Deciding Under Uncertainty

This is a narrative theme with a typological map of techniques, synthesising how Annie Duke, Howard Marks, Bill Miller, Morgan Housel, Shreyas Doshi, Alex Komoroske, and Matt Dixon each approach the same problem — making good decisions when the outcome is unknown — across poker, investing, and product.

The core claim

A good decision is not a good outcome. Judging a decision by how it turned out — resulting, in Annie Duke‘s term — is the central error, because every outcome is a mix of decision quality and luck. A professional poker player who goes all-in on the statistically correct hand and loses did not make a bad decision; a founder who rode a lucky market to a successful exit and concludes they have superior judgement may have made no decision at all.

The disciplines for deciding well under uncertainty recur across domains. They divide into three groups: techniques for forming the decision itself cleanly (Kill Criteria, Eigenquestion, Nominal Group Technique); techniques for calibrating how much risk to carry (Risk Posture, Knightian Uncertainty); and techniques for surviving long enough for the process to pay off (survival-first compounding, reversible bets via Do Half).

Resulting: the error that connects the domains

Annie Duke‘s Thinking in Bets names the universal failure mode. Resulting is the habit of inferring decision quality from outcome quality — praising a bad process that happened to work, penalising a good process that happened to fail. It corrupts learning across every domain.

In poker, resulting produces bad calls: a player who called a bluff and got lucky will call bluffs too often in future. In investing, it produces survivorship bias: the manager who loaded up on tech in 1998 and happened to exit before the crash looks prescient; the manager who made the same bet but held too long is written off as reckless, even if both made the same decision under the same information. In product, resulting corrupts retrospectives: a launch that shipped late but succeeded by market luck generates the wrong lessons for the next launch.

The corrective is identical across all three domains: judge the process, not the outcome. Ask ‘given what we knew at the time of the decision, was this the best available bet?’ rather than ‘did it work out?’. That shift requires pre-commitment — writing down the reasoning before the outcome is known, so the outcome cannot rewrite the reasoning.

Pre-commitment: deciding before you know

The reason Kill Criteria work is not the pre-mortem that generates them but the pre-committed action attached to each signal. Duke’s observation: teams routinely run pre-mortems, list the early warning signs of failure, and then proceed largely unchanged when those signs appear. The pre-mortem’s value is that it generates the criteria; the pre-commitment is the active ingredient.

A sales team’s kill criterion: if a prospect will discuss price but not agree to a demo, the lead is dead. Pre-committing that action before the deal starts removes the in-the-moment rationalisation (‘maybe they just need more time’) that sunk-cost bias produces at the worst moment. See Annie Duke on Better Decisions, Kill Criteria, and When to Quit.

Howard Marks’s version of the same discipline is his five major macro calls in fifty years. Each time he shifted significantly around his default Risk Posture, he was acting on a pre-formed observation — investor behaviour had moved to extremes that were plain to see — rather than a real-time forecast. The rarity of the calls is the discipline. If he had made five thousand calls instead, his record would have been fifty-fifty. See Howard Marks on Avoiding Disaster, Risk Posture, and the AI Bubble.

Shreyas Doshi runs the same logic in product via the pre-mortem: before a launch, imagine the project has already failed and ask what went wrong. The prompt creates psychological safety for surfacing concerns that optimism culture suppresses — and builds a shared vocabulary (tigers, paper tigers, elephants) that persists into future team meetings without the leader present. See Shreyas Doshi on Product Management Frameworks.

Knightian Uncertainty: the irreducible residue

Not all risk can be pre-committed away. Knightian Uncertainty — Bill Miller‘s operating concept, drawn from Frank Knight’s 1921 Risk, Uncertainty and Profit — names what remains: situations where no probability distribution can be assigned, because the distribution itself is unknown.

Risk has a known distribution; you can insure against it, price it, hedge it. Uncertainty is irreducible: no amount of data resolves what probability to assign. Amazon in 2002, with bonds yielding 20–30% and the equity in single digits, was a case of genuine uncertainty — the distribution of outcomes was simply unknowable. Miller held forty to fifty per cent of his personal portfolio in Amazon for decades not because he had resolved the uncertainty but because he was more comfortable than most with sitting inside it. His competitive advantage is accepting the discomfort of not knowing; most institutional mandates require risk to be bounded and explainable, which means managers cannot hold positions where the distribution is genuinely open. See Bill Miller on Amazon, Bitcoin, and Buying at a Discount to Future Value and Knightian Uncertainty.

Alex Komoroske’s adjacent possible framing maps this directly onto product strategy: the genuinely reachable options right now are small, and each action taken reveals the next set. The mistake is assuming the adjacent possible is large and jumping to the end state. Combine a low-resolution North Star with the discipline of taking only the next available step. The uncertainty about whether the North Star is reachable is real and irreducible; the North Star is not a forecast, it is a direction. See Alex Komoroske on Strategy and Complexity.

Risk posture: calibrate, do not maximise

Howard Marks frames the posture question as the speedometer: zero is no risk, one hundred is maximum possible risk. Every investor should determine their default position — calibrated on age, wealth relative to income, dependants, and intestinal fortitude — and recalibrate only when the evidence is compelling. Not constantly; that produces hyperactivity. Rarely, deliberately, and based on observable investor behaviour rather than macro prediction.

The asymmetry of losses is the reason this matters. A fifty per cent loss requires a hundred per cent gain to recover. Consistent avoidance of the worst outcomes allows compounding to work uninterrupted. The General Mills pension fund was never above the twenty-seventh percentile or below the forty-seventh in equities for fourteen years; the fourteen-year result was fourth percentile. Marks’s conclusion: ‘If you can avoid the losers, the winners will take care of themselves.’

Morgan Housel reaches the same conclusion from a different angle. Ninety-nine per cent of Warren Buffett’s net worth was accumulated after his sixty-fifth birthday — that is simply how an exponential curve works. The investor’s primary job is not to generate returns in the first decades but to avoid being forced out. Survival is the strategy. See Morgan Housel on Contentment, the Independence Spectrum, and Why Survival Is the Only Strategy.

Reducing the decision to its Eigenquestion

Before carrying out any of the above, the decision has to be clearly framed. Shishir Mehrotra’s Eigenquestion identifies the smallest set of questions whose answers determine the most important choices — the question that drives the answer rather than being driven by it. His teleportation device exercise forces it: with only two questions allowed, which two do you ask? The constraint eliminates noise and surfaces the dimensions that actually partition the strategic options.

In practice: before a product bet, identify whether the market exists and whether you can reach it economically — two folders that contain all the strategic decisions downstream. Before a long position, identify whether the business is one of the roughly four per cent of companies that generate all net equity returns over time, and whether the current price reflects the future distribution of that value. The Eigenquestion does not reduce uncertainty; it concentrates attention on the uncertainty that matters.

Indecision as a distinct failure mode

Matt Dixon‘s large-scale research on B2B sales adds a dimension the other speakers do not address directly: indecision is its own category, distinct from a bad decision and from inaction through preference for the status quo. In the dataset from 2.5 million sales calls, forty to sixty per cent of qualified pipeline is lost not to a competitor but to no decision — and fifty-six per cent of those no-decision losses represent customers who want to buy but cannot act. The driver is FOMU: fear of messing up, of being personally blamed if the purchase fails. FOMO intensifies it.

The JOLT method addresses this by reducing the irreversible character of the commitment — narrowing choices, sharing the psychological burden of recommendation, and taking risk off the table with under-promised ROI, opt-out clauses, and pre-signature introductions to implementation teams. These are, structurally, the same moves that appear elsewhere in this theme: reversibility (cf. Do Half), pre-commitment to a recommendation, and making the worst-case outcome survivable. See Matt Dixon on the JOLT Effect, the Challenger Sale, and Overcoming Indecision and JOLT Method.

Reversible bets: make the downside survivable

Do Half — Scott Belsky’s discipline of building fewer features, serving a narrower market, and killing existing functionality to concentrate usage — is, in decision terms, a bet-sizing strategy. It ensures that no single product choice is catastrophic. Small bets, reversible bets, and cheap bets all reduce the cost of being wrong. Belsky’s empirical finding at Behance: removing features caused core-action frequency to increase. Each removal concentrated attention. The product equivalent of Marks’s fewer-losers approach: remove the things that diffuse; concentrate on what compounds.

The common thread: across poker, investing, and product, the discipline is to ensure that no single bad decision — even a bad decision on a good process — can end the game.