A 10-year payments engineer published a post on Friday called “LLMs are eroding my software engineering career and I don’t know what to do.” It hit the top of Hacker News within hours — 602 upvotes and 551 comments by the end of day one.
The post resonated because the numbers behind it are real. Software-development job postings on Indeed are down 68.8% from their February 2022 peak. US programmer employment fell 27.5% between 2023 and 2025, according to IEEE Spectrum. Q1 2026 logged 52,050 tech layoffs — the highest first-quarter total since 2023.
But the post is only half right. The same data set that shows “programmer” roles collapsing also shows AI-skilled engineers getting hired 2.3x faster than the average. The skill ladder is not melting. It’s being rewritten — fast — and the people who notice early are the ones who keep moving up it.
What the post actually argues
The author lists three professional advantages he says LLMs have commoditized:
- Domain expertise. Payments, PCI compliance, idempotency, double-entry ledgers — knowledge that took years to internalize is now “promptable” once it’s been written down anywhere.
- Distributed debugging. Race conditions, lock contention, retry storms — he estimates Claude 4.5+ with MCP tooling now one-shots roughly 90% of bugs that used to take a senior engineer a day.
- Architectural taste. Once a senior differentiator. Now, in his words, “dismissed as mere taste” because the model can produce a reasonable design from a half-decent prompt.
His core line: “I have no domain expertise that another Sr. engineer steering an LLM cannot match.”
That is the part that landed. Half of HN was saying yes, exactly. The other half was saying you are describing the easy 80% of your job, not the hard 20%. Both halves are right.
The data the author leaves out
Recent labor stats tell a more complicated story than “engineers are being replaced.” They tell a story of redistribution.
| Metric | Change | Source |
|---|---|---|
| Software dev job postings (Indeed) | −68.8% from Feb 2022 peak | Indeed Hiring Lab |
| US programmer employment | −27.5% (2023→2025) | IEEE Spectrum / BLS |
| Entry-level dev roles (ages 22–25) | −20% from late-2022 peak | Indeed / BLS |
| AI skills in job descriptions | 8% → 42% (2022→2026) | Indeed Hiring Lab |
| AI/ML engineering postings | +50–100% YoY | LinkedIn Workforce Report |
| Senior engineers with AI skills | hired 2.3x faster | LinkedIn Workforce Report |
| “Write code” roles | −27.5% | BLS |
| “System design” roles | −0.3% | BLS |
Why it matters: the headline number (“programmers down 27.5%”) is doing a lot of work that the data does not actually support. Roles labeled “programmer” are collapsing. Roles labeled “software engineer,” “system designer,” or “platform engineer” are barely moving. The job is not disappearing. The title is.
Where the post overstates the case
A few claims in the original deserve a sharper edit.
“90% one-shotted.” Even taking it at face value, that is the success rate on bugs an engineer has already framed correctly enough to paste into a model. Framing the bug is the work. Anyone who has spent two hours figuring out the right four sentences to put in front of Claude knows the framing is most of the labor — and the model does not do that part.
“Domain knowledge is promptable.” Half-true. The named concepts of PCI-DSS or escrow flows are absolutely in any frontier model’s training data. The political knowledge — who at Stripe owes whom a favor, which regulator in which country actually enforces a rule, which legacy ledger inside your own company silently rounds the wrong way — is not. That is the part that stays scarce.
“Architectural taste is dismissed.” Only by people who do not have to maintain the architecture. A model can output a reasonable design. It cannot tell you that your team will quit if you adopt event sourcing, or that your CTO has a personal vendetta against gRPC. Taste is partly judgment about people. Models do not see the org chart.
Sam Altman, of all people, put the honest version on the record in early June: “There’s some AI washing where people are blaming AI for layoffs that they would otherwise do.” That is closer to the truth than the dystopian framing.
What LLMs actually do to a senior engineer’s day
Recent developer surveys point in one direction. AI-coding-tool favorability has fallen — from 77% in 2023 to 60% in 2026 (Stack Overflow Developer Survey). Only 33% of developers say they trust AI code accuracy. And 63% report spending more time debugging AI-generated code than they used to spend writing the same code by hand.
Translation: the model is fast at producing code and slow at producing correct code in unfamiliar contexts. The bottleneck moves. It used to be typing. Now it is verification, integration, and ownership.
Anthropic’s June 5 disclosure that 80% of merged production code at the company is authored by Claude is the most-cited number in the discourse. But it comes with a qualifier the headlines ignore: every line of that code is reviewed, edited, and accepted by a human engineer who knows the system. The model is the junior on the team. It is not the team.
What this means for your career — practically
If you are a working engineer, three things are true at once:
- Pure code-typing is being commoditized. Anything that looks like translating a clear spec into clean code is the part the model genuinely is good at.
- Judgment work is appreciating. Framing problems, reading ambiguous requirements, owning incidents, and saying no to bad ideas have moved up the value chain.
- Speed of adoption is becoming a hiring signal. Engineers comfortable wiring up MCP servers, agentic workflows, and eval pipelines are getting hired noticeably faster than ones who are not.
Skills to deepen
- Reading and writing specifications that an LLM cannot infer on its own — security boundaries, data contracts, regulatory edge cases.
- Debugging the model itself: eval harnesses, prompt regressions, retrieval failures, tool-call faults.
- System architecture that survives a year of change, not a week. Models still cannot see two quarters ahead.
- Code review at LLM scale — reviewing 8x more PRs per day is the new craft, per GitHub’s 2026 developer trends report.
Skills to deprioritize
- Hand-rolling boilerplate frameworks you can pull from a prompt.
- Cramming framework trivia that the model will summarize in a second.
- Defending architectures by aesthetic preference alone.
Signals to send
- A GitHub history that includes at least one agent or eval project.
- Public writing — even a single Substack post — about a real LLM failure you debugged.
- The phrase “AI” on your resume, in context, not as buzzword padding. The 8% → 42% job-description shift cuts both ways.
For more on this transition, see our analysis of the Anthropic 80%-Claude-merged disclosure and our coverage of the 2026 layoff wave in tech.
What to watch next
- Whether the 2.3x hiring multiplier holds for senior AI-skilled engineers, or whether the wage premium compresses as more people pick up the skills.
- How Anthropic, OpenAI, and Google publish their internal merge ratios through Q3 2026. Each disclosure resets the discourse.
- Whether the “programmer vs system designer” split shows up in formal BLS reporting, or stays a labor-economist artifact.
- The next viral post — every cycle has produced one. The follow-up post by the same author already landed on HN within 48 hours.
The post that started this week is going to be remembered less for its argument and more for the response it generated. The honest reading is somewhere between “this changes everything” and “this changes nothing.” LLMs are not eroding software engineering. They are eroding one specific shape of software engineering — the shape that looked a lot like a translation job. Everything above that line is appreciating, fast.
FAQ
What was the original “LLMs are eroding my career” post?
A blog post by a 10-year finance and payments engineer, published June 6, 2026 on a Bear blog called the human in the loop. It hit the top of Hacker News within hours with 602 upvotes and 551 comments on day one.
What three skills does the author say LLMs commoditize?
Domain expertise (e.g. PCI compliance, payments), distributed-systems debugging (race conditions, lock contention), and architectural judgment. He argues all three are now “promptable” by another senior engineer steering an LLM.
Is the data behind the post accurate?
The headline numbers (68.8% drop in dev postings, 27.5% drop in programmer employment, 52,050 Q1 2026 layoffs) are real and verifiable through Indeed Hiring Lab, IEEE Spectrum, and BLS reporting. The interpretation — “engineers are being replaced” — is more contested.
Are software engineers actually being laid off because of AI?
Some. But Sam Altman himself acknowledged in early June 2026 that some layoffs blamed on AI are layoffs companies would have made anyway. The honest framing is that AI is accelerating a correction that was already coming after 2021–2022 over-hiring.
Which roles are growing right now?
AI/ML engineering roles are up 50–100% year over year. Senior engineers who list AI skills get hired 2.3x faster. “System design” employment is down only 0.3%, vs −27.5% for “write code” roles.
Is “90% of bugs one-shotted by Claude” accurate?
It is the author’s personal estimate, not a published benchmark. Even taken at face value, it counts bugs after a human has framed them well enough to paste into the model. Framing the bug is most of the work.
What about Anthropic’s 80% Claude-merged-code claim?
Anthropic disclosed on June 5, 2026 that 80% of merged production code inside Anthropic is authored by Claude. Every line is still reviewed and approved by human engineers. The model is acting as a very productive junior — not as the team.
What should I actually do about all this?
Three moves: (1) double down on judgment work — specs, reviews, architecture, ownership; (2) get fluent with agentic tooling and evals so you show up on the right side of the 8%→42% job-description shift; (3) ignore the apocalyptic framing in the discourse and watch the labor data.
Want the next dispatch on AI-augmented engineering as it lands? Subscribe to the TechDaily360 weekly briefing.