Using Local LLMs to Discover High-Performance Algorithms

This experiment details using local large language models (LLMs) on a MacBook to discover and optimize high-performance algorithms. The specific goal is to generate a more efficient, native Rust implementation for matrix multiplication, a core operation in AI. The process utilizes a multi-agent system orchestrated by Microsoft Autogen, where agents propose, verify, code, and test new algorithms. This system is informed by a vector database created from academic papers on matrix multiplication, which are processed using semantic chunking to improve contextual accuracy.