The post is a project writeup on decompiling Snowboard Kids, a late-1990s Nintendo 64 game, into source code that can be rebuilt and studied. The notable twist is heavy use of LLM agents, plus workflow tricks like explicit deadlines for agent tasks, to compress a job that usually takes years into 84 days. In this world, “decompilation” does not mean making a rough readable clone. It usually means producing new source code that compiles back to the original binary, or close enough to validate that the behavior has been recovered exactly. That matters because these projects are often the bridge to PC ports, mods, bug fixes, and long-term preservation of games whose original source code is lost or inaccessible.
The strongest reaction was excitement that AI is changing the economics of reverse engineering. People working on other game decomps said the hard parts now look tractable, especially naming, labeling, build iteration, and chewing through thousands of functions. Several comments treated this as one of the few clearly useful AI applications with low tolerance for hand-wavy output because a matching binary gives you an objective correctness check. At the same time, the thread did not buy “AI all the way down” as a free pass. The practical consensus was that LLMs are useful for recovery work, but taste, testing, and cleanup still decide whether a decomp turns into a usable modern port instead of unreadable slop.
The other recurring theme was law, and nobody pretended it is settled. Clean-room reimplementation and matching decompilation were treated as very different things. A clean-room project tries to avoid exposure to copyrighted code, while these decomp efforts work directly from the shipped binary and intentionally recreate it. That makes the legal footing much shakier even if the source is newly written, distributed for free, and often requires users to provide their own
ROM or assets. Comments also pointed out why publishers often do not do the obvious thing and ship upgraded retro releases themselves. Rights for code, art, music, voice, motion capture, and platform-specific distribution can be fragmented or lost, especially after studio closures and acquisitions. Put bluntly, the technology is moving faster than the legal model, which is why fan-led preservation and porting keeps outrunning official rereleases.
blog.chrislewis.au