Dara Khosrowshahi on replacing Uber drivers — and himself — with AI

22:44We were able to charge a premium, give it to the driver, essentially to up reliability. and we started building the habit of this is an on-demand service to actually this is more than an on-demand service and i'm going to think about scheduling things in my life often having to do with travel now what we're finding is actually some people are hacking reserve if you want to call it that for reliability so if you're in west jessica county and armank and the liquidity for Uber is lower. You may not want to use on-demand for your commute, but you can use reserve for your commute as well. So what started as, let's try this for travel, is now being used to hack reliability to some extent.

23:31That insight of reserve building, and we've been at it for four to five years, reliability is not perfect perfect, but it's 99 % now. And we're always kind of working that trade-off between reliability and price because we want the price premium to be as low as possible, but you can't lose too much reliability. That insight led us to believing that you actually can move from on-demand to scheduled. And the offerings, the Uber One kind of discounts, we think will hopefully over a period of time change behavior. So you actually come to Uber to reserve your booking in advance. We don't think this is going to be a last minute thing.

24:12Like if you get to a city and you don't have a hotel, I mean, there is something wrong. Maybe it'll be there on a cancellation basis, but we are trying to drive reservation behavior and we've demonstrated previously that we can.

24:26Nilay Patel:Yeah. I feel like hardcore travelers who know to reserve an Uber, who are some of your best customers, they like price shopping hotels. Yes, totally. And there's a lot of credit card points. And my sister's a credit card points person. Yes. It's frankly a little terrifying, but she's really good at it. How are you going to compete with that? Because that's the customer. In my mind, the customer who knows to book a hotel and Uber is also the person with five different credit cards trying to get the best deal. And they know that this portal is where they need to go at this time. How do you compete with that?

24:58So I actually had an earlier interview with the points guy. and I asked him, what's the best credit card for travel? Because I was curious. Yeah, yeah. Turns out Amex Platinum, according to the points guy, is the best credit card for travel. And by the way -

25:13Nilay Patel:I don't believe you because this worked out too well. I'm just letting you know this worked out too well. It was amazing. And we have a great relationship with Amex where you get benefits and free bookings on Ubers as well. So it's actually, there's a lot of layering that we're doing. If you've got Delta, SkyMiles, You can get Delta SkyMiles for booking on Uber. We have a relationship with Marriott Bonvoy. We've got travelers using Uber all the time. We've got the Amex Platinum card, the best card for travelers as well. So I think we have kind of the right elements coming together to get some percentage of our Uber One members to try the booking experience.

25:53And then we'll go from there. And I do think that this would be a failure if it ends with hotel booking. You know, one of the pieces of magic that Uber brings is it's actually the back-end experience. You know, one of my learnings when I was at Expedia, it was basically a booking. You know, after the booking, there weren't that many services that Expedia offered other than if something went wrong and, you know, you do everything you can to help the customer. But actually what we can do is kind of connect all these logistical elements of your travel. uh so obviously you know you're uber to the airport if you did your hotel booking we already know where your uber is maybe we'll give you a discount uh to the hotel and i'm hoping that as we build out travel we can actually improve the in market experience i don't know about you but like why do i need to check into a hotel like what's the deal with that right like you know i've got my phone and if you have a hotel booking like maybe you can walk into the hotel and we can give you all the information and you can just go up to your room and maybe your app can act as a key, et cetera.

27:04There's a lot more that we want to do in terms of the in-market experience. And it's something that Uber is uniquely positioned to do because we're already in market in almost every city that you're going to want to travel to.

27:17Nilay Patel:There are competitors in these markets. Expedia is an interesting partner because you used to be the CEO of Expedia. I assume you just made a phone call and said, Hey, what's up? It's me. So actually, uh, I had to recuse myself from the process, uh, from the process entirely. The idea, the strategy, let's get deeper into travel. Obviously I was, uh, I was involved with, but because of the conflict, I'm still on the Expedia board. I had to recuse myself from the process. The team ran it and I'm like, guys, what's going on? They're like, we can't talk to you. So they got to Expedia one because of the great job that that team did they got no help from me i'm sorry the ceo of expedia wasn't like i got a board member breathing down my neck but it wasn't i had to like recuse myself in those discussions it was a little awkward but it all worked out well so obviously expedia would be a competitor but they're your partner there are other competitors there are the hotel loyalty programs booking.com exists they say the same sorts of things that you say of course they've been on the show saying literally the same sorts of things connected trip i think they talk about right all the time yeah uh why I think a lot of people like checking into the hotel.

28:20Nilay Patel:The free water especially is very useful when you arrive in a new hotel. That piece of the puzzle where you're going to connect everybody's backend systems together and build one unified experience where the Uber app is the primary interface. I could abstract that away and say, well, that's everything. That's what OpenAI would like to do. That's what Google would like to do. Sure. Why is Uber going to win that fight? Well, I think it's a different question or service offering in terms of offering the availability of the service, but to the extent that you can actually deliver it in market. OpenAI is an incredible company.

28:56They build a lot of cool things, but they don't live in the probabilistic real world that we live in. There's a Mike Tyson saying is like everything is theory until you get punched in the face. Everyone has a plan until you get punched in the face. Everyone has a plan. And, you know, we get punched in the face daily, which is drivers are canceling, riders are having issues, et cetera. Deliveries are late. And so we already deal with this probabilistic world on the back end where things go wrong all the time. And it's one thing to try to chain all of these events together and get the logistics right.

29:36but to adjust to real-world traffic conditions, cancellations, road closures, all of that stuff we do daily. So I just think we're much better equipped to actually fulfill this seamless, delightful end-to-end experience from planning to booking, making it incredibly easy, and then to delivery, the actual experience on the ground.

30:02Nilay Patel:You know, your partnership with Marriott, for example, Marriott wants those to be their customers. You're the app that everyone's doing everything in that relationship gets intermediated. Is that a tension? I mean, it's a tension at the same time. It's a tension that everyone deals with, right? Marriott competes with Expedia. To some extent, you could argue that they compete with us, although we're a much smaller player today in travel. Maybe we'll get bigger. We work with Starbucks at Uber Eats. and of course they'd rather have people come direct to their app but the fact is that uber eats brings them a lot of incremental demand as well so this coopetition theme is something that many many players have been uh comfortable with for many many years comfortable with for many many years is in one context right everybody has an app and it doesn't really matter you're all going to open the apps and maybe we can get you to open our app now you're in a world where you're going to open an app and maybe an agent's going to go off and do something for you and the idea of being the everything app in that context, Uber is describing this as a step to being in everything.

31:02Nilay Patel:Totally. It's in the press materials. Yeah, yeah. Brian Chesky was on the show. Airbnb is going to do concierge services for travel. Yes. And they're going to get way out of their lane. And maybe that's working. I haven't talked to Brian in a minute about it. OpenAI wants to be in everything. X, famously, is already the everything app, as you know. Yeah. We're all using X all day long for everything. Do you think the pressure on needing to be that interface is going up because of AI? I think the pressure is going up to some extent, but I think AI is making it possible in a way that it wasn't possible previously.

31:35One is these models can adjust to real-world conditions in a way that determinants that code can't, right? That's really cool. Whereas you had to build UI interfaces that were tight and relatively limited, AI is allowing for an interface that is unlimited, essentially. You know, you can just tell the app what you want. And you can have agents then take that and break up that request and try to deliver it as best you can. So AI is making possible now. And by the way, you can just build much faster. So to going to smart risk, the cost of taking risks is going down. So I think all of that is coming together in an opportunity set that I think a lot of companies recognize, including us, including Airbnb and the other companies.

32:28And it's going to be a race to many of these new markets. And we're confident. We've raced before. We love competition. But this is another trillion-dollar-plus opportunity. And we've done well with mobility. We've done well with delivery. All of these businesses have been built organically. So I think there's kind of a builder mindset at Uber and we're going to give it a shot. And so far, the signal is pretty, pretty damn good.

32:54Nilay Patel:We have to pause here for a quick break. We'll be right back.

33:05Nilay Patel:Support for the show comes from Upwork. Think of the fastest growing businesses, you know, you might think they've gotten where they're at just by doing more. But that's not always the case. Chances are they're just delegating smarter. Upwork helps you bring in expert freelance help fast so you can delegate and keep moving. Upwork is a one-stop platform to find, hire, and pay expert freelancers. Find specialized talent across web and software development, data and analytics, marketing, business operations, and more. Upwork also has Business Plus, which gives you access to the top 1 % of talent on their platform.

33:42Nilay Patel:With AI-powered shortlisting, you get matched to the right freelancer in under six hours. No endless searching required. And when it comes down to contracts and payments, don't sweat it because Upwork has the operational stuff covered. It's free to sign up and posting a job is easy. Visit upwork.com right now and post your job for free. That's upwork.com to connect with top talent ready to help your business grow. That's U-P-W-O-R-K dot com. Upwork dot com. Support for Decoder comes from Adobe. For every big idea, your documents folder tells a story. Let's say you've just finished pulling together a brief.

34:26Nilay Patel:So you hit export on finalversion.pdf. But then you open the file and you immediately notice a typo. Several versions later, you're exporting finalv4.actualfinaldraft.pdf. Adobe Acrobat can save you the digital clutter with PDF Spaces. It takes your documents and turns them into a living project that you can engage with, get insights from, and collaborate with others on. You can gather all your files into one workspace and have a whole conversation with your AI assistant about it and ask questions to get deep insights about your project. You can even invite people to your PDF space and let them add files, comments, notes, and more.

35:09Nilay Patel:You could doodle in the margins or even turn your project into your own personal podcast episode. Acrobat lets you generate an audio overview of your project in just one click. Learn more at adobe.com slash do that with Acrobat. Support for this show comes from Doppel. Maybe that ping you just got is an urgent message from your CEO. Or maybe it's a deepfake trying to target your business. Doppel is the AI-native social engineering defense platform fighting back against impersonation and manipulation. As attackers turn to AI to power increasingly sophisticated strikes, Doppel uses it to fight back.

35:48Nilay Patel:Their digital risk management dismantles attacker infrastructure, while human risk management builds team resilience through simulation and training. With automated takedowns, multi-channel coverage, and AI defenses that build intelligence with every fight, Doppel works relentlessly to protect people, brands, and trust. Doppel, outpacing what's next in social engineering. Learn more at doppel.com. That's D-O-P-P-E-L dot com.

36:24Nilay Patel:welcome back i'm talking with uber ceo darakasoshawi about how everyone seems to want their agentic ai to call you an uber why that's maybe not such a great idea last time you're on the show we talked a lot about agents and accessing uber as a service inside of an agentic workflow i will tell you i asked a lot of ceos that time yeah yeah this question everybody who had a physical product was like, we'll be fine. And then it was Amazon who has an interface to a bunch of dropshippers that is like filing the lawsuits, right? They have a virtual product. Everybody who is in the world of Adams was like, go ahead and try, like try to make another Uber.

36:59Nilay Patel:You just give it a shot. We'll be here when you're waiting. That was very much your attitude. What you said to me was the price of calling an Uber and chat GPT should be zero until they prove it's valuable. And then I'll figure out what the rate should be. It's been a year. Have you seen any meaningful uptake of calling Ubers from ChatGPT? No, no. And it doesn't seem to be at this point a priority for a lot of the foundation model companies, whether it's ChatGPT or Gemini. I think they're experimenting with it. But I think the enterprise market is growing much faster than anyone thought that it was going to.

37:37So I think there's been a pivot towards enterprise. And by the way, Rightly so, based on the growth rates that we see based on our internal usage of these foundation models. So at this point, that part of the market hasn't developed. And the cool thing is we're building some really cool products. You know, you can scribble a shopping list. You can take a picture of food that either looks really tasty and we'll put together a shopping list for you. if you tell us what merchant you want to go shopping at, we'll put together a list for you and we'll get it delivered automatically. So a lot of these experiences that I think people thought you'd find on OpenAI, et cetera, you're actually going to find first on an Uber.

38:22I wouldn't be surprised if it's built over a period of time, but right now enterprise is coming first and you could argue rightly so.

38:29Nilay Patel:Uber is a favorite of agentic demos. You pop up all the time. I'm just going to go down the list. Is that right? Yeah. It's kind of an everyday use case. It's great. Google and Samsung announced Gemini Task integration on the newest Samsung phones, where the model will literally open the Uber app in the background in a virtual container and click around it to get you a car. Have you seen any meaningful rides from that integration? Not yet. Not yet. But we'd be delighted to see it. We want to bring more experimentation, more opportunity for our drivers. It's just really small now. It doesn't mean it's not going to be big 10 years from now.

39:08Nilay Patel:I'm just, we had a whole year of these. Totally, totally. Alexa, has Alexa sent you any meaningful rides? No, small, very, very small. Okay, and I can keep going, but it seems like the answer - Have you used any of these products? I have to, I'm required. How is it? You know, I think they all have the problem. They're slower than me just doing it myself. Like kind of down the line, they're slower than me just doing it myself. Also, I'm only ever calling a car from work to home or home to work or to the airport. The app is one tap away for all of those experiences. And it's pretty easy to use. Now, I do think that one area that, for example, we are looking at is while the front end, the initial demand may come from any agent, I am going to want our pixels in front of you.

39:53So for example, I'm perfectly fine with OpenAI calling Uber, but then I want in that web interface and within that ChatGPT app, kind of the Uber Pixels and the Uber brand so that you know who is fulfilling that ride for you. So, you know, we'll see how things turn out. If you're an Uber One member, you're going to want to use our product, especially for travel.

40:16Nilay Patel:I mean, again, this is the fight that I've seen coming where getting people out of your app and just using Uber as a backend service, as a commodity against every other service, pure play or not, nobody's going to want this. But it seems like they've all pivoted to enterprise so fast that that fight is delayed or maybe never coming? I think it's delayed. It's going to happen because I think the size of the prize is too big. Now, if you talk about kind of history rhyming, not repeating itself, there's some of what I went through in my former job at Expedia. If you remember during those times, there was a big debate about Metasearch, right?

40:54There were these Metasearch players, Kayak, TripAdvisor, Trivago, that were amalgamating a bunch of travel content and there was a point at which meta search was quite powerful in terms of customer acquisition etc but as the supply consolidated really the value started accruing to the suppliers much more than uh the meta players and you know the travel business consolidating to xpedia booking.com airbnb there's more but the three very very big players so i do think also on the supply side, when you look at mobility, when you look at delivery, there's usually two or three players in every market.

41:33So even if you get that front end being particularly big in a consolidated, let's say supply marketplace and with our size and scale, multi-platform, all the countries that we operate in, I think we're going to be more than okay in terms of kind of the leverage and the negotiations that happen. I always try to push the negotiations to the back end, build a great experience, figure out kind of the balance, the economic balance later. But sometimes you've got to figure that stuff out up front.

42:04Nilay Patel:This is a slight difference in the last time you were here. And I just noted that companies are all different, not Uber, but the AI companies, they're all in a slightly different posture than they were a year ago. Yeah, totally. Right. They're racing towards IPO. They are constantly calling code reds. Like every week, it's a code red at OpenAI. It's a cool thing to do. Yeah. I mean, we've had CEOs come on the show and say they've called it code red. I'm like, did you actually do it? And they're like, no. We've definitely had our share of code reds. And there's a danger of code red fatigue in companies too, because then it becomes meaningless.

42:37So it's a real issue.

42:39Nilay Patel:OpenAI was a partner of yours. You've obviously launched things with them. You've used the products. As you broadly think about, okay, we're going to build AI services, we need a model provider, do they feel like a stable partner? Yes. Their products are excellent. For example, we've used, I think, ChatGPT 5.5 for some of the cool stuff that we demoed today in terms of a shopping list or taking pictures, et cetera. Codex is something that a lot of our devs use. OpenAI has been a strong partner in whatever drama that you see in the markets isn't showing up in terms of the quality of their product.

43:19They continue to be first right.

43:21Nilay Patel:The drama in the market is all encompassing. As you and I sit here today, Sam Altman and Elon Musk are in a courtroom arguing with each other. Listen, it used to be Uber when I was looking at joining the company. It reminds me of that. And we got through it. We got through it, and it's a great company now. And I think that it's an adjustment that every company has to go through. So many people are interested in how OpenAI does because it's an important company in the world. So they'll get through this. Do the model companies feel interchangeable in a way that has always seemed like a small danger here?

43:54I think interchangeable is a little bit too strong a word. I mean, I do think that what Anthropic is building, Claude, is it's spectacular. Like our developers are using it all the time. Codex is definitely picking up use of our developers. Now, what we do do is we use some of the frontier models and some of the more advanced models to pilot, build demos if you want to build something quickly. And then what we do look to do is we have, it's much more than an API layer, but we've got a platform, Michelangelo, that has all the data feeds. And then essentially you're able to switch models. And early on, when we're trying to explore something, we will use some of the more advanced models.

44:40But then once you get up to larger volumes, we will try to switch out either cheaper models or open source models to control kind of the costs and the token costs on the backend. Interchangeable is too strong a word, but we definitely experiment with various ones. And at this point, nothing is hard-coded into our systems. And frankly, we're going to make sure that none of them are hard-coded into our systems. Right.

45:06Nilay Patel:That seems like a hedge against the companies and their needs and also cost, right? The cost of tokens is still quite high. Yeah. I mean, you never want to be overly dependent on one technology unless you're highly confident or it is very, very, very proprietary. And part of it is that all this stuff is so new. I mean, you and I were talking about Cursor last year, right? And Claude wasn't a thing, at least internally. Now Claude is really, really increasing at incredibly surprising rates internally. So early on, as this market is developing, we want lots of experimentation. And we want to give our devs the freedom to try a bunch of stuff.

45:47I don't want this to be top-down. They'll shout here or there. Of course, there's going to be optimization, but right now there's a lot of experimentation going on internally.

46:00Nilay Patel:We have to take another short break. We'll be back in just a minute.

46:21Nilay Patel:Starting a side hustle or building the next big thing, Hostinger lets you go live in minutes, not weeks. Hostinger is an all-in-one platform that brings everything into one place. Your domain, website, email marketing, AI tools, and AI agents. You can create websites, online stores, and even custom apps without coding or design skills. Then use AI agents to automate tedious tasks and help grow your business. Turn your one day into day one. Go to hostinger.com slash decoder to bring your idea online for under$3 a month. Plus, get an extra 20 % off with promo code decoder. That's less than the price of a cup of coffee per month.

47:09Nilay Patel:That's hostinger.com slash decoder. Promo code decoder for an extra 20 % off.

47:20Nilay Patel:Support for today's show comes from CNN. Do you want to live forever? Influential journalist Kara Swisher is taking a hard look at the longevity industry to separate the influencer hype from evidence-backed science. In her new CNN original series, Kara's talking to Silicon Valley power players and trying out the latest in anti-aging technology to see what works and what's a waste. Kara Swisher wants to live forever. New episodes streaming Sundays with a CNN subscription. Go to cnn.com slash subscribe to start watching. Support for the show comes from AWS. How much of your workday is actually work?

48:04Nilay Patel:And how much is just hunting for information? The answer you need is buried in a Slack thread. The data's in Salesforce or in an email from two weeks ago. By the time you've pulled it all together, half your morning is gone. That's the problem Amazon Quick was built to solve. Quick is an intelligent workplace assistant that connects to all of your systems, your documents, your dashboards, Salesforce, Jira, Slack, email, and gives you complete answers in seconds. Not links to dig through, actual answers with full context. And here's where it gets interesting. Quick doesn't just find answers, it turns them into action.

48:43Nilay Patel:Create a deck, update a ticket, send a message, right there in the conversation without switching tools. It's AI that actually works the way you do. Learn more at aws.com slash quick.

49:05Welcome back.

49:06Nilay Patel:I'm talking with Dara Koshashawi, CEO of Uber. about just how weird the experience of running a software company is getting right now. Let me ask you about running a software company in 2026. This is the thing I was most excited to talk to you about. It is true. The last time you were here, we were talking broadly about AI and had all these questions about agents and the big labs coming for you with their consumer chatbots. Maybe that's not happening yet. The thing we did end up talking about, just as you were walking out, is you said, we had GitHub Copilot, but all the engineers want to use Cursor.

49:39Nilay Patel:and now you're saying and cursors around but they're all using quad code or maybe they're using codex they're the increase in clawed code usage and sometimes the replacement of cursor usage is fairly remarkable we use both they're both terrific products and then there's there's a group that's that's using codex and they're all really good and i'd say the big change is with with cursor it was you know coding and coding um assist so to speak complete, but now these agents and agentic coding is something that is, it's just blowing people away. It's very, very cool. And when you say blow people away, I would say many of your peers have gone crazy.

50:21Nilay Patel:Like they have seen agentic coding, it's looked them in the eye and they have responded by losing their minds and saying that the entire structure of a company should change around this. I'll give you some examples. Meta is reportedly going to have teams or 50 people report to one manager. Jack Dorsey can't lay off enough people fast enough. And his goal, he said this out loud, he wants all 6 ,000 people agentically assisted to report to him at Block. I don't even know how you would... It's a show about org charts. And I read that and I thought, well, our show is going to keep going for another decade.

50:56Nilay Patel:I know what kind of... We're on the cusp of the weirdest org charts in history. Yeah. Yeah. Are you there? Are you saying, okay, okay, agentic coding is going to fundamentally change how you construct a software company? We have not gone and examined the fundamental org chart of the company yet. I'm not saying it won't happen. We are pushing the company hard. And by the way, I've got to push a company harder to go to first principles in terms of how you work, period. What we found is, and again, And our culture is like bottoms up, let people do a bunch of stuff. And listen, the engineers are using it, the debugging, like all the cool stuff is happening as it should.

51:43But what we saw is like in sales, right? Salespeople now use agents to summarize information on a client that they're going to call to build out a really cool presentation. communication we're using agents and ai though i would describe around the edges of how we work so that's one and and we're not kind of thinking about well let's think about the sales function from the bottoms up customer service is another example where you know we've got agents who generally follow policies there's a policy if you're an uber one uh uber one member and you your order is delayed by 20 minutes, we're going to give you 15 bucks back because you're a loyal customer, et cetera.

52:32That's a policy that's in place. And there are agents that are following those policies, et cetera. Human agents. Human agents. Human agents. And we then said, well, let's build virtual agents to follow those policies. And it turns out that actually our policies on a global basis, the documentation is complete crap, to use a technical term. And what happens is an agent, a human agent, I'll be sitting next to you and be like, hey, what does this policy mean? It's kind of unclear. And you coach me and then I figure it out. Like humans are quite flexible. When we had AI agents go through these policies, they just went nuts, right?

53:09And so one approach was let's rebuild all the documentation and policies the right way. And then let's have the agents work based on these policies. But why do we put those policies together in the first place? It was to get to goals and outcomes based on standardized ways to get to those outcomes. I don't want to go bankrupt, but I want to keep you, the Uber One customer member, happy. And so we made a policy to approximate the optimal outcome for the population. But now I can just tell the agent what that outcome is. I want actually to be fair to a person. I want Uber One members to be happy, et cetera.

53:54I don't want to go bankrupt. So the approach that we're taking now within customer service is throw away the policies, describe to the agent what you're trying to accomplish, and then let the agents go. And obviously train them on good interactions, bad interactions, and give them feedback, et cetera.

54:11Nilay Patel:Wait, can I ask you just a foundational philosophical question? Yeah. Why trust computers to make those determinations and not people? Uh, because the model can learn based on the population of everything that is happening versus an individual human, just learning based on the experience that he or she is happening that, uh, having that day. Uh, and models are easier to track and tune than humans are to train. Okay. So this is a scale answer. It can see all the data. Yeah. So you can just describe a generalized outcome and we'll just, you can, you can retrain based on that data and you have perfect visibility into the actions reactions and you know the retraining output you don't have perfect visibility into but you can kind of iterate around that so it does demand a different approach and it's a little bit back to what you and i were talking about which is a smart risk it's a riskier approach like we got to throw stuff out and just completely rebuild in a different way and i'm really glad like it wasn't in this case it wasn't me who pushed the costops team to throw everything out.

55:13They were frustrated with the results of it that they were seeing early on. They're like, we have to be able to do better. We're going to try this out. The signal looks really promising, but I can't tell you it's actually going to work in the end.

55:24Nilay Patel:That kind of dynamic customer response. In terms of pricing, people are making it illegal in this country to do dynamic pricing in that way because it feels unfair. Yeah, we're not going to – that is actually an issue, which is what we don't want to do is have different outcomes based on targeting you versus another person versus another person. But you can have different outcomes because there were circumstances that were different. So for you, if your food was 15 minutes late, another person, you're both Uber One members, another person's food was 45 minutes late, you could actually have different outcomes because actually the circumstances are different.

56:07So it's not based on targeting or optimizing based on targeting. It's optimizing based on context.

56:13Nilay Patel:That's really interesting. It strikes me that we could probably do another whole hour on we wrote a bunch of rules for humans, and now we have to write a system prompt that isn't the rules. It's actually the outcomes that you're trying to get at. Yeah. Yeah. That's another hour. We'll see if it works. We'll see if it works. You're going to come back next year and I'm going to ask you if it works. Yeah, exactly. But let me ask you just more at the base level. When I think about software companies generally, the creative tension of any software group is you have a PM, you have a designer, you've got some engineers.

56:41Nilay Patel:Yeah. They all want to be in charge. They all think they're going to do it right. And they all need to work together. Yeah. And if you can get that right, it's magic. it feels like with the power of vibe coding everyone is going to try to do everyone else's job and no one's going to be good at it and that is we're all it's all a mess i can see it happening totally are you rethinking that basic triad inside of your book so it depends on the kind of project that you're working on there are some larger projects that you need design you need proper planning etc but we are having some product team members whereas previously if there were some simple bugs in the code or very, very simple features, they would have to then prioritize it with their engineers, et cetera.

57:26Now they're just going in and they are vibe coding and an engineer is going to review it, the code, but essentially the product person is going direct into the code base, so to speak, or going direct with an agent into the code base. So I do think for simpler problems, smaller problems, the dynamics are going to change. We're going to try it out. We're going to see what happens.

57:50Nilay Patel:When you look at a company like Meta, which seems to just be in the midst of endless rolling layoffs, they're saying it's because AI is making everybody more productive. It might be because they're just freeing up CapEx to go spend on whatever they're spending CapEx on to whatever end that Meta is going to do AI. Super intelligence, I'm told. Are you in that same spot where you're like, we're getting more productive, I need less people? No. um we are my view is if an engineer is going to be 50 or 200 more productive i want more engineers like there there are the list of ideas in terms of what we want to build so outscales our our throughput at this point that generally we are looking to add more engineers uh to our employee base.

58:37Now, there is a trade-off, and we are dealing with a trade-off right now as we speak, which is, I don't know if you saw it, but our CTO was talking to a reporter and made a comment, which is true. We have blown through our AI token and infrastructure budget for the whole year in about three to four months. And it was a big thing when that happens, but it happened. And the trade-off is going to be headcount. So we are budgeting differently. Previously, you would have a headcount budget or plan. You know, doesn't mean it would actually happen, but it's a plan going in, you would have an infra budget.

59:20Now there's an active trade-off going on between the two. And to the extent that we have overages in terms of token spend or infra spend, which theoretically those overages are products that are being built and our productivity that's being added to our engineers, we're going to hire less aggressively, so to speak. That is a live trade-off. How far it's going to go, I don't know at this point.

59:43Nilay Patel:Are you all the way at, I'm spending so much on tokens that it's costing me more than hiring one junior engineer? We are spending a lot on tokens. I haven't done the math yet, but it's significant. But the throughput is really accelerating. So at this point, it's something that needs to be managed. And I do think it's just taking different muscles. The way that we're managing budgets, it's just, especially on tech, is fundamentally different than how we did three, four years ago. All right. Well, and that's the AI question that I want to talk to you about autonomy, which is also AI, but in a very different way.

1:00:14Physical world, AI, yeah.

1:00:15Nilay Patel:You were on Dair of a CEO, and you said the employees at Uber have created an AI version of Dara to practice presenting and pitching to. Is that real? And how close are we to AI replacing the CEO? So it is real. I have not witnessed the Dara AI, but it is real. People have done it. Honestly, I don't know how good it is. It's clearly not as good as a real thing. I mean, come on, how is that possible? Decoder listeners, every time we do an AI episode, they say the AI should replace the CEO. It is a reflexive comment we get. I'm not there yet. I think that the AI-powered CEO is going to be better than the AI CEO.

1:00:57I think there's a magic in terms of teaming up humans with AI and with agents. And based on what I see, that is a superior product than pure play AI or pure play human.

1:01:11Nilay Patel:You should recuse yourself from this. You have a deep conflict of interest here. Of course I do. I'm hoping the board sees it that way as well. Maybe the board is planning this and I had no idea. I mean, that would be, in keeping with the Uber story, that would be there. How is AI changing our board processes? I've got to think about that one. Oh, trust me. I get those pictures. They're very good. You don't want anything to do with those. Let's talk about robots. Yes. Actual robots. Actual AI in the world. Uber has made a bunch of big investments in robotaxis. I want to start with Rivian. It's over a billion dollars.

1:01:40Nilay Patel:I think it's$1.2 billion in total commitment to Rivian over some number of years. I just have a really basic question. You announced partnership in March. You're going to buy up to 50 ,000 fully autonomous R2 robotaxis by 2031. But it's also called an investment. And I'm just doing the math. I'm like, that's the price of the R2 platform. You're just buying a bunch of cars. Is buying a bunch of cars an investment or actually getting equity in Rivian? So we actually invested in Rivian equity and we've invested a number of our partners. Usually we will invest in our partners in a Lucid, in a WeRide, in an AvRide, for example.

1:02:16So it is an investment and it's a vehicle commitment as well. It's both. And it's based on deliverables, obviously. They've got to deliver and based on everything that we've seen from RJ and team, putting together a first first class uh ai team and and we're confident that they can deliver on

1:02:32Nilay Patel:those r2s yeah uh the deliverables are very vague i'm just gonna read you the press release uber will invest up to 1.25 billion in rivian through 2031 subject to and i quote the achievement of certain autonomous milestones by specific dates well they are very specific contractually this into five different AI systems and no one can tell me what they are. And they're fairly fuzzy as far as what you know. What are the autonomous milestones? I could tell you, but then I'd have to kill you. The reason I'm asking is not, I mean, I desperately want to know the specifics. I'm looking at this industry in total.

1:03:07Nilay Patel:And I will tell you that we've thrown out whatever autonomous milestones we used to have, the level system that everyone used to talk about. That's all gone. No one cares about this anymore. No one's like, we shouldn't do level four. We're doing it. And I can't quite tell you when a car, what milestone an autonomy platform has to hit before I can say this is a robo-taxi. So, I mean, usually I'll give you examples of milestones, not specific to Rivian. Usually there's a milestone, for example, if you release in market with a vehicle operator. Usually one other milestone may be if you take the vehicle operator out.

1:03:44You can only take the vehicle operator out to the extent that you complete a safety case that we put together along with the autonomy provider. then another deliverable might be delivering a certain number of cars that are NVO capable that have a redundancy at a certain bill of materials as well at a certain cost. So those are examples of deliverables that have to do with either capability or economics because ultimately this is about going to market with a product that's proving to be a very, very popular product.

1:04:19Nilay Patel:Your big partner in the past was Waymo. Yes. Waymo has gotten there in many cases to some of the kinds of milestones you're describing. You're obviously different side away from Waymo. You've got the Rivian deal. You mentioned Lucid. You're going to buy at least 35 ,000 Lucid vehicles designed exclusively for use as part of Uber's Robotaxi. Yeah, and a partnership with Nuro. And a partnership with Nuro, which is the platform there. Yes. And overall, you're going to commit some$10 billion to autonomous efforts. You launched Uber Autonomous Solutions. that feels like a bet on this is happening, but we don't know who's going to win.

1:04:54Nilay Patel:You're diversifying. So it's a little bit different from that in that we believe that it is going to happen. And we believe that just like there isn't going to be one foundation model to rule them all, there isn't going to be one physical world foundation model to rule them all. And all the evidence that we see is, Yes, Waymo is past the finish line. They are the leader. They are, in many ways, an inspiration for many, many companies in this industry. They're a great partner of ours in Atlanta and Austin. There are many other companies that are getting to the finish line. A WeRide, for example, or a Pony.ai or a Baidu, these are Chinese companies, are already at the finish line.

1:05:40And we are in market, for example, with WeRide in the Middle East. and there are players like a neuro or a wabi or an avride or a wave all of whom are accelerating to the finish line and if anything the speed of getting to the finish line is accelerating one model capabilities are much much better now used to be kind of deterministic you know kind of code that you had to slog through now obviously it's learning ai models sim capability is much better so that data will go much further in terms of model training. And what we're trying to do with AV Solutions is we're trying to build out the whole necessary ecosystem around these companies so that they can focus on what they do best, which is training these models to get them to be superhuman safety.

1:06:33We can help them get there, for example, with data collect. and we can both kind of then get to market as quickly as possible. So it's not, I would say, a diversification bet. It's a bet that there are gonna be many players. And as a platform, we've always been supply-led, which is the way to grow our platform is to build out supply, whether that's more drivers or whether that's more restaurants or more hotels. then we're able to, as we build out liquidity of supply, demand shows up. And just like we want every safe human driver on the platform, we want every safe robot driver on the platform, whether that's a Waymo driver or a Neuro driver or an AvRide or a WeRide.

1:07:18It's a bet that we're making, which is there won't be one physical AI model to rule them all.

1:07:24Nilay Patel:There's some real confidence in this bet. I've talked to a lot of rideshare CEOs over the years, a lot of autonomy CEOs over the years, and it's always been 10 years away. The confidence I'm hearing from you is, oh, this is happening. We're spending a lot of money to get there faster. All the evidence we see is that it's happening. And, you know, you can, Waymo has shown the way. A lot of Waymo engineers now are working in other companies. For example, the Chinese players have shown the way. And you've seen it, the speed of foundation model development, whether it's digital foundation models or physical foundation models, you know, NVIDIA is betting on this as well.

1:08:05So, you know, these are big bets made by capable companies and we think we're on the right track here.

1:08:11Nilay Patel:In the context of our conversation, I'm going to bring up the tradeoff. Sure. By saying it's going to be more real, you no longer get to kick the can on we're not going to have drivers in the cars, which famously got Travis Kalanick in a lot of trouble by saying I want to get the driver out of the car. Long, long ago. Yeah. Because autonomy was so far away, we just didn't have to solve this problem. Mm-hmm. You have been on podcasts recently saying, oh, this problem is here. I don't know what's going to happen to 9.5 million Uber drivers when autonomy comes. You literally said, I don't know to Stephen Bartlett.

1:08:41Well, if you don't know, you should say it. You know, it's now here's what I know. 10 years from now, I am 90 % certain that we're going to have more drivers on our overall platform than we do today. Now, I don't know if that's going to be true in San Francisco. but with the way that the business is growing and the capability of building these cars at the right bill of materials in all the markets that we operated not just the high cost markets we're going to have plenty of drivers and and we also are actively looking to build out more use cases for drivers that are more complex you know one of the announcements that we made was personal shopper right?

1:09:23It was Courier. People started hacking Courier, asking Uber Couriers to go shop for them. So we decided to productize that as well. That's a very, very complicated interaction. It's a random store, take a picture of the goods. This is what I want. So we're building out much more complex use cases for humans to migrate onto as more of the work is being automated. 20 years from now i don't know what that's going to look like because then you really start increasing capabilities and i think these are big societal questions it's going to be true of white collar workers and it's going to be true of certain kinds of blue collar work as well and you know i think ceos should talk about this not not in a way to like scare people but we should also be honest about it, which is I've never seen a wave of technology that has direct impact on how companies work and how people have worked with the accelerated pace that I'm seeing today.

1:10:24Doesn't mean that society can't adjust, but the pace of change here, it's pretty remarkable.

1:10:31Nilay Patel:One of my theories about the extremely negative polling on AI is that it's fundamentally an enterprise technology. You've described this even in this conversation. The frontier models, those companies are moving to enterprise use cases. You at Uber are using them in enterprise context. And there are not great consumer products in front of people. Not yet. Yeah. I haven't seen them. Maybe they're coming. I mean, listen, we're trying to do that. And it's these moments of surprise and delight where, you know, you can talk to your Uber to get an Uber, lots of complex situations. You can transcript a shopping list, take a picture.

1:11:08Nilay Patel:I don't think that stuff is going to change the overall polling on this is a threat that will take my job away. Listen, if it's your job, I think you're right. Yeah. And so this dynamic of everybody is showing up saying the jobs are going away and mostly because it's so good at writing code, right? Like this is a weird kind of disconnected dynamic for regular people. Uber needs customers. You need people with money to want to ride around. How do you see that economy playing around? So I think that it, right now, the talk is louder than what we see in the market, right? The economy remains robust.

1:11:45The consumer remains robust. We don't see white collar people out of work at this point. So I just don't see it in market. Now, the fear that you see might be a leading indicator of what's to come. but at this point i see no signal in our actual business that it's having an impact on

1:12:05Nilay Patel:consumers at large what do you ascribe the extremely negative pulling around ai2 um i do think that it's some fear-mongering from the press they love the drama yeah are you are you part of the press or no a little bit okay can i call can i have this level of influence You can point at me all you want. But listen, it's a conversation that people are constantly having. It's a dramatic conversation. And I do think machines replacing humans has been a theme for eons, right? And what you do see in manufacturing, for example, with automation is that machines complement humans. And then there are other capabilities that humans always adjust to.

1:12:50It's just things are moving so fast now that I think the fear is it's out there. I've got 14-year-old twin boys and two other older kids. My 14-year-old kid is like, dad, why should I study? I'm not going to be able to have a job. Like we had a, and I was just blown away. My 14-year-old is asking me now, maybe he don't want to study. This does feel like the main thing 14-year-olds say. Yeah, exactly. So it's in the ether. You see signal. There are some companies, like you mentioned, who are acting on it. We'll see what happens in the next two years. But I don't see how it's going to reverse. Once we get more data, maybe the reality will be less dramatic than some make it out to be.

1:13:39And then we'll see. We'll do our best.

1:13:40Nilay Patel:I mean, I would love to be real that it's the press. The media history is not in a moment of intense strength right now. It is contracting. But, you know, there has been some – I do think that the media is incentivized sometimes to over-dramatize these things. Could be real. Maybe it's not. I do think that there is a reality in it. The question is how quickly is a change going to happen and will we be able to – will society be able to adjust fast enough? Look, I get all my news from X, the Everything app, which assures me on the daily that AGI is just around the corner. I want to ask you the question I ask every time I talk to you.

1:14:19Nilay Patel:I always take an Uber to come see you. It's just my little tradition. And I always ask the driver. Thank you. The drivers always have the same question. So I have the same question every year. Sure. And then at this time, I actually got a very detailed follow-up question. Oh, cool. All right. Drivers all want to know, how are they going to get paid more? Well, they are going to get paid more by some of the newer jobs that we're giving them. You know, shopping, for example, on a per hour basis can pay more. But I do think that driver pay is based on kind of what market rate pay is, essentially, right?

1:14:57The local pay goes up and down based on the cost, kind of the spot cost of labor in a particular market. So I think the way that drivers are going to get paid more is the cost of labor generally goes up or goes down. Right now, the cost of labor is fairly steady and driver pay has been fairly steady nationwide. It's probably$32,$33 per utilized hour. here in New York City is over$50 per utilized hour. So drivers are making decent money. Of course, they're going to want to make more money.

1:15:25Nilay Patel:They'll want to make more money. Of course. Do you think autonomy changes that rate?

1:15:33I don't think significantly. I think that drivers are going to probably take longer trips. When we see autonomous inventory coming into a market, we slow down driver recruitment because we want the drivers who are in market making as much. So this point in markets like Atlanta, like Austin, where we have a significant autonomy presence, because we've reduced recruiting, driver pay is actually up. And I'm hoping that we can continue those trends for a long time.

1:16:03Nilay Patel:I'm glad you brought up utilized hours because this is the very detailed file. It's actually good because you brought up all the keywords of this question. So you mentioned Westchester. I live in Westchester. The drivers in Westchester are allowed to drive into New York City. They are not allowed to pick up in New York City and drive back to Westchester. So they lose, it's an hour, they literally lose one utilized hour. So I've been directly requested that you go and lobby the city and state so that they can go home with a utilized hour instead of an empty run. We have already been lobbying. Some of these regulations have unintended consequences.

1:16:38New York is unfortunately one of the most highly regulated markets out there, a significant amount of your fare goes to the city, etc. I think Ubers are too expensive here. And I think regulation sometimes goes over the top. It's something that I will absolutely take to the powers that be.

1:16:59Nilay Patel:The power that be in this city is Zeran Mondami. Have you met with Zeran Mondami? I have seen him speak. I have not met him one-on-one yet, but I look forward to that dialogue. Well, here's my tips. One, say you love New York City. He loves it when you say you love New York City. Cool, cool. I do. And two, tell him the drivers want the return trips from both the airports and the city. I will absolutely relay that to him. Maybe he listens to your podcast. You never know. We know some people. The same thing. I can't tell you. I can't tell you what the milestones are. All right, cool. Dara, this is always a pleasure.

1:17:29Nilay Patel:Thank you so much for coming. Thank you. Really appreciate it. I'd like to thank Dara for taking the time to join me on Decoder. And thank you for listening. I hope you enjoyed it. If you'd like to let us know what you thought about this episode or really anything else at all, drop us a line. You can email us at decoderatheverge.com. We really do read all the emails. Or you can hit me up directly on Threads or Blue Sky. We're also on YouTube. You can check out full episodes at DecoderPod. It's the same handle on TikTok and Instagram. If you like Decoder, please share it with your friends and subscribe wherever you get your podcasts.

1:17:56Nilay Patel:Decoder is a production of The Verge and part of the Vox Media Podcast Network. The show is produced by Kate Cox and Nick Stat. This episode was edited by Xander Adams. Our editorial director is Kevin McShane. The Decoder music is by Breakmaster Cylinder. We'll see you next time.

1:18:36and fees extra. Default terms at MintMobile.com.