There’s so many AI slop decomps right now, it fucking sucks. There’s a lot of people that doesn’t even understand the basics of coding doing it just to get these stuff running and it’s so obvious it is gonna be abandoned because the prompter doesn’t even know how to fix the bugs people file for their projects.
This is specially bad in the game decomp scene. Good thing there are actual good recomps/decomps being made, like SymphonyRecomp and MediEvilRecomp.
I get using AI to help with the more repetitive stuff and under really strict supervision, but there’s way too many stuff just entirely vibecoded right now.
There’s so many AI slop decomps right now, it fucking sucks. There’s a lot of people that doesn’t even understand the basics of coding doing it just to get these stuff running and it’s so obvious it is gonna be abandoned because the prompter doesn’t even know how to fix the bugs people file for their projects.
Yes and the existence of the vibe coded project deters new projects with real effort from even occurring because people don’t want to compete with other projects or stir drama by being a competing project. Additionally you know anything the serious project does is just gonna end up contributing to the vibe coded project which is a good reason for people not to want to do it. The very existence of this shit reduces the existence of better things.
Welcome to being an artist in the age of AI art!
Eh, ‘Yet another’ is like at the core of the computing community. Nobody’s out there complaining about a lack of Doom ports or Linux Distros.
Those weren’t competing for attention and brainspace on every online repository. We’re going to have to make “non-shit” lists for everything into the future because some mouthbreather discovered he can make some simulcrum of a finish job for like 12 things in a day.
Before if you wanted to make a fake game trailer for a game that didn’t exist it took time to make it.
It’s 2032, you’re the last person on Earth subsisting on a stockpile of cans of proteined peaches and nutri-soda. You have a working internet connection to the human knowledge Xanadu Satellite Network. You only have an Intel Atom-powered netbook (has a disc drive somehow), a Philips CD-i, and one CD-RW. You want to play Doom on the CD-i, but thetr are 72 decomps on the forever-archived Github repository and all of the archived comments on their pages are spun-up bot networks accusing the others of being vibecoded.
You decide to download CDoom-i [NOT VIBECODED].
You burn it and start it up. The D-pad inputs are 90-degrees rotated to the right.
It’s vibe-coded.
I’ve tried the whole “vibecoding” thing on a personal project and, honestly, the thing that would have made it work better would have been talking to humans. especially in open source, if you’re expanding or modifying a project you’re better off asking people who know more about the software than you.
I bet that project is/was full of weird bugs and problems that would’ve never occurred if the author consulted the WINE community.
Damn that must suck, the only time I used vibecoding was on very simple scripts where I got stuck and needed help and once when I didn’t feel like doing it, but I always made sure to understand what the code was doing anyway.
What you’re describing doesn’t sound like vibe coding, but a sensible approach to programming with LLM assistance.
The vibe coding part was that I do not have experience programming in golang, so I effectively gave full autonomy to the agent to write the solution, and I simply had to assume it worked correctly. I got an output at the end, but no idea how I reached that point, and even then it didn’t fully work (and I wasn’t about to spend my $2 of deepseek credit finding out) and I had no idea what to do beyond “here’s error logs pls fix”.
When I’m doing what you describe, I’m in a domain that I understand, and I can make decisions on the solution. Even then, I rarely tell an agent to do the coding - I just copy a snippet from copilot or press tab on the autocomplete hint.
Yeah I worded it weirdly, I did do vibecode one script once, where I just told the AI to create it and just trusted it since it was very simple. For the other scripts I only used it as a tool to learn myself what to do, specially since I didn’t quickly found out what I needed to do when searching the internet.
Ahhh I get you!
Thanks for informing me about those decomps. These look awesome. There are lots of sites with listings of ports/decomps, thankfully this pcgamingwiki site lists if AI is used or not.
Oh great I was wondering if there was a site that did that
It’s getting one foot in the door though. Humans can deslop the code later on
“one foot in the door” is as useful as an “ideas guy”, which is not at all. Producing slop themed around a project provides no value for anyone else who actually wants to work on it. If anything, skilled humans avoid working for free on projects or ideas associated with LLMs.
As you are an engineering student, you should be embarrassed for posting this comment.
I call BS, I’ve used AI as a coding assistant semi-often and while its output isn’t the best, it more than makes up for that in the speed increase. The impressive part of Photocraft is that it was made in under a fucking week.
If anything, skilled humans avoid working for free on projects or ideas associated with LLMs.
Yeah, that’s why their GitHub has over 270 contributors and 57k stars. I have nothing to be embarrassed about, don’t let your hate for AI cloud your judgement.
It’s way, way harder to undo bad code than it is to write good code first.
Not necessarily true in my experience. Using AI as a code assistant ends up being faster and relatively the same difficulty for things that I’ve worked on, with the exception being niche use cases and libraries.
install GIMP
I already do. It’s subpar at best, feature and UI wise. And yes I’m aware of the PhotoGIMP UI thing.
deleted by creator
But it’s baaaaaaaaad
My hope is that this type of shit leads to better decompile techniques not using LLMs. I watched a video that said that one of these decompiled vibecoded apps cost $180k in tokens on a $200/month plan
Who is dropping $200k to “decompile” photoshop
nobody, “$180k in tokens” doesn’t actually mean $180k is changing hands at any point in this process
It does mean that though, in expense in data centres and energy.
In this case, Anthropic. They let someone paying $200/mo use however many tokens they would charge $180k for at their current API billing.
Since Anthropic subsidizes even API tokens at something like 10:1, this decompile may have cost Anthropic on the order of $2 million.
I watched a video that said that one of these decompiled vibecoded apps cost $180k in tokens on a $200/month plan
For that much you could probably pay a team of coders.
No one paid that amount: it’s the dollar equivalent of the billions of tokens that were used by the LLMs. The “devs” only used a $200/month subscription to do the decompiling. This business model is why AI companies aren’t profitable since compute is wildly more expensive than what they’re charging users.
AI slop ends up costing almost as much as a senior developer if you use tokens/hr as your metric. Except a senior dev is responsible and trains juniors.
Ah. Hell even for $200 I’m sure you could hire someone freelance though
For an hour maybe. This kind of work takes a lot of talent. Or nowadays… a prompt 🤷♀️
IDK this stuff is only gonna get better. It’s gonna be crazy seeing what code is being put out in a few years by these tools.
So vibe-code LLM decompile is a thing now? You feed the compiled binaries into the LLM and it makes source code based on statistical models I’m assuming is how this is supposed to work?
Seems like just using LLM hacking agents to get the source code from the owners is probably a more effective approach. Desloppiled Code is just going to be a huge mess every time I’d imagine.
The dev claims they are making a clean room replacement not a decompilation.
Clean room usually means having one person/team reverse engineer a product, write a detailed specification and have another person/team implement that specification so that it’s not contaminated with copyrighted material.
I have my doubts about the clean room part but it’s believable that it’s not a decompilation project that aims to come close to the original code structure.
Desloppiled Code is just going to be a huge mess every time I’d imagine.
For sure. No human is ever touching that code.
Oh, I recall seeing a post months ago about this LLM clean room approach being promoted as a service to “help remove license restrictions” on open source projects. They would deliver you a new chunk of code as a drop-in replacement for “restrictive license” existing code.
So now a similar approach is being taken to do the same for proprietary applications and games. Fucking hell this is so damn bleak.
Wasn’t that website a bit? Not that it’s hard to believe that this would exist (as a commercial service, we know people are doing this with coding agents).
Perhaps I ate the onion on that one, but it’s still a legit and active issue despite the service being a farce.
Wasn’t that website a bit?
The site is https://malus.sh/ and it is definitely satire but I think it also actually sort of does what it says it does (albeit surely not as reliably as they imply). 404media says that it is “satire but also fully functional”, at least.
The creators of the site launched it with this (CW: slop-heavy) FOSDEM talk: Let’s end open source together with this one simple trick
It’s like if Jonathan Swift ended A Modest Proposal by plugging his actual newly-opened butcher shop.
It’s impossible to hate these people enough
I thought that was a joke though, wasn’t it?
I clicked through to the site that was linked in the post, and it seemed pretty legit looking, but maybe it was a ruse. I can’t recall the site/service name now though.
A person sent in a patch to get adobe working in wine recently but it might be till the end of this year when its merged in wine proper. They apparently consulted an LLM but it was as simple as making XML parse properly.
Actually existing software is a fuck.
AES 😀
AES 😧
actually existing software
This might be the weirdest thing I’ve read on Hexbear this week.
A decompile of sonic the fighters came out for pc and it was only up for a day before being taken down. It was vibe coded and didn’t even work right.
If the app compiles to the same binary, it is not vibecoded, just the original code, I guess they just fed the model how the app is supposed to work and the binary so the LLM just guesses the library dependencies and some strings if the Adobe binary is not obfuscated
I feel like GUI coding would actually be pretty LLM friendly if it were paired with hard constraints and interaction trees (that were done manually or with a framework, the latter being much better for accessibility). Unfortunately people seem to mostly just take whatever it gets and say, “looks nice” and ship it.
