The day before, he broadcast live in the Jiyue car, and the next day he "called" for the large model tool, and Baidu CEO Robin Li seamlessly connected. On April 16, he spent about an hour explaining Baidu's latest AI development tools, once again complaining about the backwardness of open source models, and stated that "the biggest application scenario of visual large models is autonomous driving".
As in previous years, Baidu's executives showed off data, features, and the future at the Create conference, but the industry landscape is very different. In March 2023, Baidu preemptively released the Wenxin model and product Wenxin Yiyan, and in October of the same year, it came up with the latest version of Wenxin Model 4.0, with more than 200 million users in one year. The competitors are not slow to move, the dark side of the moon has sprung up, Sora's performance is shocking, and Wenxin Yiyan's position as the leader of AI applications is threatened. Robin Li said that this year, Baidu developed AI native applications and stepped on the pit and paid tuition. On the same day, Baidu opened on the Hong Kong Stock Exchange at HK$95 and its share price closed at HK$94.15.
Three tools
At 9 o'clock in the morning, Robin Li started to speak, and the title of "Everyone is a Developer" is very suitable for the developers and technology enthusiasts in the audience.
"In the past, developers used code to change the world, but in the future, natural language will become the new general-purpose programming language, and you can become a developer as long as you can speak," Robin Li emphasized. ”
Based on the above judgment, Baidu has launched new "development tools": AgentBuilder for agent development, AppBuilder for AI native application development, and ModelBuilder, a model customization tool of various sizes.
Robin Li believes that intelligent twins may be the most recent and mainstream way for everyone to use large models in the future. More than 30,000 agents have been created, more than 50,000 developers and tens of thousands of enterprises have settled on Baidu's upgraded Wenxin intelligent twins platform.
The progress of the agent is also the content that Wang Qinglin, manager of Ruidaheng Research Institute, is concerned about, "It has a broader development prospect and can be applied to various scenarios on the C-end for content generation." Due to the low threshold for use of this model and the many scenarios, it can take into account work and daily life, individuals and enterprises, and can involve multiple links in many formats, which seems to have more development potential, but how to monetize business is a challenge."
AppBuilder is used to develop AI applications, which require three steps: giving the application a name, describing the specific requirements in the role instructions, and adding a code interpreter to the tool assembly to help with the operation. ModelBuilder can customize models of any size according to the needs of developers, and further fine-tune the SFT model according to the subdivision scenario.
When demonstrating this function,Robin Li asked Xiaodu a series of questions,Including news、Schedule, etc.,"It looks like it's communicating with the same,In fact, different models are called behind it",Some are Wenxin large models4.0,Some are fine-tuned models based on ERNIE Lite。
Don't look at open source
"We can tailor a smaller model suitable for various scenarios according to needs, taking into account various considerations such as effect, response speed, and inference cost, and support fine-tuning and post pretrain (a transfer learning method for pre-trained models). In this way, the model cut out through dimensionality reduction is significantly better than the model directly retrieved from open sources, and the effect is significantly better under the same size, and the cost is significantly lower under the same effect. Robin Li further said, "Everyone used to think that open source was cheap, but in fact, in the large model scenario, open source is the most expensive. So the open source model will fall further and further behind. ”
This is not the first time he has complained about open source large models. According to reports, Robin Li recently said in an internal speech that the closed source of large models has a real business model, which can make money, and can eventually gather computing power and talents.
Coincidentally, this month, Alibaba Cloud Tongyi Qianwen disclosed the latest open source progress, open-sourced 32 billion parameter model Qwen1.5-32B, and previously open-sourced 6 large language models with 500 million, 1.8 billion, 4 billion, 7 billion, 14 billion and 72 billion parameters. A relevant person from Alibaba Cloud told a reporter from Beijing Business Daily that the above large models have been upgraded to version 1.5. Compared with the Tongyi Qianwen 14B open-source model, the 32 billion parameter model is more capable in agent scenarios, and the inference cost is lower than that of the Tongyi Qianwen 72B open-source model.
Robin Li also emphasized the cost, "Through the combination of this large and small model scheduling method, Xiaodu has completedthe'brain change action',InstalledAIThe native operating systemDuerOS X,Compared with the flagship version that uses all Wenxin large models,The response speed is increased by 2 times,The cost is reduced99%。
In fact, the combination of large and small models is not unique to Baidu. "Today, a lot of open source software gives a model of tens of billions (level), which cannot be compared with ChatGPT, but in the case of sufficient internal enterprises, its deployment cost can be reduced from the original tens of millions of dollars to tens of millions of yuan or even millions of dollars. I think this really achieves the goal of 'pulling the big model off the altar'. Zhou Hongyi, founder of 360, said in August 2023.
Cognition and landing
Another business that will always occupy a large part of the Baidu Create conference is related to autonomous driving and automobiles.
On April 15, 2024, Robin Li emphasized that "car robots are equivalent to drivers and assistants" and "large models understand everything, and they can finish when they talk".
On April 16, Robin Li mentioned driving on the multimodal large model. In his view, "multimodal large models, or the integration of text, pictures, voice, video, etc., is a very important long-term development direction of the basic model, and it is the only way to AGI (artificial general intelligence)." The biggest application scenario of the vision model is autonomous driving."
"There is no problem with Robin Li's cognition, but correct cognition does not mean that it can be successfully implemented", Wang Chao, founder of Wenyuan Think Tank, told Beijing Business Daily, "Baidu has a good accumulation and start in artificial intelligence, but it lacks the ability to commercialize it on a large scale."
In 2023, Baidu's revenue will be 134.598 billion yuan, an increase of 9% from 2022, and its non-GAAP net profit will be 28.7 billion yuan, an increase of 39% from 2022. According to the information revealed in Baidu's conference call, benefiting from general artificial intelligence and basic models, Baidu's cloud business generated 656 million yuan in revenue in the fourth quarter of 2023, and this figure will grow to billions of yuan in the whole of 2024.
Look at the data from a different perspective. In December 2023, the monthly active users of Baidu App were 667 million, and the number was 677 million in June 2023.
According to data from QuestMobile, in the field of AIGC (artificial intelligence generated content), in January 2024, Douyin's Doubao, Baidu's Wenxin Yiyan, and Kunlun Wanwei's Tiangong are the top three apps in China in terms of monthly active users, with 17.569 million, 12.125 million, and 8.014 million monthly active users respectively. With more than 200 million users, what about the monthly active data? A reporter from Beijing Business Daily interviewed relevant people from Baidu on Wenxin Yiyan's user data, "which has not been announced to the public," the person said.
Beijing Business Daily reporter Wei Wei