A while ago Tipa, who's probably experimented with AI more widely, deeply and for longer than anyone else in this part of the blogosphere, posted her conclusions on its future. As she summed it up
"The AI revolution, as it currently stands, is just a generator of garbage. You can find gems in garbage, but you have to dig through a lot of garbage to find it."
I wanted to do at least one post on AI for Blaugust. Actually, what I really wanted was to get one or more of the AIS to write me a post for Blaugust. Maybe I already have... (I haven't.)
As Tipa so rightly points out, though, it's not quite as simple as that. I've had various AIs write me stories and posts and they've all been...fine. Coherent, cogent, bland as heck. Nothing I'd want to put up here and hope no-one noticed the difference, that's for sure.
Still, I do question the wisdom that's beginning to be received that says no-one would want that kind of output instead of the innately superior work of a human. Possibly the most disturbing thing about AI generated text is that it reads exactly like the kinds of dull, spiritless writing I've been reading for all my life.
It used to come through the door in what we called junk mail back when mail was a thing. For the last twenty years it's been filling up screens instead. Until very recently all of that stuff must have been written by someone because the technology didn't exist for it to be otherwise.
Presumably someone got paid something for every tedious, soul-sucking word of it. I see absolutely no reason why AI shouldn't be able to replace whoever was grinding the stuff out and it probably already has. I imagine a living, breathing human has to read it over at some point and make a few corrections but that's editing, not writing.
Tipa also said "... if your work is such that ChatGPT could do it… you need to find better work." I can't argue with that but I do think there have to be a lot of people whose work is such that ChatGPT could do it. It's the work none of us would miss if it wasn't there in the first place but we surely can't pretend it's not there at all.
As the technology exists right now, I can't see much chance of a purely AI-generated novel winning the Booker or the Pulitzer. Purely AI-generated scripts aren't going to be up for any Oscars, actors' and writers' strikes notwithstanding.
When it comes to art and music, though, I wouldn't be quite so sure. The quality there is much harder to argue with. Some of the images generated by the likes of Midjourney or SDXL are stunning. AI cover versions are racking up huge viewing figures on YouTube.
In the arena of audio-visual AI, the concern isn't quality. It's provenance. There's not just a question of ownership, there's one of attribution, too. Leaving aside the ethics of having trained the the AIs on vast bodies of human-generated images and sounds, there's huge suspicion around whether some of the images are just being pulled verbatim from existing work. As for music, it's the same damn issue as sampling, isn't it?
The music industry, always keen to put the interests of musicians and artists first, is already maneuvering to nail down a deal whereby rights owners (By which they mean record companies.) can get a cut. Now they've seen that people like it, they don't want to stop it, they just want to own it and sell it, just like they've been doing these last hundred years or so.
All of this is geting old. It was a lot of fun a year ago but as with all new technologies it's rapidly losing its rebel cred as the mainstream moves to assimilate and flatten out everything that was interesting or exciting about it.
The way I look at it, if we ever do get to the stage where I can't tell whether something was created by an AI or a human, it might as well be a human. The thrill of it all used to be the alien quality AI brought to everything it created. Without that, why the hell bother?
Well, to make money, obviously. Clearly, if AI can produce work that sells as well as work produced by humans, or performs whatever role it needs to perform as well as a human can, whoever signs the checks (Hah! Checks!!) is going to be ecstatic. We're a long way off from AI-generated movie stars but if AI means not having to deal with a bunch of extras clogging up the parking lot then plenty of executives must be thinking "bring it on!"
But I didn't come here to talk about Hollywood and its problems. Oh, boy. That's a whole other post.
It is, however, very imagistic, as much teen poetry tends to be. It seemed like it might make an interesting test for a bunch of AIs, to see what they'd come up with.
The excellent NightCafe, for whom, perhaps ironically, you may consider this post an entry in the week's Blaugust theme of Creator Appreciation, gives me access to more than a dozen AI art generators. Some of them are free to use and some need the credits you get when you subscribe. Luckily, NightCafe also gives away a generous supply of free credits for this and that and I have more than 250 stashed away, not counting an extra ten for the exclusive, subscription-only, top-of-the-range SDXL.1.0.
I spent a while before I started this post playing around with the idea and guess what? Tipa was right! If I just enter a line or a phrase from my poem, no matter how descriptive or visual, the resulting image is... garbage.
Garbage as in it bears almost no relation to the image in my head or, I would guess, anyone else's head if they read the same prompt. Either that, or it fixates with laser-like intensity on a single word and just puts out a picture of that.
Some of the images are gorgeous. Some of them aren't. Some are every bit as weird as I hoped they'd be. Some are just dull. Most of them are functionally useless although there is one that does pretty much work. One out of eight. It's not a great ratio.
I expect you'd like to see them now. Okay, but please bear in mind none of these images came from my mind, even if I did write the poem the prompts came from.
Prompt #1 : Teeth of neon, black like the night
I mean, seriously, where do you even start? Stable Diffusion at least has some neon in there and conceivably those upright pillars could, metaphorically, be "teeth". I can't really fault it for not being "black like the night". I don't believe black neon is a physical possibility in this universe, only in the adolescent mind.
The other three, though... They look like three album covers. The bird-thing's probably a rackety mid-western screamo band, the scratchy cityscape is a compilation of no-wave covers by a bunch of Japanese bands you never heard of and the two guys posing in front of some reastaurant's flock wallpaper are Kraftwerk wannabes from sometime in the twenty-teens. Not a one of 'em has anything to do with the prompt.
Prompt 2: Tiger will come for you all
Now, I think that's definitely a much more attractive set of images and not just because tigers make for an intrinsically appealing subject. The worst by a mile is the DALL-E 2, which just looks like a badly-cropped stock photo from a zoo promo. The Stable Diffusion is a much better photograph but again it seems to have been lifted verbatim from some collection of wildlife photography.
The other two are both gorgeous images. I notice the SDXL Beta seems to have an artist's signature in the bottom right corner. That happens quite a bit with these things. I'd love to know if it's the genuine signature of the real artist who painted that picture and has had it "borrowed" by the AI or whether the AI has just added a signature because it has some training that says pictures of that kind are usually signed.
The SDXL 1.0 image is the only one of the eight I can imagine using as an illustration for the prompt. It is a tiger, it is coming for you. Literal, sure, but accurate. Also, a very powerful and appealing image. I particularly like the slightly oversized, club-like right paw. It gives the picture just that hint of wrongness I feel ought to be the reason for using an AI image in the first place.
As for using AI to speed things up... this post took me three hours to write and I haven't even edited it or "created" the illustrations I'll use in the introduction yet. If AI's going to do my job for me it's going to have to pull its finger out. I'm doing all the heavy lifting here.
And now I'm off to make myself a coffee, another thing AI won't do for you. Oh, wait...
Posts
Those first couple of cartoons on your post look like something that came from the mind of R. Crumb, only with more expressionism.
ReplyDeleteAnd to be honest, I've thought for at least a decade that an AI could write as good as some documentation writers. Doesn't make the documentation good by any stretch --and I typically want to claw my eyes out after an extended session of trying to understand all the permutations found in software documentation these days-- but it's such low level drivel that I figure an AI such as Chat GPT couldn't be any worse at it. I get that documentation and legalese is supposed to be thorough and not necessarily engaging, but wouldn't it hurt just a little bit for this sludge pile to be a bit engaging? It would at least make my days a little easier to navigate.
I can see a bit of underground comix in there, too. It's surprising, since my actual prompts specified either "Hanna-Barbera", "Saturday morning cartoons" or "Scooby Doo". It was only when I made it as specific as "Scooby Doo" that I started to get what I was after, too. If you just want something generic like "A beach at sunset" you can be fooled into thinking the AIs will do all your work for you but as soon as you have an actual image in mind it's a lot of work to get them to bring it to life.
DeleteAs for ChatGPT and the rest, they're astoundingly good at producing coherent prose. Better than most humans. At producing something you'd actually *want* to read, though, you could throw a stone in any schoolyard and hit a kid who'd do *much* better. (I'm not recommending you try it!) The ironic thing is, for fiction the AIs can't provide the spark of interest but for techical documentation they can't be trusted on the accuracy of the information. Either way, some human is going to have to get in there and fix it before it goes out.
Bit of an internally inconsistent ramble coming . . .
ReplyDeleteThe funny thing about most AI applications I have played with is that how interesting what you get out of them tends to be directly related to how interesting the prompt is. I can see a future where it might lead to a society where pure technical skill level and spare time stop standing in the way of artistically compelling ideas coming into being. However, we aren't quite there yet by any means.
I have a real love-hate relationship with ChatGPT. I can see where it could seriously accelerate some types of writing. But holy cow is the prose that you get out of these things bland and generic, as you point out. Now when I am watching some generic CW superhero show where the script sounds like a string of cliches stapled together, pretty much the worst insult that sometimes goes through my mind is "This script could have been written by an AI."
And of course entertainment industry moguls are thinking the same thing, but not in a "This just goes to show why we still need humans in the loop" way . . .
That's pretty much where I'm at, too. I'm reading a book about the music scene in 1982 and I see a lot of similarities between the way the development of technology then (programmable drum machines and synths especially) opened up access to people who would previously have been thought of as "non-musicians". A lot of truly great music came out of that but at the time it was certainly felt by some - probably many - more traditional musicians that it wasn't really "music". Listeners, however, felt otherwise, and time has proved them right.
DeleteI believe the current level of "AI" means nothing of value is likely to appear without a human having put a great deal of work into shaping and polishing it. That was part of what I was trying to demonstrate in this post. Just giving even the better AI art generators a straight-up prompt produces wildly varying results. A human has to sort them to get something appropriate and then iterate on it to get the exact image that's required.
For example, I love the warped, distorted, freakish images in the large illustrations in this post but clearly they couldn't be used in may other contexts than as illustrations of why AI can't replace humans. When you see the extremely detailed versions of something totally recognizeable, like the "What if Wes Anderson..." stuff that's been so popular on YouTube, you can bet an actual person spent many hours working with an AI to get it just right.
I keep making comparisons to sampling and that's because I think the real issue here is attribution. The AIs are just another instrument like samplers were but because they use snippets of pre-existing work there's a perceived need to credit that and also pay for its use. I have much more radical opinions on creative ownership that don't really include any residual rights beyond first publication and I'd love to see AI fracture and eventually destroy the existing rights structure but for the time being I accept we need some form of licensing and attribution just to avoid an endless round of takedowns and lawsuits.
On syths and such in the 80s, I was very much into the Art of Noise in the late 80s. I was considered pretty weird among my teen peer group because of it, but I think they were ahead of their time. They really anticipated a lot of electronica years later, though most of it is primitive by modern standards. However, "In no sense? Nonsense!" still holds up pretty well imo.
DeleteAs far as remixing snippets of other things goes, I feel like DJ culture provides a useful analogue. Some DJs are really just stringing together other people's songs, with the same level of thought that goes into a mix tape and some real some artistry (if limited) that involves slipping the beats from one song into those of another so that there is no pause between tracks. Others like DJ Spooky are able sample snippets and layer their sources in such a complex way that it's completely new music to all intents and purposes. Between those two extremes there is a lot of variation.
To me ChatGPT is still working closer to the dance DJ than the "highly original artist that happens to work with samples" end of the spectrum more often than not.
PS: If you goodle DJ Spooky his biring dance stuff tends to come up first. I am thinking more of things like Songs of Dead Dreamer. Here's a track from that:
Deletehttps://www.youtube.com/watch?v=08MXmoL3rYk
Lot more going on there than just stringing tracks together.