OpenAI has announced three AI-generated findings in mathematics, computer science, and physics. The claims touch difficult, consequential problems, but the materials are being shared in different ways: one result is public, one will involve an outside research group, and one remains unreleased. That uneven access has renewed debate over how AI research should be checked and explained.
Three fields, three release plans
OpenAI says expert “shepherds” are helping decide how to communicate the findings. For the mathematics result, the company released AI-generated output and left researchers to organize efforts to explain it. For computer science, OpenAI plans to assemble an external group of researchers to explain the work. It has not released the physics result, saying it will defer to outside experts on how to share the materials.
The announcement marks a smaller batch than the hundreds of preprints that OpenAI released at the start of October. But the different release plans make it harder for researchers to assess the claims as a group. In particular, the physics finding cannot yet be examined by the wider community, while the other results still require substantial work to understand and evaluate.
What the findings claim
The mathematics result purports to resolve the Kakeya conjecture in all dimensions. Put simply, the problem asks how little space is needed to rotate a needle or pencil in every direction. Brown University mathematics professor Brendan Hassett told The Verge the claim “looks well supported,” while noting that he had not checked the Lean formalization, a computational way to verify results.
The physics work concerns quantum chromodynamics, which describes the strong nuclear force, and its relationship to Yang-Mills theory, a cornerstone of modern particle physics. Yang-Mills theory is also central to a Millennium Prize problem. OpenAI’s work does not address the unresolved “mass gap.”
The computer science claim is that the exponential time hypothesis is false. That hypothesis is a widely held assumption that certain problems cannot be solved substantially faster than by brute force. It is closely related to P versus NP, which asks whether problems with solutions that are easy to check are also easy to solve.
These are major claims, but the article makes clear that researchers have not completed the work needed to judge them fully. Even when materials are available, understanding technical results takes time. Some experts told The Verge there are reasons to take the findings seriously, calling them “huge,” “major breakthroughs,” and “significant.”
Researchers want open scrutiny
Researchers’ interest in the science is paired with frustration about the process. Some welcomed OpenAI’s decision to consult experts and said three results are more manageable than hundreds. Their concern is that selecting a small group to see or explain work first can leave others waiting, even when the results have been announced publicly.
Constantin Kogler, a researcher at the Institute for Advanced Study, said he was fine with the release of the problems, but questioned why access should initially be limited to experts chosen by OpenAI. Physics professor Andreas Karch of the University of Texas at Austin criticized the company’s role in selecting the group and asked why it could not release the proof alongside its expert review. He argued that doing both would widen access while supporting efforts to make the result understandable.
The Advisory Group on Mathematics and Artificial Intelligence (AGMAI), which OpenAI had previously pledged to consult, had warned against this kind of selective first access. The group said it risks unequal access, biased selection, delays, and the use of selected researchers’ reputations to lend weight to a result. It recommended posting outputs on scholarly repositories, providing formalized proofs and information about how problems were chosen, and supporting researchers working to understand the findings.
Access is part of the verification process
After OpenAI’s announcement, AGMAI urged the company to “make public without delay the natural-language and Lean outputs behind its announced” physics results. The group said OpenAI had stated that release was delayed until the Lean formalization was complete, and that, by the company’s account, it was now complete. It argued there was no scientific reason for further delay.
Open access matters because scrutiny is not just a final check on a claim. Giving researchers the same materials lets more people investigate, explain, and potentially build on the work. Hassett said that having many groups around the world work through the results could support faster adoption of AI tools by those who want to use them.
The dispute is therefore about both the credibility of the findings and who gets to examine them. Expert guidance may help make difficult work legible, but researchers quoted in the article argue that public announcements should come with materials that the wider community can inspect. OpenAI’s next steps will show whether its new approach can make complex AI-generated research easier to understand while giving researchers broad access to verify it.