2026/10/03
我曾经是一个看到数学就摇头的人。
I used to shake my head at math.
现在回头看,那份怕来自太想把事情做好。想做好本身是好的,但它带来了急切:急着比较,急着出结果。
Looking back, I can see where the fear came from: I wanted too much to do things well. That wish is good, but it made me hurry. I hurried to compare, and I hurried to get results.
数学公式需要的是安静。急和静一拉扯,小细节就容易算不清。我当时知道,却选择了跳过,心里一直觉得有点虚。
Math needs a quiet mind, and hurry pulls against quiet. So I lost the small details. I knew it, but I skipped them anyway. On the outside I moved fast; on the inside I felt hollow.
有一段时间,我也认同过网上那种说法:「平时工作用不到,为什么面试还要问这种基础,为什么要刷 LeetCode。」现在看来,有这种想法的人不止我一个。
For a while, I even agreed with a common opinion online: "We never use this at work. So why do interviews test the basics? Why practice LeetCode?" I was not the only one who thought this way.
但当你沉下心来,慢慢看公式,你会发现很多道理就藏在这些公式里。复杂的往往不是问题本身,而是我们的心。
But when you slow down and read the formulas carefully, you find many lessons in them. Often the problem is not complex. Our mind is.
借用我喜欢的柯南:推理取决于心,但真相只有一个。
As Detective Conan, one of my favorite stories, says: reasoning depends on the mind, but there is only one truth.
当你克制地计算公式时,因为对真理的承认和诚实,那些纷乱的念头就像被水洗去了。
When you calculate with patience, you must be honest about what is true. And with that honesty, the noise in your mind washes away.
你知道 R = E(Rt+1 | S, A),就不会过度期望;你记得正态分布的推导过程,就知道抓大放小;你明白优化下界,就知道怎么把问题简化到最低限度。
If you know R = E(Rt+1 | S, A), you stop expecting too much. If you remember how the normal distribution is derived, you keep the large terms and let the small ones go. If you understand how to optimize a lower bound, you know how to make a problem as simple as it can be.
数学也像人生。
Math is a lot like life.
我们花很多力气去面对失败,但面对微小的成功,常常先是看不到;看到了,又容易放大或缩小。
We spend much effort learning to face failure. But small successes are hard too. Often we do not see them. And when we do, we make them too large or too small.
放大容易让人傲慢,缩小容易让人自卑。
Too large, and we become arrogant. Too small, and we lose confidence.
所以我在算数学、算公式的过程中,慢慢练习一种平衡。就像张一鸣说的:保持在一个轻度愉悦与轻度丧的状态。
So as I work on math, I also practice balance. Zhang Yiming said it well: stay mildly happy and mildly down.
2026/10/23
基础不牢,地动山摇。AI 可以做很多事,于是大家开始追求速度。我最近反过来,回头重看 Transformer 等很多基础算法。
When the foundation is weak, everything on top shakes. AI can do so much now that everyone wants speed. I went the other way: back to Transformer and other basic algorithms.
很多人见面会和我聊最近有意思的新技术、新算法。其实我这段时间在看最简单的 CNN,甚至是 softmax、Dijkstra 这些基础。
People often want to talk with me about the newest techniques. But lately I have been reading the simplest things: CNN, softmax, even Dijkstra.
在反复看的过程中,我重新整理,从头思考。
As I read them again and again, I organize what I know and think from the start.
Diffusion 的反向去噪,当时为什么会选 U-Net?
Why does the reverse denoising step in diffusion use a U-Net?
当时是怎么一步一步出现了 diffusion?
How did diffusion appear, step by step?
模型需要处理的信息是什么?
What information does the model really need to process?
怎么从不同的角度用同一个原理解决问题?
How can one principle solve a problem from different angles?
Transformer 的关键也许不只在 attention,还在一次大胆的减法;并行计算的需求,又推动了 Mamba 的出现。
The key to Transformer may not be attention alone. It may also be a bold subtraction. And the need for parallel computation helped give birth to Mamba.
我也会回想:当年我是怎么看这些算法变化的?那时的理解,有没有带着偏差?
I also look back: how did I see these changes then? Was my understanding biased?
文字表达的效率有限,我开始越来越喜欢用公式表达。
Words have limits, so I now prefer to think in formulas.
曾经看着就累的数学推导,现在反而喜欢:乘法变为加法,加项或者减项,优化下界。
Long derivations used to tire me. Now I enjoy them: turn multiplication into addition, add a term, remove a term, optimize a lower bound.
回头看,以前对推导的抗拒,很大一部分来自学生时代的竞争氛围:总在看别人,看不清自己该怎么玩。我想不止我一个人有过这样的感受。能在还年轻的时候想明白这一点,我觉得很幸运。
Looking back, my old dislike of derivations came mostly from competition at school. I was always watching other people's games and could not see my own. I think many people know this feeling. I am lucky that I understood it while I am still young.
热闹的词语和社交媒体的声音很多,有时会让人看不清真正的问题。
Social media is full of loud words, and loud words can hide the real problem.
静下心来,去看那些躺在河底、质朴到不可思议的问题与解法,往往才是出路。
The way forward is often quieter: slow down and look at the plain problems and solutions that lie at the bottom of the river.
比如 Mamba 和 linear attention,从 2020 年前后开始,很多篇论文推导到最后,在公式上和梯度下降连在了一起。看懂这一步,我更能理解为什么有些新方法会热闹一阵,又慢慢安静下去。
Take Mamba and linear attention. The work started around 2020. Paper after paper did the derivations, and in the end the formulas met gradient descent. Once I understood this step, I could see why some new methods shine for a short time and then fade.
大家都很期待天才和突破,这种期待可以理解。但期待本身不会解决问题。
Many people hope for geniuses and breakthroughs. This hope is natural. But hope alone does not solve problems.
在我看来,解决问题需要安静,需要沉下心来琢磨,而不是一边做一边想象做成以后的样子。
To me, solving a problem needs quiet. It needs slow thinking, not dreams of success while you work.
古人做事前会斋戒几天,也许就是这个道理。
In ancient times, people fasted for a few days before important work. Maybe this is why.
后退的算法里,也许才有真正的前向。
Perhaps real forward progress is hidden in a step back.