
In early October, OpenAI dropped a small mountain of mathematics onto the world: 719 manuscripts of claimed solutions to hard open problems, generated by its internal AI models. Before anyone could digest them, the mathematicians' verdict arrived — not on whether the proofs are correct, but on whether they count as mathematics at all. Only 10 of the 719 manuscripts included the model's chain of thought. Only about 42% of the proofs had been formalized in Lean, the proof-checking language that is supposed to guarantee correctness. "There is not human understanding of them at the point of release," Harvard mathematician Melanie Wood told TechCrunch, "and now the work begins."
That is the strange new frontier of the AI era: machines that can solve problems no human has solved, while producing results no human currently understands. The debate has split the mathematical world — and it is playing out, in its own distinct register, inside China's developer forums too.
The numbers, as reported by TechCrunch and other outlets: 719 manuscripts, reportedly covering roughly 372 problem families across pure mathematics and theoretical computer science. Just 10 included the model's chain-of-thought reasoning. About 42% had been formalized in Lean — meaning the majority of these claimed solutions arrived without the machine-checkable verification that the field increasingly treats as table stakes.
The timing was awkward. Weeks earlier, at the end of September, the Advisory Group on Mathematics and AI (AGMAI) — a nine-researcher body hosted by the Institute for Advanced Study — had published guidelines for exactly this kind of industrial-scale AI math. OpenAI said it had consulted the group to avoid repeating earlier controversies. Here's how the release measured up against those guidelines:
| AGMAI guideline | What the release did |
|---|---|
| Stop testing advanced math problems on proprietary models | The release explicitly describes evaluating proprietary models on open research problems |
| Human understanding should follow each result | Only 10 of 719 manuscripts included chain-of-thought; most proofs were not formalized |
| Include machine-readable metadata correlating natural-language and formal artifacts | Not done |
| Help fund the human mathematicians needed to make results meaningful | No commitment announced |
OpenAI did follow some of the principles — releasing results quickly and, in a minority of cases, documenting how the models reached their conclusions. But on the central demand, human understanding, the group left the judgment to the community. In its own statement, AGMAI said it is up to mathematicians to assess how well the recommendations were followed. Terence Tao's assessment was blunt: "Problems are being solved autonomously by AI prompters who have no interest in the broader field itself once their initial target is 'solved.'"
Tao's criticism goes deeper than one lab's disclosure habits. In September he joined 27 other Fields Medalists in publishing a public declaration on AI in mathematics (mathandai.org), which argues that the AI industry's habit of treating open problems as a benchmark is "severely misaligned" with the goals of the field. The declaration's case is worth restating plainly, because it is the philosophical core of everything that follows.
Solving a famous problem, the signatories argue, was never just about establishing true or false. It was a signal that a new insight or method had arrived — one the community would then spend years unpacking through talks, simplifications, and textbook presentations until any graduate student could learn it. Mass-producing true/false statements at machine speed, they write, could "destroy fertile ground instead of breathing life into new ideas" — stripping fields bare without planting the understanding that makes future work possible.
This is the distinction the whole controversy turns on: verification is not understanding. A Lean-checked proof tells you the steps are logically valid. It does not tell you why the theorem is true, which is the part mathematics exists to deliver. An AI that hands you a correct proof you cannot learn from has solved the problem and failed the science at the same time. Tao has taken to calling the result "proof indigestion" — more proofs than the field can absorb, correct or not.
There is a second, more technical crack in the foundation. AI models typically produce a proof in two stages: first a natural-language explanation, then a translation of that explanation into Lean code, which the machine checks. A paper released this week by mathematicians at the University of Cambridge and King's College London documents at least two discrepancies between the natural-language proof and the Lean code behind OpenAI's solution to a problem derived from the Navier-Stokes equations — the million-dollar fluid-dynamics problem that has haunted mathematics for decades.
The discrepancies don't necessarily disprove the solution. But they expose the weak link: the model translates its own reasoning into the formal language, and that translation can drift. As the paper's authors conclude, machine-written formal proofs shouldn't be trusted on arrival — they need the same peer review and scrutiny as any other proof. Which brings the whole project full circle: the formalization was supposed to replace human trust with machine checking, but the formalization itself now needs human checking.
While Western mathematicians write declarations, China's developer community is processing the same event the way it processes everything: in long, blunt, funny forum threads. A discussion on LINUX DO — one of China's largest developer forums, posted October 9 and drawing roughly 140 replies and 500 likes — took Tao's criticism as its starting point, via a Chinese tech-media writeup, and ran with it.
Forum users describe the thread in the familiar register of developer gallows humor: jokes about mathematicians being "sent to deliver food like programmers," riffing on the old anxiety that automation eventually comes for everyone, even the geniuses. But underneath the jokes, the same philosophical fault line opens up. One line of argument in the thread asks, in effect, whether we are handing our brains over to AI — 把脑子交给 AI 吗 — and whether a field that outsources its reasoning to machines keeps anything that is recognizably its own. Others push back, in the pragmatic voice the forum is known for: if the machine gets the right answer, who cares how it got there?
What makes the thread worth reading for non-Chinese developers is precisely that split. The Western debate is institutional — declarations, advisory groups, guidelines. The Chinese forum debate is personal and existential: what happens to us, to the people who trained for years to do this kind of thinking, when the thinking gets automated? Both debates are really about the same thing — the gap between getting answers and having understanding — but the forum version is rawer, funnier, and closer to how most working developers actually experience AI: not as a benchmark score, but as a question about what their own skills are for.
You might not care about the Navier-Stokes equations. You should care about the pattern, because it is coming for software next — arguably it is already here.
AI coding agents now produce pull requests at a volume no review culture was built for. The same "proof indigestion" dynamic applies: a flood of plausible, machine-generated artifacts, each one technically checkable, none of them carrying the understanding that makes a codebase maintainable. A test suite passing is verification; a reviewer who understands why the code is right is understanding. The math world is simply the first profession to formalize the complaint.
There is also a lesson in how the Chinese developer community responds to AI hype. Long before this story, forum users built their own folk benchmarks to keep AI honest — a world-map drawing test that detects when a model has quietly gotten dumber, and a 3D pelican test that stress-tests code generation with an absurd one-shot prompt. The instinct is the same one Tao is voicing at the institutional level: don't trust the scoreboard, check the understanding. Communities that build their own evaluations, rather than accepting vendors' benchmarks, end up with a much clearer picture of what AI can actually do.
The 719 manuscripts will eventually be checked, formalized, and absorbed — or quietly forgotten. That is the ordinary machinery of science, and it works. What won't be settled so easily is the question the mathematicians and the forum users are both circling: if AI can produce correct results that nobody understands, do we still get to call the activity that produced them research? Or does it become something else — a kind of industrial answer-mining, valuable but no longer the thing we meant?
Mathematics spent centuries as humanity's record of what it means to understand something. The machines are now very good at producing the record without the understanding. What happens next depends, as the Fields Medalists' declaration puts it, on the decisions of the humans in control of the technology — including the humans who decide what counts as a result worth publishing, and who is responsible for understanding it.