Why Larger Language Models Are Few Shot