Toby Ord: 100x inference compute goes 20-80%, but RL training needs ~10,000x
tobyordoxford · x · 2026-11-14
Toby Ord (Oxford) shares an empirical comparison in a thread with Ramez Naam, Ben Todd, and Allan Dafoe on buying benchmark gains with compute:
- Inference compute: going from 20% to 80% accuracy typically takes only 100x compute, implying odds ∝ inferencecompute^0.6. He cites the ARC-AGI-1 leaderboard cost-vs-score curve across OpenAI models as evidence.
- RL post-training compute: on several reasoning benchmarks, the same 20%→80% jump often requires 10,000x compute, i.e. odds ∝ RLcompute^0.3 — an order of magnitude less efficient. An OTIS Mock AIME 2024-25 chart shows base models barely improve with more tokens, while RL-trained models track the 100x-for-20-80% slope.
Takeaway: test-time scaling is roughly 100x more compute-efficient than RL post-training scaling for lifting benchmark performance.
Related event: Oxford's Toby Ord: RL Compute Needs 10,000x Boost to Lift Benchmark Scores(17 posts)→
More from Research
- Cambridge mRNA therapy regrows heart muscle after attack, boosts pumping 10% in mice — kimmonismus · 2026-11-14
- Data curation RL is just SWE envs: code pretraining data increasingly consists of Claude-built, less-sloppy repos — willcb · 2026-11-14
- NeurIPS 2025 best paper: diffusion models start memorizing at ~300 steps per image — Mastbubbles · 2026-11-14
- TestPrune: pruning regression tests to unblock AI-driven autonomous bug fixing — burkov · 2026-11-14
- Toddlers learn new verbs only from responsive partners; a call to rerun the test with AI — AnnaCiaunica · 2026-11-14
- Context beats model choice: Databricks benchmark shows 91% vs 58% for data agents — CautiousUse8597 · 2026-11-14