Do we want to train the model during inference? I would think we want to be very particular about the data used to train - notably after OpenAI's notes on the subject.