A Self-Taught Builder Hits the Wall: Seemann on Programming After LLMs

Mark Seemann publishes his replies to a reader who, with no formal CS background, used LLMs to assemble a substantial TypeScript system spanning APIs, PostgreSQL, and multi-model pipelines. The reader’s problem is now familiar: turning the prototype into a production product means fixing one AI-generated error only to trigger the next, and eventually confronting the fact that he has built something above his own level of understanding. When it works, the comprehension gap is invisible; when it breaks, ownership evaporates and every next step requires asking another model.

Seemann answers from an openly skeptical stance — he leans toward disliking LLMs while conceding they may be unstoppable, and says the tools bother him most precisely when they impress him. Drawing on his background as an economist, his larger worry isn’t personal but structural: if AI drives knowledge-worker unemployment to 30–40%, he doubts society absorbs it gracefully. The reassurance that past technological shifts always created new jobs is only half true, he argues — coal miners didn’t become programmers, and the jobs created after China joined the WTO largely didn’t land in the West.

On the practical questions, he’s glad he learned fundamentals across three decades before LLMs existed, and says a strong foundation is what now lets him ramp into unfamiliar territory — even RISC-V assembly — faster than a novice could. His long-standing heuristic still holds: understand the abstraction layer directly above and below the one you work in, and you can troubleshoot most problems. For someone with a large AI-built system and weak foundations, he favors stepping back to learn the basics systematically, as he did early in his own career — while candidly admitting he’s no longer sure that path pays off the way it did in the 1990s. Only half-joking, he says a beginner today might do well to consider a trade requiring hand-eye coordination, since manual labor looks harder to automate than code.