Y Combinator Startup Podcast
00:42:10 · Aaron Epstein, Stephen Haney
[00:42:10] S1: Yeah. It's interesting. I I was having a very similar conversation with someone who leads design at a multibillion dollar design forward company about... There's a thought process that goes through a human's head where they're reacting to something and a criteria they're evaluating. Yeah. And it's very, implicit for us as humans right now. And the question is, can you verbalize what those things are? Like, what is the process of you deciding whether something is good or not or looks good or not or Very subjective. Hits hits the bar or not. Can you literally go through and quantify all of that and teach that to a model to the point where it has taste like a human would? And then the question becomes, well, if all the models were able to do that, does everything just look the same again? And then you always need the humans to be drive... To finding the next thing that's actually gonna be new and stands out and [00:43:03] S0: Yeah. [00:43:03] S1: And pushes the limits more.
00:43:03 · Aaron Epstein, Stephen Haney
[00:43:03] S1: And pushes the limits more. [00:43:05] S0: We see that even within human... Pre... Premodels as trends take over. Yeah. And everyone kind of looks like linear for a couple years, and then somebody else figures out something. Right? I do think the models will get good at tactics. Mhmm. They'll get better at not making everything bold. Mhmm. You know? And I think that... That's great, that'll be really that'll be really nice. But the... Just the thought, the product... There's so much of design. Again, it's like, was this showing value prop or not? That stuff's more important. This site, for instance, that I think is a little too wild, maybe it's really important for them to be really fun and exciting. Maybe that's what their market needs. You know? And so maybe it's a higher order thing for them to be very exciting. And can a model make that decision? Can it... Know, maybe with enough information, you could start to get there. But I think we're pretty far away from that. You know? Maybe ten years, something like that. [00:43:47] S1: Yeah. [00:43:47] S0: May... I don't know. Maybe I'm too pessimistic. Maybe it'll come sooner. But I I still see humans as just this essential piece of the of the puzzle. [00:43:53] S1: Yeah. We certainly hope so. [00:43:55] S0: Yeah. Let's hope so. [00:43:56] S1: I'm curious. How how are you guys using agents internally at Paper? In some ways, we
00:43:56 · Aaron Epstein, Stephen Haney
[00:43:56] S1: I'm curious. How how are you guys using agents internally at Paper? In some ways, we [00:44:00] S0: use them a lot. And then in other ways, we use them... We're kind of like dinosaurs in some ways. Like like, we don't have agents in our Slack. And we don't have... We use agents for coding quite a bit, but we still read our code. We read all of our code, every single line. Every line. Usually, multiple humans reads every single line. [00:44:12] S1: Why is that?
00:44:13 · Stephen Haney
[00:44:13] S0: Well, one is design tools are really hard to build. Mhmm. You know, you can't vibe code a design tool at the level of a Figma today. Maybe you can eventually. And so it takes a lot of a lot of really precise, really careful work. It's very multidimensional Mhmm. In terms of the systems. And so we just have to be really careful. And we wanna make sure our experience... You know, designers need a 120 FPS. It needs to be fast. It needs to work every single time. And the agents just aren't there for that yet. The other thing is a small elite team. Right? So 12 people to keep the talent bar extremely high. I think when you're building companies, that's how you should try to do it these days. Mhmm. And use the agents to accelerate the the people. But if you have less people, you have less communication burden. You have less of these other costs to go through. One example of how we use agents, is Agu, our brand designer. He's a brilliant brand designer, great at visuals. He codes a little bit, but I don't think he'd call himself a programmer. And he made our entire website in about a week before this last launch Wow. Including shipping it, including all of the animations. Mhmm. And he designed everything in Paper. And this is the kind of design... You wouldn't wanna do this with prompting. Right? You need this... He has he has shaders in the background subtly coming through, these motifs. He did everything in Paper, and then he just prompted... And this was, you know, six months ago now, so different models. He just prompted the animations on top of his of his Paper file. It's like, know, take take my Paper designs, turn it into a code base. This is the Next.js site, and put it together for me. No one else looked at this. So Agu did this end to end in a week. Because he's a designer, he puts in these Easter eggs. There's little, like, fun extras.
00:45:38 · Aaron Epstein
[00:45:38] S1: And it's all just interactive.
00:45:39 · Stephen Haney
[00:45:39] S0: It's interactive. And this is something he would not have done. He wouldn't have had the time. Yeah. I'm sure he could have... You know, skill wise, he could have got there. But, like, he wouldn't have the time to do this. So I think that's really, really cool. And and so I think a lot of times, it shows up more not in our core product where we're actually pretty human based still for precision reasons, but in all the other stuff, all the videos we need to make, all the marketing site, all of that other stuff where we can, you know, have agents help us help people go faster. So we do quite a bit of that. And then, of course, the coding agents. We do use Cursor extensively. We use Claude Code extensively. We use BugBot for our PR reviews, which has been really helpful. It's a little too pedantic sometimes, but it still catches catches bugs, which is good. There's a little bit of a FOMO right now with with agents, where it's like, oh, every company is shipping without looking at their code. Are you are you going too slow? You know? We are kind of taking the approach that quality software takes time. Mhmm. And we have competitors that have every single feature, but they don't have the attention or people don't care. And for whatever reason, whatever that is, it's like people want paper to be better. They don't want the competitors that have every single feature to be better. Mhmm. And I think that it's because we care, and we're putting in that precision and that care about these things. So even if we're moving a little slower, people still care more about it. And so I think that quality matters