OpenAI's Agents Produced a Proof for a Millennium Prize Problem in 88 Hours. Mathematicians Are Skeptical
Imagine an AI producing a huge module of code in a couple of days, and you cannot understand why it works. That is roughly where mathematics stands now. OpenAI set thousands of agents on the Navier-Stokes equation, one of the seven Millennium Prize Problems, and got a proof 88 hours later. Harvard professor Michael Hopkins called such papers "nonsense," while panelists at MIT argued that the ability to verify and explain a result is worth more. Teams working with agents should budget time for verification.
AI-processed from MIT News; edited by Hamidun News
Imagine an AI agent writing a huge module over the weekend. The tests pass, but nobody on the team can explain why it works. That is roughly the situation mathematics finds itself in now. At MIT, a panel called "AI and Society: The Future of Mathematics" brought together professors from MIT, Harvard, Northwestern, and New York University to discuss what to do when machines solve problems faster than humans but it is not always clear how. For developers and founders, this is not a distant story: the same questions about verifying and trusting AI results will soon confront every team that hands serious work to agents.
What happened: from olympiad to Millennium Prize
Not long ago, AI solved problems whose answers were already known. In the summer of 2025, Google DeepMind models earned gold at the International Mathematical Olympiad: they quickly handled most of the problems that are given to the world's best high school students. Later it became clear that the models sometimes found solutions slightly better than human ones. In 2026, according to the panelists, there was a jump: AI began tackling open research questions with no known answer. Bryna Kra of Northwestern said: "Mathematics has changed more in the last two months than in the last hundred years."
In September, OpenAI directed thousands of its agents at a single problem related to the Navier-Stokes equation. It describes the flow of fluids and gases and is one of the seven "Millennium Prize Problems." After 88 hours, the agents produced a proof.
The timing was unfortunate for Tristan Bakhmaster, one of the panelists: after a year of work, he and a colleague from Anthropic were close to a solution. A debate followed: whether credit had been distributed correctly, whether mathematicians' data had leaked, how the agents arrived at their conclusion, and whether it made sense at all. A month later, OpenAI published 722 manuscripts across different areas of mathematics, claiming that AI had fully solved many major open problems.
The term "matocalypse" stuck to this episode.
The proof exists, the understanding does not
The main objection is that obtaining a proof and understanding it are different things. Michael Hopkins of Harvard said that "the papers themselves are nonsense," full of "strange ideas," and added that there is still a difference between AI mathematics and human mathematics. Bakhmaster agreed: for mathematics to matter, it must rest on human understanding.
This resembles code review. An agent submits a patch of ten thousand lines; it compiles, but if nobody can explain the logic, you have not received a solution, you have received debt. Mathematicians reached this point before everyone else.
Kra proposes abandoning the race for priority, the prestige of being "first to prove" something. Speed should no longer be the goal, and understanding needs to be assessed differently. Hopkins countered that people will still want to be pioneers. Kra replied that one can also be a pioneer in explaining how a proof came about. She helped prepare the Leiden Declaration on AI and mathematics, which mathematicians from various countries signed. The document calls for a response to the risks AI poses to mathematical research.
Journals, hiring, and students are already changing
The consequences are visible in concrete processes. At a specialized journal where Kra works as managing editor, the number of papers with AI-assisted results has grown sharply, and the editors had to hire another managing editor. In response to the flood of AI "slop," some publications now place more value on exposition: the section of a paper that explains the logic of a hypothesis or proof. Experimental journals are emerging that publish specifically human contributions, such as new explanations of old results. One of them is the video journal Mathematical Discourse, which features recorded talks.
Hiring is changing too. According to Kra, interviews are becoming more important: it is not enough to show written ideas, you must demonstrate that you understand them and can explain them to another person. Jacopo Borga of MIT reminded the audience that about 90% of a mathematician's evaluation rests on journals, and proposed devising a new metric.
Students face the greatest difficulty: they have to learn the fundamentals amid a flood of AI shortcuts. In Hopkins's group there are "AI vegans" who do not use AI at all, and those who do. He singled out a student with good intuition but weak execution, who can use prompts to pull interesting ideas out of AI. Perhaps this is a "new type of mathematician."
What to take from this in practice
A caveat: these are the views of participants in a single panel, not an independent review of OpenAI's work, and the dispute over the correctness of the 722 manuscripts continues. But the conclusion applies beyond mathematics. When obtaining a result becomes cheap, verification, explanation, and accountability become expensive. If you delegate large tasks to agents, decide in advance who will read and check their output, and how. Budget time for this, and require from the agent not only an answer but also a clear account of its reasoning. Borga put it more optimistically: "The castle is collapsing, but space is opening up for research."
- Nov 21, 2026OpenAI Disbands Robotics Team: GPT-6 Astra Solved 74.8% of Tasks in Simulation vs. 51%
- Nov 21, 2026US Government Unions Win AI Protections as Nearly 70% of Agencies Already Use the Technology
- Nov 21, 2026OpenAI Lets Four Outside Auditors Into Its Offices to Test Its Models
- Nov 21, 2026AI Agents Can Now Trade Tokens on OneDex Without an API Key or Fees
Key Questions Answered
What happened: from olympiad to Millennium Prize?
Not long ago, AI solved problems whose answers were already known. In the summer of 2025, Google DeepMind models earned gold at the International Mathematical Olympiad: they quickly handled most of the problems that are given to the world's best high school students. Later it became clear that the models sometimes found solutions slightly better than human ones. In 2026, according to the panelists, there was a jump: AI began tackling open research questions with no known answer.
What are the key details on The proof exists, the understanding does not?
The main objection is that obtaining a proof and understanding it are different things. Michael Hopkins of Harvard said that "the papers themselves are nonsense," full of "strange ideas," and added that there is still a difference between AI mathematics and human mathematics. Bakhmaster agreed: for mathematics to matter, it must rest on human understanding. This resembles code review.
What are the key details on Journals, hiring, and students are already changing?
The consequences are visible in concrete processes. At a specialized journal where Kra works as managing editor, the number of papers with AI-assisted results has grown sharply, and the editors had to hire another managing editor. In response to the flood of AI "slop," some publications now place more value on exposition: the section of a paper that explains the logic of a hypothesis or proof. Experimental journals are emerging that publish specifically human contributions, such as new explanations of old results.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.