My musings on how traditional medicine utilizes artificial intelligence

Haha, lately it’s been too hectic, with seeing patients, memorizing classical texts, organizing materials, and other tasks. I’ve been wanting to find some time to write an article, but haven’t been able to until now. Today, I’ll use this space to summarize some of my recent thoughts and plans.

The public account “Random Thoughts” will try to maintain daily updates, serving as a unique form of diary. As for the “Case Analysis” series, I’ll pause it for now. I’ll first focus on completing some study notes quickly using tools like Deepseek, including various experts’ interpretations of the clauses in Treatise on Cold Damage and Synopsis of Prescriptions of the Golden Chamber, as well as their understanding and application of formula patterns and herbal indications. I’ll also create some visual content like mind maps and tables. After completing these more labor-intensive tasks, I’ll write a batch of case analyses all at once, which will be a process of concentratedly strengthening and optimizing my diagnostic and differentiation thinking and details. Hopefully, these artificial intelligence tools can greatly reduce my workload.

1. Initially, when I heard about DeepSeek’s immense popularity, I was dismissive. I thought, how could something as unintelligent as artificial intelligence replace the human brain or traditional medicine practitioners? It seemed impossible. It wasn’t until a doctor friend, who is both a mentor and a friend, shared their positive assessment that I started to slowly try using it.

2.As I gradually got used to using DS, I found it incredibly powerful. It could find any information I wanted in a very short time, even original classical texts, and could also organize and summarize them. During this stage, I was amazed by its magic and efficiency, and even fantasized that it could greatly shorten my in-depth study and advancement in the field of traditional medicine. That was until recently when I discovered about a dozen instances of fabricated content in one article.

3.At this point, I seemed to develop a deep suspicion of DS again, wondering what it was actually useful for. It didn’t seem to be helping me. I still had to verify the authenticity of the literature it found one by one, which felt like it would have been better to just look it up in books myself from the beginning. Haha, I became very distrustful of it. It was from here that I rekindled the idea of memorizing classical texts, planning to thoroughly memorize the Guilin ancient edition of Treatise on Cold Damage, and firmly decided to seriously read books and diligently take notes.

4.After a period of hesitation, I tried to observe, analyze, and understand DS’s thought process: what tasks it can do, what it cannot, and what its strengths and weaknesses are. At the same time, I also tried to categorize my own work and study content. I slowly began to discover where DS could be helpful.

DS’s working principle and approach are roughly as follows:

It has a large inference and analysis model, which is good at some simple logical reasoning (this is similar to human thinking patterns and habits). Then, it has an initial database, which I suspect only has a small portion open-sourced, given that it’s free (this is similar to human common sense).

Next, it searches relevant web pages and organizes and categorizes the viewpoints on these pages (this is equivalent to humans learning the theoretical principles of a new subject).

Finally, it uses its own thinking patterns (the designed model) to analyze and reason about our questions and tasks based on the initial database and the collected viewpoints.”

First, it can only handle simple reasoning; it can’t manage anything too complex.

Second, it assumes that its database and the viewpoints it gathers from web pages are the truth and correct, rather than discerning the validity of this content.

Third, its analysis process often over-optimizes, taking a simple question and making it overly complicated and cumbersome, or conversely, summarizing very detailed content into a single sentence.

Fourth, it thinks and answers questions based on the user’s identity and thinking habits; it tailors its responses to the individual, like this example below: