METR: AI R&D capabilities improving fast but getting harder to measure
ben_j_todd · x · 2026-11-21
METR published its Q4 capability update covering models up to Opus 5.5 and GPT-6 Astra, focused on AI R&D-relevant capabilities. Key finding: capabilities are improving rapidly while measurement is getting harder. Pooled 50% time-horizon estimates are 40h (GPT-6 Astra) and 39h (Opus 5.5), with 95% CI upper bounds of 2645h and 1253h; METR assigns 5% odds horizons already exceed 1000 hours.
More from AGI Musings
- Academic says AI has redrawn the domain of human achievement; "not in my discipline" is likely wrong — soumitrashukla9 · 2026-11-21
- Engineer's rant: mathematicians stop at "a recipe exists" and never taste the cake — gerardsans · 2026-11-21
- Wharton MBA deck adopts 'Technology Bottlenecks' paper: same task, 23 vs 83 pressure by job — soumitrashukla9 · 2026-11-21
- 'Humans built this': from LSTM to Transformer in 29 years, no credit to the machines — teortaxesTex · 2026-11-21
- Kevin Hartnett's new book "The Proof in the Code" charts Lean's history transforming math and AI — ipeirotis · 2026-11-21
- Gary Marcus recorded his Anthropic IPO-skeptic podcast days before the $83/share offering — GaryMarcus · 2026-11-21