From analyzing cats to Erdős' theorems: artificial intelligence increasingly tackles the pinnacles of mathematics

AI models are moving from humanities tasks to solving complex mathematics

*Currently most artificial intelligences were originally developed for working with text and images, but their developers increasingly recognize the potential to apply them in mathematics. This opens two important directions:*

1. Scientific progress – new models allow quickly finding solutions that were previously considered unsolvable.

2. Demonstration of AI capabilities – successes in mathematics serve as a vivid proof of technology effectiveness.

Examples of successes
- A Cambridge student used an OpenAI model and solved the Erdős problem, previously deemed inaccessible.

- Models show high results at the International Mathematical Olympiad and other specialized contests.

- Former board member Helen Tuner notes: “We have already moved beyond simple tasks like distinguishing cats from dogs; now AI solves high-level problems.”

Specialized developments
Company Model Task DeepMind (Google) AlphaProof Mathematics DeepMind (Google) AlphaGeometry Geometry
These models received recognition on Epoch AI benchmarks, which measure speed and accuracy of solutions. Initially large language models were considered unsuitable because they generate text “based on probability” and often “hallucinate.” However, the introduction of reinforcement learning and reasoning architectures has significantly increased their reliability.

Strengthening the scientific team
OpenAI attracted two outstanding mathematicians:

- Ernest Ryu – University of California, Los Angeles
- Mehtaab Sawhney – Columbia University

These specialists help improve models and their abilities to solve complex tasks.

Mathematics as a “verifiable” test
Mathematical proofs can be automatically verified, making it an ideal field for AI experiments. This also promotes software development:

- Anthropic invests in Claude Code – an assistant that generates programming code.

What’s next?
To tackle truly difficult scientific questions, AI must rely on existing results and not count on a single “intermediate” session. Currently models can effectively aggregate information from different disciplines, accelerating the discovery of new ideas. Experts are confident: in the near future this will become a key driver of scientific progress.

> *In mathematics, AI has already proven its effectiveness.*