It's been 7 months, and the Chinese AI model can't learn from ChatGPT

Source: Titanium Media

Author: Lin Zhijia

Image source: Generated by Unbounded AI

Similar to the artificial intelligence (AI)-related conferences in the previous few weeks, at the 2023 Global Digital Economy Conference Artificial Intelligence Summit Forum, entrepreneurs and academics discussed the impact of large-scale models and generative AI technologies on society and industry.

"In the past few months, everyone has been discussing when OpenAI and China will be able to make their own large-scale models. In the past few months, large-scale models have emerged one after another. I see that many investors are starting to be anxious." Zhou Hongyi, chairman of 360, said in a speech at the forum He said that the real opportunity for large-scale models lies in China, especially the enterprise-level market, including governments and cities, and China's large-scale model should seize the opportunity of industrial Internet development.

Tang Wenbin, co-founder of Megvii Technology, Zhang Peng, CEO of Zhipu AI, Zhou Bowen, founder of Lianyuan Technology and Huiyan Chair Professor of Tsinghua University, and other guests also agreed with this view. They generally believed that large models should no longer be consumed A lot of money is spent on general large-scale models with hundreds of billions of parameters such as ChatGPT. Instead, it is necessary to build large-scale industry models with a billion-level industry, go deep into To B industry solutions, and solve practical problems in vertical fields.

In fact, as ChatGPT set off an upsurge in the investment and industrial development of China's AI large models, companies have followed ChatGPT and strived to become the "Chinese version of OpenAI". Now more than 80 large models have been publicly tested.

**But the problem is that many investors and entrepreneurs found that the technical threshold of large-scale models is not high after seven months of "rolling". Commercial solutions, while ChatGPT’s “$20 per month” Plus membership service can only become “free” in China, and the continuous AI technology investment of OpenAI is not feasible in China. **

**In the end, it was discovered that only customers in several vertical fields such as government affairs, finance, and medical care can pay for the "big model". **

Titanium Media App edits and statistics the latest information on domestic AI large-scale model data

Enterprises no longer compete with AI large-scale model technology, but return to commercialization

Over the past six months or so, large-scale model technologies represented by ChatGPT and GPT-4 have led a new round of global AI innovation. Large-scale model research at home and abroad is iterative at a high speed, various models emerge in an endless stream, and the performance of the models continues to improve.

In China, from Baidu Wenxin Yiyan, Ali Tongyi Qianwen, Huawei Pangu large model, to 360 smart brain, Zhipu AI ChatGLM, etc., all walks of life are building large models. According to Jiang Guangzhi, director of the Beijing Municipal Bureau of Economy and Information Technology, more than 80 large-scale models have been released in China so far, of which Beijing accounts for about half (more than 40).

"We seem to have climbed a mountain for a long time, and finally saw a new peak and a new dawn." Jiang Guangzhi said in his speech that in the past two decades, we have experienced the Internet, smart phones, cloud computing, big data, The Internet of Things, a series of major changes in IT technology, has profoundly changed all aspects of our production and life, and now we have ushered in a new era of general AI.

However, compared with the payment model of consumer-level platforms such as ChatGPT, China's AI large models are generally free for internal testing, and users can obtain a license for use as long as they apply for internal testing on the platform. Not only that, under the "Hundred Models War", everyone makes large models, and it is difficult to form barriers to the enterprise itself and the industry. In addition, the cost of AI computing power is high, and the commercialization of large models is not as easy as imagined.

To put it simply, the domestic AI model is far from OpenAI's business model and technology cost. OpenAI can invest hundreds of millions of dollars in computing power training costs, which cannot be achieved by domestic small and medium-sized AI companies, especially most of which are concentrated on the application side.

"The amount of parameters is actually a dialectical issue. Hundreds of billions of trillions of parameters just represent your highest ability, but high parameters mean higher requirements for thinking ability and high computing power, and customer scenarios do not necessarily require such high computing power requirements. Because not all scenarios can accept the cost of 100 billion parameters.” Dai Wenyuan, the founder and CEO of 4Paradigm, told Titanium Media App in May this year that under various constraints and companies looking for profits, To B customers began to choose cost-effective Controlled vertical AI mockup.

**At the 2023 Global Digital Economy Conference Artificial Intelligence Summit Forum, Zhou Hongyi, Zhang Peng, Zhou Bowen and other guests generally mentioned the "industrialization" and "verticalization" of large models, and went deep into government affairs, finance, cultural and creative industries, medical care and other industries Expansion, and discussions around model security, credibility, and controllability. **

Zhou Hongyi believes that the real opportunity for large-scale models lies in the enterprise-level market, and China should seize the opportunity of industrial development when it comes to large-scale models. However, when the current public large models are used in enterprise-level scenarios such as governments, cities, and industries, there are four shortcomings, including lack of industry depth, hidden dangers in data security, inability to guarantee credible content, and high training and deployment costs.

"There are many data security risks in the public large model. Each enterprise's own internal Know-how is the core asset, and the public domain large model will definitely not be trained; the use of the public large model will cause data leakage, because many ideas and plans You have to tell it, so that it can write a good article; the public large model is a generative AI, and the characteristic of its own algorithm is that it can talk nonsense, and it is serious and confident nonsense. This feature is used to make novels and scripts , as a chat robot, the effect is very good, but it is very problematic in an enterprise-level scenario. If you really use the prescription prescribed by the medical model, do you dare to believe it or take it; the public large model cannot realize the cost. It is controllable because the high cost of the large model is also exaggerated. It costs 10 million US dollars to train one time. How much computing power and how many graphics cards are needed. To make a vertical large model within the enterprise does not need to pursue comprehensive knowledge or comprehensive capabilities. Tens of billions of models may be enough, but the parameters from 100 billion to 10 billion seem to be ten times smaller, and the savings in training and deployment costs are far more than ten times."

Zhou Hongyi bluntly stated that the large-scale model that the enterprise-level market really needs in the future must conform to the characteristics of industrialization, enterpriseization, verticalization, miniaturization, and proprietaryization. Not only that, but in his opinion, the construction of enterprise-level large models must adhere to the following three principles:

1. Security: First of all, the principle of security. Large models have risks such as network security, data security, and algorithm security. Especially in terms of content security, some people have used AIGC to generate various fake content for fraud. Search engines are much more complicated. Therefore, the issue of artificial intelligence security needs to be studied from now on.
2. Credibility: The second is the principle of credibility, which can solve the "illusion" problem of large models and the problem that knowledge cannot be updated in time through the correction of enterprise knowledge base and search. "How to solve the problem of accurate output content? Now it can be corrected by vector database, enterprise search and external knowledge base."
3. Controllability: The last is the principle of controllability. Zhou Hongyi believes that when the large model is still a bit unreliable, it is suggested that when using the large model, enterprises and governments can take a small step at the beginning and do not open APIs and plug-ins to it. And the function pattern, still have to insist that it is a helper, the last person is on the "loop" of decision-making. He mentioned in the PPT that monitoring and auditing methods should be adopted to "lock the large horizontal type in a cage".

Regarding how enterprises use large models, Zhou Hongyi emphasized that enterprise large models must first make good use of general capabilities and give full play to the best and most mature capabilities of large models; focus on the pain points and rigid needs of enterprises, small cuts, large depth, such as information analysis and decision-making, Office scenarios such as enterprise knowledge search and management, office collaboration and intelligent customer service are the most suitable entry points.

At the same time, Zhou Hongyi also suggested that enterprises should first enable the "assistant" and "co-pilot" modes when using the large model, so that the large model can remain relatively independent from the existing business system and maintain isolation, which is also safer and more controllable for the enterprise. In addition, the large enterprise model should be used by both leaders and employees, so that AI is inclusive.

"It is always difficult for large models to complete a job independently, and many employees are resistant to large models. It is still difficult for large models to complete a job independently. More positioning is human Good tools and knowledge assistants, so the development of large models should be people-oriented.” Zhou Hongyi mentioned that ease of use is the first principle of the development of large models.

Zhou Hongyi judged that the digital human will become an important entrance and carrying form of enterprise-level large models. The "soul" digital assistant released by 360 Smart Brain can solve the problem of ease of use of large models. At the scene, Zhou Hongyi also showed the customized "Beijing Customer Service Can't Ask" and "Beijing City Merchants", two large-scale digital humans trained for Beijing to solve some core needs in the government and enterprise fields.

Not only Zhou Hongyi, but Zhipu AI CEO Zhang Peng mentioned that from the perspective of commercial implementation, Zhipu AI proposed the MaaS (Model As A Service) large model service concept, hoping to make hundreds of billions of high-precision large models be updated. Multiple individuals and enterprises can enjoy AI empowerment.

"We have three versions, one is end-to-end model training service, and has helped you complete some model migration training on the self-built computing power platform; the other is to provide model building services and licenses; the third is to cooperate with cloud computing vendors, Proposed API calls and Model Instance services to help everyone quickly build powerful infrastructure capabilities." Zhang Peng said.

Titanium Media App learned that Zhipu AI is developing a new ChatGLM2 large-scale model product, which reduces the number of parameters but improves data quality. Ability has been greatly improved. In multiple evaluations, ChatGLM2 scores more than GPT-4 and ChatGPT.

Zhou Bowen, founder of Lianyuan Technology and Huiyan Chair Professor of Tsinghua University, said that whether Al can be fully integrated with business is the key factor that determines whether Al can realize economic value. Only with a business-oriented AI strategy design, a complete supporting structure, sufficient AI talents, and a sound internal training mechanism can AI be fully integrated with business development needs and maximize economic benefits.

Fang Han, CEO of Kunlun Wanwei, believes that at present, large-scale models are in a state of reducing costs at the B-side and increasing efficiency at the C-side.

He believes that China's B-end service companies can see that it is difficult for any company to monopolize the entire B-end service. Due to the large model’s demand for industry data, it is inevitable that every large model company can achieve first-mover success in one or two industries. No company can succeed in all industries, and it is difficult for any company to obtain The data of the whole industry; and the C-end is bound to be fragmented. Because the payment habits of the Chinese market are actually very different from overseas, the free habits of Chinese users are very obvious. All companies that provide services to C-end users must be based on The free mode is the main mode, and the VIP mode is supplemented.

"The big tide will rise, and landing is king. We hope that this wave of entrepreneurship and investment in general artificial intelligence will be different from the previous wave, and it can land faster and generate users and income." Fang Han's theory is still biased Commercialization of large models. He mentioned that the AI-generated music products developed by Kunlun Wanwei have already landed in the cultural and tourism scene, and signed an agreement with Beijing Dongcheng District in April this year.

Zhang Xin, deputy general manager of China Telecom Group's Big Data and AI Center, announced the TeleChat large model developed by him on the spot, which supports outputting codes and writing speeches. Zhang Xin mentioned that the research and development goal of China Telecom Digital Technology Co., Ltd. is to build a 10,000-level AI algorithm cabin to become a 10-billion-level AI service provider. The products cover AI algorithms, platforms, applications, hardware, and large models.

This is the first time that China Telecom has announced its large-scale model products. Zhang Xin also said that compared with other large models, most of TeleChat products use domestic Chinese big data, 90% of which are mainly domestic, and based on China Telecom's Tianyi Cloud and cloud-network fusion base, TeleChat is used for model training parameters. At that time, the capacity can reach 47%, and the efficiency of model training and model algorithm capabilities are still improving.

In the application scenario, China Telecom's TeleChat model has begun to be intelligentized by manufacturers to solve the problems of talent shortage and insufficient coverage of some voice operators. Through the improvement of AI technology, the modernization of social governance capabilities can be realized and cost consumption can be reduced.

Guo Fan, Vice President of Unisound’s Innovation Division, once mentioned that the demonstration application of the outpatient medical record generation system based on the large model of mountains and seas jointly developed by Unisound and Beijing Friendship Hospital, in the field of smart medical care, based on Unisound’s 70 billion parameter-scale automatic Research the "Mountain and Sea" large model, combined with front-end sound signal processing, voiceprint recognition, speech recognition, speech synthesis and other full-stack intelligent voice interaction technologies, it is expected to improve the efficiency of doctors' electronic medical record entry by more than 400%, saving a single patient consultation time More than 40%, improving the efficiency of doctors' outpatient services by more than 66%.

In fact, large models are typically a winner-take-all area. More money, more computing power, and better talents are needed. Because better computing power means more people use it, more people use it means more data, and more data means better computing power results. The big model must be a battleground for the giants, who have money, technology, and more importantly, data.

But start-up companies have too many projects and their funds are too scattered, especially at the enterprise end, they end up consuming funds to buy Nvidia A100 cards and cloud services, without a quantified enterprise development process. Especially when AI companies and SaaS companies are generally difficult to make profits and make blood, start-up companies need to commercialize and make profits with large models.

Therefore, the current domestic AI large-scale model is taking shape based on the general-purpose large-scale model base and computing power center established by large companies such as Ali, Tencent, Baidu, Huawei, and SenseTime, as well as including Momo Zhixing, Tianyancha, Yunzhisheng, and China The large vertical or industry domain models established by companies such as Kewenge and Megvii only focus on one or two areas to solve core problems.

The government supports the landing of large-scale model scenes

At this forum, the Beijing Economic and Information Bureau once again announced the second batch of members of the Beijing General Artificial Intelligence Industry Innovation Partnership Program, and 63 companies were selected.

It is understood that as of June 30, a total of 416 large-scale model R&D and application companies inside and outside Beijing have applied to join the second batch of "Partnership Program". In the end, 63 companies, including Beijing Baidu Netcom Technology Co., Ltd., China Power Data Service Co., Ltd., and Beijing Jizhi Future Artificial Intelligence Industry Innovation Base Co., Ltd., were selected and announced. Among them, there are 10 computing power partners, 10 data partners, 10 model partners, 24 application partners, and 9 investment partners. In addition, 30 model observers were evaluated.

According to the estimates of Titanium Media App, as of now, more than 80 companies and institutions have been selected into the Beijing General Artificial Intelligence Industry Innovation Partnership Program.

**At present, Beijing is rapidly promoting the construction of AI large-scale models and industrial layout. **

On May 21, the Beijing Municipal People's Government issued a notice on the "Implementation Plan for Beijing to Accelerate the Construction of a Globally Influential Artificial Intelligence Innovation Source (2023-2025)". By 2025, Beijing's artificial intelligence technology innovation and industrial development will enter a new era. In the stage of development, basic theoretical research has made breakthroughs, the influence of original innovation achievements has continued to increase, and the scale of the artificial intelligence industry has continued to increase, forming an industrial cluster with international competitiveness and technological dominance.

On May 23, the General Office of the Beijing Municipal People's Government issued a notice of "Several Measures to Promote the Innovation and Development of General Artificial Intelligence in Beijing", requiring that the government's guiding role and the catalytic role of innovation platforms be fully utilized to integrate innovation resources, strengthen the allocation of elements, and create innovations. ecology, attach importance to risk prevention, and promote Beijing's general artificial intelligence to achieve innovation leadership and rational and healthy development.

Titanium Media App noticed on the spot that there are more and more cases and corporate participants about the implementation of AI large-scale models in the field of government affairs. Companies such as Ali, Huawei, Autohome, and Shibui Technology are all participating.

A person in charge of the Beijing Municipal Administration Service Bureau mentioned in a live speech that in order to support the construction of the platform "Jingce", the government needs to implement general large-scale model technology in the field of scenarios, so as to improve policy management and precise service capabilities. "In-depth mining and analysis of massive data on citizens' appeals provides stronger support for leadership decision-making, grassroots governance, and urban governance."

The person in charge mentioned above mentioned that in terms of scene advancement, an open small interface model will be used in the early stage, and privatized in-depth applications will be carried out later. In the long run, they will open up high-quality and credible government data sets by establishing digital protection mechanisms and technical evaluations under the condition of data security and controllability.

Wang Zhangsheng, head of Zhongke Wenge Delivery Center, mentioned in an exchange with Titanium Media App that government customers have high demands for large AI models, and this field also requires data security and data governance in the application of large models. Conditions, especially high-quality data training, so as to better solve practical application problems in this field.

Midu CTO Liu Yidong told Titanium Media App that the company started using Huawei Cloud infrastructure products last year to launch business in government affairs, media and other fields, and now the company is about to release large-scale model products for vertical industries that focus on online public opinion.

"The government has begun to pay for the 'big model'. On the one hand, it is policy guidance and demands for practical application scenarios. On the other hand, enterprises themselves need to solve cost problems and commercialization problems through large models." An industry insider analyzed to Titanium Media App, The domestic large-scale model industry is returning to the period when AI companies and the government are combined to form commercialization and receive government subsidies.

Jiang Guangzhi said that at present, Beijing is accelerating the construction of the National New Generation Artificial Intelligence Innovation and Development Experimental Zone and the National Artificial Intelligence Innovation Application Pilot Zone to create a globally influential source of artificial intelligence innovation. The specific measures include the following three points:

One is to strengthen policy innovation and standard guidance. Solve the problems of relatively scarce data quantity, difficult quality assurance, high cost of collection and labeling, and compliant use. Plan and introduce the "computing power voucher" policy to support small and medium-sized enterprises based on landing application scenarios to obtain diversified, low-cost and high-quality computing power, and support artificial intelligence enterprises in this city and related industry organizations in the formulation of national standards, industry standards, and local standards for artificial intelligence Play a leading role and participate in the formulation of technical standards in terms of model performance, data security, and privacy protection. The second is to increase the opening of scenarios and take the lead in implementing benchmark applications. Promote the city's government agencies, public institutions, state-owned enterprises and other organizations with the function of managing public affairs to actively purchase and use safe and reliable large-scale model-related products and services, take the lead in implementing applications in government services, smart cities and other fields, and improve urban governance capabilities level of modernization. The third is to promote the deepening and solidity of the partnership plan. In the current partnership program, more than 10 computing power partners plan to provide no less than 4000P low-cost and high-quality computing power for small and medium-sized artificial intelligence enterprises in Beijing to carry out large-scale model training and reasoning. 10 data partners have released 18 high-quality data sets of nearly 500T for large-scale enterprise training. At the same time, it will also accelerate the promotion of large-scale model industry applications, give full play to the role of the partnership program, a market-oriented docking and cooperation platform, and create a good ecology for innovation, cooperation and application of large-scale models in Beijing.

Jiang Guangzhi emphasized that the Beijing Economic and Information Bureau will continue to strengthen the allocation of high-quality resource elements, effectively integrate innovative resources, actively create an innovative ecology, and lay a solid foundation for the development of the artificial intelligence industry.

"Large models cannot be made by one company. It is best for everyone to form several large ecosystems at the levels of computing power, models, and data. Especially under the guidance of government departments, companies can form partnerships in the ecosystem, so that Everyone is more about cooperation than competition. In the future, the large model may form a relationship with ecology, partners, and win-win cooperation.” Ji Haiquan, executive director of Legend Capital, said.