Dialogue with Zhipu AI CEO Zhang Peng: The technological revolution is already fast enough, don’t just focus on the results of “super applications”

At the 2024 World Artificial Intelligence Conference held by the Huangpu River, entrepreneurs, developers, and technology enthusiasts unleashed an enthusiasm that was even hotter than the 38-degree heat in Shanghai. Opinions about the revolutionary nature of large models and the incompatibility of practical applications also collided. Create new sparks.

How to convert the computing power of large models into productivity?

When large model, chip, cloud computing, embodied intelligence, autonomous driving and other manufacturers bring together the latest achievements, forming a miniature landscape of China's AI landscape, everyone who comes here hopes to find some grassroots leading to AGI. Snake gray line.

"This is the company that looks a lot like OpenAI." I was in front of the Zhipu AI booth and heard an exhibition visitor introduce it to his companions. This is probably the view of many practitioners.

This Chinese AI unicorn has general large models of various sizes. It recently released the GLM-4-9B model, surpassing the Llama 3 8b model. The multi-modal model GLM-4V-9B achieves the same capabilities as GPT-4V with a parameter amount of 13B. Yesterday, Zhipu AI also released the 4th generation CodeGeeX code model CodeGeeX4-ALL-9B at WAIC.

Zhipu AI's pace of commercialization is also ahead of most of its competitors. Today, there are more than 300,000 agents active in Qingyan APP available for use. The Zhipu AI large model open platform currently has more than 400,000 registered users, and the average daily call volume reaches 60 billion Tokens.

Zhang Peng, CEO of Zhipu AI, believes that the current AI craze caused by large models is different from before. In the past, AI technology solved some practical problems, but today's development of large models has brought more important human-like cognition. ability.

Large models can provide generalization capabilities on one model to solve the diverse needs of a series of scenarios and applications, thereby solving the problem of cost and benefit balance. This is its essential feature.

Zhang Peng also accepted interviews with APPSO and other media on the spot, talking about topics such as the implementation of large models, super applications and the future curve of technology, covering some key issues from large model research to commercialization.

The following is a transcript of the conversation. You can see some profiles of how Chinese AI companies implemented large models.

Big models are implemented, don’t just focus on super applications

Q: Regarding TPF (Technology-Problem Fit) for large models, how does Zhipu combine technology with products and then implement it? The industry has not yet reached a consensus on this.

Zhang Peng: I never think this matter needs to be controversial. The implementation of any new technology requires a cycle. This is a natural law, but this cycle may be long or short, and when a revolutionary technology like a large model is implemented, In the process, there will definitely be greater challenges and more problems that we need to solve.

From an objective historical perspective, the implementation of this technological revolution has been fast enough, but precisely because it is so fast, everyone's understanding of this matter is still somewhat uneven, and the variance of the distribution is relatively large.

Technology still needs to be iterated and updated quickly, but in terms of application, we cannot wait for it to be fully mature before implementing it .

Q: Can you share some experience in implementing large models?

Zhang Peng: First of all, you must have a deep understanding of the ability to recognize this model. You should try to grasp its advantages and avoid taking advantage of its shortcomings. For example, if you ask the model to calculate very accurate physical models or mathematical formulas, it is just like the human brain. It is not good at it, so you should not do it. Harsh.

For example, it is not appropriate to use it to replace the calculator that you are used to, so you have to find a suitable angle to give full play to its advantages, and do not go in the opposite direction, or block the space for its development path, as the model capabilities may be compromised at any time. The iterations were crushed .

Q: How far are we from the AI ​​super applications that everyone has been discussing recently? In fact, there are very few AI applications with more than 10 million daily active users.

Zhang Peng: Everyone’s definition of super application is very vague. Everyone has their own set of ideas. Is ChatGPT considered a super application?

Q: Let’s see what analogy it is compared to.

Zhang Peng : Yes, you always have to have an analogy. ChatGPT is already the fastest product in history to have over 100 million monthly users. If this cannot be called a super application, what can be called a super application?

Don’t just look at this conclusion prematurely. If you observe the process, you will find that it has developed very fast, so be patient. The emergence of super applications is not entirely a technology-driven thing, but also considers many factors such as the market. and whether the user is ready .

Let me give you another simple example. The number of users of the Google search engine is large enough. How long did it take for it to go from becoming the world’s number one search engine to finding a successful commercial path? 6 years. It also took 6 years for the current Meta and the original Facebook.

Q: It also took longer from the emergence of mobile Internet to WeChat and TikTok.

Zhang Peng : So why can’t everyone wait a little longer? Might as well give it a try. Just like when we played the game of Arkanoid when we were kids, you want to aim at the gap and hit it into a gap very accurately. To do this, you first have to find where the gap is? Where is the path?

Many things need to be explored one after another, and this process is very important. Don't just see the final results, but more importantly, we take action. I think this is what everyone should pay more attention to at the moment.

Q: Some people in the industry think that innovative applications may be implemented in the next few years. He gave a very clear time of about three years. What do you think?

Zhang Peng: Maybe tomorrow. As I just said, this matter requires comprehensive consideration of different factors. One is the maturity of the technology itself, and the second is whether the market and users themselves are ready. The third is the discovery of demand, and even a little luck is added. There are too many variables, and it is difficult to predict this kind of thing with a simple neural network like my brain.

Q: There is a hotly debated opinion recently, saying that a basic model without application is worthless.

Zhang Peng : This matter itself is divided into two levels. First, technological innovation is meaningful in itself, because we have so many scientific researchers who are constantly exploring what causes human intelligence. We How to make machines approach human intelligence is of great significance.

The second level. If the result of this exploration is that we can engineer and productize it, turning it into a more valuable productive force, it will be of greater significance .

This matter is not a choice between two, but a series of issues. Application is of course very important. We do hope that technology can be transformed into more new productivity today, but it does not mean that our pursuit of technological innovation and essential exploration is worthless. Don't go to either extreme, they are mutually reinforcing relationships.

Q: There is also a view that the open source model is not suitable for most application scenarios, and the commercial closed source model is the most effective. Some time ago, Zhipu also released the open source version of GLM-4. What do you think about the issue of open source and closed source models?

Zhang Peng : We have always believed that open source and closed source have different essential goals and meanings. Closed source is considered more from a commercial perspective. It is a business path to provide better services and safer products. As for the open source of large models, its purpose is mainly to enrich the ecology and promote technological innovation.

If any technology simply follows a closed and monopolized development path, it will either lack vitality or become a state that is not friendly to the entire ecology . Just like the biosphere, a certain degree of diversity must be maintained. Open source is more about maintaining technological innovation and technological diversity, so that the open source community can also invest in the core of technology.

Q: You have mentioned before that Zhipu’s commercialization focus is on ToB. Which industries are ToB customers currently focusing on? What specific applications do you help customers with?

Zhang Peng: It just means that our current main source of income is still on the ToB side, but it does not mean that our commercialization path is only ToB.

At present, the B-side customers we serve cover more than 10 industries, including finance, education, the Internet, retail, automobiles, energy, traditional manufacturing, etc.

We are currently helping industry customers get started, and I mainly use several paths.

First of all, we have our own open platform, which can help our customers quickly access model capabilities at a relatively low cost. They can quickly try innovations, and then update themselves and iterate their own products and AI empowerment.

The second one is for some large companies that have relatively high data security and privatization requirements. We will provide cloud privatization solutions and local privatization solutions, and then we will be specific for some scenarios, and then there will be Chinese solutions. For small enterprises like this, we will provide integrated software and hardware solutions.

At present, our open platform has more than 400,000 enterprise users, including some small developer teams, who have registered and used our model API on our platform. The daily service volume now exceeds 60 billion tokens. The service volume is growing very rapidly.

Q: OpenAI recently stopped providing API services to Chinese developers, and Zhipu soon launched a relocation service. How is the current situation of users migrating here?

Zhang Peng: From our observation, there is an increase, but the response of the entire market has a process. I also asked my friends about the situation. In fact, everyone has observed an increase.

Q: Do you have any plans to go overseas next?

Zhang Peng: We are already laying out our international business lines and are currently negotiating some business.

Where is the next step for big models?

Q: GPT-5 has been delayed. The industry believes that the iteration curve of large models is slowing down. Is Scaling Law coming to an end?

Zhang Peng: The early Scaling Law was very simple and only focused on the parameters of the model. However, later everyone discovered the connotation of Scaling Law. The size of its parameters was only one of the factors and variables. It also included other factors such as We talked about the amount of data used for training and the number of tokens, and later found that it was also related to the amount of calculation, so the connotation of Scaling Law itself is also constantly changing.

What is closer to the truth of Scaling Law may be the amount of calculation. The amount of calculation combines computing power and data, as well as parameter scale. The final result may be a comprehensive variable, which can better represent Scaling Law. From the perspective of calculation amount, we believe that Scaling Law is still effective.

A side example to prove this is that the United States now restricts the export of AI technology. Its restriction standard is no longer, for example, the computing power of the chip, or the amount of parameters and data of the model, but the amount of calculation. Draw a line based on 10 to the 24th power. If the calculation amount of this model exceeds this line, it will not be allowed, so you can see that it is also moving in the direction closer to the truth.

But what is its essence? We are still exploring, because Scaling Law itself is an observed phenomenon and a rule obtained. It is not the connotation of a truth .

Q: In the future, will Zhipu AI prefer the large model on the client side or on the cloud?

Zhang Peng: We believe that the path of the current big model is to move towards general artificial intelligence, or even super intelligence that surpasses humans. This is a relatively reliable path at present, but we will not be limited to the cloud or the client side. choice .

We believe that the development of this technology has its own stages. At a certain stage, for example, the cloud may have the highest level of intelligence due to computing power and other reasons. If we want to move to the next stage of general artificial intelligence or Super artificial intelligence advances. It may still be centralized in the cloud to provide this stronger capability.

▲Humanoid robots on WAIC.

However, in some specific scenarios, such as on-device mobile phones, cars, and robots, it may be necessary to use end-side computing power and end-side models to cooperate with the cloud. It may be a combination of end-cloud and end-side. With this model, we may be able to achieve intelligence as smart as the current cloud on mobile phones in the future, but this involves many comprehensive factors, such as chips, computing power and energy, and a series of issues.

Any technology has stages. At this stage, the plan is like this, but from a longer time scale, is it (device side and cloud) the ultimate answer? Definitely not, it will definitely develop further in the future.

Q: Where do you think the next development of large models will be?

Zhang Peng: At present, the language and writing abilities of large models are close to or even slightly exceed the average human level. Next, we hope to use one word to describe it, "moving away from virtuality to reality", that is, it is no longer limited to being a brain in a vat, but can enter actual life and work to create actual productivity.

To achieve this goal, in addition to language ability, many other abilities are required, such as visual ability, auditory ability, and the ability to use hands and feet to grow hands and feet. We hope it will become a multi-modal A state-of-the-art model that can understand human intentions, decompose human intentions into logical execution steps, and use tools to connect to the physical world to complete these tasks.

Since we want it to have a stronger ability to interfere with the physical world, security becomes more important to prevent it from doing harmful things in the real physical world. This is the same in the digital world. We need to improve security. And do more with alignment, we call it super intelligence and super alignment.

# Welcome to follow the official WeChat public account of Aifaner: Aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo