Andrej Karpathy's autoresearch project — an autonomous AI agent that iteratively modifies and trains a GPT script — was deployed on Red Hat OpenShift AI with H100 GPUs and left to run unsupervised for 24 hours. The setup involved containerizing the project using Red Hat AI base images with PyTorch/CUDA, deploying via Kubernetes manifests, and using Claude Code Opus as the agent. Over 198 experiments, the agent achieved a 2.3% improvement in validation loss, discovering that smaller batch sizes maximize steps per training window, wider MLPs outperform deeper ones, and value embedding regularization provides late-run gains. A CUDA driver mismatch issue with OpenShift AI v3.4.0 is documented with a one-line fix. The full deployment code is publicly available on GitHub.
Table of contents
From bare metal to oc applyH100 vs. A100: same cluster, different nodeSelectorWhat the agent discovered in 24 hoursThere's a catchTry it yourselfSort: