Large Language Models have given us an AI that reads, writes, and codes, yet they have not given us a robot that can fold laundry. Closing this gap is the next major frontier in AI, and it is precisely the problem world models exist to solve — systems that learn from raw observation how the world looks, how it evolves, and how it responds to an agent's actions.

The critical question is where such systems originate. While the popular view treats world models as a novel paradigm distinct from computer vision, I argue that world models are inherently rooted in computer vision. The large-scale representation learning, generative modeling, and spatiotemporal attention machinery driven by the past decade of vision research are exactly what we need to perceive, simulate, and act. World modeling is not what comes after computer vision; it is what computer vision becomes when we unify reconstruction, simulation, and action.

I will make this concrete through three interconnected works. First, LingBot-Map establishes the perception backbone, replacing hand-crafted SLAM with an end-to-end attention model that embeds geometric priors, recovering 3D structure from continuous video at scales legacy pipelines cannot achieve. Second, LingBot-World introduces the dynamics layer through neural world simulation, moving beyond static geometry to model the causal and physical evolution of scenes. Finally, LingBot-VA translates this world model into action by jointly modeling dynamics and control within a single autoregressive framework, turning a passive representation into an embodied controller.

Ultimately, world modeling represents the next phase of AI, and computer vision is the discipline that builds it.