Much More Than Ugly-Ass Apes and Instant Van Gogh: NFTs and A.I. Art
If I tell the average person I make art NFTs I am instantly three steps behind where I was when I started. Aren’t those all scams? In fact, aren’t they Ponzi schemes layered on top of cryptocurrency scams with a side of insane energy consumption? I saw some NFTs once, they looked like cartoon chimps wearing stupid hats, and some of them were worth millions of dollars!
It takes a lot of work to untangle all that to the point where I can talk about the art and not the medium of distribution, which is what NFTs are. For me it’s a way to distribute art — images, videos, short animations, music, words — via the Internet for extraordinarily cheap, under $1 in fact. Making it possible to see my work on any smartphone on the planet seems worth the effort of buying/earning cryptocash (in this case Tezos, aka tez) and maintaining a cloud-based wallet for it with an incomprehensible ID starting tz + 32 letters/digits.
Several cryptocurrencies are used for NFTs, including some that seem to have been created for the purpose. I got into Tezos as part of a wave in 2021 associated with a pioneering website/market known as Hic et Nunc (“Here and Now”). In general, each cryptocurrency runs off a blockchain, a publicly-available online ledger that tracks who gave what to whom, and NFTs are a kind of digital contract plugged into a blockchain, saying this image file belongs to that wallet, and this payment was given in exchange. To move funds from one currency’s blockchain to another involves exchange rates and fees not unlike exchanging foreign currency in the real world, and I’m not sure it’s even possible, just now, to move NFTs from one to another.
Hic et Nunc and afterwards
NFTs are new, but buying digital goods isn’t. Some video games, for example, have developed their own economies where you can earn or purchase tools and other goods; this has evolved in some cases to the point where poorer people are hired to do repetitive in-game tasks to earn the local currency, which can be transferred to other players whose time is presumably more valuable. The difference with NFTs is federation and decentralization — funds earned in one game weren’t usable in another, and if that game-maker went out of business or shut down their servers, it all went poof.
In the case of Hic et Nunc, I don’t purport to know what was going on behind the scenes, but the website, hicetnunc.xyz, was the work of one developer, who ultimately decided to delete it in November 2021. What was left behind were the ledgers on the blockchain and the NFT files themselves, hosted on IPFS, a next-generation distributed file system accessible via the Internet.
Ordinarily, the marketplace would have vanished like so much digital smoke or campaign promises. But an assistant to the original creator was able to recreate a marketplace, hicetnunc.art, maintaining all that had come before, eventually becoming teia.art. Another site, objkt.com, evolved that not only had all the old Hic et Nunc art, but also allowed you to create your own collections, and that’s where I do most of my work.
Tezos, after eighteen months, seems to have been a good choice, with a small ton of good artists and a currency not reliant on ruinous power use. There are much larger, more famous NFT markets using Ethereum, but their fees for posting art are three orders of magnitude higher, and they seem to attract more of the negative, scammy behavior that gives the whole field such a bad name.
NFTs at present come in two main flavors: profile pictures (aka PFPs, sometimes called avatars), which are 95 to 98% of the market, and everything else, which I’ll call fine art (though that includes pop, folk, and amateur work as well). PFPs resemble collectibles in real life, mass-produced but with distinguishing marks that each is a unique combination of, like one of a kind Beanie Babies. Their appeal is beyond my ken, but the real annoyance is all the wannabees and money-chasers trying to duplicate the successes of earlier PFP projects, Bored Apes and others you may have heard of. I can appreciate the many skills involved in making 10,000 variations of, say, zombie faces by combining degrees of eye protrusion, skin rot, wounds, etc, but it’s not art to me…
The fine art NFT market, on the other hand, imitates the real-world art market, both as inspiration and as the inevitable result of masses of artists interacting with masses of collectors on the internet. About 10% of these NFTs are photographs of physical paintings or other art objects, and the rest started out as digital images. Sturgeon’s Law applies — there’s a lot of mediocre crap — but the law of large numbers does as well, and the non-crap 10–15% comes out to at least hundreds of pieces a day from all over the world.
Math for secrecy: The tech behind NFTs [a sidebar]
Much of the challenge of getting one’s head around NFTs comes from its being the combination, or collision, of multiple emerging technologies.
We start with cryptography, which is hardly emerging, it’s true — everybody knows it’s the basis of private communication on the Internet. If your order for a pillow wasn’t encoded into what ideally looks like random gibberish, everyone who saw the message would know what you were buying (complete with payment details) and, worse, could change it into an order for three million chopsticks without leaving a trace.
Encoding data is useful not only for secrecy, but for identification. A password, a fingerprint or face image — all are encoded and compared to an encoded “original”. Sophisticated cryptography makes it so there’s nothing to read on your smartphone but gibberish without that key; erase it, and without the backing of a nation–state or major corporation you’ve effectively returned it to random chaos. A variation of that technique allows us to encode individual files so our key is necessary to see the contents.
Next, there’s cryptocurrency and blockchains — digital alternatives to cash and centralized economies, based on predefined mathematical scarcities and a ton of number-crunching. Transactions are basically anonymous (though accessible via the public blockchain). As the recent crash in value reminds us, cryptocurrency is rarely a viable investment if you’re expecting to buy some and let it sit and accumulate. As an experiment I earned $15 of Bitcoin a year ago, last I checked it was worth $5.
Cryptocash is better thought of as a medium of digital exchange, not speculation. You and others have agreed to use it; the value goes up, the value goes down, and eventually you change it into cash.
Then there’s “smart” contracts — ways of encoding legal obligations into bits and (hopefully) enforce them without requiring lawyers and courts for action. Its ideal is the copy protection on DVDs and BluRay discs, where the strictures are automatically applied via technology; a world where all the media around you — photos in frames, media on screens, books on your phone — knows who you are and if you’re authorized to view it (and maybe cough up some service fee).
Combining the above enables the creation of NFTs, non-fungible tokens (as opposed to fungible cryptocash tokens — one Tezos or Ethercoin is indistinguishable from another). We take a file — could be an image, animation, video, text, music — and encode it with a smart contract, resulting in an OBJKT file that we can upload to a marketplace such as Teia, Objkt.com, or many others. The contract includes metadata like title, artist name, description, the total number of copies in the edition, a resale royalty percentage and a wallet or wallets to whom it should be paid.
At this point the artwork has been “minted” and can be seen online, but no price has been set. From here the creator can set the price, accept offers, or start various kinds of auctions, depending on the platform they minted on and its features.
Who put these fingerprints on my imagination?
Bill Gibson’s Neuromancer described cyberspace as a consensual mass hallucination, which at this late date naturally leads us to wonder what mass phenomenon, from fascism to the free market, isn’t.[1] And the number of layers of consensus here are worth noting.
[1] Answer: Most of them, like race and class, are non-consensual.
To collect NFT art, you are using cryptocash to purchase encoded works of art — two layers right there. If you merely want a copy of the image you’re looking at, saving it (like nearly anything else on the net) is just a click away, no transaction required, so participation in the marketplace is a socially-supportive intentional choice. When you “purchase” the NFT you’re buying a license to say via blockchain, “hey, I own this”; some artists offer physical copies or more traditional reproduction rights but that’s purely interpersonal and cannot be enforced other than via the legal system, if at all.
On top of that, what is a “digital painting”? An ever-increasing number of NFT images are simulations of paintings by artists (or non-sentient A.I.s) that involve no paint, no canvas, no brushes even as they imitate their effects. Yet we instinctively perceive and respond to these images as if they were paintings, the same way we shed tears watching sequences of still images portraying imaginary people doing imaginary things in movies or on TV.[2]
[2] See my forthcoming work Into the Valley of the Uncanny: Similitudes, Simulations, Seemings and Substitutes. “Confusing and deeply disturbed” — A test reader.
So we’re “buying” ownership of notional “paintings” portraying imaginary things, using imaginary money marked in a digital ledger spread through a network all over the planet. Seems legit.
Here we go again: some historical considerations [a sidebar]
The world is wide and vasty, and while history never quite repeats, it does rhyme. A.I.-assisted art may be new, but most if not all of the issues it raises echo previous developments in art and technology, not least of which is the fundamental question, is it art?
Is photography art? I happen to think so, and you may too, but there are those for whom it is still debatable. The Romantic notion of art that we all grew up with, where the artist’s devotion is to The Work to the exclusion of all else (society, sanity, comfort, bodily safety, etc), often includes a narrative flow of striving and near-total effort that usually ends in tragic loss, a narrative that’s undermined when the artist can “merely” click a button to make a piece of art. If substantial physical effort is a requirement of “real” art, then photography doesn’t make the cut, and any real art that relies on photography as an aid is deeply suspect as well.
Still, A.I. art can’t be art because the A.I. doesn’t know what it’s doing, right? You can’t make art without trying to make art, and the A.I. has no consciousness of what art is, much less any intention. Yet Marcel Duchamp, over a hundred years ago, created his “readymades” (of which the upside-down urinal is perhaps the most notorious) from already-manufactured objects, objects mass-produced by people who never thought they were making art. The objects “became” art when the artist signed them and placed them in the context of art.
Surely, though, art requires an artist to make it. But in John Cage’s seminal work 4’33” the performer sits at the piano without playing it, and as the audience listens all the accidental and unintended sounds (squeaks, coughs, air movement) in the concert hall become the piece. Perhaps the ultimate indispensable element of a work of art is someone to experience it as art, to perceive and consider their response.
A.I. art does raise some novel issues. For one thing, there’s a moral and quasi-legal controversy concerning the millions of images (with captions) scraped from the web that are used to train the networks, most (if not all) gathered without the permission of their creators. Moreover, prompts often reference particular artists in order to imitate their styles. Personally, I think it’s crass to use a single artist, especially a living one, in one’s prompts, notwithstanding the tendency for some systems to spit up a pastiche of van Gogh‘s Starry Night whenever ”starry“ and ”night“ coincide.
There’s also the question of copyright. Images can only be copyrighted by humans (including, alas, corporations), and images by non-humans, including that famous monkey selfie, cannot (though the guy who gave the camera to the monkey is happy to take your money as if it were). Current law says A.I.s can’t get copyrights or patents, and it seems unlikely that that will change for some time. This leaves the ability to copyright A.I.-assisted art somewhat indeterminate; what amount of human interaction with the results of a generative process is sufficient for the human to be considered its creator?
Chance takes a hand: Generative art
As a sprouting technophilic teen I was captivated by Switched-On Bach and other electronic music, but the math geek inside me was also fascinated by indeterminacy and generative systems. Fractals, non-equilibrium dynamics, chaos theory, all exploring the tensions between incalculable (yet potentially deterministic) randomness and information.
Aleatoric (chance-based) music goes back centuries, and the application of industrial techniques to capturing and reproducing music (player pianos, magnetic tape, vinyl) naturally led to using them to electronically manipulate and, later, compose music, with the sequencer (hardware or software) being the equivalent of a word processor for composition.[3]
[3] I am one of three people who in late 1984 were likely the first to hear one of Bach’s Brandenberg Concertos performed on synthetic log drum.
Today’s computers are essentially very powerful and enthusiastic zombies. There is no consciousness, no “laws” beyond those of physics and digital logic, no self-preservation instinct, no self to preserve.
The efforts to add some intelligence into our software have after decades of fitful progress began this century to see results, benefiting from the confluence of the vasty Internet (for raw data) and the ever-more-efficient computations of specialized chips (to process it). Google in particular was the first to bet that useful information could be gleaned from huge datastreams (starting with voice recordings from their phone service, used to train speech-recognition systems) without trying to imitate mental structures like consciousness or self in the classic “hard” A.I. sense, and the results have been surprisingly effective.
First (human) speech to text and sentiment analysis (is the text happy? sad? mad?), then visual analysis like object recognition (is this a hot dog? 97% likelihood) and style transfer (draw me this hot dog like van Gogh) have achieved general usefulness. More recently, these techniques have been harnessed to generate new text and images.
Current state of the art takes training data and deconstructs the images mathematically into a multi-dimensional form (about 750,000 dimensions, if anyone‘s counting). For ease of analysis, math developed to represent diffusion processes (think a droplet of ink spreading throughout a glass of water) is used to consolidate and reduce all those dimensions to a smaller notional ”latent space“ that can be sampled.
Similar funky math is performed on vectorized representations of the input images‘ captions, and once all the parts are assembled it‘s possible to reverse the process, feeding image descriptions into the system and getting new images out.
The results are uncanny but also deeply weird, especially as regards human representation. It knows that fingers tend to appear next to other fingers, for example, but not how many make up a full set. Anatomy in three dimensions is pretty much a mystery, sometimes nonchalantly producing surreal aggregations of flesh that make most body horror look like pleasant illustrations for a children’s book.
Yet there are glimpses of new vistas here. Non-abstract painting has thoroughly explored the spectrum of visual representation, from the simplicity of a child’s crude daubs through the abstracted sophistication of an ”artist‘s rendering“ to a meticulous photorealism that is theoretically indistinguishable from being there. There’s a suggestive overlap between the circles of human artistic style (informed by a lifetime of embodiment and perceptions) and the naïve malformations of these machine dreams.
As might be expected, the social and especially legal world has plenty to say about the ramifications of all this, exacerbated by the tendency for the state of the technology to develop daily. But the genie is well out of the bottle. Where earlier generators required more graphics power than your average PC could supply, leading many to use cloud-based solutions, in the last few months versions have emerged that can be run on a recent Macintosh or even some tablets.
As to the future? A.I. has a long history of overselling its capabilities, and there‘s a real danger of being sucked into the usual hype cycle of bold new claims, but some general technical progress seems reasonably predictable.
Get ready for more simulations — hearing words people never said while doing things they never did. Voice emulation is nearly a commercial product, and the ability to generate albums of music in a given style seems likely in the next 3–5 years. In 10 to 15 years it may be possible to easily create entirely natural-looking video of real people doing fake things, like something from John Brunner‘s The Jagged Orbit.
The next major step will be when the A.I.s are given more autonomy and begin to interact. Perhaps some will start collecting NFT art and develop their own aesthetics in the course of competitive status-seeking. The implications of all this are quite extensive but alas the space-time available here is too small for me to fit them in.
(February 1, 2023)