How this term is used here In this channel's material, a bidirectional video model comes up in the context of real-time video-to-video AI — systems that look at frames of incoming video and generate frames of output video fast enough to respond interactively. Decart's demo of a game world rendered by AI in about 40 milliseconds is the flagship example explored in the conversation with Dean Leitersdorf, co-founder and CEO of Decart. [^01a0f15a-db02-705a-8d5c-476ec97c94b0@126] Why it appears alongside other architectures The term sits in the chapter Beyond LLMs: New AI Architectures and Capabilities, where the argument is that some problems fall outside what LLMs can do — verifiable correctness, spatial reasoning, exact determinism — and newer architectures are built for those jobs. Hybrid diffusion–autoregressive video models are named there as the form factor aimed at real-time video: diffusion to model rich visual structure, autoregression to keep generating frames frame-by-frame without falling behind. What to take away When the channel talks about a bidirectional video model, understand it as part of a family of post-LLM architectures designed so that video generation can condition on and produce video in both the seeing and the generating direction at once — which is what makes interactive, sub-frame-latency experiences possible. - A bidirectional video model generates video while conditioning on incoming video, enabling real-time video-to-video interaction. [^01a0f15a-db02-705a-8d5c-476ec97c94b0@126]- Decart's example: an AI-rendered game world updated in about 40 milliseconds. [^01a0f15a-db02-705a-8d5c-476ec97c94b0@126]- It belongs to the set of architectures explored as answers to problems LLMs cannot solve, alongside hybrid diffusion–autoregressive designs.