GitHub - karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically

autoresearch is an open-source framework by Andrej Karpathy that uses AI agents to autonomously run LLM pretraining experiments overnight on a single GPU. The agent modifies a single training file (train.py), runs 5-minute experiments, evaluates results via validation bits-per-byte (val_bpb), and iterates. The setup is intentionally minimal: one file to edit, one metric, one GPU. It builds on a simplified version of nanochat and supports any AI agent (Claude, Codex, etc.) via a program.md instruction file.