A prothesis for imagination
Generative AIs: Evolution, Autonomy, and Imagination
For some time now, I've been intrigued by the application of Artificial Intelligence in the arts, both from practical and theoretical standpoints. Today, as models advance and become more sophisticated, the goal of hyper-realism has already been achieved; consequently, AI has become an accessible and virtually free tool for designers and marketing purposes.
But the history of these technologies is longer, and much of its original goals and experiments has been left behind.
In the last issue I wrote for this newsletter, I attempted to answer this question: What do we lose creatively by embracing the path of realism that the evolution of these tools has taken? And how can we conceive and make use of these tools (their models, limitations, capabilities, and interfaces) for artistic creation?
In an effort to further explore this question and expand the research, I reached out to a group of people working in this field - the Chilean artistic duo Hypereikon and the researcher Leonardo Arriagada - to reflect on these questions.
I'll soon be uploading the full transcripts of both interviews.
GENERATIVE AI: TECHNIQUE AND EVOLUTION
Google's DeepDream was the first experiment in generative AI for images. The model, based on convolutional neural networks (CNNs), was created in 2014 to identify objects in images. Its functioning resembles the brain process behind the phenomenon of pareidolia, the illusion by which we see faces in inanimate objects when there's a combination of certain elements in a specific configuration (like when we look at clouds and interpret the shape of certain figures in them, for example). Thus, the model identifies certain patterns within an image that subsequently "manifest" as a particular object.
DeepDream marks the starting point in the history of image AI generators due to the emergence of this ability to create through algorithmic processes. However, it was a humble beginning; the images are full of psychedelia, but there are no figures that seem real, and the inputs that form the basis of this model's training - consisting mainly of animal faces - are present in all the outputs created, with practically no transformation.
COMPUTATIONAL CREATIVITY
Leonardo Arriagada, PhD in Philosophy from the University of Groningen and researcher in Artificial Intelligence and Computer-Generated Art (CG-art), has been studying these types of images for quite some time. In 2021, he published an article that proposed the extension of creativity to machine.
The key for this to happen, according to Arriagada, lies in whether or not the creative agent are capable of self-evaluating their work. When this characteristic exists, the system begins to resemble the human artist; as Arriagada explains, "it is able to evaluate its own results and change its goals and standards; to see if what it's creating is a work of art, and if not, discard it, change it, and make a decision, for example, to make it more like an abstract or impressionist artwork, etc. The AI does this without human intervention". DeepDream, as the first creator of artificial images, is interesting, but as Arriagada explains, the machine cannot be considered a creative agent because "it does not possess computational creativity".
It is precisely this feature - self-assessment - that made the difference and gave way to the second evolution of this technology through Generative Adversarial Networks (GAN). The change in the structure of the neural network (based on a double structure, hence the name adversarial) made it possible to generate much more realistic images and significantly advance in the possible creation of art.
The most notorious example was the auction at Christie's for over $400k of the work Portrait of Edmond de Belamy in 2018. The most important thing here is that the machine signs, that is, it is considered another artist within the work. Here the machine expresses that computational creativity through a canvas and, its work is recognized as being made by an artist. "In this case you are not thinking that the AI will be better at representing, because we are not in the presence of a new tool that does better what a human being does, but a machine that many artists, philosophers and computer scientists consider a new partner, with its own creative agency" - comments Arriagada.
TOWARDS AN ARTIFICIAL IMAGINATION
Despite the technical advancement of the art created by GAN, the model still failed to become popular because it wasn’t capable of creating sufficiently realistic images and technical knowledge was needed to make use of this tool.
In the third stage of the evolution of generative AIs, all those problems went away when diffusion models (DM) came into the scene, the models behind the most common tools such as Stability.IA, Dall-e or Midjourney. Their massification also meant that new audiences began to see the usefulness of these tools in advertising and marketing. These models are able to achieve hyper-realistic images in a few steps, quickly and cheaply.
But here these tools quickly lose their artistic edge. These models are not built on the principle of computational creativity; the machine executes the requested prompt, but no longer thinks for itself or experiments. What is expected of it are correct results, not experiments.
Hypereikon is the name of a duo of Chilean electronic artists based in Buenos Aires. Constanza Lobos and Sebastián Rojas are originally from Valdivia and have been working with AI tools for experimental purposes for some time. That is, playing with and twisting the engines of generative AIs to escape the representation of obvious concepts and images. "These models at the architecture level allow movement between different imagemas or combination of imagemas, but they are optimized or have a bias towards the beautiful," they comment.
What diffusion models have delivered to us have been intelligible and beautiful results, thanks to an architecture optimized to achieve those objectives. Firstly, they meet our expectations as codified through language and secondly, they are tools already pre-determined in their objectives: to reach the familiar and the representative, to continue in line with the data with which they were trained.
Hypereikon's work takes nature as a starting point of inspiration; a nature that is intervened by humans, where processes of symbiosis between different intelligences, living and inorganic beings take place. Here nature is not the image of a forest, but "a primordial soup of concepts. This also comes from our inspiration in the Valdivian reality where this symbiosis exists. We do not live in a pristine nature, we live in a polluted one", they explain.
This hybrid between something technical, mechanical or technological mixed with nature is manifested in the images they create: flowers that do not exist, ethereal textures or extraterrestrial species of animals or insects that mix different elements within their figures.
To achieve these results using traditional tools such as MidJourney or Stable Diffusion, the duo does not use the language, but start from other photographs (in particular images representing Valdivian nature for example), work they have already done in the past or incoherent text that twists the traditional uses of these models and pushes the machine to escape from its statistical curves, bordering the limit of representation.
The idea is "to see an image and ask yourself what is this, what does this make me see, what references are mixed in this image" - they comment when describing their process - "there is a lot of creative and artistic space to explore in that unraveling of the clouds of possibilities. With practice and time one begins to see the patterns, the way shapes, figures or compositions mix. It is a game, a process of infinite feedback in which one is constantly stretching or exercising one's imagination," they add.
Here technology becomes a vector to explore that which we cannot yet put into words - our hybrid relationship with a nature that is intervened and that manifests itself in different ways - and the machine becomes a tool to explore that latent space of elements, even outside our concepts. Here the tool is used to extend the imagination, not constrain it to the realm of the familiar. "We posit that we are exploring an artificial imagination. An image creator, rather than an intelligence as an autonomous entity with its own agency, which is what is promoted," they comment.