Pushing this idea to the limit, why not have it deal with all of the output tokens in parallel? If you've been watching my channel for a while, you know where this is going. Diffusion LLMs do exactly that. As you already know, auto-regressive LLMs generate one token at a time until the end of sequence is produced, or a maximum token count is reached. So the number of iterations is dynamically determined by the model for each prompt. Diffusion LLMs are a completely new paradigm. They no longer generate one token at a time, but rather one draft at a time. They start with a fixed context window of complete nonsense and iteratively clean it up into a coherent response. Similar to auto-regressive models, they're still sequential. Each draft needs to wait for the completion of the previous one. However, they perform a fixed number of iterations. This is a hyper parameter that acts as a knob between speed and quality. More iterations take more time, but also improve the response. In a previous video, I covered ways of bringing this number down to tens or hundreds without compromising the final output. For an average LLM response with a few thousand output tokens, a diffusion model will perform fewer iterations than a traditional one. But of course, in case of a yes or no answer, the diffusion chain is longer.