The Shoggoth Factory
I’ve been using Bing Image Creator to make shoggoths.
The exact prompt I’ve been using is something along the lines of “gelatinous shoggoth lurking in sewer. 1920s pulp illustration”. Different adjectives and locations. It’s easy and free to try this yourself and see how it goes.
AI is itself a shoggoth. A complicated pseudo-biological system that mimics human intelligence without actually being alive. At least we hope. At once creepy and banal.
We can’t fully understand what goes on inside a neural network. We do understand though the principle by which relatively simple mathematical components produce complex behaviour on a large enough scale.
Each node in a neural network follows a simple set of rules. The basic math was worked out in the seventies. All they really do is add together a bunch of inputs. We couldn’t build them until now because we didn’t have powerful enough computers to simultaneously model the trillions of operations per second required.
This is not “intelligence”, just an eerie imitation of it. Like a parrot saying words. Whether it could become intelligent - whether a sufficiently powerful computer could write this blog better than I can, just by getting performance up to a few quintillion flops - remains an open question.
Whether such a computer would have a mind - whether it would be sapient, experience qualia, know love - is also an open question. The Chinese Room problem.
It’s helpful to me to think of shoggoths as smart matter. What if every cell in your body was a neuron, and also anything else you wanted it to be? The first eukaryotic life on Earth was Vaalbara, the Dreaming Continent, and shoggoths are just minute extensions of its mind.
AI art gets boring surprisingly quickly. For a week or two I was using Bing every day. I couldn’t wait to burn through my fifteen credits each morning. Then I lost interest. I felt I’d more or less explored the space.
The neural network performs a statistical analysis of all existing digital images. Then it produces new images situated “in between” those images. Latent in the permutation space.
“Complete this permutation space” is a hugely important artistic technique. Especially in games. Try to get every relevant fantasy cliche onto this list of d10 encounters. I’ve always enjoyed the exercise.
AI works for horror because of the uncanny valley. I’m not sure why this is but the eyes of my shoggoths are always cruel and dead. What quality is it that a human artist, even a shit one, adds to an eye? I don’t know.
It can’t model three-dimensional space (hasn’t done enough Loomis) so it can’t build a fight scene. The AI is not capable of inferring from the information it has (a long list of numbers which represent image files) that 3d space exists. It can approximate the rules of perspective for sure but it can’t really position bodies in relation to each other.
If you actually teach yourself to draw you discover that this is the most important thing.
I go to life drawing sometimes. It rules. It’s good to stare at someone and try to understand how the mechanism of the human body works. Joints, muscles, pivots. You have these things as well.
You are a triploblastic life form with a rigid internal structure. A good spooky alien would not be. What if a sponge was alive? What if a fungus? There’s a reason Cthulhu is a mollusc. The cephalopods are the largest and most neurologically advanced invertebrates - they’re like nothing else on Earth.
You could go the other way. What about a quadroblastic life form that thought we were unreasonably floppy?
The AI doesn’t know what a shoggoth is. This is good. It has to make stuff up. There’s no universal consensus on what a shoggoth looks like, no clear set of principles defining how you have to draw one. Nobody has ever dissected a shoggoth to study its anatomy.
It’s not a tiger. It’s formless. The AI will change up its body structure depending on the adjective you give it. Gelatinous, luminous, bull-headed, crystalline.
Other prompts don’t work as well. I tried “tcho-tcho” and it gave me some fucked up little dwarves. Bing has just the right amount of information about shoggoths. Not overfitting, not too loose. You’re never quite sure what you’re going to get.
“Homunculus” came out alright.
I was going to use more shoggoths to illustrate this game but I changed my mind pretty fast.
One, people seem to hate AI. I think of it as just another tool but it’s inherently creepy and I understand why it freaks people out. Of course I’m deliberately trying to freak people out but there’s a point of diminishing returns.
Two, it’s limited in scope. I have about four hundred of these things now and I don’t feel the need to do any more. At least not at the moment. I think all the AI shoggoths are sincerely really good and this has been a successful project, but you’d get bored and pissed off if I kept adding them to every post.
Three, you can make your own shoggoths, just as easily as I can. Probably better. What’s the point of posting images that took me ten seconds to generate when you could just as easily build your own? All I did was experiment with prompts.
What I’ve been doing instead is using Google Image Search (worse than it used to be) and Yandex (good at finding similar-looking images) to scour the web for relevant photographs. This is something RPG blogs have always done.
Setting a game in the real world, and in a specific moment of history, makes it much easier to curate a photo set and create a mood. You don’t learn exactly what the monster looks like, but hopefully you understand the kind of space you’re going to be moving through.
But this is also an algorithmically generated form of art. Google’s search algorithm (and now Yandex, since Google really has gotten worse) scours the full set of all 1920s photographs available online. There’s a technique to figuring out what prompts / search terms give good results.
Picking apart and decontextualising the entire space of human history. All of the people you see in these photos were alive and are now dead.
I think the reason the photographs work better than the shoggoths is they do imply a world. You look at an old picture of a guy picnicking atop the Pyramids and you understand you’re not getting the full story. What did he just say to that tour guide? Was he in the war?
Good art has depth. It has humanity. It gives the impression of being the way it is because a whole human life was lived behind it. Go to the Rijksmuseum and check out any Rembrandt. Look at the eyes.
When we say a painting “captures” someone, what we mean is that the brushstrokes could only exist on canvas in the way they do because of the entire complex procedure of human life and thought that led to the model’s eyes being formed in that particular way.
A mountain’s the way it is because of millions of years of geology. You can’t paint an honest mountain unless you can envision the particular set of crags and curves that mountains are allowed to have. You don’t have to understand the science but you need to be able to feel it in the picture.
The theory of superintelligence is that a sufficiently powerful computer would be able to develop a complete model of the world - geology, anatomy, emotional intelligence - from statistical analysis of patterns in the numbers that make up computer files.
When you frame it that way it sounds absurd. But Donkey Kong is just a bunch of ones and zeroes, added and subtracted very quickly. And this produces the wonderful ape, Donkey Kong! So how can you say where it will end?
I am fortunately only interested in this in the context of designing roleplaying games. I would say I feel in my heart that it’s all fake.
My main conclusion is that you can squeeze cool pictures out of Bing Image Creator, if you turn its limits into strengths. Bold colours, flat compositions. Dense organic environments where it doesn’t matter if the trees are misplaced. Finding just the right level of abstraction in the prompt.
I am honestly not sure why the “1920s pulp illustration” aesthetic works so well. You can use “1920s photograph”, too, although it looks a little plasticky and fake.
With AI art, even shoggoth art, I find it hard to feel the story behind something. I don’t get that sense of wanting to know more about the world. I don’t ask myself how the monster could move, what it could eat, what it would say.
I’m not seeing the skeleton. The muscles underneath the skin. Andrew Loomis says that a drawn head isn’t finished until you can locate the ear on the other side. An AI monster doesn’t have that carefully worked-out anatomy and depth.
Which is fine, it has other virtues. And maybe the computer will learn.
Although I kind of hope it won’t.