This week, I became a patron of the arts.

Matt Round, who feels like one of the last people left doing fun stuff on the internet for its own sake, runs, a website for, well, that. He hosts the world's premier stock photo site for images of crisp sandwiches, ButtyStock, created Scunthorpe Sans, a font that automatically blocks swearing, and also sponsored the toilet paper emoji 🧻.

He's also been playing around with neural networks. Nothing new there – most creative technologists have, because the technology is one of the most incredibly rich artistic tools going right now. Until recently, the most impressive output was This MP Does Not Exist, a roster of headshots of 650 MPs who, yes, don't exist. 

A general facial recognition model, fine-tuned on the real portraits of parliamentarians, the thing I like most about This MP Does Not Exist is how absolutely certain I am that I can accurately guess the political affiliation of every single one.

Look at this guy and tell me he's not a Tory:

On Wednesday, Matt released his latest project: GANksy. GAN stands for generative adversarial network, and is the basic AI technology used in This MP Does Not Exist, and a lot of other generative art. GANksy, as the name suggests, is an AI trained on a library of Banksy artworks, which then generates new art.

But there's two things that stand out for me about the project.

The first is its structure. Rather than generating new images every time a user visits (financially unsustainable for an independent artist hoping for even a modicum of viral success), GANksy is 256 pre-generated images, hand-picked by Matt from the infinite array of possibilities.

That approach is basically the same as was taken for This MP…, but unlike that project, this one has been allowed to keep the structure imposed on it: while you can click through at random, and are indeed encouraged to, all 256 images in the set have been given a name and a number.

They've also been given a price, sort of. Matt is selling each image in the collection to one exclusive buyer. Purchasing it gets you a higher res, "signed" jpg (it has GANksy stamped on it), your name on the page, and the rights to "print/sell/etc". It's yours, you can do what you want with it. And the price is just £1. Plus £1 for every other GANksy that has been sold before you.

There's only 256 images, so this isn't going to make Matt rich even if they all sell, but it's impressive to note that the current going price for a GANksy original, is £46, which, if my maths are correct, means the project has pulled in a grand in a day. That's not bad for a contemporary artist. 

It's also the perfect model for encouraging you to dive in and make the purchase if you are at all interested. Appropriately for a week in which the Nobel Prize in Economics was won by a pair of professors who researched auction mechanics, this system guarantees that the price will only go up, so why not jump in now?

That's part of why I'm the proud owner of my own GANksy original, 10110111: victorious, which I bought for £12.

I'll show it to you in a second, but firstly I need to talk about the other thing I find fascinating about GANksky:

The works look absolutely nothing like Banksy. 

A lot of what we've seen from algorithmic artists, particularly those working with neural networks, has been about mimicry. This MP Does Not Exist is a good example: the project is successful the more the finished pictures look like plausible MPs. If they look like warped mockeries of the human form, it might get chalked up as an interesting failure, a few examples saved for conference talks, but ultimately it's no more the point than a landscape artist tripping and spilling paint thinner over their canvas is pleased with the result.

Mimicry is great if the point of your art is "computers are powerful now", but the next step for many has been scale. One of the first physical artefacts of the neural network age I ever bought was from novelist Robin Sloan. Sloan's Mageframe project used GPT-2 to generate fantasy stories based on prompts from subscribers, then also generated a map based on the story, and automatically printed and posted them as real objects. It was beautiful and weird and digging through my papers now I am terrified I may have thrown out my copy. Shit.

But Mageframe still needed to spit out text similar to its training corpus. Trained on fantasy novels, it was no good if it produced alphabet soup instead. With GANksy, Round has taken the decision to ignore that constraint.

It makes sense: Banksy artworks are hard to generate. They are, at heart, puns. The visual language of a Banksy is only half the story, and reproducing the stencil/spraypaint style of his works while not understanding the semiotics of, say, a man preparing to hurl a bouquet of flowers as though it were a Molotov cocktail, wouldn't be a fake Banksy, it would just be a shit photoshop filter.

So instead, what GANksy appears to have done is just… absolutely fucked up its dumb robot brain trying to find links between Banksy works and decided that really they're all Francis Bacon at heart.

Here's 10110111: victorious

I love it. It terrifies me. I'm going to print it out and hang it on my wall for the three minutes it will take for my partner to ban it from our house.