Rethinking Interpretability in the Era of Large Language Models