On the evening of July 4th, Li Yanhong, founder, chairman and CEO of Baidu, said at the main forum of the 2024 World Artificial Intelligence Conference and the AI Global Governance High-Level Meeting on Industrial Development that he should jump out of the thinking logic of the mobile era and avoid falling into the "super application trap". Not only 1 billion DAU applications are successful. At the same time, Li Yanhong pointed out that the intelligent body is the easiest AI application to develop, and it is also the development direction that we are most bullish on for AI applications.

In the speech, Li Yanhong once again talked about open source and closed source large models, saying that open source large models have value in specific scenarios such as academic research and teaching fields, but they are not applicable to most application scenarios. "When you are in a fiercely competitive environment, you need to make your business more efficient and cheaper than your peers. At this time, the commercialized closed-source model is the most playable."

Li Yanhong pointed out that open source and closed source large models have been a controversial topic since this year, but many people confuse the concepts of model open source and code open source. He pointed out that model open source can only get a bunch of parameters, and also needs to do SFT and security alignment. Even if you get the corresponding source code, you don’t know how much of it, and what proportion of the data was used to train these parameters, and you can’t iterate development on the shoulders of giants.

He bluntly said that under the same parameter scale, the ability of open source models is inferior to that of closed source models. "If open source wants to catch up with closed source in terms of ability, it needs to have a larger parameter scale, which means higher inference cost and slower response speed. Many people use open source models for modification, thinking that this can better serve their personalized needs, but they don’t know that this has become a standalone model, which cannot benefit from the continuous upgrading of the basic model and cannot share computing power with others.

Li Yanhong admitted that open source models have value in some academic research and teaching fields and can be used to study the working mechanism of large models and form theories. But open source models are not applicable to most application scenarios. In a fierce business environment, if you want to make your business more efficient than your peers and lower than your peers, the commercialized closed-source model is the "most playable." He cited Baidu's practice in novel creation as an example. When it turned from open source model to lightweight model and then to Wenxin large model 4.0, the availability and quality rate of novel generation were greatly improved, making online literature authors more powerful.

In his opinion, the focus of large models is still on "rolling applications", and "there is no application, there is only a basic model, whether it is open source or closed source, it is worthless."

Regarding AI applications, Li Yanhong pointed out that in the AI era, "super capable" applications are more important than "super applications" that only look at DAU, and "we must avoid falling into the "super application trap" and think that an APP with 1 billion DAU must be successful, this is the thinking logic of the mobile era." He believes that as long as it can bring great benefits to the industry and application scenarios, the overall value is already greater than the mobile Internet.

He cited the express delivery industry as an example. By using the capability of large models to process orders, express delivery companies have achieved "one picture, one sentence to send express delivery", eliminating the need for other cumbersome processes, and the time has been shortened from more than three minutes to 19 seconds. "Moreover, more than 90% of after-sales problems are solved by large models, and the efficiency improvement is very significant."

In more general fields such as code generation, Li Yanhong said that Baidu's Wenxin Kuaicoding has gradually penetrated into various fields. Within Baidu, about 30% of the code has been generated by AI, and the code adoption rate has exceeded 44%.

"Applications are not far away from us," Li Yanhong said, and applications based on basic models have begun to gradually penetrate into various industries and fields. Two months ago, Baidu announced that the daily call volume of Wenxin large model exceeded 0.2 billion. Recently, the call volume has exceeded 0.5 billion. "In just two months, the call volume has undergone such a big change, which shows that it represents real demand, and someone is using it and really benefiting from large models, gaining value." (Wen Meng)