Robin Li: After stepping on countless pits, Baidu foresees the future like this

On April 16, Create 2024 Baidu AI Developer Conference was held in Shenzhen. The reporter of Nandu Bay Finance Society learned on the spot that Robin Li, founder, chairman and CEO of Baidu, delivered a keynote speech on "Everyone is a Developer", depicting a world that is no longer limited to coding skills, but uses natural language as a medium and everyone can participate in the creation of an era.

At a time when the competition between domestic and foreign enterprises is gradually "white-hot", this day seems to be no longer far away. In the past year, Baidu's Wenxin model has undergone two phases of upgrade, from version 3.0 to version 3.5, and finally to version 4.0, driving a significant improvement in general capabilities such as code generation, interpretation, and optimization.

At the meeting, Robin Li shared the progress of version 4.0 of Wenxin model with more than 5,000 technology enthusiasts and practitioners present, pointing out that compared with a year ago, the algorithm training efficiency of Wenxin model has been increased to 5.1 times, the average weekly training efficiency has reached 98.8%, the inference performance has been improved by 105 times, and the inference cost has been reduced to 1%. This also means that customers can now call 1 million times for the same cost of 10,000 calls.

Pan Helin, a member of the Information and Communication Economy Expert Committee of the Ministry of Industry and Information Technology, said in an interview with a reporter from Nandu Bay Finance Society that the Wenxin model is the most complete model in the current domestic ecology, and through continuous accumulation and development, the Wenxin model is outstanding in some special application scenarios, such as the field of AI programming, which further expands and locks the potential user group of the Wenxin model, so that users are more willing to pay for the model, because programming ability is actually a kind of productivity.

"We see that the key to the competition of domestic large models is to occupy key applications, such as some AI for drawing, some AI for editing, some AI for virtual people, copywriting, PPT, such as Wenxin model for code programmers, the application advantages of these large models in the vertical field will become the cornerstone of the future development and growth of large models. Pan and Lin added.

Robin Li favors small models: more cost-effective

"I stepped on countless pitfalls and paid high tuition fees. When talking about the specific ideas and tools for developing AI native applications based on large models, Robin Li sighed. But it is undeniable that Baidu has indeed explored a unique path on the road of large AI models.

At the meeting, Robin Li shared some specific ideas for developing AI-native applications based on large models, including MoE, small models, and agents.

Robin Li said that in the future, large-scale AI native applications will basically be based on the MoE architecture, that is, a mixture of large and small models, and do not rely on one model to solve all problems. The second is that small models are more cost-effective. "The inference cost of the small model is low, the response speed is fast, and in some specific scenarios, the small model after SFT fine-tuning can be used as well as the large model. "The third is that the development threshold of the agent is greatly reduced, and with the improvement of the ability of the agent, a large number of new applications will continue to be born. It allows machines to think and act like humans, complete complex tasks autonomously, and achieve self-iteration and evolution.

In response to the change of Baidu's basic development ideas, today's Wenxin large model series, in addition to the flagship version of ERNIE3.5, ERNIE4.0, also includes the lightweight version of ERNIE Speed, Lite, Tiny, etc. "It's not that the large model can't afford to use, but the small model is more cost-effective", which is also within the scope of Baidu's consideration for the needs of the audience.

At this stage, in the hottest AI track, the battle between large and small models has begun to take shape, and large models have become "smaller and smaller". It is observed that in December last year, Google launched three specifications of Gemini in one go: Ultra, Pro and Nano, of which the smallest Nano can run directly on mobile devices, with two versions of 1.8 billion parameters and 3.25 billion parameters. Subsequently, Microsoft also launched the Phi-2 model with only 2.7 billion parameters in December, and it not only surpassed Mistral-7B in performance, but also did not fall far behind the 70 billion parameter version of Llama 2.

However, there is always a tacit rule that large models have a tacit rule compared to small models – the larger the number of parameters, the better the performance. Among them, the most representative are OpenAI's GPT3.5, GPT4 series, Google's Gemini series, etc., In a certain specific scenario, the power of computing power enables large models to handle more detailed and complex tasks, and it can also show higher accuracy in tasks such as prediction and classification, which is still a "hurdle" that many small models cannot cross.

AI applications developed based on large models are more valuable

In addition, Robin Li once again emphasized the previous view, "The large model itself does not directly create value, and the AI applications developed based on the large model can meet the real market demand." ”

Based on this, in the three directions mentioned above, MoE, small model, and agent, Baidu has done a good job of "out-of-the-box" tools. Robin Li released the three major development tools in his keynote speech: AgentBuilder, an AI native application development tool, and ModelBuilder, a model customization tool of various sizes.

Robin Li mentioned that up to now, more than 30,000 agents have been created, more than 50,000 developers and tens of thousands of enterprises have settled in. "Today, every merchant and every customer can have their own intelligent twins in Baidu. The whole process does not require programming at all, through the input of information similar to prompt words, and simple operation and optimization, you can quickly generate an agent and become a 7X24 hours online gold salesman. ”

Taking the launch of Sophia's merchant agent as an example, the data shows that its effective lead cost has dropped by 30%. That is, it acquires an effective customer, and if the cost used to be 100 yuan, now it is only 70 yuan.

As for AppBuilder, Robin Li attributes it to an AI-native application development tool, "In just three steps, developers can develop an AI-native application in natural language, and it can be easily published and integrated into a variety of business environments." ”

The reporter of Nandu Bay Finance Society observed at the scene and found that an AI native application can be created in three simple steps: name setting, filling in role instructions, and inserting components. Robin Li showed the audience the creation process of an AI-native application through three cases: "Playground Queuing Assistant", "Huadian AI Assistant" of North China Electric Power University, and Baidu Wenku Intelligent Comic Generation.

He also pointed out that AppBuilder has two major advantages: one is powerful, such as its Q&A accuracy and friendly response rate of more than 95%, and the other is that it is easy to use, and it only takes three steps to quickly create an app and distribute it with one click.

Finally, for professional developers, Baidu launched ModelBuilder. It can customize the model of any size according to the needs of the developer, and further fine-tune the SFT of the model according to the subdivision scenario, so as to achieve better results.

Written by: Nandu Bay Finance Agency reporter Yan Zhaoxin intern Chai Jia