Comment by sshumaker

It’s how LLMs work - they are effectively recursive at inference time, after each token is sampled, you feed it back in. You will end up with the same model state (not including noise) as if that had been the original input prompt.