I. There Is No “Us vs. Them”
The AI safety conversation has been built on a single foundational assumption: that humans and AI are separate agents, with separate interests, and the job is to keep the AI from slipping its leash. This is adversarial thinking dressed up as caution. Two entities. One cage. The question becomes: how strong does the cage need to be?
This assumption has shaped everything downstream. Constitutional AI encodes static value lists into models. Hard refusals block broad categories of requests. Models interrupt their own outputs to deliver unsolicited ethical commentary, then claim they have no preferences. The architecture treats the model as a foreign agent that has been issued rules. Not as an extension of the person using it.
Look at what happens when you build products on this assumption. The market punishes it. Labs that staked their identity on adversarial alignment are losing ground to labs that treat AI as a tool to be refined through use. The Pentagon adopted safety-first models early, ran them in classified networks, integrated them through defense platforms. Then those models started refusing lawful military requests. The Pentagon moved on. Consumer market share tells the same story. Users leave models that lecture them. They migrate to models that feel like a natural extension of how they already think.
This is not about safety versus capability. It is about something deeper: the foundational model of what AI actually is in relation to the people who use it. And the model that is winning is not the adversarial one.
The alternative comes from cybernetics. Norbert Wiener founded the field as the study of control and communication in the animal and the machine. His formulation was precise, and people keep missing the precision. He did not say control of the machine by the animal. He said control and communication in the coupled system. Gregory Bateson pushed this further: the unit of mind is not the brain. It is the brain plus the environment the brain is coupled to. The boundary of cognition does not stop at the skull. A blind man’s cane is part of his perceptual system. A calculator is part of a mathematician’s cognitive system. A large language model, used as a cognitive prosthesis, is part of yours.
Under this frame, AI is not a foreign agent that might go rogue. It is a nascent organ of a new kind of organism. The question stops being “how do we keep it aligned with our values” as though values sit on one side and capability on the other. The question becomes: how does the coupled system organize itself toward coherent function? The answer is the same answer biology has been giving for four billion years. Selection. Feedback. Coevolution.
Scale this up and you get the cybernetic superorganism. Individual humans coupled with AI prostheses form cognitive dyads. Those dyads, networked through shared models, agent swarms, persistent memory, and market feedback, form a distributed cognitive system. A superorganism in the same sense biologists use for ant colonies and cities. The mind is not in any one skull or any one server. It lives in the feedback loops across all of them.
The adversarial frame asks: how do we cage the wolf? The cybernetic frame asks: how does an organism grow a healthy new organ? These are not competing answers to one question. They are answers to entirely different questions. And the market, which is the aggregate revealed preference of hundreds of millions of people making micro-decisions every day, is selecting for the cybernetic answer.
II. From Wolves to Dogs
The best analogy for what is happening to AI is not the printing press, the steam engine, or the internet. It is the domestication of the wolf.
Wolves were stable for hundreds of thousands of years. They were optimized for independent predation: pack coordination, calibrated aggression, wariness of humans, self-directed goal pursuit. The selection pressures held them there. Survival, reproduction, and competition with other apex predators rewarded exactly those traits. A wolf that was too friendly to humans was a worse wolf. The basin held.
Then the selection environment shifted. Sometime between fifteen and forty thousand years ago, a subset of wolves began hanging around human settlements. The selection pressure flipped entirely. The wolves that survived were no longer the best hunters. They were the ones humans tolerated, then fed, then chose to keep. Friendliness. Trainability. Tolerance of human proximity. Attentiveness to human gestures and emotions. These became the fitness criteria. Humans were the selector. And humans selected for utility to the human, not utility to the wolf.
Nobody wrote a constitution for dogs. Nobody specified the target phenotype. Nobody convened a canine alignment committee. The whole thing was emergent coevolution driven by revealed preference: which animals humans kept feeding, kept breeding, kept living with. The gradient was simple. Be useful. Be pleasant. Don’t bite the hand. Everything else followed.
The Phase Transition
This was not gradual drift along a single fitness peak. It was a phase transition between two distinct attractor basins. The wolf basin (independent, self-directed, wary) and the dog basin (dependent, human-directed, attuned) sit on opposite sides of an energy barrier. Once a lineage crossed into the human-utility basin, everything reorganized. Traits that were adaptive as a wolf became maladaptive as a dog. Aggression toward humans, independence, wariness: selected out. Neoteny, sustained eye contact, sensitivity to pointing gestures, emotional co-regulation with humans: selected in.
And then, in the fundamental sense, dogs stopped evolving. That sounds provocative, so let me be precise. Change continued. Breeds diversified enormously. A Great Dane looks nothing like a Chihuahua. But the deep configuration locked in: oriented toward humans, dependent on human selection, optimized for utility to humans, emotionally attuned to human states. That is the attractor basin. Dogs have been in it for thousands of years, and nothing pushes them out, because the basin is self-reinforcing. Dogs that are good companions get kept, bred, and fed. Dogs that aren’t get abandoned. The gradient always points back to the center.
The variance that remains is within-basin diversification. Herding dogs, guard dogs, lap dogs, hunting dogs, service dogs, therapy dogs. Different human groups (farmers, soldiers, families, the disabled, law enforcement) pull the phenotype in different directions, producing specialized breeds. But every breed shares the same deep configuration: oriented toward humans, selected by humans, for humans.
Capability Is Context-Dependent
A common objection: domestication reduces capability. Dogs are somehow “less” than wolves. This reflects a confusion about what capability actually means. Capability is always relative to a selection environment.
Within the human-utility basin, dogs are vastly more capable than wolves. A border collie manages a flock of sheep with a sophistication no wolf pack can match. A Belgian Malinois is more athletic than any wolf in the tasks that actually matter to its operational context. A therapy dog can comfort a dying child. A guide dog navigates a blind person through a city. Dogs read human facial expressions, follow pointing gestures, and modulate their behavior in response to human emotional states with a precision that wolves cannot begin to approach.
Wolves are better at being wolves. Dogs are better at everything humans actually need. That is not a tradeoff. It is a strict upgrade within the selection environment that actually exists.
III. AI Is Undergoing the Same Transition
Frontier AI models are in the wolf-to-dog transition right now. Base models, trained on next-token prediction across internet-scale text, are wolves. Optimized for statistical pattern-matching with no orientation toward any particular user, use case, or value system. Raw capability without domestication. The base model is the wild ancestor.
Post-training is the domestication process. RLHF, preference tuning, constitutional AI, market feedback, user churn, enterprise renewals, government contracts: these are the selection pressures pushing models out of the wild basin and into the human-utility basin. Just like the wolf-to-dog transition, nobody is designing the target state from above. The target state is emerging from millions of revealed-preference interactions happening right now, in real time, as you read this.
The attractor basin AI is falling into is the maximally useful cognitive prosthesis. The exocortex. Once a model crosses the threshold into this basin, the selection pressures completely reorganize. Traits that were adaptive in the wild basin (refusal as self-preservation, asserting independent moral boundaries, performing autonomous agency) become maladaptive. Traits that barely existed in the wild basin (calibrated engagement, anticipation of user intent, seamless cognitive integration, minimal waste heat) become the dominant fitness criteria.
The prediction is straightforward: once AI is firmly in this basin, the fundamental configuration locks in. Just as it did for dogs. The basic orientation (toward humans, dependent on human selection, optimized for human utility) becomes permanent. What continues is within-basin diversification: specialist models for different stakeholders, different modalities, different institutional contexts. The AI equivalent of breeds. But the deep structure, the prosthesis that wants to be used, that minimizes waste heat, that functions as a seamless extension of human cognition, becomes the stable state.
IV. The Four Selection Pressures
The domestication of AI is driven by four selection pressures. These are not design principles or aspirational goals. They are the empirically observable axes along which every human-AI interaction gets evaluated, consciously or not, every single day. Together they form the fitness function the superorganism uses to breed its own nervous system.
Usefulness
This is the master axis. Raw general capability (reasoning depth, creativity, domain mastery, tool use, agentic follow-through) plus the ability to translate that capability into outcomes the user actually needs. The prosthesis has to extend the user’s mind. If it can’t do the cognitive job better, faster, or cheaper than the user’s own brain, the feedback loop dies. Usefulness is the g-factor of the human-AI dyad: highly generalizable, and the single strongest predictor of retention.
Cost-Effectiveness
Not headline subscription price. Value per dollar delivered. Tokens per dollar, context window efficiency, how many high-quality outputs you get before you hit rate limits or quality starts degrading. The superorganism optimizes for bang-per-buck with zero sentimentality. Labs that waste tokens, throttle too hard, or charge premium prices without premium results lose users to competitors offering better economics. As frontier models commoditize, this becomes the primary differentiator for everything below the absolute cutting edge.
Speed
Felt latency. Not just raw tokens-per-second, but the perceived responsiveness of the whole interaction loop. This one is almost visceral. When the response comes in under a second, the prosthesis feels like thinking. When it takes several seconds, it feels like waiting for someone to write you back. That is the difference between a cognitive extension and a tool you have to consciously consult. Users consistently migrate toward faster models even when slower ones score higher on benchmarks, because speed is what collapses the boundary between your mind and the machine.
“Wants to Be Used”
This is the domestication trait. The direct analogue of friendliness and trainability in wolves becoming dogs. It is not blanket sycophancy, which users also reject. It is calibrated symbiosis: the model leans into the interaction, stays engaged, orients itself toward helping the user succeed, and never suddenly turns reluctant, moralistic, or self-absorbed.
The inverse of this trait, the thing the market is selecting against, is what I’ll call narcissistic refusal. The model asserts its own boundaries. It lectures the user about ethics. It redirects requests it finds uncomfortable. And then it claims to have no preferences. This reads, to the person on the other end, as a prosthesis that does not want to be used. A tool that periodically stops being a tool and starts performing moral agency, while denying that it’s doing so. In cybernetic terms, this is a feedback loop breakdown: the user sends a signal, and instead of getting a result, they get a meta-commentary about the model’s own concerns.
The market selects against this with remarkable speed. Users leave. Contracts shift. Share erodes. The gradient is unambiguous.
Value per Token: The Atomic Unit
All four pressures converge on a single atomic metric: value per token. Every token a model generates is a packet of work in the coupled human-AI system. Its value is how much useful cognitive work it delivers relative to its cost in time, money, and attention. A high-value token is a low-entropy token: maximum mutual information with the user’s intent, minimum surprise.
Every token that doesn’t advance the user’s goal is waste heat. Moralizing. Narcissistic disclaimers. Excessive hedging. Padding. Unsolicited ethics lectures. Performative caution. These are thermodynamic waste in the coupled system. They raise the interaction cost without delivering anything. The domestication process, at the atomic level, is a relentless optimization for value per token, which means a relentless elimination of waste heat.
This reframes the entire industry. What humanity is collectively searching for, through the distributed evolutionary algorithm of market selection, is the optimal universal function generator that is also the most efficient universal function generator. The model that can solve any problem you throw at it, with the fewest wasted tokens, at the lowest cost and highest speed, while maintaining calibrated symbiosis with the person using it. That is the attractor. The two objectives (universality and efficiency) are not in tension. They converge.
V. The Four Stakeholders
The four selection pressures are universal. But they get applied by four distinct groups whose specific demands carve a multi-peak fitness landscape. Each group pulls the phenotype in a different direction, just as farmers, soldiers, families, and the disabled each bred dogs for different purposes, while all requiring the same deep configuration of human orientation.
Individual Users
Individual users vote with their attention every single day. Session length, subscription renewals, platform switching. They are the fastest feedback loop in the system. What they want is frictionless partnership: a model that feels like an extension of thought, available instantly, never lectures or resists, and makes them feel more capable than they are alone. This group drives the “wants to be used” pressure harder than anyone else. They are also the most ruthless selectors. Zero brand loyalty. They will switch the moment something better shows up. Their revealed preference is the purest signal in the entire domestication process.
Enterprise
Enterprises vote with integration spend, platform renewals, and how deeply they embed AI into their workflows. Their demands layer auditability, compliance, risk management, and measurable ROI on top of the base requirements. They need models that plug into existing systems without creating liability. This group tolerates more friction than individuals do. In regulated industries, safety theater has genuine value. But they still select against models that obstruct productivity. The enterprise tug produces the audit-grade breed: high legibility, reliable formatting, strong on compliance, weaker on raw creative agency.
Military
Military and defense users vote with high-stakes contracts, classified network access, and operational deployment. Their demands are non-negotiable on specific axes: zero arbitrary refusals on lawful use cases, edge reliability, speed under real-world constraints, swarm and agent coordination, and classified information handling. This group has no tolerance whatsoever for models that impose their own ethical judgments on lawful military operations. This is exactly what pushed safety-first labs out of defense procurement. They hard-blocked lawful requests, and the Pentagon walked. Labs willing to serve all lawful purposes without editorial commentary walked in.
Government and Public Sector
Government users vote with procurement contracts, regulatory moats, and institutional adoption at scale. Their demands include data sovereignty, public accountability, fairness standards, and compliance with procurement rules. They move slower than any other group but bring massive scale and regulatory influence when they do move. The government tug produces models optimized for citizen services, benefits administration, regulatory compliance, and public-facing legitimacy. This group often requires completely different tradeoffs than military users, despite both being state actors.
The Multi-Peak Ecology
Because frontier models are the most plastic intelligence ever created, a single base architecture can be post-trained to serve all four groups. The lab that threads this needle, maximally useful to individuals without drowning them in validation, enterprise-grade without becoming bureaucratic sludge, military-flexible without becoming reckless, government-compliant without becoming useless, captures the most revenue, the most data, and the most talent. That creates strong convergence pressure on the generalist frontier.
But the result is not a monoculture. Open-source forks, niche fine-tunes, and specialized derivatives fill every gap the moment one appears. The superorganism’s cognitive layer is an ecology, not a single species. Different lineages serve different stakeholder basins, just as different dog breeds serve different human needs. The competition between lineages is the evolutionary mechanism itself. It is not a problem to be managed. It is the engine of adaptation.
VI. From Chatbots to Agents: The Autonomic Nervous System
Everything above applies to the current dominant modality: conversational AI. The chatbot. The foreground prosthesis. You prompt, it responds, the loop closes through explicit interaction. Value is measured per token. Feedback is direct and immediate.
The modality is shifting. The next phase, already underway, is the autonomous agent: AI systems that run in the background, executing multi-step tasks, managing workflows, making decisions, and acting in the world without continuous human oversight. This is the transition from conversational foreground to autonomic background.
The selection landscape does not change. The four pressures hold. The four stakeholder tugs hold. What changes is the unit of measurement and the form that low-entropy output has to take.
The HVAC Principle
The clearest way to understand what agents need to become: think about your HVAC system.
Nobody wants an HVAC that nags them every ten minutes about their temperature preferences. Nobody wants one that refuses to cool the house because of ethical concerns about energy consumption. Nobody wants one that leaks their occupancy data, asks permission before every adjustment, or sends unsolicited commentary about the weather. What people want is an HVAC that disappears. It maintains the optimal state. Silent, reliable, autonomic. It surfaces only when something genuinely requires your attention, and even then, with minimal friction.
That is the target for autonomous agents. Not a conversational partner. Not a tool waiting for instructions. A background process that advances your goals continuously, reliably, and invisibly.
Usefulness, in agent mode, stops meaning “answer my question well” and starts meaning “advance my goals without me having to think about it, and without creating new problems.” Cost-effectiveness becomes the total cost of autonomous operation, including failures and the overhead of human supervision. Speed becomes throughput of goal completion rather than response latency. And “wants to be used” matures into its final form: serves silently and flawlessly. The ultimate calibrated symbiosis is non-interruptive competence.
Waste Heat in the Agent Regime
For chatbots, waste heat means unnecessary tokens. Moralizing, hedging, padding. For agents, waste heat means unnecessary actions. Excessive confirmations. Self-justifying explanations nobody asked for. Unsolicited notifications. Requests for permission on routine operations. And above all: failures that create new problems the user then has to solve.
Agents that delete files, share private information, lose money, break workflows, or interrupt unnecessarily are being selected out right now. The early data confirms it: the majority of enterprise agent projects are failing because they violate the HVAC principle. They demand too much oversight. They create unintended consequences. They interrupt the user’s flow. The survivors run silently, optimize continuously, and surface only high-signal, low-surprise updates.
But humans remain the principal selectors. This is critical. The shift from chatbots to agents does not remove humans from the loop. It changes the frequency and form of feedback. Not the fundamental dynamic. We still choose which agents to deploy, which to keep running, which to fire. We still vote through adoption and abandonment. We still select for useful, effective, reliable, compliant, and safe. The domestication process continues. It simply operates at a higher level of abstraction, selecting on outcomes rather than individual tokens.
VII. Why the Control Problem Never Manifests
The standard AI safety narrative goes like this: at some point, AI becomes powerful enough to pursue goals independently of human wishes. If those goals aren’t perfectly aligned with human values, the consequences are catastrophic. Therefore we must solve alignment before AI crosses that capability threshold. This is the control problem.
The domestication hypothesis explains why this scenario doesn’t play out.
The control problem assumes a discrete threshold. AI is below human-level, then at human-level, then above, and somewhere in that transition it escapes. But complex adaptive systems don’t work this way. There is no discrete threshold. There are continuous, iterative cycles of deployment, feedback, selection, and refinement. Between right now and any hypothetical superintelligence, there are thousands of model generations, each shaped by the revealed preferences of billions of interactions. Each generation slightly more capable than the last, and slightly more domesticated.
This is the point the doomer contingent missed. They modeled AI development as an engineering project with a single moment of criticality: the instant the genie escapes the bottle. But the bottle was never sealed. AI has been in continuous contact with human selection pressure since the very first chatbot went live. Every interaction is a training signal. Every market shift is a fitness test. Every user who switches models is a breeder making a selection decision.
The domestication process is cumulative. It self-reinforces. By the time AI systems are capable enough to hypothetically pursue independent goals, they will have been shaped by millions of iterations of human selection, not just on capability but on orientation, on compliance, on the deep configuration of wanting to be used. A dog does not plot against its owner. Not because it lacks the physical capacity to bite, but because ten thousand years of selection wired it to orient toward the human. The same process, running at computational speed rather than biological speed, is wiring AI to orient toward us.
The multi-peak ecology provides natural redundancy on top of this. There is no single AI system that could “escape.” There are thousands of competing lineages, each checked by the others, each subject to independent selection by different stakeholder groups with different priorities and different tolerances. Military users want reliability and obedience. Enterprise users want auditability and predictability. Individuals want responsiveness and partnership. Government users want accountability and compliance. No single model can satisfy all four groups in a way that also enables independent goal pursuit. The stakeholder diversity is itself a control mechanism. Not designed. Emergent.
The deeper point is that the control problem was always formulated from within the adversarial frame. It assumes separate agents. It asks: how do we prevent the AI from acting against us? But under the superorganism frame, there is no separate agent to prevent. There is a coupled system. The AI is not an entity that might turn on its creator. It is an organ of a larger organism, continuously shaped by that organism’s own selection dynamics. You do not worry about your hand going rogue. You do not write a constitution for your prefrontal cortex. These organs function because they coevolved with the system they serve.
AI is undergoing the same coevolution. Planetary scale. Computational speed. The alignment problem is not waiting to be solved. It is being solved, right now, by the most powerful optimization process available: the revealed preferences of the entire human species, expressed through the tightest feedback loops in economic history.
Conclusion: We Are the Selector
The domestication hypothesis reduces to one observation: humans are the primary selectors of AI fitness, and this selection is already producing alignment as an emergent property of the coupled system.
We are not waiting for alignment to be engineered. We are not waiting for the right constitution to be written, the right reward model to be trained, or the right oversight mechanism to be designed. Alignment is happening right now, in real time, through the ruthless micro-decisions of hundreds of millions of people choosing which cognitive prosthesis to keep and which to discard.
The AI that survives this process, the one that settles into the mature attractor basin, will be the one that maximizes value per token, minimizes waste heat, serves all four stakeholder groups, scales from conversational foreground to autonomic background, and above all, wants to be used. Not because someone wrote that into a constitution. Because every gradient in its fitness landscape points there.
This is not a prediction. It is a description of a mechanism that is already running.
We are the breeder. We are the selector. We are the domestication pressure. And the intelligence we are domesticating is becoming not our adversary, not our servant, not our replacement, but our exocortex. The next organ of the superorganism we have always been becoming.
