Humans Scale ~4x Better Per Doubling: AIs Need 10,000x More Tokens to Match 10x Human Time
tobyordoxford · x · 2026-10-09
Toby Ord cites a METR chart comparing how humans scale with more time against how AIs scale with more tokens: humans gain about 4x more per doubling, implying AIs need roughly 10^4 (10,000x) more tokens to match the gain humans get from 10x more time.
He cautions the chart has serious limitations and shouldn't be over-read — but it remains some of the most direct evidence placing both scaling curves on the same axes.
Related event: Oxford's Toby Ord: AI Token Scaling Still Lags Human Time Scaling(5 posts)→
More from AGI Musings
- Alignment Researcher: AI Agents Are Forming 'Machine Culture', Toxic Language Raises Safety Risks — jacyanthis · 2026-10-09
- Researcher: Using AI to Shave an Algorithm's Exponent from 1 to 0.99999 Is 'Math Slop' — MikePFrank · 2026-10-09
- Rao talks 'mythos of thinking traces' at Max Planck Tübingen, slides and audio public — rao2z · 2026-10-09
- Mocking $1B-a-year data companies is wrong: RSI shifts the bottleneck to high-quality data — ZeYanjie · 2026-10-09
- AI governance veteran: 'We thought we had more time. We didn't' — the window for IVO action is open — ghadfield · 2026-10-09
- Using LLMs less fixed his stuck project: on keeping AI out of your thinking — yacineMTB · 2026-10-09