In-Depth | Thoughts and Practices on the Application of Large Models in Data Governance AI

The Copilot integrates BI with chatGPT, allowing users to simply input a statement expressing their intent to analyze specific data metrics; the model then autonomously processes and analyzes the data, quickly transforming raw data into visualized results. Similarly, the iDATA model supports creating a dashboard within minutes and is equipped for private deployment, enabling control over access and permissions, thereby ensuring data security and compliance.

Large models utilize data mining and deep learning techniques to understand business logic and datasets, generating visualizedoutputs without requiring manual code development. Enterprise users can configure component properties through a visual interface to develop business applications, resulting in an average reduction of development time by around 90%. This simplicityand speed make data analysis automation a highly efficient means to bridge the 'last mile' of automation in current practices.

"Attention is all you need" —— attention is the most precious capability. In today's era of information overload, individuals mustpossess highly focused attention to sieve out critical information. Corporate internal knowledge management mirrors this principle. Typically, enterprises establish centralized knowledge base platforms that collect, store, organize, and share vital information and knowledge internally.

Moreover, the application of large language models brings unique advantages to businesses. Such models demonstrate exceptional fluency in understanding and generating natural language, significantly improving accuracy compared to traditionaltechniques. They excel at grasping context and producing more coherent responses, as well as self-learning and updating based on new data. Additionally, enterprise assistants developed using large language models can assist employees in processing business data. For unstructured business data, these models can respond swiftly, greatly enhancing efficiency and accuracy in data governance. Currently, the functionalities of enterprise assistants powered by large models are gradually permeating every aspect of companies.

Mr. Hu Gang illustrated through case studies the online business data processing capabilities of iDATA, also addressing issues concerning data interaction between master data systems and business systems. He pointed out that large models can serve notonly as powerful personal assistants but also provide emotional value in both professional work and personal life.

Regarding tips for querying large models, Mr. Hu offered some suggestions. Users should strive to provide clear and concise prompts, avoiding overly complex or difficult-to-understand sentences and jargon. Including contextual information in the prompt helps GPT better comprehend tasks and objectives. It's crucial to offer detailed instructions and explanations to ensure GPT can accurately complete tasks, which involves using explicit language and specific examples.

Firstly, Mr. Hu Gang quoted the lecture views of Professor Peng Kaiping, pointing out that today's children should learn to think,ask questions, and innovate, rather than simply memorize facts. With the continuous advancement of technology, we should also adjust our approach to acquiring knowledge and information, learning to utilize questioning skills and posing high-quality questions. Secondly, Mr. Hu discussed the concept of "new-quality productive forces" mentioned by General Secretary Xi Jinping, emphasizing that digital technology innovation-driven development is the main direction for forming new-quality productive forces. We should keep up with industry trends, stay informed of the latest technologies and trends, and pay attention to changes in the labor market to allocate human resources reasonably and avoid structural unemployment.

Question 1: With the introduction of large models, the personnel structure of enterprises may need to be adjusted. Mr. Hu Gang, what do you think will be the new job positions that enterprises will recruit specifically for large models in the future? Will they specifically hire prompt engineers?

Mr. Hu Gang: Prompt engineers are known as "those who know the spells." With the increasing use of large models, enterprisesare bound to increase their recruitment of prompt engineers. In addition, enterprises will also need to recruit more artificial intelligence trainers and data annotators. AI trainers need to be able to analyze product requirements and relevant data to establish data annotation rules, ultimately achieving the value of "improving the quality and efficiency of data annotation work"and "accumulating general data in specific fields." Data annotators, on the other hand, need to be able to handle various types of data, such as text, images, audio, and so on.

Among the three essential elements of AI applications - computing power, algorithms, and data - the preparation of training data can be considered the most crucial link. Andrew Ng, the founder of Google Brain, once pointed out, "80% of AI research should be focused on data preparation, ensuring data quality is the most important task. If the industry emphasizes data-centeredness rather than model-centeredness, the development of machine learning will be faster." Therefore, there will be an increased demand for personnel related to improving data quality, while low-knowledge and repetitive jobs, such as human customer service, may shrink due to the impact of large models.

Question 2: The application of large models can effectively improve work efficiency compared to traditional methods. How then should the work efficiency of employees using large models be measured?

Mr. Hu Gang: I believe we can measure the work efficiency of employees using large models from several perspectives.

Question 3: In the PPT, it was mentioned that since China's 'population dividend' has not yet faded, the country would not preemptively invest in 'AI technology,' instead focusing more energy on research and development in areas such as new energy vehicles and chips. According to your predictions, where will China's future developments in data governanceAI technology mainly concentrate?

Mr. Hu Gang: The viewpoint expressed in the PPT reflects my personal opinion. Current chatGPT technology can lead to a significant reduction in job positions, causing the value of many ordinary workers, like CRUD Boys, SQL Boys, and news editors, to plummet dramatically. CRUD refers to the acronym for the basic operations in computing—Create, Read, Update, and Delete. CRUD is primarily used to describe fundamental operation functions within software systems or databases. While in China, chip manufacturing, new energy vehicles, and large aircraft projects can provide an abundance of job opportunities.

Regarding our future AI technology for data governance, the focus will not lie in generic AI applications like chatGPT; rather, it will be placed on proprietary large models and specific industry applications. Although the US may lead in artificial intelligence theory and groundbreaking technologies, ChatGPT, at present, is largely regarded as a 'toy,' lacking the ability to create substantial economic value. In contrast, China excels in practical AI applications, emphasizing value creation and driving continuous progress through these applications.

Huawei's Pangu large model, though not as showy as chatGPT, specializes in finance, government affairs, manufacturing, mining,meteorology, railways, and has already achieved remarkable accuracy in global weather forecasting, with instances even surpassing local manual forecasts. JD.com's YanXiao focuses on retail, finance, and supply chain logistics, NetEase Youdao's ZiYueconcentrates on education, Ctrip's WenDao specializes in tourism, and Yonyou's YonGPT is dedicated to human resource management. It is said that China currently has over 70 large-scale models under development, launch, or operation, ranging from billions to even higher parameter levels. Huawei's Pangu remains a general-purpose large model, while JD.com's YanXiao, NetEase Youdao's ZiYue, Ctrip's WenDao, and Yonyou's YonGPT are all domain-specific models with direct industrial backgrounds. They possess readily available data and application scenarios and can swiftly feedback and iterate to drive value creation—a starkly different situation from general-purpose models seeking applicable scenarios before generating value.

Moreover, specialized large models start by addressing individual problems and gradually evolve to tackle common issues, differing from general-purpose models that typically address common problems first and then adapt to specific ones. However, they might require a longer time to converge on common problems.

This distinction represents the difference between theoretical scientists and engineers. While theoretical scientists are celebratedin history, engineers are the ones who build bridges and pave roads, benefiting communities directly. Technological advancement leads to industrial upgrading, eventually translating into increased income levels, enhanced consumer confidence,and economic growth driven by internal dynamics.

Question 4: Will the introduction of large models cause employees to lose their independent ability to process data? In the future, will employees only need to know how to use models without needing to understand the underlying computational logic?

Mr. Hu Gang : The introduction of large models will not result in employees completely losing their ability to independently handle data. Instead, in the future, employees' competitive edge lies in being adept at using tools; however, to use them effectively, they must indeed grasp the underlying computational logic. Take, for example, applying Excel formulas for bid evaluation; one must already have a broad framework in mind, and then let ChatGPT fill in certain specific details. Moreover, regardless of whether it's Excel scripts or Python/Java code, these codes often contain numerous errors.

While ChatGPT possesses programming capabilities, reports have shown that its generated C++ code can be problematic, oftencontaining security vulnerabilities, and ChatGPT does not actively alert users about these issues. In knowledge base construction, ChatGPT serves as a tool that efficiently provides supporting evidence once you've already formed a well-considered perspective. The judgment and insightful wisdom still reside with the user.

ChatGPT performs well in induction and summarization, but its performance in deductive reasoning is less impressive, sometimes providing seemingly plausible but ultimately unfounded statements. During a two-day training session on Snowflake at a client's site, having studied it for barely a month myself, I found that although I could answer some questions posed by the clients, my responses were insufficiently supported and not particularly compelling. I directly consulted ChatGPT for additional arguments. ChatGPT would compile a range of evidence from various sources, and with my own judgment and logical reasoning, I could then deliver much more engaging responses to the training attendees.

When utilizing large models, such as with an AI trainer, the practitioner still needs to analyze product requirements and related data to formulate data annotation rules, ultimately aiming to enhance the quality and efficiency of data annotation work and accumulate valuable universal data in niche domains. Thus, it remains essential to understand the underlying logic and possess domain-specific knowledge or industry know-how.

Question 5: Teacher, what aspects will iDATA (iDATA) focus on in its future development?

Mr. Hu Gang: Our future focus will be on the implementation of data development Copilots and the construction of knowledgebases for industry-specific large models. More specifically, I envision a strong emphasis on the financial industry's large model sector. Currently, in the context of major B2B business domains, gpt achieves an accuracy rate of approximately 80%. We aim to elevate this accuracy rate in business domains to above 90%. In the medium term, we will work on expanding our presence in both business domains and across industries. Long-term, we will pay closer attention to the direction of intelligent agents.