# I Used AI to Write a DeepSeekV4 Paper Review, but I Do Not Think I Should Publish It I just used AI to write a paper review of DeepSeek-V4. The workflow was smooth. I dropped the PDF in, asked AI to extract the content, organize the structure, generate illustration ideas, and then turn everything into a technical blog post that looked fairly complete. The key innovations and the interpretation of evaluation results were both pretty decent. I put the outline below. It looks quite good, doesn't it? ![](https://leafw-blog-pic.oss-cn-hangzhou.aliyuncs.com/screenshot_1777095270.png) But after reading it, I felt uncomfortable. It was a pretty good paper review, yes, but it was not something I had written after truly understanding the paper. More precisely, AI had helped me turn the paper into an article whose general meaning I could understand. There is a difference, and it is not a small one. --- ## I Realized I Have Regressed a Little I have used a lot of AI over the past two years. Writing code, searching for information, organizing documents, reading papers, generating scripts, building small tools. For many things, I can ask AI to produce a first version. The efficiency is real. Something that used to take an afternoon may now produce a usable result in half an hour. The problem is exactly there. I am getting used to letting AI give me the answer first. When answers arrive too quickly, people become lazy. Especially when facing a technically dense paper, my first reaction has become: let AI summarize it first. At first, this is reasonable. Papers are long. Having AI extract the structure saves time. But if I stop at \"extract the structure\" every time, or even treat an AI-generated review as my own understanding, then it becomes a problem. I felt this while reading DeepSeek-V4. I recognized many terms, but I had not really kept up. Muon was the clearest example. I imagine people who follow Kimi or read Su Shen's blog often are already familiar with it, but honestly, I still have not carefully studied how this optimizer works. What am I even doing? My understanding of large model training is still stuck around two years ago. Back then I would still take courses, read books, and run small models on Colab, trying LoRA, QLoRA, PEFT, and similar things. I was not good at it, but at least I ran things myself. How loss decreased, how memory exploded, how batch size affected training, how models hallucinated. Those experiences were real. What about now? I have bought quite a few books about large models, but most of them have only been opened for a few pages. I do not even remember the last time I opened Colab. For many new training techniques and technical changes, I only know the names. As a hobbyist, I cannot realistically touch large model training directly because I do not have the resources. But not being able to understand current papers is still embarrassing. --- ## AI Is Great for Assisting Learning, but It Can Easily Replace Learning I do not think there is anything wrong with using AI to learn. On the contrary, AI is a very good learning tool. It can quickly explain concepts, organize context, identify a paper's structure, and fill in background knowledge. In the past, if I got stuck on a concept while reading a paper, I might need to search through several blogs and textbooks. Now I can ask AI to give me an entry-level explanation first. But there is a line. When AI lowers the barrier to entry, that is assisted learning. When AI generates an article that I myself have not fully digested, it starts to replace learning. The problem with replaced learning is subtle. It does not immediately make you feel weaker. Instead, it creates the illusion that \"I seem to understand this.\" You can ask AI to explain a concept clearly, list the meaning of variables in a formula, and summarize every section of a paper into bullet points. It reads smoothly. Smooth does not mean understood. Real understanding is usually not that smooth. You get stuck on a symbol. You find that definitions do not line up. You realize you have forgotten linear algebra. You reread the method section three times because of one figure. The process is slow and annoying. But that process is learning. --- ## I Am Going to Use AI Differently So I am not going to publish that AI-written DeepSeek-V4 review. There are already many reviews online, including many professional ones. What is the point of publishing one written by AI? I want to put AI back in the position of a tool. First, let AI generate a mind map. I need to understand the paper's high-level structure first. The structure of the generated review was actually useful, and I also used NotebookLM to generate a mind map, which worked well too. Second, focus on core concepts such as CSA, HCA, mhC, and Muon. For these questions, AI can provide initial explanations. They do not need to be long. Ideally, they should give intuition, point to the relevant formula locations, and include background context. Third, return to the paper with those concepts in mind. This is the key step. AI's explanation can only be preview material. Real understanding still has to return to the original paper, especially the method and experiment sections. Symbols, tables, figures, ablations, and system assumptions cannot be understood only through second-hand summaries. I used to read papers from beginning to end. Now I realize that may not be the most efficient approach. For someone like me who has fallen behind for a while, it may be better to build a map first, then read with questions. --- ## Using AI to Implement an Idea Does Not Mean You Own the Ability I have been feeling this more and more recently. Being able to use AI is indeed a skill. Writing prompts, breaking down tasks, letting agents run code, using skills, and building automation workflows all have value. But this kind of ability is easy to overestimate. Knowing how to use skills does not mean you understand the underlying technology. Installing a local AI client does not make you an AI engineering master. Getting AI to write a DeepSeek-V4 review does not mean you truly understand DeepSeek-V4. I am not denying tools. Tools should be used, and used well. But \"I can make this with a tool\" cannot be directly converted into \"this is my ability.\" There is a layer in between: my own understanding, judgment, and practice. Without that layer, the stronger the tool, the stronger the illusion. AI has developed too quickly over the past few years. I learned a little fine-tuning two years ago. Looking at model training and inference systems today, many things have changed. MoE, sparse attention, low-precision training, long-context serving, agentic RL, OPD. If these remain only names to me, I will quickly fall behind. In technical learning, if you do not move forward, you move backward. That is brutal. You may think you are merely not improving, but many old pieces of knowledge are also expiring. The more troublesome part is that AI can package outdated knowledge so that it still looks usable. --- ## Read More, Practice More, and Study Good Examples The method I can think of now is simple. Read more. Do not only read AI summaries. Read the original text. Even reading one small section a day is better than collecting a pile of summaries. Practice more. When I see a training technique, can I run it on a small model? There are many mini large-model training examples on GitHub. Why not actually learn from them? Even a small Colab experiment is more concrete than only reading concepts. Study good examples. See how others write technical explanations, reproduce experiments, analyze benchmarks, and admit uncertainty. Good examples calibrate my own writing and expose my shallowness. There was a period when I would mess around on Colab. Looking back, although I did not understand many things at the time, that kind of tinkering itself was important. Things failing to run, memory exploding, environment conflicts, unstable results. These problems force you to truly understand something. If learning only happens through AI conversations, knowledge easily becomes something your eyes think they have learned while your brain asks: what exactly do you know? --- ## Math Really Matters This is not entirely a feeling from reading papers. I have always felt that math has a strange charm. At my age, when I start exploring hobbies, math feels like a perfect zero-cost hobby: it kills time and gives both a sense of achievement and frustration. And I am not taking the gaokao or preparing for graduate school exams. What pressure do I have? If I learn badly, nothing happens. From my point of view, it is no different from playing DOTA2 badly for so many years. I am not a professional player. When I first started learning AI-related technology, I spent some time catching up on math, but now I have forgotten much of it again. I need to make up for it. Knowledge really is use it or lose it. Having a stronger math foundation makes reading current large-model research much easier. Otherwise it feels like reading scripture, and I only become less willing to read papers. --- ## Finally So I have decided to treat that AI-written DeepSeek-V4 paper review only as reading support, and slowly read the paper myself. This reflection, though, is something I can publish. Because it is at least a real problem I am facing right now: while using AI to improve efficiency, I am also slowly losing the patience to sit down and learn deeply. This is not something so serious that it requires confession, but it is worth staying alert to. AI is powerful. Precisely because it is powerful, I need to see my own boundaries more clearly. If I am weak, then I should read more papers, write more code, and study more math.