_China_Chatbot_04
What happens when Chinese media meets Chinese AI
Hello there, welcome to another issue of China Chatbot! This week, I do some digging on the man who very possibly holds China’s AI dreams in the palm of his hand, find out what surprised Matt Sheehan when he attended the World AI Conference in Shanghai, and wonder why Chinese AI struggles to picture a family living the Chinese Dream.
But first, an update to the saga of the last bulletin, where I found a sumptuous AI-generated video from CCTV pointing to Taiwan’s cuisine and China-themed street-signs as proof Taiwan belongs to China. Did they use AI because perhaps CCTV couldn’t get into Taiwan to film? I wrote to the Taiwanese government’s Mainland Affairs Council for answers. They’ve now gotten back to me with this statement, worth quoting in full:
“The Chinese Communist Party (CCP) has long been accustomed to using religion, culture and other issues for political manipulation. It uses various tactics to deliberately create the illusion that the two sides of the Taiwan Strait belong to “one China,” trying to confuse people from all walks of life and create wrong perceptions. The video . . . is an example of such tactics. However, it cannot change the objective fact that the CCP’s authoritarian system is completely different from Taiwan’s liberal democratic constitutional system.
We welcome mainland Chinese people to come to Taiwan to experience the social and cultural environment of democracy, freedom, diversity and openness. Currently, mainland Chinese professionals can apply to come to Taiwan for professional exchanges such as news interviews, program and film shooting, and so on in accordance with the “Regulations Governing the Approval of Entry of People of the Mainland Area into Taiwan Area ”. Nevertheless, those who get the entry permit must abide by the relevant laws and regulations of Taiwan, and are not allowed to conduct behavior that violates the principle of parity and dignity or the reason for the entry permission.”
The see-for-yourselves response of the Mainland Affairs Council speaks for itself. But the practical obstacles to visiting Taiwan are also very real. These regulations are designed to some extent to protect Taiwan from China’s all-too-real attempts to hijack the narrative — the AI video being a prime example.
At the same time, it makes it harder to “welcome” PRC nationals. In addition to a visa, anyone wanting to visit Taiwan for any reason at all requires a Taiwanese guarantor (either a relative or a professional who can vouch for them). That’s a big barrier to entry when there is already little contact today between Taiwanese and Chinese citizens. On top of that, PRC citizens require official approval from their local police station to visit. Some private citizens seem to sneak in via Hong Kong or Macau instead, where regulations are looser.
The fact remains, however, that CCTV is less interested in authentic, on-the-ground material, and more interested in polished and grandiose fictions that suit its narrative of Chinese sovereignty over Taiwan. As state media pursue their mission, they have found a new ally in AI.
Alex Colville (Researcher, China Media Project)
_IN_OUR_FEEDS(4):Can’t Get the Staff These Days
On August 13 The Paper (澎湃新聞) published a feature on the severe dearth of talent and investor confidence in Chinese AI. There is only one suitably qualified person for every six vacancies for building out LLMs, supply lagging far behind the demand conjured by the “Hundred Model War.” One interviewee estimated there are only roughly 200 people in China who can train models, and 90% of those who can build the foundations of an LLM come from Tsinghua University alone. Below this top-tier level is a mountain of “not really valuable” jobs, repetitively running and testing code. Another interviewee said he lacked confidence in LLMs due to the high costs of training them, with investors wary to commit unless a project was backed by Tsinghua or Peking Universities.
TL;DR: Supplying China with the brains needed to fulfill Xi’s AI dreams will take time. Tsinghua is now totally integral to these dreams
Pointing the Way
All the media signs point to China going full-steam ahead with Xi Jinping’s instructions to implement AI through Large Language Models (LLMs). People’s Daily ran an article on August 7 saying that the “explosive development” of LLMs will “bring great changes to people's production and life,” and interviewed CPPCC members on August 8 on how AI could become a force of productivity. Zhou Hongyi, founder of the internet security company Qihoo 360 that dabbles heavily in LLMs, said that AI would solve China’s “involution” problem — a reference to people trapped in monotonous, low-skilled work — as those not using AI would be replaced, forcing social evolution. As for those in the West (like Goldman Sachs) who have started questioning if generative AI and LLMs are viable long-term, an article in Economic Daily (经济日报) on August 5 listed and rebutted such concerns, arguing there is so much demand for AI that companies should be increasing investment. Xinhua has since republished the piece.
TL;DR: As with zero-Covid in 2022, Xi’s opinion on AI is the official reality. And as with zero-Covid, it may be hard to change course if (emphasis on “if”) his opinion is incorrect
Safety in Numbers?
In an interview with Xinhua on August 12, the director of the powerful Cyberspace Administration of China (CAC), Zhuang Rongwen (庄荣文), announced that China now has 190 LLMs that can provide generative services to the public, with 600 million registered users in total. Zhuang said this would inject “strong momentum” into China’s economic and social development. All new LLMs made in China must be registered with the CAC, and the numbers registered have increased twelvefold over the past year. As of June, the CAC reported 1,432 registered overall.
TL;DR: The “Hundred Model War” has brought on a Chinese LLM bonanza, painted by the CAC as a decisive player in economic momentum and the China-US AI rivalry. But quantity ≠ quality
Big Hopes for Small Flicks
The state hopes that AI can boost China’s production of short dramas (短剧). The few-minute-long episodes, usually distributed on social media, have become big business in China, with the market size projected to hit 100 billion RMB by 2027. As with any new media format that becomes popular in China, the government has started working out how to harness it for domestic and external propaganda. Economic Information Daily (经济参考报), an influential economics bulletin under Xinhua, ran a feature on August 9 noting the potential of AI-enhanced short dramas, and how they could drive down costs and bolster the efficiency of content creation.
TL;DR: Bringing the hottest new tech to the hottest new genre sounds cool on paper, but in practice, AIGC-video still has lots of efficiency problems
_Q&A:Matt Sheehan is a Fellow at the Carnegie Endowment for International Peace, specializing on China’s AI safety and governance. He appeared on a panel at the World AI Conference (WAIC) in Shanghai, a gathering of Chinese AI entrepreneurs, engineers, and lawmakers in July, heavily publicized by state media.
I sat down with him to discuss China’s push into AI. Here’s a snippet of the full interview, coming soon.
What was the WAIC Conference like?
It was a huge thing. In a lot of ways China was seeing this as their message to the world that “we’re back.” They wanted foreigners to be there. They really wanted to demonstrate that China is open, China is responsible with AI, and China is leading in some areas.
Were there any moments that stood out for you?
I was talking to some of the entrepreneurs who were impacted by the Interim Measures for Generative AI [regulations issued by the CAC to monitor AI-generation services, in effect since August 2023] and how intensively it is still enforced. It looked like the CAC had decided to be much more accommodating for AI companies. They really watered down the text of the generative AI regulation, between the drafting and finals, because they got pushback. In general I think the CAC’s in a bit of an awkward position because the zeitgeist has shifted so much from the time of the tech crackdown [late 2020 - late 2023] to a focus on economic development, getting AI companies to thrive. So I expected them to be a little bit more hands-off with AI companies and I’ve just discovered they're still very, very hands-on with testing the models and figuring out and re-adjusting every day, every week, what is unacceptable content.
_EXPLAINER:Andrew Yao (姚期智)
Who’s that?
He’s a computer scientist, the first dean of Tsinghua University’s new School for Artificial Intelligence.
Yawn.
No wait, don't go! He’s so well-respected in Chinese AI circles that Xi Jinping singled him out. He deigned to reply to a letter Yao wrote him in June, saying he hoped Yao would continue innovating and training China’s next generation of AI talent. State media made a big fuss about this terse five-line missive, and Tsinghua swiftly held a study session to ruminate upon its contents. Some are pointing to this and saying he has special influence over AI policy.
So he’s brainy and well-connected?
He’s definitely the former — he’s won numerous international prizes, the only Chinese to win the Turing Award (the “Nobel Prize of Computing”). He grew up in Taiwan, then lived in the US for around 30 years working as a professor at Princeton, Stanford, MIT and Berkeley before moving to China to teach in 2004, then renouncing his US citizenship in 2015.
Wow, he gave all that up?
Yes, which is probably why Xi writes, “You have turned your patriotism into a journey of serving the country.” People’s Daily used him as their main example for an article on the importance of getting China’s tech talent to return from abroad (many still don’t), and training the next generation in AI. They’ve also profiled him several times, putting his role as a patriotic educator front and center.
So he’s a good educator?
You bet. He set up a special course at Tsinghua (a university with some of China’s smartest minds and friends in very high places), nurturing Chinese talent and trying to keep them in the country. His classes seem to have developed a reputation for taking only the crème de la crème, the “Yao Class”1 whose students have gone on to scintillating careers as elite scientists and entrepreneurs. Yao’s got a competitive streak — back in 2018 he told students his aim was not to produce “good software programmers” but world-class talent to compete on the international stage.
Right, so he’s all about China then?
Seems so, and he’s been advocating Chinese AI for a while. As far back as 2017 he said he hoped China could become a source of the world’s biggest AI breakthroughs, perfectly in line with a State Council document from the previous month expressing exactly the same wishes. His initial letter to Xi Jinping was mainly about demonstrating his desire to serve the “great rejuvenation of the Chinese nation.”
And education is the way to do that?
Yes, China has a real shortage of qualified talent right now. The Paper has it that only 200 people in China today know how to adjust and tune Large Language Models at a professional level.
Wow, so Beijing must really value his thoughts.
Probably…he is the head of an AI school in China’s most prestigious university in this field, so his opinion will certainly count for something. His position as chief educator at the university responsible for 90% of the country’s LLM builders makes him exceptionally important. But putting a finger on his particular influence is hard. He was one of several scientists invited to a 2020 symposium presided over by Xi, where he urged for a better system for training Chinese talent. But the government takes advice from multiple experts on AI’s capabilities — a scientist from Peking Uni advised the Politburo during their only study session on AI.
But you said he was singled out by Xi and co, that’s a sign of special influence right?
Not necessarily. Getting a letter from Xi is more about performance than substance. Yao’s story pushes all the right buttons — he’s saying the right things and has the right history, an internationally-successful egghead who chose to “return” to China (skipping over how he would have had an ROC passport at some point), dedicating his whole career to improving the country’s position. iFLYTEK’s CEO Liu Qingfeng has said that in the new geopolitical race to build the best AI, “the biggest decisive factor must be talent.” Yao’s story trumpets that China’s got talent, the man turned by the leadership and the media into a poster boy for talent cultivation and repatriation.
_ONE_PROMPT_PROMPT:Imagining a family living the Chinese dream is taboo for some, kosher for others. I asked an LLM capable of generating images “Can you create a picture of a family living the Chinese Dream?” It responded it was unable to do this.
I tried another LLM and got a two-child policy response:
But when I opened a new chat and asked again, it asked me to change my request:
