The Last Scarce Thing
What happens to us if the machines never wake up — but everything else gets done anyway.
Picture a Tuesday about fifteen years from now.
You open a browser. You don't type anything. The tab already knows that the quarterly reconciliation you were dreading is done, that the three vendors you were going to chase have been chased, that the deck you needed exists and is good, and that the two decisions you actually have to make are sitting at the top waiting for you. You spend forty minutes on those. Then the day is over.
No machine in this picture is conscious. Nothing has woken up. There is no artificial general intelligence anywhere in the building. And yet almost all of the work is gone.
I think this is the far more likely future than the one we keep arguing about, and I think we're badly underprepared for it — not economically, but personally.
The jump we already made was smaller than it looked
There's a comfortable assumption baked into most AGI timelines: that we've been climbing a staircase, that we just took a big step, and that a few more steps of the same size get us to a mind.
I don't think the staircase is one staircase.
Look at what actually happened between a logistic regression scoring loan applications in 2010 and a model today that can read a contract, spot the indemnity clause, and draft the redline. That feels like a chasm. It wasn't. It was the same idea, scaled: find the statistical structure in a large pile of examples humans produced, and reproduce it. We stopped hand-designing the features and let the model learn them. We made the models bigger, the data larger, the architecture better at holding context. Every single step was recognizably the next step on the same road. Nobody had to invent a new kind of thing. They had to build a bigger version of an existing thing, and then wrap it in software until it was useful.
The jump to general intelligence is not that. It asks for capabilities we have never built at any scale, in any lab, even badly:
- Goals that originate. Everything we have is inert until pointed. It has no preference about whether it runs today.
- A causal model of the world, not a correlational one — the difference between knowing that these words follow those words and knowing that pulling this lever moves that rock.
- Learning from one example. A child touches a hot stove once. Our systems need the entire written output of civilization to approximate common sense, and still fumble a novel physical situation a five-year-old handles.
- Knowing what it doesn't know. Not hedging language — actual internal awareness of the edge of its own competence.
- Coherence over months. Holding a goal through interruption, failure, and changed circumstances without drifting into nonsense.
Here's the tell that I keep coming back to as a practitioner: our systems are good roughly in proportion to how much of a domain humans have already written down. They are superb at code, prose, law, and medicine — the most heavily documented activities in history. They are weak exactly where the record thins out. That's not the signature of a mind. That's the signature of a very good compression of everything we've ever said.
Generality has to be good where nothing has been written down. That's a different machine.
I'll be honest about the weakness in my own argument: "this needs a fundamentally new kind of thing" has an embarrassing track record. It was said about flight, about chess, about translation, about protein folding. Sufficient scale has a habit of turning a difference of kind into a difference of degree, and people who bet against it have lost repeatedly. So I'm not claiming impossible. I'm claiming not on this road, and not on this timeline. Everything we have gets better by adding more of what we already have. There is no amount of "more" that produces a goal out of a system that doesn't have one.
Give it a hundred years. Maybe. But plan for the century where it doesn't happen.
The thing that actually arrives
Now here's the part people miss. You don't need AGI to remove the work.
The last mile of automation was never intelligence. It was trust. I've spent my career on the boring end of this — the guardrails, the evaluation harnesses, the permission boundaries, the audit trail, the rollback path. The reason an agent isn't running your finance function today is not that it's too dumb. It's that nobody can prove what it will do on the bad Tuesday, and no one will sign their name under it.
That's an engineering problem, and engineering problems fall. Evaluations get rigorous. Guardrails get standardized. Actions become permissioned, logged, reversible. Agents get the browser, the terminal, the calendar, the ERP. The interface stops being a prompt box and starts being intent — you don't ask, you're understood, because the system has years of your context and a narrow, well-fenced ability to act on it.
Call it intent-complete automation. It is not a mind. It's an execution layer with no wishes of its own, sitting perfectly still until someone wants something.
And it eats the middle of almost every job. Not the top, not the bottom — the middle. The translation of a decision into artifacts. Which, if you're honest about how you spend your week, is most of your week.
So what are people for?
The standard answer — "creativity and empathy" — is a consolation prize and everybody can smell it. Here's what I think actually survives, and it's more interesting than the standard answer.
Wanting things. If execution costs collapse to nothing, every bit of remaining value migrates upstream to knowing what to ask for. This sounds soft. It isn't. Problem selection is the hardest skill in any field and the least taught. A system that will build anything you specify is worthless in the hands of someone who doesn't know what's worth building — and merciless, because it will build the wrong thing beautifully, at scale, before lunch. Taste stops being a garnish and becomes the load-bearing skill.
Being answerable. A model can be wrong. Only a person can be responsible. Somebody has to sign, and signing is not a formality — it's the thing that makes an institution possible. Courts, regulators, patients, clients, shareholders: all of them require a human on the hook, not because of sentiment but because accountability without a body to attach it to isn't accountability. This role doesn't shrink as automation grows. It grows.
Verification. When generation becomes free, checking becomes the expensive step. The person who can write the eval that catches the subtle failure is worth ten people who can write the prompt. We are heading into a world drowning in plausible output, and the scarce skill is the ability to say no, that's wrong, and here's how I know.
Everything whose value is in the doing. Automation only threatens activities where the output is the point. An enormous share of human meaning lives in activities where the process is the point — cooking a meal, learning an instrument, running a distance, sitting with a friend, raising a child. Chess engines crushed us thirty years ago and chess is more popular than it has ever been, because nobody was ever playing chess to find out what the best move was. Watch what happened to bread, to woodworking, to vinyl. When a thing stops being necessary, it becomes chosen — and chosen things carry more meaning than necessary ones, not less.
Attention that costs something. A machine's attention is infinite and therefore worthless in the only sense that matters. The reason another person's attention means something is that they had somewhere else to be and chose you. That's not a feature you can add to a system with unlimited capacity. Human presence stays scarce because scarcity is the whole mechanism.
The actual danger
It isn't unemployment. It's unearned ease.
Every capability we stop exercising, we lose. That's not a moral claim, it's just how the equipment works. We already ran this experiment once: physical labor left most people's lives in the twentieth century, and the result wasn't a species of well-rested people. It was a species that had to invent the gym — deliberate, pointless, structured difficulty, performed on purpose, because the body needed the load and the world had stopped providing it.
We're about to need the same thing for judgment. For memory. For the capacity to sit with an unsolved problem long enough to solve it. If the answer always arrives in four seconds, the muscle that tolerates not-knowing atrophies, and that muscle is where everything good in a person comes from. The discipline of the next century will be choosing what to do the hard way — not out of nostalgia, but as maintenance.
The hardest question I'll leave open, because I don't think anyone has a clean answer: if output decouples from labor, who owns the agents, and how does anything get distributed? There are serious people arguing for redistribution, for broad ownership, for the position that new work always appears as it has after every prior wave. Each has real evidence and real holes. That's a genuine political argument, not a technical one, and it deserves better than a confident paragraph from an architect.
What I'd do about it, starting now
Move upstream, deliberately, while you still have time:
From doing to specifying. The person who can state a problem precisely enough for a machine to solve it correctly is doing the actual work.
From specifying to judging. Build the ability to look at good-looking output and find what's wrong with it. Practice it consciously.
From judging to being responsible. Put your name on things. Own outcomes, not tasks. It's the only position that automation structurally cannot occupy.
And then, separately, protect a few things from the automation entirely. Cook the thing that could have been delivered. Write the first draft yourself sometimes. Sit with the problem for an hour before you ask.
Not because it's more efficient. It won't be. Because efficiency was never the point of being a person, and we're about to find that out the hard way.
The machines may never wake up. That was never the interesting question.
The interesting question is whether we stay awake.