Baidu Big Model: Users exceed 100 million, Flying Paddle developers exceed 10 million, one year review

In 2023, the last wave of breakthroughs in large AI models will come.

Wen Xinyiyan’s user base has exceeded 100 million, and the number of Feipar developers has reached 10.7 million.

In just two months, the overall effect of Wenxin Large Model 4.0 has improved by 32%.

On December 28, 2023, at the just past WAVE SUMMIT 2023 Deep Learning Developer Conference, Baidu unveiled a series of new developments in the Wenxin large model and the deep learning platform Fei Paddle.

At the conference, Baidu’s demonstration presented new methods and new ideas for AI native application development based on large models.

We can use the "Multi-Tool Intelligent Orchestration" development model of the Galaxy Community Large Model Tool Center to develop from scratch a multi-modal and full-featured "Travel Assistant" that integrates image and text recognition, question and answer, translation, broadcasting, etc. application, and can be implemented with just one line of code.

If you want to travel to Switzerland, you only need to upload a travel guide and your own travel plan to build a knowledge base about this trip, and then let it answer all your questions about the trip. kind of problem.

Of course, this travel assistant can also perform more operations. It is based on the Wenxin large model system and can be equipped with various tools such as OCR and speech synthesis, so it can help you identify German prompts in photos. This way, you can solve puzzles more easily.

#Or provide an AI explanation for photos of scenic spots.

#The evolution of large model capabilities and ecological construction have brought the development of basic models to a new stage. The era of "customized personal AI applications for all" seems to be coming.

Wen Xinyiyan, who has surpassed hundreds of millions of users and continues to evolve his capabilities, has recently continued to gain high reputation in the context of the "AI arms race" among global technology companies, reflecting his strong technical strength. At this conference, the technological and ecological progress revealed by the Wenxin model and the flying paddle are more related to the personal experience and rights of millions of developers.

Wenxin University’s multi-tool intelligent orchestration to build stronger AI applications

On October 17, 2023, Wenxin University has the strongest comprehensive strength so far Model 4.0 has been launched, and its four major capabilities of understanding, generation, logic, and memory have been significantly improved. The large language model is bringing the dawn of general artificial 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 Wen Xinyiyan’s user insights. She said that in 2023, Wenxinyiyan had completed 3.7 billion words of text creation and generated 300 million lines of code.

In the past year, the Wenxin Large Model, the basic model of "Wenxin Yiyan", has continuously released two major versions, 3.5 and 4.0, and its effects have been improving rapidly.

After the breakthrough of large model technology, the concept of agents represented by AutoGPT quickly entered people's attention. The development of multi-task agents (Agents) that can solve and adapt to complex tasks has become an important goal of researchers. AI agents are crucial for the application of large models. They can connect to a large number of apps and complete tasks autonomously, greatly improving the intelligence level of the system.

How is Wen Xinyiyan’s intelligence ability constructed? Specifically, Wen Xinyiyan now has two systems: System 1 is based on models and memory, giving users direct response generation such as perception; System 2 strengthens understanding, planning, reflection, evolution, etc. Ability.

With the support of System 2, Wen Xinyiyan is now better at using knowledge and various tools flexibly, able to help users analyze problems step by step, and has stronger ability to communicate proactively.

Based on the technical concept of agents, Baidu has developed Wen Xinyiyan’s agent model, which will be available for invitation testing for professional version users from now on.

At WAVE SUMMIT in August 2023, Baidu has released a new development paradigm based on "Wen Xin Yi Yan". So far, the number of applications based on Wen Xinyiyan has reached more than 4,000, covering various application scenarios. This time, Baidu focused on the community ecology, hoping to empower AI native application developers, and made a series of new upgrades to the Galaxy community.

The Galaxy community provides heterogeneous computing power support and more efficient common components. It has upgraded the flying paddle industry-level model library and full-process development tool chain for developers to achieve low-cost development of AI. application. The newly launched Galaxy Community Large Model Tool Center brings developers powerful AI native application building capabilities.

Wu Tian introduced that the newly released Galaxy Community Large Model Tool Center integrates the construction achievements of Baidu Artificial Intelligence over the years, including the Flying Paddle industrial-level model library, Baidu Brain AI capabilities, and Wenxin Yiyan tools etc., while supporting access to ecological tools, providing a very user-friendly visual interactive interface, with flexible and diverse parameter configurations for free selection, and real-time preview effects.

This series of upgrades allows the Galaxy community to provide developers with "all elements of AI native application innovation", including integrated services of development, experience, promotion, communication, and learning.

In terms of ecological co-creation, Baidu has previously launched the Wenxin large model "Galaxy" co-creation plan, which builds a comprehensive ecosystem with AI applications, tools, and data resources related to large models. At this conference, Baidu Recent advances in data are highlighted.

In order to enhance professional capabilities, Wen Xin Yiyan officially became a mentor. The first batch of 10 "Wen Xin Mentors" are top scholars and experts in their respective industries. Help Wen Xinyiyan enhance her knowledge in various professional fields. In addition, there is a special mentor - "Cihai". Wen Xinyiyan has deepened cooperation with Shanghai Dictionary Publishing House. The huge data of "Cihai" has been integrated into the Wenxin base model, which enhanced Wenxin Yiyan's knowledge and made it more Provide users with good services.

In the era of large models, use intelligent development tools

The technological breakthroughs of large models have increased the intensity of cutting-edge research, and at the same time Lowering the threshold for ordinary people to use AI.

As Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Research Center for Deep Learning Technology and Applications, said, artificial intelligence technology like Wen Xinyiyan is essentially a tool to improve productivity and will also become a universal Empower the platform to accelerate the intelligent transformation of the industry and create huge business value.

The development tool chain supported by large models brings three new development paradigms, lowers the threshold of AI technology, and will give birth to more and more native AI applications.

Ma Yanjun, general manager of Baidu AI technology ecology, elaborated on the scene with three cases.

The first is the comprehensively upgraded Baidu intelligent programming assistant Comate. We learned that more than 20% of Baidu’s internal code is written by Comate. In addition, about 8,000 companies are using the Comate SaaS version of the service, and the overall code generation adoption rate exceeds 40%.

However, the most eye-catching feature of Comate this time is the new function AutoWork, which can dismantle complex tasks based on the Wenxin large model, greatly shortening the entire process from putting forward requirements to completing coding, and the efficiency is natural get promoted.

Baidu demonstrated on-site in 2 minutes the development of a program to receive Comate trial rights. Developers only need to put forward requirements, and Comate's AutoWork completes the rest, formulates plans, and generates code.

In addition to the new features of Comate AutoWork, PaddleX v2.2, a low-code development tool, has also been officially released. PaddleX v2.2, a low-code development tool, is available on PaddleX Based on the capabilities of the development kit and fully combined with the Wenxin large model, it can complete industrial pain point problems that were difficult to solve before with high quality, greatly improving the development effect and efficiency; by providing a graphical interface development mode, it further reduces the threshold for using AI technology. Currently, It has supported more than 40 industry-level selected models, covering 10 major mainstream AI tasks, and adapted to domestic and foreign mainstream AI hardware, supporting cloud and local offline use.

The on-site demonstration is the extraction of key transaction information of bulk commodities. Focusing on professional terminology in the coal field, it solves the problem of inaccurate extraction of key information and achieves both improvement in development efficiency and effectiveness.

It turns out that AI developers can use the large model Prompt method to significantly improve the effect of tasks that are particularly difficult for AI developers.

Finally, the ecologically oriented Wenxinyiyan development mechanism can greatly improve the development effect, bring a great user experience, and at the same time help make development more efficient and convenient Develop more creative AI native applications.

For a task that does not seem particularly complicated at first glance, such as "making a chart of the dynamic ordering of the resident population of the top 10 provinces in the country over time", Ma Yanjun directly applied the Wenxinyiyan plug-in "Code Interpreter" on site and prompted input requirements , the code and implementation will be automatically generated.

Now developers are borrowing Wen Xinyiyan’s development mechanism to make a very easy-to-use application through similar functions very simply, and the experience looks no less than professional. Applications developed by engineers.

The upgrade of Wen Xinyiyan’s development mechanism is actually Baidu’s further lowering the threshold for AI application development. This involves four aspects of problems, from service development, registration and access, effect optimization to online distribution. Baidu has the strength and willingness to do this in the hope that more and more high-quality applications will emerge.

Based on Wenxinyiyan’s open development mechanism, no matter what type of developer you are, no matter what field the technology stack is in, you can use this development mechanism. Develop plug-ins, do multi-tool intelligent orchestration, and develop your own high-quality applications.

Ma Yanjun said, "With the change in development paradigm, I believe this is the best era for developers, and more and more high-quality AI native applications will emerge in the future." Obviously, Baidu has already Be prepared for this.

Flying Paddle Open Source Framework v2.6 realizes full-process optimization of large model suites

Wen Xinyiyan’s capabilities are in the first echelon of the industry’s large model field, in addition to In addition to being deeply involved in leading AI technology, we also owe it to Feipiao's industrial-grade deep learning open source open platform. At WAVE SUMMIT 2023, Baidu announced that the Flying Paddle open source framework has been upgraded to the latest version 2.6, which has made a series of hard-core improvements to support large model development.

Ma Yanjun said at the conference that the Flying Paddle open source framework v2.6 has achieved high scalability IR, adaptive graph construction mechanism and unified dynamic and static automatic parallelization in terms of development experience improvement. Core functions such as programming.

After improving basic capabilities, the Flying Paddle Open Source Framework v2.6 has optimized the entire process for the construction of large models. In short, on the Flying Paddle open source framework v2.6, all processes of large model suites from pre-training, fine-tuning, compression, inference to deployment have been comprehensively optimized.

For large model technology, it is crucial to make full use of hardware computing power. Fei Paddle has upgraded its hardware adaptation solution this time, which can better support products from different hardware manufacturers, enable flexible customization, and in-depth optimization of software and hardware collaboration.

Combined with the adaptation and optimization of Wenxin's large model, Feipiao and hardware manufacturers are jointly building the "Hardware Transformer Large Operator Acceleration Library" to accelerate the improvement of the industry's software stack system.

Conclusion

In the era of large models, technology develops very fast. In every release of WAVE SUMMIT, we Everyone can feel this fast-paced progress: Since March 2019, Wenxin Large Model has experienced a rapid development process from 1.0 to 4.0, and the biannual conference has also been held for the tenth time today.

We have also witnessed the continuous prosperity of Feipiao deep learning open source platform. So far, the Flying Paddle platform has gathered 10.7 million developers, served 235,000 companies, and built 860,000 models on it. On this increasingly powerful platform, Baidu collaborates with all parties to promote the prosperity and development of AI technology and ecosystem and accelerate the implementation of applications.

Baidu’s forward-looking judgment on artificial intelligence technology and industry trends is constantly guiding the direction of technological innovation and industrial practice.

Wang Haifeng said in the opening speech of this conference: "The deep learning platform plus large models connects the entire AI industry chain from hardware adaptation, model training, inference deployment, to scenario application, and consolidates the Industrial intelligence base. The emergence of large language models this year has brought the dawn of general artificial intelligence."

From calling the deep learning framework as the "operating system of the intelligent era", From the industrial intelligence accelerated by cloud-intelligence integration to the entire AI industry chain from hardware to applications, Baidu has now leveraged its technological advantages to build a development system that covers the entire industry and has low thresholds, and has fully utilized its capabilities in the era of large models. Take advantage.

In the process of continuous technological innovation and industry-enabling development, Feipiao itself is also constantly upgrading, from a deep learning framework to a platform ecosystem, developing into a technologically advanced and feature-rich industrial-level deep learning platform.

Perhaps it won’t be long before the changes brought about by this round of AI breakthroughs will touch more people, and we will witness the subversion of productivity and innovation by generative AI.

In this revolution, I believe we will see more and more AI native applications, from Baidu and Wen Xinyiyan.

The above is the detailed content of Baidu Big Model: Users exceed 100 million, Flying Paddle developers exceed 10 million, one year review. For more information, please follow other related articles on the PHP Chinese website!