OpenAI published a 67-page paper claiming its AI model found a way to refute the Exponential Time Hypothesis (ETH), a 1999 computer-science conjecture. The hypothesis suggests that the 3-SAT logic puzzle cannot be solved faster than by testing every combination. OpenAI's model developed a deterministic algorithm that counts satisfying assignments in subexponential time.
The AI discovered the disproof on October 16, which was formalized in Lean software on October 17. The paper was published on October 17 and revised on October 25. OpenAI has committed to publishing all raw data and funding a conference on the findings. Computer scientists Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams are currently conducting an independent examination of the work.
Anthropic also released a paper on October 26, credited to its Claude model, which addresses the same 3-SAT problem. The company noted its result is purely asymptotic and does not offer practical solutions. OpenAI clarified that its own findings do not resolve the P versus NP problem, noting that turning theoretical advances into practical improvements requires significant further research.
OpenAI recently contributed $5 million to the Simons Foundation. This funding is intended to support workshops, conferences, and programs designed to help researchers analyze and build upon results generated by artificial intelligence.
