Slop Machines: on the interaction between feed recommender systems and GenAI (notes.hella.cheap)

"Slop Machines" analyzes how modern feed recommender systems—modeled as contextual bandits—interact with generative AI, and why that interaction creates both powerful personalization and worrying failure modes. Feeds optimize a reward function built from user actions (watch time, swipes, app closes), and bandit algorithms (Thompson sampling, UCB, epsilon‑greedy) must trade exploration for exploitation. Unusually, human reward schedules make partial uncertainty itself engaging, so a tuned exploration policy can increase engagement without the usual “cost” of exploration. Recommenders scale by projecting millions of items into low‑dimensional feature spaces via collaborative filtering or content embeddings (PCA, random projections, neural encoders). The key technical pivot is that modern generative models (autoencoders, U‑nets, diffusion models) let systems synthesize content from predicted feature vectors—i.e., interpolate directly between known liked items. That enables “exactly what the model predicts” personalization, but only by blending existing signals, producing homogenized, low‑creativity “slop” that amplifies reward loops. For ML practitioners this raises practical concerns: stronger feedback loops, sample‑efficiency incentives to overfit to short‑term engagement, diversity collapse, and ethical risks of exploiting vulnerable demographics (e.g., an aging population with high spending power). The piece urges awareness of these system dynamics when designing objectives, exploration policies, and safety mitigations for generative recommender systems.