The reporter learned from recent research that AI (Artificial Intelligence) large model is equivalent to "Super Brain" and is becoming a "new highland" of artificial intelligence. AI big model is expected to realize the transition of artificial intelligence from perception to cognition, redefine the industrial model and industrial standards of artificial intelligence, and bring major changes to some industries. China has a large AI model application market, but it faces multiple challenges in the development process, such as weak technology, scarce talents and high cost. It is urgent to guide and support relevant technology R & D and industrial layout. "Super Brain" of "learning rich and five cars" —— "If you are unhappy, you can listen to some happy songs, watch comedies, go to exercise or have a good rest." —— "A bad mood is a very normal mood, so don't blame yourself too much." An artificial intelligence platform gives different answers to the question "what to do if you are in a bad mood today". Enter the platform through the website link, and input questions randomly in the "man-made Q & a" area. The machine can support and play different "man-made" and give multi angle answers to the questions. The platform is built based on the large AI model. Zhou Ming, chief scientist of Innovation workshop and founder of Beijing Lanzhou technology, introduced that AI big model, also known as artificial intelligence pre training model, imports massive data into models with hundreds of millions or even hundreds of billions of parameters. The machine deeply learns the characteristics and structures contained in the data by doing tasks such as "Gestalt filling", Finally, it is trained into artificial intelligence with the ability of logical reasoning and analysis. Generally speaking, AI big model is equivalent to making countless sets of "super brains" about knowledge exercises and simulation problems in various fields. It is well versed in the internal logic and problem-solving ideas of knowledge in various fields. It can not only understand the knowledge system of the human world, but also produce new knowledge. Since 2018, artificial intelligence has entered the "big model era". From the scattered and inefficient situation of repeated development and manual workshop artificial intelligence, that is, "there are 1000 small models in 1000 application scenarios", it has moved towards a new path of industrialization and integrated intelligence, and brought dawn to general artificial intelligence with a big model. For example, in order to produce intelligent audio, all brands compete to develop their own AI small models. In the future, all brands only need to develop on the basis of the same AI large model. "It used to be 'fragmented' artificial intelligence, and different models should be developed for different application scenarios. Workshop development consumes a lot of resources and costs, and is inefficient. In the future, when landing in various vertical fields, it only needs to fine tune its parameters based on a large model, so as to create a universal artificial intelligence." Tian Qi, chief scientist in the field of Huawei cloud artificial intelligence and academician of the International Eurasian Academy of Sciences, introduced that the AI model will redefine the industrial model, industrial standards and even industrial ethics of artificial intelligence. Interviewees cited examples. The previous small models were like "partial machines", which only learned limited data for specific application scenarios, and lacked the ability to draw inferences from one instance. Some intelligent products were ridiculed as "artificial mental retardation" by users from time to time. The AI big model learns all kinds of data from all walks of life and becomes a "generalist" with good knowledge transfer ability. It has wider application fields, higher quality and more intelligence. At the same time, Zhou Ming pointed out that the development of large models does not blindly pursue a large number of parameters, but focuses on the high performance of models. "While allowing large models to learn and master more data, we should make large model parameters relatively lightweight as far as possible, and improve energy efficiency while 'reducing the burden'." It can be decorated with multiple corners and has a wide range of application scenarios Automatic consultation, product marketing, novel continuation... Because it understands human language and has professional knowledge in subdivided industries, a large AI model can be divided into multiple roles. It can be a doctor, a writer, an emotional blogger and an e-commerce customer service AI large model industry has a wide range of applications and has a wide range of application scenarios in the fields of medical treatment, finance, retail, meteorology, news communication, literature and art, etc. Zhou Ming introduced that after inputting control attributes such as keyword, theme and text style in the large model, text content meeting constraints can be generated, which can be used for product marketing copywriting, e-commerce intelligent customer service, etc. In finance and other fields, large models independently complete screening and analysis by capturing industry data, and independently write and generate industry reports to assist employees in making decisions. Tian Qi introduced that the AI large model has been applied to the daily line detection of the power department of the State Grid. "In the past, to identify defects on various power transmission lines, it was necessary to make a small model for each type of defect. Hundreds of models were needed for hundreds of defects. Now, the identification of various defects can be completed by using a large model." It is worth noting that the current technological progress of AI large model is a key link in the evolution of artificial intelligence from perception to cognition. Liu Zhiyuan, an associate professor in the Department of computer science and technology at Tsinghua University, said that all articles and materials at all times, at home and abroad, can be used as data for training large models. By analyzing and "internalizing" a large amount of knowledge, intelligent products not only have higher performance, such as more accurate machine translation and more vivid and logical machine writing; It can also taste human emotions for emotion labeling. In addition, "it may recognize 'overtones' and even hope to discover knowledge that human beings have not mastered." The development of this technology may also bring significant changes to some industries. Zhou Ming said that a new generation of search engines will be produced in the future. "At present, most search engines are based on keyword recognition and only 'carry' the results containing keywords in the library to users. The new generation of search engines based on large model can independently generate answers by understanding human language, and support multimodal search such as text, picture, audio and video." Tian Qi said that the new generation of search engines have creative ability, such as searching for things that do not exist in reality, such as "winged cats" and "flying pigs", and search engines can also generate corresponding photos according to the description. Multiple development problems to be solved At present, American openai, Google, Microsoft, Facebook, NVIDIA and other overseas companies have laid out AI large model industry one after another. Domestic large enterprises such as Huawei, Alibaba and Baidu, as well as universities and scientific research institutions have also joined the R & D track. AI big model is becoming a "new highland" of artificial intelligence. According to the reporter's investigation, technically speaking, the domestic AI model has been preliminarily mature. In the next step, we will focus on industry promotion and iterate the model in combination with specific problems. Tianqi predicted that the large-scale application of AI model "may take two to three years", and in the future, AI model is expected to be integrated on the chip, which is more convenient for landing application. Experts pointed out that China has a large AI large model application market and has advantages in the application field, but it also faces some challenges in the development process. Liu Zhiyuan introduced that many core technologies are still in the hands of developed countries. Training AI large models requires a large number of GPU (graphics processor) chips, but the current domestic GPU chip technology is relatively weak. At the same time, China's original innovation ability in algorithm also needs to be improved. Many respondents said that at present, only a few universities, scientific research institutions and enterprises are engaged in the research and development of AI large model and can produce innovative results. Compared with foreign countries, the cultivation of high-level basic talents in relevant fields in China needs to be strengthened. In addition, the training of large models requires high power costs and equipment costs. Due to the high cost, many small and medium-sized enterprises and scientific research institutions are unable to build computer rooms with sufficient computing power, facing the problem of insufficient computing power. Take multiple measures to seize the "new highland" Many experts said that AI big model is expected to realize the transition of artificial intelligence from perception to cognition, and will more efficiently enable AI industrialization and industrialization. The effective implementation of relevant technology R & D and industrial layout guidance, supporting policies, reasonable supervision and other measures is very important. First, promote the standardization of AI large model. The interviewed experts pointed out that the training process of the large model has the characteristics of high energy consumption. After the standard is formulated, many work can be adapted and redeveloped based on the developed standardized large model, and there is no need to train from scratch every time to reduce energy consumption. At the same time, the development of the large model may also derive risks such as the generation of illegal information, privacy disclosure and dissemination of false information. Standardizing the source data used for training the model is conducive to avoiding relevant risks. Relevant national departments can lead or guide enterprises, research institutes and universities to jointly define the standards of large models. Secondly, open the training data appropriately. Respondents said that training large models requires massive data. Out of concerns about data privacy and security, it is difficult to obtain data in some fields, forming a data island. It is suggested that on the premise of national institutional supervision and macro control, the data of all parties can be appropriately opened to white list enterprises, institutions and universities, so as to enhance the strength of China's AI model while ensuring the safe use of data. For example, the appropriate opening of data by hospitals and other institutions will be conducive to the training of a large model with a better understanding of medical care and enable the development of the field of health medicine. Thirdly, encourage the sharing of computing power. According to the reporter's research, at present, only some large enterprises can afford the computing power cost required for super large model training. The interviewed experts called for guiding the "national team" of artificial intelligence with strong computing power to provide more small, medium-sized and micro enterprises and scientific research institutions with the computing power support required for large model training, explore the computing power sharing mechanism with reasonable payment, and jointly promote the development of technology and resource conservation. Finally, increase the guidance and support of original innovation. Many experts pointed out that China should strengthen research and development, strengthen basic innovation research for the model framework, guide more scientific research institutions and universities to actively cultivate relevant talents, and encourage more social forces to join the wave of the information revolution. (Xinhua News Agency)