At the Baidu World Conference a year ago, Robin Lee "hand-in-hand" taught users to do AI native applications on the spot, calling on the industry not to continue to roll up models, but to devote energy to application development.
But in the past twelve months, the super application that everyone had hoped for did not appear. Instead, more and more people began to discuss whether this round of global big model craze was a new technological revolution or a new round of foam.
Robin Lee apparently heard these questions. On November 12, he said at the Baidu World Conference: "As the flag bearer of China's AI, I think we are qualified to answer this question."
He provided the latest call data for the Wenxin big model, which currently has an average daily call volume of over 1.5 billion and has grown 7.5 times in less than half a year.
In June, when the data reached 200 million yuan, Robin Lee once discussed with Baidu executives how to call the future of the big model "how to become". At that time, the conclusion was that the call volume could increase 10 times in a year.
Now, half a year later, Baidu has approached this growth rate. Robin Lee believes that this shows that users really need a big model, and practitioners can rest assured.
In order to minimize the threshold for developers to join their own big model ecosystem, Baidu announced that it will launch the multi-agent collaboration tool "Seconda" in January next year. Robin Lee claimed that this product can enable people who do not understand the code to develop an application product.
"We will usher in an unprecedented era of making money only by thinking", Robin Lee made his bold prediction at the center of the stage.
Eliminating illusions is the primary prerequisite
Baidu wants to persuade developers to build applications based on the Wenxin big model, and eliminating illusions is the primary prerequisite.
"If the big model is always serious nonsense, no one will believe you or have applications. It is not intelligence but intellectual disability", Robin Lee said that the biggest change in the big model industry in the past 24 months is the basic elimination of illusion.
In his view, as a probabilistic model, the content generated by the large model already has uncertainty. After adopting RAG (Retrieval Augmented Generation) technology, the content quality and accuracy of the large model have been greatly improved.
This year, the competition among top model manufacturers has shifted from the field of text to graphics and videos. The most well-known product is undoubtedly OpenAI's Sora video tool, which has also driven a large number of domestic companies to focus on multimodal fields and benchmark against OpenAI.
Shooting: Interface News
Robin Lee was particularly confident about this. He said, "When the whole Chinese Internet was beating Sora's chest, we decided to solve the illusion problem of image generation. Without solving it, there would be no application."
The new technology for generating images launched by Baidu is called iRAG (Image Based RAG).
Its basic principle is to combine Baidu's image resource library with Baidu's basic magical ability to generate images with less "machine flavor". The official claims that its effect far exceeds that of Wenshengtu's native system.
Robin Lee said that currently, the pictures generated by the Wensheng Map system based on the big language model on the market are usually "false at a glance", or even illogical. While iRAG has no illusion, is ultra real, and has low cost, which is of high commercial value to enterprise customers.
From the cooperation case between Volkswagen and Baidu displayed on site, it can be seen that using AI to generate poster images can reduce the cost from the traditional tens of thousands of yuan per group to nearly zero, and the generated vehicle models are consistent with Volkswagen's real-life models.
The potential application scenarios of this technology are currently mainly concentrated in the fields of film and television works, anime works, and poster production.
Where does AI go: intelligent agents and industrial applications
In the past year, Robin Lee has stressed the importance of agents in many internal and external speeches, saying that there are not many technology companies like Baidu that regard agents as an important strategic direction worldwide.
At today's Baidu World Conference, he provided additional explanations for this viewpoint.
Robin Lee believes that the agent will become a new carrier of content, information and services in the AI native era. To be an agent in the AI era is like being a website in the PC era and an We Media account in the mobile era. The difference is that the agent is more like a real person.
On the one hand, the threshold for doing it is low enough, and Robin Lee even thinks that anyone can start to be an agent. On the Baidu Wenxin intelligent agent platform, 11 year old elementary school students are already making intelligent agents and successfully distributing them through search and other channels.
On the other hand, the ceiling of intelligent agents is very high, which is sufficient to support developers in creating very powerful applications. Especially multi-agent collaboration can solve more complex problems.
Official
Robin Lee mainly demonstrated the application of four types of agents at the scene: corporate agents, role agents, tool agents and industry agents.
Corporate intelligent agents mainly replace the functions of traditional official websites, providing company introductions, product images, and one-on-one services; Role based intelligent agents mainly refer to products based on large-scale models, which are applied in fields such as live streaming, consulting, and education; Tool based intelligent agents are currently more concentrated at the tool level, such as writing assistants and design assistants; Industry based intelligent agents mainly target vertical fields such as law, healthcare, finance, and travel.
From the perspective of classification, the agent application scenarios Robin Lee described for Baidu are basically areas with more real needs, but the key is how to persuade developers to participate and give them enough rewards.
At the industrial application level, Baidu's customers are mainly concentrated in the fields of finance, energy, education, and public services. On the Qianfan model platform of Baidu AI Cloud, there are 33000 fine tuning models and 770000 enterprise applications.
In the industry case introduction, Baidu presented a cooperation case with Yum China. The scope of cooperation between the two parties involves store location selection, intelligent customer service, and internal large model service products. Baidu also provides the Qianfan platform and basic computing power services to Yum! Brands. In the future, the two companies will seriously consider the technical feasibility of using large model ordering on the ordering screen.
Undoubtedly, well-known brands like Yum! China are benchmark customers urgently needed by Baidu to promote its own large model services. At this year's World Congress, Robin Lee highlighted the cases of leading companies such as Volkswagen, BYD, Yum China and Zhilian Recruitment. Baidu clearly hopes to use them to attract more small and medium-sized developers and join Baidu's big model ecosystem.
To this end, Baidu has also implemented an official recommendation strategy and released a list of "Top 100 Industrial Applications" to recommend these high-quality industry applications developed based on the Wenxin big model to the outside world.
Robin Lee said that Baidu does not want to launch a "super application", but wants more enterprises to create millions of "super useful" applications.
Do you want to eliminate programmers?
Whether promoting intelligent agents or industry applications, lowering the threshold is an urgent problem that Baidu needs to solve.
Robin Lee imitated Apple on the spot and also launched a "One More Thing" product, that is, "Second Da". In its official positioning, this is a software that can make applications without writing code, consisting of a basic large model and intelligent agents.
"It is by far the most complex multi-agent collaboration tool in human history", Robin Lee said, this product can enable a person who does not understand the code at all to have the ability of a programmer to quickly and low-cost turn any idea into an application product.
Shooting: Interface News
Although not many peers are making such products, which makes this "most human" somewhat redundant, Baidu's approach to trying this product is very clear.
When an ordinary user puts forward a demand for it, the No. 0 intelligent agent of Secondclick will analyze the instruction and issue a scheduling plan, dividing the task into three steps: demand determination, content production, and engineering development, and summoning multiple intelligent agents to collaborate to complete all development tasks, and ultimately output the application product that the user needs.
In the live demonstration, the invitation system for a press conference takes only 3 minutes from the user's request to the end of product development, and the user does not need to write any line of code.
Robin Lee said that Seconda has the advantages of code free programming, multi-agent cooperation and multi tool invocation. Users can complete a whole set of system construction with the help of natural language interaction.
This product sparked significant interest among many attendees during the live demonstration, and several enterprise developers encountered by Interface News were curious about its capabilities and whether it could truly replace ordinary programmers. Unfortunately, there was not much detailed demonstration on site.
Seconda will be officially released in January next year. Robin Lee said that with the improvement of basic model capabilities, Seconda will be able to do more complex things. In the future, enterprises may not need to recruit project managers, designers, developers and testers, but can independently command multiple agents to develop applications in collaboration.