dbl001 comments on interpreting GPT: the logit lens