In the summer of 2023, when I was filling out my college applications, I did not choose a single major related to artificial intelligence.
I am now studying five-year clinical medicine at a medical school in Henan. I enrolled in 2023, and this semester I am entering my fourth year. Before I enrolled, my entire relationship with AI was basically just having heard the word in the news.
If someone had told me back then that three years later I would spend every day immersed in Claude Code, OpenClaw, and Codex, that I would go into companies to implement AI, and that I would stand on stage explaining how these systems work, I probably would not have believed it.
But that is how things turned out.
In late 2025, short-video platforms and Bilibili (a Chinese video platform) were suddenly flooded with Gemini 3. Google officially released Gemini 3 on 11 · November 18, and review videos came out one after another. Everyone was saying it was the strongest model at the time.
I had only one thought: how strong could the strongest really be?
After a fair amount of effort, I finally got to use it in the official Gemini app. After trying it, honestly, I was a little shaken. Not because it gave me some magical answer, but because you could clearly feel that this thing could already understand complex problems and do quite a lot of work that used to take a long time.
For the first time, I felt a very concrete anxiety: if I did not understand it at all, I might really be at a disadvantage a few years later.
So I started looking for AI-related content all over short-video platforms. Scrolling through them, I came across WaytoAGI. It is an open-source knowledge base and learning community for Chinese-speaking AI learners. People there organize courses and case studies, and some also organize offline meetups in different cities.
At that time, I did not have any real learning path. I just spent every day immersed in it, watching how other people used AI to get things done.
In January 2026, I attended an offline WaytoAGI event in Zhengzhou. I met someone there. It was only a brief encounter, and at the time I did not think anything would come of it.
I only later realized that this person became my guide for the next half year.
After going home for winter break, I kept following the community content and came across Claude Code. At the time, I connected GLM-4.7 to Claude Code. Simply put, you express what you need in natural language, and it reads the project, changes files, runs code, and hands the result back to you for verification.
Before that, I mostly used DeepSeek and Doubao (ByteDance's AI assistant) day to day. They could answer medical questions, but I quickly ran into a problem.
Medical textbooks have editions.
Ask about a certain topic, and it might answer according to an old edition, with a very confident tone. Yet diagnostic criteria, classification schemes, and drug descriptions can change between editions. For a medical student, the most troublesome thing is not that it does not know, but that you cannot tell which edition it is drawing from.
So I wondered: could I put in the PDF of the textbook I was using and make it answer only according to the edition I specified?
That became the first thing I made, called Ling Lan. It could answer questions based on the textbook content and could also generate mind maps and review cards. The web page and retrieval flow were mainly AI-assisted; I was responsible for supplying the textbooks, designing the study flow, and repeatedly testing whether it drifted off.
In the end, Ling Lan was never really adopted for long-term study. It was just a runnable demo, still far from a mature product. I later wrote a separate retrospective about that detour.
But it mattered a lot to me.
It confirmed for the first time that even without knowing how to code, I really could use AI-assisted development to turn what was in my head into something real. But making something real is still a long way from getting anyone to actually use it.
That same winter break, Clawdbot started appearing constantly in posts from the AI bloggers I followed. It later went through the name Moltbot before becoming today's OpenClaw, a history that is also documented in the project's official description.
At that time, many people were writing skills and plugins for it. As a medical student, I had always been interested in how the human brain remembers and forgets, so I followed my own understanding and made it a memory system that stored information in layers, let memories decay over time, and periodically consolidated them.
The code was likewise AI-assisted. I was responsible for deciding the rules, how memory should be layered, what should be forgotten, what should stay, and the final testing and revision.
After finishing it, I wrote a WeChat Official Account article and shared it in the local WaytoAGI community group.
Soon someone contacted me and said it was pretty well done and the thinking was quite ahead of its time.
It was the same person I had met at the offline event in January.
During that period I was still working part-time as a teaching assistant at a tutoring center. Model subscriptions and API calls all cost money, and I needed income to cover those expenses. At first I agonized over even a single coding plan. Later, I spent more and more on AI, but I never again felt the same shock as the first time I had Claude Code help me get something running.
After seeing that this person had taken the initiative to contact me, I asked directly: do you have any AI-related work over there?
The reasoning was simple. If I could learn while doing real work in a real environment, it would certainly get me closer to where I wanted to go than continuing as a teaching assistant.
I even told him that as long as he was willing to take me, I could come over right away.
Later we met and talked for a whole evening. The next day, I got a part-time AI implementation job that paid by the day. I quit the tutoring assistant job and reported directly.
In February, I started doing AI application implementation at a medical aesthetics chain.
Some people might wonder: what can a medical student who does not know how to code actually do in a company?
I was nervous about that too at first.
Once I actually got into the work, I realized that problems in a company usually do not appear in technical language. People on the ground just tell you that customer messages cannot be answered fast enough, the data does not match, spreadsheets have to be copied manually every day, or leads from a certain platform always enter the system much too late.
You have to first understand where they are actually stuck, then break the problem down into processes, data, interfaces, and acceptance criteria. AI-assisted programming filled in the part I could not implement on my own; understanding requirements, designing processes, testing, troubleshooting, and iterating became what I was actually responsible for.
But a real company and Ling Lan were completely different worlds.
If Ling Lan broke, I could just delete it and start over. Systems in a company are connected to real customers and colleagues' work; if you change one part, someone else might have no data to use the next day. When I first joined, I was often anxious, and when I went home at night I would keep replaying whether what I had changed that day might cause problems.
In March, discussion of OpenClaw suddenly picked up. I was among the people who had made memory plugins for it relatively early, so people started inviting me to do offline talks.
When I first stood on the stage at a local AI industry event, my palms kept sweating. I knew very clearly that there might be quite a few people in the audience who had actually written code for many years, while the systems I was talking about were largely AI-assisted, with their main implementation code generated by AI.
I struggled with this for a while.
Later I gradually made peace with it. Being able to write code directly is one kind of ability; being able to break down business requirements clearly, judge whether what AI generates is usable, and know where to look when something goes wrong is another kind of ability. The two cannot be conflated, and I do not need to package myself as a traditional programmer.
Starting that month, I developed a habit. Every time I made a technical choice, I noted why I chose it at the time. Every time I messed up, I also noted exactly where the problem was. Later, as these notes accumulated, I turned them into the Pitfalls and Reflections Wall on this site.
April was my busiest stretch.
The offline talks were concentrated in those weeks; I gave several in a row. I also received a certificate appointing me as a training mentor in the AI Implementation in Practice track at a local industrial park. Several local media outlets covered these activities, and the public materials are all on the media page of my personal site.
From the outside, it looked like I suddenly had quite a few results.
The actual reality was: classes during the day, projects at night, going to bed at 3 or 4 a.m., sometimes lying down when the sky was already getting light, and then having to be in class only a few hours later. That rhythm lasted for a while, and I was basically running on fumes.
Also in April, I ran into an incident I still cannot forget.
A data synchronization job that ran automatically every day had been failing silently many times. There were no alerts; on the surface everything looked normal. For several weeks, data did not flow downstream as expected, and I only found out it had stopped long before when the business team came to ask.
That really stunned me.
When I made demos myself, I used to feel done once they could run. For a system that real people are actually using, you have to keep asking: if it breaks, how will I know?
Later I spent time adding monitoring, alerting, success heartbeats, and data read-back to the system. I have also written a separate complete technical post-mortem for that incident. Since then, whenever I evaluate any automated workflow, I first ask whether anyone would notice quickly enough if it failed.
In mid-to-late May, as final exams approached, I basically returned to school. The company part-time work became occasional visits, and the rest of my time went entirely to studying and exam preparation.
Clinical medicine is my main work. Failing a course is not an option for me.
During that period, what worried me most was not actually the exams, but whether I could pick things back up after pausing. The systems were still running, but I could not keep watching them. What if something went wrong?
Fortunately, the monitoring I had been forced to add after that April incident really paid off now. When minor issues came up later, the alert usually went off first, and then I handled them remotely. The fall I had taken earlier ended up protecting the later version of me.
In July, summer vacation began. With no classes and less commuting, I could finally focus on one thing.
My output that month was much higher than before. The personal site went live, the customer service system kept iterating, and I also produced a batch of enterprise automation workflows. Compared with the start of the year, model capabilities had also improved a lot. I spent less time going back and forth with models, and many ideas landed much faster.
Also in July, I looked back at the past half year in full for the first time, and only then realized that I had gone from someone who just watched reviews in short videos to someone working inside real systems.
August, which is now.
I have already ended my part-time job at that company.
Not because I am no longer doing AI. On the contrary, my interest in AI has not changed. I left only because the environment and the direction I want to take next began to diverge. The projects there did solve real problems and are still running, but further in, I found it increasingly hard to say where I was heading in the future.
I paused to think this question through again.
By now, I have accumulated thousands of collaboration sessions with coding agents and have worked on more than a dozen projects large and small. I have debugged at 4 a.m. and stood on stages I would never have dared imagine before.
But I still cannot detach from AI and write a system of the same scale entirely on my own.
What I do is think through the requirements, break the process down correctly, supervise the AI as it completes the implementation, test, give feedback, accept the result, and then judge which layer to investigate when something goes wrong. This ability is not the same as traditional programming, and I do not intend to describe them as the same thing.
Even more importantly, I have not switched careers.
Medicine is still my main work. AI has just given me an extra ability to turn ideas into real things. During the past half year, there was also a stretch when almost all of my time went to exam review. The time spent doing AI at high intensity was not as long as it might seem; things just came in dense bursts.
If you are also in medicine or another non-computer-science major and want to try making something with AI, my own starting point was actually very clumsy. I did not first finish learning a whole programming curriculum, and I did not first study AI principles in depth. I just seized on a real problem I had encountered myself, confusing textbook editions, and tried to solve it.
In the end, Ling Lan did not become a product.
But without that failed demo, the memory system, enterprise projects, offline talks, and this personal site probably would not have happened.
When I filled out my college application, I did not choose any major related to artificial intelligence. Three years later, I have not left medicine either.
It is just that over this past half year, AI opened an extra path for me.
If you have read this far, maybe you are also wondering whether to start. Where are you now? Or is there a particular month or project you want me to break down in more detail? Leave a comment and tell me, and I will pick one out and write a separate retrospective.
The personal experiences and timeline in this article are my own true record. Company and school information has been anonymized according to public boundaries, and the role of AI-generated code is described as it actually happened.