What happened
OpenAI has announced that its internal models have produced proofs for three long-standing scientific problems: the local smoothing conjecture in mathematics, the Exponential Time Hypothesis (ETH) in computer science, and the existence of certain interacting quantum field theories in physics. To manage these results, the company is piloting three distinct collaboration models with academic experts to verify, explain, and publish the findings.
In mathematics, OpenAI's models produced a proof for the local smoothing conjecture in all dimensions, which also resolves the Kakeya and Stein’s restriction conjectures. The proof relies on a novel auxiliary partial differential equation. OpenAI has released the natural language output and a Lean formalization publicly.
In computer science, the models produced a disproof of the Exponential Time Hypothesis (ETH), a foundational conjecture in computational complexity. This result has been formalized in Lean. A group of experts, including Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams, is leading the effort to examine the implications of this finding.
In physics, the models provided a mathematically rigorous existence proof for interacting quantum field theories in four dimensions, including Quantum Chromodynamics (QCD) and pure Yang–Mills theory. This work is being shepherded by a group of prominent physicists, including David Gross and Michael Douglas. Unlike the other results, the physics group has opted to hold off on releasing the AI-produced artifacts while they conduct their own write-up.
OpenAI is testing three different engagement models: releasing AI output directly for community review (math), forming an external group to explain the result while making data public (computer science), and deferring entirely to an external group for the release of materials (physics).
Why it matters
These results represent a significant shift in the role of AI in theoretical research, moving from assisting with data analysis to generating formal proofs for foundational conjectures. The refutation of the ETH, if verified, could fundamentally alter computational complexity theory, while the physics proof addresses a Millennium Prize problem. By experimenting with academic 'shepherding' models, OpenAI is attempting to navigate the tension between rapid AI-driven discovery and the need for rigorous, community-led peer review and verification.
The refutation of the ETH could have profound implications for design, potentially opening doors to faster solutions for problems previously thought to be computationally intractable. The ETH is a cornerstone of modern complexity theory, used in hundreds of papers to establish conditional hardness.
The physics result addresses the Quantum Yang–Mills Theory Millennium Prize problem. If the proof holds up to rigorous academic scrutiny, it would mark a historic milestone in theoretical physics, establishing a mathematically rigorous foundation for the Standard Model's gauge fields and fermions.
The initiative highlights a new paradigm for AI-human collaboration in science, where AI acts as a generator of complex, formal proofs, and human experts act as 'shepherds' to translate, verify, and contextualize these outputs for the broader scientific community.
OpenAI has explicitly stated it does not seek academic credit for these discoveries, aiming to position its models as tools for scientific advancement rather than primary authors, though the involvement of AI in such foundational work raises ongoing questions about attribution and the future of peer review.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
The primary focus is the ongoing verification process by the appointed academic groups. Specifically, the physics community's assessment of the Yang–Mills existence proof is critical, as it touches upon a Clay Mathematics Institute Millennium Prize problem. Additionally, the success of the three different collaboration strategies—ranging from open public release to expert-led gatekeeping—will likely influence how OpenAI and other AI labs disclose future scientific breakthroughs.
The community's reception of the ETH disproof is a key indicator of how the field will handle AI-generated results that contradict long-held beliefs. The planned conference, funded by OpenAI, will be a central venue for this discussion.
The physics collaboration's eventual release of their findings will be closely monitored, given the high stakes of the Yang–Mills problem and the prestige of the researchers involved.
OpenAI's commitment to funding independent examination through the Simons Foundation suggests a long-term strategy to integrate its models into the academic research , which may lead to more frequent, high-impact scientific announcements.
The effectiveness of the 'shepherding' model remains an unknown. The company has acknowledged that these approaches involve tradeoffs, such as the potential for information asymmetry and the influence of initial shepherds on the direction of research.