On the Semantics of Large Language Models