OpenAI could of course send all your chat history to a DB and then use that on a future training round. Some of it is free RLHF data that needs a bit of cleaning up. Because the user side of the conversation has clues about whether the answer was correct.

Also are we that far from fine tuning in real time? Maybe next year?

I think while the current model might be what he says the model in 2025 could be vastly different.