Math benchmark success shows log-linear diminishing returns with test-time compute, says ramez
sebkrier · x · 2026-09-09
Ramez Krieheheli notes that success rates on math problems exhibit log-linear diminishing returns as test-time compute increases, a pattern seen repeatedly in the field — sometimes a power law, sometimes logarithmic. The takeaway: piling on more inference compute yields progressively smaller gains.
More from Models
- DeepSeek reportedly running four V4 model versions, user says culling is coming to cut inference costs — teortaxesTex · 2026-09-09
- DeepSeek Accused of Silently Swapping Models Behind Its API, Sparking Debate — teortaxesTex · 2026-09-09
- Dev complains about over-blocking safety filters, asks Sam Altman for compute on Collatz — mertdumenci · 2026-09-09
- Dev claims turning off thinking mode makes AI models 5x faster, urges default non-thinking modes — Daniel_Farinax · 2026-09-09
- Yoav Goldberg: OpenAI's "cannot rule out" user data use is a lawyered admission — MannyKayy · 2026-09-09
- Heavy user on Astra: beats a junior hire on cost, but still fumbles simple tasks — RachelVT42 · 2026-09-09