The large model promotes the digital and intelligent transformation of thousands of industries

With the rapid development of artificial intelligence technology, large models have been continuously integrated with various fields with powerful digital processing capabilities and deep learning capabilities, and have gradually become the key grasp of industrial innovation and the key engine for driving new quality productivity. How does the big model change our lives, and how does it empower thousands of industries? Where will the future development go? At the 2023 Chinese Artificial Intelligence Industry Annual Conference held recently, experts and scholars attending the meeting had an in-depth discussion on this.

Embrace more application scenarios

At the forum, Xu Mai, a professor at Beihang University, opened with a speech generated by Kimi, a large model of China: large models and general artificial intelligence are a field that is dynamic and leads the future...... Large models make the interactive experience richer and more realistic, which can be called artificial intelligence in the true sense.

As the most powerful artificial intelligence technology at present, the "bigness" of the large model lies not only in its large number of large-scale parameters, but also in the huge potential and broad application scenarios it contains. In addition to common content generation, large models have shown advantages in many fields such as autonomous driving, smart healthcare, and industrial Internet, truly empowering thousands of industries.

Zheng Qinghua, an academician of the Chinese Academy of Engineering and president of Tongji University, said that large models have become the pinnacle of artificial intelligence, and that predictions about the exact structure of proteins in the past may take months and thousands of people to complete, but now with the support of large model analysis, results can be generated in just a few minutes.

Kimi, the dark side of the moon, Baidu Wenxin Yiyan, iFLYTEK Spark model...... General large models have sprung up, and the application of the large model industry has been further accelerated. Liu Cong, vice president of iFLYTEK and dean of the research institute, introduced that as a representative of the domestic general model, the Xinghuo model has a strong performance in empowering the industry, and has cooperated with Chery Automobile to create a large model cockpit, taking the lead in making the domestic large model applied in the field of intelligent cockpit and producing practical results.

For specific scenarios, not all enterprises need the "omnipotence" of a general large model, but the accuracy of the model is more needed. Zi Ran, head of the security GPT business of Sangfor Technology Co., Ltd., said that compared with the general large model, if you want to realize the commercial operation of the technology faster and make users willing to pay for it, the vertical large model may have more advantages in landing.

As a practitioner who has been deeply involved in the field of network security for many years, Zi Ran believes that current network attackers have been using large models to improve their attack methods, which are difficult to identify with traditional detectors, while the security GPT detection large model surpasses the general large model in terms of malicious code understanding ability, attack and defense confrontation understanding ability, and security basic knowledge ability, and has low inference cost, high accuracy, and easy to land. At present, the vertical model developed by his team has served more than 130 enterprises.

Environmental pollution control has also become one of the main areas for the implementation of vertical large models. The team led by Qiao Junfei, vice president of Beijing University of Technology, has made many innovative achievements in intelligent feature modeling, self-organizing control and multi-objective dynamic optimization of pollution prevention and control processes. Qiao Junfei said that in the past, it was difficult to establish an analytical model for environmental pollution prevention and control, mainly relying on manual decision-making, while the landing of vertical large models provides technical support for environmental pollution control, relying on data to make more objective and accurate judgments on pollution treatment, and achieve scientific and precise pollution control.

There are also multiple abilities that need to be tempered

While bringing great changes to production and life, artificial intelligence models still face challenges in terms of data, computing power and algorithms.

On the one hand, computing power is the "fuel" of large AI models. Gao Wen, director of Pengcheng Laboratory and academician of the Chinese Academy of Engineering, who won the "Wu Wenjun Highest Achievement Award for Artificial Intelligence", said that artificial intelligence cannot be done without computing power. If you want users to use computing power as easily as electricity, you need to solve problems such as data security and result transmission. In this regard, Pengcheng Laboratory is taking the lead in promoting the R&D and construction of "China Computing Network", and the "Pengcheng Cloud Brain" built by it has produced good results in super computing nodes, and has played an important role in the follow-up construction of domestic computing power ecology.

On the other hand, high-value data provides a constant supply of "raw materials" for large models. The larger the model parameter size, the higher the requirements for the size and quality of the data. Zheng Qinghua said that high-value data is like mineral resources that are not inexhaustible. Some experts predict that by 2026, the mineable value of large-scale corpus will be basically exhausted, and it will be difficult to retrain valuable data on big data. In addition, the current proportion of Chinese corpus is not high, and some data show that in ChatGPT training data, the proportion of Chinese corpus is less than one thousandth.

At the same time, the catastrophic oblivion of large models has also attracted widespread attention in the industry. Catastrophic amnesia is when training on a new task impairs the performance of the previous task. During the problem-solving phase, it is not possible to remember the data or scenarios that have been processed. For example, in autonomous driving, people have memories of road conditions, but autonomous driving cannot remember them, and they have to recalculate each time, so they consume a lot of computing power and electricity.

Zheng Qinghua suggested that we can start from three technical routes: one is to rely on big data, large computing power and strong algorithms to promote the development of large models, that is, to extend the original technical route, and the other is to use the "neural + symbol" collaborative way to combine the powerful learning, universality and interpretability of symbolic reasoning of neural networks. The third is the machine memory intelligent model inspired by human brain memory, which is expected to solve some inherent defects of the current large model.

Actively explore development paths

When it comes to the future development trend of large models, many experts have mentioned the combination of embodied intelligence, and humanoid robots are one of the best carriers. What is the relationship between large models and robots?

Liu Cong gave an example: if you want to go to a drawer and take a bag of potato chips and put it on the table, the large model will make a plan based on the understanding of this matter, that is, open the drawer, take out the potato chips, and then close the drawer, which is more like a kind of detached intelligence. However, if it is combined with embodied intelligence, it will decompose the instruction just now, and combine each instruction accordingly, and realize the action by controlling the robot.

Zhao Jian, a researcher at the Institute of Optoelectronics and Intelligence of Northwestern Polytechnical University, said that the current artificial intelligence model is constantly expanding to the multi-modal field, and the combination of embodied intelligence and multi-modal large model can achieve a higher level of perception, understanding and decision-making, so that artificial intelligence can effectively respond to various complex situations in the real world, bring broad application prospects to many fields, and become a bridge and link between AI and real life.

This year's "Government Work Report" proposes to deepen the R&D and application of big data and artificial intelligence, carry out the "artificial intelligence +" action, and build a digital industrial cluster with international competitiveness.

In response to the "artificial intelligence +" action, Zheng Qinghua believes that from the construction of the discipline system and talent training, it is not only about concepts and methods, but also about practical actions. First, it is necessary to put forward clear measures to empower artificial intelligence as an important means to promote innovation and development and achieve new quality productivity. By transforming the traditional training program curriculum system, artificial intelligence technology has become a basic ability. The second is to transform the traditional experimental platform and experimental methods, so that you can really study algorithms and experience the whole process of artificial intelligence from design, development to application.

From the perspective of industry development, Zi Ran said that practitioners from all walks of life should learn and embrace large models as much as possible, explore development paths, and make large models better empower the development of thousands of industries.

Zheng Qinghua said that in the process of moving towards cognitive intelligence, many theoretical and technical problems still need to be broken through. In the future, through continuous research, we hope to develop a new machine model inspired by human brain memory, and finally break through the upper limit of the ability of large models, achieve self-reliance and self-reliance, and make new contributions to our artificial intelligence base and model with independent intellectual property rights. (Economic Daily reporter Li Siyu)

Source: China Economic Net