ML Reading #31: LLM-based Evolutionary Code Agent for Mathematics and Algorithmic Discovery
AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms - Google DeepMind
Join us at Mox to discuss AlphaEvolve, Google DeepMind’s new evolutionary coding agent designed to automate scientific and algorithmic discovery. Expanding significantly on the capabilities of FunSearch, AlphaEvolve orchestrates a pipeline of LLMs to iteratively evolve entire codebases through automated feedback loops.We will explore how AlphaEvolve helped discovering a rank-48 algorithm for a 4x4 matrix multiplication, the first improvement over Strassen’s algorithm in 56 years, and establishing new bounds for open mathematical problems including analysis, geometry, and combinatorics.
Beyond theory, we will examine its practical deployment in optimizing Google’s data center scheduling and hardware accelerators, to address the future of self-improving code and AI-driven science and their implications.
Discussion at 20:00, (optional) quiet reading from 19:00.