<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xml:base="https://tilde.club">
  <title>Read Write Compute</title>
  <subtitle>记录想法以供参照。</subtitle>
  <link href="https://tilde.club/~xuqi/feed.xml" rel="self"/>
  <link href="https://tilde.club/~xuqi/"/>
  <updated>2026-05-27T00:00:00Z</updated>
  <id>https://tilde.club/~xuqi/</id>
  <author>
    <name>徐栖</name>
    <email>me@expoundite.net</email>
  </author>
  <entry>
    <title>真正的兴趣</title>
    <link href="https://tilde.club/~xuqi/log/finding-the-real-interest/"/>
    <updated>2026-05-27T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/finding-the-real-interest/</id>
    <content type="html">&lt;p&gt;我的浏览器里经常有几个打开很长时间的 tab，都是我在 HackerNews 之类地方看到，但又抽不出时间读的文章。我过去会把它们一直开着，直到不得不重启电脑。现在我开始意识到，这样做并不必要。&lt;/p&gt;
&lt;p&gt;首先，人都有多了解一点信息的本能，如果一个事情看起来有点意思，我们都想看个究竟。所以过去在集市上，即使牵个猴子卖艺，也有不少人看。但以现在各种媒介和出版渠道的发达程度，即使加上完全免费这个限定条件，让一个受过高等教育的人多少会有点兴趣的内容，也远远超过了一个人能阅读的总量。&lt;/p&gt;
&lt;p&gt;其次，各种推荐系统的原材料只能是平台上的内容和用户的行为。这就决定了它们推荐的内容很难和你真正的兴趣“对齐”。用户能看到的绝大部分内容，都是为获得流量而不是个人的兴趣优化的。而行为上，你可能只是很偶然地点进一个页面，甚至像我有时候做饭腾不出手一样，让视频自动播放了一下，却都会被推荐算法当成依据。&lt;/p&gt;
&lt;p&gt;第三，即使是一些公认的好内容，什么上万转发、上百万浏览，甚至是各种评测和欣赏者都一致说好的东西，也未必就能踩在你的兴趣点上。今天 Ars Technica 游戏栏目的一篇文章里，作者就承认他花了好几年，和一百个小时的游戏时间，也&lt;a href=&quot;https://arstechnica.com/gaming/2026/05/the-witcher-3-is-a-good-game-but-that-doesnt-mean-you-have-to-like-it/&quot;&gt;一直没能把《巫师3》打完&lt;/a&gt;。我也一样，只能通关很少数的 RPG 游戏，即使是《博德之门》和《旧共和国武士》这样公认的神作，也一直打不过第一章。&lt;/p&gt;
&lt;p&gt;强迫自己把一个东西看完，可能最后只是减少了你发现自己真正会感兴趣的东西的机会。&lt;/p&gt;
&lt;p&gt;我发现我个人就只对研究做一些事情的方法和工具感兴趣。一个事情如果能自动化，让它的边际成本变得很小，我就有兴趣研究这个自动化的过程。目前看来，这是我比较有兴趣做的事情。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>“AI”律所大失败</title>
    <link href="https://tilde.club/~xuqi/log/dotai-suffix-doesnt-make-you-competent/"/>
    <updated>2026-05-19T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/dotai-suffix-doesnt-make-you-competent/</id>
    <content type="html">&lt;p&gt;&lt;em&gt;D’Ambrosio v. Meta&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;在这场诉讼中，芝加哥男子 Nikko D’Ambrosio 起诉 Facebook 的运营公司 Meta，声称后者给一篇给他“开盒”的帖子以流量支持。他聘用的律所是 MarcTrent.AI。&lt;/p&gt;
&lt;p&gt;MarcTrent.AI 的创始人 Marc Trent 信心满满，称自己的律所一切都用 AI 做，连 Meta 庞大专业的律师团队也可以挑落马下。在案子被地区法院退回后，D’Ambrosio 再次上诉到第七巡回法院。法官&lt;a href=&quot;https://arstechnica.com/tech-policy/2026/05/legal-fail-dont-use-ai-to-sue-facebook-users-for-calling-you-a-bad-date/&quot;&gt;认为&lt;/a&gt;，诉状不仅写得“轻浮无聊”，没能提出有效的法律诉求，而且充斥着错误和捏造的法条及判例，显然是“误用生成式人工智能”的结果。&lt;/p&gt;
&lt;p&gt;这家律所可能将会因三方面原因受罚：在证据真实性上误导法庭；使用 AI 引入虚假和具有欺骗性的案例，以及滥用上诉权。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>创建一个新 tag：anti-LLM</title>
    <link href="https://tilde.club/~xuqi/log/starting-new-tag-anti-llm/"/>
    <updated>2026-05-16T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/starting-new-tag-anti-llm/</id>
    <content type="html">&lt;p&gt;我对生成式 AI 的接受度以时间为横轴画出来，是一个驼峰形的曲线。在最初的怀疑之后，有一段时间我也天天都用这类工具，包括这个博客最初的代码都全是 vibe coding 写的，还一度&lt;a href=&quot;https://tilde.club/log/genai-energy-economics/&quot;&gt;本地部署&lt;/a&gt;&lt;sup class=&quot;footnote-ref&quot;&gt;&lt;a href=&quot;https://tilde.club/~xuqi/log/starting-new-tag-anti-llm/#fn1&quot; id=&quot;fnref1&quot;&gt;[1]&lt;/a&gt;&lt;/sup&gt;过 Gemma 4。但在这个相对来说高强度使用生成式 AI 的阶段过后，我就转到了对这一技术的反对立场，简述如下：&lt;/p&gt;
&lt;p&gt;第一，如上面链接的篇目所说的，最终决定 token 的交换价格的是它的生产成本，在这里，成本最终是由人类开采和利用能源的效率决定的。而 LLM 是一种效率如此低下的利用能源的方式，我认为在除了 vibe coding 外的大部分场合，它的成本在褪去补贴之后都将显著高于已有的自动化手段。它的持续大规模应用，只能建立在对环境和部分人的持续大规模剥削上。&lt;/p&gt;
&lt;p&gt;第二，即使 vibe coding 等少数应用中生成式 AI 带来 x5、x10 的效率提升，我们又到哪里去找这多出来的 400% ～ 900% 的产品的市场？在经济史意义上，我认为生成式 AI 不过是资本主义经济在重复它试图不断并无限地提高生产力、降低生产成本、实现利润增长的固有行为模式，而这又必然带来消费能力的下降和新一轮的市场崩溃以及萧条。&lt;/p&gt;
&lt;p&gt;第三，现在各种 pro-GenAI 的叙事，无一例外都是站在资方的立场上。反对者都是劳方（包括我这样的 &lt;em&gt;de facto&lt;/em&gt; 失业者）。如同卢德派当年&lt;a href=&quot;https://www.marxists.org/chinese/hobsbawm/mia-chinese-hobsbawm-1952.htm&quot;&gt;砸毁&lt;/a&gt;的机器一样，生成式 AI 是资方把利益进一步转移给自身，夺取劳方权益的手段。在这点上，它和其前身，相对来说更具技术中立性的机器学习技术有本质的区别。&lt;/p&gt;
&lt;p&gt;第四，生成式 AI 在多数场合给出的输出，没有反映其训练资料的积累过程和来源，这除了著作权法律上的侵权问题外，也导致人的进一步原子化。仍以 vibe coding 为例。过去，当一个人遇到自己无法解决的编程问题，他会上 Stack Overflow 和邮件列表提问，在这个过程中他可能会了解到一个新的领域，更重要的是可能发现一个新的社群，至少可能认识一个这方面有知识、有见地的人。现在，使用 vibe coding 工具，表面上他得到了同样的结果，但人际互动这个环节被取代了。原本知识工人之间互相帮助的叙事，被转换成工具能力和个人能力的叙事，导致每个人都更加孤立。这是正在发生的事情。许多开发者正在离开 GitHub，转向其他平台甚至技术，正是对 GitHub 在上述过程中的推动作用的反抗行为。&lt;/p&gt;
&lt;p&gt;在中文媒体中，对生成式 AI 的报道几乎和 OpenAI、谷歌、Anthropic 的宣传口径完全一致，即以声称生成式 AI 技术的必然性和必要性为主，辅以用科幻式的担忧为幌子，实际上夸大生成式 AI 技术能力现状和前景的叙事。但不论民间或是机构中，都还有很多反对的声音，他们的声量值得被放大。&lt;/p&gt;
&lt;p&gt;而且不论这一波技术浪潮的结果如何，它带来的观念争执、情境推演，各种看好或看空的预测，人们对过去和未来劳动关系的反思，值得被记录。在已经有很多人表达这项技术的积极前景的今天，我想用这个 tag 来记录那些消极的、悲观的、犹豫的，但不一定是不正确的话语。&lt;/p&gt;
&lt;hr class=&quot;footnotes-sep&quot; /&gt;
&lt;section class=&quot;footnotes&quot;&gt;
&lt;ol class=&quot;footnotes-list&quot;&gt;
&lt;li id=&quot;fn1&quot; class=&quot;footnote-item&quot;&gt;&lt;p&gt;此篇没有技术细节，全是吐槽。 &lt;a href=&quot;https://tilde.club/~xuqi/log/starting-new-tag-anti-llm/#fnref1&quot; class=&quot;footnote-backref&quot;&gt;↩︎&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/section&gt;
</content>
  </entry>
  <entry>
    <title>McPhee 怎样搜集资料</title>
    <link href="https://tilde.club/~xuqi/log/mcphee-on-being-ignorant/"/>
    <updated>2026-04-29T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/mcphee-on-being-ignorant/</id>
    <content type="html">&lt;blockquote&gt;
&lt;p&gt;He has found that imagining he knows a subject is a disadvantage, for that prejudice will limit his freedom to ask, to learn, to be surprised by unfolding evidence.&lt;/p&gt;
&lt;p&gt;[...]&lt;/p&gt;
&lt;p&gt;Informants may be timid or hostile, unless they feel superior or equal to their interviewer. By repeating and even fumbling their answers, McPhee encourages people to embroider a topic until he has it entire.&lt;/p&gt;
&lt;p&gt;-- Introduction to &lt;em&gt;The McPhee Reader&lt;/em&gt;.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;把自己当成一张白纸，完全不带任何预设地去接收新的信息，这样才能在采访和资料搜集工作中获得最全面的认识。&lt;/p&gt;
&lt;p&gt;McPhee 在自己的报道中几乎从不出现，似乎把 ego 压缩到极小。他表现得对自己正在采访的话题一无所知，以致一些被采访者以为他不太聪明。但也正因为这样，他笔下的人物总是能大放异彩。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>我将心安理得地待在我的舒适区里</title>
    <link href="https://tilde.club/~xuqi/log/stay-in-my-comfort-zone/"/>
    <updated>2026-04-27T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/stay-in-my-comfort-zone/</id>
    <content type="html">&lt;p&gt;作为内向者，我曾经在工作中不断被提醒，要走出舒适区，不然职业不会有发展，个人不会有进步。如今看来，这些话不过是对个人的操纵——过去十多年里我都在试图跳出舒适区，变成新的我，但“到了年龄”，没有人会因为我做出的这些改变多看我一眼，他们只会觉得这个人的简历好花，不是他们要的那种从2022年起就深耕 AI 的人才。&lt;/p&gt;
&lt;p&gt;在自然界中，每一种动植物都有一个最适合自己的生境。各种气候、资源条件定义了这个生境。当这些条件发生变化的时候，动植物便会用自己的方式去适应。植物通过各种生长激素改变自身对水、光、热的需求，动物则往往一走了之，走不了的，才爬到树上，钻进水里，以求一时的平安。但它们试图去做的，都是回到舒适区，而不是离开舒适区。而近年来造成动植物的生境变化的，越来越多地是由于人类引起的气候变化，或者人类出于自身需求而侵占它们的栖息地和资源，更深一层的原因还是人类希望扩大自己的舒适区。&lt;/p&gt;
&lt;p&gt;对任何生物来说，待在舒适区里都是必要的，因为只有这样才有可能延长生存的时间，才能把能量和资源更多地用在生长和繁衍，而不是用在被迫适应环境上。在这个基础上，个体才有去探索新领域、培养新能力的可能。&lt;/p&gt;
&lt;p&gt;我能通过面试，在一个岗位上工作，就说明这是我的舒适区，那些让我从这个舒适区走出去却连诱饵都懒得提供的要求，比如让我去带货做直播，无非是提出要求的人达成两个目的的一种手段。第一个目的，是他们借此表现出对上位者的认可与服从，借着牺牲手下，可以更安全地让自己留在舒适区里。第二个目的，这类似于以学游泳为借口，让我从开得半快不慢的船上跳进海里，表面上是对我天赋和能力的检验，实际上是看我够不够温顺服从。&lt;/p&gt;
&lt;p&gt;这种在职业语境中盛行的舒适区的说法，不过是用“不肯离开舒适区”中暗含的对人格的羞辱，迫使人到一个很可能并不适合他的环境中自生自灭。之所以要个人这么做，也无非是组织和企业不愿意承担跟上市场和趋势的成本，比如遣散费和新人的培训，于是把转型甚至转轨的重任丢给个人。情节相同、台词不一样的剧本，其实我们的父辈也在更大规模上出演过。&lt;/p&gt;
&lt;p&gt;而这一切幕后的导演也没有变过。他在对自己的宣传中，喜欢强调个人对自己的境遇负有完全的责任，混得不好是因为对自己要求不够严格。在社会层面，他的影响是构成一种单向度的评价标准：你存在的价值在于你身价几何，而这又取决于你对你公司的 bottomline 构成了什么样的影响。这个导演的名字就叫 neoliberalism。&lt;/p&gt;
&lt;p&gt;所以，当有人批评我不肯跳出舒适区的时候，他批评的其实是我不肯为公司和组织的 bottomline 扭曲自己，不肯以信仰之跃的姿态跳进即使拼命挣扎也很可能无法变成良田的泥淖。那泥淖可能对另一些人来说是旷野甚至迦南，但公司更愿意为了降低 costs 而劝我承担起新自由主义下的个人责任，把自己变成看似多面实际上更加单调的角色——一个为了公司利益鞠躬尽瘁，哪里需要拧哪里，哪怕面目全非也在所不惜的螺丝钉。&lt;/p&gt;
&lt;p&gt;如果真的认同了这种把服从安排当成个人能力乃至个人价值的叙事，个人可能会永无宁日，只能跟着公司的安排去扮演一个又一个角色，主观上还觉得自己适应能力挺不错。让别人定义什么是自己的舒适区，一个人眼里就只有自己不愿放弃又不好意思承认的现状和那个人家非要他跳、不跳不是英雄好汉的泥坑。但其实一个人有很多的面向，在这个舒适区待得无聊了，或者别人威胁不跳泥坑就不让待了，大可以在另一个维度寻找自己的舒适区，同样有可能找到一片新天地。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>认识我自己</title>
    <link href="https://tilde.club/~xuqi/log/knowing-thy-self/"/>
    <updated>2026-04-26T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/knowing-thy-self/</id>
    <content type="html">&lt;p&gt;最近我决定无限期停更播客，觉得身上卸下了千斤重担，连之前纷纷变白的头发也一根根黑回去了。之前每天抓耳挠腮想选题，又因为不论怎样换形式也激不起一点水花而苦恼，这事做得极其痛苦。偏偏社交网络上又不断有人说，要做成一件事情，最重要是熬过这段没有反馈的时间云云。现在我终于想通了，事情除了半途而废的，还有就不应该由我来做的，播客不过是这一大类事情中的一件。&lt;/p&gt;
&lt;p&gt;这一大类事情，就是试图讨好大众。做自媒体、搞大众书出版、以数据为目的混社交网络，都可算在这一类事里。&lt;/p&gt;
&lt;p&gt;我早就应该认识到的，是我的脑回路和大众的并不一样，或者至少我不愿意承认这一点。我关心的并不是大众思想市场里会讨论的问题，也不是人们在更个人的空间里会去思考或关心的问题。&lt;/p&gt;
&lt;p&gt;我关心的大概是宇宙和社会中那些 amoral 的规律，以及个人应该怎样以道德的方式，在这些规律下生活。在这两个尺度相差极大的层面之间，我必须承认，没有经过充分讨论和书写的联系。这个探询的过程，必然是不足为外人道的。外人既不会懂，也不感兴趣。&lt;/p&gt;
&lt;p&gt;反过来说，我也不觉得自媒体上的站队和表态，出格与炫耀，能让人保持长久的兴趣。这些都不过是维持 engagement 的调味料而已。没有想法的支撑，它们没有口语文化之外的意义。但当今人们似乎最能接受这样的表达方式。&lt;/p&gt;
&lt;p&gt;而我不能以这样的方式思考，自然也无法按照这样的逻辑去表达。所以我会彻底放弃通过向大众表达，让大众理解我心中所想的期望。我还会继续写东西，也会以博客之类的形式发表出来，但是不会在这个过程中再考虑别人能不能看得懂或者愿意转发之类的问题。写下来纯粹是为了让自己在这个过程中能把问题想得再清楚深入一点。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>生成式 AI 的能源（不）经济性</title>
    <link href="https://tilde.club/~xuqi/log/genai-energy-economics/"/>
    <updated>2026-04-11T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/genai-energy-economics/</id>
    <content type="html">&lt;p&gt;最近我在本地部署了 Gemma 4，作为没网时的 vibe coding 备选。我是在笔记本电脑上部署这个模型的。当我试着问它一个简单的非编码问题，我的电脑风扇开始轰鸣——过去在用云上模型的时候，我一直都是用的节能模式，都忘了电脑上还有风扇。&lt;/p&gt;
&lt;p&gt;今天我又测试了一下，让 Gemma 4 推荐歌单和拟一份工作邮件，这两个非编码问题分别用了161秒和163秒，工作邮件还没完全按照我的要求来。再试着让它用 Bash 列出我硬盘上最大的10个文件，不知道这次是碰到了它的专长领域还是热身完毕的缘故，只用了124秒，几乎快了四分之一。整个过程中，模型都用上了我电脑的全部 CPU 和 GPU 时间。&lt;/p&gt;
&lt;p&gt;任何考虑使用生成式 AI 来提高生产力的企业，都应该让管理层在本地部署一个大模型，并且实际测量一下，当大模型在执行各种管理层希望交给它的任务时，总共花费了多少时间和电力。&lt;/p&gt;
&lt;p&gt;当通过网页或者 CLI 客户端使用云上大模型的时候，用户对每一次 query 的能源消耗缺乏感知。这使人们容易产生错觉，好像每个月200美元的生成式模型订阅真的可以取代每个月成本在10000美元以上的程序员。没有体现在订阅费用中的，是生产、调配、输送和转化能量的成本。包括算力成本中的非能源部分，即设施的折旧，归根到底也在很大程度上由能源的使用来决定。如果把这些成本考虑进来，再加上公司仍然需要人员来核实生成式 AI 的输出这个事实，AI 能否实现生产力的数量级放大或者成本的可观削减，要打上一个问号。&lt;/p&gt;
&lt;p&gt;当前这些成本因为外部化或者补贴，没有被反映在生成式模型的用户订阅价格中。比如美国的大模型厂商和一些能源公司签订的协议，实际上是把成本交给普通消费者分担，人们只觉得自己家的电费涨了，却不知道它和每天都用的 AI 有关系。又比如 OpenAI 和微软签订了协议，实际上是接受微软对其算力和能源成本的补贴，但即使如此，OpenAI 也难以打平收支，还需要继续找 AWS 这样的合作方进一步摊薄成本，哪怕冒着被微软&lt;a href=&quot;https://www.reuters.com/technology/microsoft-weighs-legal-action-over-50-billion-amazon-openai-cloud-deal-ft-2026-03-18/&quot;&gt;起诉&lt;/a&gt;的风险。&lt;/p&gt;
&lt;p&gt;加上一些澄清需求、调整格式的对话，我和 Gemma 4 的这次会话一共让模型工作了15分钟。以笔记本平均功耗65W和我国平均电价0.7元/度计算，我这次实验的电力终端成本大约是1分钱多一点。但这只是一个26B参数的“小”模型，而且它给我的结果还几乎都不能用。实际上能干活的商用云端模型，即使只是处理文本输入输出任务，其成本都可能数倍于此。&lt;/p&gt;
&lt;p&gt;《麻省理工科技评论》去年5月的文章里做了个&lt;a href=&quot;https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/&quot;&gt;测算&lt;/a&gt;。一个以半业余方式使用 AI 的用户，用 AI 生成10个版本的一张图像，一段5秒的视频，至少会消耗大约2.9度的电力——足够驱动一辆轻型电动车行驶100多公里。仅从能源成本的角度，如果完全没有外部化和补贴，使用 AI 的成本还不如一个在台式机上工作的新媒体编辑，他干一天可能只同样消耗两度多电，产出通常还不止一张图和一段视频。要是他在笔记本电脑上工作，电力成本至少还能降低一半。&lt;/p&gt;
&lt;p&gt;上面这个估计还比较保守，因为 GPT-5 等闭源模型的参数量可能更大。反过来说，如果一个新媒体编辑完全使用 AI 工作，一天出二三十秒视频和一百多版图像，这一天的 AI 能源成本可能高达几十元人民币，而他和公司的几名同事可能还在共用一个月费200美元的订阅——也就是说仅从产生输出的能源消耗角度，大模型提供商都是亏本的，更不要说算上高昂的模型训练和公司其他运营成本了。&lt;/p&gt;
&lt;p&gt;这些成本没有直接地反映到使用模型的费用中，是因为外部性和补贴，但经济系统是一个整体，能源的生产和供应无法像 AI 部署一样放大和规模化，这些成本一定会以某种形式反馈给经济中每一个参与者。比如我们已经看到的，内存价格的上涨导致电脑整机价格的上涨，还有个人不容易看到的，比如能源价格上涨导致航运价格上涨，又导致食品价格的上升，普通消费者的钱袋子更紧，导致你的获客成本上升。&lt;/p&gt;
&lt;p&gt;而大模型提供商的补贴不会永远持续下去，正如不存在永动机。或迟或早，提供商们都将不得不让定价更真实地反映实际成本。到那时，我们可能会发现生成式 AI 带来的“降本增效”作用是非常有限的。&lt;/p&gt;
&lt;p&gt;和蒸汽机、电力、工业自动化这些科技进步不同，AI 产生在一个生产力已经大范围过剩的时代，这或许是我们在大模型技术已经问世将近10年之后，还在摸索到底能用它来干吗的原因。在全球南方，AI 并不能满足那些“线下”的需求。在把 AI 应用于农业生产提高产量之前，还有一大堆基础设施和体制的问题需要在那些地方先得到解决，否则哪怕多生产出10倍的玉米，也只能大量拉到当地的市场上，摧毁当地的价格体系，导致农民再也不想种玉米。&lt;/p&gt;
&lt;p&gt;而在发达国家，微软、谷歌在用大量的算力在 Outlook 和 Google Docs 里帮你起草邮件和公关稿，不过是在高效地生产早已&lt;a href=&quot;https://hbr.org/2023/05/reducing-information-overload-in-your-organization&quot;&gt;过剩&lt;/a&gt;的文本而已。&lt;sup class=&quot;footnote-ref&quot;&gt;&lt;a href=&quot;https://tilde.club/~xuqi/log/genai-energy-economics/#fn1&quot; id=&quot;fnref1&quot;&gt;[1]&lt;/a&gt;&lt;/sup&gt;你用 AI 瞬间生成邮件和文档，另一端又瞬间让 AI 把它总结成一句话，并不会让你和对方能在一天里处理更多的事情，它仅仅是让人用更多的时间生成和处理文本而没有提高生产力——也就是促进内卷。&lt;/p&gt;
&lt;p&gt;而以 Agent 的方式使用 AI 的能力，需要网络、自动化方面的基础设施作为前提，更需要人把过去放在自己脑袋里的隐式知识转换成 skill、harness，或者还没出现的其他什么词儿所指代的文本。包括 vibe coding 在内，AI 这些新出现的用法仍然需要一个人来判断和评估真实世界的状态和变化，但 AI 本身并没有提供可靠、高效地让这个人来做这些工作的工具。&lt;sup class=&quot;footnote-ref&quot;&gt;&lt;a href=&quot;https://tilde.club/~xuqi/log/genai-energy-economics/#fn2&quot; id=&quot;fnref2&quot;&gt;[2]&lt;/a&gt;&lt;/sup&gt;人的脑力仍然是瓶颈，而这一次，AI 并没有像计算机的发明那样提供可以带来实际改变的“&lt;a href=&quot;https://www.themarginalian.org/2011/12/21/steve-jobs-bicycle-for-the-mind-1990/&quot;&gt;自行车&lt;/a&gt;”。&lt;/p&gt;
&lt;p&gt;一方面，我们并没有发现在 AI 能做的事情中，有什么是我们当下就有相对于生产能力有以10倍计的需求的，可能在今后的很长一段时间内我们都发现不了。同时对 AI 的产出做把控这个人，能获得的效率提升通常小于 5x。最近 cURL 这个在所有消费级 IT 产品中几乎都要用到的项目是个典型的例子，说明人的脑力和精力才是瓶颈：虽然 vide coding 带来了大量的 bug report 和代码修改方案，但不经审核采纳这些修改建议中的任何一个，都可能导致涉及数以亿计的用户无法正常上网，项目的维护团队又找不到熟悉项目的新维护者加入，只能暂停接受新的 bug report。可以想见在软件开发中也是如此，虽然编码能力提高了，但是要提高做功能、修 bug 的能力，还是得多招三个臭皮匠或者一个诸葛亮。综合起来，当前生成式 AI 能够带来的生产力提升最多也就是 x1.5、x2，而且集中在软件开发这个从计算机的自动化能力受益最多的领域。当 AI 算力（能源消耗）的外部性消失，我们可能会发现如果任何人想以目前的用户支付费用提供同样的生成式 AI 能力，他都只能当掉底裤然后去裸泳。&lt;/p&gt;
&lt;p&gt;另一方面，这种生产力的提升即使真的存在，也没有被整个社会享受到，受益的只是岗位越来越少的公司里越来越“资深”的一小部分人。同样卖70美元的游戏，在 AI 的帮助下开发投入从1万人月变成了2000人月，当然挺好，但这后面可能是公司原本聘用的1000人中有800个都要被解聘或者降薪，~增加收益~节省成本的大头被少数人拿走了。而玩家并不会因为开发周期的缩减，在一年里买更多的游戏。相反，这些被解雇的人自己可能还是游戏玩家，他们一段时间内不会再买老东家或者其他厂商的游戏了，整个市场都会缩水。生成式 AI 的应用虽然提高了产能（capacity），却让消费能力大大降低了。这简直就是我们再熟悉不过的制造大萧条的配方，那是集中在少数人手中的金钱的味道。这个现象加上现在由整个社会各行各业承担的能源和基础设施外部性，构成了 AI 崩盘最可能的促动因素。&lt;/p&gt;
&lt;hr class=&quot;footnotes-sep&quot; /&gt;
&lt;section class=&quot;footnotes&quot;&gt;
&lt;ol class=&quot;footnotes-list&quot;&gt;
&lt;li id=&quot;fn1&quot; class=&quot;footnote-item&quot;&gt;&lt;p&gt;我们甚至不能说它在生成信息。信息这个概念暗示着认识和判断，这些都是生成式 AI 不具备的能力，它的输出只是看起来像那么回事，正确、有用与否需要人的判断，也就是说实际提供信息的还是人。 &lt;a href=&quot;https://tilde.club/~xuqi/log/genai-energy-economics/#fnref1&quot; class=&quot;footnote-backref&quot;&gt;↩︎&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id=&quot;fn2&quot; class=&quot;footnote-item&quot;&gt;&lt;p&gt;以当前生成式 AI 最具备实用价值的 vibe coding 为例，它大量借助的是已有的范式和命令行工具，也就是说其实只是在帮你写命令和脚本，而且还需要你来确认这些输出不会产生灾难性的结果。这个确认和修订的过程需要的时间，很多时候并不比人自己琢磨出结果的时间短多少。 &lt;a href=&quot;https://tilde.club/~xuqi/log/genai-energy-economics/#fnref2&quot; class=&quot;footnote-backref&quot;&gt;↩︎&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/section&gt;
</content>
  </entry>
  <entry>
    <title>The scarcity in the LLM era</title>
    <link href="https://tilde.club/~xuqi/log/the-scarcity-in-the-llm-era/"/>
    <updated>2026-03-20T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/the-scarcity-in-the-llm-era/</id>
    <content type="html">&lt;p&gt;I&#39;m not a designer. This does not change the fact that I can build professional-looking interactive prototypes in minutes using an AI agent.&lt;/p&gt;
&lt;p&gt;Google has come up with a new vibe-designing tool called Stitch. From the demo, it seems to render even my half-ass &amp;quot;design&amp;quot; process antique and dumb. Why am I still generating prototypes page by page, while I can just tell Stitch what kind of app I am building and it instantly creates all the pages one can expect in such an app, along with complete design system documentation that my coding agents can use?&lt;/p&gt;
&lt;p&gt;Yet this does not change the fact that I know very little about design, either. I guess the real problem is that it makes even less sense for me to know more about design now.&lt;/p&gt;
&lt;p&gt;The same is happening to everyone who knows Gemini is not just the constellation and Claude is not only the name of their French neighbour. There was no way in the past that a programmer also knew like the back of their hands how a hedge fund manager worked, or vice versa. But this all-trade knowledge is expected of everyone now. At least from where I stand, the job market expects many candidates know everything from programming to product management to how specific industries (aviation, health care, robotics, logistics, etc.) operate.&lt;/p&gt;
&lt;p&gt;I guess the rationale behind this is &amp;quot;at least the programming part is taken care of&amp;quot;. Now, design skills are added to the &amp;quot;taken care of&amp;quot; list.&lt;/p&gt;
&lt;p&gt;Now, as everyone claims they can do this and that (with a tongue-in-cheek &amp;quot;or at least I know what to ask ChatGPT&amp;quot;), it becomes really difficult for anyone who is not in a school, college or university to learn how or why a thing is.&lt;/p&gt;
&lt;p&gt;And know-how and know-why are going to be in short supply when all human knowledge are stored, coded and embedded in all sorts of models.&lt;/p&gt;
&lt;p&gt;A hedge fund manager is now able to use an agent to make an agent that embodies their knowledge in the financial market, but no one can learn anything by watching this agent work. And the agent would not be able to change itself to come up with new solutions when the market environment changes. To do that, one would have to know how various stakeholders and market components interact. A large part of that know-how is not possible to be encoded in prose or program. Does your LLM agent know who to call when the market melts down, or when a panic sell crashes your asset price?&lt;/p&gt;
&lt;p&gt;This can happen in any sector or industry. At best, what you have is a text generator that mimics what a professional would say when presented with an expected problem. It is not impossible that such a generated answer ignites a eureka moment that helps you come up with a genuine solution, but that scenario is not guaranteed, and even impossible if you don&#39;t understand the hows and whys in your domain.&lt;/p&gt;
&lt;p&gt;In times of crisis, you need someone who can think out of the box and also know what&#39;s at stake. Neither is possible for generative LLM agents. And would not be possible if a real person has had no opportunity to know how and why things are done in the way they are done.&lt;/p&gt;
&lt;p&gt;Even if not in such a dire situation, not knowing hows and whys stymies innovation. Yes, vibe-coding can make you an app in no time, but don&#39;t expect it to come up with genius touches like Quake III&#39;s fast inverse sqrt operation. And Stitch (to a even bigger extent, Figma Make) cannot churn out something that is not routine, bland, soulless, although admittedly still pleasing to the eyes.&lt;/p&gt;
&lt;p&gt;It is tempting for organizations and managers to vertically merge positions in expectation of a fullstack everybody, given what LLM agents can do today. Well, don&#39;t, unless you honestly think the job can be done by a robot. For any person working on such a position, they will find it difficult to know anything profoundly beyond processes and routines. Your product quality won&#39;t improve over time and cannot adapt when the market changes.&lt;/p&gt;
&lt;p&gt;Corollary: for individuals especially those who desperately need a job in today&#39;s labor market, don&#39;t heed to &amp;quot;opportunities&amp;quot; that require you to do work that used to be what we call professional work and you don&#39;t grasp enough how or why, including but not limited to coding, designing, law services, finance, and HR. You won&#39;t be able to hone your existing skills and would not know more about those professions.&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>McPhee 写作法和 Emacs (Howm)</title>
    <link href="https://tilde.club/~xuqi/log/mcphee-emacs/"/>
    <updated>2026-03-12T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/mcphee-emacs/</id>
    <content type="html">&lt;p&gt;&lt;a href=&quot;https://jsomers.net/blog/the-mcphee-method&quot;&gt;The McPhee Method&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;HN Discussion: &lt;a href=&quot;https://news.ycombinator.com/item?id=45013535&quot;&gt;https://news.ycombinator.com/item?id=45013535&lt;/a&gt;&lt;/p&gt;
&lt;h2&gt;McPhee 写作法的简单总结&lt;/h2&gt;
&lt;ol&gt;
&lt;li&gt;
&lt;p&gt;收集笔记&lt;/p&gt;
&lt;p&gt;写作者先做完写作需要的全部&amp;quot;报道&amp;quot;工作。&amp;quot;报道&amp;quot;在这里是资料搜集和研究过程的统称，包括阅读、采访、现场考察等等一系列工作。McPhee（以下M）会用大量时间来做这个工作，而且是用参与式观察来搜集信息。&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Good facts are rare, so you must give yourself enough time to acquire them.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;这阶段的多种信息，都要保存到笔记中。有的是用自己的话，书中的话则原样摘录，采访录音转成文字后也放进笔记里。这一步不需要做任何组织整理工作。有人用一个巨大的 Markdown 文件来保存这些笔记。&lt;/p&gt;
&lt;p&gt;这个过程中，我们实际上在告诉自己什么是重要（noteworthy）的。当前的写作课题会让你自然地注意到各种和它有联系的事物和想法。一些作家在这一步积累的笔记就会达到整本书的文字量。&lt;/p&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;归档&lt;/p&gt;
&lt;p&gt;在这个阶段，McPhee 会把所有笔记都读一遍。这个过程中他可能会产生一些新的想法，同样会记录在新的笔记中，并和所有这个选题下的笔记保存在一起。阅读笔记不再让他产生新的想法时，他会试着写出文稿的前几段，找到文章的组织角度。&lt;/p&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;整理&lt;/p&gt;
&lt;p&gt;关于如何组织材料和文章的想法会开始浮现。最后的文章将由章节、场景和重要场面（set pieces）组成。McPhee 会给这些组成元素分别编码，并在笔记卡片上标注这些编码，说明它们的用处。除了把笔记放进预设的框架里，有时候也可能发现一些笔记可以组成一个意义单元。&lt;/p&gt;
&lt;p&gt;McPhee 对所有整理到各个主题里的笔记还会再用同样的方式整理一遍：重读笔记，感受它们的节奏，安排顺序，标注编码。&lt;/p&gt;
&lt;p&gt;上面这两步可以大大减轻写作的心理负担，因为文章的立意不需要作者预先设定，而是在阅读整理笔记的过程中逐渐发现的。&lt;/p&gt;
&lt;p&gt;这个过程完成后，文稿的两层结构就随着笔记的定位而初步确定了。上面较粗的一层是文章的章节结构，下面较细的一层是叙事的起伏节奏。&lt;/p&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;撰稿&lt;/p&gt;
&lt;p&gt;材料准备和整理完毕后，撰稿的过程就会顺畅许多。不过对于一个场景或者一段表述，完整、流畅的句子也可能要花很多工夫才能打磨好。这时可以用&amp;quot;伪写&amp;quot;的办法，先把大致要表达的意思写在方括号【】里，回头再参照相应的笔记内容来措辞。&lt;/p&gt;
&lt;p&gt;感觉这个很像 HtDP 和（似乎是）Think Python 里教的构造函数的方法：先把函数的输入输出想好，再用注释的形式写出实现步骤，然后逐一写出真正的实现。&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;So much of writing is managing your own emotions. The virtue of &amp;quot;pseudowriting&amp;quot; is that it helps you preserve hope for as long as possible---hope that what you will eventually put in place of those square brackets will be good. Hope keeps you coming back: it is more pleasant and low-stakes to pseudowrite than to fix actual language into the draft; and it is less daunting to return to a document where it feels like all that&#39;s left is for you to fill in some blanks. Do that enough times and you will, in fact, end up with something you can read top to bottom.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;blockquote&gt;
&lt;p&gt;If you&#39;re interested in the McPhee method, you should start by reading William L. Howarth&#39;s introduction to &lt;em&gt;The John McPhee Reader&lt;/em&gt;, then the McPhee articles themselves, and finally McPhee&#39;s own &lt;em&gt;Draft No. 4: On the Writing Process&lt;/em&gt;.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;h2&gt;Org-mode&lt;/h2&gt;
&lt;p&gt;这篇博客的作者 James Somers 写了一个 Org-mode 的小扩展，方便把不同 heading 中的笔记 refile 到用字母缩写编码的章节&amp;quot;bucket&amp;quot;中。除此以外，Org-mode 本身提供的各种功能似乎都让它非常适合用于 McPhee 写作法。在不同的 heading 间移动本来就是 Org-mode 最常用的操作之一，调整内容的顺序更是它的绝技。&lt;/p&gt;
&lt;p&gt;有心折腾的话，还可以让 capture 流程更自动化一些，省掉很多复制粘贴的操作。&lt;/p&gt;
&lt;p&gt;之前我觉得一条笔记不能同时出现在两个 heading 下是个功能缺陷，但仔细一想，在一篇文章里本来就应该避免一件事情说两次，这个限制反而能迫使我想清楚应该怎样使用材料。相比 Lem、TiddlyWiki、Notion 等这反倒成了优点。&lt;/p&gt;
&lt;h2&gt;Howm&lt;/h2&gt;
&lt;p&gt;对于前两个步骤，即快速搜集信息而不加组织，以及把笔记放入不同的&amp;quot;bucket&amp;quot;中，howm 对于懒得折腾的人来说更有优势。各种缩写标签可以用 goto link 来实现，而 from link 和关键词系统可以用来说明各个 bucket 具体是什么意思。&lt;/p&gt;
&lt;p&gt;Howm 还支持多级 menu，我们可以在 &lt;code&gt;C-c , ,&lt;/code&gt; 呼出的菜单中添加到次级菜单的链接，把某个写作计划下的笔记按照多种方式列出来。理论上说，也可以按照特定的排序原则排列。&lt;/p&gt;
&lt;h2&gt;小结&lt;/h2&gt;
&lt;p&gt;McPhee 写作法相当值得一试。目前看起来，还是 Org-mode 方式最可行。每个写作项目可以放在一个 Org-mode 文件中，从笔记到终稿都可以保存在里面。&lt;/p&gt;
&lt;p&gt;但用 Howm 来做最初的笔记工作也不错，如果 tag 只打到项目层面的话，就不会有太多名字冲突的问题。第二步本来也要把笔记通读一遍并做取舍，可以在这一步再把要用的笔记放进 Org-mode 文件里。&lt;/p&gt;
&lt;p&gt;另外，结合 ox-hugo 或 orginfo.js 这类包的话，也可能实现从 Org-mode 输出一个完整的电子出版物的效果。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>“使命召唤”战役的改变及其原因</title>
    <link href="https://tilde.club/~xuqi/log/cod-campaign-evolution/"/>
    <updated>2026-03-11T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/cod-campaign-evolution/</id>
    <content type="html">&lt;p&gt;CoD4 单人战役给人留下深刻的印象，其中的诸多关卡近二十年来都在不断地被玩家回味、讨论，整体上这个战役更是在架空局部低烈度战争的背景下，讲述了一个高度类型化从而不失亲切，同时又以对现代战争（和政治格局）的深刻推想而引人深思的故事。Griggs 和 Gaz 的死是很多人多年的意难平。人们在游戏中和游玩后的情感投入，甚至超过了对很多 RPG 游戏的投入。&lt;/p&gt;
&lt;p&gt;我一直在试图弄明白，之后的《使命召唤》系列单人战役为什么一直不能复制4代的成功，特别是看起来只是规模更大、更喧闹了一点，又只隔了两年的同工作室同世界观续作《使命召唤6》。&lt;a href=&quot;https://www.bilibili.com/video/BV1Ev411r7rg/&quot;&gt;这部视频&lt;/a&gt;提到的4代和6代（《现代战争2》）的一个不同让我似乎又更接近了一点最后的答案。&lt;/p&gt;
&lt;p&gt;CoD4 和 CoD6 我都很久不玩了，后者更是没给我留下太多印象。所以当把 CoD4 和&amp;quot;使命召唤&amp;quot;系列续作对比时，我更多地是把它和2019年的 remake 《现代战争》做粗浅的比较。当时我得出的结论是在《使命召唤：幽灵》的失败之后，这个系列放弃了大战场的刻画，转而描绘&amp;quot;破门行动&amp;quot;这一层级的特种部队 CQB 战斗。当任务变成一个个小房间或者箱庭的组合，自然很难讲好故事，也很难给人留下深刻的印象。&lt;/p&gt;
&lt;p&gt;看过 Jetwow 的视频后，我意识到我的这个解释也不算错，但只是用现象解释现象。视频里提到在 CoD4 和之前的系列作品中，你扮演的是一个普通士兵，需要在敌方 AI 压倒性的火力下，临场规划战术，从地图以微妙的方式提供给你的几条可选路线中选择一条，然后顶着枪林弹雨，一个路口一个路口、一个拐角一个拐角地站住脚和完成任务。在我看来，这个设计形象地诠释了什么叫&amp;quot;拱卒&amp;quot;，但低调的多重路线设计又让扮演小兵的你需要承担棋手的工作，思考我这颗小卒可能遇到什么危险、怎样能不被吃掉，而不是仅仅当一个装在枪上等着接收大场面的摄影机。&lt;/p&gt;
&lt;p&gt;2008年的 Treyarch 本质上还是个 mod 工坊，所以《使命召唤：世界大战》（CoD: WaW）虽然在试图学 CoD4 的这个设计，却没有学像。玩家操纵的角色要面对数量上具有压倒优势的敌人，但粗糙的行为脚本导致这些敌方角色只会在前一个人死后跑到同一个位置补位，玩家能看到他他就开枪，玩家躲在掩体后他就扔手雷。在低难度下，我经常有敌人排着队等我枪毙的感觉，高难度下则只能背版，在掩体后调整好准星位置，&amp;quot;闪狙&amp;quot;敌人，然后在如雨点般落下的手榴弹爆炸前仓皇逃到另一个掩体。关卡的设计上，如果我们把 CoD4 的地图比作由咽喉要道和演出舞台组成的网络，WaW 的地图只能说是一条崎岖的华山道，因为路上的景致重复而更加成为一种折磨。&lt;/p&gt;
&lt;p&gt;到了 CoD6，按照 Jetwow 视频里的说法，最大的变化是玩家扮演的从普通士兵，变成上天下海无所不能的超级战士。故事也因此从粗砺悲凉的战场回忆摇身变成摧枯拉朽的 Power Fantasy。恐惧缘于火力不足，而当玩家可以随时召唤炮火支援，趴下就能躲避反步兵跳雷的时候，只要会找掩体，不太需要思考战术就可以一路碾压过去。即使游戏中后期最难的两关，&amp;quot;Whiskey Hotel&amp;quot;和&amp;quot;Loose Ends&amp;quot;，也没有给人考虑战术的空间，而回归了类似 WaW 的用火力点数量在单一路径中制造难度的方式。尤其是高难度下的&amp;quot;Loose Ends&amp;quot;，玩家基本上只能在楼梯上守点，然后沿着一条路线被人追着打，固然困难得让人印象深刻，但也失去了重玩的价值。游戏方式上一个重要的不同是，CoD6 有了更多的载具登场，包括玩家必须乘用的快艇、直升机，和不能乘坐但为玩家提供火力支持的装甲车、坦克等等。另外，关卡里还给玩家安排了自动加特林哨戒枪、空对地导弹等等武器。少数 CQB 战斗被简化成失误惩罚不大的慢镜头交互操作。乍看上去，这些只是无伤大雅的点缀，无非是给玩家增加一点枪械之外的体验。但它们实际上是把 CoD4 经过深思熟虑后非常克制地给玩家的伤害输出和移动速度等优势，反复多次地加到玩家角色身上。CoD 1 代和 2 代面对德军虎式坦克的恐惧已成过去，棋盘式逐步推进的任务在载具和高科技武器的加持下变成肾上腺素飙升但转瞬即逝的过山车体验。&lt;/p&gt;
&lt;p&gt;CoD6 以说唱为主题音乐的预告片，似乎在向世界宣告：&amp;quot;这才是我们想表现的现代战争，酷就完了！&amp;quot;但整个单人战役玩下来，再回顾这个预告片，人们多少会有一点 Infinity Ward 为了这几个名场面的醋，包了一锅饺子的感觉。甚至可以说，为了把雪山突袭、空间站爆炸、机场屠杀等等大场面给揉到一个游戏里，制造一篇超级爽文，IW 不惜牺牲了故事的连贯。大多数任务都有截出图来能吓死一打游戏评论家的场景，但这些任务也确实像锅里那些各自漂浮着的饺子，组成不了 CoD4 那样清晰的叙事。&lt;/p&gt;
&lt;p&gt;那么现在的问题就是，为什么这碟醋对动视来说那么重要，比本来应该端给玩家的年夜饭更加重要？&lt;/p&gt;
&lt;p&gt;2007年，CoD4 的发布在当时看来是个改变游戏行业的事件，但当时，它并不是游戏行业最引人注目的现象。所有人都在关注的，是另一个著名游戏系列的新续作：《光环3》。&lt;/p&gt;
&lt;p&gt;虽然《光环3》源自一个现象级的游戏系列，但它本来也不至于整个改变 IW 的游戏制作思路。让它成为关注焦点的原因，是它登上了 Xbox 360 平台。当时宽带和无线网络在英美两国家庭的普及率迅速提高，Xbox 360 背靠经过多年运营逐渐成熟的 Xbox 网络服务，又提供了 Netflix 等等视频平台的应用，这就使它不再仅仅是让人对你另眼相看的宅男玩具，而是可以方便又坦然地放在客厅的娱乐家电。而由于 Sony 的设计和营销失误，Xbox 360 的销售完全碾压了性能相近的 PS3。借着主机和网络普及的东风，以及第一方游戏的待遇，《光环3》成了当年游戏市场当之无愧的霸主。最让动视受刺激的是，即使以 CoD4 饱受赞誉的多人游戏模式，以及 PC 和两大主机平台共同发售的优势，它的玩家粘性也远不如《光环3》。&lt;/p&gt;
&lt;p&gt;在 Xbox 360 和《光环3》之前，游戏的多人模式是游戏的增值部分，但撇开 CS、Quake 这些电竞专用游戏不说，几乎没有人会把它看作和单人战役平起平坐的内容版块。虽然&amp;quot;使命召唤&amp;quot;系列一开始就有联机功能，但那是为 PC 玩家设计的，要求玩家在电脑椅上正襟危坐，在固定的几个服务器上和熟人玩，一次至少投入一下午的时间。相对缓慢的节奏允许玩家在游戏中反复争夺点位，和队友商量战术、磨练枪法，是输是赢都不过消磨时间而已。联机功能的作用类似书的精装版，只是让发行商可以给游戏的定价增加15美元。&lt;/p&gt;
&lt;p&gt;《光环3》改变了这一切。在它的联机模式中，你可能会随机匹配上来自另一个半球的玩家，游戏机制必须简单直接，让没有耳麦的队友也知道该干什么；你很可能不再是16岁不知怎样度过漫长暑假的少年，而是28岁的上班族，只想在一天工作后窝在沙发里放松一下，还不想错过20分钟后自己最喜欢的球队比赛转播，所以游戏模式变成节奏超快的夺旗，地图也脱离对单机任务的模仿而为联机游戏的要求全新制作，演化出经典的三路模式，同时去掉了所有的死胡同和单门房间，确保所有人都得一刻不停地转移位置；你不再有时间和意愿制作地图包和 mod，也不再有和你共享这些兴趣的朋友，所以 Bungie 会持续为你和其他玩家制作对战地图包，5张新地图和新的模式只要 800 点 MS Point （约合10美元）就可以买到。&lt;/p&gt;
&lt;p&gt;最后这一点，让50美元的游戏变成80甚至100美元的游戏。我们有理由相信正是这一点最让动视眼红。CoD4 只比《光环3》晚上市两个月，最多只能在 DLC 上做点文章了。但 CoD6 刚开始开发，于是动视给 IW 设定了目标：把《光环3》及其续作拉下马。&lt;/p&gt;
&lt;p&gt;光吸引 CoD4 的玩家是不够的，新作要有更多能在 YouTube 上反复被观看的片段，吸引休闲玩家。多人游戏的节奏要更快，所以要引入新模式，地图设计上也要让双方有更多机会交战。要在至少一年内吸引玩家不停地玩，就要加入更多的武器、更难达成也更具威力的连杀奖励、更多的可定制选项。&lt;/p&gt;
&lt;p&gt;这样一来，CoD6 的多人联机模式终于从游戏变成了&amp;quot;内容&amp;quot;。&lt;/p&gt;
&lt;p&gt;这个变化如何影响单人战役呢？首先，总的开发时间是固定的，多人游戏的复杂度几乎增加了100%，对单人战役的推敲必然会受影响。多人游戏要做那么多地图，肯定得复用单人战役的图，那单人战役就不能再做影响节奏的攻坚式和棋盘式地图。其次，多人游戏里的连杀变得更复杂、威力更大了，那单人游戏里肯定要安排相应的情节，不能让好不容易达成12连杀的玩家拿到掠食者无人机却不会用。最后，多人游戏里的众多枪械也得让玩家都有机会用一下，要多安排一些不同类型的任务和换枪享受点名快感的机会，不能让人像真的大头兵一样一把 M4 打完整个游戏。&lt;/p&gt;
&lt;p&gt;这就使 CoD6 的整个单人战役节奏更快、地图更大，但同时更扁平、更喧闹，因为它的地图、机制、武器和敌人配置都得为多人游戏的短对局、陌生人匹配和持续运营的模式铺路。后来系列的各部作品，更是在这条路上越走越远了。可以说，CoD4 成为&amp;quot;使命召唤&amp;quot;系列单人战役的绝唱，是 PC 游戏的大众化和客厅游戏到来的必然结果。&lt;/p&gt;
&lt;p&gt;总之，&amp;quot;使命召唤&amp;quot;单人战役和游戏其他部分的主次关系，在《光环3》之后完全颠倒。CoD4 及之前，多人游戏是为了让你和好友们可以一起重回单人游戏里传说中的诺曼底和卡朗唐。这之后，单人战役和预告片一样，是多人游戏的 teaser。它看起来很像那些黑人说唱歌手的 MV，目的也类似：用一趟暴力和威力（权力）的迷幻之旅，让你为了能持续享受这种大金链子随便戴的迷醉感而反复掏腰包买专辑和单曲（新的游戏和更炫的皮肤）。能让人只看一遍就掏钱的 teaser 才是最好的 teaser，从这个意义上说，单人战役的浅薄、直白、短命、缺乏重玩价值，正是它最大的价值。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>推想一种新的出版</title>
    <link href="https://tilde.club/~xuqi/log/speculating-new-publishing/"/>
    <updated>2026-01-14T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/speculating-new-publishing/</id>
    <content type="html">&lt;ul&gt;
&lt;li&gt;未来的出版大概会分成两类。
&lt;ul&gt;
&lt;li&gt;一类是给机器看的，这里的机器可能是现在备受关注的 agents，也可能是别的什么实现方式，但它们的作用都会像魔镜/水晶球/Jarvis，给人提供信息和建议，提示可能的风险和机会，以及代人去操作机器/系统，来改变其周围世界和关系。我们把这种机器叫做内容机器。&lt;/li&gt;
&lt;li&gt;另一类是给人看的文本内容。是的，这个古老的范式会继续存在。如果它消亡的话，人类可能都没机会等到第一类内容的普及成为现实。人类和机器处理抽象概念和复杂问题的能力还是源于这一类内容。
&lt;ul&gt;
&lt;li&gt;首先它方便当前人工智能范式下，抽象概念的机器训练和学习。&lt;/li&gt;
&lt;li&gt;其次，面向复杂问题的创新思路需要把看似不相关的东西联系在一起，在可见的未来，只有人可以在阅读、掌握多方面的内容后实现这一点。&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;对应的服务人群相比今天会扩大。
&lt;ul&gt;
&lt;li&gt;&amp;quot;大众读者&amp;quot;即单纯的兴趣类读者会减少。小说不会彻底消失，但文学类纸书出版会变成一个很小的、很粉丝向和艺术品向的环节。&lt;/li&gt;
&lt;li&gt;但不论在哪个行业，都会变成某种专门化内容机器的用户。&lt;/li&gt;
&lt;li&gt;第二类内容将服务最高端的读者，即那些对世界建模和调优模型及机器的人。&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;出版业和内容的竞争方式也会不同。
&lt;ul&gt;
&lt;li&gt;对于第一类内容，出版业的竞争在于谁能掌握更全面的数据，谁有更好的 &lt;strong&gt;世界模型&lt;/strong&gt;&lt;sup class=&quot;footnote-ref&quot;&gt;&lt;a href=&quot;https://tilde.club/~xuqi/log/speculating-new-publishing/#fn1&quot; id=&quot;fnref1&quot;&gt;[1]&lt;/a&gt;&lt;/sup&gt;，以及谁的 agent 更能及时准确地解决用户的问题，更有 &lt;strong&gt;主动性&lt;/strong&gt;&lt;sup class=&quot;footnote-ref&quot;&gt;&lt;a href=&quot;https://tilde.club/~xuqi/log/speculating-new-publishing/#fn2&quot; id=&quot;fnref2&quot;&gt;[2]&lt;/a&gt;&lt;/sup&gt;。谁都有通用的 LLM 和 agent，出版者的 agent 如果在上面这些方面不是更有优势，特别在内容上不是更突出的话，这个产品就挣不到钱。&lt;/li&gt;
&lt;li&gt;第二类内容的竞争涉及到出版社怎样确定作者和作品的&amp;quot;PageRank&amp;quot;。可以设想这样一种情景，作者未来不仅是给出版社交稿，他可能更多地负责给出版社的 agent 和世界模型调优，以此获取报酬。第二类内容更多地像是个通告，告诉读者，这个作者参与了本公司的内容生产。越是重要、权威的作者，就越能给出版机构背书，同时将读者引向这个品牌。在某些领域，独占性的协议和关系会变得更昂贵和有价值。&lt;/li&gt;
&lt;li&gt;竞争中，话语权会向少数出版机构和作者集中吗？有趣的是，虽然上面这个出版模式的价值在于减少风险和不确定性，让任何人都可以得到可行动的知识，但世界内生的不确定性才是这个模式继续运行下去，让作者和读者不断用真实世界的观察纠正和更新既有认识的推动力。这个过程中，相信也会有能更好地解决特定问题的小机构、新机器的出现，最终也许会形成和20世纪的出版业类似的格局。&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;今天的出版机构如果想继续活下去可以做什么。
&lt;ul&gt;
&lt;li&gt;既有内容的语义化而不是简单的电子化。&lt;/li&gt;
&lt;li&gt;确定内容方向
&lt;ul&gt;
&lt;li&gt;一个能够建立起护城河的突破方向，优先集中发展。&lt;/li&gt;
&lt;li&gt;指导将来拓张、发展的内容版图。&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;建立专家、顾问团队而不是作者团队。
&lt;ul&gt;
&lt;li&gt;世界模型的专家&lt;/li&gt;
&lt;li&gt;内容机器专家&lt;/li&gt;
&lt;li&gt;产品顾问&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;了解读者，或者更好地，培养&amp;quot;亲生&amp;quot;的读者群。&lt;/li&gt;
&lt;li&gt;最终这还是要回到内容的质量和呈现上。
&lt;ul&gt;
&lt;li&gt;营销不要什么书都推，只推和你（想要）的读者群有关的。&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;Footnotes&lt;/h2&gt;
&lt;hr class=&quot;footnotes-sep&quot; /&gt;
&lt;section class=&quot;footnotes&quot;&gt;
&lt;ol class=&quot;footnotes-list&quot;&gt;
&lt;li id=&quot;fn1&quot; class=&quot;footnote-item&quot;&gt;&lt;p&gt;即描述相关因素如何相互作用的模型。未来，这些模型会变得非常复杂，而且互操作性------即是否能通过统一的输入输出让多种不同类型的模型协作------这个指标将变得非常重要。 &lt;a href=&quot;https://tilde.club/~xuqi/log/speculating-new-publishing/#fnref1&quot; class=&quot;footnote-backref&quot;&gt;↩︎&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id=&quot;fn2&quot; class=&quot;footnote-item&quot;&gt;&lt;p&gt;和&amp;quot;人工智能&amp;quot;一样，agent 目前也是个略有夸大的说法，因为它还无法与用户的需要对齐，还无法做到真的主动。 &lt;a href=&quot;https://tilde.club/~xuqi/log/speculating-new-publishing/#fnref2&quot; class=&quot;footnote-backref&quot;&gt;↩︎&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/section&gt;
</content>
  </entry>
  <entry>
    <title>Reading a piece of RAG</title>
    <link href="https://tilde.club/~xuqi/log/latex-as-rag/"/>
    <updated>2026-01-12T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/latex-as-rag/</id>
    <content type="html">&lt;p&gt;A few days ago, someone submitted to &lt;a href=&quot;https://news.ycombinator.com/item?id=46535074&quot;&gt;Hacker News&lt;/a&gt; a link to &lt;a href=&quot;https://sciencestack.ai/&quot;&gt;sciencestack.ai&lt;/a&gt;, but received very little fanfare.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;2 points by cjlooi 4 days ago&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;But what the product is trying to do is of much interest to me. As an editor, I have imagined a future for publishing where authors write in Markdown or some more semantic markup languages so more interactive ways of content consumption can build on them. And as a product manager working on a scholar AI agent product, I want to see how people approach the problem of how to use agents to facilitate reading and understanding of science papers.&lt;/p&gt;
&lt;p&gt;It is a bit hard to believe, given the huge progress LLM products achieved in 2025, that LLMs still don&#39;t read PDFs semantically. They can extract text alright, even fixing some problems like line breaks along the way, but they don&#39;t know easily where to find &amp;quot;Figure 3.2&amp;quot;, or which line is a section heading, as we would imagine. So sciencestack chooses to only process papers with their LaTeX code available, which are almost exclusively on ArXiv as of now.&lt;/p&gt;
&lt;p&gt;By parsing LaTeX codes, you get much more semantic and structural information on the text than by looking at (text extracted from) PDFs. So sciencestack can provide an API where you can query a paper as if it is a web service, and make AI agents who can answer your reading questions better.&lt;/p&gt;
&lt;p&gt;They combined these into an interactive paper reader. The user can see immediately what &amp;quot;Section 3.2&amp;quot; or &amp;quot;Figure 5&amp;quot; or &amp;quot;Reference [2]&amp;quot; is, by hovering the cursor over the highlighted/linked text. Also, thanks to the LaTeX source code, they can copy math equations in one click.&lt;/p&gt;
&lt;p&gt;For each section, LLM (GPT-5 Nano) generates a summary. The user can view them as a list in a sidebar, or click the annotation icon in the right margin, just like in MS Word. But they cannot interact with these annotations further, like commenting etc. In this case, I think it is better if these notes are integrated into the table of contents to the left.&lt;/p&gt;
&lt;p&gt;If you select some text in the paper, you get a popup toolbar which lets you copy LaTeX/Markdown presentation of the text selected, add annotation, or ask AI about it, with the latter two features only available to signed-in users.&lt;/p&gt;
&lt;p&gt;By the way, when I select the text, there are now &lt;em&gt;3&lt;/em&gt; popup toolbars offering to do something with the selection with LLMs. They are respectively provided by the instant message app, the browser, and sciencestack. This reminds me of late 1990s when all kinds of dashboards and toolbars are competing for your desktop computing and web surfing experiences.&lt;/p&gt;
&lt;p&gt;Sciencestack is more like a prototype or proof of concept, than a product that convince you that a new paradigm is on the horizon. It also reminds me that the intelligent reader I want to build may be still too computationally expensive/difficult to be feasible. At the current state, it may help one read a paper better, or more conveniently extract data, but the improvement is very limited, not enough to justify an independent product and all those tokens you have to buy.&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>Affordability Crisis: A Reader</title>
    <link href="https://tilde.club/~xuqi/log/affordability-crisis/"/>
    <updated>2026-01-03T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/affordability-crisis/</id>
    <content type="html">&lt;h1&gt;Causes&lt;/h1&gt;
&lt;h2&gt;Continued inflation&lt;/h2&gt;
&lt;p&gt;Inflation rate has been high since before COVID-19. Although it has been much slower in recent months, it remained relatively high (2.7% in Nov 2025).&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Put another way: The inflation rate since Q4 2019 has risen about twice as fast as the Federal Reserve&#39;s optimal rate, according to Mark Zandi, chief economist at Moody&#39;s.&lt;/p&gt;
&lt;p&gt;-- &lt;a href=&quot;https://www.cnbc.com/2025/12/24/affordability-is-a-buzzword-right-now-these-charts-show-why.html&quot;&gt;Affordability is a buzzword right now --- these charts show why&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;And businesses generally don&#39;t cut prices even if inflation has decelerated.&lt;/p&gt;
&lt;h2&gt;Health insurance subsidies expiring&lt;/h2&gt;
&lt;p&gt;The issue may last into 2026.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Enhanced premium subsidies for 22 million Americans enrolled in health insurance via the Affordable Care Act marketplace are poised to disappear at the end of 2025, leading insurance premiums to more than double for the average recipient.&lt;/p&gt;
&lt;p&gt;-- &lt;a href=&quot;https://www.cnbc.com/2025/12/24/affordability-is-a-buzzword-right-now-these-charts-show-why.html&quot;&gt;Affordability is a buzzword right now --- these charts show why&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2&gt;War in Ukraine&lt;/h2&gt;
&lt;p&gt;Some increases in food and fuel prices may be attributed to the Ukraine War.&lt;/p&gt;
&lt;h2&gt;Shortage in housing supply&lt;/h2&gt;
&lt;blockquote&gt;
&lt;p&gt;Today, both rental and homeowner vacancy rates---the share of total housing stock available for rent and sale respectively---are below those seen in the two decades preceding the (2007-2009) crisis and the housing market collapse that triggered it.&lt;/p&gt;
&lt;p&gt;These measures of housing stock availability point to a growing gap between supply and demand---the root of the affordability problem. Rents and mortgage payments are taking a bigger chunk of income, according to multiple measures, and the trend accelerated post-pandemic.&lt;/p&gt;
&lt;p&gt;-- &lt;a href=&quot;https://www.goldmansachs.com/insights/articles/the-outlook-for-us-housing-supply-and-affordability&quot;&gt;The Outlook for US Housing Supply and Affordability&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;One important factor causing limited supply of housing is land use regulations in the US. These regulations often are decided locally, making it difficult to reform at a large enough scale.&lt;/p&gt;
&lt;p&gt;Near the city centers, land supply is also declining, pushing up development costs. Lack of skilled labour in the construction industry also makes it more expensive to build houses. An opinion is that AI could have positive impacts in this regard, for it pushes would-be new white-collars into the construction business and other industries that need young, strong workers. The irony is not hard to see.&lt;/p&gt;
&lt;h1&gt;Impact&lt;/h1&gt;
&lt;h2&gt;Democrats victories&lt;/h2&gt;
&lt;blockquote&gt;
&lt;p&gt;Democrats have ridden the affordability issue to victories in recent elections, including the mayoral race in New York and gubernatorial contests in New Jersey and Virginia.&lt;/p&gt;
&lt;p&gt;-- &lt;a href=&quot;https://www.cnbc.com/2025/12/24/affordability-is-a-buzzword-right-now-these-charts-show-why.html&quot;&gt;Affordability is a buzzword right now --- these charts show why&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2&gt;Middle-class cost of living problems&lt;/h2&gt;
&lt;p&gt;&lt;img src=&quot;https://tilde.club/~xuqi/images/share-of-middle-class-families-that-can-afford-basic-necessities-in-160-us-metro-areas.png&quot; alt=&quot;&quot; /&gt;&lt;/p&gt;
&lt;p&gt;For other graphs, see &lt;a href=&quot;https://www.brookings.edu/articles/in-every-corner-of-the-country-the-middle-class-struggles-with-affordability/&quot;&gt;In every corner of the country, the middle class struggles with affordability&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The affordability crisis make the middle-class families more susceptible to future shocks, and affects the long-term viability of the US economy.&lt;/p&gt;
&lt;h1&gt;Response&lt;/h1&gt;
&lt;p&gt;Most of these are policy recommendations instead of actions or implementations. Current administration makes it very unlikely that these measures can be taken nationwide.&lt;/p&gt;
&lt;h2&gt;Increase income&lt;/h2&gt;
&lt;blockquote&gt;
&lt;p&gt;To increase incomes, policymakers could increase the federal minimum wage, which since 2009 remains stagnant at $7.25 per hour (or $2.13 per hour since 1996 for tipped workers).&lt;/p&gt;
&lt;p&gt;-- &lt;a href=&quot;https://www.brookings.edu/articles/in-every-corner-of-the-country-the-middle-class-struggles-with-affordability/&quot;&gt;In every corner of the country, the middle class struggles with affordability&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2&gt;Decrease living costs&lt;/h2&gt;
&lt;blockquote&gt;
&lt;p&gt;At the national and state level, lowering costs for middle-class households could include universal health and child care as well as increased investments into public transportation and Supplemental Nutrition Assistance Program (SNAP) benefits.&lt;/p&gt;
&lt;p&gt;-- &lt;a href=&quot;https://www.brookings.edu/articles/in-every-corner-of-the-country-the-middle-class-struggles-with-affordability/&quot;&gt;In every corner of the country, the middle class struggles with affordability&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;
</content>
  </entry>
  <entry>
    <title>The year of indifferent thinking</title>
    <link href="https://tilde.club/~xuqi/log/the-boring-decade/"/>
    <updated>2025-12-29T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/the-boring-decade/</id>
    <content type="html">&lt;p&gt;As one who earn one&#39;s income partly through editorial work, I have subscribed many Wechat channels operated by publishers and publish industry media. For the past year, none of the stories they published managed to make me interested, not to mention excited, about any of the books. Many others must have a similar take about books: everyone in the publishing business know 2025 was the worst year in their career so far, no matter if their career is short or long. For several consecutive years, there have been small but significant decrease in book sales and revenue across genres. Non-fiction were hit especially hard.&lt;/p&gt;
&lt;p&gt;One of the authors I know mentioned on social media last week, that she haven&#39;t bought a single book in 2025. We editors and authors are supposed to be book devourers! And it&#39;s not only books that fail to excite people. When was the last time a movie trailer made you want to go to the cinema?&lt;/p&gt;
&lt;p&gt;An explanation to this dilemma is that the publishers of books and producers of movies are trying too hard to play safe. They want to minimize the risk of a flop, to maximize their survival chances in this very uncertain decade.&lt;/p&gt;
&lt;p&gt;Superficially, this makes sense. But the demand side must have something to do with the situation as well. If playing safe brings absolute success, we should be seeing &lt;em&gt;Avengers 5&lt;/em&gt; in its &lt;em&gt;Avengers 3&lt;/em&gt; clone form already, as well as that it become the most profitable one in history. The ultimate &amp;quot;play-it-safe&amp;quot; movie was &lt;em&gt;Avatar&lt;/em&gt;, whose success was not reproducible perhaps for the rest of the century. It is that the current publishers/producers reacted to the loss of interests by playing safe, not the other way round.&lt;/p&gt;
&lt;p&gt;Fewer people are paying for books and movies. The market size is shrinking. That&#39;s why books and movies are playing it safe. When the market cannot support so many SKUs, any manufacturer would limit themselves to the tried and true formula. So would publishers and filmmakers.&lt;/p&gt;
&lt;p&gt;Yet, that was also only superficially true. In 1998 there was the &lt;a href=&quot;https://en.wikipedia.org/wiki/1997_Asian_financial_crisis&quot;&gt;Asian Financial Crisis&lt;/a&gt;, and we have &lt;em&gt;The Matrix&lt;/em&gt; in 1999 and tLoR around 2000. In 2008 there was a bigger recession thanks to greedy banks, and we have &lt;em&gt;Avatar&lt;/em&gt; the next year. Things were not boring then and people had not interests but &lt;em&gt;hypes&lt;/em&gt; on everything Hollywood, Penguin and even EA produced.&lt;/p&gt;
&lt;p&gt;Business is bad but that usually don&#39;t stop people from paying for stuff that can bring them laughters, tears, and maybe some soul-searching.&lt;/p&gt;
&lt;p&gt;The reason that people find today&#39;s content boring, is that we are not so sure about that the truth exists, and that we can identify the truth, anymore. If we don&#39;t believe a book or a movie contains some truth that we want to know or identify with, we lose interest in it.&lt;/p&gt;
&lt;p&gt;We need to believe the truth is in there, to ask the right or wrong questions, and the why questions. These questions are the reasons behind the story&#39;s allure.&lt;/p&gt;
&lt;p&gt;Social media was a heavy blow to the belief. The fact that anyone can tell anyone anything about anything, means that it is anyone&#39;s job to identify truth. It is a tiring job that exhausts many even if it is paid. After 20 years of social media, everyone is tired. We don&#39;t share things because they are true but because we find them agreeable, or just fun.&lt;/p&gt;
&lt;p&gt;And the LLM technology may be the second blow. Suddenly everyone can access from their pocket something that is mostly right, and almost factful. We have our doubts but choose to believe it when it churns out answers for ourselves. The doubts are reserved for news, books and movies we read and see. Somehow we believe the LLM chatbots are directly addressing us, and are more reliable than &amp;quot;the (mass) media&amp;quot;, which we can&#39;t tell whether LLMs are involved or not.&lt;/p&gt;
&lt;p&gt;So here we are. Apathy is a condition that we should not ignore. But luckily, as long as something is not completely AI slop, we can find something truthful in it, if not in the content itself, then the way it is packaged or delivered, or why we give it our time in the first place.&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>北京，水晶之城</title>
    <link href="https://tilde.club/~xuqi/log/beijing-city-of-quartz/"/>
    <updated>2025-12-24T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/beijing-city-of-quartz/</id>
    <content type="html">&lt;p&gt;Mike Davis 在 &lt;em&gt;City of Quartz&lt;/em&gt; 中，引用了德国地理学家 Anton Wagner 的作品 &lt;em&gt;Los Angeles&lt;/em&gt; 中的一段话：&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Here, one wants to create the Paris of the Far West. Evening traffic on Hollywood Boulevard attempts to mimic Parisian boulevard life. However, life on the Boulevard is extinct before midnight, and the seats in front of the cafes, where in Paris one can watch street life in a leisurely manner, are missing .... At night the illuminated portraits of movie stars stare down from lampposts upon crowds dressed in fake European elegance --- a declaration that America yearns to be something other than American here .... Yet, in spite of the artists, writers and aspiring film stars, the sensibility of a real Montmartre, Soho, or even Greenwich Village, cannot be felt here. The automobile mitigates against such a feeling, and so do the new houses. Hollywood lacks the patina of age.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Patina 大概可以译为&amp;quot;包浆&amp;quot;，就是时间在古代金属物件上留下那层薄氧化膜的样子。不过，Wagner 观察到的现象，也许不能简单归因于洛杉矶缺乏历史。因为号称古都的北京，到了夜晚，同样给人一种&amp;quot;life ... is extinct&amp;quot;的感觉，而且越是历史悠久的内城越是如此。&lt;/p&gt;
&lt;p&gt;在北京工作过的人都能说出一个很有力的理由，就是北京太大了，无论贫富，大部分人的居住地不在中心城区。所以人们没有所谓夜生活的习惯，如果不尽早回家料理家务，第二天的工作生活就会受影响。洛杉矶想必也类似，正如 Davis 想要说明的，它一直没有形成真正的城市空间，而只是在资本推动下，在沙漠上不断扩张的一片建成环境。&lt;/p&gt;
&lt;p&gt;但对北京，我们还可以追问，为什么这座城市在几百年时间里都有一个定义清晰的城区范围，却也始终没有形成巴黎那样的都市文化。我猜测，正因为它几百年间一直是大一统国家的国都，所以城市一直是被规划和整体兴建的，没有一个生长的过程。另外，东亚诸国的生活方式都由农业文明塑造，习惯日落而息，强调家庭价值和邻里相互监督，这些都导致从城市到其居民，都不怎么习惯夜生活。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>AI Bubble</title>
    <link href="https://tilde.club/~xuqi/log/ai-bubble/"/>
    <updated>2025-12-18T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/ai-bubble/</id>
    <content type="html">&lt;p&gt;当前有大量唱衰 AI （大语言模型）投资的声音，而我也认为它们的判断是有道理的。我详细读完的文本包括 a) Jerry Kaplan 的 &lt;a href=&quot;https://substack.com/home/post/p-181041717&quot;&gt;&lt;em&gt;There&#39;s An AI Bubble. For Real.&lt;/em&gt;&lt;/a&gt; 和 b) 2 &amp;amp; 20 的 &lt;a href=&quot;https://www.youtube.com/watch?v=oT896h76s-Y&quot;&gt;&lt;em&gt;The AI Bubble is Worse Than You Think&lt;/em&gt;&lt;/a&gt; 。&lt;/p&gt;
&lt;p&gt;它们认为当前的 AI 投资是泡沫的主要理由包括：&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;相对于各种 AI 应用的盈利能力，当前的 AI 投资规模过大和过于集中，以至于几乎都没有在自然人的有生之年收回投资的希望；&lt;/li&gt;
&lt;li&gt;类似 OpenAI - MSFT - Nvidia 这种把投资款又折成产品或服务，直接或间接地还给投资方的循环投资，带来极高的风险；&lt;/li&gt;
&lt;li&gt;不断增长的投资规模要求 AI 企业的营收达到目前的3倍以上，但 AI 技术目前基本上只能用来提高个人生产力，没有满足企业要求的产品，而个人市场很快就会见顶，各家企业还要为了争夺有限的市场份额拼命降本打价格战（昨天 Google &lt;a href=&quot;https://blog.google/products/gemini/gemini-3-flash/&quot;&gt;发布&lt;/a&gt; Gemini 3 Flash，一上来就强调成本优势）；&lt;/li&gt;
&lt;li&gt;服务于专业领域的系统很快就可以在笔记本级的硬件上运行，使得当前基于大规模云算力的 LLM 服务不再有必要。&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;我认为这些理由可以归为一个，即 LLM 并不能像人类的前几次技术革新那样，带来 &lt;strong&gt;社会&lt;/strong&gt; 生产力的实际提高，因此它无法给投资者带来实质收益。当最后的接盘侠发现无法把手中的股份以更高价格卖给下一个人时，泡沫就会破裂。上面第2点提到的循环投资，也可以认为是他们几家都信心不足的体现。&lt;/p&gt;
&lt;p&gt;社会生产力的提高体现在，人们能够更有效地搬运或转化物质/能量/信息，从而降低利用资源或劳动力的成本。比如蒸汽机和炸药把过去无法开采的矿产带上市场，内燃机和远洋货轮降低了运输货物的单位成本，互联网使传输和发布信息的成本几乎降为0，等等。&lt;/p&gt;
&lt;p&gt;但 LLM 显然无法给我们利用物质和能量的方式带来积极的影响，而在信息的利用和生产方面，它也没有带来成本的降低和效能的显著提升。表面上，借助 LLM，我做设计的速度提高了20倍不止，但是制约产品开发和上市的瓶颈从来就不是设计的速度。设计师与企业内或甲方的沟通和讨价还价才永远是他们工作中占大头的部分。LLM 只是让画图这部分工作变得更不值钱了而已，很多时候，它还造成了更高的沟通和实施成本。&lt;/p&gt;
&lt;p&gt;事实上，当人们说 AI 提高了生产力，他们往往是作为白领工人和信息工作者的雇主这样说。他们看到 LLM 仅用几分钟就能吐出看起来和人花几小时做出来的东西长得差不多的 slop，就以为效率在几十倍地提高。但他们看不到的是这些东西需要人花更多的时间改，才勉强可用。而且即使最好的情况下，你能得到的也不过是平庸之作。&lt;/p&gt;
&lt;p&gt;只有在狗屁工作的层面上，LLM 才能在保持质量的前提下带来速度的提高。如果你真心觉得 LLM 让你的工作变容易了，且不担心工作的质量下降，那恭喜你，你可能一直在做一份狗屁工作。&lt;/p&gt;
&lt;p&gt;我认为 LLM 是泡沫，还有一个理由。我认为，从根本上说，人类社会的经济活动是由能量决定的（我没有看过《能量与文明》这类书，但如果 Vaclav 的观点和我的大致相同，我不会吃惊）。最终决定你我收入和财富几何，日子过得舒心还是窘迫的，是我们利用能源的现状和前景。而大量的金钱和能源投入 LLM，只是为了让上亿份年终总结和 PPT 更能讨老板的欢心，这和你花钱租跑车带女孩兜风，最后还没追上人家差不多，只是规模大了数千万倍。&lt;/p&gt;
&lt;p&gt;我们正在把大量的能量投入 LLM，它却只生产出似是而非的文章、无法核实的数字和滥竽充数的图像。这个过程因此阻碍了其他经济领域的存在和增长，这在能源成本越来越难以转嫁给外部性的今天，是不可持续的。LLM 和其他经济系统，总有一个要因此崩溃。而今天的经济环境下，两者中任何一个的崩溃也会带来另一个的倒塌。&lt;/p&gt;
</content>
  </entry>
  <entry>
    <title>Research is refining an idea</title>
    <link href="https://tilde.club/~xuqi/log/research-is-refining/"/>
    <updated>2025-09-29T00:00:00Z</updated>
    <id>https://tilde.club/~xuqi/log/research-is-refining/</id>
    <content type="html">&lt;blockquote&gt;
&lt;p&gt;As a matter of fact, research as an inquiring drive towards understanding given phenomena is in many instances articulated as a process of adjusting and revising one&#39;s preliminary idea. As a researcher puts it:&lt;/p&gt;
&lt;p&gt;&#39;I get an idea [...] of how I might analyze something against the backdrop of the knowledge I have of the subject, um, and then I explore it in order to see if it&#39;s tenable. But not only if it&#39;s tenable, because it is a vague hypothesis, you know ... but also to flesh out the details in order to ... get smarter myself.&#39;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Inquiry-based reading -- Towards a conception of reading as a research method&lt;/p&gt;
&lt;p&gt;SAGE Publications; SAGE Publications Inc. (ISSN 1474-0222), &lt;em&gt;Arts and Humanities in Higher Education&lt;/em&gt;, #1, 19, pages 58-75, 2018 feb 27&lt;/p&gt;
&lt;p&gt;Katan, Lina; Andreas Baarts, Charlotte&lt;/p&gt;
</content>
  </entry>
</feed>
