Nonscientific: if roughly feels like as models get bigger they absorb more “wisdom” and that lowers the error-generation probability.