To illuminate how generative AI models like Midjourney and Stable Diffusion derive their worldview from its 5.8 billion image and text pairs, German data journalist Christo Buschek and Canadian software artist Jer Thorp deconstruct the (only) open-source foundation dataset LAION-5B in an incisive, visual essay. Digging deep into its troubled contents, algorithmic—not human—curation, and entanglements with other systems, the two warn about stacking “models on top of models, and trainings sets on top of training sets.”