Demanding and Designing Aligned Cognitive Architectures — LessWrong