What if we let chatGPT constantly talk to itself in between talks to humans and update his current system prompt. We spark random parts of past conversations of the day and let gpt talk to itself - reflecting. That mimics what happens in our head all the time.
At the end of the day all the system prompt is fed as a training input to the model itself with a higher weights than usual. That mimics our sleep process.
Updated model wakes up in the morning and starts all over.
What do you think?