Two AI labs broke the same law of computer science within hours. Coincidence?

A 27-year-old assumption fell twice in one day, and no human has actually read the proof yet.

Ok, you have to hear this one. Back in 1999 three researchers wrote down something called the Exponential Time Hypothesis. The short version: a classic logic puzzle called 3-SAT, basically "can you flip these true/false switches so the whole formula comes out true", cannot be solved meaningfully faster than exponential time. Nobody proved it. Everybody just assumed it. And more than 800 papers quietly built on top of that assumption.

On 26 October OpenAI announced that one of its internal models had found a sub-exponential algorithm. Hypothesis dead. They put the proof on GitHub with a formal version in Lean, and the original authors of the hypothesis are among the people checking it.

Then, a few hours later, Anthropic dropped its own proof. Different approach, same result. Claude did it on its own, no human in the loop, burning about 14 million tokens. It even took down the stronger version, SETH, along the way. The paper lists its author as "Claude (Anthropic)".

Hold on. Two labs, a problem that sat untouched for 27 years, and they both crack it in the same week and publish on the same day? Seriously?

There are two ways to read it, and neither is a conspiracy. One: the top models crossed roughly the same capability line at roughly the same time, so whoever pointed theirs at this problem first was going to find something. Anthropic dates its discovery to 16 October, ten days before going public. Two: once one lab announces, the other has every reason to publish immediately, because in science second place gets almost nothing. Put those together and "same day" looks less like magic and more like a race where both runners were already near the finish.

So what does it mean for the rest of us?

Practically, not much yet. The new algorithm is still painfully slow in real life, so your bank's encryption is fine and P versus NP is still open. The real shock is somewhere else. Anthropic admits that the 1,050 Lean modules were written and checked by machines, and that no person is known to have read the proof. A fundamental piece of human knowledge just changed, and for now the only ones who have verified it are other computers.

That is the bit to sit with. We used to ask whether AI could do real science. The new question is whether we can keep up with checking it.