On July 6, the 2023 World Artificial Intelligence Conference was held in Shanghai. Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industry Business Group, gave a keynote speech at the plenary session of the conference-Industrial Development Forum. He said that the general large model has strong capabilities, but it cannot solve the specific problems of many enterprises. Enterprise large-scale model applications need to comprehensively consider factors such as industry expertise, data security, continuous iteration, and comprehensive costs. Based on the large industry model, building your own exclusive model may be a better option for enterprises.
"The general large model can solve 70%-80% of the problems in 100 scenarios , but it may not be able to 100% meet the needs of a certain scene of the enterprise." Tang Daosheng said that the general large model is generally based on a wide range of public documents and networks Information is used for training, and a lot of professional knowledge and industry data are not accumulated enough, resulting in insufficient industry pertinence and accuracy of answers. However, users have high requirements for the professional services provided by enterprises and low fault tolerance. Once enterprises provide wrong information to the public, it may cause serious consequences.
He said that if an enterprise conducts fine-tuning based on a large industry model and its own data, it can build a dedicated model and create highly available intelligent services. Moreover, the model parameters are less than the general large model, the cost of training and reasoning is lower, and the model optimization is easier.
At the same time, large-scale industry models and model development tools can also prevent leakage of enterprise sensitive data caused by model training and use through privatized deployment, authority control, and data encryption.
In addition, the landing process of the large model also needs to go through a series of links such as algorithm construction and model deployment, and each link cannot be "lost the chain". The model needs to be continuously iteratively optimized in the future, which requires the use of systematic and engineering tools.
In response to these problems, Tencent Cloud has also recently released a panorama of Tencent Cloud MaaS services. Based on the Tencent Cloud TI platform, it has built a large-scale industry model selection store, providing 10 major industries such as finance, cultural tourism, government affairs, medical care, media, and education, and more than 50 services. One solution ; launch the industry large model fine-tuning solution to help model developers and algorithm engineers, one-stop solve tasks such as model invocation, data and label management, model fine-tuning, evaluation testing and deployment, and reduce the pressure of creating large models . On the basis of these models and tool platforms, enterprises can quickly generate their own "exclusive models" only by adding their own scene data.
"With the development of the big language model, the industry and society will also move from digitization and networking to intelligence. In this process, Tencent always believes that the fundamental goal of artificial intelligence development is to land in the industry and serve people. Enterprises that truly solve user needs and are closer to scenarios and data will have the future of big models. Tencent will join hands with all partners to use high-quality models and surging computing power to make the "golden data" of each enterprise play a high role. Help the innovation and development of the industry." Tang Daosheng said.
The following is the full text of the speech:
Hello everyone! It is a great pleasure to participate in the World Artificial Intelligence Conference. Today I am here to throw a brick to attract jade, and I invite all experts to correct me.
It has been 5 years since Tencent participated in the World Artificial Intelligence Conference. During these five years, we have participated in and witnessed the construction of the artificial intelligence "Shanghai Heights". Tens of thousands of servers have been put into production in Tencent's data center in Qingpu and the Yangtze River Delta Artificial Intelligence Supercomputing Center in Songjiang. One of our AI laboratories, Youtu Lab, is located in Shanghai and has obtained more than 1,600 global patents. The Keen Security Lab, also established in Shanghai, has reached the international leading level in areas such as AI security, and is known as the "dream team" of China's cyber security.
In the past 5 years, Tencent's artificial intelligence technology and products have also taken root in various industries: our digital smart people have "joined" more than 130 industries, serving as "digital smart employees" such as financial customer service and virtual anchors, providing users with Personalized service. Our industrial AI quality inspection has been used in many manufacturing lines such as 3C parts and lithium batteries, helping enterprises reduce costs and increase efficiency while improving the yield rate. With the help of AI, game engine and other game technologies and China Southern Airlines virtual image display technology, we generate a virtual flight environment to provide civil aviation pilots with more efficient and safe flight training.
In the past few months, the development of large language models and generative AI has achieved powerful language understanding and reasoning capabilities, and can generate complete paragraphs, exquisite pictures, videos, and even codes according to prompts, making AI a more powerful individual assistant.
Many physical enterprises are both excited and anxious, and can't wait to embrace large-scale model technology, promote the intelligent upgrade of design, sales, service and other links, and improve production, operation, and management efficiency.
In fact, although the general model is very powerful, it may not be able to solve the specific problems of many enterprises. Are large models really reliable and usable in industrial scenarios? How can we protect corporate data property rights and privacy? How to reduce the cost of using large models? These are practical issues that enterprises need to consider.
It may be a better option for enterprises to build their own exclusive models based on large industry models. At the same time, it is also necessary to use efficient professional tools to continuously optimize and iterate the model to meet the continuously changing needs of enterprises and markets.
First of all, with the help of large industry models, we can provide users with precise services more efficiently.
General-purpose large models are generally trained based on extensive public literature and network information. The information on the Internet may contain errors, rumors, and biases. Many professional knowledge and industry data are insufficiently accumulated, resulting in insufficient industry pertinence and accuracy of answers. , the output information is also relatively broad. Although the overall level of the general-purpose large-scale model is constantly improving, the strategy is a bit like boiling the ocean (Boil the Ocean), and it is not focused. The general-purpose large model can solve 70%-80% of the problems in 100 scenarios , but it may not be able to 100% meet the needs of a certain scenario of the enterprise.
However, users have high requirements for professional services provided by enterprises and low fault tolerance. Once enterprises provide wrong information to the public, it may cause huge legal liabilities or public relations crisis. Therefore, I think that each enterprise can build a unique "exclusive model" based on the industry's large model trained by professional knowledge and data, and fine-tune the enterprise's own data, so as to create a high-availability model more efficiently. smart service.
At the same time, the enterprise-specific model based on the industry's large model has fewer model parameters than the general-purpose large model, lower training and inference costs, and easier model optimization.
Secondly, with the help of a dedicated model, enterprise data is protected and data security is ensured.
Data is the raw material of a large model, and the model must eventually be implemented in a real scene to achieve an ideal service effect, often requiring the use of the enterprise's own data. If the data is not properly protected during the process, it may cause the leakage of enterprise core data and sensitive data.
Large industry models and model development tools can be deployed through privatization to make model training more secure, and can also prevent leakage of corporate sensitive data when employees access the model.
If the model service is user-oriented, user feedback data can also be used to optimize the exclusive model and continuously improve the service experience.
Thirdly, with the help of efficient platform development tools, the continuous optimization of the model can be realized quickly and at low cost.
The implementation of models in the industry is a complex and systematic project, which must go through a series of links such as data processing, algorithm construction, and model deployment, and each link must not be "off the chain".
At the same time, the application of the enterprise model does not end after one-time deployment. It also needs to be constantly adjusted according to new data during use, so that the model can keep up with the changing market and user needs. In the process, it is necessary to manage a large amount of data and labels, and constantly test and iterate the model. This requires the use of systematic and engineering tools to ensure the continuous operation of the model.
Based on the consideration of the practical problems and needs of these enterprises, last month, we also officially announced the panorama of Tencent Cloud MaaS services.
Based on the Tencent Cloud TI platform, the selected large-scale industry model store provides 10 major industries such as finance, cultural tourism, government affairs, medical care, media, and education, and more than 50 solutions. Based on these capability models, partners only need to add their own unique scene data to quickly generate their own "exclusive models". We can also help enterprise users protect their own data and feel more at ease when using models through privatized deployment of models, authority control, and data encryption.
We also launched a fine-tuning solution for large-scale industry models based on the Tencent Cloud TI platform. Help model developers and algorithm engineers to solve tasks such as model calling, data and label management, model fine-tuning, evaluation testing and deployment in one stop, and reduce the pressure of creating large models.
For example, we and the top domestic online travel companies have created robot customer service based on the "cultural tourism model". When a user inquires about a vacation itinerary, if it is a customer service robot based on a general large model, it can only give some simple introductions to scenic spots. But when we fine-tune the model by adding enterprise data based on the large industry model, the customer service robot’s answers become more accurate, available, and detailed. It can plan transportation, scenic spots, and hotel arrangements, and even directly provide booking links and coupons. and other information. It not only realizes considerate service, but also has a stronger sales transformation ability. This is what businesses need.
We also integrate the industry's large-scale model capabilities into Tencent's own enterprise-level applications to help customers improve work efficiency through smarter services.
For example, the new generation of Tencent Qidian's intelligent customer service is based on industry models, and is trained and fine-tuned in combination with customer business needs. It can provide more accurate and detailed answers, and the user experience is more humanized. At the same time, with the help of the Qidian analysis platform, sales staff can achieve accurate business analysis by asking questions in natural language, without spending a lot of time learning complex software and making Kanban.
Our Sapiens have also improved the reproduction speed of digital images because of the integration of AI generation algorithms. The production of 2D digital intelligence only needs to record a 3-minute live broadcast video. With the help of the multi-modal processing capability of the platform, real-time modeling and high-definition portraits can be produced. Within 24 hours, a "digital intelligence" similar to real people People", the cost is greatly reduced.
Ladies and gentlemen! With the development of large language models, industry and society are moving from digitization and networking to intelligence. In this process, we always believe that the fundamental goal of artificial intelligence development is to land in the industry and serve people. Enterprises that can truly solve user needs and are closer to scenarios and data will have the future of big models.
Tencent will join hands with all parties to use high-quality models and surging computing power to make the "golden data" of each enterprise play a high role and help the innovation and development of the industry.
Tencent's artificial intelligence industry practice will also be shared with you in detail in the Tencent sub-forum and game AI sub-forum. In addition, everyone is very welcome to come to the Tencent booth to experience technologies such as recording a 3-minute video to make your own Sapiens.
thank you all!