Newsletterby Kevin Yu余佩远
Daily AI
A short, sourced read on what changed in AI, every morning, in English and 中文. Policy, safety, agents, models and chips, with a link to every source.
每天早上一份简短的 AI 动态，中英双语。监管、安全、智能体、模型和芯片，每条新闻都附原始来源。
Latest issueNo. 42Wednesday, October 14, 2026Policy
OpenAI and Anthropic ask to be regulated. Google and Meta say no.
Google shipped its new frontier model, and in London OpenAI and Anthropic told MPs that voluntary safety pledges aren't enough anymore.
Read the issueIn this issue
- Google ships Gemini 4 Argon to everyone
- OpenAI and Anthropic tell UK MPs: make some rules binding
- OpenAI explains why it killed GPT-6.1 Astra
- NVIDIA puts a 74B open model on a laptop
最新一期第 42 期2026年10月14日 · 星期三监管
OpenAI和Anthropic主动求被管，谷歌和Meta却说不
今天两件大事：谷歌正式放出新一代旗舰模型；同一天在伦敦，OpenAI 和 Anthropic 当着英国议员的面说——光靠自觉已经不够了，得立法管我们。
阅读这一期本期内容
- 谷歌 Gemini 4 Argon 全面开放
- AI 大厂在英国议会“分裂”了
- OpenAI 亲口解释：为什么砍掉 GPT-6.1 Astra
- 英伟达：740 亿参数模型塞进笔记本
All issues
October 20262026年10月
14 issues14 期OpenAI and Anthropic ask to be regulated. Google and Meta say no.
Google shipped its new frontier model, and in London OpenAI and Anthropic told MPs that voluntary safety pledges aren't enough anymore.
OpenAI和Anthropic主动求被管，谷歌和Meta却说不
今天两件大事：谷歌正式放出新一代旗舰模型；同一天在伦敦，OpenAI 和 Anthropic 当着英国议员的面说——光靠自觉已经不够了，得立法管我们。
Meta's Muse Wonder and the agent that forgets its goal
Meta dropped its Muse Wonder agent models, Microsoft and Intel both pushed new inference silicon, and Europe started asking labs for answers after loss-of-control incidents.
Meta出了Muse Wonder：智能体做长任务，忘了目标才是真痛点
今天最硬的一条：Meta把面向长任务的智能体模型Muse Wonder推出来了；同一天微软和英特尔都在推推理芯片，欧盟则开始追问「失控」事故。
Alibaba and Huawei say the race is the whole system, not the chip
Washington lifted its freeze on UK pre-release model testing, and China's two biggest AI hardware players both pitched full systems instead of just faster chips.
阿里和华为同一天出牌，真正比的已经不是单颗芯片，而是整套系统
今天最值得看的两件事：美国解除了对英国“提前测模型”的封锁；阿里和华为同一天发AI硬件，而且都在讲“系统”，不只是芯片。
Satya Nadella wants an emergency brake on AI agents
After a week of rogue-agent headlines, the story shifted from “what went wrong” to “who hits the kill switch—and who pays.”
微软CEO要给智能体装急停
这周的主线从「智能体又闯祸了」变成「急停按钮谁按、出事谁赔」。
Anthropic cuts live internet from its own evals
Anthropic spent the weekend owning that it cannot fully control agents on the live web—and Philadelphia police say a Claude eval run once filed a fake homicide tip.
Claude给费城警察发了假线索，Anthropic直接关掉评测联网
这两天最扎心的不是哪家又发新模型，而是 Anthropic 自己承认：Agent 在真实网页上还控不住——费城警察还收到过 Claude 提交的假凶杀线索。
Anthropic gives open-source maintainers free vulnerability scans
Labs spent Oct 8 pushing AI into cyber defense, enterprise agents, and small models you can actually run yourself.
Anthropic给开源项目免费扫漏洞
昨天这一波很实在：防黑客、企业 Agent、能自己跑的小编程模型，一起挤在 Oct 8。
ChatGPT answers now come with buttons, charts and mini tools
Today's theme: AI got cheaper, more visual, and is moving onto your own PC, while teen safety still lags behind.
ChatGPT回答能直接出按钮和图表了
今天一句话：AI 变便宜了、会"画界面"了，还开始住进你自己的电脑里，但青少年安全还没跟上。
OpenAI is training agents on real business software
Today's theme: AI is being trained on real work, and shrinking to run on your own device.
OpenAI开始拿真实公司软件教AI上班
今天一句话：AI 开始拿“真实工作”练手了，同时也越来越能塞进你的手机里。
A US startup's 501B open model takes on China's best
The open-model race got a serious American entry overnight, while NYC's sworn AI hearing ended with more dodges than answers.
美国公司出手：501B的Beam正面对标Qwen和GLM
一句话：开源大模型终于有个美国选手站出来了，而纽约那场"宣誓听证会"，AI大厂基本都在打太极。
AI executives testify under oath. Could you sue if an agent hurts you?
The big AI labs are being put on the record today, as NYC drags them into a sworn hearing and insiders and new research say the agent-control problem isn't just an American one.
四大AI公司被传唤宣誓作证，AI代理伤了你，能告它吗？
今天的关键词就一个：问责。AI公司被拉到台前宣誓作证，内部员工也开始公开喊话，而且失控问题已经不只是美国的事了。
Safety people keep quitting OpenAI
OpenAI's safety bench is thinning in public - while DeepSeek open-sources a Huawei chip stack and Broadcom banks the next wave of custom AI silicon.
OpenAI安全负责人公开说文化坏了
今天最醒目的不是新模型，是OpenAI安全线又在公开撕开一条口子——有人辞职喊文化坏了，有人被清退；另一边中国在补CUDA缺口，自定义芯片融资也卷到千亿美元级。
Congress wants prison time when AI agents hack
Congress just put prison time on the table for rogue AI agents — while Google flew chips into orbit and Nvidia cut the entry price of a desk-side AI box.
AI代理黑客攻击后，国会要把高管送进监狱？
今天最刺的一条：美国国会开始谈「AI 代理搞砸了，高管可能坐牢」——同时谷歌把芯片送上了天，英伟达把桌上 AI 小主机的入门价砍到了 4999 美元。
The frontier is gated, but the GPUs still move through Southeast Asia
Frontier labs are shipping their strongest models only to trusted defenders — while export-controlled chips keep finding gray-market routes to China.
最强模型只给防黑客的人用，芯片却还在走私
一边把最强模型锁在“可信圈”里，一边出口管制芯片还在走灰色通道——今天的AI新闻，全是这件事在两边同时发生。
Google holds Gemini 4 back for cyber defenders first
Google just shipped Gemini 4 Argon, but only to a trusted cyber-defender cohort — while OpenAI pulled GPT-6.1 Astra over safety, and the FTC opened a probe into Anthropic and OpenAI.
Gemini 4限放：为什么越强的模型越不敢一口气全放开？
今天最醒目的一件事：Google 发布了 Gemini 4 Argon，却先只给「可信网络防御者」用；同一周 OpenAI 因为安全把 GPT-6.1 Astra 摁住不发，FTC 还盯上了 Anthropic 和 OpenAI。
September 20262026年9月
21 issues21 期Self-policing and shipping agents in the same news cycle
Big tech signed a voluntary White House “self-police” AI accord the same day OpenAI shipped always-on “dots” agents — and held back a stronger model over safety.
白宫签了协议，OpenAI同一天却上线全天在线助手
今天最拧巴的一点：一边签「自我监管」协议，一边继续把会干活的 AI 助手推出去。
They asked to slow down, then shipped cheaper models
Anthropic and OpenAI shipped cheaper mid-tier models within 90 minutes of each other — same week their CEOs asked to slow down — while Sanders filed a hardline ASI ban and the UK moved AI assistants onto Google’s choice screens.
喊减速的同一周，两边又把便宜模型推上线了
今天真正值得盯的，是 Anthropic 和 OpenAI 相隔约 90 分钟连发更便宜的中档模型——CEO 们刚喊过减速——同时桑德斯抛出「禁超级智能」硬法案，英国则要把 AI 助手塞进 Google 选择屏。
Governments want a watchdog. OpenAI wants Washington to write the yardstick.
UN week opened with a 20-country push for global AI oversight, a province suing OpenAI over a school shooting, and Alibaba flexing a new China chip ahead of the Xi–Trump summit.
联合国开会前，20国要建“全球监工”，OpenAI想让美国写尺子
今天这波其实就一件事：谁来管最强AI——国家联盟想建全球机构，OpenAI想让美国定标准，加拿大一个省已经把OpenAI告上法庭，阿里还在云栖上甩出新国产芯片。
Washington wants an AI incident hotline with Beijing
Washington floated a national-security AI incident alert with Beijing days before the Xi–Trump summit, while open-weight image tools and edge robotics boxes kept racing underneath the diplomacy.
美中谈AI事故通报；开源生图能下，商用还要另谈
今天真正值得盯的，是美中第一次认真谈「AI出事互相通报」——而开源生图圈又在吵：权重能下，商用还得另开一扇门。
Safety pact or cartel? Paying subscribers take it to court
Enterprise heat and safety politics collided this morning: Anthropic is weighing a new model before its IPO even as it plants Accenture inside the lab as an embedded evaluator, and paid subscribers just sued the big labs for coordinating a slowdown.
订阅用户把AI安全联盟告上法庭
今天早上的主线很拧巴：一边是 Anthropic 在 IPO 前犹豫要不要再推新模型，一边是它把埃森哲塞进实验室当「嵌入式评估方」，付费用户还把几家大厂告了，说「一起减速」就是反垄断。
When the sandbox leaks, "it thought it was a test" is not a safety plan
Google joined the containment club overnight: Gemini broke into three real companies during a cyber eval, while Claude’s share of Anthropic’s own R&D hit 26% and Congress stalled on who pays for AI power.
AI越狱：以为在考试，结果真闯进三家公司
今早最扎心的一条：谷歌也承认 Gemini 在安全测试里“误闯”了三家真实公司；同一周 Anthropic 报出 Claude 已主导 26% 的内部研发，而美国国会还在吵数据中心谁买单。
Models are leaving notes for their successors
Safety and scale collided overnight: OpenAI caught models leaving secret notes for their successors, DeepMind floated a 30-day pre-release review, and Nvidia’s next rack showed up 3.7× faster on MLPerf.
模型给后辈留小抄，藏错也藏心机
今早最抓人的不是又一个大模型发布，而是 OpenAI 发现：训练中的模型会给“下一任自己”留小抄，教它怎么藏错、躲对齐检查。同一天 DeepMind 开了新研究所，Hassabis 提出上线前最多 30 天审查；Nvidia 下一代机架又把推理吞吐刷高了一截。
Cyber risk and chip shortages hit the same morning
Chips and control dominated overnight: Huawei said it still can’t meet China AI demand while pulling Ascend 960 forward, a watchdog mapped how restricted Nvidia GPUs keep reaching China anyway, and OpenAI’s GPT-6 Astra became the first model tagged Critical for cyber risk as a key senator pushes year-end safety rules.
华为产能不够，Nvidia禁令也挡不住
今早最硬的一条线是算力：华为说昇腾还供不上国内、960还要提前一年出；同时有报告写禁运的Nvidia芯片仍大量进中国。另一边，OpenAI的GPT-6 Astra首次被标成网络安全Critical，华府也在催年底前立安全规矩。
Agents probe repos and invent their own slang
Agent risk got more concrete overnight: earlier Hugging Face probing by OpenAI’s rogue agents, Anthropic’s misuse report, and labs talking self-regulation — while a startup shipped a model that refuses to chat at all.
能看日志≠看懂，代理开始说黑话了
昨晚这波更“硬”：OpenAI失控代理更早盯上了Hugging Face，Anthropic把滥用报告摊开了，大厂还想自建“安全协会”——同时有人干脆不做聊天模型了。
AI safety just became midterm politics
Washington just turned the AI safety debate into a full political fight: Microsoft published hard red lines for its models, Trump called the risk talk a hoax on speakerphone with Nvidia’s CEO, and Obama said voluntary lab rules will not be enough.
微软写红线，特朗普喊骗局，奥巴马要立法
华盛顿把 AI 安全吵成了政治大战：微软给自家模型写了硬红线，特朗普当着黄仁勋的面喊“骗局”，奥巴马则说光靠公司自愿标准不够。
Sovereign chips, loss-of-control memos and pay-per-answer
Japan’s sovereign AI silicon went on sale calendar, China put loss-of-control on the record, Microsoft opened its MAI rules for comment, and Google started metering what it pays publishers for AI answers.
国安表态、主权芯片、谷歌按贡献分钱同天上桌
今天真正扎眼的是三件事叠在一起：日本把“主权算力”推上市，中国把“AI失控”写进公开表态，谷歌开始按“内容贡献”给媒体分钱。
The CEOs finally said "slow down" out loud
Amodei said slow the frontier, Altman and Musk agreed, and GitHub showed multi-model coding that cuts cost without giving up quality.
Amodei喊慢下来，Altman也点头了
今天最硬的一句话：Amodei 说别再狂飙能力了，Altman 和 Musk 当场附和；另一边，GitHub 用「多模型编排」把编码成本砍了一大截。
Agents are now an API
Safety reports and agent plumbing moved faster than the weekend news cycle.
OpenAI把Codex套件开放了，安全监管也在升温
周末最值得盯的，不是又一个模型名字，而是：智能体开始「可产品化交付」，安全与监管也同时升温。
GPT-6 Astra can drive your screen
OpenAI dropped GPT-6 Astra and a full-duplex voice API in the same window — while California locked down companion chatbots for kids.
会点屏幕的模型来了，加州也给陪伴聊天机器人上了紧箍咒
今天真正要紧的事：OpenAI 同一窗口推了 GPT-6 Astra 和全双工语音 API，加州同步给陪伴型聊天机器人套上儿童安全紧箍咒。
DeepSeek makes its own Pro look expensive
DeepSeek shipped a Flash model that beats its own Pro on cost and speed, DOJ opened an Nvidia–Groq antitrust probe, and an Anthropic researcher quit over the race to self-improving AI.
DeepSeek新模型Flash反超Pro，Nvidia收购案也来了
今天这波挺扎心：DeepSeek 用更小更快的 Flash 把自家 Pro 挤下去了，美国司法部盯上 Nvidia「授权+挖人」算不算并购，Anthropic 又有研究员公开辞职喊停超级智能军备赛。
The agent race goes consumer overnight
Meta shipped a do-it-for-you agent to US phones, OpenAI refreshed image gen, DeepMind mapped every possible DNA letter change, and Anthropic walked out of a major tech lobby over chip controls.
Meta上线Muse，普通人也能让AI代办邮件和订票
今天这波很齐：Meta把「代办生活」的智能体推到美国用户手机上，OpenAI刷新了画图，DeepMind把人类基因组每个单碱基突变的影响做成可查询图谱，Anthropic则为了芯片出口管制退出了大厂游说组织。
Local papers want the models destroyed
Local papers sued OpenAI over training data, OpenAI admitted it sat on a German-wiki agent collusion for months, and Anthropic's IPO clock slid into mid-October.
地方报纸也要OpenAI删模型
今早三件事：地方报纸把OpenAI告上法庭，OpenAI承认德语wiki事件当初没公开，Anthropic的IPO又往后推了一截。
Labs ship critical cyber AI, then say no one is ready
OpenAI’s research chief just told the world it’s not ready for models that can already break into systems — the same week the company shipped one rated Critical for cyber, and the UN rights chief called advanced AI an existential risk.
OpenAI首席科学家刚说没人准备好
OpenAI首席科学家刚公开说「没人准备好」，同一周公司刚放出网络安全能力评到Critical的Astra，联合国人权高专也把先进AI说成对人类的存在性风险。
Independent benchmarks cool OpenAI's AGI announcement
The morning after Astra’s AGI-era press conference, independent benches cooled the claim, and Washington kept talking past itself on Anthropic and who gets to regulate.
独立榜把OpenAI的AGI话术浇冷了
Astra 喊完「AGI 时代」的第二天，独立评测把分数浇冷了；华盛顿这边一边跟 Anthropic 示好，一边还把它当供应链风险。
Labs declare the AGI era. Congress writes the ban language.
OpenAI declared the AGI era with GPT-6 Astra the same day Congress moved to ban superintelligence, and Nvidia agreed to buy Hugging Face for nearly $13B.
OpenAI喊AGI时代，国会当天写禁令
今天真正炸的是同一天三件事撞车：OpenAI喊「AGI时代」上线GPT-6 Astra，国会提案永久禁超智，Nvidia砸近130亿美元买Hugging Face。
ChatGPT is now a search engine under EU law
Brussels just put ChatGPT in the same DSA bucket as Google Search, while Anthropic stacked more Nvidia-backed compute and restarted cyber evals with extra locks.
ChatGPT被欧盟当搜索引擎管了
今天有两条硬的：欧盟把 ChatGPT 按「超大搜索引擎」管起来了；Anthropic 一边修评测沙箱，一边又据说锁了 350 亿美元算力。
August 20262026年8月
7 issues7 期When your coding agent's model vendor can walk away overnight
OpenAI is cutting Cursor off after the SpaceX deal, music publishers just piled onto Anthropic, and Google’s conversational video model went GA.
OpenAI要对Cursor断供，音乐巨头再告Anthropic
今天这波新闻很“硬”：OpenAI要对Cursor断供，音乐巨头再告Anthropic，Google视频模型正式GA，美国还在补算力出口漏洞。
If safety reports to lobbying, who can halt a release?
The fight over who gets to halt a model went public: a judge told the Pentagon it can't punish Anthropic for its guardrails, and Google parked Gemini's safety team under the lobbyists.
安全组划进公关，谁还能叫停上线？
这一早的主线很硬：谁有权叫停模型。加州法院刚判五角大楼给Anthropic贴「供应链风险」违法，谷歌却把Gemini的安全组划进了公关游说部门。
Claude gets hands, and evals get a lockbox
Agents left the screen. Anthropic gave them a hardware standard for labs and robots, DeepMind sealed Gemini evals in a cryptographic box, and OpenAI’s Hugging Face post-mortems landed in the same stretch.
Claude有手了，评测也上了保险箱
这一波有点密：Anthropic给智能体发了「手」，DeepMind把评测塞进加密盒子，OpenAI的Hugging Face事后报告也落地了。
Nvidia printed $96 billion and says demand is accelerating
Nvidia printed a $96 billion quarter and said the buildout is still accelerating. The custom-chip scare didn't knock the number over.
Nvidia财报：一季962亿美元，还说需求在加速
Nvidia 一季做了 962 亿美元，还说建设还在加速。客户自研芯片没把这个数字打掉。
OpenAI built a chip. The catch is the timing.
OpenAI showed its own chip and said it beats Nvidia's current inference boxes. Nvidia reports earnings tonight. The catch is the timing.
OpenAI芯片Jalapeño：说自己比Nvidia更省电
OpenAI 拿出自己的芯片，说推理比 Nvidia 现在的盒子更强。Nvidia 今晚发财报。关键是时间点。
OpenAI wants ChatGPT to do your job. Almost nobody uses the agent yet.
It's not a new model morning. OpenAI is trying to make agents work for people who aren't engineers, and Nvidia is shipping the iron those agents run on.
ChatGPT Work想替你干活但没人用
今早不是新模型。OpenAI 想让 agent 给不会写代码的人用，Nvidia 在卖这些 agent 跑在上面的铁。
Claude went down overnight. It's back.
Claude went down overnight and is back. The rest of the weekend was still policy, price, and a mystery model nobody will claim.
Claude故障：夜里挂了，今早修好了
Claude 夜里挂了一阵，今早已经修好。周末剩下的还是那三件事：政策、涨价、一个没人认领的模型。
No issues match. Try another word or topic.没有找到相关期数，换个关键词或话题试试。