Morning brief

Tuesday, July 28

1 people · 0 companies · 4 topics

The signal

  1. Anthropic publishes its open-weights position; Amodei calls for chip controls, anti-distillation rules, mandatory safety testing What happened: darioamodei published “Our position on open-weights models” today, Anthropic’s first direct statement since the company sat out the Nvidia-led open-weights letter. He says Anthropic never sought a ban; instead proposes three policies: tighter advanced chip/equipment exports to authoritarian governments, enforcement against industrial-scale distillation, and mandatory safety testing for all sufficiently capable models regardless of whether they are open or closed. Why it matters: The three asks land very differently in practice. Mandatory safety testing on open-weight releases creates a compliance burden that closed-weight labs with existing safety pipelines absorb far more easily than independent researchers or smaller labs. The post hit #1 on HN today with active skepticism in the thread about Anthropic’s competitive motivations. Confidence: Confirmed Sources: Anthropic (primary) · Axios (independent) · TechCrunch (independent) · HN (community, 85 pts / 52 comments)

  2. China MOFCOM calls US Kimi K3 distillation sanction threats “AI hegemonism,” warns of countermeasures What happened: China’s Ministry of Commerce issued a formal statement July 27 rejecting US threats to sanction Moonshot AI over alleged distillation of Anthropic’s Fable to build Kimi K3. Beijing called the US position “lacking factual and legal grounds,” warned it would take “all necessary measures,” and counter-accused US AI firms of training on Chinese model outputs. The Bureau of Industry and Security is formally investigating Moonshot for distillation and banned chip access; no action taken yet. Anthropic previously disclosed that Moonshot generated over 3.4 million interactions with Claude through fraudulent accounts. Why it matters: This is the first formal government-to-government escalation over AI model IP. A BIS finding could restrict Moonshot’s compute access and chill cross-border open-weight collaboration more broadly; analysts say the dispute risks derailing ongoing US-China AI safety talks. Confidence: Confirmed Sources: CSET/MoFCOM statement (primary) · Hong Kong FP (independent) · The Hill (independent) · InsideRetail Asia (independent)

Tools worth touching

virgiliojr94/book-to-skill

  • Does: converts any technical book PDF (also EPUB, DOCX, Markdown, HTML) into a Claude Code skill, extracting frameworks, principles, and anti-patterns rather than producing a flat summary; loads only relevant chapters per query, not the full book
  • Why now: trending today at +366 GitHub stars; the output skill format is compatible with Claude Code, GitHub Copilot CLI, and Amp
  • Try if: you frequently reference a specific technical book while coding and want it available as a slash command instead of pasting excerpts
  • Skip if: your reference material needs live web data or structured retrieval – this is offline and static
  • Links: github.com/virgiliojr94/book-to-skill · guide

Moves on the board

Player Move Strategic implication
darioamodei / Anthropic Published formal open-weights position: no ban, but calls for chip controls, anti-distillation policy, and mandatory safety testing for capable models Mandatory safety testing on open weights creates compliance overhead that favors labs with existing safety infrastructure over independent researchers; clarifies Anthropic’s policy posture without changing its product behavior
China MOFCOM Formal rejection of US sanction threats over Kimi K3; warned of “all necessary measures”; counter-accused US firms of training on Chinese data First government-level escalation over AI model IP; BIS investigation ongoing; could affect compute access for Moonshot and chill broader open-weight cross-border collaboration

Other

  • moeru-ai/airi: +796 GitHub stars today; self-hosted AI companion with real-time voice chat, Minecraft and Factorio gameplay, Web/macOS/Windows; supports 25+ LLM backends including local Ollama. github.com/moeru-ai/airi
  • usestrix/strix: +507 GitHub stars today; open-source AI penetration testing tool; runs coverage against app vulnerabilities using automated AI-driven attack chains. github.com/usestrix/strix

Watchlist

  • Kimi K3 llama.cpp KDA kernel work: preliminary discussion open at ggml-org/llama.cpp#26041; KDA not yet implemented; no merge timeline (opened 2026-07-23)
  • OpenAI ExploitGym / Kill Switch Act: BIS now formally investigating Moonshot for distillation and banned chip access; China MOFCOM issued countermeasures warning July 27; 50-company industry opposition letter on record; no Senate companion to the Kill Switch Act yet; AISI cross-lab cheating investigation still pending (opened 2026-07-23)
  • TRAINS August 1: 4 days out; Meta still absent; no formal consequence mechanism for holdouts confirmed; CISA liability protection (enabling information sharing) expires September 30 adding secondary pressure (opened 2026-07-25)
  • EU AI Act August 2: 5 days out; GPAI/Article 50 enforcement goes live; member-state competent authority designations incomplete (“enforcement gap” at national level); high-risk obligation delay pending Official Journal publication of Digital Omnibus (opened 2026-07-26)