Karpathy just built a self-improving AI research lab (in 630 lines of code)

Andrej Karpathy released a project called AutoResearch.

You give an AI agent a training setup, point it at the code, and let it run experiments overnight. It modifies the architecture, trains for 5 minutes, checks if the model improved, keeps or discards, and repeats. About 100 experiments while you sleep.

Shopify's CEO tried it on day one. Woke up to a smaller model that outperformed a larger one he'd built by hand.

19% improvement, zero human intervention.

Karpathy's README starts with a note about how AI research will eventually be done by "autonomous swarms of agents running across compute cluster megastructures in the skies."

He says this repo is how it all began.

I break down what this means, how it works, and why this might be the first real step toward recursive self-improvement.

your favorite “meat computer”,

Wes “can it autonomously improve ME” Roth