Where Will the Next Seniors Come From?

In April 2025, 37signals (the company behind Basecamp, famously allergic to hiring sprees) opened one junior programmer position. Salary stated in public: $145,849, pitched at the top ten percent of San Francisco junior salaries, for a role explicitly capped at less than three years of professional experience. Two things about that posting are worth an essay. The first is the reasoning behind it. The second is what happened next: over 2,200 people applied for the one job, and David Heinemeier Hansson noted, with visible pride, "No AI screening here." They read the applications by hand.

Hold both facts at once. A company confident enough in juniors to pay senior-adjacent money for potential. And a market so hostile to juniors that two thousand two hundred of them converged on a single opening like it was a lifeboat. Both facts are real, and the gap between them is where this essay lives.

The lifeboat is not a metaphor

The numbers around entry-level tech hiring have gone from anecdote to dataset. SignalFire's talent research, built on tracking career movements across the industry, found that by 2025, new graduates made up just 7% of hires at Big Tech companies, down by more than half from pre-pandemic levels, with startups no better. Their 2026 follow-up reports entry-level hiring at the large tech firms down on the order of 65% versus 2019.

And it is no longer only a tech story. Stanford's Digital Economy Lab, working with ADP payroll data covering millions of workers, published what may become the canonical study of the moment: "Canaries in the Coal Mine?" by Erik Brynjolfsson and colleagues. Its headline finding: workers aged 22 to 25 in the occupations most exposed to AI have seen a roughly 13% relative decline in employment, while older workers in the same occupations held steady. The decline concentrates precisely where AI automates work rather than augmenting it. Their early-2026 update reports the erosion continuing at several percent a year. The canary is not feeling well.

The mechanism is not mysterious, and anyone who codes with these tools daily can describe it from the inside. The tasks that used to be a junior's first two years (the well-scoped ticket, the CRUD endpoint, the test suite, the first-draft anything) are exactly the tasks a frontier model now does in minutes for cents. In the agent essay I measured a machine doing a junior analyst's audit for $0.04. The economic logic that follows is brutal and locally correct for every individual firm: why pay a junior a year's salary to do slowly, and wrongly at first, what the machine does instantly?

Locally correct. Globally insane. Because the industry that reasons this way has quietly misunderstood what a junior is.

A junior is not cheap output. A junior is a senior in year zero.

Here is the syllogism the market is walking into with open eyes. Every senior engineer (every person you trust to review the architecture, smell the bad abstraction, catch the machine's confident nonsense) was manufactured by exactly one process: being a junior, for years, near real work and real consequences, with someone senior watching. There is no other factory. Nobody has ever hired a senior that some other company didn't first grow from seed. The entire senior layer of this industry is the output of an apprenticeship system that we are now, with quarterly precision, dismantling, because the seed-stage work got automated.

The industry is eating its seed corn and calling it margin.

DHH has been the loudest voice refusing this logic, and he is worth quoting in his own words, as promised. On why AI doesn't change the need: "However good AI gets, we're always going to need people who know the ins and outs of what the machine comes up with. Maybe not as many, maybe not in the same roles, but it's truly utopian thinking that mankind won't need people capable of vetting the work done by AI." That word, vetting, should sound familiar to readers of the synthesis essay: it is the judgment address, stated by a man running a profitable software company rather than a philosophy. The machine's output is cheap; the person who can tell whether it's right is not; and that person has to come from somewhere.

And back in 2018, retweeting an essay title he endorsed without reservation: "If you don't hire juniors, you don't deserve seniors". Isaac Lyman's line, DHH's amen. It reads like a slogan. It is actually a supply-chain observation.

Hiring for the slope, not the intercept

What makes the 37signals posting more than a gesture is the selection philosophy attached to it, from the companion post with the wonderful title "Why we won't hire a junior with five years of experience":

"Maybe it sounds a little harsh, but a programmer who's been working professionally for five years has likely already revealed their potential. The trajectory by which they improve has already been plotted."

The whole post is an argument for hiring on slope rather than intercept. A fresh programmer, he argues, is essentially all potential: you draw their line from a few early dots, and the line can be steep. They are, in his phrase, hiring for "peak promise", and paying a junior more than most because they are "looking for a junior who's better than most. Not better in 'what do they already know', but in 'how far could they go'."

Notice what this really is: a bet that the person (their curiosity, their taste under correction, their rate of becoming) is the asset, and the current skill inventory is nearly noise. That bet was always the right way to hire juniors. AI has just made it the only way, because the skill inventory a junior brings ("I can write a React component") is precisely the part the machine commoditized. What the machine cannot commoditize is the trajectory: the thing that, ten thousand corrected mistakes later, turns into the judgment everyone suddenly agrees is scarce.

What the apprenticeship becomes now

I don't think the answer is nostalgia: hiring juniors to hand-write the CRUD the machine writes better, as a kind of historical reenactment. The old ladder's bottom rungs are gone; pretending otherwise produces neither juniors nor seniors, just theater. The honest question is what the new bottom rungs are made of, and I think the shape is already visible to anyone using these tools seriously:

The junior's first job is no longer producing the draft. It is interrogating the draft. Give the junior the machine's output and make them find what's wrong with it, because something usually is, and finding it is trainable. Every pull request the model writes is now a free apprenticeship exercise: here is plausible work, decide whether it's true. That is a better judgment gymnasium than the old ticket queue ever was, because the old queue taught typing with occasional judgment, and the new work is judgment with occasional typing. In the OrbitSpec essay I argued reliability starts before the answer, with the question; a junior who spends two years watching where machine answers fail acquires exactly the reflex the terms of service (see who pays when the AI is wrong) assign to the human: you must verify. Verification is a career now. Someone has to be trained into it.

The senior's job description now includes running the nursery. If the market won't fund juniors doing automatable work, then the case for juniors has to be made on the only ground that survives: they are the firm's future verification capacity, hired at the price of potential, developed on purpose. That costs real senior time. It always did; the cost used to be hidden inside code review, and now it has to be chosen deliberately, which, if I am honest, mostly means chosen by founders and owners, because no quarterly incentive will ever choose it.

And the junior's own move is to sell the slope. If you are twenty-three and reading this: the market is telling you, in the ugliest possible way, that your entry-level output is worth little. It is not telling you that you are worth little. It is telling you to stop selling output. Sell evidence of trajectory: things you built and then corrected, judgments you made and can defend, machine output you caught being wrong and can show. Two thousand two hundred people applied to 37signals with, mostly, similar CVs. The ones who stood a chance were the ones who could demonstrate not what they know but how fast they become.

The address at the end of the ladder

In the four-address picture I keep returning to, this is the person address wearing work clothes. Machines make production cheap; judgment, recourse, and attention move to people. But people are not found, they are formed. Every profession that survived a technological revolution did it by rebuilding its apprenticeship around the new tools: surgeons still train on procedures machines assist, pilots still log hours in cockpits that mostly fly themselves, precisely because everyone understood that the day the machine fails is not the day to start growing your first expert.

Software is currently running the opposite experiment: keep the seniors, skip the juniors, let the machine do the middle, and assume the senior layer of 2036 will somehow exist anyway. It will not. It is being decided now, one hiring freeze at a time, in the entry-level jobs that quietly stopped being posted. The Stanford canaries are not a labor-market curiosity: they are a forecast of who will be left to check the machine's work in ten years, and the current answer is: fewer people than we will need, all of them tired.

Whatever a machine can make gets cheap. What no machine can do is tell if it's right, and the people who can tell are not a natural resource. They are made, slowly, expensively, one hired junior at a time, by companies willing to pay today for judgment that only exists tomorrow. DHH's little job posting was not a hiring event. It was a company betting on where seniors actually come from, and 2,200 applications were the sound of an entire generation agreeing, and waiting for the rest of the industry to remember.