Of Mice and Models
What happens when anyone can use AI to build software?
Replit co-founder and CEO Amjad Masad joins Plaid’s Zach Perret.
Amjad Masad
I think we're approaching a point where LLMs are more reliable than humans. I know that's probably controversial. It'll go viral and people attack me on Hacker News. That's fine. I'm used to it.
Narrator
Amjad Masad learned to code by watching his dad use the family computer. Now he's building a company on the idea that AI means most people won't have to. Welcome to Intelligent Finance, a Plaid podcast with the people building the financial products you use every day. Plaid CEO Zach Perret sits down with Replit founder and CEO Amjad Masad to talk about how far AI can go in building software and what financial services should build next. Here's Zach and Amjad.
“I think we’re approaching a point where LLMs are more reliable than humans.”
Zach Perret
All right, Amjad, thank you for being here with us. Welcome to Intelligent Finance.
Amjad Masad
My pleasure. Thank you for having me.
Zach Perret
I'd like to start at the beginning. Like, where did you grow up? What things from your childhood really influenced you as an adult?
Amjad Masad
I grew up in Jordan. And I got into computers really early on, which was odd and lucky because at the time when I got my first computer, it must have been 1993, I was 6 years old. And it was like probably the first computer in our neighborhood, if not one of the few in the city. And my father, he was like a government engineer. He wasn't like very well off financially, so he had like to take a loan to kind of buy that computer. But he had this intuition that this thing matters for the future. You know, it's like one of my earliest memories was watching him setting it up and like, you know, pulling up this, like plopping this big manual on the desk and like opening it. And like, you had to like learn to program in order to use a computer at the time. And I was like, oh, that's really cool. And, you know, like I would watch him try to use it and I would like learn by just like watching him. And then like pretty quickly, by the time I was like 7, I could like do everything on the computer and he couldn't do much of it.
Zach Perret
So it became your computer?
Amjad Masad
It became my computer. And started programming. And the first thing I built was for my younger brother to learn math. It was kind of like simple algebra. Like there's like 1 plus blank equals 2. What's in the blank? He puts 1, he gets a clap. He puts something else, he gets a boo. And it worked out well 'cause my brother works for us now at Replit as a programmer.
Zach Perret
Nice.
Amjad Masad
So I was always kind of building things for other people to kind of learn and tools. I was like really interested in like, building tools for people. It was so clear to me from the moment I was a toddler that I was just gonna do something in computers. Nice. Yeah.
Zach Perret
Cool. So fast forward a little bit, you get to college, you're studying computer science. Talk me through that story and kind of how you ended up working in technology companies. It seems like a natural extension of your childhood.
Amjad Masad
Yeah, actually, funny story. I got into reading about AI back then, I think even early 2000s. I feel like there was an AI hype. There's a lot of predictions that computers are getting powerful really quickly. Year over year, you could feel the processing power is getting better. And so there's tremendous excitement about the potential for thinking machines. And philosophically, I felt like programming is going to get automated. And I felt like, okay, I learned programming when I was such a young kid. I don't think that this is a lasting skill. I feel like AI should be able to do it. And so I went into computer engineering actually, because the thinking was if AI could program, someone still needs to build the machines for it, which actually might end up being true because AI is really good at software, really sucks at hardware so far. So far. Yeah. But pretty shortly after that, I rediscovered kind of the beauty of programming. I started reading Paul Graham's essays.
Zach Perret
This is what year?
Amjad Masad
This was 2007, 2008, something like that. And so I got really into programming languages. I quickly hit a problem, which is every time I want to kind of play around with a programming language or even like do some homework or do some coding, I would have to download gigabytes worth of software, spend a lot of time kind of setting up the development environment. And then I'd have to do that again when I go home or when I go to school, when I go to work. I had different computer at different places. I didn't have a laptop at the time, couldn't afford it. And so I was like, surely there's like a website I can like— code on? And there wasn't. And so I started looking into it. Well, pretty quickly I found out, well, I had a prototype pretty quickly. I just like put a JavaScript interpreter in the browser and like just evaluating that. But that in itself like gave me an idea that, wow, this is really cool. I had like a Nokia Symbian phone and then suddenly I can like write code, write some JavaScript on it and like run it and see the results. And just like having like a REPL in the browser kind of felt like really powerful. I had finished college, started working at Yahoo for a little bit, Yahoo in Jordan.
Zach Perret
So 2011, working at Yahoo, you had this kind of crazy idea, but did you think that that was gonna become a company at that point, or that was just a fun project you had on the side?
Amjad Masad
I thought of it as infrastructure. I thought of it as infrastructure for the web, but I just thought it needed to exist as just infrastructure because code is gonna be everywhere. People are gonna be writing a lot of code. A lot of people are gonna be learning how to code, and surely you needed just super accessible development environments. So the natural thing for me was to open source it. And then the thing that happened is a bunch of learn-to-code companies in the US started building on top of the open source project that I built, most famously Codecademy, but even at the time Udacity and others were kind of building certain products and experiences around it.
Zach Perret
And what was it called at that point?
Amjad Masad
So the infrastructure project was called js-repl, which is like how to get a REPL in any language in JavaScript. But I already they had the domain Replit, and there was like a demo page for JSReplit. I was like, oh, this is like the demo kind of, but they were using the open source project. And so I ended up joining Codecademy. The founders came out of Jordan to kind of recruit me, and I joined them. They were just out of YC, so it felt like I'm like part of a founding of a company. And I came to the States.
Zach Perret
Wait, so they had built on top of your open source library? This is when you were just doing the open source?
Amjad Masad
Yeah, yeah, yeah. So they flew out, It was like, at least let's come meet you, whatever.
Zach Perret
I love that they did that.
Amjad Masad
Yeah, yeah, Zach Sims did that. And we just hang out and we just like really liked each other. And I showed him, we went to the Dead Sea, we did a bunch of stuff. And like every day he would give me like a bigger offer. I actually signed the offer at the airport because I just kept saying no. And then whenever we got to the airport, it was like he gave me an offer that I felt like I couldn't say no to. And so I signed that. And so then they got me an O-1 visa and I came to New York.
Zach Perret
And you moved to New York at that point?
Zach Perret
And so you had a couple stops. You were at Codecademy, you were at Meta for a little while.
Amjad Masad
It was called Facebook.
Zach Perret
It was called Facebook back then. Yes, that's fair. And yet this dream for Replit or following up on the open source project that you've been doing kind of stayed throughout. Why not go pursue that as a company at any point along the way?
Amjad Masad
Well, I think as an immigrant, there are a few difficult things. One of them is immigration. So I think I just didn't have a green card. That's one. It wasn't like there are ways around that. But the other thing, I was really broke and I didn't know because at Codecademy, I didn't care about the salary. So I was getting paid $70K. I was living in a studio with someone else, but actually my friend from school who happened to be in New York as well. I didn't care about money at all. I was like, I just need enough to eat and mattress to sleep on. But when I left Codecademy, I just didn't have any money at all. So actually, when I joined Facebook, they told me you have to fly out to Menlo Park for training, for boot camp. I didn't have money for the ticket to fly out there. We had to rent our apartment on Airbnb, get the money, and then fly out. Fly out too. But yeah, it wasn't clear that I had any sort of stability to be able to start a company. Not enough runway to even incorporate it. And I was a little burnt out as well. I think you know this. I think it's just torturous to start a company. It's the worst thing you could do. I mean, obviously there's a huge upside, but as you're running the company, you're never enjoying the upside. So I was right. It's hard. But the dream kind of lived on, like you said. But I was really excited by Zuck's vision for internet.org. The idea of connecting a lot more people around the world, this thing that I keep coming back to, which is we have an opportunity as civilization, as humanity, to truly bring about equal opportunity. There's always this dream since the Enlightenment and so on that anyone can participate in capitalism and wealth creation, and everyone has a sort of an equal shot at that. Now, the problem has always been that it required capital and resources. I think the internet and computers and now AI brings that way down. Like, we're going to zero, right? Now, a kid anywhere could build something great and be able to make money online very, very quickly. So I was always inspired by that vision. And obviously, a precondition to that is connecting everyone. That's partly why I joined there, but I ended up working on open source. I just find myself building tools all the time. So I ended up working on React.js and tools like that, just like building JavaScript infrastructure to make web development better. 'Cause I often, I go try to build products, I'm like, man, this sucks. You guys live like this? I need to go fix all the tools. And yeah.
Zach Perret
Well, as a user, thank you. Yeah. Thank you for doing that. Thank you for having that mindset. But then eventually you left and decided to get over the hump and launch the company that you'd been, it sounds like, dreaming about since college.
Amjad Masad
Yes.
Zach Perret
How did that happen? What finally got you there? You started YC around that time, right?
Amjad Masad
No, YC came after. We actually were rejected from YC before I quit Facebook. We applied to YC a couple of times, we got rejected. And I think my role at Facebook sort of came to a natural conclusion. We kind of built a lot of the tools. Like we built React, we built React Native, we built like the testing tools, we built like GraphQL and all of that. Like JavaScript stack was complete like back in 2016, but like still to this day, everyone uses the same stack that we built back then. So I decided to leave. And before I decided to leave, like we, my wife and I actually like picked up kind of Replit, that demo site we built like a on the open source project a while back. And she decided to kind of give it a revamp and design as a side project. We were always working on side projects because she's a designer. And I decided, Docker had just come out, and I decided to rebuild everything using Docker because doing virtual machines in the browser is really bad. It still doesn't work very well, 10 years later, but at the time, definitely didn't work, although I was one of the first people to do that. So I decided to kind of build a, like Go had just come out, like there's a lot of really cool cloud infrastructure. So it was like a really cool project to kind of go build that. So we did that and suddenly Replit started growing. It like always had like 5,000, 10,000 users, went to 100,000 users within a month of kind of like making it better.
Zach Perret
What was the product at that point?
Amjad Masad
It was literally like you go to a website and there's like a bunch of programming languages. You can start coding in any programming language, you can run the program, you can save it, you can share it. We had just added user accounts. So you have a user account, you can save all your programs. And the main use case was still people learning to code and hobbyists. And there's some scientists running certain scripts or math things.
Zach Perret
And there was a ceiling on the complexity that you could have?
Amjad Masad
Yeah, at that point we had just added files.
Zach Perret
So pretty hard caps.
Amjad Masad
Yeah, pretty hard caps. But it was clear that the kind of the theorem that had emerged in my head is that there is an exponential relationship between how easy you make programming and how many people would want to do it. And that has borne out to be true, especially now with AI. It's like, just remove code entirely, you get an explosion of people that want to do it. And so I was like, okay, this is something you can totally start a business on, even if you don't know what the business model is. There's a clear sort of generate a function for continuously just making programming more accessible. Because I thought surely we want to fix the development environment, then we want to move to production environments, then we want to do testing, then we want to do CI, then we want to just take the entire software lifecycle, just put it in one place and make everything one click and you should be able to solve software. At the time, the idea of automating code was still kind of very early.
Zach Perret
Talk to me about kind of the linear progression here. So this was what year that you kind of went full-time at building it?
Amjad Masad
2016.
Zach Perret
Up until maybe the first 5 or 6 years of the company, it sounds like it was largely this style of product. So kind of build code on the internet, there's a ceiling on the complexity, you're writing the code, it's not automating the code. When did that really start to shift for you all?
Amjad Masad
I mean, we've always wanted to do the whole thing, but it was so hard. Another sort of flaw slash, I think, potentially talent is like I underestimate the difficulty of everything. So you end up kind of approaching a lot of really hard problems with optimism.
Zach Perret
So at the beginning you were saying we need to automate it all. You just couldn't do it
Amjad Masad
yet. Yeah. It felt like I could imagine the whole thing, but imagining the whole thing and actually doing it. For example, the thing we ran into when we hit scale, you see all this stuff that's happening now with the OpenAI, the Hugging Face attack, and what's going on. We actually faced a version of that early on where it wasn't agents, but it was script kiddies. It was kids using Replit and then like, oh, this is cool, let me try to break it. Or like, this is cool, let me mine crypto on it. And the amount of abuse, you have a sandbox environment and people are trying to escape it, people are trying to hack it, people are trying to commandeer it to hack other services, sort of like what happened with Hugging Face. A lot of work went into anti-abuse security, and we built a lot of things over the years that ended up being quite advantageous now, because when you drop in the agent, you would want to control it and make sure it doesn't do crazy things. But so there was a lot of distractions along the way or problems that we had to solve before we move on from the development environment to the production environment. So I think the time we had a real production environment was really '22, '23. Mm-hmm. We had built a really cool deployment experience. I still think it's still the best where you write a little bit of code and deploy in the same way you run and it just creates an image and deploys it, and it scales down to zero, scales to millions of users, and a database alongside with it, and everything is serverless. And it's like a really cool infrastructure. It was kind of like the first of its kind. And then we started seeing people build businesses for the first time. So it took—
Zach Perret
What was the first one that you saw? First real business?
Amjad Masad
First real business was Magic School.
Zach Perret
Okay, what's that?
Amjad Masad
Magic School is now, I think, worth half a billion dollars. It's like a teacher in Colorado. And during COVID he kind of left and he was kind of searching for meaning and trying to figure out— he wasn't happy with the way education was happening. One of his big things was in the United States, we really burn out our teachers. There's so much we pay them so little, give them so much work. And there was a lot of teachers that would just quit because of that. And he's like, this AI thing that was like '22, we had GPT-3, this new AI thing, maybe I can build some tools with it. So he started using Replit. At the time, you still had to code. He learned to code. He started using Replit to build the initial idea, the MVP. And he quickly got adoption because he knew how schools worked. And so he built a bunch of generators for assignments and ways to do grading and things like that. And he got to a million ARR fairly quickly. It was one of the first times where I was like, oh, this AI thing, you can scale revenue pretty quickly, even at the hardest. I tried to sell to schools. It's the hardest thing you could do. And he was successful because he's a domain expert. So I was like, okay, well, if you give everyone the ability to make software, I mean, those domain experts, they're They're everywhere. They're not in Silicon Valley, but they're going to be building amazing businesses. But obviously it takes like a really special founder and he is. And yeah, he ended up scaling very quickly and now it's like a, I think they're on their way to $100 million ARR.
Zach Perret
Wow.
Amjad Masad
Yeah.
Zach Perret
So you had a, what I think is a very interesting business with very good growth for many years. And then at some point, I don't know, '23, '24, the growth just inflected and went vertical. And that's because you changed the product. What was the product change and how did you know that was the right thing?
Amjad Masad
Actually getting people all the way to deploying a real thing and building a business on it, that was always a dream. And we felt like we were falling short of it. People like Adeel who built Magic School, he has enormous amount of grit to have gone to where he's gone.
Zach Perret
It was a power user product.
Amjad Masad
Power user. So can we make it so that You don't need to be that big of a power user in order to do it. I just thought, like, you know, we've solved a lot of the things. What's left is coding. We had like this 100 Days of Code, like Python environment where like Python classes where I was like, hey, do you wanna use our product? You don't know how to code? Okay, here's 100 days of learning.
Zach Perret
You can become a customer, but first 100 days, 100 days of work. Yeah. And then you can pay us some money.
Amjad Masad
So that was obviously stupid. And so I was like, okay, how do you solve that? And again, this kind of optimism, I was like, well, I think AI agents, at the time there wasn't many examples of it. There was one example, AutoGPT, if you remember that one, which was really cool. And you put a GPT in a loop and it does a bunch of tool calls. I was like, okay, this hints that this is probably coming if we start building it now. Our product gets ready and technology gets there. At some point they meet and it works. And that is exactly what happened. We started building it late '23, early '24. It took us about 7 months. Our initial harness, they weren't called harnesses back then, we would just call them software. But our initial AI harness is 10x the complexity of what it is today. Over time, the harnesses are just getting thinner. Yep. Now you can have 100 lines of code of harness essentially. But it took us 7, 8 months of research to figure out how to build good harnesses. And the product wasn't all that good at the end because the models were really bad. The models barely would— we had this problem of laziness. Models would actually— you would ask them to write a big file of code. In the middle, they will say, to-do, do this later. Yeah. It's like, what? And so we had to write— the harness had to detect that and recall the LLM and do all of that. But when we launched in September '24, most people in the world have never experienced an AI agent. And I think the thing that also made it really special is that it would set up the database for you. It would set up the deployment for you. Completely. People have never seen that at the time. And so we were stalling at $2 million ARR. We released it in a few days, we were at $10 million. in a few days. And I was like, wow, this is like explosion of potential product market fit. But you sort of like, there's still a lot of work to be done. And it was like really, I think I came to one of those like Strategy Capital meetings when like early on when this was exploding and I was like in the middle of it. But yeah, that's basically what happened.
Zach Perret
So we actually met in the room of one of these Strategy Capital meetings. For those that don't know what Strategy Capital is, it's a hedge fund, but it's run by a guy named Hamilton Helmer who wrote, in my opinion, I mean, one of the most interesting business books called 7 Powers really focused on competition. I do remember we were sitting in one of these meetings that they do every now and then and we got to chat and I heard you talking about it. I'm not sure if it was that meeting or maybe a different one, but I remember you guys were on a crazy steep growth trajectory, almost unexpectedly so.
Amjad Masad
Yes.
Zach Perret
But one of the things that really stands out to me though is you were able to go for 8 years and get to $2 million of ARR and keep going. And then you had this huge inflection. How did you deal with that as a leader, both the grit to stay focused in the early days and then the speed of scale?
Amjad Masad
I think the hardest part was to keep people motivated. And a lot of it was the mission, the vision, the kind of— we've always had stories, stories like Ideal. We've had stories of kids building, a COVID vaccine delivery program in India having a huge impact and that going viral. So we always were producing good in the world. It's like a nonprofit almost. The feeling was like we're burning VC money, not a lot by the way. We've always been fairly capital efficient, but we're doing good things in the world. And so there was a fuzzy good feeling about working at Replit. Mm-hmm. Is that we're giving people tools to learn how to code and to create things and to have impact on themselves and their community and potentially build businesses or build skills. And so that was one I think kept people motivated. The other was there was a constant user growth. There was some revenue growth. It wasn't crazy, but we'll see signs of things to come, right? And it's important to sort of tell yourself that story and highlight it to the world. Like the first time we got a company to use Replit, it was like huge and like making a big deal out of it and talking about what's the potential for the enterprise business. And so there was always momentum and also our team was always good at like being early to things. So we trained the first open source coding model in 2022. Yeah. It was a state-of-the-art open source model. It outperformed Google's like PaLM Code at the time. Also, people working at Replit were always at the cutting edge still. is at the cutting edge. And so we attracted a lot of really good engineers. For me personally, I sort of come from nothing and all of that is great. The alternatives, so I already won, so the rest is gravy.
Zach Perret
So let's fast forward to today. What is the product today? How would you describe it to the outside world?
Amjad Masad
Yeah, so the product is like The user experience of it is you go in there and you can basically build anything by describing it using natural language. And I think that there's a lot of different products that do the same thing right now. What differentiates Replit is we can take you the distance. You can rely and trust Replit to create secure software. You can scale with Replit. We have a lot of different apps that are generating tens of millions of dollars, a few that are over 100. We have apps that have millions of users. So you could trust Replit to run real infrastructure unlike many other vibe coding platforms. And it's like a full-service product.
Zach Perret
When you first started taking off, you were a consumer company, right? Or a prosumer company maybe. It was people that wanted to build their own personal app.
Amjad Masad
I would say pure consumer. Because prosumer implies that people paid for it.
Zach Perret
But you do charge.
Amjad Masad
We do charge, but it's like 0.1% of the population paid for it, or my users. Yeah.
Zach Perret
But then you made this decision to shift to enterprise. Why not just stick with consumers, let consumers build tons of stuff for themselves?
Amjad Masad
Well, there are a few things. Over time, I've realized that the real value we're providing is creating economic value, creating commercial value, right? Like the story of that teacher building a company, whether it's SMBs or enterprises, like making it so that, and by the way, it's like entrepreneurship at the heart of it, both in enterprise, in consumer and prosumer. Because like, for example, the other day someone from the government visited Replit. They naturally from the— they work in the federal government and they organically adopted Replit and they took a workflow that was requiring of like a 5-person team and they automated it. And initially they were worried about their job, but then they shared it with their superiors. And they got super excited about it and they became this internal innovation team. They were originally like a comms team that were scouring the web for information, writing memos. They automated all of that with AI agents and they became an internal development team. And now they're saving the government tens of millions of dollars because a lot of things you can build with Replit in $20 typically goes to an RFP and a contractor, government contractor charges you $20 million. And those guys are getting promoted. And so whether it's the individual or inside a company or starting their own company, ultimately it's about this idea of grit, empowerment. And I think the important thing about the moment we're in in AI is that AI properly understood is a huge commodifying force. So it's like an avalanche that's commodifying everything in its way. By the way, including a lot of what the big labs do, it's commodifying. At some point, the big labs were charging you for fine-tuning. Now you could use Astra and tell it, hey, grab a Quinn and fine-tune it to do this. So it's commodifying everyone's businesses in interesting ways. And I think the basic vibe coding is, I think it's fully commoditized now. You can ask Muse to make you a simple website. We get back to the Hamilton strategy, things like that. You have to see at least a year ahead and understand where the value is going to be and try to build something a little more valuable. Even if you don't have long-term moat or defensibility, I think one thing you could do is always be a year ahead.
Zach Perret
As you think about deploying this infrastructure into the enterprise, one of the things that I'm sure people are worried about is how reliable is it? Is it going to do the right thing? Is it going to hallucinate and create something that's either dangerous or just flat out incorrect. How do you get enterprises comfortable with it?
Amjad Masad
I think we're approaching a point where LLMs are more reliable than humans. I've made a prediction in 2025 that I said, I don't think you need to learn to code anymore because you're not going to be reading or writing code. People went crazy about it. But I think that we're approaching— so I'll give you an example. Rust, great language. People don't like using it. We wrote a lot of our application software in TypeScript. Some system software we wrote in Rust, because mistakes are really expensive there, like a null pointer exception, you could lose customer data. Rust makes it impossible to make memory issues. It's a memory-safe language. So the other day I was talking to my brother actually, and he's building a Slack bot at Replit. I expected him to be writing in Python or TypeScript. Obviously, it's just an app, and he's writing in Rust. And I asked him, how come you're writing in Rust? He's like, well, I don't have to write Rust myself. I barely have to read it. I review the code, but I don't read it all that much. And so why not just get the memory safety from the LLMs writing this low-level language? Because I'm not interacting with the code directly. And so Obviously that app is going to be more reliable than if we had created it like, you know, 4 years ago using TypeScript or Python. There's a clear sense in which you can create more reliable software right now because they are so— they don't get bored. They're like very resourceful. You can also run continuous scanning. Replit right now, as you're writing code, we have a code reviewer. So as one agent is writing code, there's an adversarial agent that's reviewing that code. That's one. Before you publish, we run a full static analysis, but also a dynamic analysis using LLMs. And we published a paper showing that once you do these 2 things together, you get much better results than just pure static analysis. And then once you deploy it, we actually have a black box testing agent that's doing pen testing on your app. Mm-hmm. And so because of all these different automations, I think you can get to a point where the code that's being generated is safer than what the human does. No, I just don't wanna make it seem that it's all rosy and mistakes never happen. I think where things are still difficult and the hallucination problem still happens is when requirements are underspecified, or the agent doesn't have enough understanding of your systems. It's on us at Replit to kind of first make sure the agents have enough context about your company. We could do a much better job at that, but also figure out ways to continuously kind of improve the user experience of having the agent come back and ask you for details and ask you to clarify. And maybe there is some way to do planning that is not just using natural language. Internally, some of our engineers are doing formal specifications. You can't get a product manager or designer to go do that, so I'm not sure we're going to do it on Replit. But I think there's a lot of white space. And by the way, this is potential for a lot of startups to emerge to figure out how can people interact with LLMs to get better at specifying the thing they want to build. without actually writing code.
Zach Perret
Whose problem is it if it's wrong? We work in financial services, we work with a lot of the banks. Let's say somewhere the bank tries to create an accounting tool and the math is off. Ultimately, that's on whoever's using the product because you're delivering a bad user experience or whoever created the product. It sounds like your view is we're not that far away from that just not happening, but in the meantime—
Amjad Masad
Or happening way less than now. It's sort of self-driving. Self-driving is already more reliable than humans, but we're actually judge it a lot harshly, and there's the march of nines that needs to happen, right?
Zach Perret
Sure. But ultimately, it's the person using the product that is ultimately responsible for the thing to deliver. And you're just trying to help make it easier for them.
Amjad Masad
Yeah. Which is why Replit is still somewhat of an IDE. There's all these chatbots now that can make software for you, but Replit actually gives you access to everything. You can go look at the Git files, you can look at the files themselves, you can go look at the components. It gives you all the tools. You don't have to learn to use all of them at once. But I do think there's a human driver, there's a driver's seat. We're not at a point where software is fully delivered autonomously, from specification all the way to delivery. Not yet. Not yet. Those questions need to be answered before we get there.
Zach Perret
What's the role of a software engineer then? If they're not writing the code, they're not having to do all that much of the systems design. You still have engineers that work at your company. Yeah.
Amjad Masad
They still do a lot of the systems design. So it used to be that we're typing every letter of the code, and then we had GitHub Copilot, there's an initial version, and it's kind of completing the lines for you. That is the inner loop of software development. And then you expanded the loop a little bit. And now, like, I think Cursor Composer was the first one that generated multiple files or like entire files, right? That's like a bigger loop. Replit Agent was the first to kind of do everything from like debugging to like setting up databases and infrastructure. That's a bigger loop. And now we're at a point where like most agents, you don't have to interact with the code at all. You can put a prompt, it'll configure everything, grab the infrastructure, get all the way to like a PR or like a deployed product. There's still other loops and it is the figuring out what needs to be built, figuring out what is the best architecture, figuring out the different trade-offs of the different systems you want to use. Like what kind of database do I want to use here? Whether the availability considerations or speed considerations and so on. I'm not saying that LLMs will not be able to do that. But right now, humans are still better at that. And it might be just because of the context. We haven't really fed the right context for LLMs. Will engineering get entirely automated? There's a chance. Yeah. I mean, what do you think?
Zach Perret
I think a lot of it will. I think the figuring out what to build, figuring out kind of the key components of how to build it.
Amjad Masad
Well, then do you need engineers for that? I mean, you need—
Zach Perret
Not necessarily.
Amjad Masad
Yeah, product people.
Zach Perret
Engineering thinking, certainly.
Amjad Masad
Yeah, engineering thinking.
Zach Perret
but not necessarily someone that's been writing software for the last 20 years.
Amjad Masad
How do you feel about that?
Zach Perret
As someone who's written a lot of software over the last 20 years, it's scary in one sense. The craft of actually writing software is fun. It's like, I've never been a woodworker, but I can imagine the craft of woodworking is fun, even though a machine can probably produce a table better than I could produce a table. And it doesn't mean that never write code again. I do it for fun. However, I think it feels incredibly empowering because the speed with which I can build things, the quality of what I can build, even just my personal weekend projects are so much more elaborate and interesting now. So I think it's great.
Amjad Masad
Yeah. That's a good mindset. And that's generally my mindset. I feel super empowered. I did have a feeling maybe a year ago of deep sadness maybe like depression for a little bit, where it was like, wow, like this thing. Do you know, like in The Matrix, that guy who is like the, basically the guy who betrays Neo? He's like, you know, I want to forget. I want you to like put me in a new character in The Matrix. I want to forget everything. It's like the amount of information in my head about like, automatic semicolon insertion in JavaScript. It's like, I have so much information and knowledge that's totally unnecessary right now. Maybe trivia, I could win some trivia programs, but that's one. But the other one is just like, yeah, I mean, it's like the thing you've dedicated a lot of your life to become a great craftsman at, it's just totally gone. Yeah. And so there was a moment that I had to get through to And I think a lot of engineers have not really stared it in the face and embraced it. And I think there's a crisis of meaning that is coming pretty soon.
Zach Perret
So within an enterprise, I was giving a talk to a fireside chat with one of our customers yesterday, actually. It was a large financial services company and their question was, how should the average person get started with AI? And it's a big company. Not everyone has been using AI as much as a startup in Silicon Valley. And my answer was just go build something. We give some examples of ways they could go build something. Could be for work, could be for personal. If you were to talk to someone that is really not technical, they don't fancy themselves an engineer, or they don't even fancy themselves a builder, but they have access to Replit at work, what would you tell them to go do?
Amjad Masad
I've actually done this where I would ask them about their day and Like, you know, walk me through what do you do at work? Like, let's say, oh, I'm a deal desk person. You know, I, you know, I work with salespeople to create quotes for the customers. And, you know, a salesperson comes to me with like, you know, these requirements and we kind of craft like a quote for the customers. Like, okay, like, can this part be automated? Like, can they, instead of coming and talking to you, can they submit a form online or can they talk to an agent? And I was like, yeah, I think that could be done. And And then, okay, can you generate a PDF or is there a standard template that you have? It's like, yeah, we have typically the first quote we send is pretty standard, so on. And then maybe people negotiate and maybe that gets into, oh, let's start the initial one. Can you sort of automate that? Can you imagine an app? And it's just like, yeah, I can imagine an app. And they'll start talking. And as they're talking, I was like, this is a prompt. What you just said is a prompt that you can put into Replit and it will create exactly what you would want. And in fact, a program like this is probably a one-shot. And surprisingly, I think the biggest thing is people can describe to another human what they want to build. Even if they're not engineers, they can talk in a fairly specified way. But when they sit in front of a computer, they feel like they have to talk differently. And I keep telling people, just talk to it like you're talking to a friend. And maybe at some point we should create a brainstorming sort of buddy where it's personalized and it's like, hey, tell me about your day. What do you do? And exactly what I've done with people. I think that would be really cool.
Zach Perret
It'd be great. I mean, you sit down in front of a blank page, it's hard to get started. You sit down and someone asks you some questions, it's much easier.
Amjad Masad
Yeah. So it's sort of like back in March, I was going to work and I saw one of our engineers that know pretty well. I was like, hey, how are you? And he didn't answer. And then he looked at me, he was like, have you heard of Nexus? And then the elevator opened and he just ran away. And I entered my company and suddenly it just feels differently, different. Everyone was sort of dazed and confused. And so we've rolled out a version of Replit internally that we call Nexus. That was before like the Grok bots of the world and all of that that happened externally. Like, which is basically this like master AI that had all the context about the business. You can ask it any question, it joins across Slack, GitHub, all our databases. It can generate any report, crunch any data, and it can generate pull requests from anywhere, from Slack, and anyone can create pull requests and not just engineers. And obviously it has all the existing Replit capabilities. Suddenly we had this empowering tool. That crisis meeting that I was just talking about, it was the first time engineers were like, oh, holy shit. A lot of people in the industry had that back in March, February, March. And our engineering output entered this steep growth period where in a matter of a couple months we kind of 3x engineering output. Customer support tickets were closing a lot faster.
Zach Perret
This all sounds great. Why was everyone dazed?
Amjad Masad
Well, everyone was worried about their job. I mean, that's the thing with AI. The moment you see the capability, the first thing you're thinking It's like, what happens to my job, right?
Zach Perret
But it turns out you probably hired more people.
Amjad Masad
We doubled. Yeah, we doubled the team since then. And so—
Zach Perret
And
Amjad Masad
Why? Why do they worry about their jobs?
Zach Perret
No, why do you hire so many more people if you're having so much more productivity?
Amjad Masad
Because we want to do more. Because we want to do more. We want to grow more. We want to like service our customers more. There's like endless things to do.
Zach Perret
We do the same thing. We become more efficient and we hire more.
Amjad Masad
Yes, exactly. So the idea of a self-driving company, it's like a very crude analogy is that you go from doing the individual tasks to putting in the address and saying where you need to go, but you need to figure out where we're going first. You need to figure out maybe figure out even how we're going to get there. But then you do this throughout the day, but the thing that is actually actually doing the work as the agent. And our company is almost entirely run that way today. And I feel like that's a transformation every company will go through.
Zach Perret
So obviously we work in financial services. There are a lot of people in financial services listening to this. What do you think should be built specifically in financial services or banking?
Amjad Masad
You might be surprised that I've like thought quite a bit about this because I got into Bitcoin pretty early on. This idea of programmable money is really powerful. One thing that I thought was missing from the internet, we're so grateful for the designers of the internet, they made something amazing, like this idea of TCP/IP and open protocols and this infrastructure that could withstand nuclear holocaust. So it's fantastic. But the one thing they missed is they didn't add a way to move value. Across that network. They
Zach Perret
intended to.
Amjad Masad
Did they? I didn't know that.
Zach Perret
Yeah, it was in the original protocol. There was some intent on actually doing a payment transaction.
Amjad Masad
Oh, okay. So when I saw Bitcoin, I was like, oh, that's like a clearly missing thing. And what kind of things could emerge, what kind of ways of creating businesses and transacting could emerge that is interesting? And I did some experimentation on Replit with that. So today, for example, you go to AWS and you kind of look at their list of all the different services and you study them one by one and maybe you ask ChatGPT, which one should I use? But there should be a world in which anyone could publish some infrastructure product, some database, some way to send SMSes, some way to— And then my agent goes and finds the best thing, tests different things, maybe even does pilots, and then negotiates, and then— Just
Zach Perret
buys it.
Amjad Masad
Buys it. And so creating software means that I'm freed from buying from one cloud provider, and I can buy from anyone. And there's a lot more innovation that comes out of that. And you could imagine a way of creating software that's totally different. Right now, open source is unmonetizable, and there's a problem with that because we've seen that with very under-resourced projects like OpenSSL, which was like half the internet was dependent on, but it was one guy's side project. And so why is critical infrastructure not monetizable? Can we get to a point where there's some automatic way? I think the reason it's not monetizable is because the transaction cost is really high. But if there's a way for agents to automatically on the fly kind of pay, I don't know, some amount of some transaction costs to use this open source package every time you install it, just pay a few cents, you could fund infrastructure projects. And so all this to say is the idea of programmable money is such a powerful tool, and it would change how we transact on the web. The idea of micropayments, I hate ads. I don't want to go to the New York Times and see ads. I'm willing to pay you. I don't want to subscribe to every magazine either. So, like, can I just, like, pay you for the things that I read? Like, I should have a wallet in my browser and, like, be able to, like, you know, send you money as I'm, like, reading stuff. Intelligent finance, programmable finance needs to exist. It's, like, way overdue.
Zach Perret
I agree. I agree. Wallets for agents and making it all simple. Okay. So at At Plaid, we read a lot of books, as you can tell from the background of the podcast, we read a lot of books. And one of the things we asked is a recommendation of a book that influenced you.
Amjad Masad
Yeah. So it's right here. Thank you for bringing it. That's really cool. Actually, when I came in, I was like, oh, they have this book. So I Am a Strange Loop by Douglas Hofstadter. A lot of programmers, a lot of people in tech might know Gödel, Escher, Bach. Have you read that one? Yeah. It's really fascinating, really weird mind that creates these books. But this is a lot more accessible, and it gets at the heart of what Hofstadter is all about, which is loops. Obviously, my company has loop in the name. An acronym. And his core idea is that the self is ultimately a loop. It's a radical idea. I'm not sure I agree with all of it, but it's the best explanation of consciousness and intelligence that I've seen and has a lot of really cool mathematics and poetry and all of that, kind of like GEB. But the core idea is that you are a strange loop. There's this strangeness about loops. The idea of self-referential thing is almost mystical thing, and the self sort of emerges out of that, and that has all sorts of implications about AI and consciousness and things like that. But the coolest thing about it is that it kind of merges, and that's the coolest thing about Hofstadter is he merges the hard science with the soul and with the humanity and with humanities. And so it's an amazing read.
Zach Perret
Awesome. Well, thank you so much for doing this and thank you for your partnership and excited to see all that comes next for Replit.
Amjad Masad
My pleasure. Thank you.
Narrator
That was Amjad Masad, Founder and CEO of Replit, with Plaid co-founder and CEO Zach Perret. New episodes of Intelligent Finance are coming soon. Follow wherever you get your podcasts or subscribe on YouTube. Thanks for listening.