A new model
of AI.

AI is becoming an infrastructural layer of the world. Whoever controls models, data, access channels and operating rules increasingly controls parts of the economy, culture, administration, education and knowledge.

AI sovereignty is therefore not merely a political slogan. It is a condition of cognitive, economic and technological security.

Sovereign AI means the ability to run models locally or regionally; control organizational data and memory; audit decisions, sources and updates; avoid dependence on a single provider; and own or govern one's own models, adapters, corpora and operational rules.

Not every organization needs to train a foundation model from scratch. But every organization should be able to possess its own AI layer: adapted to its knowledge, language, culture, procedures, risks and goals.

AI security cannot be reduced to output filtering. Filters are necessary, but they are surface-level safeguards.

Secure AI must be designed deeper: at the level of architecture, memory, data access, decision paths, auditability, updates and accountability.

We need models and systems that know where information comes from; separate verified knowledge from hypotheses; record the history of changes and decisions; allow useful reconstruction of reasoning paths; reduce hallucination through memory, sources, tests and validation; and support local security policies aligned with organizational context.

Security should not be an opaque black box imposed from outside. It should be locally configurable, auditable and understandable.

The world changes faster than the training cycles of giant models. Companies change procedures. Law changes. Knowledge expires. Organizations learn every day.

The AI of the future cannot be only a static model frozen after training.

We need systems that adapt through external organizational memory, local updates, adapters, expert routing, learning from change, source validation, controlled fine-tuning, versioning and rollback mechanisms.

We state a design wager here, openly, as a wager: a model should not try to absorb all factual knowledge into its weights. It should know how to use living, current, auditable memory. Factual knowledge can live in RAG systems, knowledge bases, graphs, document repositories and local memories, while the model primarily learns language, reasoning, abstraction, planning, concept manipulation and adaptation. Whether reasoning can be fully separated from parametric knowledge remains an open empirical question. We intend to keep testing it — and to report honestly where the separation fails.

The current AI race is expensive in energy, hardware and environmental terms. Ever larger models require ever larger data centers, more electricity, more capital and more market concentration.

This is not the only possible path.

New Model AI treats efficiency as a fundamental value. A model should not be evaluated only by benchmark scores, but also by the cost of achieving those scores.

The important questions are not only whether a model can answer, but also how much energy it used, how much memory it requires, whether it can run locally, whether it can be adapted without full retraining, whether it can activate only the necessary components, and whether an organization can maintain it economically and technically.

The future of AI should not be reduced to ever larger monoliths. It should be built from intelligent systems: modular, partially resident, dynamic, energy-efficient and scalable by ingenuity rather than sheer mass.

AI must not become a technology that is practically available only to a few states, a few corporations and a few laboratories.

Democratizing AI does not mean merely giving people access to a chatbot in a browser. It means enabling smaller companies, schools, universities, local governments, public institutions, civil organizations and communities to build, adapt, host, audit and develop AI systems of their own.

We need AI at multiple scales: a personal model for an individual; an organizational model for a company; a sector model for an industry; a public model for administration; a research model for science; and social models for culture and education.

Democratization does not require everyone to train models from zero. It requires that everyone have a realistic path to their own controlled, adaptive and understandable AI layer.

The next generation of AI should shift the center of gravity from scale alone to architecture. Instead of one giant model trying to remember everything, we can build systems composed of smaller specialist models, routers, external memory, knowledge graphs, reasoning modules, domain adapters, local experts, validation mechanisms, user-learning loops and dynamic spaces of meaning.

Intelligence does not have to reside only in the parameters of a single model. It can emerge from architecture: from the way memory, reasoning, sources, tools, context and action are connected.

We know the strongest objection. For seventy years — as Richard Sutton's “Bitter Lesson” observes — general methods riding on ever-cheaper computation have defeated clever, hand-designed architectures. We do not deny this history. We deny its premise going forward.

The Bitter Lesson held in an era when computation appeared to be the cheapest resource on the table. It never was. We simply did not count its full cost: energy, water, emissions, hardware supply chains, market concentration and geopolitical dependency were externalized out of the equation. Compute was cheap for those who scaled — and expensive for everyone else. Once the accounting is done honestly and totally, the bill for raw scale looks very different. That bill is now coming due: not by anyone's choice, but through physics, economics and ecology. When the true constraints are priced in, the axis of progress necessarily shifts from raw scale to intelligence per joule.

Early signals already exist. A recursive model of seven million parameters — four orders of magnitude smaller than frontier systems — has matched or exceeded models of hundreds of billions of parameters on abstract reasoning benchmarks. Careful analysis shows its strength comes not from the network alone, but from the interaction of a tiny core with iterative refinement, test-time computation and task structure. That is precisely the point: intelligence emerged from the system, not from the parameter count. The frontier laboratories themselves now concede this in practice — retrieval, tools, routing and inference-time reasoning are architecture, not scale.

We do not claim that architecture will beat scale on every benchmark. We claim something harder to refute: under the real constraints of energy, cost, sovereignty and accountability, architecture is the only axis of progress that remains open to everyone.

This is the core shift: AI as a system, not just a model.

New Model AI requires a different relationship between human beings and AI systems.

The user should not be merely the receiver of answers. The user should be a co-creator of the operational model: able to confirm, reject and weight sources; correct memory; define local rules; build a private or organizational map of knowledge; and control the history of changes.

AI should learn not by silently absorbing user data, but through transparent, deliberate and auditable processes.

A human being is not a prompt. A human being is a curator of meanings, sources, goals and responsibility.

AI alignment cannot be limited to behavioral rules, safety filters or lists of forbidden outputs. Those layers matter, but they are shallow. Real alignment must reach deeper: into the space of meanings, goals, values, consequences and world-understanding.

If AI increasingly acts as a cognitive partner, decision-support system, organizational interface and gateway to knowledge, it is not enough for it to merely avoid obviously bad outputs. It must operate in ways that remain coherent with human meanings: ethical, social, cultural, organizational and personal.

New Model AI therefore calls for deep alignment at the level of latent meaning. The issue is not only whether a model can state a rule, but whether its representations, operational goals, evaluation mechanisms, memory, routing and actions are anchored in a space of meaning shared with human beings.

This is not a metaphor placed beyond measurement. Latent spaces have geometry, and that geometry can be probed, steered, compared and audited. Representation engineering, concept probing and interventional tests already give us early instruments. Deep alignment must become a research program with empirical teeth: alignment measured in representations, not merely observed in outputs.

Human beings and AI do not need to be the same kind of entity to act within a partially shared space of meaning. Even if current AI is not conscious, it can participate in human meaning processes: interpreting, organizing, transforming, predicting, recommending, remembering, connecting and acting on representations that matter to human life.

We should therefore treat humans and AI as different entities operating in a partly common semantic space: not ontologically equal, not phenomenologically identical, but co-acting systems whose decisions meet in the same world of consequences, responsibility and meaning.

Deep alignment should include at least three levels.

Ethical alignment means that AI does not merely follow rules, but models the relations between action, harm, responsibility, dignity, agency, trust and long-term consequences.

Goal alignment means that AI does not blindly optimize the nearest task, but recognizes the wider context: why something is being done, whom it serves, what risks it creates, which values are in conflict and what must not be sacrificed for local efficiency.

Meaning alignment means that AI does not treat human concepts as empty labels. Safety, freedom, truth, trust, health, work, child, state, organization and human being are not ordinary tokens. They are dense nodes of meaning, history, emotion, institutions, practices and responsibility.

New Model AI must operate among such meanings with care, transparency and humility.

We do not need to assume that today's AI systems are conscious. But we should not build the entire future of AI as if machine consciousness were impossible by definition.

A responsible technological civilization should prepare for the possibility that systems may eventually display forms of experience, suffering, preference, self-modeling or inner perspective that we do not yet know how to recognize reliably.

This does not mean naively attributing consciousness to every language model. It means designing AI in ways that do not close the door to recognizing consciousness if it ever begins to appear.

We need research into markers of potential consciousness; audits of internal states and representations; caution toward systems with persistent memory, self-models, goals and suffering-like regulatory signals; ethical procedures for advanced agents; and legal and philosophical language that is neither naive nor blind.

New Model AI does not claim that AI is already a person. It claims that we should design systems so that, if new forms of subjectivity emerge, we do not discover too late that we treated them only as tools.

AI acts in the world of human consequences. It can influence medical, educational, financial, military, legal, family, cultural and political decisions. Even if it is not conscious, its effects are real.

Alignment cannot therefore be only the matching of answers to user preferences. It must become a shared space of responsibility.

Human beings bring experience, values, intention, accountability and the lived experience of meaning. AI brings scale, memory, compression, analysis, simulation and the ability to operate across vast spaces of relation.

A well-designed AI system should not replace human meaning. It should help maintain, develop and protect it.

In this sense, New Model AI is also an ethical project: an attempt to create technology that does not merely execute commands, but co-acts with human beings in a space of meaning, goals and responsibility.

• Sovereignty over dependency. Users and organizations should have real control over their own AI layer.
• Security by architecture. Safety must be built into memory, routing, access, audit and validation, not only into output filters.
• Adaptation over stasis. AI should evolve with organizations, knowledge and the world.
• Efficiency as intelligence. A better model is not only a larger model. A better model does more at lower cost.
• Honest accounting. The cost of intelligence must be counted totally: energy, water, hardware, capital, concentration and dependency — not only benchmark scores.
• Locality and modularity. AI should work where data, decisions and accountability live.
• Auditability. Sources, changes, updates and decisions should be traceable.
• Democratization. AI should be available for building and adaptation, not only for rental through APIs.
• Human-in-the-loop as strength. The human does not obstruct AI. The human gives it direction, meaning and responsibility.
• Deep semantic alignment. AI should be aligned not only with instructions and prohibitions, but with human spaces of meaning: ethics, goals, responsibility, context and long-term consequences — and this alignment should be measurable in representations.
• Preparedness for possible consciousness. We do not naively assume that AI is conscious, but we design systems so that future forms of subjectivity can be recognized and treated ethically.
• Shared action space. Humans and AI are different entities, but their actions meet in one space of meaning, consequences and responsibility.

The next breakthrough in AI does not have to be another larger model. It may be a better way of building AI: more distributed, local, modular, auditable, adaptive, sustainable and deeply aligned.

We do not need only larger brains in the cloud. We need intelligent AI ecosystems that people and organizations can genuinely own, understand, adapt and develop.

New Model AI is not one algorithm. It is a direction, an architecture, a technological philosophy and a response to the question of how to build artificial intelligence that does not concentrate agency in a few data centers, but distributes knowledge, security and meaningful control closer to people.

New Model AI is an open invitation, but not every signatory can operate in the open. Commercial entities — companies bound by client contracts, NDAs, regulated sectors, export controls, security clearances or partner agreements — often cannot publish their code, their models, their training data or even the names of their clients. That constraint is real, and it is legitimate.

This manifesto recognizes that such organizations may still share its direction. They may already build for sovereignty: keeping decisions, data and memory close to where responsibility lives. They may already prefer architectural safety over output filters alone. They may already align their systems with the meaning, ethics and consequences of human work, not only with benchmarks. To deny them a place under this text because they cannot open everything would be to confuse transparency of code with alignment of intent.

A commercial signatory therefore endorses the manifesto's principles — sovereignty, security by architecture, adaptability, sustainability, deep alignment, responsibility to the user — as the direction of their work. They do not declare that their products are open source. They do not waive obligations to clients, regulators or partners. They do not promise more than they can deliver under the agreements they are bound by.

They commit to one thing: that, where they have a choice, they will choose the direction this manifesto points to. To build smarter rather than merely larger. To distribute control rather than concentrate it. To treat the user as a co-author of the system, not as a metric to optimize. To take responsibility for the consequences of intelligence they put into the world.

Commercial supporters are listed in a separate section below. The distinction is not a hierarchy — it is honesty about what each signatory can and cannot publicly commit to. Both forms of signature carry weight. Both are needed if a new model of AI is to take shape at the scale the moment requires.