Zhang Chi, a former engineer on ByteDance's Seed team and now an assistant professor at Peking University, revealed on the podcast Into Asia that it takes roughly six months for ByteDance to complete a full round of large model training (pre-training plus post-training), while Google is reportedly able to do it in three months. He believes iteration speed is one of the core reasons Chinese companies struggle to catch up. Zhang spent about a year at ByteDance on the math team, which he described as more research-oriented, with a positioning "more for publicity," unlike the pre-training and post-training teams responsible for model delivery.
Zhang described Seed's internal "benchmaxxing" culture: team leads evaluate performance based on the benchmarks they are responsible for, and everyone is chasing higher scores. "But that doesn't translate into a good experience in actual use," he said. On paper, he noted, the models from big Chinese companies can match US frontier models, but in practice they are "not good enough." Seed's goal was to be globally top-tier, "but unfortunately, I don't think we caught up." Even the goal of being number one in China "wasn't achieved." At the end of 2024, Seed believed it had matched GPT-4o, but then DeepSeek was released, and the team realized the gap still existed. When Zhang joined, the entire group was urgently pivoting to reinforcement learning.
In the same interview, Zhang made a blunt assessment of the US-China AI gap: "I don't even agree with the statement that China is catching up. I think we are still far behind. The gap is widening, and that's unfortunate." He said his colleagues and students generally agree, but acknowledged that leaders at listed companies like Zhipu AI and MiniMax would not share that view.
He attributed the gap to three factors. First, shortcutting via distillation: many Chinese companies directly use outputs from Claude, GPT, or Gemini as training data. "Claude recently said it detected a large number of distillation attempts. I guess that's the shortcut some companies take," Zhang said. However, he acknowledged that DeepSeek demonstrated real architectural innovation with V3 and R1. Second, the missing user feedback flywheel: US models are good, so they attract more users, whose feedback makes them even better; Chinese models start poorly, attract fewer users, and don't get enough data, creating a vicious cycle. Third, infrastructure disparity: during his internship at Google, he found the infrastructure "fantastic — code runs very smoothly," unlike the massive gap he saw at ByteDance.
Zhang also discussed ByteDance's chip situation. The company primarily uses NVIDIA H20 (the compliant version for the Chinese market), with the fastest chips allocated to pre-training and post-training teams, while other teams use H20. He also mentioned that ByteDance is procuring NVIDIA's next-generation chips (he said "B300 or something," uncertain about the exact name), "but definitely not in mainland China." Regarding domestic chips, Zhang said ByteDance Seed does have some, but "I don't think any team that wants to speed up iteration is actually using them." Training "definitely doesn't" use them, though some teams might consider domestic chips for inference deployment. When asked whether DeepSeek uses Huawei chips, he replied, "That's what the news says, but I'm not sure." He emphasized that every major Chinese company faces chip procurement challenges. These are Zhang's personal observations based on his experience at ByteDance and do not represent the situation for all Chinese AI companies.
Zhang said he personally uses Claude Code and GitHub Copilot's Agent mode daily. His students use Claude Code for training monitoring: it automatically checks GPU status, debugs and restarts if it finds errors, and runs a patrol every 30 minutes — "so we can go to sleep."
When asked whether all Chinese AI practitioners use Claude Code, Zhang first said "not all" — he himself is more accustomed to Copilot — but then confirmed that Chinese AI practitioners widely use US-made agent tools. He bluntly said that Chinese models' coding agents are "completely impractical," citing his personal experience rather than systematic evaluation. ByteDance's overseas teams can directly use Cursor, leading Zhang to speculate that ByteDance's future large models may be partially built with the help of US agent tools. He also shared a personal dilemma: Claude Code is so good that he's considering "whether I still need to train PhD students," but worries that if he doesn't cultivate the next generation of researchers, there will be no one left to do research.