The problem, stated without machines
Describe a robot lifting a cup off a table. The lift changes the cup's height, and the description handles that: one fact updates. But the table is still there. The room is still lit. The door is still shut. None of these facts were mentioned by the action, and all of them, presumably, remain true. A reasoner working from a formal description of the action needs some way to know that.
The obvious fix is to write down the persistence explicitly: lifting the cup does not move the table, does not switch off the light, does not open the door. This works for one action and three facts. It fails as a strategy the moment the world has more than a handful of things in it, because the number of persistence statements needed is the product of every action and every fact that action might have touched but didn't. A world with a hundred objects and a dozen actions needs, in principle, over a thousand explicit statements just to say what stays the same. Add a new fact to the world and every existing action needs a new clause added to its description, saying that this action doesn't disturb the new fact either. The bookkeeping does not converge.
The deeper trouble is not the volume of bookkeeping. It is that persistence cannot generally be derived from the facts you have. Logic tells you what follows from what is stated. Nothing follows about a fact that was never stated in connection with the action at all. A cup lifted says nothing, logically, about a table's position, a door's latch, a light switch. You can add axioms asserting that most things stay put by default, but a default is a guess wearing the clothes of a proof, and it fails exactly in the cases where something genuinely did change and nobody thought to mention it. The problem is not computing consequences of an action. It is bounding what a description is silent about, and logic gives you no principled way to bound silence.
Where it came from
John McCarthy set the terms in 1963, in a paper called "Situations, Actions, and Causal Laws," introducing the situation calculus: a way of writing actions and their effects as statements in first-order logic, indexed by successive "situations." It was meant to let a reasoning machine plan by proof. In 1969, McCarthy and Patrick Hayes, extending that framework, gave the specific difficulty a name: the frame problem. They had found that stating what an action changes is straightforward. Stating everything it leaves alone is not, and the gap grows without limit as the described world grows.
Philosophers noticed the problem generalised. Daniel Dennett, writing in 1984, treated it as a problem of relevance in cognition generally: any reasoner, biological or artificial, faced with a changed world, must somehow ignore the astronomical majority of facts that the change didn't touch, without first checking each one. That is a broader claim than McCarthy and Hayes made, and it drew its own objections, but it kept the core insight intact: derivation is expensive where confirmation should be cheap.
Ray Reiter offered a formal answer in 1991: successor state axioms, which compress the persistence problem into a single axiom per fluent, stating exactly the actions that affect it and asserting, in effect, that everything else leaves it alone. This works, and it is elegant. It works under one condition: that the axiomatiser has correctly and completely listed every action that could possibly affect that fluent. That is the closure assumption. It is exact in a blocks world with ten objects and four verbs. It degrades sharply anywhere the set of possible actions is not fixed in advance, which is almost everywhere else.
The turn
Here is where the concept stops being a puzzle in formal logic and starts being an argument about intake. The frame problem is, at bottom, a statement about the cost of establishing non-change from a static description. That is exactly the situation a system is in whenever its knowledge of the world is a record fixed at some earlier point and it must now say something about the present.
A Large Language Model is such a system by construction. Its corpus is frozen at a training cutoff. Everything after that cutoff is, from the model's position, an unmentioned fact: not stated to have changed, not stated to persist, simply absent. The model has no channel by which the world could report otherwise, so it has only one option, structurally: assume persistence and speak as though the frozen state still holds. This is the frame problem in its maximal form, inherited whole, not by design flaw but by the nature of a corpus.
A Large World Model changes the picture, but only locally. It observes a scene directly, and while that scene is within its sensors, the frame problem simply does not arise for it: the table's presence is not inferred from an earlier lift, it is seen, continuously, for as long as the camera holds it in frame. This is why embodied perception feels so much easier than symbolic reasoning ever did — the hard cases were always the unmentioned ones, and here nothing needs mentioning because everything relevant is in view. But the moment attention moves — the camera pans, the sensor cone shifts — the scene behind it reverts to an unmentioned fact. The frame problem does not disappear. It moves to the edge of the frame, literally.
A Large Universe Model is the position defined by refusing to let that edge exist by policy: every relevant stream stays open, and every belief carries a timestamp and a source. Persistence is no longer assumed by default and no longer derived by axiom. It is re-witnessed, continuously, and a belief that has gone unwitnessed for some interval is marked stale rather than silently carried forward as true. This does not eliminate the underlying difficulty that logic identified in 1969. It relocates the difficulty from an unanswerable question — does the unmentioned still hold? — to an answerable one: when did we last check, and by what.
What narrows the claim
Three objections deserve to be taken at full strength, because each identifies real ground the argument does not cover.
The first: observation is always partial, so the boundary of the sensed set is where the frame problem returns, and everything outside that boundary still has to be assumed persistent. This is correct, and it is not a minor concession. The gain from continuous intake is not omniscience. It is that the boundary between witnessed and inherited belief becomes visible and measurable rather than invisible and assumed. A system with provenance can say, of any given belief, how it knows this and how long ago it last checked. That converts an undecidable logical question into a bounded, priceable one: a coverage gap with a number attached, rather than a silent assumption with none.
The second: planning requires reasoning about futures that have not happened, and no sensor reports on a future. This is entirely true, and it marks the honest limit of the whole argument. Continuous streams do nothing for counterfactual persistence — what a contemplated action will disturb before it is taken. What changes is the correction loop: predictions about what an action will leave alone can be checked against the streams moments after the action, rather than left to stand uncorrected for as long as the record is frozen. The frame problem in prediction is not dissolved. It is converted into a statistical estimation problem with fast feedback, which is a different, and more tractable, kind of hard.
The third: Reiter's 1991 successor state axioms already solved the representational frame problem within logic, given a well-specified action theory, so the case for more intake is overstated. This is true within its closure assumption and false the instant that assumption is violated. Blocks worlds hold still. The qualification problem — that no finite list of preconditions ever fully covers when an action succeeds — and the ramification problem — that actions have indirect effects the axioms never anticipated — survived Reiter's solution untouched. The famous Yale shooting problem showed that even inside a toy world, default persistence can produce the wrong answer to a simple question about a loaded gun. Formal closure buys correctness relative to a model of the world, and only for as long as the model stays complete. Streams buy correctness relative to the world itself.
The misreading to disown
The common error runs the other way from all three objections: it says continuous observation abolishes the frame problem outright. It does not, for the reason the second objection gives in full — prediction is not perception, and no stream reports on what has not yet happened. Nor is the frame problem a computational bottleneck that faster hardware eventually overwhelms; it was never a speed problem. It was a problem about deriving an absence of change from an incomplete description, and no quantity of processing over that description fixes an incompleteness intrinsic to it.
What this does and does not establish
The concept establishes an asymmetry, not a solution. Non-change is intractable to derive from a closed record and cheap to confirm by watching. That asymmetry explains why intake, not raw inferential power, is the axis that matters across the lineage from Large Language Model to Large World Model to Large Universe Model, and why the third position is the last rung on that particular ladder: once every relevant stream is open and every belief carries its own provenance and decay, there is no further category of access left to add. What remains — coverage, latency, cost, trust in the provenance itself — are quantities to argue over, not a new kind of intake waiting to be discovered.
It does not establish that open streams make a system a good planner, a good predictor, or a good reasoner about actions not yet taken. It does not establish that partial sensor coverage becomes complete coverage by virtue of being tracked. It only establishes where the frame problem sits once persistence stops being assumed and starts being witnessed: at the honest edge of what is currently being watched, instead of hidden inside a description that has gone quiet.