AI 资讯简报

## 背景 ChatGPT 免费版一直是 OpenAI 吸引新用户的重要入口，但过去聊太多就会碰到“限流”。现在 OpenAI 决定把纯文字聊天变成无限量，等于把使用门槛又拉低一大截。对普通用户来说，ChatGPT 会更像微信或短信那样随叫随到，而不是一个聊几句就要等的工具。 ## 细节 从下周开始，免费版和 Go 套餐的纯文字聊天不再限次数，但上传文件、发送图片这类多模态消息依然有配额。本周，免费用户默认模型升级为 GPT-5.6 Luna；下周还会上线一个“思考”按钮，专门用来处理更复杂的问题。付费的 Plus/Pro 用户则会在周四拿到更新后的 GPT-5.6 Sol，OpenAI 说它“更实事求是”，尤其在涉及日期、数字、来源、规则或假设时，能更好利用找到的资料，减少错误，回答也更直接、更精简。Plus/Pro 用户还能用一个新滑块控制 ChatGPT 在回答前“想多久”。 ## 影响 对 OpenAI 来说，免费放开文字聊天是一步棋：既能让更多人养成用 ChatGPT 的习惯，也为未来的付费转化蓄水。同时，图片、文件仍然限流，说明 OpenAI 在算力成本和用户体验之间找平衡。“思考”按钮则是在免费层给出一个“深度推理”的样品，吸引用户升级。更深一层看，聊天正在变成默认入口，OpenAI 想用免费文字拿下最多的日常使用场景。

## Background ChatGPT's free tier has long been a way to try OpenAI's chatbot, but heavy users often hit rate limits. OpenAI now wants to make text chats unlimited for free and Go users, a move that lowers the barrier for casual users and makes ChatGPT feel more like a default messaging app. The change arrives as OpenAI pushes newer models across all tiers. ## What's changing Starting next week, text-only chats will no longer count against usage limits for free/Go users. However, multimodal messages with file uploads or images will still be throttled. This week, free/Go users also get the GPT-5.6 Luna default, and next week a “Think” button appears for harder questions. For Plus/Pro subscribers, GPT-5.6 Sol gets a reliability upgrade, especially for facts involving dates, numbers, sources, rules, or assumptions. OpenAI says the model will use sources more effectively and give more direct, tightly formatted answers. Plus/Pro users also get a slider to control how much reasoning ChatGPT performs. ## Impact The unlimited text tier is a clear signal that OpenAI wants to keep free users engaged and dependent on ChatGPT. It could increase usage and training data, while nudging those who need advanced features to subscribe. But the company is also being careful: keeping image/file limits preserves compute costs, and the new “Think” button gives free users a taste of deeper rea

## 背景 这项由麻省理工领衔、发表在《自然·医学》上的研究，聚焦皮肤病诊断。目前，已有一些医生在现实中用AI辅助看皮肤病，同时患者也越来越容易通过AI搜索产品接触到这类工具。研究者测试了好几种“可解释AI”的方法：只给预测和可信度、给相似图片、给热力图，以及用大语言模型生成的文字解释。 ## 主要发现 非专业人士使用任何一种解释方法后准确率都提高了，但主要靠的是听模型的。模型一旦出错，他们的表现下滑幅度比模型正确时带来的提升还要大。其中大语言模型生成的解释最让人盲从，甚至那些含糊、空泛的解释反而更有说服力。而初级保健医生则不一样——他们在只拿到AI预测、不带解释的情况下表现最好，大语言模型的解释对他们的帮助最小。研究还发现，如果先让用户看解释、再让他们自己判断，人们会更加依赖AI。 ## 影响 这项研究说明，AI解释并非放之四海而皆准。有经验的医生可以用解释来对照自己的判断，而新手可能利用同样的解释来形成观点，结果被带偏。这之所以重要，是因为越来越多患者会拿AI来咨询健康问题。研究还提到，一个公平性约束模型可以减少肤色偏倚，但非专业人士对它的依赖也很强。因此，界面设计中的时机安排——比如让用户先给出自己的初步诊断，再展示AI建议——可能有助于减少这类“自动化偏差”。

## Background This study, led by MIT researchers and published in Nature Medicine, focuses on dermatological diagnosis — an area where AI tools are already used by some clinicians and increasingly reach patients through AI-powered search products. The team tested several explainability methods: a simple prediction with a confidence score, similar images, heat maps, and LLM-generated explanations. ## Key findings Non-experts improved their accuracy with every explainability approach, but the gains mainly came from deferring to the model. When the model was wrong, their performance dropped more than it improved when the model was right. LLM explanations caused the strongest reliance, and users even found vague or generic explanations more convincing. Primary care providers, on the other hand, were more resilient: they did best when they got the AI prediction without any explanation, and LLM explanations gave them the smallest accuracy boost. The researchers also found that showing the explanation before users made their own diagnosis increased deference. ## Impact The study shows explainability is not one-size-fits-all. A clinician can use an explanation to check against their own experience, while a novice may use it to form an opinion and get pulled toward a wrong answer. This matters because patients increasingly use AI for health advice. The research a

## 背景 现在的二十多岁年轻人对“左滑右滑”的约会软件越来越烦。Ditto 的创始人 Allen Wang 和 Eric Liu 自己就是 Gen Z，从加州大学伯克利分校退学后创业，想做一个“不用滑”的约会工具。他们发现同龄人已经厌倦了 Tinder 和 Hinge 上没完没了的刷人和尬聊，真正想要的是线下见面、真实连接。Ditto 把 AI 红娘装进短信里，大学生只需用 iMessage 发一个代码就能注册。AI 会自我介绍说：这是由大学生做的约会项目，不刷卡片、不尬聊，每周三晚上 7 点安排真实约会。 ## 细节 用户注册时，AI 会先问姓名、性别、生日，再深入了解性格、兴趣和恋爱偏好。有人还会上传明星心动照片，帮 AI 判断“你喜欢的类型”。Ditto 的算法不是按表面爱好匹配，而是看爱好背后的性格信号：比如一个爱攀岩跳伞，另一个爱说唱滑板，表面爱好不同，但骨子里可能都是爱冒险、很个性的人。创始人说，化学反应的来源并不是共同爱好，而是更深层的特质。 每周三晚上 7 点，Ditto 会把匹配对象、时间、地点一起发给你。用户只需要准时出现，约会结束后再反馈，AI 再进一步学习你的偏好，下次推荐更好的人。目前有 15 万人注册，匹配后真的去约会的比例大概 20%。在创始人看来，这个比例已经不错，因为传统滑动软件上的配对真正线下见面的概率通常更低。它还做了一个年轻人喜欢的拼贴风介绍图，告诉双方为什么你们可能合适，这种风格和 Sonder 等新 app 很像。 ## 影响 现在 Ditto 只在几十所大学运行，使用 .edu 校园邮箱来确认学生身份，这在一定程度上能降低安全风险。团队在旧金山，目前有 12 名全职员工，还搞了个“首席游艇官”的职位来办派对营销。公司已经拿到 920 万美元种子融资，投资方包括 Gradient、Peak XV 和 Scribble，连多邻国联合创始人 Severin Hacker 也主动联系过他们。不过，要把这种“高效牵线”的模式扩大到更多学校和毕业后的用户身上，还要解决好安全和规模化的问题，这对任何约会软件都是挑战。但 Ditto 的思路表明，AI 正在从“帮你找工作”延伸到“帮你找对象”。

## Background People in their 20s are fed up with swipe-based dating. Ditto's co-founders Allen Wang and Eric Liu are Gen Z themselves and dropped out of UC Berkeley to build something different. They noticed friends were tired of the endless swiping and small talk on Tinder and Hinge, and wanted more genuine, real-life connections. Ditto turns dating into a text conversation: college students text a special iMessage number with a code, and an AI chatbot named Ditto walks them through signup. The bot describes itself as "a dating initiative built by college students who get it... no swiping, no awkward DMs, just real dates every Wednesday at 7 pm." ## Details During onboarding, the AI asks basics like name, gender, and birthday, then digs deeper into personality, interests, and dating preferences. Some users even upload photos of celebrity crushes to help the AI understand their "type." Ditto's algorithm does not match people by similar hobbies on the surface; instead, it looks at what those hobbies say about who someone is. A rock climber and a skateboarder may have different interests but share underlying adventurousness and individualism. Co-founder Wang argues that chemistry is actually predictable, once you have the right signals. Every Wednesday at 7 pm, users receive a match, a time, and a place. They just show up for the date, then give feedback. Ditto uses that feedbac

背景： 特斯拉和SpaceX都是出了名的“算力大户”。搞自动驾驶、做卫星网络，样样都需要海量AI算力。这篇文章说，这两家公司宣布首期联合投入168亿美元，在美国得克萨斯州格赖姆斯县建一座叫Terafab的AI芯片制造基地。为什么这么干？因为他们掐指一算，未来几年自己需要的算力，全球现在所有芯片厂加起来都供不上。与其等别人产能升级，不如自己动手，丰衣足食。 细节： Terafab不是普通的芯片厂，它要覆盖芯片生产的全部环节：设计、制造、存储芯片、先进封装，全都集中在一个园区里。马斯克说，两家公司的合计算力需求预计会突破1太瓦，远超全球当前产能。所以这个项目就是来填这个缺口的。他还表示，这会为得州带来数千个高薪岗位，并把“尖端制造带回美国”。另外，特斯拉今年四月已经开始在得州超级工厂旁边建一座研发晶圆厂，算是给Terafab提前铺路。 影响： 这件事的意义不小。首先，它暴露了AI巨头对算力短缺的焦虑——连造车、造火箭的公司都嫌芯片不够用。其次，它可能带动更多公司考虑自己造芯片，而不是完全依赖外部代工厂。从更大的视角看，这也是半导体制造业回流美国的一个标志性例子，得州正逐渐成为马斯克的科技大本营。这笔巨额投资也再次证明，AI算力已经成为一种战略资源，谁掌握它，谁就握住了未来的主动权。

Background: Tesla and SpaceX are known as heavy users of AI, from self-driving to satellite communications. As these technologies advance, their demand for computing power is skyrocketing. This article reveals that the two companies have announced a joint initial investment of $16.8 billion to build Terafab, a state-of-the-art AI chip manufacturing campus in Grimes County, Texas. The move is a direct response to a predicted shortage of global chip supply for their future needs. Details: Terafab is designed to be a one-stop chip hub. It will cover the entire semiconductor production chain, including logic chips, memory chips, and advanced packaging, all on one site. Musk says the combined computing needs of Tesla and SpaceX will likely exceed 1 terawatt, far beyond what current worldwide capacity can produce. By making chips in-house, they aim to close that gap. The project also promises thousands of high-wage jobs for the state. Notably, Tesla already began construction of an R&D wafer fab in April, which sets the stage for Terafab. Impact: This is a major strategic bet. It shows that even the world's most valuable companies feel the pinch of limited chip supply. It could also inspire other embattled AI players to consider vertical integration. On a broader scale, it contributes to the regional shift of advanced manufacturing back to the U.S., especially in Texas, w

## 背景 苹果在7月起诉OpenAI，说它搞了个计划，从苹果前工程师那里搞走机密硬件信息。苹果声称这些工程师跳槽到OpenAI后，带走了关于零部件制造、产品测试、供应商关系、销售渠道这些“商业秘密”。这周苹果还要求法院加快取证，说内部调查发现可能还有更多前员工掺和了。 ## 细节 OpenAI的驳回复议避开了“有没有偷”这个直接问题，转而攻击苹果自己的安全漏洞。OpenAI指出苹果让员工用个人iCloud账户处理工作，而且人离职后权限也没收回，甚至提交了聊天记录，显示一位苹果经理在工程师刘畅离职后，还登着他的个人iCloud传文件，后来还问他苹果项目的技术问题。OpenAI还说苹果根本没点名到底哪些算“商业秘密”，只列了一堆“零部件制造”“产品测试”这种泛泛的类别，等于没说。它认为苹果打官司就是为了阻止OpenAI挖人，因为“OpenAI不需要苹果的秘密，它在做全新的东西”。 ## 影响 这个案子如果OpenAI赢了，会让科技公司以后很难靠“商业秘密”去告跳槽员工和新东家。它也给所有大厂提了个醒：自己内部数据管不好，就别指望法院帮你撑腰。这场诉讼也反映出AI初创公司和传统硬件巨头之间的人才争夺战越来越激烈，未来类似的官司可能会更多。

## Background Apple filed a lawsuit in July accusing OpenAI of orchestrating a scheme to obtain confidential hardware information from former Apple engineers. Apple claims these engineers, now at OpenAI, took trade secrets about component manufacturing, testing, vendor relationships, and distribution channels. This week, Apple asked the court to speed up discovery, suggesting its internal probe found more ex-employees may have been involved. ## Details OpenAI's motion to dismiss argues that Apple's own security practices undermine the case. It points out that Apple let employees use personal iCloud accounts for work and failed to revoke access after they left. It submitted text messages showing an Apple manager stayed logged into defendant Chang Liu's personal iCloud after he left, even asking him later for technical help on Apple projects. OpenAI also says Apple's complaint only lists "generic categories" and doesn't name specific trade secrets. It calls the lawsuit baseless and pretextual, meant to punish OpenAI for hiring top talent that chose to leave Apple for more innovative work. ## Impact This case could set a precedent for how courts treat trade secrets when the original company had weak security. If successful, it would make it harder for big tech companies to sue ex-employees' new employers for poaching. For the industry, it highlights the growing tension between AI

## 背景 热带气旋（飓风、台风）是地球上最致命的天气现象之一，过去50年造成超过70万人死亡、1.4万亿美元经济损失。以前预测有个两难：预测台风往哪走需要粗尺度的全球模型，但预测它有多强需要细尺度的局部模型。WeatherNext用一个AI模型同时解决了这两个问题。 ## 关键细节 WeatherNext用大约20TB的全球大气数据和近5000个历史风暴记录训练而成，在TPU上不到一分钟就能生成15天预报。它只需要28x28公里的分辨率，比传统模型粗糙100倍，但准确率还更高。更小版本的WeatherNext 2-mini甚至可以在免费Colab笔记本上运行。在2025年飓风季，这个模型帮助美国国家飓风中心对飓风梅丽莎的快速增强和登陆牙买加做出提前预警。现在他们从一次生成50个预测扩展到一次生成1000个可能情景。 ## 影响 把代码和模型权重开源，意味着任何人都能在此基础上继续开发。这可能会加速全球天气预测能力，帮助地方气象机构、可再生能源规划和灾害应对。这次进步相当于气象学十年进展。AI不会取代人类预报员，但能给他们争取更多准备时间。

## Background Tropical cyclones are among the deadliest weather events, causing over 700,000 deaths and $1.4 trillion in losses in the past 50 years. Forecasters constantly race against time. Traditional models struggle with a trade-off: to predict where a cyclone goes (track) you need coarse global models, but to predict how strong it gets (intensity) you need fine local models. WeatherNext bridges this gap with a single AI model that handles both. ## Key details WeatherNext was co-trained on 20 terabytes of global atmospheric data plus historical cyclone records from nearly 5,000 past storms. It can produce a 15-day forecast in under a minute on a TPU. The model only needs 28x28km resolution data, 100x coarser than traditional models, yet still beats them. A smaller version, WeatherNext 2-mini, runs on a single TPU in a free Colab notebook. During the 2025 hurricane season, the model helped the National Hurricane Center make a historic early warning for Hurricane Melissa's rapid intensification and Jamaica landfall. Now they're scaling from 50 predictions at a time to 1,000 possible scenarios per cyclone. ## Impact Open sourcing the code and model weights means anyone can build on this technology. This could accelerate weather prediction globally, helping local forecasters, renewable energy planning, and disaster preparedness. The advance is equivalent to about a deca

## 背景 The Verge的政策记者Gaby Del Valle最近去了佛罗里达州埃尔南多县，那里刚刚通过了一项为期一年的数据中心建设禁令。组织抗议的是一个叫“人类优先”的保守派草根团体，创始人是早年茶党运动的组织者。这件事不是简单的左右对立，更像是民粹派和科技精英派的对立。无论是民粹左派还是民粹右派，都对大科技的发展充满怀疑，而数据中心变成了他们能实际抗议和阻止的目标。 ## 细节 不同地方的人反对理由不太一样。佛罗里达人说这里太潮湿，不适合建需要低温干燥环境的数据中心，还担心地下水和PFAS污染以及噪音。亚利桑那州的人却说沙漠太干，根本不该建。甚至还有“数据中心附近出生的牛都是死胎”这种传言，虽然源头只有一个未证实的农场主。很多居民知道数据中心是必须的，但他们反对的是把超大规模数据中心建在家门口。他们也不相信数据中心能带来多少工作，因为这类设施后期维护根本不需要多少人。还有人把数据中心和监控、移民执法联系起来，比如图森市的AWS项目就引发了关于ICE的担忧。对这些普通人来说，数据中心就是AI带来失业、垃圾内容和社区疏离感的实体象征。 ## 影响 这场反弹正在打乱传统的政治阵营。佛罗里达州长德桑蒂斯一方面支持让数据中心承担水电费和再生水要求，另一方面又推动给超大规模数据中心税收优惠。共和党州长初选热门人物Byron Donalds拿了AI和数据中心行业的钱，还得到特朗普背书，但一些特朗普的铁杆支持者说，如果他赢初选，他们宁愿投票给民主党。数据中心问题很可能影响中期选举乃至2028年大选。它已经成为少数能让左右两派站到一起的议题，大家都讨厌这些又大又吵的新建筑。

## Background The Verge policy reporter Gaby Del Valle recently visited Hernando County, Florida, where the county commission unanimously approved a yearlong moratorium on data center construction. The protest was organized by Humans First, a conservative grassroots movement whose organizer was also an early Tea Party figure. This makes the fight not left vs right, but more populist vs technocrat. Both populist left and populist right are suspicious of big tech development, and data centers have become the physical thing people can rally against. ## Details Local concerns vary from place to place. In Florida, residents point to the humidity and worry about groundwater, PFAS, and constant noise. In Arizona, people say it doesn't make sense to build in a dry desert. Some rumors are extreme, like claims that cattle born near data centers are stillborn, though the original source appears to be one unverified farmer. Many residents understand that data centers are needed, but they don't want hyperscale buildings in their neighborhoods. They also don't believe the jobs promise, since these facilities don't need many permanent workers. Some people connect data centers to surveillance and immigration enforcement, like in Tucson where an AWS project has raised concerns about ICE. For many, data centers are a visible symbol of what they dislike about AI, including job loss, socia

## 背景 OpenAI 这次公开的是新版 ChatGPT 语音系统 GPT-Live 的底层工程改造。实时语音交互里，延迟是最伤体验的。文字回答慢几百毫秒影响不大，但音频一卡就会让人立刻觉得“不自然”。OpenAI 花了 6 个月重做语音系统，文章里最核心的数据是：新系统的 p95 音频帧延迟降到了旧系统 p50 的水平。意思是，原来最慢的那些延迟帧被大幅压住，绝大多数音频都能稳定按时到达。 ## 细节 这套改造有几个关键思路。第一，不再把语音当“语音转文字、模型推理、文字转语音”的单体流程，而是参考人类神经系统的多级通道，把声音放上快速通道，搜索、工具调用、日志保存挪到异步后台。第二，网络层用自研 WARP 协议把 WebRTC 建立从 6 次网络往返压缩到 1 次，路由信息被直接写进连接协议，省掉跨网络查询。第三，把部分媒体前端和推理逻辑从 Python asyncio 改成 Go，降低高并发下调度和内存回收带来的不可控停顿，还优化了 Linux 内核的 UDP 收包。第四，模型持续接收声音，不再等静音结束才应答，发言权判断交给模型本身；模型实例还可以带着上下文迁移，切换实例时旧实例继续运行，新实例 Prefill 完成后才切换，避免长时间停顿，但代价是短时间占用双份算力。 ## 影响 OpenAI 没有公布持续推理的具体成本、误抢话概率、多次压缩后的信息损失和后台任务过期比例。这些还是未知数。但这次拆解给行业指出了新的竞争点：实时语音的瓶颈往往不在 GPU 推理，而是日志、状态存储这类辅助组件会先饱和。开发者不用只等模型变快，更需要在网络握手、内存管理和状态机这些工程细节上发力。对普通用户来说，ChatGPT 的语音会更接近真人对话，卡顿和抢话会明显减少。

## Background OpenAI published an engineering deep dive on GPT-Live, the voice system behind new ChatGPT. Real-time voice has a hard latency problem: text can be slow without killing the experience, but one broken audio frame breaks the illusion of human conversation. OpenAI says it spent six months rebuilding the system. The headline metric: p95 audio-frame latency in the new system now matches p50 in the old one. The slowest 5% tail has effectively been tamed, so almost all audio frames arrive on time. ## Key details The redesign does several things. It abandons the old pipeline view of speech as "ASR, model inference, TTS" running in one service. Instead, GPT-Live uses a multi-path architecture: instant voice feedback travels on a fast path, while search, tool calls, and logging move to async background tasks. The network layer gets a custom protocol called WARP, which merges DTLS handshake, SCTP setup, and data channel negotiation, cutting WebRTC setup from 6 round trips to 1. Route hints go directly into the ICE ufrag, so relays can map packets in memory without a remote Redis lookup. The team also rewrote parts of the media front end and inference logic from Python asyncio to Go, reducing unpredictable pauses from garbage collection and thread scheduling, and tuned Linux kernel UDP handling. Another shift: the model no longer waits for silence before answering. It

## 背景 苏剑林发布文章复盘 Kimi K3 架构。K3 总参数 2.8万亿，有 896 个路由专家，但每次只激活16个。这意味着模型容量非常大，但单次计算成本没有同步膨胀。之所以要这么设计，是因为大模型不能只顾着变大，还要看每次处理实际花多少钱。 ## 关键细节 传统 MoE 要把 Token 发给不同 GPU 上的专家，通信常常先成为瓶颈。K3 用 LatentMoE 把隐藏层从 7168 维压到 3584 维再送进专家，专家变“窄”，通信和计算都更省。省下来的预算用来扩充专家数量：从“448选8”变成“896选16”，让更多窄专家组合干活，能力分工更细。 规模大了之后，稳定性和负载均衡也要跟上。K3 加了 RMSNorm 和 SiTU-GLU，防止数值波动在低精度训练（如 BF16/FP8）下被放大。用 Quantile Balancing 代替以前的固定步长调偏置，直接估计每个专家的选择门槛，让大家都不闲着。 注意力部分，K3 没有只用一种方案。它把 KDA 和 Gated MLA 交替排列，KDA 负责用固定大小的状态持续更新记忆，MLA 负责定时回查完整历史。因此 MLA 不再需要 RoPE 也能建模位置信息，这叫“广义 RoPE”。模型还把 RoPE 分支留着，不是为了好看，是为了复用成熟的训练和部署基础设施；Per-Head Muon 则让优化器按注意力头分开更新，减少互相干扰。 ## 影响 K3 的取舍说明，万亿参数模型的新竞赛不是单纯堆参数。更关键的是把参数、记忆和计算调度起来，让每步计算只碰该碰的部分。当然，K3 也有没说清的问题：低维空间可能丢信息，固定状态会遗忘细节，更多专家不一定都有清晰分工，而且效率依赖专用 Kernel 和通信优化，别人直接照抄架构不一定能得到同样效果。但对整个行业来说，它展示了一条“更大但更省”的可行路线。

## Context Su Jianlin (creator of RoPE) published a teardown of Kimi K3. The model has 2.8T total parameters and 896 routed experts, but activates only 16 per token. That is the core design philosophy: keep a huge parameter "knowledge space", while preventing compute cost from exploding as the model gets bigger and context grows longer. ## Key details Traditional MoE may need to send token hidden states to different GPUs. As experts grow, communication often becomes the bottleneck before matrix math. K3 uses LatentMoE: compress hidden state from 7168 dims to 3584 dims before sending to routed experts. That reduces each expert's compute, weights, activation, and cross-device traffic. The saved budget is then used to expand the expert pool: instead of 448 experts selecting 8, K3 can choose 16 out of 896 experts. More, narrower experts can combine for fine-grained division of labor. It also keeps 2 shared experts for general language functions. Bigger scale creates stability problems. K3 uses RMSNorm after expert aggregation to normalize scale. SiTU-GLU uses soft cap to suppress outliers that could be dangerous for BF16/FP8 low-precision training. Quantile Balancing estimates the selection threshold position for each expert, instead of fixed-step bias updates, to keep load balanced across 896 experts. Attention is also mixed: KDA and Gated MLA are alter

## 背景 Jeff Dean 是谷歌最早的“基础设施之神”和“AI 教父”级人物。1999 年加入时谷歌只有 25 人，到离职时已超过 19 万人。他早期和搭档 Sanjay Ghemawat 做出 GFS、MapReduce、BigTable，直接开创了大数据和云计算时代；2011 年后他转向 AI，和吴恩达、Greg Corrado 一起创办 Google Brain，后来又牵头 TensorFlow、TPU，并参与 Gemini 大模型。2026 年 8 月 5 日，这位效力谷歌 27 年的传奇人物正式官宣离职，随即与 Sanjay Ghemawat、Oriol Vinyals、Quoc Le 等老搭档创办新公司 Discovery Loop。核心方向是“AI for Science”，把机器学习、科学发现和工程开发全流程自动化，做一个能自己读论文、设计实验、跑验证的“AI 科学家”。 ## 细节 Jeff 离职时流出的路演 PPT 中，一份“34 位创始人”名单非常震撼。名单里包括 Anthropic 创始人 Dario Amodei、SSI 创始人 Ilya Sutskever、Character.AI 创始人 Noam Shazeer、Cohere 创始人 Aidan Gomez、Mistral AI 创始人 Arthur Mensch、Perplexity 创始人 Aravind Srinivas，以及 Physical Intelligence、Ricursive Intelligence 等机器人/芯片赛道公司的创始人；还有 xAI 的戴子航、张国栋等华人 AI 人才。杨植麟也曾在 Google Brain 与 Quoc Le 合作，后来创办了月之暗面。所有这些人都和 Google Brain 有直接关系，Google Brain 一度扩张到 1001 人，堪称 AI 界“黄埔军校”。 Jeff 还带火了一个关键项目：Brain Residency 驻留计划。这个计划不要求候选人必须有 AI 或计算机博士学历，招收了不少物理、数学、生物医学背景的“跨界异类”，考核的是量化思维、对 AI 的狂热自驱力和跨界视角。入选者拿谷歌全职员工薪资和无限算力，还能直接跟 Jeff Dean、Geoffrey Hinton、Quoc Le 等泰斗级人物做研究。正是这种不按常理出牌的人才机制，为现代 AI 行业输送了大量创始人。 ## 影响 这次创业的深层含义是：AI 人才正在从大厂向创业公司大规模流动。Jeff 的目标不只是做一个新模型，而是 “递归自我改进”，让几个人的科学家团队租用自动化 AI 服务后，能在研发效率上超过全球最庞大的科研机构。如果这条路走通，AI 将不再只是工具，而会真正变成科研主体。对行业来说，Google Brain 系人才已经渗透到 OpenAI、Anthropic、xAI、欧洲 Mistral/Cohere 等主要玩家，谷歌系人才仍占 AI 创业半壁江山。Jeff 离职的消息还带动了 Alphabet 股价波动，后续市场会继续关注 Discovery Loop 能否兑现“自动化科学家”的宏大叙事。

## Background Jeff Dean is arguably Google’s “infrastructure god” and “AI godfather.” When he joined in 1999, Google had only 25 employees; by the time he left, it had more than 190,000. He helped build GFS, MapReduce, and BigTable, which created the modern big data and cloud computing era. After 2011 he turned to AI, co-founded Google Brain with Andrew Ng and Greg Corrado, drove TensorFlow and TPU, and later worked on Gemini. On August 5, 2026, he officially left and started Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The core mission is “AI for Science”: automate machine learning, scientific discovery, and engineering, and build an “AI scientist” that can read papers, design experiments, and run checks on its own. ## Details The leaked pitch deck reportedly includes a stunning list of 34 founders who came out of Google Brain: Dario Amodei (Anthropic), Ilya Sutskever (SSI), Noam Shazeer (Character.AI), Aidan Gomez (Cohere), Arthur Mensch (Mistral AI), Aravind Srinivas (Perplexity), founders of Physical Intelligence and Ricursive Intelligence, plus xAI’s Zihang Dai and Guodong Zhang. Yang Zhilin, founder of Moonshot AI, also worked with Quoc Le at Google Brain. Google Brain grew to about 1,001 people and became a “talent academy” for the AI world. A key part of this story is the Brain Residency program Jeff started in 2016. It did not require an AI

## 背景：一次“面向未来”的大调整 谷歌这次官宣，几乎是给AI圈扔了一颗炸弹。DeepMind创始人哈萨比斯不再管日常运营，转去当DeepMind董事会主席和Alphabet首席科学家，专注AGI和研究；DeepMind现任CTO Koray Kavukcuoglu接棒。谷歌老将、第30号员工Jeff Dean则和另外三名顶级研究员（Sanjay Ghemawat、Quoc Le、Oriol Vinyals）一起离开，成立新创业公司Discovery Loop，打算用AI自动化大量科学和工程实验，涉及药物研发、硬件设计、清洁能源和材料设计。 ## 细节：被掩盖的“政治”和道德冲突 官方说法是“为下个阶段做准备”，但多名知情人士说，变动背后没那么单纯。一个关键背景是谷歌接下了美国国防部的合同，可能把AI用在大规模监控和致命自主武器上。此前Anthropic因为拒绝这些“红线”被五角大楼拉黑，谷歌却签了“任何合法用途”，填上了空缺。Jeff Dean一直是内部敢说话的人，公开反对监控和杀人机器人，还支持过Anthropic的诉讼，也批评过ICE处决平民事件。他和哈萨比斯早年都签过公开承诺，反对把决定人生死的权力交给机器。专盯AI行业的非营利组织Midas Project创始人说，军方合作已经导致DeepMind人员流失和工会运动，Dean和哈萨比斯很可能都反对过。一位在职员工说，Dean离开是“谷歌理想主义的残余彻底消亡”，也会加重人才流失。 ## 影响：速度优先，研究让路 行业普遍认为，谷歌在传统巨头里AI实力最强，但就是比OpenAI、Anthropic慢半拍。哈萨比斯一直对长期科研兴趣更大，对快速推产品没那么上心。这次调整被不少人解读为哈萨比斯影响力下降，谷歌要转向更快推产品的路线。CEO皮查伊也说，必须继续加速、明确目标。一名刚离开谷歌的人说，过去十年没见过谷歌内部AI产品领域竞争这么激烈，公司正在想方设法招能加快产品上线的人。这次变动短期内会让谷歌更“卷”产品，但也可能让愿意谈道德的声音更少，留下“平庸化”的隐忧。

## Background: A huge reshuffle framed as looking ahead Google’s announcement rocked the AI world. DeepMind founder Demis Hassabis is stepping back from day-to-day leadership to become DeepMind chair and Alphabet’s chief scientist, focused on AGI and research; DeepMind CTO Koray Kavukcuoglu will lead DeepMind. At the same time, Jeff Dean, Google’s 30th employee and 27-year veteran, is leaving with three other top researchers, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, to start a new AI startup called Discovery Loop, aimed at automating thousands of science and engineering experiments in drug discovery, hardware design, clean energy, and materials design. ## Details: The politics and ethics behind the move The official story is about preparing for the next chapter, but sources told The Verge the shift has messier roots. A key backdrop is Google’s Pentagon deal, which could allow AI to be used for mass surveillance and lethal autonomous weapons. After Anthropic refused those red lines and was blacklisted, Google signed on for “any lawful use.” Dean had been one of the most visible internal critics of such uses, and in the past he and Hassabis both signed public pledges against delegating life-or-death decisions to machines. Tyler Johnston of The Midas Project said the deal has already caused departures and unionization at DeepMind, and there are signs Dean and Hassabis

## 背景 Mirendil是一家由Anthropic前研究员创立的AI初创公司，研究方向是递归自我改进，也就是让AI能自己不断迭代变强。像Anthropic这样的大实验室一直在探索这个方向，近期也冒出了几家追逐同一目标的创业公司。但要取得实际进展，需要极其庞大的算力，这正是谷歌云要参与的原因。 ## 细节 据TechCrunch报道，这笔交易金额超过1亿美元，Mirendil将获得谷歌TPU、英伟达GPU以及受管理的训练集群。联合创始人Benham Neyshabur表示，他们的软件层能帮客户更好地利用谷歌的硬件。谷歌则把它看成一桩战略合作——不只是卖芯片，而是要构建“计算智能系统”，突破扩展的物理限制。 ## 影响 这桩合作反映了两大趋势：云巨头正在用大额基础设施投资来拉拢AI初创公司，AI公司也在疯狂锁定算力合同以保证自己不掉队。对Mirendil来说，这是一次大输血——公司6月底种子轮估值刚达到10亿美元，这笔合同相当于融资额的一半。对谷歌云而言，这能让他们获得一个专注前沿AI的战略伙伴，未来还可以把相关技术推销给企业客户。如果一切顺利，自我进化AI未来可能自动化科学研究、帮助寻找阿尔茨海默病等疾病的疗法，甚至独立完成整个AI实验室的工作。

## Background Mirendil is a fresh AI lab founded by ex-Anthropic researchers. It's working on what's called recursive self-improvement — AI that can iteratively improve itself. Big labs like Anthropic have been exploring this, and a few startups are now chasing the same goal. To make real progress, you need huge amounts of compute, and that's where Google Cloud comes in. ## Details According to TechCrunch, the deal is worth over $100 million and gives Mirendil access to Google's TPUs, Nvidia GPUs, and managed training clusters. The startup's co-founder Benham Neyshabur says their software layer helps customers get more out of Google's hardware. Google sees this as a strategic partnership — not just selling chips, but building 'systems of intelligence' that can break through scaling limits. ## Impact This deal reflects two big trends: cloud giants are investing heavily in AI startups with infrastructure commitments, and AI companies are locking in compute deals to stay competitive. For Mirendil, it's a huge boost — the company was valued at $1 billion after its seed round in late June, and this contract is about half of that amount. For Google Cloud, it's a chance to have a strategic partner working on frontier AI that could eventually be sold to enterprise customers. The potential payoff? If self-improving AI works, it could automate scientific research, help find c

## 背景 过去几年，Spotify 之所以总能猜中你想听什么，靠的是一套叫 Vector AI 的行为预测系统。它不只看你以前听了什么，而是实时判断你接下来想干什么。据报道，这套系统支撑了 Spotify 约 90% 的推荐，服务 8 亿用户。现在，参与打造这套系统的三位工程师 Sidd Motwani、Ian Anderson 和 Shivaditya Sinha 从 Spotify 出来，成立了新公司 Malachyte，打算把同样的能力用到电商上。公司刚宣布完成 1000 万美元种子轮融资，由 Bessemer Venture Partners 和 Gradient 共同领投，Harpoon Ventures 参投。 ## 细节 Malachyte 想解决的问题很简单：现在的网上商店大多‘千人一面’，新访客看到的是通用首页，老顾客推荐的也只是‘你以前买过什么’，而不是‘你现在需要什么’。它的平台用一套名叫‘双头 Vector AI’的系统，在页面加载的瞬间就开始预测顾客想找什么，不需要账号，也不需要历史记录。比如你搜‘重型靴子’并点了两次钢头靴，页面就会自动把工装裤、手套往前排，把皮鞋往后推。每一次点击、滚动、搜索都会让系统更懂你当下的意图。 公司从 2024 年开始开发和测试，合作过 20 多家旅游、生鲜和零售企业，最终聚焦电商。平台 2025 年秋季在 Fun.com 上线，2026 年 6 月起向 Shopify 商家开放，大商家可以通过 API 接入。 ## 影响 如果这套系统真的普及，网上购物会从‘记住你’变成‘懂你此刻’。对商家来说，这可能是转化率的新机会；对用户来说，购物体验会更像真人导购。创始人也提到，更大的机会是把商品陈列和营销统一到同一套对用户行为的理解上。不过，目前它还是种子轮阶段，实际效果有待更多商家验证。

## Background Spotify's ability to predict what you want to hear comes from a behavioral intelligence system called Vector AI. Rather than only looking at what you listened to in the past, it figures out what you're likely to do next. The system reportedly powers about 90% of Spotify's recommendations for its 800 million users. Now, three engineers who built it — CEO Sidd Motwani, Ian Anderson, and Shivaditya Sinha — have left to launch Malachyte, bringing the same technology to e-commerce. The startup announced Thursday that it raised $10 million in seed funding, co-led by Bessemer Venture Partners and Gradient, with Harpoon Ventures participating. ## Details Malachyte wants to fix a familiar problem: most online stores treat everyone the same. First-time visitors see a generic homepage, and returning customers get suggestions based on what they bought before, not what they need right now. The platform uses a 'two-headed Vector AI' to predict what a shopper wants, starting the moment a page loads. It does this without requiring an account or purchase history. For example, searching for 'heavy-duty boot' and clicking on steel-toed boots is enough to push work pants and gloves up the page and dress shoes down, in real time. Every subsequent click and search sharpens the system's guess. The company has been testing since 2024 and worked with more than 20 travel, groce

## 背景：机器人要变聪明，光“看”不够 具身智能指的是让机器人真正理解和操作物理世界。现在的模型不能只靠视频“看”，还要理解手指抓握力度、人体关节运动、三维空间中的动作轨迹、人与物体交互状态。这些高精度多模态数据正是下一代机器人通用大脑的“燃料”。黎曼动力计划到2026年底完成100万小时具身数据采集与训练。 ## 合作细节：两个伙伴补上数据和验证闭环 黎曼动力和光轮智能合作，把Riemann-1.0具身世界动作模型、Matrix-Game 3.5交互式世界模型，和光轮的EgoSuite人类数据平台、RoboFinals评测平台、RoboStack部署反馈平台打通。先看Riemann-1.0在真实任务中哪里不行，再用能力短板反推数据需求，而不是盲目堆数据。和诺亦腾机器人合作，则主要采集人体动作轨迹、关节运动、力觉反馈、交互对象状态等数据，重点解决多模态数据在时间和空间上的高精度同步问题。 ## 影响：赛道从拼模型开始转向拼数据闭环 Riemann-1.0之前已经在国际机器人基准和真实机器人任务中有成绩。在RoboCasa-365这个公认很难的家务基准中，它以62.6%平均成功率拿下第一，比原来行业领跑水平高8.4个百分点。这次合作说明黎曼动力希望从单一模型突破，走向“数据生产、模型训练、真实验证、反馈迭代”的闭环。对行业来说，机器人公司不再只是发布更大模型，开始比拼谁能持续拿到高质量真机数据，并让数据反过来驱动模型迭代。

## Background: Robots need more than just vision Embodied AI is about getting robots to truly understand and operate in the physical world. To do that, models can't only rely on visual information; they need to understand finger grip forces, human joint motion, continuous trajectories in 3D space, and dynamic interactions between people, objects, and the environment. These high-precision multimodal signals are the essential "fuel" for building the next-generation robot brain. That's why Riemann Dynamics plans to complete 1 million hours of embodied data collection and training by the end of 2026. ## What the partnerships actually do With Lightwheel AI, Riemann will connect its Riemann-1.0 embodied world action model and Matrix-Game 3.5 interactive world model with Lightwheel's EgoSuite human data platform, RoboFinals evaluation platform, and RoboStack deployment feedback platform. The key move is to first identify weak points from real-world task performance, then use those capability gaps to drive data requirements, rather than simply piling up more data. With Noitom Robotics, the focus is on collecting human motion trajectories, joint states, force feedback, and interactive object states, and solving the high-precision synchronization of multimodal data in time and space. ## Impact: The race is moving from models to data loops Riemann-1.0 has already shown strong