OpenAI releases AI-generated proofs for math, physics, and computer science

Three new scientific frontiers

OpenAI recently announced three major scientific discoveries generated by its models in the fields of mathematics, computer science, and physics. These findings follow a previous release of hundreds of manuscripts that claimed to solve long-standing math problems. The company is now looking for ways to better integrate these outputs into the formal scientific process.

In mathematics, the model provided a demonstration of the local smoothing conjecture across all dimensions. This specific result also resolves the Kakeya conjecture and Stein’s restriction conjecture in all dimensions. These are complex problems that have occupied researchers for years, with some recently receiving prestigious awards for solving specific cases of them.

The computer science discovery involves a deterministic algorithm that calculates the number of satisfying assignments for fixed-width CNF formulas using subexponential time and polynomial space. This specific finding refutes the Exponential Time Hypothesis (ETH). This hypothesis has been widely believed to be true since it was originally proposed in 1999.

Physics and quantum fields

The physics discovery involves a mathematically rigorous existence proof for certain interacting quantum field theories. These theories contain gauge fields and fermions in four dimensions. This includes Quantum Chromodynamics (QCD) and a unified completion of the Standard Model of particle physics.

Unlike the math and computer science outputs, the physics information has not been released in its entirety. OpenAI stated that David Gross will lead a group of experts to shepherd this work. This group includes Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov.

The term 'shepherding' in this context refers to human experts guiding the results and ensuring they are correctly interpreted. This approach highlights the difficulty of verifying high-level physics results that may be too complex for a single person to validate quickly. It suggests a collaborative model between AI generation and human oversight.

New verification protocols

OpenAI is experimenting with different ways to release these results to the public. For the computer science proof, an independent working group of scientists will examine and explain the findings. This group includes Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams, who were among the original proposers of the hypothesis.

The company is making these materials public immediately so others can examine them as well. This includes the AI output write-up and a Lean proof. Lean is a formal proof language that allows computers to check the logical consistency of mathematical statements.

The company admits this approach could have its own issues. A panel of math advisers recently recommended that AI companies avoid designating specific people to receive outputs first. They cited concerns regarding unequal access, borrowed reputation, and biased selection as potential problems with the current model.

The path toward formal science

The math results were released directly for researchers to work through their own explanations. This contrasts with the more guided approach taken for the physics and computer science discoveries. OpenAI is seeking to establish norms for how the community should deal with AI-generated breakthroughs in the future.

The math advisers suggested that labs should instead release output on scholarly repositories. They recommended including Lean-formalized proofs and clear information on how the specific problems were chosen. They also suggested providing funding for people who are tasked with understanding the results.

By providing these materials on GitHub, OpenAI is inviting the community to participate in the verification process. They have released an open letter asking for input on how to examine the claims. This move signals a transition toward a more structured, peer-reviewed style of AI-assisted research.

Impact on the research cycle

These discoveries represent a significant leap in the types of problems AI can tackle. Moving from simple data processing to proving complex conjectures in physics and math changes the role of the AI. It moves the tool from a creative assistant to a primary generator of technical evidence.

For the scientific community, this creates a need for new infrastructure. Researchers will need to develop ways to audit these proofs quickly to ensure they are correct. The speed of AI generation may outpace the speed of human peer review, creating a bottleneck in the validation phase.

As these models continue to produce more complex outputs, the focus will likely shift toward verification. The goal is to create a system where AI can propose a solution and a team of humans can confirm it. This collaborative loop is the current frontier for how AI will interact with academia.