A simple example would be when models get better at following instructions, the frantic and somewhat insane-sounding exhortations required to get the crappier model to do what you want can cause the stronger model to be a bit too literal and inflexible.