Researchers use control theory to shed unnecessary complexity from AI models during training, cutting compute costs without sacrificing performance.