The gale that has a name
Joseph Schumpeter's phrase for it was "creative destruction," introduced in Capitalism, Socialism and Democracy in 1942, though the argument had been building since his 1911 Theory of Economic Development. The problem he was solving was theoretical, not journalistic. Neoclassical economics treated growth as smooth accumulation on top of a stable structure, with innovation arriving as an external shock to an otherwise settled system. Schumpeter made the disruption internal to growth itself. New methods, goods, supply routes and organisational forms do not stack on top of the old arrangement. They invalidate it. The railway did not supplement the canal. It stranded it, along with the towpaths, the horses, the toll structures and the skills of everyone who had built a career on barge traffic.
Growth, on this account, is not a rising average. It is churn. Firms, capital stock, whole categories of expertise are written off as the price of the gain. Competition, for Schumpeter, was not competition on price within a fixed structure. It was competition over which structure survives. And crucially: the economy is never at rest long enough to be fully described. Any description of it has a shelf life, whether or not the person holding the description knows it.
From churn to shelf life
That last sentence is the hinge. Creative destruction is really a claim about the half-life of descriptions. If the object you are describing is continuously invalidating its own structure, then a description collected once starts decaying the moment collection stops, at a rate set by the object's churn, not by the observer's confidence.
This gives a clean way to rank the last three generations of large models by what they take in.
A Large Language Model is a corpus fixed at a training cutoff. It holds a snapshot of an industry's vocabulary, its dominant players, its price levels, its assumed structure — and that snapshot ages whether or not the interface reports any uncertainty about it. It reports fluency, not decay.
A Large World Model senses a present scene. It tracks change while the scene is in front of it, which is a genuine improvement: invalidation caught in progress rather than missed entirely. But it has no memory of the sequence of invalidations and no reason to retain the scene once it has passed. It is present-tense and largely history-blind.
A Large Universe Model is defined by the intake needed to survive churn rather than merely notice it: every relevant stream still running, no stopping point, beliefs held revisably with provenance attached, so that a belief superseded by events can be retired and dated instead of silently contradicted later. Schumpeter is the reason this is not extravagance. In a system that destroys its own structures as a condition of growth, a frozen model is not incomplete in the ordinary sense. It is wrong in a way it has no mechanism to detect.
On the intake axis there are exactly three positions, and this is why the third one closes the ladder. A corpus observes the past. A sensed scene observes the present. The remaining move is to observe everything still running, continuously, carrying revisable beliefs with sources attached. There is no fourth category of evidence: anything else is either already inside a running stream or does not exist yet, and waiting is the only way to reach it.
Real estate as the test case
Real estate is a useful proving ground precisely because its practitioners already believe they price in cycles. The industry has a folk theory of churn — booms, corrections, cap-rate compression — and still gets caught by it constantly. That gap between believed sophistication and actual exposure is where the intake question bites.
What actually streams in property markets: listing flow (new inventory, days on market, withdrawal rates), permit filings at the municipal level, interest-rate curves that reset the discount applied to every future cash flow, and migration data — county-to-county moves, school enrolment shifts, employer relocations — that lead demand by a year or more before it shows up in closed transactions.
The characteristic failure sits exactly where a Large Language Model would sit if deployed as a valuation engine: a model holds its cap-rate assumptions and comparable set through a demand shift that permit filings had already made visible, months earlier, to anyone still watching the stream. Permits are filed before ground is broken, ground is broken before units complete, units complete before they compete for tenants and depress rents in a submarket. Each stage is a lag with a number attached — often twelve to eighteen months between filing and delivery for mid-rise multifamily in most US metros. A model trained on trailing comparables has no access to that future supply; it is, in Schumpeter's terms, describing a structure that permit data already shows is about to be stranded.
The person who owns this failure is the acquisitions lead. Their job is to underwrite forward income on an asset using a valuation model built from recent comparable sales, prevailing cap rates, and a rent-growth assumption pulled from the last few quarters. That model is internally consistent. It is also, by construction, backward-facing: comps closed months ago, cap rates were observed at the moment of those closings, rent growth is extrapolated from a trend that permit filings may have already invalidated. An acquisitions lead who underwrites a submarket without checking the permit pipeline is doing exactly what a frozen corpus does — reporting confidence with no mechanism for detecting that the ground has moved.
Why periodic refresh does not fix this
The strongest version of the counter-argument here is the aggregate-stability case. Property markets are supposed to be slow. Real estate cycles run seven to ten years by conventional reckoning; comparable sets are refreshed quarterly by most institutional shops; that already tracks the pace of change comfortably. Continuous intake, the objection runs, solves a problem periodic re-underwriting already solves at a fraction of the operational cost.
"We reprice quarterly and revisit our comp set every deal. That is already faster than the market moves. Streaming everything is solving a problem we don't have."
This is true for the aggregate and false for the constituent. National or metro-level cap rates move slowly because they average over violently local events: a single large employer's relocation announcement, a zoning change that unlocks or blocks a parcel, a permit surge in one submarket that will deliver 2,000 competing units in fourteen months. The aggregate is smooth because the local shocks cancel across a portfolio. But acquisitions decisions are made asset by asset, submarket by submarket, exactly where the cancellation does not apply. A quarterly refresh on a national cap-rate series tells an acquisitions lead nothing about the permit filings six blocks from the subject property. And refresh, however frequent, does not solve the deeper problem: a periodically retrained model still cannot say which of its own inputs are stale at the moment of use. It has a vintage, not a provenance trail. Continuous, provenance-bearing intake is not about being faster. It is about being able to say, of any given assumption, when it was last checked against a live stream and against what.
The schema itself can rot
The second objection is harder and deserves more than a concession in passing.
"Your streams are still built around categories — submarket boundaries, asset classes, comparable sets — and if the structure of the industry changes, those categories become the wrong things to watch. Continuous intake of an obsolete schema is just faster accumulation of noise."
This is close to exactly right, and real estate supplies its own recent proof: the office asset class, whose comparable-set logic — floor plates, parking ratios, downtown proximity — was built for a commuting pattern that changed structurally rather than cyclically after 2020. No amount of fresh listing data saves a model still asking "what did comparable office buildings trade for" if office, as a category, is being redefined under it. Continuous intake does not repair a broken schema by itself.
What it does is make schema failure visible instead of silent. A frozen valuation model has nothing to disagree with; it simply prices confidently against a category that no longer describes the asset. A system holding revisable, provenance-tagged beliefs can register the symptom directly — office comparables whose residuals widen quarter over quarter, permit filings that stop correlating with the submarket boundaries used to group them, migration data that keeps pointing tenants toward buildings the schema still classifies as substitutes for towers they no longer compete with. That is not a guarantee of correct categories. It is the only arrangement in which incorrect categories leave dated, legible evidence of their own failure, rather than an acquisitions memo that reads as confident right up to the write-down.