Former ByteDance engineer: The gap between China and the US in AI is widening, mainly due to the absence of distillation shortcuts and feedback flywheels.

According to Beating Monitoring, Zhang Chi made a direct assessment of the AI gap between China and the U.S. in the same interview: “I even disagree with the idea that China is catching up. I believe we are still far, far behind. The gap is widening, which is very unfortunate.” His colleagues and students around him generally agree, but he also acknowledges that the leadership of publicly listed companies such as Zhipu and MiniMax would not share this view.

He attributes the reasons to three aspects. First, shortcutting through distillation: he believes many Chinese companies directly use the outputs of Claude, GPT, or Gemini as training data. “Claude recently said it detected a large number of distillation attempts—I guess that’s how some companies take shortcuts.” However, he also admits that DeepSeek has shown genuine architectural innovation in V3 and R1. Second, the lack of a user-feedback flywheel: the U.S. models are useful, so there are more users; user feedback then makes the models better. China’s models start off not good enough, so there are fewer users, and it’s harder to obtain data, forming a vicious cycle. Third, infrastructure gaps: during his internship at Google, he felt the infrastructure was “so good—the code runs extremely smoothly,” and the gap compared with ByteDance is enormous.