① 工程：敏捷估点用的「计划扑克」本质就是一套保护独立性的群体智慧机制——所有人同时翻牌，就是为了不让先开口的资深工程师锚定全场。若改成轮流报数，独立性瞬间崩塌，估值收敛到第一个声音上。② AI：让一个模型对同一问题独立采样多次再投票（self-consistency）能提升准确率，但前提是各次采样去相关；同模型、同 prompt、温度为零地跑 N 遍，N 个答案几乎完全相同，等于只问了一次——没有独立性，就没有群体智慧。
English Summary
Wisdom of Crowds — a crowd's aggregated judgment beats experts when four conditions hold together: diversity, independence, decentralization, and an aggregation mechanism. Independence is the load-bearing wall and the first one reality demolishes: once people see each other's answers, social proof correlates their errors and the crowd collapses into a herd. So "discuss first, then decide" often lowers collective intelligence — deliberation manufactures correlation. The aggregation rule is design, not detail: simultaneous anonymous voting ≠ sequential, anchored speaking; medians resist outliers, means don't. Same root as Page's theorem (crowd error = average error − diversity), but adds: diversity only pays off if the errors are independent. Practical test: are these judgments given independently, or have people already influenced one another? If the latter, you don't have N samples — you have one, amplified.
AI Prompts
中文提示词
我要从一群人/多个 AI 那里聚合出对 [问题] 的判断。请用群体智慧的四个条件审查我的流程：
① 多样性、独立性、分散性、聚合机制，哪一条最可能被破坏？
② 我现在的收集方式（[描述：开会讨论 / 匿名投票 / 轮流发言…]）会不会让误差相关化？
③ 给我一套具体的聚合机制（提问方式 + 汇总规则），最大化独立性、抑制从众。
English Prompt
I want to aggregate a judgment on [problem] from a group of people / multiple AIs. Audit my process with the four conditions of crowd wisdom:
1. Of diversity, independence, decentralization, and aggregation, which is most likely being broken?
2. Will my current collection method ([describe: open discussion / anonymous vote / sequential turns…]) correlate the errors?
3. Give me a concrete aggregation mechanism (how to ask + how to combine) that maximizes independence and suppresses herding.
① 决策：与其在评审会上问团队「这版能按时上线吗」（人人点头让老板高兴），不如开一个小型内部盘口让大家匿名下注。声誉与积分一旦上桌，真实概率立刻从一片「没问题」里浮出水面。② AI 超级个体：给自己的预测标价。对每个判断写下概率、事后用校准分数（Brier score）结算，等于给自己办一个一人预测市场——让信念付出代价，是最快的去偏方法。这与机制设计的「激励相容」（见 Day 49）同构：当说真话恰好最有利时，真话才会涌现。
English Summary
Prediction Markets — people bet on whether events will happen; the contract price is the market's probability estimate. They beat polls and experts by aggregating differently: weighting by information and pricing conviction, not counting heads. They solve the "cheap talk" problem — opinions are free and noisy, but bets are costly, forcing people to convert private information and true confidence into positions (the more you know, the more you stake). This compresses knowledge dispersed across many minds into one number — Hayek's "price as information." The soul of the mechanism is skin in the game: a natural honesty filter where truth-telling pays. Caveats: thin markets are manipulable and noisy; it only works for resolvable questions; long-shot bias persists. Not a crystal ball — an aggregator that's sharp when the right question is asked.
AI Prompts
中文提示词
我想知道 [某事件/某判断] 的真实概率，但怀疑相关的人都在说场面话。请帮我把它设计成一个预测市场：
① 把问题改写成一个「可裁决、有明确结果」的下注命题；
② 设计计分方式（声誉 / 积分 / 校准分数），让说真话对参与者最有利；
③ 指出这个市场可能太「薄」或被操纵的风险，以及如何缓解。
English Prompt
I want the true probability of [event/judgment], but I suspect everyone involved is giving polite talk. Help me design it as a prediction market:
1. Rewrite the question into a resolvable, clearly-settled betting proposition.
2. Design the scoring (reputation / points / calibration score) so that truth-telling is what pays participants most.
3. Flag where this market could be too "thin" or manipulable, and how to mitigate it.
① 工程：蚁群优化是真实算法，用来解路由与调度；设计多 agent 系统时，与其让一个主控逐一派活，不如让 agent 通过一块共享的工作区（共同的代码库、一份公共草稿）间接协调——这正是共识主动性的工程版。② 意识：你关注的「中央的我是否存在」，在这里得到一个可操作的隐喻——心智可能就是一个神经元蜂群，没有中枢，只有涌现。把无我从哲学命题变成系统直觉：自我是模式，不是指挥官。
English Summary
Swarm Intelligence — ants find shortest paths, bees pick nest sites, birds turn in unison, with no individual holding the global picture and no central controller. Intelligence is a property of the whole system, emerging from simple agents following local rules. The core mechanism is stigmergy: agents don't message each other directly; they modify the environment (pheromone trails) and the environment guides others. Its computing power comes from a balance of positive feedback (trail reinforcement — good paths grow stronger) and negative feedback (evaporation — bad paths fade). Without evaporation the swarm locks prematurely onto the first mediocre path: forgetting is a feature, not a flaw. Isomorphic to distributed systems (no master, local rules, emergent global behavior) and to consciousness — there may be no "central self," only an emergent pattern of a neuron-swarm, echoing the Buddhist anatta (no-self).
AI Prompts
中文提示词
我想用「蜂群 / 共识主动性」的思路重新设计 [一个团队协作 / 多 agent 系统 / 流程]，现在它依赖一个中央指挥很容易堵塞。请：
① 设计每个个体只需遵守的 2–3 条局部规则；
② 设计一块大家都能读写的「共享环境」，让协调通过它间接发生，而非点对点指令；
③ 加一个「蒸发 / 遗忘」机制，防止系统过早锁死在第一个凑合的方案上。
English Prompt
I want to redesign [a team workflow / multi-agent system / process] using swarm intelligence and stigmergy; right now it depends on a central controller that bottlenecks easily. Please:
1. Design 2–3 local rules each individual agent must follow.
2. Design a shared environment everyone can read and write, so coordination happens through it indirectly, not via point-to-point commands.
3. Add an "evaporation / forgetting" mechanism to stop the system from locking prematurely onto the first mediocre solution.
① 工程：git + Pull Request 评审就是维基模式——之所以敢让众人往主干提交，是因为每个改动可审、可回滚，下行被封顶。把「撤销成本」做低，是开放协作能成立的前提。② AI 超级个体：把你的知识库 / 第二大脑当成「一个人的维基」来运营——把自己记录的摩擦降到最低（贡献成本低），并让每条都可修改可回退，它就会像维基一样靠累积纠错单向变好。别追求一次写对，追求让它容易被改对。
English Summary
The Wiki Model — large-scale open collaboration (Wikipedia, open source) produces quality not despite openness but because of it. The engine is an asymmetry: a bad edit's cost is capped by cheap reversibility (revert in seconds), while good edits persist — bounded downside, unbounded upside, so the system ratchets one-directionally toward quality. Same skeleton as convex bets (cap losses, leave gains open); openness is safe precisely because reversal is cheap. It's an evolutionary process: edits are variation, reverts/consensus are selection, retained versions are inheritance — quality without a designer. Plus Linus's Law: with enough eyeballs, bugs are shallow. Failure conditions: contributions must be modular and independently submittable, reversal cheap, and there must be a convergence norm. Most contributions come from a tiny minority (power law), but the long tail catches the errors.
AI Prompts
中文提示词
我想用「维基模式」让 [一个文档 / 代码库 / 知识库 / 社区产出] 靠群体长期自我改善。请：
① 诊断我现在的贡献成本和纠错（撤销）成本哪个太高，怎么压低；
② 检查贡献是否足够「模块化、可独立提交」，不够的话怎么拆；
③ 设计一个收敛规范（标准 / 共识机制），防止开放退化成混乱。
English Prompt
I want to use the wiki model so that [a doc / codebase / knowledge base / community output] self-improves through the crowd over time. Please:
1. Diagnose whether my contribution cost or my error-correction (revert) cost is too high, and how to lower it.
2. Check whether contributions are modular and independently submittable; if not, how to break them down.
3. Design a convergence norm (standard / consensus mechanism) to keep openness from degrading into chaos.