A world model earns its keep when the system can rehearse possible futures inside a scene before it acts, or before it recommends an action.
Language models continue and coordinate symbolic information. That is the shape of the product. You ask for a memo; the model produces a memo. Modern language models also reason about consequences in words, plan multi-step actions, and orchestrate tool calls, so the picture is richer than pure completion. But the core capability is still symbolic continuation, and its output is the artifact you paid for.
World models, at least the ones worth taking seriously, are aimed at a different capability. They produce a short predicted trajectory — sometimes as latent representations, sometimes as generated video, sometimes as interactive frames — of what would happen if a particular action were taken in a particular scene. The trajectory is the material for a downstream step: a policy scoring alternatives, a reviewer choosing among candidates, a controller committing to a motion.
I use "rehearsal" as shorthand for planning inside a learned model, imagination training, and predicted rollouts. It also draws a bright line between systems that look impressive on stage and systems whose predicted trajectories a downstream process can actually use. Most of this essay is about that line.
What rehearsal actually looks like
Take the most literal example. A warehouse robot needs to pick a box off a shelf. The shelf is crowded. The box is at an awkward angle. Two other packages are stacked nearby.
Without any world model, the robot follows a script. Move to this position, close the gripper, lift. If anything is different from what the script assumed, the script fails and something falls or the robot stops.
With a world model, the robot could do something structurally different. It could run a possible motion inside a predictive model of the scene first, generate alternative trajectories, and score them against a goal before committing. This warehouse example is illustrative, not a reported deployment. The Dreamer research line from Danijar Hafner and colleagues studies the same underlying loop; the most recent step, Dreamer 4, trained an agent in a learned Minecraft environment using offline data and reported real-time interactive inference on one H100, a high-end data-center GPU (paper). That is a demonstrated result inside a game world. It does not, on its own, generalize to a real warehouse.
The point is not that rehearsal is settled. It is that a class of systems exists whose primary output is not a finished artifact for a human reader, but a trajectory an agent, controller, or reviewer can act on.
Four properties that do not travel together
The most common failure mode in current executive coverage of world models is to conflate four separate properties. These do not have to appear together, and in the current generation of systems they usually do not.
Visual plausibility. Does the generated sequence look right to a human viewer? OpenAI's Sora technical report and DeepMind's Genie 3 announcement are strong on this axis. Sora produces beautiful clips. DeepMind reports Genie 3 rendering interactive 720p environments at 24 fps with visual consistency over several minutes. That is real, and it is not the same as any of the three properties below.
Controllability. Does the model respond predictably to the actions or instructions given to it? Wayve's GAIA-2 is aimed here, generating controllable multiview driving scenarios for offboard development and evaluation. Runway's GWM Worlds 2 is aimed here in the interactive-media direction. Both are vendor-reported capabilities; both matter, because a rehearsal you cannot condition is a demo you cannot plan with.
Physical executability. Do the predicted transitions respect the physics of the actual system — momentum, contact, mass, occlusion, object permanence — closely enough that an action selected on the prediction survives contact with the real world? This is a much harder property. The 2026 evaluation literature — RoboWM-Bench, RoboPhys-3D, WoW-World-Eval — is being built precisely to grade this axis, and the picture is fragmented and task-specific.
Downstream task utility. Does using the model's rehearsals to inform decisions actually improve the outcomes that matter? WorldArena 2.0 and the V-JEPA 2 paper measure task and planning success. Executives should also ask about safety margins, cost per successful action, and recovery time, but these papers do not establish those operating KPIs. The V-JEPA 2-AC experiments used less than 62 hours of robot video for zero-shot planning, and success rates vary materially by task and object.
The four properties are related. They are not the same. A system can be strong on visual plausibility and weak on physical executability. A latent predictive model can be strong on downstream task utility inside a bounded domain and generate nothing a human would want to watch. Treating them as one axis is where a lot of the current buying error lives.
World models
Rehearsal matters where collecting experience is costly, rare, or unsafe
Robotics, driving, interactive worlds, and industrial work use different model families. Each is a candidate for exploring trajectories before real execution, subject to setting-specific evidence.
Four world-model use cases are shown: comparing robot motions, generating rare driving conditions, training inside interactive generated worlds, and rehearsing industrial procedures.
Compare motions before contact
Roll out grasp or manipulation candidates before the robot commits to one.
Generate the rare condition
Create controllable weather, traffic, and edge cases for training and evaluation.
Train inside generated experience
Let people or agents act inside an environment that responds frame by frame.
Rehearse a consequential procedure
Test alternatives against a grounded twin or simulator before touching production.
Where rehearsal appears to pay, and where the evidence is thin
The domains where rehearsal is a reasonable planning bet are legible, and they cluster in places that already invested in simulation-based engineering.
Robotics and industrial automation, where hand-scripted behavior struggles with novel objects and situations.
Autonomous-driving development, where the space of edge cases is too large to enumerate by hand. Wayve positions GAIA-2 as an offboard scenario generator for development and evaluation — an important role, and not the same as evidence of on-road autonomy.
Warehousing, logistics, and fulfillment are candidate settings because the environment can be bounded and the cost of small mistakes accumulates fast at scale. The evidence here is still primarily research, simulation, and vendor-reported capability rather than independent proof of scaled deployment.
Interactive worlds for training, simulation, and creative production, where Genie 3 and GWM Worlds 2 are visible vendor-reported examples.
Reconstructive and persistent spatial systems, where Marble and early-access Atlas create or simulate 3D environments without yet proving the same live interaction profile.
Bounded industrial-operations and other constrained control environments, where NVIDIA's Cosmos 3 has been packaged as an open, converged reasoning, world, and action model family for developers to build on.
What the current evidence supports is a moderate-confidence forecast: bounded physical autonomy is likely to expand first in structured environments with measurable objectives and constrained action spaces. Independent deployments showing repeatable task success, safe fallback, and economic value across multiple sites would confirm it; simulator bias and long-horizon error keeping systems confined to development, media, and operator-assistance roles would falsify it.
What the current evidence does not support is a general claim that world models are, today, carrying industrial workloads at scale. Vendor demonstrations are real. Third-party evaluation on the specific settings that matter is much rarer, and any pitch that a public demo or an open-source model release proves dependable industrial autonomy should be read with the primary evaluation literature in the other hand.
Notice what is not on the list above. Rehearsal against a world model does very little for a memo, a slide, an email, a customer service transcript, or a legal brief. Language models already handle those competently, and there is no physical outcome to rehearse against. The world-model conversation is not a threat to the chatbot conversation. It is a different conversation, for a different portfolio, spent by a different budget owner.
The demo trap
A visually stunning clip is not, on its own, evidence that the underlying model can be controlled predictably. Controllability is not evidence that the predictions respect physics. Physically plausible predictions are not evidence that using the model to select actions improves outcomes. Each of the four properties has to be measured separately, on the setting that matters, before it can be assumed.
The current generation of world models produces visually plausible sequences that vary materially in how well they satisfy the other three properties. My engineering expectation is that early production systems will combine learned models with established simulation, safety controllers, human oversight, and constrained action spaces. That is a design inference, not evidence that the pattern is already widespread.
If you are being pitched a world model as a substitute for existing simulation, controller software, or human oversight in a safety-relevant setting, the primary sources do not clear that bar today. If you are being pitched it as an offboard tool for scenario generation, developer testing, synthetic data, spatial reconstruction, or bounded operator assistance, the primary sources support a much more useful conversation.
Limitations, and a fair objection
The strongest objection to the argument above runs like this. Real physical systems already have simulators. Aerospace has aerodynamics simulators. Automotive has crash simulation. Semiconductors have process simulators. Grids have power-flow tools. None of those needed a learned world model to be useful, and several of them are far more accurate inside their domain than any learned model available today. So what exactly is being added?
That objection is largely correct on its own terms, and it is where I would slow down anyone getting too excited.
What learned world models add is not accuracy inside a domain that already has a good engineered simulator. What they add is coverage in situations the engineered simulator was never built for — unstructured environments, novel objects, human behavior, dense sensor input — at a compute cost that scales with data rather than engineering time. That is a real addition. It is smaller than the frontier vocabulary tends to imply.
So the defensible positioning is not "world models replace physics simulators." It is "world models extend rehearsal into environments the physics simulators never covered well, and are increasingly combined with them." The version I would push back on hardest is the reverse claim that nothing structurally new is happening. Latent predictive models, interactive generative environments, and world-action stacks are architecturally distinct from language-first systems, and they are aimed at prediction problems the language-first path was not built to solve.
Close
Rehearsal is where the world-model story becomes interesting.
If a system can predict how a bounded scene will evolve, well enough that comparing two candidate actions produces better outcomes than not comparing them, the loop earns its compute. Whether any specific system clears that bar in any specific setting is the honest question, and it depends on all four properties — visual plausibility, controllability, physical executability, downstream task utility — landing well enough on the axis that matters.
Part 3 turns to substrate. A rehearsal is only as informative as the data it was trained on and the operational record it can compare itself against. Document retrieval, the enterprise AI substrate of choice for two years, is insufficient for that job, and the next essay says why in plain enough language for a CIO to act on.
Sources
- Ha, David, and Jürgen Schmidhuber. "World Models." arXiv preprint, March 2018. Source. Established the plan-inside-a-learned-model vocabulary this essay uses.
- OpenAI. "Video Generation Models as World Simulators." Technical report accompanying Sora, February 2024. Source. Vendor-reported framing.
- Google DeepMind. "Genie 3: A New Frontier for World Models." August 2025. Source. Vendor-reported interactive generation.
- Wayve. "GAIA-2." March 2025. Source. Vendor-reported offboard driving-scenario generator; not evidence of on-road autonomy.
- Meta AI. "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning." Paper, June 2025. Source. Success rates are task- and object-specific.
- Danijar Hafner et al. "Dreamer 4." Project and paper, September 2025. Source. Minecraft result; do not generalize to robotics.
- Runway. "Introducing GWM Worlds 2." September 2026. Source. Vendor-reported indefinite-interaction claim.
- NVIDIA. "Develop Physical AI Reasoning, World, and Action Models with NVIDIA Cosmos 3." May 2026. Source. Vendor-reported converged model family.
- World Labs. "Marble" (November 2025) and "Atlas" (September 2026 early access). Marble. Atlas.
- Jiang et al. "RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation." CVPR Workshops 2026. Source.
- "WorldArena 2.0." arXiv preprint, May 2026. Source.
- "RoboPhys-3D." arXiv preprint, August 2026. Source.
- "WoW-World-Eval." arXiv preprint, January 2026. Source.