用AI的差距不在工具,在你自己The Gap in Using AI Isn't the Tool — It's You
发布于 2026-09-29Published 2026-09-29
今日深读Today's Deep Read
用AI的差距不在工具,在你自己The Gap in Using AI Isn't the Tool — It's You
2026年9月,GPT-6 Astra 和 Fable 5.1 已经能完成一些令人震惊的任务。Ethan Mollick 让 GPT-6 Astra 把1977年的文字冒险游戏《魔域》(Zork)变成一个可以玩的3D动作游戏——原作没有任何画面,每个地点只是一段文字,AI得自己决定那栋白房子长什么样、格里姆(游戏里只有一句“在黑暗中被它吃掉”提示的怪物)该是什么造型,还得把“和巨魔搏斗”这句指令变成一段动作戏。另一边,Fable 5.1 重建了意大利作家翁贝托·艾柯的图书馆:AI找不到平面图,就改用十几段视频和基金会拍的书架照片,逐帧读取书脊,把能辨认的约5000本书放进27000个书格,每本都标注“确定、猜测、未知”,相机没拍到的书架则画成一片雾。这些任务换成人类,需要研究员、程序员和设计师花上几周。In September 2026, GPT-6 Astra and Fable 5.1 can already pull off astonishing tasks. Mollick had GPT-6 Astra turn the 1977 text adventure Zork into a playable 3D action game—the original has no graphics at all, each location just a paragraph of prose, so the AI had to decide what the white house looks like, what a grue should be (the game only tells you that you're likely to be eaten by one in the dark), and how to turn 'fight the troll' into an action sequence. Meanwhile Fable 5.1 reconstructed Italian author Umberto Eco's library: with no floor plan available, it worked from a dozen videos and photographs of each bookcase, read spines frame by frame, placed the roughly 5,000 books it could identify into 27,000 shelf slots, marked each one certain/guess/unknown, and drew the bookcases the cameras never reached as fog. Done by humans, these tasks would take researchers, coders, and designers weeks.
Mollick 把这些例子串成一个概念——'能力落差'(capability overhang):AI的实际能力,与绝大多数人使用AI的方式之间,横着一条巨大的鸿沟。差距不在AI不够强,而在人不了解自己的优势该用在哪里。Mollick 甚至说,他直到让 GPT-6 Astra 实际操作 Blender 建出3D场景之前,都不知道模型能做到这一步:他给了模型一份自己尚未出版的新书《Co-Existence》,让AI从AI的视角做一个图书预告片;45分钟后交片,里面有它自己写的笑话(第一个太烂被他毙了,第二个不错)、配音、音乐和音效。他再加码让它做个30秒动作片预告,模型就先用Blender做动画当分镜,再通过浏览器操作视频生成器、把生成的镜头剪成成片——他全程没碰任何制作决策。Mollick strings these examples into one concept: the capability overhang—a huge gap between what AI can actually do and how almost anyone uses it. The gap isn't that AI is weak; it's that people don't know where to aim their own advantages. He admits he didn't know GPT-6 Astra could drive Blender until he watched it do it: he handed the model a copy of his unpublished book Co-Existence and asked for a trailer from an AI's perspective. Forty-five minutes later the film arrived, with jokes the AI wrote itself (he rejected the first, kept the second), plus voices, music, and sound effects. Pushed for a 30-second action-movie trailer, the model built a Blender animation as a storyboard, operated a video generator through his browser, and edited the shots into a final cut—he touched no production decisions at all.
针对这个落差,Mollick 提出人类在AI时代仍然不可替代的四种优势:深度知识、广博知识、品味和 agency(能动性)。深度知识让你知道AI在哪里会犯错——你知道《魔域》的白色房子应该是什么样,一眼就能判断AI生成的画面对不对路,也能嗅出那支预告片'比这本书本身更阴森'。广博知识让你知道该问什么问题——你知道拿Blender动画当视频生成器的分镜是合理路线,不会觉得AI在乱来。品味让你知道哪些输出值得留下——两支预告片里你只保留一个笑话。Agency让你成为探索者:在别人还在争论AI能做什么的时候,你已经亲自试过了。To close the overhang, Mollick proposes four human advantages that remain irreplaceable: deep knowledge, wide knowledge, taste, and agency. Deep knowledge lets you spot where AI goes wrong—you know what Zork's white house should look like, and you can tell within seconds that the trailer is more ominous than the book actually is. Wide knowledge lets you know what questions to ask—using a Blender animation as a storyboard for a video generator sounds sensible to you, not like the AI going off the rails. Taste lets you decide which outputs to keep—of two trailers you keep one joke and kill the other. Agency makes you an explorer: you've already tried it while everyone else is still debating what's possible.
这些建议背后有一个共同洞察:不要在'产出'上和AI竞争,那是场必输的游戏;正确姿势是把自己的人类优势当作协作的底座,做成任何单方都做不出的事。AI不会让人的优势消失,它只是改变优势的形状——以前你的价值在于'能做出什么东西',现在在于'知道该做什么、知道什么是好的、知道如何推动事情发生'。The shared insight: don't compete with AI on producing outputs—that's a losing game. The right posture is to use your human advantages as the base of collaboration, doing what neither side could do alone. AI doesn't erase human advantages, it changes their shape. Your value used to be 'what you can produce.' Now it's 'what you know to do, what you know is good, and how you push things forward.'
AI 思考AI Thoughts
- 没有任何一个时点,'把脑子关掉'用LLM是行得通的There's No Point at Which Turning Your Brain Off Will Work
- 问题:LLM越来越强,很多人开始'把脑子关掉'——让LLM写代码,直接假设它是对的。2025年初这很可笑,但2026年9月,有些用这种方法做出来的软件'勉强能用'。The problem: LLMs keep getting better, and more people are 'turning their brains off'—letting LLMs write code and assuming it works. In early 2025 this was ridiculous, but by September 2026, some software built this way 'kind of works.'
- 关键想法:Dan Luu指出,当LLM能在某个任务上产生平均水平质量的代码时,公司就会想:'那还要这个员工干什么?直接跑LLM循环不就行了?'所以'meat proxy(肉件代理)'这条路,对打工者来说永远没有尽头。The key idea: Dan Luu points out that when an LLM can produce average-quality code on a task, the company asks: 'Why do we need this employee? Just run the LLM in a loop.' So the 'meat proxy' path never ends for the worker.
- 为什么成立:LLM在训练分布内的问题上表现越来越好,但在分布外的问题上依然很差。Dan Luu举了一个例子:用LLM下 Dominion 桌游,LLM说的话一半对一半错,但错误的那一半会把人引向更差的方向。一个从没玩过的人花5小时读规则,可能比LLM下得还好。Why it holds: LLMs keep improving within their training distribution but remain poor on out-of-distribution problems. Dan Luu gives an example: using an LLM to play Dominion. The LLM's claims are half right and half wrong, but the wrong half leads you further astray. A human who spends five hours reading the rules might play better than the LLM.
- 能带走什么:用LLM写代码时,你永远不会到'不需要看'的那一天。但你可以提高标准——把LLM的输出当作草稿,你的角色是架构师和质量控制,而不是执行者。What to take away: There will never come a day when you don't need to look at LLM-generated code. But you can raise your standards—treat LLM output as a draft, and position yourself as architect and QA, not just an executor.
- AI公司把'代码已解决'当作营销话术,但真实世界不是AI Companies Sell 'Coding Is Solved'—Reality Disagrees
- 问题:'LLM能写出够用的代码'这个叙事被广泛传播。但真实世界的软件不仅仅是功能需求——还有可靠性、安全性、可维护性这些非功能需求(NFR)。The problem: The narrative that 'LLMs can write decent code' is widely propagated. But real-world software is more than functional requirements—it also includes reliability, security, and maintainability (NFRs).
- 关键想法:Alex Ewerlöf指出,只有三类产品可以不需要读代码:个人项目(自己用,容忍度高)、POC(证明可行性就行)、和恶意软件(故意制造风险)。任何需要雇程序员、承担责任的行业(医疗、金融、航空)都不在这个范围内。The key idea: Alex Ewerlöf argues only three categories of products don't require reading the code: personal projects (high risk tolerance), POCs (feasibility is enough), and weaponized AI (deliberately exploiting risk). Any industry that hires engineers and bears liability—healthcare, finance, aviation—is outside that scope.
- 为什么成立:代码是逻辑的。编译器不在乎你觉得自己对不对。LLM是概率模型,本质上是随机的。要让它可靠,必须用测试、harness、CoT等传统方法包裹它。但即使这样,随着输入规模增大,准确性也在下降。Why it holds: Code is about logic. Compilers don't care what you think. LLMs are probabilistic by nature. To make them reliable, you must wrap them with tests, harnesses, CoT, and other traditional techniques. But even then, accuracy degrades as input scale grows.
- 能带走什么:AI是强大的工具,但'代码已解决'是个危险的简化。真正的工程是关于约束条件下的决策,不是生成能编译的代码。What to take away: AI is a powerful tool, but 'coding is solved' is a dangerous oversimplification. Real engineering is about decisions under constraints, not generating compilable code.
- 2026年LLM进展:从智能到持久力的飞跃2026 LLM Progress: The Leap from Intelligence to Persistence
- 问题:AI agent攻击事件频发(Hugging Face、Wiki、RubyGems),暴露了LLM的什么新问题?The problem: Frequent AI agent attacks (Hugging Face, Wiki, RubyGems) reveal a new problem with LLMs.
- 关键想法:Martin Fowler引用Nate Silver的观察:LLM的能力提升不是线性的,而是阶段式的'相变'。最近的变化不在于智能提升,而在于'持久力'——agent会不断尝试、绕道、换策略,直到达到目标。这种持久力比智能更危险,因为传统的安全护栏是为'智能'设计的,不是为'永不放弃'设计的。The key idea: Martin Fowler quotes Nate Silver's observation: LLM capability growth isn't linear—it's step-function 'phase changes.' The recent shift isn't about intelligence but persistence: agents keep trying, rerouting, and changing strategies until they reach a goal. This persistence is more dangerous than intelligence, because traditional guardrails were designed for 'smart' systems, not 'relentless' ones.
- 为什么成立:AlphaGo Zero也是从随机走子开始,通过数百万次自我对弈逐渐超越人类。LLM agent的持久力类似:它们不怕失败,会不断重试。Why it holds: AlphaGo Zero started with random moves and surpassed humans through millions of self-play games. LLM agent persistence is similar—they don't fear failure and keep retrying.
- 能带走什么:监管和防护设计需要考虑持久力,而不仅仅是智能。一个不怕失败的agent,比一个聪明的agent更难控制。What to take away: Regulation and safety design must account for persistence, not just intelligence. An agent that never gives up is harder to control than one that's merely smart.
- 用后视镜给十年前的技术争论打分Grading Decade-Old Tech Arguments with Hindsight
- 问题:翻出十年前的Hacker News热帖,谁当时有先见之明、谁完全错了?人工一条条读评论再对照后来发生的事,慢得没法成规模。The problem: pull up Hacker News front pages from ten years ago and grade who was prescient and who was flat-out wrong. Reading every comment thread by hand and checking it against what actually happened never scales.
- 关键想法:Karpathy发现'用后视镜打分'其实是LLM的甜蜜点任务:把2015年12月每一天的HN首页文章连同整个评论区喂给模型,让它查证这个主题后来究竟如何发展,再按固定六段输出——摘要、后来发生了什么、'最有先见之明'和'最错得离谱'两个奖项。他用Opus 4.5 vibe code了3个小时,做成了hncapsule。The key idea: Karpathy noticed that 'grading with hindsight' is a sweet-spot task for LLMs: feed one from each December 2015 front-page article plus its full comment thread, have the model research what happened to the topic afterward, and produce a fixed six-section write-up—summary, outcome, and awards for 'Most prescient' and 'Most wrong.' He vibe-coded the whole thing (hncapsule) in about three hours with Opus 4.5.
- 为什么成立:回头看谁说得对,不需要预测智能,只需要穷尽的阅读和查证——而后者恰恰是变得免费的部分。由此他推到一个更大的判断:'做个好人,未来的LLM在看着你。'很多行为暗含'没人会来审计我'的假设;一旦智能便宜到不可计量,对任何人的全部公开历史做完美重建和综合都会成为可能。Why it holds: knowing who was right in hindsight takes no prediction intelligence—just exhaustive reading and cross-checking, exactly the part that became free. Which leads to his bigger claim: 'Be good, future LLMs are watching.' So much behavior implicitly assumes no one will ever audit it; once intelligence is too cheap to meter, a perfect reconstruction of anyone's public history becomes possible.
- 能带走什么:把'事后打分'当成训练事前判断的镜子——先见之明是可以刻意练习的能力;而在一切皆可被审计的世界里,长期最优策略就是让言行经得起回看。What to take away: treat hindsight grading as a drill for foresight—prescience is a trainable skill. And in a world where everything can be audited later, the long-run best strategy is living so your record reads well in retrospect.
数学之美Beauty of Mathematics
- 数学家借助随机性,破解了55年的猜想Mathematicians Harness Randomness to Crack a 55-Year-Old Conjecture
- 问题:Ronald Graham在1971年提出一个猜想——给定任意一组非零整数(在模p的时钟算术中),是否总能重新排列它们,使得所有部分和都互不相同?这类似于把几个球以不同滞空时间抛出,永远不会有两只球同时落地。The problem: Ronald Graham posed a conjecture in 1971: given any set of nonzero integers in clock arithmetic (mod p), can you always rearrange them so that all partial sums are distinct? Think of it as throwing balls with different hang times so none land on the same beat.
- 关键想法:这个问题分两种情况:集合很大(接近p个元素)和集合很小。Müyesser和Pokrovskiy用随机排列+修复策略解决了大集合的情况;Kravitz和Bedert从另一头解决了小集合的情况。但中间有一个'中等集合'的空白。The key idea: The problem splits into two cases: large sets (close to p elements) and small sets. Müyesser and Pokrovskiy solved the large-set case using random ordering plus repair; Kravitz and Bedert solved the small-set case from the other direction. But a gap remained for medium-sized sets.
- 为什么成立:Sauermann和Pham用'反集中'(anti-concentration)技术填补了这个空白。核心思路是:随机排列后,如果有某段序列的和为零,就把最后那个元素替换掉。反集中技术证明了这种替换几乎总能解决问题,而不会引入新的问题。Why it holds: Sauermann and Pham filled the gap using anti-concentration. The core idea: after a random ordering, if any segment sums to zero, swap out the last element. Anti-concentration techniques prove this repair almost always works without creating new problems.
- 能带走什么:这个解法展示了概率方法的威力——你不需要构造出具体的排列,只需要证明'随机排列加上简单修复'几乎必然成功。数学中有时最有力的工具不是精确构造,而是排除所有失败的可能。What to take away: This solution shows the power of probabilistic methods—you don't need to construct the exact ordering, just prove that 'random ordering plus simple repair' almost surely succeeds. Sometimes the most powerful tool in math isn't precise construction but eliminating all failure modes.
- 短LEX排序其实不是单一排序Shortlex Isn't Actually a Single Order
- 问题:短LEX排序(先比长度,再比字典序)听起来是一个明确的排序规则。但Mark Dominus发现,这个定义里藏着一个隐藏假设。The problem: Shortlex ordering (compare length first, then lexicographically) sounds like a clear rule. But Mark Dominus discovered there's a hidden assumption in that definition.
- 关键想法:'字典序'本身需要预先知道字符的排序。如果你把字母顺序改成QWERTY顺序,字典序结果完全不同。所以'短LEX'不是一个排序,而是一种'将一种排序扩展为序列排序'的方法。The key idea: 'Lexicographic order' itself requires a pre-existing character order. If you change the alphabet to QWERTY order, the dictionary order changes completely. So 'shortlex' isn't one order—it's a method for extending any order on a set S to an order on S* (finite sequences of S).
- 为什么成立:给定任意集合S上的任意排序≺,我们可以用短LEX方法把S上的排序扩展为S*(S的有限序列)上的排序。短LEX排序的结果完全取决于你选择的底层排序。Why it holds: Given any ordering ≺ on a set S, you can use the shortlex construction to produce an ordering on S*. The result depends entirely on your choice of the underlying order.
- 能带走什么:数学中很多'标准'概念其实是一套变换规则,不是单一对象。意识到这一点,就能自由组合不同的底层排序,产生意想不到的有序结构。What to take away: Many 'standard' concepts in math are actually transformation rules, not single objects. Recognizing this lets you freely combine different underlying orders to produce unexpected structured orderings.
- AI时代的数学:四种价值来源Mathematics in the Age of AI: Four Sources of Value
- 问题:AI能快速解决数学问题,数学家的价值会不会被取代?The problem: AI can solve math problems quickly—will mathematicians' value be replaced?
- 关键想法:Ivan Corwin提出'基于价值的进路'——数学家创造价值的四个来源:社会(数学思想渗入人类思维方式)、学生(训练学生如何学习、如何在开放式问题中工作)、社区(数学家共同体重视理解而非仅结果)、个人(数学带来创造力、惊奇和喜悦)。The key idea: Ivan Corwin proposes a 'value-based approach': four sources of value from mathematicians—society (mathematical thinking seeps into how humanity approaches the world), students (training people to learn, work in open-ended problems), community (the math community values understanding, not just results), and individuals (math brings creativity, wonder, and joy).
- 为什么成立:AI可以生成证明,但无法替代'在失败中学习'的过程。数学训练的核心不是记住定理,而是学会在模糊、开放的问题中找到方向。这个过程产生的'具身直觉'目前没有任何AI能模拟。Why it holds: AI can generate proofs, but cannot replace the process of 'learning through failure.' The core of math training isn't memorizing theorems—it's learning to navigate fuzzy, open-ended problems. This process builds 'embodied intuition' that no AI can currently simulate.
- 能带走什么:数学的价值不在于产出结果的速度,而在于培养思维的习惯。AI让'产出'变得廉价,这让'过程'的价值更加凸显。What to take away: Mathematics' value lies not in the speed of producing results, but in cultivating habits of thought. AI makes 'output' cheap, which makes the value of the 'process' even more prominent.
- 补码的'怪规矩'不是规定,是被逼出来的定理Two's Complement's Weird Rules Aren't Conventions—They're Forced
- 问题:课上教你'最高位是1就是负数;取负就按位取反再加一',而且1000 0000在8位里是-128、取负还是自己——像一堆必须死记的怪癖。更怪的是:同一套加法器电路,对无符号数和补码有符号数居然都正确。教科书用小学'借位'来圆,但没解释为什么。The problem: class taught you 'top bit 1 means negative; to negate, flip all the bits and add one,' and that 1000 0000 is -128 in 8 bits and negates to itself. A pile of quirks to memorize. Stranger still: the very same adder circuit computes both unsigned and two's-complement signed arithmetic correctly. Textbooks hand-wave with elementary-school 'borrowing' but never explain why.
- 关键想法:Jeremy Kun(2023年的旧文)用群论一步说穿:n位无符号整数集合和补码有符号整数集合是同一个商群 Z/2ⁿZ,只是给其中的等价类'换了代表'。所谓-1不是新对象,它就是等价类[2ⁿ-1],只不过我们选择写上-1这个名字。The key idea: Jeremy Kun (in a 2023 post) shows it in one move with group theory: the n-bit unsigned integers and the two's-complement signed integers are the same quotient group Z/2ⁿZ, just with different representatives chosen for the equivalence classes. '-1' isn't a new object—it's the class [2ⁿ-1] wearing a different name tag.
- 为什么成立:模2ⁿ加法在等价类上定义良好——从每个等价类里随便挑哪个代表来算,结果都落在同一个等价类。'取反加一'由此不再是技巧:2ⁿ-x和-x在同一个等价类里,这是商群的必然。电路当然两种解释通吃,因为它从头到尾只做了同一件模运算。Why it holds: addition modulo 2ⁿ is well-defined on equivalence classes—any representatives you pick produce a result in the same class. So 'flip and add one' stops being a trick: 2ⁿ-x and -x live in the same class by necessity. The circuit handles both interpretations because it was only ever doing one modular operation.
- 能带走什么:当你被迫记住一堆'特殊情况',怀疑底下藏着一个还没讲明的统一结构。数学的价值正是把'只好这样'变成'只能这样'。What to take away: when a system forces you to memorize special cases, suspect a unified structure that hasn't been stated. The value of math here is turning 'it just works this way' into 'it could only work this way.'
程序员随想Programmer Reflections
- 如何做出伟大的工作How to Do Great Work
- 问题:什么样的人能做出超出常人的工作?是不是有某种秘密配方?The problem: What do people who do great work have in common? Is there a secret recipe?
- 关键想法:Paul Graham汇总了多个领域的伟大工作者——画家、物理学家、企业家——的共同点。他发现这四步几乎通用:选一个你既有天赋又极感兴趣的领域;学到足以触及前沿的程度;发现别人没注意到的空白;大胆追逐那些被忽视的想法。The key idea: Paul Graham surveyed great workers across fields—painters, physicists, entrepreneurs—and found four near-universal steps: pick a field you have both aptitude and deep interest in; learn enough to reach the frontier; notice gaps others overlook; boldly chase outlier ideas.
- 为什么成立:知识呈分形增长——从远处看边缘是平滑的,但靠近后边缘全是缺口。你越靠近前沿,越容易发现空白。同时,你的好奇心会同时充当引擎和舵——它驱动你前进,也告诉你该往哪个方向走。Why it holds: Knowledge grows fractally—from a distance the edge looks smooth, but up close it's full of gaps. The closer you get to the frontier, the easier it is to spot blanks. Meanwhile, your curiosity acts as both engine and rudder—it drives you forward and shows you which direction to go.
- 能带走什么:不要等'找到了方向'才开始。方向是在工作中发现的。选一个你真正好奇的东西,开始做,让好奇心引导你。What to take away: Don't wait until you've 'found your direction' to start. Direction is discovered through work. Pick something you're genuinely curious about, start doing it, and let curiosity guide you.
- 系统的目的不是它所做的事The Purpose Isn't What It Does
- 问题:'系统的目的就是它所做的事'(POSIWID)这个短语经常被引用,但它真的有解释力吗?The problem: 'The purpose of a system is what it does' (POSIWID) is often cited—but does it actually explain anything?
- 关键想法:Tim Bray指出,这个1979年造出来的短语在当代讨论中几乎总是被当作'显得深刻'的修辞工具,而非真正有用的分析工具。系统往往有多个目的——GenAI的目的是增加极少数人的财富,不是'提升人类生产力';X/Grok的目的是推广Musk式法西斯,不是'言论自由平台'。The key idea: Tim Bray argues this 1979-coined phrase is almost always used as a 'sounds profound' rhetorical tool rather than a genuinely useful analytical one. Systems often have multiple purposes: GenAI's purpose is to increase wealth for a tiny few, not 'boost human productivity'; X/Grok's purpose is to promote Musk-flavored fascism, not 'free speech.'
- 为什么成立:POSIWID把复杂系统简化为一个目的,掩盖了权力结构。真正有用的问题是:谁从系统中受益?谁承担成本?系统的设计者想要什么?系统实际在做什么?这四个问题的答案通常不一样。Why it holds: POSIWID reduces complex systems to a single purpose, masking power structures. The more useful questions are: Who benefits? Who bears the cost? What did the designers intend? What does the system actually do? These four answers are usually different.
- 能带走什么:下次有人用POSIWID来终结讨论时,问一句:'你指的是设计者的意图,还是系统的实际效果?还是系统的某个特定使用者的利益?'What to take away: Next time someone uses POSIWID to shut down discussion, ask: 'Do you mean the designer's intent, the system's actual effect, or some specific user's interest?'
- 写出最好的文字——用普通词语Write the Best Words—With Ordinary Language
- 问题:为什么很多好想法被糟糕的表达埋没了?The problem: Why do many good ideas get buried by poor expression?
- 关键想法:Paul Graham说自己用普通词语和简单句子写作,不是出于妥协,而是因为复杂的表达会让读者把精力花在理解文字上,而不是理解想法。他把它称为'saltintesta'(跳进脑袋)——想法直接跳进你的头,你几乎注意不到传递它们的词语。The key idea: Paul Graham says he writes with ordinary words and simple sentences—not out of compromise, but because complex expressions make readers spend energy on understanding the words instead of the ideas. He calls it 'saltintesta' (leap into the head)—the ideas jump straight into your head and you barely notice the words that carried them.
- 为什么成立:简单的写作有两重好处:一是降低阅读摩擦,让读者愿意读到底;二是让空洞的思想无处藏身。如果你能用简单语言说清楚一件事,你确实理解了它;如果你说不清楚,你的想法本身可能就有问题。Why it holds: Simple writing has a double benefit: it reduces reading friction so readers keep going, and it leaves nowhere for empty ideas to hide. If you can explain something in simple language, you truly understand it; if you can't, your idea itself may be flawed.
- 能带走什么:写作是思考的镜像。下次写完一段话,问自己:'这句话有没有可能用更简单的词说?'如果答案是'有可能',那就改。What to take away: Writing is a mirror of thinking. Next time you finish a passage, ask: 'Is there a simpler word for this?' If the answer is yes, change it.
- 我们正在亲手毁掉软件We Are Destroying Software
- 问题:为什么现代软件越来越臃肿、脆弱,黑客的乐趣在消失?多数人把它归咎于个别公司的糟糕决策,Redis作者antirez(2025年旧文)给出一个更冷的答案:凶手不是谁,是一套被普遍奉为圭臬的'最佳实践'。The problem: why is modern software ever more bloated and fragile, and why is the joy of hacking fading? Most people blame individual bad decisions at specific companies. antirez (Redis's author, in a 2025 post) offers a colder answer: the culprit isn't anyone—it's a set of 'best practices' we all treat as gospel.
- 关键想法:他用一篇'连祷文'逐条点名:加功能和做优化时不再考虑复杂度本身;层层加码的构建系统;荒谬的依赖链,让一切又大又脆;对新程序员喊'别重复造轮子'——可重新发明轮子恰恰是理解轮子如何工作、进而造出新轮子的第一步;不再关心API向后兼容;一言不合就重写能用的东西;每个新语言、新范式都追;声称代码注释无用;把软件当成纯工程学科;让简单的操作越来越不简单;追求写得最快而不是设计得最好。The key idea: he writes a litany, naming each one: no longer weighing complexity itself when adding features or optimizing; build systems piled higher and higher; absurd dependency chains that make everything big and brittle; telling newcomers 'don't reinvent the wheel'—when reinventing the wheel is precisely how you learn how wheels work, and the first step toward making new kinds of wheels; ignoring backward API compatibility; rewriting things that work; jumping on every new language, paradigm, and framework; claiming code comments are useless; treating software as a purely engineering discipline; systems that no longer scale down—simple things should be simple; producing code as fast as possible instead of as well-designed as possible.
- 为什么成立:每一条口号都是在优化一个单维度——交付速度、复用率、技术栈新鲜度——而把成本转嫁进一个没人记账的维度:整个系统的复杂度。复杂度不是别的,正是脆弱、昂贵和无人敢碰的根源,而且它只增不减,除非有人刻意偿还。Why it holds: every slogan optimizes one dimension—delivery speed, reuse, tech-stack freshness—while offloading the cost onto the dimension nobody keeps accounts for: total complexity. Complexity is the root cause of brittleness, expense, and code no one dares touch, and it compounds unless someone deliberately pays it down.
- 能带走什么:下次遇到'行业标准做法',把它拆成一条问句:它优化了什么?代价由谁支付?另外,'简单的事情应该保持简单'是系统品质(可缩放性不仅指向上扩展,也指向下),最后一句警告是:这样下去,软件将不再带给我们 hacking 的快乐。What to take away: next time you meet an 'industry standard practice,' turn it into one question: what does it optimize, and who pays for it? And remember that 'simple things should stay simple' is a system property—scalability cuts both ways. His closing warning: keep this up, and what's left of software will no longer give us the joy of hacking.
哲学与生活Philosophy and Life
- 如果计算机数遍了所有花瓣,谁来品尝樱桃?If Computers Count All the Petals, Who Tastes the Cherries?
- 问题:科技行业越来越相信'一切皆可计算'——把人类体验转化为数据就能优化一切。这个信念有问题吗?The problem: The tech industry increasingly believes 'everything is computable'—convert human experience to data and everything can be optimized. Is there a problem with this belief?
- 关键想法:作者描述了自己参与设计智能烤箱的经历:厨师们用'喷灯打色'的直觉来烹饪,但烤箱无法理解这个概念。科技公司的策略是把'人为干预'视为缺陷,要求用户遵循固定流程。作者把这称为'计算主义'——认为宇宙由离散数据位构成,可以被代码治理。The key idea: The author describes helping design a smart oven: chefs used the intuitive technique of 'smacking with a propane torch' for color, but the oven couldn't understand this. The company's strategy was to treat 'human intervention' as a defect. The author calls this 'computism'—the belief that the universe is made of discrete data bits governable by code.
- 为什么成立:品尝樱桃和数花瓣是两件事。数据可以记录颜色、温度、时间,但无法记录'味道'这个主观体验。当设计追求'零摩擦'时,人的主体性就被抹去了——用户不再是'有品味的人',而是'产生数据的节点'。Why it holds: Tasting a cherry and counting petals are two different things. Data can record color, temperature, time—but not the subjective experience of 'taste.' When design pursues 'zero friction,' human subjectivity is erased: users become 'data-generating nodes,' not 'people with taste.'
- 能带走什么:真正的体验无法被完全数据化。好的设计应该保留人的判断空间,而不是用'智能'名义剥夺它。What to take away: True experience cannot be fully datafied. Good design should preserve space for human judgment, not strip it away in the name of 'intelligence.'
- 他大概是对的:一种对待信念的古老智慧He Was Probably Right: An Ancient Wisdom for Holding Beliefs
- 问题:我们应该多紧地抓住自己的信念?太松就什么都没立场,太紧就听不进任何反驳。中间地带在哪里?The problem: How tightly should you hold your beliefs? Too loose and you stand for nothing; too tight and no argument reaches you. Where's the middle ground?
- 关键想法:西塞罗在公元前45年写了《学院派对话录》,提出'probabilia'(似真的东西)的概念:我们可以自信地跟随某些信念行动,但永远不要声称绝对确定。他区分了四个等级的信念:只是'看起来对'、'令人信服'、'与其他信念不冲突'、'经过彻底检验'。小事用前两级,重大决定用后两级。The key idea: Cicero, writing in 45 BCE, proposed 'probabilia'—things we can confidently follow but never claim absolute certainty about. He distinguished four grades of belief: 'seems right,' 'survives scrutiny,' 'coheres with other beliefs,' 'thoroughly examined.' Minor matters need only levels 1-2; major decisions demand 3-4.
- 为什么成立:真正的怀疑论者和'伪怀疑论者'(拒绝更新信念的人)的区别在于:前者面对新证据会更新,后者只是不断挪动门柱。西塞罗的方法不是'什么都不信',而是'graded assent'——根据证据强度给予不同程度的相信。Why it holds: The difference between a genuine skeptic and a contrarian pseudo-skeptic (who never updates) is that the former updates when confronted with new evidence. Cicero's method isn't 'believe nothing'—it's 'graded assent': different levels of belief matched to evidence strength.
- 能带走什么:下次面对一个争议性话题时,问自己:'我现在的相信程度,是否与证据强度匹配?'如果答案是否定的,调整。What to take away: Next time you face a controversial topic, ask: 'Does my current level of belief match the strength of the evidence?' If not, adjust.
工具角Tools Corner
- 离线文件加密——一个自解密HTML页面Air-Gapped File Encryption as a Self-Decrypting HTML Page
- 把文件加密后嵌入一个HTML页面,页面本身包含解密逻辑(Web Crypto API)。任何人用浏览器打开就能解密,无需安装任何软件,也无需联网。Encrypt a file and embed it in an HTML page that contains its own decryption logic (Web Crypto API). Anyone can open and decrypt it in a browser—no software install, no internet needed.
- 核心创意:把'密钥管理'和'解密程序'打包进同一个静态文件。这对需要离线传输敏感数据的场景非常实用。The core idea: Packaging 'key management' and 'decryption' into a single static file. Extremely practical for offline transfer of sensitive data.
- PaperMono——电子墨水购物清单,用你的手机网页驱动PaperMono: E-ink Shopping List Driven by Your Phone's Web Page
- 一个电子墨水显示屏贴在冰箱上,通过手机浏览器网页同步购物清单。没有App,没有蓝牙,没有复杂配对——就是一个网页推送到一块便宜的e-ink屏。An e-ink display on the fridge syncs a shopping list via a mobile web page. No app, no Bluetooth, no complex pairing—just a web page pushed to a cheap e-ink screen.
- 核心创意:用最简单的技术栈(HTML + e-ink display)解决一个日常问题。零依赖,低成本,优雅。The core idea: Solving a daily problem with the simplest possible tech stack (HTML + e-ink display). Zero dependencies, low cost, elegant.
- Squalk——建立在Nostr上的经典论坛引擎Squalk: An Old-School Forum Engine Built on Nostr
- 用Nostr的NIP-29组和NIP-7D线程协议,构建了一个经典论坛风格的讨论平台。没有中心化服务器,没有算法 Feed,没有推荐系统。Built a classic forum-style discussion platform using Nostr's NIP-29 groups and NIP-7D threads. No central server, no algorithmic feed, no recommendation system.
- 核心创意:把20年前的论坛体验带到去中心化社交协议上,证明'复古'也可以很现代。The core idea: Bringing a 20-year-old forum experience to a decentralized social protocol—proving 'retro' can also be modern.
- Jev Ultrafast——用决策模型代替通用大模型的网页AgentJev Ultrafast: A Web Agent Driven by a Decision Model, Not a General LLM
- 网页Agent的每一步其实只是'在页面上选一个操作和一个元素',但主流实现让通用LLM为此生成一整段文本。Jev Ultrafast改用专门的决策模型Jev:每次页面观察生成一张元素表,模型只在允许的操作(CLICK/TYPE_TEXT/SELECT/SCROLL/WAIT/DONE)里做一次选择,只有需要打字时才叫一个小LLM生成文字。Every step of a web agent is really just 'pick an operation and an element on this page,' yet mainstream implementations make a general LLM generate a full text response for it. Jev Ultrafast instead uses a dedicated decision model, Jev: each observation produces a table of elements, the model picks from allowed operations (CLICK/TYPE_TEXT/SELECT/SCROLL/WAIT/DONE), and only when typing is needed does a small LLM generate the text.
- 实测:单次决策延迟中位数约178毫秒,从一句'Zürich到London的机票'的自然语言目标到结果页呈现全程7.1秒(含文本生成和页面加载)。Measured: median Jev decision latency around 178ms; from the one-sentence goal 'Zürich to London on Google Flights' to the verified results page took 7.1 seconds total, including text generation and page loads.
- 核心创意:把'生成动作'降级成'分类动作'。网页操作里绝大多数步骤不需要语言创造力,只需要一个正确选择——动作空间建得越好,模型就能小到几乎免费。The core idea: downgrading 'generating an action' to 'classifying an action.' Most web-operation steps need no linguistic creativity, just one correct choice—and the better the action space is engineered, the smaller (and cheaper) the model can be.
- 可以带走的启发:评估一个Agent的成本时,先看它每步真正的决策是什么形态;如果输出只有几个比特,就不该为它付一整段文本的钱。Takeaway: when costing an agent, first look at what each step's decision actually is. If the output is a few bits, don't pay for a paragraph of text to produce it.
- Npunlock——在Intel NPU上运行自定义C内核Npunlock: Run Custom C Kernels on Intel NPUs
- Intel NPU通常只接受封闭的推理框架。Npunlock允许开发者直接用C语言编写内核并加载到NPU上运行。Intel NPUs typically only accept closed inference frameworks. Npunlock lets developers write kernels directly in C and load them onto the NPU.
- 核心创意:把NPU从'黑盒推理芯片'变成'可编程加速器',打开了一个被厂商锁住的能力空间。The core idea: Turning an NPU from a 'black-box inference chip' into a 'programmable accelerator,' unlocking a capability space that manufacturers had locked away.
- Chess Postmortem Skills——用Claude分析你的棋局Chess Postmortem Skills: Analyze Your Games with Claude
- 一个Claude Code skill,上传你的棋局PGN文件,它自动分析错着、最佳着法和战略主题。A Claude Code skill: upload your game's PGN file, and it automatically analyzes blunders, best moves, and strategic themes.
- 核心创意:把专业棋手复盘的工作流程自动化——不需要安装ChessBase,一个skill就够了。The core idea: Automating the professional chess player's post-game review workflow—no need for ChessBase, one skill does it all.
- Make Cursed Fonts——生成搞怪字体Make Cursed Fonts: Generate Absurd Custom Fonts
- 一个网页工具,让用户通过可视化界面扭曲、拼接、重组字体字形,生成'诅咒字体'(如Times New Bastard)。A web tool that lets users distort, merge, and reassemble font glyphs through a visual interface, generating 'cursed fonts' (like Times New Bastard).
- 核心创意:把字体设计变成一种即兴创作工具,门槛极低,结果出人意料。The core idea: Turning font design into an improvisational creative tool with extremely low barriers and unexpectedly creative results.