Students who won where their teachers lost: A new definition of brilliance is born

July 2025, in an Australian exam room, as the clock ticks relentlessly, hundreds of teenagers attempt to solve some of the most difficult and challenging problems in mathematics. They have no access to calculators or the internet. Just paper, pencils, and their ingenuity. What was extraordinary this year was not just the human talent assembled at the International Mathematical Olympiad (IMO), but the fact that, for the first time, artificial intelligence achieved something unthinkable just a year ago. And while the machines shone like never before… humans won, at least for now.

Another question this story raises is whether, as AI becomes an increasingly powerful presence in academia, an inevitable question arises: what does it mean to be brilliant in an era where AI could solve almost anything?

The IMO is no ordinary competition. Over two days, students face six problems that require not only logic and technique, but also creativity and mathematical intuition. The questions, which range from algebra to number theory, are designed to be unique. If they resemble something already solved anywhere in the world, they are discarded.

This year, participants included some of the brightest young people on the planet, and invited guests included top tech companies, highlighting the AI model, Gemini Deep Think from Google DeepMind. This AI was evaluated using the same criteria and time as humans: four and a half hours per exam, without external assistance. IMO tackles original, novel, and unconventional problems , while an AI is trained with huge amounts of problems and repeats the process. Nevertheless, the AI won the gold medal, perfectly solving five of the six problems and achieving 35 out of a possible 42 points, but these were not the best scores of the event.

Twenty-six students outperformed the machines. From the American team, which finished behind the Chinese, Alexander Wang stood out, a student from New Jersey who won his third consecutive gold medal, being, in the words of The Wall Street Journal, one of the most decorated young mathematicians of all time. The other was Qiao "Tiger" Zhang , who bravely faced the dreaded Problem 6, the toughest one this year. This last problem, a combinatorics problem, stumped both the AIs and 569 of the 630 contestants. Only six students completely solved it.

Zhang wasn't one of them, but his partial solution was worth more than the machines' complete lack of ideas. Nevertheless, Zhang himself predicted that next year the AI could achieve a perfect score. "The day an AI can solve Problem 6," Zhang said, "I would start to worry."

And when Thang Luong, leader of the Google DeepMind team, was asked if this will be the year that goes down as the last year humans surpass AI, “It could very well be,” he said.

The creators of these AIs celebrated their achievements. Google hailed Gemini's performance as a historic leap, not only because of the score it obtained, but because, unlike the previous year, the model was able to reason directly in natural language, without needing to translate the problems into code. It also met the test time, a feat that would have been unthinkable just a year ago. OpenAI, for its part, tested its model outside the official event, but with independent evaluators and the same problems. The result was the same: gold.

A few months before this competition , in May 2025, an unusual meeting took place on the campus of the University of California at Berkeley. Thirty of the world's most prestigious mathematicians gathered in a secret conclave, not to debate each other, but to face off against an artificial intelligence: o4-mini. A state-of-the-art language model developed by OpenAI, capable of reasoning with unprecedented speed and accuracy. They attempted to pose highly complex problems that would overwhelm the AI, but it displayed an unprecedented problem-solving ability. Participants remarked that it was like being in front of an extremely competent PhD student, "even more so."

Does this mean that these researchers were defeated where the young ones triumphed? The answer is no, because the context and problem-solving are different, but there is an interesting background.

If we ask Gemini itself, its conclusion is that in the IMO, the students beat the AI in solving problems designed for competitively trained individuals, using creativity, rigor, and mathematical intuition in a limited time. The IMO problems are designed to stimulate ingenuity and lateral thinking, not just brute logical force, and AIs still struggle with these types of problems that require non-obvious leaps of reasoning. But AI increasingly wins in formal, structured, or massively exploratory problems, such as those of pure research .

Accomplished mathematicians work to discover new theories, not just puzzles. The comparison shows that the path of human mathematics is not linear; you can excel by solving what others pose to you (IMO), and then you can try to create problems that no one has ever solved. So it's not that some lose and others win. It's that they are at different stages of the same process.

Far from seeing this evolution as a threat, some of the same humans who designed these machines see it as a new form of collaboration. “This AI is like a new calculator,” said Thang Luong, leader of the DeepMind team. “A tool that can take us further, not replace us.”

Zhang himself, despite having defeated AI, doesn't see it as a rival, but rather as an incentive to think better, more deeply, and more boldly. Former gold medalist and current DeepMind researcher Junehyuk Jung agrees: truly complex problems—like the famous Problem 6—will continue to challenge machines for years.

The mathematicians at the secret meeting pondered whether they might eventually become "question setters," guiding AIs toward new discoveries. And if so, they don't yet know what that would entail.

A few years ago, being brilliant in mathematics meant, among other things, being able to solve problems no one else could. Today, that's no longer enough. When machines can reach those same levels, the new realm of genius lies elsewhere: in the ability, above all, to formulate new questions. Because while AI can skillfully solve problems, it doesn't yet know what's worth asking.