OpenAI releases math and CS breakthroughs with expert shepherds

OpenAI announced significant breakthroughs in mathematics, computer science, and physics on Monday, introducing a new communication strategy that relies on expert "shepherds" to manage the release of its findings. According to The Verge, the company is attempting to course-correct after a turbulent period in October when it unceremoniously dumped hundreds of preprints on the research community. This latest batch includes three results concerning some of the most consequential unsolved problems in their respective fields, each with a different disclosure strategy based on the advice of these external experts.

The mathematics result was released immediately, with OpenAI stating it is "leaving researchers free to organize their own efforts to explain the result." For the computer science finding, the company plans to assemble an external group of researchers to help explain the work. The physics result, however, has not been released at all; OpenAI says it will defer to an external group of experts on how to share the materials. This tiered approach reflects the company's stated goal of balancing scientific rigor with community access, though it has sparked considerable debate among the very experts it is consulting.

The details

The specific results involve deep questions in theoretical physics and mathematics. The physics findings relate to the existence of quantum Yang–Mills theory, which is part of a Millennium Prize problem, and the construction of interacting quantum field theories in four dimensions, a goal of mathematical physics for decades. The Advisory Group on Mathematics and Artificial Intelligence (AGMAI), a group of experts OpenAI had previously pledged to consult, issued a statement urging the company to "make public without delay the natural-language and Lean outputs" behind these announced results. AGMAI noted that OpenAI had delayed release until the Lean formalization was complete, but since that is now done, there is "no scientific reason for delaying the release any further."

Constantin Kogler, a researcher at the Institute for Advanced Study, welcomed the more manageable batch of three results compared to the hundreds released previously. However, he questioned the decision to restrict access to experts of OpenAI’s choosing, stating, "It doesn’t make any sense that first only a few distinguished scientists can look at them." Andreas Karch, a physics professor at the University of Texas at Austin, was more critical, telling The Verge, "The fact that they handpick the group is a disaster... They should not be picking winners and losers." Karch suggested that OpenAI could simply release the proof now to democratize access while still showing care for making the results meaningful.

Context

This announcement follows a period of significant friction between OpenAI and the mathematical community. In early October, OpenAI released a large volume of AI-generated preprints, which researchers described as an "exhausting barrage." The new "shepherd" model is an attempt to mitigate the impact of such releases by ensuring that results are vetted and explained by trusted experts before wider dissemination. However, AGMAI warned against this practice just a day before the announcement, arguing that selecting certain people to receive output first is "unnecessary" and risks unequal access and biased selection. Instead, AGMAI recommended that AI labs release output on scholarly repositories with formalized proofs and provide funding for people working on understanding the results.

Stanford mathematician Jared Duker Lichtman acknowledged the difficulty of the situation, saying, "My sense is that for any approach, there is going to be someone who is upset. So it’s just a question of finding the least offensive option." This highlights the broader tension between the rapid pace of AI-driven discovery and the traditional norms of open, peer-reviewed science. The company is experimenting with different approaches, acknowledging that each has tradeoffs, but the current strategy of gatekeeping through handpicked experts remains controversial.

What it means for developers

For developers and researchers working with AI tools for scientific discovery, this situation underscores the importance of transparency and open access to data and proofs. The use of Lean formalization for the physics results is a notable technical detail, as it allows for machine-verifiable proofs, which can be a powerful tool for ensuring correctness. However, the restricted access to these proofs means that developers and researchers outside the selected group cannot immediately verify or build upon these results. This could slow down the adoption of new tools and methods, as noted by Hassett, who argued that "having many groups around the world wrestle with these results supports rapid adoption of new tools by those that would like to use them."

Developers should be cautious about relying on AI-generated scientific results without independent verification. The difficulty in reading and verifying these papers, as noted by Xiaoyu He, a mathematics professor at the Georgia Institute of Technology, suggests that even experts are struggling with the complexity. He stated, "The results are very impressive, but I think they are making a show of trying to be responsible and thoughtful about how they release results while not being able to change the main issues: that fundamental results are being solved so quickly without human contribution, and that the papers are difficult to read and verify."

What to watch

• Release of Physics Results: Watch for OpenAI to release the physics results and the Lean formalization, as urged by AGMAI. This will be a key test of the company's commitment to open science.
• Community Reaction: Monitor the response from the broader research community to the "shepherd" model. Will it be seen as a positive step, or will it further alienate researchers?
• Adoption of AI Tools: Track how quickly researchers and developers adopt the new tools and methods demonstrated in these results. Open access will likely accelerate this process.
• Future Release Strategies: Observe how OpenAI handles future announcements. Will it continue to use the "shepherd" model, or will it shift to a more open approach in response to criticism?