OpenAI’s math and physics claims could be historic, and handpicked access is slowing their verification

The Verge reports that OpenAI announced three results touching some of research’s hardest open problems. One claims to resolve the Kakeya conjecture in all dimensions; a proof of the three-dimensional case was among the work that earned NYU mathematician Hong Wang a Fields Medal this year. Another claims to disprove the exponential time hypothesis, which is closely tied to P versus NP (our analysis of the ETH claim). A third concerns quantum chromodynamics and Yang-Mills theory, which sits at the center of a Millennium Prize problem, though OpenAI’s work does not address the unresolved mass gap. The physics work hasn’t been released.

Experts told The Verge the work deserves to be taken seriously. Brown’s Brendan Hassett said the math claim looks well supported while noting he hadn’t yet checked the formal verification. If the results hold, AI systems are now contributing to problems that have stumped researchers for decades.

Science runs on scrutiny, and scrutiny needs access. The release plan relies on handpicked expert “shepherds,” giving different fields different levels of access. Announcing a result while restricting materials to experts the company picks slows verification and lets a powerful lab decide who gets first credit and influence. UT Austin physicist Andreas Karch said the big labs already have enough power and shouldn’t be picking winners and losers.

The Advisory Group on Mathematics and AI, which OpenAI had pledged to consult, warned against exactly this a day earlier, then urged the company to release the physics outputs now that its formal verification is complete. Lean, a system for formally verifying proofs by computer, has become central to checking AI-generated math, which makes withholding formalizations harder to justify.

OpenAI’s previous approach drew criticism too. Earlier in the month, researchers were swamped by hundreds of preprints at once. A smaller, more considered batch is an improvement, and some researchers said so. Stanford’s Jared Duker Lichtman noted that any approach will upset someone.

Georgia Tech’s Xiaoyu He raised the underlying problem: no release strategy fixes the fact that fundamental results are arriving quickly, with little human contribution, in papers that are hard to read and verify. If that scales, fields could depend on a few labs both to generate results and to decide how they’re explained.

The fix is straightforward: release the materials publicly, keep expert groups to help explain them, and fund the people doing verification. Openness turns these results into trusted knowledge rather than corporate announcements. It is the same disclosure argument raised about lab safety evaluators (our analysis). More in our research coverage and OpenAI coverage.