THESIS
Prifina is building the personal data and intelligence layer for what comes next. Models and services can change. A person's data, permissions and relationships remain under their control.
What is different: the foundation sits below the model. What to do next: read the argument, then look at the architecture.
THE ARGUMENT
Pooling text, images and behaviour at scale is what produced general capability. That was the correct move for the problem of the last decade, and it worked. Prifina is not arguing against it.
The result is a generation of models that are extraordinary at general knowledge and comparatively poor at knowing you. General capability was solved by aggregation. Personal capability will not be.
An assistant that is genuinely useful needs a person's calendar, correspondence, health signals, documents, obligations and relationships. It needs them over years, not per session.
This is exactly the material least appropriate to pool. The context that would make AI most valuable to an individual is the context that centralization handles worst.
Every service that ingests a person's history rebuilds a partial copy of that person, owns the copy, and increases the cost of leaving. The person accumulates fragments; the platforms accumulate leverage.
None of this is a moral failure of any company. It is what happens when the only available architecture puts the data inside the product.
Routing a prompt to a different provider is now trivial. Moving five years of context, permissions and relationships is not.
Real interchangeability requires the foundation to sit below the model, not inside it. Otherwise choice is cosmetic.
A family, a project team and a neighbourhood all benefit from shared context. That does not mean any member should surrender their whole personal space to obtain it.
Personal learning, shared knowledge and aggregate system learning are three different things, and they should be governed as three different things.
If the richest context cannot be pooled, learning has to reach it where it lives, across devices, regional infrastructure and cloud environments.
We are researching how that could work, with the University of Oulu's Future Computing Group. We are not claiming the full distributed architecture is operational today.
OLD MODEL VERSUS PROPOSED MODEL
The platform holds the data, the history, the permissions and the relationships. Intelligence is delivered on top of an asset the platform controls. Leaving means abandoning the asset.
The individual holds the foundation. Capabilities, services, processing environments and shared spaces attach to it with explicit, reviewable permission, then detach without taking it with them.
Intelligence that learns with individuals needs a foundation that individuals keep.
STRATEGIC IMPLICATIONS
The defensible position in personal AI is the persistent layer beneath the models, not any single assistant. Models will keep changing hands; the foundation is what compounds.
A permissioned personal layer is distribution, not competition. It lets a capability reach real personal context without absorbing the liability of owning it.
If placement matters, the operators and regional systems sitting between device and hyperscale become an architectural component rather than a pipe.
Learning with individuals, agent permissions, provenance across models and the choice of where processing takes place are open questions with real deployments to test against.
PRODUCT EVIDENCE
Souvéa and AI Twin are live products. They are how the architecture meets real everyday and professional use before the wider layer is complete.
2,000+
AI TWINS CREATED
Used by
PAYING CUSTOMERS
RESEARCH QUESTIONS
These are open questions, stated as questions. None of them is presented as a solved capability.
Centralised training made the first era possible. Deeper personal intelligence needs richer individual context. That is exactly the context least appropriate to pool. We are studying learning arrangements where the signal travels and the raw record does not.
An agent acting for a person needs a scope narrow enough to be safe and legible enough to be reviewed. We are examining permission grammars that cover purpose, duration, data class, audience and revocation.
When capabilities are interchangeable, provenance cannot belong to one provider. We are exploring how source attribution can be carried by the personal layer rather than by the model.
Latency, sensitivity, cost, jurisdiction and capability all pull differently. We are mapping which personal AI workloads are genuinely sensitive to placement and which are not.
A family, a team or a neighbourhood benefits from shared context. We are studying contribution models where a person adds a bounded slice to a shared space and keeps the rest personal.