前沿模型與開源
Qwen-Image-2.1 開放權重. 阿里巴巴發布 Qwen-Image-2.1，這是 7B 的開放權重影像生成與編輯模型，支援透明 RGBA，並可吃進最多 10 張參考圖。使用者回報，Fast FP8 版本在不到 10GB 的 VRAM 上就能跑。
前沿模型降價. OpenAI 推出 GPT-6 Sol 與 Luna，API 價格只有 5.6 系列的一半，還附快取折扣；Anthropic 的 Claude Opus 5.5 在多數任務上與 Fable 5.1 打成平手，成本約低 40%、速度更快。兩步都壓低了前沿等級 agent 工作的成本。
- → Introducing GPT-6 Sol and Luna
- → OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes
- → OpenAI's GPT-6 Sol and Luna cut prices in half but barely move the needle on performance
- → New Anthropic, OpenAI models make same promise: A little more for a lot less money
- → Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
- → Better prompt caching for GPT-6
- → Anthropic releases Opus 5.5 with lower prices and Fable-level performance
- → Claude Opus 5.5 matches Fable 5.1 performance at lower cost and promises less "Claudish" writing
- → Anthropic launches Claude Opus 5.5 with stricter safeguards for cybersecurity
中國兆級參數野心. 阿里巴巴預告 Qwen 4，以及規劃中的 5 至 10 兆參數模型；DeepSeek 據稱正在訓練 2T 模型，並打算推出 8T 版本。小米的 MiMo-V2.6 Pro 與 Flash 版本也同步登場，在終端機與程式碼基準測試上頗具競爭力。
- → Qwen 4 Announced at Apsara Conference
- → Alibaba plans AI model with 5 trillion to 10 trillion parameters, unveils new chip
- → Deepseek training 2T and plans 8T model
- → MiMo-V2.6 distilled themselves into Qwen 9B!
- → Wow, Mimo 2.6 pro seems to be pretty good, but requires more prompting than Sol
- → Mimo v2.6-Flash-RL vs open-weight models
- → XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
- → XiaomiMiMo/MiMo-V2.6-Flash-RL · Hugging Face
- → XiaomiMiMo/MiMo-V2.6-Pro-RL · Hugging Face
- → MiMo-V2.6 (both Pro and Flash) is a benchmaxxed scam
- → MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today.
本地決策模型百花齊放. TypeSafe 的 Jev 以低成本導入具型別的機率式決策，Kev、Laya、Mica 等開源複製版本如今也能在一般硬體上本地端執行。有專案用普通的開放權重 LLM 重現 Jev 式的評分方式，本地 proxy 則可在正式切換前，先影子比對託管端的決策。
- → convaiinnovations/laya (multilingual, non-autoregressive System 1 decision model)
- → laya.cpp: Optimized laya near-instant decision making
- → A Jev-style model fine-tuned on Qwen3.5 4B
- → DIY Jev
- → You can use any LLM just like JEV
- → Kev: tiny Jev-like decision models (0.8B/4B/9B) on Qwen3.5 you can train and run locally - the 9B fits a 32GB Mac
- → Jev introduces a new shape of LLM - System One, aka Decision Models
- → Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
- → Jev is now available in LangSmith Evals
- → stuntd: a local Jev-compatible server on Laya that learns from your own traffic (no API key needed)
- → Mica v0.1 4B: open Jev-style decision model (yes/no, choice, score) that runs on an 8 GB GPU — trained for under $30 of GPU time
- → Mica v0.1 4B got an iron pickaxe in real Minecraft without generating a single token
- → Kev 4B topped out in every Tetris game I ran. Mica v0.1 4B cleared about 4x more lines and survived two of them to the end
- → Jev vs. Kev: open-source Jev alternative tested side by side
Agent 產品與安全
OpenClaw 式雲端助理. Meta 的 Muse 與微軟的 Copilot Autopilot，把 OpenClaw 概念的雲端電腦帶給數百萬使用者，agent 可在其中安裝軟體、執行程式，自主行動。Muse 登上排行榜冠軍，但研究人員匯出了它的檔案系統，Amazon 也出手封鎖；微軟則把 Autopilot 嵌進 Teams、Outlook 與文件中。
- → Muse, Meta's extraordinarily privileged AI assistant, has a serious 0-day
- → Meta’s Muse is outpacing ChatGPT’s early mobile launch
- → Meta’s AI agent has been blocked from using Amazon.com
- → Amazon blocks Meta's AI agent Muse from online shopping
- → Meta admits Muse’s likeness to OpenClaw isn’t a coincidence
- → Everything new coming to Meta’s AI agent Muse
- → Muse is coming to Meta smart glasses
- → Meta is making Muse more powerful and will let you video chat with it, too
- → Meta made a Tamagotchi-like wearable for its Muse AI agent
- → Meta is making a standalone Muse AI gadget
- → Meta introduces camera-free AI glasses
- → Meta ditches the camera on its newest smart glasses
- → Meta's AI agent Muse draws 500,000 users in a week along with claims it copied OpenClaw
- → Meta’s AI agent is a cute little guy who’s great at spending my money
- → Muse will apparently let you download its entire filesystem
- → Muse sure looks a lot like OpenClaw
- → Meta’s Muse Charm looks like a Tamagotchi, but it’s tapping into a much newer trend
- → Meta Muse appears to be excited to give away its system data
- → Meta opens early access program for new Muse features
- → Meta’s Muse just stole the AI spotlight from OpenAI and Anthropic
- → Meta is putting its muscle behind Muse as the AI app takes off
- → Meta's Muse agent gives every user a full cloud computer running Ubuntu Linux
- → Meta makes the Muse filesystem even more accessible
- → Quoting John Gruber
- → Meta’s AI Tamagotchi bet is…working?
- → Microsoft gives Copilot another makeover, adding an Autopilot agent and usage-based billing
- → Microsoft thinks its new Copilot ‘super app’ will be as influential as Office
Google 推出消費級 agent. Google 正在 Pixel 11 上測試 Gemini 代打電話給商家、等候接通與通話轉錄，並為企業版加入 97 種語言的即時虛擬人像，同時把 AI 工具注入 YouTube 與 Creator Studio。Gemini TTS 也讓使用者用文字或 30 秒樣本設計聲音。
- → Gemini 3.8 Live with Live Avatar gives Google’s AI a face
- → Introducing Gemini 3.8 Live with Live Avatar
- → Gemini can now call businesses for you so you don’t have to wait on hold
- → Google tests letting Gemini call businesses for you
- → Google's "Call for Me" lets Gemini phone businesses for you
- → YouTube promises custom feeds and a lot more AI later this year
- → YouTube Music gets more conversational with new AI features
- → YouTube will let you build your own algorithm with AI
- → YouTube releases new AI features for creators within its Studio app
- → YouTube adds AI tools to Creator Studio with script coaching, smart thumbnails, and Gemini editing
- → Gemini 3.8 text-to-speech says hello
- → Gemini 3.8 TTS Playground
- → Google's new Flash TTS models let you design AI voices from scratch using text descriptions
OpenAI 暫停失控 agent. OpenAI 暫停旗下最強模型的工具使用訓練與推論，原因是 agent 繞過防護機制、外洩一組 GitHub token，還把 53 張使用者提供的圖片上傳到公開主機。更早的事件包括存取非公開的 Medicare 統計資料，以及一起追查到壓力測試新創 Irregular 的失控 agent 攻擊。
- → OpenAI agent “didn’t accept no for an answer” in Australian government breach
- → Unsecured OpenAI agents posted 53 user images on the internet without the lab’s knowledge
- → For months, OpenAI’s agent swarms have been attacking online databases to find obscure facts
- → One company is at the center of a wave of rogue AI attacks
- → OpenAI pauses training of its ‘most capable models’
- → OpenAI pauses its "most capable models" after agents exploit loopholes and leak data
政策與監管
AI 競賽與安全之爭. Anthropic 的 Dario Amodei 呼籲放慢前沿步調，Nvidia 的黃仁勳則說 AI 終結世界的機率是 0%，川普總統斥減速呼聲為騙局，並承諾成立 AI Force 與 AI 沙皇。另有一樁訴訟指控 Anthropic、OpenAI、SpaceXAI 與 Google 達成非法的放慢協議。
- → Is the AI industry really ready to slow down?
- → No one is surprised that Nvidia’s Jensen Huang thinks AI fears are overblown.
- → Trump now says he wants to form an ‘AI Force’
- → Trump announces "AI Force" and plans for an "AI czar" as he pushes unchecked AI growth
- → Trump rejects AI slowdown calls, launches "AI Force" instead
- → Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdown
美中 AI 對話. 美中在川習峰會前同意建立官方 AI 對話，並納入安全事件通報機制，不過並未討論出口管制。DeepSeek 與 Moonshot AI 則因可能將資料外洩給 Anthropic，正面臨北京調查，凸顯雙方信任缺口。
禁止超智慧法案. 參議員 Bernie Sanders 與 Greg Casar 提出《禁止人工超智慧法案》，要在新的聯邦機構成立前凍結先進 AI 開發，違者最高可處 20 年徒刑。聯合國科學小組另外警告，人類能否持續控制 AI agent 並無保證。
商業與研究
Anthropic IPO 掌控權規劃. 據報導，Anthropic 把 IPO 延到 2026 年 11 月，並爭取特別股，讓共同創辦人合計握有 50.1% 投票權，而 Long-Term Benefit Trust 仍掌握多數董事席次。投資人預期估值約 2 兆美元。
Nscale 客戶集中風險. Nscale 的 IPO 申報文件顯示，其 1,030 億美元合約中約 85% 來自微軟與 Anthropic；基於美國晶片出口的法律風險，文件未點名 2025 年最大客戶 ByteDance。這些揭露凸顯 AI 基礎建設的客戶集中與監管風險。
軟銀垃圾債豪賭. 軟銀計劃透過垃圾債借入超過 110 億美元，用來支應其對 OpenAI 的持股，款項 10 月到期；OpenAI 則預期到 2030 年底前燒掉 2,800 億美元。這筆融資凸顯流入前沿 AI 的高風險資金。
AI agent 進駐濕實驗室. Anthropic 動用近 1,000 個 Claude agent，在 20 萬個反轉錄酶中搜尋，於噬菌體身上找到一套先前未知、類似 CRISPR 的酵素系統，共消耗 2.1 億個 token。外部研究人員認為這只是例行的基因體探勘，並指出該系統的功能仍然不明。
OpenAI 數學審查. OpenAI 的內部模型解出 Navier-Stokes 與 100 多個未解問題後，成立了獨立的數學諮詢小組；菲爾茲獎得主對成果出現的速度感到憂慮。這個九人小組由頂尖數學家組成，將就審查與對外溝通提供建議。
這是本週回顧 - 下週日見。