Overview

• DeepMind's large language model (LLM), FunSearch, has generated a novel solution to the decades-old 'cap set puzzle', a complex mathematical problem.
• FunSearch combines a pre-trained LLM with an automated evaluator, which checks the LLM's outputs for accuracy, reducing the risk of false information.
• The LLM and evaluator work in a continuous cycle, refining initial solutions into new knowledge.
• FunSearch's solutions are presented as programs, revealing how the solutions are constructed.
• The researchers believe FunSearch could inspire further insights in the scientific community, driving a cycle of improvement and discovery.