Alicia · Blog

A gift to myself

2026-10-03T00:00:00.000Z

Live

The website is live today, on the first day of being 20, after being refined day and night. It is a birthday gift to myself.

What's here

A homepage, a few blog posts, and a notebook that has just started to grow.

The homepage displays the basic information of myself.
The blog holds complete articles across 2 sections - learning & life.
The notes form a knowledge base, grouped by topic, with full-text search.

The technical bits

The website is constructed using React, TypeScript and Vite. Formulas are rendered with KaTeX, diagrams with Mermaid, and code with highlight.js. Markdown compilation, math and diagram rendering, full-text search, RSS — the content pipeline is watched over by 79 automated tests, and deployed automatically on every push. Article dates come from Git history, and every article gets its own share card.

What isn't finished

The content is the biggest unfinished part: three posts(except this one), one note. The plan is to keep writing the notes, and to work through the list I made in the eve-of post, one item at a time.

Besides, currently the website is a static one: every page is built ahead of time and served as plain files from GitHub Pages. Later, I plan to rent a server, add a backend, and turn it into an interactive, dynamic website.

Welcome

Come take a look, exchange a friend link if you like, or subscribe to the RSS.

See you in the comments, or in my inbox!

写在20岁前

2026-10-02T00:00:00.000Z

现在是10月2日，早上的0点11，星期五。...(以下略去原文)¹

intro

虽然最近确实常常低落且十分焦虑，不过作为20岁的前一天的一篇文章，或许还是应当写得积极一点，不是嘛。或许这样可能被认为是报喜不报忧？嗯嗯，说的很对，确实就是这样。

summary of the past

作为第一篇或许可以称得上是年度总结的文章(?)，这里或许应当贴上这样那样的图片，展示过往取得的种种成就，以及多姿多彩的生活；不过很遗憾，笔者在前18年里只是一个名副其实的做题区，在最近两年里，once again，也只是一个名不符其实的做题区(笑)。丰富的社会生活没有，多样的才艺没有，似乎在这么多年里，唯一能拿得出手的也就只有学历了。上了大学倒是加了一些社团，也有了一些很好的朋友，不过因为近期的学业上的压力，啊，这个暂且不提，也没有很多的精力去参加这样那样的活动了；这也导致了申请奖学金的时候，也只敢去申请学业相关的类别了。嗯嗯，毕竟申请的材料上也只能填上成绩这一项了嘛，0社会参与，以及暂时还没有什么结果的科研，不也就只能这样了嘛；就像以后出去找工作，简历上除了学历一栏，其他的全都是一片空白一样。

这么说来，或许截至目前，最大的成就就是上了清华大学？虽然不是很想承认这一点而且为此感到很伤心，但是目前来看，结论似乎好像就是这个样子的啦，www。欸，但是为什么在这里没有找到录取通知书的照片，而不是像很多其他校友的朋友圈里的那样呢？那就是另一个故事了，或许在之后会补上吧。

哎哎，好像说了很多丧气话呢，抱歉啦。虽然和其他的大佬们比起来，经历还是显得非常匮乏，但是在过去的一年里，还是有一些小小的闪光点来着，比如第一次见到课本里的人物，cpp程序设计语言的创造者，Bjarne Stroustrup，甚至还要到了签名以及合影，又比如，在上半年第一次尝试SRT，以及随后在暑假末以及本学期正式参与课题组的研究项目，并拿到了第一份工资(如果这个可以算作工资的话)。虽然第二个也是最近压力以及焦虑感的主要来源之一，笑。

plan of the future

好像暂时找不到别的什么值得纪念的东西了，或许和笔者不怎么喜欢记录生活有关系吧，就连这篇文章也是一时兴起才写的。好了，那么转入下一环节，未来的目标(虽然目前为止立下来的"远大志向"好像向来实现情况堪忧，但总得有那么一个目标吧，嗯嗯)。

首先第一条自然得留给明年的保研了，希望能保上目标的方向。
希望今年能在科研上投入更多的时间，也希望能做出一些有意义的成果。
希望26秋季学期的成绩不要太烂，因为真的选了太多课了www，但这真的不是我的错!(好吧，确实有一部分是个人原因啦)

outro

谢谢你愿意看到这里。我的文笔很烂，每次想写些什么，写出来的东西总是不尽人意(好像英语写作要稍好一点)；如果你认为这篇文章逻辑不通结构混乱，原谅我吧。

好的，暂时就先用这篇不尽人意的文章，为年龄还是1开头的最后一天，以及不尽人意的前19年，收一个尾吧。

Footnotes

改编自热带风味冰红茶

↩

Moments in life

2026-09-24T00:00:00.000Z

December 11, 2025

A moment to remember.

The signed book(The cpp programming language).

The signed notebook.

Added on Sep 27th, 2026.

September 23, 2026

The 1st paycheck I've ever earned.

Added on Oct 2nd, 2026.

Some thoughts on the use of Artificial Intelligence

2026-09-23T00:00:00.000Z

Introduction

A few years ago, when generative AI just came into most people's view, I encountered a Chinese reading comprehension passage which claimed that while AI could imitate literary styles, it could never produce writing of the same caliber — because it lacked the lived experience of the original author, Li Bai being the example given. The conclusion was unequivocal: AI would never replace human writers.

As the time goes by, the assertion that passage made became less and less convincing, however. With the iteration of models, ai is getting more and more competent in writing passages. Does the so called "lived experience" really matter? If something looks like a duck, acts like a duck, then it is probably a duck. The same law may apply to writing as well.

And if AI can write, what else can it do? That question led me somewhere I didn't expect — into my own experience of learning and doing research, and into an uncomfortable realization: the parts we're losing may be the parts that matter most.

Main text

This is the first post of my blog, and the second of the whole website. What I want to discuss here is the phenomenon of using Artificial Intelligence in various areas, to be more precisely, in learning and scientific research.

The origin

The first AI agent I've ever known is ChatGPT. That was back in 2023, when I was a high school student. As the most powerful AI back then, though it showed impressive competence in dialogue, its limitations were also apparent. That is, AI hallucination, a problem that has yet to be solved. It made up things and events that actually didn't exist. And it got common sense wrong. And when it came to mathematical problems, it seldom gave the correct answer. I only considered it as a toy at that time, and I thought it would take at least a decade before it became capable enough to replace humans. Then I quickly forgot about it, as the college entrance examination was coming near.(And as a student of a senior school in a small county town, I had limited access to AI, both in time and device.) I had no idea how fast things would change.

Phase 0

A year went by, and I was in university. My first semester felt no different from before. I took notes in class, did my homework on my own, and looked things up online when I needed to learn more. Back then, I only used AI in liberal arts courses — writing and communication, and ideological & political education, to be more exact, and for information searching. I hate writing about things I don't like, even now. That's also why I hated Chinese class in high school. Honestly speaking, AI helped me a lot in this aspect, freeing me from spending hours and days producing rubbish painfully.

Phase 1

Things changed a little bit, at the end of that semester. We were required to complete a project which was quite challenging for us at that time - to make a Tetris game using the given framework. All that we learned in class was only the basics of Classic C syntax and some fundamental data types. So, we had to learn more stuff ourselves, including rendering, event handling, to name but a few. Firstly I asked AI what I needed to learn to complete this job and where I could find related learning articles, then I checked whether the information it gave was right and learned it, implemented it for the first month starting from when the work was issued. Then I happened to come across Github Copilot, a coding agent made by microsoft. That's the very first AI I've ever used in programming. It could handle small projects quite well, but when the codebase got larger, it started losing track of how different parts fitted together and making up functions that did not exist, because models at that time did not have such a long context window. So I only used it for hints, for info searching, or for generating templates that I would adjust according to the project myself. Anyway, it did help me a lot, saving me plenty of time looking up related information and docs online, and thus accelerating the development of the project. And I earned an A+, the highest rank, for that course in the end. By the way, I have got A+ in only two classes till now(2026-09-23). At that time, I began to consider ai as a helpful assistant, but only for work that did not require much reasoning. After all, the L in LLM is for Language, rather than Logic.

Then at the start of 2025, DeepSeek-R1(DS for short in the following passages) came out. I quickly fell in love with it, especially in handling some assignments that I did not like(Outline of Modern Chinese History, for example). Firstly I downloaded the ppt of the class online, then uploaded it to ds, copied the requirements of the work and pasted it in the dialogue box, clicked "send" and I got the original article. After about 1h's polishing, all the work was accomplished. All of these happened only in 2h.(Now I even do not need to review the article myself, so it is faster, costing about 20 min.) As for coding, I hoped I could learn something, so I disabled the copilot extension in vscode and did the work myself. I still remember the days I spent hours debugging the program back then - setting breakpoints, constructing test cases, stepping through the code line by line, which has become a luxury today. I only used ai when necessary: to get some hint after hours of trying, to read compiler or runtime errors to help me find bugs, and, when all else failed, to upload the source code and ask it to find the problem.

This lasted until the first few months of my sophomore spring semester. Though the ai tools I used changed several times(Google Gemini for the most of the time. Gemini has been my first choice till April, 2026. Then I switched to ChatGPT/Codex with the suggestions of some friends, and I have been using it ever since then.), the workflow remained the same: I did the work myself for the courses that I considered important, and only used ai for assistance. And for courses that I thought was useless but required, I simply relied on ai to quickly finish it so as to avoid wasting too much time. In 2025, despite the huge strides AI has made in programming, it still struggles with mathematics — often falling into circular reasoning or producing false proofs. So I never thought of the possibility that AI could be used to do mathematical homework, which has become a reality now.

Phase 2

By the end of April, 2026, that approach was getting more and more unsustainable. That was because of the Software Engineering course. The main goal of the course was to cooperate with teammates to create an instant message system. And I together with another classmate was in charge of the front-end part. The language used was TypeScript and the framework was React(The same used by this website.). Before that course, the only web development experience I 'd had was writing some basic html sites. And the course DID NOT TAUGHT THOSE THINGS, IT ONLY TAUGHT THE CLASSIC DESIGN ARTS IN THE SOFTWARE ENGINEERING, which meant that you had to do that yourself. BUT I DID NOT HAVE ENOUGH TIME TO LEARN SO MANY THINGS, starting from html, to css, then to javascript/typescript, and finally, to React. I came to rely more and more heavily on the use of ai, and at last, almost all the codes were generated by ai, and what I did in this process was to understand the code written by ai and to make some suggestions for further changes, which made me feel disturbed. After all, it was I that took the class and learned the knowledge, not the ai. I can't in good conscience treat ai's capabilities as my own. Later on I learned that I was not alone - the whole class had turned into a competition of ai, rather than a competition of students' own competences. And the same phenomenon also occur in many other scientific competitions. That's quite natural, since if you do not use ai, you will be surpassed by others easily. But as for me, that is no different from doping in sports. I do not think it a good manner, but I had to do it, for I did not had enough time to learn it from the very start, and if you did not use it, you would fall behind others soon.

Phase 3

Things got even worse. This summer term, I chose a research-oriented course, in the hope of getting an experience of doing research. The subject of the project was exactly Large Language Models. And it lasted 2 weeks. THE FIRST TWO LESSONS OF THE COURSE TAUGHT THE BASIC SYNTAX OF PYTHON. All the knowledge, including RNN, CNN, and Transformer was taught only in less than 1 week. Then we were required to come up with an idea and complete the project in the remaining 1 week. Believe it or not, the teacher thought that IN JUST TWO WEEKS, we could go from COMPLETE PYTHON BEGINNERS to RESEARCHERS CAPABLE OF INDEPENDENTLY CONDUCTING AI RESEARCH. Sometimes I will think, the current curricular system is ENCOURAGING THE RELIANCE ON AI.

Actually the whole work loop ends up looking like this:

View this diagram on the website.

The ONLY thing we do in this process is the verification work. Or a little more, acting as the bridge between AI and the experiment device(A server), if you can call that work. The most important and demanding parts in the field of scientific research are simply missing: coming up with innovative ideas and figuring out how to implement them. And even if it does produce results, I would not consider that mine. This is no different from plagiarism, in my opinion.

Conclusion

I am not against AI. I use it every day, and I will keep using it. But there is a difference between using a tool and letting it use you. When I let AI write my code, I did not learn to write code. When I let it generate my research ideas, I did not learn to think. The output may look like mine, but the competence behind it is not. And if that competence is never developed, it will not be there when I actually need it.

And I do hope, in a time when everyone is impatient, we can still find the patience to learn things properly — to struggle with a problem, to sit with not knowing, to earn the understanding we claim to have.

Revision

2026-09-23: fix some typos & add some supplementary information.