"Design for cheap training but rich inference The diffusion model framework was attractive to Inception's founder not just for image quality but for a specific architectural property: 'having something that is cheap to train yet very powerful at inference time.' During training you only need a single neural network evaluation; at inference time you chain many evaluations together, producing a very deep computation graph. Anyone designing systems or workflows can apply this principle — front-load simplicity in setup so the system can compound power when it actually runs. Source: Arash Vahdat, Latent Space Podcast, Diffusion LLMs with Inception AI"