AI agents running research on single-GPU nanochat training automatically
Browse cluster: Cluster 93 →autoresearch is an autonomous AI research system that allows an AI agent to experiment with language model training overnight on a single GPU. The agent iteratively modifies the training code, runs experiments within a fixed 5-minute time budget, evaluates results using a validation metric (bits per byte), and automatically keeps or discards changes. The project is built in Python with PyTorch, implementing a simplified single-GPU version of nanochat that includes a GPT model, optimizer, and training loop. Rather than directly editing Python files, researchers provide high-level instructions through a `program.md` Markdown file that guides the autonomous agent, allowing the system to conduct approximately 100 experiments unattended while a human sleeps.