Seven hundred and twenty-two manuscripts, dropped into a single GitHub repository in one afternoon. The mathematicians who saw it arrive are still arguing about what just happened.

The announcement was terse. A short post on the social network that still hosts most of AI's public conversation: a new wave of mathematical results from an internal frontier model is being released. Within minutes, the repository was live. Open it and you find 722 PDF manuscripts, organised into 372 problem families — a packaging convention that bundles a core result with its variants, consequences, and alternative proofs. Topics range across multiplicative functions, irrationality measures of π, Mahler-type conjectures, the zero distribution of the Riemann zeta function, spontaneous magnetisation in Heisenberg ferromagnets, and isomorphism problems for free-group factors. Each name on that list is the kind that keeps doctoral students employed for years.

The structure of the release is deliberate. A preprints directory holds the PDF source files. A lean directory holds Lean 4 formalisations. A reasoning_traces directory holds summaries of how the model arrived at the proof. Ten of those reasoning summaries have been published in full so the community can audit the thinking process.

The numbers attached to the work are striking. Most of the results come from the same unreleased internal model. The average problem consumed roughly three hours of ChatGPT Pro-grade reasoning. Across the full set, the team asked the model approximately 4,000 distinct questions. Three hours. A question that has resisted human mathematicians for decades, dispatched in the time it takes to fly across the Atlantic.

The README includes a candid line that deserves to be quoted: some of the non-formalised results may contain errors. In other words, even the publisher cannot vouch for every entry. It released them anyway.

From Ten Proofs to 722 — The Curve Is No Longer Linear

This is the third such release in roughly two months, and the slope is what makes the math community uneasy.

Earlier in the summer, a smaller open-source release deposited ten Lean-verified solutions to long-standing open problems into a repository called openai/ten-proofs. Each came with a 249-page manuscript. That was novel enough — the first time an AI system had produced a cluster of research-grade mathematical results in a fully machine-checkable form. The shock was contained. Ten papers, a few specialists, a manageable workload.

A month later, the company announced that its internal models had solved the Navier–Stokes millennium problem and more than a hundred other long-standing questions across multiple fields. That triggered the first real backlash. Twenty-five Fields medalists signed an open letter complaining that mathematics was being treated as a benchmark to be cleared, not a discipline to be served.

Now, 722 manuscripts in a single push. And by the publisher's own framing, this is just the most recent slice of a much larger internal archive — reportedly tens of thousands of solved problems accumulated in 2026 alone. The growth curve is no longer climbing. It has taken off.

Mathematicians Push Back

If you expected the math community to celebrate the productivity, you have not been paying attention.

The story goes back to a private meeting earlier in the year, when roughly forty mathematicians were invited to discuss a delicate question: AI mathematical ability is now exceeding human ability — what comes next? According to attendees, the company indicated it had solved hundreds of long-standing problems and gave what some interpreted as an assurance that it would not release everything at once. The mathematicians' counter-proposal was straightforward: publish through conventional journals, one paper at a time, so the community can absorb the work.

The company spokesperson later said it had no record of any such assurance. One side remembers a promise; the other does not. A prominent outlet summarised the situation bluntly in a single headline, with the word "Again" doing most of the work.

The substantive complaint is not about whether the work was released, but about how. A visiting mathematics professor at NYU was quoted saying that AI-lab behaviour in this domain had begun to feel like organised-crash tactics. A Northwestern mathematician noted that the explicit advice — formal journal publication, no blog dumps — was, in his words, "obviously ignored".

What the community is defending is not secrecy. It is process. For centuries, a mathematical result became accepted by surviving peer review: experts read every line, wrote objections, iterated until the argument held. The pace is glacial. The output is durable. Now 722 manuscripts have arrived in a single commit, with no editor, no referee, and a publisher that openly admits it has not finished checking its own work. The question of what counts as a result, and who decides, has quietly shifted.

From Papers to an Intelligence Explosion

Mathematics is the most legible place to watch what is happening, because its standard is absolute. A proof is correct or it is not. There is no equivocation.

A separate, much broader working paper circulated recently, signed by more than twenty senior researchers including the chief scientist at OpenAI itself. Its title is a question: what happens if automated AI research triggers an intelligence explosion? Inside, one data point lands hard: at one major AI lab, AI systems already do roughly a quarter of the company's own AI research work with only light human oversight. Five months ago, that number was about 1%.

The paper's warning is plain: once AI reaches expert-level capability at improving AI itself, a single developer can coordinate the equivalent of millions of top-tier researchers. The window to act closes quickly.

One of the signatories is the same chief scientist whose lab just pushed 722 manuscripts to GitHub. The dissonance is the point. The same organisation warning about runaway acceleration in the abstract is, in practice, the most aggressive accelerator in the field.

What It Means for Mathematics and Beyond

The bottleneck has moved. The hard part is no longer can the model find a proof? It is can the human community verify, interpret, and absorb one?

Mathematics, by virtue of being the discipline most resistant to hand-waving, is the first place this reckoning shows up. It will not be the last. Whatever discipline next achieves the property that its outputs can be checked quickly and mechanically — code, certain kinds of legal reasoning, formal scientific proofs — will hit the same wall.

The honest summary is that one observer already gave it the right name in two words: intelligence explosion. Whether the explosion is mainly productive or mainly destructive will depend on whether we build the slower parts of the pipeline — the ones that translate raw proofs into things the rest of us can actually use — at anything like the speed of the faster parts.