"World models": when artificial intelligence learns to understand the world - World leading higher education information and services

Artificial intelligence is transitioning from generating text and images to modeling physical reality through a new frontier known as world models. Unlike previous generative systems that rely on statistical correlations, these models aim to equip machines with a form of common sense regarding physics, space, and logic. This shift seeks to move beyond imitation toward genuine understanding of how the world functions. Current prominent AI systems, such as ChatGPT or Claude, excel at predicting the next word or pixel but lack a consistent internal representation of physical reality. This limitation often leads to hallucinations, where models assert incorrect facts about the biological or physical constraints of the real world. Experts like Yann LeCun of AMI Labs and Fei-Fei Li of Worldlabs argue that moving past stochastic parroting requires architectures capable of modeling causes and effects. The concept draws inspiration from neuroscientist Kenneth Craik, who suggested in 1943 that the brain constructs small-scale models of reality to anticipate events. Practical application gained momentum following pioneering work by David Ha and Jurgen Schmidhuber in 2018, who demonstrated that AI could learn in virtual environments through internal simulations. These simulations allow agents to test strategies without interacting directly with the physical world. Recent proposals highlight the potential of this statistical approach. Meta’s V-JEPA model learns complex physical interactions by watching videos, while Google DeepMind unveiled Genie, which creates interactive virtual worlds from single photographs. These technologies synthesize observable information into abstract representations that facilitate planning and decision-making in uncertain situations. Applications extend to robotics, autonomous vehicles, and healthcare, where companies like Wayve claim to use models of the world to anticipate pedestrian behavior. Digital twins also simulate disease evolution or vehicle behavior. According to Julien Perez, a Lecturer in AI and Machine Learning at EPITA, these models remain in an advanced experimental phase. Large-scale adoption faces significant technical and regulatory challenges regarding robustness and security in complex real-world environments.