For two years the consensus held that physical intelligence could only be learned from real interaction data. Simulation hit a wall: as Sergey Levine argued, stronger models get better at detecting the seams in surrogate data, so the skills that transfer to reality keep shrinking. Real data was the only path, and real data is slow, expensive, and scarce. That framing is breaking, and two things changed at once. World models now learn “what happens next” from internet-scale video, giving robot...