OpenAI and Anthropic Claim to Disprove Long-Standing Exponential Time Hypothesis

OpenAI said on Monday that its latest research disproves the Exponential Time Hypothesis (ETH), a theory that has guided computer-science thinking for nearly three decades. Hours later, Anthropic published its own findings that reach the same conclusion. OpenAI has invited external computer scientists to conduct an open review of the work, noting that the participants’ involvement does not constitute endorsement.

MIT professor Ryan Williams, one of three independent algorithm-design experts reviewing the claim, cautioned that he has not verified the results. He wrote on X that, if validated, the discovery could increase interest in the possibility that P = NP and might lead to faster algorithms, though he has previously expressed skepticism about ETH.

Gary Marcus, an emeritus professor at NYU who often critiques AI developments, described the claim as potentially significant if it proves correct. Applied-mathematics consultant John D. Cook, in a blog post, said the research does not resolve the central P = NP question, noting that while 3-SAT may not require exponential time, it still appears to exceed polynomial time.

Anthropic AI-safety fellow Joshua Fonseca Rivera posted that the announcement highlights how little is known about the limits of machine intelligence. Fields Medalist Timothy Gowers, a professor at the Collège de France, called the finding “very surprising and consequential.”

Other industry voices, such as Dan Shipper, CEO of the startup Every, remarked on the broader impact of the work on the tech community.