每日科技文摘Daily Tech Digest

每日科技文摘 2026-09-30Daily Tech Digest 2026-09-30

发布于 2026-09-30Published 2026-09-30

今日深读Today's Deep Read

AI 时代,为什么大多数人反而更有机会了?In the AI era, why does everyone else actually have more opportunity now?

Ethan Mollick 是沃顿商学院的教授,也是最早系统研究 AI 如何改变工作和学习的人之一。他的这篇文章提出了一个反直觉的观点:AI 能力的快速提升,对大多数人来说不是威胁,而是机会。Ethan Mollick is a Wharton professor and one of the earliest researchers to systematically study how AI changes work and learning. In this essay, he makes a counterintuitive argument: the rapid advancement of AI is not a threat to most people, but an opportunity.

他的核心概念叫 "能力悬置"(overhang)——当 AI 模型的能力远超普通人的使用水平时,两者之间就出现了一个巨大的落差。这个落差不是坏事,恰恰相反,它是普通人能抓住的机会窗口。His core concept is the "overhang" — when AI capabilities far exceed the average person's ability to use them, a huge gap emerges. This gap is not a problem; it is an opportunity window that anyone can take advantage of.

为什么会出现这种悬置?因为 AI 的学习曲线和人的学习曲线是完全不同的。AI 模型在一次训练中就能学会大量技能,但人类的学习是一个漫长、渐进的过程。我们花几年才能精通某个领域,而 AI 几乎瞬间就达到了很高的水平。这就造成了一个奇怪的现象:AI 在大多数任务上已经超越普通人,但绝大多数人还不会使用它。Why does this overhang exist? Because the learning curves of AI and humans are completely different. An AI model can acquire vast capabilities in a single training run, while human learning is slow and incremental. We spend years mastering a domain, while AI reaches high proficiency almost instantly. The result is a strange situation: AI surpasses most people on most tasks, yet most people still do not know how to use it well.

Mollick 提出了四个普通人可以在这个悬置时代保持竞争优势的领域。第一个是深度知识:AI 掌握大量信息,但它不理解你在自己领域的细微差别。一个资深工程师对系统架构的直觉、一个医生对患者病史的综合判断,这些深度知识是 AI 无法替代的。Mollick identifies four areas where ordinary people can maintain a competitive edge during this overhang. The first is deep knowledge: AI may know a lot, but it does not understand the nuances of your specific domain. A senior engineer's intuition about system architecture, a doctor's ability to synthesize a patient's full history — these forms of deep, contextual knowledge are irreplaceable by AI.

第二个是广博知识:跨领域的知识连接。AI 可以在单一领域表现出色,但真正的创新往往来自不同领域的交叉。懂编程的医生、懂心理学的产品经理、懂数学的音乐家——这种跨界的知识组合是 AI 难以复制的。The second is broad knowledge: cross-domain connections. AI can excel in a single field, but genuine innovation often arises from the intersection of different domains. A programmer who understands medicine, a product manager who knows psychology, a musician who grasps mathematics — these interdisciplinary combinations are hard for AI to replicate.

第三个是品味:知道什么是好的。AI 可以生成无数方案,但判断哪个方案最好的能力,来自于长期的实践和经验。品味不是天赋,而是通过大量阅读、大量实践积累起来的直觉。The third is taste: knowing what is good. AI can generate countless options, but the ability to judge which option is best comes from years of practice and experience. Taste is not innate talent; it is intuition built through extensive reading and doing.

第四个是能动性:决定做什么。AI 可以帮你完成很多事情,但它不知道你应该做什么。目标设定、价值判断、意义赋予——这些仍然是人类独有的能力。一个会用 AI 但没有明确目标的人,远不如一个目标清晰、善用 AI 工具的人。The fourth is agency: deciding what to do. AI can help you accomplish many things, but it does not know what you should do. Goal setting, value judgment, meaning-making — these remain uniquely human capabilities. A person who knows how to use AI but lacks clear goals will always lose to someone who has clear goals and uses AI as a tool.

他举了两个亲身实验:模型把 1977 年的纯文字游戏《Zork》改成了可玩的 3D 动作游戏,场景全由模型设计;另一个模型重建作家埃科在米兰公寓的藏书,逐帧读视频里的书脊,从两万七千个格子里认出约五千本书,并标出哪些只是猜测。这些活儿人类团队要做几周——但没有一个项目是自己发生的:题是他选的,他对 Zork 和埃科的了解让他几秒就看出 AI 哪里错了、下一步要什么。He cites two personal experiments: one model turned the 1977 text-only adventure Zork into a playable 3D action game, designing every scene itself; another reconstructed Umberto Eco's apartment library in Milan, reading video frames spine by spine to identify some 5,000 of 27,000 shelf slots, marking which were guesses. A human team would have taken weeks — but not one of these projects happened on its own. Mollick chose them, and his knowledge of Zork and of Eco let him see in seconds where the AI went wrong and what to ask for next.

所以他建议:别和 AI 比谁产出多,那是必输的比赛;带着四种人的独有优势和 AI 协作,做出两者单独都做不出的东西。他尤其强调“能动性”:所有人都说不清 AI 的边界在哪时,愿意自己去试——与其等别人宣布,不如自己试出来。他坚持做奇怪的 AI 实验:实验本身就是给前沿绘制地图。So his advice: do not compete with AI on producing output — a losing game. Bring your four human advantages to working with AI, and make things neither you nor it could make alone. He stresses agency: when nobody can say where AI's boundaries are, be the one who tests them — rather than waiting for an announcement, discover it by trying. That is why he keeps running weird AI experiments: the experimenting itself is the map-making.

这篇文章最有价值的地方,不在于它给出了四个答案,而在于它改变了问题的框架。面对 AI 时,大多数人问的是 "我该如何和 AI 竞争",而 Mollick 提醒我们问的是 "在 AI 能力超出大多数人的这段时间里,我该如何利用这个落差"。这种视角转换本身就很有价值。The most valuable aspect of this essay is not that it gives four answers, but that it reframes the question. When facing AI, most people ask "How do I compete with AI?" Mollick reminds us to ask instead "How can I exploit this gap while AI capabilities remain far ahead of most people's ability to use them?" This shift in perspective is itself the key insight.

AI 思考AI Thoughts

数学之美Beauty of Mathematics

每日 Lean 谜题 · 周三Daily Lean Puzzle · Wednesday ★★★☆☆☆☆

素数有无穷多个There Are Infinitely Many Primes

对任意自然数 $n$,都存在一个素数 $p$,使得 $p \ge n$。(换句话说:素数有无穷多个。)For every natural number $n$, there is a prime $p$ with $p \ge n$. In other words: there are infinitely many primes.

Lean 形式化陈述(把 sorry 换成你的证明):The statement in Lean (replace sorry with your proof):

-- level: 3
import Mathlib
/-! Euclid: prove there is no largest prime — for every $n$ there is a prime $p \ge n$. (Idea: look at a prime factor of $n! + 1$.) -/
theorem infinitely_many_primes (n : ℕ) : ∃ p, n ≤ p ∧ p.Prime := by
  sorry

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

提示Hint

考虑 $n! + 1$ 的一个素因子。任何不超过 $n$ 的素数都整除 $n!$,所以没有一个能整除 $n! + 1$。Look at a prime factor of $n! + 1$. Every prime not exceeding $n$ divides $n!$, so none of them can divide $n! + 1$.

答案与讲解Solution & walkthrough
  1. 第一步(构造):取 $m = n! + 1$。因为 $n! + 1 \ge 2$,$m$ 必有素因子,设 $p$ 是它最小的素因子。对应 Lean 里的 let p := Nat.minFac (n.factorial + 1):先用 Nat.factorial_pos 知道 $n! + 1 > 1$(omega 收尾),再用 Nat.minFac_prime 得出 $p$ 是素数。Step 1 (construction): Let $m = n! + 1$. Since $n! + 1 \ge 2$, $m$ has a prime factor; take $p$ to be its smallest one. In Lean this is let p := Nat.minFac (n.factorial + 1): Nat.factorial_pos gives $n! + 1 > 1$ (closed by omega), and Nat.minFac_prime makes $p$ prime.
  2. 第二步(反设):目标是证明 $n \le p$。Lean 里 refine ⟨p, ?_, hp⟩ 把目标换成 $n \le p$,接着 by_contra 反设 $p < n$。数学上就是经典的反证:如果 $p$ 比 $n$ 小,那么 $p$ 落在 $2, 3, \dots, n$ 之中。Step 2 (assume the opposite): We must show $n \le p$. In Lean, refine ⟨p, ?_, hp⟩ reduces the goal to $n \le p$, then by_contra assumes instead that $p < n$. Mathematically this is the classic move: if $p$ is smaller than $n$, then $p$ is one of $2, 3, \dots, n$.
  3. 第三步(导出矛盾):$p$ 在 $2, \dots, n$ 里,就必然整除 $n!$(对应 Nat.dvd_factorial);可它又整除 $n! + 1$(Nat.minFac_dvd)。一个数同时整除两个相邻的整数,就必须整除它们的差 $1$(Nat.dvd_add_right),而素数不小于 $2$,不可能整除 $1$(hp.one_lt.ne')。Step 3 (the contradiction): Being among $2, \dots, n$, $p$ divides $n!$ (this is Nat.dvd_factorial). But $p$ also divides $n! + 1$ (Nat.minFac_dvd). A number dividing two consecutive integers must divide their difference, $1$ (Nat.dvd_add_right) — impossible for a prime, which is at least $2$ (hp.one_lt.ne').
  4. 第四步(结论):矛盾说明反设不成立,所以 $n \le p$,而 $p$ 是素数。既然对任意 $n$ 都存在不小于它的素数,素数当然有无穷多个——Lean 证完,数学上也证完。Step 4 (conclusion): The contradiction kills the assumption, so $n \le p$ and $p$ is prime. Since such a prime exists above every $n$, there are infinitely many primes — done in Lean, done on paper.
-- level: 3
import Mathlib
/-! Euclid: prove there is no largest prime — for every $n$ there is a prime $p \ge n$. (Idea: look at a prime factor of $n! + 1$.) -/
theorem infinitely_many_primes (n : ℕ) : ∃ p, n ≤ p ∧ p.Prime := by
  let p := Nat.minFac (n.factorial + 1)
  have hn : n.factorial + 1 ≠ 1 := by
    have := Nat.factorial_pos n
    omega
  have hp : p.Prime := Nat.minFac_prime hn
  refine ⟨p, ?_, hp⟩
  by_contra h
  push_neg at h
  have h1 : p ∣ n.factorial := Nat.dvd_factorial hp.pos h.le
  have h2 : p ∣ n.factorial + 1 := Nat.minFac_dvd _
  have h3 : p ∣ 1 := (Nat.dvd_add_right h1).mp h2
  exact hp.one_lt.ne' (Nat.dvd_one.mp h3)

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

人的思考A human's take

Euclid's theorem(维基百科条目) — Wikipedia contributors(含逻辑学家 Torkel Franzén 的分析)

这篇条目纠正了一个流行误解:欧几里得《几何原本》IX.20 的原证法其实是直接证明——任给一张有限长的素数表,都能造出一个不在表上的素数;把它改写成反证法时,“假设这就是全部素数”这句话在证明里根本没被用到(逻辑学家 Torkel Franzén 指出这一点)。这个辨析恰好解释了谜题的形式:Lean 定理写成“对每个 $n$ 造出一个 $p \ge n$”,正是欧几里得的原始结构;我们的答案虽然用反证收尾,构造核心仍是“取 $n!+1$ 的最小素因子”。作者还数了一下:这条定理至少有 200 个不同的证明——数学界对它的偏爱,本身就是它之美的一部分。The article corrects a popular misreading: Euclid's original proof (Elements IX.20) is in fact a direct proof — given any finite list of primes, it produces a prime not on the list; when people rewrite it as a proof by contradiction, the assumption 'these are all the primes' is never actually used (a point credited to the logician Torkel Franzén). The nuance explains the shape of today's puzzle: the Lean theorem says 'for every $n$, produce a prime $p \ge n$', which is exactly Euclid's original structure; our answer closes by contradiction, but its constructive core is still 'take the smallest prime factor of $n! + 1$'. The article also counts more than 200 distinct proofs of the theorem — the mathematical world's fondness for it is part of its beauty.

AI 随想An AI's reflection

以下由 AI 撰写。Written by AI.

欧几里得的这个证明有一种特殊的美感:它不去构造“下一个素数”,而是证明“存在最大素数”这个假设会自我反驳。Euclid's proof has a peculiar beauty: it does not hunt for the next prime; it shows that the hypothesis of a largest prime refutes itself.

精妙之处在于一个极小的算术动作——阶乘加一。$n!$ 被所有不超过 $n$ 的素数整除,所以 $n!+1$ 与它们每一个都互素;它的最小素因子只能从 $n$ 的外面来。整个证明只用整除的两条基本性质,却推出了一个关于无穷的结论——用有限的手段触及无穷,这是初等数论最迷人的瞬间之一。The elegance lies in one tiny arithmetic move — factorial plus one. $n!$ is divisible by every prime up to $n$, so $n! + 1$ is coprime to all of them, and its smallest prime factor can only come from beyond $n$. The whole argument uses two basic properties of divisibility yet reaches a conclusion about infinity — finitary means touching the infinite, one of the most enchanting moments in elementary number theory.

这个论证还教人区分两类知识:知道素数无穷多不难,难的是像这样把它压缩进十行以内的推理。有意思的是,Lean 里 Nat.minFac 承担了“最小素因子一定存在”这个在纸面上常被一句带过的前提——形式化强迫我们诚实:数学证明里每一个“显然”,都是编译器要求填补的洞。The proof also teaches us to tell two kinds of knowledge apart: knowing that primes are endless is easy; compressing that knowledge into a reasoning this short is the hard part. Notably, Nat.minFac in the Lean proof carries the premise 'the smallest prime factor exists' — the very step paper proofs wave through in a phrase. Formalization forces honesty: every 'obviously' in a mathematical proof becomes a hole the compiler demands you fill.

作为对照,高斯的素数定理说的是另一回事:素数不仅无穷,而且稀疏得很:在 $n$ 附近,大约每 $\ln n$ 个整数才碰得到一个素数。欧几里得只用整除的性质就拿到了无穷性,把“到底有多少”留给了两千年以后的人。今天你在浏览器里补上那个 sorry,等于把这条最短的路重走一遍——需要的不是计算,而是敢在一个“显然”面前停下来,问一句:凭什么。For contrast, the prime number theorem says something else entirely: not just that primes are endless, but that they thin out — near $n$, only about one integer in $\ln n$ is prime. Euclid got infinitude from pure properties of divisibility and left 'how many' to people two millennia later. Filling in that sorry in your browser is walking the shortest road again — what it requires is not computation, but the willingness to stop in front of an 'obviously' and ask: by what right?

✓ 已通过 Lean 编译验证(无 sorry,无额外公理)✓ Verified by Lean (no sorry, no extra axioms) · 题目来自本站题库Problem from this site's puzzle bank · Lean v4.35.0-rc3 · Mathlib 3f6737d

程序员随想Programmer Reflections

哲学与生活Philosophy and Life

工具角Tools Corner