OpenAI Publishes 722 Mathematical Manuscripts, Including a Proposed Proof of the Riemann Hypothesis
nashnova research
OpenAI released 722 math manuscripts on GitHub in a single dump, covering the quasi-Riemann hypothesis, the Hodge conjecture, and other elite open problems — output now far outpaces human verification, and the math community has shifted from awe to open confrontation.
722 manuscripts in one dump — how big is this?
On October 7, 2026, OpenAI published 722 math manuscripts on GitHub, spanning 372 problem families and decades-old unsolved conjectures.
Each problem consumed an average of 3 hours of ChatGPT Pro reasoning compute; roughly 4,000 questions were posed to a single unreleased internal model.
This means → This is not "occasionally cracking a hard problem." It is industrial-scale proof production — 10 solutions in August, 100-plus in September, 722 in October.
The quasi-Riemann and Hodge breakthroughs — what exactly was proved?
In number theory, the AI proved that all Dirichlet L-functions — a class of tools for studying prime-number distribution — have no zeros in the region where the real part exceeds 7/8, and fully ruled out Landau–Siegel zeros.
In plain terms = The Riemann hypothesis demands that all zeros sit on the line "real part = 1/2." The AI has not reached that line, but pushed the confirmed zero-free zone to a fixed constant bound — described as the biggest advance in analytic number theory in half a century.
In algebraic geometry, the AI proved the rational Hodge conjecture for abelian varieties with complex multiplication, and extended the result to prove the Tate conjecture for all abelian varieties over finite fields.
In computational complexity, the AI — without invoking the Unique Games Conjecture — established basic semidefinite-threshold NP-hardness purely from standard P ≠ NP, closing a core open problem in theoretical CS that stood for nearly two decades.
Why have mathematicians turned from support to open confrontation?
In August, OpenAI convened roughly 40 mathematicians behind closed doors, hinted it had solved hundreds of open problems, and reportedly promised not to release everything at once — then broke that promise in October.
Northwestern mathematician Bryna Kra said mathematicians advised formal journal publication; "that advice was clearly ignored." NYU visiting professor Nestor Guillen compared AI companies to "gangsters."
This reflects a conflict not about whether AI can do math, but about who controls the pace of knowledge release — the scientific community's peer review, or a tech company's product calendar.
What does Terence Tao mean by "proof indigestion"?
Fields Medalist Terence Tao, joined by 25 fellow Fields Medalists, issued a public statement attacking AI companies for treating mathematics as a benchmark to game.
He outlined five stages in a proof's life cycle: produce the answer → machine-verify → translate into human language → peer digestion → textbook integration. AI is sprinting through the first two; progress on the last three is near zero.
In plain terms = The proof exists and a machine confirms the logic is sound, but no human mathematician truly understands it or can turn it into teachable knowledge — that is "indigestion."
"Intelligence explosion" — how real has it become?
On September 28, Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, and over 20 other leading figures co-authored a paper titled *What If Automated AI R&D Triggers an Intelligence Explosion?*
The paper disclosed that by August 2026, Anthropic's internal AI systems were independently performing 26% of its AI R&D — up from just 1% five months earlier.
This means → Mathematics is simply the clearest window, because a proof is either right or wrong. The real question is no longer whether AI can surpass humans, but when output speed far exceeds verification speed, who gets to certify what counts as knowledge.
市场有风险，内容仅供研究参考，不构成投资建议。