Because of it's unpredictability and massive dependence on the training data, when LLMs start hallucinating most of the time the only fix these "engineers" have is to feed it another LLM... The genius was the transformer architecture, and evidently none of us have a damn clue how it works