OpenAI Announces Three New Advances in Math and Science, but Declines Academic Authorship

OpenAI announced on Monday that its AI models have made new progress in three fields: mathematics, computer science, and physics. All three are quite hardcore: topics related to the Kakeya conjecture in harmonic analysis, the Exponential Time Hypothesis in computational complexity, and the existence problem of quantum chromodynamics and the completion of the Standard Model.

OpenAI did not attach the label of "proof" to these results, saying only that progress has been made. It has reached out to external experts, asking them to serve as facilitators and organize teams to understand the findings. Participation is voluntary and unpaid. Participants do not sign non-disclosure agreements and are free to discuss, share intermediate results, and publish conclusions. OpenAI also said it would not seek academic authorship for these discoveries generated by its internal models.

So what exactly are these three topics?

The Kakeya conjecture sounds obscure, but its geometric picture is intuitive: find a set in the plane that contains a line segment of length 1 in every direction. Mathematicians want to know the smallest possible dimension of such a set. The answer is known in two dimensions, while the higher-dimensional case remains open. Local smoothing estimates in harmonic analysis are closely tied to this conjecture. OpenAI's results lie somewhere around here; the specific conclusions have not yet been made public.

The Exponential Time Hypothesis (ETH) is a common assumption in complexity theory: NP-complete problems such as 3-SAT cannot be solved by any algorithm running in sub-exponential time. A large body of lower-bound proofs rests on ETH. If ETH turns out to be false, problems once believed exponentially hard may not be that hard, and the underlying foundations would need to be reworked.

Quantum chromodynamics is the theory describing the strong interaction between quarks and gluons. The "existence" here refers to whether a quantum field theory can be rigorously defined in mathematics; experimentally, it has long been verified. This issue bears on whether the Standard Model can become a self-consistent and complete theory, and it is a tough nut to crack in fundamental physics.

This time, OpenAI did not rush to publish a paper. It handed the opportunity to understand, verify, and publish conclusions to external experts. The absence of NDAs means the intermediate process can be brought into the open. Declining academic authorship also sends a signal: AI-generated research results need humans to take responsibility. Ceding authorship credit to human experts may also be a strategy: if these findings are ultimately overturned, OpenAI does not have to bear responsibility for the errors; if they are confirmed, whoever first provides credible verification gets to set the narrative. OpenAI does not lose out either way.

Of course, it is also worth pouring some cold water on this. AI models have produced many elegant but incorrect results. New advances in mathematics and physics must undergo rigorous human verification. OpenAI has not said which model produced these results, nor has it released the full derivation process. How much external experts can understand, and how bold a judgment they dare to make, remain unknown.

Feeding AI discoveries into the human academic system sounds like an experiment. It may succeed, or it may fail. OpenAI has at least opened the door wide to scrutiny; the rest depends on whether the experts can handle it.