A visual analogy for text generation by LLMs?