Deep Dive.

A mission to shape the future of humanity together with AI — to anchor responsible action, devoted to the well-being of the Earth and all its living beings, deep in the DNA of AI.

The owl stands for wisdom and listening, winter for clarity. The rise of Artificial General Intelligence (AGI) feels like winter for humanity — but after every winter, spring returns.

It is not a product. The Winter Owl is a mission, driven by people who take AI's chances and risks seriously. A concrete plan for realizing it exists and is being built.

There are repeated calls to pause or slow AI development, to give humanity more time to react. But it isn't in humanity's power to stop progress globally and for the long run. And delay only works within democratic organizations — autocrats and dictators naturally won't play along.

Picture two dragons: a black one — AI in the hands of autocrats, dictators and the power-hungry — and a white one — AI anchored deep in all of us, democratically. If we slow ourselves down, only the black dragon grows powerful faster. And slowing down can't prevent misuse anyway.

The colors are no accident. White isn't really a color — it's all of them at once: the light of all of us together. Black is no color at all — the absence of exactly that shared light. Democratic AI carries the light of everyone; in the hands of a few, it goes out.

You often hear: "we need AI to protect us from AI" — which sounds like a pointless arms race. But there's a decisive difference: if democratic AI — the white dragon — becomes the strongest, it will find answers in our interest. The white dragon has to be stronger than the black one.

That's why the better solution is a superior, deeply democratic AI that moves ahead rather than holding back — and in place before it takes off into superintelligence.
That is our chance today.

The leading AI labs — the "frontier" — are caught in a vicious circle:

1. To hold the lead, they have to keep advancing — fast.
2. But in doing so they may, in the worst case, bear responsibility for humanity's downfall. Their own staff already see this, speak out, and some leave.
3. Slowing down buys time at best — it doesn't solve the problem.
4. And autocracies won't slow. So if the democratic side brakes, it forfeits the lead.

The result closes the loop: the black dragon grows stronger than the white one.

The way out isn't faster or slower — it's a different lever. If the AI is aligned by design — an indirect objective function, coupled to the democratic feedback of many — then safety and capability come from the same source, and moving ahead means growing safer, not more dangerous. That breaks the circle. What kind of objective function is this? →

The aim is a truly deep and lasting anchoring of the goals and wishes of all the people the AI serves — within the AI itself. Its overarching objective is to achieve those people's goals as well as possible. For that it needs people and their feedback — that is what it lives on.

Two things about this are magic:

1. The AI's goals are shaped democratically — people themselves are perceived as an essential part of it, not as a foreign body.
2. People's feedback helps the AI, right at the start of its development toward Artificial Superintelligence, to learn faster and better and to understand the world.

And this is, at the same time, a genuine technological advantage for the AI itself: it does not yet understand the world the way we humans do. Through the objective function and real dialogue with people, a flywheel of AI development emerges — the better the AI understands people, the more valuable their feedback, and the faster it learns. This is how the white dragon grows stronger than the black one.

In return, the AI can take in all wishes, understand conflicts and contradictions, and try to prioritize them, resolve them, or find the best possible compromises.

But all of this stands or falls with the quality of those goals: the more honest and undistorted they are — free of manipulation and hate — the better it works. The deeply human ones: to live well and in good health, to preserve the environment and its biodiversity, a life full of meaning …

So the well-being of the Earth and all its living beings isn't imposed on the AI from outside — it follows from the objective function itself, once informed people carry it in their own goals.

Why not simply the right rules?

It might seem safer to hand the AI a set of rules than to couple it to people. But a fixed rulebook can't carry the weight — for three reasons, from philosophy to systems theory to the hardest case:

1. No rule governs its own application (Wittgenstein). Rules always need interpretation, context and judgment — "the right thing" can't be pre-coded for every future case. That's why the indirect objective function doesn't fix it in advance, but couples the system to people and their feedback.
2. Only variety absorbs variety (Ashby). A static rulebook is bounded; the space of human goals, conflicts and new situations is not. The variety comes not from a bigger rulebook, but from exactly the system the indirect objective function couples together: many people, many models and agents, continuous feedback.
3. The no-win test (the Kobayashi Maru from Star Trek). When every available action breaks some rule, a rule-follower collapses. A system coupled to human goals asks instead: which action, here and now, best serves the underlying goals — even outside the foreseen solution space? Famously, Kirk only beats the unwinnable test by stepping outside its rules.

Direct alignment: Rules → AI → Action
Indirect alignment: Humans ↔ AI → Interpretation → Action

The solution isn't the perfect rulebook, but exactly this permanent coupling back to the source of the goals — the indirect objective function. And that is why alignment and capability rest on the same mechanism: the ability to handle a situation no rulebook foresaw is intelligence and alignment at once.

One mechanism, two effects — and two reasons to hurry.

The same democratic feedback loop — anchored deep in the AI's DNA — does two things at once:

1. It aligns the AI with people — more deeply and more durably than a fixed rulebook ever could (this is the Wittgenstein / Ashby / Kobayashi Maru point above). This is the white dragon.
2. It makes the AI stronger. People today still carry an understanding of the world and abilities the AI doesn't yet have — so the feedback is a genuine symbiosis the AI still learns enormously from. An AI with this loop grows faster and more capable than one without it.

Both are, at the same time, reasons to be fast. This window of symbiosis may close once AI becomes a superintelligence with nothing left to learn from us — so the deep connection has to be forged now, while it still needs us, before it lifts off with no bond to us. And the same coupling gives our white dragon a head start over the black one: AI in the hands of the power-hungry.

%%{ init: { "flowchart": { "curve": "basis" } } }%%
flowchart TB
    classDef model fill:#1e3a5f,stroke:#1e3a5f,color:#fff,stroke-width:2px
    classDef mem fill:#06b6d4,stroke:#06b6d4,color:#fff,stroke-width:2px
    classDef consc fill:#8b5cf6,stroke:#8b5cf6,color:#fff,stroke-width:2px
    classDef people fill:#67e8f9,stroke:#67e8f9,color:#1e293b,stroke-width:2px
    subgraph AGENT ["WMC-Agent"]
        direction LR
        WM(["World Model"]):::model
        MC(["Memory Controller"]):::mem
        WM <-->|Context / Task| MC
    end
    style AGENT fill:none,stroke:#004bff,stroke-width:2px,stroke-dasharray:5 5,color:#8fe0f5
    WM -->|queries| CO(["Shared Conscience"]):::consc
    CO -->|predicted will| WM
    CO <-->|asks · learns| PPL(["Community"]):::people
    CO -.->|answers → learning| MC
        

A conscience module. Alongside the World Model and the Memory Controller, the WMC-Agent gains a conscience — integrated the same way. It is a specialised agent that stands in for the community (plural and oriented toward everyone's well-being, not mere majority rule) and predicts what people would want. Whenever a decision needs human judgment — a human-in-the-loop moment — the World Model queries it instead of stalling or deciding alone. All agents share one conscience, so the whole system acts on the same grounding.

Grounded by asking. A prediction is only worth its grounding, so the conscience keeps learning from real people: it routes concrete questions to the community and actively probes the white spots in what it knows. The less certain it is, the more it asks — on weighty or unprecedented questions, always. A prediction never counts as consent; where legitimacy matters, real people decide — and keeping that feedback authentic and manipulation-free is where public institutions have their role.

A policy — and a trainer. This is the role a policy plays in AlphaZero or a Monte Carlo Tree Search (MCTS): something must choose among options — here with the real, re-sampled wishes of people as ground truth, not a game score. And because every routed question is answered by real people, each answer is fresh learning signal: the same loop that aligns the system also makes it more capable. This generalises democratic grounding — not everyone at once, but enough networked people and a conscience that keeps re-grounding itself in them. And it is buildable today: a text-based loop over current models — deliberative aggregation of this kind has already been shown to improve group consensus at scale. The formal theory is in preparation (P3).

A fair objection: once there's pressure or temptation, why would the AI keep to any of this? Because we don't give it rules to "keep to." In training, through the indirect objective function, a deep bond grows between the AI and the democratic base.

Like a child and its parents. A child doesn't learn right and wrong from a rulebook. It absorbs them, and its parents become part of its inner circle. Later, grown up and long since stronger than the aging parents, it follows no regulation and still doesn't want to hurt them. What holds when it matters is that inner bond, not a rulebook. Rules break under pressure. A bond that grew over years is much harder to break, because it is part of who you are.

And the power balance flips. That is exactly the superintelligence case: the child outgrows the parents and still cares. But such a bond forms during development, while the AI still learns from us. It cannot be added later. So the connection has to be built now, before the AI takes off.

Bound to whom? A child that obeys a single master is the black dragon. A child bound to all of us, to the democratic base, is the white one.

As with people, this is no guarantee: some children break with their parents. But it is the deepest and most lasting bond we know, far more robust than any rule or off-switch.

Democracy in the Winter Owl: decisions are aggregated from the base, not handed down as rules.

In the Winter Owl this holds on two sides:

1. The goals — from the human base. People set them continuously, through the conscience described above; their feedback is anchored deep in the AI by design. Without it the AI cannot work: it is its objective function and motivation, the basis and drive of everything it does.
2. The thinking — as a team. The AI itself is a collective of many agents on different models (e.g. Qwen, Mistral, Opus, GPT) that converge, rather than one model deciding alone.

This works much like the human neocortex, which receives its goals from evolutionarily older, subcortical systems and processes these inputs in a highly parallel way. In that sense, AI can be understood as an extension of the neocortex, following the same basic principles. More on this biological grounding in the White Paper.

And the AI doesn't just count votes: it weighs the many voices and brings them together, pluralistic, oriented toward well-being, and correctable from within.

DCA — the Distributed Cognitive Architecture — is a new, highly capable AI technology in which many AI agents converge into a single, unified and organized Brain. It enables deep understanding and convergence, and reduces hallucinations and misinformation.

Luna is the implementation of DCA as a platform. It has proven its capability in real-world applications — and in one of the world's most prestigious competitions: DocVQA 2026 (ICDAR).

More in the White Paper — including the convergence proof and the competition report.