Author: Congbin Wang

As the world witnessed the first-ever AI Developer Conference, Baidu’s Create has successfully hosted five editions to date. Contrary to just showcasing updates about the Wenxin large-scale model, this conference pinpointed a focus on the developer community, aiming to assist developers in efficiently utilizing the Wenxin model.

“Everyone is a Developer” was the theme of Baidu’s founder, chairman, and CEO, Robin Li’s speech this year. He believes that large models and generative AI will revolutionize the developer community.

In the past, developers used code to change the world. Moving forward, natural language will become the universal programming language. If you can speak, you can become a developer and use your creativity to change the world.

Baidu’s founder, chairman, and CEO, Robin Li

Moving Forward with the Practical Experience of Wenxin Yiyan

Robin Li first shared the annual achievements since the launch of Wenxin Yiyan: reaching over 200 million users, with more than 200 million daily API calls, serving 85,000 users, and over 190,000 AI-native applications developed on the Qianfan platform.

The algorithm training efficiency of the Wenxin model has now increased by 5.1 times, with an average weekly training effectiveness of 98.8%, and inference performance has improved by 105 times.

This means customers, who initially made 10,000 calls a day, can now make a million calls at the same cost.

Thanks to Baidu’s full-stack layout across chips, frameworks, models, and applications, and through end-to-end optimization, the cost of inference has been reduced to 1% of its original.

Li also emphasized that large models do not create value directly. It’s the AI applications developed based on these models that meet the real market demand.

At the conference, Baidu officially released the Wenxin large model 4.0 toolkit edition, including document Q&A, data analysis, and code execution.

Li also shared the valuable experience gained through numerous challenges and high costs.

Firstly, the future AI-native applications of large models will adopt an MoE architecture: a mix of large and small models tailored to different application scenarios;

Secondly, in specific scenarios, finely-tuned small models can perform as well as large models: compressing and distilling a basic model from a large one, then training with data, is much more effective than starting from scratch;

Thirdly, the enhancement of intelligent agent capabilities will continuously generate a plethora of new applications: including understanding, planning, reflecting, and evolving, this mechanism enables machines to think and act like humans, autonomously completing complex tasks, learning persistently, and achieving self-iteration and evolution in their environments.

Instantly Accessible, Three Major AI Development Tools

For MoE, small models, and intelligent agents, Baidu provides “ready-to-use” development tools, including AgentBuilder for intelligent agents, AppBuilder for AI-native application development, and ModelBuilder for customizing models of various sizes.

AgentBuilder

Intelligent agents might be the most mainstream way to use large models in the future. Based on a strong foundational model, intelligent agents can be generated in batches for diverse scenarios.

Today, every merchant and customer can have their unique intelligent agent on Baidu. The whole process requires no programming. With inputs like prompts and a few simple steps for optimization, an intelligent agent can be quickly created to become a 24/7 online top salesperson.

To date, over 30,000 intelligent agents have been created, with more than 50,000 developers and thousands of enterprises joining the platform. Currently, more than 10,000 Baidu clients have their own business intelligent agents, covering over 30 industries including education, real estate, machinery, and business services.

Moreover, the Wenxin intelligent agent platform also offers developers a pathway to monetize traffic, not only through Baidu search but also integrating intelligent agent capabilities into other products within Baidu’s ecosystem, such as Xiaodu, Maps, Tieba, and vehicle systems, addressing developers’ traffic distribution issues.

AppBuilder

AppBuilder significantly lowers the development barrier by encapsulating and pre-setting various components and frameworks needed for developing AI-native applications. With just three steps, developers can use natural language to create an AI-native application and easily publish and integrate it into various business environments.

“AppBuilder stands out for its powerful functionality and ease of use,” said Li.

ModelBuilder

Designed for professional developers, ModelBuilder allows customization of models of any size based on developers’ needs and further fine-tunes models for specific scenarios to achieve better results.

ModelBuilder is capable of producing models efficiently and cost-effectively. Li highlighted that the smaller-sized models generated from the powerful Wenxin 4.0 are significantly more effective and less expensive than those obtained directly from open sources at equivalent sizes.

“The open-source model will increasingly fall behind,” Li stated, debunking the myth that open-source is cheap, particularly in the context of large models, where it turns out to be the most expensive.

ModelBuilder offers the most comprehensive and diverse selection of large models, totaling 77, making it the development platform with the most extensive collection of large models in China. It includes flagship models ERNIE3.5 and ERNIE4.0, three lightweight models ERNIE Speed, Lite, Tiny, and two models tailored for specific scenarios—ERNIE Character and ERNIE Functions, suitable for role-playing and conversation or Q&A scenarios, respectively.

During the event, Baidu’s Executive Vice President and President of Baidu Intelligent Cloud, Shen Dou, also launched the new generation intelligent computing operating system “Wanyuan,” covering a heterogeneous hardware architecture, cloud-native platform, Baidu Baige, ModelBuilder, AppBuilder, and AgentBuilder.

In conclusion, Li particularly emphasized the role of developers in the AI era, stating that today everyone is a developer and creator.