Who the Intelligence Becomes

AI sovereignty is usually discussed at the wrong altitude.

The conversation rises immediately to nations, data centers, chips, open weights, closed labs, federated training, export controls. That layer matters. If the assistant becomes the interface to knowledge, then the question of who trained it is no longer academic. It becomes political, cultural, commercial, and personal.

This is the force inside Yann LeCun's Tapestry argument. If AI assistants increasingly mediate the information diet, then a farmer in India or a philosopher in France should not be forced to see the world through a model trained around the assumptions of California or Shenzhen. The problem is not language coverage. It is representation. It is the quiet distortion that happens when the medium through which you ask questions does not contain your culture, your values, your local knowledge, your inherited distinctions.

Tapestry answers that problem at the foundation-model layer. It asks who owns the base intelligence.

That is the right question for nations, language communities, institutions, and cultures.

But there is another question below it. Once the base intelligence exists, who does it become?

An open model trained on the world is still not your archive. It is not your research record. It is not your company's memory. It is not your essays, your interviews, your decisions, your arguments, your refusals, your taste. Even if the substrate is sovereign, the agent may still be generic.

That is where Selflet begins.

A selflet is not a clone of a person. That framing is too crude and misses the actual object. The object is not the person. The object is the corpus. The archive. The knowledge base. The authored body of thought left behind in writing, speech, transcripts, research, images, decisions, and accumulated judgment.

A selflet is a generative fork of that corpus.

The word fork matters. A fork is not a copy. It is not a summary. It is not an imitation pasted onto a chatbot. A fork carries forward a line of development. It inherits constraints. It preserves enough of the source structure that new material can be generated without dissolving into average output.

This is why the model is not the product. The model is the substrate. The product is the transformation of owned source material into a bounded intelligence that can continue the work without betraying the source.

That continuation requires more than facts. Facts are the easy part. They can be retrieved, cited, chunked, embedded, refreshed. The harder problem is judgment. What belongs? What does not? Which analogy is native to the source and which one is merely plausible? How long should the answer be? When should the system refuse? When should it admit the corpus does not know? When should it stop?

That is taste.

Taste is not decoration. It is not tone. It is not style in the shallow sense. Taste is the pressure inside generation that makes one path feel necessary and the adjacent path feel false. Most AI lacks this pressure. It continues because it can continue. It fills the space because the space exists. It samples fluency without consequence.

A selflet has to do something narrower and harder. It has to generate with source-native constraint. It has to speak from inside the corpus without pretending to be the person. It has to extend the archive without becoming a ventriloquist act. It has to be useful without becoming generically helpful.

Generic helpfulness is the enemy.

It is the warm bath that dissolves everything distinctive. Every writer becomes clear and balanced. Every investor becomes a cautious consultant. Every founder becomes a product manager. Every thinker becomes a list of takeaways. The model smooths the edge because smoothing is what the generic assistant has been trained to do.

Selflet runs in the opposite direction. The point is not to make the corpus more palatable. The point is to preserve the edge that made it worth forking.

This is also why boundedness is not a limitation. It is the product.

A useful selflet should not answer everything. It should know what it has authority to answer. It should know when the corpus is strong, when retrieval is thin, when a question asks it to step outside its source, when a quote would be inappropriate, when a polished answer would be less truthful than a refusal. The boundary is not a defect. The boundary is what makes the agent trustworthy.

The industry still confuses breadth with intelligence. Bigger context. More tools. More modalities. More memory. More everything. But intelligence without taste becomes noise with confidence. The more it can say, the more important it becomes that it knows what not to say.

That is the Selflet Factory problem.

The Factory is not a pipeline for ingestion and fine-tuning. It is a discipline for converting corpus into governed generativity. Provenance matters because the source matters. Retrieval coverage matters because gaps become hallucinations. Evaluation matters because voice without gates becomes performance. Boundary enforcement matters because an agent that cannot refuse will eventually betray the corpus it claims to represent.

The Santiago experiment made this visible in miniature. Fine-tuning could move the model from talking about Santiago to speaking from Santiago's world. That was the breakthrough. But the failures were just as instructive. The model could overperform the voice. It could quote too closely. It could comply with formatting requests it should have resisted. It could break the frame. Voice capture alone was not enough. The fork needed taste, gates, and restraint.

That lesson scales.

A corporate selflet should not become a generic customer-service bot with brand flavoring. A research selflet should not turn into an internet-average analyst. An author selflet should not become a style mimic. A founder selflet should not become a motivational assistant. Each failure is the same failure: generativity without fidelity.

Tapestry and Selflet belong to the same larger movement, at different layers.

Tapestry is sovereignty of substrate. Selflet is sovereignty of representation.

Tapestry asks whether the base model should be owned by a small number of centralized companies or opened into a federated global substrate. Selflet asks what happens after that substrate exists. Whose corpus shapes the agent? Whose judgment constrains it? Whose taste gives it proportion? Whose boundaries stop it from becoming generic?

These questions are not competitive. They stack.

If open sovereign models win, Selflet becomes more useful, not less. Better open substrates make the Factory stronger. They reduce dependency on closed providers. They make local deployment realistic. They allow the value to move where it belongs: away from the base model and toward the owned corpus, the manufacturing process, the evaluation discipline, and the agent that emerges.

The open model may become the road. The selflet is the vehicle built for a particular terrain.

That terrain may be a writer's archive, a research firm's memory, a company's institutional knowledge, a teacher's body of lessons, a founder's operating philosophy, an estate's licensed corpus. In every case the problem is the same. The buyer does not need a generic intelligence that knows everything badly. The buyer needs a bounded intelligence that knows this body of work well and can extend it with taste.

This is personal AI sovereignty in its practical form.

Not the slogan version. The product version. The right to own the corpus. The right to decide what enters it. The right to inspect the lineage. The right to update the knowledge base. The right to constrain the agent. The right to deploy a representation that does not default back to the consensus machine.

The consensus machine will always be there. It will get better. It will get cheaper. It may become open. It may become part of the background infrastructure of the world.

Good. Let it.

The value will not disappear. It will move upward.

The model is not the product. The corpus is not even the product by itself. The product is the factory process that turns a corpus into a bounded, taste-preserving intelligence capable of generating new work without losing the source.

Tapestry asks who owns the base intelligence.

Selflet asks who the intelligence becomes.

That second question is where the product lives.