In this lecture from Stanford\'s CS231n, from the time 26:00 to 28:00 (approximately), it is saying that they will sample from the output (prediction) for the next input.
