Weapons of Mass Distraction

The 2017 paper that introduced the transformer architecture was titled Attention Is All You Need. The claim was technical, but it has since become something closer to a description of the visual culture transformers produce — images optimized for immediate legibility, calibrated to the patterns a model has already learned to recognize as worth attending to. A visual monoculture, arriving at the same temperature.

Weapons of Mass Distraction is made by intervening in that mechanism directly. During image synthesis, I selectively suppress or remove the activations a generative transformer has identified as salient — the features it would otherwise amplify. The model does not stop. It cannot stop. It repairs itself with what remains, generating meaning from increasingly little, in directions no prompt could have requested. Some layers continue doing what they were trained to do; others have been altered upstream. What arrives is the result of that conflict — an image produced with full conviction in an unintended direction, with semantic and stylistic channels decoupled from one another.

The reference point is Claude Shannon. Brains and models alike are pattern-completion engines that cannot tolerate the absence of stimulus. Place a perceiving system in an isolation chamber and it will manufacture signal — heartbeat, hallucination, confabulated structure. The same is true of the model. There is no escaping it. One can only develop strategies for working alongside it.

These images are one such strategy. They are what a system produces when the things it has been trained to find interesting have been taken away from it, and it continues anyway.

To me these objects look like time bombs covered in slop, and I have no idea which wire to cut to defuse them.