Work Got Faster. Hiring Didn't.

Right now about 3 million Americans have been out of work and looking for more than 15 weeks. That's 43% of everyone who's unemployed, up from about 40% a year ago. Almost 2 million of them have been looking for more than six months.

At the same time, a model that could barely put together a decent essay three or four years ago can now build out an entire codebase, frontend and backend, and reason about it well enough that plenty of engineers just let it drive.

Those two facts sitting next to each other are what I can't stop thinking about. We've gotten very good, very fast, at making work faster. We haven't gotten much better at getting people into the work.

The “somehow” in the labor market

The textbook version of a labor market is simple. People want to work, companies need people, and wages and hours settle somewhere in the middle. But the way the two sides actually find each other is basically “somehow,” and that word is doing a lot of work.

I spent five years in recruiting and interviewed thousands of candidates, so I've seen the “somehow” up close. A person has to find the right opportunity, figure out if it's even real, guess what the company actually needs, prove they can do it, and survive the process. The company is doing the same thing from the other side, usually with a resume and a 30-minute call. Most of the time both sides are guessing.

One quick caveat, because these numbers get misused a lot. The average ongoing spell of unemployment is 26.3 weeks, about six months, but BLS is clear that's not how long it takes the average person to find a job. It only measures people who are still looking. The honest read is narrower and still bad: more of the people who get stuck are staying stuck for a long time. Meanwhile companies have work that needs doing. That gap is the inefficiency.

The obvious answer, and why it's too easy

In engineering there's this idea that you build the thing that makes your current job redundant, and then you go work on the next level up. We went from horses to cars, cars to electric cars, and now cars that drive themselves. None of those steps ended transportation work. They changed what the work was.

AI is the same kind of shift, and the obvious take is that people get augmented. You get something close to Jarvis, where one person directs a set of tools and produces way more than they could alone. I think that's true. What I don't buy is the jump from there to “AI will do the jobs, so placement matters less.”

Anyone who's actually built with agents knows how fast that falls apart. Take a simple agentic loop in a VM trying to fill out a form. One form changes its fields. Another uses a different component. A third asks something the agent's never seen. Frontier models keep getting better with bigger context windows and more training data on edge cases, and it's still often not enough. You need a person, or a really reliable system, to decide what to do, catch mistakes, and own the outcome. Now scale that from a form to a company, a customer, or a patient. Work isn't a pile of tasks you can finish in isolation.

So I expect most knowledge roles to change rather than disappear. The routine pieces, information processing, first drafts, a lot of communication, get handled by models. People move into the loop as editors and orchestrators: directing the work, QAing it, escalating when something's off, and integrating it into something that actually matters. The ILO's research points the same way. About one in four workers worldwide are in jobs with some exposure to generative AI, and most of those jobs are more likely to be transformed than eliminated. Karpathy's job market visualizer is a fun way to see which occupations are exposed, as long as you treat its AI-generated scores as a way to explore the question and not a list of who gets fired.

Where it actually hurts

This doesn't mean the transition is painless, and the place I worry about most is the entry level. If AI handles the routine parts of knowledge work, what happens to the jobs where people used to learn by doing exactly those parts? Stanford researchers found a 13% relative drop in employment for 22 to 25 year olds in the most AI-exposed occupations. I wouldn't pin all of that on AI, but it's a pattern worth taking seriously. Add in companies that can now hire talent anywhere in the world for work that used to need to be local, and the first rung of the ladder gets a lot harder to reach.

It also forces a harder conversation about education. The New York Fed puts unemployment for recent grads at about 5.6%, which is rough but not a catastrophe. The number that gets me is underemployment: 42% of recent grads are working jobs that don't typically need a degree. That's four years and a lot of money spent preparing for work the market isn't asking for.

I don't think that means degrees stop mattering everywhere. You don't want a lawyer who doesn't understand law or a doctor who can't operate on you. Those fields need real training, standards, and a credential that means something. But for a lot of undergraduate paths, the system is preparing people for tasks that are getting automated, instead of teaching them to use AI to do more than they could before.

Proof of work is the new resume

If the routine work gets cheap, what gets valuable is what you can actually ship. Not how many tokens you burn or how much you spend on AI, but what you finished, whether it's good, whether you can explain your decisions, and whether it did what it was supposed to do. Made money, saved time, solved the problem.

For engineers and designers this is already happening. The good ones plug AI into everything: their design systems, their stack, their MCPs, their day-to-day workflow. They ship faster and at a higher quality, and they can show it. But it's not only tech. I have friends in medicine using AI to process lecture and meeting notes and pull up information fast, which gives them back something like twenty hours a week of manual work. That's the same augmentation, it's just harder to see on a resume.

That's the gap. We're decent at recognizing proof of work for people who build software. We're bad at it for almost everyone else. And if one person with AI can do the work of a much bigger team, their judgment and ability to execute is worth more, which means our ways of evaluating people have to catch up with what people can now produce.

What an efficient labor market would feel like

This is personal for me because of Yara. I think a lot about what it would take to help billions of people get hired. At that scale you can't treat hiring as “send more applications and hope some get through.” AI already made applying almost free, and all that did was bury recruiters in volume and make it harder for good candidates to get seen. More noise isn't the fix. Better matching is.

In my head, an efficient labor market looks like this. Unemployment stays low. Finding the right job takes a couple of weeks, not months of someone's life. Wages are actually livable. Companies find people who can solve their problems, and people find work that fits what they're good at and supports how they want to live. When someone's skills or a company's needs change, moving to the next thing is easy instead of terrifying.

That kind of market changes what people worry about. A few hundred years ago most people's worries were about survival. The point of all this progress should be that the worries of your life are fundamentally different. Months of silence after hundreds of applications shouldn't be one of them.

Better matching alone won't create jobs or guarantee good wages. But placement is a huge part of the problem, and it's one we can actually fix. If AI can help us do the work faster, it should help us figure out who should be doing it, show why they're a fit, and get them there faster. That's the shift I want to work on.