Answers grow one piece at a time

An LLM does not invent a whole paragraph in one flash. It builds the reply the same way Helper B chats: guess the next piece, add it, guess again.

That next piece is a token. The loop is: look at what you have so far → pick a likely next token → append → repeat.