AI Disproves the Exponential Time Hypothesis (Twice, Same Day)

Two students hand in the same homework, on the same day, with the same surprising answer. At school, that gets you sent to the principal. In AI, it gets you a press release.

On October 26, 2026, OpenAI said one of its internal AI models had disproved the Exponential Time Hypothesis, a rule computer scientists have relied on for almost 30 years. A few hours later, Anthropic, the company behind Claude, posted a paper saying its own model had cracked the same problem.

The hypothesis was proposed in 1999 by Russell Impagliazzo and Ramamohan Paturi, two computer scientists at UC San Diego. It says certain hard logic puzzles can't be solved much faster than by trying every possible answer. Add more pieces and the time needed explodes.

Hundreds of research papers treat that as a given. If the AI is right, a lot of them need a second look.

Now the spicy part. Both companies say their AI found the key trick on the same date, October 16. Anthropic says Claude got there with no human input.

And one of Anthropic's own mathematicians, Levent Alpöge, hinted that OpenAI's solution leaned on existing work.

It's the group project where both people swear they did everything, except the group is two of the biggest AI labs on the planet.

The people checking the work aren't random either. Impagliazzo and Paturi, with MIT professor Ryan Williams, will review the claim in the open. Williams says the group signed no NDAs and isn't getting paid, and that the proof has "two or three alien-looking parts."

Personally, I find this more impressive than any chatbot demo. A machine told the people who wrote the rule that their rule is wrong, and now they're the ones grading its homework.

But careful: nothing is confirmed yet. Both proofs were checked with Lean, a program that verifies every logical step, but human experts still have to go through them. This doesn't solve the famous P versus NP problem, and it won't make your laptop faster tomorrow.

Encryption should be fine too. OpenAI says it checked and is reasonably confident nothing breaks. For more AI meets security, read how Anthropic built an AI that finds security holes in the internet.

For 27 years, nobody could break this rule. Then two AIs broke it on the same day. The only thing left to prove is who gets the credit.

Photo: Anthropic