Bilawal Sidhu argues that video models are learning an implicit representation of physics rather than merely producing attractive clips. When a system can predict the next frame quickly enough, generated video starts to function as a responsive environment.
He contrasts autoregressive generation with diffusion pipelines that render a fixed sequence through many denoising steps. A rolling context window lets the autoregressive system remember recent spatial state, accept camera or action inputs and continue the scene in real time.
The result is an early world engine with visible limits in long-term memory and consistency, but strong implications for games, simulation, robotics and embodied agents. Bilawal expects better memory and control to turn the visual generator into a reusable environment model.
Watch on YouTube
