Implications of GPT-2 — LessWrong