Computing Life

This article discusses the significant transformation GPT has brought to the field of knowledge management. First, let's understand what knowledge management systems are. I have a personal interest in productivity and knowledge management and have conducted many experiments in this field. Typically, knowledge management systems help us accomplish three tasks: 1. Acquire knowledge; 2. Store and analyze knowledge; 3. Access and serve knowledge.

Acquiring knowledge usually involves recording our thoughts in the form of blog posts or notes, ensuring that they exist not only in our minds but also on mediums such as paper, computers, or Internet. However, isolated knowledge points are not enough; the core of knowledge management lies in managing these points. A good knowledge base should establish connections between isolated knowledge points, forming a network. This way, when a certain piece of knowledge is needed, we can quickly find past experiences and lessons to apply to other fields or share with others, building a knowledge system.

GPT has brought profound changes to every aspect of this system. The first is knowledge acquisition. As I mentioned in my previous article, I created a note-taking tool that combines voice recognition and GPT organization. After using this tool, I found its importance and convenience far exceeded my expectations. For instance, during meetings, if no one takes notes, the knowledge of the participants will dissipate quickly. Therefore, having a painless, automatic, dedicated secretary to record the key points of discussions and convert them into documents is crucial for knowledge serialization or persistence. This significantly lowers the barriers to knowledge acquisition, increasing the total amount of knowledge and addressing the first challenge in building a knowledge system.

Additionally, GPT not only has recording capabilities but also has the ability to summarize and condense. I added a new feature to the tool that logs all conversations with GPT, and then runs a small program every day, asking GPT to read the day's articles, classify them by topic, and expand on the key points within each category. In this way, my thoughts are systematically summarized daily. This cross-temporal summary is key to building a knowledge network, and GPT's text comprehension and a certain degree of logical capabilities have already endowed it with such ability.

The third aspect is knowledge access and serving. Traditional tools often rely on manual link building or tools like OneNote and Notion, which depend on nested indexing structures or nested folder structures similar to Yahoo. However, in this era, approaches like Gmail's may be smarter. It no longer relies on manually building knowledge connections but instead relies on the content of knowledge itself. By leveraging text similarity, search, and semantic understanding, it dynamically constructs a small personalized knowledge network when users ask questions and presents it to the user. Although I have not yet built this part of the product, I believe it should be easily achievable through search, understanding, and GPT's API.

In a sense, this could become a critical factor in dismantling existing knowledge management systems. Knowledge management no longer requires manual entry, manual organization, or manual link building but instead relies on dictation and dynamically building connections when users query. This demonstrates two core changes GPT has brought to knowledge management systems: lowering the barriers and raising the ceiling. Lowering the barriers means that taking notes with GPT is very easy; we no longer need to open our phones, open an app, and start typing as before, but can simply speak to the phone. This change is crucial; for example, I am currently exercising while dictating this article. Raising the ceiling means that GPT, combined with a computer's powerful storage capacity, can delve deeper into data, providing previously overlooked inspiration, discovering unknown unknowns, and thus increasing the utilization of knowledge.

Furthermore, while building this tool, I found that the user interface is very meaningful. Initially, I used a web app, but I encountered various issues during use and eventually switched to a Telegram bot. This approach has several advantages:

1. In this way, users can view past chat records without adding any additional code.
2. It naturally introduces the user concept, eliminating the need to maintain our own user database and worry about password leaks, defending against attacks, or combating malicious registrations, while achieving natural separation between different users.
3. It adds extra stability. When using the web app, request failures often occurred due to the instability of the OpenAI API. Due to the app's limitations, users had to repeat their previous voice input. However, Telegram comes with the ability to store sent voice files, so we only need to resend the voice file to improve fault tolerance.
4. It transforms the synchronous process into an asynchronous one. For example, in the web app, we cannot close the screen or switch to another program while waiting for the API to return, as this would terminate the browser's operation and prevent the API results from being displayed. However, Telegram has push notification functionality, allowing plugins to run in the background, communicate after the calculations are completed, and provide users with real-time push notifications.

This leads us to contemplate the changes in how users and computers will interact in the future. Although the traditional approach is a graphical user interface (GUI), as models like GPT become increasingly popular, conversational interactions may become the mainstream, at least in these application scenarios. However, designing conversational user interfaces also faces challenges. For example, in our voice recognition tool, if we want to manually add a topic to a piece of speech, we might design two input boxes in a graphical interface: one for the topic and another for the regular voice input. But in a conversational environment, with only one dimension, we need to consider how to design interactions, what to say, and what format to use to make it feel natural for humans while being easily understood by machines.

In conclusion, we may be at the decline of traditional knowledge management systems and the starting point of a new interaction method, with many novel concepts waiting for us to explore. The above content was dictated in just ten minutes while I was exercising and listening to a video by the class representative, and then organized by GPT.