The latest frontier in artificial intelligence research revolves around the pursuit of world models, simplified representations of reality that allow machines to reason, predict, and test decisions before acting. The idea is that an AI carrying a computational “snow globe” of its environment could operate more intelligently and safely. Leaders such as Yann LeCun, Demis Hassabis, and Yoshua Bengio argue that these internal models are essential to achieving artificial general intelligence (AGI).
The concept is not new. Psychology, robotics, and machine learning have long relied on mental or computational simulations. Kenneth Craik, a Scottish psychologist, laid the foundation in 1943 when he proposed that organisms carry a “small-scale model” of external reality in their heads, enabling them to weigh options and act more competently. This notion linked cognition to computation and influenced the cognitive revolution of the 1950s. Early AI systems attempted to mimic this approach. The program SHRDLU in the 1960s, for example, used a simplified “block world” to answer questions about objects, impressing observers with rudimentary reasoning. Yet these handcrafted systems failed to scale, and by the 1980s, roboticist Rodney Brooks declared that explicit world models were a dead end, preferring to let machines respond directly to the environment.
The rise of deep learning revived Craik’s idea. Neural networks trained on vast data could develop internal approximations of their surroundings, proving effective in specific domains like simulated racing. More recently, large language models (LLMs) demonstrated surprising emergent behaviors, such as inferring patterns they were never explicitly trained on. For experts like Geoffrey Hinton and Ilya Sutskever, these feats suggested that some form of embedded world representation might be emerging within the networks.
However, closer inspection tempers the excitement. Instead of cohesive world models, today’s AIs seem to learn disconnected heuristics — bags of rules of thumb that cover isolated tasks but often contradict each other. Researchers liken this to the parable of the blind men and the elephant: models grasp fragments of reality without perceiving the whole. A striking experiment at Harvard and MIT showed that an LLM could give nearly perfect directions in Manhattan, yet collapsed when 1% of the streets were blocked. Without a coherent, map-like representation, it could not adapt. This highlights why robust models of reality matter: they enable flexibility and resilience when conditions change.
The stakes are high. Reliable world models could help extinguish AI hallucinations, improve reasoning, and make systems more interpretable. Yet how to build them remains unresolved. DeepMind and OpenAI are betting on multimodal training — feeding systems not just text, but also video, 3D simulations, and diverse sensory data — in hopes that consistent models emerge spontaneously. Yann LeCun envisions an entirely new architecture designed specifically for predictive modeling, departing from today’s generative approaches.
The search for effective world representations may not yield the mythical El Dorado of AGI, but it could produce practical advances that make AI safer, smarter, and more reliable. In this sense, the quest for world models represents both a return to the roots of cognitive science and a bold step toward the future of artificial intelligence.
www.quantamagazine.org/world-models-an-old-idea-in-ai-mount-a-comeback-20250902/