12.5 World models and video generation

What this is and why it exists

Video generation produces clips that look convincing, and the interesting question is not whether they look good — it is whether these systems have learned how the world works or a very good sense of what usually follows what. That debate is live, it matters for robotics and planning, and it is worth following honestly rather than settling from either direction. The reason to be careful is that evaluation here is genuinely unsolved, and curated clips are the least informative evidence available.

The vocabulary

• Temporal consistency — objects and scenes remaining coherent across frames.
• World model — a learned model of how a situation evolves.
• Learned simulator — using such a model to predict consequences of actions.
• Rollout — running a prediction forward several steps.
• Compounding error — small prediction errors accumulating over a rollout.
• Embodied AI — systems acting physically in the world.
• Plausibility — looking right; correctness — being right.

The mental model

Generating one good frame is a solved problem; making a hundred agree is the hard one. The diffusion machinery from the deep learning module produces convincing images. Extending it to video means every frame must be consistent with every other: an object keeps its shape and colour, a person's face stays the same face, something moving continues along a sensible path, occluded things reappear where they should, and the lighting does not change for no reason.

Temporal consistency is the field's main technical difficulty, and the failures are recognisable because they are the ones a viewer notices last and an engineer notices first: objects that morph slightly frame to frame, things that vanish behind something and come back subtly different, hands and text that are unstable, and motion that is smooth but does not conserve anything physical.

The approaches are what you would expect from the earlier modules — attention across time as well as within a frame, generating in a compressed space so the whole clip is tractable, and conditioning later frames on earlier ones. All of them work partially, and the cost is high: a video is a great many frames and the computation shows it.

The interesting proposal goes further than generation. A system trained to predict what happens next has, in some sense, a model of how things evolve. If that model is good, you can use it to plan: imagine the consequences of an action, imagine the consequences of a different one, and choose. That is a learned simulator, and it is a much stronger thing than a video generator — it is the same idea the classical AI module's search topics rested on, with the successor function learned from observation rather than written down.

Whether these systems have genuinely learned physics or learned plausible-looking correlations is the live debate, and both sides have real evidence.

For: the outputs respect a great deal that nobody encoded — objects persist, gravity acts approximately, liquids behave roughly as liquids, occlusion resolves sensibly. That is more structure than pure surface correlation would evidently produce, and probing the internal representations sometimes finds quantities that correspond to physical properties.

Against: the failures are exactly what a correlation account predicts. Quantities are not conserved when nobody is looking. Objects pass through each other in unfamiliar configurations. Counting fails. Rare situations, unlike anything in the training data, produce confident nonsense rather than uncertainty. And the systems are much better at what is common than at what is correct, which is the signature of learning the distribution rather than the mechanism.

The honest position is that it is a matter of degree and of domain, and that "does it understand physics" is a poorly posed question compared with "does its prediction hold up under this specific intervention" — which is answerable, and is what the careful work in the area actually asks.

The robotics link is direct and is where this stops being entertainment. A system that predicts consequences can choose actions, and the appeal for robotics is large: simulators are expensive to build and never quite match reality, while a model learned from recordings of the real world does not have that gap by construction.

The obstacle is compounding error. Predictions are approximate, and running one forward feeds its own output back in, so errors accumulate — a rollout that is convincing for a second becomes nonsense over ten. Every practical use is therefore short-horizon: predict a little way ahead, act, observe what actually happened, and predict again. That is a control loop rather than a simulation, and it is a reasonable way to use an imperfect predictor.

Two further difficulties, worth stating because they are frequently omitted. Physical action needs prediction fast enough to act on, and these models are slow. And the consequences of being wrong are physical — a wrong video frame is an artefact, a wrong prediction that a shelf will hold is a broken thing or a hurt person — which is why anything embodied needs the confirmation gates and bounded capability from the agent module, with more force rather than less.

Evaluation is unusually hard here, and it is the reason to hold announcements loosely.

Plausibility and correctness diverge, which is the core difficulty: a clip can look excellent and be wrong about everything that matters, and the automatic measures in use largely reward looking like the training distribution — which is what the model was trained to do, so they are close to circular.

Human evaluation is expensive and inconsistent: watching video takes real time, and viewers agree about "looks good" and disagree about "is right".

And the informative tests are the ones nobody shows. Long rollouts rather than short clips. Physical consistency probed deliberately — is the count preserved, is the object the same size after it passed behind something, does the liquid volume make sense. Prompts describing situations unlike the training data, where a learned mechanism generalises and a learned distribution does not. And, for anything claiming to be a simulator, whether a policy trained inside it works in reality, which is the only test that settles the question and the one most rarely reported.

So the reading advice this topic ends on: judge by demonstration reels and you will be wrong. Curated clips are selected, are short, and show what worked. Ask what the failure rate was, how long the rollouts were, what the unusual prompts produced, and whether anything trained inside it transferred out. What is worth learning here is the shape of the problem — that time is the dimension these approaches handle least well, and that plausibility is not evidence — rather than any particular system, because the systems will have changed by the time it matters and the shape will not.

What you should now be able to explain or do

Say why temporal consistency is harder than single-frame quality and name its characteristic failures. Explain what a learned simulator is and how it connects to search. Give both sides of the physics-versus-correlation debate with real evidence, and reframe the question as answerable. Explain compounding error and why practical use is a short-horizon control loop. Name the two further obstacles for embodied use. Say why plausibility and correctness diverge and why automatic measures are close to circular. Name the four informative tests, including the one that settles it. Read announcements accordingly.

Check yourself