Toby Ord pushes back on METR's claim that AI acceleration isn't driven by training compute
tobyordoxford · x · 2026-11-07
Oxford philosopher Toby Ord debated METR's un hedged claim that the acceleration in AI task-completion length was "not due to an acceleration in the scale of training compute." Ord argues he isn't claiming compute scaling accelerated — rather there was another step on the staircase of compute scaling, and part of the pre-train jump came from better data and algorithms. He contends METR's slope interpretation implicitly assumes the new pre-train run had zero level effect, which the shown data can't rule out.
More from AGI Musings
- perrymetzger frames AI as a race between totalitarian control and progress — basedjensen · 2026-11-07
- "Abstraction laundering": a new name for the 'you can't ban linear algebra' fallacy — AaronBergman18 · 2026-11-07
- Pierce Lilholt: the ability to design life doesn't grant the wisdom to decide how it should be designed — PierceLilholt · 2026-11-07
- Two scaling laws collide in robotics: exploding data collection meets self-improving AI — Scobleizer · 2026-11-07
- Musk: Going all-in on autonomy is a blindingly obvious move, everything else is 'horse carriage variants' — elonmusk · 2026-11-07
- TimnitGebru mocks AI safety teams by resurfacing longtermism's rich-country-lives-first quote — PolarBearby · 2026-11-07