Developer essays dissect two rival philosophies for AI-assisted coding
As AI code generation becomes mainstream, developers debate whether reading AI output should remain central to software practice.
What to know
- Two distinct approaches to AI-assisted coding are emerging: accelerators who maintain full code understanding and vibecoders who delegate implementation details to AI systems.
- Developers who read generated code report catching systematic issues like overengineering and loss of design coherence that testing alone misses; vibecoders trade this oversight for faster feature delivery.
- The long-term costs—cognitive debt for accelerators, intent debt and vendor lock-in for vibecoders—remain unclear as teams scale these practices across shared codebases.
- Current practice lacks agreement on what each approach asks of teammates, creating friction as developers collaborate across philosophical divides.
The dispute Whether understanding implementation details is a necessary professional responsibility or an outdated burden that AI should eliminate. · positions read across 20 posts and comments
Reading generated code remains essential to catch subtle design and complexity problems that automated tests miss.
-
“I do not read every line of code agents produce. I still read the parts that matter.”
Alex · Blog post ↗
Delegating implementation to AI and focusing on specification enables faster delivery and shifts cognitive load appropriately to the model.
-
“Vibecoders expect language models to abstract away implementation, putting them in the same category as compilers and frameworks: there should no longer be a need to understand technical details.”
Zanlib · Blog post ↗
Zanlib Software developer and essayistAlex Software developer and essayist
How it unfolded 2 developments, newest first · click a bar or a number to jump posts
-
2
Zanlib frames AI coding as choice between accelerators and vibecoders
Zanlib publishes essay distinguishing two prevalent AI development philosophies: accelerators retain understanding of AI-generated code and invest in explanation and change capability; vibecoders delegate implementation to AI like compilers and focus on specification and evaluation. Argues neither approach has proven sustainable yet.
“We have begun sharing codebases without necessarily agreeing on what either path asks of our colleagues.”
— Zanlib -
Two things happened recently in an open source codebase I contribute to: Last week, after our PI gave a colloquium talk, we got a series of drive-by PRs from a rando GitHub account that claimed to be run by a Microsoft employee, making shallow changes that attempted to fix some marginal issues in our codebase. We wound up closing them all, and are…
2 more of the top 3 · 20 posts in this stretch
-
My biggest fear is that this attitude permeates, or will soon permeate, every applicable profession. Your government, your insurer, your bank, your children's school, maybe even your doctor, might soon be run by people who think like that 19 y.o. or their interviewer. God help us all then. That's how AI kills human civilization. Not by means of…
-
Not sure if it's your intention but this reply reads as pretty hostile.
-
-
1
Developer outlines case for selective code reading with AI
Alex publishes essay arguing that while not reading every line AI produces, careful review of types, shared abstractions, and complexity remains essential to prevent overengineering and maintain design coherence. Describes workflow combining adversarial plan-review, agent reasoning, and human code inspection.
“I do not read every line of code agents produce. I still read the parts that matter.”
— Alex
What people are saying 17 voices from 1 site · best of 20 · verbatim
- Sep 20
-
I think you might be missing the point of what I mean by "pure applied philosophy" in this case, although it is, admittedly, a bit of a mental shortcut. I am firmly of the opinion that programming is a branch of engineering, and engineering is effectively applying science (which is a branch of philosophy) to the material world. The thing about…
- Sep 17
-
So it’s kind of a farce? This is not the case where I work although we all use AI pretty heavily.
- Sep 16
-
PRs need one signoff, indeed, but I am pretty sure the code review that led to the signoff is the result of `claude '/code-review medium'`.
-
Where do you work? Everywhere I’ve worked merges have been gated on at least one approval and often a few more based on code owners. If you don’t review your own PRs does someone else do it?
-
I realized yesterday that I was still making 'I read the code' part of my Serious Engineer identity as I committed my fifth PR in a row that I hadn't read. Partially it's Opus 5 being, just, *really* smart; partially it's the PRs being refactoring-heavy or rushed for demo; but partially it's that once you start you forget you had started anything…
-
Oddly, this is very model dependent, even within a single company. In my experience, Fable is capable of one-shotting a 2,000 line Rust utility or a small Android app, and writing a reasonable enough README.md with screenshots. But Opus 5 writes _horribly_, very nearly as badly as your parody. I wouldn't trust either of them to maintain a 10,000…
-
Meanwhile, the first line in the README: "This is a server, not a local app. This README is load-bearing to explain why foo.c:55 is a closed manifold and not a seam. Do not remove!"
-
Indeed. I think the scariest part is that even with the scenario laid out this way, a large fraction of humans would still willingly choose that path.
- Sep 15
-
I've been thinking this as well for a lot of this year. It seems far more likely to me that AI "destroys" human civilisation through the total destruction of all intellectual work and knowledge, rather than the rationalist school of AI murderbots or whatever. One can imagine a future in which all knowledge is gated behind paid LLMs and all code is…
-
I'm not saying you're wrong, but I don't see any hostility in it, personally.
-
Of course I still read the code. How else would I know what atrocities the agents are committing? And because I read the code, I know that Fable is actually an _excellent_ software engineer up to about 3,000 lines or so. But once it exceeds a context window, it tends to cut corners on maintainability. At which point, you know, I've spent decades…
-
;-) ;-) ;-) fixed ;-) ;-) ;-) My 58 year old boss posts those same emoticons btw ;-) ;-) ;-)
-
> I used to describe myself as a Programmer. Your profile here still describes you this way. ;-)
-
> What I am certain of, though, is that there is one obviously harmful practice: putting people who prefer different approaches on the same team without establishing expectations and boundaries beforehand, and then keeping to them. I think this is the punchline of this article. I've been feeling this pain more and more as $WORK introduces more and…
-
"I wrote before about how programming is essentially pure applied philosophy" "Programming is one way to sort the grains of sand of reality" "code was never the hard part" (not from this blog post, but you've probably seem those posts around) I've realized that programming is nothing at all. It's language, if anything. It's just a way for smart…
-
this is the proper take on vibecoding. but i would not compare vibecoders to compiler/framework users, unless the vibecoder is willing to learn some programming before/while vibing. even the world's most ergonomic compilers/frameworks can't hide all the abstraction. any programmer worth their salt must learn some hardware/theory behind the…
- Sep 14
-
Thanks for writing this, I shared it with my colleagues as it resonated with me a lot.