AI big model planting grass artifact saves banks millions of dollars per second？ 0 threshold／4 hours training／5 minutes deployment
Editor: editorial department
Just some time ago, after ChatGPT ushered in "iPhone Moment", OpenAI planned to launch the LLM version of the App Store.
To put it simply, OpenAI’s unique "big model application store" is to pull the big models on the market into a two-way docking platform.
In this way, developers sell AI models customized based on ChatGPT according to special purposes. And enterprises can quickly find an adaptive LLM according to their own needs.
OpenAI’s move means that the ecosystem will be integrated, and it will be easier for everyone to obtain and use various applications without repeating "making wheels".
Coincidentally, the domestic wave information is taking such an ecological opening road.
More interestingly, Inspur Information created the AIStore as a content bearing platform for the meta-brain ecology. After LLM exploded, OpenAI also built the APPstore.
Meta-brain AIStore, the "App Store" in the era of intelligent computing.
How is the ecological opening of Inspur Information?
One vivid example is the meta-brain AIStore platform it created.
Meta brain AIStore,,,
Exploding the AI big model and planting grass artifacts will immediately open the AIGC era.
Here, the "left-handed partner" is responsible for providing scene-based applications and technologies; When the "right-handed partner" needs business innovation, he can look for cooperation on the platform according to the information provided.
At the same time, "right-handed partners" can also post their own requirements on the platform, which will be undertaken by left-handed partners and then provide customized services.
As a "platform", the AIStore is to connect the two, relying on the full-stack ecological advantages such as computing power, algorithms and resource platforms to help partners realize the dual integration of technology and business, and help everyone to cooperate and market more efficiently.
If the little red book is the grass content ecology of the C-side, the AIStore platform can be understood as an online community mall for B-side applications.
Yuannao Ecology is committed to holding hands with partners to export joint solutions with leading technology and strong landing, continuously improving the serviceability, openness and easy-to-purchase of the solutions, and helping partners to obtain more business opportunities while meeting customer needs.
In this process, connection, trust and cooperation are embodied, which is the value recognition of this cooperation model. It is also in this practical exploration that the number of ecological partners of Inspur Information is increasing. It is understood that by the end of 2022, the number of ecological partners of Inspur Information has reached more than 20,000.
Nowadays, on the platform of Yuannao AlStore, we can already see many solutions for different AI business scenarios, many of which are jointly developed by more than two partners.
In addition, in order to make it easier for customers to get a more intuitive experience, Inspur Information also added demo of some solutions to the meta-brain AIStore platform.
For example, if you open "OCR", you can try out the identification of various bills and certificates with one click.
According to reports, at present, Yuannao AlStore has settled in more than 1,200 partners, more than 150 online products and solutions, and dozens of partners’ product solutions have passed Yuannao certification.
In fact, as early as 2019, Inspur Information proposed a new ecological paradigm under the era of intelligent computing-meta-brain ecology.
Meta-brain AIStore is the content bearing platform of Meta-brain Ecology. Not only that, AIStore is also the marketing platform for partners.
The industrial demand generated by the explosion of generative AI at the end of 2022 has verified the forward-looking nature of this route.
Landing technology and establishing an open industrial ecological environment
ChatGPT detonation technology innovation has completely brought about breakthroughs in new generation technologies such as generative AI.
How to integrate technology and scenes to empower the industry has become a focus of the new round of scientific and technological revolution.
So under such a background, how will the ecology be reconstructed? Where is the innovation of "change" and "unchanging"?
This is the dilemma faced by the new industrialization in the last mile.
We see that the development of large-scale language models is becoming more and more mature, and how to realize the transformation is the key link to show the real strength of large-scale models.
The relationship between technology and industry is like the relationship between fur and leather. With the existence of leather, fur is possible and valuable.
"Skin" is an industry, but also a market demand. Only by truly empowering the scene and solving the market demand can technology continue to gain profits and develop.
For some specific scenarios, if there is no customized adaptation, the value of technology cannot be 100%.
In other words, if there is no supporting production line and demand market, the technology made will also deviate from the original intention of industrialization.
The LLM "application store" that OpenAI plans to build is precisely to promote the formation of a new ecology of technology empowerment scenarios.
For example, Khan Academy, an educational APP manufacturer, developed a personalized AI tutor Khanmigo on the basis of ChatGPT.
Aquant, another enterprise AI platform, used unique data to fine-tune the ChatGPT model and created a chat bot application Service Copilot. The developed products can intelligently answer customers’ questions about equipment maintenance and repair.
It can be seen that facing the development of future industries, the ecology needs to change according to the situation, and it needs to be reformed and reconstructed with the attitude of "competition, cooperation and openness" in the changing situation. The advantages of this are:
First of all, improve the innovation ability of technology from scene practice. Technology promotes industrial development, and then feeds back technological innovation.
Cross-border integration can promote the sharing and collaboration of data, technology, knowledge and other services, thus promoting innovation in different fields.
Enhance the competitiveness of individual partners with an open and win-win attitude, so as to enhance the efficiency and competitiveness of the whole industry.
Secondly, optimize the industrial layout and give play to the competitive advantage of a single link in the ecology.
The future industrial ecology needs to maximize the value of a single partner according to the development characteristics and advantages of different links, and realize industrial optimization and upgrading and sustainable development.
The meta-brain ecology of Inspur Information has been on the road of this new paradigm of cooperation since 2019.
That is, using technology to empower the scene, using "smart computing" to gather the power of ecology, establishing a scene community, and getting through the last mile of industrial AI landing.
Meta-brain ecology, a new ecological paradigm in the era of intelligent calculation
Meta-brain ecology is to open an ecological paradigm of a new era.
Now, there are more than 500 left-handed partners and more than 4,500 right-handed partners in the meta-brain ecology created by Inspur Information.
The "convergence of industrial power" they always advocate refers to the "left-handed partner" with the core competence of AI development and the "right-handed partner" with the overall solution delivery capability of industry AI.
In short, the meta-brain ecology does not only belong to a certain enterprise, but is a combination of three elements: left-hand partner+right-hand partner+inspur information, which is a tropical rainforest-style cooperation paradigm, so as to realize the situation of complementary advantages, mutual needs and win-win cooperation.
In other words, the meta-brain ecology will become a fertile soil to promote the integration and development of the AI industry.
This ecological construction has achieved remarkable results in concrete practice.
In the financial industry, many businesses have achieved digital transformation and improved efficiency.
I have to admit that there are still some problems that cannot be solved urgently with AI.
For example, in bill processing, many banks still use manual entry, which is not only time-consuming but also expensive, and Liuzhou Bank is one of them.
At present, with the development of OCR technology, characters in image files can be quickly recognized and converted into text.
Moreover, under the blessing of the algorithm, even if the bill background is complex, the clarity is not high, and the text is inclined, it can realize intelligent extraction.
Hehe information (left-handed partner), which has been deeply involved in the field of intelligent character recognition, is a typical representative. It has solved the problem of traditional OCR application recognition in intelligent character recognition technology, and launched TextIn Studio intelligent character training platform, which can provide hundreds of document image processing services.
In addition, unlike other industries, the banking characteristics of "high network security, high data backup requirements, high monitoring and auditing capabilities, and high emergency response capabilities" make the challenge of the financial industry more difficult.
Kelibang (right-hand partner), which has been deeply involved in the field of financial privacy computing for many years, can provide overall IT and financial information solutions.
Based on the meta-brain ecology, it can fully utilize the capabilities of AI computing platform, AI resource platform and AI tool platform provided by Inspur Information.
Then, with the AI brain of this industry, how do the right-hand partners of Inspur Information play their strengths and create a new ecological paradigm in the case of Liuzhou Bank?
In the construction of OCR intelligent platform of Liuzhou Bank, we are faced with an extremely difficult practical problem:
On the one hand, banks are faced with the core bottleneck of numerous data and complex scenarios. With the high explosion and high growth of data, higher requirements are put forward for model training and reasoning.
On the other hand, considering the dimension of user data security, all data labeling and training of customized models must be completed in the bank intranet.
As a result, in view of the practical difficulties of Liuzhou Bank, such as "less computing resources and more scene applications", Hehe Information and Klebon jointly created an industry-leading one-stop intelligent OCR solution.
In this way, through the ecological synergy, we can cross the gap between industries and get through the last mile of financial digital intelligence transformation.
Then, how do they help Liuzhou Bank solve the pain points in its business based on Inspur information meta-brain ecology?
First of all, from the aspect of computing power.
With the advent of the AIGC era, many large language models put great demands on computing power when training and reasoning based on massive data sets.
On the technical side, the traditional banking business requires high-explosive and high-growth business data processing and a large number of customized model training.
Only the bottom computing system is powerful enough to meet this extremely high requirement.
However, when the hardware performance has reached the peak, it is necessary to use the software platform to achieve more precise scheduling of computing power. By optimizing the bottom resource scheduling, the maximum utilization of computing power is realized.
According to the tidal characteristics of Liuzhou bank’s business flow, Hopewell Information and Klebon quickly transplanted and optimized the computing power allocation of the solution, so as to realize the balanced allocation of computing power resources.
Specifically, based on the characteristics of peak and trough of business traffic, the automatic elastic contraction of computing power can improve the resource utilization rate by about 40% on the premise of ensuring that the business is not affected.
In addition, based on the conventional concurrent requirements of multi-business scenarios in banks, one card can be used for multiple purposes through fine-grained segmentation of computing resources.
Maximize the utilization rate of computing power, and increase the utilization rate of computing power to 3-5 times with almost zero performance loss (≈1%).
Secondly, from the perspective of algorithm framework.
According to the problems of Liuzhou Bank, such as insufficient initialization samples, lack of data labeling, and complicated voucher plate, Hopewell Information has customized the model, and then conducted targeted training for complex documents and tickets.
Therefore, a high-sensitivity reinforcement learning and training mechanism is constructed, which greatly improves the accuracy and efficiency of OCR ticket recognition.
This scheme will maximize the algorithm recognition ability of Hehe information and realize one-stop coverage of end-to-end AI services.
The AIStation, a resource platform under the meta-brain ecology, helps it realize one-click deployment of the model, and can quickly embed AI applications such as witness verification, image loss determination and document recognition into banking business processes.
It is worth mentioning that in the whole process, the model training time has also been greatly shortened, from 2 days to 4 hours, and the deployment time has also been shortened from 2 days to 5 minutes.
One-stop intelligent OCR solution enables financial institutions to develop OCR at the "zero threshold". Based on dozens of ticket samples, developers can complete OCR model development and realize rapid deployment and training.
Obviously, this can greatly speed up the development and online application of intelligent OCR in financial institutions. At the same time, it can meet the needs of intelligent OCR applications in specific scenarios to the maximum extent.
Practice has proved that the one-stop intelligent OCR solution can improve the ticket entry efficiency of Liuzhou Bank by more than 100 times, save millions of yuan in human resources costs and speed up business.
In the future, this OCR solution is not only limited to Liuzhou Bank, but also applicable to financial institutions such as securities and insurance.
In addition to computing power and algorithms, in terms of intelligent platform operation and maintenance, the AIStation platform based on Inspur information continuously strengthens the business security guarantee capability of upper-level applications by establishing a stable intelligent fault-tolerant mechanism.
Really realized the escort for the last mile of financial digital intelligence transformation.
It can be seen that a robust artificial intelligence application system is essential in the process of project landing.
The intelligent business production innovation platform launched by Inspur Information, AI Station, meets this demand.
Specifically, AIStation is an end-to-end platform specially designed for the development and deployment of artificial intelligence, which can realize one-stop efficient delivery from the whole process of model development, training, deployment, testing, release and service.
Based on resource management and scheduling, platform process support capabilities, it has carried out a lot of matching and certification work with Meta-brain partners at different levels, such as heterogeneous chips, innovative production tools, industry partner applications and customer solutions, and accumulated a lot of experience, models and solutions, which has become an important bearing platform for Meta-brain ecology.
For example, the FlagAI one-stop big model tool of Zhiyuan Research Institute is localized and deployed based on the AIStation platform.
As we all know, the training of large model needs to build a systematic and distributed training environment including computing, network, storage and framework. The traditional decentralized management makes the overall coordination of the platform poor and the training efficiency low.
AIStation realizes the unified pool management of heterogeneous computing power clusters, and automatically configures the computing, storage and network environment at the bottom of the training through an adaptive system. Through a variety of efficient resource management and scheduling strategies, AIStation can realize millisecond scheduling of Wanka cluster and improve the overall resource utilization rate to over 70%.
Through the data caching mechanism, AIStation can improve the model training efficiency by 200%-300%. Moreover, it also meets the strong requirements of large model training such as robustness and stability.
According to the change of service resource demand, AIStation can adjust resource allocation in time, realize second-level service expansion and contraction, and support large-scale AI reasoning service scenarios with millions of levels of high concurrency. The average delay of service response is less than 1ms, and the response efficiency of sudden access peak is improved by 50%.
On June 25th, AIStation, supported by professional AI development and deployment capabilities, effectively lowered the threshold of configuration and maintenance of large-scale distributed AI computing platform in the large-scale model era, and won the "Product Gold Award" at the 2023 Global Artificial Intelligence Product Application Expo.
Everything in water conservancy, Pratt & Whitney AIGC
From the perspective of meta-brain ecology, Inspur information layout industry AI is the first to deploy and enter the market. From 2019 to now, it has won the support of many partners in just four years.
AIGC only started to come out of generate this year, before which everyone was in the exploratory period.
In practice, we can find that everyone is facing the problem of landing commercialization ways, such as quantitative financial intelligent customer service, big model+digital people and so on.
Many customers have very ideal and beautiful needs, but it is not so easy to realize them.
Some customers have data and resources, and their appeal is to mine based on their own data, so as to make it smarter; Some customers have scenes, hoping to make the scenes more intelligent; Some are middle integrator software service providers, hoping to be empowered.
This puts forward an all-round test for ecology.
Relatively speaking, the meta-brain ecology is not a centralized ecology, and it is more open without emphasizing who is attached to whom.
The gap and dilemma faced by different enterprises when their products land are completely different. Under the background of rapid iteration, ecology needs to be integrated and reshaped.
This is also the essence of the so-called "water conservancy everything" in the meta-brain ecology.
Unexpectedly, many customers showed unexpected enthusiasm after the Yuannao Ecology was really launched.
This is because, during this period, customers have discovered the unique advantages of Inspur Information-
1. The computing algorithm provided around the big model has a full-stack infrastructure solution service capability.
Especially in the super-large-scale computing cluster, the algorithm of migration parameters can be optimized.
Inspur information, as an infrastructure manufacturer, has a unique advantage in this respect.
I have calculation ability, but I know how to use it well. Just like a coach in a football match, he arranges tactics according to the opponent’s situation before the game, and adjusts the personnel and position in real time according to the progress of the game on the spot. In addition, we should be good at stimulating the potential of each player.
2. As a diversified computing platform, Inspur information provides a more open and diverse platform.
It can not only support the current international leading GPU, but also take the lead in supporting dozens of domestic computing power.
Whether it is a super-large cluster training scene around a large model or a specific AI reasoning scene, it has more diversified computing products and adaptability.
3. Create an ecological chain in the AIGC era through meta-brain ecology.
Whether it is the computing algorithm service promoted by Inspur Information itself, or around more partners, based on this ecology, the partners’ capabilities are openly shared with more end customers.
In this process, maybe someone is the lock and someone is the key. Once the lock finds the key, it is the business model of generate Center.
The business in the field of To B is being reshaped one by one in the era of AIGC. Maybe new business opportunities are not far away.
In the face of the mushrooming model, are we really ready for the Nuggets AIGC?
In addition to facing the challenges of technology, the importance of forming synergy through ecology has gradually emerged on the balance of technology and industry.
It is the best answer at the moment to gather the power of ecology, benefit AIGC, and make AIGC really benefit everything like water.
And this is also what Inspur information meta-brain ecology has always been practicing.
References:
https://mp.weixin.qq.com/s/x0AETTy_461fWh8NKJMg3A