OpenAI creates an external group to advise on its mathematical advances

What OpenAI announced—and what the claim of more than 100 problems means

OpenAI announced an external advisory group on mathematics and artificial intelligence after claiming that one of its internal models had solved more than 100 open math problems. That figure is a statement by the company. On its own, it does not establish which problems were tackled, what criteria were used to count them as solved, or whether the solutions have been independently evaluated.

The distinction matters because “solving a problem” can describe results with different scopes. To assess a mathematical claim, readers need, at a minimum, the precise problem statement, a proof that other specialists can examine, and enough information to check that the argument answers the question as posed. If a result concerns a particular version of a problem, that formulation and its conditions also need to be made explicit.

The available sources support the fact that OpenAI made the announcement and that the group was created. They do not provide proofs for the more than 100 results or an independent review of them. The figure should therefore be described as a company claim, not as a set of solutions whose validity has already been publicly established.

Who is in the group, and what is its remit?

The group describes itself as independent of OpenAI and is hosted at the Institute for Advanced Study (IAS) in Princeton. According to OpenAI’s announcement and the group’s and IAS’s institutional pages, it has nine founding members. The sources provided confirm its initial size and institutional affiliation, but the information included here does not list all nine names, so they are not specified in this article.

The stated remit is to advise on the relationship between artificial intelligence companies and mathematical research, with particular attention to how results are coordinated, presented, and communicated. The group may also offer unsolicited advice and make its comments public. This gives it room to express its own views, but does not make it a governing body within OpenAI.

One limit is explicit: the group will not advise OpenAI on the pace of its internal mathematical research. The IAS has also said that the panel has no decision-making power at AI companies; each company remains responsible for its own decisions. Independence and decision-making authority are therefore different things. The group can make recommendations, but the sources do not give it the power to require changes or halt a project.

Advising on communication is not the same as certifying a proof

The central question is whether the group will review mathematical proofs or simply help determine how results should be communicated. The available descriptions refer to advice on dissemination, presentation, and engagement with the mathematical community. They do not confirm a formal mandate to audit every proof, certify its correctness, or issue scientific approval before publication.

That does not prevent members from commenting on specific results if they consider it appropriate. The group says it can offer unsolicited advice and make that advice public. But that possibility is not the same as a systematic review process, with defined criteria, timelines, and conclusions for each result. No details provided here establish that such a process exists.

Nor should the group’s location at an academic institution be confused with an institutional evaluation of the solutions. Its affiliation with IAS describes where the project is hosted; by itself, it does not show that the institute has checked the proofs or endorses every claim made by OpenAI. Reporting should keep the panel’s independence, its advisory remit, and the mathematical verification of individual results distinct.

What public evidence is available about the solutions?

The available information does not include the technical detail needed to assess the more than 100 problems mentioned by OpenAI. An Infobae report says those results had not yet been published with the technical detail associated with other advances by the company. That is a description of what was publicly available at the time of the report, not proof that the solutions are either correct or incorrect.

Separately, OpenAI has published material about a solution it attributes to the Navier–Stokes existence and smoothness problem. Xataka reports that the company presented a mathematical paper and a formalization in Lean, a proof assistant that can check arguments expressed in a formal system. Public technical material makes that case more open to examination than an undocumented claim, but it is not the same as a completed independent evaluation.

The scope of this example also needs care. According to Xataka’s analysis, OpenAI addresses formulations of the problem that allow a designed external force, and does not necessarily answer the narrower question of whether the equations generate a singularity without that force. The point here is not to decide whether the proof is valid, but to specify which version of the problem is claimed to have been solved. A proof may address one formulation without answering a closely related question.

For that reason, the Navier–Stokes example cannot automatically be used to draw conclusions about the more than 100 other problems. Each result needs its own statement, proof, and examination. The sources provided do not identify all the problems or offer a complete set of proofs that could be compared.

What is still missing to assess the announcement’s full scope?

To assess the claim about more than 100 problems, OpenAI would need to publish an identifiable list of the results, the statements used, and complete proofs. It would also be important to know what review has taken place so far, who conducted it, and whether any objections or limitations have been acknowledged. Without those elements, it is not possible to independently verify the scope of the figure or compare the individual cases.

An advisory group can help improve the clarity and responsibility of public communication. It can also identify risks for the mathematical community, suggest publication practices, or challenge how a result is interpreted. But as long as its remit is described as advisory and no specific review procedure is made public, it should not be presented as a seal of approval for OpenAI’s solutions.

The news therefore concerns two related but distinct matters: the claim that a model solved more than 100 problems, and the creation of an external panel to advise on the relationship between AI and mathematics. The second could improve the discussion about how advances are disclosed; it does not, by itself, confirm the first. Scientific evaluation depends on concrete proofs that other specialists can examine.

Questions remain: which problems are included in the figure, what “solved” means in each case, what documentation the company will publish, and whether the advisory group will assess any particular proof. Until those questions are answered and examinable material is available, the most rigorous approach is to attribute the result to OpenAI and describe the panel according to the remit documented in the sources.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.