播客·中文精读

马斯克最新 All-In 访谈(完整视频)

Elon Musk New All-In Interview (FULL VIDEO)

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

马斯克在 All-In 峰会上被问到,在华盛顿待了几个月,学到了什么。

他的回答很短:

政府基本上是无法修复的。(马斯克)

接着他补了一句更要命的。美国国债的利息支出已经超过了国防部(他临时改口叫"战争部")的预算。如果 AI 和机器人解决不了国债问题,我们就完蛋了。

这句话把整场访谈的骨架交代清楚了。后面聊的 Optimus、AI5 芯片、星舰、Grok、火星城市,在他脑子里不是几个平行的野心,是同一条链条上的不同环节。机器人干活产生产出,产出填财政窟窿;芯片和模型决定机器人能不能干;星舰决定人类有没有第二条命。

这次访谈的背景新闻也不少:xAI 以换股方式收购了 X,特斯拉第一批 Robotaxi 正式上路,董事会给马斯克提了一份新的薪酬方案,最高可以拿到接近一万亿美元的股票,但达不到业绩目标就一分没有。他本人五月份之后就没再去过华盛顿。

手是整个机器人项目里最难的部分

聊到 Optimus,马斯克给了一个很高的定位。

我认为 Optimus 会成为人类有史以来创造的最伟大的产品。(马斯克)

他给出了判断标准:现在市面上所有人形机器人公司,都缺三样东西。

第一,接近人类的灵巧度,也就是一只非常复杂的手。第二,一个能导航、能理解现实世界的 AI 大脑。第三,极大规模的量产能力。这三样都做到,才算过关。

他也承认自己现在的精力分配已经倾斜得厉害,花在 Optimus 上的脑力,可能超过任何其他单一项目。要解决的是现实世界 AI、机电一体化、供应链和量产,一堆硬问题同时压过来。

主持人追问:手为什么这么重要,为什么执行器这么难做,为什么别人做不出来,为什么必须从这儿开始。

马斯克的回答从人体讲起。人类的手是进化出来的精密机器,你仔细看看自己的手,能做的事多得离谱。挥棒球棒、穿针引线、弹钢琴拉小提琴、拆装一辆汽车。而且手的大部分肌肉不在手上,在前臂。手更像一个木偶,靠前臂的肌肉拉动肌腱,肌腱网络极其精妙。按不同的数法,一只手有 27 到 28 个自由度。

要做通用人形机器人,就必须解决手的问题。这话不是修辞。手加前臂,占了整个机器人工程难度的大部分。机器人其他部分不是不重要,但难度权重摆在那儿。

供应链的情况更极端。特斯拉想买执行器,出多少钱都买不到。市面上的电动机有成千上万种,但能装进 Optimus 的一个都没有。每一条手臂有 26 个执行器,每个执行器都包含电机、齿轮箱和电力电子,这些全部要从物理第一性原理重新设计。

人形机器人这个行业,连一条现成的供应链都不存在,得从零重建,这意味着大量的垂直整合。

有人问他,跟以前做过的项目比,这事有多难。特斯拉的 Model Y、Cybertruck、内华达的超级工厂、被叫做"fabergé egg"的 Model X。

他的排序是:比 Model X 难,比星舰容易,夹在中间。

硬件和软件哪个更卡?现阶段卡在硬件的最终设计上,主要是手。

一百万台一年,成本两万美元

价格是所有人都关心的。马斯克给的数字是:达到年产一百万台的规模,边际生产成本大约在两万美元这个量级,可能两万五,具体取决于 AI 芯片花多少钱。

芯片是大头。一颗 AI 芯片的物料成本可能就要五六千美元,甚至更多。剩下的要靠执行器的大规模效率提升来摊薄。

至于售价,他说得很直白:价格是需求的函数。成本不等于售价。

有人问什么时候能买一台放牧场干活。这次访谈里没给具体年份,只给了成本曲线和产能条件。

为什么非得是人形

这个问题问得很好:完全可以造一个比人更强的专用机器,或者比人简单的机器去干特定任务,为什么一定要做成人的样子?

马斯克的回答分两层。

第一层,如果你想干人类能干的全部事情,那就需要人形机器人。只干一部分,会容易得多。人类进化成现在这个形状和能力,是有原因的。四根手指加一根拇指确实有价值,连小拇指都挺有用。脚趾嘛,他自己也表示存疑。

第二层是主持人补的,马斯克立刻认同:世界是人设计的,为人类设计的。造出人形机器人,它天生就与这个世界的所有设施"向后兼容"。门把手、楼梯、螺丝刀、驾驶座,全部现成。

至于大脑那一层,他说基本不用额外操心。大模型的能力可以直接搬过来,机器人能听懂指令、能对话,这部分"没问题"。

访谈里还有个插曲。主持人提到他最近在网上跟 Grok 的虚拟角色 Ani 互动得挺多,马斯克笑着承认"可能有点过头了"。主持人顺势说,那可以给 Ani 一个身体。这个玩笑背后是认真的产品逻辑:虚拟角色和实体机器人会合流。

AI5 的 40 倍是怎么算出来的

特斯拉手里有两条芯片线。Dojo 负责训练,AI4 是推理芯片,已经装进所有在售车辆。AI5 的设计正在收尾。

马斯克给了一个数字:按某些指标衡量,AI5 比 AI4 强 40 倍。

40 倍听着像营销话术,他解释了拆解方式。

AI4 最严重的瓶颈之一是 softmax 运算。目前要靠模拟方式跑大约 40 个步骤,而 AI5 用几个步骤就能原生完成。AI5 还能动态处理混合精度模型。

硬件本身的账面提升是:原始算力 8 倍,内存约 9 倍,内存带宽约 5 倍。因为软硬件团队在同一张桌子上、细到极致的颗粒度上协同设计,他们清楚知道瓶颈在哪儿。把 8 倍的算力提升,乘上针对 AI4 具体短板的 5 倍优化,才是 40 倍这个数。

这种协同设计是特斯拉自己造芯片的核心理由。买现成芯片,你只能猜瓶颈在哪儿;自己设计,你知道。

还有个更重要的事实:不需要等 AI5,现在车上的 AI4 就足以做到比人类安全 2 到 3 倍的自动驾驶,甚至可能是 10 倍。

支撑这个判断的是软件。下一个版本 V14 是自 V12 以来最大的一次升级,参数量提升一个数量级,用了大量强化学习。

马斯克用了一个说法:可以把 AI 理解成对现实的一种压缩。有些压缩步骤损失太多信息,V14 修的就是这些有损环节。

到今年年底,你的车会感觉像是有意识的。(马斯克)

170 亿美元买频谱,两年后手机能直连卫星

SpaceX 花了大约 170 亿美元从 EchoStar 手里买无线频谱。马斯克开玩笑说这是"一点零钱"。

这笔钱买的东西很明确:让星链卫星直接和手机通信,而且是高带宽。

但有个前提条件,手机得改。这些频率目前的手机不支持,芯片组必须修改才能加上对应频段。时间表是两年左右,能使用这批频谱的手机大概两年后开始出货。

与此同时,SpaceX 在造支持这些频率的卫星。两条线并行推进,等卫星和手机都就位,握手就能跑出高带宽。

最终效果他自己总结:你可以在任何地方用手机看视频。

现场主持人问室内能不能用。答案是普通住宅可以,如果建筑有很厚的金属屋顶就不行。跟现在的手机覆盖逻辑差不多。

有人关心是不是要跟运营商说再见。马斯克的回应比较克制:其他运营商不会因此倒闭,因为他们手里握着大量频谱。但用户确实可以开一个星链账户,家里的天线提供 Wi-Fi,手机在外面直连卫星,一套账户打通。

被问到会不会干脆收购运营商,比如 Verizon,他的回答是"不排除这个可能"。

星舰剩下最后一块硬骨头:隔热瓦

星舰的进展,马斯克给了一个明确节点:明年回收飞船。

节奏是这样的。V2 设计只剩一套助推器和飞船,还会再飞一次。之后全面转向 V3,换装猛禽 3 发动机,火箭上几乎所有东西都变了。他提前打了预防针,这么彻底的重设计,初期出些毛病很正常。

V3 的目标参数:全复用状态下,入轨运力超过 100 吨。

对比一下现在最强的商业火箭。猎鹰重型在侧助推器复用的情况下,大约 40 吨。星舰是它的两倍半,而且箭体全部回收。

他说,除非遇到重大挫折,SpaceX 明年会演示完整的复用能力:助推器和飞船都被接住,并且把 100 吨以上的载荷送到有用轨道。

从 2002 年公司成立算起,这条路走了 23 年。他把造出全复用轨道火箭称作史上最难的工程问题之一,"至少能站上领奖台"。

现在的技术卡点,他指得很具体:飞船的隔热盾。

没有人做成过全复用的轨道级隔热盾。航天飞机的隔热盾每次飞完要修九个月。这个对比一摆出来,难度就清楚了。

这是材料科学加工程的双重难题。要同时满足几条互相拉扯的要求:扛得住高温、足够轻、不能把热量传到主结构、上升穿过雨云时瓦片不能被泡坏。最难的一条是检测,几万块瓦片,不能像航天飞机那样一块一块检查翻新,必须做到装上去就能信。

关于上次发射失败后的快速复飞,他没讲太多细节。监管的问题和审查都经历了,SpaceX 团队把下一套飞船和助推器测试完、送上发射台、再飞起来,速度很快。他把功劳全给了团队。

Grok 正在重写人类的知识库

话题转到 xAI。马斯克讲了一件很多人还没意识到的事:下一代 Grok 的训练,不打算从网页爬来的原始语料直接开始。

做法是用大量推理算力,把人类知识的全部语料过一遍,逐条判断:这条是真的、部分真的、假的,还是缺了关键背景。然后重写。删掉虚假内容,修正半真半假的说法,补上缺失的上下文。

他拿维基百科举例:让 Grok 看一个页面,判断哪些真、哪些半真、哪些假、哪些缺失,然后改写这个页面。这个方法适用于书、PDF、网站,所有形式的信息。

主持人当场提议,干脆把这个成果公开发布,做一个"Grokipedia"。马斯克说会跟团队谈这件事,尤其提到自己的传记页面"一团糟"。

他对维基百科的评价很直接:偏见严重,是一场持续的战争。你改对了一处,五分钟后就有一群人来改回去,页面上到处都是活动人士。

算力方面,孟菲斯的 Colossus 超级集群已经在扩建第二个。

算力翻十倍,智能翻一倍

关于扩展定律还有多少空间,马斯克给了一个粗略的经验法则:算力和智能大致是对数关系,10 倍算力大约能让智能翻倍。

他特意说明这只是个大概的拇指规则,但即便按这个效率,从 100 的智商到 200 的智商,依然是天大的事。

更长的时间尺度上,他认为智能会一路扩下去,直到太阳的大部分能量都被用于计算,再往后是银河系尺度,卡尔达肖夫 II 型、III 型文明那个量级。

所以他建议换个视角:不要把 AI 看成一个要抵达的终点,它只是智能升级进程的一部分。

人类智能本身也是这么涨上来的。人口增长,能存储的信息越来越多,整体智能随之扩展。现在人口在下降,人类智能趋于平台期,接下来会往下走。

他给了两个时间点:

我认为最快明年,AI 就可能在任何单件事情上比任何单个人类都聪明。到 2030 年左右,AI 可能比全人类加起来还聪明。(马斯克)

"西方的行为与自杀无法区分"

访谈在这里转向了最不技术的一段,也是马斯克说得最重的一段。

主持人提到会议前几天的一场对话,主题是"自杀式的同理心"、西方 declining 的出生率、开放的边境。三个症状,会不会是同一件事?

马斯克的回答是:西方的行为与自杀无法区分。

他列了几条证据。出生率低于替代水平,如果无限延续,西方在字面意义上无法繁衍出足够的人来替代自己。边境开放到社会结构开始解体的程度,欧洲尤其明显,英国、法国、德国的原住民文化面临被不同化的移民群体取代的风险。犯罪问题摆在那儿,精英媒体却拒绝报道。

他举了一个具体案例:一位年轻女性 Ina 在地铁上被无意义地杀害。这件事本身已经够可怕,更让他愤怒的是主流媒体集体沉默,好像这事没发生过。

"我们几乎是在否认螺旋的现实。"

他给了一个判断标准:你上次跟一个住在欧洲的欧洲人聊天,他表现得乐观,是什么时候?主持人的回答是很久没有过了,几十年。马斯克说这很罕见。

他的推论是:如果人们对未来没有乐观感和目标感,自杀可能就是自然结果。生孩子本身就是对未来的一次乐观表态。不乐观,就不会生。

宗教在其中扮演什么角色?马斯克认为过去宗教确实起了作用。自然厌恶真空,把宗教拿掉,填补空缺的东西往往比原来更糟。他用的词是"觉醒思想病毒",说这些准宗教性质的意识形态填进了宗教留下的洞,而且高度自毁。

那他自己的答案是什么?他说需要一套连贯的、让人兴奋的哲学。他自己的那套叫好奇心哲学。

我好奇宇宙的本质,我想让人类走出去,去探索群星,也许遇到外星文明,也许只看到某个存在了一千万年的文明留下的废墟。就像《星际迷航》里那些非反乌托邦的科幻作品。我就是想知道到底发生了什么。(马斯克)

要弄明白宇宙,就得扩大意识的规模和范围,就得让人类增长、向外延伸。

他提到《银河系漫游指南》,说那是一本伪装成幽默的哲学书。书里真正的观点是:难的是提出问题,答案就是宇宙本身,就是你周围看到的一切。我们甚至不知道该问什么问题。

他自己想知道的问题列了几个:标准模型对宇宙起源的描述对不对?宇宙真的 138 亿岁吗?宇宙会怎么结束,热寂还是别的路径?我们是不是活在模拟里?

他补了一句,生命是美好的,值得过。

火星能不能活,取决于补给船停了之后

月球基地还在计划里。马斯克的说法是,文明应该往更高的地方走,但去月球的意义在于建一个永久科研基地,而不只是插旗。月球有些部分可能比地球的某些部分更古老,那里做科学,也许能更接近宇宙本质的答案。

再往后是火星,目标是自给自足的城市。

他给了一个非常清晰的验收标准:如果地球的补给船因为任何原因停止发射,火星还能不能继续繁荣?能,就成了。不能,就还没成。

跨过这条线,意识的预期寿命会大幅提升。因为人类不再依赖地球上的每件事都顺顺利利。地球上有自我毁灭的风险,有超级病毒,有小行星,恐龙就是这么没的,化石记录里写着多次大灭绝。

他真正在担心的是一个时间赛跑:文明的弧线还能上升多久,能不能在它开始下降之前,把火星做成自给的。把生命变成多行星物种的窗口,在地球 45 亿年历史上第一次打开。

这也解释了他为什么说首批火星任务没那么重要。重要的是吨位,是能不能往火星送足够多的东西,让它自己转起来。不只是建一座芯片厂,而是要有能力建造这一切的完整工业基础。

时间尺度他给得很具体。如果每个火星转移窗口(每两年一次,行星对齐)运往火星的吨位都指数增长,大概 10 到 15 个窗口,也就是 25 年左右,火星就有可能自给。往宽了说,30 年内能成。

至于火星由谁管、怎么管,他觉得这个问题被高估了。形式不重要,自给才重要。

可以从今天开始做的几件事

  • 看人形机器人公司,别只看演示视频里的走路和翻跟头。问三个问题:手有多少自由度、执行器是不是自研、有没有量产路径。三个都答不上来,这家公司大概率在讲故事。
  • 如果你在制造业或硬件行业,把"执行器、减速器、丝杠、灵巧手"这条链子拉出来看一遍。人形机器人的供应链现在还不存在,谁先卡进去,谁就在未来十年有个位置。
  • 买车或换车时,留意一下车内算力和软件版本。硬件预埋决定了未来几年 OTA 的天花板,V14 这种参数量提升一个数量级的升级,老芯片跑不动。
  • 换手机时留意频段支持。卫星直连的终端大概两年后上市,如果你的换机周期是三年,现在这一台很可能就是过渡机。
  • 碰到重要信息,学 Grok 那套流程自己跑一遍:这条是真的、半真的、假的,还是缺了背景?养成习惯之后,你对绝大多数热搜的免疫力会明显提升。
  • 评估任何算力投入的回报时,用对数思维而不是线性思维。10 倍投入换 2 倍效果,是这类系统的常态,预算规划按这个来,别按线性外推。
  • 如果你关心长期风险,把"冗余"这个词用到自己的生活里。单一收入、单一住所、单一技能,本质上都是没有备份的地球。
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)。

0:00[Music] I believe Optimus is going to be the greatest product ever created by humanity.

0:05Elon Musk and his XAI startup have built the largest and most powerful artificial intelligence training supercomput in the world.

0:15As far as I know, there's only one person in the world who could do that. You know,

0:18this is an arms race of epic proportions.

0:21He's a big thinker. You guys went on Fox the other day with the Doge team. You saw Elon's face nodding while they were speaking with a grin ear to ear. He was proud.

0:32XAI has acquired X in an old stock transaction.

0:36Tesla's first robo taxis are officially on the road.

0:39The company's board proposed a new compensation package for the CEO that could pay him just about a trillion dollars in stock.

0:45He gets nothing if he doesn't hit the numbers. SpaceX will buy wireless spectrum licenses from Echoar for its Starlink satellite network for about 17 billion.

0:583 2 1. [Music] There's a splash down. How do you have time? this I I I've never understand you.

1:20Yeah. Well, I do work a lot.

1:23Ladies and gentlemen, please welcome Elon Musk.

1:28All right. All right. Where are you? Palo Alto.

1:34You're in Palo Alto and not Washington DC.

1:38I'm I'm at Tesla Global Engineering Headquarters in Palo Alto.

1:42Yeah. So, no more Washington DC. You're back at work. You're focused. Yeah.

1:48yeah. I haven't been to DC since May. Okay.

1:52that was a that was a hell of a side quest.

1:55That was a good Any lessons from your time in Washington DC?

2:03the government is basically unfixable.

2:10Elon O only. I quote David's noble efforts and this it's good to it's good to have talented people in the administration. but at the end of the day if you look at our national debt which is insanely high the interest payments exceed the u defense department I guess sorry war department budget. and

2:38Nikki Bryzy, so if AI and robots don't solve our national debt, we're we're toast.

2:47Which is a great segue. Optimus is I think going to be the greatest

2:56product in the history of humanity. What's the progress like? And how much of your how many of your cycles are going specifically to Optimus? What's the timeline? I think you're on version three, maybe four. Tell us everything.

3:14well, yeah, everything would take a long time.

3:19We've got time. we're we're finalizing the design of Optimus version 3 and that that really is going to be a very remarkable robot. it will have the essentially the manual dexterity of a human. So meaning a very complex hand. a be a a an AI mind that can navigate and comprehend reality and it will be made in very high volume. those are the three things that are missing. Like if you see any other robotics company, they're missing those three things. Those are the three really hard things.

4:10and I I I spend actually at this point it it might be more of my mental cycles than anything anything else any other single thing on Optimus. that's that's that that's solving for real world AI all of the electro mechanical issues of Optimus the the supply chain and production challenges of it because we have there is no supply chain that exists for humanoid robots. So it has to be we have to recreate it from scratch and which requires doing a lot of vertical integration.

4:50none of the actuators in Optimus are available from an existing supply chain. so but I I think it is accurate to say that if successful Optimus will be the biggest product ever

5:06and the cost of it at scale 2030 $40,000 a robot. What what do you think the first wave of them will cost? And yeah, when will we be able to buy one to work on the ranch?

5:24I think that the the marginal cost of production once you hit a million units per year is probably around the $20,000 range. it it it sort of depends on how much you spend on the AI chip in the in the robot. and you need to achieve a lot of efficiencies in the actuators. there are 26 actuators per arm like 26

5:55electric like motors, gearboxes and power electronics. so so but but the the the AI chip will be pretty expensive like that that might be like55 or $6,000 of the of the bill of materials, maybe more. and but but so I but I think at volume at a million units a year the the production cost is probably on the order of $20,000 maybe 25 something like that. and price will be as a function of demand.

6:32Elon can you maybe explain to everybody why the hand is so important to get right and why you know the actuator design is so unique and you know why it's so difficult why nobody makes it and why you have to start there almost to build the rest of the the robot properly. Well, it turns out the human hands are incredibly they've evolved to this to be this incredibly sophisticated machine.

6:59Like the your hand is an in actually a remarkable thing. It's look look closely at your hands and and think of all the things you can do with your hands which is a lot.

7:12I can think of many things.

7:14Yeah, I was just thinking about something.

7:17You know, your hands are very versatile instrument. Yeah, you can give him a high five.

7:23Very versatile. you know, you you you can swing a baseball bat, you can thread needles, you you can you put thread in a needle, you can play the piano with violin. you know, you could disassemble or assemble a car. The hands are incredibly versatile instruments. and most of the muscles of of the hand are are actually in the forearm. So your hand is kind of like a like a like it's like a puppet. Like it's mostly a puppet. The mus the muscles are coming from the forearm and they're pulling the tendons which are you know also human tendon designs or human tendon evolution is incredibly good.

8:10so you you've got this web of tendons. You you you've got I think I think the the human hand is something like depending on how you count it, 27 or 28 degrees of freedom per you know in in the hand. It's it's amazing. So in order to create a robot that can be a generalized humanoid, you you must solve the hand the hands problem. Yeah, we had we had Ario.

8:40It's got hands, needs, hands.

8:42And so is it like when you were first building Tesla where the supply chain doesn't exist and now you have to go out and find folks to work with and you know build all this vertical integration, get support. Is it is it literally like it's just nowhere to be found and you're going to have to build all of this stuff up?

8:59Yes, we we we could not actually buy the actuators for any amount of money. they assembly didn't exist even though there are I don't know 10 20,000 electric motors out there of various sizes and shapes we've had to design every electric motor gearbox and of and the controlling electronics from scratch basically from physics first principles

9:20the good news is you've got a lot of experience with factories over the last couple of decades so

9:26how challenging is this versus cybert truck model Y

9:31Model Gigafactory, you know, the Yeah. The Fabra egg known as the Model X. Yeah. Right.

9:42it's hotter than any any of those things. Okay. Yeah.

9:46Much harder. Significantly. Yeah. Starship.

9:49Yes. Well, more No, not Starship's harder. Okay.

9:56So, somewhere between a Model X and a Starship.

9:59Yeah. Is it is the what's harder, the hardware or the software?

10:06Right now, we're struggling with the the final design of the hardware. Like I said, it's really primarily the hand. Not to just just dis dismiss the rest of the robot. The rest of it's also important, but but the hands are the hands inclusive of the forearm are a majority of the engineering difficulty of the entire robot.

10:26And then let's assume you get past the hardware challenges. How much do you sort of get for free based on all the progress that's happening with LLMs? Will, you know, will consumers just be able to interact with this, talk to the robot, ask to do things, it'll understand and sort of Oh, yeah. Yeah, no problem.

10:45You're spending a lot of time with Annie, I noticed online.

10:48Not not that long. maybe I went a little over the top from Bunning Grom Imagine, but

10:54well, but in all seriousness, those characters and these robots that seems to be, you know, like maybe they

11:03you could get the embodiment of Annie, I suppose.

11:05Yeah. Why Why the human form factor, Elon? You could make something that's maybe better than a human or maybe simpler than a human to do specific tasks and maybe better than a human to do more things than a human can do. How did you decide to make it just like a human?

11:21Well, if you wanted to do all the things that a human can do, it turns out you need a humanoid robot. so if you wanted to do a subset, it that's much easier. but it turns out humans evolved to this the shape and capabilities that we we we have. it it it for for good reasons. there actually is that there is a like there's value to having quite you know four fingers and the thumb. and even the pinky actually is is quite useful. toes are much more question mark but but but the fingers

12:06well also humans humans have designed the world as well. So we designed it for us. So if you can make a humanoid robot, it'll be immediately backwards compatible with what we've built the world for. Precisely,

12:17Elon, there's another there's another part of the robot.

12:21So there's the LLMs, there's the actuation in the hands, but also there's the the silicon that runs it. And there was, you know, Dojo, I think you you posted on X AI5 and AI6. And it just seemed like you were incredibly excited about the direction in which the silicon layer was also going. Can you tell us about that and what that is and what what what what are we what are we building here? What is being built? Is it a complement to everything that exists in the world? Is it a potential long-term competitor? What is it?

12:59yeah. So at at Tesla, we basically had two different chip programs. one dojo and one dojo on the training side and then what we call you know AI 4 it's just our inference chip that the AI4 is currently shipping in all vehicles and we're finaliz finalizing the design of AI5 which will be an immense jump from AI4 by some metrics the improvement in AI5 will be 40 times better than AI4

13:36wow So 40% 40 times and and this is because we work so closely at a very fine grade level on the AI software and the AI hardware. So we know exactly where the limiting factors are and and and so effectively the AI hardware and software teams are co-designing the chip.

13:59so a 40x improvement in the silicon I think then as it as everybody here in the audience experiences it is that just an almost like an order of magnitude increase in the quality of FST and the safety that you experience as a Tesla driver and then the quality of the robot like where does it all manifest when you when you you know bring it up and actually get it into production?

14:23Yeah, to be precise the 40x is on if you say like compared to the worst limitation on AI4 which is running the softmax operation. Yeah,

14:34we currently have to run softmax in around 40 steps in emulation mode whereas that'll be just be done in a few steps natively in AI5. AI5 trip will also be u easily handle mixed precision u models. So you don't have it'll dynamically handle mixed precision. There's a bunch of sort of technical stuff that AI will do a lot better. in terms of of nominal sort of raw compute, it's it's eight times more compute. about nine times more memory roughly five times more memory bandwidth. so but because we're

15:16addressing some core limitations in AI4, you multiply that by that that 8x computer improvement by another 5x improvement because of of optimization at a at a at a very fine grain silicon level of things that are currently suboptimal in AI4. That's where you get the 40x improvement.

15:33You had keep going, keep going. so now now that said I I'm I am confident that the current chips AI AI4 chips that are in the cars will achieve self-driving safety that is at least two to three times that of of human and and maybe even 10x.

15:54and the software that will be released for that is is coming out over the next few months. So version 14 will be the biggest upgrade in Tesla software since version 12. we are increasing the parameter count by an order of magnitude. the there's there's there's a lot of reinforcement learning that's been used.

16:20there's we we there there there's like you can think of AI sort of as a way of compressing reality and and and some of those compression steps we were too lossy and and we addressed the lossiness in the compression steps. so the these are all software updates that'll that'll go out. So just over there updates your car is going to feel like it is sentient by the end of the year.

16:50Yeah, it feels that way already to be honest. I saw in the trades that you spent about $17 billion on some spectrum and that

17:00yeah so some couch change to enable your satellites and the Starlink network to connect directly with phones. What will that look like in a year or two? Are we going to drop our Verizon account and just expand our Starlink account? thank you.

17:21We're kind of hoping cuz Verizon kind of sucks. How

17:25How many of you want a Starlink phone?

17:27Who wants a Starling phone?

17:31Is it Is it technically possible?

17:33I know you can't see it, but it's everyone.

17:35Yeah. All right, cool. so this is a kind of a long-term thing. it it it will allow SpaceX to h deliver high bandwidth connectivity directly from the satellites to the phones. but there are hardware changes that need to happen in the phone. So the since these frequencies are not supported in current phones that the chipset has to be modified to add these frequencies.

18:09and that probably is a two-year time frame. So, the phones that are able to use the spectrum that was acquired probably start shipping in around two years. and and then we also need to build the satellites that are going to communicate on those frequencies. So, in parallel, we're building the satellites and working with the handset makers to add these frequencies to the phones. and then the the satellites and the phones will then handshake very well to achieve high bandwidth connectivity. But the net effect is that you should be able to watch videos anywhere on your phone. Wow.

18:47And it's going to be crazy.

18:48And what and do these do these frequencies would they work indoors inside buildings, you know, like like your phone currently does? Okay.

18:56And so will you be able to have basically

18:58if you if if you're in a building with a with a like a a thick metal roof, then no. But

19:03no, the same the same types of of

19:05Yeah. Normal normal homes. Yes. Yes.

19:08Elon, is your vision for this that instead of, you know, having an AT&T account or and then roaming when you're in the UK or you're in India, it's just we could have one direct deal with Starlink, it works all over the world eventually, not today, but at some point. Is that the end goal that basically we don't need a regional carrier, we have a global carrier, and that would be you.

19:30that that would be one of the options. To be clear, we're not going to put the other carriers out of business. They're still going to be around because they they own a lot of Spectrum. So, there's but but yes, you you should be able to have a Starlink like you have like you have an AT&T or T-Mobile or Verizon or whatever, you should be you could have a you know account with Starlink that works with your you know Starlink antenna at home for

19:58free Wi-Fi as well as on your phone and yeah it would be a comprehensive solution for high bandwidth at home and for high bandwidth direct to sell.

20:08Could you buy some carriers to have more spectrum?

20:12Maybe you could buy Verizon.

20:14Not out of the question. I suppose it that may happen.

20:17Let's talk about let's talk about Starship. You just had a really what appeared to be a phenomenal launch. H how close is it to, you know, being predictable and ready to go in a commercial setting? I I I think we'll recover the ship next year. we've got one more launch of the Starlink version two stack,

20:49but there's only one one booster and ship left. That's in the version two design. and then thereafter it's it's version three, which is a gigantic upgrade cuz that's got Raptor 3. and pretty much everything changes on the rocket with version 3. so version 3, you know, might have some initial teething pains, because it's such a radical redesign.

21:15but, it's it's capable of over 100 tons to orbit fully reusable. and I think it's I think I think unless we have unless we have some very major setbacks, SpaceX will demonstrate full reusability next year catching both the booster and the ship and being able to deliver over 100 tons to a useful orbit.

21:42What is the best rocket in the world do now in terms of tonnage to space? well in terms of sort of commercial rockets there's there's Falcon Heavy. Yeah.

21:56which will do in with with side booster reuse will do about 40 tons.

22:07So this is five times bigger. Yeah.

22:11Well two and a half times bigger in but but Starship would be full reuse full reusability.

22:16Got it. Okay. So everything comes back

22:18Elon after after the explosion that happened with the the the the failed launch there was a lot of sorry which which failed

22:30oh the more recent one the more recent the Starship the big boom yeah

22:34the big boom on the base and and and there was a lot of

22:37there was a lot of proclamations that there's going to be environmental and FAA and all these other sorts the recovery back to the launchpad again was incredible. fast. How did you get back so fast? Not just technically and work-wise, but just like regulatory clearance-wise because they said there were going to be all these questions and reviews and so on. How did how did you guys manage that?

23:01well, there were a lot of questions and reviews. We got through them all. and credit to the SpaceX team. They worked incredibly hard and they got the next ship and booster tested and on the pad and and flown and yeah, huge credit to the SpaceX team. Very proud of them for

23:19doing doing such a job, a great job recovering.

23:23I mean creating a fully reusable orbital rocket is one of the hottest engineering problems ever. It's certainly, you know, a candidate for most difficult engineering project ever. You know, it's on the podium at least. so it's a that that's been the goal of SpaceX from the beginning from 2002.

23:48and here we are 23 years later. So, it's it's a long journey and with with a a super talent like by far the I think the most talented group of rocket engineers that ever been assembled. and and we're finally next year I think we'll be able to achieve full reusability.

24:10Elon, what are the big technical blockers that you're focused on there between now and that full reusability? Are there some showstoppers where you're just kind of literally just obsessing over trying to figure out still or is it more about getting through a laundry list of your learnings and just integrating it into the next launch?

24:33Well, that the for for full reusability of the ship, there's still a lot of work that remains on the heat shield. So, no one's ever made a fully reusable orbital heat shield. Like the shuttle heat shield had to go through nine months of repair after every flight, right?

24:53so, no one has ever made a fully reusable orbital heat shield.

24:58And is that a material science problem or is that an engineering problem or both?

25:03yeah, I mean it's a material science engineering problem. So, it's but we really are looking at the fundamental physics here. again physics first principles and trying to figure out how do we make something that is you know can can withstand the heat is very light doesn't transmit the heat to the the primary sh Yeah.

25:31primary structure and whose integ

25:42and then as you ascend, if you hit some rain, you know, the tiles don't dissolve in rain. There's there's a lot of different issues and and then you really need to know that these tiles are working. You can't you know, go through this laborious inspection. So, it really needs to be we're, you know, these these tens of thousands of tiles all work and don't

26:09need to be refurbished or checked one by one as was the case with the shuttle.

26:15Can we maybe switch now? It's I mean who who else were you talked about Tesla then you go to SpaceX. Yeah. Now I' I'd like to ask you some questions about Grock and XAI.

26:28you want to just give us an update? I think you you kind of talked about where the nextG model is and you said something incredible. I still don't think people really understand it which is you know there's going to be a next training run where you expect you know not to start from the you know common web and common crawl where you expected an enormous amount of synthetic data.

26:49Just tell us about how the evolution of Grock is going and this innovation and why it's so important. Yeah. So we're we're running a lot of using a lot of of inference compute and and reasoning to look at all of the source data which is really the corpus of human knowledge and then thinking about each piece of information and then adding mod adding what's missing and correcting correcting mistakes and removing falsehoods from the from that training data. So it's it's it's like if you take say the Wikipedia as an example, but this really applies to to books, PDFs, the websites, every form of information. the

27:41the Grock is using heavy amounts of inference compute to say to look at as an example a Wikipedia page and say what is true, partially true or false or missing in this page. Now rewrite the page to in to correct the remove the falsehoods correct the halftruths and add the missing context.

28:13Well, Elon, by the way, could you just publish that? Could we create like a groipedia? I mean, that would

28:17Yeah, especially for our bio pages, which are a disaster.

28:20Yeah, Wikipedia is so biased and it's it's a constant war. you know, if something gets corrected, five minutes later, there'll be an army of people trying to

28:29I mean, it's become hyperartisan and there's activists all over it.

28:35So, if you do fix, for example, Wikipedia as a source of truth,

28:39it'd be great to publish that just so the world has it.

28:44All right, I'll talk talk about that. So, talk to the team about that like Grapedia or whatever. This here's the Groedia version.

28:53Be interesting. Yeah. and then just have it out there for just a few minutes

28:56where in terms of people here like it in terms of training Gro 5 you're you're scaling up your supercluster in Colossus and in Memphis

29:08can yeah we have a second one

29:11yeah can could you give us an update on that and then also as part of that where are we in the scaling laws if you scale a bigger cluster do you get a more powerful AI model is there a point of dimin diminishing returns or like how much more compute if you throw twice as much compute at it do you get a 10% better model do you get 100% better model like is it log linear what what I guess how much more juice is there left in scaling hardware do you think

29:42I think I think there's a natural logarithmic function associated with the amount of compute so then like say for argument sake like 10x more compute will double the intelligence. Maybe that's that that might be a rough rule of thumb, but you know, that still means that, you know, you go from 100 IQ to 200 IQ. Still pretty pretty big deal.

30:06so I and and I think we'll see intelligence continue to scale all the way up to where, you know, most of the power of the sun is harnessed for compute and then ultimately most of the power of the galaxy, you know, sort of cautev 2, cautev 3 scale compute. so I guess once you think of artificial intelligence not as sort of this you know a destination that you reach but really as part of the overall escalation of intelligence that that that we are are aware of.

30:46you know human intelligence has also scaled as you've have as the population has increased and we've been able to store more and more information. human intelligence has scaled. Now human because of population declines and low growth rate, human intelligence is is somewhat plateauing and will actually decline. And I my guess is that

31:11I I I I think that we might have AI smarter than any single human at anything as soon as next year. Wow.

31:21yeah. and and and then and then probably within five like say 2030 probably AI is smarter than the sum of all humans.

31:29Do you think do you think humans are on the decline because the AI is evolving? Do you think there's this evolution of the ecosystem on Earth that's underway that we don't really understand the structure of what's going on? But

31:46maybe yeah, maybe we implicitly know that it's coming. yeah.

32:00I I I I mean I hope the birth rates turn around. I'm a I'm a big proponent of increased birth rate. obviously

32:09Well, you doing anything about it or no?

32:12Yeah. I'm trying to set a good example. We, you know, we had a big conversation at this conference we didn't expect, which is suicidal empathy, the West,

32:24this declining birth rate. I noticed you've been pretty active about it and open borders

32:30and open borders is like let the invaders in. Could all three of those be the same thing? It

32:35it seems like there's a number of symptoms of the West being suicidal. The most obvious one being the birth rate is not a replacement level. So obviously if that continues indefinitely then the west will literally not reproduce enough to replace itself. But there's other things too. There's the fact that the borders were totally opened to the point where western culture the social fabric start to come apart. And you see this especially in Europe where they're you know the indigenous cultures of the UK or France or Germany are starting to potentially be taken over by by cultures of people who are brought in and aren't assimilating. You have crime

33:13where, you know, we have this case on social media right now, this young woman, I Ina, who's just

33:19killed in a senseless way on a subway.

33:22which is horrific enough in and of itself. But then in addition to that, the elite media just for whatever reason just refused to cover it, like it didn't exist.

33:30so you have this issue of crime that's not being addressed or even acknowledged

33:35and no acknowledgement of this. like it's almost like we're trying to deny the reality of the spiral and this Yeah.

33:44So you have the you have all these data points that seem to suggest that the west is suicidal or doesn't you know doesn't seem to want to defend itself or propagate itself. look I think everyone in this room thinks that life is awesome right? I mean pretty and I think worth living.

34:04Yeah. And when when Alex Karp was here earlier today defending the West, that got some of the loudest applause at the conference. So, I guess we probably don't really understand what's going on. We don't really

34:15Yeah. What's your take, Elon? Cuz you you know,

34:17what's your take on the suicide of the West? Yeah. What's what's

34:21I'm very worried about it. Yeah.

34:24I'm very worried about it. you know, I think there's there let's just say that the actions of the West are indistinguishable from suicide. So, but it's and look, at least in America, there's there's there's generally a sense of optimism, but when's the last time you you talked to someone from Europe who lives in Europe who's optimistic? Not for a while. Yeah. Decades, like even one.

34:55It's rare. So I I think unless people have a sense of optimism and purpose about the future, they suicide might be just what happens. like like like having a child is an act of optimism about the future. So if you're not optimistic, this Yeah. So, so I think we need to maybe give people a sense of optimism and excitement about the future and and a belief that the future will be better than the past. and they'll be more interested in having kids.

35:33Did did religion play a role in the past, Elon, to kind of plate and make folks feel that way when they

35:40Yeah, I think so. nature abhores a vacuum and if you take away religion then I think you actually you you you get something in it place which is actually worse than what was there before I mean it's like destructive basically you get you get like the white work mind virus filling filling the hole that religion used to have taking the place of of of religion

36:08you get these dystopian de facto religions that that that are very very self-destructive. so I I think perhaps some some sort of re revival of religion or at least what we need is is some coherent philosophy that people can get excited about. you know I mean for me it's a philosophy of curiosity. I'm curious about the nature of the universe and I want to go out there and I want humanity to be out there exploring the stars. maybe meeting alien civilizations. maybe in some cases we we see the ruins of a long dead alien civilization but they were they were very strong for 10 million years. you know the kind of stuff that you see in Star Trek in in a

36:59non-dystopian sci-fi book or or movie or show. and so I'm I'm just I have I have a philosophy of curiosity of of like I just want to know what's going on. And and in order to know what's going on, we we must have u an an increase in the in the scope and scale of consciousness, we must we must expand consciousness. We must grow. We must grow humanity and we must extend humanity in order to comprehend the and to to understand the universe or even what what question should we should ask about the answer that is the universe. you know Doug Douglas Adams book the hitchhiker's guide to the galaxy is actually a a deep book on philosophy disguised as humor. and

37:46what the point he was trying to make in that book was that the questions are the really the hard part. The answer is the universe. Like the answer is everything you see around you. But but but one of the questions that we don't know to ask. Yeah.

38:00now now some of the questions I guess I we I do know I'd like to know is the standard model of physics correct about the origins of the universe? Are we actually 13.8 billion years old? how does the universe end? Does it end in a heat death or in some other way? you know, a black hole. We might be. Elon, can you talk about

38:23the whole sort of simulation question? Are we a simulation? Maybe.

38:26Where does the where do you think we find the answer first? In AI or in the stars because you're pursuing both obviously.

38:38Yeah. I I I don't know if if I I hope I hope more people can get behind a philosophy of curiosity. Yeah.

38:50Because I think it's very exciting. Yeah.

38:54and and and and inherently optimistic. you like because there's there's this amazing sense of wonder about the nature of the universe. And when you just when you uncover some secret to the universe, that's amazing. And you're like a whole world of understanding is opened up. I mean, we we used to not even know where all the continents were. you know, used to be like just the map would be there be dragons and like all we know is that when they sailed in that direction, they didn't come back. I mean, the moon base,

39:32that's all that's all they knew. I I kind of feel like the moon base or just going to the moon for real this time would be a big step in the right direction. You still have the moon planned. What's the status of that? Is is that still on the agenda?

39:46Yeah, I I think it I think having I think we want to try to reach new heights as a civilization. Yeah.

39:53So, I think it's it's fine to go to the moon, but but we should go to the moon in order to establish a lunar base, like a a lunar research base.

40:01Yeah. I mean there are parts of the moon that are perhaps older than parts of of Earth. and we we we might understand more about the nature of the universe if we had a science base on the moon.

40:13that would be very cool. And then we we obviously want to go beyond the moon to Mars and build a self-sustaining city on Mars. the I I I do think that that that there is a fork in the road of human destiny where if we can establish a self- sustaining city on Mars with the the key test being if the resupply ships from Earth stop coming for any reason does Mars continue to to prosper or does it die out

40:43at the point at which Mars is able to prosper and grow on its own the probable lifespan of consciousness is dramatically greater. because we are no longer dependent on everything going right on Earth. You know, there's there's always some possibility of self annihilation on Earth with the World War II or or a super virus or or or a meteor like

41:07extin, you know, that destroyed the dinosaurs. We know from the fossil record that there've been many mass mass extinction events. So the question that I sort of was wondering about is will civilization will the civilizational arc continue to ascend such that we can make Mars self- sustaining before the civil civilizational arc descends

41:30because the the the window of opportunity to make life multilanetary exists now for the first time in the 4 12 billion year history of Earth.

41:40Yeah. Elon, let's assume that we get there and you're there.

41:46you know, you'd be the elder statesman. You'd have the moral authority of Mars. How do you run Mars?

41:57But I there's this point that I I think I I want to just emphasize again that that's that's more important than the form of governance on Mars or who's there in the early days. What really matters is that Mars is self- sustaining that we are truly a multilanet species and s such that we've achieved planetary redundancy so that that if if something and obviously we should do everything possible to make sure life on Earth is great but there's always some risk that of an annihilation event on Earth. Yeah.

42:32like I said self annihilation or some natural disaster. and and so the the probable lifespan of consciousness increases dramatically as soon as as soon as we are multilanet species with the key test being can Mars survive if the resupply ships stop coming. So so getting like the first missions to Mars are not that important. The what matters is can you get sufficient tonnage tonnage to Mars such that Mars can prosper on its own.

43:04and that means it has to have all of the ingredients of civilization. It it it's not just that you need to build, for example, a chip factory on Mars or ship fab on Mars, but you you need the ability to build.

43:16Do you do you have a sense of the time scale? Like, let's assume Starship is at a state starting in, you know, 2026. Then there's going to be a bunch of testing. Obviously, there's going to be a bunch of early testing. We only have certain launch windows. So, there's a bunch of time constraints. Is that is this a 50-year thing in your mind? Is it a 150 year thing? Is it something that is for our generation or is it our children's generation? Where do you see that point if it's optimally possible?

43:43You know, if things go and break our way,

43:47I think it can be done in in 30 years.

43:51wow. So if provided there's an exponential increase in the in the tonnage to Mars with each successive Mars transfer window, which is every two years. So every two years the the planets align and you can you can transfer to Mars. so I think in roughly 15, but maybe as few as 10, but some 10 to 15ish Mars transfer windows.

44:18If you're seeing exponential increases in the tonnage to Mars with each Mars transfer window, then it should be possible to make Mars self-sustaining in in about call it roughly 25 years.

44:33Amazing. That's incredible. All right, ladies and gentlemen, Elon Musk, we'll see you when we're back in town. We miss you. We'll

44:41see you in person next time. Thank you, brother. All right. That's

资源与出处

相关视频

同频道、同主题或相邻年份