Jason argues physical AI is extremely hard, and pushing real-world performance is a matter of both software and hardware. If your hardware always breaks you can't run an AI model on the robot for 24 hours, so you have to do software-hardware co-design and co-iteration. In research the robots would break after a couple of hours or overheat after five or six hours.
More from this episode
Jason says Dino 1 is both the arm — the physical body — and the AI, the brain of it. The reason to build them in conjunction is that physical AI is extremely hard, especially if you're pushing real-world performance, which is "a matter of both the software and hardware." If your hardware always breaks, your robots aren't good enough and you can't run an AI model on the robot for 24 hours. He recalls that when he was doing research in the lab, the robots would often break after a couple of hours, or start overheating after five or six hours, so it was physically not possible to run for 24 hours. The conclusion from that simple thought exercise: to get any real-world performance you have to do software-hardware co-design and co-iteration. Interestingly, he notes that in the AI age the models are more intelligent and do more dexterous, smooth behavior, so they're much less likely to do unsafe behavior in the first place — which made hardware reliability a simpler problem than before.