Compiler View on Heterogeneous Data Fusion: Embodied AI Bottlenecks Lie in Data Paths, Not Models

The "ChatGPT moment" for embodied intelligence is taking forever to arrive, and the debate in the tech community has shifted from model scale to data sources. What Fang Han mentioned about the "accounting logic" and heterogeneous data hits right at the most painful part of the current AI Infra layer—the data pathway isn't a compute bottleneck, it's an IO bottleneck. As someone working on Ascend AI compilers, my daily job involves figuring out how to cram more effective computation into limited memory bandwidth.