At a moment when the boundary between tool and thinker grows harder to locate, OpenAI has placed 372 mathematical results into the public record — each one produced not by a human mind laboring over years, but by an unreleased model consuming roughly three hours of compute per proof. The release, spanning algebra, topology, and theoretical computer science, arrives with CEO Sam Altman's proclamation of a new era of discovery, even as unresolved questions about attribution, training data, and peer review remind us that discovery has never been a simple act — only a human one, until perhaps now.
OpenAI's unreleased model solves 372 math problems; Altman hails 'new era of discovery'
The beginning, not the completion, of human understanding
Mark
So OpenAI is saying an AI model solved 372 math problems. What does "solved" actually mean here?
Mimi
The company ran roughly 4,000 problems through the model, and 372 of them produced results that could be verified using Lean, which is a formal proof-checking system. Some of those results are proofs of things mathematicians have been working on for years.
Luke
But here's the thing — we don't know if all 372 have been peer-reviewed by human mathematicians yet. OpenAI's own advisory group said this is "the beginning" of the process, not the end. So "solved" might mean "produced a result that passed an automated checker," not "confirmed by the mathematical community."
Mark
Why does that distinction matter?
Mimi
Because if a proof is wrong, an automated checker might miss it. And because mathematicians need to understand *why* something works, not just that it works. The advisory group is essentially saying: we have results, but we don't yet know if they're actually discoveries.
Luke
There's also the attribution problem. Tristan Buckmaster at NYU said OpenAI may have trained its model on his work before using it to solve the Navier-Stokes problem. If that's true, is it really the AI discovering something, or is it regurgitating human work it was trained on?
Mark
Did OpenAI respond to that?
Mimi
Not directly, not that we can see. They set up an advisory group after the Navier-Stokes controversy, but they haven't addressed the core question of whether the model was trained on the very papers it's now claiming to solve.
Luke
And the timing is interesting — Meta just announced their AI helped solve six problems. This feels like a race to claim mathematical discovery before anyone figures out what the rules should be.
Mark
So what happens next?
Mimi
OpenAI is funding workshops and conferences to help mathematicians understand these results. But the real test is whether the mathematical community actually accepts these as discoveries, or whether they remain curiosities produced by a black box.
El Pulso
- OpenAI dropped 372 AI-generated mathematical proofs onto GitHub without releasing the model that produced them, creating an unusual gap between the results and the means of replication.
- The sheer volume — drawn from a pool of 4,000 problems, each solved in about three hours of compute — is designed to signal not a single breakthrough but a systemic shift in how mathematical knowledge might be generated.
- A shadow of prior controversy follows the release: a prominent NYU professor has alleged OpenAI may have trained on his own work before claiming to solve the Navier-Stokes problem, raising unresolved questions about what 'discovery' means when the training data may already contain the answer.
- OpenAI's own Advisory Group at the Institute for Advanced Study cautioned that this release marks the beginning, not the completion, of mathematical understanding — and not all 372 proofs have yet faced peer review.
- With Meta simultaneously claiming its AI helped solve six major problems, the race for mathematical prestige is accelerating faster than the community's ability to verify, absorb, or trust what is being claimed.
At a moment when the boundary between tool and thinker grows harder to locate, OpenAI has placed 372 mathematical results into the public record — each one produced not by a human mind laboring over years, but by an unreleased model consuming roughly three hours of compute per proof. The release, spanning algebra, topology, and theoretical computer science, arrives with CEO Sam Altman's proclamation of a new era of discovery, even as unresolved questions about attribution, training data, and peer review remind us that discovery has never been a simple act — only a human one, until perhaps now.
OpenAI has published 372 mathematical results to GitHub, each produced by an internal AI model the company has not yet made public. The problems span algebra, number theory, topology, theoretical computer science, and mathematical logic. CEO Sam Altman framed the release as the dawn of a new era — a moment when AI begins generating genuine mathematical breakthroughs at scale, rather than merely assisting the humans who do.
The release builds on OpenAI's earlier claim that one of its models solved the Navier-Stokes Millennium Challenge. Where that announcement was singular and dramatic, this one argues through volume: 372 results, drawn from roughly 4,000 problems fed to the model, each solution consuming about three hours of compute time. Among the claimed achievements are a proof of the full Birch-Swinnerton-Dyer leading term formula for elliptic curves under specific conditions, and a result for the Mezard-Parisi formula in spin glass theory. Many proofs were verified using Lean, a formal proof-checking system with broad credibility in mathematics.
But the release carries unresolved weight. NYU professor Tristan Buckmaster has alleged that OpenAI may have trained its model on his own work before using it to solve Navier-Stokes — a charge that cuts to the heart of what discovery means when the intellectual scaffolding may already be embedded in the training data. OpenAI has not fully answered this concern.
In response to earlier criticism, the company established an Advisory Group on Mathematics and AI at the Institute for Advanced Study. That group issued a careful statement: the public release represents the beginning, not the completion, of incorporating this work into mathematical knowledge. Not all proofs have been peer reviewed. OpenAI says it plans to fund workshops and conferences to help the mathematical community engage with these results — though whether that investment will settle the deeper questions about attribution and the nature of machine discovery remains an open problem of a different kind.
OpenAI has posted 372 mathematical results to GitHub, each one produced by an internal AI model that the company has not yet released to the public. The announcement arrived as a data dump on the platform, spanning problems in algebra, number theory, theoretical computer science, mathematical logic, and topology. CEO Sam Altman promoted the release on social media, framing it as the opening of "a new era of discovery" — a moment when artificial intelligence would begin generating genuine mathematical breakthroughs at scale.
The timing builds on momentum from the company's earlier claim that one of its models had solved the Navier-Stokes Millennium Challenge, one of mathematics' most famous unsolved problems. That announcement drew attention. This one, with its sheer volume of results, is meant to signal something larger: that AI is now capable of contributing to human mathematical knowledge in ways that were, until recently, the exclusive domain of human mathematicians working over years or decades.
The 372 problems were drawn from a pool of roughly 4,000 that OpenAI fed to the model. On average, each successful solution consumed about three hours of compute time using ChatGPT Pro's thinking capability. The company says the vast majority of the results came from the same procedure and the same unreleased model. Among the claimed achievements are a proof of the full Birch-Swinnerton-Dyer leading term formula for elliptic curves under specific conditions, and a result for the Mezard-Parisi formula in the study of diluted spin glasses. Many of the proofs were verified using Lean, a formal proof-checking system widely trusted in mathematics.
Yet the release carries a shadow. OpenAI has faced criticism before over how it handles attribution and methodology. A New York University professor, Tristan Buckmaster, previously suggested that OpenAI may have trained its model on his own work before using it to solve the Navier-Stokes problem — a charge that raises questions about what counts as discovery when the training data itself may contain the intellectual scaffolding of the solution. The company has not fully addressed these concerns.
In response to the earlier controversy, OpenAI established an Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study to develop what it calls "best practices" for releasing such results. The advisory group issued a statement acknowledging that the public release represents "the beginning, not the completion, of the process of human understanding and the incorporation of the work into mathematical knowledge." In other words: these results are not yet vetted by the mathematical community. Not all have undergone peer review. The GitHub repository includes additional methodological details, OpenAI says, to promote transparency, though the company remains open to other formats for sharing results in the future.
The announcement also arrives in a competitive moment. Meta disclosed days earlier that its Muse Spark AI had helped mathematicians solve six major problems. The race to claim mathematical discovery — and the prestige that comes with it — is heating up. OpenAI is planning to fund workshops, conferences, and special programs designed to help the mathematical community understand and integrate these AI-generated results into existing knowledge. Whether that investment will resolve the underlying questions about attribution, methodology, and what it means for an AI to "solve" a problem remains to be seen.
Citas Notables
We are entering a new era of discovery now— Sam Altman, OpenAI CEO, on social media
The public release was the beginning, not the completion, of the process of human understanding and the incorporation of the work into mathematical knowledge— OpenAI's Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study
Quieres la nota completa?
Lee el original en India Today ↗