OpenAI and Anthropic say their models refuted the Exponential Time Hypothesis
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
If verification confirms the result, SI will have helped overturn a foundational assumption of computational complexity theory and open a path to algorithms long thought impossible.
OpenAI published work with an algorithm that, according to the company, solves 3-SAT problems in subexponential time; Anthropic later released its own refutation of the same hypothesis. The Exponential Time Hypothesis (ETH) holds that the time to solve such problems inevitably grows exponentially with the number of variables; it is used as a basis for complexity estimates of many computational problems. This is not a proof that P equals NP, but according to MIT professor Ryan Williams, the result could change views on the difficulty of such problems. Williams, along with researchers Russell Impagliazzo and Ramamohan Paturi, who formulated the ETH in 1999, are organizing a discussion of the work and of norms for publishing results obtained with SI. According to TechCrunch, OpenAI checked the implications for cryptography and believes the algorithm does not make existing systems insecure.
Sources 2