Looks like end to end FSD likely models future actions rather than future frames. And by focusing on predicting something less noise and more signal, it builds a good world model for a physical world generally intelligent agent.

→ View original post on X — @aravsrinivas