The third generation of artificial intelligence is about to develop, and the AI big model promotes industrial transformation

With the advent of ChatGPT, the AI industry is also changing.

"The emergence of GPT may prompt us to develop the third generation of artificial intelligence industry." On July 13, at the keynote speech of the 2023 JD Global Science and Technology Explorer Conference and JD Cloud Summit, Zhang Cymbal, academician of the Chinese Academy of Engineering, pointed out that in the era of basic models, AI is ushering in the dual opportunities of scientific and technological revolution and industrial transformation.

Zhang believes that compared with the first and second generation of artificial intelligence, the third generation of artificial intelligence needs to build explainable and robust (robust) AI theories and methods, and develop safe, controllable, credible, reliable and scalable AI technologies. In addition, Zhang stressed that the third generation of artificial intelligence also needs to use the four elements of knowledge, data, algorithms and computing power to promote the innovative application and industrialization of AI.

In terms of promoting the industrialization of AI, technology companies have continued to explore. On the same day, Jingdong officially launched the 100-billion-level Yanxi model. Cao Peng, chairman of the technical committee of Jingdong Group and president of the cloud business unit of JD.com, pointed out that the Yanxida model not only uses standard corpus in the training process, but also uses a large amount of industrial data accumulated by JD.com, and the model also has good generalization and security. Yanxi AI development and computing platform will be officially launched in August, and appointment registration has been launched.

"All technology is a tool, a means, but not a goal for us." Recently, in an interview with the media, including the 21st Century Business Herald, Cao Peng said, "We really want to use this technology to bring about industrial change, and this is what we want to see." ”

The third generation of artificial intelligence

Since it was first proposed in 1956, artificial intelligence technology has been developed for more than 60 years, but it is still difficult to say that it is used on a large scale today.

Zhang pointed out that limited by knowledge-driven and limited computing power, the first generation of artificial intelligence industry has not developed. In contrast, AI technology is widely used in the second-generation artificial intelligence industry, but there are still problems such as limited scale, data security, and algorithm security. From one side, you can get a glimpse: the artificial intelligence industry has not yet appeared such as IBM, Microsoft and other similar giants, which also means that the development of the artificial intelligence industry has a long way to go.

The reason for this can be glimpsed by comparing the development path of artificial intelligence and the information industry. Zhang Cymbal analyzed that the development of the information industry has gone through the process of establishing theory, industrialization of information technology (the emergence of general computer hardware and software), and industrial (industry) informatization landing industrial application. It is also in this process that the emergence of software and hardware is universal, and then a huge market is formed, thereby supporting the growth of leading companies such as Microsoft, Intel, and IBM.

The artificial intelligence industry, on the other hand, is not. "So far, there is no theory in the artificial intelligence industry, only algorithms and models, and there are certain flaws." Zhang cymbal said. At the same time, AI software and hardware developed by relying on algorithm models belong to special fields, and the application and market scale are limited, "so the artificial intelligence industry must be closely integrated with the application field and deeply cultivate the field to form a real artificial intelligence industry." ”

In this regard, Zhang pointed out the necessity of developing the third generation of artificial intelligence, including building explainable and robust (robust) AI theories and methods, developing safe, controllable, credible, reliable and scalable AI technologies, and promoting the innovative application and industrialization of AI.

Zhang believes that the popularity of AI models represented by ChatGPT is a step towards general artificial intelligence (AGI). Specifically, ChatGPT achieves the goal of behaviorist AI in dialogue, that is, it is close to the point of dialogue with real people, and reaches the field of opening in dialogue, that is, achieving commonality that is not related to the field.

In the era of basic models, AI is also ushering in new opportunities. Zhang believes that while the basic model brings a common platform, it will also provide a technical foundation for various applications.

In this regard, JD.com is also actively exploring. "The big model really realizes its own value, and it must be in industrial application." Xu Ran, CEO of JD.com, said that the value of the large model = the calculation power of the algorithm× the ×square of the data × the thickness of the industry. "The first three indicators are important, but the key is to apply technology in industrial scenarios and create actual value. When industrial efficiency and industrial boundary expansion are qualitatively improved, the big model has more important practical value and significance, which will be no less than another industrial revolution. ”

Industrial big model

Different from other general-purpose large models, the characteristics of the Yanxi large model are around the industry. Specifically, it integrates 70% of general data and 30% of the native data of the digital intelligence supply chain during training, which can focus on knowledge-intensive and task-oriented industrial scenarios and solve real industry problems.

"In addition to the dataset common to training standard large models, we include JD.com's own data. Differences in datasets allow us to provide differentiated model capabilities in the industries in which we excel. Cao Peng said.

He Xiaodong, president of Jingdong Exploration Research Institute and president of Jingdong Technology's intelligent service and product department, also pointed out in an interview with the media including the 21st Century Business Herald that from retail, logistics, health, finance, and then to industrial products, Jingdong's supply chain has a whole chain, one end is connected to the industrial Internet, the other end is connected to the consumer Internet, "from procurement and sales, trading, logistics and distribution, including a series of terminal services, a large amount of data will be generated every day." These data and scenarios ensure that our large models are directly oriented to scenarios and industry applications. ”

Jingdong's long link, complex collaboration, and more dynamic data return scenarios have also become the best "training ground" for large models.

"The development of JD.com itself comes from the experience of the scene, which also accumulates rich industrial data and industry know-how for us, so that we have the ability to make a large model that meets the needs of the industry and can solve the pain points of the industry, and continue to experience in the real scene to form an effective and good cycle." Xu Ran said.

In addition, the AI technology capabilities accumulated by JD.com's continuous investment over the years are also one of the important advantages of Yanxida's model.

According to reports, as early as 2021, JD.com launched a billion-level model K-PLUG, at that time, the product copy generated by K-PLUG could cover more than 3,000 categories of JD.com, generating a total of 3 billion words, and the manual review pass rate exceeded 95%. In 2022, JD.com will launch Vega, a ten-billion-level model, which can be widely used in a variety of downstream natural language processing tasks such as sentiment analysis, semantic matching, grammar correction, intelligent question answering, and common sense reasoning.

Now, on the basis of years of research, Jingdong's new generation of 100-billion-level large model Yanxi has officially appeared. "Based on such scenarios, data and years of technology accumulation, JD.com has become a frontier highland for building large models, and it is also the best position to produce large models in the industry." He Xiaodong pointed out. (Intern Shi Jie also contributed to this article)

For more information, please download 21 Finance APP