全球每年生产大约一亿辆汽车,路上跑着的存量大约二十亿辆。马斯克被问到人形机器人量产会做到什么规模,他先把这组汽车的数字摆出来当参照,然后给出自己的猜测:人形机器人的用处大得多,所以他猜是每年十亿台以上。
一亿到十亿,差了一个数量级。而特斯拉现在的人形机器人的产量是零。
这段对话出自莱克斯·弗里德曼(Lex Fridman)的播客,2024 年。马斯克在里头聊了算力、数据、手的解剖结构、量产曲线,看上去是四个不相干的话题,其实是同一条链子上的四节。
算力是排量,能买到的都不算壁垒
弗里德曼先搭了个框架。训练算力相当于发动机的排量,马力越大越好,你得先在这一项上做到最好。然后看你怎么高效使用这些算力,推理的时候怎么算得省。这两件事最后都落到人身上。再往后,是你有没有别人拿不到的数据。
马斯克没有推翻这个排序,但他的落点全在最后一项。
这个排序里,前几项都有明码标价。算力可以下单买卡,人才可以加薪挖走,效率靠组织能力慢慢磨。数据入口没办法用钱在短时间内买到,尤其是那种独家、持续、别人复制不了的入口。
所以他接下来花了大段时间讲数据。故事的起点是一个让人有点扫兴的事实。
推特的存量数据,早就被爬干净了
弗里德曼问,推特的数据对 xAI 有用吗。
马斯克的回答先往下压:领先的 AI 公司早就把推特的数据爬完了。"不是我觉得,是他们已经做了。"
这句话等于承认,公开语料的争夺战已经结束。存量没有意义,谁都能拿到的东西不算优势。你花力气去清洗一亿条旧推文,别人手里也有同一批。
有价值的是增量,而且是秒级的增量。
往前看,有用的是它精确到秒。这个别人很难实时抓取。所以 Grok 已经有一个即时性的优势。(马斯克)
通用爬虫有周期,快的话几小时,慢的话几天。推特上发生的事情,几分钟就会形成讨论。突发事件、舆论转向、某个产品的第一波真实反馈,全都长在这个时间窗里。模型的语料如果以天为单位更新,它就永远在回答昨天的问题。
这是马斯克讲数据的第一层意思:数据的新鲜度本身可以变成护城河。不是数据的量,是数据的延迟。
现实会规模化
第二层意思要大得多。
特斯拉有几百万辆车在路上跑,每辆车都带着摄像头。这个数字未来会涨到几千万辆。擎天柱(Optimus)机器人如果做起来,可能是几亿台,甚至几十亿台,它们全都在从真实世界里学东西。
马斯克说,这会是最大的数据来源。不是推特,不是网页,是机器人。
因为现实会按照现实的尺度扩展。(马斯克)
这句话听起来像同义反复,但它说的是一个量的对比。弗里德曼接了一句,人类的文字数据其实没多少。马斯克顺着这个往下讲:
看看人类到底生成了多少万亿个可用的 token(词元),把垃圾、重复的东西都去掉之后,这个数字并不大。你很快就会发现不够用。(马斯克)
"万亿"这个量级听着吓人,但模型训练的实情是,高质量的、不重复的、有信息量的文本比想象中少得多。模板文案、转载稿、SEO 农场、互相抄的问答,全都占位置但几乎不提供新信息。业内这两年反复讨论的"数据墙",说的就是这件事:人类写下来的东西,快要被读完了。
马斯克用了一个词来形容这种感觉,谦卑(humbling)。人类的全部书写遗产,喂不饱下一代模型。
而物理世界没有这个问题。现实不会重复,每一次抓取、每一次移动、每一次碰撞都是新的样本。
车只能待在路上,机器人哪儿都能去
特斯拉车队的数据再多,也有个天花板:车在道路上。道路是一套高度规范化的场景,车道线、红绿灯、交通规则,全都被设计成可预测的样子。
马斯克说得很直接:擎天柱可以去任何地方,而路以外的现实更多。
这个差别决定了数据的天花板在哪儿。自动驾驶解决的是"如何在既定规则下安全移动",人形机器人解决的是"如何在任何环境里完成任何动作"。前者是收敛的,后者是发散的。
发散到什么程度,马斯克举了个特别朴素的例子。拿杯子,看它有没有用正确的方式抓起来。往杯子里倒水,看水有没有倒进去。洒了没有。就这么点事。
弗里德曼追问,倒水这件事为什么值得反复验证。他的意思是,容器是千变万化的,你不知道面前这个杯子是什么材质、多厚、多滑、装了多少水。人类在这些变量面前毫不费力,是因为我们从小摔了无数个杯子。
就是这么简单的事。但它可以乘以十亿的规模,从现实里生成有用的数据。因果性的数据。(马斯克)
"因果性的数据"是这段话里最关键的四个字。文本语料告诉模型"人类通常这样描述倒水",传感器数据告诉模型"这样倾斜,水会洒;那样倾斜,水会流进去"。前者是描述,后者是因果。机器人是少数能自己动手做实验、自己看结果、自己修正的 AI 载体。
十亿台这个数字是怎么算出来的
弗里德曼问,人形机器人要做到量产,需要什么。
马斯克的回答是:跟车一样。
全球汽车产能大约一亿辆一年,需求也差不多是一亿辆一年。存量约二十亿辆,车的寿命大约二十年,一亿乘以二十约等于二十亿,正好是稳态。这套算术他讲得很顺,因为汽车工业已经跑了上百年,所有变量都被磨平了。
人形机器人套同样的模型,但换一个数字。因为用处大得多。
所以我猜人形机器人更接近每年十亿台以上。(马斯克)
十亿台一年,意味着什么?意味着这个产品的产能要比全球汽车工业大十倍。意味着供应链、产线、材料、装配、测试,全部要重新长出来一遍,规模还要更大。
他没有给出时间表,只给了数量级。数量级本身已经足够说明他怎么看这件事的终局:人形机器人不是汽车的一个品类,是一个比汽车大一个数量级的品类。
但紧接着,弗里德曼提了一个反差。在 Optimus 出现之前,人形机器人被公认为极难的问题。马斯克接了一句双关:
是啊,散步一样轻松。我是说,Optimus 现在连在公园里散步都有点费劲。(马斯克)
"walk in the park"在英语里既是"在公园散步",也是"轻松的事"。他一边否认这件事简单,一边承认机器人现在的水平。能走,公园这种地形不算太难,但要覆盖各种地形还早。抓取已经能做了,通用物体还不行。
一半的工程量在一只手上
整段对话里最实在的一段,是讲手。
弗里德曼说了一句挺有意思的话:人类的智能,很大一部分花在了手对世界的操作上。他还给了一个定义,智能是在世界里安全地操作物体。
马斯克顺着这个往下讲,讲出了一个具体的比例。
光是手,就可能占掉 Optimus 全部工程量的一半。从机电的角度看,手大概占一半。(马斯克)
一半。不是 AI 算法,不是电池,不是行走,是一双手。这个比例对外行来说很难相信,对做过机器人的人来说一点都不奇怪。
他还提了一个神经解剖学上的常识:大脑皮层里负责手的那块区域,面积大得不成比例。彭菲尔德(Wilder Penfield)当年在清醒开颅手术中用电流刺激病人大脑,画出了身体各部位在皮层上的映射图,那图上的手和嘴唇占了巨大面积,躯干和四肢反而很小。人脑给手分配的算力,和手在身体上的体积完全不成比例。
换句话说,人类智能里有一大块是专门为操纵物体长出来的。想让机器人做人的事,就得把这一块补上。
你的手,是前臂牵着的一具骨架木偶
马斯克讲了一件大部分人从来没想过的事:你手指的肌肉,长在前臂上。
手掌本身几乎没有肌肉,只有几条很小的。真正驱动手指的,是前臂里那一大团肌肉,它们通过长长的肌腱,穿过手腕上一个很窄的通道,牵动手指。
你的手其实就像一具骨架做的肉傀儡,靠缆绳驱动。(马斯克)
那条通道叫腕管(carpal tunnel),由一小圈腕骨和韧带围成,里面挤着九条屈肌腱和一条正中神经。通道极窄,所以一旦压迫到神经,就是腕管综合征。打字打多了手麻,根子在这儿。
这套结构是演化出来的工程方案:把沉重的驱动部件放在靠近身体的地方,用轻的缆绳去远端执行。好处是末端轻、灵活、惯量小。
Optimus 的第一版没这么干。特斯拉试过把执行器直接塞进手掌里。
结果就是那种特别大的手,看起来很怪。而且它们的自由度不够,力量也不够。然后你才明白,哦,原来执行器得放在前臂。(马斯克)
这是硬件研发里最典型的教训:你以为可以优化掉的东西,是几十亿年演化试出来的最优解。手掌装不下足够的电机,装得下也不够力,够力了手就变成一只巨大的钳子。
所以新版的手臂改了。22 个自由度,之前是 11 个,直接翻倍。执行器全部挪到前臂,像人一样,用缆绳穿过一条窄通道去驱动手指。
更狠的是,这些执行器和传感器全部从头设计,按物理第一性原理推,不是拿现成货改。
所有的执行器都是从物理第一性原理重新设计的。传感器也是全部重新设计。(马斯克)
这句话的分量在"全部"两个字。市面上买不到能塞进机器人前臂、又满足力量、精度、散热、成本四个约束的执行器。想做人形机器人,就得自己造电机、自己造减速器、自己造反馈传感器。这也解释了为什么人形机器人公司的护城河不在软件,在执行器和传感器的自研能力上。
小指为什么不能和大拇指一样长
对话里最有趣的一段,是马斯克自己抛出来的一个问题:为什么手指长度不一样。
他的推理是这样的。从能量或者演化的角度看,让五根手指一样长并不更贵。既然不贵,为什么没长成一样?
因为不一样长更好。手指长度不同,你的灵巧度更高,能做的事情更多。(马斯克)
他专门点了小指。小指看着没什么用,但少了它,灵巧度会明显下降。精细动作的稳定性,很大程度上靠小指在掌侧提供的那个支点。
这段对话看起来是闲笔,其实是在讲一个设计哲学:生物形态里那些看起来随意的细节,往往是功能约束解出来的最优解。做机器人做到深处,会不断撞上这类细节。手掌的弧度、手指的锥度、关节的阻尼、皮肤的摩擦系数,每一样都影响能不能拧开一个瓶盖。
"尽可能简单"是个很高的门槛
弗里德曼问,那你是不是得同时想清楚怎么量产,所以要做得尽可能简单。
马斯克的回答有点扫兴,但很诚实。
实际上会相当复杂。"尽可能简单"这个要求本身门槛就很高。你想要一个人形机器人做人能做的事,这个标准非常高。(马斯克)
他说,一个"能完成人类大部分工作"的机器人,哪怕砍掉一部分能力,剩下的部分依然极其复杂。不简单。非常难。
这段话值得停在原地想一想。过去几年人形机器人的演示视频满天飞,走路的、后空翻的、搬箱子的。演示和能力之间隔着一条巨大的鸿沟:演示是在受控环境里重复一个动作,能力是在陌生环境里第一次就把事情做对。
而量产的难度又叠在能力上面。手占一半工程量,手臂 22 个自由度,执行器和传感器全部自研,还要做到每年十亿台的产能。这四件事同时成立,才算把这件事做成。
马斯克在整段对话里没有给时间表,没有给价格,没有说"明年就能买到"。他给的是一个规模判断、一个技术瓶颈的位置、一个解剖学上的类比。这三样东西凑在一起,比任何发布会上的日期都更能说明他心里的难度排序。
把这段对话拆成能用的判断
如果把马斯克这段话当成投资或者职业判断的输入,能落地的部分大概是这些:
- 看 AI 公司时,先看它的数据是不是独家的、是不是实时的。公开语料已经被瓜分完,能形成长期差异的是别人拿不到的时间窗,或者别人造不出来的采集硬件。一个模型说自己训练了多少万亿 token,这个数字的边际价值在快速下降。
- 机器人赛道里,把注意力从"大脑"挪到"手"上。手占整机工程量的一半,而这个位置上的供应商极度稀缺。执行器、减速器、腱绳传动、触觉传感器、微型编码器,这些细分环节的技术门槛比整机公司更硬,被替代的可能性更低。
- 自研能力是硬指标。判断一家人形机器人公司,问一句执行器是不是自己设计的,传感器是不是自己做的。答不上来或者答"用供应商方案",量产阶段会卡死。市面上的货满足不了前臂安装的体积和功率密度要求。
- 数据采集要按"简单动作乘以大规模"来设计。不用一上来就追求复杂任务。抓取、放置、倾倒、擦拭,这类动作的因果信号清晰,可以自动化标注,可以用遥操作批量采集,可以靠规模滚起来。复杂任务是这些原子动作的组合。
- 评估一台人形机器人,先数自由度和看执行器位置。手臂 11 个自由度做不了人的大部分精细活,22 个是及格线。执行器如果堆在手掌里,说明散热和力量问题没解决,大概率还停留在演示阶段。
- 如果你是做具身智能的,把数据管道的实时性当成产品需求来设计。模型能不能拿到五分钟前的真实世界反馈,决定了它在开放环境里能不能自我修正。离线数据集训练出来的策略,在现场几乎一定要重新标定。
0:02So the training compute is kinda like the engine, how many is this horsepower of the engine. So really, you wanna try to do the best on that. Then how efficiently do you use that training compute? And how efficiently do you do the inference, the use of the AI? So obviously, that comes down to human talent.
0:24And then what unique access to data do you have? That also plays a role. - You think Twitter data will be useful? - Yeah, I mean, I think, I think most of the leading AI companies have already scraped all the Twitter data. Not I think they have. So on a go forward basis, what's useful is the fact
0:47that it's up to the second. That's hard for them to scrape in real time. So there's an immediacy advantage that Grok has already. I think with Tesla and the real time video coming from several million cars, ultimately, tens of millions of cars, with Optimus, there might be hundreds of millions of Optimus robots, maybe billions learning a tremendous amount from the real world.
1:15That's the biggest source of data I think ultimately is sort of Optimus. Optimus is gonna be the biggest source of data. - Because- - 'Cause reality scales. Reality scales to the scale of reality. It's actually humbling to see how little data humans have actually been able to accumulate.
1:36Really, you see how many trillions of usable tokens have humans generated, where on a non-duplicative, like discounting spam and repetitive stuff, it's not a huge number. You run out pretty quickly. - And Optimus can go, so Tesla cars can unfortunately have to stay on the road. Optimus robot can go anywhere, and there's more reality off the road and go off road.
2:07- I mean, except for the store, where I can like pick up the cup and see, did it pick up the cup in the right way? Did it pour water in the cup? Did the water go in the cup or not go in the cup? Did it spill water or not? - [Lex] Yeah. - Simple stuff like that. But it can do at that scale times a billion, so generate useful data from reality.
2:33So cause and effect stuff. - What do you think it takes to get to mass production of humanoid robots like that? - It's the same as cars, really. I mean, global capacity for vehicles is about a hundred million a year. And it could be higher. It's just that the demand is on the order of a hundred million a year.
2:56And then there's roughly two billion vehicles that are in use in some way, which makes sense. Like the life of a vehicle is about 20 years, so it's steady state. You can have a hundred million vehicles produced a year with a two billion vehicle fleet roughly. Now for humanoid robots, the utility is much greater.
3:15So my guess is humanoid robots are more like at a billion plus per year. - But until you came along and started building Optimus, it was thought to be an extremely difficult problem. I mean, it still- - Well, it is. - Extremely difficult. - So walk in the park. I mean, Optimus currently would struggle to walk in the park.
3:37I mean, it can walk in a park. The park is not too difficult, but it will be able to walk over a wide range of terrain. - And pick up objects. - Yeah, yeah. It can already do that. - [Lex] But like all kinds of objects? - Yeah, yeah. - All foreign objects. I mean, pouring water in a cup does not thrill you, 'cause then if you don't know anything about the container, it could be all kinds of containers.
4:00- Yeah, there's gonna be an immense amount of engineering just going into the hand. The hand might be, it might be close to half of all the engineering in Optimus. From an electromechanical standpoint, the hand is probably roughly half of the engineering. - But so much of the intelligence, so much the intelligence of humans goes into what we do with our hands.
4:22- Yeah. - It's the manipulation of the world, manipulation of objects in the world. Intelligence is safe manipulation of objects in the world, yeah. - Yeah. I mean, you start really thinking about your hand and how it works. - I do all the time. - The sensory control homonculus is where you have humongous hands.
4:39- [Lex] Yeah. - So I mean, like your hands, the actuators, the muscles of your hand are almost overwhelmingly in your forearm. So your forearm has the muscles that actually control your hand. There's a few small muscles in the hand itself, but your hand is really like a skeleton meat puppet.
5:00And with cables. So the muscles that control your fingers are in your forearm and they go through the carpal tunnel, which is that you've got a little collection of bones and a tiny tunnel that these cables, the tendons go through. And those tendons are mostly what move your hands.
5:21- And something like those tendons has to be re-engineered into the Optimus in order to do all that kind of stuff. - Yeah, so like the current Optimus, we tried putting the actuators in the hand itself, but then you sort of end up having these like- - Giant hands? - Yeah, giant hands that look weird.
5:37And then they don't actually have enough degrees of freedom and/or enough strength. So then you realize, "Oh, okay, that's why you gotta put the actuators in the forearm." And just like a human, you gotta run cables through a narrow tunnel to operate the fingers. And then there's also a reason for not having all the fingers the same length.
6:00So it wouldn't be expensive from an energy or evolutionary standpoint to have all your fingers be the same length. So why not do the same length? - Yeah, why not? - Because it's actually better to have different lengths. Your dexterity is better if you've got fingers at different length.
6:12There are more things you can do. And your dexterity is actually better if your fingers are a different length. Like there's a reason we've got a little finger. Like why not have little finger this bigger? - Yeah. - 'Cause it allows you to do, it helps you with fine motor skills.
6:28- This little finger helps? - It does. - Hmm. (laughs) - But if you lost your little finger, you have noticeably less dexterity. - So as you're figuring out this problem, you have to also figure out a way to do it so you can mass manufacture it. So it's to be as simple as possible.
6:43- It's actually gonna be quite complicated. The as possible part is it's quite a high bar. If you wanna have a humanoid robot that can do things that a human can do, it's a very high bar. So our new arm has 22 degrees of freedom instead of 11 and has the actuators in the forearm.
7:03And all the actuators are designed from scratch, from physics first principles. The sensors are all designed from scratch. And we'll continue to put a tremendous amount of engineering effort into improving the hand. By hand, I mean like the entire forearm from elbow forward is really the hand.
7:26So that's incredibly difficult engineering actually. And so the simplest possible version of a humanoid robot that can do even most, perhaps not all, of what a human can do is actually still very complicated. It's not simple. It's very difficult.
资源与出处
- 视频在 YouTube 收看
- 档案silicon-valley.tech 档案页本页视频与文稿数据来自 Elon Musk Archive
- 精读本篇精读 Markdown 版