On November 17, the "2024 Great Bay Area Science Forum - Artificial Intelligence Sub Forum and the Fifth China (Guangdong) Artificial Intelligence Summit Forum" was held in Nansha, Guangzhou. Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, shared the latest progress of current artificial intelligence technology and industrial application from the perspective of industrial practice. She stated that with the continuous advancement of technology, artificial intelligence is gradually becoming an important force driving socio-economic development, and big models are reshaping the industrial intelligence engine.
In the keynote speech section of the forum, Wu Tian shared that generative AI based on large models has pushed artificial intelligence to a new climax. With the breakthrough of self supervised algorithms, large models can undergo basic training from massive amounts of data, possessing a basic understanding of the world, and can be adapted to scenarios to produce very good results in a wide range of applications.
At the end of 2018, Baidu was conducting research and development on the technology of big models. In March 2019, it released Wenxin Big Model 1.0, and in June of this year, it released Wenxin Big Model 4.0 Turbo. The basic model continues to iterate, and the knowledge enhancement, retrieval enhancement, and intelligent agent technologies behind it are also constantly evolving. Among them, the intelligent agent technology based on the large model is a composite AI system constructed between the basic model and industrial application scenarios. Wu Tian believes that through composite AI systems such as intelligent agents, large models can be deployed in industrial applications to complete highly complex tasks in practical industrial scenarios.
As early as October last year, Baidu released the intelligent agent mechanism and developed System 2, which is similar to the human thinking system. Its core is the thinking model, including understanding, planning, reflection, and evolution, which can achieve reliable execution, self evolution, and to some extent, white box the thinking process, allowing machines to think and act like humans, autonomously complete complex tasks, and continuously learn and evolve in the environment. In order to complete complex tasks, multi-agent systems also collaborate through different organizational models such as centralization, decentralization, and assembly lines, effectively enhancing the overall efficiency of the intelligent system.
On site, Wu Tian demonstrated the application of code intelligent agents. Taking the operation activity system implemented based on the Wenxin big model as an example, the model first understands user needs, designs the system, and then plans relevant tasks to execute. Next, it generates code according to the plan and gradually implements system construction and service deployment. If the deployment fails, the model can automatically reflect and iterate based on error messages until the deployment is successful and the service runs smoothly.
With the increasing universality of artificial intelligence, its application in various industries is also deepening, empowering the transformation and upgrading of agriculture, manufacturing, energy, transportation, finance, education, healthcare, media and other industries.
For example, in the industrial field, Baidu and Zhongtian Steel have collaborated to build databases and knowledge bases for the steel industry. Based on this, the Zhongtian Nantong large model has been deployed and implemented in multiple application scenarios such as production, finance, and personnel, bringing about a series of efficiency improvements; In the field of government affairs, Baidu and the Haidian District Government of Beijing have collaborated to create a government affairs model based on the Wenxin model, reducing the previously 3-day work in scenarios such as data search and indicator calculation to 1 minute, and reducing the previously 5-day work in scenarios such as icon drawing and report writing to 30 minutes.
In addition to systematic applications in the aforementioned industries, large-scale models have also been implemented in more segmented scenarios such as retail. For example, in e-commerce live streaming scenarios, complex content creation agents based on large models can generate multi style live streaming scripts, or witty and elegant literary styles, or passionate and infectious grass planting styles, to meet the needs of different live streaming scenarios. It is reported that the operating cost of live streaming on Baidu platform has decreased by 80%. In the context of intelligent shopping guidance, traditional official websites can only display statically, while intelligent websites built on intelligent agents can provide active recommendations, timely responses, and one-on-one services, greatly improving the efficiency of interactive marketing. For example, after BYD's official intelligent agent was launched, the sales conversion rate increased by 119%.
Public data shows that the average daily volume of ERNIE Bot's large model has exceeded 1.5 billion, and the number of Wenxin Yiyan users has reached 430 million. Currently, various industries are implementing large-scale models based on their own scenarios, unique experiences, standards, data, etc., forming a huge application ecosystem.
Wu Tian finally stated that technological change drives application innovation, application explosion creates commercial value, and the formation of commercial ecology in turn further promotes technological iteration, which has become a very important driving cycle for the development of artificial intelligence models. In the future, more industries and application scenarios will rely on big model technology to generate more innovation.