Story
Oct 27, 2026
OpenAI’s AI Breakthroughs Now Face the Harder Test: Human Understanding
OpenAI says its models have cracked major problems in mathematics, complexity theory and physics. But its new release experiment puts verification, access and academic credit in the hands of outside scholars—and exposes the tensions that follow.
OpenAI has announced what it calls three new discoveries in harmonic analysis, computational complexity and quantum field theory—then immediately made the way it released them part of the experiment. Sam Altman amplified the announcement from OpenAI researcher Dan Roberts, who described findings spanning “mathematics, computer science, and physics.”
1
The company says the results include a proof of the local smoothing conjecture in every spatial dimension, with implications for the Kakeya and Stein restriction conjectures; a claimed disproof of the Exponential Time Hypothesis; and a rigorous existence result for quantum field theories including quantum chromodynamics. OpenAI frames the outputs not as finished scholarship but as “a form of raw research data” requiring expert interpretation.2
Its timeline begins with the models’ reported work: OpenAI says it obtained the harmonic-analysis output on October 18, 2026 and formalized it in Lean the next day. It says the ETH disproof was produced on October 16 and formalized on October 17. For mathematics, the company released the model output publicly, allowing researchers to organize their own scrutiny.2
For complexity theory, OpenAI instead convened an external group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams, while publishing the natural-language proof and Lean formalization immediately. The group is meant to test and explain the claim that ETH—an assumption used across hundreds of hardness results—is false. For physics, external collaborators will first write up the work, and the underlying AI artifacts remain unreleased for now.2
OpenAI says it will not claim academic credit and that participating scholars should receive recognition for the explanations and papers they produce. It also insists the groups are independent, voluntary and free to publish their conclusions.2
Yet the company concedes the central dilemma: selecting initial “shepherds” creates an information imbalance. The AGMAI panel of mathematics advisers went further, recommending that AI companies not designate people to receive outputs first, citing “unequal access, borrowed reputation, and biased selection.”3 The breakthroughs may be AI-generated; whether they become science will depend on a far more human contest over who gets to inspect, explain and own the understanding.