Er, then what is the "already trained" model? I thought pre-training was the gradient descent through the internet part of building foundational models.