Having been a coder for decades before AI came on the scene, I don’t understand how inexperienced programmers could possibly write a serious amount of working code with AI.
It’s wrong, like, at least half the time, but as an experienced coder, I can look at the “code” it generated and know what it was trying to do, and then write it correctly. I do find AI useful when I’m not sure how to go about solving a particular code-related issue, but … it just gives me something to think about, not an answer I can use directly.
It’s like google-coding in 2010; nothing you search for is exactly what you need, but it could help you see why your code isn’t working.
I really don’t dig that comparison. When you look up a snippet on stackoverflow, for example, you can immediately see the quality of the answer, as well as feedback from real people
Yeah like if you start coming across snippets that aren’t even properly indented, you know you’re digging the real bottom of the barrel (been there while struggling to fix email templating I knew nothing about back in the day). Now, the code you get from the LLM looks totally legitimate to the untrained eye, and it may even generate a convincing explanation.
You won’t have any indication when it’s dead wrong until you try to run it. And even then, it may be “working” in a way unintended because you don’t actually understand what you copy+pasted, because neither does the LLM ofc.
I can’t even imagine the spaghetti bowl you can get yourself into if you just keep vibe coding yourself deeper and deeper, while understanding nothing.
The spaghetti bowl is the real problem. You can make something that works, but it’s so fragile because the solution is rarely general and never elegant. The snippet might be surprisingly elegant, but it will reimplement the same code 3 different ways in 3 different places and the whole thing turns into a mess
You can see the quality if you’re an experienced coder. My comment lacks personal context in that I was in school in 2010 and there were plenty of my classmates who would plug snippets into their projects without fundamentally understanding what it did or learning what the project was supposed to teach us. Similar to a shortcut with AI in 2025.
There are definitely people who cut & pasted from stack overflow in the work environment, too. The difference is that I, as the clean-up crew, could google their code and find the post it came from … and then I could read the comments and figure out wtf they thought they were trying to do. When they paste LLM-generated code in, there’s no trace of where the dumbfuckery came from.
Just thinking about it makes me glad I’m near retirement.
that was exactly my point, for the “non experts” googling and using AI is very much not the same, as googling provides them with a lot more actual information (quality, alternatives)
deleted by creator
I can look at the “code” it generated and know what it was trying to do, and then write it correctly. I do find AI useful when I’m not sure how to go about solving a particular code-related issue, but … it just gives me something to think about, not an answer I can use directly.
So glad to see others that do that. Still haven’t really tried to understand what vibe coding is, as I try and ignore passing terms, but I was starting to think it was just using the AI assistants in any way. I use it in the same way as you and find it perfectly fine for that purpose but I can’t imagine using it for anything more.
I tried using chatgpt to write a basic batch file, it ended up such a horrendous mess that i gave up halfway through. Fucker got told four times, still kept putting the REM on the same line as actual code.
This is the opposite of that 'teach a man to fish and he’ll never grow hungry" etc.
Build a man a fire and he will be warm for a day, light him on fire and he will be warm for the rest of his life.
I had to google Vibe Coding. Seems like it’s not actual coding and you’d then have to check the code yourself and at that point why bother? Easier to start with something that makes sense then the understand and fix a cluster fuck.
Nah, that would be programming with AI.
In vibe “coding”, you ask the AI for the code and just run it. If it doesn’t do what you want it to do, you just ask the AI again, or another AI. Ad infinitum.
Check the code yourself? That’s like 5th century pleb work, vibe “coders” would be wasting their precious time when they can just ask another AI to do it.
“you’ll own nothing and be happy” applies to skills now.
It’s very helpful that there are a handful of nonsense phrases that AI has scraped by reading journal articles wrong. They’re commonly published in magazine format with a bunch of narrow columns, so there’s some gibberish that AI scraped by reading across the page instead of down the columns. I want to make a database of those nonsense phrases so that I can just Ctrl+F in a journal article to see if I should just skip reading it because it’s AI garbage.
It’s worse than that.
The goal isn’t to sell coding superpowers to programmers. It’s to drive a wedge between employer and employee. Make both of them dependent on an intermediary instead of each other.
Think DoorDash but for coding gigs. You don’t have a job, but a series of push notifications offering a chance to review an 18-line PR for $3.81.
Remember to respond within the next 90 seconds to maintain your priority status, and don’t decline too many offers.
Edit: See also, chickenized reverse-centaurs.
This 11 year old adult swim comedy video doesn’t even feel that ridiculous anymore.
It’s the same cycle since the '70s. Whether it’s COBOL or VB.NET or vibe coding, the premise hasn’t changed.
There’s three broad categories of code:
- Monkey code (random applets that are almost entirely business logic and non-critical)
- Actual code (most things)
- Crazy shit like kernel or browser code.
I can see vibe coding, situationally, lower the barrier to entry of (1). But also that’s no different from COBOL or VB.NET which both promise “MBAs can now write code”, which conveniently never extends to maintaining said code. And vibe coding doesn’t help with that either, ChatGPT is an awful debugger.
Your boss thinks ChatGPT will help with (2), but it either won’t or only very slightly as an advanced autocomplete. For any problem-solving that requires more specific domain knowledge than can automatically find its way into their tiny context windows, LLMs are essentially useless.
… So I’m not worried. Today’s vibe coders are yesterday’s script kiddies.
the amount of mistakes and and hallucinations ai has makes it actually take longer to code.
it’s the same old garbage in, garbage out….
it can kinda help you get started but that only saves you 10 minutes of reading documentation that you have to read anyway to make sure it didn’t make something up.
It seems OK at spewing out a bit of code it found on StackOverflow, or even joining two bits of code together, but it really falls apart when you poke at the edges of it’s knowledge.
And the problem is, neither you nor it knows where those limits are, and it very quickly goes from confident copy and paste to confident bullshit.
It even knows what excuses smell like, so it’ll give you one at random when you call it out.
yep. i’ve tried it a bit and the errors are blended in so well and seem so plausible, it’s worse than stack overflow….
even when just getting default arguments for a function it makes stuff up.
i do see it getting better at errors like that, but not much better….
Debugging is the hardest part, and now you get to spend all your time doing it
I’ve been using chatgpt to help me build a Bubble website. That is, I am doing all the work, I just bounce questions of how to achieve things and structure conditional statements correctly.
Because I’m basically sanity checking everything it says vs copying blindly, it’s interesting to see just how much it gets caught in a loop of misinformation. I’m lucky to be one of those learners who just needs an example, even if it’s a shitty one, to figure it out myself, so I often find myself using it simply to see how it’s NOT done.
But yeah, I know jack shit about coding but I’m sure AI code sucks ass.
Good for you to want to learn a new skill and taking things that LLMs spit out with healthy skepticism. I’m afraid future generations will lack such motivation.
100%. Half the time I see the first couple lines of AI code and I’m like, nah, that’s not right. Let’s do it myself lol
Here’s a fun thing. Using the latest AI to code backend and front-end code. Every couple of weeks, have to stop, go through every line and module, and throw out pretty much 90% of the code, manually refactor, and rewrite it.
It offers a good starting point, but the minute things get slightly complicated, you have to step in. I feel bad for people who think this will make it so they don’t need experienced developers and architects. They’re in for a rough ride.
An interesting point I heard the other day: if AI can replace entry level jobs, doing simple scripts that AI can definitely do (because it essentially just spits out the stack overflow/Reddit/etc training data verbatim), then companies no longer need entry level programmers.
If they don’t need entry level programmers, how do you get future senior programmers? Skipping directly to advanced stuff without getting practical experience on the simple stuff is incredibly hard.
What happens when the current senior programmers retire in larger numbers, and there’s very few replacements because the ladder is gone?
That’s a problem for Q72 and they’re incapable of looking past Q4. Besides, they’ll have already jumped ship by then, what do the execs care if they make this quarter just ever so slightly more profitable
They’re incapable of looking past a single quarter, let alone 4.
What happens when the current senior programmers retire in larger numbers, and there’s very few replacements because the ladder is gone?
I don’t have a solution. I’m just stocking up on physical paper books, so I’ll have something to entertain me while nothing works, until someone figures it out.
(I’m sort of joking, and sort of serious. I do expect Internet service outages to become a lot more common. But I actually just like books, anyway.)
Agree. Software engineering is a marathon - not a sprint. These AI tools are useful to get something up real quick, but I have a hard time seeing how they can be useful for long term maintenance work.
Software engineering is a marathon - not a sprint.
Oh BOY do I have this ‘brand new shiny’ thing called Agile at almost every fucking company ever.
It’s still a marathon, even if the name ”sprint” is used. The point is the same: software engineering is about ensuring long term maintenance. It’s about building software that can sustain through multiple sprints.
The typical code from an AI agent can barely sustain a single sprint without having to restart from scratch.
I know, but in most companies they don’t give a fuck.
What’s done is done, sure there can be some minor maintenance, but goodness forbids you need to rewrite something that handles the 10x throughtput that built up over the years.
I am usually able to get some cleanup tasks in, but from what I’ve heard, not many people are.
It’s just sad, that some think ‘sprint’ means ‘this is done and dont dare to tell me you need more time, what have you been doing the last X sprints?’.
I am usually able to get some cleanup tasks in, but from what I’ve heard, not many people are.
If the company you work for truly does not value this effort, then do not do it.
It’s not your code base. It’s theirs. You are not being rewarded for saving them from themselves. Don’t work for free.
Agile doesn’t claim that a project can be completed in a sprint.
Tell that to middle management
My middle management knows this. I don’t stick around for shit management, we don’t deserve each other. There are always other opportunities.
Plus “getting something up real quick” is the fun part.
The first draft is fun.
The second draft is pain.
The third draft is cathartic.Figure out features, add add add.
Add/change features, realise the spaghetti mess and poor design decisions you made in the first draft.
Clean everything up with better design and code.
i think the rough ride is a necessary learning experience
Every couple of weeks, have to stop, go through every line and module, and throw out pretty much 90% of the code
It offers a good starting point
It doesn’t sound like a good starting point if you have to throw out 90% of it every couple of weeks.
Drag feels schadenfreude for them. If they’re going to fire their workforce to chase trends, it would be fun for them to go out of business about it.
As usual, people assign conspiratorial motives and strategies to behavior that’s really an extremely simple straight line between two points: “AI software has a lower apparent cost than hiring another developer, so let’s use AI.”
If you run your AI, point doesn’t matter. However, what matters more is the fact that if you don’t use a skill, you just straight up lose it and that’s what AI is doing to developers. Mfs straight up forget how to write code
I learn a lot debugging the code I get from AI, and occasionally, I learn a thing or two.
Really wish they’d be a direct link to the source, not solely a screenshot. Is this the Web?
Well if it helps for y’all to know, if I can’t put my measly webpage making skills to decent use in the course of a weeks time, I’ll be buying the services of a freelancer because hoooooly shite am I rusty.
(I need to update my basic website and am terribly lazy. Maybe making some extra cash would make a kid somewhere happy.)
((Don’t message me here though I don’t check messages))
On the other side, if it’s “deskilling” to do vibe coding instead of real coding isn’t this person saying that the barrier to entry for coding has been lowered?
Either vibe coding is not effective and is therefore not taking away the skill of coding or it is effective enough to replace aspects of coding that you would otherwise need to develop the skill to do.
Like if I’m an engineer or a real estate agent or a business…dude, and I want to use coding in my field but I don’t have the time or desire to start learning a whole skill (anywhere from having children to just learning too many skills already) I assume vibe coding is my best friend.
I think it can do some stuff, especially some entry level tedium.
So far I haven’t seen a single success on the specific things I’ve tried it for, even when pretty short, other than exceedingly trivial things like reminding me whether this language has a join as a string method or as an array method of o don’t use it that often.
I do see potential for an awkward gap between unskilled and skilled where an entry level person doesn’t have as clear a path to getting actually better. In math this generally happens in school, where they keep students from using the most effective tools until they prove they can do without it. So education might have to go a bit further into programming skills rather than delegating quite so much to the professional workplace that may be less inclined.
Im not going to lie, I totally vibe code. Ive been using it to build guis that help speed up repetitive processes. Vibe coding has been helping me learn too code. I think people abuse it for sure. The code still needs to be checked since LLMs are about as trustworthy as Quora.
I have no idea what vibe coding is, can someone ELI5 it to me?
I have tried AI to get some rough C# for my hobby game but even that was unusable.
Vibe coding is basically having no idea about coding and using the AI to make snippets of Code for you
Like if you want to programm snake, you would prompt it:
- Tell me what parts of code are required to programm snake in python
then it would tell you like:
- you need a programm to make a grid system
- you need an array which can go down a tickrate
- etc pp
so you tell it like:
- Generate me code, that does xy
- Generate me code that takes the input of xy and does z with it
and so forth, then you just paste everything into a txt and ask the AI to debug it for you and hope it works
The people who need vibe coding shouldn’t be using it. And the people who can use it, don’t need it.
Idk about the last bit. I’ve done some vibe
codingdebugging to fix game mods written in languages and frameworks I don’t know and have no interest in learning at the moment. I still look over the output, but given a lack of knowledge, I’d still consider it vibe basedI don’t have the bandwidth to know enough about everything I encounter to be passable, and sometimes I just want to make some random thing work with the minimal amount of effort so I can get back to the actual task at hand.
This sounds terrible, lol! Are there any examples that can be pointed to? I’d love to see one of these constructs.
On tilvids.com some dude called picopixl is doing tutorials about this
https://tilvids.com/w/oyddhsnfHUFToBEmpEZpEg
And yeah, its pretty great what it could do, but for someone who (is his own words) can tweak the code so it works, it tool longer to make a Prompt than just coding the Game yourself
Also, Tetris in JS is like Babys first JS project, so even if you really wanted to just get Tetris from somewhere, you could have just git pulled any github project
‘Vibe coding’ is where you code only with prompts and never look at the generated code.
Seems like a great way to create insecure unmaintainable code if you ask me.
Also I just dont get why you would ever generate code
Like, you have no idea how to code something? Sure, just ask it about methods how to do it. But generating code too? Cant you RTFM?
I think you’re severely underestimating how lazy some people are, lol. I totally get what you’re saying, and from a logical perspective it makes sense. It’s just that if you survey enough people, i really think you’d be surprised at how little effort some are willing to put forth for just about anything
Isn’t the reason obvious? To save time? I’m not saying it’s a good thing but it seems prettyyyy obvious why people are doing it.
But it’s going to take hours of debugging every time. If you actually learn how to write code, you’ll get better at it over time and reuse common functions. It’ll take less time as you get better.
Well… Just because you can code does not guarantee you will find it enjoyable. It’s pretty common for people to like certain aspects of coding but not others. For instance, I personally find writing unit tests boring. So if something came by that made writing them less mundane I would certainly be enticed.
Yeah no. For example microcontrollers, which are how I learned it. There are so absurdly many traps to fall into that even within the first 10 things I did I ran into some obscure detail of the ATmega328p. And the kept happening ever since, each time lots of googling and trial and error. Now with GPT you know how much time this saves? Not just the coding itself, but also these absurd details that only an expert knows. Yes perhaps it does the same error, but after reiterating it usually sees the problem. I can also throw some datasheet for some chip at it and get exactly how to program it with what setting etc. It enables me to do FAR more advanced things. And the new 3o and 4o mini are really much better again. Code often works out of the box now.
I love how people simply downvote me for saying these things. As if me repairing an E-Bike is somehow a bad thing, just because an LLM was what enabled me to actually do it. As is programming for some reason should be an arcane art. As if technology has not always been changing in exactly the same way. With people saying “The way you do it is wrong!” and 20 years later nobody does it the old way. LLM have exploded in a few years and have come extremely far, what do you think they are capable of in 10 years? All those silly little mistakes they still do, gone.
I think we’re still in the Win95 plug and play phase. Granted I have been using Claude 3.7 thinking and I will try o3 mini or whatever I have access to via Kagi tomorrow, but today I spent 2.5 hours trying to get AI to vibe code me a bash script that could read a system group member list and write it into a users .k5login file when run. It came up with lots of stuff that looked good but in the end still made an empty file. I even tried to help it with what I guess are oddities of our system but it didn’t work. I tried editing it enough to where I felt like I was very close to just writing it myself with documentation hints and ran out of time.
Idk if it’s me not prompting right (but if so then I’m going to have to learn a proxy skill rather than just learn the skill which seems silly when it seems so opaque to learn the proxy skill. Like if I need to ask for help to prompt and it’s complicated why not just learn the code?) or just that as soon as you hit a non public even slightly customized environment AI just doesn’t have the context necessary and there’s no easy way to get it to it securely yet or maybe Claude isn’t that good (but it was pretty loved for like 6 months for writing code)… Or the hype works in very limited ways and IDK when that’ll change.
deleted by creator
Save Time where? If you want to code more than snake, you need to have a basic knowledge of coding anyway, and once you know how to code, you will want to code in your own style. And if you just want to make basic programs, just fork someones github project and change a few lines.
You’re saying this with your understanding of the field. The people pushing this are either untrained (and thus don’t know what’s going wrong) or are trying to milk money out of the former.
Not all code needs to be held to the highest standard. Sometimes you really just want a throwaway script.
