AI made decompiling old games easy, so decomp.dev shut its leaderboard

Rebuilding the source code of an old video game by hand used to take teams of volunteers years of work. AI models now do much of it, and decomp.dev, the site that tracked hundreds of those fan projects and ranked their progress, has switched off its public index and stopped accepting new projects. Its administrator says the tools work. What broke is the community.

Matching decompilation is the hobby’s name for a strict kind of reverse engineering: writing C code that, once compiled, produces exactly the same machine code as the original game, byte for byte. That rebuilt code is what lets fans run a Nintendo Mario classic natively on a PC, hunt glitches for speedruns, mod games, port them to new hardware or recover code that may otherwise be lost.

According to the announcement, written by the site’s pseudonymous admin, Encounter, recent frontier models such as GPT-6-Astra crossed a line. They follow complex directions, cross-reference symbols and debug information, and write code that reads naturally. They are also “exceptionally good permuters”, writing small scripts that test many versions of a function until one compiles to an exact match, and they stay focused across long tasks for less money than earlier models.

The damage is social. Projects listed on decomp.dev were open source to invite contributors, and Encounter describes “hostile forks” that take human-led projects and “finish” them with AI, plus opportunists who race to release half-working PC ports built on decompilations they never helped write. The progress bars, designed to keep people motivated when every fraction of a percent took real effort, became what Encounter calls “a hollow numbers game” for people who “burn tokens to make number go up.”

The shutdown is narrower than it sounds. decomp.dev keeps serving existing project pages and historical data; only the main index and new submissions are gone. Encounter also refuses to blame AI as such, writing that it “can help tremendously when used correctly”, and chose not to ban AI-assisted projects because one hobbyist cannot define and police “slop” across hundreds of them. The legal question stays open too: decompiled code still comes from copyrighted games, and faster tools widen that grey zone without settling it.

The pressure reaches beyond games. As 404 Media reported, AI-built clones of paid software such as Adobe’s Creative Cloud apps and Microsoft Office have circulated in recent days, and small game developers say AI “clean room” copies of their games keep appearing. One tool aimed squarely at this hobby, decomp.wtf, promised to point a user’s AI agent at a game they own and turn it back into C one function at a time. Its creator has since discontinued it, writing that much of the decompilation community opposes AI decompiling of that kind. The code stays on GitHub under the MIT license.

decomp.dev launched in 2024 to centralize progress tracking. Encounter dates the turn to early 2026, when Opus 4.6 and GPT-5.4 appeared to know the community’s own tools without being told, though their results still needed heavy human cleanup and cost too much per task. The site announced the change on 15 October. Its own code, which Encounter says was written entirely by humans, has never carried an open-source license, so nobody may reuse it to rebuild the leaderboard somewhere else. The games have never been easier to open up; the people who used to do it together just lost their meeting place.

Tags: AI coding, decomp.dev, decompilation, GPT-6-Astra, Nintendo 64, reverse engineering