OpenAI chief executive Sam Altman has admitted that the company's AI mathematics experiments 'got out of control' for one week, but insists the research should continue.
Speaking on a podcast released on 29 October 2026, he defended the deliberate decision to push the technology's limits despite criticism from mathematicians over transparency, verification and academic recognition.
During the Cheeky Pint interview with Stripe co-founder John Collison, Altman acknowledged the disruption AI discoveries could cause for scientists who have devoted their careers to solving difficult mathematical problems.
His remarks follow OpenAI's release of more than 700 mathematical research files in October. The company also faces calls to publish supporting evidence behind separate claimed breakthroughs in mathematical physics.
OpenAI's AI Maths Experiments Cost Hundreds of Millions
During the podcast interview, Altman acknowledged that OpenAI had deliberately given researchers considerable freedom to explore what its AI models could achieve.
'Look, we did let it get out of control for one week on the math,' Altman said. He explained that researchers needed to test the technology's limits before finding more efficient ways to conduct similar experiments.
One experiment involved the Navier–Stokes equations, which describe how fluids such as air and water move. On 8 September, OpenAI announced what it described as a solution to the longstanding Navier–Stokes existence and smoothness problem.
The company released a written proof and a formalisation using Lean, a mathematical proof-assistant system. However, publication and computer-assisted verification do not, by themselves, establish broader scientific acceptance of the claimed solution.
Altman estimated that the Navier–Stokes work cost approximately $10 million (approximately £7.47 million), possibly more.
When Collison suggested that OpenAI's wider mathematical research had cost hundreds of millions of dollars, Altman broadly agreed. However, the figure was an estimate rather than a confirmed total. He said the expenditure also covered unpublished findings and suggested that improved AI models could substantially reduce future costs.
Hundreds of AI Proofs Raise Verification Concerns
On 6 October, OpenAI released a collection of mathematical results produced by an internal AI model. The repository contained more than 700 files, raising questions about how researchers could assess such a large volume of material.
Mathematical proofs require examination for correctness, originality and significance. Computer-assisted verification can establish whether a formal argument follows specified rules. However, it does not automatically determine whether a result is genuinely new, important or properly credited.
The independent Advisory Group on Mathematics and Artificial Intelligence, based at the Institute for Advanced Study, has advocated transparent, community-led examination of AI-generated results.
The Association for Human Mathematics criticised OpenAI more sharply in a 7 October statement. It accused the company of disregarding established research practices and argued that releasing more than 700 files simultaneously demonstrated technological power rather than scholarship.
OpenAI has acknowledged the need to improve how it shares research.
On 15 October, it announced a $5 million (approximately £3.73 million) contribution to a Simons Foundation initiative helping mathematicians examine and build on AI-generated results. The foundation will manage funding for academic programmes, workshops and conferences.
Why OpenAI's Physics Claims Face Transparency Questions
The disagreement intensified on 26 October when OpenAI announced three additional claimed discoveries involving mathematics, computer science and physics. Its physics research concerned quantum chromodynamics, Yang–Mills theory and possible mathematical constructions related to the Standard Model.
The existence of quantum Yang–Mills theory forms part of a Millennium Prize Problem. However, OpenAI's announcement does not establish that the full prize problem has been independently validated or officially recognised as solved.
OpenAI proposed different publication arrangements for the three discoveries. For the physics research, it invited external experts to organise a collaboration and left decisions about releasing supporting material to that group.
The company said participants would remain independent, could publish their conclusions and would not be required to sign non-disclosure agreements. It also stated that it was not seeking academic credit for discoveries generated by its internal models.
The advisory group challenged the decision in a 26 October statement.
It called for the immediate release of OpenAI's natural-language proofs and Lean outputs relating to the physics claims. The group argued that researchers deserved equal access to the evidence, particularly after OpenAI reported completing its Yang–Mills formalisation.
Altman Says Humanity Deserves the Discoveries
Altman acknowledged that the dispute extends beyond mathematical correctness.
Scientific discoveries also shape researchers' careers, professional recognition and personal fulfilment. He said OpenAI needed to consider the people whose work had advanced mathematics over generations.
Nevertheless, he maintained that scientific progress should continue, even when new technology disrupts established research practices.
'I think, on net, the world deserves to understand how the universe works,' Altman said.
He also confirmed that OpenAI possesses additional mathematical findings beyond those already published. The company is considering how to share future results while addressing concerns raised by scientific communities.
For critics, the outstanding questions involve transparency, fair recognition and opportunities for independent examination. Producing mathematical results rapidly is not the same as establishing their significance or securing scientific acceptance.
Altman's position remains clear: OpenAI intends to continue exploring mathematical discoveries with AI. Whether it can maintain researchers' confidence will depend partly on how openly its future claims are presented, tested and understood.
