“Current AI alignment techniques are developed iteratively and based on conceptual motivations, such as "evaluation is easier than generation," rather than a formal theory.”