OpenAI announces three big science claims for researchers
OpenAI has announced three major research claims in mathematics, computer science and physics. Outside experts are helping review the work in computer science and physics. The company says it is changing how it shares results created by AI after recently releasing a set of mathematical papers.
According to The Verge, two sets of research materials are available on GitHub. OpenAI has not released all the physics materials. Researchers and the wider academic community are being asked to consider how claims like these should be checked.
Experts help review two of the claims
In mathematics, OpenAI says its work is about local smoothing for the Euclidean wave equation. The company says the proof uses an additional partial differential equation to show that Sogge’s local smoothing result holds for waves in every space dimension of two or more. OpenAI also says the result proves the Kakeya maximal and dimension conjectures, along with Stein’s spherical restriction estimates across the full range of cases expected by the conjecture.
For this mathematical result, OpenAI says it shared the model's output directly so researchers could study it and work out explanations. This is different from the process used for the other two fields, where outside experts are helping review the work.
The computer science paper claims to disprove the Exponential Time Hypothesis, an idea that many researchers have long believed to be true. OpenAI says the paper describes a deterministic algorithm that counts all satisfying assignments of fixed-width CNF formulas exactly, using subexponential time and polynomial space. The company says the result for 3-SAT is what disproves the hypothesis.
An independent working group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will study the proof and explain it. OpenAI says the group asked for the materials to be released right away so other people could examine them. The company says the group also decided to publish the AI-generated explanation and the Lean proof on GitHub.
The full physics research has not been published
OpenAI describes its physics result as a mathematically rigorous proof that certain interacting quantum field theories with gauge fields and fermions exist in four dimensions. The company says the work includes Quantum Chromodynamics and a unified version of the Standard Model of particle physics. The Verge reported that the full research materials have not yet been released.
OpenAI says David Gross will lead a group that includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. The group will help guide the physics research. According to The Verge, OpenAI did not provide the complete research materials in the release it described.
How these results are shared has also become part of the debate. OpenAI’s group of mathematics advisers, AGMAI, recommended that AI companies should not give selected people access to results before everyone else. According to The Verge, the group raised concerns about unfair access, people gaining credibility from others' reputations and bias in choosing who gets early access.
Instead, the advisers recommended publishing the work in scholarly repositories and including proofs written in Lean, details about how the problems were chosen, and funding for people who study and explain the results. These steps would help researchers check not only the main claims but also the evidence behind them and how the results were produced.
OpenAI also published an open letter asking the research community for views on giving people early access to AI-generated results, revealing conflicts of interest and using unpublished work in later research. It asked how researchers should receive credit for discovering results as well as checking, correcting, simplifying, explaining and building on them. It also asked how students and researchers can be supported as research methods change.
These questions matter in everyday academic work. Outside experts can help check difficult claims, but giving only a few people early access may prevent others from examining the results or building on them. OpenAI’s latest release therefore includes both scientific claims and an ongoing debate about who should be allowed to review AI-generated research and when they should get access.
Source: theverge.com