A robotics dataset with 10 million nearly identical demonstrations may be less useful than it sounds. Scale matters, but diversity determines what that scale actually represents. This is why the Task Generation Engine inside Axis Robotics caught my attention. Instead of leaving diversity to chance, Axis describes a system that can vary multiple dimensions of a task, including: • robot embodiments • objects • spatial configurations • semantic conditions • visual environments Think about something as simple as picking up a cup. Changing the cup, its position, lighting, surrounding objects, robot configuration, or environment can turn one basic skill into many different learning situations. That matters because real homes and workplaces are not standardized simulation labs. A robot that only succeeds under familiar conditions has not really learned to generalize. So when I look at @axisrobotics, I am increasingly less interested in asking: "How many trajectories can they generate?" The better question might be: "How much meaningful diversity can they create inside those trajectories?" For Physical AI, millions of examples are useful. Millions of genuinely different learning situations could be far more valuable.