Around 30 BCE a Roman architect named Vitruvius sat down and explained how to make concrete. He named the ingredient, volcanic dust from Puteoli near Naples. He described what it did, setting hard even underwater.
His book survived. It is still in print.
And for roughly a thousand years after Rome fell, nobody in Europe could make the stuff.
Not because the formula was secret. It was published.
Because the people who knew how to do it had scattered. Because the trade in that volcanic ash had stopped. And because a recipe is not a workshop. A workshop is full of people who have made the mistake a hundred times.
The Pantheon stayed standing the entire time. A dome 142 feet across, unreinforced, still the largest of its kind in the world. Everyone could see the proof that it had been possible. Nobody could do it.
Hold that, and look at this year’s layoffs.
The number, and how much to trust it
Employers cited artificial intelligence in 120,136 announced American job cuts between January and September. That is about 21 percent of all cuts, and the most common stated reason.
In all of 2025 the figure was 54,836. So it has more than doubled in three quarters.
Now the part that complicates it. In May alone, AI was cited in 38,579 cuts, 40 percent of everything announced that month. In September it was cited in 3,961, about 9 percent, the fifth most common reason.
A reason that swings from forty percent of layoffs to nine percent in four months is not measuring the technology. It is measuring a story companies find convenient.
So treat 120,136 as the number of times a reason was given, not the number of jobs a machine took. The aggregate labor data does not show an economy-wide AI shock, and anyone claiming otherwise is reading press releases.
What is real, and where
Underneath the noise there is a genuine signal, and it is narrow.
Stanford researchers find a 13 to 16 percent relative fall in employment for 22 to 25 year olds in the most exposed occupations. British data shows job adverts in exposed roles down 38 percent.
The damage is concentrated at the entry rung. Not across the profession. At the door.
Which makes sense, because that is the work AI does best. Verifiable output, the context sitting in one place, a known answer somebody can check. The junior analyst’s first two years. The paralegal’s document review. The graduate developer’s unit tests.
And here is the problem that nobody is pricing. Those tasks were never only output. They were the training.
Rome again
This is the Vitruvius problem exactly.
Roman concrete was not an unwritten secret. The method was published, by a named author, in a book we still have.
What the book could not hold was the rest of it. The feel of a mix about to go wrong. The judgment of when the ash is right. The thousand small corrections a builder makes without being able to say why.
That lived in people. When the trade routes failed and the builders dispersed, it went with them.
The documentation survived perfectly and the capability died anyway.
A profession that stops hiring juniors still has its seniors. For about fifteen years. Then it has a Pantheon.
Alexandria, which did not burn
The second case is the one everyone gets backwards.
The Library of Alexandria is remembered as a fire. Historians now largely agree it was a slow institutional death across centuries, driven by politics and money, with fires as punctuation rather than cause.
The turn came in 145 BCE. Ptolemy VIII, settling scores after a succession fight, expelled the foreign scholars from the city. Aristarchus of Samothrace, the head librarian, fled to Cyprus and died shortly after. The post he vacated stopped going to eminent scholars and started going to political allies.
The scrolls stayed on the shelves. The people who knew what was on them left.
And the thing to notice is the timescale. No year in that story looks like a catastrophe. Each one looks like a budget decision, or a personnel decision, or a reasonable response to a political problem. The loss is only visible from a distance of centuries.
This site has made that argument about the CDC and the research agencies, and about the unit that prosecutes public corruption. It is the same shape every time. Capacity built over decades, removed in a quarter, and not restorable by restoring the budget.
Plato got there first, and was half right
The third case is the oldest and the most useful, because it is the objection itself.
In the Phaedrus, written around 370 BCE, Socrates tells a story about the Egyptian god Theuth presenting the invention of writing to King Thamus. Thamus is not impressed. Writing, he says, will produce forgetfulness in the souls of those who learn it, because they will trust external marks instead of remembering. Students will receive quantities of information without instruction, and will seem to know a great deal while knowing nothing.
They will have the appearance of wisdom rather than wisdom.
That is, almost word for word, the case against AI in 2026.
And Thamus was wrong about the outcome. Writing did not hollow out Greek thought. It produced more of it, and better, and preserved what would otherwise have vanished.
But he was right about the mechanism. Writing really does separate having information from being able to use it. The reason the separation did not destroy anything is that Greece built institutions to close it back up. The Academy. The Lyceum. The relationship between a teacher and a student that does what a scroll cannot.
And the thing AI is displacing is precisely the institution that was doing that job. The entry-level position was not a cheap way to get work done. It was an apprenticeship that happened to pay for itself.
So what actually works
The question is how to keep the leap without draining the pipeline. Four answers, in rough order of how much evidence sits behind them.
| What to do | Why it works | Who has done it |
|---|---|---|
| Separate training from output | If juniors are kept only while cheaper than a machine, they go. Fund the apprenticeship as training rather than as production, and it survives the comparison. | Medical residency. Pilots. Trades with levies. |
| Make somebody bear the cost | A firm captures the saving from cutting juniors now and never pays for the missing senior in 2041. That is a textbook externality and markets do not solve those unaided. | Apprenticeship levies, training tax credits. |
| Keep humans where judgment forms | Checking an answer teaches far less than producing one badly and finding out why. Review seats do not make practitioners. | Aviation, which refused to let autopilot remove hand flying. |
| Watch the entry rung, not the total | The aggregate reports health while the pipeline drains, because the seniors are still there. By the time the average moves it is fifteen years too late. | Nobody yet. This is the cheapest one. |
None of that requires slowing the technology down, and this dispatch is not arguing for that. The gains are real and refusing them is its own kind of loss.
What it requires is noticing that a firm firing its juniors is making a decision with a fifteen year lag and a balance sheet that runs to next quarter.
The scarcinality reading
The framework asks one question. Is the thing genuinely absent, or present and out of reach?
Knowledge, right now, is more present than it has ever been. Every text, instantly, to anyone. By any measure of availability this is the least scarce knowledge has ever been.
The scarce thing is the person who can use it.
And that scarcity is being manufactured, quarter by quarter, by rational decisions that each look correct in isolation. Nobody is burning a library. Everybody is declining to fund a librarian.
What would prove this wrong
Three things, all checkable.
If the entry-level employment gap for 22 to 25 year olds closes while AI adoption keeps rising, the pipeline worry is wrong and the market is handling it.
If firms that cut juniors hardest show no seniority gap in five to seven years, then the apprenticeship was less load-bearing than argued here, and AI genuinely does substitute for experience.
And if the AI-cited share stays near September’s 9 percent rather than climbing back toward May’s 40, the spring was a narrative cycle. The displacement underneath it is far smaller than the headlines claimed.
Two cautions. The Challenger figures count stated reasons in announcements. That is a record of what employers said, not a measurement of cause, and here it is the point rather than a flaw. The historical cases are analogies rather than evidence. Rome lost its concrete for reasons that included collapsed trade routes and a vanished empire, and no analogy carries the weight of a dataset.