My favorite commercial airplane, and the one I keep reading about. I built a whole page for it: an annotated plan view you can pick apart, and a wing you can bend to the 150% load it was certified against.
Language models had the internet to learn from. Robots have nothing close, which is why most of them still only do the one task they were built for. π0 takes the other route: start from a vision-language model that already understands the world, add continuous motor control through flow matching, and train it across eight different robot platforms, so that teaching it a new task takes only a modest amount of data.
The part that pulls me in is that physical AI removes the safety net. A bad sentence costs nothing. A bad motor command at 50 Hz breaks something. I want to end up working on models that have to be right about the world, not just convincing about it.