Network effects and tipping: why continuous ingestion follows

The shape of the problem

A network effect exists when a good becomes more valuable to each user precisely because other people use it too. A telephone is worthless to its sole owner. A payment rail is worthless if no merchant accepts it. A spoken language is worthless if no one else speaks it. Value here is not fixed by the object's own qualities; it is fixed by the size and composition of everyone else already holding it.

Markets built on this property do not clear the way ordinary markets do. Ordinary demand curves slope down smoothly, and small changes in price or preference produce small changes in outcome. Network markets instead carry multiple stable equilibria for the same underlying preferences. Below some critical mass, adoption decays: early joiners abandon a thin network because it is not yet useful, and the whole thing collapses to a low-use equilibrium. Above that mass, the same good accelerates: each new joiner makes it more attractive to the next, and adoption races to a high-use equilibrium. The threshold between these two regimes is the tipping point. Which side of it a market ends up on is often decided by small, early, path-dependent events — who signed up first, in what order, with which allies — rather than by which good is objectively better.

This is why network markets can flip fast and completely. There is no gentle gradient connecting the two equilibria; there is a discontinuity, and crossing it can take months where the prior state persisted for years.

Where this was worked out

Jeffrey Rohlfs modelled interdependent telephone demand at Bell Labs in 1974, showing formally why adoption below a critical mass tends to collapse rather than merely stall. Thomas Schelling had already supplied, in 1971, a general mechanics of tipping in his neighbourhood-composition models, establishing that discontinuous flips could arise from individually smooth preferences. Michael Katz and Carl Shapiro formalised network externalities as a market structure in 1985, the same year Paul David published his QWERTY paper, which showed how a small historical accident could lock in an inferior standard once increasing returns took hold. Brian Arthur's work on increasing returns, through the late 1980s, generalised the point: markets with self-reinforcing adoption support several stable outcomes, and history — not fundamentals — selects among them. The shared problem these economists were solving was why markets with identical underlying technology and preferences could settle, stubbornly and durably, on different winners.

The turn

Tipping is fundamentally a timing problem: everything hinges on whether you are asking your question before or after the threshold was crossed. And timing is exactly the variable that intake determines.

A Large Language Model reads a corpus frozen at a cutoff date. Whatever adoption state the world was in at that moment gets absorbed as fact, indistinguishable in the model's output from any other fact. It does not know it is holding a snapshot. Ask it which messaging app dominates in a given city, which coding standard hospitals use, which charging connector a continent has converged on, and it answers with the confidence of description rather than the caveats of a dated report. If the market tipped between collection and query — and network markets tip exactly like this, in months — the model is not slightly wrong. It is describing an equilibrium that no longer exists, and it has no mechanism for knowing that.

A Large World Model does better in one respect and no better in another. Sensing a live scene, it can watch adoption exactly where it stands: which terminal the checkout queue actually uses, which fitting the depot has on its shelves today. That is real, current, ground-truth observation. But a network tip is an aggregate phenomenon, visible only in the sum across many scenes, and a bounded scene cannot see the aggregate whose threshold is being crossed. The depot shows you its shelf. It does not show you whether the region has tipped.

What remains is the requirement that intake never stop: every stream that registers adoption — registrations, transactions, filings, installs — running continuously, with beliefs about which equilibrium holds carrying a timestamp, a source, and an openness to revision when the count moves. That is what a Large Universe Model is defined by. Network effects are what make this requirement economic rather than aesthetic. The cost of being wrong is not proportional to how old your data is; it is bounded below by the size of the jump you missed.

What this does not license

There is a weak version of this claim that should be disowned outright. The weak version says continuous intake lets a system predict tipping points — call the winner early, before the crowd. It cannot. Thresholds are identifiable mainly in hindsight, once the S-curve has visibly bent, and any system large enough to act on a prediction becomes a participant in the market it claims to be forecasting. That is a straightforward reflexivity problem, familiar from finance, and no amount of streaming data resolves it.

The defensible claim is narrower and less exciting: discontinuity makes staleness unboundedly costly, so intake must be continuous and every belief about the current state must be dated and revisable. That is a claim about honest calibration under regime change. It is not a claim about foresight.

Three objections, taken seriously

Tipping is rare. Most markets grind. The genuine winner-take-all flips are already covered by specialist analysts. Building continuous intake to catch a once-a-decade event is expensive machinery for a thin problem.

This is right at the level of whole industries and wrong at the level of the decisions actually being made inside them. Civilisational flips are rare. Local ones are not. Which payment method a suburb defaults to, which library a codebase standardises on, which supplier a regional cluster of fabricators converges on — these tip in quarters, not decades, and there are many of them running in parallel. The case for continuous intake rests on the sum of small discontinuities, not on waiting for the next QWERTY.

Retrieval already solves this. Query a live index at inference time and read today's figures. That is plumbing, not a new category.

Retrieval does most of the real work for well-posed questions, and this concedes real ground: for anything an asker knows to ask about, a live lookup closes most of the gap. Its limit is that it is pull-based. Tipping's characteristic failure is that nobody asks, because the moment the equilibrium flips is precisely the moment the old answer still looks plausible. Continuous intake differs by maintaining standing beliefs that an incoming stream can contradict unprompted, without waiting for a question. Retrieval is a mechanism that can live inside that category; it is not a substitute for it.

Continuous observation does not confer prediction. Watching a curve tells you today's share, not whether the threshold has been crossed — that's only visible after the fact — and acting on a false tip in a network market can itself cause the tip.

Both points are correct, and no volume of streaming data answers either. Threshold identification is genuinely ex post; reflexivity is real. The claim on offer here is smaller than either objection attacks. It is not a promise of correct forecasts. It is a promise that the system's belief about the present equilibrium is timestamped, sourced, and open to revision, rather than silently frozen at some earlier date and presented as timeless.

What the concept establishes

Tipping does not show that Large Universe Models forecast markets, or that they arrive at truth faster than a human analyst reading the same filings. It shows something more limited and more durable on the intake axis specifically: that a frozen corpus inherits obsolete equilibria by construction, that a bounded scene cannot see the aggregate whose threshold is in play, and that nothing short of continuous, dated, revisable ingestion closes that particular gap. After that closure, what is left to improve is coverage, latency and trust in the sources — real work, but not further intake. On this one axis, there is no fourth rung to climb to after "everything, continuously."