"Karpathy’s autoresearch agent automatically modifies training code, runs fixed-budget experiments, evaluates results, and retains or discards changes."
Karpathy's public autoresearch repository describes an experimental loop in which an agent edits a model-training script, trains for five minutes, checks validation bits-per-byte, and decides whether to retain the modification. The editable scope includes model architecture, optimizer, hyperparameters, and the training loop, establishing the actual range of actions available to the agent.
View Source