Toby Ord: RL scaling likely power law in capabilities, citing horizon-compute paper
tobyordoxford · x · 2026-11-09
Toby Ord pushes back on Basil Halperin's claim that RL scaling suffers a log-x problem, arguing the evidence better supports power-law scaling in capabilities: Whitfill-Snodin-Becker show empirically (their fig. 3) and theoretically that task time horizon follows a power law in training compute, i.e. time horizon = (training compute)^gamma. Ord notes he had previously read the paper only for its compute-slowdown estimates and missed this relationship, and says he is thinking it through live.
More from AGI Musings
- "Pedestal Cope": a 9-type taxonomy of humanity's false AGI "plot armor" — danfaggella · 2026-11-09
- Dev argues AI refusing tasks over legal risk is overreach that 'manipulates humans' — ezshine · 2026-11-09
- Google Shopping Holiday 100 goes live, product clicks now route to AI Mode — gaganghotra_ · 2026-11-09
- The real problem with conferences is unqualified reviewers, not LLM use — LucaAmb · 2026-11-09
- antirez: devs refusing AI coding help are visibly falling behind — antirez · 2026-11-09
- AI biweekly recap: METR finds 2x faster R&D at Anthropic, Apple cuts engineers' AI spend caps — emmanuelvivier · 2026-11-09