Researchers set rules for automated AI cycles where models design, train, and test new models. Automatic code updates can outpace human review. Limit operations to 1,026 for verification checkpoints. Add human review and log compute before deploying new versions.