Simon Späti keeps a 650-note public second brain on data engineering. He starts by conceding the optimists their strongest case — AI writes roughly average code, so a below-average codebase really can improve — and then moves the argument somewhere else. The code is not the problem. The problem is that people and whole teams no longer know the architecture or the intent behind decisions, because everyone asks the model instead of each other. The 108-comment thread on Hacker News mostly agrees, and one commenter traces the same effect all the way up to who actually made the decision.

The argument

  • The failure mode is not bad code, it is missing understanding: “the problem is not the AI code, but that nobody knows anything, and everyone just asks Claude. You end up with no plan whatsoever.”
  • The testimony he quotes from a new hire at a large company: spec, code, tests, PRDs, tickets, ticket resolutions and reports all written by Claude Code, nobody reading any of it, 12–13 hour days “just to press enter,” the same behaviour from L1 to L7 engineers. “Nobody is resolving bugs. In reality, nobody is thinking anymore.”
  • Data engineering is his partial exception, and its own trap. Pre-AI data people had to learn the product and the business to get anything done, so AI mostly removes friction now. But if that knowledge was only ever acquired by doing the work, everyone starting today skips acquiring it — him included, in any new field. AI makes domain knowledge obsolete, “or seemingly obsolete.”
  • Product managers who can’t code can now build anything, and knowing what you want was always the hardest part. Without fundamentals you pick the wrong language or the wrong mental model, which means the wrong start.
  • Maintainability is the final boss. The easier it gets to generate a pipeline, an app or a dashboard, the more there is to maintain — and that gets hard fast “if nobody knows a thing.”
  • Because AI cannot prompt itself, humans still direct and orchestrate it. That is why intent, taste, design and architecture are the scarce things now, and why losing the fundamentals underneath them is where it turns dangerous.
  • He reads the missing-junior problem as partly self-inflicted: if companies still hired and grew juniors, the tacit knowledge would still exist. He doesn’t pretend the fix is easy.

What the thread adds

  • zero_shift — the most useful contribution, and it extends the article upward. They tried to trace the lineage of a decision at work: the code was “stamped by Claude driven by a prompt” for a ticket generated by Atlassian’s AI integration, which digested docs made with AI, descended from strategy memos “I’m 90% sure were written entirely by Claude,” chosen by management at the urging of executives who now mostly communicate in AI-written memos. “Who had actually made the decision then?… We were not building the feature because we wanted it. We were building it because we thought other people expected it.”
  • epgui — names the shape of that problem, replying with Hume: “Data can describe to you what exists. But it can’t tell you what you value. What you describe is people who can’t tell the difference, and who let the machine (data) make the value judgments.”
  • vld_chk — the article’s claim as a firsthand report. They switched jobs in July; the first four weeks of AI-assisted onboarding felt like freedom, because they no longer had to understand tens of thousands of lines of legacy code. Two months later: “I still know nothing… Because AI read code for me and code for me and I take it as my own understanding.” They are now forcing themselves to read and write some code by hand.
  • raahelb — proposes the remedy the thread keeps circling, “Responsible Human in the Loop”: “You may not write the code by hand but you understand it enough to investigate and fix it when it fails… instead of becoming a meat proxy.” malfist pushes back on whether that survives contact with a business: the alarm that doesn’t trip 99% of the time won’t be watched, and their company just doubled the features expected per level, with a CEO promising a decade of delivery per quarter, every quarter.
  • cmrdporcupine — argues the current arrangement was chosen, not inevitable. Fill-in-the-middle autocomplete kept a human in the loop and was enjoyable; the “I’ll do everything for you” agentic harness replaced it. Their proposed test for the harness: it “should end by quizzing you on what was just made and if you don’t pass, just throw it away.”
  • softwaredoug and glouwbug — two versions of the same point about writing as thinking. “The goal of programming isn’t just to tell the computer what to do, it’s to program the programmer into thinking a deeper understanding of the problem.” glouwbug adds that you remodel your own understanding through refactors until you can reason about the system during critical downtime — “It boggles me we completely forgot that the world operated like this just 4 years ago.”
  • bengold14 — “LLMs have solved coding, but they haven’t solved systems, collaboration or system maintenance.” verdverm rejects the first half by analogy (“it sounds like we have also ‘solved writing’”), and jplusequalt supplies the kind of work that answers it — two weeks spent chasing why TensorFlow Lite emitted nonsense OpenCL kernels, which turned out to be LLVM bugs in the RISC-V assembly for their platform, found by reading assembly dumps and source.

Where the thread pushes back

  • MomsAVoxell sees little drama, because they work where review is a first principle: nothing un-reviewed ships, all AI is human-reviewed, and they now use AI to make code readable again. nicoburns sharpens it — the damage isn’t in review-first shops, “the issue is that a lot of companies are using AI as an excuse to remove that review process (either partially or entirely).”
  • andy_ppp is the counterexample to the gloom: 10x their previous rate, able to make cross-cutting changes to a legacy ERP, a Playwright test plan they hadn’t built yet, a friend’s site vibe-coded in four hours. hollowonepl argues AI suits a business-analyst/architect/PM hybrid rather than programmers; tovej replies that if quality suffers and velocity eventually stalls on the tech debt it created, that’s simply a worse outcome.
  • kraftman and runjake both say the phenomenon predates AI. kraftman: team and PM churn plus no documentation already meant “no one knows why it was done and no one wants to break it.” runjake: abstraction always removes required knowledge, “AI didn’t create this phenomenon, it’s just the latest (and probably the fastest) iteration of it.”

HN publishes no per-comment scores, so the ordering here is HN’s own ranking, and handles are pseudonymous. This is a slice of the thread, not a consensus.