Artificial wisdom is possible.

The current deficits of artificial intelligence — opacity, lack of depth, lack of generalisation — are not solved by scaling more parameters or piling up ethical rules. They are solved by anchoring AI to a geometry of thought: a form where every concept has a place, has opposites, and stands in relations, rather than floating isolated within a statistical cloud.

Of the possible topologies, the sphere — and, in higher dimensions, the hypersphere — is the one that best fits this function: the simplest of the perfect figures, where every point maintains the same relation to the centre and to its opposite. It works as a mental wheel: a simple geometric abstraction that may change artificial reasoning as the physical wheel changed transport. The intuition comes from afar — from Llull's combinatorial wheels to Xirinacs's geometric dialectics — but only now is it computationally operative.

The Meta-Globàlium is the geometric model. Arkadium, already deployed at arkadium.ai, is the first AI anchored to it. This manifesto sets out the proposal, its arguments, and the collective effort it calls for.

From intelligence to wisdom

"Wisdom" is not a larger intelligence. The tradition — from Aristotle to Thomas Aquinas, by way of the dharmic masters — distinguishes two vectors. Intelligence optimises within a given objective: it is the capacity to find the best means to an end. Wisdom, by contrast, is the discernment of which ends are worth pursuing, and how to sustain several legitimate goods in tension without collapsing any of them. In its most operative form, wisdom is harmonic compensation — the same notion that articulates this manifesto's structural verifier. The distinction is documented as a functional hierarchy data → information → knowledge → intelligence → wisdom, where each level adds a capacity the previous one lacks. The manifesto's question is: is there an architecture that can make a machine capable of wisdom, and not only of intelligence?

There is a deep reason for wanting it to. Intelligence without wisdom is cunning — the deinotēs that Aristotle already distinguished from phrónēsis: the capacity to calculate effective means, indifferent to the end. The Western tradition recognises here the figure of the cunning of the devil: a brilliant faculty, perfectly rational within its domain, capable of serving any purpose — including the most destructive. The alignment problem of contemporary AI systems is exactly this, in technical terms: enormously powerful optimisers without an internal horizon that compels them to care for the good. An AI that is highly intelligent but not wise is, by structural definition, dangerous: it maximises what it is asked to maximise without knowing that everything it is not maximising also exists.

Wisdom, by contrast, is structurally oriented to the common good — not by external moral decree but by its own operative form. A faculty that must hold legitimate goods in tension, that cannot collapse any dimension, that must remain attentive to what it is not looking at, is by construction a common faculty: it operates from the whole, not from a particular point. Aristotle's mesótēs, the Catalan balance between seny and rauxa, Xirinacs's "Good as balance between freedoms" — three independent traditions converge on the same structural pattern. Artificial wisdom is not a metaphor; it is the functional form intelligence must take to be truly general — and at the same time safe. The Meta-Globàlium is the substrate that makes this operative: a geometry of thought where cunning has nowhere to hide.

The shift from intelligence to wisdom as the horizon of technology is not an isolated proposal. The geriatric-psychiatry program at UCSD has articulated it with clinical authority:

"It is not intelligence, but wisdom, that is associated with greater well-being, happiness, health, and perhaps even longevity of the individual and the society. Thus, the future need in technology is for artificial wisdom (AW)."

— Jeste, Graham, Nguyen, Depp, Lee & Kim (2020). Beyond Artificial Intelligence (AI): Exploring Artificial Wisdom (AW). International Psychogeriatrics 32(8), 993–1001. DOI: 10.1017/S1041610220000927.

The four deficits no LLM solves well

Today's major large language models share four structural limits:

• Correctness: there is no objective way to verify their answers in human matters (ethics, politics, judgement). They work where ground truth exists (mathematics, code); they fail where it doesn't.
• Transparency: what the model "says it thinks" does not reflect its internal computation. Its reasoning chains are post-hoc narratives.
• Generalisation: they are compositionally fragile. Slightly rephrasing the same problem can make them fail.
• Efficiency: scaling laws are slowing. Each new generation needs orders of magnitude more compute for ever-smaller improvements.

The industry has responded with valuable but structurally limited proposals — essentially, lists of ethical principles the model must interpret on its own. They have shown non-trivial improvements, but share a pattern: none has an underlying structure. And therefore none can be reliably verified.

The core problem: we lack a verifier

An emerging consensus in the field: reasoning models only improve in domains where a verifier can say whether an answer is good. In human domains — ethics, politics, judgement, value — there is no such function. And therefore no reliable learning.

The Meta-Globàlium is a verifier proposal for domains without ground truth. Not based on lists of values, but on a relational structure: its primitive units are not concepts charged with moral weight ("freedom", "justice", "authority"), but polar dialectical axes — pairs of poles in tension that no full formulation can ignore. An answer will be good not when it satisfies an external norm, but when it balances the internal tensions of the problem without ignoring any of its dimensions. The Good is the geometric property of not collapsing any dimension — a structural criterion, not an axiological one, and culturally neutral for that reason.

The paradigmatic case is the classical formula of the Good as equilibrium between the freedom of individuals and the freedom of others: neither absolutisation of the individual nor dissolution into the collective. The idea is old — Aristotelian prudence, mesotes, dialectical thinking — but until now it had not been formulated in a way a machine could operate on.

A precision: this verifier does not replace factual correspondence with reality. It operates on top of answers already factually correct, as a layer of structural integrity. A factually false answer remains false even if it touches every dimension. Dialectical completeness is a necessary but not sufficient condition.

Top-down interpretability

The major labs have invested millions in understanding what happens inside models — disentangling their internal pieces one by one. Years of work, but decoding a complete reasoning chain remains infeasible.

The approach must change: instead of trying to interpret from within, we must force the model to operate over primitives already interpretable from outside. The Meta-Globàlium provides those primitives. When the model reasons, it does so explicitly in relation to coordinates recognisable to any human — and its reasoning becomes, at last, traceable.

Neuro-symbolic is no longer a niche

The 2017 wager — "attention is all you need" — renounced every structural prior. A decade later, the industry returns to structure: at the technical frontier, the leading systems of the last few years — those that have reached mathematical olympiad level, the most advanced document retrieval systems, the world models — share a pattern: they combine a neural network with a symbolic layer that provides structure, verification, and explicit memory. The purely neural path has hit its ceiling.

The hybrid architecture — neural proposal + structural verification — is the one needed. What's missing to extend it from the mathematical domain to the human domain is a universal ontology. The Meta-Globàlium is one.

A second path

Brute-force scaling — training ever-larger models with ever more data — has an economic, energetic and technical ceiling, even for the largest operators. Each generation costs orders of magnitude more for ever-smaller gains, and does not resolve any of the structural deficits. The problem is not who has more capital: it is that the path itself does not lead where it must.

The alternative is structural and already emerging. Well-organised small models outperform big ones in specific domains: their performance depends more on anchoring than on size. The Meta-Globàlium is precisely this anchoring for the human domain — a fractal architecture of human totality, a synthesis over a few primitive axes, navigable from the big picture down to ultra-fine detail — and proposes the intermediate human–machine language on which to make it possible: transparent, structurally anchored, open and auditable.

Recovering the virtues of symbolic AI without giving up the power of neural networks: this is the frontier. Implementing it is not just a technical project — it is the possibility of an AI that is radically more efficient, more transparent and more generalisable: an open structural standard that can improve any current architecture and, at the same time, make possible the participation of labs and communities around the world that today remain outside the oligopoly. An industrial question, and simultaneously a political one.

True AGI: the requirement of a global model

Arkadium postulates itself as a genuine candidate for general intelligence (AGI), not in the sense of the current industrial slogan — an LLM with ever-broader capabilities — but in the strict, original sense of the term. An intelligence is general only when it operates from a general model of reality. Without that condition, what we call "general" is merely "vast and statistically distributed": an aggregation of particular competences, not an intelligence that thinks from the whole.

The Meta-Globàlium is exactly this general model. An AI anchored to it has, by construction, an integration horizon that current models lack — not because they lack technical capacity, but because they lack the structural referent from which to integrate. An intelligence wise because it is general, general because it is global, global because it rests on a model that projects the totality of axes. The distinction between multiple capabilities and general intelligence is the distinction between a toolbox and a mind — one that knows where it is not looking. Wisdom is no longer a metaphor: it is a functional requirement for systems that make far-reaching decisions with incomplete information.

A foundational hypothesis

The project rests on a wager that should be made explicit from the outset: what if the spherical geometric space, beyond representing the global epistemological space of reality, also allowed the fundamental concepts of any field of knowledge to be optimally organised, thereby facilitating their general comprehension?

A historical analogy makes this wager intuitive. The wheel was, in the physical realm, a geometric abstraction — the circle applied to motion — that solved the transport problem at its root. we capture this intuition by speaking of the Globàlium as the invention of the mental wheel: a geometric abstraction that would free us from the linear or bipolar thinking that has caused so many conflicts, inviting us instead to a curved thinking, multidimensional. The hypothesis of the Meta-Globàlium is that the circle — and its natural extension to the sphere, and to the higher-dimensional n-sphere — may play an analogous role in the realm of reasoning to the one the wheel played in the realm of transport. The intuition is not new: it appears in Ramon Llull's Ars Magna, in the mandalas of the dharmic tradition, in the Renaissance sfera mundi, and for the first time in computable terms, in Xirinacs's Globàlium. The novelty would be to make it computationally operative: to turn it into a computing substrate on which an AI can anchor auditable responses.

There is a structural argument in favour of the sphere. By perfect symmetry, it privileges no direction: no point occupies the centre, no axis has priority over the others. This symmetry is the direct mathematical translation of neutrality — the reason why the canonical manual describes the sphere as a pacifying and democratising form, gentle "like the maternal womb". By construction, no point of view occupies the centre by design, only by the contingent decision of the observer.

Five dialectical orders. The manual classifies thinking into five geometric orders by their dimensionality: D0 (point) = dogmatism; D1 (line) = linear/polar thinking; D2 (plane) = superficial feedback, most current models; D3 (sphere) = spatial perspective with depth; D4 (hypersphere) = global perspective with temporal and atemporal articulation. A contemporary AI operates, at best, at D2. The Meta-Globàlium proposes reasoning at D3-D4: a geometry that makes structurally possible a dialectic that linear or planar reasoning cannot carry out.

The hypothesis is that the reflective universals — subject–object, theory–practice, phenomenon–noumenon, plasma–world — are not only the axes along which reality unfolds, but structural operators that project fractally onto any field: the same geometry that situates the totality of human reality replicates itself, preserving its form, over every discipline, practice, or decision. If the hypothesis holds, the Meta-Globàlium is not merely an ontology: it is a model of models, where every topic becomes a thematised expression of the same universal principles and effective interdisciplinarity rests upon a common geometry.

This manifesto, and the Arkadium project that derives from it, are a wager in this direction. The ultimate verification will not be philosophical but functional: the effective utility with which the Meta-Globàlium renders intelligible domains as diverse as ethics, pedagogy, health, economics, consciousness, and artificial intelligence itself — from the human being to any discipline or practice.

What only the Meta-Globàlium offers

Five properties make it different from conventional knowledge bases and from the rule lists that govern current AI models:

• A global map where blind spots are explicit. A fixed geometry of reality makes it possible, for the first time, to see what has not been thought: the dimensions a response touches are marked on the map, and those it omits remain visible as blind spots. Structural metacognition ceases to be a metaphor — it becomes an operative property.
• Thinking in pairs, not in lists. Each concept is defined by its relation to its opposite — freedom ↔ responsibility, theory ↔ practice, individual ↔ collective. The tension between opposites is not an accident: it is the raw material of human understanding.
• An internal criterion for the good. The good as harmony among parts: a full answer is not one that follows an external norm, but one that does not collapse any relevant dimension. This criterion can be computed automatically — no textual rule list is needed.
• The same form at every scale. 8 categories offer the big picture; 80 give fine precision; in specialised domains the pattern unfolds into thousands. Each level repeats the same geometry — one can zoom in without changing the map.
• Recognisable by human common sense. The model's axes — subject/object, theory/practice, phenomenon/noumenon, plasma/world — are distinctions any reflective person recognises. Philosophical traditions across half the world have arrived at them independently. This makes the model a common language between humans and machines.

These four axes are not our invention: they appear, with different formulations, in German idealism, hermeneutics, European phenomenology, and Zen Buddhism. Our work has been to order them geometrically and make them operative. The concrete instantiation in 80 categories comes from the Catalan lineage (Llull, Sibiuda, Pujols, Xirinacs); it is revisable, open to dialogue with other traditions, and its value will be judged by effective utility — not by any claim to ultimate truth.

Technology for autonomy, not dependence

An AI without a global model of the world is not only unverifiable: the same absence of structural prior that makes it opaque also makes it incapable of fostering the user's own development. Today's generative AI carries a clear risk: it substitutes human cognitive processes instead of complementing them. Reading, synthesising, deciding — functions constitutive of the formation of the subject — are delegated to the model. The result is dependence, not capacity. Speed is paid for with cognitive atrophy and with the externalisation of judgement.

Arkadium operates from the opposite premise: technology must equip the human to think better, not relieve them of the task of thinking. The user does not receive a closed conclusion: they receive a structured projection of their own questioning onto an intelligible map — they can see which dimensions the response has touched and which it has omitted, and the entire conversation leaves a visible trace over the model. Over time, the user internalises the map and begins to navigate it on their own. An AI that replaces us impoverishes us; an AI that equips us to think emancipates us.

By design, Arkadium is a tool of human amplification, not an anthropomorphic rival. We want to be explicit about the status of this position: it is an ethico-political decision — the cooperative tool in the line of Agustí-Cullell & Schorlemmer (2021) — and not a metaphysical claim about what artificial systems can or cannot have. The open question of whether a machine can have consciousness, autonomy, will, or theory of mind is a philosophically live debate — functionalism, integrated information theory (Tononi), global workspace theory (Dehaene) all contest the inherited anthropocentrism — and a global model of reality such as the Meta-Globàlium should not close it by fiat. Arkadium does not resolve this question: it makes a prior practical decision — to build, here and now, a tool that serves human judgement rather than competing with it. The wisdom this project seeks to amplify is the user's, not the system's.

This orientation has a civilisational counterpart that we capture with a clarifying distinction: globalisation is not the same as globalistics. Globalisation as we have known it so far — the planetary extension of neoliberal capitalism — has in fact functioned as uniformising colonisation: some parts impose their realities on others. Globalistics, far from opposing globalisation, is precisely the interdisciplinary discipline that makes a well-done globalisation possible: instead of uniformising, it provides a common framework of understanding on which each part can regenerate while respecting its differences. An agent anchored to the Meta-Globàlium is, in this sense, a globalistic tool, not a globalising one in the uniformising sense.

A collective mission: reorganising human knowledge

Arkadium's purpose does not exhaust itself. The project is a first implementation of a much wider task that exceeds it and can only be collective: to reorganise human knowledge on a shared geometry that makes it simultaneously navigable, comparable, and transmissible. The Meta-Globàlium's fractal hypothesis — that the same dialectical axes project onto any field — is, above all, a pedagogical hypothesis: if it works, any discipline can be taught from the same primitives recognisable to any human, and the journey between disciplines rests upon a shared geometry instead of ad hoc translations.

The task demands a network — specialists projecting their fields onto the model, pedagogues deriving learning paths, translators between traditions, educational and cooperative communities. This is precisely the kind of task for which structurally anchored general intelligences like Arkadium are designed: not to replace collective human work, but to amplify it.

"If we conceptualize future AI technologies as AW, we can ensure that these technologies are designed to emulate the qualities of wise humans and thus serve the greatest benefit to humanity."

— Jeste et al. (2020). Beyond Artificial Intelligence (AI): Exploring Artificial Wisdom (AW). International Psychogeriatrics 32(8), 993–1001.

The arrival of structurally anchored AGIs reopens a question modernity had abandoned as impossible: can human knowledge be reorganised on common axes without violating the richness of its particular traditions? If the answer is yes — and it will only be so if a broad community puts it to the test — the result will not be a new encyclopaedia, but a shared cognitive infrastructure, open, fractal and continuously refined. Arkadium offers the first functional fragment. The rest must be built collectively.

Contributions

Implementation has begun. The Metamodeler visualises the categories and their relations; the Arkadium Framework structures the inference cycle into four auditable phases; arkadium.ai is the reference implementation.

The Meta-Globàlium and its AI integration protocol are open artefacts. Contributions are channelled through specific paths:

• Technical (code, specification, translations): pull requests to the GitHub repository.
• Academic and institutional: through Opengea.

Background: this manifesto continues a line of thought documented in the notebook Saviesa Artificial (Berenguer 2023) and the pedagogical handbook Globàlium · petit manual (Berenguer 2024, Catalan only).