Mathematics has always moved slowly. Proofs take years. Careers are built around single conjectures. The discipline prizes rigor above all else, and that rigor has a cost: progress is measured in decades, not quarters. That rhythm is now being disrupted. OpenAI recently announced that an internal model called Astra had produced solutions to 10 long-standing mathematical problems, some of which had resisted resolution for decades. For the mathematicians who have spent their careers on problems like these, the announcement landed with the weight of something that cannot be undone.
Ten Problems, Decades of Failure, One Model
The problems Astra solved were not selected for their simplicity. They spanned a wide range of mathematical fields, from abstract structures to questions with direct practical consequences. One result addressed how tightly spheres can be packed in dimensions beyond three, a question connected to how efficiently data can be encoded and transmitted. Another advanced the theory of error-correcting codes, which help recover information from noisy signals. A third resolved questions about how complex connected networks can become before structural patterns emerge. Other results touched on quantum game theory and the geometry of high-dimensional grids, with implications for post-quantum cybersecurity.
Yang-Hui He, a fellow at the London Institute for Mathematical Sciences, put the significance plainly: solving any one of these ten problems, he said, would be enough to earn an academic job. He had just returned from a research conference in South Korea where many attendees described a sense of “phase transition” over the previous six months, with AI producing advances that researchers considered genuine and meaningful. James Maynard, a professor at the University of Oxford and a Fields Medal winner, noted that until recently, AI breakthroughs in mathematics tended to involve problems that had attracted little serious attention from researchers. That is no longer the case.
Credit, Correction, and the Messiness of Collaboration
The most scrutinized result in the collection concerned non-sofic groups: infinite mathematical structures that, roughly speaking, cannot be approximated by finite ones. Whether such structures existed at all had been an open question for decades. OpenAI’s original announcement described the collection as addressing problems that had seen “no progress on the main result for at least a decade.” That framing drew immediate pushback.
Francesco Fournier-Facio, a mathematician at the University of Cambridge, said he and colleagues believed the announcement minimized the contributions of researchers Andreas Thom and Gábor Kun, whose recent work had helped lay the groundwork for the non-sofic groups result. Kun, a researcher at the Alfréd Rényi Institute of Mathematics in Hungary, told The Verge that OpenAI had contacted him by email shortly before publishing. He described the sweeping language in the original announcement as “rather comical,” noting that the attached research paper “clearly said that it builds on my results from 2016 and 2019.” He called the oversight “rather sloppy.”
OpenAI subsequently updated the announcement to say it was sharing “results, each of which resolves or makes substantial progress on a long-standing open problem,” and confirmed the change was made to better reflect the prior research the results build upon. No correction note was added to the page. Kun said he wondered whether similar attribution gaps might exist in the other results, which fell outside his areas of expertise. The episode illustrates something that tends to get lost in announcements of this kind: AI systems do not generate results from nothing. They build on the accumulated work of researchers, and how that debt is acknowledged matters.
What It Means When a Discipline Loses Its Monopoly on Discovery
This is what most coverage of AI and mathematics misses. The question is not simply whether AI can solve hard problems. It clearly can. The deeper question is what happens to a field when the core activity that defines it, the slow, human pursuit of proof, is no longer exclusively human.
Maynard described spending significant time over the past year in “soul searching” about the future of mathematics as a discipline. He is not alone. The mathematicians who spoke to The Verge expressed a mixture of excitement and apprehension, with some describing something closer to despair about what accelerating AI capability means for people who have dedicated their lives to the field, and for future generations who might have done the same.
The concern is not simply about jobs. It is about meaning, identity, and the structure of a discipline that has always rewarded patience and depth. When AI can compress decades of failed attempts into a single model run, the human relationship to mathematical discovery changes in ways that are not yet fully understood. Few researchers doubt that a profound upheaval is already underway. Most are still working out what to think about it.
In Short
OpenAI’s Astra model solved 10 long-standing mathematical problems spanning fields from sphere packing to quantum game theory, problems serious enough that solving any one of them would, according to researchers, support an academic career. The announcement also surfaced a real tension around credit and attribution, with at least one researcher whose work the results built upon describing the original framing as sloppy. Broader than any single result, the episode signals that AI is now capable of genuine mathematical discovery, and that the field is only beginning to reckon with what that means.
Based on reporting from The Verge.