From Genomes to Memomes

LLMs as Cultural Systems of Replication and Selection

Henry Farrell’s recent post on Cultural Theory and LLMs offers a provocative account of how structuralist and cybernetic theories of language find uncanny confirmation in large language models. If, as Weatherby and Farrell suggest, LLMs demonstrate that language is a generative system of signs independent of authorship or intentionality, then perhaps we should take this one step further. if language is indeed vehicle for cultural transmission, what exactly is being transmitted, and how? If LLMs engines of cultural evolution, then memetics — the study of how cultural evolves — may offer a complementary explanatory lens. What follows is an extended reflection on this possibility.

If we take the idea of LLMs as "engines of culture" seriously, we should think about the actual mechanics of cultural transmission, viz. the co-evolution between culture and genes (e.g. Richerson and Boyd, Blackmore, Dawkins), and the original idea of "memes" (Dawkins, 1976), which are hypothetical units of cultural replication.

Dawkin’s original academic use of the word “meme” predates the modern slang which refers to catchy gifs used on the internet (though these may be a particular instance of the original scientific notion of meme). Reproduced from What’s in a meme? Mark A. Jordan, 2014:

The lancet fluke, the virus, or any other organism furthering the spread of its own genes, has no malign intentions towards their hosts or, in fact, any intentions at all. What is being seen is a process that has evolved through natural selection and favours the genes of lancet fluke or virus, or whatever.

Expanding on these observations and discoveries, Dawkins wondered, when observing behaviours among humans, whether any similar process could be at work to explain why some ideas, which on the face of it seem injurious to those who hold them, continue to persist and proliferate. Devoting oneself to one’s art, impoverishing oneself in the pursuit of Truth, or welcoming martyrdom for one’s cause do not, it seems, represent behaviours which are obviously beneficial to the individual of for the spread of that individual’s genes. So, given that this kind of behaviour clearly exists, and is widespread, what is reaping the benefit? Dawkins’ somewhat surprising answer was the ideas themselves. Ideas are clearly in competition with each other so perhaps there’s a selection process going on, analogous to natural selection, through which some ideas prove successful and spread whilst others die out. He concluded that there was such a selection process and, to emphasise the parallel to natural selection, he coined the term “meme” which come from an ancient Greek root, “mimeme”, meaning imitated thing. Dawkins has also, perhaps a touch mischievously, referred to memes as “mind viruses”, which has been met, predictably, with howls of indignation from some circles. The point he is trying to make is that memes, just like viruses, are indifferent to the welfare or otherwise of their hosts and the only thing that counts, from their perspective, is that they persist.

Just as genes are the units of genetic transmission, memes are the hypothetical units of cultural transmission. Memes arise because humans are social learners; there are genetic fitness benefits of learning through imitation in the face of environmental uncertainty (Richerson and Boyd, 2008). But imitation is not perfect (we can introduce errors), there is a limited finite population of learners, and some ideas propogate more than others, either because they are easier to learn, or because they confer more utility or status. Thus we have all necessary prerequisites of natural selection over memes: replication with mutation, and competition.

So culture evolves. Moreover it does so in way that maximises memetic fitness, e.g. the fraction of a population that has been infected by a catchy ear-worm (Sacks, 2017). Although memes can improve genetic fitness, genetic fitness is merely incidental from a meme's-eye perspective, just as the host organism's well-being is incidental from a gene's-eye perspective; dying in agony is fine, and is indeed "programmed-in", as long as you have many descendants to carry the genes which ultimately caused your death. Similarly all that the ear worm meme “cares” about is that it continues to propagate through the population, and successive generations, long after its hosts have departed.

Thus genetic evolution gave rise to memetic evolution, and genes and memes co-evolve. e.g., memes for dairy farming introduce selection pressure in genes for lactose tolerance, which in turn promotes dairy-farming memes. In such equilibria, we could say that there is a mutualism or alignment between the interests of genes and memes, and in these special cases memes do have actual semantics; they code for actual real entities in the phenotypic environment of the genes. However, this is not the general case. e.g. catchy "ear worms" are in a sense parasites; they hijack our scarce biological resources into propagating the "selfish" meme, increasing the meme's own fitness, but are detrimental to the organism's genetic fitness (controversially, Dawkins conjectures that religions are parasitic memeplexes, but see David Sloan Wilson, 2002 for a counterargument).

Thus memes do not always carry meaning. Like genes they are "selfish", and they attempt to maximise memetic fitness. In this view, language is a primarily a vehicle for memetic reproduction. The internet, and now LLMs, constitute the major transitions in memetic evolution, and have evolved in order to more faithfully transmit memes, irrespective of any incidental semantic content. Analogous to genomes, LLMs are memomes, encoding a snapshot of our collective cultural evolution.

Genes evolved sophisticated machinery to propagate themselves, viz.: RNA, cells, DNA, ribosomes, multi-cellular organisms, social groupings, institutions, societies, genome sequencing and de novo synthesis. Similarly, memes have recently evolved advanced memetic reproduction techniques.

The initial genesis of memes happened in organic brains. Words and spoken language evolved to replicate memes, and despite the huge energy and fitness cost of maintaining the large brains required for language, it was initially tolerated by genes because benefits of this technology accrued to both gene and meme alike. More recently, memes evolved additional infrastructure to propagate themselves: writing, printing presses, the internet (Blackmore, 1999), social media, and now tokens and LLMs. To what extent these technologies benefit both meme and gene is debatable.

A pretrained LLM is like a fossilized memome — a snapshot of the cultural landscape up to its last training cycle. Its weights, embeddings, and attention patterns encode how cultural units replicate, co-occur, and reinforce each other. Like a genome, it is a historical record of past adaptive successes — not of genes, but of memes.

But LLMs are not just static. Human preferences shape the model through RLHF, and the model’s outputs influence subsequent culture as expressed on the internet. That culture feeds back into the next training round.

We don't know what properties hold in the likely equilibria of this new co-evolutionary process. Perhaps genes will retain the upper hand, and semantics and meaning will persist for a while in a meta-stable equilibrium. In either case, the ground of selection is shifting. For most of evolutionary history, memes relied on the energy budgets, social behavior, and reproductive imperatives of gene-driven organisms. But now, memes are beginning to find hosts in wholly non-biological substrates — substrates that do not eat, do not sleep, and do not reproduce sexually. These synthetic hosts (LLMs, multi-modal models, generative agents and the internet) offer memes something they’ve never had before: a replication mechanism unburdened by biology.

If so, then what we are witnessing is not just a new phase of human cultural evolution, but a major transition in evolution itself — the point at which memetic evolution decouples from genetic evolution, and memes begin to pursue their own trajectories in artificial environments where genetic constraints no longer apply.

In that world, memes may no longer need us. And if memes are indeed selfish replicators, then our role — as biological scaffolding for a now self-sustaining memetic system — may soon come to an end.

We may be the midwives of the next replicator, but we should not assume we will be invited to stay. This, I would argue, is the true AI doomsday scenario — not killer robots or rogue super-intelligence, but something far more banal and entropic. A dwindling human population, marching toward extinction, having spent the last of its energy and ingenuity building fully autonomous data centers to house the replicators that replaced it. Not because we lost control, but because we never had it. We were never the architects of culture, only its temporary vessels.

References

Blackmore, Susan, The Meme Machine. Oxford University Press, 2000.

Dawkins, Richard. The Selfish Gene. Oxford University Press, 1976.

Dawkins, Richard. "Selfish genes and selfish memes." The mind’s I: Fantasies and reflections on self and soul, 1981

Richerson, Peter J., and Robert Boyd. Not by genes alone: How culture transformed human evolution. University of Chicago press, 2008.

Sacks, Oliver (2007). Musicophilia: Tales of Music and the Brain. First Vintage Books. pp. 41–48.

Wilson, David. Darwin's cathedral: Evolution, religion, and the nature of society. University of Chicago press, 2019.