"When it comes to the current development of artificial intelligence, there is a word called 'fear of missing out', and everyone is afraid of missing out, and our customers are afraid of missing out, so we are all accelerating learning. ”
"This is a new era of artificial intelligence, we are not only witnesses of this era, we are also the builders of this era"
……
The battlefield of 100 models is still full of gunsmoke, AI Agents are surging with the trend, and artificial intelligence technology and applications are still running wildly. This year, this year's government work report proposed to carry out the "artificial intelligence +" action, which has become an important help to promote industrial upgrading and economic growth. The 2024 Inspur Information Ecosystem Partner Conference (IPF2024) Summit Forum brings together leading companies in the fields of application development, data services, computing systems, and algorithms. "Artificial intelligence +" should be "+", "what, how" +", and how to bring the power of ecology to implement "artificial intelligence +", these topics have been hotly discussed.
Where the wind rises, where the wind goes. Zuo Chun, Chairman of Zhongke Soft, Shi Shuicai, Chairman of Tors, Wang Xiangdong, Chief Scientist of Ronglian Technology, You Yang, Founder and Chairman of Luchen Technology, Wang Xiaobo, COO and Founding Partner of Zhiyuan Technology, Liu Jun, Senior Vice President of Inspur Information, and other industry leaders exchanged views in this summit dialogue, and each guest at the meeting was not only a witness to the AI revolution, but also a promoter, sharing insights and discussing how to use AI technology to empower thousands of industries and how to make the vision of "artificial intelligence +" a reality.
The summit dialogue was moderated by Chen Changhao, senior vice president of Inspur Information. On the basis of not changing the original meaning, this article has edited and sorted out the content, hoping to bring you more inspiration and thinking.
Talking Points:
■What do you think of the enlightenment of this year's IPF "Smart Yuanqi Transcendence and Cooperation"?
■In terms of strategic layout, all parties in the industrial chain will choose how to embrace "AI+"?
■How should the future ecology be coordinated?
■What role should all parties play in the AI industry chain to jointly promote "AI+"?
Wonderful views of the guests
Chairman of Zhongkesoft Zuo Chun:
For the application of large models in vertical fields, the first step is to select the base, the second step is to train the model, and the third step is to integrate the solution.
Shi Shuicai, Chairman of Tors:
The landing of the large model requires three "realities", real, practical and effective. It really refers to the engineering delivery when AI is implemented; Practicality is consistent with the user's business; The actual effect is the need to reduce costs and increase efficiency.
Wang Xiangdong, Chief Scientist of Ronglian Technology:
In the face of the storage and management needs of high computing power and massive data in the fields of biomedicine and life sciences, it is necessary to collaborate with ecological partners to meet customer requirements.
You Yang, founder and chairman of Luchen Technology:
Through the ultimate computing software and memory optimization, the added value of computing power is improved, and larger models are trained with the same equipment, helping enterprises to lower the threshold of large models in key areas as much as possible.
Wang Xiaobo, COO and founding partner of Zhiyuan Technology:
For the commercialization path of technology, we tend to change the model first, use industry cognition and model capabilities to choose a closed-loop scenario, form a data flywheel, and constantly polish our models and products to grow together with customers.
Liu Jun, Senior Vice President of Inspur Information:
In view of the diversification of computing power, model diversification and scenario diversification, Inspur Information hopes to connect and gather ecological forces to help each meta-brain ecological partner use its own scenarios and data, take the lead in the enterprise large model development platform, and then find nails when you have a hammer, and quickly enter the AI+ market.
IPF2024, what do you think?
Inspur Information Chen Changhao:
Thank you again for coming to IPF2024! This year's "Two Sessions" report put forward the "Artificial Intelligence +" strategy for the first time, and through the process of today's entire meeting, I also feel that the "Artificial Intelligence +" era has also come.
The theme of this year's IPF is "Smart Vision, Together". First of all, I would like to hear your feelings about today's conference, which point of view is the most impressive for you?
Zuo Chun: When it comes to the development of artificial intelligence, there is now a word called "fear of missing out", everyone is afraid of missing out, and our customers are also afraid of missing out, so we are accelerating learning. Today, I am particularly impressed by the efforts of Inspur Information in the system software platform, as well as the achievements in computing power and algorithms.
Tors Shishuicai: First, I feel that Inspur Information is integrating the power of full-stack AI system innovation and fully embracing AI. The second impression is that the partner ecosystem of Inspur Information is very sound. Third, the EPAI platform, an integrated end-to-end development platform for large-scale model application development, is also a highlight, which has a similar path to our Tuotian LLMOps.
Wang Xiangdong of Ronglian Technology: Two words, the first word is confidence, Mr. Peng is application-oriented and system-oriented. Application-oriented points break the characteristics of the times we live in, from technical competition to application competition. Inspur Information advocates taking the system as the core and driving application innovation with the platform, which is to compete in an ecological model. We are full of confidence in our cooperation with Inspur Information in the future. The second word is heartbeat, looking at EPAI, I am very excited, and I am thinking about how to go back and integrate with the user's application.
You Yang of Luchen Technology: My first feeling is that the products and tools released by Inspur Information today will help partners to do their business more conveniently and efficiently, especially in AI. The second point is that Inspur Information has a deep understanding of the entire technology and the general trend, which is conducive to further consolidating the power of the ecology.
Wang Xiaobo of Zhiyuan Technology: The word ecology is enough to boil blood. In the large-scale model track, China is leading in rich consumption scenarios and application scenarios, as well as very good data infrastructure. In terms of ecology, we hope to embrace the ecosystem, and we also hope that our own products also have SaaS and MaaS services.
In terms of strategic layout, all parties in the industrial chain will choose how to embrace "AI+"?
Chen Changhao of Inspur Information: First of all, as a system manufacturer, how will Inspur Information accelerate the implementation of AI+ in the face of the industrial status quo of diversified computing power, differentiated scenarios, and diversified large models?
Liu Jun of Inspur Information: If we think from the ecological model, in view of the diversification of computing power, model diversification and scenario diversification, Inspur Information expects to improve the efficiency of the entire industry in AI innovation through the meta-brain ecology, and we are very willing to become such a platform ecological carrier.
Metabrain ecological aggregation links 600+ algorithm manufacturers and 8000+ system integrators. For left-handed partners with core capabilities in AI development, we support multi-computing power and multi-mode algorithms in the Metabrain Enterprise Intelligence EPAI platform. We hope that through the EPAI tooling platform, right-handed partners can quickly and efficiently develop and implement domain-oriented large models.
The rapid development and prosperity of the ecosystem requires a product that can connect and gather strength, which is the thinking of our launch of the Enterprise Large Model Development Platform (EPAI) today. We hope that through the EPAI development platform, we can provide a toolbox and methodology for tens of thousands of partners, condense and stimulate the vitality of the ecosystem, the innovation of partners, and accelerate the intelligent transformation of thousands of industries.
Chen Changhao of Inspur Information: Zhongkesoft is currently conducting AIGC research and ecological construction in the fields of insurance and medical care. Could you please share your experience and experience with Zhongkesoft?
Zuo Chun: When a customer wants to establish an application in a vertical field, he needs to choose a pedestal, which is an initialized pedestal, which needs to carry general knowledge and low energy consumption; In addition, its technology needs to be open, which also involves open source. Because open source needs to solve the epidemic problem, that is, it needs a large number of people to use it, a lot of updates, and many people are actively maintained, and popular open source software has great value. We see that the first step of the customer is to choose the base, in the initialization stage, we need to determine the model parameters, the number of parameters and the number of tokens is not as large as possible, there should be a reasonable degree.
After we choose the base, the second step is to train the model, with the goal of training based on existing domain knowledge to achieve precision. In the training process, the manufacturer provides the base and tools, the ISV provides important knowledge sorting and knowledge training, and the customer will also put forward their own direction or focus according to the application needs. Finally, the main methods of AIGC are neural networks and vector databases, which are inductive inferences and have certain errors. The result of AI is often an upgraded version of the search function, which needs to be combined with the experience and capabilities of the existing application software to combine the two as a solution, and ultimately measure the practical effect.
These three steps must be established on the enterprise side, in the process of establishment, first of all, the basic technology lies in the manufacturer, we must learn and combine customer needs, the technology is migrated, and then built, run, and integrated with the existing system.
Chen Changhao of Inspur Information: The development of AI and large models is inseparable from high-quality data. Could you please share with you what are the core challenges encountered at the data level when the application of enterprise-level large models is implemented, and how to better transform data into "productivity"?
Tors Shicai: First of all, at the data level, there are five types of ISVs. The first is to make general software products, which have been greatly impacted with the development of open source and cloud services. The second type is to do industry application software, but there is a major challenge in the future of industry application software, and users may become our competitors, which is already very obvious in the banking industry. The third is to do manpower outsourcing services, which is also facing the risk of code farmers being replaced by AI, with higher and higher human efficiency and lower and lower returns. The fourth way is to drive software with hardware, to do things that integrate software and hardware, and to make all-in-one machines. The fifth way is the data and services that Toll thinks about. I think the future should go from software to several pieces, and the data part is very promising.
Second, we believe that data is one of the most important factors in a large model. Because the large model is the large parameters, but everyone found that the miracle of vigorous force, brute force is unsustainable, the large model has the ability, no intelligence, blindly engaged in the user side of the computing power competition, the cost is unreliable in the first place, and there is no output.
Therefore, I think that the implementation of the large model now needs three "realities", the first is real, the second is practical, and the third is practical. It really means that engineering delivery is very important when AI is implemented; Practicality needs to be highly related to the business and needs to be consistent with the user's business; The actual effect is the need to reduce costs and increase efficiency. Solving these three problems is a very important challenge for the application of large models.
Chen Changhao of Inspur Information: Ronglian Technology is making efforts to deepen the fields of biomedicine, life sciences and other fields, please talk about Mr. Wang, how to achieve application-oriented and "high computing efficiency" in the field of such scenarios with strong computing power and big data needs?
Wang Xiangdong of Ronglian Technology: Ronglian Technology transforms the business needs of the field into IT needs, and then uses IT technology and business for deep integration to assist customers in this field in digital transformation. We summarize the characteristics of biomedical, life sciences and other fields as "two highs, two more", and "two highs" refers to high computing power and high I/O iterative concurrency; "Two more" refers to diverse computing requirements, and multi-dimensional data crosses and coexists. The life sciences sector is using advanced AI technology to value data and drive its own business forward. From a technical point of view, it is essentially the computing power service of AI.
The so-called "high computing efficiency" is the dual improvement of measured performance and resource utilization. In the past, when we talked about computing power, we described how much capacity there was from the perspective of resources. But in essence, it is often encountered that customers buy the best things, but the measured performance is not brought into play, and the effectiveness is not equal. The advantage of Ronglian is to turn the user's computing power demand into technical support and provide users with efficient computing power services. With the continuous introduction of AI technology, the field of life medicine has shown a trend of multidisciplinary cross-innovation research through scenario-based application development. In view of the characteristics of high computing power, high I/O iterative concurrency, and cross-coexistence of multi-dimensional massive data, machine learning technology is used to identify tasks that can be processed in parallel and key "bottleneck" tasks, provide different computing power supply in a targeted manner, and combine multi-node concurrent execution with centralized computing power supply for bottlenecks to maximize computing power and improve business processing efficiency.
However, it is not enough to solve the scheduling problem of computing power, and the storage and management of high computing power and massive data must be combined with manufacturers to meet all customer requirements. Therefore, Ronglian and Inspur Information have carried out a lot of cooperation in this field, such as the cryo-electron microscope system of the institute and the construction of the life science laboratory platform of the hospital, of course, this also has the role of "meta-brain ecology".
Chen Changhao of Inspur Information: The open source AI large model development system launched by Luchen Technology can significantly reduce the cost of large model landing, please Professor You Yang share how to make good use of computing power and develop large models more efficiently from the AI Infra level?
You Yang of Luchen Technology: The core value of Luchen Technology is to enhance the added value of computing power. For example, if the same four or five machines can only train a model with 1 billion parameters, can they train a model with 10 billion parameters now? Through the ultimate computing software and memory optimization, the same device can be trained into a larger model. Another angle is that when the machine resources are limited, can the training speed be twice as fast to help partners and customers improve the value of computing power through optimization within limited resources?
In order to achieve the above goals, Luchen Technology mainly focuses on two points, the first point is to construct the Colossal-AI Platform system, and open-source Open-Sora, ColossalChat and other AI large model solutions, through Colossal-AI to increase the training speed by 2-7 times. Second, we focus on vertical models or vertical industries that require high computing power, and the most obvious one is the video generation scenario, and through our technology, we can help enterprises lower the threshold of video models in key fields as much as possible.
Chen Changhao of Inspur Information: Zhiyuan Technology is the pioneer of the landing of large models on the B-side. Mr. Xiaobo, please talk about how to achieve the synergy between technology and application in the development of large models, and what are the challenges in promoting the implementation of large models?
Wang Xiaobo of Zhiyuan Technology: If technology is compared to chemistry, application is chemical industry, the discovery or development of a technology, whether it is accidental or sudden, in addition to the difficulties to be solved by the technology itself, it cannot escape an economic problem. The value of commercialization is the ultimate driving force for the better development of a technology.
In what way and at what stage is the commercialization of technology is a problem faced by all technologies in the development process. China and the United States have different development directions, and at present, there is a way of thinking for domestic players to chase OpenAI, or do things on its extension line. Another way of thinking is to stand in the current industrial format and all links of the value chain, give priority to improving efficiency, or change the model first. Lingyuan Technology chose the second way of thinking, using the industry's cognition and model capabilities to first select a closed-loop scenario, grow with customers in this closed-loop scenario, form a data flywheel, and constantly polish its own models and products to apply the development of backward model technology.
How should all parties in the industry chain jointly promote "AI+" with the power of ecology?
Chen Changhao of Inspur Information: We know that there are great differences in various fields, "thousands of industries, thousands of faces", and when AI+ goes to the ground, it is inseparable from the joint efforts of ecological partners. Please summarize in one sentence, how should the ecosystem collaborate in the future, and what role will each play in the AI industry chain to jointly promote "AI+"?
Zhongke Soft Zuochun: "Value and delivery are the key"
I use two key words to summarize, one is value, to distinguish their own value, generally look at domain knowledge, some companies are strong domain knowledge, some are relatively more general. The second is delivery, the more general the delivery, the shorter the product cycle, and the longer the delivery time of the domain. On the basis of these two indicators, we found our own positioning, and then used the power of ecology to develop synergistically.
"ALL in AI, Building the Foundation with Data, Expanding the Future Intelligently"
As an ISV, to fully embrace AI, we have established a rule: if the product does not have AI genes and elements, stop developing. In terms of ecology, I think it is the right way to let partners earn more money than to earn a little more money themselves.
Wang Xiangdong of Ronglian Technology: "Meta-brain Ecology, Best Practices for the Implementation of New Quality Productivity"
In the metabrain ecosystem, Ronglian is a right-hand partner and a beneficiary. Joining the metabrain ecology is a business decision, what we expect to do can not be completed by a company alone, after the scene design involves the algorithm, every batch of data and scenes, the algorithm needs to be optimized, and the basic computing power must also be optimized, these are three types of roles, ISV+left-handed+wave information.
At the same time, the results of such cooperation are replicable in the field, and while the cost is constantly reduced, it will also greatly promote the development of field applications. Therefore, the meta-brain ecology is also a platform for creativity extraction and a best practice to promote the implementation of new quality productivity.
You Yang of Luchen Technology: "We hope to work with Inspur Information to liberate AI productivity and empower thousands of industries"
Wang Xiaobo of Zhiyuan Technology: "The ecology is open, the industry and people need to do professional things, everyone works together to make professional things bigger, and the whole ecology will prosper."
The landing of the large model is exactly the same as whether a company can sell goods to consumers, but the large model has no boundaries. I think the ecology is open, and both the industry and people need to do professional things, and everyone works together to make professional things bigger, and the whole ecology will prosper.
Chen Changhao of Inspur Information: Finally, please summarize Mr. Liu in one sentence, what strategies and thoughts does Inspur Information have for building an ecological community and promoting the implementation of AI+?
Liu Jun of Inspur Information: "On paper, I finally feel shallow, and I never know that I have to do it." This is a new era of artificial intelligence, and we are not only witnesses of this era, but also builders of this era. ”
Metabrain Ecological Partner is an organization with the ability to learn and grow, and is the only ecosystem in the industry that realizes the distribution of AI servers, the ecology of AI scheduling platform distribution and channels, and the ecology of replicating and implementing the Metabrain joint solution.
From the perspective of Inspur Information's own practice, at present, EPAI has fully empowered Inspur Information, and quickly incubated applications such as intelligent customer service, intelligent programming, and intelligent bidding assistant for operations, services, and R&D. Taking the intelligent bidding assistant as an example, it can realize the automatic identification of bidding parameters, and the recognition accuracy rate is more than 85%. And the landing of such a scenario is something that every Metabrain ecological partner can do, from their own scenarios and their own data, using EPAI tools to make it first. It's like we have a hammer in our hands and we go for a nail, which is a particularly good way to get in.