The AI industry’s crash course in large language models may be giving way to a new obsession: world models. These systems aren’t designed to chat. They’re built to simulate the physical world—or at least a convincing enough facsimile to be useful for robotics, research, and generating 3D assets. Ars Technica recently spoke with three key practitioners in the space: Vincent Sitzmann from MIT, Anastasis Germanidis from Runway, and Ben Mildenhall from World Labs. Their insights reveal a fundamental inversion of the LLM product playbook. While chatbots like ChatGPT started with a universal interface and then went hunting for a use case, the leading world model efforts are beginning with specific, hard applications and working backward toward the tool.
That’s a significant strategic difference. It suggests the early wave of world model products won’t look like another text box. They’ll likely emerge inside robotics labs, video generation suites, and professional 3D content pipelines. The architecture has echoes of LLMs, and the scaling philosophy—throw more data and compute at the problem and watch capabilities emerge—is just as fervent. But the goal is spatial and physical reasoning, not linguistic fluency.
This shift also provides a convenient off-ramp for researchers disillusioned with the LLM path to AGI. Yann LeCun, Meta’s former chief AI scientist, captured that sentiment bluntly in Wired earlier this year, calling the idea that scaling LLMs alone will reach human-level intelligence “complete nonsense.” He’s far from alone. A sizable contingent of the field sees world models as a necessary corrective, a way to ground AI in the causal, three-dimensional reality that text merely describes.
The catch? No one quite knows what the interface for a world model will ultimately look like. The immediate outputs might be controlled video, navigable 3D scenes, or robot trajectories, not paragraphs of text. The tooling is raw, the interfaces undefined. For an industry that’s grown accustomed to the simple elegance of a chat window, that’s both a daunting engineering challenge and a refreshingly honest starting point.
