A marketer I know built a customer dashboard last month. Not a mockup — a working dashboard with live data, filters, and automated alerts. Two years ago, she'd have put in a ticket to the analytics team and waited three weeks.

She didn't learn to code. She described what she wanted to an AI, iterated on the output, and had something functional by lunch. And it's good. Not "good for a non-programmer." Actually good. It solved her problem, her team uses it daily, and the analytics team never knew it happened.

This is happening everywhere. Designers are shipping code. Developers are generating production visuals. Product managers are building prototypes that used to require a sprint team. The floor of what one person can do has risen so fast that the org chart hasn't caught up.

The engineer who stopped engineering

Andrej Karpathy — co-founder of OpenAI, former Tesla AI lead — coined "vibe coding" in February 2025. The idea: fully give in to the vibes, embrace exponentials, forget the code even exists. Just describe what you want and let the AI build it. Collins Dictionary named it Word of the Year. Non-programmers everywhere started shipping apps.

A year later, Karpathy doesn't code at all.

Not "mostly uses AI." Not "codes less than he used to." He describes not writing code directly "99% of the time," instead orchestrating AI agents and reviewing their output. One of the most respected software engineers in the world — the person who built Tesla's entire autonomous driving stack — has handed the actual writing of code over to machines. "I really am mostly programming in English now," he wrote in December 2025, adding that it "hurts the ego a bit."

And he's more productive than ever.

That's the part that gets lost in the discourse. This isn't a story about a programmer lowering his standards. It's a story about someone operating at a higher level — thinking in systems, reviewing architecture, orchestrating agents — while the mechanical act of writing code happens beneath him. The craft didn't disappear. It migrated upward.

But Karpathy is also honest about what this demands. He describes AI models as "like a slightly sloppy, hasty junior dev." They don't seek clarification when confused. They overcomplicate, bloat abstractions, don't clean up after themselves. Left unsupervised, they produce what he calls "slopacolypse" — masses of almost-right code flooding production systems. His answer isn't to stop using them. It's to be a better reviewer, a better architect, a better orchestrator. He proposed rebranding vibe coding as "agentic engineering" — same practice, more oversight.

Simon Willison, the Django co-creator, drew the sharpest line: "If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding — that's using an LLM as a typing assistant." The difference isn't who writes the code. It's whether anyone understands it.

That's the real pattern. AI hasn't made expertise less important. It's made a particular kind of expertise — the ability to evaluate, direct, and improve AI output — more important than ever.

When good becomes the baseline

Karpathy's story is dramatic because of who he is. But the same shift is playing out across every discipline, at every level.

Stack Overflow's 2025 survey found 84% of developers using AI tools, up from 76% the year before. PwC's 2026 predictions explicitly recommended hiring "generalists and agent orchestrators" over narrow specialists. The barrier to entry across disciplines hasn't just lowered — for practical purposes, it's gone.

A designer can now ship working code. A developer can produce polished marketing copy. A product manager can build a functional prototype without filing a single ticket. And the output isn't a rough approximation — it's genuinely useful. The marketer's dashboard works. The developer's landing page converts. The designer's script runs in production.

Call it the 80% floor. AI reliably gets you to 80% quality in domains that aren't your own. Not as a gimmick or a party trick, but as a genuine, usable, solve-real-problems level of competence. That's transformative. It means one person can now cover ground that used to require a team.

So if good is now the baseline — available to anyone with the right tools — where does exceptional come from?

The last 20%

GitHub Copilot has a 46% code completion rate, but only about 30% of its suggestions are actually accepted by experienced developers. Not because the suggestions are wrong — most of them work. But experienced developers are selective in ways that juniors aren't. They're not just asking "does this work?" They're asking "is this the right approach? Will this scale? What happens at the edges?"

That selectivity is expertise. And it's exactly what separates the 80% floor from the ceiling.

GitClear's analysis of over 153 million lines of code found that AI-assisted development increases code churn and duplication. The code ships — but it accumulates debt. Edge cases, security vulnerabilities, architectural coherence, performance under load — this is where depth still wins, not because AI can't help, but because knowing which questions to ask requires understanding that goes beyond pattern matching.

Addy Osmani, a senior engineering lead at Google, called this "The 80% Problem in Agentic Coding." AI handles the straightforward majority brilliantly. The remaining 20% — the part where things get subtle, contextual, and consequential — is where the real value lives.

But that 20% has always been where the real value lived. The difference is that the 80% used to be hard too. It used to take years to get good enough at a craft to be merely competent. Now competence is a starting point. Which means the people who go deeper — who develop genuine judgement, not just skill — become disproportionately valuable. Not because AI threatens expertise, but because AI made everything else abundant.

Specialisation, reshaped

There's a viewpoint I've held for years: focus and specialisation beat being a jack of all trades, especially for growing businesses. An engineering company with an aerospace specialism in private jets will beat a firm that "does any engineering." That hasn't changed — and it won't.

What has changed is what a specialist can do beyond their core. A specialist aerospace engineer with AI tools can now handle their own documentation, build internal dashboards, prototype customer interfaces, and analyse market data — without becoming a shallow generalist. Their specialism is intact. Their reach has expanded.

This is the T-shaped professional, amplified. Deep in one domain, broad enough — with AI — to operate across adjacent ones. The deep expertise differentiates. The AI-enabled breadth eliminates bottlenecks and dependencies.

PwC describes this as a shift from pyramid workforces to diamond shapes: fewer entry-level specialists doing narrow tasks (AI handles those), more mid-level people who combine domain depth with cross-functional capability. Not generalists replacing specialists. Specialists who can also do the other things — and do them well.

Smaller crews, bigger reach

Think about it through the Viking ship principle. Small, skilled crews achieving outsized impact. Every crew member could row, navigate, fight, and trade — but they still had primary roles. The navigator was the navigator. They were just also capable of everything else in a pinch.

AI extends "everything else in a pinch" dramatically. Your marketing lead can now build their own analytics. Your developer can generate their own visual assets. Your product manager can prototype without filing tickets. The crew doesn't get bigger. Each member gets more capable.

But — and this is the Klarna cautionary tale — you can cut too deep. Klarna went from 5,500 to 3,400 employees on the AI efficiency thesis, then quietly started hiring again when quality collapsed. The CEO admitted it went too far. The lesson isn't that AI doesn't deliver. It's that removing people isn't the same as removing dependencies. The goal isn't fewer people. It's fewer handoffs — less time coordinating, more time creating.

What this actually means

If you're running a team or a business, the implications are practical:

  • Hire for depth plus adaptability. The best candidates aren't the ones with the longest list of skills. They're the ones who go deep in something valuable and are curious enough to use AI across everything else. T-shaped, not flat.
  • Restructure around outcomes, not functions. If your marketer can build their own dashboard, do they need to wait for the analytics team? Rethink handoffs. Reduce coordination. Let people with AI tools own more of the end-to-end.
  • Invest in the last 20%. Training, mentorship, deep expertise development — these matter more now, not less. When anyone can reach 80%, the people who can get to 95% become disproportionately valuable.
  • Build review into everything. Not as bureaucracy — as quality assurance. AI output is good, often surprisingly good. But someone with domain expertise needs to be the final pair of eyes. Karpathy doesn't skip code review just because an AI wrote it. Neither should your team.
  • Watch the Karpathy pattern. Your team will adopt AI tools faster than your processes adapt. That's fine — as long as someone's paying attention. The right response isn't to slow adoption. It's to build the oversight that lets people move fast without accumulating debt.

The takeaway

The specialist era isn't ending. It's shapeshifting. AI has raised the floor so high that the old boundaries between disciplines feel arbitrary — but the ceiling, the part where deep expertise creates real value, hasn't moved at all.

The winners in this new landscape aren't generalists and they aren't narrow specialists. They're people who go deep enough to be genuinely expert in something that matters — and broad enough, with AI's help, to do everything else without waiting for someone else to do it for them.

Karpathy spent a year trying to name this. First it was vibe coding. Then it was agentic engineering. But the real shift is simpler than either label suggests: good became the baseline, and the people who know the difference between good and great are the ones who'll thrive.