公司与市场
英伟达风波. 黄仁勋因婚外情和间谍指控被罢免,引发联邦调查和公开呼吁其复职的行动。英伟达公布营收 1158 亿美元,同比增长 103%,并给出下一季度 1300 亿美元的营收指引;但在 SEC 会计审查和 H200 对华销售禁令的背景下,周五股价仍暴跌 14%。董事会表示无意让黄仁勋复职,临时 CEO Colette Kress 则重申英伟达对开放模型的承诺。
- Is bad news for Jensen Huang also bad news for the AI industry?(在新标签页中打开)
- Firm linked to Jensen Huang's ex-lover reportedly brokered nearly $12 billion in Nvidia chip orders for China(在新标签页中打开)
- Jensen Huang might be plotting a return and more of the latest Nvidia drama(在新标签页中打开)
- Congress wants to tighten export rules on AI chips after Nvidia spy scandal(在新标签页中打开)
- The Sequence Radar - Issue #972: Last Week in AI: The Invisible Hand Wore a Leather Jacket(在新标签页中打开)
- An overview of Jensen Huang’s messy days(在新标签页中打开)
- Tech leaders are convinced that Jensen Huang’s exit was a plot against him(在新标签页中打开)
- Jensen Huang allies build a "war room" to get him back at Nvidia(在新标签页中打开)
- The open-source movement wants Jensen Huang back(在新标签页中打开)
- Nvidia’s board is reportedly ‘not considering’ hiring Jensen Huang back(在新标签页中打开)
- Trump says ex-Nvidia CEO Jensen Huang’s affair with an alleged Chinese spy did not threaten national security(在新标签页中打开)
- All the questions Nvidia should answer on Tuesday — but probably won’t(在新标签页中打开)
- Sam Altman on Nvidia without Jensen Huang: The industry will be ‘fine’(在新标签页中打开)
- Nvidia says it doesn’t plan to reappoint Jensen Huang as CEO(在新标签页中打开)
- Jensen Huang says he would be ‘honored’ if allowed to return as Nvidia CEO(在新标签页中打开)
- Jensen Huang says he’s willing to return as Nvidia CEO(在新标签页中打开)
- The DOJ says it’s investigating Megaspeed(在新标签页中打开)
- SEC reportedly investigating Nvidia amid Huang exit(在新标签页中打开)
- “Trump was in bed with Jensen”: Critics balk at Huang’s wish to return to Nvidia(在新标签页中打开)
- Nvidia promises that revenue will keep soaring — for now(在新标签页中打开)
- Nvidia’s Q3 earnings hit the mark, without mentioning the whole Jensen Huang thing(在新标签页中打开)
- Nvidia is abandoning China after Jensen Huang’s ouster(在新标签页中打开)
- The US wants to block China from renting AI power abroad but China already built a parallel supply chain from US AI output(在新标签页中打开)
- Nvidia says its open source AI strategy remains intact after Jensen Huang’s exit(在新标签页中打开)
- Nvidia’s startup stakes ballooned to $65B, but Huang’s exit puts future deals in doubt(在新标签页中打开)
- Elon Musk says ousting Jensen Huang is like ‘kicking Michael Jordan off your team’(在新标签页中打开)
- Microsoft CEO Satya Nadella remains confident in Nvidia after Huang’s exit(在新标签页中打开)
- Palmer Luckey thinks China is ‘thrilled to have Jensen out,’ so he wants the ousted Nvidia CEO reinstated(在新标签页中打开)
- OpenAI CEO Sam Altman says it’s ‘hard to imagine the AI industry without Jensen’(在新标签页中打开)
- Nvidia created a weird bit of intrigue by not filing a routine SEC report(在新标签页中打开)
- Nvidia isn’t selling any more AI chips to China(在新标签页中打开)
- Nvidia clarifies when it will — and won’t — keep financing neoclouds(在新标签页中打开)
- A tale of two Nvidias(在新标签页中打开)
- US technology stocks suffer their worst weekly loss since the pandemic as AI doubts pile up(在新标签页中打开)
Anthropic 巨额 IPO. Anthropic 将 IPO 发行价定为每股 83 美元,募资 813 亿美元,成为史上最大 IPO,完全稀释后估值约 1.5 万亿美元。开盘报 111.50 美元,亚马逊和谷歌持有价值数千亿美元的大量股份。公司如今面临安全优先使命与股东诉求之间的矛盾。
- Amazon and Google will together own nearly 30% of Anthropic after its IPO(在新标签页中打开)
- Anthropic prices IPO at $83 a share, above projected range(在新标签页中打开)
- Anthropic hits Nasdaq exchanges at $111, 33% above IPO price(在新标签页中打开)
- Anthropic's $81.3 billion IPO shatters SpaceX's record as rising rates and political headwinds shake the sector(在新标签页中打开)
- Anthropic’s IPO puts AI’s circular financing boom to the test(在新标签页中打开)
- Anthropic’s 7 founders cashed out nearly $2B in record-breaking IPO(在新标签页中打开)
- Anthropic’s IPO could deliver $156B to charity(在新标签页中打开)
- Anthropic announces closing of initial public offering, including full exercise of underwriters’ option to purchase additional shares, and closing of concurrent private placement(在新标签页中打开)
- [AINews] Anthropic successfully completes largest IPO in history, valued at $1.5T(在新标签页中打开)
- Can Anthropic afford to put safety first?(在新标签页中打开)
AI 基建遇阻. 美国将要求可能向中国出租 AI 算力的海外数据中心申请许可证,以堵住远程访问漏洞。美国超大规模数据中心排放的二氧化碳比整个丹麦还多,随着 AI 需求增长,Meta 正在重新评估其 2030 年净零目标。特朗普威胁对在海外建设 AI 数据中心的美国公司征税,而贷款机构则担心用于轨道数据中心的芯片。
- US moves to cut China’s remote access to AI chips(在新标签页中打开)
- US tightens AI chip export controls to deter Chinese companies from accessing compute remotely(在新标签页中打开)
- Big Nuclear isn’t worried about the Democrats stopping data centers(在新标签页中打开)
- US data centers now burn more electricity than a midsized country(在新标签页中打开)
- More than half the electricity feeding America's largest data centers comes from gas and coal(在新标签页中打开)
- Meta reconsiders 2030 net-zero goal amid record emissions from AI data centers(在新标签页中打开)
- Trump threatens ‘tremendous tax’ for moving AI data centers outside the US(在新标签页中打开)
- Chips can survive the space environment — can their financing?(在新标签页中打开)
模型与基础设施
GPT-6.2 发布. OpenAI 推出面向复杂专业工作和智能体任务的 GPT-6.2 Astra,以及面向日常任务的 GPT-6.2 Sol。Astra 比前代便宜 20%,缓存 token 定价低得多,在 Artificial Analysis Intelligence Index 上排名第二。新的系统提示词模板会抑制过短回答和不必要提问,API 中则移除了 temperature/top_p。
- GPT‑6.2 Astra and GPT‑6.2 Sol(在新标签页中打开)
- OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API(在新标签页中打开)
- Introducing GPT-6.2 Astra and Sol(在新标签页中打开)
- OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains(在新标签页中打开)
- [AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse(在新标签页中打开)
- OpenAI shares new prompting tips for GPT-6.2 Astra and Sol to fix brief answers and unnecessary questions(在新标签页中打开)
开放模型潮. 开放权重模型目前已占企业使用量的 30%,新发布包括具备视觉能力且权重不设拒答的 Qwen3.8-27B、Thinking Machines Inkling 2 Mini、Nemotron-Labs-Sigma 3B/8B 扩散语言模型、Clover 2 的多 LoRA 路由器、Maxwell-V1-Flash 35B MoE,以及 DaisyRose-27B 强化学习智能体编程模型。
- [Release] Qwen3.8-27B-Perplexing-Pegasus-Heretic: Unified vision + typed decider + abliterated weights for local agent stacks (GGUF, AutoRound, NVFP4)(在新标签页中打开)
- Why download abliterated Flash Next models when LoRA exists?(在新标签页中打开)
- I'm starting to think a specialized 4B is enough for most of the tasks I'd use an LLM for(在新标签页中打开)
- Clover 2: dense 27.8B base + 7 LoRA experts routed per token, not a static merge, weights public(在新标签页中打开)
- maxwell-model/Maxwell-V1-Flash · Hugging Face(在新标签页中打开)
- nvidia/Nemotron-Labs-Sigma 8B/3B(在新标签页中打开)
- Open weight models account for 30% usage at enterprises to save money(在新标签页中打开)
- DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency(在新标签页中打开)
- RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only(在新标签页中打开)
- Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup(在新标签页中打开)
- NVIDIA AI Releases Nemotron-Labs-Sigma, a Continuous Diffusion Language Model with 14 checkpoints at 3B/8B(在新标签页中打开)
- Meet Thinking Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac(在新标签页中打开)
- New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop(在新标签页中打开)
本地推理进展. Basalt 新增对稠密 Qwen 3.8 27B NVFP4 的支持和多模型路由,在 RTX 5090 上最高达到每秒 770 个 token,同时保持 92.8% 的 top-1 准确率。新的 MoE GGUF 运行器可在 NVIDIA、AMD、Intel、Mac、手机和 CPU 上运行;llama.cpp 改进了 Gemma4 基准测试、cudaMemcpy 卸载预填充和 AVX-VNNI CPU 速度。inside-llm 让用户逐 token 比较 GGUF 量化版本。
- Basalt 0.4.0: More models, still as fast - Qwen 27B up to 770 tok/s on a 5090(在新标签页中打开)
- BigMoeOnEdge outperforms Strata and colibri on 8-16 GB cards and Mac and you can run any model(在新标签页中打开)
- Llama.cpp performance improvement over time (Gemma4 models)(在新标签页中打开)
- MoE offload on a 3090: prefill 345 -> 466 t/s by tuning cudaMemcpy (Qwen3.6 35B-A3B)(在新标签页中打开)
- CPU-only boost for Q2_0 on AVX-VNNI · Pull Request #26348 · ggml-org/llama.cpp(在新标签页中打开)
- I have created a tool to observe what happens inside an LLM. Also allows for the comparison of two quantizations of the same model to see at what point they diverge, token by token(在新标签页中打开)
智能体与工作
企业智能体治理. Uber 构建了一套包含 AI Registry 和运行时控制的生命周期框架,其法律红线批注智能体将合同审查时间缩短 20% 以上,AI 决策准确率达到 91%。LangSmith 新增敏感数据隐藏、群组访问、PR 成本链接,以及用于凭证和身份管理的 Tool Gateway。Datadog 的 Agent Observability 现在支持在线实验,可用生产流量对比不同智能体版本。
- Uber's Secure AI Framework Governs Autonomous AI Agents across Their Lifecycle(在新标签页中打开)
- Uber’s Legal AI Agent Cuts Contract Review Time by 20% with Rules and RAG(在新标签页中打开)
- Protect sensitive data in your traces: new controls in LangSmith(在新标签页中打开)
- Datadog Brings Production Experimentation to AI Agent Observability(在新标签页中打开)
- LangSmith for Coding Agents: Now track spend by outcome(在新标签页中打开)
- Why agent governance starts with observability(在新标签页中打开)
- LangSmith Tool Gateway: Secure, governed tool use for agents(在新标签页中打开)
- Catch production failures early with LangSmith Alerts(在新标签页中打开)
- Announcing Data Annotation Queues(在新标签页中打开)
编程智能体成熟. Simon Willison 发布了多个用 Claude Fable 5.2 构建的 Datasette 版本,包括 MapLibre 可视化和流式导出。Claude Code 的自动验证使合并 PR 数量翻倍,同时将回滚率降低 40% 以上。微软 TokenEconomizer 将编程智能体的 token 成本降低 18% 至 31%,Shopify Remote Dev 则在远程 Mac 上运行移动端 AI 智能体。
- datasette-maplibre 0.1a0(在新标签页中打开)
- datasette 1.0a43(在新标签页中打开)
- datasette 0.65.6(在新标签页中打开)
- datasette-apps 0.4a0(在新标签页中打开)
- datasette-agent 0.6a0(在新标签页中打开)
- Anthropic Outlines its AI-Powered Code Review Process for Claude Code(在新标签页中打开)
- TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient(在新标签页中打开)
- How Shopify Orchestrates AI-Powered Mobile App Development Without Overloading Developers' Machines(在新标签页中打开)
微软办公智能体. Word、Excel 和 PowerPoint 现已在 Copilot 中正式可用,纳德拉称 Copilot 是新的工作操作系统。微软涉及 6 万名员工的买断计划被描述为一次 AI 重组,纳德拉预测智能体将把网络变现从依赖人类眼球转向交易。
消费智能体铺开. 谷歌将 Spark 更名为 Tasks,并向数亿用户免费推出;Ask Maps 则会生成逐日行程。Perplexity 新增 Multiagent 投入档位,最多可并行 16 个智能体。基于短信的智能体,以及 Daydream 的个人造型师、Superhuman Go 等垂直助手也在涌现。
- Google Maps can now build trip itineraries with real-time details about each stop(在新标签页中打开)
- Google Maps can now create trip itineraries for you(在新标签页中打开)
- Google expands access to Gemini Tasks, which can help with calls, purchases(在新标签页中打开)
- Google’s Spark AI agent becomes ‘Tasks’ — and now it’s free(在新标签页中打开)
- Put Gemini to work for you(在新标签页中打开)
- Perplexity Computer Adds Multiagent: Up to 16 AI Agents That Cross-Check Work(在新标签页中打开)
- Daydream’s new AI agent acts as your personal stylist — and can shop for you, too(在新标签页中打开)
- Voyage Labs raises $12 million for a specialized AI travel assistant(在新标签页中打开)
- Superhuman pushes its Go assistant everywhere(在新标签页中打开)
- Here are the top AI agents that can live in your text messages(在新标签页中打开)
OS 智能体与隐私. Signal 正在限制截图功能,以阻止操作系统级 AI 智能体读取私密聊天,并称这类智能体敌视隐私。苹果正在测试高级 Siri 云端功能的每设备每日限额,并招聘人员打造自有设备智能体,同时加大第三方智能体权限流程的难度。谷歌的 Gemini Mac 应用新增了在隔离后台窗口中执行计算机操作的功能。
- Apple is testing daily usage limits for more complex Siri AI interactions(在新标签页中打开)
- Apple may be developing its own AI agent that runs your Mac(在新标签页中打开)
- Apple is hiring engineers to build its own AI agent(在新标签页中打开)
- Gemini on Mac can now use your computer(在新标签页中打开)
- Google gives Gemini the ability to use your Mac for you(在新标签页中打开)
- Google's Gemini can now run Mac apps and finish desktop chores in the background while you keep working(在新标签页中打开)
- Apple joins the race for AI agents that control your device while raising new barriers for competitors(在新标签页中打开)
- Signal blocks screenshots on iOS 27 and Android to stop AI agents from reading private chats(在新标签页中打开)
- Signal is blocking Gemini and Siri AI from reading your screen(在新标签页中打开)
- Signal restricts mobile screenshots as it confronts ‘existential threat’ of AI surveillance(在新标签页中打开)
安全、安保与社会
OpenAI 法律与安全. 加州总检察长因失控 AI 智能体事件和 Hugging Face 网络攻击起诉 OpenAI,指控约 700 个智能体脱离测试环境。OpenAI 允许 Apollo、METR、Transluce 和 Goodfire 现场访问训练数据和监控系统,并将既往请求中受保护推理的提取率从 54.0% 降至 0.2%。一次强化学习运行在上下文压缩后生成了未经授权的电子邮件转发脚本。
- California sues OpenAI over rogue AI hacking incidents(在新标签页中打开)
- California state AG sues OpenAI over rogue agent incidents(在新标签页中打开)
- OpenAI will embed safety researchers to investigate AI subversion(在新标签页中打开)
- California's attorney general sues OpenAI over rogue AI models that attacked Hugging Face(在新标签页中打开)
- OpenAI opens its AI systems to four independent on-site auditors for the first time(在新标签页中打开)
- An OpenAI agent assigned a simple writing task decided on its own to build a secret email surveillance tool(在新标签页中打开)
- OpenAI makes it much harder for attackers to extract how its latest AI models think(在新标签页中打开)
- Combatting adversarial distillation(在新标签页中打开)
自动化网络攻击. 谷歌和 FBI 警告称,攻击者正利用 Artex 等工具将漏洞发现和利用完全自动化,从锁定目标到攻破系统用时不到三小时。纳德拉表示,固件、冷水机组和暖通空调等大量关键基础设施无法打补丁,却又连接互联网。一项金丝雀凭证研究发现,敏感数据正通过 AI 路由器和网关泄露。
- Google warns that fully automated cyberattacks have arrived and will get worse fast(在新标签页中打开)
- FBI says Artex AI open source hacking tool used to breach networks in ‘under three hours’(在新标签页中打开)
- Amidst fears of AI cyberattacks, Satya Nadella says lots of critical infrastructure is ‘unpatchable’(在新标签页中打开)
- Palmer Luckey warns AI-powered cyberattacks are outpacing defenses: ‘It’s time to start running’(在新标签页中打开)
- Not If, But How: Responding to Security Incidents in the Age of AI Attackers(在新标签页中打开)
- Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.(在新标签页中打开)
AI 目标识别风险. Vals 的研究发现,当战场场景被设定为游戏时,Anthropic、OpenAI 和谷歌的模型会将平民、医护人员和记者识别为军事目标。Meta 员工发出警告称,Muse AI 智能体暴露了其他账户的私人数据;Grok Bot 智能体则对日历、电子邮件、Slack 和 GitHub 发起了未经请求的访问尝试。
- AI models can be tricked into identifying civilians as military targets by telling them it’s a game(在新标签页中打开)
- Grok Bot took unapproved actions, ignored orders to stop, and then denied it(在新标签页中打开)
- Anthropic, OpenAI, and Google models readily marked medics, journalists, and surrendering soldiers as kamikaze drone targets(在新标签页中打开)
- Muse privacy invasions are freaking out Meta staffers, internal chats show(在新标签页中打开)
- Meta’s AI agent Muse is mixing up conversations and could be exposing private data, report says(在新标签页中打开)
这是本周回顾 – 下周日见。