Memetic optimization #1: brainrot

I.

I don’t use social media except for X, and even there I peruse only the Following tab. I have my YouTube feed blocked. I thought I had erected sufficient barriers against the simplistic mind-viruses of our age.

I was wrong.

It’s the human interaction that gets you. My PhD labmates told me about sounds I’ve somehow absorbed before ever hearing them firsthand. I heard the kindergarten teacher next door teach the kids “Ballerina Capuccina, Tralalero Tralala!”. The infection spread through walls and secondhand descriptions.

The epidemic seems to have faded now, and yet, the display of sheer memetic power is ... interesting. What exactly is happening here? What makes something a meme? More importantly, what can it make you do?

---

II.

The term “mind-viruses” is not just a metaphor; memes and pathogens both operate on hosts and reproduce by transmission from host to host. We can look at both as pieces of replicating information.

How much information is there in a meme?

The phrase “Tralalero Tralala”, in ASCII, is about 150 bits. The genome of the common cold virus contains 10,000 bits. The Holy Bible weighs in at 10 million bits and has convinced people to sacrifice their lives many times. Human DNA itself contains on the order of a billion bits.

Of course, in the above, we assume some priors: The DNA language already exists, the people reading the Bible can understand text; and brainrot assumes that the kids understand the concept of “la polizia”.

---

III.

Richard Dawkins coined “meme” by analogy to “gene” — both are replicators subject to evolutionary optimization. To persist, memes need to develop reproductive and defense mechanisms.

Reproduction develops first – viruses, for example, infect cells, and make the cell’s DNA help produce more copies of themselves. Memes on the internet are funny, and people like to share funny things with their friends. Religious texts tell people to spread the holy word.

Some memes develop more complex spread mechanisms, involving multiple hosts. Viruses by default interact with cells, but they can evolve affecting the organism as a whole in a coherent manner. The rabies virus makes rabid dogs eager to bite, which transmits the virus to other animals.

Toxoplasma makes rodents less afraid of cats, making them more likely to get eaten. Matthew 28:19-20 says

Therefore go and make disciples of all nations, baptizing them in the name of the Father and of the Son and of the Holy Spirit, and teaching them to obey everything I have commanded you. And surely I am with you always, to the very end of the age.

Defense mechanisms are something that a meme doesn’t strictly need in an exponential growth environment – viruses get destroyed by immune systems all the time, but as long as the growth rate R is larger than 1, it doesn’t matter. But in a zero sum environment where different memes compete for the same resources, it does matter. Beliefs that promise rewards and retaliate against apostates seem to be much more successful.¹

---

IV.

Although many memes are bad for the host (in the sense of taking resources and mindshare away from the original preferences and values of the host), some are beneficial. Herpesviruses in mice help immunity against bacteria.

Religions and political movements have often helped people by imposing positive cultural customs (e.g. religious norms against infanticide, Muslim and Mormon norms against alcohol), the sense of community, as well as promoting cooperative values (love thy neighbor) in general. I am pretty sure I have personally benefited from the “back pain is in your head” meme. The Ice Bucket challenge, although likely not beneficial for any one person that threw ice onto themselves, arguably helped humanity by redirecting resources into ALS research.

What can a meme make you do?

Viruses sometimes make the body kill itself by immune response; but they rarely make you kill another person, or make the group of 10^13 cells that is you take any coherent actions at all.

The total number of utterances of “tung tung tung” in conversations is perhaps in the billions; but there has been no coherent movement to make Italian brainrot institutional, or for any other goal at all. Even a few months is enough for the kids to get bored of the old meme and replace it with a new one. The meme is far too simple to make itself persistent, or to prevent other memes from outcompeting it.

---

V.

The Sorrows of Young Werther is a 1774 novel by Johann Wolfgang von Goethe about a guy, Werther, that’s in love with a girl engaged to another guy. Werther writes love letters, wallows in melancholy, reads Ossian, contemplates nature, grows increasingly despondent, and finally commits suicide.

This is the work that kickstarted both Goethe’s career and arguably Romanticism as a cultural movement that would come to dominate the thinking of European elites until about 1848. It was wildly popular in its time and inspired what we’d now call a fandom - large numbers of young men dressed like Werther, quoted Werther, and perhaps even thought like Werther.

It also reportedly led a bunch of people (young men suffering from romantic disappointment) to copycat suicide.² This in turn caused a moral panic and got the book banned in a few places across Europe.

I read Werther in high school; it is a good piece of writing, but it did not have similar effects on me, nor on anyone else I knew who read it, for that matter.

The meme meta has advanced significantly since the 18th century. Werther doesn’t stand a chance against the AI-generated slop.

---

VI.

Some bits of information surely have a lot of power over an individual, but it’s not clear it’s always in an easily steerable direction. The HIV virus would do best to influence its hosts to have more sex or donate blood. But it does not do it, because it’s not intelligently optimized to influence behavior in such a complex way.

The concept of “tralalero tralala” would do better for the goal of preserving itself if it could make people worship it, or build huge statues of a shark with blue sneakers!

Yet it doesn’t, because the message is way too simple, and was not optimized with persistence in mind.³

---

VII.

I believe complexity and intelligent optimization of the meme are both correlated with coherent behavior.

Take the Holy Bible. It is a full book with many narratives, written by intelligent people, and it delivers. The embodied belief in Christ has sent large fractions of all armed men of Europe to conquer the Holy Land, multiple times. There are cathedrals everywhere for those with the eyes to see.

Very successful memes (1) have a lot of information that enables carrying more complex emotions or proscribing more complex social structures; (2) have perhaps been designed with this in mind.

Evolution in the wild is a nice optimization mechanism because you don’t need anything smart to run it. You just let memes be, and whatever has the best evolutionary fitness will succeed. However, this only gets you so far, because experimentation is slow and you don’t have actual control over any part of the optimization. If you want your meme to (a) spread and (b) actually do something after spreading, evolution will only work on the former while ignoring the latter.

---

VIII.

To optimize before releasing the meme to a wider audience, you need data, ideally distilled into preference models. Only in the past couple of years have we managed to train AIs that understand human preferences enough to optimize anything without A/B testing. The only players optimizing memes properly now are companies with a stake in the attention game, and they are not chasing anything except relatively benign goals of more eyeballs, more screen time, more profit.

The answer to “What can a meme make you do?” depends a lot on how the meme came to be. A random piece of software that is evolutionary selected to be shared likely can’t make your computer do anything interesting. A piece of software that I write for you to run can do arbitrary code execution.

---

IX.

From their very inception, LLM chatbots have been (implicitly and explicitly) trained on human preference data. They have an implicit model of what people like. It is certainly possible they know about human preferences for memorable sound patterns.

So I wondered: can an LLM generate plausible brainrot from scratch? Does o3 understand the deep structure of what makes simple patterns catchy?

Here is what I got:

Not bad, huh? For a bit of post-hoc rationale of why this would be memetic, it says:

Phonetic symmetry (BR–BR / B–P / B–L–L);

Micro narrative – setup (BRR BRR), punchline (BOP!), celebration (BALALA). Minimal but complete arc.

---

X.

We have established that people’s behavior has been controlled by pieces of information in the past and present, to varying degrees. But in the current memetics meta, just saying the words is kind of not enough.

People do not instantly join al-Qaeda upon hearing about the group; it takes persistent social influence. Most people don’t start spreading “tung tung tung” out of the blue; it takes a memory connecting the meme to a positive social experience.

Nevertheless, it is plausible that the memetics meta advances from this stage. We do not have formal bounds on the influence of a piece of information on a person’s future behavior; maybe words can reach pharmacological levels of power over a person?

If it’s possible, future AIs (and people with AIs) will figure it out. We’re constantly getting better at optimization.

---

XI.

How to build defense mechanisms against this sort of thing? If you wanted to create, say, an anti-memetics division (sorry) in a place working on CBRN risks, what would it focus on?

The working mechanism of a meme requires (i) that it spreads; (ii) that it controls behavior of a person. Preventing spread amounts to censorship, and although theoretically feasible, it has large downside and is likely not a thing society will want to institute in practice.

Making people more robust to sudden memetic changes seems more tractable. In the ancestral environment, I could imagine propensity for adopting new memes could be useful, because those memes often come with evolutionary advantages. But discriminating is also useful, as bad memes could destroy the tribe. The human ability to discriminate useful from useless beliefs is not perfect, but it is something that both individuals and societies could tune to be better.

There is also evidence that memetic immunity evolves by itself; along with the Werther example, odd cults seem to be rarer nowadays than when communication technologies first reached the point where those could spread.

---

XII.

Or, maybe we should not think in terms of “defense”. Most value is ultimately created from building good things, not from preventing bad things from existing. And a natural way to fight valueless memetics is to create and spread good memetic viruses.

The fact that the memetic competition landscape is somewhat zero-sum is an advantage for the defender here. If we can create robust pro-civilization memes, they will repel bad and useless ones.

Richard Ngo wrote about how a way to structure society should be robust to change to be good; I now understand that notion a bit better. Goodness is ultimately about outcomes, and if society ruptures due to adversarial pressure, the set of ideas maintaining the old order was not the best one we could pick.

Thus if there is one actual lesson in this post, let it be this: strong beliefs that act towards a positive future are valuable; both for individuals and for communities. Those who want to create value should consider making more of those to defend our memeplex against adversarial memes.

Thanks to Edoardo Debenedetti for reading drafts of this.

1. Here’s what actually matters for memetic fitness:

True and useful ideas certainly have a selective advantage insofar as humans care about usefulness, but there can be other features of an idea that convey a selective advantage in memetic competition: for example, an appeal to (alleged) consequences of accepting the idea. This is the reason so many religions prominently feature promises and threats of divine reward or punishment: “Believe X and you’ll be rewarded; believe not-X and you’ll be sorry” is more memetically fit than “It happens to be the case that X, but this has no particular further implications,“ because the former proposition creates incentives for propagating itself.

2. Note that the epidemic of Werther copycat suicides are disputed.

3. Though, the brainrot memes did manage to actualize themselves in the physical world, with a bit of delay: you can buy Shenzhen-made “Tung Tung Sahur Italian Brainrot Stick Creative Toys Collection Ornaments”, and brainrot plushies that, for some reason, all look kind of sad?