Open any feed and watch yourself for a minute. You scroll past a hundred things and stop on one — maybe you share it, maybe you just linger. Why that one? Almost never because it was the most accurate thing on the screen. Because it made you feel something, or it said something about your side, or it was simply easy. Now multiply that split-second choice by everyone, all day long. That is the machine this chapter is about. Chapter 5 explained how an idea gets reshaped as it travels; this chapter is the prior question — which ideas get to travel at all. That turns out to be the half with a steering wheel, which is why the rest of the book runs on it.
Chapter 5 kept pointing past itself. Every few pages it described transport doing its work — the lossy hop, the compression forced by transit, the network ceiling — and then caught itself: that is one half of the machine, the other half is selection, and the other half is this chapter. So here it is. It is the half Chapter 5 kept deferring, and in the digital era it is the more powerful half.
The deferral, said as plainly as I can. Transport explains what shape an idea takes as it moves. It does not explain what moves at all. Chapter 5 leaned on the telephone game — ten kids, a sentence around a circle, garbage out the far end — and then admitted the game cheats: it forces every player to pass the message. That rule is the whole of the parlor game, and it is exactly the rule real networks do not have. In a real network each receiver decides whether to pass a thing on, to whom, and in which of its many circulating forms. That decision is not a re-encoding. It is a gate with criteria. Selection is the sum of those decisions, and this chapter is about what the criteria are, who gets to set them, and why that last question turns out to be the one the rest of the book runs on.
The selection claim, said plain: what crosses a network is not what transport left intact — it is what selection chose; and because selection chooses by criteria, selection, unlike transport, can be aimed. The second clause is the load-bearing one. A mechanism that works by criteria is a mechanism with a steering wheel. Most of this chapter is about that wheel: where it is, whose hands are on it, and what it would take to grip it differently.
Three operations, not two
transport-vs-selection and Chapter 5 both describe a pipeline stage as doing two things: transport re-encodes, selection picks. Drafting this chapter forced a correction. A stage does three things, and the missing one matters.
Start with a gap the two-mechanism picture skips. Selection picks from a set. It ranks candidates and promotes a winner — but ranking is not generating. Selection never invents the options it chooses among. So where does the set come from? “Transport re-encodes” does not answer it. Transport takes one message and scrambles it; run enough hands across enough copies and what you get is not one output but a cloud of variants drifting every way at once. Information Evolution in Social Media, tracing over 460 million instances of Facebook memes, found exactly this: variants accumulate by a neutral Yule process, “most meme variants are neutral with respect to the parent’s meme,” edit distance compounding as copy is made from copy is made from copy. That directionless cloud is the raw material.
The blindness is not a quirk of memes. Luria and Delbrück pinned it down for biology in 1943: their fluctuation test showed bacterial mutations arise before the selective pressure that later rewards them, thrown off with no foreknowledge of what will prove adaptive. Variation proposes in the dark, and only afterward does anything choose — the generate step has run blind for eighty years before anyone watched memes do it.
But not every variant in the cloud can even be expressed at the next stage. A medium with a 280-character limit cannot carry the long, qualified variant. An image medium cannot carry the variant whose entire content is a subordinate clause. Something stands between the cloud transport throws off and the winner selection promotes — something that fixes which variants are available to be chosen at all. That something is the medium.
So, three operations, present at every stage. Transport generates: lossy re-encoding scatters the idea into a cloud of variants. The medium bounds: it fixes the option space — the set of variants that can be expressed here at all. Selection picks: it ranks the viable set against its criteria, and one variant travels on. Generate, bound, pick. Picture a single rumor: it gets retold in a thousand slightly different ways (generate); only the versions short enough to fit in a text message actually get sent (bound); and of those, the juiciest one is what people actually forward (pick). The transport-vs-selection note works through why that is the right carving; what matters here is that the rest of this chapter is the third operation — with the second, the medium, turning out near the end to be the most important of the three.
What the gate rewards
A selection gate is not a mysterious thing. It is a receiver — a person, an editor, an algorithm — deciding whether a given variant is worth passing on, by criteria local to them. Stack those decisions across a whole network and an emergent fitness landscape falls out: what gets passed is whatever keeps winning local decisions. Call the criteria memetic fitness — with the warning that nothing in the word “fitness” implies the criteria are good, coordinated, or pointed at truth.
A caveat on “meme” before I lean on it. The word carries a contested theory — that culture comes in discrete, copyable units, replicators analogous to genes — and Memetics — a critique attacks exactly that, arguing cultural transmission is reconstructive, not copy-based. I think the critique largely lands, and the good news is this chapter does not need the strong claim. It needs only that receivers re-transmit some variants more than others by local criteria, which is observable whatever the underlying ontology turns out to be. “Selection” here is a claim about a sorting behavior, not about discrete replicators.
What are the criteria? The empirical literature is unusually blunt. Psychology of Virality lists the variables that predict spread on social media: “out-group animosity, incivility, negativity, moral outrage, anger, misinformation, indignant disagreement, and high-arousal information.” Information Evolution found the textual fragments that confer measurable replicative advantage — 4-grams recurring across unrelated memes with a fitness ratio above 1.4 — and they are not subtle: “allusions to competition (‘see how many people’), persistence cues (‘status for at least’), or conditions that are easy to match or identify with (‘if you love your’, ‘paste if you agree’, ‘proud to be a’),” and the self-flattering challenge, “95% of you won’t copy this, but the 5% who [have a positive attribute] will.” These are the gate’s criteria, made of measured text: identity, belonging, a small action to perform, a reason to feel like one of the good ones. None of them is truth, and none of them is even aimed near it. A friend, in general-theme, put the floor of this bluntly — “good memetics reduces down to evoking raw strong emotions.” That is what an engagement-tuned gate converges on, and Chapter 7 takes up why it is a stable floor. The point here is narrower: the gate has criteria, the criteria are nameable, and nobody who set them was asked whether you end up with a true picture of the world.
Transport is a cost; selection is a choice
Here is the asymmetry the whole chapter rests on, and it is not the one I expected when I started.
For a long time I wanted to say selection is primary — that transport is a substrate and selection the real dynamic. transport-vs-selection worked through why that is wrong: transport and selection are co-equal and parallel, neither downstream of the other. But “co-equal” is not “identical,” and the way they differ is the point. Transport and selection differ in modality: transport is an un-steerable cost, and selection is a criterial choice.
Transport is a cost; it scrambles. The word I had to fight my way to is un-steerable, because for months I wrote directionless instead, and that was wrong: the most directional process in all of physics is a cost. Entropy only ever increases, the second law never once points the other way, and you still cannot aim it at anything you choose. That is the distinction the slogan needs. Having a direction and being steerable toward a target you picked are different properties, and a pure cost can have the first and still lack the second. So transport is not directionless; it is un-steerable, and what makes it un-steerable is that it is blind to content: it scrambles a variant the same way no matter what the variant says. You can make transport cheaper or dearer, with a share button or a character limit, but you cannot make it favor one idea over another, because favoring means reading what the idea says, and reading-what-it-says is not a thing a blind cost does. Selection is different in kind. Selection has a what-it-rewards. It is criterial, and a criterion is exactly the kind of thing that can be set, tuned, captured, and contested. You cannot aim a cost. You can aim a choice.
One honest correction, because the tidy form of the claim is false. I have been writing as if cost and choice were two kinds of stuff. They are not; a choice is a cost too. Maxwell’s demon never pushes a molecule (it only opens the door for the fast ones), and for a century that looked like getting order for free, until Landauer and Bennett showed the demon pays in the end: the bits it must eventually forget cost kT ln 2 apiece to erase. Natural selection keeps the same books; its direction is bought with differential death. The real split is not cost against no-cost but blind cost against criterial cost — two ways to spend the same dissipation, one indifferent to what it acts on, one coupled to a criterion. Aiming is never free: a criterion aims a cost by coupling to it, and the coupling gets billed. This is also where the political economy of the later chapters comes from instead of getting asserted, because the criterion is the cheap, portable, low-mass layer riding on a bulky dissipative substrate, which is exactly why capture always reaches for the criterion and never the pipe.
None of this is new math, which is the reassuring part. Population genetics has carried the same split for more than fifty years, in George Price’s 1970 covariance equation: the change in any trait decomposes into exactly two terms, a selection term that is the covariance between the trait and fitness, and a transmission term that is everything imperfect copying does on the way to the next generation. The aim lives entirely in the first term. Fitness is the criterion; the covariance is the steering; and if the trait does not covary with fitness — no criterion bearing on it — that term is zero and nothing is aimed. The second term, transmission, is the blind cost wearing its formal clothes. And the equation is honest about the thing I had to walk back above: a biased transmission shows up right there in the second term, a cost that scatters unevenly without ever consulting fitness. Anisotropic, still blind, still un-steerable. The slogan is just the second term and the first term, read out loud: you cannot aim the transmission, you can aim the covariance.
That is why the book centers selection, and it is worth being exact about why, rather than waving at “selection matters more.” Selection is not central because it comes first, or does more work, or explains more cases. It is central because it is the mechanism with a steering wheel. Every question the later chapters care about — manipulation, political economy, repair — is a question about that wheel. You cannot manipulate transport; there is nothing to grip. You cannot repair transport; lossiness is not a bug, it is what abstraction is. You manipulate selection, and you repair selection — by changing what the gates reward, or who is allowed to set it. Transport is the terrain. Selection is the steering. A book about a vehicle that has gone somewhere bad is going to be mostly about the steering.
This is also where the historical claim slots in cleanly instead of competing. Transport cost has collapsed — Information Evolution watched it happen, the moment a “share” button replaced manual copy-and-paste and made replication free and exact. When transport was expensive it was also a brake: plenty of ideas could not afford the trip, and which-ideas-could was a real constraint on what spread. Collapse the cost and you do not delete a mechanism — transport still scrambles — you remove the brake. Everything can afford the trip now. So the whole question of what travels has devolved onto the one mechanism that was always a choice. Transport collapsing did not make selection the steering wheel. Selection always was. Collapse just handed it the whole wheel.
The medium is frozen selection
Go back to the second operation — the medium, bounding the option space. I called it the most important of the three and then walked past it. Here is the argument.
The medium note established that the medium sets a gate’s criteria — that the medium is the selection criteria. This chapter can sharpen it. Drawing the boundary of the option space — ruling that the long careful variant cannot be expressed and the short alarming one can — is itself a criterial act. It keeps some forms and forbids others; it has a what-it-allows, which is a what-it-rewards with the contrast turned all the way up. So the medium is not a third kind of thing alongside transport and selection. The medium is selection — performed once, in advance, and then frozen into the substrate until it stops looking like a choice.
A 280-character limit was somebody’s decision, made in a meeting, on a date. To everyone who has ever posted inside it, it is simply the shape of the world — not a gate, not a criterion, just how the medium is. That is what “frozen” buys. A runtime gate is contestable: you can watch it rank things and argue about its weights. A medium does not look like a gate at all; it looks like physics. Which makes it the most powerful selection surface in the whole picture and the least visible — because a runtime gate has to actively suppress the variant it does not want, and a medium never has to. A variant the option space cannot hold never has to be suppressed, because it never forms. The most effective censorship is the kind where nothing is censored, because the disfavored thought was never sayable to begin with.
So the medium and the runtime gate are not two things. They are one continuum of criterial constraint run at two speeds. At the slow, frozen end: the medium, set once, bounding what can be said. At the fast, live end: the recommendation algorithm — and here the point about algorithms lands hard. A recommendation algorithm is not lossy transport and it is not a neutral pipe. It is a selection gate with adjustable weights, re-tuned continuously, and Psychology of Virality gives the worked example — an audit of the Twitter/X algorithm found it “amplified political content that did not align with people’s stated preferences.” That is the live end of the same continuum the 280-character limit sits at the frozen end of. Both are selection. One was just set so long ago, and so far upstream, that we forgot a choice was ever made.
Worth making the word aim precise, because “selection can be aimed” has been hiding a spectrum inside a single verb. A criterion is aimable exactly to the degree that it sits in some actor’s causal reach, and reach varies wildly along this same continuum. The runtime gate is the easy end: its weights are right there, re-tunable in an afternoon by whoever owns the server. The medium is far harder to reach and far higher-leverage, which is the whole point: almost nobody is positioned to redraw the option space, and the few who are do it once and walk away. Harder still is the evolved human gate, reachable only slowly, by training, across thousands of turns of the want-loop. At the far end sits something like a fixed environment: natural selection has a what-it-rewards and yet offers no console anyone can sit at. So “you can aim a choice” was never a yes-or-no. It says a choice has somewhere to put an aim — a criterion — and how hard that somewhere is to reach is the whole distance from nudging a slider to capturing a platform.
Hammer of the Witches vs. Copernicus, properly this time
Chapter 5 told this story and deferred the real reading to here. The two books are Harari’s pair from the early printing-press era: The Hammer of the Witches, a short, emotionally loaded manual for finding and prosecuting witches, which sold out edition after edition and burned its image of the witch so deep into European culture you can still feel the outline; and Copernicus’s On the Revolutions of the Heavenly Bodies, which Arthur Koestler called “an all-time worst seller.” Same press. Same century.
The pure-transport reading — the one Chapter 5 gave and called incomplete — is that the Hammer was the compressed form and Copernicus the complex one, and the compressed form travels. True as far as it goes. But notice what it cannot explain: both books were printed on the same press. Their transport costs were identical. Whatever sent one to every town in Europe and left the other on the shelf, it was not the price of the trip. Run the three operations instead. Transport: the press generated copies of both, cheaply and equally. The medium: a printed book could carry either — the option space held both. The entire difference is the third operation. Selection.
And against the selection criteria of sixteenth-century Europe the two books had wildly different fitness. The Hammer offered fear, urgency, a crisp in-group and out-group, an action to perform — find the witches — and a complete narrative with the reader cast as the hero. Copernicus offered a redrawn cosmology with no action implications, no emotional payload, and no enemy. Hand both to a gate built — as every human gate is partly built — to reward threat, agency, and belonging, and the outcome is not close. The compressed-versus-complex distinction is real, but it is downstream: the Hammer is compressed because it was selected for, and the criteria it was selected against rewarded exactly the qualities a fear-manual has and a cosmology lacks. Transport did not pick the Hammer. A whole continent of gates did, the same gates over and over, and transport merely carried what they kept choosing.
The manipulation surface is the gate itself
Chapter 5 gave the transport version of the manipulation surface: bigger network, lower complexity ceiling, ideas sit closer to raw emotional payload, easier to weaponize. Real, but the shallow version. The selection version is worse.
If selection runs by criteria, and criteria can be set, then the manipulation surface is the criteria themselves. You do not have to craft a perfect viral object and inject it. You tune the gate, and the network does the rest — every receiver downstream now ranks by criteria a little friendlier to what you wanted. Information Evolution names the disanalogy with biology precisely: “evolution is a blind process in biology, but in social media, there can be a conscious effort to create mutations that will intentionally spread a meme’s presence.” Memetic selection, unlike natural selection, has an operator’s console. And the frozen-selection point makes it worse: the highest-leverage move is not tuning the live gate, which someone might audit. It is shaping the medium — the option space, the defaults, the affordances — so the variant you do not want is never well-formed. That manipulation leaves no fingerprints, because it produces no suppressed post and no demoted link. It just produces a world where the inconvenient idea is hard to say.
This is also where manufactured content — astrology, propaganda, invented statistics, AI-generated text — meets selection. A selection gate scores a variant on fitness, not on provenance. It has no input for was this ever measured against reality. So a manufactured claim and a genuine finding compressed into something equally punchy arrive at the gate looking identical, and the gate, asking only “does this score,” promotes them with equal enthusiasm. Selection does not merely fail to stop manufactured content; it cannot, in principle, see the difference. Who tunes the gates, and whether anything in the criteria points at truth, is the whole game — and it is Chapter 10’s.
Is “complexity” doing two jobs? Yes — so split it
Chapter 5 flagged a crack in its own foundation and handed it here. It defined complexity as “the number of preconditions a receiver needs,” then kept catching itself using the word for other things — depth, subtlety, informational entropy — and worried the real gating variable might not be preconditions at all but handle-ability: how many existing mental models an idea plugs into, how much emotional or identity content it carries, whether it tells the receiver what to do.
The two-mechanism structure resolves this, and the resolution is not “pick one.” It is: “complexity” was doing two jobs because the book had one word for what is really two variables, one per engine. Precondition count is transport-complexity — how much outer message, how much decoding key, must survive the hop for the idea to land intact. It is what sets the complexity ceiling; it is a property of the idea-in-transit; Chapter 5’s precondition framing is exactly right for transport. Handle-ability is selection-complexity — how the idea scores at a gate: emotional load, identity fit, action affordance. It is a property of the idea-at-a-gate. The two were never going to reconcile into one axis, because they belong to two different mechanisms. “Complexity” only looked like a single variable while the book still thought the pipeline had a single engine.
Splitting the term earns its keep at once, because the two variables come apart in the wild. An idea can be low transport-complexity — cheap to carry, few preconditions — and still die, because it is low handle-ability and scores at no gate. (“Inflation eased two-tenths of a point last quarter” is trivial to carry and goes nowhere.) And an idea can be high transport-complexity and still spread, if its compressed form is high handle-ability — which is the Hammer dynamic exactly. The precondition ceiling and the fitness criteria are two separate filters an idea must pass, and a theory with one word for both could never say so cleanly. This is provisional: I am not sure handle-ability is itself a single variable rather than a small bundle. But the move from one “complexity” to two, split along the transport/selection seam, is one I am fairly confident in.
The strongest objection: people are not that gullible
The honest pressure on this whole chapter comes from Hugo Mercier’s Not Born Yesterday and, from another angle, Cailin O’Connor and James Owen Weatherall’s The Misinformation Age. I have to take both seriously or the chapter is propaganda for its own thesis.
Mercier’s argument is that humans are not the soft targets the easy-manipulation story needs. We come with epistemic vigilance — what Psychology of Virality, citing the same literature, calls “mental immune systems”: evolved machinery that screens incoming claims for plausibility and sources for reliability. People “do not passively consume information, but interpret, transform, or ignore” it. If that is right, a chapter saying selection gates push whatever-scores into receivers who dutifully believe it is too crude.
I think Mercier is right, and I think he sharpens this chapter rather than breaking it — once you notice that epistemic vigilance is itself a selection gate. The receiver is not a passive endpoint. The receiver is a gate, and vigilance is one of its criteria. So Mercier’s correction is not “there is no selection,” it is “the receiver’s own gate is more truth-correlated than the cynical story assumes.” Good — that is true, and it is the thing Part IV’s hope rests on. Mercier actually re-labels Sperber’s term as [[open-vigilance|open vigilance]] to mark that the system is paired — receivers are simultaneously open and vigilant, with the two having co-evolved. But it does not save us, for two reasons. First, vigilance is one criterion competing against all the others — identity, threat, belonging — and Information Evolution’s fitness data shows those others routinely winning; a person can have a working immune system and still pass the meme that flatters their tribe. Second, and deeper: the receiver-gate’s criteria are not fixed. They are trained by the medium over thousands of repetitions — the want-loop. A medium can, slowly, select for receivers who pass things on without checking. Mercier describes the immune system we evolved. He does not describe an immune system that cannot be worn down. Ch 7 works through the full Mercier engagement and surfaces a deeper challenge the chapter here only gestures at — the justification-market reframe where behavior drives belief rather than the reverse, which would partially redirect this chapter’s “selection installs beliefs” reading toward “selection supplies demand-side justifications.” The two readings compose; Ch 7 carries the working-out.
O’Connor and Weatherall point at the other half. Their network models show that what corrupts collective belief is often not lossy transmission at all but selection on social criteria — conformity, the credibility we extend to people like us — operating even among agents who each reason perfectly well. That is not a counter to this chapter; it is this chapter, arriving from formal modeling: the gates, not the hops, are where truth is won or lost. The sharpest single piece of evidence is the one Psychology of Virality calls the paradox of virality — across the political spectrum, people report they “do not want” high-arousal negative content to go viral, and it goes viral anyway. Read naively, that paradox looks like it breaks any selection story: if selection reflected what people want, the unwanted thing would lose. But it is not a break. It is the frozen-selection point caught in the act. What spreads diverges from what receivers want precisely because the gate that decided was not the receiver’s gate. It was the medium’s — the algorithm, the option space — tuned by someone whose interest is not the receiver’s. The paradox of virality is not evidence against selection. It is evidence of whose selection is running.
One flag worth keeping: Mercier and O’Connor & Weatherall are load-bearing for this section, and the engagement with both is sourced — page-anchored evidence, supporting and opposing, sits behind the claims here rather than paraphrase from memory. Where their work cuts against the book, the cut is recorded, not smoothed. The deeper Mercier engagement happens in Ch 7, which works through the timescale-distinction, the justification market, and the intuitive belief split that this chapter only gestures at.
What selection does to truth
This chapter has said selection picks the variant that travels. It has not said what travelling does to that variant’s truth, and it should not, because that is Chapter 5c’s, and 5c already does it. The one sentence of handoff: transport, scrambling without direction, produces a cloud of variants that are merely off; selection, choosing by criteria, is what gives the error a direction — and whether a compressed claim ends up preserved, inverted into its opposite, or emptied into a tribal flag is set by which variant the gate reaches for. Selection is the engine; 5c is what the engine does to the truth of what it carries. The two chapters are one argument, and this is the seam.
One more sentence the seam needs, now that Chapter 2 has made the book say which reality is in transit. Everything above is the objective-content case: a gate filtering reports about facts it cannot touch. When the cargo is intersubjective — currencies, nations, creeds: truths generated by the network’s agreement rather than by reality — the same gates are doing something categorically different. They are not filtering descriptions of a fact; they are weighting votes in what becomes the fact, because for that cargo, distribution participates in constitution. The intersubjective-truth note works the case out. And the book’s oldest case study has been carrying the distinction unmarked: the Hammer’s claims about its victims were objective and false, but the judicial category it fed — the witch, prosecutable, confessable, burnable — was intersubjective and lethally real. The gates that kept choosing the Hammer were not just spreading errors about the world. They were voting a new thing into it.
Where I land
Selection is the other engine, and the one the book is really about. Transport explains the shape of what moves; selection explains what moves. The two run in parallel at every stage, neither downstream of the other — but they are not symmetric. Transport is an un-steerable cost and selection is a criterial choice, and that single asymmetry is why every later chapter is a selection chapter.
The diagnosis, stated as a selection claim: the breakage this book keeps circling is not that networks select — they always have, they must — it is that the dominant medium’s gates have been tuned, by people whose interest is engagement, to reward identity and threat and arousal over anything correlated with truth; and because the medium is selection frozen into the substrate, most of that tuning no longer even looks like a choice. That sets the prescriptive target precisely. There is no fix that routes around selection — a network with no gates is not a network. The fix, if there is one, is gates tuned differently and a medium whose frozen-in option space does not quietly forbid the complex form. A harder target than “stop the compression,” and a more honest one.
What this changes for the book
- Chapter 7 (emotional memetics) inherits its floor from here: “raw strong emotions win” is not a law of crowds, it is the equilibrium of a gate tuned for engagement. Stated that way, the floor becomes contingent — a different gate has a different floor.
- Chapter 10 (political economy) stops being optional. If selection is the steering wheel, “who owns the medium” is “who sets every criterion in the network.” Political economy is not a side chapter; it is the chapter about whose hands are on the wheel.
- Chapter 11 (AI) gets a sharp question instead of a vague one. An LLM is a gate and a medium at once — it ranks, and it bounds the option space. The question is not “is AI good or bad” but “what are this gate’s criteria, and who froze them in.”
- Chapter 12 (infrastructure) gets its brief: build gates whose criteria are truth-correlated, and media whose option space does not silently forbid the complex form. The two conditions are the spec.
Where I’m still uncertain
- Handle-ability might not be one variable. I split “complexity” into a transport half (precondition count) and a selection half (handle-ability), and I am confident in the split. I am not confident handle-ability is a single axis rather than a small bundle — emotional load, identity fit, and action affordance may not move together, and if they do not, “selection-complexity” has to become a short list, not a word.
Mercier and O’Connor & Weatherall are argued, not sourced.Closed in subsequent extraction phases — The Misinformation Age and Not Born Yesterday are both now in the graph with substantial interpretive backing. The deeper engagement with Mercier’s specific arguments lives in Ch 7 and is downstream of the timescale-distinction the book commits to there.- Emergent versus engineered tuning. I have said gates “are tuned” and let the passive voice carry it. A gate that drifts toward engagement under an ad-revenue incentive and a gate someone deliberately set to invert a specific finding are different things, and “tuned” hides the difference. Telling them apart is Chapter 10’s job; until it is done, the chapter’s villain is a little blurry.
- The receiver-gate deserves its own treatment. I introduced epistemic vigilance as “the receiver is a gate too” and used it to absorb Mercier in a paragraph. That is a lot of weight on one move. Within-receiver selection — what a single mind attends to, retains, re-transmits — may need its own section, or its own note.
Virality Trade-off · Chapter 5c: Truth, Compression, and When Each Wins →