GPT's training data covers global languages, academic papers, tech forums, philosophical essays, social media, novels, and more. Training data for Chinese models: Chinese internet data has high redundancy; lacks high-quality academic texts and original thought content. Chinese AI generally doesn't 'read' much, and what it 'reads' is too narrow. When people don't read much or read narrowly, it leads to AI not 'reading' much or reading narrowly too, hahaha.