From what I understand, these LLMs deliberately incorporate randomness. So it's not surprising at all that some of their answers would be imperfect. Something

From what I understand, these LLMs deliberately incorporate randomness. So it's not surprising at all that some of their answers would be imperfect. Something else to consider is that these models work on tokens, not characters, so that might also account for why it made a mistake.