large token models are coming for everything, because everything can be made a token.

the detour via language here is not needed, these models can speak geometry more and more fluently

This seem like a great observation, a lot of negative reactions to AI generated data seem to come from limitations of using language, thereby denying good creative input from its users.

Right, like word2vec blew everyone's mind back in the day, but 3D models have always existed in a "vector space".