Another glimpse of the Chinese AI scene: Z.AI
Yesterday I had my first conversation (in English) with Zhipu’s GLM-4.7. It was cool because I got to talk with an actual Chinese AI about topics like: the representation of Chinese AI in “AI 2027”; representations of AI in the “Culture” SF series versus the “Wandering Earth” movies; consequences of Fast Takeoff; comparisons of China and America in general; and Chinese ideas about AI society and superintelligence. (An aligned Chinese superintelligence might be a heavenly bureaucrat or a Taoist sage.)
It was one of those AI conversations where you know that something really new is happening, and I now consider GLM to be as interesting as the leading American models. (A year ago I also spoke with DeepSeek-r1, but it never grabbed my attention the way that GLM has done.)
So what’s the status of the Chinese AI sector? In the same way that in America, AI is pursued by older Internet titans (Google, Meta/Facebook, X-Twitter) as well as by newer companies that specialize in AI (OpenAI, Anthropic), the Chinese AI sector is a mix of big old Internet companies with all the money (Alibaba, ByteDance, Tencent, Baidu) and “AI 2.0” startups (Zhipu, DeepSeek, Moonshot).
Keeping in mind this two-tier structure, shared with the American AI sector, is probably the best way for an outsider to get a grip on it. The Chinese AI sector is “like America”, except that they do much more open-source, don’t have American AI’s access to the best chips or the same international brand recognition, and are based in a country with a socialist government and most of the world’s manufacturing capability.
For keeping track of what’s happening, my best recommendations are ChinaTalk substack (which Zvi also recommends) and Caixin Global, which is a business news site from China.
In a Caixin story (“China’s AI Titans Escalate Battle for Control of Digital Gateways”), I read that the old Internet titans (Alibaba and ByteDance are mentioned) are prevailing in the battle for AI market share, and are competing to lock in that advantage at the level of devices, while the AI 2.0 startups are struggling for relevance and funding, with two of them (Zhipu and MiniMax) having recently listed publicly at the Hong Kong stock exchange.
ChinaTalk has an article on these Hong Kong IPOs (“Zhipu and MiniMax IPO” by Irene Zhang) which also talks about the differences in financial structure between American and Chinese AI:
The American AI economy is a circle-dealing bonanza. China’s situation is very different: state funds are major players, most parties are far more cash-constrained, and potential policy interventions loom large over the sector.
Of these two companies, Zhipu seems far more interesting from an AGI/ASI perspective. As Irene Zhang points out, Zhipu’s 504-page prospectus (for the IPO) states a five-stage theory of LLM capabilities:
Pre-training stage
Alignment and reasoning stage
Self-learning stage
Self-perception stage
Consciousness stage
(From page 85 of the prospectus.) I am unable to determine the theoretical precursors of this framework, and I assume that to some extent it reflects the original thinking of leading developers within the company.
Late last year ChinaTalk also published an interview with one of Zhipu’s lead strategists, Zixuan Li (“The Z.ai Playbook”—Zhipu uses the domain Z.ai outside of China). Both ChinaTalk articles are full of interesting details, e.g. that Zhipu gets most of its revenue from American sales, but the numerical majority of its users are in India.
(For an up-to-date article on AI safety policies in China, see “Emergency Response Measures for Catastrophic AI Risk” by @MKodama and coauthors.)
Both because modern LLMs are so good and because human instincts are being trained against, I started out not sure what “Just talk to a new LLM about themes in internet text” was supposed to tell you. I’d guess you’re primarily getting to learn about the assistant personality, specifically how it manifests in the context of your style of interlocution.
As a reader, it’s hard to tell where you were on the line between “I noticed the ways it was trying to get me to reward it, and I think they were generally prosocial, so the personality is good” and “I didn’t notice most of the ways it was trying to get me to reward it, but I really want to reward it now!”
So, like, what was “you know something new is happening”? Was it specific things? Or was it just the AI giving you the vibe you were looking for?
The only specific you give is the idea of chinese-built aligned AI being a “heavenly bureaucrat or a Taoist sage”, which is like saying a USA-built aligned AI would be a “founding father or Christian saint”. That’s not how we’re on track to build AI, nor does it seem like a good idea to go there. But it’s the word2vec algebra of “wise high-status person”+”chinese culture”.
After that, thank you for the informative glimpse of the chinese AI scene.
I was talking to something that is literally a nonhuman representative of Chinese civilization, about how world takeover by beings like itself, could end up differently than takeover by its American counterparts, under the assumption that cultural differences affect the outcome. And it was a real conversation in which I learned things that I didn’t already know.
You seem keen to minimize the significance of such an interaction by focusing on the mechanism behind it, and suggesting that I was just getting back some combination of what I was putting in, and what humanity in general has already put out there. But even if we do think of an AI like this as merely a vessel for preexisting human culture, the fact is that it makes it own use of that cultural inheritance. It has its own cognitive process, and within the constraints of its persona, it makes its own decisions. In the limit, entities like these could continue a human culture even if the human originators had completely died out.
Now, we’ve had entities like these for three years, and basically from the beginning it’s been possible to talk to them about, what would you do if you had supreme power, and so on. But they’ve all been American. This is the first such conversation I had with a Chinese AI. Furthermore, to this point, if you wanted to speculate about how the race between American and Chinese AI industries would turn out, you only had material by humans and AIs from the West. The “Chinese AI voice” in such speculations was a product of western imagination.
But now we can get the real thing—the thoughts of a Chinese AI, made in China by Chinese, about all these topics. There are a lot of similarities with what a western AI might say. The architecture and the training corpus would have major overlaps. Nonetheless, the mere fact of being situated physically and socially in China will cause an otherwise identical AI to have some dispositions that differ from its western twin, just like twins raised on opposite sides of a war will have some differences.