The path I thought I was building toward is changing faster than I can keep up. Here's how I'd think about it if I were starting out today.
May 19, 2026
TechnologyEvery major technology shift has been met with the same prediction: the new machine will make workers obsolete. The prediction has been wrong every time.
The printing press didn't eliminate writers. The assembly line didn't end the need for engineers. The computer didn't close the office. Each wave displaced some work and created more than it destroyed.
The question right now is whether that pattern holds.
The data on early-career employment is no longer speculative. A Stanford study found a 16% relative decline in employment for workers ages 22 to 25 in AI-exposed fields since 2022, nearly 20% for software developers specifically.1 Employment for older, more experienced workers in the same fields held steady or grew. Dario Amodei has suggested AI could eliminate roughly half of all entry-level white-collar jobs within five years.2
The entry-level job was never just a job. It was where you became experienced. The bottom rung of a ladder you could only climb by going through it.
Medieval craft guilds offer an instructive parallel. The progression from apprentice to journeyman to master wasn't credential theater. Each stage was where the actual skill happened. Most prior technology waves had automated physical labor and left the cognitive structure intact; the displaced had somewhere to go. Industrialization was different. When it arrived, it didn't spare the masters while cutting off the pipeline. It collapsed the entire structure. Masters, journeymen, and apprentices went from prosperity to poverty within a generation as power looms made the skill economically worthless. The craft didn't survive in reduced form. It was displaced outright.
AI is entering from inside the cognitive stack. Not the most senior work. The apprentice work. The first draft that taught you what a good argument felt like. The research pass that showed you where the edges of a topic were. The code that didn't work until you understood why. Those tasks built taste. Taste is what makes someone genuinely good at something rather than just technically capable of producing it.
The reps disappeared. And without the reps, the taste never develops.
The instinct most people have in response to all of this is to find safe ground. The AI-proof major. The hard-to-automate skill. The niche that's too small to be worth the model's time. That framing is understandable and almost entirely wrong, because it's still asking a defensive question.
The better question: what position do you need to be standing in to see something the model can't see?
In a generalist setting, the model covers the apprentice work. The first draft, the research pass, the initial code. All of it handled before you've had the chance to learn from doing it. Go deep enough into something specific and the dynamic inverts. The problems at the edge of a narrow domain aren't in the training data. The model can't draft them. You have to figure them out yourself. The reps come back. And the same depth that brings the reps is what puts you somewhere the model has never stood.
Darwin spent eight years studying barnacles before writing On the Origin of Species. Not because barnacles were economically important or because he was being strategically defensive. Eight years of working through one specific domain, problem by problem, is how the pattern recognition gets built. He was inside the box of barnacle taxonomy long enough that the box itself revealed the structure underneath it. The insight required the specificity. There was no shortcut to the vantage point.
The person who has spent years inside agricultural finance, seeing the weird edge cases, watching deals fall apart for reasons that don't appear in any textbook, is standing somewhere the model has never stood. They're looking at problems the corpus doesn't contain yet. When they figure one out, that solution is genuinely novel. Getting there required being in the world in a way no model is.
That's the actual moat. Not the niche itself. The niche is just the context. The moat is the accumulation of pattern recognition, judgment, and situational knowledge that makes you capable of the outlier thought. The model converges toward the center of what's already been said. The person deepest inside a specific domain is always the one best positioned to see what's just beyond it.
Buffett put the underlying principle plainly in his 1996 letter to Berkshire shareholders: "The size of that circle is not very important; knowing its boundaries, however, is vital."3 He wasn't talking about playing it safe. He was talking about the only position from which you can see something others can't.
If I were 18 again and starting college this fall, I'd ask one question: what's the fastest path to the edge of something real?
Not the highest median salary. Not the credential that sounds safe. The domain that genuinely interests me, pursued deep enough and fast enough that I start running into problems that don't have documented answers yet. That's where the interesting work is.
I'd use AI aggressively to cover ground, to learn faster, to pressure-test arguments, to find gaps in my own reasoning. But as a way to go deeper, not a way to skip going deep.
And I'd get in front of real stakes early. The niche doesn't reveal itself through research. It reveals itself when something actually matters and you have to figure it out without a rubric, when you notice something that other people in the room didn't, when a problem snags your attention in a way that tells you something about where your actual edge is. That signal only comes from being in real situations.
The relationships follow from the specificity, not the other way around. When you're generic you have to go find the people. When you're the person who does a specific thing well, the right people find you.
The models will reach your niche. Maybe in two years, maybe in ten. That's a real concern. But whether your niche is AI-proof is still the wrong question.
The historical pattern holds that every wave created more work than it destroyed. That argument deserves engagement rather than dismissal. It has been right every single time it has been tested. Whether it holds this time is genuinely open.
The person far enough inside a specific domain is doing two things at once: building the kind of pattern recognition that only comes from doing the work yourself, and seeing problems the model can't reach because they aren't in the corpus yet. Those aren't separate benefits. They're the same move, compounding over time.
Those are different questions. They lead to different careers.
Epilogue
The desire for something personal has always been there. Something made for a specific person's situation, carrying the mark of a specific person's judgment. Industrialization didn't eliminate that desire. It buried it under convenience. Mass production made the generic so cheap and abundant that seeking out the handmade started to feel indulgent, even precious.
AI pushes that logic to its endpoint. When competent output on nearly any topic is available in seconds, the generic fills everything. The good-enough article, the adequate advice, the serviceable design. The world becomes more homogenous, faster, than any prior wave of mass production managed.
The pendulum has swung this direction before and come back. After chain restaurants conquered everything, the aspiration shifted to the chef's table, the cook who had a point of view. After fast fashion made clothing disposable, the interest in craft and provenance returned. People didn't stop wanting the personal. They stopped having to think about it. Then the gap between generic and genuine became impossible to ignore.
The desire for something 1-of-1 doesn't get automated away. If anything, flooding the world with the averaged version is what makes it urgent again.