Episode Details

About twenty percent of AI-generated podcast episodes suffer from a strange repetition loop — the scriptwriter generates a section, then generates it again, producing duplicated content. In this episode, we diagnose the three interacting causes: attention collapse as context windows fill, low-temperature sampling that locks onto patterns, and word-count targets that incentivize padding. We also explore why the obvious fix — a review agent — introduced its own failure mode by leaking commentary into scripts. If you're building production pipelines on LLMs, this is a masterclass in debugging emergent failures.