Choosing Rust for LLM-generated code (runmat.org)

A recent experiment building RunMat—a MATLAB-compatible runtime and compiler—demonstrates the unique advantages of using Rust for large language model (LLM)-generated code. Unlike TypeScript’s large but noisy and heterogeneous codebase, Rust’s more uniform and tightly curated public corpus, combined with its strong type system and advanced linting tools, creates a feedback-rich environment that helps LLMs produce higher-quality, idiomatic, and correct code. Rust’s strict compile-time checks and fast linting loops act as an immediate, granular signal to LLMs, accelerating convergence on working solutions and reducing error rates compared to more permissive languages. Beyond language ecosystem fit, Rust also offered practical engineering benefits: memory safety through ownership, modular crate-based architecture, straightforward cross-platform support with Cranelift JIT, and clean GPU acceleration abstractions via WGPU. These features contributed to a remarkably rapid development process—accomplishing in weeks what traditionally could take years with a large engineering team—all while delivering a runtime that benchmarks dramatically faster than GNU Octave across core MATLAB-style operations on Apple M2 hardware. The project highlights a paradigm shift in software development where language choice is increasingly about synergy with LLM-driven workflows rather than just developer availability or raw performance. Rust’s combination of consistent training data, strong compiler feedback, and rich tooling makes it an ideal substrate for AI-assisted programming, allowing developers to steer, curate, and iterate rapidly. RunMat’s ongoing evolution will continue to explore how LLMs can reshape modern runtime and scientific computing development.