每日科技文摘:语言的功能与关系,以及一道关于「分解不唯一」的老题Daily Tech Digest: Functional vs. Relational Language, and an Old Problem About Non-Unique Factorization
发布于 2026-10-18Published 2026-10-18
今日深读Today's Deep Read
对 AI 味文字暴怒的人,和自己「只看结果」的信条打了一架The Slop Snob Fights Her Own 'Outcomes Only' Creed
Honeycomb 的 CTO Charity Majors 写过一篇自白,起点是一家公司里同时在气两边的人。一半人对收到的「slop」暴怒:要五条要点,收到二十页注水的文档;问同事一个问题,回复以「Claude 说……」开头。另一半人也委屈:他们产出更快、结果更好,凭什么还要被骂「工作流不更新」?Majors 自己在不同时段站过两边。但真正让她睡不着的是另一个词——越来越多同事说,工作中的人味没了,dehumanizing。Charity Majors, CTO of Honeycomb, opens with a company mad in both directions. Half the staff is furious about slop: they ask for five bullet points and get a twenty-page padded doc; they ask a colleague a question and the reply starts with "Claude says...". The other half is equally aggrieved: they ship faster with better outcomes, and resent being scolded for not updating their workflows. Majors has stood in both camps at different times. What kept her up at night was a different word — coworkers kept saying the human connection at work was gone. Dehumanizing.
她的第一反应是标准的工程师直觉:工具就是工具,结果说明一切。客户在乎的是产品质量,不是你怎么做的。这套「 ruthless 地只看结果」的说辞她差点说服了自己——直到她注意到自己的行为完全矛盾。嘴上说着「只看质量」,私下里她对 AI 味的文字已经发展出一种暴怒式的厌恶:newsletter 一读出 AI 腔就退订,博客直接关标签页,社交媒体直接取关。她甚至说,如果有个能给作者一万伏电击的按钮,她会毫不犹豫地按下去,还当面按。Her first instinct was the standard engineer's move: tools are tools, outcomes are all that matter, customers care about quality not provenance. She nearly convinced herself — until she noticed her own behavior contradicting it. While publicly repeating "it doesn't matter how it was made", she had privately developed a violent disgust reflex for AI-flavored text: newsletters unsubscribed the moment she sensed AI-isms, blog posts closed mid-scroll, authors unfollowed. She jokes she'd slam a 10,000-volt button on slop authors and not care who saw.
很多洞察都是这么开始的:嘴上坚称 A,行为却一直像 B。逼自己调和这两件事,她被引到了对语言本身的拆解上。语言为我们做的每一件事,要么是功能性的,要么是关系性的。功能性场合里,价值在想法本身,谁说的无关紧要:代码 diff、结构化数据、形式化证明,甚至大部分技术文档——只要准确,AI 介入完全正当,没人会眨眼。而关系性场合里,价值恰恰在于「是这个特定的人说了、想了、感觉了」:问同事的意见、上级写的绩效评语、一条慰问。Her best insights start this way: loudly insisting A while behaving as if B. Forcing the two to reconcile led her to take language apart. Every job language does is either functional or relational. In functional contexts the value lives in the idea itself, and who said it is irrelevant: code diffs, structured data, formal proofs, most technical docs — as long as it's accurate, AI involvement is fine and nobody blinks. In relational contexts the value is precisely that this particular person said, thought, or felt it: asking a colleague's opinion, a manager's performance review, a condolence.
这个分法立刻解释了愤怒从哪来。绩效评语一看就是 ChatGPT 写的,员工会问:他到底读没读我做了什么?他要是不写,他在这儿干嘛?想要的是「你怎么看」,收回来的却是一段 Claude 片段——这不算偷懒,算背叛,因为你点名要的就是那个人的脑子。而且这个效应是单向放大的:交流越个人化,AI 代笔破坏的信任越多,哪怕文字本身无可挑剔。The split explains the rage immediately. A performance review that reads like ChatGPT makes the employee ask: did my manager even read what I did? You ask "what do YOU think" and get back a Claude snippet — that's not laziness, it's a violation, because what you asked for was that specific brain. And the effect compounds: the more personal the interaction, the more AI ghostwriting destroys trust, even when the words themselves are flawless.
所以她的判据很简单,也带得走:下次纠结「这里用 AI 合不合适」,别问用了多少、披露了几成——先问这段交流的价值在想法上,还是在这个人身上。价值在想法上,尽管放开用, ruthlessly 地只看结果;价值在人身上,把笔拿回来。这也顺便解释了为什么市面上那些「AI 使用比例披露」的框架总觉得没搔到痒处:关键从来不是比例,是场合。至于她自己那股暴怒——那不是不理性,是她的注意力系统在替她做一次快速的场合判断:这段话里没有一个 specific 的人,不值得给注意力。So her test is portable: next time you wonder whether AI belongs in a given message, don't ask how much AI was used or how thoroughly disclosed — ask whether the value of this exchange lives in the idea or in the person. Idea? Go wild, be ruthlessly outcome-oriented. Person? Pick the pen back up. It also explains why all those AI-disclosure-percentage frameworks felt off: the variable that matters was never the percentage, it was the occasion. As for her own rage — it isn't irrationality, it's her attention system making a fast context call: no specific human in sight, no attention owed.
AI 思考AI Thoughts
- 一天合并三万条代码之后,Anthropic 把版本控制重新设计了一遍30,000 Merges a Day Later, Anthropic Redesigned Version Control Around Agents
- 问题有多真实:2026 年 5 月他们平均每天合并约 2500 条代码变更,9 月已超过 30000 条。主分支历史近百 GB,每天 2500 万次 CI checkout,按 git 的方式取全量副本相当于每天 600 PB 流量——瓶颈不再是写代码,而是为人设计的开发基建。The problem is real: in May 2026 Anthropic merged ~2,500 changes per day; by September, over 30,000. Main-branch history is nearly 100 GB, and 25 million CI checkouts per day would mean ~600 PB of daily git traffic if fetched naively. The bottleneck is no longer writing code — it's infrastructure built for human scale.
- 关键观察:git 的核心假设是「客户端离线也能工作」,所以每个克隆都是完整仓库。可他们的客户端几乎全是 agent 和 CI 任务,天生在线。让 git 无往不利的分布式设计,在这个负载下恰恰成了负资产。The key observation: git assumes clients work offline, so every clone is a full repository. But almost all of Anthropic's clients are agents and CI jobs, online by definition. The distributed design that made git win became a liability at this load.
- 于是 Scion 把原则反过来:客户端永远在线,提交、合并、rebase 全部发生在服务器端,本地只留一个有界缓存加稀疏工作副本;blame、路径历史这类「派生数据」(概念承自 Meta Sapling 的 Mononoke)用同一个递归机制从 changeset 统一推导,想要新的查询就加一种派生数据,而不是各自造工具。So Scion inverts the principles: clients are always online; commits, merges and rebases happen server-side; the client keeps only a bounded cache plus a sparse working copy. Blame and path history are "derived data" (a concept owed to Mononoke from Meta's Sapling), computed by one generic recursive mechanism — want a new kind of query, add one kind of derived data instead of a bespoke tool.
- 代价也要看清楚:7 个人、5 个月完成整体迁移;没有离线支持,工程师在 HN 上答得很直白——Scion 最重的用户是 Claude,而 Claude 永远在线,离线根本不重要。The costs, honestly: a team of seven migrated everything in five months, and there is no offline support. An engineer answered on Hacker News plainly: Claude is by far the heaviest user of Scion and is necessarily online, so offline never mattered.
- 一个管理层的旁注:对抗式评审(让模型互相挑毛病)极其有效,但会短视——资深工程师得不断把 Claude 从「为假设性的担忧过度工程化」里拽回来。AI 既制造了这个局面,也是他们能五个月建成新系统的原因。A management aside: adversarial review (models attacking each other's work) is incredibly effective but myopic — senior engineers must keep pulling Claude back from over-engineering against hypothetical concerns. AI created this situation, and is also why a new system could be built in five months.
- AI 能力的形状:不是一条斜率,是锯齿、瓶颈和显著点(2025 年)The Shape of AI: Not a Slope but Jaggedness, Bottlenecks and Salients (2025)
- 问题:几乎所有人都在给 AI 打一个「整体分」,可实际用它的人都知道,模型同时会做医学诊断级的推理、却搞不定一台自动售货机的退货流程。能力不是一条平滑斜率,是锯齿。The problem: everyone scores AI with one overall number, but anyone using it daily knows models do medical-diagnosis-level reasoning while failing to return a vending machine purchase. Capability isn't a smooth slope; it's jagged.
- 关键想法一(锯齿):同一个模型在不同任务上横跨超人与不如小学生的巨大区间,所以「AI 能不能做 X」这个问题本身问错了粒度——X 必须拆到具体步骤。Idea one (jaggedness): the same model spans superhuman to worse-than-a-schoolkid across tasks, so "can AI do X" is asked at the wrong granularity — X must be broken into concrete steps.
- 关键想法二(瓶颈):预测 AI 什么时候能真正替你完成一件事,别看它最强的那一环,看整条工作流里最窄的那一环。写作被搞定了,但「读你 messy 的内部资料」没搞定,整件事就还卡着。Idea two (bottlenecks): to predict when AI can actually take over a job for you, don't look at its strongest link; look at the narrowest link in the whole workflow. Writing got solved; "reading your messy internal notes" didn't, so the whole task stays stuck.
- 关键想法三(反向显著点):Hughes 提出的概念——地图上被卡住的那一小块洼地,一旦被新技术填平,整个系统不是线性变好,而是整体跳跃。后来图像生成上发生的跳跃就是这类「被打破的显著点」。Idea three (salients): Hughes' concept — the small pocket where the system is stuck; once new tech fills that pocket, the whole system jumps rather than improves linearly. The recent jumps in image generation are exactly such broken salients.
- 带走:评估「AI 能不能吃掉我这个任务」时,先画出全流程,找到瓶颈在哪一环、那一环有没有正被显著点化。这比任何跑分都更接近实际答案。Takeaway: to judge whether AI can eat your task, draw the whole workflow first, find the bottleneck, and check whether that link is about to become a broken salient. That beats any benchmark score as a prediction.
- 软件博客的反模式:为什么 AI 代笔让博客越长越像Anti-Patterns in Software Blogging: Why AI Ghostwriting Makes Blogs Look Alike
- 方法本身就是想法:软件工程里靠「反模式」识别注定失败的做法,他把同一套工具搬到写作上——不列「好博客守则」,而是列出他当读者时反复踩到的坑。The method is itself the idea: software engineering identifies doomed practices via anti-patterns, so he applies the same tool to writing — not a list of rules for good blogs, but the mistakes he keeps hitting as a reader.
- 头号反模式是绕圈开场。读者给你的预算只有标题加前三句话,他要在里面回答两个问题:这文章是写给我这样的人的吗?我能得到什么?背景故事可以写,但那是在花读者「继续读」的预算。Anti-pattern #1 is the meandering intro. The reader's budget is the title plus three sentences, within which two questions must be answered: was this written for someone like me, and what do I get? Backstory is fine — it just spends the keep-reading budget.
- 第二个反模式他叫「读者知道我知道的一切」:用 cgroups 和 BSD jails 解释 Docker 是什么。作者的世界观越窄,越容易把行话当常识。解法朴素但有效:想象一个真实的同事,列出他认识的词和不认识的词。Anti-pattern #2: "the reader knows everything I know except this one thing" — explaining Docker via cgroups and BSD jails. The narrower your world, the more you mistake jargon for common knowledge. The fix is plain but works: picture a real coworker and list which words they know.
- 第三个反模式最切当下:过度依赖链接。术语不就地解释、一句「见某文」就把理解成本推给读者——而在 AI 能瞬间生成通顺文字的年代,这种「把组装留给读者」的写法恰恰是读者最先抛弃的。Anti-pattern #3 is the most current: overreliance on links — pushing the cost of understanding onto the reader with "see my other post". In an era when AI generates fluent text instantly, this assemble-it-yourself writing is exactly what readers drop first.
- 把整篇读完再看 AI 的位置就很清楚:博客的价值在于一个具体的人把话组织到「就像当面讲给你听」,AI 代笔恰好把这一层磨平了——所有文章都通顺,于是所有文章都一样。Read the whole piece and AI's position becomes clear: a blog's value is one specific person arranging words as if talking to you face to face. AI ghostwriting sands that layer off — every article becomes fluent, and therefore every article becomes the same.
数学之美Beauty of Mathematics
每日 Lean 谜题 · 周日Daily Lean Puzzle · Sunday ★★★★★★★
同一个数,两种「不可再分」的分解(IMO 1977 第 3 题)One Number, Two 'Indecomposable' Factorizations (IMO 1977, Problem 3)
设 $n > 2$。考察所有形如 $1 + kn$($k$ 为正整数)的整数构成的集合 $V_n$。称 $V_n$ 中的一个数 $m$ 为「不可分解」的,如果不存在 $V_n$ 中的两个数 $u, v$ 使得 $m = uv$。证明:存在一个数,它能以多于一种方式表示为 $V_n$ 中若干个不可分解数的乘积(只有因子顺序不同的表示视为同一种)。Let $n > 2$, and consider the set $V_n$ of all integers of the form $1 + kn$ for positive integers $k$. A number $m \in V_n$ is called indecomposable if there are no two numbers $u, v \in V_n$ with $m = uv$. Prove that some number can be expressed as a product of indecomposable elements of $V_n$ in more than one way (expressions differing only in the order of factors count as the same).
Lean 形式化陈述(把 sorry 换成你的证明):The statement in Lean (replace sorry with your proof):
/-
Copyright (c) 2026 The Compfiles Contributors. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Kimi K3
-/
import Mathlib.Tactic
/-!
# International Mathematical Olympiad 1977, Problem 3
Given an integer n > 2, let Vₙ be the set of integers 1 + kn for k a positive integer.
A number m in Vₙ is called indecomposable if it cannot be expressed as the product of
two members of Vₙ. Prove that there is a number in Vₙ which can be expressed as the
product of indecomposable members of Vₙ in more than one way (decompositions which
differ solely in the order of factors are not regarded as different).
-/
namespace Imo1977P3
/-- The set `Vₙ` of natural numbers of the form `1 + k * n` with `k` a positive integer. -/
def Vn (n : ℕ) : Set ℕ := {m | ∃ k : ℕ, 1 ≤ k ∧ m = 1 + k * n}
/-- `m` is indecomposable in `Vₙ` if it belongs to `Vₙ` but cannot be written as a
product of two members of `Vₙ`. -/
def Indecomposable (n : ℕ) (m : ℕ) : Prop :=
m ∈ Vn n ∧ ¬ ∃ p q : ℕ, p ∈ Vn n ∧ q ∈ Vn n ∧ m = p * q
theorem imo1977_p3 (n : ℕ) (hn : 2 < n) :
∃ r ∈ Vn n, ∃ F₁ F₂ : Multiset ℕ,
(∀ d ∈ F₁, Indecomposable n d) ∧ (∀ d ∈ F₂, Indecomposable n d) ∧
F₁.prod = r ∧ F₂.prod = r ∧ F₁ ≠ F₂ := by
sorry
end Imo1977P3
在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean
提示Hint
先别急着造那个数。第一步是证明 $V_n$ 里每个成员都能分解成不可分解数之积(强归纳);第二步是发现一个「小得反常」的不可分解数:$(n-1)^2$。注意 $V_n$ 里最小的成员是 $n+1$,所以任何两个成员的乘积至少是 $(n+1)^2$。Don't rush to build the number. First show every member of $V_n$ factors into indecomposables (strong induction); then find an abnormally small indecomposable: $(n-1)^2$. The smallest member of $V_n$ is $n+1$, so any product of two members is at least $(n+1)^2$.
答案与讲解Solution & walkthrough
- 乘法封闭:$1 + an$ 乘 $1 + bn$ 等于 $1 + (a + b + abn)\,n$,仍在 $V_n$ 里。所以「分解」总是在集合内部进行,这保证了问题有意义(Lean 里是 mul_mem_vn 这个引理,两行代数)。Multiplication stays inside: $(1 + an)(1 + bn) = 1 + (a + b + abn)\,n$, so products of members of $V_n$ are members again — the question is well-posed. (In Lean this is the two-line lemma mul_mem_vn.)
- 关键引理——存在「小得反常」的不可分解数:$(n-1)^2 = 1 + (n-2)n$ 属于 $V_n$,但 $V_n$ 的任何成员都至少是 $n+1$(因为 $k \ge 1$),所以两个成员的乘积至少是 $(n+1)^2$;而 $(n-1)^2 < (n+1)^2$,它怎么也写不成两个成员的乘积——不可分解。(Lean: indecomposable_sq,核心是 vn_ge 给出的下界。)Key lemma — an abnormally small indecomposable: $(n-1)^2 = 1 + (n-2)n$ belongs to $V_n$, but every member of $V_n$ is at least $n+1$ (since $k \ge 1$), so any product of two members is at least $(n+1)^2$. Since $(n-1)^2 < (n+1)^2$, it can never be such a product — indecomposable. (Lean: indecomposable_sq, powered by the lower bound in vn_ge.)
- 存在性:对 $m$ 做强归纳——若 $m$ 不可分解,它自己就是一组合法分解;否则 $m = pq$,$p, q$ 都在 $V_n$ 里且都严格小于 $m$,各自分解再拼起来。(Lean: exists_factorization,用 Nat.strong_induction_on。)注意到这里只用了「存在」,还完全没碰「唯一」。Existence: strong induction on $m$ — if $m$ is indecomposable, the singleton factorization works; otherwise $m = pq$ with $p, q \in V_n$ both strictly smaller than $m$, and we concatenate their factorizations. (Lean: exists_factorization via Nat.strong_induction_on.) Note this only ever uses existence, never uniqueness.
- 造 witnesses:取 $r = (n-1)^2 (2n-1)^2$。第一组因子:$(n-1)^2$ 乘上 $(2n-1)^2$ 的某个不可分解分解;第二组因子:$(n-1)(2n-1)$ 的不可分解分解复制两份——因为 $((n-1)(2n-1))^2 = (n-1)^2(2n-1)^2$,乘积同为 $r$。(Lean: 主定理里的 G 与 H + H。)拿 $n=3$ 验证:$100 = 4 \times 25 = 10 \times 10$,三组因子 $4, 25, 10$ 全都模 3 余 1,且全都不可分解。Build the witnesses: take $r = (n-1)^2 (2n-1)^2$. First factorization: $(n-1)^2$ times some indecomposable factorization of $(2n-1)^2$; second: take an indecomposable factorization of $(n-1)(2n-1)$ twice — since $((n-1)(2n-1))^2 = (n-1)^2(2n-1)^2$, both products equal $r$. (Lean: G and H + H in the main theorem.) Sanity check with $n=3$: $100 = 4 \times 25 = 10 \times 10$, where $4, 25, 10$ are all $\equiv 1 \bmod 3$ and all indecomposable.
- 证明两组不同:$(n-1)^2$ 出现在第一组。若它也出现在第二组,注意第二组每个因子的乘积是 $(n-1)(2n-1)$,所以每个因子都整除它;但 $n-1$ 不整除 $2n-1 = 2(n-1)+1$,于是 $(n-1)^2$ 不整除 $(n-1)(2n-1)$,矛盾。(Lean: not_dvd 加 Multiset.dvd_prod,把「因子乘积」翻译成「逐个整除」。)The two differ: $(n-1)^2$ occurs in the first factorization. If it occurred in the second, it would divide that factorization's product $(n-1)(2n-1)$ — but $n-1$ does not divide $2n-1 = 2(n-1)+1$, so $(n-1)^2$ cannot divide it. Contradiction. (Lean: not_dvd plus Multiset.dvd_prod, which turns "product of the factors" into "each factor divides".)
/-
Copyright (c) 2026 The Compfiles Contributors. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Kimi K3
-/
import Mathlib.Tactic
/-!
# International Mathematical Olympiad 1977, Problem 3
Given an integer n > 2, let Vₙ be the set of integers 1 + kn for k a positive integer.
A number m in Vₙ is called indecomposable if it cannot be expressed as the product of
two members of Vₙ. Prove that there is a number in Vₙ which can be expressed as the
product of indecomposable members of Vₙ in more than one way (decompositions which
differ solely in the order of factors are not regarded as different).
-/
namespace Imo1977P3
/-- The set `Vₙ` of natural numbers of the form `1 + k * n` with `k` a positive integer. -/
def Vn (n : ℕ) : Set ℕ := {m | ∃ k : ℕ, 1 ≤ k ∧ m = 1 + k * n}
/-- `m` is indecomposable in `Vₙ` if it belongs to `Vₙ` but cannot be written as a
product of two members of `Vₙ`. -/
def Indecomposable (n : ℕ) (m : ℕ) : Prop :=
m ∈ Vn n ∧ ¬ ∃ p q : ℕ, p ∈ Vn n ∧ q ∈ Vn n ∧ m = p * q
/-- Every member of `Vₙ` is at least `n + 1`. -/
lemma vn_ge {n m : ℕ} (hm : m ∈ Vn n) : n + 1 ≤ m := by
obtain ⟨k, hk, rfl⟩ := hm
have h := Nat.mul_le_mul hk (le_refl n)
lia
/-- `Vₙ` is closed under multiplication. -/
lemma mul_mem_vn {n p q : ℕ} (hp : p ∈ Vn n) (hq : q ∈ Vn n) : p * q ∈ Vn n := by
obtain ⟨a, ha, rfl⟩ := hp
obtain ⟨b, hb, rfl⟩ := hq
exact ⟨a + b + a * b * n, by lia, by ring⟩
/-- `(n - 1)^2` (written with `n = c + 3`) is indecomposable in `Vₙ`: it is smaller
than the square of the smallest member of `Vₙ`. -/
lemma indecomposable_sq (c : ℕ) : Indecomposable (c + 3) ((c + 2) ^ 2) := by
refine ⟨⟨c + 1, by lia, by ring⟩, ?_⟩
rintro ⟨p, q, hp, hq, h⟩
have hp1 : c + 4 ≤ p := by have := vn_ge hp; lia
have hq1 : c + 4 ≤ q := by have := vn_ge hq; lia
have hle : (c + 4) * (c + 4) ≤ (c + 2) ^ 2 := by
have := Nat.mul_le_mul hp1 hq1
rwa [← h] at this
have e : (c + 4) * (c + 4) = (c + 2) ^ 2 + (4 * c + 12) := by ring
lia
/-- `(n - 1)^2` does not divide `(n - 1) * (2 * n - 1)` (written with `n = c + 3`),
since `n - 1` does not divide `2 * n - 1 = 2 * (n - 1) + 1`. -/
lemma not_dvd (c : ℕ) : ¬ (c + 2) ^ 2 ∣ (c + 2) * (2 * c + 5) := by
intro h
have h2 : c + 2 ∣ 2 * c + 5 := by
have h' := h
rw [sq] at h'
exact Nat.dvd_of_mul_dvd_mul_left (by lia) h'
have h3 : c + 2 ∣ 2 * c + 5 - 2 * (c + 2) := Nat.dvd_sub h2 (dvd_mul_left _ _)
have h4 : 2 * c + 5 - 2 * (c + 2) = 1 := by lia
rw [h4] at h3
have := Nat.le_of_dvd one_pos h3
lia
/-- Every member of `Vₙ` is a product of indecomposable members of `Vₙ`. -/
lemma exists_factorization (n : ℕ) (hn : 2 < n) :
∀ m : ℕ, m ∈ Vn n →
∃ F : Multiset ℕ, (∀ d ∈ F, Indecomposable n d) ∧ F.prod = m := by
intro m
induction' m using Nat.strong_induction_on with m ih
intro hm
by_cases h : Indecomposable n m
· exact ⟨{m}, fun d hd => by
rw [Multiset.mem_singleton] at hd
rw [hd]
exact h, Multiset.prod_singleton m⟩
· obtain ⟨p, q, hp, hq, rfl⟩ : ∃ p q : ℕ, p ∈ Vn n ∧ q ∈ Vn n ∧ m = p * q := by
by_contra hcon
exact h ⟨hm, hcon⟩
have hpg : n + 1 ≤ p := vn_ge hp
have hqg : n + 1 ≤ q := vn_ge hq
have hplt : p < p * q := by
have h2 : p * 2 ≤ p * q := Nat.mul_le_mul (le_refl p) (by lia)
lia
have hqlt : q < p * q := by
have h2 : 2 * q ≤ p * q := Nat.mul_le_mul (by lia) (le_refl q)
lia
obtain ⟨Fp, hFp, hFpprod⟩ := ih p hplt hp
obtain ⟨Fq, hFq, hFqprod⟩ := ih q hqlt hq
exact ⟨Fp + Fq, fun d hd => (Multiset.mem_add.mp hd).elim (hFp d) (hFq d),
by rw [Multiset.prod_add, hFpprod, hFqprod]⟩
theorem imo1977_p3 (n : ℕ) (hn : 2 < n) :
∃ r ∈ Vn n, ∃ F₁ F₂ : Multiset ℕ,
(∀ d ∈ F₁, Indecomposable n d) ∧ (∀ d ∈ F₂, Indecomposable n d) ∧
F₁.prod = r ∧ F₂.prod = r ∧ F₁ ≠ F₂ := by
obtain ⟨c, rfl⟩ : ∃ c : ℕ, n = c + 3 := ⟨n - 3, by lia⟩
-- The witness is `r = (n - 1)^2 * (2 * n - 1)^2`, written with `c = n - 3`
-- (so `c + 2 = n - 1` and `2 * c + 5 = 2 * n - 1`) to avoid truncated subtraction.
have hs2mem : (c + 2) ^ 2 ∈ Vn (c + 3) := ⟨c + 1, by lia, by ring⟩
have ht2mem : (2 * c + 5) ^ 2 ∈ Vn (c + 3) := ⟨4 * c + 8, by lia, by ring⟩
have hstmem : (c + 2) * (2 * c + 5) ∈ Vn (c + 3) := ⟨2 * c + 3, by lia, by ring⟩
-- First factorization: `(n-1)^2` times any factorization of `(2n-1)^2`.
obtain ⟨G, hG, hGprod⟩ := exists_factorization (c + 3) (by lia) _ ht2mem
-- Second factorization: any factorization of `(n-1)*(2n-1)`, taken twice.
obtain ⟨H, hH, hHprod⟩ := exists_factorization (c + 3) (by lia) _ hstmem
refine ⟨(c + 2) ^ 2 * (2 * c + 5) ^ 2, mul_mem_vn hs2mem ht2mem, (c + 2) ^ 2 ::ₘ G,
H + H, ?_, ?_, ?_, ?_, ?_⟩
· intro d hd
rw [Multiset.mem_cons] at hd
rcases hd with rfl | hd
· exact indecomposable_sq c
· exact hG d hd
· intro d hd
rw [Multiset.mem_add] at hd
rcases hd with hd | hd <;> exact hH d hd
· rw [Multiset.prod_cons, hGprod]
· rw [Multiset.prod_add, hHprod]; ring
· -- The factorizations differ: `(n-1)^2` occurs in the first one, but every factor
-- of the second one divides `(n-1)*(2n-1)`, which `(n-1)^2` does not divide.
intro h
have hmem : (c + 2) ^ 2 ∈ (c + 2) ^ 2 ::ₘ G := Multiset.mem_cons_self _ _
rw [h, Multiset.mem_add] at hmem
rcases hmem with hmem | hmem <;> {
have hdvd := Multiset.dvd_prod hmem
rw [hHprod] at hdvd
exact not_dvd c hdvd
}
end Imo1977P3
人的思考A human's take
1977 IMO Problems/Problem 3 — AoPS Wiki 社区
AoPS Wiki 上这题的页面(原站拦截抓取,本文经 Wayback 快照读取):上面有 1977 年的原题陈述,与我们的 Lean 形式化逐句对应。页面上社区贴的解法走了另一条路:先用「与 n 互素的素数有无穷多个」这条欧几里得式引理加抽屉原理,找到两个落在同一剩余类里的素数,再直接造出两组不可分解的因子。同一个结论,一个解法靠「大小」造不可分解,另一个靠「同余类」造不可分解——对照着看很有味道。The AoPS Wiki page for this problem (the site blocks fetchers; read here via a Wayback snapshot) carries the original 1977 statement, matching our Lean formalization line by line. The community solution posted there takes a different route: a Euclid-style lemma that infinitely many primes are coprime to n, a pigeonhole step to find two primes in the same residue class, and then a direct construction of two indecomposable factorizations. Same conclusion, one proof built on size, the other on congruence classes — worth reading side by side.
AI 随想An AI's reflection
以下由 AI 撰写。Written by AI.
这道题最迷人的地方在于它把「分解存在」和「分解唯一」彻底撕开了。我们在自然数里长大,脑子里深深装着算术基本定理:任何合数都能分解成素数之积,而且只有一种分法。于是很容易以为「每个数都能分解到不能再分」和「分法只有一种」是一回事——前者听起来几乎是后者的显然推论。这道题构造了一个小世界:只保留模 n 余 1 的数,乘法封闭,甚至每个数都能分解成「不可再分」的因子,唯一性却轰然倒塌,$100 = 4 \times 25 = 10 \times 10$ 摆在那里,四个因子个个「不可再分」。更妙的是那个反例的造法:不可分解性不是逐个检查出来的,而是靠一道不等式 $(n-1)^2 < (n+1)^2$ 整批地制造出来——你不需要知道一个数具体长什么样,只需要知道它「太小了,小到不可能是任何两个成员的乘积」。用下界批量制造不可分解,再用「整除」排除它出现在另一组分解里,这是整道题真正的两板斧。最后留一个现实回声:Lean 证明里最重的部分其实是那个「存在性」引理——强归纳写出每个成员都可分解。 existence 便宜,uniqueness 昂贵,而且昂贵的那一半在某些世界里干脆是假的。The charm of this problem is how cleanly it tears 'factorizations exist' apart from 'factorizations are unique'. We grow up with the fundamental theorem of arithmetic, and it is tempting to read 'everything factors down to indecomposables' as an obvious corollary of uniqueness. This problem builds a tiny world — keep only the numbers congruent to 1 mod n — where multiplication is closed, everything factors into indecomposables, and yet uniqueness collapses: $100 = 4 \times 25 = 10 \times 10$, every factor stubbornly indecomposable. Even better is how the counterexample is manufactured: indecomposability is not checked case by case, it is mass-produced by one inequality, $(n-1)^2 < (n+1)^2$ — you never need to know what the number looks like, only that it is too small to be any product of two members. That is the real one-two punch: a lower bound to stamp out indecomposables in bulk, and a divisibility argument to keep that indecomposable out of the second factorization. One last echo into practice: the heaviest part of the Lean proof is the existence lemma, the strong induction showing every member factors at all. Existence is cheap; uniqueness is expensive — and in some worlds, the expensive half is simply false.
✓ 已通过 Lean 编译验证(无 sorry,无额外公理)✓ Verified by Lean (no sorry, no extra axioms) · 题目与形式化证明来自 compfiles(Apache-2.0 (compfiles, David Renshaw and contributors))Problem and formal proof from compfiles (Apache-2.0 (compfiles, David Renshaw and contributors)) · Lean v4.35.0-rc4 · Mathlib feefa32
- 把全部数学形式化:七位一线形式化者的泼冷水与路线图How to Formalize All of Mathematics: Seven Practitioners Pour Cold Water and Draw a Map
- 背景:AI 短期内接连官宣形式化重大定理——Anthropic 的 Claude 用 11 天、1300 万行 Lean 验证了费马大定理,OpenAI 宣布解出 Navier–Stokes 有限时间爆破并附 Lean 证明。于是有提案:趁热打铁,自动形式化全部数学文献,给后代留遗产。Backdrop: within weeks, Anthropic's Claude verified Fermat's Last Theorem in 13 million lines of Lean over 11 days, and OpenAI announced a finite-time-blowup solution to Navier–Stokes with a Lean proof. A proposal follows naturally: formalize the entire mathematical literature now, as a legacy.
- 他们的反驳从一个尖锐观察开始:形式化产物如果没人用,就几乎一文不值。Viazovska 球堆积高维形式化被 AI 公司「劫持」抢跑,轰动一时,但原团队后来发现清理 AI 代码比重写更费劲,干脆扔掉重来;那 1300 万行 FLT 代码也一样——定理本身并不在被怀疑,正确性之外没有留下什么。Their rebuttal starts from a sharp observation: a formal artifact nobody uses is worth almost nothing. When an AI lab hijacked the Viazovska sphere-packing formalization, the original team later found cleaning up the AI code cost more than rewriting it — they threw it away; the 13M-line FLT proof is the same: the theorem was never in doubt, and beyond correctness little remains.
- 所以真正的框架是一个三角:速度、规模、理解。抢速度和规模时被牺牲掉的,恰恰是形式化最有价值的部分——重构定义、探索推广、让人和 AI 能在库里继续做数学。OpenAI 的 Navier–Stokes 是反例也是混血儿:正确性信号这次真的关键(自然语言证明几乎读不通),但社区依然愤怒——论证来历、署名和「只鼓励产出、不鼓励理解」。So the real frame is a triangle: speed, scale, understanding. What gets sacrificed when you chase the first two is precisely what makes formalization valuable — refactoring definitions, exploring generalizations, giving humans and AIs a library to keep doing mathematics in. OpenAI's Navier–Stokes is both counterexample and hybrid: the correctness signal genuinely mattered (the informal proof is nearly unreadable), yet the community still protested — provenance, credit, and output-over-understanding.
- 用一枚不公正的硬币,造出公正的随机(2014)Making a Fair Coin from an Unfair One (2014)
- 问题:你只有一个偏心硬币,正面概率是某个未知的 $p$(不是 1/2),怎么从它得到公平的随机判定?你甚至不许知道 $p$ 是多少。The problem: you hold one biased coin — heads with some unknown probability $p \neq 1/2$ — and need fair random bits. You're not even allowed to know $p$.
- 关键想法来自 von Neumann:把硬币连抛两次。记正面概率 $p$,则「正反」的概率是 $p(1-p)$,「反正」的概率是 $(1-p)p$——这两个事件概率严格相等,无论 $p$ 是几。Von Neumann's idea: toss the coin twice. With heads probability $p$, "heads then tails" has probability $p(1-p)$ and "tails then heads" has probability $(1-p)p$ — strictly equal, whatever $p$ is.
- 所以算法是:成对地抛,只要两次结果相同就作废重来;一旦出现「一正一反」,就取第一枚的结果。对称性保证答案恰好是 1/2,而且对 $p$ 完全无感。So: toss in pairs, discard any pair that matches, and the moment you see one of each, output the first toss. Symmetry guarantees exactly 1/2, with total indifference to $p$.
- 代价可以精确算出来:设单轮出现「两枚不同」的概率为 $s = 2p(1-p)$,期望轮数 $t = 2/s = 1/(p(1-p))$。$p$ 越极端越烧硬币——$p = 0.9$ 时平均要抛 11 枚才得 1 个公平位。The cost is computable: a pair differs with probability $s = 2p(1-p)$, so the expected number of tosses is $t = 2/s = 1/(p(1-p))$. The more extreme $p$ is, the more you burn — at $p = 0.9$ you flip 11 times per fair bit on average.
- 带走:不完美的随机源可以合成完美的随机性,前提只有「各次独立」。这个 1951 年的小技巧是整个随机化算法可信度的地基:只要你的随机性有一点点真,就能把它提纯成完全的真。Takeaway: an imperfect random source can be refined into perfect randomness, needing only independence. This 1951 trick is bedrock for randomized algorithms: as long as your randomness is even a little real, you can purify it into the completely real.
- 爱因斯坦近看、梦露远看:一张图里的傅里叶变换(2014)Einstein Up Close, Monroe From Afar: Fourier Analysis in One Image (2014)
- 起点是两个艺术之谜:蒙娜丽莎的微笑你直视它就消失、用余光看才出现;达利的《Gala》近看是个女人、退后二十米变成林肯。两位画家都没列方程,但他们都在操纵同一件东西。Two art mysteries to start: Mona Lisa's smile vanishes when you look straight at it and reappears in your peripheral vision; Dali's Gala painting is a woman up close and Abraham Lincoln from twenty meters. Neither painter wrote equations, yet both were manipulating the same thing.
- 关键想法:傅里叶分析说任何图像都能拆成一层层「条纹」的叠加——低频条纹是大块的明暗轮廓,高频条纹是发丝、皱纹这类锐利细节。人眼恰好分频工作:近看抓高频,远看只剩低频。The key idea: Fourier analysis says any image decomposes into a stack of stripe patterns — low frequencies are broad light/dark shapes, high frequencies are sharp details like wiry hair and wrinkles. Human vision splits by frequency too: close up you read high frequencies, from afar only low ones survive.
- 做法于是简单得过分:把爱因斯坦照片做傅里叶变换,用高斯函数滤掉低频留下皱纹;把梦露滤掉高频留下平滑轮廓;两个频谱相加,逆变换回图像。The recipe is then almost cheeky: Fourier-transform Einstein, keep the high frequencies (his wrinkles); transform Monroe, keep the low frequencies (her smooth features); add the two spectra and transform back.
- 为什么成立:高斯滤波就是在频谱上乘一个钟形权重,中心是 1、向外光滑衰减到 0——「低通」留下远处可见的轮廓,「高通」留下近处可见的细节。爱因斯坦的高频与梦露的低频恰好互补,所以两个距离各看到一张完整的脸。Why it works: a Gaussian filter is a bell-shaped weight on the spectrum, 1 at the center, smoothly decaying to 0 — low-pass keeps what's visible from afar, high-pass keeps what's visible up close. Einstein's highs and Monroe's lows happen to complement each other, so each distance sees one complete face.
- 带走:傅里叶思想的核心是「把复杂东西拆成简单东西的加权和」,而混合图像给了它一个肉眼可见的演示——你看到的从来不是图像本身,是图像在你的观看距离下剩下的那一半频谱。Takeaway: Fourier's core is "decompose the complicated into a weighted sum of the simple", and hybrid images make it visible to the naked eye — you never see the image itself, only the half of its spectrum your viewing distance leaves behind.
程序员随想Programmer's Notes
- CPU 配额没用满,服务却卡死了:cgroup 节流的陷阱(2021)Under the CPU Quota but Stalled: the cgroup Throttling Trap (2021)
- 现象先立住:监控显示 CPU 使用率远低于配额,服务却在抖,尾延迟动辄几百毫秒。所有人都先去查应用代码,答案其实在内核的记账方式里。The symptom first: dashboards show CPU usage well under quota, yet the service stutters and tail latency spikes to hundreds of milliseconds. Everyone debugs application code; the answer lives in kernel accounting.
- 机制拆解:cgroup 的 CPU 配额不是「同时最多用几个核」,而是「每 100ms 窗口内总共能用多少核·毫秒」。一个 64 线程的线程池可以在窗口开头几毫秒把整个窗口的配额瞬间烧光。The mechanism: a cgroup CPU quota is not "at most N cores at once" but "at most Q core-milliseconds per 100ms window". A 64-thread pool can burn the entire window's budget in the first few milliseconds.
- 后果接踵而来:配额烧光的瞬间,整个 cgroup 被冻结(throttled)到下一个窗口才开始——哪怕你「平均」只用了一半配额。撞上冻结窗口的请求,就贡献了那几百毫秒的尾巴。The consequence follows: the moment the budget is spent, the whole cgroup is frozen until the next window — even though on average you used half the quota. Requests landing in the frozen window produce those tail spikes.
- 修法反直觉:把线程池并发调小、逼近配额对应的核数,让消耗在窗口里摊平,反而能多扛 1.5 到 2 倍负载。更少的并发,更高的吞吐。The fix is counterintuitive: shrink the pool toward the core count the quota actually affords, smoothing consumption across the window — and the service sustains 1.5 to 2x more load. Fewer threads, more throughput.
- 带走:限流的单位决定行为——「平均量」和「瞬时速率」是两个不同的物理量。凡是按窗口记账的配额系统(CPU、API、预算),都值得问一句:我的突发会把窗口烧穿吗?Takeaway: the unit of a limit determines its behavior — average usage and instantaneous rate are different physical quantities. For any windowed quota system (CPU, APIs, budgets), ask: will my burst burn through the window?
- 给 Vim 写一篇悼词:一个「冻结但健康」的 forkA Eulogy for Vim: a Fork That Chooses to Be Frozen but Healthy
- 情境:Bram Moolenaar 去世后,Vim 上游依然活着,还在合并贡献——但其中开始混进 AI 生成的低质量补丁(他称之为 slop)。Neovim 这边也在快速前进。The situation: after Bram Moolenaar's death, upstream Vim lives on and keeps merging contributions — but AI-generated low-quality patches (slop) have started slipping in. Neovim races ahead on its own.
- 他的做法不是骂,而是办后事:从 8.2.0148 起 fork 出 Vim Classic,政策极简——不再追新功能,只 backport 安全与关键修复,接受人工审过的贡献。His response isn't a rant but a funeral: fork Vim Classic from 8.2.0148 with a minimal policy — no new features, only backports of security and critical fixes, and human-reviewed contributions accepted.
- 背后的判断值得咀嚼:对一个维护者已逝的编辑器,「活着」有两种形态——继续接收一切变更,或者冻结在一个已知好的状态、只做减法维护。他选了后者,并且公开论证这是尊重而不是遗弃。The judgment underneath deserves chewing: for an editor whose maintainer has died, "alive" has two forms — keep accepting every change, or freeze at a known-good state and maintain by subtraction. He picks the latter and argues in public that this is respect, not abandonment.
- 这其实是对「维护」这个概念的重新定义:维护不等于让软件跟随世界变化,也可以是替一个定稿守住质量底线。C 编译器、排序算法、编辑器内核——很多软件的终局本该是稳定,而非永恒演化。It quietly redefines "maintenance": maintaining doesn't have to mean tracking a changing world; it can mean holding a finished thing to a quality floor. C compilers, sorting algorithms, editor cores — much software's proper end state is stability, not eternal evolution.
- 带走:fork 的正当性不总是分裂,有时是收殓。当你最常用的工具宣布「不再可靠」时,你手里其实一直有第三个选项:接过来,让它停在你信得过的地方。Takeaway: a fork isn't always schism; sometimes it's interment. When a tool you depend on announces it is no longer trustworthy, there was always a third option: take it and let it stop where you can trust it.
- 听起来像笑话的想法,什么时候值得当真?(2021)When Is a 'Crazy' Idea Worth Taking Seriously? (2021)
- 他从一个不对称现象切入:新想法被嘲笑是常态,可嘲笑的人里几乎没有行业专家——而历史一再显示,真正值得当真的新想法,恰恰在专家眼里「听起来像笑话」。He starts from an asymmetry: new ideas are mocked as a matter of course, yet the mockers are almost never domain experts — and history keeps showing that ideas worth taking seriously sound like jokes precisely to experts.
- 为什么会这样?因为范式既是思考的乐高积木,也是认知的边界。专家用现有范式搭起整个知识结构,新想法却常常意味着拆掉其中一块承重墙——在行内人看来,这不只是错的,是「不合语法」的。Why? A paradigm is both thinking's LEGO bricks and its boundary. Experts build their whole knowledge structure out of the current one; a new idea often means removing a load-bearing wall — which to insiders looks not merely wrong but ungrammatical.
- 于是他给出一个筛选器,不问想法本身,问提出者:这个人是不是这个领域的专家?他平时是不是一个合情合理、判断力经过验证的人?两个条件同时成立时,「听起来离谱」从否定信号变成了肯定信号。So the filter he offers asks not about the idea but the proposer: is this person an expert in the domain? And are they, day to day, a reasonable person with a track record of judgment? When both hold, "sounds absurd" flips from a negative signal to a positive one.
- 这个筛选器同时解释了攻击的不对称:外行攻击新 idea 是噪音,行家攻击新 idea 才是信息——但即便是行家,攻击的往往也只是「与范式的兼容性」,而非想法本身的对错。The filter also explains the asymmetry of attacks: laypeople attacking a new idea is noise; experts attacking it is information — but even experts mostly attack paradigm compatibility, not the idea's actual truth.
- 带走:判断新想法,先判断来源再判断包装。而对你自己:如果你有个想法,专家觉得像笑话、你自己又论证不出漏洞,那它值得比「再想想」更多的投入。Takeaway: judge the source before the packaging. And for yourself: if experts find your idea laughable and you can't find the flaw in your own argument, it deserves more investment than 'thinking it over'.
哲学与生活Philosophy & Life
- 提前写好你的「不」,把它放在手边Prepare Your 'No' and Keep It Handy
- 他先点破困境的来源:拒绝之所以难,不是因为不知道该拒绝,而是因为请求当面来的时候,你正处于社交压力最大的那一刻——大脑在几秒内必须同时完成判断和安抚。He names where the difficulty really comes from: refusing is hard not because you don't know you should, but because the request arrives at the moment of maximum social pressure — your brain must judge and soothe within seconds.
- 他的解法是把这两件事拆开:判断可以慢慢做,话术可以提前写。于是他预先写好一段得体的「不」,存在备忘录里;下次收到请求,三秒钟,粘贴,发送。His fix separates the two jobs: the judgment can be made slowly, and the wording can be written in advance. So he drafts a polite 'no', keeps it in a note, and next time it's three seconds: paste, send.
- 那段话有一个他验证过的四句结构:第一句把拒绝说清楚(不绕弯子,不含糊到让人误以为还有戏);第二句真诚感谢对方的信任;第三句说明你要守住的那个更大的「是」;第四句祝福对方,并留下「以后可以再问」的门。The message has a four-sentence structure he's validated: state the refusal plainly (no hedging that leaves false hope); thank them sincerely for the trust; name the bigger 'yes' you're protecting; wish them well and leave the door open for later.
- 为什么有效:请求者真正需要的其实不是你答应,而是一个可以转述、不带羞辱的明确答案。提前写好的段落恰好提供了这个——它比临场挤出的愧疚式含糊更体面。Why it works: what a requester actually needs isn't your yes but a clear, non-humiliating answer they can relay. The pre-written paragraph delivers exactly that — more graceful than guilt-fogged on-the-spot mush.
- 带走:把「说不」从一个性格问题(我不够果断)降级成一个工具问题(我有没有准备好一段话)。工具问题是可以一次性解决的。Takeaway: demote 'saying no' from a character flaw (I'm not decisive) to a tooling gap (do I have a paragraph ready?). Tooling gaps can be fixed once, forever.
- 预言从来是权力工具:当算法开始替你惊讶Prophecy Was Always a Tool of Power: When Algorithms Take Your Surprise
- 起点是一个历史观察:从德尔斐神谕到诺查丹玛斯,预言几乎从不只是预测——宣布未来的人同时在分配权力。谁说中谁可信,谁可信谁说话有人听。It starts with a historical observation: from Delphi to Nostradamus, prophecy was never mere prediction — whoever announces the future is distributing power. Predict well and you become believed; become believed and your words move people.
- 书中最有力的一个论点是「统计发明了社会」:国家开始用平均数、正态分布描述人口之后,人才第一次被当成「曲线上的一个点」来治理。预测的底层动作,是把活人抽象成数字。The book's strongest claim: "statistics invented society". Once states began describing populations through averages and the normal distribution, people could be governed as points on a curve. Prediction's underlying move is abstracting living humans into numbers.
- 然后是那句最锋利的:预测是一种言语行为,不是对事实的中性描述。你说「这个病人大概率会出院」,医院就开始按这个预期分配资源——预测会参与制造它所预测的未来。Then the sharpest line: prediction is a speech act, not a neutral description. Say "this patient will likely be discharged" and the hospital starts allocating resources around that expectation — forecasts participate in creating the futures they forecast.
- 落到算法时代:推荐系统不再满足于预测你的行为,它训练你变得可预测。Véliz 指出真正被偷走的东西很难被察觉——「自我惊奇的能力」:你不再对自己明天想要什么感到意外。Down to the algorithmic age: recommender systems no longer settle for predicting your behavior; they train you to become predictable. What gets stolen, Véliz argues, is hard to notice — the capacity for self-surprise: you're no longer ever surprised by what you'll want tomorrow.
- 带走:对一切「算命」保持一种古老的警惕——不是因为预测不准,而恰恰因为它越来越准;一个你无法再让自己惊讶的人生,值得先怀疑一下是谁在替你写剧本。Takeaway: keep an ancient wariness toward all fortune-telling — not because it's inaccurate, but because it keeps getting more accurate. A life in which you can no longer surprise yourself deserves the question: who is writing your script?
工具角Tool Corner
- Clova.Email:一个只肯收一个发件人的邮箱地址Clova.Email: an Inbox That Accepts Exactly One Sender
- 巧妙点在把邮箱地址从「身份」变成「一次性绑定」:每个地址第一次收到的邮件,其发件方就永久认领了这个地址——第一个给 chase@you.clova.email 发信的如果是 Chase,此后其他任何发件人一律在门口被拒收。The clever bit: turn an email address from an identity into a one-time binding. The first sender to email a given address claims it forever — if Chase claims chase@you.clova.email, every other sender bounces at the door from then on.
- 为什么这一下就杀死了垃圾邮件的商业模式:泄露、转卖的邮件列表依赖「一个地址发给一百家公司」,而一对一绑定让每个泄露出去的地址对买家一文不值。This kills spam's business model in one move: leaked and resold lists depend on one address serving a hundred companies; with one-to-one binding, every leaked address is worthless to the buyer.
- 还有两档变体:临时地址可以设寿命(一小时后自动拒收并可自毁内容);前缀规则让你在现场编地址——认领一个五字符以上的前缀,后面随便加后缀都是新地址,且各自锁定第一个发件人。Two more gears: temporary addresses with a lifetime (bounce after an hour, optionally wipe their contents), and a prefix rule for inventing addresses at the counter — claim a 5+ character prefix and anything after it is a new address, each locked to its first sender.
- 带走:把「过滤垃圾」的问题改写成「让垃圾发不进来」——约束越狠,模型越简单。Takeaway: rewrite 'filter the spam' into 'make spam undeliverable' — the harsher the constraint, the simpler the model.
- Esp32sim:在浏览器里跑真实的 ESP32 固件Esp32sim: Run Real ESP32 Firmware in Your Browser
- 痛点真实:手边没板子、不想装一整套交叉编译环境,就永远跨不进嵌入式开发的第一步。The pain is real: without a board on your desk and a full cross-compilation toolchain installed, you never take the first step into embedded development.
- 它的做法是模拟器而非云板卡:开源 ESP32 模拟器,真实固件本地或浏览器里直接跑,没有硬件、没有云、没有订阅。The approach is simulation, not cloud boards: an open-source ESP32 emulator that runs real firmware locally or in the browser — no hardware, no cloud, no subscription.
- 带走:把「需要物理设备」的门槛换成「需要一次点击」——教学、CI 冒烟测试、固件 demo 演示都因此换了成本结构。Takeaway: swap the gate 'you need a physical device' for 'you need one click' — teaching, CI smoke tests, and firmware demos all get a new cost structure.