Promoting AI technology from "germination" to "bearing fruit" is becoming the breakthrough goal for industry organizations.

Recently, several tech manufacturers are competing to start with AI coffee ordering, providing services that are closer to life and more practical for people. At the same time, the service scenarios of AI are gradually expanding. Recently, the reporter learned from the World Internet Conference that nowadays, people can interpret physical examination reports through AI health butlers, and use AI cultural tourism assistants for scenic area navigation and restaurant guides nearby.

However, the reporter also learned in the interview that the current AI vertical services still have some shortcomings in the interaction mode. To solve this problem, Ant Group is combining knowledge learning and retrieval-augmented generation (RAG) to make AI large models not only reserve vertical scenario knowledge but also get closer to the interaction logic in real scenarios.

Eric Jing, Chairman and CEO of Ant Group, said at the meeting that the transformation brought by AI will be reflected in three aspects: AI will fundamentally change all industries, reshape industrial forms and economic landscapes; AI will help humans have "super abilities" to solve complex problems that could not be solved in the past; AI will promote the rapid development of robotics technology and become an assistant to humans. AI is about to usher in an era of large-scale personalization in the service industry.

Advertisement

Specifically, the reporter learned from the conference site that based on Ant Group's independent research and development of "cognitive decision-making intelligent body technology innovation and application", Alipay is focusing on building three AI applications: AI health butler, AI life butler "Zhi Xiao Bao", and AI financial butler "Ma Xiao Cai". Taking medical and health services as an example, Alipay has already served more than 600 million individual medical insurance payments nationwide and provided digital services for more than 3,600 hospitals. By jointly launching "An Zhen Er" with the Zhejiang Health Commission, it provides personalized digital companionship services, which have served more than 1,000 hospitals and more than 10 million patients, especially convenient for elderly patients and out-of-area medical groups.

The common perception in the industry is that AI has weaknesses in domain cognition ability and poor complex reasoning ability in vertical scenarios. So how should cognitive decision-making intelligent body technology innovation make up for the above shortcomings?

Zhang Zhiqiang, the person in charge of graph learning and knowledge graph at Ant Group, told the reporter that large language models generally still have the problem of knowledge illusion, while financial, medical, or life application scenarios are more rigorous industries. These scenarios have a higher threshold of understanding and rigorous logic. Therefore, the application of large models in these scenarios not only needs to learn knowledge but also needs to conform to the industry and interact according to the corresponding logic.

Zhang Zhiqiang gave an example: "When we use large models, it is often that we ask questions to it, and the model gives us an answer. But when we go to see a doctor, the doctor will ask us questions, such as what specific symptoms, the start time of the symptoms, etc. We hope that the application of large models in the medical scenario can conform to this interactive characteristic. To this end, we have open-sourced a knowledge-enhanced generation framework based on the knowledge graph, supplemented by retrieval-augmented generation (RAG), to supplement some knowledge that the language model does not know or is unclear about. At the same time, we use the internal logic of the medical or financial, government-industry to constrain the model, making it interact logically, and ultimately forming the 'brain' of the model."

Zhou Jun, the person in charge of Ant Group's Bailin large language model, believes that to use large models in vertical scenarios, it is also required that large models can call the tools closest to user needs from the thousands of tools that Ant Group originally had, that is, "hands and feet". This is also a major difficulty for large models. Different from human cognition, various tools or knowledge exist in the form of tokens in the model's cognition. When there are many callable tools, it is more difficult for the model to decide when to call which specific tools.

"Therefore, we are enhancing the function call ability of language models through various methods, and also endowing the model with usable tools at the appropriate time, which is already much more powerful than the original model's ability." Zhou Jun said.

The reporter said: "Taking the scenario of using large models to order coffee, behind the seemingly simple steps, it requires calling multiple collaborative systems upstream and downstream, involving not only the product information and payment information of the coffee shop but also the information of the delivery platform. Taking life services, medical, and financial scenarios as the entry point, we are also continuously optimizing and debugging to continuously enhance the ability of large models, making the interaction more in line with the industry's interaction logic while expanding knowledge, thereby enhancing the user experience."