Princeton mathematics professor Sergiu Klainerman published an essay titled “AI will not make mathematicians obsolete” in the journal Inference (Vol. 9, No. 2) on September 21, 2026 — a direct response to OpenAI’s announcement on September 8, 2026, in which an internal system of approximately 10,000 agents produced a solution to the Navier–Stokes problem in 88 hours, complete with an analytical note and formalization in Lean. According to Klainerman’s assessment, the machine did not solve the original formulation, but only a weakened version with an artificial external force, while the actual problem remains open. This case is important beyond pure mathematics: a vendor’s loud claim was narrowed down in the first public expert review because machine formal verification does not check whether the right problem was formalized.

What happened

On September 8, 2026, OpenAI announced that its internal system of approximately 10,000 agents prepared a solution to the Navier–Stokes problem in 88 hours: an analytical note and a formalization of the proof in Lean, published in the open repository openai/NavierStokesAndEuler and available for independent machine checking. On September 21, 2026, the essay “AI will not make mathematicians obsolete” by Princeton mathematics professor Sergiu Klainerman was published in the journal Inference (Vol. 9, No. 2), written in response to this announcement. Klainerman argues that only the weakened forced version of the problem was solved, in which a singularity — infinite flow velocity at a finite time — arises under an external artificially selected force from a state of rest. The true unforced Navier–Stokes problem, according to his assessment, remains open.

Context

The Navier–Stokes problem is one of the seven “Millennium Prize Problems” of the Clay Mathematics Institute, and this is the first case in which a machine has claimed a solution to a “Millennium Prize Problem” — and immediately received a public professional review from a leading mathematician. The dispute is not about the logic of the constructed proof, but about which formulation of the problem to count: the variant with an external force goes back to the formulation permitted by Charlie Fefferman, and it is precisely this that Klainerman now calls a mistake. Formalization in Lean turns the proof into a machine-checkable object, however, the formal verifier confirms the correctness of the inference only within the formalized formulation and does not answer the question of whether the right problem was chosen. Klainerman’s general conclusion about recent AI mathematical achievements fits the metaphor of exhaustive search: the machine manages to “find the needle when someone else built the haystack and pointed to it.”

Why this matters for the industry

For the industry, the main practical outcome is not in the headline, but in the run architecture: the combination of a swarm of LLM agents with formal checking in Lean has proven itself as a “generation — formal verification” pipeline and has de facto become infrastructure for searching for examples and counterexamples. In this case, the verifier did its job — the specification failed: a weakened version was formalized and solved, not the original problem. Hence the procedural consequence: in agent proof pipelines, a separate stage of formulation audit is needed before launching a mass search, and a formal verifier should be set as a mandatory quality gate, remembering that “passed verification” is not the same as “the right problem was solved.” Any vendor claim that “AI solved problem X” must now go through the question “in which formulation?”, and teams selling agent research should reasonably build an independent audit of the formulation directly into the funnel and pricing. There is nothing to deploy in production yet: the system is internal and works without a public API, prices, or metrics, so the practical benefit for teams is the open repository openai/NavierStokesAndEuler — it can be used to analyze the template and transfer it to their own domains where there is a machine-checkable criterion.

Why this matters for users

It is useful for the reader to know that the loud “AI solved a Millennium Prize Problem” is structured more subtly than it sounds: the solution was obtained for a weakened version with an artificial external force, and the true unforced Navier–Stokes equations have not been touched. Klainerman’s essay and OpenAI’s publication with Lean formalization are open, so everyone can compare both positions themselves and form their own opinion — even “solved” AI problems require expert review. The essay also answers the frequent fear of young mathematicians “will I be replaced?” — Klainerman’s answer is reassuring: according to his assessment, AI will strengthen mathematicians, not replace them.

What is still unknown / limitations

From open data, it does not follow that OpenAI intended to fit the result to a weak criterion: formally, the system solved exactly the problem that was formalized, and the allowance of the formulation with an external force is a dispute within mathematics itself, since the corresponding formulation was once permitted by Charlie Fefferman, and only in hindsight does Klainerman call this a mistake. The position of the mathematical community and the Clay Mathematics Institute on the issue of the forced and unforced formulation has not yet been expressed, and independent reviews of the Lean formalization are only expected. An open empirical question remains whether the “agent swarm plus formal verifier” pipeline can work where the haystack was not built by a human in advance, and OpenAI’s system is internal and has no public API or metrics, so conclusions about its capabilities are based on the company’s own publications.

Sources

Author

Look at AI, editorial team