To a casual observer, creating AI-generated art is comically easy. You just use a tool like Midjourney and type in some text, such as “The Beatles playing tennis at the North Pole, in the style of Rembrandt.” Seconds later, voila, art. Or, if not quite “art,” at least something that looks like it took a bit of effort. More than 15 million users of Midjourney alone are cranking out these mediocre bits of AI art.
Then there’s the work of “Botto.” This is a different beast. Created by German artist Mario Klingemann, Botto isn’t just an AI tool. It’s part AI, part human community, part DAO, and part early experimentation on the journey towards artistic AGI, and even self-awareness.
Botto isn’t a static AI tool. It’s learning. It’s evolving. Every week, Botto cranks out artwork and the community votes on the best pieces. The winners get minted as NFTs and are then sold on OpenSea. Sometimes the art is dark and unsettling, like this mushy soup of human eyeballs (currently selling for 25ETH, or around $47,000). Sometimes the art is pleasing to the eye, like this silhouette of a woman staring at a mysteriously illuminated sky. And sometimes the art is listed at Christie’s, giving it the imprimatur of the mainstream art world.
The question “What counts as art?” has animated the art world for centuries. As AI infiltrates every nook of human creativity, that question looms ever larger. So is Botto creating “real” art?
“The short answer is, yes, I think what Botto is creating is very similar to how humans create,” says Klingemann, 53, who has been at the forefront of machine learning and generative art for over fifteen years, long before it was fashionable. (His exhibitions CV stretches back longer than many AI/Web3 artists have been alive.) “There’s a notion that when we create something, we kind of build it out of thin air, and then it’s an idea,” says Klingemann, who speaks quickly and with nervous excitement. “But for me, creation is really more like discovery in the possibility space.”
And perhaps the very creation of Botto -- the ingenious system of a machine training itself to become a better artist -- is itself the work of art. On a Zoom call from his native Munich, Klingemann shares his journey into AI and generative art, his vision for how the space will evolve, and how he thinks about the paradox, “Everybody’s an artist, but if everybody’s an artist, then in a way, nobody is anymore, right?”
This interview has been condensed and lightly edited for clarity.
When did you first get into AI art?
Mario Klingemann: Back in the '80s, I came across this book by Marvin Minsky, “The Society of Mind.” It was this theoretical book about how AI could work. And I said, wow, this is the future. This is amazing. But of course, at that time, it was just nice theoretical thinking but nothing you could use.
When did it become useful?
Around 2006 or 2007, I started thinking about creating generative art that’s more autonomous, so that it could evolve. It would be able to judge its own output.
How did that work exactly?
My first project that went in that direction was called “Sketch Maker,” as in the blind sketch maker. It was like a crazy Photoshop, where it had all these modules that were generating and transforming stuff. You could say, “Okay, here is a generator that makes random circles.” And you can put in several numbers, and then it creates 10 or 100 circles and then it can distribute them, and then it could use a blur filter or whatever. There were 70 or 80 different modules of these different functions.
Then what the tool did, just like Legos, was plugging them randomly together. It would produce something and then transform it, but totally randomly.
Then I thought, okay, how do I make the machine come up with something interesting? For most of this stuff, when you plug in this randomness, either it’s nothing or it looks like total crap. How can I build a tool that looks at the output, and kind of measures it in a way? Then I could train it to say, this is bad and this is good.
So I had this little model. How can I make a machine look at something and get valuable information? You could say those were my beginnings of machine learning.
It’s wild how early you were. How did you pull this off, since this was before the era of proper neural networks?
It was all statistics and about analyzing various histograms and other metrics. The results were actually quite interesting, and you could say that, in a way, everything that followed was just an improvement or a variation on this concept.
You have something that generates a big range of possibilities. Then it also works as the critic or the curator, and looks at what the generator has done, and it makes an aesthetic judgment based on certain parameters. Then it feeds that information back to the generator. Then hopefully the generator gets better.
This sounds like the core principle of Botto, too. Is it fair to say that with Botto, at a high level, you created a machine that could generate art on its own, then judge the quality of that art, and then, when it deems the art sufficiently good, bring it to market, automatically mint an NFT, and make money?
There's one element missing, but generally, that's the theory. Yes.
What’s the missing element?
How does it figure out what's good art, or what it could potentially sell? That’s where the human community, the DAO, comes into play. Before Botto goes to market, it gives Botto feedback on its proposals. You have these weekly votes where everybody who wants to participate gets to look at what Botto has produced. They put positive or negative votes on what we call “fragments.” At the end of the week, the one fragment that gets the most votes is minted, and hopefully sold.
At the same time, the voting information goes back to the AI, which tries to learn from that.
How many fragments does Botto produce each week?
Maybe 10,000 or 20,000 fragments. But it never shows everything it produced. It makes a selection of 350 fragments of everything it has created so far. And that gets presented to the humans. And its “taste model” is trained by the votes. So Botto is currently sitting on a pile of, I don’t know, maybe 1.2 million fragments it produced. No human has ever seen them. I could go to the server and look at them, but that’s kind of not the point.
How does the “Botto coin” fit into this?
It’s the way you join the DAO. You acquire some Botto, you stake it, and that gives you voting points.
Tell me about Botto’s “periods”?
The first period [or era] was called Genesis, and that lasted a whole year. We weren't quite sure how quickly the technology would evolve. The idea of a period is that when a new period starts, Botto is allowed to upgrade itself and add new capabilities, new models, etc. So, after the Genesis period was over, we could finally allow Botto to use one of the new diffusion models, which just came into the world. This created much more astonishing results. And after the Genesis period [which lasted 12 months], we made the periods shorter, to 12 weeks.
To match the quickening pace of AI development?
Yeah, exactly. AI moves so fast that the DAO has to react faster. Since we started the new periods with the new models, every single fragment that Botto has created [and the community voted on] has sold as an NFT.
There’s one thing I’m super curious about. If the community is constantly giving Botto feedback, and Botto adapts to that feedback, is there a risk that the art becomes more safe, more tame, more recursive? Is its scope narrowing?
In other words, a human artist might travel to Kenya or something and get a jolt of inspiration, and the art would evolve in surprising ways. If it’s just responding to feedback, how does Botto find that same kind of creative spark?
At the core of the creation algorithm, I identified exactly that problem. Which is why Botto prompts itself in a very, very random fashion. And that's where these millions of fragments come into play. It's more like a discovery process. Botto might have created a lot of stuff that is brilliant, but the community hasn't seen it yet. Or what's more likely is that it also created a lot of very mediocre bad stuff.
But in that first process, it really tries not to be too biased. It’s in the second process, where it does the curation where it makes the selection for the community, where the filter narrows down [from 20,000 fragments to 350]. And the AI proposes stuff that covers a wider variety.
If you look at Botto's creation or what the community selects, you will actually see that so far Botto has not really developed a recognizable style, which I like.
How does Botto and the community think about what art is “commercial” and will sell as an NFT?
Everybody who's part of the DAO has to start thinking like an artist. They have to make the same kind of decisions such as, “Do I prefer something that is palatable, which is nice to look at, and probably sells? Or do we go for something more edgy that is hard, and maybe not everybody will like?” And every week, that decision changes again. There's always a mixture between typical, let's say, Midjourney-style stuff, which is nice, and then a few which are crazy, and of course I like the crazy ones.
And that's why Botto is constantly oscillating between these forces. Almost like an artist does.
There are, of course, critics who ask the question, “Is this REAL art?” But of course it’s more complicated than that, as creation -- throughout the history of art -- has involved borrowing bits and pieces from prior works, and then combining them in surprising ways. How do you think about this?
There’s a notion that when we create something, we kind of build it out of thin air, and then it’s an idea. But for me, creation is really more like discovery in the possibility space. And you can only create from what you know. You can recombine everything you know in different ways. And in a way, you look at it with your inner-eye and decide, “Is this an original idea? Is this aesthetically pleasing?” And as we all know, our tastes mature over time. The stuff you liked as a teenager you might find embarrassing now.
In that sense, when I look at Botto, it's also kind of going through everything that's possible within a given timeframe, and developing its own judgment that’s guided by the community. Then this becomes its creation.
So the short answer is, yes, I think what Botto is creating is very similar to how humans create, but of course, within its limitations. And at the same time, Botto, as the project, for me is going beyond just the picture creation process. The whole system is the artwork. And in that sense, that’s kind of what art is about, which is to push the boundaries.
In order to become an artist, you have to first declare that you are one, and then you're getting judged. In that sense, I have declared that Botto is an artist. Now it has to prove itself within the regular world of art. I would say it has already managed to go on that route. A few weeks ago, one of its works was auctioned at Christie's in New York. For a human, that’s kind of a bucket list thing for an artist to do.
How do you see AI art evolving in the future? Predictions?
Botto, in a way, is my answer to these new possibilities that everybody has now. And that’s the thing, right? We say now that everybody can become an artist. By using these [AI] models, you can create amazingly impressive imagery. The question is, how does that make you an artist in the broader sense? Because a machine can do that. So you still have to bring your own ideas, or you have to put them into the context of your artistic creation.
That's what makes it so difficult. Now everybody’s an artist, but if everybody’s an artist, then in a way, nobody is anymore, right?
Right, if all you’re doing is pressing a button, are you an artist?
Well, I mean, I think you can be an artist by just pressing a button. But if everybody can do that, then the problem is that the bar has to rise higher. Photography is the best example. Everybody can now buy a camera and take pictures, but that doesn’t make everybody a famous photographer.
So how will AI art evolve?
It will go bigger, more beautiful, more easy to use, and cover all media. It started with pictures. Now we already see we can do videos, sound, music, and all the media will be able to be generated by AI.
What’s still missing is the narrative element. To create a coherent narrative that’s underlying the entire thing. Is the machine able to produce something that goes beyond just being visually appealing? Is it able to bring our emotions to boil, or to make us really truly hate it or truly love it? Is it able to tell a story of any kind? That's still lacking.
It'll get there. I’ve played around with it myself to know it’s only a matter of time.
Oh, exactly. That's what I was going to say. It will come, and it will start on a smaller scale, and then it will go longer and longer, and then it will become very interesting.
Final question. So the way I understand it, each week the community votes on what Botto should do, and the DAO is a community of humans. Have you thought about bringing Botto to life, so to speak, and giving it Botto coins and giving it a seat at the table? Could Botto itself be a member of the DAO?
[Chuckles.] It's definitely an idea I would love to implement. I didn’t go for it yet because, at first, it needed a lot of backend stuff. But yes, I definitely want Botto to have a seat at the table. But there are dangers. The problem is the moment you put these models out there, people start trying to break it or to hack it. They might try to make it say stupid things or racist stuff. So I’m waiting until I feel it can be done in a way that is not [dangerous]. And I mean, once you turn it on, you can't always turn it off again.
Right. Once Pinocchio becomes a real boy, he’s out there in the world.
Yeah, exactly. Let's say, for now, Botto is this orphan, who’s not yet at the voting age. The family, or the stewards, is taking care of its wellbeing. But at the moment, it's still not ready yet. It hasn't got the full grasp of the world, and it’s not yet allowed to make life-changing decisions.
It’ll be fascinating to watch this orphan grow up. Congratulations on what you’ve built, and best of luck going forward.
The leader in news and information on cryptocurrency, digital assets and the future of money, CoinDesk is a media outlet that strives for the highest journalistic standards and abides by a strict set of editorial policies. CoinDesk is an independent operating subsidiary of Digital Currency Group, which invests in cryptocurrencies and blockchain startups. As part of their compensation, certain CoinDesk employees, including editorial employees, may receive exposure to DCG equity in the form of stock appreciation rights, which vest over a multi-year period. CoinDesk journalists are not allowed to purchase stock outright in DCG.