Impossible Triangle: What's Next for Pre-trained Language Models?