Learning Programming in an Age of LLMs

URL SCAN: On learning programming in an age of LLMs
FIRST LINE: A reader recently wrote me a long letter with lots of questions about learning programming in this age of LLMs.

THE DISSECTION

This is a farewell letter dressed as career advice. Seemann, a seasoned practitioner, performs the cultural ritual of the knowledgeable elder dispensing wisdom to a newcomer — but the real content is the author's own quiet terror at watching the ground dissolve beneath his feet. The essay is structurally a confessional, not an analysis. It confesses uncertainty, admits to dread, gestures vaguely toward hope, and ends with a personal operational preference (ask LLMs falsifiable questions only). The reader's letter it responds to is more honest: "I may have built a system that is above my own level of understanding." That sentence is the autopsy. The rest is the living trying to process the death of their own category.

THE CORE FALLACY

The apprenticeship fallacy: that fundamentals are the moat, and that time plus experience will reliably compound into immunity.

Seemann correctly identifies that his reader built something they don't understand. His proposed solution — step back and learn foundations systematically — is advice calibrated for a world where the bottleneck was access to knowledge. That world is gone. The problem now is not access. The problem, as Seemann inadvertently confirms, is cognitive bandwidth and time-to-competency. He acknowledges this directly: "Reaching a level of competency high enough to recognize your past confidence as clearly lying on the too-ignorant-to-realize-it portion of the Dunning-Kruger curve took decades. Do you have that much time today?"

This is the critical admission. He recognizes the problem exists, then immediately pivots to a non-answer: learn faster with directed questions. The bottleneck, he concedes, "is how fast a human brain can absorb new knowledge." Exactly. That bottleneck is structural. It doesn't scale. AI does. The gap between AI capability and human learning speed is not a tactical problem — it's a mathematical constraint that compounds against the human every year.

The historical analogies he deploys (stocking frame, steam engine, coal miners) actually support the Discontinuity Thesis more than he realizes. He correctly notes that new jobs were created "often not for those people who lost their jobs" and that China WTO created jobs "just not in the Western world." These are not reassuring precedents. They are evidence that technological transitions are asymmetric by design — the gains concentrate, the losses distribute. He's already seen this pattern. He's describing the mechanism. He's just not calling it by its name.

HIDDEN ASSUMPTIONS

1. That the question "what to learn" is still the right question. For the individual, maybe. For the category of "programmer" as a mass economic function, the question is whether the labor market for that function survives at all, not whether the individual can learn it more efficiently.

2. That "understanding" will remain economically necessary. The entire essay assumes that deep comprehension of systems is a prerequisite for viable economic participation in software. This was true in 2015. It is increasingly questionable in 2025. If AI-generated systems can be maintained, monitored, and iterated by people who understand them at a shallower level — or by AI itself — then the economic value of deep understanding drops even if the epistemic value remains.

3. That the author's survival confidence is transferable. He explicitly notes he's "old enough, and have had enough success already" to survive disruption. This is a survivor's hedge, not a general principle. He's describing his own Position 1 (Sovereign-adjacent) status and mistaking it for a general lesson about competence.

4. That asking falsifiable questions of LLMs is a stable strategy. His operational heuristic — only ask questions with verifiable answers — is reasonable epistemically, but it describes a verification layer on top of AI output, not a path to independence from it. The reader is already doing this: they're debugging AI-generated code, which is exactly the falsifiable-question workflow. The problem isn't that this strategy doesn't work for verification. The problem is that verification labor is itself being automated.

5. That there will be a "world" worth surviving into. His concluding hope — "I hope that there's a world for them, too" — is the tell. The entire essay is built on an unexamined assumption that the transition will produce a survivable world. The DT framework does not guarantee this.

SOCIAL FUNCTION

Cultural lag theater. This essay performs the intellectual honesty of acknowledging disruption while simultaneously performing the emotional work of delaying grief. It acknowledges the cliff exists, admires its height, and then recommends leaning into the climb anyway — with better fundamentals, better questions, better habits. This is the appropriate genre for a 30-year veteran writing to a confused reader on a developer forum in 2026: not denial, not despair, but managed mourning. The function is to keep the cognitive community functional and emotionally legible long enough for the transition to resolve itself — one way or another.

It is also elite self-exoneration in slow motion. Seemann explicitly positions himself as safe ("I can survive unemployment") while genuinely mourning the potential immiseration of everyone else. The essay does not say: here is what you should build toward. It says: here is how I would try to survive, knowing I'm already in a better position than you.

THE VERDICT

The essay is an honest diagnosis wrapped in false comfort. Its core insight is accurate: AI allows people to build faster than they understand, and this creates fragile systems built on invisible gaps. The response — learn fundamentals, ask falsifiable questions, go slower — is advice calibrated for the previous technological regime.

Under the Discontinuity Thesis, the relevant question isn't how to learn programming better in the AI era. It's whether the economic category of "person who programs" survives at mass scale, and for how long. The author's own historical examples suggest the honest answer: asymmetric transition, concentrated gains, distributed losses. He documented this pattern in the article. He just didn't follow it to its conclusion.

The reader asked: "Am I partly building the appearance of a product?" The answer is yes. But the more important question they didn't know to ask is: "Does the product matter if the market for people who make it is being structurally eliminated?"

VIABILITY SCORECARD (The Reader's Position):

The Survival Plan available to this reader: The Sovereign path requires ownership of AI capital, not programming skill. The Servitor path requires becoming indispensable to those who own AI capital — which likely means domain expertise + AI coordination, not programming fundamentals in isolation. Learning data structures and operating systems is the equivalent of studying horse dentistry in 1915. The horse population is declining.