Toby Ord: Chinchilla-optimal adjustment makes the compute-horizon power law fit even better
tobyordoxford · x · 2026-11-09
Toby Ord points out that Whitfill-Snodin-Becker follow their figure with a version adjusting training compute to Chinchilla-optimal-equivalent, controlling for over- or under-trained models. The power law fit improves — shared slope R² rises from 0.906 to 0.949 — further supporting a power law between time horizon and training compute.
More from Research
- Even top 3D-gen models keep making the same basic mistakes, researcher notes — teortaxesTex · 2026-11-09
- ELLIS seminar: trajectory-aware training for masked diffusion LMs, Nov 23 — AnnaCiaunica · 2026-11-09
- Berkeley-led paper: evolved agents gain 5.2 reliability points while losing 10 tasks they once solved — rohanpaul_ai · 2026-11-09
- Community Chinese translation of the Ultimate MoE Training Handbook is out — Nouamanetazi · 2026-11-09
- Stanford CCRMA DSP legend Julius O. Smith gives 3-hour podcast interview — teropa · 2026-11-09
- Hardware-aware LDM implementation cuts memory traffic via state unwinding and windowed writes — SonglinYang4 · 2026-11-09