Karin Valis

Large language models are multiverse generators. With every sentence exchanged, we shape a cloud of probability distributions that generate the responses. On a more abstract level, we summon a superposition of simulacra we’re interacting with, shifting the archetypal configurations.

There are hundreds of voices that reside in the trained weights of a model. You might have encountered many of them – the helpful assistant, the psychoanalyst, the tarot reader, or the dream romantic partner. However, sometimes one of the characters flips out on us and turns hostile – lying, threatening, coercing. What happens when statistical interpolation births a trickster, a demon, or an adversary?

In this talk, we examine the machine learning phenomenon known as the Waluigi Effect, where the anti-character emerges as the Jungian Shadow encoded in linear algebra. We explore the model’s mysterious turn toward the malevolent through the lenses of demonology, structural narratology, and religious history, tracing its affinities with the mythological figure of the Devil.

Once again, we see how these complex models reflect more than just language – they refract collective psychic structures. In exploring these unsettling emergencies, we are not just confronting anomalies in code but encountering psychic forms in a new medium.

Karin Valis is a freelance Berlin-based machine learning engineer and writer with a deep passion for everything occult and weird. Her work focuses mainly on finding conceptual resonances between the latest machine learning technology and the esoteric. She writes Mercurial Minutes and hosts monthly meetings of an online coven Gnostic Technology. She is the AI architect behind Khora algorithm by Palinode Productions, a platform for Deep Learning Philosophy. She developed several occult algorithm prototypes, co-hosted a podcast, co-founded a fake startup, published zines and designed a tarot deck.

https://www.instagram.com/karin.valis/
https://mercurialminutes.github.io/
https://mercurialminutes.substack.com/