Karpathy's autoresearch repository introduces autonomous AI agents that conduct machine learning research overnight, eliminating the need for human intervention in experimental iteration. The system runs on a single NVIDIA GPU, where agents modify training code, execute 5-minute experiments, and evaluate results automatically. This approach transforms traditional research workflows by letting AI agents handle the tedious trial-and-error process.

nanochat training forms the foundation, with the repository containing just three key files: prepare.py for data preparation, train.py for the model and training loop, and program.md for agent instructions. The fixed 5-minute time budget ensures experiments remain comparable regardless of architectural changes, while the validation bits per byte metric provides a fair comparison across different model configurations. This self-contained system requires only PyTorch and basic dependencies.

The project represents a significant shift in how AI research might be conducted, with agents potentially discovering optimizations that human researchers might overlook. By automating the experimental process, researchers can wake up to dozens of completed trials and potentially better models, dramatically accelerating the pace of development. The repository's simplicity and focus on single-GPU training make it accessible while demonstrating the potential for autonomous research workflows.