Jim Fan, Nvidia: On foundation models for embodied agents and scaling data

So Jim is raising the argument we should train models via videos of the task as a superior route to training, makes sense given Nvidia has a robot warehouse simulator and wants to sell (pickaxe's) to business.