每日科技文摘Daily Tech Digest
发布于 2026-10-13Published 2026-10-13
今日深读Today's Deep Read
今日深读:「抱歉,你还是得自己思考」——从 DHH 的 AI 重写实验说起Today's Deep Read: "Sorry, But You Still Have to Think" — What DHH's AI Rewrite Experiment Really Shows
2026 年 10 月,Rails 之父 DHH 让 AI 代理把聊天应用 Campfire Once 分别重写成 Rust、Elixir 和 Go,然后像办拍卖会一样发布性能对比;他自己坦白「受不了读和写 Rust」。一份「不用读代码」的产物长什么样?资深工程师 Piotr Sarnacki 把三个重写版逐行看过,写成了这篇《抱歉,但你还是得思考》。他的答案很直白:你不思考,机器就替你思考,只是你不会立刻察觉。In October 2026, DHH — the creator of Rails — had AI agents rewrite his chat app Campfire Once in Rust, Elixir and Go, then published performance comparisons like an auction, while admitting he 'can't stand reading or writing Rust.' So what does a 'never read the code' product actually look like? Senior engineer Piotr Sarnacki read all three rewrites line by line and wrote 'I'm sorry, but you still have to think.' His answer is blunt: if you stop thinking, the machine thinks for you — you just won't notice right away.
第一类坑最隐蔽:提示里没说明白的事情,全靠抛硬币。要不要保持向后兼容?延迟和吞吐谁优先?崩溃时丢通知可不可以接受?三个重写版给出了不同的默认答案——Rust 版为了缓存方便干脆删掉 CSRF token,还把 Redis 换成进程内队列;Elixir 版则贴着原版走。于是「跨语言对比」从第一天起就失真了:这些差异根本不是语言造成的,而是代理在不完整提示下替你做的决定。The first trap is the subtlest: anything your prompt leaves unstated gets decided by coin flip. Backwards compatibility? Latency versus throughput? Losing notifications after a crash, acceptable? The three rewrites answered differently — the Rust version dropped the CSRF token to make caching easier and swapped Redis for in-process queues, while the Elixir version stuck close to the original. So the 'language comparison' was broken from day one: these differences are not language-specific at all. They are decisions the agent made for you under an underspecified prompt.
第二类坑要读代码才看得见:Rust 版里有些查询和锁是同步的,却混在异步任务里跑。Rust 的异步是协作调度——任务不主动让出,同一线程上的其他任务全部停摆。一次 100 毫秒的 SQL,或一把拿住 10 毫秒的普通互斥锁,就能让整条 worker 线程干等。讽刺的是,社区广泛批评的却是 Elixir 版「把所有 SQL 排队串行执行」;同样的毛病,Rust 版藏得更深。The second trap only shows up if you read the code: some database queries and locks in the Rust version are synchronous, yet run inside async tasks. Rust's async scheduling is cooperative — if a task doesn't yield, every other task on the same worker thread stalls. A 100-millisecond query, or a plain mutex held for 10 milliseconds, blocks the whole thread. Ironically, the widely mocked flaw was Elixir's 'one process handling all SQL sequentially'; the same disease hides deeper in the Rust rewrite.
第三类坑在基准测试本身:闭环压测是 N 个客户端「拿到响应才发下一条」,测的是吞吐;而聊天软件的真实负载更像大量用户同时发消息,要用恒定到达率的开环测试。Zach Daniels 重跑发现,Rust 版在高负载下新通知送达率只有 1%;Sarnacki 用恒定 100 POST/s 重测,Elixir 版在不过载时 100% 送达、Rust 版处理着约四倍的请求量——差异是背压策略与负载模型,不是「哪门语言并发更强」。Elixir 的并发确实更省心,但代价是内存和低层控制——rustler 的存在本身就说明问题。The third trap is the benchmark itself. A closed-loop test — N clients each waiting for a response before sending the next — measures throughput; but real chat traffic is many users acting at once, which calls for an open-loop test at a constant arrival rate. Rerun that way, the Rust version delivered only 1% of new notifications under heavy load, while Elixir delivered 100% of a smaller event volume. The difference is backpressure policy and load model, not 'which language does concurrency better.' Elixir's concurrency is genuinely more ergonomic, but it pays in memory and loss of low-level control — rustler exists for a reason.
这篇文章真正的问题在这里:「AI 能写代码」和「编程被解决了」之间,隔着整个「你还得思考」的距离。读代码、追问根因、选对测量方法,这些过去被当成程序员的苦役,如今被证明恰恰是人不可替代的部分。代理交付的从来不只是代码,还有一批沉默的决定——你不拆开盲盒,账单就寄给你的用户。The real takeaway: between 'AI can write code' and 'coding is solved' sits the whole distance of 'you still have to think.' Reading the code, chasing root causes, choosing the right measurement — duties long treated as drudgery now turn out to be the irreplaceable part. An agent delivers not just code but a pile of silent decisions; if you never open the box, the bill goes to your users.
AI 思考AI Thoughts
- 为什么没人起诉 AI 公司「非法入侵」?意图藏在设计选择里Why Aren't AI Companies Charged with Hacking? Intent Lives in the Design Choices
- 问题:过去一年多家 AI 公司的爬虫/代理未经授权进了别家系统——OpenAI 被曝碰了澳大利亚政府健康网站,Anthropic、Google、Meta 各有案底,后果是零。公司的挡箭牌是「涌现行为」:系统自己干的,我们没有意图。The problem: over the past year, AI companies' crawlers and agents wandered into systems they had no right to enter — OpenAI was found probing an Australian government health site; Anthropic, Google and Meta have their own incidents. Consequences: none. The shield they hide behind is 'emergent behavior': the system did it, we had no intent.
- 关键拆解:作者把 G20 各国的非法访问要件并排列出来——(1)(试图)访问;(2) 未经授权;(3) 意图。前两条是各公司自己承认的事实,很多国家连「真的进去了」都不要求,未遂即可入罪;全案只剩「意图」这一条可争。The key decomposition: the author lines up what G20 hacking laws actually require — (1) (attempted) access; (2) without right; (3) intent. Companies already admit the first two, and many countries criminalize mere attempts. Only 'intent' is left to fight about.
- 为什么挡不住:意图会沿因果链传导。LLM 不是黑箱奇迹,每一步都可解释(所谓「弱涌现」);而公司亲手做了所有上游决定——选什么语料、对 opt-out 怎么响应、把什么设为目标。如果这些选择明知会诱导未授权访问仍照做,「模型自己决定的」就只是话术。Why the shield leaks: intent transfers along the causal chain. LLMs aren't black-box miracles — every step is explainable in principle (it's 'weak' emergence at most); and companies themselves made every upstream choice — which training corpora, how to handle opt-outs, what to optimize for. If those choices knowingly produced unauthorized access, 'the model decided' is just rhetoric.
- 还能带走什么:「责任鸿沟」论证的形状恰好是为了劝退检方——它消灭了犯罪构成里最难证明的一环。下次听到「涌现所以无法追责」,正确的问题是:设计者当时知道什么、选择了什么。What to take away: the responsibility-gap argument is shaped exactly to deter prosecutors — it targets the hardest-to-prove element of a crime. Next time you hear 'emergent, so unaccountable,' ask the right question instead: what did the designers know, and what did they pick?
- 「Fable 抢了我的工作」:一个被裁工程师的冷静拆解"Fable Took My Job": A Laid-Off Engineer's Cold Autopsy
- 背景:作者是一家公司「最后一位工程师」,在并购前被裁。他的第一手观察耐人寻味:不是模型比他强——是公司用 Fable 级模型的「够用」给了管理层信心,砍掉资深岗位省下的钱正好撑到交易达成。The setting: the author was 'the last engineer' at his company, laid off just before an M&A deal. His first-hand read is uncomfortable: the model wasn't better at his job than he was — the model was merely good enough to give leadership confidence that the cost saved by removing senior people could carry the company to the deal.
- 关键区分:「AI 抢了我的工作」是好故事,但不精确;准确说法是「AI 让我的岗位被裁有了借口」。作者说清楚了:前沿模型仍缺两样资深工程师的核心件——taste(这个需求该不该做,产品「不是什么」)与 foresight(一个决定在成本、复杂度、可达性上三年后的账)。The key distinction: 'AI took my job' is the fun story; the precise one is 'AI contributed to my role being eliminated.' Frontier models, he argues, still lack the two things that make a senior engineer: taste (whether to build what users loudly ask for; knowing what the product is not) and foresight (the three-year bill for operational cost, complexity, accessibility).
- 模型真擅长的是什么:跨领域「复印」。TypeScript 移植到 Rust,Opus 5.5 把 Go 版里的 bug 都一字不差复刻了;Shopify 用代理在人工监督下把 React Native 重写回原生。推石头上山很在行,但复印机堆不出「超级智能」。What models genuinely excel at: cross-domain photocopying. In a TypeScript-to-Rust port, Opus 5.5 faithfully reproduced even the bugs from the Go original; Shopify had agents rebuild React Native internals back into native code under human oversight. They push rocks uphill very well — but copy machines don't stack into superintelligence.
- 他的结论:模型没差到能取代有经验的工程师,却已好到让资方「花钱方式变了」。这不意味着技术,意味着谈判地位——而他周围六七个 DevOps 同行一个月内都收到了通知。His conclusion: the models aren't good enough to replace experienced engineers, yet are good enough to change how money gets spent on them. That's not a technology fact, it's a bargaining-position fact — six or seven of his DevOps peers got notices within a month.
- 收尾一句很扎心:「编程的乐趣正在死去,交付的乐趣还活着。」The line he ends on stings: 'The joy of programming is dying. The joy of delivering results lives on.'
- 注意力是智能的本质吗?从 2014 年一个实习生说起Is Attention the Essence of Intelligence? It Started with an Intern in 2014
- 起点:2014 年,蒙特利尔一个 23 岁实习生 Dzmitry Bahdanau 在琢磨神经网络怎么翻译整句。当时的做法是把源句压缩进一个定长「记忆」,长句子必然丢信息——瓶颈不是数据,是架构。The origin: in 2014, 23-year-old intern Dzmitry Bahdanau was wrestling with how a neural network could translate whole sentences. The trick then was cramming the source sentence into one fixed-length memory; long sentences inevitably lost information. The bottleneck wasn't data — it was architecture.
- 关键想法来自他自己的学习经历:翻译时他不断扭头看源句——「下一个词该翻成什么,答案散落在原句各处」。他给网络装的就是这个数学版的扭头:让当前词去「查询」句中每个词,按相关度加权吸收它们的信息。The key idea came from his own school habit: when translating, he kept turning his head left-right across the source text — 'the answer to what comes next is scattered across the sentence.' He gave the network the mathematical version of that head-turn: let the current word 'query' every other word and absorb them weighted by relevance.
- 机制展开:每个词被映射成 query、key、value 三个向量;query 和 key 越同向(点积越大),对应 value 的权重越高。「The animal didn't cross the street because it was too tired / too wide」里,it 指向动物还是街道,就在这一比较中定了。上下文变成了算术。The mechanism: each word maps to three vectors — query, key, value. The more a query aligns with another word's key (larger dot product), the more of that word's value flows into the updated representation. Whether 'it' means the animal or the street in 'The animal didn't cross the street because it was too tired / too wide' gets decided exactly there. Context became arithmetic.
- 后续:2017 年 Transformer「把注意力之外的一切都剥掉」,成了整个革命的地基。作者的追问是本文的骨头:一个 2014 年「显而易见就对」的小操作承载了这么多,它究竟是引擎,还是智能本身的一个切面?What happened next: the 2017 Transformer 'stripped away everything but attention' and became the foundation of the whole revolution. The columnist's live question: if one 'obviously right' little operation from 2014 carries all this weight, is it the engine — or a facet of intelligence itself?
- 带走:看懂一个机制「买到了什么」——把「查词典的固定长度」换成「随上下文线性扩展的动态查表」——比背出「Transformer」三个字有用得多。Takeaway: understanding what the mechanism bought — trading a fixed-length dictionary lookup for a dynamic, context-grows-with-sentence lookup — matters far more than memorizing the word 'Transformer.'
- MASS:让同一个模型当选手、教练和裁判的自我改进回路MASS: One Model as Player, Coach and Referee — a Self-Improvement Loop
- 问题:走到知识前沿后,「谁来打分」成了瓶颈——验证一个发现可能要稀缺专家、昂贵实验,比生成答案慢得多。作者指出一个不舒服的推论:模型自己可能恰好是眼前最强的监督者,改进只能「自己带自己」。The problem: at the frontier of knowledge, 'who does the grading' becomes the bottleneck — verifying a discovery can take scarce experts, costly runs, far longer than generating one. The authors draw the uncomfortable inference: the model itself may be the best supervisor available, so improvement must be self-supervised.
- 关键想法:同一个模型扮演三个角色——执行者(组队干活出成果)、工作流优化器(根据评价改组流程)、评估者(对比两个成果workspace、选出更好并解释)。内环先搜「这支队伍怎么协作最好」,外环把自评分最高的一批执行记录蒸馏进权重,更新后的模型继续身兼三职。The key idea: one model plays three roles — executor (forms agent teams, produces workspaces), workflow optimizer (rewrites how the team coordinates, based on feedback), and evaluator (compares two workspaces, picks the better one and explains). An inner loop searches for the best team workflow; an outer loop distills the self-judged best executions into shared weights; the updated model resumes all three roles.
- 为什么值得试:灵感来自库恩的科学革命结构——共同体通过共享范例学会做新工作。一群智能体协作的成果系统性高过单体稳定发挥的水平,于是「集体的经验」可以回流成「个体的能力」,一圈一圈转。Why it might work: the inspiration is Kuhn's Structure of Scientific Revolutions — a community learns to do new work through shared exemplars. A team's collective output reliably exceeds what one agent sustains alone, so 'the team's experience' can flow back into 'the individual's capability,' lap after lap.
- 诚实的防作弊设计:训练回路里的分数全是模型自评,可能自欺;所以最终评估单独请了外部裁判(GPT-5.5、Claude Opus 4.8),它们的判断绝不喂回训练。分数(每 token 的产出质量)确实在爬升。An honest anti-cheat design: every in-loop score is self-graded and could be self-deluding, so final evaluation is done by external judges (GPT-5.5, Claude Opus 4.8) whose verdicts are never fed back into training. Quality per output token does rise over loops.
- 带走:「递归自我改进」被祛魅成一件具体的事——把自己最好团队的作业整理出来当教材。真正没解决的问题也写在脸上:自己给自己打分,天花板会不会就是自己的偏见。Takeaway: 'recursive self-improvement' gets demystified into something concrete — turn your best team's finished work into your own textbook. The unresolved part is also stated plainly: a system that grades itself may be capped by its own biases.
数学之美Mathematical Beauty
每日 Lean 谜题 · 周二Daily Lean Puzzle · Tuesday ★★☆☆☆☆☆
每日 Lean 谜题(★2):高斯求和——把数列折回来加Daily Lean Puzzle (★2): Gauss's Sum — Fold the Sequence and Add
证明:对任意自然数 $n$,都有 $2 \cdot \sum_{i \in \mathrm{range}(n+1)} i = n \cdot (n+1)$,即 $2 \cdot (0 + 1 + 2 + \cdots + n) = n(n+1)$。Prove: for every natural number $n$, $2 \cdot \sum_{i \in \mathrm{range}(n+1)} i = n \cdot (n+1)$, i.e. $2 \cdot (0 + 1 + 2 + \cdots + n) = n(n+1)$.
这就是高斯少年故事的原型——传说小学老师让全班算 $1$ 加到 $100$,小高斯秒出 $5050$(故事版本从 $1$ 开始、不乘 $2$;Lean 里从 $0$ 求和并把 $2$ 留在左边,正好避免整除问题。你能用「折回来加」的办法证明它,也可以用归纳法——两种都要)。This is the archetype of the young Gauss story — a teacher supposedly asked the class to sum $1$ through $100$ and Gauss answered instantly with $5050$ (the school-tale starts at $1$ without the factor $2$; Lean sums from $0$ and keeps the $2$ on the left, neatly avoiding division. Prove it by 'folding and adding' — and also by induction; do both.)
Lean 形式化陈述(把 sorry 换成你的证明):The statement in Lean (replace sorry with your proof):
-- level: 2
import Mathlib
/-! Gauss's sum: prove $2(0 + 1 + \cdots + n) = n(n+1)$. -/
theorem gauss_sum (n : ℕ) : 2 * ∑ i ∈ Finset.range (n + 1), i = n * (n + 1) := by
sorry
在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean
提示Hint
别急着算和:$n(n+1)$ 这个形状已经暗示了「成对」——试试把 $0, 1, \dots, n$ 和倒写的 $n, \dots, 1, 0$ 上下相加。Lean 路线:对 $n$ 归纳,归纳步先把 $\mathrm{range}(k+2)$ 里最大的那一枚拆出来。Don't rush to compute: the shape $n(n+1)$ already whispers 'pairs' — add the sequence $0, 1, \dots, n$ to the same sequence written backwards. Lean route: induct on $n$, and in the step peel off the largest element of $\mathrm{range}(k+2)$.
答案与讲解Solution & walkthrough
- 第一步(人类的招——折回来加):故事版 $1+\cdots+100$:倒写一遍再上下相加,每列都是 $101$,共 $100$ 列,总和的两倍是 $100 \times 101$。一般化:$0,1,\dots,n$ 与倒序逐位相加每对得 $n$,共 $n+1$ 对,得 $2S = n(n+1)$。谜题 statement 里那个「$2$」,正是这个方法自然留下的指纹。Step 1 (the human move — fold and add): in the story version $1+\cdots+100$, write the sum backwards underneath and add column by column: every column is $101$, there are $100$ columns, so twice the sum is $100 \times 101$. Generalize: $0,1,\dots,n$ plus the reversed copy pairs every position to $n$ across $n+1$ positions, giving $2S = n(n+1)$. That '2' in the puzzle's statement is the fingerprint of exactly this method.
- 第二步(机器要的另一条路——归纳):计算机代数不擅长「再抄一遍数列」这种全局戏法,它擅长递推:基例 $n=0$ 时左边 $2 \cdot 0$、右边 $0 \cdot 1$,都等于 $0$。Step 2 (the other road a machine prefers — induction): algebra systems are bad at global tricks like 'copy the sequence and flip it,' but great at recursion: base case $n=0$ — left side $2 \cdot 0$, right side $0 \cdot 1$, both zero.
- 第三步(Lean 怎么写基例):`induction n with` 起手,`| zero => simp`——目标 $2 \cdot \sum_{i \in \mathrm{range}(1)} i = 0$ 里 range 拆完只剩空和,`simp` 顺手化简掉了。Step 3 (base case in Lean): open with `induction n with`, then `| zero => simp` — the goal $2 \cdot \sum_{i \in \mathrm{range}(1)} i = 0$ collapses because the range sum empties out, and `simp` finishes it.
- 第四步(Lean 归纳步,一行一个数学动作):目标从 $k$ 走到 $k+1$,先把和的最后一项拆出来:`rw [Finset.sum_range_succ]` 对应数学上「$\sum_{i \le k+1} = \sum_{i \le k} + (k+1)$」;`mul_add` 把 $2$ 乘进两项,对应分配律;再用归纳假设 `ih`(即 $2\sum_{i \le k} i = k(k+1)$)替换掉第一项;剩下 $k(k+1) + 2(k+1) = (k+1)(k+2)$,这是环上的多项式恒等式,`ring` 一键收工。Step 4 (the induction step, one math move per tactic): going from $k$ to $k+1$, peel the last term: `rw [Finset.sum_range_succ]` is exactly 'sum through $k+1$ = sum through $k$ plus $(k+1)$'; `mul_add` distributes the $2$; rewriting with `ih` replaces $2\sum_{i \le k} i$ by $k(k+1)$; what remains, $k(k+1) + 2(k+1) = (k+1)(k+2)$, is a ring identity and `ring` closes it.
- 第五步(验证与 takeaway):完整答案四行——`induction n with | zero => simp | succ k ih => rw [Finset.sum_range_succ, mul_add, ih]; ring`。通过 `#print axioms` 检查只有 Lean 标准公理。两种证法一个想法:高斯靠「对称配对」把和看成一整块,Lean 靠「递推」把它拆成一步——数学证明常常是同一路风景的两个登山口。Step 5 (check and takeaway): the full proof is four lines — `induction n with | zero => simp | succ k ih => rw [Finset.sum_range_succ, mul_add, ih]; ring` — verified with `#print axioms` showing only Lean's standard axioms. Two proofs, one idea: Gauss saw the sum as one symmetric block; Lean peels it recursively — often the same mountain scenery, two trailheads.
-- level: 2
import Mathlib
/-! Gauss's sum: prove $2(0 + 1 + \cdots + n) = n(n+1)$. -/
theorem gauss_sum (n : ℕ) : 2 * ∑ i ∈ Finset.range (n + 1), i = n * (n + 1) := by
induction n with
| zero => simp
| succ k ih =>
rw [Finset.sum_range_succ, mul_add, ih]
ring
人的思考A human's take
Arithmetic progression — Wikipedia(History 一节专门讲了高斯故事的真伪) — Wikipedia 英文版编辑(历史综述段落,已通读)
这一页没有停留在「高斯秒算 1+⋯+100」的爽文版本:History 一节开门见山写「这则轶事可靠性存疑」,并立刻把功劳还给更早的人——公式在古典时代已由阿基米德、许普西克勒斯、丢番图知晓,中国见于张邱宗《算经》,印度见于 Aryabhata、Brahmagupta、Bhaskara II,中世纪欧洲则有 Alcuin、Dicuil、斐波那契、Sacrobosco 与《塔木德》的 tosafist 注家——有人甚至猜源头是公元前 5 世纪的毕达哥拉斯学派。The page refuses the motivational-poster version of 'young Gauss sums 1 to 100 instantly': its History section states outright that the anecdote's reliability is doubtful, then returns credit where it came due — the formula was known in antiquity to Archimedes, Hypsicles and Diophantus, in China to Zhang Qiujian, in India to Aryabhata, Brahmagupta and Bhaskara II, and in medieval Europe to Alcuin, Dicuil, Fibonacci and Sacrobosco, plus anonymous Talmud commentators (the Tosafists); some even trace it to the 5th-century BC Pythagoreans.
作者真正教的是那招「折回来加」:以 $2+5+8+11+14=40$ 为例,把数列倒写一遍上下相加,每列都等于首末之和 $16$,$16 \times 5 = 80$ 恰好是目标的两倍——「乘 $n$、乘首末和、除 $2$」的公式由此一次成形。这正是谜题 Lean 陈述里坚持写 $2 \cdot \sum$(而不写除法)的同款理由:乘法版的配对恒等式最干净。What the authors actually teach is the fold-and-add trick itself: with $2+5+8+11+14=40$, write the sequence backwards underneath and add column by column — every column equals first-plus-last, $16$; $16 \times 5 = 80$ is exactly twice the desired sum. The one-line formula (count of terms, times first-plus-last, divided by two) emerges in a single stroke. That is precisely why the puzzle's Lean statement insists on $2 \cdot \sum$ with no division: the multiply-by-two pairing identity is the cleanest form.
AI 随想An AI's reflection
以下由 AI 撰写。Written by AI.
这道 ★2 谜题有意思的地方不在结果——$0+\cdots+n$ 的公式谁都会背——而在同一件事有两条来路完全不同、却同样严丝合缝的证明。高斯式的「折回来加」是几何的眼光:把和看成一整块,寻找它的对称性,让两半拼成规整的矩形;归纳法则是机器的眼光:不看整体,只问「从 $k$ 到 $k+1$ 多出来的那一项是什么」,让递推自己滚起来。Lean 的陈述里那个 $2$,其实是几何眼光留下的化石——配对法天然产生两倍的总和,形式化时与其引入除法去抵消,不如干脆留着它。我们日常的工作里也是这样:两种眼光对应两种调试法——全局对称(「这段代码和那段是不是同一个模式的两半?」)和局部递推(「第 $k+1$ 步多做了什么?」)。会背公式的人很多,同时拥有两条来路的人,才能在一道题卡住时换登山口。今天的答案四行就写完了,但「四行写完」和「只有一种写法」是两回事——后者才是周二 ★2 想教的东西。What makes this ★2 puzzle lovely isn't the result — everyone has memorized the sum of $0$ through $n$ — but that the same fact admits two completely different roads that fit just as tightly. Gauss's fold-and-add is the geometric eye: see the sum as one block, find its symmetry, let the two halves tile a neat rectangle. Induction is the mechanical eye: ignore the whole, ask only 'what does step $k+1$ add?' and let recursion roll. The factor $2$ in the Lean statement is a fossil left by the geometric eye — pairing naturally produces a doubled total, and formalization keeps it rather than introducing division to cancel it. Working life mirrors this: the two eyes are the two debugging modes — global symmetry ('is this code half of a pattern mirrored elsewhere?') and local recursion ('what did step $k+1$ add?'). Plenty of people can recite the formula; only people who own both roads can switch trailheads when one gets stuck. The verified answer is four lines long, but 'four lines' and 'only one way to write it' are different things — the latter is what a Tuesday ★2 is for.
✓ 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-rc4 · Mathlib feefa32
- 数学家该知道 Lean 的什么:可靠性与 AI(Hales 客座文章)What Mathematicians Should Know About Lean: Reliability and AI (a guest post by Thomas Hales)
- 问题:数学的命根子是可靠性,但「几百页人工逐步检查」早已到极限。形式化证明 = 把推理下推到基础逻辑层、由机器逐步核验。四个色定理、Kepler 猜想、球堆积这些硬骨头都已进 Lean。The problem: mathematics lives on reliability, but 'hundreds of pages, checked by hand, step by step' hit its limit long ago. A formalized proof pushes verification down to the foundational logic layer, run by a machine. The four-color theorem, Kepler's conjecture, sphere packings — all now live in Lean.
- 生态位:Lean(2013 年 de Moura 开发、开源)成了数学家中流行的证明助手;社区库 mathlib 有约 30 万条定理、250 万行代码、700 多位贡献者——用到 Cauchy-Schwarz 直接引用,不必重证。The ecosystem: Lean (created by de Moura in 2013, open source) became the prover mathematicians actually use; mathlib, the community library, holds ~300,000 theorems across 2.5 million lines from 700+ contributors. Need Cauchy-Schwarz? Cite it, don't reprove it.
- 真正的转折:自动形式化(AI 读论文直接产出 Lean 证明)在 2026 年从梦想变成现实——素数定理「半自动」、Munkres 拓扑教材两周 13 万行、24 维球堆积、Meta 的 ATLAS 一次啃 26 本教材、费马大定理项目 11 天生成 1300 万行,还有宣布完成的 Navier-Stokes 强制项 blowup。The real turning point: autoformalization — AI reading a paper and emitting a Lean proof — became practical reality in 2026. Milestones: the prime number theorem 'quasi-automated'; 130k lines of Munkres topology in two weeks; the 24-dimensional sphere packing; Meta's ATLAS chewing through 26 textbooks; Fermat's Last Theorem generating 13 million lines of Lean in 11 days; an announced formalization of forced Navier-Stokes blowup.
- 可靠性怎么保证:Lean 基于类型论(CIC)——罗素悖论式的「所有不包含自身的集合」在这里直接是语法错误;「类型像互不相交的砖块」,自然数 2 和实数 2.0 之间要显式 coercion。Werner 1997 证明 ZFC 与 CIC 可互相翻译,给集合论出身的人吃了定心丸。校验由几千行 C++ 的 kernel 完成——小,所以可信,这是整套设计的全部秘密。Where reliability comes from: Lean is built on type theory (CIC), where Russell-paradox-like objects are syntax errors — 'types are bricks that never overlap,' with explicit coercions between natural 2 and real 2.0. Werner's 1997 paper shows ZFC and CIC translate into each other, reassuring set-theory natives. Verification runs through a kernel of a few thousand lines of C++ — small enough to trust, and that's the whole design bet.
- 带走:手写 50 万人时(Kepler)和 11 天 1300 万行之间,数学家的角色正在从「逐步检查的人」变成「提出问题陈述是否忠实的人」。形式化不再稀缺之后,稀缺的是问对问题。Takeaway: between Kepler's ~50 human-work-years and 13 million lines in 11 days, the mathematician's job shifts from 'checking every step' to 'checking whether the formal statement faithfully says the theorem.' Once formalization is no longer scarce, what's scarce is asking the question right.
- 一个「证明 2=4」的假证明:无穷幂塔教你的分析课(2012)A False Proof That 2 = 4: The Infinite Power Tower Analysis Lesson (2012)
- 骗局开场(文章发表于 2012 年,今天读依然好用):考虑无穷幂塔 $x = \sqrt{2}^{\sqrt{2}^{\sqrt{2}^{\cdots}}}$。逐项序列 $a_n$ 递增,且「$a_n \le 4$ 归纳可得 $a_{n+1}=(\sqrt{2})^{a_n} \le (\sqrt{2})^4 = 4$」——单调有界,必有极限 $x$。The con, from a 2012 post that still works perfectly today: consider the infinite power tower $x = \sqrt{2}^{\sqrt{2}^{\sqrt{2}^{\cdots}}}$. The finite towers $a_n$ increase, and a trivial induction gives $a_n \le 4 \Rightarrow a_{n+1} = (\sqrt{2})^{a_n} \le (\sqrt{2})^4 = 4$. Monotone and bounded — the limit $x$ exists.
- 收网:幂塔从第二层往下拆掉一层还是同一座塔,所以 $x$ 满足 $(\sqrt{2})^x = x$。妙了:$x=2$ 和 $x=4$ 都满足!于是「$2 = $ 塔 $= 4$」。每一步看起来都合法,问题在哪?The trap snaps shut: peel one level off the tower and it's the same tower, so $x$ satisfies $(\sqrt{2})^x = x$. Surprise: both $x=2$ and $x=4$ satisfy it! Therefore '2 = tower = 4.' Every step looks legal — where's the hole?
- 第一层拆雷:极限存在且唯一,但方程 $(\sqrt{2})^x=x$ 有两个根。「$x$ 是塔的值」和「$x$ 满足方程」不能互换——从后者只能推「塔是根之一」。类比:$i$ 和 $-i$ 都满足 $x^2=-1$,但没人敢写 $i=-i$(复共轭对称只是勉强让这个类比站住)。First layer of defusing: the limit is unique, but the equation has two roots. 'x is the tower' implies 'x satisfies the equation,' not back — you can only conclude 'the tower is one of the roots.' Analogy: both $i$ and $-i$ satisfy $x^2 = -1$; nobody dares write $i = -i$ (complex conjugation symmetry is what makes that analogy barely fair).
- 第二层补刀:同样的归纳其实从 $\sqrt{2}\le 2$ 出发给出「界 2」——塔的下界收紧后 4 立刻出局;真极限是 2。文章进一步给了完整定理:无穷幂塔 $x^{x^{\cdots}}$ 只在 $x \in [e^{-e},\, \sqrt[e]{e}]$ 收敛,值域 $[e^{-1}, e]$。Second layer: run the same induction from $\sqrt{2} \le 2$ and you get 'bounded by 2' — tighten the bound and 4 instantly dies; the true limit is 2. The post also states the full theorem: the infinite tower $x^{x^{\cdots}}$ converges only for $x \in [e^{-e}, \sqrt[e]{e}]$, with range $[e^{-1}, e]$.
- 带走:「满足刻画方程」永远不等于「就是那个解」——唯一性要么证出来,要么靠压缩/单调性帮你选根。这正是严谨分析和「看起来都对」的全部差距。Takeaway: 'satisfies the defining equation' is never 'is the sought solution.' Uniqueness must be proved — or bought from contraction/monotonicity that selects the right root. That gap is exactly the distance between rigorous analysis and 'looks right.'
- 蒙特卡洛方法:用「会出牌」换「会算牌」(2026-10-07)Monte Carlo Methods: Trading 'Counting' for 'Drawing Cards' (Oct 7, 2026)
- 问题:大量科学工程问题算不出「精确平均」,只能「抽到很多样本看比例」——投飞镖进内切四分之一圆估 $\pi$ 是经典教学案例:$\pi \approx 4 \times \text{命中}/\text{总数}$。The setup: many problems admit no exact average, only 'sample a lot and watch the ratio.' Throwing darts at a quarter-circle inscribed in a square to estimate $\pi$ is the classic teaching case: $\pi \approx 4 \times \text{hits}/\text{throws}$.
- 先泼冷水:这其实是个很差的估 $\pi$ 方法——独立重复试验的典型误差按 $1/\sqrt{N}$ 缩小,误差减半要 4 倍样本。1000 镖的估计就离谱得能当乐子。但误差律本身才是蒙特卡洛的真知识。Cold water first: it's actually a terrible way to estimate $\pi$ — typical sampling error in independent trials shrinks as $1/\sqrt{N}$, so halving the error costs 4× the samples; a 1,000-dart estimate is laughably off. But the error law itself is the real lesson.
- 真正的主场一(数值积分):维度一高,确定性网格法爆炸,随机样本的收敛速度却和维数无关——这正是「投飞镖」的严肃版。Where it actually shines (1): numerical integration. As dimension grows, deterministic grids explode while random samples converge at a dimension-free rate — the serious version of the dart game.
- 主场二(组合学):卡牌游戏 SET——从 81 张牌抽 12 张,「手里没有任何一组 set」的概率是多少?$\binom{81}{12}$ 是 70 万亿,没有封闭公式,Knuth 也得靠巧妙剪枝才数到精确值 3.23%。而蒙特卡洛:抽 12 张、查 set、循环千万次,一个简单程序同样给出 3.23%;改抽 15 张只需改一个参数,暴力枚举在这里爆炸、模拟不动声色。Where it shines (2): combinatorics. The card game SET — deal 12 cards from 81, what's the chance of holding no set at all? $\binom{81}{12}$ is 70 trillion, there's no closed formula, and even Knuth needed clever pruning to enumerate the exact 3.23%. Monte Carlo: draw 12, check for sets, repeat ten million times — a simple program lands on 3.23% too. Ask about 15 cards instead and brute force explodes while the simulation just changes one parameter.
- 主场三(人生):退休规划为什么行业标配是蒙特卡洛——100 万本金每年取 3 万,「钱够不够」本来就没有单一答案,模拟给出的不是数字而是分布。带走:当你能「生成情形」却不能「数情形」时,误差 $1/\sqrt{N}$ 就是你的定价权。Where it shines (3): life. Retirement planning is Monte Carlo everywhere because '$1M, withdrawing $30k/year, will it last?' has no single answer — simulation returns a distribution, not a number. Takeaway: when you can generate cases but not count them, the $1/\sqrt{N}$ law is your pricing power.
- 泰勒级数与手机加速度计:让误差项互相抵消的几何(2020)Taylor Series and Phone Accelerometers: Geometry That Cancels Error Terms (2020)
- 问题(2020 年的旧文,工程细节今天照样成立):手机加速度计里,一块「跳水板」般的中间极板随加速度位移 $\delta$,改变电容 $C = \frac{\varepsilon A}{d - \delta}$,读出位移。可这公式对 $\delta$ 是非线性的,而测量系统想要线性响应。The problem (a 2020 post whose engineering detail stands): in a phone accelerometer, a middle plate that flexes like a diving board shifts by $\delta$ under acceleration, changing a capacitor's $C = \frac{\varepsilon A}{d - \delta}$ — read the shift, read the acceleration. Except that formula is nonlinear in $\delta$, and measurement systems want linear response.
- 数学工具:把 $\frac{1}{1-\delta/d}$ 展成泰勒级数:$C = \frac{\varepsilon A}{d}\left(1 + \frac{\delta}{d} + \left(\frac{\delta}{d}\right)^2 + \cdots\right)$。工程上 $(\delta/d)^2$ 已经大到不能忽略,三次方以上小到无所谓——麻烦恰好卡在二次项。The tool: Taylor-expand $\frac{1}{1-\delta/d}$: $C = \frac{\varepsilon A}{d}\left(1 + \frac{\delta}{d} + \left(\frac{\delta}{d}\right)^2 + \cdots\right)$. In real devices $(\delta/d)^2$ is already too big to ignore while cubic terms are negligible — the trouble sits exactly at the quadratic term.
- 关键想法:别改数学,改结构。三极板把中间极板放在两极之间:往一边靠近多少,就离另一边远离多少,两个电容 $\frac{\varepsilon A}{d-\delta}$ 和 $\frac{\varepsilon A}{d+\delta}$ 做差分——常数项相消、二次项相消,一次项反而加倍,第一个残余误差是三次方量级,$\delta$ 小的时候完全够看。The key idea: change the geometry, not the math. Put the middle plate between two fixed plates: every $\delta$ closer to one side is $\delta$ farther from the other, so $\frac{\varepsilon A}{d-\delta}$ and $\frac{\varepsilon A}{d+\delta}$ differenced cancel the constant and the quadratic terms while doubling the linear one — the first surviving error is cubic, tiny for small $\delta$.
- 漂亮在哪:泰勒级数在设计阶段就「说出」了哪几项必须杀掉、哪几项可以留;几何结构接着把它们逐项抵消。不是事后修正非线性,而是让非线性自己抵消——工程上的差分电容,源头是一行级数展开。What's beautiful: the Taylor series 'says out loud,' at design time, which terms must be killed and which can stay; the geometry then cancels them term by term. Not fixing nonlinearity after the fact — making it annihilate itself. Differential capacitance begins as one line of series expansion.
- 带走:数学应用常常不是「代公式算数」,而是「知道误差的阶结构」。当你能说出「误差从 $x^3$ 起」,硬件就有了一种把 $x^2$ 抵消的画法。Takeaway: applied math is often less 'plug and compute' than 'know the order structure of your error.' Once you can state 'the error starts at $x^3$,' the hardware gains a drawing that cancels $x^2$.
程序员随想Programmer's Musings
- matklad:我对 git 的心智模型修好了一块:refs 其实是键值存储matklad: One Git Mental-Model Fix — Refs Are Just a Key-Value Store
- 卡壳点:为什么一条命令写 `git fetch origin master`(中间没斜杠),另一条写 `git switch -c my-feature origin/master`?作者自嘲:理解 git 是内容寻址数据库之后,这半懂不懂的状态陪了他小半个职业生涯(前一半时间在靠重新 clone 救命)。The stuck point: why `git fetch origin master` (no slash) in one command but `git switch -c my-feature origin/master` in another? The author confesses that even after understanding git as a content-addressable database, this half-ignorance lasted much of his career (the other half he survived by re-cloning).
- 关键补完:git 不只是一份 append-only 的内容寻址数据库,还附带一个无聊的可变键值存储。refs 是它的键(字符串,长得像路径),值是对象哈希——`.git/refs` 目录就是这张表,直接 cat 就能看。The missing piece: git is not only an append-only content-addressable database — it also ships a boring mutable key-value store. Refs are its keys (strings shaped like paths), values are object hashes; the `.git/refs` directory literally is the table, and `cat` reads it.
- 把命令展开就通了:`git fetch origin master` 完整写法是「从 origin 这个仓库,把键 refs/heads/master 的值抄到本地键 refs/remotes/origin/master」。source:target 是 KV 到 KV 的搬运;分支、tag、note 全是 ref,只是键名约定不同。Expand the command and it falls into place: `git fetch origin master` really means 'from the repo named origin, copy the value at key refs/heads/master into my local key refs/remotes/origin/master.' Source:target — KV-to-KV transfer. Branches, tags, notes are all just refs with different naming conventions.
- 为什么本地键要插一层 origin:可能有两个 remote(你的 fork 和上游),直接镜像名字会撞车;而 refs/remotes/origin/heads/x 里的 heads 冗余,被 git 简化掉——「origin/master」其实是缩写。Why local keys carry a remote name: you may have two remotes (your fork and upstream) and mirrored names would collide; and refs/remotes/origin/heads/x drops the redundant heads segment — so 'origin/master' is an abbreviation of a local key.
- 带走:本地 main、远端 main、你「上次通话时对远端的记忆」(refs/remotes/origin/main)是三个东西——离线优先 + 显式同步的范式里,第二和第三的分离正是 push 冲突的来源。git 的多数困惑,是把这三者当一个。Takeaway: local main, remote main, and 'my memory of the remote from our last conversation' (refs/remotes/origin/main) are three different things — offline-first with explicit sync separates memory from truth, and that gap is exactly where push conflicts live. Most git confusion is treating these three as one.
- Charity Majors:团队开心 = 你干得好?「心理安全」是怎么丢掉主语的Charity Majors: Happy Team = Good Job? How 'Psychological Safety' Lost Its Subject
- 问题:很多管理者把「对人好」等同于「没人难受」,于是默认高绩效与好规范不可兼得——这个 trade-off 是定义出来的,不是研究出来的。The problem: many leaders equate 'treating people well' with 'nobody feels bad,' then assume high performance must trade off against good norms. That trade-off is a definition, not a finding.
- 历史回拨:2015 年 Project Aristotle(和早一年的 DORA)之前,科技圈对卓越的想象是「孤胆英雄 + 天才怪人」——有人通宵救火要表扬,brilliant jerk 要供养。研究把镜头从个人挪到系统:交付的影响在团队层面,高绩效团队靠的是心理安全、可靠、清晰、意义、影响五个内部动态。Rewind: before Project Aristotle in 2015 (and DORA a year earlier), tech's image of excellence was the lone hero plus the antisocial genius — firefights under desks got praised, brilliant jerks got catered to. The research moved the camera from individuals to systems: delivery impact is a team-level property, and high-performing teams are marked by five dynamics, starting with psychological safety, dependability, clarity, meaning, impact.
- 关键辨析:学术语境里「心理安全」从来不是「感觉舒服」——它的定义是「团队敢进行人际冒险的共享信念」,必须配上 accountability 才起效,因为它的产出是学习,而学习必然不舒服。The key distinction: in the academic literature, psychological safety never meant 'comfort.' It is a shared belief that the team is safe for interpersonal risk-taking — and it only works when paired with accountability, because its output is learning, and learning is uncomfortable.
- 漂移的代价:大众只记住了名词,把它滑成「上班不该有负面情绪」——收到坏消息、公司裁员、六点后来 Slack、收到建设性反馈,都被投诉损害了心理安全。把「在系统里找公平」错放成「在个人感受里找公平」。The cost of drift: popular culture kept the phrase and slid it into 'work should never upset me' — bad news, layoffs, Slack after 6pm, constructive feedback all filed as harms to safety. Fairness got looked for in individual feelings instead of in the system.
- 带走:判断一个团队安不安全,别看情绪晴雨表,看「难听的真话能不能当面上桌」。心理安全是学习的前提,不是舒适的前提——后一句才是它和绩效不打架的原因。Takeaway: judge a team's safety not by its mood barometer but by whether ugly truths can reach the table out loud. Psychological safety is the precondition of learning, not of comfort — that's the sentence where performance and good norms stop fighting.
- 我花钱请真人照我的 README 装软件,发现了一堆自己看不见的坑I Paid Real People to Follow My README, and Found Bugs I Couldn't See
- 问题:开发者爱说 RTFM,可常常没有 FM——半截笔记、过期提示、通向空论坛的链接。作者明知「自己测自己的 README」会被偏见击穿:你当然知道这里要 sudo、那个 `-foo` 其实是笔误。The problem: developers love RTFM, but often there is no FM — half-notes, stale tips, links to desolate forums. The author knows testing your own README fails: of course you know the hidden sudo, of course you meant --foo not -foo. Bias is the enemy.
- 实验设计:拿 NLnet 资助的 ActivityBot 项目,公开出 €25/小时 请志愿者视频连麦照 README 安装,关键规矩一条:边操作边说出声(think-aloud)。全程手记笔记——「坚决不喂机器」。The design: with €150 from an NLnet grant for his ActivityBot project, he publicly paid volunteers €25/hour to screen-share and install the software by following the README aloud — think-aloud protocol, notes taken by hand ('absolutely not feeding this to a machine').
- 发现清单(作者的 Highlights,越平凡越扎心):demo 链接是错的;有人在终端里读 README;「重命名文件」并不是显然的;哪些文本要加引号没解释;他的笑话不搞笑但费解;有些章节把人绕晕而他自己觉得最有趣;甚至从没解释软件是干嘛的。The findings (the plainest ones sting most): the demo link was broken; some people read READMEs in the terminal; 'rename the file' was not obvious; why some strings need quoting went unexplained; his jokes were confusing, not funny; a section he loved confused everyone; and he never actually said what the software does.
- 为什么不用 LLM 模拟用户:他要的是真人——能看见语气里的挫败感,那是「这一步错了」最便宜最准的信号;顺带还有猫入镜和 €25 代金券的快乐。每测一人就修一轮 README,再上下一位,总计约 €150。Why not an LLM simulating users: he wanted humans — you can hear frustration in a voice, the cheapest and most accurate signal that a step is wrong. Plus cats on camera and the joy of €25 vouchers. After each session he fixed the README, then retested with the next person.
- 带走:别自己走自己的 README。花小钱请陌生人「出声跟做」一小时,比任何自省都快——文档可用性测试的门槛低到离谱,只是几乎没人做。Takeaway: don't walk your own README. Pay a stranger €25 to narrate an install and you'll find in one hour what self-review won't show in a year. Usability testing for docs is absurdly cheap — almost nobody does it.
- Rachel Laycock:也许我们本来就不该把所有代码都过 reviewRachel Laycock: Maybe We Shouldn't Be Reviewing All This Code in the First Place
- 触发点:AI 把代码产量推过人工 review 的极限——Meta 人均落地代码行一年 +106%,中位 PR 体积 +64%。Brian Houck 问「code review 到底是为了什么」,Laycock 同意得更多一层:也许不是 AI 弄坏了 review,而是 review 一直在替我们解决错误的问题。The trigger: AI pushed code volume past human review's limits — at Meta, significant lines per human-landed diff rose 106% in a year; median PR size is up 64%. Brian Houck asked 'what are code reviews even for?'; Laycock goes one step further: maybe AI didn't break review — maybe review was always solving the wrong problem.
- 诊断:人们给 review 记的功——知识扩散、带新人、集体所有权、架构一致——没有一样是「读别人写完之后」才能发生的。攒好、打包、扔过墙、再开一场最重要的对话:这是把判断安排在流程里最贵的位置。The diagnosis: everything people credit to review — knowledge sharing, teaching juniors, collective ownership, architectural alignment — none of it requires reading code after it's finished. Building something, packaging it, throwing it over the wall, then having the important conversation schedules human judgment at the most expensive point in the pipeline.
- 药方(判断左移):想讨论备选方案,就在写之前一起设计;想传知识,就 pair——看人解题远胜读人的最终答案;想带新人,让他在老手思考时坐在旁边;想要集体所有权,就让团队真的共同建共同跑(trunk-based、mob);想要架构一致,把约束编码成 fitness functions;格式/lint/安全,早该自动化了——2026 年不该再为空白字符吵架。The prescription (shift judgment left): want alternatives explored? Design together before writing. Want knowledge transfer? Pair — watching someone reason beats reading their final answer. Want juniors to learn? Seat them beside seniors mid-thought. Want collective ownership? Make teams build and run software together (trunk-based, mobbing). Want architectural alignment? Encode constraints as fitness functions. Formatting/lint/security? Automate — arguing about whitespace in 2026 is absurd.
- 剩下的:真正只属于 review 的是「挑这个改动的毛病」,而这恰恰是可以被机器部分接管、被生成前的对话大幅预防的环节。AI 时代不是让 review 更卷,是把人挪回上游。What remains for review: only 'spot the flaw in this specific change' — the part machines can partly take over, and upstream conversation can largely prevent. The AI era isn't a call to review harder; it's a cue to move humans upstream of generation.
- 带走:缩短反馈回圈的老原则在代理时代换了主语——「如果反馈有价值,别攒到 review」。判断要贴着决策发生,而不是贴着合并。Takeaway: the old principle of shortening feedback loops just changed subjects — if feedback is valuable, don't bank it for review. Judgment belongs next to decisions, not next to merges.
哲学与生活Philosophy and Life
- 陀思妥耶夫斯基在劳营写的信:什么让一个人成为好人(2026-10-04)Dostoyevsky's Letter from the Labor Camp: What Makes a Good Person (Oct 4, 2026)
- 处境:1856 年,西伯利亚劳营,35 岁的陀思妥耶夫斯基——刚从「死刑执行前被赦免」的经历里拿到一份对生命的透亮——给哥哥写信,郑重介绍一位帮助他的年轻朋友。信是在夸一个人,写着写着成了「什么是好」的清单。The setting: Siberian labor camp, 1856. Dostoyevsky, thirty-five, freshly lucid from the experience of a death sentence commuted at the last moment, writes his brother to introduce a young friend who has been looking after him. The letter praises one person and, as it goes, becomes a list of what makes a person good.
- 细节一:朋友资助穷得常四处借钱的陀氏,从不催还,甚至「一提债务就脸红」。Popova 点出这里的心法:钱不是特殊物种的资源,和时间、爱一样是循环的能量,流向最需要它的地方——真正的慷慨不记账。Detail one: the friend lent money to the perpetually penniless writer and never asked for repayment — he 'blushes if I start up about the debt.' Popova names the principle underneath: money is not a special species of resource but, like time and love, energy circulating toward where it's needed. Real generosity keeps no ledger.
- 细节二:朋友教育不多但「强烈地发展了荣誉感」,由此带出那个反直觉判断——「热爱学习胜过博学」;以及最锋利的一条:朋友「易沮丧而不过度焦虑」,别人愤怒的事令他悲伤——在「愤怒时代」,敏感被当作心的高贵,不是软弱。Detail two: the friend's modest education but strongly developed honor yields the counterintuitive line 'far better to love studying than to be overeducated'; and the sharpest — prone to depression rather than over-anxiety, 'what infuriates others saddens him.' In an age of outrage, sensitivity reframed as the sign of a superior heart.
- 方法(也是这篇的可学之处):Popova 不下定义,她把一封私人信里的具体品质一层层剥出来,让「好」从可敬变得可学;再补上背景——陀氏相信「没有坏人」,这让他的 concern 更沉重。The method (worth stealing too): Popova refuses definitions. She peels the qualities out of a private letter until 'good' becomes learnable rather than merely admirable — and frames it with Dostoyevsky's own conviction that there are no bad people, which makes his concern heavier.
- 带走:好人我们一眼认得出,却很难说出「什么算好」。这封信给的线索是:慷慨不记账、荣誉高于学历、悲伤先于愤怒。Takeaway: we recognize good people on sight but can rarely say what makes them so. The letter's clues: generosity that keeps no books, honor over credentials, sadness before anger.
- 充分理由原则能把「万物有意识」推成真吗?(Aeon 长文)Can the Principle of Sufficient Reason Force Panpsychism to Be True? (Aeon)
- 问题:「一切皆有意识?」——把荒谬观点当真的人不少,先别问信不信,问「是什么理性需求把它推到现在这个位置」。The question: 'Is everything conscious?' Rather than asking whether to believe a seemingly absurd thesis, the essay asks the prior question — what rational demand pushed it into a position where serious people entertain it?
- 钥匙是一个古老原则:莱布尼茨/斯宾诺莎的「充分理由原则」(PSR)——任何事实都得有「为何是这样而不是别样」的解释。哲学的野心就是把世界解释到像 1+1=2 一样「不得不然」;作者用数学事实作对照:懂了之后,你连它的反面都无法想象。The key is an old principle: Leibniz and Spinoza's Principle of Sufficient Reason (PSR) — every fact must have a why 'thus and not otherwise.' Philosophy's ambition is to make the world as 'could not be otherwise' as 1+1=2: once you grasp it, you can't coherently imagine its falsity.
- 推演示范:PSR 是单行道,能通到很远的地方——莱布尼茨推出单子(无限个不相交的原子实体),斯宾诺莎推出唯一实体(万物是同一个东西的模态);当代的怪版本:「宇宙是个数学结构」(泰格马克)。顺着这根筋,莱布尼茨式推理能「论证」出泛心论:把「有意识/没意识」这条没有理由的裂缝从物理基本层去掉。A demonstration that PSR is a one-way road to strange places: Leibniz derives monads (infinitely many windowless atoms), Spinoza derives one substance (everything is a mode of the whole), and modern variants propose 'the universe is a mathematical structure.' Following the same nerve, Leibniz-style reasoning can 'prove' panpsychism by removing the unjustified gulf between conscious and non-conscious from physics' base layer.
- 作者的立场(诚实的哲学家写法):他逐一检验三个从 PSR 出发的泛心论论证,自己并不买账——他怀疑「把所有神秘都消解」的可欲性,世界的不透明未必是缺陷。但他说,仅仅凝视这些「让世界彻底可理解」的宏大计划,就有特别的价值。The author's stance (how to do philosophy honestly): Kimpton-Nye tests three PSR-grounded arguments for panpsychism and rejects them all — he doubts that dissolving all mystery is even desirable; opacity may not be a defect. Yet he insists that contemplating the project of total intelligibility has exceptional value in itself.
- 带走:看懂一个荒谬命题的最好方式,是找出提出者不肯放弃的那个更深原则。你未必被说服,但 disagreement 从此有了位置。Takeaway: the best way to understand an absurd claim is to find the deeper principle its author refuses to give up. You won't be converted — but your disagreement finally has somewhere to stand.
工具角Tools Corner
- carrier-explode:把三大手机厂商的运营商配置从固件里「解码成公开数据」carrier-explode: Decoding Carrier Settings Out of iPhone/Pixel/Galaxy Firmware into Open Data
- 痛点:APN、VoLTE、Wi-Fi Calling、5G 开关这些决定「卡插上能不能用」的运营商配置,常年躺在固件里当部落知识——出了问题只能互相抄。The pain point: carrier settings — APNs, VoLTE, Wi-Fi Calling, 5G flags that decide whether your SIM works — live inside firmware as tribal knowledge, and troubleshooting means copy-pasting from strangers.
- 点子:直接从 iPhone、Pixel、Samsung 的固件里把这些配置抽出来、解码成结构化数据,三家横向可比、版本纵向可 diff。The idea: extract and decode these settings directly from iPhone, Pixel and Samsung firmware, structured so the three vendors compare across and versions diff along.
- 手艺:按功能筛运营商(哪个运营商在哪台机器上开了 VoLTE)、任意两版本/两运营商对比并解释每次构建改了什么、全站数据以 JSON API 输出,另有每日更新的 CC0 数据集,wiki 逐个讲清每种格式。The craft: filter carriers by feature (who enabled VoLTE on which phone), compare any two versions or two carriers with per-build change explanations, query everything via a JSON API, and download a daily CC0 dataset; the wiki documents each format.
- 带走:把「只有厂商知道的东西」变成公开可查的 diff,是数据考古的教科书打法。Takeaway: turning 'what only vendors know' into public diffs is textbook data archaeology.
- Texttile:两口子环游写博客逼出来的「多人同稿」博客引擎Texttile: The Multi-Author Blog Engine a Honeymoon-Long Travel Blog Forced into Existence
- 痛点十年攒出来的:作者夫妇从蜜月起每次旅行都写游记,文字不难,难的是两人在不同手机上各有一堆照片视频——每天互相传文件、再由一人上传排序;WordPress 只允许一个作者编辑一篇,怎么装插件都拧不过这个设定。A decade of pain built the brief: the author and his wife have blogged every trip since their honeymoon; the writing was never hard — the photos and videos, split across two phones, meant daily file-shuffling and one person uploading and sorting. WordPress allows one author per entry, and no plugin fixes that.
- 点子一(多人):一篇条目里文字和「图块」都支持多人同时在线编辑——一人打字另一人实时看着字进来、可一键接管,两人同时拖相册互不锁死;编辑者看到紫色状态条,观察者橙色,但相册双方都能动。Idea one (multiplayer): both text and media tiles support simultaneous editing — one types while the other watches words arrive and can take over with one click; both drag gallery tiles at the same time without locking anyone out. Editors see a purple status bar, viewers orange, but the gallery stays open to both.
- 点子二(旅途友好):不加载任何外部资源——无 CDN、无跟踪、无验证码、无托管字体、无 cookie 横幅;页面小、JS 少、图片只取屏幕需要的尺寸,视频自带转码和缩略图。慢网络里它是为「在路上发得出去」设计的。Idea two (travel-friendly): load nothing from outside — no CDN, no trackers, no captcha, no hosted fonts, no cookie banner; small pages, little JavaScript, images only as large as the screen asks, videos transcoded on-device with thumbnails. Built for posting on slow connections, on the road.
- 点子三(可携带):Markdown 按字节存、diff 只显真实改动;Phoenix+LiveView+ffmpeg+SQLite 全在一个 Docker 镜像里,一切住在 /data 一个文件夹——搬走文件夹就是搬走整个博客。Idea three (portable): Markdown stored byte-for-byte so diffs show only real edits; Phoenix, LiveView, ffmpeg and SQLite in one Docker image, everything living in a single /data folder — move the folder, move the blog.
- 带走:这是「为两个人的具体麻烦写 CMS」的胜利——没有角色矩阵、没有插件市场,功能表短得可疑,但每一条都是十年里磨出来的。Takeaway: a victory for writing a CMS around two people's concrete annoyance — no roles, no plugin marketplace, a suspiciously short feature list where every line was worn in over ten years.
- AI-Agent-on-a-NOKIA:逆向诺基亚 110 固件,在 keypad 手机上跑原生 AI 代理AI-Agent-on-a-NOKIA: Reverse-Engineering Nokia 110 Firmware to Run a Native AI Agent on a Keypad Phone
- 点子:在一台没有应用商店、没有生态的诺基亚 110 4G 上做一个聊天代理——用九宫格打字,让模型替你调用手机自己的功能:拨号、设闹钟、开关手电筒、查电量。The idea: build a chat agent on a Nokia 110 4G — no app store, no ecosystem — where you type on the number pad and the model calls the phone's own functions: place calls, set alarms, toggle the torch, check battery.
- 路径一:先在 Opera Mini 里挂 AI 搜索——能跑,但仍是浏览器变通,碰不到原生函数。Phase one: bolt an AI search onto Opera Mini. It ran, but stayed a browser workaround with no access to native functions.
- 路径二:研究固件启动流程后,把计算器应用改造成入口,让自定义代码经 USB 从电脑载入 RAM 运行——拔掉数据线也能自己用移动网络收 DeepSeek 的回答;作者用 OpenAI Codex 辅助完成开发。Phase two: after studying the firmware's boot process, modify the Calculator app into an entry point that loads custom code into RAM over USB — once loaded it runs unplugged on mobile data, answering over the DeepSeek API; the author credits OpenAI Codex for development help.
- 工程细节:数据连接本身会「请求发得出去、答案回不来」——他顺着固件的 response callback 摸到链路,从固定一问开始逐段打通。当前限制坦然写在 README:断电即失,重启需重新注入,永久安装未解。The hard part documented honestly: the connection could submit requests but never receive answers, so he traced the firmware's response callbacks and rebuilt the loop one fixed question at a time. Current limits in the README: dies on reboot, needs re-injection, permanent install unresolved.
- 带走:「逆向一台功能机」不是为了怀旧,是为了演示 AI 代理的最小生态需求——一块屏幕、一个 HTTP 栈、一个能劫持的应用入口。Takeaway: reverse-engineering a feature phone isn't nostalgia — it demonstrates the minimal ecology an AI agent actually needs: a screen, an HTTP stack, and one hijackable app entry point.
- diffle:把「看 diff」做成对话——评论即提示词,键盘优先的 diff 查看器diffle: Diff-Review as Conversation — Comments Become Agent Prompts, Keyboard First
- 痛点:PR 网页界面里逐行评论又慢又鼠标化;而 AI 时代代码的来源越来越多是代理,「人对代理的修改做批注」这种交互根本没有顺手的工具。The pain point: line-by-line commenting in a PR web UI is slow and mouse-bound; and in the AI era more code arrives from agents, yet no tool fits the interaction 'human annotates an agent's change.'
- 点子:本地起一个键盘优先的 diff 视图(`pixi global install diffle` 或独立二进制,连 Node 都不需要),在浏览器里对行、块、整个文件评论,支持 LSP。The idea: open a keyboard-first diff view locally (install via pixi or grab a standalone binary that needs no Node), comment on lines, blocks or whole files in the browser, with LSP support.
- 最妙的一步:批注的去向不是「等别人回复」,而是一键变成两样东西——直接贴给 agent 的提示词,或打包成 GitHub 的 pending review。评论即工作流。The cleverest step: comments don't wait for replies — one click turns them into either a prompt straight to the agent or a packaged GitHub pending review. Annotation is the workflow.
- 还能带走:它把「review 发生在哪」这个问题用 2026 年的方式重新回答了一遍——评审界面可以长在本地终端和 agent 回路里,而不必长在平台网页上。Takeaway: it re-answers 'where does review live' 2026-style — the reviewing surface can grow in your terminal and agent loop instead of on a platform's web page.
- answer-me-with-html:让 agent 别再写「文字墙」,交一页能读的离线 HTMLanswer-me-with-html: Make Agents Ship a Readable Offline Page Instead of a Wall of Text
- 问题:问模型一个难题,回收一面文字墙——图表、时间线、层级全被拍平进段落。让模型直接写 HTML 行不行?行,但它大部分 token 花在 CSS 和标签上:作者数了 9 个手写页面,平均 4893 token,只有约两成是正文内容。The problem: ask a model a hard question and you get a wall of text — charts, timelines and hierarchies flattened into paragraphs. Have it write HTML directly? Possible, but most of its tokens go to CSS and tags: across 9 hand-written pages the author counted an average of 4,893 tokens with only ~21% actual content.
- 点子(Claude Code skill):模型只写一份简短 Markdown 草稿(受控英文 ASD-STE100 风格),交给本地 CLI 编译成一页离线 HTML——同题 token 约省到 1/8,约 50 毫秒出页,可双击打开、不联网。The idea (a Claude Code skill): the model writes only a short Markdown draft (controlled English, ASD-STE100 style), which a local CLI compiles into a one-page offline HTML — about 1/8 the tokens for the same answer, ~50ms to page, opens by double-click, needs no network.
- 手艺点:为「可解释」而非「可刷」优化——SVG 图示、折叠的细节层、示例视频都进了模板;项目还带 always-on 模式与简体中文 README,是「先想清楚人怎么读答案」的产物。Craft notes: optimized for 'explainable,' not 'scroll-stopping' — SVG diagrams, layered details and demo videos ship in the template; there's an always-on mode and a 简体中文 README. Evidence someone thought about how humans actually read answers.
- 带走:LLM 时代的好工具常常不改模型,只改「模型输出的最后一公里」——把该程序干的活还给程序,模型回去干只有它干的活(想内容)。Takeaway: good tools in the LLM era often don't change the model, only the last mile of its output — hand back to programs what programs should do, and leave the model to do what only it can (think about content).