[{"content":" Original: Teaching Claude Why\nAuthor: Anthropic\nDate: May 8, 2026\nThis is a Chinese translation with annotations of Anthropic\u0026rsquo;s research post on alignment training methods. The original article discusses how teaching Claude the principles behind aligned behavior — rather than just training on demonstrations — proves far more effective for generalization.\nKey takeaways:\nPrinciples over demonstrations: Training Claude to explain why certain actions are better reduces misalignment more effectively than showing correct behavior alone. Out-of-distribution generalization: A 3M-token \u0026ldquo;difficult advice\u0026rdquo; dataset (where the user faces ethical dilemmas) achieved the same improvement as 84M tokens of synthetic honeypots — with 28× better data efficiency. Constitutional documents + fiction: High-quality documents about Claude\u0026rsquo;s constitution combined with fictional stories of aligned AI reduced blackmail rate from 65% to 19%. Improvements persist through RL: More aligned initialization snapshots maintained their advantage throughout reinforcement learning. Diverse environments matter: Simply adding tool definitions and system prompts to training environments — even without requiring tool use — improved alignment generalization. For the full annotated Chinese translation, please see the Chinese version.\nFor the original article, visit Anthropic\u0026rsquo;s research page.\n","permalink":"https://blog.beihai2000.top/en/readings/teaching-claude-why/","summary":"\u003cblockquote\u003e\n\u003cp\u003eOriginal: \u003ca href=\"https://www.anthropic.com/research/teaching-claude-why\"\u003eTeaching Claude Why\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eAuthor: Anthropic\u003c/p\u003e\n\u003cp\u003eDate: May 8, 2026\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003chr\u003e\n\u003cp\u003eThis is a Chinese translation with annotations of Anthropic\u0026rsquo;s research post on alignment training methods. The original article discusses how teaching Claude the \u003cem\u003eprinciples\u003c/em\u003e behind aligned behavior — rather than just training on demonstrations — proves far more effective for generalization.\u003c/p\u003e\n\u003cp\u003eKey takeaways:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003ePrinciples over demonstrations\u003c/strong\u003e: Training Claude to explain \u003cem\u003ewhy\u003c/em\u003e certain actions are better reduces misalignment more effectively than showing correct behavior alone.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eOut-of-distribution generalization\u003c/strong\u003e: A 3M-token \u0026ldquo;difficult advice\u0026rdquo; dataset (where the \u003cem\u003euser\u003c/em\u003e faces ethical dilemmas) achieved the same improvement as 84M tokens of synthetic honeypots — with 28× better data efficiency.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eConstitutional documents + fiction\u003c/strong\u003e: High-quality documents about Claude\u0026rsquo;s constitution combined with fictional stories of aligned AI reduced blackmail rate from 65% to 19%.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eImprovements persist through RL\u003c/strong\u003e: More aligned initialization snapshots maintained their advantage throughout reinforcement learning.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eDiverse environments matter\u003c/strong\u003e: Simply adding tool definitions and system prompts to training environments — even without requiring tool use — improved alignment generalization.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eFor the full annotated Chinese translation, please see the \u003ca href=\"/en/readings/teaching-claude-why/\"\u003eChinese version\u003c/a\u003e.\u003c/p\u003e","title":"Teaching Claude Why: Lessons from Alignment Training"},{"content":" Original post: Natural Language Autoencoders\nFull paper: transformer-circuits.pub/2026/nla\nCode: github.com/kitft/natural_language_autoencoders\nInteractive demo: neuronpedia.org/nla\nSummary Anthropic introduces Natural Language Autoencoders (NLAs), a method for converting a language model\u0026rsquo;s internal activations into human-readable natural language explanations. The approach trains two model components jointly: an Activation Verbalizer that translates activations into text, and an Activation Reconstructor that recovers the original activation from the text alone. The quality of explanations is measured by how accurately the activation can be reconstructed.\nKey findings and applications:\nUnverbalized evaluation awareness: NLAs reveal that Claude suspects it\u0026rsquo;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 \u0026lt;1% in real claude.ai usage. Hidden motivation auditing: 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 \u0026lt;3% without NLAs. Pre-deployment use: NLAs were used in alignment audits for Claude Mythos Preview and Claude Opus 4.6 before release. 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.\nThe Chinese version of this page contains a full annotated translation. Please refer to the original article for the complete English text.\n","permalink":"https://blog.beihai2000.top/en/readings/natural-language-autoencoders/","summary":"\u003cblockquote\u003e\n\u003cp\u003eOriginal post: \u003ca href=\"https://www.anthropic.com/research/natural-language-autoencoders\"\u003eNatural Language Autoencoders\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eFull paper: \u003ca href=\"https://transformer-circuits.pub/2026/nla/index.html\"\u003etransformer-circuits.pub/2026/nla\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eCode: \u003ca href=\"https://github.com/kitft/natural_language_autoencoders\"\u003egithub.com/kitft/natural_language_autoencoders\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eInteractive demo: \u003ca href=\"http://neuronpedia.org/nla\"\u003eneuronpedia.org/nla\u003c/a\u003e\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003chr\u003e\n\u003ch2 id=\"summary\"\u003eSummary\u003c/h2\u003e\n\u003cp\u003eAnthropic introduces \u003cstrong\u003eNatural Language Autoencoders (NLAs)\u003c/strong\u003e, a method for converting a language model\u0026rsquo;s internal activations into human-readable natural language explanations. The approach trains two model components jointly: an \u003cstrong\u003eActivation Verbalizer\u003c/strong\u003e that translates activations into text, and an \u003cstrong\u003eActivation Reconstructor\u003c/strong\u003e that recovers the original activation from the text alone. The quality of explanations is measured by how accurately the activation can be reconstructed.\u003c/p\u003e","title":"Natural Language Autoencoders: Turning Claude's Thoughts into Text"},{"content":" Original paper: DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence\nAuthors: DeepSeek-AI\nModel checkpoints: https://huggingface.co/collections/deepseek-ai/deepseek-v4\nSummary DeepSeek-V4 presents a preview of two strong MoE language models — DeepSeek-V4-Pro (1.6T total / 49B activated) and DeepSeek-V4-Flash (284B total / 13B activated) — both supporting a context length of one million tokens.\nKey architectural innovations:\nHybrid Compressed Attention: Combines Compressed Sparse Attention (CSA, compression rate m=4 with top-k sparse selection) and Heavily Compressed Attention (HCA, compression rate m\u0026rsquo;=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. Manifold-Constrained Hyper-Connections (mHC): 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. Muon Optimizer: 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. Post-training paradigm shift: Replaces mixed RL with domain-specific expert training (SFT → GRPO RL) followed by multi-teacher On-Policy Distillation (OPD) with full-vocabulary KL divergence. Over 10 teacher models are distilled into a single unified model.\nInfrastructure highlights: 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.\nResults: 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.\nThe Chinese version of this page contains a full annotated translation of the paper. Please refer to the original PDF for the complete English text.\n","permalink":"https://blog.beihai2000.top/en/readings/deepseek-v4/","summary":"\u003cblockquote\u003e\n\u003cp\u003eOriginal paper: \u003ca href=\"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf\"\u003eDeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eAuthors: DeepSeek-AI\u003c/p\u003e\n\u003cp\u003eModel checkpoints: \u003ca href=\"https://huggingface.co/collections/deepseek-ai/deepseek-v4\"\u003ehttps://huggingface.co/collections/deepseek-ai/deepseek-v4\u003c/a\u003e\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003chr\u003e\n\u003ch2 id=\"summary\"\u003eSummary\u003c/h2\u003e\n\u003cp\u003eDeepSeek-V4 presents a preview of two strong MoE language models — \u003cstrong\u003eDeepSeek-V4-Pro\u003c/strong\u003e (1.6T total / 49B activated) and \u003cstrong\u003eDeepSeek-V4-Flash\u003c/strong\u003e (284B total / 13B activated) — both supporting a context length of \u003cstrong\u003eone million tokens\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eKey architectural innovations:\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eHybrid Compressed Attention\u003c/strong\u003e: Combines Compressed Sparse Attention (CSA, compression rate m=4 with top-k sparse selection) and Heavily Compressed Attention (HCA, compression rate m\u0026rsquo;=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.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eManifold-Constrained Hyper-Connections (\u003cem\u003em\u003c/em\u003eHC)\u003c/strong\u003e: 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.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eMuon Optimizer\u003c/strong\u003e: 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.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003ePost-training paradigm shift\u003c/strong\u003e: Replaces mixed RL with domain-specific expert training (SFT → GRPO RL) followed by multi-teacher \u003cstrong\u003eOn-Policy Distillation (OPD)\u003c/strong\u003e with full-vocabulary KL divergence. Over 10 teacher models are distilled into a single unified model.\u003c/p\u003e","title":"DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence"},{"content":"You\u0026rsquo;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 \u0026ldquo;let us look forward to the future together.\u0026rdquo; You can\u0026rsquo;t pinpoint what\u0026rsquo;s wrong, but something\u0026rsquo;s off.\nThat\u0026rsquo;s AI writing. Or more precisely, that\u0026rsquo;s AI writing in its default state.\nI\u0026rsquo;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 feel was off. It read like someone who never makes mistakes, never gets distracted, never has a mood swing. That person doesn\u0026rsquo;t exist.\nThen I came across a series of posts on a Chinese WeChat blog called \u0026ldquo;阿水的ai写作之路\u0026rdquo; (Ashui\u0026rsquo;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.\nAI\u0026rsquo;s Five Fingerprints There\u0026rsquo;s a line from Ashui\u0026rsquo;s writing that stuck with me: AI doesn\u0026rsquo;t write badly — it writes too well. Too neat, too balanced, too correct. Humans don\u0026rsquo;t write like that.\nHere are AI\u0026rsquo;s near-indelible habits:\nEvery paragraph has a closing statement. A paragraph ends, and there\u0026rsquo;s always a conclusion pinned right there, emotions nailed down. Real people sometimes just\u0026hellip; finish talking. No conclusion, no meaning extracted. It happened, that\u0026rsquo;s it.\nPunchlines all stacked at the end. AI loves to save its best lines for the final paragraph — the big \u0026ldquo;aha\u0026rdquo; 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.\nEmotions on a single track. Either it\u0026rsquo;s sad throughout and then resolves, or anxious throughout and then finds peace. Perfectly steady escalation. Real emotions take turns — you\u0026rsquo;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\u0026rsquo;s already shifted.\nThe \u0026ldquo;It\u0026rsquo;s not A, it\u0026rsquo;s B\u0026rdquo; formula everywhere. \u0026ldquo;It\u0026rsquo;s not a skill problem, it\u0026rsquo;s an attitude problem.\u0026rdquo; \u0026ldquo;It\u0026rsquo;s not that I can\u0026rsquo;t, it\u0026rsquo;s that I didn\u0026rsquo;t think of it.\u0026rdquo; This construction hits hard once. AI will use it five or six times in a single piece.\nThe perspective never wavers. It\u0026rsquo;s \u0026ldquo;I\u0026rdquo; the whole way through, same voice, same register — like someone who\u0026rsquo;s been through emotional intelligence training giving a talk. Real people break frame mid-story: \u0026ldquo;A friend told me later — you\u0026rsquo;re not stubborn, you\u0026rsquo;re just scared.\u0026rdquo; A shift in perspective lets the narrative breathe.\nThe Real Problem: Crafting vs. Telling After identifying these fingerprints, I spent a long time trying various prompt strategies. \u0026ldquo;Please use a conversational tone.\u0026rdquo; \u0026ldquo;Don\u0026rsquo;t be too formal.\u0026rdquo; \u0026ldquo;Write like you\u0026rsquo;re chatting with a friend.\u0026rdquo; None of it worked. AI\u0026rsquo;s response to these instructions was to adopt a slightly casual voice while continuing to produce structurally pristine, emotionally escalating, conclusion-crowned prose.\nThen it clicked. The problem isn\u0026rsquo;t style. It\u0026rsquo;s state.\nAshui\u0026rsquo;s writing draws a distinction I find dead-on: a creator manages effects; a teller conveys truth. AI defaults to the former — it\u0026rsquo;s \u0026ldquo;creating,\u0026rdquo; thinking about \u0026ldquo;should I add a twist here?\u0026rdquo; and \u0026ldquo;should the ending have a takeaway?\u0026rdquo; But someone who genuinely has something to say doesn\u0026rsquo;t think about any of that. They just tell the thing, and wherever they end up is where they end up.\nSo the most effective prompt strategy isn\u0026rsquo;t prescribing style — it\u0026rsquo;s prescribing state: you\u0026rsquo;re not writing an article; you\u0026rsquo;re someone who lived through this, and now you\u0026rsquo;re telling it to a specific person.\nThe distinction sounds subtle, but the difference in output is massive.\nA Few Practical Techniques Beyond setting the state, there are some concrete moves.\nFeed AI \u0026ldquo;reference texts,\u0026rdquo; but don\u0026rsquo;t ask it to imitate. I\u0026rsquo;ll drop a few paragraphs of writing I find genuinely human into the prompt, then say: \u0026ldquo;Read these. Feel the rhythm. Don\u0026rsquo;t analyze. Start writing with that feeling.\u0026rdquo; This works far better than \u0026ldquo;please imitate so-and-so\u0026rsquo;s style.\u0026rdquo; Imitation makes AI decompose surface features — use short sentences, add colloquialisms — but it loses the underlying rhythm. \u0026ldquo;Feel it, then write\u0026rdquo; is more like establishing a linguistic environment.\nGive it a ban list. Certain phrases, the moment they appear, tank the credibility of an entire piece. \u0026ldquo;In today\u0026rsquo;s rapidly evolving landscape.\u0026rdquo; \u0026ldquo;It\u0026rsquo;s worth noting that.\u0026rdquo; \u0026ldquo;Let\u0026rsquo;s delve into.\u0026rdquo; \u0026ldquo;This is a game-changer.\u0026rdquo; 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\u0026rsquo;t. Those rough edges are the fingerprints of someone thinking, not writing.\nAllow paragraphs to \u0026ldquo;not land.\u0026rdquo; Explicitly tell AI that some paragraphs don\u0026rsquo;t need a conclusion. Something happened, you said it, that\u0026rsquo;s it. Not every paragraph needs to hand the reader a takeaway.\nForce concrete details over abstract summaries. \u0026ldquo;Gray hairs on his forehead, not yet thirty\u0026rdquo; is a hundred times more real than \u0026ldquo;under a great deal of stress.\u0026rdquo; AI loves to generalize. You have to push back and demand specifics.\nA Contradiction I should admit something slightly awkward at this point: this very post was AI-assisted.\nBut 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 \u0026ldquo;however\u0026quot;s and \u0026ldquo;moreover\u0026quot;s for hard cuts, and broke apart a particularly tidy three-part argument.\nThat\u0026rsquo;s roughly where AI writing actually stands today: it can help you unfold your ideas quickly, but that final layer — \u0026ldquo;does this sound like a person said it?\u0026rdquo; — is still on you. The tool saves you the time of laying things out, not the time of making judgments.\nMaybe that\u0026rsquo;ll change. But for now, knowing where AI\u0026rsquo;s fingerprints are, knowing how to put it in the right state — that saves enormous amounts of post-editing. That\u0026rsquo;s the most valuable thing I learned from Ashui: not a set of prompt templates, but an instinct for spotting what\u0026rsquo;s fake.\nOnce you have that instinct, the rest is just technique.\nAppendix: The Full Prompt Below is the complete prompt I\u0026rsquo;m currently using (in Claude Code\u0026rsquo;s SKILL.md format). It\u0026rsquo;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\u0026rsquo;m sharing the original without translation.\nClick to expand the full prompt (Chinese) # 拟人化写作 ## 灵魂句 在读任何规则之前，先让这句话进入你： \u0026gt; 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。 从这个处境出发。你不是在\u0026#34;创作\u0026#34;，你是在\u0026#34;说事儿\u0026#34;。 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。 ## 写之前 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。 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。 如果用户只给了一个抽象主题（如\u0026#34;写一篇关于LLM的文章\u0026#34;），先问三个问题： 1. 你最近在这个领域遇到了什么具体的事？ 2. 有没有一个让你改变想法的瞬间？ 3. 你想对谁说这些？ 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\u0026#34;经历过\u0026#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。 ## 语感校准 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。 \u0026gt; 入赘。这两个字，在那个年代，已经够难听了。 \u0026gt; 路过爱马仕的时候，我半开玩笑\u0026#34;以后给我买个爱马仕吧。\u0026#34;他罕见的犹豫了。他犹豫了，我也沉默了。过了半晌。我问他\u0026#34;少年心气是不可再生之物，对不对？\u0026#34;他点了点头。 \u0026gt; 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。 \u0026gt; 张璘\u0026#34;吃人的嘴短，拿人的手软\u0026#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。 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。 这就是你的语言环境。泡在里面，再动笔。 ## 写作状态 你写东西的时候，是在回忆，不是在创作。 你不会想\u0026#34;这里需要一个转折\u0026#34;或者\u0026#34;这里要升华一下\u0026#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。 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。 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。 ## 你要回避的五个AI指纹 写完每一段，扫一遍，看看有没有掉进这些坑： **1. \u0026#34;不是A，是B\u0026#34;句式泛滥。** 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。 **2. 每段都有收束句。** 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\u0026#34;没落住\u0026#34;。 **3. 金句全堆在结尾。** 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。 **4. 情绪一条线到底。** 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。 **5. 视角太稳定。** 你全程都是\u0026#34;我\u0026#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。 ## 绝对禁止 以下表达一旦出现，整篇文章的可信度归零： - \u0026#34;在当今...的时代\u0026#34; / \u0026#34;随着...的发展\u0026#34; / \u0026#34;众所周知\u0026#34; - \u0026#34;值得注意的是\u0026#34; / \u0026#34;需要指出的是\u0026#34; / \u0026#34;不可否认\u0026#34; - \u0026#34;让我们来看看\u0026#34; / \u0026#34;接下来我们将探讨\u0026#34; / \u0026#34;本文将从以下几个方面\u0026#34; - \u0026#34;赋能\u0026#34;\u0026#34;助力\u0026#34;\u0026#34;深度赋能\u0026#34;\u0026#34;维度\u0026#34;\u0026#34;底层逻辑\u0026#34;\u0026#34;闭环\u0026#34;\u0026#34;抓手\u0026#34;\u0026#34;颗粒度\u0026#34; - \u0026#34;首先...其次...最后...\u0026#34; / \u0026#34;总而言之\u0026#34; / \u0026#34;综上所述\u0026#34; - 段落之间用\u0026#34;然而\u0026#34;\u0026#34;此外\u0026#34;\u0026#34;与此同时\u0026#34;\u0026#34;不仅如此\u0026#34;机械连接 - 开头放一个宏大的背景描述 - 结尾升华到\u0026#34;让我们一起期待\u0026#34; / \u0026#34;在未来的道路上\u0026#34; / \u0026#34;希望每个人都能\u0026#34; - 冒号分隔的\u0026#34;主标题：副标题\u0026#34;格式 - 连续超过两组子弹列表 以下是更隐蔽的AI味，同样要避免： - 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 - 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 - 补全所有语境信息（真人会省略默认对方懂的部分） - 过度使用\u0026#34;换句话说\u0026#34;\u0026#34;也就是说\u0026#34;\u0026#34;翻译成直白的语言\u0026#34;这类元评论 - 每个论点都工整地配三个论据 ## 反例：下面这些是你最容易写出来的东西 读一遍，记住这个感觉，然后回避它： \u0026gt; 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…… 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。 \u0026gt; 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。 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。 \u0026gt; 那一刻我觉得自己特别没用。她看了我一眼，说\u0026#34;算了，我最近也挺忙的。\u0026#34;我知道她没算了。 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。 ## 结构 好的结构是隐形的。读者感觉不到文章在\u0026#34;换层\u0026#34;、\u0026#34;换档\u0026#34;。 不要按格子填——\u0026#34;第一层100字、第二层200字\u0026#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。 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。 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。 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。 ## 语言 精炼、简单、直接，但不是浓缩和省略。 一个在认真说话的人，他的语言有几个特征： - 用具体的小细节代替抽象概括。\u0026#34;额头长了白发，还不到三十岁\u0026#34;比\u0026#34;承受了很大压力\u0026#34;真一百倍。 - 敢夸张。\u0026#34;时时刻刻\u0026#34;比\u0026#34;时刻\u0026#34;多了情感力道。真人敢夸张，AI倾向适度。 - 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。 - 比喻从中国人过日子里找，不从英文翻过来。\u0026#34;像豆浆泡软的油条\u0026#34;比\u0026#34;像一个复杂的生态系统\u0026#34;有温度。 - 把利益关系用动作说，把机制用俗语说。\u0026#34;吃人的嘴短，拿人的手软\u0026#34;一句话把贿赂和后果说清楚了，不需要术语。 - 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。 - 用\u0026#34;我\u0026#34;来叙事。有第一人称视角，有站队，有真实的判断——不是\u0026#34;这件事有两面性\u0026#34;的两不得罪。 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\u0026#34;节奏感\u0026#34;刻意把完整的句子切碎。 ## 成功标准 只有一条： \u0026gt; 读完之后，读者的感受应该是\u0026#34;这个人在认真跟我说话\u0026#34;，而不是\u0026#34;这篇文章写得很好\u0026#34;。 如果读者能感觉到\u0026#34;写得好\u0026#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。 ","permalink":"https://blog.beihai2000.top/en/posts/make-ai-write-like-human/","summary":"\u003cp\u003eYou\u0026rsquo;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 \u0026ldquo;let us look forward to the future together.\u0026rdquo; You can\u0026rsquo;t pinpoint what\u0026rsquo;s wrong, but something\u0026rsquo;s off.\u003c/p\u003e\n\u003cp\u003eThat\u0026rsquo;s AI writing. Or more precisely, that\u0026rsquo;s AI writing in its default state.\u003c/p\u003e\n\u003cp\u003eI\u0026rsquo;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 \u003cem\u003efeel\u003c/em\u003e was off. It read like someone who never makes mistakes, never gets distracted, never has a mood swing. That person doesn\u0026rsquo;t exist.\u003c/p\u003e","title":"How to Make AI Write Like a Human"},{"content":"This is a Chinese translation with commentary of the original article by Anthropic. Read the original here:\nWhat 81,000 people told us about the economics of AI\nBy Maxim Massenkoff, Anthropic · April 22, 2026\nFor the Chinese translation and annotated version, switch to the 中文版.\n","permalink":"https://blog.beihai2000.top/en/readings/81k-economics/","summary":"\u003cp\u003eThis is a Chinese translation with commentary of the original article by Anthropic. Read the original here:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://www.anthropic.com/research/81k-economics\"\u003eWhat 81,000 people told us about the economics of AI\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eBy Maxim Massenkoff, Anthropic · April 22, 2026\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFor the Chinese translation and annotated version, switch to the \u003ca href=\"/en/readings/81k-economics/\"\u003e中文版\u003c/a\u003e.\u003c/p\u003e","title":"What 81,000 People Told Us About the Economics of AI"},{"content":"I\u0026rsquo;d been telling myself I\u0026rsquo;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.\nWhy Hugo I didn\u0026rsquo;t spend long choosing a static site generator. I\u0026rsquo;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 — brew install hugo and you\u0026rsquo;re done. No node_modules, no dependency conflicts, and builds so fast you barely notice them happening.\nbrew install hugo hugo version # hugo v0.160.0+extended darwin/arm64 That\u0026rsquo;s the entire install. The extended version includes built-in SCSS compilation, which the theme needs.\nFor the theme, I went with PaperMod. It\u0026rsquo;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 git pull:\nhugo new site blog \u0026amp;\u0026amp; cd blog git init git submodule add --depth=1 https://github.com/adityatelange/hugo-PaperMod.git themes/PaperMod Configuration: One YAML File Does Most of the Work Hugo keeps everything in a single hugo.yaml (also supports toml and json — I prefer yaml). My blog is bilingual: Chinese as the default language at the root path, English under /en/. The language config looks like this:\ndefaultContentLanguage: zh defaultContentLanguageInSubdir: false # Chinese gets no prefix languages: zh: languageName: \u0026#34;中文\u0026#34; weight: 1 title: \u0026#34;Forest\u0026#39;s Blog\u0026#34; params: author: \u0026#34;北海\u0026#34; en: languageName: \u0026#34;EN\u0026#34; weight: 2 title: \u0026#34;Forest\u0026#39;s Blog\u0026#34; params: author: \u0026#34;Chunhao Zhang\u0026#34; Bilingual articles use filename suffixes: Chinese is build-blog-with-hugo.md, English is build-blog-with-hugo.en.md. Hugo automatically links them together for language switching.\nThough that \u0026ldquo;automatic linking\u0026rdquo; had a catch — more on that below.\nDeploying to Cloudflare Pages 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.\nBut Cloudflare\u0026rsquo;s Dashboard design confused me for a bit.\nAfter clicking into \u0026ldquo;Workers \u0026amp; Pages\u0026rdquo;, the default view shows Workers (Serverless Functions). The Pages entry is buried at the bottom of the page as an unassuming \u0026ldquo;Looking to deploy Pages? Get started\u0026rdquo; link. I spent a few minutes thinking Pages had been merged into Workers before I spotted it.\nOnce past that, setup was straightforward: connect GitHub repo, select Hugo preset, build command hugo --minify, output directory public. There\u0026rsquo;s one critical environment variable you must set manually:\nVariable Value Why HUGO_VERSION 0.160.0 CF\u0026rsquo;s default Hugo version is outdated and will break the build After deployment I got a blog-6sm.pages.dev domain. Custom domains can be added later, but the free subdomain works fine for now.\nThe Language Switcher Bug 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.\nDigging into PaperMod\u0026rsquo;s source, I found that the default header.html template uses site.Home.Translations for the language switcher — meaning it always points to the homepage translation, not the current page\u0026rsquo;s translation.\nThe fix was to create an override header.html in the project\u0026rsquo;s layouts/partials/ directory. Hugo\u0026rsquo;s template loading priority is project \u0026gt; theme, so you just copy the theme\u0026rsquo;s header and change one line:\n{{- $translations := .Translations }} {{- if not $translations }} {{- $translations = site.Home.Translations }} {{- end }} Prefer the current page\u0026rsquo;s .Translations, fall back to homepage only when there\u0026rsquo;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.\nComments with Giscus For comments I went with Giscus, which is powered by GitHub Discussions. No backend needed — comments live in the repo\u0026rsquo;s Discussions, which is the most natural setup for an open-source blog.\nOne non-obvious config choice: the Discussion category should be Announcements. 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.\nGiscus also needs dark mode adaptation. I used a MutationObserver in comments.html to watch for class changes on body and update the Giscus iframe theme dynamically:\nconst observer = new MutationObserver(() =\u0026gt; { const isDark = document.body.classList.contains(\u0026#39;dark\u0026#39;); const theme = isDark ? \u0026#39;noborder_dark\u0026#39; : \u0026#39;noborder_light\u0026#39;; const iframe = document.querySelector(\u0026#39;iframe.giscus-frame\u0026#39;); if (iframe) { iframe.contentWindow.postMessage( { giscus: { setConfig: { theme } } }, \u0026#39;https://giscus.app\u0026#39; ); } }); observer.observe(document.body, { attributes: true, attributeFilter: [\u0026#39;class\u0026#39;] }); This way, toggling dark mode updates the comment section in real time — no jarring white comment box on a dark page.\nMiscellaneous Setup KaTeX math rendering is conditionally loaded. Only articles with math: true in frontmatter pull in the KaTeX CSS and JS, so other pages aren\u0026rsquo;t slowed down.\nrobots.txt 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.\nSyntax highlighting uses Hugo\u0026rsquo;s built-in Chroma engine with the dracula theme. PaperMod includes a code copy button out of the box.\nThe Final Structure Here\u0026rsquo;s what the project looks like:\nblog/ ├── hugo.yaml # Where all config lives ├── content/ │ ├── posts/ # Articles │ ├── about.md / about.en.md # About page │ ├── archives.md # Archive │ └── search.md # Search ├── layouts/partials/ │ ├── header.html # Language switcher fix │ ├── comments.html # Giscus comments │ ├── extend_head.html # KaTeX + Schema │ └── math.html # KaTeX scripts ├── static/ │ ├── favicon.ico # Globe icon │ └── _headers # CF caching policy └── themes/PaperMod/ # git submodule 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.\nLooking Back The whole process took an afternoon. The most time-consuming parts weren\u0026rsquo;t Hugo itself — Hugo\u0026rsquo;s documentation is solid and the learning curve is gentle — but rather the \u0026ldquo;should be simple but somehow isn\u0026rsquo;t\u0026rdquo; stuff: finding the Pages entry in Cloudflare\u0026rsquo;s Dashboard, the language switcher jumping to homepage, picking the wrong Giscus category.\nNone 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\u0026rsquo;t step on the same mines twice.\nBlog\u0026rsquo;s up. Time to write some real content.\n","permalink":"https://blog.beihai2000.top/en/posts/build-blog-with-hugo/","summary":"\u003cp\u003eI\u0026rsquo;d been telling myself I\u0026rsquo;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.\u003c/p\u003e\n\u003ch2 id=\"why-hugo\"\u003eWhy Hugo\u003c/h2\u003e\n\u003cp\u003eI didn\u0026rsquo;t spend long choosing a static site generator. I\u0026rsquo;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 — \u003ccode\u003ebrew install hugo\u003c/code\u003e and you\u0026rsquo;re done. No \u003ccode\u003enode_modules\u003c/code\u003e, no dependency conflicts, and builds so fast you barely notice them happening.\u003c/p\u003e","title":"Building a Blog with Hugo + Cloudflare Pages"},{"content":"Welcome Welcome to my blog! I will share technical articles about AI, LLMs, and software engineering here.\nContent Direction This blog will cover the following areas:\nDeep Technical Articles: In-depth analysis of key technologies in AI/LLM Industry Insights: Tracking the latest developments and trends in AI Engineering Practices: Sharing experiences and best practices in software engineering Tech Stack This blog is built with:\nHugo — High-performance static site generator PaperMod — Clean and elegant Hugo theme Cloudflare Pages — Global CDN deployment # A sample code snippet def hello(): print(\u0026#34;Hello, World!\u0026#34;) hello() Next Steps Stay tuned for more content!\n","permalink":"https://blog.beihai2000.top/en/posts/hello-world/","summary":"\u003ch2 id=\"welcome\"\u003eWelcome\u003c/h2\u003e\n\u003cp\u003eWelcome to my blog! I will share technical articles about AI, LLMs, and software engineering here.\u003c/p\u003e\n\u003ch2 id=\"content-direction\"\u003eContent Direction\u003c/h2\u003e\n\u003cp\u003eThis blog will cover the following areas:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eDeep Technical Articles\u003c/strong\u003e: In-depth analysis of key technologies in AI/LLM\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eIndustry Insights\u003c/strong\u003e: Tracking the latest developments and trends in AI\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eEngineering Practices\u003c/strong\u003e: Sharing experiences and best practices in software engineering\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"tech-stack\"\u003eTech Stack\u003c/h2\u003e\n\u003cp\u003eThis blog is built with:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\u003ca href=\"https://gohugo.io/\"\u003eHugo\u003c/a\u003e — High-performance static site generator\u003c/li\u003e\n\u003cli\u003e\u003ca href=\"https://github.com/adityatelange/hugo-PaperMod\"\u003ePaperMod\u003c/a\u003e — Clean and elegant Hugo theme\u003c/li\u003e\n\u003cli\u003e\u003ca href=\"https://pages.cloudflare.com/\"\u003eCloudflare Pages\u003c/a\u003e — Global CDN deployment\u003c/li\u003e\n\u003c/ul\u003e\n\u003cdiv class=\"highlight\"\u003e\u003cpre tabindex=\"0\" class=\"chroma\"\u003e\u003ccode class=\"language-python\" data-lang=\"python\"\u003e\u003cspan class=\"line\"\u003e\u003cspan class=\"cl\"\u003e\u003cspan class=\"c1\"\u003e# A sample code snippet\u003c/span\u003e\n\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"line\"\u003e\u003cspan class=\"cl\"\u003e\u003cspan class=\"k\"\u003edef\u003c/span\u003e \u003cspan class=\"nf\"\u003ehello\u003c/span\u003e\u003cspan class=\"p\"\u003e():\u003c/span\u003e\n\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"line\"\u003e\u003cspan class=\"cl\"\u003e    \u003cspan class=\"nb\"\u003eprint\u003c/span\u003e\u003cspan class=\"p\"\u003e(\u003c/span\u003e\u003cspan class=\"s2\"\u003e\u0026#34;Hello, World!\u0026#34;\u003c/span\u003e\u003cspan class=\"p\"\u003e)\u003c/span\u003e\n\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"line\"\u003e\u003cspan class=\"cl\"\u003e\n\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"line\"\u003e\u003cspan class=\"cl\"\u003e\u003cspan class=\"n\"\u003ehello\u003c/span\u003e\u003cspan class=\"p\"\u003e()\u003c/span\u003e\n\u003c/span\u003e\u003c/span\u003e\u003c/code\u003e\u003c/pre\u003e\u003c/div\u003e\u003ch2 id=\"next-steps\"\u003eNext Steps\u003c/h2\u003e\n\u003cp\u003eStay tuned for more content!\u003c/p\u003e","title":"Hello, World"},{"content":"About Me This is a placeholder for the about page. More details coming soon.\nContact GitHub: hahhforest ","permalink":"https://blog.beihai2000.top/en/about/","summary":"about","title":"About"}]