At the start of 2023, large-scale AI models emerged as a beacon of promise, illuminating the vast potential for artificial intelligence technology to drive transformative changes in the economy and society. After more than a year of technical advances and market refinement, a proliferation of such models has surfaced, their capacity for innovation commanding attention. Today, technological superpowers like China and the United States stand at the forefront of this era, spearheading the robust growth of the large-scale model industry.
In our country’s tech sector, Baidu has taken the lead with the launch of its Wenxin large-scale model, achieving remarkable results. At the recent Create2024 Baidu AI Developer Conference, Baidu’s founder, chairman, and CEO, Robin Li, announced that the user base of Wenxin Yiyan has now exceeded 200 million, making Wenxin the most advanced and widely applied AI foundational model in China.
Furthermore, Robin Li expressed that in the future, users of large-scale models will be able to become developers of these same models. The intersection of large-scale models and generative AI is set to radically transform the developer community. “In the past, developers used code to change the world; in the future, natural language will become the new universal programming language, enabling anyone who can speak to become a developer and use their creativity to change the world.”
01
Release of Wenxin Large-Scale Model 4.0 Tools Edition
Empowering Every User to Become a Developer
On March 16, 2023, Wenxin Yiyan was officially launched. Just a year and a month later, the user count has surpassed 200 million, with a daily average of 200 million API calls, serving 85,000 customers, and the number of AI-native applications developed using the Qianfan platform has exceeded 190,000.
The underlying support for Wenxin Yiyan is the Wenxin Large-Scale Model. Compared to a year ago, the algorithmic training efficiency of the Wenxin Large-Scale Model has increased by 5.1 times, with an average weekly training efficiency of 98.8%, and its inferencing performance has improved by 105 times, reducing the cost of inference to 1% of the original. Today’s launch of Wenxin Large-Scale Model 4.0 Tools Edition provides users with an advanced “code interpreter” function, enabling them to process and analyze complex data and files through natural language interaction, truly achieving improved performance, reduced costs, and an optimized user experience.
Robin Li once again called for the robust development of AI-native applications in the industry: “Large models themselves do not create value directly—it is the AI applications developed based on large models that can meet the real market demands.”
02
Achieving “Ready to Use” Status
Baidu Releases Three Major AI Development Tools
Over the past year, Baidu has encountered numerous challenges and paid significant tuition fees on the path of developing AI-native applications. Robin Li articulated that Baidu’s goal has always been to provide users with efficient developmental tools to boost societal creativity. He pointed out three key aspects of developing AI-native applications based on large models.
One is MoE—Mixture of Experts. Most AI-native applications will utilize a mix of large and small models rather than relying on a single model to solve all problems.
Two is the use of smaller models. These have low inference costs and fast response times. In specific scenarios, finely tuned small models can perform comparably to large models after fine-tuning.
Three is the concept of agents. Agent mechanisms include understanding, planning, reflection, and evolution, allowing machines to think and act like humans, autonomously completing complex tasks, continuing to learn within an environment, and achieving self-iteration and evolution, thus spawning more novel applications.
In these three areas, Baidu has introduced “ready to use” development tools: AgentBuilder for intelligent agent development, AppBuilder for AI-native applications, and ModelBuilder for customizing models of various sizes.
By using AgentBuilder, merchants can swiftly create an intelligent agent from a natural language prompt, which can be further optimized to become a “gold medal salesperson” available online 24/7. Additionally, the Wenxin intelligent agent platform linked to the merchant’s agent provides developers with a channel for monetization, ensuring developers do not worry about traffic distribution, and driving a virtuous cycle of distribution-diagnostic-revenue for better quality, traffic, and profits.
AppBuilder greatly lowers the barrier to entry for developing AI-native applications by pre-packaging various components and frameworks required, as Robin Li explains. “In just three steps, developers can create an AI-native application using natural language, and easily publish and integrate it into various business environments.” Besides integrating Baidu’s rich AI components, AppBuilder also supports open-source SDKs to facilitate secondary development.
ModelBuilder allows developers to create models of any size according to their needs and to fine-tune models for specific scenarios SFT (Sparse Fine-Tuning), enabling cost-effective and efficient model production. In addition to Baidu’s ENRIE 3.5 and 4.0 flagship general-purpose large models, several lightweight and industry-specific large models, ModelBuilder also supports 77 models from both domestic and international third parties, making it the domestic platform with the most extensive array of large models.
“1 billion users are also 1 billion developers,” Robin Li concluded, asserting that the creative power unleashed by AI will also accelerate the arrival of the AI era. “Using natural language as the universal programming language, everyone can use their creativity to change the world.”
END
Author: He Jia
Editor/Design: Bei Beibei
Reviewers: Wang Tao, Mei Yaxin
Producer: Liu Qicheng
[Communication World New Media Matrix]
Central Videos | Microsoft MSN | Video Channels | Weibo | Today’s Headlines | Baijiahao | NetEase | Sohu | Tencent News
Sina Points | Xueqiu | TikTok | Kuaishou | iQIYI | Zhihu | Bilibili | Migu Video | CSDN | 36Kr
[New Media Team]
Producer | Liu Qicheng
Reviewers | Wang Tao, Mei Yaxin
Editors | Wang Yurong, Bei Beibei
Video Editors | Huang Yangyang, Lu Ruixu
Operation | Lin Song