Daniel Kokotajlo: pretraining loss is a checkable task proxying nearly all downstream skills
nitarshan · x · 2026-11-13
Former OpenAI researcher Daniel Kokotajlo argues that "checkable tasks" are central to AI progress:
- Pretraining loss is a checkable task that serves as a great proxy for all sorts of downstream skills — a model with lower pretraining loss will likely perform better across the board.
- AIs have been improving at pretty much everything, including domains they weren't trained on: Claude Plays Pokemon (never trained on Pokemon) and the highly adversarial, noisy social deduction game Blood on the Clocktower.
- "Checkable tasks" isn't binary or a natural kind — it's a continuous variable measuring how expensive and time-consuming it is for an AI company to verify a task.
A substantive contribution to the debate on verifiability-driven progress.
More from AGI Musings
- Debate: why MIRI treated human values as intractable and bet on corrigibility — jd_pressman · 2026-11-13
- Has anyone's p(doom) actually gone down in the past five years? — danfaggella · 2026-11-13
- AI Agent Traffic on Cloudflare Up 1,700% in a Year as Over Half of Web Traffic Turns Non-Human — PeterDiamandis · 2026-11-13
- jd_pressman argues MIRI-style alignment of an AI godhead is distinct from intellectual history of the good life — jd_pressman · 2026-11-13
- ChatGPT's 1.2B WAUs make it the sleeping giant of consumer agents, argues omooretweets — omooretweets · 2026-11-13
- Ben Todd: no evidence AI is slow on fuzzy tasks, doubling-time data doesn't correlate — ben_j_todd · 2026-11-13