communicoupling

Why the worse one
wins forever.

QWERTY was laid out in the 1870s so that mechanical typebars would not jam. The jams vanished within a generation; the layout is on the phone in your pocket. VHS beat the better tape. The meeting sits on Tuesday because, once, it had to. Everyone involved can see the better option from here — so why can nobody move us onto it?

The pattern is path dependence, and W. Brian Arthur gave it mechanics: under increasing returns to adoption — when each user makes an option more valuable to the next — early random events are amplified instead of averaged away; the process locks in, and what it locks in can be the worse option. Arthur proved it for competing technologies (1989); Paul David told it as the story of QWERTY (1985); beneath both sits a toy from 1923, the Pólya urn.

Arthur's contrast is the sharp edge. Under diminishing returns, markets forget their accidents: an early lucky lead invites correction and washes out. Under increasing returns they remember forever — the process is non-ergodic, and history is the destiny rather than noise around it. Below, two engines run live: the bare urn, where accident hardens into fate, and Arthur's adoption market, which prices what that costs when one option was genuinely better.

increasing returns · rvalue each adopter adds for the next
0.050
quality gap · ghow much objectively better B is
+0%
switching cost · spain of abandoning a choice made
0.20
ensemble · Mparallel histories per burst
200
this history
adopter 0 / 300
worse tech wins
equal quality — nothing is the worse option
unpredictability
bits of outcome entropy · 1.00 = coin flip
Engine one · the Pólya urn — history's skeleton
draw a ball, return it with a twin of the same colour · every urn settles on a share; which share is itself pure chance
0 urns
this urn — share of gold balls past urns right: final shares after 400 draws — flat
Engine two · Arthur's adoption market
300 adopters arrive in random order and choose A or B by taste + quality + r × installed base · right: where the histories ended
history 1 · adopter 0/300 · ensemble 0
this history — share using A past histories, coloured by winner lock-in round B's late upgrade
Watch what happens
Two equal technologies under real network pull. Each history decides itself in its first few dozen adopters, then spends the remaining ones confirming. Press Run 200 histories and watch the coin land both ways across the ensemble.
U(option) = quality + r · adopters(option) Arthur's increasing-returns payoff, tastes spread ±0.4. At r = 0 every choice stands alone and accidents cancel; at r > 0 each adopter makes the same choice easier for the next, and the opening coin-flips compound into the standard.

History amplified

Two engines, one skeleton.

The urn is the whole idea with the economics stripped off. Start with one gold ball and one blue. Draw at random, return the ball with a twin of the same colour, repeat. An early run of gold makes gold likelier to be drawn, which adds more gold, which makes gold likelier still: success breeds success, mechanically. Two facts hold at once, and holding them together is the concept. Every single urn settles — its share converges and stays. And the share it settles on is itself random: for this urn, uniform, meaning any final split is exactly as likely as any other. Run the ensemble and the histogram comes out flat. The average across urns says one half; almost no individual urn says that. A process like this is called non-ergodic — the ensemble average tells you nearly nothing about the history you are actually inside.

Arthur's market dresses the urn as a standards contest. Each arriving adopter has a private taste and reads the installed base; utility is quality + r·n. The moment one option's lead satisfies r · lead ≥ taste + gap, even the laggard's most devoted possible fan does better joining the leader — that is the lock-in round, marked in gold on the canvas, and every choice after it is ceremony. At r = 0 the market forgets: deviations self-correct and quality decides every run. Raise r and the market starts remembering — and what it remembers, the worse-tech readout shows, is sometimes the wrong thing, permanently.

What to try

Three experiments in frozen accident.

01

Randomise destiny

Press Run 200 urns two or three times. The histogram of final shares stays flat — a 90/10 urn is as likely as a 50/50 one — while every individual line visibly settles. Certain convergence, random destination: that pairing is the whole lesson.

02

Toggle the physics

Set the gap to +20%. At r = 0, burst the ensemble: the worse tech wins 0% — always. Now r = 0.08 and burst again: about one history in twenty ends with the worse standard entrenched. Quality unchanged; only the returns regime moved.

03

Try to defect

Run the fair fight until the gold lock line appears, then drag switching cost to 0. Exit is free — and nobody leaves. With a 20-adopter lead at r = 0.05 the network term is 1.0, more than any taste (±0.4) can outweigh. The wall is everyone else.

The Tuesday meeting

Most of the furniture of life is a frozen accident.

Increasing returns is the normal condition of social life, because "everyone already knows it" is a form of quality. A convention gets more valuable with each person who holds it: the keyboard layout, the office suite, the language of air-traffic control, the citation classic every later paper must cite to be legible. Smaller and closer to home: the couple's division of chores, settled in the first exhausted month and never renegotiated; the committee's agenda order; the workplace tool nobody would choose today that everyone must learn tomorrow. Each of these was once a live contest among alternatives. Each locked when its lead made the alternatives academic — and the lock-in round rarely announces itself. It looks like an ordinary Tuesday.

The ensemble view is the concept's gift to judgement. An installed base is evidence of arrival order, and only weakly of merit: Betamax was never refuted, merely outnumbered. That reading cuts both ways — humility about incumbents, and mercy for the losers, including the excellent colleague whose better method arrived after the department had standardised on a worse one. It also explains a fact about commerce that otherwise looks irrational: launches are fought like wars because, under high r, the entire contest is the first dozen adopters. Being right early beats being right. And it sets the price of escape honestly: after lock-in, persuasion is aimed at individuals for whom switching alone is genuinely irrational. What breaks a lock is coordination — everyone moving at once — which is why standards fall to consortia, regulators, and generations, and almost never to arguments.

The mapping

Mechanism ↔ social life.

In the modelIn the world
the urnAny process where success breeds success — standards, skills, cities, habits, reputations.
r · adoptersCompatibility, network effects, training, "everyone already knows it".
arrival orderHistorical accident: who happened to show up early, and in what mood.
lock-in roundThe moment the alternatives became academic; every later choice is ceremony.
the ensembleThe histories that didn't happen — the worlds that got the better standard.
quality gap gThe margin by which the loser was better.
the late upgradeWhy being right early beats being right.

Where it tears

Limits.

Lock-in is not a life sentence.

Real standards do fall: technological discontinuities reset r to zero (the touchscreen dissolved QWERTY's mechanics, though the layout swam across), consortia and governments coordinate switches, and generations simply stop inheriting. VHS beat Beta and then died itself. The model contains no such shocks unless a preset injects one — its "forever" holds only inside a fixed regime, and regimes end.

Proving real-world inefficiency is hard.

Showing that an actual standard is inefficiently locked — rather than merely resented — requires a counterfactual measurement nobody has. David's QWERTY story is itself contested: Liebowitz and Margolis argued the typing-speed evidence for Dvorak was far weaker than the legend. The simulation demonstrates that inefficient lock-in is possible and cheap to produce; it cannot certify any particular standard as an instance.

Increasing returns is a scope condition, not a law.

Congestion, saturation, and boredom put plenty of choices under diminishing returns, and those markets genuinely forget their accidents — the r = 0 dial is a real place, where restaurants, crops, and commodity suppliers live. Diagnosing lock-in everywhere is its own comfortable error: sometimes the incumbent keeps winning because it is better.