For three lessons you’ve been watching beetles get darker. You’d be forgiven for thinking this was a course about bugs. It was never about bugs. The beetle was a teaching dummy — the cleanest possible body to watch the algorithm run on. This lesson is the payoff, the moment the whole course has been pointing at: we lift natural selection clean out of biology and watch it run on things that have no genes, no offspring, and no blood. Markets. Web pages. Jokes. Software. And, yes, the inside of your own body, right now, today.
Here’s the claim, stated bluntly: natural selection is substrate-neutral. A substrate is just the stuff a process runs on — the material underneath it. “Substrate-neutral” means the process doesn’t care which material that is. Long division gives the same answer whether you do it in pencil, in your head, or on a calculator; the procedure is indifferent to the hardware. Natural selection is exactly like that. It’s an algorithm — a fixed, repeatable set of steps — and an algorithm doesn’t care what it’s running on. Feed any system the three ingredients and you get the same blind grind toward things that fit, with no designer in sight.
So here’s the superpower you’re about to install. Walk up to anything — a company, an industry, a rumour, a chunk of code — and ask three questions: What varies? What selects? What gets copied? If you can answer all three, you’ve found a selection engine, and you can predict that it will self-design without anyone steering it. As always, guess before you peek.
Before you read — take a guess
We've spent three lessons on beetles. Now suppose someone claims the *exact same* variation-selection-heredity machine explains why successful smartphone features (big screens, fingerprint unlock) spread across every brand, with no industry committee ordering it. Is that a legitimate use of the model, or a loose metaphor?
The three ingredients, stripped of biology
Before we go touring other fields, let’s say the recipe in its most abstract form — the version with no mention of beetles, genes, or birds. This is the template you’ll match against everything else:
- Variation — the population’s members differ. Something varies. (Raw material: no differences, nothing to choose.)
- Selection — some variants persist, spread, or get reproduced more than others, by some pressure. Something does the choosing. (Direction: no selection, only aimless drift.)
- Heredity — the surviving variants get copied forward, so the next round starts from them. Something carries winners onward. (Memory: no copying, every round resets.)
- Repeat — let the three turn, round after round. (Compounding: one round barely moves; thousands carve a canyon.)
Notice what dropped out: genes, DNA, birth, death. None of those were ever the point. They were biology’s particular way of supplying the three jobs — raw material, direction, memory. Any other system that supplies those same three jobs, by any means, runs the same algorithm. Heredity doesn’t need sex; it just needs copying. Selection doesn’t need death; it just needs some variants doing better than others.
The universal three-question probe
Point this at anything and you’ll know in thirty seconds whether selection is at work: (1) What varies? (2) What selects — what makes some variants do better? (3) What’s inherited — how do winners get copied into the next round? Three yeses means you’ve found a self-designing engine, and you can stop looking for the “designer.” There usually isn’t one. The fitted-looking result was selected, not planned.
Markets select companies
The analogy. An economy looks designed. Shelves are stocked with roughly what people want; prices land in roughly sensible places; useless businesses are rare. It’s tempting to assume someone’s in charge — a planner who arranged it all. Nobody is. A market is a giant selection engine running on companies, and the “fitted” look is exactly the beetle’s camouflage: a filter applied over and over, not a blueprint.
The three ingredients in this domain.
- What varies? Business models, products, prices, strategies. At any moment thousands of firms are trying slightly different things — different bets on what customers want.
- What selects? Customers, profit, and investors. A firm whose product people buy earns revenue and survives; a firm whose product they don’t runs out of money and dies. “Being bought” is this domain’s version of “being eaten less.”
- What’s inherited? Successful practices get copied and imitated — by the firm itself (do more of what works) and by rivals (steal what works). The winning recipe propagates. No genes; pure imitation.
Worked example — the discount-airline copy. One airline tries a stripped-down model: no free meals, one cabin class, fast turnarounds, dirt-cheap fares. Customers, selecting on price, flock to it; it earns money and survives. Rivals watch a competitor thrive and copy the model — the practice is “inherited” across the industry by imitation. A decade later, low-cost carriers are everywhere, and the whole sector “looks designed” around cheap fares. Nobody designed it. Variants (airline models) were tried, customers selected the cheap one, rivals copied it, repeat — and the algorithm sculpted an industry. Meanwhile the airlines that couldn’t compete on price and offered nothing else didn’t get a stern memo; they were quietly selected out — bankrupt, gone, their slots and gates absorbed by survivors.
That phrase — selected out — is the market’s version of “eaten before it could breed.” Firms that fail don’t merely underperform; they’re removed from the population, and their bad practices die with them. That removal is half of what makes the surviving economy look so sensible: the dumb stuff is constantly being deleted.
Why 'the market' looks intelligent and isn't
The eerie thing about markets is that they produce fitted, sensible-looking arrangements that no single mind designed or could design — exactly like the beetle’s camouflage. There’s no committee setting the price of bread; the price is selected by millions of buy/don’t-buy decisions. This is why economists talk about an “invisible hand”: it’s not magic and it’s not a planner, it’s the same blind algorithm you watched grind on beetles, now grinding on firms. When something looks designed but you can’t find the designer, suspect selection.
A struggling restaurant chain ignores changing tastes, keeps its tired menu, loses customers to nimbler rivals, and finally goes bankrupt and closes. In selection terms, what just happened — and what's the analogy to the beetles?
When to use it
Reach for this whenever you see an industry, a market, or an economy that looks suspiciously well-organized and you’re tempted to ask “who arranged this?” Usually nobody did — ask instead what’s being selected for? (Price? Convenience? Status?) The selection pressure tells you which firms will thrive and which will be selected out next. The pitfall: don’t assume “survived the market” means “good for everyone.” Selection only optimizes for what the pressure rewards — if customers select on cheapness, you get cheap, even if cheap means worse working conditions or thinner products. The market builds what it’s selected to build, not what anyone wished for.
A/B testing: selection on purpose, run fast
Everything so far has been selection happening to people — blind, unplanned. But once you understand the algorithm, you can grab the controls and run it deliberately. That’s the leap this section makes: humans harnessing selection on purpose, which is the direct descendant of the oldest trick in the book.
The classic bridge — artificial selection. For ten thousand years, humans have run selection by hand. A farmer with a field of wheat keeps seeds only from the plumpest plants and replants them; a breeder keeps puppies only from the friendliest, woolliest dogs. This is artificial selection (also called selective breeding): the exact same algorithm, except a human plays the role of the selection pressure instead of the environment. The result is staggering — every dog breed on Earth, from chihuahua to Great Dane, was carved out of wolves in a few thousand years by humans choosing which variants got to reproduce. Variation (puppies differ), selection (the human picks), heredity (traits pass to pups), repeat. Darwin opened On the Origin of Species with pigeon breeders precisely because artificial selection is the same machine running fast and on purpose, where you can see the hand doing the choosing.
A/B testing is artificial selection for the digital age. An A/B test is when you ship two versions of something — say, two designs of a webpage’s “Buy” button, version A and version B — show each to half your visitors, and measure which one performs better. Then you keep the winner and kill the loser. Map the three ingredients:
- What varies? The variants you deliberately create — button colours, headlines, layouts, prices, email subject lines.
- What selects? A chosen metric — clicks, sign-ups, purchases, time-on-page. The variant that scores higher “survives.”
- What’s inherited? The winner becomes the new baseline — it’s kept and copied forward into the next test, which spawns new variants against it. Repeat.
Worked example — the button that earned millions. A shopping site tests two checkout buttons: A says “Register,” B says “Continue.” They split traffic 50/50 and measure completed purchases. “Continue” wins — more people finish buying. They keep B, discard A, and “Continue” becomes the new baseline. Then they test B against a new green variant, keep that winner, and so on. No designer reasoned out the perfect checkout flow from first principles; they evolved it, one selection round at a time, letting real users be the selection pressure. Companies run thousands of these, and the polished products you use daily were, in large part, bred, not designed.
It’s the artificial flavour — a human chooses the selection pressure (the metric) and sets up the contest on purpose. But that’s a difference in who holds the filter, not in the algorithm. In wild natural selection, the environment blindly does the choosing; in artificial selection, a human does. The three ingredients — variation, selection, heredity — and the grind toward “things that fit” are identical. This is the deep unity: dog breeding, crop improvement, and A/B testing are all the same engine you watched grind on beetles, with a human hand swapped in for the bird’s beak. Once you see that, you realize humans have been harnessing this algorithm for millennia without a name for it.
Select EVERY example below that is a case of artificial selection — the same algorithm as natural selection, but with a human playing the role of the selection pressure. (More than one is correct.)
When to use it
Use this whenever you face a problem too complex to reason out a perfect answer in advance — which UI converts best, which recipe tastes best, which ad works. Instead of designing the answer, set up a selection engine: generate variants, pick a metric to select on, keep winners, repeat. You’ll often evolve a better solution than you could have designed, and you’ll discover surprises no expert predicted. The pitfall is the same as in markets: you get exactly what you select for. A/B test for clicks and you may evolve a manipulative, clickbait-y product that’s great at getting clicks and terrible at everything else. The metric you pick is the selection pressure — choose it as if your product’s whole future shape depends on it, because it does.
Ideas, memes, and culture
The analogy. Think of the world of ideas as one enormous, noisy population — billions of jokes, songs, slogans, rumours, beliefs, and habits, all competing for the scarcest resource there is: space in human minds and conversations. Most die in obscurity. A few catch, spread, and end up in everyone’s head. That sorting isn’t random, and it isn’t (usually) anyone’s plan. It’s selection running on ideas.
The term. A meme (coined by biologist Richard Dawkins in 1976, deliberately rhyming with “gene”) is a unit of culture that spreads by being copied from mind to mind — a tune, a catchphrase, a recipe, a fashion, a way of doing something. The internet later borrowed the word for funny images, but the original meaning is broader: anything cultural that gets copied. Map the ingredients:
- What varies? Ideas differ wildly — a million jokes, a thousand versions of a rumour, endless tunes. Variants are constantly thrown up, deliberately or by accident (a story mutates a little each retelling).
- What selects? How memorable and transmissible an idea is. Catchy, emotional, surprising, easy-to-repeat ideas get passed on; boring, complicated, forgettable ones die in the first mind they land in. “Being repeated” is this domain’s “being eaten less.”
- What’s inherited? Imitation. You hear a catchy phrase and repeat it; that is the copying step. The idea propagates from mind to mind, no genes required.
Worked example — the unkillable jingle. A four-note advertising jingle gets stuck in your head and you find yourself humming it for days, maybe even singing it to a friend (who now has it too). Ask the three questions. It varied (one of thousands of jingles written that year). It was selected (its simple, repetitive, sticky structure made it easy to remember and hard to shake — that’s the pressure it won on). It’s inherited (you hummed it, spreading it to the next mind). The jingle “won” a selection contest. But won it for what? Here’s the trap.
Selected-for-spread is NOT selected-for-true
The most important and most missed point in this whole section: an idea’s selection pressure is how well it spreads, which has almost nothing to do with whether it’s true, good, or useful. A catchphrase you can’t stop repeating won a contest for stickiness, not correctness. A juicy false rumour outspreads a boring true correction every time, because “shocking and shareable” beats “accurate but dull” in the only contest that matters — getting copied. This is why misinformation, urban legends, and catchy-but-wrong slogans thrive: they’re superbly fit for transmission, and transmission-fitness is blind to truth. When an idea is everywhere, that tells you it spread well. It tells you nothing about whether it’s right. Never confuse “everyone’s saying it” with “it’s true” — you’re just observing the winner of a popularity contest, not a truth contest.
A frightening but completely false health rumour spreads through millions of social-media shares, far outpacing the accurate, boring correction that follows it. Using the selection-on-ideas model, what's the cleanest read?
When to use it
Pull this out whenever you notice an idea, slogan, or “everyone knows” fact that’s suspiciously widespread, and ask: did this win for being true, or for being sticky? High spread is evidence of transmissibility, never of truth. It also flips into a design tool: if you want a true and useful idea to spread, you have to make it fit for transmission — memorable, concrete, repeatable — or it’ll lose to the catchy nonsense every time. Good ideas don’t win automatically; they have to be built to survive the copy machine.
Machine learning and genetic algorithms
Here the algorithm comes full circle: humans, having understood selection, now write it into software and let it design things no human could. The clearest case is a technique called a genetic algorithm.
The term. A genetic algorithm is a problem-solving method that runs variation, selection, and heredity inside a computer to evolve a solution, rather than having a programmer design one. You don’t tell the computer how to solve the problem; you tell it how to recognize a good solution, and then you let selection breed one. Map the ingredients — they’re literally coded as steps:
- What varies? A population of candidate solutions, each a slightly different attempt — random at first. (Say, a few hundred random shapes for an antenna, or random strategies for a game.)
- What selects? A fitness function — a chunk of code that scores how good each candidate is at the goal (how well the antenna transmits, how many games the strategy wins). High scorers “survive.”
- What’s inherited? The best candidates are combined and copied into the next generation — mixed together (a digital echo of breeding) and tweaked with small random changes (digital mutation). Then the whole loop repeats, often for thousands of generations.
Worked example — the antenna nobody designed. NASA famously needed a tiny radio antenna for a spacecraft with unusual requirements. Instead of a human engineer designing one, they ran a genetic algorithm: start with random antenna shapes, score each by how well it would transmit (the fitness function), keep and recombine the best, mutate, repeat — for many generations. Out came a bent, asymmetric little shape that looks like a paperclip a toddler mangled — and it outperformed the human-designed alternatives. No engineer drew it. No engineer could have justified its weird kinks in advance. It was evolved — selected, generation by generation, for the one thing that mattered: transmitting well. The same approach evolves delivery schedules, factory layouts, game-playing strategies, and circuit designs that look bizarre and work brilliantly.
Why the evolved solution looks 'designed by an alien'
Evolved solutions have a signature: they work but look nothing like what a human would draw. They’re full of weird asymmetries and inexplicable kinks, because selection doesn’t care about elegance, symmetry, or being explainable — only about scoring well on the fitness function. This is the same reason real biology is full of clumsy, jury-rigged “designs” (your eye’s backwards-wired retina, the absurd detour of a giraffe’s recurrent laryngeal nerve): selection optimizes blindly for what works, not for what looks designed. When you see a solution that’s effective but baffling, suspect it was selected, not reasoned out.
You’ll also hear that much of modern AI is “trained.” At a high level, training a model is also a selection-flavoured process: countless candidate adjustments are tried, the ones that reduce error are kept and built on, and the rest are discarded — iteration toward “what fits the data,” round after round. The details differ from a textbook genetic algorithm, but the family resemblance — generate, keep-what-scores-better, repeat — is the same blind grind toward fit. Keep that loose; the genetic algorithm is the crisp, undeniable example.
The immune system: selection inside your body, in days
Now for the example that brings selection home — quite literally inside you. The other domains were biology’s algorithm running on non-biological stuff. This one is real, fast, blood-and-cells selection happening within a single body, on a timescale of days. It’s the most vivid “selection in real time” you’ll ever meet, because it’s saving your life as you read this.
The setup. When a new pathogen — a flu virus, say — invades, your body needs an antibody: a protein shaped to grab onto that specific invader and flag it for destruction. But your body has never seen this exact invader, so it can’t have a pre-made antibody for it. It can’t design one either. So it does something cleverer: it runs selection.
The three ingredients, in your bloodstream.
- What varies? Your immune system generates a vast, random library of antibody shapes — millions upon millions of different B-cells, each making a slightly different antibody, by deliberately shuffling gene segments. Pure blind variation, generated before any specific invader shows up.
- What selects? Gripping the invader. A B-cell whose random antibody happens to stick well to the pathogen gets a powerful “you’re the one — multiply!” signal. The ones that don’t grip get nothing. The pathogen is the selection pressure.
- What’s inherited? The winning B-cells clone themselves rapidly — copy, copy, copy — and even mutate slightly as they divide, producing variants that grip even better, which are then selected again. This is a second round of variation-and-selection nested inside the first, tuning the antibody to a razor fit in days.
Worked example — beating a cold. A cold virus invades. Among your millions of random antibodies, a handful happen to weakly grip this particular virus — pure luck, those shapes existed before the virus arrived, exactly like the antibiotic-resistant bacteria from lesson 2. Those few B-cells get the “multiply!” signal and clone like mad, mutating as they go; the best-gripping mutants are selected again and clone harder still. Within days you’ve gone from a handful of mediocre antibodies to an army of finely-tuned ones, and the infection is crushed. Variation, selection, heredity, repeat — running in your bloodstream over a single week. You didn’t design those antibodies. Nobody did. They were selected, in fast-forward, from a random library, by the invader itself.
Evolution you can feel
This is the closest you’ll ever come to feeling natural selection happen. Every time you fight off an infection and recover, you’ve just run the algorithm: a random library, a selection event (the pathogen), and explosive copying of the winners — Darwin’s machine, sped up from millennia to days, inside your own ribs. And this is also why vaccines work: a vaccine is a harmless preview of the invader that lets your body run this selection in advance, so the finely-tuned antibodies are already on file when the real thing shows up. Selection isn’t a distant fact about finches on islands. It’s part of how you’re alive right now.
Mapping the domains: who plays which role?
You’ve now met five wildly different domains running the same three-part machine. Here’s the whole tour in one table — the same three rows, five different bodies for them to run on:
| Domain | What varies | What selects | What’s inherited (copied forward) |
|---|---|---|---|
| Beetles (biology) | Beetle colour | Birds (camouflage survival) | Genes — offspring resemble parents |
| Markets | Business models, products | Customers, profit, investors | Successful practices, copied by imitation |
| A/B testing | Page variants you ship | A chosen metric (clicks, sales) | The winning variant becomes the baseline |
| Ideas / memes | Jokes, rumours, songs | Memorability + transmissibility | Imitation — you repeat what stuck |
| Genetic algorithm | Candidate solutions in code | A fitness function (a score) | Best candidates recombined + mutated |
| Immune system | Random antibody shapes | Gripping the pathogen | Winning B-cells clone themselves |
The columns never change — raw material, direction, memory — only the costume each one wears. That’s substrate-neutrality made concrete. Now lock it in by matching each domain to the role it makes vivid:
Match each selection engine to the thing that plays its SELECTION role — the pressure deciding which variants win. (Pick a domain, then click its selector.)
Pick a domain on the left, then click what does the selecting in that domain.
Each phrase below is the 'heredity' step — the way winners get COPIED FORWARD — in one of these domains. Sort each into the domain whose copying mechanism it describes.
Place each item in the right group.
- The highest-scoring candidates are recombined and mutated for the next generation
- B-cells that grip the invader clone themselves rapidly
- You hear a catchy phrase and repeat it to a friend
- Rival firms imitate a competitor’s winning strategy
- A successful business copies its own winning practice across more locations
- A rumour propagates by being re-told, mutating slightly each time
Fill in the big idea that ties every domain together:
Pick the right option for each blank, then check.
Natural selection is , meaning it doesn't care what material it runs on. That's because it's an — a fixed set of steps — and the steps only need three jobs filled: something that (the raw material), something that (the direction), and something that (the memory). Supply those three to companies, web pages, ideas, code, or antibodies and you get the same blind grind toward things that fit — with required.
Recap
You came in able to watch the algorithm run on beetles. You’re leaving able to spot it anywhere:
- Natural selection is substrate-neutral — it’s an algorithm, and an algorithm doesn’t care what material it runs on. Strip away genes and biology and you’re left with three jobs: variation (raw material), selection (direction), heredity (memory). Any system that fills those jobs runs the same machine.
- The universal probe: ask what varies? what selects? what’s copied forward? Three answers means you’ve found a self-designing engine — and you can stop hunting for a designer who usually isn’t there.
- Markets select companies: models vary, customers/profit select, practices are copied by imitation, and unfit firms are selected out. The “designed-looking” economy was selected, not planned.
- A/B testing and selective breeding (dog breeds!) are artificial selection — humans harnessing the algorithm on purpose by choosing the selection pressure. Same machine, human hand as the filter.
- Ideas / memes are selected for spread, not truth — a sticky catchphrase or false rumour wins a transmissibility contest, which is blind to correctness. Widespread ≠ true.
- Genetic algorithms run variation + selection + heredity in code to evolve solutions (the bizarre, effective NASA antenna) no human designed. AI training rhymes with the same loop.
- Your immune system runs real selection inside you in days: random antibodies vary, the ones gripping the pathogen are selected, and the winners clone themselves. Evolution you can feel.
Check yourself: the algorithm everywhere
What does it mean to say natural selection is "substrate-neutral," and why does it matter for using it as a thinking tool?
Check your answer to continue.
Where this goes next
Here’s the strangest consequence of everything you just learned. The algorithm is so good at producing fitted, polished, purposeful-looking results — the camouflaged beetle, the sensible economy, the perfectly-shaped antibody, the alien NASA antenna — that your brain takes one look and insists, someone must have meant this. Surely the eye was designed for seeing. Surely the market was arranged by someone. Surely evolution is trying to get somewhere, climbing toward a goal, improving the species, building up to us.
Every one of those is the same beautiful illusion, and it fools nearly everyone — because spotting fitted design and inferring an intender is exactly what human brains evolved to do. Lesson 5, What Evolution Is Not, is the cleanup crew: it dismantles the three misreadings that the algorithm’s eerie competence plants in your head — that selection is goal-directed, that it works “for the good of the species,” and that it’s a ladder of progress climbing toward humans. This is where most people’s intuition quietly lies to them, and it’s the last thing standing between you and genuinely getting the model. Then comes the exam you can’t undo.