Pretty sure many latent spaces are not trained to represent 3D motions and some detailed physics of the real world. Those in pure text LLMs, for example.