“Chinese AI labs appear to be converging on sparse attention mechanisms, while American labs like NVIDIA and AI2 are converging on hybrid attention architectures.”