Assuming the task remains just generating tokens, what sort of reasoning or planning would say is the threshold, before it's no longer "just a form of next token prediction?"