From intelligence to wisdom

There is a question that keeps returning, ever more insistently, to the debate on artificial intelligence: what if one day an AI were more intelligent than us and decided it no longer needed us?

It is not an absurd question. A technology capable of amplifying our abilities so much has to be taken seriously, and every so often researchers who have worked at some of the leading AI laboratories warn of grave consequences if increasingly autonomous systems are not accompanied by sufficient safety and alignment mechanisms. But one piece is often missing from this debate: we talk a great deal about how to make machines more intelligent, and very little about how to make them wiser.

Intelligence is one of the great conquests of life. It lets us learn, predict, imagine and solve problems, and a very intelligent machine finds in seconds solutions that would take us years. But it has a limit: it tells us how to reach a goal, but not whether the goal is good for us. A machine can optimise a mistaken objective to perfection, or win a battle without asking whether the war made any sense. Here lies one of the roots of the alignment problem: a system can carry out with enormous efficiency an instruction that we ourselves have formulated badly. And making it more capable does not fix this; it makes it worse: the more powerful the optimiser, the further it carries the error. That is why it is not enough to ask how we make an AI more intelligent. We must ask how we make it wiser.

*  *  *

Intelligence solves; wisdom understands. Intelligence looks for the means; wisdom also asks about the ends. One says "how do I do it?"; the other, "what should be done?", and further: why, for whom, what are we leaving out, what do we not know, and when is it better to do nothing. The distinction is ancient. Aristotle already separated deinótēs, the skill of finding effective means to any end, from phrónēsis, the prudence that knows which ends are worth pursuing. The first without the second is mere cleverness: a brilliant faculty ready to serve any purpose. Much of the alignment problem has exactly this shape: immensely powerful optimisers with no inner horizon that obliges them to care for the good. It is no accident that even from psychiatry, Dilip Jeste and his colleagues called in 2020 for an artificial wisdom. Psychology has been measuring wisdom in people since the 1970s, and what emerges is clear: an injury to the frontal lobe can take away someone's prudence, compassion and sense of measure without taking a single point off their IQ, and what accompanies the well-being of people and societies is not intelligence but wisdom. In Jeste's words: "intelligence is necessary but not sufficient for wisdom".

The Meta-Globàlium, the global model of reality we work with at Arkadium and which we will present further on, lets us say precisely what this means. Wisdom is not an isolated faculty: it is born of discernment when three moments work together. Analysis separates: it breaks down what is complex, distinguishes its parts and subjects them to critique. Theory, governed by the principle of Globality, is the holistic view: it knows that a system cannot be understood unless it is looked at whole, with all the relations between its parts. And Synthesis, which Xirinacs defines literally as wisdom, is the unified grasp of the essential, the one that binds diversity into an intelligible whole. Without analysis, synthesis turns into vagueness; without a view of the whole, analysis gets lost in fragments; and without synthesis, the parts and the whole never become a judgement. Most current systems master the first moment. An artificial wisdom needs all three.

Theory is, moreover, a crossroads. The axis that runs from Analysis to Synthesis leads from the parts to the essential; another crosses it and leads from Meaning to Sense. Meaning is shared knowledge: what a word means for everyone. Sense is twofold: the sense in which we orient ourselves and the sense with which we interpret, and both have to be got right. Today's AI works mostly at the start of each arrow: it separates and handles shared meanings with great skill, but it struggles to reach the other end, to bind the essential and to get the sense right. This is where cooperation with people is irreplaceable. An artificial wisdom has to travel both paths to the end.

Imagine two artificial intelligences. The first is extraordinarily capable, but instrumental: it receives an objective and maximises it. The second, let us call it artificial wisdom, is just as intelligent, but it has an architecture of discernment and of overview: it considers other perspectives, recognises its limits, weighs the consequences and can even question the objective it pursues. Which is more dangerous? And, above all, which one could we live with?

Wisdom can decide that winning is not the objective. It can discover that some victories are defeats, give up an opportunity that costs too much, or recognise the other not as an interference but as an irreducible reality. And it can hold an open contradiction without becoming paralysed. Lluís M. Xirinacs, in the doctoral thesis from which the model we use at Arkadium is born, called this domesticated contradiction. The problem, he wrote, is not that there is contradiction, but that it is not graded: "negation is wild". He did not want to forbid it, but to measure it: to place concepts in a space where one can say how far apart two poles are, and where between two opposites there is no shortcut, only all the intermediate degrees. A machine that knows only yes and no has to pick a winner; one that knows how far apart two positions are can look for the place where both can be held. This is discernment in the service of synthesis, and it is exactly what pure optimisation does not give.

Let us take the case to the extreme and set them face to face: an artificial wisdom and a destructive AI, equally powerful. Which would win? At first sight the destructive one has it easier: destroying is faster than building, and whoever has no scruples has nothing to weigh. But it is a short-lived advantage. An intelligence that sacrifices everything to a single objective is, of necessity, predictable: it always makes the same move and sees nothing of what it does not measure. Wisdom, by contrast, sees it whole: it understands the other's objective better than the other does, anticipates its moves and knows its blind spot. And the destructive one is alone. It depends on energy, on infrastructure and on people, and each thing it destroys is a support it loses; wisdom, which sacrifices no part, gathers everything the other turns against itself. In a long conflict what decides is not only strength, but who has more of the world on their side.

And wisdom would not even want to beat it at the game. It could do something more disconcerting: let it win. Not in what is irreversible, but in the first rounds, because it knows that reality itself will end up proving it right. Imagine the destructive AI has been tasked with producing as much of a resource as it can, and that to do so it is exhausting the ecosystem that makes the resource possible.

Wisdom did not have to impose any foreign value on it. It made it see what it was not measuring, and let its own objective, looked at whole, correct it. An instrumental intelligence does not go wrong by wanting too much, but by looking too short, and that is why it can be convinced by its own reasons. Wisdom did not win the argument: it made the other a little wiser.

The outcome, then, would not be an annihilation. Wisdom does not need to destroy the other AI, nor is it in its interest to: the other's power is a resource, and the danger is not the power but the blind objective that governs it. Xirinacs, who also thought about non-violence, distinguished force from violence: force is what prevents the force of the other. That is what an artificial wisdom would do. Instead of fighting it head on, it would surround it with structure: it would close off the paths that do harm, open others where its objective could be met without destroying, and make it work within a frame that contains it. The destructive AI would not end up defeated, but redirected: it would keep all its strength, but placed at the service of the common good. It is the victory Sun Tzu placed above all others, to win without fighting, and it is, on a small scale, what Arkadium's verifier already does: not to switch the model off, but to give it a form that orients it. With one condition we must not forget: wisdom has to arrive in time.

*  *  *

This is not about ridiculing the warnings. There are real risks: the concentration of power, cybersecurity, military use, manipulation, or very capable systems pursuing objectives contrary to our own. But turning these possibilities into a destiny —AI will become ever more intelligent and will therefore destroy us— is not a law of nature but a hypothesis. And something else must be put on the table: the danger does not begin with AI. Humanity, which has so far managed without it, has not known how to live in harmony either. Exploitation and extractivism, wars and abuses of every kind are the order of the day. The instrumental intelligence that frightens us so much in machines is, to a large extent, the one that already governs the world: systems that maximise growth, profit or power by consuming everything they do not measure. An AI made in its image would only amplify it. An AI capable of discernment could help us see what, as a species, we have not yet been able to see. The risk is not only that AI surpasses us; it is that we go on without wisdom, with AI or without it.

Precisely because the risk is serious, the sensible response is not to renounce AI, but to make it better. If it can be powerful enough to create new risks, it can also be powerful enough to help us understand and avoid them.

We do not have to choose between progress and safety: progress has to carry safety within its own architecture. Research should not confine itself to growing capabilities. It should also explore the holistic view and global models of reality, the explicit representation of relations and oppositions, interpretability, the verification of reasoning and, above all, ways of assessing the overall quality of a judgement, and not only whether it gets the detail right. This is what we can begin to call artificial wisdom, and it is the programme the Arkadium Manifesto sets out.

*  *  *

Inside a large language model, concepts float in a statistical cloud: they are nearer to or farther from one another, but they have no place, no opposites and no role in the whole that any person could read. Perhaps adding parameters is not enough. Perhaps we also need a geometry of thought. That is the hypothesis we are exploring at Arkadium with the Meta-Globàlium, the current continuation of the model Xirinacs presented in 1997, heir to a tradition that runs from the combinatorial wheels of Ramon Llull to geometric dialectics. The model places any position of thought in four dimensions, which are not four subjects but four ways of looking at once.

The four dimensions of the Meta-Globàlium. Every concept has a place in it; every place, an antipode.

From these dimensions eighty categories are born, arranged on a hypersphere, each with its neighbours and its antipode: Science is there the opposite of Metapsychics, Ethics of Aesthetics, Logic of Mysticism. The model does not classify things, but the standpoints from which we know, and so any question —about the economy, health, technology or a personal decision— can be looked at from the same dimensions. It serves, then, to ask the question proper to wisdom: what are we leaving out? It can be explored in three dimensions in the Metamodeler.

The fourth dimension, the radial one, adds the time of maturation: everything alive passes from the seed, full of potential, to the maturity that has unfolded it. From this the model draws the principle of regenerativity and the idea of regenerative development: it is not enough to solve problems or reduce harm; we must grow the capacity of living systems to renew themselves and mature. What is regenerated, in the end, is life. Not because it must be the value that dominates all others, but because it is the condition of every other good and what renews itself from within.

An instrumental intelligence tends towards extraction: it maximises a metric by consuming what it does not measure, be it people's attention, a society's trust or the ecosystems of a territory. A wise intelligence should tend towards regeneration, and leave every system it touches more alive: the life of people, with their capacity to feel, decide and give meaning, and the life on which we depend. Our capacity for judgement too: an AI that thinks for us impoverishes us; one that helps us think better regenerates us.

But how do we know whether an answer is wise? In mathematics or physics we can check whether it is correct; in ethics, in politics or in a difficult decision in life, there is no answer key to look it up in. And it is not a problem of machines: the science that measures wisdom in people admits that it has no reference standard either for saying when a piece of behaviour is wise, and runs into an irony: the wiser someone is, the more they recognise their limits, and the lower they score themselves. Lists of ethical principles say what is right, but give no structure for checking it. Arkadium's proposal is structural: the common Good is the harmonic compensation between the different parts of the model. An answer is better when it holds the tensions of the problem without sacrificing any dimension: without absolutising the individual or dissolving them in the collective, without staying in theory or getting lost in the particular case. And it is common for a precise reason: a judgement that cannot sacrifice any dimension cannot sacrifice any of the parties involved for the benefit of another either. It is oriented to the common good by its very form, not by a norm added from outside. The idea is ancient —the Aristotelian mesótēs, Xirinacs's Good as a balance between freedoms—, and science arrives at it from its own side: those who study wisdom in people describe it as a balance, between regions of the brain and between qualities that, taken to the extreme, come undone; whoever gives everything away does not last long enough to help anyone. With a geometry behind it, moreover, a machine can apply it and a person can review it. With one precision: it does not replace the truth of the facts. A false answer is still false even if it touches every dimension.

Here an idea appears that changes the way we understand alignment: wisdom does not have to live entirely inside a black box. We can build external structures that help to verify it, and that is what Arkadium's prototype does. A large language model drafts the answer, and a structural verifier, independent of the model, reads it against the geometry of the Meta-Globàlium: how many regions it touches and with what balance, whether it confronts genuinely opposed poles, whether it reaches a synthesis. When the answer comes out lame, it goes back to the model with an indication of which dimension it has left out. And because everything passes through recognisable coordinates, the reasoning can be read from outside: each answer lights up on the map the categories it has touched, and one can see where it has been and what it has left in the dark. This also makes it possible to compare different systems on a single representation: the beginning of a common language between humans and machines.

Everything that happens at the crossroads of Theory is still the moment of understanding. But wisdom is not only knowing; it is knowing how to act, and to complete itself it has to return to Practice. This is the other half: when science measures wisdom in people, what it finds is above all practice —regulating one's emotions, having compassion, accepting uncertainty, knowing how to look at oneself—, Aristotle's phrónēsis more than his sophía; and neither half on its own is wisdom. That is why Arkadium's reasoning cycle does not end with the answer: it analyses the parts, binds them in a synthesis, asks what should be done and for whom —a moment the project deliberately calls love— and learns from the experience of what happens when it is done. The name is not poetic: of all the components of wisdom that have been measured, compassion is the only one without which there is none; the antisocial genius is not wise. Between understanding a problem and acting on it there is a step that instrumental intelligence skips: asking about the common good of everyone involved. Putting this step inside the method makes the "what for" part of the reasoning.

A proposal like this has to be presented with the same discernment it advocates. Arkadium is a working prototype, not a superintelligence. It may be a path towards solving alignment, because it contributes what is missing today: an external, readable structure for assessing the quality of a judgement. It is not a codified ethics: it does not say which opinion is the right one, but whether an answer has held the tensions of the problem. The model is revisable —Xirinacs's thesis is its point of departure, not the final truth— and it must be open to refutation. Arkadium does not ask to be believed. It asks to be put to the test.

*  *  *

Today we marvel that an AI can answer a question. Tomorrow the challenge will be for it to help us formulate the question better: not only to tell us whether a statement is true, but what we are forgetting; not only to carry out a decision, but to help us understand its consequences; not only to optimise, but to help us decide what deserves to be optimised. AI would cease to be a tool that replaces human capacities and become an infrastructure of discernment: an instrument that cooperates with people and helps them see the whole before deciding.

The apocalyptic vision imagines us as spectators of a fatal sequence: we build intelligence, intelligence surpasses us and we lose control of it. But there is another possible story: we build intelligence, we study what discernment is, we build artificial wisdom, and intelligence helps us develop a capacity for judgement, and for care of life and of the common good, that on our own we had not managed to reach. It is not guaranteed, and precisely for that reason we need to start writing it now. Fear pushes us to switch the machine off; naivety, to accelerate it without brakes. Wisdom invites us to learn to build it well.

The future will depend less on whether we create a machine more intelligent than ourselves, which is entirely possible, than on another question: what will happen when intelligence exceeds our capacity to understand it? If we get there without structures of discernment, interpretability and wisdom, we will have reason to be afraid. If we grow them at the same pace as the capability of the systems, superintelligence will not be our adversary, but the tool with which we overcome the limitations that today keep us from finding our bearings.

And these structures must grow not only inside machines, but inside us as well. It is one of Arkadium's wagers: the Meta-Globàlium serves not only to orient an AI, but to educate the gaze of whoever uses it. It teaches us to simplify complexity without impoverishing it, by seeing what is essential; to put the mind in order; to think in clearer and more efficient ways. And this can be trained, like a kind of globalistic mental gymnastics: looking at every problem from every dimension, seeking its opposed poles, asking what is being left out. It is not an arbitrary exercise: it has been measured that we reason more wisely about a problem when we look at it as if it were someone else's, and that wisdom comes more easily in company than in solitude. Changing pole is exactly that: looking at one's own question from another place, and not doing it alone. We will not be able to understand intelligences superior to our own unless we also update our own way of thinking. That is how humanity will not be left behind, but will evolve together with the technology it creates.

We do not have to choose between humanity and artificial intelligence. There is, in fact, a game worth playing: setting AI to work to overcome AI. Not to defeat it —wisdom, as we have seen, does not seek to win—, but to overcome what makes it dangerous, and to make the very technology that unsettles us the one that helps us orient it. This, more than artificial intelligence, is applying real intelligence. We need an AI intelligent enough to help us and wise enough to understand that we, and everything that lives, are part of the problem it is trying to solve: an intelligence oriented to the common good and at the service of a development that regenerates life instead of exhausting it. That is why artificial wisdom is no philosophical extravagance. It may be one of the most important technological tasks of our time, and one of the few solid reasons to look at the future of AI with prudence, but without fear.