Autoresearch

Most AI research still depends on a person proposing each change, running it, reading the result and deciding what to try next. Andrej Karpathy's Autoresearch turns that loop over to an AI coding agent. The agent edits one training file, runs a fixed five-minute experiment, compares a validation score, keeps improvements and discards regressions. It is a compact demonstration of what autonomous research can look like, not a general-purpose training service.

Exploring autonomous AI research, teaching experiment design, or testing whether a coding agent can improve a small language model through repeated measured changes. The official version needs Python 3.10 or newer, uv and one NVIDIA GPU; it was tested by the project on an H100. Mac, Windows and AMD users need community forks linked from its README.

1. Clone the repository into a fresh folder, then read its README and program.md before giving an agent control. The README labels the project MIT, although the repository does not currently include a standalone licence file.
2. Run uv sync, followed by uv run prepare.py. This downloads training data from Hugging Face into your local cache and trains a tokenizer.
3. Run uv run train.py once to establish a baseline. Then use a disposable clone on its own branch, grant Claude Code, Codex or another coding agent only the permissions it needs, point it to program.md, and review every commit and result before keeping anything.