I wonder when we'll get to a point that it'll be against the OpenAI TOS to 'fine-tune' your own model with their model's outputs since it really lets you cut down on the hardware requirements to bring up a competitor.