Toby Ord: METR's sensitivity analysis weighs a vertical jump from the Mythos pre-train
tobyordoxford · x · 2026-11-06
Toby Ord notes METR's full write-up includes a strong sensitivity analysis of which assumptions its measured speedup relies on, including an alternative hypothesis of a vertical capability jump caused by the Mythos pre-train. He adds that before reasoning models and RLVR training, rare big new pre-trains were where most new capabilities arrived in sudden jumps, so calculating real per-annum capability gains requires zooming out across multiple jumps.
More from AGI Musings
- Reddit criticizes AI-danger debate: "AI can't create hardware" dodges Yampolskiy's point — MustBeSomethingThere · 2026-11-06
- AI is coming for analysts who only wait for clean data, says data lead — Nasereliver · 2026-11-06
- 'Why learn if AI can do it' is dangerous — deep engineers stay most valuable — Abhishekcur · 2026-11-06
- AI stans can't accept informed criticism: "criticism is a narrative violation" — menhguin · 2026-11-06
- Science: cracks emerge in Lean certificates backing AI-generated math proofs — neurovium · 2026-11-06
- AI 2027 tracker claims 85% prediction accuracy, says progress is ahead of schedule — connoraxiotes · 2026-11-06