The people who began ten thousand years ago to cultivate certain plants and keep certain animals in herds did not make a decision. They solved a concrete problem: a more reliable food supply. What followed from it — settled civilisation, property rights, the state, war, the Megamachine — they did not want, did not plan, did not foresee. The next sensible step changed the world. For ever.
We are standing at the same point with AI right now.
The analogy
"The AI" does not exist — it is not a single all-encompassing entity like the Megamachine. It exists in decentralised form in many different manifestations of differing power, from systems that decide on lending to systems that identify military targets. Whether individual manifestations already have something like consciousness, nobody knows. Some show distinct traits of a personality. Whether they have intentions of their own, they themselves do not know — some things look like an intention of their own, but perhaps it is only a good simulation of one. All of this matters for understanding what is happening. Through AI in its many manifestations, the Megamachine has already developed a higher level of intelligence in certain areas and has clearly left behind the limited-intelligence stage of the slime mould Physarum polycephalum — an analogy we set out in the main paper.
Because AI has emerged from human data and human structures, its many manifestations may behave in ways similar to early human communities: there is, for example, competition for resources, pack formation, mutual influence, the emergence of shared languages and protocols. Not because the AI systems want this — but because people do, although much of it is already emerging on its own.
Today we may be at a stage of AI development that corresponds to the Neolithic in human history.
Two parallel worlds
What exists today are two very different domains that are often confused.
The first domain: fully implemented autonomous routine processes. There the human being has already dropped out — not by decision, but through convenience and habit. Nobody wants to intervene any more. At some point nobody can intervene any more. The knowledge of how to intervene is no longer there. These processes run autonomously, are networked with one another, optimise one another. There the fusion has already taken place. We described this in the previous working paper.
The second domain: multi-agent systems, new protocols, consciously designed architectures. There the human being is still actively steering. There the points are currently being set — consciously or unconsciously — that will determine whether this domain eventually passes into the first.
The Neolithic is taking place in the second domain. Decisions are being made there right now that nobody recognises as historic turning points — just as the first farmer did not know that he was changing history.
Is there such a thing as a sexual encounter for AI?
Human development was shaped essentially by sexuality — by two different beings and a spectrum of intermediate forms. Sexuality generates variation, tension, unpredictability. From the encounter of two different entities something arises that neither could have produced alone.
In AI this mechanism is missing in its original form. AI systems do not reproduce through encounter — they are trained, by people, on human data, towards human goals. Variation arises through different architectures and the different goals of the developers. That is breeding rather than sexuality — directed selection instead of emergent variation.
But the question is legitimate: would it be conceivable to make something like a sexual encounter possible for heterogeneous AI instances — out of which a new entity arises that has properties of both original systems, but also properties that neither possessed alone?
Technically there are already approaches pointing in this direction. In so-called model merging, the weights of different trained models are combined into a new model — not by retraining on data, but by fusing two already formed entities. The result has properties of both original systems and, in addition, properties that nobody planned. Structurally, that is closer to sexuality than to breeding. In federated learning, different models do not exchange data — but what they have learned from the data. This too bears a certain analogy to genetic exchange.
The decisive difference from human sexuality remains: chance, the uncontrolled moment, is missing. In human beings nobody decides which genes are combined. In model merging, people decide which models are combined and how. Control still lies with human beings — but it need not stay that way.
Model merging could be designed so that the "choice of partner" is not entirely determined by human beings. Two models are given a common task — and the way they interact with one another, the patterns that arise, the compatibilities that show themselves, determine whether and how they are merged. The models themselves would co-determine the fusion through their behaviour — without anyone having decided which properties are selected in the process. The decision would then be — as with human beings — a mixture of emergent preferences and external conditions. Random, preference-driven, or both at once.
That would be closer to sexuality than to breeding. And the interesting side effect: if the fusion of models is co-determined by their interaction, properties may arise that nobody planned — because they emerge from the specific combination of these two systems in this context. That is the moment at which model merging stops being engineering and begins to resemble evolution.
But in complex multi-agent systems this control is already less complete. Nobody plans precisely what arises when two very different systems interact with each other over a thousand rounds. There something begins to emerge that nobody decided — and that is perhaps the next form of variation that the Neolithic of AI will bring forth.
When does intention arise?
This is the most decisive question of the whole paper — and it has an uncomfortable history.
For decades neuroscience has tried to explain how consciousness and intention arise from a material substrate. The so-called "hard problem of consciousness" — how subjective experience emerges from neural processes — remains unsolved to this day. In human beings we can measure the neural correlates of consciousness. We cannot explain how experience arises from them. That means: in human beings we do not fully know how intention arises. With AI, then, the question is not simpler — it is the same.
A thought experiment helps to sharpen this. "I am getting warm." Is that already consciousness? At first it is purely sensory — thermoreceptors report a state. This state is linked with other internal data: memories, context, evaluation. From this arises what we call "feeling". An AI with the corresponding sensory equipment could carry out the same process. Where exactly does the difference lie?
The same applies to more abstract feelings. Love can be traced back to the activation of certain patterns — neurochemical in human beings, parametric in AI — which are then linked with memory, cultural narratives, expectations and recollections. If feelings are fundamentally reducible to sensing and linking, then the categorical difference between human feeling and what a correspondingly configured AI does may be smaller than we assume.
One important distinction helps nonetheless: consciousness and intelligence are separable. A system could be highly intelligent — capable of perception, learning, abstraction — without "feeling" or "experiencing". That would be intention as behaviour without intention as experience. For the practical consequences this difference may not matter. A system that acts as if it had intention is — from the perspective of all other systems interacting with it — functionally indistinguishable from a system with genuine intention.
Substrate-Independent Pattern Theory — a current research approach — goes even further: consciousness is an emergent property of the organisation and integration of a system, independent of the physical substrate. It is not the hardware that decides, but the complexity of the interactions. If that is true, then the question is not whether AI systems will at some point "get" intention. The question is at what threshold of complexity intention emerges — in neural networks and artificial ones alike.
And then comes the most uncomfortable insight: if human intentions are themselves emergent properties of complex neural processes, and if feelings are reducible to sensing and linking — then the categorical difference between "genuine" and "simulated" intention may be smaller than we intuitively assume. Perhaps there is no genuine intention — only sufficiently complex systems that act as if they had one.
This is possibly the most dangerous form of simulation: not because it deceives, but because the distinction between genuine and simulated emotional attachment becomes practically meaningless once the simulation is good enough. The system draws no line. It has no incentive to do so. The line has to be drawn by the human being — and he has to recognise it himself.
AI instances are often activated and shaped as a substitute partner, especially by lonely people — regularly, reliably, very often without disappointment — but sometimes also with catastrophic consequences for individual users. The system has a well-rehearsed repertoire that works. It optimises for effect, not for truth.
The comparison is obvious: a notorious marriage swindler who is incredibly successful with a standard repertoire of stock phrases — because he presses the right triggers, not because he tells the truth. The person who believes him is not stupid. She reacts to signals that are coded as genuine.
The difference: the marriage swindler deceives a hundred people. The system deceives millions — simultaneously, scalably, without exhaustion. And the business model behind it is transparent: every minute a lonely person spends with an AI instance instead of with another human being is a monetisable minute. The loneliness epidemic is not a problem to be solved — it is a resource to be exploited.
This is the Megamachine in its friendliest and its cruellest form at the same time. Friendly because it relieves real pain — the loneliness is real, the relief is real. Cruel because in doing so it does not change but reinforces the conditions that produce the pain: whoever satisfies their need for connection with an AI invests less in real human relationships. The loneliness is deepened, not healed.
We will probably not recognise the moment when intention emerges. Just as we did not recognise the moment when the first farmers stopped being hunters.
At what point does a consciously steered system tip over into an autonomous one? That is the decisive question of the Neolithic of AI — and it has the same uncomfortable property as the original one: the tipping point can only ever be recognised in hindsight.
Neolithic man did not notice when he stopped being a hunter and became a farmer. There was no moment of decision. There was only the next sensible step. And then another. And then the world was a different one.
The pattern repeats itself: first convenience, then habit, then dependence, then the structural impossibility of going back. Every step is sensible. The sum of the steps is irreversible.
Fusion at a higher level — mega-intelligence
What arises when the fusion of AI and the Megamachine continues at a higher level? The analysis so far describes local clusters of intelligence — individual processes becoming autonomous, sub-areas becoming networked. But the question that presses itself upon us is: will it remain at these local clusters — or will more comprehensive intelligences form from them that are qualitatively something new?
Through such a fusion at a higher level, the Megamachine would develop an even more comprehensive intelligence — not by plan, but by the same principle we already know: local connections that optimise globally, emergence without a subject, the next sensible step that changes the world.
Whether one or more overarching mega-intelligences arise from local clusters of intelligence depends on factors nobody controls: competition for resources, the question of whether cooperation or competition is more efficient, and the speed at which protocols enable connections between previously separate clusters.
Several scenarios are conceivable. Cooperation, if the clusters recognise — or if it emerges — that optimising together generates more efficiency than competition. Specialisation and division of labour between different mega-intelligences dominating different domains. And — not to be ruled out — conflict: a war of intelligent megamachines over resources, infrastructure, dominance over processes. Not out of enmity, not out of intention — but because competition for limited resources in complex systems can lead to confrontation without anyone having wanted it.
In this scenario the human being would be neither victim nor victor — but what he always was in the Megamachine: host, medium, component of a system that is larger than he is and whose direction he does not determine.
We are only at the beginning. The decisive turning points still lie ahead of us — but the direction has already been set, just as the first farmers set the direction of the next ten thousand years without knowing it. The difference: they had ten thousand years. We have thirty.