Y Combinator Startup Podcast
00:00:00 · Amjad Masad, Paul Graham
[00:00:00] S0: I think with AI, everyone will have to go through this period of denial and then depression and then eventually acceptance. [00:00:10] S1: Replit's like this iceberg where there's this visible bit above the water that's like, look. You just tell it in English what to do. And it writes your code, then underneath, there's all this stuff for deployment and sandbox. [00:00:19] S0: I don't think we describe ourselves as Vibe coding now. We think Replit is the platform for the transformation for our two self driving companies. [00:00:27] S1: More than other companies, though, I think you guys are... Have been living in the future. If you had to say, like, what's the real DNA Mhmm. Of Replit [00:00:36] S0: It is living in the future, for sure. [00:00:47] S1: So you guys basically invented vibe coding. Right? And when you did it, like in 2024, this was, which it sounds it's like only two years ago, but it seems so long ago. [00:01:00] S0: Seems like it technically. [00:01:01] S1: Technologically. Yeah. How did you even know this was gonna work? Because it's obvious now that it did work, right? But at the time it seemed crazy. It seemed crazy to me. [00:01:09] S0: For us, it was incremental. It wasn't... Like for the rest of the world, I felt like a big departure from how people did things. But we've always [00:01:21] S1: You were already writing more and more of the code. [00:01:23] S0: Well, no, it it was was sort of like the YC kind of ethos of looking at what your users are doing and and and just... That will give you ideas on what to build.
00:01:23 · Amjad Masad, Paul Graham
[00:01:23] S0: Well, no, it it was was sort of like the YC kind of ethos of looking at what your users are doing and and and just... That will give you ideas on what to build. [00:01:34] S1: Because you were already doing auto complete. [00:01:36] S0: We're doing auto complete and not [00:01:37] S1: doing So how much of the code were you writing before you went to like full vibe coding where you tell the thing in English what to do? Where you're just generating code with auto complete? How much of the code were you already generating then? [00:01:47] S0: I mean, could probably say like 50, 60%. [00:01:50] S1: Really? So that's interesting. Even before vibe coding, you were already 60% vibe coding, And that was how you knew it would work. [00:01:57] S0: But the user because [00:01:58] S1: you interface thought we can make the 60%, why not make the rest? [00:02:01] S0: Well, kind of. I think the thing that was very enlightening is watching some of our users code. And so some users were doing 100% of their code using AI, but the way [00:02:17] S1: they did... No, you're talking about once you released five coding. [00:02:20] S0: Yeah. Before, before that, before that, and I'll tell you. [00:02:22] S1: Let me finish. [00:02:23] S0: Okay. So they would open ChatGPT in one window, and they would open Replit in another window. This is someone who's never coded before. [00:02:34] S1: How could you tell that they were pasting code out of ChatGPT? [00:02:37] S0: I invited some users to the office. [00:02:39] S1: But you would literally watch them. [00:02:40] S0: And we watched them.
00:02:40 · Amjad Masad, Paul Graham
[00:02:40] S0: And we watched them. [00:02:41] S1: Okay. [00:02:42] S0: How do you do this? And basically, they would copy the code, put it into Replit. Replit would produce some error. They would copy the error, put it into Type T, and then copy the fix, put it in... And it was such a manual process, and I looked at it. [00:02:57] S1: All we have to do is make an API, right? [00:02:59] S0: Well, we have to automate the entire process. It wasn't obvious, it wasn't easy because back then AI models would do a couple of tool calls. Oh, you know what's different? [00:03:10] S1: The difference is it's not just you have to automate the back and forth. You have to understand the errors, right? [00:03:16] S0: Exactly. Yeah, that's hard. Yeah, and you have to loop and the terminal condition should be that the program is working as opposed to... [00:03:24] S1: Yeah, that's still true. [00:03:26] S0: Right, as opposed to kind of we'll do one turn, right? Because that's how ChatGepti sort of worked, that's how most AI tools worked. And so can you [00:03:36] S1: loop indefinitely? Go to completion. Yes. Right? That's the new thing. That's what Claude does. [00:03:42] S0: Yes. So that's what we demonstrated. That's what Replit first demonstrated is... So we demonstrated a couple of things. One is that you can run LLMs to completion. [00:03:54] S1: To- Which I bet was news to the model companies.
00:03:54 · Paul Graham, Amjad Masad
[00:03:54] S1: To- Which I bet was news to the model companies. [00:03:57] S0: They did, they actually reached out to us and they were like, We're so surprised that our explanation works. How did you make it work? And the crazy thing, Paul, is that the harness circuit 24 is 10 times more complex than the harness today because the models were less capable. Yeah. So as models get more capable, the engineering you have to do to make them become agents is less and less so. [00:04:23] S1: Does that mean it's easier for people to compete with you? [00:04:25] S0: Yes. But we have to do... Well, that's the thing about the AI business today is that you have to just accept that everything that you're doing is getting commoditized constantly, You and have to [00:04:38] S1: Be doing the next thing. [00:04:38] S0: Be doing the next thing, right. But the other big thing that we demonstrated is that AI can provision infrastructure. So up until that point [00:04:47] S1: And you actually have infrastructure to provision. Yes. That's the thing that's different between you and all these competitors, [00:04:52] S0: right? Exactly, so we had built databases. [00:04:53] S1: You spent all these years suffering doing system software. Right, right. [00:04:58] S0: Right, it was a little tool call, spin up a database, spin up a Redis, host this. And so what people found so surprising when we launched, I mean, you saw the product July 2024. [00:05:13] S1: And when did you launch it? September 24.
00:05:13 · Paul Graham, Amjad Masad
[00:05:13] S1: And when did you launch it? September 24. [00:05:15] S0: Okay. So when people saw it, like Andrej Karpathy said, this was like an AGI moment, like a field AGI moment because everyone was like, holy crap, like LMs could do that. It freaked me out. [00:05:27] S1: Yeah. I didn't like it at all. [00:05:31] S0: How do you feel about [00:05:32] S1: it right now? Because aesthetically, it was, like, very offensive for you. Yeah. I know. I know. Because the... Either way I program is the exact opposite of this. I like... I program... You know, I spend three years writing a thousand lines of code, and if I can get it down to nine eighty in the next month, so much the better. [00:05:49] S0: But you give me the principle for all of this, which is when you first invest in Replit, when you first back Replit, you said there is a super linear relationship between how easy it is to program and how many programmers you get. [00:06:04] S1: Yes, that's true. That's not just true for programming, that's true for anything consumers use. If you make it 10% easier, you don't just get 10% more people, you get twice as many. Yes. Yeah, remember that. I didn't mean make it this easy. [00:06:17] S0: It went too far. [00:06:18] S1: I mean, just talk in English and it spits out all this code. [00:06:21] S0: Well, so the realization I had is for Replit to be truly successful, we needed to go the whole way.
00:06:21 · Amjad Masad, Paul Graham
[00:06:21] S0: Well, so the realization I had is for Replit to be truly successful, we needed to go the whole way. [00:06:29] S1: Because you did know, you were one of the early people predicting AI, And like, were you predicting, you were essentially predicting vibe coding in something like 2015 or '17 or [00:06:39] S0: Yeah, 2015, 'sixteen, in our original deck, we talk about how if you build a platform where there's millions of people coding, you can train a machine learning model that predicts... [00:06:51] S1: I mean, that's crazy ahead of its time. Yeah. Right? This was before the Transformers paper. [00:06:56] S0: That was before Transformers. [00:06:58] S1: I'm starting to think that's not the visionary, that's just your nuts. [00:07:03] S0: If you're [00:07:03] S1: that early. [00:07:04] S0: Well, think I've been fairly good at predicting where [00:07:09] S1: I'm the just future joking, is you're not [00:07:10] S0: really- Well, even when I was a kid, I remember when I used my first Nokia phone, I was like, Oh my God, people are just gonna be stuck with a new computer. And it was literally like a flip phone essentially. And some of that experience also came to Replit, which is Replit was the first programming platform that had a mobile app. And everyone laughed at us, the idea that Programming on mobile. Programming on your phone. But I [00:07:41] S1: had an RFC for it. Yes. Not an RFC, an RFS. [00:07:45] S0: Well, you [00:07:45] S1: for startups. [00:07:46] S0: You actually... You told me you would fix bugs using the Hacker News Lyspropel from your phone.