Karpathy's autoresearch turns a coding agent into an overnight ML researcher
Andrej Karpathy published autoresearch, a 630-line, single-GPU simplification of his nanochat training repo built around one idea: point a coding agent at a Markdown file instead of Python and let it experiment by itself.
Only three files matter — prepare.py with fixed constants and data prep that is never touched, train.py with the model and training loop that the agent edits, and program.md holding the human-authored instructions. Each cycle trains for exactly five minutes on a fixed wall-clock budget, checks val_bpb, then keeps or reverts. That works out to roughly 12 experiments an hour and about 100 overnight, with every experiment directly comparable because the time budget is constant.
WHY IT MATTERS
The clearest small-scale example of an agent harness running a real research loop unattended — a pattern you can copy for any experiment, measure, iterate task, not just LLM training.
