The Price of Knowledge Friedrich Hayek made two arguments,…

The Price of Knowledge

Friedrich Hayek made two arguments, three decades apart. Genuine Bitcoin and AI are the two halves of the machine he was describing.

By MrBen

In 1945 Hayek published "The Use of Knowledge in Society." The problem he set was not technical but epistemic. The knowledge a society runs on is not held in one place. It is scattered across millions of minds, much of it tacit, particular to a time and a place, impossible to state in full to anyone else. No central authority can gather it, because most of it stops existing the moment you try to write it down. The price system, he argued, is the mechanism that coordinates this dispersed knowledge without anyone needing to possess it. A price is a message. It carries just enough of what everyone else knows for you to act well without knowing what they know. Mises had reached the same wall first, in 1920: without prices, a central planner cannot compute an efficient allocation, because it has no access to the information prices carry.

In 1976 Hayek published "Denationalisation of Money." A different subject, or so it appears. Currency should be taken out of the state monopoly and issued by competing parties, disciplined not by political promise but by the same competitive pressure that disciplines any other product. Bad currency would be abandoned for good. The argument is usually filed under monetary economics, a curiosity, a footnote with a libertarian odour.

These are taught as two separate Hayeks. The knowledge one and the money one. They are not separate. They are one argument about coordination, told twice: once about knowledge, and once about the medium that knowledge needs in order to coordinate at all.

The central planner returns as a model

The dominant project in artificial intelligence rests on a premise worth stating plainly: that intelligence improves as knowledge is concentrated. Gather enough of the world's information into one system, apply enough computation, and the result will exceed the dispersed judgement of the people the information came from. The bigger the model, the closer to the whole.

This is the calculation problem rebuilt in silicon. Mises told a generation of planners that they could not compute the economy, because the information they needed did not exist in any form they could collect. The same limit holds here. A model trained on everything that has been written knows what has been written. It does not hold the tacit knowledge in a working engineer's hands, the local knowledge of a market that exists only at a particular counter on a particular morning, the proprietary knowledge a firm will never publish. The result is a system that knows a little about almost everything and understands the particular almost not at all.

I want the strong form of this, not the easy one, because the easy one is false. The claim is not that large models are useless, or that centralized intelligence is impossible in principle. They are remarkable tools. The claim is narrower and harder to escape: a single system cannot aggregate dispersed and tacit knowledge, and the attempt to do so concentrates not intelligence but ownership. The question is not whether the model is clever. It is who holds it, who is paid for it, and whether the alternative has anywhere to stand.

The market alternative needs a medium

Hayek's answer to dispersed knowledge was never a better planner. It was a market: many specialists, each acting on what only they know, coordinated through prices. Carry that across to machines and the shape is immediate. Not one model that knows everything, but many agents that each know something, built by the people who hold that knowledge, competing and cooperating and checking one another, priced by whether they work.

But a market is not a market without settlement. A price is only a message if something changes hands when it is accepted. Human markets hide this, because settlement is slow, batched, and handled out of sight by banks. A market of machine agents transacting thousands of times a second, each service metered and paid for as it is rendered, has no such luxury. It needs settlement at the speed and the granularity of the transactions themselves. A price signal with nothing to settle in is not coordination. It is conversation.

This is the hinge. The 1945 argument, taken to machine scale, generates a requirement it cannot satisfy on its own. It needs a medium.

The medium cannot be the state's

What kind of medium? Not state currency. Not for reasons of ideology but for reasons of mechanics. Settlement at machine scale must be permissionless, because no agent can wait for an account to be opened. It must cost almost nothing per transaction, because the transactions are tiny and constant. It must be final, because there is no court sitting between two machines. And it must be neutral, because the moment the medium is controlled by one party, the market clearing on top of it is no longer free. State currency f…