Today, I finally found a little free time and pushed myself to start doing some of the things I have wanted to do for a long time, but never got around to for one reason or another.
As the saying goes, the hardest part is getting started. I began writing with the same attitude I bring to writing papers: “So what if what I write today is a complete piece of shit?” To start with, I want to think about why I am starting a blog. If I understand why I want to do something, it will feel more meaningful and be easier to keep doing.
Recording experiences and feelings
The first reason is that I want to record some of my experiences, along with the thoughts and feelings I have at particular moments.
Ever since I was in school, I have thought I was not particularly good with words. As a child, I was terrified of writing essays. My mind would go completely blank; I would rack my brain and manage to squeeze out only a few sentences. Then I would be rewarded with a trio of comments from my Chinese teacher: “empty,” “bland as plain water,” and “all fluff.” To be honest, as I remember it, I did not really understand what those comments meant until around middle school, close to high school. The teacher would pass judgment without even explaining what those words meant.
But gradually, as I gained more life experience, I started to find that I could sometimes write things that both I and other people considered fairly good. My fear of writing slowly went away, and I stopped resisting it. I also think that, compared with people my age, I have had quite a few unusual experiences. They are precious to me, and I do not want those experiences and the feelings that came with them to fade with time. So I want to give myself a reason to occasionally write down what I am thinking.
Building up my research and technical understanding
The second reason is that I want to study and write about research and technical topics in depth. Scientific research takes time (see Slow Science Org). I believe the development of science is closely tied to the personal growth of researchers. A researcher first has a fairly basic understanding of a field or topic. With time and continued reflection, they may gradually reach the frontier of a particular direction. Then, after countless experiments and failures at that frontier, they may arrive at insights and understanding of their own.
I know we are now in an era when LLMs can replace a large part of intellectual labor, and the community tends toward restlessness and short-term gains. I accept this reality to some extent, and I do not think it is necessarily a bad thing. I just feel that I am still in the process of learning and building up my knowledge. Compared with the task-driven learning I started with, I now need to deepen my understanding of what I know and think from the perspective of a researcher.
In this respect, I have been greatly influenced by Scientific Spaces by Su Jilin, a leading voice in our field. When I first started learning machine learning, I read many of his articles and discovered that mathematics did not have to be boring. The same formulas and principles could be understood from other perspectives. That was a huge inspiration to my younger self, and it sparked my interest. I hope I can patiently write down and develop my understanding of technical topics, perhaps starting with model compression, which I know well.
Another thing is that top AI conferences are now widely facing DDoS-like submission attacks, while autoresearch is producing more and more content. I hope we can slowly return to the core of research questions and offer genuinely thoughtful, high-quality ideas. (The anonymous comments feature is still being developed. Please bear with me; even with AI assistance, this is not a simple task for me.)
Leaving behind original human writing
The third reason is that, in an age when generative AI can pass the Turing test with ease, I want to keep creating original human work.
The current LLM era is exciting. Even small open-source models have absorbed a huge amount of the world’s text, and many kinds of human work can now be done through generation. So why should we continue to produce human writing? Current LLMs are trained using maximum likelihood estimation. When generating text, a model chooses the most probable next token, so the passage it produces is the most probable continuation under a given condition. In other words, what it produces is consensus.
Yet, as we know, all of humanity’s groundbreaking scientific discoveries have been statistical outliers: Copernicus’s heliocentric model, Einstein’s theory of relativity, and even Turing’s Turing machine. They challenged the consensus of their time. For now, only humans can keep pushing the boundaries of what we know. On the other hand, the development of LLMs has freed us from a great deal of labor. We should use that extra time to focus on thinking itself, so that we do not gradually become slaves to AI.
(This post was originally written in Chinese. I will use AI tools to translate it into English. Following the principle of getting started first, I want to lower the barrier to beginning this project. In the future, I may write blog posts directly in English.)
Why put the blog on my personal website?
Why host the blog on my personal website instead of on knowledge-sharing platforms such as Zhihu or Xiaohongshu? I have a few reasons.
First, it gives me a sense of ownership. On my own website, I can decide where blog posts appear and present them however I like. That gives me a little freedom to do things my own way.
Second, I want to face as little content moderation as possible. My blog will focus on myself and my research, but publishing on a centralized platform means being subject to its content moderation rules. That conflicts with the spirit of free expression I value. Whether there is censorship and whether my content passes that censorship are two different matters. Of course, GitHub, where my personal website is hosted, is not perfect in this respect either, but it is the best among the available options.
Third, platforms recommend content to users, which can shift my attention. When I post on those platforms, I may find myself paying attention to likes and saves, even though they have nothing to do with the content itself and bring me no real benefit.
There is one more reason, much like the time scientific research takes: I want to spend a long time polishing a high-quality piece, instead of casually posting short, fragmented content (like short videos and interview clips).
I think this is enough for a start. Stay tuned.
Shengkun Tang
, Nantou Ancient Town, Shenzhen
Comments
Comments are not open yet.