A former ByteDance engineer says China’s AI industry is still far behind the US.<\/p>\n
Zhang Chi said China’s model benchmarks perform well but don’t translate into real-world results.<\/p>\n
Tech leaders like Elon Musk and Jensen Huang believe China is catching up and could even pull ahead.<\/p>\n
For all the talk that China is catching up to the US in AI, a former ByteDance engineer believes it’s actually falling further behind.<\/p>\n
“I don’t even agree with the assumption that Chinese models are catching up \u2014 I believe we’re still far behind,” Zhang Chi, a research scientist and assistant professor at Peking University, said on an episode of the “Into Asia” podcast. “I guess the gap is getting larger, very sadly.”<\/p>\n
Zhang, who said he spent about a year at ByteDance working on AI models<\/a> before returning to academia, said the differences go beyond the perceived rapid progress of Chinese startups.<\/p>\n While models from companies like ByteDance, TikTok’s parent company<\/a>, and Alibaba may score well on benchmarks, he said that doesn’t mean they work as well in the real world.<\/p>\n “On paper, every big tech company in China has a good model,” Zhang said. “But I don’t think they’re good enough.” He added that many teams are focused on “benchmaxxing” \u2014 optimizing for test scores rather than practical performance.<\/p>\n ByteDance and Alibaba have rolled out high-profile AI models \u2014 from video generators like Seedance<\/a> to open-source systems like Qwen \u2014 but they’ve also faced backlash over deepfakes, copyright disputes, and whether those models hold up in real-world use.<\/p>\n A key issue, Zhang said, is speed. He said top US companies can iterate on models far more quickly.<\/p>\n “Google can train or perform a full round of LLM training, both pre-training and post-training, in three months,” he said. “But ByteDance \u2014 probably we can only do one iteration in half a year.”<\/p>\n Zhang also pointed to China’s structural disadvantages, including access to advanced chips, weaker infrastructure, and lower-quality training data.<\/p>\n “There’s a huge difference between the infrastructure at Google and ByteDance,” he said. “I don’t think we’re getting high-quality data.”<\/p>\n He added that some companies rely on distilling outputs from leading US models rather than building their own data pipelines \u2014 a shortcut that may limit long-term progress.<\/p>\n Zhang said US firms also benefit from stronger user feedback loops. Products like ChatGPT, Claude, and Gemini<\/a> improve through constant interaction with users, which helps refine their models over time.<\/p>\n Chinese models, by contrast, risk falling into a negative cycle. “Chinese models started not as good, so no one really uses them for really important things,” Zhang said. “And the models continue to be not that good.”<\/p>\n Others say China is catching up<\/p>\n Zhang’s view contrasts sharply with some of the most prominent voices in tech, many of whom say China is rapidly closing the gap and could even pull ahead.<\/p>\n Nvidia CEO Jensen Huang has warned that the US risks falling behind<\/a>, while Elon Musk has said that China’s advantage<\/a> in energy and compute could help it overtake rivals.<\/p>\n AI pioneer Geoffrey Hinton has also said the US lead<\/a> may be narrower than it appears, cautioning it could erode over time.<\/p>\n Others take a more nuanced view. Alibaba chairman Joe Tsai has said the race will be decided less by model strength<\/a> and more by how quickly AI is deployed.<\/p>\n Still, Zhang’s assessment reflects a more pessimistic perspective from inside China’s AI ecosystem \u2014 one that suggests the gap with the US may be widening.<\/p>\n “To be fair, I don’t think any Chinese company can catch up with them soon,” he said.<\/p>\n