Peter “Durante” Thoman has shared a detailed, first-hand account of using large language models (LLMs) in game development. His take is that AI can be useful, but it brings plenty of extra work and some surprising costs.
Thoman became famous in 2012 with DSfix. The unofficial mod fixed the PC release of Dark Souls: Prepare to Die Edition, which was stuck at a low internal rendering resolution. It also added extra graphics and configuration options.
He later co-founded PH3, a studio that specializes in PC ports and technical optimization. Its work includes the PC versions of Falcom’s The Legend of Heroes: Trails of Cold Steel III and Ys IX: Monstrum Nox.
Porting is much more than getting a game to run on another platform. It means adapting graphics systems, improving performance, supporting different resolutions and controllers, and testing everything without breaking gameplay. In other words, Thoman knows the technical side of games very well.
The comments are the hard part
Thoman’s most concrete example isn’t about code at all. It’s about the comments inside generated code.
He says some coworkers share his frustrations with LLM output. Getting the AI to write acceptable code comments turned into a project of its own.
“We probably spent as much effort on tooling and instructions to just make those non-terrible as we spent on the things most people would initially deem ‘important,'” he wrote. “And I’d argue this is still less successful than the coding part.”
He describes the generated code itself as having “validated functionality.” The written explanations around it were the weaker part. He also made it clear that he dislikes typical LLM prose.
This shows that generating code is only one step. A team still has to read it, understand it, and maintain it.
Not all AI use is the same
Thoman’s main argument is that people should judge each use of AI on its own merits.
He contrasted two scenarios. In one, an indie developer uses generated material to add variety to unimportant background assets. In the other, a hugely profitable publisher replaces most of its engine team or brags about “vibecoding.”
He added that some generated photorealistic textures would be hard to spot unless developers disclose them. He also stressed that much of his post explains how he thinks things work, rather than approving of every result.
Modding came up too. Thoman admits AI-assisted mods can be both bad and good. His worry about “vibe-modding” is that it lets individuals make their own mods easily, which could reduce the reason to team up on shared community projects.
