The generation we're not training
Theory and evidence: why 94% exposure produces 0% unemployment
Entry-level collapse and the seniority gap we're building in real time
How the SWE role went from T-shaped to Y-shaped
Anthropic's 81K study: chosen learning works, assigned learning atrophies
Freelancers, the laws of physics, and being measurably stupid
The O-Ring theory says your job is safe. Every link in the chain must hold. AI can't replace the whole chain. Comforting — until you realize nobody is training the people who were supposed to replace you.
This issue follows the data: not the panic headlines, not the capability scores, but what's actually happening to the youngest workers, the freelancers, and the ladder everyone else is standing on.
The O-Ring will replace your job.
You know the line. 'AI won't replace you, but someone using AI will.' Cute. Also wrong.
A new report by Massenkoff and McCrory finally gives us what we've been starving for: not vibes, not capability scores, not another 'X% of jobs are exposed' headline. Actual observed usage data cross-referenced with employment outcomes.
The headline finding: a gargantuan gap between what AI can do and what it is actually doing.
94% vs 33% Computer and math occupations. Theoretical exposure: 94%. Observed exposure: 33%. A 61-point gap between panic and reality.
Programmers sit at 74.5% observed exposure, highest of any occupation, and yet... mass unemployment? Nowhere in the data. Zero divergence from unexposed workers since ChatGPT launched.
Why? The O-Ring theory. Named after the part that destroyed the Challenger. In complex knowledge work, every task in the chain must succeed for the output to have value. Automate 90% of a deal and you still need the human who reads the room when the buyer hesitates. Partial automation does not equal job destruction. It equals job concentration.
Remove the bottom rung and the ladder still stands. But nobody new is climbing it.
— Thordur Arnason
A decade from now, we will desperately need humans with deep domain judgment. And we will have prevented an entire generation from acquiring it.
So relax? No. Here's where it breaks.
A 14% drop in job-finding rates for 22-to-25-year-olds entering highly exposed professions. Not older workers. Only the youngest cohort. Companies aren't firing senior programmers. They're cancelling the junior req and buying a $20/month API key instead.
This is the junior trap. Some of us have been banging this drum for years.
The entry-level tasks that score highest on AI capability are the exact tasks 22-year-olds rely on to learn the trade. Remove the bottom rung and the ladder still stands. But nobody new is climbing it.
We are building a seniority gap in real time. A decade from now, we will desperately need humans with deep domain judgment to oversee and correct AI output in complex O-Ring workflows. And we will have prevented an entire generation from acquiring that judgment.
The O-Ring won't replace your job. It will protect it. Right up until we realize nobody trained the people who were supposed to replace you.
That's not a prediction. It's already in the data.
Companies aren't firing senior programmers. They're cancelling the junior req.
— Thordur Arnason
The SWE role has fundamentally changed over a few short months. I have been coding for decades. This is different.
I have lived through many 'this changes everything' moments in software. Most did not. This one did.
Six months ago, a senior engineer's value was measured in code written. Today, it is measured in code directed. Andrej Karpathy went from roughly 80 percent manual coding to 80 percent agent coding in a matter of weeks. Boris Cherny at Anthropic ships more than twenty five PRs a day, all written by Claude. Not reviewed by Claude. Written.
I have felt this shift myself over the past year. The muscle memory built over decades. Syntax. Patterns. The mechanical fluency of fingers on a keyboard. It is not worthless, but it is no longer the bottleneck.
The errors have changed. Not syntax mistakes that a linter catches. Conceptual errors. Wrong assumptions. Overcomplicated abstractions. Missing the simple or elegant solution. Catching these requires something implementation skill never really trained deeply. Judgment. Taste. Architectural intuition.
The question is no longer 'can you write this?' It is 'should this exist, and is this the right way to build it?'
T → Y The old T shape rewarded deep implementation skill in a single domain. The Y-shaped professional has a new vertical stem: AI fluency. The ability to direct, evaluate, and collaborate with AI systems.
Anthropic now hires mostly generalists. Not because depth no longer matters, but because the kind of depth that matters has shifted. Strategic judgment. Quality discrimination. Knowing what to build and why.
Most people are not software engineers. Most people are knowledge workers. This just happened to engineers. Where does it happen next?
The 10x engineer is not 10x faster at typing. They are 10x better at deciding what should exist.
— Thordur Arnason
From the Anthropic 81K study:
Two: Voluntary learning works. Forced learning creates atrophy.
Educators report witnessing cognitive decay in their students at 2.5 to 3 times the baseline. Tradespeople learning on their own terms? Almost zero atrophy (4%, less than half average) and some of the highest learning gains of any group.
The pattern is precise: AI learning works when chosen. When assigned, it becomes a shortcut machine that hollows out the skill it was supposed to develop.
Every enterprise mandating AI enablement programs should think hard about this finding.
AI learning works when chosen. When assigned, it becomes a shortcut machine.
— Thordur Arnason
Three: Watch the freelancers.
Anthropic's own language: 'the exposed middle.' Freelance creatives: 23% lived benefit, 17% lived precarity. The one group where upside and downside nearly cancel out. AI is both their tool and their competition.
Corporate employees have institutional buffers. Solopreneurs bet everything on leverage. Freelancers get the gains and the squeeze with nothing between them.
If you want to see what AI does to labor markets before the corporate layer catches up, watch the freelancers. They're running the experiment in real time, on themselves.
Reading this made me want to quit today and found three companies. One that hires electricians. One that hires plumbers. And one that scoops up all the brilliant junior talent nobody is hiring because the market is too busy optimizing Q2 headcount to notice it's eating its own seed corn.
Two are protected by the laws of physics. The third by the fact that everyone else is being profoundly, measurably stupid.
Everyone else is being profoundly, measurably stupid.
— Thordur Arnason
All texts by Thordur Arnason.
Originally published on LinkedIn, 2025–2026.
Editor in Chief: Lena Thorsmaehlum
Publisher: Gervi Labs