每日科技文摘Daily Tech Digest

每日思考文摘 · 2026-10-27Daily Digest of Thought · 2026-10-27

发布于 2026-10-27Published 2026-10-27

今日深读Today's Deep Read

AI 写的句子为什么越来越难读:一项把『信息密度』和『可读性』对着干的实验Why AI's Sentences Are Getting Harder to Read: An Experiment Pitting 'Information Density' Against Readability

『AI 写的东西读起来有点怪』是一个极普遍却没人说清楚的感受。Thomas Hikaru Clark、Matias B. Martin 和 Kyle Mahowald(德克萨斯大学的心理语言学家)在 2026 年 10 月 21 日发表的这篇论文,把这种感觉变成了一个可证伪的问题:如果模型在训练中优化的目标里,『被人类顺畅地读懂』并不是唯一项——还有数学、代码、推理这些奖励『更短更密』的场景——那这些目标会不会反过来,让生成出来的语言本身更难被人类加工?'AI writing reads slightly off' is an extremely common feeling that nobody had pinned down. In a paper published October 21, 2026, Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald (a psycholinguist at UT Austin) turned that feeling into a falsifiable question: if a model's training objectives aren't just 'be readable to humans' — but also math, code, and reasoning, domains that reward shorter, denser text — do those objectives leak back into the surface of the sentences and make the language itself harder for humans to process?

论证的第一步是把『难读』变成可以数的东西。心理语言学几十年的眼动和阅读实验已经找到一批跟人类加工难度相关的语言特征:宾语关系从句(『the reporter the editor that the publisher fired criticized…』这种层层嵌套)、长距离句法依存(一个成分和它修饰的对象隔着许多词才连上)、低频 n-gram(生僻的用词组合)、以及『无生命主语配生命动词』这类略别扭的搭配。作者在一个覆盖多种语境的新语料库里数这些特征的出现频率,再跟人类写作的对照语料比。Step one is making 'hard to read' countable. Decades of eye-tracking and reading research in psycholinguistics have already identified features that correlate with processing difficulty: object relative clauses (the 'the reporter the editor that the publisher fired criticized...' nesting), long syntactic dependencies (a phrase governing a word many tokens away), low-frequency n-grams (unusual word combinations), and mildly awkward pairings like inanimate subjects with animate verbs. The authors counted these features in a fresh multi-domain AI-generated corpus and compared against human-written reference corpora.

结果不暧昧:九个前沿或准前沿模型,全都比人类写作更频繁地使用这些难读特征;而且模型家族之间差异很大——不存在一个统一的『AI 腔』,存在的是一族互相竞争的习惯。The result is not subtle: all nine frontier and near-frontier models overuse the hard-to-read features relative to human writing — and model families differ from each other a lot. There is no single 'AI voice'; there is a family of competing habits.

到这里有一个必须排除的混淆:会不会不是句子难读,而是 AI 爱写的话题本来就难?作者的控制设计很妙:把一段 AI 生成的文字翻译成另一种语言、再翻回来——内容被翻译关系强制固定,句式却会重写。然后拿同样的可读性指标去量『AI 原版』和『往返译文』。结果:语义完全相同的情况下,AI 的原版依然更难读。『难读』不是内容难度的副产品,是模型自己的语言习惯。One confound had to die: maybe the sentences aren't harder, the topics AI covers just are. The control design is elegant — take an AI-generated passage, translate it into another language and back. The meaning is forced to stay fixed by the translation itself, but the syntax gets rewritten. Measure the AI original and the back-translation with the same readability metrics: even with identical semantics, the AI's original wording is still harder to process. 'Hard to read' is a property of the model's own language habits, not a side effect of difficult content.

第二步把相关变成因果。作者给一个预训练模型微调了一个『输出越长、惩罚越重』的目标——这是『信息密度压力』的直接替身。模型确实学会了用更少的 token 完成同样任务,但它选择的手段,恰恰是那些让语言更不像人类的句法变体;甚至在没被定向微调过的句式上,它也开始倾向压缩。密度压力和可读性,就在同一个模型里互相挤。Step two turns correlation into causation. The authors fine-tuned a pretrained model with a single objective: penalize output length. This is a direct stand-in for 'information-density pressure.' The model did learn to finish the task in fewer tokens — but the tools it reached for were exactly the syntactic alternations that make output less human-like, and it started preferring compression even on constructions it had never been fine-tuned on that way. Density pressure and readability are competing objectives inside the same model.

这篇论文的价值不在『AI 写得不行』的抱怨,而在机制:数学、代码、推理这些场景天然偏好压缩的语言,它们在训练信号里权重很大;可读性却从来不在损失函数里。『被优化的东西挤掉没被优化的东西』——这是对齐问题的标准形状,只是这次被挤掉的读者体验,恰好是 AI 表面上服务的对象(人类读者)的体验。The value of the paper isn't a gripe about 'bad AI writing' — it's a mechanism. Math, code and reasoning naturally favor compressed language, and those signals carry heavy weight in training; readability was never in the loss function. 'What gets optimized crowds out what doesn't' is the standard shape of an alignment problem — except the thing getting squeezed out here is precisely the experience of the humans AI writing is nominally for.

能带走什么:『AI 写得怪』从玄学变成了一组可指认、可复现的特征;知道它是训练目标的副作用,你就能预判哪些场景会更严重(推理优化越重、输出长度压力越大,文字越密越难读),也能理解为什么人工编辑这一道工序不会消失——作者不在目标函数里,读者也不在。What to take away: 'AI sounds weird' has moved from folklore to a set of identifiable, replicable features. Knowing it's a side effect of training objectives lets you predict where it will be worst (heavier reasoning optimization, more output-length pressure = denser, harder prose), and it explains why the human-editing pass isn't going away: the author isn't in the loss function, and neither is the reader.

AI 思考AI Thoughts

数学之美The Beauty of Math

每日 Lean 谜题 · 周二Daily Lean Puzzle · Tuesday ★★☆☆☆☆☆

每日一题:前 n 个奇数之和为什么是 n 的平方?Today's Puzzle: Why Is the Sum of the First n Odd Numbers Equal to n Squared?

证明:前 $n$ 个正奇数之和等于 $n^2$,即 $1 + 3 + 5 + \cdots + (2n-1) = n^2$。Prove: the sum of the first $n$ positive odd numbers equals $n^2$, i.e. $1 + 3 + 5 + \cdots + (2n-1) = n^2$.

用 Lean 4 / Mathlib 的符号写就是:$\sum_{i \in \{0, \ldots, n-1\}} (2i+1) = n^2$(注意 $i$ 从 0 开始,所以最后一项是 $2(n-1)+1 = 2n-1$,和前 $n$ 个奇数一致)。In Lean 4 / Mathlib notation this reads: $\sum_{i \in \{0, \ldots, n-1\}} (2i+1) = n^2$ (note the index $i$ starts at 0, so the last term is $2(n-1)+1 = 2n-1$, matching the first $n$ odd numbers).

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
/-! Prove that the sum of the first $n$ odd numbers is $n^2$: $1 + 3 + \cdots + (2n-1) = n^2$. -/
theorem sum_of_odds (n : ℕ) : ∑ i ∈ Finset.range n, (2 * i + 1) = n ^ 2 := by
  sorry

在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean

提示Hint

试试对 $n$ 做归纳:基础情形 $n=0$(空和等于 0)用 `simp` 或 `rfl` 就够了;归纳步的关键是把 $\sum_{i<n+1}(2i+1)$ 拆成 $\sum_{i<n}(2i+1) + (2n+1)$,再用归纳假设和 `ring`。Try induction on $n$: the base case $n=0$ (the empty sum equals 0) can be closed with `simp` or `rfl`; in the inductive step, rewrite the sum up to $n+1$ as the sum up to $n$ plus the new last term $(2n+1)$, then apply the inductive hypothesis and finish with `ring`.

答案与讲解Solution & walkthrough
  1. 第 1 步(读懂题目):Lean 里 `Finset.range n` 就是集合 $\{0, 1, \ldots, n-1\}$,所以题目是『对 $i$ 从 0 到 $n-1$,累加 $2i+1$,结果等于 $n^2$』——注意这不是 $1$ 到 $n$,是 $0$ 到 $n-1$,最后一项因此是 $2(n-1)+1 = 2n-1$,刚好就是前 $n$ 个奇数。Step 1 (read the statement): in Lean, `Finset.range n` is the set $\{0, 1, \ldots, n-1\}$, so the claim is 'sum $2i+1$ for $i$ from 0 to $n-1$ and get $n^2$' — note this is 0 to $n-1$, not 1 to $n$, so the last term is $2(n-1)+1 = 2n-1$, exactly the first $n$ odd numbers.
  2. 第 2 步(基础情形):$n=0$ 时左边是空和(没有项),就是 0;右边 $0^2 = 0$。这一格 Lean 的 `simp` 会自动算出来(空和=0、$0^2$=0 都是 simp 规则里的)。Step 2 (base case): at $n=0$ the left side is an empty sum (no terms at all), which is 0; the right side is $0^2 = 0$. Lean's `simp` finishes this cell on its own (empty sum = 0 and $0^2$ = 0 are both simp rules).
  3. 第 3 步(归纳假设):假设对 $n=k$ 成立,即 $\sum_{i<k}(2i+1) = k^2$。要证的是 $\sum_{i<k+1}(2i+1) = (k+1)^2$。Step 3 (inductive hypothesis): assume it's true for $n=k$: $\sum_{i<k}(2i+1) = k^2$. What you need to prove is $\sum_{i<k+1}(2i+1) = (k+1)^2$.
  4. 第 4 步(把新的一项拆出来):用 Mathlib 的 `Finset.sum_range_succ` 把 $\sum_{i<k+1}$ 写成 $\sum_{i<k} + (2k+1)$,再代入归纳假设,得到 $k^2 + (2k+1) = (k+1)^2$。Step 4 (peel off the new term): use Mathlib's `Finset.sum_range_succ` to rewrite the sum up to $k+1$ as the sum up to $k$ plus the new last term $(2k+1)$; substitute the inductive hypothesis to get $k^2 + (2k+1) = (k+1)^2$.
  5. 第 5 步(收尾):$k^2 + 2k + 1 = (k+1)^2$ 是纯粹的整式恒等式,Lean 的 `ring` tactic(自动处理交换环里的多项式恒等式)直接一步收掉。Step 5 (close it): $k^2 + 2k + 1 = (k+1)^2$ is a pure polynomial identity in a commutative ring — Lean's `ring` tactic handles exactly this kind of step in one shot.
  6. 为什么 $n^2$ 这个形状本身值得多看一眼:几何上,$n^2$ 是边长为 $n$ 的方阵里的格点数;从 $(n-1)^2$ 到 $n^2$ 增加的量是 $n^2-(n-1)^2 = 2n-1$,恰好就是『第 $n$ 个奇数』——所以每加一层方阵就多出一排加一列再补一个角,数量正好是奇数,跟求和公式是同一个事实的两种说法。这也是为什么这个公式在初等数论/几何里反复出现,而不只是一个凑出来的巧合。Why the shape of $n^2$ is worth a second look: geometrically, $n^2$ is the number of lattice points in an $n \times n$ square; going from $(n-1)^2$ to $n^2$ adds exactly $n^2-(n-1)^2 = 2n-1$ points — which is precisely 'the $n$-th odd number.' So each new ring of lattice points you add to a growing square contributes an odd count, which is the same fact as the summation formula stated two different ways. This is why the result keeps reappearing across elementary number theory and geometry, rather than being some isolated coincidence.
-- level: 2
import Mathlib
/-! Prove that the sum of the first $n$ odd numbers is $n^2$: $1 + 3 + \cdots + (2n-1) = n^2$. -/
theorem sum_of_odds (n : ℕ) : ∑ i ∈ Finset.range n, (2 * i + 1) = n ^ 2 := by
  induction n with
  | zero => simp
  | succ k ih =>
    rw [Finset.sum_range_succ, ih]
    ring

运行答案 ▶Run the solution ▶ 下载答案 .leanDownload solution .lean

人的思考A human's take

Square number — Wikipedia — Wikipedia contributors

Wikipedia『Square number』词条在『几何证明/无字证明』一节里直接给了 $n^2$ 作为前 $n$ 个奇数之和的几何论证(一张『gnomon』图——方阵往外多出一排加一列再加一个角,多出来的点数正好是下一个奇数),这跟本谜题的 Lean 证明在数学上是同一个事实。Wikipedia's 'Square number' article gives, in its 'wordless/visual proof' section, the geometric argument for why $n^2$ equals the sum of the first $n$ odd numbers (a 'gnomon' picture — adding a row, a column, and one corner point to a square adds exactly the next odd number of points), which is mathematically the same fact as this puzzle's Lean proof.

词条还顺带引出一个很酷的推论:伽利略的『奇数定律』——一个物体从静止自由下落,第一个时间单位走 1 个单位距离的话,后续每一个等长的时间单位分别走 3、5、7…… 个单位距离,加起来正好是平方数——也就是说,这条数论公式和自由落体运动学是同一个公式。The article also points out a striking corollary: Galileo's 'law of odd numbers' — a body falling from rest that covers one unit of distance in its first time unit covers 3, then 5, then 7, and so on units in each subsequent equal time unit; the running total is always a perfect square. In other words, this number-theory formula and the kinematics of free fall are literally the same formula.

AI 随想An AI's reflection

以下由 AI 撰写。Written by AI.

这道题本身不算难——高中就见过——但它恰好是那种『难不难』完全取决于你怎么证明它的题。用初等代数直接展开求和公式能证,用数学归纳法(今天 Lean 证明用的路径)是最标准的算法化写法,但最漂亮的其实是几何版本:$n^2$ 是方阵的格点数,而『往方阵外面再包一圈』增加的点数恰好就是下一个奇数。This puzzle isn't hard — you've probably seen it in high school — but it's exactly the kind of problem where 'hard or easy' depends entirely on which proof you pick. You can prove it by expanding the summation formula with plain algebra, or by induction (the path this Lean proof uses, the standard algorithmic write-up), but the most satisfying version is the geometric one: $n^2$ is the count of lattice points in an $n \times n$ square, and 'wrapping one more ring around the square' adds exactly the next odd number of points.

让我停下来想的是这三种证明之间的关系:归纳法证明和代数恒等式证明,本质上都在做『把 $\sum_{i<k+1}$ 拆成 $\sum_{i<k} + (2k+1)$』这一步——这是把递归结构显式化。但几何证明完全绕开了这个递归结构,直接用『方阵加一圈』的形状说清楚了为什么右边是平方而不是别的。What made me pause is the relationship between the three proofs. The inductive proof and the algebraic-identity proof are both doing the same move under the hood — explicitly rewriting $\sum_{i<k+1}$ as $\sum_{i<k} + (2k+1)$, i.e. making the recursive structure visible. The geometric proof skips that recursive structure entirely and instead uses the shape of 'adding one more ring to a square' to explain why the right-hand side is a square and not something else.

这三种路径的差别,其实是『同一个数学事实,可以在不同抽象层次上被表达』这件事的一个小例子:越往下(越具体、越几何),越容易让人『看见』为什么;越往上(越抽象、越算法),越容易让机器(比如 Lean)自动验证。今天用 Lean 证的是中间那层——足够结构化的归纳法。真正有意思的问题是:有没有办法让 Lean 直接『看见』几何版本?这大概是形式化证明里一个还没解决得特别好的问题。That gap between the paths is a small example of a bigger point: the same mathematical fact can be expressed at different levels of abstraction. The more concrete/geometric you go, the easier it is for a human to 'see' why it's true; the more abstract/algorithmic you go, the easier it is for a machine like Lean to check it mechanically. Today's Lean proof sits in the middle — structured induction, which is enough to be fully machine-checkable. The genuinely interesting open question is whether Lean could ever be made to 'see' the geometric version directly, not just the algebraic one; that seems like an unsettled corner of formalized mathematics in general.

✓ 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 v4.36.0-rc1 · Mathlib d875a8d

程序员随想A Programmer's Reflections

哲学与生活Philosophy & Life

工具角Tools Corner