Why Strong Technical Founders Fall Into the Advantage Migration Trap

The real risk is not being less intelligent

The most dangerous moment for a technical founder is rarely when the product cannot be built. It is when the product is finally built and nobody will change their behavior for it. In a Y Combinator Startup School talk, Francois Chaubard offers a sharper explanation: many top researchers do not fail at startups because they are insufficiently smart. They fail because the abilities they were once most proud of become obstacles under a new evaluation system. He describes this as an identity shift—to become a real founder, you have to kill part of your former self. 1 2

Two evaluation systems

DimensionThe researcher’s defaultThe founder’s question
What counts as goodNovel enough to pass peer reviewUseful enough to a specific person
How value is provenPapers, citations, experimental resultsCustomers change behavior and pay a real cost
How problems are chosenThe harder and deeper, the more research valueWhich problem is important enough that someone will pay for it?
How time is usedWait for sufficient evidence before concludingAct on limited information, then correct
The boundary of abilityIndividual depth in a narrow fieldAn organization’s ability to understand, deliver, and capture value

This is not a rejection of professional training. A researcher’s sensitivity to causality, evidence, and boundary conditions remains irreplaceable in safety, compliance, and irreversible decisions. The problem is that a startup transfers the authority to decide what is worth doing from peer review to customers, markets, and cash flow. Killing part of your former self means, more precisely, switching off the defaults that work only in an academic environment.

Originality asks who did it first; business asks why anyone chooses you

In research, insufficient novelty can get a paper rejected. In business, an excessive desire to be original can lose you the market. Chaubard cites Joe Lonsdale’s memorable taxonomy: an MIT-style startup invents a breakthrough technology first, then looks for the problem it can solve. A Stanford-style startup finds a real, specific customer problem first, then keeps improving the solution. This is not a claim that people from those two universities can only start companies that way. It is a claim that two starting points carry two different risks. 1

Two startup paths

PathStarting pointCore activitySuccess criterionMain risk
MIT-styleA new technology or inventionFind a use case for the technologyTechnical leadership and a defensible moatThe technology is new, but the problem is not painful
Stanford-styleA specific, recurring customer problemIterate around the same problemAdoption, retention, willingness to paySlow execution, or a faster competitor gets there first

Delivery, ride-hailing, and hotel booking were not invented out of thin air by a single startup. Instacart was not the first company to offer online grocery delivery either. Founded in 2012, it built its advantage by making selection, ordering, picking, delivery, and fulfillment consistently more reliable—not by inventing the idea of bringing groceries to someone’s door. For a founder, the key distinction is this: originality answers “has anyone done this before?” Business answers “why would a customer choose you now?” 1 3

Technical founders often mistake being different for being valuable: because the technology has never existed, customers should presumably pay for it. But customers do not pay for your technical résumé. They pay for their problem. Real innovation does not have to be a new invention; it can be a cheaper, faster, or more reliable way to solve an old problem. Originality can become a moat, but it is not demand by itself.

A hard problem is not necessarily an important one

Researchers like hard problems because the deeper they go, the more likely they are to produce new papers, methods, and reputation. But customers do not pay for research interest. An algorithm may gain ten percentage points of accuracy and excite the technical team while leaving customers cold. What they actually need might be one fewer manual check or a slightly lower operating cost. A hard problem is not necessarily an important one, and complexity is not value.

Chaubard paid that tuition himself. His company, Focal, spent three months training a computer vision model to analyze customer movement in retail stores. Retailers initially all said the problem was important. Once the model was ready, the team told them: at lunchtime, customers cluster around the deli counter and checkout, so queues form. The technology had succeeded. But when Chaubard asked whether they would pay for that conclusion, the answer was a flat no—the retailers already knew it. 1

What would manual validation have revealed? There was no need to train a model first. A few people labeling a handful of videos, or simply observing the store at lunchtime for a few days, could have produced the same conclusion—and surfaced the customer’s “I already knew that” much earlier. This is not an argument against automation. It reverses the validation order: first prove at the lowest possible cost that the information has value, then decide whether it is worth automating with AI.
  1. Confirm that the problem is real and painful enough.
  2. Deliver the complete outcome once with people, scripts, or a semi-automated process, and see whether the customer will pay for the result.
  3. Invest in models and automation only when repetition costs are high enough and quality is stable enough.
  4. After launch, keep calibrating against customer behavior instead of substituting technical metrics for commercial judgment.

AI keeps lowering the cost of building a product. Software that once took six months may now be one of ten prototypes built in a week. As production capacity rises, so does the capacity to produce the wrong thing. If the direction is wrong, greater efficiency produces greater waste. The biggest risk for an AI startup is sometimes not that the product cannot be built, but that it is built, and only then do you discover that nobody wants it.

Praise is not demand; payment is

The claim that customers need you is rarely proven by “that sounds interesting.” Between a good idea and a good business lies real customer action. Finding something interesting does not mean needing it; finding it useful does not mean adopting it; saying you would try it does not mean paying for it. The feedback that matters is not how many people praise the product, but how many will change an existing process and bear a real cost for it.

How much a signal proves

What the customer says or doesWhat it provesWhat is still missing
“That sounds interesting.”AttentionIntensity of need; willingness to adopt
“The technology is impressive.”Technical respectWhether the customer will change behavior
“I’d try it.”Willingness to exploreActual usage and switching cost
“I’d pay for that.”Stated willingness to payWhether they actually pay, and keep paying
They changed a process and bought againValidated value creationScalability and unit economics

Two kinds of approval need to be separated. An investor putting in money shows that they see an opportunity. A customer paying shows that the product has created value. The two are not interchangeable. Chaubard repeatedly insists that founders leave the office and meet customers directly—not only investors, but the places where customers live and work, to hear why they refuse, why they hesitate, and why they eventually buy. 1

Act at 70 percent, but do not change direction every week

Even after finding demand, technical founders face another barrier: perfectionism. Research emphasizes sufficient evidence and rigorous conclusions; in business competition, time itself is a cost. Chaubard cites Bezos’s “70 percent rule”: most decisions that can be tested quickly and reversed should be made with roughly 70 percent of the information you would ideally want. If you wait until you are nearly certain, a competitor may already have shipped the product and won the customer. 4 1

Does the 70 percent rule mean being sloppy? No. In his 2016 shareholder letter, Bezos calls many decisions reversible “two-way doors”: reversible decisions can be made lightly and quickly, while irreversible ones deserve more care. Chaubard adds that decisions involving major safety risks or irreversible consequences require greater rigor. The 70 percent rule applies to things that can be tested quickly and corrected if wrong. Its premise is not “information does not matter,” but “waiting costs more.”

What makes business brutal is that the best technology is not guaranteed the largest market. Chaubard points to the early database wars: Michael Stonebraker’s QUEL had technical advantages, but Larry Ellison’s Oracle brought SQL to market faster, and SQL eventually became dominant. According to Chaubard’s retelling, Ellison said that QUEL might be better, but SQL would win. Victory depended not only on technical merit but on shipping speed, user adoption, and ecosystem accumulation. Engineers ask which option is better; founders must also ask which option is easier to adopt. 1

But move fast without constantly changing direction. Researchers often like to explore many directions at once; a startup is not a sequence of papers. Shipping the product is not the finish line, and neither is the first customer using it. Real company building means improving around one important problem for years. The distinction that is easy to miss: continuous validation is not frequent pivoting. Good founders are patient about the problem and flexible about the solution. 1

Iterate quickly around the same problem

Fast experimentation happens at the level of the solution, not by changing the problem every few weeks. Only when customer behavior disproves the original problem should you return to the start and choose a new one.

flowchart TD
    A["Choose one important problem"] --> B["Form a testable hypothesis"]
    B --> C["Run the cheapest possible experiment"]
    C --> D{"Will customers change behavior or pay?"}
    D -->|"Yes"| E["Keep what works and improve the solution"]
    E --> B
    D -->|"No"| F{"Does the problem still hold?"}
    F -->|"Yes"| G["Adjust the solution or channel and try again"]
    G --> B
    F -->|"No"| A

You cannot outsource commercial judgment

Many technical founders say: “I’ll handle R&D and hire someone with business experience to run sales.” It sounds reasonable. But sales is not just moving product; the act of selling is itself a way to understand the market. Why customers refuse, why they hesitate, and why they eventually buy—those questions contain the product’s real value. If the founder never meets customers directly, they may never learn what customers are actually paying for. Sales can be divided; customer understanding cannot be outsourced. 1

This also requires another ability: expressing value. Research prizes precision; business needs people to grasp value quickly. You may spend half an hour explaining an algorithm when one sentence about how much time it saves would do more. Chaubard cites GEICO’s insurance ad: fifteen minutes could save you 15 percent or more on car insurance. What makes it effective is not a complete explanation of insurance pricing, but a direct statement of the result customers care about. Of course, concision is not exaggeration, and it is never a license to invent facts. Good commercial communication carries the real value the shortest possible distance. 1

But creating value is not enough; founders must also learn to capture it. Many researchers are naturally uncomfortable with commercial competition: they are good at solving problems but not at pricing, selling, competing for customers, and building margins. The market does not hand you revenue just because your technology is excellent. Value creation determines whether you are useful; value capture determines whether you can keep operating. A company can have superb technology and still fail because it cannot price or sell. Commercialization is not an extra credit question after the technology is finished; it is a required question from day one.

Chaubard ends with one of his bluntest lines: nobody cares how smart you are. In academia, you can prove expertise with knowledge and papers. In front of a customer, a complex explanation, advanced metrics, and an impressive résumé are not reasons to buy. What customers care about is how well you understand their problem. Proving that you are smart may earn applause; understanding what someone needs is what earns a customer. 1

What can be delegated, and what cannot? Sales execution, channel operations, finance, and legal work can all be delegated to specialists. What cannot be outsourced is direct customer understanding: who buys, in what situation, for what reason; why they refuse; at what price they will keep buying. The founder does not have to be the best salesperson, but must build that judgment personally. Otherwise the team brings back a revenue spreadsheet, not market knowledge.

The advantage migration trap: when the scorecard changes

Originality, technology, demand, sales, competition, listening—these apparently separate issues point to one deeper pattern. This article calls it the advantage migration trap: when people succeed in one environment because of a certain ability, they instinctively assume that ability matters just as much in the next one. Researchers trust originality, engineers trust technology, managers of large companies trust process. None of those abilities is wrong. But when the evaluation system changes, the strongest past advantage can become today’s biggest obstacle.

How an advantage turns into an obstacle

An old environment rewards an ability; the ability produces success; success hardens a belief. In the new environment, the same action is repeated, but the feedback gets worse. Interpret that as “not executing hard enough,” and you double down—making it even harder to turn.

flowchart LR
    A["Old environment"] --> B["Rewarded ability"]
    B --> C["Success and status"]
    C --> D["Belief: 'this is the right way'"]
    D --> E["New environment"]
    E --> F["Repeat the familiar action"]
    F --> G["Feedback gets worse"]
    G --> H{"How is it explained?"}
    H -->|"Try harder"| F
    H -->|"Reset the scorecard"| I["Learn a new way to judge"]

AI is making this migration more common. In the past, a software company with strong R&D could build a high technical barrier. Now AI keeps lowering the barrier to product development, execution itself becomes cheaper, and competition shifts toward “is this worth doing,” “who is it for,” and “who will pay after it is built.” That is why the stronger AI gets, the cheaper execution becomes and the more expensive judgment becomes. More people will have creative capabilities that once belonged only to elite technical teams. But the spread of creation does not make commercial success easy. AI can help you solve a problem; it cannot prove that the problem is worth solving.

The scorecard has changed

The old high-scoring answerThe new critical question
No one has done this beforeWho will change behavior for it
The solution is more complex and advancedDid it produce a better outcome
I proved I am smartDid I actually understand the other person
The product was delivered successfullyDid customers recognize the value and pay
I gathered enough informationCan I learn and correct faster

From proving yourself to understanding the world

The real danger is not the pursuit of originality; it is treating originality as proof of value. Inventing something new does not mean the world needs it. Solving a complex problem does not mean customers will pay. Having a capability others lack does not mean you have created something irreplaceable. The most important founder shift may not be from a smart person to a smarter person, but from proving yourself to understanding the world. What ultimately decides a product’s fate is not how strongly you proved your own ability, but how well you proved why the world needs it.

  1. Whose problem is this? When did they last spend time or money on it, and in what situation?
  2. How do they solve it now? What exactly is better about my approach—better enough to justify replacing the current one?
  3. Will they bear one real cost—a trial, a prepayment, a process change—rather than just saying “not bad”?
  4. What must be true for this to work? Which assumption can I test first at the lowest cost?
Does this argument have limits? Yes. This is not anti-intellectual, and it does not say technology is unimportant. In deep tech, medicine, aviation, and energy, technical barriers and rigorous validation remain central. Decisions involving safety, regulation, and irreversible consequences cannot use the 70 percent rule as an excuse. More precisely, the argument opposes treating the academic scorecard as the only scorecard in business. Technical ability determines what you can build; customer understanding and value capture determine whether it can last.

Sources

  1. Y Combinator · 10 Habits PhDs Must Unlearn To Build a Startup
  2. Y Combinator · Meet YC’s Newest Visiting Partners
  3. Instacart · About Us and Company History
  4. Amazon · 2016 Letter to Shareholders