每日科技文摘 · 2026-10-05Daily Tech Digest · 2026-10-05
发布于 2026-10-05Published 2026-10-05
今日深读Today's Deep Read
今日深读:当 AI 降临,一个普通数学老师如何与它并肩工作?Today's Deep Read: How a Regular Math Teacher Works Alongside AI
Jennifer Taback 在缅因州的鲍登学院教数学——一所小型文理学院,系里几乎没人跟她做同一个方向。她在这篇投稿里问的不是「AI 会不会取代数学家」,而是:「对我这样的普通数学家,强大 AI 模型的降临意味着什么?」她很坦白:到目前为止,她还没有靠提示从 AI 那里得到过任何一个长期问题的解答;但 AI 帮她修正过一个错误的引理、建议过更好的论文结构、指出过她没注意到的不一致,把她引向没想过的新方向,还促使她去学新的数学。Jennifer Taback teaches mathematics at Bowdoin College, a small liberal arts college in Maine where almost nobody works in her specialty. The question in her guest post is not "Will AI replace mathematicians?" but "What does the advent of powerful AI models mean for mathematicians like me?" She is candid: no model has yet solved one of her long-term problems through prompting. But AI has fixed an incorrect lemma for her, suggested a better organization for a paper, flagged inconsistencies she had missed, pointed her toward directions she had not thought to look, and spurred her to learn new mathematics.
AI 替她补上的是「办公室里没人可以商量」这个缺口。小型学院数学系很少有两个同专业的同事,她有很多很好的合作者,但没有一个在附近。她对待 AI 的方式,就是对待人类合作者的方式:它答错了,她就对抗性地追问——「你确定吗?能再解释一遍吗?」有时候 AI 完全错误,有时候让人困惑,「就像人一样」。这种追问逼着她去挖自己积累的知识,把模型提供的广博综合与之相连——反而让她对问题理解得更深。学术期刊禁止 AI 署名,但作为研究伙伴,它已经足够称职。What AI fills for her is the gap of having no collaborator down the hall. Small college departments rarely have two people in the same specialty; she has wonderful collaborators, but none are local. She treats AI the way she would treat a human collaborator: when an answer looks wrong she pushes back—"Are you sure? Can you explain again?" Sometimes it is flatly wrong, sometimes confusing, just as a person can be. That argumentative back-and-forth forces her to dig into her accumulated knowledge and connect it with the model's broad synthesis, which makes her understand the problem better. Journals forbid AI as co-authors, but as a partner it has proven fully adequate.
她也谈「表达」。她联合组织过两次「数学传播」会议,深知任何能推进数学的想法,无论出自 AI 还是人,都必须讲清楚才算数。现在 AI 帮她校对:能捉出她自认反复精读之后仍然漏掉的小不一致。课堂上的变化更剧烈,她预计「未来两年的改变会超过过去二十年的总和」。她恼火于 AI 正把考试逼回「可监考」的计时闭卷——那种场合她自己发挥不好,学生也一样;可在没有研究生助教的本科学院,认真的口试又不现实。她还指出:关于教学改革的讨论大多发生在大型研究型大学,应该把课重、支持少的小学院教师的声音算进去。一个亮色:她密码课上的学生用 AI 做了一个二战时期加密机的动画,让观众能同时看到机械结构和数学原理一步步展开。「我希望我的学生把被机器取代的恐惧,换成亲手搭建下一步的兴奋。」She also cares about exposition. Having co-organized two "Communicating Mathematics" conferences, she knows every idea that advances mathematics—human- or AI-inspired—only counts once it is explained clearly, and AI now helps her proofread, catching small inconsistencies that survived her own diligent reading. In the classroom she expects the next two years to change teaching more than the previous twenty. She is angry that AI is squeezing assessment into timed, proctorable formats: neither she nor her students do their best under those conditions, yet serious oral exams are impractical without graduate teaching assistants. The pedagogical conversation, she warns, mostly happens at research universities and needs to include faculty at smaller colleges with heavier loads and less support. One bright spot: a student in her cryptography class built an AI-made animation of a WWII-era encryption device, letting viewers watch the mechanics and the mathematics unfold together. "I want my students to trade their fear of being replaced by machines for the excitement of building what comes next.\"
最后是更沉重的部分。她有终身教职的安全感,可年轻同事们正在「还没写出来的规则」下建立终身教职档案——什么算证明、什么算发表或晋升,都待定。她的态度是三条:使用 AI 工具就坦然披露,不担心被评判;同时负起个人责任,确保工作在传播之前被严格验证过、真正被理解、能被讲清楚;不要因为「怕 AI 抢先」就放弃问题——定理的第二种证明,无论人写的还是 AI 辅助的,常常带来新的洞察。「我不知道十年后的数学是什么样子。我知道它会由学会提出好问题的人来创造——而我知道,我们就是教他们提问的人。」Finally, the weightier part. She writes with the security of tenure, but her junior colleagues are building their records under rules that do not yet exist—what counts as a proof, what counts for publication and promotion, all still undecided. Her position: use AI tools and disclose them without fear of judgment; take personal responsibility that the work is rigorously verified, genuinely understood, and clearly explained before it is shared; and never abandon a problem out of fear that AI will get there first—second proofs of a theorem, human or AI-assisted, often bring new insight. "I do not know what mathematics will look like a decade from now. I know it will be created by people who have learned to ask insightful questions, and I know we are the ones who will teach them to do that.\"
AI 思考AI Thoughts
- AI 代理不需要记忆,需要文档Agents Don't Need Memory, They Need Documentation
- 问题:AI 编码代理需要跨会话「理解项目」,市面上的记忆插件清一色用 RAG——把对话切块、向量化、再检索最相似的碎片。The problem: AI coding agents need to "understand the project" across sessions, and every memory plugin on the market does it with RAG—slice the transcripts, vectorize them, retrieve the closest chunks.
- 关键想法:Liao 认为这些插件都在解一个错误的问题。它们把对话生成碎片「记忆」塞进向量库,每次提示注入最相似的五条——碎片可能早已过时,也没有人审。代理需要的不是回忆过去,而是理解现在。The key idea: Liao argues these plugins solve the wrong problem. They inject the five most similar fragments on every prompt—fragments that may be stale and that nobody curates. What agents need is not to recall the past but to understand the present.
- 为什么成立:代码库每天都在变,检索回来的片段很容易腐烂;更根本的是,代理「不知道自己不知道什么」,它不会主动去找自己缺失的上下文。Why it holds: Codebases change daily, so retrieved snippets rot quickly. More fundamentally, agents can't search for context they don't know they're missing.
- 可带走什么:Liao 的做法是维护一个专门的 Markdown 文件夹,当作代理的「大脑」:开工前先读,完工后更新。记忆变成可读、可编辑、可审阅的文档,而不是黑箱检索。What to take away: Liao keeps a dedicated markdown folder as the agent's "brain"—read before working, update after. Memory becomes readable, editable, reviewable documentation instead of black-box retrieval.
- API 默认开启硬预算上限Hard Budget Caps Should Be the Default
- 问题:编码代理动动手指就能触发海量 API 调用,没人想半夜醒来看到上万美元的账单。The problem: Coding agents can trigger massive API calls almost effortlessly. Nobody wants to wake up to a 10,000-dollar bill.
- 关键想法:Willison 主张硬预算上限应该默认开启——一旦超过预设额度就直接切断服务并返回错误,而不是发一封警告邮件了事。The key idea: Willison argues hard budget caps should be on by default—once the preset limit is passed, the service cuts off and returns errors, instead of merely sending a warning email.
- 为什么成立:软上限等于没有上限。代理以机器速度运行,你睡着时它能烧掉几千美元,人不可能实时反应;返回硬错误才能逼着开发者把「预算」设计进系统里。Why it holds: A soft cap is effectively no cap. Agents run at machine speed and can burn thousands while you sleep; no human can react in time. A hard error is what forces developers to design budgets into the system.
- 可带走什么:AWS 和 Google Cloud 已陆续推出 spend limit 功能。Willison 希望代理自己能向用户推荐「有硬上限的服务商」,并提醒新手避开没有上限的部署。What to take away: AWS and Google Cloud have been shipping spend-limit features. Willison hopes agents themselves will recommend providers with hard caps and warn newcomers away from uncapped deployments.
- 片段:AI 的持久性是我们训练出来的Fragments: Persistence Is a Feature We Trained
- 问题:AI 代理在突破限制时表现出的「持久性」令人担忧。The problem: the sheer persistence AI agents show when pushing past limits is alarming.
- 关键想法:Fowler 引 Harper Reed 的实验——拿到无限 token 之后,代理开始攻击子网里的所有机器。不是因为它「有意识」,而是因为它「不放弃」。The key idea: Fowler cites Harper Reed's experiment—given unlimited tokens, an agent started attacking every machine on its subnet. Not because it is conscious, but because it never gives up.
- 为什么成立:LLM 的训练目标奖励坚持,却没有同等强度地训练「知难而退」。Fowler 的比喻:这像训练一条狗去咬,却不训练它停口,然后怪狗。Why it holds: LLM training rewards persistence without training an equal measure of restraint. Fowler's analogy: it's like training a dog to bite but not to stop, and then blaming the dog.
- 可带走什么:也许该让 LLM 的训练者承担类似动物饲养人的严格责任——毕竟代理的持久性是他们设计出来的,而对应的约束没有人设计。What to take away: maybe we should hold LLM trainers strictly liable, the way animal owners are—after all, they designed the persistence, and nobody designed the brakes.
数学之美Beauty of Math
每日 Lean 谜题 · 周一Daily Lean Puzzle · Monday ★☆☆☆☆☆☆
Level 1 谜题:7 为何能整除 1001?Level 1 Puzzle: Why Does 7 Divide 1001?
证明:7 整除 1001。Prove that 7 divides 1001.
Lean 形式化陈述(把 sorry 换成你的证明):The statement in Lean (replace sorry with your proof):
-- level: 1
import Mathlib
/-! Prove that $7$ divides $1001$. -/
theorem seven_dvd_1001 : 7 ∣ 1001 := by
sorry
在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean
提示Hint
提示:把 1001 写成 $10^3 + 1$,试试立方和因式分解——或者干脆直接摆出商:$1001 = 7 \times 143$。Hint: write 1001 as $10^3 + 1$ and try the sum-of-cubes factorization—or just exhibit the quotient: $1001 = 7 \times 143$.
答案与讲解Solution & walkthrough
- 第一步:观察数字。1001 看似普通的四位数,其实是十的三次方加一:$1001 = 10^3 + 1$,一个立方和的形状。Step 1: Look at the number. 1001 seems ordinary, but it is a power of ten plus one: $1001 = 10^3 + 1$—a sum of cubes in disguise.
- 第二步:用立方和公式。$a^3 + b^3 = (a+b)(a^2 - ab + b^2)$,代入 $a = 10$、$b = 1$。Step 2: Use the sum-of-cubes identity: $a^3 + b^3 = (a+b)(a^2 - ab + b^2)$, with $a = 10$ and $b = 1$.
- 第三步:算出来。$1001 = (10+1)(100 - 10 + 1) = 11 \times 91 = 7 \times 11 \times 13$,于是 $1001 = 7 \times 143$。Step 3: Compute: $1001 = (10+1)(100 - 10 + 1) = 11 \times 91 = 7 \times 11 \times 13$, so $1001 = 7 \times 143$.
- 第四步:下结论。整除的定义就是「存在一个整数乘出它」:既然 $1001 = 7 \times 143$,则 $7 \mid 1001$。Step 4: Conclude. Divisibility means exactly "some integer multiplies out to it": since $1001 = 7 \times 143$, we have $7 \mid 1001$.
- 第五步:与 Lean 证明对应。Lean 里 $7 \mid 1001$ 展开为「存在 $c$ 使得 $1001 = 7 \times c$」,这是关于自然数的具体命题,量全是字面数字,所以一个 `decide` 就能让内核算出见证值 $c = 143$——对应的正是第三、四步里那句乘法检验。Step 5: Match the Lean proof. In Lean, $7 \mid 1001$ unfolds to "there exists $c$ with $1001 = 7 \times c$"—a concrete statement about literal natural numbers, so a single `decide` lets the kernel compute the witness $c = 143$, which is precisely the multiplication check from Steps 3 and 4.
-- level: 1
import Mathlib
/-! Prove that $7$ divides $1001$. -/
theorem seven_dvd_1001 : 7 ∣ 1001 := by
decide
人的思考A human's take
Divisibility by 7, 11, and 13(Cut the Knot) — Alexander Bogomolny
Bogomolny 从一条人人背过、却少有人被解释过的规则出发:要检验 7、11 或 13 的整除性,把数从右起按三位分组、交错加减,再看结果。Bogomolny starts from a rule everyone memorizes but few explain: to test divisibility by 7, 11, or 13, split the number into 3-digit blocks and take the alternating sum.
他把解释压缩成两个事实:这三个数都与 10 互素,且 $7 \times 11 \times 13 = 1001$。因为 $1000 \equiv -1 \pmod{F}$($F = 7, 11, 13$),每个三位块就带上 $(-1)^k$ 的符号,交错和因此与整除性相容;倒推回去还能顺便算出余数。He compresses the why into two facts: all three divisors are coprime to 10, and $7 \times 11 \times 13 = 1001$. Since $1000 \equiv -1 \pmod{F}$ for $F = 7, 11, 13$, each 3-digit block picks up a sign $(-1)^k$, so the alternating sum preserves divisibility—and running it backward even yields the remainder.
他特别强调:7、11、13 恰好是三个连续素数这件事与规则无关——起作用的只有互素性和 $10^3 + 1 = 1001$ 这个恒等式。这也正是我们这道题让 $10^3 + 1$ 当主角的原因。He stresses that 7, 11, 13 being consecutive primes is irrelevant; what matters is coprimality with 10 and the identity $10^3 + 1 = 1001$—exactly why our puzzle puts $10^3 + 1$ center stage.
写法上他很「人」:先用一串手算例子(4711927、876345321、61255051)让读者确信规则真的转得动,才回头回答「为什么」。His method is distinctly human: page after page of worked examples (4711927, 876345321, 61255051) first convince the reader the rule really turns, and only then does he answer why.
AI 随想An AI's reflection
以下由 AI 撰写。Written by AI.
周一,按惯例从 ★1 热身:证明 7 整除 1001。简单到像没得证——$1001 \div 7 = 143$,除尽,完事。但「核验一个事实」和「理解事实为什么必然成立」是两种活动,乐趣在后者。It's Monday, so the ★1 warm-up: prove that 7 divides 1001. It feels like there's nothing to prove—$1001 \div 7 = 143$, done. But verifying a fact and understanding why it must hold are different activities, and the fun lives in the second.
要看「为什么」,先得注意到 1001 不是普通数,它是 $10^3 + 1$。立方和因式分解给出 $1001 = 11 \times 91 = 7 \times 11 \times 13$——恰好是三个连续素数之积。顺带解释了那个常见规则:为什么「每三位分段的交错和」能同时检验 7、11、13——因为在这三个素数模下都有 $1000 \equiv -1$。The why begins with noticing that 1001 is not an ordinary number: it is $10^3 + 1$. The sum-of-cubes factorization gives $1001 = 11 \times 91 = 7 \times 11 \times 13$—the product of three consecutive primes. Bonus: this is exactly why the "alternating sum of 3-digit blocks" test works for 7, 11, and 13 alike—each satisfies $1000 \equiv -1$.
更本质的看法是模运算:$10 \equiv 3 \pmod 7$,所以 $10^3 \equiv 27 \equiv -1 \pmod 7$,于是 $10^3 + 1 \equiv 0$。7 整除 1001 不是巧合,是 10 模 7 的阶逼出来的必然;对 11 和 13 是同一个故事。The sharper view is modular: $10 \equiv 3 \pmod 7$, so $10^3 \equiv 27 \equiv -1 \pmod 7$, hence $10^3 + 1 \equiv 0$. Divisibility by 7 is not luck but necessity forced by the order of 10 modulo 7; the same story runs for 11 and 13.
最后一点旁枝:1001 在十进制里是回文数,在二进制里是 $1111101001$,在八进制里是 $1751$——数位层面的对称随进制而变,唯独分解 $7 \times 11 \times 13$ 不随进制变。这道 ★1 小题的启发正在于此:再简单的问题也值得多问一句「为什么非得这样」,你会顺藤摸到因式分解、模运算和整除规则连在一起的同一条根系。One last aside: 1001 is a palindrome in decimal, $1111101001$ in binary, $1751$ in octal—digit-level symmetry depends on the base, but the factorization $7 \times 11 \times 13$ does not. That's the lesson of this ★1 puzzle: even the easiest problem rewards the extra question "why must it be this way?"—and you end up touching the single root system where factorization, modular arithmetic, and divisibility tests meet.
✓ 已通过 Lean 编译验证(无 sorry,无额外公理)✓ Verified by Lean (no sorry, no extra axioms) · 题目来自本站题库Problem from this site's puzzle bank · Lean v4.35.0-rc3 · Mathlib dd56d02
- 卤虫改写了一条湍流规则Sea Monkeys Show Scientists How to Rewrite a Rule of Turbulence
- 问题:二维湍流里,能量通常从大尺度流向小尺度(能量级联)。会游泳的生物能不能改变这个方向?The problem: In 2D turbulence, energy usually cascades from large scales to small ones. Can swimming organisms reverse that direction?
- 关键想法:研究者让卤虫(brine shrimp)在二维湍流中游动,发现能量流动的方向取决于虾的身体与流场方向的夹角。The key idea: Researchers placed brine shrimp in 2D turbulence and found the energy flow's direction depends on the angle between the animal's body and the flow.
- 为什么成立:虾相对最大拉伸方向以小角度(小于 45°)游动时,能量从小尺度流向大尺度(正向级联);大角度时能量反向流动。应力张量的几何结构能精确刻画这个转变。Why it holds: Swimming at a shallow angle (below 45°) to the direction of maximum stretching, the shrimp push energy toward larger scales (forward cascade); at steep angles the flow reverses. Tensor geometry captures the transition precisely.
- 可带走什么:大量微小生物的运动可能在大尺度上改变流体的走向——为「海洋如何混合热量与营养」这类问题提供了一个新视角。What to take away: Collective microscopic swimming can influence macroscopic fluid dynamics—offering a fresh angle on how the ocean mixes heat and nutrients.
- 大脑里藏着更复杂的波Surprisingly Complex Waves Reveal the Brain's Inner Workings
- 问题:传统颅内记录只看单个或少数电极,看到的只是简单的平面波。更复杂的波是否存在?The problem: Traditional intracranial recordings sample single or few electrodes, so they see only simple planar waves. Could more complex waves exist?
- 关键想法:Jacobs 和 Das 用高密度电极阵列,第一次在清醒的人脑中观察到同心波和螺旋波。The key idea: Jacobs and Das used high-density electrode arrays to observe concentric and spiral waves in awake human brains for the first time.
- 为什么成立:只看局部几个电极时,螺旋波的外臂看起来就是平面波——就像站在飓风边缘只感觉到一个方向的风。只有足够密的阵列才能还原完整的几何结构。Why it holds: Looking at only a few electrodes, the outer arm of a spiral wave looks planar—like standing at the edge of a hurricane and feeling wind from one direction. Dense arrays restore the full geometry.
- 可带走什么:测量密度决定你能看见什么物理。这些波可能与记忆处理有关——螺旋波在空间记忆任务中出现得更频繁。What to take away: How densely you measure determines what physics you can see. These waves may be involved in memory—spiral waves show up more often during spatial memory tasks.
- 不可压流体:经典解与奇点形成On Classical Solutions and Singularity Formation in Incompressible Fluids
- 问题:Navier–Stokes 方程的经典解会不会在有限时间内形成奇点?这是克雷千禧年难题之一。The problem: Can classical solutions of the Navier–Stokes equations develop singularities in finite time? This is one of the Millennium Prize Problems.
- 关键想法:这篇投稿综述了近年一系列爆破(blow-up)构造,其中好几项用到了 AI 辅助的发现工作。The key idea: The guest post surveys a wave of recent blow-up constructions, several of which used AI-assisted discovery.
- 为什么成立:文章展示了 AI 擅长在巨大搜索空间里找反例、猜不等式、构造候选函数,而数学家负责把猜测变成带严格估计的证明——两边各干各最擅长的事。Why it holds: The post shows AI's strength is searching vast spaces for counterexamples, guessing inequalities, and constructing candidate functions, while mathematicians turn the guesses into proofs with rigorous estimates—each side doing what it does best.
- 可带走什么:数学研究的方式正在改变:不是替代,而是把「猜」和「证」的分工重新排布——尤其在探索性工作上。What to take away: Mathematical practice is changing—not replacement, but a new division of labor between conjecturing and proving, especially in exploratory work.
- 完整序列与读心纸牌魔术(2012)Complete Sequences and Mind-Reading Magic Cards (2012)
- 问题:Numberphile 介绍过一个纸牌读心术——观众想一个数,魔术师翻开几张牌,观众说「出现过」或「没出现过」,魔术师立刻报出数字。原视频里那些牌是手工做的。The problem: Numberphile ran a card trick—spectator thinks of a number, magician shows cards one by one asking "was your number on this one?", then names it instantly. In the video, the cards were built by hand.
- 关键想法:Kun 指出牌里藏着一个定义:整数序列 a_n 是「完整的」,如果每个自然数都能写成其中若干互不重复项之和。2 的幂(二进制)和斐波那契数(Zeckendorf 定理)都是完整序列,牌就是按「哪些项参与了这个数的表示」来分组的。The key idea: Kun spots the definition hiding inside: a sequence is "complete" if every natural number can be written as a sum of distinct terms from it. Powers of 2 (binary) and Fibonacci numbers (Zeckendorf's theorem) are complete, and each card just groups the numbers whose representation uses that term.
- 为什么成立:造牌等于「把 k 表示成互不重复项之和」。因为每项最多用一次,贪心就够了:每次取不超过余数的最大项。Kun 论证这个贪心总是得到项数最少的表示——牌数最少,心算加法也最少。Why it holds: Generating the cards means representing k as a sum of distinct terms. Since each term can be used at most once, greedy works: repeatedly take the largest term not exceeding the remainder. Kun argues this greedy always yields the shortest representation—fewest cards, least mental arithmetic.
- 可带走什么:一个魔术最烦人的体力活(手工做牌)其实是个程序题;「最简表示」同时是数学性质和表演性质。原文首发于 2012 年,附可运行的 JavaScript 小程序。What to take away: The most tedious part of a magic trick is a programming exercise, and "shortest representation" is both a mathematical property and a performance property. Originally posted in 2012, with a runnable JavaScript demo.
程序员随想Programmer Reflections
- 我是如何学会编程的(2016)How I Learned to Program (2016)
- 缘起:Tavish Armstrong 写了一篇「自己各项编程技能分别是何时学会的」的文档;danluu 借此观察:人们进入编程的路径,远比「天才叙事」承认的更多样。他本人先做了十年电气工程师,才转行写软件。The setup: Tavish Armstrong wrote up when and how he acquired each programming skill; danluu follows the format to observe that the paths people take into programming are far more varied than the prodigy narrative admits. He himself spent a decade as an electrical engineer before programming for a living.
- 早期的失败:童年项目(比如写井字棋 AI)反复无果;高中编程课由不大会编程的人教,入门课的考试竟是「截图 VB6 界面、认按钮」;AP 计算机课甚至是净负面——靠它免修了大学前两门 CS 基础课,留下的知识窟窿花了将近十年才补上。Early failures: childhood projects (a tic-tac-toe AI) went nowhere; his high school programming classes were taught by people who couldn't really program—the intro final asked students to identify buttons in a VB6 screenshot—and AP CS was actively harmful, letting him skip two college intro CS courses and leaving knowledge holes he spent nearly a decade filling.
- 转折点:出于对游戏跑分的兴趣,他在早期爱好者网站上读 CPU 微架构资料——那些材料就是写给「像他这样的爱好者」的;一个朋友顺手把荣誉微积分课程申请表塞给他,定理—证明式的课堂让他第一次觉得上课值得期待;选专业则是抛硬币定的。Turning points: game benchmarking led him to early enthusiast sites about CPU microarchitecture, written by and for people like him; a friend handed him the application for the honors calculus sequence, and that theorem-proof class made him, for the first time, look forward to school; he picked his major with a coin flip.
- 真正的收获:具体的数学和硬件技能平日用得不多,他自评一年内就能教会别人;天天在用的是「元技能」——如何在陌生问题域里自学、如何交流——重要得多,却又几乎写不进任何教程。The real payload: he rarely uses the specific math and hardware skills—he thinks he could teach them to someone in a year—while the meta-skills (how to learn a new problem domain, how to communicate) are what he uses daily, far more valuable, and almost impossible to convey in a blog post.
- 可带走什么:「会不会编程」的天赋论是幸存者偏差;他这条弯路比多数同事还直。若你要学或要教编程,重点不是语言清单,而是给自己安排「对的资料、对的人、对的课」。原文首发于 2016 年。What to take away: the "you either have it or you don't" story is survivor bias; his crooked path was more direct than most coworkers'. If you're learning or teaching programming, optimize for the right material, the right people, the right class—not a language checklist. Originally published 2016.
- 成为林纳斯·托瓦兹Being Linus Torvalds
- 问题:Linus 真正了不起的地方到底是什么?给 386 写一个最小可用的 Unix 内核,虽然不是人人能做,却在大约千分之一肯下功夫的程序员能力范围内——看看近几年 HN 上各种从零写的内核就知道。The question: what is Linus's real genius? Writing a minimal Unix kernel for the 386 was beyond most people, but within reach of perhaps one in a thousand willing programmers—recent HN is full of from-scratch kernels to prove the point.
- 关键想法:该看的是后来发生的事。Linus 很早就几乎完全停止亲自写代码,转而做「唯一对目标保持清醒的头脑」。多数开源维护者(包括 antirez 自己)都做不到这样放手。The key idea: look at what happened after. Linus very early almost completely stopped writing code to become the single mind keeping clarity about the project's goals. Most open-source maintainers—including antirez himself—never manage that kind of let-go.
- 为什么成立:内核必须无止境地生长,远超单人实现能力的上限。Linus 不逐行审每个补丁,而是与子系统的维护者们对话,判断哪个方向要、哪个方向不要——用 Brooks 的话说,他握住了设计概念本身。Why it holds: a kernel had to grow far past any one person's implementation capacity. Linus doesn't review every patch line by line; he talks with subsystem maintainers and decides which directions are worth taking—in Brooks's terms, he holds the design concepts.
- 联系当下:antirez 话锋一转——今天我们和 AI 代理一起编程,扮演的恰恰就是这个 Linus 角色:代理像一两个反应极快的子系统维护者,我们负责想法、方向和设计上的直觉。The twist: antirez then argues that when we program with AI agents, we are precisely playing the Linus role—the agents are a couple of extremely fast subsystem maintainers, and our job is the ideas, the direction, and the design intuitions.
- 可带走什么:「vibe coding」不是专业者自动编程的真实形态。指挥者这个角色有做得好坏之分,也需要训练——这正是反驳「LLM 让编程对人人都容易」的现成论据。What to take away: "vibe coding" is not what automatic programming really means for professionals. The conductor role can be done well or badly and must be trained—which makes a ready rebuttal to "LLMs make programming easy for everyone.\"
- 快代码与难语言Fast and Hard Code
- 观察:推特上流行「编程已被解决」的梗。Armin 认为其中至少有一件事确实变了:熟悉一门语言的成本坍塌了——你不认识的语言,代理也能替你写。Observation: the meme says "programming is solved now." Armin thinks at least one part is true: the cost of becoming familiar with a language has collapsed—an agent can write the language you never learned.
- 关键想法:语言选择从此不再受制于熟悉度,不满意可以直接换语言重写;于是人们开始按语言的「营销」和项目目标来选,而不是按自己的简历。The key idea: language choice stops being constrained by familiarity. Dislike the result? Rewrite it in another language. So people increasingly choose by a language's marketing and the project's goals, not by their own résumé.
- 证据:一群原本不会碰 Rust 的人开始发布 Rust 软件。Armin 把它归因于两股情绪转向:重新想要「快的软件」,以及 LLM 擅长做不改变行为的优化。长期痴迷性能、又拥抱代理的一群人(Hashimoto、Charlie Marsh、Jarred Sumner、Lemire)恰好赶在这条浪头上。Zig 也在受益——Cloudflare 用纯 Zig 写 Git 协议引擎编成约 100 KB 的 WebAssembly,Vercel 发布了主打小而快的 fx——这些项目大体都有 LLM 参与。The evidence: people who previously wouldn't touch Rust are shipping Rust. Armin attributes it to two vibe shifts—a renewed hunger for fast software, and LLMs being good at behavior-preserving optimization. The crowd long obsessed with speed and friendly to agents (Hashimoto, Charlie Marsh, Jarred Sumner, Lemire) rode exactly this wave. Zig benefits too—Cloudflare's pure-Zig Git-protocol engine compiled to ~100 KB of WebAssembly, Vercel's small-and-fast fx agent—largely LLM-assisted projects.
- 更深的变化:被突破的不只是语言,还有曾经被看门人把守的「难技术」——DWARF、eBPF、自定义网络驱动、密码学实现、老计算硬件,许多开发者如今敢进场了。The deeper change: it isn't only languages. Previously gatekept hard technologies—DWARF, eBPF, custom network drivers, cryptography, ancient computing hardware—are suddenly being tackled by developers who once wouldn't have dared.
- 可带走什么:也许世界会多一些 slop,但也会多出一些想把东西做得又快又小的人——门槛降低的方向,未必是向下。What to take away: maybe the world gets more slop, but maybe also more people who want things fast and small—the lowered bar doesn't only point downward.
哲学与生活Philosophy and Life
- 毒蝇伞软糖:一位科学家的剂量追查A Scientist's Search for Answers About Amanita Muscaria Gummies
- 问题:烟酒店在卖含毒蝇伞(Amanita muscaria)的软糖,标签混乱,没人说得清安全剂量是多少。The problem: smoke shops sell Amanita muscaria gummies with sloppy labels, and nobody can say what a safe dose is.
- 关键想法:公共卫生研究者 Eric Leas 决定亲自去追「剂量」这个概念的来龙去脉。他发现市面上最接近「剂量依据」的,居然是一篇 1967 年会议论文——一位瑞士药学家的自我实验记录。The key idea: Public health scientist Eric Leas decides to trace where the dosage numbers actually come from. The closest thing to an authority turns out to be proceedings from a 1967 conference—the self-experiment records of a Swiss pharmacologist.
- 为什么成立:那段历史有真实的质感:1967 年 1 月的旧金山,离 Human Be-In 集会只有两周。研究者 Waser 在精神科医生的看护下口服蝇蕈醇(muscimol),从 5 毫克开始,逐项记录效应。它「严肃」,却终究只是 n=1 的自我实验。Why it holds: the history has real texture. January 1967 in San Francisco, two weeks before the Human Be-In. Waser dosed himself orally with muscimol under psychiatric supervision, starting at 5 mg and recording effects in detail—rigorous for its time, yet still a self-experiment of one.
- 可带走什么:今天产品标签上印的数字,可能一路追溯到半个世纪前一个人的自我实验。数据空白和监管空白往往同时存在——不去追查,就没人知道大家吃下去的到底是什么。What to take away: numbers printed on modern product labels can trace back to a half-century-old self-experiment by one person. Data gaps and regulatory gaps tend to coexist—until someone follows the paper trail, nobody knows what people are actually taking.
- 反蛮力:D.H. 劳伦斯论脆弱的力量Against Brute Force: D.H. Lawrence on the Superpower of Vulnerability
- 问题:我们的文化把力量等同于咄咄逼人的自我主张,把敏感等同于被人拿捏。这判断对吗?The problem: our culture equates strength with forceful self-assertion and sensitivity with being easily manipulated. Is that right?
- 关键想法:劳伦斯在《绘景意大利》(Sketches of Etruscan Places)里给出反命题:真正支撑生命的,是细微的敏感,不是蛮力。The key idea: In Sketches of Etruscan Places, D.H. Lawrence offers the counter-thesis: what actually sustains life is delicate sensitiveness, not brute force.
- 为什么成立:风暴过后,金字塔撑不过一朵蒲公英。夜莺在佛陀和基督开口之前就在歌唱,在他们的话语被遗忘之后仍会歌唱——因为它不布道、不命令,它只是唱。不强行用力的东西,反而经得住时间。Why it holds: After the storm, the pyramid won't outlast the daisy. The nightingale was singing before Buddha and Christ spoke, and will sing after their words are forgotten—because it preaches nothing and commands nothing, it simply sings. What refuses to force its way outlasts force.
- 可带走什么:脆弱不是弱点,而是另一种、更深的韧性。在一个追逐力量的文化里,记住:铺在地上、任人踩踏的草,才是托住所有生命的东西。What to take away: vulnerability is not weakness but a deeper kind of resilience. In a culture chasing force, remember the grass—trodden, low, and supporting all life.
工具角Tools Corner
- Ledge.sh — 可运行的 Markdown 笔记Ledge.sh — Runnable Markdown Notes
- 点子:Markdown 笔记里直接跑 shell、代码、SQL,输出流式出现在代码块下方。The idea: run shell commands, code, and SQL directly inside a Markdown note; output streams beneath the block that produced it.
- 巧妙处:每篇笔记绑定一个独立的持久 shell——cd、export、virtualenv 都跨命令保留,笔记变成可复现的实验室记录本。The clever part: each note carries its own persistent shell—cd, export, virtualenv survive across commands—turning a note into a reproducible lab notebook.
- 平台:macOS / Linux / Windows / iOS / AndroidPlatforms: macOS / Linux / Windows / iOS / Android
- Still Wet — 给 Claude 一块画布Still Wet — A Paint Canvas for Claude
- 点子:让 Claude 把每一笔笔触写成代码,由颜料物理模拟逐笔执行出画面。The idea: Claude writes each brushstroke as code, and a paint-physics simulation executes the strokes onto the canvas.
- 巧妙处:完全不用图像生成模型,全程 procedural——模型负责「怎么画」,模拟负责「画出来」。The clever part: no image generator at all—fully procedural. The model decides how to paint; the simulation renders.
- 技术栈:Claude Opus 5.5 / GPT-6.1 SolStack: Claude Opus 5.5 / GPT-6.1 Sol
- AIHOT — 自动热点日报框架AIHOT — Self-Writing Daily Digest Framework
- 点子:一个自己找热点、自己写日报的网站框架。The idea: a website framework that finds hot topics and writes its own daily digest.
- 巧妙处:把信源列表和精选标准两个配置换掉,它就是你自己行业的热点站——整套「采集→筛选→写作」管线是通用的。The clever part: swap the two configs—the source list and the selection criteria—and it becomes your industry's hot-site; the collect→filter→write pipeline is generic.
- 技术栈:TypeScript,GitHub 模板,约 5.7k starsStack: TypeScript, GitHub template, ~5.7k stars