Two lessons ago we stood over a lump of uranium and watched a single kilogram decide the fate of a city. Last lesson we did the daring bit: we lifted the idea clean off the reactor floor and showed that any reinforcing loop has a per-step gain — call it k, the average number of new events each event causes — and that critical mass is just k crossing 1. Below 1 the process loses more than it gains and fizzles; above 1 it gains more than it loses and runs away.
That was the abstraction. This lesson is the payoff: the grand tour. We’re going to walk the same threshold through epidemics, startups, marketplaces, crowds, chemistry, forests, and a painfully awkward meeting — and every single time it’s the same shape wearing a different costume. By the end you should be slightly unable to stop seeing it.
Before you read — take a guess
A disease has a basic reproduction number of R0 = 3 — each infected person, in a fully susceptible population, infects about three others on average. Public health wants to push the outbreak sub-critical by immunising enough people. Roughly what fraction of the population needs to be immune to tip it below the threshold?
That pretest is the whole lesson in miniature: you reasoned about a virus using a rule you first met in a bomb. Let’s do it properly now, one domain at a time.
Epidemics — the reproduction number
Start with the domain that made “R number” a dinner-table phrase. The amplifying unit is an infection, and its k has a famous name: the basic reproduction number, written (say “R-nought”). It answers exactly the k question — on average, how many new people does one infected person infect, in a fully susceptible population?
- The amplifying unit: one infected person.
- k here: — secondary infections per case.
- Below 1: each case produces fewer than one new case, so each generation of the outbreak is smaller than the last. It shrinks to zero. The outbreak fizzles.
- Above 1: each case produces more than one new case, generations grow, and — compounding — you get an epidemic.
The clever public-health move is that you don’t have to touch itself to win. You only need to push the effective reproduction number — call it , the number actually achieved once some people are immune — below 1. Every immune person is a dead end for the chain: a neutron that escapes instead of splitting the next atom. Immunise enough of the population and the average infection can’t find enough fresh hosts to reproduce itself. That’s herd immunity — not everyone protected, just enough dead ends that the chain reaction can’t sustain.
How many is “enough”? Here’s the worked version. The chain sustains when . Set that equal to 1 and the critical susceptible fraction is ; everything above that must be immune. So the herd-immunity threshold is :
| Susceptible fraction that just sustains it () | Fraction you must immunise () | |
|---|---|---|
| 2 | 1/2 = 50% | ~50% |
| 3 | 1/3 ≈ 33% | ~67% |
| 4 | 1/4 = 25% | ~75% |
| 5 | 1/5 = 20% | ~80% |
| 12–18 (measles) | ~6–8% | ~92–95% |
Read the last row and you understand why measles is the boss fight of vaccination: its is so high that you need almost everyone immune to keep the effective number under 1. Let coverage slip a few points and pops back over the threshold — the same knife-edge you dragged the slider across two lessons ago.
R0 is not destiny — it's a starting number
describes a virus meeting a fully susceptible population. The number you actually live under is the effective reproduction number , which behaviour, immunity, and season all bend. Distancing, masks, and vaccines don’t lower the virus’s intrinsic ; they lower the achieved k below 1. Same threshold, different lever.
Viral content and growth — the viral coefficient
Swap atoms for shares and you get the growth-hacker’s version. The amplifying unit is a sharer (or an inviter), and its k is the viral coefficient: on average, how many new sharers does each sharer recruit?
- k below 1: every wave of shares is smaller than the last. The thing spreads only as far as you pay to push it — turn off the ad spend and it dies. This is most content.
- k above 1: every wave is bigger than the last, so it spreads on its own, no fuel required. This is the rare genuinely “viral” hit.
The difference between “we have to keep buying reach forever” and “it grew while we slept” is, once again, a number sitting a hair above or below 1. Marketers obsess over the viral coefficient for precisely the reason physicists obsess over k: it’s the switch between decay and runaway.
Network effects, moats, and the adoption S-curve
Here critical mass stops being about how fast something spreads and starts being about whether the thing is worth anything at all. A network effect means the product gets more valuable to each user as more users join — one telephone is a paperweight, a million telephones is infrastructure.
- The amplifying unit: a user who, by joining, makes the product more valuable and so attracts more users.
- k here: how many additional users each new user pulls in (via the value they add).
- Below the critical mass of users: the product is nearly worthless — nobody to message, nothing to buy — so people churn out faster than they arrive. k < 1, and it bleeds to death.
- Above the critical mass: each user makes it valuable enough to attract more than one more, who attract more still. k > 1, and it compounds into something people can’t leave.
That “can’t leave” is where durable moats come from (a model with its own course nearby): once a network is past critical mass, a competitor has to lure users away from something that’s valuable precisely because everyone is already there — a coordination problem that protects the incumbent. The moat isn’t a feature; it’s the far side of a tipping point.
But nothing compounds forever, and this is the part people forget. Plot adopters over time and you don’t get a straight line to infinity — you get the adoption S-curve: a long slow crawl while you’re sub-critical, then an explosive near-vertical rise once k clears 1, then a flattening as the pool of not-yet-adopters runs dry. That last bend is the same physics as the fission grid saturating once most of the fuel is already spent: runaway growth needs unburnt fuel, and eventually there’s none left. k falls back below 1 not because the loop broke but because it ran out of things to amplify.
Three phases, one curve
The S-curve is critical mass told as a story in time. Sub-critical crawl (k < 1, or barely above and starved of adopters) → super-critical explosion (k > 1, plenty of fuel) → saturation plateau (fuel spent, k drops back under 1). Lesson 4 lives inside the first two phases — why that long flat crawl fools us into calling a winner dead.
A messaging app and a broadcast-only news app both launch in the same city. The messaging app is useless with 200 users and unstoppable with 2 million; the news app is exactly as useful to its very first reader as to its millionth. Which has a genuine critical-mass / network-effect threshold, and why?
Crowds — Granovetter’s threshold model
Now the most human costume, and the one that best repays study. In 1978 the sociologist Mark Granovetter asked why two nearly identical crowds can behave completely differently — one erupts into a riot, the other goes home — and built the threshold model to answer it.
The idea: each person carries a personal threshold — the number (or fraction) of others who must act before they’ll join in. A firebrand has a threshold of 0 (needs no one). A cautious soul has a threshold of 50 (won’t move until half the room already has). The amplifying unit is a person joining, and k is roughly how many additional people each joiner tips over their own threshold. What decides whether a spark cascades or fizzles isn’t the average temperament — it’s the distribution of thresholds.
Granovetter’s classic contrast makes it vivid. Picture 100 people in a square:
- Crowd A: thresholds are 0, 1, 2, 3, … 99 — one person at every level. The person with threshold 0 starts. That makes one actor, which trips the threshold-1 person, making two, which trips the threshold-2 person… and it unzips all the way to a full-blown riot.
- Crowd B: identical except the one person with threshold 1 is swapped for a second person with threshold 2. Now the threshold-0 person starts — one actor — but nobody has threshold 1, so the chain stalls at one. The instigator stands alone and sheepishly stops. No riot.
Same 100 people, near-identical dispositions, one person’s threshold different by a single point — and one square burns while the other yawns. The unsettling takeaway is pure critical mass with a nod to emergence: the aggregate (“an average person of moderate militancy”) tells you nothing about the outcome, because the behaviour lives in the chain of interactions, not in the individuals. You cannot read the riot off the people; you have to run the cascade.
The same shape wears many crowd costumes:
- A standing ovation needs a critical mass of early risers before the seated majority feels the pull; below it, the brave few sit back down.
- A bank run is a threshold cascade over “withdraw before it collapses” — each withdrawal makes collapse likelier, tripping the next depositor.
- A riot, a bandwagon, a fashion trend — all the same: personal thresholds, a spark, and a distribution that either carries the cascade or strands it.
Don't read the crowd off the people
The seductive error is to infer temperament from outcome: “they rioted, so they must be an angry mob.” Granovetter’s point is that a mob and a docile crowd can be statistically identical in disposition — the outcome hinges on the arrangement of thresholds, i.e. on the interaction structure, not on how militant people are on average. Same trap as reading Schelling’s segregated city as proof of bigotry. The macro pattern is not a magnified copy of the micro intention.
Autocatalysis — chemistry’s self-lighting fire
Chemistry has the purest version of all. An autocatalytic reaction is one whose product catalyses its own formation: making some of substance X speeds up the making of more X. The amplifying unit is a molecule of product; k is how many further reactions each product molecule accelerates into being. Below the threshold the reaction crawls; once enough product accumulates to catalyse itself, it takes off — a chemical chain reaction. This is a leading candidate for how life bootstrapped: a set of molecules that collectively catalyse each other’s formation is a self-amplifying loop that, once over critical mass, sustains and grows itself with no designer required — critical mass at the origin of everything.
Forest fires — the percolation threshold
Back to something you can picture burning. Take a landscape and vary the density of trees. A fire starts at one tree and can only jump to neighbours within reach.
- The amplifying unit: one burning tree.
- k here: how many new trees each burning tree ignites — which depends almost entirely on how tightly packed the trees are.
- Below a critical density: the gaps between trees are too wide, the fire can’t reliably jump, each burning tree lights fewer than one more, and the fire dies in a small scorched patch. k < 1.
- Above the critical density: neighbours are close enough that fire reliably hops tree-to-tree, each burning tree lights more than one more, and the blaze sweeps edge to edge. k > 1.
The sharp line in tree density where “small scorch” flips to “total burn” is a percolation threshold — the point where local connections suddenly link up into one system-spanning cluster. It’s the exact same phase change as the fission grid, just with the knob being how connected the fuel is rather than how fissile each unit is. This is why forest managers thin trees and cut firebreaks: they’re deliberately dragging the landscape back below its percolation threshold, breaking the connectivity so k stays under 1.
You can feel this in the grid from lesson 1. Seed it sparse and run — the reaction can’t bridge the gaps and dies. Seed it dense and run — it percolates across the whole board. Same rule, same seed strength; only the connectivity changed.
Below the threshold, above the threshold
Percolation: connectivity is the knob
Each square is a scrap of fuel; the coloured ones start the reaction. A reacting cell fires at its neighbours, triggering on average k new reactions before it burns out. Set k below, then step or run — hunt for the tipping point where “fizzles” flips to “runs away”.
Step 0: 13 cells reacting now; 13 ever activated (3% of the grid). Reacting…
The everyday one — the silent meeting
Lest you think this is all bombs and epidemics: you sat inside a tipping point at your last bad meeting. The presenter finishes, asks “any questions?”, and nobody speaks — not because nobody has one, but because each person has a private threshold for how many others must speak up first before they’ll risk looking foolish. The room is sub-critical: everyone’s waiting for the amplifying unit that hasn’t fired.
Then one person asks the awkward question. Suddenly a second feels safe, then a third, and the room that was dead silent won’t stop talking. One brave soul with a threshold of zero tipped the whole room over critical mass — the standing ovation, run in reverse and in miniature. The model isn’t only for physicists; it’s why “just say the thing” is genuinely high-leverage, and why the first honest voice in a nervous room is worth ten later ones.
Same shape, many costumes — the master table
Every domain above is the identical machine with the labels swapped. Line them up and the pattern is impossible to unsee:
| Domain | The amplifying unit | What “k” is | Below 1 (fizzles) | Above 1 (runs away) |
|---|---|---|---|---|
| Nuclear fission | A splitting atom | Neutrons that go on to split another atom | Sub-critical lump — reaction dies | Bomb / self-sustaining reactor |
| Epidemic | An infected person | / — new infections per case | Outbreak shrinks to zero | Epidemic grows |
| Viral content | A sharer | Viral coefficient — new sharers per sharer | Needs constant paid pushing | Spreads on its own |
| Network product | A user | Extra users each user attracts | Product worthless, users churn out | Compounds into a moat |
| Crowd action | A person joining in | People each joiner tips over their threshold | Instigator stands alone | Riot / ovation / bank run |
| Autocatalysis | A product molecule | Reactions each product accelerates | Reaction crawls and stalls | Chemical chain reaction |
| Forest fire | A burning tree | Trees each burning tree ignites | Fire dies in a small patch | Percolates edge to edge |
| The silent meeting | A person who speaks up | People each speaker emboldens | Awkward silence holds | Room won’t stop talking |
The transfer heuristic falls right out of the table. Whenever you meet anything that might spread — a disease, a rumour, a product, a protest, a habit, a fire, a fashion — run the same three-part probe:
- What’s the amplifying unit? (The thing that, by existing, tends to create more of itself.)
- What’s its k? (On average, how many new units does each unit cause?)
- Which side of 1 are we on? (Below → it dies; above → it runs away; near 1 → it’s on a knife-edge and small changes decide everything.)
Get those three and you can predict the shape of the trajectory before you know a single domain-specific detail. That’s the whole point of a mental model: portability.
Each scenario is a self-amplifying process. Sort it by whether its k is below 1 (will fizzle out) or above 1 (will run away).
- A measles outbreak (R0 ≈ 15) hitting a community with only 60% vaccine coverage.
- A protest square where the lone instigator finds nobody else has a low enough threshold to follow.
- An app whose every user, on average, invites 1.4 friends who join and also invite.
- A trading floor where each seller triggers, on average, two more panicked sellers.
- A marketplace past its critical mass, where each new buyer draws in more than one seller.
- A newsletter that only grows when you buy ads; each subscriber refers 0.3 new ones.
- A sparse woodland after heavy thinning, trees too far apart for fire to reliably jump.
- A disease with effective reproduction number Rt = 0.8 in the current population.
Every self-amplifying process has an 'amplification factor' — the average number of new events each event causes. Match each domain to what plays the role of k there.
Tipping-point transfer test
A virus has R0 = 5. A health authority wants to drive the effective reproduction number below 1 through vaccination. Roughly what fraction of the population must be immune, and what's the underlying logic?
Check your answer to continue.
Key takeaways
- One shape, many costumes. Epidemics (), viral coefficients, network effects, crowds, autocatalysis, and forest fires are all the same threshold — k, the new-events-per-event factor, crossing 1.
- The herd-immunity number is . Remove that fraction of susceptibles and each case causes fewer than one new case: the outbreak goes sub-critical. , .
- Network effects create moats past critical mass, and the adoption S-curve is critical mass told over time: crawl, explosion, then a saturation plateau as the fuel runs out.
- Don’t read the crowd off the people. Granovetter’s threshold model shows near-identical crowds can riot or disperse purely on the distribution of thresholds — the macro outcome is not a magnified copy of the average disposition.
- The transfer heuristic: for anything that spreads, ask what’s the amplifying unit, what’s its k, and which side of 1 are we on?
Where this goes next
You can now spot the threshold in the wild, whatever it’s dressed as. But spotting it and feeling it in time are different skills — and humans are catastrophically bad at the second. A super-critical process spends most of its life looking like nothing is happening, because exponential growth hides in the noise until it doesn’t, and our stubbornly linear intuitions get ambushed every time.
That’s lesson 4: Why Change Feels Sudden. Why the long quiet phase is indistinguishable from failure, why “nothing, nothing, nothing — then everything” is the native rhythm of a threshold rather than a surprise, and why the people who see change coming are simply the ones who learned to trust the maths over their gut. See you there.