Deep learning curriculum for large language model alignment — LessWrong