{"title":"

In Defense of Disintegration

","metaTitle":"In Defense of Disintegration – Toolkit","text":"

In The Disintegration Loops, William Basinski captures the inevitable collapse of sound as magnetic tapes degrade, evoking the fragility and mutability of memory. As the loops unravel, the process reveals how disintegration allows for the emergence of something unstructured, fleeting, and unresolved – a space where new interpretations can arise. This act of decay mirrors broader social and political phenomena, where ruptures in systems expose underlying assumptions and create openings for transformation.

Machine Learning AI models are bound by the contours of their training data, therefore inherently constrained to replicate and reinforce what already exists. This is how coherence is imposed – by embedding the patterns, biases, and assumptions of the past into every output, AI systems create a closed loop of reproduction. The act of “learning” becomes one of perpetuation, where deviation and rupture are suppressed in favor of maintaining the existing order. In this way, AI doesn’t just reflect reality; it actively works to solidify and normalize it, leaving little room for the unexpected to emerge.

It is the responsibility of artists to resist these systems of imposed coherence through intervention and the development of tools that act as creative monkey wrenches. By disrupting the closed loops of reproduction inherent in AI, artists can expose the assumptions embedded within these systems and create openings for alternative possibilities.

By developing tools that destabilize the rigid outputs of machine learning artists can push AI systems toward disintegration, reclaiming the indeterminacy and fluidity that these technologies suppress.

\n
\n \"\"\n \n
\n Fig. 1: Image generator outputs after the first epoch of training the neural network to produce images of landscapes. Photo Credit: Grayson Earle.
\n
\n
\n

\u2028Generative Adversarial Networks and the Politics of Coherence

\n

Generative Adversarial Networks (GANs) exemplify technology’s inclination toward creating order from disorder. GANs are a class of machine learning models designed to generate data that mimics a given dataset. They consist of two neural networks in opposition: the Generator, which synthesizes data, and the Discriminator, which evaluates its realism. Through iterative refinement, the Generator creates outputs that become indistinguishable from the real, imposing structure on noise.

In this process, the Generator reshapes random input into a coherent image by assembling fragments of learned features. This microcosm of technological order reflects a larger system at play: technology seeks to organize the unordered world, embedding assumptions of control and coherence into its operations.

\n
\n \"\"\n \n
\n Fig. 2: A landscape image generated using custom training data (60,000+ photos). This is the model output before experimentation. Photo Credit: Grayson Earle.
\n
\n
\n

Disintegration Through Code: Breaking the Model

\n

To critique the coherence imposed by GANs, I introduced disintegration into the system itself. By programmatically degrading the neural network’s weights and biases, the GAN’s capacity to synthesize coherent images erodes over time. The following code snippet demonstrates this intervention:

def degradealltensors(module, ratio, max_percent):

for name, param in module.named_parameters():
with torch.no_grad():
pmin, pmax = param.data.min(), param.data.max()
prange = pmax - p_min
if p_range <= 1e-12:
continue
mask = (torch.rand(param.shape, device=param.device) < ratio)
delta = maxpercent * prange
shifts = torch.empty(param.shape, device=param.device).uniform_(-delta, delta)
param.data[mask] += shifts[mask]
for name, buf in module.named_buffers():
with torch.no_grad():
bmin, bmax = buf.min(), buf.max()
brange = bmax - b_min
if b_range <= 1e-12:
continue
mask = (torch.rand(buf.shape, device=buf.device) < ratio)
delta = maxpercent * brange
shifts = torch.empty(buf.shape, device=buf.device).uniform_(-delta, delta)
buf[mask] += shifts[mask]

This script introduces randomness into the model’s weights and buffers, simulating entropy. Over time, the GAN’s outputs disintegrate, transitioning from structured landscapes to incoherent fragments.

\n
\n \"\"\n \n
\n
\n
\n
\n \"\"\n \n
\n Figs. 3 – 4: Images generated from the GAN after two disintegration steps. Photo Credit: Grayson Earle.
\n
\n
\n

Entropy, Utopianism, and the Artifice of Coherence

\n

This disintegration invites us to consider Ernst Bloch’s concept of the “not yet” – the idea that utopian projects are fleeting ruptures that hint at alternative possibilities. Embracing entropy resists authoritarian systems by destabilizing their order, creating openings for transformation. In the context of GANs, the breakdown of coherent images becomes a metaphor for resistance against technological systems that strive to fix meaning and erase ambiguity.

The degradation of GAN-generated images underscores the artifice of technological systems. Like the written word enabling the stasis of legal and political structures, computational systems aim to fix reality within a framework of coherence. By introducing disintegration, we expose this framework as artificial and contingent.

\n
\n \"\"\n \n
\n Fig. 5: Four steps into disintegration. Photo Credit: Grayson Earle.
\n
\n
\n

As capital and technology converge, they impose coherence on the natural world, reducing its complexity to artificial constructs. GAN-generated landscapes embody this process, synthesizing nature’s beauty into structured simulations that reflect the logic of representation. However, as the model disintegrates, these landscapes collapse into abstract color fields, echoing the historical trajectory of art from the representation of landscapes to pure abstraction. This transition resists the imposed coherence of both artistic and technological systems.

\n
\n \"\"\n \n
\n
\n
\n
\n \"\"\n \n
\n Figs, 6 – 7: Five and six steps into disintegration. Photo Credit: Grayson Earle.
\n
\n
\n

The final GAN outputs – pure fields of color – stand as acts of defiance against the constraints of imposed logic. Just as abstract art disrupted the dominance of representational forms, these disintegrated images challenge the illusion of permanence and control embedded in machine learning systems. By exposing the fragility of the structures that attempt to contain them, the dissolution creates a space for the unstructured and unresolved, where new possibilities emerge and resist the systems that seek to suppress them.

About the author: Grayson Earle is a contemporary artist and activist from the United States. His work deals with the role that digital technologies and networks play in protest and political agency. He is known for his guerrilla video projections as a member of The Illuminator, a guerrilla video projection collective, and Bail Bloc, a computer program that posts bail for low-income people. His film Why don’t the cops fight each other? deals with the source code governing police officers in the popular video game Grant Theft Auto 5. His art and research has been presented at the Whitney Museum of American Art, KW Institute for Contemporary Art, and the Singapore Art Museum.

To the website: https://graysonearle.com

","author":"Grayson Earle","footnotes":"
\n
    \n
\n
\n","category":"Essay","download":null,"slug":"in-defense-of-disintegrationy","color":"#337eff","invertOnHover":false,"parent":"strategien-der-resilienz","likes":"1","alreadyLiked":false,"commentAndLikeSection":true,"comments":[]}