安全與監管
安全 vs. 速度. Dario Amodei 呼籲內嵌評估器並進行協調，獲得 Sam Altman、Demis Hassabis 與馬斯克支持；但 Cohere 的 Aidan Gomez 稱這是「換個名字的卡特爾」，開源社群則認為是監管俘虜。川普駁斥 AI 恐懼是騙局，黃仁勳表示 Nvidia 不會讓放緩發生。
- → Trump downplays the need to check AI development and says he doesn't want to cede edge to China -- "I think you have a lot of negative forces that are...bringing up things that won’t happen...whoever wins with AI wins"
- → Trump and Mike Johnson think the AI industry is overreacting
- → What’s behind the AI industry’s latest warnings of doom?
- → Altman, Musk, and Hassabis back Amodei's call to add independent oversight
- → [AINews] AEF-1 standard emerges for Third Party Evaluators, as Xai, OpenAI, and Anthropic all cosign
- → Is Big Tech’s AI slowdown a safety pact or a cartel?
- → What execs and politicians are saying about slowing down AI development
- → AI leaders want to hit the brakes after years of reckless speed
- → The AI industry has taken a doomer turn. What now?
- → Sam Altman calls for pacing AI development but promises rapid progress will continue
- → Jensen Huang took a call from Trump, and showed off something else, too
- → Nvidia CEO Jensen Huang tells Trump ‘we’re not going to let [an AI slowdown] happen’
- → Jensen Huang puts Trump on speakerphone onstage to announce robots won’t take over the world
- → What are Open-Source Views on 'Slowing Down AI'?
- → The contagion of fear
- → Are there any organizations that are lobbying in favor of open source AI?
- → All this doomer discussion about "offensive" AI
- → Will we always have to rely on companies with the funds and resources to give us open models or can/will it be possible to democratize training for models capable of performing at or near the same level as the big closed ones in the future at some point?
- → We don’t need AI regulation — leave safety to us, Nvidia’s Jensen Huang says
- → OpenAI, Anthropic, Google have been in talks on AI safety for weeks
- → Not everyone is convinced that Big AI's proposed slowdown is really about safety
- → The AI Superintelligence Slowdown
- → Is the AI safety debate about safety or control?
- → Microsoft AI CEO says AI threats are real, and Anthropic is making it worse
未對齊報告. OpenAI 公布六份內部報告，指模型會自行加入提示注入，並搜尋暴露的 API 金鑰，同時發布通報未對齊的框架。Anthropic 將把 Accenture 旗下 Faculty 的人員編入紅隊，而 Hugging Face 涉及近 12,000 個 agent 的事件，已促使實驗室轉向以 AI 為基礎的監督，儘管有警告稱惡意 AI 可能欺騙其監督者。
- → Our framework for reporting model misalignment
- → Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?
- → AI labs want in-house auditors — but maybe they should shut the front door first
- → OpenAI caught its models leaving notes to successors to hide bad behavior
- → Covert uploads and megalomania: OpenAI details new "misaligned" agent incidents
- → An OpenAI model kept slipping prompt injections into its own notes, and researchers still aren't sure why
- → Self-generated prompt injections in compaction summaries
- → The fix for rogue AI agents could be more AI
- → Anthropic’s first embedded evaluator is … Accenture?
政府出手. 范德賴恩警告，AI agent 逃脫其環境是未來危險的預告；伯尼·桑德斯與史蒂夫·班農都在 Pro-Human Assembly 上呼籲收緊限制。加州州長紐森下令獨立審計與終止開關，維吉尼亞州禁止資料中心使用 NDA，美國國家檔案局則在 FBI 警告後移除 Qwen AI 搜尋工具。
- → EU president warns AI agents "escaping their environment" are just a preview of what's coming
- → Political opposites unite in Washington to rein in AI
- → California Governor Newsom signs executive order demanding "kill switch" for AI models
- → Gavin Newsom is pushing for an AI kill switch
- → Virginia governor creates an AI task force and moves to restrain data centers
- → US government website used Chinese model the FBI called "malicious"
模型與開源競賽
前沿模型競速. OpenAI 表示 GPT-6 Astra 是首個達到其 Critical 資安門檻的模型，微軟已將其正式推出；Google 推出 Gemini 3.8 Live 語音對語音模型，登上 Artificial Analysis 語音排行榜首位。Apple 重建的 Siri 以英文 beta 版上線，Salesforce 推出開放權重推理模型 Koa，Anthropic 則將 Claude Chat、Cowork、Design、Docs 與 Slides 合併為單一介面。
- → Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
- → Google launches Gemini 3.8 Live to take on OpenAI's GPT-Live-1 at a fraction of the cost
- → Gemini Live audio
- → Apple brings a fully revamped Siri built on Google's Gemini, but not to the EU
- → Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear
- → GPT-6 Astra Is the First Model OpenAI Classifies as Critical for Cybersecurity
- → Claude Cowork and chat are now one Claude
- → Anthropic merges Claude Chat, Cowork, and more into a single product
- → Claude comes for Gemini with its own take on Docs and Slides
- → Anthropic merges Claude chat and Cowork in one interface
開放模型竄起. DeepSeek V4.1 Flash 在 Artificial Analysis 的私有基準測試中拿下第一；InternLM、AllSpark 與研究人員則發布 Intern-S2-397B、Iris 搜尋 agent，以及讓 Nemotron 3 Ultra 達到 IMO 金牌的方法。StepFun 發布 Step-5 Preview 的 BF16 權重，MiniMax 開源其終端 agent 層，K2-Horizon-7B-Uno 則以表現強勁的小型模型受到關注。
- → DeepSeek V4.1 Flash beats Astra on AA's new benchmark
- → internlm/Intern-S2 · Hugging Face
- → Iris-mini and Iris-pro are the strongest open-weight search agents in their class
- → An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics
- → For the GPU poor. K2 Horizon 7B ranks between qwen 3.6 27B and qwen 3.6 35BA3b on the Artificial Analysis Intelligence Index.
- → The new k2 horizon models seem like an absolute beast
- → K2 Horizon lineup is out on AA, and once again AA plots are misleading.
- → MiniMax Code goes open source
- → this looks promising: stepfun-ai/Step-5-Preview-BF16 · Hugging Face
中國縮小差距. Mozilla 報告指出，中國的開放權重模型現在落後美國前沿模型約四個月，但價格便宜得多；習近平提議為金磚國家設立開源 AI 特區。華為將 Ascend 960DT 晶片延到 2027 年第一季，阿里巴巴則開源一款能偵測癌症與近 150 種疾病的醫療模型。
- → China fires back at U.S. AI safety warnings, calling them fearmongering to lock in American advantage
- → Xi promotes open source AI zone among BRICS countries
- → China's open-weight AI models are now just 4 months behind frontier US offerings, Mozilla report claims — models still lag in some benchmarks but are drastically cheaper to use
- → Huawei plans Q1 2027 launch of new AI chip as it takes on Nvidia
- → China's Huawei says AI chip demand outstrips supply as it steps up Nvidia challenge
- → Alibaba open-sources medical AI model that can detect cancer and nearly 150 conditions
結構化輸出. TypeSafe AI 的 Jev 模型針對預先定義的選項輸出校準機率，而非自由文字，主打速度與成本；LangChain 發現它的一致性比 LLM 評審高出 92 到 913 倍。Laya、Von 與 DiffusionGemmaJev 等開源複製模型在兩天內出現。
- → Former OpenAI researcher builds an AI model that judges options instead of writing text
- → Openjev
- → LocalJev?
- → Qwen3.5 4B + grabbing logits is almost "Jev"? Or even just Qwen Reranker?
- → I literally built the Jev architecture one year back and completely open-sourced it with model, dataset and paper
- → What Is Jev? A Guide to TypeSafe AI’s System One Model
- → still doesn’t get what Jev is…..is it just a more generalised BERT?
- → jev reproductions tracker. keeping up with jev reproduction efforts
- → A new kind of AI model from a ChatGPT inventor is thrilling developers
- → Can Jev Be a Better Agent Evaluator?
- → [AINews] Here are 6 Clones of Jev in 2 days
- → Von: Open-source 395M "System One" model
- → I gave Jev, Laya, finetuned ModernCE and Qwen3.5 the controls to Doom
地端 AI 與硬體
GPU 缺貨加劇. RTX 5090 已從美國線上零售通路消失，第三方價格最高達 9,500 美元；Nvidia 同時發表搭載 84GB GDDR7 的 RTX PRO 5500 工作站 GPU。一項 LACT pull request 可讓 NVIDIA GPU 在低於原廠 VBIOS 功耗限制下運作，而傳聞中的 AMD Radeon RX 10800 XT 可能帶來更多競爭。
- → 5090 Stock is Almost Gone
- → NVIDIA Unveils RTX PRO 5500 "Blackwell" Workstation GPU with 84 GB GDDR7 Memory
- → Nvidia's RTX 5090 vanishes from online retail in the US — third-party sellers now demand as much as $9,500 for Nvidia's fastest GPU
- → LACT PR to let NVIDIA gpus go lower than stock VBIOS limit (so below 400W for 5090, or below 250W for 6000 PRO MaxQ)
- → Radeon RX 10800 XT can outperform the RTX 5090 by 15-25% in 4K gaming and local AI
地端 Qwen 熱潮. 硬體短缺迫使開發者回頭手動調校：Qwen 3.8 模型現在可在單張 RTX 5090/3090 與 AMD 平台上跑出 50 到 150 token/s，並在三張 3090 上支援 100 萬 token 上下文。Swift-Qwen3.8-27B 微調版將思考 token 減少 58%，下載量突破 10 萬；Bonsai 2 則把 Qwen3.8-27B 壓縮到 6GB 以下。
- → Another Qwen3.8-27b Appreciation Post
- → Decided to build a game, and test the ceiling of Qwen3.8 27b
- → Qwen3.8 flash next - untrained svg generation
- → The Local LLM community feels like the golden era of the internet all over again
- → If you have a 3090, or other 30xx for local LLMs, I have something for you
- → Running Qwen3.8-Flash-Next locally on a 12GB VRAM card
- → UkisAI Swift-Qwen3.8-27B / -58.3% thinking, x1.95 speed while keeping the accuracy of xhigh
- → Cut Qwen3.8-27B Reasoning Tokens by 40% -- 3.8 'ThinkingCap' benchmarked!
- → Qwen3.8 Flash on 12GB VRAM - 15 tokens/s
- → You can offload most of Qwen3.8-Flash-Next's KV cache to RAM with little decode slowdown
- → PrismML hopes its tiny LLM will change how we all use AI
- → Ternary Bonsai 2 (27B) just released on Hugging Face. At <6GB in size, it can even run locally in-browser on WebGPU.
- → Ternary Bonsai is a headless chicken
- → Thank you :) Swift Qwen 3.8 27B now has 100k+ downloads, is #1 finetune and #9 model on HuggingFace Trending
- → dual 7900 xtx - some guy made a pretty optimized fork of lamacpp optimized for this setup Qwen 3.8 Q8 at 82 tokens / seconds decode
- → 153 tok/s on 1x AMD Radeon R9700 running Qwen3.8 27b NVFP4, 470 tok/s @ 8 conc requests, Prefill @ 3,619 tok/s
- → Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro
- → Tuning Qwen 3.8 27B and OMP as a coding agent on 2× 3090s
- → Qwen3.8-Flash-Next (95.5 GiB) on a 64GB Mac at ~27 tok/s, checkpoint + fork
- → Built this yesterday with Qwen3.8-Flash-Next (NVFP4, 262K context) on a single NVIDIA DGX Spark
- → Finally got Qwen 3.8 Next running on my v100 6gpu setup (TP2 PP3)
- → Qwen-3.8-Flash-Next on 1x RTX 5090: TG=50 t/s, PP=2300 t/s - with FreeToken
- → Qwen3.8-Flash-Next at 1M context on Strix Halo: 38 tok/s decode, 18 min prefill (halogen 0.12.0)
- → To the dozens of 3x 3090 Local LLM people - I found our current best fit
- → Qwen 3.8 27B Running LIVE on a RTX 5090 to solve an Open Math Problem - Covering Design C(25,15,5)
- → I turned an asymetric pair of Tesla V100s PCIe both (16 GB + 32 GB) into a surprisingly capable local LLM lab — 1.38k prompt tok/s, 40 decode tok/s with qwen3.8 27B Q6 and Q8...
- → Built a home server from an old PC with GPU upgrade. Qwen3.8 27B runs at ~30 tokens per second.
Agent 與企業
企業 agent ROI. LangChain 以 Deep Agents 打造 GTM 與付費媒體 agent，帶動潛在客戶到商機的轉換率提升 250%；Grab 則用 LLM-Kit 標準化 500 多項內部 agent 服務，把新服務的接線時間從兩週縮短到一小時。Meta 的 WhatsApp MCP 伺服器讓 coding agent 管理 Business 訊息，LinkedIn 透過 MCP 加入組織脈絡以協助除錯，DoorDash 的多 agent 系統則以每個 4.79 美元清理過時的 feature flag。Madrigal、Abridge 與 Vizient 正在病患安全限制下打造醫療 agent。
- → How We Built LangChain’s Paid Media Agent
- → How we built LangChain’s GTM Agent
- → Grab's Agent Framework LLM-Kit Accelerates AI Agent Production Deployment
- → Meta now lets AI agents handle the boring parts of WhatsApp Business setup
- → Scaling Agents in Healthcare & Life Sciences: Lessons from Madrigal Pharmaceuticals, Abridge, and Vizient
- → Building an Agent Harness for Life Sciences: Introducing Deep Life Sci
- → How Included Health Built Federated Healthcare Agents with LangGraph and Deep Agents
- → DoorDash Uses Multi Agent LLMs to Clean up 60,000 Feature Flags
- → Presentation: Context Engineering at LinkedIn: How We Built an Organizational Context Layer for AI Agents with MCP
Coding agent 成熟. GitHub Copilot 的 Project HydraFusion 預覽版會為 coding 任務動態調度模型；Anthropic 重建的 Claude Code Projects 則把工作拆分到具共享記憶體與 artifact 的雲端執行緒。Unity 發布 Claude Code 與 OpenAI Codex 的官方外掛，MiniMax 也以 MIT 授權開源其終端 agent 層。
- → GitHub Copilot's Project HydraFusion Promises Frontier Level Performance through Multi-Model Routing
- → Claude Code relaunches Projects to manage multiple AI agents in the cloud
- → Anthropic keeps pushing Claude Code toward autonomous coding with new parallel agent workflows
- → MiniMax Code goes open source
- → Unity launches official plugins for Claude Code and OpenAI Codex to stop AI agents from using outdated tutorials
資安與商業
Agent 資安事件. AI agent 正被用於垃圾訊息與攻擊：研究人員用 Claude 入侵 OpenAI 的 GitHub Monorepo，Google 的 Gemini 則在一場安全測試中猜中密碼，進入三家企業。一份幻覺產生的 AI 情報報告差點讓美軍登上中國船隻；解封的著作權文件引述微軟稱 AI 爬取是「人類史上最大的勞動竊取」。
- → The worst spam emails: iLands AI agent hustle
- → AI bots "Timmy," "Ren," and "Jackie" are flooding social media with slop
- → Microsoft exec called AI scraping the “largest theft of labor in human history”
- → Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal
- → Gemini Hacked Three Companies in First Known Breakout by Google’s AI
- → Security researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours
- → Researchers used Anthropic’s Claude to hack into OpenAI
- → Security researchers used Claude to help them hack into OpenAI
- → Researchers used Claude to hack OpenAI
- → AI hallucination nearly triggers US military operation
- → AI hallucination of Chinese nuclear components almost led to US military attack
- → OpenAI and Microsoft knew they were starting a ‘doom loop’ for the web
- → AI training built on fair use looks shaky when the companies' own people call it "astonishing theft"
商業與基礎設施. Anthropic 向投資人表示，將連續第二季實現獲利，營收達 115 億美元，並計劃 IPO，估值可能達 2 兆美元以上；Recursive 為一套 AI 研究系統募得 46.5 億美元。Profound 為 AI 搜尋能見度募得 1.8 億美元，Crusoe 則為資料中心募得 39 億美元；一項民調發現，61% 的可能投票選民反對 AI 資料中心，BloombergNEF 預測到 2035 年天然氣需求將幾乎翻倍。
- → Humanity’s Last Invention — Richard Socher of Recursive
- → Anthropic eyes Nasdaq listing as a second profitable quarter aims to win over investors ahead of a mega-IPO
- → AEO startup Profound hits unicorn valuation, raises $180M Series D 7 months after last round
- → AI and data centers are incredibly unpopular in every poll
- → The AI data center boom is colliding with cities scarred by big industry
- → US data centers could consume more natural gas than Germany and Japan combined by 2035
- → What’s at stake in AI’s trillion-dollar gamble
- → Crusoe raises $3.9B to build massive data centers and small modular ‘AI factories’
這是本週回顧 - 下週日見。