Toby Ord on METR extrapolation: pretraining jumps would mean slope change below 2x
tobyordoxford · x · 2026-11-10
Toby Ord discusses METR task-horizon extrapolation with Yafah Edelman and Peter Wildeford: on priors, both a jump and a slope increase seem likely, but any jump from bigger/better pretrains would mean the slope change is noticeably lower than METR's 2x estimate. He questions why METR focused on a one-off slope change rather than mild convexity, which hasn't appeared; he finds the 60%–85% probability range reasonable.
Related event: Philosopher Toby Ord says METR task-length extrapolation may be optimistic(2 posts)→
More from AGI Musings
- Labs are keeping their most capable cyber, math and bio models internal, warns lvwerra — TheZachMueller · 2026-11-10
- Outside the Western bubble, AI use is widespread and positive, even in official orgs — technollama · 2026-11-10
- 'An expert in every domain in your pocket': Reddit user calls out pure anti-AI resentment — chessboardtable · 2026-11-10
- The History of the History of Algorithms: Practitioner Histories and Teleology — fkasummer · 2026-11-10
- Rat hippocampus maps other rats on the same cells it uses for self — AnnaCiaunica · 2026-11-10
- X users roast OpenAI's obscene profits from an 'accidental' transformer discovery — akbirthko · 2026-11-10