Agreed. There is both explicit and implicit feedback in user responses. Did they continue or stop? Clarification? Pushing back? Correction? so many signals. Eve

Agreed. There is both explicit and implicit feedback in user responses. Did they continue or stop? Clarification? Pushing back? Correction? so many signals. Even more, you can judge a response by looking at how the conversation went after that moment. Hindsight analysis to predict their usefulness. The more interesting thing is when the user tests the AI ideas. Like running some code or applying advice for your hobbies. The user returns again and again communicating the outcomes in order to get new advice. The model can collect all chats across many days by topic, and judge them in retrospective. Which answers were good or bad? The user base is so large, 200M users means lots of perspectives, lots of ideas, chatGPT puts trillions of tokens in human brains per month. They influence on a grand scale. After a while some article will describe the outcomes, and that data percolates back in the next training set. A feedback loop. And an indirect agent, working through people. An experience flywheel, collecting from hundreds of millions and serving back.