Taekeun Kim

A Practice in AI Abstraction

For a long time I have been interested in how memories fade away. My audiovisual roadmap back to my home in Korea goes through various iterations as it not only loses its form, but blurs, collapses, and implodes into a permanent state of deformation. "Fading away" is a misleading notion. It implies erasure, which is hardly ever achieved.

Our distance to it is fixed. It is its reiterative deformation that projects an image of distance. Yet an ordinary trigger can jerk it forward. It was never as distant as it seemed, just suspended indefinitely.

Technology has been my medium, its imperfection unknowingly mirroring my own. The latent space of an LLM is not only a blueprint of its design but another mystery entirely.

These images were generated using a procedural prompt system I built to systematize computational failure modes: buffer overflows, lossy compression, bit-rot, semantic drift. The prompt engine below lets you configure subject, error type, data density, and iteration logic to generate prompts.

prompt engine

Select a subject, failure mode, and parameters to generate prompts.

Subject Error Params Iteration Generate

fire_1

A "starving" latent space render of a Fire experiencing a temporal buffer overflow. The AI must only use 0.15% of its structural memory for "light." A high-key white-out where the Fire is reduced to "ghosting smears", "jittery overlapping lines", "interlacing artifacts". The palette is limited to "bit-crushed" charcoal and absolute white.

forest_1

A "starving" latent space render of a forest. The AI is restricted to a 1-bit depth memory of "bark" and "twigs." A skeletal, high-key white-out where organic forms are reduced to clusters of high-frequency black stipple dots and jagged, hair-thin filaments. Extreme negative space. Heavy dithering artifacts and "digital dust" in a blindingly white void.

forest_2

A "starving" latent space render of a forest with a depth-map segmentation fault. The AI is trying to reconcile "distance" with only 1% of its spatial data. Fragmented, high-key white-out where "branches" are razor-thin black slashes hovering in an impossible, non-Euclidean void. Secondary prompt: Abstract further. The image is afraid of its own destiny.

ocean_1

A starving ocean render where the AI uses its own diffusion noise as the primary subject. A high-key white-out where the only shapes are "hallucinated" textures of salt and static, mimicking a neural network eating its own tail. Secondary: apply dimensional flattening. The AI is forbidden from calculating Z-axis depth — the ocean's volume is crushed into a single hyper-compressed horizontal plane of black bit-noise.

Zoom

Operational images: thermal drone feeds, histopathological scans, LiDAR point clouds, synthetic aperture radar. Images made by machines for machines, never intended for human eyes.

These images were never made to be seen. They exist to be processed, classified, acted upon. They belong to systems, not viewers.

Zoom enters these images at the threshold of their legibility and moves through them recursively, frame by frame, using Google Flow's generative engine as a kind of forced zoom: not into the subject, but into the image's own logic of representation. What begins as readable data progressively loses its function. Structure gives way to texture. Texture gives way to noise. The image no longer knows what it was of.

After Harun Farocki's operational images and Hito Steyerl's poor image theory, this work asks what happens when an image made to see: to target, to diagnose, to map: is instead watched until it forgets itself.

Each series begins with a single operational image fed into Google Flow. The output is cropped, then fed back as the next input, a recursive loop that forces the model to continuously re-interpret its own generation. The grid reads left to right, top to bottom: twenty-five iterations of a single image forgetting what it was.

UAV / Thermal + Aerial