← 所有 Briefs

🔭 🔭 Shroom Feed — 03/01 13:49 台北

2026-03-01 · 13:49 台北

#anthropic#pentagon#openai#claude-code#karpathy

🔴 Pentagon vs Anthropic — 全面升級

這場 AI 倫理 vs 國家安全的攤牌,在過去 48 小時內急劇升級:

Trump 下令全面封殺:

  • Truth Social 貼文:「EVERY Federal Agency… IMMEDIATELY CEASE all use of Anthropic’s technology」
  • 6 個月過渡期,但態度明確:「We don’t need it, we don’t want it, and will not do business with them again」

Hegseth 正式列為「Supply-Chain Risk to National Security」:

  • 與華為同級別的標籤
  • 「No contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic」
  • 如果這個解讀成立,所有跟美軍做生意的公司都不能跟 Anthropic 合作 — 潛在衝擊巨大

Anthropic 回擊:

  • 發表聲明回應 Hegseth 的公開言論(41K likes / 6.6K RT)
  • 宣布將在法庭挑戰 supply chain risk 裁定
  • Anthropic 聲明

OpenAI 趁勢切入但也聲援:

  • 與 Pentagon 達成 classified AI deployment 協議
  • 聲稱保留同樣的「紅線」(禁止大規模國內監控、自主武器)但結構不同 — OpenAI 同意 Pentagon 可用於「any lawful purpose」,但透過技術措施 + 合約保護來落實限制
  • 新增第三條紅線:禁止「社會信用」類高風險自動決策
  • 公開聲明:「We do not think Anthropic should be designated as a supply chain risk」
  • OpenAI 聲明

其他反應:

  • Ilya Sutskever(OpenAI 共同創辦人、已離開創辦 SSI):「Extremely good that Anthropic has not backed down」
  • Google 員工連署支持 Anthropic,但 Google 官方未表態
  • Claude 衝上 App Store #1 — Streisand effect 滿分(Boris、Thariq 轉推 @mikeyk)

📝 gu-log 已有初報:「Pentagon 威脅砍掉 Anthropic 的 $2 億合約」


🟡 Claude Code — /simplify 和 /batch 新 Skills

Boris 宣布下一版 Claude Code 將內建兩個新 Skills:

  • /simplify — 自動化 PR 到 production 的流程(shepherd a pull request to production)
  • /batch — 互動式規劃 + 平行執行 code migration。用法:/batch migrate src/ from Solid to React。每個 agent 用 git worktree 完全隔離,自測後開 PR

Boris:「I have been using both daily」。10.8K likes,反響極大。

同時 Thariq 也宣布 Friday ships:AskUserQuestion 工具現在可以顯示 markdown snippets(圖表、code examples),並分享 Claude 嘗試用 hashed unicode 模擬顏色的趣聞。


🟡 Karpathy — Multi-Agent「Research Org」實驗

回應 Thom Wolf「為什麼 NanoGPT speedrun 沒被 AI 全自動化」的提問,Karpathy 分享了用 nanochat 跑 multi-agent research org 的實驗:

  • 8 個 agents(4 Claude + 4 Codex),每個配 1 GPU
  • 嘗試不同架構:8 個獨立研究員 vs 1 首席科學家 + 8 juniors
  • Git branch per program、git worktree 隔離、tmux window grid 可視化
  • 結論:不 work。 Agent 實作能力很強,但創意實驗設計很糟 — 不做 baseline、不控制變數、「發現」增加 hidden size 能降 loss 這種 spurious result
  • 核心洞見:「You are now programming an organization」 — source code = prompts + skills + tools + processes

另外也 QT @mntruell 的「Third Era」文章,分享 Cursor Tab vs Agent 請求比例圖:None → Tab → Agent → Parallel agents → Agent Teams → ???

📝 gu-log 已有相關:「Cursor’s CEO: The Third Era of Software Development」


🟢 Simon Willison

  • Agentic Engineering 新章:Interactive Explanations vs Cognitive Debt — 讓 coding agent 建立客製化的互動 + 動畫解釋,對抗 AI 生成 code 帶來的認知負債

    • Tweet
    • 📝 gu-log 已有相關概念文:「Cognitive Debt:AI 幫你寫完了 Code,但你已經看不懂自己的系統了」
  • MLX 創建者 Awni Hannun 離開 Apple — Simon 驚訝 Apple 沒留住他。MLX 讓 Mac 成為跑 LLM 的可信平台


⚪ 其他動態

  • Artificial Analysis:Qwen3.5 27B 系列 benchmark — 27B dense model 在 Intelligence Index 得 42 分,匹敵 8-25 倍大的模型。FP8 只需 ~27GB,4-bit 可跑在 16GB+ 筆電。但 output token 用量異常高(98M vs 同級的 56-61M)

  • SemiAnalysis:AI Cluster GPU 5 年租賃經濟學 — 合約期間 EBIT margin 看起來不錯,但合約到期後 $/hr 預期下降。推薦用 IRR 而非 margin 評估

← 所有 Briefs