- AI and expertise are now deeply linked — senior engineers use coding agents far more effectively than juniors, and the gap is widening fast.
- The relationship between AI and expertise mirrors what happened when calculators replaced human ‘computers’ in the 1970s.
- Only a fraction of new CS graduates have the coding intuition needed to work effectively alongside today’s AI agents.
- Everyone — not just engineers — should develop basic domain knowledge to unlock what cheap, abundant AI assistance can actually do.
AI and Expertise Are Rewriting the Hiring Playbook
The conversation around AI and expertise in software engineering has been loud, but most of it has been wrong. Yes, coding agents are displacing certain kinds of work. Yes, fresh CS graduates are facing one of the worst job markets in recent memory. But the story isn’t simply “AI is replacing developers.” It’s something more interesting — and more uncomfortable — than that.
A widely discussed essay published in May 2026 on Modern Descartes makes an argument that deserves serious attention: the engineers thriving in the age of AI agents aren’t the ones who skipped the hard work. They’re the ones who did it first.
Meanwhile, companies like OpenAI and Anthropic — the very firms building the tools supposedly making junior engineers redundant — are competing aggressively for entry-level talent. That contradiction alone should give anyone thinking about AI and expertise serious pause.
A 50-Year-Old Lesson Nobody Learned
To understand what’s actually happening, it helps to rewind to the 1970s. There was once a literal job title called “calculator” — a human being whose professional value was fast, accurate arithmetic. These people calculated artillery trajectories, optimised ship hull geometries, and balanced corporate ledgers. Then the scientific calculator arrived, and that job ceased to exist.
Today, no serious STEM programme has dropped calculus or linear algebra from its curriculum just because Wolfram Alpha exists. Students still grind through multivariable calculus, differential equations, and statistics — even knowing they’ll reach for a numerical solver the moment they graduate. Why?
Two theories compete here. The first is pure signalling: a STEM degree is a filter for people who can endure four years of hard material, and the actual content matters less than the credential. The second is that the struggle itself builds a kind of mathematical intuition — a mental scaffolding — that makes you dramatically better at using the tools once you have them.
The honest answer is probably both. But the balance is shifting toward the second explanation. And that shift has direct consequences for how we should think about AI and expertise in software development today.
Why Senior Engineers Are Pulling Away From the Pack
The practical evidence is sitting right in front of us. Senior engineers — those with five or more years of writing code by hand, debugging miserable stack traces, and wrestling with architecture decisions that didn’t work — are substantially better at directing coding agents than their junior counterparts. Not marginally better. Substantially.
This isn’t just about knowing more syntax or frameworks. It’s about having internalised what “correct” looks like. When a coding agent produces a plausible-looking but subtly wrong solution, a senior engineer catches it. A junior engineer — or worse, someone who used AI to shortcut their way through their CS degree — often won’t. The gap between AI and expertise levels has never been more visible than it is in these moments.
The essay’s author puts the current threshold for effective AI-assisted coding at roughly the five-year experience mark. That’s the point at which someone has developed enough computing intuition to prompt agents productively, verify their outputs, and push back when something feels off. That bar will only rise as the agents improve.
What this means for the job market is stark. The essay suggests that perhaps half of new CS graduates simply won’t build enough intuition quickly enough to stay competitive. Some senior engineers will eventually fall behind too, despite their head start. The hiring market is already pricing this in: senior engineers are finding jobs with relative ease; entry-level roles have contracted sharply.
The Two-Tier Future of Software Work
What’s emerging looks less like a single developer job market and more like two distinct tiers pulling apart from each other. Understanding AI and expertise as a pairing — rather than AI as a standalone force — is the key to making sense of this split.
At the top, a relatively small group of highly capable engineers — people with deep intuition, strong judgment, and the ability to orchestrate AI tools at scale — will command increasingly high salaries. The competition for this cohort is why OpenAI, Anthropic, Google DeepMind, and a handful of elite startups are still fighting hard for the best junior talent. They’re not hiring for what those graduates can do today. They’re betting on the ones with enough raw ability to get to that five-year threshold in two or three years instead of five.
Below that, a larger, expanding tier of software consultants and lower-end engineers will continue to exist — AI assistance makes them more productive than they’d otherwise be — but salary growth there will be modest. The value of “I can write some code” has been compressed. The value of “I understand what the code should do, why it works, and when the AI is lying to you” has gone up considerably.
This bifurcation isn’t unique to software. It’s playing out across law, medicine, finance, and any other field where AI can now produce plausible-sounding outputs. The people who understand a domain deeply enough to audit those outputs will be essential. Everyone else will be working for them or competing with the AI directly. In every one of those fields, AI and expertise operate as a multiplier — but only when the expertise is real.
Why Everyone Still Needs to Learn the Basics
Here’s where the conversation about AI and expertise extends beyond professional developers. The ability to get genuinely useful results from AI tools — not just surface-level responses — is gated by some base level of domain knowledge in whatever field you’re asking about.
If you think of computers as appliances that do a fixed set of things, you’ll never think to ask an AI to automate something for you. The same principle holds for medicine, tax law, home repair, data analysis — you name it. Cheap, capable AI assistance is available to almost anyone for around $20 a month. But “knowing how to ask” isn’t trivial, and it scales with how much you understand the subject.
The essay sketches out a rough progression worth taking seriously:
- One to two weeks: Enough vocabulary to ask basic questions and understand what kind of answers are possible.
- One to two months: A feel for when to ask, what to specify, and how to frame the problem well.
- Four to six months: The ability to actually check whether the output is correct — using external sources where needed, but with enough context to know what to look for.
That last stage is the one that matters most and takes the longest. Generating an answer is easy. Knowing whether to trust it is the skill. This is precisely where AI and expertise intersect most practically for non-engineers.
The essay gives a clean example from data science: a colleague struggling to make sense of a correlation matrix was told to ask Claude to “make it prettier using NMF” — Non-negative Matrix Factorisation. Clusters immediately appeared in the data. The prompt required knowing what NMF is, when it’s appropriate to apply it, and whether the output made sense. Someone who’d never touched data science couldn’t have done any of that, no matter how capable the AI on the other end.
The Worst Possible Response to AI
There’s a nihilistic streak running through some of the discourse around AI and education — the idea that since AI can do your homework, there’s no point doing it yourself. Use the tools, skip the struggle, get the credential faster.
It’s an understandable reaction to a genuinely disorienting moment. But it’s almost certainly the worst strategy available. The engineers who will thrive over the next decade aren’t the ones who optimised away the difficult parts. They’re the ones who went through the difficult parts and came out the other side with judgment. The relationship between AI and expertise isn’t one of substitution — it’s one of compounding returns on prior effort.
Just as teachers banned calculators in middle school maths for a reason — building the mental arithmetic foundation that makes advanced work possible — the same logic applies to coding and AI today. The intuition you build by struggling through problems manually is the same intuition that makes you effective when the AI takes over the execution. Skip that step, and you’re not faster. You’re just less capable, with a better-looking output you can’t actually evaluate.
The calculator didn’t make mathematicians obsolete. It made mathematicians who understood mathematics more powerful than ever before. The same thing is happening with AI and expertise in software engineering right now — and the window to be on the right side of that gap is narrowing.
Source: https://www.moderndescartes.com/essays/ai_and_expertise/
