Introduction#
Recently, OpenAI launched the ChatGPT plugins system. Many products developed based on the OpenAI API were directly affected. The ChatGPT-based product I wanted to develop was also impacted by OpenAI (fortunately, I hadn't started development yet 🤣). This seems to be OpenAI's App Store moment, providing an official paradigm for the future development of LLM. However, I pessimistically believe that, in terms of ChatGPT plugins, there are not many opportunities for small development teams. Here is my discussion on the reasons for this viewpoint and the possible directions for LLM product development for small development teams.
Limitations of ChatGPT Plugin Development#
Taking the App Store as an example, the thriving development ecosystem of the App Store is based on the fact that the iOS platform itself has only basic universal functionality and a highly interactive hardware system. Users have different preferences and needs, and the hardware of the iPhone is not fully utilized. When these two factors are combined, they provide the conditions for major platforms and developers to solve and create demands, resulting in a vast number of client applications being developed. However, ChatGPT Plugin does not have such characteristics. Because LLM itself can provide personalized services, I also believe that it is the most personalized computing platform available to users. The interaction between users and LLM is based on multimodal information flow such as text, graphics, voice, and video, and LLM outputs target content based on this information flow. This process does not require various buttons, options, and layouts, only an input box is needed. Users can freely convey demand information in various languages and ways. This personalized interactive experience cannot be compared to the common components and design systems of today's UI.
In this process, the Plugin cannot influence the interaction method, but can only shape the capabilities in specific domains that the basic LLM model is not good at by influencing the output and input content. The functionality of the Plugin is compressed into part of the LLM output as a black box, greatly weakening users' understanding of the product. This also results in the development of the Plugin being result-oriented, and result-oriented development leads to data-oriented development. Applications with a large amount of accumulated data and business capabilities can produce better results, while applications from small and medium-sized development teams will be at a disadvantage due to a lack of data accumulation. Moreover, when LLM's training data accumulates to a certain quantity and its capabilities evolve to a certain level, most of the functionality of Plugins without data support can be completed by LLM itself. Therefore, I believe that the Plugin is only a transitional phase in the development process, and LLM will definitely develop into a giant platform, but the giant products within this platform will only be LLM itself.
Above the Platform#
LLM is already very powerful, but if I were to compare it to an operating system, it is currently in the DOS stage. Comparing Prompt Engineer now to programmers back then, one group of people is learning to use the Prompt, while another group is memorizing command line instructions... Although the interaction method of the Prompt has greatly reduced the barrier to entry, it is still quite difficult to maximize the use of LLM. Perhaps we have overestimated our language abilities. Clearly, accurately, and completely conveying our needs is not an easy task for most people.
Therefore, just as the command line interface evolved into a graphical interface, LLM will also have its own "graphical interface". I believe that the multimodal capabilities of LLM are the first step towards achieving this "graphical interface", but it is definitely not enough. Just like Win32 to Windows, and the App Store to iOS, the core is to build an application ecosystem utilizing various capabilities of LLM. These applications are not part of the LLM platform itself, but utilize the "APIs" provided by LLM to expand its interaction with the outside world. These applications rely on LLM but serve as an abstraction layer for users to interact with it, building a development ecosystem specific to LLM.
Conclusion#
As an amateur developer, what can I do in the AI era? My recent experience using ChatGPT to assist in development has shown me that its capabilities are infinite. However, the existing problem, as mentioned above, is that many people, including myself, are unable to fully utilize its capabilities or require additional effort in writing Prompts. This problem tells me that the assistance function of the Prompt is definitely in high demand. I will soon attempt to create a product that meets this need.