One — A tool that sharpens itself
Most tools don't improve themselves. A knife doesn't sharpen itself; a hammer doesn't sit in the shed designing a better hammer. If we want a better tool, we go back to the workbench and make one. The tool waits for us. That has been true of nearly every tool human beings have ever made.
So it's worth slowing down over a stranger possibility: a tool whose job is to make tools, or even to work on itself. With a tool like that, a question opens up that simply doesn't apply to a knife. What if it got better at the very work of improving? Then each improvement would make the next one a little easier, and the gains would start to stack.
The clearest everyday picture of gains that stack is the way money grows in an account that earns interest. At first almost nothing seems to happen. But the interest starts earning interest of its own, and the growth begins to feed on itself — it compounds. For a long while it looks slow, and then the curve bends upward and climbs faster than most people expect.
That bending curve is the heart of recursive self-improvement: a process that feeds its own result back into itself, the thing it produces becoming the next thing it works on. Put plainly, a system improves itself, the improved version is now a little better at the work of improving, and so it makes the next improvement, round after round — the improver and the improved being one and the same.
Why does this draw so much attention? Because almost nothing in human history works this way. Knowledge and tools usually advance at a human pace — one person teaches the next, one generation builds on the last, slowly. A process that compounds would not move at that pace. It would move like the interest curve: slow, slow, then steep.
For a long time this was just a thought experiment, argued over by a small circle of people. What changed — and the reason a major AI company recently published a report taking the idea seriously — is that the people building these systems now say the first turns of the loop have actually begun. The machines, they say, have started to help build the next machines. The rest of this piece is an attempt to say plainly how true that is, and where it leads.
Two — The doing and the deciding
Start with the plainest version of the claim: the machine is now writing much of the software that the machine is made of.
Every AI system is, underneath, a vast body of code — the instructions that train it, run it, and test whether a change made it better or worse. Engineers write that code, and for most of this field's short history a human sat and typed it line by line, the way a writer types sentences.
According to the company's report, that has shifted, and fast. As of May 2026, most of the new code going into their live systems — more than four lines in five — is written by the AI itself rather than typed by a person. Two years earlier that share was close to zero. The human engineer hasn't disappeared, but the work has changed: instead of writing the instructions, the person now describes what's needed, sets the system to work, and checks what comes back.1
We should be careful with a number like "four in five," because it can sound more exact than it is. Counting lines of code measures how much got written, not whether it was wise or well made — and the report says so itself, which is to its credit. So the figure isn't "the AI does four-fifths of the thinking." It's a rough signal that a large share of the labor has moved from human hands to the machine.
That word — labor — points to the most important distinction in this whole subject, and we'll lean on it the rest of the way. Consider an example from the report. A routine update one day began crashing tens of thousands of the company's computing jobs. An engineer pointed the AI at the live problem with little more than a short description, and the system worked through it on its own — testing one possibility at a time, tracking the fault to a single obscure setting, and confirming a fix. It did in about two hours what would normally take a skilled person two or three days.1
Look at what actually happened there, because two different kinds of work are tangled together in it. One is the doing: the patient, grinding labor of testing, tracing, confirming. The other is the deciding: a human looked at everything going wrong that day and judged that this was the problem worth handing over. The machine did the doing. The person did the deciding.
Hold onto that — doing versus deciding — because almost everything turns on it. Building these systems has always been a mix of the two: the labor of carrying out a task, and the judgment of choosing which task is worth carrying out at all. We see the difference in any workplace. A brand-new employee is handed a specific job: "the export button is broken, please fix it." A senior one is handed a goal and works out the approach. The most senior people barely touch the daily work; their job is to decide what the team should be doing this year, and which efforts are worth anyone's time.
By that measure, what the report describes is a system that has become very strong at the doing and is still a junior employee at the deciding. It can be handed a problem and solve it impressively. It is much weaker at the harder thing: knowing which problems matter, which results to trust, and when a whole line of effort is a dead end.
This is why it's fair to say the loop has started turning but hasn't fully spun up. For it to truly take off, the machine would need to handle not just the doing but the deciding — to choose its own improvements, not only carry out the ones a person chose. Today it does the labor while a human still supplies the judgment. Whether it ever crosses that line is the real open question, and it's worth seeing how thin the evidence on either side actually is.
The people building these systems offer two hints that the deciding might be starting, and the honest thing — which their report does — is to show them with the limits attached, because the limits are most of the story. In one, they set the AI loose on a real unsolved research problem, and by one measure it closed most of the distance to a good answer while two human researchers given a week closed only about a quarter. Impressive — until you read that a human had chosen the problem and a human had written the scoring sheet that defined what "good" meant. The two decisions that mattered most were still ours. In the other, they found past moments where a person had taken a wrong turn, showed the AI only the work up to the fork, and asked what it would do next; lately it picks the better next step about two-thirds of the time. But they had deliberately chosen moments where the human stumbled. On moments where the human's move was already good, the machine came out ahead only about one time in five.1
So the honest reading isn't "the machine now decides better than people." It's narrower and stranger: in the particular spots where people fumble, the machine is increasingly the steadier hand. That is a real signal, and it is not the same as judgment. Whether judgment is one more human knack the machine is bad at for a while and then isn't — the way reading another person's mood, or explaining why a joke is funny, each looked uniquely human until it didn't — nobody yet knows. The builders don't know. That uncertainty is the truthful center of the whole subject, and most of the loud talk around it is an attempt to paper over it in one direction or the other.
Three — The few who can run the loop
Here is the thing to notice, and it's the crux of the piece: the most consequential question doesn't wait on whether the deciding ever gets automated.
Suppose it does, fully — the loop closes, the machine improving itself round after round with no human left in the deciding seat. Or suppose it only goes partway, which is the more likely near-term case. Either way, the first thing that changes isn't what the machine does. It's who ends up holding it.
Return to the interest curve. Picture two people saving, one starting a single year before the other. Early on the head start looks tiny. But because the growth feeds on itself, that small early lead doesn't stay small — it widens on its own, year after year, without the first saver doing anything special. The gap grows because of the shape of the curve.
Now put a self-improving tool in place of the savings. Whoever gets the loop turning first doesn't just get a one-time head start; they get a head start that compounds. Their system improves, which lets it improve faster, which widens the distance from everyone still doing it the old way. In an ordinary business a rival can catch up by working harder or hiring more. Against something that compounds, working harder isn't enough, because the leader's rate of improvement is itself improving. That is the real reason the people building these systems are so gripped: it points toward a lead that might not be catchable.
And a second fact narrows the field further. Running this loop is not something a clever person can do in a spare room. It takes staggering amounts of computing power — vast warehouses of specialized machines, drawing enormous amounts of electricity, costing sums only a handful of companies on earth can assemble. The fuel for the loop is scarce, expensive, and already concentrated in very few hands. You can read the stakes in plain numbers: the company that published the report was recently valued near a trillion dollars and has moved to sell shares to the public, and its rivals sit in the same range. Part of what investors are buying is precisely the bet that the lead compounds.2
So a self-compounding engine, built — and this matters — on the collected knowledge, writing, and work of effectively everyone who ever put anything into the common record, is poised to be owned by a very small number of people. That is the situation. Everything else is a question about what we think is owed, and to whom.
The most hopeful voices give a clear answer, and it deserves to be stated at its strongest rather than waved away. Their advice to an ordinary person is energetic and not foolish: don't be afraid, learn everything you can, treat the upheaval as opportunity. Their answer to the harder worry — what about the people the wave washes over? — is a specific bet. It's that the same wave that destroys jobs creates better ones, and that workers will be lifted above the loop, the way bookkeepers once stopped adding figures by hand and moved to running the machines that do it. The worker isn't discarded, on this view; the worker is promoted. And for those who can't make that leap, the fallback is a second bet: that the gains will spread well enough on their own, and that the job is mainly to raise the floor beneath the least fortunate high enough that the disruption stays bearable.
It's worth seeing exactly what that second bet quietly accepts. A rising floor is fully compatible with an enormous and still-widening gap between the few who own the thing that compounds and the many who receive a comfortable share of what it throws off. You can lift the bottom and concentrate almost all the gains at the top at the same time; the two don't pull against each other. So the cheerful framing isn't cruelty. It's a wager — that growth plus a decent floor is enough, and that the size of the gap doesn't much matter. That is a position about what is owed to whom. It is not a neutral description of how things must be. And the very same facts lead other people to the opposite conclusion.
Four — A check, or a deed
When people try to answer that question — who should hold this, and what is owed to the many who won't — their proposals sort into two families. The two sound almost identical, and most of the public conversation runs them together. But the whole argument turns on telling them apart, and it's a difference anyone can feel.
Picture two tenants in the same building. The first has a generous landlord who, in a good year, hands back a little of the rent. The second doesn't rent — she owns a share of the building itself. In a good year both come out ahead by about the same amount of money. But they are not in the same position, and the difference has nothing to do with the size of the check. The first tenant's good fortune lasts exactly as long as the landlord's goodwill; she has no say in whether the building is sold, or torn down, or run into the ground. The second has a vote, a claim that doesn't depend on anyone's mood, and a piece of whatever the building becomes.
The difference is the difference between a dividend and a deed — between a slice of the profits handed out to owners and a piece of the thing itself. You can be handed a dividend without ever holding a deed, and that gap is the entire debate about who benefits from AI.
The first family of proposals — call it the check — comes, strikingly, from inside the industry. The leaders of the largest AI companies have begun to float ways of sharing some of the proceeds: a public fund to spread the gains across the population, or taxes on AI profits used to pay people a basic income. This is worth taking seriously, not sneering at — a check is real money, and money matters to people who need it. But it's worth naming precisely what it is. In every version, the ownership and the control stay exactly where they are. The public receives a share of the money; the deed never changes hands.
The second family — the deed — has its clearest example in a bill introduced in the summer of 2026 by Senator Bernie Sanders, the most concrete ownership proposal anywhere in mainstream American politics. The American A.I. Sovereign Wealth Fund Act would place a one-time tax on the largest AI companies — those taking in more than $200 million a year from AI — equal to half their value, paid not in cash but in stock.3 (A sovereign wealth fund is a pool of assets owned by the public rather than by private shareholders — Alaska has run one on its oil money for fifty years, mailing every resident a yearly check.) That stock would go into a public fund, which the bill's authors project could be worth several trillion dollars and pay every American something over a thousand dollars a year.4 The reasoning is direct: these systems were built on the gathered knowledge and creative work of everyone, so the wealth they make should belong to everyone.
Here is the part that makes it a deed and not just a bigger check, and it's the part most coverage skips. The fund wouldn't only collect dividends. It would hold voting shares and seats on the companies' boards — the power to block decisions and push for others.3 That is the hand on the wheel the renting tenant never gets. Whether you favor the idea or not, it aims at a different target than the industry's own proposals: not a richer payout, but a share of control.
Sanders put the worry about the first family bluntly: the danger, he said, is that the wealthy decide to buy off the public with a payment, precisely so that nothing deeper has to change.5 You don't have to agree with him to see that the two families answer two different questions. One asks how large the check should be. The other asks whose name is on the deed.
Five — Ownership by whom, and what we mean by abundance
We've reached the question the whole piece has been walking toward, and the honest thing is to leave it standing as a question, because that is what it still is.
Notice first that the deed family has a hard problem of its own, one it can't dodge: ownership by whom? Sanders' bill routes the deed through the federal government — a fund run by political appointees. That is a form of public ownership, but it is a particular form, and "the government owns half the company and mails you a check" is not obviously the same thing as the people owning it. It might be a third thing, with risks of its own, sitting somewhere between the deed and the check. We raise this not to settle it — we can't — but because any honest version of the ownership argument has to carry it, including ours. A deed held in the public's name is only as good as the public's actual grip on it.
But step back from the mechanics, because there's a simpler point underneath, and it's the one we'd ask a reader to leave with.
The word doing a great deal of quiet work in all of this is abundance. We are told, by nearly everyone, hopeful and worried alike, that this technology could bring an abundance the world has never seen — enough to go around, and then some. Take that promise seriously. The question this piece exists to sharpen is not whether the abundance is coming. It's what kind of abundance, and on whose terms.
Because there are two very different things the word can mean. One is an abundance handed down — a generous share of what a few owners decide to release, lasting exactly as long as they decide to release it. That is the tenant's abundance: real money, real comfort, and not one inch of say in whether it continues. The other is an abundance held — a share of the thing itself, with a vote and a claim that doesn't depend on anyone's goodwill. The check and the deed, one more time, now pointed at the future rather than at a tax bill.
It would be easy to miss how much rides on which one we mean, because in a good year they can feel the same. The difference shows up later — in the bad year, or the year someone decides the arrangement no longer suits them. An abundance you do not own is an abundance someone else can end.
So we'd put the question to the reader plainly, and then get out of the way. A machine that may, before long, improve itself faster than any of us can follow is being built right now, out of what all of us know, by a handful of people who will own it. Some of that wealth will almost certainly be shared. The only question that finally matters is whether it is shared as a check or as a deed — handed down, or held in common — and that question is not technical, and not yet answered, and not going to answer itself. It is the one decision in this whole story that was never the machine's to make. It is ours.