Diffusion models did not replace energy-based models. They are energy-based models, written differently. The noise-prediction network in a diffusion model is the gradient of an energy with respect to its input, and when data is perturbed with Gaussian noise the denoising score-matching loss coincides with the diffusion training loss. The literature states it plainly: the two differ only in parameterisation, energy against score. Now consider what each form costs to use. To use an energy you need samples, which need a chain. To use a score you evaluate a network once and you have the gradient directly, because the network was trained to be the gradient. Both describe the same distribution, and they differ in what a gradient costs. The form with the cheaper gradient is the one carrying public models at scale, which is the mechanism of lesson five appearing in a current field rather than a historical one.