OpenEvolve AI coding agent built a better algorithm

UC Berkeley researchers used OpenEvolve, an open-source AI coding agent, to optimize a load balancing algorithm for large language models, achieving a 5x speedup over existing implementations. The AI-driven approach replaced inefficient loops with vectorized tensor operations and a zig-zag partitioning scheme, reducing runtime from 19.6ms to 3.7ms at a cost of less than $10. This demonstrates AI's potential for algorithmic discovery and optimization in systems research, with researchers predicting widespread adoption for performance tuning in production systems.