Make me think the next version of GPT will learn all the rubbish prompts people use and rewrite them to be more effective. The next version of RLHF. I guess they're already using ChatGPT data to do this like Google uses researches to improve the initial results.