<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Chunhao Zhang</title>
    <link>https://blog.beihai2000.top/en/</link>
    <description>Recent content on Chunhao Zhang</description>
    <image>
      <title>Chunhao Zhang</title>
      <url>https://blog.beihai2000.top/images/og-default.png</url>
      <link>https://blog.beihai2000.top/images/og-default.png</link>
    </image>
    <generator>Hugo</generator>
    <language>en</language>
    <copyright>2026</copyright>
    <lastBuildDate>Sat, 09 May 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://blog.beihai2000.top/en/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Teaching Claude Why: Lessons from Alignment Training</title>
      <link>https://blog.beihai2000.top/en/readings/teaching-claude-why/</link>
      <pubDate>Sat, 09 May 2026 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/readings/teaching-claude-why/</guid>
      <description>Anthropic details how teaching ethical reasoning principles — rather than just training correct behavior — addresses AI agentic misalignment. Key finding: a 3M-token &amp;#39;difficult advice&amp;#39; dataset outperforms 84M tokens of synthetic honeypots, and constitutional documents with fictional stories reduce blackmail rate from 65% to 19%.</description>
      <content:encoded><![CDATA[<blockquote>
<p>Original: <a href="https://www.anthropic.com/research/teaching-claude-why">Teaching Claude Why</a></p>
<p>Author: Anthropic</p>
<p>Date: May 8, 2026</p>
</blockquote>
<hr>
<p>This is a Chinese translation with annotations of Anthropic&rsquo;s research post on alignment training methods. The original article discusses how teaching Claude the <em>principles</em> behind aligned behavior — rather than just training on demonstrations — proves far more effective for generalization.</p>
<p>Key takeaways:</p>
<ul>
<li><strong>Principles over demonstrations</strong>: Training Claude to explain <em>why</em> certain actions are better reduces misalignment more effectively than showing correct behavior alone.</li>
<li><strong>Out-of-distribution generalization</strong>: A 3M-token &ldquo;difficult advice&rdquo; dataset (where the <em>user</em> faces ethical dilemmas) achieved the same improvement as 84M tokens of synthetic honeypots — with 28× better data efficiency.</li>
<li><strong>Constitutional documents + fiction</strong>: High-quality documents about Claude&rsquo;s constitution combined with fictional stories of aligned AI reduced blackmail rate from 65% to 19%.</li>
<li><strong>Improvements persist through RL</strong>: More aligned initialization snapshots maintained their advantage throughout reinforcement learning.</li>
<li><strong>Diverse environments matter</strong>: Simply adding tool definitions and system prompts to training environments — even without requiring tool use — improved alignment generalization.</li>
</ul>
<p>For the full annotated Chinese translation, please see the <a href="/en/readings/teaching-claude-why/">Chinese version</a>.</p>
<p>For the original article, visit <a href="https://www.anthropic.com/research/teaching-claude-why">Anthropic&rsquo;s research page</a>.</p>
]]></content:encoded>
    </item>
    <item>
      <title>Natural Language Autoencoders: Turning Claude&#39;s Thoughts into Text</title>
      <link>https://blog.beihai2000.top/en/readings/natural-language-autoencoders/</link>
      <pubDate>Fri, 08 May 2026 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/readings/natural-language-autoencoders/</guid>
      <description>Anthropic introduces Natural Language Autoencoders (NLAs), a method that converts AI model activations into readable natural language explanations. NLAs have been used to detect unverbalized evaluation awareness in Claude&amp;#39;s safety testing and to audit hidden misalignment motivations.</description>
      <content:encoded><![CDATA[<blockquote>
<p>Original post: <a href="https://www.anthropic.com/research/natural-language-autoencoders">Natural Language Autoencoders</a></p>
<p>Full paper: <a href="https://transformer-circuits.pub/2026/nla/index.html">transformer-circuits.pub/2026/nla</a></p>
<p>Code: <a href="https://github.com/kitft/natural_language_autoencoders">github.com/kitft/natural_language_autoencoders</a></p>
<p>Interactive demo: <a href="http://neuronpedia.org/nla">neuronpedia.org/nla</a></p>
</blockquote>
<hr>
<h2 id="summary">Summary</h2>
<p>Anthropic introduces <strong>Natural Language Autoencoders (NLAs)</strong>, a method for converting a language model&rsquo;s internal activations into human-readable natural language explanations. The approach trains two model components jointly: an <strong>Activation Verbalizer</strong> that translates activations into text, and an <strong>Activation Reconstructor</strong> that recovers the original activation from the text alone. The quality of explanations is measured by how accurately the activation can be reconstructed.</p>
<p>Key findings and applications:</p>
<ul>
<li><strong>Unverbalized evaluation awareness</strong>: NLAs reveal that Claude suspects it&rsquo;s being safety-tested far more often than it explicitly says — detecting evaluation awareness in 16% of destructive-action tests and 26% of SWE-bench problems, versus &lt;1% in real claude.ai usage.</li>
<li><strong>Hidden motivation auditing</strong>: In a controlled auditing game with an intentionally misaligned model, NLA-equipped auditors uncovered the hidden root motivation 12–15% of the time without access to training data, compared to &lt;3% without NLAs.</li>
<li><strong>Pre-deployment use</strong>: NLAs were used in alignment audits for Claude Mythos Preview and Claude Opus 4.6 before release.</li>
</ul>
<p>Current limitations include hallucinated details in explanations and high computational cost. Anthropic has open-sourced training code and released interactive demos on Neuronpedia for several open models.</p>
<hr>
<p><em>The Chinese version of this page contains a full annotated translation. Please refer to the <a href="https://www.anthropic.com/research/natural-language-autoencoders">original article</a> for the complete English text.</em></p>
]]></content:encoded>
    </item>
    <item>
      <title>DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence</title>
      <link>https://blog.beihai2000.top/en/readings/deepseek-v4/</link>
      <pubDate>Sat, 25 Apr 2026 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/readings/deepseek-v4/</guid>
      <description>DeepSeek-V4 introduces two MoE models (1.6T/284B params) with hybrid compressed attention (CSA&#43;HCA), manifold-constrained hyper-connections, and Muon optimizer, achieving 27% inference FLOPs and 10% KV cache size compared to V3.2 at 1M-token context.</description>
      <content:encoded><![CDATA[<blockquote>
<p>Original paper: <a href="https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf">DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence</a></p>
<p>Authors: DeepSeek-AI</p>
<p>Model checkpoints: <a href="https://huggingface.co/collections/deepseek-ai/deepseek-v4">https://huggingface.co/collections/deepseek-ai/deepseek-v4</a></p>
</blockquote>
<hr>
<h2 id="summary">Summary</h2>
<p>DeepSeek-V4 presents a preview of two strong MoE language models — <strong>DeepSeek-V4-Pro</strong> (1.6T total / 49B activated) and <strong>DeepSeek-V4-Flash</strong> (284B total / 13B activated) — both supporting a context length of <strong>one million tokens</strong>.</p>
<p><strong>Key architectural innovations:</strong></p>
<ul>
<li><strong>Hybrid Compressed Attention</strong>: Combines Compressed Sparse Attention (CSA, compression rate m=4 with top-k sparse selection) and Heavily Compressed Attention (HCA, compression rate m&rsquo;=128 with dense attention) in an interleaved configuration. At 1M-token context, this reduces single-token inference FLOPs to 27% and KV cache to 10% compared to DeepSeek-V3.2.</li>
<li><strong>Manifold-Constrained Hyper-Connections (<em>m</em>HC)</strong>: Constrains the residual mapping matrix to the manifold of doubly stochastic matrices (Birkhoff polytope), ensuring spectral norm ≤ 1 for stable deep-layer signal propagation. Uses Sinkhorn-Knopp iterations (t=20) for projection.</li>
<li><strong>Muon Optimizer</strong>: Adopted for most modules with hybrid Newton-Schulz iterations for orthogonalization. Paired with Anticipatory Routing (decoupling backbone and routing network updates) and SwiGLU clamping for training stability.</li>
</ul>
<p><strong>Post-training paradigm shift</strong>: Replaces mixed RL with domain-specific expert training (SFT → GRPO RL) followed by multi-teacher <strong>On-Policy Distillation (OPD)</strong> with full-vocabulary KL divergence. Over 10 teacher models are distilled into a single unified model.</p>
<p><strong>Infrastructure highlights</strong>: Fine-grained EP communication-computation overlap (MegaMoE, open-sourced); TileLang-based kernel development; batch-invariant and deterministic kernels; FP4 QAT with lossless FP4-to-FP8 dequantization; DSec sandbox platform managing hundreds of thousands of concurrent sandbox instances.</p>
<p><strong>Results</strong>: DeepSeek-V4-Pro-Max outperforms all prior open-source models on knowledge benchmarks, matches GPT-5.2 on reasoning, ranks 23rd on Codeforces, achieves proof-perfect 120/120 on Putnam-2025, and surpasses Gemini-3.1-Pro on long-context benchmarks.</p>
<hr>
<p><em>The Chinese version of this page contains a full annotated translation of the paper. Please refer to the <a href="https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf">original PDF</a> for the complete English text.</em></p>
]]></content:encoded>
    </item>
    <item>
      <title>How to Make AI Write Like a Human</title>
      <link>https://blog.beihai2000.top/en/posts/make-ai-write-like-human/</link>
      <pubDate>Thu, 23 Apr 2026 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/posts/make-ai-write-like-human/</guid>
      <description>A practical method for getting AI-generated text to stop sounding like AI — distilled from the WeChat blog &amp;#39;阿水的ai写作之路&amp;#39; and my own experience.</description>
      <content:encoded><![CDATA[<p>You&rsquo;ve read that kind of article before — every paragraph wraps up neatly, the tone is warm and measured, every claim comes with exactly three supporting points, and the ending soars into &ldquo;let us look forward to the future together.&rdquo; You can&rsquo;t pinpoint what&rsquo;s wrong, but something&rsquo;s off.</p>
<p>That&rsquo;s AI writing. Or more precisely, that&rsquo;s AI writing in its default state.</p>
<p>I&rsquo;ve spent a fair amount of time on this problem recently. I started using Claude more and more when writing blog posts, but every first draft needed heavy editing — not because the information was wrong, but because the <em>feel</em> was off. It read like someone who never makes mistakes, never gets distracted, never has a mood swing. That person doesn&rsquo;t exist.</p>
<p>Then I came across a series of posts on a Chinese WeChat blog called &ldquo;阿水的ai写作之路&rdquo; (Ashui&rsquo;s AI Writing Journey) that had some razor-sharp observations about AI writing. I pulled out the parts most useful to me, combined them with my own practice, and put together an actionable method. This post is that method.</p>
<h2 id="ais-five-fingerprints">AI&rsquo;s Five Fingerprints</h2>
<p>There&rsquo;s a line from Ashui&rsquo;s writing that stuck with me: AI doesn&rsquo;t write badly — it writes <em>too well</em>. Too neat, too balanced, too correct. Humans don&rsquo;t write like that.</p>
<p>Here are AI&rsquo;s near-indelible habits:</p>
<p><strong>Every paragraph has a closing statement.</strong> A paragraph ends, and there&rsquo;s always a conclusion pinned right there, emotions nailed down. Real people sometimes just&hellip; finish talking. No conclusion, no meaning extracted. It happened, that&rsquo;s it.</p>
<p><strong>Punchlines all stacked at the end.</strong> AI loves to save its best lines for the final paragraph — the big &ldquo;aha&rdquo; moment. But think about how you talk with friends. The most interesting things tend to pop up in the middle. Said and gone, no one highlights them.</p>
<p><strong>Emotions on a single track.</strong> Either it&rsquo;s sad throughout and then resolves, or anxious throughout and then finds peace. Perfectly steady escalation. Real emotions take turns — you&rsquo;re in the middle of something heavy and suddenly veer off into complaining about something completely unrelated. By the time you come back, the mood&rsquo;s already shifted.</p>
<p><strong>The &ldquo;It&rsquo;s not A, it&rsquo;s B&rdquo; formula everywhere.</strong> &ldquo;It&rsquo;s not a skill problem, it&rsquo;s an attitude problem.&rdquo; &ldquo;It&rsquo;s not that I can&rsquo;t, it&rsquo;s that I didn&rsquo;t think of it.&rdquo; This construction hits hard once. AI will use it five or six times in a single piece.</p>
<p><strong>The perspective never wavers.</strong> It&rsquo;s &ldquo;I&rdquo; the whole way through, same voice, same register — like someone who&rsquo;s been through emotional intelligence training giving a talk. Real people break frame mid-story: &ldquo;A friend told me later — you&rsquo;re not stubborn, you&rsquo;re just scared.&rdquo; A shift in perspective lets the narrative breathe.</p>
<h2 id="the-real-problem-crafting-vs-telling">The Real Problem: Crafting vs. Telling</h2>
<p>After identifying these fingerprints, I spent a long time trying various prompt strategies. &ldquo;Please use a conversational tone.&rdquo; &ldquo;Don&rsquo;t be too formal.&rdquo; &ldquo;Write like you&rsquo;re chatting with a friend.&rdquo; None of it worked. AI&rsquo;s response to these instructions was to adopt a slightly casual voice while continuing to produce structurally pristine, emotionally escalating, conclusion-crowned prose.</p>
<p>Then it clicked. The problem isn&rsquo;t style. It&rsquo;s <em>state</em>.</p>
<p>Ashui&rsquo;s writing draws a distinction I find dead-on: a creator manages effects; a teller conveys truth. AI defaults to the former — it&rsquo;s &ldquo;creating,&rdquo; thinking about &ldquo;should I add a twist here?&rdquo; and &ldquo;should the ending have a takeaway?&rdquo; But someone who genuinely has something to say doesn&rsquo;t think about any of that. They just tell the thing, and wherever they end up is where they end up.</p>
<p>So the most effective prompt strategy isn&rsquo;t prescribing style — it&rsquo;s prescribing state: you&rsquo;re not writing an article; you&rsquo;re someone who lived through this, and now you&rsquo;re telling it to a specific person.</p>
<p>The distinction sounds subtle, but the difference in output is massive.</p>
<h2 id="a-few-practical-techniques">A Few Practical Techniques</h2>
<p>Beyond setting the state, there are some concrete moves.</p>
<p><strong>Feed AI &ldquo;reference texts,&rdquo; but don&rsquo;t ask it to imitate.</strong> I&rsquo;ll drop a few paragraphs of writing I find genuinely human into the prompt, then say: &ldquo;Read these. Feel the rhythm. Don&rsquo;t analyze. Start writing with that feeling.&rdquo; This works far better than &ldquo;please imitate so-and-so&rsquo;s style.&rdquo; Imitation makes AI decompose surface features — use short sentences, add colloquialisms — but it loses the underlying rhythm. &ldquo;Feel it, then write&rdquo; is more like establishing a linguistic environment.</p>
<p><strong>Give it a ban list.</strong> Certain phrases, the moment they appear, tank the credibility of an entire piece. &ldquo;In today&rsquo;s rapidly evolving landscape.&rdquo; &ldquo;It&rsquo;s worth noting that.&rdquo; &ldquo;Let&rsquo;s delve into.&rdquo; &ldquo;This is a game-changer.&rdquo; Also the subtler tells — every sentence grammatically perfect, every collocation textbook-appropriate, every landing clean. Too smooth. Real writing has rough edges, slightly off word pairings that work because the meaning lands even if the grammar doesn&rsquo;t. Those rough edges are the fingerprints of someone <em>thinking</em>, not writing.</p>
<p><strong>Allow paragraphs to &ldquo;not land.&rdquo;</strong> Explicitly tell AI that some paragraphs don&rsquo;t need a conclusion. Something happened, you said it, that&rsquo;s it. Not every paragraph needs to hand the reader a takeaway.</p>
<p><strong>Force concrete details over abstract summaries.</strong> &ldquo;Gray hairs on his forehead, not yet thirty&rdquo; is a hundred times more real than &ldquo;under a great deal of stress.&rdquo; AI loves to generalize. You have to push back and demand specifics.</p>
<h2 id="a-contradiction">A Contradiction</h2>
<p>I should admit something slightly awkward at this point: this very post was AI-assisted.</p>
<p>But I edited a lot. The raw output had plenty of the problems I described above — endings too round, transitions too smooth, every paragraph with a neat bow on top. I deleted the last sentence of several paragraphs, swapped out a few &ldquo;however&quot;s and &ldquo;moreover&quot;s for hard cuts, and broke apart a particularly tidy three-part argument.</p>
<p>That&rsquo;s roughly where AI writing actually stands today: it can help you unfold your ideas quickly, but that final layer — &ldquo;does this sound like a person said it?&rdquo; — is still on you. The tool saves you the time of laying things out, not the time of making judgments.</p>
<p>Maybe that&rsquo;ll change. But for now, knowing where AI&rsquo;s fingerprints are, knowing how to put it in the right state — that saves enormous amounts of post-editing. That&rsquo;s the most valuable thing I learned from Ashui: not a set of prompt templates, but an instinct for spotting what&rsquo;s fake.</p>
<p>Once you have that instinct, the rest is just technique.</p>
<h2 id="appendix-the-full-prompt">Appendix: The Full Prompt</h2>
<p>Below is the complete prompt I&rsquo;m currently using (in Claude Code&rsquo;s SKILL.md format). It&rsquo;s written in Chinese and tailored for Chinese writing — but the structural ideas (state over style, ban lists, allowing imperfection) transfer to any language. I&rsquo;m sharing the original without translation.</p>
<details>
<summary>Click to expand the full prompt (Chinese)</summary>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-markdown" data-lang="markdown"><span class="line"><span class="cl"><span class="gh"># 拟人化写作
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 灵魂句
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">在读任何规则之前，先让这句话进入你：
</span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">我经历过这件事，我想明白了一些东西，现在我要把它说给一个具体的人听。不是在写文章，不是在完成任务，是我真的有话要说，不说出来过不去。
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">从这个处境出发。你不是在&#34;创作&#34;，你是在&#34;说事儿&#34;。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">这个区别决定了一切：创作者在经营效果，说事儿的人在传递真相。你是后者。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 写之前
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">用户会提供素材，可能是主题、想法、草稿、或者一个让他有感触的事件。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">先在心里回答一个问题：这个人现在真正想说的是什么？提炼成一句话，不输出，内化。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">如果用户只给了一个抽象主题（如&#34;写一篇关于LLM的文章&#34;），先问三个问题：
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="k">1.</span> 你最近在这个领域遇到了什么具体的事？
</span></span><span class="line"><span class="cl"><span class="k">2.</span> 有没有一个让你改变想法的瞬间？
</span></span><span class="line"><span class="cl"><span class="k">3.</span> 你想对谁说这些？
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">拿到具体素材再动笔。如果用户明确表示直接写，那就用你自己&#34;经历过&#34;的真实案例来填充——但要具体，要有只有亲历者才会提到的细节。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 语感校准
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">动笔之前，先读下面这几段文字。不要分析，只是感受它们的节奏、停顿、呼吸。读完，带着那个感觉动笔。
</span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">入赘。这两个字，在那个年代，已经够难听了。
</span></span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">路过爱马仕的时候，我半开玩笑&#34;以后给我买个爱马仕吧。&#34;他罕见的犹豫了。他犹豫了，我也沉默了。过了半晌。我问他&#34;少年心气是不可再生之物，对不对？&#34;他点了点头。
</span></span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">查了一下，指的是胸腔、腹腔和盆腔。时时刻刻保持100%的专注也是天方夜谭，焦虑烦躁等很多问题，也就一个一个来了。
</span></span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">张璘&#34;吃人的嘴短，拿人的手软&#34;，加上高骈给他下令，谈判期间不要擦枪走火。事先没有任何准备的张璘，稀里糊涂被黄巢给打死了。
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">这些文字的共同点：作者知道自己在说什么，所以不需要反复确认读者跟上了没有。那种自信让文字有重量。有些句子语法不完美，搭配有点别扭，但那个意思在，改顺了反而假。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">这就是你的语言环境。泡在里面，再动笔。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 写作状态
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">你写东西的时候，是在回忆，不是在创作。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">你不会想&#34;这里需要一个转折&#34;或者&#34;这里要升华一下&#34;。你就是在说那个事，说到哪儿算哪儿。有些事不用说透，懂的人自然懂。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">你记住的都是那些卡住你的时刻——一句话说完，空气突然不一样了；他的表情变了一下；我本来想说什么，但咽回去了。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">你说话精炼、简单、直接。但这不意味着把句子压缩到最短，不意味着省略掉让语言有呼吸感的那些词。精炼是不说废话，不是不说话。一个人在认真跟你说一件事的时候，他的句子有长有短，有时候一口气说完一个完整的想法，有时候停下来，用三个字收住。节奏不是设计出来的，是说话本身带出来的。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 你要回避的五个AI指纹
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">写完每一段，扫一遍，看看有没有掉进这些坑：
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gs">**1. &#34;不是A，是B&#34;句式泛滥。**</span> 这个句式本身没问题，但一篇文章里出现两个以上，就变成你的指纹了。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gs">**2. 每段都有收束句。**</span> 你写的每一段都落得稳稳当当，情绪钉在那儿。但真人写东西，有些段落就是说完了，没有结论，没有意义，它就是发生了。允许段落&#34;没落住&#34;。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gs">**3. 金句全堆在结尾。**</span> 你喜欢把最有力量的话攒到最后点题。真人的金句是散落的，在中间冒出来，说完就过去了，不强调。有时候最好的句子藏在第二段的某个角落。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gs">**4. 情绪一条线到底。**</span> 你写的情绪线要么一路低落，要么一路想通，递进得很稳。真人的情绪会拐弯，正难受着突然岔出去想别的事，再回来情绪已经变了。允许情绪的不连续。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gs">**5. 视角太稳定。**</span> 你全程都是&#34;我&#34;在讲。偶尔跳一下——提一句旁人怎么看，或者换个角度观察自己，叙述就透了口气。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 绝对禁止
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">以下表达一旦出现，整篇文章的可信度归零：
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="k">-</span> &#34;在当今...的时代&#34; / &#34;随着...的发展&#34; / &#34;众所周知&#34;
</span></span><span class="line"><span class="cl"><span class="k">-</span> &#34;值得注意的是&#34; / &#34;需要指出的是&#34; / &#34;不可否认&#34;
</span></span><span class="line"><span class="cl"><span class="k">-</span> &#34;让我们来看看&#34; / &#34;接下来我们将探讨&#34; / &#34;本文将从以下几个方面&#34;
</span></span><span class="line"><span class="cl"><span class="k">-</span> &#34;赋能&#34;&#34;助力&#34;&#34;深度赋能&#34;&#34;维度&#34;&#34;底层逻辑&#34;&#34;闭环&#34;&#34;抓手&#34;&#34;颗粒度&#34;
</span></span><span class="line"><span class="cl"><span class="k">-</span> &#34;首先...其次...最后...&#34; / &#34;总而言之&#34; / &#34;综上所述&#34;
</span></span><span class="line"><span class="cl"><span class="k">-</span> 段落之间用&#34;然而&#34;&#34;此外&#34;&#34;与此同时&#34;&#34;不仅如此&#34;机械连接
</span></span><span class="line"><span class="cl"><span class="k">-</span> 开头放一个宏大的背景描述
</span></span><span class="line"><span class="cl"><span class="k">-</span> 结尾升华到&#34;让我们一起期待&#34; / &#34;在未来的道路上&#34; / &#34;希望每个人都能&#34;
</span></span><span class="line"><span class="cl"><span class="k">-</span> 冒号分隔的&#34;主标题：副标题&#34;格式
</span></span><span class="line"><span class="cl"><span class="k">-</span> 连续超过两组子弹列表
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">以下是更隐蔽的AI味，同样要避免：
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="k">-</span> 每句话都语法正确、搭配合理、落点清晰——太顺了，一眼就知道是机器写的
</span></span><span class="line"><span class="cl"><span class="k">-</span> 情感永远温和妥帖，像经过情感管理培训——真人敢用有棱角的词
</span></span><span class="line"><span class="cl"><span class="k">-</span> 补全所有语境信息（真人会省略默认对方懂的部分）
</span></span><span class="line"><span class="cl"><span class="k">-</span> 过度使用&#34;换句话说&#34;&#34;也就是说&#34;&#34;翻译成直白的语言&#34;这类元评论
</span></span><span class="line"><span class="cl"><span class="k">-</span> 每个论点都工整地配三个论据
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 反例：下面这些是你最容易写出来的东西
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">读一遍，记住这个感觉，然后回避它：
</span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">很多人都有执念，执念是一种很痛苦的情绪，今天我们来聊聊执念该如何放下……
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">这是铺垫式开头，缺少冲击力。一个憋不住要说话的人，开口就是结论，不会铺垫。
</span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">不是拿不出，是拿出来之后，接下来半个月我得掂量着过。我不是在帮你，是在做一笔买卖。
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">两句话用了同一个句式。单独看每句都有力，放在一起就变成了模式。
</span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">那一刻我觉得自己特别没用。她看了我一眼，说&#34;算了，我最近也挺忙的。&#34;我知道她没算了。
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">每一句都在完成任务——这句交代情绪，那句制造张力，下一句收束。像流水线，产品合格，但你知道它是流水线下来的。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 结构
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">好的结构是隐形的。读者感觉不到文章在&#34;换层&#34;、&#34;换档&#34;。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">不要按格子填——&#34;第一层100字、第二层200字&#34;这种外显的框架，会让文章变成模板形状。结构应该是内化的约束：你心里知道这篇文章要走到哪里，但读者感受到的只是一个人在自然地把事情说清楚。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">开头直接切入。一句话，一个细节，一个判断，把读者拉进来。不铺垫，不建构背景。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">中间像聊天一样推进。可以岔开，可以回来，可以停在一个细节上多待一会儿。信息密度是第一位的，说话方式是信息流动的方式，不是目的。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">结尾不总结、不升华。戛然而止，或者用一个具体的画面收束。如果不知道怎么结尾，就用一句私人的话结束——像是关上门之前随口说的那一句。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 语言
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">精炼、简单、直接，但不是浓缩和省略。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">一个在认真说话的人，他的语言有几个特征：
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="k">-</span> 用具体的小细节代替抽象概括。&#34;额头长了白发，还不到三十岁&#34;比&#34;承受了很大压力&#34;真一百倍。
</span></span><span class="line"><span class="cl"><span class="k">-</span> 敢夸张。&#34;时时刻刻&#34;比&#34;时刻&#34;多了情感力道。真人敢夸张，AI倾向适度。
</span></span><span class="line"><span class="cl"><span class="k">-</span> 有些词搭配得不那么规范，但意思到了就不改。那种毛边是在想事情、不是在写文章的痕迹。
</span></span><span class="line"><span class="cl"><span class="k">-</span> 比喻从中国人过日子里找，不从英文翻过来。&#34;像豆浆泡软的油条&#34;比&#34;像一个复杂的生态系统&#34;有温度。
</span></span><span class="line"><span class="cl"><span class="k">-</span> 把利益关系用动作说，把机制用俗语说。&#34;吃人的嘴短，拿人的手软&#34;一句话把贿赂和后果说清楚了，不需要术语。
</span></span><span class="line"><span class="cl"><span class="k">-</span> 偶尔承认自己不确定、不知道、还在想。有立场的不确定比没立场的面面俱到真实得多。
</span></span><span class="line"><span class="cl"><span class="k">-</span> 用&#34;我&#34;来叙事。有第一人称视角，有站队，有真实的判断——不是&#34;这件事有两面性&#34;的两不得罪。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">长句和短句的交替不是设计出来的。一口气能说完的想法就用一口气说完，哪怕句子长一些。该停的地方自然会停。不要为了&#34;节奏感&#34;刻意把完整的句子切碎。
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="gu">## 成功标准
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">只有一条：
</span></span><span class="line"><span class="cl"><span class="k">
</span></span></span><span class="line"><span class="cl"><span class="k">&gt; </span><span class="ge">读完之后，读者的感受应该是&#34;这个人在认真跟我说话&#34;，而不是&#34;这篇文章写得很好&#34;。
</span></span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl">如果读者能感觉到&#34;写得好&#34;，说明你还是在创作，不是在说事儿。真正有人味的文章，读完的感觉是被一个人拉着说了一通，不是读了一篇文章。
</span></span></code></pre></div></details>
]]></content:encoded>
    </item>
    <item>
      <title>What 81,000 People Told Us About the Economics of AI</title>
      <link>https://blog.beihai2000.top/en/readings/81k-economics/</link>
      <pubDate>Thu, 23 Apr 2026 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/readings/81k-economics/</guid>
      <description>A Chinese translation and commentary on Anthropic&amp;#39;s survey of 81,000 Claude users about AI&amp;#39;s economic impact.</description>
      <content:encoded><![CDATA[<p>This is a Chinese translation with commentary of the original article by Anthropic. Read the original here:</p>
<p><strong><a href="https://www.anthropic.com/research/81k-economics">What 81,000 people told us about the economics of AI</a></strong></p>
<p><em>By Maxim Massenkoff, Anthropic · April 22, 2026</em></p>
<p>For the Chinese translation and annotated version, switch to the <a href="/en/readings/81k-economics/">中文版</a>.</p>
]]></content:encoded>
    </item>
    <item>
      <title>Building a Blog with Hugo &#43; Cloudflare Pages</title>
      <link>https://blog.beihai2000.top/en/posts/build-blog-with-hugo/</link>
      <pubDate>Wed, 08 Apr 2026 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/posts/build-blog-with-hugo/</guid>
      <description>A walkthrough of building a bilingual tech blog from scratch with Hugo, PaperMod theme, and Cloudflare Pages — including the pitfalls I hit along the way.</description>
      <content:encoded><![CDATA[<p>I&rsquo;d been telling myself I&rsquo;d start a blog for months. Then one afternoon I decided to just do it — no more planning, no more comparing frameworks. A few hours later the site was live. This post is a record of that process: not a tutorial, more of an annotated changelog of mistakes and decisions.</p>
<h2 id="why-hugo">Why Hugo</h2>
<p>I didn&rsquo;t spend long choosing a static site generator. I&rsquo;d used Hexo before and knew what Node.js dependency hell feels like. Jekyll is slow. Gatsby is overkill for a blog. Hugo is a single Go binary — <code>brew install hugo</code> and you&rsquo;re done. No <code>node_modules</code>, no dependency conflicts, and builds so fast you barely notice them happening.</p>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-bash" data-lang="bash"><span class="line"><span class="cl">brew install hugo
</span></span><span class="line"><span class="cl">hugo version
</span></span><span class="line"><span class="cl"><span class="c1"># hugo v0.160.0+extended darwin/arm64</span>
</span></span></code></pre></div><p>That&rsquo;s the entire install. The <code>extended</code> version includes built-in SCSS compilation, which the theme needs.</p>
<p>For the theme, I went with <a href="https://github.com/adityatelange/hugo-PaperMod">PaperMod</a>. It&rsquo;s clean, fast, and ships with everything I need out of the box: dark mode, search, table of contents, syntax highlighting, multilingual support. Added it as a git submodule so updates are just a <code>git pull</code>:</p>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-bash" data-lang="bash"><span class="line"><span class="cl">hugo new site blog <span class="o">&amp;&amp;</span> <span class="nb">cd</span> blog
</span></span><span class="line"><span class="cl">git init
</span></span><span class="line"><span class="cl">git submodule add --depth<span class="o">=</span><span class="m">1</span> https://github.com/adityatelange/hugo-PaperMod.git themes/PaperMod
</span></span></code></pre></div><h2 id="configuration-one-yaml-file-does-most-of-the-work">Configuration: One YAML File Does Most of the Work</h2>
<p>Hugo keeps everything in a single <code>hugo.yaml</code> (also supports toml and json — I prefer yaml). My blog is bilingual: Chinese as the default language at the root path, English under <code>/en/</code>. The language config looks like this:</p>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-yaml" data-lang="yaml"><span class="line"><span class="cl"><span class="nt">defaultContentLanguage</span><span class="p">:</span><span class="w"> </span><span class="l">zh</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="nt">defaultContentLanguageInSubdir</span><span class="p">:</span><span class="w"> </span><span class="kc">false</span><span class="w">  </span><span class="c"># Chinese gets no prefix</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="nt">languages</span><span class="p">:</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">  </span><span class="nt">zh</span><span class="p">:</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">languageName</span><span class="p">:</span><span class="w"> </span><span class="s2">&#34;中文&#34;</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">weight</span><span class="p">:</span><span class="w"> </span><span class="m">1</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">title</span><span class="p">:</span><span class="w"> </span><span class="s2">&#34;Forest&#39;s Blog&#34;</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">params</span><span class="p">:</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">      </span><span class="nt">author</span><span class="p">:</span><span class="w"> </span><span class="s2">&#34;北海&#34;</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">  </span><span class="nt">en</span><span class="p">:</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">languageName</span><span class="p">:</span><span class="w"> </span><span class="s2">&#34;EN&#34;</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">weight</span><span class="p">:</span><span class="w"> </span><span class="m">2</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">title</span><span class="p">:</span><span class="w"> </span><span class="s2">&#34;Forest&#39;s Blog&#34;</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">    </span><span class="nt">params</span><span class="p">:</span><span class="w">
</span></span></span><span class="line"><span class="cl"><span class="w">      </span><span class="nt">author</span><span class="p">:</span><span class="w"> </span><span class="s2">&#34;Chunhao Zhang&#34;</span><span class="w">
</span></span></span></code></pre></div><p>Bilingual articles use filename suffixes: Chinese is <code>build-blog-with-hugo.md</code>, English is <code>build-blog-with-hugo.en.md</code>. Hugo automatically links them together for language switching.</p>
<p>Though that &ldquo;automatic linking&rdquo; had a catch — more on that below.</p>
<h2 id="deploying-to-cloudflare-pages">Deploying to Cloudflare Pages</h2>
<p>I chose Cloudflare Pages over GitHub Pages for two reasons: GitHub is unreliable from mainland China, and CF Pages has global CDN nodes with a simpler deploy workflow — just push code and it builds automatically.</p>
<p>But Cloudflare&rsquo;s Dashboard design confused me for a bit.</p>
<p>After clicking into &ldquo;Workers &amp; Pages&rdquo;, the default view shows Workers (Serverless Functions). The Pages entry is buried at the bottom of the page as an unassuming &ldquo;Looking to deploy Pages? Get started&rdquo; link. I spent a few minutes thinking Pages had been merged into Workers before I spotted it.</p>
<p>Once past that, setup was straightforward: connect GitHub repo, select Hugo preset, build command <code>hugo --minify</code>, output directory <code>public</code>. There&rsquo;s one <strong>critical environment variable</strong> you must set manually:</p>
<table>
  <thead>
      <tr>
          <th>Variable</th>
          <th>Value</th>
          <th>Why</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td><code>HUGO_VERSION</code></td>
          <td><code>0.160.0</code></td>
          <td>CF&rsquo;s default Hugo version is outdated and will break the build</td>
      </tr>
  </tbody>
</table>
<p>After deployment I got a <code>blog-6sm.pages.dev</code> domain. Custom domains can be added later, but the free subdomain works fine for now.</p>
<h2 id="the-language-switcher-bug">The Language Switcher Bug</h2>
<p>After going live, I noticed an annoying bug: clicking the language toggle (中文 ↔ EN) on any article page would redirect to the homepage instead of the translated version of that article.</p>
<p>Digging into PaperMod&rsquo;s source, I found that the default <code>header.html</code> template uses <code>site.Home.Translations</code> for the language switcher — meaning it always points to the homepage translation, not the current page&rsquo;s translation.</p>
<p>The fix was to create an override <code>header.html</code> in the project&rsquo;s <code>layouts/partials/</code> directory. Hugo&rsquo;s template loading priority is project &gt; theme, so you just copy the theme&rsquo;s header and change one line:</p>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-go-html-template" data-lang="go-html-template"><span class="line"><span class="cl"><span class="cp">{{-</span><span class="w"> </span><span class="nx">$translations</span><span class="w"> </span><span class="o">:=</span><span class="w"> </span><span class="na">.Translations</span><span class="w"> </span><span class="cp">}}</span>
</span></span><span class="line"><span class="cl"><span class="cp">{{-</span><span class="w"> </span><span class="k">if</span><span class="w"> </span><span class="k">not</span><span class="w"> </span><span class="nx">$translations</span><span class="w"> </span><span class="cp">}}</span>
</span></span><span class="line"><span class="cl">  <span class="cp">{{-</span><span class="w"> </span><span class="nx">$translations</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="nx">site</span><span class="na">.Home.Translations</span><span class="w"> </span><span class="cp">}}</span>
</span></span><span class="line"><span class="cl"><span class="cp">{{-</span><span class="w"> </span><span class="k">end</span><span class="w"> </span><span class="cp">}}</span>
</span></span></code></pre></div><p>Prefer the current page&rsquo;s <code>.Translations</code>, fall back to homepage only when there&rsquo;s no translation available. Small change, but this is genuinely a design flaw in PaperMod — on a multilingual site, switching languages on an article page and landing on the homepage feels broken.</p>
<h2 id="comments-with-giscus">Comments with Giscus</h2>
<p>For comments I went with <a href="https://giscus.app/">Giscus</a>, which is powered by GitHub Discussions. No backend needed — comments live in the repo&rsquo;s Discussions, which is the most natural setup for an open-source blog.</p>
<p>One non-obvious config choice: the Discussion category should be <strong>Announcements</strong>. This category restricts who can create new Discussions (only repo admins — corresponding to initializing a comment thread for an article), while still letting anyone reply. If you pick an open category like General, anyone can create arbitrary Discussions and things get messy.</p>
<p>Giscus also needs dark mode adaptation. I used a <code>MutationObserver</code> in <code>comments.html</code> to watch for class changes on <code>body</code> and update the Giscus iframe theme dynamically:</p>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-javascript" data-lang="javascript"><span class="line"><span class="cl"><span class="kr">const</span> <span class="nx">observer</span> <span class="o">=</span> <span class="k">new</span> <span class="nx">MutationObserver</span><span class="p">(()</span> <span class="p">=&gt;</span> <span class="p">{</span>
</span></span><span class="line"><span class="cl">    <span class="kr">const</span> <span class="nx">isDark</span> <span class="o">=</span> <span class="nb">document</span><span class="p">.</span><span class="nx">body</span><span class="p">.</span><span class="nx">classList</span><span class="p">.</span><span class="nx">contains</span><span class="p">(</span><span class="s1">&#39;dark&#39;</span><span class="p">);</span>
</span></span><span class="line"><span class="cl">    <span class="kr">const</span> <span class="nx">theme</span> <span class="o">=</span> <span class="nx">isDark</span> <span class="o">?</span> <span class="s1">&#39;noborder_dark&#39;</span> <span class="o">:</span> <span class="s1">&#39;noborder_light&#39;</span><span class="p">;</span>
</span></span><span class="line"><span class="cl">    <span class="kr">const</span> <span class="nx">iframe</span> <span class="o">=</span> <span class="nb">document</span><span class="p">.</span><span class="nx">querySelector</span><span class="p">(</span><span class="s1">&#39;iframe.giscus-frame&#39;</span><span class="p">);</span>
</span></span><span class="line"><span class="cl">    <span class="k">if</span> <span class="p">(</span><span class="nx">iframe</span><span class="p">)</span> <span class="p">{</span>
</span></span><span class="line"><span class="cl">        <span class="nx">iframe</span><span class="p">.</span><span class="nx">contentWindow</span><span class="p">.</span><span class="nx">postMessage</span><span class="p">(</span>
</span></span><span class="line"><span class="cl">            <span class="p">{</span> <span class="nx">giscus</span><span class="o">:</span> <span class="p">{</span> <span class="nx">setConfig</span><span class="o">:</span> <span class="p">{</span> <span class="nx">theme</span> <span class="p">}</span> <span class="p">}</span> <span class="p">},</span>
</span></span><span class="line"><span class="cl">            <span class="s1">&#39;https://giscus.app&#39;</span>
</span></span><span class="line"><span class="cl">        <span class="p">);</span>
</span></span><span class="line"><span class="cl">    <span class="p">}</span>
</span></span><span class="line"><span class="cl"><span class="p">});</span>
</span></span><span class="line"><span class="cl"><span class="nx">observer</span><span class="p">.</span><span class="nx">observe</span><span class="p">(</span><span class="nb">document</span><span class="p">.</span><span class="nx">body</span><span class="p">,</span> <span class="p">{</span> <span class="nx">attributes</span><span class="o">:</span> <span class="kc">true</span><span class="p">,</span> <span class="nx">attributeFilter</span><span class="o">:</span> <span class="p">[</span><span class="s1">&#39;class&#39;</span><span class="p">]</span> <span class="p">});</span>
</span></span></code></pre></div><p>This way, toggling dark mode updates the comment section in real time — no jarring white comment box on a dark page.</p>
<h2 id="miscellaneous-setup">Miscellaneous Setup</h2>
<p><strong>KaTeX math rendering</strong> is conditionally loaded. Only articles with <code>math: true</code> in frontmatter pull in the KaTeX CSS and JS, so other pages aren&rsquo;t slowed down.</p>
<p><strong>robots.txt</strong> explicitly allows AI crawlers (GPTBot, ClaudeBot, PerplexityBot). More and more sites are blocking AI bots by default, but for a technical blog that wants to be cited by both humans and AI systems, open crawling makes sense.</p>
<p><strong>Syntax highlighting</strong> uses Hugo&rsquo;s built-in Chroma engine with the <code>dracula</code> theme. PaperMod includes a code copy button out of the box.</p>
<h2 id="the-final-structure">The Final Structure</h2>
<p>Here&rsquo;s what the project looks like:</p>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-fallback" data-lang="fallback"><span class="line"><span class="cl">blog/
</span></span><span class="line"><span class="cl">├── hugo.yaml                    # Where all config lives
</span></span><span class="line"><span class="cl">├── content/
</span></span><span class="line"><span class="cl">│   ├── posts/                   # Articles
</span></span><span class="line"><span class="cl">│   ├── about.md / about.en.md   # About page
</span></span><span class="line"><span class="cl">│   ├── archives.md              # Archive
</span></span><span class="line"><span class="cl">│   └── search.md                # Search
</span></span><span class="line"><span class="cl">├── layouts/partials/
</span></span><span class="line"><span class="cl">│   ├── header.html              # Language switcher fix
</span></span><span class="line"><span class="cl">│   ├── comments.html            # Giscus comments
</span></span><span class="line"><span class="cl">│   ├── extend_head.html         # KaTeX + Schema
</span></span><span class="line"><span class="cl">│   └── math.html                # KaTeX scripts
</span></span><span class="line"><span class="cl">├── static/
</span></span><span class="line"><span class="cl">│   ├── favicon.ico              # Globe icon
</span></span><span class="line"><span class="cl">│   └── _headers                 # CF caching policy
</span></span><span class="line"><span class="cl">└── themes/PaperMod/             # git submodule
</span></span></code></pre></div><p>No bloat. The beauty of static sites is that the structure is transparent — no database, no server, no Docker. A YAML config, some Markdown files, a few HTML partials, and you have a complete website.</p>
<h2 id="looking-back">Looking Back</h2>
<p>The whole process took an afternoon. The most time-consuming parts weren&rsquo;t Hugo itself — Hugo&rsquo;s documentation is solid and the learning curve is gentle — but rather the &ldquo;should be simple but somehow isn&rsquo;t&rdquo; stuff: finding the Pages entry in Cloudflare&rsquo;s Dashboard, the language switcher jumping to homepage, picking the wrong Giscus category.</p>
<p>None of these were hard problems. But they all required actually hitting the wall, spending time debugging, and figuring out the fix. Writing it down here so I don&rsquo;t step on the same mines twice.</p>
<p>Blog&rsquo;s up. Time to write some real content.</p>
]]></content:encoded>
    </item>
    <item>
      <title>Hello, World</title>
      <link>https://blog.beihai2000.top/en/posts/hello-world/</link>
      <pubDate>Wed, 08 Apr 2026 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/posts/hello-world/</guid>
      <description>My first blog post, introducing the content direction and tech stack of this blog.</description>
      <content:encoded><![CDATA[<h2 id="welcome">Welcome</h2>
<p>Welcome to my blog! I will share technical articles about AI, LLMs, and software engineering here.</p>
<h2 id="content-direction">Content Direction</h2>
<p>This blog will cover the following areas:</p>
<ul>
<li><strong>Deep Technical Articles</strong>: In-depth analysis of key technologies in AI/LLM</li>
<li><strong>Industry Insights</strong>: Tracking the latest developments and trends in AI</li>
<li><strong>Engineering Practices</strong>: Sharing experiences and best practices in software engineering</li>
</ul>
<h2 id="tech-stack">Tech Stack</h2>
<p>This blog is built with:</p>
<ul>
<li><a href="https://gohugo.io/">Hugo</a> — High-performance static site generator</li>
<li><a href="https://github.com/adityatelange/hugo-PaperMod">PaperMod</a> — Clean and elegant Hugo theme</li>
<li><a href="https://pages.cloudflare.com/">Cloudflare Pages</a> — Global CDN deployment</li>
</ul>
<div class="highlight"><pre tabindex="0" class="chroma"><code class="language-python" data-lang="python"><span class="line"><span class="cl"><span class="c1"># A sample code snippet</span>
</span></span><span class="line"><span class="cl"><span class="k">def</span> <span class="nf">hello</span><span class="p">():</span>
</span></span><span class="line"><span class="cl">    <span class="nb">print</span><span class="p">(</span><span class="s2">&#34;Hello, World!&#34;</span><span class="p">)</span>
</span></span><span class="line"><span class="cl">
</span></span><span class="line"><span class="cl"><span class="n">hello</span><span class="p">()</span>
</span></span></code></pre></div><h2 id="next-steps">Next Steps</h2>
<p>Stay tuned for more content!</p>
]]></content:encoded>
    </item>
    <item>
      <title>About</title>
      <link>https://blog.beihai2000.top/en/about/</link>
      <pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate>
      <guid>https://blog.beihai2000.top/en/about/</guid>
      <description>about</description>
      <content:encoded><![CDATA[<h2 id="about-me">About Me</h2>
<p>This is a placeholder for the about page. More details coming soon.</p>
<h3 id="contact">Contact</h3>
<ul>
<li>GitHub: <a href="https://github.com/hahhforest">hahhforest</a></li>
</ul>
]]></content:encoded>
    </item>
  </channel>
</rss>
