"Artificial intelligence has narrowed our imagination of the future, and the science fiction culture of the past 100 years has narrowed AI itself."
In March last year, OpenAI opened AIGC's Pandora's box, and the large model became a "panacea" to the future overnight. China has also set off a "100-model war", from volume parameters to low prices, the AI industry presents a hot and confused complex scene.
Since last year, Baidu Robin Li has proposed that "the volume model is not as good as the volume application". During the 2024 World Artificial Intelligence Conference, he said bluntly: "Without application, large models will be worthless." ”
So, where will the next battleground for AI applications appear?
Abroad, OpenAI has launched GPTs, which Altman believes are the "initial form" of AI Agent. Apple, on the other hand, released Apple Intelligence. Cognition AI, a start-up, has launched the world's first AI software engineer. In China, companies represented by Baidu are also focusing on intelligent twins.
Recently, during the 2024 World Artificial Intelligence Conference, Robin Li, founder of Baidu, Yang Yudong, editor-in-chief of CBN Media Group, and Chen Qian, founder of "Silicon Valley 101", conducted a roundtable interview. He believes that agents represent the future trend of the AI era and are exploding.
01
The volume model is not as good as the volume app
Since last year, Robin Li has shared many times that the volume model is not as good as the volume application.
At WAIC this year, he said bluntly: "Without application, the large model will be worthless." ”
I have to admit that the industry is still focusing more on the basic model. All day long, running scores everywhere to brush up the charts, who surpassed GhatGPT again, and Open AI launched Sora again. Today's release, tomorrow's update, but few people pay attention, where is the application? Who has benefited?
We have sorted out that there are two main routes for AI application at present:
First, from 0 to 1 "super apps" or "killer apps", this kind of apps have never been born in the past and are completely based on AI capabilities. For example, ChatGPT, Sora, and Midjourney abroad; Domestic Wenxin Yiyan, Kimi, Keling, etc.
Second, AI Copilot combines existing services with AI, puts AI in the position of "co-pilot", and coordinates the working mode of AI and human collaboration. Microsoft, a foreign country, is a firm believer in Copilot. The domestic Baidu library also uses actual data growth and user feedback to prove that this route is indeed extremely valuable. Since the beginning of this year, after AI reconstruction, Baidu Library, which has AI functions such as generating PPT, comics, and mind maps, has had about 26 million paying users.
The agent may become the final form of carrying these two routes.
02 The Next Battlefield: Dead "Agent"
In his WAIC speech, Robin Li bluntly predicted: "As the basic model becomes more and more powerful, it will become easier and easier to develop applications. The simplest is the agent, which is also the direction of AI development that we are most optimistic about. ”
Microsoft founder Bill Gates once pointed out that who can dominate the personal assistant agent is the big deal. "Because you're never going to search for sites, you're going to be productive sites, you're never going to Amazon."
Andrew Ng, a well-known expert in the field of AI and a professor at Stanford University, also believes that AI agent workflows will promote AI to make great progress, and may even surpass the next generation of basic models.
So, why are intelligent bodies so favored?
In the interview, Robin Li explained: "An agent is an application based on a large model that can be almost 'universally applicable'. Most of today's AI native applications, you can make them in the way of agents, and the results are also good. Since the threshold is low enough, you may not even need to program it to make a good agent. ”
In Robin Li's view, the agent in the AI era is like the Internet website in the mid-90s, due to the extremely low production threshold, millions of websites were born in the 90s, and after the big waves, some very good websites finally came out, such as Google, Facebook, etc.
Nowadays, a novice or first-year student can easily make an agent, and the ultra-low threshold can allow more and more people to participate. ”
Robin Li believes that the capabilities used in AI Agent are still very rudimentary, and in the future, they will produce Agent capabilities that we can't think of today. But the creation of these capabilities relies on millions of developers to develop a wide variety of applications. The process of generating new requirements and solving these needs in the process of using them is an innovation process, which is the process of AI agent evolution.
For example, Uncle Alex, a content creator with tens of millions of fans, has created a real-life interactive avatar through the Wenxin intelligent twin platform, providing all fans with an "exclusive companion electronic girlfriend", through which everyone can communicate with Uncle Alex online at all times, from a one-to-many information release type to thousands of "one-to-one" interactive communication.
We found that as early as September 2023, Baidu released the "Spirit Matrix" Wenxin Yiyan plug-in ecological platform, which has since been upgraded to an agent platform. On July 5 this year, Baidu Wenxin Intelligent Twin Platform announced that it would open Wenxin Model 4.0 for free. At present, 200,000 developers and 63,000 enterprises have settled in.
03
Search will become the largest entry point for agent distribution
In the future, with the emergence of millions of agents, how to accurately distribute agents to match user needs will become the key to unleashing the potential of agents and promoting their wide application.
Search, obviously, is a more efficient channel. Robin Li also said that "search will become the biggest entry point for agent distribution".
Baidu is already one step ahead. In terms of distribution, Baidu Search has launched the "AI Assistant" channel to create a natural field for agent distribution. The "AI assistant" can not only accurately interpret the user's intention and push the most matching agent to the user, but also brings together a large number of professional agents for various scenarios, which can be called by users at any time to solve personalized problems and make the agent really run.
Taking Baidu College Entrance Examination Agent as an example, rather than letting a large model write a college entrance examination essay, helping candidates fill in their volunteers is the real need. In the process of choosing a university or major, every candidate will have a variety of different questions. Now, they can open the college entrance examination agent through Baidu search for one-on-one Q&A. At its peak, the gaokao agent had to answer more than two million questions a day.
In addition, the Wenxin Intelligent Twins platform also supports the use of APIs, SDKs and other templated access forms to open up Baidu's extraterritorial scenarios such as independent apps, WeChat applets and sites, so that the distribution field of the agent can be maximized, as much growth momentum as possible, so that developers can avoid the worry that no one cares about the applications they make.