On September 5, the 2023 Baidu Cloud Intelligence Conference was held in Beijing. In order to help enterprises use and make good use of large models faster, Baidu Smart Cloud announced four major initiatives at the conference: a new upgrade of Baidu Smart Cloud Qianfan large model platform; reconstruction of digital government, finance, industry, and transportation based on Wenxin large models Four major industry solutions; released 11 AI native applications for general scenarios covering the three major areas of service marketing, office efficiency improvement, and production optimization; launched a new ecological policy for large models to provide partners with funds, computing power, technology, marketing, etc. All-around strong support.
At the meeting, Hou Zhenyu, vice president of Baidu Group, said in his keynote speech on "Large Models Drive Cloud Computing Innovation and Change", "Large models will drive innovation in cloud computing and reshape the industrial structure of cloud computing. The development of large models is different from the past. The iteration of AI technology has simultaneously driven the reconstruction of the underlying IT infrastructure and brought about changes in the upper-layer application development model."
Hou Zhenyu, Vice President of Baidu Group
The following is the transcript of the speech:
Dear guests and friends, good morning.
Just now, Dr. Shen Ji has fully shared the technological revolution brought about by large models, released Baidu Smart Cloud Qianfan Large Model Platform 2.0, and also introduced the successful implementation of large models in thousands of industries. The model era is also full of expectations.
What I want to share next is to return to cloud computing itself, and share the disruption and changes that large models will bring to cloud computing. We believe that large models will drive cloud computing innovation and reshape the cloud computing industry landscape.
The development of large models is different from previous iterations of AI technology. It also drives the reconstruction of the underlying IT infrastructure and brings about changes in the upper-level application development model.
Let’s take a look back at the technological developments over the past few years.
Since 2012, deep learning has gradually become the mainstream algorithm for artificial intelligence. In mobile applications, deep learning shines, and algorithm model capabilities have become the core competitiveness of many mobile Internet companies. However, deep learning only stops at empowering applications, and does not substantially change the R&D paradigm of applications.
Since the birth of classic cloud computing, the virtualization of computing network storage has made computing power a basic service, and the pattern of digital infrastructure has been changed by cloud computing. Mobile applications are designed with a cloud-native architecture concept that is more friendly to the cloud, which greatly improves the development and iteration efficiency of mobile applications and helps the prosperity of the mobile application market to a certain extent.
We found that the three eras of mobile applications, deep learning and cloud computing overlap. The boom of the mobile application market has also benefited from the development of deep learning and cloud computing. However, applications, AI technology and IT infrastructure still evolve independently in three parallel lines. . In the AI native era opened by large models, these three parallel lines finally meet:
>>At the application layer, the unique understanding, generation, logic, and memory capabilities of large models will be implemented in the scene as AI native applications.
>>At the same time, large models will become a universal service, that is, MaaS, which will significantly lower the threshold for AI implementation and achieve true AI inclusiveness.
>>As an infrastructure, cloud computing will be driven by the development of large models and led by AI native applications to develop into an AI native cloud, reshaping the industrial landscape of cloud computing.
Generative AI will give birth to a new research and development paradigm, which is reflected in the fact that the implementation of AI native applications requires new architectural guidance, the research and development and reasoning of AI model capabilities require new service carriers, and the infrastructure will also develop new ones with AI as the core. Computing architecture.
Let’s introduce them one by one:
First, the new architecture. AI native applications are designed for "big models" and are designed with large models as the core. They must give full play to the new characteristics of large models, so new architecture requirements are put forward for applications.
We believe that building AI native applications requires several key technical components:
>>The first is Model, which is the model capability, which will provide services in the form of API calls. The models here include basic models as well as customers' own fine-tuned models.
>>Then there is Prompt, which helps users get better answers from the model.
>>There are also Chain and Agent to realize static orchestration and dynamic orchestration, and use the unique capabilities of large models to realize chain calls.
These are model capabilities that can be better used, and are changes brought about by large language models. Based on the changes brought about by these capabilities of large models, applications naturally need to reshape their data flows and business flows.
Second, new services. Model capabilities will become the new basic service MaaS, and MaaS needs to have the three characteristics of being rich, easy to use, and native to AI. This requires:
>>First of all, the large model platform should provide a wealth of large models for customers to choose from. Because we believe that applications must be "combination of general and specialized", it is impossible for one model to solve all problems in the future, and a combination of models is needed to meet the needs of different customer scenarios.
>>Secondly, a large model platform needs to be simple and easy to use, and needs to have a full set of tool chains, covering the entire life cycle from data collection and annotation to model development, training and evaluation, to model online inference and optimization.
>>Finally, everyone knows that data is crucial to AI models, and the platform needs to help customers establish their own data closed-loop capabilities to better support customer model iterations.
Third: new calculations. Large models require high-density calculations of large amounts of data, which brings new requirements to the computing architecture.
We are seeing an acceleration in the migration of computing workloads to heterogeneous computing, and the scale is getting larger and larger. Microsecond-level interconnection has become a key capability for scaling computing power. These require us to solve them from the level of the entire system structure using an integrated approach of software and hardware.
Next, I will further introduce the new architecture of AI native applications, new model services, and new computing infrastructure for AI.
AI needs to be implemented within the scene, which will be carried by an application. I think everyone is very concerned about how to design an AI native application and what service components it contains.
Let me share here what the general business call flow looks like. The large model application process starts with a request from the user. The system first disassembles and arranges the user's intentions. The obtained multiple subtasks are usually implemented by means of domain knowledge enhancement and search enhancement. The output content is synthesized into a complete result through the large language model, and finally After being audited by the security module, it is returned to the user. The infrastructure provides support for task orchestration and debugging, system log monitoring, etc.
To summarize briefly, AI native application development requires three key capabilities: first, an efficient application development environment , Baidu will provide low-code tools, rich vertical application templates and visual debugging tools; second, rich domain enhancement support , Baidu provides self-developed vector databases, data lakes, search enhancement and other services; third, complete content security , we provide a variety of means to ensure the security of output content.
Based on the above capabilities, customers can easily build an AI native application.
As Dr. Shen Dou shared just now, everyone has seen the super capabilities of our Baidu Intelligent Cloud Qianfan Large Model Platform 2.0, which helps enterprises bridge the gap in large model applications. The Qianfan Large Model Platform is built based on the new AI-native technology stack. We provide rich functions and components at each layer of data, model, prompt, chain, and agents to make large model applications more convenient.
The characteristics of Baidu Intelligent Cloud Qianfan large model platform can be summarized as follows:
>>First, it provides multiple easy-to-use large models , including Baidu’s self-developed Wenxin large model, but also provides multiple third-party large models. At the same time, visualization tools such as retraining and fine-tuning are provided for these models to support fast and convenient model training.
>>The second is to provide a wealth of ecological tools and cooperate with the overall Baidu Intelligent Cloud services to prepare for the implementation of large-scale model applications.
>>The third is to preset rich data sets , and also support application data to be fed back to model iterative updates, so that data can actively drive business iterations.
Baidu Intelligent Cloud Qianfan large model platform provides full life cycle support for the research and development and service of large models.
At the infrastructure layer, Baidu has developed the Baige heterogeneous computing platform. The Baige platform provides: stable and highly reliable systems, high-performance training inference services and data exchange capabilities based on high-speed networks for large models and AI native applications.
>>In terms of high reliability , the stability of large-scale parallel training is one of the key infrastructure capabilities that customers are concerned about. Baige has strong fault tolerance capabilities and can achieve second-level fault sensing and minute-level automatic fault tolerance, and can achieve 10,000 Card training is uninterrupted on a monthly basis, fully meeting customers' large model training needs.
>>In terms of high performance , after Baidu’s years of accumulation in AI infrastructure, Baige has achieved a 30%+ improvement in training performance on public model libraries and a 10-fold increase in inference throughput.
>>In terms of high-speed networks , we have self-developed high-speed networks that support lower communication delays and greater communication throughput, laying the foundation for the expansion of computing power scale.
Facing the historical opportunities in the AI native era, Baidu Smart Cloud has created a full range of products. Here we show the product panorama of Baidu AI native cloud. You can see that the content I shared is only a small part of it. More product sharing will be in the "Intelligent Computing & Large Model Technology" sub-forum in the afternoon. Everyone is welcome to Come to our sub-forum from time to time.
It is our belief to realize the inclusive benefits of AI. We have been committed to allowing everyone and every organization to enjoy the convenience of AI technology, breaking the digital divide, realizing fairness and progress, making intelligence within reach, and using technology to make complex tasks more accessible. The world is simpler, thank you all!
Click "Read the original text" to review the details of the conference