HN Debrief

Replies to comments on my "LLMs are eroding my career" post

  • AI
  • Programming
  • Economics
  • Labor
  • Developer Tools

The post is a response to criticism of an earlier essay arguing that LLMs are eating away at software careers. The author doubles down that this is not just another tool cycle. They claim newer coding agents, better harnesses, and agent-oriented documentation have already made a big jump in real workplace usefulness, that Jevons-style demand expansion will not save software jobs forever, and that the same pattern will move into law, finance, design, and other knowledge work as model wrappers improve. The core claim is blunt: if code production becomes cheap enough, much of today's paid programming gets commoditized.

Treat AI coding gains as real operational leverage now, especially on repetitive implementation work. Reorganize teams around product judgment, system ownership, and verification, because those are the parts people still think remain scarce even if code output gets dramatically cheaper.

Discussion mood

Uneasy and mostly resigned. Many commenters think coding agents are already good enough to squeeze routine programming work, even if the grandest AI takeover claims are speculative or overmarketed.

Key insights

  1. 01

    Programming fits LLMs unusually well

    Because software offers formal syntax, low ambiguity, abundant documentation, and cheap automated tests, coding is a cleaner target for LLM automation than most white-collar work. That framing explains why people are seeing real gains in programming before similar gains in accounting, law, or medicine, and why extrapolating from coding success to every profession is too glib.

    Expect software workflows to change faster than adjacent knowledge jobs. Do not use coding-agent progress alone as evidence that every expert role is about to fall at the same pace.

      Attribution:
    • queenkjuul #1
  2. 02

    Accountability still beats autocomplete

    The durable value case was not "humans know facts." It was that humans own outcomes. People hire a senior engineer or contractor partly because they can say no, absorb blame, test carefully, communicate tradeoffs, and adapt when the situation stops matching the docs. Current models help most when a capable person steers them away from average-case answers and takes responsibility for edge cases.

    Push AI toward draft generation and acceleration, but keep named humans responsible for sign-off on risky systems. In hiring and org design, weigh judgment under uncertainty more heavily than raw coding throughput.

      Attribution:
    • RugnirViking #1 #2 #3
    • altmanaltman #1
  3. 03

    Ticket takers are more exposed

    A sharp distinction emerged between people who turn assigned tasks into code and people who identify problems worth solving, shape requirements, and convert ambiguous pain into projects. If LLMs keep swallowing implementation work, the first group gets compressed hard while the second group becomes the control layer over more automation.

    Developers who mainly execute pre-scoped tickets should move upstream now. Build skills in product discovery, prioritization, system design, and stakeholder management before implementation becomes the cheap part.

      Attribution:
    • alfalfasprout #1
  4. 04

    LLMs create competent-looking impostors

    Several comments pointed at a nasty second-order effect. Models can produce output that passes as competent without transferring understanding to the user. That raises the risk of teams full of people who can ship plausible changes, pass interviews, and still lack the system intuition needed to debug failures or judge when the model is wrong.

    Tighten evaluation around debugging, tradeoff reasoning, and postmortem quality, not just feature throughput. If you rely on AI-heavy hiring or development workflows, add checks for genuine understanding.

      Attribution:
    • grebc #1
    • kristjank #1
  5. 05

    Tooling and process may matter more than models

    One of the strongest pro-automation claims came from someone arguing that benchmark obsession misses the real story. Multi-model pipelines, automated review loops, and agent-specific workflows can push coding performance far beyond what a bare chat interface suggests. If that is true, many teams are underestimating near-term impact because they are evaluating the models, not the surrounding system.

    Audit your stack for workflow leverage, not just model quality. Teams that build reliable agent pipelines may get a much larger productivity jump than teams that simply hand developers a chatbot.

      Attribution:
    • pixel_popping #1

Against the grain

  1. 01

    The inevitability story is still weak

    Several people pushed back on the tone of certainty, not on the existence of progress. They argued that strong claims depend on continued capability gains, continued capital to fund infrastructure, and a stable enough economy to absorb mass displacement. None of that is guaranteed. Current neural-network approaches may improve for years, but that is still very different from proving that broad human replacement is close or inevitable.

    Plan for disruption without betting the company on a single acceleration narrative. Keep optionality in your staffing and product roadmap instead of assuming either full automation or a near-term plateau.

      Attribution:
    • ryanackley #1
    • alfalfasprout #1
    • philipwhiuk #1
  2. 02

    Cheaper software can expand custom demand

    A credible minority argued that commoditized code does not automatically mean less software work overall. Lower development costs can make custom systems economical for smaller businesses that today limp along on generic products. That shifts the work mix rather than simply deleting it, though it may still support fewer traditional developers and more operator-builders.

    Watch for market expansion at the low end, especially in vertical software and bespoke internal tools. The opportunity may move from large engineering teams toward smaller groups shipping narrower, tailored systems.

      Attribution:
    • mexicocitinluez #1
    • rowbin #1
    • jameshart #1
    • waffletower #1
  3. 03

    Some displacement is already visible

    One commenter cut through the abstract forecasting with examples from a security operations manager and a UX designer who lost work after LLM-enabled tooling arrived and had stayed unemployed for months. It is anecdotal, but it grounds the debate in actual labor-market pain rather than distant speculation.

    Track adjacent roles, not just engineering, when assessing AI impact on your org. Headcount pressure may appear first in support functions and design before it shows up clearly in software teams.

      Attribution:
    • jmpman #1

In plain english

LLM
Large language model, a machine learning system trained on large amounts of text that can generate and analyze language and code.

Reference links

AI cost and capability trends

Books and fiction

Essays and media references

  • Taleb's turkey problem explainer
    Linked to warn against assuming past stability predicts future outcomes.
  • xkcd 1425: Tasks
    Used to poke at the tendency to treat AI tasks as solved when they remain hard in practice.
  • xkcd: Physicists
    Referenced to criticize software engineers assuming other domains are easier than they are.

Economic and social context

Examples of AI slop in the wild