Artificial intelligence has reached a dividing line.
Over the past decade, machines have learned to understand language, recognize images, and generate content. These abilities grew mainly in the digital world. Information was already recorded, problems often had clear boundaries, and the cost of error mostly stayed on a screen.
The physical world follows another set of rules. It changes continuously. Information is never complete. Every action has real consequences. The real world has no exhaustive samples, and it will not hand a machine the answer key.
To enter the physical world, a machine must form something of its own — belief.
01
I. Belief
A robot may see a cup and still not know what is inside; it may hear a sound and still not know where it came from; it may finish the action in front of it and still have to face the chain of changes that action sets in motion.
Belief is the judgment an intelligent system forms, from limited observation and prior experience, about the world's present state, its latent regularities, and the consequences of action. It contains certainty and doubt. It can persist across time, and it can update with new evidence. The machine therefore knows what it has seen, infers what it has not, estimates how much its judgment can be trusted, and decides what to do next.
A world model is the foundation of belief. It organizes scattered observations into a state of the world, distills limited experience into laws of change, and unfolds one action into many possible futures. The machine thereby gains a capacity beyond present perception: to rehearse consequences before acting, to revise judgment after the situation changes, and to find a new path when no ready answer exists.
Belief is how a machine begins to take a position on the world. A world model is how that position can be built, tested, and continually renewed.
02
II. Path
We believe that world models are the only path to general embodied intelligence.
Implementations will keep changing. Model forms will keep evolving. Whatever the architecture, a general agent that can enter the open physical world must understand how the environment changes, infer states that are not currently visible, predict the consequences of its own actions, choose among multiple futures, and correct itself from real feedback.
These capacities converge on the world model.
Imitation can acquire motions quickly. Data can cover common scenes. Rules can constrain known risks. Yet the open world will always produce new states beyond the training data. Furniture is moved. Objects wear. People change their intent. A seemingly small action can take the environment into a branch that has never appeared. Finite data cannot enumerate the infinite variation of reality. General embodied intelligence must grasp the laws that generate those variations.
The core of a world model, therefore, is whether it can extract regularities from little experience, remain consistent over long-horizon imagination, form a machine's belief about reality at this very moment, and let that belief enter action.
The world model sits at the junction of perception, memory, prediction, planning, and action. It decides whether a machine can move from “I have seen this, so I can do it” to “I understand this, so I can handle what I have not seen.”
Any path to general embodied intelligence must, in the end, arrive here.
03
III. Thresholds
For a world model to become the foundation of physical intelligence, it must cross three thresholds. These thresholds are also the standard by which we judge whether a world model has truly been achieved.
Sparse experience
Data in the real world is scarce by nature. Collecting it with robots is slow, expensive, and bound to costs of hardware, time, and safety. Even with enough resources, one cannot cover every state and action in an open environment.
The first value of a world model is to amplify limited experience. If a world model is built only by the data-fitting techniques of large models, it can only reproduce worlds the data has already covered. It can hardly generate new experience or explore new states.
We want world models to discover stable structure from a few real episodes: to understand the links among objects, relations, actions, and outcomes, and to transfer what is learned in one scene to many others. After one episode, a machine does more than remember a trajectory. After one failure, it can revise a class of judgments. After mastering one regularity, it can imagine a great many situations that have not yet occurred.
The most valuable experience in the future will be generated by the model and verified by reality. Real data provides the ground. The world model extends the frontier. Robot feedback continually corrects the model. Experience thus ceases to be data that is spent once, and becomes an asset that keeps compounding.
Forming belief
An embodied system always faces a partially observable world. Sensors offer only local evidence. Objects may be occluded. The environment may change out of view. The robot's estimate of its own state may drift. It must maintain a global judgment of the world across continuous action, keep multiple possibilities, and express its own uncertainty.
This belief must enter the system's ongoing operation. It must persist after an object leaves the field of view, update when new evidence appears, drive the machine to look when evidence is insufficient, and change the action when judgment may be wrong. The robot must form beliefs about the environment, and also about its own competence, the reliability of its sensing, and the risk of its actions.
We will turn the world model into a belief system that runs in real time — so that at every moment the machine can answer three questions: what state the world might be in now, what might happen next, and how confident I am in these judgments.
The world model thereby acquires a first-person view. The machine thereby gains a basis for continually understanding reality.
Action in concert
Belief must, in the end, enter action.
Action in the physical world cannot pursue only the success rate of a single task. A robot must also understand what changes the environment can bear, which futures the present decision opens or closes, which errors can be repaired, and which risks are irreversible. It must choose among efficiency, safety, information, and long-term outcome — and take actions that are in concert with the environment.
To be in concert means the machine can act in ways fitted to the structure, rhythm, and uncertainty of the environment. It seeks evidence first when information is lacking, leaves room when risk is high, replans when the environment changes, and asks for help when its own competence is not enough. It does not consume reality through repeated trial and error, nor does it demand that reality run under preset conditions.
The world model unfolds the future. Belief judges which reality the machine inhabits. Action hands the judgment to the real world to be tested. When the three form a closed loop, intelligence becomes able to run for the long term in an open environment.
Sparse experience builds a world. The world model forms belief. Belief produces action in concert with the environment. This is the path Cirquar has chosen.
04
IV. Cirquar
Cirquar's technical foundation comes from nearly a decade of continuous research on world models at the LAMDA Laboratory of Nanjing University, led by Academician Zhi-Hua Zhou — among the earliest and most deeply engaged world-model teams in China.
Before world models became an industry catchphrase, we were already working on the most fundamental questions: how a world model can imagine reliably over long horizons, how to control error that accumulates through imagination, how to keep representation and dynamics causally consistent, and how to enter policy learning and actually shape decisions.
We were the first to show that compound error in world models can be dissolved, and proposed methods that are both efficient and low in compound error. We were among the earliest to focus on the causal consistency of world models, offering systematic solutions at the levels of representation and imagination. We were among the first to explore world models in a unified modality, to build frameworks in which world models drive policy training, and to bring these capabilities into real business environments such as e-commerce, mobility, and local services.
These efforts have always followed one line: whether a world model can learn regularities from finite experience, remain correct over long-horizon imagination, and enter an agent's decision loop.
Today we take this accumulation from the laboratory into the physical world, and found Cirquar. We hope, through original work that is long-term and systematic, to build world models that are truly generalizable, deployable, and continually evolving — with home service as the core place they take root.
05
V. Transformation
Past machine intelligence mainly acquired ability from data left by humans. Language, images, video, and code record a world that has already happened. By learning these records, machines compress the past and recombine the past.
World models will open a new source of machine intelligence. A machine can construct possible experience, pose hypotheses about the world, compare different futures before acting, and continually update its own judgment through real-world feedback.
Intelligence thereby obtains a self-growing loop rooted in reality. Reality produces experience; the model amplifies experience. The model forms belief; belief guides action. Action changes reality; reality continues to verify the model. Once this loop begins to run, every deployment becomes the start of a new round of learning, and every robot becomes an entrance through which the system understands the world.
Human intelligence is bounded by an individual's time, body, and history. Machine intelligence can generate experience in parallel, transfer it across embodiments, gather it across systems, and accumulate it through continual deployment. What it forms will be a kind of world experience that is computable, shareable, and scalable.
This is the future that world models truly open.
The measure of machine intelligence will gradually shift from how many given tasks have been mastered, toward whether it can understand change, generate experience, form belief, and keep growing in the unknown. Datasets will no longer naturally constitute the boundary of ability. Every interaction with reality may extend the boundary of intelligence. Machines will grow from tools that invoke existing abilities into intelligent systems that continually build cognition in the world.
When the world model becomes infrastructure, the physical world will possess a continually running layer of intelligence: it turns reality into belief, belief into action, and action into new experience. Different machines, scenes, and tasks can share, on this foundation, an understanding of how the world changes.
We foresee that whoever first establishes a world-model platform that is generalizable, deployable, and continually evolving will have a greater chance to define the foundation layer of the next generation of physical intelligence.
Today we start Cirquar, and we start a cause that will take a long time to complete. We know this path must pass through layer upon layer of difficulty among mathematics, models, systems, and real machines. We also know that every long-horizon imagination, every real-world deployment, every action's feedback, brings us closer to the same goal:
Let finite experience grow into an understanding of the world. Let the world model form a machine's belief. Let machine intelligence keep evolving in reality.
General embodied intelligence begins in belief about the world, grows through imagination of the future, and is finally expressed as action in concert with reality.
Cirquar — let machines form belief, let intelligence grow in the real world.