That's a separate issue. That discussion is about training the model's responses through fairly standard RL (human-generated in this case), not providing it wit

That's a separate issue. That discussion is about training the model's responses through fairly standard RL (human-generated in this case), not providing it with a new corpus of data from which to draw its responses.