Tencent officially joins the big model melee! At the 2023 Tencent Global Digital Ecosystem Conference on September 7, the Tencent Hybrid Model was officially unveiled and announced that it would be opened to the public through Tencent Cloud. So far, the large model products of Internet manufacturers have all been unveiled.
Talking about Internet companies catching up with me in the progress of large model release, Tang Daosheng, senior executive vice president of Tencent Group and CEO of Cloud and Smart Industry Business Group, said in an interview with Nandu reporters after the meeting that there is no "catch-up" in this track, far from the stage of clear market share, "ToB technology and market penetration may be in ten years, the big model is a marathon, and now it may only run to one kilometer." ”
With more than 50 service accesses, Tencent fully embraces the big model
According to reports, Tencent Mixed Element Model is a practical-grade general large model developed by Tencent, which has been connected to more than 50 Tencent businesses to test and achieve preliminary results, including Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Financial Technology, Tencent Meeting, Tencent Documents, WeChat Sou, QQ Browser and other businesses and products.
"Tencent's mixed-element model is trained from zero from the first token." Jiang Jie, Vice President of Tencent Group, introduced that Tencent has mastered the full-link self-research technology from model algorithms to machine learning frameworks to AI infrastructure. At present, the parameter scale of Tencent Hybrid exceeds 100 billion, and the pre-training corpus exceeds 2 trillion tokens, which has strong Chinese creation capabilities, logical reasoning capabilities in complex contexts, and reliable task execution capabilities.
Jiang Jie also showed the actual application of Tencent Meeting, Tencent Docs, Tencent Advertising and other businesses after accessing the Tencent Hybrid Model.
For example, Tencent Meeting's AI assistant based on Tencent Hybrid only needs simple natural language instructions to complete complex tasks such as conference information extraction and content analysis, and can also generate intelligent summary minutes after the meeting. In terms of document processing, Tencent's hybrid model supports dozens of text creation scenarios, which have been applied in the intelligent assistant function launched by Tencent Docs. At the same time, Tencent Hybrid can also generate standard format text with one click, be proficient in hundreds of Excel formulas, support natural language generation functions, and generate charts based on table content. In advertising business scenarios, Tencent's Hybrid Model supports intelligent creative creation, which can adapt to industry and regional characteristics, meet the needs of thousands of people, and realize the natural integration of text, pictures, and videos.
Jiang Jie, Vice President of Tencent Group
"Our goal in developing large models is not to get high scores on evaluations, but to apply the technology to real-world scenarios. Tencent will fully embrace the big model. Jiang Jie said. In Jiang Jie's view, only full-link self-research can achieve better than the industry level and achieve technological innovation breakthroughs, and full-link self-research will also make large models more stable when running on massive and high-concurrency applications, which is the only way for large models and generative AI to achieve large-scale applications, and Tencent, which has more than one billion users, needs to be repeatedly polished to ensure the user experience of massive users. "From the lowest level of servers, network cards, to the entire high-speed networking, including platforms, models, and algorithms are all based on Tencent's self-research, which will definitely allow us to gradually accelerate in subsequent iterations." On the other hand, Tencent is a massive and high-concurrency business, and the open source architecture does not adapt to Tencent's volume, so we must take the road of independent research and development to cope with the impact of massive and high-concurrent services. Jiang Jie said.
At the beginning of the ten-year run, solving the actual problems of the industry is the key
On September 1, the CAC released the second batch of deep synthesis service algorithm filing information, and a number of generative synthetic large models, including Tencent's mixed-element large model, passed the filing. This means that the application of domestic large models has officially begun.
"We talk to customers and see that most of them are multiple models being plugged in and tested. People have high expectations for big models, a lot of ideas, but they are not so clear about what problems can be solved. Tang Daosheng told Nandu reporters that Tencent hopes to use big models to really help them solve problems.
It can be seen that the current main battlefield of Tencent's mixed element model, which appeared together with Tencent business scenarios such as meetings, documents, and advertisements, is still the B-side scenario.
"Now everyone likes to pay attention to and use the general large model from the perspective of TOC. For example, many people spend a lot of energy to test and let some general models talk nonsense, rather than solving the actual problems and pain points of the industry. Tang Daosheng said that the most pragmatic application of large models is still based on industry pain points, using industry large models to solve problems, "may just start using the version, can only solve 80% of the problem, but because you have a very clear use scenario, user feedback can form a feedback, let you continue to polish your industry large model, so that the accuracy of problem solving, step by step." ”
In Tang Daosheng's view, the industrial application of large models is not necessarily a fantastical scene, but it is simple and useful, and can answer customers' questions faster. In the face of the competitive landscape of the 100-model war, Tencent's path is to structure and polish products based on its own business scenarios, and open the products to industry customers after achieving satisfactory results.
For the scenario where the large model is prioritized, Tang Daosheng said to the Nandu reporter that from the perspective of the Wen Sheng literary scene, the most widely used large language model is the communication and exchange scene, which can reduce costs and increase efficiency or improve user experience, such as customer service, after-sales, etc., and the scene with more dealings with users, the big language model can play a value in more places. The high-frequency use scenarios of Wenshengtu's capabilities are advertising and marketing, "We will screen some industries to see which industries are more willing to use AIGC capabilities such as Wensheng Tu to improve their delivery efficiency." ”
In addition to the general big language model, Tencent launched the "Tencent Cloud MaaS Service" in June, where enterprises can choose the hybrid and open source large models, or industry large models covering more than 20 industries such as finance, healthcare, cultural tourism, and energy as the base, and use Tencent Cloud Intelligence's TI platform to import unique professional documents and enterprise data for further training and fine-tuning, and quickly generate more targeted exclusive large models.
Written by: Nandu reporter Ma Ningning