Researchers from the California Technological Institute have developed a new model of artificial intelligence, capable of finding solutions that require millions of steps. These algorithms can not only promote mathematics, but also help in predicting rare, but catastrophic events, such as hurricanes and financial crises.
Interest in the use of AI in mathematics is growing. For example, AlphaProof from Google Deepmind showed the level of the silver medalist of the 2024 International Mathematical Olympiad, and O3 from Openai demonstrated strong results in mathematics, natural sciences and programming.
But researchers from the California Technological Institute went further, trying to solve problems that have remained open to professional mathematicians for decades. They compare their approach to chess: if standard mathematical evidence requires 30-40 steps, then their new system solves problems that require thousands and sometimes millions of steps.
One of the areas of work was the Andrews -Curtis hypothesis – a difficult task of combinatorial theory of groups proposed 60 years ago. The researchers have not yet been able to prove the hypothesis itself, but their algorithms have refuted a number of potential counter -frauders that remained unresolved for 25 years.
Unlike existing models, such as ChatGPT or O3, which cope well with typical solutions, the new AI is looking for unexpected, complex ways. To do this, the team used the training method with reinforcement, starting with simple tasks and gradually complicating them.
What is especially important, researchers have developed new algorithms that can be used not only in mathematics, but also in other areas. They are able to identify rare and abnormal events, the so -called “black swans”, which can have destructive consequences. This opens up prospects for forecasting natural disasters, financial crises and other rare phenomena.
The results of the work of the team Published February 13 on the ARXIV platform, and now scientists continue to explore other unresolved mathematical problems. Perhaps their methods will once help solve the tasks of the level of the millennium.