Scaling Karpathy's Autoresearch: Claude Code Runs 910 ML Experiments on a GPU Cluster in 8 Hours

SkyPilot Blog·high signal

SkyPilot engineers gave Claude Code access to 16 GPUs on a Kubernetes cluster and ran ~910 experiments in 8 hours (176pts, 77 comments on HN), improving model validation loss from 1.003 to 0.974 (2.87%) at ~$300 GPU compute + $9 API cost — roughly 9x the throughput of sequential human-guided runs. Most striking: the agent autonomously discovered H200 GPUs completed 9% more training steps than H100s in the same budget and self-developed a two-tier screen-on-H100/validate-on-H200 strategy without being instructed. This directly extends Karpathy's single-GPU autoresearch paper to cluster-scale, validating the approach as practical and cheap.