Comment by bytefactory

This doesn't seem like a major difference, since LLMs are also choosing from a probability distribution of tokens for the most likely one, which is why they respond a token at a time. They can't "write out' the entire text at a time, which is why fascinating methods like "think step by step" work at all.