At the "Baidu World 2024" conference, Baidu's Chief Technology Officer, Wang Haifeng, announced that the user base of Ernie AI has reached 430 million. Since its launch on March 16th of the previous year, Baidu's Ernie large model has undergone continuous upgrades, with technologies such as knowledge enhancement, retrieval enhancement, and intelligent agents evolving constantly. Wang Haifeng detailed the latest technological advancements and applications of the Ernie model, including retrieval-enhanced text-to-image technology, the no-code product Miaoda, and Ernie QuickCode.
In terms of retrieval enhancement, Baidu has developed a "comprehension-retrieval-generation" collaborative optimization technology, enhancing the effectiveness of large model technologies and applications. The text-to-image technology can generate precise images and also perform generalized image generation. Baidu also introduced an intelligent agent mechanism and developed System 2, whose core is a thinking model that includes understanding, planning, reflection, and evolution, capable of reliable execution, self-evolution, and to some extent, making the thinking process transparent.
Wang Haifeng also shared applications of multi-agent systems in coding scenarios—Miaoda and Ernie QuickCode. Miaoda employs a centralized agent model, improving the quality and stability of complex code generation. Ernie QuickCode has been upgraded to version 3.0, showcasing capabilities for complete system development, code testing, and issue fixing. Throughout the entire development process, multiple agent pipelines collaborate, significantly enhancing the quality and efficiency of programmers' work.
Baidu has crafted a matrix of Ernie large models tailored to industrial needs, including flagship models like ERNIE 4.0 Turbo, lightweight models like ERNIE Speed, and a series of thinking and scenario models based on foundational models to meet diverse application requirements. The Ernie model and the PaddlePaddle deep learning platform have been jointly optimized, improving model effectiveness, training efficiency, and inference performance. Wang Haifeng revealed that the number of PaddlePaddle Ernie developers has reached 18.08 million, serving 430,000 enterprises and creating 1.01 million models.
Baidu also released the latest data on AI talent cultivation, having cultivated 5.92 million AI professionals for society. Wang Haifeng disclosed that the Ernie model is still undergoing continuous training, with more powerful new versions on the horizon.