This is productivity in the era of large models!

Wenxin and Flying Paddle showed us the productivity of leading large models.

What is the extent of the large model application volume? A few days ago, what we saw was writing articles, drawing pictures, and answering math questions. Now some people have used it like this:

If a long conversation is forwarded to another group chat, AI can automatically generate a summary. After getting the data, directly analyze it with evidence and pictures.
From market analysis, brand building, to outputting video advertisements, the entire process can be completed with a simple conversation with AI. This is a series of technologies, products and ecological achievements displayed by Baidu at the just-concluded WAVE SUMMIT conference, including the Wenxin model, the flying paddle platform, and AI native applications such as Liuliu.

Recently, large language models have made shocking technological breakthroughs. Artificial intelligence represented by big language models is penetrating thousands of industries, accelerating industrial upgrading and economic growth. Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Research Center for Deep Learning Technology and Application, said that the big language model has the core basic capabilities of artificial intelligence such as understanding, generation, logic, and memory, bringing the dawn of general artificial intelligence.

Wang Haifeng further stated that artificial intelligence has many typical abilities, among which understanding, generation, logic, and memory are the core basic abilities. The stronger these four abilities are, the closer they are to general artificial intelligence.

In the face of this important change in artificial intelligence, the two core technologies of Baidu, Flying Paddle and Wenxin Large Model, have come up with a series of leading releases.

Wenxin Large Model: Far Ahead

In the domestic AI field, Baidu has always been at the forefront of technology, dating back to the release of ERNIE 1.0 in March 2019. In March of this year, Baidu took the lead in unveiling the self-developed knowledge-enhanced large-scale language model "Wen Xin Yi Yan". Tips and other technologies have technical advantages such as knowledge enhancement, retrieval enhancement and dialogue enhancement.

The latest version of Wenxin Large Model is version 3.5 released not long ago. Wu Tian, ​​vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, said that Wenxin Yiyan has mastered more than 200 creative genres, covering almost all writing needs, and the richness of content is 1.6 times that of the initial period. The chain length is 2.1 times that of the initial stage, and the coverage of knowledge points is 8.3 times that of the initial stage.

In the new version, Wenxin Big Model further innovates in core technologies such as basic model, knowledge enhancement, and retrieval enhancement, and realizes improvements in basic model, fine-tuning technology, knowledge point enhancement, logical reasoning, and plug-in mechanism.

Among them, on the basis of knowledge and retrieval enhancement, Wenxin large model 3.5 proposes "knowledge point enhancement technology", which enables the model to better use fine knowledge points to improve understanding and generation capabilities, and enhance the grasp and application of world knowledge.

In terms of reasoning, through large-scale logical data construction, logical knowledge modeling, multi-granularity semantic knowledge combination and symbolic neural network technology, the Wenxin large model improves the performance of logical reasoning, mathematical calculation and code generation tasks.

What can bring us a more obvious perception is the plug-in. We know that in practical applications, large models sometimes face the problems of limited data and insufficient capabilities. In version 3.5, a plug-in mechanism has been added to the Wenxin large model. The official plug-ins that have been launched include Baidu Search, Scroll Documents, Yijing Liuying, Shuo Tu Jie Hua, and E Yan Yi Tu. in:

• Baidu search is the default plug-in, which enables Wenxinyiyan to obtain real-time and accurate information.

• Browsing documents can fully understand and locate the format, layout and other information of the document with the help of the document intelligent model and search system, breaking through the limitation of the large model on the understanding of document length. Now we can use Wenxin to "dialogue" with documents to solve the needs of summarizing, asking and answering and creating documents.

• Relying on Wenxin 's cross-modal large model, Yijing Liuying has broken through technical problems such as semantic alignment between different modalities, and innovatively integrated a series of technical capabilities such as text, vision, voice, and cross-modality. Users only need to simply input text, You can get the full video in 1 minute.

• Said diagrams are connected to Wenxin's cross-modal large model, realizing the ability to understand pictures, not only allowing AI to "see pictures and talk", but also to deeply understand the atmosphere and emotions of pictures. Users can upload and upload pictures to meet the needs of posting with pictures, e-commerce with pictures, etc., and to help you inspire.

• Eyan Yitu realizes the requirement of converting text requirements into visual charts. Only need simple data chart requirements, or input the data content of the icon to be generated to generate interactive charts to assist users in completing data analysis, insight and interactive presentation of chart information. It is understood that E-YiYiTu already supports the generation of 7 types of charts, including data charts, pie charts, line charts, radar charts, funnel charts, mind maps, and scatter charts.

At the scene, Wu Tian demonstrated the way of applying plug-ins in Wenxinyiyan. In the process of talking with AI, you can now let the big model summarize the content of long articles, display data in charts, read images, generate copywriting, and even Can synthesize video with voice. It took only 5 minutes for Wenxin Yiyan to complete the work of a complete scene from industry research, brand analysis and selection to generating promotional videos.

The plug-in further expands the capability boundary of the large model, and is also crucial to the Wenxin large model ecology. Baidu said that Wenxin Yiyan will launch more high-quality official and third-party plug-ins, and gradually open up the plug-in ecology to help developers create AI-native applications based on Wenxin's large model.

To achieve this goal, Baidu provides a plug-in development tool set based on the core technology of "Wen Xin Yi Yan", which can support the development of various types of plug-ins such as information services, tools, and innovations based on large language models. After the development is completed, it can also be closely integrated with the application layer ecology through the plug-in access platform.

The plug-in capability is officially launched and invited to test: yiyan.baidu.com/developer

At the same time, the latest upgrade of Flying Paddle AI Studio (Galaxy Community) officially launched the Galaxy Large Model Community. In the Galaxy Large Model Community, developers can obtain an integrated large-scale model development experience. At present, the Xinghe large model community has accumulated more than 300 large model creative applications, and the community also provides a wealth of functions to facilitate developers to communicate.

At the meeting, Baidu also released the latest Wenxin large-scale model "Galaxy" co-creation plan. With rich large-scale model resources and multi-level industrial ecological resources, Baidu will join hands with developers and ecological partners to activate the value of data resources and jointly build Large model plug-ins, extensive innovative AI applications.

Flying paddle open source framework v2.5, embracing large models

The reason why the Wenxin large model can subvert productivity is not only because of the innovation at the AI ​​​​algorithm level, but also the optimization of the deep learning framework.

Baidu is one of the few companies with a full-stack layout in the field of artificial intelligence, covering everything from chips to applications. At the framework level, the Paddle deep learning platform supports the production of large models upwards, improving the efficiency and flexibility of model deployment, and adapts to various hardware downwards, improving the efficiency of hardware adaptation and reducing costs.

At today's WAVE SUMMIT, the Flying Paddle open source framework officially released version 2.5, completing a comprehensive architecture upgrade and bringing new features in large model training, inference, and multi-hardware adaptation.

The key point is the joint optimization with the Wenxin model.

Through a series of new technologies in large model training, reasoning, hardware adaptation and other aspects of the Flying Paddle deep learning framework, the training and reasoning efficiency of Wenxin's large model has been greatly improved. A set of figures was given on the spot: through collaborative optimization, the training speed of Wenxin Large Model 3.5 is three times that of before optimization, and the inference speed is more than 30 times faster.

Among them, in terms of large model training, Fei Paddle and Wenxin have carried out collaborative optimization on the hardware cluster, which has increased the proportion of effective training time. At the chip level, the chip, storage, and network collaborative optimization have been carried out to improve the training throughput speed.

In terms of software, the collaborative optimization of flying paddles and model algorithms improves the efficiency of model convergence. Especially in large model training, the optimized convergence efficiency and stability greatly reduce the training time and achieve twice the result with half the effort.
Many technology companies are optimizing the training of large models, but in terms of reasoning, we are facing greater and more severe challenges. Li Yanhong once said: "When others just started to think about how to train, we have already rushed a long way in reasoning."
In terms of large model reasoning, Flying Paddle has three key links: model compression, reasoning engine, and service deployment. Carried out a full range of collaborative optimization.
In addition to adopting an adaptive Shift-SmoothQuant compression algorithm, a hybrid quantitative inference scheme combined with scenarios, and dynamic insertion batch processing technology, etc., Flying Paddle also continues to combine methods such as operator fusion acceleration and variable-length input processing acceleration to allow Wenxin large-scale model reasoning The speed is more than 30 times faster than before optimization.
In order to better support the production and application of large models, the large model suite of Flying Paddle has opened up the entire process, and has been upgraded around the six stages of large model development, training, fine-tuning, compression, reasoning, and deployment, reducing the cost of large models. Model development and application costs.
The optimization of the flying paddle framework for computing large models is also inseparable from the continuous improvement of the software and hardware collaboration capabilities. Flying Paddle provides a unified solution for the deployment of Wenxin large-scale models on various hardware, and also promotes the establishment of national standards for software and hardware adaptation.
Previously, Baidu, Sugon, Phytium, and Inspur jointly drafted the national standard "Technical Specification for Multi-Hardware Platform Adaptation of Artificial Intelligence Deep Learning Framework" led by the China Institute of Electronic Technology Standardization. Based on this standard, Paddle and more than 30 hardware manufacturers have carried out in-depth optimization of software and hardware collaboration, which has greatly improved the efficiency of software and hardware adaptation.

On top of this, the Wenxin large model has been adapted with 12 hardware partners including Nvidia, Cambrian, and Huawei, covering various types of hardware on the cloud and on the device side. At present, 25 hardware partners have jointly built the AI ​​Studio hardware ecological zone, introducing multiple ecological computing power to the AI ​​Studio large model community, and supporting developers in AI Studio-based large model development and diverse application experience.
At the basic level, the flying paddle has also completed an important upgrade. By building the basic operator system and combined operator mechanism, Paddle better integrates the neural network compiler CINN with the main framework, and realizes more general performance optimization with the help of its general compilation and optimization capabilities. Looking at the above, the automatic differentiation of the basic framework is also more complete, realizing a unified dynamic and static high-level automatic differentiation development interface, which can realize high-level automatic differentiation capabilities at a lower cost.
Ma Yanjun said that using the Paddle Compiler can achieve better performance than other mainstream frameworks in the industry.
Based on the capability upgrade of the flying paddle framework, especially the high-order automatic differentiation capability, the flying paddle open source platform has released open source AI for Science tools such as PaddleScience, PaddleHelix, and Paddle Quantum, which support complex shape obstacles and structural stress Numerous field calculation examples such as strain analysis and material molecular simulation widely support scientific research and industrial applications in cutting-edge directions such as AI + computational fluid dynamics, biological computing, and quantum computing.
After the two major upgrades of the basic framework, the new training framework of Flying Paddle has been initially formed, which not only maintains the advantages of unified dynamic and static, one-line code dynamic to static training deployment, but also further reduces the marginal cost of model performance optimization through compiler technology.
After solving various problems in the process of developing and deploying large models, the Flying Paddle platform has now achieved a lower threshold for the development of AI models, better results, and more standardized processes.

Disrupt productivity

At the conference, Baidu demonstrated the combination of large language models and intelligent work, reshaping the paradigm of people's work.

The ability of Wenxin Yiyan has been applied to Baidu's internal workflow through the intelligent work platform "Ruliu". At the scene, Li Ying, vice president of Baidu Group and chief information officer of Baidu Group, released Ruliu "Super Assistant".

It can solve most of the problems in your work. At the conference, Baidu conducted a demonstration.

Document processing is a rigid need for productivity work, and it often takes a lot of time to find and jump among a large number of documents. After the large model appears, you only need to issue instructions to the super assistant, and it can immediately find the relevant documents. If you need to learn new knowledge, the big model can generate detailed answers. If you click on the attached reference link and find that it is an English paper, you can also let the big model generate a Chinese summary.

It is said that many employees of Baidu are now using Ruliu super assistant, and AI can double the efficiency in many small details.

When it comes to improving efficiency, how to write code better is very important in technology companies. Li Ying demonstrated the Comate X intelligent programming assistant, a coding tool based on the Wenxin model, which currently supports more than 30 languages ​​and More than 10 IDEs, even some very niche languages, like automotive hardware languages.

Taking code generation as an example, Comate can generate corresponding code snippets based on natural language descriptions, and also supports automatic code generation based on comments in the code editing area. In terms of code testing capabilities, Comate can generate unit test cases for selected code, greatly reducing the time for engineers to write unit test cases and improving code quality.

Developing AI-native applications requires not only code tools, but also a development kit. Baidu proposed the Comate Stack, which includes three tools: evaluation platform iEValue, AI application development platform IPlayground and dataset hosting platform iDateSet.

Using this system, it only takes two steps to develop a vacation policy plug-in, and you don't need to input the rules, just feed the document to AI directly.

Now, the large model capability has become the AI ​​assistant of Baidu employees, and Comate has helped 80% of Baidu engineers improve programming efficiency and subvert the program development model. And this revolutionary productivity has attracted the interest of more than 100 partners.

Facing all scenarios and covering multiple industries

The application practice of Wenxin large model has covered the largest industrial scale in China.
During this period of time, the development process of large models is measured by the sky, papers and technologies are emerging one after another, and applications are constantly being updated. Baidu has always remained at the forefront of this competition - the constantly updated version of Wenxin has been applied to search, information flow, and Internet Among products such as disks and smart speakers, it is open to more ordinary users, and the results for enterprises are also very impressive.
Wenxin large model has established a complete set of large model system, in which the basic large model includes NLP (natural language understanding), CV (computer vision), cross-modal large model, and the task large model includes dialogue, search, information extraction, biological Typical tasks such as calculations.

At present, the Wenxin large model has the largest industrial scale in China. More than 150,000 companies have applied for the internal test of Wenxin Yiyan. Among them, more than 300 ecological partners have achieved test results in more than 400 specific scenarios, covering office work and improving efficiency. , knowledge management, intelligent customer service, intelligent marketing and other fields. Baidu has also jointly released 11 large-scale industry models with enterprises such as State Grid, Shanghai Pudong Development Bank, Taikang, and Geely.

Last year, Wang Haifeng pointed out that the deep learning platform plus the large model will penetrate the entire AI industry chain from hardware to scene applications, further accelerating intelligent upgrades. Today, Baidu's large-scale model AI technology stack has achieved a comprehensive layout, and deep learning and enlarged model technology have allowed AI to truly enter the stage of industrial mass production.

At the same time, Flying Paddle also announced the latest ecological data: the entire platform has gathered 8 million developers, served 220,000 enterprises and institutions, and 800,000 models based on Flying Paddle have been created.

People say that the wave of large-scale models triggered by ChatGPT during this period has brought about changes in productivity. While the large-scale model technology continues to evolve, the new capabilities of AI will eventually be implemented in the practice of various industries.

Baidu has already taken the lead in this process.

Article source: Heart of the machine official account

Author: Zenan