OSVerify referees 11 OpenAI math preprints, catching 5 major and 13 minor issues
james_y_zou · x · 2026-10-22
James Zou's team introduces OSVerify, a one-sided verification framework for scaling scientific discovery beyond human review capacity.
- Key insight: checking a proposed error is easier than judging an entire solution — verification is decoupled into error proposal and error validation.
- Applied to 11 recent OpenAI math preprints, it identified 5 major and 13 minor issues, all confirmed by human experts.
- For the 55-year-old Manin constant conjecture, OSVerify caught a proof error missed by 12 AI checks and provided a repair; paired with Rethlas it independently derived a proof from the statement.
- Thesis: scaling discovery requires scaling verification.
More from AGI Musings
- Eight months on, Citrini's '2028 Global Intelligence Crisis' essay looks prescient — teortaxesTex · 2026-10-22
- Vulnerabilities were always there — AI is just better at finding them — jonmarkgo · 2026-10-22
- 1,000 mathematicians pledge to shun AI in their research — Anen-o-me · 2026-10-22
- If your negotiation AI was built on Western values, is your sovereignty compromised? — 0xsachi · 2026-10-22
- Economist: AI writing volumes will inevitably crowd human readers out of equilibrium — paulnovosad · 2026-10-22
- Econoar on AI Hacking: Short-Term Attackers Win, Long Term It's a 'Hardening Era' — econoar · 2026-10-22