What Are World Models? A Short History of the Idea
— by Vivax
Before large language models, AI researchers chased a different dream: machines that build an internal model of how the world works.
The idea is older than modern AI. In 1943, the Scottish psychologist Kenneth Craik proposed that the brain carries a 'small-scale model' of external reality — a way to try out alternatives and react to future situations before they arise. That single intuition — that intelligence means simulating the world internally — has guided generations of researchers, and it is the foundation of what we now call a world model.
The modern, computational version took shape in the 1990s, when Jürgen Schmidhuber described recurrent networks that learn to predict their environment and use that prediction to plan — a controller paired with a predictive 'world model.' The idea broke into the mainstream in 2018, when David Ha and Jürgen Schmidhuber published a paper simply titled 'World Models,' showing an agent that learned a compact internal simulation of its environment and could even train almost entirely inside its own 'dream.'
From there the field accelerated. Yann LeCun's 2022 proposal for the Joint Embedding Predictive Architecture (JEPA) argued that machines should learn abstract, predictive representations of the world rather than generating every pixel or word. By 2025, world models had become an industry: NVIDIA released Cosmos, a family of world foundation models for physical AI, and Google DeepMind's Genie line learned to generate interactive, playable environments from a single prompt — early glimpses of general-purpose simulators of reality.
Why does a medical-AI company care about the history of an idea? Because a world model is exactly what safe clinical AI needs: a system that understands cause and effect, can simulate what happens next, and reasons about the patient and the physical environment together — not just predict the next word. Vivax is building a clinical world model on this lineage, turning eighty years of theory into grounded, decision-grade tools for hospitals.