Comment by Izkata | Hacker News Reader

Comment by Izkata | Hacker News Reader

Parent Full thread

I don't use it.

I know my mind fairly well, and I know my style of laziness will result in atrophying skills. Better not to risk it.

One of my co-workers already admitted as much to me around six months ago, and that he was trying not to use AI for any code generation anymore, but it was really difficult to stop because it was so easy to reach for. Sounded kind of like a drug addiction to me. And I had the impression he only felt comfortable admitting it to me because I don't make it a secret that I don't use it.

Another co-worker did stop using it to generate code because (if I'm remembering right) he can tell what it generates is messy for long-term maintenance, even if it does work and even though he's new to React. He still uses it often for asking questions.

A third (this one a junior) seemed to get dumber over the past year, opening merge request that didn't solve the problem. In a couple of these cases my manager mentioned either seeing him use AI while they were pairing (and it looked good enough so the problems just slipped by) or saw hints in the merge request with how AI names or structures the code.

I've been using ChatGPT to teach myself all sorts of interesting fields of mathematics that I've wanted to learn but never had the time previously. I use the Pro version to pull up as many actual literature references as I can.

I don't use it at all to program despite that being my day job for exactly the reason you mentioned. I know I'll totally forget how to program. During a tight crunch period, I might use it as a quick API reference, but certainly not to generate any code. (Absolutely not saying it's not useful for this purpose—I just know myself well enough to know how this is going to go haha)

How do you get chatgpt to teach you well? I feel like no matter how dense and detailed i ask it to be or how much i ask it to elaborate and contextualize topics with their adjacent topics to give me a full holistic understanding, it just sucks at it and is always short of helping me truly understand and intuit the subject matter.

Yes, this is my experience as well. At some point you would be better off find something written by a human, because AI would just take you in circles.

This is an interesting usecase, and I want to learn more about your workflow. Do you also use Lean etc. for math proofs.

I’m the same way. But I took a bite and now I’m hooked.

I started using it for things I hate, ended up using it everywhere. I move 5x faster. I follow along most of the time. Twice a week I realize I’ve lost the thread. Once a month it sets me back a week or more.

I repeatedly tried to use LLMs for code but god they suck. I've tried most tools and models and for me it's still way faster to write things by hand.

I'm a magical tool, it's almost like if I knew what I wanted to do ! Don't have to spend time explaining and correcting.

Also, a good part of the value of me writing code is that I know the code well and can fix things quickly. In addition, I've come to realize that while I'm coding, I'm mostly thinking about the project's code architecture and technical future. It's not something I'll ever want to delegate I think.

I use AI to discuss and possibly generate ideas and tests, but I make sure I understand everything and type it in except for trivial stuff. The main value of an engineer is understanding things. AI can help me understand things better and faster. If I just setup plans for AI and vibe, human capital is neglected and declines. I don't think there's much of a future if you don't know what you're doing, but there is always a future for people with deep understanding of problems and systems.

I think you are right, deep understanding of systems and domains will not become obsolete. I forsee some types of developers moving into a more holistic systems design and implementation role if coding itself becomes quite routinely automated.

The atrophy of manually writing code is certainly real. I'd compare it to using a paper map and a compass to navigate, versus say Google Maps. I don't particularly care to lose the skill, even though being good and enjoying the programming part of making software was my main source of income for more than a decade. I just can't escape being significantly faster with a Claude Code.

> he can tell what it generates is messy for long-term maintenance, even if it does work and even though he's new to React.

When one can generate code in such a short amount of time, logically it is not hard to maintain. You could just re-generate it if you didn't like it. I don't believe this style of argument where it's easy to generate with AI but then you cannot maintain it after. It does not hold up logically, and I have yet to see such a codebase where AI was able to generate it, but now cannot maintain it. What I have seen this year is feature-complete language and framework rewrites done by AI with these new tools. For me the unmaintainable code claim is difficult to believe.

have you tried using AI generated code in a non hobby project? one that has to go to production?

it just allucinates packages, adds random functions that already exist, creates new random APIs.

How is that not unmantainable?

We use it daily in our org. What you’re talking about is not happening. That being said, we have fairly decent mono repo structure, bunch of guides/skills to ensure it doesn’t do it that often. Also the whole plan + implement phases.

If it was July 2025, I would have agreed with you. But not anymore.

I used to experience those issues a lot. I haven't in a while. Between having good documentation in my projects, well-defined skills for normal things, simple to use testing tools, and giving it clear requirements things go pretty smoothly.

I'd say it still really depends on what you're doing. Are you working in a poorly documented language that few people use solving problems few people have solved? Are you adding yet another normal-ish kind of feature in a super common language and libraries? One will have a lot more pain than the other, especially if you're not supplying your own docs and testing tools.

There's also just a difference of what to include in the context. I had three different projects which were tightly coupled. AI agents had a hard time keeping things straight as APIs changed between them, constantly misnaming them and getting parameters wrong and what not. Combining them and having one agent work all three repos with a shared set of documentation made it no longer make mistakes when it needed to make changes across multiple projects.

Yes, all the time. Yes, those go to production. AI has improved significantly the past 2 years, I highly recommend you give it another try.

I don't see the behaviour you describe, maybe if your impression is that of online articles or you use a local llama model or ChatGPT from 2 years ago. Claude regularly finds and resolves duplicated code in fact. Let me give you a counter-example: For adding dependencies we run an internal whitelist for AI Agents; new dependencies go through this system, we had similar concerns. I have never seen any agent used in our organisation or at a client, in the half year or so that we run the service, hallucinate a dependency.

So where does your responsibility of this code end ? Do you just push to repo, merge and that's it or do you also deploy, monitor and maintain the production systems? Who handles outages on saturday night, is it you or someone else ?

FWIW I mainly use Opus 4.6 on the $100/mo Max plan, and rarely run into these issues. They certainly occur with lower-tier models, with increased frequency the cheaper the model is - as for someone using it for a significant portion of their professional and personal work, I don’t really understand why this continues to be a widespread issue. Thoroughly vetting Plan Mode output also seems like an easy resolution to this issue, which most devs should be doing anyways IMO (e.g. `npm install random-auth-package`).

We use it for 100s of projects and what you say hasn't happened for a while.

LLMs rarely if ever proactively identify cleanup refactors that reduce the complexity of a codebase. They do, however, still happily duplicate logic or large blocks of markup, defer imports rather than fixing dependency cycles, introduce new abstractions for minimal logic, and freely accumulate a plethora of little papercuts and speed bumps.

These same LLMs will then get lost in the intricacies of the maze they created on subsequent tasks, until they are unable to make forward progress without introducing regressions.

You can at this point ask the LLM to rewrite the rat’s nest, and it will likely produce new code that is slightly less horrible but introduces its own crop of new bugs.

All of this is avoidable, if you take the wheel and steer the thing a little. But all the evidence I’ve seen is that it’s not ready for full automation, unless your user base has a high tolerance for bugs.

I understand Anthropic builds Claude Code without looking at the code. And I encounter new bugs, some of them quite obvious and bad, every single day. A Claude process starts at 200MB of RAM and grows from there, for a CLI tool that is just a bundle of file tools glued to a wrapper around an API!

I think they have a rats nest over there, but they’re the only game in town so I have to live with this nonsense.

Reply on news.ycombinator.com