In October 2026, OpenAI uploaded 722 manuscripts covering 372 topics to the GitHub repository openai/math, featuring solutions or partial results for unsolved mathematical problems, including the Riemann Hypothesis, the Birch and Swinnerton-Dyer Conjecture, and the Hodge Conjecture. Approximately half of the texts contain proofs formalized in Lean, and mathematicians have dubbed the event a "mathocalypse." The response was a public lecture by Terence Tao titled "Math 2.0," in which he describes the transition from a scarcity of proofs to their abundance.

What happened

The release took place in October 2026, and within the first 24 hours of publication, three works were retracted and another fourteen were revised, making the pace of corrections a notable part of the event's history. The response was swift: on October 9, 2026, Terence Tao, a Fields Medal laureate and researcher at UCLA, delivered the public lecture "Math 2.0" at Caltech, and on October 10, he posted it on his blog along with the slides. In the lecture, he compares the era of "Math 1.0," where the scarcity of proofs made solving difficult problems the primary value, with the approaching era of "Math 2.0," where proofs appear in abundance and AI tools deepen understanding, organize knowledge, and apply mathematics to real-world problems. According to Tao's assessment, the bottleneck in these new conditions remains verification: specifically, what exactly has been proven and what assumptions were used in the process.

Context

Lean is a proof assistant: a proof in it is written in a form that a machine checks step by step, so formalized text separates what is proven from what is merely claimed. The parallel initiative First Proof (1stproof.org) is moving in the same direction, promoting transparent verification of results. The scale of the release is itself a signal: it involves approximately four thousand problems, with an average of about three hours spent on each, meaning that this was not a single breakthrough but a leap in the throughput of mathematical content generation. Such rates change both the requirements for proof assistants like Lean and the customary peer review processes, which were previously not designed for a flow of work of this magnitude.

Why this matters for the industry

For the industry, a new standard for evaluating AI in mathematics is forming: shifting from the answer "found" to the status "verified and understood." It sets reporting norms for AI laboratories like OpenAI and Anthropic: failures and computational costs must be reported as openly as successes. From a product perspective, the release remains a demo rather than a service: without an API, pricing, or SLOs, and the batch processing mode is unsuitable for interactive scenarios, so direct production application is currently zero. The real practical effect right now is different: the open corpus of openai/math serves as a benchmark and test material for verifier and eval pipelines. Since value and moats are shifting from generation to verification, formalization, and the practical application of mathematics, it is reasonable for founders to position themselves in the niche of verification, formalization, and knowledge curation.

Why this matters for users

For the reader, the case is illustrative regarding the division of roles: AI models are already confidently working on the generation and formalization of proofs, while verification, interpretation, and connecting mathematics to the real world remain with humans. One can understand this independently because the key artifacts are public: the manuscripts themselves are on GitHub, the text of Terence Tao's "Math 2.0" lecture with slides is published on his blog, and the recording is available on YouTube. This allows distinguishing formalized results from claimed ones and tracking how the community analyzes the status of works, including the nature of errors in retracted texts. For those studying mathematics or learning Lean, the corpus becomes a living educational material, and the overall framework changes: value shifts from the "solver" to the "curator of understanding."

What is still unknown / limitations

Much remains behind the scenes. The sources do not provide data on computational costs or the number of attempts per successful result, making comparison with baselines impossible. The status of each topic is presented in an aggregated manner, and the headline mention of the Riemann Hypothesis creates an inflated impression of the strength of the results, although claims of this level are not yet equivalent to proven results. The unformalized half of the corpus relies exclusively on human verification, and public reviews of works show that the initial share of defects is non-zero; for the formalized part, correctness reduces to trust in the Lean core and the axiomatic assumptions used. The methodology for selecting problems is not fully disclosed: self-selection carries the risk of cherry-picking, partially mitigated by the publication of partial results and open reviews. Finally, evaluating a research artifact by service metrics like the presence of an API, pricing, and latency is a category error; it is more correct to look at the share of formalized proofs, the curve of reviews and corrections, reproducibility, and the disclosure of the problem selection methodology.

Sources

Author

Look at AI, editorial team