.@bilawalsidhu reveals how the Seoul World Model generates realistic training data for robotics: "Seoul World Model is a really interesting paper, very similar to what Google's doing with Street View grounding. If you ask Genie to give you a representation of the Palace of Fine Arts, it'll give you a plausible reconstruction, but it ain't exactly it." "One way to constrain these models is, we could just do what we do in LLMs, which is RAG, retrieval-augmented generation." "Can we do a form of spatial RAG where you just pull in the nearest reference image of reality to condition the generation? That lets you do a lot of these use cases that are interesting for robotics." "If you wanna create a bunch of training data for downtown Seoul in different lighting conditions, really cool way to go about doing that. That's exactly what you see folks like NVIDIA trying to do with Omniverse as this 3D simulator and Cosmos."