I still hate the “vibe” terminology.
What I would have liked it to mean: While coding, put on some music, and zone out to coding.
What it means now: Prompt an AI to generate working code and solutions.
I don’t get where the “vibing” comes in. I guess you don’t have to think about the technical details? And that’s vibing? Maybe it’s just unfamiliarity and lack of practice, but poking the AI via prompting and thinking about how you can influence it better doesn’t feel like you could zone in to or “vibe”.
Maybe it’s about letting go of reasoning and just going for it? Vibing in the sense of going with the flow?
It’s not the first terminology I find unfitting. I’m trying to accept that it is what it is, and that it just is what “we collectively” have decided to call it (or ran with).
I assume the term is referring to the person knowing only the vibes of the program they want to code rather than comprehending the details of what it needs to do.
I hope vibe coding is a joke like “I just paste from stack overflow”
But I have met people who do just paste from stack overflow, mixed indentation and casing and all.
As a software engineer who was introduced to the field in high school, over two decades ago: it’s real. People are doing it. I’ve met people doing this at my job, and directed focused scorn at them. You know the idiom of “knowing just enough to be dangerous”? Like, you’re learning C and you’re just playing around and you discover pointers, and you’re like “oh this is interesting”, but you haven’t learned or internalized that it is real easy to Fuck Things Up if you don’t use them very carefully (there’s a ton of stuff like this)? LLM codegen being used by novices is an absolute shitshow because the codegen will often create nonsense, broken, logically flawed, or deprecated code, and the novice user is just going to accept it at face value instead of understanding that it’s subtly wrong.
That is just horrendous. I am not a coder by trade or education, but I do a fair amount of it in my job. I may heavily use copy and paste, but I at least make sure it matches the structure of my current code, and I always include the URL to where I got it in a comment above the code.
My org is encouraging it…
In the meantime, I am interviewing juniors who can’t write a while or for loop without looking on the internet…
The future is looking grim.
Should ask them to make an if loop and see what happens
As a senior programmer I can’t write a for loop without the internet.
I can’t remember the last time I saw a job listing that didn’t expect me to be an expert in at least 5 languages. The best part is that halfway through the interview you learn that they are no longer using half the languages listed, and are “transitioning” towards 2 others that aren’t even listed. You want me to whip out examples written in Fortran, C++, Rust, JS, and some random word you claim is a language in 2 hrs without the internet? Bitch, I don’t even think I could get prewritten “Hello World” examples compiled in 5 different languages in 2 hrs, much less on machine that I have never seen before.
This seems like more time and trouble than just writing the code yourself.
(And while the same could often be said of assigning tasks to junior programmers, in that case it is an investment in the person rather than strictly a waste.)
I swear, I wholeheartedly agree. By the time I get even something slightly like what I want I could have wrote it myself so I ended up saving zero time; there are instances where I actually save time, but it’s about tedious and simple tasks (I had to search a K-tree) or to refactor certain parts that are more or less the same. All the other times either the code sucks so much I end up writing the whole thing myself, or i know the steps myself so what’s the point of detailing a prompt that will require my intervention instead of just writing the code? Line completion is the other good thing about this wave of AI at least
I have no idea how people can use LLM-generated code. In my experience they’re absolutely terrible. However it can be good for giving some insights time to time.
I only use it when I know exactly the code I’m trying to produce, but just saving time if it can write it for me. Somewhere I saw this described as ‘toil’ vs. ‘domain knowledge’, and it definitely reduces toil even if I have to correct it. Anywhere that I wouldn’t know how to correct it, I don’t trust it.
I only used LLM-generated code once. I tested and made modifications to see if it was I want. It worked out.
Probably because they’re not good enough to know any better.
It’s definitely improving. I thought the same as you but I looked through my recent ChatGPT prompts and it’s actually decent now, at least at simple/throwaway tasks. It doesn’t stand a chance at the niche domains of my actual job.
I find that using AI can sometimes help me get the ball rolling. Rather than having a blank page, I have some (sometimes shitty) code that I can start from. I have never used 100% of generated code in production.
I use it to generically explain things to me. I give it limited things and ask it to break it down or give me alternatives.
No different than googling skills, how I ask gets different results, and sometimes it’s far faster than digging and digging through forum posts or textbook style documentation. Other times it’s a waste of my time and I quickly move on.
Yeah I agree. It’s often easier to start from something that’s wrong than a blank page.
I often refer to ai as an index for human knowledge. It’s pretty good at expanding on the context around what you ask it. It’s great for getting started or to help point you in other directions
I often compare vibe coding to lord of the rings. Saruman blocks the fellowship’s path with difficult challenges. So too does solving hard problems in programming. So Gandalf decides they will take the mines of moria (vibe coding). He knows better but does so anyway. The rest of the fellowship naively follows him down (junior devs). Most of the path is just minor hiccups and the juniors fumbling around. But they get to a certain point and things start to get too heated. The hordes of goblins being the bugs introduced by the LLM as they keep changing the code via different prompts. Then they inevitably awaken the Balrog… the monstrous Complexity Demon that was brewing behind the vibe coding sunshine and roses.
The bosses don’t deserve high quality work, fuck them. Micro-sabotage the machine anywhere you can.
This… Is actually unironically the best argument I’ve heard in favor of AIs so far, that I haven’t thought about.
Still - the thing you’d be doing instead is feeding money and attentention to AI bros, and that’s probably even worse than any job you could be micro sabotaging.
Enough micro-sabotage by enough people will in time corrupt the databases enough that people will stop trusting (and at some point stop using) them. It is happening already if reports of deliberate disinformation in the LLMs can be believed.
I don’t think you need any active sabotaging in this regard. I’m not really worried about the future of LLMs, because we are already at a point of feedback cascade where thanks to LLMs, more and more of content they steal from the internet has been AI generated by them anyway, which will eventually cause the models to collapse or stagnate. And besides, you wouldn’t be able to sabotage at a scale required for this. Thankfully, the spread of fake AI generated websites and content it has enabled is so massive, that it works as well.
I’m looking forward to that.
I tried it (Mostly cursor, cline, claude code and the new copilot agent mode) a few times, maybe it will get better at some point but right now I am incredibly slow doing it and the mental load reviewing changes and trying to get it to do what I want is incredibly exhausting.
It is an excuse. It’s just a bad one
