>There is another emerging paradigm which is still small(er) scale but showing remarkable results. That's full multi-modal training with embodied agents (aka robots). 1x, Figure, Physical Intelligence, Tesla are all making rapid progress on functionality which is definitely beyond frontier LLMs because it is distinctly different.
Tesla is selling this view for almost a decade now in self-driving - how their car fleet feeding training data is going to make them leaders in the area. I don't find it convincing anymore