How to Come Up With Startup Ideas: Five Founders on Where Theirs Came From

How to come up with startup ideas is a question that usually gets answered with a list, and a list is the one thing nobody needs now that a model will write you fifty before breakfast. So this lesson asks five founders a narrower question instead: where did yours actually come from? Harshil Mathur, Mukund Jha, Mati Staniszewski, Jorn van Dijk and Varun Vummadi each answer it from a different interview, with a visible timestamp, so you can hear the line in its own context. Their answers do not rhyme, and two of them prescribe opposite orders of operations for the same first month. For the wider map this sits inside, start with our pillar on AI for startups, and for the question that comes after this one, see how to start a startup.

Where startup ideas actually come from, in five answers

• Harshil Mathur, co-founder and CEO of Razorpay, on Y Combinator: start from the problem that blocked you, then check whether it blocked everyone else.
• Mukund Jha, co-founder and CEO of Emergent, on Y Combinator: tinker at the edge of the technology with no objective, and let the idea fall out of what you learn.
• Mati Staniszewski, co-founder and CEO of ElevenLabs, on Sequoia Capital: pick the domain other people have written off as a niche, starting from something you already know is broken.
• Jorn van Dijk, co-founder of Framer, on YC Root Access: work in the area where you are already the expert, and expect the first ideas to be wrong.
• Varun Vummadi, co-founder and CEO of Giga, on Y Combinator: stop grading ideas and find out who will pay, because the idea is not the scarce part.

One naming note before the quotes. Giga was called GigaML when this interview was recorded, and Vummadi refers to it that way on tape. Only Staniszewski has a founder playbook on this site, so his is the one name above that links out.

Harshil Mathur: the idea arrived as an obstacle, not as an idea

Mathur did not go looking for payments. He was building a social crowdfunding side project, needed to accept money for it, and ran into how hard that was in India at the time. What turned a personal annoyance into a company was the second step, which is the part most founders skip. He went into the Bangalore and Pune startup groups on Facebook to see whether anyone else was stuck in the same place. At 3:05:

"I saw almost everyone had the same problem when they were starting up that payments is extremely hard in India."

Harshil Mathur, at 3:05

That is the whole method: a problem you hit yourself is a sample of one, and the check is cheap. He is candid that the answer was not obviously attractive. Payments in India was a hard, regulated business to get into, and he says outright that if he knew then what he knows now, he might not have started it.

Years later he frames what he was doing in general terms. Asked at the end of the same talk what advice he would give an audience of people about to start, at 30:24 he does not talk about ideas at all:

"the core aspect of entrepreneur is the ability to connect with a problem deeply and spend all of your time and effort in solving that problem"

Harshil Mathur, at 30:24

In the same answer he names the failure he thinks the current tooling encourages. At 30:32:

"with AI, my worry is that because it's easier to build, it's easier for people to latch onto problems that they don't really care about"

Harshil Mathur, at 30:32

And he closes on the test he would apply. At 31:18:

"The only thing that's going to work is is there a problem statement that you can spend all of your time and effort for next 10 years solving for"

Harshil Mathur, at 31:18

Read that as a necessary condition rather than a sufficient one, which is how he states it: he is contrasting a decade of company building against how cheap building has become, not arguing that conviction can stand in for demand. He ran the demand check himself, in those Facebook groups, before he had any conviction to defend.

So what: start from the problem that blocked you, then check whether it blocked everyone else. The first half gives you a problem you can actually feel, and the second half tells you whether it is a market or a diary entry. What Razorpay was later defending, once the idea worked, is a separate question we took apart in competitive moats in the AI era.

Mukund Jha: six months of tinkering with no objective

Jha had already built and left one company, Dunzo, when he went into what he describes as pure downtime. He was not searching for a startup idea. He was following whatever new model release looked interesting, ten or twelve hours a day, and at 14:30 he calls that a luxury rather than a strategy:

"So I actually got this luxury of 6 months of like just pure tinkering on things that I really liked with no sort of objective in mind."

Mukund Jha, at 14:30

He is explicit that the absence of a goal is the active ingredient, not a charming detail. At 15:03:

"it was just pure joy of tinkering, pure joy of just building something without any pressure um that sort of led us to sort of think of this idea"

Mukund Jha, at 15:03

What that period bought him was timing. He came out of it convinced that coding was the space about to be disrupted, and Emergent, which he started with his brother Madhav in mid 2025, was at a $120 million annualised revenue run rate about a year later, as TechCrunch reported. His own account of why it worked is about position rather than insight. At 19:23:

"living on the edge, you actually discover these problems much early on"

Mukund Jha, at 19:23

Read that as a claim about information, not about vibes. If you are running the newest models daily, you meet their failures months before the people who read about them, and a failure you met yourself is a candidate problem nobody has queued up for yet. Worth saying plainly: on tape it is the interviewer, not Jha, who first calls this living at the edge, and Jha picks the phrase up. The idea is his; the label started across the table.

So what: tinker at the edge of the technology with no objective, and let the idea fall out of what you learn. This is the least reproducible route on the page, because six unpressured months is a privilege, and it is the reason it belongs next to the other four rather than on its own.

Mati Staniszewski: the annoyance he grew up with, in the domain nobody wanted

Staniszewski's idea is the oldest one here. In Poland, foreign films are traditionally narrated by a single voice reading every part, deliberately flat. At 0:59:

"If you if you watch any foreign movie in Polish language, all the voices, whether that's a male voice or a female voice, get narrated with one single character."

Mati Staniszewski, at 0:59

Everyone in the country knows this. What made it a company was the second condition, and it is a market condition rather than a personal one. On his account 2022 was the year of crypto and the metaverse, with the AI work that was happening concentrated on text and image models, and at 2:56 he describes what that left open:

"audio as a domain was still considered a big niche with so few researchers in the space working on on that work"

Mati Staniszewski, at 2:56

He treats that as repeatable rather than lucky. At 5:06, looking at the field as it stands now:

"there's still so many of those niches that people don't tackle that that you can start with and then step-by-step start opening them them up"

Mati Staniszewski, at 5:06

So what: pick the domain other people have written off as a niche, starting from something you already know is broken. One correction to the tidy version of this story, from the Forbes profile we cite below rather than from the tape: the annoyance on its own was not the idea. Staniszewski and Piotr Dabkowski were already experimenting with AI while he was at Palantir, and Forbes reports that they realised one of those projects, a promising public speaking coach, could be pointed at the lektor problem. By May 2022 they had quit their jobs to work on it full time. So the sequence was a known annoyance plus a model that happened to be able to reach it, which is a less romantic route and a more repeatable one. His playbook on how ElevenLabs was built from there is at Mati Staniszewski.

Jorn van Dijk: stay inside your craft, and expect the first ideas to be wrong

Van Dijk and his co-founder left Facebook intending to start something, without knowing what. His filter was not a market, it was the ground he already stood on. At 12:10:

"if you are either a subject expert on a certain topic or you're just insanely passionate about you know something XYZ is likely a good idea"

Jorn van Dijk, at 12:10

He finishes the thought by saying that spending time in that area gives you a higher chance of succeeding, which is a modest claim and the right size of one.

Then the useful part, which is how long it took. Framer is a design tool today, and that was not the plan. At 13:12:

"it took us a while and it was not not that obvious. I think in that year we were basically testing a bunch of different ideas"

Jorn van Dijk, at 13:12

The discarded ideas are worth naming, because they show what a year of searching actually looks like from the inside. They built a Bitcoin wallet and threw it away, judging it too small and the wrong kind of company to run, because it would have meant a banking licence and the overhead that comes with one. They built a fashion app that let you try clothes on yourself, and threw that away too. The product they kept was the tool they had been using to build the other two: they were prototyping everything, so they made a prototyping product, and designers at other companies picked it up in the first year.

Later, stuck at a few million in revenue, they ran the search a second time, deliberately. At 25:55 he describes talking to customers without pitching anything they had already built:

"just interviewing people and just asking more questions like what's a problem that you have"

Jorn van Dijk, at 25:55

So what: work in the area where you are already the expert, and expect the first ideas to be wrong. The expertise is what makes your third attempt smarter than your first, and it is why the by-product you built for yourselves is worth looking at. If you are going to run interviews like his second search, run them properly: the method is the entire subject of The Mom Test.

Varun Vummadi: the idea is the cheap part

Vummadi is the counterweight to everything above, and he is blunt about it. Giga applied to Y Combinator with an edtech idea, was told by people who had built edtech that it was a bad market, and pivoted a month into the batch. Several more pivots followed before AI customer support stuck. Asked how young people should put themselves in a position to find great startup ideas, he opens with what he got wrong. At 15:02:

"The biggest thing we made at least a mistake even after getting into YC is uh we worked on a lot of stupid ideas, which didn't make any revenue or doing anything uh for a long time"

Varun Vummadi, at 15:02

Then, at 15:17, he disqualifies the question itself:

"people just want have like a lot of ideas. I can just go into ChatGPT and get like 10 ideas on what to do, right? Uh it's never about the idea. It's about if somebody is willing to pay you money for it."

Varun Vummadi, at 15:17

His replacement test has one input, and he scopes it carefully when the host pushes back that a student has nobody to charge. At 16:08:

"If it's an important enough problem, people would pay. Either with money or with time, I would say. I mean, social media networks are just time. They don't charge you any money. But in general, for any single B2B company, if the problem is important enough, people should be willing to pay money for it. Otherwise, like you're just solving a fake problem."

Varun Vummadi, at 16:08

Take the qualifiers seriously, because they are the difference between a rule and a slogan. The money version of the test is scoped to business software, and outside it he counts attention as the payment. What does not move is the shape: an important problem extracts something scarce from the person who has it, and if nothing is being extracted you are looking at a fake one. The staffing bill that comes with selling that way is its own subject, and we took it apart in the forward deployed engineer.

So what: stop grading ideas and find out who will pay, because the idea is not the scarce part. The uncomfortable corollary is that your attachment to an idea is not evidence about the idea.

Where they disagree: tinker with no objective, or price it before you build

Jha and Vummadi both prescribe what to do with your first months, and they prescribe opposite orders of operations. Jha starts at the technology, with no buyer in the room. Vummadi starts at the buyer, with nothing built.

Jha does not offer his six months as a private accident. Mukund Jha generalises it into a claim about what the absence of an objective produces. At 15:28:

"when you're you're building things just for the pure joy of it, just for the um you know, because because you want to solve a problem, right? I think I think that allows you to go really deep into the problem and bring insights that is otherwise very hard to get."

Mukund Jha, at 15:28

Vummadi prescribes the reverse sequence, and he is describing what his company still does, not only what it learned. At 15:41:

"even for our new products that we build, we make sure that the customer can actually pay it. And we predict like this is the amount you're going to pay and get a commitment from the customer and then go and build it"

Varun Vummadi, at 15:41

You get one first month and you cannot spend it both ways. Jha's route cannot produce a priced commitment before building, because in an objectiveless period there is no product to price and no customer named yet, and naming one is the objective he says does the damage. Vummadi's route forbids the objectiveless period by construction, and he says as much when he names his biggest mistake: working on ideas that made no revenue, for a long time, even after getting into Y Combinator (at 15:02, quoted in his section above).

Before writing this section we tried to make the conflict dissolve, because a resolution that costs nobody anything is the signature of a split that is not there. The obvious attempt is sequencing: tinker freely first, then price before you build. It does not dissolve anything, because the tinkering period is not a neutral prelude for Vummadi. It is the specific error he regrets, and a rule that schedules it in advance hands him back the months he says he wasted.

The rule we would apply, and what it costs: give the objectiveless period a budget and a closing date, and after that date build nothing without a named customer and a number. Concretely, decide up front how many months of runway the window gets, run it with no deliverable and no objective, and on the closing date every candidate that came out of it has to earn a priced commitment from a specific person before it gets built.

That is not free for either of them. It costs Jha the thing he says is the active ingredient: a window with a budget and a date is a pressure and an objective, and he had neither. He is describing an open-ended period he could afford because he had just left a company, and our rule replaces that with a smaller, timed version of it. It costs Vummadi the discipline he applies uniformly: the rule licenses in advance the exact months he regrets, and it exempts everything discovered inside the window from the pricing test he runs on every new product.

And it can still fail you, in both directions, though only one of those directions has a documented case and we are not going to pretend otherwise.

Start with the payment half, where the case is real and it is on this page. Framer's prototyping tool passed the payment test in its strongest evidential form: named customers, real money, designers at Dropbox, Twitter and Microsoft inside the first year, and four to five million in revenue. It did not pass the temporal form Vummadi actually prescribes, since they built the product first and sold it after, so read this as a limit on what payment evidence proves rather than as a counterexample to his sequence. Then the ceiling arrived. Framer then went to the other designers inside those same companies looking for the next ten times the business, and at 22:16 van Dijk describes what came back:

"a lot of like soft reasons I think for for us just not really hitting product market fit"

Jorn van Dijk, at 22:16

Paying customers had told them the truth about the product they already had and nothing about the size of the market for it, and the search had to be run a second time from scratch. A number from a named customer is a floor, not a verdict.

The other half of the rule, the closing date, is where we have to be careful, because the example that looks perfect does not survive checking. At 20:18 Mukund Jha describes the problem he decided to skip:

"there were like at least 20 or 30 YC companies that were solving the exact same problem, JSON parsing, right? And we took this view that, hey, like, you know, like the next model will be able to solve this. So, we just completely skip that problem"

Mukund Jha, at 20:18

Read that as what it is. He states a bet in the future tense rather than reporting an outcome, and he never comes back to say the bet paid off. He also says nothing about whether those 20 to 30 teams had customers or revenue, so we cannot claim they would have passed a payment test and been misled by it. What the passage does show is the shape of the risk our rule carries: a window with a date on it forces you to decide, on the day it closes, what the next model will be able to do, and nothing inside the window tells you whether you are early or simply wrong. We have no verified case of a timebox killing a good idea, and rather than invent one we will say plainly that this half of the rule is unproven.

What they agree on

• Generating candidates is not the constraint. Vummadi can get ten ideas out of a chat window, van Dijk built and discarded two products, and Jha ran a new idea on the board every week for three months. Nobody on this page ran short of ideas.
• The idea comes out of contact, not out of a session. Every route here involves being somewhere specific: inside a broken payment flow, inside a new model, inside a craft, inside a domain nobody else was in. What they do not agree on is which contact you owe your first months to, which is the split above.
• The first version is usually not the company. Razorpay started with education fees, Framer started as a prototyping tool, Giga started as edtech. All three moved after contact with real customers.
• Position beats originality. Staniszewski's problem was famous in his country and Jha's insight was available to anyone running the same models; what was scarce was standing there early.

What to do this week

• Write down the last three things that blocked you personally in your own work, with the date each one happened. Do not filter for market size yet.
• For the most annoying one, spend an hour finding out whether other people hit it too. Search the forums and group chats where your kind of person complains, the way Mathur searched startup Facebook groups.
• Pick the domain in your reach with the fewest serious builders in it, and write one sentence on why it is thin: too small, too boring, or too early.
• Take the strongest candidate and ask one real potential customer what they would pay for the fixed version, before you build it. A number, from a named person, this week.
• Write down your own ten-year answer and the date you will check the payment test against it, so future you cannot quietly move either one.

AI Operating System for Startups

Sources

• "Harshil Mathur: AI Is Compressing Every Moat", Y Combinator, May 6, 2026. Harshil Mathur quotes at 3:05, 30:24, 30:32 and 31:18.
• "Emergent: How Six Months of Tinkering Led To A $100M ARR Company", Y Combinator, June 6, 2026. Mukund Jha quotes at 14:30, 15:03, 15:28, 19:23 and 20:18.
• "ElevenLabs' Mati Staniszewski: How Voice Becomes the Interface for AI", Sequoia Capital, May 6, 2026. Mati Staniszewski quotes at 0:59, 2:56 and 5:06.
• "CEO of Framer: Why Designers Should Become Founders", YC Root Access, February 11, 2026. Jorn van Dijk quotes at 12:10, 13:12, 22:16 and 25:55. On the title he is called co-founder here, because that is the claim every source agrees on. The sources disagree about the rest: the episode and Y Combinator's own promotion of it call him CEO and co-founder, while TechCrunch in August 2025 named Koen Bok as the CEO and third-party company data lists van Dijk as CPO. We found no announcement of a transition either way, and his own site says only that he runs Framer and Festina with his co-founder Koen.
• "Why Two IIT Engineers Turned Down $550K Jobs To Build A Startup", Y Combinator, May 29, 2026. Varun Vummadi quotes at 15:02, 15:17, 15:41 and 16:08. The company is now called Giga; he refers to it as GigaML on tape.
• TechCrunch, "Indian AI coding startup Emergent becomes a unicorn with $130M Series C", July 15, 2026. Body read: it calls Mukund Jha the co-founder and chief executive of Emergent, puts the company at an annual run-rate revenue of $120 million with more than 200,000 paying customers, and says he started it with his brother Madhav Jha, the CTO, in June of the previous year.
• Forbes, "How A Tiny Polish Startup Became The Multi-Billion-Dollar Voice Of AI", December 1, 2025. Body read: it describes the Polish lektor convention, carries Staniszewski's own line about it, and reports that he acts as chief executive while noting the firm uses no traditional titles.