Sander Dieleman, Research Scientist (Director) at Google DeepMind, explains how much modern diffusion models depend on guidance. Turn it off and the output falls apart, with problems in both the global structure and the fine detail. Sander argues that same dependence helps explain why diffusion models can run with far fewer parameters than large language models. In this clip, Sander covers: • What happens when you sample from a diffusion model without guidance • Why both global structure and fine-grained detail break down • How guidance lets diffusion models get away with being smaller • How diffusion model size compares to large language models