autoresearch
Karpathy's overnight research loop: one editable file, a frozen evaluator, one number. Only changes that beat the best result are kept.
How each pass goes
- Give the agent one file it may edit and an evaluator it may not.
- It proposes a change and runs the evaluator.
- Better score: keep. Worse: revert. Repeat all night.
Stops when
The time or run budget is spent; you wake up to the best version.
Starter prompt
You may edit only train.py. evaluate.py is frozen and prints one score (higher is better). Loop: make one change, run the evaluator, keep the change only if the score beats the best so far, otherwise revert. Log every attempt and score to results.md. Run until 100 attempts.