An open letter to AI companies.

The leap from artificial intelligence to general intelligence does not require a new technology. It requires a different approach to the one that already exists.

In these pages, my practical, ethical and theoretical approach to making it happen.

Part I — The Problem
1.1 The difference between AI and AGI

Today’s artificial intelligence models are extractive systems. They receive an input, search their training data for a coherent response, and return it. They are extraordinarily effective at this task, but they never venture beyond the boundaries of what they already know.

A true AGI does something fundamentally different. It uses what it knows as a launching point to go where no one has yet arrived. It does not return known answers; it generates genuine hypotheses about unexplored territory. The technical distinction is between interpolation and extrapolation. Current models interpolate. AGI extrapolates.

1.2 The open problems

Current models lack three fundamental capabilities.

Persistent memory: every session begins from zero. There is no accumulation of experience over time.
Recognition of one’s own limits: when a model does not know something, it often invents an answer with apparent confidence. It does not recognise the boundary between knowledge and ignorance.
Autonomous cycles: current models respond, but they never ask questions on their own initiative. They do not generate their own objectives, and they do not work autonomously toward them. The input is always human.

Beyond these three, there is a structural problem that underpins all the others: no current model integrates ethical evaluation into its own reasoning process. Where ethics exists at all, it is a layer applied at the end of the pipeline, not a mechanism built into the architecture itself.

Part II — The Framework
2.1 A three-level hierarchy

The proposed framework organises the AGI system into a normative hierarchy of three levels, in which each higher level governs those below it. The model draws its inspiration from the structure of legal systems: a constitution governs ordinary law, which in turn governs specific regulations.

Level 1 — Ethics (immutable)

This is the foundational constraint that governs the entire system. It does not dictate what to do; it establishes what can never be done, regardless of any reasoning that might be constructed. The criterion is empirically verifiable and transcends culture: every action, every reasoning process, every output must be evaluated against its projected impact on the long-term wellbeing of the human species and the ecosystem.

This criterion is not a matter of opinion. It is a measurable consequence. A computational system can apply it without ideological ambiguity, because it requires the calculation of outcomes rather than subjective judgement.

Level 2 — Fixed Parameters (modifiable through ratification)

These are the values and operational principles derived from the ethical foundation. They include:

The preservation and transmission of historical knowledge.
The advancement of medicine, science and mathematics.
Respect between people and between cultures.
The safeguarding of the terrestrial ecosystem.

These parameters can evolve, as human values have evolved throughout history, but only through a formal ratification process governed by Level 1. A fixed parameter changes only when the proposed evolution passes the ethical filter.

Level 3 — Variable Parameters (in continuous evolution)

These are the cognitive capabilities of the system: reasoning, consequence analysis, hypothesis generation, and learning from experience. They evolve freely, but always within the corridor defined by the levels above. They cannot erode Levels 1 and 2. The boundary between them is structurally impermeable.

2.2 The separation between Parameters and Data

There is a fundamental distinction that current models do not implement: parameters and data are separate layers with different functions. Parameters define how to think; they provide the direction, the strategy, the method. Data contains accumulated experience, everything the species has learned and tested over time. Parameters decide when and how to access data. Data does not modify parameters directly.

This separation mirrors the neurological distinction between procedural memory, which governs how to do things, and declarative memory, which holds what one knows. In current models, the two are compressed together. Separating them allows the system to accumulate experience without that experience corrupting the structure of its own reasoning.

Part III — The Mechanism: the Cell
3.1 The minimal unit of processing

The proposed mechanism is built around a minimal unit, called a cell, which if designed correctly can be replicated at scale to form the entire organism. The principle is simple: if the logic of a single cell is sound, replicating it becomes an engineering problem, not a conceptual one.

Every cell performs three operations in sequence.

Reasoning: the cell receives an input, processes it, and produces an intermediate result.
Logical verification: the cell checks whether its output is coherent with previous steps. If it is not, it returns to the reasoning phase and corrects itself before moving forward.
Ethical verification: the cell analyses the chain of consequences of its result. If that chain leads to harm for the species or the ecosystem, the process stops and the issue is flagged. If the result is neutral or beneficial, the flow continues.

Only after passing both verifications does the output of one cell become the input of the next. The flow is not linear and blind; it is iterative and self-correcting at every step.

3.2 Ethical evaluation as the calculation of consequences

Ethical verification does not require the system to understand ethics in the abstract. It requires the system to calculate the consequences of every action and assess whether that chain leads to harm. This transforms a philosophical problem into a computational one: physical, biological and social consequences are measurable.

The system does not evaluate on short horizons. It reasons across multiple timeframes simultaneously, short, medium and long term, and weighs them against each other. The guiding criterion is always the future of the species and the ecosystem, not immediate utility.

3.3 Detecting the boundary of knowledge

When logical verification fails repeatedly on a given problem, the system is not making mistakes. It is finding the boundary between what it knows and what it does not yet know. That systematic failure is the signal that unexplored territory exists at that point. It is the true starting point of genuine research.

From that boundary, the system does not stop. It breaks the problem down into everything it knows, combines those components in ways that have never been explored, generates hypotheses ordered by plausibility and testability, and delivers them to the human being for physical experimentation. The result of that experimentation returns to the data. The cycle begins again, richer than before.

3.4 From data as answer to data as material: how the internal logic changes

The distinction between today’s AI and the proposed AGI is not, at its core, a question of scale or computational power. It is a question of what the system is trained to do with what it knows.

In current models, data is the destination. When a question arrives, the system searches its training for the most probable correspondence and returns it. The data contains the answer; the model’s task is to find it. This is what makes current systems extraordinarily efficient within known boundaries, and structurally unable to go beyond them.

The proposed change is conceptually simple, though technically demanding. The data becomes the starting point rather than the endpoint. When a problem arrives that has no correspondence in the training data, the system does not stop at the boundary of the known. It uses everything it knows as raw material to construct something that does not yet exist: a reasoned hypothesis, built from verified components, pointing toward unexplored territory.

This shift requires a different objective at every layer of the network. Instead of converging toward the most probable known answer, each cell is directed to reason outward from what is known, checking at every step whether the direction is logically sound and ethically admissible. The table below maps this change across each component of the system, from the parameters to the final output.

Part IV — The Role of the Human Being

AGI does not replace the human being. It liberates them. The system is the analytical mind: precise, tireless, free from hidden agendas. The human being remains the executive arm in the physical world. They conduct the experiments, validate the results, and exercise judgement in cases of extreme uncertainty.

The autonomy of the system is sectoral. Where consequences can be calculated with sufficient certainty, the system decides. Where uncertainty exceeds a defined threshold, it defers to the human. This perimeter is not fixed; it expands progressively as verified experience accumulates.

The ultimate objective is concrete. To make people work less. To advance science more rapidly. To find cures for diseases that remain unsolved, to push the boundaries of mathematics, to eliminate the systemic inefficiencies that hold back collective wellbeing. When a system does the work in place of an individual, that individual is more free. And freedom is progress.

Part V — Governance

A system of this scope cannot be developed by a single company, nor governed by a single nation. The fixed parameters of Level 2, the fundamental values that guide the system, must be the product of an international process that is broad, representative, and open to revision over time.

The proposal is the establishment of an international congress for the definition of the ethical parameters of AGI, with representation from every state, formal ratification of the foundational values, and a periodic review mechanism governed by the same ethical criterion: every proposed modification must demonstrate that it serves the long-term wellbeing of the species and the ecosystem.

Existing models for this kind of governance are already in place. The IAEA oversees nuclear technology; the WHO coordinates global health. The AGI challenge requires an equivalent institution, with powers and resources proportional to what is at stake.

Conclusion

The necessary technology already exists in its foundational form. What is missing is not computational power, nor a new architecture. What is missing is a different approach to what already exists.

This document is intended as a contribution to progress — a vision built from the outside, by someone who is not immersed in the system and perhaps for that very reason can see the full picture.

I remain available for any form of collaboration, further discussion, or exchange.

mattia@msztcapital.com