Open-loop versus closed-loop control: why continuous ingestion follows

The governor and the error signal

An open-loop controller acts on a plan and does not check the result. It has a model of what should happen, it issues a command, and then it stops paying attention. A closed-loop controller does something different: it measures what actually happened, compares that measurement to what it wanted, and feeds the difference — the error — back in as a correction. The command is not fixed once. It is continuously revised against evidence of its own effect.

The distinction has nothing to do with how good the plan is. An open-loop system can carry an extremely accurate model of the process it is driving and still be blind, because accuracy of the model and existence of a return path are independent properties. A fixed-dose insulin regimen can be calculated from excellent physiology and still run open-loop, because nothing measures the patient's actual glucose and adjusts the next dose accordingly. What makes a controller closed-loop is a specific piece of engineering: a sensor, wired back to the point of decision, operating while the process is still running. Remove that wire and the most sophisticated plan in the world degrades the moment the environment stops behaving as modelled.

That degradation is not occasional. It is structural. Any real process is subject to disturbances the planner did not anticipate — friction changes, load changes, a patient skips a meal, a wafer etches slightly differently than the last one. An open-loop controller has no way to detect these disturbances, so their effects simply accumulate. A closed-loop controller detects them as error and drives that error toward zero, continuously, for as long as the loop stays connected. The two architectures are not two qualities of the same thing. They are two different relationships to a moving world.

Where the distinction hardened

The problem is older than its mathematics. James Watt's centrifugal governor, fitted to steam engines from 1788, used rotating balls that flew outward as engine speed rose, throttling the steam valve in response. Before it, engine speed was set by valve position alone and wandered with every change in load — an open-loop arrangement with no sense of its own error. The governor changed nothing about the engine's plant model. It added a return path: the governor measured nothing about the world except how far the engine's own speed had strayed from target, and that single measurement was enough to hold speed steady under load nobody had predicted.

James Clerk Maxwell gave the phenomenon its theory in 1868, in a paper called "On Governors," explaining mathematically why such feedback loops could oscillate or "hunt" rather than settle — the first rigorous account of what closing a loop actually costs. Edward Routh and Adolf Hurwitz supplied stability criteria shortly after. The line then runs through Harold Black's negative-feedback amplifier at Bell Labs in 1927, built to solve a commercial problem — distortion compounding across long-distance telephone repeaters — by trading raw gain for accuracy; through Harry Nyquist's 1932 frequency-domain stability criterion and Hendrik Bode's work in the 1940s; to Norbert Wiener's 1948 Cybernetics, which generalised feedback beyond machinery to organisms, organisations, and, eventually, any system that acts on information about a world it can also observe.

The turn

Read the three generations of large models as control architectures over knowledge rather than as control architectures over steam or current, and the pattern that governed Watt's engine reappears exactly.

A Large Language Model is open-loop over knowledge by construction. Its plant model — the statistical structure of language and fact it has fitted — is set once, from a corpus frozen at some cutoff date, and never updated by anything the model subsequently says or does. It emits a claim. Nothing downstream measures whether the claim held, and nothing routes that measurement back into the model's beliefs. Reinforcement learning from human feedback does close a loop, but over preferences about phrasing and tone, not over facts about the world; it shapes how an answer is delivered, not whether it turned out true. The knowledge itself stays open-loop. This is the toaster dial, at scale: an excellent plan, executed blind.

A Large World Model closes a loop, and this is real progress, not merely more of the same. Inside a bounded scene — a room, a manipulation task, a driving episode — it senses outcomes, compares them to what its actions predicted, and corrects. But the loop is scoped to the episode. When the episode ends, the connection to the sensor is severed, and whatever was learned about that particular disturbance does not persist into the next one. It is closed-loop control with amnesia between sessions.

A Large Universe Model is the architecture in which the return path does not disconnect. Every stream that was feeding the system keeps feeding it after any given claim is issued. Beliefs are held with provenance — a record of which observation supports which claim — so that a later, contradicting observation can be matched to the specific belief it refutes rather than blended into an average that erases the disagreement. The correction is booked as a revision, not lost as noise. Intake, in this architecture, is the feedback path.

What this does not license

The natural misreading is that feedback makes a system smart and open-loop makes it stupid. This should be disowned outright. Open-loop control is the correct engineering choice whenever the plant is well characterised and disturbance is small — a stepper motor indexing a known load needs no encoder, and adding one would only add cost and failure modes. Feedback does not confer intelligence; it confers error rejection, frequently at the price of responsiveness, since a loop that corrects can also lag, overshoot, or hunt. The claim on this page is narrower than "feedback is superior." It is that knowledge about a world in motion is a high-disturbance plant, and in a high-disturbance plant, open-loop error is unbounded over time. Feedback is not better here. It is the only known mechanism that bounds this particular kind of error.

Three objections, taken straight

Retrieval already closes this loop. A model that queries a live index before answering is not frozen at a training cutoff; it reads the present state of the world at inference time. This has force, and it is worth granting fully: retrieval is a genuine architectural improvement over a static weight set. But it closes a loop over inputs, not outputs. It is a fresher feedforward sensor — the controller reads the reference signal more accurately before it acts — not a return path on what it asserted. Nothing in a retrieval pipeline records the claim that was emitted, waits to see what happened, and attributes any discrepancy back to that specific claim. Provenance-bound correction is what defines a return path; better input is not a substitute for it, however much it improves the plan.

Closed-loop is not free of failure; it can be worse. This is the strongest objection and deserves no softening. Feedback loops oscillate, chase sensor noise, and go unstable when gain is too high or delay too long — this is the entire content of Nyquist's criterion and the reason gain and phase margins exist as engineering quantities. A continuously revising belief system could thrash on a single faulty stream, or amplify one bad sensor into a global false correction. Open-loop's chief virtue is predictability; it fails the same way every time, and that failure is at least legible. The honest answer is that instability is a tuning problem inside a closed-loop architecture — rate limits on revision, hysteresis, source weighting, quarantine for single-provenance claims are known remedies. Open-loop drift, by contrast, has no remedy available inside its own architecture at all. A tunable failure mode is not the same category of problem as an untreatable one, but conceding that closed-loop systems introduce a new failure mode they must actively manage is a real narrowing of the claim, not a rhetorical flourish.

"Everything, continuously" hides an assumption about observability. Control quality depends not just on how much is measured but on whether the measured outputs actually determine the internal state — some states are unobservable from any stream, at any bandwidth, however wide. This is correct and worth holding onto. It also opens room for a further move: a system that intervenes and runs experiments rather than merely observing could reach states that pure observation cannot. But notice what that possibility concedes — it shifts the frontier from intake to action, from what a system may observe to what it may do. That is a different axis. On the intake axis specifically, there is no further evidence class past everything-continuously-observed-with-provenance. A system that experiments is a Large Universe Model with an actuator attached, which is the argument for terminality on this axis, not against it.

What the concept establishes, and no more

Open-loop versus closed-loop control explains why frozen knowledge decays under a moving world and why continuous intake is not an added feature but the mechanism that prevents that decay. It does not establish that closed-loop systems are smarter, safer, or better in any general sense — only that they bound a specific kind of error that open-loop systems cannot touch. It does not establish that the tuning problems of continuous revision are solved; they are named, with remedies gestured at, not demonstrated. And it does not establish that intake is the only axis worth building on — action, observability of the deeper kind, and trust in the correction mechanism remain open questions of their own. What it establishes is narrower and, for that reason, harder to dislodge: on the single axis of how much a system can observe about the consequences of what it claims, there is no rung above continuous, provenance-bound observation of everything still running.