Google DeepMind has developed FunSearch, a large language model designed to generate provably correct solutions to complex mathematical problems rather than plausible but inaccurate ones. Using a self-checking mechanism that scores and refines outputs, FunSearch produced a previously unknown solution to the cap set problem and found a more efficient algorithm for the bin packing problem. The system runs a version of Google's PaLM 2 model and generates computer code rather than raw mathematical outputs, making results easier to verify. The breakthrough points to a new role for LLMs in advancing formal mathematical research.