A new kind of artificial intelligence called "world models" is being explored as a potential next step beyond large language models (LLMs), which are trained on text to predict the next word in a sequence. Unlike humans, LLMs have never physically interacted with the world, such as knocking a glass off a table, which may limit their capabilities. World models are systems that learn through observation to simulate the consequences of actions in the real world, attracting interest for their promise in autonomous robotics and offering a potential route to artificial general intelligence (AGI), or machines with human-level reasoning. However, the term "world model" is confusingly elastic, becoming "a kind of shorthand for all the things that current AI systems can’t do well," according to Melanie Mitchell at the Santa Fe Institute in New Mexico. It is unclear what these systems should model or whether internal representations of physical reality will be sufficient for generalisable intelligence. The concept traces back to 1943, when psychologist Kenneth Craik wrote that the human mind "carries a small-scale model of external reality and of its own possible actions," allowing us to "try out various alternatives" and "react to future situations before they arise." The idea has evolved with the theory of predictive processing, which posits that perception relies on the brain generating predictions about the external world and updating them with sensory data.