AI Breaks Into Mathematics: OpenAI's Problem-Solving Sparks a Schism in Academia

When OpenAI's internal models successively cracked open problems that mathematicians had agonized over for decades within a matter of weeks, the entire mathematics community was thrust into turmoil over identity, value, and the future. The Vergecast host David Pierce and reporter Robert Hart sat down for an in-depth conversation that reconstructs this unfolding disciplinary rupture.

OpenAI's Mathematical Raid: From "Impossible" to "Overnight"

Mathematics has long been regarded as the symbol of the "intellectual pinnacle" for AI labs—being able to solve math problems means touching some upper limit of intelligence. According to Robert Hart's reporting, over the past few months, OpenAI's internal models suddenly leaped from "barely usable" to "on par with top human mathematicians," so fast that even the most enthusiastic early adopters were caught off guard.

The impact of this turning point lies in the magnitude of the problems: from obscure small problems to challenges that senior mathematicians spent entire careers failing to crack, the models produced answers within days. After solving one of the Millennium Prize Problems, OpenAI publicly stated that "there are over 100 more in the queue"—a sentence that became a boulder pressing on the hearts of the mathematics community.

The Millennium Prize Problems were established by the Clay Mathematics Institute in 2000—seven problems, each with a $1 million prize, intended to record the greatest challenges that the mathematics community had failed to solve at the end of the 20th century. The Poincaré conjecture was solved by Perelman around 2003, but the remaining six—including the Riemann Hypothesis, P vs NP, and the Navier-Stokes equations—remain unresolved to this day. What these problems share is that their statements are extremely concise, yet the tools required may span multiple branches of mathematics, and proofs are often hundreds of pages long and require years of peer verification. So when the article claims the model "solved" one of them, the shock in the mathematics community came not only from the answer itself but from doubts about what "solved" even means—whether an answer counts without a verifiable proof is itself an open question.

The Mathematics Community's Complex Emotions: From Insomnia to Rushing Publications

Among the mathematicians Hart spoke with, some told him directly, "I can't sleep," while others questioned, "Is there any point in continuing?" The spectrum of reactions extended from despair to a nearly defensive sprint—rushing to publish as much of their own work as possible before being "scooped" by OpenAI.

This anxiety is not evenly distributed. The field of mathematics is enormously diverse internally: some still cling to blackboards and chalk, while others have long been proficient with AI tools. But even heavy users generally did not anticipate that the timeline would compress so dramatically. One mathematician described it as the "Eye of Sauron" suddenly turning toward their discipline.

The Core of the Controversy: Process Matters More Than Answers

The mathematicians' dissatisfaction is not fundamentally about the fact that "AI got stronger," but about OpenAI's misunderstanding of the nature of mathematics. Mathematics is not just a checklist of problems to be solved—the process of proof itself generates new tools, new methods, and new problems; it is the engine of the discipline's self-perpetuation. Simply providing answers is like removing the ladder to the next problem.

According to Hart's reporting, OpenAI recently released nearly 400 results scattered across more than 700 preprints, none of which had undergone peer review. Some mathematicians estimate that even a preliminary review would take years. More awkwardly, three of the papers were retracted due to "sign errors"—the kind of mistake that would be laughed off in school—and their appearance in a public release of this scale was interpreted by multiple interviewees as "a lack of care."

A preprint refers to a version of a research paper that authors voluntarily upload to public servers (such as arXiv or SSRN) before formal peer review. Its original purpose is to accelerate the dissemination of results, establish priority, and allow the community to provide early feedback. A preprint itself is not equivalent to "published," and unreviewed content is typically regarded in academia as preliminary work awaiting verification. However, when nearly 400 results are released at once in the form of more than 700 preprints, the problem is dramatically amplified: the community simply does not have enough manpower to complete screening and verification within a reasonable timeframe. More critically, the three papers retracted due to "sign errors" exposed a structural risk—large-scale automated output can easily overwhelm a verification system centered on manual diligence, while the cost of retraction is borne by the entire community.

Resource Asymmetry and the "Head Start" Controversy

The detail that most stunned the mathematics community came from OpenAI's own admission: after hearing that a certain research team was close to a breakthrough on a Millennium Prize Problem, they invested reportedly tens of millions of dollars, launched approximately 10,000 agents, and completed the solution in about 88 hours. While this "all-out crushing because someone else was about to win" approach is not entirely unprecedented in academia, the scale of resources is incomparable.

The more fundamental issue is fairness. OpenAI repeatedly stated that it wants to "enable scientists to achieve great things," but it uses internal models that outsiders cannot access or verify. On the other side, academic suspicions about whether "conversation data was used for training" remain unresolved—OpenAI's initial response was "we can't deny it," and although it later denied it, the wording was extremely precise, leaving room for interpretation.

In this context, an "agent" refers to a large language model system capable of autonomous planning, tool use, and multi-step task execution—not a single-turn Q&A chatbot. The article's mention of "approximately 10,000 agents over 88 hours" means OpenAI deployed large-scale computing power in parallel, allowing each agent to independently or collaboratively explore proof paths, with the system then aggregating candidate results. This kind of "brute-force parallel search" is not an entirely new idea in mathematics (distributed computing has long been used for enumeration and verification), but its scale and cost—reportedly tens of millions of dollars—are beyond the reach of any individual mathematician or academic team. It pushes mathematical discovery from a model of "individual inspiration plus long-term accumulation" toward a resource-intensive industrialized model, which is the essence of the fairness controversy.

Governance Attempts: Advisory Committees and "Hijacked" Independence

OpenAI has not been entirely inactive. It approached about nine top mathematicians in an attempt to form an advisory group to discuss how to release results responsibly. But according to Hart, some of those approached preferred to remain independent and formed their own group—after which OpenAI, with its larger megaphone, "hijacked" the announcement, leading outsiders to mistakenly believe the group was affiliated with OpenAI, which actually undermined its credibility.

In addition, OpenAI and Anthropic each contributed $5 million to support a Simons Foundation initiative aimed at promoting the mathematics community's understanding of AI results. But given the scale of investment and the impact caused, multiple interviewees expressed skepticism about the sincerity of this money. The most critical demand—stopping the use of frontier models to "sweep" the entire field of mathematics—was explicitly rejected by OpenAI on the grounds that "to develop tools that serve scientists, we must test them extensively."

Spillover Effects: Who's Next

Mathematics is only the first discipline targeted by the "Eye of Sauron." Hart said that researchers in economics, computer science, engineering, and other fields have begun to express similar concerns. Anthropic's practice of having Claude produce research results in a lab environment, as well as the controversy surrounding how its findings were publicized, both foreshadow the same pattern repeating across more disciplines.

What makes mathematics special is its purely theoretical nature—there is no laboratory verification step, making it even harder to adjudicate whether an answer is right or wrong. This both makes AI's "problem-solving" seem more impactful and shifts a large share of the verification cost onto the human community.

A "lab environment" in AI research typically refers to a model running under controlled conditions, equipped with additional tools, retrieval capabilities, or human oversight, differing from the public-facing product version. Anthropic having Claude produce research results in a lab environment means its performance may exceed that of the version available to ordinary users, but outsiders can hardly reproduce or verify the conditions under which these results were generated. This is the same type of problem as the "inaccessible, unverifiable internal models" mentioned later in the article: when AI companies keep model details closed in the name of "safety" or "competition" while publicly claiming scientific breakthroughs, academia faces an asymmetric information structure—the claimant holds all the evidence, while the verifier can only rely on trust. This structure itself conflicts with the norms of the scientific community.

Editorial View: Speed Is Killing Trust

The essence of this conflict is a head-on collision between two cultural rhythms: on one side, the Silicon Valley logic of "move fast, break things, don't look back"; on the other, an academic tradition that emphasizes diligence, accumulation, and peer verification. OpenAI may indeed be driving genuine scientific progress, but its release methods—bulk dumping, absent verification, ambiguous attribution—are systematically eroding the mathematics community's trust capital.

What deserves vigilance is the generational problem: if the research topics of a cohort of postdocs and PhD students are "solved" overnight, how will the discipline's talent pipeline be sustained? When AI needs mathematicians to verify, interpret, and disseminate its results, is it also undermining the very group's capacity for reproduction?

The optimistic version still exists: AI becomes a powerful tool for mathematicians, accelerating conjecture verification and cross-domain connections, while humans focus on asking better questions. But to reach that version, OpenAI needs to prove that it is not just "trying to win," but that it genuinely cares about the field it is charging into.

One-line Summary

OpenAI models rapidly solve mathematical problems, shaking the mathematics community's sense of value, fairness, and academic trust.

Paragraph Summary

Through a conversation between The Vergecast host and reporter Robert Hart, the article reconstructs the turmoil and division the mathematics community experienced after OpenAI's internal models successively solved open mathematical problems within weeks. The mathematicians' anxiety stems not only from the speed at which AI has improved but also from OpenAI's release methods: nearly 400 results scattered across more than 700 unreviewed preprints, with some papers already retracted due to sign errors. The controversy also involves resource asymmetry—OpenAI admitted investing heavily to race ahead on a Millennium Prize Problem after hearing that others were close to a breakthrough—and the "hijacking" of an advisory committee's independence by PR. The article points out that the core conflict is a head-on collision between Silicon Valley's "move fast" logic and academia's tradition of diligence, and warns that economics, computer science, and other fields may be next.