Zhou Hongyi: The development of large models must have AI beliefs

Innovation depends on the entrepreneurial spirit, and lying flat is contrary to the entrepreneurial spirit.

Text: "Chinese Entrepreneur" reporter Zhao Dongshan

Editor|Li Wei

Image source: meeting site

On December 9~10, the "21st Annual Meeting of Chinese Business Leaders" sponsored by China Entrepreneur magazine was grandly held. In the closing speech session held on December 10, Zhou Hongyi, founder, chairman and CEO of 360 Company, made a keynote speech.

The key takeaways are as follows:

1. Don't overestimate the capabilities of the current large model, but don't underestimate the potential of the future development of the large model.

2. All the entrepreneurs who come today are very smart, but we can't connect our brains together for superposition today, and large models can, as long as you have the money to buy more graphics cards, chips, memory, and computing power from NVIDIA, its capabilities can continue to improve.

3. The large model is no less than the birth of the Internet in 1995 and the emergence of personal computers in 1982, so American investors only look at AI-related projects. For companies that don't have AI concepts, AI functions, and AI components, they don't look at it at all.

4. China's opportunity is the industrialization, industrialization, verticalization, and in-depth customization of large models.

5. Establish AI faith: Believe that AI is true AI, believe that AI is an industrial revolution-level technology, believe that AI will reshape all businesses, and believe that companies and individuals who do not embrace AI will be eliminated.

6. Three questions to consider in AI: Upper and lower - whether everyone in the organization is using AI, internally, what happens if the internal business process is transformed, and externally, what happens when products and services are blessed by AI?

7. In the future, the company's prospects should be measured by "AI content": how many links in the business are optimized, empowered, and transformed by AI.

The following is a transcript of Zhou Hongyi's speech (with abridgement):

Long-termism

When it comes to entrepreneurship, first, we should always innovate and find new opportunities in the market through innovation; second, we should persist in the long-term principle of innovation.

The most innovative thing in 2023 is the breakthrough of artificial intelligence large models, which by no means belong to a few companies that do large models, but in various scenarios in the business of our entrepreneurs, and the combination of these scenarios and large models will bring many dividends in the future.

The theme of this conference is called "Tribute to Long-termism", why was the big model and OpenAI born in the United States? OpenAI uses open source technology, I am ashamed to say that many entrepreneurs in China, including me, are still a little lacking in long-termism, and we are more realistic to use artificial intelligence models to solve the problem of click-through rate of their own advertisements and solve many of their immediate problems.

OpenAI is truly long-term. They wonder whether human beings can make general-purpose artificial intelligence, and whether they can train human knowledge into a large model. In fact, this large model has been around for a long time, Google and Facebook have also existed, but no one has ever thought of training all human knowledge into a large model, and then they tried it, and the result was a miracle, using a very violent method to train a lot of human knowledge into a large model to produce GPT.

Although ChatGPT has been around for a year, there are still many people who do not believe that the large model is a real breakthrough and question that the large model is not real artificial intelligence. We have been doing search engines for 20 years and have done a lot of natural language processing work, and if you really use these large models, you will find that this time it is indeed a wolf, and it is indeed an inflection point for general artificial intelligence, because it is the first time that human beings have allowed computers to understand, store, and reason all human knowledge, and have a complete understanding of human language.

Why is language so important? Because the biggest difference between humans and animals is that we use language to describe the world, so once the machine understands human language and can talk to humans freely, it means that the machine also understands the model of the world, and many problems will be solved on this basis.

Before this breakthrough, what we did in artificial intelligence was not called artificial intelligence, it was called artificial intellectual disability, and many people suffered from it, for example, if you compete with the smart speaker at home or Siri in the Apple mobile phone for a long time, it will not understand if it is slightly complicated, because there is not a lot of prior knowledge reserves. This big model really achieved this breakthrough, and I think this is a successful example of long-termism.

Don't underestimate the potential of the future development of large models

Many people have gone to the other extreme, saying that we should make a large model of pig breeding, and everyone is full of infinite yearning and unrealistic worship for the large model. I don't think we should overestimate the capabilities of the big models now, but we shouldn't underestimate the potential of the big models for the future. The breakthrough of the technical route of the large model is only a few years, and I think there are still many shortcomings, such as the lack of knowledge in many fields, the most typical is that there is often nonsense, in fact, the biggest characteristic of people is also nonsense.

I hope you have a correct understanding of the large model, and now you can start to combine it with your business, but it is not yet completely comprehensive to take over the business, we still have to make use of the strengths and avoid the weaknesses, and give full play to the strengths of the large model. Human beings have probably spent 60 years studying artificial intelligence in the 60s of the last century, and the actual curve is very flat, but once it reaches the inflection point of the breakthrough in 2023, the future development is exponential.

All the entrepreneurs who come today are very smart, but we don't have a way to connect our brains together for superposition today, and large models can, as long as you have the money to buy more graphics cards, chips, memory, and computing power from NVIDIA, its capabilities can continue to improve.

For example, Xi if you use books as a metaphor, there are about 100 million kinds of books, and now the large model should have Xi learned 50 million to 80 million books, so many people are worried that the large model will soon have no books to read, how to surpass human wisdom?

Recently, OpenAI has been fighting in the palace, and they have said internally that they have made a breakthrough, like AlphaGo, they play chess with themselves, fight each other left and right every day, and make rapid progress. ChatGPT has invented a method where artificial intelligence generates content to train artificial intelligence, and if this is true, it is self-produced and self-sold, and it produces its own information.

What is the biggest breakthrough of Google Gemini recently? What impressed me the most is the ability of multimodality, it can not only read books, but has become obsessed with watching short videos, can understand pictures, listen to sounds, and see movies. Google has YouTube in its hands, and now the younger generation of children may no longer read books and directly watch videos to learn Xi knowledge. It has the ability to multimodal, Xi the movies made by humans in the video, all the videos, and our future cameras will be taken over by it, and we will meet here every day, and the machine can Xi.

Imagine today's large model, whether it is Gemini or GPT-4, I think it has surpassed our human individual, and the breadth and depth of his knowledge have surpassed that of the individual, and I think it is a matter of time before the next step to further improve his ability.

Three directions for the development of large-scale model technology

After the large model understands natural language, the following important ability is multimodality. I've just talked about the direct processing and production of pictures, sounds, videos, and music. Why do I talk about three directions? Many people always think that big models can only play a role in language, and language is to generate web pages, search engines, but this is a misconception.

First of all, after helping the machine understand our world, it will bring a huge boost to the work of the robot, because in the past, you couldn't train the robot, for example, you told the robot, go to the first row and bring me a cup, when the robot does not have a lot of knowledge, it does not understand what the first row is, it does not understand what a cup is, there is no way to perform this operation.

Second, there has been no breakthrough in intelligent driving, because everyone has been arguing about how many cameras and lidars to install in autonomous driving, or to work at the perception level, but when people are driving, they are at the cognitive level, and when a box appears in front of me, I will not hesitate to hit it, and there is a stroller in front of me, so I have to stop the car. Therefore, in the next year or two, we can see that automatic driving with the blessing of large models will make human autonomous driving truly achieve a breakthrough, which is why Elon Musk said on the one hand that everyone should stop doing artificial intelligence, and on the other hand, he bought 10,000 cards to make an artificial intelligence company.

Third, the important development direction of the final large model is to become a tool for scientific research, whether in biology, chemistry, new matter, or physics, it will become a good assistant for human scientists, because only when human beings make breakthroughs in basic science can we go to the sea of stars.

Recently, I took advantage of the east wind of APEC to run around the United States twice, I haven't been there for four years, and I have seen many investors and entrepreneurs, in fact, investors think that this is no less than the birth of the Internet in 1995, and no less than the emergence of personal computers in 1982, so American investors are now investing in non-AI projects, and all entrepreneurs no matter what business they do, they must at least say that I use OpenAI's API blessing, and they must have AI content, and companies without AI content will think you are out.

I had dinner with a lot of classmates, many of whom have been in the United States for twenty or thirty years, working in some large companies in the United States, and they are now stocking up on Nvidia's graphics cards within their companies.

If the United States bets that artificial intelligence is a new revolution, many contradictions will be solved, and other countries will not compete with it in one dimension.

This reminds me of the Plaza Accord back then, when the United States and Japan were competing for the first place in the world economy, and the United States forced Japan to sign a Plaza Accord, and Japan collapsed. I don't know if the story is true, but what I see is that Japan has missed the opportunity to upgrade the PC and Internet industries. Therefore, I think that for our country, the biggest insecurity is not development.

Regarding the positioning of the large model, there is a methodology in China, which believes that the large model is the operating system, and there are only two or three sets in the future, which reminds me of IBM in the fifties and sixties of the last century, which said that the world needs 5 computers enough, and it turns out that today there is more than one computer per capita in the world.

I think the real development of large models is industrialization and verticalization, so large models will be ubiquitous and will become the standard configuration of digital systems.

Qualcomm and Apple have recently released a new CPU, which supports large models that can be deployed on computers, and I think that in two years, all connected cars will have a large model of the vehicle, so in the future, every larger company and larger government agency will have its own large model.

After having an open-source large model, the biggest challenge is no longer the research and development of the large model, the most important task is to find the scene of your own work, integrate it with the large model, find the appropriate data, and how to train your own large model.

China's opportunity is the industrialization of large models

There are always people who are worried that large models threaten human beings, and my point of view is: first, the development of large models is far from reaching the time when they can threaten human beings; second, there are companies like 360 in the development process of large models, why do we want to make large models, that is, we must be determined to solve the security problems of large models; third, there is still a gap between China and the United States, including the gap in the capabilities, computing power and cognition of large models.

Today, the ability of domestic large models has reached the level of about 70 points of GPT-4, and if you use your vertical field to cultivate a vertical expert, this ability is enough. Therefore, the opportunity for China is the industrialization, industrialization, verticalization, and in-depth customization of large models.

Today, China's entire national strategy is digital transformation, industrial digitalization, and industrial Internet, and we have the most complete industrial chain in the world, the most complete industrial categories, and we have the most enterprise scenario dividends.

At present, it is generally said that there is no way to start a financial model and an education model, and only by decomposing the financial scene into 50 goals and dividing the education scene into 100 subtle scenarios, can you make changes to the industry.

There are three companies in the United States that have made money on AI, namely Microsoft, Adobe and Salesforce, they have not used AI to do anything new, they have used AI in the business chain of their existing products to make the products upgraded.

AI Faith

There are still a lot of people who think of AI as a toy, but I think:

First, AI faith. Do you believe that strong AI is real AI? Do you believe that AI is an industrial revolution-level productivity tool? Do you believe that AI will reshape all your products and technologies? Companies and individuals who don't embrace AI may be eliminated by their peers who use AI in the next few years, so you will not be eliminated by AI, you will be eliminated by those who are good at using AI.

After having AI beliefs, all in AI, all in AI is not about how many graphics cards you buy, not how many AI experts you poach, not to mention how many AI experts you have dug up, not to mention that you throw all the company's money into AI, I think it is a kind of spiritual All in, that is, you should use your core team to think about a few questions: First, AI is different from many other digital technologies, not a thing only for the boss, but the enterprise from the inside out, Second, your internal business processes, which links can be transformed by AI, and if you don't, what will happen if your opponents do it.

For example, Salesforce is the world's largest customer management company, and it has done one thing quietly, and after refining a large model by itself, in its customer service system, it can help you write an email to a customer, and it has carried out AI optimization and empowerment in very detailed links.

I feel that after doing a lot of cases, don't think that doing AI is to make a new product with AI, and doing AI means that you are not creating new things now, but from the top to the bottom of your company's organizational structure, from the inside to the outside of your business process, your customers to your products, not necessarily a comprehensive transformation, not necessarily a large transformation, but to find some small incisions for noise reduction.

Finally, I propose a measurement indicator called AI content, that is, how many employees are familiar with AI, your products, and how many details of your business process can be blessed by AI.

For many large enterprises, there are three steps: first, deploy a privatized general model and use it for their own internal employees, second, find some special scenarios on the basis of the general model and train the vertical model, and third, combine the vertical model with the company's digital business through the agent framework.

I don't think entrepreneurship is just a slogan, and I suggest that all companies not only embrace AI, but also innovate through AI. If China wants to develop a large model, entrepreneurs are the main body of innovation, and innovation depends on the entrepreneurial spirit, which is contrary to the entrepreneurial spirit.

Entrepreneurs are different from businessmen, businessmen have the opportunity to do a hand, no chance to rest, entrepreneurs should be a little idealistic, and at the same time adhere to long-termism.