OpenAI Just Shattered Mathematics. Here's What We Found.

Summary

OpenAI’s Release

• On October 6, 2026, OpenAI published a large collection of AI-generated mathematical manuscripts and supporting materials on GitHub. The results came from an unreleased internal model.
• OpenAI said the model processed about 4,000 tasks and initially produced 722 manuscripts, grouped into 372 families of related results. These figures do not mean that 722 or 372 separate problems were solved: a family may contain a main result, supporting work, consequences, or alternative proofs.
• The bulk of the work reportedly used the same procedure, averaging about three hours of ChatGPT Pro computing time per result. OpenAI also released brief descriptions of the model’s reasoning for ten results, but not full records of its attempts.
• OpenAI later retracted three manuscripts and edited the collection. It also said that 300 of the remaining 719 manuscripts had main results formalized in Lean—about 42%.

Notable Mathematical Claims

• Progress related to the Riemann hypothesis: OpenAI claimed a proof of a weaker “quasi-Riemann hypothesis” result. Rather than showing that every non-trivial zero of the zeta function lies on the critical line with real part (1/2), the result claims that none lie to the right of (7/8). By symmetry, this confines the zeros to the band from (1/8) to (7/8). A separate companion article used a weaker cutoff of (11/12). These claims would be major progress if established, but they do not solve the full Riemann hypothesis.
• A special case of the Hodge conjecture: OpenAI claimed the conjecture holds for complex abelian manifolds with complex multiplication, across every dimension and codimension. This is a significant claim about a particular family of cases, not a solution to the broader Millennium Prize problem.
• The release also included claims concerning approximations to (\pi), faster large-integer multiplication, and improved bounds for matrix multiplication.

What Lean Verification Does—and Does Not—Establish

• Lean is a proof-checking language and assistant. It can verify that a formal proof follows from its stated definitions and assumptions, step by step.
• However, a verified Lean proof does not automatically establish that the prose in a paper accurately describes the proof or that the formalized statement matches the intended mathematical problem.
• The video discusses an article, “Navier-Stokes: Lost in Translation,” whose authors compared parts of an earlier OpenAI proof with its Lean formalization and found discrepancies. One example involved an estimate: the written argument said four additional input derivatives were sufficient, while the formalized estimate examined by the authors used five.
• That difference does not by itself disprove the theorem; the weaker estimate might still suffice. The authors said they were identifying translation errors, not judging the correctness of the overall proof.
• The central lesson is that formal verification is useful evidence, but readers must still check whether the definitions, assumptions, theorem, and written explanation correspond to one another.

Reactions and Concerns

Reactions ranged from enthusiasm about the mathematical progress to concern about how the work was released and who must verify it.

• Some commentators argued that AI-driven mathematical discovery is inevitable and that producing more results should take priority over protecting the traditional mathematical profession.
• Kevin Buzzard expressed optimism while acknowledging that the papers were difficult to assess and that some had already been retracted. He suggested that truth and understanding would eventually emerge, and that AI could improve how mathematical results are explained.
• Ryan Alveys (as rendered in the subtitles) called the results exciting and argued that mathematics may need to adapt its institutions and practices to a period of abundant evidence.
• Enrico Bertini objected to a perceived double standard: mathematicians are being asked to spend their own time cleaning up company-produced manuscripts, potentially creating value for those companies.
• Tristan Umber described the personal impact of OpenAI’s claimed proof of the Kato conjecture on entropy, a problem central to his dissertation and research plans. He said the announcement caused distress and anger, particularly because the paper was difficult to read and he felt the community was being asked to verify it for free.
• The Association for Humanities Mathematics criticized the mass release, calling it a demonstration of power rather than science, and urged mathematicians to stop collaborating with OpenAI and put human understanding at the center.

Terence Tao’s Vision for “Mathematics 2.0”

Tao’s view, as presented in the video, combines optimism about AI’s potential with criticism of treating problem-solving output as the only measure of progress.

• He describes open problems as “lighthouses”: they guide research into areas where the journey itself can produce useful methods, connections, and insights. If AI reaches these goals too quickly, it could discourage exploration of the surrounding territory.
• He argues that AI should help people understand and develop mathematics—for example, by explaining difficult proofs, connecting papers, building formal libraries, and uncovering new principles—not merely by generating answers.
• The video’s concluding lesson is that getting an answer is different from understanding why it works or learning how to solve the next problem. The challenge, Tao suggests, is not only technical but also one of imagination: deciding how to use AI to support mathematical exploration and human understanding.

Sponsor Segment

The video includes an advertisement for Medly, an AI-powered learning platform. The narrator highlights its adaptive question mode, error tracking, dependent grading, photo-based work checking, new courses, and UK mock exams.

Speakers and Sources Featured

• Ellie Sleightholm — narrator and presenter.
• OpenAI — source of the manuscript collection, repository notes, and mathematical claims.
• Kevin Buzzard — quoted mathematician.
• Ryan Alveys — quoted as an enthusiastic commentator; name as rendered in the subtitles.
• Enrico Bertini — quoted mathematician.
• Tristan Umber — quoted researcher describing the effect of the Kato conjecture announcement on his work.
• Terence Tao — discussed and quoted on “Mathematics 2.0.”
• Association for Humanities Mathematics — source of a statement criticizing the release.
• Authors of “Navier-Stokes: Lost in Translation” — cited for comparing OpenAI’s written proof with its Lean formalization.
• A Reddit commenter — cited for the view that mathematics may increasingly become a domain for large language models.
• Wilder Pugh — credited for a website tracing human mathematicians whose work was cited in the manuscripts.
• Anthropic and Google DeepMind — mentioned as other AI organizations working on mathematics.
• Medly — sponsor featured in the video.