Oxford's Toby Ord: AI Token Scaling Still Lags Human Time Scaling

Oxford scholar and philosopher Toby Ord published a new essay, "AI Scaling vs Human Scaling," on October 9, systematically examining a widely assumed analogy: whether giving AI more tokens (longer reasoning compute, chain-of-thought, and tool calls) is equivalent to giving humans more working time. This is the same study relayed across multiple posts (m1–m5), with consistent core content.

Confirmed

• Ord reviewed the limited available evidence, including comparison charts from METR placing human time-scaling and AI token-scaling on the same axes.
• The preliminary conclusion is an efficiency ratio of about 4: the capability gain humans get from doubling their time investment is roughly 4x what AI gets from doubling token consumption.
• Converted, AI needs roughly 10,000x more tokens to match the progress humans achieve with 10x more time.
• Ord also noted the gap is narrowing: AI's token scaling still underperforms human time scaling, but the difference is shrinking.

Why it matters

• The analysis directly bears on expectations for returns from scaling inference compute: if token scaling is significantly less efficient than human time scaling, the cost of simply piling on compute in exchange for capability may be underestimated.
• The author himself stresses the evidence is limited, and the 4x ratio should be treated as a preliminary estimate based on METR charts rather than a conclusion; further comparative experiments may revise the number.

2026-10-09 ~ 2026-10-09 · 5 related posts

• [source] AI scales worse with more tokens than humans do with more time — but is catching up — tobyordoxford · 2026-10-09
• Humans Scale ~4x Better Per Doubling: AIs Need 10,000x More Tokens to Match 10x Human Time — tobyordoxford · 2026-10-09
• Toby Ord critiques AI-vs-human scaling claim: early models and confusing scaling math — tobyordoxford · 2026-10-09
• Toby Ord: 10x tokens may not equal 10x human time — new essay quantifies AI scaling gap — tobyordoxford · 2026-10-09

1 near-duplicate retellings: tobyordoxford