Y Combinator Startup Podcast
00:13:15 · Amjad Masad, Paul Graham
[00:13:15] S0: Right. Yes. [00:13:16] S1: Right? Like way better than some nerdy program online Mhmm. Kinda company really needs to have. [00:13:21] S0: And it helped a lot with our initial sort of audience, which are like... A lot of them were like teenage hackers. [00:13:26] S1: I thought it was really cool. Mhmm. It made Replit different. Yes. Still have... Replit still... There's something very distinctive about it, you know? Has a very... Replit, like a lot of very successful companies, it has a very strong sort of character of itself, which sounds meaningless, Mhmm. You know? You know what I mean, though? It's like a very sort of a strong solution. It's not a weak... Like in the chemical sense. Yeah. It's not a weak solution. Replit is a strong solution. [00:13:55] S0: Our conceptual understanding of what we build and our product is we always think about engines and hard mechanical things. It is not. [00:14:07] S1: That's a lot of what that graphic design was about, Yes, wasn't [00:14:10] S0: yeah, yeah. It's like there's something that's sort of hardcore about it, but it's also paradoxical in that we slave away so much to make the UX so simple for users, but on the back end, just really ugly engineering is happening. [00:14:27] S1: Systems programming, real systems programming, That's part of your DNA.
00:14:27 · Paul Graham, Amjad Masad
[00:14:27] S1: Systems programming, real systems programming, That's part of your DNA. [00:14:33] S0: That's part of our DNA. It's been always sort of challenging because a lot of programmers will look at Replit and sort of dismiss it because it looks very simple. Yeah. But then even people who end up coming to work at Replit are just like, oh my god, I didn't [00:14:48] S1: know There's so much stuff behind the scenes. Do system administration for people's sins, right? [00:14:55] S0: Yeah, I mean, we do it so they don't have to do it. [00:14:58] S1: Yeah. Yes. Do you think that will be the thing that helps you prevail over competitors? Because that is one difference, the depth of stuff that you've built and the depth of understanding you have of how to deploy programs online. [00:15:12] S0: Yeah, so that is helping us win enterprise in our category because they are a lot more sensitive to reliability, security, all of that stuff. It's really hard to get consumers to care about security, for example. A lot of our competitors do really reckless things with regards to building apps that are unsecure and so on. [00:15:40] S1: Because if you have agents writing all this code, who knows what it's doing? [00:15:42] S0: Yes. Right, yeah. And so we put a lot of guardrails, which makes our products sometimes slower, sometimes more expensive, but enterprises really like that.
00:15:42 · Amjad Masad, Paul Graham
[00:15:42] S0: Yes. Right, yeah. And so we put a lot of guardrails, which makes our products sometimes slower, sometimes more expensive, but enterprises really like that. [00:15:54] S1: And so these newly arrived opportunists who don't have the same sort of depth of infrastructure like Replit's like this iceberg, right, where there's this visible bit above the water that's like, look, you just tell it in English what to do, right? And it writes your code, then underneath, there's all this stuff for deployment in sandbox. Yeah, like we have [00:16:11] S0: kernel operating system engineers. We built our own file system from scratch to be transactional and fully reversible. I remember hearing all these stories. Back in [00:16:24] S1: the days before Vibe coding, that's what all the updates from you would be up about lately. We made our whole new file system. I hope somebody appreciates it. [00:16:31] S0: Yeah. I mean, when wanna make like a coding collaborative, like you had to invent a lot of core primitives that... [00:16:40] S1: So that's an interesting question. Did the stuff that you did in sort of version one of Replit, it's just online programming and you made... And it's sort of collaborative, did the stuff that you did to make it collaborative also help with Vibe coding? [00:16:54] S0: 100%. Mean, it's sort of surreal how plug and play it was. So when you put an agent into a rapport as [00:17:01] S1: we call it It's like you're collaborating with the agent.
00:17:01 · Paul Graham, Amjad Masad
[00:17:01] S1: we call it It's like you're collaborating with the agent. [00:17:03] S0: Yes. The agent is another user. So the initial version was basically the agent was just another user on the same system. [00:17:11] S1: And you had everything else built. [00:17:12] S0: Everything else ready for it. And even even you and the agent could work on the same project and it will just work. [00:17:20] S1: So it's no wonder you discovered Vibe coding first because you were perfectly positioned. You had all the other ingredients. Right? [00:17:27] S0: I mean, this surreal thing, even the way the company was organized was so perfect. So we had the, like, infrastructure team that handles, like, code execution team. Right? This is like sandboxing team, whatever. And then we had the sort of the deployment and cloud services team. This, like, you know, you can, you know, spin up databases and deploy VMs and whatever. And then we had the team that was building the glue between the sort of infrastructure and the IDE, and then we had the IDE team. And so where does AI fit into all of that? So we ended up creating a new agent team and we put it in the middle. And all these people are building now for the agent. So the infrastructure team is exposing tools for the agents. The IDE team, instead of building IDE for humans, now they're building edit tools for the agents. Really? [00:18:22] S1: What sort of IDEs do agents even need? I thought they just generated code. [00:18:26] S0: Well, initially in '24, editing code was actually a hard problem.
00:18:26 · Amjad Masad, Paul Graham
[00:18:26] S0: Well, initially in '24, editing code was actually a hard problem. [00:18:32] S1: You mean having it go back and fix mistakes? [00:18:34] S0: Yeah. So there was no native edit tool. Now there's the native edit tool every foundation model give you. But back then, we had to do our... We had to create our own like, sort of edit tool. When you give it, like, a chunk of text, how... Where to replace. You couldn't rely on line numbers because they would hallucinate line numbers. There's all these different things that [00:18:58] S1: You had to make it ridiculously robust in the early days because the the LLMs were so bad. [00:19:03] S0: Yeah, they would also be lazy. So they would write like a function and be like, get lazy in the middle. It's like, To do, finish it later. [00:19:10] S1: Which is not executable. [00:19:12] S0: Was Although it is now. Was insane. And that was both OpenAI and Anthropic had the same bug. And I think it's because of the chat model. I think they always wanted the chat to kind of halt early, not produce mountains of text. So that was an artifact of that. And so we had to detect that and loop over. So even editing text was a lot of work. And now there's a native edit tool. [00:19:39] S1: Is there anything like an IDE for agents though? [00:19:42] S0: So in the initial days, there was a lot of research on... [00:19:45] S1: No, mean now, [00:19:45] S0: the way agents No, no, no. [00:19:47] S1: Okay, now the agents just like generate the code and that's
00:19:47 · Paul Graham, Amjad Masad
[00:19:47] S1: Okay, now the agents just like generate the code and that's [00:19:49] S0: Now you can just give them a shell. You just give them a shell and they'll just write shell command. They can do everything with like cat and basic shell commands. They're very good [00:20:02] S1: at But you needed to have the IDE early on because the agents were bad. Yes. Yes. And so that was one of the things that enabled you to be so early. [00:20:10] S0: Yes. [00:20:11] S1: Could you have done it any earlier? No. Because you got it just at the point where it barely worked at Yes, [00:20:20] S0: and it barely worked. And like when we launched it, it was in beta, and I would say 50% of it. [00:20:26] S1: I remember I would go over to your house and you would show me things and they would not work. [00:20:30] S0: Yeah, maybe 50% is too harsh. Would say like, yeah, 30% of the time it would not work. [00:20:38] S1: Okay, so why would users tolerate that? [00:20:42] S0: Because it was such a magical experience. [00:20:44] S1: Oh, and also presumably when they were cutting and pasting stuff back and forth between chat, GPT, that also wouldn't work either. But if they couldn't program at all, then programming 70% of the time was