What Career Skills Actually Matter When AI Can Do Everything?

Most career advice right now falls into two camps: panic about AI replacing your job, or cheerfully insist that AI is “just a tool.” Neither camp is very useful if you are trying to figure out what to actually do with your career in the next five years.

John Berryman has a more specific answer. He is the founder of Arcturus Labs, an early engineer on GitHub Copilot, and co-author of the O’Reilly book “Prompt Engineering for LLMs.” He left a safe position at one of the most consequential AI products in history because staying inside one company meant his skills were narrowing while the world was widening. Now he consults across dozens of different AI projects, each one in a completely different domain.

The Fork in the Road

Berryman frames the current moment as a literal fork. “I think we are at the fork in the road between utopia and dystopia.”

For people close to the end of their career, the disruption may be navigable. For people in the middle, the window for adaptation is shrinking. The people who are already augmenting their work with AI have a compounding advantage over those who are not. And the gap is widening faster than most people realize.

“If you’re in the middle part of your career, you better be adapting pretty quick so you don’t end up in trouble, know, 10 years left to work and not employable.”

This is not a vague warning about the future. Berryman watched the software stock market have its worst week since 2008 during the recording of this conversation. He was personally invested and feeling the impact. The fear driving those sell-offs is real, even if the timeline is uncertain.

Two Skills, Not Twenty

When asked directly what the most important skill is, Berryman does not give a list of ten things. He gives two.

“Abstract general thinking and entrepreneurship.”

Abstract thinking means the ability to take a messy problem, decompose it into smaller pieces, and figure out which pieces matter. This is the skill that made Berryman effective across wildly different consulting projects — from therapist training simulations to model fine-tuning to RAG architecture. The domain changes every few months. The meta-skill of breaking problems apart stays constant.

Entrepreneurship does not necessarily mean starting a company. It means the mindset that nobody is going to hand you a career path anymore. The old model — specialize deeply, climb a ladder within an institution — becomes fragile when AI can match or exceed specialist knowledge across every field.

“The world of the specialist might be usurped to a large degree by an AI that is smarter than any PhD Nobel Prize winner in all the fields.”

The power is not in having specialist knowledge. It is in wielding that knowledge. If you can direct a system that has PhD-level expertise in every domain to solve a novel problem, you have extraordinary leverage. But you need to be able to see the problem, decompose it, and orchestrate the solution. That is what abstract thinking plus entrepreneurship looks like in practice.

Why He Would Not Build a 2022 Startup

Berryman applies this framework to the startup landscape. He would not build a startup today that resembles anything from 2022 — because the infrastructure, interfaces, and basic applications from that era are already commoditized. You could vibe-code most of them in a weekend.

The interesting problems are at a higher abstraction level. They are meta-problems — things that wield AI better, things that are genuinely unreachable without the current technology, things that resist commoditization because they require novel thinking rather than execution speed.

The startup dilemma he identifies is sharp: if a product satisfies 80% of what a customer wants, in a few years that customer can instruct an AI to replicate it with their specific preferences for free. The generalized SaaS model faces an existential compression that most founders have not internalized.

Teaching It to His Kids

Berryman is not just theorizing about this framework. He is teaching it to his eight-year-old and eleven-year-old. He homeschools them. His wife vibe-coded a Revolutionary War education game. His son has a website arcade with roughly 100 games built using Cursor, automatically published to GitHub Pages.

But the real lesson is not about coding tools. “I wanted them to see me visibly struggling and failing and being confused and getting up and trying again.”

He left GitHub and became an independent consultant in front of his kids — deliberately. The feast-or-famine reality of consulting, the marketing challenges he freely admits he has not solved, the nervous periods between projects — they see all of it. The lesson is not “learn AI.” It is “learn to navigate uncertainty, and failure is not the thing you think it is.”

FAQ

What career skills will matter most in the next 5-10 years?

Abstract general thinking and entrepreneurship. The ability to decompose messy problems into smaller components and the mindset to build solutions independently — rather than waiting for institutional career paths. Specialist knowledge becomes less defensible as AI matches expert-level performance across fields. The leverage comes from wielding AI effectively, not competing with it.

Should I specialize deeply or become a generalist in the AI era?

Berryman argues for generalism with depth — the ability to work across domains while thinking from first principles. He left GitHub specifically because working on one fixed project meant missing developments across the broader AI ecosystem. As a consultant, he works on a completely different project every two months. The meta-skill of decomposing unfamiliar problems matters more than domain expertise.

Is it still worth starting a startup in 2026?

Yes, but not one that looks like a 2022 startup. Infrastructure, interfaces, and basic AI applications are already commoditized — you could vibe-code most of them in a weekend. The startups worth building operate at a higher abstraction level: genuinely novel problems, infrastructure that AI agents depend on, and value that resists the hyper-personalization threat where users can replicate 80% of a product for free.

How do I adapt my career to AI disruption if I am mid-career?

Start augmenting your work with AI tools immediately. The gap between AI-augmented professionals and non-augmented ones is compounding. Focus on developing problem decomposition skills and the ability to orchestrate AI tools rather than competing with AI on execution tasks. Berryman frames the urgency plainly: adapt quickly or risk being unemployable with a decade of career left.

How is AI changing the value of a PhD or deep specialization?

The specialist knowledge that took years to develop can now be matched by AI systems across every field. Berryman describes this as “kind of humbling” — an AI that is smarter than any Nobel Prize winner in all fields simultaneously. The value shifts from holding specialist knowledge to wielding it: knowing which specialist questions to ask, how to validate AI output, and how to combine insights across domains.

What should I teach my kids about careers and AI?

Berryman teaches his children two things: comfort with failure and familiarity with AI tools. He deliberately left a stable job to consult independently where his kids could watch him struggle, fail, and succeed. His eight-year-old uses Cursor to build games published to GitHub Pages. The lesson is not technical proficiency — it is that trying, failing, and trying again is the normal state, not the exception.

How does AI consulting work as a career path?

Berryman describes AI consulting as feast-or-famine with extraordinary variety. Every project is a completely different domain — from therapist training simulations to model fine-tuning to RAG architecture. The upside: constant learning and breadth. The downside: “there’s sometimes where I’m not busy and I get nervous.” Marketing is the hardest unsolved problem for technical consultants.

What is the biggest mistake people make when thinking about AI and careers?

Treating AI as either an existential threat to panic about or a neutral tool to ignore. Both framings miss the point. The real question is whether you are building the skills to wield AI — problem decomposition, orchestration, entrepreneurial thinking — or competing with it on tasks it will eventually do better. One path has leverage; the other faces compression.