Autoresearch for macOS
Karpathy’s autonomous ML research agent, running natively on Apple Silicon.
Autoresearch is Andrej Karpathy’s experiment in fully autonomous machine learning research: an agent that reads a problem, forms hypotheses, runs training loops, measures results, and iterates. This project ports it to macOS so the whole loop runs locally on Apple Silicon.
The work was mostly in the performance layer: getting PyTorch to run well on the MPS backend, which is the Metal-based GPU path on Macs. With those optimizations, research cycles that would normally demand a cloud GPU reservation fit on a laptop or a Mac Studio.
Besides being useful, it’s a nice existence proof for a broader idea: you don’t always need the biggest compute. A specialized setup, matched tightly to the hardware it runs on, goes a long way.