访谈·中文精读

完整版:CMG独家专访Tesla CEO马斯克

Full: Exclusive CMG interview with Tesla CEO Elon Musk

中文精读由 AI 基于原始视频与文稿改写,非官方翻译,可能存在偏差或错误;表述以原视频为准。说明

2026 年前几个月,中国的发电量超过了美国、欧洲和印度加起来。这句话出自马斯克,说在 Tesla 全球工程总部的采访间里,语速很平,像在报天气。

他往前又推了一步:中国现在的发电量大约是美国的三倍,如果把人均拉到和美国一样的水平,这个数字会变成四倍,正好对上两国的人口比例。

把这句话和同一场访谈里的另一句放在一起,味道就出来了。他说中国 AI 公司在算力受限的情况下,做到了每单位算力性能的全球第一。电最多,算力最紧,效率最高,三件事同时成立,构成了这场对话最硬的背景音。

这场专访发生在 2026 年 10 月。Cybercab 已经在德州跑商业运营,Grok 4.7 刚发,Optimus 在镜头前跟记者玩石头剪刀布,星舰准备下个月月底试着把飞船夹住。马斯克一边说 AI 快得让他头晕,一边把未来十年的账本摊开讲了一遍。下面按话题把他的判断捋清楚,哪些是数据,哪些是预测,哪些他自己也没把握。

他对中国的判断,最后都指向人

记者问他 5 月随团访华、见到中国领导人是什么印象。他的回答没有绕,先讲任期内的变化,再讲自己亲眼看到的东西:一年之间、甚至一年之内,新建筑、新基础设施,还有街上普通人生活水平的提升,都很直观。

中国的繁荣程度提升得非常大,这发生在他担任中国领导人的时期。所以他显然做得非常出色,也非常有能力。(马斯克)

这段评价很短,信息量在后面。他把中国的变化归到"可见"这个词上,看得见,不需要靠报告来相信。一个做制造业的人用这种方式判断一个国家,逻辑是自洽的:工厂不会撒谎,基建不会撒谎,街上的人过得怎么样也不会撒谎。

谈到上海超级工厂,他给的答案更出人意料。自动化产线有多先进,产能爬坡有多快,这些他一句没提。他把功劳归给了"中国"这个前提。

中国有一种非常强的、内在的、把制造业做好的能力。这不只是今天如此,过去两三千年都是如此。说句实话,Tesla 上海的魔力来自中国。(马斯克)

记者追问,那是不是因为在中国的缘故。他答"是的",然后才补上团队的部分:有才华、勤奋、值得信任,质量优秀,效率优秀。在他的排序里,土壤在前,人在中,设备和工艺在后。

他还主动提了一件很多工厂老板不会主动提的事:照顾工人,保证好的医疗、好的伙食,让大家愿意来上班。"这很重要",他说了两遍。

上海工厂被叫作"宝石",这个词选得很准

他用的原词是 gem,宝石,还加了一句"其实很漂亮,去过的人都这么说"。

为什么值得掂量这个词。Tesla 在全球有多个工厂,德国柏林、美国德州、内华达、加州弗里蒙特,每一座他都夸过。上海得到的评价里多了一个维度:漂亮。漂亮是个主观词,落在工厂上,指的是布局、动线、秩序感这些东西,是长期迭代才会长出来的样子。

更实在的一层在产量。上海工厂是 Tesla 全球最大的出口基地,供应欧洲和亚太。他多次在财报会上把上海团队单拎出来表扬,这次在镜头前又重复了一遍。一家美国公司的 CEO,在中国官方媒体的专访里,用一个带感情色彩的名词形容自己在本土的工厂,这个选择本身就是态度。

再往深一层看,他夸的是"两三千年"的制造传统。这个说法在商业语境里很罕见,因为大部分跨国公司的 CEO 谈海外工厂,谈的是成本、关税、供应链半径。他谈的是文明尺度上的能力积累。这种表述对中国的听众当然友好,它同时也是一种真实的判断方式:他相信一国制造能力的上限,由这个国家长期形成的工程文化决定,短期内砸钱砸不出来。

Cybercab 已经在路上跑,方向盘没了

记者提到 9 月是"Cybercab 月",网上的视频传得飞快,人群围着车拍。马斯克说设计之初就是要让它看起来像来自未来,还顺带吐槽了一句街景不该停滞不前,汽车的审美、甚至衣服的审美都该往前演化,五六十年代、七八十年代都各有一轮快速变化,现在该轮到下一次了。

真正值得记的是运营进度。这车没有方向盘、没有踏板、没有侧后视镜、没有后视镜,目前在德州已经商业运营,很快会进佛州和内华达,明年年中大概会进加州。

这几个地名连起来是一条监管路径。德州是美国对自动驾驶最宽松的州之一,佛州和内华达紧随其后,加州最严也最大。能在加州拿到商业运营许可,等于拿到了全美市场的通行证。他给的时间点是明年年中,语气没有用"希望"这类词,用的是"大概会"。

这背后还有个产业含义容易被忽略。方向盘和踏板被拿掉,意味着这辆车在设计阶段就已经假定"人类不接管"。这跟 L2、L3 那一套渐进路线的哲学完全不同,后者永远给人类留一个兜底位,前者从第一张图纸起就把人排除在控制回路之外。工程上,这条路线没有回头路。

做 AI 的人也被 AI 弄得头晕

记者说现在每天刷社交媒体都能看到 AI 的新突破。马斯克接了一句很诚实的话。

AI 软件改进的速度、重大发布的密度,快到我脑袋发晕。我自己就在 AI 行业里,还是发晕。(马斯克)

他描述了自己的日常:睡觉时出现一个突破,醒来又有一个,到午饭时还有第三个。这句话里没有修辞成分。这就是他每天的工作体感。

一个同时管着 xAI、Tesla 自动驾驶、Optimus 三条 AI 线的人,说自己跟不上,这句话的分量比任何行业报告都重。

然后他讲 Grok 4.7。评价非常克制:是个扎实的工作马模型,但不如刚发布的 Opus 5.5。他没回避差距,还给了原因:xAI 做 AI 只有 3 年,Anthropic 做了 6 年。

接下来这句是全场最有信息量的预测之一。

我预计我们大概会在明年某个时候追上前沿。最有可能。(马斯克)

敢给时间点的 AI 公司 CEO 不多。他给得干脆,也给了缓冲词"最有可能"。旁边还有一个数字:Grok 的助手产品月增长约 100%,也就是一年十几倍。这个增速在消费级 AI 助手里属于第一梯队。

这里有个细节。他对自家模型的评价用的是"工作马"(workhorse)这个词,不是"最强",不是"领先"。工作马的意思是:不算最聪明,但能扛活、成本可控、可靠性够。一个把火箭送上天的工程师,评价软件时用的还是他最熟悉的那套标准。

他手里那张别人没有的牌:现实世界的工程数据

记者问到一个流传已久的话题:Grok 已经在用 SpaceX 的真实工程数据训练,这会不会是冲破 AGI 天花板的秘密武器。

他的回答里有一半是否认的:目前只用了一点点。另一半是完整的战略说明。

Anthropic 做得极好的一件事,是让 AI 在软件工程上变得出色。但还没有人让 AI 在现实世界的工程上变得出色。这就是 Grok 的机会,用 SpaceX 和 Tesla 的训练数据,让 AI 在现实工程上做到极致。(马斯克)

这句话把当下大模型竞赛的盲区说清楚了。编程、写作、数学、考试,这些任务的共同点是:数据在网上,可以爬,可以合成,可以拿来自我对弈。而现实工程的共同点是:数据在车间里、在发射台上、在风洞里、在一次又一次失败的爆炸里。

火箭为什么在某个温度下振动超标,电池包的冷却流道为什么要那样绕,发动机涡轮泵的密封圈在多少次循环后会疲劳,这些东西不在互联网上。它们只存在于那些真的把东西造出来、并且失败过的公司内部。

他说"目前只用了一点点",意思是这张牌还没打完。如果 3 年追上 6 年的差距靠的是算力和人才,那追平之后靠什么拉开差距,答案就藏在这句话里。

顺着这个逻辑还能往前看一步。软件工程之所以能被 AI 拿下,是因为反馈回路极短:写代码、跑测试、报错、改代码,一轮循环几秒钟,可以无限重复。现实工程的反馈回路以月甚至年计,一次试车、一次发射、一次碰撞测试,成本高到不能随便做。谁能把这种低频、高成本的反馈变成模型可用的数据,谁就握住了下一个阶段的入场券。这也是为什么他反复强调火箭复用:复用把发射成本打下来,同时也把试验频次推上去,两者对 AI 训练是同一件事。

中国 AI 的第一名,是每单位算力

谈到中国模型,他的评价比很多人预期的高。在智能排行榜上,中国模型表现优秀,尤其是在算力有限的前提下。

如果按每单位算力对应的性能来排,中国远远是第一。(马斯克)

这个评价需要翻译一下。单位算力性能,指的是同样的芯片、同样的电、同样的训练时长,谁做出的模型更聪明。这个指标衡量的是效率,是工程能力,是被约束逼出来的优化能力。芯片被卡住的一方,只能在这个维度上下死功夫。

他接着给了一个时间表。

中国会解决算力受限的问题,速度比大多数人预期的快。我粗略的猜测是,中国大概两三年内会在光刻和芯片制造上解决这个约束。(马斯克)

两三年。这是全场唯一一个他主动给出、且具体到年份的技术预测。他没展开论证,但从他后面谈电力的态度看,他相信底层工业能力可以被组织起来攻克,光刻机不比造车难到不可想象的地步,只是需要时间和投入。

真正的瓶颈已经从芯片挪到了电

9 月的 G20 视频会议上,他公开提过一个警告:算力和电力的短缺。这次他把数据摊开了。

中国今年早些时候的发电量,超过了美国、欧洲、印度加起来的总和,而且还在显著增长。如果按人均对齐,会从美国的三倍变成四倍。他用的词是"这只是一个事实陈述"。

为什么电力是 AI 的硬约束。一座十万卡级别的训练集群,功耗在几百兆瓦量级,相当于一座中型城市。芯片可以花钱买,厂房可以加班盖,电不行。电厂、输电网、变电站的建设周期以年计,任何一个环节卡住,集群就是一堆发热的金属。

这也解释了他为什么把中国 AI 的效率排在第一位之后,紧接着谈电。算力受限的解法有两条:一是把芯片做出来,二是把每瓦电榨出更多算力。中国同时在两条路上跑,而且电力这条路上本来就领先。两三年这个时间点,是把两条曲线叠起来看的结果。

还有一层更长远的影响。AI 训练的规模扩张,成本结构正在从"买卡"转向"买电"。当电力成为主要边际成本,拥有廉价、稳定、大规模电力的地区就会获得结构性优势。这跟过去二十年制造业向哪里转移的逻辑是同一个逻辑。

AI 安全不能只有单边立法

记者问 AI 安全领域期待什么样的共识。他给了一个具体的组织形态:一个工作委员会,专门研究安全。

安全这件事要成立,不能只有美国的监管,也不能只有中国的监管。问题是,怎么把监管公平地适用于任何国家生产的 AI。真正重要的确实是中美两国。(马斯克)

这段话的内在逻辑是:AI 模型是跨境的,模型权重可以下载,API 可以调用,监管如果只在一边生效,结果就是模型训练搬到监管松的地方去,谁都不安全。他用了"公平适用"这个词,指的是规则要对所有生产者一致,不能变成针对某一方的工具。

他还提了一个前提:安全监管要服务于全人类的安全。这话听起来空,放在他的处境里却很好理解。他同时在中美两边有巨大的商业利益,Tesla 上海是最大的工厂,xAI 要和中国的模型公司竞争。他需要一个两边都能接受的框架,否则任何一边的监管收紧都会直接切到他的业务。

换个角度看,他的提议很保守。工作委员会、监管对齐、公平适用,这些都是最低限度的技术性安排,谈不上什么宏大架构。他很清楚在当前的国际环境里,能谈成的只有这种颗粒度的东西,谈不成的是价值观层面的共识。先让两边的技术团队坐在一起定义什么叫"危险能力",比先吵原则有用得多。

机器人赢你,靠的不是聪明

采访前,摄制组和 Optimus 玩了石头剪刀布。前两局人赢了,剩下的全被机器人拿下。

马斯克听完笑了,然后解释得很实在:机器人反应时间快。

人类受生物学限制,机器人只要改执行器、改传感器、提高相机帧率、加算力,就能在物理层面做得比人快得多。这是结构性的优势。(马斯克)

这个解释拆掉了很多人对人形机器人的浪漫想象。它赢你不需要理解你在想什么,不需要读你的微表情,只要在二十毫秒内完成识别和动作输出就够了。人手从看到信号到做出动作,反应时间在两百毫秒量级。差一个数量级,猜拳就是必输的游戏。

他还提到美国社交媒体上疯传的机器人比赛:拳击、摔跤、跑步、体操。他说好看,紧接着补了一句关键的话:如果机器人只是摔倒在地,那就不好看了。能打、能跑、能翻,说明关节力矩、平衡控制、实时决策这一整套东西在真实环境里成立了。

这一段值得多想一下。娱乐性和技术成熟度在这里是同一件事。拳击要求机器人在被击中、重心偏移的瞬间恢复平衡,摔跤要求全身接触下的力矩协调,跑步要求落地冲击的吸收和能量回收。这些能力在实验室里做不出效果,必须靠大规模试错和硬件迭代。观众觉得好看,是因为他们潜意识里知道"这东西真的做到了"。

十亿台这个预测,他自己说是"容易的"

记者提到他此前说过十年内人形机器人数量会超过十亿台。他的回应是:这个预测很容易做。

不晚于十年,可能更早。他给的推演方式是复利:如果每年翻一倍,三到四年就是一个数量级,那么十五年就是百亿量级,再往后是千亿。这个算法很粗糙。他想让人看见的是曲线的形状。

对比一下人类这边的产能约束,差距就更明显了。人类需要大约 20 年长大,最后 20 年退休,中间大约 40 年是生产期,占寿命的一半。而这 40 年里还要睡觉、吃饭、上厕所,实际能投入工作的时间大概每周 40 到 50 小时。

机器人乐意每周工作 168 小时,连续不断。而且它一出生就是工作年龄,没有退休这回事。(马斯克)

168 小时是 24 乘 7。这个对比不需要任何形容词,摆在那里就够了。

不过这个预测里藏着一个容易被忽略的前提:机器人数量不等于机器人产能。真正决定经济影响的是"机器人小时数",也就是数量乘以有效工作时间。一台每天故障四小时、需要两人维护的机器人,账要重新算。他在访谈里没展开这一层,但从 Tesla 造车的经验看,他自己比谁都清楚产能爬坡的坑有多深。

"钱可能不重要",这句话该怎么读

他讲到富足时代的时候说了句容易被断章取义的话:钱在未来可能都不重要了。

完整逻辑是这样的:如果商品和服务的产出规模,被扩张到远超地球人口能够消费的水平,就进入了一个富足时代。而人类过去所有的进化史,都是一部约束史,缺食物、缺燃料、缺保暖。我们的本能、制度、经济模型,全部建立在稀缺之上。当稀缺消失,我们的脑子很难真的理解那是什么状态。

他把结果分成两块:90% 是繁荣的好结局,10% 是坏结局。他强调不能对那 10% 掉以轻心,AI 安全必须盯紧。他选择先讲好结局,因为那是更可能发生的那条路。

好结局的样子他描述得很具体:每个人都有自己的 C-3PO 和 R2-D2,比电影里更能干。可以照顾年迈的父母,可以看护孩子,可以做你孩子的私人教师,一对一那种。你能想到的活儿它都能接。会出现大量只有一个人的公司,配几百上千个机器人,有物理的也有数字的。一个人的生产力被放大几个数量级。最终是事实上的普遍高收入。

这里有个值得留意的措辞:"普遍高收入"(universal high income),跟他之前常说的"普遍基本收入"(universal basic income)差了一个词。基本收入是兜底,高收入是提升。这个用词变化说明他对结果的判断比几年前更乐观了。

会不会被几家巨头垄断,他的答案有点反常识

记者问了一个很尖锐的问题:如果钱不再是交换工具,怎么防止少数科技巨头控制所有资源,普通人怎么拿到自己那一份。

他没谈税、没谈再分配、没谈监管,直接给出的答案是产能会过剩到荒谬的程度。

机器人会饱和人类的需求。它们会做到超出任何人能想到的事情,然后会发现没那么多事可以为人类做了,于是转而为机器人自己做事。你想要一座城堡,机器人就给你建一座城堡。(马斯克)

这个论证的力量和弱点都很清楚。力量在于:如果每台机器人的边际成本趋近于电费和材料费,那么拥有机器人的人和没有机器人的人之间的差距,会被产能本身稀释。弱点在于:从"总产出足够"到"每个人都能拿到",中间隔着一整套分配机制,而分配从来不会自动完成。

分配这件事,他绕过去了,整段回答落在产能上。这两件事在一个足够长的时间尺度上可能合并,也可能不合并。这是他整场访谈里最乐观、也最没被追问到底的一段。

还可以补一句他话里的另一层意思。"机器人会饱和人类的需求"这个说法,暗示了一种需求侧的极限。人类的欲望确实有天花板,一个人能住的房子、能吃的饭、能旅行的次数,都是有限的。历史反复证明,需求会随着供给一起长。今天的普通人拥有的东西,一百年前的贵族也想象不出来。所以"饱和"这个词用得有点早。

中美在太空最实在的合作,是别撞上

记者问太空探索领域中美能怎么合作。很多人期待他谈联合探月、数据共享、共同建站,他给的答案非常工程师。

太空合作最重要的那件事,是中美发射卫星时密切协调轨道,避免潜在碰撞。这是最重要的合作领域。(马斯克)

为什么这件事排第一。近地轨道的卫星数量在过去五年涨了一个数量级,星链、千帆、一网,各家都在往天上堆。轨道资源在物理上有限,一次碰撞会产生成千上万块碎片,碎片再撞出更多碎片,最后整条轨道带几十年不能用。这叫凯斯勒效应,是航天界最不想看到的多米诺骨牌。

这件事的妙处在于它不需要信任。两家公司互相不喜欢也能谈,因为不协调的代价对双方一样致命。技术层面只需要交换轨道参数和机动计划,不涉及任何敏感信息。在所有可能的合作里,这一项的政治阻力最小,实际收益最大。

星舰真正的门槛,是飞回来还能再飞

谈到多行星物种,他把话说得很重:完全且快速的复用,是让生命走向多行星的根本突破。

他用猎鹰 9 做对比。这枚火箭回收了助推器和整流罩,但没有回收二级。

这相当于每飞一次,就扔掉一架中型客机。你可以想想,如果飞机每飞一次就要扔掉,航空业根本不可能存在。(马斯克)

这个类比之所以有效,是因为它把百分比换成了画面。一枚猎鹰 9 的二级占总成本的三成左右,听起来还能忍。换成"每次航班扔一架 737",荒谬感立刻出来了。火箭行业几十年来的高成本,本质上就是这个荒谬感的常态化。

星舰的进度他也给了:下个月月底希望能把飞船夹住,今年晚些或者明年初复飞。他补了一句:我相信中国到某个时候也会解决这个问题。

这句话说得客气,分量不轻。完全复用是航天领域的圣杯,谁先做到,谁就把进入太空的单位成本压下一个数量级。他承认中国会做到,等于承认这个领先窗口是有限的。

两个理由,一个是保险,一个是理由本身

他把支持多行星的论证分成两条线,讲得很清楚。

第一条是防御性的。意识如果能分布在多个天体上,文明的预期寿命会大幅提高。月球、火星、地球,三个地方都有人类文明,任何一处出问题,整体存活的概率都高得多。他打了个比方:这是给生命本身买的一份保险。

第二条是激励性的,他说这一条更能驱动他自己。

生活不能只是解决一个又一个悲伤的问题。总得有一些事,让你愿意早上醒来,让你期待未来,让活着这件事值得。(马斯克)

这条论证经常被当成鸡汤,其实它有很硬的内核。人类历史上所有大规模的集体行动,战争、宗教、大航海、工业化,驱动力都来自对好结果的奔赴,很少来自纯粹的避险。只靠恐惧维持的文明会萎缩,会进入防守姿态,会把资源全部投在保险上而不投在探索上。他想要的是后者。

人机之间真正的瓶颈,是每秒 100 比特

谈到 Neuralink,他给了一组数字,这组数字比任何宣传都更能说明问题。

人类输出的带宽,峰值大约每秒 100 比特。打字、说话、手势,全部加起来就是这个量级。输入的带宽高得多,因为视觉,眼睛本质上就是两台相机,每秒几兆比特。而计算机之间的通信是每秒万亿比特。

一个每秒万亿比特在通信的 AI,跟你说话,就像你在跟一棵树说话。你对着树说,树要很久才回应。(马斯克)

这个对比解释了为什么"AI 友好"这件事本身不够。哪怕 AI 无比善意,一心想让你满意,沟通速度的差距也会让协作变得极其低效。你花十分钟描述一个需求,AI 一秒就理解了,剩下九分五十九秒都在等你把话说完。

Neuralink 要解决的就是这个带宽约束。他把这件事归到"生物增强"这一类里,和可持续能源、太空、互联网、AI、基因编辑并列。这是他关心的第六条线,也是最少被讨论的一条。

给 20 岁的人的建议,落脚在"会提问"

记者问,在 AI 大规模替代的过渡期,一个 20 岁的年轻人怎么实现自我价值而不被甩下。

他的建议第一条是教育要尽可能宽。艺术、科学、工程、通识,都要碰。他给的理由很具体:通识决定你能不能把问题问出来。

为了知道该向机器人提出什么要求,你得有能力把问题组织出来。通识越宽,你越能把问题问对,也就是现在说的 prompt engineering。(马斯克)

这个逻辑值得多想一层。当执行的成本趋近于零,稀缺的东西就变成了"知道要什么"。而知道要什么,前提是你见过足够多的东西,知道什么是好的,什么是没意义的,什么是前人试过并且失败的。这些全靠知识面的宽度,不靠单点技能的深度。

他还说了一句听起来很狂的话:未来会有那么多机器人和那么多 AI,AI 会非常渴望听到你提出任何请求,因为请求会被立刻满足。这句话的另一面是:需求会成为稀缺资源。能想出值得做的事的人,会比能做事的人更稀缺。

顺便说一个他没展开但很明显的事实。相变发生的时候,被淘汰的往往是把单一流程做得最熟练的那批人,因为单一流程最容易被自动化;留下来的是能在不同领域之间搭桥的人。通识在这个意义上是抗风险资产。

他会带美国企业家去看兵马俑

最后一个问题很轻松:如果能带一位从没来过中国的美国企业家来看看,最想让他体验什么。

他的回答是:能挤出多少时间就给多少时间,中国有太多可看的东西。至少要去上海、北京。他特别推荐西安和兵马俑,说自己是坐高铁从北京到西安的,体验非常好,火车站令人惊叹。兵马俑是世界奇迹之一。

然后他补了一句很马斯克的话:记得看窗外,别玩手机,可看的东西太多了。

访谈最后,他对着镜头说了这样一段。

总的来说我想说,中国很了不起,是一个不可思议的国家。我鼓励全世界的人都去中国看看,亲自体验一下。(马斯克)

他还说,任何他能说出来的话,都不足以描述中国那种惊人的壮丽。

这句话放在整场访谈的结尾,和开头那段关于发电量、关于三千年制造能力、关于上海工厂是宝石的表述,构成了一条完整的线。他的判断方式始终一致:看东西造出来没有,看街上的人过得怎么样,看电表转得多快。

一个工程师看世界的方式,大概就是这样。

可以从这场访谈里拿走的几件事

  • 看 AI 竞争别只盯跑分榜。真正会拉开差距的是训练数据的独占性,尤其是现实世界工程数据。软件工程的活已经被做得很好,物理世界的活还空着。如果你在制造业、能源、航天、生物这些行业,你手上的数据可能比你以为的值钱。
  • 把"会提问"当成一项要练的技能。执行成本在快速下降,问题定义能力在快速升值。练法很朴素:多读跨领域的东西,多问"这件事为什么是这样",多把模糊需求拆成可执行的句子。
  • 关注电力,不只关注芯片。算力扩张的下一个卡点是发电和输配电。做投资、做选址、做基础设施规划的人,把电力的可获得性放到和芯片同等重要的位置去算账。
  • 对 10% 的坏结局保持警觉。他自己说 90% 是好结局,同时强调不能对剩下那 10% 掉以轻心。落到个人行动上,就是别把 AI 治理当成别人的事,看到具体的监管提案时认真读一遍。
  • 想理解机器人产业,盯三个指标:量产节奏、单台成本、真实场景的连续工作时长。发布会上的后空翻参考价值有限,能连续干 168 小时不出故障才算数。
  • 如果做跨境业务,学他看中国的方法:不看二手报告,去看工厂、看基建、坐一趟高铁、看窗外。判断力和信息的一手程度高度相关。
  • 把时间尺度拉长一点做规划。他给的技术预测大多落在三到十五年这个区间,两三年解决光刻,十年内十亿台机器人,十五年到百亿量级。这个尺度对个人职业选择是有意义的参照,短于三年的判断容易被噪声带偏。
YouTube 平台字幕(自动或人工),可能有识别错误,人名与数字请以视频为准。已按语句合并成段落,并去掉 um、uh 等口头填充词。 文稿来源:YouTube(自动/人工字幕)。

0:00Thank you so much Elon for joining me, and also having us here at Tesla Global Engineering Headquarters. It's so amazing to see kind of new technology innovations, all talented engineers here as well. Yes.

0:11In this May, you were also on the delegation to China as well. You have met with President Xi. What's your [music] impression about President Xi as a leader?

0:18Well, I it seems like a great leader. I mean, he's under his tenure, [music] China has prospered immensely. I mean, I've seen it firsthand. Like you you if you visit China [music] and you see it the the amount of progress that happens from 1 year to the next or even within a year. [music] you can it's very visible. Like you can see the the new buildings, the new infrastructure, [music] and and it's also obvious to see that I just walking around the streets, [music] the increased prosperity of the average Chinese citizen has increased tremendously. And that's hap- well, that's you know, [music] it's happened under President Xi's, you know, while

0:53he's been the leader of China. So so he's clearly done a very, you know, done an excellent job [music] and is very capable.

0:59Tesla Gigafactory Shanghai has always done an incredible job. You praise your teams many times. Yes. [music]

1:05So what kind of advantage or breakthroughs do you think that to making the Shanghai factory a very important position in Tesla's global blueprint? And what ultimate kind of confidence [music] do you think this will, you know, unfold especially in the market?

1:19Well, I mean, China just has a very strong inherent ability [music] to do manufacturing well. And this has been true not just in present days, but [music] for the past two or 3,000 years. I think that to be totally frank, the magic of Tesla Shanghai is [music]

1:35because of Chinese to be totally

1:39Because it's in China at least, right?

1:41Yes. We what we have is the Tesla China team [music] is just incredibly talented and hard working and trustworthy. [music]

1:51and they have the the quality is excellent and the efficiency is excellent. So the Tesla Shanghai factory is a is a gem. You know, it's it's and it's beautiful it's beautiful actually if anyone's ever [music] been there. and we also care a lot about taking care of the workers and making sure they, you know, they're good health care [music] and good food and enjoy coming to work. It's very important. So,

2:16Absolutely. I get I I I think we can tell like all your stuffing Shanghai factory they're very enjoying the working there. [music] So, we're talking about next hot topic recently. We probably we can call September is the Cybercat month. Have you seen the videos like all the people gather around that? I guess it's

2:33It's It's goes very very quickly.

2:35Yeah. I mean, it's pretty cool cuz we designed the car to look futuristic. Yeah.

2:38Cuz I think you know, from a fashion standpoint, we don't we don't want as the the the aesthetics of the street to stay stagnant. You You want to evolve the what what do cars look like? And also, I think maybe even clothing. Like we should be Like there was there was rapid fashion evolution in the '50s, '60s, '70s, '80s, '90s, and probably time to make things look more futuristic. Yeah. [laughter]

3:06And then how you see the kind of market potential for Cybercat? You know, it's it's [music]

3:11it's like, you know, it's really future futuristic, but how long it can really come to the reality or real life?

3:18Well, we have the Cybercab which has as you know, has no steering wheel or pedals or side mirrors or rearview mirror operating commercially in Texas [music] right now. And it will soon be operating commercially in Florida and Nevada and a number of other states. And probably by middle of next year, it'll be operating commercially in in California.

3:40Absolutely. I think all the changes is happening really fast. You know, it's just like everyday we have to to look social to see the kind of improving or breakthroughs about AI or other old tech topics. So,

3:53Yeah, I mean, AI is just so crazy. It's I mean AI the rate of AI software improvements so and and major announcements is so rapid that it's it's makes my head spin. You know, and I'm [music] in you know, I'm in AI and it still makes my head spin.

4:06industry but you still have to

4:08It's still [music] shocking. Yeah.

4:09you know, sometimes I'll like when I go to bed there'll be some big AI breakthrough and then when I wake up there's an AI breakthrough and by the time it gets to lunch there's another AI breakthrough. Exactly. Yeah.

4:19So that's why even during this break you have to check some others.

4:21yeah, I probably have to check some others.

4:23Yeah. So talking about like AI models, Tesla had the latest what version, right? The latest release about the Grok 4.7. Yes.

4:32I think this is your most focused these two days. I I just saw you post on social saying that you know, the Grok bot is already in your Tesla. Tell us like what's the reactions, what's the latest you have received from?

4:44Yeah, I mean Well, we just [music] we were we we release a new model, you know, roughly every month or two. Yeah. so [music]

4:53and it it's a cool Grok 4.7. It's a solid workhorse [music] of a model. it's not as it's not as good as say Opus 5.5 that just got released. but also we're we've only been around for about 3 years doing AI whereas Anthropic's been around for 6 years. So so I I would expect that we probably catch up to Frontier sometime [music]

5:22next next year. Most likely. In the meantime, I think Grok is very useful for many things. and Grok bot that's kind of like a personal digital assistant. has [music]

5:36been very popular. and is growing about 100% a month. So it's growing very rapidly. [music]

5:44I think you are very humble you saying that like Grok 4.7 is not a perfect wine, but we know that although have experienced a little bit delay, but you know that people are already very fascinated or interested to know because Grok 1.7 already being trained by [music]

5:59SpaceX real-world engineering data. So, do you think

6:03Only a little bit so far.

6:04A little bit. Okay. So, do you think that this will be one of your secret weapon to kind of break through the AGI ceiling cuz currently, you know, that the models are have the challenge for data.

6:13Yes. Well, I think we've got a real opportunity [music] with the SpaceX training data and maybe with the Tesla training data to make AI extremely [music] capable in real-world engineering. So, what Anthropic did extremely well was make AI excellent [music] at software engineering.

6:32but no one has yet made AI excellent at real-world engineering. And I think that's an opportunity for Grok using the training data from SpaceX and Tesla to be excellent at real-world engineering.

6:45You mentioned many times like previously, Chinese models also doing a fascinating job. So, currently, what's the most impression breakthroughs for Chinese models in your understanding?

6:55Well, they've they've been the Chinese [music]

6:59AI models are generally outstanding, actually. I mean, you can see on the sort of [music]

7:03intelligence rankings the es- especially given the limited amount of compute [music] that the Chinese AI companies have, the the performance is excellent. I'd rate probably China as being by far the best in terms of performance per unit of compute. And as I've said publicly you know, I think [music] China's going to address the compute limitation in I think faster than most people expect. [music]

7:30my rough guess would be that China addresses the compute constraints with lithography and chip making in about two or three years. [music]

7:37Two or three years. We will really expecting to this time. So, what are the challenges facing for AGI in early September in the G20 visual session, you have raised a very serious concern about the compute and electricity shortages. You mentioned that probably China have way more capability about electricity infrastructure. Yes.

7:58And can you elaborate a little bit more about this gap? And you know, what areas do you think the Chinese industry and capability can help to get a new breakthrough for the specifically on the electricity

8:10infrastructure for the whole world. It was just a statement of fact that China think is a China like earlier this year, China passed [music] the electricity output of the United States, Europe, and India combined. Wow.

8:22Yeah. And and continuing [music] to increase significantly. So, if if China reaches [music]

8:31the same GDP electricity output of the US, which I think is likely, then China would would go from where it is now roughly three three times the US electricity production to probably four times, which would be proportionate to the population difference.

8:44Talking about the topic like AI safety, what kind of crucial consensus do you expect?

8:48Yeah, I think maybe there could be a working committee to figure [music] out safety. because it's I think in order for safety to work, it's it can't just [music] be regulation in the US or regulation in China. It's got to be we you know, how do we how do we apply [music] apply regulation fairly to AI produced in any country, [music]

9:09but really China and and the US are the the the ones that really matter here. So, I think a really important [music] thing that would be good for the safety of AI for all of humanity.

9:21Yeah. So, let's talk about the human right robotics cuz I think this is also a kind of industrial field US-China can still also working together leading the world. And also I I can give you one example for the Optimus. It's really amazing, really impressive. So, before interview, we have a a kind of interactions. We play rock, paper, scissors with Optimus. Oh, yeah. Okay.

9:43Yeah, we just won the first two rounds, and Optimus won all the remaining. That's really impressive.

9:48Well, robots have a very fast reaction time, so it's going to be very difficult to beat a robot for reaction time. cuz we we are somewhat constrained by our biology. But, a robot, you can just improve the actuators, improve the sensors, improve the the frame rate of the cameras, increase the amount amount of AI computer is, and it's going to just fundamentally be much better, and have faster reactions than humans.

10:14I enjoyed watching the the robot games. I mean, everyone in America was fascinated by that, but Yeah, you said it.

10:20all over US social media.

10:22Yeah, you said it's very entertaining to watch, right? Yeah.

10:24Yeah, but other than that

10:26Well, but it's it's entertaining to watch, but it also shows you the progress of humanoid robotics.

10:31Exactly. like, it wouldn't be entertaining to watch if the robots just fell over. Of course.

10:35so, but you know, the robots like boxing, and then like wrestling

10:39wrestling, running, doing gymnastics. it's pretty cool.

10:43So, I mean, the future's going to have a lot of robots. Yeah. Like, a lot a lot.

10:47Yeah. You mentioned that you're seeing another 10 years, right? There will be Yes.

10:51the population for humanoid robots will surpass 1 billion. Yes. So, that's that's

10:56I think that's an easy prediction to make.

10:57Easy prediction, maybe within 10 years. [music]

10:59I'd say no later than 10 years, I'd say.

11:01yeah. So, what what Let's talk about what's the real life will look like in 10 years, you know, when there are so big you know, number of population of humanoid robots, and also you mentioned that one humanoid robots will have the five times of production productivity for human. Pretty much.

11:17Yeah. So, what what the life will be, and then I mean, specifically like like ordinary people, how we can cope with the the living, you know, with all AI kind of dominated the world?

11:29Well, I think my So, I I I let me focus on the what's the 90% likely outcome, which is the prosperous, the good outcome. Mhm.

11:37We don't want to get complacent regarding the 10% but may probably of a bad outcome. So, we need to It's important for us to pay close attention to AI safety. But let me tell you about the 90% positive outcome. The 90% positive outcome is that everyone's got robots that are their their buddies.

11:56and if you think of like Star Wars, we had C-3PO and R2 D2.

12:00and they were friendly, helpful helpful robots. [music] Well, wouldn't it be great if everyone could have their own personal C-3PO R2-D2? Mhm.

12:09But even better than that. Yep. Like more productive. Mhm.

12:11that could take care of elderly parents or be a good guardian and watch over the children or or be a teacher, like a individualized tutor to your child. Mhm. [music]

12:25and and taking care of any jobs that that that you can think of. So, that would be pretty cool. Exactly.

12:32everyone would like like to have a future where that they can have their own personal robots. And And then I think you'll see many companies where it's one person with hundreds, maybe thousands of robots, but like physical robots as well as digital robots. so, the productivity of one person would be greatly amplified.

12:53and and my prediction is that is that there will be effectively universal high income. Mhm.

13:00in fact, it's not clear to me that money will even matter in the future. like if you if you scale the economy the output of goods and services to a level much greater than could possibly be consumed by the number of humans on the planet, then you you have an age of abundance that is that is difficult for us to wrap our minds around because [music] our entire evolutionary history has been an evolution of constraints, [music] of of where we didn't have enough food or we didn't have enough you know, heat or and you know, so that's certainly going to be true if the output of goods and services is constrained by the number of humans.

13:41Because you know, humans take about 20 years to grow up and then they are generally retired for the last 20 years of their life. So there's about 40 years, maybe half the life of a human that's productive. But then humans also need to sleep and they need to do the eat and go to the rest restroom, you know, all the basic stuff.

14:00and and and so the actual number of hours that that most humans can work is maybe [music]

14:07around 40 to 50 hours a week, something like that. whereas the robot will be happy to work 168 hours a week. Continuous. [music]

14:17and this and and the robot is instantly at working age and does not have retirement.

14:22Exactly. Yeah. So you give us a very beautiful picture about how the life or how highly efficiency about the productivity by the humanoid robots. But for example, you just mentioned that maybe money will not be that kind of concern at that time. If Yeah, if money will not be a for example like a tool for exchange, in that kind of world, do you how do you think that we can try to prevent you know, the only a handful of few [music] tech giants in you know, controlling the all the resources and is specifically how ordinary people do you think they can claim their share in this kind of full abundance?

14:57Yeah, what I'm trying to say is that I think that there will be a level of abundance that [music] is more than anyone could possibly consume. I mean, I think it's going to get pretty wild like where if you want a if if you want a castle, you know, the robot will build you a castle.

15:16so it's not it's not like it's it's not as though like the robots will be hoarded by a small small number of It like actually the robots I think will be they will saturate on human demand, meaning that [music] they will be they will do more than anything humans could possibly think of. Mhm. and [music]

15:34and and and they'll run out of things to do for humans. Yeah.

15:39because there is a finite amount of things you could do for humans [music] within reason. so I think the robots will actually run out of things to do for humans [music] and then they will do things for themselves because they they've run out of things to do with for humans. So and and probably probably there will be I mean like my prediction is like as you talked about at least a billion humanoid robots within 10 years. Yeah. Probably less than 10 years. [music]

16:08I think that continues to grow and probably if if if if if if you assume that things roughly double per year then then that's about every three or four years you have a [music] 10x increase. Yes.

16:22So you know, maybe that would mean say in 15 years there are 10 billion robots.

16:3010 billion just 15 years. years 100 billion.

16:34Exactly. So this that's why we we always call it like AI world is changing improving exponentially. Continue about the cooperation between US China or what areas we can do together. The space exploration cuz currently many Chinese commercial space companies they are strive to you know, to explore more use after the successful land recovery. what's kind of areas or avenues do you think possible avenue do you think between communications or corporations can can get between US and China?

17:06Well, I mean in space I think the most important thing to achieve cooperation with is to make sure that when satellites are launched from the US or China that we we closely coordinate where what orbits the satellites are going to to avoid a potential collision. I think that's the most important area of cooperation.

17:23and then in in in general I I would say that the important thing for space launch, the fundamental breakthrough that's needed is full and rapid reusability. which we're hoping to achieve with the Starship design [music] by catching the the ship hopefully around the end of next month [music]

17:42and then re-flying the ship either later this year or early next year. and that will be a very important breakthrough and I'm sure others will solve that that problem as well and I'm pretty sure China will will at some point solve that problem, too. but it's it's [music] the full and rapid reusability of a rocket is the fundamental breakthrough needed to make life multi-planetary.

18:02Continue to talk about multi-planetary species. You always you have been always believe that human should become a multi-planetary species. So, tell us what's the current crucial step for your Mars mission.

18:13Well, like I said, the the most profound and and fundamental in- invention necessary for making life multi-planetary is rapidly and completely reusable rockets. Mhm.

18:23Oth- otherwise the costs are impossible. you know, with with our Falcon 9 rocket we we recover the booster stage and the fairing [music] but not the upper stage. And this is like throwing away a medium-sized jet on every flight. and you can imagine if in air flight [music] if you had to throw away the planes after every flight, you're you're way far too expensive [music]

18:44to and and and air flight would not be a realistic option. Well, that's been true in [music]

18:50in rockets for for a very long time. So, I can't emphasize enough and I've said this so many times over the years. full and immediate reusability of of rockets like aircraft [music]

19:04is the is the fundamental breakthrough that's needed to extend life beyond Earth [music]

19:08to make life multi-planetary and to create self-growing cities on the moon and Mars. Mhm.

19:15And and if we're the as soon as we're able to create a self-growing city on the moon or Mars, the probable [music] lifespan of consciousness is dramatically greater. because now [music] we don't have all our eggs in one basket. so if something would have something wrong would go wrong with say the [music] moon or or Mars or Earth, but we've got human civilizations on all three of those places, then our chances of surviving are much [music] greater.

19:42Of course. So three different planets. So that's why we all have to do our part. Follow

19:48And also it's just super cool.

19:49Exactly. We have to follow all the tech giants' steps, try to make our own self, try to be the become multi-planetary species.

19:57Yes, exactly. So I mean I classify these generally as like there's two lines of main lines of argument. One is the defensive line, which is that it that having having consciousness on multiple heavenly bodies greatly improves the probable lifespan of civilization. [music]

20:13that's the defensive argument. It's you can think of it as life insurance for life as a whole. Yeah. [music]

20:18the but the one I find personally that drives me more is actually the the inspiration argument. because you see that that like life cannot just be about solving [music] one sad problem after another. There must also be things that make make you excited to wake up in the morning. Exactly.

20:36That make you look forward to the future. Mhm.

20:39that make life worth living. And the being a spacefaring civilization is one of those things.

20:45I totally agree with that. So So you know you had you you once list like five areas you care about most. Like sustainable energy, space, and also the internet, AI, and gene editing. Looking ahead for another 10 years, are there any new frontiers you want to that list?

21:04Well, I I I biological enhancement with like Neuralink, this company I created with some other people, where where we we want to try to solve the bandwidth constraint that humans have. so, so so cuz you can imagine like even in the most benign scenario where the AI is super friendly and wants to do everything you can it wants to make you happy, just our rate of communication is too slow.

21:34So, our like the average the peak human bandwidth is only about 100 bits per second, roughly. and our input bandwidth is much higher because of vision. So, there we maybe have it's a debatable point, but you know, a a megabit per second or a few a few megabits per second from our eyes, which are effectively like cameras.

21:59But, our output bandwidth is really low, maybe 100 bits per second, whereas computers are communicating at trillion bits per second. So, anything we can do to increase the communication speed of humans will improve the I think the alignment of humans and machines. cuz otherwise the the the to an AI that is communicating at [music] at a terabit a second, talking to a human will be like talking to a tree. Exactly.

22:31you talk to a tree, it's it's takes a long time to respond.

22:34Let's talk about the challenge for young people. You [music]

22:36mentioned that. You also talk about education like in the AI world. And also I had I watched that in your headquarters here, you have a sign like you can come back to school in Tesla as well. Means which means that in every age you can have the opportunity to come back to to learn always to follow the latest kind of bricks breakthrough in AI world. But, what what about kind of suggest suggestions or your predictions? At that time, how for example like a 20-year-old young people, they can get [music] their fulfill their self-worth rather than being left left behind, you know, during [music] this massive transition?

23:12My recommendation for education [music] would be to get as broad-based an education as possible. And that includes arts and sciences, engineering, wide general knowledge. [music] Like in order to know what to ask the robots for, you need to be able to formulate the question. And with the broadest [music] possible general knowledge, you'll be better able to formulate the questions or what they call prompt engineering, I guess.

23:39You got to convey your what your wishes [music] to the robot. And as crazy as it sounds, we will run [music] out There will be so many robots and so much AI that the AI will actually be eager to hear any request that you could possibly have. Mhm.

23:55because [music] that request will be fulfilled immediately.

23:58Yeah. Your last visit to China is in May. If you you got to have an opportunity to bring one American entrepreneur for visiting China for the very first time, what will be the most experience you want him to have first?

24:10Well, I I generally recommend I I mean, as much time as as as that person can devote to China, they should go because China has got so many things to see. so I don't mean to I mean, but like usually people don't have they have they don't have infinite time, so at a minimum, they they should see obviously Shanghai, Beijing. I recommend Xi'an, the Terracotta Warriors. I took the bullet train from Beijing to Xi'an. That was a great experience. Train stations are amazing. [music]

24:40the Terracotta Warriors is one of the wonders of the world. Mhm.

24:44and And just And just make sure you look out the window and And you're not on your phone. Cuz there's so much [music] to see.

24:52Yeah, like you always recommend everyone to experience to go and see China by themselves.

24:56Yes. Because any words I I any words I say do not do justice [music] to the incredible majesty that is China. It's awesome.

25:05Really appreciate for your words. Thank you so much Elon for sitting down with CCTV Business Channel and Elon we truly hope to see you again very soon in Beijing. Thank you.

25:14Thank you Elon. Thank you so much.

25:17Generally, I just like to say China is awesome and an incredible country and I encourage people [music] from around the world to visit China and experience it for themselves.

资源与出处

相关视频

同频道、同主题或相邻年份