What happens when AI helps you on a task? You usually use its outputs to do something, and if that works, you come back for further assistance. If it doesn't work, you come back to correct the model. Either way a signal from the real world gets captured in the chat logs.
When AI provides a response it is possible to judge that response in hindsight. You look at the next 20 messages or sessions from next days and judge based on what followed. The chat logs provide a way to do long range credit assignment.
When AI provides a response it is possible to judge that response in hindsight. You look at the next 20 messages or sessions from next days and judge based on what followed. The chat logs provide a way to do long range credit assignment.