OpenAI CEO Sam Altman says he feels for the mathematicians who are buried under his company’s AI-generated results. He also says OpenAI isn’t going to slow down. The Verge AI reports that on an episode of the Cheeky Pint podcast released Thursday, Altman said OpenAI has more unreleased discoveries. When asked whether there’s more beyond the hundreds of proofs already public, he answered: “We do. Of course.”
This is one of the first times Altman has spoken directly about what OpenAI’s math push is doing to the research community. His position is simple. “Humanity deserves fundamental knowledge, and we should not be sitting on a major fundamental science discovery,” he said.
📌 The Quick Version
- What happened: Altman said he has empathy for researchers struggling to keep up, but he believes the public’s right to new knowledge “outweighs” the disruption.
- What’s next: OpenAI has a stockpile of unreleased math results, and it plans to keep releasing them.
- What it cost: Altman estimated OpenAI has spent hundreds of millions of dollars on math problems, including about $10 million (“maybe a little more”) on its Navier-Stokes work.
- The admission: He said the math push did “get out of control” for a while, but called that deliberate and necessary for testing what the models can do.
- The contradiction: OpenAI is holding back details of a claimed Yang-Mills result, which is exactly the kind of sitting on findings Altman says it shouldn’t do.
🔍 Why Mathematicians Are Frustrated
The results themselves aren’t the problem. Many researchers call them genuinely impressive. The problem is how they’re delivered.
Mathematicians who spoke to The Verge listed the same complaints again and again:
- Weak attribution: It’s often unclear who or what deserves credit for a result.
- Little context: Proofs show up with short writeups that researchers describe as poorly written and hard to follow.
- Not enough verification: The slow job of checking whether a proof actually holds lands on outside researchers.
This month alone, OpenAI released hundreds of mathematical results. Peer review usually moves at the pace of months per paper. When a lab can produce proofs faster than the field can check them, verification becomes the bottleneck, and academics end up doing that work for free.
⚖️ The Yang-Mills Contradiction
This is the part that stands out. Altman says OpenAI shouldn’t sit on discoveries, but researchers this week accused the company of gatekeeping. OpenAI recently announced three major breakthroughs in mathematics, physics, and computer science, and it hasn’t published the details of one of them. That one involves Yang-Mills theory, the mathematical foundation of modern particle physics.
OpenAI says it’s following advice from a hand-picked group of experts on how best to release the result. But that advice goes directly against what an earlier group of mathematicians told the company. OpenAI calls this a new approach to communicating its findings. Critics see a lab that picks which advisers to listen to depending on the situation.
💰 The Money Problem
The spending figures may matter more than anything else Altman said. Hundreds of millions of dollars on math research is far beyond what any university department or national grant program can spend. Altman also described an internal Slack channel called “Model Name-Hold My Beer,” where employees “were just throwing that at the hardest problems they could.”
That’s a fun bit of lab culture. It also shows that frontier labs can now aim huge compute budgets at the hardest open problems in science. Researchers worry they could be priced out of their own fields.
🧭 What This Means for You
- If you’re a researcher: Expect more releases, not fewer. Communities that build shared ways to verify AI proofs, like formal proof checkers and coordinated review groups, will cope best.
- If you’re watching the AI race: Math has turned into a flagship benchmark. Expect rival labs to make their own scientific discovery claims, and judge them on whether they’ve been verified, not on the headlines.
- If you’re thinking about policy: Questions about attribution and release norms for AI-generated science are still open. Right now OpenAI is effectively writing the rules on its own.
🔮 Looking Ahead
Altman agrees OpenAI needs to communicate better, and that’s a start. But he made it clear that slowing down isn’t on the table. The real question now is whether the math community can build verification systems quickly enough to keep up. The full report with more detail is available from The Verge AI.