Toby Ord and Ramez Debate Scaling Curves: Halved Slope, Massive Generational Leaps
Oxford philosopher Toby Ord and Ramez Naam examined the scaling curves of new reasoning models, revealing a paradox: the slope flattens even as generational leaps grow enormous. Ord plans a detailed writeup, and both speculate on how RLVR gains are distributed.
Confirmed
- Ord cites his rule of thumb: every 10x more RLVR compute cuts the tokens needed to reach a given performance level by 3x (paper linked).
- The new model's leap over older ones equals 100x tokens of prior models, which Ord frames as roughly four generations (e.g., o1→o3→GPT series).
- The new model's reasoning scaling slope is about half of o1/o3 on AIME: going from 20% to 80% requires 10,000x compute (tokens).
- Ord stresses that logarithmic inference scaling still holds as a fundamental law.
Unconfirmed
- Ord speculates the halved slope may stem from greater difficulty variance in the new dataset, but this cannot explain the large generational jumps.
- Ramez guesses the model was heavily RL-trained on formal math, skewing gains toward math; Ord also notes gains seem concentrated in math. These remain unverified speculation.
Why it matters
- Ramez argues the huge jumps are worth 2-3 orders of magnitude of extra token spend despite the worse slope.
- If gains are indeed math-biased, general-domain improvements may be overestimated — relevant for evaluating real model capability and future RLVR investment.
2026-09-14 ~ 2026-09-14 · 5 related posts
Primary sources
- New scaling curve has half the slope: 10,000x compute for 20%-to-80%, but bigger generational jumps — tobyordoxford ·
- New generation's jump equals 100x tokens on old model, like four generations in the o1-to-GPT-5 era — tobyordoxford ·
- ramez: generational jumps pay for 2-3 orders of magnitude of tokens, suspects heavy math RL — ramez ·
- [source] New scaling curve has half the slope: 10,000x compute for 20%-to-80%, but bigger generational jumps — tobyordoxford · 2026-09-14
- [source] ramez: generational jumps pay for 2-3 orders of magnitude of tokens, suspects heavy math RL — ramez · 2026-09-14
- Toby Ord: 10x more RLVR compute cuts tokens-to-target ~3x; gains may be math-specific — tobyordoxford · 2026-09-14
- [source] New generation's jump equals 100x tokens on old model, like four generations in the o1-to-GPT-5 era — tobyordoxford · 2026-09-14
- Halved slope may be an artifact of higher-difficulty-variance dataset, log scaling still holds — tobyordoxford · 2026-09-14