Pod Insights Y Combinator Podcast Replit CEO Amjad Masad: Coding Agents, Autonomy, and the Future of Work About this episode Amjad Masad started Replit to make programming accessible to anyone, anywhere. What began as a tool for learning to code has grown into a platform pushing the limits of AI-assisted software creation and recently surpassed $100M in ARR. On The Breakdown with Tom and Dave, Amjad shares the journey from his early days in Jordan, working on open-source projects, to leading Replit through major pivots from "teach a billion people to code" to "let anyone build software." He discusses the evolving nature of programming, the future of work, and the next generation of human-computer collaboration.
Transcript 0:00 Speaker 00 I think the dystopian view of AI or AI as like taking all our jobs and all that is like not correct. And I think the future of work is more human, is more interactive, is more multimodal. Once the making of things gets easier, the ball like goes back to like how many ideas you can have. I wouldn't put learning code as the top thing. I would put learn to make things. Learn to make things with code, learn to make things with video, learn to make like anything with AI.
0:37 Speaker 01 Welcome back to the Breakdown. I'm Dave, this is Tom, and today we're joined by Amjad Masad. Amjad is founder and CEO of Replit, which was a YC 2018 company. Welcome to the show.
0:47 Speaker 00 Thank you.
0:48 Speaker 01 So today, we want to talk about kind of your beginnings in helping people learn to code, what you're doing now in replacing a lot of that work, and then where we see the future going in terms of how people make things. So Replit started, I guess, in 2016. You were in YC in 2018. You started out as a tool to make it really easy for people learning to program, to set up development environments purely on the web. But now you're really taking off in AI assisted coding. Tell us the latest.
1:18 Speaker 00 Yeah, so the mission has always been to make programming more accessible. Then we sort of, you know, after YC, we sort of updated it to be a little more ambitious. We start talking about a billion software developers. It just sounded absurd at the time.
1:33 Speaker 01 And this was all pre LLM, right?
1:35 Speaker 00 Yeah, pre LLM. There was a moment of time, I don't know if you remember that, but like 20, in the like 2015 AI hype, there was like a bit of a NLP, natural language hype as well. There's a bunch of AI companies that ended up being just like humans behind the scene. They all went broke. But there was like a glimpse of potentially doing NLP on code. So we had an idea that, you know, it's probably coming. I even had it in my seed deck that like at some point, like we'll collect enough data to like train models. It wasn't until like GPT-2 in 2020 that it felt like it was going to be possible. So we, you know, we had built all these primitives. Like the development environment, the sort of hosting environment, all the other stuff around it.
2:21 Speaker 00 And it felt like if we just like added like AI agents, like it will be able to orchestrate all this stuff. And it'll be, it'll be really great. But agents were like every time we tried them. I think the first time we tried it, like 2021 didn't work, 22 didn't work. And then 2024, like early, early in the, in the year, we just felt like it's getting close. Even GPT four O could like be coherent for like two minutes. Somehow just ended up doing a big bet. The company was not doing super well.
2:55 Speaker 02 Yeah. Was, I was about to ask, was there like a bet the company moment where it's like, we were teaching people to code. Now we're no longer doing that. Now we're going to enable everyone to build apps.
3:02 Speaker 00 Yeah, yeah, yeah. We had grown the company quite a bit. And we were burning way too much money and we decided to do a layoff. And so we, we cut like perhaps 50 people and then like another like 15, 20 people left. And so we were like less than half of the company. And so I just put everything on, on Repli Agent. I just felt like this is the thing. It has to work. It felt like it was close. Yeah, I don't know. I was just like burn the boats, boats kind of moments. It's just like, we gotta do this. I mean, it's, I felt like this was the, this was the, the thing that would make the company work. And honestly, like if Claude 3.5 hadn't come out mid us building agent, we would have probably failed.
3:45 Speaker 00 Because like I said, like GPT four O would like stay coherent for two, three minutes. Claude 3.5 was the first one that like could work for like, you know, five to 10 minutes and actually, actually work. And the code generation is all.
3:56 Speaker 02 And that's sort of the approach that I think is most successful for startups, kind of building on the very edge of what is possible. You start out on a mission. It's the technology's not quite there yet. You start building and you sort of, you, you skate where the puck is going. You know, it sort of catches up with you.
4:09 Speaker 00 Yeah. And you want, you want to keep, keep doing that.
4:11 Speaker 02 Totally.
4:12 Speaker 00 Yeah.
4:12 Speaker 01 Yeah, I remember you, you came to speak at Y C like six months ago and right after you had launched this. And at the time you said, we're very far from fully automated software development.
4:22 Speaker 02 Do you still feel that way?
4:23 Speaker 00 Not at all. I mean, I feel like I've, you know, every time I make a sort of some kind of prediction, it feels like bold and like, you know, but I've been, I've been consistently wrong. And like the way things are moving is a lot faster. Just the level of autonomy, like I said, like maybe 3.5 was like five to 10 minutes, like 3.7 perhaps can get up to like four to five minutes an hour. The 4.0 in the, in the system card and OPUS, they said they made it work for seven hours.
4:53 Speaker 02 Wow.
4:54 Speaker 00 Seven hours.
4:55 Speaker 02 That's incredible.
4:55 Speaker 00 That's always been the limiting factor for making agents is like, can they stay coherent? Can, is the context length actually useful to reason over? If an LLM can work for, for seven hours, that's basically like a human worker.
5:08 Speaker 02 Yeah. And presumably they're working at much faster speed than a human would do. So they're kind of completing perhaps a week's work in those seven hours.
5:14 Speaker 00 The only thing that I think is, is missing and a big limiting factor for actually automating a lot of work is computer use. Computer use kind of sucks.
5:23 Speaker 02 Yeah.
5:24 Speaker 00 I'm sure you've, you've tried it. And this is the difference between Replit Agent, really giving it one prompt, going all, all the way in an app versus having to kind of babysit it a little bit and like test with it.
5:36 Speaker 02 And we had a company in the last batch called Browser Use that's, that's working browser automation.
5:40 Speaker 00 Browser Use is great. Yeah.
5:40 Speaker 02 And another one called Pig, which is doing the same for kind of Windows desktops.
5:43 Speaker 00 Yeah.
5:44 Speaker 02 And I think advice I would give founders today is taking either browser use or that Windows automation with Pig and trying to apply that into enterprise, into vertical, into a vertical industry.
5:55 Speaker 00 The moment this technology worked, those two companies are just gonna.
5:58 Speaker 02 Totally. And I think, I think we're like weeks or perhaps single digit months away from it working really, really well. And so now is the time to get started using those technologies.
6:06 Speaker 00 Absolutely. A hundred percent. And this is, this is what we're focused on. So, you know, Replit Agent, V1 to V2 was a huge jump in autonomy. V3 is the most autonomous thing. So we're already working on it. And the interesting thing for us is the underlying technology and how it would enable autonomy. There are like a few important things. One is, you know, being transactional, the ability to roll back.
6:32 Speaker 00 Is is very important. So you you would want you'd want it to be safe for agents in the same way that gets made it safe for human programmers. To kind of experiment and create branches, whatever. You want the same thing for for agents. Like if an agent can messes up a DB migration, it should be able to roll back. And also it should be able to like sample across different different paths. I think this is very, very important for autonomy. If you look at. Like, for example, when when Anthropic publishes their Sweet Bench score, they publish one without sampling and one with sampling. And it goes from seventy percent to eighty percent.
7:09 Speaker 02 And so this is the idea that you sort of you spawn multiple agents, each one will take a shot at it, and then you figure out which one works and choose that that branch effectively.
7:17 Speaker 00 We built a infrastructure that is fully transactional and moves in lockstep, meaning that a file system is a snapshot fail based file system. The DB is a snapshot based DB. And so as you're we're making commits to the entire system, including the virtual machine as you're going. And so what we can do is also we can we can fork it and branch it out. So if we had computer use that's working well, you can you can just like. Right now, what they do with the sampling is they do they have some kind of judge that does ranking, which, you know, it's not a true verifier. A true verifier is a is a is a test, right? And a computer use test. So you you can sample out and really pick the best branch that's actually working.
8:02 Speaker 02 That's incredible.
8:03 Speaker 00 And and repeat that on and on and reliability gets really, really high.
8:07 Speaker 02 And so fast forward six or 12 months, you're not spawning one agent, you're spawning, is it five or 10 or a million? Like, how?
8:14 Speaker 00 I think this is where it's going to get interesting, which is you'd want to give the user the ability to compute budgets. I think we're starting to see that with with the old style models of like, here's here's how much budget. But but look, I mean, if you give us a thousand dollars, we'll spend them.
8:33 Speaker 02 That's really cool. I've not thought of that before. But the idea is like spawning multiple branches and then just picking the best one and being able to do that in parallel. I just the human brain doesn't naturally work like that. We think sequentially. We can't.
8:43 Speaker 01 It reminds me of this legend. I don't know if this is actually true, but at Apple, when Steve was running it, he would intentionally have teams doing basically the same things and then see which one did the best job. It's like.
8:53 Speaker 00 I've heard OpenAI does that now.
8:54 Speaker 01 Interesting.
8:55 Speaker 00 I've heard that like the Codex project like had multiple different teams. So in the literature, you'll see that small models sampled will beat larger models. So Sonnet sampled is probably better than Opus. Companies haven't tried that where instead of hiring one senior engineer, you hire like 10 junior engineers. And like, give them the same task and then.
9:20 Speaker 02 It's just very expensive with humans, but with LLMs it's relatively cheap.
9:23 Speaker 00 That's right.
9:23 Speaker 02 You can just do that. Do a hundred of them and just pick the best one every time.
9:26 Speaker 01 So how are people using Replit Agent? And who are the people using it?
9:30 Speaker 00 You know, it goes back to sort of our early vision that if you make programming easy, then more and more people would want to do it. Actually, this is one of the first thing that PG and I sort of connected on. Before we got into YC, actually PG found us on Hacker News and so we started this email relationship. And so he told me there's like a super linear relationship with how easy programming is versus how many people would want to do it. The optimizing function of Replit has always been like just keep lowering the barrier to entry and that's how you grow users and you grow customers. Right now, basically people from kind of every walk of life, we've seen users.
10:13 Speaker 00 Product managers tend to be a very, a great use case and users for us. And we've had customers, product managers, who are able to make significant impact on the business without talking to engineers at all, like you know, running A B tests or optimizations or things like that. I mean, it's empowering. I mean, it gets us to think about the really the seams between these roles, like you know, product manager, designer, engineer. We actually just created a new product group and typically, you know, product, you know, head of product has a bunch of like product managers reporting to them. But we're actually having this group have has designers, engineers and PMs.
10:57 Speaker 00 And the idea is they're all using AI all the time to prototype. In some cases, go all the way to production. There isn't this sort of waterfall model where you know, there's a lot of inefficiency, like communication problems between these different teams, they can move incredibly fast. So I think that that starts to change how tech companies work.
11:15 Speaker 02 My experience of this was whenever I was working in a startup, the list of ideas in the backlog would always be infinitely long, basically. And the bottleneck was always engineering time. And now I'm doing my own personal projects. I write my to-do list or my idea list, and then I get to work and the ideas just get done. And suddenly, the bottleneck is like my ability to have ideas.
11:34 Speaker 00 Yeah, yeah.
11:35 Speaker 02 And it's just such a weird experience, looking at a to-do list that's just empty, being like, what do I do next?
11:39 Speaker 00 I've heard from a team that has like a really large Replit deployment in their company, that their founder is using Replit. And that's stressing the engineers out because, oh, I did this in a weekend. Like, can't you do it in a month? What are you guys to show for yourself?
11:53 Speaker 02 And so are these previously technical people and non-technical people, they're building the first version, are they deploying that production? Or are they giving to engineers and say, hey, build this? Like, what do you see typically?
12:03 Speaker 00 We advise to like work with engineering, but that doesn't always happen. I think it's understandable that a lot of PMs and designers want to go straight to users. What we're seeing is they go to like beta users and test users. And I think this works really well. But in some cases, they just like put it in production. And so right now, we're having discussions with all these companies, especially the engineering leaders are unhappy about this. Like, who's on call for these services? Who's responsible for the bug? Who's responsible for it? There's a lot of questions that are coming up. The obvious answer to all these questions is like, agents are responsible.
12:38 Speaker 02 So what's the limiting factor today? I buy code a thing and I'm like YOLOing it to production. Like, what do the engineers typically object? Like, what's going wrong?
12:48 Speaker 00 Security is a big one. Like, LLMs are fallible, like humans are. They tend to write some, there's some components that they do terribly at. Like, for example, OFF, like, they all kind of suck at it. They all use like, you know, old methods of assaulting and hashing. And that's been, that's been a big sticking point. And we've seen a lot of examples out there right now of some catastrophes. Luckily, it hasn't been, like, there hasn't been a major catastrophe. I think it's coming. But there's been like solo founders who would leak, you know, API keys or would make it really easy to get around the login security protections.
13:31 Speaker 00 And there are a lot of tools out there that are like really not trying to take responsibility for that and saying, oh, it's the user's problem, the user's fault. For us, we think that as a platform that is marketing for the non-developer, you actually have a responsibility. We're trying to take away some of the things that we think LLMs should not do today. So auth, for example, we have a built-in auth. So if you go to a repl, just say add auth. It will pull in an auth component that we built from scratch. It has captcha on it. It has all the security bells and whistles. So you don't have to worry about all of that. We integrate it with your database. You have a user management portal on your site.
14:15 Speaker 00 As much as possible, building it in these components that are tricky, I think payments is another one. I don't think you'd want the LLMs to kind of do the payments.
14:24 Speaker 02 And they're not that different, right? Like the sort of checkout, you know, you might have one-time checkout, you might have a subscription, pay as you go. They're not unlimited versions of payments. There's like two or three or four.
14:33 Speaker 01 And this is, the analogy is today, humans don't write their own version of payments or their own version of auth. They use providers that have built these components already. So it seems like we're seeing the same thing play out and that's the best way to do it.
14:44 Speaker 00 Yeah, 100%. And then the other thing we added, we partnered with is a great security company called SEMCRIP. And so right now, when you go to deploy a repl app, we run a security scan. We run a code security scan. We give you like a report of warnings and errors and things like that. And the agents can try to fix them for you.
15:04 Speaker 02 Okay. And so security is one of the big kind of bottlenecks to fully deploying into production. I can imagine there must be other things that coming down the line, like sort of scalability, you know, looking for like N plus one database queries, performance bottlenecks. What do you have in your mind, like a clear set of blockers that need to be overcome before this is truly like one click to deploy?
15:23 Speaker 00 Yeah, I mean, the big, the biggest one is just going to be humans like social, like, you know, there's just going to be mistrust. And I think it's just going to have, has to play out. So, you know, leaving that aside, like enterprises has to adapt to all of this stuff. Having some way to also scan for scalability, like figuring out, you know, like doing fuzzing or.
15:46 Speaker 00 Whatever it is, or having some kind of adversarial agent that's trying to break your app. I think that's one big thing. I think integrating with a company's ecosystem. So one thing we're adding is the ability to bring in your design system.
16:00 Speaker 01 That's cool.
16:01 Speaker 00 So in addition to kind of us providing these components, if you go in a company, they have a lot of these components built out. As Rapplet is getting deployed into these large companies, how can we hook into their internal systems?
16:13 Speaker 01 This makes me think about the kind of spectrum of these different coding tools. On one of the spectrum, you have kind of what I would call the power tools, the cursors, the windsurf that let developers use this as a way to amplify their efforts. And on the full other end of the spectrum, you have more of the like consumer facing, hey, if you want to make an app, like you can now make an app. And it sounds like you guys are kind of in the middle. You're helping companies get stuff done, but you're doing it for people who don't look like the traditional developer.
16:43 Speaker 00 Yes.
16:43 Speaker 01 How do you see this world playing out? Are there going to be like N different tools? Or will we converge to one place on that spectrum?
16:50 Speaker 00 AGI is convergence, obviously. But leaving that aside, it's really hard to plan for that world. I think that the battle for how to incrementally make engineers more productive is just, it's an obvious one. The market there is obvious. There's a lot of companies going after it. Both the application companies like Cursor, the underlying model companies are trying to go there. I mean, Claude Code is competing with Cursor. Cursor uses Claude. I think it's a bit of a bloodbath there, but the market is obvious and the market is really large. I would guess that there's going to be more of a consolidation there. Maybe it's not one, but it's probably two or three at best. I think the sort of the market we're in is a lot larger.
17:36 Speaker 00 So the addressable humans is a lot more. You know, we're talking about a billion, but it could be more. Like really any knowledge worker should be able to like solve problems with software. My conception of Rapplet right now is the, what we want it to be is a universal problem solver. It'll solve problems in your personal lives or solve problems in your work. And all of that. And so I think that that market is it will probably be like a little more diverse. And I think companies will kind of figure out where where they slot in. We're trying to solve autonomous programming with the focus on the non engineer. We want you to not worry about security, not worry about systems, not worry about any
18:19 Speaker 00 of that. We want you to really come in with your ideas to to Replit and be an agent's manager. And we're trying to do it in a way that is like as human as possible and it kind of fits into into into the workflow. One big difference between engineers and sort of non engineers, be it executives or product managers, you're not on your desk like eight hours a day. Mobile is a big part of that. We have like a really great mobile app. And we're thinking about this way of like ambient building. Maybe you start an app on your in your desktop, you go away with your phone, you're in boring meeting, you get a notification from the agent saying, I'm done with this. Do you want something else? You kind of texted. And so we're trying to kind of build that thing.
19:01 Speaker 02 Yes, a question I had related to that point is about the user interface. So for something like cursor or windsurf, it's pretty obvious. The primary UI element is like the code. You see code, you have a little chat window, but primarily it's about diffs. It's about kind of changes to code. And for a tool like Replit, the primary interface is like the graphical user interface. You know, it's the it's the buttons and the the font.
19:27 Speaker 00 WYSIWYG, like you see what you're building.
19:28 Speaker 02 Exactly. And that's great for building user interfaces. But when you're trying to build more complicated sort of logical flows, I found it a little bit difficult because I couldn't, there was no way I couldn't see the code. I can't visualize what's going on behind the scenes. And it's sort of, it's almost a black box. Fast forwarding here, if you're trying to build more complex internal workflows, how do they, how does a product manager or an operations manager of a company visualize that workflow and the kind of logical branching that happens?
19:52 Speaker 00 Yeah, so if you look back in the history of computing, there was always this vision of visual programming. It never worked very well because ultimately it's about Turing completeness. Like, these systems are not universal computing devices. And now we go to code gen. Obviously, code gen is Turing complete, but you're interfacing with it primarily via natural language. Natural language is fuzzy. It's really hard to know whether it's doing the right thing. I think the synthesis of these two things is probably coming, where you are interfacing with natural language, but you can, instead of like just staring at code, there's maybe an interface or like a different view on top of code.
20:37 Speaker 00 You can imagine being able to... Do you know Smalltalk?
20:41 Speaker 02 vaguely.
20:41 Speaker 00 Yeah, so Smalltalk is this... It's the first object-oriented programming system. And, you know, Alan Cave would say it's like, it is actually OOP where everything comes after it, it's not. But the interesting thing about it and Smalltalk, you can actually, the way you interact with code is not via files, but via objects, via like logical objects. And so there's some kind of prior art there. And I think the world we're headed in, where there's some kind of abstraction over code that allows people to like, understand it.
21:13 Speaker 02 Yeah, I think that's really interesting. There's like an open space there, whether it's like pseudocode, who's like, looks like English, but it's a little bit more structured or it's a visual drag and drop, I don't know.
21:22 Speaker 01 Yeah, I think back to building products with a team of engineers, designers and other folks. The interactions that I would have as the like product lead with those teams was verbal. It would be written, like abstracted ideas. We would draw stuff on the whiteboard together. We'd make system diagrams. We'd look at the results of the app and we would test it and point out like, oh, this thing is broken, that's too slow. And it feels to me like that sort of interface, which is very multimodal and very flexible, is probably the best end result. Like, I think we will get to something like that, where the author of products will be doing that, but the teams they're talking to are not other humans, they're agents doing these things.
22:00 Speaker 00 Has there been an attempt in sort of PM land to have a little more formalism around communication?
22:08 Speaker 01 Yes, and I would not say it's been good. The main thing is just like the PRD, right? The product spec, right? And it's just become, in my opinion, oftentimes a performative work artifact that you create just so that you can have a thing to get your promotion, right? If you're at a big company. But they're not actually that useful. To me, the most useful interactions are just these like whiteboard conversations, right? What do we want it to do here? Oh, we got to think about that. Oh, we didn't consider this. Okay, let's rethink this whole thing. Like, those sorts of conversations.
22:36 Speaker 00 Yeah, I can play a role in that as well. So I, you know, this startup Granola that allows you to kind of record meetings. And they released this like team version that like all the meetings get transcribed and go there. And they have a mobile app right now where like you can put it on the table and like it's trying to. So I was thinking, maybe we should go Granola maximalism. Where, look, you shouldn't fight the trend in which companies are becoming increasingly oral as opposed to like written. And because like, you know, people are talking on Slack, people are in meetings, people like are communicating via prompts with agents. But you would want a set of AI tools
23:19 Speaker 00 that is actually like creating that record in the background that's like searchable and organisable and all of that.
23:28 Speaker 02 Yeah, I wonder when we'll have our first AI in like oral meetings. You know, you're jamming with your designer and the AI is like chips in and say, well, how about this idea?
23:37 Speaker 00 Yeah, and this is where I think the like dystopian view of AI or AI as like taking all our jobs and all of that is like not correct. And I think the future of work is more human, is more interactive, is more multimodal, is more fun, in my opinion.
23:54 Speaker 03 YC's next batch is now taking applications. Got a startup in you? Apply at ycombinator.com/apply. It's never too early and filling out the app will level up your idea. Okay, back to the video.
24:07 Speaker 01 All right. So last time we chatted, you had launched Replit Agent and things were growing really crazy. I imagine that has continued. Anything you can share there?
24:16 Speaker 00 Maybe soon we'll share some numbers, but since Replit Agent launched, we're growing 45% compound monthly average.
24:24 Speaker 01 These are like metrics that we tell YC companies during the batch when they have a base of like no users to try to achieve. And you're doing this at larger scale.
24:32 Speaker 00 Yes. But you know, it put a lot of strain on the company and our systems. We're still relatively relatively small. I feel like it can get to your head and you can start optimized for the wrong thing. It's very easy in AI to increase ARR while users are not happy because they're spending a lot more and like not getting the results. And in some cases, maybe it shouldn't grow that that fast because like you'd want users to get a better experience for less money. And so it's one thing that we try not to obsess. We actually don't have ARR goals at Replit. We have like more product goals, retention goals, just like other metrics.
25:10 Speaker 02 Yeah, I can see the sort of bad pattern with some AI companies. You grow top line revenue very, very quickly, but the churn is like approaching 100%. And eventually that just catches up.
25:17 Speaker 00 And the gross margins are horrible, too. And so it's like, you know, the more growth you have, the worse the company is doing financially.
25:25 Speaker 02 So how do investors see this space? You must have talked with a bunch. Can they tell the difference?
25:30 Speaker 00 It's all kind of a blur for them because investors, when they, I mean, I'm just going to generalize here, but when they start looking at the space, they'll use everything for three minutes and everything for three minutes looks the same. So I think it will start to clarify and these products will start to, I suppose, converge, diverge more with the different focuses in the areas we're talking about. I think over the next year, it'll be like a little more clearer. But I think a lot of them are just like, you know, when we talk to them, they're super confused. They don't understand these systems. They don't understand where they're going.
26:06 Speaker 02 I think that's a great segue into the sort of technology underlying some of these tools. I'd love to dig in a little bit deeper. We've seen some announcements from Cursor and Windsurf about how they're kind of layering, whether it's Claude or Gemini or the OpenAI models with then. Their own layers on top, the fast apply APIs, stuff like that, fast apply models, I guess. Can you give us an overview of how it works at Replit?
26:29 Speaker 00 Yeah, a lot of it is you're patching problems with the underlying frontier models.
26:34 Speaker 02 Yeah.
26:35 Speaker 00 So the reason fast apply was important because none of the models were doing very well at diffs.
26:41 Speaker 02 Can you explain what fast apply is?
26:43 Speaker 00 When you're trying to like edit a file as an L M, the best thing to do is to like create a create a diff. But these models are actually not very good at creating this. They're actually not very good at counting the lines in the source code. So they get confused about a lot of these things. So for a while, a lot of companies were just like rewriting the entire entire file.
27:04 Speaker 02 And so diff for for non-tech users is sort of like remove these three lines and inject these other three lines instead, like find and replace almost.
27:12 Speaker 00 That's right. That's right. And so you needed some way to make the model generate something of a diff. But it's actually not good enough to merge it. And so you need another model to actually do the merge. So.
27:27 Speaker 02 Instead of rewriting like it's a thousand line file, just like output the entire thousand lines.
27:31 Speaker 00 It's very slow. It's very expensive.
27:33 Speaker 02 Yeah.
27:33 Speaker 00 So you prompt the model to be as lazy as possible.
27:36 Speaker 02 Yeah.
27:37 Speaker 00 But in that case, it's really hard to apply. So you need another model that's like doing the application.
27:41 Speaker 02 Okay.
27:42 Speaker 00 You can train the model or you can use Gemini Flash or some of these smaller, smaller models. And so in some cases, you need to to train a model or fine tune a model to do a better job at it. Now the cases, you just also patch a bunch of other models to do it. It's engineering. It's engineering as opposed to like research, I would say.
28:01 Speaker 02 And I noticed you guys don't expose the underlying model to the user with other tools like Curse and Windsurf. There's a dropdown. It's like, I want to, you know, let's see what Gemini thinks of this problem. You don't do that. Why is that?
28:13 Speaker 00 A big part of our research efforts is in evals. And I think this is like an underrated part of like, you know, AI coding. So we spend a ton of time evaluating new models, writing evals, generating evals, just data crunching, trying to figure out what the users are getting or feeling. We built a lot of systems to try to understand how these systems are performing. The moment a new frontier model lands, we are evaluating almost immediately. Gemini, for example, like, you know, a couple of months ago, we're making the rounds. Really great model, really awesome at like one-shotting, in some cases better than Claude.
28:54 Speaker 00 Agentic work, it wasn't like a tool calling, you know, some of the things like that. But users just see the hype and they're like, oh, like, yeah, give me Gemini.
29:06 Speaker 02 How much notice do you have? Do they like drop it on you and you're like scrambling? Or do you have like days or weeks before that to try to, you know, test it out?
29:12 Speaker 00 Well, we have good partnerships with these companies. We have a great partnership with Google. We have a great partnership with Anthropic, even with OpenAI. We have a close relationship. So a lot of them give us, give us a heads up, give us some early checkpoints and we try, we, we can play with all of them. And in the case of Anthropic, we're always like the first day kind of launching, because we end up building. A lot of times we're sort of spinning where they're going.
29:35 Speaker 02 Yeah.
29:36 Speaker 00 Because you kind of like, you can tell like 3.5, 3.7, there's some direction. You could tell like where 4.0 is going to land. It's a start architecting the systems. But you know, ultimately, a lot of our engineering efforts still infrastructure, like the distributed network file system, the snapshot based network file system. Like it took us two years to build. Like there's nothing off the shelf to do that. A lot of the security stuff is really, really difficult. Like Replit is one of the few places in the world where you can get like a virtual machine in the cloud. You know, by just like creating an account. And that's like a, it's really hard to kind of run this and protect against this. You have crypto miners, you have all sorts of stuff like that.
30:18 Speaker 00 And Replit also uses Nix OS under the hood. Nix OS is a Like fully declarative, also transactional, uh, operating system generator, as it were. We have like this multi-terabyte hard drive in all the different regions that we have compute. That that that cached all the packages in the in the world and that gets attached to every container. And and again, all of that, I mean, I keep coming back to this idea of transactionality. Like you would want the system to be fully functional. You would want it to be, um, you would want it to be safe in order to like experiment with something. Go back, do the sampling with with agents. So really, a lot of the, a lot of the work is just like the engineering of this infrastructure.
31:02 Speaker 00 And it's like not as apparent, not as sexy as sort of like we're training this model. Here's the, um, but but I think this is where you can build like a, a bit of a lead.
31:13 Speaker 01 Yeah, I, when VCs talk about moats, the thing that I translated to in my head is what is the compounding advantage that you might be able to achieve?
31:21 Speaker 00 Yes.
31:21 Speaker 01 Right? It's not really a defensive mode. It's more just I'm ahead and by being ahead in some vector, it allows me to continue to move faster.
31:28 Speaker 00 That's right.
31:29 Speaker 01 And it sounds like this is a great example of that.
31:30 Speaker 00 True moats are often not obvious until like decades, perhaps many decades into into the company. Like you wouldn't know what like Netflix's mode, uh, but obviously they have one like Disney tried to like compete with them. And everyone was like, you know, down on on Netflix, but turned out they have a mode as this, uh, content production system that they that they built.
31:49 Speaker 01 So Amjad, you you started with the mission of like making it easier for people to learn to code. You've accomplished a lot of that. Now you're pushing the envelope of what it even means to code. I've got young kids. I want them to be productive creators in the world. What should I tell them to do? Should they learn to code? What does that even mean?
32:07 Speaker 00 Look, I think if you wanna go the like professional software developer route, I think getting a computer science degree and like learning fundamentals makes sense.
32:18 Speaker 00 But if you want to be a creator, if you want to be a journalist in this world, I don't think it's necessary anymore to learn a code in the more like traditional ways of learning code. I think you pick it up by osmosis almost like like go to Repland and you know and start using it. And at some point, you're gonna run into some issue where you're gonna have to look at code or you're gonna have to look at logs just by being resourceful, having to Google around and all that. You'll start picking it up. And by the way, this is how our generation kind of learned how to code.
32:46 Speaker 01 Yeah, when you were talking, I'm like, that's what I did.
32:49 Speaker 00 Yeah, and somehow it just became very industrial and very formal over the years. Like the way we made web apps is like you just like start a notepad with a, you know, HTML file. And now you have to like learn like webpack or whatever. Yeah, it's like, so I think the future of work is not really clear. What is you, you know, we can sort of like have some like, you know, idea of like where the world is headed. And so with with my children, I want them to have like as broad based knowledge as possible. I want them to be as generalist as possible. I want them to be as generative as possible, like being able to like create a lot of ideas because once the making of things gets easier,
33:32 Speaker 00 the ball like goes back to to to like how many ideas you can have. And so I wouldn't put learning a code as the top thing. I would put learn to make things, learn to make things with code, learn to make things with video, learn to make like anything with AI.
33:49 Speaker 02 And so what do you think happens with SaaS generally? If we're very soon able to say, create me a version of Google Calendar or clone DocuSign, what happens to SaaS, do you think?
34:00 Speaker 00 We have stories today of a lot of people replacing hundreds of thousands of dollars worth of SaaS with Repland. I heard a story from someone who, you know, they got quoted $150,000 for a piece of software. He went and made it in Repland, sold it to his employer for $32,000, cost him $400. Companies that have a platform developer community around it and plugin ecosystem and things like that, I think those are safe. You're not going to be able to vibe code Salesforce. I think the the vertical SaaS is in trouble. And I think it's my guess, it's already probably showing in some of the metrics.
34:45 Speaker 02 Okay, last question. What advice would you give to founders starting out right now?
34:51 Speaker 00 You know, the best advice is what you actually pointed out earlier is work on the edge of what's possible. Because one evolution of AI or models will make your business viable and sell only your first in market. I find it rare to see founders that are actually like sitting down trying to actually predict the future. And maybe that's something that was, you know, ill-advised in the past. But I think right now, trying to actually figure out where things are headed is a very, very important skill to have. So like make some prediction, figure out like how to create like a crappy product that would get better immediately as you switch the model.
35:35 Speaker 00 I mean, computer use is a great example.
35:38 Speaker 02 It's been an absolute pleasure, Amjad. Thank you so much for coming on.
35:40 Speaker 00 My pleasure. Thank you.
35:41 Speaker 02 See you next time. Thanks.
Previous John Jumper: AlphaFold and the Future of Science Next Chelsea Finn: Building Robots That Can Do Anything Page updated October 26, 2026 · transcript and summary are machine-generated
