While there’s a lot of talk about China catching up to the US in AI, a former ByteDance engineer believes it’s actually falling further behind.
“I also disagree with the assumption that the Chinese model is catching up. I believe we are still far behind,” Zhang Chi, a research scientist and assistant professor at Peking University, said on an episode of the “Into Asia” podcast. “Very unfortunately, I think the gap is getting even bigger.”
Zhang, who said he worked on developing AI models at ByteDance for about a year before returning to academia, said the difference goes beyond the Chinese startup’s perceived rapid progress.
Models from companies like TikTok’s parent company ByteDance and Alibaba may score well in benchmarks, but that doesn’t mean they’ll perform as well in the real world, he said.
“On paper, all the big tech companies in China have good models,” Zhang said. “But I don’t think they’re good enough.” He added that many teams focus on “benchmaxing,” or optimizing test scores rather than actual performance.
ByteDance and Alibaba have deployed high-profile AI models, from video generators like Seedance to open source systems like Qwen, but have also faced backlash over deepfakes, copyright disputes, and whether their models stand up to real-world use.
The key issue, Chan says, is speed. He said top U.S. companies can iterate on models much more quickly.
He said, “Google can train or run a full round of LLM training within three months, both pre-training and post-training.” “But with ByteDance, you can probably only do one iteration every six months.”
Zhang also pointed to China’s structural shortcomings, including access to advanced chips, weak infrastructure, and low-quality training data.
“There are significant differences in the infrastructure between Google and ByteDance,” he said. “I don’t think we’re getting high-quality data.”
He added that some companies rely on extracting output from major US models rather than building their own data pipelines, which could limit long-term progress.
Chan said US companies are also benefiting from stronger user feedback loops. Products like ChatGPT, Claude, and Gemini are improved through continuous interaction with users, which helps us refine our models over time.
In contrast, the Chinese model risks falling into a negative cycle. “The Chinese model wasn’t that good to begin with, so no one was actually using it for anything really important,” Zhang says. “And the model is still not that good.”
Some say China is catching up.
Zhang’s views stand in sharp contrast to some of the tech industry’s most prominent voices, many of whom argue that China is rapidly closing the gap and could even take the lead.
Nvidia CEO Jensen Huang has warned that the US risks falling behind, while Elon Musk has said China’s advantages in energy and computing could help it overtake rivals.
AI pioneer Jeffrey Hinton also said the U.S. gap may be narrower than it appears, warning that it could narrow over time.
Others take a more nuanced view. Alibaba Chairman Joe Tsai said the race will be won more by the speed of AI deployment than by the strength of the model.
Still, Zhang’s assessment reflects a more pessimistic view from within China’s AI ecosystem, suggesting that the gap with the United States may be widening.
“To be fair, I don’t think Chinese companies will be able to catch up anytime soon,” he said.