See pricing
How To Go From Researcher to Founder
Y Combinator Startup PodcastFounders

How To Go From Researcher to Founder

Francois Chaubard spent three months training a people-tracking model that retailers asked for and then refused to pay for. He explains the research habits that quietly kill technical products, and why a human in the loop should come before any model.

November 4, 2026 · 10 min listen · 7 min read · Francois Chaubard
0:00
–:––

Context

Francois Chaubard is a YC Visiting Partner who did research in Fei-Fei Li's lab and later founded and ran Focal Systems, a retail computer vision company. In this short Startup School talk, he lists 10 mindset shifts that researchers need to make to become founders. His argument is that many habits academia rewards, like novelty, rigor and breadth, work against building something people pay for. The lessons carry over directly to PMs on technical or AI teams, where a strong capability can easily be mistaken for a product users want.

The Big Idea

Start from what customers will actually pay for and work backward to the technology, not the other way around.

Chaubard's own costly mistake was building a model before checking that its output had value. His fix is to validate demand cheaply, often with a human doing the work, before spending months or millions on the technology.

Key Insights

1. Say versus pay

  • Academia trains people to chase what Chaubard calls vanity metrics: grades, citations and the number of talks given.
  • The founder question is different: what will people actually pay money for, as opposed to what they say they would pay for?
  • A team may believe users want its new algorithm, but Chaubard says the honest answer is often no.
  • Teams usually only discover this once they start talking to people and asking tough questions.

2. Interesting is not important

Research rewards digging into why something works, because that gets published. Chaubard says that work often has no effect on anyone's daily life. Founders, by contrast, must focus only on what is important to their users. That means accepting an empirical result that works even when the team cannot fully explain why, and shipping it.

3. Novelty is not required

  • In academia, a paper that is not fully original gets rejected.
  • In startups, Chaubard says copying a proven idea is often a good strategy, and novelty is often not even preferred.
  • He cites Joe Lonsdale's split between two startup styles:
  • MIT-style startups start from a brand new invention, a patented black box, and rarely become massive companies because few people want the black box.
  • Stanford-style startups start from one boring customer problem, like groceries or a local hotel room, and iterate on it for years.
  • Instacart did not invent online grocery delivery, since many startups tried it first. It has spent about 15 years executing on that single problem better than anyone.
  • Other pairs he names where a later entrant competed on execution: Lyft and Uber, Anthropic and OpenAI.

4. Validate before you build

At Focal Systems, every retailer told the team that tracking where shoppers go in a store was a major problem. The team spent three months collecting data and training a people-tracking model. The result told stores that shoppers spend a lot of time at checkout and at the deli counter at lunchtime. When Chaubard asked whether they would pay for it, they said absolutely not, because it delivered no value. He says manually labeling a few videos a day would have produced the same conclusion and saved three months.

5. Customer empathy cannot be delegated

Chaubard warns researchers against hiring someone to "do the business stuff" while they keep doing research. A founder is responsible for everything, from hiring and cash runway to sales and marketing. His reasoning is that stepping away from sales and marketing means never building empathy for customers. Without that empathy, a team cannot build a product customers really want.

Mental Models & Frameworks

Human in the loop first

Chaubard's advice for AI companies: before spending heavily on compute to train a model, start with a human in the loop. A person does the work by hand and delivers the output to the customer. If the customer will not pay for the hand-made version, a trained model will not change that answer. Use it whenever a team is about to invest in automation whose value to the customer has not yet been proven.

The 70% rule

Chaubard quotes Jeff Bezos: most decisions should be made with around 70% of the information you wish you had. Waiting for full certainty means a competitor launches first.

  • Example: in the early database market, Michael Stonebraker had a technically superior query language called QUEL.
  • Larry Ellison, founder of Oracle, pushed SQL into the market faster and said he did not care if QUEL was better.
  • SQL won, which Chaubard sums up as speed beating perfect information.

Trade-offs & Nuance

Punchy claims, not false ones

Chaubard says obsessing over rigor in marketing will destroy a company, and points to the insurance slogan "15 minutes could save you 15% or more on your car insurance" as a famously inexact but hugely successful line. He also draws a clear limit: by no means should anyone lie. The trade-off is precision for excitement, not truth for excitement.

Focus versus breadth

Researchers often work on five ideas at once, which suits academic exploration. Chaubard says a startup needs a long-horizon vision and daily work on the same problem for years. Spreading effort across many ideas, or getting bored and switching, distracts the team from its core mission.

Common Mistakes

Mistake: Asking investors instead of users

Chaubard says that talking to people means talking to those who have the problem, not investors. In his view, most investors cannot tell what will be a great product. The better question goes to the person with the problem: if we did this specific thing, would you pay for it?

Mistake: Leading with your intelligence

Chaubard says academics often explain their expertise unprompted, and that this fails in sales, product development and fundraising alike. What works is curiosity about people and listening from the user's perspective. That applies whether the buyer is an IT department, a procurement officer or a consumer.

Practical Application

Fake the model before building it

Before a team trains or buys an AI capability, have a person produce its output by hand for a few real customers. Then ask those customers directly whether they would pay for that output. This is the test that would have saved Focal Systems three months of model work on shopper tracking that retailers refused to buy.

Turn stated problems into prices

When a customer says something is a major problem, follow up with a concrete offer: if we delivered this exact output, would you pay for it, and how much? A stated problem is not the same as willingness to pay, which is the gap Chaubard hit with retailers.

Decide at 70% confidence

For a roadmap or launch decision, write down what information is still missing. If the team already has roughly 70% of what it wishes it had, Chaubard's advice (borrowed from Jeff Bezos) is to decide now rather than wait for a competitor to ship first.

Sit in on sales calls

PMs on research-heavy or technical teams should regularly join sales and customer calls rather than leaving them to someone else. Chaubard's point is that customer empathy comes only from direct contact, and a product cannot be built well without it.

Bottom Line

A strong technical capability is not a product until a customer will pay for its output. Validate that cheaply, often with a human doing the work, then focus on one customer problem and move before you have perfect information.

Resources Mentioned

ResourceTypeWhy it was mentioned
SPIN SellingBookChaubard recommends it to researchers who want to build a founder's sales mindset.
High Output ManagementBookChaubard recommends it for researchers learning to run a company.
The 7 Habits of Highly Effective PeopleBookChaubard recommends it as part of quieting the "academic brain."
How to Win Friends and Influence PeopleBookChaubard quotes Dale Carnegie on being interested in others and ties it to how sales and product work.

Notable Quotes

"I could have manually labeled a few videos a day and given them the same data and gotten the same conclusion." (Francois Chaubard)

"If you abstract yourself away from sales and marketing, you will never develop empathy for your customers." (Francois Chaubard)

"In founder mode, you have to be okay with saying, it just works." (Francois Chaubard)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free