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

On September 7,2023 Tencent Global Digital Ecology Conference was successfully held in Shenzhen. At the meeting, Wu Yunsheng, vice president of Tencent Cloud (Cloud), intelligent person in charge of Tencent Cloud and head of Youtu Laboratory, delivered a keynote speech, focusing on sharing the latest capability upgrading and landing practice of Tencent Cloud industry large model.

He mentioned that industry practice has become an important criterion for measuring the value of large models as they step out of the "early stage" of exploration. "In the second half of intellectualization, Tencent Cloud (Cloud) Intelligence has been focusing on industrial practice while insisting on investing in AI technology research and development, emphasizing the pragmatic feasibility of technology and the safety and reliability of products. "

Based on this, in the past period of time, Tencent Cloud (Cloud) MaaS capability has been continuously upgraded, hoping to help customers apply large models faster and more economically, so as to seize the development opportunities in the era of AI large models.

(Wu Yunsheng, Vice President of Tencent Cloud (Cloud), Intelligent Director of Tencent Cloud and Director of Youtu Laboratory)

Wu Yunsheng believes that the application of AI model is gradually penetrating into all links of thousands of industries, including design, R & D, production, marketing and service. From conversational customer service to marketing content generation to cross-modal retrieval and risk control, the application of large models is demonstrating greater capabilities, higher efficiency, wider application scenarios and deeper potential. A "productivity revolution" driven by the "big model" is quietly taking place.

Under this background, Tencent Cloud (Cloud) starts from the actual demand of industrial customers, based on the self-developed Tencent mixed large model, supports the mainstream open source model application in the industry, continuously upgrades MaaS capability from three dimensions of technology base, platform capability and intelligent application, helps customers to build exclusive large models and intelligent applications, and promotes the implementation and application of large models in the industry.

At the conference site, Wu Yunsheng also carried out detailed explanation one by one for the three major upgrade points. The upgrade of technology base enables Tencent Cloud (Cloud) to provide customers with more powerful and diversified basic models; the upgrade of TI platform tool chain ensures efficient and stable operation of training tasks, and provides task management, model optimization and deployment support, making it easier for customers to implement model landing applications; Through the upgrade of intelligent applications, Tencent Cloud provides customers with a variety of plug-and-play products, which combine AI large model technology with actual business to provide customers with more efficient, natural and intelligent solutions.

As an important part of Tencent's industrial Internet ecosystem, Tencent Cloud (Cloud) Intelligence always emphasizes the importance and necessity of "cloud-intelligence integration." Committed to achieving a balance between technology research and development and industrial practice, and constantly expanding the "depth" of AI application in the industrial field.

Wu Yunsheng said,"In the future, we will continue to polish MaaS capabilities, join hands with customers and partners, and jointly embrace the 'efficiency revolution' in the era of AI large models." "

The following is the full text of the speech:

Good afternoon everyone!

I am Wu Yunsheng of Tencent. Today, my sharing theme is "Tencent Cloud (Cloud) MaaS Upgrade, Deepen Industry Promotion" Efficiency Revolution."

Just now you also saw Tang Daosheng and Mengniu Lu total number of Homo sapiens interaction.

After accessing AI large model capability, digital intelligence clone can generate answers in real time according to questions, and match corresponding expressions and actions. This is also one of the results of our MaaS capability upgrade in the past period.

More than two months ago, we officially released the Tencent Cloud (Cloud) industry model solution, dedicated to building a model selection store and providing one-stop MaaS services for customers. We believe that the industrial scenario is the best training ground for large models, and this direction soon becomes the consensus of the industry.

In the past period of time, we have also continued to iterate and upgrade MaaS capabilities, hoping to help more customers, quickly and low-cost landing large models, and seize the development opportunities of the AI large model era.

At present, the development and application of AI large models have brought new increments to the entire cloud service market. The application of large models is gradually running through the whole process of enterprise design, R & D, production, marketing and service, helping enterprises innovate, helping employees improve efficiency and accelerating the process of cloud intelligence integration.

At present, the development of large models in China can be described as "blooming." According to incomplete statistics, there are more than 70 large models with parameters above 1 billion yuan; after a period of development in the past, the exploration of enterprises in the direction of large models has passed the "early stage," and "industrial landing" has become an important standard for evaluating the value of large models. At this morning's ecological conference, Tencent also officially announced its self-developed universal model, providing a solid base capability for our industry's large model service.

Especially for the in-depth application of the large model in various fields such as finance, retail, culture and tourism, government affairs, travel, etc., the effect of "improving quality and increasing efficiency" has gradually appeared. The application scenarios of the large model also cover the core production and operation links of enterprises such as dialogue customer service, marketing content creation, cross-modal retrieval, risk control, etc., showing stronger ability, higher efficiency, wider scenarios and deeper application. A productivity revolution driven by the "big model" is taking place.

Relying on Tencent Cloud's years of experience in deeply cultivating industrial Internet, we continuously upgrade MaaS capability from three dimensions of technology base, platform capability and intelligent application, assist customers to build exclusive large models and intelligent applications, and accelerate the implementation of model industry based on self-developed Tencent mixed model from industrial customer demand scenario and support mainstream open source model application in the industry.

We create a model selection store that offers customers models of different sizes. By upgrading large model training and fine-tuning the tool chain, customers can build their own models faster and at lower cost. At the same time, Tencent Cloud (Cloud) has a very wide range of PaaS and SaaS products, which are very good test fields for our industry model.

By upgrading the technology base, we can provide customers with more capable and diversified base models.

The large model of Hunyuan builds a solid base for Tencent Cloud MaaS. Today, Hunyuan is also officially open to industry customers. Customers can experience it directly on Tencent Cloud through front-end interaction and API call; they can also use Hunyuan as the base model and adjust it directly in public cloud.

At the same time, Tencent Cloud TI platform is also fully connected to mainstream open source models such as Llama 2, Falcon, Dolly, etc., supporting direct deployment and invocation of series models, and the application process is simple.

Customers can complete model deployment with one click on TI platform according to their needs, test the effect of these open source models, and realize model selection.

Customers can also fine-tune these open-source large models with one click of the TI platform. Tencent Cloud TI platform provides the ability of multi-machine multi-card and breakpoint continuous training, which can ensure the stable and efficient operation of training tasks; training monitoring ability, which can monitor training indicators and resource utilization rate in real time; task management ability, which can automatically manage the training tasks of multiple iterations of the model, record the super-parameter snapshot of each training task, and quickly restore the training site.

By upgrading our platform capabilities, we can provide our customers with a more complete tool chain.

The upgraded TI platform tool chain has high-performance fine tuning and deployment capabilities for large models, covering the entire life cycle of large model development and application. We have greatly upgraded the computing power and storage at the core of fine tuning of large models. For example, in terms of computing power, Tencent Cloud HCC high-performance computing node and Xingmai high-performance computing network are connected, providing 3.2Tbps network bandwidth; Turbo-CFS distributed storage is connected, with the maximum bandwidth reaching 100GiB/ s, greatly improving the multi-machine multi-card training performance across machines.

With the brand-new upgraded TI platform tool chain, customers can quickly try a variety of large models by adding their own unique scene data and fine tuning training, and fine-tune their own model solutions in a short time according to their own business scenario requirements.

In terms of acceleration, the brand-new upgraded Taiji Angel large model acceleration framework improves performance by more than 30% compared with common solutions in the industry through asynchronous scheduling optimization, memory optimization, calculation optimization, etc.

Provide customers with rich, plug-and-play products through intelligent application upgrades.

Tencent has long-term investment and accumulation in the field of AI technology. In the past, we have a lot of AI capabilities, which are the first to be applied in internal products, which is the endogenous kinetic energy of Tencent AI.

Today, our large model capabilities are also reshaping Tencent Cloud's extensive PaaS and SaaS products. These exploration practices also further verify the availability and ease of use of our MaaS capabilities.

Focusing on customer service scenarios, Tencent Enterprise released a brand-new generation of intelligent customer service in June this year, introducing large model capabilities to improve dialogue interaction, manual assistance and knowledge construction. In the past two months, we have further upgraded our management configuration capabilities, improved knowledge building efficiency, and optimized the C-end user experience. At the same time, we actively carry out industrial practice with all walks of life, and set benchmark cases in the fields of culture and tourism, government affairs, real estate, etc.

Based on the analysis model, we upgraded and launched Tencent Enterprise Point Analysis AI Assistant, covering four scenarios: conversational analysis, auxiliary data configuration, intelligent conclusion extraction and one-click report generation. In September this year, the AI assistant of enterprise point analysis officially opened the internal test application to large customers. In the future, everyone can have professional data analysis capabilities to make data run faster and provide decision support for the front line of business.

Before that, we launched a number of Homo sapiens factories. Today, with the support of AI large model technology, we have released a brand-new version of "universal mouth shape" version of small sample number Homo sapiens. After uploading materials, the production cycle of small sample number Homo sapiens is shortened from 24 hours to only 1 hour, greatly improving the production efficiency of Homo sapiens. With the help of large model technology, our digital homo sapiens can achieve more efficient, natural and intelligent effects.

In the area of content creation, we have partnered with Reading Group in the inference deployment and acceleration of graphic and graphic images.

In the scenario of Wensheng graph and graph generation graph, it provides high QPS and low time consumption scheme for C-end users. Moreover, the deployment cost of reasoning model itself is high, and there are bottlenecks in reasoning throughput and delay. Based on the Angel inference acceleration engine carried by Tencent Cloud TI platform, the inference ability is increased by 30% on average, and the cost is reduced to one quarter of the original in the field of literary and literary works. In addition, TI platform provides dynamic batch and automatic service scheduling mechanism, which greatly improves service throughput.

As a digital assistant, we will continue to polish MaaS capabilities and join hands with customers and partners to embrace this "efficiency revolution" in the AI big model era.

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.