Post

The reason why procgen in games is, as we like to say, a great way of getting twice the content with four times the work is that you can't fight Shannon's laws any more than you can fight the laws of thermodynamics. You can't give the audience more information than you were willing to put in.

Sometimes this means manually creating a lot of material that is going to be recombined, sometimes this means spending an unreasonable amount of time defining the exact parameters and structure of a grammar or a generator.

2 reposts89 likes

It's a magic trick of the same kind as the ones where something large comes out of a small box; it's done by making something that *unfolds* itself. "Generation" is almost a misnomer, really, it's *decompression.*

Any aspect of the system that isn't downstream from those choices is, essentially, noise. A little bit of noise is texture, a lot of noise makes the semantic content of the piece collapse out.

This is the trap that LLMs and other generative AI systems present in this context; if you're generating vast amounts of content from a tiny prompt, what you're doing is supplying your audience with almost nothing and filling in the empty space with the semantic equivalent of sawdust.

Ultimately I think the first few games with LLM chatbot NPCs are an interesting novelty, but there's no particular long-run future in this because every game that does this has functionally the same NPCs. it's all just a fungible commodity.

Honestly, I think it's even worse than it seems. These systems create outputs which would be described by the maximum number of people as the inputs. Part of the implication to that is anything not in the input *must* become un-memorable and unworthy of mention. Conceptually invisible to the mind.

Hmm, interesting. There's a depth to hand-crafted characters precisely because there are things they won't or can't do, and figuring out "what are the limits of this character's range", or "what are the limits of this dungeon" is only interesting if there *are* limits.