Robin Lee, chairman and CEO of Baidu, remembered that the first time he came to the World Artificial Intelligence Conference (WAIC) was in 2022. The theme of that conference was related to the metauniverse. The sponsor sent a message to him to tell him about the metauniverse. He replied, "I said I should talk about AI, I can't talk about the metaverse.". At that time, he mentioned in his speech his judgment on the trend of AI: the technological development path of AI has undergone a directional change, shifting from discriminative artificial intelligence in the past to future generative artificial intelligence.
The above prediction occurred in the summer of 2022, five months after the release of GhatGPT, followed by another wave of artificial intelligence sweeping the world at an unprecedented speed.
"In two years, the whole world has changed, and AI can be said to have subverted the cognition of most people." Robin Lee sighed in today's WAIC speech.
He believes that in 2023, there will be a hundred model wars in China, which will cause huge waste of social resources, especially computing power. However, it will also help us establish our ability to catch up with the world's most advanced technological models.
He said that when Baidu announced the release of Wenxin 4.0 in October last year, it was mentioned that Wenxin 4.0 was no less than GhatGPT, which was not widely accepted. However, many closed source models in China now claim to have caught up and surpassed GhatGPT.
He mentioned that there are also some laymen who confuse the concepts of model open source and code open source. "Open source models require a lot of parameters, but it's still necessary to do secure alignment. If you don't know how these parameters come from, you won't be able to achieve the same level of success as others. Even if you get the corresponding source code, you don't know how much data was used or how much proportion of data was used to train these parameters. Therefore, obtaining these things doesn't allow you to iterate and develop on the shoulders of giants."
In today's speech, he once again mentioned the debate between open source and closed source approaches in the field of large models. He believes that under the same parameter scale, closed source models have better capabilities than open source models. If open source wants to have the ability to catch up with closed sources, it needs to have larger parameters, which means that the inference cost will be higher and the reaction speed will be slower.
"A lot of people use open source models to change their models, thinking that they can better serve personalized applications. But they don't know that creating orphan models like this can't benefit from the basic model, nor can they share computing power with others," Robin Lee said.
However, at the same time, he also acknowledges the value of open source big models in specific scenarios such as academic research and teaching, but they are not suitable for most application scenarios. "When you are in a fiercely competitive market environment, you need to make your business more efficient and cost-effective than your peers. At this time, commercial closed source models are the most effective."
Compared to the open-source and closed source approach, he believes that the more important thing is the implementation of applications. He mentioned that the focus of the industry is now on the basic model, "running scores and brushing rankings all day long, who is who surpasses GhatGPT4, and Open AI comes out Sora..." However, in fact, without applications, having a basic model, whether open or closed, is not worth a penny.
He called on the industry not to roll up models anymore, but to roll up applications. Applications are actually not far away from everyone. Based on benchmark models, they have gradually penetrated into various fields and industries. He cited the call volume data of Wen Xin Yiyan, which was still 200 million two months ago, and now has reached 500 million, indicating that the big model represents the real demand, and some people really benefit from the big model.
He gave an example in the field of express delivery, where large models were used to help process orders, achieving the goal of sending packages one by one without the need for other cumbersome processes. The time was shortened from 3 minutes to 19 seconds, and over 90% of problems were solved by large models, resulting in a significant improvement in efficiency. In the field of novel creation, online authors have gained wings like tigers with the help of AI. And software similar to code generation is gradually penetrating into various fields. Robin Lee revealed that about 30% of Baidu's internal code has been generated by AI, and the code adoption rate has exceeded 44%.
Therefore, Robin Lee proposed that the industry should avoid falling into the "super application trap", and believed that the success of an app with 1 billion DAUs must be achieved. This is the logic of thinking in the mobile era. In the AI era, the rule may not be the same. "Super capable" applications are more important than DAU's "super applications". As long as they can generate large gains in industry and application scenarios, the overall value is much greater than that of mobile Internet.
Intelligent agents are his most promising direction for AI development. He judged that in the future, various intelligent agents will be created in fields such as healthcare, finance, education, manufacturing, transportation, agriculture, etc. based on their own scenarios, unique experience rules, data, etc. There will be millions of intelligent agents appearing in the future, forming a huge intelligent agent ecosystem.
AI is penetrating into various industries at an unprecedented speed. Many people are worried that humans will no longer have job opportunities? Robin Lee said that he heard a lot of complaints, few constructive opinions, and few people were committed to exploring new job opportunities brought by generative AI.
Robin Lee said that he has seen some new work emerging, such as the word engineer. This profession does not require programming in the future, but requires strong logic to articulate workflows and use prompt words to tune models. With the emergence of a large number of intelligent agents, this job demand will also soar. "These job opportunities usually have a relatively low threshold, and what you do can generally support your family. If you do well, the upper limit can be an annual salary of one million."