Here's a weird contradiction.
Anthropic published a major labor market study this week showing AI could theoretically handle 94% of computer science tasks... but is only being used for 33%. Legal? Theory says ~90%; reality is barely 20%. And crucially: they found no systematic increase in unemployment for highly exposed workers since late 2022.
Meanwhile, Citadel Securities released a chart showing software engineering job postings are up 11% year-over-year, even as overall postings stay flat.
Wait, so AI ISN'T taking jobs?
Not so fast.
What the Data Actually Shows
Anthropic's study (full PDF, appendix) introduces something called "observed exposure" (a new metric that combines what AI could theoretically do with what it's actually being used for). They cross-referenced this against the O*NET task database covering ~800 U.S. occupations and BLS employment projections through 2034.
The headline findings sound reassuring:
- Computer programmers are the most exposed occupation at 75% task coverage.
- Customer service reps and data entry keyers follow close behind.
- There's been no systematic increase in unemployment for highly exposed workers since late 2022.
But buried in the same data: hiring of young workers (ages 22-25) into AI-exposed occupations has dropped ~14%, echoing separate findings from ADP payroll data. And workers in the most exposed roles tend to be older, female, more educated, and higher-paid. If that sounds like you, you better read the rest of this...
A separate résumé and job posting study found that firms adopting GenAI reduce junior headcount entirely through slower hiring (not layoffs), providing the first large-scale evidence of AI as what researchers are calling "seniority-biased technological change."
So companies aren't firing people. They're just... not replacing the ones who leave. And they're especially not hiring the young ones.
Although in some cases, they ARE firing them. Block just cut nearly half its workforce to under 6,000 employees, with Jack Dorsey explicitly saying AI enables smaller teams for the same productivity.
The Productivity Is Real. The Macro Impact Isn't. Yet.
Alex Imas' research documents a growing body of micro studies showing real productivity gains from generative AI: 14-55%+ in coding, customer support, writing, and even mammography. Often the biggest gains go to less-skilled workers. But those gains haven't yet shown up in aggregate productivity statistics.
His explanation: micro studies test workers who are told to use AI and trained on it. In the real world, only 36% of workers feel properly trained (per BCG). Workers adopt AI unevenly, use it for unproductive tasks, and face organizational frictions that prevent lab-demonstrated gains from scaling. This mirrors what happened with IT in the 1980s and 90s, when Robert Solow famously observed "you can see the computer age everywhere but in the productivity statistics."
Citadel's own report makes a similar observation: the St. Louis Fed's real-time population data shows no inflection in daily AI use for work. And they note that displacing white-collar work "would require orders of magnitude more compute intensity than the current level of utilization."
So the displacement narrative is premature. But the wage narrative is already here.
The Adoption Gap: Who Benefits and Who Doesn't
In a follow-up study, Imas and Soumitra Shukla dig into who actually uses AI in the real world. The picture is uncomfortable.
AI adoption is concentrated among higher-skilled, higher-income workers and richer countries, and the gap isn't narrowing. Using Anthropic's own Economic Index data, they show the countries and workers who adopted earliest remain ahead. There's no sign of convergence. A ChatGPT Pro subscription costs the equivalent of 38.6 months of income in low-income countries.
Within companies, the pattern is the same. BCG found that managers use AI at nearly twice the rate of front-line workers (they call it the "silicon ceiling"). And in one study of 100,000 engineers across 500 firms, even 18 months after companies bought AI coding licenses, only about half of engineers actually used them. When they did, productivity went up 8-15%. But that didn't translate into increased output or changes in employment.
The researchers call this the "adoption margin." In controlled studies where everyone is given the tool and trained on it, AI compresses the skill distribution; less experienced workers benefit most. In the real world, where adoption is voluntary and training is spotty, the advantages tilt toward people who already have the judgment, resources, and motivation to use AI well.
This is the mechanism by which wage compression happens. Not mass layoffs. There's a growing gap between those who use AI to multiply their output and those who don't. And when the layoffs do happen, it will create increased competition for fewer roles actually available amongst those who are newly unemployed.
What Businesses Actually Are (and Why Software Won't Eat Itself)
On the a16z podcast this week, Atlassian CEO Mike Cannon-Brookes offered a framework that clarifies where AI fits into all of this.
His argument: businesses aren't databases. They're accumulated knowledge and processes.
The whole history of software from 1960 to 2022, according to a16z's Alex Rampel, was taking filing cabinets and turning them into databases. There were benefits, but it didn't make the world dramatically more efficient. You still needed a human to retrieve the file, interpret it, and act on it.
"The cool thing about everything that's happening in AI land is that the filing cabinet can do work."
QuickBooks can accomplish a task by itself now. Workday can call three references for a job candidate. Your IT service desk can resolve a password reset without a human touching it. The database isn't just storing information; it's executing processes.
Alex Rampell then invoked David Ricardo's theory of comparative advantage (from 1817, yes), which Atlassian CEO Mike said explains why "vibe coding your own Workday" is, in his word, "preposterous." Even if you could theoretically build your own HR system, you have a comparative advantage doing something else. You should run your business and let specialized software handle payroll, the same way Ricardo argued England should make cloth and Portugal should make wine, even if England could technically do both.
They then made a distinction that matters for the wage question, breaking business processes into two types:
- Input-constrained processes (like customer service, legal review, reference checking): Your customers ask a fixed number of questions. If AI answers them 10× faster, you don't get 10× more questions. You just need fewer people. This is the Zendesk/Sierra story: the number of support seats a company needs could approach zero.
- Output-constrained processes (like marketing, engineering, creative work): Here, you can theoretically do unlimited work. You're constrained by creativity and bandwidth, not incoming demand. For these, companies take the AI efficiency gain and do more output rather than cutting input. You don't fire the marketers; you have them produce 5× more campaigns.
The wage implications differ. Input-constrained roles face headcount cuts. Output-constrained roles face wage pressure, because if one person with AI can do what three did before, the market rate for that work adjusts. Both categories lose pricing power. The difference is whether it shows up as fewer jobs or lower pay per job.
And here's where Ricardo's comparative advantage gets interesting in the AI age: the individual's comparative advantage is no longer about which task they can do fastest. It's about which tasks only they can judge the quality of. AI can write code, but can it decide what to build? AI can draft marketing copy, but can it understand your customer's emotional trigger points from 10 years of face-to-face sales? The comparative advantage shifts from execution to judgment, from doing to knowing what's worth doing
The Builder's Path
Ryan Carson taught over a million people to code through Treehouse, then built and sold multiple companies. On The Neuron's livestream this week, he offered a view from the trenches that cuts through both the optimism and the doom.
"I just closed a $2 million seed round," he said. "I'm not hiring anybody."
Not because he doesn't need work done. Because AI tools let him do the work of what used to require a team. He described running Codex agents in parallel, automating testing with end-to-end browser runs, and using visual debugging tools to iterate on his frontend without manual test scripts.
When asked if he'd hire a junior developer, his answer was immediate: "No. And things are going to get weird on that."
But Carson is an optimist. His bet is that the people who would have been junior developers at Company X will instead build things for themselves. The math he offered is simple and powerful:
To make $5,000 a month (a solid living), you need 250 people to pay you $20.
That's it. Find a problem that's worth $20/month to 250 people anywhere in the world, build the solution with AI tools, and you have a business. No investors. No corporate middleman.
"I think we're going to move into a world where we have an explosion of interesting small apps," Carson said. "People can get 150K, 200K a year revenue out of one thing, and that's pretty great."
This connects directly to the comparative advantage framework from the a16z conversation. You won't necessarily work as a software engineer. But you'll use AI tools (think of them as digital employees who can software engineer) to build software that serves your unique understanding of a problem nobody else is solving. Your comparative advantage isn't coding. It's knowing what 250 people will pay $20 for.
The Counterpoint: Are We Losing the Ability to Judge?
Jeremy Howard, the fast.ai founder and creator of ULMFiT (the foundational technique behind modern AI fine-tuning), offered a sharp warning on ML Street Talk this week that cuts right at this assumption.
His argument: AI-assisted coding is becoming a "slot machine." You pull the lever (type a prompt), the AI spits out code, and if it looks right, you ship it. The dopamine hit is real. But unlike actually writing code yourself, you don't build the deep mental models that let you catch subtle bugs, understand system architecture, or make good judgment calls when things go wrong.
Howard invokes a concept from cognitive science called "desirable difficulty." Learning works best when it's effortful (we wrote about this!). The struggle of debugging, the frustration of a failing test, the slow process of understanding why something works... that friction is what builds genuine expertise. Remove the friction (by having AI write everything), and you get the illusion of competence without the substance.
He compares it to what happened in radiology: radiologists started using AI as a second opinion, but studies showed that over time, human radiologists got worse at catching things independently. They stopped building the pattern-recognition skills that come from effortful practice.
This is a real tension. If the only durable advantage in a wage-compressed world is genuine expertise and judgment, and AI tools are simultaneously eroding that expertise in the people who rely on them too heavily... you can see the problem.
The counter: Howard himself concedes that AI coding tools are genuinely useful in "verifiable domains" where you can test the output (code compiles or it doesn't, tests pass or they don't). The danger is in domains where you can't easily verify, where the slot machine pays out often enough that you stop checking. The takeaway isn't "don't use AI." It's "use AI, but maintain the friction that builds real understanding."
Putting It All Together
Here's where all these data points converge into a single picture:
- AI makes software dramatically cheaper to build, so demand for software explodes. Every small business that couldn't afford a dev team can now build digital products. New use cases emerge that nobody bothered with before because they were too expensive. Hence Citadel's rising job postings.
- But each unit of software requires fewer people. When one developer with AI tools can do the work that used to require three (or when a non-developer like Carson can build what used to require a team), the market rate for each hour of work adjusts. Not immediately. Not uniformly. But persistently.
- The adoption gap determines who wins. Imas and Shukla's data shows that the workers who invest in learning AI tools gain 3-5× leverage. The ones who don't (half of engineers don't use tools their companies already paid for) are the ones whose relative value is falling. Wage compression doesn't happen because AI fires everyone. It happens because the adopters become so much more productive that non-adopters compete at a disadvantage, in a labor pool that just got a lot bigger.
- Comparative advantage shifts from execution to judgment. As Alex Rampel said, the filing cabinet can do work now. The question is no longer "can you do the task?" but "do you know which task to do, and can you verify it was done correctly?"
- But judgment atrophies if you're not careful. Howard's warning about desirable difficulty is real. Use AI as a force multiplier, but maintain the friction that builds genuine understanding. The slot machine is addictive precisely because it feels like expertise. Take the time to struggle and learn what your AI is doing so you can better steer and judge it in the future.
The end result = more software jobs, more demand for digital work, but persistent downward pressure on wages per unit of output. Companies hire, but they don't need to pay as much. The labor pool expands. And the people who thrive are the ones who can create value independently; who understand a problem deeply enough to know what to build, and can use AI to build it without a corporation's permission.
What To Do About It
This is why we keep pushing you to learn to code, to build with AI, to ship your own projects, and to learn to run your own open models. Learning these skills is insurance. The ability to create your own value, independently, without relying on a single employer to pay you for tasks that get cheaper every quarter... that's the most durable career move you can make right now.
As Ryan Carson said on our livestream: in order to make a solid living (call it a wage of $5K a month), all you need is 250 people to pay you $20. If you can solve a problem worth $20 to 250 people, you can make a living working for yourself. No outside investors or corporate middleman needed.
The tools are there. Codex is free to start, and just launched on Windows. Claude Code starts at $20/month. Replit handles deployment with one click (so does v0, which is this author's choice amongst those two).
The barrier to building a product and putting it in front of paying customers has never been lower. The barrier to doing it well is taste, judgment, and domain expertise, the things AI can't replace. Howard is right that you need to maintain real understanding, not just accept AI output on faith. Ryan recommended a Boris podcast that suggested you go one layer deeper to understand large language models. Cannon-Brookes and Rampel are right that accumulated knowledge and process design are where the real value lies. And Carson is right that the math now works for individuals in a way it never has before.
Waiting for your current company to figure out what to do with AI is the riskiest move of all.
