"The first few months everyone was talking about OpenAI and when China would be able to make its own big model, and in these months big models are coming out all over the place, and I see a lot of investors are starting to get anxious." Zhou mentioned at the 2023 Global Digital Economy Conference Artificial Intelligence Summit, and he argued in his speech that big models still have opportunities and in China, on the industrial side.

Zhou believes that the big model is not a so-called bubble, but a new industrial revolution. The big model cannot be compared to an operating system, but like a database, it becomes standard for every digital system.

To develop big model in China, we should ride on the wind turbine of industrial digitization.

"The real opportunity of big model in China, I think, is in the enterprise market, including the government and cities, and China should seize the opportunity of industrial Internet development most to do big model." Zhou Hongyi said in his speech.

Here is the transcript of Zhou Hongyi's speech.

Building a "secure, trusted, controllable and easy-to-use" enterprise-level AI model

Dear leaders and guests, good afternoon! Due to time constraint, I will share some of the application prospects of AI big models.

In the past few months, we have been discussing OpenAI and when China will be able to make its own big models, and in the past few months, there have been a lot of big models, and I see that many investors are getting anxious. Two days ago, Zhu Xiaohu and Fu Sheng had a fight online, and the topic was where are the business scenarios for big models for many entrepreneurs and companies?

After listening to Jiang's speech, I feel that this page is redundant, but I have to tell people when we meet outside, many people do not believe that the big model is true AI, strong AI or general AI, these are the premise of all our discussions. What makes me very emotional is that, for the first time, humans have allowed computers to understand human language in its entirety, and understanding language allows for understanding the knowledge and models of the world.

Some people think that the big model is not a windfall and bubble like the popular concept before? I think it should not be, but a new industrial revolution. Big Model directly improves the labor productivity of every person and every organization, and at the same time achieves a kind of general AI capability empowerment. Think about it, when the computer was first invented also did not drive the industrial revolution, at first only some physics research, weather forecasting, high precision fields to use the computer, far from ordinary people, ordinary enterprises, and then the PC into the home and business to change the world. The industry should strive for this goal, only the big model into thousands of households, empowering a hundred lines and a thousand industries, in order to truly promote this revolution brought about by artificial intelligence.

Some people think that if we imitate OpenAI, there may be one or two big models in China and the whole world in the future, so we compare big models to operating systems, and most companies may not have a chance. I think the future trend of big models will not be just one, but will become standard in every digital system, like databases, from small deployments on cell phones, to medium-sized deployments in cars, and of course, within enterprises and governments.

One of the important strategies of the country at present is industrial digitization, which is to use digital technology to help traditional industries, especially manufacturing industries, to empower transformation. To develop a big model in China, we should follow the trend and empower the industry digitally. Recently, I often traveled and communicated with many provincial and municipal officials, and everyone's perception is the same, the real opportunity of big model is in China, I think it is in the enterprise market, including the government and cities, and China should seize the opportunity of industrial Internet development to do big model most.

But when we go to the government, cities and enterprises with generic big model APIs, we will find that there are many problems with just having big models and using them directly. Just having big models is generalist, but it lacks industry depth. chatGPT was amazing at the beginning, and it felt like it could answer everything, but if really industry experts ask very deep questions, its ability is limited. The really valuable knowledge inside the industry and inside the enterprise is not available in public channels, and there is no way to meet the requirements of vertical expertise of the industry and enterprise scenarios with big models alone.

President Zhu Jun just said that there are many data security risks in the public large model. Each enterprise's own internal Knowhow is the core asset, certainly not trained to the public domain large model. The use of public big model will have the problem of data leakage, because many ideas and plans have to be told to it for it to write a good article. The public big model is generative AI, and the characteristic of the algorithm itself is that it will talk nonsense, and it will talk nonsense in a serious and justified way. This feature is used to make novels, write scripts, and as a chatbot to make fun of the effect is very good, but put in the enterprise scenario is a big problem, if really take the medical big model prescription, you dare not believe? Dare to eat? The public large model can not achieve cost control, because the high cost of the large model is also rendered very powerful, training once to ten million dollars, how much computing power, how many graphics cards are needed, the enterprise internal do not have to pursue the comprehensive knowledge of the large vertical model, but also do not have to pursue the ability to do everything, ten billion models may be enough, 100 billion to ten billion seems to be a small parameter to do ten times, saving training and deployment costs can be much more than Ten times. Therefore, for many enterprises, the cost of training enterprise-level large models has been dramatically reduced, and our goal is to pull the large models down from the altar and turn them into something that every enterprise and government department feels can be used directly.

In response to these problems, let's think about what kind of big model enterprises, governments and markets really need. It can be simply summarized as five.

Industrialization. There must be industry depth training data to have value.

Enterprise. Need to cooperate with the enterprise internal knowledge base, and to do real-time iterative updates, so as to ensure that the large model better understand the enterprise.

Verticalization. We should not try to use one big model to solve all the problems. The future landing form of big models in enterprises must be a combination of multiple vertical models, and vertical models are more capable of solving professional problems. Recently, there was a rumor that ChatGPT 4.0 was very powerful and might be a big model with super trillion parameters, so that the cost and fine-tuning of such a big model would be difficult and move the whole body with one hair.

The trend of Google Lamma's development direction is to desperately make the big model bigger, less than a trillion are embarrassed to say they are doing a big model, but there is another trend is to make the parameters of the big model architecture, training data set smaller. One of the latest Apple computers can operate a large-scale model of more than 30 billion. The future of small-scale large models may also be a trend, the future if a large model management of autonomous driving, intelligent cockpit, not in the cloud instantaneous response, there may be a large model architecture for each car.

Proprietary. The Chinese market will not be all public cloud market, public cloud will account for part of the market, many large central enterprises, state-owned enterprises, government agencies care a lot about Location, where the data is stored to look more important, so the proprietary deployment can ensure security and control.

A few days ago, I attended the World Internet Conference in Shandong, there are also three other principles proposed, that is, the industrialization program needs to follow the security, trustworthy, controllable.

Security. Just now Mr. Zhu Jun has mentioned, I will not mention more, the traditional network security, network attacks, vulnerabilities, algorithm security, data security problems can find ways to solve, the most terrible is to generate content is not safe. The earliest I also despise others to record the screen, each time the live demonstration I also apprehensive, do not know what the next sentence will answer the results. It is true that some people use AIGC to generate a variety of fake content fraud, this governance is much more complex than the management of the search engine public, so artificial intelligence security issues from now on to start research.

Insert an advertisement, 360 is also the first security company in China, we are also doing a big model to open the black box of artificial intelligence, so the big model is also the direction we want to work together with everyone to attack to ensure the bottom line of the development of artificial intelligence.

Trustworthy. People do not need to hit eighty percent of the model, the other twenty percent nonsense, need to be 100 percent accurate and precise. How to solve the problem of illusion? How to solve the problem of accurate output content? Corrections can now be made through vector databases, enterprise search, and external knowledge bases.

Controllable. The big model is still a bit unreliable at the moment, the government and the scenario use up smaller steps, do not come up with the main driver, plug-in and function model. Do not easily open the API to the big model, because it is only an assistant, and in the end, people have to make decisions.

Enterprise big model how to land to find the scene? We have to use the generic capabilities.

First of all, we have to use the capabilities that the big model is currently the best and most mature. Take the current big model as an example, generation and creation, code ability, logical reasoning, knowledge Q&A, summarize down, the best is actually two points: one is the Q&A dialogue, one is the writing generation. Government and enterprise use should first cut from a few mature angles, come up and enterprise business systems closely coupled together, into a very complex application, my experience is afraid that it is difficult to converge.

Secondly, from the pain point just need to choose the scene, small incision, large depth. Really use the big model, you will calculate every day Ou math problems, brain teasers? In fact, not, the real need of the scene is still related to most people, indicating that the office scene is just in demand. For employees there are enterprise knowledge search, knowledge management and training, for leaders there is information decision-making and opinion analysis, internally there is office generation office collaboration, and externally there is intelligent customer service to improve user experience. At present, the digitalization and intelligence of enterprises is not high, these scenes are the pain points of the enterprise office, but also the place where the big model can most improve efficiency.

The big model is not a panacea, some entrepreneurs are excited to say that with the big model, the ERP and CRM in the enterprise can be discarded, in fact, many of the core business of database storage can not be replaced, the big model can only play the auxiliary. Microsoft gives you a good example, all the scenarios are co-pilot, can give you navigation, give you advice, give you pointers, but will not grab the steering wheel indiscriminately. I think let the big model just start to turn on the co-pilot mode first, and the existing business system to maintain a relative degree of independence and isolation, so that the safety and control, safe landing, fast implementation. If you have to combine with the existing system, many units of the API may not be maintained by anyone.

AI pervasiveness. Can not only be used for the leaders, many business systems are given to the leaders to do a large screen display, which is of course also very important, but the real value of artificial intelligence is that everyone should be able to use, technology affirmative action. Many leaders asked me how in the end to use artificial intelligence empowerment? I said if every employee of the enterprise, every product manager, every business backbone are not familiar with artificial intelligence, have not used the big model, generate a lot of ideas is simply impractical. If an enterprise from the leadership to the middle and grassroots, all of them can simply get started with the basic functions of the big model of artificial intelligence, after six months, many people will gradually become familiar with the big model will take the initiative to think about how to combine with the business system, how to improve their work, as a two-way empowerment of people and the big model.

Big model development should be people-oriented, good and easy to use. There are always people rendering the application of big model to lay off staff, which makes many employees resist to big model. I strongly oppose this view, the big model at present want to complete a job independently or difficult, more positioning is a good tool for human, knowledge assistant, so the development of the big model to be people-oriented. But does the model really work well? You experts may feel that they write Prompt are very battlefield, but expect to become experts in Prompt, the big model used to eighty or ninety points is also very difficult. 360's proposal is to create a digital assistant with the soul on the basis of the big model with the enterprise, defined as digital employees, digital experts, digital consultants. Because dealing with generic dialog boxes is still not as good as chatting with various business-defined digital employees to meet daily habits, we further guide people to easily use the big model to complete their work by means of digital people. Digital people are not the same as the digital people who read a fixed script from beginning to end in the current online live broadcast, but behind them must be digital assistants driven by big models and with professional background and long-term accumulation. Who can define the digital person based on the big model well in the future may be the key to use the big model well within the enterprise.

Finally, to sum up: the development of big model should be really combined with national strategy, on the one hand, to develop core technology, on the other hand, to find various application scenarios. As an investor, a pure technology can create continuous business value only when combined with application scenarios. The big model has great potential in the process of digital transformation of cities, industries and enterprises to intelligence, and the evolutionary iteration has just begun, and I believe it will become the standard for digital systems in the future. The future may not be a hundred-model war, but a ten-thousand-model poor dance, and there are many opportunities for both To B, To G and SaaS-based enterprises. We will work with our ecological partners to create credible, controllable and secure large model solutions to escort the construction of a strong network and digital China.

"The first few months everyone was talking about OpenAI and when China would be able to make its own big model, and in these months big models are coming out all over the place, and I see a lot of investors are starting to get anxious." Zhou mentioned at the 2023 Global Digital Economy Conference Artificial Intelligence Summit, and he argued in his speech that big models still have opportunities and in China, on the industrial side.

Zhou believes that the big model is not a so-called bubble, but a new industrial revolution. The big model cannot be compared to an operating system, but like a database, it becomes standard for every digital system.

To develop big model in China, we should ride on the wind turbine of industrial digitization.

"The real opportunity of big model in China, I think, is in the enterprise market, including the government and cities, and China should seize the opportunity of industrial Internet development most to do big model." Zhou Hongyi said in his speech.

Here is the transcript of Zhou Hongyi's speech.

Building a "secure, trusted, controllable and easy-to-use" enterprise-level AI model

Dear leaders and guests, good afternoon! Due to time constraint, I will share some of the application prospects of AI big models.

In the past few months, we have been discussing OpenAI and when China will be able to make its own big models, and in the past few months, there have been a lot of big models, and I see that many investors are getting anxious. Two days ago, Zhu Xiaohu and Fu Sheng had a fight online, and the topic was where are the business scenarios for big models for many entrepreneurs and companies?

After listening to Jiang's speech, I feel that this page is redundant, but I have to tell people when we meet outside, many people do not believe that the big model is true AI, strong AI or general AI, these are the premise of all our discussions. What makes me very emotional is that, for the first time, humans have allowed computers to understand human language in its entirety, and understanding language allows for understanding the knowledge and models of the world.

Some people think that the big model is not a windfall and bubble like the popular concept before? I think it should not be, but a new industrial revolution. Big Model directly improves the labor productivity of every person and every organization, and at the same time achieves a kind of general AI capability empowerment. Think about it, when the computer was first invented also did not drive the industrial revolution, at first only some physics research, weather forecasting, high precision fields to use the computer, far from ordinary people, ordinary enterprises, and then the PC into the home and business to change the world. The industry should strive for this goal, only the big model into thousands of households, empowering a hundred lines and a thousand industries, in order to truly promote this revolution brought about by artificial intelligence.

Some people think that if we imitate OpenAI, there may be one or two big models in China and the whole world in the future, so we compare big models to operating systems, and most companies may not have a chance. I think the future trend of big models will not be just one, but will become standard in every digital system, like databases, from small deployments on cell phones, to medium-sized deployments in cars, and of course, within enterprises and governments.

One of the important strategies of the country at present is industrial digitization, which is to use digital technology to help traditional industries, especially manufacturing industries, to empower transformation. To develop a big model in China, we should follow the trend and empower the industry digitally. Recently, I often traveled and communicated with many provincial and municipal officials, and everyone's perception is the same, the real opportunity of big model is in China, I think it is in the enterprise market, including the government and cities, and China should seize the opportunity of industrial Internet development to do big model most.

But when we go to the government, cities and enterprises with generic big model APIs, we will find that there are many problems with just having big models and using them directly. Just having big models is generalist, but it lacks industry depth. chatGPT was amazing at the beginning, and it felt like it could answer everything, but if really industry experts ask very deep questions, its ability is limited. The really valuable knowledge inside the industry and inside the enterprise is not available in public channels, and there is no way to meet the requirements of vertical expertise of the industry and enterprise scenarios with big models alone.

President Zhu Jun just said that there are many data security risks in the public large model. Each enterprise's own internal Knowhow is the core asset, certainly not trained to the public domain large model. The use of public big model will have the problem of data leakage, because many ideas and plans have to be told to it for it to write a good article. The public big model is generative AI, and the characteristic of the algorithm itself is that it will talk nonsense, and it will talk nonsense in a serious and justified way. This feature is used to make novels, write scripts, and as a chatbot to make fun of the effect is very good, but put in the enterprise scenario is a big problem, if really take the medical big model prescription, you dare not believe? Dare to eat? The public large model can not achieve cost control, because the high cost of the large model is also rendered very powerful, training once to ten million dollars, how much computing power, how many graphics cards are needed, the enterprise internal do not have to pursue the comprehensive knowledge of the large vertical model, but also do not have to pursue the ability to do everything, ten billion models may be enough, 100 billion to ten billion seems to be a small parameter to do ten times, saving training and deployment costs can be much more than Ten times. Therefore, for many enterprises, the cost of training enterprise-level large models has been dramatically reduced, and our goal is to pull the large models down from the altar and turn them into something that every enterprise and government department feels can be used directly.

In response to these problems, let's think about what kind of big model enterprises, governments and markets really need. It can be simply summarized as five.

Industrialization. There must be industry depth training data to have value.

Enterprise. Need to cooperate with the enterprise internal knowledge base, and to do real-time iterative updates, so as to ensure that the large model better understand the enterprise.

Verticalization. We should not try to use one big model to solve all the problems. The future landing form of big models in enterprises must be a combination of multiple vertical models, and vertical models are more capable of solving professional problems. Recently, there was a rumor that ChatGPT 4.0 was very powerful and might be a big model with super trillion parameters, so that the cost and fine-tuning of such a big model would be difficult and move the whole body with one hair.

The trend of Google Lamma's development direction is to desperately make the big model bigger, less than a trillion are embarrassed to say they are doing a big model, but there is another trend is to make the parameters of the big model architecture, training data set smaller. One of the latest Apple computers can operate a large-scale model of more than 30 billion. The future of small-scale large models may also be a trend, the future if a large model management of autonomous driving, intelligent cockpit, not in the cloud instantaneous response, there may be a large model architecture for each car.

Proprietary. The Chinese market will not be all public cloud market, public cloud will account for part of the market, many large central enterprises, state-owned enterprises, government agencies care a lot about Location, where the data is stored to look more important, so the proprietary deployment can ensure security and control.

A few days ago, I attended the World Internet Conference in Shandong, there are also three other principles proposed, that is, the industrialization program needs to follow the security, trustworthy, controllable.

Security. Just now Mr. Zhu Jun has mentioned, I will not mention more, the traditional network security, network attacks, vulnerabilities, algorithm security, data security problems can find ways to solve, the most terrible is to generate content is not safe. The earliest I also despise others to record the screen, each time the live demonstration I also apprehensive, do not know what the next sentence will answer the results. It is true that some people use AIGC to generate a variety of fake content fraud, this governance is much more complex than the management of the search engine public, so artificial intelligence security issues from now on to start research.

Insert an advertisement, 360 is also the first security company in China, we are also doing a big model to open the black box of artificial intelligence, so the big model is also the direction we want to work together with everyone to attack to ensure the bottom line of the development of artificial intelligence.

Trustworthy. People do not need to hit eighty percent of the model, the other twenty percent nonsense, need to be 100 percent accurate and precise. How to solve the problem of illusion? How to solve the problem of accurate output content? Corrections can now be made through vector databases, enterprise search, and external knowledge bases.

Controllable. The big model is still a bit unreliable at the moment, the government and the scenario use up smaller steps, do not come up with the main driver, plug-in and function model. Do not easily open the API to the big model, because it is only an assistant, and in the end, people have to make decisions.

Enterprise big model how to land to find the scene? We have to use the generic capabilities.

First of all, we have to use the capabilities that the big model is currently the best and most mature. Take the current big model as an example, generation and creation, code ability, logical reasoning, knowledge Q&A, summarize down, the best is actually two points: one is the Q&A dialogue, one is the writing generation. Government and enterprise use should first cut from a few mature angles, come up and enterprise business systems closely coupled together, into a very complex application, my experience is afraid that it is difficult to converge.

Secondly, from the pain point just need to choose the scene, small incision, large depth. Really use the big model, you will calculate every day Ou math problems, brain teasers? In fact, not, the real need of the scene is still related to most people, indicating that the office scene is just in demand. For employees there are enterprise knowledge search, knowledge management and training, for leaders there is information decision-making and opinion analysis, internally there is office generation office collaboration, and externally there is intelligent customer service to improve user experience. At present, the digitalization and intelligence of enterprises is not high, these scenes are the pain points of the enterprise office, but also the place where the big model can most improve efficiency.

The big model is not a panacea, some entrepreneurs are excited to say that with the big model, the ERP and CRM in the enterprise can be discarded, in fact, many of the core business of database storage can not be replaced, the big model can only play the auxiliary. Microsoft gives you a good example, all the scenarios are co-pilot, can give you navigation, give you advice, give you pointers, but will not grab the steering wheel indiscriminately. I think let the big model just start to turn on the co-pilot mode first, and the existing business system to maintain a relative degree of independence and isolation, so that the safety and control, safe landing, fast implementation. If you have to combine with the existing system, many units of the API may not be maintained by anyone.

AI pervasiveness. Can not only be used for the leaders, many business systems are given to the leaders to do a large screen display, which is of course also very important, but the real value of artificial intelligence is that everyone should be able to use, technology affirmative action. Many leaders asked me how in the end to use artificial intelligence empowerment? I said if every employee of the enterprise, every product manager, every business backbone are not familiar with artificial intelligence, have not used the big model, generate a lot of ideas is simply impractical. If an enterprise from the leadership to the middle and grassroots, all of them can simply get started with the basic functions of the big model of artificial intelligence, after six months, many people will gradually become familiar with the big model will take the initiative to think about how to combine with the business system, how to improve their work, as a two-way empowerment of people and the big model.

Big model development should be people-oriented, good and easy to use. There are always people rendering the application of big model to lay off staff, which makes many employees resist to big model. I strongly oppose this view, the big model at present want to complete a job independently or difficult, more positioning is a good tool for human, knowledge assistant, so the development of the big model to be people-oriented. But does the model really work well? You experts may feel that they write Prompt are very battlefield, but expect to become experts in Prompt, the big model used to eighty or ninety points is also very difficult. 360's proposal is to create a digital assistant with the soul on the basis of the big model with the enterprise, defined as digital employees, digital experts, digital consultants. Because dealing with generic dialog boxes is still not as good as chatting with various business-defined digital employees to meet daily habits, we further guide people to easily use the big model to complete their work by means of digital people. Digital people are not the same as the digital people who read a fixed script from beginning to end in the current online live broadcast, but behind them must be digital assistants driven by big models and with professional background and long-term accumulation. Who can define the digital person based on the big model well in the future may be the key to use the big model well within the enterprise.

Finally, to sum up: the development of big model should be really combined with national strategy, on the one hand, to develop core technology, on the other hand, to find various application scenarios. As an investor, a pure technology can create continuous business value only when combined with application scenarios. The big model has great potential in the process of digital transformation of cities, industries and enterprises to intelligence, and the evolutionary iteration has just begun, and I believe it will become the standard for digital systems in the future. The future may not be a hundred-model war, but a ten-thousand-model poor dance, and there are many opportunities for both To B, To G and SaaS-based enterprises. We will work with our ecological partners to create credible, controllable and secure large model solutions to escort the construction of a strong network and digital China.