QWE AI Academy The #1 mistake people make after reading a post like “LLMs are eroding my software engineering career and I don’t know what to do” is treating it as a feelings post. It’s not. It’s a diagnostic – and if you read it carefully, the prescription writes itself.
The post dropped on June 6, 2026 and was the top story on Hacker News within a day. If you’re a working developer and it hit a nerve, this is a hands-on playbook for what to do this week, not a pep talk.
The Post Everyone’s Sharing – and What It Actually Says
The author is a backend engineer with 10 years of professional experience, who moved from frontend into backend and ended up specializing in finance, bookkeeping and payment processing – PCI compliance, double-entry ledgers, escrows, reconciliation, payment lifecycles, bank transfer idempotency. That’s a deep, expensive specialization. He thought it was his moat.
Then his company changed how it hired. Job listings went from “Software Engineer – Area” to just “Software Engineer,” with team assignment happening after the offer is accepted. Domain expertise stopped being a differentiator. And in his own words, all the finance, debugging, and distributed-system knowledge he earned through hours of sweat and tears is now promptable.
Read the second pillar he names carefully – it’s the actual tell. He says nobody needs A or B-grade codebases anymore because they’re being made for LLMs, not for humans to read. That’s not a claim about model capability. That’s a claim about a quality bar his employer chose to lower.
Why “Just Upskill into AI” Is Bad Advice
The standard tutorial response to this post is some variation of “learn system design, learn MLOps, become a generalist who steers AI.” That advice ignores the actual market data – and the data is genuinely contradictory.
Nadella said 30% of Microsoft code is now written by AI. Over 40% of Microsoft’s May 2025 layoffs targeted software engineers. Token costs for AI inference dropped 280-fold over the past two years (as of mid-2026), per tech-insider.org – meaning the economic case for delegating coding to models gets stronger regardless of whether the models themselves improve. Enterprise spending on AI dev tools hit $12 billion in 2025 and is projected at $18 billion in 2026. The pressure to delegate is structural, not cyclical.
Flip it over: WhatJobs analysis citing Challenger, Gray & Christmas data shows only about 5% of the 1.17 million 2025 tech layoffs were directly attributed to AI automation – most were overhiring corrections that used AI as a narrative cover. A separately cited MIT study found 95% of companies that adopted AI haven’t seen real productivity gains. LinkedIn data shows new software engineering postings declined 15% in early 2026 versus 2025.
Here’s what’s actually happening: those two clusters of statistics describe the same market from opposite sides. The 30%-of-code headline is supply-side – what companies are shipping. The 5%-of-layoffs headline is demand-side – what they’re telling HR lawyers. Reading only one gives you a wrong career bet. The market is bifurcating, not collapsing, and generic “learn AI” advice tells you nothing about which side of the split to land on.
Think of it like the shift from film to digital photography. Most photographers who “just learned Photoshop” ended up doing the same commodity work for less money. The ones who survived either moved up into direction and creative strategy, or moved into niches – forensic photo analysis, archival work – where a digital camera still can’t stamp a legal chain of custody. The analog skill didn’t vanish. It just stopped paying in the commodity lane.
The Repositioning Playbook (Do This in 7 Days)
The original author’s pillars eroded specifically because his work didn’t have external accountability. When the only person checking quality is an internal manager, the quality bar can be dropped overnight. When the checker is a regulator, an auditor, or money moving between banks – it can’t. That asymmetry is the whole game.
The job isn’t to escape LLMs. It’s to attach yourself to work where someone outside your company will catch the mistakes.
Day 1-2: Audit your current work
List every task you did last sprint. For each one, ask: who notices if this is wrong? If the answer is “my manager,” that task is at risk. If the answer is “a regulator, an auditor, a paying customer’s CFO, or a security review board,” that task is your moat. Circle the second category.
Day 3: Pressure-test an LLM on your domain
Open Claude or ChatGPT and ask it to generate code for the most regulated thing you’ve ever shipped. Then have it explain the compliance reasoning. Look for hallucinations on purpose.
You are a senior payments engineer.
Write the SQL and application logic for a chargeback
reconciliation flow that satisfies PCI-DSS 4.0 and
Reg E (US) for a $250 disputed transaction.
Cite the exact regulation clause for each control.This isn’t busywork. A fintech engineer in the HN thread reported that the most bleeding-edge model hallucinated what a regulation actually required – and they only knew because the code had already been reviewed by human counsel. Using these tools to ship compliant financial products unsupervised would be totally mad. You want to be able to spot exactly that failure mode, on demand, in an interview.
Day 4-5: Rewrite your résumé around accountability
Replace “Built X using Y” bullets with bullets that name the external checker. “Designed escrow ledger reviewed by Big-4 audit with zero findings” beats “Designed escrow ledger in Python.” The first one is unpromptable. The second one is a Cursor session.
Day 6: Pick one adjacent domain with regulatory gravity
Healthcare (HIPAA), payments (PCI/PSD2), aviation (DO-178C), automotive (ISO 26262), pharma (21 CFR Part 11), defense (ITAR). Pick the one closest to what you already do. Read one official spec end-to-end. Yes, the actual PDF.
Day 7: Ship one public artifact
A blog post, a GitHub repo, a conference CFP submission – something with your name on it that demonstrates judgment a model can’t fake. Sean Goedecke nails the underlying reason this works: an LLM has no skin in the game, so trust has to be built purely on track record. Executives trust engineers because they know engineers will face consequences if they get it wrong – their bonus, promotion, or in the extreme case being fired. Your public track record is the asset that outlasts any particular model generation.
Rewriting the Author’s Own Situation
The author has payments domain depth. He writes Design Docs good enough that his company wanted PMs to be able to read them. Under the “learn LangChain” playbook, he grinds through a skill that’s commoditizing even faster than his current one.
Under this playbook, the move is the opposite: double down on payments-with-external-accountability. Reposition from “backend engineer who knows payments” to “payments engineer who can defend an architecture to an external auditor.” Productize one Design Doc per quarter as a public artifact – anonymized, with the trade-offs and the regulatory citations spelled out. Move toward roles where the buyer is a CFO or a compliance officer, not a VP Engineering chasing velocity. And document every time his current employer drops the quality bar – in writing, in the PR description – so that when something breaks, the decision trail exists.
Notice none of that requires learning a new framework. The expertise is already there. The repositioning is about changing who can see it and who it’s accountable to.
Pro Tips From the Trenches
Pro tip: If your manager has ever said “that’s good enough, ship it” about code that touches money, identity, or safety, you are not safe – your manager is the one who’ll be replaced first, and your work will be re-graded by someone stricter. Get ahead of that by writing the stricter review yourself, today, and putting it in the PR description.
Stop treating “taste” as a defensible skill. The original post’s author lists code-quality taste as his last standing pillar. Taste without an external buyer is a hobby. Taste that maps to a regulation, a SOC 2 control, or a contractual SLA is a job.
Also worth sitting with: token costs dropped 280-fold as of mid-2026, and there’s no reason to expect that curve to flatten. If inference is effectively free by 2027, the economic argument for delegating another layer of coding work gets made automatically – no new capability needed. Skill choices made today should probably assume that’s the baseline, not a stretch scenario.
The junior pipeline problem deserves its own honest framing.Stephanie Kirmer’s widely-shared argument is that senior engineers are sought after because they solve problems juniors and LLMs can’t – but they start out as juniors. Hollow out that entry layer and there’s no pipeline. That’s a real structural problem for the profession. Individually, though, you still need to not be in the hollowed-out layer right now. The pipeline argument is cold comfort if you’re the one being thinned out.
What does it mean for a skill to be “safe” in a market that’s repricing this fast? Probably not what it meant three years ago. External accountability is one anchor. Public track record is another. But the honest answer is that nobody knows which specific roles survive five years of models that are 280x cheaper to run than they were two years ago. Anyone selling certainty here is selling something else.
FAQ
Is the viral post overreacting?
No. The author is describing a specific, observable shift inside his company – not making a general claim about the industry. The overreaction would be generalizing from his single data point to “all of software engineering is dying,” which the post itself doesn’t actually claim.
Should I quit and go back to school for an ML PhD like the author considered?
Probably not. The post floats this idea and then dismisses it – frontier labs are flooded with applications, and a PhD is a 4-6 year bet on a market that’s repricing every quarter. The repositioning playbook above is a 4-6 week bet you can run in parallel with your current job. Most working engineers with domain depth are better served doubling down on that depth in a context with external accountability than starting over from scratch in ML, where the competition is people who’ve been doing nothing else.
What if I’m a junior and don’t have 10 years of domain depth to reposition with?
Harder, honestly. Pick a regulated domain on day one and refuse to leave it. Take the boring fintech job over the exciting consumer-app job. Get your name on a compliance audit, a security review, or a published threat model before your second year. The layer the market is thinning is specifically juniors who skipped the “take a beating from external reviewers” stage. One concrete next action: open your last three PRs right now and ask, for each one, “who outside my company would have caught it if this was wrong?” If you can’t name someone for any of the three, that’s your week’s homework – not another framework tutorial.