ChatGPT - AI text

[Ser Not Appearing]

I have a friend who's really into AI stuff and he's been playing with art generators for a while now. He recently started playing with a chat bot and indicated that it's actually pretty good if you work at it. I decided to mess with it to see what I could do in terms of writing book scenes.

I didn't put a whole lot of work or thought into it but tried a few things. All in all I spent maybe a half hour, including account creation.

The AI generated text is certainly not great but it has the ability to generate some important elements of books. It had a bigger vocabulary and structured things much better than my 9-year-old could do, for instance.

One of the cool things it can do is how there's the ability to iterate with it. For instance, I had it generated scene where a man walks into a bar in order to drink. It didn't describe the man or the bar at all and so I asked it to add that. It gave me this description:

"The male customer strode into the inn, his demeanor brusque and impatient. He was a tall, broad-shouldered man with a rough, unshaven face and a scowl that seemed to be permanently etched on his features. A heavy cloak was thrown over his shoulders, and a sword hung at his side, a clear indication of his status as a warrior of some sort."

I then ask it to make the man appear sketchy and it gave me this:

"The male customer slunk into the inn, his demeanor furtive and shifty. He was a tall, thin man with a gaunt, angular face and a perpetual scowl. He was dressed in threadbare clothes, and a dirty cloak was thrown over his shoulders. A thin sword hung at his side, its rust-covered blade a testament to its lack of use."

It also changed some of the ways that other people reacted to him, where they were suddenly hesitant and suspicious, which is really interesting to me. Incorporating that description change into the narrative is pretty cool.

In another attempt, I had it generate a fight scene where one guy defeats three others. It gave background on how confident the three others were how they found out that they were wrong (motivations that I didn't input). It didn't describe any of the actual blows, so I asked it to add blow by blow details. It included round houses and unexpected punches and fast blows. It wasn't the level of detail I wanted but it definitely knew what I was getting at and added some stuff ... showed some awareness of the fact that books wouldn't just describe events but also describe what other people are thinking.

I did a third attempt where I asked it to generate a scene where someone walks into a council chamber to meet a leader and recognizes that the leader is someone other than who he says he is. It did that but it also discussed that the council was occurring because the grave threat to the union ... which isn't anything I had in the prompt ... it also decided that the man recognized the leader but couldn't remember where he knew him from ... another detail I didn't include.

---

Has anyone else used these tools? Any thoughts on concerns about creative content and the future?

I can imagine good uses where a writer with writers block can enter some prompts and get ideas for inspiration.

I can also imagine challenges where these tools could make the years-long process of writing an actual book non-profitable if they get good enough to just generate a book on demand based on reader prompts.

[Ser Not Appearing]

To get it to produce this stuff, I did include at the outset of the prompt that I wanted these in narrative book form in the style of a medieval fantasy.

I haven't yet played with whether or not it can write in different narrator perspectives but I think that'll be the next test I run when I have time. I'll generate something from an omniscient point of view and then ask it to change narrator type just to see how that affects things.

[JGP]

I'm not too dialed in on ChatGPT, but this is more typical example of an AI-dudebroartist-Pinocchio attitude toward art scraping:

 

 

I am a current drawslave, aspiring paintpig [digital painting] and I'm kind of concerned.

From what I've seen, won't touch them myself, you can turn out some neat stuff with prompts. Which is fine I suppose if you're just spitting whatever, but once you're client facing, minor adjustments required to complete the contract might be what kills that, but hard to say.  

 

 

[Underfoot]

Can we get it to write winds of winter 

[JGP]

Then there's stuff like this thread.

Pretty sure this shitpickle didn't check with anyone who understands ethics. 

Or people.

 

I haven't given it a lot of thought, but I'm not sure how we're going to regulate AI of these types.

[Ser Not Appearing]

3 hours ago, JGP said:

I'm not too dialed in on ChatGPT, but this is more typical example of an AI-dudebroartist-Pinocchio attitude toward art scraping:

 

 

I am a current drawslave, aspiring paintpig [digital painting] and I'm kind of concerned.

From what I've seen, won't touch them myself, you can turn out some neat stuff with prompts. Which is fine I suppose if you're just spitting whatever, but once you're client facing, minor adjustments required to complete the contract might be what kills that, but hard to say.  

 

 

 

I played with both midjourney and Dall-E 2 a bit. Originally, I was thinking I might be an easy way to get a logo for a podcast but that turned out to be an entirely fruitless endeavor. Since then, it's been a somewhat interesting thing to play with just to see what I can get it to generate and how different prompts change things. The kids also love coming up with oddball things just to see what it spits out. We had some really interesting images generated by asking it to draw the scariest man on Earth or a shopping cart eating a child. It can be fun.

My experiences do suggest that, at this point, AI does a really bad job if you want something super specific but that it does a great job if you want something themed or with a certain color palette or perspective or style and then a general subject matter. Some of the art you can get it to generate is legitimately amazing. But it's not easy to get it to follow through on even simple specifics like "a man plugging his ears."

What I think merits a lot of thought (other than copyright issues), and what I was getting at here with the chatbot part, is a recognition that these uses are very much in their infancy and they already do surprisingly well in some regards. Given time, there's every possibility that these will turn into something super powerful where you could generate a full book on demand based on prompts. That's something of a scary proposition if you're an author.

Granted, it could also be the type of thing like flying cars where it seems like you might go from cars and planes to flying cars but somehow that just never materializes. Maybe AI never quite gets to the point where it can mimic the creative process well enough to give us works that feel authentic and original. Maybe it stays something of a sophisticated mimicry. I don't know, but it's been really interesting to consider more and more over time.

[JGP]

16 minutes ago, Ser Not Appearing said:

 

I played with both midjourney and Dall-E 2 a bit. Originally, I was thinking I might be an easy way to get a logo for a podcast but that turned out to be an entirely fruitless endeavor. Since then, it's been a somewhat interesting thing to play with just to see what I can get it to generate and how different prompts change things. The kids also love coming up with oddball things just to see what it spits out. We had some really interesting images generated by asking it to draw the scariest man on Earth or a shopping cart eating a child. It can be fun.

My experiences do suggest that, at this point, AI does a really bad job if you want something super specific but that it does a great job if you want something themed or with a certain color palette or perspective or style and then a general subject matter. Some of the art you can get it to generate is legitimately amazing. But it's not easy to get it to follow through on even simple specifics like "a man plugging his ears."

What I think merits a lot of thought (other than copyright issues), and what I was getting at here with the chatbot part, is a recognition that these uses are very much in their infancy and they already do surprisingly well in some regards. Given time, there's every possibility that these will turn into something super powerful where you could generate a full book on demand based on prompts. That's something of a scary proposition if you're an author.

Granted, it could also be the type of thing like flying cars where it seems like you might go from cars and planes to flying cars but somehow that just never materializes. Maybe AI never quite gets to the point where it can mimic the creative process well enough to give us works that feel authentic and original. Maybe it stays something of a sophisticated mimicry. I don't know, but it's been really interesting to consider more and more over time.

Well... I believe you, and on the level don't feel like you're wrong. The copyright issue is going to be interesting, even if a little nauseating. Very, very few artists who share their traditional or digital work online via portfolio type interfaces or whatever consented to having their work scraped. Of course it can spit out some beautiful pieces. It is scraping and recompositing art. And hey, maybe they'll eventually figure out how to code it so it doesn't also share some sometimes real live artist's signature lol

Referring to the twitter thread I shared above, and what they did? I don't know if my camp is empty, but it's firmly on this shit needs to get regulated sooner rather than later. I mean what, while the world is burning down around us we might one day be able, with the computing power to render, etc, your own whole movie.

While I'm thinking about it, we need some kind of watermark for spoofing of any type when it comes to fraud.

Like, now.  

[Ser Not Appearing]

As a somewhat cynical person living in the US, I imagine the rest of the world might have hope of legislating some reasonable protections and controls. We don't.

[JGP]

That's heavy. 

The cynical part of me wonders if some techbro or bros hasn't already soft sold it as sure, slow roll legislation and we'll have enough by then anyway. The digital representation of our selves might be on the line here. 

Not sure how I feel about Canada's resolve looking ahead at this now myself.

  

[IFR]

I think the implications of this are exciting. I really look forward to humans being removed from the equation in the production of art. Humans have always been a poor cog in the mechanism of production.

And I think the belief that humans have some rare insight into the "soul" of art is pure narcissistic foolishness. There's nothing special about Ulysses or Moby Dick. What ever humans can do, with time it is likely AI will be able to do better. I do wonder what people will make of an AI smarter than they who can conjure an obscurist work better than Joyce, can create metaphors more subtle and clever than Melville, a write beauty in language better than Nabokov or McCarthy? It will be interesting to see that resistance once again as people try to cling to the idea that humans are special.

It would be nice to be able to read a custom made AI story that conjures the most satisfying narrative to your personal preferences. It will be especially wonderful if the presentation of this story can be augmented by virtual reality. 

[williamjm]

13 hours ago, Underfoot said:

Can we get it to write winds of winter 

I suspect if it did that then the quality of the plotting might be more Benioff and Weiss level than GRRM.

[Proudfeet]

11 hours ago, Ser Not Appearing said:

Granted, it could also be the type of thing like flying cars where it seems like you might go from cars and planes to flying cars but somehow that just never materializes. Maybe AI never quite gets to the point where it can mimic the creative process well enough to give us works that feel authentic and original. Maybe it stays something of a sophisticated mimicry. I don't know, but it's been really interesting to consider more and more over time.

I don't know, maybe I'm behind the times but I feel like AI is just a sophisticated glorified flow chart. It's a result of very specific design and inputs. Machine learning is just pattern recognition increasing its stock of both correct and incorrect answers so that it can more reliably land on the correct answer. Its limited by the ability of the programmers.

I'll respect AI's command of language more when they don't botch simple translations. And creativity is another hill to climb after that.

[IFR]

28 minutes ago, Proudfeet said:

I don't know, maybe I'm behind the times but I feel like AI is just a sophisticated glorified flow chart.

This describes almost every blockbuster movie and successful Netflix show.

[Proudfeet]

6 minutes ago, IFR said:

This describes almost every blockbuster movie and successful Netflix show.

I'm sure it describes most tasks. My point is that the bottleneck is the programmers. AI isn't going to independently improve by themselves.

[JGP]

28 minutes ago, Proudfeet said:

I'm sure it describes most tasks. My point is that the bottleneck is the programmers. AI isn't going to independently improve by themselves.

Or be capable of actual art on its own, beyond still life I suppose. 

Inspiration doesn’t fall within the realm of the soul [whatever that it] but certainly consciousness, which AI is far from.

Even the language bots like ChatGPT, they’re still aping the thought labor of live human beings. 
 

[IFR]

22 minutes ago, Proudfeet said:

I'm sure it describes most tasks. My point is that the bottleneck is the programmers. AI isn't going to independently improve by themselves.

Not in the near future, I suppose. AI is highly parameterized for now. But AI self-improvement (independent of programmer influence) is a very likely eventuality.

My point is that most of what constitutes "creativity" is really a prosaic thing and within the grasp of even a parameterized AI with coaxing by programmers.

3 minutes ago, JGP said:

Inspiration doesn’t fall within the realm of the soul, but certainly consciousness, which AI is far from.

I don't think there's anything that establishes "creativity" in art as particularly deep and outside the grasp of an advanced automaton. Does Hamlet kill his uncle or doesn't he; and does he soliloquy afterward - what is dramatically more satisfying and "layered"? It's simple pattern recognition. An AI trained on other works to recognize symbolism, dramatic twists, and character arcs then producing some iteration thereof is really no different than a person doing the same thing. The "thought" that went into it may be alien to the conventional approach, but the end result will converge to the same thing. As with all new technology, there will be conspicuous flaws at first, and then uncanny valley territory, and then it will be to the point most people won't be able to detect the difference, they'll have to train AI to detect other AI.

[JGP]

Just now, IFR said:

I don't think there's anything that establishes "creativity" in art as particularly deep and outside the grasp of an advanced automaton. 

Guess we'll see. If say, some international organization with clout banned art scraping [and recombination thereof] these AI programs would take a prompt and go where, precisely? When an AI produces a piece of art [of whatever form or medium] unprompted, I'll take the above quoted line of thought seriously.

 

[IFR]

16 minutes ago, JGP said:

Guess we'll see. If say, some international organization with clout banned art scraping [and recombination thereof] these AI programs would take a prompt and go where, precisely? When an AI produces a piece of art [of whatever form or medium] unprompted, I'll take the above quoted line of thought seriously.

 

I think there's a slight bit of miscommunication. I doubt AI will do anything too far out of the parameters of its programming (yet). But I do think it will be capable of operating within its paramaters indistinguishable from a human, eg. if a group of experts were given new samples from the most acclaimed living artists and samples of art produced by AI, the experts' ability to determine which was which would be no better than randomly guessing. The human artists produced their work from "creativity" and "inspiration" and their exposure to previous art, and the AI created the art by extrapolating from art it scanned, so the pathways to that end point were different, but it was the same end point.

I doubt legislation will have much of an effect curtailing this sort of technology. It rarely does.

But like you said, we'll see.

[JGP]

47 minutes ago, IFR said:

I think there's a slight bit of miscommunication. 

There's not. The failure to communicate is probably mine.

You keep finger quoting creativity and inspiration [re: art, not invention] like they're not real things that're purposefully intrinsic to conscious beings. 

I write as well [polishing something for the Clarion workshop, also working on a book] and I've been drawing since I got my first crayon before that, so clearly biased, but you and I, I think, are exemplar of left vs right brain. On this in particular and also in general.

[Ran]

14 hours ago, JGP said:

recompositing art.

They don't contain any databases of images from which to recomposite works. Stable Diffusion and Midjourney AIs work by basically learning much the way a person learns: they see images that are described to them and form ideas of what constitutes "pop art" or "dog" or "gouache", learning the underlying "rules" behind these things, and when you ask it for a "pop art image of a dog painted in gouache" it takes a completely random noise-filled image and attempts to manipulate it over multiple iterations to create an image that it thinks conveys those ideas. 

The models they use are full of weights in a "latent space" of thousands of dimensions, but there's no images contained in it, just the "rules" for various concepts which it then tries to run through, concept to concept (and sub-concept to sub-concept) to get at an end point. Human artists are much the same -- our brains contain a conceptual cloud from which we can pluck meaning, so that we can imagine strange things and put them to paper/canvas/screen.

Now, it's the case that extremely famous pieces of art can be reasonably produced by these models -- the Mona Lisa is the obvious one -- but that's, well, the Mona Lisa, and I guess if asked many artists can conjure up in their heads the Mona Lisa and homage it.

The problem for artists comes mostly when they learned that among the thousands of dimensions in the latent space were things that reasonably constituted styles -- it learned "rules" that told it, "This is a D&D cover style splash art of an Orc" or "This is a Nazi demon in the style of Mike Mignola" and so on... and if the base models don't understand these concepts very well, well, anyone with access to a GPU can train a model to understand those and other concepts better. Not by having it store up actual images, but just showing it images that it starts to "understand" the rules behind... But of course, artists can and do mimic the styles of other artists, to study, to homage, or to make a buck.

Style, like design, isn't protected by copyright, hence all the Patreon artists doing Ghibli-style or Disney-style commissions. If it's okay for people to do, I don't see why it's not okay for an AI to do at the direction of people. 

That said, I think the real power of AI image generation is that it's going to be a tool that helps artists improve their workflow and be able to do simply more of what they want with less of the drudgery. You can come up with a rough idea for something you want to do and then have an AI generate a hundred variations on that idea to help spark your creativity. You can take a rough sketch and create variants to see possibilities you hadn't considered. You could have an AI that acts as your personal inker or colorist, taking your black-and-white pencil work and turning it into something finished with little outside input-

There's a lot of people at the very basic level of artistry -- the Fiverr artists doing Chibi-style character commissions  and so on -- who are going to find commissions drying up as Joe Schmoe can collaborate with an AI to get closer to their vision of some character than they can ever get through another person, because the AI doesn't really have its own opinions or creativity, it just has concepts and how to turn them into iamges. But there's many serious artists out there who are, I think, going to be perfectly fine for the forseeable future. 

In the end, I think these models work under appropriate legal principles. Whether the law should be changed to allow humans to look at and learn from art but restrict machines from doing the same, well, I'm opposed purely because I think it's nonsensical and backwards-looking.

(FWIW, we've been playing with Stable Diffusion to create images for characters that have gone many years without depictions on the Wiki, and to tool about and create portraits for some of the characters on Blood of Dragons MUSH. In the case of the wiki, we're explicit that these are to be replaced by art from human artists should anyone ever make them that are ASoIaF (rather than TV show) specific, meet a basic level of quality, and we get permission to use them. In fact, when we posted a take on Bellenora Otherys, an editor ended up remembering or discovering that an artist had recently depicted her and got permission and put that one in there instead. Which is great (and I think led to their work being featured on a few other character pages).)