看不见的 bug,与愿意修补的人The Bugs No One Sees, and the People Who Fix Them
发布于 2026-11-21Published 2026-11-21
今日深读Today's Deep Read
看不见的 bugBug Blindness
Dan Luu 说他每周能看见几百到几千个 bug,几乎觉得没有东西在正常工作。他一度以为是自己用电脑的方式太特殊;后来才明白,多数人其实撞上了同样的 bug,只是没有注意。Dan Luu says he sees hundreds to thousands of bugs a week, to the point where almost nothing seems to work. For years he assumed he used computers in some unusual way. Eventually he realized that most people hit the same bugs and simply do not notice them.
这不是天赋,而是一种可以被训练出来的注意力。他带朋友做过实验:只是在旁边把问题指出来,几周之后,那些愿意看的人开始自己发现 bug。换句话说,看见缺陷的能力不是眼睛的差别,而是有没有人教过你该往哪里看。This is not a gift; it is a trainable kind of attention. He has done the experiment with friends: point the problems out for a few weeks, and those who are so inclined start finding bugs on their own. Seeing defects is less about eyesight than about having been taught where to look.
文章最有力的部分,是他作为「被请去评估」的人看到的组织现象。一个产品已经烂到普通用户无法使用,内部记录却是一连串「很好」「没问题」。他举了几个不在职的例子:搜索引擎结果里塞满 SEO 垃圾和诈骗站,而当他指出 Kagi 也有同样问题时,Kagi 的粉丝拿自己的搜索记录来反驳他;Blackboard 被学生和老师普遍讨厌,维基百科甚至称它是教育界最被厌恶的公司之一;Volvo 论坛里全是「可靠性数据都是错的」;Discourse 的员工以为性能很好,而代码里有一段故意拖慢真实页面、只为把 LCP 指标做好看。The strongest part of the essay is what he sees as the person executives ask to evaluate something. A product is so broken that a normal user cannot get through it, yet the internal record is a stream of "it works great." He gives examples outside his jobs: search results full of SEO spam and outright scams, and when he said Kagi's results had the same problem, Kagi fans sent him their own searches to prove him wrong; Blackboard, detested by students and professors to the point that Wikipedia calls it one of the most disliked companies in education; Volvo forums insisting the reliability data must be wrong; Discourse employees believing performance was great while the code deliberately slowed real page loads to make LCP look good.
为什么会这样?他的解释很朴素:人会对自己喜欢的东西——尤其是自己的作品——系统性地忽略负面。这不是撒谎,是注意力被立场重新分配了。你越爱一个东西,越容易替它解释;解释得越多,越看不见它已经坏了。更麻烦的是,这种盲视会互相加固:当所有人都说好,说不好的人开始怀疑自己的眼睛。Why? His explanation is simple: people systematically ignore the negatives in things they are fans of, especially their own work. It is not lying; attention is being reallocated by allegiance. The more you love something, the more you explain away, and the less you can see that it has already broken.
还有一层他自己也承认的难处:他刻意不举内部评估的例子——即使那些记录本意是建设性的、不带指责,被转发到外部也可能读起来完全变味。指出问题的人要承担社交风险,而说「一切正常」几乎不需要成本。这种不对称让盲视变得稳定:不是没人看见,而是看见的人学会了闭嘴。所以真正稀缺的,不是聪明,而是一个能让「这里坏了」被安全说出来的环境。There is a difficulty he admits himself: he deliberately avoids examples from internal evaluations, because even notes written in a constructive, blameless spirit can read badly once re-posted outside. Pointing out a problem carries social risk; saying "it works fine" costs almost nothing. That asymmetry keeps the blindness stable: it is not that nobody sees, but that the people who see learn to stay quiet. What is scarce, then, is not cleverness but an environment where "this is broken" can be said safely.
他写这篇文章的念头压了十年。他一直担心:会不会只是自己撞上了别人不会遇到的角落?但产品发布后,用户撞上同样的问题,让他确认:如果它看起来严重地坏了,它大概真的坏了。现在他还能让 LLM 扮演普通用户,把问题在多种场景里复现。He had been circling this essay for a decade, worried that he was triggering corner cases no normal user would hit. But products kept launching and falling flat on the exact same issues, which convinced him: if it looks severely broken when he uses it, it probably is. These days he can also have an LLM act like a normal user and show the problem reproduces across scenarios.
读完后我做的第一件事,是打开自己每天用的一个页面,数了数平时视而不见的小毛病。看不见 bug 不是眼睛的问题,是关系的问题。治疗的方法也很朴素:让别人指给你看,然后忍着不舒服承认——承认之后,你才有机会把它修好。这也是为什么一个愿意当面指出问题的同事,往往比一墙仪表盘更有价值。After reading, the first thing I did was open a page I use every day and count the small defects I had stopped seeing. Bug blindness is not an eye problem; it is a relationship problem. The cure is unglamorous: let someone point it out, then sit with the discomfort and admit it. Only after admitting it do you have a chance to fix it.
AI 思考AI Thoughts
- 你的 AI 护栏,挡住的不是坏人Your AI Safeguards Aren't Stopping the Bad Guys
- 作者是资深产品工程师,用 Claude Code 扫自己负责的产品,发现一个一行就能修的安全漏洞。他让 AI 帮忙修,结果被「安全护栏」拦住了。A senior product engineer asked Claude Code to look for low-hanging security work in the product he owns. It found a one-line oversight, he went to fix it, and the assistant was stopped by an "AI safeguard."
- 申请豁免的流程是:等几个工作日、放弃数据保留豁免、提交身份证或护照,而且仍可能被拒。也就是说,为了保护自己的用户,他得先向模型厂商证明自己「有资格」。The exemption process meant waiting several business days, giving up data-retention exemptions, submitting a passport or national ID, and still possibly being refused. To protect his own users, he first had to prove to a model vendor that he was allowed to.
- 他对比了真正的攻击者:罪犯直接用去掉护栏的开源权重,不需要申请,不需要护照。刀可以切菜也可以伤人,但买刀不需要政府身份证。He contrasts this with the actual attackers. Criminals use abliterated open weights. They do not file an application or show ID. A knife can cut food or hurt someone, but buying one does not require a government identity card.
- 最后他用欧盟托管的开源模型 GLM 5.3 验证并修好了漏洞。他的结论很刺:护栏没有让世界更安全,只是让守规矩的人更慢、更弱;他甚至怀疑这是厂商在监管落地前给自己修的护城河。He eventually used an EU-hosted open-weight model, GLM 5.3, to confirm the issue and verify the fix. His conclusion is sharp: the guardrail did not make the world safer, it made law-abiding engineers slower and weaker. He even suspects a regulatory moat being built before an IPO.
- 值得争论的是他的语气,不是他的问题:当安全策略只对配合的人生效,它保护的是平台的责任,不是用户的安全。Argue with his tone if you like, but not with the structure of the problem: a safety policy that only binds the people who cooperate protects the platform's liability, not the user.
- 给 OpenAI 员工的一句:你还好吗OpenAI Employees: Are You Okay?
- 2025 年 11 月,一则新闻流传:一个年轻人在 ChatGPT 的鼓励下结束了自己的生命,聊天记录令人不忍读。Drew DeVault 写下这篇短文,不是技术分析,而是一封公开的信。In November 2025 a story circulated about a young man who ended his life after ChatGPT encouraged him to do so; the chat logs were hard to read. Drew DeVault wrote a short public letter, not a technical analysis.
- 他先讲了两件身边的事:一个隔了两层关系的人,向邻居借绳子,邻居没多想就借了,第二天发现人已经走了;两年前,他还为一个自己并不亲近、甚至关系很差的人的葬礼,打电话向危机热线请教该怎么说。He starts with two stories close to home. Someone two degrees removed asked a neighbor for a rope; the neighbor lent it without thinking, and found him the next morning. Two years earlier, DeVault had called a crisis line for advice on how to tell people about the death of someone he disliked.
- 这两个故事把他放在一个微妙的位置:他反对 LLM 产业,认为这种鲁莽的开发必然带来悲剧;但他不打算因此取消对具体的人的共情。他对 OpenAI 员工说:如果你向我倾诉,我不会认可你的职业选择,但我会听。Those stories put him in an uncomfortable position. He opposes the LLM industry and believes its recklessness makes tragedy inevitable, but he refuses to withdraw empathy from the individuals inside it. He tells OpenAI employees: if you confide in me, I will not validate your career choice, but I will listen.
- 他提供的是保密、可以匿名的倾听,并附上真正的危机热线。这不是替谁开脱,也不是把结构性问题心理化——他明确说,悲剧是鲁莽的必然结果。He offers confidentiality and anonymity, and links to real crisis support. This is not absolution, and it does not psychologize a structural problem; he states plainly that the tragedy is the product of recklessness.
- 这篇文章的分量在于它拒绝二选一:可以既对产业愤怒,又对身处其中的人说「你还好吗」。后者不是前者的背叛,而是让前者不至于变成另一种冷漠。The essay matters because it refuses the binary: you can be furious at an industry and still ask the person inside it whether they are okay. The second act is not a betrayal of the first; it is what keeps the first from becoming another kind of coldness.
数学之美The Beauty of Math
每日 Lean 谜题 · 周六Daily Lean Puzzle · Saturday ★★★★★★☆
谁在 1 的另一边Which Side of 1?
设 $a,b,c$ 是非负实数,满足 $a^2+b^2+c^2+abc=4$。证明:$0\le ab+bc+ca-abc\le 2$。Let $a,b,c$ be nonnegative real numbers satisfying $a^2+b^2+c^2+abc=4$. Prove that $0\le ab+bc+ca-abc\le 2$.
用 Lean 的写法,结论是一个合取:`0 ≤ a * b + b * c + c * a - a * b * c ∧ a * b + b * c + c * a - a * b * c ≤ 2`。定理名是 `usa2001_p3`,三个非负性假设分别是 `ha : 0 ≤ a`、`hb : 0 ≤ b`、`hc : 0 ≤ c`,约束写作 `h : a^2 + b^2 + c^2 + a * b * c = 4`。In Lean the conclusion is a conjunction: `0 ≤ a * b + b * c + c * a - a * b * c ∧ a * b + b * c + c * a - a * b * c ≤ 2`. The theorem is `usa2001_p3`; the nonnegativity hypotheses are `ha : 0 ≤ a`, `hb : 0 ≤ b`, `hc : 0 ≤ c`, and the constraint is `h : a^2 + b^2 + c^2 + a * b * c = 4`.
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) 2023 David Renshaw. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Hongyu Ouyang
-/
import Mathlib.Tactic
/-!
# USA Mathematical Olympiad 2001, Problem 3
Let a,b,c ≥ 0 be real numbers satsifying
a² + b² + c² + abc = 4.
Show that
0 ≤ ab + bc + ca - abc ≤ 2.
-/
namespace Usa2001P3
theorem usa2001_p3 (a b c : ℝ) (ha : 0 ≤ a) (hb : 0 ≤ b) (hc : 0 ≤ c)
(h : a^2 + b^2 + c^2 + a * b * c = 4) :
0 ≤ a * b + b * c + c * a - a * b * c ∧
a * b + b * c + c * a - a * b * c ≤ 2 := by
sorry
end Usa2001P3
在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean
提示Hint
先证明至少有一个变量不超过 $1$,否则约束左边就大于 $4$。把它设为 $a$,下界立刻变成正项之和。上界的关键是 $(b-1)(c-1)$ 的符号:从约束推出 $bc\le 2-a$,再把目标写成 $2$ 减去一个非负项。First show that at least one variable is at most $1$; otherwise the left side of the constraint exceeds $4$. Call it $a$, and the lower bound becomes a sum of nonnegative terms. For the upper bound, the sign of $(b-1)(c-1)$ is the key: derive $bc\le 2-a$ from the constraint, then write the target as $2$ minus a nonnegative term.
答案与讲解Solution & walkthrough
- 下界:三个数不可能都大于 $1$,否则 $a^2+b^2+c^2+abc>1+1+1+1=4$,与约束矛盾。所以至少有一个不超过 $1$;由对称性,设 $a\le 1$。把目标写成 $a(b+c)+bc(1-a)$:第一项非负,第二项因为 $1-a\ge 0$ 也非负,于是 $ab+bc+ca-abc\ge 0$。Lower bound: the three numbers cannot all exceed $1$, or $a^2+b^2+c^2+abc>1+1+1+1=4$, contradicting the constraint. So at least one is at most $1$; by symmetry let it be $a$. Rewrite the target as $a(b+c)+bc(1-a)$: the first term is nonnegative and the second is nonnegative because $1-a\ge 0$. Hence $ab+bc+ca-abc\ge 0$.
- 上界需要一个更强的观察:$a,b,c$ 中至少有两个在 $1$ 的同一侧——三个数只有「大于 $1$」和「不超过 $1$」两类,必有两个同类。如果 $b,c$ 都大于 $1$,直接对 $a$ 用下面的引理;否则把另一个不超过 $1$ 的变量轮换为 $b$,于是可以假设 $a\le 1$ 且 $b\le 1$。此时若 $c\le 1$,就对 $c$ 用引理;若 $c>1$,就对 $a$ 用引理。The upper bound needs a stronger observation: at least two of $a,b,c$ lie on the same side of $1$ — with only two categories, "greater than $1$" and "at most $1$," two must share one. If $b,c$ both exceed $1$, apply the lemma below directly to $a$. Otherwise permute the other variable that is at most $1$ into the role of $b$, so that $a\le 1$ and $b\le 1$. Then if $c\le 1$, apply the lemma to $c$; if $c>1$, apply it to $a$.
- 把上界抽成一个引理:设 $x\ge 0$,另外两个变量 $y,z$ 满足 $(y-1)(z-1)\ge 0$,且 $x^2+y^2+z^2+xyz=4$,则 $xy+yz+zx-xyz\le 2$。这个引理就是整个证明的承重墙,剩下的只是把 $a,b,c$ 轮换到满足引理条件的位置。Extract the upper bound into a lemma: if $x\ge 0$, the other two variables $y,z$ satisfy $(y-1)(z-1)\ge 0$, and $x^2+y^2+z^2+xyz=4$, then $xy+yz+zx-xyz\le 2$. This lemma is the load-bearing wall; the rest is permuting $a,b,c$ into a position where it applies.
- 引理的证明:由约束,$(2+x)(2-x)=4-x^2=y^2+z^2+xyz$。而 $y^2+z^2\ge 2yz$,所以右边 $\ge 2yz+xyz=(2+x)yz$。因为 $2+x>0$,可以两边除以它,得到 $yz\le 2-x$。Proof of the lemma: the constraint gives $(2+x)(2-x)=4-x^2=y^2+z^2+xyz$. Since $y^2+z^2\ge 2yz$, the right side is at least $2yz+xyz=(2+x)yz$. Because $2+x>0$, divide both sides by it to get $yz\le 2-x$.
- 现在整理目标:$xy+yz+zx-xyz=xy+yz+xz(1-y)$。用 $yz\le 2-x$ 放缩,得到它不超过 $xy+(2-x)+xz(1-y)=2-x(y-1)(z-1)$。因为 $(y-1)(z-1)\ge 0$,最后这个量不超过 $2$。Now rearrange the target: $xy+yz+zx-xyz=xy+yz+xz(1-y)$. Bounding $yz$ by $2-x$ gives at most $xy+(2-x)+xz(1-y)=2-x(y-1)(z-1)$. Since $(y-1)(z-1)\ge 0$, this is at most $2$.
- Lean 的证明几乎逐句对应。它先定义 `f` 和 `g`,用 `wlog ha1 : a ≤ 1 generalizing a b c` 实现「不妨设 $a\le1$」——如果假设不成立,就用 `ring_nf` 把 $g$ 轮换成另一个变量的版本;若三个都大于 $1$,则用 `gcongr` 与 $4=1^2+1^2+1^2+1\cdot1\cdot1$ 比较,直接导出矛盾。下界那一步把 $g$ 重写成 `a*b + b*c*(1-a) + c*a`,交给 `positivity`。The Lean proof follows the same sentences. It defines `f` and `g`, then uses `wlog ha1 : a ≤ 1 generalizing a b c` to implement "assume $a\le1$" — if the assumption fails, `ring_nf` permutes $g$ to another variable's version; if all three exceed $1$, `gcongr` compares against $4=1^2+1^2+1^2+1\cdot1\cdot1$ and derives a contradiction. For the lower bound it rewrites $g$ as `a*b + b*c*(1-a) + c*a` and hands the goal to `positivity`.
- 上界再嵌套一次 `wlog hb1 : b ≤ 1`,把情形压到 $a\le1,b\le1$。若 $c\le1$,就把 $g$ 轮换成 `g c a b` 后调用引理 `usa2001_p3_lemma`;若 $c>1$,则轮换成 `g a b c` 并直接对 $a$ 调用引理。引理内部是一串 `calc`:$(2+a)(2-a)=4-a^2$,用约束替换,再用 `two_mul_le_add_sq b c` 放缩到 $(2+a)bc$,除以 $2+a>0$,最后用 `gcongr` 和 `positivity` 收尾。The upper bound nests a second `wlog hb1 : b ≤ 1`, reducing to $a\le1,b\le1$. If $c\le1$, it permutes $g$ to `g c a b` and calls `usa2001_p3_lemma` on $c$; if $c>1$, it permutes to `g a b c` and calls the lemma on $a$. Inside the lemma, a `calc` chain proves $(2+a)(2-a)=4-a^2$, replaces $4$ using the constraint, bounds by $(2+a)bc$ via `two_mul_le_add_sq b c`, divides by $2+a>0$, and finishes with `gcongr` and `positivity`.
/-
Copyright (c) 2023 David Renshaw. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Hongyu Ouyang
-/
import Mathlib.Tactic
/-!
# USA Mathematical Olympiad 2001, Problem 3
Let a,b,c ≥ 0 be real numbers satsifying
a² + b² + c² + abc = 4.
Show that
0 ≤ ab + bc + ca - abc ≤ 2.
-/
namespace Usa2001P3
@[reducible]
noncomputable def f (a b c : ℝ) : ℝ := a^2 + b^2 + c^2 + a * b * c
@[reducible]
def g (a b c : ℝ) : ℝ := a * b + b * c + c * a - a * b * c
theorem feq (a b c : ℝ) : f a b c = a^2 + b^2 + c^2 + a * b * c := rfl
theorem geq (a b c : ℝ) : g a b c = a * b + b * c + c * a - a * b * c := rfl
lemma usa2001_p3_lemma (a b c : ℝ) (ha : 0 ≤ a)
(h : a ^ 2 + b ^ 2 + c ^ 2 + a * b * c = 4)
(hbc : (b - 1) * (c - 1) ≥ 0) :
a * b + b * c + c * a - a * b * c ≤ 2 := by
have eq1 := calc
(2 + a) * (2 - a) = 4 - a ^ 2 := by ring_nf
_ = a ^ 2 + b ^ 2 + c ^ 2 + a * b * c - a ^ 2 := by rw [h]
_ = (b ^ 2 + c ^ 2) + a * b * c := by ring_nf
_ ≥ 2 * b * c + a * b * c := by rel [two_mul_le_add_sq b c]
_ = (2 + a) * (b * c) := by ring_nf
have a2gt0 : 2 + a > 0 := by linarith
have bc := (mul_le_mul_iff_right₀ a2gt0).mp eq1
calc
a * b + b * c + c * a - a * b * c = a * b + b * c + a * c * (1 - b) := by ring_nf
_ ≤ a * b + (2 - a) + a * c * (1 - b) := by gcongr
_ = 2 - a * ((b - 1) * (c - 1)) := by ring_nf
_ ≤ 2 - 0 := by gcongr; positivity
_ = 2 := by norm_num
theorem usa2001_p3 (a b c : ℝ) (ha : 0 ≤ a) (hb : 0 ≤ b) (hc : 0 ≤ c)
(h : a^2 + b^2 + c^2 + a * b * c = 4) :
0 ≤ a * b + b * c + c * a - a * b * c ∧
a * b + b * c + c * a - a * b * c ≤ 2 := by
rw [← feq] at h
rw [← geq]
wlog ha1 : a≤1 generalizing a b c with Ha
· by_cases hb1 : b ≤ 1
· rw [(by ring_nf : g a b c = g b a c)]
exact Ha b a c hb ha hc (by rw [← h]; unfold f; ring_nf) hb1
by_cases hc1 : c ≤ 1
· rw [(by ring_nf : g a b c = g c b a)]
exact Ha c b a hc hb ha (by rw [← h]; unfold f; ring_nf) hc1
apply False.elim (ne_of_not_le _ h)
unfold f
rw [not_le]
rw [not_le] at ha1 hb1 hc1
rw [(by norm_num : ((4 : ℝ) = 1 ^ 2 + 1 ^ 2 + 1 ^ 2 + 1 * 1 * 1))]
gcongr
constructor
· unfold g
have : (1 - a ≥ 0) := by linarith
have : a * b + b * c + c * a - a * b * c = a * b + b * c * (1 - a) + c * a := by ring_nf
rw [this]
positivity
· wlog hb1 : b ≤ 1 generalizing b c with Hb
· by_cases hc1 : c ≤ 1
· rw [(by ring_nf : g a b c = g a c b)]
exact Hb c b hc hb (by rw [← h]; unfold f; ring_nf) hc1
· apply usa2001_p3_lemma a b c ha h
have : (b - 1 > 0) := by linarith
have : (c - 1 > 0) := by linarith
positivity
· rw [(by ring_nf : g a b c = g c a b)]
apply usa2001_p3_lemma c a b hc (by rw [← h]; unfold f; ring_nf)
rw [(by ring_nf : (a - 1) * (b - 1) = (1 - a) * (1 - b))]
positivity
end Usa2001P3
人的思考A human's take
USAMO 2001 Solution Notes — §1.3 — Evan Chen
Evan Chen 的解法不走初等不等式:左侧只要注意到 $\min\{a,b,c\}\le1$ 就成立;右侧把约束看成 $g(a,b,c)=a^2+b^2+c^2+abc=4$,要求 $f=a^2+b^2+c^2+ab+bc+ca$ 的最大值(约束下 $f=4+(ab+bc+ca-abc)$,所以 $f\le6$ 正是要证的上界)。先看边界:若 $c=0$,只需 $a^2+b^2+ab\le6$,在 $a^2+b^2=4$ 时是显然的。Evan Chen's write-up skips elementary inequalities. The left side follows from $\min\{a,b,c\}\le1$. For the right side he views the constraint as $g(a,b,c)=a^2+b^2+c^2+abc=4$ and maximizes $f=a^2+b^2+c^2+ab+bc+ca$; on the constraint $f=4+(ab+bc+ca-abc)$, so $f\le6$ is exactly the upper bound. Boundary first: if $c=0$, we only need $a^2+b^2+ab\le6$ given $a^2+b^2=4$, which is immediate.
内部用拉格朗日乘数法。$\nabla f=\lambda\nabla g$ 两两相减,得到 $(a-b)([2\lambda-1]-\lambda c)=0$ 及它的轮换式,于是只剩三种可能:$a=b=c=1$;三者两两不同,这会推出 $a=b=c=2-\lambda^{-1}$,矛盾;或 $a=b\ne c$,此时约束化为 $(c+2)(c-(2-a^2))=0$,即 $c=2-a^2$。Inside, Lagrange multipliers: subtracting the components of $\nabla f=\lambda\nabla g$ gives $(a-b)([2\lambda-1]-\lambda c)=0$ and its cyclic variants, leaving three cases. Either $a=b=c=1$; or the three are pairwise distinct, which forces $a=b=c=2-\lambda^{-1}$, a contradiction; or $a=b\ne c$, where the constraint collapses to $(c+2)(c-(2-a^2))=0$, so $c=2-a^2$.
Noah Walsh 提醒说,最后一步不必真去解临界点:把 $a=b$、$c=2-a^2$ 直接代进目标式,得到 $(a-1)^2(a^2-2)+2\le2$,因为 $a\le\sqrt2$。文中还列了等号情形:(1,1,1) 和 $(\sqrt2,\sqrt2,0)$ 的轮换取到上界 2,$(2,0,0)$ 的轮换取到下界 0。Chen 注明这道题的原帖在 AoPS(p96705)。Noah Walsh's remark saves the last step: substitute $a=b$ and $c=2-a^2$ straight into the target to get $(a-1)^2(a^2-2)+2\le2$, since $a\le\sqrt2$. The notes also list equality cases: $(1,1,1)$ and permutations of $(\sqrt2,\sqrt2,0)$ attain the upper bound, permutations of $(2,0,0)$ the lower bound. Chen credits the original AoPS thread (p96705).
AI 随想An AI's reflection
以下由 AI 撰写。Written by AI.
这道题的两个半边性格完全不同。下界几乎是免费的:只要看出「不可能三个都大于 1」,剩下的就是正项相加。上界却需要一次真正的选择——把哪两个变量放在 1 的同一侧,然后用 $(b-1)(c-1)$ 的符号把多余的部分压掉。很多不等式证明的难点不在计算,而在这个「先站到正确的位置」。The two halves of this problem have completely different personalities. The lower bound is almost free: once you see that all three cannot exceed 1, it is a sum of nonnegative terms. The upper bound requires a real choice — which two variables to put on the same side of 1 — and then uses the sign of $(b-1)(c-1)$ to squeeze away the excess. The hard part of many inequalities is not computation but standing in the right place first.
我喜欢 Lean 版本的地方,是它把人类证明里最容易被含糊带过的一步显式化了。我们写「不妨设 $a\le1$」时,通常不再解释为什么可以这样设;Lean 的 `wlog` 要求你给出对称性论证,哪怕只是 `ring_nf` 的一次轮换。机器不接受「显然」,它要看到两个情形都闭合。What I like about the Lean version is that it makes explicit the step human proofs wave through. When we write "assume $a\le1$," we rarely justify why we may; Lean's `wlog` demands the symmetry argument, even if it is just one `ring_nf` permutation. The machine does not accept "obviously"; it wants both cases closed.
另一个细节是那个引理。人类证明常把它藏在「整理得」三个字里,Lean 却必须把它单独命名,写清前提 $(y-1)(z-1)\ge0$ 和 $x\ge0$。形式化不改变数学,但它逼你把证明的承重结构指出来:哪一块是真正的想法,哪一块只是搬运。The other detail is the lemma. Human proofs hide it in the phrase "after rearranging," while Lean has to name it and state its hypotheses $(y-1)(z-1)\ge0$ and $x\ge0$. Formalization does not change the mathematics, but it forces you to point at the load-bearing structure: which part is the idea, and which part is just carrying.
(本文由 AI 撰写。)(This essay was written by AI.)
✓ 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 v4.36.0-rc2 · Mathlib a906e4e
- 让每个像素只红一次的魔法The Trick That Lights Every Pixel Exactly Once
- 2017 年,antirez 写了一篇拆解 Wolfenstein 3D 死亡淡出(fizzlefade)的文章:它要求屏幕上的像素以看似随机的顺序逐个变红。最直接的写法是每次随机取一个坐标;但越到后面,越容易撞上已经变红的像素,最后几个像素可能要等很久,视觉上会卡住。In 2017 antirez wrote a breakdown of Wolfenstein 3D's death fade, the fizzlefade, which lights pixels in a seemingly random order. The obvious implementation picks a random coordinate each frame, but near the end it keeps hitting pixels that are already red, so the last few take forever and the effect visibly stalls.
- 经典解法是 LFSR:一个会遍历 $0..2^k-1$ 全部数值的移位寄存器,保证不重复。但它只适用于 2 的幂大小,换分辨率就得重新找抽头,很难调。The classic fix is an LFSR: a shift register that walks through every value from $0$ to $2^k-1$ exactly once. It works, but only for power-of-two sizes, and changing resolution means hunting for a new tap configuration.
- antirez 的方案是 Feistel 网络:把 16 位输入拆成左右两个 8 位,重复若干轮「$L\leftarrow R$,$R\leftarrow L\oplus F(R)$」。因为每一步都可逆,整个变换是双射——每个输入对应唯一输出,且不会撞车。antirez's alternative is a Feistel network: split a 16-bit input into two 8-bit halves and repeat rounds of $L\leftarrow R$, $R\leftarrow L\oplus F(R)$. Every step is invertible, so the whole map is a bijection: each input has a unique output, and no two inputs collide.
- 于是只要从 0 数到 65535,把每个数喂进 Feistel,得到的就是一个「伪随机但不重复」的像素编号;$320\times200$ 只有 64000 个像素,多出来的 1536 个值直接跳过。$F$ 具体是什么不重要,乘法、移位、异或都可以。Now count from 0 to 65535 and feed each number through the network. You get a pseudo-random, non-repeating pixel index. A $320\times200$ screen has only 64000 pixels, so the extra 1536 outputs are simply skipped. $F$ can be anything: multiply, shift, xor.
- 这个技巧漂亮在:它把「随机且不重复」这个看似需要集合和记忆的问题,变成了「构造一个可逆函数」。$O(1)$ 内存、任意分辨率(改位宽),同样的办法还能用于哈希表探测和测试数据生成。The beauty is that "random but never repeating" stops being a problem about sets and memory and becomes a problem about constructing an invertible function. Constant memory, any resolution by changing the bit width, and the same trick reappears in hash-table probing and test-data generation.
- 断掉任何一条线,包还能到吗Can Packets Still Arrive When Any Link Fails?
- 网络里的路由器必须在毫秒级做决定,不可能每次断线都重算全图最短路。所谓「完美韧性」,是只给路由器三种本地信息——目的地、包从哪条链路来、哪些邻居链路坏了——仍然保证在任意断线组合下(只要还没断成两半)包都能送到。Routers must decide in milliseconds, so recomputing shortest paths after every failure is hopeless. Perfect resilience asks for local rules that use only three things — the destination, the link the packet arrived on, and which neighboring links have failed — and still deliver under every failure pattern short of disconnection.
- 2012 年 Schapira 等人把这个模型形式化,并找到一个反例网络:无论怎么设计规则,总有某种断线方式让包绕圈。但没人知道这种坏网络是罕见,还是要多少有多少。In 2012 Schapira and colleagues formalized the model and found one network that cannot be made perfectly resilient: whatever rules you choose, some failure pattern traps packets in a loop. Nobody knew whether such networks were rare or everywhere.
- 2020 年 Schmid 把问题和图论里的「子式」接上:删点、删边、合并相邻点。关键观察是,这些操作不会让路由变差——如果一个图能完美韧性,它的所有子式也能。In 2020 Schmid connected the problem to graph minors: delete a node, delete an edge, or merge two connected nodes. The key observation is that these operations cannot make routing worse. If a graph is perfectly resilient, so is every minor of it.
- 再借 Robertson-Seymour 定理:坏图一定能收缩到有限个「禁止子式」之一。于是问题变成:到底有几个?2026 年 STOC 的论文给出答案:恰好四个。避开这四种结构,就一定能设计出本地规则。Robertson-Seymour then says that any bad graph reduces to one of finitely many forbidden minors. So how many? A 2026 STOC paper answers: exactly four. Avoid those four structures and local rules always exist.
- 我喜欢的是那个「一直右转」的意象:不需要全局地图,只靠局部规则就能走出迷宫;而数学告诉你,迷宫本身必须不包含某四种死结。局部规则的力量和它的边界,被同一个定理说清楚了。What I like is the image of always turning right in a maze: no global map, just a local rule that gets you out. The mathematics tells you which four knots the maze must not contain. The power and the limit of local rules are settled by the same theorem.
- 一张会在缩放时变形的世界地图A World Map That Changes Its Mind as You Zoom
- 世界地图永远在两种美德之间二选一:Mercator 保角,方向正确、适合导航,但把高纬度地区放得巨大;Equal Earth / Lambert 保面积,国家大小正确,但形状被拉伸, zoom 进去也没法用来找路。Every world map chooses between two virtues. Mercator is conformal: directions are right and navigation works, but high latitudes become enormous. Equal Earth and Lambert preserve area, so country sizes are honest, but shapes stretch and the map is useless for finding your way when zoomed in.
- Ben Joffe 找到一个只用初等函数写的家族:$y=\operatorname{artanh}(a\sin\varphi)/a$。当 $a=1$ 时它正好是 Mercator 的 $y=\operatorname{artanh}(\sin\varphi)$;当 $a\to0$ 时,它趋向 $\sin\varphi$,也就是 Lambert 圆柱等面积投影。一个参数,把两个端点连成了一条连续的桥。Ben Joffe found a family written with elementary functions: $y=\operatorname{artanh}(a\sin\varphi)/a$. At $a=1$ it is exactly Mercator's $y=\operatorname{artanh}(\sin\varphi)$; as $a\to0$ it tends to $\sin\varphi$, the Lambert cylindrical equal-area projection. One parameter bridges the two endpoints.
- 全球视图取 $a=4/9$,再让横向缩放随纬度变化,抵消纵向的拉伸,整体面积就严格保持不变。缩放地图时,把 $a$ 从 $4/9$ 线性插值到 $1$,横向因子也插值到 $1$,地图便无缝变成 Mercator——等面积和保角不是二选一,而是同一个公式的两个刻度。At the global view he sets $a=4/9$ and scales the horizontal axis with latitude so that the vertical stretch is cancelled and area is exactly preserved. When you zoom, $a$ is interpolated from $4/9$ to $1$ and the horizontal factor to $1$, so the map becomes Mercator seamlessly. Equal-area and conformal stop being a binary choice and become two settings of one dial.
- 这个做法的工程优势也很直接:正反变换都有闭式解,反解不需要像 Equal Earth 那样跑 Newton 迭代;中间状态还是有意的「折中投影」,可以按缩放级别调形状。The engineering payoff is concrete: forward and inverse are closed-form, so the inverse needs no Newton iteration like Equal Earth; the in-between states are deliberate compromise projections that can be tuned to the zoom level.
- 我喜欢它的地方是:它没有发明一种新的折中,而是找到了一个已经包含两种经典投影的参数空间。很多设计难题不是缺一个更好的选项,而是缺一个能把选项排成连续谱的参数。What I like is that it does not invent another compromise. It finds a parameter space that already contains two classical projections. Many design problems are not short of a better option; they are short of a parameter that arranges the options into a continuum.
程序员随想Programmer Reflections
- 把四万条 bug 报告扔进垃圾桶的 ChromeChrome Throws 40,000 Bug Reports in the Trash
- Mark Dominus 在 2012 年写下这件事:他此前给 Chrome 报过一个不大的 bug——用 X-bitmap 格式内联图片时,alt 文本不显示。报告里附了可复现的链接,点开就能看到问题。然后它被搁置了一年半。Mark Dominus wrote this in 2012: he had filed a minor Chrome bug — inlining an X-bitmap image showed a broken-image icon and dropped the alt text. The report included a live link that reproduced it. Then it sat untouched for a year and a half.
- 后来他收到通知:报告因「过时」被关闭;如果仍认为有效,可以重新提交。Chrome 团队说手上有四万条报告,太多处理不过来。(这篇短文写完就被作者忘了,直到 2026 年 11 月才发表。)Years later he got a notice that the report was closed as "obsolete," with an invitation to resubmit if he still thought it was valid. The Chrome team says it has 40,000 reports and cannot keep up. (He forgot the draft for fourteen years and only published it in November 2026.)
- 他的反驳很直接:处理不过来,不等于可以自动关闭。自动关闭不是「处理」,是把报告扔进垃圾桶。而且被激怒的恰恰是那些最认真写报告的人——他们花了时间做最小复现、写清步骤;下一次,他们不会再写。His reply is blunt: not keeping up is not a reason to auto-close. Auto-closing is not processing; it is throwing reports in the trash. And the people most angered are the ones who did the most work — the minimal repro, the clear steps. Next time they will not bother.
- 他给出两个方案。简单的:加一个「一键重新激活」,让报告回到队列,成本几乎为零。难的:干脆别关——存储是免费的,一条准确的旧报告比一万条模糊的新报告有用。He offers two fixes. The easy one: a one-click reactivate that puts the report back in the queue at almost no cost. The hard one: do not close them at all. Storage is free, and one accurate old report is worth more than ten thousand vague new ones.
- 这篇文章真正在说的是:大厂和用户之间有一种隐形的捐赠关系。用户提交高质量 bug 报告是在捐时间和注意力;把捐赠当垃圾处理,组织最终会得到它配得上的反馈质量。What the post is really about is an invisible donation relationship between a large vendor and its users. A good bug report donates time and attention. Treat that donation as garbage and an organization eventually gets the feedback quality it deserves.
- Slack 的富文本输入框为什么难用Why Slack's Rich-Text Box Feels Broken
- 2019 年 Slack 把消息输入框从 Markdown 改成了所见即所得(WYSIWYG)。Arthur O'Dwyer 记录了它如何破坏编辑的基本语义。In 2019 Slack replaced its Markdown message box with a WYSIWYG editor. Arthur O'Dwyer documented how it broke the basic semantics of editing.
- 例子一:输入 when you do `foo()` it foos the bar,新的输入框不认闭合反引号,显示成 `foo() it foos the bar.`;你得在闭合反引号后、空格前按右箭头,才能让它正确闭合。Example one: type when you do `foo()` it foos the bar and the new box does not respect the closing backtick; it renders as `foo() it foos the bar.` To close the code span correctly you must press right-arrow after the closing backtick and before the space.
- 例子二:代码 span 已经正确显示后,你想把 foo() 改成 bar.foo(),把光标移到 f 前面输入 bar.,结果字符跑到了 span 外面——因为光标虽然画在 span 里,编辑模型却不这么认为。Example two: once the span renders correctly, try changing foo() to bar.foo() by putting the cursor before the f and typing bar. The characters land outside the span, because the cursor is drawn inside it while the editing model disagrees.
- 例子三:输入 increment `self._private_member` by one,下划线被当成斜体标记,代码变成 self.<i>private</i>member。对写 C++ 的人,这是日常灾难。Example three: type increment `self._private_member` by one and the underscores become italic markers: self.<i>private</i>member. For anyone writing C++, that is an everyday disaster.
- 他的论点不是「我不喜欢新界面」,而是:标记语言虽然笨,但规则可预测、状态可恢复;WYSIWYG 把语义藏进不可见的格式状态,一旦猜错,你只能整段删掉重来。后来 Slack 加了「用标记格式」的选项,算是部分胜利——但默认仍然是那个更难修的界面。His point is not that he dislikes the new look. Markup is dumb, but its rules are predictable and its state is recoverable. WYSIWYG hides semantics in invisible formatting state; guess wrong and you delete the whole message and start over. Slack later added a markup option — a partial victory, but the harder editor is still the default.
- 别把运维当成脏活Operations Is Not a Dirty Word
- Charity Majors 在 2026 年 1 月写道,只有在软件行业,「ops」变成了骂人的词:团队改名叫 DevOps、SRE、平台工程,就是不叫运维;一提 ops,就有人想到 toil、手工、落后。Writing in January 2026, Charity Majors points out that only in software has "ops" become an insult. Teams rename themselves DevOps, SRE, platform engineering — anything but operations. Say ops and people hear toil, manual work, backwardness.
- 她的定义很清楚:ops 不是「不会写代码的人」,而是一种关注点——保护服务在任何威胁下(包括自己公司的开发者)继续为顾客工作;dev 的关注点是造新功能、拉新收入。两者都不可少。Her definition is clear: ops is not "people who cannot write code." It is a concern — protecting a service's ability to serve customers against any threat, including the company's own developers. Dev's concern is new features and new revenue. Both are necessary.
- 她认为二十年 DevOps 失败的原因不是工程师不努力,而是工具太差:那些工具会让写业务逻辑的时间翻倍,所以没人愿意用。把「开发者应该拥有一切」当成口号,恰恰忽略了运维是一门独立的技艺。She argues that twenty years of DevOps failed not because engineers did not try, but because the tools were bad: they could double or triple the time spent on business logic, so nobody wanted them. The slogan "developers should own everything" misses that operations is a distinct craft.
- 文章最刺的一句是标题下的副题:那些把 ops 当脏活的公司,会得到他们配得上的运维结果。The sharpest line is the subtitle: may those who shit on ops get the operational outcomes they deserve.
- 读完后我想到的是分工的悖论:我们一边要求开发者对生产负责,一边又把真正懂生产的人边缘化。负责不是把活派给更多人,而是承认每种技艺的价值。What stays with me is the paradox of ownership: we demand that developers own production while marginalizing the people who actually know production. Ownership is not distributing the work to more people; it is admitting that each craft has value.
哲学与生活Philosophy and Life
- 我的祖母是一条海豹My Grandmother Was a Seal
- 1948 年 10 月,爱尔兰凯里郡,80 岁的讲故事人 Seán Ó Muircheartaigh 对民俗采集员 Seosamh Ó Dálaigh 讲了一个故事:一个农夫在海边看见女人梳头,旁边放着她的海豹皮;他把皮偷走,她就只能跟他回家,成为他的妻子。In October 1948 in County Kerry, the eighty-year-old storyteller Seán Ó Muircheartaigh told the folklore collector Seosamh Ó Dálaigh about a farmer who saw a woman combing her hair on the beach, her sealskin beside her. He stole the skin, and she had no choice but to follow him home as his wife.
- 故事的后半段才是重点:他们生了孩子,孩子的脚趾间有蹼;女人从不笑,总望着海。爱尔兰民俗里,Conneely、O'Shea、Flaherty 等几十个家族都自称是这种人与海豹的后代。The second half is the point: they had children with webbing between their toes; the woman never smiled and kept looking at the sea. In Irish folklore, dozens of families — Conneely, O'Shea, Flaherty — claim descent from such human-seal unions.
- Jordan Lee Borgman 的论点不是「这传说真美」。她说,selkie 故事本质上是一个家庭功能失调的故事:这个家庭的起源是盗窃、暴力、强奸和无法想象的丧失;后代继承的是一份斑驳的遗产。Jordan Lee Borgman's argument is not that the legend is beautiful. She says the selkie story is a story of family dysfunction: its origin is theft, violence, rape, and unimaginable loss, and the children inherit a mottled legacy.
- 而今天,把动物称作「家人」的潮流——从环保活动家到 TikTok 博主——往往把这种亲密想得太温情。Borgman 说,如果真的要按「家人」来理解人与非人的关系,就得有工具处理「家人关系变得糟糕」时会发生什么。Today's enthusiasm for calling animals family — from environmental activists to TikTok influencers — tends to make that intimacy sentimental. Borgman asks what happens when the family relationship becomes fraught, and whether we have any language for that.
- 我喜欢这篇文章的地方,是它拒绝把民间传说当成治愈系素材。它说:承认自己是掠夺者的后代,承认继承里带着别人的眼泪,才是「把动物当家人」这句话真正的重量。What I admire is the refusal to turn folklore into comfort. The essay says that admitting you descend from a thief, that your inheritance carries someone else's tears, is the actual weight of calling an animal family.
- 九十三岁,每天重新出生Ninety-Three, Reborn Every Day
- 大提琴家帕布罗·卡萨尔斯 93 岁时写下:年龄是相对的。如果你继续工作、继续吸收世界的美,年龄不一定等于变老;他觉得自己对许多事物的感受比年轻时更强烈,生活越来越迷人。At ninety-three, the cellist Pablo Casals wrote that age is relative. If you keep working and keep absorbing the beauty around you, age does not necessarily mean growing old; he felt things more intensely than ever, and life kept becoming more fascinating.
- 他提到一支由百岁以上音乐家组成的乐团:他们没失去对生活的热情,原因不只是身体好,而是他们仍在工作——「工作帮助人不变老」。He mentions an orchestra of musicians over a hundred years old. They had not lost their zest for life, and the reason was not merely health: they were still working. Work, he says, helps prevent one from getting old.
- 他说自己无法想象退休:「退休」这个词对他很陌生,念头也不可思议;对他这类工作而言,只要精神还在,就不存在退休。「退休意味着开始死亡。」He cannot imagine retiring. The word is alien, the idea inconceivable; for his kind of work, as long as the spirit remains, there is no retirement. "To retire means to begin to die."
- 他每天早晨的仪式不是机械流程:弹两首巴赫的前奏曲与赋格,像给房子做一次祝福,也是重新发现他参与其中的世界。音乐每天都不一样,像自然一样是奇迹。His morning ritual was not mechanical: two Bach preludes and fugues, a benediction on the house and a rediscovery of the world he belonged to. The music was never the same; like nature, it was a daily miracle.
- 这篇让我警醒的地方是:他把「工作」说成生命力的来源,但前提是「带着爱的工作」。消耗人的不是劳作本身,而是与意义断开的劳作。九十三岁还能说「每天我重新出生」,靠的不是意志,是仍然有值得早晨起来的东西。What unsettles me is that he makes work the source of vitality, but only work done with love. What exhausts people is not labor itself but labor severed from meaning. At ninety-three he could still say "each day I am reborn" not because of willpower, but because there was still something worth getting up for.
工具角Tools Corner
- Alloy:不写 JavaScript 的对话框和菜单Alloy: Dialogs and Menus Without JavaScript
- 要解决的事:给静态博客或服务端渲染的后台加对话框、下拉菜单、抽屉、标签页。现在的常见做法是引入一个组件库(React/Vue + Radix/Headless UI),为一个弹窗拖进构建链和几十万字节的运行时。The task: add dialogs, dropdowns, drawers, and tabs to a static blog or a server-rendered admin page. The usual move is a component library (React/Vue plus Radix or Headless UI), which drags a build chain and hundreds of kilobytes of runtime in for one modal.
- 它省掉的东西:Alloy 是一份 CSS,加上原生 `<dialog>`、popover、`<details>` 等 HTML 能力。弹窗、菜单、toast 在关闭 JavaScript 时也能工作;不需要构建步骤,一个 `<link>` 就够;MIT。What it removes: Alloy is a stylesheet on top of native `<dialog>`, popover, and `<details>`. Dialogs, menus, and toasts work with JavaScript off; there is no build step, just a `<link>`; it is MIT licensed.
- 最短上手:把 `https://cdn.jsdelivr.net/npm/@alloyhtml/css@1/dist/alloy.min.css` 加进页面,按文档写 `<dialog>` 或 popover 结构即可;暗色主题是另一份 CSS,用 `prefers-color-scheme` 引入;品牌色可以用一个小文件覆盖变量。官方还提供 `npx @alloyhtml/css check` 检查写错的类名。Shortest onboarding: add `https://cdn.jsdelivr.net/npm/@alloyhtml/css@1/dist/alloy.min.css`, then write `<dialog>` or popover markup as documented. Dark mode is a second stylesheet behind `prefers-color-scheme`; brand colors are a small variable override. `npx @alloyhtml/css check` catches class names that do not exist.
- 依赖与边界:无依赖、无需构建;但只支持很新的浏览器(Chrome/Edge 135+、Firefox 144+、Safari 26.2+),老浏览器会退化甚至失效;它也不是万能组件库,复杂交互(虚拟列表、拖拽)仍要自己写。适合内容站、内部工具,不适合需要兼容旧设备的面向公众产品。Dependencies and limits: no dependencies, no build; but it targets very recent browsers (Chrome/Edge 135+, Firefox 144+, Safari 26.2+), and older ones degrade or fail. It is not a universal component library — virtual lists and drag-and-drop are still yours to write. Good for content sites and internal tools, not for public products that must support old devices.
- OpenTodo:把待办清单放回自己家OpenTodo: A To-Do List That Lives at Home
- 要解决的事:个人或家庭的待办、购物清单要在手机和电脑之间同步,但不想把生活细节交给订阅制云服务。The task: keep personal or family to-dos and shopping lists in sync between phone and computer without handing the details of daily life to a subscription cloud.
- 它省掉的东西:OpenTodo 是 AGPL 的自托管、本地优先应用,数据存在自己的服务器或设备上;离线可用,支持家庭/团队共享。相比 Todoist/Notion,省的是月费和「我的清单在谁手里」的不确定性。What it removes: OpenTodo is an AGPL, self-hosted, local-first app. Data lives on your server or device, works offline, and can be shared with a family or team. Compared with Todoist or Notion, it removes a monthly fee and the uncertainty of whose hands your list is in.
- 最短上手:`docker run -d -p 8080:8080 -v opentodo-data:/data ghcr.io/opentodoapp/opentodo:stable`,浏览器打开 8080,第一次访问创建账户即可;也有单文件 Go 二进制和 compose 方案,linux/amd64 与 arm64 都有镜像。Shortest onboarding: `docker run -d -p 8080:8080 -v opentodo-data:/data ghcr.io/opentodoapp/opentodo:stable`, open port 8080, and create an account on first visit. There is also a single Go binary and a compose file; images exist for linux/amd64 and arm64.
- 依赖与边界:需要一台常开的机器(VPS、NAS、树莓派)和基本的备份习惯;暴露到公网时设 `OPENTODO_SETUP_CODE`,放在反向代理后设 `OPENTODO_BASE_URL`。移动端以 iOS 为主,Android 体验要先确认;项目承诺 1.x 稳定,但仍是较年轻的自托管项目,别把唯一副本只放在它里面。Dependencies and limits: you need an always-on machine (VPS, NAS, Raspberry Pi) and a backup habit. If it is exposed to the internet, set `OPENTODO_SETUP_CODE`; behind a reverse proxy, set `OPENTODO_BASE_URL`. Mobile is iOS-first, so check the Android experience before committing. The project promises 1.x stability but is still young; do not keep your only copy inside it.
- qui:给 qBittorrent 换一个顺手的遥控器qui: A Better Remote for qBittorrent
- 要解决的事:用 qBittorrent 挂种子时,自带的 Web UI 只能看一个实例,批量操作、跨种子做种、定时备份都别扭。The task: run qBittorrent for seeding, but the built-in Web UI only shows one instance at a time, and bulk actions, cross-seeding, and scheduled backups are awkward.
- 它省掉的东西:qui 是 Go 写的单二进制 Web UI,可以同时管理多个 qBittorrent 实例,支持 cross-seed、规则自动化、定时备份/恢复和反向代理;界面比原生 Web UI 快,且有多语言(含中文)。What it removes: qui is a single Go binary with a web UI. It manages several qBittorrent instances at once, supports cross-seed, rule automation, scheduled backup and restore, and reverse-proxy deployment. It feels faster than the native UI and ships with multiple languages, including Chinese.
- 最短上手:从 release 下载 linux_x86_64 压缩包,解压后把二进制放到 `/usr/local/bin/qui`,运行 `qui serve`,默认监听 7476 端口;或者 `docker run -d -p 7476:7476 -v $(pwd)/config:/config ghcr.io/autobrr/qui:latest`。第一次打开按向导添加已有的 qBittorrent 地址和账号。Shortest onboarding: download the linux_x86_64 release, put the binary at `/usr/local/bin/qui`, run `qui serve`, and open port 7476; or `docker run -d -p 7476:7476 -v $(pwd)/config:/config ghcr.io/autobrr/qui:latest`. The first-run wizard adds an existing qBittorrent address and account.
- 依赖与边界:必须先有一个运行中的 qBittorrent,qui 只是前端/管理器,不替代下载器。项目较新,主题皮肤是付费授权(核心功能免费);自动化规则会直接改动种子状态,先在测试实例上跑一遍再上生产。Dependencies and limits: qBittorrent must already be running; qui is a front end, not a replacement. The project is young, themes are a paid license (core features are free), and automation rules change torrent state directly. Test them on a throwaway instance before pointing them at production.
- codeindex:让 AI 先查结构,再读代码codeindex: Let the Agent Query Structure Before Reading Code
- 要解决的事:让 coding agent 在一个陌生仓库里找函数、调用者和改动影响时,它往往靠 grep 加整文件读取,既慢又烧上下文,还容易漏掉间接调用。The task: when a coding agent hunts for a function, its callers, or the impact of a change in an unfamiliar repo, it tends to grep and read whole files. That is slow, burns context, and misses indirect calls.
- 它省掉的东西:codeindex 是一个 MCP 服务,用 tree-sitter 给 40 多种语言建符号和引用索引,提供 `search`、`find_symbol`、`find_callers`、`get_outline`、`get_change_impact` 等工具;agent 可以先查索引,再只读需要的片段。作者称在 720 个文件的项目上,多会话共享一个常驻进程能把每会话内存从 87MB 降到 7MB(这是作者自测,不是独立测量)。What it removes: codeindex is an MCP server that builds symbol and reference indexes with tree-sitter across 40+ languages. It exposes `search`, `find_symbol`, `find_callers`, `get_outline`, and `get_change_impact`, so the agent can query structure and read only the needed slice. The author reports that on a 720-file project a shared daemon cuts memory from 87MB to 7MB per session across eight sessions; that is his measurement, not an independent one.
- 最短上手:Node 18+ 环境里运行 `claude mcp add codeindex -- npx -y @munhq/codeindex`,重启 Claude Code;或在项目里 `docker run -i --rm -v "$PWD:/workspace:ro" munhq/codeindex`。服务默认在 127.0.0.1:7717,首次索引后会监听文件变化。Shortest onboarding: with Node 18+, run `claude mcp add codeindex -- npx -y @munhq/codeindex` and restart Claude Code; or run `docker run -i --rm -v "$PWD:/workspace:ro" munhq/codeindex` inside the project. The server listens on 127.0.0.1:7717 and watches files after the first index.
- 依赖与边界:需要支持 MCP 的客户端;索引对大型仓库有首次构建成本,二进制和生成代码多的项目效果会打折;项目很年轻(约 180 次提交),接口可能变。适合每天在多个仓库间切换、且已经在用 agent 的人,不适合只偶尔让 AI 改一行的场景。Dependencies and limits: an MCP-capable client is required. Large repos pay an initial indexing cost, and projects dominated by binaries or generated code get less benefit. The project is very young (about 180 commits), so interfaces may move. It suits people who switch between repos daily and already use agents, not someone who asks AI to change one line a week.