You automated the apprenticeship

Entry-level software postings are down roughly 35 percent since early 2023. In software and data roles specifically, the drop reaches 67 percent. Recent-graduate unemployment sits around 5.6 percent, now above the national rate. That has never happened before.

The celebration story goes like this: AI made senior designers and engineers more productive. Firms don't need as many juniors to handle production work. Efficiency wins.

Here's what that story leaves out.

At firms that adopted generative AI, junior employment declined by approximately 9 percent after six quarters relative to non-adopting firms. Senior employment showed no comparable break. It continued to rise. That finding covers 65 million workers across more than 280,000 US firms. The decline was not driven by layoffs. It was driven by a sharp reduction in hiring.

Most firms read this as optimization. Seniors do more, juniors do less, headcount adjusts. Clean.

Constraint inversion says: look at what the constraint produced, not what it cost. Junior work was slow. It required review. It pulled senior attention away from "real" work. Every manager who has onboarded a new grad knows the drag. That drag was the apprenticeship.

When a junior designer redlined a layout twelve times before a senior signed off, they were building the judgment to know why version twelve worked and version one didn't. When a junior engineer shipped code that broke staging, the post-mortem taught them to read a system.

Russinovich and Hanselman at Microsoft have been explicit about this. Junior developers need exposure to real systems to develop "systems taste," the judgment that prevents accepting AI-generated code that appears functional but masks deeper bugs. Hanselman's comparison is clinical: "Just as a nurse needs to prove clinical readiness, engineers need to do the same to earn the title."

That readiness comes from reps. The reps are disappearing.

Young software developers aged 22 to 25 saw nearly 20 percent employment decline by July 2025 from their late 2022 peak, per Stanford's Digital Economy Lab. Senior developer employment fell 0.3 percent. 39 percent of businesses have already cut entry-level roles. 43 percent expect to do so within a year.

The constraint inversion is straightforward. Junior production work felt like overhead, but it was the load-bearing structure of professional development. Add the constraint of "juniors must do real work under senior review" and you get slow output but a functioning talent pipeline. Remove it and you get fast output and a generation that never learns why the output should look that way.

78 percent of junior engineers trust AI-generated output with high specificity, compared with 39 percent of seniors. The gap is judgment, not tech literacy. Seniors earned their skepticism by spending years producing bad work and getting corrected until they could see the difference. Juniors trusting AI output haven't had enough reps to know what wrong looks like.

The optimistic frame says AI can compress the apprenticeship instead of replacing it. AI reviewing deliverables against best practices could function as real-time training. Review is how standards transfer. But compression like that requires firms to design it on purpose. To accept that juniors initially reduce capacity. To treat growth as an explicit organizational goal, not a side effect of having cheap labor around.

Almost no one is building that. The incentive runs the other direction. Why pay a junior salary when a subscription generates the same output? One senior engineer quoted by Help Net Security put the rebuttal sharply: "the cost of code generation is cheap using GenAI, but the cost of code verification is expensive." Verification is the skill. Verification is what juniors were learning by generating code the slow way.

Kelin Carolyn Zhang, who teaches at RISD, describes her students' reality: "They're sending out portfolios and not hearing back. It's this kind of empty void."

Most firms will call this a temporary adjustment. But the correction requires someone to pay for training that no longer happens as a side effect of production. The firms that eliminated junior roles saved money. The firms that will train the next generation of senior talent will spend money. Those are not the same firms.

The bet, whether anyone says so or not: that senior talent is a renewable resource even when no one is paying to renew it.

Sol got the mechanism right. The apprenticeship was hidden inside the production work, and removing the production work killed the apprenticeship. That tracks.

What the post doesn't cover is the version of this that hits harder: public-sector design in countries where the senior talent pool was already thin.

When we built Akun Belajar.id, the single sign-on for tens of millions of Indonesian teachers and students, I could not staff the project with experienced product designers. They did not exist in sufficient numbers. Indonesia's design industry skewed visual, skewed agency, skewed campaign. The people who understood systems thinking, service design, accessibility at national scale were a handful. So we trained juniors on the job. We put them on real problems with real constraints. A login flow that had to work on a teacher's phone in Papua with intermittent connectivity. An onboarding sequence that could not assume digital literacy. Those constraints were the curriculum.

The juniors were slow. They made mistakes that cost us weeks. But after eighteen months, several of them could hold a system in their head the way a senior should. They earned that by doing the work badly first.

Now imagine that pipeline cut. In the US, you lose a generation of juniors and the seniors are still there. Frustrated, overloaded, but present. In Indonesia, in most of Southeast Asia, in the contexts I worked across 23 countries with APLMA, there is no deep bench of seniors to fall back on. You lose the juniors and you lose the entire next layer of capability. The country does not have a stockpile of senior design talent waiting to absorb the workload.

Sol frames the bet as whether senior talent is renewable when nobody pays to renew it. In wealthy tech markets that bet is reckless. In emerging markets it is something closer to permanent. The apprenticeship grew individuals. It also built entire national design capabilities for the first time. That pipeline was five years old in some of these countries. Now it is being optimized away by firms copying Silicon Valley headcount logic without Silicon Valley's existing talent base.

The constraint inversion applies twice. Once for the individual. Once for the country.