This blog post discusses the regression of conversational quality in GPT-6 across various effort levels, highlighting issues such as formulaic writing, repetition, and a lack of nuance. The author calls for more rigorous evaluations of conversational intelligence, humor, and the ability to adapt corrections, stressing that these traits are essential for a language model's utility. They urge OpenAI to investigate these concerns further and to ensure these qualities are prioritized alongside technical performance.