Shulou(Shulou.com)11/24 Report--

On September 7, 2023 Tencent Global Digital Ecology Conference was held at Shenzhen International Convention and Exhibition Center, focusing on the new trend of the future development of the industry and the latest progress of self-research technology products, showing a full range of industry frontiers and intelligent insights. At the special session on industry models and intelligent applications, Wu Yongjian, vice president of Tencent Cloud, head of intelligent research and development of Tencent Cloud, and head of research and development of Tencent enterprises, delivered a keynote speech on "MaaS helps upgrade applications to reshape productivity in the intelligent era". Based on AI large model technology, Tencent Cloud updates and upgrades nearly 10 intelligent applications and solutions to further speed up the landing of the model industry.

Wu Yongjian, vice president of Tencent Cloud, head of intelligent research and development of Tencent Cloud, and head of research and development of Tencent enterprises.

Wu Yongjian believes that the application of AI large models has given rise to an "efficiency revolution", allowing intelligent applications to show stronger capabilities and higher efficiency. For example, in terms of AIGC applications, Tencent Mianyuan model is officially opened to the public through Tencent Cloud. Users can input instructions to the model through natural language to improve the efficiency of copywriting. Relying on the AI painting model developed by Tencent, Tencent Cloud AI painting can intelligently generate images from input text and pictures, support more than 25 painting styles, jointly provide support for high-quality and efficient content creation, and speed up landing in content communities, e-commerce and other scenes.

At the same time, relying on industry large model solutions, Tencent Cloud has improved the production efficiency of a variety of intelligent applications such as image recognition, data analysis and human-computer interaction, and introduced more diversified plug-and-play products and services to the market. reshape productivity in the intelligent era.

Wu Yongjian said that in the field of image recognition, the TI-OCR training platform integrates OCR large models to help customers solve specific scene recognition problems with end-to-end algorithms, reducing training costs and training time; in the field of data analysis, Tencent's point analysis AI assistant realizes zero-threshold dialogue analysis by analyzing large models, so that everyone can have professional data analysis capabilities and provide decision support for the business front line. In terms of human-computer interaction, through the industry large model hub, we can access a large amount of knowledge and tools, realize the understanding of text, language, video and other multimodal intentions, generate intelligent and accurate answer output, and support more intelligent, efficient and natural intelligent customer service, AI assistant, and digital Homo sapiens applications.

Wu Yongjian said that in the future, Tencent Cloud will continue to promote advanced technologies such as AI models to be applied to internal PaaS and SaaS products to verify the availability and ease of use of MaaS services. On the other hand, we will continue to create more out-of-the-box AI model products and services, break through the whole chain of AI industry landing, and help customers to achieve digital intelligence upgrade.

The following is the content of Wu Yongjian's keynote speech:

Hello, online viewers and media friends. I'm Wu Yongjian from Tencent, and the theme I share today is "MaaS helps upgrade applications to reshape productivity in the smart era."

In June this year, we officially released the Tencent Cloud industry model solution to help customers apply the AI model to the industry scenario. At the same time, we also promote internal practice and launch two products, Tencent Enterprise Intelligent customer Service and Enterprise AI Assistant, based on the industry model, to solve the problem of "unintelligent" intelligent customer service and high threshold for data analysis through technological innovation.

Subsequently, we apply the AI large model capability to a wider range of PaaS and SaaS products. On the one hand, verify the availability and ease of use of our MaaS services. On the other hand, more out-of-the-box products and services will be opened to customers to promote the effectiveness of the large model in the actual production and operation scenarios.

Today, I would like to take advantage of the Digital Ecology Conference to share with you our latest practice.

Relying on Tencent Cloud's years of experience in cultivating the industrial Internet, we proceed from the needs of the industry, based on the self-developed Tencent mixed meta model, and support mainstream open source model applications. High-quality industry models are built into the TI platform to provide customers with model services of different sizes. By upgrading the large model training and fine-tuning the tool chain, we can help enterprises quickly generate their own models, upgrade intelligent applications, and speed up the landing of the model industry.

First of all, I would like to share with you our exploration of AIGC application based on Tencent's mixed meta model.

In the past, in the copywriting scene, the traditional creation took "manual" as the core, which was easy to produce some problems, such as lack of creativity, lack of timeliness, uncontrollable quality and so on. Today, Tencent Mianyuan model is officially opened to the public through Tencent Cloud. Users can generate instructions for the input content of the model through natural language to improve the efficiency of copywriting. Through parameter selection and multiple rounds of input, the generated content is controlled and tuned to improve the effect of copywriting. According to the needs of different scenarios such as marketing, creation and efficiency, the corresponding proprietary content is generated to improve the usability of copywriting. At the same time, it can also connect quickly with advertising platform, UGC platform and SaaS services, accelerate landing, and provide more accurate and efficient commercial copy creation tools for all kinds of scenes.

Similarly, traditional painting creation also needs to spend a lot of time and energy, and heavily depends on the creator's professional level and subjective experience, which brings greater operating costs to the enterprise. We launched Tencent Cloud AI painting, which can intelligently generate images by entering text and pictures, supporting more than 25 painting styles. It is worth noting that relying on the AI painting model developed by Tencent, Tencent Cloud AI painting has a strong Chinese understanding ability, which can better support the image generation of Chinese elements, such as ancient poetry understanding, ink paper-cut and so on. We have also greatly reduced the API interface time, which is 50% lower than the industry average, and provides support for high-quality and efficient content creation.

Going deep into the industry, we find that in the content community, e-commerce and other scenes, the combination of picture and text can provide richer, more personalized and more creative advertising information, and effectively improve the conversion rate. Based on the two methods just introduced, users only need to enter the topic and product description to generate recommended copy, product description and other content with one click. AI painting can also provide creative assistance in content matching pictures, publicity posters and other links, so as to achieve the double efficiency of intelligent marketing and intelligent creativity. Now, you are welcome to apply for a trial of Tencent Mianyuan model, or you can log on to Tencent Cloud's official website to try AI painting.

In addition, in June this year, we also announced a large model solution for Tencent Cloud industry.

The large-scale model of exclusive industry, which is constructed by a large amount of professional knowledge learning and strong reasoning ability, has stronger expressiveness and adaptability in specific fields. Based on this, we further improve the production efficiency of intelligent applications such as image recognition, data analysis, digital Homo sapiens and customer service, and bring more diversified plug-and-play products and services to the market.

In the field of image recognition, in the past, OCR character recognition needs to go through many processes, such as picture input, detection, recognition, structure and so on. On the one hand, multi-process operations can easily lead to error accumulation and reduce the accuracy of the model; on the other hand, the image recognition requirements of each business scenario are different, which makes it difficult for a single model to be reused in the new scene, and each customization needs to increase the cost.

After upgrading to an "end-to-end" OCR model, the recognition result is generated in one step from image to text. For example, for a bill example, you can directly ask "when is the commission date", and the big model gives an answer after understanding the picture, which greatly reduces the recognition process and improves the recognition accuracy. In the case of mixed invoice sticking, one model can support multiple types of image recognition, and the generalization is further improved.

Compared with the traditional model, the OCR large model breaks through the technical limitations and effectively reduces the landing cost. For example, for the problem that handwriting and complex forms are difficult to identify, the OCR large model can improve the recognition accuracy by fully understanding the picture and directly extracting the relevant fields; for pictures with complex background, the OCR large model can use end-to-end algorithm to avoid interference information and answer relevant questions directly. In view of the training cost, in a specific scenario, the traditional model training improvement index needs to rely on more than 2000 training samples for 48 to 60 hours of training, while the large OCR model only needs 50 to 100 training samples for 2 hours to achieve a 3 to 20% effect improvement.

At present, the large OCR model has been integrated into the TI-OCR training platform for customers to fine tune small samples according to their own business scenarios.

In the field of new energy, for example, customers mark pictures through TI-OCR training platform, fine-tune the large OCR model with small sample enterprise data, realize the recognition of curved wire characters, watermark nameplates and complex watt-hour meter pictures, and promote automatic information audit. With the assistance of the large OCR model, the character reading accuracy is more than 98%, and the automated audit is expected to save customers more than 80% of the manpower.

In the data analysis scenario, when users want to know the "recent sales situation", the traditional data analysis process is complex and highly dependent on professionals, and the conclusion can only be generated through the disassembly of analysis ideas and the introduction of data indicators. the link is long and inefficient. Faced with the same problem, the general large model is easy to generate redundant or invalid answers because of its lack of analytical expertise and insight.

To this end, we have created a large analysis model that supports zero-threshold conversational data analysis.

Compared with the general large model, the analysis large model internalizes the professional data analysis knowledge and the analysis index system of different industries, and can carry out intention understanding, train of thought disassembly, data reasoning and generate conclusions from the perspective of analysts according to the characteristics of the industry.

For example, by analyzing the intention understanding ability of the large model, the word sales volume will be associated with "sales" and "number of orders", and "recent" will be associated with "nearly 7 days". Clearly analyze the trend of the number of orders for nearly 7 days.

By analyzing the data reasoning ability of the large model, users will be further advised to analyze the reasons for the changes in the number of orders from different sales channels.

Based on the analysis model, we upgraded and launched Tencent Enterprise Analysis AI Assistant, covering conversational analysis, assisting data configuration, extracting intelligent conclusions, and generating reports with one click. In September this year, Enterprise Analytics AI Assistant officially opened internal test applications to major customers. In the future, everyone will have the ability to analyze professional data, make the data run faster and provide decision support for the front line of business.

In the aspect of human-computer interaction, we use the industry large model center to access a large amount of knowledge and tools to achieve text, language, video and other multimodal intention understanding, to generate intelligent and accurate answer output. Next, I will introduce in detail the exploration of intelligent customer service, voice assistant and digital Homo sapiens.

Around the customer service scene, Tencent Enterprise released a new generation of intelligent customer service in June this year, introducing large model capabilities to improve dialogue interaction, manual assistance and knowledge construction. Recently, we have further upgraded the configuration capability of the management side to improve the efficiency of knowledge construction and optimize the C-side user experience. At the same time, we actively carry out industrial practice with various industries, landing benchmarking cases in the fields of literature, tourism, pan-politics, real estate and so on.

For example, in the aspect of knowledge construction, the closed loop of "import-verification-tuning" of enterprise exclusive knowledge can improve the effect of reading and understanding of complex documents, further optimize the cold start efficiency and reduce the cost of operation and maintenance. On the basis of text documents, we add access to complex documents such as mixed arrangement of pictures and text, multi-column typesetting and so on. Through semantic chunking, vectors are generated to enrich the question and answer combination and content with a highly available and scalable vector database scheme; through the rapid generation of question-and-answer pairs and the backtracking of the original text, the efficiency of question-and-answer verification is improved; by providing dialogue testing and operation tools, operators can quickly achieve tuning and improve the accuracy of question and answer.

At the level of dialogue and interaction, the ability of large models to iterate vector databases and search engines and intelligently generate anthropomorphic answers can enable intelligent customer service to cope with more complex requirements. For example, traditional customer service cannot support complex requirements such as "procedure explanation". Through the completion of context information and question-answering reasoning based on the enterprise knowledge base, the new generation of intelligent customer service can increase the resolution rate of complex problems by 30%.

In terms of user experience, the blessing of large models also makes intelligent customer service closer to "real-life customer service". For example, users' emotions can be accurately identified and appeased in a timely manner. Can distinguish task-based, knowledge-based, chat-based topics, through the way to solve task requirements. In the picture example, we can see that even if the user inserts the topic of "small talk" in the process of business processing, it will not cause the task to be interrupted and ensures the success rate of business processing.

In addition to intelligent customer service, the introduction of the travel model has also greatly improved the semantic understanding ability and interactive experience of the car voice assistant.

The traditional vehicle voice assistant understands the user's request in the way of rule + small model, but there are some problems, such as unnatural dialogue, unemotional mechanical reply, single function and so on. After adding the large model, more natural dialogue ability is provided through complex intention recognition, and more emotional voice interaction is achieved by setting human settings. At the same time, with the iteration of large model capabilities, more travel scenario capabilities will be supported.

For example, based on the large travel model, the voice assistant can automatically generate itinerary planning by asking simple questions. According to the situation of the vehicle, the voice assistant will actively prompt for refueling and add gas stations along the way in the navigation. At the same time, the large model also integrates the body signal information, which can be used for vehicle fault diagnosis for users.

In the field of digital human, the development of digital intelligence technology promotes the digital human from customization to universal benefit. Through the industry model, we further reduce the threshold of digital human application and improve production efficiency and interactive experience.

In April this year, we launched a Homo sapiens factory, which can generate small samples of Homo sapiens in 24 hours. Today, with the blessing of AI large model technology, we once again improve production efficiency and release the "universal mouth shape" version of small samples of Homo sapiens, which can be obtained within 1 hour after uploading materials without training, and the production efficiency of Digital Homo sapiens has been improved again. Although the process and time are greatly simplified, the image effect is still very lifelike.

The blessing of large model technology also enables Homo sapiens to achieve more efficient, natural and intelligent results. Our newly upgraded "interactive" small sample Homo sapiens can pause and change gestures according to the real conversation scene, which is closer to the state of human interaction. In the future, small sample Homo sapiens with interactive ability will no longer be limited to one-way broadcast scenarios such as short video production, but can also be used in interactive scenarios such as service consulting.

In the field of 3D digital Homo sapiens production and operation, relying on AI large model technology, we promote photo-based generation of 3D images, change the form of manual work, and greatly improve the speed of generation. It can also achieve intelligent action-driven, automatically matching actions for 3D digital Homo sapiens based on semantics, so that the service experience of "intelligent employees" is close to that of human employees.

It can be seen that AI large model applications are giving birth to an "efficiency revolution", allowing intelligent applications to show stronger capabilities, higher efficiency, and more scenarios. In the future, we will continue to apply AI models and other advanced technologies to internal PaaS and SaaS products, break through the whole chain of AI industry landing, and reshape productivity in the intelligent era. Let more out-of-the-box AI large model products and services help customers to achieve digital intelligence upgrade.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.