Hacker News Reader

Comment by datameta | Hacker News Reader

The more I use LLMs to code, the farther I stray from the joy of coding.

I won't regurgitate the common sentiment of it being like pair programming with an oddly spiky profile junior engineer savant. I find my emotions being activated too often in having to recontextualize or reign in or re-orient or repeat.

If I wrote code that doesn't work, that is on me. If I have to read code that doesn't work, I didn't write it, and it's veracity or accuracy are only uncoverable by me - but are already being treated as gospel by my AI peer, and the LLM keeps asserting issues with "my" code that I don't yet feel I have ownership over because I have not ingested it fully is supremely tiring. The whiplash feels almost like an abusive relationship.

I want to get back to feeling accomplished once finished with a coding session, not like I just went through a kafkaesque ringer.

View on HN

> The more I use LLMs to code, the farther I stray from the joy of coding.

I agree. The way I put it is that it feels like programming has been turned into the job of an editor, when it used to be an author. Of course, editing was always a huge part of programming, probably the biggest part, but I always kinda felt like reviewing and editing was more the "brussel sprouts" part of the job and authoring the "ice cream". I was fine to eat my brussel sprouts if I got to have some scoops of ice cream sprinkled throughout, but now it feels like just endless plates of brussel sprouts.

You can call it editor or you can call it director/architect. Instead of tapping out characters you spend more time making interesting and more high level decisions.

I totally disagree with this, and that's because (like this blog post says) the output can't be trusted - you still need to review every line of generated code. The recent post I linked in another comment about the Cloudflare OAuth implementation had a catastrophic bug that would have been "game over" if the engineer wasn't such an expert in the field.

So it's like an architect but only if you need to triple check every nail, bolt and screw because otherwise the building will probably collapse.

What was the bug, didnt read the CF article

https://news.ycombinator.com/item?id=44167006

It turns out that the process of "tapping out characters" is (a) never any kind of bottleneck for meaningful product velocity in any kind of meaningful system; and (b) actually important for engineers to do in order to understand the system that they're expected to maintain.

If (a) isn't true for you, then you're operating in a pathological environment, context, organization, etc. which isn't representative of any kind of broader industry experience.

And (b) is usually clear to anyone who's, for example, taken a university course, and compared the efficacy of manual note-taking vs. (say) automatic transcriptions from lectures. The process of parsing the lecture through your brain and into notes that you write yourself, turns out to be essential to information retention and conceptual understanding.

There's a big difference in criticality of different kinds of software. If the code you're writing will be 'the thing that runs the actual business logic in production' of course you should understand and own the code. And that code is very meaningful in the way you describe.

But so much of the software we need to solve various problems is not mission critical. Like tooling. Or any low stakes software that can be easily replaced. Or some script that makes your life a little bit easier. Or maybe even a small gui app that helps you compose some specific configuration for that other software. Tools where the output is easy to verify and can stand on its own and can't be used by an attacker to exploit your systems.

If you put a lot of effort into tooling that you can throw away the moment better tooling appears you can make the mission critical software leaner.

> actually important for engineers to do in order to understand the system

It is one tool for it and an important one. But it is not the end all tool for understanding or trusting code. If that were the case you'd have to rewrite all code you ever had to maintain. You rely on a stacks of software all the time that you do not write.

AI indeed makes mistakes but so does humans so we have to validate the code we use with multiple such tools.

In any kind of minimally-effective engineering team, engineers will only ever be working with and maintaining code that's "mission critical" in the sense you're describing here, there won't be any meaningful amount of "non-mission-critical" code worth considering or optimizing.

This is in some sense a didactic assertion: if it's not the case, then your engineering team isn't providing any value beyond what a bash script, gluing-together JIRA tickets and GitHub PRs via the LLM-du-Jour, could do autonomously.

I have never worked in an engineering organisation that lives up to that level of efficiency that you seem to describe. Once you scratch the surface there's always a mountain of things that would help people get more efficient that never gets prioritized.

And I've never worked in an engineering org that _didn't_ live up to the level of efficiency that I'm describing. I guess, seek better employment?

I've been struggling with the same thing. It's the same reason I didn't really want to go on the management path because it loses the authoring part. But now it seems to be forced on us.

That being said, there are lots of brussel sprouts in authoring sometimes - the boilerplate and wrote stuff we've done a billion times. You can make Claude eat those brussel sprouts which is nice.

also, very nice analogy, but don't besmirch the good brussel sprout name. I think in real life, cooked properly, I might even prefer them to ice cream.

Tangential fun fact!

Good tasting brussel sprouts are a very recent invention. They used to be a lot more bitter, but a new cultivar was engineered in the 90's that all but completely eliminated the bitter notes. Over the next 30 years, pretty much everybody has switched over, so now brussel sprouts are just drastically better than they were when they got the reputation for being gross.

Also, it's brussels sprouts. The name comes from the city of Brussels, BE.

I definitely hear you on this. I, like most people my age came into coding when it was mostly high level (Javascript, python, etc), yet I watch YouTube videos about things like NES programming using 6502 assembly. Part of me yearns for a time I didn't get to experience where we were closer to the hardware. Screaming at an LLM for giving me outdated methods to a library I'm trying to implement isn't in my ideal workday.

So far, the AI revolution has only given me more work. People have come at me with applications that are 80% done (really more like 50%) that they "vibe coded" and need an actual programmer to finish. These apps would most likely just be a spark in someone's imagination pre-AI.

In a way, this is a positive. I can't say it's been more fun though.

I can definitely relate to wanting to feel accomplished after a coding session and how AI can sort of strip you of that feeling at times. For me personally, I started to find the joy in the AI + coding relationship once I realized I just had to reframe my thinking a bit.

It's less about code generation for me and more about opportunities to learn new things. It could be as simple as having it optimize my code in a way I hadn't even thought of or it could be me wanting to learn a new complex architectural pattern I haven't had the time to deep dive but have wanted to. Now I can spin something up and have a base understanding of it in minutes. That's exciting to me more than anything else. In a way, it takes me back to way earlier in my career when every day felt like I was learning something new and cool on the job from more seniors devs. I think as you get more senior and experienced, those "cool" learning moments start to happen a little less, so having AI be able to reignite that is exciting in a lot of ways.

I still experience what you outline for bootstrapping understanding, or bringing the knowledge horizon closer, so to speak. Part of the learning though is through the vigilance of parsing and correcting.

When I had a small passive circuit in my head that I wanted to one-shot solder on a protoboard and didn't want to get bogged down in KiCAD, talking through it with an AI and repeatedly correcting its understanding really solidified my own. It's like mentoring while being tutored.

So I still see value in use for smaller novel projects, but using LLMs to hit a deadline for production is not something I want to do any longer, for the time being.

do you let it generate lots of code before review? i find this feeling you describe if i let it run wild and then have to sit there and figure out what it did and slog through every line.

the happy medium for me is either giving it very defined small tasks that i can review (think small edits or refactors over a few files). or i baby sit it and review every change its about to make. doesn't always pan out but it's basically a super powered auto-complete and i can "steer" it the direction i want. i can validate chunks much faster and it's typically of good quality.

that or i make mental notes of what i want it to fix after the first pass and run it thru again with the edits. in that way it does feel more collaborative and less like reviewing some unknown piece of code from someone else.

I lost the joy of coding years ago. Using an LLM is novel and interesting, in way that programming hasn't been for over a decade

Yes, it seems to decrease the fun/satisfying parts of software development and increase the unpleasant parts.

Reply on news.ycombinator.com