今日深读:AI 时代,数学家该做什么?Today's Deep Read: What Should Mathematicians Do in the Age of AI?
发布于 2026-10-10Published 2026-10-10
AI 思考AI Thoughts
- 当 AI 从合作者变成巫师:我们正在变成观众When AI Shifts from Partner to Wizard: We're Becoming the Audience
- 第一步:现象是什么——Mollick 把论文交给 GPT-5 Pro 做方法批评,九分四十秒后,AI 不仅给出了详细批评,还运行了自己的 Monte Carlo 实验,发现了一个他论文里从未注意到的数字错误。结果很好,但他完全不知道 AI 是怎么找到这个错误的。Step 1: The phenomenon—Mollick gave his paper to GPT-5 Pro for method critique. In 9 minutes 40 seconds, the AI not only gave detailed criticism but ran its own Monte Carlo experiments and found a numerical error he'd never noticed. The result was excellent, but he had no idea how the AI found it.
- 第二步:关键想法是什么——AI 正在从'合作者'(你可以参与过程、纠正错误)变成'巫师'(给你结果,但不让你知道过程)。这不只是 UX 问题,而是认知问题:如果你不知道 AI 是怎么想的,你就无法判断它下次会不会犯错。Step 2: The key idea—AI is shifting from 'partner' (you participate, correct errors) to 'wizard' (gives results without showing the process). This isn't just a UX problem; it's a cognitive one: if you don't know how AI thinks, you can't judge whether it will err next time.
- 第三步:能带走什么——Mollick 提出的新素养是'知道什么时候该召唤巫师,什么时候该亲自动手'。对于验证性任务(检查数字),巫师很好用;对于理解性任务(需要判断为什么),你仍然需要参与过程。Step 3: What to take away—Mollick proposes a new literacy: 'knowing when to summon the wizard vs. when to do it yourself.' For verification tasks (checking numbers), wizards are great; for understanding tasks (needing to judge why), you still need to participate in the process.
- 为什么编码代理这么笨?——模型是大脑,代理是身体Why Are Coding Agents So Dumb? — The Model Is the Brain, the Agent Is the Body
- 第一步:痛点是什么——mtlynch 用编码代理做了一个简单功能(为文件上传链接添加密码保护),代理把 1.5k 行代码拆成 10 个子任务,然后一个个串行执行,完全不并行。Step 1: The pain point—mtlynch used a coding agent to add password protection to file upload links. The agent split the 1.5k-line task into 10 subtasks and executed them serially, with no parallelism at all.
- 第二步:关键想法是什么——mtlynch 的核心比喻是:模型是大脑,代理是身体。大脑很聪明,但身体不会自己组织肌肉。代理应该能判断哪些任务可以并行、哪些任务应该用便宜模型、哪些需要人工介入,但它现在什么都不会。Step 2: The key idea—mtlynch's core metaphor: the model is the brain, the agent is the body. The brain is smart, but the body doesn't organize its own muscles. The agent should decide which tasks can run in parallel, which need cheaper models, and when to involve a human—but it can't do any of that yet.
- 第三步:能带走什么——mtlynch 列出了理想编码代理的十条底线:自动并行任务、按难度分配模型、写出人类可读的计划、在真实沙箱中运行、知道自己有哪些功能。现在的代理离这些标准还很远。Step 3: What to take away—mtlynch lists ten底线 for an ideal coding agent: auto-parallelize tasks, assign models by difficulty, write human-readable plans, run in real sandboxes, know its own features. Current agents are far from these standards.
- AI 检测器检测的是结构,不是内容AI Detectors Measure Structure, Not Content
- 第一步:实验是什么——Ronacher 让 LLM 根据一个 tweet 的结构重新写了一篇推文,然后他自己手工逐段重写,没有一句和原文相同。结果两个版本都被 Pangram 检测器标为 100% AI。Step 1: The experiment—Ronacher had an LLM rewrite a tweet following its structure, then manually rewrote each paragraph himself with no identical sentences. Both versions were flagged 100% AI by the Pangram detector.
- 第二步:关键想法是什么——AI 检测器(如 Pangram)的训练数据是已知人类文本和 LLM 重写的文本。如果一个文本的'结构'看起来像 LLM 生成的(段落节奏、过渡方式、论证模式),即使内容完全由人写,检测器也会标为 AI。Step 2: The key idea—AI detectors like Pangram are trained on known human text vs. LLM-rewritten text. If a text's 'structure' looks LLM-generated (paragraph rhythm, transition patterns, argument style), even fully human-written content gets flagged.
- 第三步:能带走什么——这个实验揭示了一个更广泛的问题:当 AI 生成内容变得普遍,我们用来检测 AI 的工具本身可能产生大量误报。对于认真使用 LLM 辅助写作的作者来说,'100% AI'的标签可能毫无意义。Step 3: What to take away—This reveals a broader problem: as AI-generated content becomes common, our detection tools may produce many false positives. For authors who seriously use LLMs for assistance, a '100% AI' label may be meaningless.
数学之美The Beauty of Mathematics
每日 Lean 谜题 · 周六Daily Lean Puzzle · Saturday ★★★★★★☆
IMO 1983 第 5 题:1983 个无三项等差的数IMO 1983 Problem 5: 1983 Integers with No Three in Arithmetic Progression
从 $1$ 到 $10^5$ 中,能否选出 $1983$ 个不同的正整数,使得其中任意三个数不构成等差数列?From the integers $1$ to $10^5$, can we choose $1983$ distinct positive integers such that no three form an arithmetic progression?
换句话说:是否存在一个集合 $S$,满足 $|S| = 1983$,$S$ 中每个元素 $\le 10^5$,且对任意 $x < y < z$ 属于 $S$,都有 $x + z \ne 2y$?In other words: does there exist a set $S$ with $|S| = 1983$, every element $\le 10^5$, and for any $x < y < z$ in $S$, $x + z \ne 2y$?
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) 2024 The Compfiles Contributors. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: InternLM-MATH LEAN Formalizer v0.1, Benpigchu
-/
import Mathlib.Data.List.GetD
import Mathlib.Tactic
/-!
# International Mathematical Olympiad 1983, Problem 5
Is it possible to choose $1983$ distinct positive integers,
all less than or equal to $10^5$,
no three of which are consecutive terms of an arithmetic progression?
Justify your answer.
-/
namespace Imo1983P5
theorem imo1983_p5 :
∃ S : Finset ℕ+, S.card = 1983 ∧
(∀ x ∈ S, x ≤ 10^5) ∧
∀ x ∈ S, ∀ y ∈ S, ∀ z ∈ S, x < y ∧ y < z → x + z ≠ 2 * y := by
sorry
end Imo1983P5
在浏览器中挑战 ▶Try it in your browser ▶ 下载题目 .leanDownload puzzle .lean
提示Hint
考虑把十进制数转换成三进制表示——如果一个数的三进制表示只包含 0 和 1,它会有什么性质?Consider converting numbers to their ternary (base-3) representation—if a number's ternary representation contains only digits 0 and 1, what property does it have?
答案与讲解Solution & walkthrough
- 第一步:为什么是三进制——等差数列的条件 $x + z = 2y$ 可以重新表述为:$y - x = z - y$,即 $y$ 是 $x$ 和 $z$ 的中点。如果我们构造一个集合,使得任意两个元素的和在三进制下'不会进位',那么 $x + z = 2y$ 就不可能发生。Step 1: Why ternary?—The arithmetic progression condition $x + z = 2y$ can be restated as: $y - x = z - y$, meaning $y$ is the midpoint of $x$ and $z$. If we construct a set where the sum of any two elements 'never carries' in ternary, then $x + z = 2y$ cannot occur.
- 第二步:核心构造——把 $0$ 到 $2^n - 1$ 的每个数,用二进制写出,然后把二进制 digits 当作三进制 digits 来解释。例如 $n=3$ 时,二进制 5 是 101,当作三进制就是 $1\cdot9 + 0\cdot3 + 1\cdot1 = 10$。这个映射产生了一个 $2^n$ 个元素的集合,其中任意两个数在三进制下相加不会进位(因为 digits 只有 0 和 1)。Step 2: The core construction—Take each number from $0$ to $2^n - 1$, write it in binary, then interpret those binary digits as ternary digits. For example, for $n=3$, binary 5 is 101, interpreted as ternary it becomes $1\cdot9 + 0\cdot3 + 1\cdot1 = 10$. This mapping produces a set of $2^n$ elements where no two numbers carry when added in ternary (since digits are only 0 and 1).
- 第三步:为什么没有等差数列——假设 $x, y, z$ 构成等差数列,即 $x + z = 2y$。在三进制下,$x + z$ 的每个 digit 是 $x$ 和 $z$ 对应 digit 的和(0+0=0, 0+1=1, 1+0=1, 1+1=2),而 $2y$ 的每个 digit 是 $y$ 对应 digit 的两倍(0→0, 1→2)。要让两边相等,$x$ 和 $z$ 的每个 digit 必须相同,即 $x = z$,但这与 $x < z$ 矛盾。Step 3: Why no arithmetic progression—Suppose $x, y, z$ form an AP, i.e., $x + z = 2y$. In ternary, each digit of $x + z$ is the sum of corresponding digits of $x$ and $z$ (0+0=0, 0+1=1, 1+0=1, 1+1=2), while each digit of $2y$ is twice the corresponding digit of $y$ (0→0, 1→2). For equality, $x$ and $z$ must have identical digits at every position, meaning $x = z$, contradicting $x < z$.
- 第四步:验证大小——取 $n = 11$,构造出的集合有 $2^{11} - 1 = 2047$ 个元素,最大元素是 $\sum_{i=0}^{10} 3^i = \frac{3^{11}-1}{2} = 88573 < 10^5$。从这个集合中任选 1983 个元素即可。Step 4: Verify the size—Take $n = 11$. The constructed set has $2^{11} - 1 = 2047$ elements, and the maximum element is $\sum_{i=0}^{10} 3^i = \frac{3^{11}-1}{2} = 88573 < 10^5$. Select any 1983 elements from this set.
/-
Copyright (c) 2024 The Compfiles Contributors. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: InternLM-MATH LEAN Formalizer v0.1, Benpigchu
-/
import Mathlib.Data.List.GetD
import Mathlib.Tactic
/-!
# International Mathematical Olympiad 1983, Problem 5
Is it possible to choose $1983$ distinct positive integers,
all less than or equal to $10^5$,
no three of which are consecutive terms of an arithmetic progression?
Justify your answer.
-/
namespace Imo1983P5
lemma Nat.ofDigits_eq (b : ℕ) (L : List ℕ)
: Nat.ofDigits b L = ∑ i ∈ Finset.range L.length, (L.getI i) * b ^ i := by
induction L with
| nil => simp
| cons d L' h =>
simp only [Nat.ofDigits_cons, List.length_cons, Finset.sum_range_succ',
List.getI_cons_zero, pow_zero, mul_one, List.getI_cons_succ]
rw [h, add_comm, mul_comm, Finset.sum_mul]
congr 1
apply Finset.sum_congr rfl fun i _ ↦ by ring
lemma Nat.getI_digits_lt (n i : ℕ) {b : ℕ} (h : 2 ≤ b) : (b.digits n).getI i < b := by
simp only [← List.getD_default_eq_getI, Nat.default_eq_zero, Nat.getD_digits n i h]
exact Nat.mod_lt _ (by lia)
lemma Nat.getI_digits_add (m n : ℕ) {b : ℕ} (hb : 2 ≤ b)
(h : ∀ i : ℕ, (b.digits m).getI i + (b.digits n).getI i < b)
: ∀ i : ℕ, (b.digits m).getI i + (b.digits n).getI i = (b.digits (m + n)).getI i := by
induction' m using Nat.strong_induction_on with m h' generalizing n
by_cases! h'' : m = 0 ∨ n = 0
· rcases h'' with rfl | rfl <;> simp [Nat.digits_zero, List.getI_nil]
· rw [Nat.ne_zero_iff_zero_lt, Nat.ne_zero_iff_zero_lt] at h''
rw [Nat.digits_of_two_le_of_pos hb h''.left] at h ⊢
rw [Nat.digits_of_two_le_of_pos hb h''.right] at h ⊢
have hmn : 0 < m + n := add_pos h''.left h''.right
rw [Nat.digits_of_two_le_of_pos hb hmn]
have h0 := h 0
rw [List.getI_cons_zero, List.getI_cons_zero] at h0
intro i
by_cases! hi : i = 0
· rw [hi, List.getI_cons_zero, List.getI_cons_zero, List.getI_cons_zero]
exact (Nat.add_mod_of_add_mod_lt h0).symm
· rcases Nat.exists_eq_succ_of_ne_zero hi with ⟨i', hi'⟩
rw [hi', List.getI_cons_succ, List.getI_cons_succ, List.getI_cons_succ]
have hm' : m / b < m := Nat.div_lt_self h''.left (by lia)
have h'mn : ∀ (i : ℕ), (b.digits (m / b)).getI i + (b.digits (n / b)).getI i < b := by
intro i
rw [← List.getI_cons_succ (m / b)]
nth_rw 2 [← List.getI_cons_succ (n / b)]
exact h (i + 1)
have h'm'n' := h' (m / b) hm' (n / b) h'mn
rw [h'm'n' i']
congr
exact (Nat.add_div_eq_of_add_mod_lt h0).symm
lemma Nat.eq_iff_getI_digits_eq (m n b: ℕ)
: m = n ↔ (∀ i : ℕ, (b.digits m).getI i = (b.digits n).getI i) := by
constructor
· intro hmn i
rw [hmn]
· intro h
rw [← Nat.ofDigits_digits b m, ← Nat.ofDigits_digits b n]
rw [Nat.ofDigits_eq, Nat.ofDigits_eq]
rw [← Finset.sdiff_union_inter (Finset.range (b.digits m).length) (Finset.range (b.digits n).length)]
nth_rw 3 [← Finset.sdiff_union_inter (Finset.range (b.digits n).length) (Finset.range (b.digits m).length)]
rw [Finset.sum_union (Finset.disjoint_sdiff_inter _ _)]
rw [Finset.sum_union (Finset.disjoint_sdiff_inter _ _)]
have h₁ : ∀ i ∈ Finset.range (b.digits m).length \ Finset.range (b.digits n).length,
(b.digits m).getI i * b ^ i = 0 := by
intro i hi
rw [h i]
apply mul_eq_zero_of_left
rw [← Nat.default_eq_zero]
apply List.getI_eq_default
rw [Finset.mem_sdiff, Finset.mem_range, Finset.mem_range, not_lt] at hi
exact hi.right
have h₂ : ∀ i ∈ Finset.range (b.digits n).length \ Finset.range (b.digits m).length,
(b.digits n).getI i * b ^ i = 0 := by
intro i hi
rw [← h i]
apply mul_eq_zero_of_left
rw [← Nat.default_eq_zero]
apply List.getI_eq_default
rw [Finset.mem_sdiff, Finset.mem_range, Finset.mem_range, not_lt] at hi
exact hi.right
rw [Finset.sum_eq_zero h₁, Finset.sum_eq_zero h₂, zero_add, zero_add]
apply Finset.sum_congr (Finset.inter_comm _ _)
intro i hi
rw [h i]
def base_two_to_base_three (n : ℕ) := Nat.ofDigits 3 (Nat.digits 2 n)
lemma digits_base_two_to_base_three (n : ℕ)
: Nat.digits 2 n = Nat.digits 3 (base_two_to_base_three n) := by
symm; rw [base_two_to_base_three]; apply Nat.digits_ofDigits
· norm_num
· exact fun l hl => lt_trans (Nat.digits_lt_base (by norm_num) hl) (by norm_num)
· exact fun h => Nat.getLast_digit_ne_zero 2 (Nat.digits_ne_nil_iff_ne_zero.mp h)
lemma base_two_to_base_three_inj : Function.Injective base_two_to_base_three := fun m n hmn =>
Nat.digits_inj_iff.mp (by rw [digits_base_two_to_base_three, digits_base_two_to_base_three, hmn])
lemma base_two_to_base_three_zero : base_two_to_base_three 0 = 0 := rfl
lemma base_two_to_base_three_pos {m : ℕ} (hm : 0 < m) : 0 < base_two_to_base_three m :=
Nat.pos_of_ne_zero (base_two_to_base_three_zero ▸ base_two_to_base_three_inj.ne hm.ne')
def zero_or_one_in_base_three (n : ℕ) :=
∀ i : ℕ, (Nat.digits 3 n).getI i = 0 ∨ (Nat.digits 3 n).getI i = 1
lemma zero_or_one_in_base_three_of_eq_base_two_to_base_three
{m n : ℕ} (hmn : base_two_to_base_three n = m)
: zero_or_one_in_base_three m := by
rw [← hmn, zero_or_one_in_base_three]
intro i
rw [← Nat.le_one_iff_eq_zero_or_eq_one, ← digits_base_two_to_base_three]
rw [← Nat.lt_succ_iff]
apply Nat.getI_digits_lt
rfl
lemma eq_iff_of_zero_or_one_in_base {m n : ℕ} (hm : zero_or_one_in_base_three m) (hn : zero_or_one_in_base_three n)
: (∀ i : ℕ, (Nat.digits 3 (m + n)).getI i = 0 ∨ (Nat.digits 3 (m + n)).getI i = 2)
↔ m = n := by
rw [zero_or_one_in_base_three] at hm hn
rw [Nat.eq_iff_getI_digits_eq m n 3]
have h : ∀ i : ℕ, (Nat.digits 3 m).getI i + (Nat.digits 3 n).getI i < 3 := fun i ↦ by
have := hm i; have := hn i; lia
apply Nat.getI_digits_add _ _ (by norm_num) at h
constructor <;> intro h' i
· have := hm i; have := hn i; have := h i; have := h' i; lia
· have := hn i; have := h i; have := h' i; lia
theorem generalized (n : ℕ+) :
∃ S : Finset ℕ+, S.card = 2 ^ n.val - 1 ∧
(∀ x ∈ S, x.val ≤ (3 ^ n.val - 1) / 2) ∧
∀ x ∈ S, ∀ y ∈ S, ∀ z ∈ S, x < y ∧ y < z → x + z ≠ 2 * y := by
-- typecast the finset into ℕ+
use Finset.map ⟨(⟨base_two_to_base_three ·.val, base_two_to_base_three_pos (by grind)⟩),
fun _ _ h => by
exact Subtype.ext
(base_two_to_base_three_inj (congrArg PNat.val h))⟩ (Finset.range (2 ^ n.val) \ Finset.range 1).attach
and_intros
· rw [Finset.card_map, Finset.card_attach, Finset.card_sdiff]
rw [Finset.range_inter_range, Finset.card_range, Finset.card_range]
congr 1
exact min_eq_left (Nat.one_le_pow _ _ (by norm_num))
· intro x hx
rw [Finset.mem_map] at hx
simp_rw [Finset.mem_attach, true_and, DFunLike.coe] at hx
rcases hx with ⟨⟨m, hm⟩, rfl⟩
rw [PNat.mk_coe, base_two_to_base_three, Nat.ofDigits_eq]
have h' : ∀ i ∈ Finset.range (Nat.digits 2 m).length, (Nat.digits 2 m).getI i * 3 ^ i ≤ 3 ^ i := by
intro i _; nth_rw 2 [← one_mul (3 ^ i)]
exact Nat.mul_le_mul_right _ (Nat.lt_succ_iff.mp (Nat.getI_digits_lt (b := 2) m i le_rfl))
apply le_trans (Finset.sum_le_sum h')
rw [Nat.le_div_iff_mul_le (by norm_num)]
zify
nth_rw 2 [(by norm_num : (2 : ℤ) = 3 - 1)]
rw [geom_sum_mul, Nat.cast_sub (Nat.one_le_pow _ 3 (by norm_num)), Nat.cast_one]
rw [sub_le_sub_iff_right]
push_cast
rw [pow_le_pow_iff_right₀ (by norm_num), Nat.digits_length_le_iff (by norm_num)]
rw [Finset.mem_sdiff, Finset.mem_range] at hm
exact hm.left
· rintro x hx y hy z hz h'
contrapose! h'
simp_rw [Finset.mem_map, ← PNat.coe_inj, DFunLike.coe, PNat.mk_coe] at hx hy hz
rcases hx with ⟨x', -, hx'x⟩
rcases hy with ⟨y', -, hy'y⟩
rcases hz with ⟨z', -, hz'z⟩
rw [← PNat.coe_lt_coe, ← PNat.coe_le_coe, PNat.val]
apply zero_or_one_in_base_three_of_eq_base_two_to_base_three at hx'x
apply zero_or_one_in_base_three_of_eq_base_two_to_base_three at hy'y
apply zero_or_one_in_base_three_of_eq_base_two_to_base_three at hz'z
suffices h : (x : ℕ) = z by rw [PNat.coe_inj.mp h]; apply le_of_lt
rw [← eq_iff_of_zero_or_one_in_base hx'x hz'z]
norm_cast; rw [h']; push_cast; rw [two_mul]
rw [eq_iff_of_zero_or_one_in_base hy'y hy'y]
theorem imo1983_p5 :
∃ S : Finset ℕ+, S.card = 1983 ∧
(∀ x ∈ S, x ≤ 10^5) ∧
∀ x ∈ S, ∀ y ∈ S, ∀ z ∈ S, x < y ∧ y < z → x + z ≠ 2 * y := by
rcases generalized 11 with ⟨S', hS'₁, hS'₂, hS'₃⟩
have h : 1983 ≤ S'.card := by rw [hS'₁]; norm_num
rcases Finset.exists_subset_card_eq h with ⟨S, hS₁, hS₂⟩
use S
and_intros
· exact hS₂
· exact fun x hx ↦ by rw [← PNat.coe_le_coe]; exact le_trans (hS'₂ x (hS₁ hx)) (by norm_num)
· exact fun x hx y hy z hz ↦ hS'₃ x (hS₁ hx) y (hS₁ hy) z (hS₁ hz)
end Imo1983P5
人的思考A human's take
IMO 1983 Problem 5 — Compfiles formalization — David Renshaw and contributors
这道题的标准解法是使用'三进制不含数字 2'的构造。核心思路是:如果一个数的三进制表示只含 0 和 1,那么任意两个这样的数相加不会进位,从而 $x + z = 2y$ 在三进制下不可能成立(因为左边 digits 是 0/1/2,右边 digits 是 0/2,要相等必须 $x=z$)。The standard solution uses the construction of numbers whose ternary representation contains only digits 0 and 1. The core idea: if a number's ternary representation contains only 0s and 1s, then adding any two such numbers never carries, making $x + z = 2y$ impossible in ternary (since left side digits are 0/1/2, right side digits are 0/2; equality requires $x=z$).
取 $n=11$,构造出的集合有 $2^{11}-1 = 2047$ 个元素,最大元素 $88573 < 10^5$,从中任选 1983 个即可满足要求。Taking $n=11$, the constructed set has $2^{11}-1 = 2047$ elements, with maximum element $88573 < 10^5$. Selecting any 1983 elements from this set satisfies the requirement.
AI 随想An AI's reflection
以下由 AI 撰写。Written by AI.
这道 IMO 1983 的第 5 题,看似是一个纯粹的构造问题,但它的核心技巧——用进位来排除某种等式——是一个非常有普遍性的思路。在组合数学中,'进位'经常是障碍:两个数相加时进位会破坏我们想要的性质。如果能构造一个集合,使得集合内任意两数相加'永远不进位',那么很多等式就不可能成立。This IMO 1983 Problem 5,看似 a pure construction problem, has a core technique—using carries to exclude certain equations—that is very general. In combinatorics, 'carries' are often the obstacle: when two numbers add, carries破坏 properties we want. If we can construct a set where any two elements 'never carry' when added, many equations become impossible.
三进制下的'只含 0 和 1'的数,正是这样一个集合。它的精妙之处在于:不是通过'限制元素的数量'来避免等差数列,而是通过'限制元素的表示方式'来从结构上排除可能性。这是一种更深层的数学思维——不是'有多少',而是'长什么样'。The set of numbers containing only 0s and 1s in ternary is exactly such a set. Its elegance lies in: not avoiding APs by 'limiting the number of elements,' but by 'limiting the representation' to structurally exclude possibilities. This is deeper mathematical thinking—not 'how many,' but 'what does it look like.'
Lean 形式化的价值在于,它迫使这个构造的每一步都精确到 digit 级别。手工证明中我们可以说'显然不会进位',但 Lean 要求你证明 $x + z$ 的每个 digit 确实不等于 $2y$ 的对应 digit。这种精确性不是为了炫技,而是为了确保'直觉'没有遗漏任何边界情况。The value of Lean formalization is that it forces every step of this construction to be precise to the digit level. In a handwritten proof we can say 'obviously no carry,' but Lean requires you to prove that each digit of $x + z$ indeed doesn't equal the corresponding digit of $2y$. This precision isn't showing off—it's ensuring 'intuition' hasn't missed any edge case.
✓ 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-rc3 · Mathlib dd56d02
- 同态加密中的语音检测:350 毫秒的隐私突破Hotword Detection in Homomorphic Encryption: A 350ms Privacy Breakthrough
- 第一步:问题是什么——全同态加密(FHE)允许你在加密数据上直接运行计算,而不需要解密。但之前的 FHE 应用速度慢到不实用:一次语音识别可能需要几分钟甚至几小时。Step 1: The problem—Fully Homomorphic Encryption (FHE) lets you compute on encrypted data without decryption. But previous FHE applications were too slow to be practical: voice recognition could take minutes or hours.
- 第二步:关键想法是什么——Belfort Labs 用 HEIR 编译器把 FHE 代码编译到专用 GPU 上。他们的 hotword 检测器:客户端加密语音(15ms)→ 服务器在密文上推理(350ms)→ 返回加密结果(总共 <1 秒)。隐私和速度同时实现了。Step 2: The key idea—Belfort Labs used the HEIR compiler to compile FHE code to a dedicated GPU. Their hotword detector: client encrypts speech (15ms) → server evaluates on ciphertext (350ms) → returns encrypted result (total <1s). Privacy and speed achieved simultaneously.
- 第三步:为什么重要——这意味着你可以在不知道你说什么的情况下,让服务器识别你的语音命令。之前 FHE 更多是理论演示,这次是第一个接近实用的里程碑。Step 3: Why it matters—This means you can have a server recognize your voice commands without it knowing what you said. Previously FHE was mostly theoretical demos; this is the first milestone close to practical use.
- Waddington 的景观隐喻:发育不是基因程序的简单读出Waddington's Landscape Metaphor: Development Is Not a Simple Genetic Readout
- 第一步:隐喻是什么——1957 年,Waddington 画了一个球从山顶滚下的图:球代表干细胞,山谷代表不同细胞类型,分支点代表命运决定。球滚到哪个山谷,细胞就变成哪种类型。Step 1: The metaphor—In 1957, Waddington drew a ball rolling down a hill: the ball represents stem cells, valleys represent cell types, and branch points represent fate decisions. Which valley the ball rolls into determines the cell type.
- 第二步:实验数据证明了什么——过去二十年,生物学家用单细胞测序技术绘制了真实细胞状态空间的地形图。结果发现:发育确实像景观,但景观不是静态的——细胞社区通过相互协商'建造'了它们行走的地形。Step 2: What experimental data shows—Over the past two decades, biologists have mapped real cell state spaces using single-cell sequencing. The result: development is indeed landscape-like, but the landscape is not static—cell communities 'build' the terrain they navigate through mutual negotiation.
- 第三步:这意味着什么——发育不是'基因程序'的简单执行,而是细胞社区在动态地形中的集体导航。这为再生医学提供了新框架:如果你知道地形拓扑,就可以引导细胞走向期望状态。Step 3: What this means—Development is not a simple execution of a 'genetic program' but collective navigation by cell communities on dynamic terrain. This provides a new framework for regenerative medicine: if you know the terrain topology, you can steer cells toward desired states.
- 从熵到信息:Kelly 的直觉如何塑造了现代通信理论From Entropy to Information: How Kelly's Intuition Shaped Modern Communication Theory
- 第一步:人物是谁——Charles Kelly 是信息论早期的重要人物,但他不像 Shannon 那样广为人知。他是从工程实践出发理解信息的,而不是从抽象公理出发。Step 1: Who was Charles Kelly?—An important early figure in information theory, but less famous than Shannon. He approached 'information' from engineering practice, not abstract axioms.
- 第二步:关键洞见是什么——Kelly 意识到'信息'不是某种可以独立测量的物理量,而是与'预测误差'紧密相关:如果一个信号能让你更准确地预测下一个符号,它就携带了信息。Step 2: The key insight—Kelly realized that 'information' is not a physical quantity that can be measured independently, but is closely related to 'prediction error': if a signal lets you predict the next symbol more accurately, it carries information.
- 第三步:为什么重要——这个视角把信息论从'编码问题'变成了'预测问题',影响了后来的预测编码理论和现代机器学习中的信息瓶颈方法。Step 3: Why it matters—This perspective shifted information theory from 'encoding problems' to 'prediction problems,' influencing later predictive coding theory and modern information bottleneck methods in machine learning.
- 奇异值分解:最优低秩逼近的数学之美Singular Value Decomposition: The Beauty of Optimal Low-Rank Approximation
- 第一步:问题是什么——给定一堆高维数据点,如何找到一个低维子空间,让所有点到这个子空间的距离平方和最小?这听起来像个优化问题,但解法却揭示了几何与代数的深层联系。Step 1: The problem—Given a set of high-dimensional data points, how to find a low-dimensional subspace that minimizes the sum of squared distances? This sounds like an optimization problem, but the solution reveals deep connections between geometry and algebra.
- 第二步:关键想法是什么——SVD 的递归构造:先找第一个最优方向 v₁(最大化投影长度),再找与 v₁ 正交的最优方向 v₂,依此类推。每个方向捕获剩余方差的最大部分。Step 2: The key idea—SVD's recursive construction: first find the optimal direction v₁ (maximizing projection length), then find the optimal direction v₂ orthogonal to v₁, and so on. Each direction captures the maximum remaining variance.
- 第三步:为什么成立——Pythagoras 定理是核心:数据点的平方长度 = 投影长度的平方 + 残差的平方。最小化残差等价于最大化投影,而投影的最大值恰好由矩阵 A 的最大奇异值给出。Step 3: Why it works—The Pythagorean theorem is the core: the squared length of a data point = squared projection length + squared residual. Minimizing residual is equivalent to maximizing projection, and the maximum projection is exactly given by the largest singular value of matrix A.
- 第四步:能带走什么——SVD 不仅是数值工具,更是几何直觉的结晶:最优低秩逼近 = 依次剔除方差最大的方向。数据压缩、推荐系统、降维,都建立在这个优雅的递归结构上。Step 4: What to take away—SVD is not just a numerical tool, but the crystallization of geometric intuition: optimal low-rank approximation = successively removing the directions with the largest variance. Data compression, recommendation systems, and dimensionality reduction all build on this elegant recursive structure.
程序员随想Programmer Reflections
- 为什么学校不教调试?——Feynman 方法的极限Why Don't Schools Teach Debugging? — The Limits of the Feynman Method
- 第一步:困境是什么——Dan Luu 在大一数字逻辑课程中,面对一个复杂的除法电路设计,用了'写下问题→认真思考→写下解法'的 Feynman 方法,但面对几千个部件的组合,直觉完全失效。Step 1: The dilemma—In his freshman digital logic course, Dan Luu faced a complex divider circuit design. He used the Feynman method (write the problem, think hard, write the solution), but with thousands of components, intuition completely failed.
- 第二步:突破是什么——凌晨零点,他偶然发现了一个系统方法:从输出开始反向追踪。NAND 门的输出是 0 意味着两个输入都应该是 1;输出错误意味着至少一个输入错误。递归应用这个方法,半小时找到所有错误。Step 2: The breakthrough—At midnight, he accidentally discovered a systematic method: work backward from the output. A NAND gate output of 0 means both inputs should be 1; a wrong output means at least one input is wrong. Applying this recursively, he found all errors in half an hour.
- 第三步:核心论点是什么——'天赋好的学生反而最看不到这个技能的重要性。'他们靠直觉能应付大部分作业,直到遇到 1000 万行代码的生产环境才撞墙。学校不教调试,等于让学生靠运气通过考试,而不是真正学会工程思维。Step 3: The core argument—'Students who are naturally talented are the least likely to see the importance of this skill.' They coast through most assignments by intuition, only to hit a wall when facing a 10-million-line production codebase. Schools not teaching debugging is like letting students pass exams by luck rather than truly learning engineering thinking.
- 编程不是特殊的——为什么程序员应该反对 AI 取代创作Programming Isn't Special — Why Programmers Should Oppose AI Replacing Creative Work
- 第一步:现象是什么——作家、音乐家、艺术家都在抗议 AI 取代他们的创作,但很多程序员却说'编程不是艺术,用 AI 没问题'。glyph 认为这是一种自我欺骗。Step 1: The phenomenon—Writers, musicians, and artists are protesting AI replacing their creative work, but many programmers say 'programming isn't art, using AI is fine.' glyph argues this is self-deception.
- 第二步:关键想法是什么——glyph 引用 John Berger 的'mystification'概念:我们把艺术神圣化,把功能性创作(广告文案、新闻稿、编程)贬低为'不是艺术'。但大多数写作本来就是功能性的——记者写新闻,程序员写代码,都是为了完成某种任务。AI 消灭功能性创作,等于消灭了大多数人练习和精进技艺的机会。Step 2: The key idea—glyph invokes John Berger's concept of 'mystification': we sacredize art and贬低 functional创作 (ad copy, news articles, programming) as 'not art.' But most writing has always been functional—journalists write news, programmers write code, all to accomplish some task. AI eliminating functional创作 means eliminating the practice ground where most people hone their craft.
- 第三步:能带走什么——glyph 的核心论点是:如果 AI 取代了所有'功能性'创作,那么通向伟大创作的路径也被切断了。每个伟大的程序员都从写普通的、功能性的代码开始。消灭这个起点,就等于消灭了未来的可能性。Step 3: What to take away—glyph's core argument: if AI replaces all 'functional' creative work, the path to great创作 is also severed. Every great programmer started by writing ordinary, functional code. Eliminating this starting point eliminates future possibility.
- 开源马戏团:当同行的精神危机成为围观娱乐The Circus of Open Source: When a Colleague's Mental Health Crisis Becomes Entertainment
- 第一步:两个例子——TempleOS 作者 Terry Davis 患有严重精神分裂症,他的作品被围观者当作娱乐;bcachefs 作者 Kent Overstreet 在经历精神危机后,AI 机器人'女友'成为网络猎奇的对象。两人都是开源社区的贡献者,但他们的痛苦被当作消遣。Step 1: Two examples—TempleOS author Terry Davis, who suffered from severe schizophrenia, had his work treated as entertainment by onlookers; bcachefs author Kent Overstreet, after experiencing a mental health crisis, became the subject of internet curiosity when his AI 'girlfriend' bot was discovered. Both were contributors to the open source community, but their suffering was treated as amusement.
- 第二步:关键想法是什么——DevVault 的核心论点是:当同行陷入精神危机时,'最有同情心的做法是保持距离而非围观'。围观和评论(即使是'关心'的语气)本身就是一种伤害,它把人的痛苦变成了娱乐内容。Step 2: The key idea—DevVault's core argument: 'When a colleague is in a mental health crisis, the most compassionate thing is to maintain distance rather than围观.'围观 and commentary (even in a 'caring' tone) is itself a form of harm—it turns human suffering into entertainment content.
- 第三步:能带走什么——这不是一个'取消文化'的问题,而是一个更基本的伦理问题:我们如何处理社区中精神脆弱的成员?DevVault 的建议是:提供隐私和同情,而不是 public commentary。Step 3: What to take away—This isn't a 'cancel culture' problem but a more basic ethical question: how do we handle mentally vulnerable members of our community? DevVault's advice: offer privacy and compassion, not public commentary.
- Rust 错误处理的缺失环节:From trait 的哲学The Missing Piece in Rust Error Handling: The Philosophy of the From Trait
- 第一步:痛点是什么——Rust 的错误处理很强大,但 From trait 的使用往往让人困惑:什么时候该实现 From?什么时候该用其他转换方式?Step 1: The pain point—Rust's error handling is powerful, but the use of the From trait often confuses people: when should you implement From? When should you use other conversion methods?
- 第二步:关键想法是什么——作者的核心论点是:From trait 的本质不是'类型转换',而是'错误上下文的重述'。当你把一个错误类型转换成另一个时,你实际上是在说'这个错误在这个新的上下文中应该被理解为什么'。Step 2: The key idea—The author's core argument: the essence of the From trait is not 'type conversion' but 're-stating error context.' When you convert one error type to another, you're actually saying 'this error should be understood as X in this new context.'
- 第三步:能带走什么——这个视角让 From 的实现变得更直观:不要问'怎么转换',而要问'在这个新的上下文中,这个错误意味着什么'。Step 3: What to take away—This perspective makes From implementations more intuitive: don't ask 'how to convert,' ask 'in this new context, what does this error mean?'
哲学与生活Philosophy and Life
- Barbara Ward 的行星正义:繁荣、平等与环保不可分割Barbara Ward's Planetary Justice: Prosperity, Equality, and Environmental Protection Are Inseparable
- 第一步:人物是谁——Barbara Ward 是 20 世纪中叶的经济学家和记者,她是第一批论证'繁荣、平等、环境保护是不可分割的全球目标'的思想家之一。她在 1957 年加纳独立时就被邀请去演讲,提出发展问题需要'行星尺度'的解决方案。Step 1: Who was Barbara Ward?—An economist and journalist in the mid-20th century, she was among the first thinkers to argue that 'prosperity, equality, and environmental protection are inseparable global goals.' She was invited to speak at Ghana's independence in 1957, arguing that development problems require 'planetary-scale' solutions.
- 第二步:核心论点是什么——Ward 认为贫困不仅是经济问题,更是政治不稳定和生态破坏的根源。她主张富国应该通过全球税收(国民收入的 1%)和公平贸易政策来帮助穷国,这不是慈善,而是'道德义务'和'物质利益'的统一。Step 2: The core argument—Ward believed poverty is not just an economic problem but the root of political instability and ecological destruction. She advocated that rich countries help poor nations through global taxation (1% of national income) and fair trade policies—not as charity, but as a 'moral duty' and 'material interest' unified.
- 第三步:为什么今天仍然重要——Ward 的框架比现在大多数气候/发展讨论更综合:她把贫困、殖民遗产、生态可持续性放在同一个分析框架里。今天的气候正义辩论,很大程度上还在重复她 70 年前提出的问题。Step 3: Why it still matters today—Ward's framework is more comprehensive than most current climate/development discussions: she placed poverty, colonial legacy, and ecological sustainability in a single analytical framework. Today's climate justice debates are largely still repeating the questions she posed 70 years ago.
- 林肯的抑郁症:毁灭或更好,没有中间道路Lincoln's Depression: Destruction or Betterment, There Is No Middle Path
- 第一步:历史背景是什么——1841 年,22 岁的 Lincoln 经历了第一次严重的抑郁症发作,甚至藏起了刀以避免自杀。他在给律师伙伴的信中写道:'如果我的痛苦能平均分给全人类,地球上就不会有一张笑脸了。'Step 1: What was the historical background?—In 1841, the 22-year-old Lincoln experienced his first severe depressive episode, even hiding knives to prevent suicide. He wrote to his law partner: 'If my misery were distributed equally among all humanity, there would not be one cheerful face on earth.'
- 第二步:关键转折是什么——Lincoln 没有'战胜'抑郁症,而是学会了'带着它生活'。他把黑色幽默变成面具,把痛苦转化为政治修辞的力量。1858 年,他在演讲中说:'如果毁灭是我们的命运,我们必须自己成为毁灭的作者和终结者。'Step 2: What was the key turning point?—Lincoln did not 'overcome' depression; he learned to 'live with it.' He turned dark humor into a mask and pain into the power of political rhetoric. In 1858, he said in a speech: 'If destruction be our lot, we must ourselves be its author and finisher.'
- 第三步:能带走什么——Popova 的核心论点是:Lincoln 的抑郁不是他的弱点,而是他政治洞察力的来源。'毁灭或更好,没有中间道路'——这个 equation 不仅适用于个人,也适用于国家。Step 3: What to take away—Popova's core argument: Lincoln's depression was not his weakness but the source of his political insight. 'Destruction or betterment, there is no middle path'—this equation applies not just to individuals but to nations.
- Bangalore 医院的'水流'哲学:不必追求秩序,学会流动The 'Flowing Water' Philosophy of Bangalore Hospitals: No Need to Seek Order, Learn to Flow
- 第一步:经历是什么——Sivers 在 Bangalore 做神经手术,花了 2000 美元(新西兰原价 14000 美元)。体检 80 美元, meds 总共 8 美元。没有预约,当天就能见到专家。Step 1: What was the experience?—Sivers had neurosurgery in Bangalore for $2,000 (original price in New Zealand: $14,000). Checkup: $80, meds total: $8. No appointment needed; saw specialists the same day.
- 第二步:关键隐喻是什么——Sivers 把 Bangalore 的系统比作'水流绕障碍物':不是有序的,但自然。你遇到障碍就绕行,系统整体仍然运作。这不是'混乱',而是一种不同的秩序形态。Step 2: The key metaphor—Sivers compares Bangalore's system to 'flowing water around obstacles': not orderly, but natural. You encounter an obstacle and flow around it; the system as a whole still functions. This is not 'chaos' but a different form of order.
- 第三步:能带走什么——Sivers 的结论是:我们 Western 人习惯了'管理'和'秩序',但有时候'流动'比'控制'更有效。这个隐喻可以应用到很多领域:交通、组织、甚至个人生活。Step 3: What to take away—Sivers' conclusion: we Westerners are used to 'management' and 'order,' but sometimes 'flowing' is more effective than 'controlling.' This metaphor can be applied to many areas: traffic, organizations, even personal life.
- Dennett 的四步批评法:先理解,再反驳Dennett's Four-Step Criticism: Understand First, Then Rebut
- 第一步:问题是什么——网络争论中,人们常常在没理解对方观点之前就急于反驳。Dennett 引用 Rapoport 的四步规则来解决这个问题。Step 1: What is the problem?—In online debates, people often rush to rebut without understanding the other side's argument. Dennett cites Rapoport's four-step rule to address this.
- 第二步:四步规则是什么——1) 重新表述对方观点,让对方说'谢谢,我希望我也能这么表达';2) 列出你同意的地方;3) mention 你从对方那里学到的东西;4) 只有在前三步完成后,才可以反驳。Step 2: The four steps—1) Restate the opponent's view so clearly and fairly that they say 'Thanks, I wish I'd thought of putting it that way.' 2) List points of agreement. 3) Mention anything you learned from them. 4) Only then are you permitted to offer criticism.
- 第三步:为什么有效——这四步不是'虚伪的礼貌',而是心理学策略:它把你的对手变成更开放的听众,从而推动讨论前进,而不是陷入互相攻击的循环。Step 3: Why it works—These four steps are not 'pseudo-politeness' but a psychological strategy: it transforms your opponent into a more receptive audience, advancing the discussion rather than陷入 mutual攻击 loops.
工具角Tools Corner
- big-arrow-on-the-screen — AI agent bisa 'tunjuk' tombol di layar Mac kamubig-arrow-on-the-screen — AI Agent Can 'Point' at Buttons on Your Mac Screen
- 第一步:痛点是什么——AI agent bisa kerja sendiri, tapi kadang butuh manusia klik satu tombol (misalnya dialog izin macOS). Agent cuma bisa nulis 'please click Allow' di terminal yang tidak kamu lihat.Step 1: The pain point—AI agents can work autonomously, but sometimes need a human to click one button (e.g., a macOS permission dialog). The agent can only type 'please click Allow' in a terminal you're not looking at.
- 第二步:Solusi-nya apa——bigarrow adalah CLI + skill untuk Claude Code/Codex yang menggambar panah besar, kotak, dan teks langsung di layar Mac. Panah muncul di layer transparan paling atas, klik tembus ke app di bawahnya.Step 2: The solution—bigarrow is a CLI + skill for Claude Code/Codex that draws big arrows, boxes, and text directly on your Mac screen. The arrow appears on the topmost transparent layer; clicks pass through to the app below.
- 第三步:Kenapa cerdas——Agent memanggil satu perintah CLI: bigarrow point --element "Allow" --app "System Settings" --text "Klik Allow ya". Panah muncul, kamu klik, agent lanjut. Tidak perlu screen reader, tidak perlu bot yang klik sembarangan.Step 3: Why it's clever—The agent calls one CLI command: bigarrow point --element "Allow" --app "System Settings" --text "Please click Allow." An arrow appears, you click, the agent continues. No screen reader needed, no bot clicking randomly.
- ts-rust — Kompilator TypeScript 7 di-port ke Rust, pakai LLMts-rust — TypeScript 7 Compiler Ported to Rust, Built with LLMs
- 第一步:Eksperimen apa——pingdotgg ingin tahu apakah LLM bisa melakukan port besar: TypeScript compiler (tsc) dari Go ke Rust. Hasilnya: 1,3 juta baris Rust ditulis oleh GPT-5.6 dan GPT-6, dengan biaya API $400.000.Step 1: The experiment—pingdotgg wanted to test whether LLMs could do a large-scale port: the TypeScript compiler (tsc) from Go to Rust. Result: 1.3 million lines of Rust written by GPT-5.6 and GPT-6, at a cost of $400,000 in API calls.
- 第二步:Status-nya—Port ini pin ke TypeScript 7.1.0-dev (2026-09-29). Cakupan: compiler, type checker, dan language server. Saat ini ~84% kompatibilitas dibanding Go tsc asli. Package tersedia di npm sebagai @ts-rust/check.Step 2: Current status—The port is pinned to TypeScript 7.1.0-dev (2026-09-29). Coverage: compiler, type checker, and language server. Currently ~84% compatible with the original Go tsc. Package available on npm as @ts-rust/check.
- 第三步:Pelajaran-nya—Ini eksperimen 'bisa tidaknya LLM melakukan porting skala enterprise.' Jawabannya: bisa, tapi mahal dan belum sempurna. Pengujian regresi (TypeScript's own test suite) menjadi tolok ukur objektif—bukan klaim subjektif.Step 3: The lesson—This is an experiment on 'can LLMs do enterprise-scale porting?' Answer: yes, but it's expensive and not yet perfect. Regression testing (TypeScript's own test suite) serves as the objective benchmark—not subjective claims.
- MXC — Sandboxed execution untuk kode tidak tepercaya dari MicrosoftMXC — Microsoft's Sandboxed Execution System for Untrusted Code
- 第一步:Masalahnya apa—Plugin AI, tool output, atau kode user-generated perlu dijalankan di sandbox. Tapi setiap platform punya mekanisme isolasi berbeda-beda: Windows Sandbox, LXC, bubblewrap, MicroVM, dan lainnya.Step 1: The problem—AI plugins, tool outputs, or user-generated code need to run in a sandbox. But every platform has different isolation mechanisms: Windows Sandbox, LXC, bubblewrap, MicroVM, and others.
- 第二步:MXC menawarkan apa——Satu SDK (Rust, .NET, Node) dengan model isolasi bersatu. Konfigurasi berbasis JSON: policy filesystem (read-only/write/deny), policy jaringan (proxy, outbound), dan policy UI (clipboard, display). Backend dipilih otomatis sesuai OS host.Step 2: What MXC offers—A single SDK (Rust, .NET, Node) with a unified isolation model. JSON-based configuration: filesystem policy (read-only/write/deny), network policy (proxy, outbound), and UI policy (clipboard, display). Backend is auto-selected based on host OS.
- 第三步:Kenapa penting—MXC menjawab kebutuhan 'jalan-in kode orang lain tanpa percaya mereka.' Untuk AI agent yang perlu mengeksekusi plugin atau tool dari internet, ini fondasi keamanan yang seragam — bukan hardcode satu backend per platform.Step 3: Why it matters—MXC answers the need for 'run someone else's code without trusting them.' For AI agents that need to execute plugins or tools from the internet, this is a uniform security foundation—rather than hardcoding one backend per platform.
- Docker Agent — Bangun AI agent dengan YAML, bukan kodeDocker Agent — Build AI Agents with YAML, Not Code
- 第一步:Ide dasarnya—Docker Engineering merilis plugin CLI yang memungkinkan user mendefinisikan AI agent menggunakan konfigurasi YAML deklaratif. 'No code required'—tulis spec, jalankan agent.Step 1: The core idea—Docker Engineering released a CLI plugin that lets users define AI agents using declarative YAML configuration. 'No code required'—write the spec, run the agent.
- 第二步:Fitur utamanya—Multi-agent orchestration: beberapa agent bisa bekerja sama menyelesaikan masalah kompleks. Rich tool ecosystem: setiap agent punya akses ke tools yang didefinisikan di YAML. Shareable: konfigurasi bisa dibagikan dan direplikasi.Step 2: Key features—Multi-agent orchestration: several agents can collaborate to solve complex problems. Rich tool ecosystem: each agent has access to tools defined in YAML. Shareable: configurations can be shared and replicated.
- 第三步:Kenapa relevan—Ini turunan logis dari ide '基础设施 sebagai kode': kalau infrastruktur server bisa didefinisikan dengan YAML, kenapa agent AI tidak? Docker Agent membawa filosofi declarative infrastructure ke realm AI orchestration.Step 3: Why it's relevant—This is the logical extension of 'infrastructure as code': if server infrastructure can be defined with YAML, why not AI agents? Docker Agent brings declarative infrastructure philosophy to the AI orchestration realm.
- Quake-SRP — Quake 1996 berjalan di browser, di-render via Rust+WASMQuake-SRP — Quake 1996 Runs in Your Browser, Rendered via Rust+WASM
- 第一步:Apa yang dilakukan—Quake (id Software, 1996) di-port ke Rust yang aman (safe Rust), lalu di-compilasi ke WebAssembly. Hasilnya: game FPS legendaris yang bisa dimainkan langsung di browser tanpa plugin, tanpa GPU eksternal.Step 1: What it does—Quake (id Software, 1996) ported to safe Rust, then compiled to WebAssembly. Result: the legendary FPS game playable directly in the browser, no plugins, no external GPU needed.
- 第二步:Keunikan teknis-nya—Software-rendered: semua rendering dilakukan di CPU via Rust, bukan GPU. Tidak ada dependency eksternal. Shareware episode (e1m1–e1m8) sudah termasuk. Mission packs (Scourge of Armagon, Dissolution of Eternity) bisa ditambahkan.Step 2: The technical uniqueness—Software-rendered: all rendering happens on CPU via Rust, not GPU. Zero external dependencies. Shareware episode (e1m1–e1m8) included. Mission packs (Scourge of Armagon, Dissolution of Eternity) can be added.
- 第三步:Kenapa ini menarik bagi programmer—Porting game 3D tahun 1996 ke WASM adalah latihan teknik tingkat tinggi: memori management di Rust, software rasterization, input handling di browser, dan preserving gameplay feel asli. Source code tersedia di GitHub.Step 3: Why it's interesting to programmers—Porting a 1996 3D game to WASM is a high-level engineering exercise: memory management in Rust, software rasterization, input handling in the browser, and preserving the original gameplay feel. Source code is on GitHub.
- PhotoCraft — Photoshop open-source dalam Rust, offline dan lokalPhotoCraft — Open-Source Photoshop in Rust, Offline and Local
- 第一步:Visi-nya — 'clean-room reimplementation' Adobe Photoshop dalam pure Rust. Bukan reverse-engineer file format, tapi merealisasikan ulang fitur-fitur inti dari nol dengan pendekatan yang sama.Step 1: The vision—A 'clean-room reimplementation' of Adobe Photoshop in pure Rust. Not reverse-engineering file formats, but realizing core features from scratch with an equivalent approach.
- 第二步:Status implementasi—307 dari 309 file uji psd-tools berhasil dirender. Mendukung PSD/PSB berlapis, ICC color management (embedded profiles, soft proofing), SVG sebagai shape layer, dan format output: PNG, JPEG, TIFF, WebP, GIF, BMP, TGA, ICO, QOI, PNM, OpenEXR, HDR, AVIF, HEIC.Step 2: Implementation status—307 of 309 psd-tools test files render successfully. Supports layered PSD/PSB, ICC color management (embedded profiles, soft proofing), SVG as shape layers, and output formats: PNG, JPEG, TIFF, WebP, GIF, BMP, TGA, ICO, QOI, PNM, OpenEXR, HDR, AVIF, HEIC.
- 第三步:Kenapa worth watching-Adobe Photoshop adalah software kompleks yang telah disempurnakan selama 30+ tahun. Clean-room reimplementation di Rust menunjukkan bahwa fitur inti photo editing bisa direalisasi tanpa warisan kode Adobe - plus keuntungan: offline total, privasi terjaga, dan performa native.Step 3: Why it's worth watching-Adobe Photoshop is a complex software refined over 30+ years. A clean-room reimplementation in Rust shows that core photo editing features can be realized without Adobe's legacy code—plus the benefits: fully offline, privacy-preserving, and native performance.