The International Mathematical Olympiad has been the graveyard of AI ambitions for years. Every July, the world’s best teenage mathematicians gather to solve problems that stump most adults with PhDs. And every year, the machines watched from the sidelines.
Not anymore. On July 25, 2024, two AI systems built by Google DeepMind — AlphaProof and AlphaGeometry 2 — scored high enough to earn a silver medal in the competition. That puts them ahead of the vast majority of human competitors, behind only the handful who take gold.
The stakes here are not academic. Mathematics is the language in which we describe physics, engineering, finance, and logistics. An AI that can reason through novel mathematical problems is an AI that can, in principle, design a stronger bridge, optimize a supply chain during a crisis, or spot a flaw in a cryptographic system before it is exploited. This is not about chatbots writing poems. This is about machines that can think their way through problems no one has taught them to solve.
AlphaProof and AlphaGeometry 2 are the products of a company that has been chasing this kind of reasoning for years. Google DeepMind was founded in the UK in 2010, bought by Google in 2014, and merged with Google’s Brain division in April 2023. It now runs research centers in the United States, Canada, France, Germany, and Switzerland. The company built its reputation on neural Turing machines — a 2014 innovation that lets neural networks access external memory, much like a person jotting notes while working through a hard problem.
That architecture matters. Traditional AI models are good at pattern matching. Show them enough cat photos and they learn to identify cats. But the International Mathematical Olympiad does not test pattern recognition. It tests the ability to construct proofs — logical chains that must be original, valid, and complete. A silver medal means these systems can do that, at least within a defined mathematical domain.
The competition itself is brutal. Hundreds of the world’s sharpest teenagers compete. Problems cover algebra, geometry, number theory, and combinatorics. Most contestants train for years. The fact that two AI systems reached silver standard in the same year suggests the pace of progress is accelerating.
What is genuinely at risk here is the boundary between human and machine capability. For decades, computers have outperformed humans in narrow, well-defined tasks — calculating, sorting, searching. But open-ended reasoning has remained human territory. That territory just shrank. If AI can reason mathematically at a silver-medal level, it is reasonable to ask what other cognitive territory might be next. Legal reasoning. Medical diagnosis. Strategic planning.
Google DeepMind has a history of making these questions urgent. The company has built AI that beats world champions at Go, predicts protein structures, and controls nuclear fusion reactors. Each breakthrough narrows the gap between what machines can do and what only humans could do before.
The silver medal is not a gold medal. The systems are not yet the best in the world. But the distance between silver and gold is small, and the distance between this year’s performance and next year’s may be smaller still. The International Mathematical Olympiad has been a benchmark for human intelligence for decades. That benchmark just moved.
