Create 2024 Baidu AI Developer Conference Showcased in Shenzhen
On April 16, the Create 2024 Baidu AI Developer Conference, themed “Creating the Future,” took place at the Shenzhen International Convention and Exhibition Center. Wang Haifeng, Baidu’s Chief Technology Officer, announced that the total user base of Wenxin’s AI has reached 200 million, with a daily call volume of 200 million, effectively meeting users’ work, life, and study needs.
Wang mentioned that intelligent agents are a crucial direction for development and will lead to more application breakthroughs. Intelligent agents involve advanced training on basic models, including supervised fine-tuning of thinking processes, preference learning for behavioral decisions, reinforcement learning for result reflection, culminating in a thinking model. Similar to humans, intelligent agents read manuals, learn how to use tools, and can then utilize these tools to complete tasks.
Reportedly, leveraging trillions of training data, the Wenxin large model has acquired capabilities in both natural language and coding, bridging the gap from thinking to execution. Building on these dual capabilities of the Wenxin large model, Baidu has developed Code Intelligent Agents and Intelligent Code Assistants. Wang Haifeng stated, “Code Intelligent Agents enable everyone to do what only programmers could do before; everyone can become a programmer. On the other hand, Intelligent Code Assistants help professional programmers write better code more efficiently, serving as programmers’ AI companions.”
During the conference, Wang Haifeng further revealed that with continuous improvements in model effectiveness, Baidu has enhanced capabilities such as context enhancement, private domain knowledge enhancement, and seamless process integration. Currently, the overall adoption rate of the intelligent code assistant Comate has reached 46%, with generated code proportions in new code reaching 27%.
Wang Haifeng also shared insights on multi-model technology. He emphasized that in the process of applying large models, effectiveness, efficiency, and costs are all crucial factors to consider.
Apart from intelligent agents, code technologies, and multi-model techniques, the Wenxin large model continues to innovate in other aspects, including a data system based on model feedback loops, large model alignment technology based on self-enhancement feedback, and multi-modal technologies. Wang Haifeng disclosed that the effectiveness of the Wenxin large model 4.0 continues to improve, showing a 52.5% enhancement within six months of its release.
Written by: Journalist Shen Zhao
Image: Provided by the interviewee