Recently there was an article about how some students at Stanford got ChatGPT to train their open source AI for a very low cost solution. I think you are right, except that it won't be human feedback, it will be AIs training other AIs, or somehow training themselves.