Can Chinese Companies Create the Next ChatGPT?

Technology Author: Qinqie He, Yongqian Yang Feb 22, 2023 10:30 AM (GMT+8)

This article discusses the possible birth of the Chinese version of "ChatGPT" and relevant challenges. Additionally, this article also tries to make meaningful predictions on AIGC tools.

AI

If you ask what the most splashing, eye-catching technology product at the beginning of 2023 is, the answer is undoubtedly ChatGPT. First, with Microsoft's USD 10 billion investment in OpenAI, the "brain" behind ChatGPT, Google urgently held a 'code-red' internal reflection conference to review the slow response to AIGC's general-purpose products and accelerated testing of ChatGPT-competing product.

 ChatGPT's overnight explosion, high usage threshold, and, most importantly, the possibility of its generalized products replacing traditional search engines have not only aroused the enthusiasm of major overseas technology companies but launched a heated race domestically. Baidu takes the lead.

Screenshot 2023-02-19 at 9.53.11 PM.png.png

 Baidu Takes the Lead

 On January 30, 2023, various news agencies reported that Baidu was developing an AI dialogue service similar to ChatGPT. As of the writing of this article, Baidu confirmed that the name of its ChatGPT-like project is ERNIE Bot (Chinese: 文心一言). It will complete internal testing by March 2023 and be open to the public afterward. Baidu's stock price ushered in a new high for more than half a year, showing the capital market's keenness towards AIGC and the anticipation from the outside world.

 Baidu has spent the past few years seeking to transform itself into an artificial intelligence company, investing billions of dollars in technologies, including self-driving cars and chips for artificial intelligence applications, making it the most promising domestic technology company to develop AI-related products first. In 2019, Baidu developed a system called Ernie (Chinese: 文心), a deep learning model similar to the technology ChatGPT is based on, which has been used to make its search results more relevant. Later, Baidu developed many Ernie models and expanded their capabilities to include image and art generation, similar to the functions of OpenAI's Dall-E.

 Since 2021, attempts have been made to combine AIGC and search. Baidu has long been using Ernie as the basis for its chatbot, training it using Chinese and English language resources worldwide. According to Baidu's open-source research paper, the resources used to train Ernie included Wikipedia, BookCorpus, Reddit, and Baidu's product ecosystem, which covers Baidu Encyclopedia, Baidu News, etc.

 Before Baidu's Create AI developer conference in January 2023, Baidu Search announced that it would upgrade its "generative search" capabilities based on Baidu's self-developed generative model and pointed out that generative AI and search engines are complementary rather than substitutes. It is further proposed that the underlying search technology and AI underlying technology are interlinked. 

Screenshot 2023-02-19 at 9.53.33 PM.png.png

 Baidu is by no means the only Chinese tech giant that has been eyeing the AI opportunity for years. All the big names, Alibaba, Tencent and ByteDance, have developed and applied AI in various business areas. As the search engine track they all once intended to build and compete against Baidu is now quietly changing its form and reopening, it is no surprise that they want to grasp the land.

 ByteDance is recently reported to be quietly preparing for the revival of Wukong Q&A (Chinese: 悟空问答), which has been shut down for two years, relaunching it as a Q&A platform with a search-based functional design. Thinking about the relationship between ChatGPT products and search tools, it is not difficult to see Bytedance's hidden ambition.

 Tencent, on the other hand, has just gained patent authorization for its "human-computer dialogue method, device, equipment and computer-readable storage medium," according to Tianyancha (Chinese: 天眼查).

 Challenges Ahead

 The current explosive success of ChatGPT is inseparable from its models, big data, and the immense computing power behind it. It uses the Transformer neural network architecture, the training data comes from a massive corpus including Wikipedia and real conversations, and there are as many as 175 billion model parameters. Through training model iteration and tremendous data training, ChatGPT's dialogue effect is more realistic, and its functions are more comprehensive. The value and prospects of ChatGPT cannot be underestimated, but whether Baidu and others can catch up and monetize remains a question mark upon multiple hurdles. 

 Data Collection Restriction 

ChatGPT automatically collects global data information to form training data. Still, there are corresponding policy restrictions on the collection and acquisition of overseas-specific information in China, and the quality of Chinese data on the entire Internet is still significantly different from that of English data. Moreover, in terms of natural language understanding, the polysemy of Chinese has a naturally higher threshold than English. On the other hand, whether the basic model can accommodate the amount of GPT-level data still needs to be verified.

 High Maintenance Cost

The high operation and maintenance cost of ChatGPT is still a problem that must be addressed. According to AI scholar Tom Goldstein, ChatGPT's daily cost of generating content based on massive GPU computing power is about USD 100,000 dollars. That was only based on the one million active users amount two months ago. 

According to Li Di (Chinese: 李笛), CEO of Xiaoice (Chinese: 小冰), the AI system developed by Microsoft, "The amount of dialogue interaction currently supported by the Xiaoice framework in one day reaches the amount of dialogue interaction of 14 human beings in a lifetime. If the ChatGPT-like method is used, the daily cost will be as high as CNY 300 million, and the annual cost will exceed CNY 100 billion."

 Commercialization

ChatGPT can achieve fluent dialogue and even discuss some professional topics. Still, more often than thought, it is a "made-up language master" with problems such as outdated data, prejudice, false information, and incorrect values. If it is used as a search tool, the results obtained need to be verified twice in traditional search engines or other places, which will reduce efficiency. 

 Aside from technological imperfection, ChatGPT has yet to show a sustainable business model or create real value. For Microsoft, ChatGPT is a tool that can be embedded in different applications, but judging from its current maturity, it has yet to bring commercial value to Microsoft. Currently, the cost of ChatGPT is not directly proportional to its rewards, and it is difficult for even giant companies to support such a long-term high-loss business. 

 Even though the industry's eagerness for ChatGPT continues, and Chinese tech companies should remain vigilant and prepared, considering the cost, lack of industry knowledge accumulation and data training expertise, there is still a long way to go before a Chinese version of ChatGPT can be truly commercialized.

 Future Trends of AIGC Technology

This article makes four predictions for AIGC technologies, aiming to shed some light on what's to come in the AI space. Firstly, the rapid rise of ChatGPT will come with its fair share of controversy, and it could be used negatively. Public trust and information security will be suffered under such situation. The second prediction is highly relevant to the first one: increasing laws and regulations of AI applications will be enacted by the government to provide a more user-friendly and positive environment. Thirdly, more investments will happen in the AIGC field following Microsoft's multi-million acquisition. The increased competition can lead to further market growth and more mature technology, benefiting society with new ways of adopting AIGC tools. 

Finally, more businesses will adopt AIGC technology as it evolves and becomes more mature. The potential of AIGC tools to be combined with traditional business models is considerable. Ultimately, the combination will lead to increased customer satisfaction and exciting innovations.