‘World Models,’ an Old Idea in AI, Mount a Comeback By John Pavlus | September 2, 2025 --- Overview The concept of "world models"—internal representations that an AI carries to simulate and predict real-world environments—has surged back into prominence in AI research, especially in the quest for artificial general intelligence (AGI). Thought leaders like Yann LeCun (Meta), Demis Hassabis (Google DeepMind), and Yoshua Bengio (Mila) see world models as essential for creating AI systems that are smart, safe, and capable of scientific reasoning. --- What Are World Models? They are simplified, computational representations of reality ("a computational snow globe"). AI systems use them to internally test predictions and make decisions before acting on real-world tasks. Humans also have world models in their brains, allowing us to predict and avoid dangers (e.g., not stepping in front of a moving train). --- Historical Context The idea dates back to 1943 when psychologist Kenneth Craik theorized about a “small-scale model” of external reality in organisms' minds. This led to the cognitive revolution in psychology and linked cognition directly with computation. Early AI, like the 1960s' SHRDLU system, used hand-coded world models to answer questions about simplified environments but could not scale. By the 1980s, AI pioneer Rodney Brooks rejected explicit world models for robotics, favoring direct interaction with the environment. The rise of deep learning revived the idea, enabling AI to build internal approximations from trial and error, such as navigating virtual environments. --- Current Status and Challenges Large language models (LLMs), like ChatGPT, exhibit unexpected abilities which some attribute to implicit world models. However, research shows these AIs often function via disconnected "bags of heuristics" rather than coherent, consistent world models. This is compared to the parable of the blind men and the elephant: partial, inconsistent pieces rather than a whole understanding. Such heuristics can still produce effective results in many tasks. But without a true, consistent model, AI systems lack robustness—demonstrated by an LLM's failure to handle minor disruptions in a simulated Manhattan navigation task. A coherent world model would allow better adaptation and reliability in complex, changing environments. --- Why World Models Matter They may improve robustness, reasoning, interpretability, and reduce “hallucinations” (errors where AI fabricates information). Developing reliable world models is a key goal for AI labs and researchers aiming for safer, more intelligent AI. There's no consensus on how to build these models or even how to detect their existence within AI yet. --- Paths Forward Google DeepMind and OpenAI hope multimodal data (video, 3D simulations) will induce spontaneous formation of world models in neural networks. Meta’s Yann LeCun believes a new AI architecture (beyond generative models) will be necessary. The field remains uncertain but motivated by potential breakthroughs that world models could bring to AI. --- Related Reading Science, Promise and Peril in the Age of AI When ChatGPT Broke an Entire Field: An Oral History How ‘Embeddings’ Encode What Words Mean — Sort Of --- Summary The idea of embedding internal world models within AI systems is gaining renewed interest after decades on the sidelines. While current AI often relies on fragmented heuristics rather than consistent models, the quest to build or discover coherent world models could unlock safer, more adaptable, and genuinely intelligent AI systems. However, fundamental questions