"I built this"... No, you didn't!
tl;dr a rant about new accounts advertising their vibe coded projects.
Since a couple of weeks I notice more and more brandnew accounts from various instances joining the threadiverse, just to post about their new service, software or game "they built".
No you didn't built this!
Every time I looked into the repo (if provided) there was a fucking AGENTS.md telling a clanker what to do.
And now you're coming in here. You don't care about the community you post in. You don't want to participate and contribute to this place. Your first post is a fucking ad for your new thing that "you've built".
But you didn't built it!
I'm not a full stack developer, but I've worked alongside developers who built amazing stuff for the free and open web. The lenghty discussions about wether a feature should be implemented this or that way, the long nights of trying to get a feature ready, only to throw it all away the next day because you realized it wasn't that good of an idea after all. It's hard work.
You just had a quick idea, told the billionaire bot what to do, and vomited out thousands of lines of code in a couple hours.
No, I won't check out your new thing that "you made".
And that's the sad part. Open source used to be fun. Discovering new stuff that others had built, giving it a try and maybe finding a small bug and opening a pull request. That feeling of working together on something that can be used by everyone, despite not knowing each other.
You don't have any of that. And you don't deserve my attention, because you didn't build anything.
36 replies
I maintain an open source project with a modest following and got my first ever LLM-generated PR this last weekend.
The code change itself was small and largely fine, if somewhat over-commented in that way that LLM code often is. It was a small improvement for scaling one particular aspect of the UI at higher display scales.
The PR description was entirely LLM generated as well. It's quite possible that I was the first human to actually read it, including the "author".
My response:
They went off like I had just personally insulted their family. Posted a huge rant about how AI was the future, I was going to be left behind, and that my project was unusable on 2K and 4K displays due to the issue this PR was addressing.
I posted a follow-up saying that I actually didn't mind the idea of accepting code that had be written with LLM assistance, mostly because I don't want to wade into the mire that is trying to weigh where every small change came from, but I expect human written PR descriptions. I'm not interested in interacting with meat proxies, if I wanted to ask AI to fix all the bugs I was aware of I could just do that myself.
I blocked them at that point, so can't tell you what they may have said in response.
Aside: Ignoring the unrelated mess that is what "2K" actually means, I've been using 1440p displays for a decade and a half, and another regular contributor runs at 4K, so I have no idea what this person's bar for "unusable" is.
This is the way.
llm stuff can be valuable, but not if the burden is just slopped onto the maintainer.
As a potential slopper, I just blew a few days budget on Argon to do a security review of an open source rust project that's important to me.
It came up with some findings, but nothing I found concerning.
Now that report might have some useful stuff with useful fixes. But I'm not just going to slop it over to the maintainers to deal with.
I'll either put in the work to understand the changes, consider if they're worth the overhead, and put together a human PR that I'm comfortable with, or I'll leave it alone.
Your work should be more complete with AI. It's mostly fine to slop up tests with minor review. It's fine to use as a tool. It doesn't replace the human, and if you think it does it's probably because you're trying to shift that burden to someone else.
Could be someone with a visual impairment. Doesn't excuse their behaviour or pushing an LLM written PR with no actual interaction with the process.... wouldn't really give them the benefit of the doubt though. Sounds like someone with a "solution" looking for a "problem".
I agree with that rant. You're right that's it's annoying.
I feel as if AI/LLM projects have had an overall negative effect on open source communities. There's a constant flood of new projects. It's overwhelming. With all those new projects is a different attitude and that attitude has changed the vibe of these communities completely.
People that use AI or aggressively promote AI care very little for existing community rules or are constantly pushing those boundaries. They get hostile and feel entitled to more when they are called out for not disclosing their AI usage. They constantly push AI on others who are asking for help or advice. They belittle and insult those who do not want AI code/documentation. They act like this technology which has an extremely short existence in comparison to human existence is inevitable and must have complete influence over our futures.
The promoters of AI give off manager vibes. Someone who desires to control someone or something else and then claim all that work as their own. There's very little thought as to how the work was completed or the quality of that work. I don't like bosses or managers so this attitude is not something I like seeing in my communities.
I think this blog post does a good job explaining the effects of AI on an open source community: https://blog.fallible.net/it-breaks-a-village/
As much as parts of the FOSS community has been toxic, especially to minority devs, it’s still one of my favourite creations of humanity. It’s a Library of Alexandria moment that LLMs have been trained on pretty much every line of FOSS code written and accessible on the internet, commodified it for billionaires, and is now being used to attack those projects it stole from in an avalanche of slop. It’s what makes Linus‘ recent statements such a betrayal to me and makes me wonder why Linux & git were ever open source if that’s his attitude to all this.
What all these AI heavy devs don't understand (or care about?) is that I don't need you as a proxy for the AI. If it keeps improving in the way they hope and expect, you will eventually be able to type "build me an app" and it will just do it.
That sounds great on paper until you realize that you either need A: a subscription to a Big Tech company who is on record saying they want to put human knowledge on a meter, or B: Access to hardware that is priced out of reach for many people (due to the actions of big tech) if you want a useful, performant, local model.
LLM tech is neat if you look at it in a vacuum. But you CANNOT do that in the real world. Most of the force behind the AI push is from people who want to put skill behind a pay wall they control. I cannot advocate that, even if claude can rewrite Linux in Rust perfectly. And it greatly upsets me that there are pillars of the OSS community (looking at you, Linus) who refuse to consider that because it is personally useful to them.
Unfortunately, we've been put in a situation where you're damned if you do, damned if you don't. And the blame is solely on the devs and big tech personalities who move fast, break things, and do not think about the effects of what they put into this world
It may be anecdotal, but the push to use LLMs seems to come more aggressively from older folks. I always have to remind my older colleagues that, if we don't open new positions because "LLM can help us do more", we end up blocking newcomers from starting a new job in IT. It's as if, since they already have a secured position as seniors, they don't really bother thinking about the rest of us...
People are too worried about "being productive" because AI is being sold as the ultimate productivity tool that will totally make you boatloads of money because you're one prompt away from making your million dollar idea into a real app/game. They're falling for it, testing it out, seeing success (or "success") and thinking it's the first step into their millionaire lives. It's the latest get rich quick scheme, only this time it's dragging half the world and its economy with it.
Me, I'm trying to learn my way around SDL3, which seems especially interesting as it also has a GPU API so programs should, in theory, automatically use the appropriate backend (Vulkan, OpenGL, DirectX, etc) without having to write in said spec. Once I have a better idea of what's going on and what I'm supposed to do, I might create a FOSS game engine and a bunch of video classes explaining how and why I did stuff; so, don't expect anything concrete within the next 6 months! I love being unproductive and those hustle-life fuckers can go fuck themselves
Speaking of fun open source stuff you can discover out of nowhere - Licar is written in pure C with almost no dependencies.
As someone who is building said Game Engine in SDL3 using SDLGPU, I will have to warn you, the GPU api is missing a lot of features and suffers from some very annoying limitations that you have to workaround. The engine is still in development. The only reason I have not published the source yet is because I have still not figured out a satisfying multithreading architecture and also because i am showcasing parts of it as a 'mini project' for college credits. It should be finally ready soon, I have managed to get a doom style FPS working satisfactorily and the next thing I am fitting the engine for is Open World style asset streaming.
That aside, it was the first big project that I tried using LLMs for, and mostly they were used for boilerplate(SDLGPU has a lot of configuration that needs to be laid out, moreover since this is an engine, there needed to be a way to change the config from the scripting layer) and generating giant switch-case trees for input handling.
beyond that I have found that if I have written two disconnected modules that I need to connect, and if i can visualise mod1->[some sanitization]->mod2->output, I can confidently let the LLM handle that piping if i described it, but if it was even a degree more complicated than that, the LLM tended to do a poor job or radiated outwards and touched modules it shouldn't(not that it didn't give me the end result, just that it over complicated the solution or introduced 'slop')
I agree with your sentiment. On a side note if English isn't your first language just so you know built/build can seem very similar. Standardly one would say I built this in past tense. If someone says you didn't built this for some reason it's wrong where I'm from, it reverts to the present of build. "You didn't build this". Not sure 100% why.
I want to build a bridge vs I built a bridge. Maybe it is because when you said "you didn't build this", it is in the present accusation... English is dumb
In my nowhere near professional opinion, I think it's because in the phrase "You didn't build this", "didn't" is in the present tense, so "build" has to follow in the same tense. If the phrase were "You hadn't built this", because "hadn't" is past tense (or at least acts that way), then "built" would follow. My only justification is that this sounds right to my ear, and it's a loose theory as to why. The only consistent thing about English is the inconsistency
You are almost correct; “didn’t” (or “did not” in its uncontracted form) is considered past-tense. Since you’ve already shifted the tense to the past, you no longer need to do so for the following verb - you are “already in” the past!
So “didn’t go” vs “didn’t went”, “didn’t build” Vs “didn’t built”. The verbs following a past-tense auxiliary verb retain their simple form.
How big are your margins with this much of side note?
Touche
Reminds me of the people who write a tl;dr that's twice the length of the thing they're summarizing.
Ha, I did indeed struggle with the grammar there. I know that "did built" is wrong, but that one must've slipped. I'm not going to edit it now, since it would take value from your comment. Thanks for the insight. English isn't my first language indeed.
Haha, no worries about if you change it. I'm considered to be crazy by enough people that it doesn't bother me. : )
I think it works. It's a way to add emphasis, kind of spitting back what they've claimed.
You're not processing it. You're not adapting what they've said to proper English.
The debate isn't wrong, but this one seems acceptable to me as a writer's flourish, consciously or not.
I’ve come across a dozen projects in the past month or so that I’m genuinely interested in looking further into, riiiiiiiiiight up until I hear about the AI in it.
I don’t comment about it or anything, just back out of the thread and tag the poster as a vibe coder and move on.
If I wanted something vibe coded, I could have a model generate it for me and tweak it to my liking at least as well as some random shmuck on the internet. Thats not what I’m interested in. I want stuff people understand enough to fix problems, and that means largely human written.
I have several games/vr spaces I would love to make, but they would require using AI (I don’t know shit about coding or development, and don’t particularly want to learn tbh, its a huge entry barrier that I won’t be hurdling any time soon) so they simply aren’t going to happen, just like the pre-clanker days. If I’m not even willing to put in the work for my idea, its probably not that good an idea to begin with. And that’s the same way I feel about vibe coders.
Was just having this thought. People who think "making" or creating with botslop is actually real, are idiots and delusional. At least its a red flag to stay away from them because they are probably out of their mind on other topics as well.
But they commissioned it themselves in a model which is non consentually and illegally trained. That has to be admired! 😆
This is how I feel about 3d printing.
All programmers who use an IDE and don't code from scratch every time are a bunch of hack frauds. No shortcuts!
https://en.wikipedia.org/wiki/Straw_man
https://en.wikipedia.org/wiki/False_equivalence
I'm sorry, did you IDE code your program for you? Go back to punch cards and get a real programing job.
I mostly disagree with your rant, as it paints ALL LLM-generated code in bad light.
Which isn't really the case. Quality heavily depends on the prompt, the person prompting the model, the model itself, the harness, the codebase, the context, the list goes on.
A skiddie asking an LLM to make "the next Facebook" will fail. Awful code quality, barely working spaghetti, impossible to maintain.
But put an actual engineer in front of the LLM and you'll get much higher quality code. Because the engineer knows how to do things right AND how to get others - including the LLM - to do the right things.
Maybe. I'm an actual engineer and I've done both. I'm experimenting and learning the difference. It can be easy to get lazy and vibe until your project is eventually spaghetti.
Most of my work stuff is high quality. My integration tests are half slopped together.
I've made a vibe coded personal project, and it feels like vibe coded crap. I'm still iterating on that, but if it doesn't work out I can just trash the whole thing. Importantly, nobody has to deal with the slop but me.
I'd recommend utilising the various engineering skills and agents.
One of the prime issues of LLMs is that they know so much, but don't know when to apply what. For basic queries sure, but once you start going into the technical nitty gritty details, it's easy for the model to start reaching and making up shit.
Well written agents and skills can keep you on track, a well developed loop replicating real life processes (e.g. the design-develop-test-verify flow of software development), so you're not chatting with a bot, but spawn an entire team of agents that are experts in their own fields, and they pass tasks around until refined, developed and tested according to parameters. Hell you can even trust entire design cycles to them if you use the right agents.
I agree as well, but what are you going to do when all three of the major operating systems are vibe coded?
Linus Torvalds posted recently that he's vibe coding the Linux kernel and he had strong words for people who are against vibe coding. He makes a couple decent points, but agree with him or not, using Linux past a certain point means that your OS is vibe coded, so do you really have a strong leg to stand on?
Microsoft has said they're using Copilot to shape the next version of Windows (even calling Copilot itself the OS), but we don't know to what extent.
Heavens help us if Apple gets Siri/Apple "Intelligence" to vibe code macOS or iOS. Because it's the derpy AI. It's like that hydra meme, you know which one Siri is, you can label the other two whatever you want. Of course if Apple were going to vibe code their OSes, they'd probably just use Claude. I don't think it's happening now but I can't rule out it happening in the next 3-5 years. I think by 2030 most software will be at least partly vibe coded.
Continue to criticize and resist it however I can just like I do with everything else. I don't accept the way the world is and try in my small ways to change it.
this is pretty controversial opinion here but i think there's difference between 'vibecoded' and 'ai assisted'
the shitty stuff like what op mentioned fits the first category, while linux and other stuff is the latter
like it or not many people/devs like to use it, especially if they're doing it for work because it makes it easier to churn out large amounts of code and features (and ofc bugs). pretty sure apple employees already use llms for developing macos/ios.
i (mostly) agree with the frustration against llm stuff.
Arguments about if how we obtained the model’s weights in an ethical way (we didn’t) aside,’using AI as a boilerplate/autocorrect/autocomplete machine is the correct way.
Prompting an entire project and then running with it without understanding it first is not.
It's hard to find the gray area between the corpos stupidly shoving ai down our throats with the idea to replace all labor and those who are pushing back by hating anything to do with AI.
They're useful. They're not a trillion dollars in the next decade useful.
They help with some problems. They cause new problems.
You can use ai as a tool without creating vibe-coded slop. And there are situations where vibe coded slop can be fine. (Usually personal, non-critical things.)
LLM generated code performs worse than human written code, and costs more. Once the investors realize the ROI is never coming, and AI companies have to charge the real cost, the narrative will change.
"what are you going to do when all the major operating systems are ... ?"
Someone must have heard a similar question back in the Unix days.