Shulou(Shulou.com)12/24 Report--

Introduction

A year on, the AGI craze triggered by ChatGPT continues unabated. It's just that there is more discussion about how to land than the popularity of the national model when it was first launched.

With the development of computing power, database, big data and other underlying technologies, the construction and application of large models in various fields are accelerating, so do these signs indicate that AGI is coming? What are the common performances of the most advanced large model technology? Back in China, what are the challenges facing the current application scenarios of the large model, and how should they be solved? What are the trends and opportunities that entrepreneurs should pay attention to?

To answer the above questions, recently, the second half of the Tencent Cloud TVP AI Innovation Seminar series, "the second half of AI, exploring Innovation and Application" was held on Tencent Binhai Mansion in Shenzhen, where a number of senior experts in the AI field were invited to share the cutting edge, and specially set up a deep grouping brainstorm for the guests, a collision of thoughts on the spot and a burst of wonderful ideas.

"I have been doing artificial intelligence for more than a decade, and I have deeply felt the rapid development of this technology in the past year. After GPT came out last year, I read the relevant papers and sorted out a lot of materials in June, and I felt that I had almost understood it. As a result, new innovations continue to emerge in the second half of the year, including various applications. I can't help but wonder, what is the limit of the big model?" Under the opening questions of the host, flickering CEO and Tencent Cloud TVP Jin Zhihui, the event officially opened.

AGI represented by large models: autonomous decision-making, self-exploration, self-iteration

"after ChatGPT came out, people found that it had passed the Turing test. So I wrote an article during the Spring Festival this year and made it clear that AGI is coming."

In his keynote speech on "the present and Future of General artificial Intelligence", Wang Wenguang, vice president of Analytic data and TVP of Tencent Cloud, first traced back to the continuation of the technology in large models from the initial introduction of the concept of neural network. Specifically, the current development of the large model mainly shows the following characteristics:

The diversification of models. During the two years from Google's release of Transformer in 2017 to 2018, representative models began to emerge in the industry. Among them, typical examples such as BERT and GPT,GPT appeared earlier than BERT, but in terms of language processing power, BERT first broke out. At that time, BERT surpassed the level of human experts for the first time in the evaluation of reading comprehension.

The outbreak of parameter scale and data scale. The excellent performance of BERT in "reading comprehension" opened the era of "barbaric growth" of model parameters, until today GPT-4 has reached a parameter scale of 1 trillion. At the same time, the data corpus has changed from hundreds of megabytes in the BERT period to several PB or even dozens of PB now.

Judging from the overall pattern of the development of the big model, the gathering place of the global big model is still in Silicon Valley, followed by China's "hundred model war". Europe obviously lags behind in this "battlefield" and there are few winners.

With regard to how the large model can be better applied in practice, Wang Wenguang particularly stressed that the large model has "hallucination" problems that cannot be solved by itself, resulting in no guarantee of accuracy and facts. Therefore, the use of it needs to be limited, which is very useful in the case of low reliability and authenticity. In view of the fact that the big model is not good at doing mathematical calculations, Wang Wenguang said: "my suggestion is that the results given by the big model can be run through the code interpreter, which is a feasible method."

According to its introduction, at present, the application of the large model is mainly concentrated in the fields related to text creation, including access to materials, marketing copywriting, and novel creation, mainly based on the strong ability of search and fabrication.

"but if we need to ensure reliability, especially if our outlook data model platform 'Cao Zhi' is oriented to industrial applications, we need to find ways to use retrieval enhancement and verify its results through the method of knowledge graph. do the credential of knowledge, the second verification, to do all the answers to the source."

Thinking about the future, Wang Wenguang believes that AGI, represented by large models, will further develop into a new era of "independent decision-making, independent exploration, and self-iteration." Compared with the difficulty of self-iteration, independent decision-making and self-exploration are being realized:

"there are only two kinds of 'output': one language and one tool. The latest development of the language is now the large language model, and the direction of training the large model with video is also worthy of attention. In the use of tools, there is no lack of Agents and tool call methods of application tools. We can also combine the large model as a brain with a robot, and the robot can operate according to instructions to achieve" embodied intelligence. "

In the more distant future, we may have to face the situation of "silicon machine day, carbon-based millennium". When the first real robot appeared in the world, although it was an artificial object, its level of intelligence was much higher than that of human beings. Human society is built on the recessive condition that "human is the only agent". "then, can the current social structure of human beings still exist? can we maintain the status quo? these problems will be left to scholars in social science, philosophy and other fields to study." Wang Wenguang said finally.

The key elements of AI's successful landing under the condition of limited capital patience

"whether it is career, technology, or life, it is an S-shaped curve, and it is impossible to grow smoothly, fall to the bottom all the time, and always follow the development path of the S-curve. The same is true of the development of AI." Shao Hao, a senior artificial intelligence expert and investment director of Chuangfang, Tencent Cloud TVP, said in a keynote speech on "AI Venture track Trends and Investment and financing".

From the introduction of the concept of artificial intelligence in 1956 to AlexNet winning the championship of ImageNet in 2012, it seems that AI has once again ushered in a new industry. Between 2013 and 2019, NLP, AR, VR and other technologies had a good momentum of development, and many concept companies received financing. But in fact, the industry is still in the hype cycle, until 2019, ushered in the first trough of this wave.

"We know that technological transformation requires long-term precipitation, but capital does not have much patience. In the United States, the patience of capital may be up to a decade, but in China, perhaps for up to seven years, shareholders need to see a return on investment."

According to the logic of capital operation, when the project is difficult to achieve the expected return in the short term, it is difficult to avoid the fate of being closed. Based on Shao Hao's personal experience in the field of AI investment, he believes that it is very difficult to choose the investment target of the big model.

So, in the case of limited capital patience, what are the key elements for AI technology to successfully land?

First of all, it is to achieve efficient application. In Shao Hao's view, whether the large language model is applied to office software, or the search ability of AI is applied to the crawling and recommendation of information, it is an excellent application case of AI in the popularization scenario. You also need to find pain points in professional areas, such as making major discoveries and finding new targets through big data's help.

"whether it is a good application or not, there are probably several factors to judge: the industry is labor-intensive, enterprise production tools, universal products, and terminal scenarios include military industry, medicine, banks, securities firms, and so on. It is easier to make sense by including these key points."

Secondly, it is necessary to form a complete industrial closed loop. For example, in the field of semiconductors, we need to form a market-oriented industrial chain from raw materials, equipment, talents, matching, production and sales. The same is true in the field of AI, if it is difficult to form a market closed loop, it is difficult to deliver the products that users really like.

On this basis, Shao Hao shared the key factors to be considered in AI investment: "take our company's' Shenzhen Oriental Investment'as an example, focusing on seven words' Guangzhou-Shenzhen high-speed travel', including the breadth of market-oriented, the depth of product innovation, high team configuration, fast growth, as well as the rationality of entry and exit prices and the feasibility of investment."

In addition, Shao Hao also gives him two suggestions for AI entrepreneurs based on his experience in starting a business:

First, AI determines the technology limit, but don't pursue it blindly. Based on the complexity and high-cost addition of AI, technology replacement can be carried out in the early stage of the project, and if it goes well, AI will be introduced in the follow-up process, which can reduce investors' risk expectations; second, in addition to technology, how new products should replace the old products in the market is also worth thinking more.

Finally, based on the prediction of the future technological development of the large model, Shao Hao put forward some specific suggestions:

Focus on the bottleneck elements and domain applications in the development of AI, such as computing power and industry data

In the application of specific fields, there are more opportunities in medical and military fields.

Choose "+ AI" projects instead of "AI+" projects guided by user needs and industry needs

Maintain a forward-looking vision, such as focusing on multimodal and other cutting-edge directions

Emphasize "industrial closed loop".

Vector database: realizing multimodal data access and human-to-data interaction

"the essence of the big language model is to shift the governance paradigm, and its function is mainly reflected in that even non-programmers can easily schedule the GPU computing power hidden behind the big language model through natural language."

In the keynote speech "Vector Database: data Hub in the AI era", Luo Yun, Deputy General Manager of Tencent Cloud Database, first put forward his views on the large model.

Furthermore, matched with strong computing power, there is a perfect data storage platform. Through the internal practice of Tencent Cloud, Luo Yun believes that the following two key points need to be tackled in building a storage platform:

Multimodal data. Including relational database data, data in the file system, and stored unstructured data, these data are value data and need to be used in an intelligent way. Among them, the key lies in the multi-modal interaction of data, to achieve the correspondence and circulation of data in different modes, such as table data, KV data, picture data and so on.

The interaction between people and data based on natural language. How to schedule the underlying storage without going through the programmer, including reading and writing data, retrieving data and other tasks.

To achieve the above two points, you need to use the "hub" and "index" function of the vector database:

Vector database is the data center. The key point of the information difference between the data lies in the inconsistency of the format, and through the "vector" way, the multimodal data can be finally expressed in a format.

The vector database is located at the index layer. Through the vector database, we can index the data in different data spaces, including database, big data, file system and so on. It is a general index layer and is easy to find.

In addition, when predicting that vector databases may face the challenge of achieving enterprise-level and intelligence, Tencent Cloud has also made a number of landmark attempts in the industry, including:

Together with the standard-setting organization of China Institute of Information and Communication, we have completed hundreds of billions of tests. " This includes achieving a maximum of 100 billion vector scale and a peak capacity of 5 million QPS, achieving 99.99% availability; we are also the first vector database in China to pass the standard test of the Institute of Information and Communication; at the same time, together with the Institute of Information and Communication, more than 50 enterprises jointly issued the first vector database standard "Vector Database Technical requirements" in China. "

The end-to-end recall rate has increased by more than 30%. Through the integration of Embedding, natural language query is realized; through the application of AI suite, end-to-end RAG application retrieval is realized, thus the recall rate is increased by more than 30%.

"usually people will use the framework to do it, but when we test it, we find that the accuracy rate is not very high. Basically, it is built through open source solutions, and the end-to-end recall rate is 50% even high. In order to improve the recall rate, we have also made many attempts internally, such as working with the PCG business team to platform some of their knowledge and apply standard solutions to the industry, so that the industry can grow faster. If you do it in this way, the recall rate can reach 700.80%. "

In fact, the data performance of Tencent Cloud vector database has been bright and growing rapidly since the middle of this year, but Rome was not built in a day. According to Luo Yun, as early as 2015, Tencent Group carried out a multi-party layout of Tencent Cloud vector database. For example, it is used in Tencent Video and Tencent News to ensure the reliability of content uploaded by users, that is, the content of UGC. "our method is to convert news and video clips into vectors and save them, and then do vector matching by checking duplicates and some sensitive data that is not allowed, so that we can quickly review the content materials."

So far, Tencent Cloud vector database has access to more than 40 businesses within the group, completing 160 billion requests per day. In addition, since July this year, the cloud on the vector database has accumulated more than 1500 enterprise users and developer users in about 3 months, with a rapid growth.

Grouping brain storm, point of view collision

After the wonderful speech of the three guests, the host CEO, Tencent Cloud TVP Jin Zhihui concluded: "I very much agree with the road to AGI, and the topic of AGI is also very attractive. If you are interested in this topic, I recommend you to read three books: life 3.0, A brief History of the Future and Thousand brains, which basically gives a good interpretation of the development trend of human AGI in the future."

Immediately after that, the in-depth discussion session of everyone's participation officially opened, and the host CEO flickering, Tencent Cloud TVP Jin Zhihui raised four hot issues about large models and AI technology. The guests at the scene chose different topics for discussion in the form of a group lottery, and then each group sent representatives to speak, and other group members could also express their views pertinently. After the hot discussion, the controversial topic "opinion competition" was specially set up, and the guests also collided with more wonderful ideas in the process of exchange.

Hot discussion

What opportunities will 1:GPTs bring to entrepreneurs and how will it evolve in the future?

Yang Fangxian, founder of 53AI and TVP of Tencent Cloud, said: "I have thought very deeply about this topic in the past few months. After a comprehensive discussion in our live group, the view is that GPT-4 is very valuable to the development of the industry, just like the Internet product" personal home page system "in 2000. Ordinary users can simply DIY their own website through the home page system." Today, the earliest individual stationmaster at that time has become the main force of Internet talents in China today.

From this point of view, GPTs's strong choreography ability and extremely low threshold for use will greatly promote the popularity of large model applications and train a large number of talents for the large model industry. In the next ten years, there will be a large number of AI applications in both To C and To B scenarios, which will be developed independently and will not be built on GPTs.

Topic 2: what are the core directions for the future development of AI and big models?

The second group of speakers, professor of Peking University, director of Shenzhen system Chip Design key Laboratory of Peking University and Tencent Cloud TVP he Jin, put forward that after group discussion, we generally believe that this issue will not be particularly clear. From the direction of core technology, the five major directions of the intelligent era include: scheduling capacity of CPU, computing capacity of GPU, storage capacity of Memory, and transmission capacity of network. And big data.

Among them, the computing and storage capacity can be integrated through the integrated chip of storage and calculation. if the perception layer is added, the three-layer fusion of sense, storage and computing can be realized, and the network transmission capacity may be further integrated in the future. From the perspective of application direction, we are more optimistic about the future development of "AI agent".

What can be the combination of topic 3:AI, large model and industry or domain, and what kind of application scenarios do you like?

On this issue, Tang Dingqing, a representative of the 1994 intelligent partner, said that taking the intelligent outbound business as an example, before there was no large model, it mainly carried out some relatively single human-computer and user interaction, relying on the structure of the main process. When there is a large model, we can combine user information, as well as many rounds of dynamic information in the interactive process to make dynamic decisions, the core is to improve the conversion of AI certificates of deposit.

In addition, the guests of each group also expressed their own views according to the application scenarios in their own field.

Yang Zhentao, director of research and development of vivo and TVP of Tencent Cloud, pointed out that in the field of software development, the big model can allow more engineers to do higher value-added "blank questions" and bring out the creative parts that people are better at. You can leave it to AI Agent to process or find a job with Bug.

Host Huoguang swaying CEO and Tencent Cloud TVP Jin Zhihui pointed out some important scenes in NLP in combination with their language education field. First, children's language learning and adult language learning, there have been many start-ups using ChatGPT dialogue technology to produce the next generation of language training scenarios. In addition, in the future elderly society, how to use AI to provide companionship and emotional value to the elderly is also an applicable scenario.

Topic 4: will China gradually level or pull further away from the AI competition between China and the United States in the next decade?

On this question, the fourth group of speakers on behalf of Meituan, senior technical director of UAV, Tencent Cloud TVP Chen Tianjian replied: after discussion in the group, we believe that although there is a gap in the future, it will not be farther and farther away. First of all, although there is a gap in arithmetic, there are not many scenarios that require such powerful computing power in real AI applications. For example, for the two Chinese and American companies in the field of self-driving, the computing power of American companies is twice as high as that of Chinese companies, but it is mainly used in vision, while Chinese companies replace it through higher-quality sensors, so it can be seen that our country can find an advantage path in the field of AI product application; secondly, in terms of original technology, China and the United States still need to compete in the dimensions of R & D investment, talent and time.

When it comes to computing power, host CEO flickering, and Tencent Cloud TVP Jin Zhihui remain optimistic. He believes that China's current breakthroughs in computing research and development are very fast, and AI talents are also constantly catching up, and China will be able to keep up with the United States in the future development of large model technology.

Opinion Competition: "effective acceleration" or "Super alignment"?

Representatives of effective acceleration theory: Sam Altman, Yann Lecun, Andrew Ng (Wu Enda)

Core idea: technological development should not be limited by other factors, it should be more open, and AI should not be regulated.

Representatives of Super alignment Theory: Elon Musk, Hinton, Ilya Sutskever

Core point of view: based on the potential threat posed by technology, the behavior of the AI system should be guided so that it meets the designer's intention, that is, it needs to be regulated.

From the positive point of view, the main reason why we are in favor of "effective acceleration" is that all countries will eventually choose to regulate, but because AGI has a great impact on people, the results of regulation are still unknown. Wang Wenguang, vice president of Analysys data and TVP of Tencent Cloud, said: AGI is coming in the future and will have a great impact. On this point, the views of the two groups are the same, but Ilya hopes to open it up for everyone to use after allowing AGI to serve human beings, while Altman hopes to develop first and seize the opportunity first. With regard to the current gap between China and the United States in the development of AI, I think we should first support the development and catch up in technology.

The opposite side, from the point of view of human nature, believes that the lack of effective supervision and "effective acceleration" under governance will lead to the situation that "buttocks determine the head", or even "personal interests override collective interests". In this regard, Zhongke hydrogen Yan CDO and Tencent Cloud TVP Zeng Yonghong said: today's society ruled by law needs ethics and order. No matter how long it takes in the future, whether in the physical world or the virtual world, it will eventually be reflected in the physical world: human beings are responsible for rules and order, and silicon-based is responsible for implementation and enforcement. For example, whether the acquisition and use of data is in compliance, if the data trained by AI is not constrained, it will cause a situation out of control, and pollution before treatment may cause irreversible harm, so the predictive regulatory framework can better develop and make use of AI.

Yang Zhentao, R & D director of Zhengfang vivo and TVP of Tencent Cloud, further pointed out that this involves a lot of topics about values, and now many people will wear the label "e / acc" to show that they are supporters of "effective accelerationism." Some people may think that this represents a kind of elite arrogance, but in fact, in the stage of technological growth, on the eve of AGI breakthrough, if there is too much regulation, it will inevitably affect the development of technology itself. From another point of view, from the social and government level, we may need to consider the general environment and values, including the improvement of the law, the ethics of science and technology, and so on, all of which will be taken into account, but at the right time, we should not spray pesticides when the new technology seedling grows new leaves, or we should give it room to grow.

Anti-PP parking founder & CEO, Tencent Cloud TVP Li Jian once again stated his point of view: what I want to express is to think about this issue from the perspective of human development, when everything is out of control, where does order come from? human beings are not only in the current era, but in the 20, 30, 100 years are fleeting, my view is the need for regulation to ensure long-term order.

Finally, Luo Yun, deputy general manager of Tencent Cloud Database, reconciled the views of the two sides, raising the thinking of this topic to a philosophical level: in my own opinion, there is no answer to this topic. I believe no one will say that our human purpose is to make AI, and the purpose of carbon-based is to make silicon-based. Our ultimate goal of using AI must be to serve human beings, help us to increase productivity and make human beings happier. Well, under this great consensus of purpose, I think the current stage should be free to run forward and make bold attempts, which is more beneficial to the development of science and technology in our country.

Conclusion

In fact, AGI is first of all a technical proposition, but it is also a social proposition and a philosophical proposition. Therefore, in this TVP AI innovation seminar, both guest speeches and offline discussions will add social thinking and philosophical thinking to AGI on top of the technology, which is also the super-expected harvest of this event.

Adhering to the original intention of "using science and technology to influence the world", Tencent Cloud TVP will also join hands with experts from all walks of life to explore the theoretical frontiers and technical practices of large models in different dimensions and fields, and jointly move towards the intelligent future of AGI.

TVP, the most valuable expert (Tencent Cloud Valuable Professional) of Tencent Cloud, is an award awarded by Tencent Cloud to technical experts in the field of cloud computing. TVP strives to create a communication platform with industry technical experts and promote effective communication between Tencent Cloud and technical experts and users, so as to build cloud computing technology ecology and realize the beautiful vision of "influencing the world with science and technology".

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.