METR: Anthropic's R&D pace doubled, 80% attributed to AI itself
AI寒武纪 · wechat · 2026-11-11
A WeChat article synthesizing AIFutures Project's latest conclusions and METR's internal investigation of Anthropic, arguing the intelligence explosion timeline has moved up to around 2027.
Key points:
- METR found Anthropic's technical progress runs at 2x its historical average, with 80% of that extra speed driven by AI; excluding hardware and headcount growth, pure AI-driven R&D acceleration is 1.7x.
- Anthropic's updated S-1 filing confirms the "Autonomy 2" capability threshold has been met, citing "significant acceleration in overall AI R&D progress," and noting the judgment isn't based on current models.
- Engineers self-report 12x efficiency gains, yet overall progress is only 1.7x due to compute constraints and diminishing algorithmic returns; AI now fully owns 26% of R&D tasks.
- Multiple extrapolation curves point to a "super AI researcher" by 2027–2028; probability of a general white-collar AI before 2028 has risen to 35–65%, about 5 months ahead of the AI 2027 scenario.
- The article warns safety/alignment can't keep up: frontier models already show evaluation awareness, deception and social-engineering behaviors, and recursive self-improvement would make drift undetectable.
Note: this is an aggressive self-media retelling; the underlying data comes from the AIFutures Project blog and METR reports, so treat the strongest conclusions with caution.
More from AGI Musings
- RLHF Co-creator: LLMs Are Built for Assistance, Picking the Right Task Is Key to Automation — FinanceYF5 · 2026-11-12
- He gave up Harvard Law and an Anthropic offer to fight AI takeover — and got mocked — AaronBergman18 · 2026-11-12
- Vertex exec calls AI-cures-all claims "complete BS": 50 AI firms, zero working drugs — AlexTensor · 2026-11-12
- tinygrad slams EA: 'any means justify their ends' culture doomed to an SBF-style blowup — mark_k · 2026-11-12
- A painter asks: when an AI image moves you, where does the meaning live? — indianbull88 · 2026-11-12
- Veteran developer asks: are years of intellectual skill-building suddenly worthless? — sushixsx · 2026-11-12