Comment by jshmrsn

It’s not the LLMs themselves, it’s the training data that they are trained on being supplemented by synthetic video from simulations / 3D rendering.