Author | Li Shuiqing
Edit | Heart
Zhidong reported in Chengdu on April 9 that at the Baidu Intelligent Cloud GENERATE Global Ecological Conference held today, Baidu Intelligent Cloud released a series of product rights and support plans around the large model technology stack, and launched a series of partner market rights.
In terms of computing power, Baidu Baige's AI heterogeneous computing platform continues to upgrade, and it is said that the training throughput of large models has increased by 30%~60% compared with the open source version, and the effective training time has reached 98%, supporting ultra-large-scale heterogeneous computing at the level of 10,000 cards. Baidu Intelligent Cloud announced the launch of the recruitment of strategic partners for intelligent computing, recruiting 3+ computing partners.
In terms of models, Baidu Intelligent Cloud Qianfan ModelBuilder provides 77 selected large models including Wenxin Yiyan series models and a full-process model tool chain, and announced the construction of the world's first Qianfan Large Model Innovation Lab to carry out joint innovation around 10,000+ application scenarios with partners.
In terms of application development, Baidu Intelligent Cloud Qianfan AppBuilder application development platform continues to upgrade, announcing that Agent supports the automation of the whole process of application creation for the first time, and will launch the workflow orchestration function next week, thereby lowering the threshold for large-scale application development.
At the same time, Baidu Intelligent Cloud has developed some applications or "semi-finished products" based on Qianfan for direct resale or further development by partners. For example, seven AI native application products, such as the intelligent coding assistance tool "Baidu Comate" and the knowledge management platform "Zhenzhi". Baidu Intelligent Cloud officially launched the industrial model base, and will launch the technology/product base for government affairs, transportation, automobile, finance and other industries in the future. In addition, Baidu Intelligent Cloud also launched the Thousand Sail Cup· Special incentive policy for developers of AppBuilder large models.
In terms of application sales, Baidu Intelligent Cloud has newly released the Qianfan AI native app store recruitment plan, providing high rebate incentives and marketing resource support, and plans to achieve 2000+ high-quality applications in the store in 2024.
It is worth mentioning that today, Baidu also released a new Baidu Intelligent Cloud Qianfan large model all-in-one machine, including the general version, Ascend version and Kunlun core version, which supports the out-of-the-box one-stop tool chain for large models, and has been given priority orders to partners.
1. More than 85,000 customers used Qianfan, fine-tuned 14,000 models, and the number of ecological partners increased by 5 times
Shen Dou, executive vice president of Baidu Group and president of Baidu Intelligent Cloud Business Group, said that the industrial implementation of large models is still in the early stage, but large models have begun to be really used.
▲Shen Dou, executive vice president of Baidu Group and president of Baidu Intelligent Cloud Business Group
According to Baidu's data, in the past year, the number of API calls on Baidu's intelligent cloud Qianfan large model platform has shown an exponential growth, with more than 85,000 customers using the Qianfan platform, with a total of more than 14,000 models fine-tuned and 190,000 applications developed.
The ecosystem of large models is also developing rapidly. In the past six months, the number of Baidu's intelligent cloud ecosystem partners has increased by 5 times, and more than 300 applications developed by partners have entered the market through the "Qianfan App Store".
Shen Dou believes that in the past, the threshold for using AI technology was very high, and development, deployment, talent, and cost were all difficult problems, which greatly inhibited the industrialization process of AI.
Today, the strong generalization capability of large models provides a general, transferable basic capability and a new AI native application R&D paradigm for cross-industry and cross-scenario AI application development, which greatly reduces the R&D threshold and improves R&D efficiency, while also bringing fundamental changes to the industrial ecological environment.
Baidu Intelligent Cloud empowers partners with AI native technology stacks, providing technology stacks including Baidu Baige AI Heterogeneous Computing Platform, Baidu Intelligent Cloud Qianfan ModelBuilder, Baidu Intelligent Cloud Qianfan AppBuilder, etc., and formulating differentiated ecological strategies for different markets.
Second, the Qianfan platform was upgraded again, and the Qianfan large model all-in-one machine was launched to create a large model service super factory
Today, Baidu's intelligent cloud Qianfan platform has achieved a new upgrade, further attacking towards the goal of "building a large model service super factory".
Baidu Intelligent Cloud announced that Qianfan Platform Agent supports the automation of the whole process of application creation, which can be one-click global optimization, automatic recommendation tools, and automatic optimization instructions. In other words, it is to support the generation of all the information of the application in one sentence.
Taking the "Agricultural Smart Answer Assistant" as an example, this application development is based on Qianfan AppBuilder to realize the automation of the whole process, and the avatar, character instructions, tool instructions, etc. are generated through the large model with one click, which greatly reduces the development threshold.
Baidu Intelligent Cloud previews that it will launch the workflow orchestration function next week to help users quickly access custom components and flexibly orchestrate business processes.
Baidu also released a new Baidu Intelligent Cloud Qianfan large model all-in-one machine today, which supports a one-stop tool chain for large models that can be used out of the box, and has been prioritized for partners to order.
In general, in terms of partner "equipment", Baidu Intelligent Cloud provides full-stack leading technical support including the AI heterogeneous computing platform "Baige", the large model development and management platform "Qianfan ModelBuilder", and the AI native application development platform "Qianfan AppBuilder" to help partners quickly build differentiated advantages in the industry.
1. Baidu Baige AI heterogeneous computing platform: training throughput increased by 30%~60%
Baige AI heterogeneous computing platform is divided into three layers: resources, construction, and large model acceleration. In terms of computing power efficiency, which has attracted much attention in the industry, Baige's training throughput has increased by 30%~60% compared with the open source version.
In terms of resource utilization, Baige provides 98% effective training time and 95% effective bandwidth utilization. In terms of scale, Baige supports ultra-large-scale heterogeneous computing at the level of 10,000 cards, and supports services such as performance optimization of the full technology stack of 25+ mainstream large models, ultra-stable infrastructure, and automatic fault tolerance assurance.
2. Baidu Intelligent Cloud Qianfan ModelBuilder: 77 models are optional, and the whole process tool chain is available
Baidu Intelligent Cloud Qianfan ModelBuilder provides a wealth of large models and a full-process tool chain to provide enterprises with a better "potency ratio".
Qianfan not only supports Wenxin large models and series of lightweight large models, but also supports many third-party large models and vertical large models, with a total of 77. In terms of tool chain, Qianfan provides the whole process of data management, model tuning, model evaluation/optimization, inference service deployment, and prompt engineering.
At the same time, Baidu Intelligent Cloud Qianfan also cooperated with Intel to launch inference acceleration based on general-purpose CPU. In addition, it also provides rich SDK and full-stack API capabilities, allowing partners to quickly and low-threshold model development.
3. Baidu Intelligent Cloud Qianfan AppBuilder: Support the full-process automation of Agent
Baidu Intelligent Cloud Qianfan AppBuilder is committed to enhancing the effect of model application and lowering the development threshold, including basic components, advanced components, Agents, multi-channel applications and distribution. The above-mentioned updates to the agent's full-process automation and workflow orchestration functions are all from this level of platform upgrades.
This time, Baidu announced the full opening of product rights and interests, and launched the recruitment of 3+ intelligent computing strategic partners, the world's first Qianfan large model innovation laboratory, and the Qianfan Cup· Special incentives for AppBuilder large model developers, etc.
3. Launch the industrial large model base, "large model + small model", and create intelligent applications such as "Dr. Electric".
2024 is considered to be the first year of the implementation of large models, and large models are entering thousands of industries.
Today. Baidu Intelligent Cloud announced that the industrial large model base product was officially released, and the large model base for government affairs, transportation, automobile, finance and other industries will be released in the future.
It is reported that the industrial model base provides ecological partners with full-stack capabilities for industrial model application development, which is divided into five layers of architecture: resource service layer, model service layer, knowledge service layer, application development layer, and business hub layer.
The resource service layer can realize the integration and opening up of customers' "research, production, supply, marketing and service" business data, Baidu maps and other characteristic data, so as to maximize the value of data assets. At the same time, it also has the ability of device, edge, and cloud collaboration, maximizes resource scheduling and utilization efficiency through cluster management, and supports hybrid collaboration of large and small models on the business application side to achieve business application results. Based on Baidu Baige · The AI heterogeneous computing platform can improve the overall resource utilization by more than 50%.
The model service layer is preset with industrial large language models, visual large models, and multimodal large models, as well as more than 300 small models of industrial scenarios, including visual models, mechanism models, and mathematical models, to jointly ensure the practical application effect of industrial scenarios.
The knowledge service layer presets the industrial knowledge base and provides service capabilities such as knowledge generation and knowledge retrieval, which can help customers and partners continuously update and precipitate knowledge assets in various business fields.
The application development layer is mainly composed of three parts: the industrial application development center, the industrial application service center, and the industrial application market, with built-in industrial application component libraries and templates, opening up a complete closed-loop path from low-threshold development of industrial applications to final commercialization. It is estimated that the application development cycle has been greatly shortened from 3 months to 10 days on average, and the out-of-the-box accuracy rate is as high as more than 85% in terms of combined model services and knowledge services.
The business hub layer provides a "leadership cockpit" for managers, a Copilot assistant for front-line employees, and scheduling center capabilities such as business operations and maintenance management, which will reshape the new production relationship in the era of large models.
Based on the base of Baidu's intelligent cloud industrial model, Shanyuan Electronics has developed the "Dr. Electricity" intelligent application to escort the power supply system for mine production. The ability to answer questions and answer questions about professional knowledge and assist maintenance of equipment provided by Dr. Dianke can quickly solve problems such as equipment failure, system failure, and safe operation and maintenance, greatly improve the efficiency of knowledge acquisition by more than 70%, and help new employees become "masters" of power supply maintenance in seconds.
In addition, Dr. Diandi can also efficiently deal with power supply system accidents, forming a business closed-loop from rapid location of power supply system faults, accurate push of disposal suggestions, immediate execution of disposal operations, to one-click disposal of disposal plans, shortening fault disposal time by more than 30% and reducing operation and maintenance costs by 30%.
Fourth, seven AI native application products have been upgraded, fully SaaS, and the combination effect is better
In addition to providing a development platform, Baidu Intelligent Cloud has also launched 7 products to help enterprises upgrade intelligently.
1. Baidu Intelligent Cloud Yinian is a large model-driven marketing content creation platform. For example, based on this application, D&S Media automatically writes scripts for a car company and generates marketing videos, exploring a new paradigm of video marketing in the AI era.
2. Baidu Intelligent Cloud Xiling is a digital human platform that comprehensively reconstructs the large model. Xiling has been relatively mature in digital human live broadcasting, video production, digital endorsement and other aspects.
3. Baidu Intelligent Cloud Keyue is an intelligent customer service platform with a comprehensive reconstruction of the large model. For example, in the "Vehicle Butler" application, the efficiency of self-service problem solving for users has been increased to 90%, and the operational efficiency of enterprise intelligent customer service has been increased by 6 times.
4. Baidu Intelligent Cloud Zhenzhi is a knowledge management platform for large-scale model reconstruction. For example, Zhen Zhi helps users "manage, check, use, and edit" in China Southern Power Grid, helping to increase the query efficiency by 50% and the writing efficiency by 2 times.
5. Baidu Intelligent Cloud Super Assistant is a super office entrance supported by a large model. Through the Agent development engine and a variety of interactive forms, Baidu Intelligent Cloud Super Assistant creates flexible, open and compatible office efficiency improvement tools.
6. Baidu GBI is a fully intelligent generative BI product. Baidu GBI provides fully intelligent generative BI services from the whole life cycle of data such as data access, data cleaning, chart drawing, problem attribution, in-depth insight, and predictive planning.
7. Baidu Comate, a full-process intelligent coding assistance tool, has been promoted with a large number of partners.
The way users can reach these AI applications is simple, either directly through SaaS or private deployment. Various products can also be used together with each other, and the combined power is even more powerful.
Fifth, the three major market differentiation ecological strategies, market rights and interests are comprehensively upgraded
For different market partners, Baidu Intelligent Cloud has formulated a differentiated ecological strategy.
Shen Dou revealed that in the past year, Baidu Intelligent Cloud's partners mainly included two types: scenario co-creation partners and user growth partners.
Among them, scenario co-creation partners are mainly composed of "comprehensive" partners such as consulting, solutions, and services, and "application-oriented" partners who are good at software development. Another type of user growth partner is composed of dealers, distributors, agents and other partners.
Baidu Intelligent Cloud divides the large model market into: head market, value market, and high-potential market. For different types of customer markets, we adopt different models of cooperation with ecological partners.
First of all, in view of the strong comprehensiveness, high complexity, and strong personalized demands of customers in the head market, Baidu Intelligent Cloud will look for partners with strong customer relationships, in-depth understanding of the customer's industry, and deep integration and delivery capabilities of industry solutions and products.
The second type is the value market, where there are more customers than the top market, and the business scenarios are more standardized and less complex. Partners can quickly replicate the industry benchmark created for leading customers, or provide relatively lightweight customized solutions. In this process, Baidu will take the principle of "partner first", do a good job in technical and business support, participate in the customized development of solutions on demand, and work with partners to develop markets and serve customers.
Third, the high-potential market has the largest number of customers, and the business scenarios are simpler, and out-of-the-box solutions are more important. For such customers, Baidu will focus on polishing standardized products and solutions, and give partners enough concessions to ensure partners' profit margins, support partners, and fully dominate sales and services.
This time, Baidu Intelligent Cloud promotes the comprehensive upgrade of market rights and interests. Different new rights and interests are provided for three types of customers, such as 1 billion yuan of targeted business opportunities and 10 million yuan of special fund support for 10+ comprehensive partners in the head market, and a joint solution co-creation fund of up to 200,000 yuan for a single application partner of 100+ application partners in the value market.
Conclusion: The large model "Circle of Friends" has been updated to reconstruct the cloud computing ecosystem
Facing the wave of intelligent upgrading led by large models, Baidu Intelligent Cloud has worked with partners to discuss the new cloud computing ecosystem in the era of large models, and formulated differentiated ecological strategies for three target markets: head market, value market, and high-potential market, and released a series of product rights and support plans around the large model technology stack.
As Shen Dou said, the strong generalization ability of large models provides a general, transferable basic capability and a new AI native application R&D paradigm for cross-industry and cross-scenario application development, which greatly reduces the R&D threshold and improves R&D efficiency, while also bringing fundamental changes to the industrial ecological environment.