# autoresearch-mlx

An Apple Silicon port of Karpathy's autoresearch project, running fixed-time autonomous training-experiment loops natively through MLX with no PyTorch or CUDA dependency.

autoresearch-mlx is an Apple Silicon port of Karpathy's autoresearch, an autonomous research loop controlled through a program.md protocol file. It keeps the same basic rules as the original: one mutable training script, one tracked metric, a fixed five-minute training budget per experiment, and a keep-or-revert decision made through git after each run, but it runs natively on Apple Silicon through MLX instead of requiring PyTorch or CUDA.

A coding agent can be pointed at the protocol file to run the loop: editing the training script, running a fixed-budget experiment, reading the result metric, and keeping or reverting the change before repeating. The project's own results log shows the pattern in action: under a fixed wall-clock budget, smaller, faster-training model configurations can beat larger ones simply by fitting more optimizer steps into the same budget, with longer overnight runs on different Apple Silicon machines converging on different winning configurations.

It requires an Apple Silicon Mac, Python 3.10 or newer, and uv.

Website: https://github.com/trevin-creator/autoresearch-mlx