RL compute scaling: 10x more RL compute may double task horizons, exponent ~0.3
tobyordoxford · x · 2026-11-09
Toby Ord (Oxford) estimates that in 2025, RL post-training compute grew 10x every 4-5 months while new pre-training likely stalled, with frontier model task horizons doubling every 4 months. Naively, each 10x in RL compute yields 2x longer horizons, implying a scaling exponent of 0.3 — about twice what a rule-of-thumb (RL exponent = 1/3 of pretraining) would predict. He also notes measured curves are snapshots with fixed algorithms/data quality, unlike the historical frontier where both improved, which could affect headline conclusions.
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
- Toby Ord: RL scaling likely power law in capabilities, citing horizon-compute paper — tobyordoxford · 2026-11-09
- antirez: devs refusing AI coding help are visibly falling behind — antirez · 2026-11-09