The Long Tail of LLM-Assisted Decompilation

AI races through N64 game code, then hits a boss fight — the comments go wild

TLDR: An AI blew through much of a Nintendo 64 game’s code and, with “find-similar” tricks, reached ~75% before stalling. Commenters celebrate a path to revive lost classics, while skeptics question correctness, note big-function limits, and push for pairing old-school tools with AI—fueling a lively preservation debate.

The dev behind an AI-powered attempt to decode a Nintendo 64 game says the bot sprinted from about 25% to 58% matching code, then crawled until a new trick—prioritizing similar functions—pushed it to ~75%. After that? A stall. And the internet showed up with popcorn and hot takes.

Hype squad first: fans are giddy, dreaming of resurrecting lost classics. One commenter cheered that this could help games “with lost source,” name-dropping Red Alert 2. Meanwhile, the skeptics brought receipts. One user pounced on the line that Claude quits on jumbo-sized functions, dubbing it the final boss of retro code. Another asked the big question: can this technique guarantee correctness, or is it just confident guessing dressed as C?

Then came the pragmatists: why not tag-team with traditional decompilers and let AI clean up the messy parts? The author’s own tweaks—finding lookalike code via similarity scores (think: code doppelgängers), plus tools like a graphics microcode helper and a code-permuter—got applause for being clever, not magic.

The community vibe: AI speedruns old games until it hits the boss, preservationists are ecstatic, engineers demand proof, and everyone agrees this is high-drama retro archaeology. One fan dubbed it “AI that helps and harms no one”—and yes, the memes wrote themselves.

Key Points

  • Initial one-shot LLM decompilation rapidly increased matched code from ~25% to 58% for a Nintendo 64 game.
  • Progress slowed, prompting a workflow shift that ultimately reached ~75% matched code before stalling.
  • Early prioritization by estimated difficulty (via logistic regression) worked until only hard functions remained.
  • Prioritizing unmatched functions with similar matched counterparts proved highly effective for decompilation.
  • Exact similarity methods, including the Coddog tool using Levenshtein distance, complemented a handcrafted metric; specialized tools like gfxdis.f3dex2 and decomp-permuter further aided results.

Hottest takes

"one of the best use cases for AI today… no one’s job is being taken" — nemo1618
"Claude struggles with large functions… gives up… on those exceeding 1,000 instructions" — decidu0us9034
"Does this technique limit the LLM to correctness-preserving transforms?" — amelius
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.