Do Large Language Models Know How Much They Know?