From everything that I’ve read the answer to “what happened?”—whether the AI did a great job or a terrible job—is “No one knows.” And the people who do have answers made them up the same way people 2000 years ago explained eclipses.
My guess: That is the actual difference in the code that generated those models. 