Manifesto
Intelligence from many small parts.
Advanced Nonlinear Technologies · London
I · The belief
Look closely at anything that thinks, and you find the same thing: not one great mind, but a multitude of small ones.
A brain is not a single vast computer. It is roughly eighty-six billion neurons, each almost trivially simple, none of them in charge, and out of their chatter comes you. An ant colony has no leader, yet a scent trail and a few local rules find the shortest path across a continent of dirt. A fern is not drawn, it is grown: one rule repeated until a whole plant stands where there was nothing. Fireflies in a mangrove flash together, though not one of them is counting the beat. Heart cells find a single rhythm. A thousand circadian clocks, each running slightly fast or slow, quietly agree on dawn. A flock of starlings turns as one body, with no bird in command. Even the frontier models everyone calls monoliths are, on the inside, colonies of experts with a router deciding who speaks.
Intelligence has always been a collective phenomenon. Simple parts, simple rules, repeated and coordinating: the oldest and most reliable recipe there is. The single giant mind is not the destination. It is the anomaly. We are building for the world where that truth wins.
II · What we refuse
Today, a business that wants to think with AI rents a brain that forgets it. Brilliant, and borrowed. It hallucinates, so every draft is edited by hand. It never learns your process, so you explain yourself again and again. It charges by the use, forever, and the bill grows with every person who touches it. Your corrections, the hard-won knowledge of how your work is actually done, leave with the session and are gone.
This is a strange bargain, and we refuse it. We refuse the premise that intelligence must be enormous, that it must be rented rather than owned, that it must forget you the moment the conversation ends. We refuse the belief that the only road forward is a bigger model in a bigger building drawing more power from the grid. There is another road, and it is the one nature already took.
III · What we build
We build Ferns: tiny specialist models, small enough to run anywhere, tuned to beat models ten times their size at the one job you need done.
A Fern does not know everything. That is the point. It is not built to win trivia; it is built to do your work, to read the invoice, answer the question, route the ticket, see what the camera sees, and to do it better, on your data, than a giant generalist ever could. It is small enough to live on a phone, a laptop, a watch. It is private by construction, because it runs where your data already is. And every time your team corrects it, it gets better at being yours.
Small, owned, and improving. That is a different kind of intelligence than the one on offer, and it belongs to the business, not to us and not to a lab.
IV · The names are the thesis
A Fern is a fractal: a few rules, repeated, that grow the whole structure. That is what our models are, small and nonlinear, grown rather than assembled.
Fernfly is a single Fern set loose through an API, for the developer who wants one specialist and nothing more.
Everyday Series is many Ferns orchestrated into Series, the sequences of steps behind your everyday work; a frond is itself a series of leaflets, and a workflow is a series of tasks. This is where your team teaches the models and the models take over the work.
And antelligent is the whole: a society of small agents, each carrying its Ferns, talking to one another the way ants talk in a colony, until intelligence emerges from coordination that no single part possesses. The names are a ladder, and every rung is the same principle at a larger scale.
V · How it compounds
Open weights and runtimes are everywhere, and we use them gladly. But a model on its own does not improve by sitting there. What compounds is the loop between the work and the model. Every organisation’s Fern is tuned on that organisation’s corrections. The more the work is done, the more corrections there are; the more corrections, the better the Fern; the better the Fern, the more work it takes on.
The data is theirs and does not transfer. The tuned model is theirs and does not transfer. Your competitor can copy everything you have built and still be years behind on the only thing that matters: the accumulated, proprietary knowledge of how a particular business actually works. Costs fall as the model moves on-device; quality rises as the corrections accumulate.
VI · Where we play
We build AI that is fast, accurate and cheap to run. Behind all three is one number: the energy it takes to get a piece of work done. We go after that number three times, at three depths.
Engineering · EverydaySeries
Waste above the model. Most token spend goes on work the step never needed: a large model asked to do a small job, context re-sent to agents that never read it. We put the practical fixes together with governance, so a company gets agents it owns, running each step on the smallest model that can do it, with an approver and a record of what happened.
Models · Fernfly
Waste inside the model. Most of what people want from AI is not reasoning; it is understanding what was asked and doing it. So we build intent-to-action models: small, specialised, under a second, and they return exactly the call you were after instead of a paragraph about it.
Substrate · Research
Waste in the hardware itself. Inside a chip today only a little of the power goes into the arithmetic; most of it goes into carrying the numbers to the place where the arithmetic happens. We are building a substrate that does not carry them at all: physical oscillators that settle into a solution rather than computing toward one. No GPU in the loop. We are aiming at two orders of magnitude less power.
Same waste, three places. The first two you can buy today. The third is why the first two look the way they do.
VII · The long game
There is a longer arc underneath all of this, and it follows from the last one. If most of the energy is spent moving numbers rather than computing with them, the deepest place to fix that is underneath the model. That is where we are digging. Today our Ferns run on standard silicon, and they run well. But we believe that to get small models that reason coherently, at very low power, without the fragility of today’s hardware, computation itself will have to change: from the clocked, digital arithmetic of the transformer to the physics of nonlinear, chaotic dynamics, where the answer emerges from a system settling rather than a processor stepping.
This is research and a very active one. In simulation, the architecture already writes coherent prose crushed to three-bit precision, where a standard transformer collapses into word-salad; it resists the runaway loops behind hallucination; it fine-tunes in minutes to bit-perfect structured output. But it also has real, open weaknesses. It is early. It is a ambitious. But it is doable. Science is there, engineering is what we are after. There is a plenty of room at the bottom.
many simple parts, coupled
the researchmany small models, each a specialist
the productmany small agents, talking
the platformVIII · How we proceed
We are disciplined about how we get there. We start narrow. Our first job is to turn scattered pilots into one repeatable motion, in the work where our advantage is a requirement and not a nice-to-have.
We sell the outcome, not a custom build. We land a single Series and let the flywheel expand it. We are realist. We keep eyes on the sky but ears on the ground. Our engineering of today is tuned to needs of today which bridges to our engineering and customer needs of tomorrow. Agentic optimization and small models today, substrate for tomorrow.
IX · The invitation
Almost everyone in this field is working on making models bigger. It is the obvious direction, it is well funded, and a great many clever people are already there. We are going the other way, and we would like more company.
It is harder than it sounds. Making something small enough to run on a phone is not a smaller version of making something large. It is a different problem, with different mathematics, and far fewer people have solved it. The substrate work may not come off at all. We think it will. We are not certain.
What we are certain about is the direction. Intelligence should belong to the people who use it. It should live on their machines, learn from their work, and stay there. Every part of that is an engineering problem, and every one of them is unsolved enough to be worth a career.
If you find the small side of this more interesting than the large side, if you would rather squeeze a model onto a device than add another billion parameters to one, you are the person we are looking for. It does not matter much whether you come as an engineer, a researcher, a customer or a critic. It matters that you think it is worth doing.
Write to us. Tell us what you would work on.
