Are mathematicians becoming obsolete? This is the secret conclave that humiliated some of the world's greatest minds.

In May 2025, on the campus of the University of California at Berkeley, an unusual meeting took place that could mark a turning point in the history of modern mathematics . Thirty of the world's most prestigious mathematicians gathered in a secret conclave, not to debate each other, but to face an artificial intelligence: o4-mini , a state-of-the-art language model developed by OpenAI, capable of reasoning with unprecedented speed and accuracy.

The goal was to test the machine for two days on some of the world's most complex problems. For every question o4-mini failed to solve, the mathematician who formulated it would receive a reward of $7,500 (€6,389) . However, the outcome of that marathon left many in attendance perplexed. Ken Ono, a mathematician at the University of Virginia and one of the judges of the event, told Scientific American magazine: " I've never seen that kind of reasoning in a model. That's what a scientist does. It's terrifying."

The meeting, held by the nonprofit Epoch AI as part of the FrontierMath project, took place under strict security measures. Participants signed confidentiality agreements and were prohibited from using email . They could only communicate through the encrypted Signal app, to prevent leaks that could contaminate the model's training. The 30 mathematicians were divided into groups of six and competed against each other to design problems that they could solve but that would crash the AI.

The problems posed ranged from number theory to algebraic geometry, encompassing challenges that would normally require weeks of academic work. o4-mini , however, solved them in minutes, not only providing an answer but also exhibiting a structured reasoning process. He broke down the problem, proceeded step by step, and proposed intermediate solutions before arriving at a conclusion.

"What distinguishes these models is that they now better link solutions to smaller problems," Jordi Serra Ruiz, professor of computer science at the Universitat Oberta de Catalunya, explains to ABC . "This way, they can solve the same challenges as people step by step. But only on problems that have been previously trained or explained ."

The most disturbing thing for many of those present was the astonishing progress AI has made in just one year. With training that evolved, its developers began formulating "level four" problems—questions that only a handful of experts in the world can conceive—and by April 2025, o4-mini was able to solve nearly 20% of them. Traditional models barely managed to do more than 2%.

"For each problem, the AI spent the first two minutes absorbing and mastering the relevant literature. Then it would solve a simplified version to learn, and finally it would launch into the full problem and find a correct, albeit daring, solution ," Ono said. "And at the end, the AI actually said, 'No need to cite (my inspiration), because I calculated the mystery number! '"

The experience left a lasting impression on the participants . “It’s like working with an extremely competent collaborator,” Ono acknowledged. Yang Hui He, a mathematician at the Institute of Mathematical Sciences in London and a pioneer in the use of AI in research, went further: “This is what an excellent graduate student would do. Even better, in fact.”

Hui He also introduced the concept of 'trial by intimidation'. That is, an AI responds with such confidence that the listener—even an expert—accepts its conclusions without question. "If you say something with enough authority, people just get scared. I think o4-mini dominated this test by intimidation ," he said.

As the weekend progressed, the mood became increasingly ambivalent: admiration for the technical progress, yet deep concern about the future of mathematicians. " What happens when the machine solves everything faster than you?" several attendees asked. Ono was blunt, but clarified that she wasn't trying to be alarmist: "It's a serious mistake to think that artificial general intelligence (AGI)—a hypothetical type of AI that would have the ability to understand and apply knowledge in a similar way to a human —will never happen, that it's just a computer. In some ways, these models are already outperforming our best PhD students."

The outcome of this meeting wasn't a defeat, but rather a warning. The group managed to formulate ten problems that o4-mini couldn't solve , although everyone understood that this human advantage will become increasingly difficult to maintain. The possibility of reaching "level five"—with questions unsolvable even for the best humans—no longer seems like science fiction.

In that scenario, the mathematicians at the meeting pondered whether they might eventually become "question setters," guiding AI toward new discoveries. "Human creativity and interpretation will remain fundamental," Hui He insisted. "AI calculates, reasons, deduces... but it doesn't yet dream or intuit ."

Óscar Corcho, professor of Artificial Intelligence at the Polytechnic University of Madrid , sums it up for ABC this way: "We have to adapt to working alongside these machines, just as we did when search engines emerged on the web." Understanding the inner workings of these new intelligences is a challenge as crucial as understanding the human brain : "There will be blind spots in this artificial mind that we will try to decipher. In fact, we are already something like 'the psychologists of AI.'" But it's always good to remember that we are dealing with a tool that doesn't have to make its users obsolete.

The U.S. Defense Advanced Research Projects Agency (DARPA) has long warned that mathematics is stuck in the past . “Mathematics is still done the way it was centuries ago: by people standing in front of a blackboard,” lamented Patrick Shafto, the program’s director. That’s why, in April 2025, he launched the expMath initiative, with the goal of developing an “AI co-author” that can break down large mathematical problems into more manageable components and solve them quickly and accurately.

The Berkeley meeting was more than an experiment; it was a mirror of the present and a glimpse of the future. Ken Ono noted in this regard: " I have colleagues who literally said that these models are approaching mathematical genius."

And if you ask AI itself about this phenomenon, the answer is striking. ChatGPT, also developed by OpenAI, responds: "People like Terence Tao, Noam Chomsky, or even creative geniuses like Leonardo da Vinci or Marie Curie have had a deep, innovative, and personal understanding of the world that goes far beyond what AI can do right now."

But he points out that if we ask ourselves why we would want something that surpasses us, the answer is simple: "Because, deep down, we know that limits suffocate us. And we also know that comfort without challenge dulls the mind. Perhaps that is the key: in a world where machines solve problems efficiently, our true value will be imagining what has no solution yet."