Palmer Luckey, the Oculus co-founder now running defence company Anduril and retro-hardware maker ModRetro, sees an unusual beneficiary in the latest surge of AI-assisted reverse engineering: the VR hobbyist. “The AI-decompilation boom will be such a boon for VR homebrew, almost unimaginably so,” Luckey wrote in a post on X on Sunday. It is a large claim, but it rests on a visible change in a once painfully slow craft.

Matching decompilation — reconstructing source code that compiles into the same machine code as an original game — has traditionally demanded years of collective, specialist work. Large language models have not abolished that work. They have made some parts of it radically faster, precisely because the machine supplies an unforgiving answer key.

For VR, the promise is not merely more old games on new screens. It is the prospect of turning preserved software into places: a classic racing game rebuilt around a cockpit, a museum room that lets visitors handle a scanned cartridge before entering its world, or a fan-made immersive adaptation that would once have required a small studio.

A machine-checkable problem The recent completion of the Super Smash Bros. Melee decompilation illustrates the shift. The project began in July 2020 and reached its 100% matching goal in September, after more than six years. Contributors said language models accelerated the final stretch. That milestone does not mean an AI simply recovered Nintendo’s original source code. It means a reconstructed C codebase can be compiled to produce a byte-for-byte match for the original game binary — a stringent technical test that gives humans and models a clear way to reject wrong answers. A separate Fable II project used static recompilation tools and a local Qwen model as a test of what AI could do. Its author said the game could be completed, but also reported texture problems and did not consider the port finished. That is the more useful measure of the moment: progress can arrive in days or weeks, while polish, compatibility and maintenance remain stubbornly human jobs.

“The AI-decompilation boom will be such a boon for VR homebrew, almost unimaginably so,” Luckey wrote on X. Luckey has spent years imagining a bridge between preservation and immersion. In a 2024 interview, he described a virtual “living museum” of retro games built from high-resolution scans of boxes, manuals, cartridges and their variants, with AI and human curation helping reconstruct settings such as old E3 show floors.

His ModRetro business, meanwhile, sells modern FPGA-based hardware for older formats, including its M64 Nintendo 64-compatible console. AI-assisted decompilation could supply a missing layer for that museum vision: software malleable enough to study, adapt and re-present rather than merely emulate behind glass.

VR’s opportunity is also its bottleneck

The new tools could lower the barrier for enthusiasts who want to make VR conversions, accessibility mods or historically minded interactive exhibits. Native or recompiled software can offer developers more scope to alter interfaces and rendering than a conventional emulation setup, though each game presents its own technical constraints. But headset ecosystems are not blank canvases.

Meta requires apps submitted to the Horizon Store to undergo review and comply with platform policies; developers are responsible for securing the rights and permissions needed for distribution. Its rules also reserve the company’s power to remove apps that violate its terms. That leaves a sharp division between the technical possibility of a homebrew project and a lawful, durable route to an audience.

In the United States, copyright law provides a narrow interoperability exception for some reverse engineering by someone lawfully entitled to use a program. EU software rules contain a similarly limited decompilation provision. Neither is a general licence to copy games, distribute their assets or defeat platform safeguards.

The legal distinction matters especially in VR. A fan can imagine a virtual archive; a rights holder may see a derivative commercial product. And firmware work on closed devices brings separate contractual and anti-circumvention questions that cannot be settled by calling a project preservation.

The flood has already begun

The scene is also arguing about whether speed has become a vice. Established recompilation developers have publicly disowned AI-heavy projects that borrow the language of their work, saying rushed releases are buggy, poorly tested and unlikely to be maintained. A catalogue launched this week to identify human-led decompilation and recompilation projects was an unusually direct response to the influx.

Their objection is not that automation has no place in reverse engineering. Matching work has always depended on tools, tests and collaboration. The concern is that a technically plausible demo can now be mistaken for a reliable port, while the careful work of documenting systems, correcting edge cases and supporting users is left behind.

Luckey’s prediction is therefore best read as a forecast of capacity, not quality.