A rumour is now a starting gun, and mathematics just fired the first one
OpenAI spent a nine-figure token budget on someone else's problem because it heard they were close, and the fix will come from research contracts rather than from labs behaving better.
The most important sentence in OpenAI's Navier–Stokes post was not about fluids. It was this one: "Our effort began on September 1st after hearing a rumor."
Within four days that rumour had become roughly 10,000 agents, 2.7 million messages and 130 billion output tokens, aimed at a problem two people had been working on quietly for a year. Whatever else the priority dispute turns out to be, it is the first documented case of a frontier lab treating a rumour about a rival's unpublished result as a work order.
Terence Tao's response is the clearest statement of what that costs. Good open problems, he argues, are the scarce resource in mathematics. Not solutions. "We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential." His image is a region dying of thirst beside an ocean. You can generate infinite questions, and almost none are worth a year of a human life; the ones that are get identified slowly, by people, and then mined out in a weekend.
Is he right? We think so, and the objection to him is weaker than it looks.
The rebuttal, fairly
Why Terence Tao is wrong on AI + Math. A thread:
1. Net negative is factually false. After a proof or disproof we know strictly more than before. No opaque process reduces the set of knowledge we have. About the human math trajectory "net negative" read the next points.
Salvatore Sanfilippo, antirez to anyone who has used Redis, posted a thread titled "Why Terence Tao is wrong on AI + Math." His first point is the strongest: "After a proof or disproof we know strictly more than before. No opaque process reduces the set of knowledge we have." A theorem is a theorem. If the machine finds it first, humanity has it sooner, and anyone can then study the proof, simplify it and extract the insight at leisure. On this view Tao is mourning a working method, not a loss of knowledge.
True, and it misses the target. Tao's claim is not about the stock of theorems. It is about the process that generates the next good question, and he brings a receipt from his own Equational Theories Project: automated sweeps found small counterexamples to implications that had stumped the team, and "if that counterexample had been found first, we might not have discovered the more interesting method."
The method the humans were forced to invent is worth more than either result, and it exists only because the answer arrived late.
antirez's other argument is that a solved problem makes nearby understanding faster, because "you are a lot more adjacent to what you seek if you already have a solution you can model, simplify, study." Also true, for the people who get to study it. Tristan Buckmaster describes the Euler write-up his own model produced as "AI slop". The hundred-page Navier–Stokes proof has been read by nobody outside OpenAI. The insight-extraction step antirez relies on is precisely the step that has not happened — and once the headline is taken, who funds it?
Three facts survive the personalities
OpenAI started because of a rumour; its own post says so. The team running the agents "collectively had no research-level expertise in fluid dynamics", by Sébastien Bubeck's account, so the human contribution was the decision to spend. And the model may have been improved by "de-identified data derived from their usage of our products", which the company "cannot rule out".
A rumour, a budget, and a training set that might contain the victims' drafts. That machine is already built and switched on.
Simon Willison drew the parallel to security, where a rumour of a bug is now enough to find the exploit. The difference is that security already assumed adversaries. Mathematics ran on the assumption that you could tell a colleague at another institution what you were working on. Buckmaster's email to OpenAI was written under that assumption. It went out on a Thursday. The prompts went out that weekend.
The fix will not come from norms
Tao proposes designating classes of problems "analysis-required", so a bare answer without extracted insight counts for little. As a norm inside mathematics, fine. As a constraint on a lab with a model to launch, it is nothing. OpenAI did not need mathematicians to accept the result. It needed a press cycle, and Noam Brown's "Lee Sedol moment" post tells you what the result was for.
The lever that exists is contractual. Every unpublished idea typed into a coding agent is, under today's product terms, potential training data unless you have opted out and trust the opt-out. Buckmaster paid OpenAI out of his research grant to help beat himself; his statement mentions "footing a large bill to OpenAI". Universities and funders can refuse to pay for tools on those terms, and some will, now that there is a case to point at. We would expect at least one major lab to offer a research tier with a contractual no-training, no-derived-data guarantee before the end of 2026, and the first demands to come from mathematics departments, because they now have the example.
So here is the call. By the middle of 2027, "we heard a rumour and pointed the model at it" is something labs deny doing rather than write into announcements. Not because they stop, but because this showed what it costs to admit it. Proving us wrong is simple: the next result of this size openly credits a rumour as the trigger. If that happens, Tao's ocean is already drained, and we were the optimists.
