China Business Daily (Reporter Zhao Yiru)"Every family has a big model" has become the most important way to show muscles in the scientific and technological circle in the first half of this year. In just a few months, nearly 100 AI models have emerged in the market. General model, industry model, vertical model, companies have chosen the best track to start, and the competition for big models has become fierce.
The 2023 World Artificial Intelligence Conference showcases the achievements of AI+ smart medical services. (Image courtesy of CNSPHOTO)
The general big model is fiercely competitive.
The first shot of the domestic "Hundred Models War" was started by Baidu.
In March this year, ERNIE Bot, Baidu’s generative AI product, officially invited for testing. Li Yanhong, founder, chairman and CEO of Baidu, said: "ERNIE Bot and generative AI represent a new technology paradigm and a great opportunity that every enterprise can’t miss. The AI market is about to usher in explosive demand growth, and its commercial value release will be unprecedented and exponential. "
As the first published AI general model in China, ERNIE Bot has made a full stack layout in the four-tier architecture of "chip-framework-model-application". This has also become one of Baidu’s competitive advantages.
In April, Alibaba Cloud’s "Tongyi Qianwen" model also began to invite tests. Zhang Yong, CEO of Alibaba Cloud Intelligent Group, said that all Alibaba’s products will be connected to the "Tongyi Qianwen" model in the future and be completely transformed.
"For the AI era, all products deserve to be redone with a big model, and all companies are on the same starting line." Zhang Yong said.
Regarding the positioning of "Tongyi Qianwen", Zhou Jingren, CTO of Alibaba Cloud Intelligent Group, said that Tongyi Qianwen is a general model and will not be optimized or customized according to a specific business problem. The ability to open up the universal question means that enterprises do not need to train large models from scratch, but can form their own exclusive models on the basis of universal question, combined with application scenarios, knowledge systems and special needs of the industry.
In May this year, Iflytek released the "iFLYTEK Spark Cognitive Model". Liu Qingfeng, chairman of Iflytek, said that Spark Cognitive Model has surpassed ChatGPT in text generation, knowledge quiz and mathematical ability, and strives to surpass ChatGPT as a whole by the end of October.
In June this year, 360 Group released the self-developed cognitive general model "360 Zhinao 4.0", and announced that 360 Zhinao has been connected to 360′ s product "Family Bucket".
"China has entered the period of the Hundred Models War. In addition to its basic ability, it is more important to compete in application scenarios." Zhou Hongyi, founder of 360 Group (hereinafter referred to as 360), said. At the C end, 360 plans to embed the AI big model capability into the core products, so that it can provide services to users like AI assistants. On the B side, the 360 plan is first applied in large-scale industry models and large enterprises.
In addition, Shangtang Technology released the "SenseNova" large model system, Kunlun Wanwei released the 100 billion-level large language model "Tiangong", and Beijing Zhiyuan Artificial Intelligence Research Institute released the Enlightenment 3.0 large model series …
"The domestic AI big model is in full bloom. First of all, because the big model is the basis for the development of the new generation of information technology revolution AI era, large Internet companies will follow up, and innovative enterprises will try their best to do it if they have the ability. Secondly, because foreign countries have successively opened up some platforms, they can save the training cost of large models, and some fine-tuning and reasoning costs are affordable for many teams, so domestic large models can flourish. " Shenyang, director of the Metacosmic Culture Laboratory of Tsinghua University School of Journalism and Communication, told the China Business Daily reporter.
"The phenomenon that the AI big model is in full bloom shows a trend, that is, generative AI may have a profound impact on the industrial structure and society in the future. Now it has just begun, and the penetration of AI into the industry is deepening. " Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, said.
The big industry model takes a different approach.
In the case of fierce competition in the general big model, some enterprises saw the advantages of the industry big model and chose to release the industry big model first.
In June this year, Tencent announced the research and development progress of Tencent Cloud’s big model industry, and jointly launched the cooperation of building a big model industry with 22 customers, and jointly launched the "Tencent Cloud Industry Big Model Ecological Plan" with 17 ecological partners.
"We believe that customers need more industry-specific industry models." Tang Daosheng, senior executive vice president of Tencent Group and CEO of Tencent Cloud and Smart Industry Group, said.
Tang Daosheng believes that at present, the universal big language model has certain limitations in dealing with the landing of industrial scenes. First of all, the training data of the general large model mainly comes from public data sets or network data, and the knowledge about the professional field of specific industries is limited. In addition, the training of general large language model requires a lot of computing resources and a long training cycle, which may be expensive and time-consuming for enterprises. At the same time, safety and compliance are also necessary considerations. Therefore, it is very important to choose to cooperate with cloud vendors with one-stop industry model service capabilities to create enterprise-specific industry model, which can reduce the cost and time investment of enterprises while being easy to use and use.
Coincidentally, Huawei also chose to release a big model of the industry.
In July this year, Zhang Pingan, managing director of Huawei and CEO of Huawei Cloud, released Pangu Big Model 3.0 at Huawei Developers Conference 2023. At present, Pangu model has played a role in many industries, such as finance, manufacturing, pharmaceutical research and development, coal mine, railway and so on.
"Pangu was born for the industry, so it is necessary to think for the industry and better serve customers in thousands of industries." Zhang Pingan said.
"Huawei has a very strong business foundation in many industries. As a big model of the industry, it can be quickly applied to various industries and directly exert economic benefits and social values. The general big model is oriented to C-end users, and its generated results are highly uncertain, which will also bring greater regulatory costs. " Shenyang said.
"If the GM model is to be really well done and quickly put into application in the C-end market, it will take time to polish its software, and it will also face the cost and technical difficulties of large-scale investment in content management. Therefore, from the perspective of research and development progress, the industry model can be used first. Huawei and Tencent have a good foundation in this respect and can quickly put it into the industry to use it. " Shenyang believes.
Vertical large model highlights advantages
Some companies choose to continue to narrow the scope of application of the AI big model and focus on the vertical field.
In July this year, Ctrip released the first vertical model of the tourism industry, "Ctrip asks". Liang Jianzhang, Chairman of the Board of Directors of Ctrip Group, expressed the hope that users would get "reliable content and reliable recommendation" from the tourism industry, including "Ctrip asks".
Ctrip’s motivation to launch the vertical big model is precisely to see the "unreliability" of the general big model.
"The opportunity of the vertical big model is that AIGC has encountered a very big challenge-the answer is not so reliable." Liang Jianzhang said. He believes that even if the planning saves half an hour, the recommended hotel or itinerary may have a 5% chance of being wrong, which is not worth the candle. As a large vertical model, "Ctrip asks" screens 20 billion high-quality unstructured tourism data, and combines Ctrip’s existing accurate structured real-time data with Ctrip’s historical training robots and search algorithms to train the self-developed vertical model.
Zhihu also launched its own big model based on its own content advantages.
In April this year, Zhihu officially released the Chinese big model of "Knowing the Chart AI". Kelly Y Zhou, founder, chairman and CEO of Zhihu, said: "With the unique advantages of application layer and data layer, Zhihu will contribute to the big language model of Chinese Internet."
The first big model function "Hot List Summary" of the Chinese big model "Knowing the Chart AI" aims to capture, sort out and aggregate the answers to the questions on the Zhihu Hot List by using the super language understanding ability, and show the answers to the users, so as to provide them with intuitive and clear answers.
"It is very necessary for companies such as Ctrip and Zhihu to launch vertical models. At present, we urgently need large industry models provided by vertical fields. The vertical large model can get more social feedback in the process of continuous use through the general base large model, plus industry data, industry fine-tuning and industry application scenarios, thus achieving a high degree of intelligence. " Shenyang believes.
"Hundred Models Battle" Competition Final
The booming AI big model industry soon ushered in industry standardization and management norms.
On July 7, at the 2023 World Artificial Intelligence Conference, the National Artificial Intelligence Standardization General Group announced the list of the first large-scale model standardization special group leader units in China. The joint team leaders are six enterprises, including Beijing Baidu Netcom Technology Co., Ltd., Alibaba Cloud Intelligent Group, Iflytek Co., Ltd., Sanliuling Security Technology Co., Ltd., Huawei Cloud Computing Technology Co., Ltd. and China Mobile Communications Co., Ltd. Research Institute.
"Standardization of large models is very important." Wu Tian told the China Business Daily reporter, "Baidu, as the co-leader, hopes to actively play its role in promoting the research and development of core technologies, industrial application and ecological construction of large models, and promote the top-level design and rule development of large models standardization. Baidu itself is an important practitioner, with first-hand practical experience, and hopes to actively participate in it to promote and promote the development of the big model industry. "
On July 15th, the National Network Information Office, together with the National Development and Reform Commission, the Ministry of Education, the Ministry of Science and Technology, the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration of Radio, Film and Television, promulgated the Interim Measures for the Management of Generative Artificial Intelligence Services, aiming at promoting the healthy development and standardized application of generative artificial intelligence, safeguarding national security and social public interests, and protecting the legitimate rights and interests of citizens, legal persons and other organizations.
Whether it is general, industrial or vertical, the number of large models has increased rapidly in just a few months. At the 2023 Global Digital Economy Conference, Jiang Guangzhi, Party Secretary and Director of Beijing Economic and Information Technology Bureau, said that with the upsurge of large-scale model development in ChatGPT, more than 80 large-scale models have been publicly released in China so far.
With so many large models, the final result of competition is inevitably curious.
"From the current point of view, it is impossible for an AI big model to dominate the world, because whether the AI big model can be widely used depends on the technical ability on the one hand and the application scenario and traffic pattern on the other." Shenyang believes, "But from the ultimate direction, when we reach the stage of general artificial intelligence, we may only need a big AI model."
"the AI big model is just a phased phenomenon. In the process of evolution, various enterprises and institutions will gradually find their own positioning and move towards market segments. " Wu Tian told reporters, "I think the final result of the big model is that only a few big models will stay in the future." Because the cost of making a big model from the bottom is very high, it takes years of accumulation, firm confidence and continuous investment to do it well. In the future, we may only rely on a few large models, and these large models will have a very wide application ecology. "
Reporting/feedback