Author: Florian Brand

Compiled by: Deep Tide TechFlow

Deep Dive Introduction: The background of this article is that SAIL (a media alliance that unites top AI writers on Substack, including Nathan Lambert, Sebastian Raschka, ChinaTalk, etc.) organized a visit to Chinese AI labs. The author, Florian, went with the group to more than a dozen companies, including Lunar Dark Side, Xiaomi, MiniMax, Zhipu, Meituan, Alibaba, Ant Financial, Moda, Zero One Things, and Unitree, and wrote this review.

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

While not exactly "famous," it has a certain level of visibility within the open-source AI community, and the perspective of foreign AI practitioners on the Chinese AI ecosystem is quite interesting.

text

Over the past 10 days or so, I had the privilege of visiting AI labs in China with my colleagues at SAIL. As someone who was visiting both China and the US for the first time in six months, I found the differences between the two places fascinating, but even more fascinating were the similarities.

What impressed me most was that the AI ​​researchers I met were all very humble.

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

Many researchers are close friends, from the same university, or share the same hometown. They discuss their work openly, and the results are published as papers a few months later.

This is one of the biggest differences between the US and Western AI circles. In the US, the atmosphere often resembles a zero-sum game. Labs are cautious about positioning. Researchers think about competition, and some have high self-esteem. Leaders insult and attack each other in leaked memos. This difference can perhaps be explained by the fact that leading US labs are closed-source, while many Chinese labs are open-source. Chinese labs are "somewhat wary" of ByteDance's Doubao, the most widely used chatbot, which is closed-source and has a significant lead.

Meanwhile, the overall atmosphere was strikingly similar to that of San Francisco. Researchers were extremely online, reading extensively on Twitter and Xiaohongshu, the latter becoming increasingly popular. They were all building their next model using Claude Code or their own CLI. Some monitored the training run while we were meeting, observing the reward curves rise. They were thinking about further scaling and complaining about insufficient computing power. They were frustrated with the current state of benchmarks.

Their primary focus is training better models. This differs from researchers in San Francisco, where they consider the political or philosophical implications of AI. They don't consider mass unemployment, permanent underclassing, or whether their models are conscious. They simply want to train excellent models.

Their eyes light up when they hear you've used their model. They're eager to fix every flaw in the current model in the next generation. They work through the night to push for the model's release and still show up in the office afterwards.

Most of the researchers I met were young, many in their early 20s or around 25. Some were undergraduates, but more commonly they were PhD students working in industry simultaneously. Their consensus was that industry is now more interesting than academia, a view I strongly agree with because I've done exactly the same thing. Labs place great importance on acquiring this type of talent, actively recruiting interns and graduate students; something Western labs don't do.

The researchers' optimism has extended to the general public, who seem even more optimistic about technology and the future of AI and robotics. During the trip, people recounted stories of their parents and grandparents using Doubao and DeepSeek for all sorts of things, including discussing mathematical theorems. This is a stark contrast to the West, where the general public is averse to AI.

Overall, this trip gave me a glimpse into this ecosystem. It's impossible to understand the culture of such a vast civilization in just a few days. As a staunch supporter of the open AI ecosystem and open research, I am very optimistic about the future of both and hope for extensive international collaboration.

I want to thank all the amazing people I met at The Dark Side of the Moon, Xiaomi, MiniMax, Zhipu, Meituan, Alibaba, AntRhythm, Moda, 01Wanwu, Unitree, and elsewhere. Thank you for your time and warm hospitality. I also thank SAIL for organizing this trip and all the participating writers and journalists. I am deeply grateful to have met so many outstanding and ambitious people in such a short time.