Sean Goedecke’s blog post “LLMs Reward Expertise” hit the top of Hacker News this week with over 1,100 upvotes, and the thesis is hard to dispute: AI coding tools don’t close the gap between novices and veterans. They widen it. The more you already know, the more value you extract from the model.

A 2026 Cortex survey of engineering leaders across North America, Europe, and Asia-Pacific found nearly 90% of teams actively using AI coding tools. The productivity numbers are real — 20 to 55% faster task completion, 30 to 60% time savings on coding, testing, and documentation. But the distribution of those gains is lopsided. Senior engineers capture the bulk of them. Juniors, who lack the mental models to prompt effectively, evaluate output, and catch subtle errors, see far less lift.

The conventional wisdom circa 2023 was that AI would democratize software engineering. The reality, three years in, is closer to the opposite. And the response from much of the industry has been a quiet, self-satisfied shrug — as if this outcome simply confirms that expertise matters, that the cream rises, that the market is working as designed.

It isn’t. The market is strip-mining expertise, and nobody is replanting.

The Junior Developer Is Becoming an Externality

Here is a sentence you hear in engineering leadership meetings now, spoken without irony: “With what Copilot and Claude Code let my seniors do, I don’t really need to hire juniors this year.”

A VP of engineering at a mid-size SaaS company told me this in a hotel bar at a conference in June. He wasn’t defensive about it. He was pleased. His top three engineers were shipping at a rate that would have required a team of eight two years ago. Why dilute that with onboarding costs, mentorship overhead, and pull requests that need three rounds of review?

The math works — this quarter. A senior engineer earning $200,000, augmented by AI, might genuinely produce more than that same engineer plus two juniors earning $90,000 each. The spreadsheet is unambiguous.

But the spreadsheet only has columns for this quarter. It has no column for “where do senior engineers come from in 2032.” It has no line item for “accumulated tacit knowledge that can only be acquired by writing bad code and having someone better tell you why it’s bad.”

Every senior engineer who is now 35% more productive was once a junior engineer who broke production, misunderstood a requirement, and learned architecture by maintaining a mess they didn’t create. That pipeline is not optional. It is the entire supply chain.

The Expertise That LLMs Reward Is Non-Renewable Under Current Incentives

Goedecke’s post focuses on the individual: if you want to benefit from AI, get good first. Fair enough as personal advice. As industrial policy for an entire sector, it is a recipe for eating the seed corn.

The dynamic is not unique to software. Law firms have been quietly running the same experiment. A litigation partner who can use AI to draft motions, summarize depositions, and surface relevant precedent in minutes needs fewer associates. The partner bills more hours at a higher rate. The associate who would have learned to draft those motions by doing them — badly, slowly, under supervision — never gets hired, or gets hired and does doc review while the interesting work stays with the partner and the model.

In both professions, the people who currently possess the expertise are extracting unprecedented productivity from it. The people who would have acquired it through apprenticeship are being routed into narrower, more mechanical roles — or not hired at all. The expertise pool is being drawn down faster than it is being replenished.

This is not a problem the market will solve on its own, because the market prices expertise as a stock, not a flow. A senior engineer’s market value reflects the expertise they have now, not the expertise they will need to train into existence for the industry to function in a decade.

The Companies That Win the Next Decade Will Be the Ones That Still Train Juniors

There is a counterargument, and it goes like this: AI tools will get better, the prompting gap will narrow, and juniors will eventually catch up. Maybe. But the gap Goedecke identifies is not primarily about prompt engineering. It is about judgment — knowing what to ask for, recognizing when the output is subtly wrong, understanding the architecture well enough to integrate the generated code without creating a tangle of technical debt.

Those are not skills you acquire by chatting with a model. They are skills you acquire by working alongside people who already have them, on real systems, with real consequences. If the industry stops creating those conditions for early-career engineers, no amount of model improvement will fill the gap. The models will get smarter, and the seniors will get even more productive, and the juniors will fall further behind.

The firms that resist the spreadsheet logic — that keep hiring juniors even when the quarterly ROI looks worse, that protect mentorship time even when AI makes it feel optional — are not being charitable. They are making a long-duration capital investment that their competitors are too impatient to make. In 2032, when the current cohort of senior engineers starts retiring or moving into management, those firms will have a bench. Everyone else will be bidding up the shrinking pool of experienced talent and wondering where it all went.

Goedecke is right that LLMs reward expertise. The question his post doesn’t ask is whether the industry’s response to that fact will ensure there is any expertise left to reward.

Sources