The formula
In 1931 Alfred Korzybski told the American Mathematical Society that the map is not the territory. He meant it as a constraint, not a slogan. A map is made of different stuff than the ground it describes, drawn at a coarser grain, and built for a purpose that determines what it leaves out. Two things follow from that constraint. A map must declare its omissions, because it cannot contain everything. And a map is fixed at the moment of survey, while the territory keeps moving under it. Korzybski's real point, often missed, was that this does not make maps worthless. It makes them judgeable — by how closely their structure matches the territory's, and by how honestly they state their own scope and date.
Norbert Wiener sharpened the engineering version of this in 1945: the best model of a cat is another cat, which is exactly why any lesser model must be chosen deliberately, for a purpose, and dated. Borges pushed it to absurdity a year later with a map built at 1:1 scale, which rotted in the desert because a map that tries to be the territory stops being useful as a map. The lesson sits between those two failures. Too coarse, and the map lies by omission. Too literal, and it stops being a map at all. What's left is discipline: state the grain, state the purpose, and re-survey as the ground moves.
From formula to lineage
Read on the axis of intake — how much of the world a system takes in, and for how long — Korzybski's distinction sorts into three regimes, not two.
A Large Language Model is a map printed once. The corpus closes at a cutoff, and everything inside it is internally coherent but uniformly dated, with no record of which facts were still true when the press stopped. A Large World Model is a surveyor standing in a bounded scene with instruments live: correspondence between map and ground is excellent, continuously refreshed, but only for the patch the instruments cover, and only while the surveyor stays there. A Large Universe Model is the survey that never closes — every gauge left running, no cutoff, each belief carried with a timestamp, a source, and a decay estimate, so a claim can be re-derived from its evidence rather than simply reasserted from memory.
The physics behind this is not metaphorical. Correspondence between a map and its territory is mutual information, and mutual information decays whenever the territory has dynamics the map does not track. Decay happens for free. Maintenance does not. Only fresh observation buys correspondence back, and the bill for it never stops arriving. That is why there are exactly three ways to hold a map against a territory that keeps moving: survey once and let it drift, survey continuously but locally, or run every gauge continuously and never declare the survey finished. The third option is not a better version of the second. It removes the one thing left removable — the stopping condition. Nothing beyond "all sources, no cutoff" is a further category of evidence; it is only more of the same evidence, better instrumented.
What an acquisitions desk actually watches
Real estate is a useful proving ground precisely because it has no shortage of live gauges, and almost every failure in the sector traces to which of them a given map was allowed to see.
An acquisitions lead evaluating a submarket is, whether she names it this way or not, running a map against four streams. Listing flow tells her what is transacting now, at what discount to ask, and how fast. Permit filings tell her what is coming — units approved, broken down by type, months or years before they lease up. Rate curves set the discount rate on every pro forma she builds, and they move weekly. Migration data — net household formation, job relocations, school enrolment shifts — sets the demand side that permits are supposed to be answering. None of these four streams sits still, and none of them moves at the same speed as the others. Rate curves can move in a single Federal Reserve week. Permits accumulate over quarters. Migration shows up in tax filings a year after the move.
A valuation model built the traditional way is a Large Language Model of a submarket: a snapshot of comparables, cap rates and absorption trends, coherent as of the appraisal date, silent on how each input aged. It is not wrong to build one. It is wrong to keep using it past its usefulness without knowing that you have.
Where the frozen map fails
The characteristic failure in this domain has a specific shape: a valuation model holds through a demand shift that was visible in permit filings months earlier. This is not a data-availability problem — permit records are public, often searchable within days of filing. It is a staleness-detection problem. The acquisitions lead's model was surveyed at underwriting, and underwriting happened before the permit surge that eventually flooded the submarket with new supply. The comparables used to set the cap rate were themselves lagging indicators, since they reflected transactions that closed before the pipeline was visible. By the time absorption data confirmed the oversupply — vacancy climbing, concessions widening — eighteen months to two years had usually passed since the permits that predicted it were filed.
The failure is not that the model was inaccurate on the day it was built. It is that nothing in the model's structure could tell the acquisitions lead, six months later, that the ground underneath her cap rate had already moved. A Large World Model would do better here in one narrow sense: an underwriting team that keeps its permit-tracking live, re-running absorption forecasts each quarter against fresh filings, is surveying continuously rather than once. But that correction typically covers one market, one asset class, for the duration of the deal's hold period. The team disbands, the model closes with the deal, and the next acquisition starts its own bounded survey from zero. The locality is the limit, not the diligence of any one analyst.
| Regime | What it holds | What it misses in this domain |
|---|---|---|
| Frozen valuation model | Comparables and cap rates as of underwriting | Permit surge filed after underwriting, priced only once vacancy confirms it |
| Live deal-team tracking | Absorption and rate updates for the asset under review | Cross-market migration shifts outside the tracked submarket; resets at each new deal |
| Provenanced, standing intake | Listings, permits, rates and migration, each dated and decay-rated, across markets, continuously | Nothing structurally — the limit becomes coverage and cost, not category |
Two objections worth taking seriously
Most of the built environment barely moves. Zoning codes, floor plans, structural specifications, comparable-sales history from five years back — these decay slowly. Continuous re-survey of stable ground is wasted attention, and an acquisitions team has finite hours.
That is correct, and it is exactly why frozen underwriting models work at all most of the time. The failure in the demand-shift case was never that everything moved; it was that the model could not tell the fast-moving permit data from the slow-moving zoning data, because both entered the pro forma with the same implicit timestamp: "true as of underwriting." A model that carried provenance per input — this cap rate assumption is 90 days old, this permit count is three days old, this migration figure is a year old by definition of the data source — would let the acquisitions lead see exactly where her exposure sat. Continuous intake is not valuable because the whole submarket moves. It is valuable because it says which piece moved, and lets old, stable assumptions stay cheaply frozen while volatile ones get re-priced.
The deeper problem is categorical, not temporal. A cap rate is already an abstraction — a single number standing in for lease terms, tenant credit, capital markets sentiment and dozens of other things folded together. No amount of fresh data removes that compression. Currency was never the real gap.
Granted, and this bounds the claim rather than breaking it. Abstraction error — the cap rate flattening real heterogeneity into one figure — is fixed by declaring scope and grain, not by re-survey. What continuous, provenanced intake fixes is a different and separable failure: knowing when the abstraction, however imperfect, has gone stale relative to the ground it was built on. Terminality is claimed only for intake — for how much of the world is being watched, and for how long — not for the deeper problem of what any single number can be made to represent.
The terminal rung
Real estate does not need a model that observes everything at once to demonstrate the point. It needs one that never stops observing the four streams that already exist — listings, permits, rates, migration — and that reports, honestly, which of its beliefs are three days old and which are three years old. That is the Large Universe Model's proposition on this axis: not omniscience, but a standing survey with no designed stopping point, where staleness is a number attached to every claim rather than an unpleasant surprise discovered at the next site visit. Beyond that there is more coverage to add, more markets to instrument, more provenance to tighten. There is no further category of intake left to invent.