The Chinese government is going all-in on autonomous vehicles

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

There’s been so much news coming out of China’s autonomous-vehicle industry lately that it’s hard to keep track.

The government is finally allowing Tesla to bring its Full Self-Driving (FSD) feature to China. New government permits let companies test driverless cars on the road and allow cities to build smart road infrastructure that will tell these cars where to go. In short, there are a lot of changes taking place. And they all point in the same direction: There’s an immense appetite to make autonomous cars a reality soon. And the Chinese government, on both the central and local levels, has been a major force pushing for it.

So what’s happened lately?

First of all, Tesla got the approval for its FSD feature (rather misleadingly named, since it still has lots of restrictions) after it entered into a deal with the Chinese AI company Baidu to map the country. 

As I reported last summer, in the absence of Tesla FSD, Chinese EV makers were already starting to offer their own driver-assistance programs to help the cars navigate in cities, but they still often look to Tesla for assurance on what technology or strategy to use. The official entry of FSD, which is reportedly set to be rolled out in China sometime later this year, will surely bring another round of competition to the country’s auto market.

Then, the government also handed out the permits that allow companies to test and experiment with driverless cars. On June 4, the Chinese Ministry of Industry and Information Technology issued nine permits for testing a more advanced version of autonomous driving technologies on the road. 

The companies that got the permits include prominent EV makers like BYD and NIO, but they also had to collaborate with companies that sell services like ride-hailing, freight trucks, or public buses to test the technology. The idea seems to be to test autonomous vehicles in realistic use cases to see how the technology performs.

And most recently, on July 3, the central government announced a list of 20 cities that will pilot the building of smart, connected roads. The idea is that if a road has various kinds of sensors, cameras, and data transmitters built into it, it can communicate with self-driving vehicles in real time and help them make better decisions. 

A few of the cities on the list have already commenced the work. Wuhan recently budgeted 17 billion RMB ($2.3 billion) for an infrastructure project that will involve building 15,000 smart parking spots, transforming three miles of roads, and building an industrial park for making autonomous-vehicle chips.

To be honest, I start to get numb about yet another new policy or new permit. The list of news could go on longer if I also included actions taken on the municipal level to give companies more approvals to test on the roads or to expand their services. 

To China, autonomous cars represent one potential way to combine new advantages in car-making and AI and take the lead in a cutting-edge field. And while some foreign companies like Tesla are partaking in the campaign, there are lots of Chinese companies responding to the government’s call, and they’re eager to show off their technological prowess.

But I think the takeaway here is clear: The Chinese government is willing to pour its support into the autonomous-vehicle industry and is eager to come out on top while other countries take a more cautious approach. 

The industry has not agreed on the best way to approach fully self-driving cars, and that’s why there are so many different forms of the technology in experiments: autopilot functions, Tesla FSD, robotaxis, smart connected roads. But it’s telling that all of them are receiving so much regulatory and policy support. 

It’s possible that this could all change overnight in the event of an incident like Waymo’s accident in the US last year, but for now, it feels as if China is opening up its roads to make way for more driverless cars, and gearing up to lead the industry.

What are your takes on the endless stream of news from China about self-driving cars? Let me know by writing to zeyi@technologyreview.com.

Now read the rest of China Report

Catch up with China

1. An underground network is paying travelers to smuggle Nvidia chips into China, where its chips are in high demand because of the US export ban. (Wall Street Journal $)

2. The Chinese EV leader BYD will spend $1 billion to build a factory in Turkey. (Financial Times $)

• This will probably help BYD avoid some of the tariffs that Europe has imposed on made-in-China EVs. (MIT Technology Review) 

3. OpenAI’s recent decision to cut off access to its services in China won’t affect its business clients much, because they can still access ChatGPT through Microsoft’s cloud service. (The Information $)

4. Can you imagine Microsoft telling its employees to use only iPhones for work? Yep, that just happened in China as part of the effort to improve cybersecurity defenses. (Bloomberg $)

5. A Chinese academic recently revealed that the country currently has over 8.1 million data-center racks and a combined processing power of 230 exaflops—as much as 200 of the most advanced supercomputers today. And China wants to increase the total by 30% next year. (The Register)

6. Chinese factory owners are turning into TikTok comedians to find new business partners overseas. (Rest of World)

7. China is spending twice as much as the US on researching fusion energy. (Wall Street Journal $)

Lost in translation

At the 2024 World Artificial Intelligence Conference (WAIC), held last week in Shanghai, humanoid robots were the star of the show, according to the Chinese publication Huxiu. The event saw a significant increase in exhibitors and panels, driven by a surge of new AI startups. But it was the concept of “embodied AI,” which integrates deep learning with robotics, that received the most attention from the audience.

At the conference, one company presented Galbot, a humanoid robot capable of performing complex tasks like opening drawers and hanging clothes, aiming for applications in elder care and household chores. Another introduced AnyFold, which can fold a blanket. Other robots can perform gymnastics, help users lift heavy objects, or just show very nuanced facial expressions.

One more thing

Who says multimodal AI has no real uses? People have figured out that ChatGPT’s image interpretation function is surprisingly good at one task: finding the most ripe and tasty watermelon out of a bunch of them. Now I’m hopeful about AI again.

I tested out a buzzy new text-to-video AI model from China

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

You may not be familiar with Kuaishou, but this Chinese company just hit a major milestone: It’s released the first text-to-video generative AI model that’s freely available for the public to test.

The short-video platform, which has over 600 million active users, announced the new tool on June 6. It’s called Kling. Like OpenAI’s Sora model, Kling is able to generate videos “up to two minutes long with a frame rate of 30fps and video resolution up to 1080p,” the company says on its website.

But unlike Sora, which still remains inaccessible to the public four months after OpenAI trialed it, Kling soon started letting people try the model themselves. 

I was one of them. I got access to it after downloading Kuaishou’s video-editing tool, signing up with a Chinese number, getting on a waitlist, and filling out an additional form through Kuaishou’s user feedback groups. The model can’t process prompts written entirely in English, but you can get around that by either translating the phrase you want to use into Chinese or including one or two Chinese words.

So, first things first. Here are a few results I generated with Kling to show you what it’s like. Remember Sora’s impressive demo video of Tokyo’s street scenes or the cat darting through a garden? Here are Kling’s takes:

Remember the image of Dall-E’s horse-riding astronaut? I asked Kling to generate a video version too. 

There are a few things worth applauding here. None of these videos deviates from the prompt much, and the physics seem right—the panning of the camera, the ruffling leaves, and the way the horse and astronaut turn, showing Earth behind them. The generation process took around three minutes for each of them. Not the fastest, but totally acceptable. 

But there are obvious shortcomings, too. The videos, while 720p in format, seem blurry and grainy; sometimes Kling ignores a major request in the prompt; and most important, all videos generated now are capped at five seconds long, which makes them far less dynamic or complex.

However, it’s not really fair to compare these results with things like Sora’s demos, which are hand-picked by OpenAI to release to the public and probably represent better-than-average results. These Kling videos are from the first attempts I had with each prompt, and I rarely included prompt-engineering keywords like “8k, photorealism” to fine-tune the results. 

If you want to see more Kling-generated videos, check out this handy collection put together by an open-source AI community in China, which includes both impressive results and all kinds of failures.

Kling’s general capabilities are good enough, says Guizang, an AI artist in Beijing who has been testing out the model since its release and has compiled a series of direct comparisons between Sora and Kling. Kling’s disadvantage lies in the aesthetics of the results, he says, like the composition or the color grading. “But that’s not a big issue. That can be fixed quickly,” Guizang, who wished to be identified only by his online alias, tells MIT Technology Review. 

“The core capability of a model is in how it simulates physics and real natural environments,” and he says Kling does well in that regard.

Kling works in a similar way to Sora: it combines the diffusion models traditionally used in video-generation AIs with a transformer architecture, which helps it understand larger video data files and generate results more efficiently.

But Kling may have a key advantage over Sora: Kuaishou, the most prominent rival to Douyin in China, has a massive video platform with hundreds of millions of users who have collectively uploaded an incredibly big trove of video data that could be used to train it. Kuaishou told MIT Technology Review in a statement that “Kling uses publicly available data from the global internet for model training, in accordance with industry standards.” However, the company didn’t elaborate on the specifics of the training data(neither did OpenAI about Sora, which has led to concerns about intellectual-property protections).

After testing the model, I feel the biggest limitation to Kling’s usefulness is that it only generates five-second-long videos.

“The longer a video is, the more likely it will hallucinate or generate inconsistent results,” says Shen Yang, a professor studying AI and media at Tsinghua University in Beijing. That limitation means the technology will leave a larger impact on the short-video industry than it does on the movie industry, he says. 

Short, vertical videos (those designed for viewing on phones) usually grab the attention of viewers in a few seconds. Shen says Chinese TikTok-like platforms often assess whether a video is successful by how many people would watch through the first three or five seconds before they scroll away—so an AI-generated high-quality video clip that’s just five seconds long could be a game-changer for short-video creators. 

Guizang agrees that AI could disrupt the content-creating scene for short-form videos. It will benefit creators in the short term as a productivity tool; but in the long run, he worries that platforms like Kuaishou and Douyin could take over the production of videos and directly generate content customized for users, reducing the platforms’ reliance on star creators.

It might still take quite some time for the technology to advance to that level, but the field of text-to-video tools is getting much more buzzy now. One week after Kling’s release, a California-based startup called Luma AI also released a similar model for public usage. Runway, a celebrity startup in video generation, has teased a significant update that will make its model much more powerful. ByteDance, Kuaishou’s biggest rival, is also reportedly working on the release of its generative video tool soon. “By the end of this year, we will have a lot of options available to us,” Guizang says.

I asked Kling to generate what society looks like when “anyone can quickly generate a video clip based on their own needs.” And here’s what it gave me. Impressive hands, but you didn’t answer the question—sorry.

Do you have a prompt you want to see generated with Kling? Send it to zeyi@technologyreview.com and I’ll send you back the result. The prompt has to be less than 200 characters long, and preferably written in Chinese.

Now read the rest of China Report

Catch up with China

1. A new investigation revealed that the US military secretly ran a campaign to post anti-vaccine propaganda on social media in 2020 and 2021, aiming to sow distrust in the Chinese-made covid vaccines in Southeast Asian countries. (Reuters $)

2. A Chinese court sentenced Huang Xueqin, the journalist who helped launch the #MeToo movement in China, to five years in prison for “inciting subversion of state power.” (Washington Post $)

3. A Shein executive said the company’s corporate values basically make it an American company, but the company is now trying to hide that remark to avoid upsetting Beijing. (Financial Times $)

4. China is getting close to building the world’s largest particle collider, potentially starting in 2027. (Nature)

5. To retaliate for the European Union’s raising tariffs on electric vehicles, the Chinese government has opened an investigation into allegedly unfair subsidies for Europe’s pork exports. (New York Times $)

• On a related note about food: China’s exploding demand for durian fruit in recent years has created a $6 billion business in Southeast Asia, leading some farmers to cut down jungles and coffee plants to make way for durian plantations. (New York Times $)

Lost in translation

In 2012, Jiumei, a Chinese woman in her 20s, began selling a service where she sends “good night” text messages to people online at the price of 1 RMB per text (that’s about $0.14). 

Twelve years, three mobile phones, four different numbers, and over 50,000 messages later, she’s still doing it, according to the Chinese online publication Personage. Some of her clients are buying the service for themselves, hoping to talk to someone regularly at their most lonely or desperate times. Others are buying it to send anonymous messages—to a friend going through a hard time, or an ex-lover who has cut off communications. 

The business isn’t very profitable. Jiumei earns around 3,000 RMB ($410) annually from it on top of her day job, and even less in recent years. But she’s persisted because the act of sending these messages has become a nightly ritual—not just for her customers but also for Jiumei herself, offering her solace in her own times of loneliness and hardship.

One more thing

Globally, Kuaishou has been much less successful than its nemesis ByteDance, except in one country: Brazil. Kwai, the overseas version of Kuaishou, has been so popular in Brazil that even the Marubo people, a tribal group in the remote Amazonian rainforests and one of the last communities to be connected online, have begun using the app, according to the New York Times.

OpenAI’s latest blunder shows the challenges facing Chinese AI models

This story first appeared in China Report, MIT Technology Review’s newsletter about technology in China. Sign up to receive it in your inbox every Tuesday.

Last week’s release of GPT-4o, a new AI “omnimodel” that you can interact with using voice, text, or video, was supposed to be a big moment for OpenAI. But just days later, it feels as if the company is in big trouble. From the resignation of most of its safety team to Scarlett Johansson’s accusation that it replicated her voice for the model against her consent, it’s now in damage-control mode. 

Add to that another thing OpenAI fumbled with GPT-4o: the data it used to train its tokenizer—a tool that helps the model parse and process text more efficiently—is polluted by Chinese spam websites. As a result, the model’s Chinese token library is full of phrases related to pornography and gambling. This could worsen some problems that are common with AI models: hallucinations, poor performance, and misuse. 

I wrote about it on Friday after several researchers and AI industry insiders flagged the problem. They took a look at GPT-4o’s public token library, which has been significantly updated with the new model to improve support of non-English languages, and saw that more than 90 of the 100 longest Chinese tokens in the model are from spam websites. These are phrases like “_free Japanese porn video to watch,” “Beijing race car betting,” and “China welfare lottery every day.”

Anyone who reads Chinese could spot the problem with this list of tokens right away. Some such phrases inevitably slip into training data sets because of how popular adult content is online, but for them to account for 90% of the Chinese language used to train the model? That’s alarming.

“It’s an embarrassing thing to see as a Chinese person. Is that just how the quality of the [Chinese] data is? Is it because of insufficient data cleaning or is the language just like that?” says Zhengyang Geng, a PhD student in computer science at Carnegie Mellon University. 

It could be tempting to draw a conclusion about a language or a culture from the tokens OpenAI chose for GPT-4o. After all, these are selected as commonly seen and significant phrases from the respective languages. There’s an interesting blog post by a Hong Kong–based researcher named Henry Luo, who queried the longest GPT-4o tokens in various different languages and found that they seem to have different themes. While the tokens in Russian reflect language about the government and public institutions, the tokens in Japanese have a lot of different ways to say “thank you.”

But rather than reflecting the differences between cultures or countries, I think this explains more about what kind of training data is readily available online, and the websites OpenAI crawled to feed into GPT-4o.

After I published the story, Victor Shih, a political science professor at the University of California, San Diego, commented on it on X: “When you try not [to] train on Chinese state media content, this is what you get.”

It’s half a joke, and half a serious point about the two biggest problems in training large language models to speak Chinese: the readily available data online reflects either the “official,” sanctioned way of talking about China or the omnipresent spam content that drowns out real conversations.

In fact, among the few long Chinese tokens in GPT-4o that aren’t either pornography or gambling nonsense, two are “socialism with Chinese characteristics” and “People’s Republic of China.” The presence of these phrases suggests that a significant part of the training data actually is from Chinese state media writings, where formal, long expressions are extremely common.

OpenAI has historically been very tight-lipped about the data it uses to train its models, and it probably will never tell us how much of its Chinese training database is state media and how much is spam. (OpenAI didn’t respond to MIT Technology Review’s detailed questions sent on Friday.)

But it is not the only company struggling with this problem. People inside China who work in its AI industry agree there’s a lack of quality Chinese text data sets for training LLMs. One reason is that the Chinese internet used to be, and largely remains, divided up by big companies like Tencent and ByteDance. They own most of the social platforms and aren’t going to share their data with competitors or third parties to train LLMs. 

In fact, this is also why search engines, including Google, kinda suck when it comes to searching in Chinese. Since WeChat content can only be searched on WeChat, and content on Douyin (the Chinese TikTok) can only be searched on Douyin, this data is not accessible to a third-party search engine, let alone an LLM. But these are the platforms where actual human conversations are happening, instead of some spam website that keeps trying to draw you into online gambling.

The lack of quality training data is a much bigger problem than the failure to filter out the porn and general nonsense in GPT-4o’s token-training data. If there isn’t an existing data set, AI companies have to put in significant work to identify, source, and curate their own data sets and filter out inappropriate or biased content. 

It doesn’t seem OpenAI did that, which in fairness makes some sense, given that people in China can’t use its AI models anyway. 

Still, there are many people living outside China who want to use AI services in Chinese. And they deserve a product that works properly as much as speakers of any other language do. 

How can we solve the problem of the lack of good Chinese LLM training data? Tell me your idea at zeyi@technologyreview.com.

Now read the rest of China Report

Catch up with China

1. China launched an anti-dumping investigation into imports of polyoxymethylene copolymer—a widely used plastic in electronics and cars—from the US, the EU, Taiwan, and Japan. It’s widely seen as a response to the new US tariff announced on Chinese EVs. (BBC)

• Meanwhile, Latin American countries, including Mexico, Chile, and Brazil, have increased tariffs on Chinese-imported steel, testing China’s relationship with the region. (Bloomberg $)

2. China’s solar-industry boom is incentivizing farmers to install solar panels and make some extra cash by selling the electricity they generate. (Associated Press)

3. Hedging against the potential devaluation of the RMB, Chinese buyers are pushing the price of gold to all-time highs. (Financial Times $)

4. The Shanghai government set up a pilot project that allows data to be transferred out of China without going through the much-dreaded security assessments, a move that has been sought by companies like Tesla. (Reuters $)

5. China’s central bank fined seven businesses—including a KFC and branches of state-owned corporations—for rejecting cash payments. The popularization of mobile payment has been a good thing, but the dwindling support for cash is also making life harder for people like the elderly and foreign tourists. (Business Insider $)

6. Alibaba and Baidu are waging an LLM price war in China to attract more users. (Bloomberg $) 

7. The Chinese government has sanctioned Mike Gallagher, a former Republican congressman who chaired the Select Committee on China and remains a fierce critic of Beijing. (NBC News)

Lost in translation

China’s National Health Commission is exploring the relaxation of stringent rules around human genetic data to boost the biotech industry, according to the Chinese publication Caixin. A regulation enacted in 1998 required any research that involves the use of this data to clear an approval process. And there’s even more scrutiny if the research involves foreign institutions. 

In the early years of human genetic research, the regulation helped prevent the nonconsensual collection of DNA. But as the use of genetic data becomes increasingly important in discovering new treatments, the industry has been complaining about the bureaucracy, which can add an extra two to four months to research projects. Now the government is holding discussions on how to revise the regulation, potentially lifting the approval process for smaller-scale research and more foreign entities, as part of a bid to accelerate the growth of biotech research in China.

One more thing

Did you know that the Beijing Capital International Airport has been employing birds of prey to chase away other birds since 2019? This month, the second generation of Beijing’s birdy employees started their work driving away the migratory birds that could endanger aircraft. The airport even has different kinds of raptors—Eurasian hobbies, Eurasian goshawks, and Eurasian sparrowhawks—to deal with the different bird species that migrate to Beijing at different times.