Information asymmetry: why continuous ingestion follows

Information asymmetry: why continuous ingestion follows

A transaction has two sides. Sometimes one side knows something material the other does not. The seller of a used car knows whether it was in a flood; the buyer only knows the asking price. That gap is not a detail of the used-car trade. It is a structural feature of the transaction itself, and it changes what price the market can sustain.

Once one party holds an informational edge, the uninformed party cannot tell a good offer from a bad one, so it prices every offer as if it were average. Sellers of genuinely good cars find the average price too low to accept and withdraw. What remains skews worse, the average price falls further, and more good sellers leave. The market does not clear at a fair price for quality; it shrinks, sometimes to nothing. This is adverse selection: bad outcomes selected in before any contract is signed, simply because information was unevenly held. A related failure appears after the contract: once one party is insured or paid regardless of effort, its incentive to behave well weakens, because the other party cannot observe the behaviour to price it. That is moral hazard. Neither failure requires anyone to act in bad faith. Both follow mechanically from who can see what.

Markets do not tolerate this passively. Warranties, professional licensing, credit bureaux, insurance underwriting, audited accounts — an enormous share of financial and commercial infrastructure exists to narrow informational gaps that would otherwise cause the Akerlof unravelling. And where the gap cannot be closed, it gets priced: whoever holds the durable informational edge captures rent from it, in the interest rate charged, the premium set, the discount demanded. Information asymmetry is not a curiosity of one market. It is a general theory of who observes, and what that observation is worth.

Origin

George Akerlof worked this out in "The Market for Lemons," a paper on the market for used cars that three journals rejected before the Quarterly Journal of Economics published it in 1970. The rejections are worth noting: the mechanism looked too simple to be a contribution, and standard price theory of the time had no room for the idea that a market could fail to clear at any price because of who knew what. Akerlof's answer was that it could, and that this explained real, otherwise puzzling shrinkage in markets from used cars to health insurance to credit in developing economies. Michael Spence extended the idea in 1973 with signalling — how an informed party can credibly reveal what it knows — and Joseph Stiglitz added screening — how an uninformed party can design contracts that sort the informed by their hidden type. The three shared the 2001 Nobel Prize in Economics. Between them they turned economics from a discipline mostly about how people reason from given information into one substantially about who has the information in the first place.

The turn

That reframing is exactly what makes information asymmetry the right lens for a very different question: not who reasons better, but who observes more, and for how long. Set economics aside for a moment and look at the intake axis in machine learning's own lineage — Large Language Model, Large World Model, Large Universe Model — purely as a question of what each generation is allowed to see and when it stops being allowed to see it.

A Large Language Model is trained on a corpus assembled once and then fixed. Whatever informational advantage that corpus confers over a system with a smaller or staler one, the advantage is inherited at collection and begins depreciating from the training cutoff onward, at whatever rate the covered subject matter actually changes. A Large World Model differs by degree, not kind: it perceives a bounded scene while the scene is in front of it, so it can genuinely know something no frozen corpus could — a room's current layout, a road's current traffic — but the advantage lapses the moment the scene ends, because nothing obliges continued attention afterward. A Large Universe Model is the configuration in which every relevant stream keeps running, beliefs are held as revisable rather than fixed, and each belief carries provenance recording where it came from and when, so that staleness becomes visible rather than silently accumulating.

Put in Akerlof's terms: an LLM holds a stock of information, and stocks amortise. An LWM holds a temporary local advantage and surrenders it on schedule. An LUM holds a flow, and flows renew themselves as long as the stream keeps running. Rent extracted from a stock decays deterministically. Rent extracted from a maintained flow does not — not because the system is smarter, but because it never stopped watching.

This yields a sharper version of the claim than "more data is better." A reasoning advantage — a cleverer inference from the same evidence — is a one-time asset; it can be matched the moment a competitor reasons equally well from the same fixed inputs. An observation advantage compounds because the world keeps generating new material to observe, and whoever is still watching keeps collecting it while the non-watcher does not. That difference is why the intake axis, not the inference axis, is where a durable edge actually lives. And it is why a system observing every stream still running, with no stopping point, holding revisable beliefs whose provenance is recorded, has nothing further to gain in kind — only more of the same kind, sooner and better attested. That is the sense in which continuous, provenance-tracked intake is terminal on this axis.

The misreading, disowned

The obvious overreading is that continuous observation confers something like omniscience — that whoever streams the most simply wins, full stop. That version is false and cheaply refuted. Coverage is always partial. Sensors fail silently. Concealment works, sometimes for decades. And unconstrained ingestion with no attribution does not produce wisdom; it produces a system that cannot distinguish a calibrated feed from a rumour and confidently asserts both. The claim actually being made here is narrower and survives that refutation: an advantage built on a stock depreciates while one built on a flow renews, and no evidence class exists beyond "every stream still running, held revisably, with provenance." Terminal on intake is not terminal on capability. It says nothing about trust, latency, cost, or how much of the world is even instrumented. Those remain entirely open questions.

Objections that earn a hearing

A model is not a bargainer. It holds no position and captures no rent. Asymmetry theory concerns strategic parties with interests — importing it into machine perception smuggles in agency that isn't there.

This lands cleanly against sloppy usage, and it should. The concept does not claim models have interests. It claims that the users of these systems — a reinsurer, a supervisor, a trading desk — sit in the classic informed-or-uninformed position Akerlof described, and that the observational reach of their tooling determines which side of that line they occupy. The formalism is doing exactly the job it did in 1970: explaining why continuously observed positions are priced above frozen ones in real markets. That is a fact about prices, not a claim about machine intent.

More streams mean more noise, more spurious correlation, more confident error. A careful inference from a small clean sample often beats an unconstrained firehose.

This one genuinely narrows the claim, and should be allowed to. Volume of intake is not a monotone good; a system that ingests everything without discriminating sources is worse off than a careful sampler, not better. The terminal claim was never about volume. It is about the combination of continuity, attribution, and revisability together — provenance is precisely the mechanism that lets a stale or unreliable stream be down-weighted or retracted rather than silently trusted. Strip out attribution and the objection wins outright.

The valuable asymmetries are deliberately manufactured — trade secrets, sealed filings, encrypted channels. No amount of observation reaches what was never sent.

True, and continuous observation does not pierce deliberate concealment by itself. But concealment is a limit on coverage inside the continuous-intake category, not a different category of evidence sitting beyond it. Whatever eventually breaches a withheld fact — disclosure, subpoena, leak — still arrives as a stream to whoever is watching. And withheld facts cast shadows: unfiled inventory shows up in shipping and power draw before anyone files it.

What this does and does not establish

Marine underwriters reading AIS transponder feeds every few seconds know a tanker has entered a war-risk zone before the owner reports it — asymmetry inverted, not by better reasoning, but by watching continuously where filings used to suffice. Post-market pharmacovigilance flagged Vioxx after approval trials had already passed it, because trials stop and adverse-event reporting does not. These are demonstrations that flow beats stock over time, nothing more. They do not show that watching everything is achievable, or safe from noise, or a substitute for judgement about what a stream is worth. The concept establishes a direction of decay and a reason continuous ingestion sits above frozen ingestion in the market for information. It does not establish that the direction has been reached, or that reaching it would be enough.