The Exchanges, every show

Q Have you been surprised by how price sensitive people are around security and code reviews?

A Not in our segment. I would say in the engineering segment, um, And we, we have some engineers use Replit, but the 75% are non-engineers. Engineers are more price sensitive because they have a lot of options. They can use a lot of different products on the, on the market. Now, when, when you're an operations manager using Replit, And you just saved 10,000 dollars on a SaaS software. You've gained, you've saved another, you know, 200,000 dollars on, on headcount. And you're spending an additional thousand dollars to just make sure that the software is more secure. That's like a no brainer. The ROI has been a hundred fold for, for, for companies we work with. On the consumer side, there's more price sensitivity, especially if I'm an entrepreneur just dipping my toes, which is why we reduce the price on our core plan. So I think there's going to be, and you, you, you hinted at that earlier, there's going to be this different models for different use cases or different parts of your journey. If you're just starting out, you don't want to be hit with a thousand dollar bill. You want to be able to play around with 20, 30 dollars before you commit.

AI assessment note: “Not in our segment. I would say in the engineering segment”

Q Are there two, or are you, are they both running or only, you're only running? So you're not using any compaction from Claude?

A No, no, no, no. Just ours is better. And I think the thing about the primitives that the labs build, They tend to be very generic, and so I think every app will need to, uh, you could start with the generic firmatives, but for it to perform really well on your benchmark, on your users, every use case is going to be different, and so we care a lot about, like, you know, certain pieces of data around the architecture of the app. There are certain things that absolutely need to remember. Like, if it forgot that it has a database, like, that's catastrophic, right? So there are things that we really care about preserving, but then there are a lot of things that are like bugs and things that you solved. Actually, if you keep them in context, Asian performs worse because it is confused by, by the history of it. So you need to be very prudent about knowing what to delete and knowing what to, what to keep in the background. We have sort of a graph-like structure of understanding the different memories, and we do compaction on, um, actually there's like multiple layers. There's compaction, but there's also writing to long-term memory, and long-term memory today is marked down files. So repli.md, I'm sure you look at it sometimes. Repli.md is an example of that, but there are other files that the agent can write to, and you can also prompt it to write its own long-term memory. You can say…

AI assessment note: “No, no, no, no. Just ours is better.”

Q You have to weigh off. Is it worth that three month advantage? No.

A Right. You, you do. And that's why it makes it so, so freaking hard and why you have to change your mind all the time, right? Because there's a lead up time, but the most important thing is optionality. So in 20, 23, when we were training models, uh, we achieved better coding performance than the, than the state of the art models at the time, GPT 3.5. Right. Um, but then since Sonnet came out of later Opus, uh, The gap is closed by a lot, and they were doing, they were spending tens of billions of dollars, if not hundred billion dollars, making agents work. And that would, that would have been a dumb strategy for us to go and try to compete on that. But now I would say the opportunity opened up again for other reasons. Uh, the, the open source models are getting really good. And we're, you know, we're approaching a certain plateau in how good coding models could get. And so, uh, you can use your data to fine tune a model specifically for your use case. We, I don't know if you saw, but, um, Intercom yesterday talked about their new model that is better at customer support than the frontier models. And so maybe their model is going to be state of the art for three to six months, and maybe six months from now, the models will like zoom back ahead.

AI assessment note: “Right. You, you do. And that's why it makes it so, so freaking hard”

Q How do you think about maintenance in this case? You have ops teams building tools, you have entrepreneurs building tools. You gotta maintain these fuckers. This is hard enough running a business. Are you going to maintain this now?

A This is where Replit shines. And, and, you know, if you talk to Jason or some of our other customers, um, Replit goes way further than any other Vibe coding product on creating more maintainable software. For example, and part of the reason Replit has been slightly more expensive than, than others is that we do a code review for every, for, for every, you know, code change that we make. We spent a lot of tokens on maintenance as much as we spent on creating that software. Replit also has a built-in tester. So if you enable all the power features, whenever you, whenever the agent writes code, goes into a testing phase, spins up a browser, tests everything in the, in the app, Goes into a code review session, reviews that, kicks it back to the coding agent, gives it feedback, you know, the test failed here, the code review is not good. And people enjoy looking at the code review agent because it's kind of a dick. It's like, this looks like AI generated slop. It'll actually say that. And then, and then it goes back. We're also building, um, agents that are sitting in production software. So we already have security agents right now that are sitting in enterprise Deployments and are monitoring activity and, uh, and they're monitoring packages, monitoring for supply chain attacks. And so the thing about AI, any problem AI creates, there's more AI that you can build to solve that prob…

AI assessment note: “Replit goes way further than any other Vibe coding product on creating more maintainable software.”

Q When you look at your model usage today, I've heard you say before, when I was obviously listening to your prior shows, that you have a preference for Anthropic. What does the model usage look like across different providers today?

A Yeah. So Anthropic, the, you know, has been the sort of workhorse for, for over, over a year right, right now. It's, it's like the, the, the core agent loop because it can run for a long time coherently. But the few things have changed. Google's Gemini's models have, um, are the best at price performance, for example. You know, given their price, where do they sit on the period of frontier, right? And so for tasks, for example, like tasks like code search, we might create a sub agent, uh, that is, that is cheaper and has good enough performance. Uh, and we offload that from the main core loop, right? So we now we use, and I wrote this thesis back in 22, I call it the society of models. Now we use models from every provider. Actually, at some point we were sending more tokens to Google than we were sending Anthropic, despite Anthropic being that kind of the core workhorse. And so there's this concept of agent labs, right? We talk about AI labs, but there's agent labs, you know, us cursor, some of these other companies. Our goal is to start with the user problem. What are we trying to fix? What are we trying to build? And walk back to the technology and use whatever model we need to use. In some cases, we build our own models.

AI assessment note: “Anthropic... has been the sort of workhorse... Now we use models from every provider.”

Q Our company is going to be so much smaller in the future. When you look at the capabilities of individual people, do you buy the ideological Silicon Valley? Oh, you know, we're just going to do more and we're going to be so much more capable. What are we actually like? No, we will have dramatically smaller engineering teams.

A I see, I see both. So, so Jason is someone who's, who wants to work with a very lean team is doing better, more than When we had, when he had people on staff. Yesterday I met an entrepreneur at a conference in D.C. that's using Replit. I think he sells board games online. And he says, it's been so transformative on my business. We've saved so much money on, on SaaS. We're selling more, uh, that I decided to use the increased revenue and efficiency to hire more people. He hired eight more people. Uh, there's a customer case study we actually published on our site, Firecrown Media. Um, like a sixty million dollar media company that owns magazines, different properties, and they've been so successful using Repplet for marketing automations and all sorts of things like that, that decided to hire more people that know how to do vibe coding in order to, uh, to, to, to, to, to sell more and do more and build more. So it depends really on the kind of company, but we see companies that were like, we want to get leaner and we want less people. And I think it'll come down to the entrepreneur, the level of ambition, how they want to run their company. For, for, for us at Rap Lit, I think what we want to try to eliminate is, uh, or reduce is a lot of supporting roles. We want builders, right? We want, uh, we want builders and we want salespeople because I think salespeople, it's like people…

AI assessment note: “I see, I see both. So, so Jason is someone who's, who wants to work with a very lean team”

Q on them, makes you more copyable, et cetera, you know, Business attributes. So one of the things that the people who listen to our podcast a lot are entrepreneurs. You're navigating a business challenge that they need to learn from. So how do you think about that? What are the things that you're doing to build unique and enduring business? What are the kinds of considerations that go into that?

A First thing I would say is user obsession. There's a massive platform called LMS, right? That platform is useful on its own. People can use it directly, but for you to build a very valuable business on top of what kind of value you're creating, One is, if you're really obsessive over certain types of users, OpenAI can't build for every type of user in the world, so pick a user that you know very well, hopefully it's you. You've had that experience, you're, you're a salesperson, you were gonna build the best L.M. for salespeople, right? That goes into UI, UX, all of that. Now, second, and this, this might comment, I'll go a little deeper on the comment about the environment around it. So, You know, I said that people in, uh, using Replit should feel like they're in safe environment. Uh, it, it should feel like anything is reversible. Now that turns out to be a huge technical challenge. Now you could easily kind of hack around it or say, you know, we'll just, you know, restart the app or do whatever. But what we did is we built a transactional file system that allows undo. Like we have proprietary file system that we spent two years building that every action is like an immutable Part of a ledger that anything can be rolled back and roll forward. That ends up being like a time travel system. And now we're finding more use cases for it. For example, one way to make, um, models bet…

AI assessment note: “what we did is we built a transactional file system that allows undo”

Q So now we're moving to our rapid fire. That doesn't mean you have to answer rapid fire. There's this questions we, we ask all of our guests. Um, is there a movie song or book that fills you with optimism for the future?

A You know, a book that I think is, is, is amazing that it pulls together everything that I love. I really love learning about theory of consciousness, theory of mind, the philosophy of that. I like computers and math and physics and all of that. I like the human aspect of human stories. Um, I am a strange loop by Douglas, uh, Hoff Satter is, is an amazing book. It's partly about his, a book about his wife dying and their kids and, and how, how, His personal story in that. There's also a book about him struggling with the concept of soul, and kind of trying to understand it in a, like, more secular way, in a, like, mathematical way. It, it goes through vignettes about, like, Gödel's incompleteness theorem, or what that says about, um, about how consciousness is potentially, like, a loop, and it's, it's, it's an amazing book that tells you that, like, really everything is interconnected, and we can't be, Technical or innovators without actually caring about all the other things that happens in our world.

AI assessment note: “I am a strange loop by Douglas, uh, Hoff Satter is, is an amazing book.”

Q And is, are these vibe coding programs or these bespoke programs that people are building with prompts? Are they in production or are they mostly hobbies that people fool around with?

A Depends on, um, first bucket is, is more hobby, personal life. Second bucket entrepreneurs. As you know, most startups die, so most startup ideas don't make it to fruition. The 10% of startups that are small businesses that get off the ground, they get the most value out of Replit. Uh, and some of them are in production now. Um, you know, I've, I've talked about a lot of these stories, but, you know, for example, we, we have this, uh, creator, his name is John Chaney. Uh, he's a serial entrepreneur. Used to take him many months and hundreds of thousands of dollars to build applications, and now he can spin up a business. And get to million dollar run rates in, in a matter of, of weeks. Obviously he has experience. Like he, he knows the formula of what it means to be an entrepreneur, but people can learn that over time. And in terms of the, um, enterprise, um, you know, we have, for example, Zillow. The CEO of Zillow recently on New York Times Dealbook talked about how everyone at Zillow is using Replit to accelerate product innovation, because product innovation no longer depends on engineers. You can have product managers do the entire iteration, getting user feedback, even without going to the engineers. So it just like increases, we have Duolingo, Um, a bunch of these customers that are really focused on innovating, building their second, third product, uh, that are now usin…

AI assessment note: “first bucket is, is more hobby... And some of them are in production now.”

Q they said founders in public, AI is writing 99% of our code. In six months, we won't need any engineers. Founders in the DMs, uh, does anyone know a good React developer? 30,000 dollar bonus. And I will name my firstborn son after you. So Can you explain that disconnect between this view that engineers are going away and this still like very intense demand for engineers in the market?

A I never made the point that engineers would go away. I make the point that entrepreneurs can start businesses without needing engineers and that we already see that. We already see, you know, I meet YC companies and, uh, Y Combinator is the most, uh, prestigious startup accelerator in the, in the world, Bay Area. And In the past Y Combinator would encourage you to go get a technical co-founder. But like we said, there's so many people with amazing ideas that don't have a technical co-founder. And so they're starting to get into YC. And what they tell us is we're just going to build this thing on replet. We're going to see how far we can get. And they often got, get really, really far. Now, if you're building a venture scale company and you want to like get to hundreds of millions of dollars of revenue and you want to, you know, become 1,000,000,010 billion, a hundred billion dollar company, you're going to have to hire engineers. But if you're trying to build, um, a company that creates a really great living for you, even, you know, you can, Potentially get rich from it. You, I think we're almost there where you can do it on your own without any developers. And so when I'm talking, I'm talking to our audience.

AI assessment note: “if you're building a venture scale company... you're going to have to hire engineers”

Q Okay, I do want to ask you about something that you didn't mention when you looked at the different factors for why prices might not be going down. There might be investor pressure, uh, there might have been this equilibrium reached, or is it possible that these models have just gotten so big and expensive to run that the fundamental economics of AI are just not working? So explain why.

A Um, uh, You can surmise the bigness of the models based on speed, token, token throughput. It's not perfect, but, but if you remember GPT, 4.5, GPT, 4.5, uh, was an experimental model from OpenAI. It was the idea, let's train a trillion parameter dense model, meaning it is not sparse, meaning all the token, all the neurons are activated on every request. And it was so slow. It's really hard to run these things. The new models, even when they're big, they're sparse models. They're called MOE, mixture of experts. So in every request, there's a router layer that takes it to the expert part of the circuit in order to answer that question. So, you know, there are models with trillion parameters But any given request is thirty-two billion active, and that's like a kind of small model. Um, and what we're seeing based on speed and things like that, it's actually probably the models are getting more efficient. I mean, Deep Seek showed that the models are getting more efficient, and if, you know, Deep Seek open source was able to make it, you better believe that the labs are also getting more efficient.

AI assessment note: “The new models, even when they're big, they're sparse models.”

Q First your own, then a little bit of YC, then a little bit of third party investors. It's not like this was an overnight success at all, and so I, I'm curious, did you ever have a moment where you were like, we are not gonna make it, and I'm not good enough, and this company's not gonna work out?

A Yeah, yeah, I mean, in, uh, 20, 24, early 24, so last year, um, the company had not been growing all that much. We're actually, we're kind of hit a plateau and started decreasing a little bit, so we were in this awkward Position where we make it easy enough to get started, but not enough to get you all the way to building a business or a real piece of software. And because our focus is on getting everyone to make applications, anyone with a dream without any skill or background in computer science should be able to come to Replit and make an app that was always the kind of the core of the mission. We're also not very good for professional developers because they needed all these tools that we didn't have. And so we're in this like awkward middle and not making any revenue and have this Really large staff, 130 people, had raised two hundred million dollars, and as you know, you know, you, uh, you talk about it in the book, like, you know, there are businesses that you could build that provide a lot of, uh, you know, profits and good living for yourself, and you have complete, you know, control of what you do, but once you take on venture capital, there's almost a contract, uh, you're implicitly kind of getting into, which is, I'm gonna build a billion dollar company. This is this, I'm going to build something really scalable, really big. And I take that responsibility seriously …

AI assessment note: “Yeah, yeah, I mean, in, uh, 20, 24, early 24, so last year”

Q of bounce around. They're using cursor one week, and then cloud code the next week, and then they're bouncing to this one. Um, what are you seeing, and what do you think's happening, and do you think there's any, uh, any merit to that fear in the market, or on X right now, that maybe the AI revenues that we're seeing are not as sticky as previous SaaS installation cycles?

A It's certainly suspicious. Uh, I mean, it, anyone who tells you differently is sort of, and I was a little worried about our growth. Like, is it, is it going to go away? What's going on? Like, this is like a little too fast. Um, and I think, uh, you know, we really start focusing on the, on the metrics that really matter, which is the revenue quality retention, things that more advanced companies are thinking about, like net dollar retention. We actually have above hundred percent net dollar attention, even for our consumer plan, which is probably unheard of. But, um, uh, you know, I, I think that, uh, I think maybe ARR is becoming a bit of a vanity metric, like how fast you can grow the ARR. I think you want to make sure you're providing an awesome service and also make sure that people just want to stay on your platform. I think, especially on the VS Code Wars, the switching costs being so low, Is, is going to be a challenge for them. You can go from Copilot to Windsurf to Cursor in like five minutes. Same project. Quite like it's one click. You can open the project. And so they're going to have to find a way to differentiate. The great thing about Replit, we provide this really great convenience. And so if ain't broke, don't fix it. You deploy something on Replit and people, people really tend to stay.

AI assessment note: “It's certainly suspicious. Uh, I mean, it, anyone who tells you differently”

Q Well, we got to ask about the Vercel drama. We talked to Guillermo earlier this show. What's going on there? What's your take? Can you break it down for us?

A Well, you know, I, uh, someone asked me, um, do you, do you plan on supporting Next.js and Repl.it agents? I said, no, because, uh, we want to bet on something like Vite that is open, like community builds, uh, and the way Next.js has been trending is that it's optimized for Vercel. They disagree, but it is actually not a controversial statement in the larger community. There's a project called Open Next that makes it so that, you know, uh, Next.js is deployable on more platform. Matt, the CEO of Netlify, commented, uh, agreeing with me. The larger community feels like, you know, Next.js is Is perhaps more tied to Vercel than they would like to admit. And I think they overreacted to my comments and, and, and they, and they strized my, my comments. I think it would have been, uh, uh, less of a, a story if they, you know, they didn't overreact to it. And I look, uh, uh, Vercel is an amazing product. It's an amazing brand. They've done a great job managing a lot of open source projects, not just Next.js. And for me, water on their bridge. Like, you know, I, uh, You know, I, I think, you know, emotions get, get high on Twitter and you start calling and, you know, I wish them really the best.

AI assessment note: “the way Next.js has been trending is that it's optimized for Vercel.”

Q Replit or, or REPL. What, what is REPL in, in REPLit?

A You know, uh, for example, a DOS or like a UNIX command line is a form of a REPL. So, uh, it comes from, um, Lisp actually in the 19 fifties at MIT, um, where, uh, you can create, uh, the simplest programming environment with like one command, which is read, eval, print, loop. So read reads the command from the command line. Eval, evaluates it, like a lisp string, uh, print prints out the results, and then goes back to the start. So this is like the simplest IDE in the world, right? Um, and so when I was building, so it, you know, the, the way sort of Repl.it was born, I, I was going to school in, uh, in Jordan, uh, studying computer science and, um, I, I didn't have a laptop. So every time I would want to do some homework, I would have to reinstall the development environment. And that was like the, you know, it was so annoying, right? You have to download gigabytes of software and packages and, and there's always something that goes wrong. Um, and so I was like, you know, I'm, I'm using everything in the browser. At the time, Google was putting docs and Gmail in the browser.

AI assessment note: “one command, which is read, eval, print, loop.”

Q And maybe explain to folks, you know, why, why are you using GitHub? Why are you pushing into GitHub?

A For sure. So, uh, GitHub is a, um, uh, version control system and a collaboration system. So if you want to, uh, put something open source so that other people can contribute to it, other people can use it and open issues, uh, for you, um, and tell you, like, here's what's missing, here's what's working and what's not working. Uh, you can, you can use GitHub for that. And so you can think about a Replit. Replit is the sort of real-time interface for creating applications. You can use Replit for deployment, for editing, even for collaboration, but it's more real-time. You can think about GitHub as the more long-term sort of storage for your app. So if you want to, uh, keep it version there, they have a lot of really interesting tools to help you manage your application. Uh, people can download your code. And again, there's this feature called per requests where people in open source or people on your team can send you, can send you a per request. Uh, this is true of GitHub and other places like GitLab. You can do all of that from Replit. So here we are, I think, in like, you know, uh, you know, 15, 20 minutes, we were able to build an application, we deployed it, we did a data migration, we moved it to SQL, we pushed it to GitHub, and we're on our way. Like, if this was a real project, I've already made real progress, uh, towards my Scenic Drives startup. Which, if you built tha…

AI assessment note: “GitHub is a, um, uh, version control system and a collaboration system.”

Q You have made a huge bet on AI as a company. Can you just talk about Ghostwriter and how it came about and the investment in this area?

A Yeah, throughout my career working on code that handles code, right? Whether it's at Codecademy for teaching programming, whether it's my own project, whether it's React and building the runtime around React Native, I always felt like our tools that were handling code, whether it's like compiling it, parsing it, minifying it, all that stuff. We're very kind of rigid and very laborious. Um, in a lot of ways you're building sort of a classical intelligence system, very algorithmic, um, and a lot of heuristics and, and all of that. And I always thought that you can probably apply machine learning to it. And I started reading around whether anyone had done it. There was this seminal paper in 2012 called on the naturalness of software. And basically a bunch of researchers try to apply NLP to code. And what they found is that actually code Can be modeled like any language. That's why they call it naturalness is because hey, code is kind of repetitive, like language. You can do things like Ingram, basic Ingram model can actually start to generate code that is compilable. That had a huge impact on me. And every year or two while starting Replit, I would go look at the state of the art on ML on code and nothing ever really worked all that well up until GPT-II. And you can like take GPT-II and fine tune it and like try to make it write code and was like kind of okay. But obviously GPT-II…

AI assessment note: “And that's when we started building what became Ghostwriter.”

Q code autocomplete with local context has become as widely adopted as it has. Like your tool and co-pilot are sort of a handful of tools that people like very much accept have changed workflow. Maybe if we just project out a little bit, where do you think that the like next big leaps in productivity in software development come from in terms of AI? What else is going to happen?

A It's an amazing tool, but I think it's still very primitive, and we haven't fully explored the full capabilities of even GPT, but also transformer models applied to code in general. There's a lot to do. I think there's a lot to build on the sort of the layer just around the models. So they're basically how to give models better context, how to give them tools, the ability for the model to actually Be able to go out and read a file, install a package, evaluate code, the ability for these models to be more agentic and be able to kind of write entire features, uh, by themselves. And so that that's all the work just around the model itself with the model in the center. And then there's a lot of work on the models themselves that haven't been, you know, fully explored. The way we train models today is we just give it a large corpus of, of, of code. But you can imagine different ways of training models, and we've played around with different techniques at Replit. There are ways to make the models, give models more intuition about how code execution happens as opposed to just static code, because you're training them on static code. They understand the structure and syntax of code, but they don't really understand the semantics of code. So you can imagine a way to train Models where you're not only feeding them code, but you're also feeding them the results of the evaluation of that c…

AI assessment note: “how to give models better context, how to give them tools”

Q Can you talk about the decision to, um, train models versus use existing APIs? Like, was it just a latency thing? How did that happen?

A So when Microsoft did it, like, they took GPT-III, they distilled it down, and they hosted it on their own infrastructure. They did a bunch of, uh, caching, a lot of things like that. If you're someone who's just using the OpenAI API, you couldn't do any of that. And it was very expensive. It was very slow. Um, and even now, you know, there's no completion models. Now all the models are chat models. You know, they're not going to be releasing any completion model going forward. And so if you want to build a completion based product, it's actually fairly difficult to do it using commercial APIs. And we felt like we can, we can make a model that's both cheap and fast and good enough for that autocomplete use case. And we found that the three billion parameter size is kind of the nice sweet spot where you get enough raw IQ from it, because like one billion felt kind of too dumb in a lot of ways. And it was going to be cheap enough and fast enough to host and be able to do a fast inference from. Um, and it was a time after Yeah, the chinchilla paper first came out and LLAMA had been announced and the idea of just training them longer had just been in the air and, and we're like, okay, what if we, we apply all these open source tokens on a smaller model? And then we applied some more tokens from replets data that gave us a 50% improvement over the model that we open source. So we op…

AI assessment note: “we felt like we can, we can make a model that's both cheap and fast”

Q Because you value their creativity. Like why is that? Unpack that logic for me.

A Because Their leverage will be, will be huge. John Carmack is one of the best software engineers in the world. He, he made Doom. He worked on VR and now he's working on AI. He's still limited by how fast he types. Right. Which is like a crazy thought, right? Like if he's not, because, because ultimately you have to type the code out, right? If the AI is actually better at typing the code out, and he's actually manipulating, uh, things at a higher level of abstraction, and he's talking to not just one or two, maybe 1000 AIs doing work for him, he's going to have an insanely higher leverage on the world. And so I think The best, most creative people will have a lot of the menial work automated for them. They wouldn't have to type a unit test anymore. The AI would do that, for example. And so a lot of their time is spent just doing more things and Um, you know, think about Elon, right? Elon is like starting companies left and right, and that's probably not the end of it. Like he'll probably start more companies in the future, right? You know, part of the reason is because he's sort of like operating at a much higher level because he's built a set of very competent people around him. Uh, and like they follow him from one company to another, whether it's like on the finance and legal side, Or it's on these sort of like engineering side. A lot of that could be AI and could be automat…

AI assessment note: “Because Their leverage will be, will be huge.”

Q For example, how are you provocative in phrasing? Cause I find everyone's like, we look for a team player who's good at communication and shared interests and alignment. And you're like, Fuck off.

A Yeah. Zuck said, um, said something very, uh, very interesting, uh, about like move fast and break things. I think a journalist was asking him about it. And he said, like one heuristic to know whether you're saying something meaningful is that like reasonable people can disagree with it. And the opposite of it is also reasonable. So like move fast and break things could be like move slow and don't break anything. And that could be like a value at IBM. Right. Uh, you know, or steady or slow, you know, you can phrase it like a little more charitably. And so if something where the opposite of it doesn't make sense, or nobody would actually, it's, it's not a value anyone would hold, they're actually not saying anything. Uh, it cancels each other out. And so to be provocative, um, you need to say something that people will disagree with, and the opposite of is also somewhat reasonable. One of our values, for example, is like seek pain. And so the idea behind seek pain is that there's a lot of painful things in, in building a startup. Talking to customers is actually extremely painful. When your product is not working or there's like some fundamental, you know, lack of product market fit or some issue you're dealing with, you know, founders and entrepreneurs typically don't want to face that. And facing that is painful. At Replit, we've, you know, had some things in the past, uh, tha…

AI assessment note: “One of our values, for example, is like seek pain.”

Q Why, why do you think we're at the new start now? I love Marc Andreessen's statement that there's no such thing as a bad idea, only the bad time. And it's like, timing is everything I've learned in investing. Why is now the time? Because everyone always says now is the time. Why is now the time?

A Well, I like, we don't have to speculate, right? Like we actually can look at the results. Andre Karpofi was the head of AI at Tesla, built a self-driving team there. He now, I think, back at OpenAI. A few months ago, he tweeted that AI is writing 80% of his code today. 80% of his code. Is written by a machine. That is unprecedented. On Replit, we see that 30 or 50 to 50% of code for Ghost Rider or AI, Ghost Rider users is written by the AI. The Ghost Rider users report that tasks sometimes are cut in half in terms of, like, how much time they needed. There was a study actually done on Copilot users, GitHub's Copilot, Uh, that showed that programmers are 55% more productive. So, okay, this is not a 10 X, right? But it is a start of something. We're really at the early innings of this technology. This technology being large language models and transformer based models. And so it's only been, uh, you know,

AI assessment note: “we don't have to speculate, right? Like we actually can look at the results.”

Q word fun. Um, and after I learned to code, I saw it in a different light. And I think that's really important is this idea where it's like, how do you make coding fun? So what are your thoughts there? How have you been able to design Replit to actually enhance people's creativity and make them want to come back and make them see this skill in a new light?

A So I think most things that are fun tend to be devoid of a lot of drudgerous routine work, right? You know, when, when you're playing a video game, you're not like building the video game every time or setting up the TV or, or, or doing some like rote IT task, right? When you are, um, doing a sports hobby, you're in the flow, you're doing the thing you're excited about doing. The problem with coding is that a lot of the maintenance around the development environment and the packages and the integration of all the different components, Was the thing that engineers were spending most of their time doing. The moments they were coding, they were an absolute bliss, but those moments were actually very little in terms of the, if you think about the pie charts of what it meant to work as a programmer. So the first thing that Repli did is remove the need to do all this setup. That in itself made programming a lot more, uh, fun. And then add the collaborative aspect. Like a lot of what we find fun in life has to do with other people. We're just social animals, right? And so like, if I can share my program with you with just a link, that's really fun. That's what makes Figma fun. That's what makes Any other collaborative tool fun is that I can just like send you a link and you're, you're in there with me, or you can play it and try it out. And then finally, there's like a lot of like exp…

AI assessment note: “So the first thing that Repli did is remove the need to do all this setup.”

Q That's awesome. Let's move on to the next one. Someone has used Replit to build over 50 projects. Now you can use your own definition of project here. Um, these don't need to be full blown startups, but have you seen someone build that magnitude of projects on Replit?

A Yeah, I, I think that's, um, that's fairly reasonable number of projects. So there's a six-year-old developer. His name is Ray Han. He's one of our most prolific programmers. He actually built a bunch of interesting projects where he reverse engineered How Replit works. And he built, he built an unofficial API. One of his unofficial APIs is a security program that searches people's repls to find discord tokens. So a lot of people build discord bots on Replit and they copy and paste the tokens in clear text as opposed to putting them in our encrypted service, the secrets manager. And so he would Find those tokens. He would invalidate them because discord has service to invalidate those. And then he would send them a notification. He would say like, Hey, like we found that you've exposed your token. Uh, so that's one example of a project he made. He w he was, and Scylla is one of the most prolific, uh, bounty hunters. And he built one of our earliest bounties, which was A startup that wanted to build like a stable diffusion based t-shirt generator. So you would share a t-shirt based on a prompt and then get it printed and sent to you. Like there isn't one week where I don't see Rayhan producing like a new piece of software.

AI assessment note: “Yeah, I, I think that's, um, that's fairly reasonable number of projects.”

Q also a decision that you made, which was to build Ghostwriter on top of your own models. So something like a co-pilot is built on top of GPT-III to my knowledge, and that's a decision to be built off another platform, but you went a different route. So can you speak a little bit more to how you made that decision and what kind of inputs led to that output?

A Well, first of all, how crazy it is that, uh, that Microsoft had a dependent on another company, whereas Replit built our own thing. There are like multiple ways to answer this. One of them is UX. UX is inherently inseparable from the infrastructure for how a product works. I think most people think as the, they're separate things, but if you're serious about making products, you know, uh, you know, famous, um, Alan Kaye quote to Steve Jobs, he told Steve Jobs, if you're serious about making software, you have to make hardware. Um, and that's why like Apple's this full stack company is because they think about everything from the transistor to the touch, right? And so I think for us, it was like, just. Like if this is going to be a core interaction with our platform, we have to be able to optimize it and we have to get the latency down to the point that we, we like, we feel it's going to be a really great user experience. And we weren't able to really get that when we're hitting something over an API. Cause the latency will be all over the place. We can get the caching ride. We can get the, um, you know, location, right? We didn't have control about any of these things. That's a huge downside of being a consumer of a mere API. And then, uh, the other part is a strategic part, which is like, if you think, if you believe that this is a primary platform shift and this is going to …

AI assessment note: “if this is going to be a core interaction... we have to get latency down”

Q then also vet their clients and say, okay, this is, You know, a good project. This is worthwhile for us to introduce into this marketplace. And then they match people. And then of course those marketplaces take some commission for doing so. And so since many of these already exist, like how do you see bounties as an ecosystem differentiating or maybe providing something new to people within the community?

A I think it's already differentiated. And the reason it's already differentiated is because the development environment is built into the system. If Uber was a marketplace to connect you to people and then they have to go get their own car, right? Like, so you have to go meet someone at a coffee shop and then you and them go get a car. That's an absurd example, but that's what happens at Upwork or some of these marketplaces. Like I asked for a piece of software. And then you go make it in something that I don't know. And then you send me a zip file. And then what do I do with that? Right. And Replit, I just sent you a link, a link to a computer that's running, that's running your application. That's like fundamental innovation on top of that. And then like all the services just being integrated right there. Like you're opening out AI API key that the cloud runtime, like all that stuff, the database, we just added, uh, Postgres support. Just Repli being this complete platform, just like makes this process a lot more efficient.

AI assessment note: “the reason it's already differentiated is because the development environment is built into the system.”

Q word fun. Um, and after I learned to code, I saw it in a different light. And I think that's really important is this idea where it's like, how do you make coding fun? So what are your thoughts there? How have you been able to design Replit to actually enhance people's creativity and make them want to come back and make them see this skill in a new light?

A So I think most things that are fun tend to be devoid of a lot of drudgerous routine work, right? You know, when, when you're playing a video game, you're not like building the video game every time or setting up the TV or, or, or doing some like rote IT task, right? When you are, um, doing a sports hobby, you're in the flow, you're doing the thing you're excited about doing. The problem with coding is that a lot of the maintenance around the development environment and the packages and the integration of all the different components, Was the thing that engineers were spending most of their time doing. The moments they were coding, they were an absolute bliss, but those moments were actually very little in terms of the, if you think about the pie charts of what it meant to work as a programmer. So the first thing that Repli did is remove the need to do all this setup. That in itself made programming a lot more, uh, fun. And then add the collaborative aspect. Like a lot of what we find fun in life has to do with other people. We're just social animals, right? And so like, if I can share my program with you with just a link, that's really fun. That's what makes Figma fun. That's what makes Any other collaborative tool fun is that I can just like send you a link and you're, you're in there with me, or you can play it and try it out. And then finally, there's like a lot of like exp…

AI assessment note: “the first thing that Repli did is remove the need to do all this setup.”

Q have been built with AI. So those are two examples of how this technology not only, you know, maybe made things faster, but actually made things possible. If that makes sense, things that weren't prior available. So have you seen any glimpses of that or any thoughts around just like having, you know, we could position it as somewhat of a superpower, having that superpower accessible. What does that change?

A Man, this is such a brilliant question. Like, I actually haven't thought about it as much, and I think I should. Like, uh, you know, your example with chess also happened with Go, you know, with the Lisey Doll game. Do you remember that move? That basically what happened is the AI made a strange move that made it, gave it a disadvantage in the near term and a huge advantage in the long term. And they were so confused about it because it's almost like alien intelligence intruding on this thousand year old game, right? And, and producing this fundamentally novel move. So I don't think we've seen that entirely in, in programming yet, but I'd be definitely on the lookout for that. What I would say we've seen is that tasks that were previously would require A ton of work, like a ton of insane amount of laborious work getting done like that. You know, for example, the parsing question. GPT-III is incredibly good at parsing. If you give it a malformed JSON, it will still parse it. Writing parsers is one of the hardest things you can do in programming. Writing parsers in GPT-III is one of the easiest thing you could do. You could spend 15 minutes in the OpenAI Playground. And so really that goes from a task That requires hours and maybe days and weeks of building and testing to something that takes a 15 minute. And so that's a fundamental phase shift in, in how we do things. That's act…

AI assessment note: “I don't think we've seen that entirely in, in programming yet, but I'd”

Q Some people ask, hey, how should I think about X versus other platforms? You know, maybe their customers aren't on X, et cetera. How do you think about that?

A Well, I, I think, uh, X is very good for the sort of Silicon Valley, for tech, for kind of the in-group messaging, for maybe one hop, Outside of Silicon Valley, which is kind of the early adopter, if you're targeting early adopters, uh, you know, so I would also cluster Hacker News within that as well. Um, but it, it actually, once you want to go to late adopters, or like early majority, if you think about this kind of bell curve of adoption, uh, Twitter becomes actually not very impactful there. It can still be very important where Um, like, you know, just this morning, uh, journalists reached out to me to, to ask to talk to me about a tweet that I made. About some of the stuff that's happening with Anthropic. Uh, and so like Twitter can still be a place where it is, it is the start of stories. And that's another thing, you know, back to the conversation of why this is important is the landscape shifted so much that journalists report on Twitter. So it's kind of starts on Twitter. So Twitter's influence is amazing, but it tends to be elite kind of inside group influence. I think Instagram, Facebook still, You know, YouTube is where you get to the early majority. And I think I was late to, to YouTube, uh, to sort of to Instagram. Like, um, you know, I just started focusing on it and like maybe, uh, February or something like that.

AI assessment note: “Instagram, Facebook still, You know, YouTube is where you get to the early majority.”

Q You have to weigh off. Is it worth that three month advantage? No.

A Right. You, you do. And that's why it makes it so, so freaking hard and why you have to change your mind all the time, right? Because there's a lead up time, but the most important thing is optionality. So in 20, 23, when we were training models, uh, we achieved better coding performance than the, than the state of the art models at the time, GPT 3.5. Right. Um, but then since Sonnet came out of later Opus, uh, The gap is closed by a lot, and they were doing, they were spending tens of billions of dollars, if not hundred billion dollars, making agents work. And that would, that would have been a dumb strategy for us to go and try to compete on that. But now I would say the opportunity opened up again for other reasons. Uh, the, the open source models are getting really good. And we're, you know, we're approaching a certain plateau in how good coding models could get. And so, uh, you can use your data to fine tune a model specifically for your use case. We, I don't know if you saw, but, um, Intercom yesterday talked about their new model that is better at customer support than the frontier models. And so maybe their model is going to be state of the art for three to six months, and maybe six months from now, the models will like zoom back ahead.

AI assessment note: “Right. You, you do. And that's why it makes it so, so freaking hard”