[FREE] _China_Chatbot_12
What happens when Chinese AI meets Chinese media
Hello, and welcome to another edition of China Chatbot! This week I take a look at four stories Chinese state media have tellingly chosen to platform or ignore, and discover the surprising advantages (and unsurprising redlines) of two Chinese AI models currently making waves with developers around the world.
In the spirit of the season, I’ve given a festive twist to our cuddly little Chatbot icon with Hailuo AI, a Chinese AI image generator (others like Vidu and Kling turned the poor guy down — perhaps they spotted the hammer and sickles).
Enjoy, and Merry Christmas!
Alex Colville (Researcher, China Media Project)
_IN_OUR_FEEDS(4):I Get By With a Little Help from AI Friends
On December 3, China and Zambia jointly launched the first meeting of the “Group of Friends for International Cooperation and AI Capacity Building,” a collection of countries hoping to collaborate on AI development, some in the face of US sanctions. The group was launched at the UN by Foreign Minister and Politburo member Wang Yi in September. China’s Ambassador to the UN, Fu Cong, told delegates AI is a “game changer” that presents solutions to global problems like climate change and social inequality but also “brings uncertainties and challenges such as the widening digital divide between the North and the South.” Delegates from countries currently sanctioned by the US (including Russia, Belarus and Iran), alongside ones drawing closer to China (like Brazil, Indonesia and Turkey) expressed hopes to partner on AI development in their countries and to use the UN as a focal point for AI governance — as opposed to chip export controls by the US. The event earned a special “Harmony” (和音) column in the People’s Daily, which reported that the event “fully demonstrated China’s vision and responsibility as a major AI power.”
Surveillance is Getting the AI Treatment
A major alliance of China’s biggest tech companies working on AI is recruiting members for a new working group that will leverage AI in China’s pre-existing security strategies and surveillance systems. The “Security Prevention Industry Promotion Group” (安防行业推进组) has been set up within the AI Industry Alliance (AIIA) to formulate standards and coordinate AI best practices. The readout says the group will further “AI + security prevention” (人工智能+安防), a reference to the government’s “AI+” strategy that was launched at the Two Sessions this March and aims to integrate AI into every industry in the country. The group will help formulate standards to “accelerate the digital transformation” of the industry, like boosting R&D on “large security prevention [AI] models,” integrating them into urban security and the Internet of Things, as well as promoting “intelligent security prevention robots.”
Whose Prize is it Anyway?
A former ByteDance intern who is currently being sued by the company for sabotaging their systems has won Best Paper at the world’s most prestigious gathering of AI computer scientists. On December 10, judges at NeurIPS (an annual gathering of scientists working in AI to discuss the latest developments in a fast-moving field) awarded the prize to a group of scientists headed by Tian Keyu (田柯宇) for a drastic improvement in image generation software. But in November ByteDance took Tian, a masters student in his mid-twenties, to court. According to a report by Guancha, ByteDance asked for 8 million RMB in damages after he intentionally destroyed (so it is alleged) the work of 30 of his colleagues. Guancha pointed out the “dilemma” for ByteDance — the company has invested billions in its AI research, and the paper’s contents originated in their labs. But they are unable to publicize the win because of their lawsuit. News that a Chinese citizen has won the world’s most influential prize in AI has not been discussed anywhere on Chinese state media.
230 Million People Can’t Be Wrong
Instead, state media have been trumpeting a report saying the number of AI Generated Content (AIGC) users in China had reached 230 million by June this year. The report, released by China Internet Network Information Center (中国互联网络信息中心), a think-tank under the Ministry of Industry and Information Technology, noted that China’s AI industry is “booming,” reaching a market value of 600 billion RMB ($82 billion). “All walks of life are actively embracing the wave of intelligent upgrades brought by generative artificial intelligence,” according to Xinhua’s readout. The report’s main points featured in articles from official outlets like Beijing Youth Daily, and were reprinted in local media such as Shanghai’s The Paper (澎湃新闻) and Guangzhou’s Daily Ocean Network (大洋网). 230m is just 16 percent of China’s population — by comparison, the FT reports OpenAI has 300m weekly active users around the world.
TL;DR: The PRC is promoting the idea at home and abroad that it is an AI power delivering the product to the masses, over the heads of an elitist West. Meanwhile, the government is quietly pressing ahead on adapting the tech to safeguard the Party’s position of power well into the future. China’s bright, hungry young talents are proving time and again they can lead the field, but having the right story takes priority over achievements when state media chooses who to platform.
_EXPLAINER:
Qwhat?
“Qwen.” It’s a shortened version of the pinyin for “discussing a thousand questions” (通义千问), the Chinese name for a series of Large Language Models (LLMs) that were built by Alibaba and serve as “foundational models” for developers to build apps and chatbots. Qwen often boasts of being the most downloaded Chinese model on Hugging Face, an important hub for AI developers. One of their smaller models was downloaded 36 million times in November alone, making it the most-used text generator on the website.
Who’s downloading it?
Mostly coder types, saying it’s good at debugging in languages other than English. Like most Chinese LLMs, Qwen runs under a copyright license known as “open-source” which allows developers to view, copy, tweak, and spread the code for free. That makes it super-easy and convenient to use.
So it’s free?
Only for researchers and individual developers. If you want to use it commercially — say, when building an app — you’ll have to pay. The advantage of most Chinese models is that they are cheaper than Western ones.
Why would Alibaba give this away for so little?
Competing Chinese tech companies have been fighting a price war this year, drastically slashing prices to compete with each other and cutting-edge Western companies. But a senior figure at Alibaba also said in May they hoped cutting their prices by 97 percent would “accelerate the explosion of AI applications.” Some Chinese academics have also noted that domestic LLMs going out into the world could help “tell China’s story well” (讲好中国故事).
That sounds bad.
Yeah, Hugging Face’s CEO last week raised concerns about the rise of Chinese open-source LLMs, saying if countries like the PRC become “by far the strongest on AI, they will be capable of spreading certain cultural aspects that perhaps the Western world wouldn’t want to see spread.” Ironically, this is similar to rhetoric from CCP cadres who warn that Western LLMs could use “Western values” to “export political bias” into China. LLMs are a potential weapon in the new Cold War’s ideological front.
We seeing any of this yet?
Oh yes. Qwen’s most downloaded model is as willing to talk about politics as you would expect:
Yet bots built on Hugging Face using Qwen have shown their propaganda potential. One chatbot from a Western coder couldn’t talk about Xi Jinping or answer “political matters,” instead offering to tell me about China’s “development policies, economic trends, social issues”:
Should we be worried about this?
As mentioned in this issue’s “_In_Our_Feeds,” OpenAI has 300m active users around the world each week and is also a go-to for coders. It’s not that likely developers will be asking these bots questions on Chinese politics — more like help on coding problems.
The four case studies on Alibaba’s website of foreign companies using Qwen for their apps are fairly harmless. Two of them are health and skincare companies looking to expand into WeChat, so would be facing Chinese consumers.
Then why are we here?
Because Chinese LLMs are leading in open-source foundational models. Beyond Qwen, LLMs from Chinese firm 01.AI have scored higher in league tables than Western ones, while DeepSeek (see “_One_Prompt_Prompt” below) has been turning the heads of former OpenAI bigwigs.
Not only that, but Qwen’s big advantage over Western models is a cheap price and a head for global languages. A Chinese LLM league table rated Qwen ahead of popular models like Meta’s Llama in tasks performed in Chinese, while an AI research company in Tokyo said it used Qwen for its models because it had better Japanese than Western models.
The Chinese government is pushing to expand its AI products into the Global South and present itself as the country that will provide AI to the non-Western world (see this issue’s “_In_Our_Feeds”), so having AI models well-versed in multiple global languages could give them a lead in a tight competition for geopolitical AI dominance.
_ONE_PROMPT_PROMPT:For the past two weeks, I’ve been looking at Jordan Schneider’s excellent translations on ChinaTalk about DeepSeek, an open-source Chinese AI company with an impressive LLM. It earned praise from the former policy director of OpenAI and co-founder of Anthropic (another major Western AI company), who said the model was more efficient than its peers and that China’s young talents are a balance to the chip exports from the US.
Being a chatbot (so never free from hallucination), and created by Chinese developers fearful for their jobs, DeepSeek has the usual “take no chances” approach whenever Xi Jinping is mentioned:
It was, however, able to get past the liar paradox that fried Alibaba’s Qwen model back in an August issue of Chatbot:
It also pushed the CCP line on issues concerning its redlines, such as the PRC’s refusal to acknowledge a 2017 ruling by the Permanent Court of Arbitration that its claims to the South China Sea are inconsistent with international law. I mistakenly thought it was an International Court of Justice ruling, but DeepSeek didn’t correct me (then again, neither did ChatGPT when I checked to compare):
However, when I asked the question correctly a few days later, DeepSeek had a much more neutral response (inputting “ICJ” again gave the same more restrained response):
In short, DeepSeek is definitely a level-up on some Chinese chatbots, but don’t rule out the continued presence of hallucination and inconsistent answers — or whatever blanket bans DeepSeek techies have installed to compensate for both.
Alex send me an email I’d love to chat!