OpenAI announces three big science discoveries and a new way to work with academics

OpenAI says it has identified three significant results spanning mathematics, theoretical computer science and physics, while changing how it plans to involve outside researchers in assessing work generated by its systems. The announcement arrives less than a month after the company released hundreds of manuscripts that it said addressed long-running mathematical questions. This latest release pairs claims of new results with a more structured effort to let specialists examine, explain and test them.

The three projects are not being handled in identical ways. OpenAI said external scientists will help guide the computer-science and physics work, using the term “shepherds” for that role. For the mathematics result, it said advisers favored publishing the underlying AI-produced material so other researchers could develop and scrutinize explanations themselves. That distinction puts review and interpretation closer to the center of the release, rather than treating a model’s output as a finished scientific account.

One mathematical claim concerns local smoothing for the Euclidean wave equation. OpenAI’s released description says its proof relies on an auxiliary partial differential equation and establishes Sogge’s local-smoothing result for waves in spatial dimensions of two or more. The company also says the work has implications for the Kakeya maximal and dimension conjectures and for Stein’s spherical restriction estimates within the interior of the proposed range. The model output was posted directly, following the approach used in the company’s October 6 release.

In computer science, OpenAI says a new result challenges the Exponential Time Hypothesis. Its description presents a deterministic method for exactly counting satisfying assignments in fixed-width CNF formulas using subexponential time and polynomial space. It says the 3-SAT case would disprove the hypothesis. An independent group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams is expected to examine the material and provide an explanation of the proof.

The computer-science materials were made public immediately at the request of that working group, OpenAI said, allowing other researchers to inspect them as well. That decision illustrates the company’s stated effort to build outside examination into the process while the arguments are still being worked through. A claim of this scale remains separate from its eventual acceptance: public release can make it easier for specialists to look for errors, identify missing steps or determine whether a proposed result withstands formal analysis.

The physics result has not been published in full. OpenAI characterized it as a mathematically rigorous existence proof involving certain four-dimensional interacting quantum field theories with gauge fields and fermions. It said the scope includes Quantum Chromodynamics and a unified completion of the Standard Model. Because the complete physics information is not yet public, researchers outside the effort have limited material with which to evaluate the claim independently.

OpenAI’s new release process follows criticism of restricting early access to selected experts. The company acknowledged that its method could create problems. A panel of mathematics advisers assembled by OpenAI, known as AGMAI, said it did not recommend the practice of assigning particular people advance access to model output. The panel identified unequal access, reliance on borrowed credibility and potentially biased selection as concerns with that approach.

Those advisers instead urged AI companies to use scholarly repositories for future releases. Their recommendations also included Lean-formalized proofs, disclosure of how problems were selected, funding for people charged with understanding the material and additional supporting information. The proposals reflect a wider question raised by AI-generated scientific claims: making an output available is not the same as supplying the tools, time and incentives needed for a research community to verify it.

The next test will be what happens after publication. The mathematical release gives researchers raw model output to interpret, while the computer-science claim is being examined by an outside group and the physics project remains only partly disclosed. OpenAI’s announcement therefore combines ambitious assertions with an experiment in scientific review. Whether the results become durable contributions will depend on independent checking, clearer explanations and the response of the academic communities asked to assess them.

Source: The Verge