With the emergence of LLM intelligence and the explosive development of APIs, all walks of life are paying attention to how to make good use of general models and how to adjust their own industry applications. The most important input to LLM is data, and the most frequent interface is the database. What impact will the popularity of model applications have on databases? What new requirements does the era of big models put forward for enterprises' data management capabilities and paradigms?
Focusing on these issues, Jian Lirong, co-founder and CEO of Kuker Data, accepted an exclusive interview with "China Electronics News" to interpret the changes and opportunities in the cloud computing and database industries in the era of large models. The following is the full text of the interview:
At present, the trend of AI reengineering industry triggered by large models has become unstoppable, especially for the underlying database supporting AI. "The rapid application of super-large language models represented by ChatGPT will lead to great changes in the cloud computing and database industries." Jian Lirong, co-founder and CEO of Beijing Kuke Data Technology Co., Ltd. (referred to as "Kook Data") recently accepted "China E-News reporter said in an exclusive interview.
The upsurge of large models will change the competition dimension of cloud computing and database markets, and accelerate the trend of enterprise IT architecture towards distributed and parallel development. At the same time, large-scale models will promote the popularization of multi-cloud, and independent database vendors with neutrality and products that support multi-cloud deployment are expected to benefit from it.
The whole link of data processing will be reshaped
The AI large model is a model based on massive multi-source data. It needs to capture knowledge from a large amount of labeled and unlabeled data through continuous training, and store the knowledge in a large number of parameters to establish an efficient processing of various tasks. Technology Architecture. It has many advantages such as general purpose and large-scale replication, and is an important direction for the realization of AGI (General Artificial Intelligence).
"'Data warehouse', 'data platform' and 'big model' are essentially to better answer decision-making problems. In a sense, they complement each other." Jian Lirong said that on the one hand, data warehouses are mature Advanced data management, cleaning, and parallel processing technologies can effectively improve the processing flow of training data required for training and fine-tuning large models; on the other hand, as a natural fact data or knowledge management platform, data warehouses can provide correct answers for generative AI The required context effectively solves the "illusion" problem that is prevalent in large models. The organic combination of data warehouse and large model can better help enterprises achieve auxiliary decision-making.
The difference is that the way large models process data is obviously different from current mainstream data warehouses and data platforms. Data warehouses and data platforms often compile raw data into two-dimensional tables, and then perform data cleaning, regularization, and completion processing, and finally realize business intelligence through complex SQL. The large model needs to continuously feed the original text information in the form of prompts, allowing it to perform deep learning, so as to achieve efficient processing of tasks, which is completely different from the traditional two-dimensional table-based storage and management of data.
Jian Lirong analyzed: "Traditional data processing consumes a lot of manpower, material resources and time, and there are many links that are very error-prone, such as data cleaning, data lineage analysis, master data management, data quality, data governance, ETL, data analysis, database operation, etc. Wei et al. The rapid application of the general artificial intelligence model represented by ChatGPT will greatly improve the automation of all links in the data processing process.”
For example, Text2SQL (that is, Text-to-SQL, refers to the process of converting natural language text into structured query language) uses the ability of large models to automatically generate structured query language based on natural language to complete BI (business intelligence) tasks , Improve the work efficiency of data engineers.
Jian Lirong said: "The emergence of large models, on the one hand, makes everyone start to think about how to use the capabilities of models to reconstruct all aspects of the data processing link to achieve a higher degree of intelligence and automation; on the other hand, it also prompts Everyone began to think about how to adapt the data processing rules of the data warehouse and data platform to the large model, so as to better support the training, tuning, deployment, reasoning and application of the large model.”
Cloud computing resource consumption patterns will be changed
As we all know, the key to large model training lies in computing power, data and algorithms. Jian Lirong believes that the cloud computing platform is the most suitable platform to provide these three elements. First, large models require a lot of computing power, especially high-end GPUs; second, massive data, especially some high-quality data; in addition, large models also need the support of algorithms, Model as a Service will become a new PaaS Serve. These are new requirements, and they are also what the cloud platform is best at. Therefore, the emergence of large models will be very effective in boosting the cloud computing market. At the same time, cloud vendors with stronger GPU computing power will have more competitive advantages.
Jian Lirong pointed out that the emergence of large models will have a subversive impact on natural language processing, computer graphics, and even autonomous driving, changing the entire software and hardware technology stack in these fields, thus bringing a new resource consumption model to the cloud computing market .
Taking SaaS services as an example, the impact of large models on low-code will be very obvious. The core value of low-code (or zero-code) is to solve the problems of slow software development and high threshold through drag and drop combination. However, the emergence of large models has subverted the entire development model of low-code. "The application scenarios that can be covered by low-code are limited, and the technology stack in the background will be completely subverted by the large model in the future." Jian Lirong said.
Large models like ChatGPT can directly create applications through natural language descriptions, and AI can generate codes much faster than humans, and can even continue to suggest improvements through dialogue. Previously, Grammarly, an AI writing tool once valued at $13 billion, saw a sharp drop in website users after ChatGPT was released.
Jian Lirong believes that the large-scale promotion of AI applications has actually increased the competition dimension of the IT industry, and different companies have different competitiveness in different dimensions. The IT layer will be more diverse, which will naturally drive the popularity of multi-cloud.
In the future, most ordinary non-tech enterprise users only need to call the MaaS service (Model as a Service) provided by cloud vendors to build their own vertical models and applications. Factors such as autonomy and controllability may be more inclined to build its own basic platform to complete proprietary model training and reasoning tasks.
Large model accelerates database distribution and parallelization
With the rise of "model fever", the huge amount of data has brought pressure on storage and computing resources, which requires the database itself to be closely integrated with cloud computing technology, and through the decoupling of metadata, computing and storage layers, so as to give full play to the cloud platform. Elasticity and scalability.
Jian Lirong believes that in the context of the explosion of large models, databases need to provide horizontal concurrent access capabilities and multi-paradigm data processing and analysis capabilities (including support for declarative language SQL, procedural language Python/R, graph computing, full-text retrieval, streaming Formula computing, high performance computing, machine learning and artificial intelligence) and storage management capabilities of massive heterogeneous data (structured data, semi-structured data, unstructured data and real-time data).
Under this trend, data warehouses based on cloud-native architecture will become an important direction for the development of the database industry in the future, and large and medium-sized enterprises usually choose a multi-cloud deployment model considering factors such as high availability and bargaining power.
"This trend also has a significant impact on our multi-cloud database enterprise. Take our core product HashData Cloud Data Warehouse as an example. At the beginning of the design, we considered the multi-cloud deployment scenario by decoupling different components of the system. The dependence on specific interfaces makes it easy to connect to various open cloud platforms and meet the needs of enterprise data flow in different cloud platforms." Jian Lirong said, "We are now developing functional components that enhance the HashData data warehouse support vector data storage and processing retrieval capabilities. Combined with the high scalability, high availability and high elasticity of cloud data warehouse, it can better support and expand the application scenarios of large models." Jian Lirong said.
At the same time, the trend of distributed and parallelization will further accelerate. "The application of large models will further accelerate the trend of distributed and parallelized database industry, and distributed and parallelized on a deeper and wider level." Jian Lirong said, "Deeper refers to more heterogeneous computing power Resources need to be parallelized, including CPUs, GPUs, DPUs, etc., and more broadly refers to projects similar to "counting from east to west", all of which require data systems to better provide sharing and collaboration capabilities."
In addition, in addition to the To C form like ChatGPT, the large model will appear in the form of personalized and independent deployment in the enterprise market, that is to say, all walks of life and even different departments of the same enterprise will have their own Instances of large models, similar to industry experts. In this context, the ability to integrate deep learning and large models in the database kernel is required.
"Whether it is all aspects of the data analysis link or the operation and maintenance of the database itself, it is very time-consuming and dependent on experience. We are trying to use large models to train expert systems in the database field, hoping to improve data analysis and database intelligent operation and maintenance capabilities. , to achieve a function similar to 'automatic driving' within the scope of the database." Jian Lirong said.