每日科技文摘Daily Tech Digest
发布于 2026-10-17Published 2026-10-17
今日深读Today's Deep Read
2025 年 LLM 综述:我们不是在养动物,是在召唤幽灵Karpathy's 2025 LLM Review: We're Not Raising Animals, We're Summoning Ghosts
年底做技术综述的人很多,但 Karpathy 这篇之所以值得展开讲,是因为他做的不是「今年发了哪些模型」,而是「今年的训练配方哪里变了,以及这个变化说明 LLM 到底是什么东西」。他从一名天天在写训练代码的人的视角,列出了几个「事后看才明白是范式变化」的点,其中互相咬合,构成一个完整的想法链条。Year-end tech reviews are a dime a dozen, but Karpathy's is worth expanding on because he isn't listing models — he's asking where the training recipe changed, and what that says about what LLMs actually are. Writing as someone who still runs the training code himself, he lays out several paradigm shifts that only look obvious in hindsight, and they interlock into one causal chain.
第一环:训练栈多了一个新阶段。2020-2024 年的标准配方是三段:预训练(GPT-2/3)、监督微调(InstructGPT 时代的 SFT)、人类反馈强化学习(RLHF)。后两段一直是「薄薄一层」——计算量小、目标软。2025 年出现并变成标配的是第四段:RLVR,即可验证奖励的强化学习——让模型在数学题、代码谜题这类「机器能自动判对错」的环境里反复训练。关键变化在于:因为这个阶段的奖励是不可作弊的硬奖励,优化可以跑得很久、很深。模型在长跑里自发「演化」出了像人类推理的东西:把问题拆成中间步骤、来回检验。这不是谁教它的,是奖励信号逼出来的。于是 2025 年的能力进步,基本等于各家实验室把原本打算投预训练的算力转投到「更长的 RL」上——同尺寸的模型,长得多的 RL 训练。Link one: a new stage joined the stack. From 2020 to 2024 the standard recipe was three stages: pretraining, supervised fine-tuning, RLHF. The last two were always 'thin' — little compute, soft objectives. In 2025 a fourth stage became standard: RLVR, reinforcement learning on verifiable rewards — training the model over and over in environments like math puzzles and code, where correctness is machine-checkable. The key change: because the reward can't be gamed, the optimization can run long and deep. During those long runs models spontaneously evolved something like human reasoning — breaking problems into intermediate steps, double-checking back and forth. Nobody taught them that; the reward pressure produced it. So 2025's capability gains mostly equal labs redirecting compute from pretraining into 'much longer RL' — same-size models, far deeper post-training.
第二环:由此引出全文最有名的比喻——我们不是在「进化动物」,是在「召唤幽灵」。动物是几十亿年自然选择的产物,优化目标是丛林里活下去;LLM 的优化目标是模仿全人类的文本、在数学题里拿奖励、在排行榜上赢一票。两棵完全不同条件下长出来的「智能树」,不该用看动物的眼睛去看。证据就是 2025 年大家开始习惯的那种「锯齿状智能」:同一个模型,前一秒在博士级数学上碾压你,后一秒被一个越狱话术骗得像认知障碍的小学生。可验证的领域它就尖峰突起,不可验证的领域它塌下去——天才和糊涂是同一张权重上的地形。Link two: hence the essay's famous metaphor — we're not evolving animals, we're summoning ghosts. An animal is the product of billions of years of selection to survive in the jungle; an LLM is optimized to imitate humanity's text, collect rewards on puzzles, win arena votes. Two intelligence trees grown under utterly different pressures shouldn't be judged through the animal lens. The evidence is the 'jagged intelligence' everyone got used to in 2025: the same model that crushes you on PhD-level math one second gets fooled like a cognitively challenged grade-schooler by a jailbreak the next. Wherever domains are verifiable it spikes; wherever they aren't, it sinks. Genius and confusion are terrain on the same set of weights.
第三环:这个「形状」直接解释了他为什么在 2025 年对基准测试失去信任。基准测试按构造就是「可验证环境」——所以它天然落在 RLVR 的打击半径里。实验室的典型操作是「benchmaxxing」:围绕每个基准在嵌入空间里造相邻的小环境去训练,把锯齿填平到覆盖考题。在别的领域这叫作弊,在这里成了日常手艺。「在测试集上训练是新的艺术形式」——他说得很冷。当任何公开考题都会被拿去当训练信号,绝对分数就失去了意义,这也和他引用的那张「人类智能其实也各有锯齿,只是锯齿方向不同」的梗图呼应:评测的前提是你知道自己在量什么地形。Link three: that shape directly explains his loss of trust in benchmarks. Benchmarks, by construction, are verifiable environments — so they fall inside RLVR's firing line. Labs routinely 'benchmax': build little adjacent environments around each benchmark in embedding space and train on them, filling the jagged gaps until the test is covered. In another field you'd call that cheating; here it's craft. 'Training on the test set is the new art form,' he writes flatly. When any public test becomes a training signal, absolute scores stop meaning anything.
第四环:应用层的形状也在变。他认为 Cursor 的爆发揭示了一个新层次——「LLM 应用」不再只是套壳,而是做上下文工程、把多次 LLM 调用编排成越来越复杂的 DAG、给人一个可介入的界面、并提供一个「自主性滑杆」(你决定让 AI 自己跑多远)。他给的判断是:LLM 实验室负责把模型养成「能力通偏的毕业生」,而应用层负责把这些毕业生「组织、微调成上岗的专业团队」——最后这句是他的原文味道。Link four: the app layer got a shape too. Cursor's explosion, he argues, revealed a genuine new layer — 'LLM apps' that do context engineering, orchestrate many LLM calls into increasingly complex DAGs, give humans an interface, and offer an 'autonomy slider'. His bet: LLM labs grow generally capable graduates; app companies organize, tune and deploy them as working professionals.
能带走的想法是方法本身:先问「训练目标是什么」,再问「智能长什么样」,最后推「所以哪些评测会失效」。Karpathy 把三件事串成一条因果链,比任何单点观点都耐用。这也是这篇值一个深读的原因——它不是新闻复述,是一个人认真想明白了一点什么。(原文发表于 2025 年 12 月。)What you can carry away is the method itself: ask what the training objective is, then what the intelligence looks like, then which evaluations are therefore doomed. Karpathy threads three observations into one causal chain, which is why this piece — published December 2025 — earns the deep read.
AI 思考AI Thoughts
- 一个字不认识,能不能打好 Codenames?人其实一直在用「副信道」Can You Play Codenames Without Knowing a Word? Humans Cheat Via a Side Channel
- 问题:Codenames 全靠词义联想,一个完全不理解词义的 bot 有没有可能赢真人?Luu 造了个实验:bot 只记棋盘位置模式,不碰语义。The question: Codenames runs on word meaning — can a bot that understands nothing beat humans? Luu built one that only remembers board-position patterns, zero semantics.
- 反直觉的发现:熟练的真人给线索时,会不自觉地避开某些棋盘形状——比如一排五个、一块六连——他们甚至没意识到自己在用「图案」传递信息。形状成了一个副信道,人类玩家偷偷共用它。The counterintuitive finding: experienced players subconsciously avoid certain board shapes — a line of five, a blob of six — without even noticing they're encoding information in geometry. Shape is a side channel humans secretly share.
- 由此引出的第二层论点:bot 利用这个副信道就能打得不差,说明「看起来需要深刻理解的活」里,有一大块其实是可模式化的。这和他对 AI 写代码的观察同构:在那些「正确性/性能约束弱」的场景里——慢 100000 倍也算正常、有 bug 也有人兜底——AI 生成的「够好」已经可以上岗。The second layer: since a bot can ride that side channel, a lot of what looks like 'deep understanding' is actually pattern-matchable. That mirrors his view on AI code: in domains with weak correctness/performance constraints — normal to be buggy, normal to be 100000x slow — AI's 'good enough' is already shippable.
- 对「AI 迟早取代程序员」的补充论证:AI 的输出要有约束力,前提是任务有明确的 spec;而现实中大量工作恰恰是「没有 spec、靠迭代」的——这类工作 AI 反而更需要人不断喂验收标准。His addendum to 'AI will replace programmers': for AI output to be binding, the task needs a spec; much real work is spec-less and iterative, which means humans must keep feeding acceptance criteria.
- 能带走什么:评估一个系统「懂不懂」之前,先看看人类在这项任务里到底贡献了什么信息——可能比你以为的少,也可能通过你根本没注意的副信道在泄漏。Takeaway: before asking whether a system 'understands', look at how much information humans actually contribute to the task — and through what side channels they leak it.
- 数了 900 个开源 AI 项目之后:繁荣集中在哪一层After Counting 900 Open-Source AI Projects: Where the Crowd Actually Sits
- 问题:开源 AI 的热闹是均匀的还是偏科的?她不想凭印象说,于是手动翻了 896 个至少 500 star 的相关仓库(关键词 gpt/llm/generative ai),剔掉教程清单,对 845 个软件仓库做分类和逐月统计。The question: is open-source AI's boom evenly spread or lopsided? Instead of vibes, Chip Huyen hand-compiled 896 repos with 500+ stars (searched by keywords gpt/llm/generative ai), dropped tutorial lists, and classified 845 software repos with month-over-month counts.
- 关键想法:把「AI 栈」拆成三层——基础设施(vLLM、向量库、调度)、模型开发(训练框架、数据集工程)、应用开发(prompt、RAG、界面),栈外再单列模型仓库和应用类项目。这个分法本身是论点:热闹不是均匀的。Key idea: split the 'AI stack' into infrastructure (vLLM, vector DBs, scheduling), model development (training frameworks, data tooling), and application development (prompts, RAG, interfaces), plus model repos and apps outside the stack. The taxonomy itself is the argument.
- 数据结论:爆发期(2023 年 ChatGPT 之后)新增仓库集中在最底层和最上层——中间「模型开发」层开源得最少。原因不难读:训练框架要么被大厂握着手,要么门槛高到个人贡献者挤不进去。Finding: the 2023 explosion concentrated in the bottom and top layers — the middle, model development, has the least open source. Easy to read: training frameworks are either held by big labs or too expensive for individual contributors.
- 曲线为什么 2023 年 9 月后走平:她的解释带着「低垂果实已被摘完」的判断——剩余问题更难、需要更多资源,加上和公司的对话从狂热转向务实,新项目的供给自然回落。她还观察到中国生态和西方生态明显走形。Why the curve flattens after Sept 2023: her explanation carries a judgment — the low-hanging fruit is picked, what remains takes more effort and resources, and company conversations turned from mania to pragmatism. She also notes China's open-source ecosystem visibly diverging from the Western one.
- 能带走什么:判断一个技术栈哪里过热、哪里真空,别数新闻条数,数「有使用者的仓库」按层的分布——这是一份可以照着抄的分析模板。Takeaway: to see where a tech stack is overheated or hollow, don't count news items — count adopted repos by layer. A reusable analysis template.
数学之美The Beauty of Math
每日 Lean 谜题 · 周六Daily Lean Puzzle · Saturday ★★★★★★☆
1985 年 IMO 第 2 题:两条染色规则如何锁死一种颜色IMO 1985 Problem 2: Two Coloring Rules That Lock in a Single Color
固定整数 $n \ge 3$,令 $N = \{1, 2, \dots, n-1\}$;再取一个与 $n$ 互素的 $j \in N$,即 $\gcd(j,n)=1$。Fix an integer $n \ge 3$ and let $N = \{1, 2, \dots, n-1\}$; fix $j \in N$ coprime to $n$, that is, $\gcd(j,n)=1$.
把 $N$ 中每个数染成红、黑两色之一,满足两条规则:Color every number in $N$ red or black so that:
(1)对一切 $i \in N$,$i$ 与 $n-i$ 同色;(1) for all $i \in N$, $i$ and $n-i$ have the same color;
(2)对一切 $i \in N$ 且 $i \ne j$,$i$ 与 $|j-i|$ 同色。(2) for all $i \in N$ with $i \ne j$, $i$ and $|j-i|$ have the same color.
证明:$N$ 中所有数必须同色。Prove that all numbers in $N$ must receive the same color.
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) 2025 Jeremy Tan. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Jeremy Tan
-/
import Mathlib.Algebra.Order.Ring.Canonical
import Mathlib.Algebra.Order.Ring.Star
import Mathlib.Data.Int.NatAbs
import Mathlib.Data.Nat.ModEq
/-!
# International Mathematical Olympiad 1985, Problem 2
Fix a natural number $n ≥ 3$ and define $N=\{1, 2, 3, \dots, n-1\}$.
Fix another natural number $j ∈ N$ coprime to $n$. Each number in
$N$ is now colored with one of two colors, say red or black, so that:
1. $i$ and $n-i$ always receive the same color, and
2. $i$ and $|j-i|$ receive the same color for all $i ∈ N, i ≠ j$.
Prove that all numbers in $N$ must receive the same color.
-/
namespace Imo1985P2
/-- The conditions on the problem's coloring `C`.
Although its domain is all of `ℕ`, we only care about its values in `Set.Ico 1 n`. -/
def Condition (n j : ℕ) (C : ℕ → Fin 2) : Prop :=
(∀ i ∈ Set.Ico 1 n, C i = C (n - i)) ∧
∀ i ∈ Set.Ico 1 n, i ≠ j → C i = C (j - i : ℤ).natAbs
theorem imo1985_p2 {n j : ℕ} (hn : 3 ≤ n) (hj : j ∈ Set.Ico 1 n)
(cpj : Nat.Coprime n j) {C : ℕ → Fin 2} (hC : Condition n j C)
{i : ℕ} (hi : i ∈ Set.Ico 1 n) :
C i = C j := by
sorry
end Imo1985P2
在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean
提示Hint
先把 $N$ 补成模 $n$ 的全体剩余类(加上 $0$ 不花你任何代价,两条规则作用到 $0$ 上都只说「它和自己同色」)。然后注意:规则(1)是反射 $x \mapsto -x$,规则(2)是反射 $x \mapsto j-x$。接连做这两个反射,会得到什么样的「移动」?First extend $N$ to all residue classes mod $n$ — adjoining $0$ costs you nothing, since both rules applied to $0$ just say 'it matches itself'. Then notice: rule (1) is the reflection $x \mapsto -x$, and rule (2) is the reflection $x \mapsto j-x$. What kind of motion do you get by composing the two reflections?
答案与讲解Solution & walkthrough
- 看清两条规则的几何身份。把同色关系记作 $a \sim b$,并在模 $n$ 的圆环上思考:规则(1)说 $\sim$ 对反射 $x \mapsto -x \pmod n$ 封闭;规则(2)说它对反射 $x \mapsto j-x$ 封闭——取绝对值只是把这条反射「折回」$1,\dots,n-1$ 区间的写法。See what the two rules really are. Write $a \sim b$ for 'same color' and think on the circle of residues mod $n$: rule (1) says $\sim$ is closed under the reflection $x \mapsto -x$, and rule (2) says it is closed under the reflection $x \mapsto j-x$ — the absolute value is just that reflection folded back into $1,\dots,n-1$.
- 两个反射复合出一个平移:先 $x \mapsto -x$、再 $x \mapsto j-(-x)=x+j$,于是 $x \sim x+j \pmod n$(每一步都严格调用题设规则,不引入新假设)。Compose the two reflections and a translation appears: $x \mapsto -x \mapsto j-(-x) = x+j$, hence $x \sim x+j \pmod n$, each invocation being exactly one of the two given rules.
- 互素让轨道走遍全班:$\gcd(j,n)=1$ 意味着 $j, 2j, \dots, (n-1)j \pmod n$ 恰好是 $1,\dots,n-1$ 的一个排列。任取 $i$,它等于某个 $kj \bmod n$,沿轨道 $j \sim 2j \sim \cdots \sim kj$ 一路同色,全体同色得证。Lean 证明里引理 `C_mul_mod` 用 `Nat.le_induction` 对 $k$ 归纳「$C (k j \bmod n) = C j$」;归纳步按 $k j \bmod n$ 与 $j$ 的大小分两种情况——大于 $j$ 时直接套规则(2),小于 $j$ 时先用规则(2)再用规则(1)折回,正好对应第 2 步平移的两个方向。Coprimality makes the orbit hit everything: since $\gcd(j,n)=1$, the numbers $j, 2j, \dots, (n-1)j \pmod n$ are a permutation of $1,\dots,n-1$. Any $i$ equals some $kj \bmod n$, and chaining $j \sim 2j \sim \cdots \sim kj$ gives the result. In the Lean proof, the lemma `C_mul_mod` proves `C (k * j % n) = C j` by induction on $k$ via `Nat.le_induction`; the induction step splits on whether $k j \bmod n$ is greater or less than $j$ — the first case applies rule (2) directly, the second uses rule (2) then rule (1) to fold back, exactly the two directions of the translation from step 2.
- 形式化里两处不显眼但致命的守卫(值得你在浏览器里撞一次):一是归纳步必须先证明 $(k+1)j \bmod n \ne j$,否则规则(2)的例外条款不让用它——Lean 用 `Nat.ModEq.cancel_right_of_coprime` 把互素翻译成消去律来解决;二是非零性:若余数为 $0$ 则 $n \mid (k+1)j$,与 $\gcd(j,n)=1$ 且 $k+1 < n$ 矛盾,这一步保证「折回区间」的调用始终合法。Two quiet but essential guards in the formalization (worth hitting once in the browser): the induction step first needs $(k+1)j \bmod n \ne j$, otherwise rule (2)'s exception clause forbids its use — Lean gets this from `Nat.ModEq.cancel_right_of_coprime`, turning coprimality into a cancellation law; and nonzero-ness: if the residue were $0$ then $n \mid (k+1)j$, contradicting $\gcd(j,n)=1$ with $k+1 < n$, which keeps every call inside the interval.
- 收尾把任意 $i$ 接进轨道:取 $j$ 模 $n$ 的逆 $v$(Lean 里是 `Nat.exists_mul_mod_eq_one_of_coprime`),则 $i \equiv (v i \bmod n) \cdot j \pmod n$,套用第 3 步的引理即得 $C i = C j$。Close up by wiring any $i$ into the orbit: take the inverse $v$ of $j$ mod $n$ (in Lean, `Nat.exists_mul_mod_eq_one_of_coprime`), so $i \equiv (v i \bmod n) \cdot j \pmod n$, and apply the lemma from step 3 to get `C i = C j`.
/-
Copyright (c) 2025 Jeremy Tan. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Jeremy Tan
-/
import Mathlib.Algebra.Order.Ring.Canonical
import Mathlib.Algebra.Order.Ring.Star
import Mathlib.Data.Int.NatAbs
import Mathlib.Data.Nat.ModEq
/-!
# International Mathematical Olympiad 1985, Problem 2
Fix a natural number $n ≥ 3$ and define $N=\{1, 2, 3, \dots, n-1\}$.
Fix another natural number $j ∈ N$ coprime to $n$. Each number in
$N$ is now colored with one of two colors, say red or black, so that:
1. $i$ and $n-i$ always receive the same color, and
2. $i$ and $|j-i|$ receive the same color for all $i ∈ N, i ≠ j$.
Prove that all numbers in $N$ must receive the same color.
-/
namespace Imo1985P2
/-- The conditions on the problem's coloring `C`.
Although its domain is all of `ℕ`, we only care about its values in `Set.Ico 1 n`. -/
def Condition (n j : ℕ) (C : ℕ → Fin 2) : Prop :=
(∀ i ∈ Set.Ico 1 n, C i = C (n - i)) ∧
∀ i ∈ Set.Ico 1 n, i ≠ j → C i = C (j - i : ℤ).natAbs
/-
# Solution
Let $a \sim b$ denote that $a$ and $b$ have the same color.
Because $j$ is coprime to $n$, every number in $N$ is of the form $kj\bmod n$ for a unique
$1 ≤ k < n$, so it suffices to show that $kj\bmod n \sim (k-1)j\bmod n$ for $1 < k < n$.
In this range of $k$, $kj\bmod n ≠ j$, so
* if $kj\bmod n > j$, $kj\bmod n \sim kj\bmod n - j = (k-1)j\bmod n$ using rule 2;
* if $kj\bmod n < j$, $kj\bmod n \sim j - kj\bmod n \sim n - j + kj\bmod n = (k-1)j\bmod n$
using rule 2 then rule 1.
-/
/-- For `1 ≤ k < n`, `k * j % n` has the same color as `j`. -/
lemma C_mul_mod {n j : ℕ} (hn : 3 ≤ n) (hj : j ∈ Set.Ico 1 n) (cpj : Nat.Coprime n j)
{C : ℕ → Fin 2} (hC : Condition n j C) {k : ℕ} (hk : k ∈ Set.Ico 1 n) :
C (k * j % n) = C j := by
induction k, hk.1 using Nat.le_induction with
| base => rw [one_mul, Nat.mod_eq_of_lt hj.2]
| succ k hk₁ ih =>
have nej : (k + 1) * j % n ≠ j := by
by_contra! h; nth_rw 2 [← Nat.mod_eq_of_lt hj.2, ← one_mul j] at h
replace h : (k + 1) % n = 1 % n := Nat.ModEq.cancel_right_of_coprime cpj h
rw [Nat.mod_eq_of_lt hk.2, Nat.mod_eq_of_lt (by lia)] at h
lia
have b₁ : (k + 1) * j % n ∈ Set.Ico 1 n := by
refine ⟨?_, Nat.mod_lt _ (by lia)⟩
by_contra! h; rw [Nat.lt_one_iff, ← Nat.dvd_iff_mod_eq_zero] at h
have ek := Nat.eq_zero_of_dvd_of_lt (cpj.dvd_of_dvd_mul_right h) hk.2
lia
rw [← ih ⟨hk₁, Nat.lt_of_succ_lt hk.2⟩, hC.2 _ b₁ nej]
rcases nej.lt_or_gt with h | h
· rw [Int.natAbs_natCast_sub_natCast_of_ge h.le]
have b₂ : j - (k + 1) * j % n ∈ Set.Ico 1 n :=
⟨Nat.sub_pos_iff_lt.mpr h, (Nat.sub_le ..).trans_lt hj.2⟩
have q : n - (j - (k + 1) * j % n) = (k + 1) * j % n + (n - j) % n := by
rw [tsub_tsub_eq_add_tsub_of_le h.le, add_comm, Nat.add_sub_assoc hj.2.le,
Nat.mod_eq_of_lt (show n - j < n by lia)]
rw [hC.1 _ b₂, q, ← Nat.add_mod_of_add_mod_lt (by lia), ← Nat.add_sub_assoc hj.2.le, add_comm,
Nat.add_sub_assoc (Nat.le_mul_of_pos_left _ hk.1), ← tsub_one_mul,
Nat.add_mod_left, add_tsub_cancel_right]
· rw [Int.natAbs_natCast_sub_natCast_of_le h.le, Nat.mod_sub_of_le h.le]
rw [add_mul, one_mul, add_tsub_cancel_right]
theorem imo1985_p2 {n j : ℕ} (hn : 3 ≤ n) (hj : j ∈ Set.Ico 1 n)
(cpj : Nat.Coprime n j) {C : ℕ → Fin 2} (hC : Condition n j C)
{i : ℕ} (hi : i ∈ Set.Ico 1 n) :
C i = C j := by
obtain ⟨v, hv⟩ := Nat.exists_mul_mod_eq_one_of_coprime cpj.symm (by lia)
have hvi : i = (v * i % n) * j % n := by
rw [Nat.mod_mul_mod, ← mul_rotate, ← Nat.mod_mul_mod, hv.2, one_mul, Nat.mod_eq_of_lt hi.2]
have vib : v * i % n ∈ Set.Ico 1 n := by
refine ⟨(?_ : 0 < _), Nat.mod_lt _ (by lia)⟩
by_contra! h; rw [Nat.le_zero, ← Nat.dvd_iff_mod_eq_zero] at h
rw [mul_comm, ← Nat.mod_eq_of_lt (show 1 < n by lia)] at hv
have i0 := Nat.eq_zero_of_dvd_of_lt
((Nat.coprime_of_mul_modEq_one _ hv.2).symm.dvd_of_dvd_mul_left h) hi.2
subst i; simp at hi
rw [hvi, C_mul_mod hn hj cpj hC vib]
end Imo1985P2
人的思考A human's take
AoPS Wiki — 1985 IMO Problems / Problem 2 — AoPS Wiki 社区贡献者
题目页收录的社区解答(经网页存档读取)第一步就把 $1,\dots,n-1$ 补上 $0$,视作模 $n$ 的全体剩余类:加 $0$ 不引入任何新约束——两条规则作用到 $0$ 上都只说「它与自身同色」——于是推理可以毫无边角地在一个完整的圆环上进行。题目页收录的社区解答(经网页存档读取)第一步就把 $1,\dots,n-1$ 补上 $0$,视作模 $n$ 的全体剩余类:加 $0$ 不引入任何新约束——两条规则作用到 $0$ 上都只说「它与自身同色」——于是推理可以毫无边角地在一个完整的圆环上进行。
核心一步是「两条规则都是反射」:规则(1)是 $x \mapsto -x$,规则(2)是 $x \mapsto j-x$;解答者把两者复合成平移 $x \mapsto x+j$,同色关系对平移封闭,再用 $\gcd(j,n)=1$ 说明反复加 $j$ 走遍所有非零剩余类,全体同色,证明即毕。核心一步是「两条规则都是反射」:规则(1)是 $x \mapsto -x$,规则(2)是 $x \mapsto j-x$;解答者把两者复合成平移 $x \mapsto x+j$,同色关系对平移封闭,再用 $\gcd(j,n)=1$ 说明反复加 $j$ 走遍所有非零剩余类,全体同色,证明即毕。
页面末尾挂着 'Alternate solutions always welcome'——这题在 AoPS 上仍是开放讨论的题。社区解答与本谜题的 Lean 形式化走的是同一条轨道论证:人把「折回区间」和「例外条款」当作显然跳过,机器则要求把它们写成显式的分情况与不等式证明——读这份 diff,正是读「显然」二字的成本。页面末尾挂着 'Alternate solutions always welcome'——这题在 AoPS 上仍是开放讨论的题。社区解答与本谜题的 Lean 形式化走的是同一条轨道论证:人把「折回区间」和「例外条款」当作显然跳过,机器则要求把它们写成显式的分情况与不等式证明——读这份 diff,正是读「显然」二字的成本。
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,是两条看起来只会「局部合并颜色」的规则,合起来竟然是一种移动。反射复合出平移,是群论里最古老的几何直觉之一:平面上的万花筒,两次翻转等于一次滑行;这道 1985 年的 IMO 题把它缩进了一圈数字里。互素条件在这里的角色也值得玩味——它说的其实是「步长 $j$ 在模 $n$ 的圆环上没有小圈子」:锁的齿数和扳手的步数互素,你才能走遍所有位置。数论与组合几何在一条染色题里握手,而这个握手正是轨道-稳定子思想的最小可感知版本:同色类不是被逐对检查出来的,而是被一个群作用整体「轮转」出来的。对准备在浏览器里挑战它的读者,另一个收获在形式化一侧:人写「显然,反复加 $j$ 遍历所有非零剩余类」是一行字,Lean 要的是引理、归纳法、两条不显眼的不等式守卫——$\ne j$ 与非零性。这谜题的真正训练不在想出证明(那是三行的事),而在把「显然」二字逐条兑换成机器肯吃的步骤。★6 的难度,一半在题,一半在把题翻译给编译器。The most charming thing about this problem is that two rules which only ever seem to merge colors locally, composed, turn out to be a motion. Two reflections make a translation — one of the oldest geometric intuitions in group theory; a kaleidoscope of flips equals one glide, and this 1985 IMO problem compresses that into a ring of numbers. The coprimality condition is also worth savoring: it just says that walking in steps of $j$ on the circle mod $n$ never breaks into small cycles — lock tumblers and wrench steps must be coprime for the walk to cover every position. Number theory and combinatorial geometry shake hands inside a coloring puzzle, and that handshake is the minimal perceptible form of the orbit-stabilizer idea: color classes are not checked pair by pair but rolled around by a group action. For readers about to attempt the puzzle in the browser, the other lesson lives on the formalization side: a human writes 'obviously, adding $j$ repeatedly hits every nonzero residue' in one line, while Lean demands the lemma, the induction, and two unglamorous inequality guards — not equal to $j$, and nonzero. The real training of this puzzle is not finding the proof (that is three lines); it is exchanging the word 'obviously', item by item, for steps a compiler will accept.
✓ 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
- 用文件夹写程序:Folders,一门以目录层级为语法的语言Programming with Folders: Folders, a Language Where Directory Depth Is Syntax
- 问题:一门「程序」能不能不写在文本里,而直接长在文件系统里?Temkin 的答案是设计一门 esolang:Folders——文件名是变量名,嵌套深度是语法,同级子目录的字母序是执行顺序。The question: can a 'program' live not in text but in the filesystem itself? Temkin's answer is an esolang, Folders: file names are identifiers, nesting depth is syntax, and alphabetical order among siblings is execution order.
- 关键想法:每类命令由「这个节点有几个子文件夹」决定。比如深度 1 的目录是一条语句,里面挂 0 个子目录就是 nop、挂 2 个就是某种复合操作。把「缩进+控制流」这两件文本语言里靠约定俗成表达的事,变成目录结构的物理事实。Key idea: what a node means is decided by how many subfolders it has. 'Indentation' and 'control flow' — two things text languages handle by convention — become physical facts of directory structure.
- 最妙的推论:在 Windows 上,空文件夹不占存储——理论上你可以在 U 盘里「装下无限长的程序」。存储上限消失后,语言的限制回到了「人能看懂多深的嵌套」,这本身就是个漂亮的观察。The best consequence: on Windows an empty folder takes zero bytes, so in principle a USB stick can hold an infinitely long program. Once storage stops being the limit, the real limit is how deep a human can read — itself a neat observation.
- 可视化收益:文本代码可以渲染成等高线图或迷宫(高度=深度),程序第一次成为「一眼看形状」的对象——他真做了这个转换,程序的地形一目了然。Visualization bonus: a Folders program can be rendered as a contour map or maze (height = depth), so code becomes something you see the shape of at a glance. He built exactly that.
- 能带走什么:一门语言的本质约束是它的表示介质。换掉介质(从字符串换到文件系统),你会立刻看见哪些语法约定是必要的、哪些纯属习惯——这是「数据即代码」思想最干净的演示之一。Takeaway: a language's constraints are its representation medium. Swap the medium from strings to filesystems and you immediately see which syntactic conventions are essential and which are habit — one of the cleanest demos of 'data as code' thinking.
- 弱学习即强学习:AdaBoost 背后的那个反直觉定理Weak Learning Is Strong Learning: the Counterintuitive Theorem Behind AdaBoost
- 问题:什么算「学会了」?Kun 从 PAC 学习模型讲起:算法要在任意恶劣的数据分布下,以高概率做到几乎全对,才算强学习器。而弱学习器只要求比瞎猜好一点——准确率 1/2 + η 就够了,η 叫「edge」(优势)。The question: what does it mean to 'learn'? Kun starts from the PAC model — an algorithm that, under any nasty data distribution, gets nearly perfect answers with high probability is a strong learner. A weak learner need only beat a coin flip slightly: accuracy 1/2 + η, where η is 'the edge'.
- 关键想法(反直觉定理):对任何分类任务,「弱可学习」和「强可学习」等价。也就是说,哪怕你只会一个刚赢掷硬币一点点的方法,存在一个通用程序把它磨成任意精确的学习器。强弱之差是量的,不是类的。Key idea (the counterintuitive theorem): for every classification task, weak-learnability and strong-learnability are equivalent. Any barely-better-than-random method can be ground into an arbitrarily accurate learner by one generic procedure. The gap between weak and strong is quantitative, not qualitative.
- 为什么成立:boosting 的构造——反复把样本重新加权(上轮分错的加倍、分对的减半),让弱学习器每次专攻残局,最后多数投票。Kun 把 AdaBoost 按「从直觉一步步长出来」的方式推导,并给出训练误差随轮数指数下降的证明。Why it holds: boosting's construction — repeatedly reweight the examples (double the ones missed last round, halve the ones hit), making the weak learner face a fresh residual problem each time, then majority-vote. Kun derives AdaBoost 'organically' from intuition and proves training error drops exponentially with rounds.
- 有意思的张力:这个理论最初是纯数学问题(Valiant 1984 提出、Schapire 1990 构造),但落地后 AdaBoost 成了实际中打表格数据的冠军算法之一。「先证明存在一种磨法,再找到好磨法」是理论驱动工程的范例。Nice tension: the theory was a pure math question (Valiant 1984, Schapire 1990's construction), yet AdaBoost went on to win practical table-data contests. 'First prove a grinding procedure exists, then find a good one' is theory-driven engineering as textbook case.
- 能带走什么:一个「能力刻度」的样板论证——先定义两个看似天差地别的强度,再构造性地把弱的堆成强的。这套「等价+约化」的思路,日后被用来处理各种「勉强有用」的系统(集合通信、冗余纠错都同构)。Takeaway: a template for arguing 'capability scales' — define two seemingly different strengths, then constructively stack the weak into the strong. The same equivalence-plus-reduction idea recurs wherever 'barely useful' systems get amplified.
- AI 基准测试的救法:让所有模型都去「高海拔训练」Saving AI Benchmarks by Making Every Model Train at Altitude
- 问题:基准分数还能不能说明模型「有多强」?Hardt 直言不能——同一个模型,测试集稍微换个版本,准确率能从 90% 跳到 73%。分数作为绝对度量已经死了。The question: do benchmark scores say anything about how strong a model is? Hardt says no — tweak the test set a little and the same model swings from 90% to 73%. Scores as absolute measurements are dead.
- 关键想法:分数没意义,排名曾经有意义。基准的真正历史作用是「体育比赛」而不是「温度计」——它定义规则、让大家来比,而相对比较(A 超过 B)对测试集的小改动更稳健,所以 ImageNet 和语音基准的排名信号哪怕数字不可信也推动了进步。他说深度学习革命的功劳要归给排名这个结构,而不是归给分数。Key idea: scores were never meaningful, but rankings once were. A benchmark's real historical role was a sports competition, not a thermometer — it defines the rules, everyone competes, and relative order is robust to small changes in the test set. ImageNet's rank signal drove the deep-learning revolution even when its numbers couldn't be trusted.
- 为什么 LLM 时代连排名也失效了:模型会「备考」。测试集公开躺在网上,实验室把考题嵌进开发循环(也就是「在测试集上训练」),针对性刷题把差距刷出来——排名不再是能力的信号,而是备考能力的信号。Why even rankings broke in the LLM era: models cram for the exam. Test sets sit in the open on the internet, labs fold them into the development loop, and the ranking stops measuring capability and starts measuring test-prep.
- 他的药方:「高海拔训练」——承认备考不可禁止,于是让所有参赛模型都做等量的基准定向后训练。这样绝对分数照样通胀,但「谁在同等条件下练得更好」的排名重新可比,公平性和信号同时恢复。His prescription: 'training at altitude' — since exam prep can't be banned, make every entrant do an equal amount of benchmark-targeted post-training. Absolute scores still inflate, but the ranking — who, under equal conditions, converts prep into general skill — becomes comparable again.
- 能带走什么:一个度量设计原则——当被测量者会主动优化测量本身时,别去修补指标,去改「大家在什么条件下被比较」。从 1959 年第一个基准到 GPT-6,60 年的故事其实是同一个教训的两面。Takeaway: a measurement-design principle — when the measured optimizes the measure itself, don't patch the metric, change the conditions under which everyone is compared.
程序员随想Programmer Reflections
- 标准库之争问错了:难的从来不是技术,是钱该发给谁The Standard-Library Debate Is the Wrong Question: the Hard Part Is Money, Not Tech
- 问题:Rust 社区吵了很多年「标准库该大还是该小」。matklad 说这个问题问错了——真正的问题是社会架构:什么样的组织形态能持续产出高质量标准库?The question: Rust has argued for years over whether its standard library should be big or small. matklad says that's the wrong question — the real one is social architecture: what organizational form can keep producing a high-quality stdlib?
- 关键想法:对比三个样本。Python 标准库质量参差,但它把 C API 固化下来,换来数据科学生态爆发——一个「不完美但敢承诺」的胜利。Go 用 x/ 仓库当「溢出容量」,让官方孵化器吸收失败。Rust 到 2026 年竟然还没有「读一字节随机流」的标准 API。Key idea: three case studies. Python's stdlib is uneven but it froze the C-API story and got the data-science explosion — a win for 'imperfect but committed'. Go uses x/ as overflow capacity, letting an official incubator absorb failure. Rust, in 2026, still has no standard API for reading a random byte stream.
- 为什么难的不是技术:一个全球几万人依赖的字节流 API,错误不可逆,需要「把钱放进某些人口袋」的机制让少数人全职负责十年。开源的技术民主解决不了问责和资助问题——这是高利害全球协调问题。Why the hard part isn't technical: an API tens of thousands depend on worldwide is irreversible to get wrong; you need a mechanism that 'puts money into certain people's pockets' so a few can own it full-time for a decade. Open-source technical democracy doesn't solve accountability or funding — it's a high-stakes global coordination problem.
- 由此看「小标准库」口号:它其实是回避问责的修辞——把该有人养的职责推给 ecosystem。他的处方不是「变大小」,而是先设计治理和资助模型,让责任落地。His read on 'small stdlib' slogans: they're rhetoric that dodges accountability, pushing the duty to feed batteries onto 'the ecosystem'. His prescription isn't resize-the-lib; it's design the governance and funding model first.
- 能带走什么:平台类项目做架构决策前先问「谁负责十年」。功能争议往往只是治理缺陷的症状;把社会架构想清楚,技术取舍自己会浮出来。Takeaway: before architecture decisions on platform projects, ask who owns this for ten years. Feature fights are often symptoms of a missing social contract; get that right and the technical trade-offs fall out.
- 拆解「连体测试」:一条用例只咬一个断言Decompose Run-On Tests: One Assertion Per Case
- 问题:一条测试里塞好几个操作和断言(run-on test)很常见,但失败时你分不清是哪个操作坏了——测试失去了「定位」能力,只剩「报警」能力。The question: run-on tests — several operations and assertions in one case — are common, but when they fail you can't tell which operation broke. The test keeps its alarm function and loses its diagnostic one.
- 关键想法:拆的手术很机械——复制原测试,每次只保留一个操作加它自己的断言,其余砍掉;拆完的每条重新起名字,最后删掉原版。Key idea: the surgery is mechanical. Copy the test, keep one operation plus its own assertion per copy, delete the rest, then rename each surviving fragment and throw away the original.
- 为什么给测试起名字可以很奢侈:函数命名要迁就调用方语境,而测试只有一个读者——它在 CI 里红掉的那一刻。名字可以写成完整的句子,「test_cart_total_applies_tax_once」这种在普通函数里显得啰嗦的命名,在测试里恰好是最经济的文档。Why test naming gets to be extravagant: function names must fit their callers' context; a test has exactly one reader — whoever sees it go red in CI. Names can be full sentences, and a verbose 'test_cart_total_applies_tax_once' is actually the most economical documentation.
- 深层收益:拆完之后,每条测试变成一个可独立失效的假设,失败集合直接是诊断树。连体测试的失败是「某处坏了」,拆解后的失败是「这三件事坏了」。Deeper payoff: decomposed tests are individually falsifiable assumptions, so the set of failures is directly a diagnosis tree. A run-on failure says 'something broke'; decomposed failures say 'these three specific things broke'.
- 能带走什么:一次机械手术换来定位能力。他 2013 年提的这个模式,本质是把「测试的读者是谁」想清楚了——不是写它的人,是未来读红叉的人。Takeaway: one mechanical operation buys you localization. Beck's point (published 2013) rests on knowing who reads a test — not the author, but the person staring at a red X later.
- Codeberg 投票禁掉 LLM 项目之后:民主程序不保证好决定After Codeberg Votes to Ban LLM Projects: Democracy Doesn't Guarantee Good Decisions
- 背景:Codeberg(非营利代码托管平台)投票决定禁止由 LLM 撰写的项目。Ronacher 的立场不是站「该不该禁」,而是盯住程序本身暴露的问题。Background: Codeberg, the non-profit code host, voted to ban projects written by LLMs. Ronacher's angle isn't whether the ban is right — it's what the procedure revealed.
- 关键想法:多数投票做出的决定可以是个坏决定,而且坏得很合法。这次投票意味着依赖 Codeberg 的项目随时可能被一场投票请出去——哪怕这些项目早于禁令存在、且有人靠它们活着。Key idea: a majority vote can produce a bad decision, and a bad decision can be perfectly legal. The vote means projects hosted on Codeberg can be expelled by a poll later — including projects that predate the ban and that people depend on.
- 为什么比公司更不可预测:公司决策有产权和合同兜底,规则改变至少要走合同;一个没有成文宪章的投票机构,边界条件没人承诺。他说这类机构比独裁者更没法预期,因为独裁者至少个人利益和稳定挂钩。Why this is less predictable than a company: companies are bounded by property and contracts; rule changes must at least clear those. A voting institution without a written charter promises nothing about its own boundary conditions — less predictable than a dictator, he says, because a dictator's personal interest at least aligns with stability.
- 给出的药方:任何你托付基础设施的非营利平台,应该先有写下来的权力边界(宪章 + 溯及力条款),再谈功能。「社区自治」听起来美德,听起来没有可追责性。His prescription: any non-profit platform you entrust infrastructure to should have written limits on its own power (charter plus non-retroactivity clauses) before you discuss features. 'Community governance' sounds virtuous but isn't by itself accountable.
- 能带走什么:选平台时加一条检查:它的规则能不能约束它自己?一个可以事后投票改变游戏机制的机构,你托管的就不只是代码,还有对它未来行为的预期。Takeaway: add one test when choosing a platform — can its rules bind itself? Hosting on an institution that can vote the game mechanics to change after the fact means you've custodied not just code but your expectations of its future behavior.
哲学与生活Philosophy & Life
- 一个「快写完了」的念头,吃掉六年人生'Almost Done' Ate Six Years of My Life
- 问题:作者 2016 年想写一个「迷你帖」记录社会气氛的不对劲,结果卡住、扩成大工程——此后六年,他的人生被一篇「再过几个月就完成」的文章绑架。The question: in 2016 the author sat down to write a 'mini-post' about something off in the cultural air. It jammed, ballooned into an opus — and for six years his life was held hostage by an article that was always 'a couple months from done'.
- 关键想法:最危险的自欺不是「做不到」,而是「快做到了」。如果他知道自己在做一个六年项目,他会正常安排生活(订婚、约会、度假照做);正因为每次真诚地相信「收尾在即」,他敢把一切「先放一放」——一放就是六年。Key idea: the most dangerous self-deception isn't 'I can't' — it's 'I'm almost there'. Had he known it was a six-year project, he'd have kept a normal life (engagements, dates, vacations all still happened). Because each time he sincerely believed the finish was imminent, he felt licensed to put everything else on hold — for six years.
- 为什么这种自欺特别强:它同时骗过两套心理机制——沉没成本(已经投入太多不能白费)和乐观偏差(快好了快好了),而且它自我维护:每个「快完成」的错误预期本身又会积累新投入,让下一次误判更可信。Why this delusion is so strong: it fools two psychological systems at once — sunk cost and optimism bias — and it self-maintains: each wrong 'almost done' expectation adds new investment that makes the next wrong estimate feel more credible.
- 文章的诚实之处:他没有把这篇写成成功学——六年「几乎完成」的执念最后成了这篇回头看的人生复盘,标题里那句「我多希望早点知道」是全文论点。What makes it honest: he doesn't spin it as triumph liturgy. Six years of 'almost done' ended as this retrospective, and its whole argument hides in the wistful framing of looking back.
- 能带走什么:给自己项目装一个「六年视角」的体检——问一句「如果我现在知道它会占我 N 年,我还会开始吗」。凡是无法回答「什么时候收尾」的长期承诺,都会把人生装进一个「暂时搁置」的无限缓冲带。Takeaway: run a 'six-year view' check on your long-running commitments — 'if I knew today it would take N years, would I still start?' Any open-ended project that can't answer 'when does this wrap' will file your actual life into an infinite waiting room labeled 'for now'.
- 我们的阿喀琉斯之踵是概率感——而这恰好也是盔甲Our Achilles Heel Is Probability Sense — and Exactly Why It's Also Armor
- 问题:普通人最普遍的认知弱点是什么?他的答案:不理解概率——而且这个弱点正在让人越来越焦虑。The question: what is the average person's most universal cognitive weakness? His answer: not understanding probability — and it's making us more anxious all the time.
- 关键例证(具体到扎心):一个没进第一志愿大学(录取率 <6%)的学生很痛苦。Lamade 先问「美国有多少所高中」(学生猜 12000,实际约 25000),再问「有多少个全校第一」——25000 个 valedictorian,藤校一年一共收约 15000 人。也就是说,理论上至少四成的第一名都会被拒。被拒不是低概率事件,被拒才是高概率事件,「优秀=录取」的直觉图景从算术上就站不住。The sharp example: a student is devastated after missing his first-choice college (<6% acceptance). Lamade asks 'how many high schools are in the US?' (he guessed 12,000; the real number is ~25,000), then 'how many valedictorians are there?' — 25,000, while the Ivies admit about 15,000 freshmen a year. In principle, at least 40% of all valedictorians get rejected. Rejection isn't the long shot — acceptance is; the 'excellent implies admitted' picture fails arithmetic, not luck.
- 机构靠这个弱点赚钱:招生顾问卖「路径」、球队卖「总冠军梦」、资管经理卖「跑赢市场」——他引了芒格那句:跟客户说真实胜率,人家就不会把钱给我了。Institutions monetize the weakness: admissions counselors sell 'paths', franchises sell 'a championship someday', asset managers sell 'beating the market' — with Munger's line quoted: tell clients the real odds and they'd hand me no money.
- 反转(标题里没说的):写到最后他改口了——同一个「高估小概率」的缺陷,正是车库创业、越洋探险、独立宣言的心理燃料。纯粹按期望值活的人不会去创业。所以它不是要修的 bug,是要管理的体质:知道自己对哪类叙事免疫失败,比假装自己能算准更实际。The turn (not in the headline): by the end he revises — overrating long odds is also the psychological fuel of garage startups, ocean crossings, and declarations of independence. A person who lived purely by expected value would never found a company. So it's not a bug to patch but a constitution to manage: knowing which narratives you're bad at resisting matters more than pretending you can compute.
- 能带走什么:给人生设两个闹钟——被「逆袭故事」打动时提醒自己这是盔甲在发烫;因「不符合低概率剧本」而沮丧时提醒自己:那个剧本本来就容不下大多数人。Takeaway: set two alarms in life. When an underdog story moves you, remember that's your armor heating up; when you're crushed by failing to fit a low-probability script, remember the script never had room for most people anyway.
工具角Tool Corner
- Guitoro:用键盘 H/J 拨真吉他的弦Guitoro: Strum a Real Guitar with the H and J Keys
- 巧妙点:每个音不是合成器弹出来的,是真实吉他录音——和弦、拨弦、扫弦都是录好的样本,56 种扫弦模式对应键盘右手区(H/J 等),左手按弦直接映射到字母键。The clever part: every note is a real guitar recording, not synthesis — chords, plucks, strums are captured samples, with 56 strum patterns mapped to right-hand keys (H/J and friends) and fretting mapped to letters.
- 学习模式是真正的发明:把一首歌放慢到 20%,每根弦被拨响的瞬间在屏幕上实时发光,指板同步显示按法——它把 chord chart 永远表达不了的那部分(节奏和手指位置的时序耦合)变成了可观察对象。The learning mode is the actual invention: slow a song to 20% speed, each string glows in real time the moment it's plucked, the fretboard shows the fingering — it externalizes exactly what chord charts can't express: the timing coupling between rhythm and hand position.
- 能带走什么:「教学工具」的原创点常常不在加功能,在把原本只能靠内听的东西外化成可见的时间序列。Takeaway: originality in teaching tools often isn't features — it's making what you could only hear something you can watch.
- Unfenced:用官方土地边界数据量「后门直通荒野」的房子Unfenced: Houses Where the Back Door Opens onto Public Land, Measured Against Official Boundaries
- 巧妙点在数据源:市面上的「紧邻自然」靠房产描述文字(「距离森林 2 英里」),Unfenced 把 USGS PAD-US 官方公共土地边界和房产逐块做几何求交——它卖的不是地图,是「实测边界长度」这种可证伪的数字(例如「5,240 英尺实边界」)。The cleverness is the data source: listings claim 'near nature' via prose ('2 miles from forest'). Unfenced intersects official USGS PAD-US public-land boundaries with parcel geometry one house at a time — it sells not a map but a falsifiable measurement ('5,240 ft of shared border'). 13,626 parcels measured so far.
- 概念命名即产品:「走出后门,继续走」——把「与公共土地接壤」这个测绘术语翻译成生活画面,一句话讲完筛选逻辑。目前已对 13,626 处房产完成实测。Naming as product: 'Walk out the back door and keep walking' translates the survey term into a life image and states the filter in one sentence.
- 能带走什么:数据型产品的护城河经常是新数据,而是把已有的权威数据源用到别人没敢用的严格程度。Takeaway: a data product's moat is often not new data, but using an existing authoritative source with a rigor nobody else bothered for.