00:10:20 · Amjad Masad
[00:10:20] S0: you remember when Marc Andreessen went viral for quote unquote retard maxing and saying that he doesn't introspect on Twitter the other day, and that was like super controversial. And I think there's like deep wisdom in that, in that as a founder, you have to kind of be stupidly executing on a day to day basis. And the question is, is the company growing? Are you adding customers? Are you selling your product? And that is true for us every single day. And if not, if there's something wrong, just talk to customers. They're gonna tell you it is competition. They're gonna tell you it is something else and handle these problems locally. And then if you do that enough days in a row, you're gonna end up in a much better place, right? So in some sense, you want to shut off the entire world and you want to focus on your customers and people using your product, right? That's number one. Number two, that is totally opposite, is every now and then try to predict where the future is headed and try to predict your place in it. I think actually predicting the future is much easier than most people think. Like my TED AI talk twenty twenty three, I came up and said, software is gonna be solved in the next couple of years. Here's what software development looks like. And I sketch out vibe coding for the next two years. We actually just completed the roadmap that I sketched out. Like one of the last elements was parallel agents, how you can put in a prompt and it becomes four different tasks. For example, Replit now, I think is the best parallel agent on the market. That was like looking forward, like three years. And so we have some deep beliefs about where the future is headed. And so we're trying to position Replit in a way strategically making few key bets that will pay off in the next six to twelve months based on where we think the world is headed. But on a day to day basis, I think if you're consumed by these questions, you're not gonna be able to execute.
00:12:29 · Aditya Agarwal, Ruchi Sanghvi, Amjad Masad
[00:12:29] S2: To kind of perhaps echo this point back to everyone here who's an early stage founder, I actually think that this is one of the key contradictory things you have to become good at in today's day and age, which is that how do you find of like north stars and kind of like essentially philosophical groundings that you feel confident about? And also ignore the day to day shit that you will hear around you but also not ignore it to the point that you are actually ignoring the good information. But I don't think anybody has figured this out. So I think if it makes you feel better, I think that this is a problem in today's day and age that all of us have to figure out as founders. And I think it's a really hard problem to figure out. [00:13:11] S3: So if you're holding both states in your head and you've just accomplished what laid out in 2023, what have you laid out for the next three years? [00:13:21] S0: Oh, I don't know if I wanna say it here. These guys are gonna compete with us. [00:13:25] S3: You just tell us one thing.
00:13:25 · Ruchi Sanghvi, Amjad Masad
[00:13:25] S3: You just tell us one thing. [00:13:30] S0: Think we're headed into a post prompting world. I think prompting... The way technology move is things get layered, right? But I think we've reached the apex of sort of what prompting is and what it could do. And the next phase is about, I think, two different things. One is new interfaces for interacting with AI. And the second is fully autonomous agents that can push instead of pull, that can actually operate in the background, understand you or your organization about how they work, and do proactive work that is useful. And we're building in both these directions. [00:14:11] S3: I'm a little confused about the interface because it will still be web based. [00:14:18] S0: I think that the multimodal capabilities of models are very compelling. You know in the movie Her, like how when he's installing the operating system, he's sitting in front of his computer, he's talking to his computer. I think we're almost there in a moment where you can sit in front of the screen and maybe even turn on your webcam and maybe there's touch component there. And it's very natural, sort of like you're dueling, that we just launched our Canvas interface. You are like moving your hands, you're talking. It was like, do this here, do that here. It kind of sort of looks like Iron Man. Did you see that meme, Iron Man vibe coding? And so I think we're headed to that world. Maybe part of it is gonna be VR or AR, but we're headed to that world where communicating with machines is just gonna feels more natural, more like how I talk to you.
00:14:18 · Amjad Masad, Haya Odeh
[00:14:18] S0: I think that the multimodal capabilities of models are very compelling. You know in the movie Her, like how when he's installing the operating system, he's sitting in front of his computer, he's talking to his computer. I think we're almost there in a moment where you can sit in front of the screen and maybe even turn on your webcam and maybe there's touch component there. And it's very natural, sort of like you're dueling, that we just launched our Canvas interface. You are like moving your hands, you're talking. It was like, do this here, do that here. It kind of sort of looks like Iron Man. Did you see that meme, Iron Man vibe coding? And so I think we're headed to that world. Maybe part of it is gonna be VR or AR, but we're headed to that world where communicating with machines is just gonna feels more natural, more like how I talk to you. [00:15:12] S1: Minority report. [00:15:13] S0: Or minority report. But I don't like to give that example because it's like so dystopian. But yes, but that interface is really cool.
00:15:13 · Amjad Masad, Aditya Agarwal
[00:15:13] S0: Or minority report. But I don't like to give that example because it's like so dystopian. But yes, but that interface is really cool. [00:15:20] S2: Yeah, it's interesting. Both of the kind of use cases that you laid out, I was just actually literally vibe coding well, I was building SPC's app kind of upstairs right before this, and I was just trying to be like move this thing over here, shorten this, kind of like rearrange the layout, make it more compact. And then the other thing I was running was trying to just, again, automate the set of things that it's supposed to do every morning for me. But I was thinking in my head, I have done the same thing the last Like, two why is it not predicting what it's supposed to do? So I think both of those, which is inferred use based on kind of my usage patterns, and then also the UI stuff can be so much quicker than what it is right now. The prompting back and forth is so slow. [00:16:01] S0: Well, you should just check out Agent 4. I have. The Canvas there's some fundamental way in which the big labs are operating that is limited. I think in trying to build that AGI, they're not really building systems that are super user friendly. Claude Code work is amazing Claude is amazing, but everything is based on this prompt, something happens, and maybe you can schedule it so that it happens every day. But it is not a fully integrated system that understands. And this is the opportunity for everyone here, is building amazing user experiences, because I don't
00:16:01 · Amjad Masad, Ruchi Sanghvi, Haya Odeh
[00:16:01] S0: Well, you should just check out Agent 4. I have. The Canvas there's some fundamental way in which the big labs are operating that is limited. I think in trying to build that AGI, they're not really building systems that are super user friendly. Claude Code work is amazing Claude is amazing, but everything is based on this prompt, something happens, and maybe you can schedule it so that it happens every day. But it is not a fully integrated system that understands. And this is the opportunity for everyone here, is building amazing user experiences, because I don't [00:16:36] S3: think I agree. I think the hardest thing, or my biggest user issue and my biggest user complaint is how do I load all the context that I wanna share to be able to build something as powerful as I want it to be? Which is like going back to what you were saying, like why hasn't it learned from my usage patterns? Why Yeah. Isn't it retaining [00:16:58] S1: So that's where we're thinking about like the pull and push information and what So would that look right now it's a lot of the user is telling the agents like, do this, do that. You don't understand me. Like, no, no, no. Like I meant that. Did you forget about XYZ? We spoke about it multiple times. I mean, and at this point, maybe you need to understand my taste and I build multiple projects, and I need it the same way. So if you're working with a real engineer or a real designer, at this point, they need should understand each other at some point, at some level. This is what we're trying to