Tarot of the Latent Spaces by Hermetechnics Inc.
Most of my content consists of verbose opinions on esoteric books, fun mythological factoids and general roaming through piles of related notes. But the fact is, my real job is writing code and playing with machine learning models. Some time ago, Fiala and I combined these tech-bro skills with our weird hobbies and founded a fake Silicon Valley startup called Hermetechnics Inc. We engage in quite a wide range of activities - workshops, talks, exhibitions, networking… Our current focus is on a project called Tarot of the Latent Spaces - machine learning generated tarot. As many people ask for updates on this front, I would like to give you a brief overview of our research.
I’ll start with some machine learning background. You've probably seen these StyleGAN models able to recognise a picture of a dog. The model is trained on many different photos of dogs: all sizes, shapes and breeds. It has its internal representation of what this concept stands for - tails and ears, fur, and other stuff. I like to think about this representation as to the essence of ‘dogness’ translated to machine language (mostly mathematical matrices).
You can ask the trained model to point to a dog when they see one. But you can also ask what this thing labelled ‘dog’ means for them - and ask them to generate a picture of one for you. Then, by gently tweaking the seed given to the model, we can swim in the sea of their internal representations, so-called latent space - move one way, the generated animal grows in size, move in the other, and the fur gets darker.1 This is the way you get all those eerie pictures of faces and animals that never existed but resembled reality in a very uncanny way.
Extracting the Archetypes
So, how do we connect all this with the promised Tarot? We can train the model in a very similar way, but instead of feeding it pictures of animals, we can try many different designs of a Tarot trump. This way, we can try and distil the essence of a specific major arcanum. We show it hundred distinct pictures of The Moon, and we ask the model: what do you see?
You can see four possible answers in the picture above. These are random samples from our model of The Moon card. How well do you think the archetype is preserved? They seem to have it all - the shadowy landscape, mysterious animals creeping on the damp floor, gloomy sky and, of course, the big round full moon silently hanging over the horizon. We removed all the symbols and semantic clues that are intentionally placed on the card by an artist. The machine finds the features by itself, yet we can still clearly recognise the mood and feel of the moon.
Not only we shall see the tarot trump through the eyes of the machine, but we have a continuous model with infinite representations of it. This means we can generate any number of different card designs, enabling us to create a completely unique tarot deck for anyone. We can also reel out beautiful audiovisual readings; you should be able to see them soon on The Future for Today or in our snippet for MonomerTV as 0e20f2.2
The deck is finished and the first test print is out of stock! You can follow us on our Tarot of the Latent Spaces Instagram for more updates
Check out some sneak peeks in the article below and drop me a message in case you want to secure yourself a copy
In case you have any questions, technical or non-technical, feel free to reach out. ✨
This ‘tweaking’ is not so explicit because the mathematical model doesn’t think in the same categories as we do; you can see this in action on some cars.
The generative video is juxtaposed with other footage.