On September 11-12, the CCS 2024 Chengdu Cybersecurity Series event was held in Chengdu. Baidu Vice President Chen Yang attended the opening ceremony of the conference and gave a keynote speech. Chen Yang stated that under the empowerment of large models, the research and development process is undergoing a profound transformation. The understanding, generation, logic, and memory capabilities of large models are combined with the field of software development, making software development more efficient and secure.
CCS 2024 is jointly sponsored by Chengdu Internet Information Office and Chengdu High tech Industrial Development Zone Management Committee, and undertaken by Chengdu Sugar Free Information Technology Co., Ltd. Under the theme of "Cybersecurity for the People, Cybersecurity Depends on the People" during Cybersecurity Week, CCS 2024 focuses on the new quality productivity in the AI era, with the theme of "Wanxiang AI, New Security", gathering experts and enterprise representatives from all walks of life across the country to discuss the new paradigm of cybersecurity in the AI era and explore the path of integrated development of artificial intelligence and digital security.
Generative artificial intelligence is bringing unprecedented opportunities to the field of software development. At the meeting, Chen Yang shared his judgment on the trend of intelligent development in software research and development, from real-time continuation of L1 code at the beginner level, to full process assistance in L2 research and development, to L3 intelligent agent combination, and finally entering the era of no code in natural language programming. The industry is currently in the stage of evolution from L2 to L3, and the role of intelligent agents in the research and development process is becoming increasingly important. In the L3 stage, end-to-end generation from requirements to products can be achieved through private domain knowledge enhancement, engineering architecture understanding, code generation testing, and requirement direct code generation, which has a direct impact on the R&D process. In the L3 stage, the proportion of AI code generation can reach 80%, further unleashing engineer productivity.
Chen Yang introduced that Baidu has been exploring the field of intelligent coding for two years. The intelligent coding efficiency improvement tool Wenxin Kuai Code has been used by tens of thousands of engineers within the company, and the intelligent research and development efficiency has been significantly improved. So far, 30% of the code generated by Baidu every day is generated by Wenxin Kuai Code, with an overall adoption rate of 46%. The application of this efficiency improvement tool has enabled engineers to achieve an overall efficiency improvement of 12%, significantly increasing the number of code submissions and business iteration speed per unit of time. More importantly, intelligent coding efficiency tools allow engineers to free themselves from complex and repetitive work and focus on more valuable and creative tasks.
While pursuing efficiency, Baidu also attaches great importance to engineering quality. Wenxin Kuai Code has integrated quality and safety considerations throughout the entire R&D process: in the R&D requirement stage, safety design is introduced through requirement clarification to ensure the implementation of safety standards; In the research stage, through module sorting and code explanation, engineers can quickly understand the architecture logic of the project and code, identify existing security logic vulnerabilities and fix them with one click. In addition, the automatic generation of flowchart patterns further improves the efficiency of this stage; During the development phase, the security enhancement mode is used to make AI generated code more secure. Sensitive information is desensitized to avoid data leakage during data transmission. In addition, vulnerabilities can be automatically discovered and fixed for engineer code; During the testing phase, the automated generation of unit tests through testing agents ensures the quality and safety of research and development; During the release phase, online business security is ensured through process security control.
The application of Wenxin Quick Code not only improves research and development efficiency, but also significantly enhances code quality. In Baidu's internal practice, the accuracy rate of scanning security vulnerabilities for Wenxin Kuai Code has exceeded 95%, and 83% of the scanned vulnerabilities have been fixed in a closed loop. In addition, not only in the research and development stage, but also in the secure operation stage, Wenxin Kuai Code can significantly improve efficiency.
Chen Yang also pointed out that although the industrial application of big models in the field of intelligent research and development has shown significant value and practical effects, the implementation still faces various difficulties and challenges, including how to make good use of private domain knowledge, how to solve the differences in industry quality and standards, how to coordinate engineering culture and operation promotion, how to achieve value measurement, and so on. At present, Baidu has summarized a standardized implementation process and best practices through the internal and external experience of Wenxin Kuai Code. By implementing the principle of human-machine collaboration and building a landing execution system, it promotes the transformation of enterprise engineering culture and data-driven value loop. (Li Yan)