Operator AI Opportunities: Expand the new track with a large model

East Digital and Western Calculation are becoming a common opportunity in various industries, and the same is true for operators.

East Digital Calculation brings a new demand for computing power network construction and universal network upgrade needs in various industries. In this context, operators also need a series of subdivided value nodes to promote business progress and open market space. In other words, there are roads in the east, and operators need to pay more attention to which cars on the road.

In the overall architecture of the east count, the value of the AI ​​pre -training large model is worthy of attention. In recent years, the big model has been the most concerned direction in the AI ​​field. Gao Lu's model effect, excellent dataset performance, and irreplaceability in the fields of cross -modular areas have made large models become the target of the industry and academic circles. For a while, the "Big Model" became the common needs of the AI ​​industry, and for this reason, it also derived a new opportunity for supporting industries.

The computing power requirements of large models are very strict, and it requires a series of software and hardware infrastructure support, such as an artificial intelligence computing center based on Shengteng AI. However, there are differences between the computing power supply between a single computing center. Therefore, in the context of the east countless calculations, the demand for the east computing power is guided in an orderly manner to the west, optimizing the layout of the data center construction, and promoting the cooperation and linkage of the east and west. Promote the development of the large model industry.

For operators, the big model brings a series of new AI market space. The operator helps large models to conduct cross -regional training, complete computing power resource adaptation, and constitutes new service opportunities. The operator itself can also improve service capabilities and industrial efficiency.

New track of operators: market imagination brought by big model training

OpenAI released a well -known investigation report in 2018. Since 2012, the global AI computing power has increased by 300,000 times in six years, and it has continued to grow. When the era of big models officially came, the global AI computing power demand further increased. AI computing power has become the main challenge to restrict the development of the industry.

In 2018, Google released the pre -training model BERT in the NLP field, and its superior ability was quickly recognized by the industry. The GPT-3 created by OpenAI has achieved the performance of leap-up in many tasks, promoting large-scale global models to usher in explosive development, and model parameters show index levels. "Big Data+Big Model" has become the most valued development path of the AI ​​industry.

But in practice, we can find that there are many challenges in training and deployment in large models. For example, the performance of the model in the deployment scenario is insufficient; the generalization capacity of large models in different scenarios is not strong; the pre -training large model parameters are large, the amount of calculation is large, the training cost is high, etc. Among them, how to complete large model training at low cost and efficiently has become a problem that needs to be solved first. In this context, model training for modeling in the east counting and west calculation system is becoming the choice of more and more enterprises and academic institutions.

For example, the Institute of Automation of the Chinese Academy of Sciences released a cross -modular general artificial intelligence platform at the 2021 World Artificial Intelligence Conference, "Zidong. Taichu". Relying on the high-scale efficient distribution training framework, it has built a well-pre-training model, speech pre-training model and visual pre-training model with excellent performance, and developed "Vision-Text-Text-Text-Text-Text- The "three -mode state of the voice uniformly stated that the three -mode pre -training model was constructed, giving cross -modular universal artificial intelligence platform multiple core capabilities. From the perspective of the application level, "Zidong. Taichu" has both cross -modular understanding and generating ability. It has a wide range of application spaces in the direction of "Tu Sheng Yin" and "Sounds". , AI voice broadcast, AI poster creation and other application scenarios.

In 2022, Qianbo Information was based on Shengteng AI and MindSpore AI framework. Using the "Zidong. Taichu" three -mode model to create a sign language teaching and examiner, to create sign language movements and signal pictures and text in a pioneered manner. Link. At present, the all -in -one machine has been launched in dozens of schools in Hunan, which relieves the shortage of sign language teachers to a certain extent and brings important social value.

In the process of "Zidong. Taichu" big model training process, the large model of the "East Digital West Training" is very able to show the context of East Nigents and Western calculations. This model is led by the relevant team of the Institute of Automation of the Chinese Academy of Sciences. Based on Shengteng AI basic software and hardware platform, training is completed at the Wuhan Artificial Intelligence Computing Center. It can be said that in advance, it has tried the combination of industry-university-research in Eastern Data-Midwest Computing Power Training-Eastern Application ".

Taking this as a case, we can see that the new AI market of operators is open quickly. The new market that allows operators to serve large models is very important.

Shengteng base: escort the operator AI track

The big model embarked on the road of "East Digital and West Training". There are several necessary industrial conditions: algorithm, data, computing power, and the software and hardware infrastructure and computing power network required for training.

From the current stage, the training algorithms of large models have become increasingly improved, and the basic threshold for large model training in international and domestic models has been decreasing. Next is data. Enterprises, universities, and scientific research institutions in the eastern region have accumulated rich data, which can effectively invest in the training of large models. This is why China has been able to catch up although it is on the road of big models.

The next is computing power. With the plungement of the artificial intelligence computing center in various places, it is possible to complete large model training with clustering and centralized AI computing power. The AI ​​computing power system provided by Shengteng AI can be used as the foundation of the construction of the artificial intelligence computing center, or it can empower all sectors of industry, university, research and research in various ways, and bring abundant and complete AI computing power supply. At the level of software and hardware supporting facilities, Shengteng AI basic software and hardware platforms, including the ATLAS series of hardware and partnership hardware, heterogeneous computing architecture CANN, full -scene AI framework MindSpore, Shengteng applications enable MindX, full -process development tool chain Mindstudio and one -stop AI development platform Modelrts. In order to meet the training needs of more and more enterprises and academic institutions, Shengteng AI has created a full -process enable system for large models. The system includes the entire process from large model planning, large model development to the industrialization of large models, which can achieve end -to -end acceleration of large model industries.

We can see that after the demand and basic technology of large models, the large -scale and industrialized AI bases represented by Ascension AI have improved. The big model is not only able to train, but also has a new cycle of simple, efficient, and low threshold training.

The last key needs of the big model are network and network supporting services, which can become a major opportunity for operators. It can be said that Shengteng's AI track, which is a operator, has played a value -protected value.

Expand the future with AI: operator innovation justice

The importance of AI technology in the business landscape of operators is rising. As early as 2020, OMDIA's ICT-ENTERPRISE Insights announced a survey data. Nearly 80%of new operators believed that using AI technology is a "important" or "very important" IT project, of which nearly 60%of them Operators plan to increase investment in AI tools.

We can see that the scenarios and values ​​of using AI and AI in operators are already very extensive. For example, in the face of the complex operation and maintenance management needs of 5G networks, mainstream operators have used AI technology to realize network intelligent analysis, real -time prediction and automated operation and maintenance. In the smart customer service and smart call platform, NLP and multiple rounds of dialogue technology have become the key to improving the efficiency of operators' services. AI technology is also an important assistance for new models such as operators' special network services. For example, during the Winter Olympics, China Unicom combined with AI, cloud computing, blockchain technology with 5G to escort the smart Winter Olympics. At the same time, operators' investment and construction standards in AI technology are also continuously strengthened.

In this context, using AI technology and serving customers to make good use of AI technology have become the key needs of operators. The Shengteng base can bring a wide and critical supporting role to operators. For example, Zhejiang Mobile relies on the nine -day platform+liter base to build an industry AI empowerment test platform that is innovatively innovated in the whole stack, providing an artificial intelligence infrastructure for the industry, which can realize data empowerment, algorithm empowerment, application empowerment Essence The platform presets the 900+ algorithm and 800+ pre -training model, which can quickly customize the algorithm model to meet the diversified scene application needs of the city. Through the open standardized interface, the combination of modularization schemes, it provides multi -scenario AI services for urban management, and fully promotes the computing power to become a public service like hydropower "a little access, that is, use immediately".

Facing the future, we can see that the training and deployment needs of large models are emerging in industry, industry, university and research. Further, large models are part of industrial AI capabilities and applications. In addition to the general models of pre -training in general, there are more AI models and AI capabilities that need to practice cross -regional training.

At the same time, the operator itself is also the main body of large models and heavy AI capabilities. On May 16th, the "Hang Xiaoyi" released by Zhejiang Mobile on Telecommunications Day is based on "Zidong. Taichu", superimposed on the scene of Hangzhou characteristic attractions -the Southern Song Royal Street, and conduct efficient localized incremental training. Digital wisdom.

Facing the future, more and more large models will embark on the road of counting. Not long ago, the "China Computing Power Network-Intelligent Comegrated Network" was officially launched, which is an important step in the construction of China's computing power network. Next, a large number of AI tasks including large models will complete training and deployment in the computing power network. On the basis of the computing power network, operators can jointly participate in the unified market of computing power networks and realize the social -level computing power services on the entire network, which is a huge opportunity for operators. Operators can become the main force of computing power network construction, and will also benefit from the network incremental increase brought by AI.

These trends and changes are constitute a new track for operators. Large model training has increased network demand, and supporting network services are more complicated. These characteristics can bring extremely high value increases to operators, and help open up the continuous market space in the context of East Nigne West.

In front of the industrial era of large models, the operator has the last puzzle of the industrialization of the big model. Operators' attention and effort, it also helps large models to move towards large -scale applications and industrial prosperity.

- END -