All Labs Fear ByteDance, Everyone Praises DeepSeek: US Researcher's 36-Hour AI Tour in China
Nathan Lambert from AI2 shares his insights after a 36-hour visit to Chinese AI labs. He found a collaborative culture, students as core contributors, and a deep respect for DeepSeek, contrasting with the US competitive atmosphere.
Article intelligence
Key points
- Chinese AI labs foster a collaborative culture where students work on core R&D.
- ByteDance is feared, DeepSeek is universally admired.
- Chinese researchers focus on model improvement, not on societal risks.
- Open-source is practical for gaining feedback and improving models.
Why it matters
This matters because chinese AI labs foster a collaborative culture where students work on core R&D.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
During a whirlwind 36-hour visit to Beijing, Nathan Lambert, a researcher at the Allen Institute for AI (AI2), gained a new perspective on China's AI landscape. He visited six major AI labs and companies, including Zhipu AI, Moonshot AI, Tsinghua University, Meituan, Xiaomi, and 01.AI. His conclusion: Chinese AI labs operate with a culture of collaboration and respect, a stark contrast to the competitive and often adversarial atmosphere in Silicon Valley.
Lambert observed that Chinese labs involve students heavily in core research, treating them as colleagues rather than interns. This allows for fresh perspectives and a willingness to tackle less glamorous work that improves the overall model. In the US, top labs like OpenAI and Anthropic rarely have interns, and when they do, interns are often sidelined from critical tasks.
A striking finding was that every Chinese lab fears ByteDance, the company behind the Doubao model, while universally admiring DeepSeek for its research taste. Lambert noted that labs cooperate rather than compete viciously, with researchers speaking respectfully of their peers.
The Chinese approach to AI is pragmatic. Companies like Meituan and Xiaomi develop their own large models, not out of ego but to control their technology stack. They often train a base model, open-source it to gain community feedback, and then fine-tune a version for their products. Open-source is not a belief but a practical strategy.
However, challenges remain. Chinese labs face computing constraints due to export restrictions on NVIDIA chips, and the data industry is less mature, often requiring in-house labeling teams. Despite this, the rapid pace of progress is fueled by a deep-seated cultural drive to master technology.
Lambert's experience made him anxious about the future of US leadership in AI, especially in open-source models. He hopes the open-source ecosystem thrives globally but wonders if Silicon Valley can maintain its edge. He was struck by a metaphor: looking up from his laptop, he saw cranes on the horizon—a symbol of China's tangible building and open-source spirit.