Exercise 1: Why an Image Model Has No Representation of a Prompt as a "Request" — Possible Solution ==================================================================== HOW A DIFFUSION MODEL IS ACTUALLY TRAINED ------------------------------ Per this chapter, "during training, the model sees millions of (image, text caption) pairs and learns a statistical association between patches of text and patches of visual pattern." The training signal is purely correlational: given this caption, what pixel arrangements tended to co-occur with it across the training set. Nowhere in that process does the model ever see a labeled example of "here is a request, and here is what compliance with that request looks like" — the entire training objective is about matching text to image statistics, not about recognizing or fulfilling intent. WHY THIS RULES OUT A "REQUEST" REPRESENTATION ------------------------------ Per this chapter, "there is no step anywhere in that process resembling 'read this sentence, form an understanding of what is being asked, and comply.' The model has no representation of your request as a request at all — only as a string of text to condition image generation on." A representation of "this text is a request I should comply with" would require a training objective built around instruction-following, which is precisely what an instruction-tuned LLM like Claude has (RLHF/instruction-tuning, per this chapter's own contrast) and a diffusion model does not. The image model's text input functions purely as a conditioning signal — a set of statistical anchors pulling generation toward associated visual patterns — not as a message to be interpreted and acted on. WHY THE LLM CONTRAST MAKES THIS CONCRETE ------------------------------ Per this chapter, "prompt1-2's own anatomy of a good prompt... rests on one load-bearing assumption: that the model being prompted can follow an instruction. Claude reads [an instruction] and does something recognizable as obeying that instruction, because Claude was trained, via RLHF and instruction-tuning, specifically to map instructions to compliant behavior." This draws the line precisely: Claude's training includes an explicit step teaching it to recognize and comply with instructions; a diffusion model's training never includes any equivalent step, no matter how instruction-shaped the input text looks. WHY THIS WORKS AS AN ANSWER ------------------------------ It traces the absence of a "request" representation directly back to the chapter's own description of the diffusion training process (caption-to-pixel-statistics association, not instruction-to-compliance mapping), and uses the chapter's own Claude/RLHF contrast to show precisely what a diffusion model's training is missing that an instruction-tuned LLM's training includes.