Anti-Meme Explosion

Most of what I write falls into one of two categories: opinions or technical blogs. This is neither. It’s more of a brain-dump, a collection of barely structured thoughts I’ve been wrestling with over the past month, ever since I stumbled onto the concept of antimemetics.

Popularized by the sci-fi book There Is No Antimemetics Division, an anti-meme is exactly what it sounds like: the opposite of a meme. While memes inherently want to jump from mind to mind and spread, an anti-meme naturally resists transmission. The term is interesting precisely because it is a little too broad; it’s almost a negative space around memes, where ideas die, resist, or become useless in the act of spreading.

What hooked me wasn’t the idea itself; I’ve always had an intuitive sense of things like this existing. What surprised me is that anti-memes are one of those concepts where, once you have a word for it, you are exponentially more likely to start seeing them everywhere and connecting ideas to them.

That’s usually a sign of a great concept.

Ever since this idea was added to my embedding table, it has been a sporadic yet constant background process in my mind. Maybe it’s just my neurodivergent brain making patterns, but you can fit a lot of things into this broad concept, things with very different characteristics. I’ve even gone so far as to try to categorize them.

There seem to be at least four distinct kinds:

• Type 1: Ideas suppressed by their environment.
• Type 2: Ideas that are inherently difficult or impossible to transmit.
• Type 3: Ideas that are actively erased by their own use.
• Type 4: Ideas that erase themselves, and take other ideas down with them.

These categories are not meant to be mutually exclusive or exhaustive; they are closer to “things one might reasonably call anti-memes”. Some overlap, others probably don’t fit anywhere, and a few may just be edge cases. That fuzziness is part of what makes the concept interesting.

Type 1 Anti-Meme: Ideas suppressed by the environment

The simplest anti-meme is an idea that would spread normally, except that something in the environment prevents it.

The obvious examples are politically incorrect ideas, or beliefs that clash with cultural norms. But the much more interesting Type 1 anti-memes are those unintentionally suppressed by external factors.

Think of typing an obscure word into software with an overly aggressive autocorrect. The word itself isn’t hard to understand, but the system keeps overriding it with the closest match in its dictionary. There’s no explicit rule banning the obscure word; the autocorrect is simply optimizing for common words. Take this train of thought a few stops further, and you arrive at LLMs.

LLMs might be the greatest anti-meme generators we’ve ever built.

As many people have said before me, LLMs are trained on a large distribution of text. By default, an LLM’s “idea distribution” is bounded by the internet’s own distribution, or by whatever distribution optimizes the particular RL environment they went through in their post-training phases. That’s why outlier ideas, or ideas simply outside the distribution, are almost impossible for an LLM to naturally generate. Sure, it can mix concepts into new things, but the mixing itself has a distribution.

Imagine a writer who filters all their drafts, and maybe even their raw thoughts or decision-making, through an LLM. What would happen to those outlier ideas? They often start out ugly and malformed, but eventually turn into genuinely good ideas. From an LLM’s perspective, those raw ideas just look like errors.

In the long run, LLMs smooth out the distribution of user ideas. Your text might have much better prose, but the strange and unique ideas get unintentionally and slowly eroded.

We’ve introduced a massive external factor to human ideas, one that is quietly generating an uncountable number of anti-memes.

Type 2 Anti-Meme: Ideas that are inherently difficult or impossible to transmit

This type of anti-meme is knowledge that is not being actively suppressed; it is just very hard to move from one mind to another.

A lot has been said about this type of knowledge over the years, sometimes under different names. One of the best examples is tacit knowledge, attributed to Michael Polanyi’s Personal Knowledge (1958), who summarizes it well: "we can know more than we can tell."

If you have ever worked in a big organization, you’ve certainly experienced this firsthand. There is a type of knowledge embedded within the topology of relationships between people in the organization. It does not exist in a single project, person, or document; it lives in the shared history, the favors, the trust, and even the grudges. Often, it’s impossible to know this knowledge even exists until a seemingly invisible problem blows up after a round of layoffs or a corporate restructuring that modifies the underlying human mesh.

This type of anti-meme knowledge is one of the reasons I think AI replacing humans in the workplace will take much longer, and be much more painful, than most CEOs want to admit. I’m not saying the transition won’t happen, but I wouldn’t be surprised if major corporations completely implode due to the sudden, unrecoverable loss of their tacit knowledge.

Type 3 Anti-Meme: Ideas destroyed by being used

This one is tricky because it’s where things start sounding like magic, or completely absurd. To be fair, thinking of concrete examples is not easy for me.

Still, there are things that might fit in this category. Suppose you discover a mispriced asset in the stock market: you hold a valuable piece of information. You can keep it a secret and do nothing, but to actually make money, you have to act on it. The paradox is that the very moment you execute a trade based on that secret, you start killing the information. As others notice the market shift, the mispricing is rapidly eliminated.

That’s the logic of the Efficient Market Hypothesis. Markets are machines that eat secrets. When someone makes use of an information asymmetry, it soon stops being an asymmetry.

There are likely other examples of this outside of finance, and speaking from experience, throwing this concept out there makes for a fantastic conversation starter.

Type 4 Anti-Meme: Ideas that erase themselves (and other ideas with them)

This is where things get completely abstract. Pointing out a specific Type 4 anti-meme is nearly impossible for me, either because they only exist in sci-fi, or because they are genuinely just that good at erasing themselves.

One thought I had is that, on a societal level, the technical knowledge required to build AGI could be a sort of Type 4 anti-meme.

The moment we achieve self-improving AI, it would be hard, if not impossible, for human AI researchers to keep understanding how all the parts of the self-evolving system fit together. Suddenly, the incentive for humans to know how to build AI evaporates. Why bother? Whatever approximations humans attempt, the AGI would do better and faster.

The same logic would apply to many other disciplines this hypothetical super AGI could work on. A couple of generations later, humans would have no incentive whatsoever to possess or build technical knowledge.

The technical knowledge that initially created the AI would end up erasing itself from humanity, surviving only in computers, and taking a lot of other knowledge with it.

It’s not as if things like this haven’t happened before; just imagine how much practical and technical knowledge about things like vacuum tubes will vanish forever once the last generation of engineers who actually needed them passes away.

In that sense, the knowledge of how to build AI is the ultimate anti-meme.