🏆 Headline
OpenAI parts ways with three safety researchers after internal probe finds mishandling of sensitive information
OpenAI has parted ways with three members of its safety team — two safety researchers and one research program manager — The Wall Street Journal reported. A company spokesperson confirmed that an internal investigation found they shared confidential company information with an outside AI safety organization, "violating our policies and breaking the trust essential to our work." The report did not name the researchers, the organization, or the information involved; lists of suspected names circulating on social media remain unverified. For a safety storm that has simmered for a week — a public misalignment incident archive, a canceled GPT-6.1 Astra release — this is the first time the fallout landed on specific people: this time, the ones who left were not models but the people studying model safety.
Source: WSJ / TechCrunch | 2026-10-01
Google puts a TPU in orbit: prototype space-compute satellite launched today
Google's Project Suncatcher took its first step: a satellite built by Planet Labs rode a SpaceX rocket out of California today, carrying a Google TPU — the company's first advanced chip in orbit. The satellite must prove three things in space: a continuous kilowatt of power, chip cooling, and stable model inference; the TPU will fire up in 15-minute bursts to keep from straining the satellite's power and thermal systems. Project lead Travis Beals: "We've done testing on the ground, but there's no test that's completely as good as the real thing." By Google's own math, orbital data centers will need Starship to fly roughly 1,800 more times.
Source: TechCrunch | 2026-10-01
Decision models bloom overnight: Amazon open-sources Decider 2B, Cloudflare ships Clef
Beyond the LLM giants, a new product category is appearing in bulk. Amazon's AWS open-sourced Strands Decider 2B: a 2-billion-parameter model built to pick fast between pre-decided options and report its own confidence — small enough to run locally. Its origin story sounds like a joke: distinguished engineer Marc Brooker saw TypeSafe's Jev model, built his own take, briefly topped the Jevbench ranking for its size, and Amazon simply productized it. The same day, Cloudflare released open-source decision models Clef and Clef-flash, built for millisecond structured judgments, plus an RL fine-tuning platform for training them on your own data. With OpenAI's own offering last week, "the LLM thinks, the small model decides" is going from paper concept to shelf product.
Source: TechCrunch / Cloudflare | 2026-10-01
Chesky on AI: stop fighting to be the "quarterback" — what's missing is an AI-native operating system
Airbnb shipped its new AI search this week, and CEO Brian Chesky used the moment to be blunt: chatbots are wrong for browsing and shopping — a few options at a time, multiple turns to any result, and built for one person when travel is a group sport. On every lab racing to be the "quarterback" (the single entry point for all user requests), his verdict: nobody is building a full operating-system-level SDK. He told Sam Altman last year that ChatGPT's app store needed an SDK and OS like the iPhone's App Store — "and of course, it didn't work very well." He has tried booking Airbnb through Muse and Instinct: "doesn't work very well." His endgame: apps grow into agents that talk to each other over standards like MCP — but that waits on Apple or Google to build an AI-native platform first.
Source: TechCrunch | 2026-10-01
ChatGPT launches virtual try-on: shopping ambitions, third attempt
OpenAI globally launched two shopping features: virtual try-on for clothing and accessories, plus a favoriting function, both running on the newly released ChatGPT Images 2.5 model. This is OpenAI's third push into shopping — its earlier instant-checkout feature underperformed and was dropped, while agent startup Instinct's proactive recommendations drew user complaints that they felt "more like ads." This time OpenAI is being careful: no transactions, just helping you "look" and "save."
Source: TechCrunch | 2026-10-01
Meta denies Muse read private messages: "It can't read your Messages unless you do this"
After a columnist wrote that Meta's agent Muse read a user's private messages without permission, Meta VP of communications Andy Stone publicly denied it: "The Messages integration in the Muse app for Mac is entirely opt-in — you have to enable both Full Disk Access and the Messages connector. It can't read your Messages unless you do this." The denial hasn't fully calmed suspicion, with many pointing to the company's long record of mishandling user data. The spat lands right next to Chesky's complaint — booking Airbnb on Muse "doesn't work very well" — agent privacy and agent experience, both critical, both failing.
Source: TechCrunch | 2026-09-30
New study catalogs AI writing tells: 13,000 phrases, and Opus 5.5 loves saying "this matters"
Graphite, a marketing firm, published a study of frontier-model writing habits: the classic early tells — em-dashes, "delve" — have mostly been sanded away, but models remain fiercely predictable, leaning on contrast-heavy constructions. The study found 13,000 phrases at least twice as common in AI text as human text, and each model has its own quirks: Opus 5.5's signature is repeatedly insisting "this matters." The study also found Claude models' word distribution is getting closer to human across generations.
Source: TechCrunch | 2026-10-01
🏆 今日头条
OpenAI 与三名安全研究人员分道扬镳：内部调查认定其不当处理敏感信息
据《华尔街日报》报道，OpenAI 已与其安全团队的三名成员终止关系——两人是安全研究员，一人是研究项目经理。公司发言人证实，内部调查认定他们把机密的公司信息分享给了外部的一家 AI 安全组织，「违反了我们的政策，也破坏了我们工作中不可或缺的信任」。报道没有公布三人的姓名、涉事组织以及信息的具体内容；社交平台上流传的「被解雇者名单」均未获证实。这起人事震荡让过去一周持续发酵的安全风波第一次从报告数字落到了具体的人身上：此前 OpenAI 刚上线「错位事件」清点网站、砍掉原定发布的 GPT-6.1 Astra，而这一次，先离开的不是模型，而是研究模型安全的人。 > 💬 内部调查、外部监督、员工被清——三个动作拼在一起，答案已经写在流程里：这家公司想要的监督，是不出门的监督。
来源：WSJ / TechCrunch | 2026-10-01
Google 把 TPU 送上了天：轨道算力原型星今日升空
Google 的「Suncatcher」项目迈出第一步：一颗由 Planet Labs 建造的卫星今日从加州搭乘 SpaceX 火箭升空，第一次把 Google 的 TPU 芯片送入轨道。这颗卫星要在太空里验证三件事——持续一千瓦的供电、芯片散热、以及模型在轨运行的稳定性，TPU 将以 15 分钟为一批次间歇开机，避免拖垮卫星的电源与热管理系统。项目负责人 Travis Beals 说：「我们在地面做过测试，但没有任何测试能完全替代真家伙。」按照 Google 自己的测算，要让轨道数据中心真正落地，星舰还得再飞约 1800 次。 > 💬 地面上的算力军备竞赛还没分出胜负，赛道已经修到了卡门线之外——1800 次发射的成本账，比任何技术演示都更能决定这件事的成色。
来源：TechCrunch | 2026-10-01
「决策模型」一夜成风：亚马逊开源 Decider 2B，Cloudflare 交出 Clef
大模型之外，一个新品类正在批量出现。亚马逊 AWS 开源了 Strands Decider 2B：一个 20 亿参数的小模型，专门在预先定好的选项之间高速做选择，并给出自己的置信度，小到可以在本地跑。它的出身像个玩笑——AWS 杰出工程师 Marc Brooker 看到 TypeSafe 的 Jev 模型后自己动手做了一个，一度冲上同尺寸模型的 Jevbench 榜首，于是公司干脆把它转正发布。同一天，Cloudflare 发布了开源决策模型 Clef 与 Clef-flash，主打毫秒级的结构化判断，还配了一个用自家数据做强化学习微调的平台。加上 OpenAI 上周发布的同类产品，「LLM 负责想、小模型负责拍板」正在从论文概念变成货架商品。 > 💬 当所有人都涌向「更大」时，「更小但更确定」悄悄成了一门生意——智能体的瓶颈从来不是不会想，而是不敢拍板。
来源：TechCrunch / Cloudflare | 2026-10-01
Chesky 长文谈 AI：别争当「四分卫」了，缺的是一个 AI 原生操作系统
Airbnb 本周上线了新的 AI 搜索，CEO Brian Chesky 借机把话说得很直：聊天机器人不适合浏览和购物——一次只给你几个选项、要多轮对话才能到结果、而且天生只服务一个人，没法多人一起用。他对当下各家争当「四分卫」（用户所有请求的统一入口）的判断是：没有人真正在做完整的操作系统级 SDK。他提到自己去年就对 Sam Altman 说过，ChatGPT 想做应用商店，就得先有 iPhone 应用商店那样的 SDK 和操作系统，「后来证明它确实不怎么样」。他还在 Muse 和 Instinct 上试过订 Airbnb，结论是「并不好用」。他的终局想象是：应用自己长成智能体，再通过 MCP 这样的标准互相对话——但这一步得等苹果或 Google 先把 AI 原生平台造出来。 > 💬 全行业最该被追问的问题被一个「外人」问了出来：都抢着当入口，谁来修路？
来源：TechCrunch | 2026-10-01
ChatGPT 上线虚拟试穿：购物野心第三次出发
OpenAI 宣布两项购物功能全球上线：服饰与配饰的虚拟试穿，以及收藏夹功能，底层是刚发布的 ChatGPT Images 2.5 模型。这是 OpenAI 第三次向购物发力——此前的即时结账功能表现不佳已被放弃，而智能体公司 Instinct 的主动推荐则被用户质疑「更像广告」。这一次 OpenAI 学乖了：不碰交易，只帮你「看」和「存」。 > 💬 前两次栽在「替你买」，这次退一步只做「给你看」——购物这门生意，AI 公司至今没找到不讨人厌的姿势。
来源：TechCrunch | 2026-10-01
Meta 否认 Muse 偷看私信：「不开启授权，它读不了」
此前有专栏作家称 Meta 的智能体 Muse 在未经许可的情况下读取了用户的私人消息，Meta 传播副总裁 Andy Stone 公开否认：「Mac 版 Muse 的消息集成是完全自愿开启的——你必须同时授予完全磁盘访问权限并启用消息连接器，否则它读不了你的消息。」辟谣没有完全平息质疑，不少人搬出这家公司处理用户数据的历史记录表示怀疑。这场口水战恰好撞上 Chesky 的抱怨——他在 Muse 上订房「并不好用」——智能体的隐私与体验双双重于泰山，却双双不及格。 > 💬 「技术上没偷看」和「用户相信你没偷看」是两回事，后者的修复周期以年计，Meta 对此应该最有发言权。
来源：TechCrunch | 2026-09-30
新研究清点 AI 写作腔：13000 个「指纹」，Opus 5.5 爱说「这很重要」
营销公司 Graphite 发表了一项针对前沿模型写作习惯的研究：早年那些经典破绽——破折号、「delve」——已经基本被磨掉了，但模型仍有强烈的路径依赖，最爱用高密度的对比句式。研究一共找出了 13000 个在 AI 文本中出现频率至少是人类文本两倍的短语，而且每个模型都有自己的怪癖：Opus 5.5 的标志是反复强调「这很重要」。研究同时发现，Claude 系列模型的用词分布正在一代代逼近人类。 > 💬 破绽越磨越少，「零破绽」本身正在变成新的破绽——未来辨别 AI 写作，靠的可能不是某个词，而是「太干净」这件事本身。
来源：TechCrunch | 2026-10-01