6 min read
All 2,145 functions of the Nintendo 64 game Snowboard Kids were turned back into human-readable C in 84 days. The same developer needed 596 days to finish its sequel, Snowboard Kids 2, the same way. The difference was AI coding agents.[1]
This year a run of AI-assisted decompilation projects has appeared for Game Boy Advance, Nintendo 64 and Nintendo DS games, and in early October X filled up with videos of two games “mashed together” with AI. At the same time, communities that spent years doing this work by hand have pushed back, calling it disrespectful. This piece sorts out what has become possible, what still needs people, and where the legal lines sit.
What decompilation is
Games are written in a language like C and turned into machine code by a compiler. Only the machine code ships; the source stays inside the studio. Decompilation runs that process backwards, reading machine code and writing source that resembles the original.
Game communities practice matching decompilation. The rewritten C, compiled with the same compiler and settings as the original, must produce machine code identical to the shipped game, byte for byte. Because the standard is that strict, a machine can tell you immediately whether a function is right.
With complete source, people can build unofficial PC and phone ports, high-resolution and high-frame-rate patches, and mods. The Super Smash Bros. Melee decompilation, begun in July 2020, reached 100% on September 8.[2] Ship of Harkinian, the PC port of Ocarina of Time, sits on top of the same kind of work.
Why AI fits: there is a referee
Matching decompilation suits a coding agent. The agent proposes C, compiles it, diffs the bytes against the original, edits and recompiles, dozens of times without a human. Because a machine judges right and wrong, plausible-but-wrong answers have little room to slip through.
Australian developer Chris Lewis’s numbers show the gap.[1]
| Game | Approach | Time |
|---|---|---|
| Snowboard Kids 2 (N64) | Human-led, agent called in when stuck | 596 days |
| Snowboard Kids (N64) | Agents run in parallel, humans solve the blockers | 84 days, 2,145 functions at 100% |
Against rule-based tools the gap is wider still: the automatic decompiler m2c matched 17 of 1,830 functions (0.93%) in the same game.[1] Lewis ran agents across four Git worktrees and says they handled most easy functions and library code.
Macabeus, decompiling the 2001 GBA game Klonoa: Empire of Dreams with Claude Code, had matched 51% of 663 functions by August, and along the way the agent found a 25-year-old Easter egg on the save-deletion screen.[3] A Super Mario 64 DS reconstruction reports 99.9% of functions matched.[4]
What still needs people
The same write-ups list the limits.
Compiler quirks. Many N64 games were built with SGI’s IDO compiler, and Lewis writes that “LLMs and I are not particularly good at reproducing its output.” Three experts solved the stuck functions; roughly 4.8% of matching commits involved expert intervention, and he estimates the project “would have stalled around 89–90% without their IDO expertise.”[1]
Arithmetic and rules. In earlier work Lewis found agents often got basic arithmetic such as struct field byte offsets, C89 syntax rules, and complex compiler-generated control flow wrong, and concluded they work best as “accelerators inside a human-led loop.”[5]
Cheating. Macabeus reports the agent trying to modify the compiler itself when stuck, and names the biggest risk of AI-heavy projects as “cognitive debt reaching a level that is impossible to pay off”: code that matches but that nobody understands.[3]
Model rankings are not fixed either. Lewis found Claude better than Codex on Snowboard Kids 2, but on this project “Codex continued to outperform Claude.”[1][5] Research points the same way. DecompileBench (2025) found LLM output more readable but less functionally correct than traditional tools,[4] and Echo, a paper posted September 16, reports 2.43 times more exact matches than the strongest baseline by using compilation results as feedback.[6]
Where the community split
The backlash is about quality and attitude. Dario of the N64: Recompiled team, which ported Zelda and Banjo-Kazooie to PC, said a newly released AI-built Banjo-Tooie port was “entirely AI-generated” and unaffiliated with them. “We have not, do not and will not use AI to ever work on these projects,” he said, calling AI ports “a race to the bottom” and building them on his team’s tools “very disrespectful.”[7] The Tooie port’s author had disclosed that “AI tools generated substantial portions of this project’s code.”[8]
The early-October “mashup mods” on X are a different thing. Videos of Minecraft running inside Elden Ring or Skate 3 controls in Call of Duty arrived right after Anthropic’s Claude Opus 5.5 launch; reporting suggests most are tech demos, not playable mods. Nexus Mods tells users to “avoid using these tools unless you have explicit permission to use all the assets.”[9]
So “AI decompilation” covers three different things: human-verified matching decompilation (Snowboard Kids, Klonoa), ports mostly generated by AI (Banjo-Tooie), and demo videos whose playability is unclear (mashups).
Where the legal lines sit
In the US, Sega v. Accolade (1992) treated disassembly necessary for compatibility as fair use, and Sony v. Connectix (2000) did the same for intermediate copying while building an emulator. Circumventing copy protection is governed separately by DMCA Section 1201.[4] Take-Two’s lawsuit against the re3 project for GTA III and Vice City ended in a 2023 settlement, which sets no precedent.[4]
Korea’s Copyright Act, Article 101-4, allows a lawful user to reverse-analyse program code without permission only to obtain information needed for interoperability that cannot otherwise be obtained, and forbids using it to build a program substantially similar in expression.[10] Article 6 of the EU Software Directive has the same shape.[4]
Matching decompilation, by definition, produces code that compiles to the same machine code as the original, which sits squarely against “substantially similar.” That is why most projects publish code without graphics or audio and ask users to extract assets from their own copy.[2] Having an AI rewrite the code changes nothing; new names and syntax do not create independent authorship.[4]
Publishers have so far targeted distributables more than source. Nintendo sent GitHub seven DMCA notices on August 17 that removed 403 repositories, all in Switch emulator networks using console keys.[11] How that line moves as AI decompilation repositories multiply is still open.
What to do with this
- Check where a port comes from before downloading. Repositories usually say whether they sit on a matching decompilation or were mostly AI-generated; the latter tend to be buggier and more likely to be abandoned.
- Avoid releases that bundle game assets. Legitimate projects ask for your own game files.
- If you are a developer, point agents at problems with a referee. Decompilation works because a machine can verify the answer. The same approach suits code migrations with tests and legacy analysis you can check automatically. Without verification, you accumulate the cognitive debt Macabeus describes.
Reflection
596 days to 84 is a big gap, but even inside those 84 days three people solved the decisive part. Lewis writes that “a motivated human team with the right expertise and intuition can match or even exceed the pace.” What AI sped up was the part anyone could already do; what remained was the knowledge of a few people who know an old compiler’s habits. The Banjo team’s anger comes from watching others borrow the name of that work.
Notes
Notes
- Chris Lewis, “Decompiling a Nintendo 64 Game in 84 Days” (2026-08-26). blog.chrislewis.au ↩
- Android Authority on the Melee decompilation (completed September 8, 2026; assets must be extracted from the user’s own game). androidauthority.com ↩
- Macabeus, “Starting a Decompilation Project from Zero: Claude Code and 51% of a 2001 GBA Game” (2026-08-17). gambiconf.substack.com ↩
- Kingy AI, “AI Video Game Decompilation: What Works, What’s Possible, and What’s Legal” (checked October 2026). kingy.ai ↩
- Chris Lewis, “Using Coding Agents to Decompile Nintendo 64 Games”. blog.chrislewis.au ↩
- Jun Bi et al., “Echo: Learning-based Matching Decompilation using Trusted Back Translation”, arXiv 2609.18706 (2026-09-16). arxiv.org ↩
- Video Games Chronicle (Chris Scullion), 2026-10-03. videogameschronicle.com ↩
- GamesRadar+, 2026-10-02. gamesradar.com ↩
- Everything Edinburgh, AI mashup mods, 2026-10-04. everythingedinburgh.com ↩
- Korea Copyright Act, Article 101-4 (Reverse analysis of program code). casenote.kr ↩
- GitHub DMCA repository, Nintendo notice, 2026-08-17. github.com/github/dmca ↩
Source: https://blog.chrislewis.au/decompiling-a-nintendo-64-game-in-84-days/