TL;DR. Two years ago the loudest voices in tech told you that by now, software engineers would be a historical curiosity. Prompt in, product out. It is August 2026, and developer employment did not collapse. But the same wave that failed to remove developers removed something else, and almost nobody is telling that half of the story properly: it deleted the routine work that turned beginners into seniors. Employment of software developers aged 22 to 25 is down roughly 20% since late 2022, while older engineers in the same occupations held flat or grew. Quantum computing is now running the identical hype cycle one lap behind — with a genuinely historic milestone that nobody outside the field is discussing, and a fake one that is everywhere.
Two things that are both true
If you graduated with a computer science degree this year, you are competing with hundreds of other applicants for a listing that says junior and then quietly asks for three years of experience.
And the profession as a whole is fine.
Both of those things are true at the same time. Most takes pick one and pretend the other does not exist, which is why the discourse has been so useless. The interesting question is not whether AI took the jobs — it demonstrably did not — but why the damage landed so precisely on the people trying to get in.
Why AI hasn't taken developer jobs
Start with what is true, because the AI hype is not nothing. AI assistance is no longer a fringe workflow; it is the default one, and a large share of new code is written with a model in the loop. So why didn't the jobs disappear?
Writing code was never the bottleneck
Think about what you actually do in a working week. How much of it is typing new code? Maybe twenty percent, on a good week. The rest is figuring out what to build, understanding a system somebody else wrote five years ago, arguing about tradeoffs, and being the person who is accountable when it breaks at 3 AM.
BCG put this well in their 2026 analysis: AI can dramatically accelerate code generation and testing, but it cannot own an outcome end to end. And you cannot cleanly split the work into "the model does the code" and "the engineer does the judgment," because in real software those two things are the same activity. The judgment is expressed as the code.
The verification tax
This is the counterintuitive one, and it is the finding we would point at first.
METR — Model Evaluation and Threat Research, a nonprofit that runs studies on what these models can actually do — ran a controlled trial in which experienced open-source developers worked on their own codebases, half the tasks with AI tools and half without. They were 19% slower with AI.
The part that should worry you is not the number. It is that they believed they had been faster. They estimated a speedup that did not exist.
That is the shape of the whole problem. Time saved generating is paid back in review, debugging, and maintenance. It does not disappear — it moves, and it moves somewhere harder to see. We spend a lot of our week inside codebases where that bill came due; it is most of what an AI code audit actually finds.
Speed you can feel is not speed you can measure
The METR result is uncomfortable precisely because the developers were experienced, working in code they knew, and sincerely reported the opposite of what the stopwatch said. Perceived velocity and actual velocity came apart. Any team deciding its AI strategy from how fast the work feels is reading an instrument that is known to be broken.
Cheaper software means more software
Every backlog at every company is a graveyard of things nobody had time to build. Make building cheaper and you do not fire the builders — you finally start clearing the graveyard. That has been the story of every tooling revolution since the compiler, and there is no particular reason this one inverts it.
A lot of "AI layoffs" were not AI layoffs
Companies cut for financial reasons and cite AI, because "we are leading the AI transformation" reads better to shareholders than "we over-hired in 2021 and then interest rates went up." There is a term for it now: AI washing. The tell is in the reversals — organizations that cut loudly on an AI narrative and then quietly restarted entry-level hiring once the balance sheet recovered.
So: AI did not remove developers. It removed a layer of tasks.
Hold onto that sentence, because that layer was the training ground.
The junior developer squeeze
Here is the number that matters. Stanford's Digital Economy Lab, working with payroll records from ADP — Automatic Data Processing, one of the largest payroll processors in the United States — across millions of workers, found that employment of software developers aged 22 to 25 is down roughly 20% since late 2022.
Now look one row down in the same dataset. Across all AI-exposed occupations, employment fell 0.2% year over year. Older workers in those same occupations were flat or growing.
The damage is not spread across the profession. It is concentrated almost entirely on the people trying to get in.
Why it happened is not really about AI being smart
The traditional path into this industry was an apprenticeship in disguise. You got hired, and for a year you did the work nobody senior wanted: fixing small bugs, writing tests, updating docs, building the boring CRUD screen. You were not profitable. Everyone knew you were not profitable. That was fine, because in twelve months you would know the codebase and become useful.
AI ate that exact work. Not the interesting work. The training work.
So companies now face a question they have not answered: how do you onboard a junior when there is nothing routine left to hand them?
It compounds in three ways
The postings lie. Listings labeled "entry-level" have grown while actual hiring into entry-level dropped — companies advertise a junior role and fill it with someone who has five years of experience.
You are not competing with other juniors. The 2024–2025 layoffs pushed a large population of mid-level engineers into the market, and plenty of them will take a role labeled junior.
Carrying a junior got more expensive. Seniors report spending meaningfully more time on review when juniors lean hard on AI assistants. The cost of a junior went up at the exact moment their obvious value went down. That is the cruel one.
What to actually do about it
Stop optimizing for the thing that got automated. Nobody is hiring you in 2026 because you can produce a React component — the model does that for free. The hiring case is the work the model demonstrably cannot do:
- Read a large unfamiliar codebase and explain why it is shaped the way it is
- Debug something in production when the logs are lying to you
- Recognize when the AI output is confidently wrong — this is now the differentiator, and it is the skill juniors are weakest at
- Own an outcome rather than a ticket
And be strategic about where you aim. The open doors are not the ones on your feed. Enterprise, healthcare tech, fintech, government, defense, mid-size SaaS, small businesses — places with old codebases, compliance requirements, and real pain. Candidates who did an internship report markedly better offer rates than those who did not, and that gap is worth more than any framework you could learn this month.
Hard is not the same as closed. But you have to read the market you are in, and most people applying right now are still applying to 2019.
Quantum computing: what to actually expect
The same hype machine is already warming up for round two, so it is worth being inoculated in advance.
The result that actually matters
Google's Willow chip demonstrated below-threshold error correction. In plain terms: for decades, adding qubits added more errors than you could correct, so scaling made things worse. Willow showed the opposite — add physical qubits, and the logical error rate goes down, exponentially.
That single result moves quantum computing out of physics and into engineering. Engineering problems get solved.
IBM is on a public roadmap toward a fault-tolerant machine, Starling, by 2029, with its current Nighthawk processor at 120 qubits running circuits with thousands of entangling gates and a target of demonstrating verified quantum advantage by the end of this year — notably backed by an open tracker so third parties can attack the claims. Microsoft and Quantinuum have logical qubits running with error rates better than the underlying hardware. Real progress, not vapour.
What it does not mean
It does not mean a quantum computer replaces your laptop. Quantum machines are not "faster computers." They are a specialized accelerator that beats classical hardware on a narrow set of problems: simulating quantum systems — chemistry and materials science — plus certain optimization and sampling problems. Your web app will never run on one. There is no quantum advantage for a REST API.
It does not mean encryption breaks next year. Breaking RSA-2048 — Rivest–Shamir–Adleman, the public-key algorithm most of the internet runs on, at a 2048-bit key length — needs thousands of logical qubits, which at today's error rates means millions of physical ones. Current systems have hundreds of physical qubits. That is not a gap you close in a good quarter. The Global Risk Institute's expert survey puts the odds of a cryptographically relevant machine within ten years at roughly 17 to 22%.
The one part that reaches your job
"Harvest now, decrypt later." An adversary can record encrypted traffic today and decrypt it in 2035. So for anything with a long confidentiality lifetime — health records, state secrets, legal archives — the migration deadline is now, not when the machine exists. We walked through the mechanics of that threat — Shor's algorithm, what it does to RSA and ECC, and why agencies are already storing traffic they cannot yet read — in our explainer on encryption.
NIST, the US National Institute of Standards and Technology, has finalized the post-quantum cryptography standards, US federal agencies are working toward a 2030-ish target, and the practical work is unglamorous: inventory where your crypto lives, and move to hybrid key exchange. If you are still auditing where TLS terminates across your estate, that inventory is the same piece of work.
That is the quantum action item for a normal developer in 2026. You do not need to learn Qiskit to keep your job.
And when you read the next headline, watch the right metric. Physical qubit count is marketing. Logical qubit count, error rate, and circuit depth are engineering. Treat every "quantum advantage" claim as provisional until classical researchers have spent six months trying to beat it on a laptop — historically, a lot of them succeed.
What about AI running on quantum computers?
It is the least real thing in this entire discussion, and the arrow points the other way round.
The input problem. To run a quantum algorithm on ordinary data, you first have to load that data into a quantum state, and for generic data the loading can cost as much as the computation. Training a language model is overwhelmingly about moving data around. Quantum is worst at exactly the thing that dominates the job.
Training does not scale. There is a phenomenon called barren plateaus: as you add qubits to a deep parameterized circuit, the gradients shrink exponentially. The loss landscape goes flat and there is nothing left to descend. That is not a footnote — it is the central open problem of the field.
Dequantization. Since Ewin Tang's 2018 result — she was an undergraduate, and she killed a famous quantum speedup by writing a classical algorithm that matched it — researchers keep finding the same thing. Give a classical algorithm the same data access the quantum version quietly assumes, and the advantage evaporates. Dequantization did not kill quantum machine learning. It drew the boundary around it.
What is inside that boundary is real but narrow: data that is quantum to begin with, plus a handful of kernel methods with a proven separation. Nothing that looks like a language model.
And here is the part we find genuinely funny. In April, NVIDIA shipped a family of open models called Ising. What do they do? Calibrate quantum processors and decode quantum error correction — substantially faster and more accurate than the traditional approaches.
Read that again. AI is being used to build quantum computers, not the other way round.
Gartner's near-term line is blunt: no enterprise AI workload running at scale on quantum hardware through 2028, and even fault-tolerant systems will not have the logical qubits for economically viable end-to-end AI by 2030.
One warning: "quantum-inspired" is a marketing term for a classical algorithm running on classical hardware. When a vendor claims twenty-times-faster AI training from quantum optimization, check whether a single qubit was involved. Usually not one.
The pattern
The AI hype was wrong about developers being obsolete — and almost nobody noticed it was right about something worse and quieter, which is that we broke the ladder that turns beginners into seniors. We solved the training-work problem by deleting the training work.
Quantum is running the same play. The real milestone — error correction below threshold — is historic, and almost nobody outside the field is talking about it. The fake milestone — encryption dies next year — is everywhere.
Learn to tell those two apart and you will be right about technology more often than most people who work in it.
If you are running a codebase where a lot of the code arrived faster than anyone could review it, that is the exact situation the verification tax describes — and it is what an AI code audit is for. We also wrote up how we build with AI and the eleven problems we find in almost every AI-built codebase.
And for anyone trying to get in right now: the ladder is broken, but the building is not. Aim at the work a model cannot do, and at the companies whose problems are old, unglamorous and real.
Ilya Nixan is Founder & Lead Developer at Nerdy Production, a Flutter-first agency that builds and maintains apps across fintech, healthcare, and retail.