How China is building a parallel generative AI universe

How China is building a parallel generative AI universe

The technological revolution is huge The leaps machine learning models have made in the last few weeks is inspiring everyone to be excited about the future. Nervous Its uncomfortable consequences. ChatGPT’s ability for intelligent conversations is becoming a cult following the success of text-to-image tools from StabilityAI and OpenAI.

China’s tech community has been closely following the West’s progress, so entrepreneurs, investors, and researchers are seeking ways to get in on the generative AI market. Open-source models are being used by tech firms to attract enterprise and consumer customers. Individuals are making a lot of AI-generated content. Regulators quickly responded to clarify how text, image, or video synthesis should be used. Concerns about China’s ability keep up with AI advancement are being raised by U.S. tech sanctions.

Let’s take a look in China at how generative AI is sweeping the globe towards the end 2022.

Chinese flavors

Stable Diffusion, DALL-E 2 and other viral art creation platforms have made it easy for everyone to notice that generative AI is now a hot topic. Chinese tech giants are half-way around the world. They have also captured the public’s attention with their similar products, adapting to the country’s taste and political climate.

Baidu, which was first discovered through search engines, has been a prominent player in the search engine market in recent years. Autonomous driving is a new area of focus for the company., ERNIE-ViLG, Model with 10 billion parameters It was trained using a data set that contained 145 million Chinese image text pairs. How does it compare to its American counterpart? Here are the results of the prompt “kids in New York Chinatown eating shumai” given by Stable Diffusion versus the same prompt (Niu Yue Tang Xiao Hai Shao Mai ) for ERNIE–ViLG.

Stable Diffusion


As someone who grew-up eating dim sum in Chinatowns and China, I would say that the results are tied. Both got the right Shumai, which in dim sum contexts is a type or type of shumai. Succulent shrimp and pork dumplings in a yellow half-open wrap. Stable Diffusion captures the atmosphere of a Chinatown dim-sum restaurant, but its shumai is off. But I can see where the machine is going. ERNIE-ViLG generates a lot of energy. A It’s a type of shumai, and it’s more common in eastern China than the Cantonese.

This quick test demonstrates the difficulty of capturing cultural nuances when data sets are inherently biased. Assuming Stable Diffusion has more data on the Chinese diaspora, ERNIE-ViLG is likely to be trained on a wider range of shumai images that are rarer in China,

Tencent’s is another Chinese tool that has made a lot of noise. Different DimensionsThe AI generator can transform photos into anime characters. The AI generator has its own bias. It was originally intended for Chinese users but it quickly gained popularity in anime-loving countries like South America. However, users quickly realized that the platform was not able to identify black and plus-size people, which is a significant missing feature in Japanese anime. This led to offensive AI-generated results.

It is also evident that the model was not adjusted for darker skin tones.

However, Different Dimension Me is the name. But they have already blocked or limited overseas users because they couldn’t handle the traffic.

— Rui Ma Ma Rui (@ruima) December 7, 2022

ERNIE-ViLG is not the only large-scale Chinese text to image model. TaiyiIDEA, a research laboratory, created. Harry Shum, a renowned computer scientist, is the leaderMicrosoft Research Asia, the largest research branch of Microsoft outside the U.S. was co-founded by, Open-source AI model has one billion users and is trained using 20 million filtered Chinese images-text pairs. Parameter.

IDEA is a local government-backed institution that has been working on cutting-edge technologies in recent years, unlike Baidu and other profit-driven tech companies. This means that the center has more freedom to research and less pressure to achieve commercial success. It’s a promising outfit that is based in Shenzhen, supported by one China’s wealthiest cities.

AI Rules

China’s generative AI tools don’t just depend on the domestic data they use, they are also shaped and influenced by local laws. As MIT Technology Review It should be notedBaidu’s text to image model filters out politically sensitive keywords. This is to be expected given that censorship has been a long-standing practice on the Chinese internet.

The new regulatory measures that are being implemented by the government to protect the future of this emerging field are even more important. They target “deep synthesis technology”, which is defined as “technology that uses deeplearning, virtual reality and other synthesis algorithms in order to generate text, images and audio. Users’ behavior is affected by the fact that prompts can be traced back to their real identities.

These rules could, however, lead to a more responsible use of generative artificial intelligence (which is already being misused elsewhere). You can produce NSFW or sexist content.. For example, the Chinese regulation explicitly prohibits people from creating and spreading fake news AI-generated. However, it is up to the service providers to decide how this will be implemented.

“It’s fascinating that China is at forefront of regulation [generative AI] As a country,” Yoav Shoham (founder of AI21 Labs) said. OpenAI rival from IsraelIn an interview. “There are many companies that are limiting AI… Every country that I know has tried to regulate AI or make sure that the legal or social system is keeping up with technology, particularly regarding the regulation of the automatic generation and distribution of content.”

There isn’t consensus on how this rapidly-changing field should yet be managed. Shoham said that he believes it’s an area where we all have to learn together. It must be a team effort. It must involve technologists who understand the technology and its limitations, as well as social scientists and the public sector.

Monetizing AI

Many Chinese are using machine learning algorithms to make a lot of money, as artists worry about being replaced by AI. They aren’t the most tech-savvy people. They are more likely to be opportunists, or stay-at-home moms looking for an additional source of income. They see that they can improve their prompts. They can trick AI to create stunning wallpapers or emojis.They can use this information to charge downloads or drive ad revenue on social media. The most skilled people are able to sell their prompts to other people who want to make money, or even train them.

Other Chinese use AI in their formal jobs, like The rest of the world. For example, light fiction writers can cheaply produce illustrations for their works. This genre is often shorter than novels and features illustrations. AI can be used to design T-shirts, nail presses, and other consumer goods. This is an interesting use case that could disrupt the manufacturing industry. Manufacturers can reduce design costs and shorten production cycles by producing large batches of prototypes quickly.

It is too early to see how generative AI will develop in China and the West. Entrepreneurs have made decisions based upon their early observations. A few founders said that professionals and businesses are happy to pay for AI as they see a direct return on their investment. Startups are eager to find industry use cases. One brilliant application was from Surrealists supported by Sequoia China (later Renamed to Movio() Hillhouse-backed ZMO.aiThe pandemic saw that e-commerce sellers had difficulty finding foreign models because China was closing its borders. The solution? The algorithms were used to create fashion models for all shapes, colors, races, and sizes.

Some entrepreneurs doubt that their AI-powered SaaS will experience the same skyrocketing valuations and meteoric growth as their Western counterparts. Jasper And Stability AIThey are enjoying. Many Chinese startups have shared the same concern over the years: China’s enterprise customers are less willing to pay more for SaaS than customers in developed countries. This is why Many of them expand overseas..

China’s SaaS market is also dog-eats-dog. You can do quite well in the U.S. by building product-driven software that doesn’t rely on human service to acquire and retain users. However, in China, even if your product is great, your competitor could steal it and hire dozens customer support staff to beat you,” said a founder from a Chinese generative AI startup. He requested anonymity.

Shi Yi, founder and CEO FlashCloud is a sales intelligence startupAccording to, Chinese companies tend to prioritize short-term profits over long-term innovation. He said that Chinese tech companies are more focused on generating quick cash and getting skilled in applications. Unidentified Shanghai-based investor said he was disappointed that the major breakthroughs in generative AI this past year were all happening outside China.

Roadblocks ahead

Even though Chinese tech companies want to invest in large neural networks training, they may not have the right tools. The September U.S. government China imposed export restrictions on high-end AI chips. Although many Chinese AI startups are focused on the applications front, they don’t require high-performance semiconductors to handle large amounts of data, basic research will require less powerful chips. This is according to an enterprise software investor at a top Chinese venture capital firm. He requested anonymity. He said that China is embracing these sanctions and investing in advanced technologies for the long-term.

Baidu, a company that claims to be a leader in China’s AI field, believes the U.S. chip sanctions on its AI business are “limited” in the short- and long-term, according to Dou Shen, the executive vice president of AI Cloud Group. Q3 earnings call. Baidu’s AI cloud business is “largely not dependent too much on the high-end chips.” In cases where it does, it has “already had enough in hand to support our business in near term.”

What about the future? “When we consider it from a mid-to-long-term perspective, we actually have our own AI chip, so-called Kunlun“Yes,” the executive assured. “By using our Kunlun chips [Inaudible] Our AI platform can recognize text and images in large language models. The efficiency of our AI platform can be improved by 40%, and the total cost has been reduced from 20% to 30%.

It will be interesting to see if Kunlun or other Chinese AI chips can give China an advantage in the generative AI race.

Continue reading