自动驾驶：从特斯拉到 Waymo 及更远

女士们先生们，欢迎来到 Lemonade Stand。今天我们有一位非常特别的嘉宾，Casser Ununice，Applied Intuition 的 CEO 兼创始人。

Ladies and gentlemen, welcome to Lemonade Stand. Today we have a very special guest, Casser Ununice, the CEO and founder of Applied Intuition.

他们做了功课。你发音对了。

They did your homework. You pronounced it right.

其实我刚刚才问过。我说，等等，姓什么来着？

I actually had to ask right before this. I was like, wait, what's the last name?

这就像我自己的一个心结。几年前我做了一个 MIT 的访谈，当时有一群 MIT 的学生在招聘，他们做了准备，然后就这样做了。所以我在开头说了那句话，很长一段时间那是我唯一的内容。所以每次我遇到人，特别是那些看过那期节目头三分钟的人，他们念名字会非常准确，我就知道他们肯定看过我在 MIT 的演讲。

That's like my own mental thing. I said years ago I did this like MIT interview. It was like a bunch of, you know, MIT kids who were recruiting from MIT and they had prepared and they did this. So I said that at the beginning for a long time that was the only like content that I had. So every time I would meet people were particularly like they would watch that first three minutes of that episode and they'd be very precise with the name and I was like I you clearly saw this MIT talk I did.

非常感谢你接受我们的采访。我们现在在 Applied Intuition 的一个车库里，身后有一些车辆。要不你先开始吧？我们想了解一下，对于可能不太了解自动驾驶或者不关心自动驾驶的普通人来说，为什么这件事从宏观上看很重要？

Thanks so much for sitting with us. We are here in one of the Applied Intuition garages with some of the vehicles behind us. Actually why don't you kick it off? We wanted to kind of see like for the average person who might not be super aware of autonomous vehicles or might not care about them. Why is something like this important on a broad level?

你能描绘一下吗？假如我是一个技术怀疑论者，自动驾驶会如何以积极的方式影响我的生活？

Could you paint a picture? If I am a say I'm even a tech cynic, you know, how are autonomous vehicles going to affect my life in positive ways?

是的。我自己的观点是，我有点像乐观的怀疑论者。所以我对人们关于技术的一些焦虑也有想法。但我尝试用稍微不同的方式来看待。在自动驾驶领域这并不难，因为自动驾驶的价值就是减少受伤和死亡。我觉得没有人会认为这不是一件好事，这很难争论。

Yeah. So, I my own view I'm like I'm like an optimist cynic. So, that's I also have, you know, thoughts around some of these let's say anxieties that people have around technology. But I think I try to approach it in a slightly different way. It's not that hard to do in self-driving because the value of self-driving is just less injuries and less deaths. And I think there's nobody who's like that's that's hard to debate that that's a positive thing.

让我稍微分解一下自动驾驶生态系统，然后你们可以随便问。好。很多人说到自动驾驶，取决于你是在哪里长大的。如果你在底特律做汽车生意，你马上会想到，这是高速公路车道保持巡航控制、自适应巡航控制。你装一些雷达和摄像头，车子就能保持在路上。这不是复杂的智能自动驾驶，而是高级巡航控制。如果你和旧金山或洛杉矶的人聊天，他们经常看到 Waymo，他们会认为自动驾驶就是大型机器人出租车。车里完全没有人类，完全自主驾驶，基本上可以在城市里任何地方行驶。如果你在澳大利亚珀斯，说到自动驾驶，那里有大量的采矿活动，西澳大利亚是大型采矿运营商的基地，他们从珀斯出发去矿区。采矿中的自主运输已经进行了至少十年，实际上是 15 到 20 年。第一次看到卡车自主地移动泥土。那种泥土移动其实相当简单，基本上就是沿着路线行驶。如果你和工厂里的人聊天，他们见过沿着地面移动的机器人。25 年前我在工厂工作过，有沿着地面移动的机器人，基本上比叉车好一点，你可以这么想，但没那么重型。所以自动驾驶的世界非常广阔。在这个背景下，让我给它一个结构。首先我们谈汽车，然后谈其他行业。

Let me break down the self-driving ecosystem a little bit and then we'll like you know well then you guys can ask what whatever you want. Okay. So, a bunch of folks when you say self-driving depending on if I grew up in Detroit. Depending on if you're in Detroit in the car business, you're immediately thinking, okay, this is like a highway lane keep cruise control, adaptive cruise control. You put some radars and you have a camera and the car kind of stays in the road. It's not sophisticated intelligent self-driving, but it's like advanced cruise control. If you talk to somebody in San Francisco or LA and they see Waymos all the time, they think self-driving it's a big robo taxi. There's no human in the vehicle at all. It's driving completely autonomously and it goes basically anywhere in the city. If you're out in Perth in, you know, Australia and you say self-driving, that's where you have tons of mining happening in Western Australia and that's kind of the home of some of the big mining operators as they jump off from Perth to to to mines. Autonomous hauling in mining has been happening for conservatively a decade, but really like 15 20 years. First time you're seeing trucks that are moving dirt, you know, autonomously. Now, that type of dirt moving in that universe, it's it's quite unsophisticated. It's just like essentially it's following a route. You talk to somebody in a factory, they've seen mobile robots that follow the ground. I worked in factories 25 years ago, you'd have robots that are following the ground and they were like essentially like slightly better than like they're like forklifts. You can almost think of it, but they're not as heavy duty. So, there's a huge universe of what self-driving is. So with that context, let me like put a structure to it. First we'll talk cars and then every other industry.

在汽车领域，简化思考自动驾驶的方式就是：驾驶员座位上是否有人。有各种 SAE 级别，比如 L2、L2++、L3，但我们不需要深入。其实就是驾驶员座位上是否有人。我们可以简化为特斯拉和 Waymo，对吧？

In cars, the way to think about self-driving just to simplify it is is there a person sitting in the driver's seat or there's not a person sitting in the driver's seat. There's all of these like Society of Automotive Engineers levels like L2, L2++, L3, but we don't need to get into that. It's just really is there a human sitting in the driver's seat? There's not. We can simplify that saying by saying Tesla and Waymo, right?

所以这是最简单易懂的思考方式。特斯拉所做的，先不提 Cybercab，是目前你能买到的特斯拉。你买了它，它基本上能为你驾驶，可以行驶几百甚至几千英里而无需干预，从 A 点到 B 点，从家到办公室，可能十次驾驶中你完全不需要介入。但有时你需要介入，可能是因为有 Uber 上下客、路上有个箱子、有人靠边停车、有施工或一些特殊情况。交通灯因下雨或停电不工作等等。这时人类介入，处理最后那一点点情况，我夸张点说 5%，但实际上可能只有 0.5% 的情况，但仍然需要人类在场。没有人类，特斯拉无法理解场景并绕过它。这里的关键和 AI 的核心在于，系统是否基于其传感器正确感知环境。特斯拉的传感器比 Waymo 少，所以它就像一个人看得少一点。另一方面，Waymo 有更多传感器，而且驾驶座上没有人，这意味着它必须应对所有可能的变化，包括警察封路，所有人突然需要倒车掉头，完全出乎意料的情况。这就是自动驾驶的场景。大家常问的问题是，我们什么时候才能拥有特斯拉那样的东西，或者什么时候才能坐上 Waymo，什么时候每辆车都能带我去任何地方？这更复杂。但让我谈谈其他人们不常讨论的自动驾驶领域。

And so like that's the most simple way and easy way to think about it. What Tesla really does is and leaving the you know cyber cab out. This is the Teslas that you can buy behind it currently. Currently currently this is like you buy it and it drives generally speaking for you and it'll go hundreds sometimes thousands of miles without you needing to intervene and it'll go from point to point and they'll navigate from your home all the way to your office and maybe you won't for 10 different drives you won't interact at all. But sometimes you will and it might be because there's somebody's there's an Uber drop off or there's a box in the road or somebody's pulled over, you know, there's some construction site or some something unique is happening. The traffic lights are not working because it rained and some, you know, there's the electricity's, you know, down or something. Then the human comes in and they basically take care of the last I'll just be exaggerate, you know, exaggerate like 5% but really it's like the last like half a percent of cases, but you still need the human there. Without a human there, the Tesla is not gonna, you know, it does not, it does not have the capability to understand what's happening in the scene and navigate around it. And the key point here and like where the like AI of all this is is does the system perceive the environment correctly to as it is based on the sensors it has. A Tesla has less sensors than a Waymo and therefore it's almost like somebody who just sees and perceives a little less. On the other side, you have Waymo, which has many more sensors and there's no driver in the seat, which means it has to tackle every variant of thing that can happen, including like a police officer shuts down the road and now suddenly everyone has to like back up and turn around. Like it's a completely out of the blue scenario. That's kind of the scene within self-driving. The big question that is always asked is, well, when are we all going to have Tesla like things or how when are we all getting Waymos, you know, when will every car drive me to wherever I want? That's more complex when we talk about that. But let me talk about all the other areas of self-driving that people don't talk about a lot.

我觉得人们不会。普通人没有接触过，甚至没有真正想过。

I think that people don't. The average person is not engaging with or doesn't even really think about.

是的。对于纯音频收听的听众来说，我们面前有一辆拖拉机，还有工程车辆，那边有一辆保时捷，我旁边有一辆卡车。你们正在努力让这些各种各样的车辆实现自动驾驶，我觉得这特别有意思。

Yeah. And it's worth like for people listening purely audio. We're sitting in front of a tractor in front of like construction vehicles as well as you know there's a Porsche over there. There's a truck right next to me. There's this huge range of vehicles that you are working to make autonomous and I think it's particularly interesting.

所以在这个背景下。

So with that context.

所以我们用特斯拉和 Waymo 举例，因为大多数人不是农民，也不是商业卡车司机。但最接近的可能是商业卡车驾驶。就是把那些通常在高速公路上运输货物的大卡车变成自动驾驶。已经有很多公司在做这个了，也有很多测试在进行。比如，我们现在就在日本高速公路上运营无人驾驶卡车——应该说是自动化卡车。它们正在运输货物，就在我们说话的此刻。

So like we I'm just using the Tesla Waymo example because everyone like you know most people are not farmers and most people are not uh working you know as commercial truck drivers. But maybe the next kind of closest is commercial truck driving. So it's taking these you know large trucks that typically move goods uh on highways and making them autonomous. There's a bunch of companies doing that already. There's a bunch of tests happening. We for example run driverless trucks, automated trucks I should say on Japanese highways right now. They're moving cargo right now as we speak.

那是完全没人的吗？没有安全员？有安全驾驶员。是的。我认为目前地球上还没有完全无人驾驶的卡车。但这很快就会发生。我们说的不是五年后、三年后，甚至可能不是一年后。我的意思是，就在我们说话的时候，绝对有公司在尝试实现无人驾驶。

Is that with no one like no one safer? There is a safety driver. Yeah. Yeah. And I think in I don't think there is a driver out truck on the planet right now. Yeah. Uh but that will happen very soon. Like that is not like uh we're not talking about like 5 years from now or three years from now or maybe even a year from now. I mean there's there are absolutely companies that are trying to get driver out as we speak.

但如果你对自动驾驶一无所知，你理解自动驾驶卡车这件事。现在，让我们谈谈更独特的东西，比如建筑工地。那其实不是真正的驾驶，或者更难以理解的是战场。那么自动驾驶在这种情况下如何工作？我用战场举例，因为它最极端。你是一名士兵，在战区受伤了，无法行动，需要车辆离开战场。你应该能告诉那辆车“我需要离开这里”，然后车辆就能自主撤离。广义上讲，自动驾驶就是把智能放入一个物理移动的机器中。

Uh but you so in your mind as you think if you know nothing about self-driving you understand the self-driving truck thing. Now, let's go to something more uh uh let's say unique. It's like a construction site. Then now that's you're not really driving and that's not really, you know, so there or maybe even like a more uh kind of a difficult thing to understand is a battlefield. So, how does self-driving work in that situation? Let me use the battlefield example because it's kind of the most out there. You're a war fighter and you're in a war zone and you are injured. you're incapacitated and you need that vehicle to leave the theater and you should be able to tell that vehicle I need to get out of here and that vehicle can leave and exit exit autonomously. Broadly speaking, you can just think about self-driving and and really it's just taking intelligence and putting it into a physical moving machine.

很多时候，当人们说物理 AI，尤其是在硅谷，他们总是在谈论人形机器人。但我们在 Applied Intuition 对物理 AI 的理解，其实就是把智能注入所有这些现有的机器。地球上现在有超过十亿台这样的移动机器，它们只是不智能而已。人类在提供智能。

A lot of times when people say physical AI, especially in Silicon Valley, they're always talking about humanoids. And I think the way we at Applied Intuition think about physical AI is actually just taking intelligence into all these existing machines. There are uh north of a billion of these moving machines on the planet right now. They're just not intelligent. The human is providing the intelligence.

你的意思是涵盖汽车、卡车、联合收割机、船只。我们确实在做海事方面的工作，也在做无人机。无人机可能是一个很好的例子，因为人类无法坐在里面，所以智能从一开始就具身其中。而且每个季度、每个月、每年，这种智能都会变得更便宜、更复杂，能做越来越多的事情。现在一切都还很简单。我认为最令人印象深刻的是 Waymo 的机器人出租车，它们基本上能处理地理约束内的任何情况。

And you mean across both cars, trucks, uh combines, boats. We do we do work in maritime. We do work on drones. You know, drones are probably a good example of where almost like because a human cannot sit in the physical thing. Intelligence is really, you know, embodied in it in in almost from the beginning. uh and as we like every quarter and every month and every year that intelligence will get cheaper and it'll get more sophisticated and so it can do more and more things. Right now everything is quite simple. I think the the the most impressive stuff is Waymo's robo taxis which are like they can basically handle anything that's thrown at them within their geographic constraint.

好的。那么，就此展开，在 Applied Intuition，你们不只是做 Waymo 正在做的事，也不只是针对特定类型的自动驾驶汽车或建筑行业。你们在构建一个操作系统，一套可以安装在任何车辆上的技术——不仅是汽车，还有拖拉机、无人机、农业设备。所以，从高层次来说，你能解释一下为什么这么做吗？对我们外行来说，这似乎很疯狂。为什么不去专注一个垂直领域？为什么只选一个行业，而你们却在所有这些不同行业部署？我们在工厂参观时看到了。为什么这么做？有什么好处？策略是什么？

Okay. So, kicking off on that, here at Applied Intuition, you are not just doing what Waymo is currently doing and trying to tackle this specific type of self-driving car or you're not just going after the construction industry. You guys are instead building an operating system uh set of technology that can be installed in any vehicle and not just again not just any driving car, but into a tractor, into a drone, into agricultural equipment. So, on a high level, could you just break down why do that? That seems, you know, to a layman like us insane. Why wouldn't you go for one of these verticals? Why wouldn't you pick one industry and instead you you guys are deploying now across all these different industries? We've been able to see them in our in our factory tour. Why do that? What's the benefit? What's the strategy?

是的。我本来是工程师，但也读了 MBA，所以我会用一些 MBA 的管理术语。你举了特斯拉或 Waymo 的例子，这些是垂直公司，它们什么都做。它们制造传感器——Waymo 不造车，但特斯拉从车到计算都自己做，对吧？所以它们是垂直整合的。垂直整合有优势，尤其是在技术子组件还不存在的时候。我们是一家水平公司，就像芯片公司。英伟达、高通或 AMD 这些公司提供技术，用于很多很多设备。在软件方面，我们就像 Android，Android 运行在数千种硬件设备上。对于在家听的工程师来说，我们既字面上制造操作系统，也比喻性地——通俗地说——这项技术存在于很多很多设备上。

Yeah. The So, I'm an engineer originally, but I also did an MBA. So, I'm going to use some of my MBA administration jargon. Yeah. which is like uh well you're talking a Tesla or a Waymo as an example. These are vertical companies. They're doing everything. They're they're they're making the sensors and Waymo doesn't make the cars but Tesla makes the cars all the way down to the compute, right? So they're verticalized. There's advantages of being verticalized especially when the technology the subcomponents don't exist. We are a horizontal company. We're like a chip company. Uh an Nvidia or a Qualcomm or AMD these companies are providing technology which goes into lots and lots of devices. On the software side, we're like an Android and Android runs across thousands of hardware devices. And for the engineers who are listening at home, we're both literally we make an operating system and also proverbally we, you know, uh like coloquially this technology sits on lots and lots of devices.

为什么这么做？我来说说我们创办公司时选择这条路的原因，以及今天的现实。公司成立时，我们并不清楚哪种自动驾驶版本会被市场最大规模采用。如果你选择机器人出租车，你就做了一个决定，被锁定了。这是一个非常昂贵的决定。Waymo 已经花了超过 250 亿美元开发这项技术。说清楚一点，地球上没有多少公司或政府花过 20 亿美元做研发。世界上最高的建筑——哈利法塔，在迪拜，造价 15 亿美元。所以当我们随口说 250 亿时，那是巨额资本。我们没有 Google 这样的靠山。我们创业时就像一支草根乐队，没有威尼斯 Inner Scope Records 那样的资助。所以我们得开始打造热门产品并分发它们。我们开始业务的方式是：嘿，我们干脆构建一套工具，让所有不同的自动驾驶公司都能用来构建他们的系统。这真正教会我们的是，水平模式其实很有效，因为我们没有押注特定的形态，而是押注整个行业。总会实现自动驾驶的，也许卡车会更快到来。

Why is that? Um I'll give you the let's say the reason when we started the company why we went went down this route and then kind of the reality of what it is today. When we started the company, we didn't quite know which version of self-driving was going to be the most consumed by the market. So, when you pick, let's say, roboaxi, you're you're making a decision. Locked in. You're locked into that, and it's a very expensive decision. I mean, Waymo uh has spent north of 25 billion dollars developing that technology. To be super clear, there's not many things on the planet that companies or governments have spent $2 billion on for research and development. You know, the tallest building in the world, the the Burj Khalifa in in I think it's Dubai, that cost 1.5 billion. So, like when you're when we just throw these numbers around like 25 billion, like that is a huge amount of capital. And so, you know, we didn't have Uncle Google. We were starting our you know, we're we're like the we're the scrappy band. We're not we're not funded by, you know, we're not Inner Scope Records in Venice. And so we got to we got to start making hits and we got to distribute them. And and the way that we started that business was applied in was like, hey, let's just build a tooling that all these different self-driving companies can use, right, to build their systems. And what that really taught us is actually horizontal actually works well because we're not then betting on a specific form factor. We're just betting the entire industry. We'll somehow get autonomous. Maybe trucks will come faster.

也许建筑行业会更快发展，也许机器人出租车会更快到来，我们当时并不知道。我觉得这有点像金融术语里的‘隔离风险’。但随着业务深入——我们公司现在快 10 年了——我们构建了操作系统，并应客户要求直接开始做自动驾驶，然后我们发现：‘哦，实际上我们可以在很多垂直领域和很多制造商那里做同样的事情。’最初只是汽车行业，很多制造商。然后发现：‘哦，实际上你可以在卡车运输、国防等领域做同样的事情。’

Maybe construction will come faster, maybe robotaxis will come faster and we didn't know. I think that almost like in some finance terms we kind of isolated ourselves from that risk. But then as we got deeper in the business, our company's almost 10 years old now, as we got deeper in the business and we built like operating systems and we started building autonomy directly because our customers asked for it, then it's like, 'Oh, actually we can do the same thing across lots of verticals and lots of manufacturers.' First it was just automotive, just lots of manufacturers. Then it was like, 'Oh, actually you can do the same thing in trucking and in defense, etc.'

嗯。

Yeah.

然后发生了一件非常重要的事情。出现了一次技术转变。我不想讲得太细，但你知道 Google 发表了一篇研究论文《Attention Is All You Need》。OpenAI 的人看到了它，这导致了 LLM 的繁荣，也就是后 Transformer 时代——一种大致能让现代聊天机器人成为可能的 AI 架构。嗯，同样的技术也进入了自动驾驶。同样的 AI 架构现在也在自动驾驶中。当你听 Nvidia 或其他公司讲话时，他们会提到‘端到端自动驾驶技术’——他们说的就是这个。所以在 Transformer 这个非常重要的时刻之前，自动驾驶的每个垂直领域实际上都是相当离散和不同的。

Then something really important happened. There's a technical technology shift. I don't want to get too much in the weeds, but you know there was this research paper 'Attention Is All You Need' that Google published. The OpenAI guys saw it, that led to the LLM boom, which is like post-Transformers, a type of architecture within AI which allows for these modern chatbots roughly. Well, that same technology also entered self-driving. That same AI architecture is now in self-driving. The way you'll hear about it when you're listening to Nvidia or somebody, they'll say 'end-to-end self-driving technology' — that's what they're talking about. So self-driving before this very important moment of Transformers, each of the verticals were actually quite discrete and different.

当 Transformer 成为一个广泛适用的系统时……你看聊天机器人，以前有专门用于金融和客户服务的聊天机器人，现在有了这种能做所有事情的通用语言模型。

When Transformers became a broad system that applied across... see chatbots, you remember you had chatbots that were like just for finance and just for customer service, and now you have this like generalized language model which does everything.

嗯。

Yeah.

这是因为底层的技术架构。所以今天，你可以输入来自矿山和人类驾驶汽车的数据，这实际上让运行在许多不同硬件上的模型变得更好。

And that's because of the underlying technical architecture. And so today you can feed in data from mines and from cars, human-driven cars, and that actually makes this model which runs on lots of different hardware better.

嗯。

Yeah.

包括在船上运行的模型和在飞机上运行的模型——我们真的让 F-16 自主飞行过。所以这几乎像是一家年轻公司的生存本能：‘嘿，我们卖给很多玩家吧，因为我们不知道什么会成功。我们只是在押注行业和真正的技术优势，我们确实很幸运并从中受益了。’

Including models that are running on boats and models that are running on planes — we literally have flown F-16s autonomously. And so there's almost like the survival instinct of a young company that was like, 'Hey, let's sell to a bunch of players because we don't know what's going to work. We're just kind of betting on the industry and the actual technological advantage, which we've certainly gotten lucky and benefited from.'

我想我们之前的参观中有一个……

I think there was an from our tour earlier.

抱歉回答得太长了。我感觉自己说了 10 分钟。

Sorry for getting long answers. I feel like I'm giving 10 minutes.

不，我能看出你非常兴奋和充满热情。是的，是的，是的。不，这很棒。

No, I can tell you're very excited and passionate. Yeah. Yeah. Yeah. No, it's great.

我们之前参观时，有机会和副 CTO 交谈。我觉得出乎意料的是，所有这些不同的垂直领域可以相互补充，从矿山获取的数据和信息对半挂卡车或普通汽车并非无用——而是有帮助的。

When we did the tour earlier, we had the opportunity to speak with the deputy CTO. I think something that was unexpected to me was this idea that all of these different verticals can be complementary to each other and that the data that they bring in and the information that you're pulling from a mine is not necessarily unhelpful to the semi-truck or the regular car — it is helpful.

是的，是的。恰恰相反。多样化的数据能更快地改进模型。所以你实际上需要数据的多样性。让我用一个更贴切的例子。假设你只收集高速公路数据，想象你是一个人，不是 AI。你只在高速公路上开车。那么，当你被扔进城市交通中，比如在卡拉奇，你会不知所措。顺便说一句，有研究表明，如果你只在美国开了很长时间车，然后去另一个国家，确实需要一两天、三天来适应当地的交通规则。

Yeah. Yeah. It's the opposite. Diverse data improves models faster. So it's like you actually want a diversity of data. Let me use a more salient example. You could have just collect highway data and imagine you're a human, not an AI. You only drive on highways. Well, then you get thrown into city traffic in, you know, in Karachi, you would be overwhelmed by that. And by the way, there are studies that show if you've only been driving in America for a long time and then you go to another country, it does take you like a day, two days, three days to adjust to the rules of the road.

是的，我在新西兰差点撞死一个人。

Yeah, I almost killed a guy in New Zealand.

那也是新西兰。剪掉，剪掉，剪掉这段。

That's also New Zealand. Cut that out. Cut that out. Edit that out.

一两天算客气了。我不负责任。你问我们为什么做这个，这就是证据 A。

One to three days is generous. I'm irresponsible. You're asking why we're doing this, you know, exhibit A.

具体来说是为了从 Doug 手中拯救新西兰人。

To save Kiwis from Doug specifically.

那个可怜人只是在过他的生活。他不知道 Doug 差点把他干掉。

That poor man is just living his life. He doesn't realize that Doug almost took him out.

当然，我在人们通常不开车的这一侧是安全的。

Surely I'm safe on this side of the road where people don't normally drive.

嗯。

Yeah.

所以我想快速了解一下这方面的基础设施，因为我之前不知道 LLM 和 Transformer 对这个行业如此关键。老实说，我没想到这一点，即使作为一个……

So I want to just quickly try to understand the infrastructure side of this because I did not know that LLMs and Transformers were so pivotal to this industry. I wouldn't have thought that to be honest, even as somebody who's...

不是 LLM 作为输出——嗯，我说的是 Transformer 架构，抱歉。是的，是的，是的。所以我想我理解得对吗：你们正在构建一个最终能在许多不同车辆上运行并理解许多不同环境的系统，当你们从所有不同测试环境中提取数据时，所有这些数据都在喂养、成长和成熟一个单一的、类似世界模型的东西？这是根本情况吗？

It's not LLMs are the output — well, I'm talking the Transformer architecture, excuse me. Yeah, yeah, yeah. So I guess am I correct in understanding that you guys are sort of building this system that can ultimately run on many different vehicles and understand many different environments, and that as you pull data from all of the different environments that you're testing on, all of them are feeding and growing and maturing a single like world model? Is that what's fundamentally going on?

是的，‘世界模型’有不同的技术定义，所以我用那个词。嗯，一个理解周围世界的物理 AI 模型。好的。做出决策，然后告诉机器去行动，比如加速或朝不同方向移动。但答案是肯定的。

Yeah, 'world model' has a different technical definition, so I use that. Yeah, a physical AI model which is understanding the world around it. Okay. Making decisions and then telling the machine to act, you know, to literally like do this thing, like accelerate or move in a different direction. But yes, the answer is yes.

你们有没有一个更大的……这想想还挺神奇的。

Do you have kind of a larger... which is like pretty amazing when you think about it.

反直觉。我的意思是，我看过‘3Blue1Brown’，如果你知道他的话。是的，是的，当然。你知道，学习 Transformer。在我的脑海里，我不理解从那里到操作建筑机械的飞跃，但它竟然能工作，这太神奇了。

Unintuitive. I mean, I've watched the 'Three Blue One Brown' if you know him. Yeah. Yeah. Absolutely. You know, learn the Transformer. And in my brain, I don't understand the leap from that to running a construction rig, but it's amazing that it works.

我的意思是，这样想吧。就想想你自己，作为一个人。你开汽车。所以一旦你开过车，如果你坐进卡车，你可能不知道如何开手动挡，或者具体像大型 A 类车辆之类的。但你有这样的理解：‘这是方向盘，那是油门，那是刹车，我要上路并沿着车道线行驶。’所以这很相似——有很多迁移学习。嗯，真正发生的是模型正在理解世界的物理规律。

I mean, think about it this way. Just think about it, you as a human. You drive a car. And so once you drive a car, if you sat in a truck, you don't know maybe you don't know how to drive a manual or specifically like a large Class A vehicle or something like that. But you have an understanding of 'this is the steering wheel, that's the gas, that's the brake, and I'm going to go on the road and in these lane lines.' And so it is similar — there's a lot of like transferred learning there. Well, what's really happening is the model is getting an understanding of the physics of the world.

这就是实际情况，也是为什么这是物理 AI。它不像大型语言模型，在那里理解概念以及它们之间的关系，这些概念是词语。在单个词语中，当我说“fall”时，根据上下文，我是在说天气、有人摔倒，还是长期资本管理公司作为对冲基金倒闭？这是三种不同的情况，但上下文会告诉你。在物理世界中，环境会告诉你什么是可行驶的表面，我能做什么，以及我期望物理环境中的其他事物为我做什么。这实际上是一个相当棘手的问题。我们假设了所有这些事情。因为我们从小就知道这张桌子不会移动。我们理解这张桌子的特性和重力。模型必须学习所有这些。所以你想让它接触多样化的数据。但这就像经典的 AI。所以，缩放定律确实有效，并且有很多努力投入到实际制造这些真正智能的系统中。但我们认为这显然是一件大事。我认为人们犯的一个错误是认为自动驾驶只会出现在新事物中。你必须买一个全新的、带有自动驾驶的东西。实际上，比如你去一个矿场，那些机器已经存在了 25 年，它们被购买来在矿场运行几十年。所以我们不能等到更新换代。然后我们可以用硬件改造这些机器，让它们变得智能。

That's really what's happening and that's why this is physical AI. It isn't like large language models where there's understanding these concepts and how they relate to each other, which are words. In individual words, how does when I say something like 'fall' based on the context, am I talking about weather, somebody tripping, or long-term capital management falling as a hedge fund? Those are three different things, but the context tells you. In the physical world, the environment tells you what's a drivable surface, what can I do, and what do I expect the other things in the physical environment for me to do. It's actually a pretty tough problem. We assume all these things. We know because everything we grew up with that this table is not going to move. We have an understanding of the properties of this table and gravity. A model has to learn all of those things. And so you want to expose it to diverse data. But it's like the classic AI. So, the scaling laws really work and there's a lot of effort that goes into actually making these really intelligent systems. But we think it's obviously a really big deal. I think that one of the mistakes that people make is like self-driving will only be in new things. You have to buy a brand new thing that has self-driving. Actually, like you go to a mine, those machines are there for 25 years that they're being bought to run for decades in that mine. So, we can't wait till the turnover. So then we can retrofit those machines with hardware to make them intelligent.

我在想，你们正在开发的平台已经部署到这么多不同类型的东西上。有没有像你们做过的船那样，离最终愿景还很远的东西？而像商用车的自动驾驶，比如我的车，可能非常困难，但实际上似乎进展很大，接近最终愿景。所以你们有没有觉得缺少或不足的地方？

Is there something I think what I'm imagining is like the platform that you guys are developing has been deployed to so many different types of things. Is there an expectation of something like the boats you guys have worked on that is very far away from like the end vision? Whereas something like self-driving for commercial vehicle like for my car is maybe very difficult but also practically seems very far along and seems close to the end vision of what that's supposed to be. So is there something where you guys feel like you're short or missing?

我不同意这种观点。即使每个人都拥有特斯拉 FSD，比如每辆车都有，但 98% 的车辆不是特斯拉。所以它们没有。但假设另外 98% 也获得了。当你问什么困难时，让那另外 98% 和公司获得，不是针对什么，但设计一辆车已经够难了。更不用说制造一辆人们想买的、有吸引力的车，现在还要以易于使用的方式在其中放入真正的智能。这需要很多很多年。所有这些事情的难点不在于技术。而是技术向这些机器的扩散。那才是真正的难点。

I would disagree with that view where even if everybody had, let's say Tesla FSD in every one of their cars, just an example, 98% of vehicles are not Teslas. So they don't have that. But let's say the other 98% got it. When you ask about what's difficult, getting those other 98% and getting the companies, and not to pick on anything, but designing a car is hard enough. Let alone making it an attractive car that people want to buy that has, but now putting real intelligence inside it in a way that's easy to use. That's going to take many, many, many, many, many years. That's the difficult part of all this stuff is not the technology. It's the diffusion of this technology into these machines. That's actually the hard part.

你们最近在这方面有了突破，对吧？比如你们宣布了与 Stellantis 的合作。你们的平台正在直接集成到一系列人们熟悉的汽车品牌中。比如玛莎拉蒂、吉普等。那是什么？未来几年会如何发展？

Well, you guys had a recent breakthrough on this front, right? Like you have this you've announced this partnership with Stellantis. And your guys' platform is being directly integrated into a bunch of these car brands that I think people are familiar with. Things like Maserati, Jeep, and stuff. What is that? How is that playing out over like the next few years?

是的。我们刚刚宣布，因为这是热点，但在全球前 20 大汽车制造商中，有 18 家是我们的客户。我们覆盖了你想到的绝大多数品牌，你在停车场看到的，我们都在与他们合作。

Yes. We just announced it just because this is topical, but of the top 20 global automakers, 18 are customers. We're in vast majority of the brands you ever think of and you look in a parking lot, we're working with them.

等等，你能深入谈谈吗？那意味着什么？如果你与这些品牌合作，他们只是使用你的软件吗？他们是否添加传感器？你的技术和软件在做什么，汽车制造商 OEM 如何改变他们的做法？

Wait, so just can you also dive into that? What does that mean? You know, if you're partnered with these brands, does that mean they're just using your software? Are they adding sensors to be able to you like what is the state of what your tech and software is doing and how OEMs the car manufacturers are changing what they're doing?

每个。我们来谈谈如何安装自动驾驶操作系统。我们能够以你从街角经销商那里买到的同款车型无法做到的方式与这些车辆交互。合作车型，你正在采用，这里有两个不同的话题。一个是座舱体验，智能座舱体验，另一个是自动驾驶。它们会混合，随着时间的推移会融合，但它们是两个几乎独立的产品组，我们可以这样讨论，我们两者都做，也就是将智能带入物理机器。好的。所以如何与汽车对话、汽车如何与你互动以及汽车如何驾驶是两件不同的事。回答你的问题，我们做什么，我们提供全谱系。我们是一家技术提供商。他们把我们看作像芯片公司一样，只是我们不卖芯片。所以你是 Stellantis 或任何你能想到的汽车公司，你想让座舱体验更好。你想成为行业最佳，但你的团队没有 AI 工程师。你的团队没有那些技能，或者你有那些技能，但非常昂贵。作为技术提供商，我们可以将这些成本分摊给许多制造商。在旧汽车业务中，一种思考方式是，我曾在博世工作。当博世，你知道，梅赛德斯可以自己做刹车，但为什么梅赛德斯从博世购买刹车？因为博世集中了全球对刹车的需求，把它们放在一个工厂，在那里制造所有刹车，这实际上降低了成本。所以博世制造的刹车比梅赛德斯自己制造更便宜，仅仅因为他们有规模。我们在这些平台上做类似的事情，无论是座舱平台还是自动驾驶。所以你是小松，制造工程设备。你实际上也想要座舱和自动驾驶的东西，对吧？然后我们可以卖给他们。

Every. There's a let's talk how to build a self-drive's operating system installed. We're able to interact with these vehicles in ways that that same model if you bought it right now at a dealership down the street you wouldn't be able to do. Partnership car that's you're taking like let me there just two different topics here. So one is just the in-cabin experience, the intelligent in-cabin experience, and the other self-driving. So they mix and over time they'll converge, but those are two separate almost product groups as you can talk about it and we do both of those things which is bring and we brought, let's call it as bringing intelligence into the physical machine. Okay. So how you talk to the car and how the car interacts with you and then how the car drives are two different things. To answer your question of what do we do, we provide that full spectrum. We're a technology provider. They think of us as just like a chip company except we don't sell chips. So you're Stellantis or whatever car company you can think of and you want to make your in-cabin experience better. You wanted to be the best in the business, but you don't have AI engineers on your team. You don't have those skills in your team or you do have those skills, but it's really expensive. When we're a technology provider, we can split those costs across lots and lots of manufacturers. A way to think about this in the old car business is I used to work at Bosch. When Bosch, you know, Mercedes can do brakes, but why does Mercedes buy brakes from Bosch? Because Bosch takes all the globe's demand for brakes and they put in one factory and they make all the brakes there and it actually lowers the cost. So Bosch will make brakes cheaper than even Mercedes could make it themselves just because they're doing volume. We're kind of doing the same thing on these platforms, both the in-cabin platforms and on the self-driving stuff. So then you're Komatsu and you'd make construction equipment. You're like actually the in-cabin stuff we also want and the self-driving stuff we also want, right? And then we can sell them that.

那么，是不是可以说，对于这些公司来说，第一步可能是设置软件以实现座舱体验，然后你们也提供自动驾驶软件系统产品？这样对吗？

So is it correct that maybe first step would be for some of these companies they set up the software so that there's this in-cabin experience and then you're also offering this essentially product for the software system which is autonomous driving. Is that okay?

是的，完全正确。现在很难一概而论，但每家公司都有不同的策略。

Yeah absolutely. Now the reason it's hard to talk about generalizations but every company has a different strategy.

好的。

Okay.

有些人会说，嘿，实际上我们想买你们的自动驾驶，但座舱部分我们自己来做。

Some people are like hey actually we want to buy your self-driving but we want to do the in-cabin stuff ourselves.

好的。

Okay.

有些人会说，我们会买现成的方案，但我们要自己做自动驾驶。我之所以提到解决方案的光谱，是因为我们也制造所有用于开发这些技术的工程工具。所以有些公司会说，把所有的工程工具给我们，我们自己来创造知识产权。从这个意义上说，我们是一个真正的技术提供商，不是针对普通用户，但这就是科技公司的核心所在。

Some people were like, we'll buy the turnkey stuff, but we'll make our own self-driving. Now the reason I talk about a spectrum of solutions is we also make all the engineering tools to make these things. So some companies will just say, give us all the engineering tools, we'll make the IP ourselves. In that way, we're truly a technology provider, not for people at home, but this is what the bolts of a technology company are.

我认为许多对技术普遍持怀疑态度的人，对 AI 有广泛的恐惧，但即使是自动驾驶汽车，短期内的失业后果会如何影响事物。我很好奇你对此的看法，以及你在工作中是否有一种责任感伴随着这种理解。

I think many people who are broadly cynical about technology have these fears of AI broadly, but even just autonomous vehicles, like the consequences of job loss in the short term, how that's going to affect things. I'm curious what you feel about that and if there's a sense of responsibility in the way that you work on things here that comes with that understanding.

是的，有几个方面。绝对有责任。我们是社会成员，不是孤立的。我在底特律长大，非常关心那里发生的事情。有两件事。一是安全方面的责任。你不想推出不安全的科技。所以安全是我们公司的第一核心价值观。然后是对下游经济或社会影响的责任。AI 在知识工作者领域实际上更难回答，比如会计师和白领领域会发生什么。在蓝领领域，长途卡车运输、农业。美国农民的平均年龄是 58 岁。未来 20 年我们对食物的需求将翻倍。所以我们需要更多食物，而农民年纪很大。以采矿为例：全球 1% 的劳动力在采矿，但 8% 的死亡事故发生在采矿。采矿极其危险，而且地处偏远。人们并不急于从事这些工作。那么如何解决农业问题、采矿问题或长途卡车运输问题？人们不想当长途卡车司机。在日本，政府和公司如此急于推出无人驾驶卡车，是因为真的没有司机。司机短缺导致运营停滞。他们不得不限制加班时间，因为司机不足，人们工作到累死。所以这个将智能赋予物理移动机器的 AI 领域，焦虑和担忧要少得多。AI 来得不够快。自动驾驶来得不够快。所以这个领域的争议要小得多。但具体到旧金山和纽约的出租车司机，他们确实想做那份工作，而现在有了机器人出租车，我认为这些都是需要解决的大问题。我不是市场原教旨主义者，不像硅谷的一些人或哈佛 MBA。我举一个我在 Y Combinator 时的例子。在这家公司之前，我在 Y Combinator，Sam Altman 是总裁，我是 COO。我们投资了 DoorDash，当时叫 Palo Alto Food Delivery。我记得当时想，Grubhub 和 Seamless 已经存在了。现在我们都用 DoorDash。从劳动力角度看，人们离开了麦当劳和塔可钟，去为 DoorDash 和 Uber 开车。餐馆说：“我们很难招到人。”部分原因是工资，但部分原因是工作更差。当你为自己开车时，你可以随时开始和结束。我在麦当劳工作过。我记得第一天，我把手插在口袋里，一个经理说：“把手从口袋里拿出来。手插在口袋里的人不是在干活。”当时没有顾客。但当你开 Uber 时，没人会这么说。所以劳动力正在从麦当劳转向 Uber 和 DoorDash。当我们投资 DoorDash 时，你可能会说这会影响餐馆，因为人们会开始开车。但经济会自我调节。最根本的问题是我们的问题是否会得到解决。这就是资本主义。在我职业生涯的大部分时间里，我没有助理。我终于有了一个，但我并没有少干活；我干的一样多，只是工作内容不同。我的乐观观点：作为一个南亚裔男性，我的家人有开 Uber 的，以前是出租车司机，会有其他工作自然出现，因为人类总是需要解决问题。我不知道这些工作会流向哪里。这是我的希望。在农业方面，则更直接。

Yeah, there are a couple of areas. Absolutely. You have responsibility. We're members of society, we're not just abstracted away. I grew up in Detroit, I care a lot about what happens there. There are two things. One is responsibility from a safety side. You don't want to feel technology that is unsafe. So safety is our first core value in the company. And then there's responsibility for the downstream economic impacts or social impacts. AI in the knowledge worker space is actually a more difficult answer, like what happens to accountants and white-collar fields. In blue-collar fields, long-haul trucking, farming. The average American farmer is 58 years old. Our need for food is doubling over the next 20 years. So we need more food and the farmers are very old. Take mining: 1% of the global workforce is in mining, but 8% of fatalities are in mining. Mining is extremely dangerous and in the middle of nowhere. People are not rushing into these jobs. So how do you fix the farming problem, the mining problem, or the long-haul trucking problem? People don't want to be long-haul truckers. In Japan, the government and companies are so intent on getting driverless trucks because there are literally no drivers. Driver shortages are shutting down operations. They had to put caps on overtime hours because people were working themselves to death due to not enough truck drivers. So this area of AI, putting intelligence on physical moving machines, has a lot less heartburn and anxiety. AI can't get here fast enough. Autonomy can't get here fast enough. So it's a lot less contentious in this area. But when you get to specific things like taxi drivers in San Francisco and New York, where they do want to do that job, and now robo-taxis, I think those are big questions that have to be figured out. I'm not a market fundamentalist like some folks in the valley or Harvard MBAs. I'll use an example from my time at Y Combinator. Before this company, I was at Y Combinator, where Sam Altman was president and I was COO. We funded DoorDash when it was called Palo Alto Food Delivery. I remember thinking, Grubhub and Seamless already exist. Now we all use DoorDash. From a labor perspective, people have left McDonald's and Taco Bell to drive for DoorDash and Uber. Restaurants say, "We have a hard time getting people to work here." Partly it's wages, but partly the job is worse. When you drive for yourself, you can start and end whenever you want. I worked at McDonald's. I remember one of the first days, I had my hands in my pockets, and a manager said, "Get your hands out of your pockets. Anyone with hands in their pockets is not doing real work." There were no customers. But when you drive for Uber, no one tells you that. So the labor pool is choosing to move from McDonald's to Uber and DoorDash. When we funded DoorDash, you could have said this would impact restaurants because people would start driving. But the economy finds itself. The most fundamental question is whether our problems will be solved. That's capitalism. For the vast majority of my career, I didn't have an assistant. I finally got one, but I'm not doing less work; I'm doing the same amount, just different work. My optimist view: as a South Asian man with family members who drive for Uber and were taxi drivers before, there will be other jobs that naturally emerge because humans always need problems solved. I don't know where those job pools will go. That's my hope. In farming and agriculture, it's more straightforward.

似乎有两种情况。一方面，最终结果会是确定的；但另一方面，在像农业这样的领域，实际上并没有你预期的那种大规模失业，因为一开始就没有那么多人从事这些工作。

It seems like there are two categories. On one hand, there will be this end result that's figured out, but on the other, in something like farming, there actually isn't this large displacement that you'd expect because there aren't that many people filling those jobs in the first place.

但即使是“失业”这个概念——我不是经济学家，而且我觉得自己总是对那些硅谷人在自己专业领域之外高谈阔论翻白眼。所以我想对自己了解的东西非常谨慎：我了解底特律，了解汽车行业，了解 Y Combinator 和创办公司、融资，也了解物理 AI。所以我先声明这一点。关于失业问题：如果你按季度看就业数据，会发现美国增加了 5 万个岗位或减少了 5 万个岗位。那实际上只是净差额。每个季度，数百万个岗位被创造和摧毁。这只是差额。而且如果我没记错的话，每个季度确实有数百万个岗位被摧毁和创造。所以在这个背景下，很多公司来了又走，很多岗位来了又走，任何单个职业代码相对于整个劳动力市场来说其实都很小。

But even that concept of displacement—I'm not an economist, and I think I always roll my eyes when I see Silicon Valley guys pontificating in areas way outside their expertise. So I want to be super thoughtful about the stuff I know: I know Detroit, I know the car business, I know Y Combinator and funding companies and starting companies, and I know physical AI. So I'm putting that caveat on there. The displacement issue: if you look at jobs on a quarterly basis, you'll see that the US created 50,000 jobs or lost 50,000 jobs. That's actually only the net difference. Every quarter, millions of jobs get created and destroyed. It's just the difference. And I think if I remember correctly, it's like literally single-digit millions every quarter that get destroyed and created. So within that context, lots of companies are coming and going, lots of jobs are coming and going, and any individual job code is actually pretty small relative to the full pool of the labor market.

但再次强调，我从你的声音里听出了犹豫，因为你并不想宣扬一切都会完美顺利。我确实认为这是新时代。如果你看看工业革命——一个常被提及的例子——那时有很多动荡。我的意思是，苏联是在工业革命中诞生的，最终导致了七八十年的巨大灾难。而这只是一场革命。作为工业化的产物，还有许多其他革命发生。美国在工业革命后发生了反垄断革命。工业革命后发生了两次世界大战。它们都是因为蒸汽机，最终是发电机吗？是也不是。有点关系。但失业问题呢？也许有点关系，也许没有。也因为人们从农业社会迁移到城市。很多事情都在变化。我不知道所有这些变量是什么，但我认为如果你看看我们的政治，确实有些不对劲。我们今天所处的政治环境截然不同。政治一直都有分歧。人们觉得 60 年代分歧更少，或者内战时期国家互相残杀分歧更少。不，我们经历过非常分裂的时期。这并不意味着美国的终结，也不意味着资本主义和民主的终结。但关键是，我认为我们都应该意识到这些事物在变化，我们需要提出新的解决方案，因为问题将是新的。

But again, I hear hesitation in your voice because you're not trying to propagate that it'll all be perfect and okay. I do think these are new times. If you look at the industrial revolution, which is an often talked about example, there was a lot of upheaval. I mean, the Soviet Union was created in the industrial revolution, which ultimately ends up being a huge calamity for seven or eight decades. And that's only one revolution. There are many other revolutions that happen as industrial outputs of industrialization. You have the antitrust revolution that happens in America in the post-industrial revolution. You have two world wars that happen post-industrial revolution. Do they all happen because of the steam engine and then ultimately the dynamo? Yes and no. Kind of. But job displacement? Maybe kind of, kind of not. It's also because people moved from agriculture-first societies to cities. A lot of things are changing. I don't know what all these moving variables are, but I think if you look at our politics, you do see something is going on. The political environment we're in today is distinctly different. Politics has always been divisive. I think people think the 60s were less divisive, or the Civil War where the country literally fought each other was less divisive. No, we've had very divisive periods. It doesn't mean it's the end of America, or the end of capitalism and democracy. But the point is, I think we should all be aware that these things are moving, and we need to maybe come up with new solutions because the problems are going to be new.

当你从事这样的技术工作时，无论是强制还是自愿，你会与政府监管机构合作吗？这涉及多少——比如在日本，这是强制性的——答案是强制性的。是的。那么有多少？他们在进行什么样的对话？因为大概人们意识到了这一点，政府也意识到了，也许在采矿或卡车运输等领域，某些国家存在直接的劳动力短缺并导致问题，那可能容易得多。

Are you, when you are working on technology like this, whether it's required or voluntary, working with government regulators? How much of this involves—like in Japan where it's required—the answer is required. Yeah. So how much? What types of conversations are they having? Because presumably people are aware of this, governments are aware of this, and maybe in areas like mining or trucking in certain countries where there's a straight-up labor shortage and it is causing problems, that's probably a lot easier.

法规在不同行业以非常不同的方式体现。例如，在采矿中，正如我之前提到的，法规主要围绕安全、安全、安全。这是一项极其危险的工作。人们经常死亡。所以世界各地政府的所有规则实际上都基于 100 年来人们死亡的经验。顺便说一句，几年前我背着背包去了玻利维亚，去了一个完全不受监管的矿井。基本上——玻利维亚有一种社会主义观点，认为国际矿业公司剥削工人，工人应该自己经营矿井。提示：那是我见过的最危险的工作场所，因为没有人遵守任何规则。所以尽管政府——我们美国人，我想，无论怎样都普遍讨厌政府——或者公司被讨厌，这在美国也很常见。他们确实会制定规则，当错误发生时，规则就会被制定。所以当你考虑法规时，它们总是向后看。我们在汽车、自动驾驶出租车和卡车上看到的法规都是关于谁能驾驶以及如何驾驶。新的规则和法规正在为自动驾驶出租车和类似事物制定。但法规总是远远落后。汽车于 1886 年在德国发明。停车标志——那个红底白字粗体的八边形——直到 1930 年才在美国统一。那是在咆哮的二十年代之后。整个咆哮的二十年代都没有统一；它们以不同的方式存在。所以最终，美国国家公路交通安全管理局（NHTSA）在 60 年代末成立。汽车在 19 世纪末发明，而这是 1960 年代。所以法规往往非常滞后。因此我认为，社会不应该期望政府能预见所有问题。它总是滞后的。但你会说，难道你指望公司自我监管吗？那也可能非常糟糕，因为公司的动机总是简单的：盈利。这是公司存在的原因。我对所有这些事情——法规、政府、公司、工会等——的看法是，它们只是一群一起做项目的人。所以当你分解这些——比如 NHTSA 或 FAA 或通用汽车或其他——分解成只是一起工作的人群，你就会对它们有更人性化的理解。就像每个人都在摸索着前进，试图弄清楚事情。

Regulations manifest themselves in different industries in very different ways. In mining, for example, as I mentioned earlier, regulations are really around safety, safety, safety. It is an extremely dangerous job. People die all the time. And so all the rules the government has around the world are all based on literally 100 years of people dying. As an aside, a couple years ago I went to Bolivia just with a backpack and I went to a mine in Bolivia which was completely unregulated. It was essentially—Bolivia has a roughly socialist view that international mining companies are exploiting workers, and the workers are just going to run the mines themselves. Hint: the most dangerous workplace I've ever seen, because there's nobody holding any rules to account. So as much as governments—we especially Americans, I think, just hate government in general no matter what—or companies are hated, that's also a common thing in America. They do bring in rules, and when a mistake gets made, rules are made. So when you think about regulations, they always look backwards. The regulations we see on cars and robo-taxis and trucks are all around who can drive and how they can drive. New rules and regulations are being made literally for robo-taxis and some of these things. But regs always are far, far behind. The car was invented in 1886 in Germany. The stop sign—just the octagon with red and white letters in bold—that finally becomes consistent across the US in 1930. That was after the Roaring 20s. We had the whole Roaring 20s and there was no consistency; they were all done in different ways. So finally, the NHTSA, the highway transportation authority for America, starts in the late 60s. The car was invented in the late 1800s, and it's the 1960s. So regs just tend to be really far behind. So I think the way society is, we shouldn't expect the government to basically anticipate all the problems. I think it'll always be lagging. But then you're saying, are you just expecting the companies themselves to self-regulate? And that can also be really bad because the company's motive is always simple: to make profits. That's the reason a company exists. The way I would think about all these things—regulations, governments, companies, labor unions, etc.—they're just groups of people that are working on projects together. And so I think when you kind of dissolve this—like NHTSA or the FAA or General Motors or whatever—and you dissolve it into just groups of people working together, then you get a more human understanding of what it is. It's like everyone's just kind of stumbling their way through and trying to figure things out.

我非常赞赏像 Waymo 这样的公司，他们做得非常好，设定了很高的安全标准，几乎成了行业标杆。我认为这非常积极。

I give a lot of credit to companies like Waymo who've done really good work and have a really high safety bar, almost setting an industry standard. I think that's been super positive.

我记得在另一次采访中，你谈到你们在海外的影响力如何增长，甚至包括现在在日本运营的卡车。你们在全球很多国家都有业务。据我所知，你们没有涉足的一个地方是中国。我认为世界……

I think it was in a different interview with you. I saw you talking about how your presence abroad has grown, the amount of even the truck that's operating in Japan right now. You have a presence in a ton of other countries around the world. One of the places you don't have a presence in, as far as I understand, is China. And I think the world has...

我认为那是我们唯一没有涉足的主要市场。

I think it's the only major market we don't play in.

中国和美国已经成为人工智能行业的主要参与者。我想知道你怎么看：Applied Intuition 是否在与中国的一些主要玩家在物理 AI 领域竞争？为什么在这么大的市场没有业务？你们与那个国家的关系如何，为什么没有涉足？

China and the US have kind of become these major players within this AI industry. And I was wondering how you see: is Applied Intuition competing with some other major players in this physical AI space in China? What's the reason for not having anything there in such a large market? Kind of your guys' relationship with that country and why there's no presence.

是的，这很复杂。像所有事情一样，我倾向于深入探讨细微差别。所以我会这样做，因为这些需要细致分析。让我分块回答这些问题。首先，我们应该把中国视为竞争对手吗？或者他们是我们的竞争者？嗯，没有哪个国家——竞争者的定义是那些从同一个池子里拿钱的人。国家不会拿钱。中国有公司会与我们竞争，但不是国家。所以你大致在问：美国是否与中国竞争？我认为人人都在相互竞争。中国与韩国竞争，韩国与日本竞争，日本与美国竞争，美国与德国竞争，但我们也都在合作。我认为这就像国际联盟的观点，我们如何制定规则和秩序。中国具体来说是一个共产主义国家。所以，他们在本质上是资本主义的，但这意味着他们政府的目标非常不同。他们公司的目标也非常不同。所以我们倾向于将自己的价值观投射到别人身上。举一个具体的例子：我们听到华为。对于不了解的人来说，它最初是一家网络公司，但现在是一家广泛的科技公司，基本上是一家消费电子公司。你会想，“哦，消费电子，那一定像三星和苹果。”实际上，华为不是那样的。“华为”这个词的意思是“中国的雄心”——这就是它的翻译。而且我认为大约每四个华为员工中就有一个是政府成员。创始人曾说过，我们的目标不是盈利。我们的目标是扩大全球市场份额和影响力。你能想象苹果的名字是“让美国再次伟大”，并且每四个员工中就有一个是某个政党的成员，他们说不关心利润，只关心美国的影响力吗？那不是一家公司。所以我想说的是，你不应该把苹果和华为相比，因为华为不是苹果，苹果也不是华为。这是非常非常不同的事物。它们都生产产品，但它们是截然不同的东西。我们在辩论和对话中犯的错误是，我们说“Applied Intuition 是否与 X 公司竞争？”或者“苹果是否与 Y 公司竞争？”它们不是同类比较。这是非常非常不同的事物。

Yeah, it's complicated. Like everything, I tend to want to get into a lot of nuance. So I will, because these need nuance. Let me answer some of those questions in separate chunks. First, should we think of China as a competition or are they our competitors? Well, no country—the definition of a competitor is somebody who is taking money out of the same bucket as you. And a country doesn't take money. There are companies in China that will compete with us, but not the country. So what you're broadly speaking is: does America compete with China? I think everybody competes with everybody. China competes with Korea, Korea competes with Japan, Japan competes with America, America competes with Germany, but we also all work together. And I think that's like the League of Nations view of how we figure out the rules and orders. China specifically is a communist country. So, they're capitalist in nature, but that means their government's goal is very different. Their companies' goal is very different. So we tend to project our values onto other people. Let's take a specific example: we hear Huawei. For the folks at home, originally it was a networking company, but now it's a broad technology company, basically a consumer electronics company today. You think, "Oh, consumer electronics, it must be like Samsung and like Apple." Actually, Huawei isn't like that. The word "Huawei" means "China's ambition"—that's what it translates to. And I think something like one out of four employees of Huawei are members of the government. The founder has said our goal is not to make profits. Our goal is to grow market share and influence around the globe. Can you imagine Apple's name is "Make America Great Again" and one out of four members are party members of a specific party, and they say we don't care about profits, we care about America's influence? That's not a company. So what I'm trying to say is you should not compare Apple to Huawei, because Huawei is not Apple and Apple is not Huawei. These are very, very different things. They both make products, but they're very different things. And the mistake we make in our debates and dialogues is we say things like "Does Applied Intuition compete with company X?" or "Does Apple compete with company Y?" They're not apples and apples. These are very, very different things.

但我想，可能有一些中国公司正在以类似的能力解决所有这些领域的自动化问题。无论他们是否追求利润，他们仍然试图在你可能需要的所有汽车中占有一席之地。

But there's probably, I imagine, there are Chinese companies that are approaching this problem of automation across all these verticals in a similar capacity. Whether or not they're in pursuit of profit or not, they're still trying to have a presence in all the cars that you might need.

当然。现在，我们不参与的原因——直接回答这个问题——是的，广义上有竞争对手，但没有一对一的竞争对手。没有一家精确的公司，但有很多。顺便说一句，在美国和欧洲也存在这种情况。但具体来说，为什么我们在那里没有办公室并参与竞争——我们在其他任何地方都有办公室。我们从汽车行业起步，中国汽车行业极其封闭，政府对中国公司有偏袒。你无法在公平的竞争环境中进入。你无法在公平的竞争环境中竞争，而且知识产权不受尊重。就像人们会偷你的想法，而你没有任何追索权。没有法律体系会让政府代表 Applied Intuition 对中国公司进行干预，并说：“好吧，Applied Intuition，这是你的知识产权，这家公司偷了它，我们将惩罚这家公司。”他们只会说：“不，他们是家中国公司。他们赢。他们总是赢。”所以我们只是不想参与一个对我们不利的环境。然后，广义上说，我认为这一点被夸大了：我们为美国做国防工作，人们认为，“哦，这就是你不在中国的原因。”那可能是最不重要的原因。坦率地说，因为我们是一家军民两用公司。我们构建的所有技术并不是说我们在构建国防专用技术。我们构建的技术实际上在很多领域都是商业可用的，然后我们将其用于国防机器，这与成为国防承包商不同。

Absolutely. Now, the reason we don't play—and just answer that question very directly—is so yeah, broadly there are competitors, and there's no one-to-one competitor. There isn't a precise company, but there are many. And that by the way exists in the US and that exists in Europe as well. But in terms of specifically why we don't have an office there and compete—we have offices basically everywhere else. We started automotive, the Chinese automotive industry is extremely insular, and the government puts their thumb on the scale for Chinese companies. You can't just go in on an even playing field. You can't compete on an even playing field, and all the way to IP is not respected. Like literally people will steal your ideas and you have no recourse. There's no legal system where the government will intervene on behalf of Applied Intuition against a Chinese company and say, "Well, Applied Intuition, this was your IP and this company stole it, and we're going to punish this company." They're like, "No, they're a Chinese company. They win. They always win." So it's like we just don't want to participate in an environment where that's not going to work for us. And then, broadly speaking, also I think this gets overblown: we do defense work for the US, and people think, "Oh, that's the reason you're not in China." That's probably the least reason. Frankly speaking, because we're a dual-use company. All the technology we're building is not like we're building defense-specific tech. We're building tech that is actually commercially available in lots of areas, and then we're putting it in defense machines, which is different from being a defense contractor.

我想我现在好奇的是我们还没有触及到的一些东西。你谈到了知识产权或策略可能被泄露。你们相对于其他公司的独特优势是什么？有很多人在尝试这个，我认为你们是目前最成功的之一，并且正在许多行业积极部署。我的理解是，在这个地区有汽车——每个软件或广义的软件系统——按一下按钮就能把车叫进车库。你们的优势是什么？为什么你们成功了？为什么你们现在是领导者之一？

I guess I'm curious now about something we haven't touched on. You talked about IP or strategy being leaked potentially. What is your guys' unique advantage over other companies? There are many people trying this, and I think you are one of the most successful right now, and it's really actively being deployed in many industries right now. My understanding is there are cars here in this area—every software or the software system broadly—called a car into the garage with a button. What is your edge? Why are you guys successful? Why are you one of the leaders right now?

我会给出我认为的真正答案，然后从非技术观众的角度来谈。真正的答案是这些系统制造起来极其复杂。就像为什么 Anthropic 和 OpenAI 是仅有的两家，或者少数几家？因为他们做的事情真的很难。这不仅仅是商业策略。也不仅仅是分销。技术本身实际上很难制造。

There's—I'll give what I think the actual answer, and then I'll talk about it from a non-technical audience perspective. The actual answer is these systems are incredibly complex to make. Like why are Anthropic and OpenAI the only two that are like that, or a handful? Because they're really hard to do what they're doing. This is not just like a business strategy. It's not just like distribution. The technology is actually difficult to make.

拥有无人驾驶出租车的国家比拥有核武器的国家还少。这些技术极其复杂。

There's less countries that have robo taxis than have nuclear weapons. I mean these are extremely complex technologies.

是的。我们往往轻描淡写，比如 Waymo 或 Tesla 所做的，其实非常了不起。作为硅谷人，我们应该为这些本地英雄公司感到自豪。我们做的都是真正困难的事。

Yeah. And we just handwave over it like what a Waymo does or what a Tesla does. It's incredible. We should be very proud as people of Silicon Valley that these companies are local hometown heroes. So we do things that are really hard.

非技术层面的答案是：我们非常了解自己的市场。我本科就读于通用汽车学院，在汽车行业长大。

The non-technical answer is we know our markets really well. I went to the General Motors Institute undergrad. I grew up in the car business.

嗯。

Yeah.

我在通用汽车工作过。我真的很热爱汽车行业，也了解它。我的联合创始人 Peter Lewig，他的父亲和祖父在汽车行业干了 28、30 年。我们骨子里就是汽车人。我们常开玩笑说，我们忘掉的汽车行业知识都比很多人知道的多。我说的不是爱好者那种知识，比如 99 款 GT3 和 992 款 GT3 Touring 的区别。我说的是真正的行业：如何造车、如何定价、如何……

I worked at General Motors. I really love the car business and I understand it. My co-founder, Peter Lewig, his father and grandfather worked in the car business for 28, 30 years. We are car guys all the way down. We used to make jokes that we forgot more about the car business than lots of people know. And I'm not just talking about enthusiast knowledge like the difference between a 99 GT3 and a 992 GT3 Touring. I'm talking about the actual industry: how do you make a car, how do you price it, how do you...

你当时在 V6 生产线上工作，对吧？

You worked on the V6 line, right?

是的，我做过。我是制造工程师，不是工人。但我也干过工人的活。是别克车，但也是帮助运营制造工厂的一部分。这不像你只是车库里停着一辆车。

Yeah, I did. I worked as a manufacturing engineer, not as labor. But I also did labor side. It was Buicks, but part of helping run factories that make things. It's like you're not just having a car in your garage.

嗯。

Yeah.

我热爱汽车行业。我认为我们非常成功的原因之一是，当你对一个行业了解得如此深入时，就像和那些在国防领域干了 25-30 年的人交谈，他们对国防的理解是外行人永远无法企及的。所以，如果你能把产品或技术与他们对市场的理解结合起来，就能做出一些非常特别的事情。非技术层面的要点是：我们真正了解自己的业务，了解我们所在的市场，也知道买家会如何购买。好的产品加上对市场的理解。

I love the car business. One of the reasons I think we're very successful is when you know an industry that deeply, it's like when you talk to people who've been in defense for 25-30 years, they understand defense in a way that you as a layman will never understand. So if you can marry product or technology with their understanding of the market, you can do some really special things. The non-technical point is we really know our business, we know the markets we play in, and we know how our buyers are going to buy. Good products and understand the market.

从 VC 的角度来看，因为每个答案都取决于你问谁，VC 可能会说：我们身处一个极度需要我们产品的市场，你们聪明又努力固然好，但关键是市场需求这些产品，就像吸力一样，这就是我们做得好的原因。这是各种因素的混合。就像为什么某些播客成功，而另一些不成功？有 50 个原因。可能是他们起步比别人早，可能是主持人有名气……

From the VC answer, because each of these answers depends on who you're talking to, from a VC perspective I think they would say we're working in a market that deeply wants our products, and it's good that you guys are smart and work hard, but it's the market demands these products and it's just a sucking sound, and that's why we've done well. It's a mix of all those things. It's like why are certain podcasts successful, why are they not? There's 50 reasons. It could be they started before everybody, it could be the hosts are famous or...

高大英俊。

Tall and handsome.

是的。高大英俊。这就是我们坐在这里的原因。但你知道，就像失败是孤儿，成功有千个母亲之类的说法。我们成功的原因有很多。

Yeah. Tall and handsome. That's why we're all sitting here. But you know, it's like failure is an orphan but success has a thousand mothers or something like that. There are many reasons that we're successful.

但如果我们失败了，就会一片寂静。

But if we were failing, you'd be like crickets.

是的。

Yeah.

我们稍微聊了你的个人经历。你在巴基斯坦长大，移民到这里，在底特律长大。我很好奇，正如你提到的，你在 YC 工作，差不多同一时期，像 Sam Altman 这样的人去创建了以软件 AI 为重点的公司。而你几乎在同一时间离开，说我们要做硬件和车辆，让它们做 amazing 的事情。你为什么选择这么做？这看起来难得多，可能更痛苦，回报也没那么直接。你的人生经历是如何导致这个选择的？

We've touched a little bit on your personal history. You grew up in Pakistan, immigrated here, grew up in Detroit. I'm curious, as you mentioned, you worked at YC around the same time that folks like Sam Altman went off to create software AI focused companies. You essentially at the exact same time went off and said we want to make hardware and vehicles do amazing things. Why did you choose to do that? Seems a lot harder, maybe a little more painful, less immediately rewarding. How does your life experience lead up to that?

是的。要说明的是，我认为我们更像一家人工智能公司。如果你看我们在算力上的投入和建筑的技术能力，它不像你在造…… 没错。我们真正的洞察是与制造商合作，这非常不同。原因就是这是我熟悉的领域。我在通用汽车、博世工作过，上过通用汽车学院。这融合了我人生的两个领域：谷歌和软件宇宙、AI 世界，以及我成长的行业。所以这更多是偶然而非计划。我记得 Peter 和我创办公司时，我们考虑过加密、语音等很多其他想法。我很高兴我们没有走那条路。

Yeah. To be clear, I think we're more like an AI company. If you look at how much we spend on compute and the technical abilities of the building, it's not like you're building... Exactly. Our real insight was to partner with manufacturers, which is very different. The reason is it's just what I know. I worked at General Motors, I worked at Bosch, I went to the General Motors Institute. It's merging the two areas of my life: the Google and software universe and the AI world with the industry that I grew up in. So it's more random than planned. I remember when Peter and I were starting the company, we were looking at ideas like crypto and voice and all these other things. I'm so happy we didn't go into that.

我认为人们觉得这…… 通常我只对创始人做这些演讲。那是我的受众。我真的很喜欢创始人。创始人基本上是小企业主，只是他们在这个概念下可以筹集资金并规模化，因为软件可以很好地规模化。这从根本上就是洗衣店和软件公司老板的区别。他们仍然是小企业主，但创始人必须非常小心，不要吸取错误的教训。从应用直觉中得出的错误教训是“哦，我们早就计划好了，全都知道”。我认为这是不诚实的。作为一家年轻公司，你要做的是获得 traction。明确地说，traction 就像你做播客，有人真的在听，一小群人，然后他们告诉其他人。在我们的业务中，就是我们了解汽车行业，可以用我们知道的软件和 AI 为那些人制造东西。一旦我们有了收入和 momentum，就能雇佣更多人。我们赚的钱用来支付工资，不会存进银行账户或分红。

I think people think this is... typically I only do these talks for founders. That's my audience. I really love founders. Founders are basically small business owners except they do it within this concept that they can raise capital and scale because software scales really well. That's fundamentally the difference between a laundromat and someone who runs a software company. They're still small business owners, but founders have to be very careful not to take away the wrong lessons. The wrong lesson from applied intuition is like 'oh we already planned this and knew it all'. I think that's disingenuous. What you're trying to do as a young company is get some traction. Traction, to be very explicit, is like if you're doing a podcast, somebody is actually listening, a small group, and then they tell other people. In our business, it's like we know the car business and we can make stuff for those guys from what we know: software and AI. Once we got a little bit of revenue and momentum, that allows us to hire more people. What do we do with the money we make? We use it to pay salaries. It doesn't go into some bank account or dividends.

这确实让我们能雇佣更多人，支付电费、食物和 GPU 的费用，让我们能继续做我们喜欢的事情，也就是硬件与软件的交汇点。

It literally allows us to hire more people to pay for the lights and to pay for the food and to pay for the GPUs and allows us to continue to work on the stuff that we like, which is this intersection of hardware and software.

Ker，非常感谢你来做客。这太迷人了。这是一个非常迷人的行业。我觉得我们还能再聊三个小时。

Ker, thank you so much for joining us. This is fascinating. This is a fascinating industry. I feel like we could have talked for like three more hours.

百分之百。我这儿还有一百个问题想问。

100%. I have like a 100 questions here I want to do.

而且我们谈得还比较宏观。这里面有很多细微之处。最后我想说，假如你是不从事 AI 工作、也不了解 AI 的人，只是不断听到这个东西，那么你该如何与之相处呢？你也许听这个节目是为了学习，那就尽可能自己去体验产品，然后你会看到它的局限性——YouTube 上有一堆视频，试图让 ChatGPT 数到一千，结果根本不行。所以我说，如果你接近这项技术并去学习，我认为这确实会降低你的焦虑。恐惧的根源是缺乏理解。所以去理解，去理解。这并不意味着没有真正的风险，也不意味着我们作为社会不需要解决我们谈到的那些复杂问题。

And there's a and we still were pretty high level. There's a lot of nuance and all of these things. The last thing I would say is like let's say you're somebody who doesn't work in AI or doesn't know about AI but are just constantly hearing about this thing and like how do you relate to this, you know, and you're trying to maybe listen to this to learn, just engage with the products yourself as much as you can and you start seeing the limitations and there are a bunch of like YouTube videos on like trying to get like ChatGPT just to count to a thousand and it's like it's hopeless. So the reason I say is you if you get close to the technology and you learn, I do think it lowers your anxiety a little bit. I mean fear, the root of fear is lack of understanding. So try to understand, try to understand. That doesn't mean there are not real risks. Doesn't mean we as a society have to figure out all these complex things we talked about.

但我认为你会稍微更处于主导地位，你知道，没有双关的意思。

But I think it you're a bit more in the driver's seat, you know, no pun intended.

非常感谢你来做客。

Thank you so much for joining us.

非常感谢大家的观看。

Thank you so much for watching.

天哪，我们从圣何塞飞回来。票卖得太快了。

Oh my god, we flew back from San Jose. Sold out so fast.