A world model predicts how an environment might change after an event or action. It works by representing a state, applying a transition rule or learned dynamics and estimating possible next states. A simulation repeats that process to explore a sequence of futures. The important boundary is that a simulated result is a prediction under assumptions, not evidence that the world actually followed the prediction.
This complete guide develops the distinction with a fictional two-step parcel-delivery model, exact probability calculations, an independence trap, simulated trajectories and an independent three-step transfer exercise. It fills topic 070 in eduKateSG’s How Super Intelligence Works reading map, following chapters on Spatial Reasoning and Sensors and Edge Intelligence.
The series uses Super Intelligence as a broad editorial name for exploring AI systems, not as a claim that contemporary models already possess artificial superintelligence. Everything called a measurement in this chapter is either explicitly sourced or clearly labelled illustrative. The question is not whether a future looks impressive; it is whether we can show what generated it and what its evidence allows us to conclude.
Chapter contents and reading routes
Start with the model · Read Parcel Lab · Check the probability · Try the transfer case · Sources and next reading
- 1. A world model predicts a possible next state
- 2. State, action and transition: the three essential parts
- 3. Observation is not the whole underlying state
- 4. The complete fictional Parcel Lab
- 5. Construct the simplest possible transition table
- 6. Compute the two-step success probability
- 7. Translate the probability without inventing observations
- 8. Independence can be the most important hidden assumption
- 9. Policy is the rule that chooses an action
- 10. A simulation is repeated use of a model
- 11. Random seeds support reproduction, not truth
- 12. Fidelity means fidelity to the decision-relevant mechanism
- 13. Current generative world-model research
- 14. The simulation-to-reality gap
- 15. Counterfactuals require changed causes, not merely edited scenery
- 16. Multi-step errors compound
- 17. Models are useful when they disagree with a naive plan
- 18. Data provenance matters for learned world models
- 19. Repairing a misleading simulation report
- 20. Independent transfer: three stages with a shared failure
- 21. Full answers to the changed-context exercise
- 22. Evaluation checklist for a world model
- 23. Frequently asked questions
- 24. Sources, specialist routes and knowledge links
1. A world model predicts a possible next state
A world model is a representation used to predict how an environment may change, especially after a chosen action. It can be a short mathematical rule, a structured simulator or a learned neural model. The important operation is conditional prediction: given what the system believes about the current state, and given an action or disturbance, what might the next state be? A model of this kind supports planning, comparison and rehearsal.
Not every world model contains photorealistic video or a complete simulation of physics. A spreadsheet forecasting inventory is a modest model of part of the world; a grid that estimates where a cart will travel is another. A video-generating model may have a much richer appearance but a less transparent relationship between visual plausibility and physical accuracy. The appropriate model depends on what decision needs support.
Distinguish a model’s proposed next state from evidence of an observed next state. Predicting that a cart reaches a shelf is not the same as observing it arrive. Prediction should be labelled modelled, its conditions listed, and its consequences checked against suitable real or independently validated evidence where the stakes demand it.
2. State, action and transition: the three essential parts
Start with state: the variables that tell the model what matters now. For a training-room cart these might include its grid position, orientation, the location of barriers and whether its battery is adequate. An action could be move one square east, rotate, stop or request more information. A transition rule describes how the state is expected to change after that action.
Some transitions are deterministic under the modelling assumptions: a turn instruction always changes a symbolic orientation by ninety degrees. Others are probabilistic: a slippery floor or unreliable connection might prevent the intended result. A model can represent uncertainty by giving probabilities over possible next states rather than producing one inevitable outcome. The probabilities are only as justified as their calibration and source data.
The same reasoning applies outside robotics. A learning system might model which prerequisite becomes accessible after a particular practice exercise; a water-management simulation might model levels after rainfall. These uses have different domain mechanisms and evidence standards. Sharing the mathematical idea of a transition does not make their assumptions interchangeable.
3. Observation is not the whole underlying state
An observer may see the cart at (2,2) while its wheel condition remains unknown. A camera image reveals surfaces and visible objects but can hide what is behind a cabinet. The full state needed for reliable prediction may therefore be partly unobservable. A system can maintain a belief or bounded set of possible states rather than inventing a single exact world.
Suppose a status message states that a door is open, but the timestamp is fifteen minutes old. The message is evidence about the past, not confirmation of the present. Before directing the cart through the doorway, a real system needs a fresh observation and the right physical safety controls. A world model should represent this difference instead of quietly freezing the last known state.
Good learners ask, ‘Which state variable comes from direct observation? Which comes from a model? Which is assumed?’ That three-way distinction is unusually powerful. It prevents a plausible simulated future from being mistaken for a measurement of the world.
4. The complete fictional Parcel Lab
Our teaching laboratory is the invented Parcel Lab, a tiny warehouse-like grid with an automated delivery cart that exists only on paper. The cart begins at (1,1), and its first task is to reach (3,1) along two eastward moves. We assume each attempted move either succeeds and advances one cell, or fails and leaves the cart where it was. No diagonals, collisions, battery effects or human traffic are represented at this stage.
The fictional model assigns each move an independent success probability of 0.9. This is a teaching assumption selected for easy arithmetic, not a measured reliability estimate for any real cart. We will compare a two-step plan, alternative assumptions and an independent transfer case. The exercise focuses on the limits of a model’s conclusions rather than claiming a deployed machine could safely execute its output.
The task statement also requires honesty about the denominator. We are initially measuring whether two consecutive attempts both advance, not whether an eventual delivery succeeds after unlimited retries. If a failed movement is retried, the number of attempts and total expected time must be modelled separately.
5. Construct the simplest possible transition table
At state (1,1), choosing east leads to (2,1) with probability 0.9 and leaves the cart at (1,1) with probability 0.1. If that first move succeeds, another east command leads to (3,1) with probability 0.9 or leaves the cart at (2,1) with probability 0.1. If the first move failed and the plan blindly issues only one more east command, the cart cannot reach (3,1) within those two attempts.
The table is a complete simplified transition model. It does not claim to explain why a movement fails. That missing mechanism might be wheel slip, a blocked path or a cancelled instruction. If causes matter to the next decision, the state must include them, or the model needs a more realistic uncertainty structure.
Notice why a small transition table can teach more than a cinematic virtual warehouse. It forces every branch to have a meaning and every probability to sum appropriately. The appearance of realism is optional; the correctness of the relationship between action and next state is not.
6. Compute the two-step success probability
To reach (3,1) in exactly two attempts under this model, both independent moves must succeed. Multiply 0.9 by 0.9 to obtain 0.81. Thus the toy model predicts an 81% chance of reaching the destination after two commanded moves, and a 19% chance of not reaching it. The complementary 19% includes failure on the first attempt or failure on the second.
The successful-first, failed-second branch has probability 0.9 times 0.1 = 0.09. The failed-first branch has probability 0.1, regardless of what happens after it, because the two attempts cannot cover the full two-cell distance if the first did not advance. The branches do not overlap and sum to 0.81 + 0.09 + 0.10 = 1.
These are calculations inside the stipulated model. They do not measure any particular robot or prove that a 90% estimate is well calibrated. A meaningful report says ‘under two independent 0.9 move-success assumptions, the predicted completion probability is 0.81’ rather than ‘this delivery robot is 81% reliable’.
7. Translate the probability without inventing observations
If the same fictional two-attempt setup were repeated one hundred times under identical independent conditions, the model’s expected number of completed two-step runs would be 100 times 0.81, or 81. Expected value is a property of the assumed probability distribution, not a record of having performed one hundred trials.
Actual results would vary. Even a well-calibrated probabilistic model does not guarantee exactly eighty-one successes in the next hundred runs. A report should keep the language of expectation separate from the language of observed count. If the author writes that eighty-one parcels were delivered, the author has turned a prediction into a fabricated measurement.
To evaluate the model, a researcher would need an appropriately designed set of observed trials, clear success criteria and careful accounting for changes in conditions. Only then could they compare forecast probability with observed frequencies and investigate where the model misrepresents reality.
8. Independence can be the most important hidden assumption
Imagine that both steps depend on the same unobserved floor condition. In nine-tenths of fictional sessions, the floor provides adequate traction and both moves succeed; in one-tenth, traction is insufficient and both fail. Each individual move still has an overall success probability of 0.9, but the outcomes are no longer independent.
Under that correlated scenario, the probability that both moves succeed is 0.9, not 0.81. The identical marginal probability does not uniquely determine the combined event when dependence changes. This is why multiplying two impressive percentages can mislead if they share a common risk.
In real operations, shared battery limits, network outages, temperature and maintenance state can connect failures. A world model that ignores those shared causes may understate or overstate joint risks. The appropriate model must represent the dependence structure or honestly bound the result when it remains unknown.
9. Policy is the rule that chooses an action
A world model describes possibilities; a policy selects what to do with them. For the Parcel Lab, a naive policy commands east twice without checking where the cart is. A feedback policy observes the cart after the first attempt and decides whether to retry, stop or change route. These are different action rules even though the initial goal and transition model may be identical.
A planner can compare policies under a metric such as expected completion time, energy cost or number of unsuccessful operations. But a plan with a high predicted reward might still be unacceptable if it violates a safety boundary. The objective must include constraints, not just an attractive endpoint.
Students should distinguish ‘what will happen if we do A?’ from ‘which A should we choose?’ The first is prediction, the second is a decision. The third question, ‘are we permitted to do A?’, belongs to human and institutional authority.
10. A simulation is repeated use of a model
A simulation advances the model through a sequence of states. It may show one possible trajectory, many stochastic trajectories or an interactive environment where the user changes actions. A simulator is not automatically a world model; it is a method of executing a chosen model’s transition rules over time.
For the Parcel Lab, a simulation could sample a random outcome for each commanded step, update the cart’s symbolic position and store the resulting path. Repeating such runs approximates probabilities that may also be calculated exactly in this simple case. In more complex environments, simulation becomes useful because exact enumeration may be costly or impossible.
Simulation output depends on the model, sampling method, random seed, parameters and stopping criteria. A thousand runs of an incorrect model produce a precise estimate of that incorrect model’s behaviour. Repetition reduces some sampling noise; it does not repair a mistaken assumption about the world.
11. Random seeds support reproduction, not truth
A pseudorandom simulator can be configured with a seed so that a particular execution produces a repeatable sequence of sampled events. That helps investigators reproduce a surprising failure, compare code revisions and audit model logic. It does not make the simulated events real observations.
If you change the random seed, the realised counts may change while the underlying assumed probability remains the same. If you change the probability or dependence structure, the predicted distribution itself changes. A report needs to distinguish these two sources of variation: Monte Carlo randomness within one model and uncertainty about which model is valid.
For a transparent learning exercise, print the rule and the expected hand-calculated result before trusting a simulation. Then use the simulated output as a secondary check of implementation, not as proof that the fictional warehouse exists.
12. Fidelity means fidelity to the decision-relevant mechanism
A simulator can display richly textured shelves, sunlight and shadows while omitting wheel friction, human unpredictability or network delays. For delivery reliability, the missing physical variables may matter more than graphic detail. The correct question is whether the simulator represents the mechanisms necessary for the particular decision.
An educational simulation of number lines does not need photorealistic scenery. It needs accurate mathematical relations and useful feedback. A robot-training environment may require believable contact forces, sensor noise and timing. A climate scenario needs appropriate Earth-system dynamics and boundary conditions. Different decisions require different forms of fidelity.
The appropriate evaluation therefore begins by naming the claim: is the model suitable for conceptual practice, path comparison or real-world deployment forecasting? Evidence supporting one use should not silently be promoted to the others.
13. Current generative world-model research
Google DeepMind’s Genie 3 research announcement from August 2025 describes interactive environments generated from prompts, with examples of real-time navigable scenes. Its later model page describes the system’s capabilities and limits, including imperfect representation of real places and constrained interactions.
These publications document a research direction, not proof that a generated environment predicts every physical consequence correctly. Video continuity and controllable scenes are achievements of a particular kind, while reliable transfer to an actual workplace demands additional tests. Marketing descriptions of ‘world understanding’ require interpretation in terms of specific measured tasks.
One can use generative environments to practise ideas, expose agents to hypothetical situations or explore alternative scene configurations. But the simulated outcome should be tagged as generated. An imagined collision-free journey is not evidence that a real cart may safely take the same route.
14. The simulation-to-reality gap
The simulation-to-reality gap is the difference between behaviour inside the represented world and behaviour in the actual environment. A robotic policy might succeed in a virtual warehouse because obstacles are perfectly detected, movement follows ideal commands and people never unexpectedly enter the aisle. On hardware, sensing delay, wheel slip and changing layouts can break those assumptions.
Testing transfer requires observations from the intended setting with appropriate controls. Training on many simulated floor conditions may improve robustness, but it does not guarantee it. The right question is which variations were represented and whether the policy was evaluated against previously unseen real conditions.
To make the gap visible, write a table of modelling assumptions and their likely failure modes. Label every untested assumption. That is more useful than concluding that real-world deployment is safe because the rendered animation looks smooth.
15. Counterfactuals require changed causes, not merely edited scenery
A counterfactual asks what might happen under a meaningful change in conditions or action. Suppose the Parcel Lab cart takes a longer route to avoid a hypothetical obstruction. The model should alter the relevant map, update the route constraints and account for the new number of moves. Simply recolouring the obstacle while retaining the old transition logic is not a causal test.
Distinguish interventions from correlations. A simulator may learn that wet floors and failed moves often appear together. To reason about what happens if the floor is dried, it needs a mechanism relating traction to movement, not just a superficial visual correlation. Causal claims demand evidence beyond association.
Independent transfer checks are powerful here. Change the success probability, change the dependence assumptions, or change the available action and ask whether the calculated result follows. If a model repeats the original output without recognising the changed condition, its apparent flexibility is weak.
16. Multi-step errors compound
A transition model may be fairly accurate for one step while becoming unreliable over a long horizon. Small prediction errors can compound as later states are generated from earlier predictions. In an interactive environment, the model may drift in object identity, geometry or timing as the sequence grows.
One way to limit drift is to re-anchor predictions in fresh observations and restrict how far ahead decisions rely on unsupported details. A different approach represents uncertainty across trajectories rather than following only the most likely path. Neither method eliminates all limits, but both make assumptions and correction opportunities visible.
The Parcel Lab illustrates the principle in miniature: each additional independent step reduces the probability that every step succeeds on the first attempt. The arithmetic does not mean every real-world model has that exact form; it highlights why long-horizon claims need their own evaluation.
17. Models are useful when they disagree with a naive plan
A world model earns value by revealing consequences the planner overlooked. Suppose a naive planner assumes a two-step route always reaches the goal. The toy model instead predicts a nonzero chance of stopping short. That does not automatically select the best response, but it motivates a feedback checkpoint after the first attempt.
An effective model can also expose a trade-off: a shorter route may have lower expected success while a longer route has better clearance. The choice depends on constraints and the cost of different outcomes. A model should explain which trade-off is represented and which is omitted.
Do not judge a model only by whether it produces a confident plan. Judge whether its predictions change rationally when the inputs change and whether a user can identify what to verify before the action becomes consequential.
18. Data provenance matters for learned world models
A learned transition model depends on its training observations and their coverage. If the data contain only dry-floor movement, the model may be poorly informed about wet-floor conditions. If recorded success labels are incomplete, the model could learn from mistaken targets. These are issues of data provenance, not simply network size.
Training and evaluation populations should be distinguished. A model tested only on repeated versions of its training scenarios may look more capable than one facing genuinely unfamiliar layouts. Record which environments, action types, sensors and conditions contributed to the evidence.
Readers should also ask whether a benchmark score refers to video prediction, geometry consistency or task success. These are different outcomes. A system that generates convincing scenery can still misplace objects; a system that predicts a grid state accurately can still lack realistic visual rendering.
19. Repairing a misleading simulation report
Here is an intentionally flawed fictional report: ‘Our simulation delivered 81 out of 100 parcels, therefore the robot has an 81% real-world delivery rate.’ Two errors are present. The toy model produced an expected count of 81, not an observed count. And the probability was assumed rather than measured in the real environment.
A defensible rewrite is: ‘Under a fictional two-step model with independent 90% movement success, the predicted chance of reaching the endpoint after two attempts is 81%. No field performance has been measured.’ If actual simulation trials had been run, their counts could be reported separately with the random seed and model conditions.
The correction is not merely cautious wording. It changes which decisions can follow. The model may help design a testing plan, but it cannot certify a physical deployment, authorise a purchase or establish the cart’s real reliability without independent evidence.
20. Independent transfer: three stages with a shared failure
Now consider an invented library trolley route requiring three consecutive progress steps. If each step independently succeeds with probability 0.8, what is the probability that all three first attempts succeed? Give the numerical answer and explain what the word independently contributes to your calculation.
Next suppose all three steps depend on one shared condition: with probability 0.8 the route is clear and all steps succeed; otherwise the path is blocked and none advances. Each step still has marginal success probability 0.8. Does the three-step completion probability still equal the product 0.8 cubed?
Finally, someone says, ‘The virtual trolley completed the route, so the building is safe for the real trolley.’ Identify which evidence is missing and propose two independent observations that would be needed before considering physical movement.
21. Full answers to the changed-context exercise
With three independent 0.8 successes, the joint probability is 0.8 × 0.8 × 0.8 = 0.512, or 51.2%. This assumes that each required move independently advances on the first attempt. It does not count retries, changes in the route or other unmodelled conditions.
In the shared-condition case, completion occurs whenever the one common clear-route event occurs. Its probability is 0.8, or 80%, not 51.2%. The problem’s marginal single-step figures are identical, yet the joint outcome differs because the dependence structure is different. That is the reason the independence assumption belongs in the visible explanation.
A safe physical assessment would require, for example, measurement of the actual path width and inspection of unexpected obstacles and pedestrian activity, alongside appropriately validated control and stopping capabilities. These are examples of needed evidence, not a complete safety certification process. The simulator alone cannot answer them.
22. Evaluation checklist for a world model
First define the task: single-step state prediction, scene consistency, counterfactual explanation, policy comparison or physical action support. Next name the environment, allowed actions and state representation. Record measured ground truth separately from generated trajectories. Measure error under changed layouts and new conditions, not only on examples resembling training data.
Then test uncertainty: does the system report ambiguity, recover from an incorrect observation and revise after a contradictory measurement? Compare relevant baselines, including simple rules or physical models. A sophisticated generator should not win solely because the demonstration looks attractive.
Finally, evaluate the complete system: sensing, prediction, action choice, permissions, controller and verification. A component can pass its own test while the application fails to deliver the user’s intended outcome. Failure modes and repair routes are part of the curriculum, not an afterthought.
23. Frequently asked questions
Is a world model the same as a video model? No. Video generation can represent changing scenes, but a world model is defined by its predictive use and relation to environmental state and action. Different systems may overlap in capabilities without being identical.
Is a simulation a real experiment? It is an experiment on a model. It can help investigate possibilities and prepare tests, but its predictions must be validated against appropriate external evidence before being treated as claims about the real world.
Do larger models guarantee better physics? No. Model size by itself does not establish correctness for contact, long-horizon dynamics or uncommon conditions. Evaluate the relevant mechanism directly.
Can world models help students? Yes. Transparent toy simulations encourage students to state assumptions, predict consequences, test changes and compare their reasoning with observations. The educational benefit comes from that cycle, not from the glamour of the visual presentation.
24. Sources, specialist routes and knowledge links
Source starting points include Google DeepMind’s Genie 3 announcement and its current world-model capability and limitations page. These are provider descriptions of a research system, not proof of universal physical accuracy. For wider modelling methods, consult eduKate’s How Simulation Works and How Intelligence Works: Counterfactual Simulation.
For the neighbouring mechanisms, read Spatial Reasoning and Sensors and Edge Intelligence. A prior companion discussion, World Models, Causality and Planning, examines the broader possibility of advanced AI prediction; this chapter specifically supplies a transparent simulation audit and numerical transfer case.
Return to How Super Intelligence Works, How X Works and The Gold Standard of Education for connected learning. All Parcel Lab states, probabilities, routes and counts are original fictional teaching fixtures. No experiment on a physical device was performed and no real deployment statistic is claimed. Sources reviewed 22 October 2026.
Continue the companion learning route with Hallucinations, Grounding, Evaluations. Each chapter examines a different point at which an AI system moves from representations and sources toward a claim that can be checked. Return to the complete Super Intelligence reading map for the numbered curriculum and unfilled topics.