People are smarter than AI.

Google DeepMind and OpenAI won gold medals at the Mathematical Olympiad, but the American teenagers on the team still scored higher than these two machines.

The IMO 2025 was held from July 10-20 in Sunshine Coast, Queensland, Australia, with 630 participants. Of those, 67 (approximately 11%) won gold medals. The fact that AI also achieved a similar feat is a historic milestone.

This is the first year the IMO competition has officially partnered with several AI developers. The IMO judges have confirmed the results from the partner companies, including Google, and requested that they publish the results by July 28th.

However, the final results showed that even the most advanced and modern machines were defeated by the world's brightest teenagers.

The last wall of humanity?

The notoriously rigorous IMO exam is held over two days. Each day, students must solve three progressively more complex problems within just over four hours. The questions cover algebra, geometry, number theory, and combinatorics.

Because of its complex and unconventional problems, the annual math test has become a useful tool for measuring AI progress year after year. In this era of rapid development, leading research labs dream of the day their systems are powerful enough to qualify for an IMO gold medal.

Although this ambition was realized when Google and OpenAI's AI won gold medals, it's noteworthy that 26 students achieved higher scores.

Among them are four stars from the US team, including Qiao Zhang, a two-time gold medalist, and Alexander Wang, who brought home New Jersey's third consecutive gold medal. With this achievement, Wang has become one of the most outstanding young mathematicians of all time, and could still win another gold medal at the IMO next year.

The US team at the 2025 International Mathematical Olympiad. Photo: WSJ.

Speaking to the WSJ , Zhang said he spent the longest time on problem number 6 during the exam. This is a notoriously difficult combinatorial problem, encompassing a branch of mathematics involving counting, sorting, and combining discrete objects.

The solution to this problem requires ingenuity, creativity, and intuition that humans can utilize but machines cannot, or at least not yet. "I would actually be a little scared if AI models could solve Problem 6," Zhang said.

Problem number 6 stumped DeepMind and OpenAI's models, but it wasn't just AI that struggled. Of the 630 contestants, 569 also received a score of 0. Only 6 received a perfect score of 7 points. Zhang, proud of his partial solution, earned 4 points, more than most other contestants.

With its continuous advancements, many have suggested that AI may be "destroying" humanity in mathematics. However, Junehyuk Jung disagrees.

A former IMO gold medalist, Jung is currently an associate professor at Brown University and a visiting researcher at DeepMind. However, he doesn't believe this is humanity's final battle. According to Jung, problems like Problem 6 will continue to trouble AI for at least another decade.

"There are things AI will do very well. However, there will still be things that humans can do better," Jung remarked.

The gold medal race in mathematics

In 2021, a student named Alexander Wei was part of a research project that asked him to predict the mathematical capabilities of AI by July 2025. When considering other predictions, Wei thought this was too optimistic.

However, this former student is now living proof of just how wrong he was. Wei is the research scientist who led the IMO project for OpenAI. What's even more impressive is how they did it.

Not only OpenAI, but Google also called the results at IMO 2025 a major breakthrough. In 2024, AI models required problems to be translated into computer programming languages ​​for mathematical proof.

In just one year, these systems were operating entirely in "natural language" without any human intervention. DeepMind also completed the exam within the IMO's 4.5-hour time limit. For comparison, in 2024, this system took several days to calculate the solution.

Many people might think of AI as a competitor. However, those behind the models view them as complementary tools.

Final results of the 2025 International Mathematical Olympiad. Photo: WSJ.

"This could potentially be a new computer that will propel the next generation of mathematicians," said Dr. Luong Minh Thang, a senior researcher at Google (USA).

Three weeks ago, Dr. Minh Thang's team successfully developed AlphaGeometry 2. The most notable feature of this version is its ability to solve geometry problems that surpass those of gold medalists at the International Mathematical Olympiad (IMO).

"If students with IMO gold medal-winning skills can solve 40 out of 50 problems, then AlphaGeometry 2 will be able to solve 42 out of 50 problems , " Dr. Thang happily shared.

Last year, this model only won a silver medal. This year, Google is using the Gemini Deep Think multipurpose model (a version that was previously showcased at the developer conference in May).

Notably, according to the WSJ , before the competition began, Dr. Thang's team was still making adjustments. Initially, Dr. Thang only expected DeepMind's model to solve all three problems on the first day.

The simplicity, elegance, and remarkable readability of those solutions astonished the mathematicians. The next day, as soon as Thang and his colleagues discovered that the AI ​​system had solved two more problems, they realized that they could absolutely win the gold medal.

As a result, DeepMind's AI successfully solved 5 out of 6 problems. Notably, all of them used empirical reasoning models, processing mathematical concepts using natural language, a stark contrast to the complex approaches previously used by AI companies.

Source: https://znews.vn/nhung-nguoi-thong-minh-hon-ai-post1572367.html