Anthropic redacts internal audit as METR assessment flags rapid AI R&D acceleration
Wes Roth · youtube · 2026-11-06
Wes Roth's video pulls together Anthropic's latest safety and self-improvement developments:
- Redacted audit: Large portions of the R&D section in Anthropic's August 2026 risk report were redacted, with co-founder Sam McCandlish publicly addressing the redactions and the METR audit, raising transparency concerns.
- Independent METR assessment: METR published an assessment of Anthropic's AI R&D acceleration plus a review of the risk report's R&D section.
- Model 4 and benchmarks: The video claims an internal model "broke all benchmarks", and covers the 2x safety line, AECI benchmarks, trend inflection, compute allocation, and deployment of 30,000 AI agents.
- Recursive self-improvement: The Anthropic Institute released "When AI Builds Itself" and work on measuring the pace of AI development; Daniel Kokotajlo commented on the S-1 RSI filing.
Overall takeaway: frontier labs' autonomous AI research capabilities are accelerating, and tensions between external auditing and public disclosure are growing.
Related event: METR evaluation of Anthropic's accelerated AI R&D shakes the community(81 posts)→
More from AGI Musings
- Daniel Kokotajlo slams METR's confidentiality norms as 'self-sabotage' — DKokotajlo · 2026-11-07
- eigenrobot on the contemporary mythos: denying beauty's reality as a developmental failure — eigenrobot · 2026-11-06
- Researchers Slam AI Peer Review: LLMs Find 100+ Limitations but Miss Critical Issues — IanArawjo · 2026-11-06
- If AGI is so close, why is everyone leaving? Inside the bets of 45 AI neolabs — The Decoder · 2026-11-06
- Professor proposes sit-in exams and rec letters to certify AI-free problem-solving skill — PTenigma · 2026-11-06
- "College kids are screwed": founder reverses stance as AI widens skill gap — RachelVT42 · 2026-11-06