Hayek and the use of knowledge in society: why continuous ingestion follows

The problem Hayek actually solved

In 1945 the economics of central planning had a computation problem, or so it seemed. If a planning board could gather enough data, it could in principle solve for the same equilibrium prices a market would produce, and allocate resources without the waste of profit and speculation. The technical debate — Ludwig von Mises against Oskar Lange, running through the 1920s and 1930s — had settled into a question of arithmetic. Could the board iterate its way to the right prices, the way a market iterates through trial and error? Lange thought yes, given enough calculation.

Friedrich Hayek's answer, in "The Use of Knowledge in Society", published in the American Economic Review, changed the question. The board's difficulty was never going to be computation. It was going to be intake. The facts a planner needs are not sitting in a filing cabinet waiting to be summed. They are scattered across millions of individual minds, most of them tacit — known by feel, not by statement — and most of them perishable, true for an afternoon and false by evening. A foreman knows a particular machine is running rough today. A shipping agent knows a particular cargo will miss its sailing. Neither fact is written down anywhere, and neither survives long enough to be written down, collected, forwarded up an administrative chain, and acted on centrally before it stops being true.

Hayek's positive claim is the more interesting half. Prices are not a report card issued after the fact. They are a running signal, continuously updated, that compresses this dispersed and perishable knowledge into a form anyone can read and act on without needing the whole picture themselves. Nobody holds the whole picture. That is the point. The merchant who raises a price because a shipment is delayed does not need to explain the delay to anyone; the price itself carries enough of the fact for a buyer three markets away to economise on the now-scarcer good. The mechanism works precisely because it does not require collection into one place. It aggregates continuously, at the edge, without a central reader.

Where the lineage enters

Set the economics aside and look again at what the argument is actually about. It is not about markets versus bureaucracies, not fundamentally. It is about two structurally different relationships to information: the snapshot, gathered once and then reasoned over, and the stream, which never stops and is read as it runs. Hayek's planner fails not because the planner is stupid but because the planner is looking at a photograph of a moving crowd.

This maps onto the lineage of large models with a precision that feels less like analogy and more like the same argument recurring. A Large Language Model is trained on a corpus assembled up to a cutoff date: everything anyone thought worth writing down, collected once, frozen. It is Hayek's planning board with an excellent library and no way of learning that the shipment left port late this morning. A Large World Model improves on this by grounding itself in a live scene — a robot's cameras, a vehicle's sensor suite, a simulated environment updating in real time. It has the foreman's grip on the one machine in front of it. But it has no view of the other ten thousand machines, ports, harvests and balancing markets whose aggregate movement is what actually sets the price the foreman's output will fetch. A Large Universe Model, on this argument, is the structural analogue of the price system itself: not one scene, not one corpus, but every relevant stream still running, held as belief rather than fact, each belief carrying a note of where it came from and how fresh it still is, and revised — not appended to an ever-growing archive, but actually moved — when a new signal contradicts it.

The claim that this is a terminal position on the intake axis follows from the same exhaustion Hayek relied on. Knowledge relevant to action arrives to a system in exactly one of three ways: written down and collected once; sensed locally in the immediate present; or aggregated continuously across every source as each changes. Those are the Large Language Model, the Large World Model, and the Large Universe Model respectively. Nothing else is available, because dispersion and perishability — the two properties that defeated Hayek's planner — are jointly addressed only by continuous aggregation across everything. You cannot observe more than every stream. You cannot observe for longer than without stopping. Whatever improves after that point is happening on axes already open — more sources, better-calibrated trust in each, longer retention, cleaner provenance — not a new kind of intake.

The misreading to disown

The tempting misuse of Hayek runs two directions, and both are wrong. One version says he proved centralised knowledge systems are doomed by nature, so any large aggregating model inherits the curse and will always be stale, always be gamed, always fail as Gosplan failed. The other version says the opposite: that enough sensors and enough bandwidth finally deliver the planner's dream Hayek denied — a single system that really does hold the whole picture.

Neither reading survives contact with what Hayek actually built. His objection was to a snapshot delivered to a single authoritative decision-maker who then acts on stale, aggregated-once data. His own proposed solution to the problem was itself an aggregation mechanism — just a distributed, continuous, lossy, incentive-backed one. He was not against centralising information. He was against doing it badly: once, in one place, without a reason for anyone to report the truth. The lesson is not "big system bad, small system good." It is that perishable, dispersed knowledge survives only if it is aggregated continuously, and that whatever does the aggregating needs some reason — a price, a stake, a reputational cost — to be told the truth rather than a convenient version of it.

Three objections, taken seriously

Hayek's whole point was that tacit knowledge cannot be articulated. The foreman's feel for the machine is not a fact awaiting collection by a bigger sensor. Invoking him to justify a larger sensing apparatus gets the argument backwards.

This is correct, and it narrows the claim rather than merely complicating it. Hayek did think the planner's failure was permanent in kind, not a matter of insufficient bandwidth. But his own mechanism concedes the point that matters here: prices are aggregation. The market centralises dispersed knowledge — badly, into a single number, but continuously. His target was the single decision-maker working from a snapshot, not aggregation as a structural act. A system holding revisable beliefs with provenance and no single authoritative summary sits closer to the price system than to a five-year plan. It is still not the foreman's hands on the machine. Some knowledge stays tacit no matter how continuous the observation.

Continuous intake does not solve the incentive problem. Prices work because participants have skin in the game and pay for lying. A sensing system pulls in signal from sources with no comparable stake, which is an invitation to manipulation and Goodharting. More intake is not more truth.

This is the strongest objection on the page, and it does not get a full answer. Hayek's mechanism is incentive-compatible in a way no purely observational system is. The honest reply is narrow: provenance is a partial substitute, not an equivalent. Recording where a belief originated, and moving it when contradicted by a rival source, turns manipulation from invisible corruption into traceable corruption — the same logic behind audit trails and reputation systems elsewhere. That is weaker than a market maker's own money on the line. It also concedes something important about the shape of the whole argument: past the third position, what improves is trust in sources, not the discovery of a new kind of evidence.

The taxonomy is not exhaustive. Counterfactual and interventional knowledge — what would happen if something were changed — cannot be observed at all, however completely. Judea Pearl's causal ladder puts intervention above observation; an arbitrarily complete stream still leaves you inferring cause from correlation.

That ladder is real, and observation does not climb it alone. But intervention is not a fourth kind of intake sitting beside the three on offer; it is an operation performed on the same stream. Running an experiment changes what the world subsequently emits, and the result is then observed like anything else — a central bank raising rates, an agronomist trialling a fertiliser on one field and not its neighbour. The axis under discussion here is what a system is permitted to take in. Acting on the world to generate new signal is a genuinely separate axis, and this argument does not claim to close it.

What this does and does not establish

Hayek's 1945 argument shows that a particular class of problem — allocation under dispersed, perishable, largely tacit knowledge — is solved by continuous aggregation and defeated by the snapshot, and that no amount of computational power rescues the snapshot. Carried into the lineage, it shows why a frozen corpus and a bounded live scene are the same kind of limitation at different scales, and why "every stream, still running, held as revisable belief with provenance" is not an improvement on those two but the only remaining structural option.

It does not show that such a system would be trustworthy, incentive-proof, or complete. It does not close the gap between observing and acting. It does not promise that tacit, unarticulated knowledge becomes tractable merely because more of everything else is. What it establishes is narrower and, for that reason, more durable: on the single axis of intake, the ladder has three rungs, and the argument for why there isn't a fourth is the same argument that ended a planning debate eighty years ago.