DoNews reported on April 16 that on April 16, the Create 2024 Baidu AI Developer Conference with the theme of "Creating the Future" was successfully held at the Shenzhen World Convention and Exhibition Center. Wang Haifeng, Chief Technology Officer of Baidu, delivered a speech on the topic of "Technology Builds the Foundation, Galaxy Shines", and interpreted the key technologies and latest progress of many Wenxin large models such as agents, code, and multiple models.
Since the release of the knowledge-enhanced large language model Wenxin Yiyan on March 16 last year, Baidu has continuously iteratively upgraded the Wenxin large model, and has continued to innovate in technological innovation, further developing knowledge point enhancement from knowledge enhancement and retrieval enhancement, and based on greater computing power, more data and stronger algorithms, relying on the paddle platform, from Wenxin 3.0, 3.5, to 4.0, the Wenxin large model has become more and more powerful, and the effect and performance have been comprehensively improved.
Wang Haifeng said that intelligent twins are an important development direction, which will bring more application explosions. The agent is further trained on the basic model for thinking enhancement, including supervised fine-tuning of the thinking process, preference learning for behavioral decision-making, and reinforcement learning for outcome reflection, so as to obtain the thinking model. Like a human, the agent's thinking model reads instructions, learns how to use tools, and can invoke tools to complete tasks.
Wang Haifeng interpreted the thinking process of the agent and demonstrated how the agent can call the tool through thinking. On the Wenxin Model 4.0 tool version, ask the question, "I'm going on a business trip to the Greater Bay Area for a week." I want to know how the weather changes so I can decide what to bring. Please help me check the temperature in Beijing and Shenzhen for the coming week, tell me what clothes I should bring on a business trip, and organize it into a table. "The agent is like a human being, through thinking and planning, dismantling the user's needs into multiple subtasks, first calling the "advanced networking" tool to query weather information, then calling the "code interpreter" to draw a temperature trend graph, according to the weather conditions of the coming week, selecting the appropriate clothing, and finally thinking and confirming the results, and automatically summarizing them into a table.
From trillions of training data, the Wenxin model has learned both natural language and code capabilities, opening up the process from thinking to execution. Based on these two capabilities of the Wenxin model, Baidu has developed code agents and intelligent code assistants.
Wang Haifeng said: "Code agents allow everyone to do things that only programmers can do before, and everyone can become programmers, while intelligent code assistants help professional programmers write better code more efficiently, which can be said to be programmers' AI peers." ”
The code agent is "writing code with a model to make complex tasks simple", and the thinking model plus the code interpreter constitutes the code agent. First, the thinking model understands the user's needs, and after thinking, integrates the instructions and related information to complete the task into a prompt and inputs them to the code interpreter; then, the code interpreter translates the user requirements expressed in natural language into code and executes them according to the prompts, and obtains the execution result or debugging information; then, the thinking model reflects on and confirms the execution result of the code interpreter, and returns the result to the user if it is correct, and continues to update it independently if it is incorrect.
Wang Haifeng demonstrated the use of code agent to customize the invitation letter for the guests of this Create conference, the code agent first understands the content of the invitation template, then generates the code and executes, fills in the guest's name in the appropriate position, the newly generated invitation file is named after the guest, and finally outputs a number of packaged invitations.
At the conference, Wang Haifeng further revealed that on the basis of the continuous improvement of the model effect, Baidu has further built capabilities such as context enhancement, private domain knowledge enhancement, and seamless process integration. At present, the overall adoption rate of Comate has reached 46%, and the proportion of new code generation has reached 27%.
Comate seamlessly integrates code understanding, generation, optimization, and other capabilities into all aspects of the R&D process, acting like an assistant to help improve the quality and efficiency of code development. Comate demonstrates the process of helping engineers take over the code, through a simple instruction, you can quickly understand the architecture of the entire code, even the specific implementation logic of each module, and can automatically generate new code to meet the requirements based on the current project code and third-party code.
Wang Haifeng also shared multi-model technology on the spot. He said that in the process of large-scale model application, effect, efficiency and cost are very important. In practical application, it is necessary to select the most suitable model based on the requirements of the scenario.
On the one hand, it is efficient and low-cost model production, and on the other hand, it is multi-model inference. In terms of efficient and low-cost model production, Baidu has developed a collaborative training mechanism for large and small models, which can effectively carry out knowledge inheritance, efficiently produce high-quality small models, and can also use small models to achieve contrast enhancement to help train large models. At the same time, the seed model matrix, data quality improvement and enhancement mechanism, and supporting tool chain from pre-training, fine-tuning alignment, model compression to inference deployment are built.
The efficient and low-cost model production mechanism helps the application speed to be faster, lower cost, and better results. In terms of multi-model inference, Baidu has developed an end-to-end multi-model inference technology based on feedback learning, built an intelligent routing model, and carried out end-to-end feedback learning, giving full play to the ability of different models to handle different tasks to achieve the best balance of effect, efficiency and cost.
In addition to agents, code, and multi-model technologies, Wenxin models also continue to innovate in other aspects, including a closed-loop data system based on model feedback, large model alignment technology based on self-feedback enhancement, and multimodal technology. Wang Haifeng announced on the spot that the effect of Wenxin Model 4.0 continued to improve, and it increased by 52.5% in half a year after its release.
The continuous and rapid evolution of Wenxin's large model is due to Baidu's full-stack layout of chips, frameworks, models and applications, especially the joint optimization of PaddlePaddle Deep Learning Platform and Wenxin. The average weekly training efficiency of the Wenxin model reached 98.8%, which was 5.1 times the training efficiency and 105 times the inference efficiency when Wenxin Yiyan was released a year ago. Up to now, the PaddlePaddle Wenxin ecosystem has gathered 12.95 million developers, served 244,000 enterprises and institutions, and created 895,000 models based on PaddlePaddle and Wenxin.
Wang Haifeng said that the cumulative user scale of Wenxin Yiyan has reached 200 million, and the average daily call volume has also reached 200 million, which effectively meets the needs of users for work, life and study.
Finally, Wang Haifeng introduced the latest progress of Baidu's AI talent plan, which proposed in 2020 to cultivate 5 million AI talents for the whole society in five years, and this goal has been achieved ahead of schedule. He said, "In the future, we will continue to devote ourselves to talent training, so that the little stars of talents will converge into a bright galaxy." In the era of intelligence, everyone is a developer, everyone is a creator, let us work together to create a better future. ”