NIE 2023 | Roundtable Forum: The Post-Generative AI Era and the Transformation of Industries with Large Models

Executive Director Cui Nan of Frost & Sullivan Greater China

Cui Nan:Since the launch of ChatGPT last year, the AI industry has shown a hot trend both in terms of capital and products. To date, nearly 200 large models have been released by domestic research institutions and enterprises. Generative AI is not only applied to C-to-C chatbots but also widely involved in industries such as enterprise services, marketing, healthcare, finance, and education. So, at this point, what is the current application status of generative AI? How is it integrated with traditional business models and products? Are there any successful innovative cases for reference?

Hu Jianlin, Vice President of Zhongguancun Kejin

Hu JianlinHe stated that since the launch of ChatGPT, generative AI has attracted widespread attention and discussion in the entire technology industry. Especially at the capital and product levels, the enthusiasm of the entire industry has risen significantly. In addition, the application of generative AI is no longer limited to basic chatbot models but has widely penetrated into multiple fields such as enterprise services, marketing, finance, healthcare, and education. As for large models, although they have attracted much attention in the early stages, as time goes by, enterprises and research institutions have begun to evaluate their actual effects and economic value more rationally.

Recently, Zhongguancun Kejin's cooperation with Noah Wealth on large models has been successfully implemented. By integrating self-developed financial domain large models, intelligent customer service, and other artificial intelligence technologies, Zhongguancun Kejin has created an intelligent knowledge base for Noah Wealth, which has the capabilities of multi-modal document analysis, automatic QA question extraction, and automatic knowledge content tagging. By integrating Noah Wealth's WeCom and its wealth management platform iNoah APP application, it provides intelligent question and answer query functions based on enterprise knowledge documents for its employees and users. By empowering intelligent customer service products with large model technology, it has significantly improved the accuracy rate of question intention recognition and response in the customer service system, with the expectation of reducing more than 70% of system operation work in the later stage and effectively helping Noah Wealth reduce costs and increase efficiency.

In addition, finance, trust, and other industries are also exploring the possibility of integrating large models with their businesses. For example, in financial shared centers, large models are used to upgrade and optimize knowledge bases, providing more accurate and efficient services. Generally speaking, current large models show great potential and value in knowledge assistance and service upgrading, and their application prospects in the future are worth looking forward to.

Zhang YahuiShe stated that currently, the application of large models in the AI field has attracted widespread attention. Although they are developing rapidly, most applications are still in the basic stage. On the one hand, large models have certain limitations, such as insufficient learning of knowledge in some vertical industries, making it difficult for them to be directly applied to production environments; on the other hand, although the current market is already quite large, as large models are combined with various businesses, market potential will be further amplified.

At present, large models have been initially applied in fields such as social media, art creation, e-commerce, healthcare, and education, but most are still shallow-level understanding AI rather than deep AI. For enterprises, optimizing customer service by combining large models, such as intelligent voice interaction and intelligent customer service, can improve service quality and lay the foundation for future deep applications. In summary, the prospects of large models are broad, but in-depth exploration and application still require time.

Dr. Deng XinyuShe stated that in the field of healthcare and big data, many companies have recently launched their own large language models, such as MedGPT from "Yilian" and Disc-MedLLM from Fudan team, as well as the layouts of Baidu, Huawei, and Tencent. Different from them, Zhiben Medical focuses on the field of tumor precision diagnosis and treatment. Using the valuable professional data accumulated over the years, it has launched the first large language model application in the tumor vertical field at home and abroad, "Zhihui Companion," based on the Zhiben Internet Hospital. The function of this application is not to replace doctors but to provide a professional assistant for patients, answering various professional questions about tumor treatment guidelines, drug information, drug medical insurance information, eligible clinical trials, real cases, etc., at any time. At the same time, it also provides a continuously updated professional information tool for doctors to assist them in making judgments.

In addition, the development speed of large language models is amazing. For example, the multi-modal model about to be launched by ChatGPT indicates that more innovative applications will emerge in this field.

Li YangHe stated that from the perspective of the capital market, since large models such as Stable Division and OpenAI made their debut last year, the enthusiasm for AI has continued to rise, but the enthusiasm has declined since the second quarter of this year.

There are two reasons: First, many investment institutions feel that there are not many differences among large model startups and it is difficult to find characteristic application scenarios; second, actual application scenarios have not brought significant improvements. In terms of technical diagnosis, there are two distinctions: one is optimization based on original AI applications, and this optimization may be an order of magnitude improvement; the other is completely based on new AI technology, subverting the original technology for industry applications. For example, in the BI industry, realizing data analysis and report generation automatically through AI is the first type of application; while knowledge base queries, taking advantage of AI's advantages in processing unstructured data, and realizing cross-database information extraction through AI belong to the second type. For the capital market, the second type belongs to an incremental market with greater investment opportunities and potential.

Executive Director Cui Nan:With the continuous rise in the popularity of generative AI technology, we must recognize that it is still an emerging technology and comes with a series of challenges and concerns. What do you think are the main challenges that generative AI technology faces in the industry at present? Among these challenges, which ones can be effectively solved in the short to medium term, and which ones may still lack clear solutions?

Hu JianlinHe stated that generative AI technology, as an emerging technology, although it has attracted much attention, still faces a series of challenges. First of all, although it performs well in language generation, there are still problems in fact judgment and process generation, especially when outputting information that seems reasonable but is actually inaccurate. In addition, although some challenges have been solved, such as enhancing fact judgment by using an external knowledge base, some problems, such as ethical and mental issues, are still difficult to solve. From the application perspective, how to effectively utilize unstructured data, such as audio and video content, and how to balance the use of models in high-frequency and low-frequency problems are also core challenges. At the same time, data security and compliance are also issues that the industry needs to face.

Generally speaking, although generative AI technology has made progress in some aspects, it still needs to follow its development trend, be combined with actual applications, and be continuously adjusted and improved.

Zhang Yahui, Co-founder and CMO of Yunbat Intelligence

Zhang YahuiShe stated that in the face of the high popularity of generative AI technology, although many enterprises are eager to embrace this new technology, they actually face many challenges in both entrepreneurship and application.

First of all, the training complexity of large models is high and requires a large amount of computing power and resources. Although this may be solved in the future with technological progress, it is still a significant bottleneck in the near term. Secondly, with the application of AI, issues such as citizen privacy and information security are becoming increasingly prominent, which requires corresponding regulations for restraint and management. In addition, when dealing with abnormal or complex scenarios, the anti-interference ability and stability of large models still need to be improved. Finally, although some enterprises have begun to explore the application of large models in vertical fields, most are still in the primary stage. How to truly create value for customers and how to deeply integrate technology with actual business are still huge challenges. Generally speaking, although generative AI technology has brought great opportunities, how to overcome the above challenges and truly realize its application value is a topic that enterprises in all industries need to think deeply about and explore.

Dr. Deng Xinyu, Head of AI at Zhiben Medical

Dr. Deng XinyuShe stated that under the current wave of generative AI technology, although many enterprises hope to embrace and apply this new technology, the challenges they face are still huge. First is the technical aspect: computing power resources and talents are current bottlenecks. Especially in the domestic environment, it has become more difficult to obtain high-performance GPUs, and cultivating talents specialized in large model research and development is still a major problem. In addition, the lack of real-time learning and long-term memory of large language models is also one of their application limitations. At present, large language models are more like a fixed version, lacking the ability to learn instantly, and there are no good solutions in the short term. Methods similar to vector library updates are only a transitional solution.

Moreover, commercial implementation is also a major problem. Although the accessibility of technology is gradually increasing, how to truly utilize it and create competitive advantages is a problem that every enterprise must face. In a highly competitive market environment, how to maintain and expand one's market position, especially when facing large companies with strong capital, is something every startup needs to think about. In the short term, an enterprise may be able to establish an advantage by relying on its professionalism in a specific field, but in the long run, how to continuously innovate and maintain an industry-leading position will be the key.

Overall, whether in terms of technology or commercial application, generative AI faces many challenges. For enterprises, how to overcome these challenges and utilize the opportunities they bring will be the key to their future development.

Li YangHe stated that from the perspective of investment institutions, the domestic AI field faces several major risks and challenges. First, although many large domestic companies have released advanced models, their actual capabilities are still lagging behind GPT-4, showing an obvious gap compared to foreign countries. Second, policy risks cannot be ignored either. The formulation of specific policies has set unclear boundaries for the AI industry, leaving many companies in an uncertain state and hesitant to easily expand their business scope. In addition, at the application level, the selection of client parties is not entirely based on technical advantages but more depends on trust in suppliers. Uncertain AI output results may lead some clients to be reserved. In scenarios where there is low tolerance for the authenticity and accuracy of AI, penetration will be slower. Finally, investors are increasingly paying attention to whether enterprises are AI-native; companies that simply rely on technological hype for simple optimization may find it difficult to obtain long-term investment.

Executive Director Cui Nan:Every challenge hides an opportunity. Against this background, if we successfully overcome these challenges, what do you think our future development trend will be? From the perspective of these big names, how will the technical development path of generative AI evolve? What do you expect the future product form and application scenarios of generative AI to be like?

Hu JianlinHe stated that from the perspective of generative AI technology, we observe that although the capabilities of large models are enhancing, future complex tasks cannot be solved by a single large model. The solution strategy will involve the integration of multiple technologies and multiple models, and an open-source framework for multi-Agent (intelligent body) provides a direction for this. In the future, product forms will mainly be divided into two categories: tool-based products based on chat, such as ChatGPT. These general-purpose tool products have wide applications but lack industry-specific characteristics; the second category is domain large models, which are targeted at specific industries such as finance, government affairs, healthcare, and law, providing more commercially valuable solutions for knowledge-intensive and highly privacy-conscious industries.

When choosing an implementation field, we consider three dimensions: task fault tolerance, domain professional knowledge requirements, and scenario value. For example, for high-risk tasks, the combination of human and machine is an initial and effective choice. Generally speaking, although current AI applications still have their limitations, generative AI is undoubtedly a huge trend. For the long-term prosperity of the industry, all parties should join hands to jointly promote the development and positive cycle of this ecosystem.

Zhang YahuiShe stated that the technical trend of classical AI in vertical fields is clearly reflected in two major directions: intelligence and anthropomorphism. Taking intelligent voice interaction and intelligent customer service as examples, future digital intelligent bodies need to more accurately understand the real intentions of customers in complex scenarios, especially in dealing with rich Chinese contexts and meanings. The pursuit of anthropomorphism is not limited to a more natural dialogue style but also involves the depth of emotional interaction, making it closer to the real-person communication experience.

With the progress of technology, it is foreseeable that in the future, enterprise official portals will be replaced by digital intelligent bodies instead of traditional text to interact with customers. In addition, telephone robot customer service may be upgraded to video form digital intelligent bodies. In specific vertical fields such as the insurance industry, digital intelligent bodies can serve as training lecturers, accumulating and transmitting enterprise knowledge. However, at present, most enterprises' solutions still remain at the level of simple model combination. In the future, the demand for in-depth exploration and application of professional scenarios will be more crucial. Our goal should be to truly transform "human intelligence disability" into "artificial intelligence" and make substantial contributions to society.

Dr. Deng XinyuShe stated that from the development trend of generative AI, the future hotspots will mainly be in two directions: one is physical intelligence, and the other is virtual metaverse. In the field of physical intelligence, large language models will be more combined with physical hardware to expand application scope. The challenge lies in how to localize large language models and effectively integrate them into physical carriers such as robots, robot dogs, or other carriers, endowing them with the ability to interact and learn with the environment. In the field of healthcare and big data, nursing and housekeeping robots show great potential.

On the other hand, the metaverse will surely return to the center stage as the focus. The key to its development lies in solving user access problems, such as the progress of brain-computer interface technology, cost control and performance optimization of AR/VR/MR technology. Once these technologies make breakthroughs, the metaverse will truly rise and drive the rapid evolution of large language model agents in the metaverse, and even strong artificial intelligence may quickly appear in virtual worlds. Facing future opportunities and challenges, in the field of physical intelligence, it is necessary to continuously improve multi-modal technology, miniaturize computing hardware, and solve the real-time learning problem of large language models. In the field of metaverse, the innovation of XR hardware and the development of brain-computer interface technology will play a key role.

Li Yang, Senior Vice President of Haishong Capital

Li YangShe stated that from the application perspective, we can break through several directions. First, the characteristics of the new generation AI lie in in-depth user semantic understanding and multi-round dialogue capabilities. This enables AI to interact efficiently with the elderly or people with low educational levels, increasing interaction efficiency by ten times or even twenty times. Secondly, AI's design ability will not replace designers but can empower strategies and processes, greatly improving efficiency. Moreover, the fragmented digital data in future society provides great opportunities, and AI has the ability to integrate and efficiently utilize these scattered fragmented information. Finally, large models can serve as a central control platform, combining expert models, internal models, and small models to build a comprehensive control system, which may be a valuable direction.