NVIDIA, wake up. A new serious Chinese whole-body control rival from @GalbotRobotics in town. It's called Humanoid-GPT, and I think it's a bigger deal than the name suggests. The field has been stuck on one trade-off. Trackers that nail fast, agile moves tend to fall apart on anything they haven't seen before. Trackers that handle new moves tend to go sloppy on the hard stuff. This paper's claim: that wall isn't real, it's just a question of scale. Older trackers learn from a few million motion clips. Humanoid-GPT learns from two billion, all mapped onto a real Unitree G1. The smart move is how they teach it. First they train hundreds of small specialists, each great at one kind of motion. Then they fold all of them into a single model that can do everything. (On the name: the "GPT" just means it predicts the next move from what it has already seen, the same way a robot can only react to the past, never the future.) Its closest rival is SONIC. Both can copy unseen motions, but SONIC runs on a design that stops improving once you feed it more data. Humanoid-GPT keeps getting better the more you give it. Here's the part to actually watch. Look at the feet in tele-op videos. Spent enough time teleoperating SONIC to know the feet are where it struggles, the timing drifts and the steps lose their crispness. Humanoid-GPT holds the footwork far cleaner. The payoff: zero-shot dance, boxing, basketball, even getting up off the floor on real hardware. No fine-tuning. Under 1.5ms to run on a single 4090. Is this the next big thing in whole-body control? Nice work @HughWang19 and team 👏