Author: Florian Brand

Translation: Shenchao TechFlow

Introduction by Shenchao: The background of this article is that SAIL (a media alliance that brings together top AI writers on Substack, including Nathan Lambert, Sebastian Raschka, ChinaTalk, etc.) organized a visitation tour of Chinese AI laboratories. The author Florian joined the tour visiting more than a dozen companies, including Moonlight's Dark Side, Xiaomi, MiniMax, Zhiyu, Meituan, Alibaba, Ant, Modao, Zero One Everything, Yushu, etc., and penned this impression.

Florian Brand is a PhD student at Trier University in Germany and the German Research Center for Artificial Intelligence (DFKI), focusing on the application and evaluation of large language models.

Not "very famous," but has some visibility in the open-source AI circle; the perspective of foreign AI practitioners on the Chinese AI ecosystem is quite interesting.

Main Text

About 10 days ago, I had the privilege of visiting Chinese AI laboratories with SAIL partners. As someone making my first visit to China and the U.S. within six months, I found the differences between the two locations fascinating, but what intrigued me more were the similarities.

The most impressive thing to me was how humble the AI researchers I met were.

They spoke highly of other laboratories and peers. DeepSeek was mentioned frequently, possibly because they had just released a model a few days before our visit, and people talked about DeepSeek's paper with genuine admiration.

Many researchers are close friends, coming from the same university or sharing the same hometown. They openly discuss their work, and their research results will be published as papers months later.

This is one of the biggest differences from the Western AI circle. In the U.S., the atmosphere often feels more like a zero-sum game. Laboratories are very cautious about positioning. Researchers think about competition, and some have a high self-evaluation. Leaders insult and attack each other in leaked memos. This difference might be explained by the fact that leading labs in America are closed-source, while many Chinese labs are open-source. Chinese laboratories are wary of ByteDance's Doubao, which is the most used chatbot, and closed-source, having a significant lead.

Meanwhile, the overall atmosphere is surprisingly similar to San Francisco. Researchers are highly online, reading extensively on Twitter and Xiaohongshu, the latter becoming increasingly popular. They all use Claude Code or their own CLI to build the next model. Some people monitor training runs while we meet, observing rewards curves rising. They are thinking about scaling further, complaining about insufficient computing power. They feel frustrated with the current benchmark testing status.

Their main focus is training better models. This differs from San Francisco, where researchers think about the political or philosophical implications of AI. They do not consider mass unemployment, a permanent underclass, or whether their models are conscious. They just want to train outstanding models.

Their eyes light up when they hear you used their model. They are eager to fix all the flaws of the current model in the next generation. They hustle to push model releases day and night and still appear in the office afterward.

Most researchers I met were quite young, many in their early 20s or around 25 years old. Some are undergraduates, but more commonly, they are PhD candidates working in the industry. Their consensus is that, compared to academia, the industry is much more interesting right now, a view I strongly agree with as I have done exactly the same thing. Laboratories place great emphasis on acquiring such talent, actively recruiting interns and graduates, which is not done in Western labs.

The researchers' optimistic mood extends to the general public, who seem more optimistic about technology and the prospects for AI and robots. Along the journey, some shared stories of their parents and grandparents using Doubao and DeepSeek for various activities, including discussing mathematical theorems. This is clearly different from the West, where the general public harbors a strong disdain for AI.

Overall, this trip gave me a glimpse into this ecosystem. It is impossible to understand the culture of such a large civilization in just a few days. As a firm supporter of open AI ecosystems and open research, I am very optimistic about the future of both and hope for lots of international cooperation ahead.

I want to thank all the wonderful people I met at the Moonlight's Dark Side, Xiaomi, MiniMax, Zhiyu, Meituan, Alibaba, Ant Group, Modao, Zero One Everything, Yushu, and elsewhere. Thank you for your time and warm hospitality. Additionally, thanks to SAIL for organizing this trip and to all the writers and journalists involved. I am very grateful to meet so many outstanding and ambitious people in such a short time.