The website decomp.dev said Thursday that it is shutting down because “the community aspect of decompilation has been broken” by AI, and that projects that once required intense collaboration between people with different skills can now be done trivially with AI. The administrator of the website said that it has become impossible to keep track of which projects are primarily human-led and which ones are done by AI.
“Historically, matching decompilation has been a brutally difficult, time consuming task, on the scale of years for any moderately-sized game with a number of active contributors,” the pseudonymous admin of decomp.dev, who goes by Encounter, wrote on the site.
But recent frontier AI models like GPT-6-Astra have made it trivially easy to decompile a game: “They are significantly better at following complex directions, cross-referencing resources for symbols or debug information, and writing natural code with enough guidance,” Encounter wrote. “It is now clear that, with the right guidance, human input, and code reviews, AI can expedite the decompilation process in a significant way. What is also now clear is that the community aspect of decompilation has been broken.”
AI-powered video game decomps and recomps (recompiling code after modifying it) are everywhere. People have, for example, used AI to put Mario into Elden Ring, port games to platforms they were never designed to work on, or to combine various games in any permutations you could possibly imagine. In recent days, tools specifically designed to decomp games with AI have popped up, such as "decomp.wtf," which promises to "point the AI agent you already use at a game you own. decomp.wtf turns it back into C, one function at a time, and checks every function byte for byte against the original."
Decomp.dev was a collaborative website that hosted file libraries and tools for reverse engineering, re-coding, and decompiling existing video games, meaning taking an existing game and recreating its source code. Decompiling a video game can be done for a variety of reasons and is popular in a few different niches of video game fandom. Speed runners decompile video games to find potential glitches or advantages they can use to beat a game faster; hobbyists decompile video games to port them to new platforms or to emulate them (for example, making Super Mario 64 run natively on PC); modders decompile games so they can more easily modify them; archivists decompile games to recreate code that may be lost; and some people decompile games just for the challenge.
Encounter wrote that the use of AI in the community has sped things up to the point where it has become overwhelming to track human-led projects, and people using AI have taken existing, human-led AI projects and have created “hostile forks” of them, where open-source code written by humans is taken and “finished” by AI: “These kinds of hostile forks put little to no effort into meaningfully contributing to existing projects, tearing apart our community fabric.” Encounter added, “anyone with an AI subscription can manage to create a port or static recompilation that functions at a basic level […] While community members put in the effort to decompile a game, opportunists have frequently raced to release a half-baked, barely-functional port from it.”
What the fuck are you on about? You're talking about abstract nonsense.
Coding AI doesn't "recreate existing code that is working with something a bit different".
If I have, say, a website or an app, and I tell the AI that on this and this page/screen there's a green button, and I want that to turn blue, it goes in, finds the button (either via code analysis or via an emulator/browser + visual identification), finds its style, and changes the colour.
If I tell it to figure out what the very same button does, it does the same flow, except it goes further by tracing that button's click event, and then all the logic that is bound to it, even through multiple layers (e.g. in a well designed app that click event is handled in the ViewModel of the specific screen, which in turn calls a Use case, which in turn incorporates business logic and calls Repositories, which in turn call the data storage or networking layers, or maybe even both...), then explains what happens.
But let me give you a real world example. Just the other day I was working on a tool, a web flasher for a handful of different ESL (Electronic Shelf Label, e-ink based price tags). It involves connecting to a USB probe, plugging the probe into the board via spring contacts, and then running specific code to install custom software on the device so it talks to my smarthome. Most of the work was lift and shift - the flashing logic existed as C/Python code, so it was quickly ported to JavaScript, utilising WebUSB.
But it was ugly. The only thing I told the damn thing was to "make it look good". You know what the AI did?
It went onto the official site of the project, grabbed the fonts, colours, sizes, etc., and put that style on the flasher page. The only "went out of its way" thing it added was a "virtual e-ink display" emulating a tag, that was actually on the project site (https://openepaperlink.de/). It didn't go around changing the flashing code, the device detection, etc., it just covered the changes I requested - which were pretty broad.
What AI does depends 100% on the prompt. Image editing is a bit finicky because most image editing models come with their own LLM beyond the visual engine itself, and that LLM tends to be super limited, without any thinking loop, so it does what it first recognises (and often doesn't recognise negation - e.g. telling it to NOT change something, will result in specifically that change, because, again, the LLM included is tiny and crappy).
But you can't compare a "hey I asked AI to do X to my image and it did twelve other things" to working with codegen AI - especially the reinforced loop models like Claude or Codex that literally goes on to work out any issues by expanding on its own task while the goal remains what you set it to. So if you tell it to design a fancy website for you, you get a fancy website - but e.g. it won't start working on creating an investment portfolio or do sports bets for you.
My own experiences with various projects aren't abstract nonsense. You said yourself the prompt matters, and a broad prompting over time on the same project does drift. You haven't seen it, then good. I can't speculate on why. I''m not even saying it always happened, only that because of how these things work, probability does come into play.
The last is a strawman, I was using image generation as an example of visual drift that's more obvious that code. Nor did I imply drift as in a totally different direction - as I stated, the new routines would still work sometimes, but would differ in their approaches or use different variable names. Not all drift is critical or breaks things, but it's there. It's not an abstract.