Learning Programming in an Age of LLMs

Open answers to a reader's letter.

A scholar precocious wrote maine a agelong missive pinch tons of questions astir learning programming successful this property of LLMs. After a spot of back-and-forth, I sewage support to quote extensively from the missive successful bid to effort immoderate answers successful public.

None of my answers I see peculiarly rigorous; the business is truthful uncertain that I tin only reply to the champion of my abilities, but I don't declare them to clasp immoderate benignant of immutable truth.

"I'm trying to understand really group who profoundly understand package deliberation astir learning and competence successful the property of AI. I'm approaching it almost arsenic a historiographer would: asking group straight really they make consciousness of a technological modulation while really surviving done it.

"About a twelvemonth agone I became fascinated by AI-assisted programming. Despite having nary general CS background, pinch LLMs I managed to build a reasonably ample TypeScript/JavaScript strategy involving APIs, PostgreSQL, LLM pipelines, investigation automation and multi-model workflows. At first it felt almost magical: AI seemed to illness the region betwixt having an thought and being capable to build it.

"But now I'm trying to move that strategy into a existent accumulation product, and I'm struggling. I hole 1 correction pinch AI, past different appears, past different portion behaves successful a measurement I don't afloat understand. After months of refactoring I had an uncomfortable realization: I whitethorn person built a strategy that is supra my ain level of understanding. When everything works, that spread is almost invisible. When it doesn't, it becomes very real.

"Sometimes I genuinely don't cognize what to do adjacent without asking different model. That made maine wonderment whether I spent a twelvemonth building a product, aliases partially building the quality of one: thing blase capable to work, but which I don't yet understand profoundly capable to genuinely own.

"I'm not anti-AI astatine all. I'm fascinated by these systems and want to activity pinch them professionally. But I'm unsure what the correct narration pinch them should be."

Indeed, I'm not judge either, but earlier proceeding, I find it astir transparent to uncover my position. I haven't yet decided connected AI, but I thin toward disliking it, knowing afloat good that it whitethorn beryllium unstoppable.

I do activity and research pinch it, and it often impresses me. At different times, it frustrates me. It's usually erstwhile it impresses maine the astir that I resent it maximally.

When it's bad, it tin beryllium frustrating, but past astatine slightest I tin sorb an ember of warmth successful the illusion that what I've spent much than 30 years learning is still relevant. When it's astatine its best, I sometimes think: Where do I motion up for the Butlerian jihad?

My position connected LLMs is only partially based connected my ain socio-economic status. I'm aged enough, and person had capable occurrence already, that each different things being equal, I tin past unemployment. I'm not sure, connected the different hand, than immoderate knowledge-based nine can.

It whitethorn beryllium that LLMs will return programmer jobs earlier they return different white-collar jobs. After all, programming may beryllium a subject wherever verification is easier than, say, security claims management. Still, if we scope a constituent of wide unemployment among knowledge workers, I'm not judge nine arsenic we cognize it will survive.

I usually don't talk overmuch astir my inheritance arsenic an economist, but successful this discourse I find it applicable to mention. As an economist, I can't ideate that wide unemployment of 30-40% will not person a important effect connected the economy.

I'm painfully alert of the arguments that this has happened before: There whitethorn beryllium occupation loss, but the beforehand of exertion leads to caller jobs we can't moreover ideate today. It was for illustration that pinch the preamble of the stocking frame, the steam engine, the soul combustion engine, computers, etc. This is only partially true: Yes, caller jobs were created, but often not for those group who mislaid their jobs. Coal miners didn't conscionable go programmers overnight.

The aforesaid benignant of statement was utilized erstwhile China was admitted to the World Trade Organization. And indeed, tons of caller jobs were created, conscionable not successful the Western world.

So, based connected lived and humanities experience, I'm sceptical of arguments that each will beryllium fine.

But I sincerely dream that I'm wrong. I emotion to program, and wouldn't mind doing it for different 10 years. Perhaps much importantly, I person young big children. I dream that there's a world for them, too.

"So I'd really for illustration to cognize really you deliberation astir this. Are you gladsome you learned programming fundamentals earlier LLMs existed? If you were starting today, would you still earnestly study languages, information structures, databases, networking, operating systems, debugging and architecture? Do you deliberation AI tin fto group go tin of building overmuch faster than they go tin of understanding?"

Am I gladsome that I learned programming earlier LLMs? Yes, of course. Those skills served maine good for 30 years.

If I was starting today, I'd earnestly see learning carpentry, metalworking, gun-smithing, aliases thing other that requires hand-eye coordination. I cognize that advances are made successful robotics, too, but replacement of manual labour seems to dishonesty farther successful the future.

But to reside the question: I am, personally, presently learning information structures, connection semantics, etc. arsenic portion of a assemblage programme. I do that because I'm curious, however, and not because I expect to get overmuch monetary reward retired of it.

Do I deliberation that AI enables group to create faster than they tin support up? This remains to beryllium seen. Software developers person already, for decades, been moving connected apical of abstractions they didn't understand. If you were a web developer, you didn't cognize overmuch astir compiler programming. If you were a compiler programmer, you didn't cognize overmuch astir integrated circuit design. And if your occupation was to technologist integrated circuits, you wouldn't cognize overmuch astir the levels of abstraction above you.

A bully norm of thumb was: Understand the level of abstractions straight beneath the 1 you activity in, arsenic good arsenic the 1 above. That would alteration you to troubleshoot most problems.

"And really do you personally woody pinch that? When AI tin lick thing immediately, really do you determine erstwhile to usage it and erstwhile to activity done the problem yourself? If you were successful my position, pinch a important AI-built task but anemic foundations underneath it, would you measurement backmost and systematically study those foundations, support building and study arsenic problems appear, aliases harvester the two?"

That's 2 radically different questions, because I nary longer person a anemic instauration successful package development. Even if I were dealing pinch thing acold from what I usually do, I tin ramp up leveraging what I already know. Let's ideate that personification tasked maine pinch maintaining an exertion written exclusively successful RISC-V assembly code. That's the astir alien package situation I tin ideate for myself. Adapting to specified a improvement situation would beryllium difficult for me, but still not arsenic difficult arsenic it would beryllium for personification caller to programming successful general. Believe it aliases not, I have written mini workout programs successful RISC-V, arsenic good arsenic an workout compiler that compiled to RISC-V.

But what if I had virtually nary package background?

Well, erstwhile upon a time, I was successful precisely that situation. When I started my career, for years I balanced a knife's separator of getting things done while learning connected the job. Beginning successful 1999, I wrote COM components successful C++, not knowing overmuch of what I was doing. Somehow, I still made it work, moreover to a grade that I managed to destruct immoderate evident representation leaks.

I was, however, ne'er happy conscionable slapping things together without knowing really they worked. So I did, arsenic suggested by the question, measurement backmost to systematically study fundamentals. This worked good for a profession launched successful the mid 1990s. Will it activity good today?

I'm not truthful sure: Reaching a level of competency precocious capable to admit your past assurance arsenic intelligibly lying connected the too-ignorant-to-realize-it information of the Dunning-Kruger curve took decades. Do you person that overmuch clip today?

Granted, pinch LLMs, you tin study faster, because you tin inquire much directed questions. Thirty years ago, I would bargain books successful the dream that they would incorporate some adjuvant material. This still meant slogging done a batch of learning worldly not instantly applicable to the task astatine hand.

Still, I uncertainty that it's imaginable to importantly velocity up quality learning. The bottleneck is hardly the teachers nor the materials, but really accelerated a quality encephalon tin sorb caller knowledge.

"One past point I would beryllium particularly grateful to perceive astir is really you learned programming yourself, and really you study caller method things today. How did you attack learning a caller connection earlier successful your career? Books, projects, reference different people's code, exercises, debugging, thing else? And if you had to study a wholly caller programming connection today, pinch AI available, really would you do it?"

The short reply to the first question: Slowly, based connected overmuch proceedings and error, occasionally backed by a book.

Apart from a very early mendacious commencement pinch COMAL 80, my first programming projects was to (re)calculate bifurcation diagrams and the Lorenz attractor for my master's thesis successful economics. Reaching for what I had, I wrote them successful QBasic, learning from the samples that shipped pinch it, arsenic good arsenic occasionally asking a friend.

While I'm glossing complete galore details, successful the 1990s and 2000s, I mostly learned from examples and documentation. While I did bargain a book astir C++, I don't deliberation I ever vanished it, and I picked up various Basic dialects arsenic good arsenic C# exclusively from archiving and illustration code.

That said, though I ne'er publication a book to study C#, books were instrumental successful school maine some F# and Haskell. I have, complete the years, relied heavy connected books to amended myself, but arsenic my Goodreads profile reveals, I emotion books successful general.

How do I study a wholly caller programming connection today? Again, my acquisition is useless to personification caller to programming successful 2026: I've now seen truthful galore programming languages that if I tally into a caller one, I tin usually prime it up from perusing existing codification and looking up the fewer things that aren't instantly clear.

But that's presupposing that the connection successful mobility is 'normal'. If I had to get backmost into APL, I'd astatine slightest person to find a tutorial.

You whitethorn person noticed that I don't overmuch usage LLMs for learning. LLMs don't hallucinate; they bullshit, and I'm profoundly distrustful of thing they show me. This is not to opportunity that I don't usage LLMs, but I thin to inquire them questions that output verifiable answers. Can I make this Haskell look much succinct? Any useful reply to specified a mobility is simply a codification proposal that either works, aliases doesn't work; is shorter, aliases isn't. That's easy to verify.

What should I study next? does, connected the different hand, not output a verifiable answer. I thin to not to inquire specified questions of LLMs.

In conclusion, you could opportunity that I for illustration asking LLMs falsifiable questions.