Astra solve-rate barely improves at max compute, undercutting the 'too smart to throttle' RL theory
zainhas · x · 2026-09-05
A heatmap shared by zainhas shows Astra's solve-rate density barely moves when scaling inference compute from low to max, contradicting teortaxesTex's theory that the model is 'too smart' for unsaturated tasks and spirals off-track at full compute (who also argued RL on the base isn't finished).
If that explanation were true, zainhas argues, you'd expect at least some sparks of brilliance on the right side of the heatmap — there are none: 'this model is so alien.'
More from Models
- Data lab bids $12.5M for Spirit Airlines operational data, betting on 'realism' as AI training frontier — Exp_Mark · 2026-09-05
- Theory: Most of Google Astra's Gains Come From Vision, Not Coding — dumquestions · 2026-09-05
- Best-Value AI Subscriptions Under $20/$200, Updated for AA Index v4.2 — popiazaza · 2026-09-05
- antirez: Astra is a big jump for software dev as RLVR-scaled models keep growing — antirez · 2026-09-05
- Heavy user: ChatGPT's Astra intuitively outclasses Claude Fable 5.1 at the same price — DynaBeast · 2026-09-05
- Fable 5.1 refuses questions on Evoscale/Biohub papers, drawing 'Opus'ed' quip — nathanbenaich · 2026-09-05