Network Security Internet Technology Development Database Servers Mobile Phone Android Software Apple Software Computer Software News IT Information

In addition to Weibo, there is also WeChat

Please pay attention

WeChat public account

Shulou

Tencent Tang Daosheng: embrace the industry model to promote the high-quality development of the industry

2025-01-18 Update From: SLTechnology News&Howtos shulou NAV: SLTechnology News&Howtos > IT Information >

Share

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

"the rise of the big language model makes it more possible for AI to benefit on a large scale. Over the past period of time, our industry large model products have been recognized by many customers, which have really helped users improve their abilities. At the same time, customer demand has also promoted the continuous upgrading of Tencent Cloud MaaS, and our large model store is constantly 'expanding' to meet the different needs of different enterprises."

On September 7, 2023 Tencent Global Digital Ecology Conference was successfully held in Shenzhen. In the special session of "Industry Big models and Intelligent applications", Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of Cloud and Intelligent Industry Group, delivered a keynote speech, focusing on the far-reaching impact of technology products such as AI models on industry development, as well as the latest strategic thinking and capability upgrading of Tencent Cloud industry models.

Tang Daosheng, senior executive vice president of Tencent Group and CEO of Cloud and Smart Industry Group

Tang Daosheng mentioned that the large model of the industry has passed the "taste period" and is currently going deep into various business scenarios to promote the intelligence of the whole chain of enterprises. At the same time, in the landing process of large model industrialization, customers not only pay attention to the size and function of the model, but also pay more attention to how to select and customize model products suitable for their own business development.

Based on the actual needs of customers, Tencent Cloud continues to upgrade large model selection stores, including self-developed general model "mixed element", more than 20 mainstream open source models and more industry large models. Enterprises can choose appropriate model products according to their own needs, and carry out training and fine tuning to meet individual needs.

In addition, Tang Daosheng also believes that the upsurge of models has led to rising costs, and embracing the big model should not only be impulsive, but also rationally consider landing costs. In view of this, Tencent Cloud also provides one-stop solutions from data processing, multi-machine and multi-card training to hardware optimization to help enterprises create and use large models efficiently and at low cost. At the same time, Tencent Cloud TI platform has also been upgraded to effectively improve the training speed and reasoning efficiency of large models.

"the development and landing of products such as AI models will have a 'qualitative' impact on thousands of industries. We will also continue to open product capacity and explore technology applications to help the industry achieve higher quality development." Tang Daosheng said.

The following is the full text of the speech:

Hello everyone!

Welcome to the Tencent Global Digital Ecology Conference "Industry Big Model and Intelligent applications" to discuss the AI-driven industrial development.

Artificial intelligence has been developed for more than 70 years. In the previous waves, although AI has landed in some industrial applications, it is still far away from the common industrial applications due to the limitations of basic algorithms, computing power, data and so on.

The development of the big language model for the first time let us see the possibility of AI landing on a large scale in the industry. Six months ago, many companies marveled at the generative ability of the GM model and couldn't wait to try to integrate with the business. However, it is soon found that the general large model is faced with many challenges in practical application, such as professionalism, accuracy, data security, cost and so on. Based on this, Tencent Cloud launched Tencent Cloud MaaS service on June 19 to help industry partners create large model products that suit them through the "industry large model + enterprise data fine tuning" program.

Over the past three months, our idea of building a large model of the industry has been generally recognized by the industry. Many enterprises have greatly enhanced the ability of content production, marketing and customer service through our large model of the industry, and customers are also with us. Dig out more innovative applications.

Technological changes and the rapid development of customer requirements have not only brought more opportunities for the development of large models in the industry, but also promoted the rapid iteration of Tencent Cloud's MaaS service capabilities. Here, I would like to share with you some of the development trends of large models that I have observed.

First of all, the industry model is going deep into all kinds of business scenarios to promote the intelligence of the whole chain of enterprises.

Six months ago, customers came to talk about large models, and the application scenarios they could think of were basically text customer service. But today, the application scenario has been rapidly extended to various fields, for example, in the financial industry, the large model has been used in account opening, business processing, risk control and other scenarios.

Based on these customer needs, we try to build full-stack product capabilities with large models, which can be used in different aspects of the business to help enterprises improve the quality and efficiency of the whole chain.

Take the financial industry as an example, the handling of a large number of documents is a prominent pain point for financial salesmen. This job is low in skill, but very important, has a low fault tolerance rate, and consumes a lot of time every day. A domestic head commercial bank came to us and used the TI-OCR platform with large model capability to identify multiple forms such as orders, invoices, applications and so on with only 50 labeled data. The accuracy rate is more than 95%, and the data required is greatly reduced. It not only saves a lot of time for business staff, but also can automatically extract core tags, generate electronic data files, and carry out follow-up business analysis.

Risk management is the lifeline of financial business. Financial institutions have to upgrade their risk control models once in a few months. The traditional modeling process is time-consuming and labor-consuming. We also put the multimodal risk control knowledge that Tencent has accumulated for many years. Precipitate into the financial risk control model. Through the hint of a small sample, financial institutions can automatically generate their own risk control models according to changes in the customer base and the market. The whole modeling process has been reduced from 2 weeks to 2 days, and the overall anti-fraud effect has been improved by about 20%.

In customer service, the application of large model is not only text dialogue, but also the combination of exclusive big model and digital Homo sapiens, so that virtual customer service can interact with customers smoothly and in real time in an anthropomorphic way; coupled with face core technology based on audio and video, and image recognition, financial business can be handled online efficiently. According to industry statistics, manpower costs account for 60% and 70% of the operating costs of financial customer service centers. Citic Construction Investment uses Tencent Cloud's intelligent solution to complete 95% of the new user account opening work.

Secondly, with the industrial landing of the large model, people pay more attention to how to find a model base that better matches their own business, and train the exclusive model that meets their own needs.

This drives us to constantly upgrade large model stores to provide new "1+N+N" services to meet the different needs of different enterprises.

"1" represents our Tencent mixed meta model. In the main forum of the conference, we officially announced the self-developed general model-Mianyuan. A number of businesses and products within Tencent have been accessed and tested, and achieved good results, and more businesses and applications are gradually being connected. Mianyuan is not only an important supporting base for Tencent industry model, but also open to the industry. Through our large model store, customers can use mixed elements to train their own models and integrate seamlessly with existing business systems with the help of API open capabilities.

Then the first "N" means that in addition to the mixed element, we provide customers with the latest and most popular open source general large models in more than 20 industries, such as Llama 2, Falcon, Bloom, and so on. At the same time, our TI platform has also supported the training and reasoning of these open source models. Based on Jupyter Notebook, customers can quickly start fine tuning of the model, and model deployment can be completed through low code operations.

The second "N" refers to our industry model. On the basis of the general model, through the re-processing of industry data, we can provide more professional and accurate services for industry customers. At present, our industry model, from the early literature and travel, pan-mutual, retail and other fields, has rapidly expanded to more than 20 industries, such as energy, consumer electronics, and so on, and covers production, sales, customer service and other links.

Enterprises can select appropriate models in the large model selection store, import enterprise-specific professional documents and enterprise data through Tencent Cloud intelligent TI platform, and do further training and fine tuning to quickly generate more targeted models to better meet the individual needs of enterprises. At the same time, whether it is built on the public cloud or privatized deployment, we can do a good job in permission control and data, so that enterprise users can feel more at ease when using the model.

Thirdly, the upsurge of models leads to rising hardware and labor costs, more and more enterprises realize that embracing large models is not only impulsive, but also rationally consider landing costs, training and reasoning efficiency.

The generation of enterprise exclusive model involves many links, such as digital asset resource management, data tagging, training, evaluation, testing and deployment. At the same time, according to the business development, the enterprise model needs to be constantly tuned and iterated, and the whole process of data processing should be repeated constantly.

How to help enterprises make good use of the model with high efficiency and low cost? Through the fine tuning solution for industry large models based on Tencent Cloud TI platform, we help model developers and algorithm engineers solve data processing problems in one stop, ensure high-quality, efficient, safe and compliant data processing, and create and use large models with high efficiency and low cost through multi-machine and multi-card training.

This time, we have made a new upgrade to the tool chain of the TI platform. The newly upgraded Taiji Angel framework has increased the training speed of the large model by 30% through asynchronous scheduling optimization, video memory optimization, computing optimization and so on, and the reasoning speedup has reached 2 times.

The big model of the industry is coming into all kinds of industries, and a large number of new scenes and new requirements have been activated. How to obtain the underlying computing power at low cost is also a prominent problem that enterprises are facing at present. On the underlying infrastructure of the large model, we constantly optimize the "iron triangle" of servers, networks and data to help customers reduce costs and increase efficiency.

Our new generation of HCC high-performance computing cluster achieves lossless release of GPU computing power with a service agreement level (SLA) of no less than 99.9%. At the same time, based on the cloud native architecture, it realizes the mixed deployment of training and reasoning business, and greatly saves the deployment and training costs of large models.

Our self-developed server inter-machine network, Xingmai, achieves the strongest 3.2T bandwidth in the industry and supports 100 million card cluster networking, so that the communication between GPU is faster, the congestion is less, and the computing efficiency is higher.

We also take the lead in launching vector databases among domestic cloud manufacturers to improve the storage and retrieval efficiency of massive unstructured data, allowing large models to pre-train data classification, de-duplication and cleaning to achieve a 10-fold improvement in efficiency. The data access time of about 1 month can be completed in 3 days, which greatly reduces the cost of the enterprise.

There is no doubt that we are entering an era of drastic changes by artificial intelligence, and AI will create more value through deep integration with the industry. Tencent will continue to open up its technology and capabilities to help the industry embrace intelligent upgrading and achieve higher-quality development.

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.

Views: 0

*The comments in the above article only represent the author's personal views and do not represent the views and positions of this website. If you have more insights, please feel free to contribute and share.

Share To

IT Information

Wechat

© 2024 shulou.com SLNews company. All rights reserved.

12
Report