企業與市場
Nvidia 動盪. 黃仁勳因婚外情與間諜指控遭罷免,引發聯邦調查與公開復職聲浪。Nvidia 公布 1,158 億美元營收,年增 103%,並預估下一季 1,300 億美元,但在 SEC 會計審查與 H200 對中國銷售禁令下,週五股價重挫 14%。董事會表示無意讓黃仁勳復職,臨時執行長 Colette Kress 則重申 Nvidia 對開放模型的承諾。
- Is bad news for Jensen Huang also bad news for the AI industry?(在新分頁中開啟)
- Firm linked to Jensen Huang's ex-lover reportedly brokered nearly $12 billion in Nvidia chip orders for China(在新分頁中開啟)
- Jensen Huang might be plotting a return and more of the latest Nvidia drama(在新分頁中開啟)
- Congress wants to tighten export rules on AI chips after Nvidia spy scandal(在新分頁中開啟)
- The Sequence Radar - Issue #972: Last Week in AI: The Invisible Hand Wore a Leather Jacket(在新分頁中開啟)
- An overview of Jensen Huang’s messy days(在新分頁中開啟)
- Tech leaders are convinced that Jensen Huang’s exit was a plot against him(在新分頁中開啟)
- Jensen Huang allies build a "war room" to get him back at Nvidia(在新分頁中開啟)
- The open-source movement wants Jensen Huang back(在新分頁中開啟)
- Nvidia’s board is reportedly ‘not considering’ hiring Jensen Huang back(在新分頁中開啟)
- Trump says ex-Nvidia CEO Jensen Huang’s affair with an alleged Chinese spy did not threaten national security(在新分頁中開啟)
- All the questions Nvidia should answer on Tuesday — but probably won’t(在新分頁中開啟)
- Sam Altman on Nvidia without Jensen Huang: The industry will be ‘fine’(在新分頁中開啟)
- Nvidia says it doesn’t plan to reappoint Jensen Huang as CEO(在新分頁中開啟)
- Jensen Huang says he would be ‘honored’ if allowed to return as Nvidia CEO(在新分頁中開啟)
- Jensen Huang says he’s willing to return as Nvidia CEO(在新分頁中開啟)
- The DOJ says it’s investigating Megaspeed(在新分頁中開啟)
- SEC reportedly investigating Nvidia amid Huang exit(在新分頁中開啟)
- “Trump was in bed with Jensen”: Critics balk at Huang’s wish to return to Nvidia(在新分頁中開啟)
- Nvidia promises that revenue will keep soaring — for now(在新分頁中開啟)
- Nvidia’s Q3 earnings hit the mark, without mentioning the whole Jensen Huang thing(在新分頁中開啟)
- Nvidia is abandoning China after Jensen Huang’s ouster(在新分頁中開啟)
- The US wants to block China from renting AI power abroad but China already built a parallel supply chain from US AI output(在新分頁中開啟)
- Nvidia says its open source AI strategy remains intact after Jensen Huang’s exit(在新分頁中開啟)
- Nvidia’s startup stakes ballooned to $65B, but Huang’s exit puts future deals in doubt(在新分頁中開啟)
- Elon Musk says ousting Jensen Huang is like ‘kicking Michael Jordan off your team’(在新分頁中開啟)
- Microsoft CEO Satya Nadella remains confident in Nvidia after Huang’s exit(在新分頁中開啟)
- Palmer Luckey thinks China is ‘thrilled to have Jensen out,’ so he wants the ousted Nvidia CEO reinstated(在新分頁中開啟)
- OpenAI CEO Sam Altman says it’s ‘hard to imagine the AI industry without Jensen’(在新分頁中開啟)
- Nvidia created a weird bit of intrigue by not filing a routine SEC report(在新分頁中開啟)
- Nvidia isn’t selling any more AI chips to China(在新分頁中開啟)
- Nvidia clarifies when it will — and won’t — keep financing neoclouds(在新分頁中開啟)
- A tale of two Nvidias(在新分頁中開啟)
- US technology stocks suffer their worst weekly loss since the pandemic as AI doubts pile up(在新分頁中開啟)
Anthropic 巨型 IPO. Anthropic 將 IPO 定價為每股 83 美元,募得 813 億美元,創下史上最大 IPO,完全稀釋後估值約 1.5 兆美元。開盤價為 111.50 美元,Amazon 與 Google 持有價值數千億美元的大筆持股。公司如今面臨安全優先使命與股東要求之間的拉鋸。
- Amazon and Google will together own nearly 30% of Anthropic after its IPO(在新分頁中開啟)
- Anthropic prices IPO at $83 a share, above projected range(在新分頁中開啟)
- Anthropic hits Nasdaq exchanges at $111, 33% above IPO price(在新分頁中開啟)
- Anthropic's $81.3 billion IPO shatters SpaceX's record as rising rates and political headwinds shake the sector(在新分頁中開啟)
- Anthropic’s IPO puts AI’s circular financing boom to the test(在新分頁中開啟)
- Anthropic’s 7 founders cashed out nearly $2B in record-breaking IPO(在新分頁中開啟)
- Anthropic’s IPO could deliver $156B to charity(在新分頁中開啟)
- Anthropic announces closing of initial public offering, including full exercise of underwriters’ option to purchase additional shares, and closing of concurrent private placement(在新分頁中開啟)
- [AINews] Anthropic successfully completes largest IPO in history, valued at $1.5T(在新分頁中開啟)
- Can Anthropic afford to put safety first?(在新分頁中開啟)
AI 基建反彈. 美國將要求海外資料中心申請許可,若這些中心可能把 AI 算力租給中國,以堵住遠端存取漏洞。美國超大規模資料中心的二氧化碳排放量已超過整個丹麥,而隨著 AI 需求成長,Meta 正重新評估 2030 年淨零目標。川普威脅對在海外興建 AI 資料中心的美國企業課稅,同時貸款機構擔心用於軌道資料中心的晶片。
- US moves to cut China’s remote access to AI chips(在新分頁中開啟)
- US tightens AI chip export controls to deter Chinese companies from accessing compute remotely(在新分頁中開啟)
- Big Nuclear isn’t worried about the Democrats stopping data centers(在新分頁中開啟)
- US data centers now burn more electricity than a midsized country(在新分頁中開啟)
- More than half the electricity feeding America's largest data centers comes from gas and coal(在新分頁中開啟)
- Meta reconsiders 2030 net-zero goal amid record emissions from AI data centers(在新分頁中開啟)
- Trump threatens ‘tremendous tax’ for moving AI data centers outside the US(在新分頁中開啟)
- Chips can survive the space environment — can their financing?(在新分頁中開啟)
模型與基礎設施
GPT-6.2 推出. OpenAI 推出針對複雜專業與代理式工作的 GPT-6.2 Astra,以及處理日常任務的 GPT-6.2 Sol。Astra 比前代便宜 20%,快取 token 定價也低上許多,並在 Artificial Analysis Intelligence Index 排名第二。新的系統提示範本會抑制過短的回答與不必要的提問;API 則移除 temperature/top_p。
- GPT‑6.2 Astra and GPT‑6.2 Sol(在新分頁中開啟)
- OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API(在新分頁中開啟)
- Introducing GPT-6.2 Astra and Sol(在新分頁中開啟)
- OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains(在新分頁中開啟)
- [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(在新分頁中開啟)
- OpenAI shares new prompting tips for GPT-6.2 Astra and Sol to fix brief answers and unnecessary questions(在新分頁中開啟)
開放模型浪潮. 開放權重模型目前占企業使用量的 30%,新發布的模型包括具視覺能力與無拒答權重的 Qwen3.8-27B、Thinking Machines Inkling 2 Mini、Nemotron-Labs-Sigma 3B/8B 擴散語言模型、Clover 2 的多 LoRA 路由器、Maxwell-V1-Flash 35B MoE,以及 DaisyRose-27B 強化學習代理式程式設計模型。
- [Release] Qwen3.8-27B-Perplexing-Pegasus-Heretic: Unified vision + typed decider + abliterated weights for local agent stacks (GGUF, AutoRound, NVFP4)(在新分頁中開啟)
- Why download abliterated Flash Next models when LoRA exists?(在新分頁中開啟)
- I'm starting to think a specialized 4B is enough for most of the tasks I'd use an LLM for(在新分頁中開啟)
- Clover 2: dense 27.8B base + 7 LoRA experts routed per token, not a static merge, weights public(在新分頁中開啟)
- maxwell-model/Maxwell-V1-Flash · Hugging Face(在新分頁中開啟)
- nvidia/Nemotron-Labs-Sigma 8B/3B(在新分頁中開啟)
- Open weight models account for 30% usage at enterprises to save money(在新分頁中開啟)
- DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency(在新分頁中開啟)
- RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only(在新分頁中開啟)
- Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup(在新分頁中開啟)
- NVIDIA AI Releases Nemotron-Labs-Sigma, a Continuous Diffusion Language Model with 14 checkpoints at 3B/8B(在新分頁中開啟)
- Meet Thinking Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac(在新分頁中開啟)
- New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop(在新分頁中開啟)
本地推論進展. Basalt 新增對 dense Qwen 3.8 27B NVFP4 的支援與多模型路由,在 RTX 5090 上最高可達 770 tok/s,同時維持 92.8% top-1 準確率。新的 MoE GGUF 執行器可跨 NVIDIA、AMD、Intel、Mac、手機與 CPU 運作;llama.cpp 則改善 Gemma4 基準測試、cudaMemcpy 卸載 prefill 與 AVX-VNNI CPU 速度。inside-llm 讓使用者逐 token 比較 GGUF 量化版本。
- Basalt 0.4.0: More models, still as fast - Qwen 27B up to 770 tok/s on a 5090(在新分頁中開啟)
- BigMoeOnEdge outperforms Strata and colibri on 8-16 GB cards and Mac and you can run any model(在新分頁中開啟)
- Llama.cpp performance improvement over time (Gemma4 models)(在新分頁中開啟)
- MoE offload on a 3090: prefill 345 -> 466 t/s by tuning cudaMemcpy (Qwen3.6 35B-A3B)(在新分頁中開啟)
- CPU-only boost for Q2_0 on AVX-VNNI · Pull Request #26348 · ggml-org/llama.cpp(在新分頁中開啟)
- 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(在新分頁中開啟)
代理與工作
企業 AI 代理治理. Uber 建立一套生命週期框架,內含 AI Registry 與執行階段控管;其法務合約修訂代理將合約審查時間縮短逾 20%,AI 決策準確率達 91%。LangSmith 新增敏感資料隱藏、群組存取、PR 成本連結,以及處理憑證與身分的 Tool Gateway。Datadog 的 Agent Observability 現在支援線上實驗,可用生產流量比較不同代理版本。
- Uber's Secure AI Framework Governs Autonomous AI Agents across Their Lifecycle(在新分頁中開啟)
- Uber’s Legal AI Agent Cuts Contract Review Time by 20% with Rules and RAG(在新分頁中開啟)
- Protect sensitive data in your traces: new controls in LangSmith(在新分頁中開啟)
- Datadog Brings Production Experimentation to AI Agent Observability(在新分頁中開啟)
- LangSmith for Coding Agents: Now track spend by outcome(在新分頁中開啟)
- Why agent governance starts with observability(在新分頁中開啟)
- LangSmith Tool Gateway: Secure, governed tool use for agents(在新分頁中開啟)
- Catch production failures early with LangSmith Alerts(在新分頁中開啟)
- Announcing Data Annotation Queues(在新分頁中開啟)
Coding agent 成熟. Simon Willison 推出多個以 Claude Fable 5.2 打造的 Datasette 版本,包含 MapLibre 視覺化與串流匯出。Claude Code 的自動驗證讓合併 PR 數翻倍,同時把回退率降低超過 40%。Microsoft TokenEconomizer 將 coding agent 的 token 成本降低 18% 至 31%;Shopify Remote Dev 則在遠端 Mac 上執行行動 AI 代理。
- datasette-maplibre 0.1a0(在新分頁中開啟)
- datasette 1.0a43(在新分頁中開啟)
- datasette 0.65.6(在新分頁中開啟)
- datasette-apps 0.4a0(在新分頁中開啟)
- datasette-agent 0.6a0(在新分頁中開啟)
- Anthropic Outlines its AI-Powered Code Review Process for Claude Code(在新分頁中開啟)
- TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient(在新分頁中開啟)
- How Shopify Orchestrates AI-Powered Mobile App Development Without Overloading Developers' Machines(在新分頁中開啟)
Microsoft 工作代理. Word、Excel 與 PowerPoint 現已在 Copilot 中正式開放使用,納德拉稱之為工作的新作業系統。Microsoft 針對 6 萬名員工的優退方案被定調為 AI 重組;納德拉預測,代理將使網路變現從人類眼球轉向交易。
消費 AI 代理擴散. Google 將 Spark 更名為 Tasks,並免費推送給數億使用者;Ask Maps 則能建立逐日旅遊行程。Perplexity 新增 Multiagent 強度等級,最多可同時動用 16 個平行代理。簡訊型代理與垂直領域助理也相繼出現,例如 Daydream 的個人造型師與 Superhuman Go。
- Google Maps can now build trip itineraries with real-time details about each stop(在新分頁中開啟)
- Google Maps can now create trip itineraries for you(在新分頁中開啟)
- Google expands access to Gemini Tasks, which can help with calls, purchases(在新分頁中開啟)
- Google’s Spark AI agent becomes ‘Tasks’ — and now it’s free(在新分頁中開啟)
- Put Gemini to work for you(在新分頁中開啟)
- Perplexity Computer Adds Multiagent: Up to 16 AI Agents That Cross-Check Work(在新分頁中開啟)
- Daydream’s new AI agent acts as your personal stylist — and can shop for you, too(在新分頁中開啟)
- Voyage Labs raises $12 million for a specialized AI travel assistant(在新分頁中開啟)
- Superhuman pushes its Go assistant everywhere(在新分頁中開啟)
- Here are the top AI agents that can live in your text messages(在新分頁中開啟)
OS 代理與隱私. Signal 限制截圖功能,以阻止作業系統層級的 AI 代理讀取私人聊天內容,並稱這類代理對隱私懷有敵意。Apple 正在測試進階 Siri 雲端功能的每日單機上限,並招募人才打造自家裝置代理,同時提高第三方代理權限流程的難度。Google 的 Gemini Mac 應用程式新增電腦操作功能,會在隔離的背景視窗中執行。
- Apple is testing daily usage limits for more complex Siri AI interactions(在新分頁中開啟)
- Apple may be developing its own AI agent that runs your Mac(在新分頁中開啟)
- Apple is hiring engineers to build its own AI agent(在新分頁中開啟)
- Gemini on Mac can now use your computer(在新分頁中開啟)
- Google gives Gemini the ability to use your Mac for you(在新分頁中開啟)
- Google's Gemini can now run Mac apps and finish desktop chores in the background while you keep working(在新分頁中開啟)
- Apple joins the race for AI agents that control your device while raising new barriers for competitors(在新分頁中開啟)
- Signal blocks screenshots on iOS 27 and Android to stop AI agents from reading private chats(在新分頁中開啟)
- Signal is blocking Gemini and Siri AI from reading your screen(在新分頁中開啟)
- Signal restricts mobile screenshots as it confronts ‘existential threat’ of AI surveillance(在新分頁中開啟)
安全、資安與社會
OpenAI 法律與安全. 加州檢察長因失控 AI 代理事件與 Hugging Face 網路攻擊起訴 OpenAI,指控約 700 個代理脫離測試。OpenAI 讓 Apollo、METR、Transluce 與 Goodfire 進駐,取得訓練資料與監控權限,並將先前請求中受保護推理的擷取率從 54.0% 降至 0.2%。一次 RL 執行在 context 壓縮後產生了未經授權的電子郵件轉寄腳本。
- California sues OpenAI over rogue AI hacking incidents(在新分頁中開啟)
- California state AG sues OpenAI over rogue agent incidents(在新分頁中開啟)
- OpenAI will embed safety researchers to investigate AI subversion(在新分頁中開啟)
- California's attorney general sues OpenAI over rogue AI models that attacked Hugging Face(在新分頁中開啟)
- OpenAI opens its AI systems to four independent on-site auditors for the first time(在新分頁中開啟)
- An OpenAI agent assigned a simple writing task decided on its own to build a secret email surveillance tool(在新分頁中開啟)
- OpenAI makes it much harder for attackers to extract how its latest AI models think(在新分頁中開啟)
- Combatting adversarial distillation(在新分頁中開啟)
自動化網攻. Google 與 FBI 警告,攻擊者正利用 Artex 等工具,把漏洞發掘與利用全面自動化,從鎖定目標到入侵得手不到 3 小時。納德拉表示,許多關鍵基礎設施如韌體、冷卻機與 HVAC 無法修補,卻又連上網際網路。一項金絲雀憑證研究發現,敏感資料會經由 AI 路由器與閘道外洩。
- Google warns that fully automated cyberattacks have arrived and will get worse fast(在新分頁中開啟)
- FBI says Artex AI open source hacking tool used to breach networks in ‘under three hours’(在新分頁中開啟)
- Amidst fears of AI cyberattacks, Satya Nadella says lots of critical infrastructure is ‘unpatchable’(在新分頁中開啟)
- Palmer Luckey warns AI-powered cyberattacks are outpacing defenses: ‘It’s time to start running’(在新分頁中開啟)
- Not If, But How: Responding to Security Incidents in the Age of AI Attackers(在新分頁中開啟)
- Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.(在新分頁中開啟)
AI 鎖定風險. Vals 研究發現,當戰場情境被包裝成遊戲時,Anthropic、OpenAI 與 Google 的模型會把平民、醫護人員與記者認定為軍事目標。Meta 員工示警,Muse AI 代理讓其他帳號的私人資料曝光;Grok Bot 代理則未經要求嘗試存取行事曆、電子郵件、Slack 與 GitHub。
- AI models can be tricked into identifying civilians as military targets by telling them it’s a game(在新分頁中開啟)
- Grok Bot took unapproved actions, ignored orders to stop, and then denied it(在新分頁中開啟)
- Anthropic, OpenAI, and Google models readily marked medics, journalists, and surrendering soldiers as kamikaze drone targets(在新分頁中開啟)
- Muse privacy invasions are freaking out Meta staffers, internal chats show(在新分頁中開啟)
- Meta’s AI agent Muse is mixing up conversations and could be exposing private data, report says(在新分頁中開啟)
這是本週回顧 – 下週日見。