preface

while the whole world is consumed with scaling inference performance for autoregressive models like llms , where we generate content based on last token and self-attention and reward and aligning based on human perferences / alignment — all this sounds nice , but is there any other way to do this. imagine what to do if you want to build a model simply denoising from its initial state to some state and you train it over and over to give some results — the models that are then formed is basically a diffusion based model.