"There are many large models in China, but there are very few AI native applications developed based on large models." On November 15, at the 2023 Xili Lake Forum in Shenzhen, Baidu founder, chairman and CEO Robin Li said when talking about the current situation of the domestic AI industry. On that day, he shared two "cold thoughts" and three "hot drivers" about the AI native era.
Robin Li pointed out in two "cold thoughts" that as of October, 238 large models have been released in China. There are too many large models and too few AI native applications developed on top of the models. He also said that many industries, companies, and even cities want to train their own special large models from scratch, but special large models do not have the ability to emerge intelligently and have very limited value.
When talking about "hot drive", Robin Li said that powerful basic large models can drive the explosion of AI native applications. Currently, the best AI native applications have not yet appeared. "In the AI native era, there will definitely be excellent AI native applications developed based on these large models." In contrast, AI native applications will also drive the development of AI technology stacks such as models and chips. "Only through more scenarios and applications can a larger data flywheel be formed and the chips be able to be used enough and easy to use."
He specifically pointed out that embracing the AI era requires top leadership, “because only CEOs will care whether new technologies have a positive impact on key indicators of their business.”
1. “We need 1 million-level AI native applications, but we don’t need 100 large models.”
"We need 1 million-level AI native applications, but we don't need 100 large models." Robin Li pointed out that in the global market, AI native applications are becoming a major trend. "Microsoft does not have its own basic large model. It cooperates with OpenAI, but it has the most successful AI native application. As everyone knows, it is Copilot for Office 365, which costs 30 US dollars a month."
As for the country, he analyzed that as of October, 238 large models had been released in China, compared with 79 in June, an increase of three times in four months. However, compared with the dozens of basic large models and thousands of AI native applications abroad, the number of domestic AI native applications is very small.
Robin Li said, "Continuous and repeated development of basic large models is a huge waste of social resources." Looking back at the PC era and the mobile Internet era, Robin Li said that various PC software were developed based on Windows systems, and the only two operating systems in the mobile era were Android and iOS. Large models are similar to operating systems, so in the end developers will only rely on a few large models to develop AI native applications.
"Without the ability to emerge intelligently, the value of dedicated large models is very limited." Robin Li analyzed that many industries, companies and even cities are buying cards, hoarding chips, and building intelligent computing centers, hoping to train their own dedicated large models from scratch. However, they do not know that the large models produced in this way do not have the ability to emerge intelligently.
"Intelligent emergence" refers to the ability of large models to learn by analogy, that is, things that have not been taught can be learned by large models. "Only when the parameter scale of your model is large enough, the amount of training data is large enough, and you can continuously invest and iterate, can intelligent emergence be produced."
Robin Li believes that "the industrialization model of large models should combine the general capabilities of the basic model with the proprietary knowledge of the industry." In other words, the large model is a small model, and the special small model has fast response and low cost. The large model is more intelligent and can be used to find out. He bluntly said, "There are more than 200 large models in China, and they are on this list and that ranking, but the number of calls is actually very small. The number of calls of Wenxin Model Company is more than the number of calls of these 200-plus large models combined."
2. Only CEOs care about the positive impact of new technologies on business indicators.
“I have seen many companies that attach great importance to this opportunity, but do not have a deep understanding of the nature of the problem.” Robin Li mentioned the current industry phenomenon. The CEO handed over the task to the person in charge of IT, thinking that “making a basic model yourself, or selecting a high-scoring large model based on the evaluation methods spread online” means embracing the AI era. This is actually a huge waste of company and social resources.
"Why do we need a top leader to drive it? Because only CEOs care about whether new technologies have a positive impact on the key indicators of their business." Robin Li said that the essence of the problem is whether the large model has a positive impact on Internet companies' DAU, duration, user retention and other indicators, and whether it has an impact on the company's revenue, profits, and costs.
He took Baidu Wenku as an example. Baidu Wenku, which has been reconstructed through AI nativeization, can generate a 20-page PPT in one minute, including chart generation, format beautification, etc., and it is almost zero cost, realizing the transformation from a content tool to a productivity tool. Because of this transformation, “Wenku’s payment rate has been significantly improved. This is what I call AI’s role in promoting key business indicators.”
"At Baidu, we have resolutely restructured each of our product lines to be AI-native," Robin Li said. He believes that the concept of AI native will definitely be accepted by C-end users and startups first, followed by small and medium-sized enterprises, and finally large enterprises. He said that large companies have a clear division of labor and need CEOs to take the initiative to lead changes.
3. AI native applications and basic large models mutually benefit and achieve mutual success
In Robin Li's view, powerful basic large models will drive the explosion of AI native applications. "China has a leading basic large model. This is a solid foundation for the development of AI native applications and is the underlying capability."
He introduced that since the release of Wenxin Yiyan based on Wenxin Big Model 3.0 on March 16, Wenxin Big Model has continued to iterate and was upgraded to version 4.0 last month, which is called the professional version on Wenxin Yiyan’s website and APP. Wenxin 4.0 has significantly improved in the four major abilities of understanding, generation, logic and memory. Since Wenxin Yiyan was opened to the whole society on August 31, the number of API calls of Wenxin Big Model has shown exponential growth.
"AI native applications are applications developed based on the understanding, generation, logic and memory capabilities generated by the emergence of large model intelligence. These capabilities were not available in the past era, so they can open up unlimited space for innovation."
In contrast, Robin Li said that technology stacks in the AI era, such as models and chips, also need to be driven by AI native applications. "Good applications will drive the market and force market changes." He used the new energy automobile industry as an analogy to say that measures such as tax exemption and exemption for new energy vehicles and unlimited number and travel restrictions on new energy vehicles have effectively stimulated the rapid growth of the new energy automobile industry. The AI industry is also driven by demand, and efforts should be made on the demand side and application side to encourage companies to use large models to develop native AI applications and use the market to promote industry development.
"Only through more scenarios and applications can a larger data flywheel be formed and the chip be made sufficient and easy to use." Robin Li said.