The Latent Space of the Middle Kingdom

We have gotten to know all the founders of the big American AI companies, sometimes even researchers like Yann LeCun or star developers like Andrej Karpathy. In France too, Arthur Mensch enjoys considerable name recognition, especially since the president started talking about Mistral on France 2's evening news. But what about China?

To paraphrase a question from the editor-in-chief of Le Grand Continent, the Paris-based geopolitics journal, who asked on television whether anyone on the panel could name three famous living Chinese people (the answer stopped at the Chinese president).

I could ask you the same question. Do you know the founder of the Chinese company that released, this month, one of the best models of the moment? And the name of the model?

In this newsletter, most of which is reserved for paying members, we explore the world of Chinese models. I have wanted to write about this for a long time, but I had the feeling that things were not always very clear.

Because we don't know much about what is being done in China. We don't know who the labs are, or what their work really looks like.

And in Europe, because we wanted to simplify the narrative, nobody has any real knowledge of how they work, which is often caricatured.

And yet Chinese models are already everywhere. In the rankings of the best AIs, at higher positions than Mistral. They are even already nestled inside the products we use every day.

Since the release of Mythos, we thought the Americans were ahead again. But, and this is probably the most shocking part, that is not the case. Worse, the lead of the best American models is tending to shrink.

On June 18, 2026, on X, Elon Musk estimated that Chinese models would reach the level of the best American frontier models in the first quarter of 2027. Jie Tang, the founder of Z.ai, replied that it would be sooner than that.

Elon Musk: "Probably Q1."
Jie Tang: "won't take that long."

And yet this week, at the G7 in Évian, during the lunch that brought together heads of state and the AI bosses, Sam Altman, Dario Amodei, Demis Hassabis, Arthur Mensch of Mistral. Not a single Chinese entrepreneur was present.

Granted, China has not been invited to these summits since 2009, back when the G8 still hosted the big emerging economies.

But setting up a meeting with Chinese entrepreneurs right after Anthropic was barred from exporting Fable would have been interesting.

For now, it seems that France wants above all to align itself with the coalition led by the United States, which Amodei and Hassabis were backing.

Europe itself has closely aligned its position with Washington's: Chinese models are a security risk. Italy blocked DeepSeek upon its release, Germany had asked for it to be pulled from app stores, the Czech Republic banned it from its administration, and France, through the CNIL (the French data protection authority), investigated its data storage strategy without banning it.

On the China question, there are two important issues.

First, training, which is done in a Chinese political environment. But it seems that post-training makes it possible to limit the ideology of the models.

Second, and this is a valid reason, the data of the online services is stored in China, where the 2017 intelligence law can compel a company to hand it over to the state. Much like the Cloud Act in the United States, for that matter.

But that risk only applies to remote use, when the models run on servers in China. Unlike OpenAI and Anthropic, though, these models are open and downloadable. Nothing prevented, and nothing prevents today, running them in European data centers, where the data never leaves.

It is a question nobody dares to ask, such is the priority given to Mistral, which is nevertheless far behind in terms of performance.

The political vision (having our own, less capable models) and the technological vision (building on open source models to grow a European strain of models) are not aligned. For now, the political vision is winning.

But it is not just a question of models. The chips that run them are also a real issue, and there things are far more complex. The Chinese chip ecosystem is not really accessible in Europe. We don't really understand how it works. We have a Chips Act based on emulating an American and Taiwanese model that we don't know how to imitate, whatever the amount of money invested. The days when Nvidia had its first chips built in France are long gone.

Upgrade to the paid version!

This newsletter matters because it compiles everything we have found and understood on the subject. If you are not yet a subscriber, now is the time. This second half of 2026 has shown that not only is everything changing (models, agents, loops) but that it may not be possible to catch up except by spending a great, great deal of time on it. Because I am working on an AI project, I spend that time, and I share the fruit of that thinking with my paying members.

Join us here.