Author | vanilla
Edit | Li Shuiqing
Zhidong reported on June 28 that today, Baidu officially released the Wenxin large model 4.0 Turbo version at the WAVE SUMMIT Deep Learning Developer Conference 2024, which will be launched simultaneously on the web and API interfaces, and the App will also be launched one after another.
▲文心大模型4.0 Turbo API开放
At the same time, the PaddlePaddle platform released a new generation of PaddlePaddle Framework 3.0 beta version, which has the characteristics of integrated training and pushing of large models, unified dynamic and static automatic parallelism, automatic compiler optimization, and multi-hardware adaptation of large models.
Baidu Comate, Baidu's intelligent code assistant, has been upgraded to version 2.5, and the speed of development, business iteration and enterprise landing has been greatly improved, and it has a new Chinese name "Wenxin Quick Code", officially joining Baidu's "Wenxin family".
According to Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Research Center for Deep Learning Technology and Application, Wenxin Yiyan currently has more than 300 million users and more than 500 million daily calls. The PaddlePaddle platform has served 14.65 million developers and 370,000 enterprises, and built more than 950,000 models.
1. The number of users exceeded 300 million, and Wenxin Model 4.0 Turbo was launched
After decades of development, AI technology has evolved from the early manual writing of rules, to the later statistical machine learning, to the era of deep learning, and then to the era of large models, and algorithms and models have become more general and unified.
Wang Haifeng said that large models bring dawn to AGI, which can be interpreted from two perspectives, one is the versatility of AI technology, and the other is the comprehensiveness of AI capabilities. Among them, comprehensive AI capabilities mainly include four parts: understanding, logic, generation, and memory.
Baidu launched Wenxin Model 1.0 in March 2019, and after several iterations, its version 4.0 was released in October last year. At present, the cumulative user scale of Wenxin Yiyan has reached 300 million, and the number of daily calls is 500 million.
Today, Baidu launched the Wenxin model 4.0 Turbo version, its web terminal and API will be open to users from now on, and the App side will also be launched one after another, and enterprise users can log in to Baidu Intelligent Cloud Qianfan large model platform to call API services.
▲Wenxin large model 4.0 Turbo was released
Wenxin 4.0 Turbo has greatly improved the generation speed and effect. Wang Haifeng demonstrated the generation effect of the 4.0 version compared to the 4.0 version, under the same prompt words, the generation speed of the 4.0 Turbo is significantly faster, and the generated content is more organized.
▲Comparison of the generation speed of Wenxin 4.0 Turbo and Wenxin 4.0
In terms of upgrading the basic model, Baidu has further developed the agent mechanism, including understanding, planning, reflection and evolution, so that it can perform reliably and evolve itself, and to a certain extent, the thinking process is white-boxed, so that the large model can think and act like a human, call tools to complete complex tasks independently, and continue to learn in the environment to achieve independent evolution.
▲Agent upgrade based on Wenxin large model
Baidu also co-created the first agricultural intelligent body "Farmer Academician Intelligent Body" with Academician Zhu Youyong and his team of the Chinese Academy of Engineering, which is based on the Wenxin Intelligent Twin platform, absorbing the research results of Academician Zhu Youyong and related agricultural knowledge, which can answer questions in production and life for farmers.
▲Farmer academician agent
Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, said that in the past six months, the average number of questions asked by users using the Wenxin model has increased by 78%, and the average length of questions has increased by 89%, and the prompts used have become more complex and specific.
A large number of applications are based on Wenxin's large model for innovation, and so far, Wenxin Yiyan has created 590 million articles, drawn 240 million pictures, generated 7 billion lines of code, and assisted 130 million people in their work.
▲The cumulative results of the Wenxin model
Since its launch in August last year, 1,000+ large model tools have been created, more than 550,000 AI applications have been created, and more than 1,000B high-quality data has been accumulated.
For example, Baidu cooperated with Changguang Satellite Technology Co., Ltd. to develop a symbiotic earth intelligent assistant, which served 31,000 users in the first month of its launch and answered more than 38,000 remote sensing technical questions.
Second, PaddlePaddle launched the 3.0 beta version, with a four-layer framework and four advantages
PaddlePaddle is an open-source and open industrial-level deep learning platform launched by Baidu, according to Wang Haifeng, it has served 14.65 million developers, 370,000 enterprises, and created 950,000 models.
Today, the PaddlePaddle Framework 3.0 beta version was officially released, which has four major features: dynamic and static unified automatic parallelism, automatic compiler optimization, multi-hardware adaptation of large models, and integration of large model training and pushing.
▲Baidu Paddle Frame 3.0 Beta
In terms of design concept, considering the development trend of large models and heterogeneous multi-core, in order to ensure the training and inference performance of large models, simplify the development and tuning process of large models, and better adapt to various chips, PaddlePaddle and Wenxin have jointly optimized in two aspects: training and inference.
▲ Paddle frame design concept
Specifically, the automatic parallelism capability of dynamic and static unity realizes hybrid parallel training of large models through simplified development, which can reduce the amount of distributed training core code by more than 50%, and the training performance of typical models is 20% ahead.
The automatic optimization of the compiler can efficiently support large models and AI scientific computing, improve the training efficiency of general models through integrated design, improve the general performance of large model inference by 30%, and lead the solution speed of typical scientific computing models by 71%.
▲Compiler auto-optimization
In terms of multi-hardware adaptation of large models, developers only need to adapt more than 30 interfaces to fully support large model training, compression, and inference through low-cost hardware access to support software and hardware collaborative optimization.
The integration of large model training and pushing is the core feature of PaddlePaddle Framework 3.0, which accelerates the RLHF training and reuse inference engine by 2.1 times and improves the efficiency of the quantitative compression and reuse distributed strategy by 3.8 times through the seamless connection of training, compression, and inference.
▲ large model training and push integration
From the perspective of the overall framework, PaddlePaddle 3.0 is divided into four layers: the presentation layer, the scheduling layer, the operator layer, and the adaptation layer. It is worth noting that the development interface of PaddlePaddle 3.0 is fully compatible with version 2.0, and users do not need to overturn the code and reinvent the code.
▲Paddle 3.0 frame
PaddlePaddle is also optimized for AI for Science, which can adapt to more than 130 models of mainstream scientific computing libraries, covering the global solution mode of data, mechanism, and mathematical integration, and can widely support scientific problem analysis, among which based on high-order automatic differentiation and compilation optimization technology, the equation solving model is 71% faster than PyTorch on average.
3. The intelligent code assistant has been upgraded to version 2.5, covering 80% of Baidu's internal programmers
Chen Yang, vice president of Baidu, released version 2.5 of Comate, an intelligent code assistant, which integrates more R&D knowledge and covers a more comprehensive process.
▲Baidu Smart Code Assistant upgraded to version 2.5
At the same time, the assistant also has a Chinese name - Wenxin Quick Code, and officially joined the Baidu Wenxin family.
Chen Yang said that the "fast" of Wenxin Express Code is mainly reflected in three aspects: fast development speed, fast business iteration, and fast enterprise landing.
▲Wenxin fast code 2.5
Specifically, in terms of development speed, Wenxin Quick Code integrates more R&D knowledge, including the practical experience of hundreds of technical experts, billion-level R&D field knowledge, and R&D scenarios of tens of thousands of enterprises.
In Baidu's internal use, the accuracy of the development framework code generated based on Wenxin Quick Code has reached 80%, and the number of codes submitted per unit time has increased by 35%.
▲The development speed of Wenxin fast code has been improved
In terms of business iteration, Wenxin Quick Code seamlessly integrates all aspects of R&D, including demand research, design, code writing, testing, release, etc.
Chen Yang said that in terms of single-point efficiency improvement in Baidu's R&D process, the writing time of a single line of code based on Wenxin Quick Code is compressed to 1.74 seconds, the automatic generation of deployment scripts reaches the second level, and the time spent on a single offline environment deployment is saved by more than 10 minutes. Overall, the overall R&D efficiency improvement for individual engineers is more than 14%.
▲The iteration speed of Wenxin fast code business has been improved
At the enterprise landing level, in response to the pain points such as differences in industry norms, enterprise private domain knowledge, and the combination of code assistants and internal tools of enterprises, Wenxin Quick Code provides core capabilities such as security and trustworthiness and business adaptation.
After one year of practice, among Baidu's more than 10,000 internal engineers, the coverage rate of Wenxin Quick Code has reached 80%, the adoption rate of its generated code has reached 46%, and the new code generation accounts for 29%.
Chen Yang also shared the landing case of benchmark customer Himalaya, which achieved 90% coverage among engineers, 44% code adoption rate, and 33% of new code generation in just one quarter.
▲Wenxin quick code enterprise landing case
Jiang Jie, CTO of Himalaya, said that the integration of Wenxin Express Code and Himalaya's accumulated R&D capabilities and knowledge has greatly shortened the time for technical research and code writing, and greatly improved the overall R&D efficiency and product quality of the enterprise.
At present, Wenxin Express provides 4 editions: Standard Edition, Professional Edition, Enterprise Edition, and Enterprise Exclusive Edition, of which the Professional Edition is open to professional developers for a limited time, and the basic and advanced functions and models are not limited.
Conclusion: AI has entered the stage of industrial mass production
From a historical point of view, throughout the first three industrial revolutions, their core driving forces, mechanical technology, electrical technology and information technology, have a strong universality. Wang Haifeng said that when they show the characteristics of standardized, automated and modular industrial production, the core technology will enter the stage of industrial production.
At present, AI technology, including algorithms, data, models, tools, etc., has also become highly versatile, and has the characteristics of standardization, modularization, and automation, promoting AI to enter the stage of industrial mass production.