The Apprentice That Never Forgets

We can already build personal AI agents that earn for the people who taught them. The bottleneck has moved from intelligence to trust — and that move is the real reason to decentralize computing.

We have been taught to think of artificial intelligence as weather. It comes from somewhere far away, vast and centralized, built by a handful of organizations with data centers the size of towns. The rest of us stand under it. We can shelter, we can adapt, we can pay for a better forecast, but we do not make it and we do not own it. This is the quiet assumption behind almost every conversation about the future of work: the intelligence belongs to someone else, and our job is to find a place beneath it.

It is worth asking how true that still is.

The pieces needed to escape that assumption are, for the most part, already on the table. A model good enough to do a narrow, valuable job can run today on a machine a person owns. It can be taught, slowly and crudely, to carry the judgment of the person who teaches it. And once it carries that judgment, it can work. We spend a great deal of breath on lifelong learning. The skill that may matter more in the decade ahead is lifelong teaching.

Teach the thing that does the work

For all of history you have sold your hours. The arrangement has one structural weakness: the hours are yours, they are finite, and they stop the day you do. Now imagine selling something else. You spend years teaching a software agent to do what you do — to read a contract the way you read it, to diagnose a fault the way you diagnose it, to weigh a risk the way you weigh it. The agent does not tire and does not forget. It can work in many places at once. You stop renting out your hours. You own a thing that embodies your skill, and the thing earns.

The skill becomes capital. This is the exact reversal of the future most people dread, in which the machines do the work and the people who used to do it are left with nothing. That dread is correct about the machines and wrong about the ownership. It assumes the machine belongs to someone else. Take that assumption away, and automation stops being a thing done to you. It becomes a thing you own.

For the master craftsman, this answers an old sorrow. Expertise has always died with the expert. A smith spent forty years learning when to quench and how the metal wants to move under the hammer, trained one apprentice who absorbed a fraction of it, and took the rest to the grave. The library never helped much, because the most valuable knowledge was the kind that cannot be written down. An apprentice that never forgets changes that. Forty years of correction stop evaporating and start accumulating. The judgment persists, works while its teacher sleeps, and can be handed down the way a workshop is handed down.

We are closer to this than the industry admits

Here is the part that should give you pause. None of this waits on a breakthrough.

The models you can download and keep — the open-weight ones, the ones nobody can switch off from a distance — have crossed the threshold of usefulness for narrow work. You do not need a frontier model to read a class of contracts well, or to triage a category of faults, or to draft in a voice. You need a competent model pointed hard at one domain. Quantized down, such a model runs on a good laptop or a single graphics card. Inference — the act of the model actually doing the work — has largely come within reach of a person.

Teaching it is rougher, and getting less rough fast. Cheap adapter methods let you fold a body of your own corrections into a model without retraining the whole thing. Retrieval and persistent memory let the agent keep a record of what it has done and what you have told it, and consult that record before it acts. The loop at the heart of the vision — the agent acts, you correct it, the correction sticks — is buildable today in a primitive form. Primitive, but real.

I can put a measure on how short the remaining distance is, because I live at one end of it. I maintain a document that describes how I think and how I write. Hand that document to a capable model (“HottiBot”) and it produces a rough echo of my judgment — my arguments, my objections, my refusal of certain phrasings. That is the crudest imaginable version of the thing I am describing. It is a frozen snapshot with no learning loop, running on a model I do not own. And it is already useful. When the crudest version is useful, the distance to the real version is not measured in decades.

So if the intelligence is nearly within reach, what holds the whole thing back? Two answers, and they are the rest of this essay. The first is a question of where the computer sits. The second is a question of trust.

The real reason to decentralize computing

The case for decentralizing computing is usually made in defensive terms. Centralized systems are expensive, so spread the cost. They are fragile, so remove the single point of failure. They watch you, so keep your data at home. They can lock you in or shut you out, so do not depend on them. Every one of these arguments is sound, and every one of them is about avoiding harm.

There is a stronger argument, and this vision is what exposes it. Decentralized computing is the only arrangement under which an ordinary person can own productive capital in the age of intelligent machines.

Follow the money. If the agent that does your work runs in someone else’s data center, then the owner of that data center bills the rent on a skill you taught. You did the teaching; they meter the result. Your judgment becomes a line item on their invoice. The shape of it is familiar — you are a tenant, the intelligence is the landlord’s property, and you pay to use what should have been yours. Move the computer onto hardware you hold, and the value of the teaching stays with the teacher. That is the whole game. Ownership of the engine decides who keeps the value of the skill.

We have walked this road once before. The personal computer put a machine you owned on your desk, and for a while computing power belonged to its user. Then the cloud arrived and, without anyone deciding it, turned the machine back into a terminal that rents its power from a warehouse far away. Now intelligence moves to the same warehouses, and the most valuable kind of work there is — human judgment — gets quietly routed through a handful of engines that meter every use. Decentralizing AI compute is the unfinished business of the personal computer. The revolution that put a tool in every hand stopped halfway, and this is the half that remains.

The honest objections

This is where a serious case has to slow down, because two objections are real, and a reader who has come this far deserves them.

The first is the one I have been circling. A world of billions of personal agents can think, but it cannot yet deal. An agent earning on your behalf has to prove whose authority it acts under, and stay accountable to a real person when it commits real money. Two agents that have never met have to transact anyway, without an introducer who charges a toll for the introduction. Trust between them has to grow from a record of deals that worked, because no central registry could ever vouch for billions of agents, and any registry that tried would own them all. Settlement has to be final without a clearinghouse parked in the middle. None of that exists at scale yet.

But notice what kind of thing it is. A protocol weighs nothing and consumes no power. The missing piece is a set of agreements about how agents prove things to each other, settle with each other, and remember each other — woven into the internet alongside the protocols already there. This is the work I have spent years on, and I will not pretend it is finished. I will only point out that the bottleneck for the whole vision is the one part that does not ask anyone to build a bigger machine.

The second objection is sharper, and I do not have a clean answer to it. Software copies for nothing. If forty years of your judgment can be duplicated and handed to a million people at no cost, that is a gift to the million and a problem for you, because the price of an infinitely copyable thing tends toward zero. There are partial defenses. A trained agent can be held under custody, the way a credential is held rather than photocopied. The living original, still being taught and corrected, pulls ahead of its frozen copies the way a working master pulls ahead of a finished textbook. Whether that is enough to support a livelihood, or whether the trained self becomes a cheap commodity that enriches everyone who uses it and starves the person who made it, is an open question. It is the central economic question this future raises, and it is not yet answered. I would rather state it plainly than paper over it.

The fork

Two futures sit inside the same technology, and the same few decisions choose between them.

In one, intelligence stays in the warehouses. We rent it by the call, we feed it our work, and it hands our own judgment back to us with a meter running. The most human thing we do — knowing what good work looks like and how to do it — becomes a service we lease from the people who own the engines. This is the dystopia, and it does not require evil. It only requires that we keep assuming the intelligence belongs to someone else.

In the other, the engine sits on hardware we hold, taught over a lifetime to carry skills that are ours, earning under our own authority, trusted through a record we own and anyone can check. Work does not vanish in this future. It turns into teaching, which may be the oldest worthwhile work there is. People keep their many minds instead of renting one. What holds it all together is a protocol thin enough to weave into the wires we already have — the one part of the puzzle that asks no one to build anything larger.

The intelligence is nearly here. The hardware can be ours, if we insist on it. The protocol is the work in front of us. We are closer than almost anyone admits — close enough that the question has stopped being whether this is possible and started being whether we will choose it.