A technical paper quietly published in mid-December has sent ripples through the mathematical community with an unprecedented claim: a large language model has produced a genuine discovery in pure mathematics. The achievement, described in a December 14 paper, marks a departure from the typical limitations of such models, which generally excel at generating plausible-sounding text but struggle with factual accuracy.
The system, named FunSearch, was developed by Google DeepMind, the London-based artificial intelligence laboratory acquired by Google in 2014. In April 2023, DeepMind merged with Google’s Brain division. FunSearch tackled problems in combinatorics and coding theory, areas where human researchers had reached impasses. Rather than simply providing answers, the model generated code to create new mathematical constructions, evaluated those constructions, and refined its own code in a continuous feedback loop.
This approach differs fundamentally from how large language models typically operate. Most such models function as predictors, determining the next word based on previously seen data. FunSearch, however, was tasked with writing a program to solve a specific mathematical problem. When the program failed, FunSearch rewrote it. This iterative process continued until the program succeeded, producing a solution unknown to the researchers who had set the problem.
Google DeepMind has pursued this capability for years. In 2014, the company introduced neural Turing machines, which allowed neural networks to access external memory similarly to conventional computers. This was an early step toward models capable of reasoning rather than repetition. FunSearch represents the next phase, combining a large language model with an automated evaluator that checks every output for correctness. The language model generates ideas while the evaluator eliminates incorrect ones, enabling the system to learn.
The problems FunSearch solved are not trivial. The cap set problem and the problem of constructing large sets of vectors with low inner products have been studied for decades and have practical applications in data storage and error correction. FunSearch produced solutions superior to any known human result, with verifiable proofs.
Google DeepMind now operates research centers across the United States, Canada, France, Germany, and Switzerland as a subsidiary of Alphabet Inc. Company leadership has stated that the ultimate goal is artificial general intelligence—a machine capable of performing any intellectual task a human can. FunSearch contributes to this objective by demonstrating that a machine can generate new knowledge rather than merely retrieve existing information.
The paper represents a single result on two problems in one narrow field. However, the method is potentially generalizable. If FunSearch can be applied to other mathematical areas such as number theory, topology, or dynamical systems, the pace of discovery could accelerate significantly. A machine that identifies patterns no human has observed would fundamentally change the role of mathematicians. The significance of the December 14 paper lies not in an AI winning a game or composing a poem, but in its discovery of something true that was previously unknown.
