The AI Coding Agent Revolution, The Future of Software, Techno-Optimism | Amjad Masad, CEO, Replit

Thank you. Excited to be here.

That's right.

For example, a DOS or Unix command line is a form of a REPL. So it comes from Lisp, actually, in the 1950s at MIT, where you can create the simplest programming environment with one command, which is read-eval-print loop. So read reads the command from the command line, eval evaluates it like a Lisp string, print prints out the results, and then goes back to the start. So this is like the simplest IDE in the world, right? And so when I was building—so the way Replit was born—I was going to school in Jordan, studying computer science, and I didn't have a laptop.

So every time I would want to do some homework, I would have to reinstall the development environment. And it was so annoying, right? You have to download gigabytes of software and packages, and there's always something that goes wrong. And so I was like, I'm using everything in the browser at the time. Google was putting Docs and Gmail in the browser.

Yeah, 2008, '09. Chrome came out, V8, the JavaScript engine. The web was becoming a real application platform. So I was like, why can't I code in the browser? And then I sort of Googled it, and I was surprised that no one had built it. So, okay, I'll build it. How hard could it be? And I actually got a prototype pretty quickly because you can eval JavaScript in the browser, right? So I put a text box, I put a button, you write a little bit of JavaScript, you run it, and it alerts the results.

And I gave it to my friends, and they were all playing with it and really excited about it. Started adding features like sharing and saving and things like that. And I was like, okay, I need to add more languages. You're studying computer science, you're probably going to be using a lot of different languages: Python, Java. And that's when I hit into problems, because the browser can only run JavaScript. And this became like a multi-year research effort, culminating in having a breakthrough where my friends and I were able to compile a bunch of languages into JavaScript.

We used a library from Mozilla. That library became Wasm, which runs sort of native programs in the browser. Figma is built on that. A bunch of really modern applications wouldn't be possible without that. But at the time, it was just a library and there wasn't anyone using it. And so we were like the first to kind of use it to run all these programs in the browser and put it up on GitHub, and it went viral on Hacker News and a bunch of other places.

And it was really surreal because, yeah, I didn't expect it to be this big. And people were talking about it in conferences. Brendan Eich, the CTO of Mozilla at the time and the inventor of JavaScript, tweeted about it. And that was like the highlight of my life at the time.

Yeah.

And then one day, as always, I'm lurking on Hacker News, and I see this startup, Codecademy, maybe a couple months after I open-sourced all the infrastructure that we built. And I click on the site, and it was going viral, right? And they built this interactive way to learn how to program. Because learning how to program at the time, it's like textbook and all of that. But they built this thing where the course kind of talks to you and gives you these exercises and you have to code with it.

It was a very novel idea. And at the time, the hype was around MOOCs, massive open online courses.

And I had a feeling that they're using my stuff. I just liked the way the system worked. And then I tried to kind of look behind the scenes, and I validated it. So I put a comment on Hacker News. I was like, oh, you're using this library, my library. And then two days later, I get an email from the CTO, Ryan Bubinski, and he's like, yeah, I'd love to chat with you.

And I don't think they knew I was like a kid in Jordan. And so we started talking.

Yeah. So at that time I had graduated. My time at the university was pretty tenuous because I didn't like sitting in class all that much initially. I also, and I'm sure we'll get more into that, was reading a lot of sci-fi. I was really into AI and really into how do you build a brain? And I was reading a lot about cognitive science and neuroscience. There wasn't a way to study that in Jordan, but that was a personal interest of mine.

And I felt like we're on the cusp of solving intelligence, and I felt like computers will be able to program themselves. So I was like, okay, if computers can program themselves, what can I do with computers? And I was like, okay, well, someone still needs to build these computers and run them. So I went into electrical engineering, computer engineering, as opposed to programming. And then two or three years in, I'm like, I don't think machines are going to be coding themselves anytime soon.

Yes. Well, I just switched back to computer science. But yeah, it took me six years, and then I graduated. I was hacking all the time.

Well, we're the YouTube generation, right?

Yeah, I mean, YouTube came out, what, 2005?

I just started watching it then. And so I think—

No, I went to British schools.

Yeah, because—

So Jordan was like a British colony, right? And so the influence was more British. And so we did IGCSE and O-levels and A-levels and things like that. But it was also lame to sound English. Although now it's cool, I think.

But I did, to be a little serious, I did work on the accent a little bit because, I don't know, I guess I just wanted to fit in. And I thought, I really want to be American, and I've always kind of wanted to be here and wanted to be in Silicon Valley. I watched—there's a movie, this very low-budget movie called Pirates of Silicon Valley, and it's about the story of Steve Jobs and Bill Gates. And it's a very overdramatized story, and half of it is wrong, but it really inspired me.

And I was like, oh, Silicon Valley is like a great place. You can fight and yell at each other and build computers and software.

So we grew up with a lot of American culture, and it was very much an interest of mine to really be here.

Yes. But I was also very, very passionate about this because when I built this open-source version of Replit, it was about education. It was about me being able to learn to code and code on the go without needing an expensive computer. And so I was like, okay, if this helped me so much, I think it's going to help a lot of people in the world. And I had this theory that programming could be the thing that democratizes access to wealth and entrepreneurship.

And I think it is playing out that way.

And so I was really passionate about joining Codecademy. And at the time, you saw a ton of people come online in Africa and India, and then Android had just started going, and it felt like everyone's going to be able to get a computer or a phone. And I always wanted to make programming work on the phone, and I tried really hard to make Codecademy work on the phone. Now, Replit, we have a mobile app, actually.

No, it's pronounced Code Academy too.

But everyone—

If you remember, at the time, it was the hot thing to do at a startup: misspell your .com.

Yeah, exactly.

Yes, Internet.org, and I was very early to that. Do you remember that?

So the idea, which Elon ended up doing, was beaming internet from the sky, essentially. And again, this fits into this vision that I had of more people connecting to the internet. And I felt that was such an important thing. I had the option to join the Internet.org team or work on mobile. So when I got there, the way Facebook worked is, you go through boot camp for like three months, and then you do team selection. So I didn't know what I was going to work on, but I picked up Android tasks.

And did really well at them. And I picked really the hardest tasks that you shouldn't pick as a boot camper. At some point, I broke the Android app for like 10 million users, the beta users. And everyone was surprised that I was able to do that as someone in boot camp because I was taking a lot of risks to move fast and break things. I joined the Internet.org team, and because we were in New York, almost all the teams wouldn't take me.

They were like, you have to move to the Bay Area.

And I wanted to stay here. So I ended up on the Photos team for about six months. And that was the only team that would take me, essentially. And then on the nights and weekends, I was contributing to React, and I was interested in this team that was trying to make mobile programming more accessible because Java was this heavy, brutal thing.

Every time you change a line of code, you need to recompile the entire app. So this project called Catalyst at the time, which became React Native, and I was one of the first three or four people to work on it. And I was tasked by the founder of React and the founder of React Native, Jordan Walke, to build the JavaScript toolchain for React Native. So I started building it here initially on my own.

And they told me, all right, go out and build a team. So I was here in New York, interviewing who's who in the JavaScript community, and I met a lot of friends that I'm still in touch with today.

Yeah, but actually, I ended up moving to the Bay Area because, we were just talking about it, there wasn't a lot of talent here. It was very competitive. And so I moved to the Bay Area, built the team there, and released React Native a year later. And it was this super viral thing, and a ton of people used it.

I sort of wanted to let go. I was like, I did this thing. I kind of burnt out while working at Codecademy, and I wanted to do other things. But actually, the project kept growing. It was like this open-source project, and we had like a demo page, and I kept getting emails. And I looked at Google Analytics, and we had still 10,000 users. I'm like, this thing refuses to die. Yeah, it refuses to die. So let me fix it a little bit.

So I started fixing it, and pretty quickly it started growing again. 10,000, 20,000, 50,000, 100,000 users. And at the time, I moved the code execution from the browser to the backend. At the time, Docker had just come out, and I felt like if we moved it to the backend, we could have packages and servers, and I could make it into a full development environment. And so it was costing me a ton of money at the time to just keep going.

It was like the user growth was insane. I was like, okay, maybe I have to sell it. And so I tried to sell it. And then I was like, my time at Facebook was coming to a natural end. I was like, I kind of want to start a company, but I don't want to start this company because I'm already done with it. But it kept growing, and people kept really getting excited about it. And I was like, okay.

Okay, I guess I need to start this company. And I looked at the landscape of online IDEs, and there were a ton of attempts at it. The problem is no one had built like a sort of a web-native experience. And the way I define a web-native experience is that it should be able to load really quickly. It should be able to be URL-addressable. So when I send you a URL, you should be able to see my project.

Yeah, it should be multiplayer by default. It should be real-time. And none of those existing IDEs were doing that.

Yeah, Figma was in stealth for four years. Exactly right. And so, yeah, we started Replit officially in 2016 because this thing just wouldn't die and kind of exploded itself onto the world. And I was like, okay, I was still passionate about the idea, but I just knew it was going to be super hard.

So we tried to get into YC even before I quit my job at Facebook, because that's what you would do to de-risk it, essentially. And really, YC de-risks a lot of entrepreneurship, mostly for the good. I think you need to take—I think some amount of risk is good, but we wouldn't get in. And so when I took the plunge, it was just us, and we were able to raise money from Bloomberg Beta, a bit of capital. And over the next two years—

Pre-seed, yeah.

Pre-YC.

Okay. And we were grinding for two years, building a lot of this infrastructure that we just talked about. But it was pretty lonely, and we weren't really able to hire a lot of engineers. We weren't part of the Silicon Valley sort of hype or buzz. Also, the dev tools at the time were kind of not an exciting space, right? That was pre-GitHub acquisition. GitHub acquisition kind of really exploded everything.

So I got used to writing about what we were doing. And so I tried to write something every couple of weeks about some of the engineering challenges we were facing. And I thought it could be a good recruiting tool, marketing tool. And Hacker News really liked my writing, and so it would always make it there. Now they hate me. But that's the kind of natural evolution of Hacker News.

Three or four times. I kind of lost count.

They wouldn't reach out to us. It was like, I would get the automated email.

So the only people that would give us money—we had a lot of developers use the platform, but they're like, "I'm not going to pay for it. It's a useful thing, but probably not going to kill me if I don't have it." But the people that were really dependent on it were the students. Again, not really surprising, because that's what I built it for, essentially, and hobbyists and people who wouldn't know how to set up a programming environment, essentially. And so we started monetizing that.

So we started selling to schools. We sold the API as well, and other companies built on top of it to kind of sell to schools as well.

I mean, I understand it. I have tons of respect for education entrepreneurs. It just takes a lot of grit.

We get checks for like $50.

So Paul Graham had already retired. If you remember, he passed the mantle to Sam in like 2015. And so, 2017, '18, he was—

Yes. Paul Graham, PG, the founder of YC, read Hacker News every day at the time. And so he found one of my posts, and a few days later, I wake up one morning and I have a message from Sam Altman. It's probably still in my DMs. I can look it up.

Yeah. He was like, "Hey, I run YC, really impressed by what you're doing." I was like, "Dude, I know who you are." And he's like, "I'd like to meet you." And so he's like, "Come to the Mission." I was like, "That's not YC's office." So I go to the Mission, and it was Neuralink on the right, and it was OpenAI on the left. And so I went to OpenAI. I was like, "Why am I here?"

Yeah.

Well, Elon just retook the building.

Okay, so now xAI is there.

I think there's some sort of message underneath that. That's like, "We're taking it." So he's like, "PG found your company, and he said it was something that he's been looking for for a long time." He's like, "Why don't you go visit him in the UK?" And I'm like, "Well, yeah, let me get my private jet." At the time, I wasn't American. To get a visa as a Jordanian to the UK would have taken me another year.

I was like, "How about I email him first?" And so I started this email relationship with Paul. By the way, PG doesn't hop on Zoom calls, or very rarely does.

So we started this, which was awesome, because I'm getting personalized PG essays, and it felt such an honor. And we just talked about programming, what's hard about it. And he had told me that one of the insights from Viaweb—Viaweb is the online store that PG built and sold to Yahoo—was like, "Oh, people want to customize the stores, and they want to put in a piece of code." And he always imagined, like, if you built a browser environment where you can code and easily generate apps, that could be amazing for entrepreneurship.

And so we really connected on that. And that's always been my passion, is like, how do you get more people to participate in the amazing wealth creation engine that's the internet? And we also connected on the technical aspect of it and programming languages. And I really enjoyed the emails. And he's like, "Look, I'm going on a trip now, but YC is about to start." I was like, literally a few days later, YC would start. "Why don't you join the batch?"

And then he went offline. I didn't hear from him. So I forwarded the message to Sam. I was like, "Hey, PG is saying we should do YC." And my co-founder, who's my wife, she's still with the company. She's head of design.

I have a brother.

My brother leads a portion of engineering. And I was like, "Should we do YC?" And it's like, we got rejected all these times. We're starting to make money. Do we need to give them 7%? And so I emailed Sam and Paul and just said, "We'll do YC as long as we get access to you two. We don't want to go in and become some random startup." And they said yes. I was like, "Okay, we'll do it." And then Sam is like, "Hey, just a formality, fill in this application."

Like, fuck, I've done this application so many times. It was kind of like triggering some trauma in me because you record the video and talk about it and send it, and you never hear back. So we did this application very quickly, and the next day I go to YC. It was literally the kickoff day, so we would have a dinner. We're sitting outside. I think we're literally the last interview. So we sat outside for two or three hours, and you see these late interview startups that are going in and out.

And so they call us in, and I go in, and it was Gustaf and Jared and all these amazing YC partners. And at the end, it was Michael. At the time, he was the CEO. Sam Altman was the president. He was the CEO. Michael is a big guy. And so I shake his hand, and I felt a squeeze. I was like, what's wrong with this guy? Why is he— I barely sat on the chair, and he was like, his face was really angry.

He was like, why did you Rickroll us? And so, because I was so sick of making these YC videos, and I thought it was really just a formality, I put the Rickroll song, the Rick Astley "Never Gonna Give You Up" song, in the YC application. Actually, if you go to our YC profile now, they publish the videos. Click on the video. It is still a Rickroll. And people discover it every once in a while. And imagine what's happening because they're reviewing the application just before the interview.

And they're all sitting together, huddling in the computer, and they click the video, they get Rickrolled. So they're all really pissed at me. And they gave us the toughest YC interview, probably in the history of YC. It was really tough and pressuring. One question from here, one question from there. Because I asked my friends from the batch, and no one had such an intense interview. So the interview is done, we leave. I tell Haya, I was like, okay, I don't think we're going to get in.

We fucked it up. We fucked up our only chance. And I call an Uber. And then I get this phone call from a random number and pick it up. And it's like, it's Adora Cheung. She was in the room as well. And she's like, you got into YC. How does that sound? I was like, are you sure? I just offended all of you, and you're really angry at me. It was like, no, come back, do the paperwork quickly because the dinner is about to start.

That's how we got into YC.

That's very true, actually. Very good observation.

Yeah, we published the seed deck somewhere on the internet, and you can see there was this master plan, and AI was part of it. So as part of my work at both Facebook and Codecademy, I worked on compilers and interpreters, and there are very kind of finicky programming exercises. So, taking a piece of code and parsing it—that's how compilers work. You parse it into a structured tree, and then the way you traverse the tree, that's how you evaluate a program, essentially.

And I always figured that you could use machine learning instead of doing all of this by hand. And NLP was getting better and better. Actually, I don't know if you remember, but in 2015, '16, we had a chatbot hype.

Period. Because there was some NLP unlock at the time, but it wasn't good enough. Everyone was faking it. There was Facebook M, which was a bunch of people.

And so I had this intuition that we were going to be able to run machine learning on code, to evaluate code, to help people learn how to code. It's just this intuition that deep learning on code will unlock a lot of avenues. And I felt like if we built this large community, we were also going to be able to collect a lot of data that's going to help with that. And so every year, we would prototype something from the start, and it was just never good enough to ship anything.

And then GPT-2 came out, and that was the first time when I felt like, okay, this feels like the unlock, because GPT-2 can write semi-coherent code. And then when GPT-3 came out, I think we were one of the first companies to build anything on top of it. The first thing we built is, like, you highlight a piece of code and you explain it.

So still a lot of people, hobbyists, students, are using the platform. And so a lot of code, you're just copy-pasting from Stack Overflow or library code, and you want to be able to understand it. And so highlighting a bit of code and explaining it, it's actually very useful. Still have it in the product. People still use it a lot. And it wouldn't have been possible without—you wouldn't have been able to even conceptualize it without LLMs. But what I really wanted to build was, like, a true autocomplete code generator system.

And it was really hard to build anything like that with GPT-3. It was expensive, it was slow. And we started talking to OpenAI about doing something with them. And then Microsoft had just come out with Copilot in beta. And I was like, we really need to do this. This is the future. Our company would die if we don't do this. Because I was like, all of coding will be that. There was a lot of skepticism even inside the company, but I felt like that was the big bet that we were going to have to do.

And so at the time, it was like maybe 2022. And so we're between a rock and a hard place. We were like, okay, we either build on GPT-3, and it really sucks because Copilot was using some kind of fine-tuned version, distilled version of GPT-3. So that was really fast. We couldn't come to any sort of agreement with OpenAI at the time. And so I was determined that we're going to have to go build it on our own. And no one at the time had done that.

Everyone was relying on OpenAI. Open-source AI was not really a thing. And we were really plugged into the news, and we found that Salesforce had a small research team, and they worked on code models. And they produced this model that was still very slow because of the model architecture, but it was decent. It could generate some code. And so we took that.

That was CodeGen, yes. So we took that, we sort of rewrote the code to make it a lot faster. We did a lot of engineering work on top of it to be able to ship it. So that was the beta version of what we called Ghostwriter. So we built the first sort of Copilot alternative, essentially. And it went super viral because it really allowed people to imagine that there's a world in which you can build AI as a startup without depending on OpenAI.

Because at the time, the feeling was like, oh, OpenAI invented this thing, and it is this magical thing. It's just like no one will be able to replicate it. It's like this dark art. And OpenAI had stopped publishing research and models. And so I think we inspired a lot of people that, oh, okay, open source is possible, open models are possible. But still, CodeGen wasn't really great. So I was determined to train our own model. And so that was the next phase.

And I wanted to use our data to do that. And again, we were—

Yeah. By the way, when we released CodeGen, it was me and, like, an intern working on it. I recruited an intern, and that was it. When it came time that we needed to train a model, we had to have more and better people. And by the way, other people in the company had joined and started learning how to do machine learning. And some people on the data science team joined. We had started talking to Google. We've always been building on top of the Google platform, and they reached out and they're like, look, we're building a GPT-3 alternative, and we want to see if you want to use it.

And we're like, oh yeah, it looks great. PaLM at the time, we would love to use it. And the talk progressed, the conversation progressed, and then it was blocked. They're like, we're not going to release it. The search team said it was too much of a legal risk, so we're not going to release any LLMs. Our counterpart, a researcher on the other side of Google, was part of this conversation. He was so frustrated by it, and he wanted to leave anyway, so he left Google.

His name is Michele Catasta, a researcher who was on the PaLM team. And I was like, we're going to train this model. Why don't you at least part-time kind of join us to train this model? And he was like, yeah, let's do something great. And I was like, look, we're going to be constrained by capital, by talent. And so we have to work within these parameters. And so we decided to train a super small model. At the time, Llama had just come out, and in the research paper, they talk about how you can train a small model a lot longer, get the same performance as a much bigger model.

So it's a lot more cost-efficient. And we trained this 3 billion parameter model. It was the first 3 billion parameter model that was state of the art in coding, in open-source coding. It did better than all the other open-source models. It actually did better than PaLM on coding. And yeah, that was another kind of big moment in the AI community. And then at the time, Ghostwriter was finally a product that we were proud of and we could sell, and we put it on the market, and that commercially started being a viable business.

Very early '23.

Yes, very early '23. I mean, around the time when ChatGPT was going viral.

Yeah, so after we released Ghostwriter and we got a lot of excitement, we did this really big deal with Google on the cloud and AI side and raised a really big round. And I was like, okay, it's time to mature as a company. I'm going to go hire all these executives, and we're going to grow the business, and we're going to go into enterprise, and we're just going to do all the things you're supposed to do, I guess, on paper. I did all of that.

At the same time, I sort of neglected the product a little bit and also the culture a little bit. Up until that point, for the past six years, I was interviewing everyone. And when I put my big-boy hat on, I was like, okay, I'm going to delegate all these things, and I'm going to just be more high-level and add more structure and management and things like that. And we were suddenly behind at the time.

Autocomplete stopped being the state of the art for how people use AI for coding. Cursor had just come out, and there were a lot of other explorations around agents and different things like that. And we had a lot of these explorations early on, but we started lagging behind a little bit because we were optimizing. And this is a lesson I take to heart right now. We were optimizing for the current generation of models. We were building, like, Ghostwriter V2, right?

As opposed to figuring out what's next. And we'll get back to that in a second. And so, we started 2024, we're like 130 people burning a ton of money. The commercial aspect kind of lagged behind how much we were burning and how big the team was. A lot of the engineers were really miserable. Even the early engineers, people started leaving because it just started to feel bureaucratic. And you know this, like, a lot of companies go through this.

And also, we weren't defining our customer set clearly. Replit was always about democratizing programming. Our first mission was making programming more accessible, and then we kind of updated it to make it a little more ambitious, which is empowering a billion software creators. But people at the company were confused because, oh, are we building for developers? Are we building for no-code people? Are we building for...? And so I felt we needed a radical change, and I felt we needed to bring the burn down, and I felt that we needed a reset.

And I felt we needed to innovate again, and we needed to build the AI product that wasn't possible today. And the thought experiment I went through is like, let's build the thing such that the next generation of models will make possible, which actually turns out to be a really hard thing. How do you project forward what is the capability of these models? And when are they going to land? And there's so much risk to that.

I said, so I gave this big talk about how programming agents are going to change the industry and make it so that anyone, even people with no technical skills, would be able to build software by just commanding these agents. And it put up a very ambitious vision for Replit and for the world to see. And I felt like, oh, maybe we'll get there by the end of the decade. And early 2024, when we kind of went into founder mode, laid off half the team, leaned up the executive team and the management structure, and brought us back to kind of like a small team with a very intense focus on one thing.

And that one thing was like, let's build some version of Agent. And so we created this thing called the Agent Task Force. And that was basically the only thing we were doing at the company. And that was like March or April 2024. And we got something done really quickly, and it was starting to work, but it was very clumsy. Claude 3.5 Sonnet came out. And that was basically the bet. It was like, we're going to build something that's not going to work today, but at some point some model will come out and it'll work.

And that's exactly what happened.

Yeah.

It was very buggy. It was very random. It could produce something. It was very slow. It would take like $100 to make, like, a to-do app or something like that. And if you watch some of the videos, like Devin was, for example, working on, they had all these problems as well. It's very slow.

That's right. That's right. And when Sonnet came out, we plugged it in and the product immediately got better. We did also cut scope quite a bit. So with Devin and others, a bunch of them got funded. Their idea is, like, fully automated agent, so you can put it in the background and it'll do its thing. With us, the kind of intuition that I had is we want to make it so that it's like this man-machine symbiosis. And that's something I actually wrote about back in 2017.

We didn't think we were at a place, and the technology was not at a place, to create fully autonomous systems. So let's build this human-in-the-loop system where the agent goes and could do 10 iterations or 20 iterations, but will come back to the user and say, was what I built what you expected? Can you test it for me? Because agents are very good, very bad now at having eyes and being able to test something. And that was another kind of piece of product innovation that we did.

So we got to September, we launched it in early access, and at the time there was no other programming agent on the market. So it was really the first and captured everyone's imagination. And the video of the launch went viral.

We thought we'd get, like, some users to test it, and we got a ton of users, which was good and bad. It wasn't ready. And so there's always this—you want to set expectations correctly. And I think we quickly made it a lot better. Like, right now, I would say it's like 5x better than what we launched in September. And I feel like it's going to get 10x better even this year.

Well, now it's more.

Yeah, 6 or 7x.

I would say there's something also between the assisted and the autonomous. And that's like the Cursor model, which is not like type-ahead, kind of like Copilot, but it's not fully—it's not like an agent. It is somewhere in between, where you can generate entire files.

And that's the thing that made them very successful.

That's exactly right. Yeah. And we wanted so that you don't have to be a programmer. I think you still need to be technically savvy, which is, I think, the kind of people we're getting: Silicon Valley professionals and largely finance and these kinds of industries, although we're starting to break out into more traditional industries that are able to use agents. But our vision, again, is like, I think we're going to be able to get to, at some point, like a billion users.

And the kind of mental model that I have, if Excel has like a billion users, anyone who's able to use Excel should be able to use Agent.

So yeah, you write a prompt, it quickly goes into generating almost like a full application for you. It'll provision a database, it'll do whatever it needs, like a programmer would need. It'll create a development environment, it will bring in libraries.

And I think till this day, we're the only company that's done that, which is like the full-stack experience from the code to the databases to the cloud deployment.

We actually stopped caring and perhaps removed a lot of the features of people going to the IDE and coding. It's still possible. But our view, and the way we track our metrics now, is that when you go into the editor and code, it's almost like the Tesla disengagement, right? So what is the main metric that Tesla tracks? Miles between disengagements, right? And, like, the V13, I don't know if you saw the chart, but it goes—it was like 10x V12 in terms of how many miles to go between.

And that's my experience too with Tesla now. I go to work without disengaging once. So when you disengage, they consider it a bug, and they're like, give us feedback. So that's how we're thinking about it. So all the stuff that we built around coding AI, I don't think it matters. What really matters is being able to chat with an agent and get an application on the other end.

Yes.

Yes. The multiplier nature of Replit, what we're seeing is you can bring in an engineer into the project, so you can invite an engineer.

Yes. So let's say you're a solo entrepreneur and you're trying to build an application, and you get 80% done, but then you get stuck on this one problem. What do you do? You could potentially go to Upwork or some of these places, but they may not know Replit, and they might confuse you even further. So we built this small community where you can hire someone from the Replit community to get your application to the finish line.

Yeah, so at the bottom layer is the Replit container, and there's a process that we call PID 1, which is like the first process in the container that orchestrates all these things. It has services such as install this package, write this file, read this file, but also has services where it is watching the file system, and then every edit is updating a vector DB there. So RAG is built into that. And so this fully encapsulated system that kind of looks like an operating system.

So it manages the entire system and exposes this very clean interface to the agent where it can expose all these tools. We can install a package, read a file, delete a file, write a file, provision a database, do all these things that a human using the Replit editor and IDE could use. And then, in terms of the architecture of the agent, it's a multi-agent system. So when you're talking to the agent in what we call the lobby, which is like when you're starting the agent session, it is mostly a conversational agent that is building a plan for you.

It doesn't have a lot of tools, only tools around retrieval. And so that builds a plan for you. And then when you start with a plan, it goes into sort of a prototyping agent, as it were, where this agent is responsible for generating the initial application. It tries to kind of be as complete as possible, but it tries to kind of reduce the ambition and the scope of the program because people put PRDs that are really big. And so we're trying to get the first thing done, and then it kicks it off to the project management agent.

And so that becomes the primary agent where you're conversing. And that agent is the agent that has the memories, has the connection with the user, and then it is the router agent. It will kick off other agents, right? So you give it a new task, add this feature, it will kind of provision an editor agent. So the editor agent gets some context, gets the files, the RAG via files, and goes into its own iteration loop where it's, like, writing the files and trying to edit the file.

By the way, editing files with models is actually quite a tough problem. Turns out they don't know how to generate diffs. They don't work with line numbers very easily. So you need quite a complicated system for edits to work. So when the editor finishes, it kind of goes back to the project manager, gives it a summary of what it was able to do, and that gets committed to memory. Then the project manager kicks off a bunch of processes to restart the application and get all the way to presenting the application to the user.

Now, along this way, it might run into an error, right? It restarts the application, runs into a compiler error. The editor maybe ran into a syntax error or runs into a runtime error or runs into a database error. So now you need to kick off some debugging loop to be able to debug that. And the debugging loop might come back and go back to the editor. And so this is where the autonomy is happening, right? It is going through these different states.

In order to get the application to the sort of final state to be able to present it to the user.

Yes. And that way you can encapsulate every agent in this clean interface because the contexts get pretty messy if you're doing everything in one place. And what we found is all these LLMs are getting marketed as very long-context models. What we found is reasoning over long context is actually not very good. It's, like, not good at all. Once you cross 32,000 tokens, that performance of reasoning just goes down to hell very quickly.

And being able to audit the memory and being able to, yeah, make sure not everything goes into the memory. Like when you have a subagent going and doing its thing, not everything comes back and pollutes your memory and context, and kind of really having this isolation.

So software agents, there's this benchmark called SWE-bench, so Software Engineering Benchmark, and that's what everyone's competing on. We don't really compete on that. And the reason is because we are building for a different audience. The people who are building SWE-bench are building agents that go from issue or Jira ticket to a pull request. What we go from is a high-level product description to an app. And that's, like, a fundamentally different problem. And so the tough thing for us is that there aren't academic benchmarks, which gives you a head start, right?

So we've had to build our own benchmarks. And as you get users, you get more data, and you're able to get some of that data and create benchmarks from it. And that's, like, quite a tough problem. So we have our own internal evals and benchmarks.

Yeah. So when we were building the Replit environment for humans, you were like, okay, this is the editor. This is where you go and type code. This is the UI for the editor. There's the line numbers, and this is what you see. There's the console. This is where you run the code. And here's the UI for the terminal and everything. Here's the files, and here's the UI. Here's how the files work. And then you want to deploy or you want to add a database.

Here's how you do it. Here's how you provision the database. Here's the UI for it. Here's the interaction, right? When you start building something like the agent, your expectation is like, okay, I'll just present all these things as APIs and it'll just work. What you find pretty quickly is that if you just do that kind of verbatim, you don't have the best performance. And there was a paper, I think, from OpenDevin or some of the people that are working on agents in open source, where they introduced the term ACI.

So the term HCI is human-computer interaction. ACI is like agent or AI-computer interaction. And the main observation from that paper is that actually you need to create tools that almost have a UI, but a UI built for language models. So the interesting philosophical thing here is that language models are actually kind of human imitation machines because they're trained on all our crap on the internet. So they become this, like, it's very much like us. And so they understand—their view of the world is closer to us than to a pure program.

And so in that paper, for example, they say it's better to, instead of giving it the entire file, the editor, it's better to give it a view of the file and have it navigate it, do page up and page down. And so you're actually building an editor for the agent. And this goes across the entire stack. For example, we give it eyes, right? We give it a screenshot tool so it can go screenshot the browser and see what's going on in the application and the kind of feedback as it's editing the code.

It gets compile errors and things like that. And so now we spend a lot of time crafting these tools for the AI, which is kind of like a weird, surreal thing to think about.

The main problem is reasoning. And the reason why I sort of predicted that it's going to take a while to get reasoning correctly, and I think I was kind of correct, it happened faster, but I still think the reasoning is not there. Obviously, O1 and these test-time compute models might change things. But sort of what models still are today is they're sort of like completion engines, right? That's how LLMs are trained, the sort of autoregressive models where they try to predict the next token.

They're still next-token prediction machines. Reasoning is different. Next-token prediction is more like intuition. It's more like, oh, I know what you're going to say next, right? Whereas reasoning is more like this happened, and potentially this caused this. And, well, here's a conjecture, there's a hypothesis, and let me think about that. Oh, but I'm wrong. Let me backtrack and think about a different hypothesis. And so it is qualitatively different than intuition. And in the sort of literature, it's called System 1 versus System 2.

That Daniel Kahneman book, Thinking, Fast and Slow. And so models today think fast. You could argue, "Oh, when is that?" But still, the jury's out whether you can build really good agents with them. We're starting to see results from that, but it's still kind of early.

I think the best way to think about it is to think about other industries where that happened. Right now, it's very easy to take an amazing photo with an iPhone, right? Apple did this whole campaign around "Shot on iPhone." There are a lot of people that are just making really great content with just iPhones. Professional photographers didn't die. Maybe the population is not growing as fast, but they exist and they have their own equipment, and it tends to be better.

I don't know if the market size of professional photographers has shrunk, but it probably kind of stayed constant. And I think that's probably what's going to happen to software engineering. The growth that we've seen in computer science over the past 10 years, I think, will slow down as the need for them slows down because companies will be more efficient. Zuck was on Rogan recently, and he talks about, in 2025, we're going to have a mid-level software engineer equivalent of a coding agent that's going to be able to commit code.

I think it's a little overstated, but I don't think he's wrong. I think we're on this trajectory where you can do a lot of junior- to mid-level engineering tasks automatically. And that's refactoring, testing, I think initially, but pretty soon just some basic features and moving UI around and things like that. So the need for software engineering will go down. And as nonprofessional software engineers are able to build their own software, I think the demand for the plethora of SaaS that we have today will also go down, and that will reduce because you'll be able to generate software on the fly that really fits your use case.

And I don't think SaaS is going away anytime soon because there's all these ways in which companies can support enterprises. That's quite important in terms of security and all that stuff. But there are a lot of long-tail SaaS that I think will just go away, and people will be able to generate the software that will reduce demand for software engineering. So I think the population of software engineering will probably be constant and not grow all that much.

That's a good question. Well, you can generalize it. How do you become an account executive if there are no AI SDRs? And it feels like there's not going to be SDRs, right? I think it's just a general question about entry jobs. I don't have an answer. What do you think?

You could also imagine AIs teaching humans. I feel like education sucks, right? Selling to education sucks. But it is one of those applications of AI that hasn't been fully explored. I think using AI to teach those junior programmers and accelerate them to becoming senior pretty quickly is a possibility.

Well, I think if you think about it as sort of like a factory line or a pipeline, the bottleneck has always been sort of like the making of the thing, right? So you have an idea and you need to make the thing, and then you need to sort of sell it and distribute it. And we have all these ways where sales and distribution has gotten easier over time. Like PLG is an example of one, ads, things like that. So there's a lot of innovations and ways in which companies can grow fast today.

So the bottleneck is like, oh, how do you create the thing? And even if you have the capital, it's actually still quite hard to make things, right? And you know that as an investor. Sometimes you invest in a company, it's still not working. You could give it as much money as you can. I mean, what is it, like all these companies, Magic Leap and Quibi or whatever. And there's a ton of examples of companies that sucked up a lot of capital and went nowhere.

But if, in fact, software creation gets dramatically easier and better, then I think the bottleneck will shift somewhere else. And my prediction is that the bottleneck will become on the idea side of things, which is counterintuitive because the Silicon Valley dogma is like, ideas don't matter, only execution. And maybe in the future it'll be more like, execution doesn't matter, only ideas. Because if you think about it, there aren't a lot of great ideas in the world. And a lot of times it takes this novel idea, and all these ideas in retrospect seem great, but sometimes a thing could have happened 10 years ago, and it just happened today because someone had a really great idea.

It's a skill.

Right. Finding problems to solve is a skill. Problem solving is a skill. But if you know how to code or if you know how to make things with Replit Agent, you still need to develop the skill of spotting problems in the world that you're going to be able to solve. And by the way, back to the bottleneck, that bottleneck is also true of science and math and every part of human invention. Assume that the labs are fully automated and we have robotics, and so what becomes the bottleneck to running experiments?

Ideas. I assume we have these amazing theorem provers and things like that, and then you can try things really fast. And then what is the bottleneck to coming up with new mathematical insights? It's going to be new things to try and ideas to try. It brings up the question of, like, how do you teach the next generation of people? How do you teach your kids to become more creative? And, yeah, I'm not sure how you do that.

I feel like venture capitalists are already intuiting that the sort of golden age of SaaS is on its way out. And you can look at the sort of American Dynamism trend, like, trend to going to the world of atoms away from bits. You have that after Anduril; like, defense is a big thing that all these VCs are getting into. Government in general, obviously, like AI, AI chips, and sort of what we used to call hard tech is becoming trendy, and maybe it becomes the majority of venture capital investment, as opposed to the other way around.

I think there are still network-effect businesses in software that are quite defensible. Like, even if you generate Facebook, you're going to have to get a billion users, right? So consumer social, those will probably become more important as categories for VCs. And perhaps there's probably going to be a consolidation of venture capitalists, especially there's, like, a bit of a squeeze where a lot of the early-stage rounds, there's, like, a lot of angels and solo capitalists that are able to do those.

And the later-stage rounds, they're, like, so expensive and capital-intensive that only a few can do them. And in terms of entrepreneurship, the other thing that you brought up, I actually think it's going to be net positive for entrepreneurship. The way we define entrepreneurship today in Silicon Valley is, like, you build a venture-scale business. The way America defined entrepreneurship pre-Silicon Valley is, like, anyone would be able to build a business. If you look at the number of firm creation, it's, like, been trending down since, like, whatever, the '60s or '70s, which is kind of crazy because everyone's talking about startups today, but new firm creation has been trending down.

Actually, there was, like, an uptick with COVID. I don't know if it stayed, but COVID got a lot of people to quit their jobs and start online businesses. So I think the form of entrepreneurship will come back as the sort of more small-business entrepreneurship.

Yeah. So I think that there's actually two ways to view the world. You're either a primitivist or a techno-optimist, because I guess the European style of, like, we're done, we're just going to distribute the wealth and regulate—I don't think it works at all, because capitalism is based on growth. And when you don't have growth, you actually have decline. And so you can say we're actually going to go back, and actually technology is bad, and degrowth is actually, I think, a more coherent view than we're done.

And so you're either degrowth or for growth. There's nothing in the middle. And I think pro-growth has its downsides. Like everything that we talked about, it's scary because you just don't know what to teach your children. Things are going to rapidly change. But I'm much more optimistic about that and about the future of humanity than a world where we sort of go back to a more primitive—

Yes.

Yeah, I've always kind of struggled to describe my worldview or outlook. And I think the best way I could think about it is, like, I'm in awe of civilization. And I feel like if you restart the world 100 times, perhaps only a minority of the time will get to the level of civilization that we're in. It's quite hard to go from a single-cell organism to a multicellular organism, and it's quite hard to get more complex. And there's all these ways in which evolution could have stopped.

But I do think no one talks about how human civilization could have stopped at sort of the Stone Age or the agricultural age. There's no reason for it to actually continue developing further.

Yeah. And I think we take that for granted. And I think this is sort of the degrowth versus growth mindset. COVID showed how fragile civilization is. In some sense, civilization is very robust. But you had the disruption in shipping, and that caused all sorts of downstream effects. And there were a few instances where I just felt like, oh, actually, you could easily disrupt the global supply of food and people would die, right? Or just the fact that there's a virus.

Going around and, like, spreads super fast. And governments are dysfunctional, don't know how to deal with it. They either become super draconian or they kind of just put their hands up and are like, we're not going to deal with it. And there was this massive distrust in governments and institutions. And I think going through that phase, I just felt like the thing I value the most, and the thing that I think we should protect—and this also gets to the techno-optimist—is to continue to develop as human civilization.

So if you look at the Biden administration and if you look at what happened to the world, just look at geopolitics. You have the Ukraine-Russia war, which I know people are very sensitive about this, but I do believe that we bear some responsibility for kind of exaggerating this issue, and we made sure that there wasn't any resolution to it. And there was such a motivation to kind of keep it going. Our relationship with China got a lot worse, which is kind of surprising because Trump sort of started that.

But Biden really took it to the next level. And there were moments where they could have gotten on the right track. And there's small things where President Xi visited the U.S. and Biden, same day, kind of insulted him. And the situation in the Middle East, I think, could have been resolved if Biden was stronger and put some pressure on both parties to kind of resolve the situation. And America actually has a lot of power to be able to do that. But the Biden administration was a combination of both being weak and sort of just not having a vision for the world that's coherent.

Obviously, Trump, there's all these downsides to him, but I do think that he's a powerful force in geopolitics because I think he's both a dealmaker and he projects American strength, such that a hostage and ceasefire deal was done even before he took office, exactly what he said. And I think it remains to be seen. There's so many ways it could go wrong.

Yeah, I think without a stable world, there's no tech. And I think if you want techno-optimism, you have to really care about geopolitics. You have to care about a stable world because it's all connected. That's the world that we built. It is a global world. It is globalist, right? So you can't return to this idea of like, oh, I'm a big fan of reindustrializing the U.S., but the dependence on China is a reality. And there's no reason for them to really be an adversary.

And so in terms of Trump's kind of view on AI, I think they already took a hammer to regulations with the executive orders. I think the Biden administration was heavily influenced by the sort of effective altruism movement, spent a lot of money on lobbying and everything, and they were really trying to pause or slow down AI development. That resulted in the executive order, and I think a lot of other things that the Biden administration was excited about doing. And now we have Project Stargate.

Yes, $500 billion.

Yeah. And it seems like Trump played a big role in that and in attracting investments. And I think it sounds like there's going to be a lot of energy around space and getting to Mars. And I think that'll also help with industrialization. It'll help with the U.S. becoming competitive again with the kind of what Elon talks about all the time, which is: you need to be able to look at the future and be excited about it.

And yeah, those are the primary things. I think just generally the mood in Silicon Valley is also better. I think we got to a point where it got very focused on identity politics. And I think that's something also the Democrats really pushed and excelled at. And you see Zuck sort of bringing Facebook back—or Meta—back to the hacker culture and reversing a lot of the sort of identity-based hiring and identity-based promotions and all of that stuff.

And he went on Rogan and he said, "We need more masculine energy in the company."

But you could tell that I came to Silicon Valley so we could build and sell computers, right? And so just bringing back the hard-charging energy of Silicon Valley. And that doesn't mean we want to alienate anyone. In fact, I think it's going to still attract the best talent from all over the world, but it needs to be energetic and it needs to be exciting. And the side effect of this administration is actually making it possible for entrepreneurs to be in founder mode.

Of course, my pleasure. Thank you.