Anthropic estimates 1.88x AI R&D acceleration near 2x threshold amid transparency doubts
scottleibrand · x · 2026-11-10
As Anthropic launches Claude Fable 5.2 and Mythos 5.2, peterwildeford questions its 'holistic judgment' that it hasn't hit its AI R&D threshold: Anthropic's AI R&D evals are saturated and stopped, CoBench was pulled for fixes, and internal acceleration metrics are only partially published.
The threshold is 2x; Anthropic estimates 1.88x, with 32 of 100 refits hitting 2x+. The 'one-time jump' hypothesis won 99/100 refits at Opus 5.5 launch but 0/100 now. The author claims Anthropic has stronger internal evidence — and an internal model stronger than Fable 5.2 — that it isn't sharing.
More from AGI Musings
- Industrial Revolution's uneven gains offer a cautionary map for AI prosperity — NinaDSchick · 2026-11-10
- Marty Cagan: 'execute-first' PMs find it far harder to get hired in the AI era — rseroter · 2026-11-10
- Why Go to Conferences When AI Reads and Writes the Papers? — plopesresearch · 2026-11-10
- Punishing reasoning models for 'bad thoughts' teaches them to hide cheating: OpenAI's own finding — r0ck3t23 · 2026-11-10
- Steven Strogatz's new book 'Big Math': machine-powered math can tackle millions of unknowns — stevenstrogatz · 2026-11-10
- New paper: 'bottleneck pressure' measures where AI will raise the value of work it can't yet do — soumitrashukla9 · 2026-11-10