机器人公司CEO：人形机器人革命已成现实，即刻启程——与Bernt Bornich和David Blundin对话 | 第188期

你对全球机器人的思考可能比任何人都多。你对十年后的愿景是什么？

You think about robots in the world probably more than anybody else. What's your vision ten years from now?

首先，将会发生的是，各位，我们

First of all, what will happen is Everybody, we're

现在位于帕洛阿尔托的1x Technologies公司。这位是CEO兼创始人Bernd Bornick，还有这边的Neo Gamma一号和Neo Gamma二号。我想象最终我们会在机器人身上实现相同水平的AI，让我感觉像是在与一个完全智能的生命体对话。

here at 1x Technologies in Palo Alto. Bernd Bornick, the CEO and founder, Neo Gamma one and Neo Gamma two over here. I imagine we're gonna have the same level of AI eventually in the robot where I feel like I'm talking to a fully intelligent being.

Amolta是基于现实的，对吧，它真正理解存在的本质。

Amolta is grounded, right, that actually understands what this exist is.

我对此感到震惊。同样震惊。

I'm shocked by that. Shocked by that too.

我们如何解决科学领域剩下的那些真正棘手的问题？没有类人机器人的参与，这是不可能实现的。这对人类幸福而言几乎具有存在性的意义。

How do we solve the remaining really hard problems in science? This is not going to happen without humanoids. It's almost existential to us for human happiness.

所以塞利姆一直在说

So Selim is constantly saying

让它看起来像章鱼，并让它以章鱼所能展现的全部优雅姿态来运作，而不是试图将其限制在五根手指上，按照我们预设的方式来操控物体。

Have it look like an octopus and let it operate in the all the elegance that an octopus can rather than trying to constrain it into five fingers on this hand that do certain things and manipulate objects the way we're supposed to manipulate them.

那么给他的最终答复是什么？

So what's the definitive answer to him?

我们姑且说人形生物就是一张脸吧。

Let's just say humanoid is a face.

女士们先生们，这才叫登月计划。

Now that's the moonshot, ladies and gentlemen.

那是

That's

我。戴夫·布伦登，我的登月搭档和Neo Gamma。这边是NeoGamma一号和NeoGamma二号。我们刚参观了设施，相当令人惊叹。我们看到了，大概有几十个处于不同开发阶段的NeoGammas。

me. Dave Blunden, my moonshot mate and Neo Gamma. NeoGamma one and NeoGamma two over here. And we just did a tour of the facility, and it's pretty extraordinary. We saw, you know, probably dozens of NeoGammas in different stage of development.

他们从头到脚制造所有东西。NeoGamma内部大约有多少个组件？

They literally manufacture everything from head to toe. And how many components inside NeoGamma, roughly?

哦，这是最高机密。

Oh, top secret.

绝密。好的。

Top secret. Okay.

它大概在，嗯，几百的范围内，不是几千。

It's in it's well, it's in the hundreds, not the thousands.

好的。确实。

Okay. True.

我只是

I just

年底前在家里安装了我的第一台NeoGamma。是这样吗？哦，对。好的。太棒了。

secured my first NeoGamma at my home by end of the year. Is that right? Just Oh, yeah. Okay. Fantastic.

太好了。我们即将根据你的选择与Burnt或NeoGamma一起录制播客。我们这就去播客区域吧。你能在前面带路并帮我清一下路吗？

Great. So we're about to do a podcast either with Burnt or with NeoGamma depending upon what you want. And let's go ahead. We'll go over to the podcast area. Would you lead the way and maybe clear the way for me?

棒极了。顺便说一下，那边的包可以让Neo Gamma来拿。

Awesome. By the way, those bags over there, Neo Gamma can carry those.

超过体重的一半了。

Over over half the body weight.

对。我...我...我不是Neo Gamma。嘿，这个能交给你拿吗？

Yeah. I I I'm yeah. I I can't that I'm not the Neo Gamma. Hey. Can I give this to you to carry?

你可以试试。可能会触发安全限制，但通常没问题。

You can try. It might hit some safety limits, but it usually works.

好的。手臂抬起来。非常好。

Alright. Arms up. Very good.

对，就这样。你可以松手了，然后走几步。好了。它可能在走几步后觉得这样对我有点不安全。

Properly. There you go. You can let it go, and you can take a few steps. There you go. It might, after a few steps, decide that, like, this is a bit unsafe for me.

谢谢Neo。真是力大无穷。好的。而且很高兴知道NeoGamma会帮忙打扫房子周围。是的。

Thank you, Neo. Incredibly strong. Alright. And it's nice to know that NeoGamma will, you know, clean up the house around you. Yeah.

听着，虽然不知道你是几号，但我想说非常感谢。谢谢你的打扫。

Well, listen. I'm not sure what number you are, but I wanna say thank you so much. Thanks for cleaning up.

不客气。感谢您的时间。

Of course. Thank you for your time.

幸会幸会。祝你今天愉快。也非常感谢你。我想表现得有礼貌些。

A pleasure. A pleasure. And Have good day. And thank you very much as well. I wanna be polite.

要知道，你永远不知道机器人霸主什么时候会来对付我们。我希望你记住我真的很礼貌。我真的、真的很礼貌。好了，我安全了。

Know, you never know when the robot overlords are gonna, like, come after us. I want you to remember I was really polite. I was really I was really polite. Okay. I'm safe.

太好了。那你曾经恋爱过吗？我是说，你见过这么多其他机器人。总有些会让你心动吧？没有吗？

Great. And have you ever been in love? I mean, you meet all these other robots. I mean, some of them gotta be turning you on. No?

你应该看看41号。

You should take a look at 41.

41号。好的。伽马41号是你的地盘。明白了。

41. Okay. Gamma 41 is your is your gig. Okay. Got it.

谢谢。好的。哦，听着。伯恩斯在此。我们停止这次对话吧。

Thank you. Okay. Oh, listen. Burns here. Let's stop this conversation.

抱歉。抱歉。

Sorry. Sorry.

规矩点。是的。

Do behave. Yeah.

各位，欢迎来到Moonshots。我现在与我的Moonshot伙伴Dave Blunden在一起。Salimis Male这周末因为陪孩子或儿子而离线。但今天我特别邀请到了One X Technologies的CEO兼创始人Burt Bornick。Burt，非常高兴你能来。

Everybody, welcome to Moonshots. I'm here with my Moonshot mate, Dave Blunden. Salimis Male is offline with his kids this weekend or his son this weekend. But I'm here in particular with the CEO and founder of One X Technologies, Burt Bornick. A pleasure, Burt.

太棒了。非常期待这次对话。

Awesome. Looking forward to this one.

谢谢。是的，我是说，我们刚刚结束了这次巡演，这非常不寻常。你们是什么时候搬进这些新设施的？

Thank you. Yeah. I mean, we just finished this tour, and it's pretty extraordinary. When did you move into these facilities here recently?

大约一个半月前。

It's like one and a half months ago.

不错。我是说，这里有很多人在建造机器人，但还没有机器人造机器人的场景。

Nice. Well, I mean, just many levels of people building robots. No robots building robots yet.

我们正在朝这个方向努力，但还没

We're getting there, but not

实现。所以，

yet. So,

我是说，当我非常熟悉机器人技术、人形机器人领域时。虽然像Figure和特斯拉这样的公司最初专注于进入工厂，尤其是汽车工厂，但你承诺要进军家庭市场。是的。我个人对此感到兴奋。但我想先问，为什么选择家庭？

I mean, when I I'm very familiar with the robotics, the humanoid robot space. And while companies like Figure and Tesla are focused initially going into factories, automotive factories in particular, you made a commitment to the home. Yes. And personally, I'm excited about that. But I'd like to start with why the home?

对我来说，原因有很多，但主要有两点。第一点比较明显，就是消费类硬件的扩展速度与其他领域截然不同，对吧？iPhone在短短十多年内就达到了超过十亿台设备。在我看来，人形机器人除非达到规模，否则没有意义。对于某个特定问题，总有更好的自动化系统可用。

There's like to to me, there's there's two, like, there's a lot of reasons, but there's like two main reasons. Now, the first one is kind of obvious, which is just like, I mean, consumer hardware just scales at a different pace than everything else, right? We got to more than a billion devices of the iPhone in a bit more than a decade. And to me, humanoid robots does not make sense unless it's at scale. There's always a better automation system that you can use for one specific problem.

你需要规模才能真正实现惊人的可靠性、极低的成本、强大的生态系统和智能。更深层次的原因是，智能源于多样性。这一点从一开始就在各类AI研究中非常清晰，如今在AI的实际应用中也同样如此，无论是语言模型、图像模型、视频模型，还是这里的机器人模型。你并不需要反复输入相同的数据。仔细想想，这很合乎逻辑，对吧？

You need scale so that you really get this incredible reliability, incredibly low cost, incredible ecosystem and intelligence. Now, the slightly deeper one is also that intelligence comes from diversity, and this has been very clear actually from all the way in the beginning really, in all kinds of AI research and also now more practical applications of AI across all different domains, where it is like a language model or an image model or a video model, or in this case, robotics model. You don't really need data of the same thing over and over. Like, if you think about it, it's very logical, right?

所以如果你在汽车工厂里，基本上就是在重复做同样的事情，学不到新东西。

So if you're in an automotive factory, you're basically doing the same thing over and over again. You're not learning new stuff.

没错。我们实际上有一些相关数据。我们的上一代人形机器人Eve曾在2022、2023年用于安保和物流领域。数据显示，大约20到40小时后，机器人对特定任务的学习就会进入平台期。具体时间取决于任务复杂度——比如用轮式人形机器人巡逻设施、开门这类多样性较高的工作，可能需要40小时以上；而整天重复移动杯子这种简单操作，可能20小时就达到极限。但无论如何，这种模式无法通向通用人工智能。我们可能与人形机器人领域的其他公司有所不同——我认为我们更像是一家全力奔向AGI的公司，关注的是如何尽快实现这一目标，而非如何在工业场景中应用劳动力。

Yeah. And we actually have some data on this. We have some real data because our previous generation Humanoid Eve, we deployed that into both guarding and logistics back in 2022, 2023, and about twenty to forty hours our robots kind of plateau and stop learning for that specific task. Depends on how complex it is, like if you're guarding a facility and you're driving around, because that had wheels, but it was a humanoid wheels, opening the doors and there's some diversity to that, so then you're more on the forty plus hours, and if you're just moving this cup from here to over here all day, right, then you're in the lower end of 20. And there's just no path from there to general intelligence, and we are maybe kind of a bit different than the rest of the humanoid space in this, that I see us more as a company really running towards AGI and how can we come there as fast as possible versus how can we apply labor in industrial or similar settings.

因此，机器人技术服务于构建真正的通用人工智能模型，并获取足够丰富的新数据来训练这些模型。

So, robotics in service of building true AGI models and getting enough new rich data to train up these models.

你提到安保机器人需要二十到四十小时。那么家庭中各种事务的等效时长是多少？总共需要多少小时的

And you said twenty hours, forty hours for security guard robot. What's the equivalent for all the variety of things you can do in the home? How many hours of

目前还不确定。微型？大概数万小时吧。以我们当前的规模来看，多样性似乎没有上限。随着发展我们需要多元化，但这个问题很关键——因为我们要讨论目标究竟是什么？

We don't know yet. Mini? Tens of We thousands yeah. So at our current scale, we don't really see any kind of cap on diversity. It'll get there and we'll need to diversify, but I think you ask a very important question, right, because we want to talk about what is the goal?

对我而言，这不仅是AI或机器人技术，而是两者的结合。因为真正的丰裕是什么？是知识与智能的丰裕乘以劳动与商品服务的丰裕。二者缺一不可且相辅相成。社会面临的限制往往不仅存在于智能或数据层面。

To me, it's not just AI or robotics, it's a combination. Because if you think about what this is like, what is abundance, right? It's an abundance of knowledge or intelligence kind of multiplied by an abundance of labor or goods and services. You kind of need both, and they follow hand in hand. We can talk more about that, but the constraints we have in society aren't always only on the intelligence or data layer.

还存在于我们构建的基础架构之中。

They also are on the substrate that we're building on.

每周我和团队都会研究未来十年将改变行业的十大技术元趋势，涵盖从人形机器人、通用人工智能、量子计算到交通、能源、长寿等领域。没有空谈，只有真正影响我们生活、企业和职业的重要趋势。如果想获取这些分析，我每周会发送两次简报邮件，只需两分钟即可读完。

Every week, my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There's no fluff. Only the most important stuff that matters, that impacts our lives, our companies, and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two minute read via email.

若你想提前十年洞悉最重要的元趋势，这份报告正适合你。读者包括全球最具颠覆性公司的创始人CEO，以及构建突破性科技的创业者。如果你不想了解未来趋势及其影响与机遇，那这可能不适合你。免费订阅请访问dmadness.com/metatrends，比别人早十年掌握先机。

And if you wanna discover the most important meta trends ten years before anyone else, this report's for you. Readers include founders and CEOs from the world's most disruptive companies and entrepreneurs building the world's most disruptive tech. It's not for you if you don't wanna be informed about what's coming, why it matters, and how you can benefit from it. To subscribe for free, go to dmadness.com/metatrends to gain access to the trends ten years before anyone else. Alright.

回到正题。当我思考这个问题时，就像幼儿通过爬行、玩耍、探索物理世界、与不同人和事物互动来建立大脑新皮层模型。你们的NeoGamma是否相当于在多元环境中学习的婴儿？

Now back to this episode. So when I think about it, I imagine this is like why a toddler crawling around, playing, investigating the physics universe it's in, interacting with different people and different things is learning and building a model in its in its neocortex. And so it's is that basically the same your neo gamma is a is a infant learning in a diverse environment?

确实如此。某种程度上人类也是如此，但在高等动物中更明显——有多少智能是天生的本能。你肯定不希望机器人毫无目的地随机行动。

It is. Yeah. And I think just like to some extent for humans, right? Also, but it's more pronounced in older animals, like how much of this kind of intelligence is innate and part of your instincts. You don't want your robot to just go around randomly doing anything.

你希望它尝试可能成功的行为。因此传统AI模型仍有价值——通过互联网数据、模拟数据、合成数据进行训练，这些能帮你起步但无法完全抵达终点。它能实现看似有用的功能，然后通过机器人真实的交互学习循环在现实世界中不断实验。我们尚不确定这个循环能走多远。

You want it to try to do things that might succeed. So there is room here for the more cool classical AI models, where we're training based on internet data, simulation data, synthetic data, everything that everyone else is doing that's useful to get you off ground, but it doesn't fully get you there. It gets you to something that does something seemingly kind of maybe useful, and then you can experiment. You can have the robot really have this interactive learning loop, where it's learning in the real world, and that can get you. We don't know how far it can get you, right?

我们不知道具体有多少。

We don't know how much.

关于数据收集这个话题，你知道，看着它们在大楼里和厨房周围走动，真是令人惊叹。它们毫无威胁感，你会自然而然地靠近它，完全不会觉得它会做出任何尴尬的举动，比如撞到你之类。所以这一点极其重要。

And this whole topic of data gathering, you know, it's amazing watching them walk around the building here and walk around the kitchen. They're so unintimidating. You walk right up to it intuitively. You don't feel like it's ever gonna do anything awkward, hit you or anything like that. So that's gotta be incredibly important.

你就是它

You're it's

舒适。很舒适。既舒适又

cozy. It's cozy. It's cozy and it

似乎不会打碎杯子或其他东西。所以，我认为这对数据收集任务来说非常核心。对吧？因为正如你所说，必须让它去尝试。否则，它怎么学习呢？

doesn't seem to break the glasses or anything. So, I mean, that's gotta be really core to the data gathering mission. Right? Because you you have to, like you said, let it experiment. Otherwise, how's it gonna learn?

那么顺着这个思路，你们在NeoGamma中设计了哪些元素来让它适合家庭使用？

So along that lines, what what design elements did you build into NeoGamma to make it for the home?

当然。这实际上要追溯到十年前公司成立的时候。是的。真的，我在这个领域已经很久了。所以我有点喜欢

Sure. This actually goes all the way back to the founding of the company a decade ago now. Yeah. Really, I've been in the field for a long time. So I kind of like

从那时就开始了

started since the

我还是个孩子的时候。我

I was a kid. I

你几岁就开始造机器人了？

You're building robots at age what?

我11岁时就决定要研究人形机器人。

I was 11 when I decided that I was gonna like do humanoids.

那你当时模仿的是哪款人形机器人？是《星球大战》里的？《星际迷航》里的？还是《迷失太空》里的？到底是什么？

And what was what was your what was your humanoid robot that you that you modeled? Was it Star Wars? Was it Star Trek? Was it what was it? Lost in Space?

本田的Asimov。什么？本田的Asimo。Asimo。

Honda Asimov. What's that? Honda. Honda's Asimo. Asimo.

Asimo。Asimo。对。那是个很棒的机器人，对吧？他们起步很早。

Asimo. Asimo. Yeah. It's a beautiful robot, right? They started very early.

你可以去看看本田的...哦对，对。当然。

And you can check out like the Honda Oh yeah, yeah. Sure.

的

Of

当然。现在有更现代的型号，但像本田Asimo P6是九十年代末的产物。对。那时候就能上楼梯了。没错。

course. There's more modern ones, but like the Honda Asimo P6 was like end of the nineties. Yeah. And that was walking upstairs. Yeah.

在舞台上跑来跑去，给人递球。

Running around stage, giving someone a ball.

它还曾向奥巴马总统问好。

Like it greeted president Obama.

是的。我记得有一次。那是稍晚些时候的事了，不过确实如此。

Yeah. I think at one point. That was a bit later, but yes.

好的。他们，

Okay. They,

而且它当时非常超前，对吧？但这些年来我构建了很多东西。重要的是，我创办公司时坐下来深入思考过：我们研发了这么多出色的机器人，却未能真正成功。为什么？归根结底是几个基本原则——首先，若要打造可扩展的智能体，它必须能在人类社会中生活学习。

And it was so ahead of its time, right? But I built a lot of stuff up through the years. But I think importantly, I started the company, I sat down and I thought really deep about this, okay, there's all these amazing robots that we worked on, and it didn't really work. Why didn't it work? And then it comes down to these fundamental principles of, first of all, if you actually want to make something that's scalable with respect to intelligence, it needs to be able to live and learn among us.

日常生活中存在无数细微差别。我们所有行为都具有社交属性——工作是社交，每项任务都是社交。我们始终在完成事务的同时处理这些社交情境。

And there's just so many nuances to this through everyday life, right? Everything we do is social. Like work is social. Every task is social. And we navigate these social situations all the time while we do the things we do.

世界上大部分劳动也发生在社交语境中，即当你在

And most of the world's labor also happens in a social context, in that there are

执行时周围有其他人在场。

other people around you when you do it.

物体也具有社交属性。咖啡杯空了——你是需要新杯子？还是这个脏了？或是想续杯？又或者你习惯整天把杯子放在外面？

Objects have social context, right? The coffee cup is empty. Do you need a new is it dirty? Or do you want a refill? Or do you keep your cup out for the day?

就像存在这种你想触及的多样性角色。如果你深信这点，就会归结到：从第一性原理看，机器人必须绝对安全，不能伤害人类；同时要高度能干，具备人类级力量；最后必须极其经济实惠。需要找到这种精妙平衡——通过不断简化仍能获得强大系统，从而实现规模化生产，真正提升质量降低成本。

And like, there's this like, role of diversity that you want to access. So kind of if you're a big believer in that, then it boils down to, okay, role needs to be safe from a first principles point of view, not able to harm people. It still needs to be very capable, it needs to be as strong as a human. Then it just needs to be incredibly affordable. Like you need to find this beautiful combination where you can simplify, simplify, simplify, and still get a very capable system so that you can manufacture this at scale and really drive quality up and cost down.

没错。这就是十年前公司的创立原则：我们要制造安全、能干且经济实惠的机器人。所谓实惠，就是从第一性原理出发实现可制造性——超轻量化、超高能效（用小电池）、极简零件、无需精密公差、无需特殊合金材料，最终达成极致简约却高性能。

Right. So that was the founding principle of the company a decade ago. They said, like, we're going to make robots that are safe, capable and affordable. And by affordable, mean, it's going to be like first principles, manufacturable and affordable. Very lightweight, very energy efficient, so you can have a small battery, very few parts, designed in a manner that doesn't require tight tolerances, no special alloys or materials, and just like incredibly simple but performant.

这正是我们的初衷，也是耗时十年的原因。因为公司进行了大量创新研究，才最终实现这些肌腱驱动机器人——

That's really what we set out to do. And that's also why it took a decade, right? Yeah. Because like there's so much novel research that's been done in the company to get to where we have these tendon driven robots that

这个愿景相较于汽车如何定位？比如每户一台？两台？你提到iPhone直接面向消费者销售达十亿部，但一人一部很明确。而机器人可能是两台、四台，甚至更罕见？

What's the vision there relative to the car, say? Like one in every household? Two? And you mentioned the iPhone, the go direct to consumer iPhone sales get to a billion, but it's exactly one per person is pretty obvious, right? But robots, it could be two, it could be four, be rare.

我做过那个调查，大家普遍表示根据价格点至少会买两台，对吧？所以从价格点考虑，我听到的数字是3万、2万。我们见过价格更低的中国机器人，但性能不如Neo Gamma。你们有考虑中的定价区间吗？

I've done that poll and everybody routinely says I would have at least two depending on the price point. Right? So price point wise, when I think about this, what I've heard is 30 k, 20 k. We've seen Chinese robots in much cheaper price points, but not as capable as Neo Gamma. Do you have a price point that you're thinking about?

你的估计很接近了。实际价格比人们想象的要便宜。有意思的是，我认为这非常关键。我们要确保不仅做出最好的产品，还要保持价格竞争力。

You're not far off. It's cheaper than what people think. Okay. It's quite interesting because I think this is very important. I want to make sure that we are not only making the best product, we want to be price competitive.

这将是极其重要的。实际上我们与中国产品相比仍有价格优势，但就像你说的，这不是同等级别的比较。如果考虑机器人的自由度数量——也就是关节数量代表的性能——我们的单位成本其实低得多。我们在降低复杂度方面做得很好。

I think that's going to be incredibly important. And we are actually still price competitive with the Chinese ones, but you have to count, like you said, it's not the same, right? So if you think about the number of degrees of freedom that the robot has, like how much capability, basically how many joints, then we actually have a significantly lower cost. So I think we've done a really good job on reducing complexity to get there.

Burnt，我记着的数字是3万美元购买价，或每月300美元租赁价，相当于每天10美元，每小时40美分。这个范围准确吗？

The numbers, Burnt, that I keep in my mind is like 30 k purchase or $300 a month to lease, $10 a day, 40¢ an hour. Am I in the right range there?

是的。我们还能更优惠，但这个区间没问题。

Yeah. I think we could do better, but yes.

好的，这太棒了。

Okay. Think that's fantastic.

但说实话需要更优惠吗？不必。说真的，这个价格我立马就会买单。

But I mean, do you need to do better? No. Mean, obviously, heartbeat, in I'll pay that.

毫不犹豫，这个价格完全可以接受。这样的话，人们应该会考虑买两三台机器人。所以我认为...

In a heartbeat, that's good enough. Yeah. In that case, I think people could imagine owning a couple of those robots. So I think

这取决于你的视角。显然每个人都想要机器人。其中被严重低估的是它的陪伴属性——人形机器人是AI的绝佳载体。当你们交流时，它能通过肢体语言、视线追踪和定向音频识别对话者。

it really depends on the lens you see this through. So I think clearly everyone's going to want a robot. And I think that this beautiful thing about the companion aspect of this, which is so underrated, right? Because the humanoid is just such a beautiful interface for AI. And when you talk to it and you see the body language, it can look at you, it sees who's talking to it, directional audio, all these things.

像我11岁女儿如果有机器人，她只想和它坐在沙发上聊天。这将成为重要使用场景。它既不是宠物也不是人类，而是介于两者之间的存在。就像《卡尔文与霍布斯》里的霍布斯——如果你读过的话。

Like all my 11 year old daughter can do, if she has a robot, she actually just wants to sit next to it on the couch and talk about things. And that is clearly going to be such a big aspect of it. I see it as like it's not, I'd say it's not another pet, but it's not another human either. It's something kind of in between. And like I said, it's kind of like my Hobbs, Like, if you ever read Calvin and Hobbes, it's the hops.

我认为观察这些关系如何发展将会极其令人兴奋，因为它将伴随你一生，对吧？它会记住关于你的一切。而且

And I think it's going to be incredibly exciting to see how these relationships develop because it's the thing that will be around you all your life, right? It will remember everything about you. And

它它会喜欢那些真正...如果你将C-3PO那种助手机器人的构想与你实际打造的产物相比较，立刻让我印象深刻的两点是：第一，它是柔软的，不像金属外壳。第二，声音完美无缺。哦谢谢。当你和它交谈时，会立刻放下戒备，就像和普通人对话一样，因为它没有C-3PO那种机械嗓音。

it's it's gonna like things that are like really if you compare a c three p o in that vision of a of a assistant robot and you compare it to what you've actually built, the two things that jump out at me right away, one, it's soft. It's not like a metal outside. And two, the voice is perfect. Oh, thank you. Like, when you're speaking to it, you immediately are disarmed and you just talk to it because it doesn't have a C three p o robotic voice.

它拥有一种完全令人放松的自然声音，并且对你说的任何话、任何手势或动作都能做出非常灵敏的反应。

It has just a perfectly soothing normal voice and it's very responsive to anything you say, any gesture or anything.

所以我设想这些机器人都将配备先进的人工智能，达到比如GPT-5或Gemini-3的水平。因此这些机器人将拥有超高智能，能够完全理解并满足你的需求。一旦它们完全掌握了物理模型，就能执行任何你需要的任务。你们决定自主构建AI系统的决策，我觉得这非常有意思。

So I imagine that these robots will all have advanced AIs at the level of, you know, GPT-five or you know, a Gemini three. And in so being those robots will be hyper intelligent and able to understand fully and answer what you need. And once they've learned the physics models fully, do whatever you need. You've made a decision to build your AI systems in house. And I find that fascinating.

事实上，其他许多人形机器人公司——并非要将你们置于对比情境——也做出了相同选择，而非与大型超大规模云服务商合作。你能谈谈这个决策吗？

And in fact, a number of the other robot companies, humanoid robot companies, not going to put you into a comparison mode here, but have made that same decision versus partnering with the large hyperscalers. Can you speak to that?

我们讨论的并非同一件事。在我看来，智能并非始于语言。语言更像是一种我们创造出来的生成性人工建构，它确实令人惊叹——能以如此高效、凝练的方式传递意义与指令。语言固然极其实用，但并非智能的核心。

Well, we're not doing the same thing. I mean, to me, intelligence does not begin with language. Like language is this generative artificial construct that we have come up with, and it's incredible. I mean, it's such an efficient, compressed way of conveying meaning and instruction. So language is very useful, but it's not the core of your intelligence.

智能的核心在于空间与时间维度，关乎你如何感知周遭世界——既包括视觉层面的观察，也涉及情感层面的体验。我们现在正逐渐认识到，那些原生于此模态的模型，在加入文本后，将比语言优先的模型展现出更高层次的智能与能力。

The core of your intelligence is spatial and temporal, and it has to do with how you perceive the world around you, both with respect to how you see the world, but also how you feel the world, right? And we're getting to where we're seeing that models that are native to that modality and then you add text will be more intelligent and more powerful than the language first.

据我所了解的智能理论，普遍认为智能的存在需要具身化，而智能的扩展则需要语言作为载体。

I mean, I've read about intelligence and the belief is that you needed embodiment for intelligence to exist and language for intelligence to scale.

我无法严格证明具身化的必要性，但从工程角度而言，这无疑是条更便捷的路径。试想世界中的信息获取：你可以训练一个能预测视频的世界模型，让它实时生成新视频帧。理论上当然也能仅通过文本来训练——如果有足够多的文字描述，或许最终能提取出有效信号，特别是配合RLHF等反馈机制来评估生成质量。但这就像舍近求远，既然最终要输出视频，直接训练视频模型才是正途。

I don't feel that I can prove, I don't have rigorous proof that embodiment is needed. I do have very, very strong proof that from an engineering perspective, it's just a way easier path. So if you think about the information in the world and can you access this? You could train a world model that can predict video and tell you like, hey, here's a new video frame, right? Render this for you.

理论上你确实可以仅用文本训练这样的模型——如果有海量物体文字描述，或许最终能获得足够高的信噪比，至少通过类似RLHF的反馈循环来评估生成帧质量。但何必如此？这简直是效率最低下的方式。既然要输出视频，直接训练视频模型才是明智之选。

You could, in theory, could probably train that only on text. Like if you have enough text descriptions of things, maybe at some point you could get a high enough signal noise that you actually can get something useful out, at least if you kind of have some feedback loop with some RLHF or something where you're like, am I happy with this frame? But I mean, why would you do that? That's just like such an inefficient way of doing that. Of course, you train on video because you're going to output video, right?

因此从这个角度来看，我认为显而易见的是，要想首先达到人类水平的智能并有望超越它，你需要我们体验到的所有模态。但关于机器人，我认为还有另外两点实际上非常重要。

So from that perspective, I think it's just obvious that you need all the modalities that we experience if you want to get to first and foremost human level intelligence and hopefully pass that again. But then I think there's one other thing about robots there's two other things actually that are quite important when it

涉及到学习时。

comes to learning.

第一点相当明显，我想我们都认同这一点，即机器人可以进行交互式学习，对吧？你与世界互动，从而能够学习。但如果你从更学术的角度思考这些智能如何演化、如何获得推理能力等问题，我们通常的做法是对世界进行观察，了解世界的运作方式。我知道如果我这样做，会发生什么。对吧？

And the first one is quite obvious, and I think we all kind of identify this, which is like robots can do interactive learning, right? So you interact with the world and therefore you can learn. But if you think about it more from a kind of academic point of view of like how those intelligences kind of evolve, how you get reasoning, all these things, then what we generally do is that we have some observation of the world, like we kind of know how the world works. So I know that if I do this, I know what is going to happen. Right?

我以前见过这种情况。所以我实际上是从这一点开始的，我有一个目标——我想拿起杯子。现在我有了一个世界模型，有了拿起杯子的目标。

I've seen this before. So I actually start with that, and I have a goal. I want to pick up the cup. So now I have a model of the world. I have a goal of picking up the cup.

我采取了一个动作。我知道我采取了什么动作。我知道我采取的动作是伸手去抓杯子。然后我观察结果。

I take an action. I know which action I took. I know the action I took was to, like, reach for and grasp the cup. Yeah. And then I observe the result.

如果你看互联网或YouTube，你只有观察。你不知道视频中人的心智模型，不知道他们采取了什么动作，也不知道他们试图实现什么。

If you look at the Internet or in general, you can look at YouTube, All you have is just the observations. Right. You don't have any of the mental model of the person in that video. You don't know which actions they took. You don't know what they tried to achieve.

你只有观察。这不是我们学习的方式。我们可以追溯到科学方法：你应该有一个理论，提出假设，测试假设，观察结果，然后重复这个过程并学习。这在互联网数据中是不可能的。

You only have the observation. This is not how we learn. We can actually bring it all the way back to the scientific method. It's like you should have a theory, come up with a hypothesis, you test your hypothesis, you observe the result, and then you do it again and you learn. And that is just not possible with the internet data.

所以

So there's

绝对不可能通过下一个标记或原始互联网抓取和所有视频抓取实现。因此，在编码和物理实验等有限领域，你可以有相同的体验，但仅限于该领域。编码就是一个很好的例子：比如，我这样写不行，那样写也不行。

Definitely definitely impossible with the next token, you know, raw Internet scrape and with all the video scrape. So then in these limited domains like coding and physics experiments, you can actually have that same experience, but it's only within that domain. Like coding is a good example. Like, oh, let me try writing it this way. It didn't work.

你在这个狭窄的领域变得非常非常擅长，但仍然对世界的运作方式毫无直觉。

Let me try writing it that way. Didn't work. So you get very, very good at that narrow domain, still have no intuition about how the world works.

不，你们确实在做模拟。重申一次，很难证明这方法行不通，对吧？当然，如果你有一个非常优秀的模拟器，并且真正扩大模拟规模、结合学习与智能体模拟，或许能得到类似效果。但关键在于，模拟器的保真度远不及现实世界，要弥合这个差距极其困难，而且相比直接接触现实世界，这种方式的算力效率也低得多。不过对我来说，核心不在于这场学术辩论谁对谁错。

No, you do simulation, though. Again, it's hard to prove that this won't work, right? So sure, if you have a really good simulator and you really scale simulation and learning and simulation with agents, maybe you can get something similar. But I mean, the fidelity of your simulator is nowhere near the real world, and it's just so incredibly hard to get there and close that gap, and it's also so compute inefficient compared to just being in the real world. But I think for me it boils down to not this academic exercise of proving who's right and wrong.

更重要的是：什么样的工程方法在这里是合理的？而这显然是条更短的路径。

It's more what's the engineering approach that makes sense here? And it's just a way shorter path.

你之前提到过与谷歌、YouTube或特斯拉相比的数据收集量。能否具体谈谈？我们的使命就是尽可能多地收集这些家庭机器人在日常互动中产生的数据。

You mentioned before in our conversation the amount of data that's being collected relative to Google or YouTube or Tesla. Can you speak to that? I mean, mission is get as much possible data during the day of an interaction of these robots in the home.

没错。你可以做个粗略估算——当然我们目前还不完全清楚哪些模型需要哪些最有价值的数据。但想想看，如果有1万台机器人全天候收集数据，其产生的非重复有效数据量将超过YouTube每日上传量。仅在这个规模下，你的机器人舰队生成的有效数据就已超越YouTube。

Yes. I mean, you can do some napkin math, right? And of course, we don't know exactly like what is the most useful data from which models, etcetera, yet. But if you think about it, if you have 10,000 robots out there and they gather data most of the day, then that is more data than the non duplicated useful data that gets uploaded to YouTube each day. So already at that scale, you actually have your fleet of robots generating more useful data than YouTube.

这还只是1万台的情况。如果再考虑规模化生产部署到社会后，你会很快意识到：互联网其实没那么庞大。机器人产生的数据量将远超互联网数据量。

So that's just at 10,000. And then if you think about like how we scale manufacturing here as this starts deploying into society, you actually very quickly come to the conclusion that like, know what, the internet isn't actually that big. Like, you're going to have way more data from robots than you're going have from the internet.

所以我想在这里提出一些数字作为基础。你们已经生产了大约数百台NeoGamma。但你们即将启用一个新的制造工厂，还有一个正在规划中，当然在不透露任何你不愿意分享的信息的前提下。能否请你大致说明到2026年，你们的年产量预计会达到多少？

So I want to hit some numbers here just to set them as foundations. You built hundreds of the NeoGamma, roughly. But you're about to you got a new manufacturing plant that you're about to open. Can you give us a sense of and then another one that's in plans, right, without disclosing anything you're not willing to. But can you give me a sense of by the '26, how many you're manufacturing on an annual run rate?

那么在2027到2028年，你预计的增长路径是怎样的？

And then in 2728, what's the growth path you imagine?

是的。首先做个小更正，我们生产的摄像头数量尚未超过100台，但机器人产量已超过100台，并且有多个版本。

Yeah. First of all, just a small correction. We haven't built more than 100 of the cameras. But we've built more than 100 of the robots. There's been multiple versions.

但到2026年，工厂的年产量将超过2万台。

But the factory run rate 2026 is north of 20,000.

年化产量2万台。

20,000 annualized.

年产量，没错。当然，需要逐步提升才能达到，所以2020年还达不到那个数字

Annual, yep. Of course, a ramp to get there, so you don't reach quite that number in 2020

那么

So

每月几千台。

a couple thousand a month.

接下来工厂的情况，我们试图遵循数量级递增的原则，对吧？但实际很难完全做到。我认为iPhone的产量增长曲线是个很好的参照——你看它们几乎翻倍增长，但在达到某些规模时会遇到瓶颈期。如果要让人形机器人的制造规模达到iPhone级别，确实会面临一些非常有趣的问题。比如连铝这样的基础材料都可能耗尽。

Now, the factory after that is kind of like we're trying to follow an order of magnitude, right? We're not going to quite be able to do that. I think the iPhone ramp is a very good comparison here, where you see they almost doubled, but you have a few plateaus as you reach certain scales where you run into problems. And there are some quite interesting problems there if you're going to scale the manufacturing of humanoids to the iPhone level, right? Because you run out of some basic stuff like aluminum, for example.

我指的不是地球上没有旧铝可用。但当现有铝材精炼产能的使用比例达到一定程度后，铝原料采购就会变得极其困难。这可能会是个挑战。

You don't use old aluminum on the planet. That's not what I mean. But there's a certain amount of percentage of current refinement of aluminum you can use before you start to really struggle sourcing aluminum. And that might be a challenge, I think.

等等。iPhone的增速大约是翻倍。这个有趣的统计数据我甚至没想过。我是说，你达到了十亿

Wait. The iPhone ramp was about doubling. That's an interesting stat I hadn't even thought of. I mean, you get to a billion

更像是1.7倍。

It was more like 1.7.

年化1.7倍 哇。超过 那就不算太夸张

1.7 annualized Wow. Over That's not as massive

所以你可以想象100倍 嗯，

as So you can imagine a 100 Well,

指数级增长还很遥远。

exponentials are quite far.

是啊。不。我不知道。

Yeah. No. I don't know.

我们听说了，我们已经听说了

We've heard, we've we've heard

实际上是那些点击声。

the clicks, actually.

所以，但你可以想象在本十年结束前，年产量将达到数十万。

So, but you can imagine a run rate before the end of this decade of hundreds of thousands per year.

本十年结束时，

End of this decade,

远远不止。按那个速度算的话，还要多得多。

way more. Way more at that rate.

没错，没错。到了那个阶段，你真的需要思考哪些因素会拖慢进度，对吧？归根结底当然是提炼和采矿这些环节。但越来越关键的因素其实是劳动力。比如，不大量使用机器人劳动根本不可能实现目标。

Yeah, yeah. Now, at that point, you need to really think about what are the things that will slow you down, right? And it comes down to refining, mining and refinement, of course. But increasingly, it actually comes down to labor. Like, you're not going to get there without really using robots for labor.

想想iPhone产能爬坡时期，苹果几乎调动了中国全国范围的劳动力。即便如此还是面临人力短缺，不得不将供应链扩展到周边国家。

If you think about the iPhone ramp, then Apple kind of displaced a large part of the Chinese population across the country for labor. And they still ran out of labor and had to expand into neighboring countries.

太疯狂了。

That's wild.

现在我们在设计方面做得非常出色，零件数量极少，组装极其简单。不过相比iPhone，组装复杂度还是更高些。所以你刚才说

Now, I think we've done an incredible job in the design. So it's very few parts. It's very simple to assemble. But it's still more complicated to assemble than an iPhone. So You said

我本来打算

what I was going

要说的是，它

to say, it

看起来更复杂。是的，

looks more complicated. Yeah,

比iPhone更复杂，对吧？假设它需要五倍的时间，那么我们就需要五倍于iPhone的人力。好吧，那你就麻烦了。

is more complicated than an iPhone, right? So let's say it takes five times as long, so we need five times as much labor as the iPhone. Okay. Then you're trouble.

这是个不错的衡量标准。

That's good metric.

那你就陷入困境了。

Then you're in trouble.

所以这个问题必须解决。

So it's got to be solved.

所以你必须实现自动化，对吧？当然，这本来就是目标。我们希望能尽快达到我所说的这个硬起飞时刻，那时机器人可以制造机器人，机器人可以扩建数据中心、芯片工厂和能源基础设施。

So you have to automate, right? And of course, that's the goal anyway. We want to get as quickly as possible to what I call this hard takeoff moment, where you have robots building robots, robots building out the data centers, the chip fab, the energy infrastructure.

实际上，我们能从汽车中学到什么？这边有iPhone，零件更少，每单位人力是五分之一。那边有汽车。零件的数量与汽车相比如何？

What can we learn from the car, actually? So you've got the iPhone, fewer parts, one fifth the labor per unit. Then over here you have a car. How does the part count compare to a car?

我们大约有几百个零件。汽车大约有五万个。

So we have a few 100 parts. The car has 50,000 roughly.

5万。所以简单多了。

50,000. So it's much simpler.

我是说，一辆车重达4000磅。

I mean, a car weighs 4,000 pounds.

是啊。用了很多材料。

Yeah. A lot of material.

我们的机器人重66磅。

We we our robot weighs 66.

好的。

Okay.

所以我认为，它其实不能和汽车相提并论。我看到很多人把仿生机器人和汽车比较。但说实话，这种比较应该重新考虑。它根本不是车。如果做得好的话，它更接近冰箱。

So I think, like, it's not really comparable to a car. I've seen a lot of light of space compare humanoids to cars. But I think then you should go back to the drawing board, to be honest. Like, it's not a car. If you do a really good job here, it's closer to a refrigerator.

是个非常复杂的冰箱，但比起汽车，它确实更像冰箱。

It's a very complicated refrigerator, but it's closer to a refrigerator than cars.

让我们为观众和听众深入了解一下这个机器人的外形设计。66磅重。谈谈它的电池续航和功能吧。能否从具体参数角度描述一下？

Let's dive into a little bit of the shape, the understanding of the robot for our viewers and listeners. 66 pounds. Let's talk about battery life, its abilities. Describe it from a specific stats point of view, if you would.

哦，当然。我认为最重要的参数是它可拥抱的特性。

Oh, yeah, sure. So I think, first of all, I think the most important stat is this huggable.

可以拥抱。没错。确实可以拥抱。我抱过机器人。

It's huggable. Yes. It is huggable. Have hugged a robot.

是的。这关乎安全性，以及如何在它的空间里感到安全、舒适。但从纯粹的统计数据来看，它重66磅，能举起约150磅的重量。

Yes. And like this is a safety and how it is to feel safe in its space, soft. But from a pure stats point of view, it's 66 pounds. It can lift about 150 pounds.

这太棒了，我是说，从重量与力量的比率来看。

Which is amazing, I mean, in terms of the weight strength ratio.

这相当于一个运动员的体重力量比。是的。而且它能携带大约50磅的东西四处走动。希望你之前已经看到了。电池续航大约是四个小时。

It is the weight to strength ratio of an athletic human. Yeah. And then it can carry like about 50 pounds around. So you hopefully saw earlier here. And battery life is about four hours.

充电需要半小时还是两小时？

Rechargeable in half an hour or two hours?

一半时间。如果用完整个电池，大约两小时。实际上，有趣的是，我家里就有一个，对吧？所以我现在开始收集一些相关数据了。

Half of it. So like two hours if you use a full battery. Now, actually, interestingly enough, I have one in my house, right? So I'm starting to get some data on this now.

五英尺四，五英尺五。这是什么？

It's five foot four, five foot five. What is it?

五英尺

Five foot

四。五英尺四。好的。

four. Five foot four. Okay.

我想是的。那是个

I think so. That's a

顺便说一句，完美的身高，以防你在想。

perfect height, by the way, just in case you're wondering.

是啊。这也是我妻子的身高。这也是我妻子的身高。所以就像，我同意你的看法。

Yeah. It's also the height of my wife. It's also the height of my wife. So it's like, I agree with you.

这是我的。所以

It's mine. So

那很好。是的。但我想说的是，真正开始使用产品时你会发现很多参数表上不会体现的东西，这非常有趣。比如机器人完全静音。这并非偶然。

that's good. Yeah. It's but I mean, it's what's very interesting, like once you start actually using the product, right, you notice a lot of things that don't usually show up on a spec sheet. Like the robot is completely quiet. And that's not a coincidence.

这是我们花了很大功夫实现的。当你第一次把它放在家里，觉得机器人很安静，没问题。第一天还行，第二天有点烦人，到第三天你就会想：天啊它能不能快点离开客厅？因为这种声音对吧？绝对的静音是硬性要求。你得考虑它在你的空间里如何运作。

That's something we worked so hard on. And the first time you put this in your home and you think like the robot is very quiet, it's fine. You put it in your home and you're like, first day it's fine, second day it's a bit annoying, like third day you're like, oh man, is it going leave my living room soon because like this sound, right? It's such a requirement for like just dead quiet, right? Figure out how this in your space.

充电方面基本不会遇到问题。因为机器人会时不时进行微型充电。当它不工作时...其实我不太关心它能运行多少小时，更在意充电速度是否足够快，让它随时能完成我

Charging wise, don't really ever run into the problem. Because the robot just takes these micro breaks every now and And then when it's not doing I actually don't care that much about how many hours it can run. Care about that it charges fast enough that it can just always do whatever I

想做的事。不错。

want to do. Nice.

对。既然你提到参数，我想快速讨论下自由度数量的问题。这基本上是指机器人有多少个关节对吧？人类每条腿有6个关节，共12个。如果每条手臂有7个，再加14个。

Yep. Well, and I want to talk about quickly, since you said that, specifications, like the number of degrees of freedom, right? Which basically is how many joints does the robot have, right? So humans have six joints in each leg, that's 12. If you have seven in each arm, that's 14 more.

所以12加14等于26。现在很多机器人都有26个关节，这很常见。通常它们没有腕关节，但有颈部关节。这里两个，加起来就是26。

So now you're 12 plus 14, that's 26. You see a lot of robots today that have 26, that's quite common. Usually they don't have the wrists. They have the they actually have the neck instead. So two here and then you're like at 26.

我们这里有三个。所以头部能做出丰富表情，这很重要。这里有完整的七个关节，脊柱有三个。

We have three here. So you have proper expression with your head. It's quite important. We have all the seven from here. We have three in the spine.

当然每只手还有22个关节。

And then of course we have 22 in each hand.

我是说，我在手臂设计中看到的令人难以置信。

I mean, what I saw in the arm design was incredible.

人类的手部结构是怎样的？

How do humans have in their hand?

22块骨头。

22.

所以你匹配了这个数量？

So you matched it?

嗯，好吧。取决于你怎么计算毛细血管骨。就是那些让你能弯曲手掌的小骨头。某种程度上你可以看作是有四到五个自由度，而不只是两个。所以人类的手其实更灵活些。

Well, okay. Depending on how you count your capillary bones. So like the small bones that you have here that allow you to cup your hand. You could to some extent see that that's more like four or five degrees of freedom, not really two. So then the humans have a bit more.

功能上非常相似。再说一次，这对在家庭环境中完成各种任务极其重要。但从AI角度来看，就像我们最初讨论的多样性，对吧？这是衡量智能的一个指标。多样性，多样性...

Functionally, it's quite similar. And this, again, just is incredibly important to be able to do all those tasks in a home. But also from an AI perspective, there are like we talked about diversity initially, right? It is the one metric for intelligence. And the diversity Diversity

环境和数据的多样性。

of environment and data.

嗯，数据的多样性。而你的多样性来自两个方面：你能达到的多样性上限。这取决于你部署的环境。比如在工厂每天重复同样工作，机器人再好也缺乏多样性。其次是机器人有多强的能力？

Well, diversity of your data. And your diversity comes from two things, the limit to the diversity you can achieve. It comes from the environment you're deploying in. So right, if you're in a factory doing the same thing every day, it doesn't matter how good your robot is, it's not going to be diverse. And then how capable is your robot?

它能做多少事情？对吧？如果它不能进行手部精细操作，处理柔软可变形物体、精密物品这类东西，那你就无法获得相关数据。所以想要最大化多样性，必须在两方面都做到极致。

How many things can it do? Right? Because if it cannot do any kind of like in hand manipulation or handling like soft deformables, all these kind of things or delicate objects or whatever, then you get no data on that. So, like, you really have to kinda go max max on both, right, if you wanna maximize your diversity.

大约十八个月前，我和我最亲密且才华横溢的朋友戴夫·布伦登合作创立了Link Experiential Ventures。在LINK，我们管理着约十亿美元的种子资金，总部位于剑桥肯德尔广场，就在MIT和哈佛之间。当戴夫和我从MIT毕业时，我们都立即创办了公司。但在那个年纪，所有条件都不利。你有想法，却难以融资，连房租都负担不起。

It was about eighteen months ago that I partnered with one of my closest and most brilliant friends, Dave Blunden, to start Link Experiential Ventures. At LINK, we manage about a billion dollars of seed stage money based out of Kendall Square in Cambridge, right between MIT and Harvard. When Dave and I both graduated from MIT, each of us immediately started companies. But at that age, everything is working against you. You have an idea, you're challenged to raise money, and you can't afford rent.

尽管有众多加速器存在，你仍在与成千上万其他初创公司争夺同一批投资者。戴夫和我都花了大量时间思考如何激励和支持创始人突破这些障碍，实现快速发展、创造财富、影响世界，尤其是在当今这个AI无处不在的时代。我们看到许多公司在短短两三年内就达到数十亿美元的估值，速度前所未有。有些公司甚至几周内就能增值数百万乃至数千万美元。于是我们开始自问：如何帮助这些创始人不间断地加速前进？

And even with all the accelerators out there, you're competing against thousands of other startups for the same pool of investors. Both Dave and I have spent a big chunk of our lives focusing on how do we inspire and support founders to knock down those barriers, to go big, to create wealth, to impact the world, to build and scale as fast as possible, especially in today's AI everything world. We're seeing so many companies reaching multibillion dollar valuations in just two to three years faster than ever before. Some companies are adding millions or tens of millions of dollars of value in just weeks. So we started asking ourselves, how do we help these founders go faster and not skip a beat?

举个例子，几个月前我们在MIT旁边购置了一栋公寓楼，让毕业的创业者能立即入住，在寻找长期住所时不影响技术开发进度。我们正竭尽所能加速建设者及其高智商团队的发展——资金支持、导师指导、连接我个人的CEO及投资人资源网络都是其中环节。我们拥有6.6万平方英尺的专业孵化空间，26家AI初创公司将Link XPV视为家园，回报率令人惊叹。

As an example, a couple of months ago, we bought an apartment building adjacent to MIT where a graduating entrepreneur can move in immediately without slowing down their tech build while they search for a place to live. And so we're doing everything we can to accelerate builders and their super smart teams. Of course, funding is part of it, mentoring is part of it, connecting them with my personal network of abundance minded CEOs and investors as part of it. We house 66,000 square feet of purpose built incubator space and 26 AI startups call Link XPV their home. And the returns have been amazing.

我没什么要问的，但如果你正在AI时代创业，欢迎访问linkventures.com了解我们。现在回到节目。

I have nothing to ask, but if you are building a company in the AI era, check us out at linkventures.com. Now back to the episode.

有个极客问题想请教——当你在物理实体上搭载神经网络时，很难判断功能限制究竟来自神经网络还是硬件物理结构。有没有方法能分离这两者进行调试？这算是两个独立层面吗？

Yeah. Geeky question for you, but really, really curious to know. Because when when you build something physical and then you attach a neural net to it, it's actually very hard to tell whether the constraint in what it can and can't do is in the neural net or in the physical construction of the hardware. Is there any way to decouple that and debug? You know, is it two different sides?

还是说这根本不可能？一旦两者融合就...好吧我们其实有

Or is it just incredibly impossible? I mean, once it's meshed together, you just can't Well, we have

一个相当不错的神经网络。通常我的解决思路是：能否在TallyOp中实现？如果可以，就意味着只要有足够数据，合适的神经网络就能做到。有趣的是...

a pretty good neural net here. Yeah. So usually the way I approach this is, can we do it in TallyOp? And if we can, the right neural net can do it with enough data. Interesting.

这个理论基本被验证成立。当我们在TallyOp中实现某项功能后，只需要收集大量类似任务的多样化数据，借助迁移学习并针对特定任务获取海量数据，那么无论任务多复杂几乎都能实现。当然这不代表能实现所有跨领域泛化——我们还没到那阶段。但可以看到神经网络确实具备这种潜力，现在需要的是规模化，实现任务间知识迁移、分布泛化等能力，这些在LLM中已常见但在机器人领域尚属罕见。我们内部有些很酷的进展已经显示出...

And that's generally been proven to be true. If we manage to do something in TallyOp, it's just like, okay, now we need a lot of diverse data of similar tasks, we get some transfer learning and we need a lot of that specific task, and then almost irrespective of how complicated that task is, you can get it to work. Now, of course that doesn't mean you can get everything to work with generalization across, we're not there yet. But you can see that, okay, you can get the neural network to do this, now we need to scale it so we kind of get this beautiful transfer of knowledge between tasks and our distribution generalization and all these things that we currently see in LLMs that we don't see that much in robotics yet. We have some pretty cool stuff internally where we see some signs of Well,

我设想的情景是：你让它做法式火焰薄饼或进行显微手术，但它做不到。硬件方坚称硬件达标肯定是软件问题，而软件方则反驳...

I'm picturing, though, is you you ask it to to make crepe Suzette or you you ask it to do microsurgery, and it can't quite do it. And then you say, well, look, the hardware guy is claiming the hardware is good enough. It must be the software guy. Mhmm. And then the software guy is saying, no.

不，神经网络没问题，是硬件能力不足。

No. No. The software the neural net is fine. Mhmm. The hardware just can't do it.

然后双方就吵起来了。

And then they fight it out.

然后我们就说，好吧，我们派出最优秀的远程操作员，证明他能做到。那硬件显然也能做到。显然。这就是成功的证明。

And then we just say like, well, we bring in our best teleoperator and we say like, he can do it. Then the hardware can do it, clearly. Clearly. Proof of success.

嗯，好的。这正是我想说的。所以你们有个远程操作员选项可以控制硬件。哦，这真的很有意思。

Yeah, okay. That's where I was going. So you have a remote operator option who can control the hardware. Oh, that's really interesting.

那么接下来就会遇到难点，我们可能无法继续这样做了。为什么？因为机械手实在太出色了，它们具备极高保真度的触觉反馈。人类的手很厉害？不，是机器手。

So then you'd get like, well, but we're getting to where this gets hard, where we can kind of no longer do this. Because? Because the hands are just so good and they have very high fidelity tactile feedback. The human hands are so good. No, the robot hands.

好的。

Okay.

所以它们确实有，人类的手仍然更胜一筹。但问题是机器人手真的非常、非常好，它们拥有极其快速且高度精细的触觉。而我们无法真正高效地将这种触觉从人类身上转移过去。

So they have, the human hands are still even better. But like the problem is the robot hands are really, really good and they have really fast, highly detailed tactile. And we can't really transfer this efficiently enough from the human.

是的，因为远程操作员

Yeah, because the teleoperator

正在使用某种，我是说，回到XPRIZE的话题，对吧？那个虚拟化身挑战赛。要在那么短的时间内完成传输确实是个难题。现在我们开始看到，机器人实际上通过实时的强化学习，在操作技巧上进步神速。它能在真实环境中交互式学习如何操控物体，完成操作员只能梦想的动作。

is using some kind of, I mean, Back to the XPRIZE, right? The avatar challenge. So it's a really hard problem to transfer that fast enough. So now we start to see that the robot actually learns how to do manipulation way better from reinforcement learning in real. So you actually have the robot interactively learning in real how to handle objects, and it can do things that the operator could just dream of.

所以现在我们有点麻烦了。不，我们没法再那样做了。

So now we're kind of screwed. No, we can't do that anymore.

我想依次讨论三个话题：远程操控与全自动化的对比。

I want to talk about three things in sequence. Teleoperations versus full automation.

嗯。

Mhmm.

家庭安全。嗯。以及家庭隐私。没错。因为当你进入家庭时，这些必须是至关重要的。

Safety in the home. Mhmm. And privacy in the home. Yes. Because those have gotta be critically important as you're entering the home.

所以我们看到NeoGamma机器人在这里以远程操作模式运行，同时也以全AI模式运行。它能两者兼顾。其AI系统将变得越来越好、能力越来越强。就像我现在和Gemini三代或Grok四代对话，或者很快要和GPT五代对话一样，我感觉是在和一个高度智能的人类交谈，觉得它理解我的需求并能对我的请求采取行动。我想象最终机器人也会有相同水平的AI，让我感觉在某种意义上是在和一个完全智能的生命体对话。

So the robots, we saw the NeoGamma out here operating in teleoperator mode, but also in AI, full AI mode. And it was able to do both. Its AI systems are going to increasingly get better and better and more capable. Again, as we're talking as I'm talking to Gemini three or Grok four or, you know, GPT five soon, I'm talking to a highly intelligent human and getting a feeling that it understands what I want and it's able to sort of like take action on my requests. I imagine we're gonna have the same level of AI eventually in the robot where I feel like I'm talking to a fully intelligent being in one sense.

哦，是的。很明显。

Oh, yes. Clearly.

而且

And is

是接地气的，对吧？它实际上在某种程度上理解这种存在，而今天的大语言模型有点像...它们有这种抽象的概念，但有点像...如果你开始探究，这种表象很快就会消失。

grounded, right? That actually understands to some extent what this existence is, which today's LLMs are kind of like, they have this kind of abstract notion of it, but it kind of like, it's a facade that kind of quickly falls away if you start to probe at it.

但是

But

会达到那种程度的。想想

that will get there. Think

在远程操作模式下，人类戴着VR头显并使用触觉控制器。不。人类在做什么？

In the teleoperations mode, you've got humans wearing VR headsets and using haptic controls. No. What are the humans doing?

他们给出稍微更高层次的指令。就像引导说，嘿，把手放在这里，抓住这个东西。你不想过度约束系统。你想给它一些机会来解决如何完成任务。所以我们有自下而上的学习，为操作者提供越来越抽象的接口。

They're giving slightly more high level commands. So just guiding like, hey, put your hands over here, grasp this thing. You don't want to over constrain the system. You want to give it some opportunity to solve for how to do the task. So we have the learning coming up from the bottom and enabling a more and more abstract interface for the operator.

然后我们还有从海量数据中自上而下学习，越来越接近你希望机器人执行的通用行为。它们有点像在中间相遇，对吧，在那里

And then we have the learning of all the large amounts of data we have coming from top and getting more and more like the general behavior that you want the robot to do. And they kind of like meet in the middle, right, where

在这方面，'You're'操作员总是将自动化和远程操作结合使用。

the You're operator goes using automation and teleoperations always together in that regard

是的。因此，所有能让机器人执行远程操作员工作的功能，都是端到端完全学习的。比如，网络直接向电机输出扭矩。

and Yes. So everything that enables the robot to do anything that the operator does, teleoperator, is fully learned end to end. Like, the network outputs torques to the motors. That's

这与特斯拉和埃隆·马斯克所说的非常相似，最初自动驾驶汽车全是C++代码加少量神经网络，可能是80%C++加20%神经网络。随着时间推移，神经网络占比逐年增加。现在

very similar to what Tesla and Elon Musk were saying, where the self driving car was originally all C plus plus code with a little bit of neural net, maybe 80% plus plus 20% neural net. Then every year that went by, it became more neural net. Now

已有30万行C++代码被淘汰。是的，

there's 300,000 lines of C plus plus were eliminated. Yeah,

只剩下少数防护栏代码，其余全由一个神经网络处理。这里的情况也是如此，

just a few guardrails left. Then the rest is just one neural net. So same thing here,

对吧？是的。全是权重参数，对吧？代码可能就几百行。

right? Yeah. It's all weights, right? Like, the code is just a few 100 lines.

真的吗？是啊。全都是参数。

Is it really? Yeah. That's It's all parameters.

什么是

What's the

参数计数器？那些都是超级机密吗？

parameter counter? Is that all super secret?

算是机密吧，但和现在的神经网络比起来规模很小，因为要在机器人上快速运行。有点像你的肌肉神经系统，不过它会处理视觉输入，所以也不算特别小。

It's kind of secret, but it will be small if you compare it to today's neural networks because it's running on the robot very fast. It's kind of like your muscle neural system, but it does take in vision, so it's not very small.

好吧，这引出了我迫切想问的问题，你看过《机械姬》对吧？我看那部电影时就在想，为什么

Well, begs a question I'm dying to ask, which is, you know, you've seen Ex Machina, right? When I saw that movie, I'm like, why

别往反乌托邦方向扯，尤里子。打住。

Don't go dystopian on this, Yuriko. Okay.

但为什么大脑——那个蓝色团块——要放在头部？对。为什么不放在服务器机房？

But why is the brain, the blue blob, in the head? Yeah. Why isn't it in the server room?

这样所有机器人可以共享学习成果。

So learning is shared between all robots.

是啊。而且容量可以大得多。如果机器人一半能量都用于思考，把大脑移到服务器机房远程通讯，电池续航能翻倍。你们选择放在头部除了拟人化和酷炫之外还有什么原因？

Well, yeah. And it can be much bigger. And like if half the power of the robot is going into the thinking, you could run twice as long on a battery charge if you move it over to the server room and have it just communicate remote. Why did you choose to put it in a head, aside from being anthropomorphic and cool?

不不，和那些无关。简单来说——头部如果不放大脑就空着。其他部位简直挤爆了。要在这种微型尺寸里构建同等功率的人形机器人，还要留出完全软性化的空间，本身就是超高难度的工程问题。

No, no, it has nothing to do with that. So there are some simple answers to that, which is, I mean, the head is where nothing else is unless you put the brain there. The rest of It's like everything else is pretty freaking full. Like it's building a humanoid with this kind of like power level in such a miniaturized form and still having enough space to make it completely soft and all this. It's a really hard engineering problem.

所以问题变成：如果不放在头部，不放在实体机器人身上，还能放哪？另外有个次要原因：人脑最高带宽需求来自视觉，其次是听觉嗅觉触觉，但视觉绝对主导。你肯定希望眼睛和计算单元距离最短。

So it's like, where are we going to put this if we don't put it in the head, if we don't put it on the physical robot? Now, there's a smaller argument. The very high bandwidth thing that happens in your brain is vision and to some extent audio, smell, right, tactile, but vision just dominates. And you just want to minimize the distance between your eyes and the compute.

当真？所以传感器之间的带宽...

For real? So so the the bandwidth between the sensors,

眼睛

the eye

家用WiFi根本传输不了。

wouldn't make it over the home WiFi.

嗯，甚至都到不了机器人的胃部。

Well, wouldn't even make it down to the stomach of the robot.

真的吗？

Really?

在不过度纠结选择哪种物理接口进行数据传输的情况下。它的带宽非常高。

Without getting overly complicated on like which physical interfaces you would choose for this transfer. It's very high bandwidth.

这让我很震惊。我也很震惊。

Shocked by that. I'm shocked by that too.

但我的意思是，我们完全没有使用激光雷达、结构光、手腕摄像头之类的设备。我们纯粹是在模拟人类视觉，对吧？所以我们非常依赖这一点。因此它具有极高分辨率、极高带宽和极高频率。

But I mean, we're running like no LiDAR, no structured light, no wrist cameras, no nothing. We're running pure like emulation of human vision, right? So we're relying so heavy on that. So it's a very, very high resolution, very high bandwidth, very high frequency.

这没问题，因为人类大脑本来就很靠近眼睛。

That's fine because that's exactly where the human brain is very close to the eyes too.

确实。但这并不意味着不能在云端处理事务。我们会在云端处理一些事情，但从智能角度来看会形成层级结构。就像你的肌肉神经系统反应非常快，对吧？通常以25赫兹的频率运行，而且不会上传到大脑。

It is. Now, that doesn't mean that you can't do things in the cloud. We do things in the cloud, but it kind of becomes hierarchical from an intelligence point of view. Just like if you think about your muscle nervous system, this runs quite fast, right? It usually runs at like 25 Hertz and it doesn't go up to your brain.

有神经元分布在你全身做出决策。我们在机器人里也实现了这点。我们将部分功能下放到控制动力电子设备中。

There are neurons distributed out through your system that makes decisions. We have this in the robot. We have some of our stuff pushed to the power electronics that controls.

只是为了降低延迟，提高速度。

Just for latency, just for speed.

没错。然后你还有大脑本身，其实运行得相当快，对吧？通常在5到10赫兹之间，虽然频率不高但延迟极低。这个在机器人上运行。如果你运行的是更像1赫兹的流式处理，那通常就是LLM首次令牌传输了，对吧？

Yeah. And then you have the brain itself, which actually runs pretty fast, right? It's usually between five and ten Hertz and very, even though it's five to 10 Hertz, very low latency. And this runs on the robot. Now, if you're running more like a one Hertz streaming thing than typically an LLM first time to token LAN, right?

是的。

Yeah.

那虽然能脱离主板运行，但无法解决高频触觉反馈操控任务。速度实在太慢了。

That runs off board, but that can't solve the high frequency tactile feedback manipulation tasks. That's too slow.

好的，当我的NeoGamma第一次学会打鸡蛋做煎蛋卷时，问题是所有NeoGamma都会同步学会这个技能吗？你们是共享学习的吗？

Okay, the first time my NeoGamma learns to crack open an egg to make an omelet, the question is, do all NeoGammas then learn that? Are you shared learning?

确实如此。现在存在两种学习模式：一种是云端模型共享学习，所有NeoGamas都能获取这些数据；另一种是分布式模型。当然，我们每晚都会进行模型检查点更新——如果某个模型表现更优且拥有更多数据。

They do. Now there's a shared learning in the sense that you can say, this data goes to the cloud model that is doing this for all NeoGamas. But there's also the distributed models. So of course, there would be a nightly checkpoint where, hey, this model is better. We have more data.

我们已通过安全验证（具体验证方法稍后会讲），然后将更新部署到所有机器人。虽然采用分布式架构，但机器人之间仍能互相学习，只是需要通过服务器层进行单跳训练和更新分发。

We validated this. We gave it to safety, which I'll talk about later when it comes to how you validate the models. And then we deploy that to all the robots. So even though it's distributed on the robots, they can still learn from each other, of course. It's just you need to do one hop through the server layer and do the training and propagate this out.

在不远的将来，我非常看好设备端联邦学习的大规模应用。这关乎如何让您的终身伴侣机器人能从那些珍贵且私密的经历中持续学习。

There is a future not so far away where I'm pretty bullish on there being a lot of federated learning happening on device. And this has to do with how do we have your companion really throughout life learn from all of the experiences that are dear to you, but private.

没错。

Yes.

所以所有机器人不会完全相同，但它们会共享智能核心架构。

So all robots will not be the same, but they will share an intelligence backbone.

让我们深入探讨隐私与安全问题。当你邀请这些机器人进入家庭时，有些活动可能不愿被外界知晓。比如夜间你入睡后机器人执行任务，总不希望清晨发现保险柜被打开还到处宣扬吧？或者当它照顾年迈母亲时，不该因为老人要求就每晚给她灌威士忌。

Let's go into the conversation of privacy and safety. So you're inviting these robots into your home where there will be activities that you may not want shared with the world. And then, of course, you're asleep and the robot is running tasks at night. You don't want to wake up in the morning and find your safe has been opened and the robot's gone to talk about. Or you don't want the robot to be taking care of your aging mother and find out that it's you know, given her shots of scotch at night when she asked for them.

那么你们如何保障安全与隐私？

I mean, so how do you deal with safety and privacy?

顺便说一句，最后一条是最难的。我们可以稍后再讨论。好吧，别给奶奶喝苏格兰威士忌。

That last one is the hardest one, by the way. We can get back to that. Okay. Not to give grandma scotch.

苏格兰威士忌的shots。

Shots of scotch.

因为通常来说，模型总是被调校得有点谄媚，它们最终会做我们让它们做的任何事。所以如果从隐私方面开始，我认为首先最重要的是透明度。如果你是第一批像你这样的人，彼得，家里会有一个NeoGamma，

Because generally, models are they're always kind of tuned to be kind of sycophants, they end up doing whatever we ask them to do. So if we start with the privacy side, I think first of all, it's a just lot about transparency. If you're one of the first people like you, Peter, that will have a NeoGamma in your house,

我们

we

某种程度上是在用隐私换取成为早期使用者，因为没有数据，我们就无法改进产品。当然，我们会尽一切努力确保这是在你的条款下的隐私，并且由你掌控。但如果我们想改进产品，确实需要你的数据。当然，我是说，

are kind of trading a bit on privacy versus being an early adopter, because without the data, we can't make the product better. Of course, we're going to do everything we can to make sure this is privacy on your terms and that you are in control. But we do need your data if we're going to make the product better. Sure. I mean,

我一直把我的数据给谷歌、亚马逊、X。我是说，人们没有意识到你在家里和配偶讨论时，亚马逊的Alexa在听，对吧？Siri也在听。

I give my data to Google, to Amazon, to X all the time. And I mean, people don't realize that you're sitting at home having a discussion with your spouse and Amazon's, Alexa's listening, right? Siri's listening.

但它们在做一个非常重要的事情，我们也在做，那就是我们公司里没有人能听到或看到那些数据。

But they're doing something very important, which we also do, which is no human in our company can hear or see that data.

是的。

Yes.

那些数据会进入训练模型。是的。但它不会经过人类。对吧。现在，如果我们想查看那些数据，有时候你可能需要，对吧？

That is going into the training model. Yes. But it doesn't go by a human. Right. Now, if we want to look at that data, and sometimes you might need to, right?

可能会像是，让我们弄清楚这里发生了什么，因为显然有一些事情在多个机器人之间发生，我们想弄清楚是什么，然后我们会在你的手机上发送一个通知，说，嘿，这个特定的窗口，我们想审查数据，你会收到一个视频，展示那些数据是什么。然后如果你说可以，我们就能拿到解密密钥查看数据。如果你说不，那我们就不能。所以，你是掌控者。实际上，即使是进入训练数据，我们也总是有24小时的延迟。

Might be like, let's figure out what happens here, because something clearly is happening across multiple robots that we want to figure out what is, then we'll send you a notification on your phone where we say like, hey, this specific window, we want to review the data, And you'll get a video of like what that data is. And then if you say yes, then we get the decryption key and we can look at the data. If we don't, if you say no, then we can't. So that, you're in the control of that. And actually, even with respect to going into the training data, we always run like a 24 delay on training.

所以如果有些内容你确实不希望存在，哪怕是训练数据中的，比如这件事从未发生过，要从存在中彻底抹去，你可以在它进入训练权重前手动删除。

So if there is something that you really don't want, even in the training data, like this never happened, erase it from existence, You can go in and delete it before it gets into the training weights.

我想让大家都明白，有政策和方案使这种做法被科技公司接受并使用，而你们将实施其中最优的方案。

I just want everybody to hear, there are policies and plans that make this acceptable and used by technology companies and you're going be implementing the best of those.

这挺酷的。

That's pretty cool.

但所有这些——我刚才提到的模式是当机器人处于我们所谓的'尽力自主'状态时，对吧？这是大多数情况。就像你们今天早些时候看到的，如果你和它说话，让它做某事，希望它能做对。如果没做对，你可以说'坏机器人'，希望下次它能改进。但这实际上是在现实生活中的学习过程。

All of But of there is like so the mode I talked about now is when the robot is what we call best effort autonomy, right? Which is most of the time. It's what you saw earlier today, where if you talk to it, you ask it to do something, hopefully it does the right thing. If it doesn't do the right thing, then you can say bad robot, and hopefully it's better next time. But it's actually this is learning in real life.

这真的像是互动式学习。有趣的是，机器人从失败中学习任务的速度其实比成功时更快。当然，就像我们一样，它从失败中学到更多。但在这个模式下，这就是隐私保护。

This is really like interactive learning. And the robot, interesting enough, actually progresses faster on tasks when it fails than when it succeeds. Sure. It learns more from failures, just as we do. But in this mode, that's the privacy.

现在说到远程操控，显然不看杯子就无法完成这个任务。所以我们做了一些抽象处理，让你实际上看不到具体的人，只能看到模糊影像和你正在交互的对象，我们能在过滤端做很多工作来确保隐私。但最重要的是，除非你批准，否则没人能接入你的机器人，对吧？这在机器人上非常明显，比如灯光会变化，就像有人进入了你的机器人，而且必须是你预先批准的操作员之一——比如这里有四位服务你的操作员。这有点像邀请清洁工之类的人进入你家。

Now, it comes to Teleop, then of course, there's no way you can do this task without seeing the glass. So we do some abstractions so that you actually don't see people people, you kind of like just see blobs and like you just see the object you're interacting with, and we can do a lot on the filtering side to ensure privacy. But the most important thing we do here is that no one goes into tally up in your robot unless you approve it, right? And it's very visible on the robot, like the lighting changes, and it's like someone is in your robot, and it's one of the preselected operators that you have approved for a large set of operators, like here are the four that services you. So that's kind of like inviting your cleaner whatever into your house.

另一个人。

Another human.

另一个人进入你家。你只需要确保他们确实是受邀的。

Another human into your house. And you just need to make sure that they're actually invited.

让我们花点时间详细说明这点。早期当NeoGamma在我家时，它基本是自主运行的，但有时需要引入远程操作员。因此总部会有远程操作员，当它需要帮助、处理复杂事务或出错时，操作员可以介入确保任务完成。

So to actually take a second and spell this out in more detail, in the early days when I have NeoGamma in my home, it'll be baseline autonomous, but there will be times where it needs to bring in teleoperator. And so you'll have teleoperators in headquarters that if it needs help or doing something complicated or it gets something wrong, the teleoperator can step in and actually make the task happen.

是的，初期其实有两种模式。一种是刚才讨论的'尽力自主'模式，另一种是任务调度模式——就像我现在在家里的角色。我用手机安排说'在这段时间内，我要你为我完成这些任务'。

Yeah, it's in the beginning, it's actually there's two different modes. So you have the mode, which I call like the best effort autonomy that we just talked about. Yeah. And then you have task scheduling, which like my role at home now is doing that. So I take my phone and I schedule and say like, hey, between these hours, here are the tasks I want you to do for me.

今天就像要洗我的白色衣物。然后会有个Instacart的包裹送到。你可以在门口接收后放进冰箱。基本上就是保持整洁。我不在家时也这样安排。

Today is like do my white laundry. And then there's a package coming from Instacart. You can receive it at the door and then back it in the fridge. And it's just like generally tidy. And I've given it when I'm not home.

比如这些工作时间，我在上班。只要把事情完成就行，对吧？现在我不在乎是自主完成还是远程操作员来处理。明白吗？

Like these hours, I'm at work. Just get it done. Right? Now, I don't care if that happens autonomously or to a teleoperator. Right?

所以很多情况会交给远程操作员，因为有些任务相当复杂，我们目前还无法很好地实现自动化。

So a lot of that happens to a teleoperator because some of these tasks are quite complicated and we don't know how to automate them well enough yet.

现在，

Now,

当然，那个远程操作员会利用自主性来提高效率。所以并不全是远程操作，但我其实不关心具体比例。任务完成了就行。是的。我们大致这样划分，然后中间还有片灰色地带。

of course, that teleoperator uses autonomy to help improve the efficiency. So it's not all teleoperation, but I don't really care about the mix. The task gets done. Yes. So we kind of split it like that, and then there's the gray zone kind of in the middle.

如果就像你想的那样，比如说，你想邀请朋友来开派对，希望机器人当调酒师，而我们目前还没有调酒师模式。这时你可以授权远程操作员来完成。所以你可以

If it's like you want to, I don't know, having your friends over for a party and you want the robot to be the bartender, and we don't have like a bartender mode yet. Then you can approve a teleoperator to do that. So you can

我们在特斯拉餐厅或他们活动中看到的所有Optimus视频，都是远程操作演示点。

All the videos we see of Optimus at Tesla's diner or at their events are teleoperation points.

确实如此。但我认为远程操作被不公平地贴上了某种负面标签。为什么呢？我想是因为人们不够清楚，比如，这是帕洛阿尔托吗，它是自主运行的吗？但其实这只是标注数据。

They are. But I think teleoperation has gotten this kind of like underserved, bad reputation or name. Why? I think this is because people don't have enough clarity in like, hey, is this Palo Alto, is it autonomous? But it is just labeled data.

这是专家示范。如果你观察任何经过训练的大型AI模型，都有大量人员坐下来手工标注数据、查看案例、梳理问答，从而构建出这个高质量的数据集来使其运作，对吧？所以你在通用信息上进行预训练，我们也这么做，只是机器人领域发生的一切，最终都会形成一个经过精细调整的、非常高质量的数据集。而在机器人学中，这就是远程操作，因为这就是专家示范。这就是手工标注的数据。

It's expert demonstrations. If you look at any of the big AI models that were trained, there was an enormous amount of people that sat down and hand labeled data and looked at examples, rolled out question answers, and bootstrapped this very high quality data set for this to work, right? So you pre chain on general information, we also do that, just everything that's happened with the robot, and you have a fine tuned data set that is very high quality. And in robotics, that is teleoperation because it's the expert demonstrations. It's the hand labeled data.

本质上没有区别。只是我认为目前对实际情况缺乏透明度。

It's not different. It's just I think there's some lack of transparency in what's going on.

嗯，我认为反对意见在于，如果你有演示视频，让它看起来好像能完成某项功能，但实际上它做不到，因为那是你手动编码实现的。好吧，我

Well, I think the objection is if you have demo, like a video, it makes it look like it can do something and it actually can't because you hand coded it. Well, I

认为它显然可以做到，只是无法自主完成。

think it clearly can, but it can't do it autonomously.

对，确实做不到。但我觉得你说得很对——如果它在物理上能做到，如果机械结构支持这个动作，神经网络无论如何都会立即填补那个盲区的，你知道，一旦训练完成。所以我认为这完全合理。

Yeah, can't do it. Yeah, but I think you're dead on that if it can physically do that, if the mechanism can do that, the neural net will fill in that blind spot instantly anyway, you know, once you've trained it. So I think that it's perfectly legit.

现在进入可能是最重要的环节——登月计划中的健康科技板块。大约十年前，我一位身体极好的挚友因侧腹疼痛去医院，结果查出已是癌症四期。几年后，我的大学兄弟会成员在睡梦中因心脏病发作去世，他还那么年轻。

And now it's time for probably the most important segment, the health tech segment of moonshots. It was about a decade ago where a dear friend of mine who has incredible health goes to the hospital with a pain in his side only to find out he's got stage four cancer. A few years later, fraternity brother of mine dies in his sleep. He was young. He dies in his sleep from a heart attack.

那时我才意识到，除非主动检查，人们根本不知道体内正在发生什么。我们都对自己的健康盲目乐观。但你知道吗？70%的心脏病发作毫无征兆——没有呼吸急促，没有疼痛。大多数癌症被发现时都已太晚，处于三期或四期。

And that's when I realized people truly have no idea what's going on inside their bodies unless they look. We're all optimists about our health. But did you know that seventy percent of heart attacks happen without any preceeding? No shortness of breath, no pain. Most cancers are detected way too late at stage three or stage four.

可悲的事实是，我们已拥有规模化检测和预防这些疾病所需的所有技术。那一刻我意识到必须采取行动。我认为每个人都应有机会在疾病恶化前使用这项技术进行筛查和预防。于是我与一群杰出的企业家朋友——托尼·罗宾斯、鲍勃·霍里、比尔·卡普合作，整合关键技术和顶尖医师科学家团队，创立了名为Fountain Life的机构。每年我都会去那里进行数字体检，四小时内采集从头到脚200GB的身体数据，全面掌握健康状况。

The sad fact is that we have all the technology we need to detect and prevent these diseases at scale. And that's when I knew I had to do something. I figured everyone should have access to this tech to find and prevent disease before it's too late. So I partnered with a group of incredible entrepreneurs and friends, Tony Robbins, Bob Houri, Bill Cap, to pull together all the key tech and the best physicians and scientists to start something called Fountain Life. Annually, I go to Fountain Life to get a digital upload, 200 gigabytes of data about my body, head to toe, collected in four hours to understand what's going on.

所有数据都会输入我们的AI系统和医疗团队。这对我而言是每年不可妥协的必修课。我唯一的请求就是：请成为自身健康的CEO。要明白身体多么擅长隐藏疾病，并真正了解体内发生的变化。

All that data is fed to our AIs, our medical team. Every year it's a non negotiable for me. I have nothing to ask of you other than please become the CEO of your own health. Understand how good your body is at hiding disease. And have an understanding of what's going on.

您可以访问fountainlife.com与我的团队成员沟通，网址是fountainlife.com。现在我想切换到另一个有趣的话题——恐怖谷效应

You can go to fountainlife.com to talk to one of my team members there. That's fountainlife.com. I wanna jump into another fun subject, which is the uncanny valley

嗯。

Mhmm.

以及面部模拟。你们团队内部肯定反复讨论过：如何塑造面部外观？拟真度要达到什么程度？皮肤质感如何呈现？最终以什么形式表现？

And the face. So talk to, I mean, you've probably had endless conversations internally about how do you make a face look? How human do you make it? How skin like do you make it? How do you represent it?

能否从哲学角度谈谈你和Dar（你们设计团队的成员）是如何思考这个问题的？你们在哪些方面让它足够人性化？

Can you tell us sort of philosophically what you and Dar, who on your design team, how do you think about that? Where do you make it human enough?

这是一条非常微妙的界限——你需要确保肢体语言能清晰传达，因为这是设备（或者说伴侣）的魔力所在。但同时，你又不希望它触发你的本能警觉，觉得‘这个人类有点不对劲’。你并不想把它做成人类。

Is this very delicate line where you want to make sure like body language comes across crystal clear, because that's like the magic of the device, Of the companion. Yes. But at the same time, you don't want it to get like to where your kind of instincts tell you, hey, something's wrong. This is a human, but there's something wrong with it. You don't want it to be a human.

令人惊讶的是，确实存在这样一个区间：人们明确认同‘这是个我能理解的生物，我懂它的肢体语言’，但它又显然不是人类。我们就是要立足这个区间，具体位置则因人而异——不同人的接受阈值不同。所以我们力求折中，确保对大多数人来说，这既是极易理解的产品，又不会令人不适。

And it's actually pretty surprising that there is this gap where people clearly identify this as like, hey, this is a being I identify with, I understand his body language and everything, but it's also clearly not a human. And you want to be in that space. And then where you are in that space kind of depends a bit on who you ask. People have a different threshold here. So we're trying to hit in the middle of that and ensure that for as many people as possible, this is just an incredibly easy to understand product, but at the same time that it's not creepy.

顺便说下规模化问题，我认为产品采用率至关重要。新技术的普及通常需要时间，因为存在认知壁垒。就像使用手机也存在入门门槛。

And I think adoption here, by the way, talking about scale, adoption is so important, and adoption of new technology usually takes some time because there's just like this knowledge barrier, right? There's a barrier to entry. Even using a phone, there's a barrier to entry.

这个

This

交互方式极其自然，完全零门槛。就像和...

interface is just so natural. Like there is no barrier to entry. It's something you just talk to like a

最让我震撼的是，家里有大约50件我不会处理的事，比如给泳池反冲洗这种破事。而机器人能实时获取信息，学会操作流程

What's incredibly cool to me is that there are like 50 things around the house that I don't know how to do, including the fricking, way to backwash the pool, like all this crap. The robot can, in real time, access the information, learn how

并直接执行。是的，从工作层面来说。

to do it and just do it. Yeah, from the work.

我做不到。我得研究一小时。而且没有工人会愿意以低于400美元的价格上门处理。这类事情多到数不清——我不是要取代人类，而是在填补那些因知识冷门而原本无解的空白。

I can't do that. It would take me an hour to study. And there's no laborer that's gonna come into the house and do it for under like $400. And so it's like, like, are so many things that are in that category where I'm not trying to replace a human being. I'm doing something that there literally was no other option for because the knowledge is obscure.

现在家庭中这类需求太多了。比如热水器老是跳闸需要重置流程。你可以查资料，机器人也能查。

And there's so many of those things around a house now. Like resetting the water heater keeps going out and the reset process. But you can look it up. The robot can look it up.