Robin Li reveals the misunderstanding of big models: open source cannot solve the efficiency problem, and the gap between big models is getting bigger and bigger

Bai Xiaojiao sent from Aofei Temple
Quantum Bit | Public Account QbitAI

The gap between the big models will become bigger and bigger!

Li Yanhong’s latest internal speech was exposed , which immediately sparked heated discussions in the industry.

After all, with various large models sweeping the test sets and many scores exceeding GPT-4o, it is easy to give people the illusion that the next GPT-4o or the next OpenAI is about to be born .

Why do you say that? Li Yanhong further explained that the ceiling of the large model is very high, and it is still far from the ideal situation, so the model must be continuously and quickly iterated, updated and upgraded.

This requires consistent investment for several years or even decades to continuously meet user needs and reduce costs and increase efficiency.

In addition, he also said that the open source model is not efficient and cannot solve the computing power problem, and intelligent agents are the most important development direction of large models.

Baidu has always been a pioneer in the application of large models, and the conversation this time by its helmsman Robin Li will undoubtedly provide a practical reference for the industry.

Let’s take a look at what he said.

Li Yanhong's internal conversation exposed: three big model cognitive misunderstandings

During the internal conversation, Robin Li raised three points of thought, which just happened to respond to the cognitive misunderstandings that are currently receiving attention: Is the gap between large models narrowing? Has the large model technology reached its ceiling? Why is the commercial model more cost-effective?

First, the gap between large models is not getting smaller, but getting bigger.

At the beginning, Li Yanhong opposed the outside world's view that the gap in the capabilities of large models has narrowed, and believed that the gap between models is still significant and will become larger. He pointed out that although the newly released models performed well on the test set, this does not prove that the gap between them and the most advanced models such as GPT-4o has narrowed.

He explained that after many models are released, through ranking, guessing test questions, and answering skills, from the rankings, the model's capabilities may be very close, "but in actual application, there is still a clear gap in strength."

On the one hand, the gap between models is multi-dimensional . The evaluation of model capabilities includes multiple dimensions such as understanding, generation, and logical reasoning, as well as the corresponding costs and reasoning speed. In addition, overfitting on the test set may lead to misunderstandings about model capabilities.

Now that big models have reached the application stage, he believes that the real criterion should be whether the model can meet user needs and generate value in actual applications. Therefore, in the actual use of Baidu, he does not allow technical personnel to compete for rankings.

On the other hand, the model has a high ceiling . What we can do today is still far from the desired effect, so the model needs to be continuously iterated and updated. Only by investing for several years or even more than ten years can the model meet user needs, scenarios, and meet the needs of improving efficiency or reducing costs. This is also the key to maintaining competitiveness.

Therefore, Li Yanhong believes that being 12 months ahead or 18 months behind is not that important. Even if you can always be 6 months ahead of your competitors, you will win .

Secondly, the open source model cannot solve efficiency problems in commercial applications .

In his speech, Robin Li emphasized that open source models require users to deploy and maintain them by themselves, which leads to low GPU utilization and inability to effectively share inference costs. However, closed source models achieve higher efficiency and effectiveness by allowing users to share resources and share R&D costs.

Currently, the GPU utilization rate of Wenxin Model 3.5 and 4.0 can reach more than 90%.

As mentioned above, there are many dimensions to evaluate a model. It is not only about looking at the multiple capabilities on the list, but also about the effect and efficiency. As large models are accelerating into commercial applications, open source models have no advantage in the pursuit of high efficiency and low cost.

Robin Li clearly stated that in the era of large models, the efficient use of computing power is the key to determining the success or failure of the model, and open source models cannot solve this problem.

Finally, intelligent agents are the most important development direction of large models, and low barriers make application transformation more direct and efficient.

What are the main stages of the development of the big model? In an internal speech, Robin Li gave a clear answer.

The first is the Copilot stage, which assists humans in operations; the next is the Agent stage, which has the ability to use tools autonomously and self-evolve; and the last is the AI ​​Worker stage, which can complete a variety of tasks independently.

Among them, intelligent agents are the most important development direction of large models. Compared with multimodality, which everyone is paying attention to, there is no industry consensus. However, in Baidu's products, such as the Wenxin intelligent agent platform AgentBuilder, the potential of intelligent agents has begun to be recognized.

The low threshold feature makes the transformation from model to application simple, prompting a large number of new intelligent agents to be created on the Baidu platform.

Li Yanhong emphasized that with the help of Baidu's user base and needs, intelligent bodies can better meet market demand and promote its further development.

Baidu's intelligent body practice has entered the deep water zone

In summary, if the first two points mentioned by Robin Li are still about the present, then the intelligent body represents the future. And the background premise of all this is not unrelated to the current stage of large-scale model development entering the deep water zone .

Today, as the pace of basic model updates slows down and large model applications gradually penetrate the industry, companies are facing more complex market environments and technical challenges. Simple technology iterations are no longer sufficient to meet the diverse needs of the market.

People's expectations and views on large models have also changed accordingly. The number of model parameters and ranking scores are no longer the core indicators of model capabilities, and it is actually not important whether it is open source or not.

The industry's demand for AI is no longer just a simple pursuit of technology. Solving practical problems is the only criterion for measuring large models. In this process, more problems and challenges cannot be ignored, such as the cost of reasoning and computing power, as well as the efficiency of processing business.

Baidu, which has been investing in the industry for a long time and continuously, naturally provides a development reference for many large models in China when facing the current proposition.

The answer is Agents .

Therefore, this internal conversation with Robin Li was not just about industry awareness, but also a powerful verification and reflection from Baidu's intelligent body practice.

Prior to this, Robin Li emphasized in many speeches that intelligent agents represent the future trend of the AI ​​era.

As a large-scale model application that is almost "universally applicable", intelligent agents not only have a low threshold, but also do not even require programming skills, allowing users to easily develop powerful applications. Robin Li figuratively compares intelligent agents to "websites in the AI ​​era", indicating that it will form a huge ecosystem of millions of people. This wide range of application potential makes intelligent agents a "Super APP" in all walks of life, promoting the popularization and application of AI technology.

Correspondingly, Baidu has made significant progress in the field of intelligent entities.

Through the Wenxin intelligent agent platform AgentBuilder , Baidu has attracted 200,000 developers and 63,000 companies to join, and will open the Wenxin Big Model 4.0 for free in July 2023. This move allows developers to flexibly choose the appropriate model version when building intelligent agents, greatly lowering the threshold for development.

And in a short period of time, Baidu's intelligent agents have shown strong potential for large-scale model applications. According to Baidu's Q2 2024 financial report, the distribution volume of intelligent agents in the Baidu ecosystem is increasing rapidly, with an average daily distribution of more than 8 million in July, doubling from May.

Popular intelligent agents include content creation, personality testing, and schedule planning, covering multiple industries such as education, law, and B2B. Baidu's intelligent agent ecosystem has attracted 16,000 merchants to participate, creating a win-win situation for users, developers, and service providers.

Robin Li emphasized that the development of intelligent agents not only depends on technological innovation, but also needs to be closely integrated with user needs. As user demand for intelligent agents continues to rise, these intelligent agents can be quickly iterated. Only when the intelligent agent ecosystem continues to expand can the in-depth application of AI technology be promoted in various fields.

As the application of large models becomes increasingly in-depth, Baidu's intelligent body practice undoubtedly provides important reference and inspiration for the industry.

The craze for large models is entering a period of reshuffle

It has been clearly felt this year that with the continuous development and in-depth application of large-scale model technology, the industry is entering a new stage. The characteristics of this stage are that the pattern of large-scale model players has basically taken shape: players with self-developed and sustainable large-scale model development capabilities have begun to gather at the top .

At the same time, the application and implementation of large models began to enter the ecological construction period .

More and more entrepreneurs who are optimistic about the prospects of big models are no longer concerned about whether to develop or build big models themselves, but are more concerned about how to use existing big models to solve actual pain points and problems.

In this process, the intelligent agent, as the smallest AI application implementation method, has shown great potential. It has a low threshold and is lightweight, and can be quickly rolled out and covered in the industry, meeting the two major requirements of efficiency and cost. With the continuous enhancement of the basic model, the application of intelligent agents can become simpler and more extensive.

This is also the core reason why Robin Li is optimistic about intelligent entities.

From Li Yanhong's speech, we can see that Baidu's strategic focus is shifting. The stage of volume-based models has passed, and now it is more important to build a rich and colorful application ecosystem through intelligent agents, so that the ecosystem can become the moat of Baidu's big model and culture.

This means that Baidu will pay more attention to the value and significance of intelligent entities in the application ecosystem, and by continuously improving the intelligent entity platform and tools, attract more developers and companies to join in and jointly create a prosperous AI application ecosystem.

In the future, intelligent agents will not only be limited to basic functions such as content creation and schedule planning, but will expand to more professional fields such as medical, financial and legal services, providing users with personalized and efficient solutions.

To achieve this vision, companies need to continue to invest resources in technological innovation and iteration, continuously optimize algorithms and improve user experience.

Of course, there are also some important issues to be faced in this process, such as data privacy and security, management and maintenance of intelligent bodies, etc. However, any technology application will face challenges of one kind or another when it enters the deep waters.

As the intelligent ecosystem continues to grow, Baidu is leading the industry towards a more intelligent and efficient future, bringing new opportunities and challenges to various industries.

Click here ???? Follow me, remember to mark the star~