A language model doesn’t write a word; it gives every word a probability, and something has to pick one. Temperature reshapes the probabilities, top-k and top-p cut off the unlikely tail, and a random number makes the choice. Repeat, one token at a time.