Two things from the past week that fit together. Physical Intelligence opened a beta API: upload your data, fine-tune their models, and deploy hosted inference on your own robots. Their suggested loop is to collect demonstrations, train, test on the robot, then use the failures to decide what data to collect next. And at GTC Berlin NVIDIA released Human 1M, a dataset of human activities with synchronized three-view video and 3D hand tracking. The announcement doesn’t say how big it is. My take: models are becoming something you can rent, and the data is what you have to bring. @vangrid_io is on that side of it, with phone captures of real places.