He Ignored Every Startup Rule and Built a $2.5B AI Company

Erik Bernhardsson (0:00)

I don't really necessarily buy learning from failure, I think you learn way more from success. We sat down for a couple of years and just had a lot of conviction if what we're building will solve a bunch of problems. Which problems? Like, we don't know.

Wouter Teunissen (0:10)

Erik Bernhardsson was Spotify's 40th employee. He built Spotify's recommendation engine as a side project and grew a 400-person engineering team at one of the fastest-scaling fintechs ever. In 2021, he started Modal with just five people and spent two years building deep AI infrastructure. Then, stable diffusion dropped, and everything changed. Revenue went from hundreds to hundreds of thousands almost overnight. And today, Modal is valued at over two and a half billion dollars and is used by companies like Mistral, Ramp and Scale AI.

Erik Bernhardsson (0:39)

Humans are the root cause of everything. Everything bad or good that happens at a company is like downstream from hiring good people that can make those decisions and do the right things. People have all these like random ideas of what they want, but often that's not exactly what they need. Don't build what they're asking for. The whole point of a startup is to come up with a solution that people didn't really think of, and then put it in front of customers.

Wouter Teunissen (0:58)

In this episode, you can expect to learn about how to build a deep AI native company, how to grow a company from 10 to 10,000 people without it destroying itself, and how to decide what problem you should solve, and so much more.

Erik Bernhardsson (1:10)

I love taking a problem that people think is really hard, that kind of scratched the surface, and I was like, yeah, it's kind of hard, but not quite as hard as people think. To me, that's really exciting. What's a moat of a startup? There is no moat, it's just...

Wouter Teunissen (1:23)

I'm super curious, you studied physics, right? That doesn't seem like the natural starting point for the company you're building now. Can you walk me through why you decided to do that?

Erik Bernhardsson (1:36)

Honestly, I grew up coding. I was coding my whole life since I was eight or something like that. And I did a lot of programming competitions. So I kind of felt like, this is maybe me being arrogant or something. But when I started university, I was like, I kind of already know how to code. So I felt like maybe I just started something else. And I mean, I was also into math and physics too. I did a lot of math and physics competitions too, but I was better at coding. And yeah, I actually don't regret it at all that I studied physics. It's a very wide, big sort of toolbox. A lot of different tools that helps you understand the word and world and thinking. Gives you a lot of mental models that are very valuable for a lot of different things.

Wouter Teunissen (2:21)

And would you do the same choice now with AI and software engineering?

Erik Bernhardsson (2:25)

I mean, probably. I think even more so, right? Because I feel like the computer science is even less, in a way, important. A lot of what you study with computer science is like, build a compiler, build a file system. I don't know if you... I think it's helpful to understand some of those stuff. I didn't take it.

You definitely don't need to study things like object-oriented programming. But all this stuff you learn in physics, like control theory, feedback loops, you learn statistics, stuff like that. I mean, especially in Sweden, physics is very math heavy, probably all over the world, actually. But I think all of those things you learn are quite valuable themselves.

Wouter Teunissen (3:01)

How does that translate to the... Because you're in a really interesting position. You've got a pretty technically difficult product, but you're also, as the founder, in a position where you have to manage people, you have to hire people, you have to manage internal stakeholders.

Erik Bernhardsson (3:14)

Yeah, those are all systems, you know?

Wouter Teunissen (3:16)

Okay, interesting. What sort of frameworks do you use then frequently?

Erik Bernhardsson (3:21)

I do feel like there's sometimes a little bit like over-mathemite.

People take them to a physics toolbox and they're trying to apply to everything. Obviously, something as complex as humans, you can't reduce math. But like, I don't know, I think a lot in terms of prioritization, how should we prioritize things? I have kind of framework-y type thing in my head. I think in terms of how do you grow teams, how do you organize stuff, work in progress, how do you minimize that, how do you get shit done quickly? What are the feedback loops? How do you learn fast from customers? All these things, I wouldn't say it's super mathematical, but I still have a quantitative approach to it, and I think it's actually quite helpful.