In this issue:

• An anaysis of Google DeepMind’s FunSearch, a model that was able to discover new algorithms in computer science and mathematics.

Discovering new science might be most complete Turing Test for the AI models. New scientific methods require complex reasoning skills, combining knowledge from many fields, constant experimentation and evaluation, and many other complex cognitive skills. Google DeepMind has been one of the AI labs pushing the frontiers of using AI to streamline our path to new scientific discoveries. Models such as AlphaGo has enabled the discovery of new proteins while AlphaTensor was able to improve classic matrix multiplication algorithms. Google DeepMind’s newest iteration in this area is FunSearch, a model that was able to create new mathematics and computer science algorithms.

FunSearch provides a clever approach to discover new algorithms by “thinking in code”. Essentially, FunSearch uses an LLM to generate computer programs based on a set of functions for a given problem and then uses an evaluator to prove the different solutions. The FunSearch named is derived from the fact that the model iteratively searches the function space.

Inside FunSearch