The question is about discovery. If an LLM (trained purely by supervised learning to mimic its input) is fed talk of discovery, then it will talk more about discovery, but it won't do discovery. Perhaps we can agree on the answer to that limited question? Anyway, I hope we are getting closer to the question. It requires a little bit of nuance. Markus Buehler and I were careful not to claim a limitation of LLMs in general as they are not even a well defined concept.