Two Small Experiments on GPT-2 — LessWrong