
OpenAI’s claim that its AI models have solved a longstanding problem in mathematical physics does not put machines close to a “theory of everything” reconciling quantum mechanics with Einstein’s gravity, three of South Africa’s leading physicists and mathematicians say. A fourth, George Ellis, doubts the OpenAI claim itself.
On 26 October, OpenAI announced AI-generated results in mathematics, computer science and physics. Among them was a proof of the existence of pure Yang-Mills theory, a stripped-down version of the theory that describes the strong nuclear force, which binds the particles inside atomic nuclei.
A “shepherding committee” chaired by Nobel laureate David Gross is reviewing the work. On 9 November, it said its checks of OpenAI’s code, written in Lean, a programming language that lets computers verify mathematical proofs, showed “that it proves the existence of pure Yang-Mills theory according to the problem statements which are standard in physics and mathematics”.
Its assessment, published the next day, adds that the work “does not prove that the infinite-volume limit is unique or that there is a mass gap”. A mass gap means the theory’s lightest particle has a mass greater than zero. Proving existence and the mass gap together is one of the Clay Mathematics Institute’s Millennium Prize problems, each of which carries a US$1-million reward.
The way AI is advancing, within the next five years, the best mathematicians and the best theoretical physicists in the world will be silicon.
Vishnu Jejjala, research chair in theoretical particle cosmology at Wits
“It’s hard to predict, given the pace at which AI is advancing, but I would say that this is highly unlikely in the near future,” said Kevin Goldstein, a high-energy theorist and professor in Wits University’s School of Physics.
The biggest obstacle, he said, is experimental. “The energy scales at which quantum gravity is relevant are not easily experimentally accessible to us.”
“It is my understanding that the path to proving the existence of Yang-Mills had already been set out, with OpenAI grinding through much of the technical detail that would have taken humans much longer. For quantum gravity, we are not sure what the correct path is.”

Vishnu Jejjala, who holds a research chair in theoretical particle cosmology at Wits and has applied machine learning to string theory, agreed. “Regarding quantum gravity, AI helping is realistic; AI solving the problem, less so,” he said.
“Taken at face value, the Yang-Mills result shows that LLMs [large language models] can carry through a long, technically challenging argument once the target is fixed. There, the theory was known, the criteria for success were written down, a strategy existed and a computer could check a proposal and verify an answer. Quantum gravity has none of these things.”
The bigger obstacle, for Jejjala, is conceptual. “Choosing the correct principles is a different task from proving a stated theorem, and nothing in this result shows whether AI can do it,” he said.
“For Yang-Mills theory, certainty without clarity is still a result. For quantum gravity, where the missing piece is a crucial insight or a set of new ideas, understanding is the result.”
‘The mathematics is really the third problem’
Jonathan Shock, associate professor of mathematics and applied mathematics and interim director of the UCT AI Initiative, said AI “is quickly making progress on some of the most difficult problems that we have ever posed”, but that unification has “some major differences”.
“One thing is that there isn’t, as far as I’m aware, an agreed-upon statement of what is to be proved,” he said. “The open questions are somewhat conceptual, like ‘What is spacetime at the very smallest scales?’ This is different from the Millennium Prize problems, which are very clearly posed.”
Without more experimental data, he said, AI could find many theories consistent with what is known, with no way of telling them apart. “So I’d say that the main issues are data and concepts, where the mathematics is really the third problem.”
“A proof that nobody understands can be checked line by line and still not tell physicists why the result is true,” Shock added. “Physicists will trust such a theory when they can explain it to each other without the machine.”

‘Publicity stunt’
Ellis, emeritus professor of mathematics and applied mathematics at UCT and Stephen Hawking’s co-author on The Large Scale Structure of Space-Time, said he was “not an expert on AI and what it can do as regards physics”. He pointed TechCentral to a post by Columbia University senior lecturer in mathematics Peter Woit titled “The Yang-Mills Publicity Stunt”.
In it, Woit calls OpenAI’s Millennium Prize framing “extremely misleading”. The Clay institute’s official problem description, written by Edward Witten and Arthur Jaffe, a member of the shepherding committee, excludes “weak-existence (compactness)” as a solution “unless one also uses other techniques to establish properties of the limit (such as the existence of a mass gap and the axioms)”.
“This is exactly the sort of ‘existence’ proof OpenAI has produced, and they have not separately established the mass gap or properties comparable to it,” Woit wrote. He also criticised the committee for keeping private the 1 147-page document explaining the proof, and urged it to “publicly explain that the OpenAI press release was misleading, release all the documents they are holding onto, and then shut themselves down”.
The committee’s assessment does say the construction satisfies the Wightman axioms, the standard requirements for a quantum field theory, which the Clay description lists alongside the mass gap.
‘Carbon chauvinists’
Jejjala, though, is bullish about AI’s longer-term prospects. “The way AI is advancing, within the next five years, the best mathematicians and the best theoretical physicists in the world will be silicon. In areas where there is an objective quality to truth and beauty, we shouldn’t be carbon chauvinists,” he said. — (c) 2026 NewsCentral Media





