Latest / Logosofia no Dia a Dia
What DeepSeek's Training Data Reveals About Model Voice
When Daniel ran his model evaluation for podcast script writing, DeepSeek V4 Pro won not on benchmarks but on feel — its dialogue simply sounded more authentic. This episode traces why: DeepSeek's training corpus is 60% English and 30% Chinese, but that Chinese portion is heavily weighted toward narrative literature — Tang dynasty chuanqi tales, modern WeChat fiction, and other voice-driven storytelling. By contrast, Kimi from Moonshot AI trains on conversational social media like Zhihu and Weibo, while Qwen deliberately filters cultural bias through data augmentation. We explore the ASR…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.