This morning, the news broke that Larian Studios, developer of Baldur's Gate 3 and the upcoming, just-announced Divinity, is apparently using generative AI behind the scenes. The backlash has been swift, and now Larian founder and game director Swen Vincke is responding to clarify his remarks.
You know it doesn’t have to be all or nothing, right?
Part of the “magic” of AI is how much of the design process gets hijacked by inference. At some scale you simply don’t have control of your own product anymore. What is normally a process of building up an asset by layers becomes flattened blobs you need to meticulously deconstruct and reconstruct if you want them to not look like total shit.
That’s a big part of the reason why “AI slop” looks so bad. Inference is fundamentally not how people create complex and delicate art pieces. It’s like constructing a house by starting with the paint job and ending with the framing lumber, then asking an architect to fix where you fucked up.
If you don’t like them, you can just chuck them in the trash and you won’t have wasted the work of an artist
If you engineer your art department to start with verbal prompts rather than sketches and rough drawings, you’re handcuffing yourself to the heuristics of your AI dataset. It doesn’t matter that you can throw away what you don’t like. It matters that you’re preemptively limiting yourself to what you’ll eventually approve.
That’s a big part of the reason why “AI slop” looks so bad. Inference is fundamentally not how people create complex and delicate art pieces. It’s like constructing a house by starting with the paint job and ending with the framing lumber, then asking an architect to fix where you fucked up.
This is just the whole robot sandwich thing to me.
A tool is a tool. Fools may not use them well, but someone who understands how to properly use a tool can get great things out of it.
Doesn’t anybody remember how internet search was in the early days? How you had to craft very specific searches to get something you actually wanted? To me this is like that. I use generative AI as a search engine and just like with altavista or google, it’s up to my own evaluation of the results and my own acumen with the prompt to get me where I want to be. Even then, I still need to pay attention and make sure what I have is relevant and useful.
I think artists could use gen AI to make more good art than ever, but just like a photographer… a thousand shots only results in a very small number of truly amazing outcomes.
Gen AI can’t think for itself or for anybody, and if you let it do the thinking and end up with slop well… garbage in, garbage out.
At the end of the day right now two people can use the same tools and ask for the same things and get wildly different outputs. It doesn’t have to be garbage unless you let it be though.
I will say, gen AI seems to be the only way to combat the insane BEC attacks we have today. I can’t babysit every single user’s every email, but it sure as hell can bring me a shortlist of things to look at. Something might get through, but before I had a tool a ton of shit got through, and we almost paid tens of thousands of dollars in a single bogus but convincing looking invoice. It went so far as a fucking bank account penny test (they verified two ach deposits) Four different people gave their approvals - head of accounting included… before a junior person asked us if we saw anything fishy. This is just one example for why gen AI can have real practical use cases.
This is just the whole robot sandwich thing to me.
If home kitchens were being replaced by pre-filled Automats, I’d be equally repulsed.
A tool is a tool. Fools may not use them well, but someone who understands how to properly use a tool can get great things out of it.
The most expert craftsman won’t get a round peg to fit into a square hole without doing some damage. At some point, you need to understand what the tool is useful for. And the danger of LLMs boils down to the seeming industrial scale willingness to sacrifice quality for expediency and defend the choice in the name of business profit.
Doesn’t anybody remember how internet search was in the early days? How you had to craft very specific searches to get something you actually wanted?
Internet search was as much constrained by what was online as what you entered in the prompt. You might ask for a horse and get a hundred different Palominos when you wanted a Clydesdale, not realizing the need to be specific. But you’re never going to find a picture of a Vermont Morgan horse if nobody bothered to snap a photo and host it where a crawler could find it.
Taken to the next level with LLMs, you’re never going to infer a Vermont Morgan if it isn’t in the training data. You’re never going to even think to look for one, if the LLM hasn’t bothered to index it properly. And because these AI engines are constantly eating their own tails, what you get is a basket of horses that are inferred between a Palomino and a Clydesdale, sucked back into training data, and inferred in between a Palomino and a Palomino-Clydesdale, and sucked back into the training data, and, and, and…
I think artists could use gen AI to make more good art than ever
I don’t think using an increasingly elaborate and sophisticated crutch will teach you to sprint faster than Hussein Bolt. Removing steps in the artistic process and relying on glorified Clipart Catalogs will not improve your output. It will speed up your output and meet some minimum viable standard for release. But the goal of that process is to remove human involvement, not improve human involvement.
I will say, gen AI seems to be the only way to combat the insane BEC attacks we have today.
Which is great. Love to use algorithmic defenses to combat algorithmic attacks.
But that’s a completely different problem than using inference to generate art assets.
Part of the “magic” of AI is how much of the design process gets hijacked by inference. At some scale you simply don’t have control of your own product anymore. What is normally a process of building up an asset by layers becomes flattened blobs you need to meticulously deconstruct and reconstruct if you want them to not look like total shit.
That’s a big part of the reason why “AI slop” looks so bad. Inference is fundamentally not how people create complex and delicate art pieces. It’s like constructing a house by starting with the paint job and ending with the framing lumber, then asking an architect to fix where you fucked up.
If you engineer your art department to start with verbal prompts rather than sketches and rough drawings, you’re handcuffing yourself to the heuristics of your AI dataset. It doesn’t matter that you can throw away what you don’t like. It matters that you’re preemptively limiting yourself to what you’ll eventually approve.
How do you think a human decides what to sketch? They talk about the requirements.
This is just the whole robot sandwich thing to me.
A tool is a tool. Fools may not use them well, but someone who understands how to properly use a tool can get great things out of it.
Doesn’t anybody remember how internet search was in the early days? How you had to craft very specific searches to get something you actually wanted? To me this is like that. I use generative AI as a search engine and just like with altavista or google, it’s up to my own evaluation of the results and my own acumen with the prompt to get me where I want to be. Even then, I still need to pay attention and make sure what I have is relevant and useful.
I think artists could use gen AI to make more good art than ever, but just like a photographer… a thousand shots only results in a very small number of truly amazing outcomes.
Gen AI can’t think for itself or for anybody, and if you let it do the thinking and end up with slop well… garbage in, garbage out.
At the end of the day right now two people can use the same tools and ask for the same things and get wildly different outputs. It doesn’t have to be garbage unless you let it be though.
I will say, gen AI seems to be the only way to combat the insane BEC attacks we have today. I can’t babysit every single user’s every email, but it sure as hell can bring me a shortlist of things to look at. Something might get through, but before I had a tool a ton of shit got through, and we almost paid tens of thousands of dollars in a single bogus but convincing looking invoice. It went so far as a fucking bank account penny test (they verified two ach deposits) Four different people gave their approvals - head of accounting included… before a junior person asked us if we saw anything fishy. This is just one example for why gen AI can have real practical use cases.
If home kitchens were being replaced by pre-filled Automats, I’d be equally repulsed.
The most expert craftsman won’t get a round peg to fit into a square hole without doing some damage. At some point, you need to understand what the tool is useful for. And the danger of LLMs boils down to the seeming industrial scale willingness to sacrifice quality for expediency and defend the choice in the name of business profit.
Internet search was as much constrained by what was online as what you entered in the prompt. You might ask for a horse and get a hundred different Palominos when you wanted a Clydesdale, not realizing the need to be specific. But you’re never going to find a picture of a Vermont Morgan horse if nobody bothered to snap a photo and host it where a crawler could find it.
Taken to the next level with LLMs, you’re never going to infer a Vermont Morgan if it isn’t in the training data. You’re never going to even think to look for one, if the LLM hasn’t bothered to index it properly. And because these AI engines are constantly eating their own tails, what you get is a basket of horses that are inferred between a Palomino and a Clydesdale, sucked back into training data, and inferred in between a Palomino and a Palomino-Clydesdale, and sucked back into the training data, and, and, and…
I don’t think using an increasingly elaborate and sophisticated crutch will teach you to sprint faster than Hussein Bolt. Removing steps in the artistic process and relying on glorified Clipart Catalogs will not improve your output. It will speed up your output and meet some minimum viable standard for release. But the goal of that process is to remove human involvement, not improve human involvement.
Which is great. Love to use algorithmic defenses to combat algorithmic attacks.
But that’s a completely different problem than using inference to generate art assets.