DeepMind AI Achieves First Scientific Discovery Using Large Language Model
DeepMind's FunSearch system uses an LLM paired with an automated evaluator to produce a verifiable new solution to the decades-old cap set puzzle, marking the first scientific discovery made by an AI language model.
DeepMind introduces FunSearch, a system that pairs a large language model with an automated evaluator to make a verifiable scientific discovery for the first time. The AI successfully finds new constructions for large cap sets, pushing the boundaries of the decades-old cap set puzzle further than human researchers ever achieve. While it does not completely solve the mathematical conundrum, it generates genuinely new and verifiable knowledge.
Unlike previous experiments where AI merely solves math problems with known answers, FunSearch uses a continuous back-and-forth process between Google's PaLM 2 model and a fact-checking evaluator. This built-in safeguard effectively prevents the AI from hallucinating false information, which historically limits the usefulness of language models in strict scientific research.
The tool also stands out because it outputs the actual computer programs used to construct its solutions rather than just providing the final answers. DeepMind researchers hope this transparency inspires scientists to build upon the AI's logic, driving a continuous cycle of human improvement and AI discovery.