Writing Rust code that's faster than state-of-the-art libraries by asking agents to make the code faster (minimaxir.com)

In a recent exploration of language models' capabilities, a researcher discovered that modern agentic LLMs could generate Rust code outperforming state-of-the-art libraries. By prompting the model to iteratively optimize Rust implementations of machine learning algorithms, significant performance enhancements were achieved—reportedly increasing execution speed by factors of 2 to 20 times, depending on the algorithm and context. Initially, the focus was on the UMAP dimensionality reduction algorithm, with optimizations including the use of SIMD operations and function fusions to enhance speed without resorting to unsafe Rust code practices. This work is particularly significant for the AI/ML community as it not only showcases the potential of LLMs to generate high-performance code but also opens avenues for creating efficient, scalable libraries for machine learning in Rust—often recognized for its speed and memory safety. The resulting code was shown to outperform existing implementations, and with further trials on various algorithms, it was demonstrated that the agent could maintain performance improvements while also enhancing code quality. This iterative optimization process, dubbed "benchmaxxing," suggests that with robust prompting and constraints, LLMs can effectively bridge the gap between algorithmic performance and practical application in software development.