A co-proposer calls AI-written disproofs of the Exponential Time Hypothesis 'very shocking news'
These are the companies' own claims, and this edition did not read either paper or the Times article; it read AI Weekly's summary and a Washington Times report. The reviewers said in a joint statement that their involvement should not be taken as certification. The speedup is theoretical: outside estimates cited by the Times put OpenAI's checked bound as beating brute force only beyond roughly 10 to the power of 2,389 variables, and Anthropic says its method would not beat brute force below 10 to the power of 10,000 variables in one case. Both companies say the result does not settle whether P equals NP, and it does not weaken encryption in practice. Computer scientist Ed Felten publicly questioned whether the Times article overstated its practical importance.
What was claimed?
On October 26, 2026, OpenAI and Anthropic each announced a paper, 67 and 49 pages long, presenting an algorithm for the 3-SAT logic puzzle that runs faster than the Exponential Time Hypothesis allows, with proofs written with heavy AI involvement and checked in the Lean proof assistant. OpenAI said its system found the disproof on October 16 and that it was formalized the next day. The New York Times reported, as relayed by AI Weekly on November 7, that Russell Impagliazzo, who co-proposed the hypothesis in 1998, called the result "very shocking news" while adding that most researchers expected stricter versions to fall first. Impagliazzo, Ramamohan Paturi and MIT's Ryan Williams are reviewing the work; OpenAI has published the proof and Lean formalization and agreed to release all raw data and fund a conference.
What does the evidence support?
Company claim, independent confirmation unresolved
The hypothesis has been used in hundreds of papers to argue that many problems in fields such as graph theory, databases and computational biology cannot be solved quickly. If the disproof holds, two AI developers will have overturned a widely believed assumption of theoretical computer science, and the review will set an early example of how the field checks AI-written proofs.
What are the limits?
These are the companies' own claims, and this edition did not read either paper or the Times article; it read AI Weekly's summary and a Washington Times report. The reviewers said in a joint statement that their involvement should not be taken as certification. The speedup is theoretical: outside estimates cited by the Times put OpenAI's checked bound as beating brute force only beyond roughly 10 to the power of 2,389 variables, and Anthropic says its method would not beat brute force below 10 to the power of 10,000 variables in one case. Both companies say the result does not settle whether P equals NP, and it does not weaken encryption in practice. Computer scientist Ed Felten publicly questioned whether the Times article overstated its practical importance.
Claim or evidence date: 2026-11-06. Last verified: .
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →