OpenAI Announces Three Scientific Breakthroughs and a New Academic Collaboration Model
Less than a month after releasing a massive batch of 722 manuscripts that claimed to solve hundreds of mathematical problems, OpenAI published late on October 26 three new results it describes as "discoveries" — a term the company has been using increasingly as its models advance into territory once reserved for humans. Unlike the previous release, this time the company adopted an experimental peer-review strategy guided by external researchers, in a move that reveals both the ambition and the increasingly visible tension between AI acceleration and traditional scientific validation.
The three announced results
The first discovery concerns harmonic analysis, a field of pure mathematics that studies how complex signals and functions can be decomposed into elementary waves. In 1991, mathematician Charles Sogge proposed the local smoothing conjecture — an assertion about the behavior of waves as they focus. The conjecture posits that the extremes of concentration must be short-lived, which has deep implications across multiple areas of mathematics. In 2019, Guth, Wang, and Zhang proved the conjecture in two spatial dimensions; in 2025, Wang and Zahl resolved the Kakeya conjecture in three dimensions. OpenAI claims its models produced a proof for all dimensions — which, if confirmed, would resolve not only the original conjecture but also the Kakeya conjecture and Stein's restriction conjecture in any dimension. The proof, generated on October 18, was formalized in the Lean proof assistant the following day.
The second discovery is even more impactful for computer science: OpenAI's models refuted the Exponential Time Hypothesis (ETH), a conjecture formulated in 1999 by Russell Impagliazzo, Ramamohan Paturi, and Stefan Wiedemann. Since then, the ETH has been used in hundreds of papers to derive computational hardness results across fields ranging from graph theory to computational biology, databases, and computational geometry. The proof that the hypothesis is false — if independently validated — would reopen fundamental questions about how fast core optimization problems can be solved. OpenAI obtained the natural-language refutation on October 16 and formalized the result in Lean on October 17.
The third discovery sits at the frontier of theoretical physics and formal mathematics: an existence proof in constructive quantum field theory. As a subproblem, the work includes a proof of the existence of pure Yang-Mills theory, which is part of the Clay Mathematics Institute's Millennium Prize problem. This represents an advance in the most difficult and long-standing problems of mathematical physics — something that the 2004 Nobel Prize in Physics awarded to David Gross, Frank Wilczek, and David Politzer, for their discovery of asymptotic freedom in Yang-Mills theory, only underscores.
The new "shepherding" academic model
What differentiates this round from previous releases is OpenAI's attempt to involve external experts directly in the validation and contextualization process. The company organized its approach around four principles: public transparency without non-disclosure agreements, no academic credit claimed by OpenAI, formal recognition for researchers who write "shepherd" analyses, and independent, unpaid participation.
For the mathematics result, OpenAI decided to release the model's output directly, leaving researchers free to organize their own explanatory efforts. For the computer science and physics results, the company formed external collaborator groups — which it called "shepherds" — to guide understanding and validation of the results, while the AI output remained publicly available on the company's GitHub.
OpenAI also announced a $5 million donation to the Simons Foundation to fund academic conferences and workshops studying AI-generated results. The idea is that these events would serve as neutral spaces where mathematicians and physicists can examine formal proofs in Lean and build on the findings.
Tensions and criticisms
OpenAI's approach has not passed without criticism. The AGMAI (Advisory Group on Mathematics and Artificial Intelligence) panel issued a warning to the mathematical community, recommending that AI companies avoid designating specific researchers to receive preliminary results. The argument: this practice creates unequal access, generates dependence on the reputations of selected researchers, and may produce selection biases that distort independent evaluation.
Additionally, there is a historical precedent: in previous releases, Gizmodo reported that an OpenAI Navier-Stokes proof was formalized in Lean in a way that verified a weaker proposition than claimed in the natural-language text. This raised questions about the reliability of automatic formalizations — an issue the mathematical community watches with caution.
The text from OpenAI itself acknowledges the tension: "choosing the initial shepherds inevitably shapes who participates first and creates an information asymmetry." The company says it is open to feedback on how to make the process more inclusive.
Why it matters
If confirmed, the three results would represent real leaps in what AI can achieve in mathematics, theoretical physics, and computer science. But the most significant may be the publication model itself: OpenAI is testing whether the combination of formal proofs in Lean + peer review guided by experts can serve as a bridge between what machines produce and what academia considers valid science.
What to watch is whether this "shepherding" model will generalise — or whether the scientific community will ultimately demand even stricter standards of transparency and universal access, as recommended by AGMAI.
Sources: The Verge, Al Jazeera, Silicon Report, OpenAI, Gizmodo
✓ Independent sources cross-checked and verified before publishing
