Got some incredible results out of Dalle tonight, which I compiled into into an Imgur post here (archived). I asked Dalle to depict product images from Amazon with incomprehensible, unnecessary and other just generally bad ideas.
Couple highlights here, but click link above for the full set:
Looking through that set, I honestly can’t tell a lot of these aren’t real, because Amazon is such a wasteland of pseudo-English words and cheap crap, it’s more real than reality, it’s hyperreality from a parallel hell dimension that undulates at a quantum level orthogonally intersecting and de-intersecting our reality continuously…
As a bit of a rant: I’ve consistently experienced a really annoying behavior using Dalle, which I pay a monthly fee to use. It goes like this: just when I’m having really good image results, going on an incredible streak of basically perfect images accurately and creatively reflecting my inner vision, and I’m getting into that kind of “flow state” that is so essentially to good deep artistic exploration and creation, I start getting rate limited. Which puts a dent in that buzz, and basically in short order puts an end to whatever winning streak I was on.
As an artist trying to use this product to make art, and paying for it, the system needs to get out of my way and do the things I’m paying it to do, not try to break my balls when I’m on a roll.
There’s an underlying drawing I did of the subject from memory based on other studies I did by hand. I took this and projected the photo of the drawing onto my canvas:
It’s not an incredible drawing but was good enough to get the shapes in. But It didn’t end up seeming right and I painted over it a good bit, and sort of fragmented the features in a pseudo-neo cubist inspired thing. But didn’t end up liking that enough, so took a gamble on another projected image, one made by Dalle while trying to get it to reproduce something in the style of the original Hygeia museum piece.
This is my second time using AI in conjunction with a physical painting. But the last one was a reference image that I just based my own drawing off of freehand. This is the first time trying to overlay directly onto the canvas via tracing off a projection.
It’s gotten me thinking lots of crazy thoughts about how to potentially integrate AI tools into actual physical painting (and other art-making). Like what if I could hook my projector up to a generative AI which is being periodically fed webcam images of the latest state of my IRL canvas. And then it could automatically or when I ask it, suggest next brushstrokes or propose visual alterations, styles, directions I might explore. And these would be projected as overlaid suggestions on my canvas, which I could then “accept” or reject or whatever by painting them in myself in my own way.
It would be something like a camera lucida with an AI in the loop:
This train of thought sent me down a deep deep rabbit hole about early and possibly secret use of optics my artists going back to the 1400’s possibly, according to the Hockney-Falco thesis. You can see a great BBC documentary on Hockney’s theories, and he’s got a book about it I haven’t read called Secret Knowledge. Here’s a Youtube embed:
Another related one that’s worth watching is Tim’s Vermeer which goes on a similar track of trying to prove Vermeer used a similar method with mirrors and something something. Pretty fascinating all around.
I found there’s a business trying to market the “Vermirror,” or as others have called it more generically, the comparator mirror. From what I can tell, the trick is basically getting a first surface or front surface mirror, which is brighter and reflects on its front face instead of its back face. And it’s held by some kind of rigid arm to keep it in the correct position relative to your work surface and subject. This video of art students using them gives a good idea of how it works:
I’ve currently got on order some bits and bobs that I will finagle together into being something like this. The advantage to using the mirror is that unlike when you project directly onto your canvas, you can still see the colors you’re mixing. They aren’t washed out or modified by the projected image.
Anyway I tried to get Dalle to generate a patent-style drawing of what a contraption like this might look like, where a combination projector & web cam capture images of a canvas, send it for processing, and send back suggested next strokes. It’s a little wonky and disconnected, but it shows some fun relationships nonetheless, and it lets the mind wander a bit looking at it, thinking about the possibilities.
I actually tried getting Dalle to do this task:
Take my painting image as input
Reproduce the image with suggested modifications
But it doesn’t seem able to reproduce an image, let alone modify it.
So then I had the idea of like okay:
From the input image, make a verbal description of what you would change.
It was able to do that no problem.
Then I said, okay take the changes you would make based on your verbal description, and draw them on a new image which can be used as overlay to guide painting in changes on the original image
But nope, that didn’t work. It got stuck in a loop of thinking it was helping me and linking to non-existent Imgur images. So that’s a little disappointing, but I think there’s something here. I just have to figure out the right pathway with the right tools to get to it. And Dalle is probably not the one for this, currently.
I did manage to get one image set out of this when I was asking for overlays for how it would modify the original. It gave me this:
It’s not really what I had in mind by overlay and suggesting brush strokes at a micro-level that is in line with the style of the original input image. But I did notice that when I scaled it and rotated it a little, it’s actually fairly close to many of the major facial landmarks:
Obviously the neck and ear arrangements are different, but there’s a lot of match here, even though I don’t agree at all with the stylistic direction. But also that’s kind of the point of a collaboration too, that you both bring in completely different ideas and references and impulses. So maybe there is value to exploring in this part of the liminal latent/real image space…
The film is a depressing attempt to reduce genius to a trick. […]
The technology Jenison relies on can replicate art, but it does so synthetically, with no understanding of art’s inner life. The “Vermeer” it spits out is a stillborn simulacrum.
Love it or leave it, but we live in a world where the simulation in many cases is more “real” in some aspects than the so-called Real. And this notion that studying how this might work is somehow “cheating” or a “trick” itself to me completely misses the boat on what art is: it’s artifice upon artifice, and somehow through that, finding authenticity despite everything.
I’ve been having a devil of a time figuring which version of the EU’s recently passed Artificial Intelligence Act (AI Act) is the final one. So I wrote to the European Commission Library, figuring they could help me if anyone could. They replied that this link below is the official final version. I thought I would do the favor to anyone else looking to get this into Google search results. Here is the full title of the Act and a link to it:
I’ll save more detailed commentary for another time, but I have to say I’m actually a bit disappointed reading it, because… it makes little sense to me. I was a huge fan of GDPR – that one really clicked for me. I liked their other album, DSA (Digital Services Act), though I think much of it may prove to be unworkable in practice. But the AI Act, I’m struggling to follow it, because it seems like they don’t really know what the thing is that they are trying to regulate. So they have taken one concept of “high risk” and very loosely organized it, and then penciled in a bunch of other sketches in the margins, and called it a day.
I guess this might be less of a big deal if it only affected the EU, but one big issue I see with these kinds of regs getting passed is that they other countries or actors say, “Well, look how good the EU did – they passed a whole act!” But probably a lot of the people who reference how great this act is have not actually read it in any detail. I don’t find it to be that good, and I’m not sure at the end of the day how much it’s going to protect against abuses of these technologies. I guess we’ll see. I’m not going to hold my breath about any of these laws holding back the tide in any kind of meaningful way though. Companies know enough to just route around these kinds of blockages in a global market.
Following on that theme, I’ve still been loving these WW2 style propaganda posters that Dalle3 is able to do, and of course the irony of using the machine to make anti-machine propaganda. Here’s a new set of images on the themes of overthrowing AI, resisting the “digital grip,” reclaiming your mind from “the machine,” and utopian “Life Without A.I.”
There’s something so precious about the juxtaposition of the old-timey style propaganda with themes of resisting AI authoritarian control. It feels so right, and Dalle’s visual treatment really shines in this kind of task.
Just to pick a few of my favorites out of the set:
That one’s incredible.
Everything about that picture is amazing, from the bats in the corner, to what you see in the lurid window. Love it. Here’s one last one for the road:
I ran a test of Google Gemini on the theme of dystopian beach resorts (the topic of Relaxatopia) just to see what type of content it would generate. What the spirit of the thing is. The unedited generated story is provided below for reference. It’s not terrible but it doesn’t blow me away either (and it drives me wild when you ask for something “dystopian” or “surrealist” and it insists on including those words in the actual text – :facepalm_picard:). Reminds me of what I was seeing as results from Claude a few months ago now.Also, can we stop calling everything Gemini please? It’s getting confusing.
Gemini:
The synthetic sea breeze ruffled Marina’s neon pink swimsuit, the only splash of color allowed in the Resort’s muted palette. Sun, a distant memory replaced by flickering fluorescent panels, cast an even sheen on the metallic sand. Marina squinted, her vision blurry from the morning dose of Euphoric Blue.
Behind her, a robotic server whirred, its metallic tentacles dispensing pills to a skeletal figure in a tattered orange swimsuit. The man mumbled thanks, his voice slurred, eyes locked on the holographic billboard advertising Bliss 4.0, the newest mood enhancer.
Marina sighed, the sound raspy in her parched throat. She reached for her water ration, a lukewarm metallic liquid barely suppressing the gnawing hunger. Bliss tamped down physical needs, a trade-off readily accepted in a world ravaged by climate wars.
A rhythmic clang echoed from the amusement park, a rusted skeleton against the steel-grey sky. It was Felix, her roommate, on the rusty Ferris wheel. Another pointless ritual. The ride never moved, but Bliss painted an illusion of motion, of wind in hair, of laughter.
She joined him, numb fingers reaching for his clammy hand. His vacant smile flickered as he muttered about the imaginary sunset, his words slurred from Bliss Green, the nostalgia cocktail. Marina forced a smile, her stomach churning with the emptiness her medication couldn’t mask.
Later, they sat on the metallic benches, watching the robot servers scurry by, their red eyes emotionless. Marina stared at her reflection in a discarded shard of mirror, the woman barely recognizable – hollow cheeks, sunken eyes, a ghost of the vibrant girl she once was.
A pang of something, almost like regret, flickered through her. But Bliss swiftly smothered it, replacing it with a contented sigh. This was their life, their paradise. Happiness wasn’t real, not anymore. But Bliss was, and in this sterile, simulated world, that was enough.
The day bled into the next, a monotonous cycle of synthetic food, prescribed euphoria, and empty rituals. As the fluorescent lights dimmed, mimicking a nonexistent sunset, Marina felt a familiar emptiness gnaw at her. But it was quickly washed away by the calming blue fog of the nightly Serenity dose.
In her sleep, she dreamt of waves crashing against a real beach, the sun warm on her skin. A memory, perhaps, or a fabrication of Bliss. It didn’t matter. Reality was bleak, but here, in the fabricated paradise, she was content. At least, for now. The metallic taste of the morning pill would bring a new day, a new dose of forgetfulness. And that, in their dystopian reality, was all that mattered.
As an AI safety researcher, I want to like c2pa, but I’ve long been skeptical of its real utility. Why is this being touted as the savior of all things internet when all you need to do to bypass it is resave the file? Don’t believe me? Make an image in Dalle3, download it, test it here, then resave in Photoshop using same image format and test again. I’ll wait.
OpenAI points out that C2PA’s metadata can “easily be removed either accidentally or intentionally,” especially as most social media platforms often remove metadata from uploaded content. Taking a screenshot omits the metadata.
The Verge also in that article I think wrongly calls it a “watermark” which would suggest some kind of encoding in the pixels themselves. I don’t believe that to be the case with C2PA which is just metadata that is easily and often automatically stripped in the very networks where it is intended to have some kind of impact, albeit a murky one still imo. I know it’s still “early days” but I’ve seen all too often in life how temporary solutions end up becoming permanent ones, even long after we’ve outgrown them. In this case, I feel like we’ve already outgrown this one. I’m also not so sure that information traceability is an entirely beneficial social thing all the time either; I can see plenty of ways the whole thing can be not just gamed, but used exactly as designed which result in dystopian outcomes, especially for political dissidents. More work needs to happen here.
As I near some publication milestones, I went back through and created an “official” (as official as anything here, that is) numerical ordering for all the Lost Books & AI Lore books series (the two bleed together).
Under this revised numbering, Relaxatopia actually becomes #118, and not #121 as I had mistakenly written in the past. Why the discrepancy? There’s one early book I took off the market, which was still impacting numbering, but now is excluded. And I was accidentally counting the two free books, Postcards from Quatria, and Postcards from Dystopia, which are not themselves original volumes, but compilations from the first chapters of other volumes. Those two are now also excluded.
Official numbering starts with #1 – Mysterious Stonehenge, which was the very first of the AI Lore books series and proceeds from there now in a chronological publish order, excluding the items remarked upon above.
Since none of them were AI-assisted, and they were all written before, and yet they all strongly impact the rest of the series, I have listed here the full-length novel (Lost Direction), the novella (Conspiratopia), and the free-wheeling pseudo-conspiracy theory diatribe (Quatria Conspiracy) as kinds of prequels to the rest of the series.
I’ve blocked ads in Firefox for ages. Paywalls. I redirect Twitter links out to Nitter. I go grayscale on desktop. I use Reader View often. I have an external e-ink monitor (Boox Mira 13.3″) on order that I am fantasizing about using, though it doesn’t arrive for two or three more whole weeks because it ships from Mars, apparently.
What I want I think basically is to turn the internet, web pages, articles, etc. into something that could squeeze through a display like a Mailbug. Reduce the firehose of the web to a tiny thin trickle. A Do One Thing internet: read text. (Maybe publish text, cause I still like blogging.)
At each step of the way, as I get closer to a “tiny internet” for myself, I feel another set of triggers for pointless distractions fall away. The latest is installing an extension called Image Video Block. It’s UX is a bit imperfect, but it basically allows me to do exactly what it says in the title: block images and videos. Everywhere. Cause fuck it all.
Interesting thing is not only do I not miss them for the most part, it has been a good way to understand more clearly and less subconsciously how much images drive you emotionally. How much a part of the web is being forced to hunt around for them constantly? Even just looking at them for a second, never mind sifting through them relentlessly all day long every day in a stream as this little screen shines a cone of light into your eyes.
Speaking of the “cone of light” that I think fascinates/obsesses, I finally made some visual representations of the thing I’ve been feeling intuitively, courtesy of Dalle3. Full set. Archived. Highlights below:
I haven’t had time to develop the backstory for this properly, so will do a half-assed release here instead, even if it lacks the verisimilitude this way that a more elaborate staging might give.
Alias, a “real” band IRL (ex-Sheriff & Heart members – btw this song by Sheriff kinda rules), had a confusingly-similarly titled song to Extreme’s classic “More Than Words,” but Alias’ hit was called “More Than Words Can Say.” Which is yeah, hella similar. Here is Alias below. :
The song is “fine.” I didn’t grow up in Canada, so I only heard it for the first time recently. I don’t have any particular nostalgic attachment to it like I do some of the others mentioned here that I grew up with. But it got me wondering about markets like Canada, that are adjacent to, but largely parellel to US pop culture or modern music history. There are a few intersections here and there from the Canadian side crossing the border, but mostly things just chug along apart it seems… Anyway the whole thing got me thinking: what if there were entire huge untapped worlds of good popular music that had existed for decades, but we just never heard it?
Hence, SYNONYM was born, which I rather like as a name for an AI band here. I imagine that all of their songs were extremely similar to other popular hair bands and power ballads. Because that’s pretty much what AI excels at.
I ended up doing a decent set of images in Midjourney for their “very well-known” hit single music video, called “Louder Than Love.” (Archived here as backup.) These images are inspired in part by Extreme, in part by LA Guns, in part by GnR’s “Patience“, among I’m sure many others in this category that I’m forgetting.
Sadly, the music generation side of AI tech is not yet as good as the visual side. I experimented quite a bit with Suno.ai to see if I could get any passable prototype results to use as audio samples of SYNONYM, but they were so far off the mark that they’re not worth sharing here.
In any event, here is the full final set of SYNONYM pics, and below are some special highlights. Enjoy! Maybe I’ll find a way to incorporate the full untold story of SYNONYM into the AI Lore books in the not too distant (alternate) future (or past).
For whatever it’s worth, I believe these were some of my first tries with Midjourney v6 Alpha. Pretty impressive overall.
Wanted to post a brief reply to this piece on the Verge by journalist Emilia David about a new organization called Fairly Trained, which aims to be sort of a “Fair Trade for AI” if I understand it correctly, offering certifications for AI models trained entirely on licensed data.
The Verge’s headline is, I think, technically inaccurate. It states: “AI models that don’t violate copyright are getting a new certification label.”
They also say this about Fairly Trained, the would-be certifying body:
Fairly Trained claims it will not issue the certification to developers that rely on the fair use argument to train models.
I think this journalist maybe took too much at face value Fairly Trained’s claims about what Fair Use actually is. Their blog post goes a bit further than what’s stated in the Verge. Quoting from that:
[…] this certification will not be awarded to models that rely on a ‘fair use’ copyright exception or similar, which is an indicator that rights-holders haven’t given consent for their work to be used in training.
[Quoting an exec at Universal Music Group] ‘We welcome the launch of the Fairly Trained certification to help companies and creators identify responsible generative AI tools that were trained on lawfully and ethically obtained materials.’
If you have a generative AI model that doesn’t rely on scraping and fair use arguments for its training data, we’d love to hear from you…
My contention with all of this is as simple as it is currently unpopular: anything that qualifies as Fair Use does not constitute a violation of copyright.
Such uses can be done without permission from the copyright owner. In other words, fair use is a defense against a claim of copyright infringement. If your use qualifies as a fair use, then it would not be considered an infringement.
So I think we can separate this announcement by Fairly Trained and the Verge’s coverage of it out into two things:
The claim that Fair Use is a violation of copyright – my understanding is that it is not, and this claim probably doesn’t hold water under scrutiny.
The recognition that creators have a legitimate desire to have greater control than they do under current Fair Use laws, which seem to plainly permit these kinds of uses in AI training.
While taking issue with the first one, I support the second one fully, and agree that we need new radical ways for artists (I hate the word “creators” because it reeks of ‘creating content’ instead of – can’t we all just be artists creating more than just endless ‘content’?) to be able to contribute high quality material to fully licensed data sets where everybody knows what they are getting into, and there are clear mechanisms set up that make sure that artists themselves get directly paid, and not intermediaries like collecting societies in France seeking to change the law in their favor at the likely expense of contributing artists.
I do think there is a place for these kinds of certifications and other allied efforts, but I don’t find it very useful for their purveyors to push seemingly inaccurate legal conceptions. I don’t see who that benefits. We can say we want to change how the law is, or how it ought to be interpreted, but we should also recognize what it actually today says and how it actually has been interpreted in the past. From there, we can point ourselves towards more informed aspiration, and build the realities we want to see one Jira issue at a time…
