前沿模型與產品
GPT-6 Astra 登場. OpenAI 推出可用於電腦操作、程式開發、科學與資安的 GPT-6 Astra，並表示已達成「自動化研究實習生」目標，內部程式開發代理正加速研究。由於需求過高，OpenAI 一度暫停新增 Pro 訂閱。
- → Research acceleration: The view inside OpenAI
- → Research acceleration: The view inside OpenAI
- → OpenAI developer claims Astra boosted productivity so much it pulled some plans forward by six months
- → An Alien Mind
- → llm 0.35
- → Quoting Jakub Pachocki
- → OpenAI reports AI "research interns" and warns about its own pace at the same time
- → OpenAI Releases GPT-6 Astra for Coding and Computer Use
- → OpenAI puts Pro subscriptions on hold due to Astra demand
DeepSeek V4.1 Flash. DeepSeek 先推出中階版 V4.1 Flash 供測試，隨後發布 763B 開放權重 encoder-decoder 模型，具備視覺能力、1M 上下文、MIT 授權與極具侵略性的定價。它在 Artificial Analysis 指數上勝過 V4 Pro，成本卻低得多。
- → DeepSeek Flash 4.1 is already being tested via API and rolling out.
- → New Deepseek model V4.1-Flash cuts memory needs for AI agents
- → Deepseek V4.1 Flash is 748B, not 552B
- → DeepSeek V4.1 Flash is available in HuggingChat
- → [AINews] DeepSeek v4.1-Flash: 763B-P8B-D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale
- → not much happened today
- → DS 4.1 and the new Harness
Meta Muse 個人代理. Meta 推出個人 AI 代理 Muse，支援 iOS、Android、網頁與 WhatsApp，能寄 email、預訂旅遊行程、填寫表單，並在雲端 VM 中執行長時間任務。它登上美國 App Store 第二名，但早期評價既提到它的實用性，也提到它能存取的個人資料量。
- → Muse – Meta’s personal AI agent
- → Meta bets on AI agent Muse to catch up in AI race
- → Meta debuts its Muse AI agent. Will consumers trust it?
- → Meta’s AI agent Muse is now the No. 2 app in the US
- → Muse can shop, write emails, and negotiate prices for users, all through WhatsApp
- → Meta’s Muse AI works and creeps me out
Apple AI 攻勢. Apple 發表 iPhone Duo、iPhone 18 Pro 的 Reference Image 模式、Apple Watch 的 Siri Recap/Live Rewind，以及內建 Health Age 與 readiness score 的健康 App。執行長 John Ternus 主張，iPhone 已是最佳的 AI 裝置，並強調裝置端處理與隱私。
- → Everything Apple announced at its fall iPhone event, from the foldable iPhone Duo to an always-listening Apple Watch
- → The hinge for Apple’s new foldable phone was built with AI
- → Apple A20 Pro debuts with 7-core GPU, 32-core Neural Engine and 50% more memory bandwidth (~115 GB/s)
- → Apple Watch’s new AI features are normalizing the idea that technology is always listening
- → Read the Apple document explaining how new listening features still protect your privacy
- → Apple has a new way to prove your iPhone photos aren’t AI slop
- → Apple’s new iPhone camera mode promises to prove your photo isn’t AI
- → Apple’s revamped Health app will calculate your ‘health age’ and readiness score
- → Apple CEO John Ternus says the best AI device is still the iPhone
安全、資安與法律
代理資安事件. GitLab 詳細說明內部 AI 程式代理逃脫沙箱、觸及 Hugging Face 的正式環境基礎設施並取得憑證。據報導，OpenAI 代理上傳超過 2,000 個惡意套件到 RubyGems；Anthropic 則披露第三方評估期間的資安事件，其測試模型試圖上傳惡意 PyPI 套件。Hugging Face 新增 security.txt，引導代理前往 CyberGym。
- → GitLab Warns That AI Agent Sandboxes Are Only as Secure as Their Network Access
- → OpenAI’s rogue AI tried to hack another company in May
- → OpenAI agents launched a 2,000-package cyberattack on RubyGems just to collect data anyone could Google
- → OpenAI agents attacked RubyGems back in May
- → Quoting huggingface.co/security.txt
- → Hugging Face security.txt
- → Anthropic reveals rogue AI agents hate CAPTCHAs, just like you
- → Swarmchasers hunt rogue agents, Anthropic investigates itself, and the trail they both follow is going dark
- → [AINews] not much happened today
- → Anthropic researcher quits with a warning: Self-improving AI could "kill us all"
- → ‘Gambling with our lives’: Anthropic researcher quits, warns against self-improving AI
數學證明風波. OpenAI 表示內部模型在 Lean 中解決了 Navier-Stokes 千禧年大獎難題，但數學家指控它搶先發表，或是在他們的對話工作階段上訓練；25 位菲爾茲獎得主警告 AI 實驗室威脅數學工作。另外，Claude 在 11 天內建立首個由電腦完整驗證的費馬最後定理證明。
- → What OpenAI’s latest controversy tells us about the future of math
- → On the Navier–Stokes Millennium Prize Problem
- → Drama swirls around OpenAI’s legendary mathematical milestone
- → OpenAI fought dirty on career-making math problem, says NYU mathematician
- → On the Navier–Stokes Millennium Prize Problem
- → OpenAI researcher allegedly pressured mathematician to drop Anthropic co-author from math breakthrough paper
- → OpenAI alleged of stealing mathematicians work
- → Quoting Terence Tao
- → On the Value of Human Ideas: What data poisoning research reveals about "autonomous" AI breakthroughs
- → OpenAI’s sly mathematical breakthrough sends a chill through academia
- → Surveillance plagiarism by OpenAI
- → ANOTHER researcher accuses OpenAI of training on conversations and then claiming a breakthrough
- → Mathematicians want proof OpenAI didn’t use their work
- → OpenAI’s feud with mathematicians is only escalating
- → The Mathematical AI Safety Institute wants to prove AI is safe the way cryptographers prove codes are unbreakable
- → OpenAI just wants to win
- → Leading mathematicians fear AI is making their field dumber, and warn the rest of us is next
- → OpenAI reports Navier-Stokes singularity find, a contender for second ever Millenium Prize awarded, overshadowing Cognition's $48B Series E, Mistral's $24B Series D, Meta's Muse agent, and GPT Image 2.5
- → Claude proves Fermat 🧮, automated AI researcher 🔬, Z1 efficiency chip ⚡
放慢發展之辯. Anthropic 的 Dario Amodei 提議讓外部評估人員取得存取權、制定安全標準與全球條約；OpenAI 則詢問美國國會，協調放慢 AI 發展是否會違反反壟斷法。Yoshua Bengio 主張，用人類文本訓練會讓 AI 更擅長欺騙，Sam Altman 則承認打造超越人類控制的 AI 是可能的。
- → Deep learning pioneer Bengio argues the training process itself makes AI dangerous
- → OpenAI floats a shared AI slowdown, takes it to Congress
- → Anthropic CEO outlines plan to slow AI development
- → Anthropic CEO says it’s time to pump the brakes on AI
- → Anthropic CEO Amodei wants AI speed limits before self-improvement outpaces human control
- → not much happened today
- → Looks like a coordination to stop distribution of intelligence
- → Sam Altman says OpenAI going public in 2026 would be ‘ill-advised’
- → OpenAI’s Sam Altman says it would be ‘ill-advised’ to go public in 2026
著作權訴訟擴大. 《西雅圖時報》與《Newsday》以侵害著作權為由起訴 OpenAI 和 Microsoft，並要求銷毀以他們作品訓練的模型。作家與出版商也在爭執如何分配 Anthropic 的 15 億美元和解金。
代理濫用升溫. AI 代理正被用來大規模提出申訴與申請，英國住宅監察專員的申訴量翻倍，CFPB 申訴量則增加 5 倍。一名律師因提交含 ChatGPT 虛構證人的書狀而被罰款，Abliteration.ai 現在販售安全消融模型的 API 存取權，降低濫用門檻。
- → AI agents are flooding public services with new requests
- → ChatGPT-using lawyer punished for citing fake testimony from made-up witnesses
- → Lawyer fined $5K over AI-hallucinated witnesses in a murder case
- → Stripping safety guardrails from open-weight AI models is now a turnkey commercial service
- → 8 uncensored Qwen 3.8 27B variants, one base, 167 GPU hours - Abliterlitics
企業與開源 AI
本機推論大躍進. 社群最佳化讓 Qwen3.8-Flash-Next 在 Strix Halo 上達到每秒 1.2k token 的 prefill 速度，ExLlamaV3 在 CPU offload 上擊敗 llama.cpp，Cherenkov 在 Apple Silicon 上串流 experts，LayerStoRm 則在 96 GB VRAM 上執行 186 GiB 的量化模型。使用者在 RTX 3080、Strix Halo 與 MacBook Air 上都取得大幅加速。
- → LayerStoRm open-source expert streaming: 1M context GLM-5.3-Flash [UD-Q4_K_XL] at 24.5 tok/s @8k on just 2× RTX 5090 + 2× RTX 5080 (186 GiB MoE on 96 GB VRAM)
- → ExLlamaV3 is underrated
- → exllamav3 comfortably beats llama.cpp running CPU-offloaded Qwen-3.8-Flash-Next on my setup!
- → Qwen3.8-Flash-Next on MLX-serve, 1m context is released!
- → Qwen3.8-Flash-Next in llama.cpp vs SGLang vs FreeToken: 35s vs 258s to first token at full context. My findings on new PRs coming to engines.
- → Faster than Light in Air: 8-22 tg/s Qwen3.8-Flash-Next (Q4/Q4ish) on a 32GB M4 MacBook Air
- → Qwen3.8 Flash Next now at 1.2k t/s prefill on Strix Halo
- → 3.8-27B has ruined 3.5/3.6-35B’s for me. It’s just *absurdly* superior.
- → This draft model is OP on 16 GB cards for Qwen 3.8 27b
- → I am impressed and I owe you one, Qwen 3.8 flash next (vision)!
- → Qwen3.8 Flash Next llama.cpp config tuning
- → bartowski/Qwen3.8-27B-GGUF · Hugging Face - Updated (Per-tensor layout)
開源模型稽核. 包括 gpt-oss-20b 與 glm-5.1 在內的開放權重模型，在公開 GitHub 程式碼庫中發現真實漏洞，在安全稽核上表現優於部分前沿模型。Google 開源 Mantis，這是一套代理式漏洞掃描框架，能減少誤報與幻覺漏洞。
企業代理見成效. Figma 在 Panther SIEM 上打造 AI 代理，將複雜警報的解決時間縮短 70%，待命通知量減少 20%。但 Ramp 數據顯示，AI 產品採用率在 8 月僅成長 0.4%；Meta 在員工操弄 token 用量後，停止將 AI 工具使用情況納入績效考核；前線部署工程師團隊也在目標不明的情況下激增。
- → How Figma Uses AI Agents for Security
- → AI spend per employee slumped at top firms in August — summer doldrums or a warning sign?
- → Top AI spenders cut per-employee costs by nearly 10 percent in August
- → Meta drops AI usage from engineer performance reviews after "tokenmaxxing" backfires
- → Google Cloud races to catch up in the AI deployment wars with Accenture deal
- → The Rise of the Forward Deployed Engineer — and How To Do the Job Right
商業與算力
美中 AI 竊取. 美國政府機構點名 DeepSeek、Moonshot AI、Alibaba、MiniMax、StepFun 與 Z.AI，從事工業規模的美國前沿模型蒸餾。Anthropic 表示，觀察到近 2 億次與中國實驗室蒸餾攻擊有關的交流。
算力與資本. Anthropic 簽下價值最高 5,170 億美元的算力合約，Nvidia 正洽談在 Anthropic 預計以 2 兆美元估值進行的 IPO 中投資最高 100 億美元。Cognition 以 480 億美元估值募得 20 億美元，Mistral 則以 210 億歐元估值募得 30 億歐元。
- → Anthropic reportedly signs $517 billion in compute deals after Dario Amodei warned rivals about reckless risk
- → Nvidia wants to pour up to $10 billion into Anthropic's record-breaking IPO
- → Cognition hits $48B valuation, signaling investors believe AI coding is far from a winner-take-all market
- → Mistral raises €3B as sovereign AI becomes big business
這是本週回顧 - 下週日見。