The annual World Artificial Intelligence Conference is regarded as the vane of the AI industry, and this year is particularly "hot". According to the statistics of the conference, the 2023 World Artificial Intelligence Conference has accumulated more than 1,400 guests to attend, with an exhibition area of more than 50,000 square meters and a total of 133 forums. As of 3 p.m. on the 8th, the number of offline visitors exceeded 177,000, the traffic of the whole network exceeded 1.07 billion, an increase of 68% over the previous session, and the exposure of the whole network reached 6.41 billion, both reaching a record high.
Looking back at the three-day 2023 World Artificial Intelligence Conference (WAIC 2023), whether at the opening ceremony, theme forums or nearly 100 industry forums, large models have become the focus of hot discussions among AI bigwigs, around the governance and landing of large models, Nandu reporters sorted out the following core trends.
1 Security and trustworthiness become the premise of AIGC
How to build safe and controllable trustworthy AI has become the basic premise and core support for the development of the artificial intelligence industry. On July 7, at the "Artificial Intelligence Innovation and Governance" forum of the World Artificial Intelligence Conference (WAIC) hosted by Southern Metropolis Daily and the Institute of International Governance of Artificial Intelligence of Tsinghua University, Zhou Bowen, Hui Yan Chair Professor of Tsinghua University, Tenured Professor of the Department of Electronics, Chief Expert of the Institute of International Governance of Artificial Intelligence and founder of Zhiyuan Technology, said that it is necessary to balance the landing of artificial intelligence technology and ethical governance in order to achieve long-term development. In Zhou Bowen's view, different countries have introduced AI governance policies, among which there are three common points that everyone is emphasizing, the first is to adhere to the parallel governance goals and development planning, and cannot only talk about governance without development. Second, the purpose of governance is to provide a safe harbor for technological innovation and development, and third, the governance model is becoming more and more market-oriented.
Zhou Bowen, Hui Yan Chair Professor of Tsinghua University, Tenured Professor of the Department of Electronics, Chief Expert of the Institute of International Governance of Artificial Intelligence, and founder of Zhiyuan Technology
At the "Trusted AI" forum held on the afternoon of July 7, He Jifeng, an academician of the Chinese Academy of Sciences, said that "AI can no longer be regarded as a simple technical work, but as an engine of change for the entire organization." Trusted artificial intelligence should be committed to ensuring data security and credibility, system behavior accountability, algorithm models explainable, network environment trust, legal ethics credibility and other issues. ”
During the World Artificial Intelligence Conference, Ant Group and Tsinghua University released the AI security detection platform "Ant Jian 2.0", and was selected as the "Treasure of the Town Hall" of this conference. According to reports, "Ant Jian 2.0" realizes the use of generative ability to detect generative models. The detection standard covers three categories: content security, data security, and ethical security, and can complete risk confrontation detection for large model generated content, including hundreds of dimensions such as personal privacy, ideology, illegal crime, prejudice and discrimination, and generate test reports to help large models optimize more targeted and continuously.
"Generative large model is a kind of 'deep black box' technology, 'Ant Jian 2.0' through intelligent game confrontation technology, simulation of black production and automatic generation of massive test sets, can realize the "generative robot" on the "AIGC generative model" induced detection and calculation, much like a 24-hour sleepless 'security hacker' looking for a large model, so as to find the weaknesses and security problems of the large model." This kind of AI-to-AI evaluation is also teaching. Just like generative examiners assess generative athletes, the two sides improve each other's abilities in the confrontation, further improving the AI safety level of the entire industry. Li Junkui, president of the technology department of Ant Group's large security business group, said.
It is worth noting that due to the current lack of trustworthy AI evaluation indicators, everyone has a "Hamlet" in their minds, resulting in a completely different definition of trustworthy AI. Under the paradigm of large models, how to clarify different dimensions, these trustworthy AIs have many challenges. In this regard, Zhou Bowen said that the landing of reliable AI requires cross-domain system collaboration. "The trustworthiness of trustworthy AI means throughout its entire life cycle, not just in the intermediate stage of algorithm development. From the user needs at the beginning of the design to the final product interaction. Due to the different concerns of all parties, there are large gaps to be explored in this area, which requires the joint efforts of academia, industry and government regulators. ”
As early as May 2022, Zhou Bowen founded and led the Collaborative Interactive Intelligence Research Center of the Department of Electronics of Tsinghua University. In his view, the core logic of the development of artificial intelligence and large models in the future is that AI will be more and more good at doing systematic work, and people need to help AI systematize on the track we want from the perspective of values and governance. The collaboration between AI and people will produce more high-value scenarios.
2 major model national standards launched, 6 enterprises entered the "national team"
With the increasing attention of society to the potential risks of AIGC, the large model has also ushered in the launch of national standards. On the first day of the 2023 World Artificial Intelligence Conference (WAIC 2023), the National Artificial Intelligence Standardization Group guided by the National Standards Commission announced the leader of the mainland's first large model standardization task force, which was jointly served by Shanghai Artificial Intelligence Laboratory and six enterprises including Baidu, Huawei, Alibaba, iFLYTEK, 360, and China Mobile.
In May this year, the state officially launched the standardization of large models, set up a large model thematic group in the overall group of artificial intelligence standardization, and began the declaration of the leader and deputy leader of the large model special group and the proposal of large model standardization research topics. It is revealed that the large model task force will undertake the standardization of large models, and the ultimate purpose is to promote the practical combination of large model technology and standardization and promote the healthy development of the artificial intelligence industry.
According to the relevant application requirements, the leader of the large model task force needs to have a high professional level in the theoretical research and industrial application of the large model, and be familiar with the international frontier scientific and technological development trend, regulations, systems and standards in the field of large model. After the establishment of the "national team", the joint leader unit needs to give full play to its theoretical and practical experience in the field of large models, and provide standardized research topic selection suggestions around prominent issues in key areas such as software and hardware bases, key technologies, and industrial applications of large models.
3 Internet giants aim at "service" and strive to be "model factories"
During the World Artificial Intelligence Conference, Zhou Jingren, CTO of Alibaba Cloud, said in his speech: "Alibaba Cloud will take promoting the prosperity of China's big model ecosystem as the primary goal, provide a full range of services to large model startups, including the most powerful intelligent computing power and development tools, and provide full support in terms of funding and commercial exploration." According to reports, Alibaba Cloud Modai Community is positioned in the "big model free market" and has gathered 1.8 million AI developers and more than 900 high-quality AI models contributed by more than 20 top AI institutions.
In 2022, Alibaba Cloud proposed the concept of MaaS (Model as a Service), proposed a development paradigm with AI models as the core, and built a set of cloud computing technology and service architecture with AI models as the core, which will all be open to large-model startups and developers, providing a full range of services including model training, inference, deployment, fine tuning, evaluation, and productization implementation.
"We want to provide a series of technical services for developers, model users, model creators, etc., which is the role of Alibaba Cloud in the MaaS ecosystem." Zhou Jingren said in an interview with Nandu that the positioning of the Tongyi big model is to serve enterprises that do not have the ability to develop models, hoping to achieve AI inclusiveness, so that enterprises can choose not to do model development to directly solve business problems, for enterprises that have the ability to do self-developed models, they can choose Alibaba Cloud's computing power platform as technical support to develop models, in the MaaS ecosystem, different models have their own strength, and ultimately model users make choices.
Also targeting MaaS are Tencent and ByteDance. On June 19, Tencent Cloud announced the R&D progress of industry large models for the first time, relying on the Tencent Cloud TI platform to build a selected store for large models in the industry, providing customers with one-stop MaaS (Model-as-a-Service) services to help customers build exclusive large models and intelligent applications. At present, Tencent Cloud has provided more than 50 large-model industry solutions for more than 10 industries such as media, cultural tourism, government affairs, and finance.
On June 28, Volcano Engine released the large model service platform "Volcano Ark", which provides enterprises with a full range of platform services (MaaS, that is, Model-as-a-Service) such as model tuning, evaluation, and inference. At present, "Volcano Ark" integrates the big models of many AI technology companies and research institutes such as Baichuan Intelligence, Mos, Moss of Fudan University, IDEA Research Institute, Lanzhou Technology, MiniMax, Zhipu AI, etc., and has launched the invitation for testing.
As early as March this year, when Wen Xin's words were released, Li Yanhong asserted that the era of big models will produce three major industrial opportunities: the first is new cloud computing companies, whose mainstream business models have changed from IaaS to MaaS. The second category is companies that carry out industry model fine-tuning, which is the middle layer between the general big model and the enterprise, they have the industry Konwhow, call the general large model ability, and provide solutions for industry customers. The third category is companies that develop applications based on large model bases, that is, application service providers.
4 Large models in the industry are blooming, and finance and medical care are popular application scenarios
During this World Artificial Intelligence Conference, industry large models have become a hot topic in dozens of forums, and the industry generally believes that moving from ridicule entertainment to industry applications is a key step in the development process of artificial intelligence large models.
On July 6, Tang Daosheng, senior executive vice president of Tencent Group and CEO of the Cloud and Smart Industry Business Group, said at the plenary session of the conference - the Industry Development Forum that the general big model has strong capabilities, but it cannot solve the specific problems of many enterprises. The application of large models in enterprises needs to comprehensively consider factors such as industry expertise, data security, continuous iteration, and comprehensive cost. Building your own exclusive model based on the large model of the industry may be a better option for enterprises.
Zhou Ming, founder and CEO of Lanzhou Technology, also believes that the big model has brought about the leapfrog development of cognitive intelligence technology, but the big model cannot be addicted to letting it tell jokes, and then share and entertain the public online, but to serve thousands of industries. "The general model is equivalent to high school graduates, the industry model is equivalent to college graduation, and the model that can solve special scenarios is equivalent to graduate school, it has to be built layer by layer, and only after adding the knowledge graph of industry scenarios can more rigorous content be output." Zhou Ming said to the Nandu reporter.
Zhang Yaqin, academician of the Chinese Academy of Engineering, chair professor at Tsinghua University, and dean of the Intelligent Industry Research Institute (AIR), also emphasized the importance of industry vertical models in the forum during the 2023 World Artificial Intelligence Conference. Zhang Yaqin said that in practical applications, although large models are important, it is necessary to develop corresponding vertical and accurate models according to different industries and scenarios. For example, in the field of unmanned driving, accurate models with low latency and high safety are required, and models that are good at writing poetry and painting are not needed. In the biological field, it is necessary to focus on developing models that can handle biological languages such as protein structure, rather than models that contain other redundant information on the Internet. Only in this way can the efficiency and accuracy of the model be improved and meet the needs and expectations of users.
At the same time, HUAWEI CLOUD held the summit forum "Pangu Big Model, Born for the Industry" at the developer conference, continuing the concept of AI for industries.
In terms of industry selection, Tencent has targeted the financial sector. During the World Artificial Intelligence Conference, Tencent upgraded the application of industry models in the fields of financial risk control, interactive translation, and digital sapiens. Among them, the financial risk control solution supported by the industry model has been 10 times more efficient than before, and through Tencent's accumulation of more than 20 years of black and gray industry confrontation experience and thousands of real business scenarios, the overall anti-fraud effect is about 20% higher than that of the traditional model. Based on the PROMPT model, enterprises can iterate risk control capabilities, from sample collection, model training to deployment and launch, achieving zero manual participation in the whole process, and reducing the modeling time from 2 weeks to only 2 days. Even with limited sample accumulation, quick setup can be done without skipping the "cold start" process.
During the World Artificial Intelligence Conference, SenseTime also highlighted and demonstrated the application practices of its large-model technology with all parties in the industry since its official release, including smart automobiles, finance, medical care, e-commerce, mobile terminals, industrial parks and other industries.
In the financial field, SenseTime cooperates with banks, insurance, securities firms and other customers, uses digital humans for intelligent customer service, smart marketing, etc., and provides new functions such as investment research analysis and research report writing by accessing the ability of large language models to achieve cost reduction and efficiency increase. In addition, after mounting the financial knowledge base, it can also output content Q&A based on the customer's product description 100%, and realize timely update of information. In the medical scenario, SenseTime has built a Chinese medical language model "Big Doctor" based on massive medical knowledge and clinical data, providing multi-scenario and multi-round conversation capabilities such as guidance, consultation, health consultation, and decision-making assistance, and will soon support multi-modal comprehensive analysis of medical images, text, structured data, etc., and continuously improve medical language understanding and reasoning capabilities, and continue to empower the rate of hospital diagnosis and treatment and patient service improvement.
"The breakthrough of large models has set off a new round of technological revolution in artificial intelligence, followed by explosive growth in industrial demand, and new application scenarios and application models are rapidly emerging." Xu Li, Chairman and CEO of SenseTime, said, "SenseTime hopes to continue to promote the improvement of AI infrastructure capabilities, not only to create a basic model with more powerful general capabilities, but also to further efficiently integrate the expertise of different vertical fields to build a professional large model that understands the industry and has more expertise, fundamentally reducing the downstream application cost and threshold of the large model, and allowing the industrial value of the large model to bloom in thousands of industries." ”
In addition, in the field of business information query, Tianyancha and HUAWEI CLOUD jointly explored the pendant trusted large model, and launched a trusted business assistant "Tianyan Sister" based on this technical solution.
Comparison of query results of "Tianyan Sister" with other open source big models.
Unlike general vertical knowledge, data and information in the business field are constantly changing. An enterprise's shareholding structure, business scope, financial strength, and operational risks may change from moment to moment, and further instructions need to be formed through complex calculation and judgment rules in combination with commercial practices and laws and regulations. The large model generated by corpus-based training, because it is still a probabilistic generative model in nature, cannot find the latest and most accurate answer in its own knowledge system, nor can it simply access the public data of the Internet to obtain conclusions. In addition, when the user's problem requires complex reasoning, the probabilistic nature of the large model leads to different results each time. This leads to repeated errors in large models when answering business questions, which can neither guarantee the reliability of the information provided nor the consistency of the same question when asked by different users.
The development of "Sky Eye Girl" is precisely to solve these problems. According to Liu Chao, founder of Tianyancha, "Tianyan Sister" has three advantages: trusted data, trustworthy algorithms and natural interaction. At present, the product has completed internal testing, and it is expected that the trial will be open to users by invitation by the end of this month.
Written by: Nandu reporter Ma Ningning from Shanghai