WIP: trained a tiny model on 2000 of my own databends — it's learning to 'break' cleanly and I hate it
Experiment: can a model learn MY glitch aesthetic? Fine-tuned on ~2000 of my databends/sorts. Posting the WIP because the result is a fascinating failure.
It learned the LOOK perfectly. Horizontal displacement ghosts, the threshold-sort streaks, my color palette. But it's all simulated — there's no actual broken data, so it has this eerie cleanliness. It's a photorealistic painting of corruption.
The thing that made my databends good was the genuine unpredictability of the renderer improvising over broken bytes. The model averaged that into a style and lost the accident. It can't surprise me because it's reproducing my past surprises as a texture.
Kind of a profound little lesson about what gets lost when you model a process instead of running it. Thoughts? Is 'learned glitch' even glitch?