karpathy/autoresearch ? reverse-engineered prompt

Hey, I want to set up that 'autoresearch' project from Andrej Karpathy. The big idea is to have an AI agent that can run its own research experiments on a language model, basically working on its own overnight.

It should be able to modify the Python training code itself, run a quick test for a fixed 5-minute time budget, check if the results got better, and then just keep iterating automatically. My only job should be editing a simple markdown file to give the agent its high-level goals and instructions.

Can you build out this setup for me? It's all based on Python and needs a single NVIDIA GPU. Start by installing the dependencies—I think it uses a tool called uv—and then run the data preparation script to get everything ready. The end goal is a system where I can just point the agent to my instructions and watch it go.