罗恩·巴伦(Ron Baron)抛出的问题很实在:如果特斯拉明年要产一百万台 Optimus,后年一千万台,再往后一亿台、十亿台,这些机器人走在人行道上,地方够吗?要不要专门修点什么?
马斯克的回答听上去像段子,数字却是真的:全世界 80 亿人,可以全部站进纽约市一层楼的地面里。人类占的空间比直觉里小得多。他补了一句,如果你坐飞机横跨美国,想往下丢一个水球砸中人,大概率砸不到,因为地面是空的。
这个开场定下了整场访谈的调子。马斯克谈机器人、谈 Grok、谈芯片、谈自动驾驶,反复用的是同一套动作:先算物理上够不够,再算成本能不能降下来,最后问一句,为什么别人不做。
顺带说一句,巴伦开场还提到另一件事。SpaceX 已经完成大约 5000 次发射,在轨卫星 9000 颗,正朝 15000 颗推进。贝索斯前一天刚刚完成火箭回收的首次成功,而帮他做成这件事的团队,很多是从 SpaceX 出去的,因为在那里跑得不够快。
Optimus 的账本:一百万台之后,成本两到三万美元
巴伦追问的核心是价格。现在工业机器人,比如日本川崎那些用在产线上的家伙,一台要 5 万到 15 万美元。马斯克描述的 Optimus,目标价是 2 万美元级别。
他的估算很具体:当 Optimus 达到年产一百万台的稳定状态,再花大约一年打磨成本,物料加工时成本(COGS)能压到 2 万到 3 万美元,按现在的美元算。他称这个估算相当安全。
为什么敢这么算?特斯拉造车那套逻辑能直接搬过来。买别人的零件,必须清楚对方的成本结构,报价离谱就自己做。这套垂直整合的路子,会用在机器人身上。
机器人有个比车更难的东西:手。
马斯克说,一只手加前臂有 50 个执行器,一台机器人两只手就是 100 个。执行器包含电机、减速器、电力电子,一整套。一只灵巧的手,背后是极高的复杂度。
巴伦问,为什么非要把手做得这么复杂,简单点不行吗?
马斯克的理由是精度。拧螺丝刀、转扳手听起来简单,穿针、弹吉他的难度极高,这些活要求接近人手的分辨率、灵敏度和自由度。真正让他愿意投入的理由,落在医疗上。
他描绘的场景是:Optimus 的精度会达到超人水平,能做的外科手术包括人类医生因为手抖或视野限制根本做不了的那些。到那个时候,每个人都能用上最好的外科医生。
这句话值得停顿。过去几十年,人们一直在说要消灭贫困、要让所有人看得起病,方案却一直拿不出来。钱解决不了这个问题,因为顶尖外科医生的数量有限,医生不会在树上长出来。
马斯克的说法是,这些医生以后会在工厂里被造出来。
机器人抢走工作之后,人靠什么活
巴伦问得很直接:如果机器人什么都能干,人还剩什么?
马斯克没有绕。他举了国际象棋的例子。你手机里不联网的棋类程序,能轻松击败马格努斯·卡尔森(Magnus Carlsen),人类最强的棋手。而国际象棋的受欢迎程度,正处在历史最高点。
机器在某件事上比人强,不代表人就不想再做这件事。体育是这样,棋类是这样,未来很多事也会是这样。
他还给了个日常的比喻:谁不想要一个自己的 R2-D2 或者 C-3PO?一个能帮你遛狗、买菜、教孩子、陪你聊天、关键时刻保护你的伙伴。这类需求是天然的,而且比《星球大战》里的设定还舒服。
至于数量,他把场景分成两块。个人消费端,可能人手一台甚至更多;工业端,用来生产产品的机器人数量大概是人类的 3 到 4 倍。加起来,全球机器人总量可能到 300 亿甚至 400 亿台。
对比一下,现在全球汽车保有量大约 15 亿辆。机器人这个量级,属于另一个数量级的事情。他也承认,造出十亿台机器人需要时间,这个过程本身会相当可持续。
Neuralink 加 Optimus:六百万美元先生,只要六万美元
访谈里有一段特别具体的故事。巴伦几年前给马斯克发过一篇文章,讲一个叫 Barrens 的年轻人,33 岁做到投资组合经理,因为细菌感染失去双腿,之前还是残奥选手。巴伦问,有没有办法让他离开轮椅。
马斯克当时的回答是:有,三到四年,给他一副 Optimus 的身体,再用脑机接口让他像正常人一样行走、奔跑、跳舞。
这次访谈里,他更新了进度。Neuralink 已经有超过 10 位患者植入了设备。这些原本手臂或腿部无法动弹、有的像霍金那样完全锁闭的人,现在可以用几乎和我们对话一样快的速度交流。
技术路径是:植入物读取大脑运动皮层的信号,同时向感觉皮层回传信号,再把这些信号传给外接的 Optimus 机械腿。结果是这个人不但能走,还能跑得比普通人更快。
马斯克拿老电视剧《六百万美元先生》(The Six Million Dollar Man)打比方,然后补了一句:现在这套东西大概六万美元就能做到,连当年那个数字的百分之一都不到。
他顺带提了一句公司的目标已经改了。过去是“加速可持续能源”,这件事基本做成了。新的目标是“可持续的丰裕”(sustainable abundance),让所有人富足,同时不破坏自然荒野。
买下推特,图的是言论自由
巴伦讲了一段两人之间的往事。当年马斯克收购推特,巴伦主动打电话说要投 1 亿美元,其中 6000 万来自基金,4000 万是他自己的钱。马斯克在电话里愣了一秒,说“真的?”巴伦说,你在特斯拉上帮我们赚了 80 亿,不支持你说不过去。
结果钱打过去当天,这笔投资就被减记了 70%。
马斯克在台上笑着回应:
这就是为什么我知道你是真朋友。友谊的考验,是谁在形势最糟、所有人都反对你的时候还站在你这边。(马斯克)
后来对冲基金开始打电话,想按原价把这些股份买走,那些人总能知道些不该知道的事。巴伦没卖。再后来,推特改名 X,业务重组,Grok 上线,算力中心建起来,那 3.5 亿美元的投资变成了 7 亿美元。
巴伦问了一个很多人好奇的问题:你买推特,是不是早就想好了要拿它的数据训练 AI?
马斯克的回答很干脆:没有这个打算。他买推特,是因为觉得它对文明有负面影响,正在推送反文明的观念,被极左翼、被旧金山和伯克利那套政治立场俘获了。平台封禁了很多右派账号,包括一位在任总统,这在美国历史上没有先例。
他的逻辑链条很清晰:言论自由是民主的基石。没有言论自由,选民无法做出知情投票;无法知情投票,就没有真正的民主。所以需要一个真正的公共广场。
他特意强调,收购之后没有任何左翼声音被封禁或压制。目标是给全国各个立场同等的权重,让人们在广场上交换观点,同时尽量避免煽动暴力。
他还说了一句更宏观的话:如果美国不够强大,做生意还有什么意义。美国是支撑西方文明的那根中心支柱,柱子倒了,一切都倒。
OpenAI 的起点:一场生日派对上的“物种歧视者”
这是整场访谈里信息密度最高的部分之一。
马斯克说,他创立 OpenAI 的原因,来自和拉里·佩奇(Larry Page)的一次对话。佩奇当时是他很亲近的朋友,但马斯克觉得佩奇对 AI 的危险不够在意。
矛盾在马斯克的一场生日派对上公开化。佩奇当着很多人的面,说马斯克是“物种歧视者”(speciesist),理由是他偏袒人类而不是计算机。
马斯克在台上的复述是:
我当时就想,拉里,你站哪边的?听起来你站在计算机那边。可你总该站在人类这一边。(马斯克)
这件事之后,他决定必须有一个能和谷歌抗衡的力量。他问自己,谷歌的反面长什么样?答案是一个开源的非营利组织。OpenAI 里“Open”这个词,本意就是开源。
他出了早期的全部资金,A 轮、B 轮、C 轮,招来了伊利亚·苏茨克维(Ilya Sutskever)这些关键人物,把自己知道的都教给了他们,还帮忙牵线,让萨提亚·纳德拉(Satya Nadella)从 Azure 拨出一些资源。
他特别强调,整个过程他没有寻求任何财务回报。后来对方提出给他股份,他拒绝了。理由是:非营利组织不该有股份,也不该成为个人致富的工具,这在道德和法律上都站不住。
巴伦补了一刀:如果你只为了钱,你不会放弃一个今天值 5000 亿美元的东西。
xAI 的三根支柱,和一块别人没有的数据
马斯克对 xAI 的定位很清醒:起步晚了两年半,从后面追,算弱势一方。他认为技术创业最终的赢家,是跑得最快的那个。
他把 AI 公司的成功要素拆成三条:
- 能不能吸引最顶尖的人才。
- 能不能最快把最多的 AI 硬件上线。
- 有没有独特的数据来源。
第二条是他最有底气的地方。xAI 在几个月内建成的 Colossus 数据中心,规模是地球上其他同类设施的四倍,用的是 10 万块 GPU,而别人还在 2.5 万台这个量级上堆 CPU。黄仁勋公开说过,全世界只有一个人能做成这件事。
马斯克在台上接了一句:
我一直跟人说我是外星人,可没人信。不过我绿卡上确实写着“外星人登记卡”,政府给我证明。(马斯克)
第三条是 X。他称 X 是全世界最好的实时数据源。别人在纯数字世界里处理文本和图像,xAI 在做的是把物理世界接进数字世界,动作、视觉、实时事件,这些别人拿不到。
他还说了一句挺有意思的话:他脑子里的想法比能落地的多,所以会走一些“棋盘上没有的棋”,别人预判不到的招。
至于商业模式,他反而不太担心。他的判断是,哪怕一家规模不大的 AI 公司,只要做成了,估值也会很高,因为它对生产率的贡献太大。这个领域会出现很多家稳定在几千亿美元市值的公司。
Grok 5:六万亿参数,10% 的 AGI 概率
马斯克给了一个明确的时间点。Grok 5 会在 2026 年第一季度出来,他认为它会以明显优势成为全球最聪明的 AI,每一项指标都是,没有例外。他也承认自己可能判断错。
更关键的是这句话:
Grok 5 是我第一次觉得,我们实现通用人工智能(AGI)的概率不再是零。概率不高,我自己估算大概 10%。也就是说,仍有 90% 的可能做不到。但这是我第一次这么想。(马斯克)
为什么是 Grok 5?他给了几个技术细节。
参数规模是 6 万亿,而 Grok 3 和 Grok 4 基于 3 万亿参数的模型。更重要的指标是智能密度,他提出两个衡量方式:每 GB 的智能,和每万亿次运算的智能。这是在看模型的效率,单纯堆参数说明不了什么。
训练数据的质量也提到了机器级。Grok 5 原生多模态,文本、图像、视频、音频一体。工具使用能力大幅增强,除了调用现成工具,还能自己造工具来更好地回答问题。
他特别强调实时视频理解:
能理解实时视频,是 AGI 的必要条件。人类当然做得到。做不到这一点的模型,根本谈不上 AGI。(马斯克)
他还提到有些“特殊的东西”不能在公开场合讲,但会让 Grok 5 用起来真的感觉像有意识。
说到现有的 Grok,他也没忘推荐一句。Grok 4 Heavy 会同时派生多个智能体并行工作,彼此比对输出,像学习小组讨论完给一个最终结论,效果一直在变好。
巴伦追问,那意识会不会一直涨?Grok 5 比 4 强,4 比 3 强,那 5 之后是 10,10 之后是一百万?马斯克说,对,会涨得非常远,远到几乎无法理解。
把知识刻进石头:银河百科全书计划
访谈里有一段不太被注意,分量却很重:Grokipedia。
这个项目未来会改名,叫“银河百科全书”(Encyclopedia Galactica),致敬阿西莫夫和道格拉斯·亚当斯,两位都在书里写过这个东西。
它的定位是:一个开源的、全人类知识的蒸馏库。任何人都能访问,任何人都能用,其他机构想拿来训练也可以。
更远的一步是复制。马斯克想把这些知识的副本散布到整个地球,再放到月球、火星,甚至深空。
他的参照物是亚历山大图书馆。那座图书馆被烧毁,是人类文明史上的一场大悲剧。这一次他想把知识刻进石头,做成类似微缩胶片的形式,广泛分布。最坏的情况下,未来的文明还能看到我们学到了什么,从这里接着往前走。
100 吉瓦的太阳能 AI 卫星,相当于美国四分之一电力
话题拉到宇宙尺度时,马斯克给了一个具体到吓人的数字。
xAI 已经有一条路径,能每年把 100 吉瓦的太阳能 AI 卫星送上轨道,这会是超大规模 AI 供电和运行成本最低的方式。
为了让人理解这个量级,他给了参照:美国所有电源加起来的平均电力负荷,大约是 460 吉瓦。也就是说,100 吉瓦相当于全美发电量的四分之一。
巴伦追问,宇宙里到底有没有别的文明。马斯克的回答很坦率:
我的哲学是好奇。我就是想知道这个宇宙在发生什么。标准物理对宇宙起源的描述对不对?热寂是不是终点?有没有其他外星文明?能不能和他们对话?还有哪些关于现实的问题,是我们不知道该问的?(马斯克)
所以他把目标定得很清楚:扩大意识的范围和规模,去理解宇宙。
工厂就是芯片:5 秒下线一台车的物理逻辑
话题回到地面。巴伦提到特斯拉的产线节拍:行业平均 40 到 60 秒一台车,特斯拉现在是 35 秒,马斯克说过要降到 10 秒,甚至可能是 5 秒。
5 秒一台车,听起来像子弹一样往外射。马斯克的解释很朴素:5 秒一台,意味着车以每秒 1 米的速度下线,这就是快步走的速度。车身长度不到 5 米,所以车是走着出来的,你甚至能跑得过它。
他顺手算了一笔账:一周 1 万分钟,7×24 小时运转,每分钟 10 台车,就是一周 10 万台。
方法才是重点。他说某天晚上站在工厂里,突然意识到可以快得多。于是他做了一件事:把工厂切成一个个立方米,算有多少立方米在干有用的事。结果是有用的比例低得惊人,体积效率很差。零件移动的速度也很慢,通常卡在装刹车灯、装座椅这类工位上。
解法是把工厂当芯片来设计。芯片提效靠两件事:把电路靠得更近、做得更小,然后提高时钟频率。工厂同理,提高体积密度,缩短物料移动距离,加快节拍。
他提到一个思维工具,由巴伦的女儿罗宾转述:把自己想象成一个比特,或者一个原子,问它在工厂里走的这条路径合不合理。不合理就改。比特的旅程不合理就改软件,原子的旅程不合理就改产线。
巴伦问了一个更根本的问题:为什么别人不做?福特的高管最近公开夸中国车企做得好,可那些能力是被特斯拉带出来的。为什么大多数公司做不到这种改进?
马斯克的回答只有一句:大多数公司是渐进主义者。管理层想的是比去年好 5% 到 10%,冒可能失败的大险不在他们的选项里。他本人没有这个心理障碍。
AI5 芯片:整个特斯拉押在这一颗上
访谈里最技术、也最能看出马斯克工作状态的部分,是 AI5 芯片。
他直说:整个特斯拉都押在这颗芯片上。它要装进下一代自动驾驶汽车,也是 Optimus 的关键。
而这个项目此前状态很糟。设计目标太激进,进度推不动,看不到成功的路径。另一个 Dojo 项目做得还行,可没法跟英伟达竞争。
他的处理方式是:把两个项目合并成一个,所有人集中做 AI5。训练继续用英伟达,推理必须用自己的芯片。
AI5 的定位是推理芯片,性能功耗比极高,绝对功耗很低。他给出的对比是:在车和机器人的推理场景下,性能至少是英伟达芯片的 2 到 3 倍,成本大概是英伟达芯片的 10%。
巴伦问他周末在干什么。他说,周六,最近周日也是,都在弄芯片。到目前这个阶段,整颗芯片的物理设计已经装在他脑子里,他能把整颗芯片可视化出来。
这句话听着像吹牛,可结合他前面讲的工厂体积效率、把工厂当芯片设计,你会发现这是一套一致的工作方式:先在脑子里建模,再让现实去对齐模型。
为什么可能得自己建晶圆厂
芯片设计完了,得有人造。巴伦问,是不是要找台积电或三星合作。
马斯克先给了很高的评价:台积电和三星都是伟大的公司,特斯拉和 SpaceX 都在用他们的产能,希望他们能多快就多快、能多大规模就多大规模。
问题出在节奏上。他问建一座新晶圆厂从开工到量产要多久,对方回答五年。他的反应是,五年对我来说是一个时代。他的时间尺度是一年、两年,第三年就往无限扩。
所以他的判断是:按这个节奏,芯片供应会成为 Optimus 和自动驾驶汽车产量的硬约束。到那时候只有两个选择,要么自己建一座超大的晶圆厂,要么产量被芯片卡死。
他也承认合作伙伴在拼命扩产。台积电在台湾和亚利桑那,三星在韩国和得州,四个厂区同时在推,从他们的角度看已经是闪电速度。可对他还是不够快。
这里的潜台词很清楚:合作伙伴并没有偷懒,问题是需求增长的速度超过了整个半导体行业习惯的扩产节奏。
FSD:100 亿英里数据,比人类安全四倍
最后回到自动驾驶。巴伦问,FSD 是不是就是让车像人一样开,无需靠代码去穷举每一种情况,比如前面有消防车、车上挂着自行车、有人边开车边遛狗。
马斯克确认了这个理解。真正的功夫在车里的 AI 软件,目标是让它的安全性超过人类司机。
他给了两组数据:
- 现有的 AI4 硬件由特斯拉设计、三星代工,有能力实现无监督全自动驾驶,安全性至少是人类驾驶员的 2 到 3 倍。这里说的无监督,指你在车里睡着了也没问题。
- AI5 硬件预计能把安全性再提升大约 10 倍。
支撑这些结论的是数据规模。到现在累计约 100 亿英里的驾驶数据,在统计上非常明确:开启 FSD 的事故率是人类驾驶的四分之一。
还有个反直觉的细节:很大比例的用户买了 FSD,却从来没试过。所以特斯拉现在出于安全考虑,会主动要求给客户演示一次,让他们知道怎么用、怎么打开。
马斯克在台上直接喊话:去任何一家特斯拉门店,他们都会演示给你看,这件事没有藏着。巴伦在旁边补了一句:你试过之后就会买。
他到底图什么
访谈接近尾声,巴伦问了一个很人性的问题:你现在有几百亿、几千亿美元,接下来打算怎么用?希望别人怎么看你?
马斯克的回答很低调:主要是需要对公司保持足够的所有权,才能继续决定公司的方向。从个人消费角度,他没有度假屋,只在奥斯汀有一套中等大小的房子,在星舰基地还有一栋 8000 美元买的小房子,朋友们来看了都以为他在开玩笑。
然后他讲了一段更长的话。他说在良性的场景下,AI 和机器人会带来一个丰裕的世界,会有一个临界点:AI 和机器人把人类所有需求都满足了,真的没别的事可做。到那时候,AI 大概得开始想它能为自己做什么。
他自己想做的,是扩大意识的边界,让意识的范围和规模大幅增长,去其他星系,去没人去过的地方,看看有没有还活着的外星文明,或者早已消亡的文明,透过他们的遗迹去理解他们曾经是什么样子。
巴伦最后回到那个生日派对的梗:拉里·佩奇说你是物种歧视者,可你确实是那个认为人类应该活下去的人。
马斯克说:
是的,我毫不掩饰地站在人类这一边。(马斯克)
这场访谈信息量很大,真正值得带走的,是他反复使用的那套方法:把问题还原到物理和成本,算清楚一个单位需要多少材料、多少空间、多少时间,然后问一句,现在为什么做不到,卡在哪,能不能换一种做法。
如果你想把这场访谈里的东西用起来,可以试试这几件事:
- 找一个你所在行业里“大家都接受”的节拍或成本,用物理方法重算一遍。像马斯克算工厂体积效率那样,把流程拆成最小单元,看有多少单元在真正创造价值。
- 用“比特的旅程”和“原子的旅程”这两个问题审视你的业务。信息从产生到被使用,中间绕了多少圈;物料或任务从开始到交付,中间走了多少冤枉路。每一段不合理的路径都是一个改进点。
- 区分渐进目标和数量级目标。5% 到 10% 的改进不需要勇气,10 倍的目标才需要重新设计系统。给自己至少一个数量级目标,逼自己换方法,别只加努力。
- 关注瓶颈环节的真实扩产周期。如果你的业务依赖某个上游产能,去问清楚对方从决策到量产要多久。这个时间比你的增长节奏慢,就得提前想替代方案,哪怕代价是自己建。
- 对新技术,先亲自试一次再下判断。FSD 有大量用户买了却从没用过。信息不足常常只是借口,真正的原因是根本没试。找一个你一直怀疑的新工具,这周花半小时真正用一遍。
- 留意“机器比人强但人还在做”的领域。国际象棋是最典型的例子。这类领域往往藏着被低估的机会,价值落在人的参与本身,效率只是其中一部分。
0:06[applause] Thank you everyone. As I said before, this is a lot different than Convention Hall in Aspberry Park. today we're privileged to have Elon Musk join us. virtually and there could couldn't be anyone more appropriate to speak to today when I'm theme of our conference is changing lives. there's no one who's changed our lives more than Elon especially changed our lives financially but changed everyone else's lives and and and that really feels like it's just beginning. Without him there'd be no electric vehicles. there'd be no FASD. Has anyone ever heard of thought of full service drive self-driving five or 10 years ago? Has anyone ever thought
1:04about that? so reflyable rockets, no one has been able to do that before. congratulations to Jeff Bezos for accomplishing it yesterday. First time. we've already done what 5,000 flights and and we have 9,000 satellites I think it is on the way to 15,000. So and the people who did that for him did it for Jeff are people who used to work at for for SpaceX
1:36but couldn't go fast enough for Elon. So now they went over to Bezos. so there would be no Starlink no refus. In fact one of the first questions I wanted to ask is I was telling my assistant about Optimus before and how the plan is to go from a million production next year to 10 million the following year to 100 million to a billion. And what she asked me was so where's the room for them on the sidewalks? Are you going to build something? Is that the idea that we have behind Boring Company to go underground for the robots to walk around or fly above? Where's the room for them?
2:21Well, actually, [clears throat] you can you can fit all of humanity on floor in the city. That's how small the humans are. and and octopus doesn't mind being packed densely. So, it's always helpful to think how how much room do people take and and and this is why I think just having it in the back of your mind that all 8 billion people on Earth can fit on one floor in the city of New York.
3:08So play another way if you met is is if you're flying across the country and your goal is to drop a water balloon on someone below you. You will fail [laughter] because it's empty.
3:25We we used to do that when we were in eighth grade.
3:28We got them though. [laughter] yeah.
3:32So So Kaitlin, don't worry. There's enough room.
3:35Plenty room. So, one of the things I think about when you talk about Optimus is that there's so many functions that it can perform. What's going to be left for humans? Is there a job that humans going to have other than just living in you know this this great abundance that that you're going to create? What what's going to happen?
3:57Well, I think there is this question of oh, how do you drive meaningful life if the robots can do everything? But we see lots of examples where even though machines can be much better, humans still enjoy like athlet athletics for example or let's take a a mental sport like chess u the computers are so good that your
4:34phone not even connected to the internet can beat Magnus calls. easily. But yet chess is at alltime highs in popularity. So really machines being better us at something doesn't mean we derive that connection and doing it.
5:02By the way guys, the connection isn't that good with a phone. Is there some way we can do something else? bit echoey.
5:11We're not The connection isn't that good. I think you keep cutting in and out.
5:14And I What if I talk to this? Talk.
5:20Okay. Okay. You do that. You do that and I'll do this. [laughter]
5:30so, so the function the function get like that close to look at my noles. the the the functions that you see these robots doing is what what are they going to do?
5:43And why is how are we going to have a billion of them when we have 8 billion people?
5:47How look it's going to take us a minute to make a billion robots. So, you know, that'll be very sustainable. but but I think we will there will ultimately be billions billions of humanoid robots on it. a way to think of it is, who on earth not want their own personal R2-D2 C3PO?
6:20like pretty much everyone was, you know, easy street, but even better. like a personal helper buddy robot. It would be great. you could teach your teach your kids, take your dog for a walk, get the groceries, you know, chat, you know, protect you when when needed. Great. and then how many robots would there be in industry providing products?
6:54Probably three or four to one relative to humans which which suggests that total number of robots will be somewhere around maybe as high as 40 billion 40 billion maybe maybe 30 billion robots. It's a lot. So, so, the Japanese company I think is Kukov, those robots that are used in manufacturing are 50,000, 100,000, 150,000. And you're describing a robot that's $20,000. Do we have to have a million a year to do that or 10 million a year to get to $20,000?
7:31And is that something that's going to be affordable for people? Are they going to be rented? They're going to be purchased corporations, are we going to get some kind of carried interest once they buy them from us? How is this going to work or don't we have the model yet?
7:45well my rough guess is that the cost of goods sold their labor and materials for Optimus after we reach a million units of steady state production. So call it a year after reaching a million units a year because it takes a lot of effort to improve the cost. but at that point I would expect the the bill the labor materials to be 20 to $30,000 in current year dollars.
8:20I think that's that's a pretty I think that's a pretty safe estimate
8:23when you're improving costs. So with cars, your idea is that everything we buy from other people to use in our cars, we know exactly how much it costs and therefore we can tell someone how much we're going to pay for what we're buying from them. And if it doesn't if they're making too much then we make that stuff ourselves. Is that the same kind of idea we have in this robots where we're going to it should be much simpler to make than a car or am I wrong because of the hand?
8:52The hand is the hand is extremely complex. there are 50 actuators in the hand and forearm. Actuator is the motor.
9:05Yeah. actuator is the motor, gearbox, and power electronics. so that's a 100 per robot. really a lot of blood actuators and sensors, right? Am I getting approximately? so it's it's it's there's a lot of complexity.
9:32why is that important? Why is it important that we have such a complex hand? so in order to do dextrous tests you you have to have a hand with the the sensitivity, precision and degrees of freedom of a human hand because you know the so something that is we find easy to do like pick up a screwdriver or turn a wrench or even say thread a needle or play the guitar actually require a lot of dexterity.
10:13and one of the one of the paths but one of the reasons we we think we achieve sustainable abundance which is sort of the that new sort of the revised version of the company's goal because it was accelerate sustainable energy which as you mentioned we've we've done that. our new goal is sustainable abundance. So that's abundance for all.
10:42and but but in a way that is sustainable that does not destroy any of the natural wilderness.
10:48How do we decide who gets what? Somebody wants to buy my house, they can just come in and start living there.
10:56Well, I'm not sure why. I mean, you do have a nice house, so [laughter] I can certainly see see the appeal. but the robots will be able to make anyone a house and you know as long as you don't insist on it being in a particular location you can have Roberts will be able to build you a castle if you want. So and but the the reason for pant dexterity is you want to be able to do like surgery and precision medical actions and so
11:36imagine a world where everyone has access to the best surgeons, literally everyone. And Optimus will have the level of precision that is frankly superhuman and will be able to do medical procedures of very sophisticated medical procedures any any medical procedure perhaps things that that really humans can't even do because they're too they're too difficult. and that will be available to anyone.
12:10People often talk about eliminating poverty and providing great medical care, but they never they never actually have a solution. and and money doesn't solve it because there are only so many there's a very limited number of of great doctors and surgeons. they don't grow on trees. but now they they will get built in factories. So, so, I sent you a year or two ago an article about a young man who was in an interview in Barrens and he was 33 at the time and he'd become a portfolio manager and he lost his legs to a it
12:47was a parolympic performer and he lost his legs to maneing bacteria and he said, "Is there anything we can do to get him out of a wheelchair?" And you said, "Yes, there is. in three or four years we can give him an optimist body and then we can use you know transistors in his head in his brain to let him function as a normal person dance and sing and walk and run. Have we been able to make progress in that area?
13:17Yes. So that's the components of two of my companies. One is being your link and the other being Tesla. so Neurolink is also making good progress. now has I think over 10 patients with Neurolink implants. and and these people who didn't have the ability to move their arms or legs in some cases were completely locked in like like Steve Corking. and they can now communicate I think as quickly almost as quickly as we're communicating right now.
13:52which is very cool and that's that's going to continue to accelerate. so what we can do is use a neural link implant that is taking signals from the motor cortex of the brain and and also receiving signals from this sematensory cortex and then you have someone who's lost their legs optimus legs and so you I mean we're really getting like the $6 million man here I mean from back in the I don't know if you watched that show, but I watched it. I watch I watched it.
14:30Yeah, I thought it was pretty fun. and and we we can actually give someone superhuman cyborg capabilities like the $6 million man, but less than $6 million [laughter]
14:41even in this day and age. I mean, $6 million back then was a fortune. These days, it's like nothing. but but for much less than that I mean like for something that that would be reasonable and affordable you know it might be more like $60,000 type of thing. and and you can take signals from the neurolink the mind that would be transmitting to the legs and transmit those to the attached Optimus robot legs and and he would actually be able to run faster than a human.
15:16just just like $6 million.
15:19You know sounds really exciting. Sounds unbelievably exciting. let's switch to XAI. so three years ago you were here and you had either just purchased or about to purchase you know Twitter which you've renamed X and you were widely criticized for that. And yes I was
15:41right. In fact, you you even mentioned here on stage that you didn't want to have anyone else be angry at you because you had enough people trying to kill you already. Yeah. Yeah. Totally.
15:52Right. So, so you were buying you were buying X and it was $42 billion I think and you were in the process of raising money and I called you up and you hadn't called me to solicit me and I called you up and I said I would like to invest with you $100 million in this and it was $60 million for one of our funds and $40 million for me and and and you said really I said yeah [laughter]
16:19you said really and I said and and you told me that you thought I would make a double. And I said, "Well, I hope so." but to me, it felt like you made us $8 billion in in Tesla, and it'd be not very appropriate if I didn't support this new venture that you were doing. So, we invested and and then the day that we paid the money, we marked it down 70%.
16:46Oh, man. [laughter] This is this is how I know you're a true friend, Aron. because this is a you know I I do regard you as as as a as a true a true and trusted friend. and you know the the test of friendship it [applause]
17:06I got more to the story. Well, [laughter] the test of friendship the test of the test of friendship is is is who supports you
17:16when the chips are down and the times are tough and everyone's against you. That's a real friend and that's you Ron.
17:22Thank you. So, so, so we did invest that 100 marked down to 30 and then about a year later we started getting phone calls from hedge funds who always seem to know things they're not supposed to know. [laughter]
17:37Yeah. Yeah. How do they do that?
17:39Right. And they started saying, "I'd like to buy your stock for what you paid for it." And I said, "I think I'd rather wait and stay in," which I did. And then and then changed the name to X and changed the configuration of business and and then from that you bought Twitter and it came together and a social network along with this and and and all of a sudden we have a business that has incredible data. Did you buy this for the data that no one knew about? But but the data with 600 million people talking with each other. So physical data that no one else has. And then you started Grock which is based on our data. and everyone else doesn't have that. They got digital stuff. And
18:24then when they have Grock then we need more data centers. And you're building those and in a space of a very short time less than months. you built a data center which is four times what as large as anyone else on the planet. 25,000 what other people had in in CPUs and we built it in GPUs 100,000 more powerful and then you know now we
18:53have now we're going to go to hundreds of thou but the bottom line is the investment that we made we put up more money and we have a total of $350 million invested over the past two or three years and now it's worth $700 million. [applause] so everything you touch is like that. It's the most unbelievable thing I've ever seen. So so everyone invest in technology and we investing in in the technology person best engineer on the planet. So thank you very much. So so the vision is the question is did you buy X did you buy Twitter because of the data? Is that did you have all this in your head before you did it?
19:35not really. No, I I just bought Twitter because I thought it was having a negative effect on civilization. and just sort of pushing ideas that were anti-ivilizational. you know, it sort of having been somewhat captured by the far left. I'd say, you know, it's fair to say the radical left. I mean they they wouldn't regard themselves as such but it was captured by a you know group of people whose political beliefs are those of you know deeply San Francisco and Berkeley which is about as you get in America. so
20:20that that that meant it wasn't a good forum for debate because they suspended many people on the right including the president as as you may recall a sitting president which is really unprecedented for president. so I think we need to have a public square where there's true freedom of speech. and freedom of speech is [applause]
20:51Freedom of speech is is the bedrock of democracy. if there's not freedom of speech, people cannot make an informed vote. And and if you cannot make an informed vote, you don't have a real democracy. So that's the purpose of acquiring Twitter was to try to bring it more to the center. there have been no no leftwing voices have been banned or or anything like that or or suppressed but what we try to do is give equal weight to all parts of the country so that they can be a public town square where people can exchange ideas and hopefully not present violence and I think that's that's fundamental to I think it's it's one of Like you said, free speech is the
21:43bedrock of democracy. It's why it's the first amendment because people came from countries where if they could be killed or imprisoned for what they said and in fact this is happening all around the world as we speak even in places like Britain. so that's anyway I I did it because I I
22:09felt like u if it's the civilizational risk had to be addressed and and I mean what if America is not strong then what do businesses matter? the the the America is the central pillar that holds up a western civilization and if that pillar falls everything falls. So, so so you were one of the founders, the two founders of ChachiBT and you know, Open AI and and you had
22:50a disagree and it was founded as a charity and it was your idea that you wanted to make sure
22:56that you know freedom of speech and all the things that you deem important for good lives on our planet were followed. safety and the the other founder what he tried to do and did was accomplished is that he got control even though it was your money and he got control and and you and he said Elon I'd like you to stay and you said I want to go I don't want to be a part of this and and and he offered you some ownership and you said I don't want it and so here you walk away from an ownership of chat GBT.
23:36So you're obviously not doing all this stuff for money. I mean you are, but I mean it's that's not but I mean if you if it were only about money, you would have never left something that's worth 500 billion dollars by itself. And so here you're forming this new entity Grock to accomplish what you wanted Chat
24:00GBT to accomplish. But you think that we have an advantage in this because of the data, because of the compute, because of what? And what are you going to do with this ultimately? You you talk about connecting physical world to digital. What does that mean?
24:16Yeah. Well, just going back to opening eye for a second. the reason I found it opening eye was because I was concerned based on my conversations with Larry Page, who used to be a close friend of mine, that he was not sufficiently concerned about the dangers of AI. this really came to a head
24:39when at my birthday party he in front of a large group of people called me a speciousist for favoring humanity over computers. that I find that troubling. [laughter] I was like Mary what side are you on? it sounds like you're on the side of the computers. but you really need to be on team humanity here, you know. so,
25:11after that I was like, okay, this is it. We got to have some counterbalance to, Google. because Larry doesn't seem to care if humans make it or not. so I thought, what's the opposite of Google? It would be an open-source nonprofit. And that's where the word open in OpenAI comes from. It means open source.
25:33and and then I I provided I provided all the money beginning like what was the series APC rounds and recruited the key people like Ilas Sky and taught them everything I know and you know I actually even got them to deal with Microsoft with with Satcha initially to got I got Satcha to donate some time from Azure.
26:03and for all that I did not seek any financial reward whatsoever. and the reason I actually took down the offer for shares is because I mean I I felt like what do you shares and why like nonprofits supposed to have shares at least last time I checked. you know they're not based nonprofits are not supposed to be a vehicle for self-enrichment. So that's why I turned on the offer of shares because it it didn't seem morally or legally transible.
26:36so then then with with with XAI we and we are starting late with XI you know we're only I don't know two and a half years old basically and we're starting from behind. you know, we are somewhat of an underdog. but you know, pretty good with technology. So, I don't want to pat myself on the back here, but I'm pretty good with technology. and and we are
27:12advancing faster than any other way. So, I think and the the for technology ventures, the winner ultimately is the one that is able to move the fastest. So, so, so we think that so we're optimizing for the best technology and we're doing something different than others. Others in in a digital world and we're physical to digital with movements and visual and other people can't
27:42match that and also we have the real-time data. What does that mean? Why should we do better than everyone else? Why are we going to win? And why are we going to at least be different than everyone else so we have a really strong business?
27:58Well, first of all, I think
28:00in terms of being a strong business, actually, I'm not too worried about that that because even even a small player that is successful in AI will be worth a lot because they will contribute so much in productivity to the economy. So and and so so it's actually pretty easy to achieve a not pretty easy but I mean it's it's not there there will be many companies that are worth sustainably several hundred billion dollars sustainably so then several question
28:37of like well how do we achieve the lead that comes down to three things. are you able to attract the best talent? are you able to bring the most amount of AI hardware online? Can you can you bring GPUs online faster than anyone else? And we've we've already demonstrated that we can do that. Jensen Monk himself said that he was blown away by how fast Xiol the license data center.
29:14Jensen said there's only one human on the planet who could have done that. That's you.
29:19Yes, he did say that. it's very very kind of very
29:27made about are you really human? [laughter]
29:31I keep I keep telling people I'm an alien, but nobody believes me.
29:36I mean, when I got my green card, it said alien registration card. So I mean you know I have have proof from the government. [laughter] I guess I have to get registered. [laughter] so let's say I've got some some like relatively rare skills these days in America in terms of of of getting hardware built.
30:06if you look at the biggest successes in manufacturing in America since World War II by far are Tesla and SpaceX. yeah.
30:21So, so, so to stay on Grock for another minute. So the idea of connecting physical and digital is that that's different than digital digital figuring out who wants to buy what. We're doing something entirely different. Is that fair?
30:41I wouldn't say we're doing something entirely different. We're doing we're doing some things that are same, some things that are different. but it's just saying like the the elements that define success for an AI company are going to be one, the talent to the hardware. How much AI hardware can you bring to bear? that that's actually a very big deal. and we've shown that we're the best at doing that at XAI.
31:07And then third, unique access to data. And for that, we've got the the X system, formerly the Twitter system, which is the by far the best source of real-time data in the world. So that that those are some pretty significant assets. and and I think we're going to come up with some very innovative ideas.
31:31but I have more ideas in my head that I know what to do with, frankly. so I think we'll make some moves that are not on the chess board. that people don't anticipate. some creative moves. and and and like I said, so so and I should point out Grock right now actually Debbie is still the smartest AI best my knowledge. It's it's I recommend trying it out. Grock 4 heavy is where we spawn several agents. They they they work in parallel and they compare their output like a study group and give you the final conclusion. And it keeps getting better. and now we've begun training on Grog 5. Grog 5, I think,
32:20will be the smartest AI in the world by a significant margin on every on every metric without exception. I might be wrong but I I think that will be the case and and that will be in Q1 sometime. Gro five.
32:39Yes. And Grock 5 is the first time where I thought well we have a nonzero chance of achieving artificial general intelligence. not that it's a high chance. I I I sort of I calculate like 10%. That's what my biological neural network comes up with, which still means 90% chance that we don't.
33:03but I've never thought that before. And so for the first time, I think like, wow, this this this really could be general intelligence, at least a small chance. Brock 5 will really be something special. and and it'll be both extremely intelligence extremely intelligent and extremely fast.
33:28so so one of the things that we're doing that I think is interesting is Groipedia. which we're going to rename down the road to incite to to be encyclopedia galactic in honor of Isaac Azimov and Douglas Adams who both mentioned that in the books and the idea behind Galactica is to
33:57create an open-sourced repository of all knowledge like a distillation of knowledge and and open source meaning anyone can access it, anyone can use it and if if other people want to train on it, they can do so. and then we want to create copies of this and distribute these copies throughout Earth. and even put them on the moon and Mars and out in deep space. as
34:29in a way sort of a a modern day Library of Alexandria. You know, it was the great tragedy that the Library of Alexandria burned down or or was burnt down. and so in order to preserve this knowledge I think we we we want to literally etch it in stone and sort of stone stone like micrfluge and and and distribute it widely. So in worst case scenario future civilization can can see what we what what we learned and we pick things up from there.
35:02So is there a major breakthrough that you can describe that allows us to do this with GR 5? Is it just speed? Is it more compute and therefore we have more analys more you know information we can train on what is the breakthrough that allows us to to to have this 10% chance for AGI is there is there a breakthrough or is just speed and access to data not just but
35:32so there's a couple things it will be the largest model to best man knowledge So this is this is a six trillion parameter model. whereas Grock 3 and four are based on a three trillion parameter model. moreover the six trillion parameters will have a much higher intelligence density per gigabyte than rock for.
36:03I think this is an important metric to think about intelligence per gigabyte and intelligence per trillion operations. so we've learned a lot. so the quality of the data that we're training on with GR 5 is machine. it's also inherently multimodal. So it's text, pictures, video, audio.
36:33it's it's going to be much better at tool use and in fact creating tools to be more effective at answering questions and and understanding the its vision will be extremely good. it'll have real be able to understand real time video which is I think a really fundamentally important thing that only other AI can understand real time video and I think if you can't do that which is humans can obviously do you're you really can't achieve AGI
37:13by the way every one of these
37:14there's some special source items that I that there's some special source items that I I can't talk about in a public forum obviously you I can't give away, you know, all all the secrets here and just between us, you know. but but but we have a few a few other special things that are that are in the works for rock 5. So it's it's
37:39it's it's really going to feel sentient.
37:46So but there is no when we rock five and you're talking about the advances There is no limit. So when we're Grock five is better than Grock four which is better than Grock three. So so it keeps going. So once we get to sentient levels we go two sentient 5 sentient 10 sentient a million. Correct.
38:08The the sentience will grow. I mean what's really mind-blowing is can how far can the sentience grow to your to your point? How far does it go? I think it goes immensely far. Almost almost incomprehensively far and almost does go incomprehensibly far. so the like like we see a path to to
38:46putting a 100 gigawatts per year of solar powered AI satellite into orbit. and and having this be actually the lowest cost way to power and operate u AI at a very large scale. for reference, the United States consumes roughly 460 gawatts on average per year because the average power load in the US is 460 g.
39:22The whole country, the whole country,
39:24all electricity of all sources in the US. Yes.
39:26And you're talking about 100 being added.
39:29Well, roughly a quarter of the US electricity output.
39:35that we have we have a plan mapped out to do that. It it gets crazy.
39:42So So there's what, a trillion planets like Earth in the world in the solar system, whatever you call it. A trillion. And in that trillion, so Big Bang was 14 billion years ago, 13 half billion years big. And planet, our planet's only four billion years old. So there must be other planets that are like ours with all the minerals, oxygen, hydrogen, silicon, carbon. so life here has been extinguished four times and presumably you feel that other places, other civilizations, other planets, then we'll get off of this ex exist. And are
40:30they planets where the beings there are part human or part carbon and part metal?
40:39Well, I think we'd like to find out. I'd like to find out. I mean, my philosophy is one of curiosity. I I I just want to know what's, you know, what's going on in this universe. is the standard of is the standard physics correct about the beginning of the universe? Is heat death the end of the universe? Are there other alien civilizations? Can we talk to them? And and what questions should we be asking about reality that we don't know to ask? So that's my question is to
41:10expand consciousness to better understand the universe.
41:16So So let's go to away from the universe back to Tesla again. And so you said that
41:23gota [laughter] be back on the ground here.
41:25Back on ground. And so you said that our expertise is in making things better, faster, cheaper than other people. And when I started investing in Tesla, when we started investing in Tesla, you were telling us that that it's the machine that makes the machine that's most important. The machine that makes So you were into machine learning, machine technology then, which is 15 years ago. And so and now the
41:55average car I think is 50 minutes or 50 seconds, 40 seconds, 60 seconds and we're now 35 seconds. Every 35 seconds the car rolls off and then you say that we're going to get down to 10 seconds and you said it's a possibility we can go to five. Five I I
42:17five seconds every car is rolling off a line. How's that happen? I I I certainly see a path to achieving a roughly 5,000 millisecond sec time or 5 seconds. which is only that's only really walking speed. That's like sort of a fast walk. 1 m per second is a fast walk. The car is less than 5 m long. So 5second cycle time the the cars will will be exiting the line at walking speed. So it's it's you can run away from them. It's not it's not going to be like they're coming out like bullets or something. [laughter] So, but as a as a rough rule of thumb that's you know it's 10,000 minutes in a week. If you run 247 operation and you get you know let's say 10 cars per
43:06minute you've got 100,000 cars a week. So, so but the question is how come we're able to do this? Do other people just not care or they think that gee if I do something if I have this idea and we try to implement it and it doesn't work then I'm not going to get promoted or I can get fired or I'll get blamed and if it works then I'm going have to do a lot of extra work that I wouldn't know how to do if it didn't work. So why don't other people have, you know, a mindset of making things better? Why the J the Chinese, they've been great at copying us and in some instances probably even done better than us after they've copied for the first time. But
43:57how how come do you think that other people haven't been able to make the advances we have? And even the guy from Ford just recently said, "Gee, those people in China wouldn't give us compliments, but he said the people in and the Chinese have copied us. those people in China are doing great, which is really a compliments us because the reason they're doing great is because of us. So why don't other people do this? Why doesn't he do that?"
44:24most companies are incrementalist. you know the management team wants to do I don't know 5% maybe 10% better than last year as opposed to take big risks that could fail. obviously I don't have a problem with taking big risks. Yeah, I say
44:54so the and I like to use the tools of physics to analyze things and you know when I was in the factory one night I was I was looking at the factory and I was like you know this this could be much more efficient could be much faster. the I just trying to just rough math trying to calculate the volumetric efficiency of the factory. So if you divide the factory into cubic meters and say how many cubic meters are doing something useful and it's a surprisingly small percentage the volumetric density is not very good.
45:34and then the speed of the cars and the parts moving is quite slow. generally limited by the speed to which people can say attach brake lights or a seat or something like that. so if you densify the factory and improve the volumetric efficiency which is helpful for for for production efficiency because and things have less distance to move just like just like a chip you densify circuits on a chip make it more efficient. You think of a factory is a lot like drift. how do you make a drift faster? Well, you you bring circuits closer together, make them smaller and you increase the clock speed.
46:22So Robin told me that you told her once that I think of myself as a bit and if I'm a bit how would I like to travel? That makes sense.
46:32Bit or an atom? If it's if it's the I'm software I software or if I'm on a ship, I say what's the journey of the bit? what am I doing? And if if this journey doesn't make sense, I need to fix it. And if the journey of the atom in the factory doesn't make sense, I need to fix it.
46:48So when you say you're working on weekends, I spend all my Sundays working on a chip. What does that mean? What do you do?
46:54it's Saturdays, but some sometimes Sundays usually recent weekends been Sundays, too. yeah, just the the AI5 chip, which is going to be a great chip. you know, all all of T all of Tesla hinges on that chip. That's that's the chip that goes that go into our next generation of self-driving cars. And it's also essential for the Optimus robot.
47:28So that that chip program was in bad shape. it it wasn't it wasn't closing because it's it's it's quite an ambitious trip design and and it really wasn't on a path to success. and then we also had the dojo program which was doing it was also doing okay but but not on a path to be competitive with Nvidia.
47:51So I I collapsed I collapsed the two programs into just one program which is to get everyone focused on AI5 chip which is essential continue to use Nvidia for training but we need the AI5 inference chip which is it's a a very powerful chip but that but it's also a very low power chip so it's it doesn't use a lot of power so it's its performance for what is extremely good.
48:17we think you know it's probably going to be at least two to three times better than Nvidia informs for what at the at the inference level in the car and the robot and you know I don't know 10% of the cost of an Nvidia chip something like that. So, so the these are very important numbers to achieve and
48:46I had to get the ship program back on track. So, put immense amount of time into it and and at this point I have the entire physical design of the chip laid out in memory. I can visualize the whole thing. But when you're when you're talking about chip chip manufacturing, you say we might have to build some kind of a giant fab. Presumably, we would have a partner with TSMC or with Samsung. We wouldn't do this by ourselves. And those guys have been doing this for we would do it by ourselves. And and these are things that cost 20 billion, 30, 40 billion.
49:21And where the people come from to do this? How could we not have partners? Is your thought to have a partner to do it by ourselves ultimate 10 years? You don't there's not going to be enough chips in the world to accomplish what we're trying to do.
49:37Yeah. First of all, I have immense respect for TSMC and Samsung and we we've we worked with both TSMC and Samsung. at Tesla and at SpaceX. so TSMC and Samsung are great companies. and and I and we want them to make our trip as quickly as they can and scale up to as high as possible volume that that they're comfortable doing.
50:07but but it's it it doesn't appear to be fast enough. and you know, when I ask how long will it take to from start to finish to get a new chip fab built, they tell me five five years to get to blind production. I'm like five years to me is an opportunity. I my my timelines to meet you by the way
50:34one year, two year [laughter] and year three scale it rails to infinity. So So I I can't even see past three years. so then I'm like, damn, okay, this is this is not going to be fast enough. So, you know, if they if they change their minds and say, "Yeah, they're going to go faster and they they want to provide us with, you know, 100 200 billion AI
51:05chips a year," in the time frame that we need them. That's great. But how could they not when they know that we're demand and we are going to use the product in our own product. How could they not want to be our supplier or how could they not want to be partners with us? I don't understand.
51:25Oh no. Well, they are partners. we're using both TSMC
51:29I mean to really expand capacity tremendously. Why don't they do that? From their standpoint, they are because you'll be using TSMC Taiwan, TSMC Korea, TSMC Taiwan, TSMC Arizona, Korea, Samsung, and the Texas back Samsung. So, we got four fab and you know, from their standpoint, they're moving like lightning.
52:02I'm just saying that nonetheless it would be a limiting factor for for us that they're going as fast as they can from their standpoint. There's metal to the metal. They've just never had someone with sense of company without sense of urgency. could so I that's man it might just be that the only way to get to scale at the rate that we want to get to scale is to to build up a really big fab
52:35and or or or be limited in output of optimize and self-driving cars by the AI chips.
52:46Obviously you're not going to Right. Those are two choices.
52:50to go to FSD. We don't have too much more time, but FSD full ser full self-driving. the you know, we're making these tremendous breakthroughs, it seems. And you said recently that you didn't want to expand capacity for making cars until you were convinced that that was the case. you are now convinced that's the case and again here we're making these exponential leaps in make in in these cars and you said that a very large percentage of people who've actually paid for full self-driving don't use it and have never tried it before yes it's pretty well
53:33yeah how yeah how we do how
53:36nice something you want to try it type of thing it's amazing
53:40yeah yeah so so we're we're now kind of insisting with customers for safety reason that we demonstrate full self-driving because the numbers are unequivocal at scale that over with with the now order of 10 billion miles driven that it's four times safer on full set driving than than than not.
54:10So, so it's actually a big improvement in in safety. And so, at this point, we're we're just insisting that we at least demonstrate self-driving to customers so they know how to use it and turn it on for safety reasons.
54:25So, it drives full self-driving drives just like a person as opposed to having to code to look for every circumstance that would happen. and you had a you know this this is a fire truck in front of us with a bicycle attached to it or someone walking his dog while he's driving along. They have to identify everything with a code. What AI does, what we do is we as I understand it is to make sure that it's just like us. Is that fair?
54:59Yes. the it the key to achieving full self driving unsupervised full self driving from safer than a human is improving the AI software in the car. We're confident that the AI4 hardware that that's a chip that we designed currently made by Samsung is capable of achieving a safety level unsupervised meaning if you're asleep in the car at least two to three times that of the average right
55:33maybe more. and then with AI5 we think can achieve probably a 10x improvement in safety. so the these are really big deals. I mean I think this one these are very profound things that you're seeing here. and I really encourage people to go out there and try Tesla self-driving and and see for yourself.
55:58It's you can just go to any Tesla store, they'll they'll show it to you. You know, it's it's not secret.
56:03See, everyone here you're benefiting if you try this and then buy it. But once you try it, you're going to buy it.
56:13You should try it. and so, so I want to close on what I mentioned this morning on CNBC was that you're not doing this so you can get enough money to buy a beach house. You're you're do you're doing this even mine. [laughter] Yeah, that's really
56:38But but but you're doing this because you know Larry Page was right. You're you're a species that you think humans should survive.
56:51Yes, I'm pro I'm unabashically prohuman. [laughter]
56:57I mean, so you're you're spending actually whether you're worth another trillion dollars or 400 bill, it doesn't really matter. What are you going to do with all this money at the end? What what's what's your plan? How do you want people to think about you?
57:14Well, you know, mostly I need to have enough sort of ownership of the companies to be able to continue to direct their activities. but it's from a personal consumption standpoint. I I don't actually own any vacation homes. and I just own one sort of, you know, mediumsiz house in Austin.
57:36so and and actually say a tiny house at Star Base that I've
57:41seen that house. It's tiny. Yeah. Yeah.
57:44Yeah. I bought actually people people like friends of mine have come to visit then they thought I was kidding. I'm like, "No, it's real." I bought it $8,000. But I I've done a lot with the place.
57:56Artificial turf in front, little white pickup fence. Yeah. So but but like like I said with with AI and robotics there will be abundance world. So it's people actually in fact in a benign scenario there's going to be interesting threshold that that AI passes and AI robotics pass where it's run out of things to do for humans
58:32literally it's it's completely satiated all human wants and then I guess it'll have to start thinking about what to do for itself or I don't know but I but you know overall I want to take the set of actions that expand consciousness into the future so that the scope and scale of consciousness grows tremendously and that we explore other star systems like in Star Trek go places nobody's ever gone before and find out if there are existing alien alien civilizations or maybe this long dead alien civilization and we can look through their ruins and try to understand what what they were like.
59:23and just generally understand the universe.
59:28Where are you now right now?
59:30I'm in I'm in Silicon Valley.
59:35Thank you very much for today. Welcome Thank you for everything you've done for me, for our shareholders, and for humanity. [applause] Thank you, Ron. Thank you, sir. [applause] Thank you. What's next, guys? [laughter]
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