OpenAI says its AI solved three big science problems, and experts will check

OpenAI announced three "major" discoveries this week, one each in math, computer science and physics, and none has been independently confirmed yet. The most striking claim says a widely believed idea about how fast computers can solve certain problems, in place since 1999, is false. The release comes less than a month after OpenAI put out hundreds of math manuscripts.

What OpenAI claims

In computer science, OpenAI says its AI disproved the Exponential Time Hypothesis. That idea holds that some hard problems cannot be solved quickly, and it has been widely believed. The write-up and a computer-checkable proof are now public on GitHub.

In math, OpenAI says the AI proved the local smoothing conjecture in every dimension. It says this also settles the Kakeya conjecture and Stein's restriction conjecture in all dimensions. Months ago, NYU professor Hong Wang won a Fields Medal for resolving Kakeya in three dimensions.

In physics, the claim is a rigorous proof that certain quantum theories exist, including the one describing the strong nuclear force. The full physics material has not been released.

Who will check the work

OpenAI calls its outside reviewers "shepherds." For computer science, a group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will examine and explain the proof. Impagliazzo and Paturi helped propose the hypothesis in 1999. For physics, David Gross will lead a group of five experts. For math, OpenAI released the raw AI output and left researchers to build their own explanations.

Even the advisers disagree

A panel of math advisers that OpenAI assembled said just a day earlier that AI companies should not give chosen people early access. It named three problems: unequal access, borrowed reputation and biased selection. The panel preferred public releases on scholarly repositories, with proofs a computer can check, notes on how problems were chosen, and funding for the people who work to understand the results.

OpenAI admits its own approach could have issues. It also published an open letter asking the research community how to handle early access, how to give credit for checking and fixing work, and how to support students and researchers.

What happens next

The key step is review by independent experts, and that can take time. Earlier this month, OpenAI pledged $5 million to help mathematicians make sense of AI-generated research. Until reviewers weigh in, these results are claims, not established facts.