Large models and databases: two-way boost in the AI era

With the advent of the AIGC era, Large Language Model (LLM) headed by GPT has become one of the hottest topics in the field of artificial intelligence today. These powerful models not only perform well in tasks such as content creative generation, language translation, and code assistance, but also have a revolutionary impact on the development of databases.

1 Big language model: a new era of human-computer interaction

Throughout the development of human civilization, language has always been an important part of the continuous evolution and progress of civilization. From the earliest oral transmission to the emergence of written words, the communication and expression methods of language have been continuously improved, allowing knowledge and ideas to be passed on across time and space.

The continuous advancement of science and technology has led to the birth of one of mankind's greatest inventions, the computer, and with it a brand new language: machine language. Machine language is a set of instructions that a computer can understand and execute. Machine language executes very efficiently inside a computer, but for humans, directly writing and reading machine language is a tedious and complex task. In order to simplify interaction with computers, humans invented assembly language, which represents machine language instructions as mnemonics, but it still requires a high level of technical skills to write and understand.

With the further development of computer technology, humans have invented high-level programming languages, which are closer to natural languages and make programming simple and humane. However, high-level programming languages are limited by compilers and interpreters, limiting their ability to express and understand complex statements. People are eager to make it easier to interact with machines, and ideally, to make machines truly understand natural language.

Under this demand, artificial intelligence emerged. For more than 60 years since its birth, people have been working hard to study natural language processing (NLP) and are committed to making machines understand natural language more accurately and execute Corresponding commands to achieve more intelligent interaction with humans.

NLP: the link between human-computer interaction

(Source: easyai.tech)

On November 30, 2022, OpenAI released ChatGPT, a large language model based on GPT technology. It demonstrated a shocking level of artificial intelligence and quickly became the focus of attention from all walks of life. Before this, there has never been a language model as powerful as ChatGPT, and its release marks a new era of human-computer interaction.

2 Powerful empowerment of major language models

The emergence of ChatGPT has triggered a new round of AI craze. In order to catch up with the wave of the times caused by ChatGPT, more and more technology companies have developed their own large language models, and the AI tools generated based on these large models are more There are countless ones, covering many fields such as programming, databases, audio, video, language translation, conversation and chat, etc.

Application areas of large models

(Source: aigeneration.substack.com)

For example, in terms of programming, Github Copilot and Mintlify are both AI code assistants based on large models. The former can generate appropriate code suggestions based on the developer's code context and comments, helping developers improve programming efficiency and quality, and reduce duplication and cumbersomeness. work and realize your ideas easily.

AI programming assistant Github Copilot

(Source: github.blog)

The latter can generate code comments based on the semantics and context of the code, reducing the burden on developers to write comments and improving the readability and maintainability of the code.

Code annotation tool Mintlify

(Source: g2.com)

In addition, large language models have also had a wide impact in other fields. In terms of writing, large language models can be used for text generation, paragraph rewriting, intelligent review, etc. In the image field, large language models can achieve functions such as image generation, image repair, and image background removal.

The big language model is not only a technology, but also an important driver of the development of the digital economy. With the vigorous development of the digital economy, data has surpassed land, labor, technology and capital to a certain extent, becoming the fifth most powerful factor of production in promoting economic growth. In the era of digital economy, massive amounts of data are generated and processed every day. Behind this, there is a technology that is particularly important. It is the "root technology" of the digital economy and an important link between upper-layer applications and underlying basic resources. It is also known as the foundation. The “crown jewel” of software is the database.

3. When large language models meet databases

Databases are core components of modern information systems and are used to store, manage, and retrieve large amounts of structured and unstructured data. With the explosive growth of data and user demands for more advanced queries and analysis, traditional database systems are facing challenges. As a result, databases began to integrate and innovate with various emerging technologies, such as cloud computing, big data, blockchain, etc., resulting in a series of new databases with more powerful functions, providing more choices and solutions for modern information systems.

So, what kind of sparks can come from the collision between large language models and databases?

3.1 Application of large models in the database field

Large language models can empower database systems in many aspects, thereby achieving better execution performance and achieving intelligence. The following are the application dimensions of some large language models in databases:

NL2SQL（Natural Language to SQL）

Traditional database interaction requires the use of Structured Query Language (SQL) or other programming languages, which may be difficult for non-technical professionals to learn and understand. NL2SQL refers to technology that converts natural language (NL) into structured query language (SQL). Its goal is to enable non-technical professionals to interact with databases using natural language without having to write complex queries.

SQL Chat is a conversational interactive SQL client tool based on a large model. It provides a user-friendly interface that enables users to interact with the database through natural language conversations.

Compared with traditional GUI mode, SQL Chat pays more attention to user-friendliness and naturalness. It simulates conversational communication between people, and users can ask it questions in a natural language-like manner without being familiar with the specific syntax and structure of SQL query statements. This chat-like interaction method allows users with non-technical backgrounds to easily communicate and query the database.

SQL Chat converts natural language into SQL query statements

By providing a more intuitive and natural interaction method, SQL Chat lowers the threshold for using SQL and provides non-technical personnel with a more convenient and friendly database operation experience. This interaction method greatly simplifies the interaction process between users and the database, and improves the usability and ease of use of the database.

Database performance optimization

Database performance optimization has always been one of the most troublesome issues for DBAs and developers. It is an extremely complex task that involves many aspects, including hardware, system design, database structure design, SQL query optimization, index strategy, and cache management. wait.

Among them, SQL query optimization is the database performance optimization method that developers are most exposed to and the most commonly used. The goal of SQL query optimization is to reduce query response time, reduce database load, and improve query efficiency through various means.

Generally speaking, the execution speed of a SQL query is related to many factors such as the quality of the SQL statement itself, the execution plan generated by the database, the database cache mechanism, the size of the data table, and the complexity of the query conditions. The execution plan of the database is related to the cache. The mechanism is determined by its own development and design specifications and cannot be easily changed. Therefore, in the same database environment, the efficiency of query execution depends on the quality of SQL query statements, the performance of high-quality SQL statements and low-quality SQL statements . The performance is worlds apart.

However, many SQL programmers are unable to write high-quality SQL statements, and even experienced DBAs spend a lot of time and energy optimizing a complex SQL query. Until the emergence of large language models, SQL tuning is no longer a nightmare for DBAs.

The large language model can analyze a given SQL query statement and provide query rewriting and optimization suggestions. It can infer potentially more efficient query methods based on the structure and semantics of query statements, and quickly provide corresponding optimization suggestions, greatly reducing the burden on developers and maintainers.

Use SQL Chat to optimize query statements

3.2 Database promotes the optimization and development of large models

Large language models are essentially language models based on neural network architectures that are pre-trained with large-scale data sets and have a huge number of parameters (usually in the billions or more). Computing power, algorithms, and data, as the three major elements of artificial intelligence, are also important factors in promoting the development of large models.

The training and inference of large language models require a large amount of computing resources. The improvement of computing power allows the model to conduct deeper training on larger data sets, thereby improving its language understanding and generation capabilities; continuously improved algorithms can optimize the model The structure and training method make it more effective in utilizing computing resources, accelerating the convergence process, and improving training efficiency; data is the key to the emergence of large model capabilities. Large language models are completely driven by data , and the training process requires a large amount of data resources. , the quantity, quality, and diversity of training data are crucial for training large language models.

As a core tool for storing and managing data, databases can provide efficient data storage and retrieval capabilities and support the training of large language models. By storing data in a database, batch reading and processing can be easily performed, improving data availability and training efficiency.

Taking ChatGPT, the most popular large language model at present, as an example, the GPT-3 model has as many as 175 billion parameters. Data shows that the total computing power consumption required for a GPT-3 model training is 3640 PF-days, costing about 12 million US dollars; What is even more shocking is that according to information collected by industry insiders, the newly released GPT-4 model has a parameter volume as high as 1.76 trillion. The larger the number of parameters, the smarter the model, but the greater the overhead. Computing power requirements are closely related to parameter magnitude. Parameter magnitude is also an important reference for currently measuring the quality of large model training. In other words, computing power is the underlying source of power for training large models. An excellent computing power base can greatly improve the training effect of large models. The success of ChatGPT benefits from the powerful cloud computing services provided by Microsoft Azure.

The computing power requirements for training large models are growing rapidly

(Source: blogs.nvidia.com)

It can be seen that for enterprises that want to have their own large models, huge data computing requirements and high computing costs are two "big mountains" standing in front of them. Even if they obtain the code of a complex large model, it is not Anyone can run. Therefore, the large language model is not only the result of complex algorithms, but also relies on the support of cloud computing services , including the supply of resources in various aspects such as computing, storage, and databases.

4 large model + database: 1+1>2

The integration of large language models and databases will promote the development of human-computer interaction and database applications. The combination of the two is a win-win situation. By leveraging the language understanding and generation capabilities of large language models, the use and management of databases will become more convenient and efficient. Intelligent; the database provides high-quality data sets and efficient data management to support the training and application of large language models. The combination of databases and large models is bound to become a major trend in the development of both in the future.

Tuoshupai large-model data computing system (πDataComputing System, abbreviated as πDataCS) will be released at the company's annual technology forum on October 24 this year. It aims to become the basic technology base of AI, with strong technological innovation and leading product strength. The industry has high hopes, believing that the large model data computing system will open a new paradigm of AI technology.