Where this term shows up In this channel's material, an autoregressive video model appears in the exploration of architectures beyond the standard text-prediction LLM — specifically in the discussion of Mirage, the real-time video-to-video system that generates a playable game world in about 40 milliseconds per frame [^08615b13-5661-5b2b-bcb0-3e4b551b9b63@504]. How the idea connects to LLMs The analogy the channel draws is direct: just as a large language model predicts the next token by feeding everything it has already produced back in, an autoregressive video model predicts the next frame the same way — each newly generated frame becomes part of the input for generating the next one [^08615b13-5661-5b2b-bcb0-3e4b551b9b63@504]. That frame-by-frame feedback loop is what makes Mirage's world generation feel continuous rather than pre-rendered. Why it matters in the broader comparison Within the channel's treatment of model architectures, the autoregressive approach is one option among several — it changes how the model produces output, and that change determines whether it fits the task. Real-time interactive worlds like Mirage lean on autoregressive frame prediction because latency matters frame to frame, whereas other architectures (such as non-autoregressive energy-based models) trade off differently [^08615b13-5661-5b2b-bcb0-3e4b551b9b63@504]. - An autoregressive video model generates video one frame at a time, using previously generated frames as input for the next prediction.- The channel frames it as an LLM next-token prediction applied to video.- It is central to understanding Mirage, the real-time video-to-video game-world engine covered in this material [^08615b13-5661-5b2b-bcb0-3e4b551b9b63@504].