OpenAI’s ETH Bombshell Puts AI Discovery on Academic Trial
OpenAI’s ETH Bombshell Puts AI Discovery on Academic Trial
OpenAI said it had made three AI-generated discoveries spanning harmonic analysis, computational complexity and quantum field theory—then handed the hardest work to outside scholars: determining what, exactly, the systems had found.
The company said its models generated a proof of the local smoothing conjecture in every dimension, implications for the Kakeya and Stein restriction conjectures, and a rigorous existence result involving quantum chromodynamics and a completion of the Standard Model. Most explosively, it said its system had disproved the Exponential Time Hypothesis (ETH), a decades-old assumption used across hundreds of computer-science papers. “The ETH has been widely believed to be true,” OpenAI said, adding that its claimed refutation had been formalized in Lean, a proof-verification language.
The chronology matters. OpenAI said its model found a natural-language disproof of ETH on October 16, 2026, and that the result was formalized the following day. Rather than simply publishing the conclusion, it recruited a group led by UC San Diego’s Russell Impagliazzo and Ramamohan Paturi—among ETH’s original proposers—and MIT’s Ryan Williams. At their request, OpenAI released the raw materials publicly and committed to fund a conference on the result’s implications.
That process reflects a shift from the lab’s earlier proof releases, which critics said left researchers scrambling to assess sweeping claims. Williams, who had previously likened an OpenAI proof dump to a “DDOS attack on mathematicians,” agreed to help organize the response. He said parts of the ETH proof appeared routine, but “two or three alien-looking parts” made clear that “the whole community needs to work on [it] together.”
If the claim holds, ETH-dependent results may need revisiting and researchers may have to reassess how plausible the broader proposition that P equals NP has become. Williams stressed that OpenAI’s algorithm is unlikely to be the final word, while also saying it could challenge long-held intuitions about which hard problems are tractable.
The immediate concern is verification, not victory laps. Williams said OpenAI had checked whether the result threatened cryptographic systems and was reasonably confident it did not. The larger question is whether an open, expert-led review can turn a startling model output into durable science.
Write a comment