IT Home On October 28, OpenAI announced on Monday local time that the company has made remarkable breakthroughs in mathematics, computer science, and physics, while also saying it is adjusting its communication methods for releasing these achievements. However, despite this apparent change in strategy, researchers who spoke with The Verge expressed increasing frustration with the company’s approach.
IT Home noticed that in the past few months, there have been continuous conflicts between OpenAI and the mathematical community. At the beginning of October, the company released hundreds of preprints at once, leaving the group of mathematicians overwhelmed. This time, OpenAI said it would cooperate with domain experts who act as “guides” to determine the appropriate way to publish its latest results.
Three achievements target several of the most important unsolved problems in the fields of mathematics, computer science, and physics. Some continue OpenAI’s previous research directions, while others open up new research paths. According to the suggestions given by these “guides”, the three achievements were released using different methods: the achievements in mathematics were directly released as AI-generated derivations, “to be interpreted by academic researchers on their own”. For the achievements in computer science, OpenAI would form an external research team to interpret them. As for the achievements in physics, they will not be made public at this time; OpenAI said it would be decided by an external expert team regarding the subsequent release method.
This mathematical achievement is quite impressive; it is said to have proven the Kakeya conjecture in all dimensions. Simply put, this problem investigates: how much space is required to make a needle or a pencil rotate in various directions. Earlier this year, mathematician Wang Hong won the prestigious Fields Medal based on her proofs of the three-dimensional version of the conjecture. Brendan Hassett, a mathematics professor at Brown University, said that the conclusion given by OpenAI “seems well-supported,” but he also emphasized that he has not yet verified the Lean formalization proof, which is the formal system relying on computer verification of the result.
The achievements in the field of physics revolve around the standard model, which is the most mature and complete theoretical framework in physics. The research focuses on quantum chromodynamics, a theory that describes the strong nuclear force, one of the four fundamental forces, as well as its relationship with the Yang-Mills theory. The Yang-Mills theory is also a cornerstone of modern particle physics and one of the seven Millennium Prize Problems. However, this research by OpenAI has not resolved the unresolved “mass gap” issue.
The last result claims to refute the exponential time hypothesis. This is a widely accepted assumption in theoretical computer science: some problems cannot be solved much faster than brute force enumeration. The hypothesis is closely related to the P vs. NP problem, which is also a Millennium Prize Problem and one of the most well-known unsolved problems in this field. Its core involves exploring whether problems whose answers are easy to verify can also be solved quickly.
Due to the varying level of transparency of these achievements, and the high verification standards for such results, it is difficult for researchers to fully evaluate them. Even if all the data are available, the verification process would take a lot of time. However, experts who were interviewed said that these achievements are worth taking seriously, and many described them as “of great significance,” “major breakthroughs,” and “of profound importance.”
Scientists are certainly excited by these scientific advancements, but at the same time, they are annoyed by OpenAI’s approach. After receiving a large number of research conclusions produced by AI in recent months, this sense of fatigue is already very familiar. Many researchers hope that the entire academic community can focus on the intensive verification work, and question why OpenAI still needs to set additional hurdles and personally select experts to oversee the entire process.
Most experts who were interviewed approved of OpenAI’s approach of seeking opinions from external researchers. However, they suggested that publishing research conclusions publicly means that the relevant materials should be made available for everyone to verify. Constantin Kogler, a researcher at the Higher Institute, said, “Compared to previously releasing hundreds of papers at once, this time only three results were published. I agree with this release pace.” But he disagreed with limiting access to only those experts selected by OpenAI. He said, “There is no reason for only allowing a few well-known scientists to view the content from the beginning.”
Andreas Karch, a professor of physics at the University of Texas at Austin, said: “It is very bad that they personally select the expert team. Large laboratories already have enough influence, and they shouldn’t decide who is qualified to participate and who is excluded.” At first, Karch proposed that researchers could be selected through a selection process, but he also admitted that the people who are ultimately selected will likely still be top scholars. Then he asked another question: Why couldn’t the proof materials be released simultaneously, along with a team of experts for interpretation? “This way, everyone would have the opportunity to access the content, and it also reflects their desire for the results to be understood and valued.”
Just one day before OpenAI’s official announcement, the AGMAI, a group of experts in mathematics and artificial intelligence, published a statement specifically warning against such information control practices. The experts said, “We do not recommend that AI companies adopt this approach.” They believe it is unnecessary to prioritize opening content to specific groups. This method relies on the prestige of selected researchers, and it also poses risks such as uneven access rights, biased selection processes, and unexplained delays in progress. The proposal put forward by the expert group is: The AI laboratory should upload the results to an academic preprint platform, include formal proofs and explanations of the research topics, and provide funding and supporting resources to help the academic community interpret these conclusions.
After the release of these new results, AGMAI urges OpenAI to “immediately publish the natural language manuscript and Lean formal proof corresponding to this physics result publicly”:
These are both fundamental problems in physics and important mathematical issues. The existence of the quantum Yang-Mills theory is a Millennium Prize Problem, and the construction of the four-dimensional interaction quantum field theory has been a goal in mathematical physics for decades.
We greatly respect the scientists invited by OpenAI to lead this project. However, as we previously stated, the interpretation of the results should be carried out in a public environment, and everyone should have equal access rights. OpenAI says that the delay in publication is to complete the Lean formalization proof. According to them, this work has now been completed, and from a scientific standpoint, there is no reason to continue delaying the publication.
Hassett believes that making the results fully public also helps promote scientific development and facilitates a wider acceptance of new tools like AI. “Multiple research teams around the world working together on these findings can help researchers who need them get started using these new tools more quickly.”
It may not be realistic to find a communication method that satisfies researchers in all fields that OpenAI is entering. OpenAI states that it is trying different approaches, and acknowledges that each method involves trade-offs. Jared Duker Lichtman, a mathematician at Stanford University, said: “My feeling is that no matter which approach is chosen, there will always be people who are dissatisfied. We just need to find a way that causes the least controversy.”
Xiaoyu He, a mathematics professor at the Georgia Institute of Technology, believes that the problem is much deeper than its presentation format. “The results themselves are impressive, but in my view, they are merely deliberately showing a responsible and cautious attitude, while avoiding the core contradiction: these fundamental challenges have been quickly solved by AI, with no contribution from human researchers throughout the process, and the proof texts produced are difficult to read and verify.”