Alex Karp says two workers are AI-proof: the trades and the neurodivergent

On the tech show TBPN this spring, Palantir CEO Alex Karp gave the bluntest piece of career advice a billionaire has offered the AI generation: “Everybody’s worried about their future, but there are basically two ways to know you have a future. One, you have some vocational training. Or two, you’re neurodivergent. And when I say neurodivergent, I mean broadly defined.” It went viral as a claim about “non-linear thinkers” and ADHD, but that is the paraphrase. Karp’s own word was neurodivergent, and he was talking partly about himself: he is dyslexic, and in December his company launched a Neurodivergent Fellowship paying $110,000 to $200,000, after a clip of him unable to sit still through a New York Times interview went viral.

Two groups, then: people who work with their hands, and people who think differently. One of those claims is measurable, so we measured it.

Group one holds: the hands-on fields show zero AI demand

Our data can’t tell you whether a robot will ever swing a hammer. But it can tell you something adjacent and concrete: whether employers are rewriting a job around AI, by whether their postings now ask for AI skills. Across 100,000-plus live postings, we tagged which occupations name LLM or agent tooling in their top-20 skill demand. The split is stark.

Say the honest limit out loud: “no AI-skill demand” means employers aren’t asking electricians to prompt a model, not that a machine could never rewire a panel. It’s a demand-side signal, not a robotics forecast. But it points the same way Karp does, and it stacks with a second moat. The safest occupations we track pair the manual work with a license: electrician, plumber, HVAC technician, registered nurse, physical therapist, paramedic: every one shows no AI-skill demand anda required credential. A model can’t pass the licensing board, and it can’t crawl the crawlspace.

The irony the trades are living: AI is causing their shortage

The 2013 Oxford study by Frey and Osborne, the one that put “47% of jobs at risk” into the culture, named exactly three things that resist automation: physical dexterity, creativity, and social intelligence. The trades sit squarely on the first: non-routine physical work in cramped, unpredictable spaces, which is Moravec’s paradox in a tool belt: the tasks easiest for a human are the hardest to automate.

And here is the part that should end the “learn to code, not to weld” era for good. The US Bureau of Labor Statistics projects roughly 80,000 new electrician openings a year, with electrician employment growing 9% against 3% for all jobs. The engine of that demand is the AI build-out itself: data-center construction, where electrical work runs 45 to 70 percent of the cost, is projected to need 300,000+ new electriciansthis decade, and a journeyman license takes three to five years and 8,000 hours to earn, so the shortage can’t be closed on demand. Fortune called the wider skilled-trades gap a “$1 trillion crisis.”

So the same technology hollowing out the entry rung of white-collar work is, physically, creating a historic shortage of the workers it can’t replace. Karp’s group one isn’t just safe. It’s where the money is moving.

Group two: a thesis, a fellowship, and a caveat

The second claim is harder, and it deserves more care than the viral version gave it. Karp’s wager is that neurodivergence— dyslexia, ADHD, autism, “broadly defined”, becomes an edge precisely as AI commoditizes the linear, in-distribution thinking it does best. Palantir put money on it: the Neurodivergent Fellowship drew over a thousand applications, and Karp framed it flatly — “the neurally divergent (like myself) will disproportionately shape America’s future.”

There is real research under the mindset half of this, and it is specific, not a superpower story. Johan Wiklund and colleagues, across the Journal of Business Venturing (2016–2017), find that ADHD traits — impulsivity, hyperfocus, a bias toward action over planning, tolerance for risk, align unusually well with entrepreneurship, and that adults with ADHD are over-represented in self-employment. Which maps onto Frey and Osborne’s second moat, creativity: divergent, cross-domain thinking is the thing generative models, trained to complete the most probable next token, are structurally weakest at.

The caveat is non-negotiable and the honest brands say it: ADHD is a recognized disability with real daily costs, not a hack. The research shows fit for particularroles (founder, creative, high-stimulation, crisis-response), not blanket immunity to automation. Karp is stating a bet, not a finding. But it’s a bet pointed at the same target as the trades: the two human capacities, the hands and the leap, that the current machines don’t have.

What to do with two AI-proof groups

Neither group is a place you simply are or aren’t. Vocational training is a route — often a short, well-paid, license-gated one, that a surprising range of backgrounds can reach, and the instrument on this site measures which trades your current skills already sit closest to. The neurodivergent edge is a working style you can lean into by choosing roles that reward it: founder over functionary, the job with novelty and stakes over the one with a checklist. Both of Karp’s answers reduce to the same instruction the rest of our data keeps giving: stop optimizing for the roles a model is quietly learning to do, and move toward the two things it still can’t: dexterity and genuine divergence.