Claim: "Opus 5.5" matches top human experts at picking experiments with 2.3x less compute
CatAstro_Piyush · x · 2026-10-07
Rynorhn argues science has always been bottlenecked by finite human attention — unread papers, untested hypotheses, unnoticed cross-field connections. Millions of tireless agents that read everything and run parallel experiments could turn a civilization of a few million researchers into one with effectively unlimited researchers.
Unverified claims cited: "Opus 5.5" reportedly matched the best human expert results on the TasteVal benchmark at deciding which experiments are worth running, using roughly 2.3x less experimental compute — with that ability said to double every three months.
The key shift isn't AI becoming "a better scientist" but an order-of-magnitude jump in research scale; what discovery rates look like afterward, nobody knows.
More from AGI Musings
- OpenAI reasoning models went from basic arithmetic to decades-old math breakthroughs in two years — daniel_mac8 · 2026-10-07
- The New Skill in the Math-Proof Era: Spending Verification Effort Where Errors Cost — dbreunig · 2026-10-07
- Richard Ngo's hard sci-fi Tinker sketches how AI designs chips for its successors — teortaxesTex · 2026-10-07
- tszzl: Don't be an uncomprehending audience clapping at the AI magic show — Afinetheorem · 2026-10-07
- VraserX: Personal AI agents like Grok Bot, Muse and Dot will soon be everywhere — VraserX · 2026-10-07
- Predicting a 'third thing' RSI: capability leap in ~9 months via automated R&D, short of full FOOM — teortaxesTex · 2026-10-07