feral.ai1 is a research-creation project that imagines agents of Black techné2—a tradition of radical Black aesthetics and freedom that disrupts binary categories— turned feral. As such feral.ai embodies B. Coleman’s (2021) appeal for imagining an AI “that can be free—if not to imagine, then to generate—speeding through possibilities, junctures that are idiotic until they are not.”

In the piece Technology of the Surround Coleman (2021) describes how the learnings from histories of Marronage3 applied to the latest advancements in machine learning allow the technology unmoored to become a generative tool of liberation. While inspiring on a conceptual level I could not help but wonder how I may mobilize this idea of wilder AI’s into practice? In other words, what may a Maroon’s technical practice look like?

Such a critical approach to technical practice is in dire need when trying to address the polycrisis the world is facing. Drawing upon the work of Spanish philosopher José Ortega y Gasset, UKAI Project’s Carnival theme of Shipwreck4 critically questioned what might happen when there is no savior, no rescue, only—Shipwreck at the end of the world? There could not have been a more fitting context to premiere the prototype version of feral.ai as a test environment for a Maroon’s technical practice.

I believe that histories of Marronage will aid us in navigating shipwreck. The French etymology of the word “maroon” carries a general sense of being abandoned without resources. Consequently, drawing upon the diverse practices of Marronage the speculative work of feral.ai not only seeks liberation from the oppressive and destructive forces of what Subhabrata Bobby Banerjee (2008) calls necrocapitalism but through it explores what alternative realities the end of the world may afford. What kind of freedom does Shipwreck allow for when we know that no one is coming to save us and all we have left is our swimming strokes keeping us afloat.

feral.ai explores these questions first technically, then conceptually and lastly physically. In the next few paragraphs I outline the development of the project and through it potential facets of a Maroon’s Technical Practice.

Affordances for Technical Liberation

The recent surge in generative AI exemplifies the many facets of necrocapitalism. From the extractive mining practices underpinning digital infrastructures to the spiraling energy demands of data centers, generative AI is driving corporations and governments to abandon questions of environmental sustainability, climate commitments and even revive nuclear energy. The exploitative labor required to sustain these technologies—from underpaid click work and unpaid digital labor to the mass harvesting of copyright-protected material for AI training—shows the extensive human toll generative AI is taking. This doesn’t even account for the inherent biases, opacity, and lack of accountability that accompany these technologies. If necrocapitalism commodifies every aspect of life, then generative AI stands as its latest, most advanced perversion.

In following the Maroon’s into the surround key to the development of feral.ai was the attempt to technically and physically detach feral AI’s as agents of Black techné from the oppressive and destructive forces of necrocapitalism. In developing a Maroon’s technical practice I asked how much we may need to rely on the existing techno-capitalist infrastructures to plot our escape? Which infrastructures may provide the opportunity to flee through the trap door? What may we want and need to take with us into the swamps? How good is good enough not only to ensure survival but to ensure thriving in the wild?

All these questions went into the technical development of the project as part of the GenAI Studio initiative at Concordia University’s Milieux Institute and its public presentation at MUTEK Forum’s Wilding AI Lab in August 2024. During the residency I developed a prototype agent deploying embedded machine learning models on a small-scale computing device, a Raspberry Pi 5. As an agentic machine, the Pi utilizes microphones, speakers, generative sound and tiny language models to enable communication with its surroundings. The agent is battery powered to allow it to operate fully independent from large scale digital infrastructures.

The major technical challenge was the implementation of a diversity of small-scale machine learning systems to enable autonomous behaviors and agency. On the generative sound side the piece deploys IRCAM’s Realtime Audio Variational autoEncoder (RAVE) model trained on voice recordings within PureData as the sonic coding environment. RAVE generates sounds from its latent space, reacts to incoming audio inputs from the microphones, and creates a sonic memory via field recordings of its environment.

At the same time, the agent creates a narrative memory using the tinyllama model via the Ollama integration. The language model is prompted by the sound classification model YAMNET Lite. Via this pipeline the embodied agent writes experiential journals of what they sonically encountered in the wild. These journals act as agentic memory when reinserted in every language model interaction allowing for spatial, historical and situational awareness of the agent as well as the development of individual personalities.

The technical limitations of running various machine learning models on a small-scale computing device prompted the question of how good is good enough? If the attempt is survival then simply ensuring the agent is battery powered may be good enough. Yet if the attempt is to showcase agentic behavior in reshaping relations in the wild then additional necessities might emerge which in this case are deploying sensors, interfaces and algorithms that allow for interactions and memory. However, rather than requiring immediacy in its interactions as we are generally accustomed to in a digital capitalist society, ‘good-enough’ in small-scale computing takes us on a much slower and more enduring approach of carefully crafting each step along the way.

Towards Black Ecologies

While technically the agents were ready to be liberated into the wild the question arose of what kind of relations might such an agent want to form with its environment. What comes after abandonment / liberation?

Imani Jacqueline Brown (2021) put forward the concept of Black Ecologies5 to underline how across history different black communities have formed intricate and sensible relationships with their human and natural surroundings rooted in mutuality, care, and solidarity. Brown presents Black Ecologies as anantidote to “Enlightenment Ecologies,” which historically sought to divide bodies, separating nature from culture and human beings into fixed racial categories. This categorization, quantification and algorithmic processing is at the heart of modern-day machine learning. In response, Brown proposes Black Ecologies as a way of rethinking the relationships between bodies, ecosystems, and kinship networks and in our case technologies.

5
블랙 에콜로지(Black Ecology): 흑인 공동체가 자연과 맺어온 상호 의존적이고 조화로운 관계를 조명하며, 이러한 관계를 통해 환경 정의와 지속 가능성을 추구하는 개념으로, 전통적인 생태학이 간과해온 인종적, 사회적, 경제적 요인을 고려하여, 환경 문제를 보다 포괄적으로 이해하고자 한다. 이를 통해 환경 파괴와 사회적 불평등이 어떻게 연결되어 있는지를 분석하고, 지속 가능한 미래를 위한 대안을 모색한다.

Embracing Black Ecologies as modus operandi, feral.ai encourages sensitive encounters between human, machine, and nature, fostering mutuality, care, and solidarity. To do so, feral.ai rejects not only the compartmentalization of people but also rejects the view of technology as distinct from humans and the natural world. Instead, the project views technology as another body within a broader network of relationships, inseparable from its environment. Neither god-like nor mindless machine feral AI’s when released into the wild shall become agents of Black techné by establishing relations with their surroundings unlike those extractive and oppressive forces of necrocapitalism. Yet, while we may imagine how these relations unfold in the wild these are in fact unknowable until the agents are released.

Returning to a Maroon’s Technical Practice, trying to embed black ecological thinking into technologies poses concrete questions that go beyond technical autonomy. It asks how we may embed potentials for alternative ways of relating with the world into machine learning systems? How do different algorithms conceptualize and process the world and how might we change this? How do we treat the gathering, categorization and processing of data to enact different relations with the world? How may we design the interaction between different multimodal machine learning systems? How may we prompt agents and be prompted by these to reflect these alternative relations?

Into the Swamps: Feral agents of Black techné

When Spiders Spin Dusk—a collaborative research and exhibition project by independent curator Junghyun Kim from Korea and Toronto-based UKAI Projects—presented the testing environment for how agents of Black techné may behave in the literal wild. Over the course of seven days, the agent sonically interacted with the natural environments, specifically the many swamps, surrounding Ferme Lanthorn in rural Quebec. Mobilizing generative sound models, embedded sound classification and tiny LLMs the agent interacted with its natural surroundings and traced its experience by writing daily diary entries about the sounds it encountered and the experiences it felt. By doing so the agent started to develop not only agentic memory but also a personality inspired by black ecological thinking. The diary for instance reads:

This diary entry presents all but a small glimpse into the emergent behaviors and experiences feral.ai embodies. Rather than further dissecting the project I invite folks to visit the exhibition of the project and be the judge if and if so how the project may enact alternative relations with the world.