正文 · AI 翻译
IT Home reported on October 28 that OpenAI announced on Monday local time that it had made striking breakthroughs in mathematics, computer science and physics, while saying it was adjusting how it communicates the release of its results. Yet despite what appeared to be a change of course, researchers interviewed by The Verge were growing increasingly frustrated with the company's approach.
IT Home noted that friction between OpenAI and the mathematics community has persisted for months. In early October, the company released hundreds of preprints at once, leaving mathematicians struggling to keep up. This time, OpenAI said it would work with subject matter experts serving as “guides” to determine appropriate ways to release its latest results.
The 3 results address some of the most important unsolved problems in mathematics, computer science and physics. Some continue OpenAI's earlier lines of research, while others open new ones. Following the guides' advice, the 3 results have different release plans: for the mathematics result, OpenAI has released the AI-generated derivation directly, “leaving researchers in the academic community to organize themselves to interpret the result”; for the computer science result, OpenAI will assemble an external research group to interpret it; and the physics result will remain undisclosed for now, with OpenAI saying an external team of experts will decide how it is released later.
The mathematics result is startling: it reportedly proves the Kakeya conjecture in all dimensions. Put simply, the question asks how little space is needed to rotate a needle or pencil so that it points in every direction. Earlier this year, mathematician Hong Wang won the prestigious Fields Medal for work including a proof of the 3-dimensional version of the conjecture. Brendan Hassett, a mathematics professor at Brown University, said OpenAI's conclusion “appears to be well founded,” but stressed that he had not yet checked the Lean formal proof, the system used to verify the conclusion by computer.
The physics result concerns the Standard Model, the most mature and comprehensive theoretical framework in physics. The research focuses on quantum chromodynamics, which describes the strong nuclear force, 1 of the 4 fundamental forces, and its connection to Yang–Mills theory. Yang–Mills theory is also a cornerstone of modern particle physics and figures in 1 of the 7 Millennium Prize Problems, although OpenAI's research does not resolve the outstanding “mass gap” problem.
The final result claims to disprove the Exponential Time Hypothesis. This is a widely accepted conjecture in theoretical computer science that some problems have no solution method much faster than brute-force enumeration. The hypothesis is closely related to the P versus NP problem, which is also a Millennium Prize Problem and 1 of the field's best-known unsolved questions. At its heart is whether problems whose answers are easy to verify can themselves be solved quickly.
Because the results have been disclosed to differing degrees and are exceptionally difficult to verify, researchers cannot easily assess the claims in full. Even with all the materials, verification would take considerable time. But the experts interviewed said the results deserved serious attention; many described them as “enormous,” “major breakthroughs” and “far-reaching.”
Researchers were excited by these scientific advances, but they were also irritated by OpenAI's approach. After months of receiving large numbers of AI-generated research claims, that sense of fatigue was already familiar. Many researchers want the academic community as a whole to concentrate on the demanding work of verification, and they question why OpenAI is adding barriers by personally selecting experts to control the process.
Most of the experts interviewed approved of OpenAI consulting outside researchers, but argued that publishing research claims means making the relevant materials available for everyone to check. Constantin Kogler, a researcher at the Institute for Advanced Study, said, “I approve of this pace of release: 3 results this time, compared with hundreds released all at once before.” But he disagreed with restricting access to experts selected by OpenAI. “There is no reason to let only a few prominent scientists see the material at first,” he said.
Andreas Karch, a physics professor at the University of Texas at Austin, said, “Having them handpick the team of experts is very bad. Large labs already have enough influence; they should not decide who gets to take part and who is excluded.” Karch initially suggested selecting researchers through a screening process, though he acknowledged that those chosen would probably still be leading scholars. He then asked why the proof materials could not simply be released at the same time as an expert interpretation group was assembled. “That would give everyone a chance to access the material while also showing that they want the results to be understood and taken seriously.”
Just 1 day before OpenAI's announcement, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) published a warning specifically about this kind of information control. “We do not recommend that AI companies adopt this model,” the group said. It argued that giving selected people early access is unnecessary, draws on the prestige of the researchers chosen, and creates risks including unequal access, biased selection and needless delays. The group's proposal was for AI labs to upload results to academic preprint platforms, include formal proofs and explanations of the problems being addressed, and provide funding and other support to help the academic community interpret the claims.
After the new results were announced, AGMAI urged OpenAI to “immediately release the natural-language manuscript and Lean formal proof for the physics result”:
These are both central questions in physics and important mathematical questions. The existence of quantum Yang–Mills theory is a Millennium Prize Problem, and constructing an interacting quantum field theory in 4 dimensions has been a goal in mathematical physics for decades.
We have great respect for the scientists OpenAI has invited to lead this work. But, as we said in our earlier statement, interpretation of the results should proceed in public, with everyone having equal access. OpenAI says it delayed publication to complete the Lean formal proof. By its own account, that work is now complete, so there is no scientific reason to delay publication further.
Hassett believes that making the results fully public would also advance science and help new tools such as AI gain wider acceptance. “Having multiple research teams around the world study these results together would help researchers who need them start using these new tools sooner.”
It may be unrealistic to find a way of communicating that satisfies researchers in every discipline OpenAI is entering. OpenAI says it is trying different approaches and acknowledges that each involves trade-offs. Jared Duker Lichtman, a mathematician at Stanford University, said, “My sense is that whatever approach is chosen, someone will be unhappy. We just have to find the one that causes the least controversy.”
Xiaoyu He, a mathematics professor at the Georgia Institute of Technology, believes the issue runs much deeper than the form of release. “The results themselves are very impressive, but to me, they are merely trying to present themselves as responsible and cautious while avoiding the central problem: AI is rapidly solving these fundamental problems without contributions from human researchers at any stage, and the resulting proofs are hard to read and difficult to verify.”