Andrej Karpathy

So I was trying to figure out how to do photorealistic AI photos, and it was ... Stable Diffusion by itself is not doing that well. The faces look all mangled, and it doesn't have enough resolution or something to do that well. But I started seeing these base models, these fine-tuned models, and people would train on porn, and I would try them and they would be very photorealistic. They would have bodies that actually made sense, body anatomy. But if you look at the photorealistic models that people use now, there's still core of porn there, of naked people. So I need to prompt out, and everybody needs to do this with AI startups, with imaging, you need to prompt out the naked stuff. You have to keep reminding the model, "You need to put clothes on the thing."

You've launched a lot of companies and built a lot of products. As you say, most failed, but some succeeded. What's your philosophy behind building the startups that you did? I think my philosophy is very different than most people in startups, because most people in startups, they build a company and they raise money and they hire people and then they build a product and they find something that makes money. And I don't really raise money. I don't use VC funding, I do everything myself. I'm a designer, I'm the developer, I make everything, I make the logo. So for me, I'm much more scrappy. And because I don't have funding, I need to go fast. I need to make things fast to see if an idea works. I have an idea in my mind and I build it like a mini startup, and I launch it very quickly, within two weeks or something, of building it. And I check if there's demand and if people actually sign up and not just sign up, but if people actually pay money. They need to take out their credit cards, pay me money, and then I can see if the idea is validated. And most ideas don't work, as you say, most fail.

You said that there's been a few low points in your life, you've been depressed and building is one of the ways you get out of that. But can you talk to that? Can you take me to that place? That time when you were at a low point? So I was in Holland and I graduated university and I didn't want to get a normal job and I was making some money with YouTube because I had this music career and I uploaded my music to YouTube and YouTube started paying me with AdSense, $2,000 a month, $2,000 a month. And all my friends got normal jobs and we stopped hanging out because people in university hang out, you chill at each other's houses, you go party. But when people get jobs, they only party in the weekend and they don't hang anymore in the week because you need to be at the office. And I was like, "This is not for me. I want to do something else." And I was started getting this, I think it's Saturn return. When you turn 27, it's some concept where Saturn returns to the same place in the orbit that it was when you're born.

So 12 startups in 12 months. Startup number one, what were you feeling? What were you ... Sitting behind the computer, how much do you actually know about building stuff at that point? I could code a little bit because I did the YouTube channel and I would make websites for the YouTube channel, it was called Panda Mix Show. And it was these electronic music mixes like dubstep or drum and bass or techno, house.

Time is limited. Actually, can we go back to the you laying in a room feeling like a loser, I still feel like a loser sometimes. Can you speak to that feeling, to that place of just feeling like a loser? Because I think a lot of people in this world are laying in a room right now listening to this and feeling like a loser. Okay. So I think it's normal if you're young that you feel like a loser, first of all.

Back to startups, you wrote a book on how to do this thing, and gave a great talk on it, how to do startups, the book's called MAKE: Bootstrapper's Handbook. Yeah.

Can you actually explain? It'd be cool to talk about some of the stuff you've created. Can you explain the photoai.com? Yeah, yeah. So, it's like fire your photographer. The idea is that you don't need a photographer anymore. You can train yourself as an AI model, and you can take as many photos as you want anywhere, in any clothes, with facial expressions, like happy, or sad, or poses, all this stuff.

What's your advice for people in general on how to learn all the state-of-the-art AI tools available, like you mentioned new model's coming out all the time. How do you pay attention? How do you stay on top of everything? I think you need to join Twitter, X. X is amazing now and the whole AI industry's on X. And they're all anime avatars. It's funny because my friends ask me this, "Who should I follow to stay up to date?" And I say, "Go to X and follow all the AI anime models that this person is following or follows." And I send them some URL and they all start laughing like, "What is this?" But they're real people hacking around in AI. They get hired by big companies and they're on X. And most of them are anonymous. It's very funny. They use anime avatars. I don't. But those people hack around and then they publish what they're discovering. They took out papers, for example. So, yeah, definitely X.

Robotics is an interesting one for that because there's been a lot of robotics companies that failed, because it's been very difficult to build a robotics company make money 'cause there's the manufacturing, the cost of everything. The intelligence of the robot is not sufficient to create a compelling enough product from wish to make money. There's this long line of robotics companies that have tried, they had big dreams, and they failed. Yeah, like Boston Dynamics. I still don't know what they're doing, but they always upload YouTube videos and it's amazing. But I feel like a lot of these companies don't have, it's like a solution looking for a problem for now. Military obviously uses. But do I need a robotic dog now for my house? I don't know. It's fun, but it doesn't really solve anything yet. I feel the same with VR. It's really cool. Apple Vision Pro is very cool. It doesn't really solve something for me yet. And that's the tech looking for a solution, but one day will.

And then there's the outcasts who go to the bottom of the spreadsheet and they would try to hide messages and they like, "I don't want to be with the cool kids up at the top of the spreadsheet, so I'm going to at the bottom." I mean, but that kind of crowdsourcing element is really powerful. And if you can create a product that used that to its benefit, that's really nice. Any kind of voting system, any kind of rating system for A and B testing is really, really, really fascinating. So, anyway, so Nomad List is great. I would love for you to talk about that. But one sort of way to talk about it is through you building hood maps. You did an awesome thing, which is document yourself building the thing and doing so in just a handful of days, like 3, 4, 5 days. So, people should definitely check out the video in the blog post. Can you explain what hood maps is and what this whole process was? So, I was traveling and I was still trying to find problems, and I would discover that everybody's experience of a city is different because they stay in different areas. So, I'm from Amsterdam and when I grew up in Amsterdam, or I didn't grow up, but I lived there in university, I knew that center is like, in Europe, the centers are always tourist areas, so they're super busy. They're not very authentic, they're not really Dutch culture, it's Amsterdam tourist culture. So, when people would travel to Amsterdam I would say, "Don't go to the center, go to southeast of the center, the Jordaan or the Pijp or something." More hipster areas. I was like, "A little more authentic culture of Amsterdam."

So one of the things you do really, really well is learn a new thing. You have an idea, you try to build it, and then you learn everything you need to in order to build it. You have your current skills, but you learn just the minimal amount of stuff. So you're a good person to ask how do you learn? How do you learn quickly and effectively and just the stuff you need? Just by way of example, you did a 30 days learning session on 3D where you documented yourself giving yourself only 30 days to learn everything you can about 3D. Yeah, I tried to learn virtual reality because this was same as AI. It came up suddenly 2016, 2017 with I think HTC Vive, these big VR glasses before Apple Vision Pro. And I was like, "Oh, this is going to be big so I need to learn this." And I know nothing about 3D. I installed I think Unity and Blender and stuff, and I started learning all this stuff because I thought this was a new nascent technology that was going to be big. And if I had the skills for it, I could use this to build stuff. And so I think with learning, for me, I think learning is so funny because people always ask me, "How do you learn to code? Should I learn to code?" And I'm like, "I don't know." Every day I'm learning. It's kind of cliche, but every day I'm learning new stuff.

What's your philosophy of monetizing, how to make money from the thing you build? Yeah. So a lot of starters, they do free users, so you could sign up and could use an app for free, which it never worked for me well because I think free users generally don't convert. And I think if you have VC funding, it makes sense to get free users because you can spend your funding on ads and you can get millions of people come in predictably how much they convert and give them a free trial, whatever, and then they sign up. But you need to have that flow worked out so well for you to make it work that you need... It's very difficult.

Man, I use GPT-4o. It's amazing. So, I have user input, I have reviews, people can review cities and I don't need to actually sign up. It's anonymous reviews and they write whole books about cities and what's good and bad. So, I run it through GPT-4o and I ask, is this a good review? Is it offensive? Is this racist or some stuff? And then, it sends message in Telegram, it rejects reviews, and I check it and man, it's so on point. It's so- Automated.

One thing you espouse a lot, which I love is the automation step. So, once you have a thing, once you have an idea, and you build it, and it actually starts making money, and it's making people happy, there's a community of people using it. You want to take the automation step of automating the things you have to do as little work as possible for it to keep running indefinitely. Can you explain your philosophy there? What you mean by automate? Yeah. So, the general theory of starters would be that when it starts, you start making money, you start hiring people to do stuff, do stuff that you like marketing, for example, do stuff that you would do in the beginning yourself. And whatever, community management, and organizing meetups for Nomad List, for example, that would be a job, for example.

I love how we're actually highlighting all kinds of interesting problems that all could be solved with a startup. Okay. So, what about the exit? When and how to exit? Man, you shouldn't ask me because I never sold my company.

So, you usually build the stuff solo and mostly work solo. What's the thinking behind that? I think I'm not so good working with other people. Not like I'm crazy, but I don't trust other people.

But it's true. You're also extremely good at what you do. I'm just looking at the interfaces of Photo AI, you would jQuery, how amazing is jQuery? But you can see these cowboys are getting... There's these cowboys. This is a lot. This is a lot. But I'm glad they're all wearing shirts. Anyway, the interface here is just really, really nice. I could tell you know what you're doing. And with Nomad List, extremely nice, the interface. Thank you, man.

Just out of curiosity, what ID do you use? Let's talk about your whole setup. Given how ultra productive you are that you often program in your underwear slouching on the couch, does it matter to you in general? Is there a specific ID you use? VS Code? Yeah, VS Code. Before, I used Sublime text. I don't think it matters a lot. I think I'm very skeptical of tools when people think they say it matters, right? I don't think it matters. I think whatever tool you know very well, you can go very fast. And the shortcuts, for example, IDE. I love Sublime text because I could use multi-cursor. You search something and then I could make mass replaces in a file with the cursor thing and the VS Code doesn't really have that as well.

Take me through the perfect ultra productive day in your life. Say where you get a lot of shit done and it's all focused on getting shit done. When are you waking up? Is it a regular time? Super early, super late? Yes. So, I go to sleep at 2:00 AM usually, something like that and before 4:00 AM. But my girlfriend would go sleep midnight. So, we did a compromise like 2:00 AM. So, I wake up around 10:00, 11:00, no, more like 10:00. Shower, make coffee. I make coffee, like drip coffee, like the V60, the filter. And I boil water and then put the coffee in and chill, live with my girlfriend, and then open laptops, start coding, check what's going on, bugs or whatever.

I got to ask you about, back to the digital nomad life, you wrote a blog post on the reset and in general, just giving away everything, living a minimalist life. Yeah.

You wrote a blog post, "Why I'm unreachable and maybe you should be too." What's your strategy in communicating with people? Yeah. So when I wrote that, I was getting so many DMs as you probably have a million times more. And people were getting angry that I wasn't responding. And I was like, "Okay, I'll just close down these DMs completely." Then people got angry that I closed my DMs down, that I'm not like, man of the people.

You become more and more famous recently. How's that affect you? It's not recently, because it's this gradual thing, right? It keeps going. And I also don't know why it keeps going.

What do you think about EAC, Effective Accelerationist movement? Man, you had Beff Jezos on. I love Beff Jezos and he's amazing. And if EAC is very needed to similarly create a more positive outlook on the future, because people have been very pessimistic about society, about the future of society, climate change, all this stuff. EAC is a positive outlook on the future. Technology can make us... We spend more energy. We should find ways to of course, get clean energy, but we need to spend more energy to make cooler stuff and go into space and build more technology that can improve society. And we shouldn't shy away from technology. Technology can be the answer for many things.

All right. Okay. So you had an incredible life. Very successful, built a lot of cool stuff. So what advice would you give to young people about how to do the same? Man, I would listen to nobody. Just do what you think is good and follow your heart. Right? Everybody peer presses you into doing stuff you don't want to do. And they tell you like parents, or family, or society and tell you. But try your own thing because it might work out. You can steer the ship. It probably doesn't work out immediately. You probably go into very bad times like I did as well, relatively, right? But in the end, if you're smart about it, you can make things work and you can create your own little life of things as you did, as I did. And I think that should be more promoted. Do your own thing. There's space in economy and in society for, do your own thing. It's like little villages, everybody would sell. I would sell bread. You would sell meat. Everybody can do their own little thing. You don't need to be a normie, as you say. You can be what you really want to be.