AI agents solve 98% of research-level theoretical physics problems at under $3 each on new TheoPhysEval benchmark
robleclerc · x · 2026-10-15
The arXiv paper TheoPhysEval (2610.16733, 34 authors including prominent theoretical physicists) introduces 50 unpublished, verifiably answerable research-level theoretical physics problems, each solvable by a world expert in about a day.
- The best agent, GPT-6-Astra harnessed in Codex, reached 98% final-answer accuracy with realistic tool access (literature search, numerical/symbolic scripting), at under $3 per problem.
- Accuracy dropped only slightly when web search and background papers were withheld.
Commentator Rob Leclerc argues theoretical physics — not physics itself — is what AI will 'cook': the physical world becomes the bottleneck for progress, making him bullish on automation, semiconductors and manufacturing, with capital saved in software redirected to unlock physical bottlenecks and turn low-margin businesses into high-margin growth stories.
More from AGI Musings
- AI workshop for 100+ lawyers: watching an associate acquire a 'founder's edge' — jkubicki · 2026-10-25
- AWS CEO Andy Jassy says AI agents will force companies to rewrite almost all their code — VraserX · 2026-10-25
- "AI can never do this": why kindness and loyalty remain a human-only job — alfred_lua · 2026-10-25
- Leaked Liang Wenfeng interview: Anthropic's lead is cyclical, Chinese AI will be systemically cheaper — teortaxesTex · 2026-10-25
- Apollo chief economist: AI has delivered "zero payoff" outside Big Tech — Polymarket · 2026-10-25
- 141-word SIGN Act proposal would jail AI CEOs who lie about model safety — StephenLCasper · 2026-10-25