which apparently hard to come up with by compiling lots of text into a statistical model for what text is most likely to come after your question