Scaling laws, which are empirical, are the closest existing theoretical tool for understanding AI improvement