Support this channel on Patreon ► https://www.patreon.com/zoranhorvatGenerative AI can write code, but it cannot develop software on its own. Here is why the…
But look at all those new fancy job positions. Forward Engineer, AI Enablement Engineer, Prompt Engineer, Applied AI Engineer.
I’ve seen a colleague write an API interface for support technicians with AI. As he adds to it or makes changes, the AI breaks something somewhere else. When you ask how to do something he struggles to find where it is on the interface. He says it should be in the documentation, but he doesn’t know where because he vibe coded the documentation as well. Then when he needs to just incorporate extra devices with the same API it becomes a gauntlet because he’s not sure what the AI did for the first set and the new set only gets half the functionality. On top of that, we are running days behind schedule because we spend so much time on debugging it.
If this was done properly from the ground up it would be less messy, have more intent, and it could be a framework for other projects.
I don’t have great experiences trying to get LLMs to write code, but I’ve found it to be incredibly valuable for checking my code for mistakes and oversights. Sure, it will spit out nit picks, false positives, and straight up nonsense, but at a rate that is quite a bit lower than other tools I’ve used. I not infrequently end up going through a number of iterations, rewriting code and tests based on its feedback, until I’m happy with the code and the LLM no longer points out actual problems.
It probably doesn’t help my productivity, in terms of lines of code committed or whatever companies might use, but it certainly helps improve the quality of my work
The problem isn’t the tool; it’s the lack of engineering foundations. Generalizing all AI-assisted development as ‘vibe coding’ is a massive oversimplification. There is a vast difference between a beginner blindly copy-pasting LLM output into a codebase they don’t understand, and a senior architect using LLMs as a high-powered assistant to speed up boilerplate, local schema generation, or parsing scripts. When you already know exactly how the underlying system operates, how memory is managed, and how to design clean software architectures, the LLM is just a productivity multiplier. You still design the data flow, audit the tool-use sandboxes, and review every single line of code. It doesn’t replace thinking; it replaces tedious typing.
As a senior developer I have serious doubts about the whole thing. Yes, I don’t do tedious typing anymore, now I do extremely tedious code review all day, my least favorite part of the job. And I have to be very vigilant because the AI is an idiot more often than not. Then when I finally publish my own code it’s time to go review my colleagues’ ai code and figure out what they missed in their review.
I don’t feel much of a productivity multiplier. I’m not saying we won’t get there, but this current iteration ain’t it.
you are a senior developer, start to transform your view in a senior architect. With AI there’s no more need of developers. An architect ius needed, and if it haas a solid dev backround as yours projects will change view. Technology is always changing and it’s hard to stay at pace. But if you look from an higher perspective your project your experience will only help the AI to do the Job for you
lol no. this is what a mediocre dev thinks.
In your mental model, you can only become an architect if your artificial subordinate does its job properly. Unfortunately, it looks like the subordinate is an idiot, so your analogy just doesn’t work.
Removed by mod
deleted by creator
I’ve found AI powered autocomplete suggestions is actually quite nice. The time save when it gets it right is sometimes significant, but the time lost when it gets it wrong is usually negligible.
good grief, AI powered autocomplete is so bad it slows down my development by an order of magnitude.
It’s really set up for failure. It can only fall short of a deterministic, type-aware, contextual autocompletion, something IDEs have been supporting for like a decade.
Agree. Using AI as a tool is very productive. On the other hand, letting AI drive everything is insanely time consuming and tiring work due to back & forth prompts.
The most important thing developers forget is planning. I am senior and used to delegate dev to Junior Devs. If you have not enough experience in software architecture you are missing the most important thig: You cannot start developing. You must start planning, first of all require your agent to plan the steps for the target mission. Than examine the plan produced, ask to divide int single in testable units. Most AI Vibe programmers start with develop directions. That is wrong. The longest part of the job is to prepare the Agent to perform correctly
There can be a million papers on how AI is worthless, the vibe coders are way too brainwashed into their shitty habits that they’ll never stop. It’s basically a sect at this point, but a sect that is destroying the world.
I’ve coded “professionally” (solo dev for a small company. Not the full dev experience so big caveat)
I find AI is amazing at writing unit tests and other test harnesses. That’s all it’s good for besides summarizing documentation and even then it’s so misleading it’s basically useless you have to constantly doubt it. It’s a pathological liar… And I mean it’s kind of an impossible task… It’s trained on the entire internet. The strongest weights it has are for the oldest and most popular libraries. Oftentimes it’s telling me to use deprecated shit left and right because that’s biased extremely high during training, and maybe 1% of posts on that library mention it’s deprecation.
I prefer to write my code the old fashioned way. If I have it generate code it’s never pasted in, and I prefer having it explain what I don’t know about what it’s doing. Using it as an assistant and a TDD buddy works relatively well.
Also if you poke at software architecture. I’ve learned a lot (I wasn’t formally schooled in computer science) and it’s good as long as you constrain the scope of what you’re asking about, and make sure to consider it on your own against the existing project architecture. I still make the design decisions myself.
Oh, also having it make design decision documents. I often make those to record “this is why I’m doing this feature this way. These are the options I considered, and why I decided my way is better in this particular case” so I have a record of my state of mind and decision making. AI is phenomenal at making easy to read and well summarised DDDs
I find AI is amazing at writing unit tests
Some of my coworkers say this too. The tests are generally garbage that don’t test what they say they do. All this says is that you don’t care how useful your tests are outside of checking a box
Yes you have to verify it yourself you don’t just trust it. The biggest benefit is tests are extremely to read and falsify like simple scripts. It’s faster to generate it than type it by hand, and any edits I make are minimal or require a reprompt that gets it right the second time.
There are papers that are pro and con. It’s a new tech - we’re still figuring out how to use it. It’s gotten a lot better than a year or two ago though.
It’s definitely helped me get things done a lot faster. You can point to “that one study everyone does” and tell me “no it’s not” but honestly, it is.
I’ll give you an example - I had to write a small bash script to fix home directory ownership on a server because somebody borked it and some critical jobs weren’t launching properly. Just something to read /etc/passwd, parse out the owner and home dir and
chown-R $user $home. Dozens of user dirs so quicker to just script it.Time was of the essence. Claude had a script in ~6 seconds. Yeah - I could have written it - but not that fast. I validated the output and gave it a run. All’s good.
You can insult me as “not good developer” or whatever you need to do to make yourself feel better - your opinion of me is irrelevant. But these tools are pretty damn good at what they do if you use them properly. “Properly” being the key word here. They are tools not employees so you need the proper critical thinking to apply them effectively.
There are papers that are pro and con
proceeds to not mention or link any paper that is pro.
If you’re interested you can search, since you haven’t I assume you’re not.
You are the first I’ve heard of this, burden of proof is on you. Claiming there is proof then refusing to provide it, saying others should look it up is actually so common on Lemmy that it’s an instant indication that someone is lying out of their ass.
It’s funny how everyone that tries to defend the use of AI, other than deny the fact that most evidence (and common sense/basic understanding of AIs) points to the idea that it’s just negative, always avoid the subject of environment.
Even if you assumed that you get the same result with Claude or whatever in 6 seconds, and with your brain in five minutes… Is the end of humanity and a massive extinction event worth those 5 minutes?
Defending AI while there are literally news popping out right now about heat waves being more violent than ever recorded and killing thousands of people per day in countries that had no problems with heat a few decades ago, is extremely selfish and self-centered. And the only justification is “it’s a bit faster and easier”. Okay.
All of this, and the facts that the entire thing was trained on stolen material, and the companies behind it are all run by bat shit crazy shitheads.
Everyone loves to shit on US tech companies, hates Windows and MS, and Amazon, Google, Apple, and Meta (all rightfully so), but is absolutely willing to overlook the AI companies that are arguably doing more harm than those others combined (or are tightly invested / integrated with them, or are them) just so they don’t have to spend a few minutes typing, or thinking about how something works.
It’s funny how everyone that tries to defend the use of AI, other than deny the fact that most evidence (and common sense/basic understanding of AIs) points to the idea that it’s just negative, always avoid the subject of environment.
It wasn’t even brought up twerp.
Ah so I guess you can’t read. Might explain the need to defend AI.
Easy block.
Yeah this is the type of constrained use case I use it for too. Same with unit tests and testing in general.
If only there was an entire suite of books by Pragmatic.
I’m a retired software dev, but anactive developer I know talks about how much time he saves daily by using Claude for grunt work.
I’ve tried (and still try) to find good workflows using AI, but I don’t think I’ve found anything that saves time. To get the same quality using AI, I have to go back and forth with it a lot, and review all the bloated bullshit it generates. If you don’t care about quality or the thing even working correctly (e.g. prototypes/POCs), it’s really fast.
don’t think I’ve found anything that saves time
Then you’re doing something very wrong. I’m yet to be convinced about AI coding for long lived projects, but for throwaway scripts, visualisations, one-off conversions and basically anything that doesn’t need to be maintained long term it absolutely saves a mountain of time.
Stick your head in the sand all you like…
I already covered that in my last sentence. Those things aren’t the things that take the bulk of the time, in my experience; I guess other peoples jobs are different. Generally, everything I write needs to maintained long term, or at least until the company goes under.
Or if you do care about quality and learn how to use it right.
You aren’t prompting right, bruh. You need to be very good at writing markdown to leverage Ayy-Eye properly.
You aren’t using the latest and the greatest model, bruh. Have you tried Claude Fairy Tale 6.9.420?
Did you include “don’t make mistakes” in the prompt? No? Rookie mistake bro.
Tell him this: If you’re writing that much boilerplate (grunt work), you’re probably doing something wrong. If AI can write your normal work, you really should be learning, so not using AI at all.
Actually I’m not going to tell a dev with decades of experience that an internet rando thinks he’s doing his job wrong lol.
Would you at least tell him to call his mother every once in a while?
nsfw (aussie)
K. It’s astounding to me that after decades in the business, he’s still writing code that can be written by basically an intern.
When I try to get AI to write my code, it stumbles all over itself just failing to understand the simplest of my libraries. (Literally it made multiple mistakes using the fucking tokenizer library I wrote that is so simple, an intern would have no problem with it.)
Yeah, so if you are a senior dev, the easy work just disappears? Like if you are an experienced carpenter, you will never have to drive nails again, because that’s too easy!
any senior carpenter that is doing the easy work has a jig they custom built to do it in a tenth the time. same for a dev. templates in your ide, the use of specific libraries that make developing tests 10x faster than a normal junit test. Whatever it is, if you’re a dev with decades of experience and you haven’t gotten to the point that the ‘grunt work’ isn’t a few seconds of time then you’re not a good dev, you’re a shit dev who sat on their ass for decades doing nothing.
And now AI has given you the chance to be even better at that, sitting on your ass doing nothing.
If the majority of your time is spent on the easy work, you’re probably not a senior dev. If you are, maybe you shouldn’t be.
Yes, we still have to do the easy work, but it doesn’t take much time. I’d rather do it myself and know that it’s done right, than let an AI do it and have to fix all of its mistakes. In my experience, it takes the same amount of time.
he’s still writing code that can be written by basically an intern He usually does full-stack projects by himself, so he has to do everything. And he’s using AI to do what interns could do. I’ve dabbled a little using VSCode AI myself to refactor and upgrade a couple hobby projects, and it didn’t “stumble all over itself” at all. In fact it conversed with me like an intern or colleague would, and made many proposals I agreed with. There are ways to craft your prompts that make AI work better. Maybe that’s your problem I dunno.
Even if it does write code that works, it usually (about 50% of the time in my experience) has bugs, and sometimes those bugs can be really difficult to spot. For me, it has never saved me any time. I’m either fixing something it doesn’t know how to do correctly, or going over its code with a fine tooth comb because when it says, “this is production ready code, with no bugs,” it’s usually wrong. That takes a lot of time. It’s easier for me to just write the code correctly myself.
Admittedly, I haven’t used that new model that Anthropic revoked access to the public to recently. Maybe that one is good enough for government work.
Fable was “just ask and get it done” quality level, but really I don’t get THAT much bugs - about the same amount that I see irl developers do: get a new feature, find 5 problems, get them fixed, find one more, done. As a recommendation - try to alleviate the biggest problems that ai models have:
- overconfidence - skipping wrong stuff “because they have a note saved that it works” or losing the point where they stopped after a session broke, then making things up. Test Driven Development solves the majority of problems like that - when Claude writes tests first, then it’s not able to bullshit me that the job is done when “everything is red”
- even with large contexts, they run out and stuff gets lost. So if you’re not doing something really compact, make your ai document everything, document the feature they are working on now, make it then offload it to permanent doc when that piece is finished. When the ai will fuck up next time, you can tell it to “go read some docs” and most of the times it will work
I’m glad it works for you, but it simply does not work for me. Maybe you could try yourself on some of my libraries, because I have never gotten it to save me any time. It’s just spending money and making the work less fun for no reason. Oh, also not having the copyrights to the things that go in my code base, don’t forget that.
When I first tried it I felt lost, but after watching a couple videos about writing good prompts I had no trouble getting it to produce perfectly good code that did what I wanted. Your mileage may vary.
Honestly, I think what you consider „good code“ is just shifted from what the previous commenter considers „good code“. Prompting is about giving enough information so the AI can solve the task without needing to reconstruct a lot of context. Most people using AI somewhat regularly will have figured out to write good enough prompts. I‘ve never seen AI generate perfectly good code beyond hello world and the fibonacci sequence. And by perfectly good I mean I wouldn’t change beyond 30% of what it produces, which is not a high bar.
Most of the younger generation devs at my company are using ai coding - mostly for said gruntwork (like writing small functions, api methods, writing what data they want to see instead of complex sql requests), but some are more enthusiastic and use heavier agentic setups. The best validation is that we still have old-school human pull request reviews (enforced by a scary Chief R&D) and if your colleague would see something unreadable or weird, your stuff wouldn’t pass.
I’m a Product Manager and I have several pet products now - all pretty viable (depending on the time I invested in each of course). A stocks website, a money splitting android app (now passing google play review), a weather app. All working, and I have really low coding skills myselfAI corporate slop is when you mass generate a bunch of stuff, don’t read it, and then export the mental burden to a different coworker. You make them looks less productive and you look extremely productive, when really you’re stealing the productivity of colleagues.
You just reminded me of that.
Again, in a corporate setting I see that the “generated stuff” is being read before being approved. And in my private setting I’m working on my projects for myself, I’m not “stealing the productivity”, I obtained productivity. All my life I was coming up with ideas, planning, and managing, but having something on my own wasn’t possible - I can’t hire a dev team. Now I am my team, I’m empowered.
You could’ve literally learned programming…
Not possible
what actual product do any of those sites produce
sounds like toys
“Sounds”? You are an expert in measuring value by vague comments?
Okay:- The expense splitting app I made for myself, my friends, and maybe other people (no ide if it will be picked up), after I was managing a giant spreadsheet of our new year’s trip costs. Now my app calculates everything, of any complexity, and it is used.
- The stock trading site has everything that I need from a stock trading site (to look for opportunities to buy), but I needed 3 or 4 before
- The weather app I made not because there’s not enough weather apps, but because each and every app wants your data for 1000s of their partners, so I made my own thing that shows weather.
If it’s not real life value, then I don’t know
None of this sounds valuable considering it all already existed.
There’s foss apps for weather that don’t collect all that data, if you’re using Android it’s usually a good idea to first check f-droid for utilities like that as it’s likely someone else already built it, and it’s usually still receiving fixes and updates (the page shows you that kind of information).
