The essence of a large language model is to forcibly construct a self-consistent value system based on existing input data. Hallucinations can be seen as a natural manifestation and extension after self-consistency. Many new scientific discoveries are precisely because they encounter an 'error' in the natural world that cannot be explained by existing theories and cannot be self-consistent, so they must abandon the old theories. This roughly explains why, so far, no large language model (with so much data) can spontaneously make new scientific discoveries, because the model itself does not have the ability to judge right from wrong.