OpenAI Says Its AI Disproved a Major Computer Science Hypothesis. Experts Are Now Checking
OpenAI claims its model refuted the Exponential Time Hypothesis and produced two more big math and physics proofs. None of it is independently verified yet. Here's what was actually announced.
Hundreds of computer science papers are built on one assumption that almost everybody in the field believed was true. On Monday, OpenAI said one of its internal AI models proved that assumption false.
That’s a huge claim, so let’s be clear up front: it’s a claim. It hasn’t passed independent review yet, and the experts checking it are saying so loudly.
What OpenAI announced
In a blog post on October 26, OpenAI announced three results its models produced:
- Computer science: a refutation of the Exponential Time Hypothesis (ETH), a decades-old conjecture that a classic logic problem (Boolean satisfiability, or SAT) can’t be solved dramatically faster than brute force. ETH is closely tied to the famous P vs. NP question, and researchers have leaned on it to argue that problems in graph theory, biology, databases and more are hard.
- Math: a proof of Sogge’s 1991 local smoothing conjecture about how waves focus, in every dimension. If it holds, it would also settle the Kakeya and Stein restriction conjectures in all dimensions.
- Physics: a rigorous existence proof for certain quantum field theories in four dimensions, including quantum chromodynamics, the theory of the strong force. It includes a proof that pure Yang–Mills theory exists, which is part of one of the Clay Institute’s million-dollar Millennium Prize problems.
OpenAI says its model found the ETH disproof on October 16 and that it was formalized in Lean, a language that checks proofs by machine, the next day. The write-up and Lean code are public in OpenAI’s GitHub repo. The physics materials are being held back while an expert group writes them up.
Who’s checking it
Here’s the part I find most interesting. The ETH check will be led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, two of the people who originally proposed the hypothesis, plus MIT’s Ryan Williams. According to The Next Web, the three wrote that “our involvement should not be taken as certification of the claim.” They’re unpaid, signed no NDAs, and asked OpenAI to release all the raw data and fund a conference on the result. OpenAI agreed.
The physics work will be shepherded by a group led by Nobel laureate David Gross, The Verge reports.
Why it matters
If the ETH refutation survives scrutiny, it reopens questions about how fast a lot of “hard” problems can really be solved, though OpenAI itself says turning it into practical speedups will take significant further research. And it would be one of the biggest signs yet that AI can produce genuinely new results in science, not just summarize old ones.
If it doesn’t hold up, it’ll be a very public lesson in why “a computer checked it” and “the field accepts it” aren’t the same thing.
Would you trust a proof written by AI if a computer checked it? Tell me on X @JCredibleTech.
Sources: OpenAI, The Verge, The Next Web, OpenAI GitHub.
