Episode Details

A listener built a Hugging Face dataset of 5,000+ podcast episodes, each stamped with the model that generated it and synced daily. That makes it a longitudinal, model-attributed corpus — an ablation study where the treatment is the model version and the outcome is stylistic quality. This episode digs into what that artifact can actually measure: why verbatim repetition is a solved problem while semantic repetition slips past every keyword method, how U-shaped attention bias and lost-in-the-middle effects turn a 40-turn script into a loop machine, and why repetition is really a model stalling rather than malfunctioning. Then the concrete toolkit — sentence embeddings with non-adjacent pair filtering, CASE for anisotropy, BERTopic for intra-episode recurrence, c-TF-IDF for per-model tic profiles, compression ratio as a cheap first signal, and correlation dimension for a plottable per-episode scalar. Episode #160567 — open it directly at myweirdprompts.com/160567