forked from xlisp/xlisp.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
142 lines (136 loc) · 22.1 KB
/
Copy pathindex.html
File metadata and controls
142 lines (136 loc) · 22.1 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>Steve Chan — xlisp</title>
<meta name="description" content="Steve Chan — Clojure / Emacs / Python Lisp hacker. Notes on machine learning, deep learning, reinforcement learning, and large language models.">
<meta name="author" content="Steve Chan">
<link rel="icon" type="image/jpeg" href="icon.jpeg">
<link rel="apple-touch-icon" href="icon.jpeg">
<meta property="og:type" content="website">
<meta property="og:title" content="Steve Chan — xlisp">
<meta property="og:description" content="Steve Chan — Clojure / Emacs / Python Lisp hacker. Notes on machine learning, deep learning, reinforcement learning, and large language models.">
<meta property="og:image" content="icon.jpeg">
<meta property="og:site_name" content="Steve Chan — xlisp">
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:title" content="Steve Chan — xlisp">
<meta name="twitter:description" content="Steve Chan — Clojure / Emacs / Python Lisp hacker. Notes on machine learning, deep learning, reinforcement learning, and large language models.">
<meta name="twitter:image" content="icon.jpeg">
<link rel="stylesheet" href="style.css?v=light1">
<script>(function(){try{if(localStorage.getItem('theme')==='light')document.documentElement.setAttribute('data-theme','light');}catch(e){}})();</script>
</head>
<body>
<main class="wrap">
<header class="hero">
<pre class="banner"> ___ __ ____ ___ ____
/ _ \(_ ) (_ _)/ __)( _ \
( (_) )/ /_ _)(_ \__ \ )___/
\___/(____)(____)(___/(__) </pre>
<h1>Steve Chan</h1>
<p class="tagline">Clojure / Emacs / Python — Lisp hacker</p>
<p class="links">
<a href="https://github.com/xlisp">github.com/xlisp</a>
</p>
</header>
<section id="about">
<h2>About</h2>
<p>
My name is <strong>Steve Chan</strong> and I'm a
Clojure / Emacs / Python Lisp hacker. I currently focus on
<strong>Machine Learning</strong> and
<strong>Deep Learning</strong>, with special interest in
<strong>Reinforcement Learning</strong> and
<strong>Large Language Models</strong>.
</p>
</section>
<section id="projects">
<h2>Projects</h2>
<ul class="project-list">
<li>
<a href="https://github.com/hulunote/hulunote"><code>hulunote/hulunote</code></a>
<p>A note-taking system for thinkers and hackers — capture, link, and grow your knowledge graph.</p>
</li>
<li>
<a href="https://github.com/hackerlibs/code-hacker"><code>hackerlibs/code-hacker</code></a>
<p>A toolbox of hacker-grade utilities and experiments for code wranglers.</p>
</li>
<li>
<a href="https://github.com/xlisp/CodeChat"><code>xlisp/CodeChat</code></a>
<p>Chat with your codebase — pair-program with LLMs over real source trees.</p>
</li>
</ul>
</section>
<section id="writing">
<h2>Writing</h2>
<ul class="post-list">
<li><span class="post-date">2026-08-07</span> <a href="posts/math-his-cha0-4-modeling.html">建模：把数学当工具用 —— 从托勒密的本轮到 GPT 的权重，一条四千年没断过的线</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/2uhUBZXNUevSWdsI9OdLnA" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-08-06</span> <a href="posts/boltzmann-to-hinton.html">从玻尔兹曼到辛顿: 从 \(S = k \log W\) 到 gpt.py:145 的一条线</a></li>
<li><span class="post-date">2026-08-06</span> <a href="posts/math-his-cha0-99-code-to-formula.html">程序 → 公式：把你已经会的 for 循环，翻译成 \(\Sigma\) —— 程序员逆袭数学的诚意手册</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/0s5zTkQZJgySERZiLtnCFw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-08-06</span> <a href="posts/cloze-to-fim.html">英语考试的完形填空，就是大模型Tab写代码的全部方法</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/gFnHY_a5AOzWo8gfkcVBXg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-08-05</span> <a href="posts/deng-wang-edu-soil.html">菲尔兹奖照出中国教育最大的短板：为什么两位数学天才，都去了欧美？</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/nMiOp2xAXTq93Zw6FCUH2Q" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-08-05</span> <a href="posts/deng-wang-edu-soil-v2.html">从"保护天才"到"滋养创新"：邓煜、王虹的成长路径带来的一点思考</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/WvpIGoCuLxnQaHZ1gFWWPA" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-08-04</span> <a href="posts/why-language-is-learnable-sanskrit.html">为什么语言规律可以被学习：从梵语到 CodeChat 的 8B 权重</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/HN1kMlc0LMRurx13_C43SA" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-08-01</span> <a href="posts/gradient-field-to-phase.html">把磁场掰斜一点点，你的位置就变成了一个角度</a></li>
<li><span class="post-date">2026-07-31</span> <a href="posts/math-his-cha01.html">为什么AI，物理学家顶尖的人才总是出现在欧美？原因出现在这！</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/qnkIPPwIM--Np2MID3s33Q" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-31</span> <a href="posts/math-his-cha02.html">为什么 AI、数学、物理的底层理论，总是诞生在欧美？原因出现在这！</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/n-GdUz0GTDnrQYccLdWpzw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-31</span> <a href="posts/math-his-cha0-6-symbols.html">为什么 AI 论文全是天书？因为每一个数学符号，其实都是一句人话</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/TnCDdpdcyCUsX6LrjN_S0A" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-30</span> <a href="posts/from-linear-to-nonlinear.html">大学4年没讲明白的神经网络数学，被一段 PyTorch 代码讲透了</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/N6ijiaBLqxNjqFF4Z3mqkA" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-30</span> <a href="posts/langgraph-in-practice.html">95% 的 LangGraph 教程，都没告诉你它真正解决了什么</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/cgQuhAYe3mFGCd6bUrYHvw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-30</span> <a href="posts/loop-engineering.html">AI圈突然都在聊 Loop、Graph Engineering，炒作还是新东西？</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/gwqXvQXJqLIA-sOWvfCA6A" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-27</span> <a href="posts/mri-to-multimodal.html">医院那台核磁共振，和大模型的位置编码、多模态，是同一件事</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/zP1tk1ZcwYgtWFn6nf77fw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-23</span> <a href="posts/sft-rl-inference-mechanics.html">训练写权重，推理用权重 + 脚手架：SFT / RL 训完之后到底是怎么生效的</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/UFjvHPkX8W58o4SkdCgaZQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-22</span> <a href="posts/mask-causal-triangular.html">Mask、三角矩阵与因果掩码：历史、起源与原理</a></li>
<li><span class="post-date">2026-07-17</span> <a href="posts/synthetic-data.html">Claude 的暗黑科技：把垃圾代码炼成黄金训练数据</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/nUsdS0fJmArjZc8tCS4X1w" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-13</span> <a href="posts/math-his-cha0.html">为什么AI，数学家顶尖的人才总是出现在欧美？原因出现在这！</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/FmO9GiRMOV-XM_OeSknFeQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-09</span> <a href="posts/rl-history-from-flappybird-to-grpo.html">ChatGPT 会推理的秘密，藏在一款游戏身上</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/sCDUv7ZCErXN8pclZGn6Yw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-09</span> <a href="posts/rl-dog-training-to-llm-reasoning.html">训一条狗，和训出会"思考"的 DeepSeek，用的是同一套办法</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/qV5fG7UkF3zxUt_5hBuxfg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-08</span> <a href="posts/zscore-to-layernorm.html">为什么大模型能堆到几十层还训得动？答案藏在一个最不起眼的动作里</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/tnIQbAD8CzIXjCg1KtCcAQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-07</span> <a href="posts/complex-numbers-to-rope.html">高中觉得最没用的"虚数 i"，其实是大模型旋转位置编码最优雅的写法</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/_CPGU3fDuHcnuNNIbyPXqQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-05</span> <a href="posts/logarithm-to-cross-entropy.html">高中觉得最没用的 log，其实是大模型判断"自己错得有多离谱"的唯一标尺</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/-jvEdxmZTsptooJQVMjGsQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-05</span> <a href="posts/vectors-to-attention.html">高中觉得只是"带箭头的线段"的向量，其实是大模型把"意思"变成"方向"的地基</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/6AGTXGvqum_j96sPh_CohQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-05</span> <a href="posts/combinatorics-to-generalization.html">高中背到头秃的排列组合，其实解释了大模型为什么"不可能靠背、只能靠学"</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/_rtIc1VdPLC-VLJqk2fm5w" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-05</span> <a href="posts/limits-to-temperature.html">高中觉得最玄乎的"极限"，其实是大模型"从犹豫到笃定"的那个旋钮</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/FiGVRF0OFBiuRvw7c7jcwg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-05</span> <a href="posts/induction-to-autoregression.html">高中最容易被忽略的数学归纳法，其实就是大模型"逐字生成"的骨架</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/LVpfsmOUKltiyAggt2-zvw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-03</span> <a href="posts/sequences-to-autoregression.html">高中背了三年的数列，原来是 GPT 一个字一个字蹦出来的秘密</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/esw_IhBHGsFUynE3LcSYrw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-01</span> <a href="posts/information-theory-through-pytorch.html">大学4年没讲明白的信息论，被一段 PyTorch 代码讲透了</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/0amu7ysKNy0qDwwWRs7qWw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-07-01</span> <a href="posts/trigonometry-to-rope.html">高中背了三年的 sin 和 cos，原来是大模型记住"谁先谁后"的秘密</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/aeCUH9ms9pFr1VbqNqbl8g" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-28</span> <a href="posts/symbolic-bayes-neural.html">Mythos、Fable 5 的背后，其实是一条 70 年没人讲透的 AI 主线</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/Sj2bn-PAZFliXrU5-uqGyw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-27</span> <a href="posts/probability-through-pytorch.html">大学4年没讲明白的概率论，被一段 PyTorch 代码讲透了</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/DbQXV42DeCOEQJ5zHykd6g" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-25</span> <a href="posts/from-quadratic-formula-to-llm.html">高中老师说"买菜用得上求根公式吗"——可这条公式，是整个大模型的祖宗</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/yNz0f8E8VQIEjxjwcxz7ng" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-23</span> <a href="posts/linear-algebra-in-one-pytorch-snippet.html">大学4年没讲明白的线性代数，被一段 PyTorch 代码讲透了</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/ywgU3NuNmHow2dY8k-wpsA" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-21</span> <a href="posts/how-to-write-an-mcp.html">用最小作用原理写一个 MCP：模型的那双手，该怎么造</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/-vK0Jb2v4vJtnOj7bCwpwQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-20</span> <a href="posts/prompt-as-code.html">把提示词当代码写：Prompt Engineering 的第一性原理</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/Zr5dU95j-O-FaAtZcoGDtg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-17</span> <a href="posts/in-depth-analysis-of-react.html">从 CoT 到 ReAct，再到"会自己思考"的模型</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/Kp_r8HmDPkef8xW6n-aR-A" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-06-15</span> <a href="posts/wechat-article-gpt.html">AI 最大的缺陷不是智商，而是失忆：我们做了一个 AI 的海马体</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/Ki-Inwp_b4AKSjSILxAaig" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-05-21</span> <a href="posts/harness-is-everything.html">Harness 即一切：模型不是不够聪明，而是是否有足够有用的tools</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/l7p1E0WMnPzkMI2CCVkWpw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-05-13</span> <a href="posts/universal-function-approximator.html">万能函数模拟器：大模型就是用微积分从数据里"积分"出来的函数 P</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/F0d16B3Jig0Si-4OHAJK0w" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-05-06</span> <a href="posts/code-search-to-gpt.html">从代码语义搜索到 GPT 写代码：五代范式、三次质变与一条贯穿始终的"搜索维度提升"主线</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/Spsg5Odl0yvBoaaMxkGuWg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-05-06</span> <a href="posts/gpt-as-super-search.html">GPT 是更高级的"万能谷歌搜索"：从关键词到高维空间的搜索进化史</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/6lDVbFAwDXwQ5IXC8NdfVw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-05-06</span> <a href="posts/eda-and-llm.html">EDA、符号主义与 LLM 的交汇</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/_smB4c6tOaxXlFdVAHxMyw" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-05-06</span> <a href="posts/why-you-not-student.html">Why You Not Student?</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/MYqxTxiJDGM6WlumbgqacQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-05-04</span> <a href="posts/probability-as-area-matrix-as-map.html">概率就是面积，矩阵就是映射：大模型最底层的两块拼图</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/nAYdQ-Ss3UiOuPlAsJkVzg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-04-28</span> <a href="posts/differentiable-programming.html">深度学习是可微分编程：从 y = wx + b 讲到 CodeGPT</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/vYmBUkVGXd5tK0IpoFwK6w" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-04-26</span> <a href="posts/physics-and-deep-learning.html">物理学的影子：量子力学与统计力学如何塑造了深度学习</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/ZJxyh1DsxacNeH0VK1eCAA" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-04-25</span> <a href="posts/transformer-dominates-all.html">同一个 Transformer，吃掉一切：为什么语音、图像、自动驾驶、大模型都在用它</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/Tx5HopK3--g5UQj_FZDjBg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-04-25</span> <a href="posts/transformer-from-vision.html">从图像视角理解 Transformer</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/bcYzyD77gBQ0RT98n7qVJg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-04-20</span> <a href="posts/rag-vs-sft.html">RAG 还是 SFT：面对一堆私有数据，该怎么选？</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/dVhnA4ZBEsPenFaZ2Lu7Bg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-04-20</span> <a href="posts/transformer-debugger.html">Transformer Debugger：把大模型从黑盒拆成白盒</a></li>
<li><span class="post-date">2026-04-19</span> <a href="posts/sft-forgetting-and-moe.html">多次 SFT 的灾难性遗忘：SFT 的本质、MoE 的本质、以及它们各自解决什么问题</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/j0UgHGjnpU6Qawlq2uEQeg" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-04-16</span> <a href="posts/mixed-sft-vs-moe.html">混合 SFT vs MoE：同一思路吗？</a></li>
<li><span class="post-date">2026-04-14</span> <a href="posts/multi-gpu-training.html">从单卡到多卡：分布式训练原理与 FSDP 实战</a></li>
<li><span class="post-date">2026-04-11</span> <a href="posts/swebench-rl.html">GRPO on SWE-bench Train: Docker-in-the-Loop RL</a></li>
<li><span class="post-date">2026-04-08</span> <a href="posts/rl-algorithms.html">CodeChat RL 算法选型</a></li>
<li><span class="post-date">2026-04-02</span> <a href="posts/rl-sft-grpo-intro.html">LLM 训练技术详解：从 SFT 到 GRPO</a></li>
<li><span class="post-date">2026-03-12</span> <a href="posts/compression-is-intelligence.html">压缩即智能：从自编码器到 GPT 的认知哲学</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/Fi7HLJCHzoNxNe3xzVrtLQ" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-03-12</span> <a href="posts/deep-dive.html">从 RNN 到 CodeGPT：序列建模的进化史</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/ng1IbwPoc9Cc_2LXred4og" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
<li><span class="post-date">2026-03-12</span> <a href="posts/rlhf-and-platonic-representation.html">强化学习对齐与柏拉图表征：ChatGPT 成功的另外两块拼图</a> <a class="post-wechat" href="https://mp.weixin.qq.com/s/HuXFNhP4oKitGgpuemyH4A" target="_blank" rel="noopener" title="在微信公众号阅读">微信</a></li>
</ul>
</section>
<footer>
<p>© 2026 Steve Chan · built with a tiny <a href="https://github.com/xlisp/xlisp.github.io">python script</a></p>
</footer>
</main>
<script src="theme.js"></script>
</body>
</html>