We are announcing three new discoveries across mathematics, computer science, and physics, spanning harmonic analysis, complexity theory, and quantum field theory.
Our release process continues to evolve based on feedback from relevant academic communities. In our view, these manuscripts are best understood as raw research data. To help transform this raw research data into scientific understanding, we are experimenting with a new approach for this release. We reached out to outside experts to serve as shepherds and organize teams to help understand these new results. They suggested three different release strategies:
- Mathematics result: Release the AI-generated material directly and let researchers self-organize to interpret the work.
- Computer science result: Assemble an external collaborator team to interpret the result, while publicly releasing the AI-generated material for anyone to inspect and build upon.
- Physics result: Assemble an external collaborator team and follow their advice on how to release the materials.
We do not yet know which of these three approaches is most effective, and we welcome feedback on what works and where we can improve.
External collaborations will follow four core principles:
- Transparency. We are transparent about the existence of the collaboration. Participants are not required to sign NDAs and are free to discuss the work, share intermediate findings, and publish their conclusions.
- No Academic Credit. We do not seek academic credit for discoveries produced by our internal models. Inviting researchers to help shepherd a collaboration does not imply that they or their working groups made the initial discovery.
- Recognition. The understanding built, reports written, and other materials created by the team should be properly recognized. The team will also help ensure that the work appropriately credits the researchers, techniques, and ideas on which it builds.
- Independence. Participation is voluntary and uncompensated. Working group members participate independently, and their conclusions reflect only their own views. They do not speak on behalf of OpenAI, nor does OpenAI speak on theirs.
We will hand over the materials and artifacts related to these results to the collaboration teams. For the computer science collaboration, the team decided to immediately make both the natural-language proof and the Lean formalization public, and we are releasing them directly to our GitHub repository. For the physics collaboration, the team chose to hold off on releasing the AI-generated artifacts for now, opting to write their own exposition first.
All of these approaches involve trade-offs. On one hand, these collaborations empower experts to thoroughly dissect the results and contextualize them within the broader academic landscape. On the other hand, choosing the initial shepherds inevitably determines who gets early access, creating information asymmetry. We welcome feedback on how to broaden participation and make this process more practical.
Local Smoothing and the Kakeya Conjecture
The mathematics result falls within the field of harmonic analysis. In 1991, Sogge formulated the local smoothing conjecture, which concerns the behavior of waves as they focus. Specifically, the conjecture posits that the most extreme focusing must be fleeting. The degree of focusing at a typical, random point in time is always far less pronounced than what can occur in the worst-case scenario. This is a recurring theme across mathematics: worst-case behavior is often far more extreme than what is observed on average. Given the fundamental nature of this conjectured phenomenon, it has broad implications throughout mathematical analysis. For example, it implies both the Kakeya conjecture and Stein’s restriction conjecture.
In 2019, Guth, Wang, and Zhang proved the local smoothing conjecture in two dimensions. In 2025, Wang and Zahl proved the Kakeya conjecture in three dimensions. Our model has produced a proof of the local smoothing conjecture in arbitrary dimensions, thereby also resolving the Kakeya conjecture and Stein’s restriction conjecture in all dimensions. This new proof is based on a novel auxiliary partial differential equation (PDE).
We obtained the AI output on October 18, 2026, and completed the formal proof in Lean the following day. We then reached out to an expert to discuss collaborating around the result. Based on their advice, we decided to release the AI output publicly, following the same strategy as our October 6, 2026 release. We have published both the AI natural-language output report generated by our model and the Lean proof in a public GitHub repository.
The Exponential Time Hypothesis is False
The computer science collaboration team will work to understand and explain a result concerning the Exponential Time Hypothesis (ETH). This is a long-standing conjecture regarding the computational complexity of Boolean satisfiability (SAT), a problem closely tied to the P vs. NP question. ETH has long been widely believed to be true by the academic community, serving as the foundation in hundreds of papers to derive conditional computational hardness results across graph theory, computational biology, databases, computational geometry, and many other areas. Remarkably, our system recently proved that…