first. So, if you want to build an intuition, then stick around. Today, state-of-the-art LLMs predict one token at a time based on the previously generated tokens. The result is left to right generation, similar to how we write English words on paper. This is the family of auto reggressive models or ARMS. Diffusion proposes a new paradigm. A diffusion model generates the entire text sequence at once, starting with complete gibberish at time zero. At each point in time, it recomputes a full draft of the sequence. Some words persist and some get replaced with better guesses. This refinement repeats over several steps like a student revising an essay. Now, I want to point out a flaw in this terminology. While the next token prediction models are singled out as auto reggressive, I would argue that diffusion models are auto reggressive too because they ingest their own output. They just differ in the granularity of the output, an entire response candidate instead of a single token. But I'll stick to the established terminology to avoid confusion. In image