Especially because a lot of its training data would have included videos/images which don't adhere to a physics model that's even internally consistent within t

Especially because a lot of its training data would have included videos/images which don't adhere to a physics model that's even internally consistent within the video, let alone representative of real-world physics. Sora is not just for photorealistic video generation, but also cartoons, animation, and fantastical nonsense. I suspect if you trained a similar model on the video feeds of every Wyze / Ring / Arlo / Nest / Blink camera, both indoor and outdoor ... then you'd eventually end up with a casually usable physics simulator that works well in most common situations. Could be a very useful model for home robotics. I don't think it's the most compute-efficient way to get "a casually usable physics simulator that works well in most common situations" ... but I do think it would work, and might have some advantages for certain applications (like home robotics).