中文精读
预测不了泡沫的终点，也可以先看清自己的判断怎样失真
嘉宾：朱宁，上海交通大学上海高级金融学院金融学教授（当期身份）
主持：曲凯
原发布：2026 年 6 月 13 日｜时长：70 分 46 秒
整理：2026 年 10 月 18 日｜官方单集
技术会改变世界，是否等于与它相关的任何价格都合理？知道一个趋势很重要，是否等于知道何时进入、何时离开？曲凯与朱宁讨论 AI 泡沫，最后不断回到一个更朴素的问题：判断可能错，而错的代价究竟是什么？
本期提供的不是逃顶时间表。朱宁反复强调，事前很难证明泡沫，识别风险特征与预测破裂时刻不同。对产品与创业读者，最有用的迁移，是分开结果、能力与环境，并在热闹叙事之外保留可检查的失败条件。
一、上涨可以同时放大自信、外推与从众
回听：01:27—03:59。 朱宁列出三个反复出现的行为：过度自信、线性外推、从众。价格上涨后，人容易说“我早就知道”，把结果记在自己的预测能力上；下一次便更愿意扩大行动。
线性外推则把最近的方向直接画向未来。昨天上涨、今天上涨，明天似乎也该上涨。它忽略条件可能改变，也忽略较长历史里的不同路径。从众提供另一种捷径：朋友都赚到钱，便把别人的结果当作自己应参与的证据。
三个机制会互相增强。周围人展示成功，强化趋势似乎延续的感觉；自己偶然获得正结果，又使人更相信判断。这不只发生在不懂技术的人身上，专业知识也可能成为解释自信的材料。
编辑推论： 产品增长同样容易出现这个回路。某个月收入上涨，可能来自提价、季节或一次大客户，而非新功能；若直接按最近增速规划招聘和支出，就把结果归因与线性预测叠在了一起。先写替代解释，再谈复制增长。
二、认出技术价值，不等于认出合理价格
回听：04:14—09:57、46:25—53:48。 朱宁把两个命题分开：AI 是重要技术；某项资产值得当前估值。即使承认前者，后者仍需算账。企业有收入、企业盈利、用户因为使用它赚到钱，又是不同的问题。
节目里关于模型公司“赚钱”的说法没有附财务资料，也未分清收入和利润，不能据此形成事实结论。可保留的是价值链检查：上游收到钱，不自动说明下游商业闭环；某个环节增长，也不证明全链条可持续。
资本与叙事并非完全虚无。它们可以吸引人才、增加基础设施、加快研发，反过来改善技术。朱宁用给树浇水、施肥作比喻：可以加速生长，不能让树无限长到天上。因此，叙事能调动资源，却不能代替技术限制与客户结果。
对创业团队，这意味着“热度是假的”和“热度证明一切”都过于简单。编辑推论： 应记录资金如何变成具体产物、产物如何被采用、采用能否支付持续成本。每个箭头都需要证据，不能只展示第一张融资或访问量图表。
三、四项条件，是提醒检查，不是触发行动的公式
回听：25:36—30:41、54:57—58:58。 朱宁归纳四个必要但不充分条件：新概念、产品或技术；宽松流动性；政府支持；缺乏经验的投资者。他同时强调，即使条件凑齐，也不必然发生崩盘，更不保证短期发生。
“必要不充分”在这里很关键：它是嘉宾基于研究与观察提出的分类框架，不是本篇已经验证的普遍定理。节目认为 AI 当时符合这些特征，属于 2026 年 6 月判断，不能当作今天已经核验的市场结论。
他还谈到价格上涨斜率、资金支持、原来质疑者转为乐观、大众集中推荐等观察。这些可以帮助提出“是否拥挤”的问题，却没有明确阈值、回测和误报率。公开预测本身也可能改变参与者行为，因此不能摘成机械信号。
主持追问跌多少才算泡沫时，他给出四五成作为个人选择，并明确不是科学概念。这个回答保留了定义的不确定，不能倒过来把任何下跌达到某个幅度都认定为同一种机制。历史故事的细节若未核，也不应被改装成逃顶教材。
四、圈层与 AI 都增加信息，也可能放大盲点
回听：10:46—19:55、31:15—37:37。 朱宁讨论时代和能力，借市场整体趋势与个人能力的区别，提醒人先问“赚的是什么钱”。同处一个时代仍有不同结果，但不能凭一个成功案例精确分配运气与能力的百分比。
他提到低谷进入行业的人可能走得更远，解释包括真心喜欢、能通过更严筛选、后来遇到复苏。这同时包含选择效应，不能理解为经济越差毕业必然越成功；节目没有附原研究来支持一般因果结论。
圈层则影响听到什么、何时听到。相同故事传到不同人手里，进入成本可能已经变化。身处核心圈也不保证正确，重复接触同一种叙事，反而可能误以为全世界已经达成可靠共识。
AI 又增加一层风险：更容易检索和总结，并不等于拥有机构的数据、流程和风险控制。朱宁说的“双重幻觉”，一层来自模型可能出错，另一层来自人误以为有了模型就没有能力差距。他关于机构与指数产品未来地位的看法是预测，不是本篇的配置建议。
编辑推论： 用 AI 做研究时，应专门要求未核事实、反方材料与结论依赖条件。更多文字如果只是在重复同一来源，增加的可能是确信程度，而非证据独立性。
五、知道偏差以后，还需要能改变动作的安排
回听：20:25—24:45、59:39—01:09:31。 对“这次是否不同”，朱宁没有给绝对答案：新技术确实可能改变许多东西，人也可能重复旧偏差。与其只花力气证明判断一定正确，可以先检查失败后果和承受能力。
后半段最有力的例子是他自己的回忆：雷曼破产前一天还在香港办聚会，周围有许多同事，却没人告诉他第二天会发生什么。研究风险、接近专业人士，都不保证识别自己所在环境的风险。这比把偏差讲成“别人不理性”更有说服力。
节目将财务投资与人力资本投入分开，后者包括学习与能力积累。但学习也有时间和机会成本，不能推出所有培训或转行都必有高回报。具体选择仍要回到目标与约束。
最后，朱宁借当下体验的幸福与回忆中的幸福，讨论财富不是生活的唯一尺度。这是帮助人重新表述目标的框架，不是本篇已经验证的心理测量。外部排名、某次赚钱和自己真正重视的生活结果，并不总是同一件事。
应用练习：为一次高不确定决策写两页纸
第一页写“我为什么相信”：事实、解释、预测各列一栏。每个关键数字写来源和时间，每个因果解释写一种替代解释，最后注明哪些材料其实重复来自同一个圈子。
第二页写“如果我错了”：计划投入多少时间与现金，哪些承诺难撤回，什么结果会影响正常交付，谁在何时重新检查。下面四个问题可以用于产品预算，而非证券买卖：
| 需要检查 | 具体问题 |
|---|---|
| 趋势外推 | 最近增长的条件是否还会存在？ |
| 自我归因 | 哪些变化来自环境，而非团队新能力？ |
| 从众来源 | 三个人的意见，是否其实来自同一篇文章？ |
| 纠偏动作 | 什么证据出现时，会缩小范围或停止投入？ |
目标不是消灭所有风险，而是让“我知道有风险”真正对应某个可执行动作。若说得出所有偏差，却没有任何复查与调整安排，知道仍未变成做到。
来源边界
依据官方介绍与完整转写整理，为访谈精读，不提供买卖、持仓或资产配置方案。宏观、就业、历史估值、研究案例和四项条件保留嘉宾归因，未独立核验。金融人名与年代损坏处不据字面补写；片尾音乐不作观点。编辑推论及练习用于一般产品决策，另行标识。
回听导航
- 01:27—09:57：三种偏差与预测局限。
- 10:46—19:55：结果、时代与能力。
- 20:25—30:41：不确定性与所谓见顶信号。
- 31:15—37:37：圈层、信息瀑布与 AI 幻觉。
- 46:25—53:48：技术、估值与叙事资源。
- 54:57—58:58：四个条件的必要不充分边界。
- 59:39—01:09:31：知道与做到、财富与生活。
相关主题：决策偏差、证据纪律、技术与资本。延伸收听：莫傑麟谈预期与交付、Character.AI 事件中的公司与产品分层。
English edition
Even without predicting a bubble's end, examine how your judgment becomes distorted
Guest: Zhu Ning, professor of finance at the Shanghai Advanced Institute of Finance, Shanghai Jiao Tong University; role at the time
Host: Qu Kai
Originally published: June 13, 2026 | Duration: 70 minutes 46 seconds
Edited: October 18, 2026 | Official episode
Does world-changing technology make every associated price reasonable? Does recognizing a trend mean knowing when to enter or leave? Discussing AI bubbles, Qu Kai and Zhu Ning repeatedly return to a simpler question: judgment can be wrong, so what would the consequences be?
This is not a timetable for escaping a peak. Zhu stresses the difficulty of proving a bubble beforehand and separates risk characteristics from forecasting a collapse. Product and startup readers can use the distinction among outcomes, ability, and environment while keeping inspectable failure conditions outside an exciting narrative.
1. Rising outcomes can amplify confidence, extrapolation, and herding
Listen: 01:27–03:59. Zhu identifies overconfidence, linear extrapolation, and herding. After prices rise, people remember having predicted them and credit their own ability. This encourages larger subsequent commitments.
Extrapolation extends the latest direction into the future: yesterday and today rose, so tomorrow appears likely to follow. It ignores changing conditions and other historical paths. Herding supplies another shortcut: friends' gains become evidence that one should participate.
The mechanisms reinforce each other. Others display success, strengthening the appearance of continuity; a fortunate result increases confidence in personal judgment. Technical expertise does not necessarily confer immunity and may supply more persuasive explanations for confidence.
Editorial inference: Product growth can produce a similar loop. Monthly revenue may rise because of pricing, seasonality, or one large customer rather than a feature. Hiring against the latest growth rate combines questionable attribution with extrapolation. Write alternative explanations before planning replication.
2. Recognizing technological value does not establish a reasonable price
Listen: 04:14–09:57 and 46:25–53:48. Zhu separates important technology from an asset deserving its valuation. Accepting the first does not settle the second. Revenue, corporate profit, and customers making money through a service are further distinct questions.
The conversation's statements that model companies make money include no financial records or consistent distinction between revenue and profit. They cannot become factual conclusions here. The useful question concerns the value chain: upstream receipts do not automatically establish downstream economics, and one growing stage does not validate the whole system.
Capital and narratives are not entirely empty. They attract talent, build infrastructure, and accelerate research, potentially improving technology. Zhu compares them to watering and fertilizing a tree: growth can accelerate without becoming unlimited. Narratives mobilize resources without removing technical constraints or replacing customer outcomes.
Editorial inference: Both dismissing all attention and treating attention as complete proof oversimplify matters. Record how funding becomes an artifact, how the artifact is used, and whether use covers continuing costs. Each connection needs evidence beyond a financing or traffic chart.
3. Four conditions invite examination rather than trigger a formula
Listen: 25:36–30:41 and 54:57–58:58. Zhu lists four necessary but insufficient conditions: a new concept, product, or technology; abundant liquidity; government support; and inexperienced investors. He emphasizes that their combination does not guarantee a collapse, particularly in the short term.
The qualification matters. This is the guest's classification from research and observation, not a universal theorem validated by this article. Calling AI consistent with these characteristics is a June 2026 judgment rather than a verified statement about today's markets.
He also discusses price slopes, funding support, skeptics becoming optimistic, and widespread public recommendations. These may prompt questions about crowding, but the episode supplies no thresholds, backtests, or false-positive rates. Public predictions can themselves change behavior. They should not become mechanical signals.
Asked how much decline constitutes a bubble, he personally suggests roughly forty to fifty percent while explicitly rejecting a scientific definition. That preserves uncertainty; it does not classify every decline of a given size as the same mechanism. Unchecked historical anecdotes should not become a peak-timing textbook either.
4. Networks and AI can expand information while deepening blind spots
Listen: 10:46–19:55 and 31:15–37:37. Zhu contrasts broad market trends with personal contribution, asking what actually produced a gain. People in the same era achieve different results, but one success cannot assign precise percentages to luck and ability.
His suggestion that people entering an industry during a trough may progress further includes selection effects: stronger interest, stricter entry, and later recovery. It does not establish that graduating into a weak economy generally causes success. The episode does not attach original research establishing that conclusion.
Networks shape what people hear and when. The same story can reach people at different entry costs. Being close to an industry does not guarantee accuracy. Repeated exposure to the same account can create an illusion of independent consensus.
AI adds another risk: easier retrieval and summaries do not provide institutional data, processes, or risk controls. Zhu's “double illusion” concerns both model errors and the user's belief that AI eliminates capability gaps. His views about future institutional and index-product dominance are forecasts rather than allocation instructions here.
Editorial inference: AI-assisted research should explicitly retain unverified facts, opposing material, and conditions supporting conclusions. More text repeating one source may increase certainty without increasing independent evidence.
5. Knowing a bias still requires arrangements that change action
Listen: 20:25–24:45 and 59:39–01:09:31. Zhu does not settle whether this time is different. New technology can change much while people repeat old biases. Instead of concentrating solely on proving a judgment correct, examine failure consequences and the capacity to absorb them.
His personal recollection is the strongest example. The day before Lehman's bankruptcy, he held a gathering in Hong Kong with many colleagues, none warning him what the next day would bring. Studying risk and knowing professionals do not guarantee recognizing risk in one's own environment. This is more revealing than describing bias as other people's irrationality.
The episode distinguishes financial investment from human-capital investment such as learning. Learning also uses time and opportunities; the distinction cannot prove every course or career change has a high return. Actual choices need goals and constraints.
Finally, Zhu distinguishes experienced happiness from happiness remembered later. This helps restate personal goals without becoming a psychological measurement verified here. External rankings, one profitable outcome, and a valued life are not necessarily the same result.
Exercise: write two pages for an uncertain decision
The first page explains why you believe it. Separate facts, interpretations, and forecasts. Give important numbers sources and dates, add alternative causal explanations, and identify materials repeating the same network or original source.
The second page asks what happens if you are wrong: time and cash committed, difficult-to-reverse promises, consequences for normal delivery, and who reviews the decision when. These checks concern product budgets rather than securities:
| Check | Question |
|---|---|
| Extrapolation | Will the conditions behind recent growth persist? |
| Self-attribution | Which changes came from the environment rather than new team capability? |
| Herding | Do three opinions actually come from one article? |
| Correction | Which evidence would cause narrower scope or stopped spending? |
The aim is not removing every risk, but connecting awareness to a concrete action. Naming every bias without arranging review and adjustment still leaves knowledge separate from practice.
Source boundary
This is a reading of the official introduction and complete transcript, without trading, holding, or allocation proposals. Macroeconomics, employment, historical valuations, research examples, and the four conditions retain attribution without independent verification. Damaged financial names and dates are not reconstructed, and closing music is not treated as argument. Editorial inferences and the exercise concern general product decisions and are marked separately.
Listening navigation
- 01:27–09:57: Three biases and limits to prediction.
- 10:46–19:55: Outcomes, the era, and ability.
- 20:25–30:41: Uncertainty and supposed peak signals.
- 31:15–37:37: Networks, information cascades, and AI illusions.
- 46:25–53:48: Technology, valuations, and narrative resources.
- 54:57–58:58: Necessary but insufficient conditions.
- 59:39–01:09:31: Knowledge, practice, wealth, and life.
Related themes: decision biases, evidence discipline, technology and capital. Further listening: Mo on expectations and delivery and Separating company and product in the Character.AI event.