Crash Team Racing, a PS1 racer that I'd argue is better than Mario Kart, finally just arrived on PC thanks to a decomp. If you've been waiting for the chaotic Naughty Dog spin-off to break free from its retro console prison, you'll probably already be making weekend plans, but this is sadly yet another project made using AI.
Dubbed Crash Team Racing: Turbocharged, the shiny new port is actually based on the existing CTR native project. The latter is the product of human coders who've been decompiling the classic kart racer since 2021-2022, whereas the new source-port uses the decomp as a base while enabling native-resolution rendering "up to 4K and beyond", alongside widescreen support, enhanced textures, and frame rates up to 240fps.
Being able to play Crash Team Racing at 4K on a PC sounds like a dream, and the fact that it's free is admittedly an irresistible cherry on top. I won't be chowing down on this specific high-spec PS1 decomp Sundae, though, as developer Camzie99 has confirmed in a lengthy YouTube comment that AI, specifically Large Language Models (LLMs) were used to create the port.
In response to a fan exclaiming "tell me this one's not 'vibe-coded' god i need this" in a comment, Camzie99 decides to challenge the definition by replying "it depends on your definition of vibe coded I suppose," before elaborating on how Crash Team Racing: Turbocharged was created.
Before we dive into that, it's worth highlighting that "vibe-coding" typically describes a development method that feeds plain English into an AI to create websites, apps, and programs rather than coding line by line. In the case of Crash Team Racing: Turbocharged, the project's creator (who apparently has been developing for 16 years), says they used a LLM to write code in "smaller chunks" rather than generating the whole thing with one prompt.
"I did use LLMs to write code, however it wasn't with vague prompts such as "Build a CTR recomp, make no mistake." elaborates Camzie99. "Instead my process is much more fine-grained and iterative. For any particular feature, I'd first plan the feature, find locations in the code that are likely to need changing, and decide at a high level what needs to change at each point. Once that was done, I would then get an LLM to actually write the code, but it would again be in smaller chunks, function by function, or limited to a single file for example so it was easy to review and ensure it was as expected."
Camzie99 goes on to include an example of how they used LLMs for implementing PGXP, an enhancement used by emulators to tackle PS1 polygonal wobbling. The argument is that this isn't a "low-effort flip" since the developer carried out all of the "decision making, planning and thought processing," but the code was ultimately generated using AI rather than written by hand.
هەفتانە هەرسکردنەکان، چیرۆکەکانی ئەو کۆمەڵگایانەی کە خۆشت دەوێن و زۆر شتی تر
سەرچاوەی فەرمی بە زمانی ئینگلیزی: https://www.gamesradar.com/hardware/retro/the-og-crash-team-racing-for-ps1-just-arrived-on-pc-but-its-sadly-yet-another-recomp-spoiled-by-ai
