DeepMind’s AI finds new solution to decades-old math puzzle

DeepMind’s language model, FunSearch, has made a significant breakthrough by discovering a novel solution to a challenging math problem known as the “cap set puzzle.” This achievement could potentially signal a new era in AI development.

FunSearch excelled in finding new constructions for large cap sets, surpassing the best-known solutions to date. While it did not solve the cap set problem definitively, it contributed to a new piece of verifiable knowledge about this scientific problem, representing the first scientific discovery using a large language model.

The use of an automated fact-checker within FunSearch differentiates it from traditional large language models, making it a promising tool for scientists. The continuous interaction between the language model and the fact-checker not only produces solutions but also reveals the construction process, potentially inspiring further insights and discoveries in scientific research.