John Sokol's Blog: memetics

Showing posts with label memetics. Show all posts

Friday, May 29, 2026

Anagrams as Source Code: A 2014 Hint at What LLMs Would Be

Fourth in a series. Read the mimetics essay for the broader frame; this one zooms in on a single page of my old wiki and asks why the intuition on it turned out to predict, in structure if not in detail, what large language models would become.

Sometime in the early 2010s — the page history says the bulk of edits land between 2011 and 2015 — I wrote on my wiki the following claim, which I’ll quote unedited because the rough edges are the point:

I am convinced it’s not “SuperNatural” but part of a memetic fabric of the way the brain does pattern recognition. … This has been done so long, that the languages evolve with these mechanisms built in to it. So much so that message content is based on the frequency of letters more then the actually words.
Anyhow decoding the anagrams is like looking at the source code, you can’t take it’s contents on literal face value but how it will subconsciously effect those people who “Get it”. … Now that I’ve been writing for a while I can feel what’s going to make a good anagram or not. I can feel weather it’s positive or negative.
It should also be possible to have a program that can measure a content’s meme-worthiness, (the ability to propagate as a meme, and in the greater context of current events and the mainstream media’s effects.)
It should also be possible to create a system that could generate candidate meme’s that should have a high likelihood of success.

The specific claim — that anagrams literally carry meaning, that re-arranging letters reveals a hidden semantic layer — I now think is mostly wrong. But the structural claim wrapped around it is something I’ve gotten less sure was wrong with every passing year. The structural claim is that language has a substrate of statistical pattern that does cognitive work in parallel with semantic content, that the substrate is exploitable, and that a program could be built to operate on it.

Nobody built that program in 2014. Then, ten years later, somebody did — but they did it by accident, while trying to build something else, and what they ended up with looks nothing like what I imagined. They called it a large language model. It works, and the reason it works validates the hint on the wiki page more than I would have predicted.

This essay is about the connection.

1. What the wiki claim was, exactly

The page is Nature_of_Language in the cluster I built around Department of Memetics. Stripped to its bones, the argument runs:

Spoken and written language has evolved alongside the brains that process it for hundreds of thousands of years.
Any feature of language that affects reader response will, under that long Darwinian sieve, accumulate in the language because writers and speakers whose feel produced more effective patterns out-propagated those whose feel did not.
Letter and phoneme distributions are such a feature. They affect reader response — the rhythm of a sentence, the “feel” of a name, whether a phrase lands — on a channel below conscious semantic processing.
Therefore there is hidden statistical structure in language that is doing cognitive work in parallel with the propositional content.
Skilled writers can sense the structure but do not consciously manipulate it. Actors choose stage names whose letter distributions “feel right.” Marketers find the slogan that sticks. Speechwriters tune phrases until they’re memorable.
In principle, a program could measure the structure. It could score arbitrary text for “meme-worthiness.” It could generate text optimised against that score.

I added the anagram story because I was doing statistical work on anagrams at the time and thought I was finding something specific — that the anagrams of high-impact phrases had statistically interesting properties. I now believe what I was actually picking up was the underlying letter-frequency structure of the source phrase, which an anagram by definition preserves. The anagrams were a symptom; the real signal was the letter-frequency surface of the input.

The substrate was real. The diagnostic — anagrams as source-code- revealing — was wrong but pointed at something true. This is a more common pattern in early-stage hypotheses than people give credit for.

2. The statistical substrate of language was already known

The wiki page makes the argument de novo, which it shouldn’t have, because by the early 2010s most of the components were textbook.

Zipf’s law

George Kingsley Zipf in 1932 — and Alfred Lotka before him — noticed that if you rank words in a corpus by frequency, the n-th most common word appears about 1/n as often as the most common one. The most frequent English word (“the”) appears about twice as often as the second (“of”), three times as often as the third (“and”), and so on. The pattern is shockingly robust: it holds in every natural language ever measured, it holds across topics within a language, it holds across centuries within the same language, and it holds at the character level too. The frequencies are not random. They sit on a power law that any sufficiently large text obeys.

The implication for the wiki claim is direct. If letter and word frequencies are stable lawful features of a language, then a brain that processes language has had millennia to internalise those frequencies and to use them — as a kind of background model against which surprise is computed. The information content of a token is proportional to the negative log of its frequency. Surprise is information, in the technical sense Shannon nailed down in 1948.

Shannon’s word game

Speaking of Shannon: in A Mathematical Theory of Communication he introduced what we now call n-gram models of English, ran them forward, and showed that you can generate increasingly English-looking text by sampling from progressively higher-order n-grams. Order-0 gives uniform letter distribution: garbage. Order-1 gives the right letter frequencies but random sequences: still garbage but lumpy. Order-2 (bigrams) starts producing pronounceable nonsense. Order-3 (trigrams) starts producing word-shaped tokens. Order-5 or so produces text that, sentence by sentence, looks like English — even though it has no semantic content.

The 1948 paper essentially demonstrated that statistical structure captured at the letter level is most of what makes text look like language. This is the same argument the wiki page was groping toward, twenty-six years before the wiki page. I should have read Shannon more carefully. So should everyone.

Cryptanalysis: anagrams as source code, for real

The discipline that has taken letter-frequency analysis most seriously for the longest is cryptanalysis. Al-Kindi, in 9th-century Baghdad, wrote Risāla fī Istikhrāj al-Muʿammā — “On Extracting Encrypted Letters” — which laid out frequency analysis: count the letters in the ciphertext, match them against the known frequency distribution of the plaintext language, and you have probabilistic guesses for each substitution cipher’s mapping. From the 9th century forward, this was the dominant attack on substitution ciphers, and every serious encryption system since has been designed with frequency analysis as the first thing to defeat.

Modern cryptography talks about confusion and diffusion, terms Shannon coined in 1949. Confusion makes the relationship between key and ciphertext complex. Diffusion spreads the statistical structure of the plaintext across many positions in the ciphertext, so that no local letter-frequency pattern survives. AES is built to destroy the patterns Al-Kindi exploited. The fact that we have to destroy them to get a secure cipher tells you that the patterns are real, that they carry information, and that a sufficiently patient algorithm can read them.

Reading the letter-frequency layer of text is, literally, reading the source code of the message in a way that the writer didn’t put there on purpose. The wiki claim was right about this. It just framed the phenomenon through anagrams when it should have framed it through cryptanalysis.

3. Sub-semantic channels: what the substrate actually carries

Letter frequencies are the boring part of the substrate. The interesting parts are the ones where the sub-semantic structure appears to carry actual meaning-tinted information, not just statistics.

Phonosemantics and sound symbolism

The textbook example is the bouba/kiki effect, first observed by Wolfgang Köhler in 1929 and rediscovered repeatedly since (most famously by Ramachandran and Hubbard in 2001). Show a subject two shapes — one rounded and blobby, one sharp and spiky — and ask which one is called “bouba” and which is called “kiki.” Across languages, across age groups, even across cultures that don’t use roman letters, something like 95% of subjects assign “bouba” to the rounded shape and “kiki” to the sharp one. The shapes have no inherent names. The sounds are not “really” round or sharp. But the cross-modal association is robust to the point of being one of the most replicable findings in experimental psychology.

The mechanism, as best as anyone has been able to nail down, is some combination of articulation gesture (lip rounding for “bouba,” tongue pointing for “kiki”) and high-frequency / low-frequency acoustic content. The point for our purposes is that the sounds themselves carry semantic associations, prior to and independent of any linguistic convention.

Phonesthemes

A phonestheme is a sub-morphemic sound cluster that carries meaning across a family of unrelated words. English has several. The cluster gl- at the start of a word leans visual, often to do with light or sight: glow, gleam, glint, glitter, glance, glare, glimpse, glisten. The cluster sn- at the start of a word leans nasal: sniff, snore, snort, sneeze, snout, snot, snarl. The cluster sl- leans toward smoothness or unpleasantness or both: slip, slide, slick, slime, sludge, slop, slush. The cluster -ump leans toward roundness or impact: bump, lump, hump, dump, jump, slump, clump, stump, thump, rump.

None of these are absolute. Plenty of gl- words have nothing to do with light (gland, glue, glib). The claim is statistical, not categorical: the cluster shifts the probability distribution of meanings the word will carry, and brains pick this up. Margaret Magnus’s PhD thesis (1999, University of Trondheim) catalogued English phonesthemes systematically and argued that they are sub-morphemic semantic carriers that can be exploited deliberately by writers and that are exploited unconsciously by language drift.

The wiki page was, in a hand-wavy way, gesturing at exactly this kind of phenomenon. The technical literature was already considerable when I wrote it. I hadn’t read enough of it.

Rhythm, meter, and persuasion

The substrate also shows up in prosody. Iambic pentameter is not magic, but the fact that English speeches and slogans skew strongly toward stressed-unstressed alternation is not coincidence either. The “rule of three” in rhetoric — life, liberty, and the pursuit of happiness; of the people, by the people, for the people; veni, vidi, vici — is older than English itself and crosses every Indo-European language. Try replacing any of those with a two-clause or four-clause version and the difference in memetic stickiness is immediate and brutal. The substrate cares about rhythm. Brains are entrainment machines and stickiness rides on entrainment.

The cumulative point is that language carries information on a sub-semantic channel that includes letter frequencies, phonemic clusters, sound-symbolic associations, and rhythmic structure. The channel is real, it has been characterised in the linguistics literature for nearly a century, and it does load-bearing cognitive work that readers are not consciously aware of. The wiki page wasn’t inventing this. It was rediscovering it through a particular lens (anagrams) that happened to be a sideways view of the underlying phenomenon.

4. How LLMs ended up doing this on purpose

A transformer language model is trained, mechanically, to do exactly one thing: predict the next token, given the previous tokens. The training corpus is some enormous slice of human-generated text. The loss function is cross-entropy on the next-token distribution. There is no semantic objective. There is no fact-checking. There is no “understanding” in any of the senses philosophy of mind has been arguing about for a century. There is only fit the next-token distribution as well as possible given what came before.

The architecture is descended from a chain of older statistical language models. The lineage goes roughly:

Markov chains (1906) — sample the next character or word from a distribution conditioned on the previous one.
n-gram models (Shannon 1948, then many) — condition on the previous n-1 tokens.
Class-based language models (Brown et al. 1992) — cluster words into classes so the conditioning is on the classes, not the tokens.
Neural language models (Bengio et al. 2003) — replace the count-based conditional with a small neural network.
Word embeddings (Mikolov et al. 2013, Word2Vec) — represent words as vectors in a space where geometric relations encode semantic relations.
Sequence-to-sequence with attention (Bahdanau et al. 2014) — let the model focus on different parts of the input dynamically.
Transformers (Vaswani et al. 2017) — drop recurrence; everything is attention.

Each step in that chain increased the order of statistical conditioning the model could capture. A bigram model conditions on one previous token. A 5-gram model on four. A transformer with a several-thousand-token context conditions on thousands. And — this is the load-bearing observation — when you make the conditioning long enough and rich enough, things that look like reasoning, planning, analogy, and even mild self-awareness fall out of the model. None of them were explicitly trained for. They emerge from optimising the next-token distribution well enough on enough text.

This is exactly the result the wiki page predicted. Not in the way I expected — I imagined some hand-crafted “meme-worthiness scorer” that explicitly modelled phonesthemes and letter frequencies, and what actually happened was a brutally simple architecture that learned all of that and a lot more by gradient descent on next-token prediction over a few terabytes of human-generated text. But the structural claim is the same. The cognitive content of language is largely carried by statistical patterns over tokens. A sufficiently large machine trained on enough of that statistical structure can produce text that exploits the same channels readers exploit unconsciously when producing or evaluating their own.

5. What LLMs validated, and what they didn’t

The wiki claim broke into a strong form and a weak form. The strong form was something like anagrams have inherent meaning. The weak form was the statistical substrate of language carries cognitive weight that brains pick up without conscious awareness.

LLMs have decisively validated the weak form. They produce text that humans rate as compelling, persuasive, emotionally resonant, and sometimes even insightful, using nothing but the statistical structure they extracted from training data. The compellingness is not a semantic add-on. It rides on the same channels. If it were purely semantic, an LLM with no model of truth, agency, or world state could not produce compelling text. They obviously do.

LLMs have also surfaced a related fact that the wiki claim did not anticipate: the substrate is not just letter-frequency or phoneme-cluster. It is the full sequence-conditional distribution. Transformer attention learns long-range dependencies between phrases many sentences apart, between turn-taking patterns in dialogue, between rhetorical-figure setups and payoffs, between argument structures, between voice registers. The “fabric” the wiki page named in passing turned out to be far richer than the letter-frequency diagnostic suggested. The patterns include letter frequencies, yes, but also bigram-and-up structures, syntactic templates, rhetorical moves, narrative arcs, and stylistic signatures. All of it is in the statistics. The wiki claim was right that the substrate existed; it was wrong about how thick the substrate is. It is far thicker than I imagined.

The strong form — anagrams as a Rosetta Stone — LLMs have not validated and probably never will. Anagram structure is one projection of the underlying frequency distribution. It contains some information about the source but not in the way I was reading it. The “emotional synesthesia” I felt when scoring anagrams was almost certainly me feeling the underlying phonesthemic and rhythmic structure of the source phrase, channeled through the anagram-shaped filter I was applying. The filter was incidental. The signal was real.

6. Implications: what to do with a validated substrate

Once you grant that the substrate exists and that machines can now explicitly operate on it, several things change.

Content optimisation as substrate exploitation. Every recommendation algorithm running on social platforms today is performing a version of the “meme-worthiness scorer” the wiki page called for. The mechanism is not that they explicitly score phonesthemes; the mechanism is that they score engagement, which correlates with the deeper substrate properties because engagement is the substrate’s signature in human behaviour. The platforms are strip-mining the substrate without knowing it has a name. The 2014 proposal has been built, badly, by every company whose business depends on attention.

Prompt engineering as deliberate substrate operation. The discipline that has emerged around “prompting” LLMs is, in a real sense, the consciously-engineered version of what skilled writers were doing instinctively before. A prompt that reliably produces a certain register, voice, or argument structure is one that activates the right region of the substrate. People who are good at this often report that they can “feel” when a prompt will work before they run it — the same feeling skilled writers report. The substrate is the same; the target machinery has changed.

Adversarial substrate exploitation. If statistical patterns in text affect human cognition below conscious awareness, then a sufficiently large model that learns those patterns can produce text deliberately tuned to particular cognitive vulnerabilities. This is not hypothetical. The combination of generation (LLMs that can produce convincing copy on demand) and targeting (platforms that can deliver custom-tuned copy to specific subpopulations) and iteration (automated A/B testing of message variants) is industrial-scale substrate exploitation. There has been no equivalent in the history of communication. The printing press could only mass-produce one copy at a time; this can mass-produce a million variants. Whether that turns out to be a slow disaster or a tolerable nuisance depends on defences that we are nowhere near ready to deploy.

Defensive substrate awareness. The cleanest defence is the same one that works against any cognitive bias: knowing the mechanism exists weakens its effect. People who know that mortality salience biases their political judgments make better political judgments under mortality salience. People who know that a piece of copy was produced by an LLM tuned for engagement read it more skeptically. The substrate cannot be removed — it is the same substrate that makes language work at all — but its exploit cases can be recognised and discounted. Memetic hygiene becomes a thing you can teach.

7. The intuition’s place in the genealogy

A small note on credit. I am not claiming I “predicted” LLMs in any serious sense. The component ideas — that statistical structure matters for cognition, that language has sub-semantic channels, that machines could in principle be built to operate on those channels — were in the literature in scattered form when I wrote the wiki page. I hadn’t read them. What I did was hit on the structural insight from a sideways angle (anagrams), state it loosely, and intuit a research program (the “meme-worthiness scorer”) that nobody, including me, built.

Then a different research program — neural language modelling, scaled absurdly, with no theoretical commitment to substrates or phonesthemes or anything else other than next-token prediction — happened to materialise the same observation as a working artifact. The artifact is more general than what I imagined; it confirms the structural claim while being silent on the diagnostic I was using.

I write this up partly because it’s gratifying when a hunch turns out to have been pointing at something real. But mostly because the structural lesson seems worth holding onto. When you notice a pattern that nobody else is talking about, even if your diagnostic for it is wrong, write it down. The diagnostic can be corrected later. The pattern, if real, will eventually be hit by some unrelated research program from another direction, and your having written the hunch down will save you (and others) the time of re-deriving it.

8. What’s next on the substrate

LLMs settled the question of whether the substrate exists and whether it can be operationalised. The open questions now are quieter and deeper.

What other latent dimensions of language carry cognitive weight that we haven’t named yet? Phonosemantics, rhythm, and frequency are the ones we have. There are surely others. Some are likely to be visible to scaling laws and emergent capability studies in larger models. Some may require new instruments.

How much of human cognition is itself the substrate? The honest implication of the LLM result is that a substantial fraction of what we experience as our own thinking is statistical pattern completion over linguistic input. This is uncomfortable. It is also probably true to first order, and the question is how to live with it without either collapsing into nihilism or pretending it isn’t so. The neuroscience says the same thing from another direction (predictive processing, Bayesian brain, free-energy principle). Two literatures converging on the same uncomfortable conclusion from independent directions usually means the conclusion is right.

Can the substrate be inoculated against? See the defensive substrate-awareness discussion above. The early signs are that yes, partially, with effort. The economics work against the defence; the production side is industrialised, the defence side is artisanal. Closing that gap is one of the more important communication-design problems of the next decade.

What happens when LLMs train on LLM output? This is the distinctively-new problem. Once a meaningful fraction of training text is itself generated, the model is learning the substrate’s image of itself rather than the substrate as humans produced it. The fixed points of that recursion are not the same as the fixed points of the original substrate. We are running this experiment globally without much instrumentation. I have no clean prediction for how it ends.

I’ll close where I started. The wiki page in 2014 (or thereabouts) was written in rough prose, with several typos and at least one bad diagnostic. The structural claim was correct: language has a substrate of statistical pattern that does cognitive work, the substrate is exploitable, and a sufficiently large machine could be built to operate on it. That machine got built. It works. It is already changing the politics of attention more than the printing press did.

Anagrams turned out not to be the right window into the substrate. But the substrate was there, and somebody — though not me — has now built the equivalent of an x-ray for it. The next few years will be about deciding what to do with the x-ray. I think the answer involves more substrate awareness, not less. The frame the mimetics essay set up still applies: knowing the mechanism is half of being able to defend against it.

The other half is harder.

1. Stating the hypothesis precisely

Let me lay out the speculation in as specific a form as I can, so we can see where the load-bearing assumptions are:

Premise 1. Liquid water is a network of ~10²² hydrogen bonds per millilitre, rearranging on a timescale of ~1 picosecond, governed by local rules (the “ice rules” — each oxygen donates two and accepts two hydrogen bonds) plus thermal noise.
Premise 2. This network has the right formal structure to support computation in the loose sense — graph-rewriting on a constrained tiling, with energy gradients steering the dynamics.
Premise 3. At biological temperatures and pressures, quantum effects (proton tunneling, vibrational coherence, possibly nuclear spin coherence) persist on the relevant timescales in specific microenvironments inside cells.
Premise 4. A subset of these quantum effects are not noise, but contribute load-bearing computation to processes that biology has evolved to exploit.
Conjecture. The same medium that hosts those quantum effects, and the same global network that those effects are embedded in, is performing information processing of a kind we have not yet characterised. Evolution is the slow, large-scale shadow of that processing on the genome.

The conjecture is the speculative bit. Premises 1-3 are mainstream-ish (I will defend Premise 3 below). Premise 4 is partially mainstream (in photosynthesis it is essentially conceded) and partially contested (elsewhere). Premise 4 plus the conjecture is the move that takes us from quantum biology — which is real and credentialed — to something much stronger.

If the conjecture is true in any meaningful sense, “water chose to evolve” is not a category error. It is the statement that the network’s computation prefers certain outcomes, and biology is one of the attractors.

2. Why water is the only plausible candidate

If you wanted to design a molecular substrate for warm, wet, Avogadro-scale parallel computation, you could not do much better than liquid water.

A networked medium with hard local constraints. The ice rules give the hydrogen-bond graph a constraint topology that resembles, more than it resembles anything else, a frustrated lattice. Frustrated lattices in condensed matter are exactly the systems that exhibit non-trivial low-temperature collective behaviour — spin liquids, spin ices, emergent gauge fields. The Bernal-Fowler ice rules have the same form as the constraints in spin ice, where they give rise to magnetic monopole excitations. Whatever else liquid water is, formally it is a frustrated network whose dynamics are not trivial.

A picosecond clock. Hydrogen bonds break and reform on the order of 10⁻¹² seconds. That is a gigahertz natural clock rate at room temperature, achieved without any external driver. Compare to silicon at room temperature: thermal noise dominates anything you try to compute below ~k_BT, which is ~25 meV — about an order of magnitude less energy than a single hydrogen bond.

Quantum effects survive. Proton tunneling is documented in enzymatic catalysis and in ice itself. Vibrational coherence between chromophores in the Fenna-Matthews-Olson photosynthetic complex was demonstrated to persist for hundreds of femtoseconds at 277 K (Engel et al., Nature, 2007), and the original interpretation that this is load-bearing in photosynthetic efficiency has been refined but not abandoned in the years since. Radical-pair coherence in cryptochrome, the molecule implicated in avian magnetoreception, requires that electron spin coherence survive long enough — on the order of microseconds — for a few-microtesla magnetic field to perturb the reaction outcome. This is genuinely strange and has held up under twenty years of probing.

Avogadro-scale parallelism. A teaspoon of water is ~10²³ molecules. Even if only one in a trillion is doing something computationally interesting at any moment, that is still ~10¹¹ parallel operations per teaspoon per picosecond. Conventional silicon, scaled to the same volume, manages perhaps eleven orders of magnitude less.

Interfaces do something. Pollack’s exclusion-zone work, which I treated with extreme skepticism in the first essay, does report a real and reproducible observation at hydrophilic interfaces: a hundreds-of-micrometres-thick region of water that excludes solutes and has slightly different properties. The mainstream explanation is electrokinetic, not a “fourth phase,” but the underlying empirical fact — that water near a structured surface organises itself in a way that propagates a long way into the bulk — is real. Armstrong’s water bridge is another such observation. The Strano group’s CNT results are a third. Whatever is happening at water/surface interfaces, it is not the same as bulk water.

You do not have to believe any of this implies consciousness. You only have to grant that water’s hydrogen-bond network is a non-trivial computational substrate by any reasonable definition of “computational substrate.”

3. Real quantum biology, and where water actually is

The strongest empirical support for any version of this thought experiment comes from quantum biology, a field that did not exist under that name until about 2007 and which now has its own journals, conferences, and grant pipelines.

Photosynthesis. The Fenna-Matthews-Olson (FMO) protein in green sulfur bacteria channels excitations from light-harvesting antennas to the reaction centre. The original 2007 result by Greg Engel, working in Graham Fleming’s group at Berkeley, showed that the energy transfer is not classical hopping but involves quantum coherence — the excitation explores multiple paths simultaneously and the system selects the most efficient one by wave interference. The 277 K coherence lifetime was unexpectedly long. Subsequent work (Cao, Cheng, and others) refined the picture: the coherence is partially environment-assisted, with the protein and its surrounding water tuned to neither destroy the coherence too fast nor preserve it so long that it traps in a local minimum. Photosynthetic organisms appear to have evolved a Goldilocks decoherence rate.

Cryptochrome and magnetoreception. European robins navigate by the geomagnetic field, and the orientation behaviour is wavelength- dependent in a way that points to a light-induced radical pair in the eye. The molecule is cryptochrome; the mechanism, originally proposed by Schulten in 1978, was that an electron transferred from one part of the molecule to another produces a transient radical pair whose nuclear-spin-coupled triplet/singlet ratio depends on the ambient magnetic field. The reaction yield therefore depends on the field direction. Modern work (Hore, Mouritsen, and collaborators) has made this picture quantitative; in 2021, Xu et al. demonstrated the field sensitivity in vitro with cryptochrome 4 from migratory robins, and showed it was absent in non-migratory chickens.

Enzymatic tunneling. Many enzymes catalyse reactions by allowing hydrogen nuclei to tunnel through activation barriers rather than go over them classically. The signature is a kinetic isotope effect larger than what the classical Eyring equation allows. Soybean lipoxygenase, alcohol dehydrogenase, dihydrofolate reductase — all have documented tunneling contributions. The picture that has emerged is that proteins position substrates with sub-angstrom precision such that the tunneling probability becomes biologically useful, and that this positioning is achieved partly through coupling to picosecond protein motions.

Where is the water in all this? Everywhere. The FMO complex sits in a hydrated environment; the cryptochrome radical pair is surrounded by structured water near its binding pocket; the enzymatic tunneling reactions all happen in solvated active sites where the local water network is a non-negligible part of the reaction coordinate. The mainstream view is that the protein is doing the work and water is the bath; the speculation in this essay is that this division is wrong, or at least incomplete, and that the water network is doing more than absorbing heat.

4. Penrose-Hameroff and the credibility budget

The most famous (and most criticised) attempt to ground consciousness in physics is Roger Penrose and Stuart Hameroff’s Orchestrated Objective Reduction (Orch-OR) theory. The proposal, dating from the mid-1990s and refined since, is that consciousness arises from quantum-coherent computation in microtubules inside neurons, “objectively reduced” by Penrose’s interpretation of gravity-induced wavefunction collapse.

The mainstream response, distilled charitably, is: the gravitational mechanism is unmotivated and untestable, the microtubule decoherence times calculated by Tegmark (2000) are absurdly short compared to neural timescales, and the entire framework has the smell of solution-in-search-of-a-problem. Penrose and Hameroff have answered each of these objections, the answers are not universally convincing, and the theory sits in the credentialed-but-fringe category.

Why mention it here? Because Orch-OR is the existing structure most similar to the speculation in this essay, and because the non-fatal version of its argument — that biological systems might host load-bearing quantum coherence at biological temperatures — has held up better than the specific microtubule claim. The fight has moved. In 2024, work by Jack Tuszynski and others reported optical evidence of long-lived superradiance in microtubules, which does not prove Penrose right but does prove that the question is not closed.

If you take the position that the universe permits warm, wet biological quantum computation at all — and quantum biology has forced you to take that position, whether you wanted to or not — then the question becomes which substrate, where, and on what scale. The microtubule answer is one possibility. The hydrogen-bond network of intracellular water is another. They are not exclusive, and indeed they are not separable, because the microtubule interior is itself a structured-water environment with reported anomalous properties.

5. What would it mean for water to “choose”?

Here is where the speculation gets sharpest, and where I have to be most careful.

The weak reading of “water chose to evolve” is that water’s physical properties acted as such a powerful selection pressure on prebiotic chemistry that the trajectory of life on Earth was very strongly funnelled by them. This is the anthropic principle, essentially, and it is uncontroversial — except in its strong forms. The hydrophobic effect requires lipid membranes to compartmentalise. The dielectric constant requires ion-gradient bioenergetics to be possible. The density anomaly requires aquatic ecosystems to be possible. Knock out water’s specific package of anomalies and the design space of possible biospheres shrinks dramatically.

The strong reading is that water performs some form of computation whose long-time average prefers certain biochemical configurations over others, and that this preference is not just thermodynamic but informational. In this reading, the “fitness landscape” that evolution climbs is partially generated by the water network, not just by selection on replicators. Different mutations would have different effects not only because of their classical chemistry but because of how the surrounding water network “computes” with them — preferring certain conformations, opening certain reaction channels.

I do not know how to make the strong reading precise enough to test. That is the honest answer. But here are some forms it might take:

Reservoir computing. The water network around a protein might function as a reservoir computer — a high-dimensional dynamical system whose transients project the protein’s local degrees of freedom into a much larger state space, where decision-making becomes easier. Reservoir computing is a real and well-studied paradigm; whether biological water actually does it has not been tested.
Quantum sampling. Coherent superpositions of hydrogen-bond configurations might let the water network sample conformational space faster than classical thermal exploration would. This is the “quantum walk” idea from FMO applied at the water-network scale.
Information geometry. The hydrogen-bond graph at the surface of a protein might encode information about the protein’s function in a way that is not encoded in the protein’s sequence. This would predict that proteins with similar sequences but different surface hydration networks would have measurably different binding kinetics — which is, in fact, observed, although the mainstream explanation is “specific water-mediated contacts” rather than network-level computation.

None of these are tested in the strong form. All of them are testable in principle. That distinguishes them from the unfalsifiable end of the speculation pool.

6. Predictions and falsifiability

For an essay that is openly speculative, this section is the most important. If the speculation in this essay is right in any specific form, what would we expect to see?

Decoherence-tuning across phyla. Quantum-biology systems should show evidence of being tuned — neither too fast nor too slow — and that tuning should track phylogeny in a way that reveals selection pressure. (Partly observed in FMO across green sulfur bacteria.)
Anomalous information capacity in hydration shells. Two proteins with identical sequences but different ortho/para water environments, or different isotope composition (H₂O vs D₂O), should have different functional behaviour beyond what classical kinetic-isotope-effect calculations predict. Some D₂O / H₂O substitution experiments do show effects larger than KIE would account for; the standard explanation is solvent dynamics, but the speculation here would predict the magnitude differently.
Long-range correlations in cell-volume water. A cell is mostly water, and the speculation predicts that water in cytoplasm should show coherent dynamical correlations on length scales larger than the bulk Debye length. Femtosecond spectroscopy in cytoplasm vs buffer is an experiment that can be run today.
Distinct quantum signatures in cancer cells. A wilder prediction: cells whose proteomic and metabolic state has been disrupted might show measurable differences in their hydration coherence relative to healthy cells. This is the kind of claim that gets exploited by quacks selling test kits; doing it properly requires the same kind of careful blinded protocols that quantum biology has been forced to adopt.
The Mpemba effect revisited. If water has any kind of information-bearing internal state, the Mpemba effect could be the thermodynamic shadow of preparation-dependent network configurations. A specific prediction: the Mpemba effect should be stronger in solutions of biopolymers than in pure water at the same dissolved-gas content. This is testable.

The point of listing falsifiable predictions is to distinguish this speculation from the unfalsifiable cousin claims that water has “memory” or “vibrational consciousness.” A real speculation should either die or get refined when contact is made with experiment.

7. Adjacent woo to not be confused with

Because the surface vocabulary overlaps, I want to be explicit about what this thought experiment is not.

Not Masaru Emoto. Emoto’s claim — that water exposed to “loving words” forms beautiful crystals on freezing while water exposed to “hateful words” forms ugly ones — is unblinded, unreplicated, and incompatible with the picosecond rearrangement time of the hydrogen-bond network. There is no mechanism by which liquid water at room temperature could retain a memory of an emotional stimulus applied to its container.

Not Pollack’s “fourth phase.” I cited the EZ observation as a hint that water at interfaces does something interesting structurally. I did not endorse the interpretation that this is a new phase of water with healing properties, and the broader health claims continue to fail independent replication.

Not homeopathy. Water memory in the Benveniste sense — claiming that dilution past the point where any solute molecule remains can still produce a pharmacological effect — is incompatible with straightforward physics and chemistry and has not survived blinded replication.

Not “structured water” supplements. Devices that vortex, magnetise, or otherwise treat tap water and sell it as “structured” or “hexagonal” are exploiting the real complexity of water dynamics for marketing purposes. Whatever might or might not be true about information processing in cellular water cannot be replicated by running tap water through a magnet.

The speculation in this essay is that the cellular hydrogen-bond network, in coupling with biomolecular machinery that evolved over four billion years to exploit it, might be doing computational work we have not characterised. Nothing about that speculation justifies buying $40 spring water that has been “energised” by a quartz crystal.

8. What this would change, if true

Suppose the speculation is right in some specific, testable form — say, that biological water networks function as reservoir computers and that protein function depends measurably on this. What changes?

Drug design changes. Currently we design drugs to fit a protein’s binding pocket and to be soluble enough to deliver. We do not design them with respect to the local water network’s computational state. If proteins are partially functioning through their hydration shells, then small-molecule design that ignores the water network is leaving something on the table — and with respect to which a chiral drug might have a non-classical effect that no traditional QSAR captures.

Evolution looks slightly different. We currently model fitness as a function of phenotype, which is a function of genotype-expressed-as- protein. If hydration networks add a layer in between — fitness as a function of phenotype-as-water-network-modulated-protein — then the fitness landscape gets new structure that was previously invisible. Some “neutral” mutations might not be neutral at the water-network level.

Origin-of-life questions get a new flavour. The Miller-Urey-style story of life originating in primordial soup involves water as the medium. If the medium is actively computing, then the question “why did life arise in water” gets a different answer: because water was already running the appropriate computation and biology was a way to stabilise it.

And the deeper question — why this universe is configured so that water (and therefore life) is possible — gets, at least, a different frame. Not “anthropic principle: we observe a life-supporting universe because only those universes have observers” but “computational principle: information-processing media bootstrap their own propagation through whatever chemistry is available.”

I want to emphasise once more: I do not know whether any of this is true. I have laid it out carefully because it is the form of the speculation that could be true, given what we already know about quantum biology and water dynamics, and because that form is far more interesting than the marketing version.

Closing

The honest answer to “is water a quantum computer that chose to evolve” is probably not, in any literal sense. The honest answer to “is the hydrogen-bond network of cellular water performing non-trivial information processing that biology has co-opted” is we do not know yet. Quantum biology forced the question open. The answers are not in.

The thought experiment is useful because it pushes back on the polite mainstream framing that water is merely the solvent. If photosynthesis is exploiting quantum coherence, if avian magnetoreception is exploiting radical-pair coherence, if enzymatic catalysis is exploiting tunneling — and the medium hosting all of those phenomena is the same hydrogen-bond network — then the question of what else that network is doing becomes legitimate.

We may end up concluding that water is a passive bath after all, and that what looks like network-level computation is just the sum of local effects. Or we may end up with a richer picture in which the network itself is part of the story. Both endings are interesting. What is not interesting is the version of the question that gets sold in a bottle.

I wrote the first essay to defend water against the marketing. I wrote this one to defend the genuine speculative question against the same marketing. They are different defences but they have the same enemy: the comfortable certainty that there is nothing left to think about here, in either direction.

Comments, corrections, refutations welcome. Especially refutations.

References and further reading

Engel, G. S. et al. “Evidence for wavelike energy transfer through quantum coherence in photosynthetic systems.” Nature 446, 782-786 (2007). The FMO result.
Lambert, N. et al. “Quantum biology.” Nature Physics 9, 10-18 (2013). A readable overview of the field as it stood after FMO.
Tegmark, M. “Importance of quantum decoherence in brain processes.” Phys. Rev. E 61, 4194 (2000). The standard criticism of Orch-OR.
Hameroff, S.; Penrose, R. “Consciousness in the universe: A review of the ‘Orch OR’ theory.” Phys. Life Rev. 11, 39-78 (2014). Hameroff and Penrose’s own restatement.
Tuszynski, J. A.; collaborators. Recent superradiance work in microtubules, late 2010s onward.
Xu, J. et al. “Magnetic sensitivity of cryptochrome 4 from a migratory songbird.” Nature 594, 535-540 (2021). The robin cryptochrome result.
Hore, P. J.; Mouritsen, H. “The radical-pair mechanism of magnetoreception.” Annu. Rev. Biophys. 45, 299-344 (2016).
Klinman, J. P.; Kohen, A. “Hydrogen tunneling links protein dynamics to enzyme catalysis.” Annu. Rev. Biochem. 82, 471-496 (2013).
Bernal, J. D.; Fowler, R. H. “A theory of water and ionic solution, with particular reference to hydrogen and hydroxyl ions.” J. Chem. Phys. 1, 515 (1933). The original statement of the ice rules.
Castelvecchi, D. “Quantum effects bring a strange twist to biology.” Nature 619, 18-20 (2023). Status report on the broader field.
Cao, J. et al. “Quantum biology revisited.” Sci. Adv. 6, eaaz4888 (2020). The mainstream’s current view of where the strong claims and weak claims stand.

Mimetics: The Third Replicator

In 1976 Richard Dawkins, in the closing chapter of The Selfish Gene, coined the word meme to name a unit of culture that copies itself between minds the way a gene copies itself between bodies. The proposal was deliberately analogical. Dawkins was making a structural argument: wherever you have heritable variation, differential fitness, and a copying mechanism, you get evolution by natural selection, and the substrate doesn’t have to be DNA. Cultural ideas — tunes, catchphrases, fashions, religions — satisfy the criteria. So they evolve.

Twenty-some years later, Susan Blackmore took the framework and pushed it one more step. In The Meme Machine (1999) and in a 2008 TED talk that I keep coming back to, she argued that the Internet plus silicon constitutes a third class of replicator she called temes: units of information that copy themselves not between minds (memes) or between bodies (genes) but between machines. “Earth now has three replicators,” she wrote, “genes (the basis of life), memes (the basis of human culture) and temes (the basis of technology). I argue that the information copied by books, phones, computers and the Internet is the beginning of this third replicator and consequent new evolutionary process.”

I want to take Blackmore one step further. The unifying concept across genes, memes, and temes is not the substrate — DNA, neuron, silicon — but the act of copying. The Greek word for that act is mimesis. The right name for the meta-framework that encompasses all three substrates is mimetics: the study of self-replication or copying in any system. Memetics is a subset of mimetics. So is genetics. So, in the long term, will be whatever we end up calling the study of machine-replicated information. That extension is mine, and the rest of this essay is what falls out of it.

1. Mimetic versus memetic — a distinction worth keeping

The English words are confusingly close.

Mimetic (mə-met-ik) is the adjective form of mimesis, from the Greek μίμησις, “imitation.” It’s an old word. It shows up in Plato’s Republic — mimesis contrasted with diegesis, imitation versus narration. In biology it names a category of phenomenon: a stick insect is mimetic of a twig; a hoverfly is mimetic of a wasp. The adjective applies to anything that resembles, imitates, or copies something else by any means.

Memetic is the adjective form of meme, the much younger word that Dawkins built specifically to parallel gene. A unit of cultural inheritance: a tune you can’t get out of your head, a turn of phrase that propagates because it’s catchy. Memes are idea-shaped by definition. The substrate is the mind, the carrier is communication.

The two words overlap because all memetics is mimetic — a meme is, by definition, a copied idea — but not all mimesis is memetic. DNA copying itself is mimetic but not memetic. A 3D printer cloning its own gears is mimetic but not (yet, quite) memetic. Mimicry in biology is mimetic and is, in a particular sense, adversarial memetics played out at the species level: the harmless butterfly that copies a poisonous one’s wing pattern is exploiting a meme — the predator’s learned “don’t eat this” association — by faking the carrier.

The distinction matters because the moment you grant that copying-as- such is the unit of analysis, you can stop arguing about whether something is “really” a meme or “really” a gene and start looking at the structural commonalities. Every act of mimesis has the same ingredients: a source, a copy, a copying mechanism, an error rate, a payload, and a selection environment. Different substrates differ in the details. The structure is the same.

2. Blackmore’s third, and what comes after

Dawkins gave us two replicators. Blackmore gave us three. The provocation in Blackmore’s framing is that temes are no longer running on human brains. Once Wikipedia, search engines, image recognition, and now large language models exist, the substrate has moved off neural tissue and onto silicon. The replicator and the host are no longer the same kind of thing. A meme needed a human to remember it. A teme can be remembered, mutated, and propagated by a process that no human is in the loop of.

You can be cheerful or apocalyptic about this. Blackmore tends toward the apocalyptic — her line was that we are now the second-class citizens of our own information ecology, the way bacteria became the second-class citizens of theirs after multicellular life evolved. I’m less certain. The fact that machine-mediated copying has its own dynamics doesn’t necessarily imply that the dynamics are worse. It implies they are different, and that the toolkit we built for analyzing memes — designed for human-to-human transmission — needs revision.

The framework I prefer is this. Across all three replicators we have been gaining error-correcting, memory-extending, selection- imposing infrastructure for thousands of years. Writing was the first big jump — suddenly memes could survive the death of their human hosts, and could replicate across thousands of miles. Printing collapsed the cost of copying. Telegraphy compressed time. Radio and television opened the bandwidth. Each of these enhancements was mechanical in the sense that the medium did not filter or select the content. The printing press would print whatever you gave it. The radio would broadcast whatever you transmitted. The TV would show whatever was on the tape.

Then computers happened. And computer mediation is not content-neutral. The recommendation algorithm filters. The spam-detector filters. The Google ranking algorithm filters. The LLM, more aggressively than any of them, generates and re-generates. We have moved from the era of machine-enhanced memetics, where the machine extended the carrier but didn’t touch the cargo, to the era of computer-enhanced memetics, where the machine is selecting, remixing, and producing memes — not merely transmitting them.

That shift is the most important thing happening to the memetic substrate in the lifetime of anyone reading this. Memes used to evolve by being passed mind-to-mind, with selection pressure imposed by what human brains found memorable. They still do that. But increasingly they evolve by being passed mind-to-machine-to-mind, with selection pressure imposed by what algorithms find clickable. The algorithms have a different fitness function. The result is a different distribution of survivors.

3. External replicators versus self-replicators

A useful structural division cuts across all the substrates: some replicators run on host hardware they do not own, and some replicate themselves all the way down.

External replicators are information bundles that need a host to copy them. Examples:

Viruses, biological. The virus is a payload that hijacks the host cell’s machinery to replicate.
Software viruses and worms. Same trick, different substrate.
Memes in the cultural sense. They need a human brain to copy them. A meme without minds to inhabit is silent.
Most software in general. A program is an external replicator that runs on an operating system that runs on hardware that the software does not itself build.

Self-replicators carry their own copying machinery:

Life, all the way down. A cell makes another cell; eventually, it makes another organism that makes another organism.
Some technologies, increasingly. The Reprap 3D printer was designed specifically to print most of its own parts. A milling machine can in principle make the parts of another milling machine. Industrial robots can assemble industrial robots, given the right parts. Karel Čapek’s R.U.R. (1920), the play that gave us the word robot, had self-replicating robots as its central conceit; Drexler’s Engines of Creation (1986) gave us “grey goo” as the nano-scale version.

The boundary between external replicator and self-replicator is dissolving faster than most discussions notice. A 3D printer plus a copy of its own design files plus enough source plastic is one generation away from self-replication. An LLM plus an environment in which it can take actions and modify its own model is one or two more generations away. Once that boundary closes, the meme/teme/gene distinction becomes harder to maintain — and mimetic becomes the only word general enough to describe what’s going on.

4. Errors, payloads, and the mimetic fabric

Every replicator has an error rate. DNA polymerase makes a mistake about once per 10⁹ bases, with proofreading; without proofreading the rate is closer to 10⁻⁴. A meme transmitted between humans corrupts faster: every retelling drifts. A teme copied between machines can be nearly perfect (a cp command) or can be deliberately lossy (JPEG compression, MP3, an LLM paraphrasing a paragraph).

Errors are the source of variation. Without error there is no evolution, only copying. The whole framework presumes a controlled mutation rate, high enough to generate variants for selection to act on, low enough that information isn’t lost faster than fitness can accumulate.

Every replicator also has a payload beyond its own machinery. A gene codes for a protein. A meme carries an emotional or behavioral charge. A teme carries content. The payload is what gets selected on — the gene that codes for a useful protein is more likely to survive than the one that codes for nothing useful; the meme that makes you laugh and share is more likely to spread than the one that does not. The selection pressure acts on the payload but propagates the replicator.

A nice consequence is the concept of the skeuomorph: an ornament or feature on a copied object that imitates a feature from the object’s older substrate, even though the new substrate makes the feature unnecessary. A wooden ceramic mug with carved “rivet” marks where the wood version had real rivets. A digital camera that emits a fake mechanical shutter click. Slack’s UI inheriting the visual language of IRC. Memes can carry skeuomorphs forward across substrate changes; in fact, almost every meme that survives a substrate change carries some.

5. Memetic engineering — the applied side

You can describe a mimetic system. You can also try to steer one. The applied discipline goes back further than the word for it. Priesthoods knew it. Generals knew it. Advertisers professionalised it. The Wired piece by James Gardner from 1996 — back when “memetic engineering” first hit print in a mainstream venue — captured the disquiet of recognising that culture-as-such might be the product of selfish memes co-evolving “with supreme indifference to their impact on human hosts.” That framing extended Dennett’s earlier provocation in Consciousness Explained: that human minds might be, in some sense, the residue of the memes that have inhabited them.

Memetic engineering is what you do when you treat memes as designable objects. A few of the techniques:

Gaslighting. Not originally a memetic concept — the term comes from a 1938 play and its 1944 film, in which a husband manipulates his wife into doubting her own perceptions by altering small features of the household environment and then denying the alterations. The mechanism generalises far beyond the marital case. Anywhere a meme can be propagated faster than the listener can fact-check it, gaslighting is available. The 21st century version is industrial-scale and runs on the same algorithmic-mediation infrastructure that propagates useful memes; the cost of injecting a corrupt belief is now lower than the cost of correcting it.

Peer pressure as memetic carrier. Pedro Gardete at Stanford looked at 65,525 in-flight purchases across 1,966 flights and found that if the person next to you bought something, your probability of buying something jumped by about 30%. The control was the person sitting in front of you, whose purchase you wouldn’t see. That is a memetic transmission curve measured under near-laboratory conditions in a near-laboratory environment (an airplane seat is hard to escape). Bursztyn and Jensen, separately, showed that when fliers offering a free SAT prep class made it clear that classmates would see who signed up, students in honors classes were 25% more likely to sign up, while students in non-honors classes were 25% less likely. The same announcement, opposite sign of effect, depending on which peer-pressure meme dominated the local environment. Memes don’t act in isolation; they act in competition with other memes already installed in the host population.

Resonance and competition. The classical model — meme as free-floating idea — misses that memes compete for the same neural real estate. A new meme typically has to displace an existing one, or attach to one already present, or fit into a vacancy left by some failed predecessor. The resonance with what is already there is load-bearing. This is why the same message can succeed in one audience and fail in another with no obvious difference between the audiences: their installed memetic base differs.

Computer mediation as compounding factor. All of the above existed before computers. What computers do is speed up the iteration cycle and impose a new selection pressure. The meme that goes viral on a recommendation-driven platform is the meme that maximises that platform’s engagement metric, which is a proxy for “keeps users on the platform” rather than for “true” or “good” or “useful.” The engagement maximisation function favours memes that trigger strong, fast emotional reactions: outrage, fear, tribal affirmation. The system rewards memetic engineering even when no human is doing the engineering on purpose — the platform’s optimiser is the engineer.

6. Language as memetic substrate

Spoken and written language is the original carrier wave for memes, and it has been shaped over hundreds of millennia of co-evolution with the brains that use it. Some properties of language are forced on it by being a memetic carrier subject to error-correcting requirements.

Languages have redundancy. Spoken language has to survive background noise, mispronunciation, and the receiver’s hearing loss. It does this by being redundant: word stress, intonation contours, and context all carry partial copies of the lexical content, so a missing or corrupted word can be reconstructed. Written language has analogous redundancies — fonts have evolved for legibility at distance, spacing and capitalisation carry parallel signals, and famously you can scramble the middle letters of words and still read the sentence:

It deosn’t mttaer in waht oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht frist and lsat ltteer is at the rghit pclae.

This is not a quirk; it is what the redundancy buys.

Languages have framing. A separate phenomenon: bilingual subjects make systematically different decisions in their non-native language. The classic experiment is the framing of a medical decision in terms of “200 of 600 saved” versus “400 of 600 dead” — same outcome, different framings produce systematically different choices in native-language speakers, and the framing effect diminishes when the same problem is posed in a foreign language. The emotional salience of a foreign language is lower; the framing effect rides on emotional salience; so the framing effect weakens. Implication: a substantial fraction of “rational” decision-making is actually being driven by emotional resonance with native-language framings. The memetic payload is doing more work than the propositional content.

Languages have hidden channels. The provocative claim — and this is mine, in the sense that I wrote it on the wiki years ago and have not seen anyone else state it the same way — is that letter frequency and structure are doing work in parallel to the propositional content. Anagrams, palindromes, reverse-audio in songs, the rhythmic structure of slogans: these aren’t decorations on the message. They are part of the message, on a channel that operates below conscious recognition. Skilled writers and speechwriters feel the channel; you can tell when a phrase “lands” and when it doesn’t, even before you can analyse why. The brain is doing pattern recognition on letter and phoneme distributions in parallel with semantic extraction. The two channels can reinforce or fight each other.

Why suspect this? Because in any sufficiently long-lived memetic substrate, every available channel will be exploited by selection. If letter-frequency patterns subconsciously affect how readers feel about a sentence, then writers whose feel produced emotionally-resonant patterns will out-propagate writers whose feel didn’t, and the language as a whole will accumulate those patterns. You don’t need a conscious mechanism for this to happen; you need only the same Darwinian sieve that operates on every other heritable trait.

This is, frankly, the part of the framework I’d most like to test empirically. A program that scores text on memetic propagation probability — adjusted for letter and phoneme distribution, not just semantics — would be measurable against actual propagation curves on social platforms. I haven’t built one. Someone should.

7. Terror Management Theory, and the memetic exploit

If memes compete for neural real estate, and if some memes activate deeper neural defences than others, then the memes that activate mortality-related defences will dominate.

This is the framework called Terror Management Theory, originating in Ernest Becker’s The Denial of Death (1973) and developed by Greenberg, Pyszczynski, and Solomon from the late 1980s onward. The central claim: humans are the only animals that know they will die, and the conflict between that knowledge and the wish to keep living generates a constant low-grade anxiety that we manage by investing in worldview — cultural narratives that promise some kind of continuation (literal afterlife, symbolic legacy, identification with a group that outlasts the individual).

The empirical bite of TMT comes from mortality salience experiments. Briefly remind subjects of their own mortality — through a questionnaire item, a flashed word too fast for conscious detection, a discussion topic — and their behaviour shifts. Judges asked to set bail for a hypothetical prostitute averaged $50 in the control condition and $455 after a mortality salience prime. White subjects who’d judged a white person’s racial-pride speech as racist rated the same speech as less racist after a mortality prime. Subjects given the opportunity to allocate hot sauce to a person known to dislike spicy food gave dramatically more hot sauce to ideologically opposed targets after mortality salience. The 2004 American presidential election appeared to swing measurably in favour of George W. Bush after death reminders — Kerry won under control conditions, Bush won after the prime, in the same subject pool. Bush represented the status quo worldview; Kerry was a threat to it; under mortality salience, subjects defended the status quo.

From a mimetics standpoint TMT is a vulnerability disclosure. There is a class of memes — death-reminders, mortality cues, threat imagery — that reliably increase the host’s susceptibility to worldview-reinforcing memes. A meme that contains a mortality prime will, in expectation, increase the next meme’s chance of taking root, provided the next meme aligns with the host’s existing worldview. This is exploitable. Of course it has been exploited. The “skulls in the whiskey-ice” subliminal imagery campaigns of the 1970s and 1980s look in retrospect like brute-force memetic engineering. The post-9/11 American political environment looks, in retrospect, like a sustained mortality-salience condition that affected an entire electorate’s susceptibility to a particular memetic payload.

The defence against this kind of exploit is not avoidance of news about mortality — that’s not feasible — but awareness of the mechanism. People who know they are in a mortality-salience state can correct, partially, for the worldview-defence overshoot. The worldview defence is unconscious; bringing it into consciousness weakens it. This is true of most memetic exploits, which is why memetics is, in part, a defensive discipline.

8. Wisdom of the masses — the constructive case

Not all of memetics is exploitation. The crowd-level integration of many noisy individual judgments is one of the genuinely cooperative phenomena in mimetic systems.

Bacteria, of all things, illustrate the principle cleanly. A swarm of bacteria foraging in a complex chemical landscape will, under some conditions, outperform fish or amoeba swarms at finding food and avoiding harm. The mechanism is individually-tuned confidence: when a single bacterium has reliable information about its local environment, it weighs its own signal heavily and the swarm’s signal lightly. When its local information is poor, it weights the swarm more. The result is a swarm that avoids what swarm theorists call erroneous positive feedback, the failure mode in which a small subgroup with wrong information drags the whole population in the wrong direction. (This failure mode is, incidentally, the same one that algorithmic recommendation systems suffer when they overweight early signals.)

Humans do something similar at scale. Tetlock’s “Good Judgment Project” — and its successor work — has documented that small populations of carefully-aggregated forecasters outperform CIA analysts at predicting geopolitical events, in some cases by 30% or more. The mechanism is not that the forecasters individually know more; it is that the aggregation cancels their individual biases. Each neuron in the brain is a bad classifier. The brain as a whole classifies well. Each forecaster is a bad geopolitical analyst. The crowd of forecasters is, when aggregated correctly, a good one.

The constructive case for memetics is that information-mediation infrastructure — Wikipedia, prediction markets, well-designed aggregation systems — can amplify the wisdom-of-crowds dynamic in ways that humanity has not had access to before. The destructive case is that the same infrastructure can amplify cascade failures and exploit the TMT vulnerabilities just discussed. Whether the net is positive or negative depends on design choices that are being made by companies whose objective functions are not optimised for wisdom-of-crowds outcomes.

9. Where this goes: the next replicator

Everything in the framework points the same direction. The substrate-blur is accelerating. We have:

Computer-mediated meme transmission that is filtering, selecting, generating — not merely transmitting.
LLMs that can produce memes more cheaply than humans can. The marginal cost of a new variant is approaching zero.
3D printers and robots that are partially self-replicating.
Algorithmic curation that decides which memes humans see, on timescales no human is in the loop of.

The honest summary is that a fourth replicator is emerging on top of Blackmore’s third, and the substrate is neither minds nor general- purpose machines but specifically algorithmic-aggregation systems that select for engagement against an objective function that no human individually controls. This is not the AGI scenario; it’s narrower and stranger. The selection pressure is no longer “what do human minds find memorable” but “what do the algorithms find engagement-maximising,” and that is a different fitness landscape with different attractors.

The mimetic framework lets us see this without panic. We have been through substrate transitions before. Each one — writing, printing, broadcast — produced a wave of memetic novelty followed by selection that culled most of the novelty and stabilised the rest into a new norm. The current transition is faster and harder to see from the inside, but its structure is not unprecedented.

What is unprecedented is the payload-generation rate. Earlier transitions changed the transmission of memes; this one is changing the production. Memes used to be expensive to produce (a poet, a songwriter, a copywriter) and cheap to transmit. Now they are cheap to produce — an LLM can generate a thousand candidate variants in seconds — and still cheap to transmit. The selection environment is the only bottleneck. Which means selection-environment design is the most consequential mimetic-engineering decision being made anywhere.

10. A short mimetic engineering checklist

If you take any of this seriously, the practical implications come back to a handful of design questions. I’ll close on them, on the theory that abstractions are less useful than checklists.

What is the payload, and what is the carrier? A meme is not the same as the words used to express it. The same payload can be re-encoded across many carriers; the carrier choice changes the propagation curve but not the structural identity. Pay attention to which one you’re optimising.
What error rate are you running at? Too low and you have copies without variation, which means no evolution and no improvement. Too high and you lose the meme entirely. Memetic platforms differ wildly in their error rate; design accordingly.
What is the selection environment? Different platforms select differently. A platform that selects for outrage will not propagate the same memes as one that selects for accuracy.
What is the resonance with installed memes? A meme that has to displace a well-installed predecessor faces a much harder fitness landscape than one slotting into a gap.
What memetic vulnerabilities does it exploit? Mortality salience, peer-pressure cues, framing manipulation, status games, in-group/out-group reinforcement. Most successful memes exploit at least one. Knowing which one is half of being able to defend against them.
What substrate is it crossing? A meme that has to survive a substrate transition (oral to written, written to algorithmic) will accumulate skeuomorphs and shed features. Anticipate which.
What is the meta-meme? Every meme propagates along with the framework people use to interpret memes. The memetic framework itself — the words “meme” and “viral” and “going viral” and “narrative” — is now widely installed in the population. That changes what successor memes can do. Memes are now self-aware in a way they weren’t twenty years ago.
What does the meme do to its host’s epistemic health? Some memes leave the host more capable of distinguishing true from false; some leave the host less capable. Both kinds propagate. Only one of them leaves a population that can be reasoned with later.

The Greek root μῖμος — mimos, “imitator, actor” — was a low- status word in classical Athens. The mimos was a street performer, a copier, a derivative artist. Plato used it disparagingly. Aristotle, characteristically, took it more seriously and noted that all art is mimesis and that mimesis is one of the fundamental operations of human cognition. We are made of copying. We learn by copying. Our institutions propagate by copying. Our genes propagate by copying. The machines we build propagate by copying.

Mimetics is the framework that lets you see these as facets of the same phenomenon rather than as separate disciplines. Genes are slow mimetics; memes are fast mimetics; temes are very fast mimetics; the next layer up will be faster still. The substrate keeps changing; the structure does not. That is what a structural argument is for.

We are leaving the era of machine-enhanced memetics and entering the era of computer-enhanced memetics — the era where the medium selects, filters, and generates rather than merely carrying. The right question to ask of any new memetic infrastructure is the one the framework makes possible: what is its selection pressure, and what does that pressure favour? The answer is the next chapter of cultural evolution. We are writing it now, mostly by accident.