autoresearch

autoresearch is an experiment published by Andrej Karpathy (GitHub: karpathy). In the README's words, the idea is to "give an AI agent a small but real LLM training setup and let it experiment autonomously overnight." In each round the agent edits train.py, trains for 5 minutes, evaluates the result and decides whether to keep or discard the change. You steer the research by editing program.md. The project is a simplified version of Karpathy's nanochat. The macOS, AMD and other versions mentioned in the README are forks by other authors, separate from this repository.

Who it's for:Researchers and developers with an NVIDIA GPU who want to watch an AI agent run its own model-training experiments. It is of no use if you don't write code.

Before you start: You need one NVIDIA GPU (the author ran it on an H100), Python 3.10 or later, and uv. CPU and Apple MPS aren't supported in the main repository. A prepare step downloads training data first. The README suggests turning off all permission prompts in Claude or Codex so the agent can keep editing files and running commands. We don't recommend this: once prompts are off, the agent runs every command and changes every file without asking you. GPU compute and the model calls behind the agent can both cost money, so estimate the bill for an overnight run yourself. If you use a cloud model, the relevant code and logs are sent to that provider. On licensing, the README says MIT, but no separate LICENSE file was found in the repository.

Based on the project's public documentation. Not tested by AI Inspiration World.