Language models predict tokens, physical reasoning needs causes, so world models will take over the parts of AI that touch the real world. The premises are mostly right. The conclusion people draw from them, that switching architectures fixes hallucination, does not follow.
What is actually being built
Three efforts get cited together, and each one does something a chat model does not.
- JEPA, the joint embedding predictive architecture associated with Yann LeCun, predicts future state in a learned representation space rather than reconstructing raw pixels. The bet is that discarding unpredictable surface detail leaves representations that support planning.
- DeepMind's Genie 3 generates interactive environments you can act inside. One consequence is that the model's implied dynamics become observable through interaction instead of staying internal.
- World Labs' Marble generates navigable 3D scenes from limited input. The target is spatial structure, not sentences about space.
LeCun's departure from Meta to found AMI Labs came with reported funding around 500 million euros at a valuation near 3 billion euros. That is not evidence that the architecture works.
Where the error goes instead
Hallucination is a property of generation without verification. Change the modality and the failure follows the generator.
A language model that invents a case citation is cheap to catch, but only under conditions: a resolvable index exists, and the claim is a citation rather than a characterization of what the law holds. Citations are discrete, addressable, and checkable against a register. Characterizations are not.
A world model that rolls out a manipulation plan in which a box's center of mass sits two centimeters off, or a gripper clips a shelf edge for three frames, produces output that a reviewer watching a rendered rollout will usually wave through, and that low grade automated inspection, meaning frame level anomaly scores and trajectory smoothness heuristics, will not flag either. The error is continuous, spread across a trajectory, and rendered persuasively.
The counterargument, which is strong
Continuous trajectories admit checks that free text does not. Two centimeters of center of mass error is exactly what a rigid body simulator catches. Joint limits, contact forces, and collision geometry are what constraint solvers were built for. Intent in a contract is checkable by nothing comparable. Judged by verifier cost, the metric this piece is arguing for, embodied domains can score better than linguistic ones rather than worse.
So here is the narrower claim I will defend. World model errors are less legible to human reviewers, because rendered output gets graded on whether it looks physical. Machine legibility is available but not free: the physics checker is itself a model with its own error, and a passing sim replay only predicts a passing execution to the extent the sim to real gap is small and characterized. Where validated physical checkers exist, embodied work is the better verification bet. Where they do not, plausible rendering conceals error more effectively than fluent prose does.
That leaves grounding and auditability as a conjectured near term tradeoff rather than a structural one. My mechanism is about how outputs get reviewed, not a theorem about representations, and it weakens as physical plausibility checkers mature.
Cheap checks do the work
The leverage sits in the rejection path rather than the inference path. Some checks that earn their cost:
- Dimensional analysis. An engineering result with mismatched units gets rejected without anyone evaluating the argument that produced it.
- Kinematic and geometric feasibility. Joint limits, reach envelopes, collision meshes. A proposed trajectory that intersects known geometry is discarded before it reaches an actuator.
- Conservation and monotonicity constraints. Mass balance in process design, non-negative inventory, energy bounds.
- Reference resolution. A legal citation either resolves to a real provision whose text matches the claim, or it does not.
- Bounded ranges. A dose outside a published range is blocked and escalated no matter how fluent the justification.
None of these are symbolic reasoning in any interesting sense. They are cheap computations with known soundness properties, and the research program label they sometimes get filed under adds nothing to the argument. What matters is cost, coverage, and whether the check is sound.
None of them care whether the proposal came from a transformer over tokens or a latent dynamics model. Inside a checker's coverage, generators are close to interchangeable. Outside it they are not: a generator whose failures fall where the check cannot see them is worse than one whose failures land inside coverage, even at identical benchmark scores. Verifier reliability is a function of the error distribution it is filtering, which is why false negative rate on your own logs is the number that matters.
Where a cheap and sound checker exists, the verifier is the load bearing component. That condition fails for a lot of work people want to automate, and that failure is the real constraint.
Pick problems by verifier cost, not by modality
Ask whether you can write a checker that is cheaper and more reliable than the generator it is checking. Linguistic versus physical is the wrong axis.
- Warehouse pick and place scores well. Geometry, payload, and force limits are machine checkable, and failure is observable within seconds.
- Structural sizing scores well. Closed form limits and code requirements give you a hard filter over generated candidates.
- Radiology triage scores partially. Cross view consistency and prior history are checkable, but ground truth still needs a clinician, so the system should route rather than decide.
- Open ended contract drafting scores poorly. Citations are checkable, intent is not, so the human stays in the loop on the parts that matter.
- Long tail driving scenarios score poorly for a specific reason: the open problem is specifying what counts as correct behavior in a rare scenario, not generating a plan for it. The checker is the research problem.
If you cannot state what would falsify an output, you do not have a high stakes deployment.
What follows, with the hedges kept in
- Language models will likely keep the interface layer and the fuzzy retrieval and normalization work. That allocation looks stable, though it is an inference from current cost curves rather than a fixed fact.
- World models are a reasonable bet for sample efficiency in embodied planning, on the theory that predicting dynamics in representation space needs fewer trials than learning policies from scratch. This is an active research bet. Public evidence so far is mostly demonstrations rather than deployed reliability numbers.
- The durable differentiator will probably be the verification stack, at least in domains where a sound checker can be written at all. Elsewhere the differentiator is probably human workflow design, which is a different and less discussed problem.
Three questions worth asking any vendor in this space
- Show me an output your system rejected, and name the check that rejected it.
- What runs after generation, and what is that check's false negative rate on your own logs.
- For the failure modes you have no checker for, who reviews them, and against what reference.
Those answers are cheap to give when a rejection path exists and hard to improvise when it does not.