How to Start a Startup: Five Founders on What Comes First

Type how to start a startup into Google and you mostly get the 2010 answer: find a co-founder, pick an idea, talk to users, raise a seed. None of that is wrong, and all of it was written before one person could ship a working product in a weekend. So this lesson takes the narrower question, what actually comes first, and puts five founders who are building right now against each other on it: Sam Altman, Pedro Franceschi, Garry Tan, Amjad Masad and Aravind Srinivas. Each is quoted from a different interview, with a visible timestamp, so you can hear the line in its own context. They do not agree, and where they split is the most useful part of this page. For the wider map this sits inside, start with our pillar on AI for startups.

What starting a startup actually takes, in five answers

• Sam Altman, co-founder and CEO of OpenAI, on Relentless: pick a target today's models cannot reach, and plan forward on the assumption that they will.
• Pedro Franceschi, co-founder and CEO of Brex, on The Light Cone: explore broadly, then judge a problem by whether you can state it inside a very small surface area.
• Garry Tan, president and CEO of Y Combinator, on his own show The Light Cone: the first ingredient is a need you personally feel, because that is the one input no model supplies.
• Amjad Masad, co-founder and CEO of Replit, on Y Combinator: start by being honest about what your early signal actually is, because early traction is easy to misread.
• Aravind Srinivas, co-founder and CEO of Perplexity, on 20VC: stop deliberating and begin, because the cost of beginning has collapsed.

Two of the five have no playbook link above, and that is deliberate rather than an omission. This site has no Sam Altman playbook, and Brex is covered through his co-founder at Henrique Dubugras, not through Franceschi.

Sam Altman: pick a target today's models cannot reach

Altman's first move is about ambition, and he is specific about why. His argument is that the market has still not priced in that capability keeps improving, which makes it rational to begin work whose core step does not run well yet. At 4:46:

"the exponential of model progress is going to continue. And it is okay to start working on things now that require smarter or cheaper models."

Sam Altman, at 4:46

He is blunt about how few founders act on it. At 1:09:04, asked by Ti Morse which of his own mental frameworks have gone stale, Altman says this:

"a startup of today still looks mostly like a startup of 10 years ago because that's what the received wisdom says you're supposed to do."

Sam Altman, at 1:09:04

So what: pick a target today's models cannot reach, and plan forward on the assumption that they will. Write down the version of your product that fails on current models, then start on the parts that do not depend on the missing capability. We wrote a whole post on running that play, build for the next AI model.

Pedro Franceschi: compress the problem until it fits a napkin

Franceschi, who runs Brex, is not arguing for a smaller ambition. He is arguing that the test of whether you understand a problem is whether you can state it small. His observation is that the companies that worked started with an absurdly narrow surface: Stripe was an API, Brex was a terminal, Airbnb was a form. At 19:51 he turns that into a test:

"if you can't minimize your surface area and and solve the problem with a very clear set of boundaries, you haven't found the right problem to solve"

Pedro Franceschi, at 19:51

The version he uses on his own team is shorter. At 20:16:

"intelligence is compression. So when someone comes to pitch me an idea in the company I'm like it has to fit in a napkin. Like great ideas fit in a napkin."

Pedro Franceschi, at 20:16

Note carefully what he does not say. He never tells you to delay starting. In the same answer he says you can use AI to compress a problem into a smaller surface area, and calls that valuable. Asked a minute later whether founders should now explore 30 ideas in parallel instead of three, he says to pick a broader universe of things for an early initial exploration and then go talk to customers. So the sequence is explore broadly, then judge a problem by whether you can state it inside a very small surface area. The napkin is a test you keep re-applying, not a gate you have to clear before you are allowed to begin.

Where does the judgment come from, if not from the model? At 21:55 Franceschi is explicit that execution has stopped being the constraint:

"The wisdom to choose is still I think the the missing bottleneck and to me that all comes from which signals are not in the models."

Pedro Franceschi, at 21:55

His reason is concrete. A customer describing a problem gives you a local optimum shaped by their own worldview and constraints, not a prompt you can hand to a model and get a winning product out of.

So what: run the exploration wide, then force the survivor down to one interaction you could draw on a napkin. If it will not compress, that is information about how well you understand it, not a reason to stop working. His companion argument, that the CEO has to own the AI rebuild rather than delegate it, is broken down in go all-in on AI: Brex CEO lessons.

Garry Tan: start from a need you personally cannot put down

Tan spent 13 years not writing code, then rebuilt his old blogging platform in about five days with a coding agent. When he explains why that project and not another, the answer is not market sizing. At 9:16:

"for me, it's like this, you know, mass sort of a shift in technology was happening and then, uh, I had a need and a want and a desire and it was a burning desire."

Garry Tan, at 9:16

He is explicit that this is the part the tooling does not cover. At 18:48, describing where the human sits in his own agent workflow:

"That's where the human, you know, vibe coder operator agentic engineer, needs to supply their understanding of what's going on, what are we building. There's not really a substitute to that."

Garry Tan, at 18:48

So what: the first ingredient is a need you personally feel, because that is the one input no model supplies. Pick the problem you would still be annoyed by if the startup failed. If you are working solo, the mechanics of turning that into shipped software are in solo founder: build like a team.

Amjad Masad: be honest about what your early signal really is

Masad's answer is the least romantic and probably the most load-bearing. Asked directly what he would do differently starting Replit today, at 31:18 he goes straight at self-deception:

"being really honest with yourself about like product market fit is very important. It's very easy to delude yourself."

Amjad Masad, at 31:18

The claim underneath that is epistemic, and it is the part that matters later on this page. He is not saying founders ignore bad news. He is saying the encouraging read is available at every moment and feels perfectly reasonable from the inside. The examples he reaches for are a first user and a first paying customer: both genuinely worth celebrating, and neither one the same thing as a market pulling the product out of your hands.

So what: start by being honest about what your early signal actually is, because early traction is easy to misread. The interviewing discipline that produces evidence rather than encouragement is the whole subject of The Mom Test.

Aravind Srinivas: the cost of beginning collapsed, so begin

Srinivas is the counterweight to all the filtering above. Asked what he would advise the large population of people who are not using AI in their existing work, at 1:02:53 his answer is two sentences long:

"Get started. First step, get started. And and and channelize your curiosity, right?"

Aravind Srinivas, at 1:02:53

That one is about adoption rather than founding. He makes the founding version of the point himself a minute later, describing someone who could never leave a job because starting a company used to mean hiring people and setting up an office. At 1:04:24:

"For the first time in history you can get started on an idea with like one or two other friends"

Aravind Srinivas, at 1:04:24

So what: stop deliberating and begin, because the cost of beginning has collapsed. The office, the hiring plan and the funding round were once prerequisites, and now they are consequences.

Where they disagree: hold the direction, or cut earlier than feels right

Altman and Masad were not asked the same question. Altman was asked which startups he wished more founders would take on, and answered about targets and planning horizons; Masad was asked what he would do differently at Replit, and answered about persisting on a path that looked like it might be working. But their instructions collide at the same decision, how long you stay on something that is not working yet, and at that decision they give opposite orders.

Altman says hold. He has just told Ti Morse he is surprised more founders do not take on the crazy thing with the new tools, and that most are building AI agents for one enterprise vertical instead. At 3:00:

"being willing to truly internalize the fact that scaling laws are going to continue and planning for the things that are not possible or economical this month but will be possible in 2 years, four years, whatever."

Sam Altman, at 3:00

Masad, asked what he would do differently, says he held too long. At 31:48:

"you keep going down that path, but in reality you should have changed directions a little earlier"

Amjad Masad, at 31:48

Put both rules against a single moment. It is month nine, the thing does not work, and you have a story about why. Altman's rule says that is the expected condition of a good bet, and the story is probably right, because the constraint is external and dated. Masad's rule says that story is the delusion, and that he ran it himself at Replit.

The obvious escape is to say they are describing different blockers, capability for Altman and demand for Masad, so you simply diagnose which one you have. That escape is weaker than it looks. Start with what Masad actually claims, which is narrower than the escape needs it to be: his point is about misreading demand signal, that the encouraging read of a first user or a first payment is available at every moment and feels reasonable from the inside. The next step is ours rather than his. That same weakness applies to the diagnosis the escape asks for, because the call between a capability that is not there yet and demand that was never there is made on thin evidence by the person with the most to lose from calling it wrong. Altman's rule, meanwhile, asks you to act for years on a belief the present cannot check. So the escape does not remove the problem. It relocates it into a judgment about yourself.

So there is no free resolution here, and any rule you adopt costs one side something. The most useful one keeps that cost visible:

Write two kill criteria before you start, and keep them apart. The first is a demand test you can run this quarter on today's models, however degraded the product has to be to run it at all. The second is the capability bet you are explicitly exempting from the kill criteria, named out loud, with the date you will re-open it. Then hold the line between them. If the demand test fails, Masad wins and you move, no matter how good the capability story sounds. If the demand test passes and only the capability is missing, Altman wins and you hold to the date you wrote down.

Be honest about what that rule costs each of them. It costs Altman the wildest version of his advice: a bet whose demand cannot be tested at all today, even badly, is one this rule will not let you begin, and some of those are the crazy things he wishes more people would attempt. It costs Masad the clean discipline he arrived at: you are deliberately carving one variable out of the kill criteria, which is the exemption he watched himself abuse. And it can still fail both of them, because a degraded demo can lose a demand test for reasons that disappear the moment the models improve. That leftover risk is how you know the disagreement is real rather than a wording problem.

What they agree on

• Building is no longer the expensive part. All five treat code as cheap and treat the choice of what to build as the scarce input.
• The founder supplies something the model cannot. Tan calls it agency, Franceschi calls it the signals not in the models, and both mean the same thing: the judgment and the need come from you.
• The received startup playbook is stale in its details. Altman says today's startups still look like startups of a decade ago out of habit; Srinivas says the office and the hiring plan are no longer prerequisites.
• Team size is a result, not a starting assumption. Franceschi says he would start by asking why it cannot just be him; Srinivas says one or two friends is now enough to have a real shot.

What to do this week

• Write your target as the version that fails today: the product you would build if models were twice as capable and half the price. Keep it to one paragraph.
• Now compress it. State the first shipped surface as a single interaction, on one line. If it takes more than one, keep cutting until it fits.
• Name the need behind it that is yours. If you would not care about the problem after a failed launch, pick a different problem.
• Write your two kill criteria down before you build: the demand test you can run this quarter, and the one capability bet you are exempting, with the date you re-open it.
• Ship that one interaction to 5 to 10 real users this week rather than spending the week on a plan.

AI Operating System for Startups

Sources

• Ti Morse, "Sam Altman - How to Start a Startup", Relentless (podcast), July 25, 2026. Quotes at 3:00, 4:46 and 1:09:04. This is the Relentless interview, and it is not the Y Combinator Startup School conversation released the same week.
• The Light Cone, "The Most AI-Pilled CEO We Know", Y Combinator, June 10, 2026. Pedro Franceschi quotes at 19:51, 20:16 and 21:55.
• The Light Cone, "Tokenmaxxing: How Top Builders Use AI To Do The Work Of 400 Engineers", Y Combinator, May 8, 2026. Garry Tan quotes at 9:16 and 18:48.
• "Replit's CEO On The Only Two Jobs Left In The Company Of The Future", Y Combinator, April 25, 2026. Amjad Masad quotes at 31:18 and 31:48.
• Harry Stebbings, "Perplexity CEO: Micron Will Be More Valuable Than Meta & How Export Controls Helped Not Hurt China", 20VC, June 15, 2026. Aravind Srinivas quotes at 1:02:53 and 1:04:24.
• How to Start a Startup (CS183B), Stanford, Fall 2014, Sam Altman's own course site as archived on November 1, 2014: "CS183B is a class we're teaching at Stanford", with Altman listed on the schedule for the first lecture on 9/23/14. The class was announced in Fortune, "Stanford to teach 'how to start a startup'", September 16, 2014.