The wave of big models sparked by OpenAI has been hot for nearly two years, with related technologies iterating and innovating at an unprecedented speed. From large companies to entrepreneurs to venture capitalists, they are all searching for super applications based on big models in the era of generative AI.
However, objectively speaking, the super applications that the industry had hoped for have not yet emerged. Some people even begin to question whether this global big model craze in the past 24 months is a new technological revolution or a new round of foam?
At today's Baidu World Conference, Robin Lee, chairman of Baidu, answered this question with a picture. In the speech, when talking about the AI foam that is hot in the industry, the screen behind him showed a curve of the daily average adjustment amount of Wenxin model, showing a steep growth. Data shows that the daily usage of Baidu Wenxin's large model reached 1.5 billion, with a growth rate of 7.5 times in six months.
"In the past 18 months, the explosion of China's big model applications can be represented by this chart or this curve." Robin Lee said that when the daily call data was still 200 million six months ago, he once said when discussing the future of the big model with Baidu executives: "If our daily average API call volume of the big model increases 10 times within a year, I think it will be. Now only half a year later, we are closer to this number."
On the same day, Robin Lee released two AI technologies: the retrieval enhanced Wensheng Graph (iRAG) technology and the codeless tool "Seconds Da". The former is mainly used to solve the illusion problem of large models in image generation and enhance practicality; The latter lowers the industry threshold and enables ordinary users to possess the skills of programmers.
Retrieval enhancement has become the consensus of the big model industry. In the past 24 months, Robin Lee believes that the biggest change for the industry is that the big model has basically eliminated illusion, and the accuracy of answering questions has been greatly improved, making AI usable and trustworthy from "serious nonsense".
He recalled that at the beginning of this year, when the whole Chinese Internet was beating its chest for Sora, Baidu decided to solve the illusion problem of image generation. The search enhanced text generated image technology released by Baidu today combines Baidu's search image resources with basic model capabilities to generate various hyper realistic images.
On site, he used the phrase "draw a realistic picture of a Volkswagen patrol car flying over the Great Wall." The generated image, when enlarged, showed no distortion of the car model or logo, and had a high degree of integration with the Great Wall background.
However, the First Financial News reporter found that this realistic image can only be said to have removed the "machine flavor" to a certain extent, and is more realistic than the "fake at a glance" AI image, but it is still far from achieving the realistic effect of being able to "confuse the real with the fake".
However, with the advancement of AI generated image technology and improved usability, the application space is also opening up. "For example, in the brand promotion scene, it used to cost ten to two hundred thousand or even hundreds of thousands to shoot such a group of posters, but now the cost of such creation is close to zero," said Robin Lee.
"The commercial value of iRAG is reflected in: no illusion, super reality, no cost, and it can be taken as soon as possible." Robin Lee then teased: "Just imagine, if the model generated by Volkswagen's poster looks like Toyota, it will be a nuisance."
With these basic model capabilities in place, he predicts that the industry will soon usher in an AI application explosion. Robin Lee highlighted two AI application directions: industrial application and agent.
Focusing on the industrial application of the big model, Robin Lee mentioned that in the past year and a half, the big model has achieved results in cost reduction and efficiency increase after combining with scenarios in finance, energy, education, recruitment, public services and other fields. Taking the cooperation with Yum! Brands as an example, the current AI customer service applications and solutions have covered Yum! Brands' entire business line. The peak daily call volume of large models reaches millions, and the "problem solving rate" of customer service robots has increased by 90%.
Building an intelligent agent is similar to building a website in the PC era or creating a self media account in the mobile era. The difference is that agents are more like people and more intelligent. Robin Lee speculates that agents may become new carriers of content, information and services in the AI native era.
He gave an example that after searching for the keyword "education and tutoring" on Baidu, these digital people can be seen on the search results page. These digital individuals are more natural and able to pause at the appropriate time to respond to questions raised by netizens on site. In today's digital live streaming, in many cases, the conversion rate has exceeded that of real people
For example, the tool based intelligent agent "Free Canvas" jointly created by Baidu Wenku and Baidu Netdisk allows users to freely drag and drop rich media materials such as documents, audio and video on a canvas like interface, generating multimodal content. The legal intelligent agent "Faxingbao" has answered 16.6 million legal questions from users, not only providing answers like professional lawyers, but also calculating legal compensation amounts, writing legal documents, and recommending suitable human lawyers.
In Robin Lee's opinion, the low threshold and high ceiling of agents can not only make everyone get started, but also make complex and powerful applications. Intelligent agents are the most mainstream form of AI applications and are about to reach its tipping point