AI cold thinking under the upsurge of "thousand models war"

Original source: International Finance News

Although the 2023 World Artificial Intelligence Conference has ended, the large-scale model boom initiated by ChatGPT will continue to surge, and the layout and exploration of artificial intelligence will also be a major theme in recent years. "Go to the table first" and "take the first move" are the industry consensus that set off this round of "100-model war" upsurge, and lead the industry into the "1000-model war".

Image source: Generated by Unbounded AI‌

At the beginning of July, there was a boom in artificial intelligence in Shanghai. The **2023 World Artificial Intelligence Conference has reached the highest number of exhibitors and exhibition area ever. Many companies announced at the conference that they will release large AI models. **The high temperature and strong convective weather failed to dissipate everyone's enthusiasm. The gate of the exhibition once attracted scalpers to sell tickets, and many people helped the old and the young to explore the cutting-edge development trend of artificial intelligence.

Under the upsurge, ** should also calmly see that the large model still faces core problems such as robustness, compliance and credibility. **Compared with developed countries, my country still has a gap in chips, computing power, data, etc. Data scarcity is a major problem affecting the application of large models. Among them, the difficulty of obtaining high-quality Chinese corpus data restricts domestic large models a major factor in development.

At the current stage when the core issues have yet to be broken through and the gap is being bridged, what kind of development path should China's AI development explore? During the three-day forum and interviews with many participating industry experts, the most answers the reporter got were "vertical integration" and "landing application". AI development trend.

"Go to the table first"

At present, the development of digital economy has become a global consensus. As a strategic emerging technology, artificial intelligence is increasingly becoming the core driving force for industrial upgrading and productivity improvement. In November 2022, OpenAI launched ChatGPT, a large-scale conversational general-purpose artificial intelligence model. A new round of AI innovation boom was launched around the world.

**At the 2023 World Artificial Intelligence Conference, the large model is the leading role. **Baidu Wenxin Yiyan, Alibaba Cloud Tongyi Qianwen, Huawei Cloud Pangu, Xunfei Xunhuo, Shangtang Rixin, Lanzhou Mencius MChat, Xinghuan Wuya Transwarp Infinity, Midu Honey Nest Series, Torsi General and vertical large models such as Tuotian and Daguan "Cao Zhi" are dizzying.

Lin Jianming, founder and chairman of Samoyed Cloud Technology Group, pointed out in an interview with a reporter from International Finance News that **AI is at the starting point of a new round of industrial trends. ** From the perspective of the layout of the large model, Baidu, Ali, Huawei and other "high-end players" carry out a "four-in-one" layout from the computing power layer, platform layer, model layer, and application layer; scientific research institutions and start-up technology companies have found another way, The entry point is to develop large-scale model algorithms and subdivided field applications.

Lin Jianming said that the current domestic large-scale model parameters are basically in the scale of 100 billion or more. In terms of application direction, most enterprises mainly focus on internal applications in the early stage, and gradually extend to B-end enterprises. **Artificial intelligence technology continues to make breakthroughs. Large manufacturers and small and medium-sized technology companies are competing for large models. Naturally, no one wants to miss the big wave of this era. Only by "going to the table first" can you grasp the "trump card" of the rules. In the context of the fading dividends of the mobile Internet, choosing to embrace large models is expected to bring new growth points.

Zhou Bowen, IEEE/CAAI Fellow, Huiyan Chair Professor of Tsinghua University, tenured professor of the Department of Electronics, and founder of Lianyuan Technology, told the reporter of "International Finance News" that **China should adopt a large-scale system based on "independent innovation, safety and controllability". The development route of language model and generative artificial intelligence technology focuses on promoting the wide application of large models with general-purpose capabilities in vertical industries. **In addition, business applications, academic innovation, and technological ecology all need to be diversified, and cannot be completely concentrated on a large model, nor should they all use one way of thinking to do things.

Multiple Challenges

Under the AI upsurge, large models still face multiple challenges such as robustness, compliance and credibility. Lin Jianming said bluntly that compared with the world, especially the United States, we still have a certain gap in AI chips, patents, algorithm research, and a mature innovation ecosystem. **The main factors restricting the development of domestic large-scale models are: first, large models require large computing power, and we have shortcomings in chips and computing power; second, lack of high-quality Chinese corpus data and industry data; third, the number of professionals Rare, basic research innovation is not enough.

"The financial industry is a special existence with very high requirements for risk management and security. The challenges of trust risk, model risk, ethics, stability, accuracy, data security, compliance and other risks faced by the research and development of financial large models are more severe .” Lin Jianming pointed out.

Jiang Ning, deputy general manager and chief information officer of Mama Consumer, said in an interview with a reporter from the International Finance News that the **AI large model still faces core issues such as dynamic adaptability, robustness, compliance and credibility in key decisions , How to eliminate noise and interfering issues, in sudden and unpredictable situations, it is especially critical to achieve continuous stability and compliance and credibility of key decisions. **

Jiang Ning pointed out that domestic large-scale models lack original breakthroughs, and there are still gaps in model reasoning ability and large-scale model generation ability. The difficulty of obtaining large-scale and high-quality Chinese corpus data is a major factor restricting the development of large-scale models in China. Specific to the financial field, it also faces many challenges such as privacy protection, continuous stability, compliance and credibility.

Zhou Bowen believes that in the current training of large AI models, the algorithm side converges to the neural network Transformer model, the computing power side relies on AI server clusters with large-scale parallel computing capabilities, and the data side needs to feed large-scale data sets with a huge amount of data. From the perspective of the three elements of AI, data scarcity is obviously a major problem leading to the implementation of large-scale model applications. Specific fields such as the financial industry, which have extremely strict requirements for data security and user privacy protection, also pose a series of challenges to large models such as trustworthiness, autonomous controllability, and strong security.

Zhou Bowen said that the industrialization of large-scale models is also facing challenges: first, the data scale is large and the data quality is uneven; second, the model is large in size and difficult to train; ** Therefore, the development of large models depends on the comprehensive support of algorithm computing power and data. Large models are the focus of future industrial development, but the business model of large models is worth exploring. Because the cost barriers of large models are very high, both large companies and small businesses have their own burdens.

Vertical integration

At the current stage when the core issues have yet to be broken through and the gap is being bridged, what kind of development path should China's AI development explore? What other development opportunities are there? Jiang Ning pointed out that building a combined AI system is a development trend, effectively combining the usability and professionalism of discriminative models in various vertical fields, and the characteristics of transfer learning and generalization capabilities of generative large models, so that it will be truly popular in the industry. Take advantage of the generalization ability of large models.

Lin Jianming pointed out that in the future, large-scale models will have great potential in the digital intelligence process of cities, industries, and enterprises. The domestic layout of large models needs to strengthen independent innovation capabilities, enhance the core competitiveness of large models from various levels such as computing power, algorithms, and talents, and also closely combine national strategic needs and industry development directions to explore industry pain points and scenarios in depth.

In addition, "it is necessary to use its own technology, scenarios, user and industry data and industry Know-How (industry secrets) to create a large vertical field model; to empower the real economy with 'general model + industry Know-How special model' and establish its own barriers advantage." Lin Jianming said.

Zhou Bowen believes that the large-scale model industry in China should start from end-to-end, and iteratively develop larger business models, which may be a more suitable approach. On the basis of having general capabilities, continuous training in the vertical field to improve the professional capabilities of large models is an important means to help the development and progress of large models in the future. **

Zhou Bowen pointed out that from a theoretical and technical perspective, differences must exist. In the development of AI, on the one hand, we are a catcher at the technical level, and on the other hand, we are likely to become innovators and even leaders at the application level. **China's AI needs to explore a new path, that is, to vertically integrate from the self-developed general model to the application and the user's full-scenario closed-loop, so as to realize the "dual landing" of generative artificial intelligence technology and commercial value. **

For entrepreneurial competition, Zhou Bowen believes that it can be divided into three routes: the first route is to build a large-scale underlying model with general capabilities, from technical algorithms to model iteration, and scene closed-loop; the second route is based on other people's models ( Such as GPT), and then combine their own industry Know-how to do training; the third route is purely for application, which is to use the model directly, and this barrier will be lower.