DoorDash 联合创始人谈智能体商务与食品配送的未来

听众朋友们，欢迎回到 No Priors。今天和我一起的是 DoorDash 的联合创始人 Andy Fang 和 Stanley Ting。我们聊一聊如何用自然语言向 DoorDash 订餐和买杂货；这对智能体式商务的未来意味着什么；他们的配送机器人 DOT；DoorDash 过去 8 年如何一直是一家机器人公司；他们网络的数据优势；以及这一切对 900 万送餐员和每年 30 亿次配送意味着什么。欢迎 Andy 和 Stanley。非常感谢你们来做客。真的很期待和你们聊聊 DoorDash 在做的一切酷事。我想先从 DoorDash 正在推进的智能体式商务开始。我觉得你们是真正用 AI 改变人们消费方式的最大规模落地之一。

Hi listeners, welcome back to No Priors. Today I'm here with Andy Fang and Stanley Ting, co-founders at DoorDash. We talk about how you can ask DoorDash in natural language for food and groceries, what that means for the future of Agentic Commerce, their delivery robot DOT, how DoorDash has been a robotics company for the last 8 years, the data advantages of their network, and what all this means for 9 million Dashers and 3 billion deliveries a year. Welcome, Andy and Stanley. Thank you so much for being here. Really excited to talk to you about all the crazy stuff DoorDash is doing. I thought we could start with what's going on with Agentic Commerce at DoorDash. I feel like you have one of the largest rollouts of actually using AI to change what people consume.

嗯。

Yeah.

那这背后的来龙去脉是怎样的？

So what was the backstory here?

说实话，这个想法开始于几年前，我们尝试在这个方向有所布局。一开始我们很看好语音这种交互方式。

It started a couple of years ago, honestly, in terms of our attempts to make a play here. Originally we were bullish on voice as the modality.

但最后发现语音并不是答案。

And that ended up not being the thing.

确实最终没有落地，但也许未来会成真，只是当时没跑通。不过让我们非常感兴趣的是这种自然的对话式体验。我们看到的是，人们可以把脑子里的想法很自然地转化为界面操作，而不是去网上做一番研究，或者搞关键词优化。人们发现这样搜索更简单，无论是更精细的餐厅发现搜索，还是杂货侧的各种任务。而且随着我们逐步推广，我们看到很多有意思的增长数据一直保持住了。

That ended up not being the thing, but maybe it will be in the future. It just didn't really land. But what was very interesting for us was this natural conversational experience. What we've seen is people being able to naturally translate what's in their head into this interface, versus trying to do some research online or do some keyword optimization. People just found it easier to search for things—either more nuanced restaurant discovery searches or different tasks on the grocery side. And we've seen a lot of interesting traction that's upheld as we've expanded the rollout.

你从用户侧看到了哪些行为变化？比如说，我的吃法或购物方式会不一样吗？

What are you seeing in terms of behavior change from the user side? Do I eat or buy differently?

是的。在餐厅这一侧，我们看到使用 Ask DoorDash 的用户中，有 50% 的下单轨迹是订了从未预订过的餐厅。这个比例很大，因为这是 DoorDash 历史上最难提升的指标之一。所以这个变化很显著。在杂货这一侧，我们看到购物篮金额明显提高，大约高出 40%。人们会拍一张冰箱里的照片，然后说：“帮我把冰箱补满”，或者做一些带有饮食限制的膳食计划，又或者“这个周末我想和家人做一顿意面晚餐”，甚至就是“帮我把常买的东西再下一单”。这比在传统界面里一路点选要省事得多。

Yeah. On the restaurant side, we're seeing that 50% of trajectories of people using Ask DoorDash for restaurants—50% of those trajectories are people ordering from places they've never ordered from before. That's huge, because that's one of the hardest metrics historically for DoorDash to move. So that's been big. On the grocery side, we're seeing much higher basket sizes—about 40% larger baskets. People will take a picture of what's in their fridge and say, 'Help me stock up my fridge,' or they'll do meal planning with dietary constraints, or 'I want to cook a pasta dinner this weekend with my family,' or even just 'Help me reorder my usuals.' That's a lot easier than tapping through the traditional experience.

这太厉害了。我从没觉得 DoorDash 难用，但这表明确实存在未被满足的潜在需求，因为去新餐厅吃饭的门槛还不够低。

That's wild. I've never thought of DoorDash as difficult to use, but that suggests there's latent demand that wasn't being served because it wasn't easy enough to eat at new places.

没错。餐厅这一侧，很多人会形成习惯，但我认为人们也希望在吃的东西上有一些多样性。我们觉得这种体验恰好为人们表达这种需求提供了自然的方式。想想这种社交货币：我朋友 Andy 帮我发现了一家真的很棒的餐厅。Andy 特别棒，对吧？所以我觉得人们看待 DoorDash 的方式也因此不一样了。

Correct. A lot of people on the restaurant side build habits, but I think people also want diversity in what they're eating. We felt this experience ended up being a natural way to allow people to express that. Think about the social currency of my friend Andy found a new really good restaurant for me. Andy's awesome, right? So I feel like that's even a different way people look at DoorDash.

另一个我们投入的方向，是把世界知识融入体验中。

And another thing we invested in was actually incorporating world knowledge into the experience.

这是什么意思？

So what does that mean here?

指的是 DoorDash 平台上之外的，关于餐厅的实时动态。比如我们会看互联网上正在流行什么，或者模型里没有的信息——因为模型的知识截止日期比较早，人们自己反而会发现的那些内容。也许是网上热门的东西，或者是各个论坛里大家在聊什么。这正好对应你说的那个点：人们想吃当下流行的事物。所以这是我们尝试融入体验的东西，目的是让用户更加信任它。

Things that are going on with restaurants outside of DoorDash. We'll see, hey, what's trending on the internet, or stuff that's not in the models—things people would find because their knowledge cutoff is too early. Maybe it's what's trending online or what people are talking about in various forums. It goes to your point of wanting to eat what's cool. So that was something we tried to incorporate into the experience to make people trust it more.

你觉得五年后，人们的购买方式或对餐厅的看法会有哪些不同？

How do you think people will buy or think about restaurants differently, like five years from now?

五年后的事情我不确定。

I don't know about five years from now.

我明白，在 AI 时代真的很难预测下一步。

I realize it's really hard in the age of AI, like next step.

所以对 Ask DoorDash 而言，我觉得接下来几个月，重点还是让人们更容易发现这个体验，并且知道该从哪儿开始。因为如果你只看到一堆可以输入的推荐查询，但不知道从何入手，可能会觉得很懵。所以要琢磨如何通过实验和打磨用户体验，来鼓励人们找到自己的使用场景。如果往更远了想，那就有点偏推测了。Stanley 和我经常聊这个。如果今天有人要从零创建 DoorDash——比如几个大学生在车库里想创办一个 DoorDash——我觉得那会非常不一样，很可能一开始就是智能体优先。有一组数据我最近总是想到：现在网上智能体的流量比人类流量还多。那么，我们要怎么打造一种 DoorDash 式的体验，去顺应这个趋势呢？这里有一些很有意思的遐想，但很难说。

So for Ask DoorDash, I would say, to start with, maybe that's the next couple months or so. I think it's making it easier for people to discover the experience and figure out what to do, because it can be intimidating if you just see suggested queries that you can type but you don't know what to start with. So figuring out how to experiment and tinker with the user experience to encourage people to find use cases for it. If I think further out, it's a little more speculative. Stanley and I talk about this all the time. If someone were to create DoorDash today—college kids in a garage trying to start DoorDash—I think it would look very different, probably more agentic-first. One stat I always like to think about: there's more agent traffic on the web than human traffic. So how do we have a DoorDash-type experience that plays into that trend? There are some interesting speculations there, but hard to say.

我的智能体能了解我在吃什么、或者在杂货方面想买什么吗？帮我理解一下，你们如何看待更丰富的上下文，或者如何在这方面做得更聪明？

What could my agent know about what I want to eat or what I want to buy from a grocery perspective? Help me understand how you think about richer context or how to be smarter there.

当然。一个很酷的例子：有人可能说，“嘿，我们办公室里，我可以在食品储藏架的上面装一个摄像头。”然后就可以这样：“嘿，当货架快空的时候，我可以向 DoorDash 发送一条指令，把货架补满。”

Sure. One cool example: someone might say, 'Hey, for our office, I can just have one of the cameras on the pantry shelf.' And it's like, 'Hey, when the shelf starts to get empty, I can fire off a query to DoorDash to stock up my shelf.'

是的。这原来是人干的活儿。没错。

Yes. This is a human being task here. Yes.

对。这差不多是——我的意思是，这些我们之后会再聊到——我们在 CLI 上的早期实验。这是一个例子，说明如何降低智能体参与这种体验的摩擦。

Yeah. That was kind of—I mean, something we talk about more later—our early experimentation with our CLI. It's an example of making it less friction for an agent to participate in that experience.

好的。既然我们在聊用户需求——我也算是 DoorDash 顶级用户了吧。我现在确实是他们的大量客户之一，但我每个周日晚上都要为大家庭做家宴，我们靠 DoorDash 吃饭，因为我没法每周都为这么多人下厨。或者说，我没办法每次都亲自做。

Okay. Well, while we're here talking about user needs—I've got to be like a top percentile DoorDash consumer. I'm, you know, a lot of customers at this point, but I host family dinner for extended family every Sunday night, and we eat DoorDash because I'm going to cook for all these people every week. Or I can't all the time.

我每次都是同样的做法，就是挨个问大家：‘好，谁要来？’

I do the same thing every time, which is to poll everyone: 'Okay, who's coming?'

哦，对。

Oh, yeah.

然后这些人有过敏或者各种忌口。我会问：‘大家有没有特别想吃的？’嗯。

And then these people have allergies and whatever else. And I ask, 'Does anybody feel like anything special?' Yeah.

然后我就下单了，我觉得能做的也就这些了。

And then I order, and I feel like that's all I can do.

你直接帮我设成自动模式了。我人到场，家人也吃得好。

You just put it on autopilot for me. I show up, my family's good.

这是一个使用场景。我觉得不完全一样，但类似的场景就是办公室订午餐。如果你是办公室经理，你不想还得盯着在某个时间点前订好午餐，不然外卖就不会到。而且每个人又有各自的过敏原或者饮食偏好之类的。

That is a use case. I think it's not exactly the same, but a similar use case is office lunch ordering. If you're the office manager, you don't want to have to make sure you ordered lunch by this time, otherwise it won't show up. And again, everyone has their own allergies or dietary preferences and stuff.

Stanley，你们在自主性和机器人技术方面也做了很多事情。作为创始人，你们对 DoorDash 的愿景显然比表面的‘这是个外卖配送网络’或者公司最初的一句话简介要更宏大、更有野心。机器人方面的工作是从什么时候开始的？

Stanley, you guys are doing a whole bunch of things on the autonomy and robotics side as well. Clearly your view of DoorDash as founders is broader and more ambitious than the surface-level view of it as a food delivery network or whatever the one-liner for the company was. How long ago did the robotics efforts start?

是的，我们实际上从比大家以为的更早就开始研究机器人技术和自主性了——其实是从 2018 年开始。当时自主性和机器人会不会成为主流还不明显。但我们觉得这项技术会给我们的行业带来变革，甚至可能是颠覆性的。我觉得创始人主导的公司有个好处：我们可以去思考那些面向未来的、更带有推测性的事情，并且不断思考如何确保自己不被颠覆。

Yeah, we've actually been looking into robotics and autonomy for much longer than people thought — since 2018, actually. Back then, it wasn't obvious that autonomy and robotics were going to be a thing. But we felt this technology would be transformative to our space and potentially disruptive. And I think that's the nice thing about being a founder-led company: we get to think about much more future-speculative things on the horizon, and constantly think about how we make sure we don't get disrupted.

就在下一个——我觉得，就像 Andy 说的，下一个 DoorDash，如果出现的话，不会是有人做出一个一模一样的 DoorDash，只是界面更好看。那样就太蠢了。

By the next — I think, like Andy said, the next DoorDash, if it comes along, isn't going to be someone who builds the exact same version of DoorDash but maybe with a better UI. That would be dumb.

对。

Yeah.

将来的东西会是这样的：我们怎么把 AI 智能体融入商业？怎么把自主性、机器人、无人机配送这些东西整合进来？我觉得，快进七八年，你会看到这一切在 AI、机器人和自主性领域逐渐成为现实。Waymo 已经上路了，我们也很庆幸在 2018 年就早早做了那些投资。

Going to be something like, okay, how do we incorporate AI agents into commerce? How do we incorporate autonomy, robotics, drone deliveries, etc.? And I think, fast forward seven or eight years later, you're seeing everything starting to play out in AI, robotics, and autonomy. You see Waymos happening, and I think we're glad we made that investment early on in 2018.

2018 年的时候这项业务还没那么惊人，不及今天这样。我觉得这么说应该没错吧？你如何看待这些非常长期押注的时机和顺序？从资本配置的角度，你会在什么时候投入这些事情？

There's an amazing business in 2018 — it was like less amazing than it is today. I feel like that's a fair statement, right? How do you think about the timing and sequencing of these very long-term bets, and from a capital allocation perspective, when you can invest in these things?

是的，我觉得这和我们在 DoorDash 投资很多东西的方式可能是一样的：一切都是从实验开始的。从某种程度上说，这就是 DoorDash 的创业故事。DoorDash 是斯坦福大学宿舍里的一个实验。它最初是一个叫 politely.com 的网站，上面有八份 PDF 菜单和一个 Google Voice 电话号码。直到我们确认这里确实有戏，才把它变成了一家公司。这就是我们过去 13 年一直秉持的理念，我们也把它用到了自主性和 AI 上。我们在 2018 年刚开始的时候，并不想启动一个巨大的机器人项目，也不想雇一个机器人专家去造硬件。其实就我和半个工程师的时间。那是一个臭鼬工厂式的项目，是一次探索性的实验：去看看外面有什么。我们当时甚至不知道自主性是什么样子，也不知道机器人会怎样影响我们的行业，但我们先去探索。去建立合作，去学习，去实验。而且一开始我们也没打算造自己的机器人。实际上，我们当时觉得自己不需要造这些技术。我们想，可以找一堆伙伴合作。那时候我们对机器人一无所知。有很多创业公司已经造出了机器人和自主技术——我们为什么不直接和他们合作呢？我们可以就做一个平台：我们建 API，我们负责配送，等等。我们确实这么做了好几年。我们和这个领域的每个人都合作过，从人行道机器人的玩家一直到自动驾驶出租车玩家。我想说我们从那段经历里学到了三件事。第一，它验证或者说确认了我们的信念：这里确实有戏。自主性的发生只是时间问题，而不是是否会发生的问题。你看看今天，Waymo 已经在路上开了。所以我们应该继续投入。第二，它让我们了解到真正实现自主性需要什么，因为结果发现你必须围绕自主性建很多东西：基础设施、生态系统。自主性如何与 DoorDash 整合？你接什么配送？运营方面怎么办。结果发现你必须围绕自主性建很多东西，才能让自主性成为可能。不是放一个机器人进去就行了，甚至不是放一个 LLM 进去，然后一切就奇迹般地发生了。它周围有很多东西，你得建立一个平台生态。所以我们最后建了一个叫‘自主配送平台’的东西。本质上就是：在一个自主性、机器人和无人机无处不在的后自主世界，你需要建哪些产品、技术、API、调度？你需要建什么？你如何和商家整合？消费者体验是什么样？第三件事，也是我们学到的可能最重要的一件事，最终让我们意识到必须自己造这些技术，就是‘围绕使用场景去构建’这个想法。

Yeah, I think it's probably the same as how we invest in a lot of things at DoorDash: everything starts out as experiments. In a way, that was the founding story behind DoorDash. DoorDash was a Stanford college dorm-room experiment. It started out as a website called politely.com with eight PDF menus and a Google Voice phone number. And it was only once we figured out, okay, there's something here, let's turn this into a company. And that's basically the philosophy we've taken throughout the past 13 years, and we've applied it to autonomy as well, and AI as well. When we first started in 2018, the intention wasn't to spin up this giant robotics program or hire a roboticist and go build hardware. It was really just me and half an engineer's time. It was a skunk works project. It was an experiment to go explore what's out there. We didn't even know what autonomy looks like or how robotics is going to impact our space, but let's go explore. Let's form partnerships. Let's learn. Let's experiment. And in the beginning, the intention wasn't to build our own robot. Actually, we didn't think we needed to build any of this technology. We thought we could just partner up with a bunch of folks. Back then, we didn't know anything about robotics. There were all these startups that had built robots and autonomy — why don't we just work with them? We can essentially be the platform: we'll build the APIs, we'll handle the distribution, etc. And we did that for actually several years. We worked with everyone in the space, from the sidewalk robot players all the way up to the robo-taxi players. I'd say there are three things we learned through that experience. One is it kind of validated or confirmed our belief that there's something here. Autonomy is a question of when it was going to happen, not if. And again, fast forward to today, you see the Waymos driving, right? It's happening. So we should keep investing. The second is it allowed us to learn what it takes to actually enable autonomy, because it turns out there's a lot of things you have to build around autonomy: the infrastructure, the ecosystem. How does autonomy integrate with DoorDash? What deliveries do you take on? The operational aspect. It turns out you have to build a lot of things around autonomy in order to make autonomy possible. It's not just you plop a robot in, or even AI, just plop an LLM in, and then things magically happen. There's a lot around it, and you kind of have to build a platform ecosystem. So one of the things we ended up building is this thing called the Autonomous Delivery Platform. Essentially, it's: what are all the products, technology, APIs, dispatch that you need to build in a post-autonomy world where autonomy, robotics, and drones are everywhere? What are all the things you have to build? How do you integrate with merchants? What does the consumer experience look like? And the last thing, which I think is probably the most important thing we learned and eventually led us to realize we had to build this technology ourselves, is really this idea of building towards a use case.

嗯。

Mhm.

没错。

Yes.

外面有很多做自主性的创业公司，但我们总觉得这些公司并不是真的关注某个使用场景。总觉得他们是先做技术。

There are a lot of autonomy startups out there, but we always felt like these companies weren't really focused on a use case. It always felt like they built the technology first.

嗯。

Mhm.

然后反过来再去找一个能套进去的问题。这些东西都是在真空中造出来的，这有点奇怪，因为在软件行业——我们当时在 YC 的时候，一直被教导要服务客户，做人们想要的东西。这种理念被反复灌输，然后你再去迭代。

And then retroactively try to find a problem to fit into. These things were all built in a vacuum, which is kind of weird, because in the software world — when we went through YC, we were always taught to serve the customer, build something people want. That's drilled into you, and then you can iterate.

但说到硬件、硬科技、AI 和机器人，大家恰恰喜欢反着来：先造技术，而不去想你要解决的用例是什么。只要这样做，最后得到的往往就是不太对路的东西。我们当时看过很多这个领域的公司，但总觉得它们不是 DoorDash 需要的。举个例子，在配送行业里，基本上有两类公司。一类是人行道机器人公司，就是那种 2 到 3 英里每小时、带轮子的饮水机——技术简单、非常高效。但我们很快发现速度和距离是巨大的限制，因为 DoorDash 平均每单配送是 3 到 5 英里，而典型配送时间是 15 分钟（不含做餐时间）。所以如果你放一个时速 2 英里的人行道机器人，那根本行不通。另一头是自动驾驶出租车公司，它们是为载人设计的：4000 磅的车，速度很快，用来载人。但载人跟载货的问题其实很不一样。如果你只是送几个卷饼，你真的需要一辆带座椅和空调的 4000 磅汽车吗？在自动驾驶出租车里，取货和卸货的问题也很不一样。你可以步行到 Waymo 那里——你坐 Waymo 的时候，有多少次被放在离目的地半条街或者一条街以外？这完全没问题，因为你可以走两步，但包裹做不到。那怎么解决我所说的“最初和最后 100 英尺问题”？餐品怎么在商家那里取？整合是什么样的？客户这一端，你怎么把餐放下？你怎么找到车道？你知道，人们期望餐品放在家门口或门廊前。所以当我们环顾四周，问自己：如果从第一性原理出发——这在 DoorDash 一直是我们的理念——从业务和客户用例出发，倒推回去，然后精确地造出我们解决用例所需的东西，那会是什么样？我们发现：实际上没有人真正在做这个东西。它不是人行道机器人，也不是自动驾驶出租车。我们觉得应该是介于两者之间的东西。对我们来说，合适的类比是：如果你想解决密集郊区里 3 到 5 英里的配送——那是大部分配送发生的场景——合适的类比大概是一辆自动驾驶摩托车、滑板车或自行车形态的车辆。它不需要是 4000 磅，大概 300 磅就够了。但它也必须比人行道机器人快很多——大概 20 到 25 英里每小时。当我们发现没人做这个，我们就决定：如果没人做，那我们就不等了，我们要自己掌握命运。我们投入进去，看看能造出什么。这经历了很多迭代。我们开始用真实的 DoorDash 订单测试，研究我们完成的 100 亿次配送，提取其中的洞察和运营经验。最终，我们推出了自己的自研自动配送机器人。这一路走来很不容易，但这也是我们希望应用到业务各个方面的东西——无论是自主性、机器人还是 AI。一切都是从实验开始，从“你解决什么客户问题、什么用例”开始，倒推回来，迭代、验证假设，然后慢慢把产品做出来。

But then when it comes to hardware, hard tech, AI, and robotics, it's people who do the opposite: they try to build the tech first and don't really think about the use case they're building toward. Whenever that happens, you end up with something that just isn't the right fit. We went through a process of looking at all these companies out there, but it always felt like it wasn't exactly what DoorDash needed. For example, in the delivery world, there are basically two buckets of companies. You have sidewalk robot companies, which are kind of these 2–3 mph water coolers on wheels—super effective, simple technology. But we quickly realized speed and distance were huge limitations, because the average delivery at DoorDash is about 3 to 5 miles, and the typical delivery time is 15 minutes, excluding the time it takes to make the food. So if you put a 2 mph sidewalk robot, it's just never going to work. Then on the other end of the spectrum, you have the robo-taxi players, which are really designed for carrying people around—4,000-pound vehicles, super fast, transporting people. But it turns out the problem of carrying people versus carrying goods is actually a little different. If you're only carrying a couple of burritos around, do you really need a 4,000-pound car with chairs and AC? The pickup and drop-off problem is also very different in robo-taxis. You know, you can walk to a Waymo. How often have you taken a Waymo where it drops you off half a block or a block away from where you need to be? That's totally fine because you can walk, but packages can't do that. So how do you solve what I call the first and last 100 feet problem? How does the food get picked up at the merchant? What does that integration look like? And then on the customer side, how do you drop off the food? How do you find the driveway? People expect their food to be dropped off at the front of their driveway or their porch. So when we looked around, we asked ourselves: if we were to start from first principles—and this has always been our philosophy at DoorDash—start from the business and the customer use case, work backward, and build exactly what we need to solve our use case, what would that look like? We looked around, and it turns out no one is really building that. It's not a sidewalk robot, and it's not a robo-taxi. We felt it was probably something in between. The right metaphor for us—again, if you're trying to solve that 3 to 5 mile delivery in a dense suburb, which is where most deliveries happen—is probably an autonomous motorcycle, scooter, or bike-profile vehicle. It doesn't need to be 4,000 pounds; it's probably about 300 pounds. But it also has to be a lot faster than a sidewalk robot—let's go 20 or 25 mph. And when we looked around and saw no one's building that, we decided: if no one's going to do it, instead of waiting around for it to happen, we're going to control our own destiny here. Let's invest in this and see what we can build. It took many iterations. We started testing with real DoorDash deliveries, looking at the 10 billion deliveries we've done, extracting the insights and operational learnings we have. That's eventually what led us to launch and ship our in-house autonomous delivery robot. So it's been quite a journey, but again, this is something we look to bring to every aspect of the business—autonomy, robotics, AI. It always starts out as experiments. It always starts out with: what is the customer problem you're solving for? What's the use case you're solving for? Work your way backward, iterate, validate your hypothesis, and slowly build the product over time.

听起来非常理性。我有一个假设，而且很有意思。我想问问你，现在所有人都用上自动配送，我们处在生命周期的哪个阶段？我有一个假设，不知道你是否认同：现在这个时代，很多人是先做技术，而不是从客户倒推。我觉得人们以为所有东西都会像 ChatGPT 那样运作。对，而且顺便说一句，当然有人做过指令微调，把它打磨成一个有用户体验的产品。但我觉得人们的思维模式是：这是一项通用技术，随便就能变成不同的应用。他们正把这个想法套用到很多领域，尤其是在自动驾驶上。我感觉到很多人是：'好，我们先把模型做出来，然后其他东西就算不容易，至少也是次要的。' 这完全不是我的看法。

That sounds extremely rational. I have a hypothesis, and it's very cool. I want to ask you where we are in the life cycle of everybody getting automated deliveries. I have a hypothesis, and I'm curious if it resonates with you: a lot of people in this era are building technology first versus customer-back. I think people think everything is going to work like ChatGPT. Right, and by the way, there was, of course, work done on instruction fine-tuning to shape it into a product that still had a user experience. But I think the mental model people have—that it's a general technology and it's just kind of free to turn into different applications—is what they're applying to lots of different things now, especially in autonomy. My sense is people are like, 'Okay, we'll make the model, and then the other stuff will be, if not easy, at least secondary.' This is not my view at all.

是的，我同意你的看法。这基本上就是我们打造配送形态的方法论。我觉得那种做法在软件领域也许行得通，但至少对我们这样的业务来说，DoorDash 是一个物理世界的业务。你把技术带到物理世界里，物理世界总是混乱得多，也复杂得多、微妙得多。我觉得人们没有意识到的其中一点就是 DoorDash 到底有多复杂。我们每年要做超过 30 亿次配送。没有两单配送是一样的，全部 30 亿单都各不相同，形态各异，覆盖不同的地理环境。旧金山市中心的配送，和达拉斯、甚至欧洲、或者赫尔辛基那种下着雪的地方完全不同。披萨和冰淇淋差别很大；晚餐和杂货订单差别很大，而我们现在扩展到零售、药品和包裹，又是另一番景象。DoorDash 发生的配送多样性如此复杂，以至于人们有时候根本意识不到这个问题有多微妙。而这就是我们在 DoorDash 需要解决的问题。

Yeah, I agree with you there. That's basically our methodology to building the delivery form factor. I think maybe that approach works in software land, but at least for a business like ours—DoorDash is a physical-world business. You bring technology into the physical world, and the physical world is always a lot messier. It's a lot more complicated, a lot more nuanced. I think one of the things people don't realize is just how complicated DoorDash is. We do over 3 billion deliveries a year. There are no two deliveries that look the same. All 3 billion deliveries look different. They come in all sorts of shapes and sizes and different geographies. A delivery in downtown San Francisco is completely different from a delivery in Dallas, or even in Europe, or in Helsinki where it's snowing. A pizza is very different from ice cream; your dinner is very different from your grocery order, which is very different now that we're expanding to retail, pharmacy, and parcels as well. The diversity of deliveries at DoorDash is so complex that people sometimes don't realize just how nuanced the problem is. And that's kind of what we have to solve for at DoorDash.

我觉得这是学习过程的一部分，尤其是在构建自主性或 AI 的时候：你要如何应对所有这些复杂性？而且归根结底还是：你是否理解使用场景。我认为我们比其他人拥有巨大优势，因为我们有别人没有的东西——DoorDash。我们有 100 亿次配送的数据可以提取。我们有所有这些消费者，每月有超过 4000 万消费者下单。我们理解当事情出错时如何处理，如何整合所有不同类型的商家。你和麦当劳或星巴克合作的方式，和夫妻三明治店非常不同；得来速餐厅和商场或主街上的餐厅又很不一样。你要如何处理这些不同的使用场景？不同的交互、不同的取货点。 嗯，我不知道你想不想补充 AI 方面。对我来说，你提到的那个类比——我会从自主性的角度去想，但也会从人形机器人领域可能如何发展的角度去想。几个月前我们还推出了一款叫 Tasks 的产品，让 Dash 配送员帮助收集数据点，用来训练一些世界模型。我觉得我们还处于非常早期，可以用很多不同的形态，对于哪种模型会成功也有不同看法。但和 ChatGPT 不同，我认为 V1 就需要投入大量成本。我想 ChatGPT 也可能花了很多钱。但我认为确实有很大压力去弄清楚如何真正提供价值——我必须比今天人们能做的更好。无论是 Dot 端到端配送，还是——我的意思是你可能投资了这个领域的很多不同玩家——但确实有压力要在质量或成本上优于替代方案。所以，是的。

And I think that's part of the learning process, especially when it comes to building autonomy or even AI: how do you manage through all that complexity? And again, it always comes down to: do you understand the use case? And I think we just have such a huge advantage over everyone else, because we have something that everyone else doesn't have: it's called DoorDash. We have 10 billion deliveries of data to extract from. We have all these consumers—over 40 million consumers ordering every single month. We understand the complexities of how to handle when things go wrong, how to integrate across all different types of merchants. The way you work with a McDonald's or Starbucks is very different than working with a mom-and-pop sandwich shop. A drive-thru restaurant is again very different than a restaurant at a strip mall or downtown Main Street. And how do you handle those different use cases? Different interactions, different pickup points. Um, I don't know if there's anything you want to add on the AI side. I mean, for me, the analogy you brought up—I think about it in terms of the autonomy thing, but I also think about it in terms of how the humanoid robotics space is starting to play out potentially. We also launched a product called Tasks a couple months ago, where we're having people in the Dash fleet help collect data points to help train some of these world models. And I think we're so early there, and there are so many different form factors you can use, and there are different opinions on what type of model is going to work versus not. But unlike something like ChatGPT, I think there's a lot of expense needed to invest in just the V1 of this. I guess ChatGPT could cost a lot of money too. But I think there's a lot of pressure though to figure out how to actually provide value—I have to be better than what people can do today. And whether it's Dot delivering something end to end, or—I mean, you probably invest in a bunch of different players in the space—but there's real pressure to be better than the alternative, from either a quality or a cost perspective. So yeah.

是的。不然我们在干嘛？

Yes. Otherwise what are we doing?

是啊，完全正确。

Yeah. Exactly.

嗯，那对我们这些不在凤凰城的人来说，DoorDash Dot 是什么？给我们讲讲它的设计吧。

Um, so for those of us who aren't in Phoenix, what is DoorDash Dot? Tell us about the design of it.

好的。DoorDash Dot 是一台自动驾驶配送机器人，完全由 DoorDash 内部自研。它重 300 磅，最高时速 20 英里，大小是汽车的十分之一。它是市面上唯一一款既可以在人行道行驶，也能上自行车道和机动车道的配送机器人。它已经在凤凰城上线，我们差不多已经实际配送两年了。它达到了 L4 级完全自动驾驶。所以如果你去凤凰城坦佩，感觉真的很像旧金山的 Waymo。

Yeah. So DoorDash Dot is an autonomous delivery robot. It's built entirely in-house at DoorDash. It weighs 300 pounds, travels up to 20 mph, and is one-tenth the size of a car. It's the only delivery robot out there designed to travel not just on sidewalks, but also on bike lanes and on the road as well. It's live in Phoenix. We've been live doing deliveries for almost two years now. It's fully autonomous Level 4. So if you come to Phoenix and Tempe, it really feels like Waymo San Francisco.

我要说件事，看你是不是同意。在机器人领域，除了要理解丰富的使用场景之外，你还需要知道自己将要面对的環境分布是什么。这对所有人来说都是个大问题。我觉得大多数熟悉这个领域的人都明白，要拿出一个精心挑选的成功任务演示并不难，对吧？

I'm going to state something and see if you agree. Even beyond understanding the wealth of use cases, you need to know what the distribution of environments you're going to be playing in is, in robotics. This is a huge problem for everybody. I think most people familiar with the area understand that it's not that hard to get a cherry-picked demo of one cool success on a task, right?

问题是要让它在任意物体或任何环境中都能工作。所以行业里有个巨大的疑问：我们要去哪里获取感觉上真实的数据？而最真实的数据其实就是真实世界的数据。所以我认为这是一个非常有趣的前提，说明你可能有资格去做这件事，而不只是为了业务质量去做。

The problem is getting it to work on any object or in any environment. And so there's this huge question in the industry: how are we going to get data that feels like realistic data? And the best realistic data is actually real-world data. So I think that's a really interesting premise for why you might have the right to go do this, besides wanting to do it for the quality of your business.

对，没错。而且我觉得这正是 DoorDash 可以发挥优势的地方。我们不必解决 100% 的使用场景。这也是我们早期做合作时学到的经验，关于如何构建自动驾驶配送平台：理解什么样的配送适合什么样的模式。我一直以来的愿景是：不要设计一个解决所有问题的东西，而是制定一个多模态策略。比如让 DoorDash Dot 负责从商场出发的 3 到 5 英里郊区配送。现在我们已经在凤凰城上线，那是 Dot 的起点。Dot 非常适合那个市场——密集的郊区，但各地点之间仍然相隔足够远。如果是农村地区，道路基础设施差，又是轻量订单，也许可以派无人机配送。如果是复杂的多步骤杂货订单，需要上下楼梯、拣货和装袋，那么仍然需要 Dasher 来完成。DoorDash 的好处在于，这不是一个全有或全无的方法。你可以随着时间逐步引入这些模式，选择适当的方案。再说一次，关键在使用场景：该解决哪些使用场景，该为每个使用场景选择哪些模式——有没有某些配送非常适合机器人而不是人类？

Yeah. No, exactly. And I think that's also where DoorDash gets to shine with our advantage. We don't necessarily have to solve for 100% of our use cases. That was part of our learning in the early years when we did partnerships for how we built our autonomous delivery platform: understanding what kind of deliveries fit into what modality. And I think the vision was always: let's not design something to solve for everything, but instead let's come up with a multimodal strategy. Perhaps you have DoorDash Dot do the 3-to-5-mile suburban deliveries from a strip mall. So right now we're live in Phoenix; that's our starting point with Dot. That's the perfect market for Dot—dense suburbs, yet things are still far enough apart. Maybe if it's a rural area with poor road infrastructure, and it's a lightweight order, maybe you send a drone delivery. If it's a complicated multi-step grocery order where we have to go up and down stairs and pick and pack orders, you're still going to have a dasher for that. And that's the nice thing about DoorDash: it's not an all-or-nothing approach. You can phase in these modalities over time and pick and choose what's right. Again, it's about the use case: what are the right use cases to solve for, what are the right modalities to fit into each use case—are there certain deliveries you can carve out that make a lot of sense for robotics versus humans?

是的。我还觉得你们能控制路线和分配非常酷，就像你们能说“我能够完成这个任务”。

Yeah. I also think that's really cool that you have control over the routing and the distribution, where you're like, 'I can accomplish this task.'

没错。而且从消费者端和商家端来看，体验完全一样。对顾客来说还是同一个 App，你什么都能用。对商家来说，只需要一次集成。你已经接入了 DoorDash，突然之间你不仅得到 Dasher，还有无人机和自动驾驶。我认为这就是 DoorDash 最终在构建的东西：一个面向本地商业的生态系统。这是很难复制的。而且要在真实世界里、在 40 多个国家、面对所有这些不同的岗位和商家去实现，这正是这门生意最难的部分。

Exactly. And then from the consumer side and the merchant side, it's the exact same experience. Still the same app for the customer—you can access everything. And for the merchant, it's just one integration. You already integrated DoorDash. All of a sudden, you get not just Dashers, but you get drones, you get autonomy. I think that's what ultimately DoorDash is building: it's really that ecosystem for local commerce. And that is something that is really hard to replicate. And it's about trying to do that in the real world across 40-plus countries, all these different jobs, all these different merchants. That's the hard part about the business.

帮朋友问的。

Asking for a friend.

一个关于你怎么走到今天的问题。现在生态里，能做机器人或应用型 AI 的研究者和人才供给不足，而你追的各种炫酷用例又很多。很多人会偏向通用方案——比如“我们可以一次搞定”。我猜你也在争取这些人。你怎么说服别人来 DoorDash 做这些事？

A question about how you got here. There’s an insufficient supply of researchers and people who know how to work on robotics or applied AI in the ecosystem, relative to all the different cool use cases you go after. A lot of people gravitate toward the general case—like, 'We can solve it once.' I assume you’re competing for some of those people. How do you convince people to work at DoorDash on these problems?

对，我的说法很简单。基本上就是：你是想去做原型和演示、待在博士实验室里，还是想做一个真正能在现实世界里交付的东西？我觉得这真的就是我们建立起来的文化，无论是在 DoorDash Labs 还是在所有 AI 项目上。我们不只是来做纯研究的。归根结底，我们要交付能产生实际影响的东西。过去十多年里，自动驾驶圈的人受够了——做了一个东西十年，却从没真正看到自己的产品在现实世界里被使用。而我们一直更注重务实、可落地的做法。我们不是要去做疯狂登月式的想法，而是要把能落地的东西先推出去，然后真正学习这些技术如何跟物理世界互动，并开始迭代。因为这些技术不是在真空里造出来的。你必须把东西放到现实世界，跟现实世界接触，然后从中学习。

Yeah, my pitch is really simple. It’s basically: do you want to go work on prototypes and demos and be at a PhD lab, or do you want to work on something where you can actually ship something in the real world? I think that has really been the culture we set up, both at DoorDash Labs and across all the AI efforts. We’re not just here to do pure research. At the end of the day, we get to ship something with real impact. People in the autonomy world over the past ten years were fed up with working on something for ten years and never actually reaching the point where they saw their products used in the real world. And for us, we’ve always been much more focused on a pragmatic, practical approach. It’s not about going after some crazy moonshot idea. It’s about getting something out that can be shipped in the real world, and then actually learning how these technologies interact with the physical world and starting to iterate. Because again, these things aren’t built in a vacuum. You have to put something into the real world, make contact with the real world, and learn from that.

我觉得我们在 DoorDash 很早就这么做了。其实，我不觉得很多人知道，我们在凤凰城做自动驾驶配送已经两年多了。我们去年公开宣布，这个项目已经做了超过两年。但一开始真的只是在学习。就像你之前提到的，做一个花哨的演示，或者让某种东西在一次性环境里跑通，是一回事；要把它变成一个真正的规模化车队、一项规模化服务、一门规模化生意，就完全是另一回事了。我经常讲这件事：做自动驾驶业务，需要的远不止是技术本身。你得在现实世界里真正做大规模——把车队规模放大。

I think we did that pretty early on at DoorDash. Again, I don’t think a lot of people know this, but we’ve actually been doing autonomous deliveries in Phoenix for over two years now. We publicly announced last year that we’d been doing it for over two years. But in the beginning, it really was just about learning. As you mentioned earlier, it’s one thing to do a fancy demo or have something that works in a one-off environment. It’s entirely different to turn that into an actual scaled fleet, a scaled service, a scaled business. I talk about this a lot: building an autonomy business takes more than just autonomy. You have to actually scale something in the real world—scale fleets.

突然之间，你就会遇到各种各样的边界情况，对吧？这些情况你平时根本看不到。当你必须一周七天、每天十小时、规模化地运行某件事时，东西就开始出问题。可能是很简单的，比如灰尘盖住了一个摄像头传感器。你的自动驾驶技术栈能有多少余量来应对这个？或者地上有些落叶——因为我们的机器人会在机动车道上行驶，但会尽量像自行车一样走右侧，或者走自行车道。如果人行道沿线的位置有落叶，可能你右两个轮子在落叶上，左两个轮子还在柏油路上。

All of a sudden, you’re running into all these edge cases, right? You just don’t see them when you have to do something seven days a week, ten hours a day, at scale, things start breaking. It could be something as simple as dirt covering one of your camera sensors. How robust is your autonomy stack to handle that? Or there are leaves on the ground, and because our robot drives on the road but tries to act like a bike, it takes the right side of the road or the bike lane. If there are leaves located along the sidewalk area, maybe half your wheels—the right two wheels—are on the leaves while the left two wheels are still on the asphalt.

是啊，突然间你要给车轮的扭矩就非常不一样了。你的自动驾驶技术栈、中间件和底层控制必须换一种方式处理。如果我只是在一个漂亮的小演示环境里开车，我根本不会想到这种事。事情就是这样开始出问题。你想怎么处理运营？大家没有意识到，为了把自动驾驶规模化，有大量非自动驾驶的运营工作。你得建 depot（站点/仓库）。这是一个实体世界生意。你得建 depot、做维护。电池怎么办？怎么充电？如果刹车系统——这是我们遇到过的一个问题。在某些情况下，车辆必须非常猛地刹车，再生制动系统会反过来压制电池，因为它会造成一种电击。这只会出现在极端边界情况下，但确实有某些情况你必须这么刹，因为这是现实世界中的事，而安全非常重要。所以如果它应付不了，你必须得想办法解决。

Yeah, well, all of a sudden the torque you have to send to the wheels is very different. Your autonomy stack, your middleware, and your low-level controls have to handle that differently. That’s something I would have never thought of if I were just driving in a nice little demo environment. Things just start to break. How do you handle operations? People don’t think about all the non-autonomy operations you need in order to scale autonomy. You have to set up depots. It’s a physical-world business. You have to set up depots and maintenance. What about your battery? How do you recharge it? What if your braking system—here was an issue we ran into. There are certain situations where the vehicle has to brake so hard that the regenerative braking system overpowers the battery, because it causes this electric shock. Again, it only happens in extreme edge cases, but there are certain situations where you have to do that because something happens in the real world and safety is super important. So if it can’t handle that, you have to figure it out.

另一个我们没想过的例子是机器人开机。当这还只是一个演示项目时，没人会去想开机时间。最初版本的机器人开机是一个简单的 Jenkins 脚本，我们的一个工程师花几个小时拼出来的。当时没问题。但现在你每天早上要给几百台机器人开机。脚本有一半时间会崩溃，还要花 30 到 45 分钟。乘以 500 台机器人，突然就变成了一个巨大的生产力问题。

Another example we didn’t think about was booting up the robots. When this was still a demo project, nobody thought about boot-up time. The original version of the robot boot-up was a simple Jenkins script that one of our engineers hacked together in a couple of hours. It worked fine then. But now you’re booting up hundreds of robots every morning. The script crashes half the time and takes 30 to 45 minutes. Multiply that across 500 robots, and all of a sudden it’s this huge productivity issue.

然后当然，你得把可靠性想清楚。现在你得开始考虑制造和供应链，当然还有运营层面：这个设备怎么和商家对接？怎么解决取货和送货问题？怎么培训商家？你甚至怎么找到顾客家里的 GPS 定位点。这听起来有点傻，但当你把地址输进 Google Maps 时，那个 GPS 定位点——尤其是去一个公寓小区的时候——从来不会正好是同一个位置。

And then, of course, you have to think through reliability. Now you have to start thinking about manufacturing and supply chain, and of course the operational side: how does this thing interface with merchants? How do you handle the pickup and drop-off problem? How do you educate the merchant? How do you even find the GPS pin of a customer’s home? That sounds kind of silly, but when you punch someone’s address into Google Maps, the GPS pin—especially if you’re going to an apartment complex—is never exactly the same spot.

确实。

Absolutely.

但如果是人，你总能想办法搞清楚。你根本不会多想。一个人类 Dasher 到了现场，他能找到餐厅在哪：这是那栋楼，这是前门。机器人做不到。

But if you’re a human, you kind of figure it out. You don’t even think about it. A human dasher shows up, they can find where the restaurant is: this is the building, this is the front door. You can’t do that with a robot.

机器人会出现在一个定位点，然后突然就好像：‘嗯，好吧，哪个……哪里……哪个临街店面？哪个前门？哪扇门？’现在想象一下 Dot 正在四处张望。

The robot's going to show up to a pin and all of a sudden it's like, 'Well, okay, which... where... which front storefront is it? Which front door is it? Which gate is it?' Now just imagine Dot looking around.

没错，正是。而且这确实是你得搞清楚的事情。但好在 DoorDash 有那些数据。比如所有的配送点，我们能看到人们实际上在哪个位置放下包裹。

Exactly, right. And again, that's something you have to figure out. But the nice thing is, again, DoorDash has that data. Like, we all the drop-offs, we can see where people are actually dropping off the package.

是啊。人类骑手历来都在哪儿放下包裹？这就是所谓的最初和最后 100 英尺问题。这些数据在其他任何地方都不存在，谷歌地图里也没有，只有 DoorDash 掌握。

Yeah. Where did the human dasher drop it off historically? And that, you know, is that first and last 100 ft problem. That data doesn't exist anywhere else. It doesn't exist in Google Maps. It only exists at DoorDash.

是的，我觉得这确实是一个非常有意思且真实存在的优势。嗯，早先在人们讨论 AI 和既有公司以及初创公司会发生什么时，我觉得很多人对既有数据优势的理解非常表面化。

Yeah, I think that is a really interesting and genuine advantage. Um, early on when people were talking about what's going to happen with AI and incumbents and startups, there were a lot of people I think had a very surface level view of like what the incumbent data advantage was.

是的。

Yes.

嗯，因为他们并没有真正去想：我们到底要做什么？用例是什么？这些智能应当完成什么？于是他们会说：‘啊，我们有客户记录和数据库’，而我会觉得，那其实和我们试图用智能体去完成的事情关系不大，对吧？而且我认为这在……中完全真实存在。

Um, because they didn't really think about like, well, what are we trying to do? Like, what is the use case? What is the intelligence supposed to accomplish? And so they'd be like, 'Ah, we have the customer records and database,' and I was like, that actually has very little to do with the thing we're trying to accomplish with an agent, right? And I think this is totally real in...

嗯，在机器人领域，我投资了一家叫 Sunday 的公司。我们深深地相信，你没法想象那种分布。一旦你接触到物理世界，就像你说的，真实世界，你会想：‘天哪，如果我们要洗碗，为什么洗碗机里有只猫？’比如你在某个人家里，主人会说：‘猫咪喜欢洗碗机。’这不是你靠想象能预见的。同样，你也不会去想象：‘我要处理这个扭矩问题，一个轮子在落叶上，另一个不在……’然后你会想：‘好吧，那这在分布中到底有多重要？’当你有足够数据的时候，你又会发现在另一台洗碗机里有另一只猫，然后你会想：我不知道这种场景有多少，但搞清楚的唯一方式绝不是工程师坐在那儿凭空设想这个机器人会遇到什么场景。那显然不是现实。

Um, in robotics, where... I'm an investor in a company called Sunday. Right. And one thing that we deeply believe in this company is you can't imagine the distribution. Right? As soon as you make contact with the physical world, as you said, or the real world, you're like, 'Man, if we're trying to do the dishes, why is a cat in the dishwasher?' And like, you know, you're in somebody's real house and they're like, 'The cat likes the dishwasher.' That's not something you're going to go imagine. Just like you're not going to imagine, 'Oh, I'm going to deal with this torque problem where one wheel is on the leaves and not...' And then you think, 'Okay, but how important is that in the distribution?' And then you find another cat in another dishwasher when you have enough data, and you're like, I don't know how many of these are out there, but the only way to find out is not by an engineer sitting and being like, let me imagine the setup and the scenario for this robot. That's clearly not going to be the reality.

我只是觉得，对于 AI 的下一个前沿，至少能让我们非常兴奋的，就是它将会如何影响物理世界。而且我到你的观点，就是你能模拟的只有那么多。你只能假装和想象各种演示场景。所以让我们非常有信心的，是把我们所拥有的世界级运营经验与世界级技术结合起来。而且我觉得很多 AI 研究者非常不愿意做大量运营方面的东西，或者他们觉得那很容易搞定。但我们在 DoorDash 这里真正强大的地方在于，我们有一支世界级的运营团队，你可以和它合作，无论是收集或标注数据，无论是弄清楚如何部署机器人，以及如何让整个机器人队伍运转起来。而且我想对于和我们聊过的很多人来说，这都极具说服力，因为就像：‘嘿，我们这儿可不是在空谈理论，你知道的。’

I just feel like for the next frontier of AI, it's, you know, at least what we're really excited about is like how it's going to affect the physical world, you know. And I think to your point, it's like you can only simulate so much. You can only pretend and imagine various demo situations. So one thing that makes us very confident is pairing that world-class operational expertise that we have with world-class technology. And I think a lot of AI researchers are very hesitant to do a lot of the operational stuff, or they think it's easy to handle. But one thing that's really powerful about what we have here at DoorDash is we have a world-class operations team that you can partner with, whether it's to collect or annotate data, whether it's to figure out how to deploy robots, and figure out how to get the fleet operations to work. And I think for a lot of people we talked to, that's very compelling because it's like, 'Hey, actually, we're not just talking hypothetical here, you know.'

你们已经在菲尼克斯实际配送了。接下来规模化会面临哪些挑战？

You're making the deliveries in Phoenix. What are the challenges from here for scale up?

我们在菲尼克斯做无人配送已经超过两年了。去年我们实现了完全自动驾驶 L4 级，我觉得那是一个超级令人兴奋的里程碑。而且实际上问题的关键是，如何把它从最初只有几台机器人、10 台，做到 100 台。我们需要爬过这座山：怎么去规模化。我认为有三个方面。首先是自动驾驶能力能否规模化。五年前的问题是完全无人驾驶是否可能，这只是一个研究项目，还是科幻？现在，尤其有了 AI，Waymo 已经实现了这种突破。我觉得 Tesla 也开始突破。我们去年也做到了。我们的整个自动驾驶技术栈都是自研的，但它是为配送而专门建造的，这有点不一样。你不能只是照搬。我觉得这是人们常常忽略的一点：你不能把 Waymo 做过的东西复制粘贴，然后放到 DoorDash 机器人上就万事大吉了。还是因为用例不太一样。这是一种自行车道尺寸的车辆，经常要在马路和人行道之间穿梭。据我所知，世界上没有第二个这样的东西，除了那个行为像 DoorDash 机器人的。但我们是针对自己的用例独特地打造的。所以自动驾驶绝对是一块。如何在不只是菲尼克斯、还有湾区更多城市持续推进规模化？我确信我们会遇到越来越多的边缘情况。但有趣的是，自动驾驶很可能越来越不构成约束或障碍。真正的问题更多在于后面两个方面。第二个是运营。如何规模化运营？菲尼克斯的餐馆和旧金山的餐馆行为不一样，伦敦和赫尔辛基也不同。如何适配所有这些不同的集成？所以是接口层，再加上车队管理。最后一块是硬件。这有点好笑。五年前我们刚起步时，大家都觉得硬件是通用品。现在看起来硬件正开始成为瓶颈。比如我们最初 100 台机器人是自己手工造的，那没问题。但接下来一千台或一万台呢？那我们就得开始考虑供应链和部件可靠性这类事情。这些东西必须能坚持很长时间。而且你不是在猜，因为你实际上能知道它们需要坚持多久，以及在现场表现如何。

I mean, we've been doing deliveries in Phoenix for over two years now. We went fully autonomous L4 last year. I think that was a super exciting milestone. And really it's just a matter of how do you take this from, again, originally it was just a couple robots, 10 robots, to 100. Again, it's just like we got to make that hill climb of how do you scale this. And I think it's really three components. First, can we get the autonomy to scale? Five years ago, the question was like, was autonomy even possible? Was this just a research project? Is this science fiction? Now, especially with AI, like Waymo's kind of made that breakthrough. I think Tesla's starting to make that breakthrough. We made that breakthrough last year. Like, our entire autonomy stack is built in-house, but purpose-built for delivery, which is a little bit different. You can't just copy. And I think this is the other thing people miss: you can't just copy and paste what Waymo's done and then plop it into the DoorDash Dot and everything works. It's, again, the use case is a little bit different. This is a bike lane profile vehicle that's constantly navigating between the road and the sidewalks. As far as I know, there's nothing else like this in the world, besides maybe the one that behaves like DoorDash Dot. But we built it uniquely to our use case. So autonomy is definitely one piece. How do you keep scaling across not just Phoenix but bring it to the Bay Area and more cities? I'm sure we're going to run into more and more edge cases. But the funny thing is, autonomy is probably increasingly becoming less and less of a constraint or a blocker. It's really more the next two, which is... The second is operational. How do you scale operations? Restaurants behave in Phoenix differently than restaurants in San Francisco, versus London or Helsinki. How do you adapt to all these different integrations? So it's the interface layer, and then the fleet management of it. And then the last piece is hardware. It's kind of funny. Like when we first started five years ago, everyone thought hardware was a commodity. And now it's starting to look like hardware is starting to become the bottleneck. Like, we hand-built the first 100 robots ourselves, which is not an issue. But then, okay, the next thousand or 10,000... Well, we're going to have to start thinking about things like supply chain and component reliability. These things have to last for a really long time. And you're not guessing, because you can actually tell how long it needs to last and how it's doing in the field.

完全正确。

Exactly. Right.

制造就像学习所有这些，结果发现在规模化时是个相当难的问题。所以我们实际做的一件事，就是与一家叫 Also 的公司合作，这是一家从 Rivian 分拆出来的微出行公司。RJ 实际上是这家公司的创始人兼董事会主席。

Manufacturing is like learning all that, and that turns out to be a pretty hard problem at scale. So one of the things we actually did was partner with this company called Also, which is a micromobility company spun out of Rivian. RJ is actually the founder and chairman of the board of the company.

我们为什么不跟真正懂车辆规模化的人合作呢？所以这就是我们达成的其中一项合作。不过有意思的是，五年前的问题是自动驾驶，现在越来越多转向运营、商业化、硬件和制造。而这正是 DoorDash 能凭借我们的规模和运营优势发光的地方：我们怎么把这件事从 0 到 1，做到像 100 到 1,000 这样的规模？

Why don't we work with someone who knows how to actually scale vehicles? So that's kind of one of the partnerships we struck up. But it's funny—the problem five years ago was autonomy. Now it's increasingly becoming more about operations, commercialization, hardware, and manufacturing. And again, this is where DoorDash gets to shine with our scale advantage and operational advantage: how do we take this thing from not just 0 to 1, but like 100 to 1,000?

10 亿到 30 亿。

1 to 3 billion.

对，10 亿到 30 亿。没错。我觉得 DoorDash 的位置非常好，可以接手这件事。我们在这方面有非常独特的优势，这也是我们想要发力的地方——发挥我们的长处。

Yeah, 1 to 3 billion. Right. And I feel like DoorDash is just so well positioned to take this on. We have such a unique advantage here, and that's where we want to play—playing to our strengths.

你们有巨大的优势。你们有网络、有现成的优质业务，还有这两个——我相信还有其他——围绕智能体商务和自动驾驶的大动作。你如何看待一家一万多人的公司，其中很多是运营，很多是技术？我相信你正在深入思考员工的生产力。比如谁负责这块、今天什么最重要、你们还发布基准——跟我们聊聊这些吧。

So you have these enormous strengths. You've got the network and the existing great business, and these two—amongst others, I'm sure—really big plays around agent commerce and around autonomy. How do you think about it, for a 10,000-plus-person company, where a lot of that company is ops, a lot of that company is technology? I'm sure you're thinking deeply about the productivity of that workforce. Like who owns it, what matters today, you're publishing benchmarks—talk about that.

我觉得在过去几年，真正要在科技行业高水平运作所需的条件已经发生了很大变化。我们去年如此兴奋地收购一家叫 Metis 的公司，其中一个原因就是想给公司注入一些 AI 原生的思维方式。对我们这样规模的公司——而且我觉得至少每家大公司都面临这一点——很多初创公司的运作方式非常不同。这一点你大概比任何人都清楚。但我们公司里很多人一直很难看到可能性，因为他们太习惯过去的做法了。所以关键在于，怎么把那些真正见过前沿可能性的人引进来，并把它融入我们的工作方式。显然，编码是最明显的转型点，我们在那里看到了很多成果。但我们也正在做整个组织的 AI 赋能工作。还有就是想清楚如何对公司各个部分进行基准测试。几周前我们发布了一个叫 DashBench 的基准，主要关注我们评估各种模型和测试工具在编码任务上的表现。对我们来说，这是一个很好的初步练习，用来计算我们花出去的这些钱的 ROI。一周前我还在看这个数据。我们 6 月的支出比 1 月增长了大约 20 倍。

I feel like in the past couple years, what was required to really operate at a high level in the technology industry has changed a lot. One of the reasons we were so excited to acquire a company called Metis last year was really to infuse some of that AI-native thinking into the company. For a company of our size—and I think every large company, at least, is facing this—a lot of startups operate very differently. You see this better than anyone else probably. But a lot of people at our company have struggled to see what's possible because they're so used to how things have worked historically. So really figuring out how to bring in people who have actually seen what's possible on the frontier and incorporating that into how we do our work. And coding is obviously the most obvious place to do transformation—we've seen a lot of gains there. But there's also work we're doing in terms of AI enablement across the entire organization. And figuring out how to benchmark various parts of the company. We announced a benchmark called DashBench a couple weeks ago, mainly focused on our ability to figure out how well various models and harnesses performed on coding tasks. That was a really good initial exercise for us to figure out how to calculate the ROI on all this money we're spending. I was looking at it a week ago. Our spend in June went up like 20x versus the spend in January.

哇。

Wow.

对。所以很明显，这些投入必须得到某种回报。显然我们看到了很多……你懂的。

Yeah. So clearly this has got to get some sort of return. And obviously we're seeing a lot of... you know.

我能问你一下吗——你不一定要回答——既然你看了这笔支出，它是下降了、持平了，还是继续增长？

Can I ask you—you don't have to answer—but since you've inspected this spend, has it come down, has it been flat, has it continued to grow?

我们看到它趋于平稳。

We're seeing it flatline.

而且我认为很大一部分是源于这些刻意的尝试。因为当人们在尝试时，尤其是在今年年初或者去年十二月，就会出现一种阶跃式的可能性变化。很多只是停留在实验层面，让人放手去做。但现在已经到了需要考虑更多的时候：第一，我们有一些简单的事情可以做，来确保我们不做浪费的事；第二，就我们发布的这个基准而言，我们确实需要开始计算 ROI。如果我们有办法最大化智能，同时把一些成本更低的任务委托给开放权重模型，我们就能获得同等水平的智能，但实际花费比只使用封闭权重模型更低。所以我们认为编码是机会很多的领域，主要是因为那部分支出大部分仍然集中在工程相关任务上。但实际上，我们组织内席位增长最快的是非技术部门，因为分析师发现它很有价值。我们的运营人员、负责战略商户的客户经理，他们想知道如何做 QBR，如何自动化很多这类工作。所以我们正在做一些工作，弄清楚如何在这些其他领域对我们的工作进行基准评估。另一件让我们感兴趣的事情是，我们与一些前沿实验室合作，比如会计任务或分析任务：最新模型表现如何？我们遇到的一个挑战是，我们会问团队：‘模型在你的任务上表现如何？’他们说：‘还行，能用。’但当我们做到……

And I think a lot of it is through some of these intentional efforts. Because when people were experimenting, especially at the beginning of the year or maybe December last year, there was a step function change in terms of what was possible. A lot of it was just experimenting and letting people run with it. But it’s gotten to a point where, one, there are easy things we can do to make sure we’re not doing wasteful stuff; and two, as it relates to this benchmark we released, we actually need to start calculating the ROI. If there’s a way for us to maximize intelligence but delegate to open-weight models for some of the cheaper tasks, we can actually get a comparable level of intelligence but pay less than if we were just using closed-weight models. So coding is where we think there’s a lot of opportunity, mainly because the vast majority of that spend is still within engineering-related tasks. But we’re actually seeing the highest amount of growth in our organization in terms of seats in the non-technical organizations, because analysts are finding a lot of value in it. Our operators, account managers who are trying to figure out how to do their QBR with strategic merchants, how to automate a lot of that. So there’s work we’re doing there to figure out how to benchmark some of the work we’re doing in these other areas. Another thing that’s interesting for us is we work with some of these frontier labs on, say, accounting tasks or analytics tasks: how well do the latest models perform? And a challenge we’ve run into is we’ll ask our teams, ‘How well do the models perform on your task?’ and they’ll say, ‘Yeah, it works okay.’ But then when we do…

然后你会说，好吧，3000 万美元的……

And you’re like, okay, $30 million of…

对，没错。问题在于成本。但当我们把某个人的数据送到实验室时，我们先得做数据清洗，然后还要搭一个强化学习环境等等。然后模型表现非常好。但我们发现这有点像你之前说的 Sunday Robotics 的例子：如果你把问题简化，模型可能做得很好。但不知道为什么，当我们用企业数据、真实场景去跑的时候，表现就没有那么好了。所以我们面临的问题是：是因为我们需要在接入框架上做一些事情，才能让模型发挥出来？还是模型本身就缺少这些东西，无论是数据分布还是能力集，导致它在会计分析或财务职能上无法实现那种阶跃式的变化？所以我觉得这是我们接下来的方向，超越编码这类事——当然编码本身还有很多工作要做。但在如何真正把这种阶跃式变化带到各类工作中，有很多值得探索的地方。

Yeah, exactly. It’s the cost. But then when we send somebody’s data to the labs, we have to do the data scrubbing, then we have to set up a reinforcement learning environment, whatever. And then the models crush it. But then we realize it’s kind of like what you’re saying with the Sunday Robotics example: if you dumb down the problem, maybe the models do well. But for some reason, when we actually run it on enterprise data with all the real stuff, it’s not performing as well. So for us, it’s a question of: is it because there are just things we need to do with the harness to get the model to perform? Or are there inherently things the models just don’t have in their data distribution or capability set that are not allowing that step function change in accounting analytics or finance functions? So I think that’s the next step for us beyond the coding stuff, which of course has a lot of work to do. But there are a lot of interesting things in terms of how we really see that step function change across the work.

长期来看，是不是所有 Dasher 都会消失，只剩地图上一个个点？然后会发生什么？

Is the long-term view that you get rid of all the Dashers and it’s just dots everywhere? What happens?

是的。嗯，我的想法，实际上我的预测是，在一个机器人、无人机和 AI 无处不在的世界里，我猜十年后，我们实际上会有更多 Dasher 在配送，而不是更少。原因很简单：DoorDash 的增长速度和运营规模都相当惊人。我不知道大家是否知道，我们有超过 900 万名 Dasher 在配送，业务同比增长 25%。如果快进十年，我们想从现在的规模翻 5 倍、10 倍，运力从哪里来？难道我们要让半个美国的人每个月都帮我们配送吗？这大概不可能。所以我们必须在其他领域寻找机会，既要引入新的运力形态，也要提升自身业务的效率。我认为机器人、无人机、Waymo 这样的自动驾驶车、人行道机器人，我们会看到一个多形态的配送队伍。我们必须尽可能去获取每一种可用的运力形态。所以我认为你不仅会看到更多的自动化和机器人，还会看到更多的人。而且随着自动化、机器人技术以及效率提升的引入，我也认为随着时间的推移，你会看到需求出现更强劲的激增，因为配送会变得更加经济实惠。

Yeah. Well, my take, my prediction actually, is that in a world where robotics, drones, and AI are everywhere, my guess is that in 10 years time, we’re actually going to have more Dashers doing deliveries, not less. Simply because, first, the pace at which DoorDash is growing and the scale at which we’re operating is pretty insane. I don’t know if people know, but we have over 9 million Dashers doing deliveries, and the business is growing 25% year over year. Fast-forward 10 years, and if we want to 5x from here, 10x from here, where’s the supply going to come from? Are we going to have half of America doing deliveries for us every month? That’s probably not going to be the case. So we have to find other areas of opportunity, both bringing in new modalities and improving efficiencies within our business. With robotics, drones, Waymo-like vehicles, sidewalk robots, I think we’re going to see a world with a multimodal fleet. We need to get our hands on every single modality we can get. So I think you’re not only going to see more autonomy and more robotics, but you’re going to see even more humans as well. And with the introduction of autonomy and robotics and efficiency gains, I also think over time you’re going to see an even stronger surge in demand, as delivery becomes even more affordable.

我期待有一天能用上这些。

I look forward to getting these one day.

太棒了。Andy，当你回顾最初在智能体商务方面的探索中学到的东西，未来人们除了点餐之外，会以怎样不同的方式购物？

Amazing. And Andy, when you think about what you’ve learned with the initial foray into agentic commerce, how are people going to buy differently in the future beyond food?

是的，我觉得其中一个让我着迷的趋势是，过去几年，Google 搜索查询的长度变长了。我对此的解读是，人们更愿意像跟真人说话一样，去和 Agent 或 App 交流。所以如果我们快进到未来，我认为我们越能让人们像和人打交道一样去和 App 或 Agent 交互，就越能降低他们下单的阻力，不管是点餐、买杂货还是零售商品。另一件我认为会成为现实的事情是，我们都需要思考‘Agent 优先’的体验应该是什么样的。我们最近发布的 DoorDash CLI 就是在测试这方面的一些东西。一旦你开始思考这件事，就会浮现出很多有趣的新的使用场景。举一个具体的例子：有个人非常兴奋地使用 DoorDash CLI，因为他说：‘让我用这个来简化我创业公司的行政管理工作。’然后当他发现 DoorDash 不只是送午餐时，他说：‘哦，等等，DoorDash 还能帮我买便利店的商品和杂货。’然后他就直接把摄像头对准了自家的食品储藏柜。

Yeah, one of the trends that I found fascinating is that over the past couple of years, Google search queries have gotten longer. And the way I’ve interpreted that is that people feel more comfortable talking to agents or apps like they would a normal human being. So if we fast forward and look ahead to the future, I think the easier we can make it for people to interface with apps or agents as they would with a person, the more it’ll reduce the friction in compelling them to place an order, whether that’s for food, groceries, retail, whatever. Another thing that I think is going to be true is that we all need to think about what the agent-first experience looks like. We’ve been testing some of that with the DoorDash CLI we launched last week. There are a lot of interesting emerging use cases that can crop up once you start thinking about this. One concrete example: someone was really excited to use the DoorDash CLI because they said, ‘Let me basically streamline my office manager use case for my startup.’ And when they found out DoorDash did more than just lunch, they said, ‘Oh, actually, DoorDash can order me convenience and groceries.’ So they just pointed a camera at their pantry shelf.

每当货架快要空了，他们就会派出智能体去补货。所以我觉得，这些你不会真正想到的用例，将会解锁一些在今天的世界里不太可行或不可能实现的有趣用例。随着我们把事情做得更自然地智能体优先，其中一些用例会变得有趣得多。

Whenever the shelf was getting empty, they would fire off the agent to restock the shelf. So I think those types of use cases that you wouldn't really think of—these are going to unlock some interesting use cases that wouldn't be as feasible or possible in today's world. As we make things more naturally agent-first, some of these use cases are going to become a lot more interesting.

太棒了。我很喜欢你们对用户体验以及 DoorDash 的范围与规模所表现出来的雄心。谢谢你们。

Amazing. I love how ambitious you guys are for both the user experience and the scope and scale of DoorDash. Thanks, guys.

是的，很高兴来到这里。

Yeah, it's a pleasure to be here.

在 Twitter 上关注我们：No Priors Pod。如果你想看到我们的面孔，请订阅我们的 YouTube 频道。在 Apple Podcasts、Spotify 或任何你收听的地方关注本节目，这样你每周都能收到新一期。此外，可以订阅电子邮件，或在 no-priors.com 上查看每一集的文字稿。

Find us on Twitter at No Priors Pod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way, you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com.