> I take it for granted that all these services are going to be free. They are a goldmine for behavioral and persuasion engineers. They are also a goldmine for

> I take it for granted that all these services are going to be free. They are a goldmine for behavioral and persuasion engineers. They are also a goldmine for LLMs. Training on human text is necessary for AIs but it has one major flaw - it is so called "off-policy". That means it portrays human behavior and human errors. While human-AI chat logs portray AI errors, so they are better material to generate training data than human text. Those LLM errors are usually corrected by the human, there is an implicit signal in there to improve the model. chatGPT is reportedly serving 10M customers and let's assume 10K tokens/month/user. Then it seems they collect ~1T tokens/month. In one year they have 12T tokens, while their original training set for GPT-4 was rumored to be 13T tokens. It's about the same size! I am expecting to see more discussion about LLM chat log datasets in the near future. What have they learned in one year from our interactions and explorations?