Review – 3
ChatGPT:
In artificial intelligence, world models refer to an AI system’s internal representation of how the world works – essentially a simulated reality inside the machine’s mind. A world model enables an AI agent to predict future states of its environment and imagine potential scenarios without directly experiencing them. This concept has roots in cognitive science: humans are thought to carry a mental model of the world, allowing us to anticipate outcomes and plan our actions. As system dynamics pioneer Jay Forrester described, “The image of the world around us, which we carry in our head, is just a model… [one] has only selected concepts, and relationships between them, and uses those to represent the real system.”. In other words, we do not mentally simulate every detail of reality, but we hold an abstract model that captures the essential structure of our world. In AI terms, a world model is typically a generative model of the agent’s environment. Formally, researchers define a world model as a system that, given the current state of the world and an action taken by the agent, can predict the next state. For example, if the current state includes the position of objects and the agent’s status, and the action is “move forward,” the world model should predict what the new positions and statuses will be after that action.
A world model is all about learning an internal model of the world’s dynamics so that an AI can understand, predict, and plan within its environment. This involves learning abstract representations (latents), using self-supervised prediction as a training signal, capturing causal relationships over time, and leveraging the model for imagination and planning. The end goal is an AI that doesn’t just react reflexively, but “knows” how the world works and can reason about what to do – much like humans who foresee outcomes using mental models.
Consider this vivid analogy from a recent essay: An LLM is like a brilliant autocomplete that can imitate understanding by stitching together patterns in language, but it has no “imagination” beyond what it’s seen in text. A world model, by design, imagines – it simulates situations that perhaps never occurred exactly the same in its training data. It’s the difference between reading about the world versus experiencing a simulated mini-world. One paper framed it elegantly: “LLM simulates the next word in human language… [A world model] simulates the next world.” [Source]. Each has its strengths – LLMs leverage vast knowledge encoded in human text, whereas world models provide an internal sandbox for truly understanding and navigating an environment. The consensus emerging is that next-generation AI systems likely need both: the knowledge and linguistic capabilities of LLMs plus the grounded predictive planning of world models. Indeed, researchers have begun experimenting with combining the two, for example by training language models with objectives that encourage predicting missing text in a way analogous to JEPA (to imbue them with a bit of world-model-like consistency) [Source].
Claude:
Kenneth Craik proposed in 1943 that minds construct “small-scale models” to anticipate events before they occur. Philip Johnson-Laird’s mental models theory (1983) suggests we reason by mentally simulating scenarios rather than applying formal logic. The predictive processing framework in neuroscience argues the brain constantly predicts sensory input, updating beliefs only when predictions fail.
World Models translate these cognitive principles into AI architecture. They comprise nine interconnected components: latent spaces that compress high-dimensional observations into 32-64 dimensional representations filtering out unpredictable noise; predictive coding that forecasts future latent states from current state plus action; model-based reasoning enabling deliberative Mode-2 thinking versus reactive Mode-1 responses; causal structure encoding cause-effect relationships; simulation capabilities for forward prediction and counterfactual “what if” analysis; memory systems maintaining both short-term hidden states and long-term associative knowledge; planning mechanisms using gradient-based optimization or search-based methods like Monte Carlo Tree Search; self-supervision through contrastive or non-contrastive learning from unlabeled data; and embodied agent capabilities closing the sensorimotor loop from perception through planning to action.
This architecture provides what LLMs fundamentally lack: sample efficiency through imagination rather than trial-and-error (humans learn to drive in 20 hours versus millions of RL episodes for pure model-free learning), common sense grounded in intuitive physics about gravity and momentum, planning and reasoning through forward simulation of consequences, generalization by capturing underlying dynamics that transfer across tasks, safety through virtual testing of dangerous scenarios, and computational efficiency by predicting abstractions not pixels. [Source]
