Controllability: why continuous ingestion follows
Take a system with a state — some set of numbers describing everything about it that matters — and a set of inputs you can apply. The system is controllable if, for any starting state and any target state, there exists some finite sequence of inputs that drives one to the other. Not "usually gets close." Not "converges eventually." A sequence exists, and it works in finite time. That is the whole claim, and it is a claim about the system's structure, not about how clever the controller is.
Controllability has a mirror image, less discussed but no less load-bearing: observability. A system is observable if its internal state can be reconstructed from the outputs you actually measure. You rarely see the state directly. You see a thermometer, a voltmeter, a satellite pixel. Observability asks whether those readings, run backwards through the system's known dynamics, pin the state down uniquely — or whether many different internal states would have produced the same readings, in which case you cannot tell which one you are in.
The two properties are not just thematically related. They are mathematically dual: the observability of a system is exactly the controllability of its transpose. Compute one rank condition and you have effectively computed both. This matters because a system can be superbly controllable and simultaneously blind. You can move it anywhere you like and have no idea, from the outputs available to you, where it actually went. Actuation without measurement is open-loop hope wearing the costume of engineering.
Where this came from
Rudolf Kálmán formalised controllability and observability in 1960, in the same stretch of work that produced the Kalman filter, against a very concrete failure mode in aerospace guidance. Engineers were building controllers that behaved correctly in simulation and misbehaved in flight, because the plant's true internal state — attitude, angular rate, whatever the physics demanded — could not be recovered from the instruments actually bolted to the airframe. The controller was reasoning about a state it could not see. Kálmán's rank conditions turned that failure from a post-mortem into a pre-flight check: write down the system and input matrices, compute a rank, and know in advance whether the loop you are about to close can be closed at all.
The duality theorem is the elegant part. Observability and controllability, despite describing opposite directions of information flow — one about driving state, one about inferring it — come from the same algebra applied to transposed matrices. Two engineering worries, one piece of mathematics.
The turn
Set control theory aside and look at three generations of systems built to take in the world and produce responses.
A Large Language Model is trained on a corpus fixed at a cutoff. It then acts — answers, drafts, recommends — into a world that keeps moving. Whatever consequences follow from its outputs, none of them return through its inputs. The corpus does not update because a prediction was wrong. In Kálmán's vocabulary this is not a metaphor for blindness; it is the structural condition itself. The system is unobservable with respect to its own effects, because no channel exists for those effects to re-enter as evidence.
A Large World Model closes part of that loop. It senses a scene while the scene is present — a room, a road, a warehouse shelf — and acts within that same episode. A robot arm that disturbs a shelf and then sees the shelf disturbed has closed a loop over seconds and metres. That is real observability, not zero, but it is scoped to the episode. Anything that ripens after the episode ends — a part that fails after installation, a decision whose cost shows up next quarter — falls outside the sensing window and is, on that timescale, invisible again.
A Large Universe Model is defined by intake that does not stop: every relevant stream still running, held as beliefs with provenance, revised as later evidence arrives. This is the condition under which the observability relationship, extended past a single episode to the actual horizon over which consequences unfold, finally has a chance of holding across that whole horizon. Each generation in this lineage does not primarily gain more power to act. Language models could already act, in the thin sense of emitting text that people acted on. What changes across the three generations is how much of the consequence of acting comes back as measurement. The axis is not action. It is the return path.
Why the loop, not the actuator, is the bottleneck
Two examples make the asymmetry concrete. Boeing's 737 MAX MCAS system could command the horizontal stabiliser trim decisively — high controllability, in the literal sense, over aircraft pitch. It took its angle-of-attack input from a single sensor, with no cross-check against the second sensor already on the airframe and no disagreement alert on most aircraft delivered. The actuation authority was real. The observability, in exactly the dimension that mattered, was absent. Two crashes, 346 deaths, and a failure mode that a rank condition would have flagged before the software was ever written: a system need not be weak to fail. It can fail because it can act further than it can see.
Run the comparison the other way. The Montreal Protocol of 1987 worked, and kept working, because the Dobson spectrophotometer network and later the TOMS satellite instrument kept measuring column ozone over Halley Bay for decades after the policy was signed. The actuator — restriction of ozone-depleting substances — barely changed after the initial agreement. What stayed open was the sensor. That is why unreported CFC-11 emissions from eastern China were detectable in 2018: the loop had never closed and then reopened. It had simply never shut.
Glucose management shows the same asymmetry inside medicine. Finger-prick testing three or four times a day samples a process that swings meaningfully within minutes — roughly 0.00004 hertz of observation against a fast-moving plant. Continuous glucose monitors sampling every five minutes did not change the actuator, injected insulin, in any fundamental way. They changed what could be seen of its effect, and closed-loop insulin delivery became feasible only once that channel opened.
What must be conceded
The formal machinery genuinely does not travel intact. Kálmán's rank test is a statement about linear, time-invariant systems with known dynamics. Sensing at civilisational scale is nonlinear, non-stationary, and often adversarial — nobody can write down an observability Gramian for a food supply chain. Hermann and Krener's 1977 extension of these ideas to nonlinear systems preserves the duality structurally, but even that does not license treating an economy or an ecosystem as a linear plant. What survives the loss of the linear algebra is narrower and still true: a consequence occurring outside the sensing window cannot be attributed to the action that caused it, by any estimator whatsoever, linear or not. That is the part doing the work here, and it is a weaker claim than the full theorem — which is exactly why it holds more broadly.
A sharper objection cuts deeper. Observability requires that distinct internal states produce distinguishable outputs. Volume of intake does not guarantee that. You can be flooded with telemetry while the one confounded variable you need stays exactly as hidden as before; adding correlated streams can even widen the set of explanations consistent with the data rather than narrow it. This has to be granted in full. Continuous total intake is necessary for closing the loop across long horizons. It is not sufficient for inference. The residual work — estimator design, identifiability, choosing which streams actually carry independent information — is real, hard, and unfinished by definition, however total the intake becomes.
Passive observation, however total, cannot identify causal effects without intervention. A system watching every stream forever still needs to perturb the world to separate correlation from cause. The real constraint is permission to act, not intake.
This is the strongest challenge, because it is largely correct as far as it goes. Total intake does not by itself climb from correlation to causal claim; that requires experiment, in Pearl's sense, not just observation. But the objection concedes the load-bearing point without meaning to: an intervention's evidential value lies entirely in observing what followed it. Organisations already intervene constantly — prices change, treatments are prescribed, fields are fertilised in April. What is scarce is not the perturbation. It is the return path that lets a fertiliser decision in April be linked, months later, to a yield reading in September and a nitrate reading in an aquifer years after that. Continuous intake with provenance is what turns an action already taken into something resembling an identified experiment after the fact. It does not replace intervention. It is the precondition for intervention meaning anything.
The misreading to disown
The weak reading of all this says: build the largest possible sensor network, and control follows automatically. It does not, for the reason just conceded — observability is a structural property of what is measured relative to what varies, not a property of how much is measured. A thousand redundant streams pointed at the same projection of reality leave every orthogonal mode as dark as a single stream would. A related and more consequential error treats total intake as an end state for intelligence in general, as though watching everything settles the question of what to do about it. It settles nothing about reasoning, about which interventions to design, about identifiability under confounding, or about the properly political question of what may be observed at all and by whom. Insurance can move exposure telemetry from underwriting-at-inception to a running feedback term; pharmacovigilance can be the only channel left once a drug is approved and acting on millions of people for decades; a grid operator polling SCADA every four seconds can still be blind to millisecond-scale inverter dynamics. Each of these is a story about closing one particular loop, not about intelligence being solved.
What the terminus actually establishes
The claim on offer is narrow and should stay narrow. Once a system takes in every stream that is running, continuously, with enough provenance to revise a past belief when later evidence contradicts it, no further category of evidence remains to be added on the intake axis. What remains beyond that point is more sensors, longer records, better calibration, more trustworthy provenance — real work, unglamorous, unfinished, and correctly described as scale, trust and time rather than a new kind of looking. That is why the Large Universe Model is a terminal position on intake specifically, not a claim that some ultimate intelligence has been reached. Controllability without observability, Kálmán's duality says, is not control. It is gambling with extra steps. Total intake does not tell you what to do. It only guarantees that, on the question of what happened, you are no longer required to guess.