Large language models seem more useful for cutting down the number of "VP"s