The year is about to pass, and looking back on 2023, we will feel that this is undoubtedly the "Year of Big AI Models." This year, hundreds of large models emerged around the world. According to relevant reports, there are more than 200 large models in China alone, forming a veritable "Battle of 100 Models."
But there is a problem, I wonder if you have noticed it? Although there are many large models, there are very few companies and industries that actually use them. The vast majority of large models are trapped in a small "spot". They can only rank, run, and publish papers, but they cannot really go into the mountains and rivers of industry applications. More or less, it has the flavor of "enclosing one's own cuteness".
The true success of large AI models depends not on data parameters and model scale, but on the ultimate depth and breadth of value. Large models need to break out of the circle and go into the hands of industries, companies, and users.
So, how to achieve the circle-breaking movement of large models? Huawei Cloud gave an answer during the development of the Pangu model.
Recently, the Huawei Cloud Pangu Model Theme Forum·Shenzhen Station with the theme of "Opening up and flying together, a win-win new era of industry AI" was successfully held. Three basic solutions based on the Pangu Model and Huawei Cloud AI overseas were released on site. plan.
During this period, Dong Libin, Director of Huawei Cloud Marketing Department, delivered a keynote speech on "AI for Industries, Open and Fly Together, Win-win Industry AI New Era". He said: "In order to enable every industry and every enterprise to quickly use and build large model capabilities, , to achieve innovative upgrades based on large models, Huawei Cloud will adhere to 'AI for Industries', with the Pangu large model as the core, and continue technological innovation; based on Shengteng AI cloud services, it will provide enterprises with a surging supply of AI computing power; Innovation methods enable scenario-based innovation and stimulate the prosperity of solutions. At the same time, Huawei Cloud will also provide large model development tool chains, AI capability calling and joint innovation application models, global collaborative ecosystems, and global promotion strategies, open to customer partners Fly together to accelerate win-win business."
Summarizing and analyzing the promotion and empowerment strategies specified by Huawei Cloud around the Pangu large model, we will find that there are three dimensions of work that need to be completed to break the circle of the large model. These experiences can be absorbed by the cloud computing and AI industries.
AI large model
Why is it so hard to break the circle?
When it comes to large AI models, we always emphasize model parameters, data scale, rankings, ratings, etc., but rarely discuss how large models are actually used by industries and enterprises. It seems that the large model that should be fully integrated with the industry has encountered some invisible walls and is trapped in a very cramped and narrow academic space.
What exactly causes this phenomenon to occur?
The physical world we live in is a three-dimensional space, and the three axes X\Y\Z construct the coordinates of the three-dimensional space. Applying this metaphor to large models, you will find that large AI models will encounter difficulties and challenges on these three axes, thus limiting their own development possibilities. These three walls include:
1. Wall of technology. When large models are applied in industries, they will first encounter a series of technical problems. These include the most famous problem of scarcity of AI computing power, as well as the challenges of lack of tools that are easily encountered in the parameter adjustment and deployment stages of large models, and the challenge of insufficient application ecological support. From computing power to application, large AI models may encounter technical stuck points at each stack layer, causing the implementation of the entire large model to become a barrel effect, that is, failure to solve a problem will lead to a series of obstacles.
2. Wall of scene. In addition to the technical dimension, large models must also take into account the needs, characteristics and knowledge of industry users themselves. The demands for large models in different industries have both commonalities and characteristics, and require complete scenario-based solutions to support them in order to achieve high-efficiency and low-cost application of large models, which is still very lacking in the current large model industry.
3. Geographic wall. Many people have not noticed that there is a geographical wall surrounding the large model. As companies go overseas and operate globally, they become new business trends. A large number of companies need consistent AI large model technical support and application experience on a global scale. This poses a great challenge to cloud computing providers’ global infrastructure construction and global operations capabilities, and is a scarce resource at this stage.
The existence of these three walls greatly limits the application range of large models. On the other hand, if a large model wants to break out of the circle, it must also break this wall and realize the multi-dimensional evolution of the large model from a point to a line, surface, and body.
The Pangu large model explores and attempts in three aspects at the same time, which is unique in the current field of large models.
X-axis broken circle
Build full-stack support from computing power to applications
The first "breaking circle" achieved by Pangu's large model is to technically break through the stuck points of large model implementation and solve a series of large model challenges from the computing power layer to the application layer.
As the core of the model layer, the Pangu model itself provides a 5+N+X three-layer decoupling architecture, which naturally has good ecological openness. Through the combination of models, computing power, tools, and ecology, Pangu Large Model can solve the diverse and complex large model needs of industry users. This kind of "technological breakthrough" is mainly reflected in the extension of the Pangu large model in both upper and lower directions.
The first is to open up the computing power base downwards.
At the computing power level, Huawei Cloud Ascend AI cloud service can effectively solve a series of problems caused by the scarcity of AI computing power and enterprises waiting in line. Huawei Cloud has built three major AI cloud computing power centers in Gui'an, Ulanqab, and Wuhu, which can provide enterprises with surging AI computing power. At the same time, enterprises and developers can also directly use the industry's mainstream open source large models, such as LLaMA, Baichuan, etc., through the "Huawei Cloud Ascend AI Cloud Service Hundred-Mode Zone".
The next step is to open up the tool and application ecosystem.
In order to better empower enterprise users and developers, Huawei Cloud also provides a series of technical and ecological empowerment solutions based on the model. These capabilities can solve the problems in the process of tuning, development, and application implementation of large models, and connect the last mile from large models to industry applications. Overall, these include:
3 full process tool chain: From computing power tuning, general AI development, to large model development, it helps enterprises accelerate the efficiency of large model development in one stop.
2 application modes: Enterprises can directly call the Pangu large model capabilities through the API, or they can customize enterprise-specific large models based on the Pangu large model and combined with their own data.
1 global collaborative ecosystem:Huawei Cloud opens a comprehensive large-model ecological cooperation path to three types of partners: software partners, service partners, consulting and system integration partners, and provides AI Gallery and cloud store KooGallery platform provide complete platform support for large model asset realization, knowledge sharing, product listing, transaction promotion, etc.
1 global promotion strategy:Huawei Cloud will accelerate the launch of Pangu large model, AI computing power and tuned open source large model at nodes in various regions around the world, and share capabilities through capability sharing. Build, share business opportunities, accelerate business, and share AI value with customers and partners.
This 3+2+1+1 model solves a series of problems for large models from development tools to application ecology, so that large models can not only be trained and deployed, but also receive complete operation and commercial support.
From Shengteng AI cloud service, to Pangu large model system, to "3+2+1+1" empowerment, a complete large model full-stack support chain has been created, and enterprise users will not be in any stack layer. Encountered stuck points with large models.
Y-axis broken circle
Go to the scene and release productivity
The next big model problem faced by industry and enterprise users is how to turn large models into scenario-based solutions they need with the highest efficiency and lowest cost. You must know that large model technology is very new, and it is very difficult for enterprises to develop solutions. At the same time, there are many commonalities among large application model scenarios in different industries, and frequent repeated development will cause a huge waste of costs.
In order to solve this problem, Huawei Cloud has created three basic solutions around the Pangu large model, which can help customers and partners quickly implement AI solution innovation for segmented scenarios, lower the threshold for large model application, and realize the integration of large model technology and industry Efficient scene fusion. These include:
1. Pangu large model + search solution.
Search is the main scenario for the application of large models in industries, especially in finance, government affairs, medical and other industries. Search can provide knowledge question and answer, document question and answer and other applications, thereby greatly releasing industry productivity. By deeply integrating the Pangu large model with the industry knowledge base, and combining technologies such as search, GaussDB vector database, and fine ranking, it can improve search capabilities in semantic understanding, generalization capabilities, accuracy, etc., and achieve real-time knowledge acquisition, precise question and answer, and The results are traceable. Taking the financial industry as an example, by applying the Pangu large model + search solution, the service efficiency of agent knowledge question and answer scenarios can be improved by 10%.
2. Pangu large model + digital human solution.
With the continuous development of digital humans, digital humans combined with large models are favored by more and more industries. It has a great productivity-improving effect in intelligent customer service, e-commerce, and corporate offices. Pangu's large model + digital human solution can provide support for various digital human application scenarios such as broadcast interaction, intelligent customer service, and office assistants. The digital human brain center blessed by Pangu's large model can provide three major capabilities: accurate intention understanding, user privacy and security protection, and plug-in center. The creation efficiency of digital humans based on this solution can be increased by 200%, and the final interactive experience of digital humans is comprehensively improved.
3. Pangu large model + RPA (intelligent process robot) solution.
Process robots have a wide range of applications and can be effectively used in government affairs, finance, legal affairs, finance, retail, human resources and other fields. The Pangu large model + RPA (intelligent process robot) solution fully leverages the core advantages of Pangu large model and RPA product WeAutomate, supports natural language interactive calling of RPA by large models, has an execution error rate of less than 0.05%, and a legal compliance degree of 100 %, significantly reducing the accuracy and compliance risks of manual operations.
The emergence and development of scenario-based solutions will further reduce the difficulty of large model application for industry users and eliminate development costs. Through Pangu's scenario-based solutions for large models, coupled with the parameter adjustment and customization capabilities of exclusive models, companies can find the best large model solution that suits their needs, thereby breaking the scene wall of large models and allowing large models to truly integrate into the industry.
Z-axis broken circle
It’s time for big models to go to sea
The combination of large models with enterprises going overseas and global operations is a topic that has not yet received effective attention. But in fact, with the world paying extremely high attention to large AI models and the climax of Chinese companies going global, global support for large models is actually crucial.
By obtaining a globally consistent large model experience, companies can more easily carry out intelligent upgrades and use large model capabilities as global corporate competitiveness, thereby building unique intelligent industry advantages. In order to help enterprises sail the sea with AI, Huawei Cloud launched the AI overseas plan.
This can be achieved because Huawei Cloud has always adhered to global development planning. Huawei Cloud KooVerse global network has created a 50ms latency circle around the world, becoming the first choice for overseas enterprises. On this basis, Huawei Cloud AI's overseas plan will gradually launch large-model full-stack technology results in overseas nodes to help overseas companies build large-model advantages.
Among them, in terms of computing power, Huawei Cloud will continue to provide AI dual-stack computing power services globally in 2024 to meet the world's diverse AI computing power needs.
At the model level, Huawei Cloud will be the first to launch Pangu natural language, vision, multi-modal, scientific computing, prediction and other large model capabilities in overseas nodes. Among them, the natural language large model supports English, Arabic, Thai and other languages.
In terms of open source large models, Huawei Cloud will successively launch more than 100 open source large models in overseas nodes next year to optimize and adapt natural language, video images, multi-modal and other categories. Meet the diverse large model needs of enterprises.
Dong Libin said that the biggest opportunity in the next ten years is artificial intelligence, and the era of large models has begun.
Through the construction of full-stack technology, the implementation of scenario-based solutions, and the development of AI overseas, large models will no longer only play a limited role at one point, but can open up boundaries and release value to industry scenarios, enterprise production, and user experience. Go among them.
Eventually, an era will come where every company can build large model capabilities quickly, efficiently, and at low cost.
The depth and breadth of the value of large models will shine in this era.
