I recently decided to dive into Information Theory, absolutely not because it’s my mandatory course. The standard “Bible” for the subject is Elements of Information Theory by Cover & Thomas. It’s a brilliant book, but honestly, reading through a dense math textbook cover-to-cover is a quick way to kill my motivation.

Instead, I decided to use a Just-In-Time (JIT) learning method: pick a project, and only open the textbook when I need the math to solve a specific problem.

The project? Building an AI that plays Wordle perfectly.

It turns out, Wordle isn’t just a vocabulary game. It is a pure manifestation of Claude Shannon’s 1948 mathematical framework for information. Here is how I translated Chapter 2 of the textbook into a Python script.

I was watching Karpathy’s tutorial nn-zero-to-hero.

When moving from a simple Multi-Layer Perceptron (MLP) language model to a WaveNet (Convolutional) architecture, a fundamental question arises: Why do we need a new layer type?

During several previous CTF competitions, I could do a bit of every category except pwn. So this winter holiday I decided to properly learn it. Turns out it’s really fun when you actually manage to pwn something.

This blog post will keep updating (hopefully).

在这个云计算和 CDN 追求“毫秒级”响应的时代，我决定反其道而行之。

手里有一台位于洛杉矶（LA）的 VPS，平时用来跑 Caddy 反代和做一些 Web 服务。众所周知，从国内直连 LA 的延迟本来就是“看脸”的（平均 200ms-400ms）。

看着终端里的光标闪烁，我突然冒出一个极其“Silly”的想法：既然 SSH 可以端口转发，那我能不能把流量在我和 VPS 之间像打乒乓球一样来回弹射，直到把延迟堆到令人发指的程度？

说干就干，我开启了一场关于 TCP Meltdown（TCP 崩溃） 的实地观测实验。