The State of AI in China
Is China winning the race to AGI?
This is the first article I’ve written that has an accompanying video! If you want to hear me chat about this article, check the video out on YouTube. Likes and subscribes are appreciated. GIF preview below.
The artificial intelligence race has huge economic and societal ramifications. Economically, any nation with superintelligent AI will have an enormous economic advantage. Socially, any national with superintelligent AI will have the ability to influence other nations and their citizens via that AI. These societal and economic concerns along with the national security and privacy issues that come with AI will force governments to get involved (as many are or already have) and turn the AI into a nationalized race.
I’ve written previously comparing the AI race to the nuclear arms race. While not a perfect comparison, it puts the scope of this race into perspective. A better comparison to understand the influence AI can have is to think about what the internet would be like if it was entirely controlled by a single country and distributed throughout the world. That single country would have immense influences over others.
To understand the worldwide AI development landscape it’s important to understand the influence other countries have on it. I want to discuss China because:
Their political landscape is very different from western countries creating a very different AI development environment.
They’re developing models at the same rate as other frontrunners around the world.
They’re often seen in opposition to the other world superpowers developing AI which makes for an interesting competitive landscape.
The best place to start understanding the state of AI in China is to have a basic understanding of how the Chinese government works. I won’t diminish the information about the Chinese government to “China is bad” like many others do in the western world, instead I’ll focus on what those differences mean for AI advancement.
Fundamental Advantages
The Chinese Constitution reads that China is a socialist country run by the Chinese Communist Party (CCP). Most western countries are primarily capitalist, which means the individual human rights between and powers of government between the two are very different. The Chinese government is set up to centralize power and enable the government to enact regulation based on what they believe is best for China as a whole. This has received criticism from western countries because it often allows the government to enact policy at the expense of the individual.
I’m not going to dig deeper into this because it’s very nuanced and I’m not an economist or politician so if I’m missing important details, please leave a comment clarifying what I missed. A really simplified way to understand China’s governance in terms of AI development is anything owned or produced by a Chinese entity (think business or university) or individual can be used by the Chinese government however they deem necessary. I’ve previously written about this when I wrote about why TikTok’s privacy policy is concerning.
If you’ve been reading my articles, I hope you can quickly realize that this has massive implications for the development of artificial intelligence. Let’s get into them.
What does this mean for AI?
There are three primary blockers to artificial intelligence development:
Data: companies need the data to train machine learning models for research and commercial applications. The largest hurdle for AI development is having clean data that models can learn from.
Regulation: Government regulation has the ability to completely stifle innovation in any field. Regulation in the field of AI is no different.
Privacy: Any time data is being used for training, privacy is always a concern. Using private data can expose information and create personal or national security risks.
In the US, we’ve seen the difficulty of all three of these:
Large tech companies have come under fire for using training data training that had been scraped against terms and conditions (see above).
Regulation like SB 1047 has the potential to stifle the AI development landscape in an attempt to make AI “safer” for consumers. In reality, it has side effects that could be far worse.
Data privacy is such a concern for consumers there are guides for how they can ensure their data can’t be used for training AI systems.
China’s centralization of power makes it much easier for all three of these to be worked around. China has made it clear that leading the AI race is a national priority. The Chinese government can use regulation to ensure companies developing AI have access to the data they need and allow those companies to use the user data of Chinese citizens as necessary.
Essentially what we’re seeing is Chinese regulation pushing their nation forward in AI development while the US is fighting regulation to keep developing AI as it threatens to stunt growth.
Disadvantages for Chinese AI
So why is China just now catching up in the generative AI model race and what hinders their future development? Many of the advantages China has in the AI race also come with disadvantages in the form of embargoes and tariffs from other countries resulting from China’s unfair trade and anti-competitive practices. China faces difficulties regarding:
Import restrictions and tariffs: President Biden has banned the export of “advanced chips” to China including the GPUs necessary for machine learning. This was mentioned in Nvidia’s 8-K filing in October 2023. This means Chinese companies aren’t able to get their hands on the same level of state-of-the-art compute as US companies.
Investment restrictions: there are investment restrictions from US-based investment firms in Chinese companies making it difficult for Chinese companies to participate in the massive AI venture capital boom.
Sanctions against using specific software and hardware: sanctions have been put on Chinese companies using certain AI software and hardware tools. For the most part, software limitations are negligible because there are alternatives or alternatives that can easily be built. Hardware limitations are a major hindrance.
Chinese AI development gets around these sanctions (as best they can) by:
Starting domestic chip development to alleviate the burden caused by import restrictions. This takes a while to ramp up, but China has the means to do it. It remains to be seen if their AI chips will be as efficient as the current state-of-the-art chips developed by US companies.
Centralizing data control to alleviate one of the biggest hindrances to AI development.
Government support for AI has been established. While the US government fights AI development, the Chinese government pushes advancements forward.
Chinese research teams and companies utilize and develop open AI models. This allows them to take advantage of the momentum created by other researchers building state-of-the-art open AI.
Currently, it seems like the methods China is using to get around their AI development difficulties has made it difficult for them to keep up with western development. However, we don’t know what the future holds and this gap seems to be narrowing as an open Chinese LLM recently topped the Open LLM Leaderboard 2. It’s important to understand the AI landscape across the world if countries want to stay competitive.
One question from a subscriber that remains after going through all this is: How will China navigate the tricky scenario of maintaining tight censorship with training self-sufficient AI systems? I don’t know how they’ll do this exactly, but it’ll have to do with restricting the AI available to Chinese citizens and maintaining control over it. I think this is something they’re still figuring out and a reason they’ve made it a national goal to lead the world’s AI development.
Originally I had planned on writing the above information and diving into the different Chinese AI models but I quickly realized that was its own rabbit hole. I apologize if anything above is unclear. If so, leave a comment with a question and I’ll get to it. :)
Thanks for reading!
Always be (machine) learning,
Logan
Resources
Here are the resources I read through while preparing this if you want to keep digging:
Chinese netizens have to resort to Roman initials to be able to go around the censor bots. That's the data available behind the ggeat fire wall