Potemkin Understanding in Large Language Models