Page's Diversity Theorem (Diversity Trumps Ability) — rests on an identity: crowd error = average individual error − prediction diversity. Diversity is a term that literally subtracts from collective error, not a feel-good extra. A team of clones of the best performer has zero diversity → group error equals individual error, no aggregation gain. This is exactly why ensemble methods (random forests, bagging) beat single strong models: averaging decorrelated errors cancels them. The diversity must be relevant (different useful models/heuristics) and the problem hard enough. Don't ask only "how able is the team" — ask "will their errors correlate?" Correlated errors mean many people reason like one.
AI Prompts
中文提示词
我要组建一个团队/AI 评审组来解决 [问题],候选成员/视角是 [列表]。请用「群体误差 = 平均误差 − 多样性」分析:
① 哪些成员的错误最可能高度相关(同样的盲区)?
② 为了提升多样性项,我应该补进哪种迥异的背景或模型,哪怕它单项能力略弱?
③ 给出一个 4 人组合,最大化「去相关」而非「平均最强」。
English Prompt
I'm assembling a team / AI review panel to solve [problem]. Candidate members/perspectives: [list]. Using "crowd error = average error − diversity":
1. Which members' errors are most likely to be correlated (shared blind spots)?
2. To raise the diversity term, which distinct background or model should I add, even if its individual ability is slightly lower?
3. Propose a 4-member set that maximizes decorrelation, not "highest average ability."
① 工程:分布式系统的不少利器都是「外行红利」——流行病学的传播模型成了 gossip 协议,控制论的反馈调节成了 backpressure。作为 AI 超级个体,刻意从主场之外进口框架,是你独有的杠杆。② 育儿:你身在局中,反而看不清孩子和你之间的某个动力循环;一个「外人」(老师、另一位家长)一句话点破,正因为他不在这套默契里。给关键难题主动找一个领域外的旁观者,是低成本的高回报动作。
English Summary
Outsider Perspective Dividend — problems that stump insiders are often cracked by people at the margin of, or outside, the field. Not because outsiders are smarter, but because they bring a different default solution library; their tools haven't been pruned by the field's conventions. A problem unsolvable with one field's standard toolkit can be routine with another's. Same root as Page's theorem: insiders' errors are correlated, so the field stalls at the same spot. The dividend is real but noisy — outsiders also propose many naive ideas, so you need a broadcast-and-filter mechanism, not faith in any one outsider. Kin to Zen shoshin (beginner's mind): expertise compresses away "irrelevant" options, but the answer sometimes lives among the discarded ones.
AI Prompts
中文提示词
我在 [领域] 卡住的难题是 [描述],本领域的标准做法都试过了。请扮演外行红利引擎:
① 列出 3 个邻近但不同的领域,它们天天处理与我这个问题结构相似的情形;
② 各给出该领域的标准解法,并把它「翻译」回我的问题;
③ 标出哪一个最可能是被我的领域惯例提前修剪掉的盲区。
English Prompt
I'm stuck on [problem] in [field]; I've tried the field's standard approaches. Act as an outsider-dividend engine:
1. List 3 adjacent-but-different fields that routinely handle problems structurally similar to mine.
2. Give each field's standard solution and translate it back into my problem.
3. Flag which one is most likely a blind spot my field's conventions pruned away early.
① 跨学科:神经科学 × 机器学习杂交出了深度学习;佛学的观照训练 × 临床心理学杂交出了正念减压。这些都不是「外行红利」式的一次借用,而是两条血统的稳定重组。② 个人定位:你自身的稀缺价值,恰恰来自你是一个少见血统的杂交体——AI / 分布式 × 育儿 × 意识与佛学。这正是「近到能整合、远到不冗余」的甜区。提醒:别去拼两种过于相近的领域(如两种 ML 流派),杂不出优势;也别奢望这种组合会自动稳定,每一次都仍需亲手培育。
English Summary
Hybrid Vigor (Heterosis) — crossing two distinct lineages often yields offspring stronger than either parent; inbreeding does the reverse (harmful recessives accumulate). As a metaphor: combining two different "lineages" of ideas/disciplines yields recombinations stronger than either pure line. Distinct from Page's theorem — that's about aggregating predictions (errors cancel); this is generative recombination (new strong variants). Echo chambers are intellectual inbreeding: shared, unchallenged "deleterious recessive" assumptions. The least-discussed key: there's an optimal genetic distance — too close (two flavors of one field) = no vigor; too far with no bridge = incompatibility (outbreeding depression). The dividend lives at "close enough to integrate, far enough to be non-redundant." And vigor shows mainly in the F1 — a brilliant cross doesn't automatically breed true.
AI Prompts
中文提示词
我想在 [主题/项目] 上做一次跨界杂交。我的主场是 [领域 A]。请:
① 推荐 3 条「遗传距离适中」的异系领域(近到能整合、远到不冗余),并排除掉太近或太远的;
② 对每条,指出 A 与它重组后最可能产生的那个强变体是什么;
③ 提醒我每个组合里可能存在的「不亲和」风险点,以及需要补的转译桥。
English Prompt
I want to cross-breed ideas for [topic/project]. My home field is [field A]. Please:
1. Recommend 3 lineages at an optimal "genetic distance" (close enough to integrate, far enough to be non-redundant); exclude those too close or too far.
2. For each, name the strong variant most likely to emerge from recombining it with A.
3. Flag the incompatibility ("outbreeding") risk in each pairing and the translation bridge I'd need to build.
① AI 超级个体:你的杠杆来自一个别人还没下的正确赌注——一套别人还不信的工作流或工具。等所有人都信了,它已是「对 + 共识」,红利消失。② 关键纪律:把它和「身份式唱反调」分开。真正的反共识价值,必须配一个你能说清的具体 edge(你掌握了别人没有的信息,或你的模型在这一点上确实更准),否则你只是落进基础率最高的「错 + 反共识」格。判断自己:我反对,是因为我真懂得更多,还是因为反对让我显得独特?
English Summary
Contrarian Value — being right isn't enough; outsized value lives only where you are both right and non-consensus. If everyone already believes it, the value is priced in (markets) or already built (startups). Map it on a 2×2: right/wrong × consensus/contrarian. Only "right + contrarian" yields edge; "right + consensus" = no edge; "wrong + contrarian" = costly. This is the harvesting condition for the first three models: diversity, outsider views, and hybridization generate non-consensus positions, but only the correct minority captures the payoff. Crucial caveat: most contrarian views are wrong (that's why they're contrarian) — the base rate of "non-consensus AND right" is low. The discipline isn't to oppose for its own sake (identity contrarianism), but to find the rare case where you have a real informational or model edge. Like convex bets: when you're right and alone, you take the whole prize.
AI Prompts
中文提示词
我有一个判断/赌注:[描述]。请用反共识价值矩阵压力测试我:
① 它落在「对/错 × 共识/反共识」哪一格?给出当下的主流共识是什么;
② 如果我宣称是「对 + 反共识」,逼我说清具体 edge:我到底掌握了什么别人没有的信息或更准的模型?
③ 诚实评估:我是不是在为反而反(身份式唱反调)?若是,警示我最可能掉进「错 + 反共识」的地方。
English Prompt
I hold a belief/bet: [describe]. Stress-test me with the contrarian-value matrix:
1. Which cell is it in (right/wrong × consensus/contrarian)? State the current mainstream consensus.
2. If I claim "right + contrarian," force me to name my concrete edge: what information or sharper model do I actually have that others don't?
3. Honestly assess whether I'm being contrarian for identity's sake; if so, warn me where I'm most likely landing in "wrong + contrarian."