An American student has developed an AI model that performs as well as the top participants in the International Mathematical Olympiad in the geometry category.
Trieu Trinh, a recent graduate from New York University, spent the last four years developing AlphaGeometry. This AI can solve various challenging geometry problems. According to Trinh, who published his work in the journal Nature, the program performs at the level of the best participants in the Mathematical Olympiad.
Trinh didn’t develop the system alone. From 2021 to 2023, he worked at Google to develop AlphaGeometry, specifically at DeepMind, Google’s AI development branch. DeepMind has previously developed programs such as AlphaGo and AlphaZero, which mastered the games of Go and chess, respectively.
‘Making a big leap’
However, solving mathematical problems is a tougher nut to crack. Trinh told the New York Times that he struggled for a long time to get AlphaGeometry working. Eventually, he succeeded. “We’re not making incremental improvements. We’re making a big leap, a significant breakthrough in terms of results,” he said in the American newspaper.
AlphaGeometry is now capable of solving random geometry questions from the Mathematical Olympiad. Out of 30 questions, the program could solve 25. For comparison, human winners of the Mathematical Olympiad manage to solve an average of 25.9 questions. The silver and bronze medals go to people who score an average of 22.9 and 19.3, respectively.
Human-like thinking
According to Trinh and the other authors of the paper, AlphaGeometry is ‘a remarkable milestone in automated reasoning at a human level.’
Other AI tools that gained popularity in the past year, such as OpenAI’s ChatGPT, are not capable of logical reasoning. A note is that AlphaGeometry was specifically developed for Olympiad questions and cannot replace mathematicians outright.