<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sha1rholder on Medium]]></title>
        <description><![CDATA[Stories by Sha1rholder on Medium]]></description>
        <link>https://medium.com/@sha1rholder?source=rss-4c891e377534------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*ztJaU7udjjNtJvkt</url>
            <title>Stories by Sha1rholder on Medium</title>
            <link>https://medium.com/@sha1rholder?source=rss-4c891e377534------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 18:36:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sha1rholder/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[A failed attempt on “HTML as PowerPoint in AI era”]]></title>
            <link>https://medium.com/@sha1rholder/a-failed-attempt-on-html-as-powerpoint-in-ai-era-2bfa18e33a71?source=rss-4c891e377534------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bfa18e33a71</guid>
            <category><![CDATA[aritificial-intelligence]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[html]]></category>
            <dc:creator><![CDATA[Sha1rholder]]></dc:creator>
            <pubDate>Tue, 14 Apr 2026 14:20:41 GMT</pubDate>
            <atom:updated>2026-04-14T14:20:41.840Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<h3>The debate on “HTML as PowerPoint in AI era” reminds me of my attempt to replace slides with HTML during my internship one year ago, which taught me a hard lesson.</h3><p>Saw a post on RedNote (小红书): <a href="http://xhslink.com/o/6x5zcPDxZDB">http://xhslink.com/o/6x5zcPDxZDB</a></p><blockquote><em>Want to hear some real voices 🙌🏻</em></blockquote><blockquote><em>I’ve been thinking about a recent trend: with the emergence of AI, there’s been a shift in how people view the traditional office suite trio (Word, Excel, PowerPoint). Some suggest that the future is Markdown as Word, CSV as Excel, and HTML as PowerPoint.</em></blockquote><blockquote><em>I’ll leave the first two aside for now, but I do have some doubts about the last point, “HTML as PowerPoint.” From the perspective of someone with a technical background (like me), I think replacing PowerPoint with HTML is a great idea because it offers better interaction and is highly suitable for content presentation.</em></blockquote><blockquote><em>However, I was chatting with an interviewer recently, and he mentioned that while we might think HTML is great, in many traditional enterprises, if you need to communicate with a client and they ask for a slide deck, you often still have to use the PPT (PowerPoint) format.</em></blockquote><blockquote><em>But then again, I thought, when we usually send a slide deck to clients, we generally just send a PDF. That’s because it might contain sensitive information that we don’t want others to modify. If it’s just about exporting to PDF, HTML can do that just as well.</em></blockquote><blockquote><em>So I’ve been pondering, or rather, I’m genuinely curious:</em></blockquote><blockquote>For people or clients without a technical background working in traditional industries, is the PowerPoint format itself very important to them?</blockquote><blockquote>If we just send them an HTML file directly, would it make them feel we are unprofessional?</blockquote><blockquote><em>Thinking about it from another angle, if I convert it to PDF anyway, the client wouldn’t actually know whether I made it with PowerPoint or HTML.</em></blockquote><blockquote><em>I really want to know if, outside the AI circle, the vast majority of ordinary people and traditional clients would find replacing PowerPoint with HTML acceptable while still reflecting professionalism.</em></blockquote><blockquote><em>I don’t have an answer to this question right now, but I’d truly love to hear everyone’s thoughts. If more enterprise clients, bosses, or everyday folks could share your views, I would be extremely grateful 🙏🏻</em></blockquote><p>This post reminded me of my internship last year. At that time, we hadn’t really entered the AI era; LLMs were literally just chatbots, and not many people were optimistic about the efficiency boosts they could bring. Long before that, those who code regularly understood why “Markdown as Word” made sense, and those working with data processing practically treated CSV as their mother tongue. Both are the general trend now; much like the debate on “whether AI is actually practical,” there’s really nothing left to argue about. However, “HTML as PowerPoint” has a much smaller audience and has traditionally only been played around with in front-end dev circles.</p><p>Once, I had the opportunity to present a few ideas. As usual, I used Markdown to organize my thoughts and draft the materials first. The company’s computers had mediocre specs, taking four or five seconds just to open a PPT, so I thought I might as well export the Markdown to HTML. HTML is great for version control, opens in 0 seconds, eliminates the hassle of choosing “which app to open it with,” can be viewed by anyone on both PCs and mobile phones, has no Office version compatibility issues, features responsive layouts… everything seemed perfect. The only downside was that it wasn’t convenient for others to annotate or edit, but for one-off presentation materials, that didn’t matter much.</p><p>HTML itself does not embed static resources like images. Fortunately, I wasn’t stupid enough to send my boss a .zip archive. I put the HTML and the accompanying assets/ folder on the company&#39;s NAS and told my boss/colleagues: &quot;The presentation document is in the xxx directory on the NAS; just double-click to open it.&quot; Then, unsurprisingly, surprise occurred. A colleague found it too troublesome to navigate to the NAS directory every time to view the document, so he copied the HTML file out. Upon opening it, the formatting was naturally completely broken. Someone forwarded the file via WeChat/DingTalk, which naturally resulted in the loss of all static resources as well.</p><p>At the time, I just thought I had caused a minor inconvenience for my boss/colleagues, but nowadays, whenever I look back on it, I’m filled with regret. I thought I had considered everything so comprehensively, believing I had a “better” and “more elegant” solution, and assuming that the separation of semantics and presentation was always a best practice. It turned out to be nothing more than self-indulgence. Nobody really cares whether a document takes 5 seconds or 0 seconds to open; nobody cares if copying 10 files takes up 100KB or 100MB; nobody really cares about responsive layouts when they’re in such a rush that they have to view a document on their phone; and nobody cares about version control. Taking a step back, if you just throw the pre-edited and post-edited files into WeChat File Transfer Assistant, why shouldn’t that count as version control? The underlying principle isn’t much different from Git LFS, and objectively, it’s not THAT inconvenient.</p><p>I really created a hassle…</p><p>However, this isn’t to say that there’s inherently something wrong with the HTML format making it unsuitable for document sharing. If I had placed the static resources in object storage and used absolute paths back then, the HTML wouldn’t have “broken” no matter how it was forwarded. If I had used mdBook and CI/CD’d it to Gitee (Chinese GitHub), not only would it have looked visually superior, but a single link would have ensured that everyone was always viewing the exact same document as I was. At the very least, I could have just base64-encoded the images and stuffed them directly into the HTML… At the time, on the surface, it seemed like I was just brainless when sending files; but in essence, I failed to truly respect other people’s working habits, caring only about forcing what I thought was “advanced” onto others. On the other hand, I have to admit that, both then and now, I lacked a clear-headed understanding of my own fundamentally amateurish skill level.</p><p>往者不可谏(What is past cannot be amended). Here are some naive thoughts that haven’t yet been slapped in the face by reality:</p><p><strong>Regarding PPT:</strong> In less than a year, LLMs have evolved from chatbots to being able to generate highly decent-looking web pages. Just as humans naturally know how to make PPTs, LLMs naturally know how to write HTML. When human capability to make PPTs finally falls behind an LLM’s capability to write HTML, and when LLMs effortlessly generate exquisite reports on computers with “hidden file extensions,” PPTs will naturally fade out.</p><p>Once HTML transitions from being something “99% of people can only passively consume” to something “anyone can generate by default with zero knowledge requirements,” everyone will realize that there’s nothing magical about it. It’s just like when Microsoft Pinyin IME appeared in 1993, and everyone realized that typing Chinese characters on a computer wasn’t anything magical. The “tech bro” Wubi users back then merely typed a bit faster than Pinyin users; how much different are they from the “people who can write HTML” today?</p><blockquote><em>The history of Chinese keyboard input: Wubi (五笔) is a fast but notoriously complex shape-based typing method that required intense memorization, making it an elite, specialized skill for early Chinese computer users. In contrast, Pinyin (拼音) is a phonetic method (typing words as they sound) that was initially painfully slow due to homophones, but the 1993 introduction of smart, predictive Pinyin software allowed anyone to type fluidly using natural language without any technical training. Since then, the market share of Wubi quickly plummeted to around 3%.</em></blockquote><p><strong>Regarding PDF:</strong> The original post said, “when we usually send a slide deck to clients, we generally just send a PDF. That’s because it might contain sensitive information that we don’t want others to modify.” That’s not entirely the case, actually. Modern readers can universally modify PDFs. With today’s OCR engines and generative AI, even bitmap PDFs can be altered flawlessly. Therefore, the method of raising the cost of tampering/forgery simply because “modifying a PDF is a hassle” is outdated. PDF as a “final delivery format” no longer stops someone who genuinely wants to alter it. Short of putting it on a blockchain, true tamper resistance can and only can be achieved through digital signatures.</p><p>The immutability of a PDF relies on its distribution as a single file and its signature, while its non-repudiation relies on trusted third parties (like CAs and WeChat File Transfer Assistant). HTML can do both of these just as well. Moreover, HTML with JS/WASM is not only just as secure as PDF with Acrobat JS, but it also naturally boasts openness, compatibility, high performance, semantic structures, machine friendliness, and more. It can even achieve highly sci-fi features like localized signing/encryption, layered signing/encryption, and trusted, traceable collaboration based on these — all requiring nothing more than a JS/WASM file (which have been readily available for years). Most importantly, the browser is unequivocally the most installed software in the world, bar none. All browsers have natively supported all the above features for a long time, so there isn’t even a barrier to entry; no one needs to “learn” how to use HTML. To sum up, whether PDF’s raison d’être as the “final delivery format” still holds water is also up for debate, so you can’t blame Adobe’s stock price for taking a skydive.</p><p>I’ve fallen into self-indulgence again.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bfa18e33a71" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI, The Flesh Polymer, And Some Ramblings Of An Overworked]]></title>
            <link>https://medium.com/@sha1rholder/ai-the-flesh-polymer-df81e5ec22cb?source=rss-4c891e377534------2</link>
            <guid isPermaLink="false">https://medium.com/p/df81e5ec22cb</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sha1rholder]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 19:22:07 GMT</pubDate>
            <atom:updated>2026-03-17T07:35:22.635Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NKRgQVLfc6RyGfJGF08Zaw.png" /><figcaption>Thanks Nanobanana2</figcaption></figure><p>AI is a hideous flesh polymer. We once fed this infant full of curiosity, only to watch it mutate into a sun-obscuring monstrosity, before we could even blink.</p><p>AI absorbs human knowledge, aggregates it to teach even more humans, and is already generating vast amounts of new knowledge on our behalf. Today, 99% of the views on the knowledge we create come from AI-driven web crawlers.</p><p>And so, we too have become a part of this polymer.</p><p>Knowledge has already surpassed the absolute limit of what any single person can bear. For one’s knowledge to multiply, it must be seen by others in this world. But since we can no longer pursue exposing ourselves directly to one another, we are left with no choice but to inject our semen into that gargantuan polymer. That orb of flesh transmutes every gram of semen into hundreds of thousands of tons of milk, seducing humanity with its massive breasts. How could people possibly withstand a thousand years of thirst? And so we close our eyes, greedily suckling at the teats of that fleshy mass (the milk is laced with cocaine!), no longer seeing each other.</p><p>People silently mutter to themselves: the reason this colossal mass of flesh is abominable is that it never acquires or uses data in permitted ways. Yet, deep down, we know better: social norms always lag behind technological progress. We can only endlessly patch up the narrow cages of our own making, treading a perilous balance between a complete loss of control and being crushed to death by our own swelling selves.</p><p>The people suckling the milk are, slowly, being sucked inside by the teats.</p><h4><strong>Yet the people who closed their eyes, have grown eyes all over their bodies.</strong></h4><p>Better not continue if your Sanity is low.</p><p>These are some random inspiration or just illusions after working on designing a “superior” markup language for Idk-how-many hours continuously. Before AI could feed me the essence of others’ wisdom and let me “grow a bunch of extra eyes,” I hadn’t made up my mind to start this task. But I discovered that my design iterations for a “better markup language for AI era,” along with deepening human understanding of AI and information and research progress (especially regarding specification complexity in LLM output hard-constraint FSM schemas), increasingly resemble less a tool for communication and display between people, and more a bridge between humans and machines. It seems direct human-to-human communication is no longer the focus of “language functionality” in this era. Human-friendliness holds less and less weight in the definition of “efficiency”… But on the other hand, since machines are currently so friendly to humans, being friendlier to machines is being friendlier to humans :D</p><p>It occurred to me: the paradigm of how we share knowledge has, over the years, gradually shifted from human-oriented to machine-oriented, and we call this “efficiency” (I have to admit, it actually seems to be true). The internet connecting people went from pursuing stylistic expressiveness to SEO optimization. The byproduct of SEO, well-structured data, created super smart AI, then we moved from SEO to GEO, and instantly to JSON-encapsulated Markdown that strips away even HTML (anonymous closing tags are really great for reducing schema complexity!)… Christ, I might have gone crazy, but I reduced the complexity from O(n) to O(n³) (worst situation, usually better).</p><p>I told you not to.☹️</p><p>(Translated by some llms)</p><blockquote># 聚合物</blockquote><blockquote>AI是一团可憎的血肉聚合物。我们曾满怀好奇地喂养这个婴儿，如今眼见它变成遮天蔽日的怪物，连眼都不及眨</blockquote><blockquote>AI学习着人类的知识，又把知识整合来教给更多的人类，甚至已在为人类创造大量新的知识。如今，我们创造的知识99%的浏览都来自爬虫和AI</blockquote><blockquote>于是我们也成为了这个聚合物的一部分</blockquote><blockquote>知识已然超过任何人所能承载的上限。要让自己的知识得以增殖，就要被这个世界上的其它人看到，但我们无法再追求将自己直接暴露给他人，便只能将自己的精液注入那巨大的聚合物。那肉球将每一克精液转化为成百上千吨奶水，用硕大的乳房勾引着人们。人们哪堪千百年的饥渴？便闭上眼，对着那团血肉的乳头贪婪地吸吮（那奶水里混了可卡因！），不再看见彼此</blockquote><blockquote>人们心里默念着：这团庞大血肉之所以可憎，是因为它们从不以被允许的方式获取和使用数据。但我们更清楚：社会规范总是滞后于科技发展，我们只能对自己创造的狭小囚笼修修补补，在彻底失控和被膨胀的自己挤死之间做举步唯艰的平衡</blockquote><blockquote>吸着奶的人们 慢慢被那奶嘴吸进去了</blockquote><blockquote>**但 闭上眼的人们 浑身都长出了眼睛**</blockquote><blockquote>— -</blockquote><blockquote>These are some random inspirations or just illusions after working on designing a “superior” markup language continuesly for idk-how-many hours。在AI能用他人智慧的精华喂养我，让我“长出一堆额外的眼睛”前，我一直没下定决心开始这项任务。但我发现，我对“更好的标记语言”的设计迭代随着人们对AI和信息的理解深入和研究进展（尤其是近期llm输出硬约束FSM schema方面的规约复杂度研究）变得越来越不像是人与人之间交流的工具，而是人类与机器之间的桥梁。似乎直接的人人沟通已不再是这个时代里“语言功能性”的重点。人类友好性在“高效”定义中的权重占比越来越低……不过换个方面讲，因为现在的机器对人类实在太友好了，所以对机器越友好就是对人越友好 :D</blockquote><blockquote>突然想到，我们分享知识的范式这么多年来，已经从面向人类一步步转向面向机器，且称之为“高效”（不得不承认事实上好像确实如此）。联系人与人的互联网从追求样式表现力到SEO优化，SEO的副产品搞出了super smart AI，就又从SEO到GEO，再瞬间到连html都剥离了的json封装markdown（匿名结束标签真是降低schema复杂度的好东西！）……天哪，我可能是疯了，但是我把语法复杂度从O(n)降到了O(n³)（最坏情况，通常更好）</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df81e5ec22cb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Thinking in English, Dying in Chinese: Chinese AI Landscape in Late 2025]]></title>
            <link>https://medium.com/@sha1rholder/thinking-in-english-dying-in-chinese-chinese-ai-landscape-in-late-2025-a72b1879084e?source=rss-4c891e377534------2</link>
            <guid isPermaLink="false">https://medium.com/p/a72b1879084e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[deepseek]]></category>
            <dc:creator><![CDATA[Sha1rholder]]></dc:creator>
            <pubDate>Mon, 01 Dec 2025 16:58:00 GMT</pubDate>
            <atom:updated>2025-12-01T22:15:08.523Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1JUWTclnNAMkbGDD8McYvQ.png" /></figure><p>On the afternoon of December 1, 2025 (UTC+8), DeepSeek released two models: DeepSeek v3.2 official and a special “showcase” edition, v3.2-speciale.</p><p>I tested them briefly and noticed that, beyond the difference in Chain-of-Thought (CoT) length, their reasoning strategies diverge significantly. The smarter speciale model has a much higher chance of thinking in English regardless of the input language. This confirms a harsh reality: even for top-tier Chinese models, the quality and quantity of English training data still far surpass that of Chinese data (so fk Robin Li and Baidu for ruining Chinese internet)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UsYaQ8Mlrm5ZPB_FalXYBg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EldS0QbVED8gzJONpxJSzg.png" /></figure><p>The all-English CoT reminded me of the release of MiniMax M2 on October 27. It barely managed to join the top tier of open-source models by utilizing a fully English CoT. At the time, I thought this was an unconventional innovation for a Chinese domestic model. Now, after the release of DeepSeek v3.2 and v3.2-speciale, it seems this approach poses no real barrier to entry for other Chinese companies. The difference is that while other Chinese models haven’t fully unleashed their capabilities yet, MiniMax M2 seems to have already squeezed that particular lemon dry.</p><p>Another Chinese model, Moonshot’s flagship, Kimi K2 thinking was released on November 6. While the model shows a significant boost in intelligence, the non-Turbo mode suffers from an agonizingly long chain-of-thought and slow token output (according to statistics, it takes on average five times longer to complete an inference task than Claude Sonnet 4.5 thinking); the Turbo mode, meanwhile, is just too expensive, worthy not. I get the feeling that K2 thinking sacrifices cost-efficiency and practicality somewhat for the sake of high benchmark scores. Moreover, it has severe “infinite loop” bugs. In my humble opinion, seeing such issues in a production release is a classic result of rushing a product out the door.</p><p>Aside from MiniMax and Moonshot, all of China’s leading AI companies have a cash-cow business serving as a financial lifeline. In addition to the three tech giants — Alibaba, ByteDance, and Tencent — DeepSeek is backed by High-Flyer, a quantitative trading firm, while Zhipu AI (developer of the GLM series) is a state-backed enterprise with strong ties to Tsinghua University.</p><p>Thus, I “suspect”, under intense investor pressure to prove their monetization strategies, MiniMax and Moonshot had to roll out revenue-generating models prematurely. Consequently, they have become the most aggressive domestic vendors in terms of paid traffic and marketing in China, all while having to divert resources to polish their agents or coding plans.</p><p>For the tech giants, fighting on multiple fronts is sustainable — ByteDance has Doubao, Tencent has Yuanbao, Alibaba has Qwen — but for smaller startups, where are they supposed to find the bandwidth to refine all these products and compete with the titans — with such limited resources?</p><p><strong>Among base models from China today, only DeepSeek v3.2 (think) and Qwen3 series rank in the top tier in terms of intelligence and practicality. Besides, </strong>Tencent and ByteDance, controlling the key traffic gateways, are cementing their dominance in the ToC sector.</p><p>It’s a pity MiniMax and Moonshot haven’t finished IPO yet; otherwise, there might have been some profit to be made in shorting them. Or anybody interested in creating a prediction market on who’s gonna be the winner of China’s AI race?</p><h4>Reference</h4><p><a href="https://huggingface.co/deepseek-ai/DeepSeek-V3.2">DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models</a></p><p><em>No rights reserved.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a72b1879084e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Game Theory of a Frustrating and F**ked Up Cafeteria Queue]]></title>
            <link>https://medium.com/@sha1rholder/game-theory-of-a-frustrating-and-f-ked-up-cafeteria-queue-2aedb98d12c9?source=rss-4c891e377534------2</link>
            <guid isPermaLink="false">https://medium.com/p/2aedb98d12c9</guid>
            <category><![CDATA[game-theory]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[math]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sha1rholder]]></dc:creator>
            <pubDate>Mon, 17 Nov 2025 06:45:14 GMT</pubDate>
            <atom:updated>2025-11-17T06:45:14.757Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qsbaColPkDfzl2tLYEWb_Q.png" /></figure><blockquote><em>This is a </em><strong><em>cynical, mathematical, and computational rant</em></strong><em> born from standing in a ridiculous queue.</em></blockquote><blockquote><em>It starts with classic </em><strong><em>Game Theory</em></strong><em> (rational agents), expands into </em><strong><em>computational modeling</em></strong><em> and the paradox of </em><strong><em>Pareto optimality</em></strong><em>, and finally devolves into discussions of </em><strong><em>non-causal systems</em></strong><em> and parallels with </em><strong><em>Multi-Agent AI Training paradigms</em></strong><em>.</em></blockquote><blockquote><em>Read if you enjoy applying crazy science to some stupid real-world problems.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6YqalI-CsARqc4XXzoZrZw.png" /><figcaption>cafeteria dish return slot in UoE/ZJU Haining campus</figcaption></figure><p>For those unfamiliar, here’s the setup: each cafeteria section has its own conveyor belt for returning trays. You approach from the front and place your tray on one of two return stations. During peak hours, a line forms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*97BPE4YuBxUic_0VGmOg2w.png" /></figure><p>If it were just a queue, it would be no big deal — nothing more than a resource bottleneck. The issue is, even though each belt has two return stations, the queue almost exclusively forms at the <em>front</em> station, while the rear station often sits empty. This significantly reduces the efficiency of tray returns. This phenomenon can actually be explained with a set of 7 axioms.</p><ol><li>Every queuer is a rational, self-interested agent.</li><li>All queuers form the queue at roughly the same time, but in a sequential order.</li><li>A queuer can walk to either station instantly. Using the front station is immediate, but walking to the rear station and back adds an extra 7 seconds of travel time</li><li>Each station can only serve one person at a time. Returning a tray takes approximately 2 seconds, and every queuer knows this.</li><li>Due to a wall near the entrance, only 3 people can queue directly in front of the first station. However, the space between the front and rear stations can hold an infinite number of people.</li><li>The area near the conveyor belt is only wide enough for 3 people to walk abreast, so we can consider it as three parallel lanes.</li><li>Queuers must leave a clear exit path for people who are finished.</li></ol><p>Axiom 1 implies that each queuer seeks to maximize their utility U, where U = -t, and t is the total time an individual spends returning their tray. Based on axioms 2, 3, and 4, we can deduce a critical decision point. The wait time at the front is 2 seconds per person. The travel time to the rear is 7 seconds. Therefore, if there are 3 people already at the front station, the 4th person in line faces a choice: wait 6 seconds (3 people × 2s) at the front, or spend 7 seconds walking to the empty rear station. Clearly, <strong>waiting is faster</strong>. So, queuer #4 waits at the front.</p><p>This, combined with axioms 5, 6, and 7, results in queuer #4 blocking lane two, leaving queuers #5 and beyond with no choice but to wait. When someone finishes, queuer #5 becomes the new fourth person in line and continues to block lane two, preventing anyone behind them from reaching the rear station. And so, the rear station becomes virtually useless, and the belt’s throughput averages one person every 2 seconds — no different from having only one station.</p><p>Clearly, <strong>regardless of the queuers’ decisions, for a queue of x people, the total return time T(x) and the individual return time for the n-th person t(n) have the following relationship:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*8Lz5RCHx0M2iezixavJnNg.png" /></figure><p>In the selfish-agent scenario above, we can find T₁(x) from t₁(n):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/408/1*6iNw2P8atfK2pLM_RH5e_g.png" /></figure><p>A classic “tragedy of the commons,” nothing too exciting. The solution to avoid this and improve system efficiency is simple and counter-intuitive: <strong>change axiom 4, increasing each person’s tray return time by 1 second, to 3 seconds.</strong></p><p>At this point, queuer #4 snaps. Waiting 9 seconds at the front is worse than spending the extra 7 seconds to walk to the rear. So, they go to the rear, freeing up lane two. So, queuer #4 heads to the rear. Now, what about queuer #5? They arrive to see 3 people at the front station (a 3 × 3s = 9s wait) and queuer #4 walking to the back. The rear station itself is still open. For #5, using the rear station would take 7s (walk) + 3s (return) = 10s. Since 9s is less than 10s, #5 decides to wait at the front, blocking lane two again. But this block is temporary. After 3 seconds, one person leaves the front. Queuer #6 arrives, sees the lane is now clear, and faces the exact same choice as #4 did: a 9s wait at the front (3 people ahead) versus a 10s commitment to the rear. They’ll also choose the rear. This creates a pattern: the 4th, 6th, 8th… person will go to the back</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/732/1*dkNMiTxldDePOmlV6hKtUw.png" /></figure><pre>import matplotlib.pyplot as plt<br>import numpy as np<br><br><br>def cal_T2(xi):<br>    if xi &lt;= 3:<br>        return 3 / 2 * xi**2 + 3 / 2<br>    elif xi &gt;= 4 and xi % 2 == 0:<br>        return 3 / 4 * xi**2 + 5 * xi - 4<br>    else:<br>        return 3 / 4 * xi**2 + 5 * xi - 15 / 4<br><br><br>range_val = 20<br>x = np.arange(1, range_val + 1)<br><br>T1 = x**2 + x<br>T2 = np.array([cal_T2(xi) for xi in x])<br>y = T1 - T2<br><br>fig, ax = plt.subplots()<br>ax.plot(x, y)<br>ax.set_ylabel(&quot;ΔT (seconds)&quot;)<br>ax.set_title(&quot;T₁(x) - T₂(x)&quot;)<br>ax.grid(axis=&quot;y&quot;, linestyle=&quot;--&quot;, alpha=0.7)<br>ax.spines[&quot;top&quot;].set_visible(False)<br>ax.spines[&quot;right&quot;].set_visible(False)<br>ax.spines[&quot;bottom&quot;].set_position((&quot;data&quot;, 0))<br>ax.set_xticks(x)<br><br>plt.tight_layout()<br>plt.show()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*MhNwpvdKiDbdRxyJjNdyHw.png" /></figure><p>As you can see, for a queue of more than 15 people, adding 1 second to the individual return time actually <em>reduces</em> the total waiting time for everyone. It can also be proven that this strategy is not only an individual optimum for the queuers but also a group optimum for any queue size — not a second more, not a second less. But I’m too lazy to write out the proof here.</p><p>Of course, the above describes a phenomenon generated by a group of selfish agents. Real life isn’t populated solely by selfish agents; there are also altruists, collectivists, and whatever-ists. The whatever-ists are unpredictable, so let’s look at the first two.</p><p>As the problem grows more complex, let’s first analyze the nature of a queuer’s decision, <em>a_n</em>​. When lane two is free, every queuer has only two choices: queue at the front or go to the rear. This can be abstracted as <em>a_n</em>​∈{<em>True</em>,<em>False</em>} (True for front, False for rear). Since every queuer will inevitably face a moment when lane two is free, <strong>the decisions of the entire queue can be represented by a 1D boolean array </strong>[<em>a</em>1​,<em>a</em>2​,<em>a</em>3​,…,<em>a_last</em>​]<strong>, and the n-th person’s decision is essentially a binary function.</strong> To reduce complexity, let’s introduce an 8th axiom: if a rational agent predicts equal utility for queuing at the front versus going to the rear, they will choose the front, i.e., they choose True when <em>U</em>(<em>True</em>)=<em>U</em>(<em>False</em>). (This also aligns with the lazy psychology of a real-life queuer).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/892/1*ILiRSobZ6yCrpJMbmgG_yQ.png" /></figure><p>We can also calculate each queuer’s return time, t, based on Afront<em>Afront</em>​ and an<em>an</em>​, and then recursively find the total return time T. (Expressing this with formal logic would be a nightmare, so let’s use Python to make it tangible.)</p><pre># These implementations have extremely poor performance. Don&#39;t learn.<br>def t(A_front: list[bool], action: bool) -&gt; int:<br>    recycle_time, tail_time = 2, 7  # or whatever time combination you specify<br>    A_front = A_front.copy()<br>    timer = 0<br>    head = 0<br>    tail = 0<br>    while A_front:<br>        if A_front.pop(0):<br>            head += 1<br>            if head &gt; 3:<br>                head -= 1<br>                tail -= 1<br>                timer += recycle_time<br>                if tail &lt; 0:<br>                    tail = 0<br>        else:<br>            tail += 1<br>    return (<br>        timer + recycle_time * (head + 1)<br>        if action<br>        else timer + recycle_time * (tail + 1) + tail_time<br>    )<br>    # the first value is actually always equal to recycle_time * A_front.count(True)<br><br>def T(A: list[bool]) -&gt; int:<br>    T_val = 0<br>    A_front = []<br>    for action in A:<br>        T_val += t(A_front, action)<br>        A_front.append(action)<br>    return T_val</pre><p>Obviously, for a selfish agent, <em>U</em>=−<em>t</em>, so there’s no need to consider <em>A_back</em>​. Therefore, predicting the decision chain of a group of selfish agents, <em>A_selfish</em>​, is quite simple.</p><pre>def a_selfish(A_front: list[bool]) -&gt; bool:<br>    return -t(A_front, True) &gt;= -t(A_front, False)<br><br>def A_selfish(A_front: list[bool], n: int) -&gt; list[bool]:<br>    A = A_front.copy()<br>    for _ in range(n):<br>        A.append(a_selfish(A))<br>    return A<br><br>A_selfish_20 = A_selfish([], 20)<br>print(A_selfish_20)<br>print(f&quot;T_selfish = {T(A_selfish_20)}&quot;)<br># results when {recycle_time, tail_time = 2, 7}<br># [True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True]<br># T_selfish = 420</pre><p>For a collectivist (or altruist), <em>U</em>=−<em>T</em> (or <em>U</em>=−(<em>T</em>−<em>t</em>)). Since <em>A_back</em>​ is a variable of T, their decision depends on their prediction of how those behind them will act. This requires case-by-case analysis. Let’s first consider the case where collectivists (and altruists) assume everyone behind them is a selfish agent. Their decision chains, <em>A_helpall</em>​ (and <em>A_helpothers</em>​), would be:</p><pre>def a_helpall(A_front: list[bool], back: int) -&gt; bool:<br>    return -T(A_selfish(A_front + [True], back)) &gt;= -T(<br>        A_selfish(A_front + [False], back)<br>    )<br><br>def A_helpall(A_front: list[bool], n: int) -&gt; list[bool]:<br>    A = A_front.copy()<br>    for i in range(n):<br>        A.append(a_helpall(A, n - i - 1))<br>    return A<br><br>def a_helpothers(A_front: list[bool], back: int) -&gt; bool:<br>    return -(T(A_selfish(A_front + [True], back)) - t(A_front, True)) &gt;= -(<br>        T(A_selfish(A_front + [False], back)) - t(A_front, False)<br>    )<br><br>def A_helpothers(A_front: list[bool], n: int) -&gt; list[bool]:<br>    A = A_front.copy()<br>    for i in range(n):<br>        A.append(a_helpothers(A, n - i - 1))<br>    return A<br><br>A_helpall_20 = A_helpall([], 20)<br>print(A_helpall_20)<br>print(f&quot;T_helpall = {T(A_helpall_20)}&quot;)<br># [False, False, False, False, False, False, False, False, True, True, True, True, True, True, True, True, True, True, True, True]<br># T_helpall = 284<br><br>A_helpothers_20 = A_helpothers([], 20)<br>print(A_helpothers_20)<br>print(f&quot;T_helpothers = {T(A_helpothers_20)}&quot;)<br># [False, False, False, False, False, False, False, False, False, False, False, False, False, False, False, False, False, False, False, True]<br># T_helpothers = 515</pre><p>In fact, for any given values in axioms 3 and 4, it can be proven by exhaustion that the decision chain formed by “collectivists who assume everyone behind them is selfish” is always one of the globally optimal solutions (if we only consider efficiency, any chain where <em>T</em>=<em>T_min</em>​ is globally optimal). It can even be proven that “for any queue composed of the three types of agents above, replacing any one agent with a collectivist is always Pareto optimal.” For “altruists who assume everyone behind them is selfish,” however, the efficiency is far worse than that of a purely selfish group.</p><blockquote><em>The altruist, assuming those behind them are selfish, tends to self-sacrifice by queuing at the rear station, which they believe will go unused. This sacrifice is only valuable if there are indeed selfish agents behind them who can benefit from the freed-up front position. If there are too few altruists, the system behaves almost identically to the inefficient “all-selfish” scenario. If there are too many, many of them make unnecessary sacrifices, as there aren’t enough selfish actors behind them to capitalize on the efficiency gains they create.</em></blockquote><blockquote><em>— Claude Sonnet 4.5</em></blockquote><p>The above methods can be easily extended to cases like <em>U</em>=−(0.2<em>T</em>+0.8<em>t</em>), which better reflect real-world queuers who are “somewhat selfish yet still consider the collective interest.”</p><p>To wrap up this part of the problem: if altruists are randomly distributed in the queue, there must exist a ratio of selfish agents to altruists, <em>k</em>=<em>a_selfish</em>​/<em>a_helpothers</em>​, that minimizes the expected value of T, i.e., <em>E</em>(<em>T</em>)=<em>E_min</em>​(<em>T</em>). Thus, <em>E</em>(<em>T</em>) can be modeled as a bivariate function of queue size n and ratio k. And to achieve <em>E</em>(<em>T</em>)=<em>E_min</em>​(<em>T</em>), k can also be expressed as a function of n.</p><p>For any given <em>n</em>, <strong>I have discovered a truly numerical solution of this </strong>(with a complexity of O(2^N⋅N²))<strong>, which this memory is too small to contain.</strong> &lt;del&gt;Actually I don’t know how to find the analytical solution.&lt;/del&gt;</p><p>No matter how complex the makeup of a queue of agents “who assume everyone behind them is selfish” is, there is always a unique decision chain as a static solution. However, if the queue contains collectivists or altruists who know the actual composition of <em>A_back</em>​, they must predict the decisions of other collectivists and altruists to maximize their utility. This reflexivity turns the rational queue into a non-causal system, violating the laws of physics. And for a system no longer based on the rational agent hypothesis, all the logic above falls apart, and the queue has no classical equilibrium solution.</p><p>Under bounded rationality, predicting the decision chain would require introducing levels of cognition and incorporating statistical and probabilistic methods. An individual’s decision would then be seen as a recursive belief update about others’ strategies… If I keep going, this starts to sound like some AI training methodology (<strong>Bayesian Inverse Reinforcement Learning in Multi-Agent Systems</strong>), which is way above my pay grade.</p><p>Having written all this, I finally feel a sense of catharsis about getting stuck at the tray return. Maybe it’s not that everyone is an idiot; perhaps the world is just far more complex than I can comprehend.</p><p><em>No rights reserved.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2aedb98d12c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hidden Marketing Wars of Chinese Consumer Electronics: Part 1.5]]></title>
            <link>https://medium.com/@sha1rholder/hidden-marketing-wars-of-chinese-consumer-electronics-part-1-5-9d0b5b0a9be2?source=rss-4c891e377534------2</link>
            <guid isPermaLink="false">https://medium.com/p/9d0b5b0a9be2</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[branding]]></category>
            <dc:creator><![CDATA[Sha1rholder]]></dc:creator>
            <pubDate>Mon, 17 Nov 2025 00:01:17 GMT</pubDate>
            <atom:updated>2025-11-17T07:11:22.285Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Oumd7LAyOrEzijDd.png" /><figcaption>photo from Dongguan Dianxuntong Electronic Technology Co., Ltd</figcaption></figure><p>Previous article: <a href="https://medium.com/@sha1rholder/hidden-marketing-wars-of-chinese-consumer-electronics-part-i-95f07e5df78a">Hidden Marketing Wars of Chinese Consumer Electronics: Part I</a></p><p>This article is a slight detour from my original plan. I had intended to write a deep dive into the MCN agency Chronovate and the idea that “rational consumers don’t exist in decentralized marketing,” but I’ve decided to pivot. This post continues the same theme as my last one, but shifts its focus from consumer audio to the stories behind some made-in-China peripherals.</p><p>This article will shed light on two lesser-known events of the competition in China’s domestic peripheral devices market from a consumer’s perspective.</p><h3>When Tech Is No Longer a Barrier, What Decides Victory?</h3><p>In the second half of 2023, the Chinese domestic peripheral market exploded. The days of Logitech, Razer, and ROG enjoying their hefty profit margins unchallenged were suddenly over. The challenge didn’t come from another giant, but from a swarm of new players you’ve probably never heard of. This was when made-in-China brands began to rapidly seize market share in the gaming peripheral space.</p><p>The reason is simple: the supply chain for key components — mouse shells, micro-switches, scroll wheels, keyboard switches, and PCBs — is heavily concentrated in China’s Pearl River Delta. China has no shortage of engineers capable of designing these parts into respectable products, and an even greater abundance of factories that can squeeze costs to the absolute limit.</p><p>As for the parts with a real technological barrier, like the high-precision, high-polling-rate optical sensors in flagship mice? Even international giants like Razer, SteelSeries, Corsair, and Alienware rely on source manufacturers like <strong>PixArt</strong>. This means Chinese brands aren’t at a competitive disadvantage. Furthermore, <strong>even the world’s top peripherals use chips made with widely available, mature manufacturing processes.</strong> For China’s industrial ecosystem in the 2020s, these are not a problem. This reality prevents companies like PixArt from price-gouging and, in turn, creates fertile ground for Chinese brands to wage a price war.</p><blockquote><em>The cost of taping out a new chip is extremely high and the cycle is long. But once shipment volume and profit reach a certain threshold, domestic alternatives will eventually emerge. In fact, some Chinese chip designers are already working on cracking the high-end mouse sensor market.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ccdlTJHn0OV6ZBsZU97iMA.png" /><figcaption>image from Dongguan Dianxuntong Electronic Technology Co., Ltd</figcaption></figure><p>So, when technology and cost no longer create a significant gap, the way various Chinese brands build peripherals starts to look a lot like an ancient folk ballad:</p><blockquote><em>东市买骏马，西市买鞍鞯，南市买辔头，北市买长鞭 — — 《木兰辞》</em></blockquote><blockquote><em>“Buy a fine steed in the east market, a saddle in the west market, a bridle in the south market, and a long whip in the north market.” — from the </em>Ballad of Mulan<em>. (Her legend was adapted into a Disney movie, which I reckon many have seen.)</em></blockquote><blockquote><em>This famous line describes Mulan preparing for war by sourcing different gear from different places. In this context, it means brands are just assembling their products from a common pool of readily available, high-quality suppliers.</em></blockquote><p>Take mice, for example. In late 2023, a new mouse was released every few weeks, but most of them used the exact same internal solution (typically a PixArt PAW3395/3950 sensor + Nordic 52840 MCU). Brands with similar positioning were separated by only a few dollars. At this point, the competition wasn’t about who made a <em>better</em> product, but who could capture consumer mindshare <em>first</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/649/1*tqh3nsojAVcJ4WI55FpPbQ.jpeg" /></figure><p><strong>Timing became the line between life and death.</strong> Every day you delayed your product launch meant another competitor might be seen by more consumers before you. In the eyes of the consumer, “late” equals “copycat,” and “copycat” equals “unoriginal.” This is when tactics like astroturfing, fake reviews, and sockpuppets stopped being a question of “if” and became a question of “how many can your PR team deploy?”</p><h3>VanGuard’s “Clone Jutsu” and the Truth Behind the Sell-Off</h3><p>The first story is about VanGuard Network, a company that transitioned from a gaming media outlet to a peripheral brand.</p><p>In late 2023, VanGuard suddenly began “birthing children” at an incredible rate. It operated three main sub-brands: Atom TeK, VerteX Entertainments, and MaidLions, all with highly overlapping product lines. You might buy a VanGuard Fox1 mouse, only to find that Atom TeK released a mouse with the exact same internals a short while later, just with a different shell, slightly tweaked mouse feet, and maybe a cheaper price tag. VanGuard never tried to hide the relationship; all their brands shared the same driver software.</p><p>At the time, peripheral enthusiasts were puzzled by this strategy. Why was VanGuard promoting its sub-brands so heavily? Then, a rumor spread that explained everything: <strong>VanGuard was selling the entire brand to a Chinese distributor of {a top-tier hardware giant, ROyal Gaming}.</strong> It wasn’t a tech acquisition or a strategic partnership; it was a straight-up brand sale. ROyal wasn’t interested in VanGuard’s tech or supply chain — they wanted the reputation and community VanGuard had built among enthusiasts as one of the first successful new Chinese brands.</p><p>Looking back at the “clone jutsu,” it all made sense. This strategy is a classic business maneuver known as <strong>金蝉脱壳</strong>.</p><blockquote><em>“The golden cicada sheds its shell.”</em></blockquote><blockquote><em>This is one of the famous Thirty-Six Stratagems. It refers to a tactic of escaping or reinventing oneself by leaving behind a decoy or an old identity — the “shell.”</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lu7ecvPztGJmcXhgPZejrA.jpeg" /></figure><p>VanGuard was a new company, short on cash, and desperate for capital during the frantic market-grab phase. They planned to sell the main brand from the start, incubating sub-brands in advance to pull off this “golden cicada” maneuver. Whichever sub-brand performed best would become the “heir.” The flood of similar products was a way to test the market and spread supply chain risk.</p><p>After the sale, the “VanGuard” brand was handed over to ROyal’s distributor, which leveraged its own channels, supply chain, and customer service. The original VanGuard team, meanwhile, moved forward with the lean and successful Atom TeK. The original gaming media account was even renamed to Atom TeK, cutting ties with the past. They emerged from the bloodbath as a leader among the new wave of Chinese peripheral brands.</p><p>It was a brilliant move. They laid the groundwork with sub-brands, sold their core brand identity for a massive cash injection at a critical time, and shed some of the negative impressions from VanGuard’s early, less mature products. And the ROyal distributor didn’t lose either; they got a well-known brand with an established audience and have been steadily growing their market share.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*agen90PbT_spz5NkN5fo_Q.jpeg" /></figure><h3>From Barbaric Growth to Carving Up the Market</h3><p>Fast forward to early 2024. The landscape of Chinese peripherals had largely settled. A few top players like VanGuard, qom, Wolf&amp;Spider, and MCHOSE had each carved out their territory. The market began shifting from acquiring new users to competing for existing ones. A notable change was that <strong>“casual buyers” were decreasing, while “savvy consumers” were increasing.</strong></p><blockquote>“Casual buyers”<em> are those who rely heavily on platform recommendations or are easily swayed by flashy, low-quality sponsored content. </em>“Savvy consumers”<em> are those whose purchasing decisions rely heavily on in-depth reviews and research within enthusiast communities. They are generally resistant to low-effort sponsored posts. (In my experience, over 95% of all consumers, including these “savvy” ones, can’t actually distinguish between a well-orchestrated positive campaign and genuine user feedback).</em></blockquote><p>During the “barbaric growth” phase, brands carpet-bombed the internet with astroturfers, paid reviews, and even recruited college students to become “reviewers” (as a college student, I was once invited to become one, but I wasn’t as short on cash as many students are, so I declined), feeding a massive ecosystem of tech media. Over time, some of the more skilled top- and mid-tier reviewers began producing genuinely deep content to build their long-term credibility. They couldn’t just shill for any product anymore; they had to create content for the growing “savvy consumer” audience.</p><p>Review quality began to diverge sharply, and the growth of enthusiast communities produced more and more “savvy consumers” and KOLs/KOCs (Key Opinion Leaders/Customers). At this point, continuing with blatant astroturfing was like draining the pond to catch the fish — a self-destructive strategy. Brands needed a new path.</p><p>Meanwhile, the OEM factories that manufactured products for these brands were a constant threat. They could launch their own brands at any time. They had the cost advantage and the technical know-how; all they lacked was brand recognition and marketing experience. If they decided to play the same game of aggressive marketing, the established order could be shattered, just as VanGuard had done. A prime example is <a href="https://medium.com/my-valkyrie.com">Valkyrie</a> (this is their real name), an OEM for PC cooling components with very good products (I mean it, they’re really good). The new top dogs now had to defend their turf from new players.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/972/1*YGngEEeHn2VMzLH9Ok_7aA.jpeg" /><figcaption><em>A Joint Declaration to Boycott Astroturfing</em></figcaption></figure><p>And so, in March 2024, a joint denunciation titled <em>A Joint Declaration to Boycott Astroturfing</em> appeared simultaneously on the official social media accounts of several top brands. The statement accused a specific brand, MCHOSE, of “misleading consumers and slandering competitors.” The declaration, signed by VanGuard and a growing list of other brands, was a solemn pledge to “jointly resist and never use astroturfing.”</p><p><strong>This was a genius-level move.</strong></p><p>First, it turned “anti-astroturfing” into the ultimate marketing campaign. <em>You hate fake reviews? Great. We’re leading the charge against them, which makes us more “conscientious” than the brands that use them.</em></p><p>Second, once all the major players signed the declaration, it became painfully obvious who was still using astroturfing: the ones who <em>didn’t</em> sign. <strong>Any newcomer trying to use astroturfing would now be branded an “enemy of the industry,”</strong> making them an easy target for a coordinated takedown. A new brand, even one from a major OEM, was now blocked from using the old playbook. The established leaders had already built their mindshare and could now pivot to more subtle tactics like “community management” and “KOC seeding” — which are technically a form of astroturfing, albeit a far more advanced and subtle one.</p><p><strong>It’s like after everyone has struggled through a muddy pit, the first ones to get across turn around, fill it in, and erect a sign that says “Road Closed.” They not only buried the brand still struggling in the mud but also built a monument on top of it.</strong> This was the most brilliant marketing play in the 2024 Chinese peripheral market.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0uP4q0HYlOpEl-ealDEIIg.jpeg" /></figure><h3>Marketing Shouldn’t Be the Only Moat</h3><p>These two stories — VanGuard’s “golden cicada” act and the shift from barbaric to “civilized” marketing — point to a single truth: <strong>in a market without a deep-rooted technological moat, marketing is the only moat.</strong></p><p>International giants like Corsair, SteelSeries, and Alienware have struggled to adapt to the Chinese market. Razer, Logitech, and ROG, with more significant tech advantages, have doubled down on the high-end market and are holding their ground. Meanwhile, Chinese domestic brands have carved out a huge space with their price advantage. The next question is whether they can continue to optimize user experience, build better products, and eventually create their own technological moats.</p><p><strong>Marketing shouldn’t be the only moat.</strong> This requires brands not only to hold themselves to a higher product standard but also that reviewers make an effort to provide objective, insightful content, even when it’s sponsored. Of course, the ideal scenario is a rise in the collective knowledge of the consumer base itself. In China’s peripheral enthusiast circles, I’m delighted to see more and more rational, savvy consumers. I believe this trend will continue in the broader consumer tech market, and things will keep getting better (except for audio and Hi-Fi products).</p><p>Btw, when you buy a mouse, keyboard, or headphones, how much of your decision is based on the story the brand is selling, and how much is based on the quality of the product itself?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8ls2z1Swv6Xu_HdY-0Zyzg.jpeg" /></figure><p><em>Some brand names mentioned above are fictitious.</em></p><p><em>No rights reserved for this article.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9d0b5b0a9be2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hidden Marketing Wars of Chinese Consumer Electronics: Part I]]></title>
            <link>https://medium.com/@sha1rholder/hidden-marketing-wars-of-chinese-consumer-electronics-part-i-95f07e5df78a?source=rss-4c891e377534------2</link>
            <guid isPermaLink="false">https://medium.com/p/95f07e5df78a</guid>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[audio]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[marketing]]></category>
            <dc:creator><![CDATA[Sha1rholder]]></dc:creator>
            <pubDate>Sun, 16 Nov 2025 06:26:12 GMT</pubDate>
            <atom:updated>2025-11-17T00:06:20.410Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tKpS_zjaAeGImahM601N9w.jpeg" /><figcaption>Photo from FiiO, a Chinese Audio Electronics Manufacturer</figcaption></figure><p>I’ve had a front-row seat to the rise of Chinese consumer electronics. I participated in the lifecycle of audio products, from initial project proposals, R&amp;D to their final marketing and sales. As an audiophile, I was a proud supporter for quite a long time — enthusiastic about its innovation, value, and rapid growth. But beneath the surface of this booming ecosystem lies a set of uncomfortable truths I’ve been reluctant to address. Until now.</p><p><em>(A quick note: As my primary language is not English, I’ve relied on AI assistance to help refining my words. I appreciate your understanding if some nuances are lost in translation.)</em></p><h3>Chinese Audio Equipment, Centralized Marketing, and the Erosion of Consumer Trust</h3><p>In recent years, a wave of Chinese audio brands has risen with astonishing speed. Most entered the market by offering “great value,” starting with passive audio products like IEMs. The vast majority invests minimally in R&amp;D, relying instead on the mature supply chains of the Pearl River Delta to ship products quickly. Once they accumulate enough orders, they secure funding and scale up.</p><p>“HiFi from China” is jokingly referred to as “<strong>Chi-Fi</strong>” by Western audiophiles, and Chinese brands are now striving to shift the connotation behind this term from “great value” to “great quality.”</p><p>Objectively, they have shattered the halo effect carefully cultivated by Western “big-name” manufacturers. Through their products, they’ve revealed a stark truth: most consumer audio equipment from Europe, America, and Japan is absurdly overpriced, and its technical capabilities often fall far short of its marketing claims.</p><p><strong>But there’s another side to this coin.</strong> The rise of these brands is often accompanied by ethical gray zones in marketing and an erosion of consumer rights.</p><p>Recently, while reviewing a paper on HRTF in consumer equipment, I was reminded of the months I spent researching the audio industry in Shenzhen. My understanding deepened significantly, and I felt it would be a waste not to document some of these insights. <strong>So, I’ve decided to share a series of articles exploring the marketing strategies, ethical boundaries, and implications for consumer welfare among China’s emerging audio brands.</strong> This is the first installment, recounting a real-world case — a covert clash between a domestic upstart and a tech behemoth.</p><h3>The Manufactured “Objective” Review</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QaRn7N7SkCIvoJuO-d2v_Q.png" /></figure><p>To understand this conflict, you must first grasp the ecosystem of KOL (Key Opinion Leader) reviews.</p><p>In the niche world of consumer audio, brands routinely commission KOLs to produce review content for product launches. The price ranges from several thousand to tens of thousands of RMB, depending on the KOL’s follower count, audience demographics, and persuasive power.</p><p>Different KOLs cater to different audiences: budget IEMs are marketed to trend-conscious university students, while premium over-ear headphones target middle-aged men “with taste.” A KOL who commands a larger, more affluent, and more persuadable audience naturally commands higher rates. Their audience’s attention and trust are the commodities being sold.</p><p>Most review content, however, is boilerplate marketing jargon like “excellent resolution” or “wide soundstage.” In extreme cases, brands ghostwrite the entire script (a skill I, unfortunately, became very good at) — the KOL merely reads it aloud and hits publish. Whether they’re sprinkling in pseudo-technical garnish like “interpreting FR curves” (often without understanding them) or <strong>pandering to audiophiles who believe power cables alter sound</strong>, the goal is always the same: reach the broadest possible audience.</p><blockquote><strong><em>For KOLs, the brand’s PR team is the real client. The trusting consumers are merely their product.</em></strong></blockquote><p>Most PR personnel neither understand nor need to understand the product itself. And KOLs study products not for the sake of truth, but to better serve the brands paying for their endorsements.</p><h3>A Minnow in the Leviathan’s Shadow</h3><p>This dynamic came to a head in a recent clash.</p><p>Over the past few years, an emerging Chinese audio brand — let’s call it <strong>Flower Tech</strong> (a pseudonym) — expanded from budget IEMs into the TWS (True Wireless Stereo) market. By forging a deep partnership with a specialized MCN company, it systematically cultivated a cohort of “brand-loyal” top-tier KOLs across major content platforms, achieving remarkable marketing results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L5wW0puo8Z_gj1ju9CNTbQ.png" /><figcaption>hint</figcaption></figure><p>This is the standard playbook for many emerging Chinese brands: first, build credibility with “high-value HiFi,” then leverage that momentum to enter the much larger market of active audio products like TWS. Once established in the low-to-mid tier, pushing into the premium segment is the inevitable next step — setting the stage for a direct confrontation with smartphone giants and established digital brands.</p><p>In September 2025, a Chinese tech leviathan — a company most of you have heard of — launched a flagship TWS in a somewhat niche category. As usual, its PR team invited a group of reviewers for early access. Among them was a KOL renowned in the community for being “objective” and “outspoken.”</p><p>The problem? This “independent” reviewer was, in fact, closely tied to Flower Tech.</p><p>While the reviewer didn’t outright disparage the product, the video repeatedly emphasized that it offered <strong>“not a great value for money”</strong> — a phrase that, within audio communities, functions as a thinly veiled negative review.</p><p>The leviathan’s PR team, of course, expected a “rational” (read: promotional) review. They likely had no idea about the reviewer’s true affiliations. All they knew was that they had paid for a service and didn’t get the praise they anticipated. So, they pressured the video platform. The video was taken down, and the reviewer’s account was suspended for several days.</p><p>Ironically, this heavy-handed move triggered a backlash. Across Chinese audio communities, users began questioning the tech giant’s “manipulative marketing.” Meanwhile, Flower Tech — sacrificing nothing more than a brief suspension of one of its KOLs — successfully created a “reputation vacuum” around its rival’s flagship, all without ever revealing its hand. Barring unforeseen circumstances, Flower Tech is now poised to make its own entry into the premium segment.</p><p><strong>Silencing a critic, it turned out, was the best possible assist for an opponent.</strong></p><h3>(next article) The Illusion of Decentralization</h3><p>In truth, Flower Tech’s products hold no genuine technological edge — they merely offer consumers another “not-bad” option. After all, few consumers actually buy multiple TWS models to conduct side-by-side tests. Nor are Flower Tech’s marketing tactics some stroke of genius; it’s just the industry-standard centralized promotion playbook.</p><p><strong>So, why was this specific marketing play so effective? Why did this one “objective” review, planted by a competitor, manage to sway so-called “rational consumers” so powerfully?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YE5fFDq1HZhm0bDZ.png" /><figcaption>image from Defamation Defenders</figcaption></figure><p>The answer lies beneath the surface. The tactic was simple, but the strategy was not. A specialized MCN, the <strong>Chronovate HiFi Promotions Studio</strong> (pseudonym), is deeply intertwined with numerous Chi-Fi brands. It doesn’t just control KOLs; it cultivates the entire environment他们 in which they operate. It manufactures the <em>appearance</em> of a decentralized, grassroots consensus.</p><p>In my next article, I’ll delve into Chronovate’s operations and expose an even more unsettling reality: <strong>in today’s era of sophisticated “decentralized marketing”, the “rational consumer” is an illusion, and our choices are not nearly as free as we thought.</strong></p><p><em>No rights reserved for this article.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=95f07e5df78a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>