The big model basically eliminates hallucinations, super applications have not appeared, smart partiesIt is the most mainstream form of AI applications to usher in an explosion point ... World 2024 conference held on November 12, Li Yanhongs many latest research on AI.
Basic elimination hallucinations
"In the past 24 months, the AI industryWhat is the biggest change? Basically, the big model is basically eliminated. "On November 12, Li Yanhong said at the Baidu World 2024 conference.
"Phantom" is a special language for AI models. Phantom problems are one of the biggest restrictions widely used in large models.Earlier many AIGC products liked "a serious nonsense", which caused AI output to be credible.
As of the beginning of November, the average daily call volume of Baidu Wenxins model exceeded 1.5 billion, compared with the 200 million disclosed in May, an increase of 7.5 times.50 million times, an increase of about 30 times.Li Yanhong said that "this growth rate exceeded expectations", indicating that AI is a real demand.He lamented that this steep growth curve represents the outbreak of Chinas large model application in the past two years.
In the development trend of AI applications, Li Yanhong said that intelligence is the most mainstream form of AI applications and is about to usher in an explosion point.He explained the four major categories of intelligence in the company, characters, tools and industry, and will be a smart body ratio to be a website in the PC era, or the self -media account in the mobile era.It is more like your sales, customer service and assistant.P>
At the meeting, Baidu released the two major AI technologies of Wenxin IRAG and non -code "seconds".Wenxin IRAG is used to solve the illusion of large models in picture generation, which greatly improves practicality; "seconds" without code technology allows everyone to have the ability of programmers, and will create millions of "super useful" applicationsEssence
The technique behind the solution to the problem of the fantasy of text is RAG, that is, the retrieval enhancement.At present, the RAG at the text level has been done well, and basically eliminates hallucinations for large models; but in terms of multi -mode state such as images, the combination of RAG is not enough, which is also the direction of Baidus hope of breaking through.
At present, Wen Shengtu based on large language models also has more serious hallucinations, especially for the Great Wall, Oriental Pearl, Pyramid, Einstein, Beethoven and other specific locations, items and characters, Often occur the illusion of Zhang Guanli, which allows the generated pictures to "see a fake", which affects the practicality of AI.
In response to the above problems, Baidu has developed the Irag (Image Based RAG) retrieved and enhanced.Generate a variety of super -real pictures, "the overall effect is far more than Wenshengtu native system, and the smell of machine is removed."
Li Yanhong showed a picture of the public to patrol the Great Wall from the Great Wall of the Great Wall of the Great Wall.Through Wenxin Irag technology, no matter whether this specific model of the car is the model of the car or the Great Wall of the background, there is no illusion of errors or deformation.At the same time, the combination of Einstein and various background attractions in the "Einstein Traveling World" picture is also highly similar to the real world, and the texture is close to the photo.
Eliminate the hallucinations of large models, which is also the basis for the outbreak of AI applications.In Li Yanhongs view, today, the basic model is ready, and the stars who are about to usher in AI applications are shining.
In addition, the codeless tool "Second" is a software that can realize arbitrary ideas without writing code, covering unclear programming, multi -intelligent collaboration, multi -tool calls, etc.Features, you can build a variety of applications as long as you speak.It can help more people and more companies create millions of "super useful" applications.Li Yanhong concluded that this means that everyone can direct multiple smart bodies themselves to cooperate to complete the task. "As long as you have ideas, you can think of things. We will usher in an unprecedented, and you can make money by thinking about your thoughts.Time.All the top technology companies in the world are paying attention to smart bodies, but there are not many companies like Baidu who uses intelligence as the most important strategic direction.In the native era, intelligence will become a new carrier of content, service and information.
He uses corporate intelligence as an example. In the traditional PC official website model, companies can only display the companys introduction and product parameters, but they are lacking actively recommending, timely recommendation, timely timeResponse and one -to -one service capabilities; while the companys intelligence can recommend the corresponding products according to the needs of customers. In terms of service, it can also respond more directly and quickly to demand, which can greatly improve the efficiency of interactive marketing.After BYDs official intelligence was launched, the sales conversion rate increased by 119%, and the September interaction rate of Lenovo AIPC Intelligence increased by 89%.
In addition, Li Yanhong showed the characteristic functions and use scenarios of many types of intelligence such as characters, tools and industries.For example, the "free canvas" jointly created by Baidu Library and Baidu Web Disk, allowing users to drag the wealthy media materials such as "canvas" on a similar "canvas" interface to quickly generate multi -modal content.
"Smart body is the most mainstream form of AI applications, and it is about to usher in its explosion point." In Li Yanhongs view, the threshold of the smart body is low and the ceiling is high, which can make people not only make people make people not only allow peopleEveryone can make complex and powerful applications.On the same day, he demonstrated the TOP100 intelligent body of the Wenxin Smart Platform, which has both the character categories such as farmers academicians, as well as the smart bodies of various scenarios such as tools, industry, workplace, emotion, and entertainment.
Responsible editor: Peng Bo
School pair: Yang Shuxin
