摘要：我认为人形机器人将成为有史以来最大的产品。需求将是永不满足的。每个人都会想要一个。这就像是，谁会不想要自己的个人版 C-3PO 或 R2-D2 呢？
马斯克：
我认为人形机器人将成为有史以来最大的产品。需求将是永不满足的。每个人都会想要一个。这就像是，谁会不想要自己的个人版 C-3PO 或 R2-D2 呢？
CNBC 主持人
你之前提到过，你的目标是在 2030 年前生产 100 万台机器人，我记得这是你公开说过的计划。
马斯克
我认为这是一个合理的目标。
CNBC 主持人
然后迈向你所说的“可持续的丰裕”（sustainable abundance）。不过，我很好奇，我们一直在讨论自动驾驶技术需要多长时间才能达到或超越人类能力。但针对这些机器人，它们需要多少训练才能执行不同类型的任务？这会需要很长时间吗？
马斯克
这需要大量计算资源和时间。但有些关键的突破我们认为可以实现——比如，如果Optimus（特斯拉机器人）能通过观看 YouTube 或其他视频来学习，就像人类通过观察学习一样，那么它的任务扩展性将会有巨大飞跃，因为它能快速学会任何事情。我们正在朝这个方向努力，但目前还没完全实现。
CNBC 主持人
也就是说，你需要大幅提升机器人的学习和训练能力。
马斯克
是的，关键阈值在于机器人能否通过观看视频学习。但目前我们还在训练 Optimus 完成一些基础任务，比如让穿着动作捕捉服的人类示范如何拾取物体、开门、扔球、跳舞等。这些训练是必要的，它们能为机器人的基础功能“启动智能”。而真正有趣的是，当机器人能“自我训练”时——就像孩子通过玩玩具学习一样。比如，孩子反复尝试把积木放进对应形状的孔里，机器人也可以通过类似的自我试错来学习，只需设定一个奖励机制（比如“成功放入正确位置”）。
CNBC 主持人
实现这些需要 AI 或计算能力的突破吗？还是现有技术已经足够？
马斯克
需要一些突破，但我认为并非不可逾越。未来几年内我们就能解决这些问题。
CNBC 主持人
所以当这些机器人量产时，它们会像你描述的汽车一样完全自主？比如，我可以直接说“去洗碗”或“遛狗”，甚至不用开口，它们就能预测我的需求？
马斯克
没错。你甚至不需要开口，它们会推测你的需求并主动执行。
CNBC 主持人
这需要多少 GPU 支持？
马斯克
需要相当多的 GPU。我们有自己的训练项目 Dojo，但它目前仅贡献了约 5%的自动驾驶训练算力（纽约的数据中心在运行）。我们仍会从英伟达、AMD 等公司大量采购
GPU——只要它们的性能优于我们自研的硬件。
英语原文
Musk
I think humanoid robots will be the biggest product ever. The demand will be insatiable. Everyone is gonna want one. It's like basically, who wouldn't want their
own personal C-3PO、R2-D2.
CNBC 主持人
I, you've said that and you know, your goal, you've also said is to produce a million robots, I think by 2030. That's what I had you on the record as saying.
Musk
I think that's a reasonable target.
CNBC 主持人
And then start towards sustainable abundance, which you can get into. But yeah, you know, I wonder, we've been talking about autonomous and how long it takes to train
the automobile to be able to be the equivalent of or exceed human capabilities. But what about these robots to the extent that, how much training are they gonna need to actually, to be able to do various different types of tasks? Is that gonna to take a long time?
Musk
So it's gonna take a lot of compute resources and it'll take time. I think there certain threshold breakthroughs that we think we can achieve where if Optimus can watch videos, you know, YouTube videos or whatever. And based on that video, just like a human can learn how to do that thing, then you really have task extensibility that is dramatic cuz then it can learn anything very quickly. So I, I think we'll get there and then we're. Not there yet, though.
CNBC 主持人
You're relying on a significant uptick in terms of learning and training.
Musk
So I'm calling it a very significant threshold would be understand the ability to learn from watching a video just so right. But at the.
Yeah, I mean, right now we're really, we're training optimists to like primitive tasks where human in a kind of a, what's called a mocap suit is up and sort of cameras on the head is moving in the way that the robot would move to, say, pick up an object or
open a door or the basic tasks, throw a ball, dance. And we need to that. I think that's needed to sort of bootstrap the intelligence so you can have the basic functions.
Then where I think it gets very interesting and very much like humans is that you want the robot to self play. So you say, how does a child learn? Well, a child has toys.
A child plays with the toys, plays with the blocks. You know, at some point figures how to put the triangle in the triangle hole in the circle hole by doing it over and over again. And this, the self play. Once you have a lot of robots, you can do this self play,
which is that you just put the robot in a room with toys and have the robot play, literally have the robot play with toys.
And you have to have a reward function. Say like, okay, your, the goal of the robot is to put the circle in the circle hole, the square and square, all triangle and triangle hole, and keep doing it until it works and the reward function is succeeding.
CNBC 主持人
And there are no advances needed to accomplish that now. I mean, no advances in AI or just compute and things of that nature that can happen.
Musk
There are some advances needed, but I don't think these are insurmountable. I think
we, I think we can we can I we can solve these things and right the next few years.
CNBC 主持人
Okay. So at that time when millions of these things are coming off of a line, like we
just, you know, saw with cars, they're gonna be fully autonomous, they're gonna
come off and I'm gonna, and there, I mean, the way you've described it, come in my
house and I'm gonna be able to say, alright, do the dishes. Now I need you to walk
the dog. Absolutely. Hold the baby.
Musk
In fact, you really won't even need to. It'll, it'll figure out what you probably want and
do what you probably want without even having to ask.
CNBC 主持人
How many GPUs are you gonna need for that?
Musk
Well, quite a few GPUs. We do have our own program called DoJo for training, which
I think will be helpful.
CNBC 主持人
Right now it's contributing about 5% towards self driving, right? So that's the one in
New York?
Musk
Yeah, yeah, yes, in New York. So, but we expect to still buy a lot of GPUs from Nvidia,so from AMD and maybe from others. And as long as a Nvidia is better than what we make, we'll keep buying from Nvidia.
来源：全产业链研究一点号
