Pocketpair Publishing boss John Buckley says we're already starting to see a flood of 'really low-quality, AI-made games' on Steam and other storefronts.
Neural networks are deterministic. In LLMs, it outputs probabilities, which are picked from via seeded RNG. Image generation tries multiple options based on different seeds, then picks the best fit as identified by a neural network and repeats. For both, if you give a specific model the same inputs, you’ll get the same output.
The public-facing interfaces don’t give seed control, which means they give a different output each time, but that isn’t an inherent property of generative AI.
I think it can - procedural generation consist of procedures, that is elements designed by humans, which are just connected into a bigger structure. Every single template, rule and atomic object (e.g. a single room in a generated house) is hand-designed, and as such no matter what comes out the elements and connections were considered by a real human. On the other hand, generative AI is almost always some sort of machine learning, that is an approximation of what a good structure of something should be, but it is only a very poor, randomised approximation. You have absolutely no guarantees nor constraints on what might pop out of the model - that is my main concern with genAI, though the whole outputted thing looks reasonable, upon closer inspection it has a lot of inconsistenties.
I think you are reading in the “designed by humans” part. Even when that is nominally true, the whole point of procedural generation is to create a level of complexity and emergence that the outputs are surprising and novel. Things no one expected are desirable. I think the distinction being drawn is not meaningful; in both cases, it is entirely possible and likely that no human being understands how a given output was arrived at.
Nonsense. Procedural generation is a rule-based deterministic system while generative AI is probabilistic and data driven. It’s fundamentally different.
Markov chains are both probabilistic and data-driven. For example. LLMs are not that far removed from markov chains. Should game developers be allowed to use latent spaces or is that too sloppy AI?
Okay, but (ignoring that procedural generation can also be probabilistic) what is the functional difference? The point I’m getting at is that you cannot banish the one without necessarily limiting the other.
The difference between “generative AI” and “procedural generation” cannot be meaningfully nailed down.
Procedural generation is theoretically deterministic, but it’s a fairly minor distinction.
Generative AI is too. Maintain your seed and you should get the same result every time.
Most of the SaaS AI tools don’t expose control over their RNG, but some self-hosted ones do.
Generative AI is by definition non deterministic.
Neural networks are deterministic. In LLMs, it outputs probabilities, which are picked from via seeded RNG. Image generation tries multiple options based on different seeds, then picks the best fit as identified by a neural network and repeats. For both, if you give a specific model the same inputs, you’ll get the same output.
The public-facing interfaces don’t give seed control, which means they give a different output each time, but that isn’t an inherent property of generative AI.
I think it can - procedural generation consist of procedures, that is elements designed by humans, which are just connected into a bigger structure. Every single template, rule and atomic object (e.g. a single room in a generated house) is hand-designed, and as such no matter what comes out the elements and connections were considered by a real human. On the other hand, generative AI is almost always some sort of machine learning, that is an approximation of what a good structure of something should be, but it is only a very poor, randomised approximation. You have absolutely no guarantees nor constraints on what might pop out of the model - that is my main concern with genAI, though the whole outputted thing looks reasonable, upon closer inspection it has a lot of inconsistenties.
I think you are reading in the “designed by humans” part. Even when that is nominally true, the whole point of procedural generation is to create a level of complexity and emergence that the outputs are surprising and novel. Things no one expected are desirable. I think the distinction being drawn is not meaningful; in both cases, it is entirely possible and likely that no human being understands how a given output was arrived at.
Nonsense. Procedural generation is a rule-based deterministic system while generative AI is probabilistic and data driven. It’s fundamentally different.
Markov chains are both probabilistic and data-driven. For example. LLMs are not that far removed from markov chains. Should game developers be allowed to use latent spaces or is that too sloppy AI?
Okay, but (ignoring that procedural generation can also be probabilistic) what is the functional difference? The point I’m getting at is that you cannot banish the one without necessarily limiting the other.
It’s less of a functional different and more of a moral one.
Content theft is a separate issue. We can agree to ban the fruits of content theft without drawing arbitrary technical taboos.
You don’t need any preexisting training data for procedural generation
Where are all these prompt based image generators that identify themselves as procedural generation?