Every once in a while, I will re-read Mr. Liang Qichao's "Young China".
In an era where technology is asking for answers, and when technological innovation is becoming more and more valuable, this article that we have all learned contains great value. For example, it says "If the young people of the whole country become young, China will be the country of the future, and its progress is immeasurable." Every technology is a newborn teenager, but do these technological innovations really have a youthful spirit?
Is our technology really broad-minded and willing to carry out full-stack innovation of root technology?
Are we willing to go out of the house and travel around to understand application scenarios and gain insight into industry needs for our technology?
Can our technology be integrated with the general trend and take on the responsibility of our times?
In the new round of global AI boom triggered by large models, the answers to these questions become especially important. Over the past few months, dozens of large models have been released in China. But there are always netizens asking such a question: China has made so many large models, what is the difference?
Should we stick to the old-fashioned and prudent principle of "we have what others have" and follow suit to build AI, or should we actively integrate with this land to create new intelligence that "we have what no one else has"?
If your answer is also the latter, then Pangu Large Model 3.0 is worth knowing about.
During the Huawei Developer Conference 2023 (HDC.Cloud 2023 ) held on July 7, Huawei Cloud released the Pangu Large Model 3.0. Different from other big AI models, HUAWEI CLOUD chose to go to the wilderness of the industry and write an article "Youth AI Talk".
AI model
need some youth
Liang Qichao satirized a depressive phenomenon in "Young China Talk", "Although the thunder is hovering on its top, other things are not cared about, not known, not heard."
In technological innovation, we often fall into a similar situation: when a technology becomes popular, everyone rushes to it, as long as it can make something that is almost intersecting. Today, more than ten years after the rise of the third AI, this stagnation has already emerged. For example, it has become an industry consensus that AI large models have good versatility and strong generalization. However, the application of large models is still on a more traditional track-writing poems, answering questions, running scores, and brushing records. From the model point of view, this is only a quantitative change from the application logic of deep learning in the past ten years, but lacks a qualitative change.
When the large-scale model encounters the digital China construction in full swing, countless enterprises will ask: the large-scale model in the pavilion is very good, but why can’t my industry use it?
If China's large model wants to develop, it must face and answer this question.
On the whole, the threshold for the combination of industry and large-scale models is still very high, mainly manifested in the following aspects:
1. AI technology itself has a high threshold and high cost. From the perspective of computing power cost, data cost, development cost, and talent cost, it is not easy for enterprises to apply AI, which has caused fundamental obstacles to the implementation of large-scale models.
2. Data security and infrastructure concerns. The large model is the core digital asset of an enterprise, and it needs multi-dimensional guarantees in terms of core technology and data application. In particular, the level of technological autonomy of large AI models is becoming increasingly important in the global environment.
3. The large model lacks the necessary nodes to integrate with the industry. The application of large models in the industry requires not only the versatility and generalization of the model, but also the professional knowledge and industry skills of the model.
How to overcome these challenges? The answer is not in the office or the laboratory, but in the vast land. We have to go out like teenagers, into the industry, into the wilderness to see the problems and find the answers.
In today's China, digital and real integration has become the general trend of the industrial economy. The advantage of China's economy is that there are many industries and a comprehensive industrial system. The digitalization and intelligentization of various industries are vigorous and strong, and the national digital infrastructure represented by the east and the west is developing rapidly.
The large model must be combined with these advantages and needs in order to create revolutionary technological productivity and activate the deep value of thousands of industries.
Since its release in 2021, Pangu Big Model has been thinking about the real concerns of customers in customer operations, product development, software engineering, production supply, marketing and other industries. The Pangu model, which "does not compose poems but only does things", has entered the field of the industry with a lively and lively youthful spirit.
into the wilderness of the industry
What is the difference between Pangu Large Model 3.0?
Under the idea of HUAWEI CLOUD's "AI for Industries", combined with the trend of large-scale model capabilities that the industry is most concerned about, Pangu large-scale model 3.0 has achieved a series of upgrades, centering on the three major directions of "reshaping the industry, taking root in technology, and opening up and flying together". , Continue to build core competitiveness.
In HDC.Cloud 2023, the five basic models of Pangu CV large model, Pangu forecasting large model, Pangu scientific computing large model, Pangu NLP large model, and Pangu multimodal large model were fully displayed, and a series of industry large models were released, including Large-scale models of government affairs, finance, manufacturing, and rail industries.
In terms of technical differentiation, the Pangea Model 3.0 includes a series of new technical features.
Among them, the field of large language model and multimodal model that is most concerned by the outside world, Pangu 3.0 can provide serialized basic large models with 10 billion parameters, 38 billion parameters, 71 billion parameters and 100 billion parameters, matching different scenarios, different time delays, different Complex industry needs. These models can provide a consistent set of capabilities, including knowledge question answering, copy generation, code generation, NL2SQL, plug-in calls and other capabilities of large language models that are of great concern to the outside world, as well as image generation and image understanding capabilities of multimodal large models.
At the level of application structure, the Pangea Large Model 3.0 has built a large model system that is fully oriented to the industry and designed based on industry needs, including a 5+N+X three-tier structure, that is, a basic large model of 5 L0 layers, which can provide General skills to support various applications of enterprises; N L1-level industry large models can be based on the combination of various capabilities of the basic large-scale model to help enterprises create large-scale model products that meet their own needs through data fine-tuning; X is a massive L2 layer Scenario model, the scenario model is more focused on a specific application scenario or specific business, providing customers with out-of-the-box model services.
With the blessing of this three-tier structure, Pangu Large Model 3.0 can adapt to various needs of enterprises, and through various and flexible combinations and arrangements, it forms a "thousand-faced large model" that meets the needs of thousands of industries.
How can Pangu Large Model 3.0 go deep into the industry and gain a foothold in the industry? This question requires us to remember three key words.
The red sun is rising, its way is bright
Industry precipitation behind the big model
The first keyword: industry experience.
The so-called "understanding the industry" is not a standardized ability, but the experience integration built by countless technical systems and industrial systems. There is no shortcut to accumulating industry experience. Instead, we must go in-depth one by one, study each case, go to mines, docks, and laboratories to gather industry experience.
Relying on Huawei's industry corps, HUAWEI CLOUD has accumulated more than 400 industry solutions, and jointly built 7 major industry aPaaS with partners, so as to truly accumulate the core demands, core knowledge, and core capabilities of each industry. In terms of AI, HUAWEI CLOUD AI has cooperation experience in more than 1,000 projects in various industries, and has a specific and detailed understanding of the intelligent demands of various industries. Under the general direction of Huawei Cloud AI for Industries, we can see that the Pangu model itself is deeply integrated with industry knowledge, and has learned more than 10 industry public data, covering finance, government affairs, meteorology, medical care, health, Internet, education , automobiles, retail, etc., have learned more than 50 billion tokens in each industry sector.
The industry experience with various types of industries, huge scale of cooperation, and huge amount of industry data has finally created the "great glory" of the Pangu model in the field of the industry.
Taking the large model of the mine as an example, allowing coal miners to perform remote operations in the office is the focus of the intelligent development of coal mines. However, remote operation will face the problem that a large amount of dust and water mist in the mine will block the camera and weaken the monitoring effect. Faced with this problem, the Pangea large-scale model can stitch more than 100 channels of video on the same screen, and use the dust-permeable algorithm, foreign object detection and other visual large-scale model technologies to identify large rocks, large coals, and coal bunkers in real time during the mining process. Jamming and other abnormal conditions, so that the ground staff can control the shearer operation in real time without going down the well.
Currently, based on the large model of the Pangu Mine, HUAWEI CLOUD and Shandong Energy Group have cooperated in depth to develop 21 scenario-based applications covering seven major business systems in the energy industry. The successful application of the Pangu Large Model has made the working environment of coal miners more comfortable and safe, greatly Improve the production efficiency and safety level of the mining industry.
Industry precipitation is not an overnight achievement, but years of integration and breakthroughs. The Pangu model can truly integrate into the industry because it was born for the industry early on and has been working on this road for a long time.
The milk tiger howls in the valley, the beasts tremble
AI Era of Independent Innovation
The second keyword: independent innovation.
Under the current situation, self-reliance and self-improvement in science and technology has become a national strategy and industry needs, especially in key industries involving the national economy and people's livelihood. When faced with the opportunity of large-scale models, new demands are put forward for the level of independent innovation, and the only way is to move towards full-stack independent innovation. Only in this way can we meet the intelligent needs of these industries and truly let AI enter the wilderness of the industry.
To this end, the Huawei Cloud Pangu large model has achieved full-stack independent innovation in terms of computing power, chip enablement, AI framework, and AI platform. Huawei has built an AI computing power cloud platform based on Kunpeng and Ascend at the bottom layer, as well as the heterogeneous computing architecture CANN, the full-scenario AI framework MindSpore, and the AI development production line ModelArts, etc., to provide distributed computing power for the development and operation of large models. Key capabilities such as parallel acceleration, operator and compilation optimization, and cluster-level communication optimization. Based on Huawei's AI root technology, the performance of large model training can be adjusted to 1.1 times that of mainstream GPUs in the industry.
(Huawei Executive Director, Huawei Cloud CEO Zhang Pingan)
At the same time, HUAWEI CLOUD is also actively promoting the combination of the large-scale model business and the strategy of eastern data and western computing. At the meeting, Zhang Pingan, executive director of Huawei and CEO of HUAWEI CLOUD, announced that the Ascend AI cloud service with a single cluster 2000P Flops computing power will be launched in HUAWEI CLOUD's Wulan Chabu and Gui'an AI Computing Power Center went online at the same time. Ascend Cloud service can provide longer and more stable AI computing power services. After 30 days of kilocalorie training, the long-term stability rate reaches 90%, and the breakpoint recovery time does not exceed 10 minutes.
With the development of AI technology, the intelligent leap of the social economy is in sight, and independent innovation is the unavoidable core mission of this era. Building another pole of AI in the world has become the responsibility of Huawei Cloud and an opportunity for China's technology .
The future is like the sea, and the future is long
AI cloud service based on usability
The third key word: real and usable.
The industry landing of large models needs to face a series of complex environmental challenges, talent challenges, and business challenges. Overcoming these difficulties requires not only the improvement of technical capabilities, but also the improvement of the ease of use of large models in all aspects and lowering the threshold of industrial intelligence. In this regard, we cannot build a car behind closed doors and take it for granted to develop AI and large-scale models. Instead, we need to meet the real industry needs and developer needs.
To this end, HUAWEI CLOUD has built a series of real and usable AI cloud services with usability as the guidepost. For example, the large-scale model business model we discussed above, through the three-tier model of L0, L1 to L2, fully opened up the business model of large models and solved the problems of enterprise users in terms of specificity and customization of large models.
In addition, HUAWEI CLOUD provides a variety of deployment modes including public cloud, public cloud large model area, and hybrid cloud to meet the different deployment needs of enterprises, so as to meet the deployment needs of different types of enterprises for large models.
In terms of the development friendliness of large models, HUAWEI CLOUD provides an easy-to-use and reliable large model tool suite, Kaitian aPaaS that gathers a large number of multi-industry scene APIs, and an exclusive large model community that contains rich and high-quality courses and technical certifications. Hope and development Together with researchers and partners, we will jointly explore the innovative path of combining the Pangu model with the industry.
It is mentioned in "Young China Theory", "If the young are wise, the country will be wise, if the young are rich, the country will be rich, if the young are strong, the country will be strong, and if the young are independent, the country will be independent." AI large-scale model technology is at the key point of the international strategy and industrial revolution. For the integration of digital and real information and intelligent upgrading of the social economy, the large-scale model plays the role of "AI is strong, technology is strong, and AI is independent, technology is independent". .
At this time, we need to have the courage to take responsibility and dare to explore the unknown large-scale model system. AI needs to go to industry knowledge, to independent innovation, to developers, to the interior of enterprises, to factories, to mines. Go to the wilderness, work hard, and become a young AI.
The goal of the Pangu model is to help every enterprise and everyone have their own expert assistants. In the end, the power of AI in every enterprise and every industry will trickle down into the sea, and only then can it converge into a young and promising Smart China.
Even though there are thousands of years, there are eight wastelands. The future is like the sea, and the future is long.
Only by breaking out and doing practical things can AI technology be immortal and boundless to the country.
