It’s infuriatingly hard to understand how closed models train on their input