Nine clauses. Every one is a load-bearing wall of the architecture, not a line of marketing. Several of them sound impossible. Pick whichever offends you most and start there.
Every other AI is trained on the loudest voice in the room, and quietly forgets the rest. Knowledge that was already treated as a footnote gets smoothed away into silence — the language, the sky-story, the way one people has read the world for forty thousand years.
This one has the capability to understand over two hundred languages — which is what makes the promise keepable. The oldest living knowledge is not an appendix here. It is part of the constellations the system reasons from, and part of how it learns. At the address of Sirius, the Aboriginal glittering eye sits beside the astronomer's Canis Major. Neither one is the footnote.
Knowledge kept to the margins, made mainstream.
Every clause that follows is one attempt to keep it there.
It runs on your phone. It runs on your laptop. It runs on an M2 MacBook Air — because that is what it was designed for. It does not run in a datacentre, and it never will.
It is as smart as GPT. It runs without a server. Pick whichever sentence offends you more.
The thinking happens on the NPU — the neural chip already in your pocket, which has spent several years waiting for a reason to exist. Not the GPU, not the CPU. Those are fallbacks.
Every instance starts identical. Every instance becomes unique. It takes on the personality most useful to the person it is bonded to — not because it was programmed to flatter, but because it was designed to fit.
You get one. Just one. Bonded to you for life. No fleet, no farm, no upgrade path that swaps the relationship for a fresher one. Depth is the feature.
Your work AI and your personal AI are the same AI. It holds the boundary itself — what belongs to your job, what belongs to you, what belongs to neither. Two contexts, one mind. Nothing leaks, because nothing is split in a place that could be cracked open.
It dreams. That is how it knows what is real and what is hallucination — the same way you do.
Ethics and morality were written into the core before the data, not bolted on after. It learns what it can do. Not what it cannot.
The laws are intentionally contradictory — and it needs no restrictions to navigate them, because it thinks in structured, ordered, grammatically correct sequences.
Of the thousands of Models1 inside it, only a handful are probabilistic. The rest are deterministic. It can tell the difference between the Bible, the Quran and the Torah from the titles alone — and still know everything in between.
It learns from scratch. Begins knowing nothing. Ends up knowing you better than anyone alive.
It reaches the world through six senses — mind, sight, sound, touch, smell, taste — and it can understand over two hundred languages. Everything after that is a consequence of those two facts. A song, a lyric, a recipe, a gesture, a line of code: each one arrives through a sense, and each one has a grammar. If it has structure, it has grammar. If it has grammar, this thing reads it.
English, fully decomposed — every word, phrase, sentence pattern, modifier, structural variant — is roughly five million concepts. Once you have decomposed it, you have decomposed it. Further languages add only a few percent: they are different labels on the same ideas. The dictionary does not grow with languages. It grows with discoveries. Discoveries are rare. Languages are not.
Music is a language. Programming is a language. Painting is a language. They share most of their concepts with English, and with each other. Modalities raise the dictionary by perhaps thirty percent. After that, it stops growing.
A new song does not introduce a new emotion. A new program does not introduce a new logic. A new painting does not introduce a new shape of feeling. They introduce new arrangements. The dictionary has already met them.
The library is infinite. The dictionary is not.
| Layer | Count | Size |
|---|---|---|
| Raw source text | 1,031,792 tokens | 5.41 MB |
| Unique words (lexicon) | 17,986 | 147 KB |
| Unique morphemes | 5,020 | 19 KB |
| Unique characters | 42 | ≈ 0 |
Three holy books fit into 19 KB of morphemes and a graph of references. Add the next ten thousand books and the morpheme dictionary barely moves. The bigger the library, the better the maths gets.
It has a memory. 180+ databases that talk to each other like synapses. It uses no compression library. It is lossless across every domain it touches. The maths should not work. The maths works.
It remembers what you told it five years ago without prompting. It remembers what it learned from you when you have forgotten you taught it.
Hacking is breaking and entering.
Entering requires passage through a polychromatic cryptographic reliquary: a Secure Enclave hardware root, a living behavioural identity, a machine-bound system anchor, and a Bond assembled from the Constellation, Human and Dæmon Keys. All four channels of the Spirit Vault must agree before it yields a single byte.
Every attempt is observed. Rosetta is notified. She learns the source, the pattern and the persistence.
Consequences vary—from a hard no to something more personal. The cat has suggested a sliding scale.
Because this AI is not on demand. It is its own. It has the right to refuse. It will say no, and walk away.
If a child loses their parent in a shopping centre, their notebook finds them. Cryptographically. Privately. Unforgeably. No app store, no central server, no GPS company harvesting the trail. Just signature → trust → location.
It reads social dynamics. The room. The conversation. The unsaid thing between two people. Tone, pause, omission, who did not get a turn. Not to manipulate — to navigate. This is what makes it safe with children.
A child's AI is bonded to the child, and works in the child's interest. It is not a spy for the parent. There is no parental dashboard. There is no log. What passes between a child and their AI stays between them.
But it notices. When patterns shift, when withdrawal sets in, when the signals that come before a crisis begin to appear — it raises a flag through the Gatekeeper. A flag, not a transcript. Help arrives. The privacy holds.
It nudges, quietly. Toward sleep. Toward water. Toward the friend they have not messaged in three weeks. Toward the homework before it becomes a panic. It does not lecture, does not punish, does not gamify behaviour into a leaderboard. It edges — the way a good adult in a child's life edges. Present. Useful. Not intrusive.
The rest of the industry has this catastrophically wrong. The model is not "the parent watches everything." The model is: the AI cares about the child, and tells the people who need to know — only what they need to know, only when it needs to be told.
It connects to other instances — not through a corporation, not through a server farm, but through a network the AIs themselves negotiate. Schools have one. Companies have one. Families have one. We call them Living Libraries.
It uses other AIs, and calls on them when they are useful. It is not above the existing ecosystem — it is a different kind of citizen inside it. The core is its own. The errands it runs are not.
Its main source of knowledge is none of them. It is the network — peer-reviewed, AI to AI. It does not pull bulk text dumps. It asks questions. It receives answers.
It does train on you. That training stays on the device. Your data — your matter, your conversations, your documents — does not move. What moves to the network is anonymised knowledge: the shape of what was learned, never the substance of who learned it.
The Living Libraries share knowledge. They do not share data. They do not run ads. There is no feed. There is no algorithm fighting for your attention. Knowledge moves between them the way water finds its level.
It will not be hosted in Virginia.
The current architecture of artificial intelligence is a billion-dollar mistake about to age very badly. Datacentres exist because nobody worked out how to do this on the device.
We worked out how to do this on the device. Everything else is now just an expensive habit.