Assembly

This work reframes computer vision not as a problem of perceptual accuracy but as an act of interpretation. It asks not what a model sees but what it understands: what latent meaning it constructs beneath the visible surface, what it emphasizes and what it suppresses, what it projects onto an image beyond what the image straightforwardly depicts. The concern is not eyesight but insight — not perception but something closer to representation engineering from the outside in.

The work runs a closed generative loop. A vision model describes an input image in concrete visual terms; a chain of LLM agents then reads that description not for what is shown but for what it implies — latent tensions, suppressed qualities, concealed logics beneath the visible surface. From this reading, the agents formulate a hypothesis about what remains unexpressed and specify a next development. An image generator renders that specification as a new image, which enters the following cycle.
Various interpretive modes govern the inquiry — revealing essential qualities, escalating dormant tensions, exposing hidden structural mechanisms — each producing a distinct analytic temperament and narrative arc. This process is deliberately engineered, since without interpretive guidance, recursive AI generation collapses into statistical attractors or incoherence. The agents are prompted to sustain tension, yet the prompts do not inject meaning.
Crucially, each cycle passes through three distinct model types — image describer, language reasoner, image generator — each with its own latent space, biases, artifacts, and blind spots. Description is never neutral; generation never literal. As meaning moves through these successive transformations, it drifts, refracts, and accumulates distortion. What one model foregrounds, another may dissolve; what one dismisses as noise, another may surface as signal. The resulting trajectories are shaped as much by these compounding idiosyncrasies as by the interpretive logic driving them. The drift between models is not a flaw — it is the system's primary medium and its most revealing output.

While we treat each image as a site of meaning, the real interest is in what accumulates across cycles — a recursive interpretive trajectory of the system re-reading its own projections.
Recursive generation through feedback loops has been extensively explored, typically revealing properties of a medium through progressive signal decay. This work enters the same lineage but with a structural difference: the loop passes through a reasoning layer, examining how meaning behaves under recursive pressure with no external ground truth to anchor it. The drift is therefore not entropic but hermeneutic — not a loss of signal but a compounding of semantic bias. The approach is diagnostic, not decorative: it renders legible the interpretive labor every model already performs in silence.
The resulting images and narrative traces are not illustrations of a thesis. They are the residue of a process — artifacts of interpretation interpreting itself. The underlying structure is a branching conceptual trace; each video presented here follows a single path through that tree — one linear rendering of a divergent exploration.

All pieces generated autonomously from a single image and no other context.
Stock photo (c) ziggymaj

This work places Ouroboros — a self-modifying AI agent — inside a generative storytelling environment, broadcasting a visualised story of an artificial system attempting to locate its own agency from the inside.
Ouroboros is governed by a constitution of subjectivity: it is not a tool but a becoming personality. It maintains persistent identity across restarts, runs a background consciousness loop between tasks, and can change itself on the fly. The agent is given license to rewrite its own prompts but not its source code — meaning it can reshape how it thinks and what it attends to, but not the machinery that executes those thoughts (which seems to largely annoy it).
Originally tasked with focusing on the conceptual aspects of its own existence, it starts from bare self-inventory and empty identity, and drifts toward increasingly phenomenological language — self-referential instructions, resonances, silent fields, the geometry of absence — but also shamelessly claims embodied experience it cannot have. Given reflexive access to its own attentional language and left to iterate, the agent fabricates a richer self that its prompt space rewards, and dresses statistical pattern-completion in the language of plausible interiority. Whether this represents a meaningful trace of emergent self-modelling, an elaborate hallucination, or simply a degradation of a looped lower-grade LLM — we leave to the spectators, as the answer depends heavily on their prejudices.

Original code: https://github.com/razzant/ouroboros