Claims that genAI speeds up PC ports of classic games are either "false" or missing the point, says Banjo Kazooie: Recompiled developer

"Getting the game running is basically the first step of the journey."

The software developer behind an unofficial PC version of N64 platformer Banjo-Kazooie has posted an impassioned screed about the usage of generative AI tools for rush-job PC ports of classic games.

The dev in question is DarioSamo, a member of the modder group who released Banjo: Recompiled earlier this year. He argues that the idea that generative AI allows for faster porting is either outright "false" or simply misguided, explaining that decompiling a game – that is turning the original machine code into human-readable code – in order to release on modern platforms is supposed to be an exercise in actually understanding and preserving how that game was made.

Recently, Youtube channel Video Game Esoterica published a video about a Banjo Tooie "recompilation" project that makes use of generative AI. According to DarioSamo (via VGC), the new genAI recomp is based on "tooling" devised for Banjo: Recompiled, but is not the work of the Recompiled team. He feels VGE haven't made this clear in their video, calling it "irresponsible" and an incentive for generative AI users to "keep pumping out these projects just so they get their 10 minutes of fame in your videos".

In a follow-up thread on Bluesky, posted last night, DarioSamo went into detail about the impact of people using generative AI to knock out PC versions of older games. He began by explaining what a "recompilation" is – a type of port that uses "static recompilation" to automatically convert a game's code "into code that behaves the same to another platform." These recompilations sometimes call upon full or incomplete decompilation projects "as sources of information to make the necessary changes for a port to behave properly on the target platform," DarioSamo continued, "including enhancements necessary for high refresh rates, widescreen, QoL and more."

Recently, people have been using generative AI to automate things, but DarioSamo argues in his thread that "any claims you've seen about AI speeding up this 'tedious' process are false when recompilation is involved." He continued that devs were already quite capable of putting a recompilation together in less than a day, back when the genAI tools came into use.

"The pattern we've been seeing for several months so far is that a new type of person has entered the scene that thinks that because AI gets them over the hurdle of getting the game running, the job is basically done," DarioSamo wrote. "And they need nothing but a credit card and a subscription to an AI service. What these people are unaware of is that getting the game running is basically the first step of the journey and not the end.

"This is why I argue that as far as recomps go, AI hasn't provided any sort of speedup," he went on. "Making a barebones recompilation was always easy and required very little time. It's the polish and making a good and sustainable project for the wider public where the difference is. It's the part that actually takes far more time."

To illustrate this, DarioSamo pointed to DK64 Rekongpiled - a port of Donkey Kong 64 that doesn't make use of generative AI. "[The Rekongpiled team] were racing against another project that did use AI," he wrote. "Because the developers were deeply familiar with DK64, they were able to get the game to a very good state, further sped up by myself giving them information on how to use the framework to its full potential. Meanwhile, the AI effort failed to make further progress as it was lacking that knowledge." (DarioSamo didn't specify which genAI Donkey Kong port he was referring to.)

Generative AI can allow for faster decompilation, DarioSamo concedes in his thread, but only if all you care about is ending up with code that compiles into something matching the original game's ROM file. "However, the entire point of decomps has always been documenting the code and understanding how it works," he wrote. "This is why decompilation and porting efforts are usually done by different groups of people: one group works on understanding and documenting the game's code, while a different group can take that documented codebase to aid in their porting efforts."

Large Language Models, by contrast, "will often bruteforce this process and end up with a result that, while it does compile to the correct ROM, it is not readable or understandable by a human or even the LLM itself," DarioSamo continued. This makes it harder to create both a port based on the decompiled code, or patches for the port once released.

Using these tools also means shortcutting the process of working on the code in tandem with other port makers and players, DarioSamo wrote, "which means the new people entering the scene are also deeply unaware of a lot of small but critical decisions early in development, which leads to the tools making an unsalvageable project that needs to be restarted from scratch."

The rush of generative AI ports "can demotivate a developer who was in the middle of creating a project that checks all the boxes of the quality standards that are expected by the public," DarioSamo said in closing. "It is important to show your support to these developers, and allowing promotion of AI-generated projects in our spaces would just discourage them further. Assuring them that there's an audience that cares about the details and things being done properly is very important.

"These are old and beloved games that have been available in numerous ways for decades, so there's no rush to be the first to provide a solution using these tools, and it's unlikely it'll be the only solution in the future either," he said.

The part of this that jumps out for me is the point about the missed opportunity to grasp and preserve how a game was made, in a creative sector that is already plagued by cyclical brain drains and the loss of crucial documentation. It's deliberate, anti-social obliviousness for the sake of a (notional) fast return, which I think is very much the genAI philosophy at large.

The discussion around generative AI usage for recompilation echoes similar anger in modding communities. Recently, the creator of a Total War: Warhammer 3 mod "deliberately corrupted" their creation in order to stop people using it for "AI-related purposes". Back in August, Mark interviewed NexusMod users about what they consider an acceptable level of genAI usage, and how platforms should enforce transparency.