AI marks a new regime of overheating, fueled by voracious demands for energy and extraction and sustained by vast infrastructures of cooling. Like a pyroclastic flow, it surges through socioscapes and climates, generating, predicting, ranking, and authorizing while producing a strange fusion of acceleration and paralysis. Even the living body, reshaped as a pyroplastic assemblage, flickers under machinic fluoroscopy, exposed within regimes of domination that render deviation and refusal visible and actionable.
ABLAZE engages AI as material, system, and counterpart – as a pharmakon, a force at once enabling and destructive. At the center of the work, two irreducibly different yet entangled bodies meet: the human body – lived, sensing, vulnerable flesh – and the constructed, distributed, operational AI body, in which power assumes the form of classifications, thresholds, and decisions.
Following a tomographic logic, the work produces partial views of the AI body by probing models from contrasting perspectives, much as CT imaging reconstructs a body from multiple projections. Yet ABLAZE does not merely image this body – it intervenes in it. Steering vectors become incisions in activation space, shifting responses to the same prompt between opposing positions of power and revealing the tension between agency and control within the intervention itself.
The resulting responses become audiovisual and narrative fragments in which authority meets deviation and subjection meets refusal. Model activations are translated into maps, scores, and moving surfaces. Visitors can enter their own prompts and test how the model’s responses shift between opposing positions of power. In this way, they encounter AI as an evolving field of relations.
ABLAZE expands resolution beyond its predominantly technical meaning and unfolds it as a multilayered artistic practice – an aesthetics of resolution. It opens a sensory and performative field in which audiences can explore how collective agency might emerge from these entanglements.
During the opening, the entire process is enacted live, forming the basis for both the exhibition and audience participation.
Note on the Artistic-Computational Approach
ABLAZE uses contrastive activation steering to investigate large language models from within their operational structure. Following the tomographic logic introduced above, it probes models across different layers and from contrasting positions, making differentiations between authority and subjection available for artistic-computational experimentation.
The project presents language models with paired situations, each seen from two perspectives: that of a person exercising power and that of a person subjected to it. The differences between the resulting activation patterns are averaged across many such pairs, producing a geometric direction between authority and subjection within the model’s internal structure.
Shifting the model in either direction along this axis can orient its response to the same prompt more strongly toward authority or subjection, revealing different expressions of domination, submission, resistance, and vulnerability.
Dimensionality reduction translates the high-dimensional activation space into two-dimensional geometries of proximity, separation, clustering, and orientation. These function as provisional “organograms” of the AI body, tracing how the tested differentiations are distributed across activation space. Projected geometries become spatial and visual scores, steering vectors become dramaturgical forces, and model responses provide material for narrative, choreographic, audiovisual, and interactive composition.
