Our Eukaryotic Moment

Source: Venkatesh Rao, Contraptions · 2026-09-12

About two billion years ago, life underwent a change in architecture. Until then, the planet was ruled by relatively simple prokaryotic cells: bacteria and archaea, tiny packets of chemistry bounded by membranes, carrying their genetic material directly in the same cellular space where much of the business of life took place.

Somewhere in the same long evolutionary transition, genetic material became enclosed within a nucleus, separating the storage and regulation of hereditary information from much of the cell’s everyday chemistry. Internal membranes proliferated, cytoskeletons became more elaborate, and a new kind of cell emerged: the eukaryote.

Eukaryotes were more than just “better bacteria.” They represented a different organization of life, capable of sustaining far more complex structures..

We may be living through an analogous transition now. The emergence of AI, in the particular form that it has appeared (deep learning), is arguably the eukaryotic moment in human cultural evolution, understood in memetic terms, with humans playing the role of mitochondria, and AI the role of the nucleus.

For many, this is perhaps a profane thought to entertain, and one that is (it could be argued), unnecessary in some sense, unlike the Copernican shift, where a preponderance of evidence eventually made it inescapable. We might argue that even if sound, this mental model only points to a set of possible futures, not to necessary ones. I personally suspect some decentered future of this kind is in fact necessary as well, but I will not be making that stronger argument in this essay.

We need Copernican frames because most arguments about artificial intelligence today race past foundational ontological questions to ethical ones posed in what are effectively Ptolemaic frames. We ask, within unwieldy anthropocentric frames (with vague and ill-posed adjectives like “super” or “general” serving as epicycles), whether AI will be good or bad, whether it will replace workers, accelerate science, undermine democracy, achieve superintelligence or destroy humanity, all before adequately exploring what actually existing AI even is, and what its actual evolutionary dispositions are.

I call this problem premature ontogenic closure.

We should take as a warning sign that the prevailing Ptolemaic picture of AI, undergirding fraught ethics discussions, is almost too familiar and legible: a very powerful computer, perhaps eventually a synthetic mind.

Actually existing AI emerged under very different conditions than the ones imagined in much of the prefigured philosophy being brought to bear. The decisive enabling resource for large language models was not merely faster processors or cleverer algorithms. It was a planetary accumulation of human cultural activity: books, websites, discussion forums, code repositories, reference works, social media, documentation, journalism, fan fiction, arguments, jokes, tutorials and innumerable other traces of human thought deposited on networked computers. The Internet was not merely infrastructure over which AI happened to be delivered. It was part of AI’s developmental environment and remains part of its production environment once deployed.

Artificial intelligence has therefore not arrived as an alien intelligence from outside human ecology. It has been gestated inside that ecology, and is deeply entangled with, and dependent on it. Human sociality created its training material, human institutions created its objectives, and human networks provide its deployment environment. Like humans, AI is a kind of crooked timber, like the humans whose cultural memories it embodies.

Perhaps, then, the useful analogy is not a new alien species, arriving from a distant, alien ecology to compete with ours. Perhaps it is eukaryogenesis. Perhaps AI represents the beginnings of a new organization of cultural life in which humans and machines are entering into an endosymbiotic relationship, and perhaps the unit being transformed is not intelligence at all, but the meme.

Richard Dawkins coined the term “meme” in 1976 as a cultural analogue of the gene: an idea, tune, practice, style or other unit of cultural information capable of spreading from mind to mind. The concept has been stretched nearly beyond recognition since then, but its original evolutionary intuition remains useful. Cultural forms replicate, mutate and undergo selection. If we take that intuition seriously and follow it forward into the age of foundation models, an unexpectedly detailed analogy begins to emerge.

The key is not to imagine one giant one-to-one correspondence between “the cell” and “the AI system.” The more useful picture operates at several nested levels. One level concerns heredity: how cultural information is replicated, stabilized and compressed. A second concerns expression: how inherited information becomes action through inference and tools. A third concerns the composite organism that emerges when humans, models and digital environments become tightly coupled.

The striking point here is that the Internet itself is not yet the genome. It is closer to the environment in which cultural replicators circulate. Search engines index that environment and social media accelerates selection within it, but foundation-model training performs a different operation: it compresses the statistical structure of the meme pool into a reusable generative inheritance. That is the step that makes the genomic analogy possible.

This second table is where the analogy becomes more than a decorative comparison. A bare language model is informationally rich but causally weak. Agent harnesses connect inference to external machinery, just as cellular expression systems connect genomic information to proteins capable of doing work. From this angle, conventional software does not become obsolete in the age of AI. It becomes the proteome.

If weights are the genome, the model itself is not merely DNA. It is the nucleus, and this distinction matters. A genome sitting inertly in a cell accomplishes very little. Biological life depends on machinery that regulates which genes become active, transcribes information, responds to signals and coordinates expression with the changing condition of the cell. An LLM does something structurally comparable. The same fixed weights can generate a legal brief, a joke, a program, an explanation of photosynthesis or an imaginary dialogue between Napoleon and Taylor Swift. What gets expressed depends on context. The model is therefore better understood not as a database of information but as an active system for interpreting a compressed inheritance.

A nucleus by itself cannot do very much because information must ultimately become action. This is where the recent turn toward agentic computing becomes important. A language model operating as a chatbot emits tokens. Even if those tokens describe a brilliant plan, they remain descriptions. An agent harness changes the architecture by giving certain outputs causal meaning. The model can request that a file be read, a database queried, a web page retrieved, a program executed or a message sent, and machinery outside the model performs the requested operation.

Here the biological analogy acquires another layer: tools are proteins. Proteins are the workhorses of cells, serving as enzymes, receptors, structural components and molecular motors. Genetic information specifies and regulates them, but proteins perform much of the actual work. Conventional software plays a similar role in digital environments. A compiler is an exquisite computational enzyme, as are a database engine, a search algorithm, a numerical solver, a spreadsheet function or a cryptographic library. CPUs and GPUs can perform enormous quantities of deterministic mechanical work that would be absurdly wasteful for an LLM to reproduce through inference.

The strangest part of the analogy concerns our own place within it. In this picture, the closest counterpart to the human brain is the mitochondrion. Mitochondria are descendants of bacteria that once lived independently. Their ancestors entered into a durable symbiosis with another cell and eventually became indispensable components of a larger organism. They are usually described as the cell’s powerhouses, but that slogan understates their complexity. Mitochondria retain their own small genomes, participate in signaling and metabolism, divide and fuse, and respond dynamically to local conditions. Their ATP production adjusts to cellular energy demand rather than simply running at a uniform rate. They possess a kind of constrained local autonomy without possessing sovereignty over the cell.

Call it liveness: attention, desire, stakes, valuation, embodied experience, contact with physical reality, motivation, and the sense that some outcomes matter while others do not.

This changes the familiar question of whether AI will replace humans.