Topic

Intelligent R&D at Baidu

With the rapid development of big model applications, all industries are actively embracing AI, among which software R&D, as a major field of big model, also regards it as an epoch-making productivity tool. Different from common natural language problems, in the field of software R&D and programming, big models have the need to match the precise syntax of the target language, identify the best path and edge cases, pay attention to many small details in the problem specification and other demands, and the practical applications need to solve the integration problems of business domain knowledge, R&D environment, internal services and so on. We believe that only the organic combination of engineering, process, and model can produce breakthroughs in the improvement of R&D efficiency. This sharing will introduce Baidu's technological thinking, solutions, and application practices in intelligent R&D, including model construction, SFT fine-tuning, RAG knowledge enhancement, engineering site perception, and openness of tool platforms, and other means, which can fully tap and release the productivity of big models and improve efficiency for engineers. Outline: 1. Opportunities and Challenges of Big Model Assisted Code Generation in Intelligent R&D Scenarios 2. Intelligent R&D Experience Sharing a. Analysis and quantification of big model code capability using evaluation tools b. Further enhancement of model code generation effect through SFT and other means c. Knowledge enhancement programme with RAG as the core to organically integrate the model with the business. 3. Comate development platform practice a. Embrace business customisation with openness to improve R&D efficiency. b. Integrate the community ecology, so that traditional tools can serve engineers more efficiently and accurately through intelligence. c. Turning passive into active, combining with the user programming site to push suggestions, truly becoming a development partner 4. Thoughts and Prospects Under the development of big model, the synergistic relationship between human and model will eventually change. Firstly, there will be a stronger sense of "companionship" between the model and the person. Instead of the person actively using the model to get answers, the model will "actively" observe the person's activities and give appropriate suggestions; secondly, just like the development of intelligent driving from L2 to L4, the model will gradually change from "assisting" to "assisting". Secondly, as intelligent driving develops from L2 to L4, the model will gradually develop from "assisting" to "cruising" to "taking over", which is more inclined to the completion of end-to-end requirements rather than the assistance of specific tasks.