"Summarizing Books with Human Feedback" (recursive GPT-3) — LessWrong