Replit CEO Amjad Masad on 1 Billion Developers: A Better End State than AGI?

[00:00] what is special about humans and what's replicable in the machines, at least in the near term. My view is that [00:11] AI is going to get [00:13] Really good. [00:14] at, um, [00:16] sort of two things, things that are highly represented in the data and things that you can construct a very good RL environment for. [00:25] So what can you construct a great RL environment for? [00:28] um, [00:29] like obviously with AlphaZero games, right? Games are famous for, you can have these self-play sort of algorithms that develop over time. Now with reasoning models, [00:42] You know, math is an environment that especially with lean learning, [00:47] It's like a code, almost like an expression of math that can be executed. [00:53] that's like a great... [00:55] you know, RL environment, I think code execution as well. So running the code and, and then, you know, um, [01:02] doing reinforcement learning on that. [01:05] And things that are already represented on GitHub and things like that. [01:09] But there's a lot of other domains where we actually still don't know what we're [01:14] how we're going to make them better, like fundamentally new ideas, new knowledge. [01:20] it's not entirely clear how we're going to get there. Can you use RL for these more software things? Perhaps you create a word model, you can approximate these things, but I feel like the...

[01:32] sort of the ideas [01:34] Um, [01:35] and the creativity and the sense of like, [01:38] coming up with really novel things and understanding the world in very complicated, intractable way and, you know, coming up with an idea that could fundamentally change how things work or change the world, I think will still be the domain of the human. [02:08] . [02:11] Hello and welcome to Training Data. I'm David Kahn and I'll be the guest host on today's episode interviewing Amdrad Massad, the founder and CEO of Replit. [02:20] Amjad's vision of the future is a world in which a billion people on the internet become developers. And in today's episode, we imagine how these billion developers will reshape the economy, society, culture, and more. A lot of people talk about AGI as a utopian vision of the future where people aren't working and there's universal basic income. But what if there's a different vision of the future? What if people are working, they're working as developers, and they radically change segments of the economy that we never thought we could revolutionize? [02:49] education, industrials, and more. That's the topic of today's episode. Enjoy. [02:55] I'm Chad. Welcome. Thanks for coming on the podcast. Thank you. My pleasure. [03:00] You and I have known each other, I think, five years now, and I had the chance to invest in you four years ago at KOTU, so I've been able to see the journey. One thing that's been true for you since the first day I met you, and I think for maybe a decade even before I met you, is you've been talking about this idea of a billion developers coming on the Internet, which is a bold idea. Maybe it's gotten more consensus as AI has come around. Maybe tell us a little bit about that idea. When did that intuition first hit you?

[03:25] And how has that journey evolved? [03:27] Before we get to that, I actually remember the first time we met, I was in our [03:31] Bryant office, which is actually kind of a home, a loft. We're in the basement sitting downstairs. And so we worked slash lived in this really small place in San Francisco. [03:43] And in terms of the billion developers, it's just like, [03:47] Ever since I was a kid, I started programming really early on. My first experience with computers was when I was six years old. [03:56] And by seven, I was trying to make things with it. [04:00] The first program I made was for [04:03] my younger brother, to learn math. [04:05] And I've done such a good job at it that he works at Replit today. [04:12] It just always felt like making software is the natural thing to do on a computer. I was actually surprised that this is... [04:20] the domain of the expert as opposed to this thing that anyone can do. Um, [04:25] And through thinking about why is that the case, it just felt like a lot of tools were complicated. Actually, they were getting more complicated over time. So if you think when I was a teenager kind of building a business, it was visual basic. [04:40] And I can like, you know, make an app in Visual Basic with database and everything, kind of shrink wrap everything into an exe and sell it. [04:48] And then the web came along and it just felt like a lot more complicated. [04:53] It was more powerful. We can deliver things over the Internet.

[04:56] And then the complexity didn't stop. Like if you think about a JavaScript application today, [05:02] You know, there's a lot of things you need to do. You need to spend perhaps hours... [05:08] If you're new, you might spend days kind of setting up the development environment and learning all these aesthetic things like, you know, what is Webpack and what is translation, compilation and all that. [05:20] And it just felt kind of worse than when I started. And there was kind of no reason. I didn't feel like there was an intrinsic reason for that. There's all these perhaps social phenomena that made it so that programming is... [05:35] a lot more complex. And one is this decentralized nature of open source. When open source took over and it was in Microsoft product managers kind of designing how programming should look like. But I felt like even then you can have the open source ecosystem being this decentralized innovation machine, but you can create... [05:58] experiences on top of that by mixing the best of the open source to create [06:03] amazing experiences. [06:05] In the old language of open source hackers, there's the cathedral and the bazaar. And the idea of the bazaar is this complicated mess. And this is where, you know, open source software lives. And the cathedral is like something like Apple or Microsoft, where you're designing it top down. [06:20] But I was like, well, that's kind of a false dichotomy. We can build cathedrals from bazaars, right?

[06:27] And that's been kind of the driving motivation for Replit. And I felt like, okay, if you make programming something that more people can do by removing this complexity, a lot more people would want to use it. [06:50] to be a user of applications, which was the original vision for computing. [06:55] Um, [06:56] This was the drive behind that. Also, the opportunities presented by being a programmer is amazing. I'm sure we'll get into my story, but the fact that I was able to make it. [07:09] all this money when I was a kid was able to [07:12] get an O and B that I get to the U.S. And I felt that opportunity could be a lot more accessible to people. Maybe take us just to start, like take us 10 years into the future, or I don't know, you tell me how many years from now it is, when there are a billion developers, what does the world look like? How does the economy look? I mean, I think you have this sort of imagination about all the ways that we're going to build software and the ways that businesses are going to be built are different. And to me, it is somewhat of a utopian vision of the future. I know, tell us a little bit about what that looks like. Yeah, I try not to be too utopian. [07:42] Uh, but, [07:43] A few things on that. [07:46] One is the nature of if essentially anyone can program and most knowledge workers would want to, [07:54] develop or make applications or solve problems using software and AI, the nature of what it means to be a company kind of changes. Because if you think about companies today, we have these roles and we have these silos and kind of the way...

[08:09] companies are structured are based on [08:12] the sort of factory pipeline from the sort of industrial revolution. Actually, if you look at society today, a lot of it hasn't really been updated since the industrial revolution. So the main collaboration slash sort of work innovation from that era is the pipeline, is the idea, you know, you do this one thing and then you pass, you know, whatever you're making to the next person. Eventually, there's a car or a toy or whatever at the end of the factory chain. [08:42] And, you know, society is sort of designed, but that's like the main design principle that we have. And so you look at the school, for example, you start at like, you know, kindergarten, you go to first grade, second grade, and it's like everything is like kind of created like that. [08:56] And even companies, it's like... [08:58] or you have the, you know, the [09:01] The product manager creating a PRD and then goes to the designer and then goes to an engineer and then goes to a release engineer and then the product is out there, goes to marketers, goes to UX researchers. [09:17] always the silos and the [09:20] and this pipeline model. [09:22] If you have journalists that can make increasingly more complicated things and can use computing to its true core and to solve problems, I think you're… [09:34] you're going to have people in the organization that can solve problem across the board. So if you're someone in sales and you want to...

[09:43] you have an idea to drive more sales. You might spin up sort of like an SDR agent that does this one specific thing that's based on an idea that you have. There's no... [09:53] There's no separation, oh, well, I need to go to my boss to go kind of hire someone to do the SDR. You can actually spin up an agent that does that exact thing that you want to do. Perhaps maybe you're on a customer call and a customer is asking, oh, well, I'm going to go to my boss. [10:11] how can I do this X, Y, and Z with your product and perhaps your API or SDK? Now, traditionally, you have to go back to engineering and you have to ask them how to do that. But in this world, you can spin up something like Replit Agent and you can say, well, yeah, prototype this thing for me and show the customer in real time. By the way, all the stories that I'm saying are real things [10:41] problem solvers. And you see that in startups, right? We see that in startups, but I think [10:46] that's going to be the case at scale. [10:49] The other thing is... [10:52] The way we construct software, [10:54] I think will change. So if you think about how we construct software, again, when you look within one company, you see this like sort of factory pipeline, but also when you look at software, [11:08] The economy in general, you also see the same sort of separation between different companies and how the products get made. For example, the supply chain, right?

[11:16] in software we have some sort of supply chain where you have a look at [11:20] you know, database company and then you have a [11:22] sort of infrastructure hosting company. And then you have, uh, [11:25] you know, whatever front end company is delivering actually goods and services. [11:31] And I think going from the industrial age to the network era, I think the way we construct things will become more like a network. And so – [11:42] You can imagine the way software is constructed is if we have something like crypto, Bitcoin, stable coins, things that are able to let people to transact without having to know each other, trust each other. You can imagine software being constructed where I am. [12:04] sitting in front of software agents and I'm going to say, okay, well, I need to create this product [12:09] and the agent is gonna be like, oh, well, I'm gonna go grab this database from this area, this... [12:14] you know, thing that sends SMS or email from this area. And by the way, they're going to cost this much. And as an agent, I actually have a wallet, I'm going to be able to pay for them. And when I'm going to publish my software, my software is monetized. And whenever there's like a dollar that comes in, it kind of flows through the entire network. And so there's this ambient services that my software agent is able to compose without necessarily having any sort of centralized system. [12:44] by hackers and people that are making money on the internet.

[12:48] And so I think the nature of software itself and the nature of companies and perhaps the nature of economy would change when everyone can be a generalist software and AI agent creator. Maybe taking that last point on the economy to its lateral extreme, macroeconomics itself changes in a world where everyone's a developer. You have all these stories. A lot of your users are abroad. They're in other countries. You have users in countries. [13:13] How many countries do you have users in now? Every country. China is harder, but we have some users there, but basically every other country. How do you think the global economy changes, the way the economy works today, the people you're empowering? How does this change how these economies work? [13:30] So I think one concrete thing, I think Peter Thiel talked about this paradox of – [13:37] the internet where the internet was supposed to be the great equalizer, yet it centralized all the wealth in Silicon Valley. And this [13:45] this one area. [13:47] Um, [13:48] And so why is the case? Like this thing that is purely virtual could have been and perhaps should have been [13:56] much more decentralized. Um, [13:59] And there are all these things you can say about network effects and things like that. But, you know, ultimately, I think that part of the problem is, yes, we think about this technology as accessible, but it's not as accessible as we think it is. [14:15] For example, there's... [14:17] university student in India that was living in some rural area, but he was going to school to study computer science and didn't have a computer or laptop. He had an Android phone at home and he would go and repl it and start programming and learning how to code on his phone because we have a mobile app.

[14:40] And then he started picking up tasks. [14:43] from our Bounties platform. So Bounties is the ability for people on our platform to provide [14:50] services to others. Thinking behind it is like we have a lot of people with time [14:57] And then we have a lot of people with money and no time. So a lot of people with time and money and it's like obvious trade. [15:03] So a lot of people learn to go on a platform and they want to earn, so we want to present opportunities for them. [15:08] And so we had an entrepreneur here in the U.S., [15:12] um, [15:12] actually a technical recruiter building a recruiting app, but he was running into problems and he was able to go on our, [15:20] Bounty's platform hire this kid. [15:23] and for this guy to kind of fix some of those problems and for this entrepreneur to be able to ship his applications. [15:29] And that kid made money, made more money than his entire family would make for an entire year. [15:34] Um, [15:35] And so, yes, the opportunities will get distributed. Obviously, this is more providing labor services, but you can imagine it actually creating a company and creating something that could be scalable and that way. [15:48] Perhaps... [15:50] that this [15:51] massive wealth generation machine that we call the intranuts will be accessible to more people in the world. Yeah, that's amazing. When I think about your vision, I think... [15:59] it's fundamentally this empowering economic vision. I think it's a vision of wealth. It's a vision of prosperity. It's a vision of, [16:06] giving opportunity to people like and we'll talk about your story but people like you and your kid um i'm curious yeah to question to you like what do you think is the macro because you know you're more than a finance uh brain like what do you think the macro implications of of everything that we just talked about i think that if we go down this route of having a billion developers the question is what are all the sectors of the economy today they're not techified if you will they're gonna get techified and so you think about health care you think about education you think

[16:31] percentages of GDP that fundamentally have not been touched by Silicon Valley and touched by Silicon Valley technology. [16:38] And [16:39] The thing that I think about is, well, the barrier to entry... [16:42] was high for these industries. And so in order to build some of these industries, you had to go sell to them. It was difficult. It was tricky. And so what were the lowest barriers to entry? [16:50] consumer, you and I can go download Facebook on our phone and e-commerce, we want to buy things. And it's pretty easy to make money buying things. And so you think about social and e-commerce to sort of the initial engines that kind of get going as the economy was techified. [17:03] And this is where the... [17:05] you think about the wage rate for software engineers being a proxy for how valuable software engineers are. Um, [17:11] And so the wage rate has continued to go up. It's sort of one of these surprising things. You would think that as more supply of software engineers came into the market, the wage rate would go down. And what that tells you is the value that software engineers drive is actually higher than the wage rate. [17:27] companies like Replit, and as you see this next wave of developers come, [17:31] you're going to have this incredible prosperity from the value that they're going to go create. And so the thing that excites me is, what are all the sectors of the economy where we can go create this value such that, [17:41] the kid in India who's using Replit is going to extract some of that value and create prosperity for his family, but also such that the consumer at the end of that experience, call it a health care experience, call it a government experience, whatever it might be. [17:54] is also accruing value. And so we've seen that value creation, but to the extent that there's 100 million people, I think, who've created GitHub accounts today, and there's going to be a billion. And so you have this incredible increase.

[18:07] I think there's a lot more to come. There's also like a cultural impact, right? Like where Silicon Valley, you know, as much as we have this global view of the world and we can – [18:18] because we have so many immigrants, we can actually relate to a lot of those people. But there's limits to that. I remember when I was working at Facebook, [18:25] We were redesigning the photos experience on desktop. [18:29] And we designed this amazing kind of vertical scrolling experience where everything was scrolling after the iPhone came out. [18:36] And everyone was excited about it. And then we went and did an A/B test and all the metrics tanked. I was like, "What is this? Did we create some crappy product that didn't work?" And then the UX researchers did some tests and then nothing came back as not working. Everyone was happy about the product. Then one product manager actually looked and dug into the metric and found that most [19:02] Like most, like, you know, the majority of Facebook desktop users use it on the netbooks or like those websites. [19:11] those laptops where it's like wire screen and they don't have a lot of vertical space. And that was mind-blowing to me. Where everyone in Silicon Valley have these like amazing sort of MacBooks, [19:25] And so there are limits to how much we can relate. And I think the products that we build is often – [19:33] not as suitable to these culture and creates this flattening of the world. Whereas when you have

[19:40] sort of innovation, decentralized, I think they're going to be able to create applications that are more local, that can like benefit their communities. And I think there's less of that today. You've architected your product this way. Maybe you could, I think this is like one of the secrets of Replit that people don't fully understand is how do you make a product? How do you make a coding product that's actually good on mobile, right? It is sort of people think of this landscape and there's a lot of hard work that goes into that. How do you make it work on all these devices, right? Maybe talk to us a little bit how you've done that. [20:10] The main breakthrough of the open source project that we've become wrapped up the company was that I was the first to compile a bunch of programming languages to JavaScript to run in the browser. So this technology becomes WASM later on. But I was on the grand floor. It was like a research project by Mosella. It was called nscripten. And me and my friend were able to compile CPython. [20:37] to JavaScript using, using this technology. And we contributed a lot to it. We had to create, um, [20:43] sort of a Unix simulation layer in the browser. It was like a very complicated project. But it captures people's imagination. We'll put it up on Hacker News and went super viral. And people got really excited about it. I remember one highlight at the time was Brendan Eich, the inventor of JavaScript, like tweeting about it. It's like, wow, you know, we're like... [21:01] kids in Jordan like building this thing and people got people, you know, [21:05] Very important. People got excited about it. [21:07] um, [21:08] And at the time, I thought that was like the best thing.

[21:12] The problem is when I tried to load them on phones, they would crash, especially low-end Android phones. I even wanted to work on these Nokia Symbian phones at the time. It's because we were downloading tens of megabytes of JavaScript. [21:32] So it worked for a lot of people. People are excited about it. But the problem with client-side... [21:38] execution is that a lot of people's machines are just not very good. And so when we [21:45] When I went to work at Code Academy, and maybe I'm skipping a little bit in our story, again, we wanted more and more people in the world to learn how to code. And we started having users in Africa and other places like that. And the computers would not handle tens of megabytes of JavaScript to download. And then so I started building the sort of back-end execution environment such that what you're using on your phone is really a thin client. [22:15] and coming back to the client. And then the... [22:18] You make it to that, you hyper optimize the JavaScript application such that it's very small, it does incremental loading. [22:26] all of that stuff with react right now is really easy. But back in 2011, it was actually quite hard to do. We had to invent a lot of kind of web technologies to kind of make it, make it work. Um, [22:37] And, you know, still to this day, like if you go to the app store and download [22:41] the Ralfwood app, it's actually one of the smaller apps on the App Store. It's less than 100 megabytes, I think.

[22:49] Whereas most apps are on-door of gigabytes. [22:52] Why do you think that is? The IDE market is one of the, I think, the last markets to move to the cloud. You talked to a lot of Silicon Valley developers today. They still want the local thing on their laptop. You cannot make me run this thing in the cloud, whereas every other piece of software we've used, people have eventually given in. Google Sheets and Figma and all these amazing cloud-native companies. Why do you think developers have been the last market to move to the cloud? [23:22] maker is walking around barefoot or something like that. At least in Arabic, there's this saying where, you know, people who make things for other people are often not satisfying their own needs. I think there's some of that and I think there's some cultural aspect. Like, you know, you can't really, I can't really pinpoint a culture. [23:42] real sort of technical problem that can't be solved or can't be mitigated that prevents people from like coding in the cloud. I think a lot of it is cultural. There's this sense of control over the stuff on my machine. And there's this old kind of joke in programming where progress in programming happens one generation at a time. So you need the old generation to sort of retire for the new [24:09] Folks that are learning to code on Replen, Codecademy, and things like that, I think they're coming into the market now. They have no problem with it. [24:17] with the inner coding on cloud products.

[24:19] Another question. Back to the origin of the company. Did you first see that we were going to have LLMs and Sonnet and all these amazing models to actually transform coding back when you started the company originally? [24:32] So when I was working at Code Academy and then I went to work at Facebook, [24:38] and before that on the open source replet, what I was doing is wrangling code. So I was writing compilers, transpilers, parsers. [24:46] Um... [24:47] And those things are very fiddly and very complicated programs. And it just felt like, [24:57] you know, it just felt like this [24:59] this laborious thing that machines should be better at. It's like parsing this character and putting it in a node and trying to understand the structure of the program. [25:09] And I read this paper in 2012, which I think sort of is an underrated paper that should be legendary sort of on the order, maybe not on the order, but kind of similar to Attentions is All You Need. It's called On the Naturalness of Software. [25:23] And this group, they're making the argument that NLP can be applied to code. [25:32] And they make a statistical argument about how code kind of [25:37] is statistically like natural language. There isn't a lot of daylight between them. And so they built an Ingram model and they use the Ingram model to do completion. And by the way, you can scale Ingram models like Ingram models were the original language models and famously Google did this trillion parameter Ingram model that was pretty good at translating the very, very simple

[26:02] simple things that's trying to predict the next character based on frequency. [26:09] uh, [26:10] They constructed that thing and then it was performing well. It wasn't performing as well as IDE completions, but then they used [26:20] static methods like IDE completions to rank the completions from the language model. And the result was superior and programmers rated it as superior and actually did more completions than the static methods on their own or the classical kind of algorithmic methods. So at that point, it just felt like, okay, that's the future. Deep learning and things like that were coming on the scene where [26:48] and your networks. And if this very simple thing is able... [26:53] to [26:54] to have such results, I'm sure we're going to have better things. And so try to [27:01] do something like that, but it didn't work really well at the time. But actually, when I pitched, you know, I think the seed deck is out there on the internet, but one... [27:12] One slide in it, I try to do the Elon Musk master plan, which everyone does these days, but basically, the Elon Musk master plan, which... [27:22] really amazing. It's like, oh, we're going to build the Roadster. It's going to be expensive. We're going to use that money to fund the next generation and that will create economies of scale and then so on. [27:32] And they did it.

[27:33] Basically, it was like we're going to create this website for hobbyist learners and teachers, and we're going to grow it that way. And that way we're going to get a lot of data about how people use and learn programming. Then we're able to train machine learning models, and it's going to be this like AI powered, you know, tool. And then that would be a tool that's like more powerful than traditional software, you know, tools. [28:03] code, make things and deploy them. Um, and that was like in 2015. Um, [28:08] And so every year I would try to do something with machine learning. And the first time that I got a glimpse of something working was GPT-2. [28:18] And there was a lot of experiments at the time. A lot of the early community around GPT kind of had that feeling that, okay, this is working. Because you saw GPT-1, you saw GPT-2. Even GPT-2, there was like two sets of weights. [28:33] one more parameters, you could tell [28:35] Dot. [28:36] Scaling is working. [28:37] And so when GPT-3 came, it was amazing. It was like... [28:41] you know, groundbreaking event. For a lot of us, it was the ChatGPT moment. Because ChatGPT was not, it was like a UI innovation, perhaps some fine tuning. But it was really clear that it was here and it was coming and I started orienting the company to take advantage of that. [28:58] Do you want to talk to us a bit about Replit Agent? I mean, that feels like a transformative moment in this company and almost like the whole company. You got these millions of users. You had this community and almost like the whole company was built for this moment. Talk to us a bit about Replit Agent.

[29:12] Yeah, so... [29:14] every year [29:16] So we implemented co-pilot like features. We were actually the first startup to train code models outside of Microsoft and OpenAI. We opened SolarCenter for a small time. There were SOTA models. There were three billion parameters. [29:31] And that was really exciting. But the thing that I thought would be transformative is agents. [29:37] I actually had a thread that went viral at the time in 2022 about how software agents will change [29:43] programming and perhaps the economy. And so every year we tried it, I remember we tried it with GPT-3, the context window wrangling that we had to do was insane. Like we had to, every time we generated like a class or a function or whatever, we'd have to like summarize it like into one sentence and put it back into context to be able to kind of iterate. [30:07] So that didn't work. [30:09] uh 2023 you had auto gpt and things like that come out and you could tell like [30:14] oh, the models are getting better at like [30:16] you know [30:18] like being coherent a little longer. [30:20] but it wasn't working very well. [30:22] October 2023, I gave a talk. [30:27] uh, at Ted talk. Uh, and I talked about how, how in the future people make software. And in it, I talked about agents. I kind of predicted, uh, [30:37] Test time scaling as well, where I was saying there's going to be a trade-off between cost, speed, and accuracy. And sometimes you want to pour more compute or time into a task. And you're going to have these plans, and the agent will iterate on these plans and build the software. And you're going to be a human in the loop where you're sort of a creative engine. It was the plan. It was the roadmap for what we're going to build.

[31:04] And 24, I think around the time 4.0 came out, um... [31:08] It felt like we're almost there. And so I essentially put the entire company on that. And in fact, we actually did a layoff. [31:17] in May 2024, partly to kind of cut Burn and focus, [31:23] But the other part is like, [31:25] I knew in my bones that that was the way to go. And we were doing a lot of other random things that were unimportant. [31:31] So between sort of March... [31:34] The first time I got a demo from someone on the AI team, Zen on our team, he showed me something and I could immediately feel like we're there. It was almost like a baby kind of agent. [31:50] So we spent [31:53] seven, eight months working on it. [31:55] And then, um, [31:58] It just felt like... [32:00] We were on this treadmill fixing one bug after the other, trying to make it coherent. The transformative thing that happened is like June 24 when Sonnet 3.5 came out and Sonnet 3.5 had two important properties. [32:16] Actually, the main property that people don't really talk about all that much is Sonnet was able to generate... [32:23] Uh... [32:24] A [32:25] uh, [32:26] like thousands of tokens. [32:28] I think up to 32,000 tokens [32:31] of [32:32] code, coherence code. So it can one-shot a repo almost. Whereas with GPT-4-0, when you're doing agents, you're writing one function at a time and then testing. If you remember the cognition demo, it's like, oh, they write one thing, they go to the browser, they write one thing, they go to the browser. And that was intractable. That was very, very expensive. And we thought that's not going to work. And the thing that's on it, bro, at the time is writing high-quality code and not being lazy about it.

[33:02] products were lazy. There's a real problem of laziness where it's a [33:07] it, it like, you ask it to do something and, and, and, [33:10] At one point it adds a comment and says the rest of the code here. [33:14] Bro, I asked you to make that program. Why are you asking me? [33:18] It is like a human. Yeah. [33:24] And so Sonnet was this huge unlock, and we felt like we were going to be able to ship soon. [33:30] Uh, you know, there's always this tension, you know, the quality and when do you own a ship? And so we set, um, we set a deadline for my birthday, uh, September 5th. And, and, and it was so contentious. Actually, at least one person in the team kind of quit because they thought we're not ready and he was partially right. Um, yeah. [33:49] And we launched Agent and there was a lot of excitement because it really is the first software agent that works on the market, the very first. And people got really excited because, oh, you can try this thing and you can get a glimpse of the future. Oh, this thing can bring in a database and like provision it for you. It can run a migration. It can like deploy my app. Yeah. [34:11] But then as it escaped the kind of early adopters, a lot of people were disappointed because it was actually kind of crappy. [34:21] And so we sprinted between September and then December. We had something we were really proud of and we exited the beta. And although we're growing well, then we grew really fast from there and people got a lot of value out of it.

[34:35] Now we have version two coming. [34:38] and it's rolled out to some beta users. [34:42] And we've done A/B tests and it is [34:45] um... [34:46] is anywhere, depending on the metric we're looking at, between 50 and 500% better. Wow. So it created 50% better daily retention, 50% conversion, 50% better... [34:58] engagement, but the metric that was really great was, um, [35:03] new users are twice more likely to deploy an app that they made with agents and power users [35:11] are five, six times more likely to deploy an app that they made. And deployment of Replit is very, very important because that's what makes Replit great is that you can go from an idea to deploy a thing and we see [35:26] people who deploy are 10 times more likely to retain on Replet. So that metric we track very well. So now I look at V1, I'm like this crappy... [35:35] thing. We got to [35:36] We got to ship V2 as soon as possible on the team, and so we're starting to roll it out now, and it's really, really good. [35:44] Can you give us some of the intuition for how you've gotten it so much better? Because I imagine it's still Sonnet 3.5. That still seems like this other model that everyone likes right now under the hood. Maybe what's the behind the scenes for how the agent actually works and how you've gotten it to be so much better when the base models are very similar? [36:00] Yeah. So, um, [36:02] At the 3.5 era, we had to build this fairly complicated multi-agent system where we're talking about land graph.

[36:12] earlier where there was a state machine component to it. [36:16] And there's a lot of [36:18] non-AI... [36:20] Uh... [36:21] state transitions that we're having to make so that the more you put into the eye, the worse decisions it often makes and less coherent. So context window was still very important because there are actually studies showing that although, you know, these models are, [36:39] kind of advertising a million tokens actually after [36:44] something like, you know, 32,000 tokens, reasoning and a lot of benchmarks just like tank like crazy, right? [36:54] And so there's a lot of context window wrangling. And so we had a manager and it had different context than the editor, than the debugger, because we had to kind of make sure they, they're isolated and have their different memories and things like that. Um, [37:11] And, you know, the tooling between all these things had to be different. We built our own sort of protocol for... [37:19] the agent to call tools because actually tool calling was not very reliable at the time. It would like hallucinate tools or hallucinate arguments or, [37:27] So there was a lot of engineering behind it to make it work. [37:30] When computer use from Anthropik came out, [37:34] So that was 3.5 v2. The moment we laid our eyes on it, we felt that there's something important that wasn't getting marketed. That was a start of a transition, which is it looks like they actually fine-tuned it for long-horizon reasoning.

[37:56] Because if you want something to do computer use, you're going to have... [37:59] all these images and contacts and it's going to have to [38:02] kind of continuously click on things, reason, think about it. [38:05] what to do next, click on the other thing. And so you can actually roll out [38:10] a [38:11] long chain of tool calls. [38:14] without [38:16] that big of a degradation in [38:18] and reasoning. [38:20] And so we started re-architecting [38:23] the [38:25] the model of the agent to be, [38:28] uh, [38:28] let's call it, less multi-agent, [38:32] and more single threaded because the models are getting good at it. And that's like a much bigger simplification. It added other challenges, but it's a much bigger simplification over [38:43] existing model [38:45] And so kind of the lesson to learn is that you have to constantly rewrite the systems. [38:50] because you want to make use of the next version of it, and you want to be able to predict what's coming down the line. [38:57] When 3.7 came out, [38:59] you know, Chris had a [39:01] big problem with integrating it. Everyone hated it on Cursor. I think they just recently fixed it. [39:07] And the reason is because, again, 3.7 is more agentic. [39:11] And so when you try to use it and it's a composer style request response, it is not very good at that. It's actually worse at that because it's trying to kind of make decisions, but you're not giving it the space to... [39:23] to kind of do this looping, [39:25] um,

[39:26] And so... [39:28] And so there's a really kind of tough lesson for engineering and product teams. If you over-optimize for the present capabilities, you're going to enter local [39:39] uh, [39:40] maxima that is quite hard to move out of. There's like an innovator's dilemma problem on the order of months. Innovator's dilemma used to be like on the order of decades, right? Where a company has an innovation, reaches a certain height of success, [39:58] based on a product that they made that people really love. But then there's a disruptive technology and they don't make use of it because it might be destructive for their current business where it cannibalizes their business. And so they tend to [40:12] not pay attention to it and fight it. The classic example is like, [40:16] Kodak had a digital camera product, but they didn't actually launch it because I was going to cannibalize their film business, which was the thing that's making the money, actually. And this kind of thing is happening on the order of months, which is really hard to wrap your head around. And so they move fast, break things. [40:33] is actually very, very important today, right? More than ever. And you have to be okay with your... [40:39] with sometimes like delivering crap experiences. [40:43] You mentioned cursor. What do you make of the entire landscape? It feels like the landscape of coding tools has just exploded in the last six months. And, you know, developers are reaching for new ideas. You know, many more people who didn't consider themselves developers before entering the market. What's your view of the market landscape and how do you think of as, you know, the ideal person that chooses replets versus one of these other tools?

[41:07] Yeah, so I think... [41:09] there are incremental [41:11] uh, innovations and there are sort of more, uh, [41:16] disruptive clean slate innovations. [41:19] and I would put roughly in the latter category. So, [41:23] So you take VS Code and you build an amazing AI experience on top of it. [41:28] ReaderCursor is like a fantastic experience. A lot of people on our team use it. [41:32] um, [41:33] the [41:34] The thing is, it is by definition, Kamal. You took... [41:37] a piece of software that Microsoft has been building for over a decade and you added [41:43] this. [41:43] you're much better experienced on top of it, but you're still [41:47] you're still kind of, it is still... [41:49] Literally. [41:50] this additional layer. [41:52] with [41:53] With Rapplet Agent, we actually took a clean slate approach to that. And I was like, okay, what do we need to make it that people aren't coding at all? [42:02] Like, you know, we want to kind of drive towards this vision of no coding, not only no coding, no DevOps, no IT. Like, we don't want you to set up a database. We don't want you to write migrations. Like, writing migrations is the worst thing ever. Like, you know, people don't talk about this, but one of the worst things about building software is being able to keep your version control, your database schema, and your, you know, sort of deployment settings and configuration in concert, in lockstep. [42:32] um, [42:33] And most, actually most averages, [42:35] When you read the postmortems from Google Cloud or AWS, it's like a configuration error. And usually it's an outdated configuration. The software changes, but they forget an environment variable configuration. And so when you have an agent actually do these things, it's actually much better at doing that. And so the system that we built for the agent is this sort of transactional lockstep system that does these transformation one at a time.

[43:05] Bye. [43:06] reverts the code, but also reverts the database changes. [43:10] and other environmental thing, environment variables, things like that. So there's all these things that we thought about where... [43:21] you as a developer coming to Replet, the thing you have to worry about the most is [43:27] is your ideas. Mm-hmm. [43:29] You know, you don't have to worry about where do I... [43:33] Get. [43:34] object storage and how do I configure my buckets? [43:37] Right? [43:38] Whereas when you're sitting in cursor, [43:40] Um, [43:41] There's either someone in your company, a backend engineer, worrying about these things, and you're basically hooking into these APIs. [43:47] or you're having to build them from scratch and like, [43:50] It would help you hit those APIs, but you still have to architect the larger system. [43:56] So I think there's like a fundamental category [43:58] deference. [43:59] between these things. [44:00] These products are hooking into existing systems and require a lot of existing support around these systems, whereas Replet is trying to be. [44:08] the [44:09] final tool you have to adopt in order to build a piece of software. [44:14] One thing I've heard from users when I talk to them about Replit Agent and deploying on Replit, one of the things that people love is that you can go straight to deployment and that Replit [44:21] does the whole thing. You start from nothing and you end up with a fully deployed system. Can you talk a little bit about that? How do you think the deployment space plays into this? You guys, in some ways... [44:31] are not comparable to a lot of these other companies in the sense that you can go all the way on Replit. How do you think about that last mile? Who cares about that? And how does that play into the long term vision?

[44:41] Yeah, so... [44:42] In the same way that I can say that cursor helps you with coding, doesn't help you with the other stuff. [44:47] First hell, hops with hosting doesn't help you with the other stuff, right? And again... [44:51] It's like... [44:52] this back to the cathedral and the bazaar, sort of like a bazaar, and you need to kind of [44:57] find all these tools and configure them and figure out [45:00] Whereas sort of Rapalit, I mean, another sushi analogy, there's a la carte and there's amakase. Rapalit is amakase. We're going to make, we think we have great taste. [45:09] or are going to make the design choices for you. [45:12] And it's not for a lot of people. If you want to make all these choices, don't come to Replet. I mean, there's a lot of other products out there. So in terms of like the deployment system, if you want to reach a billion people eventually, like making software, [45:30] a priori, most of them will not know how to set up a [45:34] deployment environment. And therefore, the Robler product needs to have a deployment environment. Basically, that's how we approach product in general. [45:42] you know [45:43] In the early days of Replit, there was always this narrative of it's students, it's young people, it's people learning how to code. It feels like in recent years that's really changed and maybe AI has enabled that. I was talking to a Netflix engineer who's telling me about how he's using Replit. Paul Graham has obviously always been a big advocate of Replit and one of the early believers in the company. [46:03] How do you think the user persona has changed and how will it continue to evolve over time? How amazing of a visionary is Paul? How can he look at crappy...

[46:12] uh toy like replet and just see that oh that this one day could could be really big or the idea that people are leaving uh you know are sleeping on mattresses and other people's homes that's going to be like a hundred billion dollar company um it's incredible yeah it is it is fascinating um [46:29] He's also an incredibly good human and he's been very supportive during very difficult times in the company when others weren't. It was frustrating for a long time that, oh, Rapaleta is a toy, a hobby toy, students, kids, teenagers. [46:46] On the one hand, I was excited by it because we had a group of users that are willing to experiment, that are willing to try to find a future of software, the future of what we just talked about, of being able to have an economy built into the software infrastructure. And we tried a lot of that, and I think it's going to work. The main missing thing was the agent, but all these things are going to work. Yeah. [47:11] there's this co-evolution that happened with these users that was very important. The other thing that [47:17] that perception was important because, you know, [47:19] um, [47:20] sort of competitors weren't paying attention. They looked at Refl and was like, oh, that's a toy thing. Why would I pay attention to it? [47:27] uh, [47:28] A very close friend of mine and someone who's looked up to in the industry, he's a very [47:34] visit us the other day and visit, you know, a lot of our friends, [47:39] and he was telling him he was using Replit to kind of prototype ideas for a new company. And everyone was like, well, isn't that like the toy thing that I'm just working on? I was like, well, no, it's very useful for me. And that's how I'm able to iterate on these ideas very quickly. So now, yeah, the perception is changing.

[47:57] Um, [47:58] And partly, look, I mean, we took explicit decisions to make the product more premium. Actually, what's fascinating now is Replit is the most expensive product on the market. [48:10] with Cursor Composer, I think one request is four cents. [48:14] With Replite Warner Quest, it's 25 cents. [48:17] And the reason we did that is like we just really want to build an agent. We don't want to build [48:23] code gen tool. And if you want to build them agents, it's going to be expensive. It's going to [48:28] iterate and and it's going to call a bunch of tools with every request and so [48:34] It did alienate some users, and that partly also shifted the perception and kind of more professionals got on board. But if we're going to reach a billion people, we need to... [48:48] We need to... [48:48] go down market again. But I think [48:51] I think we started in one end and now we're starting in another end. We'll meet in the middle, I guess, at some point. [48:58] I guess that Raptoid is sort of the more roadster now. [49:02] And as we drive costs down by... [49:05] Perhaps using more open source models and things like that, it's going to become more and more acceptable to people. [49:12] What do you think of Vibes Coding? [49:14] I don't really like this term. I think you shouldn't fight it because [49:18] Uh, [49:19] I actually didn't like Gen AI either. [49:22] It's just, I don't know. Same. Yeah. [49:27] I don't know, it just like cheapens the...

[49:30] possibilities. Yeah. It's like, oh, this thing that generates things. [49:35] Well, actually you can build agents and not just generate things. We can actually reason. [49:39] Um... [49:41] And vibe coding is... [49:44] Makes sense if you're sort of starting from a position of coder and you're Andre Karpathy and you don't want to kind of worry too much about the code and you keep hitting enter or whatever. But if you're starting with Rathlet, you're actually not starting from a position of code. You're starting from… [50:02] an idea, you're starting from an idea that you're iterating on and, uh, and then you go in and the agent is unfolding this code in front of you. Actually, when you're using Rapplet Agent, you don't have, uh, [50:15] the luxury to look at the code. So now we have another product called Assistant. [50:21] and assistant is for more advanced people. And that you can do more by coding there because it's like a request response and you can like kind of review the code and do all of that. [50:32] um, [50:34] If I were to explain Replet, it's just vibe. [50:37] Like Dunk Court vibe. [50:38] not five couldn't just five you've always been a bit of a contrarian one idea that's very popular now in Silicon Valley is AGI there's this idea there's going to be no software engineers and you sort of have the opposite view there's going to be a billion software engineers and [50:51] Talk to how do you think about AGI? What is AGI in your view? And how do you what would you say to all the people who say I talked to a lot of software engineers and they're like, well, I'm not gonna have a job. I'm worried about not having a job. I'm worried about no one's gonna have a job. We're all just gonna be on universal basic income and all this stuff like.

[51:07] Replit is in some ways the opposite vision, right? Of like, we're actually all going to have jobs. They're just going to be very powerful and we're going to be able to do [51:13] all these amazing things. I think it's like a fundamental, it's philosophical difference. Like what, you know, what is special about humans and what's replicable in the machines and at least in the near term. My view is that AI is going to get [51:34] Really good. [51:36] at, um, [51:37] sort of two things, things that are highly represented in the data and things that you can construct a very good RL environment for. [51:46] So what can you construct a great RL environment for? [51:49] Um... [51:50] like obviously with AlphaZero games, right? Games are famous for, you can have these self-play sort of algorithms that develop over time. Now with reasoning models, [52:03] You know, math is an environment that especially with lean learning, [52:08] It's like a code, almost like an expression of math that can be executed. [52:14] that's like a great... [52:16] you know, RL environment, I think code execution as well. So running the code and, and then, you know, um, [52:23] doing reinforcement learning on that. [52:26] And things that are already represented on GitHub and things like that. [52:30] But there's a lot of other domains where we actually still don't know what we're doing. [52:35] how we're going to make them better, like fundamentally new ideas, new knowledge.

[52:41] it's not entirely clear how we're going to get there. Can you use RL for these more software things? Perhaps you create a word model, you can approximate these things, but I feel like the... [52:53] sort of the ideas [52:55] Um, [52:56] And the creativity and the sense of like, [52:59] coming up with really novel things and understanding the world in a very complicated, intractable way. [53:07] Coming up with an idea that could fundamentally change how things work or change the world, I think will still be the domain of the human. And AGI will have AGI, but it would be a functional AGI, meaning it would do… [53:20] the jobs that a lot of humans are doing today by virtue of the, [53:26] training data being available and by virtue of some of these jobs, having ground truth that you can train on. And the reason I call it functional AGI is because... [53:35] is fundamentally not general in that you can throw it in a super novel environment and for it to efficiently learn [53:46] things, especially when they're [53:49] when there's not explicit feedback and be able to [53:54] uh, [53:55] to be successful in that environment, which, uh, which, uh, [53:59] which, you know, the definition of the universal AI, which... [54:03] It doesn't feel like we're trending that way, but I do think that you can reach something... [54:07] So the definition of AGI at a lot of these companies is doing economically useful activities in front of a computer.

[54:15] It's like a remote worker is what AGI. I feel like we're going to get there. Uh, but, um, [54:21] If I have a remote worker, I'm going to create 100 more workers and implement all my ideas. Yeah. [54:27] And it's still a tool. It's useful for me. Is it going to replace me? Well, if I am like a code monkey, it's going to replace me. But if I see my... [54:37] place in the world as someone who can generate ideas and create products and services because I understand what people want and how... [54:47] the economy works and all of that, I think that's still irreplaceable. [54:50] I want to talk to you a little bit about your life story. You touched on coming to the U.S., the ON visa growing up in Jordan. Maybe start at the beginning for people who don't know your story because it's a really inspiring story. And in some ways, your whole life has built up to replicate. It's not just a company. I mean, your wife works at the company. Your cousin works at the company. This is like your whole life and your whole mission. All in. Bring us to the beginning. How did this all come together? How did you get to the U.S.? How did you get into Silicon Valley? Yeah. [55:18] One of my first memories, just as a child, I don't know if you remember your first memory, but I remember very vividly that my father kind of getting this machine and opening this machine. [55:32] this box and like putting it together. And I was fascinated by it. [55:35] even before I knew what it does. [55:38] And I walk over him and I kind of look over his shoulder, him sitting at the keyboard. [55:45] sort of

[55:47] Finger... [55:48] Instead of typing, he had a big manual. You throw it at someone, it could really hurt them. He was reading into it, but one by one, he was finger typing, CD, MKDIR, these DOS commands. [56:05] And I remember feeling that it was just, wow. [56:10] This is like a... [56:12] This is a machine that you can talk to. [56:15] Like, you know, what is a DOS? It is like a REPL. You know, this is where the REPL name comes from, read, eval, print, loop. Like you can... [56:22] Uh... [56:24] have a essentially a conversation with this machine uh so my father would like go and like attend uh classes to learn uh computer computers and he had like you know every night he would come in with all these notebooks uh he's like such a such a big nerd um the apple doesn't fall far from the tree yeah [56:45] I'm different than my father. My father is a Palestinian refugee, grew up with this intense focus on education. If we're going to make it, we're going to have to be the best educated. We're going to have to be better than anyone else. So he's like a... [57:05] A students and working really hard. [57:08] My mom is sort of the opposite. My mom is like a sort of a free spirit. She was into poetry. She taught me a lot of poetry when I was a kid. And I was actually able to, it was like there's an art to telling Arabic poetry. And so,

[57:31] I was this mix of two things because I can be very intuitive. I can be, uh, I can go, uh, [57:38] purely in intuition and take a lot of risk based on some kind of idea or vision or something that I feel good about. And I also have this very... [57:47] analytical mind can certainly be such a [57:51] uh... [57:52] pain in the butt about being so rigorous about certain things. And that's really my father's inspiration. And so one day my father comes home and I'm sitting in front of the computer and he was mad. He put all the savings into this machine. And I was like, [58:14] you know, opened the computer apart, took the parts out, put them back in. And I was like, you know, don't get angry. I know exactly how to use it. I know exactly how it works. [58:24] And I... [58:25] I showed him how to use it. I showed him essentially how to create things, how to open applications that he's been studying all this time. And it was like, okay, this is fascinating. You know, [58:35] Right. It's yours. And the first program that I wrote, I think I mentioned it earlier, was to teach my younger brother math. And I thought, okay, I can use this thing, this conversational machine. I still think about it that way. Obviously, with AI, that really happened. Yeah. And this idea that you can sit in front of it, learn something. [58:58] play games, do something fun. [59:01] And so that was the first program that I built.

[59:04] And then when I was a teenager, I was really – [59:08] obsessed in Counter-Strike. So I would go to these internet and LAN gaming cafes and I would like [59:16] you know, [59:16] whatever money scraps of money i have i would like put it into that and just play play a lot of counter-strike and strategy games and things like that i got very good at them to the point that it was like a source of income i was actually winning tournaments and things like that uh yeah so it's got into esports uh early on so that's like another bench of my life but um [59:39] One of the things I noticed about those businesses is they were running on pen and paper. I'm like, you have all these computers. You can just write software for it. [59:50] So I did write this client-service software that did accounting, that get people accounts and username and password, and manage their time in the system. And I also did security so that people can't format their computers or install malware and start selling that. And I made a lot of money on that. I sold it to a lot of businesses in Jordan. [1:00:16] Um, [1:00:17] By the time I got to college, I had this idea that AI is going to get so good, we're not going to have to write software. It just felt like… [1:00:27] Software is this pedestrian thing that you do. It's not that interesting. You just do it to make things.

[1:00:36] And at the time, like these wizards of code generations were coming out from Microsoft. [1:00:40] So when I went to school, I actually studied... [1:00:44] more on the electrical engineering side. [1:00:46] especially because my father thought that computer science was not a real field, because the Engineering Association of Jordan would not admit you, [1:00:56] as an engineer if you don't have electrical engineering. And by the way, my father is the vice president of this organization. I didn't know that. I learned something new. He really loves that engineering organization. So, [1:01:10] But throughout my college experience, I got really into programming languages. I started reading Paul Graham, started reading Hacker News. Paul Graham wrote a lot on... [1:01:18] on Lisp. He actually has a book called OnLisp. And Paul Graham's view of [1:01:24] Program languages is more of an art. [1:01:28] rather than science. These artifacts are aesthetic artifacts and not just functional artifacts. [1:01:36] And that also played into... [1:01:39] the reason to create Rappler because I wanted to try all these programming languages. And there wasn't a place on the internet to be able to try all these programming languages. [1:01:47] And after I created Repl.in, had that breakthrough that I talked about earlier, [1:01:52] it went viral in the U S a bunch of, um, [1:01:55] tech companies started adopting it. If you remember 2010-11, there was the MOOC [1:02:01] kind of hype. You know, we have the AI hype now. There's a like blip period where there's like the massive online courses, right?

[1:02:08] Udacity came out at that period, Coursera came out, Code Academy. [1:02:12] Uh, and I got a bunch of, um, they all started using replet by the way. And I got a bunch of offers. [1:02:18] I eventually decided to come to the US, got an O1 visa, [1:02:22] because a lot of my work was published in the news and everywhere else. And so it was possible to go no one visa [1:02:29] Landed in New York. [1:02:31] early 2011 and [1:02:32] uh, [1:02:33] The only money that I had [1:02:35] was taken from me at the airport. [1:02:38] And the reason is... [1:02:39] in the airport in Amman. [1:02:42] I had like perhaps $700. Those are the money that I'm going to the U.S. with. And [1:02:49] The folks at the airport did not know what a no one visa is. Apparently, no one in Jordan had ever gotten a no one visa to get to the U.S., [1:02:58] And they didn't think it was like a... [1:03:01] a resident visa. They thought it was like a visitor visa and that I needed a ticket back. And I was like, no, it's like, I can go there. I can work. It's a work visa. And they didn't believe me. I was like, go look it up on the internet. They didn't believe me. And they made me buy a ticket back. And that was something like 500, 600 dollars. So I arrived there with about a hundred bucks. Yeah. [1:03:21] And my salary was $80,000, $70,000 or $80,000 in New York City. [1:03:26] That's pretty vague. Yeah. [1:03:28] Well, not in New York City, when your rent is like $2,000. [1:03:32] I think at some point you had someone offered to buy the company for a lot of money. How did you decide to turn that down? That seems like that was a pretty big decision. Coming from this background, you grew up, you...

[1:03:40] made it to the U.S., you got this job, and then you have an offer to buy your company for... [1:03:44] life-changing amount of money. Insane amount of money. We were like six people at the time and you know, [1:03:50] The numbers that were thrown around is between $500 million to $1 billion. We're six people. [1:03:55] um, [1:03:56] And so it would have made me insanely rich, right? [1:04:01] So, uh... [1:04:02] And [1:04:04] And it was a tough time. [1:04:08] I wasn't really happy about the culture then. We grew to six, seven people or something like that from the three that are family, essentially. [1:04:15] And we hired people that I didn't like very much and the culture was changing and [1:04:20] And I kind of wanted to do a reset. [1:04:24] And at the same time, my mom was diagnosed with cancer back in Jordan, and I had to go back. The entire family had to go back, which is half the team, to spend time with my mother, and it was a very stressful time. [1:04:41] um... [1:04:42] And, um, [1:04:44] The thing that [1:04:46] made me, which is like the absolute rational thing to do is to stake the money, go home, and, and, um, [1:04:55] live, I guess, happily ever after or something. [1:04:58] But, um... [1:05:00] I felt two things. [1:05:01] One is [1:05:05] my dreams, right? I've had the dream of being in Silicon Valley for so long. Like the first time I knew about Silicon Valley is through a,

[1:05:13] low-budget movie called The Pirates of Silicon Valley. [1:05:17] where it's the [1:05:18] dramatized fight between Steve Jobs and Bill Gates. [1:05:24] And I was like, wow, this Silicon Valley place is like, [1:05:26] There must be like flying cars and like really... [1:05:30] advanced technology, of course, you come here and it's like [1:05:33] That's like the suburbs. [1:05:36] But I always wanted to be here. I thought the innovation... [1:05:39] I was reading about all these entrepreneurs. [1:05:43] I felt like this is the most important thing and those people are heroes. [1:05:48] And I felt like the weight of Silicon Valley also on my shoulders because [1:05:52] Um, [1:05:54] Paul Graham, you know, Mark Andreessen, [1:05:57] and other people that I really respect invested in the company. And they all have very high hopes for it. And they're all really excited about it. And I felt like I don't want to let people down. [1:06:11] And I felt like if I sold the company, it would have been [1:06:14] I wouldn't have achieved the potential of it. [1:06:20] and maybe I would regret it in the future. [1:06:23] And, um, [1:06:24] And it's like, [1:06:26] okay, being rich is good. I think money is actually great. And [1:06:31] and improves your lives in many ways, allows you to focus on things you love. [1:06:37] But... [1:06:38] What are you going to be? You're going to be another... [1:06:41] you know rich silicon valley dude and there's a lot of them you know you know like write invest do angel investing is like your life has gone a little boring and and you never kind of want to take the pain again unless you're ale musk to kind of like go start a start another company and put your life's force and energy into it and so for all these reasons uh we decided to to turn it down

[1:07:04] Now you have 40 million users? [1:07:06] Yes. [1:07:07] and, and, uh, [1:07:09] you know, we achieved evaluation over the, uh, [1:07:12] what were we going to sold for? And I think the company is actually underpriced now. [1:07:15] It's a lot of grit. [1:07:16] to get here from six people. [1:07:18] I want to talk a bit about your management style. You have a unique management style. I remember seeing one time on the internet, you said, like, I'm now the VP of engineering of Replit. I feel like over the years, there's always been... [1:07:27] you're a hands-on leader, [1:07:29] Your leadership style maybe has become more popular, but it wasn't popular six, seven years ago when you were doing this. How did you develop your leadership style? What advice do you have for founders as they think about running their companies? In some way, it's a deficiency. [1:07:59] all of that [1:08:00] um, [1:08:01] And in some way, I would say it is a perhaps lack of skill in terms of like how to do traditional management. [1:08:10] The management style that I have is similar to how I would lead an open source project or how I would lead a sports team back home. I've always been a leader and I've always been sort of a hands-on leader where it is this... [1:08:29] duality of [1:08:31] uh, [1:08:32] micromanaging [1:08:33] and trusting people. It is actually not at odds...

[1:08:38] to do these things. And the most inspiring leaders both can go dig into details and give very precise direction, but then really trust people to deliver on those things and also trust people to have their own innovation, their own ideas. [1:08:53] Um, [1:08:54] The way I managed the company from the start is I had [1:08:56] the, um, [1:08:58] a text file inside Rap Lit. I use it as a set of a notebook. [1:09:01] And the text file had everyone's names and the one thing that I think they should be working on or the one thing that I expect them to deliver on. [1:09:09] And every week when I meet everyone, we would go around the table and I would tell them, did you do this thing? [1:09:15] And it's either yes or no or something that will go right or something like that. [1:09:21] And then the other question, what are you going to do next week? [1:09:24] And so it's like, [1:09:25] okay, they did that last week or they didn't do it. [1:09:27] Why did they fail? What happened? And here's what they're going to do the next week. By the way, my execs still... [1:09:34] that everyone in the company [1:09:36] Every week. [1:09:37] On Friday they write, [1:09:39] Here's what I got done this week. Here's what I'm planning to do next week. [1:09:43] And so, uh, [1:09:45] And so [1:09:46] I can still keep in my head [1:09:48] what most of the company can do or is working on. Like I can walk around and tell you this guy's working on this. This person is working on this. [1:09:58] Um, so partly as I can keep a lot of complexity in my head and I can like really be able to kind of, um, [1:10:05] make these, um, these, you know, very deep decision about how, like a,

[1:10:12] one button should work inside the product or how certain marketing ideas we should run and [1:10:21] And so I can go [1:10:22] between all these different departments and be able to go all in into the details and then kind of zoom back out. [1:10:30] Um... [1:10:31] And also, you know, [1:10:33] just having really high expectations. If someone, if week over week, that person who said they were going to get that thing done and they couldn't, it's like an obvious... [1:10:43] reason to let them go. It's like not that complicated. And so [1:10:47] Replit has a high attrition rate, especially in the first few months of people joining. [1:10:54] 20, 30% of people are going to either leave or get let go. [1:10:59] because they're not able to keep pace of the environment or they get confused about how to work in that kind of environment. [1:11:06] The other thing I think that's unique about your culture, I remember I attended some all-team dinner at an all-you-can-eat barbecue restaurant. I don't know if you remember this. And there's all these young... [1:11:14] people around the table and some of them didn't even go to college, graduate college. I mean, you seem to recruit for raw talent. Yes. And that's something that a lot of people talk about doing, they want to do, but it's hard to do. How do you do that? How do you filter people? How do you find these people? How do you give them leeway? How do you train them? Talk to us a little bit about how that, how that happens. Yeah. First of all, I am, you need to be able to work with weirdos and misfits. [1:11:40] um, [1:11:42] And I was...

[1:11:43] able [1:11:44] to do that, whether it's instinctively or whether I relate, [1:11:47] to them being sort of myself, Orido and Misred, [1:11:51] And, uh, and I can see talent, certain people, like even back in, in my college years, like finding those people that like have hidden talents and being able to harness it somehow. I had this feeling of like. [1:12:04] If someone has a talent that is not being put to good use, I feel like a sense of waste. This needs to be harnessed somehow and I think that's sort of a good… [1:12:16] management skill [1:12:18] Um... [1:12:19] But in terms of... So first of all, you shouldn't have allergy towards these people. I remember... [1:12:27] uh... [1:12:28] my wife and co-founder Haya when we hired Mason, he was like 18 year old kid. He, um, he was [1:12:35] I joke that he's a runaway kid from Santa Cruz, which is partly true. He lost Santa Cruz in high school. He wasn't happy. He wasn't happy with his family, with his school. He went up to the city. He went to... [1:12:47] Um, [1:12:48] one of those boot camps, [1:12:51] Those bootcamps were using Replit, all of them. And one of them was sending me bugger parts. I'm like, look, I don't have a lot of people who work on this. Why don't you send me your best engineers? [1:13:01] It's like, look, I'm going to send you this kid who's really awkward, but he's one of our best. [1:13:05] So he came in and I remember Haya one day was like, oh, this kid like doesn't. [1:13:10] doesn't keep the door open behind him. He literally slams the door on me. He's like, well, he's not really thinking about you. He's thinking about code as he's walking around.

[1:13:20] And still to this day, I see people doing that. [1:13:25] the first thing is most people I think just reject these people outright when they can't communicate with them or they can't relate to them and then the other thing like [1:13:37] uh, [1:13:38] you know, they're going to spike on certain things, and then they're going to... [1:13:42] be not great at some other things, right? And so, but, you know, [1:13:48] you can construct a team where it all fits together. Hmm. [1:13:51] where the peaks of someone is the values of someone else, right? Someone is really good at [1:13:59] you know, [1:14:00] Shipping. [1:14:01] thousands of lines of code and someone else is really good at testing and very methodical about code reviews and [1:14:07] If you have one who's like a cowboy slinging code and another is like a little more careful, [1:14:12] and like meticulous and rigorous and perhaps... [1:14:15] a little annoying of how meticulous they are. Put these two things together. There's a lot of tension, but at the end of the day, you get a good product. So you want to... [1:14:26] balance the team in a way. [1:14:30] And you... [1:14:32] want to go to hire from places that other people aren't going to. And, you know, that's, [1:14:38] that gives you an advantage because everyone's competing over the existing talents on the market. [1:14:44] So for us, a lot of it was going to Repl.it. A lot of it was running hackathons or running sort of prizes on Repl.it and

[1:14:52] Uh, and, you know, some kids win and we fly them out to, to SF to work with us. And, you know, we've had anywhere from 16, 17, 18, 21 years old, uh, you know, join the team. And even on the older side, um, yeah. [1:15:07] people who haven't had traditional jobs before or were indie game hackers that joined the team. [1:15:15] Um, [1:15:17] So yeah, I mean, it's sort of a diversity of sources. If I am starting a company today, I would try to like find – [1:15:24] um, [1:15:24] Niche communities [1:15:26] This is where you'll find some really underrated talent, whether it's a niche group community around [1:15:32] a niche crypto project or a niche, um, [1:15:35] Program language. [1:15:37] um... [1:15:38] That's where I would look first. [1:15:39] You mentioned your wife, Haya, a couple of times. What's it been like building a company with your wife? [1:15:45] Um... [1:15:46] Depends on the day. [1:15:48] Ah. We... [1:15:51] We've worked together, you know, [1:15:54] We met at Newark back in Jordan. [1:15:57] We were working at this company. We were recruited... [1:16:00] actually a foreigner... [1:16:03] came to Jordan as part of a job from Belgium, and he decided that he loved Jordan a lot. He loved the desert, and he wanted to build a company there. So I was the first employee there, and Hyah was the second. She was a designer. I was an engineer. [1:16:16] And he pitched us on this idea. We're going to do consulting [1:16:19] And then we're going to build a product on the side and then we're going to become a product company.

[1:16:24] By the way, every company that has this idea never becomes a product company. I'm sure you've seen some of them. [1:16:28] And so the company was very dysfunctional. He was non-president. And so it was like the inmates were running asylum. [1:16:36] So I and I ended up working on a lot of projects together just for fun. [1:16:42] And then we started dating, which dating in Jordan is not... [1:16:47] Real dating, like God, for a cup of coffee, she had like a seven... [1:16:50] 7 p.m. curfew [1:16:52] And so... [1:16:55] And then by the time I started working on the open source replet, she actually contributed the... [1:17:03] the logo and a few other designs. She helped with a few other things. [1:17:07] And then when I got the... [1:17:10] physio.com the us um we we got married and we're very young i was like 24 at the time [1:17:17] And then she joined me in the US. We continued working on projects the whole time. We've worked on art projects, worked on games, [1:17:25] all these other things. [1:17:26] Um, [1:17:27] And then when the time came to [1:17:30] to start the company [1:17:31] I was kind of like looking for co-founders, [1:17:34] Because I'm like, oh yeah, YC says that you need to have co-founders and [1:17:39] I guess it's best to have co-founders and [1:17:41] I was like going in these co-founder days and trying to meet people and all that. [1:17:45] And then she was like, "Well, I can start the company with you." [1:17:50] I was like, it's going to be really painful. [1:17:52] Like, it's really, really painful. Like, are you sure you want to go through that? And she's like, yeah, yeah, of course. Like, I saw you, you know, with Code Academy. I'm like, you know, I can do that.

[1:18:02] three or four months later, she's like, [1:18:04] I underestimated how hard this thing is, but she obviously lived up to the challenge. And I think overall... [1:18:13] it, um, [1:18:14] it strengthened our relationship because when I was at Code Academy and I was, you know, working, you know, [1:18:22] 12 hour, 14 hour days, she doesn't really relate. Like what kind of job [1:18:29] requires that kind of commitment. [1:18:30] and then with rap lead. [1:18:32] we were both doing that. So it's not like, you know, your, your partner's going off in the morning, not returning until midnight on the weekend. They're, [1:18:40] they're wasted and they can't really do anything fun. And so you don't really have this great relationship with them. If you have a startup founder and you're kind of, you have a regular job, but if you're both founders, [1:18:52] You're actually going through this together. [1:18:54] And I think that ends up being good. And at the time, it was actually working against us because venture capitalists do not like [1:19:04] husband, wife, founders. There wasn't a lot of examples at the time. Obviously, the most [1:19:11] popular things example is, um, [1:19:14] Pogram and Jessica. [1:19:16] Uh, but then now there's like, they get Canva founders that a bunch of other companies that, that made it work, but there are challenges too. Like, um, [1:19:25] Right now, we're fully sort of [1:19:27] fully committed to the company, but also we have kids. [1:19:30] And so how do you [1:19:31] How do you manage this? That adds quite a bit of stress and pressure on us.

[1:19:37] Um... [1:19:38] And then the other thing is when something is going wrong, [1:19:42] in the company. [1:19:43] Like a lot of people can go home and can disconnect. [1:19:46] because they find comfort in talking to their spouse about other things. [1:19:51] But when Hai and I go home, we're talking about the problems and the problems kind of percolate and they end up seeming bigger and worse. And so you start to need rules around when do you actually talk about work and you're constantly breaking those rules, obviously, because you're thinking about something. So I think it is a challenge and I think it is at the same time, there's something very important. [1:20:15] uh, [1:20:16] very unique and good about it. Yeah. Well, it's impressive how well you made it work and you guys have a really great dynamic. Thank you. All right. Lightning round. We have some lightning round questions. Are scaling laws going to hold? [1:20:29] Yes, it is. [1:20:31] scaling different things. We're just going to keep finding things to scale. [1:20:35] What is the best piece of advice you've gotten? [1:20:39] Tollgram asked me this question. [1:20:42] I was stressed at the time and he told me, [1:20:45] is this your life's work? Like, are you going to be working on this 10, 20 years from now? [1:20:50] And I said, yes. And he said, well, why are you worrying too much about the daily... [1:20:55] tribulations. If they just settle into the fact that even if [1:21:01] Things are not great today. You're going to be able to fix them and you're going to be able to... [1:21:05] Turn things around. [1:21:06] as long as you're not dying.

[1:21:09] That's why Paul always talks about not dying and surviving as a company. [1:21:14] What is your favorite new AI app? [1:21:16] I'm kind of a Luddite because I make a lot of things myself and I use replet to make them. [1:21:22] I use the basics like, you know, chachup tea, perplexity, but then I spin up [1:21:28] I spent a... [1:21:28] pieces of software every day [1:21:31] uh, to, to use, I guess, uh, Manus is an interesting demo. Hmm. [1:21:37] Yeah. Um, [1:21:38] It's actually pushing on this idea that we talked about of how long can models work while saying coherence. I think they showed that they can go for an hour. [1:21:49] with some coherence. [1:21:51] Since you brought up Replit apps, what's a cool Replit app? [1:21:54] that you've seen recently? Yeah. [1:21:55] Uh, there are a lot of business things that, that are, are, you know, are, are fascinating. Um, yeah. [1:22:02] One cool thing that I went to New York, I was meeting a few investors and customers, and I met with Sears Home Services. [1:22:12] That's a century-old company. I didn't know it still existed, but apparently the home services department still existed. I'm with this very cool, trendy team. They kind of easily work at a startup. They're working at Sears. And they told me that six months ago, they finished the Cobalt Migration. [1:22:29] And it took them six months. [1:22:31] And then I start asking him, like... [1:22:34] what kind of other tools do you use? Do you use any SaaS software? They don't have any ERP software, any of these modern SaaS software.

[1:22:43] and they leaf frogs. [1:22:45] an entire generation of [1:22:47] of software that, [1:22:50] to start using Grapplet to create agents to manage their business. That's really cool. [1:22:57] Yeah, so they have these field workers that every morning they wake up and they're like, oh, how do I... [1:23:03] optimize my routes to kind of earn the most and service the most customers. So they built like an AI tool that actually gives them the most optimal route that they use every day. [1:23:13] For example, yeah. By the way, that team is non-technical. They're all operations people. [1:23:18] It's insane that you go from Cobalt to Replit and you skip everything in between. I mean, it does... [1:23:23] We were talking about this at the beginning of the conversation, but it gives you a sense of the sectors of the economy that are going to get transformed by this technology. Favorite book? [1:23:30] I am a strange loop. Douglas Hofstadter. Yeah. I actually disagree with the conclusion slash premise perhaps of the book, but it's still like one of the best books exploring concepts like AGI concepts like [1:23:45] consciousness and the soul and what it means to be a human versus what it means to be a machine intelligence. [1:23:51] Um... [1:23:52] What is the foundation model that you're, you like most or most impressed by? Uh, the, I mean, uh, Claude is the, uh, is the best. I mean, 3.7 is, is, [1:24:02] really the best at doing agent stuff. [1:24:04] person who's influenced your life. [1:24:06] I think my mom had this... [1:24:09] uh, [1:24:10] supernatural belief in me. Like, she...

[1:24:13] she would... [1:24:15] talk about all the great things that I was going to do even when I was a very small kid. [1:24:21] She gave me this, again, [1:24:24] Uh, [1:24:25] unwarranted at times confidence in myself. [1:24:29] Recommended reading on AI. [1:24:31] Um... [1:24:33] Twitter. It's really awesome. I think it got degraded recently, but there's so many papers that you find on Twitter or so many snippets of information. Just like... [1:24:46] Like, go read those papers, skim them, and I think you'll learn about what's coming down the line by just... [1:24:51] reading the literature. [1:24:53] Last question. Who is the most underrated person in AI? [1:24:57] I think... [1:24:58] Michaela Gattasa, the head of AI, now president at Replit, [1:25:02] um, [1:25:03] I don't think he's very public, but he was... [1:25:07] one of the early pioneers of LMS for code. And he's a great leader as well and visionary around where AI for code is headed. [1:25:21] Super impressive guy. I love every conversation I have with him. [1:25:25] All right. Well, thank you for coming on the podcast. I really appreciate our friendship. Thank you for doing this. Thank you. I mean, I appreciate your belief in us. And I think you saw something special a real long time ago. And like all these things that you noticed about what's special about Replit. And I really appreciate that. Thank you. We're still 4% there. 40 million users. We got 960 million to go. Yeah. Exciting times ahead. Thanks, John. Thank you.

[1:25:47] Thank you. [1:26:11] Thank you.