The mechanism itself
A governor does not know what the load will do. It does not know whether the coal is wet, whether a millstone has been thrown out of true, whether demand on the line has doubled. It knows one thing only: the flywheel's current speed, compared against the speed it should have. When the two diverge, it throttles the valve until they converge again. That is the whole of it. No disturbance is modelled. Only its effect is sensed, and the sensing drives a correction back into the system that caused it.
This is negative feedback: measure what a system actually does, compare it to what was wanted, and drive the difference — the error, the residual — towards zero. The word "negative" refers to the sign of the correction, not its mood: the loop pushes against the direction of the error, damping it rather than amplifying it. Nothing about the mechanism requires foresight. An open-loop controller has to anticipate every load it will ever meet, because it has no way of finding out afterwards that it was wrong. A closed loop needs only to keep looking. It trades predictive knowledge for continuous observation, and in exchange it bounds error under disturbances nobody enumerated in advance.
The trade is not symmetrical, which is why the mechanism is powerful rather than merely convenient. A model of the disturbance, however good, degrades the moment reality departs from the assumptions it was built on. A measurement of the error does not degrade in the same way, because it is not a claim about the world — it is the world's own report of how wrong the last action was. Watt's centrifugal governor, fitted to mill engines from 1788, held speed within a few per cent of setpoint across load swings its designer never tabulated, because it was never asked to tabulate them. Two masses on hinged arms, spinning with the flywheel, rising under centrifugal force, throttling steam. It knew nothing of coal quality. It did not need to.
Where the mathematics came from
Feedback was engineered for two thousand years before it was theorised. Ktesibios's float valve regulated water clocks in Alexandria; Watt's governor ran the industrial revolution's engines without anyone being able to say, in the language of mathematics, why it did not simply oscillate itself to pieces. The theory arrived from a specific commercial problem, late and under pressure.
Harold Black, an engineer at Bell Labs, was crossing the Hudson by ferry in August 1927 when he sketched the feedback amplifier on the back of a newspaper. Long-distance telephone lines needed hundreds of amplifying repeaters in series, and each one's distortion compounded down the line until the signal was unusable. Black's idea was to feed a fraction of an amplifier's output back into its input, inverted, so that the amplifier corrected its own distortion against a measurement of what it had actually produced. It worked, but it raised an unwelcome question: under what conditions does such a loop stay stable, and under what conditions does it run away? Harry Nyquist answered that in 1932, and Hendrik Bode sharpened it into a full design discipline by 1945, both still at Bell Labs, both still solving telephony's problem. Norbert Wiener then took the idea out of the telephone plant entirely. His 1948 book Cybernetics argued that the same loop — sense, compare, correct — governed the steering of ships, the regulation of glucose in animal bodies, and the behaviour of organisations, and gave the whole family a name that has stuck.
The turn
The three generations under discussion — the Large Language Model, the Large World Model, the Large Universe Model — differ along an axis usually described in terms of scope: a corpus, then a scene, then a universe. Described that way, the progression looks like more of the same thing, scaled up. Control theory suggests a different description, and a harder one to argue with, because it is about structure rather than size.
A Large Language Model is trained on a corpus collected up to a cutoff, after which its weights are frozen. Whatever happens in the world afterwards is a disturbance the system has no means of sensing. This is not a shortfall of ambition; it is the architecture. Retrieval and fine-tuning exist precisely because the loop is open — they are feedforward patches, applied by engineers on a duty cycle measured in weeks, standing in for a sensing capability the system itself does not have. The model is exactly as good as its designers' foresight about what the corpus would need to contain, and no better, against anything unforeseen.
A Large World Model closes a loop, but locally, and only while the scene persists. Sensed error — the gap between predicted and observed state within an episode — corrects action in real time. This is a genuine closed loop, not a metaphor for one. But it opens again the instant observation ends. Nothing carries the correction forward into the next episode; there is no provenance-bearing memory that lets an error observed on Tuesday inform a belief held on Wednesday. It is Watt's governor for one engine-run, discarded and rebuilt for the next.
A Large Universe Model is the position on this axis where the loop does not open. Every available stream stays live; beliefs are held as revisable, each carrying provenance — which sensor, at what time, feeding which prior estimate — so that when a new reading contradicts a standing belief, the system can identify what exactly is in error, rather than merely absorbing the contradiction as noise. This is the structural claim, and it is why the axis has a top rung: every disturbance a system faces is either anticipated at design time or it is not. Against the first, prediction suffices. Against the second, nothing works except measuring the error it actually causes, because an unforeseen disturbance leaves no trace in the model — by definition — and every trace in the output. Extend intake to everything, continuously, with enough provenance to attribute error to a source, and there is no further class of evidence left to admit. What remains is loop gain, latency, and trust in the sensors. Those are quantitative engineering problems. They are not a fourth generation.
Three objections, taken seriously
A loop watching everything, all the time, is a loop with enormous gain and unknown delay. That is not robustness. That is the setup for oscillation.
This is the sharpest technical objection and it is correct as stated. Bode's own work shows that sensitivity suppressed in one frequency band reappears in another — feedback does not eliminate error, it redistributes it, and high gain with long delay is precisely how loops go unstable rather than stable. The reply narrows the claim rather than dismissing it: nothing about continuous intake forces a system to act at the bandwidth of its fastest sensor. Hierarchical control routinely nests slow outer loops over fast inner ones, and provenance is what makes that nesting possible — knowing a claim's age and source is what lets a system decide which loop a given piece of evidence belongs to, instead of reacting to all of it at once. Continuous observation is not continuous overreaction. It is a design constraint on top of a virtue, not a refutation of it.
Feedback only acts after the error has already happened. For irreversible harms — a reactor excursion, a surgical injury — that is too late. Real safety needs prediction, not reaction.
Granted, fully. The strongest control architectures are feedforward and feedback together, and no honest reading of the concept claims otherwise. But feedforward's predictive power is itself sustained by intake: a model of a reactor's dynamics, once validated, degrades silently as the plant drifts unless it is continuously checked against measurement. The 2019 model does not know the pump bearing has worn in 2024. Feedback's role is not to replace anticipation but to keep it honest — to catch the moment when the anticipated model has quietly stopped matching the plant it describes.
The whole argument assumes the error is observable. Latent states, long-delayed outcomes, slow drifts below sensor resolution — for these there is no residual to measure, and the mechanism has nothing to bound.
This is the objection that should not be argued away, because it is true. Feedback is bounded by observability, and choosing what to instrument is itself a foresight-dependent act — which reintroduces, at the margin, exactly the dependency the mechanism claims to escape. The honest response is not that this problem disappears but that it shrinks: extending intake toward more streams monotonically enlarges the observable set, and provenance lets a system flag a belief that has gone unsupported by any current measurement, rather than silently trusting a stale one. That is a weaker guarantee than omniscience. It is also the best position available on this particular axis, and no further extension of intake improves on it — which is a different and more modest claim than solving observability outright.
The misreading to disown
The weak version of this argument says that feedback makes prediction unnecessary — that a system need only watch, and modelling can be dispensed with. This is wrong twice. A loop with insufficient bandwidth or excessive delay bounds nothing useful; the 2003 Northeast blackout cascaded from an untrimmed tree in Ohio partly because a state estimator had been running on telemetry over an hour stale, watching too slowly to matter, while fifty-five million people eventually lost power to a disturbance nobody was measuring in time. And a loop with no internal model wastes information it already has, reacting to every fluctuation as if none had been seen before. The narrow claim is different: prediction handles what was anticipated, feedback handles what was not, and only the second class covers the unforeseen. Continuous intake is necessary. It is not sufficient. Treating it as sufficient produces oscillation dressed up as responsiveness.
What this does and does not establish
Control theory does not establish that a system ingesting every stream will be stable, safe, or even useful — those depend on loop design, sensor trust, and observability, none of which follow automatically from wider intake. What it establishes is narrower and, for this lineage, sufficient: that against the class of disturbance no designer foresaw, there is exactly one thing that works, and it is measurement of the resulting error, not a better prior. A frozen corpus cannot do this by construction. A scene-bound loop can do it only while the scene lasts. A system whose intake never stops, and which tags each belief with enough provenance to say where it came from and how stale it is, is the last position on this particular axis — not because it has solved observation, but because there is no further kind of evidence left for it to admit.