US Researcher’s 36‑Hour China AI Lab Tour Highlights Culture and Open‑Source Edge

Observations from a 36‑hour tour of Chinese AI labs

Nathan visited six leading Chinese AI research organizations—Zhuiyi, Moonlight, Tsinghua, Meituan, Xiaomi, and Qianyi (also known as 月之暗面, 智谱, 清华, 美团, 小米, 零一万物)—and met researchers, graduate students, and executives.

All labs expressed concern about ByteDance’s growing influence while simultaneously admiring DeepSeek for its research quality.

Cultural traits that accelerate progress

Researchers readily take on unglamorous tasks that directly improve model performance.

Many core contributors are current graduate students who work as full‑time collaborators rather than peripheral interns.

Students are less encumbered by previous AI hype cycles, allowing rapid adoption of new paradigms such as MoE scaling, RL‑based scaling, and agent‑centric designs.

Lower emphasis on personal ego reduces internal friction and makes organizational scaling smoother.

This contrasts with U.S. labs (e.g., OpenAI, Anthropic, Cursor, Google) where top talent often avoids internships or is isolated from core work, and where internal politics can hinder model‑level trade‑offs.

Open‑source as pragmatic strategy

Companies such as Meituan and Xiaomi develop their own large language models not for hype but to retain control over core technology. Their typical workflow is to train a general‑purpose base model, open‑source it for community feedback, and then fine‑tune a proprietary version for internal products. Open‑sourcing is treated as a practical means to obtain external validation and accelerate iteration rather than as an ideological commitment.

Infrastructure constraints

Nvidia GPUs remain the de‑facto standard for training; most labs report chip shortages that limit experiment scale.

The domestic data industry exists but varies widely in quality. Labs often build their own reinforcement‑learning training environments or rely on internal annotation teams (e.g., at ByteDance and Alibaba) to ensure data reliability.

AI market dynamics

China’s AI demand splits between a relatively small SaaS ecosystem and a massive cloud‑computing market. Consequently, AI services are expected to align more closely with cloud offerings than with traditional SaaS products.

Geographic and collaborative ecosystem

Beijing’s AI landscape resembles a dense “Silicon Valley” where competing labs are within walking distance, fostering an ecological network of mutual respect rather than a hostile tribal rivalry.

Overall, the combination of collaborative culture, extensive student involvement, pragmatic open‑source practices, and a cloud‑centric market orientation gives Chinese AI labs a distinct advantage in rapidly advancing large‑scale model development.

Code example

往
期
推
荐
1、
Spring Boot 4.0.6 发布，紧急修复了 8 个安全漏洞，赶紧看看自己的 Spring Boot 版本！！
2、
古法编程时代已结束？ClaudeCode 创始人：2026 年，我几乎不再手写代码了～
3、
火速吃瓜：Kimi K2.6设计能力超越Claude Design
4、
极简落地！SpringBoot3 注解式敏感词高效处理
5、
把 Tokens 消耗全部换成 DeepSeek V4，我每个月要花费多少钱？
6、
告别if-else噩梦！流程编排技术真的太香了！
点
分
享
点
收
藏
点
点
赞
点在看