Baidu Intelligent Cloud Wenxin Big Model Technology Exchange Conference held in Shenzhen recently, Baidu, Tuworm Creative/Photo Source: IDC Zhou Jingyu/Cartography
Securities Times reporter Yan Cui
Artificial intelligence (AI) models are developing at an unprecedented pace, playing the song of "ice and fire".
On the one hand, the number of domestic AI large models has increased significantly, related applications have accelerated, the future development path of the industry has gradually become clear, and the algorithms and models of the large models themselves have continued to be iteratively optimized. On the other hand, AI-related scams occur frequently in China, and some companies have even banned new applications after they have been launched.
The reporter of Securities Times · e company exchanged with domestic experts in the field of AI large models and found that the industry has reached a consensus on the future prospects of AI large models, and the development is unstoppable, and large models have brought the "iPhone moment" of artificial intelligence. However, at the same time, the risks, safety and other issues caused by AI cannot be ignored, which has attracted the attention of enterprises, governments and other circles.
Some companies have achieved full control over generative AI input and output content, while others are reverse designing products that decode and resolve AIGC (generative AI) outputs. Experts collectively call on the society to increase the supervision and review of artificial intelligence applications, improve the ability to distinguish and prevent artificial intelligence, give full play to the advantages of large models, and overcome its defects.
Rapid development
Domestic AI large models show a rapid development trend.
On May 29, at the Zhongguancun Forum, Zhao Zhiyun, director of the China Institute of Science and Technology Information and director of the New Generation Artificial Intelligence Development Research Center of the Ministry of Science and Technology, said that according to incomplete statistics, 79 large models of more than 1 billion parameters in China have been released, and Beijing and Guangdong have the most from a regional point of view, with 38 and 20 large models respectively, and natural language processing is the most active field for large model research and development.
"China's large models are showing a vigorous development trend, and various technical routes are making breakthroughs in parallel, especially in natural language understanding, machine vision, multi-modality, etc., and a number of influential AI models in the industry have emerged." Zhao Zhiyun said.
Xin Zhou, general manager of Baidu Intelligent Cloud AI and Big Data Platform, told the Securities Times e company reporter that since the launch of the internal test on March 16, the performance of Wen Xin Yiyan has been continuously upgraded, and many major version iterations have been completed, and the inference performance has been greatly improved.
On May 18, iFLYTEK Chairman Liu Qingfeng said that iFLYTEK is now only one step away from ChatGPT, only two points away from 100 points, and will surpass it by October this year. Previously, Liu Qingfeng said that iFLYTEK Xinghuo cognitive big model has surpassed ChatGPT in three major capabilities: text generation, knowledge question answering, and mathematical ability.
"Large models and generative AI have two major features, one is to change the way people and machines communicate, allowing people to interact with machines or various interfaces through natural language, which is a disruptive innovation." Xin Zhou told reporters, "Another major feature is that it will continue to iterate and optimize, and future products will do everything possible to combine with generative AI natural language interaction, and relevant interaction data can form a closed loop and continue to iterate to further improve interaction efficiency." ”
Increased commercial use
The application of AI large models is accelerating, and the industry generally believes that there will be more and more applications of AI large models in the future.
Liu Qian, general manager of Baidu Intelligent Cloud AI Application Product Department, revealed to reporters that at present, more than 150,000 companies have applied to test Wenxin's words, of which more than 300 ecological partners have achieved test results in more than 400 specific scenarios, whether it is government affairs, finance, media, e-commerce, retail or other industries, there are a large number of application needs, content creation, intelligent customer service, intelligent marketing, enterprise knowledge management, office efficiency improvement and other aspects are frequently occurring application scenarios.
Xu Hao, General Manager of the Skydome Platform Solution Division of Kingdee Software Co., Ltd., said that Kingdee is working with Wenxin Qianfan Big Model Platform to explore the landing of large models and enterprise management software in AI application scenarios such as intelligent financial management, intelligent human resource management, and intelligent operation management.
Liu Qian also introduced that Baidu Intelligent Cloud's six intelligent product series will also be comprehensively upgraded based on the Wenxin model, including: government affairs, finance, intelligent creation, intelligent customer service, digital human live broadcast platform, enterprise knowledge management, etc., which will be launched after the security assessment is completed in the future. Taking intelligent customer service as an example, after the upgrade, the knowledge productivity is increased by 9 times, and the cost of multi-round dialogue construction is reduced by 65%, which can enable end users to obtain a smarter and more person-like customer service experience.
Recently, the Ministry of Education and other 18 departments jointly issued the "Opinions on Strengthening Science Education in Primary and Secondary Schools in the New Era", requiring that school teaching and services should be improved, and experimental teaching should be explored to explore the use of artificial intelligence, virtual reality and other technical means to improve and strengthen experimental teaching, making up for the lack of high-quality education and teaching resources, and also boosting confidence in the future development and application of generative AI to a certain extent.
In addition, OpenAI, Huawei, Ali and other enterprises are accelerating the implementation of AI large models, among which OpenAI took the lead in launching ChatGPT applications for Apple mobile phones on May 19, taking an important step in promoting the landing of ChatGPT applications.
"Both incremental markets and emerging opportunities are expected to bring new inflection points to AI market growth in the next 12-24 months." Lu Yanxia, research director of IDC China, said, "It takes more than 12 months for the new generation of AI to form a market scale, but its landing speed and possible industrial impact cannot be underestimated. Market participants and industry users alike need to carefully consider the impact it may have on their current business and actively introduce the next generation of AI. ”
Accelerate penetration
In addition to the accelerated landing of B-side and C-end applications or the upgrade based on AI large models, large models have also spawned another business model, that is, the serviceization of large models, empowering thousands of industries, and building industry large models that meet the characteristics of various enterprises and industries, so as to reduce the cost of developing AI large models and related applications in the whole society, and accelerate the penetration of AI large models in all walks of life.
After the emergence of this model, application developers can pay less attention to the underlying architecture or development logic of the application and focus on the end-to-end connection and interaction between the user side and the enterprise development side. Baidu Intelligent Cloud's Wenxin Qianfan Big Model Platform adopts this model, which not only provides large model services including Wenxin Yiyan, but also provides various tool chains and complete environments for developing AI applications, supporting various third-party open source and closed-source large models, and becoming a distribution center for large model production and distribution.
In other words, after the launch of the AI big model, the industry rules of the cloud computing industry have been completely changed, and the business model of the cloud service market has added a new business model MaaS (model as a service) in addition to the previous Iaas (infrastructure as a service), PaaS (platform as a service), and SaaS (software as a service). Cloud service providers can directly make large model products and services to earn profits, and enterprises that are interested in developing or training large models only need to purchase the MaaS of cloud service providers, select the appropriate model and purchase method, and can realize the development and training, inference, etc. of their own large models at a lower cost, without the need to build and train basic large models starting from 0.
"In general, we do not recommend that enterprises train large models from 0 because the cost is very high. For example, the new large-scale language model LLama released by Meta used about 45T of data, 1.2 million card hours, 120 servers, and more than a month of training, and the entire cost may cost millions of dollars. For enterprises, it is important for enterprises to train a proprietary large model, but it is not necessary to start from 0. Xin Zhou, general manager of Baidu Intelligent Cloud AI and Big Data Platform, said.
As for the external concern about whether the public cloud deployment data is secure, Xin Zhou told reporters that if the data exists in the public cloud, Baidu will have a standard cloud-compliant security mechanism, and there will also be physical isolation methods such as resource isolation mechanisms to protect user data security.
Zhou Jingren, CTO of Alibaba Cloud Intelligent Group, also said recently, "The model-centric development paradigm (MaaS) has become the industry standard, and the entire link of future application development will be based on this concept. ”
Zhu Yong, vice president of Baidu Intelligent Cloud, told the Securities Times reporter, "In the past, enterprises chose cloud vendors mainly to consider basic cloud services such as computing, storage, and network, and today we should look at whether the model is good, whether the framework is good or not, and whether the collaboration between the four layers of the architecture, chip, model, framework and application, is efficient." In the future, our AI developers will mainly develop based on large models, and do not need to pay too much attention to other lower-level technical details. ”
"AI applications mainly have two costs, one is infrastructure costs, such as computing power, storage, including servers, etc., the second is AI application development costs, which will require a lot of manpower, algorithm experts, etc. After the emergence of the Wenxin Qianfan model platform, the threshold for enterprises to develop AI applications has been significantly reduced, and the cost has also been greatly reduced." Zhu Yong said that with the rise of applications and the HP of large model platforms, the cost of AI large model training, inference and related AI application development will decline rapidly.
"In the future development of AI big models, only a few companies may develop basic large models, and other industries will develop their own large models according to their own business needs, data volume, etc. to access the large model platform, and the development path of the entire industry has become increasingly clear." An AI industry source told reporters.
Face the risks
The future of AI large models can be expected, but the risks or security issues brought about by rapid development cannot be ignored.
Recently, the police have intensively released AI fraud cases, and they were deceived 2.45 million yuan in 9 seconds and 4.3 million yuan in 10 minutes... Many victims in Anhui, Shanghai, Fujian and other places received video calls from "relatives and friends" asking for money transfers for various reasons. However, the video call turned out to be a video synthesized by fraudsters using AI, and the victim was deceived because the video "confirmed" the identity of the other party.
In addition, while OpenAI has pioneered the introduction of ChatGPT for Apple phones, Apple has banned employees from using ChatGTP and other AI tools. People familiar with the matter revealed to the media that this is because Apple is worried that the company's confidential data may be leaked during use.
It can be seen that the risks and safety issues related to AI have caused certain negative impacts on society and should cause great concern.
Zheng Lei, chief economist of Samoyed Cloud Technology Group, told reporters that AI fraud is an inevitable problem in the development of technology, and the industry is also aware of the importance of this problem, and is designing products that reverse decode and distinguish AIGC output results, but this product will lag behind the development of AIGC technology and needs to be constantly filled.
"There will always be corresponding solutions in the future, and other methods are currently needed, such as industry self-discipline and legal penalties, or requiring that AIGC output must have built-in special identification that can indicate its source and can be traced." Zheng Lei said.
"Counterfeiting is a practice that only a very small number of people benefit and most people suffer, and it is necessary to form a correct understanding of this problem in the whole society and improve people's ability to identify and prevent AI counterfeiting." Zheng Lei believes that "this issue is very important, if people are highly suspicious of AIGC, it will limit the healthy development of the industry." ”
AI algorithm expert and serial entrepreneur Huang Song told reporters that artificial intelligence algorithms continue to improve and upgrade, and more and more can "fake the real", but because these technologies have not been fully recognized by the public, many people do not understand that artificial intelligence has such capabilities. "It is precisely because of this cognitive error that some criminals find opportunities to commit fraud. Of course, other new technologies are also abused in the early stage, and all sectors of society should take active action to avoid or reduce such phenomena. Huang Song said.
Huang Song believes that technology practitioners, information content supervision departments, public security departments, and media should work together to formulate response strategies. At the legislative and law enforcement level, relevant personnel need to fully study the new situation and new criminal methods, and timely adjust and supplement relevant laws and regulations to deter lawbreakers; At the level of scientific and technological means, relevant personnel should actively invest in the development of tools and methods for detecting and limiting the abuse of AI technology, so as to suppress potential criminal space and provide detection tools and means; In terms of public opinion publicity, the society should actively publicize the rapid changes in new technologies, and at the same time remind the public of the risks and scams.
An Guangyong, an expert from the Credit Management Committee of the All-Union M&A Trade Union, Hong Yong, an expert at the think tank of the China Digital and Real Integration 50 Forum, Chen Jia, an independent international strategy researcher, and Zhang Xuefeng, a financial commentator, also expressed similar views to the Securities Times reporter, calling on all levels of society to work together to prevent AI risks and jointly promote the healthy development of artificial intelligence.