Source: Zhongjin Dian Dian, Author: Yu Zhonghai, Wang Zhihao, et al

summary

OpenAI released the ChatGPT Plugin, and the “operating system” of the AI era first took shape.On March 24, OpenAI officially released the plug-in function ChatGPT Plugin. With the 11 officially released plug-ins, users can already perform functions such as calculating recipe calories and ordering ingredients and air tickets online on ChatGPT. Specifically, ChatGPT Plugins is the beginning of a further ecological transformation. Improvements based on ChatGPT include the ability to access real-time data from the Internet, create and compile code, and call and create third-party programs. We believe that ChatGPT Plugins is expected to become the core entry point of the AI era, shifting from being a technology enabler of big model technology to an important ecological entry point for the platform economy. On the one hand, it can access and empower applications, and on the other hand, it can call and operate applications. Using the big model, it can also act as an “operating system” to accelerate the fine division of labor with other application layers in overseas ecosystems.

GitHub released an upgraded version of CoPilot X to comprehensively improve developers' work efficiency.On March 22, GitHub Copilot X was released. On the basis of the code generation of the original Copilot, interactive code generation, modification, interpretation, and speech generation capabilities were added, and GPT-4 was fully enabled in scenarios such as pull requests and terminal interfaces commonly used by developers. In the short term, we are optimistic that development efficiency will be greatly improved under the power of GPT; in the long run, we believe that through natural language interaction, GPT-4 is expected to develop into a “no-code platform” in the true sense of the word, leading to the evolution of the business model of the software industry.

In the future, the big model side or the Matthew effect will be remarkable, and the application side is still expected to flourish. I am firmly optimistic about the development opportunities on the application side.Under the big model route, multi-dimensional requirements such as cost, computing power, scenarios, and data create a high threshold. We think it may be difficult for small to medium participants lacking resources to keep up with the pace of technological development, and the big model pattern may become concentrated. While exploiting the ecological first-mover advantage, large domestic model manufacturers still face the risk of lower thresholds and ecological shocks brought about by potential open source GPT models. On the other hand, the big model empowers the industry with a wide range of application scenarios, and we think the application side is expected to flourish. We believe that under the big model technology revolution, we should pay more attention to data-rich closed-source application scenarios (e-commerce, search, industry, finance, architectural design, etc.), and pay attention to the progress of overseas industrialization implementation and domestic benchmarking results.

risks

Technological progress fell short of expectations, and domestic application implementation fell short of expectations.

Main text

OpenAI releases ChatGPT Plugins, and the “operating system” of the AI era is beginning to take shape

On 2023/3/24, Beijing time, OpenAI officially released the plugin function ChatGPT Plugins.ChatGPT supports plug-ins, complements the shortcoming of data timeliness, and has a series of capabilities such as online access to the latest information, perform mathematical calculations, run code, or directly call third-party API services. With 11 officially released plug-ins, users can already calculate recipe calories, order ingredients, air tickets, etc. online on ChatGPT.

Chart: OpenAI debuted 11 official plug-ins for ChatGPT Plugins

Case 1: Accessing real-time data from the Internet

For questions related to real-time information, ChatGPT has enabled the Internet to obtain information and then generate answers.Before the plugin was launched, ChatGPT was criticized the most about training data as of 2021, and the latest news could not be obtained. However, based on previous work such as WebGPT, the ChatGPT Plugins were able to browse Internet information, greatly expanding the scope of data. For example, users can ask ChatGPT questions about the latest Oscars. ChatGPT will search online and then output the results in a way that everyone is already familiar with. Furthermore, as a language model, ChatGPT can still do creative work on language, such as linking Oscar actors and movies with poetry.

Chart: ChatGPT Plugins uses the Browsing plug-in to access the internet to answer Oscar-related questions and write poetry

Case 2: Creating and compiling code

ChatGPT can run the Python compiler to process uploaded or downloaded code and files.ChatGPT allows users to upload files to the session workspace, execute the Python compiler to run code, and persist in the session for subsequent calls. This function can help programmers improve workflow efficiency, solve quantitative and qualitative mathematical problems in actual use by users, perform data analysis and visualization, and convert file formats.

Chart: ChatGPT writes code to solve math problems

Chart: ChatGPT interacts with and visualizes data files

Case 3: Invoking and Developing a Third-Party Program

ChatGPT Plugins greatly lower the threshold for calling and developing third-party programs.Currently, you only need to enter natural language, and the ChatGPT model can understand the user's needs to call the appropriate plug-in API to achieve the user's intention. In addition to calling existing plug-ins, ChatGPT also supports rapid generation of custom plug-ins through natural language interaction (such as the OpenAI official website example for building plug-ins to manage to-do items). Developers only need to describe the plug-in's functions in natural language to obtain the corresponding plug-in, greatly lowering the plug-in development threshold.

Figure: Example Manifest file for a plugin that manages to-do items

Chart: ChatGPT calls the OpenTable plugin to make restaurant reservations

Chart: ChatGPT calls the InstaCart plugin for online shopping

ChatGPT Plugins enable mutual use between models and applications, and are expected to become the “operating system” of the AI era in the future.We believe that ChatGPT Plugins may become the core entry point of the AI era, shifting from being a technology enabler of big model technology to an important ecological entry point for the platform economy. On the one hand, it is reflected in the ability to access and empower applications, and on the other hand, it can call and operate applications. Using the big model also acts as an “operating system”, accelerating the fine division of labor with other application layers in overseas ecosystems. We think Microsoft's previously released Business Chat in Copilot validates the idea of an ecosystem portal.

From Windows in the PC era to GPT in the AI era, the Microsoft card is an important ecological entry point, building high ecological barriers.When Microsoft worked as a Windows operating system in the PC era, it attached great importance to ecology, and quickly cultivated a large number of Windows-based application vendors with its first-mover advantage. In fact, ecological barriers are one of Microsoft's core barriers in the operating system field. In the AI era, we think Microsoft's ecological layout ideas are similar. In the early days of big model releases and ecological cards, the ecology was “platformized and upstream”, continuously turning first-mover advantages into ecological barriers.

In the future, the big model side or the Matthew effect will be remarkable, and the application side is still expected to flourish. We are firmly optimistic about development opportunities on the application side.Under the big model route, multi-dimensional requirements such as cost, computing power, scenarios, and data create a high threshold. We think it may be difficult for small to medium participants lacking resources to keep up with the pace of technological development, and the big model pattern may become concentrated. While exploiting the first-mover advantage of the ecosystem, large domestic model manufacturers are still facing the risk of lowering the threshold and impacting the ecosystem of big model manufacturers due to potential GPT big model open source. On the other hand, the big model empowers the industry with a wide range of application scenarios and application tasks, and we think the application side is expected to see a flourishing situation. We believe that under the big model technology revolution, we should pay more attention to data-rich closed-source application scenarios (e-commerce, search, industry, finance, architectural design, etc.), and pay attention to the progress of overseas industrialization implementation and domestic benchmarking results.

Chart: Sorting out cutting-edge overseas scenarios and domestic benchmarks

GitHub upgrades CoPilot X, GPT integrates IDE to transform software industry productivity

GitHub CoPilot accelerates code development by empowering programming with AI.GitHub CoPilot is an AI-enabled programming tool based on the GPT-3 model jointly launched by Microsoft and OpenAI in June 2021. It can automatically recommend or generate code for programmers to use. It was officially launched in June 2022 (priced at $10 per month or $100 per year). According to the GitHub CEO, since its release, GitHub has participated in writing 46% of new code through automatic comments and code generation, helping developers write code faster by 55%.

Empowered by GPT-4, the upgraded version of GitHub Copilot X fully empowers the entire development process.On March 22, 2023, GitHub launched the CoPilot X Program, integrating the GPT-4 model into the IDE. Through dialogue and terminal interaction interfaces, it supports interactive code writing, correction, explanation, speech generation, etc. during the code writing process. Copilot X can also achieve deep empowerment in the development of commonly used fields such as pull requests (pull requests), command line use, and knowledge document search, to help developers improve efficiency throughout the process. Currently, Copilot X has not been publicly tested, and developers who subscribe to Copilot can wait in line for trial by joining the shortlist.

Supports conversational code writing, correction, interpretation, and speech generation to comprehensively improve programming efficiency

GitHub Copilot Chat became the “code version” of ChatGPT.It allows developers to get a ChatGPT-like experience in the editor, and GitHub announced that Chat will be integrated into integrated development environments (IDEs) such as Visual Studio and VS Code one after another. Through the chat window, developers can write and modify code dialoguously, and can also allow Chat to explain selected codes, and even support voice code generation functions. We believe that CoPilot X, empowered by GPT-4, has greatly improved developers' productivity.

Figure: GitHub Copilot X supports code generation through conversational interactions

Participate in processing code push and command line usage to comprehensively optimize the developer user experience

GitHub Copilot X enables the code pull (pull requests) process.Pull requests (pull requests) are one of the core functions of GitHub for developers. By submitting pull requests, developers push the written code to the repository administrator, who decides whether to merge the code into the main branch. Copilot X helps developers automatically write request descriptions (which can be quickly filled in via the tab key), thereby speeding up code review and merging for the entire team.

GitHub Copilot X enables quick search of terminal command line commands.The terminal (Terminal) is also an interface often used by developers, and the terminal has a large number of command lines, making it difficult even for experienced developers to remember all the commands accurately. GitHub Copilot CLI supports fuzzy queries to find terminal instructions needed by developers by writing commands and loops.

Figure: GitHub Copilot X automatically writes a request description (left) and quickly finds command line commands (right)

The application of GPT in code development is reshaping productivity in the software industry.Currently, GPT-4 applications can cover many simple programming tasks, and can also empower developers in more complex programming processes. In the long run, we think that through natural language interaction, GPT-4 is expected to develop into a “no-code platform” in the true sense of the word, and that users with no development experience are also expected to become developers; in the short term, we think it still exists as an auxiliary tool for developers.

Empowered by GPT, we believe that companies in the software and service industry are expected to usher in profound industry changes:

► For IT service companies:In the short term, we believe that IT service companies are expected to invest fewer development or implementation personnel to achieve the same quality of delivery, reducing costs and increasing efficiency on the cost side; on the revenue side, since IT service companies usually use a daily quotation model, the overall price of the project may be suppressed to a certain extent when implementing personnel is reduced, but overall, we think that the benefits of AI improving development efficiency outweigh the disadvantages for IT service companies as a whole. In the long run, we believe that the continued decline in developers is expected to usher in a continuous decline in development costs and a continuous increase in gross margin, and IT service companies are expected to usher in changes in business models.

► For product-oriented companies:Cost reduction is second, and development productivity is the core. The iterative speed of software products is one of the core competencies of product-based software companies. We believe that with GPT, the iterative development speed of software products is expected to reach the next level. For software companies, it is expected that the speed of response to customer needs and feedback and the speed of application of new technology will be further improved; for customers, rapidly iterating products are also expected to help improve the user experience.

edit/lambor