00:09:12 · Aditya Agarwal, Tony Xu
[00:09:12] S1: You know, one of the things that is tricky about building for the real world is that, you know, margins are thinner and kind of like... You know, sometimes when you when you talk to entrepreneurs who are kind of building in hardware or the real world, you realize what a great business model software is. Right? Like, if you have 75, 80% gross You [00:09:31] S0: have a lot of experience with that. [00:09:32] S1: Yeah. Exactly. So 75, 80% gross margins, you can fill a lot of up and still your business will be alive, generally. Right? It's much harder with, like, of real world businesses. So, guess, how do you balance in some ways... How do you create a system within these companies that allows you to move fast while also realizing there is a cost to actually making mistakes that is much larger than your average software company. And obviously, there's a very famous example that you've talked about, about the Stanford football game, I think. But how do you not let that experience intrinsically just slow you down? So how have you found that balance at DoorDash? Yeah.
00:10:13 · Tony Xu
[00:10:13] S0: I I I think that, you know, this is probably, you know, the answer that would be true honestly in building great software too, which is if you care a lot about a craft, you should care a lot about efficiency. And I would argue every craft, every function has some efficient frontier. There's always some version of the optimization where you're trying to maximize quality subject to some constraint. The constraint could be time. It could be budget. It could be something else, but every function has that. And I think everyone who I've always admired, you know, in terms of their pursuit of their craft excellence, always gets that. So there's no, like, one thing, you know, Aditya, to answer your question, but it's almost like every project, if you think about, you know, the unit of work at any company, including your own, probably boils down to some project. That project boils down to some set of workflows. And for each one of those projects and workflows, there is probably this efficient frontier. I think even the fact that you institute that and then measure it and then hold people accountable to it, you know, that ultimately becomes a system that, you know, reinforces in real practice, not just culturally, but actually operationally how you actually build a system that cares about quality and cares about efficiency.
00:11:40 · Aditya Agarwal
[00:11:40] S1: Moving on to perhaps just kind of, like, you know, tackling obviously what everyone here is working on directly, indirectly, obviously, various parts of the AI stack. Sure. Maybe to think first about AI at the application layer. I think one of the tricky parts is that it's weird, right? Like, you would think that we would have moved on from a basic kind of, like, text IRC interface to AI, which kind of, you know, came about in November 2022 with ChatGPT. But it's actually pretty... I mean, if you use Claude code in terminal mode, it's literally exactly that right now. So I think... I guess what I'm trying to... Do you have thoughts on where does this... How does this UI evolve? Because Yeah. I don't actually think we know yet as an industry. Like, you know, voice mode, text mode, they all seem quite basic. So I'm wondering if you are seeing something within DoorDash that you might have an idea to share.
00:12:34 · Tony Xu
[00:12:34] S0: I definitely agree with you that we're all trying to figure it out and that everyone knows that, you know, whatever we have today is probably the worst version of what we're gonna have in the future. Right? And so I think in some ways, that's... You know, hopefully should bring a sense of optimism. I mean, these tools have already brought tremendous productivity gains and, you know, frankly, ideas, new ideas, etcetera. I think the thing that sometimes gets missed, though, in in the search, if you will, for some of these... And and I think this will be the answer or the... At least the approach that we're thinking about at DoorDash is you gotta figure out what the end to end job ultimately for the customer is. And I think sometimes it's easy to lose sight of, you know, what that is. You know, for for DoorDash, well, you know, the end to end job is to bring you something, whether it's a T shirt, a water bottle, a hamburger, whatever it might be. And our job's not finished until that whole sentence gets completed. Some of that happens in a digital interface. Some of that can happen through, you know, photos and video, maybe voice, maybe search, maybe an assistant that does a complicated curie for you. But some of it has to get finished ultimately offline and physically. Actually, many things have to happen offline physically in order for that one sentence to be completed. And I think that's why it's hard. It's because the end to end job, you know, for any user, depending on what the job is, may actually require a different interface. And I think that's what we're all trying to figure out. We're we're we're kinda solving for different use cases here, different use cases there. On the flip side, you know, a lot has already taken off on just some of these chat interfaces. I mean, the usage, I think, speaks just by probably the usage in this room. And so that's where we're at right now, and I think, you know, the main, you know, at least advice I give our teams is what's the ultimate end to end job? And then from there, you know, let's figure out what the interfaces ought to be. And I and I think right now, I think every company is is is in some some some exploration of that.
00:14:44 · Aditya Agarwal
[00:14:44] S1: As a prominent company that is obviously integrating AI, how do you make sense of the model landscape? Are you just happy that there are multiple model providers that are kind of like essentially trying to win your business? Do you also feel as though that you need to be moving down the stack yourself to own more components around the context, the memory, the long horizon planning? Like, what is your kind of like chess play there?
00:15:07 · Tony Xu
[00:15:07] S0: It goes back to a bit the the the previous question where I think the name in the game is to understand what is it that we're actually doing for the customer. Instead of starting from what is the thing I must protect for my business. It's almost like if you can run your business with metrics that only a customer would care about. You know, for example, a customer doesn't care about how many users you have or how much revenue you generate or whatever your ARR is or your weekly growth rate. I actually think that that will give better insight as to potentially, you know, what to actually work on in terms of, you know, producing a road map. And so, you know, for us, you know, the first thing is you gotta play with all the tools and understand what are the actual true capabilities versus what their promise is. And I would argue, you know, that's different by different types of areas in coding, like you're saying, Aditya, like, totally agree. I think the progress of reality versus promise, that is getting very quickly closed. In some other areas where you're trying to solve an end to end workflow that may have multiple systems that need to be touched, that requires human judgment, that that does not have a verifiable good response, that's not a great place right now, you know, in terms of, you know, for for AI agents where the gap between reality and the future potential is still fairly sizable, I think. You know, at at least the way that we've thought about it is what... How are we gonna make the most sense of what the end to end job is? How do we make the most sense of what are the tools today? What can they be used for? And then how do we actually, you know, marry those two things to deliver something unique? It's not, you know, it's not rocket science, you know, to to to answer your question. It's also not starting, though, from a place of, okay. Here is the, you know, big strategy, and and and and, therefore, these are the things that we must own. Because I think, candidly, I would say that the field is moving so quickly that it's very, very hard to know exactly, you know, certainly within the next five years, where the technology is gonna go and how far in in terms of what it can do will be.