World Model

Partially. Sora understands basic physics (gravity, reflections, fluids) but doesn't have complete world understanding. OpenAI describes it as a 'world simulator'.

Robotics (understanding physical environment), autonomous driving (predicting other vehicles), and video generation (realistic physics).

A world model is an internal, learned representation of how an environment works, letting an AI predict what happens next from its current state and actions. Instead of reacting only to raw inputs, the system simulates plausible futures. This idea underpins model-based reinforcement learning, robotics and modern video generators, where the model captures dynamics like motion, physics and object permanence.

A large language model predicts the next token in text and excels at language tasks. A world model predicts the next state of an environment, often visual or physical, capturing dynamics like motion and collisions. They overlap: video generators like Sora and Genie blend both ideas. The key difference is what they model, sequences of words versus how a world evolves over time through actions.

Notable examples include DeepMind's Dreamer, which learns to act by imagining outcomes inside a learned model, and Genie, which generates playable, interactive environments from images or text. OpenAI's Sora is described as a world simulator for video. Yann LeCun's JEPA family learns abstract predictive representations rather than pixels. Together they show world models spanning robotics, gaming, video and research.

An agent with a world model can imagine the consequences of actions before taking them, mentally rolling out possible futures and choosing the path with the best outcome. This is called planning in latent space. Because the agent rehearses inside its model instead of the real world, it learns faster and more safely. Dreamer-style systems use exactly this loop to master control tasks efficiently.

Robots act in a physical world where real trial and error is slow, costly and risky. A world model lets a robot predict how objects move, fall or resist, so it can plan grasps and motions inside simulation before acting. This reduces real-world data needs and improves safety. It also helps the robot generalise to new situations by reasoning about physics rather than memorising fixed responses.

A traditional simulator is hand-coded by engineers with explicit physics rules. A world model is learned from data, inferring the rules of an environment by observing it. This makes world models more flexible for messy, real situations a human could not fully script, but also less precise and harder to verify. Many modern systems combine both, using learned models where hand-written physics falls short.

Today's world models still struggle with consistency over long horizons, often drifting into impossible physics, vanishing objects or broken cause and effect. Video systems like Sora can mishandle complex interactions or object permanence. They are expensive to train, hard to interpret, and may learn surface patterns rather than true physical understanding. Reliable long-term planning and faithful real-world dynamics remain open research challenges.

Many researchers think so. Figures like Yann LeCun argue that learning a predictive world model is essential for common sense and reasoning, since humans understand the world by anticipating it. By grounding AI in how environments behave rather than only in text, world models could enable agents that plan, adapt and act reliably. They are widely seen as a key building block toward more capable, general systems.