The use of FunSearch represents a shift from DeepMind's previous tools, AlphaTensor and AlphaDev, which solved math problems by treating them as puzzles in games like Go or chess. Unlike these tools, FunSearch can theoretically be used to find solutions to a wide range of problems as it produces code, a recipe for generating the solution, rather than the solution itself. The results are also easier to understand, making it a promising paradigm for leveraging the power of large language models in scientific research.
Key takeaways:
- Google DeepMind has used a large language model called FunSearch to solve a famous unsolved problem in pure mathematics, marking the first time a large language model has been used to discover a solution to a long-standing scientific puzzle.
- FunSearch, which is built on top of DeepMind’s game-playing AI AlphaZero, combines a large language model called Codey with other systems that reject incorrect or nonsensical answers and plug good ones back in.
- FunSearch was able to come up with code that produced a correct and previously unknown solution to the cap set problem, a complex mathematical problem.
- FunSearch also demonstrated its versatility by approaching another hard problem in math: the bin packing problem, and came up with a solution faster than human-devised ones.