Jacob Tsimerman, who received the Fields Medal last month, said at the award ceremony that he is set to join OpenAI to work on AI safety research.
In early August, Tsimerman and OpenAI research scientist Sébastien Bubeck brought about 40 mathematicians to OpenAI’s San Francisco headquarters for a private summit. One presentation at the meeting was titled “The End of Mathematics.”
A closed-door summit framed the stakes
Daniel Litt, a professor at the University of Toronto, told attendees: “It is possible that we end up in a world with no high-quality mathematical research and the complete disappearance of human mathematical expertise.”
He also said that such an extreme outcome was unlikely, but argued that mathematicians need to act.
Drew Sutherland, a mathematician at the Massachusetts Institute of Technology, put it more bluntly: “Maybe we’re just the canaries in the coal mine, and mathematicians get hit first.”
Math was chosen as the testing ground because it is treated as a proxy for high-level intellectual activity. A core idea running through the summit was that whatever AI can do in mathematics is likely to be repeated in other fields sooner or later.
OpenAI and Anthropic have been releasing results in quick succession
That burst of results formed the immediate backdrop to the summit.
In May, an unreleased OpenAI model overturned the unit distance conjecture. The problem, which concerns arrangements of points on an infinite plane, had remained unresolved for years. According to the report, the AI solution used a known but difficult technique from algebraic number theory, effectively pulling a tool across subfields.
Tsimerman said at the time that if the result were submitted to any journal, he would “accept it without hesitation.” Many mathematicians viewed it as the first major AI breakthrough in mathematics.
Then in August, OpenAI announced 10 new AI-generated findings in mathematics and computer science, spanning several subfields that usually take years to even enter.
Anthropic was active as well. Levent Alpöge, an Anthropic employee, used Claude to construct orthogonal vector pairs on a 668-dimensional hypercube. Mathematicians had long believed such a construction existed, but no one had actually produced it. Alpöge revealed the result in a 24,000-character post on X made entirely of plus and minus signs.
Link: https://x.com/__alpoge__/status/2087504785952182273
AI results are starting to bypass the standard publication pipeline
Traditional mathematics follows a fairly stable process: workshop discussion, circulation among peers, posting to arXiv, and publication in a journal after peer review.
AI-generated work is starting to skip parts of that pipeline.
Dmitry Rybin, an entrepreneur based in Shenzhen, asked ChatGPT to “make a breakthrough” and overturn a network flow hypothesis. Each time the model failed, he pushed it to keep going. ChatGPT eventually produced a counterexample. After verifying it, Rybin posted the result on X while he was out watching a movie with friends.
Link: https://x.com/DmitryRybin1/status/2079904005652893709
Mathematician David Bessis questioned that style of release: “You’re boiling pasta and tweeting that you solved a major problem? Of course that goes viral. But what happens after that? Who is going to sort through this pile of stuff and figure out what’s actually going on?”
Last week, Anthropic also said one of its employees asked an internal version of Claude to “seriously try” the Riemann hypothesis. The hypothesis has been open for more than a century and carries a $1 million prize. Under repeated prompts such as “continue” and “think again,” the model produced a new result on a related problem, according to the report.
More than 3,000 signatures did not settle the debate
The response across mathematics has been split.
In June, an open letter called the Leiden Declaration was released and signed by more than 3,000 mathematicians. It called for researchers to use AI tools responsibly and to make sure results are verified and citations are handled properly.
Another group of mathematicians wants to go much farther and reject AI and AI companies entirely in order to preserve a human place in mathematics.
Bryna Kra, a mathematics professor at Northwestern University, attended the summit and is also one of the 16 initiators of the Leiden Declaration. She said: “The heart of mathematics is understanding a result, not just proving a result. A proof that is not understood does not become part of the literature.”
That pushes the debate to a harder question: AI may be able to prove something, but can it explain how it did so?
Explanation remains a weak point for top models
Harvard mathematician Melanie Matchett Wood helped write a human-readable version of the proof behind OpenAI’s unit distance conjecture result. In that process, she found a shared weakness in top AI models: they say a lot about the easy parts and pass quickly over the hard parts.
“Top AI models still can’t identify the genuinely difficult parts of an argument and explain them clearly,” she said.
German mathematician Andreas Thom drew a sharper distinction. One of the 10 new OpenAI results involved a new construction related to non-sofic groups, filling a gap between two papers he wrote with co-author Gábor Kun in 2016 and 2019.
The pair later published a follow-up paper that simplified and extended the AI finding. “AI is solving problems in a very clever and substantial way,” Thom said. “But new concepts emerged only after humans participated in the proof process. At this point, that is not something AI is doing on its own.”
Bubeck outlined four futures, but the meeting ended without a consensus
After the summit, Bubeck described four possible paths ahead.
- Mathematics could start to resemble software engineering, with AI helping hundreds of people work together on the same problem.
- It could look more like physics, built around scaling compute and AI models in place of particle accelerators.
- It could become closer to museum curation, with AI generating output and humans selecting and interpreting it.
- Or mathematicians as a group could shift into AI safety.
Bubeck said: “We must put people, and mathematicians, first. Mathematics has meaning only when mathematicians learn something from it. A civilization without humans has no meaning.”
Still, the 40 attendees did not reach a common view.
Tsimerman said: “There is no settled answer yet on what the mathematics community wants. The purpose of this meeting was more to open a conversation than to make a decision.”
A Fields Medalist moving into AI safety has become part of the story
Tsimerman has not formally started at OpenAI yet.
Still, the fact that a Fields Medal winner has chosen to move into AI safety work at an AI company now stands as a snapshot of where mathematics finds itself. The San Francisco summit did not produce a single answer on how AI will reshape mathematical research or what role mathematicians will hold inside that process.
What it did make clear is that the argument has started in earnest.
The reference cited in the source material is a Washington Post article published on Aug. 19, 2026, titled “mathematicians ask what’s left for humans when AI can do math research.” The MarsBit post also says the piece came from the WeChat account Xinzhiyuan, written by ASI启示录 and edited by 马可.