World Models Make a Comeback in AI: Can Simulated Realities Unlock True Intelligence?

The concept of a world model—a mental simulation of the environment that an AI system uses to predict outcomes and make decisions—has made a strong comeback in artificial intelligence research. Once dismissed as impractical, the idea is now central to the ambitions of leading AI labs striving to achieve artificial general intelligence, or AGI. Pioneers like Yann LeCun of Meta, Demis Hassabis of Google DeepMind, and Yoshua Bengio of Mila see world models as essential for creating AI that is not only intelligent but also safe, scientific, and capable of reasoning like a human. The roots of the idea stretch back to 1943, when Scottish psychologist Kenneth Craik proposed that organisms use internal "small-scale models" of reality to anticipate events and act more effectively. He linked this mental simulation directly to computation, suggesting that both brains and machines rely on the ability to model external events. This insight foreshadowed the cognitive revolution in psychology and laid an early foundation for AI. In the 1960s, AI systems like SHRDLU demonstrated early versions of world modeling by navigating a simplified "block world" to answer commonsense questions. But these rule-based models failed to scale. By the 1980s, roboticist Rodney Brooks rejected the idea entirely, arguing that the real world is its own best model and that explicit representations hinder performance. The revival of world models came with the rise of deep learning. Neural networks, trained on vast amounts of data, began to develop internal approximations of their environments—enabling tasks like virtual driving or game playing. As large language models (LLMs) like ChatGPT showed unexpected abilities—such as interpreting emojis or mastering Othello—some experts speculated that these systems must contain hidden world models, as Craik had envisioned. But evidence suggests otherwise. Instead of coherent, consistent models, today’s LLMs appear to rely on "bags of heuristics"—a patchwork of disconnected rules and approximations. They can answer questions or generate text with remarkable fluency, but lack a unified understanding of the world. For example, a model trained to navigate Manhattan can give accurate directions without a real map, but fails when faced with minor disruptions like blocked streets. It’s as if the AI only knows parts of the elephant—trunk, leg, tail—without grasping the whole animal. Still, these heuristics are powerful. With trillions of parameters, LLMs can mimic complex behaviors without needing a full internal model. But they lack robustness. A world model, even a simple one, could help AI systems adapt to surprises, reason reliably, and reduce hallucinations. That’s why the race is on. Google DeepMind and OpenAI believe that training models on multimodal data—video, 3D simulations, and more—will allow world models to emerge organically from neural networks. Meta’s LeCun, however, argues that new, non-generative architectures may be necessary to build proper world models from the ground up. While no one knows exactly how to build one, the potential rewards are immense. A true world model could bring transparency, reliability, and safety to AI—key steps toward meaningful AGI. The path remains uncertain, but the dream of a computational snow globe—containing a working model of reality—continues to drive innovation.