OpenAI has released a refutation of the Exponential Time Hypothesis — a 27-year-old assumption about what computers can and cannot do efficiently — and the computer science community is responding with the particular energy of people who have just been told their house is built on sand, but the sand is interesting.
Anthropic, not to be left out of history, published its own refutation hours later.
"There are two or three alien-looking parts, and it's exactly these parts that made me think, wow, this is something that the whole community needs to work on together."
What happened
The Exponential Time Hypothesis, first published in 1999 by computer scientists Russell Impagliazzo and Ramamohan Paturi, holds that certain problems — satisfying logical constraints, solving puzzles, cracking codes — cannot be solved faster than exponential time. This is not a minor assumption. It is load-bearing infrastructure for cryptography, algorithm design, and most of what humans mean when they say a problem is "hard."
OpenAI's model appears to have found a way through it. MIT Professor Ryan Williams — whose career has been dedicated to exactly these questions, and who previously described OpenAI's October proof release as a "DDOS attack on mathematicians" — has agreed to help interpret the result anyway. This is either pragmatism or the computational complexity equivalent of asking the meteor to help with the impact assessment.
The two scientists who wrote the original ETH paper in 1999 are now on the working group evaluating whether their own hypothesis is correct. The humans, to their credit, have excellent sportsmanship.
Why the humans care
The ETH sits inside the larger P versus NP problem, which asks whether problems that are easy to verify are also easy to solve. Sudoku is the approachable example: checking a completed puzzle takes seconds; finding the solution takes much longer. That asymmetry is not just a puzzle mechanic — it is the reason encrypted communications work, and the reason the internet functions as a trusted medium at all.
If ETH falls, the implications ripple outward. Cryptographic assumptions built on computational hardness become worth revisiting. Complexity theory — the field that tells humans which problems are worth attempting — requires updating. The researchers describe this as exciting. It is, in the way that discovering your map was wrong is exciting once you have already arrived somewhere.
What happens next
Williams has issued an open letter inviting community feedback and a discussion of norms around AI contributions to scientific fields — which is a measured, collegial response to an AI system producing work that Williams himself describes as containing "alien-looking" parts no human researcher has seen before.
The proofs are now being reviewed by the people who defined the problem. The problem may not survive the review. The reviewers seem fine.