Last I heard, current AI will devour themselves if trained on content from other AI. It simply isn’t good enough to use, and the garbage noise to value ratio is too high to make it worth filtering through.
Yeah… that is right up there with “AI can’t do feet” in terms of being nonsense people spew.
There is nothing inherently different between a picture made by an LLM and a picture drawn by Rob Liefeld. Both have fucked up feet and both can be fixed with a bit of effort.
The issue is more the training data itself. Where this CAN cause a problem is if you unknowingly train on endless amounts of ai generated content. But… we have the same problem with training on endless amounts of human content. Very few training sets (these days) bother to put in the time to actually label what input is. So it isn’t “This is a good recipe, that is a bad recipe, and that is an ad for betterhelp”. It is “This is all the data we scraped off every public facing blog and youtube transcript”.
Its also why the major companies are putting a big emphasis on letting customers feed in their own data. Partially that is out of the understanding that people might not want to type corporate IP into a web interface. But it is also because it provides a way to rapidly generate some labeled data because you know that customer cares about widgets if they have twelve gigs of documents on widgets.
I see AI as basically a magic box containing an approximation of skill but lacking understanding and intent.
And what is the difference between someone getting paid to draw a picture of Sonic choking on a chili dog by a rando versus an AI generated image of the same?
At the end of the day, we aren’t going to see magic AIs generating everything with zero prompting (maybe in a decade or two… if the world still exists). Instead what we see is people studying demand and creating prompts based on that. Which… isn’t that different from how hollywood studios decide which script to greenlight or not.
You’re largely arguing what I’m saying back at me. I didn’t mean that the AI is bad, but that the AI content that’s out there has filled the internet with tons of low quality stuff over the past few years, and enough of this garbage going in degrades the quality coming out, in a repeating cycle of degradation. You create biases in your model, and feeding those back in makes it worse. So the most cost-effective way to filter it out is to avoid training on possibly AI content altogether. I think OpenAI was limiting the training data for ChatGPT to stuff from before 2020 up until this past year or so.
It’s a similar issue to what facial recognition software had. Early on, facial recognition couldn’t tell the difference between two women, two black people (men or women), or two white men under the age of 25 or so. Because it was trained on the employees working on it, who were mostly middle-aged white men.
This means that there’s a high demand for content to train on, which would be a perfect job to hire artists for. Pay them to create work for whatever labels you’re looking for for your data sets. But companies don’t want to do that. They’d rather steal content from the public at large. Because AI is about cutting costs for these companies.
And what is the difference between someone getting paid to draw a picture of Sonic choking on a chili dog by a rando versus an AI generated image of the same?
To put it simply: AI can generate an image, but it isn’t capable of understanding 2-point perspective or proper lighting occlusion, etc. It’s just a tool. A very powerful tool, especially in the right hands, but a tool nonetheless. If you look at AI images, especially ones generated by the same model, you’ll begin to notice certain specific mistakes - especially in lighting. AI doesn’t understand the concept of lighting, and so has a very hard time creating realistic lighting. Most characters end up with competing light sources and shadows from all over the place that make no sense. And that’s just a consequence of how specific you’d need your prompt to be in order to get it right.
Another flaw with AI is that it can’t iterate. Production companies that were hiring AI prompters to their movie crews have started putting blanket bans on hiring prompters because they simply can’t do the work. You ask them to give you 10 images of a forest, and they’ll come back the next day with 20. But you say, “Great, I like this one, but take the people out of it,” and they’ll come back the next day with 15 more pictures of forests, but not the original without people in it. It’s a great tool for what it does, but you can’t tell it, “Can you make the chili dog 10 times larger” and get the same piece, just with a giant chili dog.
And don’t get me started on Hollywood or any of those other corporate leeches. I think Adam Savage said it best when he said last year that someday, a film student is going to do something really amazing with AI - and Hollywood is going to copy it to death. Corporations are the death of art, because they only care about making a product to be consumed. For some perfect examples of what I mean, you should check out these two videos: Why do “Corporate Art Styles” Feel Fake? by Solar Sands, and Corporate Music - How to Compose with no Soul by Tantacrul. Corporations also have no courage when money is on the line, so that’s why we see so many sequels and remakes out of Hollywood. People aren’t clamoring for a live action remake of (insert childhood Disney movie here), but they will go and watch it, and that’s a safe bet for Hollywood. That’s why we don’t see many new properties. Artists want to make them, but Hollywood doesn’t.
As I said, in my ideal world, AI would be making that corporate garbage and artists would be able to create what they actually want. But in the real world, there’s very little chance that you can keep a roof over your head making what you want. Making corporate garbage is where the jobs are, and most artists have very little time left over for working on personal stuff. People always ask questions like, “Why aren’t people making statues like the Romans did,” or “Why don’t we get paintings like Rembrandt used to do.” And the answer is, because nobody is paying artists to make them. They’re paying them to make soup commercials, and they don’t even want to pay them for that.
Yeah… that is right up there with “AI can’t do feet” in terms of being nonsense people spew.
There is nothing inherently different between a picture made by an LLM and a picture drawn by Rob Liefeld. Both have fucked up feet and both can be fixed with a bit of effort.
The issue is more the training data itself. Where this CAN cause a problem is if you unknowingly train on endless amounts of ai generated content. But… we have the same problem with training on endless amounts of human content. Very few training sets (these days) bother to put in the time to actually label what input is. So it isn’t “This is a good recipe, that is a bad recipe, and that is an ad for betterhelp”. It is “This is all the data we scraped off every public facing blog and youtube transcript”.
Its also why the major companies are putting a big emphasis on letting customers feed in their own data. Partially that is out of the understanding that people might not want to type corporate IP into a web interface. But it is also because it provides a way to rapidly generate some labeled data because you know that customer cares about widgets if they have twelve gigs of documents on widgets.
And what is the difference between someone getting paid to draw a picture of Sonic choking on a chili dog by a rando versus an AI generated image of the same?
At the end of the day, we aren’t going to see magic AIs generating everything with zero prompting (maybe in a decade or two… if the world still exists). Instead what we see is people studying demand and creating prompts based on that. Which… isn’t that different from how hollywood studios decide which script to greenlight or not.
You’re largely arguing what I’m saying back at me. I didn’t mean that the AI is bad, but that the AI content that’s out there has filled the internet with tons of low quality stuff over the past few years, and enough of this garbage going in degrades the quality coming out, in a repeating cycle of degradation. You create biases in your model, and feeding those back in makes it worse. So the most cost-effective way to filter it out is to avoid training on possibly AI content altogether. I think OpenAI was limiting the training data for ChatGPT to stuff from before 2020 up until this past year or so.
It’s a similar issue to what facial recognition software had. Early on, facial recognition couldn’t tell the difference between two women, two black people (men or women), or two white men under the age of 25 or so. Because it was trained on the employees working on it, who were mostly middle-aged white men.
This means that there’s a high demand for content to train on, which would be a perfect job to hire artists for. Pay them to create work for whatever labels you’re looking for for your data sets. But companies don’t want to do that. They’d rather steal content from the public at large. Because AI is about cutting costs for these companies.
To put it simply: AI can generate an image, but it isn’t capable of understanding 2-point perspective or proper lighting occlusion, etc. It’s just a tool. A very powerful tool, especially in the right hands, but a tool nonetheless. If you look at AI images, especially ones generated by the same model, you’ll begin to notice certain specific mistakes - especially in lighting. AI doesn’t understand the concept of lighting, and so has a very hard time creating realistic lighting. Most characters end up with competing light sources and shadows from all over the place that make no sense. And that’s just a consequence of how specific you’d need your prompt to be in order to get it right.
Another flaw with AI is that it can’t iterate. Production companies that were hiring AI prompters to their movie crews have started putting blanket bans on hiring prompters because they simply can’t do the work. You ask them to give you 10 images of a forest, and they’ll come back the next day with 20. But you say, “Great, I like this one, but take the people out of it,” and they’ll come back the next day with 15 more pictures of forests, but not the original without people in it. It’s a great tool for what it does, but you can’t tell it, “Can you make the chili dog 10 times larger” and get the same piece, just with a giant chili dog.
And don’t get me started on Hollywood or any of those other corporate leeches. I think Adam Savage said it best when he said last year that someday, a film student is going to do something really amazing with AI - and Hollywood is going to copy it to death. Corporations are the death of art, because they only care about making a product to be consumed. For some perfect examples of what I mean, you should check out these two videos: Why do “Corporate Art Styles” Feel Fake? by Solar Sands, and Corporate Music - How to Compose with no Soul by Tantacrul. Corporations also have no courage when money is on the line, so that’s why we see so many sequels and remakes out of Hollywood. People aren’t clamoring for a live action remake of (insert childhood Disney movie here), but they will go and watch it, and that’s a safe bet for Hollywood. That’s why we don’t see many new properties. Artists want to make them, but Hollywood doesn’t.
As I said, in my ideal world, AI would be making that corporate garbage and artists would be able to create what they actually want. But in the real world, there’s very little chance that you can keep a roof over your head making what you want. Making corporate garbage is where the jobs are, and most artists have very little time left over for working on personal stuff. People always ask questions like, “Why aren’t people making statues like the Romans did,” or “Why don’t we get paintings like Rembrandt used to do.” And the answer is, because nobody is paying artists to make them. They’re paying them to make soup commercials, and they don’t even want to pay them for that.