Google DeepMind AI Discovers Solution to Long-Standing Math Problem
Google DeepMind introduces FunSearch, a tool that pairs a large language model with evaluative systems to uncover a new mathematical solution. This marks the first time a large language model produces verifiable, previously unknown scientific information.
Google DeepMind uses a new AI tool called FunSearch to solve a famous unsolved problem in pure mathematics. By pairing a code-focused large language model named Codey with systems that filter out bad answers, the researchers create a method that generates entirely new and verifiable information. This achievement marks the first time a large language model successfully discovers a solution to a long-standing scientific puzzle.
FunSearch takes a different approach compared to previous DeepMind tools like AlphaTensor and AlphaDev, which treat math problems like games of chess or Go and remain limited to specific tasks. Instead, FunSearch prompts the language model to write computer code and automatically discards the majority of nonsensical or incorrect outputs. The system keeps only the successful pieces of code and feeds them back into the process to guide future attempts.
The researchers start by outlining the target problem in Python and allow the AI to iteratively search for mathematical functions. Even though the team admits they do not fully understand why the process works so well, the resulting gold among the garbage proves that large language models can do more than just hallucinate facts. This breakthrough opens up new possibilities for using AI to drive future discoveries in fundamental science.