Baidu's Large Model in 2023: Wenxin Yiyan Becomes the First Domestic Model to Reach 100 Million Users
-
Wenxin Yiyan's user base has exceeded 100 million, and PaddlePaddle's developer community has grown to 10.7 million.
In just two months, the overall performance of Wenxin Large Model 4.0 improved by 32%.
On December 28, 2023, at the recently concluded WAVE SUMMIT+ 2023 Deep Learning Developer Conference, Baidu unveiled a series of new advancements in the Wenxin Large Model and the deep learning platform PaddlePaddle.
At the conference, Baidu's demonstrations showcased new methods and approaches for developing AI-native applications based on large models.
To develop an AI-native application, not a single line of code is needed: Using the "Multi-Tool Intelligent Orchestration" development mode in the Galaxy Community's Large Model Tool Center, we can create a fully functional "Travel Assistant" application from scratch, integrating multimodal capabilities such as image-text recognition, Q&A, translation, and broadcasting.
If you're planning a trip to Switzerland, simply upload a travel guide and your itinerary to build a knowledge base about your journey. Then you can ask it any questions related to your travel plans.
Of course, this travel assistant can do much more. Based on the Wenxin large model system and equipped with tools like OCR and speech synthesis, it can help you understand the content of German signs in photos:
Or provide AI-powered commentary on scenic spot photos.
The evolution of large model capabilities and ecosystem construction has propelled foundational models into a new stage, making the era of 'personalized AI applications for everyone' appear imminent.
Breaking through the milestone of hundreds of millions of users, the continuously advancing capabilities of Wenxin Yiyan (ERNIE Bot) have recently garnered increasing acclaim amid the global tech companies' 'AI arms race,' showcasing formidable technical prowess. At this conference, the technological and ecosystem progress disclosed by the Wenxin large model and PaddlePaddle holds significant implications for the practical usage experience and rights of millions of developers.
Wenxin Large Model + Multi-tool Intelligent Orchestration for Building More Powerful AI Applications
On October 17, 2023, the most comprehensive and powerful Wenxin large model 4.0 was unveiled, with significant enhancements in its four core capabilities: comprehension, generation, logic, and memory. Large language models are now heralding the dawn of artificial general intelligence.
At WAVE SUMMIT+ 2023, Wu Tian, Vice President of Baidu Group and Deputy Director of the National Engineering Research Center for Deep Learning Technology and Applications, shared insights into the usage of Wenxin Yiyan. She stated that in 2023, Wenxin Yiyan has completed 3.7 billion words of text creation and generated 300 million lines of code.
Over the past year, the foundational model of Wenxin Yiyan, the Wenxin Large Model, has successively released two major versions, 3.5 and 4.0, with continuously rapid improvements in performance.
Following breakthroughs in large model technology, the concept of intelligent agents, represented by AutoGPT, quickly entered the public's view. Developing multi-task agents (Agent) capable of solving and adapting to complex tasks has become an important goal for researchers. AI agents are crucial for the application of large models, as they can connect numerous apps, autonomously complete tasks, and significantly enhance the intelligence level of systems.
How is the agent capability of Wenxin Yiyan constructed? Specifically, Wenxin Yiyan now features two systems: System One is based on models and memory, providing users with direct response generation like perception; System Two enhances a series of capabilities including understanding, planning, reflection, and evolution.
With the integration of System 2, Wenxin Yiyan now excels in flexibly utilizing knowledge and various tools, enabling it to analyze problems progressively and engage in more proactive communication.
Based on the concept of agent technology, Baidu has developed the agent mode for Wenxin Yiyan, which is now available for invitation testing for professional users.
At the WAVE SUMMIT in August 2023, Baidu introduced a new development paradigm based on Wenxin Yiyan. To date, over 4,000 applications have been built on Wenxin Yiyan, covering a wide range of scenarios. This time, Baidu aims to empower AI-native application developers by upgrading the Xinghe Community ecosystem.
The Galaxy Community provides heterogeneous computing support and more efficient general components, upgrading PaddlePaddle's industrial-grade model library and full-process development toolchain for developers to enable low-cost AI application development. The newly launched Galaxy Community Large Model Tool Center offers powerful AI-native application building capabilities.
Wu Tian introduced that the newly released Galaxy Community Large Model Tool Center integrates Baidu's years of achievements in artificial intelligence, including the PaddlePaddle industrial-grade model library, Baidu Brain AI capabilities, and the ERNIE Bot tool, while also supporting the integration of ecosystem tools. It provides a user-friendly visual interactive interface with flexible and diverse parameter configurations and real-time preview effects.
This series of upgrades enables the Galaxy Community to provide developers with "all elements for AI-native application innovation," including integrated services for development, experience, promotion, communication, and learning.
In terms of ecosystem co-creation, Baidu's previously launched ERNIE Large Model "Galaxy" Co-Creation Program has built a comprehensive ecosystem around large model-related AI applications, tools, and data resources. At this conference, Baidu highlighted the latest progress in data-related initiatives.
To enhance its professional capabilities, Wenxin Yiyan has officially 'taken mentors.' The first batch of 10 'Wenxin Mentors' comprises top scholars and experts in their respective fields, who will help Wenxin Yiyan strengthen its understanding across various professional domains. Additionally, there is a special mentor—'Cihai.' Through deepened collaboration with Shanghai Lexicographical Publishing House, the vast data from 'Cihai' has been integrated into Wenxin's foundational large model, enriching Wenxin Yiyan's knowledge and improving its service to users.
In the Era of Large Models, Using Intelligent Development Tools
The technological breakthroughs in large models have intensified cutting-edge research while simultaneously lowering the barrier for the general public to use AI.
As Wang Haifeng, Baidu's Chief Technology Officer and Director of the National Engineering Research Center for Deep Learning Technology and Applications, stated, AI technologies like Wenxin Yiyan are fundamentally tools to enhance productivity. They will also serve as a universal empowerment platform, accelerating the transformation of industrial intelligence and creating substantial commercial value.
The development toolchain powered by large models introduces three new development paradigms, lowering the threshold for AI technology and fostering the emergence of increasingly native AI applications.
Ma Yanjun, General Manager of Baidu's AI Technology Ecosystem, elaborated on these points with three case studies during the event.
First is the comprehensively upgraded Baidu Intelligent Programming Assistant, Comate. It has been revealed that over 20% of Baidu's internal code is now written by Comate. Additionally, around 8,000 enterprises are using Comate's SaaS version, with an overall code generation adoption rate exceeding 40%.
However, the most impressive aspect of this Comate update is the new AutoWork feature. It can decompose complex tasks based on the Wenxin large model, greatly reducing the entire process for developers from proposing requirements to completing code, naturally improving efficiency.
Baidu demonstrated the development of a program to claim Comate trial benefits in just 2 minutes live. Developers only need to propose the requirements, and the rest is handled by Comate's AutoWork, which formulates plans and generates code.
In addition to Comate AutoWork's new features, PaddlePaddle's low-code development tool PaddleX v2.2 has been officially released. Building upon the capabilities of the PaddlePaddle development kit and fully integrating the ERNIE large model, PaddleX v2.2 can effectively address previously challenging industrial pain points, significantly enhancing development outcomes and efficiency. By offering a graphical interface development mode, it further lowers the barrier to using AI technology. Currently, it supports over 40 industry-grade selected models, covering 10 major mainstream AI tasks, and is compatible with both domestic and international mainstream AI hardware, supporting both cloud and local offline usage.
The live demonstration showcased the extraction of key trading information for bulk commodities. By addressing the inaccuracies in extracting key information related to professional terminology in the coal industry, it achieved improvements in both development efficiency and effectiveness.
Previously, tasks that were particularly challenging for AI developers can now achieve significant improvements in performance using large model Prompt methods.
Finally, the ecosystem-oriented development mechanism of Wenxin Yiyan can greatly enhance development outcomes, delivering an excellent user experience while facilitating the more efficient and convenient creation of innovative AI-native applications.
For a seemingly straightforward task like "creating a dynamic ranking chart of the top 10 provinces by resident population over time," Ma Yanjun demonstrated the use of Wenxin Yiyan's "Code Interpreter" plugin on-site. By simply inputting the requirement as a prompt, the system automatically generated and executed the code.
Now, developers can leverage Wenxin Yiyan's development mechanism to easily create highly functional applications through similar features, with an experience that appears no less impressive than those developed by professional engineers.
The upgrade to Wenxin Yiyan's development mechanism represents Baidu's effort to further lower the barrier to AI application development. This addresses challenges across four key stages: service development, registration and integration, performance optimization, and deployment. Baidu has both the capability and the intention to pursue this, aiming to foster the emergence of increasingly high-quality applications.
Based on the open development mechanism of ERNIE Bot, developers of any type and technical stack can utilize this system to create plugins, perform intelligent multi-tool orchestration, and develop high-quality applications.
Ma Yanjun stated, "With the shift in development paradigms, I believe this is the best era for developers. We will see an increasing number of high-quality AI-native applications emerge in the future." Clearly, Baidu is well-prepared for this trend.
PaddlePaddle Open-Source Framework v2.6 Achieves Full-Pipeline Optimization for Large Model Suites
ERNIE Bot's capabilities rank among the top tier in the industry's large model domain. This achievement stems not only from Baidu's continuous advancement in cutting-edge AI technologies but also from the PaddlePaddle industrial-grade deep learning open-source platform. At WAVE SUMMIT+ 2023, Baidu announced the upgrade of PaddlePaddle to version 2.6, introducing a series of significant enhancements for large model development support.
Ma Yanjun stated at the conference that PaddlePaddle Open Source Framework v2.6 has implemented core features to enhance the development experience, including highly extensible IR, adaptive graph construction mechanism, and unified static-dynamic automatic parallel programming.
After improving foundational capabilities, PaddlePaddle Open Source Framework v2.6 has undergone comprehensive optimization for large model development. In short, the framework now offers end-to-end optimization for large model suites, covering all processes from pre-training, fine-tuning, compression, inference, to deployment.
For large model technologies, fully utilizing hardware computing power is crucial. PaddlePaddle has upgraded its hardware adaptation solutions to better support products from various hardware manufacturers, allowing for flexible customization and deep optimization of software-hardware collaboration.
In conjunction with the adaptation and optimization of the ERNIE large model, PaddlePaddle and hardware manufacturers are jointly building a "Hardware Transformer Large Operator Acceleration Library" to accelerate the improvement of the industry's software stack system.
Conclusion
In the era of large models, technological advancements are rapid. With each release at WAVE SUMMIT, we can feel this fast-paced progress: since March 2019, the ERNIE large model has evolved from version 1.0 to 4.0, and the biannual conference has now reached its tenth edition.
We have witnessed the continuous prosperity of the PaddlePaddle deep learning open platform. As of now, PaddlePaddle has gathered 10.7 million developers, served 235,000 enterprises, and built 860,000 models on its platform. On this increasingly powerful platform, Baidu collaborates with various parties to promote the prosperity of AI technology and ecosystem, accelerating application implementation.
Baidu's forward-looking judgment on AI technology and industry trends continues to guide the direction of technological innovation and industrial practice.
Wang Haifeng stated in his opening speech at this conference: "The deep learning platform combined with large models has connected the entire AI industrial chain from hardware adaptation, model training, inference deployment to scenario applications, solidifying the foundation for industrial intelligence. The emergence of large language models this year has brought dawn to general artificial intelligence."
From proclaiming deep learning frameworks as the "operating system of the intelligent era" to cloud-intelligence integration accelerating industrial intelligence, and further to establishing a complete AI industry chain spanning from hardware to applications, Baidu has now utilized its technological advantages to construct a development system that covers all industries with low barriers to entry, fully demonstrating its strengths in the era of large models.
In the continuous process of technological innovation and industrial empowerment, PaddlePaddle itself has been constantly upgrading, evolving from a deep learning framework to a platform ecosystem, and developing into an industry-leading deep learning platform with advanced technology and rich functionalities.
Perhaps it won't be long before the transformations brought by this wave of AI breakthroughs reach more people, and we will witness the disruptive impact of generative AI on productivity and innovation.
In this transformation, we believe we will see more and more AI-native applications emerging, from Baidu, from ERNIE Bot.