10x the parameters, but validation is still done by people

After watching XPeng's second-gen VLA sharing session, what stuck with me was that line: "AI is starting to understand time." Sounds mystical, but what it actually means is that the model is starting to take temporal relationships seriously—and for inference scheduling on the domain controller, that's a real change.

But another line caught my attention more. They said themselves that doing physical AI well isn't just about the model itself—you also have to solve two problems: first, how to efficiently use real-world data; second, how to efficiently validate the model's behavior in the real world. That's a pretty honest thing to say. Parameters are up tenfold over the previous generation, model capability is climbing, but the validation workload is climbing right along with it.

I feel this deeply when testing domain controllers. The more end-to-end the model looks, the fewer observable interfaces there are in between—when something goes wrong you can only fall back on replay and scenario re-runs, and a whole round gets eaten up just constructing edge cases. They say this is the first mass-produced physical-world large model, capable of action generation as well as understanding and reasoning. The word "mass-produced" carries weight—it means this thing has to hold up against the long tail under automotive-grade conditions. Scaling laws keep working on VLA, I believe that, but scaling laws don't tell you which category of scenarios you haven't covered yet.

The bottleneck for physical AI is going to be on the validation side next, not the training side. Whoever gets the validation efficiency of real-world behavior up first is the one whose mass-production cadence will hold steady. Validation work—short term, it still has to be done by people running it round after round.