In 1903 the person most likely to build the first working aeroplane was Samuel Langley. He ran the Smithsonian, he was a respected astronomer, and the War Department had given him $50,000 to do it.
His Aerodrome was launched off a houseboat on the Potomac twice that year. Both times it went straight into the river, the second time on 8 December.
Nine days later Wilbur and Orville Wright, who ran a bicycle shop in Ohio, flew for twelve seconds at Kitty Hawk.
They had no degrees and far less money. What they had was years of thinking about bicycles. Everyone else treated flight as an engine problem: build something powerful enough and put wings on it. The Wrights treated it as a balance problem, because a bicycle falls over unless the rider keeps controlling it. So they built a plane the pilot had to keep balancing, and spent their time on the controls. Langley knew more about aerodynamics than either of them. He was just working on the wrong question.
Why specialising used to be right
For most of the last hundred years the advice to specialise was correct. Adam Smith opens The Wealth of Nations with a pin factory. One worker on their own could hardly make twenty pins a day. Ten workers who split the job into separate steps made about forty-eight thousand.
That idea moved from factories into careers. Pick a field, do the degree and the training, and spend twenty years getting better at one thing. In 2008 Malcolm Gladwell turned research on deliberate practice into the ten thousand hour rule, and it became something everyone repeated.
It made sense because expertise was slow to move. It lived in a small number of heads and took years to put into anyone else's. If you needed a tax answer you had to go and find a tax person, and that was slow enough that being the tax person was a good career.
What changed
A lot of what experts get paid for is knowing things. Which of a few hundred problems usually causes this one. Which clause in a contract is normal and which one is somebody trying it on. Which of the plausible answers actually holds up. That used to take years to learn, because it was spread across textbooks, old cases and other people's experience.
Now you can get most of it in an afternoon. AI does not know everything, and it is not about to replace people who are genuinely expert. What it has done is shrink how long it takes a smart outsider to become useful in a new field, from years to a few weeks.
So a specialist's head start does not last as long as it used to, and more of the value goes to people who can hold two or three fields in their head and see where they connect.
A chemistry Nobel without any chemistry
Demis Hassabis was a master-level chess player at 13 and co-designed the video game Theme Park at 17. He spent years in the games industry, went back to do a PhD in cognitive neuroscience, and then started DeepMind.
None of that is chemistry. For about fifty years, working out the shape a protein folds into was one of the big unsolved problems in biology, and people spent whole careers on it with X-ray crystallography and a lot of patience.
In 2020 DeepMind's AlphaFold2 entered CASP, the field's own blind test, and its predictions were accurate enough that the organisers said the problem had largely been solved. In 2024 Hassabis and John Jumper shared the Nobel Prize in Chemistry with David Baker.
The difference was how they framed it. The field mostly treated folding as physics: simulate the forces until the chain settles. DeepMind treated it as a pattern problem. There were around 170,000 known protein structures, so they learnt the mapping from those. That is an obvious move if you have spent your life in machine learning and a strange one if you have spent it in a lab. Nobody in the field was being stupid. They were just too close to it.
DeepMind still hired biologists. Being a generalist was never about not needing experts. It is about working out what kind of problem you actually have.
Palantir hires for it on purpose
Palantir's best known role is the forward deployed engineer. They take someone good at a lot of things, put them inside a hospital, a bank or an army unit, and expect them to understand that world well enough to build for it within a few weeks. They are not hired for knowing healthcare or defence. They are hired for being able to land somewhere unfamiliar and be useful quickly.
The company is run by Alex Karp, who has a PhD in social theory from Frankfurt and a law degree from Stanford, and is not an engineer. That makes more sense than it sounds. Most of what Palantir is really arguing about is how institutions should use data and where the limits are, which is a question about power before it is a question about databases.
You see the same thing elsewhere once you look. Palmer Luckey built a VR headset in his parents' garage and now builds autonomous defence hardware at Anduril. A lot of the people behind modern machine learning came from physics rather than computer science, because the work needed people who were comfortable with heavy maths and big messy systems.
The fair objection
You would not want a generalist doing your heart surgery. That is true. Some depth cannot be shortcut. You cannot read your way into knowing how tissue behaves under a scalpel. And someone new to a field will sometimes remove a rule that looks pointless and turns out to have been holding everything up.
But surgery is doing a lot of work in that argument. It is one of the few jobs where the bottleneck is a trained pair of hands. Most jobs are closer to this: work out what is actually going on, find the people who know the parts you do not, and get something out before things change again. Being broad is not a handicap in that kind of work. It is most of the job.
A generalist who is not good at anything is not a generalist either, just someone with a lot of interests. The version that works has real depth in one or two things and uses each one to question the other.
Why this matters to me
This is not abstract for me. I design and I build, and at Toasted Beans I was also the one cold calling law firms and walking into accountancies to sell what I had made. A few years ago that would have meant being average at three different jobs. Once I started using Claude Code I went from websites to web apps, Mac apps and iPhone apps in a few months, and it was the design and the selling that made any of it useful to anyone.
What is worth being good at now
If knowing things is cheap, the expensive part is knowing what to ask. A model will answer anything you put to it, straight away and confidently. It will not tell you that you are looking at the wrong quarter, or that the real problem is in a department you have not spoken to yet.
Taste matters more too, meaning being able to tell a good answer from one that only sounds good. Anyone who has used these tools properly has been handed something well written and quietly wrong. The only defence is having been caught out before, in something, somewhere.
So does learning fast, and being able to go from knowing nothing to useful again and again without hating the first week. And so does carrying ideas between fields, which is what the Wrights did with bicycles and what Hassabis did with games and neuroscience.
Depth is not dead. It is just not the safe bet it used to be, because narrow, well defined work is exactly what these tools get good at first. Langley had the money, the team and the title. The Wrights had a bicycle shop and a better question.