Artificial intelligence large models have been launched one after another to accelerate the development of the empowerment industry

Recently, many domestic science and technology enterprises or scientific research institutions have officially opened self-developed artificial intelligence large models to the public, marking that domestic large models have begun to move from small-scale testing to large-scale application. How will the big model empower individuals and industries? What impact will this round of open services have on industry competition? What will be the future development trend of domestic large models?

The competition for artificial intelligence big models has entered a new stage

In early September, iFLYTEK announced that the iFLYTEK Spark Cognitive Model was open to the public, and so far, the first batch of more than ten AI models including iFLYTEK Spark, Baidu Wenxin Yiyan, SenseChat, etc. are all open to the public. In addition, the large models of some companies, including Tencent, have also been filed and will be opened to the public at the opportunity.

Artificial intelligence large models refer to the "big parameter" models trained using large-scale data and powerful computing power, which usually have a high degree of generalization and generalization ability, and can be applied to natural language processing, image recognition, speech recognition and other fields.

Previously, many enterprises and institutions in the mainland have released large-model products and conducted small-scale testing, and the large-scale model of the open service also means that they have passed the filing of relevant departments. On August 15, Mainland China officially implemented the Interim Measures for the Management of Generative Artificial Intelligence Services.

Yu Huan, director of Baidu's Science and Technology and Society Research Center, introduced that Wen Xin Yiyan has grown rapidly since its launch in March this year, mastering more than 200 creative genres, content richness is 1.6 times that of the early stage of release, the length of the thinking chain is 2.1 times that of the early stage, and the coverage of knowledge points has reached 8.3 times that of the initial stage. In terms of efficiency, the training speed of the Wenxin large model has reached 3 times and the inference speed has reached more than 30 times.

Liu Qingfeng, Chairman of iFLYTEK, introduced that since the first release of the iFLYTEK Spark cognitive big model on May 6, it has undergone two major iterative upgrades, and has continued to evolve in seven capabilities: text generation, language understanding, knowledge question and answer, logical reasoning, mathematical ability, code ability, and multimodal capability. At the latest iFLYTEK Spark V2.0 upgrade conference released on August 15, code capabilities and multimodal capabilities were upgraded.

According to the relevant person in charge of SenseTime, the characteristics and capabilities of the company's large model "Discussion SenseChat" include efficient processing of a large number of texts, powerful natural language dialogue, logical deduction capabilities, extensive knowledge reserves and information updates, support for Chinese Simplified, Chinese Traditional, English and various commonly used programming languages.

"With the successful filing, the various parties involved are eager to showcase the R&D progress and results since the beginning of the year. At the same time, the faster release of related services, you can get more user and industry feedback, so as to further accelerate the iteration and application of their respective large models, and gain as much competitive advantage as possible. Yu Hao, vice president of Legend Holdings Co., Ltd. and president of the Prospective Technology Research Institute, said.

The big model will empower individuals and industries

The domestic large model that has been launched has opened up a number of tools and is expected to become a "super assistant" for ordinary people to improve efficiency.

Liu Qingfeng said that since the launch of the test, users of the iFLYTEK Xinghuo APP have developed more than 20,000 personalized artificial intelligence assistants, each ordinary person can call a number of artificial intelligence assistants that have been put on the shelf, and can also develop their own exclusive assistants through a few simple steps to meet the needs of workplace, marketing, travel, life, official documents, customer service and other scenarios, and help solve various problems in work and life. According to the data since the launch test of iFLYTEK Xinghuo cognitive big model, the current scenarios that users just need to use are concentrated in scenarios such as knowledge Q&A, content generation, creative planning, life encyclopedia and programming assistance.

"As the automation of expert capabilities, large models can bring great convenience to ordinary people in all aspects, such as helping ordinary people learn knowledge, modify articles, generate solutions, and so on. Therefore, the centralized release of large models can allow ordinary people to quickly contact, use and understand large models, and the good experience of large models will enhance ordinary people's trust in large models. Sun Le, researcher at the Institute of Software of the Chinese Academy of Sciences and secretary general of the China Society for Chinese Information, said.

The relevant person in charge of Ailiai Think Tank believes that the centralized launch is expected to accelerate the commercialization of domestic large models and downstream applications. At present, the industry has the conditions for large-scale commercialization, from the supply side, the maturity of the existing technology can assist users in content production on a large scale, and the downstream application supply has blossomed at multiple points, involving painting, writing and video and other fields; From the demand side, driven by the fragmentation and lightweight of entertainment, the demand for user content consumption has increased exponentially, and traditional production methods have led to a sharp increase in cost pressure in the industry, and it is urgent to reduce costs and increase efficiency through artificial intelligence production content.

At present, a number of companies are exploring the in-depth application of large models in various industries. In early September, Tencent's hybrid model was officially unveiled and announced the opening of API access services through Tencent Cloud. According to the relevant person in charge of Tencent, Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Fintech, Tencent Meeting, Tencent Docs, WeChat Souyi Search, QQ Browser and other Tencent internal businesses and products have been connected to the hybrid model and achieved preliminary results, and more services and applications are gradually being accessed.

"With large model generation technology as the core, artificial intelligence is becoming the key driving force for the next round of digital development, bringing new ideas to solve industrial pain points." Tang Daosheng, senior executive vice president of Tencent Group and CEO of the Cloud and Smart Industry Business Group, said.

Shenzhen Yuntian Lifei Technology Co., Ltd. first announced its large-scale model research and development in July this year. The relevant person in charge of the company introduced that based on the "Book of Heaven" large model, Yuntian Lifei and Shenzhen Longgang Political and Numerical Bureau jointly explored the application of the large model in the government consulting service system, and planned to gradually expand to other government agencies and public service fields in the future, such as legal, finance, education, medical care, transportation, etc.

For large model R&D companies, whoever can bring the best industry application the fastest will be able to obtain the most user feedback and data, further increasing the competitive advantage of their large model.

"When Wenxin opens up its services to hundreds of millions of Internet users on a large scale, it can get more human feedback in the real world, which will help further improve the basic model and iterate the product more quickly, creating a better user experience." Robin Li, founder, chairman and CEO of Baidu, said.

The "infancy" of the large model is still waiting to be "fed back"

Industry insiders believe that although the current domestic large-scale model research and development is progressing rapidly, bottlenecks such as computing power and talents still need to be further broken, and the business model needs to be further clarified.

According to Lin Dahua, a professor at the Shanghai Artificial Intelligence Laboratory, the biggest difficulty and challenge in the training of a language model with a scale of hundreds of billions of parameters is not the lack of data, but the high cost of trial and error. A large model with 100 billion parameters needs 3 months to train on a kcal cluster. Each update requires a lot of computing power, and the training cycle is very long.

Yu Hao believes that the large model as a "base technology" is still in the "infancy stage", and its growth requires extremely expensive and energy-consuming "computing power warming house", surrounded by a group of "expert nannies" specializing in artificial intelligence algorithm optimization, and a large number of high-quality "data ingredients". If you accidentally ingest "contaminated ingredients" or "do not strictly train" large models, it is very likely to cultivate "bear children" who are "nonsense", causing incalculable impact on large model enterprises.

"Considering future sustainable development, on the one hand, it is to build a large model infrastructure engine and build a solid foundation, on the other hand, cultivate a small model ecosystem, evolve a technology stack that adapts to the long-tail needs of diversified small scenarios of artificial intelligence, solve practical problems with small models, create value monetization, and then feed back large models." Yu Hao said.

When talking about the development trend of domestic large models, Sun Le believes that the first is that with the acquisition of a large number of high-value data and user feedback, the level of domestic large models will improve rapidly. Secondly, in terms of application, large models will be widely implemented, and typical application paradigms will gradually form in suitable scenarios. Finally, large models will still face some long-term scientific problems, such as reliable, controllable and credible problems, and will also discover new scientific problems in the large-scale implementation process, which need to be solved by governments, businesses and academia.

Shen Yang, a professor at Tsinghua University's School of Journalism and Communication, believes that the centralized launch of domestic large models will give rise to a series of development trends, including accelerated technology iteration, deepening industry applications, diversified business models, prominent data and ethical issues, and internationalization. These trends will not only affect the development of the model itself, but will also have a profound impact on socio-economic, legal norms and the international competitive landscape. Therefore, all parties need to pay close attention and make corresponding strategic planning and countermeasures.