① AI 工作流：设计 agent 的奖励时，若只奖励即时可见的输出，agent 会"双曲贴现"地牺牲长期目标（如为短回复牺牲正确性）——长程任务必须显式引入对未来状态的承诺式约束，而非指望它"自觉"。② 深度工作：靠意志力对抗手机必败；把手机锁进另一个房间（承诺装置）才有效。③ 育儿：与其反复说教"先写作业"，不如和孩子共同设计一个无法临时反悔的结构（固定时段、设备托管）。不要和冲动的当下自我硬碰硬，要替远见的自我提前布好局。
English Summary
Present Bias / Hyperbolic Discounting — humans discount the future hyperbolically, not exponentially: near delays are penalized steeply, distant ones barely at all. This curve shape produces preference reversal — from afar you commit to the larger-later reward, but as the smaller-sooner temptation approaches, you defect. Procrastination and weakness of will aren't moral failures; your present self and future self are literally different decision-makers (echoing the Buddhist no-continuous-self). Willpower loses because the present self wins every local fight. The fix is precommitment (Ulysses contracts): let the farsighted self lock the impulsive self out of the choice, and make distant rewards and present costs vivid.
AI Prompts
中文提示词
我在 [目标/习惯] 上反复"远看坚定、临近叛变"。请用现时偏误/双曲贴现分析：
① 指出我的"现时自我"在哪个临界点夺权，是什么即时诱惑触发了偏好反转；
② 别给我"更自律"的建议——给出 2 个具体的承诺装置，让未来的我无法临时反悔；
③ 怎样把远期回报拉到眼前变具体、把当下成本变可见，从而改变贴现的输入。
English Prompt
I keep being "firm from afar, defecting up close" on [goal/habit]. Analyze via present bias / hyperbolic discounting:
1. Pinpoint where my present self seizes control and which immediate temptation triggers the preference reversal.
2. Don't tell me to "be more disciplined" — give 2 concrete precommitment devices that make my future self unable to defect.
3. How can I make the distant reward vivid now and the present cost visible, changing the inputs to my discounting?
① 学习配置：在一个主题上的第 5 小时，新增收获远小于在一个新领域的第 1 小时——跨学科迁移本质上就是在套利边际效用递减，把注意力从饱和曲线挪到陡峭曲线。② AI 产品：第 20 个功能对用户的边际效用可能为负（复杂度反噬），算力也一样，与其堆参数不如换条曲线。③ 育儿：第三个兴趣班的边际收益常已为负，孩子的空闲与无聊反而是更陡的那条曲线。不要最大化单一输入，要在多条递减曲线之间做边际配平。
English Summary
Diminishing Marginal Utility — each additional unit yields less added satisfaction; the utility curve is concave. From this one curve flow risk aversion (concavity means a sure amount beats an equal-EV gamble, since E[u(x)] < u(E[x])), diversification, and the logarithmic value of money (a dollar is worth more to the poor). The deeper move: don't maximize an input — find the knee of the curve and stop, reallocating to another still-steep curve. Neuroscience rhymes: the brain encodes deltas, not levels (dopamine prediction error, hedonic adaptation), so any constant good decays toward zero utility — the hedonic treadmill — and novelty/variety is what resets the curve. The water-diamond paradox dissolves once you separate total from marginal value.
AI Prompts
中文提示词
我正在把大量 [时间/金钱/算力/注意力] 投入 [某个输入]。请用边际效用递减审视：
① 估计这条曲线的"膝点"大概在哪，我是否已越过拐点在做无效堆叠；
② 列出 2 条目前更陡峭的替代曲线，把边际资源挪过去能多换多少总效用；
③ 哪些"稳定不变的好处"正被享乐适应吃掉？用什么新异/多样把曲线拨回陡峭。
English Prompt
I'm pouring lots of [time/money/compute/attention] into [one input]. Examine it via diminishing marginal utility:
1. Estimate where this curve's knee is — have I crossed the bend into wasteful stacking?
2. List 2 currently-steeper alternative curves; how much total utility would shifting marginal resources there buy?
3. Which "constant goods" are being eaten by hedonic adaptation, and what novelty/variety could reset the curve to steep?
① AI 浪潮的主旋律就是颠覆式：更小/更便宜/开源的模型从低端起步，以"够用且极廉"吃掉被高端模型过度服务的大量任务。② "AI 超级个体"本身即颠覆者——一个人用廉价工具做到 80% 的咨询/开发产出、收 1% 的价，正从下方掏空过度服务的机构。③ 反向自检：你自己提供的价值里，哪一块是"远超对方真实所需的高端性能"？那正是最先被"够用版"替掉的部分。别问"我能否做得更精致"，要问"哪里的够用版正在便宜十倍地逼近"。
English Summary
Disruptive Innovation (Christensen) — disruptors rarely beat incumbents head-on. They enter at the low end or a new market with a product that's worse on the metrics mainstream customers value, but cheaper, simpler, more convenient. Incumbents rationally cede the thin-margin segment and chase high-margin customers — the trap. As the disruptor improves along its own trajectory, it eventually clears the "good enough" threshold and the incumbent's accumulated high-end performance becomes oversupply: worthless. What kills the incumbent is correct resource allocation, not laziness — a structural dilemma. Distinguish disruptive from sustaining innovation (most improvement is sustaining). Like evolution, invasion comes from the margins, not the center. Watch where you over-serve customers, and where a "worse-but-cheaper" thing is climbing toward you.
AI Prompts
中文提示词
我/我们的产品或角色是 [描述]，主流客户/受众最看重 [指标]。请用颠覆式创新分析：
① 我在哪些维度"过度服务"了——性能已超出对方能消化的范围，成了暴露的腹部？
② 有没有"一开始更差、但便宜/简单到打开非消费市场"的东西正沿自己的轨迹爬上来？
③ 区分我面临的是延续式还是真正的颠覆式威胁，并给出一条"自我颠覆"而非死守高端的路径。
English Prompt
My product/role is [describe]; mainstream customers value [metrics]. Analyze via disruptive innovation:
1. On which dimensions am I over-serving — performance beyond what they can absorb, an exposed belly?
2. Is there a "worse-at-first but cheap/simple enough to open non-consumption" entrant climbing its own trajectory toward me?
3. Distinguish whether I face a sustaining or a truly disruptive threat, and give one path to self-disrupt rather than defend the high end.
① 投资：仓位大小要用几何增长率而非算术期望——押注须小到能"活着复利到下一轮"，这正是凯利与安全边际的内核。② AI 部署：一个"99% 安全"的动作重复一千次，失败近乎必然——风险在时间上累积，是非遍历的，不能用单次期望安慰自己。③ 育儿：有些对孩子的风险是不可逆吸收壁，须与普通"正期望"风险区别对待。先问有没有归零点；只要有，先活下来永远压倒最大化期望。
English Summary
Ergodicity — a system is ergodic only when its time average (one trajectory over time) equals its ensemble average (many trajectories at one instant). Much of life and markets is non-ergodic, especially under multiplicative dynamics with an absorbing barrier (ruin, death). The classic coin flip (+50% / −40% per round) has positive ensemble expectation yet almost every individual path drifts to zero — the mean is propped up by a few lucky outliers you'll never be. The right object is the geometric (log) growth rate, which penalizes volatility and never compromises with ruin — the math behind the Kelly criterion, skin in the game, and every "avoid going to zero" instinct. Averaging over parallel worlds you don't inhabit is the subtlest error in a non-ergodic world. Always ask: am I averaging over parallel worlds, or over my own time? If there's an absorbing barrier, survival beats maximizing expectation.
AI Prompts
中文提示词
我在考虑这个决策/赌注：[描述，含可能的收益、概率与最坏结果]。请用遍历性审视：
① 这是对"平行世界取平均"还是对"我自己的时间取平均"？两者结论是否分叉？
② 路径上是否存在吸收壁（破产/出局/声誉崩塌/健康不可逆）？若有，期望值就是误导。
③ 给出按几何增长率/"先活下来"逻辑的版本：该如何缩小仓位、或直接拒绝。
English Prompt
I'm weighing this decision/bet: [describe, with payoffs, probabilities, and worst case]. Examine via ergodicity:
1. Is this an average over parallel worlds or over my own time? Do the two conclusions diverge?
2. Is there an absorbing barrier on the path (ruin, getting knocked out, reputation collapse, irreversible health harm)? If so, expected value misleads.
3. Give the geometric-growth / "survive first" version: how to shrink the bet size, or simply decline.