In 1903 the smart money was on Samuel Langley. He was Secretary of the Smithsonian, a respected astronomer, and he had fifty thousand dollars of War Department money to build the first flying machine. He had the credentials, the institution and the team.
His Aerodrome was catapulted off a houseboat on the Potomac twice that autumn. Both times it went into the river. The second ducking was on the eighth of December.
Nine days later, two brothers who repaired bicycles in Ohio flew for twelve seconds at Kitty Hawk.
The Wrights had no funding, no degrees and no aeronautics department. What they had was a bicycle shop, and the bicycle shop was the whole point. Everyone else was treating flight as a power problem: build a big enough engine, bolt some wings to it, say a prayer. The Wrights had spent years thinking about balance, because balance is what a bicycle is. A bicycle is an unstable machine that a person learns to control. So they built an unstable machine and worked out the controls: three axes, warping wings, a pilot who steers it like a rider.
Langley knew more aerodynamics than either brother. He lost anyway, because he brought more rigour to the wrong question.
The century that bet on depth
For most of the last hundred years the opposite advice was correct, and it was correct for good reasons.
Adam Smith opened The Wealth of Nations in 1776 with a pin factory. One man working alone, he reckoned, could make perhaps twenty pins in a day. Ten men splitting the job into eighteen separate operations made forty eight thousand. Specialisation was not a preference, it was arithmetic, and the arithmetic was overwhelming.
That logic climbed steadily up the value chain. It went from pins to professions: the doctorate, the residency, the tenure track, the twenty year career in a single vertical. By 2008 Malcolm Gladwell had turned a body of research on deliberate practice into a number everyone could repeat at parties, and ten thousand hours became the folk wisdom of a generation. Pick a lane at twenty two. Dig. The depth is the moat.
It worked because expertise was expensive to move. It lived in a small number of heads, it took a decade to install in a new one, and there was no way to borrow it for an afternoon. If you wanted a competent tax opinion you had to go and find a tax person, and finding a tax person was slow enough that being the tax person was a career.
What actually changed
The moat was never really the thinking. It was the lookup.
Take any expert job apart and most of it is retrieval. Not the glamorous part, the part that pays: knowing which of four hundred things usually goes wrong here, which clause is standard and which one is somebody trying it on, which of the plausible answers survives contact with reality. That knowledge was slow to acquire because it was scattered across textbooks, precedent and other people's scar tissue, and the only way in was to absorb it by proximity over years.
The cost of borrowing it has fallen off a cliff. This is the part people either overstate or refuse to look at. The overstatement is that AI knows everything and expertise is finished, which is not true and is not going to be. What actually happened is narrower and more disruptive: the time it takes a capable outsider to go from knowing nothing about a field to being usefully dangerous in it has collapsed from about a decade to about a fortnight.
That does not make the specialist worthless. It makes the specialist's lead temporary, and it moves the prize to whoever can hold two or three fields in their head at once and see where they join.
A Nobel Prize in a subject he never studied
Demis Hassabis reached master strength at chess as a teenager, wrote a large chunk of a video game about running a theme park when he was seventeen, spent a decade in the games industry, went back for a PhD in cognitive neuroscience, then started an AI lab.
Nowhere in that sequence is there any chemistry.
Protein folding had been an open question since the early seventies. How a chain of amino acids arrives at the specific three dimensional shape that makes it work is the sort of problem that consumed entire structural biology careers, and for fifty years the field ground forward with X-ray crystallography and enormous patience.
In 2020 AlphaFold2 turned up at CASP, the discipline's own blind assessment, with predictions accurate enough that the organisers called the problem substantially solved. In 2024 Hassabis and John Jumper shared the Nobel Prize in Chemistry with David Baker.
A chemistry Nobel, to a man whose early CV includes a theme park simulator.
The reframe is what did it. Structural biology had treated folding as a physics problem, a matter of simulating forces until the chain settles into place. DeepMind treated it as a pattern problem: here are a hundred and seventy thousand known structures, learn the mapping. That is an obvious move if you have spent your life in machine learning and a strange one if you have spent it at a beamline. Nobody inside the field was being stupid. They were standing too close to see a different frame.
And DeepMind hired biologists, obviously. The generalist advantage was never about doing without the domain. It is about deciding what kind of problem the domain actually has.
Palantir hires for this on purpose
If you want to see the thesis written into an org chart, look at Palantir.
Their signature role is the forward deployed engineer. You take a strong generalist, drop them inside a hospital group or a bank or an army unit, and they learn that world well enough within a few weeks to build against it. They are not hired as healthcare people or defence people. They are hired as people who can land somewhere unfamiliar and be useful quickly, which is a real skill and mostly an unteachable temperament.
The company is run by Alex Karp, who has a doctorate in social theory from Frankfurt and a law degree from Stanford, and who is not an engineer. A philosopher runs one of the most consequential defence software companies in the world. That is not an amusing footnote. Palantir's entire pitch is an argument about how institutions should use data and where the line sits, and that is a question about power and legitimacy long before it is a question about databases.
The pattern repeats once you start looking for it. Palmer Luckey built a VR headset in a garage and now builds autonomous defence hardware. A striking share of the people who built modern machine learning came out of physics rather than computer science, because the job needed people comfortable with heavy maths and large messy systems, and physics produces those in bulk.
The strongest version of the other side
You do not want a generalist doing your heart surgery.
That objection is correct and worth taking seriously. Some depth really is irreducible. You cannot read your way into knowing how tissue behaves under a scalpel, or how a bridge sounds the week before it fails. There is also the Chesterton's fence problem: the newcomer sees a rule that makes no sense and removes it, and the specialist is the one who knows the rule was load bearing. Confidence without the scar tissue is how you get an outsider who is fast, plausible and catastrophically wrong.
But notice how much work surgery is doing in that argument. It is one of the few fields where the bottleneck is a trained hand rather than trained judgement. Most jobs are not surgery. Most jobs are: read the situation, work out what actually matters, find the people who know the parts you do not, and ship something before the situation changes again. Breadth is not a handicap in that work. It is the job description.
And a generalist who knows nothing is not a generalist. He is unemployed with a wide range of interests. The version that wins has real depth somewhere, usually in two somewheres, and uses each one to interrogate the other.
What the job is now
If the lookup is cheap, what is expensive?
Knowing which question to ask. This was always the scarce thing and it is now very close to the only scarce thing. A model will answer whatever you put to it, instantly and in confident prose. It will not tell you that you are asking about the wrong quarter, or that the real problem is upstream in a department you have not spoken to yet.
Taste, which is telling a good answer from a merely plausible one. Anyone who has used these systems seriously has been handed something fluent, well structured and quietly wrong, and the only defence is enough judgement to feel the wrongness before it ships. That feeling comes from having been burned before, in some field, about something.
Speed of acquisition. Going from zero to dangerous, repeatedly, without finding it unpleasant. Some people find the first week in a new domain exhilarating and some find it humiliating, and that difference is now worth a great deal of money.
Carrying patterns across. The Wrights carried balance over from bicycles. Hassabis carried pattern recognition over from games and neuroscience. The move is nearly always the same shape: take a well understood idea from one field and notice that another field has been solving the same problem badly, in its own private vocabulary, for thirty years.
None of this means depth is dead. It means the returns have moved. For a century the safe choice was to pick a lane early and dig, and that was genuinely good advice in a world where knowledge had to be grown one head at a time. It is now the risky choice, because narrow lanes are exactly what these systems eat first, and digging is the part machines turn out to be best at. The choice that looks reckless, staying curious across several fields and getting very good at moving between them, is the one that compounds.
Langley had the grant, the team and the title. The brothers had a bicycle shop and better questions.
Back the bicycle shop.