Xinhua Finance and Economics, Beijing, September 13 (Reporter Li Tangning) At the "Digital Trade Development Trend and Frontier Forum" of the 2024 China International Trade in Services Fair, which opened on September 12, Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, said that the AI big model has strong enough versatility, shows the ability to think and plan, and will reconstruct the industrial digital engine. Intelligent agents built on the foundation of large models can handle more complex tasks in application scenarios, which is driving the explosion of industrial AI applications and generating significant commercial value.
In her keynote speech at the forum, Wu Tian stated that big models have become the mainstream development direction of artificial intelligence. ERNIE Bot, a big language model of knowledge enhancement launched by Baidu based on the accumulation of artificial intelligence technology over the years, is the core technology of knowledge enhancement, which ensures that the model can flexibly use the stored complete knowledge system, produce more accurate and valuable content, and provide substantive help to users. Wenxin intelligent agent further conducts thinking enhancement training on the basic model, constructs a thinking model, and has stronger understanding, planning, reflection, and evolution abilities. It can autonomously call tools to solve problems according to various demand scenarios.
On site, Wu Tian demonstrated multiple intelligent agent applications, including creative agents, code agents, and image generation agents. For example, in terms of creation, let ERNIE Bot write an outline of the application analysis report of the big model in the field of international trade. The creative agent needs to first think about the actual intention of the disassembly user, and then call creative thinking tools, online search tools, outline tools, etc. according to the implementation needs after disassembly, and finally generate a hierarchical, logical, professional and comprehensive report outline. In terms of code, code agents can assist users in writing the entire system level code, including code architecture design, code generation, and code execution.
Baidu said that ERNIE Bot currently has more than 300 million users, the average number of daily calls of Wenxin's big model exceeds 600 million, and the average daily processing of Tokens text is about 1 trillion. The Wenxin Big Model has generated a total of 7 billion lines of code, created 590 million articles, compiled millions of professional research reports, answered 170 million learning questions, and assisted 130 million people in their work.
On the forum, Wu Tian also stated that with the increasingly powerful basic models, developing applications is becoming simpler, and "intelligent agents are the most promising direction for the development of AI applications
In industrial practice, taking the enterprise service scenario as an example, the Baidu merchant intelligent agent developed based on the Wenxin big model can provide professional answers and in-depth communication to potential user questions through technologies such as dual end understanding, knowledge accumulation, and demand stimulation. It can be online at any time and respond 100%. Data shows that currently, the number of professional answers provided by merchant intelligent agents has exceeded 2.5 million per day. For individual users, deep communication has become a reality, with the highest number of conversation rounds reaching 288, "said Wu Tian.
In the e-commerce live streaming industry, in response to challenges such as high costs, limited time, and unstable quality, intelligent agent technology based on large models can flexibly generate digital human live streaming intelligent agents. It is reported that Baidu's e-commerce digital person live streaming "Hui Bo Xing" can generate a sales intelligent agent within 5 minutes. It can not only be online 24/7, helping merchants reduce 80% of operating costs, but also achieve high quality, stability and controllability, with a 75% increase in conversion rate for high-quality answers. Behind the commercial value lies the improvement of technical capabilities such as live script generation, live Q&A, and intelligent control brought by digital human live streaming intelligent agents based on the Wenxin big model.