Essay argues METR cannot meaningfully check Anthropic's pace commitments
kyliebytes · x · 2026-09-16
Very Sane AI Newsletter publishes a pointed rebuttal to Dario Amodei's September essay "We Must Pace the Frontier," which proposed giving embedded third-party evaluators like METR employee-like access to verify safety commitments. The essay's verdict: no.
Key arguments:
- METR is not meaningfully independent: it is closely tied to Anthropic, some staff have strong social ties to AI lab employees, and it works out of Constellation, a shared research center hosting AI lab staff
- METR is insufficiently staffed for embedded oversight
- METR holds no authority over Anthropic that cannot be revoked at Anthropic's discretion — unlike banking-industry regulatory supervisors
The author argues that framing embedded METR as verifiability for pacing commitments is so suspicious it looks like an attempt to evade oversight and sabotage oversight efforts in general; if Dario doesn't truly believe METR can check Anthropic like a government regulator checks a bank, the policy has been badly formulated and communicated.
Related event: Debate Erupts Over METR's Independence in Auditing Anthropic(4 posts)→
More from Companies & People
- Musk explains why Terafab must exist: Taiwan chip risk plus capacity ceiling — elonmusk · 2026-09-16
- Meta's Alexandr Wang: Alignment Is a Prerequisite for the Agent Economy — summeryue0 · 2026-09-16
- Two solid robotics courses recommended: Majumdar's and oier_mees' open resources — micoolcho · 2026-09-16
- Lily Ray warns at SEO conference: automating SEO with AI tools is "incredibly risky" — gaganghotra_ · 2026-09-16
- Google engineers can now use Claude internally alongside Gemini — VraserX · 2026-09-16
- ElevenLabs in one week: UMusic deal, Music v2.5, MCP multimodal, $1M AI SDR — lukeharries · 2026-09-16