We think methods like they are promising simply because language models presently find out a good deal about human values for the duration of pretraining. Learning about human values is not really as opposed to learning about other topics, and we must always assume larger models to have a additional correct image of human values and to uncover them