K Allado-McDowell co-leads the Artist + Machine Intelligence program at Google Research, facilitating collaboration between Google AI researchers, artists, and cultural institutions. Alongside twenty years of experience in software design and engineering, K also has an MFA in photography and regularly performs acoustic and electronic music. In this exchange, we speak about the relation between divination practices and machine learning, multidimensionality of neural nets as an alternate thinking schema, as well as the role of predictive algorithms in constructing the plausible instead of the possible.
[DI]: How did you become interested in esotericism and divination practices?
[KAM]: I became interested in esotericism and divination as a result of studying esoteric physiology. In my early twenties, I had experiences doing hatha yoga that forced me to learn about the energy maps of the body that exist in Vedic philosophy. As a result, I spent decades obsessed with meditation, yoga, and esoteric knowledge systems from Asia, South America, and Europe. Nondual systems like Kashmir Shaivism, Dzogchen Buddhism, or Taoism most accurately describe the experiences I've had. I frequently use the I Ching for reflection on important decisions I make in my daily life. Working with plants is also an important path for me.
[DI]: How did you begin working with Machine Learning and what brought you to Artists and Machine Intelligence at Google?
[KAM]: I came into my current role through a conventional path that became strangely serendipitous. I was working as a user experience prototyper on a design team embedded in AI Research. I worked on speculative product prototypes that used AI, and helped formulate models of AI interaction in products. When trippysquirrel.jpg escaped onto the internet mid-2015, Google released DeepDream, and my director Blaise Aguera y Arcas formed a group to exhibit work made with DeepDream. As the only person with an MFA in a then 150-person research team, I volunteered to lead the group. We converted a one-night event budget into an informal artist grant and collaboration program, and officially launched the Artists + Machine Intelligence program in February of 2016.
[DI]: With regards to esotericism and practices of divination, such practices often have an anticipatory relation to the future, becoming methods of foresight. At the same time, Machine Learning(ML) is overwhelmingly used for the purposes of prediction (predictive text, predictive policing, targeted advertising etc.) how do you see the relation between these modes of anticipation?
[KAM]: Older divination systems are based on cosmologically foundational symbols within a culture. For example, we can look to the yin and yang lines of the hexagrams that govern the I Ching and the eight elemental trigrams they compose, or the archetypal symbols of the Tarot. To work, these symbols must lock into underlying patterns of nature or the mind. Current predictive systems for policing, advertising, co-writing, etc. are trained on data sets arising from social systems positioned within a singular historical milieu: modernism, with its colonial roots. With several centuries or millennia of integrative practice, new ML systems might absorb the foundational symbolic potency of older divination methods.
[DI]: There is also something about interpretation, legibility and agency that differs a lot between older divination methods and current predictive systems. I find that in older divination practices the level of interpretation can be performed both by the reader and the person who is being read, and the divination method is a system in place to structure that interaction. Agency in that case is distributed along all things and people partaking the divination. On the other hand, in current predictive systems the receiver of a prediction is not afforded the chance to do the interpretation; leaving them in a position of little if any agency. Do you agree with that position? Do you think that there is a way of overcoming that logic
[KAM]: Of course it’s difficult to generalize, but you are correct about the role of interpretation and introspection in readings. Working with the I Ching requires an absorption of Chinese philosophy that reshapes the reader through introspective application of the hexagram and commentary. This process bears fruit outside of the specific question addressed by the oracle as one becomes attuned to Taoist and Confucian ways of approaching problems. This is as much a part of the process as receiving an answer to a question. Although, the I Ching has also given me great stock tips.
These older systems and their integration of inner perception and outer manifestation are impossible according to materialism. In a materialist view, an oracular reading is random or purely a psychological projection. The causal power of a predictive policing system is much more clearly mappable in contrast, and its creators are our contemporaries, not our ancestors. These might be the deepest rifts between divinatory and predictive systems and a very good reason to distinguish them from each other.
[DI]: You have stated before that “In the 21st century art, technology and spirituality will merge, it is just really messy right now.” In which way do you think that these will merge and how do you understand the role of corporate interests that lead technological development being carried forward into this merging?
[KAM]: We can read spirituality, art, and technology, as ways of mediating relations with the unknown or unknowable. Spirituality actively seeks it out, art translates it into form, and technology (at least in the American west coast formulation that dominates digital platforms now) extracts an infrastructure from cultural experimentation. The social and conceptual experiments that gave rise to the 1960s counterculture, and ultimately Silicon Valley, are an example of this process.
This translation of the unknown into infrastructure through cultural experimentation is not well articulated or understood. The resulting ambiguity has allowed for conflation of engineering and problem solving with revolutionary social good. Such identification has animated Silicon Valley for decades, despite the obvious blind spots it produces. Tech is now approaching health, consciousness and pharmacology as areas for developing social infrastructure. This infrastructure could become an apparatus of deep control or a platform for expansion of human consciousness and potential. Whether this plays out as integration or extraction will have everything to do with the role of women, indigenous, and earth-centered people and cosmologies in that process. This is critically important.
[DI]: I would think that it can be an apparatus of deep control as well as a platform for expansion of human consciousness, simultaneously. In some ways it already is both in the sense that large tech firms operate as apparatuses, exercising control over various situations, such as the Facebook–Cambridge Analytica scandal, yet technological infrastructure has already become an extension of the thinking of many people.
[KAM]: Yes, it is and will be both, and will alternate between modes of extraction and integration, as the internet does.
[DI]: What is an example of integration or extraction? I wonder whether any system that aims to integrate all epistemologies doesn’t at the same time transform them, echoing colonial practices.
[KAM]: Precisely, because agency is still centered in the platform. The appropriation of technical platforms and programs by a local perspective and through a local cosmology (after Yuk Hui’s cosmotechnics) is one way to imagine an anti-colonial alternative.
[DI]: I find very interesting the way you have spoken about machine learning in relation to multidimensionality and hallucination. It seems to me that you are suggesting machine learning can propose a paradigm or rather a schema which will allow humans to incorporate this multidimensionality in their thinking. Could you expand on the relation between machine learning and multidimensionality? Do you see neural nets as a way of overcoming the binaristic basis of computing? Also how do you see multidimensionality as a form of cognition?
[KAM]: The underlying geometries of our technologies become subconscious influences on our lives. In the 90s Marshall McLuhan's axiom "the medium is the message" was a mantra for evangelists of the network. Now networks profoundly influence our economy, elections, global political movements, aesthetic trends, etc. providing a metastructure that conditions everything. I see high-dimensional machine learning models poised to perform a similar transformation on culture.
Works generated with ML hallucination can be described as movements through the latent or high-dimensional space of a neural net. Where a 3D spatial system has dimensions for x, y, and z, neural nets have much higher degrees of dimensionality, which can be imagined as a type of space. For example, an image recognition system might have hundreds of thousands of “axes”, and an address could be generated with a location on each one.
I began to develop a felt sense of these high-dimensional spaces when viewing visual artworks made with AI. But the same sense can be generated around non-visual data. When I saw a mapping of high-dimensional survey data about gender identification and expression created by my director at Google AI, Blaise Aguera y Arcas, it became clear to me that high-dimensional systems could help us experience ourselves outside of restrictive binaristic categories of identity. A binaristic model allows two choices and a spectral model allows for a continuum. But models that take into account many aspects of gender expression (sartorial, behavioral, linguistic, physiological, etc.) can be mapped in high-dimensional space then visualized in two and three dimensions using algorithms like UMAP or t-SNE, to reveal a complex space of possible gender expression that doesn’t conform to a simple binary or even a continuum.
In general, I would characterize multidimensional cognition as thinking with patterns rather than data points. This pattern-based cognition scales to very high levels through organic methods like meditation or co-cognition with other living intelligences like plants, animals, ecosystems, and planetary and star systems. Ideally, can reflect this to ourselves through machine intelligence.
[DI]: Could you give me an example of a work with ML hallucination that had an impact on you?
[KAM]: Jenna Sutela’s work nimiia cétiï hallucinates language through the motions of a slime mold controlling a neural net. This type of interaction between human and non-human intelligence, while very avant-garde in style and presentation, points to an opening with broad social implications. If we can think of our world as one inhabited by many forms of non-human intelligence, we may begin to see the living beings that surround us and we may start to respect their intelligence. If we can’t perform this simple act of seeing, we are too myopic to responsibly create synthetic non-human intelligence.
[DI]: Hallucination seems to be a theme that you return to, both through your work in Artists and Machine Intelligence, but also you have spoken about it in the context of hallucinogenic drugs. Do you consider hallucination as a means of re-arranging the normative limits of perception?
[KAM]: Current neuroscience shows that psychedelic hallucinations produce greater degrees of connection between neurons, an increase in dimensionality. This type of hallucination is clearly a means of re-arranging perception, and the neuroplasticity it generates has proven to be healing for disease like PTSD.
[DI]: Why is hallucination a prefered means? I am also thinking of the increasing use of hallucinogenic drugs among tech-workers, either by microdosing LSD at work or taking ayahuasca to reach higher levels of ‘innovation’ or ‘creativity’.
[KAM]: The simple introduction of greater degrees of neuroplasticity will produce creativity or innovation. But this refactoring of indigenous practices into usable strategies for platform capitalism is at the heart of the extractive relationship Silicon Valley has with counterculture and the “others” it mediates into mainstream culture (like Eastern meditation practitioners or South American curanderos). This connection between colonized peoples and the center of Western technical dominance can be another point of extraction or it can be a lever into the hearts and minds of the people creating global infrastructure, depending on which voices are centered.
[DI]: Listening to one of your talks you expressed that it is your “responsibility to study magic because it is the oldest, more established tool to deal with the universe as it really is.” Could you expand on why do you think magic is an important area to study within the current socio-political structures?
[KAM]: We need to radically reframe our relationship with Earth if we are to survive the effects of our technologies. Part of the matrix of belief that limits our perception are Christian and colonial notions of time and futurity that inform modernism and rational materialism. Jailbreaking our minds from these limiting fundamental beliefs is a good first step toward empathizing with subjects excluded from the global power structure. For example, one common side effect of engaging magical systems is an increased understanding of our ancestral relationships and the perception of ourselves as ancestors. This disconnection from our own ancestry and ancestorhood makes it possible for us to consume resources the way we do.
[DI]: How does your study of magic inform your thinking of machine learning and more broadly AI?
[KAM]: Seeing symbolic and predictive systems as potent, empowered, extensions of mind into matter brings a gravity and sense of responsibility to one's work. If we take seriously the idea that our intelligences engage with a living, enchanted, planet, society, and cosmos, we can find a sacred duty in our work with technology. Ethics are an important first step, as they open the door to a deeper consideration of the nature of our technical heritage. But we can go further. What if technology were a prayer?
[DI]: In my understanding of ML as a form of computing that evolved out of statistics, its primary focus is to identity the probable, the plausible. I wonder how ML’s ubiquitous use in processes that vary in scale and scope—from assigning the amount of refugees in each European country, to suggesting text for this email I am writing to you—constructs the plausible rather than the possible. Is there a way that ML does not become a means of instantiating the plausible according a Western view of the world with its taxonomies and stereotypes?
[KAM]: This is a great question. From a cultural point of view, we can begin by expanding the perspectives and epistemologies that feed our supervised ML systems, and who it is that creates them. This basic first step would mark an important and necessary change, and requires critical consideration of the processes and power dynamics involved in creating ML.
From a technical point of view, I see you describing the closed aspect of what are called supervised learning systems. Supervised learning systems are trained on labelled data sets; they can only know what we've told them, and we can only tell them what we can know and express in written language. Reinforcement learning systems, on the other hand, are goal-driven, and derive functions for a system to achieve an outcome. This, to my mind, promises a closer mapping of intelligence as such. To put it in esoteric terms, does consciousness exist in a taxonomy? Or does it exist in the world? And if it exists in the world, what is required for intelligence to absorb consciousness into itself?