Physical AI makes the case for adaptable models more concrete than a generic open-versus-closed debate. NVIDIA’s Ming-Yu Liu explains on Practical AI that robots have different cameras, sensors, action spaces, latency limits, and power budgets; a cloud API cannot always supply the control or customization the device needs. He separates policy models that choose actions from world representations that support prediction and explanation, and contrasts classical equation-based simulation with learned approximations. The operational payoff he describes is a faster development loop: compare candidate driving or robot policies in simulation, narrow the hypotheses, and still test the promising ones in the real world. That last step matters—plausible generated futures are not certificates of physical correctness. Liu also explains why an embedded batch-size-one workload cannot assume the utilization advantages of a shared service batching many users, making intelligence per watt and architecture choices central. His description of Cosmos and NVIDIA’s open resources remains a vendor account, not an independently tested release matrix. So does the ambitious factory scenario of specialized agents coordinating inventory, transport, assembly, and inspection: he explicitly says such a future could be decades away and stresses unresolved verification and hardware limits. The strongest takeaway is bounded and useful: model access can improve adaptation and experiment selection, but neither openness nor simulation removes the obligation to measure what happens outside the model.
“Innovation, freedom.” — Ming-Yu Liu, 06:51
“One is rule based, but the other is more like a data driven approximation.” — Ming-Yu Liu, 28:43