AI can already do part of the work in 80% of American jobs.
The scarce thing was never electricity or steel or even money. It was a competent person paying attention to your problem.
That is why the good doctor has a waitlist, why the good school sits inside a catchment you cannot afford, and why the permit office is the way it is. Expertise does not scale, so it pools where the money already is. Almost everything that runs badly runs badly for the same reason underneath: not enough people who know what they are doing, and no way to afford more of them.
Cheap, abundant machine intelligence is the first thing that has ever credibly threatened to change that. Not by being smart. By being cheap. The price of getting a model to a given level of competence fell about 280-fold in eighteen months1. Whatever you think the ceiling is, the floor is dropping through it.
So one of two things is about to happen. Either competent attention stops being scarce and the parts of life that run badly because thinking is expensive start running well. Or the cheapest competent attention in history gets owned by a few companies and rented back to everyone else. The technology does not care which. That indifference is the whole problem, and it is why this is a political question, not a technical one.
The public takes a real share of the upside, in ownership, not a press release. Lab workers can organize and speak without losing their equity for it. The safety net moves at the speed of the layoffs, not the speed of a benefits queue.
Permitting, procurement, and public software that actually work. Everything else here is a fantasy if the government can't deliver it.
Upzone, fix permitting, and stand up a public developer that builds homes at scale. The market half and the public half. Not one or the other.
Transit that is fast, frequent, and cheap. Libraries, parks, and pools that are excellent and open to everyone. What can't be sold to you should be the best thing you have.
The mechanism the optimists skip is substitution. A model that does a task for a dollar competes with a person who does it for forty, and the firm buying the work is not sentimental about the difference.
What is new is the direction. Every automation wave before this one started at the bottom and climbed. Engines replaced horses, mills replaced weavers, containers and cheap flights moved the factories overseas. The deal underneath was always the same: the machine took the dull physical work, and the human moved up into the thinking work, which stayed expensive because thinking is hard.
This wave starts at the top. The work most exposed to current models is high-wage, high-education, cognitive work: writing, analysis, coding, translation, illustration2. The escape hatch every earlier generation used, move up into something that takes judgment, opens onto the one room the machine is already standing in. Tell a displaced worker to learn to write or draw or code, and you are pointing them at the three things the model does for a dollar.
You do not have to believe every job disappears. That is the weak version, and you should not make it; the people predicting mass technological unemployment have been wrong for two hundred years straight3. The strong version is quieter and harder to dodge. Workers have always had one piece of leverage: capital needed them, because there was no other way to get the work done. Give capital a credible substitute at the margin, not for every job, just enough of them, and that leverage is gone. Wages soften. The share of the economy that goes to people who work, rather than people who own, keeps sliding4. It happens without a single dramatic headline about robots taking all the jobs.
And the gains from the substitution do not scatter. They collect in one place: wherever the models are owned.
Frontier AI is the most capital-intensive thing the software industry has ever built. The four biggest cloud companies were on track to spend more than $400 billion on capacity in 2025, heading toward roughly $600–700 billion in 20265. The three leading labs are collectively valued near a trillion dollars while most of them lose money.
Capital intensity is a wall. You cannot build a competitive frontier model in a garage anymore. You need data centers, power-grid interconnects, years of chips, and the kind of money only a handful of firms and states can assemble. Whatever the rhetoric about democratizing AI, the structure pushes relentlessly toward concentration. The constraint is not the algorithm. It is power and capital.
If “ownership decides the outcome” sounds like a slogan, look at what is already happening. Who is allowed to use the most capable models is not set by the market. It is set by the government, by name, in writing.
Since 2022 the US has used export controls to decide which countries and companies can buy the chips that train frontier models. A 2025 rule went further and restricted access to the model weights of the most advanced American systems6. Then a new administration rescinded that rule, wrote a different one, and began approving specific chip sales to specific Chinese buyers case by case, reportedly for a cut of the revenue. Like or hate any of these calls. The point is that they are calls, made by people in a room, about who gets the most powerful tool of the decade. The room exists. The only open question is who is in it.
You can tell when honesty is missing, so here is ours.
We do not know the speed. Estimates of how much cognitive work is exposed run from a tenth to well over half, and exposed is not gone. We do not know the shape of the transition. The skeptics have a real case: measured productivity gains across the whole economy are still small, small enough that serious economists think the entire effect on growth this decade rounds to a rounding error7. They might be right. So far the aggregate data is closer to their side than to the people selling a revolution.
None of that resolves the question this project is about. A slower transition is still a transition, and it ends in the same place: cheap competence, owned by a few, sold back to the rest. Speed changes the timeline, not the destination. And the bar for “this matters” is low. An unemployment shock of even five points would rival the worst recession since the war8. You do not need the science-fiction scenario. You need the boring one, arriving slightly faster than institutions can adjust, which is the one we are clearly getting.
The variable that decides how this goes is not how fast or how total. It is who owns it.
Spend a minute on what cheap competence makes possible, because the doom is not the interesting part.
Almost every public thing that runs badly runs badly because doing it well takes more skilled attention than the public could ever afford to buy. The benefits office is hostile because there are too few caseworkers and the ones there are drowning. The permit takes a year because the queue is long and the expertise is thin. The bus is late because planning a good network is hard and the people who can do it are scarce and expensive. This is not lazy government. It is rationed competence.
That is exactly the constraint about to loosen. The clearest early evidence is unglamorous and points the right way: give frontline support workers an AI assistant and the biggest gains go to the least experienced ones9. It pulls the floor up. Multiply that across every counter, every form, every queue where a public service is bad because skilled help is expensive, and you get the least glamorous and most valuable thing on offer: the DMV that works. The permit that takes a week. Benefits that reach the people they are for without a fight.
This is the oldest practical tradition on the left, and it used to have a rude name. A century ago the people who ran Milwaukee were mocked as “sewer socialists”10, too busy with clean water, public power, and working streetcars to bother with revolution. They kept winning elections for forty years because the sewers worked. That is the bet. Not a utopia. A government that is good at the boring, decisive things, finally able to afford the competence to do them.
We will be straight about the evidence: it is thinner on this side. There is strong data that AI displaces private-sector labor, and only early, partial data that it makes public services better. We think the second is real and worth building toward. We are not going to pretend it is already proven.
The people who understand this technology best are almost entirely absent from the politics of it.
The left, mostly, has decided the technology is the enemy, that the correct posture toward the most consequential tool of the decade is to be against it. That is a choice to forfeit. Refuse to touch the thing and you hand every decision about who owns it to whoever does. Inside the labs the politics come in two flavors: philanthropy aimed at far-future catastrophe, and the conviction that the best thing to do with a powerful technology is build it faster and sort out the consequences later. Both leave the ownership question exactly where it is.
There is a third place to stand. It is empty, and it needs the people who actually build the things. You can believe this technology is real and powerful and believe the public should own a real share of what it becomes. Those are not in tension. Holding both is the entire idea.
At the frontier labs you are hired as a Member of Technical Staff. We think some of us should also be Members of Socialist Staff. You can do both. The contradiction is the point.
Membership is anonymous by design. We like our jobs and our colleagues, and the surest way to keep a member list from leaking is to never collect one.
Use a personal email, not your work one.
We write when there is something worth sending. We store your email and nothing else — no name, no employer, no trackers — and we will never sell it or give it out. Unsubscribing deletes you.