OpenAI Claims Breakthrough on 100 Open Math Problems While Launching Advisory Panel

A Sudden Acceleration in Mathematical Claims

OpenAI has announced that an internal artificial intelligence model has resolved more than 100 previously unsolved problems spanning most areas of mathematics. The disclosure arrived on 21 September alongside the formation of a new Advisory Group on Mathematics and Artificial Intelligence, to be hosted at the Institute for Advanced Study in Princeton, New Jersey.

The timing is striking. The announcement follows weeks of tension within the mathematics community after OpenAI published a solution to the Navier-Stokes Millennium Prize problem, one of seven challenges carrying a million-dollar reward from the Clay Mathematics Institute. That release, which OpenAI characterised as "abrupt" in its own statement, drew sharp criticism from researchers who felt the company had bypassed traditional peer review and academic scrutiny in favour of speed.

The new advisory group appears designed to address that friction, though its mandate is carefully circumscribed. Nine prominent mathematicians will evaluate the significance of results and coordinate their public release, but the group will have no authority over the direction or tempo of OpenAI's mathematical research. According to OpenAI, the group "will not be responsible for advising us on how to pace our internal progress on mathematics."

A Bridge with Limited Structural Support

The Institute for Advanced Study underscored the advisory group's constraints in its own statement. "Although we will give advice, we do not have decision making power at any AI company, and the responsibility for the decisions made by any company will rest with that company," the institute said.

Members of the panel will not receive compensation, though they retain the right to offer unsolicited advice, speak publicly about their views, and control their own membership. That structure grants a degree of formal independence, but it also places the group in a reactive posture: assessing what OpenAI has already produced rather than shaping what it pursues.

Of the nine initial members, only one, Camillo De Lellis of the Institute for Advanced Study, was among the 25 Fields Medal laureates who signed an open letter earlier in September. That letter argued that artificial intelligence laboratories are jeopardising intellectual work in mathematics as they compete to solve famous problems ahead of human researchers.

The signatories expressed concern not only about the pace of AI-generated results but also about the lack of transparency in how models arrive at solutions, the absence of traditional verification processes, and the risk that commercial incentives are distorting the priorities of mathematical inquiry.

The Scale of the Claim

OpenAI's assertion that a single model has resolved more than 100 open problems is extraordinary by any historical measure. Open problems in mathematics are questions that have resisted solution for years or decades, often because they require novel techniques or insights that existing frameworks cannot supply. The Riemann Hypothesis, the Birch and Swinnerton-Dyer Conjecture, and the Navier-Stokes existence and smoothness problem are canonical examples; each has occupied generations of mathematicians without yielding a complete proof.

At Opentechwire, we have tracked the deployment of large language models and reasoning systems in formal theorem proving, and the trajectory has been steep. Models trained on mathematical corpora can now generate candidate proofs, verify logical steps using proof assistants such as Lean or Coq, and explore combinatorial spaces that would take human researchers months to traverse. Yet the jump from incremental progress on formalisation tasks to the resolution of more than 100 open problems suggests either a profound leap in capability or a broader definition of what constitutes an "open problem" than the mathematical community typically employs.

OpenAI has not published a list of the problems it claims to have solved, nor has it detailed the verification process applied to the proofs. That opacity makes independent assessment difficult and feeds the unease among mathematicians who worry that commercial AI laboratories are prioritising announcements over rigour.

The Navier-Stokes Shadow

The Navier-Stokes problem concerns the existence and smoothness of solutions to equations that describe fluid flow. It is one of the most famous unsolved questions in applied mathematics, with implications for turbulence modelling, weather prediction, and aerodynamics. OpenAI's decision to release a solution without advance notice to the broader mathematical community was widely criticised as a departure from norms of collaborative review.

Mathematicians pointed out that extraordinary claims require extraordinary scrutiny. A proof of a Millennium Prize problem would normally undergo years of review by specialists before being accepted, and even then, the process often uncovers errors or gaps that require revision. The abruptness of OpenAI's announcement suggested that the company was more concerned with establishing priority than with ensuring correctness.

The new advisory group is positioned as a corrective, a mechanism to slow the release of results and subject them to expert evaluation before they reach the public. Yet the group's lack of authority over OpenAI's internal research means it cannot prevent the company from continuing to generate proofs at whatever pace its computational resources allow. The advisers will see the results after they are produced, not before.

Independence and Influence

The structure of the advisory group raises questions about how much influence it can exert. Members can speak publicly, which means they could in theory criticise OpenAI's approach or warn the mathematical community about premature claims. That freedom is meaningful, but it also places individual mathematicians in the position of acting as whistleblowers if they believe the company is overreaching.

The group's control over its own membership is another form of independence, allowing it to add or remove members without OpenAI's approval. That could enable the panel to maintain credibility even if its relationship with the company becomes strained. But without decision-making power, the group's primary function is to serve as a reputational buffer: a signal to the public and to mathematicians that OpenAI is engaging with expert opinion, even if it is not bound by that opinion.

This dynamic is not unique to OpenAI. Technology companies have long convened advisory boards to demonstrate accountability while retaining operational autonomy. The effectiveness of such arrangements depends on the willingness of advisers to push back publicly when their advice is ignored, and on the willingness of the company to accept reputational costs if the advisers resign or criticise its decisions.

The Competitive Pressure

The open letter signed by Fields Medallists earlier in September was explicit about the competitive dynamics driving AI laboratories. The signatories argued that companies are racing to solve famous problems in order to demonstrate the power of their models, and that this race is distorting the ecosystem of mathematical research. Problems are chosen not because they are scientifically important but because they are recognisable to a general audience and can generate media attention.

That incentive structure is visible in OpenAI's announcement. The Navier-Stokes problem is one of the most famous in mathematics, and its solution would be a headline event regardless of who achieved it. The claim of more than 100 additional solved problems amplifies that narrative, positioning OpenAI's model as a transformative tool for mathematical discovery.

Yet mathematicians have noted that the value of a proof is not solely in its correctness but in the understanding it provides. A proof that arrives as a black-box output from a neural network, even if verified by automated tools, may offer little insight into why a theorem is true or how its techniques could be applied elsewhere. The process of constructing a proof is often where the most valuable ideas emerge, and that process is opaque when the work is done by a model.

What Verification Means in This Context

Formal verification using proof assistants is a rigorous process, and when a proof is fully formalised in a system like Lean, it can be checked mechanically for logical errors. But formalisation is not the same as understanding, and it does not address the question of whether a proof is interesting or whether it advances the field.

Moreover, many open problems in mathematics are stated informally or involve concepts that are difficult to formalise without significant human judgement. A model that generates a candidate proof must still rely on human mathematicians to determine whether the proof addresses the intended question and whether its assumptions are reasonable. That step is where errors and misunderstandings often arise, and it is precisely the step that OpenAI's advisory group is meant to oversee after the fact.

The advisory group's role in assessing significance is therefore critical. A proof may be technically correct but trivial, or it may solve a problem that was already close to resolution. Without access to the proofs themselves and detailed context about how they were generated, the broader mathematical community cannot judge whether OpenAI's claims represent a genuine breakthrough or an artefact of how the company is counting solved problems.

The Institute's Position

The Institute for Advanced Study is one of the most prestigious centres for theoretical research in the world, and its decision to host the advisory group lends credibility to the arrangement. The institute has a long history of independence from commercial interests, and its faculty includes some of the most influential mathematicians alive.

Yet the institute's statement was careful to clarify that hosting the group does not imply endorsement of OpenAI's research or control over its activities. The advisory group is a separate entity, and the institute's role is to provide a neutral venue and administrative support. That distinction matters because it protects the institute's reputation if the advisory group proves ineffective or if OpenAI's mathematical claims do not hold up under scrutiny.

For OpenAI, the association with the institute is valuable. It signals that the company is engaging with elite academic institutions and that its work is being taken seriously by leading researchers. But the structure of the arrangement limits the institute's exposure if the relationship becomes contentious.

The Broader Pattern in AI Research

OpenAI's formation of an advisory group follows a pattern seen across the artificial intelligence industry. As models become more capable and their outputs more consequential, companies are creating oversight mechanisms that offer some degree of external input without ceding control over core decisions.

These mechanisms vary in their effectiveness. Some advisory boards have resigned en masse when their advice was ignored; others have functioned as intended, providing a channel for expert feedback that shaped company policy. The key variable is whether the company is willing to bear the cost of public criticism from its advisers, and whether the advisers are willing to use their platform to push back when necessary.

In the case of mathematics, the stakes are somewhat different from other domains where AI is deployed. A flawed medical diagnosis or a biased hiring algorithm can cause immediate harm to individuals. A flawed proof, by contrast, is primarily a reputational problem for the institution that produced it and a waste of time for the mathematicians who attempt to verify it. The harm is to the integrity of the field rather than to individual people, but it is harm nonetheless.

What Happens Next

The advisory group will begin its work with a backlog of more than 100 claimed proofs to assess. The pace at which those assessments proceed will be an early indicator of the group's capacity and independence. If the group moves quickly to validate OpenAI's claims, it may face scepticism from mathematicians who believe the review process is being rushed. If it moves slowly or raises concerns about the quality of the proofs, it will test OpenAI's willingness to tolerate public disagreement from its advisers.

The group's composition will also matter. With only one signatory of the Fields Medallists' open letter among the initial nine members, the panel may not reflect the full spectrum of concern within the mathematical community. Future additions to the group could shift its posture, particularly if members are chosen who have been vocal critics of AI laboratories' approach to mathematics.

For the broader AI industry, OpenAI's move is likely to be watched closely. If the advisory group succeeds in building trust with the mathematical community and slowing the release of unverified results, other companies may adopt similar structures. If it fails, or if mathematicians conclude that it is primarily a public relations exercise, the backlash could extend beyond OpenAI to the entire enterprise of using AI for mathematical research.

The fundamental tension remains unresolved. OpenAI has built a model that can generate proofs faster than human mathematicians can review them, and the company has commercial and reputational incentives to publicise those proofs as quickly as possible. The advisory group can slow the public release of results, but it cannot slow the internal generation of those results, and it has no power to redirect the research if it believes the company's priorities are misaligned with the interests of mathematics as a discipline.

That asymmetry means the group's influence will depend less on its formal structure and more on the credibility of its members and their willingness to speak candidly when OpenAI's claims do not meet the standards of the field. The coming months will show whether that credibility is enough to bridge the gap between the pace of AI research and the deliberative norms of academic mathematics.