Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster

Researchers gave Claude Code access to 16 GPUs on a Kubernetes cluster to scale Karpathy's autoresearch project, which uses a coding agent to autonomously improve neural network training. Over 8 hours it ran ~910 experiments in parallel, discovering that model width scaling mattered most and learning to exploit heterogeneous hardware (H100s for screening, H200s for validation), achieving a 2.87% improvement in validation loss over baseline.