How Can AI Large Models Break Through the Circle?
-
As the year comes to an end, looking back at 2023, we can undoubtedly call it the 'Year of AI large models.' Hundreds of large models have emerged globally this year. According to related reports, China alone has seen over 200 large models, creating a veritable 'Hundred-Model Battle.'
But have you noticed a problem? Despite the abundance of large models, very few enterprises and industries are actually using them. Most large models remain confined to a small 'point,' limited to topping leaderboards, benchmarking, and publishing papers, unable to truly venture into the vast landscape of industrial applications. To some extent, it feels a bit like 'circling the wagons and amusing oneself.'
For AI large models to truly succeed, it’s not about competing in data parameters or model scale but about the ultimate depth and breadth of value. Large models need to break through the circle and reach industries, enterprises, and users.
So, how can the breakthrough of large models be achieved? Huawei Cloud has provided an answer through the development of its Pangu large model.
Recently, the Huawei Cloud Pangu Model Forum with the theme 'Open Collaboration, Winning the New Era of Industry AI' was successfully held in Shenzhen. The event unveiled three foundational solutions built on the Pangu model and announced Huawei Cloud's AI global expansion plan.
Image source note: The image was generated by AI, with image licensing provided by Midjourney.
During the forum, Dong Libin, Director of Huawei Cloud Marketing Department, delivered a keynote speech titled 'AI for Industries, Open Collaboration, Winning the New Era of Industry AI.' He stated: 'To enable every industry and enterprise to quickly utilize and build large model capabilities, achieving innovation and upgrades based on large models, Huawei Cloud will adhere to 'AI for Industries,' with the Pangu model at its core, continuously driving technological innovation. Based on Ascend AI cloud services, we will provide enterprises with robust AI computing power. Through co-creation, we will enable scenario-based innovation and foster solution prosperity. Additionally, Huawei Cloud will offer a large model development toolchain, AI capability invocation and joint innovation application models, a comprehensive collaborative ecosystem, and a global promotion strategy, working openly with customers and partners to accelerate mutual business success.'
Analyzing Huawei Cloud's promotion and empowerment strategies centered around the Pangu model reveals three key dimensions of work required to break through industry barriers with large models. These insights can be valuable for the cloud computing and AI industries to absorb.
AI Large Models
Why is it difficult to break through?
When discussing AI large models, we often emphasize model parameters, data scale, leaderboard rankings, and scores, but rarely explore how these models can be practically applied in industries and enterprises. It seems that large models, which should be fully integrated with various sectors, have encountered invisible barriers, confining them to a narrow academic space.
What exactly causes this phenomenon?
The physical world we inhabit is a three-dimensional space, with the X, Y, and Z axes forming the coordinates of this space. Applying this metaphor to large AI models reveals that they encounter difficulties and challenges along these three axes, which limit their developmental potential. These three walls include:
1. The Technical Wall. When large models move toward industry applications, they first encounter a series of technical issues. These include the well-known problem of AI computing power scarcity, as well as challenges such as the lack of tools during model tuning and deployment phases, and insufficient support from the application ecosystem. From computing power to application, AI large models may encounter technical bottlenecks at every layer of the stack, turning the entire process of model deployment into a barrel effect—where one unresolved issue can lead to a cascade of obstacles.
2. The Scenario Wall. Beyond the technical dimension, large models must also accommodate the needs, characteristics, and knowledge of industry users themselves. Different industries have both common and unique requirements for large models, necessitating comprehensive scenario-based solutions to support efficient and cost-effective application. Currently, such solutions are still very scarce in the large model industry.
3. The Geographic Barrier. Many people overlook that there's also a geographic barrier surrounding large models. As corporate overseas expansion and global operations become new business trends, numerous enterprises require consistent AI large model technical support and application experiences worldwide. This poses a significant challenge to cloud computing providers' global infrastructure construction and operational capabilities, currently representing a scarce resource.
The existence of these three barriers greatly limits the application scope of large models. Conversely, for large models to break through and expand their influence, they must overcome these barriers, achieving multidimensional evolution from a single point to lines, surfaces, and volumes.
The Pangu Model has made explorations and attempts in all three aspects simultaneously, standing out uniquely in the current large model landscape.
X-axis Breakthrough
Building Full-Stack Support from Computing Power to Applications
The first "breakthrough" achieved by the Pangu Model is to technically address the bottlenecks in deploying large models, solving a series of challenges from the computing power layer to the application layer.
As the core of the model layer, the Pangu Model itself provides a three-layer decoupled architecture of 5+N+X, naturally possessing excellent ecological openness. By combining the model with computing power, tools, and ecosystems, the Pangu Model can meet the diverse and complex demands of industry users for large models. This "technical breakthrough" is mainly reflected in the extension of the Pangu Model in two directions.
First, it connects downward to the computing power foundation.
At the computing power level, Huawei Cloud's Ascend AI Cloud Services effectively address a series of challenges such as AI computing scarcity and enterprise queuing delays. Huawei Cloud has established three major AI cloud computing centers in Gui'an, Ulanqab, and Wuhu, providing enterprises with robust Ascend AI computing power. Simultaneously, businesses and developers can directly access industry-leading open-source large models like LLaMA and Baichuan through the "Huawei Cloud Ascend AI Cloud Services Hundred Models and Thousand States Zone."
Next is the upward integration of tools and application ecosystems.
To better empower enterprise users and developers, Huawei Cloud offers a series of technical and ecosystem empowerment solutions on top of the models. These capabilities address challenges in fine-tuning, development, and application deployment of large models, bridging the last mile from large models to industry applications. Overall, these include:
3 comprehensive full-process toolchains: From computing power optimization and general AI development to large model development, helping enterprises accelerate large model development efficiency in a one-stop manner.
2 Application Modes: Enterprises can directly call Pangu model capabilities via API, or customize their own enterprise-specific large models based on Pangu by incorporating proprietary data.
1 Comprehensive Collaborative Ecosystem: Huawei Cloud opens full-spectrum large model ecosystem collaboration paths for three types of partners (software partners, service partners, and consulting & system integration partners). It provides AI Gallery and cloud marketplace KooGallery platforms to support model asset monetization, knowledge sharing, product listing, and transaction promotion.
1 Global Promotion Strategy: Huawei Cloud will accelerate the deployment of Pangu models, AI computing power, and optimized open-source large models across global regional nodes. Through capability co-creation, opportunity sharing, and commercial acceleration, it aims to share AI value with customers and partners worldwide.
This 3+2+1+1 model resolves the series of challenges in large model development from tools to the application ecosystem, making large models not only trainable and deployable but also providing complete operational and commercialization support.
From Ascend AI cloud services to the Pangu large model system, and then to the "3+2+1+1" empowerment, a complete full-stack support chain for large models is established, ensuring that enterprise users encounter no bottlenecks at any stack level.
Y-axis Breakthrough
Moving Towards Scenarios, Unleashing Productivity
The next challenge for industries and enterprises with large models is how to transform these models into scenario-specific solutions with the highest efficiency and lowest cost. It's important to note that large model technology is very novel, making it extremely difficult for enterprises to develop solutions. Additionally, there are many commonalities in applying large models across different industries, and frequent redundant development can lead to significant cost waste.
To address this issue, Huawei Cloud has developed three foundational solutions around the Pangu large model. These solutions help customers and partners quickly innovate AI solutions for specific scenarios, lower the application threshold of large models, and achieve efficient integration of large model technology with industry scenarios. These include:
1. Pangu Large Model + Search Solution.
Search is a primary application scenario for large models in industries, especially in finance, government affairs, and healthcare. It enables knowledge Q&A, document Q&A, and significantly enhances productivity. By deeply integrating with industry knowledge bases and combining technologies like search, GaussDB vector database, and fine-ranking, Pangu large model improves semantic understanding, generalization capability, and accuracy, achieving real-time knowledge acquisition, precise Q&A, and result traceability. For example, in the financial sector, applying Pangu large model + search solution can improve service efficiency by 10% in customer service knowledge Q&A scenarios.
2. Pangu Large Model + Digital Human Solution.
With the continuous development of digital humans, those integrated with large models are gaining popularity across industries. They significantly enhance productivity in intelligent customer service, e-commerce, and enterprise office scenarios. The Pangu large model + digital human solution supports various applications such as broadcast interaction, intelligent customer service, and office assistants. The digital human brain empowered by Pangu large model provides three key capabilities: precise intent understanding, user privacy and security protection, and a plugin center. This solution can improve digital human creation efficiency by 200% and enhance the overall interactive experience.
3. Pangu Large Model + RPA (Intelligent Process Robot) Solution.
Process robots are widely applicable and can effectively operate in government affairs, finance, legal services, accounting, retail, human resources, and other fields. The Pangu Model+RPA (Intelligent Process Robot) solution leverages the core strengths of the Pangu Model and the RPA product WeAutomate, supporting natural language interaction with the large model to invoke RPA. The execution error rate is below 0.05%, with 100% compliance, significantly reducing the risks associated with manual operations in terms of accuracy and compliance.
The emergence and development of scenario-based solutions will further lower the difficulty of large model applications for industry users and eliminate development costs. Through the Pangu Model's scenario-based solutions, combined with the customization capabilities of dedicated models, enterprises can find the best large model solution tailored to their needs, thereby breaking down the barriers of large model applications and truly integrating large models into industries.
Z-axis Breakthrough
The Right Time for Large Models to Go Global
The integration of large models with corporate globalization and overseas operations is a topic that has not received adequate attention. However, given the global high interest in AI large models and the surge in Chinese enterprises going abroad, the global support of large models is actually crucial.
By achieving a consistent large model experience globally, enterprises can more confidently pursue intelligent upgrades, leveraging large model capabilities as a competitive edge worldwide, thereby building unique intelligent industrial advantages. To assist enterprises in navigating the global market with AI, Huawei Cloud has launched the AI Global Expansion Plan.
This achievement is made possible by Huawei Cloud's consistent global development strategy. The Huawei Cloud KooVerse global network has already established a 50ms latency circle worldwide, making it the preferred choice for enterprises expanding overseas. Building on this foundation, Huawei Cloud's AI Global Expansion Plan will gradually deploy full-stack large model technological achievements on overseas nodes to help global enterprises establish competitive advantages in large models.
In terms of computing power, Huawei Cloud will provide dual-stack AI computing services globally in 2024 to meet diverse AI computing needs worldwide.
At the model level, Huawei Cloud will be the first to deploy Pangu large model capabilities including natural language processing, computer vision, multimodal, scientific computing, and prediction on overseas nodes. The natural language large model supports multiple languages including English, Arabic, and Thai.
Regarding open-source large models, Huawei Cloud will deploy over 100 open-source large models across various categories such as natural language, video/image, and multimodal on overseas nodes next year, meeting enterprises' diverse large model requirements.
Dong Libin stated that the greatest opportunity in the next decade is artificial intelligence, as the era of large models has already begun.
Through full-stack technology construction, the implementation of scenario-based solutions, and the development of AI going global, large models will no longer be limited to playing a single-point role. Instead, they will break through boundaries, releasing value into industry scenarios, enterprise production, and user experiences.
Ultimately, an era will arrive where every company can quickly, efficiently, and cost-effectively build large model capabilities.
The depth and breadth of the value of large models will shine brilliantly in this era.