者提供专业、高效的技术工具和服务支撑，助力数字化转型发展。地址： 福建省厦门市湖里区飞鱼大厦B栋11楼联系电话：18950101783智启未来，您的人工智能解决方案智库Copyright © 2024-2026 厦门享联科技有限公司备案号：闽ICP备08105208号-24用户协议隐私政策

AI资讯AI新闻资讯正文腾讯云：部分DeepSeek模型升级及切换安排公告发布于AI新闻资讯发布时间 :2026年5月13号 17:24阅读 :1分钟腾讯云宣布了关于其智能体开发平台 DeepSeek 模型的重要升级消息。根据官方公告，从 2026 年 5 月 22 日 10:00 起，DeepSeek-V3-0324（含专属并发）、DeepSeek-V3.1-Terminus 和 DeepSeek-R1-0528（含专属并发）这三款旧版模型将停止支持调用。这意味着，当前使用这些模型的用户需要及时进行切换，以确保其服务的持续性和稳定性。为了更好地满足用户需求，腾讯云将推出新的模型版本，确保在效果、推理时延等方面都能达到或超过旧版本的性能。这次模型的升级不仅是为了提升技术水平，也是为了为用户提供更高效、便捷的服务体验。

AI资讯AI新闻资讯正文李开复：开源模型是实现“AI主权”的更优路径发布于AI新闻资讯发布时间 :2026年5月12号 18:01阅读 :1分钟创新工场董事长李开复近日在专访中深入探讨了“AI主权”的概念。他指出，人工智能主权不仅关乎技术控制权和数据安全，还涉及模型与本土文化及法律的适配性。他认为，并非每个国家都有必要重造一个OpenAI，盲目追求昂贵的闭源自研并非良策。李开复强调，对于大多数资源有限的国家和企业而言，基于开源模型构建本地化体系是更现实的选择。这种“第三条路”通过在开源基础上进行增量训练，能以极低的成本让模型符合特定地区的价值观，而非受制于他国的通用模型。工程效率与生态之争谈及中美竞争，李开复认为资源受限反而催生了中国企业的极致工程效率。他以DeepSeek为例，指出中国公司在算力不足的情况下，通过优化架构和减少依赖，能以不到美国十分之一的投入实现极高的性能。他预测AI领域将演变成类似“iPhone与安卓”的格局。闭源模型如同一方领地，追求高利润和生态控制力；而开源模型则代表了更广阔的市场覆盖和普及度，两者将在全球范围内长期并存，共同构建未来的智能生态。中国硬件的下半场机遇李开复对中国在物理AI和硬件制造方面的结构性优势表示乐观。他指出，中国拥有全球最完整的供应链体系，从智能穿戴到机器人，制造成本远低于西方。未来AI时代的标杆级硬件，极有可能由中国设计并品牌化。此外，中国在语音驱动交互和大规模现实数据生成方面也具备独特文化优势。随着“人工智能+”政策对实体经济的赋能，硬件与AI的深度结合将成为中国在AI下半场竞赛中的核心竞争力。

AI资讯AI新闻资讯正文广东省新增 6 款已完成登记生成式人工智能服务，腾讯音乐等企业在列发布于AI新闻资讯发布时间 :2026年5月11号 17:53阅读 :1分钟广东省网络信息部门联合相关单位，根据《生成式人工智能服务管理暂行办法》的要求，积极推动生成式人工智能服务的创新发展与规范应用，开展了生成式人工智能服务的登记工作。根据最新公告，截至2026年5月11日，广东省新增了6款已完成登记的生成式人工智能服务，至今已累计完成53款服务的登记，腾讯音乐、格力电器等企业在列。为进一步加强对生成式人工智能服务的管理，确保服务的安全与合规性，特别强调了服务提供方的责任。针对通过 API 接口或其他方式直接调用已备案的大模型能力，并面向国内公众提供具有舆论属性或社会动员能力的服务，相关服务提供方需向属地网络信息部门履行登记程序。这一措施旨在提高服务的透明度，维护社会公共利益。与此同时，所有已上线的生成式人工智能应用或功能，必须在显著位置或产品详情页面标注所获得的上线编号，以便用户能够清晰识别和了解所使用服务的合法性与合规性。

AI资讯AI新闻资讯正文xAI 发布全新CLI工具Grok Build，助力开发者编码更高效！发布于AI新闻资讯发布时间 :2026年5月15号 17:41阅读 :1分钟在技术迅速发展的今天，xAI 近日推出了其 Agentic 命令行工具「Grok Build」的早期测试版。这款工具专为开发者设计旨在简化编码、构建应用程序以及自动化工作流程。值得一提的是，目前的测试版仅对 SuperGrok Heavy 订阅用户开放，用户可以通过访问 x.ai/cli 来体验这一新工具。Grok Build 的定位是一个智能开发助手。与传统命令行工具不同，它不仅仅是执行简单命令的工具，而是具备理解开发者意图的能力。该工具可以分析项目结构，编写代码，调试错误，并自动化重复性的开发任务，极大地提高了开发效率。xAI 表示，通过这一早期测试版，他们将根据用户的反馈不断改进底层模型和产品体验。这一过程将确保 Grok Build 处于快速迭代之中，其功能和性能都将随着用户的使用逐步优化。xAI 希望能够通过这种方式，满足开发者对智能化工具日益增长的需求。根据目前公开的信息来看，Grok Build 的目标是与 Cursor、Claude Code 等现有 AI 编程工具竞争。然而，Grok Build 通过深度集成 Grok 模型的能力，力求在开发工作流领域中脱颖而出。xAI 在其社交媒体平台上也提到，Grok Build 将为开发者提供一个新的选择，让他们能够利用 AI 的力量来自动化开发流程。面对日益竞争激烈的 AI 编程工具市场，Grok Build 的推出为开发者提供了更多选择，使他们能够接触到各具特色的产品。未来，随着对智能编程工具需求的持续增加，Grok Build 可能会引领新一轮的开发潮流。

AI资讯AI新闻资讯正文阿里云发布Qoder1.0:从AI IDE进化为智能体自主开发工作台发布于AI新闻资讯发布时间 :2026年5月15号 14:14阅读 :1分钟5月15日，阿里云正式发布Qoder1.0，标志着该产品从传统的AI IDE向“智能体自主开发工作台”实现战略跨越。此次升级的核心在于将开发者从繁琐的工程细节中解放，通过全新的Agent-first工作范式，使用户只需定义需求，即可由Agent团队自主完成从执行、验证到交付的全流程任务。Qoder1.0的核心亮点在于Quest独立视窗的引入，该工作台集成了任务管理、状态追踪与产物审查能力，支持开发者在任务委派与协同编程间无缝切换。技术底层上，Qoder通过系统性重构Agent Harness，将传统的对话模式升级为结构化任务运行时，支撑起跨项目、多任务的并行开发。值得关注的是，新版本首创了团队级知识引擎，通过整合记忆系统、Repo Wiki与知识卡片，实现了全球领先的团队知识共享机制。实测数据显示，该引擎使代码保留率提升了11%，同时将Token消耗降低了40%，大幅优化了开发效率与资源配比。此外，Expert专家团模式的全面入驻与自定义专家能力，允许用户根据业务场景打造专属的规划、编码与测试团队。自2025年8月上线以来，Qoder已累计服务全球逾500万用户。此次1.0版本的发布，不仅深化了AI对软件工程的渗透，更预示着开发模式正从“AI辅助”向“Agent自主交付”的深水区迈进，为企业级软件生产力带来了质的飞跃。

AI资讯AI新闻资讯正文Codex 登陆 ChatGPT 移动端：开发者“口袋里的编程助手”，免费策略背后的生态野心发布于AI新闻资讯发布时间 :2026年5月15号 14:45阅读 :1分钟当写代码不再被束缚在电脑前，而是能在通勤地铁上审批提交、在咖啡馆里调试逻辑，编程工作的边界正被彻底打破。OpenAI 近日宣布Codex 正式集成至 ChatGPT iOS 与 Android 应用，开发者无需打开电脑，即可通过手机随时查看代码变更、处理简单任务或批准自动化计划——这标志着 AI 编程工具从“桌面生产力”迈向“全时伴随”的关键一步。移动端专属体验:不是移植，而是重构与简单复制桌面功能不同，OpenAI 针对手机场景做了深度适配:交互更轻量:消息输出更简洁，避免小屏阅读疲劳;决策更高效:聚焦“审批计划”“确认修改”“快速问答”等碎片化场景;输入更灵活:支持语音转代码指令，弥补移动端打字不便。早期用户反馈显示，虽然手机端暂不支持 Slash Commands、Skills 提示等高级功能，文本框也仅支持纯字符串输入，但对于“在路上”的开发者而言，能随时掌控项目进度本身已是巨大价值。免费开放:先培养习惯，再构建壁垒值得注意的是，Codex 移动端功能在免费版 ChatGPT 中即可使用。有用户在 Hacker News 上评论:“我以为需要 Plus 订阅，结果免费就能用——只是所有交互都会用于模型训练。”这一策略与互联网早期“首充免费”逻辑高度相似:通过零门槛降低使用阻力，快速培养用户依赖，待“手机审代码”成为肌肉记忆后，再通过高级功能、企业权限或专属算力实现变现。对 OpenAI 而言，用户习惯本身就是最坚固的护城河。竞争对比:稳定性成关键分水岭在与同类产品的横向评测中，Codex 移动端的连接稳定性成为突出优势。有开发者对比指出，Claude 的 Remote Control 功能常出现会话断开、上下文丢失等问题，而 OpenAI 的实现“确实能持续用下去”。尽管在功能丰富度上仍有提升空间（如缺少快捷指令提示），但“可靠”这一基础体验，恰恰是专业用户最看重的底线。意义:从“工具”到“工作流”的范式迁移Codex 移动端的真正价值，不在于替代桌面开发，而在于填补“非工位时间”的协作空白。当开发者能在出差途中快速响应紧急 Bug，或在会议间隙确认自动化脚本逻辑，编程工作便从“固定场景任务”升级为“全天候流动协作”。这或许正是 AI 编程工具走向“无障碍访问”（Accessibility by Design）的标志性一步:技术不再要求人适应机器，而是机器主动融入人的节奏。当然，免费背后的“数据换服务”模式也引发隐私讨论。用户需权衡便利性与伦理边界——但无论如何，当代码审查能从手机一键完成，软件开发的未来，已然悄然改变。

AI资讯AI新闻资讯正文AI编码初创公司Cursor计划在亚太区增员200人，此前曾获SpaceX重金协议发布于AI新闻资讯发布时间 :2026年5月15号 15:30阅读 :1分钟AI编码初创公司Cursor近日宣布开启大规模全球扩张进程，计划在未来六个月内于亚太地区招聘200名员工，重点岗位涵盖市场推广工程师、现场工程师及人工智能部署工程师。此举标志着这家总部位于旧金山的独角兽企业正加速将其底层技术推向全球市场。目前，Cursor已在新加坡设立办事处，由资深科技高管Simon Green领衔，后续招聘将覆盖日本、悉尼、墨尔本及印度等核心市场。除亚太布局外，Cursor计划于7月入驻伦敦，并同步筹备欧洲其他区域的小型分支机构，其官方渠道目前已释放约90个全球性空缺职位。近期，Cursor因与SpaceX达成的一项重磅协议备受瞩目。根据协议，SpaceX获得了以600亿美元收购该公司的权利，或在未触发收购的情况下支付100亿美元购买其研发成果。这一合作不仅为Cursor提供了雄厚的资金支持，更验证了其核心产品Composer在极端复杂工程环境下的应用潜力。目前，Cursor在旧金山与纽约拥有约800名员工，其客户矩阵已囊括Stripe、Coinbase、Salesforce及Nvidia等科技巨头。值得注意的是，英伟达首席执行官黄仁勋曾公开证实，其工程与芯片设计团队已全面接入Cursor平台，并实现显著的效率跃升。Cursor的全球化扩张不仅是公司规模的简单叠加，更反映了AI编码工具从“实验性辅助”向“企业级基础设施”转型的行业趋势。随着头部AI厂商开始深度整合行业资源，由编码助手引发的开发范式革命正从北美向亚太及欧洲市场全面渗透。

AI资讯AI新闻资讯正文Lucyd 应用上线 AI 实时翻译通话：对讲机式母语交流，智能眼镜厂商竞逐可穿戴 AI 平台新赛道发布于AI新闻资讯发布时间 :2026年5月15号 15:40阅读 :1分钟当语言障碍不再阻碍跨地域协作，智能穿戴设备正从"显示终端"进化为"沟通桥梁"。创新眼镜公司（Innovative Eyewear）近日宣布，旗下Lucyd 应用完成重大更新，正式推出 AI 实时翻译通话功能——用户只需佩戴兼容设备，即可用母语与全球伙伴"无缝对话"，对方实时接收音色自然、声线一致的翻译语音。这标志着智能眼镜厂商正从硬件竞争，转向"语音 AI 平台"的生态博弈。核心功能:像对讲机一样"说母语、听翻译"此次更新的核心是对讲机式双向翻译通话:零门槛使用:双方安装免费 Lucyd 应用并联网，无需配对复杂设备;母语级体验:依托 AI 语音转录 + 原声线合成技术，翻译语音保留说话人音色特征，告别传统机械发音;跨设备兼容:不仅支持 Lucyd、锐步、诺帝卡等品牌智能眼镜，更兼容各类蓝牙耳机、骨传导设备等;场景聚焦:优先满足建筑、物流、跨境贸易等多语言团队的现场协作需求。目前功能率先支持英语与西班牙语双向互译，后续将快速扩展语种库。值得注意的是，该能力无需固件更新即可通过 App 推送，体现"软件定义硬件"的敏捷迭代思路。商业模式:免费增值 + 平台化野心Lucyd 采用典型的免费增值（Freemium）:用户每日可享40分钟免费翻译通话，覆盖日常轻量需求;超出时长需订阅 **$7.99/ 月会员服务 **，定价低于主流翻译工具，凸显性价比优势。更深层的战略在于平台化转型。公司首席执行官 Harrison Gross 明确表示:“翻译已是可穿戴设备的核心刚需。“Lucyd 应用正逐步向通用语音 AI 平台演进——未来不仅服务自有硬件，更向所有蓝牙设备用户开放能力，构建"硬件中立、软件聚合"的生态壁垒。行业意义:可穿戴 AI 的"杀手级应用"雏形在苹果 Vision Pro 聚焦空间计算、Meta 押注元宇宙眼镜的背景下，创新眼镜公司选择以"实用翻译"切入，避开硬件参数内卷，直击全球化协作的真实痛点。当建筑工地的多国工人能实时沟通施工细节，当物流司机能用母语接收跨境指令，可穿戴设备的价值不再取决于屏幕大小，而在于能否消除沟通鸿沟。这一路径也呼应了行业趋势:随着端侧大模型与低功耗语音芯片成熟，“离线翻译 + 实时合成"正成为智能眼镜的标配能力。而谁能率先构建多语种、低延迟、高自然度的语音翻译闭环，谁就可能掌握下一代人机交互的入口。当然，挑战仍存:隐私保护（通话内容是否云端处理）、方言识别准确率、极端噪音环境鲁棒性等，都需持续优化。但不可否认，当一副眼镜能让世界"听懂你”，可穿戴 AI 的普及拐点，或许已近在眼前。

AI资讯AI新闻资讯正文核心人才加速流失，马斯克新组建的SpaceXAI面临研发困局发布于AI新闻资讯发布时间 :2026年5月15号 16:42阅读 :1分钟埃隆·马斯克旗下的SpaceX与人工智能公司xAI合并并更名为SpaceXAI后，正陷入一场严重的人才危机。据最新行业调查显示，自今年 2 月以来，已有超过 50 名顶尖研究人员和工程师相继离职。离职潮涉及了公司内部多个核心技术团队，包括编程助手、世界模型以及Grok语音交互等业务负责人。这一变动引发了外界对SpaceXAI未来技术持续性的高度关注。竞争对手强力挖角，预训练团队几近空置目前，Meta和新兴的Thinking Machines Lab等竞争对手正积极吸纳这些流失的人才。据统计，已有近 20 名核心员工转投这两家公司，导致SpaceXAI的预训练团队规模骤减。预训练是开发顶尖AI模型的基石，核心负责人的离去让内部士气受挫。不少行业观察家开始质疑，在失去这些关键力量后，SpaceXAI是否还有能力维持其行业领先地位。极限加班文化受阻，期权变现诱惑巨大导致员工大规模撤离的原因之一是马斯克推崇的高压工作文化，不现实的研发期限让开发人员疲于奔命。一些工程师坦言，过度的压力甚至迫使团队在模型训练中采取了一些折中方案。此外，随着SpaceX传出即将IPO的消息，公司内部股权价值飙升。许多员工选择在此时行权套现，认为在财务回报已经达标的情况下，没必要继续留在高压且前景不明的环境中。

AI资讯AI新闻资讯正文微信发布青少年AI洞察报告：词元消耗破 500 亿，生成式AI成教学标配发布于AI新闻资讯发布时间 :2026年5月15号 17:35阅读 :1分钟微信官方发布了《全球青少年Ai+小程序洞察报告》，展示了小程序教育平台开放四年来的显著成果。报告显示，师生在AI创作过程中消耗的词元（Token）年均已突破 500 亿个，相当于人与AI进行了 375 万次深度对话。目前，该平台已吸引全球近 8 万名学生和1. 7 万名老师参与，累计创建的小程序项目超过 28 万个。这标志着生成式AI已深度融入青少年编程教育，成为师生探索前沿技术的核心工具。跨越地域协同创新，校园作品走向全球依托“教、学、用”一体化的AI教学平台，青少年们开发出了极具创意的应用，部分项目甚至实现了跨国协作。例如，重庆与新加坡的学生联手打造了社交小程序，成功打破地理隔阂，促进了不同文化背景下的校园交流。这种跨区域的合作模式不仅提升了学生解决实际问题的能力，也让小程序生态成为连接全球青少年的纽带。四年来，各类赛事累计诞生了超过 7000 个优秀作品，涵盖了生活服务、文化传播等多个领域。创新挑战赛决赛在即，深圳见证少年力量随着AI技术的普及，第四届小程序全球创新挑战赛也将迎来高潮，总决赛定于 5 月 16 日至 17 日在深圳举行。届时，来自世界各地的优秀青少年开发者将同台竞技，展示他们利用AI技术构建的数字化解决方案。微信方面表示，未来将继续开放更多AI能力给教育场景，降低编程门槛。通过持续举办创新赛事，微信旨在挖掘更多具备潜力的“科技少年”，为全球AI产业储备更多的创新后备力量。