Gary Marcus Slams OpenAI's Vague Math Proof Report: Zero Details, Won't Pass Peer Review
GaryMarcus · x · 2026-10-08
Gary Marcus argues the real news in OpenAI's unreleased-model math proof announcement is what wasn't disclosed:
- The report offers only "Same procedure" and "using an unreleased model"—no details on whether proofs were one-shot and Lean-verified or iterative, no failure rate, no training/post-training/data augmentation info. It would never pass peer review.
- With no methodology, there's no way to gauge generalizability beyond math: it could be a real step toward AGI or just clever leverage of Lean and synthetic data in a verifiable domain with no generality.
- He also criticizes X discourse as an "ignorant cheering section" applauding without asking basic scientific questions.
Related event: Gary Marcus Slams OpenAI Math Proof Report as Too Vague for Peer Review(2 posts)→
More from AGI Musings
- Vets on Claude-built LagosLife: coding was never the moat, product design and marketing are — mfckr_eth · 2026-10-08
- Worldpay's Agentic Commerce Report 2026 surveys 8,000+ consumers across 7 markets on AI-agent shopping — mikeflache · 2026-10-08
- Superexponential publication boom measured across 5M papers and 2.5M researchers — k7agar · 2026-10-08
- After 80 years of AI, humans stopped writing the rules — AI now builds AI — ghadfield · 2026-10-08
- Rejected by his crush, he used Bayesian reasoning to become an ML researcher — tensorqt · 2026-10-08
- What would we do during an AI pause? Let frontier lab staff finally think — JacquesThibs · 2026-10-08