(The materials are all online materials, if there is any infringement, please contact to delete immediately)
Lead:
At the Tencent Global Digital Ecosystem Conference in 2023, Tencent officially unveiled the mystery of the mixed element language model, announced its official debut, and opened it to the public through Tencent Cloud. This general-purpose big language model not only has powerful Chinese creation capabilities, but also has logical reasoning capabilities and reliable task execution capabilities in complex contexts. In this article, we will explore the story behind Tencent's hybrid model and its potential and impact in the market.
In the past half a year when domestic cloud vendors have set off a "hundred model war", Tencent has kept a low profile and is not in a hurry to show its general large models.
This move has led people to speculate: why is Tencent not in a hurry to release a general large model? What exactly is Tencent doing in the days when the market is out of sight?
The reason why Tencent is not in a hurry to show the GM model is because they have always adhered to the style of playing steadily and speaking with strength. Pony Ma, chairman and CEO of Tencent, said they see GM's large model as a centuries-old opportunity, similar to the industrial revolution that invented electricity, and is in no hurry to bring semi-finished products to market early. He stressed that for the industrial revolution, taking out the light bulb a month earlier was less important over a long period of time.
Tencent pays more attention to the practicality and durability of large models, rather than the short-term heat.
In the days when the market is out of sight, Tencent has been working to improve the underlying capabilities of large models. Since 2021, they have launched NLP sparse large models with hundreds of billions and trillion parameters, breaking the records of the three major lists of CLUE and achieving a new breakthrough in Chinese understanding ability. At the end of last year, after ChatGPT set off a wave of large models, Tencent's large model route was further firm, and it used its advantages of diversified and rich products, data, and scenarios to iterate multiple versions.
Unlike other companies, Tencent chose to test large-language models with its own products, rather than through chat scenarios. Jiang Jie, vice president of Tencent Group, said that Tencent has more than 20 years of development history and massive user data, and using these to test large models will have better results. Therefore, Tencent's product matrix has become the "nourishment" for the corpus training of the mixed element large model, and it is also the "grinding stone" for scene applications.
At present, the hybrid model has been connected to more than 50 Tencent businesses, including Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Financial Technology, etc. This includes not only C-end applications, but also B-side scenarios, making Tencent's large-model applications and tests comparable to or even surpass other large models.
Tencent officially announced that the hybrid model will become the foundation of Tencent Cloud MaaS services, and customers can call the hybrid element through APIs or use it as a base model to build exclusive applications for different industry scenarios.
Tencent's full-link self-research of the hybrid model is one of its highlights. This means that Tencent has fully mastered its independent research and development technology from model algorithms to machine learning frameworks to AI infrastructure, starting from the first token. This includes large-scale, high-quality, diversified corpus, innovative large-model algorithms and training methods, self-developed Angel machine learning framework, and powerful computing infrastructure.
This full-link self-developed technology has brought obvious results. Tencent's mixed-element large model uses the "truth detection" algorithm to correct the facts in the pre-training stage, which reduces the "illusion" by 30% to 50% compared with the mainstream open-source large model, especially in the field of lower error tolerance or more complex tasks. In addition, Tencent's mixed-element large model also enables the model to identify trap problems through reinforcement learning methods, improves the effect and performance of processing ultra-long texts through positional coding optimization, and proposes a new strategy of thinking chain, so that the large model can be reasoned and made in combination with practical application scenarios.
Tencent has also developed its own machine learning framework Angel, which has significantly improved the speed of training and inference.
Fulllink Self-Research has obtained the highest score in the standard compliance test of the China Academy of Information and Communications Technology, and has also performed well in mainstream evaluation sets, especially in the fields of science, college entrance examination questions and mathematics in the Chinese.
Tencent's hybrid model not only performs well in evaluation, but also
Importantly, they successfully applied the technology to real-world scenarios. Tencent's Hybrid Model has already achieved initial results in a number of businesses, including Tencent Meeting, Tencent Docs, Tencent Advertising, etc.
These applications demonstrate the great potential of mixed-element large models to improve the efficiency and quality of user interaction services.
Taking Tencent Meeting as an example, the hybrid model has created an AI assistant for it, which can complete complex tasks through simple natural language instructions, including conference information extraction and content analysis. Users ask questions in meetings, and the hybrid model can respond quickly and generate intelligent summary minutes. This function has been highly recognized by users in terms of instruction understanding, in-meeting Q&A, and conference summary.
In the advertising field, Tencent's Hybrid Model supports intelligent creative creation, which can adapt to the characteristics of different industries and regions, meet personalized needs, and realize the natural integration of text, images and videos.
In addition, intelligent shopping guide can improve the service quality and efficiency of merchants in scenarios such as enterprise WeChat.
Tencent President Martin Lau said that Tencent's mixed-element model has a wide range of application prospects in generative AI technology. This is not limited to the Q&A experience, but also includes improving the efficiency and quality of user interaction services, promoting ad targeting, data targeting capabilities, and improving content productivity. Therefore, Tencent will continue to develop generative AI technology and apply it to more fields to realize the effect of multipliers.
Tencent's hybrid model still has more room for imagination.
It enables continuous training, lifelong learning, and constant updating of the latest knowledge to improve performance and accuracy. At the same time, the hybrid model will also release more market potential for Tencent in model-as-a-service solutions.
The release of Tencent's hybrid model is a start, and Tencent promises to continue to evolve its capabilities and continue to surprise users. The hybrid model is no longer a semi-finished product, but Tencent has always adhered to high standards and pursued more perfect products. Tencent's full-link self-research and continuous technological innovation have enabled it to maintain its leading position in the field of large models, laying a solid foundation for future development.
In short, the official debut of Tencent's hybrid model marks China's great progress in the field of large language models. This universal big language model not only has powerful technical capabilities, but also has been successfully applied to multiple practical scenarios, providing users with more efficient and intelligent services. With the continuous evolution and application expansion of the hybrid model, it will become an important tool for Tencent in the field of artificial intelligence, and will also bring innovation and development opportunities to more industries and fields. Tencent's hybrid model can be expected in the future.
This article is the first publication of today's headlines, and all other accounts are porters.