Horsting argues that generalist engineers fluent in frameworks and cloud tooling are precisely what LLMs replace most efficiently, while engineers who carry deep contextual knowledge — like the 40-page actuarial paper behind an insurance CSV column — remain irreplaceable. He uses his own insurance software career as evidence that AI tools cannot infer regulatory, actuarial, or industry-specific context from surface-level artifacts.
The editorial frames the shift as an inversion of fifteen years of meta-career advice that rewarded going horizontal (React, then Kubernetes, then LLM tooling). With GPT-class models one-shotting CRUD apps in Cursor, the constrained supply of full-stack generalists that justified premium compensation is being commoditized around domain specialists.
The editorial acknowledges Horsting is restating a thesis Joel Spolsky wrote variations of in 2002, but argues the argument lands differently now that LLMs have made the 'everything else' around domain expertise cheap. The novelty isn't the claim that context matters — it's how thoroughly the surrounding skills have been commoditized.
Bret Horsting's essay, "Domain expertise has always been the real moat," hit 382 on Hacker News and surfaced a thesis the comments section couldn't stop chewing on: the generalist software engineer — fluent in three frameworks, a cloud, and a queue — is exactly the profile that LLMs are most efficient at replacing. What they can't replace, Horsting argues, is the engineer who knows that a Medicare Advantage risk-adjustment claim has 21 valid HCC categories, that the SEC's Rule 15c3-3 customer reserve formula needs to be recomputed every Friday, or that a semiconductor tape-out has a specific failure mode when the DRC deck disagrees with the foundry's LVS rules.
The argument is blunt: code is the cheap part now; context is the moat. Horsting points to his own career arc inside insurance software — the kind of domain where a CSV column called `loss_dev_factor_2009` has a 40-page actuarial paper behind it — and notes that none of his AI tools can infer that paper from the column header. The HN thread filled with parallel anecdotes: a defense engineer on ITAR-bounded radar signal processing, a tax-software lifer who described U.S. partnership returns as "a 1,200-page spec written by Congress at gunpoint," a healthcare integrator pointing out that HL7v2 still runs hospital billing in 2026.
The piece is not new in spirit — Joel Spolsky was writing variations of this in 2002 — but it lands differently when GPT-class models can one-shot a working CRUD app in a Cursor window. The shift in 2026 isn't whether domain expertise matters. It's how much of *everything else* just got commoditized around it.
For the last fifteen years, the meta-career advice has been the inverse of Horsting's: go horizontal. Learn React, then learn Kubernetes, then learn whatever LLM tooling is hot this quarter. The reward was real — generalist engineers got priced at $400K TC because the supply of people who could ship a full stack was constrained. That supply constraint is the exact thing LLMs dissolved first. A junior with Claude and a clear ticket can now produce the React component, the Postgres migration, and the Terraform module in an afternoon. The premium for being the human who could do all three has compressed.
What hasn't compressed — and arguably can't — is the premium for tacit knowledge that lives in regulated, adversarial, or physically-constrained industries. Consider the asymmetry. An LLM trained on the open web has read every Next.js tutorial ever written. It has read approximately zero internal memos about how Anthem actually processes a denied prior authorization, how Lockheed structures a CDR package for the DoD, or why your bank's wire-transfer system has a 3pm cutoff that's actually a Fedwire deadline plus 90 minutes of internal float reconciliation. The training data for these domains is locked behind NDAs, paywalls, and the inside of someone's head — and no amount of scale fixes that.
The counterargument from the HN thread is worth steelmanning: domain expertise has *always* been valuable, so what's actually new? Two things. First, the relative value shifted — when generic engineering was scarce, it commanded a premium; now that it's abundant, the scarcity rent moves to whatever is still scarce. Second, the leverage shifted. A domain expert with LLM tooling now ships at the productivity of a small team. The old constraint was that the actuary couldn't code and the coder couldn't actuary; AI collapses one side of that wall, and the side it collapses is the cheaper-to-acquire side.
There's also a brutal corollary nobody likes to say out loud. If you spent the last five years becoming a great generalist engineer, you optimized for the skill that just got the largest efficiency gain from AI — which is the same as saying you got the largest wage compression. The HN thread had several variations of "I'm a staff engineer at a FAANG and I'm seriously considering going to learn claims adjudication." That's not a joke. That's a market signal.
Three practical implications for engineers thinking about the next five years.
First, pick a vertical and go deep before you pick your next framework. The half-life of "senior React developer" as a job title is collapsing. The half-life of "engineer who understands FDA 21 CFR Part 11 and can ship a compliant audit log" is not. If you're inside a domain company already — fintech, medtech, legaltech, defense, energy — the right move is probably not to leave for a bigger paycheck at a horizontal SaaS. It's to spend two years absorbing every weird edge case your domain has and then become the person who codes against that knowledge.
Second, stop building generic developer tools and start building domain-specific ones. The graveyard of YC-backed "AI for X" companies in 2025 was full of teams who shipped a generic LLM wrapper into a vertical they didn't understand. The winners — Harvey in law, Abridge in clinical documentation, EvenUp in personal injury — won on domain depth, not model choice. Your competitive advantage is no longer your prompt engineering; it's your access to people who'll tell you why the prompt is wrong.
Third, interview for the boring industries. Insurance, logistics, manufacturing ERP, claims processing, supply-chain compliance. These are the companies whose codebases LLMs can't autocomplete because the rules aren't on the open web. They pay less than FAANG today. They will pay more than FAANG in five years because they're the last places where engineering productivity actually depends on humans who've read the regs.
The quiet bet here is that the next decade of software hiring inverts the last one: domain experts will hire generalist engineers as commodity inputs, not the other way around. The actuary, the radiologist, the supply-chain planner, the structural engineer — armed with LLMs — become the bottleneck their employers fight over. The full-stack generalist becomes the COBOL programmer of 2035: not extinct, just no longer where the premium lives. Horsting's essay is a useful warning shot. The question to ask yourself this quarter isn't "what framework should I learn next." It's "what industry's ugly details do I actually understand better than a model trained on the public internet." If the honest answer is none, that's the project.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.