Multimodal generation industry 'black pill': diffusion models are just rendering shells, intelligent souls are all in language According to the dynamic observation of Beating, Ethan He, the former world model leader of xAI, threw out the "Black Pill" in the field of visual generation in the Latent Space podcast: the current video and image generation model does not really understand the physical world, and the diffusion model is essentially a pixel level renderer that does not have physical cognitive ability. Without the support of descriptive text with extremely high semantic details, diffusion models cannot output logical dynamic images at all. Taking the NVIDIA Cosmos model as an example, the core diffusion model responsible for image rendering only has 7B parameters, while the true intelligent hub is actually the Prompt Rewrite, which is served by the Large Language Model (LLM). The rewriter is responsible for extending short human instructions into ultra long and detailed descriptions that include light and shadow, materials, environment, and actions. The logical quality and matching degree of the final video depend almost entirely on the rewriting quality of the prompt words by the language model, rather than the diffusion model itself. This decoupling of vision and language heralds a complete reconstruction of human-computer interaction. With the sharp decline in inference costs, future terminal interactions will evolve into Generative UI, and traditional front-end code such as React will be eliminated. The big language model performs logical reasoning in the background, while the real-time video diffusion model directly generates personalized interfaces at the pixel level based on instructions. [Original link]