Full Transcript for Replit CEO on The Self-Driving Company: How AI Agents Learn to Operate Themselves While Humans Set the Destination | Amjad Masad | E317

I heard you say that not knowing how to code can give you an advantage. I think people who have been coding with AI for a while are also at a disadvantage because most recent generations of AI model like the Frontier are actually better with very simple prompts. What do you think was that pivotal moment where you felt like you already hit that product market fit? The group of people, about like 30 of us, that were working on Replit Agent, which is really the first coding agent on the market, there's no examples of coding agents at the time.

But the moment we started seeing it work internally, we just had tremendous conviction in it. Now, still, we didn't know that we're going to 100x revenue in a year. How far do you think we are from having mainstream adoption where companies can confidently say, okay, we have some automated feedback loops here? I would say like, you

hey this is carlos ceo at product school and your host on the product podcast my guest today is amjad massad ceo of replit 10 years in and he says the first eight were just finding product market fit then they shipped the agent 100x revenue in a year and now they're at 50 million users and a 9 billion dollar valuation chasing a billion in revenue over 75 percent of those users are not engineers and he corrects me early on nobody in ai has recurring revenue It's run rate revenue.

Here are the things we'll cover. Why he says chat was never the thing code is. Not knowing how to code is now an advantage. His prompts have gotten lazier on purpose. The self-driving company and why he thinks it's mid next year. Frontier performance at 40% of the cost through model routing. Let's get into it. AI helps us generate ideas, draft plans, and spin up work in minutes.

But faster doesn't necessarily mean better. When everyone has their own AI tools, their own outputs, and their own context, you get fragmentation, fractured version control, scattered review processes, decisions happening in parallel, not in concert. Rebuilding context between people and platforms costs you time and momentum. That's where Mural comes in.

Mural gives teams a shared visual canvas to bring work together. You can generate structured starting points right on the canvas instead of moving outputs between tools. Keep all of your context in one place. And compatible AI agents can contribute alongside the team, bringing in context from the tools they already know. The future of work is still human.

Learn more at mural.co slash mural dash AI. Angel, welcome. Thanks for having me. To ProductCon and to the Product Podcast. My nine-year-old son couldn't be here, but he asked me to say thank you. Of course. Because he's been bycoding some video games on Replit. That's awesome. Video games have gotten so good with AI. It's crazy. I was just actually at a booth.

Main page at 3:07

One of our sales engineers is always making video games with Replit. He just showed me FIFA that he made. I'm like, this is better than the FIFA I played growing up. And it was like one prompt. Wow. Well, but not everything were rainbows and sunshines. You've been grinding for 10 years, right? And the first eight kind of were the beginning of finding product market fit.

Right. So what's been happening? What do you think was that pivotal moment where you felt like you already hit that product market fit? Yeah. In the history of startups, we've always had this idea that like, you know, companies like you spend two years in C, two years in A, you get your product market fit and you scale from there. But I think that sort of was only like a SaaS era phenomenon.

If you look at old businesses, like it takes about five, six, seven, eight years to build something meaningful. And even recently, when you look at businesses like Notion or Figma or some of our peers from that era, it took them a while to get there. So I think like I've had the patience because I just felt like we are riding on certain trends that are going to become more and more important in the future.

One of them was cloud. Two was AI. Three was like mobile. And four was like collaboration and kind of the future of work. So I've always had conviction. Obviously, there are times when there's a lot of self-doubt. But when we were working on Repl.Agent, We kind of knew it was the future. Like, it was a rough time for the company. We had just done a layoff because we were spending a lot of money and not making any.

But the group of people, about like 30 of us that were working on Replit Agent, which is really the first coding agent on the market, there's no examples of coding agents at the time. But the moment we started seeing it work internally... we just had tremendous conviction in it. Now, still, we didn't know that we're going to 100X revenue in a year.

That was like, no one predicts that. It's really hard to predict. But we knew it was the future. So a few years forward, now you are with over 50 million users, $9 billion valuation. You're aiming for a billion dollars in annual recurring revenue by the end of this year. How is that going? By the way, no one in AI has annual recurring revenue.

So we need to get used to saying RRR, run rate revenue. Yeah.

But yeah, things are going well. I think there are a few things that are interesting. I think that enterprise adoption of AI has really advanced this year. So if you look at the last, like, you know, since GPT-3, enterprises were not really using AI properly. ChatGPT sort of started to change that and then Microsoft Copilot. But if you look at most enterprises today, they just use chat.

Most people here who work at large companies, if you're not an engineer, engineers are typically on the cutting edge. But if you're not an engineer, you just get access to chat and maybe some RAG system or whatever. But now this year, we're starting to see like big banks sign up to Replit, big financial institutions, Visa, Amex, some of our customers, governments, the US government is a customer of Replit right now.

And so we're finally starting to see how AI could transform

work and how we do work. Because I don't think chat really has a big impact. And there was like this big question, oh, we're investing so much in AI, we're not seeing impact. I don't think chat is the thing. I really think ultimately code is the thing. The idea that everyone can make software, can create code, and can like work with agents, and most agents are coding agents, essentially, to get their work done.

That's quite transformers on the economy. And we're in the early innings of that, but the compounding nature of it is quite astounding. I want to learn more about it because it seems like over 75% of your users are non-engineers. And I heard you say that actually not knowing how to code can give you an advantage. I think people who have been coding with AI for a while are also at a disadvantage because most recent generations of AI model like the Frontier are actually better with very simple prompts.

my prompts have gotten lazier over time. I use sort of like less skill files, less... So models are getting more and more capable. So the more you try to micromanage them, the worse they are. You know, there's like... this wisdom in Silicon Valley that micromanagement is not great, which is true, right? Like if you have a talented engineer, you shouldn't be micromanaging them.

But like most engineers, their instinct is to micromanage the AI. Whereas Someone who doesn't have that instinct might actually get better results. And you're starting to see that. I saw that with Replit Agent, you're able to spawn multiple sub-agents, right? And the next thought that comes to that is how do you actually, let's call it orchestrate or oversee, I don't want to say micromanage, but to make sure that at least you have a good sense of what's going on.

As a human user? Uh-huh. So the UI obviously needs to give you so, you know, with parallel agents, we have like a thread for each one of them. You can sort of like dig into each thread of work and see what's going on. Also, agents come back with questions and we have the mobile app. If you install it, it'll like send you a message. It'll be like, hey, like I'm going to go over budget or I'm going to use this new package or I'm going to do this refactor.

And so you have that sort of oversight. And also the agent kind of learns your pattern of usage. And so if you're someone who is totally hands-off, the agent kind of learns that. But if you're an engineer who is micromanaging the AI and that's what you prefer, more power to you and Replit will sort of adapt to that. So you recently coined this term of the self-driving company.

How can companies learn to learn? So I think there is a compounding nature of learning to learn. Very much like

everything we generate can go back into context. The problem now with most companies is most people are using Cloud or ChatGPT on their desktop. And that is actually not feeding back into context. At Replit, most of our AI usage happens in Slack and a lot of it in public channels. And so everything is self-referential. We write documents, they kind of go back to the context.

Main page at 9:52

And so collaboration becomes very important. I mean, we spend a lot of time figuring out how to do collaboration with AI on Replit, like having multiple people on the same project. None of the big labs are focused on that, but I think that's like a mistake. Like you would want like I think it's more important right now to approach, you know, building companies more collaboratively and doing work more collaboratively because we're going to have less work on our plate to go do as individual contributors.

And where new ideas sort of emerge happens from brainstorming, happens from building on each other's ideas. You almost want to approach building a company like an open source community, where in the open source community, things might be a little bit chaotic and there's no centralized leader. But everyone's building on top of each other's things.

Everyone's learning from one another. And I think the future of how companies work As companies become more autonomous, when a lot of the engineering is autonomous, when the support is autonomous, when a lot of the operations are autonomous, you have to ask, what is the role of the human? And ultimately, I think it is knowing what the customer wants, I think figuring out what's the best thing to build, figuring out the major decisions, kind of setting the destination, and then trusting the self-driving autonomy of the company to do it.

to do the rest. A huge unlock for me was to be able to start using Replit within Slack. And I see that you also have a similar integration with Muse or with Cloud Code. So I'm curious to know how you're thinking about those type of external user interfaces as a way to enhance the collaboration.

Yeah, I mean, I think we're all sort of trying to discover what that looks like. I don't think anyone has figured out what's the best kind of user interface. I think operating system is the right question. I think we're all using Slack as an operating system in some ways for companies. but it is not a good one. It is basically dumping ground for all of text.

And it has become very hard for the human user to absorb everything, which is why we're having to rely on agents more and more to summarize more of the communication, more of what's happening. You might want to at some point have like a human only sort of interfaces, because I think a lot of the stuff that agents are generating Or like, kind of, you can think of it as chains of thoughts that you might not want to worry about.

So anyways, there's a question of unlegibility here, but I think it's a very open question. And it's like one of the biggest startup opportunities today is figuring what that is out. And then for the technical user, for the engineer, one of the things that I noticed is that when you publish on your blog, that you are able to increase the output of your code through output without sacrificing quality and other maybe traditional trade-offs that you would expect as a result of just putting more gas into the fire.

So how are you able to combine that so you can really increase that productivity and that throughput without making the quality suffer? So early on, when the whole vibe coding thing was happening, Replit was probably the only player that was compute maximalist. So we were always trying to generate the best software, the most secure. And we built a lot of the...

things you can string together in the market, but like we have built-in code review, we had built-in security agents, and it made Replit a little more expensive, a little slower, but you can trust the results more. And so now internally at Replit, we all use Replit to build Replit. So internally, we are a lot more comfortable than most companies opening pull requests to designers, to even salespeople.

The reason why a lot of companies might not want to open that up is because there's still like a safety issue. There's still like code quality issue. If you have really good systems and guardrails, I think most companies would want to open up

their GitHub repos to non-engineers. So are those guardrails and systems created by default or do companies need to set up something just to make sure that the quality is there? Bigger companies need to do that. I think smaller companies don't. But now with bigger companies, we started doing, I mean, everyone's doing it now, but the FDE model where we go in, we figure out the requirements and we work with them to encode their guidelines and style guides and In many cases, in banks and big financial institutions, there's a lot of compliance, regulatory issues, and those are very tough.

And this is a big reason why I'm sure a lot of people who work at these companies would know. This is a big reason why AI has not seen tremendous adoption. There's so many ways in which things can go wrong. And so, yeah, they require a little bit more handholding. You also have, what, like 80%, 85%, Fortune 500 adoption?

Main page at 15:12

Yeah, this is like PLG in terms of like enterprise contracts. It's significantly less than that, but it's catching up. Got it. I mean, the other big piece, I think, is the model routing, right? Like some of the enterprises will have their own criteria in terms of what are the options for their teams. Like, how do you go about that type of routing? Yeah, so we promise our customers that we'll get you the cheapest

best token for any given task. So because we've accumulated a lot of data on usage over the past three, four years, we train models that can figure out what is level, not just model size, but also reasoning efforts and so on. And so we've published benchmarks on our website showing that we can get frontier performance at 40% of the cost pretty consistently.

So when you buy Replit, we sort of, In most cases, again, there's like big enterprises and governments that have certain restrictions. But in most cases, we would want the customer to trust us with a model selection. And we have zero-day retention contracts with all the different providers. Because if we have that freedom, we're always going to be at frontier performance, but also significantly cheaper.

Yes, I think there is like a market dynamic situation where you are collaborating with OpenAI, Anthropic, and at the same time, you could be considered a competitor, right? So how do you navigate that dichotomy? You know, we've always had this issue. Like we've been Google Cloud customer for a long time. They've been trying to compete with us ever since I remember.

Like they had like an online ID called like IDX and they had something before that and they're probably all dead now. And now they have like the Google Studio. I mean, it never really bothered me. The way I view these big companies is that there's like two or three things that they can't lose on. And then they're going to do 100 or 1000 things that are less important to them.

I mean, initially, when everyone was adding vibe coding to their platforms, you know, a lot of investors were really worried about about us. And, you know, I kept saying that we just do this one thing, which is, we're one stop shop for creating software. And the depth of the platform is very important. Like the consumer use cases, yeah, anyone can make sort of any sort of website anywhere.

You can ask an agent to make that. But if you want to build a business on Replit or if you want to operate a business, I think we're still the best at that because we provide all the infrastructure. We provide the security guarantees, you know, compliance, privacy, and so on. It doesn't seem like... You know, if you build a website on, you know, Muse or Codex or whatever, they send it to Cloudflare, some external provider to like host it for you.

With Replit, it's all one thing. And so the agent has access to the entire environment, development and production. And we also have better control. Like when you trust Replit, you only trust one player as opposed to trusting like three or four different providers. Totally. I think as the models commoditized, the value got pushed upstream and the application layer has a more differentiator.

Plus, you do not have to do the heavy investments in building the infrastructure. Yeah. I mean, that's the main thing. I mean, so, you know, our power users tend to be either entrepreneurs who are like building, you know, multimillion dollar businesses, many times solo. And then like, you know, corporate employees that are like doing amazing things with Replit as well.

And the common thing between these two groups is that I don't want to manage Infra. Yeah. Someone told me that the difference between micromanagement and attention to detail is that it depends on how much agency you give to the other person on the other end. Yeah. So anyway, going back to the self-driving company, I think that's a fascinating concept that still feels futuristic.

Main page at 19:15

How far do you think we are from having mainstream adoption where companies can confidently say, okay, we have some automated feedback loops here? I would say like mid next year,

companies will be comfortably running agents doing real work. Right now, most agents happen on employees' laptops. Very few companies... perhaps outside of tech, that are using agents to do real workflows. So I think mid-next year, it'll feel like you really have digital coworkers. I think in the last few weeks, we're also starting to see a massive adoption from these AI assistants like Muse, Instinct, Crockbot.

How does that super cycle impact your business? It is good in that more people knowing the power of these tools will demand their companies to have those tools or bring some of these tools to their companies. It is challenging in that I don't think consumer vibe coding is a viable market anymore. And it's something we've made some money on.

But from the start, I knew that was going to happen. So that's why... I always thought that, you know, we should monetize enterprise and businesses and not really monetize consumer, which is now we have a free, we have a free, like you subscribe to Rebel for $20 and you can just do code forever. So we have like a free or like unlimited plan now.

But I think it's all good. Like the more people understand how impactful these tools can be in their lives or their work, then adoption will happen faster. And that's good for our business. Well, we're going to leave it here. I would just say that it would be awesome if one year from now, six months from now, we're sitting next to each other and we say, remember what you said? Yes, there's so much adoption of this self-driving company.

I think it's one of the best times to be alive. It's really, there's a lot of things to build. A lot of people are trying to convince you that there's no reason to create startups or businesses anymore. There's tremendous opportunity. And at least for the next two years, it's like an amazing time. Thank you, Amjad. Of course.