Interpreting GPT: the logit lens – LessWrong(2020) | Dark Hacker News