A visual analogy for text generation by LLMs? — LessWrong