As a longtime procgen dev and enjoyer, it's been a weird experience to look at present-day GenAI talking points, because they were all settled long ago in the space of Proc Gen.
- Proc Gen is a cost saving measure. Would Notch, Redigit, Derek Yu, and a whole slew of last decade's indies been able to punch above their weight and challenge the AAA's without it?
- Proc Gen can be soulless. But only if your designer decision was to push a button to get back a perlin heightmap, or Binary Space Partition and call it a day. Bad (or just old) roguelikes used to do that. Now, there is so much that an actual designer would do on top of it.
- Proc Gen is unintentional, that's the point- to see new behaviors emerge that weren't programmed in and surprise us in unexpected ways.
We've even had more experimental indies play with text-generating markov chains that were trained on centuries-old public domain text. Or level-generating WaveFunctionCollapse trained off of internally developed level presets.
All of this was an open and shut case in the last decade, where the advantages were clearly recognized, the flaws were acknowledged constructively, and the quirks were taken by the crazier artists to go into their own wild directions. The discussion was healthy.
If anything, those similarities between last decade's proc gen and this decade's GenAI should be a guiding light for how this new technology is supposed to be applied.
- Don't use it to empower AAA's to cut their teams of 10 designers to 1. Use it to empower indies where one person can now generate a world it would've taken 10 designers a decade to make by hand.
- Don't settle for just pressing a button with a one sentence prompt and call it a day. Give more nuanced details, train a fine-tune/LoRA with work you intentionally selected, and weave the results together with your own handcrafted content to make something that shows your specific style.
- Don't look at the quirks and hallucinations and declare it unfit for your use case. Take what you find as a novelty, and build your use case around it. Maybe it's useful, maybe it's not- but it's that exploration that constitutes what art is itself.
The only thing that's changed between last decade's markov experiments and this decade's GPT was that people started training models off things that weren't theirs. The answer is simple: don't. Developers have done it before, and they continue to do so. They provide the correct example.
Decades ago, designing a dungeon layout with a couple of setpieces and enemies used to be considered creative. We broke down all the techniques that made "good layouts" entertaining, and packaged them to run into infinity, for anyone else to use. If it were to be used by someone else, that program doesn't reflect the user's intentionality- there's no understanding of the nuances that went into it. But that's the nature of middleware- they'll use it to make something of their own. The automation didn't destroy the art of creating. We, the humans, just got more creative.
