huge technology The leap that machine learning models have shown in the past few months has made everyone excited about the future of AI, but also worried about its unpleasant consequences. After all the buzz around Stability AI and OpenAI’s text-to-image tools, ChatGPT’s ability to have intelligent conversations has become a new obsession in every field.
Entrepreneurs, researchers, and investors in China, where the tech community has always followed progress in the West, are looking for ways to tap into the field of generative AI. Technology companies are devising tools built on the open source model to attract consumer and enterprise customers. Individuals are benefiting from AI-generated content. Regulators have moved quickly to define how text, images, and video compositing can be used. Meanwhile, US tech sanctions have raised concerns about China’s ability to keep up with advances in AI.
As generative AI takes the world by storm towards the end of 2022, let’s take a look at how this explosive technology is unfolding in China.
Chinese flavor
Thanks to viral art creation platforms like Stable Diffusion and DALL-E 2, generative AI is suddenly on everyone’s lips. On the other side of the world, Chinese tech giants are also wooing the masses with similar products, with twists to suit national tastes and political climates.
Baidu, which has made a name for itself in search engines and has stepped up its self-driving game in recent years, has developed ERNIE, a 10 billion parameter model trained on a dataset of 145 million Chinese image-text pairs. – We operate ViLG. How fair are you to your American counterparts? Below is the result of the prompt “Children eating dumplings in New York’s Chinatown” given to Stable Diffusion for the same prompt in ERNIE-ViLG’s Chinese (纽约唐人街小孩吃烧卖).
stable diffusion
ERNIE-ViLG
As someone who grew up eating dim sum in China and Chinatown, I think the result is a draw. Neither could get proper shumai. In the context of dim sum, it’s a type of juicy shrimp and pork dumpling that comes in a half-open yellow packaging. Stable Diffusion sets the tone for a Chinatown dim sum restaurant, but the shumai is off (although you can see where the machine is going). While ERNIE-ViLG produces a A type of dumpling, more common in eastern China than in the Cantonese version.
A quick test reflects the difficulty of capturing cultural nuances when the datasets used are inherently biased, with Stable Diffusion having more data on the Chinese diaspora and ERNIE- We assume that ViLG is trained on a larger variety of shumai images, which are probably more rare outside of China.
Another popular Chinese tool is Tencent’s Different Dimension Me, which can turn a photo of a person into an animated character. AI generators exhibit their own biases. Although it was aimed at Chinese users, it unexpectedly became popular in other anime-loving regions such as South America. I’ve noticed that I can’t identify groups that lack , leading to unpleasant AI-generated results.
Apart from ERNIE-ViLG, another large-scale Chinese text-to-image model is Taiyi. This is the brainchild of IDEA. IDEA is a laboratory headed by renowned computer scientist Harry Shum, who co-founded Microsoft Research, Microsoft’s largest research arm outside the United States. Asia. The open-source AI model was trained on his 20 million filtered Chinese image-text pairs and has 1 billion parameters.
Unlike Baidu and other profit-oriented tech companies, IDEA is one of the few institutions working on cutting-edge technology with the support of local governments in recent years. That means the center likely enjoys more research freedom without the pressure to drive commercial success. , is a remarkable and up-and-coming organization.
AI rules
China’s generative AI tools are not just about domestic data they learn from. They are also shaped by local laws. As MIT Technology Review pointed out, Baidu’s text-to-image model filters out politically sensitive keywords. It’s no surprise, given that censorship has long been ubiquitous on the Chinese internet.
More important to the future of the fledgling sector is a series of new regulatory measures aimed at what the government calls “deep synthesis technology.” This means that “just like other types of internet services in China, from texts, images, games to social media, using deep learning, virtual reality, and other synthetic algorithms, users will be able to The fact that the prompt can ascertain the user’s true identity inevitably has a limited impact on user behavior.
But on the bright side, these rules could lead to more responsible use of generative AI. Generative AI is already being exploited elsewhere to produce NSFW and sexist content. For example, Chinese regulations explicitly prohibit the generation and dissemination of AI-generated fake news. However, how it is implemented is up to the service provider.
“It is interesting that China is at the forefront of regulation.” [generative AI] as a country,” Yoav Shoham, co-founder of Israel-based OpenAI rival AI21 Labs, said in an interview. “There are various companies that are imposing limits on AI… every country I know either regulates AI or in some way has a legal or social system on the regulation of technology, especially auto-generated. We make an effort to verify the content.”
However, there is still no consensus on how a rapidly changing field should be managed. “I think it’s an area that we’re all learning together,” Shoham admitted. “It has to be a collaborative effort. Technologists who really understand technology and its functions and functions, public sector, social scientists, people affected by technology, and commercial and legal issues. We need to involve governments, including regulatory aspects.”
Monetizing AI
As artists fear being replaced by powerful AI, many artists in China leverage machine learning algorithms to generate revenue in a variety of ways. They don’t come from the tech-savvy crowd. Rather, they are opportunists or stay-at-home moms looking for additional sources of income. They realized that by improving their prompts, they could trick the AI into creating creative emojis and stunning wallpapers that they could then post on social media to generate more advertising revenue or charge for direct downloads. I’m here. Those who are truly skilled may even sell prompts or train for a fee to others who want to join the game to make money.
Other countries in China, like the rest of the world, are using AI in their formal jobs. For example, light fiction writers can cheaply mass-produce illustrations for their work. This genre is shorter than fiction and often features illustrations. An interesting use case that could disrupt the manufacturing realm is using AI to design prints for his t-shirts, press nails, and other consumer goods. By rapidly generating large batches of prototypes, manufacturers save design costs and shorten production cycles.
It’s too early to know how generative AI is developing differently in China and the West. However, the entrepreneur made a decision based on early observations. Several founders have told me that companies and professionals are generally willing to pay for AI because they see a direct return on their investment. There was his one of his successful applications from Sequoia China-backed Surreal (later renamed to Movio) and his Hillhouse-backed ZMO.ai. ZMO.ai has found that e-commerce sellers are struggling to find foreign models during the pandemic as China continues to close its borders. solution? The two companies worked on algorithms to generate fashion models of all shapes, colors and races.
But some entrepreneurs don’t believe AI-powered SaaS will see the kind of explosive recognition and rapid growth enjoyed by its Western counterparts such as Jasper and Stability AI. Over the years, many Chinese startups have told me they have the same concerns. Chinese corporate customers are generally less willing to pay for his SaaS than companies in developed countries, so many companies are starting to go abroad.
Competition in the Chinese SaaS space is also gluttonous. “In the US, you can do pretty well by building product-driven software that doesn’t rely on human services to acquire or retain users. But in China, even if you had a great product, Even so, your rival could steal your source code overnight and hire dozens of not-so-expensive customer support staff to outrun you.
Shi Yi, founder and CEO of sales intelligence startup FlashCloud, agreed that Chinese companies often prioritize short-term returns over long-term innovation. “When it comes to talent development, Chinese tech companies tend to focus on building application skills and making quick money,” he said. A Shanghai-based investor, who declined to be named, said he was “a little disappointed that all the big breakthroughs in generative AI this year are happening outside of China.”
obstacle ahead
Even if a Chinese tech company wants to invest in training neural networks at scale, it may lack the right tools. In September, the U.S. government slapped China in restricting exports of high-end AI chips. But for companies doing basic research, using less powerful chips means computing takes longer and costs more, enterprise software investors said. A top VC firm in China who wishes to remain anonymous. The good news, he argued, is that such sanctions are pushing China to invest in advanced technology in the long term.
As a self-proclaimed leader in China’s AI space, Baidu believes the impact of US chip sanctions on its AI business will be “limited” in both the short and long term, Baidu said. said executive vice president and head of the AI Cloud Group. , Dou Shen at the third quarter earnings call. That’s because “most” of Baidu’s AI cloud business is “less dependent on advanced chips.” And if you want a high-end chip, “actually, we already have enough in stock to support our business in the near future.”
what about the future? “In the medium to long term, we are actually developing our own AI chip named Kunlun,” said a confident executive. “By using our Kunlun chips, [Inaudible] For large-scale language models, the AI platform performed text and image recognition tasks with a 40% increase in efficiency and a 20% to 30% reduction in total cost. ”
Only time will tell if Kunlun and other homegrown AI chips give China an edge in the generative AI race.
Story updated to clarify that Yoav Shoham is the co-founder of AI21 Labs.
