00:17:29 · Aditya Agarwal, S2
[00:17:29] S1: Yeah. I mean, I I love that answer, particularly the part about everybody should just be using these technologies. Yeah. And I often... I'm just like, you know, you kind of... There's there's still a little bit of an inertia where you go in, you're like, I'm gonna have to learn how to use this thing. I'm gonna suck at it in the beginning. But these tools are just like... Honestly, all of us are capable of using these tools. 100%. By the way, one of the things... One of the notes I have here is that Tony is customer obsessed, not competitive focused. And it's pretty amazing. You've kind of brought up, like, having a customer centric point of view multiple times. [00:18:00] S2: And it's clear that that's how [00:18:02] S1: you kind of like frame a lot of your kind of thinking about both what to build and what DoorDash should be doing. It's pretty amazing. Yeah. And generally, there are these huge platform shifts is when both old and new incumbents tend to get displaced. Right. Or at least they are under the most threat. Right. So what do you I mean, do you worry that there is a more AI native version of DoorDash out there? And if so, like, what would that look like? You don't have to give out the ideas, but, like, I'm curious, like, what is your... What would be your... Like, what is a more AI native version of kind of an interface to the physical world? What would that look like?
00:18:43 · Tony Xu
[00:18:43] S0: You know, I I always go back to what is the end to end problem that we're trying to solve and and what are the different component parts. Right? One of the challenges we found in building DoorDash is it's it's not just an app. It's not just a logistics algorithm. It's not just, you know, a lot of customer service issues that we have to resolve within usually seconds or minutes. It's all of these things coordinated together in order to, you know, just do something simple, like bring you lunch or bring you dinner. And, you know, the question I always ask is, where can a new technology, to your point... And it doesn't have to be, you know, AI even though I'm sure that's getting more than a 100% of the airtime these days, whether it's autonomous technologies, you know, things that are physical, like hardware. But regardless of the technology, regardless of, like, the specific implementation details, like, what is this thing actually capable of in terms of which part of the stack can it actually do better? Because in truth, like, I don't view any digital solution that's going to, you know, replace an end to end system that has physical parts if it can't also do the physical part. Right? But maybe there's parts of the digital experience in which it can. You know, certainly, for instance, you can write much better queries on chat assistance with greater intelligence that right now the DoorDash app can do. Okay. That might be an area in which we look into. Right? But that alone would be insufficient to actually bring the customer what they want, which is maybe lunch or dinner or, you know, the the latest pair of shoes or, I don't know, lipstick that they forgot because they're on a work trip or something like that. And you kinda have to do all of those things. And so the way I think about it is, well, where are we most vulnerable? Or where do we have the most opportunities? The same version of the same question, the other side of the question, where it's actually gonna help again. If I looked at every metric at the company from a customer's perspective, not whatever the output metrics might be in terms of what we report to Wall Street, well, very quickly, you're gonna know, I I think, like... And then I'll just evaluate the new technology against each one of these areas to see where can I do better?
00:20:55 · Tony Xu, Aditya Agarwal
If we can do better, we're gonna be the fastest or the greatest adopter of that tool. [00:21:00] S1: Are you doing anything interesting in terms of accumulating kind of a of a a data moat that, you know, perhaps in the call it in the coming two, three years would allow you to essentially build your own kind of, like, RL models, like, know, or your own whatever the version is of kinda doing post training, kinda whatever their current regime is, of, like, in three or four years. Are you doing anything special in terms of readying yourself for that?
00:21:29 · Tony Xu
[00:21:29] S0: Yeah. So one of the things, if you think about what we've had to do in terms of building DoorDash is we basically have to digitize the physical world. And I think while, you know, LLMs can obviously train on all of the Internet, if the data is not available on the Internet, that would be very difficult to perform that task. You know, data such as where can we find the last parking space near this building in San Francisco? Data such as how many Coke cans are available to buy at the Safeway on 4th Street nearby here? Or where can I find the cheapest pair of shoes in size ten and a half, you know, for my upcoming, you know, five k race this weekend? Well, those answers don't exist on the Internet because those answers don't even exist sometimes in the physical world in a digitized format. And and so if you think back about, like, you know, what... One of the things that DoorDash is building, we're building lots of inventory systems effectively, where we're trying to catalyze... Catalog information for the first time that's never been available on the Internet. You know, this is why we have to use our own, sometimes, mapping systems, our own inventory systems. It's not because we don't trust third party systems. It's because third party systems either don't exist or the information is very incomplete. And so for us, you know, that information can be very useful. And in the way I always think about how you build, you know, technology is how do you actually grow the pie so that you can get more participants into the into the ecosystem? And, you know, this is a perfect example where we're building lots of this information, and we probably will be a very large, you know, partner or participant where we will team up with some of these companies building these foundational models to actually build very useful models of information that doesn't exist that it would be impossible to train on because it... Because they don't have it.
00:23:27 · Aditya Agarwal
[00:23:27] S1: Yeah. No. That's actually a huge... Both a huge asset and an opportunity. Yeah. That's amazing. What agents are you personally using? Yeah. A personal or professional capacity?
00:23:41 · Tony Xu
[00:23:41] S0: In person... Oh, I play with pretty much all of them. I I literally personally play with all of them. You know, currently, I've been very impressed with what Gemini has shipped. Oh, it's amazing. And you asked earlier, I think, a question about, like, you know, why do we not yet have the right set of interfaces? Why are we still stuck with text? Well, I think part of it is because I think some of these companies are trying to figure out, well, what does that mean? Like, for instance, if I'm doing a set of queries where I'm trying to learn about my health, as an example, where I may wanna input a lot of, you know, images that are held in arcane systems like an MRI or, you know, an X-ray or, you know, a blood test or... You know, these are in a lot of different systems that it has to pull from. And then there's obviously the consumer privacy component to that, which, you know, adds a layer of of complexity specifically in health care. But I could give the same kind of, you know, example if we talked about what if you wanted to buy a house, or what if you wanted to, you know, think about investing for the first time? Or I'm a parent. You know, I get a lot of parenting advice for some of these, assistants, and that's a... Sometimes you do wanna, you know, input, like, screaming audio from a child or a toddler. You know? And the point that I'm trying to make is, you know, each one of these tasks actually require a pretty different interface to offer a best in class experience versus what's provided today. And look. I think these providers know that, and they're working towards it. But those are some of my use cases, and I would say that, you know, those are some of the things I've experienced while using them.