Why can't a next token prediction machine not predict a train of reasoning?