What's amazing is that given how LLMs receive input data (as tokenized streams, as other commenters have pointed out) it's remarkable that it can ever answer this question correctly.