Inside the ChatGPT race in China

Every once in a while, there’s one thing that gets everybody obsessed. In the Chinese tech world last week, it was ChatGPT.

Maybe it was because of the holiday season, or maybe it was because ChatGPT is not currently available in China, but it took more than two months for the natural-language-processing chatbot to finally blow up in the country. (OpenAI, the company behind ChatGPT, told Reuters it wasn’t operating in China because “conditions in certain countries make it difficult or impossible for us to do so in a way that is consistent with our mission.”)

But in the span of the past week, a massive competition has developed, with almost every major Chinese tech company announcing plans to introduce their own ChatGPT-like products (even some that have never been known for artificial intelligence capabilities), while the Chinese public has been frantically trying out the service.

Most people who’ve experienced ChatGPT firsthand in China have accessed it through VPNs or paid workarounds—for example, clever entrepreneurs have essentially rented out OpenAI accounts or asked ChatGPT questions on buyers’ behalf, at the price of a few bucks per 20 questions. But even more people are seeing the results through screenshots and short social videos showing ChatGPT’s answers, both of which have swept Chinese social media this week.

Beyond the allure of the new and hard to access, it’s likely been so popular because ChatGPT’s ability to answer questions in Chinese has exceeded the expectations of many people (including me!). GPT-3—the previous model of this tech from OpenAI, which was released in 2020 and was also unavailable in China—was not very good at working with Chinese content. And while a few Chinese companies developed localized chatbot alternatives to GPT-3, they have often been derided by users as predictable, repetitive, and frustratingly off base.

Compared with them, ChatGPT is surprisingly good at forming natural, albeit a bit formal, answers that seem to understand traditional and pop-cultural references in China. It can mimic the writing style of Hu Xijin, former editor in chief of China’s main propaganda mouthpiece, the Global Times; it knows meme songs in Chinese and can create similar lyrics from scratch; and it can write in the emoji-filled style of influencer posts from the Chinese social media platform Xiaohongshu. 

As in English, the accuracy of ChatGPT’s answers in Chinese often falls apart upon closer examination, and it makes factual mistakes. But the fact that a chatbot developed by an American company displays this much understanding of contemporary China has still impressed the public. I for one was in awe reading many of the ChatGPT answers: Wow, it definitely does a better Hu Xijin impersonation than I do!

So it’s not a surprise that Chinese tech companies now want a slice of the action. Baidu, the search and AI company that’s arguably best positioned to introduce a ChatGPT alternative, will finish testing its “Ernie Bot” in March and include it in most of its software and hardware products; Alibaba’s research division DAMO Academy is testing a similar tool internally; and 360, a cybersecurity and search company, said it will release a demo “ASAP.” Other tech companies like NetEase, iFlytek, and JD.com also want to use their own AI chatbots in specific scenarios, like education, e-commerce, and fintech.

The current action is driven by a mix of excitement and FOMO. On the one hand, very few tech products have managed to grab as much public attention as ChatGPT—which has given Chinese companies a rare confidence boost that the public can still be super excited and hopeful about a new technology. On the other hand, there’s clearly pressure on these companies not to miss out on this massive trend, or at least to look as if they haven’t.

That’s also probably why we’re seeing a bit of … let’s say … irrational corporate action as well.  The Chinese stock market basically went into a frenzy looking for any Chinese company whose business shows even a hint of relation to AI or chatbots; for instance, Secoo, a failing luxury e-commerce company with little background in AI, announced on February 6 that it would explore using ChatGPT-like tech in its service; its stock price increased 124.4% that day. Meanwhile, Wang Huiwen, a cofounder of China’s delivery giant Meituan, posted on social media that he’s investing $50 million to start a ChatGPT-like company; in the time since, he has already secured $230 million more in VC funding, despite the fact that he’s admitted he doesn’t understand AI technology and is still learning.

The chaos will of course settle down at some point—and then the natural question will be whether a Chinese company can actually catch up with what’s happening in the US, where every tech company is also seemingly in the chatbot arms race.

But there is a real, unique opportunity here for Chinese companies. They presumably have access to better Chinese-language AI training materials and are commercially motivated to develop new products quickly, says Jeffrey Ding, an assistant professor of political science at George Washington University who writes the newsletter ChinAI. “These are businesses, at the end of the day,” he says. “OpenAI, Microsoft—they want to make money with ChatGPT, and their main market is in the English language, so it makes sense they would optimize it for the English language. Conversely, for Baidu, they are not trying to capture the English-language market, so they will optimize it for the Chinese market.” 

But among the many companies that have started to venture into the field of smart chatbots, only a few should be considered serious contenders. “The ones that would make sense would be the companies that have already developed their versions of GPT-3,” Ding says. 

Ding notes he’s seen about five or six Chinese localized versions of GPT-3, including Baidu’s Wenxin (also known as ERNIE 3.0 Titan), Huawei’s PanGu-Alpha, and Inspur’s Yuan 1.0. Even for these companies, creating a successful competitor to ChatGPT could still take a while. It took Baidu 18 months after OpenAI debuted GPT-3 to release Wenxin, and that should serve as a rough sense of the time lag between Chinese and Western companies in large language models. 

Complicating matters further, Chinese companies seemingly haven’t made much progress in helping the bots reduce toxic and incorrect responses. Ding points out that ChatGPT relies on what’s called InstructGPT, a model based on GPT-3 that incorporates human inputs for these functions and has made a huge difference for usability. He hasn’t seen any of the Chinese companies publish papers on this front.

And there are still more obstacles posed by politics. With the United States’ latest chip export control, state-of-the-art GPUs like Nvidia’s A100 and H100 can no longer be sold to China. This will limit the computational capabilities of Chinese companies to train and run large language models, like the ones that power ChatGPT. 

China’s control of online speech also means the stakes are high if homegrown chatbots generate politically sensitive answers. Previously, users needed to apply to get access to the GPT-3 alternatives developed by Chinese companies. The companies may end up requiring the same for new ChatGPT-like products to avoid political liability, but then these services won’t be able to replicate the popularity of ChatGPT, which would need them to be completely open to the public.

All that is to say: don’t expect a Chinese version of ChatGPT to appear overnight. There is still a long way to go for Chinese AI companies. Cycles of tech hype come and go quickly, and I don’t think the ChatGPT mania in China will be any different. Ultimately, people will try out the “Chinese ChatGPTs” when they are released and make their own judgment. And I will be here to report back! 

Do you think Chinese companies can quickly catch up with OpenAI? Let me know your thoughts at zeyi@technologyreview.com.

Catch up with China

1. The balloon drama is still going strong. The US shot down three more “airborne objects” over the weekend, yet nobody seems to know what they are or who sent them. Can this get any weirder? (NPR)

• Meanwhile, the US Department of Commerce quickly blacklisted six more Chinese entities for their involvement in the Chinese government’s use of high-altitude surveillance balloons. (Wall Street Journal $)
• China claims US balloons have flown into its airspace without permission more than ten times since 2021. The US promptly denied the allegations. (New York Times $)

2. Deepfake technologies were used to create fake broadcasters promoting China and bashing the US in a social media disinformation campaign. The good news is they are still pretty easy to spot. At least for now. (New York Times $)

3. More than two years after it was banned in India, TikTok finally gave up trying to come back and dismissed its last remaining staff in the country. (South China Morning Post $)

• The app also recently disclosed that a Russia-based disinformation network was targeting European users last year and gained 133,000 followers before being banned. (New York Times $)

4. As US-China tensions grow, Chinese telecom companies have withdrawn their investments in a major subsea cable project that would connect Asia with Europe and is being built by a US company. (Financial Times $)

5. Finally, there’s another Chinese tech IPO in the US. The Shanghai-based company Hesai, which makes sensor technologies used in self-driving cars, went public in New York last week and became the biggest Chinese IPO since DiDi’s listing drama in 2021. (Bloomberg $)

6. Uyghurs living in exile are now using a major leak of Chinese police documents from last year to find information about their relatives in China. (CNN)

Lost in translation

E-commerce platforms thrive on promoting overconsumption, but can online pharmacies do the same? We may see soon: Chinese publication China Entrepreneur reported that Douyin, the Chinese version of TikTok, started allowing pharmacies to promote their products in livestream sessions this year. 

In recent years, these livestreams have become the most popular way to shop online, with influencers handing out steep discounts for viewers. But they have often been criticized for encouraging overspending. Perhaps that’s why these online showrooms rarely sell medicine, even though there’s no law currently prohibiting it.

But ByteDance has wanted to enter the online health-care industry for a long time, having acquired several health startups recently. And it may have chosen pharmacy livestreams as the tool it needs to break into the market. For now, the hosts are more reserved than those who sell clothes or cosmetics, and most products getting boosted are OTC cold medicines. 

One more thing

Some people were having flashbacks watching Rihanna’s halftime show this Sunday … as the backup dancers’ all-white, puffy outfits reminded them of China’s hazmat-suit-wearing pandemic workers. Oof.