Replit CEO 谈编程的职业生涯、AGI 与 Vibe Coding 之战——与 Amjad Masad、Dave B 和 Salim 对谈

Throw away most of the advice that your parents and your society sort of give you cuz it's such a dynamic world right now and conforming is the worst thing that you could do. I'm John is founder and CEO of Replet which was a YC 2018 company. I am not a programmer. I am not a coder but I can now create software. I came to Silicon Valley to start the company and that was Replet. We knew for a fact that there are a lot of people a lot of ideas and there's so many barriers in their way. the barrier of programming and the power how hard it is to to write software. We're very quickly going to get through all of the code out the legacy garbage part of society. The idea is to be able to talk ideas into creation. Programs are no longer static. You can buy a machine with zero software in it and you can vibe code all of it by just like talking to the machine. The career of the future is entrepreneur. That's the only career that I think is going to survive. If you have access to computers and the internet, you should be able to build something great. And we're going to find talents all over the world. Those who are having the most impact will rise to the top. What a profound opportunity for everybody in the world to make a massive difference. Now that's a moonshot.

Welcome, everybody. Allow me one second to bring onstage my 2 moonshot mates, Dave Blundin. Dave is the creator and CEO of Link Ventures. Link Ventures runs a little over $1 billion out of MIT—no comment to Stanford here—investing in companies in the AI space. An amazing guy, he built one of the very first, if not the very first, neural network machine-learning companies, exited for $1 billion, and it's been uphill since then. He and I and Mike Saylor were roommates together at our fraternity, so it was quite a corner of the house.

Salim is over here, another moonshot mate. He's the creator of Exponential Organizations. He was my, effectively, co-founder along with Ray Kurzweil at Singularity University and has been the author of a number of extraordinary books. We'll learn more about him. Amjad, a pleasure to have you here. You're the founder and CEO of Replit.

Nice to meet you. It's such a pleasure for me.

If I were to summarize your life, I would say: coding in an internet café in Jordan, rejected 3 times by YC, finally gets in, builds Replit, and is now helping 1 billion people code. Is that a—

That's right. That's a good summary.

Not bad. One of the things that Dave, Salim, and I talk about a lot, because it's what we care about deeply, is helping entrepreneurs take moonshots and having a moonshot that's part of a massive transformative purpose. We talk about the idea that an individual today can possibly impact the lives of 1 billion people.

Mm-hmm.

What an extraordinary world to be in. I love the fact that your moonshot—and it sounds like your massive transformative purpose—has been to enable 1 billion people to code. When did you first have that vision?

I used to go code in internet cafés back in Jordan. One thing I realized was that they had computers all over the place, but they were not using software to manage the system. Everything was paper and pen, and I decided to build software to manage the store. That was my first company.

The thing that I noticed was that the idea was so powerful, and I felt like I could build it very quickly. It took me 2 or 3 years to get the software done, and it took me another few months, up to a year, to figure out how to actually deploy the thing: how to burn it onto CDs, how to get it to the café, how to install it there, and how to manage the thing.

I felt there was such a big barrier between having an entrepreneurial idea and actually deploying it into the world. That's when I started thinking about how to make that process easier.

Can I show you how easy it is? Can we get the slide up on the screen here? This is me this morning. I'm flying from Santa Monica Airport. I've got the laptop open in my lap, obviously. You can see a Starlink antenna on the dashboard over there, and I'm coding on Replit on the way up here. This is the future.

Yeah, that's awesome.

I tweeted this out. Hopefully, Elon reaches out and talks about the Starlink–Replit partnership. I think that would be perfect. But I said, “Life doesn't get any better than this: the ability to create on demand, anywhere, at any time.”

Did air traffic control know about this?

That's a win. Hopefully, Elon reaches out and talks about the Starlink–Replit partnership. I think that would be perfect. But I said, “Life doesn't get any better than this: the ability to create on demand, anywhere, at any time.”

And so, in college, I ran into another problem: every time I went to do a little bit of programming in a different language, I would have to install the development environment again and again and again. I thought, “We're moving everything to the web, everything to the internet, except programming.” So I started working on what would become the first browser-based coding sandbox.

That went viral on Hacker News and GitHub. I remember one of the important moments was Brendan Eich, the inventor of JavaScript and part of the Mozilla browser, tweeting about my project. Here I am, a 20-year-old kid in Jordan, in my parents' basement, and the inventor of JavaScript—the language that I used—is talking about my application. That was huge for me.

Then I started getting these job offers. Peter Norvig had just started to create Udacity, and he reached out to me: “Hey, we're trying to integrate this program you built so we can teach people programming in the cloud.” Then another company called Codecademy started up on the technology that I built, and Codecademy became the number 1 coding school in the world. How many people here learned how to code on Codecademy? A lot. I thought it didn't work, but we taught 15 million people how to code.

Afterward, I left the company and sat back, reflecting on my own journey. If a kid from Jordan is able to come to the US and invent something that affects millions of lives, there are probably millions of other kids like that all around the world. There are, and so that became my mission: how do you make it so that anyone can build something that could impact millions of people?

Love it. Dave, where's yours?

So, really, really interesting. The journey from Jordan to Silicon Valley, I think, is worth spending 1 second on because I have a follow-up question on that. But you skipped over that in your storyline. How did you end up actually physically here? Because I know you live here now.

I got an O-1 visa to go to New York, and I was a founding engineer at Codecademy. Then I left Codecademy, and I couldn't work at the time. I had 60 days to stay in the country.

By the way, can we talk about how important the O-1 visa is and how critical it is? One of the things I've been tweeting and saying is that we should be stapling a green card on top of every degree that comes out of Stanford, MIT, or Harvard to enable those entrepreneurs to stay here, work here, and contribute to the US economy.

I think this part of the story is really important with you in particular, because you didn't come over to go to Stanford or to go to one of the universities here. You had already finished university in Jordan.

Yes.

Yeah.

A lot of my colleagues, a lot of other people from my country, typically find their way into the US via university.

Yeah.

But I actually wasn't a great student because I spent all my time hacking and programming, so my grades were kind of crappy. I had to come through the open-source route.

I heard the story of you hacking your university—

Yes.

—to change your grades.

Yeah. I felt like it was unfair that—

Better test anyway, isn't it? I mean, that should be the actual grade.

Yeah. I was actually getting failed for attendance because I wasn't showing up.

Because you were doing work.

I was doing work. Actually, one thing I tried to do—I had a Nokia Symbian phone—and I thought, “Can I code on the desk so I can do something more interesting than listening to this teacher blabbering?”

I love the fact that you hacked the Y Combinator application and put a Rickroll video in the application.

That's true degeneracy.

Right?

Yeah, that's right. There's this feeling of these structures, these systems that we built around civilization, that are really meant to make people conform.

Yes.

Put you in the box.

Yeah, put you in the box.

I always felt like the first time it worked was when I thought differently and did something different, and that allowed me to go through the back door, which I think is a more meritocratic way of doing it. I just embraced that, and partly, it kind of always works out. I don't know if I'm just lucky, but even quitting Codecademy, not knowing what to do, I got bored and wanted to do something else. I had only 60 days to stay in the US, and I went and applied to a bunch of companies.

One company I was really excited about at the time was Facebook. Mark Zuckerberg was talking about the vision of Internet.org: the idea that we were going to connect everyone in the world, and everyone would have access to the internet. That’s an extension of my mission: if you have access to computers and the internet, you should be able to build something great, and we’re going to find talent all over the world. So I applied to work at Facebook, and the lawyers told me, “Well, you have to go back to Jordan so we can apply for a new visa.”

I was like, “Screw you. I’m not going to do that.” On my birthday, I got a green card in the mail because I had applied for one, and something worked. The idea is that if you take a chance, the universe kind of rewards you for that. Don’t be too stupid about it, but I think that’s kind of how it works out. I don’t know if I’m just lucky, but even quitting Codecademy—I didn’t know what to do. I just got bored and wanted to do something else.

That is kind of a scary story, though, because that green card could have gone in any direction. That’s totally random.

Yeah. And yet here you are, and here’s Replit. Think about what a difference that makes for the world. When I joined Facebook, I was sort of a nobody, but I was really excited about this idea of Internet.org. I started working on Android because I thought Android was the main device that most people were going to use to connect to the internet.

Yeah.

I ran into this problem with Java. Java had the job of being this really crappy—no offense to any really crappy programming language—resource hog, and it was really hard to program. Then I happened upon a new idea: Can we make the mobile platform more programmable? I started working on that and found other people at the company working on it. That became React Native, and React Native made it so that anyone could write code once and run it on all the different devices.

Now that has also touched billions of people, because part of Facebook is written in it, Discord is written in it, and so on. I think the idea of not overplanning your life, just going through life, finding interesting problems to solve, and solving them will usually net out to the right thing.

That experience, that life experience, is a perfect segue into my question: using Replit to discover talent all over the world. There are 2 things I’m dying to ask you. One of them is this: there are 2 meta-questions. I can use my own product to discover and recruit talent, and I can use my own product to build my own products. We’ll get to the second one later.

It’s really unique to products like yours that you can actually see a billion people, or whatever the number is, coding, right? You being discoverable in Jordan through that vehicle means you can now make that a much more scalable process, because there’s talent all over the world.

And it’s not naturally going to include a grade-point average, a degree, or whatever. So how are you going to find those incredibly talented people? Oh, wait—they’re right here using my product every day. I can see them writing incredible code and building incredible products.

Our first employee—we were just my co-founder and I. My co-founder is my wife, and we were struggling to hire at the time. We weren’t in Y Combinator yet, and Y Combinator rejected us 3 or 4 times. That’s why they had to get Rickrolled.

Wait, was it 3 or 4? You’re not going to forget that.

I actually kind of forgot, because I would apply every time. I would record a video every time. You know, the YC videos are like a hostage situation. You’re sitting there. How many of you have applied to Y Combinator 3 or 4 times?

I love that idea that you have a discovery engine for talent. One of the things that’s very important today is that talent is global, genius is global, and the tools are available and democratized globally. It’s amazing how fast it has changed.

I remember talking to Dave earlier about when we were both undergrads at MIT. There was a course called 6.111.

6.111.

I would literally build my computers with AND gates, NAND gates, and NOR gates, and hexadecimally code them for finite-state machines. That was coding, and that was hard. It’s amazing how far it has come.

You only need a NAND gate, by the way. But we eventually got to EPROMs, and it was extraordinary how far it has come.

I want to talk about the future of coding for a second. I had Emad Mostaque on my Abundance stage 2 years ago, and we were talking about this. He was saying that coding would be done by AI systems, and the front page of all the newspapers in India was, “Emad Mostaque announces coding is dead.” He got a bunch of hate mail for that.

What is coding in the future?

I think that to understand coding in the future, we have to understand it in the past. In the early era of computing, maybe in the 1930s and 1940s and a little bit into the 1950s, computers were fancy calculators. In order to change the programming of computers, you had to literally rewire them. There were no programs.

Turing, in his 1936 paper, invented the Turing machine and showed that you can build a universal computer. But it wasn’t until von Neumann in 1945 that he invented the concept of a stored-program machine, which is what you have in your pocket today. Every computer uses a von Neumann architecture.

That was a huge leap forward. The idea that you could program a computer was suddenly a big thing, because computers weren’t programmable. People forget that. Then you had another leap, when Grace Hopper invented the compiler.

One interesting quote that is actually very reminiscent of this era—and I’m sure she got a lot of hate mail for it—is that we used to program in machine code, but it’s not really programming. It’s similar to rewiring the machine, and she wanted people to program in English.

She called the C programming language and the high-level programming languages that she invented “English.” It’s similar to what Andrej Karpathy recently said: “English is a hot new programming language.” It’s not hot or new.

Grace Hopper had actually thought about it, and she was like, “The specialists are not going away, but we’re going to introduce programming to millions of people, because programming is going to be in English.”

Yes. Of course, that vision sort of went away, and programming became an industry. You had to go to Stanford for 4 years to be able to do it.

But now we’re at a moment, I think, as big as the invention of the compiler—perhaps even bigger—and as big as the invention of the stored-program computer. Programs are no longer static. They’re malleable. You can buy a machine with 0 software in it, and you can vibe-code all of it by just talking to the machine.

That’s been the vision of computing since the start, which is why I think vibe coding undersells that vision. The idea is to be able to talk ideas into creation.

But your idea generator gets compiled into code?

That’s right. That’s right. What ultimately matters is thinking clearly, being able to break problems down into individual components, and then being able to communicate them clearly.

I grew up doing C++ and Pascal coding, and you learn a certain structure: navigating data structures, thinking through and architecting the overall software before you start building it so you’re more efficient, and then going around optimizing everything.

As people are vibe-coding, those skills aren’t there. It’s the same argument we used to make with assembly language: Who wants to be writing assembly, machine language, or hexadecimal? We’ll just do it at a higher level, and you’re then articulating those ideas at a higher and higher level. Is that essentially the direction you see this going?

I think so. Every time someone has made an argument in the history of computing that something is too high-level—that you’re abstracting away the NAND gates, the registers, and all of that, and you’re not going to be able to do good work because it’s too high-level—it has been proven wrong. Every time, what we thought was too high-level becomes low-level.

Now C programming is low-level. At the time, the machine-code people were saying, “What is this? This is [__]. This is [__].”

Interesting twist on that, though: I curse you for not being around when I was doing software development.

That’s all I’m going to say.

Just curse.

I didn’t know Grace Hopper invented the compiler. If her original vision was, “You’re going to do it in English,” then you look at where Python ended up, or where high-level languages ended up, and you can’t really easily specify persistence, pointers, indirection, or stuff like that. So it kind of went to some middle ground.

Now, with Lovable, you can build stuff like boom. With Replit, you can build stuff like boom. And I’m sure the thing you built on your plane worked, right?

It did. I built an app that would assess my mindset every morning and then upload it. I would upgrade it based upon a series of prompts. I have a book I’m finishing called Mindset Mastery, and I wanted that to become something that would just be usable for myself. It was great. It was beautiful. Actually, in addition to being great, it was beautiful.

I want to hit something that I think is important. Right now, there are 150 million GitHub accounts. I looked at the numbers, and if you look at programmer salaries over the last 3 years, they’re up 50% on average. At the same time, the number of programmers has gone up 50% over the last 3 years, while the average programming salary increase has gone up 24%.

That’s counterintuitive if there’s an oversupply, which tells me that the value of programmers is massive, and the exponent is greater than 1 on this. What happens when we have 1 billion programmers, 2 billion programmers? What’s your vision for that world of abundance?

You read Abundance early on. Thank you for your kind words. You said earlier that it is one of the mechanisms for unleashing creativity and increasing entrepreneurship. We talk about it all the time: The career of the future is entrepreneur. That’s the only career that I think is going to survive. I don’t know if you agree with that or not.

Yep.

So what is it like when we have 1 billion or 2 billion programmers and coders in the world?

Yeah. I think the world will trend into a more meritocratic society, and not because there’s someone who has an ethical vision of the future and they’re going to impose meritocracy. I think meritocracy will impose itself because those tools are becoming more available to more people. Those who are having the most impact will rise to the top naturally, right?

I think that’s true in every aspect of society. We talked about our YC story. When we entered YC, it was ultimately because Sam Altman and Paul Graham saw Replit on Hacker News.

So, just to tell the story a little bit—which I’ve heard—you get rejected 3 times. Paul Graham and Sam basically reach out to you. They’ve seen you on Hacker News, and they’re like, “Okay, you’ve got to apply again.”

Yeah. So what happened was, I woke up one morning in late 2017. We had started the company in 2016, although it had been a project for a lot longer than that.

Overnight success after 9 years of hard work.

We’d been grinding. It was 2 or 3 of us. We were selling to schools and whoever could buy our software.

Mom, would you buy this?

Anyone, really. We were door-to-door programming-environment salesmen.

I got this Twitter DM, and it was Sam Altman, with his lowercase username, saying, “Hey, I run YC.” I was like, “Dude, I know who you are.” He said, “We’re interested in what you’re doing. We’d love to meet.” Then he gave me this address. I was like, “That’s not YC’s address in Mountain View.”

So I showed up there, and it was Neuralink and OpenAI. I didn’t know about either. Maybe I’d heard about OpenAI a little bit. I went into OpenAI, and Sam was sitting there. He said, “We’re really interested in what you’re doing. By the way, Paul found you on Hacker News.”

At the time, I was building the system and blogging about it. I’d written some interesting blog posts. Sam said, “You should go see Paul Graham. Unfortunately, he’s retired. He’s not in the Bay Area. You should go to London and meet him.”

I was like, “Okay, let me grab my private jet and go.” The thing about rich people in Silicon Valley is that once they’re rich, they think everyone else is rich, too. I couldn’t even get a visa to go there. So I said, “How about you give me his email first, and we’ll go from there?”

I started an email relationship with Paul. For 2 months, we wrote back and forth.

Paul’s a great writer.

He’s a great writer. At some point, I’ll get his permission to publish the writings that we had. He would write me essays about our shared vision for how programming should be.

The reason we started YC is that we knew for a fact that there are a lot of people and a lot of ideas, and there are so many barriers in the way—not only the barrier of capital and the barrier of getting attention, marketing, and all that, but even before that, the barrier of programming and how hard it is to write software.

We went through all that, and at the end, he was like, “Yeah, I think YC is starting next week. I’m going on a trip, but you should talk to Sam about going into YC.” I was like, “Okay.” So I emailed Sam, and he was like, “Yeah, yeah, I would love to have you in. You should join.”

I said, “We have the kickoff tomorrow.” He said, “Okay.” Then he sent me another email: “As a formality, you should still apply.” I was like, “Ah, [__]. I’m not going to go through the YC application all over again.”

How long is the YC application?

If you want to do a good job at it, it’ll take you an hour. I’m very lazy.

You couldn’t have had Replit do it for you, huh?

No. At the time, I couldn’t. There was no LLM, so I had to do it manually. I did a very lousy job at it because I was thinking, “Hey, guys, you’re recruiting us. I’m just not going to put in the effort.”

Then the video came up, and I thought, “Shit, I’m not going to record the video.” So I pasted in a link, and we went to the next day. The next day was the kickoff. They did the late interviews. They always have this tradition of letting a few startups in on the same day.

We waited the whole day, and finally the door opened and they told us we could come in. We went in, and it was Gustaf Alströmer and Michael Seibel, who was the CEO of Y Combinator at the time. He was a big guy, right? I was shaking their hands, and I felt Michael squeeze my hand a little bit. I don’t know what was going on there.

The moment I sat down in the chair, his face was red. He was like, “Why did you Rickroll us?” It turned out that while we were sitting outside, they were getting Rickrolled inside. They clicked on the video and got this Rickroll video playing. If you guys haven’t seen it, just go Google it.

Yeah, there’s a meme. You can find it easily.

He was very upset, and he started giving us a really hard interview. The questions were really good.

What were you thinking at this point?

I thought we were screwed. We were done. I actually exited the room and ordered an Uber. My co-founder was like, “We have to wait.” I was like, “We’re not getting in. After we did that, we’re not getting in.”

But then I got a phone call from Adora Cheung, and she said, “You got in.” So we went in that same day, and Sam Altman did the kickoff and all of that.

How important was the YC experience to you? Would you have accelerated and gotten to this point without it?

I think one property that I have is that I would never quit, so I would have kept going. I think we probably would have made it, but YC probably accelerated our progress by years. It opened a lot of doors, and just being around these amazing people—being around Sam and Paul—you raise your ambition a lot by being around ambitious people.

The Silicon Valley thing in general.

I think that’s what matters. The question about talent is very important. There’s talent everywhere, but I think the density of the network at places like Stanford is what matters.

Paul Graham has that great quote that being a startup founder is only about 1 millimeter away from being unemployed. So when you’re telling your parents what you’re doing, you need like-minded people all around you as your support network. Otherwise, it’s really difficult. Doing it in Jordan would have been emotionally difficult, and you obviously pulled it through.

Just to put a pin in that story, it’s an incredible story. If anyone from the Stanford administration is in the room, I know Eric Bolson[?] and others are saying your application process is a complete failure, but the pull process of discovery is a complete win.

When you’re building your own company, you see all these people and what they’re doing. You don’t need them to apply to come and work for you. You have the data. You can analyze the data and pull. The world is clearly going to move from applications to big-data-based pull.

To discovery.

To discovery.

Yeah. And that’s what I’m saying. The world is going to be forced to be more meritocratic.

Yes.

I think the problem with institutions like YC is that they’ve gotten better, and Stanford is very much about pedigree.

Yeah.

You know where you're from. And it's not—I’m not faulting them for it; it's a heuristic.

Yeah. If you're a Stanford dropout, Stanford did the work for us to choose you, and therefore we're just going to piggyback on that. But in reality, talent comes from everywhere. Philip Rosedale, the creator of Second Life—I don't know if you know him.

One day, Philip says, “Do you know why there are more startups in San Francisco than any place else? Why is the success rate so high?” And he goes, “I want to tell you what I found out.” So he ran a script on top of LinkedIn where he measured the density of technical founders per square kilometer. Yeah.

And he found that San Francisco had the highest technical density.

Right. There was Austin, there was Cambridge, and Miami starts to pop out. But the notion is, if you're taking the risk of starting a company and you fail, instead of going back to your mom or dad's bakery to work, your friend down the street has another startup you can go and join. And so it is a density issue.

Yeah. But increasingly, I think that is moving online. I know you're good friends with Balaji and his idea of the network state and all of that. I was in Jordan last week, and I saw a company there that is ostensibly headquartered in San Francisco. They run all their engineering there, but they come here quite often. They meet with VCs, and they incorporate in Delaware. When they go into meetings, it's almost like they're headquartered in San Francisco. San Francisco is going to the cloud in many ways. Yeah, Salim.

So, this evolution of where things are going with the future of software development is pretty profound. But at the business level, I use the analogy of Kickstarter. When Kickstarter appeared, for the first time in business history, you could get market validation for a product without building the product. That was kind of incredible. What you're doing is essentially providing an environment where people can write software, fully finish that software, and then get market validation and figure out what's needed at almost zero cost. That's right.

So this is taking out the marginal cost of all of this stuff, which is what we talk about a lot with Abundance, with Exponential Organizations, Singularity University, et cetera. As you accelerate that, essentially you get to a point where society is just one huge amount of code being written on a nonstop basis. How do you think about the future of society in the context of that?

Yeah, it's a great question.

Let's zoom out here. Huh?

Yeah. So, I think if you think about entrepreneurship, I don't just think about it in terms of quitting your job and starting a business. It's about being able to solve problems in the world and taking initiative without your boss telling you to do that.

We're seeing it a lot in the enterprise as well. We've seen cases where a product manager at Zillow was able to deliver millions of dollars in bottom-line revenue for the company. Although he wasn't an engineer, he was able to increase the conversion of home buyers at the company. Suddenly, everyone heard about that story.

Now we have 500 licenses at Zillow, and everyone is encouraged to vibe-code because they become more entrepreneurial. That person got agency.

Agency. It's how you—

It's the biggest challenge that large corporations have today. They're so structured, and they're so fragile in that structure. I mean, Peter, you talk about this all the time, right? It's enabling your employees.

You try anything disruptive in a big company, and the immune system attacks you, right?

Standard problem.

But here's why vibe-coding is important: You can do it anyway, and you can show the results. The results will speak for themselves, right? This is a huge breakthrough.

Permissionless innovation can happen inside the organization as well. We call this PDI—permissionless disruptive innovation—because in the past, whenever you wanted to do something really disruptive, you had to get permission from the Medici family, the church, the corporation, the government, or an investor. Somebody had to bless you to do this.

Today, for the first time in human history, you can do very disruptive things without any permission. My favorite poster child here is Vitalik Buterin. An 18-year-old kid gets together with a few friends, ignores his professors, and now, boom, you have Ethereum—a $600 billion ecosystem that nobody understands.

And so that's kind of incredible, that you can do that. What a profound opportunity for everybody in the world to make a massive difference.

Yeah.

And what you're doing is now providing the scripting language.

Yes.

To take anybody's massive purpose, as we call it, and start making it an actual reality.

So here's where it gets even weirder, stranger, and more abundant: It is not going to be just you doing the disruption. It's going to be your agents.

You'll be able to program agents that are running in the organization, finding the inefficiencies, and actually making progress on your behalf as well. You can also run a startup that is producing these agents. Let's say you're someone who's worked in HR for 30 years. You know all the HR processes. You have a gold mine in your head—an unmonetized gold mine in your head.

What you do in the near future—we're actually launching this tomorrow—Agent 3, you heard it here first, will allow you to create other agents. You'll be able to create an HR agent.

How long have you been working on that?

It's been about 6 months since Agent 2. Every 6 months, commensurate with the progress in LLMs, we're trying to predict what the next set of capabilities will be that unlocks the next version.

We already know what Agent 4 is going to look like. We take a bet, we try to build it, and over time, the first version of it is crappy, but then the foundation models and everyone else sort of catch up.

Yeah. So, I always tell entrepreneurs: Build crappy products.

So agents can basically spin off their own agents.

There’s going to be an agent—

What could possibly go wrong?

Look, I don't focus on what goes wrong. That's other people's problem. But—

It just—it’s going to be mind-boggling. It's almost like a singularity moment. What's going to happen once we have agents that can also transact with each other? I can have an agent—

Which brings in crypto.

Which has a crypto wallet that can go hire other agents.

Sure.

I'm going to raise my agent to be a trillionaire.

Exactly.

Yeah. We're probably going to be seeing autonomous agents with their own wallets, net worth, and all of that.

One thing I heard you speak about is the future of the company. I want to talk about that because we have a lot of entrepreneurs in the audience and among the people who will listen to this around the world. Most corporations today have been built in a very siloed, very factory-production mechanism.

You've railed against this over and over. John Hagel talks about this. All big corporations are designed for 2 things: They're designed for predictability and efficiency.

Right. But in today's world, it's so volatile. You need to be architected for flexibility, agility, adaptability, and speed.

That's right. Therefore, I actually did a talk with you onstage, Peter, a few years ago in Toronto, and the title of the talk was “The Death of the Corporation.”

We're seeing corporations go to platforms, and platforms go to ecosystems.

Yeah. Right.

There's an economist—I think his name is Ronald Coase—

Ronald Coase.

Ronald Coase. Yeah, the theory of the firm, right?

Yeah. So Ronald Coase wrote a 9-page paper in the 1930s citing that the reason we have bigger and bigger companies is that transaction costs inside a big company are cheaper than outside a big company. He won the Nobel Prize in Economics for that 9-page paper.

In the book that Peter and I just wrote, the second edition of Exponential Organizations, we actually declare Coase's law dead because the transaction costs inside a big company are way higher than doing it on the outside.

Exactly. Exactly. And therefore, big companies as a category will start to decline very rapidly.

That's right. My takeaway from that paper is that full-time employment is a bug of the system, not a feature.

Yeah.

In reality, if you're doing what you're told, that's a bug.

Doing what you're told is a bug as well.

Honestly, employment is a bug. Tell my kids, “Just go out.”

Yeah. So the reason why firms have to hire people, train them, and then put them in a somewhat draconian hierarchy is because the transaction costs of going out into the market tend to be cost-prohibitive.

If you lower that—for example, Uber lowered the transaction cost of looking for a taxi from that to a couple of clicks of a button—

If an agent can hire a human or hire another agent, then we're going to see companies shrink, a lot more entrepreneurship, and a more dynamic market. I think more people will get rich, and I think there's going to be a fast life-and-death cycle of companies. Dave, do you want to jump into your question from earlier about Replit coding?

Do you want the easy one or the hard one first?

Yeah, give me the hard one.

All right. I'd love to get your insight on this very specific topic. We're very quickly going to code out the legacy garbage part of society, right? Right now, you can use Replit to build things so quickly. There's all this garbage software that was cost-prohibitive to build because it was too expensive to hire enough engineers, but now you can just do it.

If we get to the point where you can create 1,000 times more code, or 10,000 times more code, then it's all going to be greenfield. What's the net-new thing we can build that adds value that we just couldn't afford to build before? You have a lot more vision into that than anyone else because you can see all these millions and millions of users and what they're starting to build.

Mm-hmm.

So what are you seeing them build today that gives you insight into what they'll be building in the greenfield future we're coming into?

I think it's agents. The reason we built what we're calling the Agent 3 stack is that it's the next evolution of software. When you look at most pieces of software, especially inside the enterprise, they're trying to solve a very specific problem. Usually, you create a piece of software, and then someone uses that piece of software to solve the problem.

That intermediate step isn't needed. You create an agent programmed and imbued with the domain knowledge to solve that problem, and then you run that agent autonomously inside the organization. That's what we're seeing our users struggle with. They're saying, "I want to build this automation process, but it keeps generating an app."

Maybe I need an interface for the chat app to talk to the agent, but I want to be able to run it based on hooks, time triggers, and so on. We're building all of that, and that's how our roadmap works. We're seeing what people are trying to do. They're looking at AI, they're looking at agents, and they're seeing that an LLM can make good judgments.

I think that's the main thing that happened in the past 6 months to a year: LLMs with agents can do this multistep reasoning, especially now with GPT-5. The scale of the thinking in GPT-5—GPT-5 base is so dumb, but GPT-5 high is so intelligent—means we're able to scale that intelligence. Right now, these agents can be nodes in very nuanced decision-making.

We had some conversation on the stage earlier today about some fairly sketchy research out of MIT that indicated that, in the real-world corporate environment, people are getting virtually no benefit out of AI-generated coding, and you're actually living it. Do you have any estimate of how much of Replit is built by Replit, or what the force multiplier is inside your own organization?

The impact of Replit on our organization is more in how we run the company, even outside of engineering, than it is inside of engineering, counterintuitively.

So you're using Replit to run Replit?

Yes, we're using Replit to run Replit. Our salespeople, for example, will write some software using Replit that uses an LLM to analyze all their call transcripts and find the main reason why people might not buy Replit. They'll take that to the product team and say, "This is the main thing we're missing—this compliance issue."

I really want to touch on this topic because you're probably the oldest guy in your Y Combinator cohort, right?

Yeah. Yeah.

Yeah. And so when you look across all these people who are vibe coding or building tools, they have virtually no large-scale management experience.

They're like, "Yeah, this helps me write code more efficiently." But it's actually even better at managing large organizations: all the planning, all of the things like who did what. They have no experience with it.

I've got about a $2–$3 million-a-week payroll, and it's a gold mine in knowing what's going on across 1,000 people.

The embedded data in that payroll is a gold mine.

It's a gold mine, and it's so much easier than the harder challenge of just SWE-bench.

You know, you're probably the only guy in that YC world—

—that's actually running and has run a large organization and sees its advantage as a management tool, not just as a leaf-node code-development tool.

Yeah, it's almost like engineers using AI is table stakes. I would certainly challenge a lot of those studies that are coming out—I think you called them dubious, and I agree—but it's table stakes. There are so many products you can pick from.

I think where Replit is really shining today in the enterprise is because of the vertical integration we do, and because you don't need prior coding knowledge to do all that. By the way, we're launching all these integrations and all the ways in which Replit can hook into your data infrastructure.

All of that will help you run the company. Your legal team—I mean, the gentleman here, we were talking earlier about his finance team at Dolby using Replit to kind of run the finance team. I think that's really profound, and not a lot of people are talking about that.

You're going to run into some data-sovereignty issues and on-premises issues, because there are a lot of people—we know a medical CEO who got excited by ChatGPT and uploaded all their patient data into ChatGPT, and now they've got massive legal issues around that. You're going to face some of that.

That's an engineering-solvable problem.

Or a lawyer-solvable problem.

A lawyer-solvable problem. What I'm excited about is that the current stack of business software—ERP systems, for example—will take certain specific use cases and automate those. My co-founder, Harish, talks about this a lot. But now you can automate every minutia level use case across a business: all the hundreds of little things that people are doing. Anybody can automate those.

The legal team will automate contract drafting, and they'll just start replicating those processes and getting them out there.

The best thing that happened over the last 10, maybe 20, maybe 30 years is the digitization of these organizations and governments, and the system of record—the ERPs, the HR systems, the CRM, Salesforce, whatever that is. That's the infrastructure for the next set of automation and the invention of the future of work. We're exactly at the right time to be able to build these systems.

All right. You want the hard question?

Yeah, let's go. I want to talk about the vibe-coding wars.

All right, let's go. Or I will.

All right. We'll start. There's the easier and the harder of the 2. Okay, the easier one first: Lovable versus Replit.

Good answer.

Good answer. All right, first of all, if I look 3 years into the future, is it all boats rising with the tide, or is one of you going to kill the other? Are you going to differentiate and become different? Do you think about it every day, or do you not focus on competition? How's it going to—

I think it's already quite differentiated. The reason people conflate all these products is because you can put in a prompt and get something out. Most of our competitors generate a front-end app for you. To connect it to a database, you need to go grab a database somewhere else. To deploy it, you need to go to AWS and deploy it.

Replit actually provisions a database for you, runs the database migrations for you, provisions a production database for you, and handles that separation for you to help with the deployment process. The platform—and the reason we've been around for 10 years—is that we've been building the depth of the platform because we knew that the challenge of programming and making software isn't just about the code. A lot of it is about the infrastructure.

I'm thinking the chart I'm looking at here is full-stack versus front-end only, right?

Right.

And you're at the full-stack side. Then there's nontechnical and technical, and you sit sweetly in the middle.

That's right.

But what about Vercel? I think of them as being a deploy company historically, and now suddenly they're—

Well, in this chart, Vercel's down here in the bottom right.

Vercel, I think, with v0, started as a design tool.

Yeah.

Now I think everyone is seeing that the value is where Replit is, because Replit is winning all the deals.

So everyone's trying to come after you.

They're trying to come after us, but good luck.

Okay. Well, that is a perfect segue into the really hard question. We were over at OpenAI headquarters a few weeks ago with Kevin Weil, and when they launched GPT-5, they brought Michael Truell on stage and made a big deal out of Cursor. But just a couple of weeks prior, they were trying to buy Windsurf.

I was like, “Okay, wait a minute.” You guys are friends today, but in an alternate world where Microsoft didn’t torpedo that deal, you guys would be the worst arch-enemies in the world today. So, you have these foundation-model companies. Are they going to come in your direction, or how is that going to play out?

The really interesting thing about our market is that it’s total war. There are no friends. Really, there aren’t. Arguably, in prior eras of Silicon Valley and tech, there were natural allies that could form. Right now, today, you’re an ally with someone else, and they’re going to come out and attack you tomorrow. You have to be super paranoid if you’re starting a company at this stage.

Are you super paranoid?

Of course. I think competition is going to come from everywhere. Google is one of our closest partners. We spend hundreds of millions of dollars with them, and they have 3 competing products. I still love them for it, and Sundar still uses Replit instead of the other ones. Shout-out to Sundar.

Yeah, I think of this as the Dropbox-Box thing. You have OneDrive, iCloud, and Google Drive, but just by focusing on that one product area wholeheartedly, Dropbox does very well. I think that will be the same no matter what other people do; their interests are so varied.

A rising tide of near-infinite demand.

Yeah, there’s that, but I also remember when I was at Yahoo. You’re managing a big company across 120 different web properties, and you have to allocate resources, et cetera, et cetera.

You were where?

I was at Yahoo, by the way, too, for 7 months.

Yes. We’re going to come over there for you.

Anyway, you’re managing in a big company across 120 different web properties. One dedicated team is always going to beat you.

Right?

Always. By the way, if you said, “Yeah, we want to be just like Dropbox—”

Yeah, Dropbox is a great company. But I think it’s a lot more brutal than those eras because, back a decade or 2 ago, you had this politeness between the big companies. Google would say—Microsoft would say to Google, “Don’t come into the office space, and don’t improve Google Docs too much. We won’t come after search too much.” There was this polite thing.

But anyway, it was all détente.

Well, because everything’s converging on AI, they all settled into their swim lanes. They said, “Okay, Apple, your phones are fine, and Bing will suck forever. We’ll just commit to that.” But now all of a sudden, it’s all-out war because everything is converging on, “Oh, wow, all that matters is AI.”

I 100% agree with Salim on the idea that customer-segment focus is a superpower, and that’s very important. That being said, I do think there’s a potential that this product will differentiate but eventually converge. Replit Agent 3—you’re going to see it tomorrow—is the most autonomous agent on the market. We ran it 2 days ago for 4 and a half hours because, right now, it can test its own code.

You and I were talking earlier about whether AGI is here or not. The thing that would make me feel like you can define the current era of models as AGI is that, if they get real and good environmental feedback, you can run them endlessly and have them actually try to solve a problem. Invariably, what we’re seeing in software is that, if you’re able to spend as many tokens as you can and they’re running in a good environment, they’re solving these problems.

I’ve heard you say that it’s this universal expansion of coding capability that you believe is going to lead us to AGI.

Yeah, I think that—

I think AGI is here already, personally.

Right?

I think we passed the Turing test and didn’t notice. I think we passed AGI and haven’t noticed. We’ll see if we notice.

I call BS on this. We’re talking about something we can’t define, can’t measure, and don’t have a test for.

Yes, that’s why I can. The Turing test had a measurable outcome, right? It was very clear. AGI—and now we’re talking about ASI—we still have no idea what any of that means. Aside from that, I’m fine with the conversation.

I have a question, though. Can I—

But just to answer that, the insight from Turing is that the Turing machine—the computer—is the ultimate problem-solving machine.

Yes.

That scales infinitely.

That scales infinitely.

And that is the milieu in which intelligence will operate in order to solve any problem. Whatever AGI system you can imagine, ultimately it’s going to be writing code to solve problems. That’s my fundamental belief.

Yes. Okay, interesting. Can I take this in a slightly different direction?

Maybe.

Just a show of hands: how many of you at this event over the last couple of days have had your minds completely blown? Just a show of hands. Okay, a bunch of you. All right, I’m going to try and go for one right now.

We have biology, which is essentially software, right? You have 50 trillion cells in the human body, each governed by DNA. Now, essentially, we can edit—

3.2 billion base pairs.

We can now edit the DNA like you would software. Essentially, a human being is now a software-engineering problem. When you apply Replit to biology code, are you seeing that happen? When do you see that happen? What do you think of the—

The name Replit sounds very biological.

Yeah, completely. Is that a coincidence? It’s right there in the name.

Ultimately, the fascinating thing about LLMs, which is similar to DNA code, is how much knowledge they compress. I downloaded an app called full moon on my phone that allows you to download open-source, 1-billion-parameter models because I was on a plane. I like to read books on the plane. I know there’s Starlink on planes right now, but I pretend it doesn’t exist—

Because there are 12 hours when I can read a book in peace.

And so I thought, “I need to look some stuff up, so let me download DeepSeek 1B.” It’s amazing how good it is. Obviously, it hallucinates, and you need to check its work and everything like that. But 1 billion parameters—is that less than 1 GB?

Yeah.

That embeds a large portion of human knowledge. If you run it for longer, like DeepSeek-R1 does, it can solve problems as well.

I can’t wait to download it into my Neuralink chip.

Yeah, exactly. These systems are efficient, and it shows you something about the nature of the universe: there’s so much data, noise, and redundancy that you can really boil things down to very small representations. The essence of things is actually very simple.

Isn’t it mind-blowing? Even if you take a big, big, big neural net, you’ve got about 1 trillion parameters. It fits easily on your laptop, and this is all human knowledge—everything on Wikipedia, everything from the internet—learned through about a $100–$200 million training process. It all compresses down, and it fits on your laptop. The processing is way too slow, so getting it to run—

But putting that aside, just the amount of storage that we have is incredible, and the amount of processing. This is why NVIDIA is worth so much: the processing.

We’ve been so inefficient, because I think the ultimate LLM will probably be about 1 billion parameters.

That’s crazy.

Back to DNA and biology, we’re seeing big results from OpenAI, obviously, and AlphaFold before that. Clearly, we have the data to train, and now, with the way we train LLMs—unsupervised learning—you can find this compressible information out of this massive sea of noise without supervision or data labeling.

I think LLMs will get really good at that. Once they’re able to generate and code biology, the thing for us would be to create these environments. That’s what I see our company as: as a tech company, we’re about creating environments. We want to be the best habitat for LLMs. Right now, the best habitat for LLMs is to be in a virtual-machine environment, but in the future—

One more question to close out the competitive-landscape topic. If I look at the companies that do coding—Replit, Cursor, Windsurf, Blitzy—you guys are probably the only ones that built foundation models because you were there early, so you actually trained these things. I think everyone else just picked up OpenAI or whatever later.

So then you’re like, “Okay, well, we’re not going to spend $100 million to train our own model. It’s way too expensive.” Now we’re on top of the other guys. Oh, wait—now our valuation is in the billions. You’re raising tons of money—it’s confidential; I just found out—so you’re raising tons of money, and you’re not actually planning to consume it, right?

You haven’t spent the money from your last funding, but you have a war chest that’s actually plenty big to train foundation models. Is there a version of your future where you get back into the foundation-model training business?

I think it’s cyclical, and the way it works is that, when we went from GPT-2 to GPT-3, there was a period of time when there wasn’t a lot of progress. The tricks were known, so we sort of knew how to do pretraining. The labs were busy cooking up Mixture-of-Experts and GPT-4 and everything else.

There was a period of time when startups could actually use their data, add a layer on top of these foundation models, and fine-tune them. Replit Code 3B, a 3-billion-parameter model, was a state-of-the-art coding model in 2022 and 2023. That put us ahead of the competition because no one else had access to a model that cheap and that easy to run.

But then the foundation models made another big jump by scaling up the models. I think we're reaching a point where it makes sense to go back into training because we have a certain data set that we think is very helpful, especially for the kind of use cases that Replit does. We're starting to find places in our agent architecture, because it's not just 1 agent; we have so many sub-agents. We have the testing agent, and we were able to build a computer-use architecture that is 3 times faster and 15 times cheaper than the state-of-the-art models from the big labs.

Wow.

I think we're going to continue finding these things until we get the next frontier model. That's what happens in the history of machine learning: the bitter lesson. Then you can pour in data and compute, consume all these use cases, and be open-minded about it.

Some companies were like, "We're going to go, we're going to die. We're going to live or die by training." But for us, we're ultimately trying to solve a problem and pursue a mission, and we'll just do whatever it needs to get there.

I really never thought in a million years that you would have the guts to say that. You have the capital to do it, but I could see that some brilliant MIT alum, Christian Bailey, here and others' eyes lit up as soon as you said that, because I think the mission statement that you just outlined will attract a ton of talent that doesn't want to go someplace boring. They want to do something super.

I want to close us out by going back to your roots. Listening to our Moonshots podcast, when the 3 of us discuss taking moonshots, there are entrepreneurs around the world who are a young Amjad someplace in the Middle East, Southeast Asia, South America, or Central America, who are dreaming big. What's your advice to them? What's your advice to inspire them, and how should they think about their next few years? Do you agree that the next few years are really the game?

Yes. Yes, it's everything. My advice will probably get them into trouble, but I think that's part and parcel.

We've had the conversation: skip university, go build a company.

Yeah. I mean, those are details. I think if you get into Stanford, get in and then drop out. What really matters is throwing away most of the advice that your parents and your society give you, because it's such a dynamic world right now.

Don't do what you're told.

Yeah, and conforming is the worst thing that you could do.

Don't do what you're told.

It'll take real self-programming to actually exit that mindset of doing what you're told and try to really think from first principles about what you find your inner purpose is.

Find your inner purpose. Why?

You know, there's a quote I love: "2 important days in your life: the day you were born and the day you found out why."

Yeah.

Yeah. I think the meme, the midwit curve, has a real truth to it, in that you can really overplan your life and overthink your life. Following your intuition and being attracted to a problem—I think there's something there. You were talking about the human spirit, right? There's something, and I agree with you, that's essential about the human spirit that is probably not going to be captured.

I tweeted last night: The 1 thing AI will not replicate or displace is the human spirit.

I think because we're humans, we live among other humans, and we can see problems that are very important for our communities and the people around us that are not going to be embedded in the machine. Following your intuition about what problems you want to solve will probably net out to a truly unique and differentiated position in the world for you to have.

I'm very blessed to have a long-term mission. I've found my purpose, and I think the way to do it is to follow your curiosity and intuition. If it takes you to university, then that's fine. If it takes you to building a company, that's fine.

The implication of this is pretty profound. Take education: we've been doing education from a supply-side perspective for 100 years, where you go get a skill and become a developer, an accountant, a doctor, or whatever, and then you go to the job market to sell those skills. All of that evaporates now, and we have to move to the demand side and say, "What problem do you want to solve?"

That's right.

Now you get the tools, the technologies, the techniques, and the repositories to solve that problem.

I think that is going to be the most exciting change that we're going to see.