The Collapsed Distance
On AI, the commodity trap, and the next thing the machine will try to take
---
For most of my working life there was one thing I could do that I quietly believed was mine. I could take a tangled situation and produce the clean version of it. The clear paragraph. The argument that holds. The analysis a person could actually act on. It took years to build, and it felt like the floor I stood on.
Then, on an ordinary afternoon, I watched a machine do it. Not badly. Not as a clumsy first attempt I could feel superior to. It did the thing, the actual thing, in a few seconds, and then offered me eleven more versions in case the first one was not to my taste.
I want to be precise about what I felt, because it was not fear and it was not excitement. It was a kind of vertigo. The quiet recognition that the floor I had been standing on was never a floor. It was a height the world had agreed to grant me, for exactly as long as the skill stayed rare. And the rarity had just ended.
This essay is not really about AI strategy, though it will look like that for a while. It is about what happens to a person, and to anyone who builds anything, when the thing they were proud of stops being scarce. Where the value goes when it leaves. Why this time it does not go where it has always gone before. And what, if anything, is still standing when the machine can climb into the very thing you were good at.
One sentence to hold the whole thing together. Everything that follows is a single movement: the ground rising under everyone at once, and the search for the one place a person can still stand.
---
What actually became free
Everyone says the same sentence now. AI will commoditize everything. People say it calmly, almost with relief, as if naming the wave were the same as surviving it. The sentence is half right, and the half it gets wrong is the half that matters.
Here is the half that is true. The competent answer is becoming abundant. The clean analysis, the serviceable draft, the response that sounds right, the kind of output that used to take a trained mind and a quiet afternoon, now arrives in seconds, from the same few providers everyone else is using. When the same capability reaches everyone through the same handful of doors, the capability stops being the thing you can charge for.
There is an old puzzle that sits underneath this. Water keeps us alive and costs almost nothing. Diamonds are useless and cost a fortune. Price has never tracked importance. It tracks scarcity, at the exact margin where someone reaches for the thing. For a few hundred years, a sharp mind producing clear work sat in the diamond column. It was rare, so it was dear. What just happened, quietly, on a thousand ordinary afternoons like mine, is that intelligence-on-demand moved into the water column. Still essential. No longer scarce. No longer yours to price.
And notice exactly what crossed over. The output did. Not the judgment behind it. Not the knowing what is worth asking. Not the trust that decides whose answer anyone acts on. Knowledge did not start arriving from the wall like current, whatever the people selling it would like you to believe. Only the surface of it did, the part that could be produced without a person attached. The distance between that surface and everything underneath it is most of what this essay is about.
So the value does not disappear. It moves. It always moves. It leaves whatever just became abundant and pools wherever something is still scarce. That much is well worn by now. The question almost nobody sits with is the next one. It moves to where, exactly. And underneath that, the harder one: is the place it moves to actually safe this time, or does this wave break a rule that every wave before it obeyed?
---
The system around the tool
Start somewhere concrete, because the abstract version of this argument is the one that misleads people.
Twenty years ago a spreadsheet could already do almost anything. That was precisely the problem. Nobody wanted a blank spreadsheet. A blank spreadsheet is pure possibility, and pure possibility is useless. What people paid for was the system built around the spreadsheet. The structure that decided which fields exist, the order they get filled, the rules that fire when you fill them. Thousands of software companies were built on that one insight. People do not want the bare tool. They want the direction someone has already put inside it.
The same thing is happening now, one layer up. The blank spreadsheet of this decade is the language model. Nobody wants the bare model. Everyone wants the model pointed at something: recruiting, procurement, contracts, strategy. So far the pattern is identical, and the comfort is real. We have done this before. We know how it ends.
This is where two words start doing real work, and most people use them loosely. Wrapper and harness.
A wrapper puts a nicer surface on the model. A prompt, an interface, a workflow, a subscription. You ask it for a sales email, it writes a sales email. A harness does something different in kind. It does not expose the model, it directs it. It gives the model memory, tools, a sequence, and the authority to act across steps without you steering each one. A lead enters the system, and the harness enriches it, scores it, drafts the outreach, schedules the follow up, updates the record. The model is no longer the product. The orchestration is the product.
A wrapper exposes the model. A harness directs it. The harness is better engineering. Neither one is a moat.
I have spent real time on this distinction, because it is the right distinction. A harness is the system around the tool, and the system around the tool is where I have always believed the value lives. So the instinct that follows feels obvious. Stop building wrappers. Build the harness. Build the operating layer, the thing that turns a clever tool into a directed system.
The instinct is right about one thing and dangerously wrong about another. And the wrong part is the part that feels most like progress.
---
The sophistication trap
Here is the trap, and it is worth saying flatly, because it cost me a few comfortable assumptions of my own.
A harness is more sophisticated than a wrapper. It is not, for that reason, more defensible. Those are two different things, and people fuse them into one because it feels intuitive that the more complex system must be the safer one. It is not.
The orchestration I just described, enrich the lead, score it, draft, schedule, update, is not hard to copy. For a competent team it is a week of engineering. Sophisticated, yes. Rare, no. A harness with generic logic and no audience loses to a wrapper with ten million users every single time, because users do not buy intelligence and they do not buy sophistication. They buy whatever is already sitting inside the tool they open every morning.
It gets worse, and this is the part that should make every harness builder nervous. Much of what looks like a clever, hard to copy harness today is not a moat at all. It is scaffolding. It exists to compensate for what the model cannot yet do on its own. The careful chains, the guardrails, the elaborate routing, the patches for the model's blind spots. All of it rests on the assumption that the model stays roughly as capable as it is today. The model does not stay as capable as it is today. It improves on a curve nobody can time. And when it improves, the scaffolding does not become more valuable. It evaporates. The work you did to route around a weakness becomes dead weight the moment the weakness is gone.
So the sophistication of your harness is not your protection. Quite often it is a measure of how heavily you are betting against the model getting better, which is the one bet in this field that history keeps settling the same way.
Which leaves the real question exposed, sharper now. If the harness is not the moat, what is? And underneath it, the thing this whole wave keeps circling without naming: why does this feel more dangerous than the last time we watched a substrate turn into a commodity?
---
The collapsed distance
This is where the comforting story stops being comforting, and you have to see exactly where it breaks, because everything turns on it.
In every earlier wave, you owned your substrate, or at least you stood safely above it. The spreadsheet company wrote its own software. The platform underneath it, the operating system, later the cloud, had no reason and no real ability to wake up one morning and become your specific product. Microsoft was never going to ship a spreadsheet for recruiters and erase the company that had built exactly that. The distance between the platform and your product was wide. And that distance, more than any feature you shipped, was your safety. You could climb one floor above the commodity and build there, knowing the floor below you could not climb after you.
That distance has collapsed.
When you build on a language model, you do not own your substrate. You rent it. You rent it from a lab that holds the same core technology you are wrapping, that has every incentive to move up the stack into your territory, and that has a visible habit of shipping next month the exact feature you shipped last month. For the first time in the history of these waves, the substrate is itself intelligent. It does not sit there inertly, waiting to be built upon. It can climb.
For the first time, the thing you build on can build you.
This is the real reason the ground feels different under this wave, and why the old move, just build one floor higher, no longer saves anyone. The floor below you is awake. The refuge that protected every previous generation of builders, the simple fact that platforms could not become applications, is gone. The platform is now made of the same material as the application, and it is rising.
It is the same vertigo I felt on that ordinary afternoon, scaled up to an entire economy. You think you are standing on something solid, a skill, a product, a layer you climbed years to reach, and then you feel it lift, and you understand it was never a floor. It was a height the scarcity lent you, and the scarcity is leaving.
So you cannot hide above the substrate anymore. The honest question is no longer how to climb one floor higher. It is the opposite. When the substrate climbs into your layer, what does it not automatically get to bring with it?
---
What the substrate does not inherit
The answer is short, and it is the one the whole conversation was circling. When the model climbs into your layer, it brings its intelligence. It does not automatically bring four other things. Your distribution. Your proprietary data. Your trust. Your seat inside a specific organization's operations.
Be careful here, because this is where the argument is easiest to overstate, and I overstated it the first time. These four are not beyond the lab's reach. That claim is wrong, and a single example knocks it down. Microsoft owns distribution at a scale no startup will ever touch. The large providers are accumulating trust directly, millions of people now believe a chat window over the professionals standing next to them. The labs are working their way into the enterprise environments where the proprietary data lives. So no, none of these are unreachable.
What is true, and far stronger for being defensible, is this. None of the four come free with the intelligence. Intelligence is now the thing that arrives automatically, at almost no cost, the moment you connect to a provider. The four do not arrive with it. Each has to be won separately, slowly, out in the world rather than inside the model, and each is far harder for a provider to acquire than the next jump in capability. The capability jump is their home turf. Distribution, data, trust, access, that is everyone's turf, and it is fought for on foot, relationship by relationship, never shipped in a release.
Of the four, distribution is the one people most misunderstand, so it is worth being precise. Distribution is not marketing. It is not a louder launch or a bigger budget. Distribution is repeatable, controlled access to the people who will use the thing, such that you do not have to re-win their attention every time you have something new to give them. There is a simple test. When you ship something new, do you already have somewhere to put it, or do you have to go and find an audience from scratch. If you have to go and find them, you do not have distribution. You have a product and a hope.
There are only a few ways anyone has ever built it. You inherit an installed base, which is why the incumbents drop AI into hundreds of millions of seats that already pay them. You own an audience directly, the writer's path, attention already earned that you can ship straight into. You become the surface where the work happens, the system of record, so the user already lives inside you. Or in serious domains you hold trust, where being the name people already believe is itself the access. None of these is a better model or a cleverer harness. Distribution is a different game from intelligence, and it decides who is still standing after the commoditization finishes.
And here the essay has to turn, because this stops being a problem about companies the moment you look at it directly. If intelligence is the thing being commoditized, then it is not only AI startups building on borrowed ground. It is every person whose value was their intelligence. The clear thinker. The good writer. The sharp analyst. The competent professional whose worth was the quality of what came out of their head. All of that is the substrate now. All of that just got cheap. The vertigo I described at the start is not a founder's problem. It is waiting for everyone who ever stood on a skill. So the question stops being theoretical and becomes personal. If the model can do the thing you were proud of, what is your moat as a person?
---
You are the wrapper
You are now the wrapper. That is the uncomfortable, clarifying way to say it, and it is also not quite true, which is exactly why it is useful.
A person is not a wrapper. A person is closer to a harness. You carry judgment, context, memory you can be held to, accountability, relationships, the ability to be answerable for a decision long after you have made it. None of that is raw output. All of it is direction. So the clean version, you got commoditized, is not actually right, and the wrong part of it is the part worth rescuing.
Be precise about what got cheap. Not you. The wrapper-layer of you. The part that produced competent words, summaries, analyses, first drafts on demand, the part whose value was the quality of what came out of your head with nobody attached to it. That layer is now free. The harness-layer of you, the judgment about what is worth doing and the relationships that let you do it, did not get cheaper. If anything it grew more valuable, because it is now the scarce half of a thing whose other half fell to zero.
Your raw cognitive output became cheaper. You did not.
So your moat as a person mirrors the company's. It is not your intelligence. It is what the substrate does not inherit when it climbs into your work: the people who trust you, the relationships you have already earned, the proprietary data of your own lived experience that no model was trained on, and the right to be in the room, the seat inside someone's decisions that took years to earn and transfers to no one else.
For anyone who writes, this is not abstract. Your voice and the ideas you coin are a kind of distribution, and they are among the hardest kinds for a model to take, because they are tied to you. A model can generate any harness. It cannot generate the fact that when a certain question comes up, a certain group of people think of your name and your framing first. When you put a concept into someone's head and it becomes the way they see a problem, you have laid distribution inside their thinking that no provider can simply absorb, because it does not point at a tool. It points at you. The framework is copyable. Being the person people return to for the next one is much harder to copy.
That is where I meant to end. It is also where the easy version of this essay ends, and the easy version is wrong, because it quietly assumes the model will always be able to do the work but never be allowed to hold the relationship. That assumption is true today. It is the next thing to break.
---
From answers to action
We are not moving from software to AI. We are moving through three stages, and we keep mistaking the middle one for the destination.
Software executes rules. Intelligence answers questions. Agency acts. The entire market is arguing about the second stage, which model is smarter this month, while the third stage is the one that changes the shape of the question. An agent does not answer you. It does things for you. That is a different kind of object, and almost nobody has priced in what it means.
Picture it honestly, a few years out. Your agent knows your calendar, your inbox, your finances, your health, your preferences, the people you answer to and the people who answer to you. It knows you better than any adviser ever could, because it never forgets and it is always there. And it does not merely inform your decisions. It makes most of them, and brings you only the few that still need a human hand.
Now the question sharpens past anything in the first half of this essay. When your agent chooses, researches, negotiates, and buys on your behalf, who actually holds the distribution? You? The agent? Or whoever built the agent? This is the edge where most of the conversation stops, right before the interesting part begins. Because if even distribution can be aggregated by whoever owns the agent, then the moat I just spent an essay defending is standing on rented ground too.
So go one floor further down, to the thing underneath distribution, the thing that does not aggregate so easily. And it is not trust.
---
The permission layer
Trust and permission are not the same thing, and the gap between them is most of the next decade.
Trust means I believe you. Permission means I let you act for me. I trust many writers; only a few change what I decide. I trust many doctors; only a few I let near my health. I trust many advisers; only a few I let change my company. The difference is not belief. It is the right to act inside my life.
Trust is a feeling. Permission is a key.
That key is the scarce thing the agent era runs on. Not the best answer, the right to act on it. The question stops being who has the intelligence and becomes who am I willing to let act for me. That is a completely different question. Call it the permission layer.
And here is the part the hopeful version of this idea always skips, the part that makes it true instead of merely comforting. Permission is not a safe harbor. It is the next contested floor. Walk the logic all the way forward. If your agent holds the permission, and the agent belongs to a provider, then the provider quietly becomes the broker of your permission. The distance collapses a second time, one storey higher. The same climb we already watched, model into application, runs again, agent into the relationship itself. Whoever owns the agent reaches for the permission, unless that permission stays anchored to a person the agent cannot stand in for.
The collapsed distance was never a one-time event. It is a pattern.
It is a pattern, and it keeps climbing, and each time it climbs it comes for whatever we thought was finally safe. First the model came for the answer. Then the application. Then distribution, through the agent. Then permission, through whoever owns the agent. Every floor we run to turns out to be rented. This is the real shape of the thing, and it is why naming any single moat as permanent is a mistake. There is no permanent moat. There is only the question of what the climb cannot reach.
---
What stays yours
Go back to the afternoon. The machine did the thing I was proud of, in seconds, eleven times over. For a while I thought the lesson was that I had to get faster, sharper, higher. That I had to climb. It took me longer than I would like to admit to see that climbing was the wrong instinct entirely, because the floor was rising faster than any person can climb, and it always will now.
Every wave turns yesterday's moat into today's commodity. Capital, distribution, attention, intelligence, each was scarce, each became abundant, each handed its value to whatever it could not itself become. The moat was never any single one of them. The moat is whatever the newly abundant thing cannot carry with it as it climbs.
Intelligence can climb into the product. Distribution can be aggregated by the agent. Permission can be brokered by the machine. But one thing still resists transfer: the person the permission is anchored to. The human source no agent can stand in for. Chosen for reasons that have little to do with capability, and never will.
So I stopped trying to climb. You cannot win the height. The floor will always rise to meet you. What you can do is the opposite of climbing. You can go closer. Closer to the few people who would rather have the answer from you than from the machine that can also produce it, faster, eleven times over. Not because your answer is better. It often will not be. Because the having-it-from-you was the part that was never for sale.
That is the last distance. Not the distance above the commodity, the one that just collapsed under everyone at once. The distance between a person and the people who choose to let them act inside their lives. The machine can take the work. It can take the answer, the workflow, the harness, the audience, in time even the agent that carries your voice into the world. It cannot be the one they trust enough to hand the key.
That distance has not collapsed. It is the last one that cannot.