Educational CAI: Aligning a Language Model with Pedagogical Theories — LessWrong