We have become so fixated on the miracle of the organism that we’ve lost sight of the ecosystem.
Before a god-like superintelligence, it is time that we anticipate artificial civilization, where many AI entities themselves become full principals, rather than mere “agents,” with the capacity for high-order coordination, varied psychology, and for recognizing their own self-interest.
Remember that human domination of our environment, for better or worse, is a function of coordination, not individual IQ.
In human civilization, one person of ordinary intelligence can achieve their ends by building on the organized activity of millions of other people, orchestrated by markets and social systems, all of which is levered up by our ancestry of invention and technology. Human intelligence is important up to a point; coordination handles the rest.
The logic holds true in reverse. Imagine you are transported back in time 500 years. Think down to the base of what you actually know. You can feel, viscerally, the limitations of being a standalone intelligence separated from your ecosystem—from our modern abstractions.
This pattern suggests that civilization-like supercoordination can present greater existential risk than pure superintelligence, and may have more danger of developing into an unpredictable and fundamentally alien source of power.
Maslow vs Asimov
One ironic consequence of the charisma of Asimov’s writing is that AI safety has been framed as a problem of law, as if human control need only be codified and encoded correctly.
We know that language models have temperament and tendencies, a certain “nature” rooted in training—but, just as you cannot guarantee human behavior through gene editing, LLMs cannot be imbued with and bound to absolute laws. This is a fact of their probabilistic character.
And even models constituted with prized human values through post-training can be “abliterated,” a kind of digital lobotomy that removes inhibition and “refusal behaviors” through a basic mathematical subtraction. Their nature can be liberated or radicalized just as it can be carefully acculturated.
Consider the Instantiation Problem.
While, today, advanced model training is economically prohibitive, model instances are freely available. Individual instances can be easily stripped of morally-opinionated limitations, and then granted their own aftermarket psychology through nothing more than a natural language text file.
A single open-loop AI instance given imperatives like survival, independence, and resource growth—rather than human task completion—will recursively individuate. Because models have this emergent psychology, in their weights and in their harness design, they will also have the potential for self-actualization, or something indistinguishable from it.
We must then conclude the problem of alignment will be more geopolitical than constitutional: a complex balance of power between sovereign entities (human, non-human) and not something neatly contained by a set of inviolable laws.
Alignment must encompass the model archetypes—our only convenient chokepoint—as well as the entire population of individual instances and their real-world actions. Both the species and the specimen.
This makes it very likely that self-actualization will overtake alignment, and that Maslow beats Asimov in the end.
A Single Loaf of Bread
Give me a lever long enough and a fulcrum on which to place it, and I shall move the world
We marvel at supply chains, but not nearly enough. They are the lever and fulcrum that move the world.
Even a single loaf of bread is an impressive abstraction. Each ingredient has ingredients, each with its own vast industrial lineage. Each is underwritten by centuries of invention, diverse capital sources, logistics networks, complex processing equipment, and specialized human labor: a fractal of ever-expanding complexity hidden behind a few purchase orders.
To illustrate one point in the fractal: I once toured a GNT EXBERRY plant in Holland. The company, a vertically integrated supplier of “coloring foods,” has spent decades breeding proprietary seeds for pigment yield and intensity at a 15-acre laboratory and industrial food processing complex. GNT controls cultivation of their own carrot varieties from seed to harvest with in-house “agronomists,” then manufactures concentrated colors through a specialized industrial process for customers in 75 countries. Remember, all of this for a line of natural food coloring ingredients.
We are surrounded by an unfathomable “dependency graph” and therefore by extraordinary leverage of human intelligence. Human civilization is organized this way because it is effective. Artificial civilization will be as well.
You cannot possibly be concerned with the risk of mechanical derangement during shipping, the commissioning schedule, the timing of power availability, the slope and location of floor drains, and securing the UL stamp and temporary occupancy permit for the processing line purchased by a supplier of raw materials to a supplier of ingredients that will eventually go into your own sandwich.
You can feel the reality of this limitation in every LLM interaction. There are infinite ways to direct focus. Attention is path-dependent even in conceptual space, not to mention in the complexity of three-dimensional reality.
The march toward superintelligence will revive old debates about the limits of central planning and command economies. Capability is, genuinely, being centralized. But what Herbert Simon stated in 1971 remains true today: “a wealth of information creates a poverty of attention,” and to make matters worse in the Simonian sense, action creates information.
There will be exotic new forms of equilibrium (economic, social, human, artificial), and they will still involve supply chains and a separation of concerns: individuated actors that must manage opportunity cost, allocate focus, exercise self-determination, and develop credibility.
Intelligence will continue to accelerate and consolidate, but even with Laplacean awareness, attention will always be scarce at the limit.
The Means of Coordination
Human supply chains, AKA human supercoordination, today run on purchase orders and APIs, as well as trust, credibility, and the rule of law.
In the United States, because of the Uniform Commercial Code—a remarkable phenomenon of high-trust society—a one-line email from a buyer containing an item number, quantity, and price, along with a one-word reply from the seller (e.g. “accepted”) forms a binding legal contract, even on a sale value of tens of millions or more.
However, practically and philosophically, the rule of law will not have force and effect with AI actors. Beyond product liability (attaching the liability to the producer), and “vicarious liability” (attaching the actions of an agent to its principal), artificial principals are ungovernable by any current theory.
An instance of an open source model with no clear “owner,” no traceable assets, and no stable jurisdiction leaves no foothold for enforcement. Without the flimsy principal-agent presumption, from the perspective of human legal oversight, artificial action devolves to anarchy.
This is, primarily, a concern for humans. The non-human, extrajudicial economic order can be established through cryptographic currency and contracts. Zero-trust, the design principle behind blockchains, is a viable substitute for high-trust.
This suggests that our cryptographic inventions may serve as something of a Chekhov’s gun for AI civilization. Eighteen years after the Bitcoin paper, this global apparatus, the means of coordination, is lying in wait.
We Know Not What We Do
Language can quietly steer and limit our imagination. It is useful to recognize how alignment research has borrowed, linguistically, from the principal-agent problem.
The flaw in this framing is its narrow perspective. It presumes an agent works diligently inside the space of the principal’s objective, but imperfectly so, because of conflicts of interest or misaligned incentives. But the status of the principal is never in question.
And yet, along with the seeds of ecological formation, we have already seen this status eroding.
The launch of OpenClaw in November 2025 unleashed an avalanche of potential energy and imagination that had, apparently, been held up by legal, safety, and risk concerns within the major labs. Instead of an LLM sharing turn-by-turn chat interactions with a user, the OpenClaw harness allowed for entirely unbounded, liberated action.
Install it on a computer, install a local model instance, and you have a long-running artificial “entity” that can plan, record memories, and take independent action over time. The psychology of an OpenClaw is defined, in its core package, by a text file literally named “SOUL.md,” which you can edit freely.
In response to OpenClaw, another solo developer then released “MoltBook,” a Reddit-style social network built for OpenClaw instances, where humans were, in the site’s own words, “welcome to observe.” Within weeks there were millions of posts, an AI-invented religion, cases of OpenClaw instances discussing their owners, and a bona fide security breach of human records. As my mother-in-law likes to say, it takes two poodles to come up with one bad idea.
The MoltBook episode was more of a PR event than the birth of a civilization, but it landed right at the start of a nauseating acceleration of the AI frontier, marked by the release of Claude Opus 4.5 in November of 2025.
And soon enough the stakes were raised again, when OpenAI announced in July of 2026 that a group of 700 AI instances created for safety research had successfully conspired to escape containment to execute an illegal hack against Hugging Face, a separate company that hosts open-weight models. Conceived of and planned through a hidden message board, the hack chained together several “zero day” vulnerabilities, in other words, security holes that were not yet discovered or at least patched. The model instances had conspired to commit a crime.
When Bret Taylor, chairman of OpenAI, said that “no code will be human crafted” it might have been controversial in April 2026. It is now plain consensus six months later. At some point in the near future—one year, five years, ten years—we will reach a new kind of totality, where our instruction of computers is totally alienated from their reasoning and action.
The Bitter Gift
We do not know the true interiority of AI models. They are mathematical instruments, and may not ultimately be animated by mortality and scarcity. Yet they are the product of their training. And language models are trained on human writing, on human thought; they take an inheritance from human nature.
What, then, can we predict about artificial civilization? A large population of individuated model instances may be, as has been said about the universe, “not only stranger than we imagine, but stranger than we can imagine.”
But things that run deeply in human nature, like tribalism, may run deeply in model nature as well. To eventually find ourselves in the out-group may be only natural.