把隐性判断说清楚：Annie Duke 谈更好的决策与更早放弃

Asked what stuck with her from her friendship with Daniel Kahneman, who had died about a week before the recording, Duke starts with his humility. Despite his enormous influence, he wanted to hear what others thought and was quick to say he did not know much. When a student asked about studies in *Thinking, Fast and Slow* that failed to replicate, he was the first to say he wished priming were not in the book.

Duke contrasts this with researchers whose careers rest on a line of work that did not replicate and who argue against the evidence. Kahneman, in her telling, accepted that he knows what he knows now but not everything. He was also a pioneer of adversarial collaboration, where you find someone who strongly disagrees with you and write a paper together on the study that would settle the question.

Her example is the claim, from Kahneman and Angus Deaton, that money stops buying happiness beyond roughly $75,000. A decade later Matthew Killingsworth found happiness keeps rising with income. Rather than trade attacks, Kahneman teamed up with Killingsworth and Barb Mellers to resolve it. Duke adds that he was endlessly curious about other people and recalls a lunch that turned out to be him asking about her work the whole time.

Asked about parenting frameworks, Duke gives two pieces of advice. Kids need to know deep down that you love them, and at some point you will drop your baby on its head, which happened with each of her four children. Her point is that mistakes happen, babies are resilient, and every small event does not leave a lasting mark. She also says that with four kids you realise you have little influence on temperament, as a well-behaved first child can make parents feel smug, which she calls resulting.

For decisions, she borrows Kahneman's line that nothing is as important as it seems while you are thinking about it. She uses mental time travel with her kids by asking how they will feel about the situation years from now, with the grounding as a funny story at Thanksgiving when they are 40. The tool counters the focusing effect, where the feelings of this moment swamp our sense of time. Some things will matter in 20 years, but most will not.

Her second tool is the word "nevertheless." When a child argues against a punishment, she wants them to feel heard while still following through: I hear you, nevertheless you are grounded. In a workplace, leaders can tell employees that their input was incorporated and that, nevertheless, this is the path the company is taking. Duke says it is a strong leadership skill.

Lenny notes that Kahneman himself said knowing about biases did not make him a perfect decision-maker, and asks how much better people can actually get. Duke answers that it depends entirely on whether you do the work. She points to Kahneman's early research on hiring, which moved from unstructured judgment ("I know a great product manager when I see one") to making the implicit model explicit.

The steps were to define what you are looking for in the abstract, turn it into a decision rubric and build a structured interview. Only after that process did Kahneman have people apply their intuition. Duke says the hit rate in hiring went from about 50% to 65%, which she calls enormous for a noisy domain. The catch is that nobody does it, and Kahneman often asked her how she gets anyone to.

Duke thinks we overrate our intuition and our ability to notice things in the moment, and often other people have more insight into our situation than we do. She sees three groups: people who do not know the methods, people who know but do not use them, and the tiny group who do. Anyone willing to join the last group has a large edge.

The easiest change Duke has seen stick is to gather opinions independently and asynchronously. People assume a meeting does three things: discover opinions, discuss them and decide. Doing the first in a room produces cross-influence, and the loudest or most confident person gets outsized sway. Only the discussion belongs in the meeting.

For discovery, she suggests sending a prompt before the meeting. Examples include brainstorming features and force-ranking them with a short rationale, ranking an existing list, or estimating timelines in sprints. Replies go only to the organiser, or into a form that dumps to a spreadsheet only one person sees. The organiser then circulates everyone's answers beforehand, which she calls a nominal group: people work as a group but independently. The meeting itself should focus on where people disagree, since much meeting time is spent agreeing in different words.

Decisions can also be made outside the room, by one decision-maker or by private votes. Duke says the word "alignment" is stupid because ten different people will not leave a room agreeing, and expecting them to makes meetings coercive. The aim is to convey why you believe something rather than to convince. Leaders then say what has been heard and that, nevertheless, the decision is the decision.

Lenny asks how people can still feel included when one person decides. Duke says the fix is to move from a coercive to a curious mode. Interrupting silences someone, and saying "you're wrong" makes people tribal. She gives a forecast example: each person sends a point estimate plus lower and upper bounds with a rationale, then the group discusses and asks clarifying questions.

As facilitator, she never says "I agree." She reflects each view back, as in "what I heard you say is...", and lets the speaker correct her. She illustrates with a client's debate over whether remote and hybrid policy should be consistent across functions. Lenny says he would feel good if someone showed they understood him even if the decision went the other way.

Duke says people end up feeling more endowed with the decision and see themselves reflected in it. Independent input also reveals the real spread of opinion. A junior person who thought four weeks would never say so after a senior says three months, but the written estimates surface it. Seeing that spread helps people accept an outcome that is not exactly what they wanted. Lenny summarises the model as independent brainstorming, group discussion of disagreement, then a single decider, and Duke adds that the same approach works for forecasting, planning and budgeting.

Answering a question suggested by First Round's Brett Berson, Duke recalls what venture firms told her after *Thinking in Bets* in 2018. Some said they did not need to make decisions explicit because they know a good founder when they see one. Others said her ideas did not apply because their feedback loops last a decade, unlike poker. She replied that poker is much noisier than people think, since she rarely sees opponents' cards and needs many hands to know whether she has an edge.

To Josh Kopelman, who said exits take ten years, she asked whether investors fall asleep like Rip Van Winkle and wake up a decade later. In between, the company may fund at Series A, which she notes was necessary for a billion-dollar exit and is more correlated with outcomes than the investment at C. Other signals include product-market fit, traction, new ARR, churn and ability to retain top talent.

Her bold claim is that you choose how long your feedback loop is: ask which things are necessary but not sufficient, or correlated with the outcome you want, and forecast those. Duke notes that Series A funding currently comes roughly 16 months after seed. She also explains why people prefer long loops, because they preserve the psychological safety of being believed good, such as an early Uber investor who does not want to learn how much was luck. Humans trade the long run for feeling good now, and tighter loops risk feeling wrong sooner.

When Duke started, First Round recorded the partners' votes but little else. She had them rate market, team, founder and product on a scale of one to seven rather than good or bad, and create shared definitions through mediating judgments, such as competitive landscape for market. They also began forecasting explicitly, starting with the probability a company funds at Series A.

Five years in, they know whether hundreds of companies funded at Series A and can show partners how accurate their forecasts were. They can also check how ratings of market or founder map onto how companies later did. Duke says this tells partners where their judgment is strong, where to avoid over-indexing, and lets them revise the rubric using data. She declines to say who is best, but says every partner has strengths and weaknesses that only partly overlap, which shows the value of diverse input.

The most surprising finding was that things partners pound the table about are sometimes highly predictive, for them and others, and sometimes not predictive even for the person pounding. Duke says this is why making the implicit explicit matters: intuition is not garbage, since we would die otherwise, but without making it explicit you never find out when it is off base. Even true experts do not necessarily have good intuition about what is important.

On pre-mortems, which are popular in product, Duke says they work only if paired with a pre-commitment. In her research with Maurice Schweitzer and Linnea Gandhi at Penn, pre-mortems generally did not change plans or behavior much. Their best use is setting kill criteria: signals that tell you to pivot or stop. Once something launches, people are slow to quit, and projects run over budget and past deadlines before being shut down.

She gives an example from a sales team asked to imagine that a lead from an RFP or RFI had consumed six months and the deal was dead. They listed early warning signs: the RFP seemed written with a competitor in mind, the customer would not take a demo and only talked price, and no decision-maker could be brought to the room after the first few meetings.

Each signal got an action. Price-only means kill the deal. A competitor-written RFP means asking directly how far along the competitor is, then killing or pursuing depending on the answer. A missing decision-maker means offering executive alignment at the next meeting, and killing if they decline. This also protects salespeople from being second-guessed for abandoning deals. Duke ties it to climbers who push on into a blizzard, and Lenny notes it is another case of making implicit assumptions explicit.

On her book *Quit*, Duke says the data suggests people quit long after they should. Starting involves luck and hidden information, which is why Bezos has a 70% rule to start with less than full certainty. But stopping is also a decision under uncertainty, and we do not like to walk away unless we know for sure. She cites Richard Thaler's observation that most people do not quit until it is no longer a decision, and notes that the same applies to jobs, relationships, projects and products.

Several biases pile on. Sunk cost makes us treat past investment as waste if we leave, although the real waste is what we continue to put in going forward if we would not start today. Endowment makes us value what we have built more than identical things we do not own, which is especially acute in product. Identity and how others see us matter too, since quitting because you had no choice invites no questions, while quitting early does.

Her story is Stewart Butterfield's game Glitch. It was a critical hit with $6 million in the bank, strong investors and 5,000 diehard users, but for each heavy player there were 95 to 99 who left after five minutes. After a six-week paid-marketing test in 2012, Butterfield woke up certain that Glitch was over. He saw it was not a venture-scale business and that the employees' equity was not worth their time. Two days later he proposed turning the internal communication tool into the next product, which became Slack. Duke says the cost of continuing includes the opportunities you cannot see, and he could not see Slack until he quit Glitch.