In the greatest secrecy, thirty of the most brilliant mathematicians on the planet gathered in Berkeley for a weekend in May 2025. Their mission seemed simple: creating mathematical problems so complex that they would make the most advanced artificial intelligence fail. But what happened upset their vision of the future of mathematics.
A challenge that turns into a nightmare for experts
The event, organized by Epoch Ai, aimed to test the limits of the O4-Mini of Openai model, a system of revolutionary reasoning capable of complex deductions. Unlike the previous AI models which failed miserably on unpublished mathematical problems, O4-Mini promised unrivaled capacities.
The protocol was rigorous: each participant had to conceive of the problems he could solve, but which had to make the robot stumble. Confidentiality was so crucial that mathematicians had to sign non-divulgation agreements and communicate only via signal (an ultra-secure messaging application), lest their exchanges by email be analyzed by other AI and contaminate training data.
The financial issue was considerable: each problem not resolved by AI brought in its creator 7,500 dollars. But very quickly, the researchers realized that winning this money would be more difficult than expected.
When the AI exceeds doctoral students
Ken Ono, renowned mathematician from the University of Virginia, experienced one of the most striking moments of this confrontation. Confident, he submitted to O4-Mini a problem of numbers of numbers which he described as “open question”-the type of challenge which normally occupies a doctoral student for weeks.
What happened then left ono speechless. In ten minutes flat, AI has not only resolved the problem, but it has also demonstrated a remarkable methodological approach. She first spent two minutes to analyze specialized literature, then announced that she wanted to practice on a simplified version of the problem. Once this step has been mastered, she attacked the main challenge and proposed a correct but daring solution.
What has most disturbed ono is the conclusion of AI: ‘No quote is necessary because I calculated this mysterious number myself!‘Faced with this disconcerting insurance, the mathematician immediately alerted his colleagues the next morning.
Results that challenge expectations
The performances of O4-Mini amazed the assembly. While traditional AI models only solve less than 2 % of unpublished mathematical problems, O4-Mini reached around 20 % success on research level issues. Even more impressive, the system demonstrated a creative reasoning capacity comparable to that of a “very good graduate student, or even more”.
Yang Huhe He, mathematician at the London Institute for Mathematical Sciences, underlined the extraordinary speed of AI: she accomplished in a few minutes what a human expert would take weeks or months to achieve.
Despite their hard efforts, the thirty mathematicians have managed to identify only ten questions capable of blocking the robot. This meager result revealed the extent of AI’s progress in just one year.
Credit: ISTOCK
Credits: Phonlamaiphoto/Istock
A worried revolution
This demonstration of force raises deep questions about the future of mathematics. If AI continues to progress at this rate, it could soon solve problems that even the best mathematicians fail to elucidate.
Ono and his colleagues already see a future where mathematicians would change role: rather than solving the problems directly, they would ask questions and supervise reasoning AI, like a leader of third cycle students.
However, this perspective is accompanied by worries. The disconcerting assurance of O4-Mini recalls what is ironically called “proof by intimidation”: when something is affirmed with enough authority, even experts can be impressed.
A historical turning point for science
This confrontation marks a decisive turning point in the history of mathematics. For the first time, rival artificial intelligence with the capacities of the best human minds in one of the most abstract disciplines.
As ono said to his colleagues: ” It was a serious mistake to claim that generalized artificial intelligence would never exist. These models already surpass most of our best third cycle students.«
Faced with this reality, higher education will have to rethink its methods to cultivate human creativity and preserve the essence of mathematics for future generations.
