OpenAI announced on October 26 that its internal models have produced three new scientific results: a proof of Sogge's local smoothing conjecture in harmonic analysis, a disproof of the Exponential Time Hypothesis in computer science, and a rigorous existence proof for quantum chromodynamics and related quantum field theories in physics. None of the three has been peer reviewed. OpenAI is unusually candid about that, calling the manuscripts raw research data and handing them to outside experts to examine, explain, and, if they hold up, turn into actual scientific understanding.
The computer science claim is the one with the widest blast radius. The Exponential Time Hypothesis, proposed by Russell Impagliazzo, Ramamohan Paturi, and Francis Zane, holds that 3-SAT cannot be solved in subexponential time, and hundreds of papers have used it to derive conditional hardness results across graph theory, databases, computational biology, and geometry. OpenAI says its system found a natural-language disproof on October 16 and formalized it in the Lean proof assistant the next day. Both the writeup and the Lean proof are public on GitHub. Fittingly, the examination will be led by Impagliazzo and Paturi themselves, alongside MIT's Ryan Williams, and OpenAI has agreed to fund a conference on the implications. If the result stands, it does not make hard problems fast overnight; OpenAI itself notes that turning it into practical algorithms will take significant further research.
In mathematics, the claimed proof of Christopher Sogge's 1991 local smoothing conjecture covers every spatial dimension, which would also resolve the Kakeya conjecture and Stein's restriction conjecture in all dimensions; human mathematicians had settled local smoothing in two dimensions in 2019 and Kakeya in three dimensions in 2025. The AI output arrived on October 18 and was formalized in Lean a day later. The physics claim is the most closely held: existence proofs for four-dimensional quantum field theories including QCD, built on Balaban's approach to constructive quantum field theory, with pure Yang-Mills existence, part of a Clay Millennium Prize problem, as a subproblem. A group led by Nobel laureate David Gross will shepherd that work, and unlike the other two, its materials are being withheld until that group writes them up.
The release model may matter as much as the results. OpenAI says it seeks no academic credit, set no nondisclosure agreements, and made participation voluntary and unpaid, with three different strategies chosen by the experts: open release for the mathematics, public release plus an external group for computer science, and expert-led delay for physics. That is a more honest template than a victory-lap announcement, and it acknowledges the asymmetry baked in: the lab still chooses who sees extraordinary claims first.
For people building with AI, the lesson is not that conjectures are falling like dominoes. It is that generation has outrun verification, again. A Lean formalization is a real check on the argument as formalized, but experts still have to confirm the formal statement matches the conjecture mathematicians actually care about, which is why none of this counts as settled until the examiners report. The same trust gap showed up in miniature on October 27, when TechCrunch reported Bank of America emailing customers not to share their banking passwords with AI assistants. Whether the claim is a disproved hypothesis or an agent asking for your login, the productive question is the same: what, exactly, has been verified, and by whom?
