The same “emergent capabilities” pattern we saw with large language models is now showing up in robotics. Physical Intelligence’s latest research shows their vision-language-action models suddenly start transferring knowledge from human GoPro footage to robot behavior once the model is large enough. Smaller models completely ignore the human data, but at scale, the representations align automatically. This is how training robots goes from “millions of dollars in specialized data” to “wear a came

Should Apple make iPhone smaller again?