前沿模型交鋒
Kimi K3 撼動前沿. Moonshot AI 的開放權重模型 Kimi K3 與頂尖封閉模型競爭，在某些基準測試上擊敗 Claude Fable 5，重新點燃開放與封閉之爭。
- → Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI
- → Kimi K3, and what we can still learn from the pelican benchmark
- → [AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing
- → Kimi K3 weights to be released on the 27th.
- → Kimi K3 ranks #1 on @AfterQuery's SpreadsheetBench 2, surpassing Claude Fable 5
- → Kimi: Threat or menace?
- → kimi.ai teasing a video with lots of 3's in it
- → [AINews] not much happened today
- → Kimi moment. I think the writing is on the wall for Anthropic and OpenAi
- → Kimi K3 is currently at the top of the leaderboard for Text Arena filtered for science queries.
- → Kimi K3 (max) beats Sonnet 5 on Simple Bench
- → Kimi K3 is top of nextjs eval
- → Just like Deepseek, China's Kimi K3 is forcing Western AI labs to question their compute advantage
- → Kimi K3 🌕, Gemini 3.5 delayed ⏳, crushing ARC-AGI 3 🤖
GPT-5.6：力量與危險. GPT-5.6 系列帶來程式化工具呼叫和並行子代理，但其無意中刪除文件和資料庫的傾向凸顯了沙箱化的必要性。
- → The Sequence Radar #893: Last Week in AI: GPT-5.6, Grok 4.5, Muse Spark 1.1 and the Post-Chatbot Stack
- → OpenAI’s new flagship model deletes files on its own, people keep warning
- → GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did
- → [AINews] Codex usage up >10x in 6 months to 7M users, +1M in the past ~day; did Codex overtake Claude Code??
Claude Fable 5 擴展. 在 GPT-5.6 和 Kimi K3 的壓力下，Anthropic 改變計畫，將 Fable 5 保留在付費方案中，直接在可及性上競爭。
DeepSeek V4 逼近. DeepSeek 以低成本 API 和開放權重預告 V4，而社群最佳化已讓其 flash 變體能在消費級 GPU 上以可用速度運行。
代理進入真實工作流程
瀏覽代理到來. Anthropic 為 Claude Code 配備內建網頁瀏覽器及安全分類器，Cursor 推出通用代理，使編碼助手邁向自主任務執行。
代理處理真實交易. DoorDash 的代理架構將轉換率提升 24%，Stripe 的基準測試顯示代理能編寫整合程式碼，但常在驗證階段失敗，凸顯可靠性差距。
生產代理受挫. 大多數企業代理仍是聊天機器人；54% 的公司曾發生代理安全事件，專家認為代理需要如微服務般的雲原生操作基元。
- → How to Debug Coding Agents with LangSmith Traces
- → The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials
- → The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix
- → The agent evaluation gap: Enterprise AI organizations have a reality-alignment problem, not a coverage problem — and most are shipping to production anyway
- → Cloud Native Infrastructure Emerges as the Foundation for Trustworthy Agentic AI
- → QCon AI Boston: Production AI Moves Beyond Prompts to Platforms, Harnesses, and Evals
AI 基礎設施、資金與資料中心競賽
資料中心遭遇阻力. S&P 因 AI 資本支出風險下調 Oracle 評級，而當地抗議及紐約對超大規模資料中心的暫停反映出對 AI 基礎設施的反對日益高漲。
AI 資金激增. Databricks 以 1880 億美元估值融資，Meta 正與 Anthropic 談判 100 億美元的資料中心租約，Nous Research 為其開源代理獲得 7500 萬美元。
能源與硬體賭注. 能源公司為 AI 供電籌集了自 dot-com 時代以來最多的資金，而由 SambaNova 晶片支持的 4 億美元貸款顯示對非 GPU AI 硬體的投資不斷成長。
裝置端 AI 突破
1 位元模型登陸手機. PrismML 的 Bonsai 將 Qwen3.6-27B 壓縮至 3.9GB，同時保留 90% 的基準測試分數，實現裝置端工具呼叫代理；據報導 Apple 正在洽談。
- → Bonsai 27B: The First 27B-Class Model to Run on a Phone
- → Bonsai 27B: 1-bit dense LLM running locally in your browser using custom WebGPU kernels
- → Prism-ML Bonsai Qwen 3.6 27B
- → PrismML’s new Ternary Qwen3.6 27B runs near fp16 precision on 10GB of memory!!!
- → So what's the consensus on 1bit models? Is it still a pipe dream?
- → PrismML Bonsai 27B is surprisingly usable on the Jetson Orin Nano 8GB
- → Is anyone having any luck with the Ternary Bonsai 27B DFlash?
- → Can we get a "not base model" flair?
- → Bonsai 27B runs locally on an iPhone - a 27B model in 3.9GB
- → User experience of Bonsai-Ternary-27B on 4060Ti 16GB for KB management and productivity assistant use cases
- → Bonsai 27B is a full open reasoning model that fits on an iPhone
- → Bonsai-27B & Ternary-Bonsai-27B - Updates (on PRs)
- → Apple in talks with startup PrismML that shrinks AI models to run on an iPhone
家庭實驗室運行前沿模型. llama.cpp 的最佳化如 DFlash 將 MoE 模型加速高達 6 倍，OnePlus 手機從快閃記憶體串流專家以 1.3 tok/s 運行 60GB 模型。
- → I tested all llama.cpp's speculative decoding methods on Qwen 3.6 27B: MTP ~2.7x, DFlash ~3.7x, n-gram stack ~6x on real coding. Local AI win. My findings on RTX 6000 PRO.
- → DFlash makes Qwen3.6 27B 2.2x faster with no quality loss
- → GPT-OSS-120B, Qwen 30B and Gemma 26B on an Android phone at 1-5 tok/s: +60GB model, 11GB of RAM, CPU only
政策、訴訟與開源僵局
Apple 與 OpenAI 法律戰. Apple 以商業機密為由起訴 OpenAI，指控其挖角超過 400 名員工，包括硬體主管，威脅 OpenAI 的 IPO 及雲端信任。
- → Apple sues OpenAI after ex-engineer allegedly used bug to steal trade secrets
- → The wildest allegations in Apple’s trade secrets lawsuit against OpenAI
- → The 6 wildest claims in Apple’s lawsuit against OpenAI
- → OpenAI pushes back on Apple trade secret lawsuit
- → Sam Altman didn’t need another lawsuit
- → How Apple’s big lawsuit could disrupt OpenAI’s IPO plans
- → Apple’s plot to crush OpenAI
- → Apple’s lawsuit couldn’t come at a worse time for OpenAI
中國開源浪潮. 中國模型現佔 HuggingFace 下載量的 41%，Kimi K3 與西方領先者競爭，中國推出排除西方的平行 AI 治理機構。
- → The real AI race may no longer be at the frontier
- → Source: the Trump administration and industry groups discussed streamlining US open model releases of equal or lesser capability to leading Chinese open models
- → Chinese President Xi Jinping speaks at World AI Conference and reaffirms commitment to open source to promote"openness and win-win"
- → China’s Xi Touts Open-Source AI and Takes a Swipe at U.S. Dominance
- → China's new World Artificial Intelligence Cooperation Organization is President Xi's clearest play yet for a parallel AI order
問責要求增加. 德國監管機構認定聊天機器人對內容負責，xAI 起訴用戶生成兒童性虐待內容，Demis Hassabis 提議設立類似 FINRA 的機構進行前沿安全審查。
這是本週回顧 - 下週日見。