Procedural generation and generative AI are separate, distinct areas under the umbrella term of generative systems.
Regardless of opinions about the respective fields, I hope we can agree on using the right terms for the right things, to foster clear communication.
Feedback on the chart is welcome. Let me know if you find anything inaccurate or misleading, though keep in mind I had to leave out many details for brevity.
That's my take, too. My first criticism of the chart is that it accepts "Generative AI" as a proper use of "AI."
@Professor_Stevens @MrBehemo AI on its own is such a broad and fuzzy term that has been used about anything that gives even the slightest impression of autonomy. As such it's not very descriptive, as it can mean almost anything.
But the combination "generative AI" has a much more specific usage. It does not refer to a lot of unrelated things. So judged on its capability for clear communication, it works fine IMO.
"AI" is fuzzy because the modern popular press is lazy. Russell & Norvig is the most widely accepted text on AI. The definition there is, “the designing and building of intelligent agents that receive percepts from the environment and take actions that affect that environment.”
One can avoid the implications of a definition by recursively calling for more definitions of the words in a definition. There is, however, more to their book than I can excerpt here. Have you read it?
Lacking a monopoly isn't the same as being of no relevance. Your chart makes a (imo, crucial) distinction between procgen and what some people call "gen AI." I am with you all the way on that! But the chart isn't just about gaming.
You asked for feedback. Mine is that you have a great chance with this format to add that not everyone even accepts "gen AI" as AI.
Man, you read fast.
@runevision Could add that procgen often uses a seed to make the results reproducible whereas GenAI's statistical model usually has non deterministic output?
The ability to give the same input to a procgen system and get the same output is helpful to debugging the algorithm and have confidence in what you can expect from the system -- essential for use in games.
I guess GenAI tends to give different output for the same input because it uses the previous input as additional input? Not sure.
@runevision The problem is that people use the term "procedural" to denote anything that is not precomputed i.e. "generated on the fly".
So generative AI would then be a subset of procedural generation ...
This is a futile battle. I remember the time when anything above pushing pixels to the screen was considered "AI" in games - even playing keyframe animations.
Or terms like "mixed reality" (nope, nothing to do with Hololens), "XR" - or "hologram" used to describe Pepper's ghost.
@runevision That's probably because the current AI tools are simply not good enough (yet?) and difficult to use for lay persons for most stuff one would want to use as "procedural content".
Apart from talking to NPCs using some LLM which is the low hanging fruit, there isn't much else - e.g. level generation, geometric asset generation, etc. even though attempts have been published. So the "classic" techniques still win here.
I didn't mean generating images - that's apart from textures not something all that useful when working on a game.
When people think about "procedural content" it is usually 3D geometry that is being generated, whether individual objects or entire levels and worlds.
And while there has been some work on creating neural networks capable of generating these, including rigged meshes for animation, it is far from something practically usable so far.
Yes, true - but outside of that context I haven't really encountered anyone making this link between generative AI and generative systems.
OTOH, what do you even define as "procedure" and "rules"?
Would an inference runtime executing a trained network not count as a "procedure" and "algorithm" only because it runs something based on statistics? I think any computer scientist would vehemently disagree there.
Why to even make such a contrived distinction?
To me that sort of distinction is rather artificial, IMO.
Moreover, procedural generation doesn't necessarily mean it must be rule-based.
It would make more sense to put it as a distinction between generative neural networks and specifically rule-based (or grammar-based) systems (which can be also non-deterministic, btw) if that's what you want to explain.
Otherwise it is overly general and risks stretching the terminology way too far.
Rune Skovbo Johansen
Nicolai Hähnle
NickBell
Professor_Stevens
havchr
David
Jan Ciger