porby comments on Pretraining Language Models with Human Preferences