A reader wrote Mark Seemann a long letter: no formal CS background, a year of LLM-assisted work, a fairly large TypeScript and PostgreSQL system with APIs and multi-model workflows. Then the uncomfortable part — trying to turn the prototype into a product, he suspects he built something above his own level of understanding. Seemann answers in public, and says up front that his answers are not rigorous.

The reader’s diagnosis is familiar to anyone who has done AI-assisted work at speed:

  • Fixing one error surfaces another; parts of the system behave in ways he does not understand
  • Months of refactoring without ever fully owning the result
  • Sometimes no next step at all without asking another model

Seemann discloses his own position before answering. He has not decided on AI but leans toward disliking it, works with it anyway, and resents it most exactly when it impresses him most. As an economist he does not buy the “new jobs will appear” reassurance — new jobs did appear with the stocking frame and the steam engine, but rarely for the people displaced, and the same argument accompanied China’s entry into the WTO.

The advice is narrower than the question, and better for it:

  • Understand the abstraction layer directly below the one you work in, plus the one above — developers have always built on layers they did not fully understand, and this covers most troubleshooting
  • LLMs do not hallucinate, they bullshit — so use them only where an answer can be checked
  • Ask falsifiable questions (“can this expression be made more succinct?”) and skip unfalsifiable ones (“what should I learn next?”)
  • Human learning does not speed up much: the bottleneck is the brain, not the teachers or the materials

What makes the piece worth reading is that he does not sell the exit. The route he actually took — step back and systematically learn the fundamentals — was viable for a career started in the mid-1990s, and reaching the point where your own earlier confidence looks like Dunning-Kruger took decades. He does not pretend that timeline still exists.

The durable takeaway is the decision rule, not the verdict on AI. Outsource the questions whose answers you can verify within a minute. Keep the ones that shape what you learn next to yourself.