New essay: AI scales worse with more tokens than humans do with more time, but is catching up
tobyordoxford · x · 2026-10-09
Toby Ord (Oxford) published a new essay asking whether AI models scale as well with more tokens as humans do with more time. Reviewing the limited available evidence — including a METR chart plotting human scaling with time against AI scaling with tokens — he concludes that AI has so far scaled less well with more tokens than humans do with more time, but appears to be catching up. The claim has direct implications for forecasting how long-task agents improve as test-time compute grows.
Related event: Oxford's Toby Ord: AI Token Scaling Still Lags Human Time Scaling(5 posts)→
More from AGI Musings
- Writing may fork in the AI era: conversational clay vs machine-readable marble — jonippolito · 2026-10-09
- Steve Hsu: physicists are embracing AI disruption far more openly than mathematicians — burny_tech · 2026-10-09
- The real AI-math risk: a slow-to-adapt field could see funding gutted worldwide — rickasaurus · 2026-10-09
- Steve Hsu: getting a life's-work solution from AI deserves joy, not dread — burny_tech · 2026-10-09
- OpenAI model cracks 90 of math's top 500 open problems in one 722-paper dump — Don't Worry About the Vase (Zvi) · 2026-10-09
- Critic blasts Anthropic's use of "abuse" to describe how humans treat software — astralmatrix · 2026-10-09