LLMs also don't work by generating probability distributions of the next word. Your explanation isn't able to explain why they can generate words, let alone sentences.