Baidu CTO Wang Haifeng Discusses Latest Advances in Large Models and the PaddlePaddle Platform at the 2024 Zhongguancun Forum

On April 27, 2024, the 2024 Zhongguancun Forum “Future AI Pioneer Forum” was held in Beijing. Baidu Chief Technology Officer Wang Haifeng attended and gave a speech, outlining the latest directions of large‑model development and announcing recent achievements of the Wenxin large model and the PaddlePaddle deep‑learning platform.

Wang noted that since the full launch of Wenxin Yiyan on August 31 last year, the user base and daily call volume have surpassed 200 million, serving 85,000 customers and more than 190,000 AI‑native applications.

He emphasized that agents are a crucial direction for large models, bringing about a wave of application breakthroughs. An agent is built on a foundation model and undergoes enhanced “thinking” training, including supervised fine‑tuning of the reasoning process, preference learning for decision making, and reinforcement learning for result reflection, resulting in a “thinking model.” Such a model can read manuals, learn how to use tools, and invoke those tools to complete tasks.

From trillion‑scale training data, the Wenxin model has acquired both natural‑language and code capabilities, bridging the gap from thinking to execution. Leveraging these abilities, Baidu has developed a code agent and an intelligent code assistant. Wang said, “The code agent lets anyone do what previously only programmers could do, making everyone a programmer; the intelligent code assistant helps professional programmers write better code more efficiently—it is an AI peer for programmers.”

He described the deep‑learning framework combined with large models as the “operating system of the intelligent era.” PaddlePaddle is China’s first independently developed industrial‑grade open‑source deep‑learning platform, comprising a core framework, a model library, development kits, tool components, and the Xinghe community for learning and training. Through co‑optimization of PaddlePaddle and Wenxin, the weekly training efficiency of the Wenxin model reaches 98.8%, a 5.1‑fold improvement over a year ago, and inference speed is 105 times faster. The PaddlePaddle‑Wenxin ecosystem has attracted 12.95 million developers, served 244 000 enterprises, and created 895 000 models.

Wang concluded that artificial intelligence has become a key driver of a new wave of scientific and industrial transformation. To accelerate AI development, Baidu hopes to base innovation on Wenxin and PaddlePaddle, turning scientific breakthroughs into industrial innovation and supporting high‑level self‑reliance in technology.