The Smallness of Large Language Models