We Are Rebuilding 1810
The word “scientist” did not exist until 1834. It was invented by a man worried that knowledge was coming apart — and given, first, to a woman who refused to let it.
In March 1834, William Whewell reviewed Mary Somerville's On the Connexion of the Physical Sciences and needed a noun that did not exist. “Man of science” was inaccurate on two counts: Somerville was not a man, and she was not a practitioner of one science. She wrote about the joints between astronomy, physics and chemistry — about how the parts held together.
So Whewell coined scientist. And the reason he reached for a new word is the part worth keeping: he was alarmed. He described what he saw in his own field as a “proclivity to separation and dismemberment.” The era in which a single person was expected to hold a working grasp of nature was visibly ending, and he wanted a name for what would come after it.
The word for the modern expert was born out of anxiety about the loss of the generalist. It was first applied to someone who had not yet been lost.
How the walls actually went up
It happened in two moves, and neither was a decision to fragment.
France, 1793. The Convention abolished the universities — and the academies with them. Not reformed: closed. A year later, in their place, came institutions built around a single purpose. The École polytechnique opened in 1794 to produce engineers for the republic, with Lagrange and Monge at the board. Purpose-built, narrow by design, and extraordinarily effective.
Berlin, 1810. Wilhelm von Humboldt founded the university that became the template for every university now standing: teaching and research fused, the seminar, the doctorate, the chair. This is the model that won, and it is the one that produced the modern department.
Here is the irony that should survive this essay. Humboldt built it for unity. The founding conception explicitly called for all disciplines to be present in one place and for their representatives to talk across the boundaries. Fragmentation was not the goal. It was the emergent behaviour of a structure designed to prevent it — because a chair needs a budget, a budget needs a boundary, and a boundary is a wall whether or not anyone meant to build one.
What the walls bought, and what they cost
They bought almost everything. The division of cognitive labour is the reason the nineteenth and twentieth centuries happened at all. Anyone arguing that specialisation slowed progress has to explain the two centuries that followed it, and they cannot.
What changed is not the rate of progress. It is the location. Look at where the large moves of the last hundred and fifty years actually came from: biochemistry, biophysics, molecular biology, cognitive science, bioinformatics. Every one of them is a hyphen. Every one of them began as a seam between two departments, and every one of them eventually hardened into a department of its own — at which point it needed a new hyphen.
The cost is not slowness. The cost is that the seams are structurally homeless. A seam has no chair, no journal, no funding line, no career track, no one whose promotion depends on it. A person who works there pays a tax that a person working inside a discipline does not. So fewer people go, and the most productive ground in the system is the ground with the worst incentives.
Two findings worth knowing, and one worth handling carefully. Benjamin Jones's work on the burden of knowledge documents what specialisation does over time: researchers narrow, they start their productive careers later, and teams grow — the frontier moves further from any single head. Uzzi and colleagues (Science, 2013) found that the highest-impact papers are overwhelmingly conventional in their references, with one unusual combination injected — depth plus a single foreign element, not novelty for its own sake. The popular claim that scientific work has become measurably less disruptive over the last decades is a real published result but a methodologically contested one, and it should not be repeated as settled.
The part that concerns AI
Now the turn.
Specialisation was a response to a biological constraint. It exists because no human head can hold physics and law and medicine and history at working depth, and no human lifetime is long enough to try. Whewell was not wrong about the problem. Every wall built between 1794 and today is a reasonable accommodation to the size of a skull.
And then something arrives that is not bound by that constraint. A language model has read across every one of those walls. It does not get tired at the boundary of its field, it has no tenure committee, and it pays no social cost for asking a chemistry question in a legal register. For the first time since the seventeenth century there exists something that can hold the joints — the exact thing Somerville was doing by hand.
And the industry's first instinct has been to cut it into specialists.
A model for law. A model for medicine. An agent for security, an agent for finance, an agent for support — each with its own scoped tools, its own domain corpus, its own evaluation set drawn from its own field. Vertical after vertical, each one a chair with a budget and a boundary.
We are rebuilding 1810, in silicon, on purpose — and we are doing it to the first entity that never had the constraint the walls were built for.
The objection, which is correct
Specialisation is not the error, and a general model is not the answer. Depth requires narrowing; a system that is shallow everywhere is useless everywhere. Scoping an agent to a domain is also how you scope its blast radius, and anyone who has watched an unscoped agent operate on production infrastructure will defend that boundary hard. Those are good reasons and they are not going away.
So the question is not whether to specialise. It is the one the nineteenth century answered badly:
Who owns the seam?
In 1810 the answer was nobody, and it stayed nobody for two hundred years. If we scope every agent to a vertical and never fund the space between them, we will get the same answer again — with the same consequence. The most valuable ground will be the ground nobody is responsible for.
What we do about it, and how little it proves
Our own attempt is small and it is worth describing precisely, because the honest version is more useful than the impressive one.
We run work through more than one agent on purpose, and we choose them to be incompatible: one that builds, one whose entire function is adversarial review, one that reasons in a different register altogether. They are pointed at the same artifact and required to disagree in writing. That is not a clever architecture. It is a manufactured seam — an attempt to buy, with structure, the thing the university stopped supplying in 1810.
The rule that goes with it matters more than the setup: convergence is not correctness. When our reviewers agree, that is information about their overlap, not evidence about the world. A field that only reads itself will reach consensus, and consensus of that kind is a measurement of the reading list.
What this does not establish: we have one operator, one workload, and no control group. It is a working practice, not a result, and we would be doing exactly what this essay criticises if we dressed it as one.
The word, again
Whewell invented “scientist” at the moment a unity was breaking, and he gave it first to the person who was holding that unity together by hand.
“AI” may turn out to be a word of the same kind — a name coined right at the point where something whole starts being cut into administrable pieces. If that is what it is, then the interesting work for the next decade is not a larger generalist and not a deeper vertical.
It is making sure somebody lives on the seam, and is paid to.
Sources: William Whewell's coinage of “scientist” appeared in his March 1834 review of Mary Somerville's On the Connexion of the Physical Sciences in the Quarterly Review; the “separation and dismemberment” concern is his own framing. The French Convention abolished the universities and academies in 1793; the École centrale des travaux publics, which became the École polytechnique, opened in 1794. The University of Berlin was founded in 1810 on Humboldt's conception of teaching and research unified across all disciplines. Research findings referenced: Jones on the burden of knowledge; Uzzi et al., Science (2013) on atypical combinations. The “declining disruptiveness” literature is noted as contested and is not relied on here.