OpenAI announced Monday that its artificial intelligence systems have produced major solutions across mathematics, theoretical computer science, and physics, including a purported refutation of a foundational computational hypothesis.
The San Francisco company made the findings public less than a month after posting hundreds of AI-generated math papers, The Verge reported. Rather than handling the new claims entirely in-house, OpenAI reached out to prominent outside researchers whom it called “shepherds” to inspect, explain, and contextualize the raw outputs.
Refuting a cornerstone of complexity theory
In computer science, OpenAI claims its internal model disproved the Exponential Time Hypothesis, a conjecture introduced in 1999 by Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. The hypothesis addresses the computational limits of solving Boolean satisfiability, a problem tightly linked to the unresolved P versus NP question, according to The Next Web. Hundreds of academic papers rely on the assumption that the conjecture holds true.
OpenAI stated that its model produced a deterministic algorithm capable of counting satisfying assignments for fixed-width formulas in subexponential time. An independent panel that includes Impagliazzo, Paturi, and Massachusetts Institute of Technology professor Ryan Williams has taken on the task of evaluating the claim. The trio chose to publish the AI output and machine-verified proof directly on GitHub so the wider community could review it.
“Our involvement should not be taken as certification of the claim. A result of this significance requires careful, independent scrutiny,” Impagliazzo, Paturi and Williams wrote in an open letter to fellow researchers.
Williams posted on social media that the group signed no non-disclosure agreements and received no compensation, according to The Next Web.
Proofs in harmonic analysis and quantum physics
In mathematics, the company stated that its model established Christopher Sogge’s local smoothing conjecture in all spatial dimensions of two or more. The result simultaneously settles the Kakeya maximal conjecture and Stein’s spherical restriction estimates, OpenAI said.
Alex Iosevich, a mathematics professor at the University of Rochester, characterized the announcement to Al Jazeera as a “historic development” that initially brought shock, optimism, and a “twinge of sadness” over the shifting role of human mathematicians.
For physics, the model supposedly produced a mathematically rigorous proof of existence for specific four-dimensional quantum field theories, including quantum chromodynamics and aspects of pure Yang-Mills theory, which connects to a Millennium Prize problem. A research group led by Nobel laureate David Gross, joined by Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov, is guiding that review. Unlike the computer science panel, the physics team opted to withhold the raw AI outputs while drafting its own human analysis.
Debate over access and academic credit
The company’s strategy of selecting specific scholars to evaluate early outputs sparked immediate scrutiny. An advisory group called AGMAI cautioned that granting early access to favored researchers risks creating information imbalances and borrowing scientific prestige, The Verge reported.
OpenAI acknowledged those concerns in its public announcement, writing that while collaborations assist with contextualizing findings, choosing initial leaders inevitably creates asymmetry. The company also announced a $5 million donation to the Simons Foundation to fund academic conferences and workshops centered on analyzing AI outputs.
Scientists worldwide have now begun the painstaking process of examining the proofs line by line to determine whether the machine-generated logic holds up under full peer review.
