Moreover, you only get an answer for a question which has already been posed and answered, hence is represented by a pattern which the large language model can

Moreover, you only get an answer for a question which has already been posed and answered, hence is represented by a pattern which the large language model can arrive at --- it's not possible to get an answer which wasn't already present in the training data.