This article is authorized for reprint by AI New Media Quantum Bit (Public Account ID: qbitai). For reprinting, please contact the source.
This article is about 3000 words long and is recommended to be read in 6 minutes.
The gap between large models will continue to widen!
Li Yanhong’s latest internal speech has been exposed, sparking heated discussions in the industry.
After all, in the current phenomenon where various large models sweep leaderboard test sets, with multiple scores exceeding GPT-4o, it is easy to create the illusion that the next GPT-4o or the next OpenAI is about to emerge.
Why do I say this? Li Yanhong further explained that the ceiling for large models is very high, and we are still far from the ideal situation, so models need to be constantly iterated, updated, and upgraded quickly.
This requires long-term investment over years or even decades to continuously meet user needs and reduce costs while increasing efficiency.
In addition, he also stated that open-source models are not efficient and cannot solve the computing power problem, while intelligent agents are the most important development direction for large models.
As a pioneer in the application of large models, Baidu’s helmsman Li Yanhong’s speech undoubtedly provides practical reference for the industry.
Let’s take a look at what he actually said.
Li Yanhong’s Internal Talk Exposed: Three Misconceptions About Large Models
In his internal talk, Li Yanhong raised three points of thought that directly respond to the current misconceptions that are receiving attention: Are the gaps between large models narrowing? Is large model technology approaching its ceiling? Why do commercial models offer better cost-effectiveness?
First, the gap between large models is not narrowing but widening.
Right from the start, Li Yanhong opposed the view that the gap in capabilities between large models is shrinking, asserting that the differences between models remain significant and will continue to grow. He pointed out that although newly released models perform well on test sets, this does not prove that their gap with the state-of-the-art models like GPT-4o has narrowed.
He explained that many models, after being released, may appear to be close in capability due to leaderboard rankings, answering techniques, and guessing test questions. However, “in actual applications, there is still a significant gap in strength.”
On one hand, the differences between models are multidimensional. The evaluation of model capabilities includes understanding, generation, and logical reasoning across multiple dimensions, as well as corresponding costs and inference speeds. Additionally, overfitting to test sets may lead to misunderstandings about model capabilities.
Now that large models are entering the application phase, he believes that the true measure should be whether models can meet user needs and generate value in real applications. Therefore, in Baidu’s actual usage, he does not allow technical personnel to focus on leaderboard rankings.
On the other hand, the ceiling for models is very high. What can be done today is still far from the desired results, so models need to be continuously iterated and updated. Only through sustained investment over years or even decades can models meet user needs, scenarios, and demands for efficiency improvement or cost reduction. This is also the key to maintaining competitiveness.
Thus, Li Yanhong believes that the notion of leading by 12 months or lagging by 18 months is not that important. Even if you can ensure that you are always six months ahead of your competitors, you have won.
Second, open-source models cannot solve efficiency problems in commercial applications.
In his speech, Li Yanhong emphasized that open-source models require users to deploy and maintain them on their own, leading to low GPU utilization and an inability to effectively share inference costs. In contrast, closed-source models achieve higher efficiency and effectiveness by allowing users to share resources and distribute R&D costs.
Currently, Baidu’s Wenxin large models 3.5 and 4.0 can achieve GPU utilization rates of over 90%.
As mentioned earlier, evaluating a model involves multiple dimensions; it is not just about looking at various capabilities on the leaderboard but also considering effectiveness and efficiency. When large models accelerate into commercial applications, open-source models do not have an advantage in the pursuit of high efficiency and low costs.
Li Yanhong clearly stated that in the era of large models, the efficient utilization of computing power is the key to determining the success or failure of a model, and open-source models cannot solve this issue.
Finally, intelligent agents are the most important development direction for large models, with low thresholds making application transformation more direct and efficient.
What stages will the development of large models go through? In his internal talk, Li Yanhong provided a clear answer.
The first stage is the Copilot phase, assisting humans in operations; the next is the Agent intelligent agent phase, which possesses the ability to autonomously use tools and self-evolve; and finally, the AI Worker phase, which can independently complete a variety of tasks.
Among these, intelligent agents are regarded as the most important development direction for large models. Compared to the multimodal focus that everyone is paying attention to, industry consensus has yet to form. However, in Baidu’s products, such as the Wenxin intelligent agent platform AgentBuilder, the potential of intelligent agents has already begun to be recognized.
Moreover, this low-threshold characteristic makes the transformation from models to applications simple, prompting the creation of numerous new intelligent agents on the Baidu platform.
Li Yanhong emphasized that leveraging Baidu’s user base and demand, intelligent agents can better meet market needs and promote further development.
Baidu’s Intelligent Agent Practice Has Entered Deep Waters
In summary, the points discussed by Li Yanhong reflect the current state of affairs, while intelligent agents represent the future. The backdrop for all this is closely related to the current phase of large model development entering deep waters.
As the speed of basic model updates slows down and the application of large models gradually deepens in industries, companies are facing a more complex market environment and technical challenges. Simple technological iterations are no longer sufficient to meet the diverse demands of the market.
People’s expectations and views on large models have also changed; the number of model parameters and leaderboard scores are no longer core indicators of model capability, and whether a model is open-source or not is actually not that important.
The industry’s demand for AI is no longer merely a pursuit of technology; solving practical problems is the only standard for measuring large models. In this process, more issues and challenges cannot be ignored, such as the cost of inference and computing power, as well as business efficiency.
Therefore, Baidu, which has been investing long-term and continuously in the industry, naturally provides a reference for many large models in China facing this proposition.
The answer is intelligent agents.
Thus, Li Yanhong’s internal talk is not only about industry cognition but also a powerful validation and reflection of Baidu’s intelligent agent practice.
Previously, Li Yanhong emphasized in multiple speeches that intelligent agents represent the future trend of the AI era.
As a nearly universally applicable large model application, intelligent agents not only have a low threshold but even allow users to easily develop powerful applications without programming skills. Li Yanhong vividly compared intelligent agents to “the website of the AI era,” predicting that it will form a huge ecosystem with millions of units. This widespread application potential makes intelligent agents the “Super APP” of various industries, promoting the popularization and application of AI technology.
Correspondingly, Baidu’s layout in the intelligent agent field is significant.
Through the Wenxin intelligent agent platform AgentBuilder, Baidu has attracted 200,000 developers and 63,000 enterprises to settle in, and in July 2023, it opened the Wenxin large model 4.0 for free. This move allows developers to flexibly choose suitable model versions when building intelligent agents, greatly lowering the development threshold.
In a short period, Baidu’s intelligent agents have demonstrated the powerful potential of large model applications. According to Baidu’s Q2 2024 financial report, the distribution volume of intelligent agents in Baidu’s ecosystem is rapidly increasing, with daily distribution exceeding 8 million in July, doubling compared to 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, creating a win-win situation for users, developers, and service providers.
Li Yanhong emphasized that the development of intelligent agents relies not only on technological innovation but also on closely aligning with user needs. As user demand for intelligent agents continues to rise, these intelligent agents can iterate quickly. Only by continuously expanding the intelligent agent ecosystem can AI technology be deeply applied in various fields.
As the application of large models deepens, Baidu’s intelligent agent practice undoubtedly provides important reference and inspiration for the industry.
The Large Model Frenzy Enters the Reshuffle Period
This year, it has become apparent that with the continuous development of large model technology and the deepening of applications, the industry is entering a new phase. The characteristics of this phase are that the landscape of large model players has basically formed: players with self-developed and sustainable R&D capabilities for large models are beginning to cluster at the top.
At the same time, large model applications and implementations are starting to enter an ecological construction phase.
More and more entrepreneurs optimistic about the prospects of large models are no longer entangled in whether to self-develop or build large models; instead, they focus more on how to utilize existing large models to solve practical pain points and needs.
In this process, intelligent agents, as the smallest AI application realization method, exhibit enormous potential. They are lightweight with low thresholds, capable of quickly spreading across industries and meeting the two major demands of efficiency and cost. With the continuous enhancement of basic models, intelligent agent applications can become simpler and more widespread.
This is also the core reason why Li Yanhong is optimistic about intelligent agents.
Through Li Yanhong’s speech, we can see that Baidu’s strategic focus is shifting. The phase of competing for basic models has passed; what is more important now is how to build a rich and diverse application ecosystem through intelligent agents, making the ecosystem a moat for Baidu’s large models and Wenxin.
This means that Baidu will pay more attention to the value and significance of intelligent agents in the application ecosystem by continuously improving the intelligent agent platform and tools, attracting more developers and enterprises to join, and jointly creating a prosperous AI application ecosystem.
In the future, intelligent agents will not be limited to basic functions like content creation and schedule planning, but will expand into more specialized fields such as healthcare, finance, and legal services, providing users with personalized and efficient solutions.
To achieve this vision, enterprises need to continuously invest resources in technological innovation and iteration, continually optimizing algorithms and enhancing user experience.
Of course, in this process, there will also be some important issues to face, such as data privacy and security, management and maintenance of intelligent agents, etc. However, any technology application entering deep waters will face various challenges.
As the intelligent agent ecosystem continues to grow, Baidu is leading the industry towards a more intelligent and efficient future, bringing new opportunities and challenges to various industries.
Editor: Yu Tengkai
About Us
Data Party THU, as a public account focusing on data science, is backed by the Tsinghua University Big Data Research Center, sharing cutting-edge data science and big data technology innovation research dynamics, continuously spreading data science knowledge, and striving to build a platform for gathering data talent, creating the strongest group in China’s big data.
Sina Weibo: @Data Party THU
WeChat Video Account: Data Party THU
Today’s Headlines: Data Party THU