Halperin derives time horizon = RL compute^(γ/3), so RL scaling may still be power law
BasilHalperin · x · 2026-11-09
Basil Halperin works through, live, whether RL scaling suffers from a "log problem" — and concludes time horizon is plausibly still a power law in RL compute:
- Whitfill-Snodin-Becker show empirically and theoretically that time horizon = (training compute)^γ; the shared figure shows clean log-log linear fits (shared slope R²=0.906) for Qwen2.5/Llama 3 across benchmarks
- Toby Ord estimates 100x pretraining compute ≈ 1,000,000x RL compute in capability gains, i.e. training compute = (RL compute)^(1/3)
- Combining: time horizon = (RL compute)^(γ/3)
More from Research
- New benchmarks DecoyCount and ScienceVision probe whether AI sees or reasons — liangchen5518 · 2026-11-09
- AI chip design demo: one prompt for a 64-bit RISC-V processor under 3W TDP — satnam6502 · 2026-11-09
- Video-guided machine translation paper lifts low-resource Indic languages by +1.75 BLEU — kalyan_kpl · 2026-11-09
- Peer-reviewed paper applies truth maintenance systems to LLM memory belief revision — letandrewcook · 2026-11-09
- A cathedral that sings: 1D acoustic sim shows voices fighting their own resonant body — Sauers_ · 2026-11-09
- Survey by 30+ authors maps modern NLP tokenization, from subword wins to its flaws — rdesh26 · 2026-11-09