Autoresearch: Agents researching on single-GPU nanochat training automatically
Summary
Karpathy's autoresearch project demonstrates running autonomous AI agents that iteratively modify a minimal single-GPU nanochat training setup. The workflow centers on three core files (prepare.py, train.py, program.md) where agents edit train.py while humans iteratively refine program.md, enabling overnight, automated experimentation. The design uses a fixed 5-minute wall-clock budget and a single metric (val_bpb) to compare results across runs.