There are battles for attention, you know, bits, and battles for what's happening in the physical world, for the atoms. And that we largely occupy ourselves in the second category, right? We are in the war for atoms because, one day ultimately, what's the point of having a personal assistant if it can't actually do real things.
Tony, thanks so much for doing this. One of the things I wanted to start with was this concept, if you'll indulge me on it, because it's something I've been thinking about a bunch, kind of around the idea of AI spend at companies. And I think the trigger sort of anecdote that I saw that I'd like to get your take on was Uber was basically saying, in the first few months of the year they blew through their whole AI budget. They were just rampantly consuming tokens, and what did they get out of it? Did they get more rides? Did they get more drivers? Did anything, their margins improve? Like what happened? And I think they were kind of saying it wasn't obvious. And you're obviously the leader of a company that's super metrics-oriented and has executed super well, so I assume you must think about this. But I'm just curious, kind of at a high level, how you think about what is a large rising line item, and as it continues to rise, what do you care about?
Yeah, I think every company right now is trying to figure that out, where the ultimate goal of any technology is actually to hopefully solve a problem and actually make things cheaper. I mean, that's the goal of technology. But as you know, especially when you have constrained resources like compute in the case of these large LLM companies, sometimes you don't get the timing of these things always exactly when you can deliver the outcomes, or you don't get the cost profiles or the efficiency profiles to serve the technology to match exactly when you can deliver the customer outcomes. And I think that's the world in which we live today. So I think every company is trying to figure this out. In terms of how we're thinking about it, first and foremost is what customer jobs can we actually solve. And so I do think when you have new technologies, there's always going to be this period of inefficiency and almost like discovery, where you got this new toy, you don't know exactly what you need it for. You probably kind of know that you don't really need a frontier model to know what the weather is, perhaps, but on the flip side, you don't also know the limits or the ceiling of what the technology can do, and you kind of want to know that. And so the way at least I think about how you do this somewhat efficiently, even though by definition you're going to accept inefficiency in the discovery process or in the invention process, you want to do it in the most contained set of ways that deliver customer outcomes. So you actually want to put customer outcomes and start there, and then actually give teams as many shots on goal towards those outcomes as possible. Doesn't mean you get there, but at least it's directed.
Well, it's funny because as a software company, which is where so much of the AI productivity is happening right now, it's actually harder in some ways for them to measure the end outcome of the work that their engineers have done than maybe yours, where you can say, hey, I can measure did we complete more deliveries? Did we get more restaurants? Do we have new users? Or all the things that you're tracking. I was sort of reflecting that over the last few years you've kind of watched this egg move through the snake in the pipeline. From first there's the buildout and there's chips, and then you get data centers, and then it's like, oh man, are people ever going to use this? And now they are using it and the token spend is ramping, but now it's like we've got to turn the token spend into burritos at people's homes. And you're in as good of a position to see that as anybody. So are you going to have certain teams try to move the metrics in a market with AI, or is it kind of just bottoms up, let the teams explore whatever they want with AI?
Yeah, I think it depends on the team. So if you're product teams, you're exactly right, you can direct teams towards metrics that matter to customers. If your merchants on DoorDash today, they're onboarding 35 to 50% faster because we can actually have AI produce all of their catalog or their menu in the case of a restaurant, help edit your photos, help set up what might be the best way to describe yourself, especially if you're a new retailer coming online into a city. So there's metrics like that that are very trackable, measurable, have immediate positive benefit to customers. Similarly with dashers, we've had AI help detect both fraudulent activity as well as safety incidences before they occur, and you can stop those and take preventative action if you knew that was actually happening. So there's things like that where you can get immediate customer benefit. But you said something earlier in your question about software teams, for example, and I think one of the things you notice there is that only certain parts of a software engineer's day is writing code. And it's great that we have models today to write the code for us, but that helps maybe the 25, 30, 50% of our time that's actually shipping code. But what about everything else that has dependencies in product reviews, design meetings, alignment with business teams, etc.? That also has to change. If that doesn't change and come together, you're not going to be able to have perhaps the productivity gain that you hope to have. And I think that's what companies like ourselves, but I'm sure a bunch of companies across the industry, are trying to figure out, and getting those workflows right so that you're not just AI native from code development, but you're AI native in how you actually operate.
Have you done anything to track different engineering and product teams inside the company, comparing how much more productive they become with AI usage, like ones that are using it more and less heavily and things like that?
Yeah, I mean, you get all these facts, and I'm sure as with all things, there's always a distribution of outcomes. Maybe the average is 50% more productive, but you have engineers who might be 35 times more productive. It's crazy, right? And it is crazy. And it's always fun to study that distribution. But again, I think what's the value of that? The value of that is really knowing what the art of the possible is, especially in understanding your workflows. But what someone like I have to do is I have to think about, well, how do you create that working environment for everyone? Because it's great that we have engineers who are 50 times more productive. How do you actually get everyone? But is that an observation to that new ceiling?
Exactly, and the question is, is that an observation and somebody's going to be far out, or is there a learning there that you can go replicate?
Yeah, exactly. And I believe it's the latter, because with all things, especially if they're newer, you're always going to get a distribution of outcomes, right? There's always, in sports, in academics, there's always a distribution of outcomes. But that doesn't mean there aren't classes and teams and good things that you can teach as a coach or as an instructor so that you can get the rest of the class to move to the higher plane.
Yeah. You know, for a lot of companies when AI came around, it was either the most petrifying new technology ever, or it was like, oh my god, thank you, tailwind. And I feel like for DoorDash, it was probably neither. And you tell me if you object to this, but I would say you're mostly operating in the world of atoms, not bits. You've got this two-sided and three-part marketplace with restaurants and dashers and consumers and all these things, and you're very related to cities and all of that stuff. And so I would think it's both insulated from AI in a good way, and then you also maybe didn't, it's not like your business started growing 400% faster because this AI thing showed up and wasn't there yesterday. So how has that been? Is that right? Is that about what it's been for you? Has it been kind of just insulated, and that's both good and bad?
Yeah, I mean, we're not selling burritos or Nike shoes or groceries as fast as tokens at some other companies. But yeah, look, I think you're largely correct. I remember, I don't know, I think this was 2021. Yeah, this is 2021, before the arrival of ChatGPT. We were playing around with some of these models. I mean, I think it was like GPT-2 or something like that, right? And just kind of getting a sense of what the world could look like. And obviously, the models back then are very, very different from the models of today. But even then, we kind of had this description of the world where there are battles for attention, bits, and battles for what's happening in the physical world, for the atoms. And that we largely occupy ourselves in the second category, right? We are in the war for atoms and moving things around. And if we can do that, and we can build a catalog, for example, for where every item exists inside of a city, or every parking spot exists, or all of this information, we can, as the kind of company's battle for attention start maturing, we can really partner and work together in very productive ways. Because one day, ultimately, what's the point of having a personal assistant if it can't actually do real things?
It's actually funny on that point. It strikes me that one of the, it's very cool, but one of the things that I think there's room for improvement is that so much of tech is just typing on our screens and we're all living in the computer. But at some point, to make lives better, you need to move stuff around. You need education. You need stuff that happens in a hospital.
Yeah. You know, and so you need real estate to be developed. It's like all this kind of stuff, but it's all physical at the end.
Yes. And so in some ways, it's like all the software ultimately does need to be in service of stuff.
Yes. I mean, and this is exactly why, I guess it would have been five years ago at this point, the strategy was very much, no, play the game that you were meant to play. The game we're meant to play is the game of atoms, and be best-in-class at that. And not only is that good for our business, but I also think it's great for our relationship over time with these companies as they finally build out some of their assistance, and build out super intelligence, and build out agents that can actually do things versus what they currently do.
Because we naturally will need one another in order to be useful to the communities that we serve. Has there been any impact on recruiting or things like that when everybody's losing their mind on AI and you're in this different thing that you're like, I know this is really important and it's going to become even more important eventually, but have you had to work through any of that?
Yeah, I mean, it's definitely a battle. We are an applied AI company, and we have been pretty much since day one of the company, whether you want to call it math or machine learning or now LLMs, we can use whatever terminology we want. But as the years have gone by, we are an applied technology company that also solves real-world problems. So yes, it's certainly difficult when you're competing in places where activity is really hot, in machine learning, in folks who are looking to do applied research. It's very competitive right now.
Yep. Are there things, as you think about your next frontier and where you want to go, and you look at the AI stuff, it does seem like some of it is pretty direct. For example, I could imagine robotics maybe could matter at some point. I could see self-driving matters at some point. What are the next parts of AI outside of the LLM on the software side that you care about?
Yeah. Well, I would say first, before we step away from some of the LLMs, we do care about what LLMs can do, and I think they have gotten materially smarter and more powerful. Not just helping with coding, which may be the most prevalent use case today and certainly the most prevalent spend consumption.
Well, it's definitely a good input for building stuff.
100%, 100%. But even, for instance, one of the things we always ask ourselves is how can a technology actually improve outcomes for customers. So if you take as an example the DoorDash app, we've grown a tremendous amount over the last even five years, forget the 13 years we've been doing this, just in the last five years, kind of the same arc in which some of these LLMs have grown up. DoorDash has moved from one product category, restaurants, one market, the US, into virtually every retail category, the leader now in deliveries of grocery, convenience items, alcohol items, across 40 countries.
There's lots of positives with that. One of the challenges for a consumer though is, boy, it's a lot harder to use an app like DoorDash because you now have restaurant things in there, you now have grocery items in there, you have retail items in there, you now have the ability to make reservations or get deals inside of restaurants. It's getting more complicated to actually use our product.
LLM really can help with that. And just as I think there will be personal agents soon, the DoorDash app should be a personal agent. It should be a personal agent to help you do anything inside of your city, right? I can't think of any other product and any other utility, frankly, greater than the number of connections that you can have between you and the businesses inside of your city. And so those are the kinds of things that you'll see us launch just with LLM. You should be able to have an easier time actually doing research on what items you want to buy or what might be great recommendations for you based on all the history that we have across tens of billions of orders.
And you should pretty much be able to get anything to your house and not have to shop for it.
Yeah. Or have the choice of getting it later when you're home, or because you may be away, or the next day because it's better for you. You're exactly right, and so I do think that there's still a lot of excitement just in the LLMs. But you're right, if you want to go beyond that, there is a lot of work around the physical space. So for example, we've been working on autonomous vehicles since 2019, and a lot of that started with traditional systems, kind of with machine learning, yet traditional ways of thinking about how to use those techniques to build things like recommendation systems. LLMs kind of put a complete new spin on it and didn't really require any of those techniques. Something similar is happening in the physical world, where it used to be if you wanted to drive autonomously inside of a market, a lot of what you would do is actually build mapping systems and almost like heuristics and rules to create an engine in which you can make good decisions and weigh them in real time. Or perhaps you can put all of this into a neural net and take similar techniques that some of the LLM companies are using and actually make even faster and better decisions, right? And so there are things like that that are going on that allow us to do some of the work that we're doing with autonomous delivery faster, as an example.
Yeah, totally. This is now going way back in time, but are you surprised how much people are willing to consume on apps? Obviously, when you started the company in 2013, this was the plan, but has it gone a lot further than you expected? People really will just order a burrito for a lot of money, and it probably, and it's worth it, like it turned out that it's worth it to people. But was that obvious to you in 2013 that it would go this way?
No, it was not obvious. People pay like 35 bucks for a burrito. For example, when we launched our first partnership with a national brand, that was July of 2015, I believe, with Taco Bell. I was actually quite skeptical whether or not it would work. On the one hand, you have this legendary brand that's never offered delivery before being offered for the first time. That was the bull case version. On the flip side, you're right, you're going to pay a premium to get that order.
And which wouldn't have been intuitive because you said, well, I can get this order for $4 in store.
Yeah. Exactly. And so that was not obvious to me in 2015 of what would happen, the bull or the bear version of that. But clearly people love getting Taco Bell delivered.
Of course. Well, that's also, I mean, it now makes sense to me in a weird way because you save a lot of time going back and forth. You can use that time to do all this other stuff. You can use that time to work. And I think more people are familiar with those kind of calculations now and all that stuff.
Yeah. I think there's a lot of things, Jack. I mean, one of the things that I remember even in 2013 looking at is just this marvelous fact that only goes in one direction. There's a few of them. One of them is that if you looked at food consumption in the 1950s when the US government used to measure this, or when they first started measuring this, something like 70 to 80 cents on the dollar was spent on grocery. This is in the 1950s. If you looked at 2013 when DoorDash was founded, that number was getting closer to 55 cents towards groceries, 45 cents towards restaurants. Today, it's closer to 55 cents towards restaurants, 45 cents towards groceries. And so over a 75, 76-year arc, yes, there's ups and downs, but if you look at the trend line, it kind of goes in one direction, which is in the direction of food prepared by somebody else.
Do you know by chance if the relative cost of a steak you made through your own groceries versus a steak prepared by somebody else, if that ratio of cost has one gotten more expensive relative to the other?
It's a great question, but I think it depends a lot on how you value your time. So this gets me to another fact that I think is pretty interesting, which is if you looked at the percentage of dual-income households, same time period, 1950 to 2020. So this goes way beyond AI or COVID or any of this sort of stuff. You find that the percentage of households has gone from like a quarter, dual income, to almost 70% present day. If you look at the number of restaurants as another example, just total number of restaurants, total count, in the 60 to 70 years in which this is measured, there's maybe two years in which the total number of restaurants in the current year doesn't exceed the previous year. So even if new restaurants, which is a very tough business as you know, I mean, having worked as a dishwasher at my mom's place, I definitely know how hard it is, yes, there's a risk of going out of business, but there's always more than enough that replenishes the bucket every successive year. I think when you put together some of these facts over 70-plus years or something, they kind of spell out that people, whether it's through dollars or time, are expressing the fact with their activity, not just their words, but their activity, that they must value if somebody else made them food.
