This article is based on the keynote speech "AIGC-Driven High-Performance Business Practice" by Guo Rongfeng, the person in charge of Sensors data intelligence business.
The following are the main points of this article:
People who understand business and can communicate with AI will become the super individual of the company and exert greater value
Sensors Data provides support for the key roles of the company by building data analysis copilot and user operation copilot, making it a super individual
The key to improving the application effect of AIGC is enterprise data. In the era of AIGC, enterprise data is becoming more and more important
1. Understanding of AIGC business applications
Regarding the paradigm of AIGC, we believe that on the one hand, it is LUI (Language User Interface, natural language user interface), and on the other hand, it is an expert model. For example, in painting, we want to draw a "beautiful morning", LUI understands the need "beautiful morning", and the drawing model draws the picture; for example, let AI imitate Stefanie Sun's voice to sing XX songs, LUI understands the need "sing XX songs", and the singing model completes the singing. The current large models can better understand natural language, and LUI is relatively mature; but at present, not all industries or scenarios have corresponding AI expert models, and some are realized through traditional methods. With the development of AIGC capabilities, various AI expert models will become more and more mature. To quote Dr. Lu Qi, we are now at the inflection point of a new paradigm where "models" are ubiquitous, and the next era is the world of models.
From the perspective of business operation, what impact will the LUI + expert model bring? Indeed, it can improve the work efficiency of employees in specific scenarios, such as automatically writing meeting minutes; but from other perspectives, is it possible to have greater value?
Let me share a story first. I have a friend in the publishing industry who mainly does illustrations for books. She is the CEO of her small company. She has not been trained in drawing for a day. After applying AIGC, she has become the most creative and productive person in the company. I asked her why, and she said: Drawing is divided into two steps. The first step is to design, that is, what to draw, and the second step is to draw, using well-trained muscle memory to draw what you want. Although she doesn't know drawing skills, her years of experience in the industry have given her a good understanding of customer needs and illustration design, and because she has rich drawing knowledge, she can clearly express design needs to AI, and then AI will complete the specific creation. Knowing business and having the ability to talk to AI makes her the most powerful productivity and the best creative staff.
In this story, the LUI + expert model replaces the painter who can only provide basic skill value, and the business leader or senior painter who understands business & knows how to talk to AI communicates with AI, directly conducts business, exerts greater value, becomes a super individual in the company, and naturally improves the efficiency and quality of the company's business.
Therefore, from the perspective of enterprise management, it is possible to build a LUI + expert model to serve key roles in the company, create a super individual in the enterprise, and improve the efficiency and quality of business operations.
2. Implementation of AIGC
Regarding the implementation of AIGC business, on the one hand, how to improve work efficiency in common scenarios? On the other hand, how to empower people with business capabilities in the company to give full play to their value, so that they can become super individuals of the company and drive business development.
In the closed loop of business consisting of perception, decision-making, action, and feedback, business leaders, as key roles, pay more attention to perception and decision-making. We set these two links as the focus of AIGC applications to help them exert their value and achieve super-individuals.
Most of the current enterprise workflow is that analysts produce data reports and reports. Business leaders see the data reports, and data analysts analyze and report abnormal data performance. It is difficult for business personnel to directly access data, which limits their insight into the overall business situation. In addition, when making user operation decisions after insight into data, how to better understand past strategies and formulate new operation strategies based on current goals also has a high threshold. In response to these situations, we can empower business leaders through AIGC.
In the perception link, the focus is on the deep integration of data analysis models and AI, building data analysis expert models, supporting conversational index queries, quickly viewing the desired data in Q&A, and helping super-individuals to easily obtain data and understand business status (what); support conversational data analysis, expert models interpret indicators and formation, and help business attribution (why). In the decision-making process, the focus is on the integration of marketing models and AI, building a marketing expert model, and supporting conversational audience selection and marketing strategy generation.
We have built an intelligent assistant, which is currently in the Demo stage, starting from the data analysis scenario, and now it is in the form of a browser plug-in. Taking the event analysis scenario as an example, use natural language to enter the data indicators to be obtained in the input box, such as the number of users who searched and clicked in the last 7 days. The GPT model converts the natural language into request query JSON and initiates the query, and displays it graphically. Why do we adopt text2json instead of text2sql? Because this has two advantages. On the one hand, it is easier to understand, which is convenient for business personnel to judge the query; on the other hand, it is easier to perform human intervention. For example, if the generated query JSON is incorrect, if you want to change the calculation method or query conditions to see how the indicators are, you can quickly adjust them.
How to achieve this? First of all, let GPT understand the schema and tasks of our data, so what we have to do is to pass our schema to GPT, but because of the length limit, the first problem we have to solve is how to make the prompt shorter. We first filter out the fields that enter the prompt from thousands of fields in the report. Secondly, the filtered schema will have a lot of fields. Too many fields will affect the accuracy and accuracy of GPT. Therefore, it is necessary to interact with GPT to let it select which fields are related to the product. Finally, generate JSON through GPT, and for complex queries, you can let it generate a structure first, and fill in the content under this structure. Therefore, the query process is relatively complicated.
In addition, it can also assist in the writing of SQL. Some analysis is difficult to do with JSON queries, so SQL is used to make custom queries. Write the desired query in natural language in the input box, and the GPT model converts the natural language into SQL and initiates the query. If the SQL is wrong, you can modify it, or optimize it through multiple rounds of dialogue.
In addition to the analysis model, we also made a help question and answer assistant to help business people learn to use the product more quickly. Users can ask some product questions in natural language, and GPT answers the questions based on our product help manual. For example, what can be done by entering Session analysis, the answer is listed above, and the reference of our product help document is quoted. GPT makes queries and generates summaries based on queries. The summarization effect and relevance are better than traditional search effects. For example: query "What is the relationship between the overview and bookmarks?" In this scenario, the GPT summary describes the definition and relationship between the overview and bookmarks in detail, while ordinary search does not have such capabilities.
3. Cognitive summary of AIGC practice
The above introduces some of the applications we have carried out around AIGC, and then discusses the problems we encountered and the cognition we gained in practice.
Let’s talk about model accuracy first. Compared with the summary generation model and image generation model, the data analysis model requires a lot of precision. The quality of summary generation and image generation can be judged intuitively, but it is difficult to judge the right or wrong of data indicators intuitively, so the accuracy requirements are very high. What are the factors that affect the accuracy of the model?
First, the application method of the model. We need to know how to write the prompt to affect the performance and effect, such as whether to step by step, whether to self-calibrate, whether to generate JSON structure first, etc.
Second, the reasoning ability of the model. The most common ones are the logic errors of the last 7 days, year-on-year, month-on-month, and some aggregated statistics, such as queries such as "the number of users who have successfully taken photos in the past 7 days and the number of times exceeds 10".
Third, the prompt length constraint problem. Our events, attributes, etc. may have hundreds or even thousands of fields in combination. We cannot stuff all the table information into the prompt, and we need to select events and attributes first.
Fourth, field understanding issues. In order for GPT to understand events and attributes, it is necessary to give the model enough information, such as the event show_take_photo_guide_popup, we need to let the model know that this is "the pop-up guide after the first photo".
Fifth, business understanding issues. There are two types of business understanding. One is the understanding of business indicators, such as conversion rate. Different companies have different definitions (numerator and denominator) of conversion rate, and the model needs to know the meaning of business indicators.
How to solve the above problems? On the one hand, we need to wait for the improvement of the ability of the large model, or make a model in a certain field by ourselves. On the other hand, regarding field understanding and business understanding, how to accumulate data becomes critical. In the case of LUI, data completeness is very important. In addition to more data accumulation, knowledge is also required, such as which of the two fields with similar semantics should be chosen? This is not data, but a problem that requires knowledge to solve. Therefore, it is necessary to build a metadata knowledge map of the enterprise. For example, there are 1,000 reports in the enterprise, which fields and attributes have been used, and what is the relationship between these fields, attributes, and indicators? These can be constructed and used for event and attribute selection reasoning.
Let's look at product design and evaluation again. After LUI, product design will change from scenario and function-driven to user-question-driven. This has brought about two impacts. On the one hand, the way of user guidance has changed. Users need to be guided to become able to ask questions and ask questions correctly. User guidance runs through before, during, and after asking questions. On the other hand, the evaluation method of products has changed from "whether there is a certain function" to "whether certain questions can be answered correctly." To solve these problems, on the one hand, it is necessary to improve the ability to guide users, and on the other hand, it is necessary to build a knowledge base of industry issues and related issues.
Based on the previous discussion, we summarize the AIGC application method. When an enterprise makes a decision, it is determined by both technical feasibility and business needs. From a business point of view, you can't do AIGC just because you want to do AIGC. You must have a business starting point, and you need to start from the scene and business application. Technically speaking, data is important, and it can improve the accuracy of large models in specific scenarios. To sum it up: the application scenario is the starting point, the data is the core, and the large model is the support.
✎✎✎
【More content】
How can data intelligence help improve sales efficiency?
10-page quick overview of "generative AI" capability boundaries and commercialization
Sensors Data announces access to Baidu Wenxin Yiyan Capabilities
▼ Click "Read the original text" to understand Sensors data