非平凡点：① 它本质上是一种信息传递机制。风险信息靠"谁承担后果"在系统里流动；一旦切断这条链路（即代理问题），代理人的动机就和委托人的尾部风险脱了钩。② 要害在下行，而不是整体得失——空谈的人和实干的人，区别就在于后者把自己也押了进去；只有判断错了要付代价，人才会真正校准。③ 它与进化同构：自然选择正是靠风险共担来运作的——押错的基因会被淘汰出基因池。市场、演化、科学的自我纠错，底层是同一台机器。④ 更高的形态是"灵魂入局"（soul in the game）：替他人承担下行，而不只是为自己。
① 技术决策：只画架构图、从不值班（on-call）的架构师，和背着 pager、半夜被自己设计叫醒的架构师，设计质量天差地别。值班就是架构师的风险共担——它让设计自动变得抗脆弱。② AI 协同：当你部署一个会自主行动却不担后果的 agent，所有尾部风险都压在你（委托人）头上。下行风险大的决策，一定要把自己留在回路里，别让"不入局"的 agent 替你下注。③ 育儿：让孩子承担小的自然后果（忘带作业→自己面对老师），比你替她兜底学得更深。过度保护=替孩子拿走了风险共担，养出的是脆弱。
English Summary
Skin in the Game — not merely incentive alignment or fairness, but a filter: decision-makers must bear the downside of their decisions, or the system silently accumulates hidden fragility (gains privatized, losses socialized) until tail risk detonates. It's an information-transfer mechanism — risk info propagates through who carries the consequences; sever that link (the agency problem) and the agent's incentives decouple from the principal's tail risk. Specifically about downside: doers who are exposed calibrate; talkers don't. Structurally isomorphic to evolution — natural selection removes bad bets from the gene pool via skin in the game. Test any advice by asking: what does the advisor lose if they're wrong? If nothing, down-weight it sharply.
AI Prompts
中文提示词
我面临一个决策/正在听取一组建议：[描述决策与建议来源]。请用"风险共担"透镜分析：
① 每个建议者/决策方，如果判断错了实际会损失什么？谁只享上行不担下行？
② 这个系统里，下行后果最终压在谁身上（找出隐性的"社会化亏损"）？
③ 给出 1 个具体设计，把决策权和下行后果重新绑定，让我和关键方真正入局。
English Prompt
I'm facing a decision / weighing advice from several sources: [describe decision and who's advising]. Analyze through the Skin-in-the-Game lens:
1. For each advisor/party, what do they actually lose if wrong? Who enjoys upside without downside?
2. Where do the downside consequences ultimately land (find the hidden "socialized losses")?
3. Propose one concrete redesign that re-binds decision rights to downside exposure, putting me and key parties genuinely in the game.
医学史是一部否定之道的教科书。在抗生素之前的漫长岁月里，医生的"积极干预"（放血、灌肠、各种猛药）杀死的病人多过救活的。人类寿命的最大跃升不来自"添加"某种神药，而来自移除有害因素——洗手、清洁饮水、隔离病原、停掉那些有害的常规操作。"首先，不要造成伤害"（primum non nocere）正是这条原则的医学誓言。
场景 · BigCat
① 生产力：杠杆最高的一招往往是做减法（砍会议、退掉承诺、卸载工具），而不是再上一套新系统。给日程做减法，比给日程做加法，收益高一个量级。② AI 工作流：当 agent 输出变差，第一反应往往是"再加几条指令"——错。先删掉冗余 context 和互相打架的提示，上下文越臃肿，模型越糊涂。③ 思考：想看清一件事，先移除你"确信却错误"的旧信念，再谈引入新框架——清空错的，比塞进对的更要紧。④ 育儿：移除障碍（过密日程、睡前屏幕），往往比再报一个兴趣班更能让孩子发展。
English Summary
Via Negativa — improve by removal, not addition. Our knowledge of what's wrong/harmful is far more robust than our knowledge of what's right; falsification beats verification, subtraction beats addition — because removing a known harm has bounded side effects, while adding something "beneficial" injects untold downstream couplings you can't compute. Negative knowledge survives time better than positive theory (Popper's falsifiability is via-negativa epistemology). What survives is what's left after removal — addition is fragile, subtraction antifragile (echoes Lindy). Humans have a strong addition bias: facing a problem, people overwhelmingly add elements and rarely consider removing one, even when removal is better. Default to asking "what can I delete?" — deletion is the default, addition must justify itself.
AI Prompts
中文提示词
我想改进 [系统/流程/生活领域]，现状是 [描述]。请用"否定之道"反转我的默认思路：
① 列出 3-5 个最该移除的东西（有害的、冗余的、增加脆弱性的），按"移除收益 × 移除成本低"排序；
② 指出我的"加法偏误"在哪——我本能想添加什么，而那其实是错的方向；
③ 给出一个"减法优先"的改进方案，并说明移除每一项的副作用为何是有界的。
English Prompt
I want to improve [system/process/life area]; current state: [describe]. Invert my default thinking using Via Negativa:
1. List 3–5 things most worth removing (harmful, redundant, fragility-adding), ranked by removal-benefit × low removal-cost.
2. Point out my "addition bias" — what I instinctively want to add that's actually the wrong direction.
3. Propose a subtraction-first improvement plan, explaining why each removal's side effects are bounded.
① 作为"AI 超级个体"的职业设计：守住一块稳定的收入基本盘（安全端），同时对新兴 AI 能力押下非对称的小赌注（凸性端）——别把全部身家压在一条"看起来还行"的单一路径上。② 时间分配：80% 投在可靠的核心能力，20% 投在可能 10 倍回报的疯狂实验。③ 阅读：读经得起时间的根基（数学、经典）+ 最前沿的论文，跳过那些读完就忘的中间地带。④ 育儿：稳固的安全基地（依恋、规律作息）+ 允许游戏里的大胆冒险探索；别去过度优化那个"安全的中间"。
English Summary
Barbell Strategy — load both extremes (extreme safety + extreme aggression), deliberately avoiding the "seemingly prudent middle": e.g., ~90% in maximal safety + ~10% in high-convexity bets. The point is convexity: the safe leg caps downside (bounded loss), the aggressive leg keeps unlimited upside — you want asymmetric payoff. "Medium risk" is a trap because it relies on knowing the risk precisely, and tail risk is unknowable; a basket of medium-risk assets correlates and crashes together in a crisis — diversification is an illusion. Transfers across domains: career (stable income + wild side bets), reading (timeless classics + frontier, skip the forgettable middle). Same family as optionality — buying many positive options cheaply. Don't ask "is this high or low risk?" — ask "is my downside capped and my upside open?"
AI Prompts
中文提示词
我要在 [领域：职业/投资/时间/学习] 做配置，现状与选项是 [描述]。请用"杠铃策略"重构：
① 指出我现在是不是落在危险的"中等风险中间地带"——有没有自以为稳健、实则尾部暴露的选项？
② 设计一个杠铃：哪部分该极度安全（封死下行）、哪部分该押高凸性赌注（敞开上行）、各占多少？
③ 列出 2-3 个低成本、高上行的非对称小赌注，说明每个的最大亏损是否有界。
English Prompt
I'm allocating in [domain: career/investing/time/learning]; current state and options: [describe]. Restructure with the Barbell Strategy:
1. Am I sitting in the dangerous "medium-risk middle"? Flag any option I think is prudent but is actually tail-exposed.
2. Design a barbell: which part should be extremely safe (cap downside), which should be high-convexity bets (open upside), and in what proportions?
3. List 2–3 low-cost, high-upside asymmetric bets, stating whether each one's maximum loss is bounded.
① 工程：对一个运行良好的生产系统不停"优化"、手痒去调参数，结果引入的故障常常比它预防的还多。反倒是那些"忍住不去碰正常系统"的值班工程师，可用性最高——别去干预一个正常运转的系统。② 管理：对一支高绩效团队搞微观管理（干预），会把它之所以高效的那套自组织亲手摧毁。③ AI：对一条已经跑通的流程反复重调提示词、过度工程化——克制往往胜过折腾。④ 育儿：盘旋在孩子头顶、替她解决每一个小困难（医源性伤害），等于剥夺了她发展韧性的机会；"观察性等待"让自然的发展规律去工作。
English Summary
Naive Interventionism — the urge to act even when doing nothing is better, while ignoring iatrogenics (harm caused by the healer/intervener). Complements Via Negativa: the latter tells you to remove harmful things; this warns that the "intervention" you blindly add is often itself the harmful thing. Driven by action bias: visible intervention is rewarded, visible inaction punished — even when inaction is correct — and the harm of intervening is delayed/invisible, so it's underweighted. In complex systems, interventions have nonlinear, lagged side effects; the more robust and self-healing the system, the more likely intervention backfires. The skill is discernment: intervene aggressively in fragile, genuine emergencies; restrain in robust, self-healing systems (cf. Daoist wuwei, "watchful waiting," Chesterton's Fence). Before acting, ask: "what happens if I do nothing?" — put non-intervention on the table as a real option.
AI Prompts
中文提示词
我正打算对 [系统/团队/孩子/项目] 进行干预：[描述我想做什么]。请用"朴素干预主义"压力测试我：
① 如果我什么都不做，会发生什么？这个系统能自愈吗，还是真的需要急救？
② 我的干预可能带来哪些延迟、隐形的医源性副作用？它会不会亲手制造它想防止的问题？
③ 判定这是"该果断出手的脆弱情形"还是"该克制的稳健系统"，并给出最小必要干预（或"按兵不动"）的方案。
English Prompt
I'm about to intervene in [system/team/child/project]: [describe what I want to do]. Stress-test me with Naive Interventionism:
1. What happens if I do nothing? Can this system self-heal, or is it a genuine emergency?
2. What delayed, invisible iatrogenic side effects might my intervention cause? Could it create the very problem it aims to prevent?
3. Judge whether this is a "fragile situation demanding decisive action" or a "robust system demanding restraint," and propose the minimal necessary intervention (or "stand pat").