Replit Founder and CEO Amjad Masad and PostHog CEO James Hawkins | Granola Firesides

Yeah.

Is this on by the way? Test.

I think we got the trends right. I think we got that the world just needs way more software, way more software engineers, way more people making software than we currently had. There were all these attempts at creating more software engineers, boot camps, all these things, but it wasn't really satisfying the insatiable need for more software. So that's one. I think we got the shift to the cloud. I'm kind of dating myself, but starting in 2016 it wasn't really clear that everything was going to be cloud-based, but cloud-based was going to be very important. And then AI. People forget, but in 2015-16 there was a language model hype period as well where people were starting to build language models that were performing reasonably well and they could code reasonably well. I've been writing compilers and doing programming language stuff for so many years, and I thought the way we write parsers, the way we handle programming was very manual, and I thought it's going to get automated. All these three things were in my deck in my Series C deck. So I think we got the trends right. But the timing was just going to take a lot longer than I expected for all these technologies to mature, including the cloud. Right now it's so easy to create a sandbox in the cloud. We have microVMs. We didn't have microVMs at the time. Containers barely worked at the time. So that had to get better. Web browsers had to get better. AI had to get massively better. The principles were right. The timing was a little early, but I think it gave us a lot of benefits, starting early.

No, I felt it even before that. I wrote the open source version of Replit back in 2010, which was just an in-browser REPL. It was the first way you could run a lot of different programming languages in the browser. I worked at Codecademy. I was the first engineer there in 2011. It was a massively viral product. The amount of interest in coding and the stories we hear every day with tools like Replit, where someone learned a little bit of how to make software and went on to create a million dollar business. We wouldn't hear the same thing. We would hear someone learned to code, made a fitness app, made $5,000, and it was amazing. And we'd hear it once a year. So you could tell there's a lot more entrepreneurship to be created. A lawyer learned a little bit of coding and automated a big part of their boring job. You would hear the stories, but they were so limited because if you ever look at retention curves of education companies, they go all the way to 1% in like two weeks. No one likes learning to code; it's just really hard. So the few people that make it all the way, the really motivated people, would get tremendous value. So it was obvious that that was the right thing. Then at Facebook I had a similar experience. I went in at Facebook really inspired by Zuckerberg's vision with internet.org. There was a huge effort to connect everyone to the internet. Google had these blimps flying around trying to beam internet down. Elon had just started working with Starlink. Zuckerberg's idea was that the rise of the Android platform was very important, and we're going to go do deals with all these telcos and fund internet for people. I thought the Android platform was a very important platform. So I went in to do Android, and I thought that was the worst programming environment ever. It was Java. You change a line of code, you wait 15 minutes for it to compile. I'm a web programmer. I built REPLs. I'm really into interactive programming. I can't wait. If the compile takes 15 minutes, I'm going to go to Twitter and forget what I'm doing. So I was like, okay, we need to make it better. We ended up inventing React Native. I was one of the founding engineers on that. We increased the population of mobile developers by multiples. Everything that I worked on had this exponential, super linear relationship between how easy you make programming and how many more people would want to do it. So it was pretty obvious to me that this thing needed to exist.

I think it's a balance. You don't want to be too big-brained and think like an investor with all these trends. What Mark Andreessen calls maxing is an approach. I wouldn't go full that or full big brain. Find somewhere in the middle. Have basic assumptions about the world, but in the day-to-day, don't overthink it. Action is important.

If you get a lot of users, you can monetize some of them. We had the benefit of getting a lot of users early on at Replit. The way we did it was by promoting the product as much as we could. At the time it was Hacker News, Reddit, Twitter. Today you would do it via influencers and Instagram and TikTok. I see a lot of founders today hiring a chief influencing officer instead of a CTO, which arguably you might not need these days, but you definitely need someone who's going to get you initial distribution because it's so hard to get any attention. So that was the thing at the time. We always wanted the product to be naturally viral, very easy to adopt. Today we just announced our self-service enterprise, which is kind of an oxymoron, but basically you don't have to talk to our sales team to adopt the enterprise product. You can do it all automatically, consumer grade. We always had that mindset of building a viral product. That helped. The harder thing was getting enterprise customers. We failed. Our first enterprise deal was going to be Block back in 2023. The wrong thing we did was try to adapt our product too much to what they needed. Ultimately, once you have true product-market fit, the market really drags you, and that's what you need to be focused on.

Definitely not. There wasn't a sales bone in our body. Marketing was always me liking to talk about what we're doing, so we always got some attention. But I think engineering and design, because my co-founder is a designer.

Probably $20 million ARR before we got our first salesperson. That was last year. Our enterprise product is about a year old.

Every cycle there's a different company that's going to kill you. In 2016, Google was going to kill you at any moment. In 2022, OpenAI was going to kill you at any moment. In 2026, Anthropic is going to kill you at any moment. You're always going to get asked, 'How come Anthropic is not going to drop this?' Those are definitely stupid questions. I think VCs think too much in a zero-sum mindset, which is surprising because their reason to exist is to create new wealth, new products, net new value in the world, but most venture capitalists don't think that way.

No, because we have the benefit of being around for so long. We've had competition come and go. You just learn not to stress too much about it and focus on the next big innovation. Replit was the first coding agent on the market, before Cloud Code, before anything like that. Even within the agent paradigm, Agent 3 was before people really realized that agents are truly autonomous. Agent 3 was October last year, and it could run for 10 hours. Over the next three months, we're going to have another big leap that will show what's really possible. If you build a culture focused on competition, you end up in a mimetic closed loop where you're copying each other and it's all incremental. Some parts of the company need to be hyperfocused on competition: your sales people, go-to-market people, some PMs. But if everyone is so focused on competition, you lose the forest for the trees.

I try to drive vision. I make predictions about where the future is headed. It's an underrated skill, and you can develop it. Write some predictions down about what LLMs are going to be doing in three to six months or a year, and back-propagate on that. Test yourself whether you're right or wrong. You can develop a skill to predict where it's headed. Being a student of history, I really like computing history, reading about past innovations and innovators and biographies. You learn a lot about how technologies develop. One thing you learn is to never underestimate shitty technology. GPT-2 sucked, but it was very interesting. You could tell it was going to be amazing. Since GPT-2, we've been really focused on LLMs. Anything that looks exponential, like scaling laws, the moment you look at them you should realize this is going to be insane. We have this machine where you put in more capital and get better performance. It's obviously going to take over the world. Having certain principles helps, and having enough freedom in the company for people to experiment. In March 2024, one of our lead agent engineers, Zen Lee, showed me a demo on his computer. It was the first time I saw an agent that was actually working. It was based on GPT-4.5, and it would work for 30 seconds to a minute and a half, then totally go off the rails. But you could tell it was doing the right tool calls, writing code, running it, looping and debugging. It was working. Vision alone would not have helped, but vision plus enough freedom inside the company for people to innovate.

Because we're based on the right assumptions about the world, we built the right things. We built this great cloud development environment, great sandbox environment. Now everyone's building sandboxes. We built this really great file system for collaboration, which turned out to be very useful because anytime you're working with an AI, you're collaborating with it. We built a great deployment environment and a bunch of cloud services around that. The big unlock was that we're making programming easier by reducing all these barriers. The ultimate barrier is coding. If you remove coding, you're going to get an explosion of growth and activity. It's the same principle I've seen over and over in my career. The moment we put Replit Agent and hooked it into all these different services, we took our IDE team, which was responsible for building the interface for people coding, and said your user is no longer humans. You are now building interfaces for the AI. We took our cloud services team building databases, object storage, and said you're no longer building ways for humans to provision them; agents will be provisioning them. We did that reorg, but it was kind of a reshuffle because we didn't move any teams around. We just created this interaction where the user is no longer the human, it's the agents. Once we released the product, it was an immediate hit because it was the first time people could try an agent that could work.

What percentage of your users are now agents?

Is it using Stripe projects?

How do they pay?

Okay.

Exactly. Agents talking to agents is going to be really interesting. How do you think about everyone spending loads of money on tokens? Are you thinking long-term models will get commoditized, or will you train your own? What's your belief?

I start from the belief that the ROI our users are getting is huge. It's orders of magnitude. The startup founder who was quoted $100,000 to make his app made it on Replit for $100. Very few technologies can deliver three orders of magnitude of efficiency. So from that perspective, it's very cheap. But from the perspective of the average individual trying it and failing, paying a lot of money to do that, and then getting good at it, you might need a couple hundred bucks before you produce something valuable. Or someone just wants to play with it. It's a little costly on that end. Hopefully that gets better. Frontier is getting more expensive. Google just launched Flash 3.5, and it's three times more expensive. GPT-5 is like two times more expensive. 4.7 was the sneakiest one because they didn't change the price, they just changed the tokenizer, making it 35% more expensive. All the frontier token prices are going up. The scope that a lot of our peers were saying that token price will go down, we never built on that assumption. We built on the assumption that we want to build a great product to deliver massive ROI, and the price we're charging will be inconsequential. Luckily, we're one of the few startups in our space that's actually gross margin positive. That being said, things are getting cheaper on the open source models like the Kimis of the world, DeepSeek is going to get better. It starts to eat into some use cases, not everything. We should have a classifier that figures out what you're trying to do, the level of complexity, and then kick it off to cheaper sub-agents, and for more complex things go to the bigger model.

The latter, unfortunately. I should feel happier, but I'm not. It's a character flaw that entrepreneurs have, but it makes people want to do more and produce more. I was like, of course I was right. That's the mindset for better or worse. But every now and then it hits you. I was at 10 Downing Street meeting with some government folks, and I was like, what am I doing in these rooms? That's odd. So there are moments where it feels surreal, typically in certain social settings. But the thing about Silicon Valley is no matter how good you're doing, someone else is doing better. You're always comparing yourself to others, and VCs are also comparing you to others. It's a rat race. Ultimately, you have to settle into a stoic mindset of going back to why you're doing what you're doing. For us, it's the users. When I walk around and someone comes up on the street and pulls out the Replit app and shows me what they built and is super excited about it, that means way more than seeing some charm.

I think agents are a new type of software. It took us many years to make making applications easy, standard, maintainable, easy to run and scale. There's no agent builder in the world that's good today. They're all brittle once you deploy them. You don't know what they're doing. Even the runtime model, we're just packaging them as apps, putting them in a container, and they run on a heartbeat and are very wasteful. The big thing for us is to figure out how to create reliable agents. I don't think anyone has cracked it yet, and I think we might be the first.

No, I'm not writing code. The other day I had a moment of sadness. The thing I grew up doing, that I really love and tied my identity to, is dead. I thought we should rent the Computer History Museum in Mountain View and do a funeral. I want to invite Donald Knuth, who wrote 'The Art of Computer Programming.' It's an amazing book about how programming is an art. We had all these ideas about programming, and now it's sort of like what Lee Sedol felt when he played against AlphaGo. You have to realize and accept that the craft you've devoted all this time to is different now. It's entirely different. I don't try to force myself to code. I just pull up my Replit app and make things and have fun with it. I do that quite a bit and try to prototype the next version of Replit inside Replit. That's a lot of fun and keeps me sharp. But my day-to-day is kind of boring. It's meetings. Every now and then I meet with someone interesting, but most meetings are just running the business.

Building, for sure. The YC model. Paul Graham wrote an essay about the manager-maker schedule, which I think still holds. Spend half your time talking to users and VCs, and half your time building, coding, designing, and creating things. I would do that for as long as possible, until you ultimately have to step back and spend a lot of time running the business.

A really cool thing is that I sometimes learn about our launches on Twitter. It's an amazing feeling because I used to be involved in every launch. It felt like birthing a child, having to push it out and yell at everyone. Then suddenly it just happens. That's a really amazing feeling.

No, it's definitely more fun back then. When you're CEO of a larger organization, you can think of it as a sieve or a filter. You get the absolute worst problems. You don't get any of the fun. You get the crap that no one was able to handle. We were handling fires most of the time: PR crises, performance issues, incidents, customers wanting to churn. There's always some problem you're dealing with. But I never felt that fun is a goal to aspire to. If fun is the goal, I would play video games. They're a lot more fun than working.

At Replit, there are two types of people: builders and sellers. But the sellers are also builders. It's really one type of person. Everyone builds. One example is our deal desk person. Once you have a big B2B business, there's someone responsible for producing quotes and managing the entire sales cycle. At our scale, it would be five people, but we have this one woman who joined us. On her first day, she wrote an app on Replit, a Slack agent that handles the process. She automated her job, and now she's basically a developer maintaining this app. We haven't had to hire anyone else in that role. It's incredibly AI-pilled. We also have our own version of Replit Agent internally that is more autonomous and produces pull requests. We have one called Eevee that reads our analytics, DataDog, or traces, and produces bets or hypotheses about what could be better, then generates pull requests on its own. It's starting to creep up in terms of fully autonomous, with no prompt involved, just based on production data. We spend about a quarter of our time building the machine that builds the machine itself. Everyone is trying to automate their jobs at Replit.

No, but you'd be surprised that there are people at these organizations that are incredibly AI-pilled. It takes one person. One of our biggest customers now is Zillow. The account grew by three orders of magnitude in the last year, driven by one person. A lot of people are using Replit internally, but he brought it in, trained everyone on it, and produced the first success. We made tens of millions of dollars, and they made tens of millions of dollars using Replit. The leadership was immediately accepting and wanted to deploy it inside the entire company. That champion is doing a lot of the work for us, training people and doing internal hackathons. It takes one person, no matter how large the organization, to be incredibly AI-pilled and drive everyone around them. Some cultures do not allow for that, but cultures that do allow for it get that sort of person.

Do something useful, but hit the trend early and be first. Getting attention is incredibly hard. When we released agents, it was a magical experience because no one had experienced coding agents before. Karpathy tweeted that this is the feel of the AGI movement, and everyone got excited. We had this edge because if we had launched three or four months later, the novelty would have gone. I'm not saying novelty absolutely needs to exist, but it gives you that initial boost. Unless you're really good at marketing, there are companies that are much better at marketing than us and can get distribution much easier. For us, our product is our marketing. If you're a product, engineering, design, management company, your innovation is your distribution channel. We don't do growth hacking. We don't know how to do it. Instead, we know how to innovate and create moments where people have to pay attention because we've done something so innovative.