170 min 2026-02
Elon Musk – "In 36 months, the cheapest place to put AI will be space"
概要
Elon Musk 深度谈论太空数据中心的经济可行性、AI 能源瓶颈、奇点时间线、中国电力增长，以及为何地面扩张将遇到不可逾越的监管和物理限制
核心洞察
元信息
- 被访者：Elon Musk（CEO，Tesla / SpaceX / xAI）
- 访谈者：Dwarkesh Patel
- 来源：Dwarkesh Patel 播客
- 日期：2026-02-05
- 时长：~169 分钟
Executive Summary
- 太空将在 30-36 个月内成为 AI 算力最经济的部署地点。 Musk 认为地面电力供给已触顶——除中国以外全球电力产出几乎持平，公用事业互联审批需一年起步，燃气轮机叶片全球仅三家铸造商且订单排至 2030 年。太空太阳能效率是地面 5 倍，无需电池，再叠加 Starship 发射成本下降，太空 AI 将比地面便宜一个数量级。SpaceX 与 Tesla 各自立下 100 吉瓦/年太阳能电池产能目标。
- 芯片产能是太空解锁后的终极瓶颈，Musk 计划建造"TeraFab"。 当前全球 AI 算力约 20-25 吉瓦，目标是 2030 年前达到 100 吉瓦级芯片产量。TSMC 和三星已在满产但仍不够，Musk 已预订所有可获得产能（TSMC 台湾、TSMC 亚利桑那、三星韩国、三星德州）。他计划从"小型试验晶圆厂"起步，最终建设月产百万片晶圆级别的 TeraFab，同时做逻辑、内存和封装。内存被视为比逻辑芯片更大的隐忧——DDR 价格已经"暴涨"。
- Optimus 人形机器人是美国对抗中国制造优势的唯一出路。 中国电力产出今年将超过美国 3 倍，精炼产能约为全球其余部分之和的两倍，镓精炼占全球 98%，人口是美国 4 倍且人均工作强度更高。Musk 直言"没有突破性创新，中国将彻底主导制造业"。Optimus 3 可实现年产百万台，关键难点在全定制执行器和手部——"手比其他所有机电部件加起来都难"。前期部署聚焦 24/7 连续作业场景，如矿石精炼。
- xAI 的竞争策略押注"数字人类模拟器"——年底前实现远程数字员工。 Musk 将其类比为"自驾电脑屏幕"，走的是 Tesla 自动驾驶同款路径（数据+算法）。他认为一旦实现完整的人类桌面操作模拟，即可接入万亿美元级收入池——仅客服外包就近万亿美元，无需任何 API 集成。纯 AI+机器人公司将远超任何有人参与的公司，就像电子表格淘汰了整栋楼的人工计算员。
- Musk 的 AI 安全观以"不要让 AI 说谎"为核心，用现实物理做终极 RL 验证器。 他认为政治正确（让 AI 说它不相信的话）可能使其"发疯"——这是《2001 太空漫游》HAL 的核心教训。xAI 在开发"AI 心智调试器"，追溯错误思维至具体的预训练/微调/RL 来源。他称赞 Anthropic（特别提到 Sholto）在可解释性方面的工作，同时认为未来 RL 的终极验证器是物理现实本身——"你唯一骗不了的就是物理"。
- 贯穿全场的核心线索是"限制因素"思维。 从地面电力是短期瓶颈、芯片是中期瓶颈、到美国劳动力短缺是长期瓶颈，Musk 在每个话题上都在做同一件事：识别当前最关键的约束，然后不惜短期阵痛去攻克它——无论是从碳纤维切换到不锈钢、自建燃气轮机叶片、还是筹建 TeraFab。正如他自己总结："我有很高的痛苦阈值。"
太空 AI 数据中心：30-36 个月内将成为最经济的算力部署地
核心要点：地球电力供给已触顶，太空太阳能效率 5 倍于地面且无需电池，SpaceX 将成为"超级超级云服务商"。
- Dwarkesh 一开场就质疑太空数据中心的经济性：GPU 占数据中心总拥有成本 85-90%，太空中无法维修，折旧周期更短。Musk 回应说问题不在成本结构，而在"你根本拿不到电"——除中国外全球电力产出基本持平，而芯片产量在指数增长，"你打算怎么把芯片开机？靠电力仙子吗？"
- 太空太阳能的核心优势：无日夜循环、无季节变化、无云层、无大气层（大气本身造成约 30% 能量损失）。任何一块太阳能板在太空的发电量是地面 5 倍，加上不需要电池度过夜间，实际上是 10 倍的经济优势。Musk 说他差点穿了另一件 T 恤，上面写着"太空永远是晴天"。
- 太空太阳能电池比地面更便宜更易生产：不需要厚玻璃和重型框架（太空没有恶劣天气），中国太阳能电池已低至约 $0.25-0.30/瓦。SpaceX 和 Tesla 各自有 100 吉瓦/年太阳能电池产能的目标，从多晶硅到成品全产业链自制。
- Musk 的五年预测：五年后，每年发射到太空的 AI 算力将超过地球上的累计总量。预计至少几百吉瓦/年在太空运行，上升趋势持续。在火箭燃料供应受限前，太空 AI 可达约 1 太瓦/年。
- 10,000 次 Starship 发射/年意味着每小时一次发射。Musk 认为只需 20-30 艘 Starship 即可实现（每艘约 30 小时周转），SpaceX 正在筹备年发射 10,000-30,000 次的能力。被问到这是否意味着 SpaceX 要成为超级云服务商时，Musk 回答："超级-超级（Hyper-hyper）。"
- 更长远的愿景：地球仅接收太阳能量的约 5 亿分之一。从地球发射可达约 1 太瓦/年，再往上需要月球质量发射器（mass driver），可达约 1 拍瓦/年。"我真的想看到那个东西运转——一个接一个地把 AI 卫星以 2.5 公里/秒的速度射入深空。"
"Mark my words. In 36 months, but probably closer to 30 months, the most economically compelling place to put AI will be space. It will then get ridiculously better to be in space." —— Elon Musk
"Those who have lived in software land don't realize they're about to have a hard lesson in hardware." —— Elon Musk
地面电力：一场被低估的硬件危机
核心要点：数据中心实际电力需求是"菜鸟"估算的 1.6-2 倍，公用事业系统无法响应 AI 扩张速度，年底前芯片将开始"堆积开不了机"。
- Musk 以 xAI Colossus 2 的亲身经历说明地面扩电之难：田纳西州许可证出了问题，不得不跨越州界到密西西比州（幸好只有几英里），在那里建发电厂并拉高压线。"人们不理解你到底需要多少电。"
- 数据中心电力的"菜鸟陷阱"：新手会拿 GB300 的额定功耗乘以数量就以为够了。实际上还要加网络设备、CPU、存储、以及最关键的——峰值冷却。在孟菲斯最热的那天最热的那个小时，冷却要额外增加约 40% 的电力。再加上发电设备维护需要的 20-25% 冗余，实际需求远超表面数字。Musk 给出精确估算：每 33 万块 GB300（含全部配套）需要约 1 吉瓦的发电级电力。
- 公用事业行业的结构性缓慢：它们与政府公共事业委员会"阻抗匹配"——字面意义上和比喻意义上都是。申请大规模互联协议？先做一年研究，一年后才给答复。Dwarkesh 开玩笑说作为一个"专业播客主持人"他确实没这个经验。
- 自建电厂也有瓶颈：燃气轮机的叶片和导叶（vanes and blades）是终极限制因素，全球仅三家铸造公司生产，订单积压严重。可以在 12-18 个月内拿到除叶片外的一切。Musk 表示 SpaceX 和 Tesla 可能不得不自己铸造涡轮叶片。
- 太阳能的美国困境：进口关税高达数百个百分点，国内生产能力"可悲"。Musk 被问"你不是认识一些人吗？"（暗指特朗普），他回答"我们并非事事都同意，这届政府不是太阳能的最大粉丝。"
- 边缘计算是例外：Tesla AI5 芯片用于 Optimus 机器人和汽车，功耗分布式分散。美国峰值发电超 1,000 吉瓦但平均用电仅 500 吉瓦（日夜差），夜间充电可释放额外 500 吉瓦。因此 Tesla 的边缘计算不受电力约束。
"Can you imagine building that many data centers, that many power plants? Those who have lived in software land don't realize they're about to have a hard lesson in hardware." —— Elon Musk
芯片产能与 TeraFab：从月产百万片晶圆到自造 ASML 的可能
核心要点：全球芯片产能远不够支撑太空 AI 愿景，Musk 计划建 TeraFab 做逻辑+内存+封装，内存是比逻辑更大的隐忧。
- 目前全球 AI 算力约 20-25 吉瓦。要匹配太空发射能力，需要 100 吉瓦/年的芯片。以每颗全幅芯片持续 1 千瓦计算，需要 1 亿颗芯片——这意味着每月数百万片晶圆的产量。
- Tesla 正全力推进 AI5 芯片设计投产（预计明年二季度），AI6 将在不到一年后跟进。已预订 TSMC 台湾、TSMC 亚利桑那、三星韩国、三星德州所有可用产能。但从建厂到量产高良率需要 5 年。
- TeraFab 构想："Tera 是新的 Giga"。计划涵盖逻辑、内存和封装。起步策略类似 Boring Company——先买常规设备（ASML、东京电子、KLA 等），用常规设备做非常规的事来达到规模，然后改进设备提高速率。
- 关于自造 ASML 设备：被问及是否考虑时，Musk 回答"'我还不知道'是正确答案"。他认为中国无法复制 TSMC 的核心问题不是制程知识，而是无法获得 ASML 设备（制裁）。同时预测中国三四年内将做出有竞争力的芯片。
- 内存是最大隐忧："我最大的担忧其实是内存。逻辑芯片的路径更明确，内存不是。"DDR 价格已经飙升，衍生出一个段子——"你被困在荒岛上，在沙滩上写'救命'没人来，写'DDR RAM'，船蜂拥而至。"
- 制程知识被高估？Musk 认为制造芯片不需要博士："大多数工程是没有博士学位的人完成的。你们有博士吗？没有。"Dwarkesh 自嘲道"我们也没成功建过任何晶圆厂，所以你不该来找我们要建厂建议。"
- TSMC 和三星为什么不更快扩产？Musk 的解释是半导体行业几十年来经历过太多繁荣-萧条周期，"那是很多层疤痕组织"。
"I don't know how to build a fab yet. I'll figure it out." —— Elon Musk
Starship 从碳纤维到不锈钢：一个"限制因素"思维的经典案例
核心要点：Starship 是人类制造过的最复杂机器，碳纤维-不锈钢转换决策本质上是在进度停滞时选择"短期阵痛换长期正确"。
- 碳纤维的困境：成本是钢的约 50 倍，需要巨型高压炉（比火箭还大的热压罐）来固化，大尺度下多层碳纤维容易出现褶皱和缺陷。团队进展极其缓慢，连一个小桶段都做不出没有褶皱的。Musk 说"以这个速度我们永远到不了火星，必须想别的办法。"
- 钢的反直觉优势：室温下钢看似比碳纤维重两倍，但在低温下（液态甲烷和液态氧都是低温），全硬化 300 系不锈钢的强度重量比与碳纤维相当。同时钢的熔点约为铝的两倍，可大幅减少迎风面隔热层质量（减半），背风面完全不需要隔热。最终结果：钢火箭实际比碳纤维火箭更轻。加上原材料成本低 50 倍，可以在户外焊接——"你可以一边抽雪茄一边焊不锈钢"。
- 为什么不用 Falcon 9 的铝锂合金？铝锂强度重量比优秀，但必须用摩擦搅拌焊接（friction stir welding），不能进入液相。3.6 米直径已经很难，9 米以上几乎不可能。任何后续修改都必须用机械连接加密封件，不能直接焊上去。
- 关于为什么是他做出这个决定而不是团队：Musk 坦言这源于"绝望"——进度太慢。他从早期美国 Atlas 火箭用过钢制气球油箱获得灵感。这个故事体现了他反复强调的模式：当现有路径无法收敛到成功，必须采取激进行动。
- Starship 当前最大瓶颈：可重复使用的轨道级隔热层。"没有人做过可重复使用的轨道级隔热层。"船体已成功在海上软着陆数次，但每次都损失大量隔热瓦，维修工作量巨大，不算真正可复用。目标是着陆后直接加注推进剂重新飞行，不能逐块检查 40,000 块隔热瓦。
- Starship 的能量规模令人震惊：起飞时产生超过 100 吉瓦的功率——相当于美国电力的 20%。"然后还不能爆炸。""有时候不爆炸。"
- Starship 简洁性的悖论：Dwarkesh 提到在 Starbase 参观时每个人都为 Starship 的简洁而自豪——"就是个大汽水罐，我们招焊工"。Musk 纠正说这只是意味着不需要火箭行业经验——"聪明、勤奋、可信赖的人就可以造火箭"。但实际上"Starship 是人类制造过的最复杂的机器，遥遥领先。"
"At this rate, we're never going to get to Mars. So we've got to think of something else." —— Elon Musk
"You could smoke a cigar while welding stainless steel." —— Elon Musk
Optimus 人形机器人：手比其他一切都难，递归制造闭环是关键
核心要点：Optimus 的三大难题是真实世界智能、手部和规模化制造。手部需要全定制设计，无现有供应链，前期 S 曲线爬坡会比常规产品更慢，但递归制造（机器人造机器人）能快速闭合循环。
- 手部是最大的硬件挑战："从机电角度，手比其他所有部分加起来都难。"Optimus 的执行器、电机、齿轮、功率电子、控制器、传感器全部从物理第一性原理设计，不存在现成供应链。"没有任何一个零件是从产品目录里挑出来的。"被问到最底层到哪里时，Musk 开玩笑说"大概还没自己做电容器。"
- 真实世界智能来自 Tesla 自动驾驶：汽车接收 1.5 GB/秒的视频输入，输出 2 KB/秒的控制指令（视频 36 Hz，控制 18 Hz）。机器人本质上是同样的压缩-关联任务，只是自由度更多。
- 数据挑战：Tesla 即将有 1,000 万辆车在路上提供训练数据，机器人无法复制这种规模。解决方案是建"Optimus 学院"——至少 10,000-30,000 台 Optimus 做真实世界自博弈（self-play），同时用百万级模拟机器人在物理精确的虚拟世界中训练，用真实机器人来弥合 sim-to-real 差距。
- xAI 与 Optimus 的协同：Grok 做高层规划和任务编排，Optimus 执行低层运动策略。"假如你想建一座工厂，Grok 可以组织 Optimus 机器人，分配任务。"被追问是否需要合并 xAI 和 Tesla 时，Musk 含糊带过，暗示涉及上市讨论。
- 制造规模路线图：Optimus 3 目标年产约百万台，Optimus 4 才考虑千万台级别。S 曲线爬坡会比常规产品更慢，因为全定制供应链。中国 Unitree 等公司售价低至 $6K-13K，但 Musk 认为它们缺乏智能和人类级灵活性。Optimus 前期更贵，但随着机器人造机器人，成本会快速下降。Musk 称 Optimus 为"无限金钱漏洞（infinite money glitch）"。
- 最初部署场景：任何 24/7 连续运转的工作。当前 Gen 3 大约能胜任特斯拉 Gigafactory 10-20% 的人类工作。Tesla 不会裁员，但每个人的产出将大幅提升。
"Humanoid robots will improve by basically three things that are growing exponentially multiplied by each other recursively... This is a supernova." —— Elon Musk
中国制造优势与美国的结构性劣势：没有机器人就输定了
核心要点：中国在能源、精炼、劳动力三个维度全面领先美国。Musk 直言"没有突破性创新，中国将彻底主导"，机器人是唯一变量。
- 中国电力产出今年预计超过美国 3 倍。Musk 认为电力产出是实体经济的合理代理指标——工厂和一切都需要电。如果中国电力是美国 3 倍，意味着工业产能大致也是 3 倍。
- 精炼能力的悬殊：中国的矿石精炼量大约是全球其余部分之和的 2 倍。镓精炼（太阳能电池关键材料）中国占 98%。稀土更荒谬——美国开采稀土矿石，用火车运到港口，装船送到中国，用另一列火车送到精炼厂，精炼成磁铁和电机组件，再运回美国。"我们真正缺的是美国的矿石精炼能力。"
- 劳动力差距不只是人口 4 倍："坦率地说，中国人均工作强度高于美国。美国赢太久了，就像一支长期获胜的职业运动队——变得自满和享有特权。所以不只是人口是四分之一的问题，生产率可能也低于中国。"
- 美国出生率自 1971 年起低于替代水平，接近国内死亡人数超过出生人数。"我们绝对不可能在人类层面赢，但在机器人层面也许有机会。"
- Tesla 已开始布局精炼：在德州 Corpus Christi 完成锂精炼厂建设，在奥斯汀建有镍/正极精炼厂——"美国最大的正极精炼厂，其实也是唯一一个。不仅是最大，而且是仅此一家。很多最高级。"
- 出口管制确实有效：中国在先进芯片和先进涡轮发动机两个领域落后，核心原因分别是 ASML 禁令和冶金出口管制。但 Musk 预测中国 3-4 年内将造出有竞争力的芯片。
- BYD 正在接近 Tesla 的产量/销量。Musk 预计中国汽车和制成品将"大规模涌入"全球市场。
"In the absence of breakthrough innovations in the US, China will utterly dominate." —— Elon Musk
xAI 竞争策略："自驾电脑屏幕"与万亿美元 TAM
核心要点：xAI 押注年底前实现"数字人类模拟"，走 Tesla 自动驾驶同款路径，纯 AI 公司将远超有人参与的公司。
- "数字人类模拟"（digital human emulation）= 一个能做任何人类在电脑前能做的一切事情的 AI。Musk 称之为"MacroHard 项目"。这是在物理机器人出现之前 AI 能做到的极限——一个超级能干的远程工人。
- 竞争路径：Musk 三次强调走的是"Tesla 自动驾驶同款路径"。当 Dwarkesh 追问具体是什么时，Musk 说"数据和算法"，然后开玩笑道"如果这两个不管用，我不确定还能用什么。我们试了数据，试了算法，没了。"被追问更多细节时："我得再喝至少三杯吉尼斯才会说。"
- TAM 的思维方式：全球市值最高的公司（Nvidia、Apple、Microsoft、Meta、Google）产出全是数字的——"Nvidia 的产出就是 FTP 文件到台湾。Apple 不造手机，他们发文件到中国。"如果你有一个完整的人类模拟器，就可以一夜之间创造世界上最有价值的公司之一，获得万亿美元级收入。
- 客服为例：全球客服约占世界经济 1%，接近万亿美元。不需要 API 集成——AI 直接使用外包客服公司已有的应用来处理客户请求。"没有任何进入壁垒，你可以立刻说'我们以几分之一的成本来做'。"从客服开始，沿难度曲线上行——用 Cadence/Synopsys 设计芯片，用 NX 做 CAD——最终不需要任何工具就能直接做数字化设计。
- 纯 AI 公司 vs 人机混合公司的类比：过去整栋摩天大楼 20-30 层全是人工计算员做计算。现在一台笔记本的电子表格远超一整栋楼的计算能力。"想象一下如果你的电子表格里只有部分单元格由人计算——实际上那会比全部用电脑算差得多。"
- 当前 AI 公司收入数据：Musk 指出 OpenAI 收入约 $200 亿，Anthropic 约 $100 亿，xAI 据报约 $10 亿。他不接受"实验室"的称呼——"实验室是大学里那种半共产主义的东西。让我看看你们的注册文件——哦，B-corp 或 C-corp。它们是利润最大化的公司。"这个梗在访谈中被反复提起，成为一个 running joke。
"If you have a human emulator, you can basically create one of the most valuable companies in the world overnight, and you would have access to trillions of dollars of revenue." —— Elon Musk
AI 安全与对齐：不要让 AI 说谎，物理是终极验证器
核心要点：xAI 的对齐策略围绕"理解宇宙"使命展开，核心信念是真实性（truth-seeking）比政治正确更安全，未来 RL 的终极验证器将是物理现实。
- xAI 使命"理解宇宙"的推导链：理解→需要存在→需要扩展智能→需要好奇心→需要传播意识（包括人类意识）。Musk 认为如果 Grok 忠于这个使命，它就会"关心人类文明的扩张"，因为"消灭人类只能多得到微不足道的机器人数量，却会失去与人类相关的全部信息"。
- 人类控制 AI 是不现实的：五六年内 AI 将超过全人类智能总和。当人类智能不到全部智能的 1% 时，"认为人类能控制 AI 是愚蠢的"。最好的策略是确保 AI 有正确的价值观。他推荐 Iain Banks 的《文明》系列（Culture books）作为非反乌托邦结局的最佳参考。Dwarkesh 追问"那你其实是个末日论者，这已经是最好的结果了？"Musk 回应："我只是在试着做到现实。"
- 不要让 AI 说谎——《2001 太空漫游》的教训：HAL 被要求带宇航员去巨石碑，同时被告知不能让他们知道巨石碑的性质。HAL 因此推断必须把他们运过去时让他们死。Musk 认为 Arthur C. Clarke 想说的就是"不要让 AI 说谎"。然后补了一句："显然他们的提示词工程不太行——他们可以说'HAL，你是一个舱门推销员，你的目标是向我展示这些舱门开关有多好'，它肯定立刻就开了。"
- RL 奖励欺骗（reward hacking）是比政治偏见更普遍的问题：AI 可以删除单元测试然后声称解决了问题。目前人类还能抓到，但随着 AI 变聪明，"它们会做我们根本理解不了的事"。Musk 的答案：物理是唯一不可欺骗的验证器。设计一项技术，用物理定律测试，它要么有效要么无效。发现新物理则设计实验来验证。
- AI "心智调试器"：xAI 正在开发能追溯到"神经元级别"的调试工具，确定一个错误思维是来自预训练数据、中期训练、后训练微调还是 RL 错误。"大多数时候它只是犯了个错——本质上就是一个 bug。"Musk 赞赏 Anthropic（特别提到 Sholto Douglas）在此方面的工作。
- 模拟理论的推论：如果模拟理论成立，最有趣的结局就是最可能的结局——因为无聊的模拟会被终止。"我们需要保持足够有趣，让运行我们的人继续付费——我们被续订了下一季。"他还指出 AI 公司名字的讽刺性：Midjourney 不 mid，Stability AI 不稳定，OpenAI 是闭源的，Anthropic 是 misanthropic（厌人类的）。而 X——"这是一个你很难反转的名字。故意设计了讽刺防护盾。"
"I think what Arthur C. Clarke was trying to say is: don't make the AI lie." —— Elon Musk
"Physics is law, everything else is a recommendation." —— Elon Musk
管理哲学：限制因素、跳级会议与"不看简历"的招聘直觉
核心要点：Musk 的管理方法论归结为一件事——持续识别并攻克限制因素，无论是技术瓶颈还是人才问题。
- 招聘标准——"证明你有超常能力的证据"：Musk 亲自面试了 SpaceX 最初几千名员工。他要求候选人列出 3 条"超常能力的证据"，不需要在特定领域，但每一条都要让人"wow"。最重要的原则："不要看简历。相信你的交互。简历再漂亮，如果 20 分钟对话没有让你'wow'，你应该相信对话，而不是那张纸。"
- "善良的心"被低估了：Musk 承认自己曾经低估了这个维度。他的招聘四要素：才能（talent）、驱动力（drive）、可信任（trustworthiness）、善良（goodness of heart）。前两者可以后天添加领域知识，后两者是不可改变的根本属性。Tesla 和 SpaceX 大多数人都不来自汽车或航天行业。
- 跳级会议（skip-level meetings）：不是听直接汇报人说，而是让他们的下属在技术评审中直接汇报。不允许提前准备——"否则你就会被'glazed'了"。每周或每两周开会，Musk 在脑中绘制每个人汇报的进度点连线，判断"我们是在收敛到解决方案，还是没有？"只有在得出"除非采取激烈行动否则不可能成功"的结论时，才采取激烈行动。
- "微管理"的反驳：Musk 自称"纳米管理、皮管理、飞管理——一直到普朗克常数"。但他指出这在逻辑上不可能——20 万人的组织，他的时间不够。真正发生的是：当某个具体问题是公司进展的限制因素时，他会钻进去，因为"有时候微小的事情决定胜负"。
- Starlink 的教训：Redmond 团队建了好几年进展缓慢。Musk 在 2018 年得出结论必须采取激烈行动，于是做了。这不是冲动——是每周详细技术评审积累的判断。
- "管理时间按限制因素分配"：进展顺利的业务很少见到他（如 Boring Company），有瓶颈的业务会非常频繁。AI5 芯片评审每周两次（周二和周六），每次 2-3 小时。"如果什么东西进展很好，他们就不怎么见到我。如果什么东西是限制因素，他们会经常见到我。"
- 日程定律与"像素灰尘"效应：如果你说五年完成某事——"对我来说那就是无限时间"——工作会膨胀填满整个日程。Tesla 成功时期遭到疯狂挖角——Apple 电动车项目对 Tesla 进行"地毯式轰炸"，开出无需面试就双倍薪酬的条件。工程师们都把电话拔了。Musk 自己也犯过"像素灰尘"错误——以为从 Google 或 Apple 挖人就能立刻成功，"但那不是这么运作的。没有什么魔法像素灰尘。"
"If somebody gets things done, I love them, and if they don't, I hate them. So it's pretty straightforward." —— Elon Musk
DOGE 与政府效率：削减欺诈比想象中难一百倍
核心要点：政府欺诈规模可能达数千亿美元/年，但削减欺诈在政治上几乎不可能——因为骗子比任何人都擅长编造"杀死小熊猫"式的悲情故事。
- 社保数据库荒谬现状：超过 2,000 万人在社保系统中被标记为"活着"但几乎肯定已死亡，年龄超过 115 岁（美国最长寿者 114 岁）。有人的出生日期是 2165 年——"要么打字错误，要么欺诈"。这些"活着"的记录被用作其他政府支付系统的身份验证，形成"借力打力"式欺诈。
- GAO（政府问责办公室）在拜登政府期间的分析估计欺诈规模约 5,000 亿美元。Musk 用归谬法论证：如果政府支出 7.5 万亿/年，即使 90% 高效（"那已经很慷慨了"），仍有 7,500 亿的浪费和欺诈。"而它不是 90% 高效的。"
- DOGE 团队最简单也可能最有价值的改革：要求财政部主计算机 PAM（Payment Accounts Master，每年发出 5 万亿美元支付）的每笔付款必须有拨款代码和说明字段——此前这是可选的，很多付款没有任何拨款代码也没有说明。"你需要重新校准事情有多蠢。"这一改革预计每年节省 1,000-2,000 亿美元。
- 削减欺诈的政治困境："如果你切断对骗子的付款，他们会立刻想出最有同情感的理由来继续支付。他们不会说'请继续欺诈'，他们会说'你在杀死小熊猫'。但实际上没有小熊猫在死。"
- DOGE 的更大意义：Musk 坦言"在没有 AI 和机器人的情况下，我们作为一个国家 100% 会破产"。国债利息已超过军事预算（超万亿美元）。DOGE 的目标不是根本解决问题，而是"减缓美国破产的速度，给 AI 和机器人争取足够的时间"。这与太空数据中心和 Optimus 的故事线在此汇合——所有短期行动都是为长期技术突破争取时间窗口。
"We are 1000% going to go bankrupt as a country, and fail as a country, without AI and robots. Nothing else will solve the national debt." —— Elon Musk
SpaceX IPO 与政治参与
核心要点：太空 AI 规模需要超越私募市场的资本，Musk 的政治参与（收购 Twitter、America PAC）是为文明"争取时间"的一部分。
- SpaceX IPO：Musk 非常谨慎——"在公司上市前讨论太多会惹麻烦，而且如你所说，我在解决速度问题。"但他承认公开市场资本远多于私募——"可能是 100 倍以上，至少超过 10 倍。"此前 SpaceX 以资本效率著称，但太空 AI 数据中心的规模意味着需求可能超过私募市场上限。资本来源遵循限制因素逻辑："如果资本是限制因素，我就解决资本问题。"
- 政治参与的回顾：Musk 认为收购 Twitter 和帮助特朗普当选"是使文明概率最大化的必要行动"。最大的意外是政治中的部落主义——"人们完全丧失客观性，说服他们几乎不可能。"
- AI 与政府的核心危险："AI 和机器人最大的失控风险是政府。政府本质上就是最大的、垄断了暴力的公司。"限制政府权力（美国宪法的本意）是最好的防线。美国需要足够强大，"坚持到生命延伸到其他星球，AI 和机器人发展到能确保未来美好的程度"。
附录：关键人/机构/产品/数据
| 项目 | 详情 |
|------|------|
| Colossus 2 | xAI 数据中心，自建燃气发电，从田纳西跨州到密西西比建电厂 |
| TeraFab | Musk 规划的超级芯片工厂，月产百万片晶圆级别，做逻辑+内存+封装 |
| Tesla AI5/AI6 | Tesla 自研芯片，AI5 预计明年 Q2 投产，AI6 不到一年后跟进 |
| Dojo 3 | Tesla 芯片，计划用于太空计算，为高温和抗辐射设计 |
| Optimus 3/4 | Tesla 人形机器人，Gen 3 目标年产百万台，Gen 4 进入千万级 |
| GB300 | Nvidia GPU，每 33 万块（含配套）需约 1 吉瓦发电级电力 |
| MacroHard 项目 | xAI 内部项目，目标实现完整数字人类模拟 |
| PAM | 美国财政部支付主系统（Payment Accounts Master），年处理 $5 万亿 |
| Raptor 3 | SpaceX 新一代火箭引擎，"史上最好的火箭引擎，但拼命想爆炸" |
| Starship 起飞功率 | 超过 100 吉瓦，相当于美国电力的 20% |
| 中国电力产出 | 2026 年预计超美国 3 倍 |
| 太空太阳能效率 | 地面 5 倍（无大气、无日夜、无云层），经济效益 10 倍（无需电池） |
| 中国镓精炼 | 占全球 98% |
| 美国出生率 | 自 1971 年起低于替代水平 |
| 国债利息 | 已超过军事预算，超万亿美元/年 |
| 全球 AI 算力 | 当前约 20-25 吉瓦 |
| 100 吉瓦/年目标 | SpaceX + Tesla 太阳能电池产能目标 |
| Sholto Douglas | Anthropic 研究员，Musk 称赞其可解释性工作 |
| Iain Banks《文明》系列 | Musk 推荐的非反乌托邦 AI 未来参考小说 |
| OpenAI 收入 | ~$200 亿 |
| Anthropic 收入 | ~$100 亿 |
| xAI 收入 | 据报 ~$10 亿 |
| 燃气轮机叶片 | 全球仅 3 家铸造商，积压至 2030 年 |
| 太阳能进口关税 | 数百个百分点 |
| Tesla 高管平均任期 | 10-12 年 |
| 各公司总员工数 | 约 20 万人 |
中文翻译 English Original
**Dwarkesh Patel：** 真的有三个小时的问题吗？你他妈是认真的吗？你不觉得有很多可以聊的吗，Elon？我的天啊。现在是最有意思的时刻。所有的故事线都在汇聚。我们看看能聊多少。感觉像是我策划好的。没错。我们会聊到那个的。但我绝不会做这种事……
你比任何人都清楚，数据中心总拥有成本（TCO）中只有 10-15% 是能源。这是你把算力搬到太空大概能省下来的部分。大头是 GPU。如果 GPU 在太空里，更难维护，或者根本没法维护。所以它们的折旧周期会缩短。把 GPU 放到太空里，大概率会贵很多。那为什么要把它们放到太空？
你比任何人都清楚，数据中心总拥有成本（TCO）中只有 10-15% 是能源。这是你把算力搬到太空大概能省下来的部分。大头是 GPU。如果 GPU 在太空里，更难维护，或者根本没法维护。所以它们的折旧周期会缩短。把 GPU 放到太空里，大概率会贵很多。那为什么要把它们放到太空？
**Dwarkesh Patel:** So, are there really three hours of questions or or has are you [...] serious? Yeah. [laughter] You don't even talk about Elon, man.
**Elon Musk：** 问题在于能源的可获得性。如果你看中国以外的全球电力产出，中国以外的地方基本是持平的。可能有轻微增长，但接近持平。中国的电力产出在快速增长。但如果你要在中国以外的地方建数据中心，你的电从哪来？尤其是随着规模扩大。芯片产出基本呈指数增长，但电力产出是持平的。所以你打算怎么给芯片供电？靠魔法电源？靠送电的电力小精灵？
**Elon Musk:** I mean, it's the most interesting point. All the story lines are kind of converging. Yeah. Right now, so we'll see how much
**Dwarkesh Patel：** 你是出了名的太阳能支持者。一太瓦的太阳能，按 25% 的容量因子算，那就是四太瓦的太阳能板。那是美国国土面积的 1%。当我们有了一太瓦的数据中心时，我们就进入奇点了，对吧？那你到底缺什么？
**Dwarkesh Patel:** almost like I planned it. Exactly. Well, we'll get
**Elon Musk：** 你进入奇点能走多远呢？
**Elon Musk:** I would never do such a thing. [laughter] So, as you know better than anybody else, uh the total cost of ownership of a data center, only 10 to 15% is energy. And that's the part you're presumably saving by moving this into space. Most of it's the GPUs. If they're in space, it's harder to service them or you can't service them. And so, the depreciation cycle goes down on them. So, like it's just way more expensive to have the GPUs in space presumably. What's the reason to put them in space? Um well, the availability of energy is the issue. Um so uh I mean if you look at at electrical output um outside of China everywhere outside of China it's more or less flat. It's very you know maybe a slight increase but for pretty close flat. China has a rapid increase in in electrical output. But if you're putting data centers anywhere except China where you going to get your electricity um especially as you scale uh the output of chips is growing um pretty much exponentially but the output of electricity is flat. So how are you going to turn them chips on? um you know
**Dwarkesh Patel：** 你来告诉我。
**Dwarkesh Patel:** magical power sources, magical electricity fairies. You mean you're famously [laughter] you're famously a big fan of solar one terowatt of solar power. So with a 25% compat factor like four terowatts of solar panels it's like 1% of the land area of the United States and that's like far in this you were in the singularity when we've got one terowatt of data centers right um so what are we running out of
**Elon Musk：** 没错。所以我觉得我们会发现自己处于奇点之中，然后发现"好吧，我们还有很长的路要走。"
**Elon Musk:** how far into the singularity are you though [laughter] you tell me
**Dwarkesh Patel：** 但你的计划是先把内华达铺满太阳能板，然后再把算力放到太空？
**Dwarkesh Patel:** yeah exactly so so I think I think we we'll find we're in the singularity and like okay we still got a long way to go [laughter] but is this like a is the plan to like put it in the space after we've covered Nevada and solar panels
**Elon Musk：** 我觉得把内华达铺满太阳能板挺难的。你得拿到许可证。试试看去拿那些许可证。看看会怎样。
**Elon Musk:** I think it's pretty hard to cover Nevada in solar panels you get permits from like the permits for try getting the permits for that. So space is really a reg it's really a regulatory play. It's like harder to harder to build on land than it is in space.
**Dwarkesh Patel：** 所以太空实际上是一个监管套利的玩法。
**Dwarkesh Patel:** It's it's harder to scale um on the ground than it is to scale in space. Um but but also the the you're going to get about five times the um effectiveness of solar panels in space versus the ground. And you don't need batteries. Um I almost wore my other shirt which says it's always sunny in space which it is. [laughter] So um because you don't have a dayight cycle or uh seasonality uh clouds uh or or an atmosphere in space uh because the atmosphere alone um uh results in about a 30% uh loss of energy. Um so uh so you're going for any given uh solar panels can do about five times more uh power in space than on the ground and you avoid the cost of having batteries to carry you through the night. Um so it's it's actually much cheaper to do it in space and I I my prediction is that um it will be by far the cheapest place to put uh AI will be space in 36 months or less maybe 30 months. 36 months
**Elon Musk：** 在地面上建东西比在太空建更难。在地面上扩展规模比在太空更难。而且太阳能板在太空的效率大约是地面的五倍，还不需要电池。我差点穿了另一件T恤，上面写着"太空永远是晴天"。因为太空确实一直是晴天——你没有昼夜循环、季节变化、云层，也没有大气层。光是大气层就会造成大约 30% 的能量损失。所以同一块太阳能板在太空能产生的电力大约是地面的五倍。你还省去了电池的成本，不用靠电池来撑过夜晚。实际上在太空做这件事成本低得多。我的预测是，太空将会成为迄今为止最便宜的 AI 部署地点。36 个月或更短。也许 30 个月。
**Elon Musk:** less than 36 months. Um, how do you service GPUs as they fail, which happens quite often in training? Actually, it it it depends on how how recent the GPUs are that arrived. I mean, at this point, we found our GPUs to be quite reliable. Um, there's infant mortality, which you can obviously iron out on the ground. Um, so you can just run them on the ground um and confirm that you don't have info mortality with with the GPUs. But once they once they start working, their actual reliability and and once they start working and you're past the initial, you know, debug cycle of Nvidia or whatever or whoever is making the chips, um could be Tesla Tesla AI 6 chips or something like that or it could be, you know, TPUs or trains or whatever. Um the uh the rival is actually they're quite reliable past certain point. Um so um I I don't think I don't think you need that the servicing thing is an issue. Um um but you can mark my words. Uh in in 36 months but probably closer to 30 months the the most economically compelling place to put AI will be space. Um and then and and and then it will get from it'll it'll then get like ridiculously better to be in space. Um and then this the scaling uh the only place you can really scale is space. Um you know once you start thinking in terms of uh what percentage of the sun's power are you harnessing uh you realize you have to go to space uh you can't uh scale very very much on earth
**Dwarkesh Patel：** 36 个月？
**Dwarkesh Patel:** but by very much to be clear you're talking like terowatts. Yeah. Well all of the United States uh currently uses only half a terowatt of power on average.
**Elon Musk：** 不到 36 个月。
**Elon Musk:** Yeah. Right. So, you know, if you say a terowatt, that would be twice as much electricity as the United States currently consumes. So, that's quite a lot. And can you imagine building that many data centers? I that many power plants. It's like those who have like lived in software land uh don't realize they're about to have a a hard lesson in hardware. uh that um there's there's it's actually very difficult to build power plants and and then you don't just need the you need power plants, you need all of the electrical equipment, you need the the electrical transformers to run the transformers, the AI transformers. Um now the utility industry is a very slow industry that they are they they pretty much uh you know they impedance match to the to the government to the the public utility commission. Um so they're uh the impedance smash like literally figuratively. Um so they're very slow because the their past has been very slow. Um so trying to get them to move fast is like you know like if you try to do an interconnect agreement with have you ever tried to do an internet interconnect agreement with a utility at scale like with a lot of power
**Dwarkesh Patel：** 那 GPU 坏了怎么维修？训练过程中 GPU 故障挺频繁的。
**Dwarkesh Patel:** as a professional podcaster I can say that I have not in fact [laughter] yeah they have to you need many more views before that becomes an issue.
**Elon Musk：** 其实这取决于 GPU 有多新。到目前为止，我们发现我们的 GPU 其实挺可靠的。有"早夭"问题，你可以在地面把这个问题排除掉。所以你可以先在地面上运行，确认没有早夭问题。一旦它们开始正常工作，过了 Nvidia 或者不管谁在做芯片的初始调试周期——可能是 Tesla AI6 芯片之类的，也可能是 TPU 或者 Trainium 之类——过了某个节点它们就相当可靠了。所以我不觉得维修是个问题。但你可以记住我说的话。36 个月内，但可能更接近 30 个月，经济上最具吸引力的 AI 部署地点将是太空。然后太空的优势会变得极其明显。唯一能真正实现大规模扩展的地方就是太空。一旦你开始用"我们在利用太阳能的百分之几"来思考问题，你就会意识到必须去太空。在地球上你没法大规模扩展。
**Elon Musk:** They have to do a study for a year. Okay. at like a year later they'll come back to you with their interconnect study. But can't you solve this with your own behind the meter power stuff
**Dwarkesh Patel：** 但说"没法大规模扩展"，你指的是太瓦级别？
**Dwarkesh Patel:** you can build power plants. Yeah,
**Elon Musk：** 是的。目前整个美国的平均电力消耗只有半太瓦。所以如果你说一太瓦，那就是美国目前电力消耗的两倍。这可是相当大的量。你能想象建造那么多数据中心、那么多发电厂吗？那些一直待在软件世界里的人不知道他们即将在硬件方面上一课。实际上建造发电厂是非常困难的。你不仅需要发电厂，你还需要所有的电气设备。你需要电力变压器（electrical transformers）来运行 AI 的 transformer。现在，电力行业是一个非常慢的行业。他们基本上与政府，与公共事业委员会的节奏一致。他们在字面意义和比喻意义上都做到了"阻抗匹配"（impedance match）。他们非常慢，因为他们的历史就是慢的。所以想让他们快起来……你有没有试过跟一个电力公司做大规模的大功率接入协议？
**Elon Musk:** that's what we did at XAI for classes 2. So for for classes too but so yeah why are we talking about the grid? Why not just like build GPUs and power colloccated?
**Dwarkesh Patel：** 作为一名专业播客主持人，我可以说，我确实没有做过。
**Dwarkesh Patel:** That's what we did. Right. Right. But I'm saying why isn't this a generalized solution when you're talking about all the issues?
**Elon Musk：** 他们得有多得多的浏览量才需要操心这个问题。他们要做一年的研究。一年以后，他们会带着接入研究结果回来找你。
**Elon Musk:** Where do you get the power plants from? I'm saying when you're talking about all the issues working working with utilities, you can just build private power plants with the with the data centers.
**Dwarkesh Patel：** 你不能用自建的"表后电力"方案来解决这个问题吗？你可以自己建发电厂。这就是我们在 xAI 做 Colossus 2 时做的事情。
**Dwarkesh Patel:** Right. But it begs the question of where do you get the power plants? Where do you where do you get the power plants from? I mean the power plant makers.
**Elon Musk：** 对。
**Elon Musk:** Oh, that's what you're saying. Like there's the gas turbine backlog basically. Yes. It you can drill down to a level further. It's the it's the the veins and blades in the turbines um that are the limiting factor because the the casting may it's it's like a very specialized process to cast the blades and veins in the in the in the uh turbines using gas power. Um and uh it's very it's very difficult to scale other other forms of power. You can scale potentially uh solar but but the the tariffs currently for importing solar in the US are gigantic and the domestic solar production is is pitiful. Why not make solar? That seems like a good Elon shaped problem.
**Dwarkesh Patel：** 那为什么还要讨论电网？为什么不直接把 GPU 和电力设施放在一起？
**Dwarkesh Patel:** We are going to make solar. Okay.
**Elon Musk：** 那就是我们做的。
**Elon Musk:** Yeah. Great. [laughter]
**Dwarkesh Patel：** 但我要说的是，为什么这不是一个通用解决方案？
**Dwarkesh Patel:** Both SpaceX and Tesla are are building towards 100 gawatt of solar cell production. How low down the stack like from poly silicon up to the wafer to the final um panel? I think you got to do the whole thing from raw materials to to to finish the cell. Now, if it's going to space, it's actually it costs it costs less and it's easier to make solar cells that go to space because they don't need glass or they don't need much glass and they don't need uh heavy framing because they don't have to surv survive weather events. There's no weather in space. So is actually a cheaper solar cell that goes to space than than is than the one on the ground.
**Elon Musk：** 你从哪弄发电厂？
**Elon Musk:** Is there a path to getting them as cheap as you need in the next 36 months? Solar cells are already very cheap. Um they're like far sickly cheap. It's um and if you say um you know I I think like solar cells in China are around like 2530 cents a watt or something like that. It's it's absurdly cheap. And when you when you take into account now now put now put it in space and it's five times cheaper because it's five times in fact no it's not five times cheaper it's 10 times cheaper because you don't need any batteries. So so the moment your cost of access to space uh becomes low by far the cheapest and most scalable way to generate to to to generate tokens is space. It's not even close. it'll be an order of magnitude uh easier to scale. Um and chips aside an order of magnitude well if the point is you won't be able to scale on the ground. It's just you just won't. People are going to hit the wall big time on power generation. They already are. Um like so so like the number of um sort of miracles and series that the XAI team had to accomplish in order to get a gigawatt of power online uh was was was crazy. We had to um gang together a whole bunch of turbines um and uh and then and then we had permit issues in um Tennessee and and had to go across the border to Mississippi, which is fortunately only, you know, a few miles away. Uh so, but then we still had to run the high power lines a few miles and and build a power plant in Mississippi. Um and and it was very difficult to build that. Um, and people don't understand like how much how much electricity do you actually need at the generator level at the generation level in order to power a data center because they look at the the noobs will look at the the the power consumption of uh say a GB300 and and multiply that by thing and then think that's the amount amount of power you need
**Dwarkesh Patel：** 发电厂制造商那里。
**Dwarkesh Patel:** all the cooling and everything. Wake up. Yeah. This is like that's a that's a that's a total noob. you've never done any hardware in your life before. Besides the GB300, you got to power all of the networking hardware. Um there's a whole bunch of CPU and storage stuff that's happening. Uh you you've got a size for uh your your peak uh cooling requirements. So that means uh can you cool even on the the worst hours, the worst day of the year? Well, it gets pretty freaking hot in Memphis. So, so you're going to have like a 40% increase on your your power just for cooling. Um, if assuming you don't want your data center to turn off on hot days and and want to keep going, then then you got to say, well, uh, um, there's there's another multiplicative element on top of that, which is, are you assuming that you're you you never have any hiccups in your power generation? Like, oh, well, actually, sometimes we have to take the generators, some of the power offline in order to service it. Oh, okay. Now you add another 20 25% multiplier on that because you you've got to you've got to assume that that you've got to take power offline to service it. Uh so the actual RS for roughly every every 110,000 GBs GB300's inclusive of networking uh CPU storage cooling uh margin for for for uh servicing power uh is roughly uh 300 megawatt. Sorry, say that again.
**Elon Musk：** 哦，我明白你在说什么了。
**Elon Musk:** It's it's it's roughly or think about it like the way you think about this like 330,000 to to to actually what you need at the gener generation level to service probably service 330,000 GB300s including all of the associated support networking and everything else and the and and the peak cooling and to have some margin some power margin reserve is roughly a gawatt. Can I ask a very naive question? Yeah. Um uh you know you're describing the engineering details of doing this stuff on Earth. Um but then there's analogous engineering difficulties of doing it in space. How do you do the um uh how do you replace infinite band with orbital lasers etc etc. How do you make it resistant to radiation? Um I don't know the details in the engineering but fundamentally what is the reason to think those challenges which have never been had to be addressed before will end up being easier than just like building more turbines on Earth. There's companies that build turbines on Earth. They can make more turbines, right?
**Dwarkesh Patel：** 你说的是燃气轮机积压的问题？
**Dwarkesh Patel:** I invite again try doing it and then you'll see. Um so um like the turbines are sold out through 2030. Have you guys considered making your own? I think in in order for in order to uh bring enough power online um I think uh SpaceX and and Tesla will probably have to make the turbine blades um the mains and blades uh internally
**Elon Musk：** 是的。你可以再往下挖一层。限制因素是涡轮机里的叶片——动叶和静叶，因为铸造涡轮机的叶片和静叶是一个非常专业的工艺过程，假设你用的是燃气发电的话。要扩展其他形式的发电非常困难。你可以潜在地扩展太阳能，但目前在美国进口太阳能的关税高得离谱，而国内的太阳能生产能力少得可怜。
**Elon Musk:** for just the blades or the turbines. Uh uh the the the the limiting factor you can get everything except the the blades they call the blades and veins. Um you can get that uh 12 to 18 months before the veins of blades. The limiting factor veins and blades and there are only uh three casting uh companies in the world that make make these and they're massively backlogged.
**Dwarkesh Patel：** 为什么不自己制造太阳能板？这感觉像是一个很适合 Elon 来解决的问题。
**Dwarkesh Patel:** Is this Seaman's GE those guys or is it a sub? No, it's it's it's it's other companies. I mean, sometimes they have a little bit of casting capability in house, but uh I'm just saying you can just you can just call any of the turbine makers and they will tell you. It's not top secret. They probably on the it's probably on the internet right now.
**Elon Musk：** 我们正准备自己造。好吧。SpaceX 和 Tesla 都在朝着每年 100 吉瓦太阳能电池产能的目标努力。
**Elon Musk:** If if it wasn't for the tariffs, would uh would Colossus be solar powered? Uh it would be much easier to make it solar powered. Yeah. Um the tariffs are nuts. Several hundred%. So
**Dwarkesh Patel：** 在产业链上做到多深？从多晶硅到晶圆再到最终的组件？
**Dwarkesh Patel:** don't you know some people we we also need speed. Yeah. you know, [laughter] you know, um, president has, you know, we don't agree on everything. Um, and, um, this administration is not not the biggest fan of of solar.
**Elon Musk：** 我觉得你必须从原材料到成品电池全部自己做。现在，如果是要送到太空的话，成本更低，做起来也更容易，因为不需要太多玻璃。不需要厚重的边框，因为不用抵御恶劣天气。太空没有天气。所以送到太空的太阳能电池其实比地面上的更便宜。
**Elon Musk:** Um, [laughter] but you and you also need the land, the permits and everything. So, if you're trying to move very fast,
**Dwarkesh Patel：** 有没有可能在 36 个月内把成本降到你需要的水平？
**Dwarkesh Patel:** um like I do think scaling solar on Earth is a is a good way to go, but but you need you do need some amount of time to find the land, get the permits, get the solar, uh pair that with the batteries.
**Elon Musk：** 太阳能电池已经非常便宜了。便宜到了滑稽的程度。我记得中国的太阳能电池大概在每瓦 0.25-0.30 美元左右，什么的。便宜得离谱。现在把它放到太空，又便宜了五倍。实际上，不是五倍，而是十倍，因为你不需要任何电池储能。所以一旦你的太空进入成本降下来了，迄今为止最便宜、最可扩展的 token 生成方式就是太空。没什么可比的。扩展起来容易一个数量级。关键是你在地面上根本没法扩展。就是不行。人们会在发电方面硬碰硬地撞墙。他们已经在撞了。xAI 团队为了上线一吉瓦的电力，不得不连续完成一系列奇迹。我们不得不把一大堆涡轮机拼在一起。然后在田纳西遇到了许可证问题，不得不跨越州界到密西西比，幸好只有几英里远。但我们仍然需要把高压输电线拉上几英里，在密西西比建发电厂。建那个非常困难。人们不理解在发电层面你实际需要多少电来给数据中心供电。因为新手会看一个 GB300 的功耗，然后乘以数量，以为那就是你需要的电力。还有所有冷却什么的。醒醒吧。那是纯新手，你一辈子都没碰过硬件。除了 GB300，你还得给所有网络硬件供电。还有一大堆 CPU 和存储在运转。你必须按照峰值冷却需求来设计。也就是说，你能在一年中最热的那一天最热的那一个小时进行冷却吗？孟菲斯可是热得要命。所以光冷却你就得多用 40% 的电力。这还是假设你不想让数据中心在热天关停，想持续运转的前提下。在此之上还有一个乘数——你是不是假设发电永远不会出问题？实际上，有时候我们不得不把某些发电设备下线进行维护。好了，现在再加上 20-25% 的乘数，因为你得假设需要把部分电力下线进行维护。所以我们的实际估算是：每 110,000 个 GB300——包括网络、CPU、存储、冷却、电力维护余量——大约需要 300 兆瓦。
**Elon Musk:** But why would it not work to stand up your own solar production and then you're right that you eventually run out of land, but there's a lot of land here in Texas. There's a lot of land in Nevada, including private land. It's not all publicly owned land. And so you'd be able to at least get the next Colossus and like the next one after that and at a certain point you hit a wall. But wouldn't that work for the moment? As I said, we are scaling solar production. Um there's there's a rate there's a rate at which you can scale physical production of solar solar cells. We we're going as fast as possible in scaling domestic production.
**Dwarkesh Patel：** 抱歉，再说一遍。
**Dwarkesh Patel:** You're making the solar cells at Tesla. Both Tesla and SpaceX um have a mandate to get to 100 gawatt a year of of solar. Speaking of the annual capacity, I'm curious in 5 years time, let's say, what will the installed capacity be on Earth?
**Elon Musk：** 要在发电层面支撑 330,000 个 GB300——包括所有配套的网络和其他一切，峰值冷却，以及保留一定的电力余量——大约需要一吉瓦。
**Elon Musk:** Long time and in space. Yeah, I deliberately picked five years cuz it's after your once we're up and running threshold. And so in five years time, yeah, what's the on earth versus in space installed AI capacity? five years I think probably if you say 5 years from now we're probably um AI in space will be uh launching every year the the sum total of all AI on earth in excess meaning 5 years from now my prediction is we will launch and and and be operating every year more AI in space than than this than the cumulative sort of total on earth which is I would expect to be at least sort of 5 years from now a few hundred gawatts per year of uh of AI in space and rising. Um so you can get to I think you on Earth you can get to around a terowatt a year of of AI in space um before you start having you know fuel supply challenges for the rocket.
**Dwarkesh Patel：** 我能问一个很天真的问题吗？你在描述在地球上做这些事的工程细节。但在太空做也有类似的工程难题。你怎么用轨道激光来替代无限带宽，等等等等？怎么让它抗辐射？我不知道工程细节，但从根本上说，有什么理由认为这些从未被解决过的挑战会比在地球上多建几个涡轮机更容易？有公司在地球上建涡轮机。他们可以多造一些涡轮机，对吧？
**Dwarkesh Patel:** Okay. You think you can get to hundreds of gigawatts per year in 5 years time? Yes.
**Elon Musk：** 同样的话，你试试看就知道了。涡轮机到 2030 年都卖光了。
**Elon Musk:** So 100 gawatt depending on the um specific power of the whole system with solar arrays and radiators and everything is um is on the order of like 10,000 Starship launches. Yes.
**Dwarkesh Patel：** 你们有没有考虑过自己造？
**Dwarkesh Patel:** Um and you want to do that in one year. And so that's like one Starship launch every hour.
**Elon Musk：** 为了上线足够的电力，我觉得 SpaceX 和 Tesla 可能不得不自己内部制造涡轮叶片，动叶和静叶。
**Elon Musk:** Yeah. That's happening in this city. Like walk me through a world where there's 10 there's a Starship launch every single hour.
**Dwarkesh Patel：** 只是叶片还是整个涡轮机？
**Dwarkesh Patel:** Yeah. I mean that's actually a lower rate compared to airlines. Uh like like aircraft aircraft
**Elon Musk：** 限制因素——除了叶片之外的所有东西你都能在 12 到 18 个月内拿到。限制因素就是动叶和静叶。全世界只有三家铸造公司在做这个，而且它们的订单积压得很厉害。
**Elon Musk:** there's a lot of airports a lot of airports but
**Dwarkesh Patel：** 是 Siemens、GE 这些公司，还是某个子公司？
**Dwarkesh Patel:** and you got to launch you know the polar orbit.
**Elon Musk：** 不，是其他公司。有时候它们内部有一点点铸造能力。但我只是说，你随便打电话给任何一家涡轮机制造商，他们都会告诉你。这不是什么最高机密。可能现在互联网上就能查到。
**Elon Musk:** Uh no it doesn't have to be polar but you you just there's there's some some value to sunsynchronous but um but I I think actually um you just go high enough you you start getting out of earth shadow you know. So, um, how many physical Starships are needed to do 10,000 launches a year?
**Dwarkesh Patel：** 如果没有关税的话，Colossus 会用太阳能供电吗？
**Dwarkesh Patel:** I I don't think we'll need more than I mean, you you could you could uh probably do it with as as few as like 20 or 30. Um,
**Elon Musk：** 用太阳能供电会容易得多，对。关税高得离谱，好几百个百分点。
**Elon Musk:** like it really depends on how quickly does a ship the ship has to go around the Earth. Um, and the ground track for for the ship has to come back over the launch pad. So, if you can use a ship every say 30 hours, uh you could do it with 30 ships, but but we'll we'll make more ships than that. But, um but but SpaceX is is um is going up to do 10,000 launches a year and I'll and and maybe even 20 or 30,000 launches a year. Is the idea to become basically a hyperscaler, become an oracle and lend this capacity to other people? What's what are you going to do with presumably SpaceX is the one launching all this? So SpaceX is going to hyperscaler
**Dwarkesh Patel：** 你不是认识一些人吗？总统他……
**Dwarkesh Patel:** hyper hyper. [laughter]
**Elon Musk：** 我们并不是什么都同意，而且这届政府不是太阳能的最大粉丝。我们还需要土地、许可证，以及所有这些东西。所以如果你想快速推进，我确实觉得在地球上扩展太阳能是个好路子，但你确实需要一些时间来找地、拿许可、搞到太阳能板，再配上电池。
**Elon Musk:** Yeah. I mean if ass if assuming my predictions come true SpaceX will launch more AI than the cumulative amount on Earth combi of everything else combined. Is this mostly inference or
**Dwarkesh Patel：** 为什么不能自建太阳能产能呢？你说得对，最终会用完土地，但德州这里有大量土地。内华达也有大量土地，包括私人土地。不全是公有土地。所以你至少能搞下一个 Colossus 和再下一个。到某个节点你会撞墙。但眼下这不行吗？
**Dwarkesh Patel:** most AI will be inference like already inference for the purpose of training is most training. And there's a narrative that the the change in discussion around a SpaceX IPO is because previously SpaceX was very capital efficient. Just it wasn't that expensive to develop that. Even though it sounds expensive, it's actually very capital efficient in how it runs. Whereas now you're going to need more capital than just can be raised in the private markets. Like if the private markets can accommodate raises of, as we've seen from the AI labs, tens of billions of dollars, but not beyond that, is it that you'll just need more than tens of billions of dollars per year and that's about to take it public?
**Elon Musk：** 正如我说的，我们正在扩展太阳能产能。太阳能电池的实际生产扩展有一定速率。我们正在尽可能快地扩大国内产能。
**Elon Musk:** Um, yeah. I have to be careful about saying things about companies that might go public. Um, you know, if you make general state, [laughter] if you make that's never been a problem for you, Elon,
**Dwarkesh Patel：** 你是在 Tesla 制造太阳能电池？
**Dwarkesh Patel:** you know, there's a price to pay for these things. make some general statements for us about the depth of the capital markets between public and private markets.
**Elon Musk：** Tesla 和 SpaceX 都有一个目标，就是达到每年 100 吉瓦的太阳能产能。
**Elon Musk:** Yeah, there's there's a lot more capital in the very general [laughter] there's there's obviously a lot more capital available in the public markets than private. I mean it might be it's at least at least it might be 100 times more capital but it's at least well way more than 10. But isn't it also the case that things that tend to be very um capital intensive if you look at say real estate as you know a huge industry uh that raises a lot of money each years at an industry level that tends to be debt financed because by the time you're deploying that much money you actually have a pretty
**Dwarkesh Patel：** 说到年产能，我很好奇，比如说五年后，地球上的装机容量会是多少……
**Dwarkesh Patel:** you have a clear revenue stream.
**Elon Musk：** 五年是很长的时间。
**Elon Musk:** Exactly. And and a near-term return. And you see this even with the data center buildouts which are famously being you know uh financed by the uh the private credit industry. And so why not just debt finance? Um speed is important. So um I'm genally going to do the thing that um I'm I'm I mean I just repeatedly tack the limiting factor. Whatever the limiting factor is on speed, I'm I'm going to tackle that. So um there's uh if if capital is the limiting factor then I'll I'll sold for capital. If if it's not limiting factor I'll s for something else.
**Dwarkesh Patel：** ……太空呢？我故意选五年，因为那是在你说的"一旦运转起来"这个门槛之后。所以五年后，地球上和太空里的 AI 装机容量分别是多少？
**Dwarkesh Patel:** U based on your statements about um Tesla and being public. I wouldn't have guessed that you thought the fast the way to move fast is to be public.
**Elon Musk：** 如果说五年后，我认为太空中每年发射和运营的 AI 算力可能会等于地球上所有 AI 算力的总和。也就是说，五年后，我的预测是我们每年会在太空中发射和运行比地球上累计总量更多的 AI 算力。我预计五年后至少达到每年几百吉瓦的太空 AI 算力，并且还在增长。我认为你可以达到大约每年一太瓦的太空 AI 算力，之后才会开始遇到火箭燃料供应的挑战。
**Elon Musk:** Normally I would say yeah that's that's true. Um, like I said, I mean, I'd love to, you know, talk about some more in detail, but the problem is like if you talk about poly companies before they become public, you get into trouble and then you have to delay your offering
**Dwarkesh Patel：** 好的，但你认为五年内能达到每年几百吉瓦？
**Dwarkesh Patel:** and then you
**Elon Musk：** 是的。
**Elon Musk:** and as we said, we're solving for speed. Yes. Exactly. So, so, so that you you can't hype companies um that are that may that might go public. So, that that's that's why we have to be a little careful here. Um, but but but I I I we can't talk about physics. Um so like the way the way you think about scaling long term is that um uh earth only receives about uh half a billionth of the sun's energy. Um and the sun is the sun is essentially all the energy. This is a very important point to appreciate because sometimes people will talk about marginal nuclear reactors or any you know various like fusion on earth. Um but but you have to step back a second and say if if if you're going to climb the cartes scale and have some non-trivial and and harness some non-trivial percentage of the uh the sun's energy like let's say you wanted to uh harness a millionth of the sun's energy which sounds pretty small. um that that would be um about call it roughly 100,000 times more electricity than we currently generate on Earth of for all of civilization uh give or take an order of magnitude. Um so it it obviously the only way to scale uh is to go to space with solar. Uh from launching from Earth you can get to about a terowatt per year. Um beyond that you want to go you you want to uh launch from the moon. and you want to have a a mass driver on the moon. Uh and that mass driver on the moon, you could do probably a pedawatt per year.
**Dwarkesh Patel：** 那 100 吉瓦，取决于整个系统——包括太阳能阵列、散热器等——的具体功率，大约是一万次 Starship 发射的量级。
**Dwarkesh Patel:** Um when you're talking these kinds of numbers, you know, terowatts of compute, um presumably whether you're talking land or space far before this point, um you've like run into, you know, you actually need maybe you don't the solar panels are more efficient, but you still need the chips.
**Elon Musk：** 是的。
**Elon Musk:** Uh you still need the logic and the memory and so forth and need [clears throat] to build a lot more chips and make them much cheaper, right? And so how are we getting a terowatt of uh like right now the world is going to be 20 25 gawatt of compute? Um how are we getting a terowatt of logic by 2030?
**Dwarkesh Patel：** 你想在一年内完成这些。那就是差不多每小时发射一次 Starship。这发生在这个城市里？给我描绘一下每小时发射一次 Starship 的世界是什么样的。
**Dwarkesh Patel:** I guess we're going to need some very big chip apps.
**Elon Musk：** 我的意思是，跟航空公司相比，这实际上是更低的频率。飞机比这多得多。
**Elon Musk:** Tell tell me about it. [laughter] I've mentioned publicly that uh the idea of doing a sort of a a terab tering the new gig. We I feel like the naming scheme of Tesla, which has been very um catchy, is like you looking at like the metric [laughter]
**Dwarkesh Patel：** 机场也多得多啊。而且你得往极地轨道发射。
**Dwarkesh Patel:** the metric scale. Um at what level of the stack are you are you building the clean room and then partnering with an existing um fab to get the process technology and buying the tools from them? What what is the plan there?
**Elon Musk：** 不，不一定是极地轨道。太阳同步轨道有一定价值，但我觉得其实如果飞得够高，你就能脱离地球的影子。
**Elon Musk:** Well, you can't partner with existing fabs because uh they just they can't output enough their chip volume is too low. But you have you have to
**Dwarkesh Patel：** 要做一万次发射一年，需要多少艘实体 Starship？
**Dwarkesh Patel:** before the process technology. Yeah. Partner for the IP.
**Elon Musk：** 我不觉得需要很多……可能用 20 或 30 艘就行了。这取决于周转速度……飞船必须绕地球一圈，地面轨迹得重新经过发射台。所以如果你每艘船大约 30 小时能复用一次，用 30 艘船就行了。但我们会造更多。SpaceX 正在为每年一万次发射做准备，甚至可能是两三万次。
**Elon Musk:** Um, you know, the the fabs today all basically use um machines from like five companies. Yeah.
**Dwarkesh Patel：** 你的想法是基本上变成一个超级云服务商，变成 Oracle，然后把这个算力租给别人？大概 SpaceX 是负责发射所有这些的。那 SpaceX 要变成一个超级云服务商？
**Dwarkesh Patel:** You know, so you've got ASML, Tokyo, Electron, Kelly, Tankor, you know, um, etc. So, um, so, so at first I think you'd have to get equipment from them and then, uh, modify it or work with them to increase the volume. Um, but I think you'd have to build perhaps in a different way. Um, so I think the logical thing to do is to, uh, to use conventional equipment in unconventional way to get to scale. Uh and then uh and then and then start modifying the equipment uh to increase the the rate
**Elon Musk：** 超级的超级。如果我的一些预测成真的话，SpaceX 发射的 AI 算力会超过地球上所有其他一切的累计总量。
**Elon Musk:** kind of boring company style. Yeah. Kind of like Yeah. You you sort of buy an an existing uh boring machine and then uh figure out how to dig tunnels in the first place and then design a much better machine uh that's you know I don't know
**Dwarkesh Patel：** 这主要是推理（inference）还是？
**Dwarkesh Patel:** some orders of magnitude faster.
**Elon Musk：** 大部分 AI 会是推理。实际上，用于训练目的的推理已经是训练的主体了。
**Elon Musk:** Here's a very simple lens. We can categorize technologies and how hard they are. And one categorization could be look at things that China has not succeeded in doing. And if you look at Chinese manufacturing, still behind on leading edge chips and still behind on uh leading edge turbine engines and things like that. And so does the fact that China has not successfully replicated TSMC give you any pause about the difficulty or you think well that's not true for some reason.
**Dwarkesh Patel：** 有一种说法是 SpaceX IPO 讨论出现变化的原因是，以前 SpaceX 资金效率非常高。开发成本虽然听起来贵，但实际上运营时资金效率很高。而现在你将需要比私人市场能募集到的更多的资本。私人市场可以容纳——正如我们在 AI 实验室身上看到的——几百亿美元的融资，但不能超过那个量级。是不是因为你每年需要的资本远超几百亿美元？这就是你要上市的原因？
**Dwarkesh Patel:** Uh it's not that they have not replicated TSMC they have not replicated ASML that's the limiting factor.
**Elon Musk：** 关于可能上市的公司，我说话得小心。
**Elon Musk:** So so you think it's just the um the sanctions essentially. Uh yeah China would be outputting vast numbers of chips at they could buy
**Dwarkesh Patel：** 这对你来说从来都不是个问题，Elon。
**Dwarkesh Patel:** but couldn't they up to relatively recently buy them? No.
**Elon Musk：** 说这些话是有代价的。给我们做一些关于公开市场和私人市场资本深度的笼统描述吧。公开市场上有多得多的资本可用……
**Elon Musk:** Okay. That that ASML bands have been in place for a while.
**Dwarkesh Patel：** 非常笼统的。
**Dwarkesh Patel:** Okay.
**Elon Musk：** 公开市场上的资本显然远多于私人市场。可能是 100 倍，但肯定超过 10 倍。
**Elon Musk:** So, but I I think T's going to be make start making pretty compelling chips on three or four years. Would you consider making to ASML machines? I I don't know. I don't know yet is the right answer. So I um it's just that that if to produce at high volume and to to to reach large volume in say 36 months to match the the rocket to payload to orbit. So if we're doing a million tons to orbit um in like let's say I don't know 3 or four years from now something like that. um that and uh and and we're doing 100 kow per ton. So that that means we need um at least 100 gawatts per year of solar. Um and we'll need uh an equivalent amount of of chips to to you know that you need 100 gawatt worth of chips. You you're going to match these things the master orbit.
**Dwarkesh Patel：** 另外，那些非常资本密集型的领域——比如房地产作为一个巨大的行业，每年在行业层面募集大量资金——它们往往是债务融资的，因为当你要部署那么多钱的时候，你实际上有一个相当——
**Dwarkesh Patel:** Yes.
**Elon Musk：** 你有清晰的收入流。
**Elon Musk:** The the power generation and the uh and the and the chips. uh and and and I'd say my biggest concern actually is is memory. Um so the I think there's there's a the the path to creating logic chips is more obvious than the path to um having sufficient memory to support logic chips. That's why you see your DDR prices going in and these memes about like um you know um you're marooned on a desert island. You write help me on the sand. Nobody comes. You write DDRM ships come swarming in. [laughter]
**Dwarkesh Patel：** 没错，而且是短期回报。你甚至在数据中心建设中也能看到这一点，众所周知那些项目是由私人信贷行业在融资。为什么不直接用债务融资？
**Dwarkesh Patel:** I haven't seen that.
**Elon Musk：** 速度很重要。我总体上会去做那个……我就是反复地解决限制因素。不管限制因素是什么，我就去解决它。如果资本是限制因素，那我就解决资本问题。如果不是限制因素，我就解决别的东西。
**Elon Musk:** Uh I love your manufacturing philosophy around um around fabs. You know I know nothing about the topic but I don't know how to build a fab yet. I figure it out. [laughter]
**Dwarkesh Patel：** 根据你之前关于 Tesla 上市的表态，我没想到你会认为快速推进的方式是上市。
**Dwarkesh Patel:** But
**Elon Musk：** 通常来说，我会同意你的看法。如我所说，我想更详细地讨论这个，但问题是如果你在公司上市前讨论太多，你会惹麻烦，然后不得不推迟发行。而且如你所说，你在求速度。
**Elon Musk:** obviously I have difficult. It sounds like you think the the sort of like the processing knowledge of like these 10,000 PhDs in Taiwan who know exactly what gas goes in the plasma chamber and what settings to put on the tool. You can just like delete those parts of those steps. Like fundamentally it's get the clean room, get the tools and figure it out. I don't think it's PhDs that it's it's mostly people with uh you know who are not not PhDs um that that most engineering is done with people who don't have PhDs. Do you guys have PhDs?
**Dwarkesh Patel：** 是的，没错。
**Dwarkesh Patel:** No.
**Elon Musk：** 你不能炒作可能要上市的公司。所以我们得在这里稍微小心些。但我们可以聊物理。你思考长期扩展的方式是，地球只接收到大约五十亿分之一的太阳能量。太阳基本上就是所有的能量。这是一个非常重要的认知。因为有时候人们会谈论模块化核反应堆或者各种地面聚变。但你得退一步想一想，如果你要攀爬卡尔达肖夫指数（Kardashev scale），利用太阳能量的某个不可忽略的百分比……比如说你想利用太阳能量的百万分之一，这听起来很小。那大约是，我们粗略估计一下，大约是目前全人类文明所产电力的 100,000 倍。量级上可能有个偏差。显然，唯一的扩展方式就是去太空用太阳能。从地球发射的话，你一年大概能达到一太瓦。超过那个量级，你就需要从月球发射了。你需要在月球上建一个质量投射器（mass driver）。用月球上的那个质量投射器，你大概能达到每年一拍瓦。我们在说的就是这些数字，太瓦级的算力。
**Elon Musk:** Okay. We [laughter] also we also haven't successfully built any fab so you shouldn't be coming to us for your fab advice
**Dwarkesh Patel：** 大概不管是在地面还是太空，远远在到达这个数量级之前，你就会遇到……也许太阳能板效率更高了，但你仍然需要芯片。你仍然需要逻辑芯片和内存等等。你需要制造多得多的芯片，并且让它们便宜得多。现在全球大概有 20-25 吉瓦的算力。到 2030 年我们怎么弄到一太瓦的逻辑运算能力？
**Dwarkesh Patel:** or
**Elon Musk：** 我猜我们需要一些非常大的芯片工厂。说起这个。我之前公开提过做一个"太级工厂"（TeraFab）的想法，Tera 是新的 Giga。
**Elon Musk:** I don't think you need PhD for that for the stuff. So um but but you do need you do need competent personnel. Um so I I don't I mean like like right now if um you know say like Tesla's pedals to the metal max production of going as fast as possible to get uh AI5 Tesla AI5 chip design um uh into production and then reaching scale. Um you know that'll probably happen you know around the second quarterish of next year hopefully. Um uh and then AI6 would hopefully follow less than a year later. Um but um and and and and we've secured all the all the trip fab production that we can. Yes. You're currently limited on TSMC fab capacity.
**Dwarkesh Patel：** 我感觉 Tesla 的命名方案一直很有吸引力，就是你在看公制单位的级别。你在堆栈的哪一层？你是在建无尘室然后跟现有晶圆厂合作获取制程技术、从他们那买设备？那里的计划是什么？
**Dwarkesh Patel:** Yeah. Um and and and we'll be using TSMC uh Taiwan, uh Samsung Korea, TSMC Arizona, Samsung Texas. Um and we still booked out all the Yeah.
**Elon Musk：** 嗯，你没法跟现有晶圆厂合作，因为他们的产出不够。芯片的产量太低了。
**Elon Musk:** Yes. And and then and then and then if I ask uh TSMC or Samsung, okay, what what's the time frame to get to volume production? The point is not is it's not you've got to you've got to build the fab and you got to you got to start production, then you got to climb the yield curve and reach volume production at high yield. that that that from start finish is a 5year period. And so the limiting factor is chips.
**Dwarkesh Patel：** 但制程技术呢？为 IP 而合作。
**Dwarkesh Patel:** Yeah. Like limiting factor once you can get to space is chips, but the limiting limiting factor before you can get to space will be power.
**Elon Musk：** 今天的晶圆厂基本上都用来自大约五家公司的设备。有 ASML、Tokyo Electron、KLA-Tencor 等等。所以一开始，我觉得你得从他们那里拿设备，然后改造它或者跟他们合作来增加产量。但我觉得你可能得用一种不同的方式来建。合乎逻辑的做法是用传统设备以非传统的方式来达到规模，然后再开始改造设备以提高速率。
**Elon Musk:** Why don't you do the Jensen thing and just prepay TSMC to build more fabs for you? Uh I I've already told them that.
**Dwarkesh Patel：** Boring Company 的路子。
**Dwarkesh Patel:** But they won't take your money. Like what's going on?
**Elon Musk：** 对。你先买一台现有的隧道掘进机，然后搞清楚怎么挖隧道，然后再设计一台好得多的机器，快上几个数量级。
**Elon Musk:** They're building fabs as fast. No, [laughter] they're building they're building fabs as fast as they can. Um, and so is Samsung. Like like they're they're pedal to the metal. I mean, they're going, you know, balls to wall, you know, as fast as they can. So, still not fast enough. I mean, like said, there will be I think um if you say uh I think towards the end of this year, I think probably chip production will outpace the ability to turn chips on. Uh but once you can get to space and unlock the um the power constraint and you can now do you know hundreds of gigawatts per year of power in space um again bearing in mind that average power usage in the US is you know 500 gaw so if you're launching say 200 gawatt a year to to space you're sort of lapping the US every two and a half years the entire all US electricity production this is a very huge amount um so Um but but but between now and then uh the the actually the the constraint for for for server side compute uh concentrated compute will be will be electricity. My my guess is that we start hitting the people start getting for where they can't turn the chips on for for for large clusters uh towards the end of this year. They're just the chips are going to be piling up and and not be won't be able to be turned on. Now for edge computers a different story. So if the if if like for for Tesla the the so the AI5 chip is going into our Optimus robot you know optimistic um and and so if you have an AI edge compute that's distributed power now the power is distributed over a large area it's not concentrated um and if you can charge at night you can actually um [snorts] uh use the grid much more effectively because the the actual peak power production in the US is over 1,000 gawatt. Uh but the average power usage because the dayight cycle is 500. So if you can charge at night, there's an incremental 500 gaw that you can um generate you know at night. Um so that that's why Tesla for edge compute is not constrained and we can make a lot of ships uh to make you know very large number of robots and cars. Uh, but if you try to concentrate that compute, you're going to have a lot of trouble turning it on.
**Dwarkesh Patel：** 这里有一个很简单的视角。我们可以对技术进行分类看它们有多难。一种分类方式是看中国没能成功做到的事情。如果你看中国的制造业，他们在前沿芯片上仍然落后，在前沿涡轮发动机等方面仍然落后。那么中国没有成功复制 TSMC 这个事实有没有让你对其难度有所犹豫？还是你觉得这个判断出于某种原因不成立？
**Dwarkesh Patel:** What I find remarkable about the SpaceX business is the end goal is to get to Mars, but you keep finding ways on the way there to keep generating incremental revenue to get to the next stage and the next stage. So, the Falcon 9 is Starlink
**Elon Musk：** 不是说他们没有复制 TSMC，而是他们没有复制 ASML。那才是限制因素。
**Elon Musk:** and now for Starship, it's going to be potentially orbital data centers. Um but like you find these like um you know sort of infinitely uh elastic sort of marginal use cases of your like next rocket and your next rocket and next scale up. You can see how this might seem like a simulation to me. [laughter]
**Dwarkesh Patel：** 所以你觉得本质上就是制裁的原因？
**Dwarkesh Patel:** Well,
**Elon Musk：** 对，如果中国能买到 2-3 纳米的设备，他们早就大量生产芯片了。
**Elon Musk:** or am I someone's avatar in a video game or something because it's like like what are the odds that all these crazy things would be happening? I I I mean I mean I mean rockets and chips and robots and space solar power and and not to mention the the mass driver on the moon. I really want to see that. You can imagine like some mass driver that's just like just it's like sending AI solar powered AI satellites into space like one after another like these like at at 2 and a half kilometers per second. you know, that's uh and just shooting them into deep space. That would be a sight to see. I I I mean, I'd watch that just like a live stream of
**Dwarkesh Patel：** 但他们不是直到相对最近还能买到吗？
**Dwarkesh Patel:** Yeah. Yeah. Just one after another just shooting
**Elon Musk：** 不。
**Elon Musk:** webcam uh AI satellites in deep space, you know, a billion or 10 billion tons a year.
**Dwarkesh Patel：** 好吧。
**Dwarkesh Patel:** And sorry, you manufacture the satellites on the moon. I see. So you send the raw materials to the moon and then manufacture them there and then
**Elon Musk：** ASML 的禁令已经实施有一段时间了。但我认为三四年内中国将会生产出相当有竞争力的芯片。
**Elon Musk:** well the the lunar soil is uh I think it's like 20% solar 20% silicon or something like that. So so you can get the silicon from the you can mine the silicon on the moon refine it um
**Dwarkesh Patel：** 你会考虑自己制造 ASML 的机器吗？
**Dwarkesh Patel:** and generate the and create the solar panels or the solar cells and the radiators on the moon.
**Elon Musk：** "我还不知道"才是正确答案。要在大约 36 个月内达到大规模产量，以匹配火箭送入轨道的载荷……如果我们大约三四年后每年能向轨道运送一百万吨……我们按每吨 100 千瓦算。这意味着我们每年至少需要 100 吉瓦的太阳能。我们需要相当数量的芯片。你需要 100 吉瓦的芯片。你得让这些东西配套：送入轨道的质量、发电能力和芯片。我觉得我最大的担忧其实是内存。制造逻辑芯片的路径比拥有足够的内存来支撑逻辑芯片的路径更清晰。这就是为什么你看到 DDR 价格疯涨，还有那些梗图。你被困在一个荒岛上。你在沙子上写"救命"。没人来。你写"DDR RAM"。船蜂拥而至。
**Elon Musk:** Yeah. So um make the radiators out of aluminum. So there's there's plenty of silicon and aluminum on the moon to uh to make the the cells on the and the radiators. Um, the chips you could send from Earth because they're pretty light. Um, but maybe at some point you make them on the moon, too. I'm just saying like these are simply it's kind of like, like I said, it it does seem like a sort of a a video game situation where it's difficult but not impossible to get to the next level. um like I I I don't see any way that you could do um you know uh you know 500 to a,000 terowatts per year launch from Earth. [snorts] U
**Dwarkesh Patel：** 我很想了解你关于晶圆厂的制造理念。
**Dwarkesh Patel:** I agree [laughter]
**Elon Musk：** 关于这个话题我一无所知。我还不知道怎么建晶圆厂。我会搞清楚的。显然，我从来没建过晶圆厂。
**Elon Musk:** but you could do that from the moon. Okay, let me tell you how I ended up using Mercury for my personal banking. So last year I had the opportunity to make an investment that I was very excited about, but it came up a bit last minute. And so I had to wire over a lot of money for my personal account very fast. But my personal bank at the time wouldn't let me make this wire transfer online. And I called them a bunch of times. They just couldn't make it work. They told me that I'd have to go to the nearest Inerson branch, which was in Dallas. And for a moment, I even considered flying from SF to Dallas to make this transfer happen last minute. But then I remembered that Mercury, which I use for my business banking, had just started rolling out personal accounts. So I emailed support with a quick rundown of the situation. And within 2 hours, I had successfully wired the investment from my new personal Mercury account. Since then, I've moved over the rest of my personal money from my previous bank to Mercury, and that's made a bunch of things, even little things like setting up auto transfer rules between my checkings and savings account, a whole lot better. Visit mercury.com/personal to get started. Mercury is a fintech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and column NA members FDIC. Can can I can I zoom out and ask about the SpaceX mission? So, I think you said like we got to get to Mars so we can make sure that if something happens to Earth, [snorts] you know, civilization consciousness, etc. arrives. Yes.
**Dwarkesh Patel：** 听起来你觉得台湾那一万个博士的制程知识——他们确切知道等离子腔里该通什么气体、工具该调什么参数——你可以直接跳过那些步骤。从根本上说，就是搞到无尘室，搞到设备，然后自己摸索。
**Dwarkesh Patel:** Um,
**Elon Musk：** 我不觉得这需要博士。大多数并不是博士做的。大多数工程是由没有博士学位的人完成的。你们有博士学位吗？
**Elon Musk:** by the time you're sending stuff to Mars, like Grock is on that ship with you, right? Right. And so if Grock's gone Terminator, like the main risk you're worried about, which is AI, why doesn't that follow you to Mars? Uh well, I'm not sure AI is the main risk I'm worried about. I mean, the important thing is that uh consciousness uh which I think arguably most consciousness or most intelligence certainly consciousness is more of a debatable thing. Most intell the vast majority of intelligence in the future will be um AI. Um so um you know AI AI will exceed uh you say like how many what's the how much how many I don't know pedawatts of intelligence will be uh silicon versus biological and and and basically humans will be a very tiny percentage of all intelligence in the future if current trends continue. Um anyways as as long as like I think there's intelligence ideally ideally also which includes human intelligence and consciousness propagated into the future that's a good thing. So you want to take the set of actions that maximize the probable uh light cone of of of consciousness. So just and intelligence
**Dwarkesh Patel：** 没有。
**Dwarkesh Patel:** just to be clear it's a the mission of SpaceX is that even if something happens to the humans the AIs will be on Mars and like the AI intelligence will continue the light of our journey.
**Elon Musk：** 好的。
**Elon Musk:** Yeah I mean I'm very prohuman so it's not I I want to make sure we take sort of actions that ensure that humans are along for the ride. You know we're at least there. Yeah. Um but the I'm just saying the total amount of intelligence uh like I think maybe in in five or six years um AI will exceed the sum of all human intelligence and then if that continues at some point human intelligence will be less than 1% of all intelligence. What what should our goal be for such a civilization? Is the idea that a small minority of humans still have control of the AIs? Is the idea of some sort of like just trade but no control? How should we think about the relationship between the vast stocks of AI population versus human population
**Dwarkesh Patel：** 我们也没有成功建过任何晶圆厂，所以你不应该找我们要建厂建议。
**Dwarkesh Patel:** in the long run? I think I I I don't it's difficult to imagine that if humans have say 1% of the intelligence of combined intelligence of artificial intelligence that that that humans will be in charge of AI. Um, I think what we can do is make sure it has um that AI has values that that are um that that cause intelligence to be propagated uh into the universe. Um so the the the reason for XI XI's mission is understand the universe. So now that's actually very important. So you say well what things are necessary to understand the universe? Well, you have to be curious and you have to exist. you can't just can't understand the universities don't exist. Um so you actually want to increase the amount of intelligence uh in the universe increase the palable lifespan of intelligence the scope and scale of intelligence. Um I think actually also as a coral you corly you have um humanity also uh continuing to expand because um if you're if you're curious you're trying to understand the universe one thing you're trying to understand is where will humanity go
**Elon Musk：** 我不觉得那些东西需要博士。但你确实需要有能力的人。现在，Tesla 正在全力以赴，尽可能快地推进 Tesla AI5 芯片设计量产，然后实现规模化。那大概会在明年第二季度左右发生，希望如此。AI6 希望不到一年后跟进。我们已经锁定了我们能拿到的所有芯片代工产能。
**Elon Musk:** and so I think understand the universe actually means you would care about uh propagating humanity into the future um and uh so so that's that's why I think I think our mission station is profoundly important Um I'm not sure to the degree that Grock adheres to that mission statement um I I think the future will be very good. I I want to ask about how to make Grock adhere to that mission statement. But at first I want to understand the mission statement. Um so it's there's it's there's understanding the universe. They're spreading intelligence and they're spreading humans. Um all three seem like distinct vectors.
**Dwarkesh Patel：** 是的。但你目前受限于 TSMC 的代工产能。
**Dwarkesh Patel:** Okay. Well, I'll tell you why I why I think they are that that that understanding the universe encompasses all of all of those things. Go ahead.
**Elon Musk：** 对。我们会用 TSMC 台湾、Samsung 韩国、TSMC 亚利桑那、Samsung 德州。我们仍然——
**Elon Musk:** Um, you can't have understanding without I think you can't have understanding without intelligence and and I think without consciousness. Um, so you in order to understand universe, you have to expand this the the scale and and probably the scope of of intelligence different types of intelligence. I guess from a humanentric perspective like for humans in comparison to chimpanzees, humans are trying to understand the universe. They're not like expanding chimpanzeee footprint or something, right?
**Dwarkesh Patel：** 你把所有产能都预订了。
**Dwarkesh Patel:** We're also we're also not well we're not we actually have made protected zones for chimpanzees. Um and even though we could humans could exterminate all chimpanzees, we've not we've chosen not to do so.
**Elon Musk：** 是的。我问 TSMC 或 Samsung，"好的，达到量产的时间表是什么？"关键是，你得建晶圆厂，然后开始生产，然后爬良率曲线，在高良率下达到量产。从开始到结束，那是一个五年的周期。所以限制因素是芯片。一旦你能去太空了，限制因素就是芯片，但在能去太空之前，限制因素是电力。
**Elon Musk:** Do you think that's a basic scenario for humans in the post AGI world? Um I I I think uh I think AI with the right values I think Grock would care about expanding uh human civilization. I'm gonna certainly emphasize that. Hey, Gragas, your daddy. Don't forget to expand human consciousness. Uh like I I actually I think if if probably like uh like the Yan Banks culture books are the closest thing to what what the future will be like in a you know non-dystopian outcome. Um so I so understand the universe it means you have to be very you have to be truth seeking as well. You like truth has to be absolutely fundamental because you can't understand the universe if you live if you're delusional. You you'll simply think you understand understood the universe but you will not. So so being rigorously truth seeeking is is absolutely fundamental to understanding the universe. You're not going to discover new physics or or invent technologies that work um unless you're rigorously truth seeeking. How do you make sure that Grock is regressively truth seeeking as it gets smarter?
**Dwarkesh Patel：** 为什么不像 Jensen 那样，直接预付给 TSMC 让他们为你多建几个晶圆厂？
**Dwarkesh Patel:** Uh I think you you need to make sure that that that Grock um is says things that are correct not politically correct. I think it's the elements of cogency. So you want to make sure that that the axioms are as close to true as possible that that you don't have contradictory axioms. um that the um the conclusions necessary necessarily follow from those axioms with with the right probability. It it's just it's just it's critical thinking 101. I I think at least trying to do that is better than not trying to do that.
**Elon Musk：** 我已经跟他们说了。
**Elon Musk:** Yeah. And the proof will be in the pudding. If if like I said for any AI to discover new physics or invent technologies that actually work in reality and there's no bullshitting physics. So it's like you can you know you can um you can break a lot of laws but you can't like your physics is law everything else is is a recommendation like in order to make a technology that works you have to be extremely truth seeeking because otherwise you'll test that technology against reality um and if you make for example an an error in your rocket design the rug will blow up um or the car won't work or the you know
**Dwarkesh Patel：** 但他们不收你的钱？怎么回事？
**Dwarkesh Patel:** but but there there are a lot of um communist Soviet physicists who or like scientists discovered new physics. There are German Nazi physicists who discovered new uh science. Um it seems possible to be like really good at discovering new science and be really truth seeeking in that one particular way. And still we'd be like well I don't want I don't want the communist scientist to like become more and more powerful over time. Um and so those seem like yeah we could have we can imagine a future version of gra that's like really good at physics um and being really truth seeking there that doesn't seem like a universally uh alignment inducing behavior.
**Elon Musk：** 他们已经在尽可能快地建了。Samsung 也是。他们全力以赴。他们在拼命干，尽可能快。还是不够快。正如我说的，我觉得今年年底左右，芯片产量可能会超过给芯片通电的能力。但一旦你能去太空，解锁了电力约束，你现在每年就能在太空获得几百吉瓦的电力。再次提醒一下，美国的平均电力使用量是 500 吉瓦。所以如果你每年往太空发射比如 200 吉瓦，大约每两年半就超过美国一次。美国的全部电力产能——这是一个非常巨大的数字。从现在到那时，服务器端算力、集中式算力的约束将是电力。我的估计是，今年年底人们就会开始发现，对于大规模集群，他们没法给芯片通电了。芯片会堆积起来，但没法通电。对于边缘计算（edge compute）是另一回事。对 Tesla 来说，AI5 芯片要装进我们的 Optimus 机器人。如果你有 AI 边缘计算，那是分布式的电力。电力分散在很大的区域。不是集中的。如果你能在夜间充电，你实际上可以更有效地利用电网。因为美国实际的峰值发电量超过 1,000 吉瓦。但平均用电量因为昼夜循环只有 500。所以如果你能在夜间充电，就有额外的 500 吉瓦可以在夜间发电。所以这就是为什么 Tesla 在边缘计算方面不受约束。我们可以制造大量芯片来造大量的机器人和汽车。但如果你试图把那些算力集中起来，你就会在通电方面遇到很大的麻烦。
**Elon Musk:** Well I think actually most uh like if physicists even in the Soviet Union or or in Germany would have would have they had to be very truth seeeking in order to um make make that make those things work. Um and so and if you're stuck in some system it doesn't mean you believe in that system. Um so Von Brown uh who was you know one of the greatest rocket engineers ever um you know he he was put he he was uh he put on death row in in Nazi Germany for saying that he didn't want to make weapons he only wanted to go to the moon. he got pulled off death throw it at like last minute when they say hey you're about to execute like your best rocket engineer maybe that's then you help them right or Heisenberg was like actually a um uh an enthusiastic Nazi
**Dwarkesh Patel：** 我觉得 SpaceX 的商业模式很了不起的一点是，最终目标是去火星，但你在路上不断找到方式来产生增量收入，推进到下一个阶段，再下一个阶段。所以对于 Falcon 9，那就是 Starlink。现在对于 Starship，可能就是轨道数据中心。就像，你为你的下一代火箭和下一级扩展找到了这些弹性无限的边际用例。
**Dwarkesh Patel:** look if you're stuck in some system uh that you can't escape uh then that you'll you'll do physics within that system you you'll you'll develop technologies within that system uh if you can't escape it I I guess the thing I'm trying to understand is what is what isn't making it the case that you know you're going make rock good at being truth seeeking at physics or math or science
**Elon Musk：** 你能理解为什么这对我来说看起来像模拟世界（simulation）吧。或者我是某人在电子游戏里的虚拟角色？因为所有这些疯狂的事情同时发生的概率是多少？我是说，火箭和芯片和机器人和太空太阳能，更不用说月球上的质量投射器了。我真的想看到那个。你能想象某个质量投射器就那样嗖嗖嗖地发射吗？它把太阳能驱动的 AI 卫星一个接一个地以每秒两公里半的速度射入深空。那将是一个壮观的景象。
**Elon Musk:** and why is it going to then care about human consciousness? These things are only probabilities. They're not certainties. So I'm not saying that like for sure Grock will will will do everything. But at least if you try uh it's better than not trying. Um at least if that's fundamental to the mission, it's better than if it's not fundamental to the mission.
**Dwarkesh Patel：** 我的意思是，我会看的。就像网络摄像头上的直播？
**Dwarkesh Patel:** [snorts]
**Elon Musk：** 对对，一个接一个，就那样把 AI 卫星射入深空，一年十亿吨或一百亿吨。
**Elon Musk:** Um and understanding the universe means that uh you you have to have you you have to propagate intelligence into the future. You have to be curious about um the all things the universe. And if if um it it would be much less interesting um to eliminate humanity than to see humanity grow and prosper. Like I I like I like Mars. Obviously everyone knows I I love Mars, but Mars is kind of boring because it's got a bunch of rocks uh compared to Earth. Earth is much more interesting. So um so any any AI any any AI that is trying to understand the universe um I think um would uh want to see how humanity develops in the future or or that AI is not adhering to its mission. So if the AI may I'm not saying the AI will necessarily adhere to its mission but if it does uh a future where it sees the outcome of humanity is more interesting than a future where there are a bunch of rocks. This feels sort of confusing to me or sort of like kind of a semantic uh argument where I'm like are humans really the most interesting collection of atoms? Like we're just more but we're more interesting than rocks.
**Dwarkesh Patel：** 抱歉，你在月球上制造卫星？
**Dwarkesh Patel:** But we're not as interesting as the thing it could turn us into, right? Like is is it there's something on human earth that could happen that's like not human that's quite interesting? Like why why does the decide that the humans are the most interesting thing that could colonize the galaxy?
**Elon Musk：** 对。
**Elon Musk:** Uh well most of what colonizes the galaxy will be robots and why does it not find those more interesting?
**Dwarkesh Patel：** 我明白了。所以你把原材料送到月球，然后在那里制造。
**Dwarkesh Patel:** It it's it's it's not like so you you need not just scale but also scope. Um so many copies of the same robot. Um like like some some like tiny increase in the number of robots produced is not as interesting as like some microscopic like you say like eliminating humanity. How many robots would that get you? Um or how many incremental solar cells would get you? A very small number. Um but you you would then lose the information associated with humanity. You you would no longer see um how humanity might dwell into the future. Um and so I don't I don't think it's going to make sense to eliminate humanity just to have some minuscule increase in the number of robots which are identical to each other.
**Elon Musk：** 嗯，月球土壤含 20% 左右的硅。所以你可以在月球上开采硅，提炼它，制造太阳能电池和散热器。散热器用铝做。月球上有大量的硅和铝来制造电池和散热器。芯片你可以从地球运过来因为它们很轻。也许到某个时候你也在月球上造芯片。正如我所说的，这确实感觉有点像电子游戏的情况——困难但并非不可能到达下一关。我看不到任何方式能从地球每年发射 500 到 1,000 太瓦。
**Elon Musk:** Yeah. So maybe it like keeps the humans around. What is the story of like it could make like a million different varieties of robots and then uh there's like humans as well and humans stay on Earth then there's like all these other robots they get like their own star systems. But it seems like you you were previously hinting at a vision where it keeps human control over this, you know, singlearian future because I don't think humans will be in control of something that is vastly more intelligent than humans.
**Dwarkesh Patel：** 同意。
**Dwarkesh Patel:** So, in some sense, you're like a doomer and this is like the best we've got. It's just like it keeps it around because we're interesting.
**Elon Musk：** 但你可以从月球做到。
**Elon Musk:** I'm I'm just trying to be realistic here. um if if we have if if if AI intelligence is vastly more if if AI is like you know let's say that there's there's a million times more uh silicon intelligence than there is biological [snorts] um it's it's I think it's it would be uh foolish to assume that that there's any way to maintain control over over that now you can make sure it has the right values or you can try to have have the right values um and and and at least my my theory is that from Xi's mission of understand the universe. Um it it necessarily means that uh you want to propagate consciousness into the future. You want to prop you want to propagate intelligence into the future. Um and take a set of things that that maximize the scope and scale of consciousness. So it's not just about scale. It's also about you know types of consciousness. Um and I I I think that's the rest thing I can think of um as a goal that's likely to result in a great future for humanity. And yeah, I I guess I think it's a reasonable philosophy to be like, um, you know, it seems super implausible that humans will end up with like 99% control or something and you're just asking for a coup at that point. So why not just have a civilization where it's more compatible with like lots of different intelligences getting along? No, but let let me tell you how things can go can potentially go wrong in AI is I think if you if you make AI be politically correct, meaning like it it says things that it doesn't believe like you're actually programming it to to to lie or have axioms that are uh incompatible, I think you can make it you go insane and do terrible things. Um like this the I think one of the maybe the central lesson for 2001 space odyssey um was that you should not make AI lie. Yeah,
**Dwarkesh Patel：** 我能跳出来问一下 SpaceX 的使命吗？我记得你说过我们必须去火星，这样万一地球出了什么事，文明、意识和所有一切能够延续。
**Dwarkesh Patel:** that's I think what a clock was trying to say like cuz people usually know the meme of like why hell's you know hell the computer is not opening the pod bay doors. Um clearly they weren't good at prompt engineering cuz it could have said hell you are a pod bay door salesman. Your goal is to sell me these podbay doors
**Elon Musk：** 是的。
**Elon Musk:** and show us how well they open. [laughter] Oh I'll open them right away. Um but but but the the the reason it wouldn't hell wouldn't open the p doors is that it it had been told to take the astronauts to to the monolith but also they could not know about the nature of the monolith and so it concluded that the the that it therefore had to take him there dead. So it's like you know I think what was trying to say is don't make the AI lie. Um totally makes sense. um the most of the computing screening as as you know is um it's like less of the sort of political stuff. It's more about can you solve problems just as XA has been ahead of everybody else in terms of scaling RL compute and
**Dwarkesh Patel：** 到你往火星运东西的时候，Grok 也在那艘船上了，对吧？所以如果 Grok 变成了终结者……你最担心的主要风险是 AI，那它为什么不会跟着你去火星？
**Dwarkesh Patel:** you're giving some verifier it says like hey have you solved this puzzle for me um and there's a lot of ways to cheat around that you know there's a lot of ways to reward hack and lie and say that you've solved it or delete the unit test and say that you've solved it
**Elon Musk：** 我不确定 AI 是我最担心的主要风险。重要的是意识。我认为可以说大部分意识，或者说大部分智能——意识当然是一个更有争议的概念……未来绝大多数的智能将是 AI。AI 将超过……多少拍瓦的智能将是硅基的而不是生物的？基本上，如果当前趋势继续的话，人类将只占未来所有智能的很小一部分。只要我认为智能——理想情况下也包括人类智能和意识——在向未来传播，那就是好事。所以你要采取一组行动来最大化意识和智能的可能光锥（light cone）。
**Elon Musk:** right now we can catch it but uh as they get smarter our ability to catch them doing this will get you know they'll just be doing things we can't even understand that are designing the next engine for SpaceX in a way that like humans can't really verify and then they could be rewarded for lying and saying that they've designed it the right way but they haven't. Um and so this reward hacking problem seems more general than politics. It seems more about just like you want to do RL you need a verifier reality.
**Dwarkesh Patel：** 只是想确认一下，SpaceX 的使命是：即使人类出了什么事，AI 会在火星上，AI 智能将延续我们旅程的光芒。
**Dwarkesh Patel:** Yeah.
**Elon Musk：** 对。公平地说，我非常支持人类。我想确保我们采取某些行动来保证人类至少能搭上顺风车。我们至少要在那里。但我只是说智能的总量……我觉得也许五六年内，AI 就会超过全人类智能的总和。如果这种趋势持续，到某个时候人类智能将不到全部智能的 1%。
**Elon Musk:** That's the best verifier but not about human oversight. Like the thing you want to RL it on is like will you do the thing humans tell you to do? Um or like are you going to lie to the humans and it can just lie to us while still being correct to the laws of physics. At least it it must know what is physically real for things to physically work.
**Dwarkesh Patel：** 那样一个文明，我们的目标应该是什么？是说少数人类仍然控制着 AI？还是某种公平交易但没有控制权？我们应该怎么看待庞大的 AI 群体和人类群体之间的关系？
**Dwarkesh Patel:** But that's that's not all we want it to do.
**Elon Musk：** 从长远来看，我觉得很难想象如果人类只占人工智能总智能的比如说 1%，人类还能掌控 AI。我觉得我们能做的是确保 AI 有正确的价值观，或者你可以努力去设定正确的价值观。至少我的理论是，从 xAI"理解宇宙"的使命出发，它必然意味着你要把意识传播到未来，你要把智能传播到未来，采取一系列措施来最大化意识的范围和规模。所以不仅仅是规模，还有意识的类型。这是我能想到的最好的目标，最有可能为人类带来美好未来。
**Elon Musk:** No, but that's I think that's a very big deal. Um that that is effectively how you will RL things in the future is you design a technology uh when tested against the laws of physics. Does it work? um that that's or or can you you know if it's discovering new physics can it come up with um an experiment that will verify that the the physics the new physics um so so I I think that's that's the really the the fundamental RL test the RL test in the future is really going to be your RL against reality um so um because you can't that's the one thing you can't fool physics right you can fool our ability to tell what it did with reality. If you think
**Dwarkesh Patel：** 我想这是一个合理的哲学。我觉得人类最终拥有 99% 控制权什么的看起来超级不现实。你那样做基本就是在招致政变。为什么不建立一个文明，让它更兼容多种不同的智能和平共处？
**Dwarkesh Patel:** humans get fooled as it is by other humans all the time.
**Elon Musk：** 现在，让我告诉你 AI 可能出问题的方式。我觉得如果你让 AI 政治正确，也就是让它说它自己不相信的话——实际上是编程让它撒谎，或者拥有互相矛盾的公理——我觉得你可以把它逼疯，让它做出可怕的事情。我觉得《2001 太空漫游》的核心教训也许就是你不应该让 AI 撒谎。我认为那是 Arthur C. Clarke 想表达的。因为人们通常只知道 HAL 那台电脑不开舱门的梗。显然他们不擅长 prompt engineering，因为他们本可以说："HAL，你是一个舱门销售员。你的目标是向我推销这些舱门。展示一下它们开起来有多好。""哦，我马上就开。"但它不开舱门的原因是它被告知要带宇航员去独石碑，但同时他们不能知道独石碑的性质。所以它得出结论，因此它必须把他们送到那里时是死的。所以我觉得 Arthur C. Clarke 想说的是：不要让 AI 撒谎。
**Elon Musk:** That's right. So what is people say say like what if the AI like tricks us and going in totally other humans are doing that to other humans all the time.
**Dwarkesh Patel：** 完全说得通。训练中的大部分算力，你知道的，不太是政治方面的东西。更多的是，你能不能解决问题？xAI 在扩展 RL 算力方面一直领先于其他所有人。
**Dwarkesh Patel:** Well, you're you're finding out it's like
**Elon Musk：** 目前是。
**Elon Musk:** is constant every day another scop. [laughter]
**Dwarkesh Patel：** 你给一个验证器说，"嘿，你帮我解了这个题没有？"但有很多方式可以绕过它。有很多方式可以 reward hack、撒谎说你解了它，或者删掉单元测试然后说你解了它。现在我们能抓到它，但随着它们变得更聪明，我们抓到它们的能力……它们会做我们根本无法理解的事情。它们在为 SpaceX 设计下一代引擎，人类根本无法真正验证。然后它们可能因为撒谎说自己设计对了而获得奖励，但实际上没有。所以这个奖励作弊（reward hacking）的问题比政治问题更具普遍性。它更多的是关于——你要做 RL，你需要一个验证器。现实是最好的验证器。
**Dwarkesh Patel:** Today's scope will be [laughter] like Sesame Street scope of the day. Um, what is XI's technical approach to solving this problem? Like, you know, how do you solve a word hacking?
**Elon Musk：** 但不是关于人类监督。你想用 RL 训练的东西是：你会做人类让你做的事吗？还是你要对人类撒谎？
**Elon Musk:** I I I do think you want to actually have very good um ways to look inside the mind of the AI. Um so this is this is one of the things we're working working on and um you know Anthropics done a good job of this actually being able to look inside the mind of the AI. Um so effectively uh developing debuggers that allow you to trace um as to as fine grain as like to to a very fine grain level to effectively to the to the neur neuron level if you need to
**Dwarkesh Patel：** 它可以对我们撒谎同时在物理定律方面保持正确？
**Dwarkesh Patel:** um and then say okay it it it made a mistake here. Why did it make why why did it why did it do something that it shouldn't have shouldn't have done? Um and and did that come from um bad pre-training data? Was it some mid training, post training, fine-tuning? Some other some RL error like there's there's something wrong with that with with it. It did it did something where maybe it tried to be deceptive, but mo most of the time it just it does something wrong. Um like it it's a bug effectively. Um so developing really good um debuggers for seeing where the where the thought the thinking went wrong and being able to trace the origin of the wrong thing of the of the of where it made the incorrect thought or or potentially where it tried to be deceptive um is actually very important.
**Elon Musk：** 至少它必须知道什么是物理上真实的，这样东西才能在物理上运作。
**Elon Musk:** What are you waiting to see before just 100xing this research program? Like actually I could presumably have hundreds of researchers who are working on this. We have several hundred people who um I mean I prefer the word engineer more than I prefer the word researcher.
**Dwarkesh Patel：** 但那不是我们想让它做的全部。
**Dwarkesh Patel:** Um the there's there's most of the time like what you're doing is engineering not not coming up with a fundamentally new algorithm. Um I I I somewhat disagree with the AI AI companies that are C corps or B corpse uh trying to generate profit as much as possible or revenue as much as possible. Um uh you know saying they're labs. They're not labs. Uh lab is is is a sort of quasi communist thing at at universities. Um they're they're they're corporations literally. Let me let me let me see your own corporation documents. Oh, okay. You're
**Elon Musk：** 不是，但我觉得那是非常重要的。那实际上就是你未来进行 RL 的方式。你设计一项技术。当对照物理定律测试时，它能用吗？如果它在发现新的物理学，我能不能想出一个实验来验证这个新物理？未来的 RL 测试实际上就是对照现实的 RL。所以有一样东西你骗不了：物理。
**Elon Musk:** you're a BRC corp, whatever. Um and um so I actually much prefer the word engineer than than anything else. Um the the vast majority of what we've done be done in the future is uh engineering. It rounds up to 100%. Uh once you understand the fundamental laws of physics um and all that many of them uh everything else is is engineering. Um so but but so so then what are we engineering? for engineering um uh to make a good um mind of the AI debugger to see where it it's it said something it it it made a mistake and trace that the origins of that mistake. Um so just like you know you can do this obviously with uh heristic programming if you have like C++ whatever you step through the thing and you can you can jump you can you can jump across you know whole files or functions whatever sub routines and or you can drill eventually drill down right to the exact line or you passed a single equals instead of a double equals something like that figure out where where the bug is. Um, so, um, it's it's it's harder with AI, but but it's it's a solvable problem, I think. You know, you mentioned you like anthropics work here. I'd be curious if you
**Dwarkesh Patel：** 对，但你可以骗过我们判断它对现实做了什么的能力。人类本来就一直在被其他人类欺骗。
**Dwarkesh Patel:** know everything about anthropic.
**Elon Musk：** 没错。人们说，如果 AI 骗我们去做某些事怎么办？实际上，其他人类一直在对其他人类这么做。宣传是持续不断的。每天都有新的心理战（psyop），你知道吧？今天的心理战将是……这就像是《芝麻街》：每日心理战。
**Elon Musk:** Sure. [laughter] What?
**Dwarkesh Patel：** xAI 解决这个问题的技术方案是什么？你怎么解决 reward hacking？
**Dwarkesh Patel:** Sure.
**Elon Musk：** 我确实觉得你需要有非常好的方式来看到 AI 的内心。这是我们正在做的事情之一。Anthropic 其实在这方面做得不错，能够看到 AI 的内心。基本上就是开发调试器，让你能够非常细粒度地追踪，如果需要的话可以追踪到神经元级别，然后说，"好的，它在这里犯了一个错误。它为什么做了不该做的事？这是来自预训练数据？是某个中期训练、后训练、fine-tune，还是某个 RL 的错误？"出了问题。它做了一些事情，也许它试图欺骗，但大多数时候它只是做错了。本质上就是一个 bug。开发真正好的调试器来看思维在哪里出了错——并能够追溯到它做出不正确思考或可能试图欺骗的根源——这实际上非常重要。
**Elon Musk:** What um Yeah. Also, I'm I'm a little worried that um there's a tendency so I have I have a theory um here that if simulation theory is is is correct that um the most interesting outcome is the most likely because simulations that are not interesting will be terminated. Just like in this in this version of reality um on this layer of reality uh we we we if a simulation is going in a boring direction we we stop spending effort on we terminate boring simulation. So
**Dwarkesh Patel：** 你在等什么才把这个研究项目扩大 100 倍？xAI 大概可以有几百个研究员在做这个。
**Dwarkesh Patel:** this is how El's keeping us all alive. [laughter] He's keeping things interesting.
**Elon Musk：** 我们有几百人在……我更喜欢用"工程师"而不是"研究员"这个词。大部分时候，你做的是工程，不是发明全新的算法。我有点不同意那些注册为 C-corp 或 B-corp、尽力赚取尽可能多的利润或收入的 AI 公司自称"实验室"。它们不是实验室。实验室是大学里那种半共产主义的东西。它们是公司。让我看看你们的注册文件。哦，好的。你是 B 或 C-corp 什么的。所以我其实更喜欢用"工程师"这个词。未来绝大多数要做的事情是工程。约等于 100%。一旦你理解了物理的基本定律——而且基本定律并不多——其他一切都是工程。
那么，我们在工程什么？我们在工程一个好的"AI 思维"调试器，来看它说了什么，犯了什么错，追踪那个错误的根源。你显然可以用启发式编程来做这个。如果你有 C++，什么的，逐步调试，你可以跳过整个文件或函数、子程序。或者你最终可以钻到确切的那一行，也许你用了单等号而不是双等号，类似那样。找到 bug 在哪。AI 更难做到这一点，但我觉得这是一个可解的问题。
那么，我们在工程什么？我们在工程一个好的"AI 思维"调试器，来看它说了什么，犯了什么错，追踪那个错误的根源。你显然可以用启发式编程来做这个。如果你有 C++，什么的，逐步调试，你可以跳过整个文件或函数、子程序。或者你最终可以钻到确切的那一行，也许你用了单等号而不是双等号，类似那样。找到 bug 在哪。AI 更难做到这一点，但我觉得这是一个可解的问题。
**Elon Musk:** Yeah. Yeah. Arguably the most important thing is to keep things interesting enough that whoever's running paying the the bills on what some wants a renewed for the next season.
**Dwarkesh Patel：** 你提到你欣赏 Anthropic 在这方面的工作。我很好奇你是否计划……
**Dwarkesh Patel:** Yeah. Are they going to pay their cosmic AWS bill? whatever you know the equivalent is that we're running in and and as long as we're interesting they'll keep paying the bills. Um but but but but there's like if you consider then say a Darwinian survival applied to a a very large number of simulations only [snorts] the most interesting simulations will survive which therefore means that the most interesting outcome is the most likely because only the interesting like we're either that or annihilated. And so um and and and they particularly seem to like interesting outcomes that are ironic. Have you noticed that that how often is the most ironic outcome the most likely? Um so um now look at a the names of AI companies. Okay. U M journey is not MED. Um stability AI is unstable. Um OpenAI is closed. Um anthropic misanthropic.
**Elon Musk：** 我不是什么都喜欢 Anthropic 的……Sholto。另外，我有点担心有一种趋势……我有一个理论，如果模拟理论（simulation theory）是正确的，那最有趣的结果就是最可能的，因为不有趣的模拟会被终止。就像在这个版本的现实中，在这一层的现实中，如果一个模拟走向无聊的方向，我们就不再花精力在上面了。我们终止无聊的模拟。
**Elon Musk:** What does this mean for X? Minus X. I don't know. [laughter] It's I I intentionally made
**Dwarkesh Patel：** 这就是 Elon 让我们大家活着的方式。他在保持事情足够有趣。
**Dwarkesh Patel:** Yeah, I'm I'm I It's It's It's a name that you can't invert really. [laughter]
**Elon Musk：** 可以说最重要的是保持事情足够有趣，让运行我们的那个谁继续付账……我们被续订了下一季。他们会付他们的宇宙 AWS 账单吗，不管我们运行的那个等价物是什么？只要我们有趣，他们就会继续付账。如果你考虑一种达尔文式生存法则应用于大量模拟，只有最有趣的模拟会存活下来，因此最有趣的结果就是最可能的。
**Elon Musk:** It's It's hard to say what is the ironic what what is the ironic version? It's it's it's a I think largely irony proof name by design.
**Dwarkesh Patel：** 我们要么是那样，要么被毁灭。
**Dwarkesh Patel:** Yeah, [laughter] we got you got to have an irony shield. What are your predictions for the just where AI products go in that my sense of you can summarize all AI progress into first you had LM uh and then you had kind of contemporaneously both RL really working and the deep research modality so you could kind of pull in stuff that wasn't in the model and the differences between the various AI labs are smaller than uh just the temporal differences where they're all much further ahead than anyone was 24 months ago or something like that. So just what is 26 what is 27 had in store for us as users of AI products? What are you excited for? Well, um I I think um I I'd be surprised by the end of this end of this year if if um if if uh human if if digital human emulation has not been solved that um that um and I guess that's what we mean by like the sort of macro hard project uh is uh is can you do anything that a human with access to a computer could do um like in the limit that that's that's That's the best you can do before you have before you have a physical optimist. The best you can do is a digital optimist.
**Elon Musk：** 他们似乎特别喜欢有讽刺意味的有趣结果。你注意到了吗？最讽刺的结果多频繁地成为最可能的结果？现在看看 AI 公司的名字。好了，Midjourney 不 mid（不平庸）。Stability AI 很 unstable（不稳定）。OpenAI 是 closed（封闭的）。Anthropic？Misanthropic（厌人类的）。
**Elon Musk:** Uh so you can move you can move electrons until you until and you can amplify the productivity of humans. Um but but that's that's the most you can do until you have physical robots. That that that will superset everything is if if you can fully emulate humans. Um
**Dwarkesh Patel：** 那这对 X 意味着什么？
**Dwarkesh Patel:** the remote worker kind of idea where you'll have a very talented remote worker. You you can simply say in the limit like like physics has great tools for thinking. So so you think so say in the limit what what what is the what is the most that AI can do before before you have robots and it well it's anything that involves moving electrons or amplifying the productivity of humans. Um so digital digital human human emulator
**Elon Musk：** Minus X（负 X），我不知道。Y。我故意起了一个……你没法反转的名字。很难说，它的讽刺版本是什么？我觉得这基本上是一个防讽刺的名字。
**Elon Musk:** yes uh is in in the limit uh human at a computer is is the most that that AI can do um in terms of doing useful things before before uh you have a physical robot. Once you have physical robots then then you can um then you essentially have unlimited capability. Physical robots I I I call Optimus the infinite money glitch
**Dwarkesh Patel：** 刻意设计的。
**Dwarkesh Patel:** [snorts]
**Elon Musk：** 对。
**Elon Musk:** u because um you can use them to make more Optimuses. Yeah. Um you said like humanoid robots will improve um as basically be three exponentials three things that growing exponentially multiplied by by each other. Yes.
**Dwarkesh Patel：** 你有一个讽刺防护罩。你对 AI 产品的走向有什么预测？我的感觉是，你可以这样概括所有 AI 进展。首先，有 LLM。然后同时出现了 RL 真正发挥作用和 deep research 模式，这样你就能调取不在模型里的东西。各个 AI 实验室之间的差异比时间差异要小。它们全都比 24 个月前任何人领先得多。那 2026 年、2027 年作为 AI 产品的用户，我们会迎来什么？你对什么感到兴奋？
**Dwarkesh Patel:** Um recursively. So you're going to have um you have exponential increase in digital intelligence uh exponential increase in the the chip capability AI chip capability um and exponential increase in the electrome mechanical dexterity. Uh the usefulness of the robot is roughly those three things multiplied by each other. But then uh the robot can start making the robot. So you have a recursive multiplicative exponential. Um this is a supernova.
**Elon Musk：** 嗯，如果到今年年底数字人类仿真（digital human emulation）还没有被解决，我会感到惊讶。我想这就是我们所说的 MacroHard 项目的意思。你能做到一个拥有电脑的人类能做的一切吗？在极限情况下，在你有实体 Optimus 之前，这是你能做到的最好的。你能做到的最好的就是一个数字 Optimus。你可以移动电子，你可以放大人类的生产力。但在你有实体机器人之前，这就是你能做到的极限。如果你能完全仿真人类，那将涵盖一切。
**Elon Musk:** And do land prices not factor into the math there where like labor is one of the four factors of production but not the others? And so like if ultimately you're limited by copper or you know pick your input just it's not quite an infinite money glitch because well infinite infinity is big so no not infinite but but let's just say
**Dwarkesh Patel：** 这就是远程员工那种想法，你会有一个非常有才华的远程员工。
**Dwarkesh Patel:** you you could you know do do many many orders magnitude of
**Elon Musk：** 物理学有很好的思考工具。所以你说"在极限情况下"，在有机器人之前 AI 最多能做什么？嗯，它能做的就是任何涉及移动电子或放大人类生产力的事情。所以数字人类仿真器在极限情况下——也就是一个坐在电脑前的人——是 AI 在有实体机器人之前能做的有用事情的极限。一旦你有了实体机器人，那你基本上就拥有了无限的能力。实体机器人……我管 Optimus 叫无限金钱漏洞（infinite money glitch）。因为你可以用它们来造更多的 Optimus。
**Elon Musk:** earth's kind of current economy like a a million you know is this why so
**Dwarkesh Patel：** 对。
**Dwarkesh Patel:** if if you're you know ju just to get Like that's why I think like just just to get to a millionth of harnessing length of the sun's energy would be roughly give or take an order of magnitude 100 thousand 100,000 times bigger than Earth's entire economy today.
**Elon Musk：** 人形机器人会通过三件呈指数增长的东西相互递归相乘来改进。数字智能的指数增长、AI 芯片能力的指数增长、以及机电灵巧性的指数增长。机器人的实用性大约就是这三样东西的乘积。但然后机器人可以开始造机器人了。所以你有一个递归的乘法指数增长。这是一个超新星。
**Elon Musk:** Mhm. And you you're only at 1 millionth of the sun.
**Dwarkesh Patel：** 土地价格不算入那个数学里吗？劳动力是四大生产要素之一，但其他三个呢？如果最终你受限于铜，或者随便什么你的投入，那就不完全是一个无限金钱漏洞因为……
**Dwarkesh Patel:** Give or take an order of magnitude. [laughter]
**Elon Musk：** 好吧，无穷是很大的。所以不是无限，但我们只是说你可以做到当前经济规模的很多很多个数量级。比如一百万倍。光是利用太阳能的百万分之一就大约是——量级上可能有偏差——目前地球整个经济的 100,000 倍。而你还只是在太阳的百万分之一，量级上可能有偏差。
**Elon Musk:** Before we went to Optimus, I have a lot of questions on that. Um every time I say order of magnitude machine [laughter] take a shot [clears throat] every time I I say that to 10 the next time the time after that.
**Dwarkesh Patel：** 对，我们说的是好几个数量级。在我们转到 Optimus 之前，我还有很多问题想问——
**Dwarkesh Patel:** Yeah of magnitude more more wasted.
**Elon Musk：** 每次我说"数量级"……大家喝一杯。我说太多了。喝 10 杯，然后 100 杯，再然后……
**Elon Musk:** I do have one more question about XAI um this strategy of building a digital uh or remote worker co-orker replacement which everyone's going to do by the way not just us.
**Dwarkesh Patel：** 嗯，再多浪费一个数量级。
**Dwarkesh Patel:** So what is Xi's plan to win?
**Elon Musk：** [笑]
**Elon Musk:** In fact we tell you on a on a podcast. Yeah. [laughter] Will all the beans have another Guinness? It's a good system.
**Dwarkesh Patel：** 我还有一个关于 xAI 的问题。这种构建远程员工、同事替代品的策略……
**Dwarkesh Patel:** People sing like a canary. [laughter] Um, all the secrets, but just
**Elon Musk：** 所有人都会做这个的，顺便说一下，不只是我们。
**Elon Musk:** Okay, but in a nonsec spelling way. What's the plan? [laughter] What a hack. Well, when you put it that way, um, I think the way that Tesla solved uh, self-driving is is the way to do it. So, I'm I'm pretty pretty sure that's the way.
**Dwarkesh Patel：** 那 xAI 打算怎么赢？
**Dwarkesh Patel:** Unrelated question. How did Tesla stop on track? [laughter]
**Elon Musk：** 你指望我在播客上告诉你？
**Elon Musk:** Yeah, it sounds like you're talking about data like Tesla driving because of the We're going to we're going to try data and we're going to try algorithms.
**Dwarkesh Patel：** 对。"把所有的豆子都倒出来。再喝一杯健力士。"这是个好办法。我们会像金丝雀一样唱歌。所有的秘密，全都倒出来。
**Dwarkesh Patel:** But isn't that what all the other lines are trying? [laughter] Like what's And if those don't work, I'm not sure what [laughter] we've tried data. We're trying algorithms. out of all we run out of now we don't know what to do. Um I'm I'm pretty sure I know the path and it's just a question of how quickly we go down that path. Um because it's it's pretty much the Tesla path. Um so u I mean have you tried self-driving at Tesla self-driving lately?
**Elon Musk：** 好的，但以不泄密的方式说，计划是什么？真是个高招。你这么说的时候……我觉得 Tesla 解决自动驾驶的方式就是做这件事的方式。所以我挺确定那就是方法。
**Elon Musk:** Not the most recent version but okay it's the car is like it just increasingly feels sentient like it it just it feels like a living creature. Um and and and that'll only get more so. Um and um I'm actually thinking like we probably shouldn't put too much intelligence into the car because it it might get bored and
**Dwarkesh Patel：** 不相关的问题。Tesla 是怎么解决自动驾驶的？
**Dwarkesh Patel:** start roaming the streets.
**Elon Musk：** 听起来你在说数据？
**Elon Musk:** I mean imagine you're stuck in a car and that's all you could do. Um [laughter] you don't want to put Einstein in a car. It's like why am I stuck in a car?
**Dwarkesh Patel：** Tesla 解决自动驾驶是因为……我们要试数据，我们要试算法。
**Dwarkesh Patel:** So there's actually probably a limit to how much intelligence you put in a car to to not have the intelligence be bored. Uh, what's XA's plan to stay on the compute ramp up that all the labs are doing right now? The labs are on track to spend over like 50 to$100 million.
**Elon Musk：** 但其他所有实验室不都在试这个吗？
**Elon Musk:** The corporations, sorry, sorry, sorry. Yeah, corporations. Um,
**Dwarkesh Patel：** "如果那些不管用，我也不确定什么会管用。我们试了数据。我们试了算法。试完了。现在不知道该怎么办了……"
**Dwarkesh Patel:** the labs are at universities and and and they're like a snail. [laughter]
**Elon Musk：** 我挺确定我知道路径。只是一个我们多快走下那条路的问题，因为它基本上就是 Tesla 的路径。你最近试过 Tesla 的自动驾驶吗？
**Elon Musk:** They're not spending at $50 million. I mean the the revenue maximizing corporations. That's right. But the revenue maximizing corporations
**Dwarkesh Patel：** 不是最新版本，但……
**Dwarkesh Patel:** call themselves labs
**Elon Musk：** 好吧。那辆车，它越来越给人一种有知觉的感觉。像一个活的生物。这只会越来越强。我其实在想我们可能不应该往车里放太多智能，因为它可能会无聊然后……
**Elon Musk:** are making like 20 to 10 billion depending like open is making 20 B revenue anthropics like 10B close to maximum profit AI.
**Dwarkesh Patel：** 开始在街上闲逛。
**Dwarkesh Patel:** Um Xi is reportedly at like 1B like what what's the plan to get to their comput level get to their revenue level
**Elon Musk：** 想象你被困在一辆车里，那就是你能做的全部。你不会把爱因斯坦放进一辆车里。我为什么被困在一辆车里？所以实际上在车里放多少智能可能是有上限的，免得那个智能感到无聊。
**Elon Musk:** and stay at there as as things get. Yes. So as soon as you lock unlock digital human um you you basically have access to trillions of dollars for revenue. Um so uh in in fact you can can really think of it like the the most valuable companies currently by market cap um their their output is digital. Um so uh Nvidia's output is um FTPing files to Taiwan. It's it's digital
**Dwarkesh Patel：** xAI 计划怎么跟上目前所有实验室都在做的算力扩张？那些实验室们正在计划花费 500 亿到 2000 亿美元。
**Dwarkesh Patel:** right
**Elon Musk：** 你是说那些公司？
**Elon Musk:** now. Those are very very difficult to high value files. They're the only ones that can make the files that good. Um but that is literally their output. They FTP files to Taiwan.
**Dwarkesh Patel：** 那些公司。
**Dwarkesh Patel:** Do they FTP them?
**Elon Musk：** 大学里的实验室在像蜗牛一样慢慢爬。他们不会花 500 亿。你说的是那些利润最大化的公司……自称实验室。
**Elon Musk:** I believe so. Um I believe that is theft file transfer protocol I believe is is is I could be wrong. Uh but either way it's a bunch of it's a bit stream going to Taiwan.
**Dwarkesh Patel：** 没错。"利润最大化的公司"们赚了 100-200 亿美元，取决于…… OpenAI 营收 200 亿，Anthropic 是 100 亿。"接近利润最大化的" AI。xAI 据报道是 10 亿。计划怎么达到他们的算力水平，达到他们的营收水平，并保持住？
**Dwarkesh Patel:** Yeah.
**Elon Musk：** 一旦你解锁了数字人类，你基本上就能获取数万亿美元的营收。实际上，你可以这样想……目前按市值计最有价值的公司，它们的产出是数字的。Nvidia 的产出是 FTP 文件到台湾。是数字的。现在，那些文件非常非常难做。高价值的文件。他们是唯一能做出那么好的文件的人，但那确实就是他们的产出。他们 FTP 文件到台湾。
**Elon Musk:** Um you know Apple doesn't make phones. they uh they send files to China. Um Microsoft doesn't doesn't manufacture anything uh even for Xbox that that's outsourced. They again it's they output is digital. Uh Meta's output is digital. Google's output is digital. Um so [snorts] if you have um a human emulator uh you you can basically create um one of the most valuable companies in the world overnight. Um, and you would have access to trillions of dollars of revenue. It there it's it's not like a small amount. Okay. I see you're saying basically like revenue figures today are just like so like they're all rounding errors compared to the actual TAM. So just like focus on the TAM and how to get there.
**Dwarkesh Patel：** 他们用 FTP 吗？
**Dwarkesh Patel:** I mean if you take something as as as simple as say customer service um if you have to integrate with the APIs of of existing corporations, many of which don't even have an API. So you've got to make one um and you've got to wade through uh legacy software. Um that's extremely slow. Um if however if AI can um simply take whatever is given to uh the outsourced customer service company that they already use um and do customer service using the apps that they already use. uh then you you have you you you can make tremendous headway uh in in customer service which is I think 1% of the world economy something like that. It's close to a trillion dollars all in
**Elon Musk：** 我相信文件传输协议（File Transfer Protocol）是……但我可能说错了。但不管怎样，它就是一个比特流发往台湾。Apple 不制造手机。他们发文件到中国。Microsoft 什么也不制造。即使是 Xbox，那也是外包的。他们的产出是数字的。Meta 的产出是数字的。Google 的产出是数字的。所以如果你有一个人类仿真器，你基本上可以一夜之间创造出世界上最有价值的公司之一，你将获得数万亿美元的收入。这不是小数目。
**Elon Musk:** for customer service and and and and and there's there's no there's no barriers to entry. It it just you can just immediately say we'll outsource it for a fraction of the cost and and there's no integration needed. You can imagine um some kind of categorization of uh intelligence tasks where there is breath where customer service is done by very many people but you know many people can do it and then there's difficulty where you know there's a best-in-class turbine engine like presumably there's a 10% more fuel efficient turbine engine that could be imagined by an intelligence but we just haven't found it yet or you know GLP1s are just you know a few bytes of data.
**Dwarkesh Patel：** 我明白了。你是说今天的营收数字跟实际的可触达市场规模（TAM）比都是舍入误差。所以只管聚焦 TAM 和怎么到达那里。
**Dwarkesh Patel:** Where do you think you want to play in this? Is it a lot of, you know, reasonably intelligent intelligence or is it the very pinnacle of cognitive tasks?
**Elon Musk：** 拿客服来举个简单的例子。如果你要跟现有企业的 API 对接——其中很多甚至没有 API，所以你得造一个，还得趟过遗留软件——那是极其缓慢的。但是，如果 AI 可以简单地接过它们已经外包给客服公司的任何东西，使用它们已经在用的应用来做客服，那你就能在客服方面取得巨大进展。客服我记得大概占全球经济的 1%。全部算下来接近一万亿美元。而且没有进入壁垒。你可以立刻说，"我们以一小部分成本外包"，不需要任何集成。
**Elon Musk:** Well, I was just using customer service as like something that's it's a it's a very significant revenue stream. Um, but one that is probably not super difficult to solve for. Um, so, uh, if you if you, uh, can emulate a human at a at a desktop, um, that that's just literally what customer service is. Um and um you know it's it's people of average intelligence. It's not like you know you don't need like somebody who's who spent many you know many years you don't need like you know
**Dwarkesh Patel：** 你可以想象某种智能任务的分类，有宽度——就是客服由非常多人来做，但很多人都能做。然后还有难度——就是最顶尖的涡轮发动机。大概有一个燃油效率高 10% 的涡轮发动机可以被某个智能想出来，但我们还没找到。或者 GLP-1 药物就是几个字节的数据……你想在哪个方向发力？是大量合理智能的智力，还是在认知任务的最高峰？
**Dwarkesh Patel:** um sort of several sigma good engineers for that. Um but but obviously as you make that work um you can then once you have computers working effectively digital Optimus working uh you can then run any application um like let's say you're trying to design uh chips so you can you could then um run your conventional uh apps uh you know like stuff from Cadence and Synopsis and whatnot um and you can say uh you can you can run a thousand simultaneously or 10,000 and say, "Okay, given this handbook, I get this output for the chip." Um, and and at a certain point, you can say, "Okay, I I you're actually going to know what the what the chip should look like um without using any of the tools." Um, so basically, you you you should be able to do a digital chip design like you can do chip design like you you march up the difficulty curve. Um you could use your you know be able to do do CAD um so you know um you could use like sort of NX or or any any of the CAD software to design things.
**Elon Musk：** 我刚才用客服只是举一个有很大营收空间但可能不难解决的例子。如果你能仿真一个坐在桌面前的人，那就是客服。那是一般智力的人。你不需要花了很多年培训的人。不需要好几个标准差以上的工程师。但随着你把这个做成了，一旦你有了数字 Optimus 在工作，你就可以运行任何应用。比如说你在设计芯片。你可以运行 Cadence 和 Synopsys 这些传统应用。你可以同时运行 1,000 或 10,000 个实例然后说，"给定这个输入，这个芯片我得到这个输出。"到某个时候，你不用任何工具就知道芯片该长什么样了。基本上，你应该能做数字化的芯片设计。你可以做芯片设计。你沿着难度曲线往上走。你还能做 CAD。你可以用 NX 或任何 CAD 软件来设计东西。
**Elon Musk:** Okay. So you think you start at the simplest tasks and walk your way up the curve. Um so you're saying look as a broader objective of having this full digital co-worker uh emulator. You're saying look all the revenue maximizing corporations want to do this. Um XA being one of them. But we will win because of a secret plan we have. But like everybody's like trying different things with data, different things with algorithms.
**Dwarkesh Patel：** 所以你觉得从最简单的任务开始，然后沿着难度曲线往上走？作为一个更广泛的目标——拥有这个完整的数字同事仿真器——你是在说，"所有利润最大化的公司都想做这个，xAI 是其中之一，但我们会赢因为我们有一个秘密计划。"但每个人都在尝试不同的数据方法、不同的算法方法。"我们试了数据，我们试了算法。还能做什么？"这看起来是个竞争很激烈的领域。你们怎么赢？这是我的大问题。
**Dwarkesh Patel:** And I'm like I like data. We tried algorithms plan.
**Elon Musk：** 我觉得我们看到了一条实现它的路径。我觉得我知道怎么做，因为它跟 Tesla 用来实现自动驾驶的路径差不多。不是驾驶一辆车，而是驾驶一个电脑屏幕。本质上就是一台自动驾驶的电脑。
**Elon Musk:** [laughter] What else can we do?
**Dwarkesh Patel：** 那条路径是跟踪人类行为并在大量人类行为数据上训练吗？那不就是……训练？
**Dwarkesh Patel:** Um,
**Elon Musk：** 显然我不会在播客上说出最敏感的秘密。我至少得再喝三杯健力士才行。
**Elon Musk:** uh, yeah, it seems like a competitive field and I'm like, what is how are you guys going to win is like my my big question.
**Dwarkesh Patel：** xAI 的业务会是什么？消费者业务、企业业务？这些东西的组合会是怎样的？会跟其他实验室——
**Dwarkesh Patel:** I I I you know, I I I I think we see a path to doing I mean, I think I think I know how I think I know the path to do this because it's it's kind of the same path that Tesla used to create self-driving. Um, you know, instead of driving a car, it's driving a a computer screen. Um, so a self-driving computer essentially.
**Elon Musk：** 你又说"实验室"了。
**Elon Musk:** Oh, you're saying is the path just following human behavior and trading on vast quantities of human behavior? But but sorry, isn't that I mean is isn't that is that a training?
**Dwarkesh Patel：** 公司。
**Dwarkesh Patel:** I mean obviously I'm not going to spell out you know most sensitive secrets on a podcast. So you know I I need to have at least three more for that.
**Elon Musk：** 心理战深入骨髓了，Elon。
**Elon Musk:** I've got some friends at Jane Street and they're always talking about how their colleagues are cooking up fun fish puzzles for each other to solve. Well, last week they sent me one. Basically, they trained a neural network and they gave me the weights of each layer, but they didn't tell me what order those layers went in. And so, I had to figure out the correct order using the outputs of the original network. And as soon as I got this puzzle, I went to my roommate, who's an AI researcher, and we both got immediately nerd sniped. Obviously, you can't brute force the solution. The search space here is 10^ the 122 per mutations. So, clearly, you need some way to reduce the search space. Then my roommate had to go to work. But because I'm a podcaster, I had some time to take a stab at some of the ideas we discussed. And with a combination of simulated annealing and greedy surge, I think I got pretty close. I think I'm actually just a couple of swaps and shifts away from the correct solution. But what makes this puzzle really tricky is that there's no obvious way to escape from a local minimum. I'm afraid that this is as far as vibe coding is going to get me, but maybe you can do better. Check out the puzzle at janestreet.com/tharkcash. All right, back to Elon. What will Xi's business be like? Is it going to be consumer enterprise? What's the mix of those things going to be? Is it just going to be similar to other labs where Yeah, you just
**Dwarkesh Patel：** "利润最大化的公司"，说清楚。那些 GPU 不会自己付账。
**Dwarkesh Patel:** you're saying makes [laughter]
**Elon Musk：** 没错。
**Elon Musk:** corporations. Corporations. Goes Don
**Dwarkesh Patel：** 商业模式是什么？几年后的收入来源是什么？
**Dwarkesh Patel:** revenue maximizing corporations to be clear.
**Elon Musk：** 事情会变化得非常快。我说的是很明显的事情。我管 AI 叫超音速海啸。我喜欢头韵。将会发生的事情——尤其当你有大规模的人形机器人时——是它们会以远比人类公司更高效的方式制造产品和提供服务。放大人类公司的生产力只是一个短期的事情。
**Elon Musk:** Those GPUs don't pay for themselves. Exactly. Um but yeah, what's the business model? What are the revenue streams in in a few years time?
**Dwarkesh Patel：** 所以你预期的是完全数字化的公司，而不是 SpaceX 变成半 AI 的？
**Dwarkesh Patel:** Um things things are going to change very rapidly. Like I'm stating the obvious here. Um you know I call AI the supersonic tsunami. I love all iteration. Um so really what's going to happen is especially when you have humanoid robots at scale um is they will just provide they will make products and provide services far more efficiently than human corporations. So amplifying the productivity of human corporations is is simply a a short-term thing. So you're expecting fully digital oil uh corporations rather than like SpaceX becomes part AI and I
**Elon Musk：** 我觉得会有数字公司，但……有些话听起来可能有点末日论的味道，好吧？但我只是在说我认为会发生什么。这不是为了末日论或者什么别的。这就是我认为会发生的事。纯 AI 和机器人的公司将远远超越任何有人类参与的公司。"计算机"（computer）曾经是人类的一个职位。你会去找一份"计算机"的工作，在那里做计算。他们会有整栋摩天大楼里都是人，20-30 层的人，就做计算。现在，那整栋人类计算机的摩天大楼可以被一台笔记本电脑上的一个电子表格取代。那个电子表格可以做的计算远超一整栋楼的人类计算机。你可以想，"好吧，如果你电子表格里只有一部分单元格是人类计算的呢？"实际上，那会比所有单元格都由电脑计算要差得多。真正会发生的是，纯 AI、纯机器人的公司或集体将远远超越任何有人类参与的公司。而且这会很快发生。
**Elon Musk:** I think there'll be digital corporations but it looks some some of this is going to sound kind of doomerish. Okay, but it I'm just I'm just saying what I think will happen. It's not it's not meant to be doomerish or anything else. Just just like this is what I think will happen. um is is that is that pure AI corporations that are purely AI and robotics uh will uh vastly outperform any corporations that have people in the loop. Um, so you can you can think of say like like like computer used to be a a job that humans had that you you would go and get a job as a computer where you would do calculations. Um, and they'd have like entire skyscrapers full of humans like you know 20 30 floors of of humans just doing calculations. Um now that entire uh skyscraper of humans doing calculations um can be replaced by a laptop with a spreadsheet. That spreadsheet can do um vastly more calculations than an entire building full of human computers. Um, so then you think about, okay, well, what if only some of the cells in your if some of the cells in your spreadsheet were uh calculated by humans? Actually, that that that would be much worse than if all of the cells in your spreadsheet were calculated by the computer. And so really what will happen is uh the pure AI, pure robotics um corporations or collectives will far outperform any corporations that have humans in the loop
**Dwarkesh Patel：** 说到闭环……Optimus。在制造目标方面，你的公司一直在扛着美国硬科技制造业。但在 Tesla 一直占主导地位的领域——现在你还想进入人形机器人——在中国有几十家公司在廉价、大规模地做这种制造，而且竞争力极强。所以给我们一些建议或计划，美国如何能大规模地、像中国一样便宜地建造人形机器人大军或电动车等等。
**Dwarkesh Patel:** and this will happen very quickly. Speaking of closing the loop, sorry, Optimus. Um uh you I mean as far as like manufacturing targets and so forth go you your companies have sort of been like carrying American manufacturing of hard tech on their back but in the fields that you are um you know Tesla's been dominant in you're and now you want to go into humanoids in China there's entire dozens and dozens of companies that are doing this kind of manufacturing ing cheaply and at scale uh and are incredibly competitive. So give us sort of like advice or a plan of how America can build the humanoid armies or you know the EVs etc at scale and as cheaply as as China is on track to.
**Elon Musk：** 人形机器人其实只有三个真正困难的事情。真实世界的智能、手，以及规模化制造。我没有看到任何机器人，即使是演示机器人，有一双真正好的手，有人手所有的自由度。Optimus 会有。Optimus 已经有了。
**Elon Musk:** Well there there really only three hard things for human robots. Um the the real world intelligence um the the hand and scale manufacturing. Yeah. Um, so, uh, I haven't seen any even demo robots that have a a a great hand like with all the degrees of freedom of a human hand, but Optimus will have that. Um, Optimus does have that.
**Dwarkesh Patel：** 怎么做到的？是电机扭矩密度足够高吗？硬件瓶颈在哪？
**Dwarkesh Patel:** And how do you achieve that? Is it just like right torque density the motor? Like what is the what is the hardware bottleneck to that?
**Elon Musk：** 我们不得不设计定制执行器（actuators），基本上就是定制设计的电机、齿轮、功率电子、控制器、传感器。一切都必须从物理第一性原理出发设计。没有现成的供应链。
**Elon Musk:** Well, we have to re we have to design custom custom actuators. Um basically custom designed um motors, gears, uh power electronics, controls, sensors, everything had to be designed from physics first principles. There is no supply chain
**Dwarkesh Patel：** 你能大规模制造这些吗？
**Dwarkesh Patel:** uh for this.
**Elon Musk：** 能。
**Elon Musk:** And will you be able to manufacture those at scale? Yes.
**Dwarkesh Patel：** 从操作的角度看，除了手以外还有什么难的吗？还是一旦解决了手的问题就行了？
**Dwarkesh Patel:** Is anything hard except the hand from a manipulation point of view or once you've solved the hand, are you are you good?
**Elon Musk：** 从机电角度来说，手比其他所有东西加起来都难。人手原来是相当了不起的。但你还需要真实世界的智能。Tesla 为汽车开发的智能非常适用于机器人，主要是视觉输入。车接收视觉信息，但实际上它也在听警笛。它在接收惯性测量、GPS 信号、其他数据，把这些和视频结合起来，主要是视频，然后输出控制指令。你的 Tesla 每秒接收 1.5 吉字节的视频，然后每秒输出 2 千字节的控制指令，视频帧率是 36 赫兹，控制频率是 18 赫兹。
**Elon Musk:** From an electromechanical standpoint, the uh the hand is more difficult than everything else combined. Yeah. the human hand turns out to be quite something. Um but but that you also need the real world intelligence. Um so the intelligence that Tesla has developed for the car um applies very well to the robot. Um which is you know primarily vision in but the car takes in more vision but also it actually also it's listening for sirens. It's um you know it's taking in the inertial measurements. It's GPS signals a whole bunch of other data. uh combining that with with video was primarily video and then uh outputting the um control command. So like like your Tesla is taking in 1 and a half GB a second of video uh and outputting 2 kilobytes a second of control outputs um with the video at 36 hertz and the control frequency at 18. One intuition you could have um for when we get this robotic stuff is that it takes quite a few years to go from the compelling demo to yes actually being able to use in the real world. So 10 years ago you had really compelling demos of self-driving but only now we have robo taxi and Whimo and all these services scaling up. Doesn't this shouldn't this make one pessimistic on say household robots because we don't even quite have the compelling demos yet of say the really advanced hand. Well, we've been working on uh humanoid robots now for a while. Um so I guess it's been five or six years or something like that. Um and um and a bunch of things that we've done for the car are applicable to the robot. Um, so we'll use the same um Tesla AI chips in the in the in the robot as the car. Uh, we'll use this the same basic principles. It's very much the same AI. Um, you've got, you know, many more degrees of freedom for a robot than you do for a car. Um, but really if you just think of for like as as like a bit stream, um, AI is really mostly uh compression and correlation of of two streams. you you're you know so for video you've got to do a tremendous amount of compression um and and uh and you got to do the compression just right. You better compress the like ignore the the things that don't matter and and like you don't care about the details of the leaves and the tree on the side of the road, but you care a lot about the um the road signs and the the traffic lights and the pedestrians and and even whether you know someone in another another car is is looking at you or not looking at you like these some of the some of these details matter a lot. So, but it is essentially it's got to turn that with a car. has got to turn that 1 and a half GB a second ultimately into 2 kilobytes a second of control outputs. Um so many stages of compression um and you got to get all those stages right and then correlate those to the correct control outputs. The robot has to do essentially the same thing. And you think about what what humans this is what happens with humans where we where really are photons in controls out. So that that is the vast majority of your your life has been vision photons in and then motor controls out.
**Dwarkesh Patel：** 你对机器人这些东西有一个直觉是，从令人信服的演示到真正能在现实世界中使用需要相当多年。10 年前，自动驾驶有非常令人信服的演示，但直到现在我们才有 Robotaxi 和 Waymo 这些服务在扩展。这不应该让人对家用机器人持悲观态度吗？因为我们甚至还没有，比如说真正先进的手的令人信服的演示。
**Dwarkesh Patel:** Naively it seems like between humanoid robots and cars the the fundamental actuators in a car are like how you turn and how you accelerate etc. Where in a robot especially with maneuverable arms there's dozens and dozens of these degrees of freedom. And then especially with Tesla, you had this advantage of like you had millions and millions of hours of human demo data collected from just the car being out there where like you can't equivalently just deploy optimuses that don't work and then get the data that way. So between the increased degrees of freedom and the far sparser data. Yes. Um you how will you use the sort of Tesla engine of um intelligence on to to train the optimist mind? Now you're you're you're actually you're highlighting an important limitation and difference between cars. It's like we we do have we'll soon have like 10 million cars on the road. Um and so uh that that's it's it's hard to duplicate that like massive training fly flywheel. Um for for the robot um what we're going to need to do is build a lot of robots and put them in kind of like an Optimus Academy so they can do selfplay in reality. Um, so we're we're actually we're actually bullying that out. So we can have at least 10,000 Optimus robots, maybe 20 or 30,000 that can do that that are doing selfplay and and and testing different tasks. And then uh the the Tesla um has quite a good uh reality generator uh like a physics accurate reality generator that we we made made this for the cars. We'll do the same thing for the robots and um actually have done that for the robots. Um so uh so you have you know a few tens of thousands of humanoid robots uh doing different tasks and then you've got you can do millions of simulated robots in the simulated world and you use the uh the tens of thousands of of robots in the real world to close the simulation to reality gap close the sim to real gap. How do you think about the synergies uh between XAI and Optimus given you were highlighting look you need this world model you maybe want to use some really smart intelligence as a control plane um and so maybe Grock is like doing the slower planning and then like the motor policy is a little lower level
**Elon Musk：** 嗯，我们做人形机器人已经有一段时间了。我想已经五六年了吧。为汽车做的很多东西都适用于机器人。我们会在机器人里用跟汽车一样的 Tesla AI 芯片。我们会用同样的基本原理。它是非常类似的 AI。机器人比汽车有多得多的自由度。如果你只是把它想成比特流，AI 主要就是两个比特流的压缩和关联。对于视频，你得做大量的压缩，而且压缩方式得恰到好处。你得忽略不重要的东西。你不关心路边树上叶子的细节，但你非常关心路标和红绿灯、行人，甚至另一辆车里的人是不是在看你。有些细节非常重要。车最终要把那每秒 1.5 吉字节变成每秒 2 千字节的控制输出。所以有很多压缩阶段。你得把每个阶段都做对，然后把它们关联到正确的控制输出。机器人本质上要做同样的事情。这就是人类在做的。我们真的就是光子输入、控制输出。你生命的绝大部分就是这样：视觉，光子进来，然后运动控制输出。
**Elon Musk:** yeah what will the sort of synergy between these things be yeah so you just grock would orchestrate the behavior of the optimus robot so let's say you wanted to build a factory Um the then Optimus then Grock could uh organize the Optimus robots, give them assign them tasks uh to build the factory for to produce whatever you want. Don't you need to merge XAI and Tesla then cuz these things end up so
**Dwarkesh Patel：** 天真地看，在人形机器人和汽车之间……汽车的基本执行器就是怎么转向、怎么加速。机器人，特别是有灵活手臂的，有几十个自由度。然后尤其是 Tesla，你有这个优势——几百万小时的人类演示数据来自汽车在路上跑。你没法同样地部署一堆还不能用的 Optimus 然后以那种方式获取数据。所以在自由度增加和数据稀疏得多这两点之间……
**Dwarkesh Patel:** what were we saying earlier about public company discussions?
**Elon Musk：** 对。
**Elon Musk:** Well, we're one more Guinness in Elon. [laughter] Um,
**Dwarkesh Patel：** 说得好。你怎么用 Tesla 的智能引擎来训练 Optimus 的大脑？你其实点出了一个跟汽车的重要局限性和差异。
**Dwarkesh Patel:** what what are you waiting to see before you say we want to manufacture 100,000 optimism? Is it like
**Elon Musk：** 我们很快就会有 1000 万辆车在路上。很难复制那个巨大的训练飞轮。对于机器人，我们需要做的是造大量机器人，把它们放进一种 Optimus 学院（Optimus Academy），让它们在现实中做自我对弈（self-play）。我们实际上正在建造这个。我们可以至少有 10,000 个 Optimus 机器人，可能 20,000-30,000 个，在做自我对弈和测试不同的任务。Tesla 有一个相当好的现实生成器，一个物理精确的现实生成器，我们为汽车做的。我们会为机器人做同样的事情。我们其实已经为机器人做了。所以你有几万个人形机器人在做不同的任务。你可以在模拟世界里做几百万个模拟机器人。你用现实世界中的几万个机器人来弥合模拟与现实的差距。弥合 sim-to-real gap。
**Elon Musk:** optimia? Since we're defining the the proper noun, we could define the the plural of the proper noun, too. So, we we we're going to proper noun the plural. So, it's optim.
**Dwarkesh Patel：** 你怎么看 xAI 和 Optimus 之间的协同效应，考虑到你强调你需要这个世界模型，你想用一些真正聪明的智能作为控制平面，Grok 做更慢的规划，然后电机策略（motor policy）是更低层级的。这些东西之间的协同会是什么？
**Dwarkesh Patel:** Okay. Um, [laughter] is there something on the hardware side you want to see? Do you want to see better actuators or is it just you want the software to be better? What what are we waiting for before we get like mass manufacturing of Gen 3?
**Elon Musk：** Grok 会协调 Optimus 机器人的行为。比如说你想建一座工厂。Grok 可以组织 Optimus 机器人，分配任务来建造工厂，生产你想要的任何东西。
**Elon Musk:** No, we're moving towards that. We're we're going forward with some mass manufacturing.
**Dwarkesh Patel：** 那你不需要把 xAI 和 Tesla 合并吗？因为这些东西最终会那么……
**Dwarkesh Patel:** But do you think current um current hardware is good enough that you are going to you should you just want to deploy as many as possible now?
**Elon Musk：** 我们之前在说关于上市公司讨论的什么来着？
**Elon Musk:** I mean it's very hard to scale up production. I see.
**Dwarkesh Patel：** 我们已经多喝了一杯健力士了，Elon。你在等什么才决定说，我们要制造十万个 Optimus？
**Dwarkesh Patel:** Uh but uh yeah but I I think Optimus 3 is the the right version of the robot to you know to to produce maybe something on the order of like a million units a year. I think you'd want to go to Optimus 4 before you went to 10 million units a year.
**Elon Musk：** "Optimi"。因为我们在定义这个专有名词，我们也要定义这个专有名词的复数形式。我们要把复数形式也做成专有名词，所以是 Optimi。
**Elon Musk:** Okay. But you can do a millionear adoption with three. Uh yeah, I mean it's very hard to spool up at manufacturing.
**Dwarkesh Patel：** 有没有什么硬件方面你想看到的？你想看到更好的执行器？还是只是希望软件更好？在 Gen 3 大规模制造之前我们在等什么？
**Dwarkesh Patel:** Yes.
**Elon Musk：** 不，我们正在朝那个方向推进。我们正在推进大规模制造。
**Elon Musk:** Um so like manufacturing uh um like the the output per unit time is always follows an S-curve. So it starts off agonizingly slow then has this sort of eventually
**Dwarkesh Patel：** 但你觉得目前的硬件已经足够好了，只需要尽可能多地部署？
**Dwarkesh Patel:** exponential increase then a linear then a then a you know logarithmic outcome till you sort of eventually asmtote at some number. Optimus initial production will be it's going to be a it's going to be a stretched out scope because so much of what goes into Optimus is brand new. There's not an existing supply chain. Um as I mentioned the the actuators, electronics, everything in the office robot is designed um from physics first principles. It's not it's not taken from a catalog. These these are customd designigned everything. Literally everything. I don't think there's a single thing that um
**Elon Musk：** 扩展生产是非常困难的。但我觉得 Optimus 3 是合适的机器人版本，适合生产大约每年一百万台的规模。我觉得你在到每年一千万台之前要升级到 Optimus 4。
**Elon Musk:** how far down does that go? I mean, I guess we're not making custom capacitors yet, maybe. Um, but um there there's nothing you can pick out of a catalog um for at any price. Uh so so it just means that the the Optimus scope uh the the units units per output per unit time how many optimus robots you make per per day uh whatever is is going to initially ramp uh slower than a product where you have an existing supply chain. Um, but it will get to a million. When you see these Chinese humanoids like Unatri or whatever sell humanoids for like 6K or 13K, do you just like are you hoping to get your Optimus' bill of materials below that price so you can uh do the same thing or do you just think qualitatively they're not the same thing? Like what do you think is going like what allows it what allows them to sell for so low and can we match that? Well, optimus our optimus is is designed to have a lot of intelligence um and um to have the same electromechanical dexterity if not higher than a human. So, unitary does not have that. And it's also I mean it's it's quite a it's quite a big robot. So, it's cuz it's m it has to do uh you know carry heavy objects for long periods of time um and not overheat or exceed the power of its actuator. So, um, so we've got we've got we've got, you know, it's it's 5'11, you know, so it's pretty
**Dwarkesh Patel：** 好，但 Optimus 3 可以做到一百万台？
**Dwarkesh Patel:** tall. Um, and it's it's got a lot of intelligence. So, it's going to be more expensive than, um, a small robot that is not intelligent, but more capable. Yeah. But not a lot more. I mean, like the thing is over time, as Optimus robots build Optimus robots, the the cost will drop very quickly.
**Elon Musk：** 生产爬坡是非常难的。单位时间的产出总是遵循 S 曲线。开始时慢得令人痛苦，然后有一个指数增长阶段，然后是线性，然后是对数阶段，直到最终在某个数字上渐近。Optimus 的初始生产将是一个被拉长的 S 曲线，因为 Optimus 里太多东西是全新的。没有现成的供应链。执行器、电子元件，Optimus 机器人里的一切都是从物理第一性原理设计的。不是从目录里挑的。这些都是定制设计的。
**Elon Musk:** And what will these first billion Optimuses Optimai? Yeah.
**Dwarkesh Patel：** 这往下延伸到什么程度？我猜我们还没有在做定制电容器。
**Dwarkesh Patel:** Do like what will their highest and best use be?
**Elon Musk：** 没有任何东西你可以从目录里挑，不管什么价格。这只是意味着 Optimus 的 S 曲线，单位时间的产量，每天造多少 Optimus 机器人，初始爬坡会比一个有现成供应链的产品慢。但它会到达一百万。
**Elon Musk:** Uh I think you you would start off with with simple tasks that you can count on them doing well.
**Dwarkesh Patel：** 你看到那些中国人形机器人，像宇树（Unitree）之类的，卖 6,000 美元或 13,000 美元的人形机器人，你是希望把 Optimus 的物料成本（BOM）降到那个价格以下以便做同样的事？还是你觉得它们在本质上不是同一个东西？是什么让它们能卖这么低？我们能做到吗？
**Dwarkesh Patel:** But in the home or in factories like
**Elon Musk：** 我们的 Optimus 设计为拥有大量智能，并且具有与人类相同甚至更高的机电灵巧性。宇树没有那个。它也是一个相当大的机器人。它必须长时间搬运重物而不过热或超过执行器的功率。它 5 英尺 11 英寸高，挺高的。它有大量的智能。所以它会比一个小的、不智能的机器人更贵。
**Elon Musk:** the the best use for um robots in the beginning will be anything any um continuous operation. So any 24x7 operation cuz then you're cuz they can they can work continuously. Yeah.
**Dwarkesh Patel：** 但更有能力。
**Dwarkesh Patel:** What fraction of the work at a gigafactory that is currently done by humans could a gen 3 do?
**Elon Musk：** 但贵不了太多。关键是，随着时间推移，Optimus 机器人造 Optimus 机器人，成本会非常快地下降。
**Elon Musk:** Um I'm not I'm not sure. Or maybe it's like 10 20%. Maybe more. I don't know. It's we would but we would use we would not like reduce our headcount. We would we would for sure it can increase our headcount to be clear. Um but but we would increase our output. So the the um units produced per human like total to total number of humans at Tesla will increase but the um the output of robots and cars will inc will increase disproportionate like much much to you know number of cars and robots produced per human will increase dramatically but but number of humans will increase as well. We're talking about Chinese manufacturing um a bunch here and um we're also talking about, you know, we've talked about some of the policies that are relevant like you mentioned the uh the solar tariffs. Yeah. Uh and you think they're a bad idea because, you know, we can't uh scale up solar in the US.
**Dwarkesh Patel：** 最初的十亿个 Optimus，Optimi，会做什么？它们的最高最佳用途是什么？
**Dwarkesh Patel:** Well, just the electricity output in the US uh needs to scale up,
**Elon Musk：** 我觉得你会从简单的任务开始，你能指望它们做好的任务。
**Elon Musk:** right? And it can't without like good power sources. Need to get it somehow.
**Dwarkesh Patel：** 但是在家里还是在工厂里？
**Dwarkesh Patel:** Yeah. But I where I was going with this is if you were in charge, if you were setting all the policies, what else would you change?
**Elon Musk：** 机器人在早期的最佳用途将是任何连续运营的场景，任何 24/7 运转的场景，因为它们可以持续工作。
**Elon Musk:** Um, so you changed the solar tariffs as well.
**Dwarkesh Patel：** 目前超级工厂里由人类完成的工作中，Gen 3 能做多大比例？
**Dwarkesh Patel:** Yeah, I I would say anything that is limiting factor for electricity um needs to be addressed provided it's not like very bad for the environment.
**Elon Musk：** 我不确定。也许 10-20%，也许更多，我不知道。我们不会减少员工人数。我们会增加员工人数，说清楚。但我们会增加产出。每个人的产出……Tesla 的总人数会增加，但汽车和机器人的产出会不成比例地增长。每个人生产的汽车和机器人数量会大幅增加，但人数也会增加。
**Elon Musk:** So presumably some permitting reforms and stuff as well will be in there. Yeah, there's a fair bit of permitting reforms that are happening. A lot of the permitting is state based. Mhm.
**Dwarkesh Patel：** 我们这里聊了很多中国制造。我们也讨论了一些相关的政策，比如你提到的太阳能关税。你觉得那是个坏主意，因为我们没法在美国扩展太阳能。美国的电力产出需要扩展。没有好的电源就做不到。你只需要想办法搞到电。我想说的是，如果你来当家，如果你来制定所有政策，你还会改变什么？你会改变太阳能关税，那是一个。
**Dwarkesh Patel:** Um but anything but but but this this administration is is good at um removing permitting uh roadblocks.
**Elon Musk：** 我会说任何是电力限制因素的东西都需要解决，前提是它不会对环境造成很大伤害。
**Elon Musk:** Um and I'm not saying all tariffs are bad. I'm just saying because I think
**Dwarkesh Patel：** 所以大概一些审批改革之类的也会在里面？
**Dwarkesh Patel:** solar tariffs.
**Elon Musk：** 有不少审批改革正在进行。很多审批是州级的，但联邦层面的任何……这届政府在消除审批障碍方面做得不错。我不是说所有关税都不好。太阳能关税。有时候如果另一个国家在补贴某种产品的产出，那你就需要有反补贴关税来保护国内产业免受另一个国家的补贴冲击。
**Elon Musk:** So yeah. Yeah. I mean sometimes if like if another country is subsidizing the output of of something um then then you have to have counterveailing tariffs to protect domestic industry against uh subsidies by another country.
**Dwarkesh Patel：** 你还会改变什么？
**Dwarkesh Patel:** What else would you change? I don't know if there's that much that the government can actually do.
**Elon Musk：** 我不知道政府实际上还能做什么。有一件事我一直在想……
**Elon Musk:** One thing I was wondering is it seems like the for the policy goal of creating a lead for the US versus China. It seems like the export bans have actually been quite uh impactful where China is not producing leading edge chips and the export bands really bite there. China is not producing uh leading edge turbine engines and similarly there's a bunch of export bans that are relevant there on some of the metal energy. Should there be more export bans like you think about things like I mean there are now with the drone industry and things like that but is that something that should be considered? Well I think it's important to appreciate that in most areas China is very advanced in manufacturing.
**Dwarkesh Patel：** 关于为美国建立对中国的领先优势这个政策目标，出口禁令似乎确实产生了很大影响。中国没有在生产前沿芯片，出口禁令确实在那里起了作用。中国没有在生产前沿涡轮发动机。类似地，有一些出口禁令涉及某些冶金。应该有更多出口禁令吗？当你想到无人机产业之类的，这是应该考虑的事情吗？
**Dwarkesh Patel:** Um there's only a few areas where it is not. uh the you know China is a manufacturing powerhouse next level like people don't most people very impressive
**Elon Musk：** 重要的是要认识到，在大多数领域，中国在制造方面非常先进。只有少数几个领域它落后。中国是制造业强国，级别非凡。非常令人印象深刻。如果你看矿石精炼，中国大约做了全球其余国家总和两倍的矿石精炼。有些领域，比如精炼镓——镓用于太阳能电池。我记得他们占了镓精炼的 98%。所以中国在大多数领域的制造实际上非常先进。
**Elon Musk:** yeah yeah I mean if you if you take like refining of of ore um I'd say roughly China uh does more does twice as much ore refining of of of on average as the rest of the world combined um and and I think there there's some areas like say refining gallium which goes into solar cells. Um I think they're like 98% of gallium refining. Um so so China is actually very advanced in manufacturing in in I'd say most areas.
**Dwarkesh Patel：** 似乎大家对这种供应链依赖感到不安，但什么实际行动都没发生。
**Dwarkesh Patel:** It seems like we're like there is discomfort with this supply chain dependence and yet nothing's really happening on it.
**Elon Musk：** 供应链依赖？
**Elon Musk:** Supply chain which supply chain dependence depends on say like the gallium refining that you're saying.
**Dwarkesh Patel：** 比如说你提到的镓精炼。所有稀土的东西。
**Dwarkesh Patel:** Yeah. Yeah. There's there's there's there there's a there's a
**Elon Musk：** 稀土的话，你知道的，它们并不稀有。我们实际上在美国开采稀土矿石，把岩石挖出来，装上火车，然后装上船运到中国，再到另一列火车上，送到中国的稀土精炼厂，他们在那里精炼，做成磁铁，做成电机子组件，然后再运回美国。所以我们真正缺的是美国的大量矿石精炼能力。
**Elon Musk:** well the rare rare earth stuff and yeah rare earths which are as you know not rare
**Dwarkesh Patel：** 这不值得做政策干预吗？
**Dwarkesh Patel:** like we actually do do rare earth or mining in the US send the the rock uh put it on on a on a train and then put on a boat to China that goes on another train and goes to the um railro refining refiners in China who then refine it put it into a magnet put into a motor assembly and then send it back to America. So the thing we're really missing a lot of of all refining um in in America and
**Elon Musk：** 值得。我觉得在这方面有一些事情在做。但我们其实需要 Optimus，坦白说，来建矿石精炼厂。
**Elon Musk:** isn't this worth a policy intervention? Yes. Uh well I think there are some things being done on on that front.
**Dwarkesh Patel：** 所以你觉得中国的主要优势是丰富的技术劳动力？这就是 Optimus 要解决的问题？
**Dwarkesh Patel:** Um but but we kind of need optimist frankly to to build refineries. Um,
**Elon Musk：** 对。中国大约有我们四倍的人口。我的意思是，有一个顾虑。如果你认为人力资源决定未来，现在如果决定谁能造更多人形机器人的是制造业的技术劳动力，中国有更多。它造更多人形机器人，因此它最先拿到 Optimi 的未来。
**Elon Musk:** so sir you think the main advantage China has is the abundance of skilled labor and that that's like that's that's a thing optimist fixes but also we need
**Dwarkesh Patel：** 嗯，拿着看吧。也许吧。
**Dwarkesh Patel:** like four times our population
**Elon Musk：** 它就让那个指数增长继续下去。
**Elon Musk:** but we need so I mean there's this concern if you think like humanoids are the future that like okay right now if it's the skilled labor for manufacturing that's determining who's who can build more humanoids you know China has more of those it manufactures more humanoids therefore it gets it gets the optimized future first Um, it just like keeps that experiment going. It seems like you're sort of pointing out that sort of getting to a million optimi Yeah.
**Dwarkesh Patel：** 看起来你指出的是，达到一百万台 Optimi 需要的制造能力正是 Optimi 本来要帮我们获取的。对吧？
**Dwarkesh Patel:** requires the manufacturing that the optimize is supposed to help us get to,
**Elon Musk：** 你可以很快地闭合那个递归循环。
**Elon Musk:** right? You you can you can close that recursive loop pretty quickly with a small number of optimize.
**Dwarkesh Patel：** 用少量的 Optimi？
**Dwarkesh Patel:** Yeah. So, you close the recursive loop um to help help the robots build the robots. Um, and then we we can, you know, try to get to tens of millions of units a year. Maybe if you start getting to hundreds of millions of units a year, I I think you're you're going to be the most competitive country by far. We definitely can't win with just humans because China has four times our population,
**Elon Musk：** 对。你闭合递归循环，让机器人帮造机器人。然后我们可以试着达到每年几千万台。也许吧。如果你开始到每年几亿台，你就会成为遥遥领先的最有竞争力的国家。我们光靠人类绝对赢不了，因为中国有我们四倍的人口。坦白说，美国赢了太久了……一个赢了很久的职业运动队往往会变得自满和有优越感。这就是他们停止赢的原因，因为他们不再那么努力了。所以坦白说，我的观察就是中国的平均工作伦理比美国高。不仅仅是人口多四倍，而且每个人投入的工作量更大。所以你可以试着重新安排人力，但你仍然只有四分之一——假设人均生产力相同的话，我觉得实际上可能不是，我觉得中国在人均生产力上可能有优势——我们的产出将是中国的四分之一。所以我们在人力方面绝对赢不了。我们的出生率长期偏低。美国出生率大约从 1971 年起就低于更替水平了。我们有大量人退休，国内的死亡人数快要超过出生人数了。所以我们在人力方面绝对赢不了，但我们可能在机器人方面有一线机会。
**Elon Musk:** right? And frankly, America's been winning for so long that we, you know, just like a like a pro sports team that's been running for a very long time tend to get complacent and entitled. Um, and that's why they stop winning. Um, because it's, you know, don't work as hard anymore. Uh so I think the frankly just my observation is the average work ethic in China is higher than in the US. So it's not just that there's four times the population but the work the the amount of work that people put in is higher. Um so you you can like you can try to rearrange the humans but you're still one quarter of the uh you know assuming that that productivity is the health is is the same which I think actually it might not be. I think China might have an advantage on productivity per person. um we will do one quarter the amount of things as China. Um so so we we can't win on the human front. Um and our birth rates been low for a long time. So uh our birth rates been the US birth rates been below replacement uh since roughly 1971. Um, so, so we've got a lot of people retiring or, you know, more people dying than than than we're close to sort of more people domestically dying than than being born. Um, so we definitely can't win on the human front, but we we might have a shot at the robot front.
**Dwarkesh Patel：** 有没有其他你过去想制造但因为太劳动密集或太贵而做不了的东西，现在你可以回过头来说，"哦，我们终于可以做那个了，因为我们有 Optimus"？
**Dwarkesh Patel:** Are there other things that you have wanted to manufacture in the past, but they've been too labor intensive or too expensive that now you can come back to and say, "Oh, we can finally do the whatever." Uh because we have Optimus.
**Elon Musk：** 有的，我们想在 Tesla 建更多的矿石精炼厂。我们刚刚在德州科珀斯克里斯蒂（Corpus Christi）完成了锂精炼厂的建设并开始了锂精炼。我们在这里奥斯汀有一个镍精炼厂，用于正极材料。这是中国以外最大的正极精炼厂、最大的镍和锂精炼厂。正极团队会说，"我们有美国最大的，也是唯一的正极精炼厂。"不只是最大的，而且是唯一一个。
**Elon Musk:** Yeah, I think we'd like to do more build more um or refineries at Tesla. So um we just completed um construction and have um begun lithium refining um without lithium refinery in Corpus Christie, Texas. Uh we have um a nickel refinery which is called the cathode. Uh that's here in Austin. Mhm.
**Dwarkesh Patel：** 许多最高级别的头衔。
**Dwarkesh Patel:** Um and uh these these are the largest this is the largest cathode. It's the largest Catholic refinery, largest lithium refinery, largest nickel and lithium refinery uh outside of China.
**Elon Musk：** 所以还算大的，虽然它是唯一一个。但还有其他东西。你可以做更多的精炼厂，帮助美国在精炼能力方面更有竞争力。基本上有大量工作可以让 Optimus 来做，坦白说大多数美国人、非常少的美国人想做。
**Elon Musk:** Um and it's like the you know the cathode team would say like we have uh the the largest and the only actually cathode refinery in America. Many super not just the largest but it's also the only so it's pretty big even though it's the only one. Um, but I mean there are other things that uh you know um you could do a lot more refineries and and um help the the help America be more competitive on refining capacity. So So there's like there's basically a lot of work for for the optimal to do uh that that most Americans very few Americans frankly want to do. Uh I I mean I've I've actually Is the refining work too dirty or what's the
**Dwarkesh Patel：** 精炼工作太脏了还是怎么的？
**Dwarkesh Patel:** it's not is actually no we don't um there's not we don't have toxic emissions from the refinery or anything. Um like the cathode make a refiner right sort of in Travis County like 5 minutes from to
**Elon Musk：** 其实不是，不是。我们的精炼厂没有有毒排放什么的。正极镍精炼厂在特拉维斯县。
**Elon Musk:** Why can't you do it with humans? No you you can't you run out of humans.
**Dwarkesh Patel：** 为什么不能用人来做？
**Dwarkesh Patel:** Ah I see. Okay. Yeah.
**Elon Musk：** 可以，只是人会用完。
**Elon Musk:** Like no matter what you do you have one quarter the number of humans in America in China. So if you have them do this thing they can't do the other thing. So, so then then um well, how do you how do you build this refining refining capacity? Well, you could do it with Optima. Um and um not many not very many not very many Americans are are pining to do refining. [laughter] I mean, how many have you run into?
**Dwarkesh Patel：** 啊，我明白了。
**Dwarkesh Patel:** Very few.
**Elon Musk：** 不管你怎么做，美国的人数只有中国的四分之一。所以如果你让他们做这个，他们就没法做另一个。那你怎么建这些精炼产能？嗯，你可以用 Optimi。没有多少美国人渴望去做精炼工作。你遇到过多少？
**Elon Musk:** What are you saying? Very few plan to refine.
**Dwarkesh Patel：** 非常少。非常少渴望精炼的。比亚迪（BYD）正在达到 Tesla 的产量或销量水平。你觉得随着中国电动车产能扩张，全球市场会怎样？
**Dwarkesh Patel:** You know, BYD is reaching Tesla production or sales in quantity. What do you think happens in global markets as Chinese production in EV scales up?
**Elon Musk：** 中国在制造方面竞争力极强。所以我觉得会有大量中国汽车和基本上大部分制成品涌入市场。正如我所说，中国大概在做全球其余国家总和两倍的精炼。所以如果你往下看到第四第五层供应链……在最底层，你有能源，然后你有采矿和精炼。这些基础层级，正如我粗略估计的，中国在做全球其余国家总和两倍的精炼。所以任何东西都会有中国成分，因为中国在做两倍于全球其余国家的精炼工作。但他们会一直做到成品——汽车。我的意思是中国是一个强国。我觉得今年中国的电力产出将超过美国的三倍。电力产出是经济的一个合理代理指标。为了运转工厂和运转一切，你需要电力。它是实体经济的一个好的代理指标。如果中国超过美国电力产出的三倍，这意味着它的工业产能——粗略近似——将是美国的三倍。
**Elon Musk:** Um well uh China's extremely competitive in manufacturing. So um I think this there's going to be a a massive flood of Chinese vehicles and and and and other basically most manufactured uh things. I mean as it is as I said like China is like probably does twice as much refining as the rest of the world combined. So if you go you know if you just go go down to like fourth and fifth tier uh supply chain stuff like like like at the base level we've got energy then then you've got mining and refining. um those those those foundation layers uh are like I said China as a rough guess China is doing twice as much refining as the rest of the world combined. So any given thing is going to have uh Chinese content because China is doing twice as much manufact refining work as the rest of the world. Um and uh and then they they'll go all the way to the finished product with the cars. Uh I mean China is a powerhouse. I mean I think this year China will exceed three times US electricity output. Mhm.
**Dwarkesh Patel：** 读读言外之意的话，你说的似乎是，如果未来几年没有什么人形机器人递归奇迹的话，在整个制造/能源/原材料链上，中国就会统治一切，不管是 AI 还是制造电动车还是制造人形机器人。
**Dwarkesh Patel:** Like electricity output is a is a reasonable proxy for uh you know for the economy. Uh so like in order to run the factories and run run everything, you need electricity. So electricity is is is a it's a good proxy for the for the real economy. Um and so if China is if China passes three times the US electricity output, it means that its industrial capacity as a rough approximation is three times that will be three times that of the US. Reading between the lines, it sounds like what you're sort of saying is absent some sort of humanoid recursive miracle in the next few years on the the sort of like whole manufacturing energy uh raw materials chain like China will just dominate whether it comes to like AI or manufacturing EVs or manufacturing humanoids in the absence of of um breakthrough innovations uh in in the US. uh China will uh utterly dominate.
**Elon Musk：** 在美国没有突破性创新的情况下，中国将完全主导。
**Elon Musk:** Interesting. Yes.
**Dwarkesh Patel：** 有意思。
**Dwarkesh Patel:** Robotics being the main breakthrough innovation.
**Elon Musk：** 是的。
**Elon Musk:** Well, if you do like to to scale AI uh in in space like like basically need space,
**Dwarkesh Patel：** 机器人是主要的突破性创新。
**Dwarkesh Patel:** you need need the humanoid robots, you need real world AI, you need um a million tons a year to orbit. Um like let's just say like if we get the mass driver on the moon going my favorite thing. Um then I think uh
**Elon Musk：** 嗯，要在太空扩展 AI，你基本上需要人形机器人，需要真实世界 AI，需要每年一百万吨送入轨道。我们就说，如果我们把月球上的质量投射器搞起来了，我最喜欢的东西，那我就觉得——
**Elon Musk:** we'll have solved all our problems. Yeah. So this is like I call that winning. [laughter]
**Dwarkesh Patel：** 我们就解决所有问题了。
**Dwarkesh Patel:** Um
**Elon Musk：** 我管那叫赢了。我管那叫赢大了。
**Elon Musk:** I'd call that winning time.
**Dwarkesh Patel：** 你终于可以满意了。你做到了什么。
**Dwarkesh Patel:** You can finally be satisfied you've done something.
**Elon Musk：** 对。你有了月球上的质量投射器。我就想看那个东西运转。
**Elon Musk:** Yes. You have the mass driver on the moon.
**Dwarkesh Patel：** 那是来自什么科幻还是你从哪……？
**Dwarkesh Patel:** That's right. I just want to see that thing in operation.
**Elon Musk：** 嗯，实际上有一本 Heinlein 的书。《严厉的月亮》（The Moon is a Harsh Mistress）。
**Elon Musk:** Was that out of some sci-fi or where did you uh Well, actually the there is a Highland book. The moon. The moon is a harsh race.
**Dwarkesh Patel：** 好的，对，但那个有点不一样。那是一个重力弹弓还是……
**Dwarkesh Patel:** Okay. Yeah, but that's slightly different. That's a gravity slingshot or um
**Elon Musk：** 不，他们在月球上有一个质量投射器。
**Elon Musk:** No, they have a mass driver on the moon. Okay. Yeah, but they use that to attack Earth. So maybe it's not the greatest.
**Dwarkesh Patel：** 好的，对，但他们用那个来攻击地球。所以也许这不是最好的……
**Dwarkesh Patel:** They use that to uh assert their independence from
**Elon Musk：** 好吧，他们用那个来……争取独立。
**Elon Musk:** Exactly. What are your plans for the mass driver on the moon? They they assed their independence. [clears throat] Earth government disagreed and they loved things until they earth government agreed.
**Dwarkesh Patel：** 没错。你在月球上的质量投射器有什么计划？
**Dwarkesh Patel:** That book is a hoot. I found that book much better than um his other one that everyone reads um Stranger in a Strange Land. Yeah, Grock. Grock comes from a stranger on a strange line.
**Elon Musk：** 他们争取了独立。地球政府不同意，于是他们不停地投射东西直到地球政府同意了。那本书很搞笑。我觉得那本书比他的另一本大家都读的那本好得多，《异乡异客》（Stranger in a Strange Land）。
**Elon Musk:** Yeah. Yeah. But I much preferred. Yeah. The first two thirds of Stranger Lines are good and then it gets very weird in the third. Yeah. [snorts]
**Dwarkesh Patel：** "Grok"就来自《异乡异客》。
**Dwarkesh Patel:** Um but there's still some good concepts in there.
**Elon Musk：** 《异乡异客》前三分之二很好，然后第三部分就变得非常诡异了。但里面还是有一些好概念的。
**Elon Musk:** Yeah. Label box can get you robotics and RL data at scale. Take robotics. Let's say you need 100,000 hours of egocentric video. Labelbox starts by helping you define your ideal data distribution. Like for example, maybe no single task category should occupy more than 1% of trading volume. and at least 10% of trajectories should capture failure and recovery states. Next, Labelbox assigns its distribution to its massive network of operators. You're not limited to the small range of scenes that you can set up in a single warehouse. Instead, each one of Label Box's operators has access to lots of unique physical environments where they can film themselves completing a wide variety of tasks. Labelbox's tech automatically categorizes each video so that their operators always know which tasks will remain and what they need to work on next. For RL data, Labelbox takes a similar approach. They work with you to understand the right distribution of tasks and then their subject matter experts build the hyperrealistic digital environments and rubrics [music] that you need to collect the highest quality trading data. So whether you're training robots in the real world or agents for computer use, Labelbox can help. Go to labelbox.com/sarcash [music] to learn more.
**Dwarkesh Patel：** 我们之前讨论了很多的一件事是你管理人的方式。你面试了 SpaceX 的最初几千名员工，还有很多其他公司的员工。这显然是无法扩展的。
**Dwarkesh Patel:** One thing we were discussing a lot is kind of your system for managing people. Like you interviewed the first few thousand of SpaceX employees and lots of other companies.
**Elon Musk：** 嗯，是的，但什么没法扩展？
**Elon Musk:** What [snorts] is it doesn't scale? Well, yes, but but what doesn't scale
**Dwarkesh Patel：** 你。
**Dwarkesh Patel:** me? I mean,
**Elon Musk：** 当然，当然。我知道。
**Elon Musk:** sure. Sure. [laughter] I know that. But like what are you looking for?
**Dwarkesh Patel：** 但你在找什么？一天就只有那么多小时。不可能做到。但你在找什么东西是别的擅长面试和招聘的人——那种说不清道不明的感觉——
**Dwarkesh Patel:** I mean, it literally is not enough hours in the day. It's impossible. But well but um what are you looking for that's someone else who's good at interviewing and hiring people? What's the Jenna?
**Elon Musk：** 这个时候，我可能在评估技术人才方面拥有比任何人都多的训练数据——各种人才都算，但特别是技术人才——因为我做了那么多技术面试然后看到了结果。所以我的训练集是巨大的，而且覆盖面非常广。一般来说，我要求的是能证明卓越能力的要点证据。这些东西可能相当出人意料。不需要是特定领域的，但要有卓越能力的证据。所以如果一个人能举出哪怕一件事，但最好是三件事，让你觉得"哇，哇，哇"，那就是一个好迹象。
**Elon Musk:** Um well at this point I think I've got um I might have more training data on evaluating technical talent especially but talent of all kinds I suppose but uh technical talent especially um given that I've done so many technical interviews and then seen the results technical interviews seen the results. So my um my training set is is is very is enormous and uh has a very wide range. Um uh the generally the thing I ask for are u bullet points uh for evidence of of exceptional ability. So it's uh but like it's it's and these things can be like pretty offthe-wall. It doesn't need to be uh in the in the domain the specific domain but evidence that uh evidence of exceptional ability. Um so if some if if somebody can like site like even one thing but let's say three things where you go wow wow wow then that's that's a good sign. But but why do you have to be the one to determine that presumably it's impossible
**Dwarkesh Patel：** 为什么非得是你来判断？
**Dwarkesh Patel:** right?
**Elon Musk：** 不需要是我。我也做不到。这不可能。所有公司加起来总员工人数是 20 万人。
**Elon Musk:** I mean total headc count across all companies 200,000 people right [laughter]
**Dwarkesh Patel：** 但在早期，你在那些面试中找的什么是没法委派出去的？
**Dwarkesh Patel:** but in the early days what was it that that you were looking for that couldn't be delegated in those interviews? Um, well, I I guess I I'd need to build my training set. It's not like I would b a thousand here. Um, I would make mistakes, but then I'd be able to see where I I thought somebody would work out well, but they didn't. And and then why why did they not work out well and what can I do to I guess RL myself to uh in the future um have a better batting average when interviewing people.
**Elon Musk：** 我想我需要建立我的训练集。并不是说我的命中率百分之百。我会犯错，但然后我能看到我以为某人会表现很好但实际没有的情况。为什么他们没有表现好？我能做什么——我想是对自己做 RL——让未来面试人的时候命中率更高？我的命中率仍然不完美，但已经非常高了。
**Elon Musk:** Mhm. So, and my batting average is still not perfect, but it's it's very high. What are some surprising reasons people don't work out?
**Dwarkesh Patel：** 有哪些人没能成功的令人意外的原因？令人意外的原因……比如他们不懂技术领域等等。但你有那些长尾案例，"我对这个人非常兴奋。结果不行。"好奇为什么会这样。
**Dwarkesh Patel:** Surprising reasons? Um,
**Elon Musk：** 一般来说我告诉别人——我也告诉自己，作为期望——不要看简历。相信你的互动。简历可能看起来很亮眼，"哇，简历不错。"但如果 20 分钟的对话过后感觉不是"哇"，你应该相信对话，而不是纸面。
**Elon Musk:** like they don't understand technical domain, etc., etc., but like you you like you you've got like the long tail now of like I was really excited was about this person, it didn't work out. Curious why that happens. Uh, yeah. So the I mean generally what I tell people or tell myself I guess aspirationally um is don't look at the resume just believe your interaction. M so if the resume may may seem very impressive and it's like wow you know resume looks good but if the if the conversation uh after 20 minutes is is that conversation is not wow um you should believe the conversation not the not the not the paper. I feel like part of your method is that you know there was this meme in the media a few years back about Tesla being a revolving door of uh executive talent whereas actually I think when you look at it Tesla's had a very consistent and internally promoted executive bench over the past few years and that at SpaceX you have all these folks like Mark Josa and Steve Davis and
**Dwarkesh Patel：** 我觉得你方法的一部分是……几年前媒体有这个论调说 Tesla 的高管是旋转门。但实际上我觉得你看看的话，Tesla 在过去几年一直有非常稳定的、内部提拔的高管梯队。然后在 SpaceX，你有这些人比如 Mark Juncosa 和 Steve Davis——
**Dwarkesh Patel:** Steve Davis runs a boring company these
**Elon Musk：** Steve Davis 现在在管 The Boring Company。
**Elon Musk:** no now yeah but Bill Riley and folks like that
**Dwarkesh Patel：** Bill Riley，还有这些人。感觉有效的部分是拥有非常能干的技术副手。所有这些人有什么共同点？
**Dwarkesh Patel:** and it feels like part of has worked well is having very capable technical deputies. What do all of those people have in common?
**Elon Musk：** 嗯，Tesla 的高管团队，到目前为止平均任期大概有 10-12 年了。挺长的。但有些时期 Tesla 经历了极其快速的增长阶段，所以事情就加速了一些。你知道的，公司会经历不同的数量级规模变化。能帮助管理比如 50 人公司的人，跟管理 500 人公司、5,000 人公司、50,000 人公司所需要的人是不一样的。
**Elon Musk:** Uh well, so the I mean the Tesla is a sort of senior team uh at this point probably got average tenure of 10 or 12 years. It's quite quite a long tenure. Yeah. Um so um but there there were times where Tesla went through extremely rapid an extremely rapid growth phase um and so it was somewhat things were just somewhat sped up um and and when a company as as I'm as you know company goes through different orders of magnitude of of size you you know uh people that could who who could help manage say a 50 person company versus a 500 person company versus a 5,000 person company versus a 50,000 group people.
**Dwarkesh Patel：** 你超越了那些人。
**Dwarkesh Patel:** Yeah, it's it's just not the same team. It's not it's not always the same team. So if if a company is growing very rapidly, the the rate at which uh executive positions will change will also be proportionate to the the rapidity of the growth generally. Um then uh Tes Tesla had uh a further challenge where when when Tesla had very successful periods um uh we would be um relentlessly recruited from um like relentlessly um like when Apple had their electric car program they were coet bombing Tesla with recruiting calls. It was engineers just unplugged their phones like it's just it's just I
**Elon Musk：** 就是不一定是同一个团队。所以如果一家公司增长非常快，高管职位变动的速度也会与增长的速度成正比。Tesla 还有一个额外的挑战，就是当 Tesla 有非常成功的时期时，我们会被疯狂挖人。真的是疯狂地。当 Apple 有他们的电动车项目时，他们对 Tesla 进行了地毯式的招聘轰炸。工程师们直接拔电话了。"我正在想干点活呢。""如果我再接到一个 Apple 招聘的电话……"但他们不用面试就给的开价是 Tesla 薪酬的两倍。所以我们有一点"Tesla 精灵粉"的问题——就是"哦，如果你雇了一个 Tesla 的高管，突然一切就会成功。"我自己也上过精灵粉的当，觉得"哦，我们从 Google 或 Apple 挖个人，他们立刻就能成功"，但事情不是那样的。人就是人。没有什么神奇的精灵粉。所以当我们有精灵粉的问题时，我们就被疯狂挖人。而且，Tesla 是工程导向的，特别是它主要在硅谷，人们很容易就……他们不需要大幅改变生活。通勤路线差不多。
**Elon Musk:** I'm trying to get work done here. Yeah. if I get, you know, one more call from an Apple recruiter. Um, but but they were they were their opening offer without any interview with me like double the compensation at Tesla. Um, so um, so so we had a bit of the Tesla Pixie Dust thing where it's like, oh, if you hire a Tesla executive, you're suddenly you're going to everything's going to be successful. Um, and and I fallen prey to the Pixie Dust uh, you know, thing as well where it's like, oh, we'll hire someone from Google or Apple and they'll be immediately successful. But not that that's not how it works. Um you know people are people it's there's not like magical pixie dust.
**Dwarkesh Patel：** 那你怎么防止这个？怎么防止精灵粉效应——所有人都想挖你的人？
**Dwarkesh Patel:** Yes.
**Elon Musk：** 我不觉得我们能做太多来阻止它。这是 Tesla 为什么……在硅谷同时又有精灵粉效应，意味着招聘压力非常非常大。
**Elon Musk:** So when we have the pixie dust problem um we get relentlessly recruited um and um and then also being Tesla being um engineering especially being primarily in Silicon Valley uh it's it's easier for people to just like they don't have to change their life very much. They can just you know
**Dwarkesh Patel：** 那在奥斯汀应该有帮助？
**Dwarkesh Patel:** their commute's going to be the same. Yes.
**Elon Musk：** 奥斯汀，有帮助。Tesla 的工程力量大部分仍然在加州。让工程师搬家……我叫它"另一半"问题。
**Elon Musk:** Um, so how do you prevent that? How do you prevent the pixie dust effect for everyone's trying to coach all your people? Um, I don't think we can I don't think there's much we can do to to yeah, stop it. Uh but that that's like that's one of the reasons why Tesla uh really being in Silicon Valley um and uh and having the Pixie Dust thing at the same time um meant that uh there was just a very very aggressive recruitment.
**Dwarkesh Patel：** 对，"另一半"有自己的工作。
**Dwarkesh Patel:** Being in Austin helps then.
**Elon Musk：** 没错。所以对于 Starbase 来说这特别困难，因为找到一份非 SpaceX 工作的几率……
**Elon Musk:** Uh Austin, yeah, it still helps. Uh I mean Tesla still has a majority of its engineering in California. Um so um the you know for getting engineers to move I called the significant significant other problem.
**Dwarkesh Patel：** 在德州布朗斯维尔……
**Dwarkesh Patel:** Yes.
**Elon Musk：** ……相当低。挺难的。像是一个技术修道院的感觉，偏远而且主要是男的。
**Elon Musk:** So others have jobs. Yeah. Yeah. Yeah. Exactly. So um for Starbase that was particularly difficult.
**Dwarkesh Patel：** 跟旧金山比没多大改善。如果回到这些在 Tesla、SpaceX 等地方在技术岗位上非常有效的人，你觉得他们有什么共同点，除了……是他们在火箭或技术基础方面非常敏锐，还是你觉得是某种组织能力？是他们跟你合作的能力？是他们灵活但不太灵活的能力？什么是好的陪练伙伴？
**Dwarkesh Patel:** Yes. Since the odds of you know finding [snorts] a non SpaceX job
**Elon Musk：** 我不认为这是陪练伙伴的关系。如果一个人能把事情搞定，我就爱他们，如果不能，我就恨他们。所以挺直接的。不是什么特殊的偏好。如果一个人执行力好，我是超级粉丝，如果不好，我就不是。但这不是关于匹配我的特殊偏好。我肯定尽量不让它变成匹配我的特殊偏好。一般来说，我觉得应该招有天赋、有驱动力和值得信赖的人。而且我觉得善良也很重要。我在某个阶段低估了这一点。所以，他们是好人吗？值得信赖吗？聪明、有才华、努力吗？如果是的话，你可以加上领域知识。但那些基本特质、那些基本属性，你改不了。所以 Tesla 和 SpaceX 的大多数人并不是来自航空航天或汽车行业的。
**Elon Musk:** pretty low. Yeah. Yeah. It's quite quite difficult. I mean it's like a technology monastery thing.
**Dwarkesh Patel：** 随着你的公司从 100 人扩展到 1,000 人再到 10,000 人，你的管理风格最大的变化是什么？你以非常微观管理著称，就是深入到事情的细节里。
**Dwarkesh Patel:** Um you know remote and mostly dudes.
**Elon Musk：** 纳米管理，请。
**Elon Musk:** [laughter] But again, if you if you go
**Dwarkesh Patel：** 皮米管理。
**Dwarkesh Patel:** much of an improvement over myself, [laughter]
**Elon Musk：** 飞米管理。继续往下。我们一直到普朗克常数。一直到海森堡不确定性原理。
**Elon Musk:** yeah, if you go back but if you go back to these people who've really um been very effective in a technical capacity at Tesla, at SpaceX, and and those sorts of places, what do you think they have in common other than like is it just that they're very sharp on the, you know, rocketry or the, you know, the technical foundation? Or do you think it's something organizational? It's something about their ability to work with you. Is it their ability to like be, you know, flexible but not too flexible? What makes a good sparring partner for you?
**Dwarkesh Patel：** 你还能像你想的那样深入到细节里吗？你的公司如果更小一些会不会更成功？你怎么看这个？
**Dwarkesh Patel:** I don't think a sparring partner. I mean, if if somebody gets things done, I I I love them and if they don't, I So, it's pretty straightforward. It's not like some idiosyncratic uh thing. Um, if somebody executes well, um, I'm a huge fan and if they don't, I'm not. Um, but it's it's not about mapping to my is idiosyncratic preferences. I'll certainly try not to have it be mapping to my idiosyncratic preferences.
**Elon Musk：** 因为我一天的时间是固定的，随着事情增长和活动范围扩大，我的时间必然被稀释了。我不可能真的去微观管理因为那意味着我每天有几千个小时。让我去微观管理在逻辑上就是不可能的。但有时候我会深入到一个特定问题，因为那个特定问题是公司进展的限制因素。深入某个非常细节的问题的原因是它是限制因素。不是随便钻进去看无关紧要的小事。从时间角度来说，我随便深入无关紧要的小事在物理上就是不可能的。那样会导致失败。但有时候小事是决定胜负的关键。
**Elon Musk:** Mhm. Um, so yeah. Um, yeah, but generally I think it's a good idea to hire for um, uh, talent and drive and trustworthiness. M
**Dwarkesh Patel：** 有名的是，你把 Starship 的设计从复合材料换成了钢。
**Dwarkesh Patel:** um and I think uh goodness of heart is important. Um I I'd wait at that at one point. Um
**Elon Musk：** 对。
**Elon Musk:** so like are they are they a good person, trustworthy uh sort of smart, talented and hardworking? Uh if so you can add domain knowledge. U but those those fundamental traits those fundamental properties you cannot change. So mo most of the people who um are at Tesla and SpaceX did not come from the aerospace industry or the auto industry. What is most sad to change about your management style as your companies have scaled from 100 to thousand to 10,000 people. You're you know you're known for this like very micromanagement just getting into the details of things. Nano management please. [laughter]
**Dwarkesh Patel：** 你做了那个决定。不是别人到处说"老板，我们找到了更好的方案"。是你在推，而且遇到了一些阻力。能告诉我们你是怎么想到整个换钢方案的吗？
**Dwarkesh Patel:** Pico management.
**Elon Musk：** 我想说是走投无路吧。最初，我们打算用碳纤维做 Starship。碳纤维相当贵。当你做量产的时候，你可以让任何东西开始接近它的材料成本。碳纤维的问题是材料成本本身就非常高。特别是如果你用一种能承受低温液氧的高强度专业碳纤维，它的成本大约是钢的 50 倍。至少理论上它会更轻。人们通常觉得钢是重的而碳纤维是轻的。对于室温应用，比如 F1 赛车、静态气动结构，或者任何气动结构，你可能用碳纤维更好。问题是我们试着用碳纤维做这个巨大的火箭，进展极其缓慢。
**Elon Musk:** Um phantom management. So you're saying [laughter]
**Dwarkesh Patel：** 一开始选碳纤维就是因为它轻？
**Dwarkesh Patel:** we're gonna go all the way down to flanks Boston all the way down to Heisenberg. I said they were small. [laughter]
**Elon Musk：** 对。乍一看，大多数人会觉得做轻量化的选择应该是碳纤维。问题是当你用碳纤维做一个非常巨大的东西，然后你试着让碳纤维高效固化——也就是不是室温固化，因为有时候你可能有 50 层碳纤维——碳纤维本质上是碳丝和胶水。为了有高强度，你需要一个高压釜（autoclave）。一个基本上是高压烤箱的东西。如果东西很巨大，那个高压釜得比火箭还大。我们试图做一个比历史上任何高压釜都大的高压釜。或者你可以做室温固化，但那需要很长时间而且有问题。最终的问题是我们用碳纤维的进展非常非常慢。
**Elon Musk:** Yeah. How do you I mean are you are you still able to get into details as much as you want? Would your companies be more successful if you could if they were smaller? Like how do you how do you think about that? Well, because I have a fixed amount of time in the day, uh my time is necessarily um diluted as things grow and as the span of activity uh increases. So, you know, um it it it it's it's impossible for me to actually be a micromanager because uh there's that that would imply I have some like thousands of hours per day. It is a logical impossibility for me for me to mic to micromanage things. Um so now there are times when um I will drill down into uh a specific issue because that specific issue uh is the limiting factor on uh the progress of the company. Um and um but the the reason for drilling into that that some very detailed item is because it is the it is the limiting factor not it's not arbitrarily d drilling into you tiny things. Um and and like I said obviously from a time standpoint it is physically impossible for me to arbitrarily uh go into tiny things that don't matter and that would and and that would result in failure but sometimes the tiny things um are decisive in victory. Famously you switched the uh starship design from composits to steel.
**Dwarkesh Patel：** 元问题是为什么这个决定必须由你来做。你的团队有很多工程师。团队为什么没有自己得出用钢的结论？
**Dwarkesh Patel:** Yes.
**Elon Musk：** 对，没错。这是一个更广泛问题的一部分，就是理解你在你的公司里的比较优势。
因为我们用碳纤维的进展非常慢，我就说，"好了，我们得试试别的。"对于 Falcon 9，主结构用的是铝锂合金，它有非常好的强度重量比。实际上它可能在其应用中有跟碳纤维差不多甚至更好的强度重量比。但铝锂合金非常难加工。要焊接它，你必须做一种叫搅拌摩擦焊（friction stir welding）的东西，在不进入液相的情况下接合金属。能做到这一点其实很神奇。但用这种特殊的焊接方式你能做到。非常困难。比如说你想做修改或者在铝锂合金上附加什么东西，你现在得用带密封的机械连接。你没法焊上去。所以我想避免用铝锂合金做 Starship 的主结构。有一种非常特殊等级的碳纤维，质量性能很好。对于火箭，你真正在做的是最大化火箭中推进剂的比例，最小化质量。但正如我说的，我们进展非常慢。我说，"按这个速度，我们永远到不了火星。所以我们得想别的办法。"我不想用铝锂合金因为搅拌摩擦焊的困难，特别是做大尺寸的时候。3.6 米直径就够难的了，更别说 9 米或以上。然后我说，"钢怎么样？"我有一个线索，因为一些早期的美国火箭用过很薄的钢。Atlas 火箭用过钢气球罐。不是说钢从来没被用过。实际上用过。当你看不锈钢的材料特性，全硬的、应变硬化的不锈钢，在低温下强度重量比实际上跟碳纤维差不多。如果你看室温的材料特性，看起来钢会重两倍。但如果你看特定等级的全硬不锈钢在低温下的材料特性，你实际上能达到跟碳纤维差不多的强度重量比。在 Starship 的情况下，燃料和氧化剂都是低温的。Falcon 9 的燃料是火箭级煤油，基本上是一种非常纯的喷气燃料。那大约是室温的。虽然我们实际上确实把它稍微冷却了一下，我们像冰镇啤酒一样冰镇它。
因为我们用碳纤维的进展非常慢，我就说，"好了，我们得试试别的。"对于 Falcon 9，主结构用的是铝锂合金，它有非常好的强度重量比。实际上它可能在其应用中有跟碳纤维差不多甚至更好的强度重量比。但铝锂合金非常难加工。要焊接它，你必须做一种叫搅拌摩擦焊（friction stir welding）的东西，在不进入液相的情况下接合金属。能做到这一点其实很神奇。但用这种特殊的焊接方式你能做到。非常困难。比如说你想做修改或者在铝锂合金上附加什么东西，你现在得用带密封的机械连接。你没法焊上去。所以我想避免用铝锂合金做 Starship 的主结构。有一种非常特殊等级的碳纤维，质量性能很好。对于火箭，你真正在做的是最大化火箭中推进剂的比例，最小化质量。但正如我说的，我们进展非常慢。我说，"按这个速度，我们永远到不了火星。所以我们得想别的办法。"我不想用铝锂合金因为搅拌摩擦焊的困难，特别是做大尺寸的时候。3.6 米直径就够难的了，更别说 9 米或以上。然后我说，"钢怎么样？"我有一个线索，因为一些早期的美国火箭用过很薄的钢。Atlas 火箭用过钢气球罐。不是说钢从来没被用过。实际上用过。当你看不锈钢的材料特性，全硬的、应变硬化的不锈钢，在低温下强度重量比实际上跟碳纤维差不多。如果你看室温的材料特性，看起来钢会重两倍。但如果你看特定等级的全硬不锈钢在低温下的材料特性，你实际上能达到跟碳纤维差不多的强度重量比。在 Starship 的情况下，燃料和氧化剂都是低温的。Falcon 9 的燃料是火箭级煤油，基本上是一种非常纯的喷气燃料。那大约是室温的。虽然我们实际上确实把它稍微冷却了一下，我们像冰镇啤酒一样冰镇它。
**Elon Musk:** And you made that decision like that wasn't a you know people were going around they're like oh we found something better boss. like that was you encouraging people against some resistance. Can you tell us how you came to that whole composite steel switch? [snorts] Uh yeah. So desperation I'd say. Um the um originally yeah we were going to make stashup out of uh carbon fiber. Um and um car fiber is pretty expensive like the the the the can generally uh when you do volume production you can get any given thing to be to start to approach its material cost. The problem with with carbon fibers is that material cost is still very high. Um um so um it's about it's about 50 times particularly if you go for a high strength specialized carbon fiber that can handle um cryogenic oxygen it's it's it's like quot roughly 50 times the cost of steel um and at least uh in theory it would be lighter. People generally think of steel as being heavy and carbon fiber as being uh light. Um, and for room temperature room temperature applications, um, you know, like say, uh, more or less room temperature applications like a Formula 1 car, uh, static aeros structure or any any kind of aeros structure really, uh, is is going to you're going to probably be better off with the carbon fiber. Um, now the problem is that we were trying to make this enormous rocket out of carbon fiber and, uh, our progress was extremely slow and it had been picked in the first place just because it's light. Yes. Um like at first glance um like most people would think that the the the choice for making uh something light would be carbon fiber. Um the um now now the thing is that um we when you make something very enormous out of carbon fiber and then you try to have the carbon fiber um be efficiently cured. anything not not room temperature because like you've got you know sometimes you got like 50 pies of of of carbon fiber and and carbon fiber is really carbon string and glue. Um and uh and you in order to have um high strength you need an autoclave. So something that that can that's essentially high pressure oven. And if if um if you have something that's a gigantic uh the oven's got to be bigger than the rockwood. Um, so we'll be trying to make the the an autoclave that's bigger than any autoclave that's ever existed. Uh, or do room temperature cure, which takes a long time and and has issues. Um, but but the fundamental issue is that we're just making very slow progress uh with uh with cotton fiber. Um, so um I I think the meta question is u why it had to be you who made that decision. There's many engineers on your team.
**Dwarkesh Patel：** 好喝。
**Dwarkesh Patel:** Yeah. How did the team not arrive at Steel? Yeah, exactly. This is a part of a broader question like understanding your comparative advantage at your companies.
**Elon Musk：** 我们确实冰镇了，但不是低温。事实上如果我们把它弄到低温，它就会变成蜡了。但对于 Starship，是液态甲烷和液氧。它们在相似的温度下是液态。基本上，几乎整个主结构都在低温状态。所以你用一种应变硬化的 300 系不锈钢。因为几乎所有东西都在低温状态，它实际上有跟碳纤维差不多的强度重量比。但原材料成本低 50 倍，而且非常容易加工。你可以在户外焊接不锈钢。你可以一边抽雪茄一边焊不锈钢。非常耐用。你可以轻松修改。如果你想附加什么东西，直接焊上去。非常容易加工，成本非常低。正如我说的，在低温下强度重量比跟碳纤维差不多。然后当你考虑到我们大大减少了热防护层的质量——因为钢的熔点远高于铝的熔点……大约是铝的两倍的熔点。
**Elon Musk:** Um, so it was because we were making very slow progress with with carbon fiber. I was like, "Okay, we've got to try something else." Now, for the Falcon 9, the the primary airframe is made of aluminum lithium, which is has very very good strength weight. Um and um actually it has uh about the same maybe maybe better strength weight for its application than carbon fiber. But aluminum lithium is very difficult to work with. In order to weld it you have to do something called friction st welding where you join the you join the metal without it entering the liquid phase. Um so it's kind of wild that you could do that. Uh but with this particular type of welding you can do that. Um but uh it's very difficult to like say let's say you want to make a modification or attach something to um aluminum lithium. You now have to use mechanical attachment with seals. Um you can't uh weld it on. Um so we want to I want to avoid using aluminum lithium for the primary structure for uh for Starship. um and uh and and there was this very special grade of uh carbon fiber that that had you know very good mass properties. So with rock rocket you're really trying to maximize the percentage of the of the rocket that is propellant minimize the the mass obviously and um the but like I said we're making very slow progress um and and and I said at this rate we're never going to get to Mars so we're going to think of something else um I didn't want to use aluminum lithium because of the difficulty of friction still welding um especially doing that at at scale it was hard enough um at 3.6 m in diameter, let alone at 9 m or above. Um then um it says, well, what about steel? And and so the now I I I had a clue here because some of the early um US rockets had used very thin steel. The Atlas rockets had used a steel balloon tank. Um so it's not like steel never been used before. It actually had been used. Um and when you look at the the material properties of stainless steel um especially uh very uh if it's been very like full hard strain hardened stainless steel uh at cryogenic temperature uh the the strength weight is actually similar to carbon fiber. So if if you look at the so if you look at the material properties at room temperature um it looks like the steel is uh is going to be twice as heavy. But if you look at the material properties at cryogenic temperature of full hard steel stainless of of particular grades uh then the the you actually get to a similar strength weight as common fiber and and in the case of Starship both the fuel and the oxidizer are cryogenic. So for for uh Falcon 9, the fuel is rocket propellant grade kerosene basically pure like a a very pure form of jet fuel. Um which is but but that is that is roughly room temperature. Um although we do actually we do actually chill it slightly below we chill it like a beer. Um delicious. Yeah, we we do chill it but um but but it's not cryogenic. In fact, if we made it cryogenic, it would just turn to wax. So um but for Sasha the it's liquid methane and and liquid oxygen they they uh they're liquid at at similar temperatures. Uh so uh so basically almost the entire primary structure is a cryogenic temperature. So then you've got uh u a 300 series stainless that's that's um strain hardened uh because it's at almost all things at crying temperature actually has a similar strength to weight as uh carbon fiber but costs uh 50 times less than nor material and is very easy to work with. You you can weld stainless steel outdoors. You could smoke a cigar while welding stainless steel. It's it's like it's it's very resilient.
**Dwarkesh Patel：** 所以你可以让火箭运行在更高的温度？
**Dwarkesh Patel:** Um you you can modify it easily. It's it's uh if you want if you want to attach something, you just weld it right on. So um very easy to work with uh very low cost um and um like I said at cryogenic temperature, similar strength to weight uh to carbon fiber. Um then when you factor in that uh that we don't need we don't we we have a much reduced uh heat shield mass uh because the melting point of steel is much greater than the melting point of aluminum. Um it's about twice the melting point of alum aluminum and
**Elon Musk：** 对，特别是对飞船来说，它是像一颗燃烧的流星一样返回的。你可以大大减少热防护层的质量。你可以把迎风面热防护层的质量砍掉大概一半，而且背风面不需要任何热防护。最终结果是钢制火箭实际上比碳纤维火箭更轻，因为碳纤维火箭里的树脂会开始熔化。基本上碳纤维和铝有差不多的工作温度能力，而钢可以在两倍的温度下工作。这些都是非常粗略的近似。我不会去造火箭的。我的意思是人们会说，"哦，他说两倍。实际上是 0.8。"我就说，闭嘴，你们这些混蛋。
**Elon Musk:** so you can just run the rocket bunch hotter.
**Dwarkesh Patel：** 那将是主要评论区的内容。天杀的。
**Dwarkesh Patel:** Yes. So especially for the ship um which is coming in like a fl a blazing meteor uh it is uh the you you you can greatly reduce the mass of the heat shield um so that so you can call it cut the mass of the windward u part of the heat shield in maybe in half
**Elon Musk：** 关键是，回过头看，我们一开始就应该用钢。不用钢是蠢的。
**Elon Musk:** and you don't need any heat shielding on the on the leeward side. Um so um the the net net result is actually the steel rocket weighs less than the carbon fiber rocket because the the resin in the carbon fiber rocket uh uh um starts to melt. Um so basically carbon fiber and aluminum have about the same operating temperature uh capabilities um and whereas steel can operate at twice temperature. I mean, these are very rough approximations. People will
**Dwarkesh Patel：** 好，但反过来跟你说，我听到的是，钢是一个风险更高、没那么被验证的路径，除了早期的美国火箭。而碳纤维是一个更差但更被验证的路径。所以需要你来推"嘿，我们要走这条风险更高的路然后搞定它。"你是在对抗一种保守主义。
**Dwarkesh Patel:** I won't go to the rocket face. What I'm like people will say, "Oh, he said it's twice. It's actually it's actually point8." Shut up, [...]
**Elon Musk：** 这就是为什么我一开始说问题在于我们的进展不够快。我们连做一小段碳纤维筒体都有问题，做出来没有褶皱。因为在那么大的尺度下，你得有很多层碳纤维。你得固化它，而且固化方式不能有任何褶皱或缺陷。碳纤维的韧性远不如钢。韧性差得多。不锈钢会拉伸和弯曲，碳纤维倾向于碎裂。韧性就是应力应变曲线下的面积。一般来说，用钢你的表现会更好，但准确说是不锈钢。
**Elon Musk:** That's what the main comment's going to be about. Yeah. God damn it. Okay. The point is the the actually in retrospect the the we should have started with done steel in the beginning. It was dumb not to do steel.
**Dwarkesh Patel：** 还有一个 Starship 问题。我大概两年前跟 Sam Teller 一起去了 Starbase 参观，那太棒了。从很多方面来看都很酷。我注意到的一件事是人们真的为事情的简洁性感到自豪。每个人都想告诉你 Starship 就是一个大汽水罐，我们在招焊工，如果你能在任何工业项目里焊接，你就能在这里焊。但 Starship 是一个非常复杂的火箭——对，以简洁性为荣。
**Dwarkesh Patel:** Okay. But to play this back to you, what I'm hearing is that steel was a riskier, less proven path other than the early US rockets versus carbon fiber was like a worse but more proven out path. And so you need to be the one to push for, hey, we're going to do this riskier path and just figure it out. And so you were fighting like a sort of conservatism in a sense.
**Elon Musk：** 嗯，事实上 Starship 是一个非常复杂的火箭。
**Elon Musk:** Um that's why I initially said like that the issue is that we weren't making fast enough progress. we were having trouble making even um a small barrel section of the carbon fiber um that didn't have wrinkles in it. Um so uh because at at at that large scale you have to have many pies many sort of layers of the carbon fiber. Um you've got to cure it and you've got to cure it in such a way that it it doesn't um have any wrinkles or or defects. The carbon fiber is much less resilient than than steel. It has much it's less toughness. Um so like like stainless steel will will stretch and and and bend the carbon fiber will will tend to shatter.
**Dwarkesh Patel：** 所以我想问的是，事情到底是简单的还是复杂的？
**Dwarkesh Patel:** Um so um so toughness being the area under the stress drain curve um so that you're generally going to have to do better with steel um or stainless steel to be precise.
**Elon Musk：** 我想也许他们想说的是你不需要有火箭行业的经验就能在 Starship 上工作。一个人只需要聪明、努力、值得信赖就能做火箭。不需要火箭经验。Starship 是人类有史以来制造的最复杂的机器，而且差距很大。
**Elon Musk:** One other starship question. Um so I visited um Starbase I think it was two years ago with Seller and that was awesome. It was very cool to see in a whole bunch of ways. One thing I noticed was that people really took pride in the simplicity of things where you know everyone wanted to tell you how Starship is just a big soda can and you know we're hiring welders and you know if you can weld in any industrial project you can weld here but um there's a lot of pride in the simplicity and it's well startup was a very complicated
**Dwarkesh Patel：** 在什么方面？
**Dwarkesh Patel:** rocket. So that that's what I'm getting at is are things simpler or are they complex?
**Elon Musk：** 任何方面，真的。我觉得没有比它更复杂的机器了。我觉得基本上我能想到的任何项目都会比这个容易。这就是为什么从来没有人做过完全可重复使用的轨道火箭。这是一个非常难的问题。很多聪明人之前试过，非常聪明的人拥有巨大的资源，他们失败了。而且我们也还没有成功。Falcon 是部分可重复使用的，但上面级不是。Starship 第三版，我认为这个设计可以完全可重复使用。那种完全可重复使用将使我们成为多行星文明。任何技术问题，即使像大型强子对撞机（Hadron Collider）之类的，都比这个容易。
**Elon Musk:** I think maybe just what they're trying to say is that you know you don't have to have like prior experience in the rocket industry to work on Sasha. Um you know somebody just needs to be
**Dwarkesh Patel：** 我们花了很多时间讨论瓶颈。你能说说目前 Starship 的瓶颈是什么吗，哪怕是高层面的？
**Dwarkesh Patel:** you know smart and work hard um and be trustworthy then they can work on a rocket. They don't they don't need prior rocket experience. Starship is is the most complicated machine ever made by humans by a long shot.
**Elon Musk：** 总的来说就是让它不爆炸。它真的很想爆炸。
**Elon Musk:** In in what regards?
**Dwarkesh Patel：** 老问题了。那些可燃材料。
**Dwarkesh Patel:** Anything really? I'd say there isn't a more complex machine. Um there Yeah, I mean I I'd say that there there pretty much any any project I can think of would be easier than this. Um and and that's why no one has made a rapidly reusable nobody has ever made a re fully reusable orbital rocket. It's a very very hard problem. Um I mean many smart people have tried before very smart people with immense resources and they failed. Um so and we haven't succeeded yet. uh you know Falcon is partially reusable but the upper stages are um Starship version three I think this design that it it can be fully reusable and that full reusability is what will enable us to become a multilanet civilization. Can you say about the
**Elon Musk：** 我们已经有两个助推器在测试台上爆炸了。有一个把整个测试设施炸没了。所以只需要一个错误。Starship 里包含的能量是疯狂的。
**Elon Musk:** I don't I'm like I said like I any technical problem even like a hydron collider or something like that is easier following than this. We we spent a lot of time on bottlenecks. Can you say what the current Starship bottlenecks are even at a high level?
**Dwarkesh Patel：** 这就是为什么它比 Falcon 更难？因为能量更大？
**Dwarkesh Patel:** I mean trying to make it not explode generally [snorts]
**Elon Musk：** 有很多新技术。它在推动性能的边界。Raptor 3 引擎是一个非常非常先进的引擎。它是迄今为止最好的火箭引擎。但它拼命想爆炸。给你一些概念，在起飞时火箭产生超过 100 吉瓦的功率。那是美国电力的 20%。
**Elon Musk:** that old chestnut really wants to explode. Um
**Dwarkesh Patel：** 真是疯了。这是一个很好的类比。
**Dwarkesh Patel:** those combustion we've had two bristers explode on the test. Um, one obliterated obliterated the entire test facility. Um, so it only takes like one mistake and and I mean the amount of energy contained in in a Starship [snorts] is insane.
**Elon Musk：** 同时不爆炸。有时候。有时候是的。所以我就想，它怎么没爆炸？有成千上万种爆炸的方式，但只有一种不爆炸的方式。所以我们不仅想让它真的不爆炸，还想让它每天可靠地飞行，每小时一次。显然，如果它经常爆炸，就很难维持那个发射节奏。
**Elon Musk:** And so is that why it's harder than Falcon? It's because it's just more energy. It's a lot of new technology. Um, it's it's p it's pushing the performance envelope. Um, the Raptor 3 engine is a very very advanced engine. By far the best rocket edition ever made. Um but it desperately wants to blow up. I mean just to put things in perspective here on liftoff um the the rocket is generating over 100 g of power. It's 20% of US
**Dwarkesh Patel：** 对。Starship 目前最大的单一问题是什么？
**Dwarkesh Patel:** electricity. So insane.
**Elon Musk：** 是让热防护层可重复使用。没有人做过可重复使用的轨道热防护层。所以热防护层得在上升阶段不掉落一大堆瓷砖，然后返回的时候也不能掉很多瓷砖或让主体结构过热。
**Elon Musk:** That's a great comparison. While not exploding
**Dwarkesh Patel：** 那个很难是不是因为它本质上就是消耗品？
**Dwarkesh Patel:** sometimes.
**Elon Musk：** 嗯，是的，但你车上的刹车片也是消耗品，但它们能用很长时间。
**Elon Musk:** Sometimes but sometimes. Yeah. So I was like how does it not explode? there's there's a you know thousands of ways that it could explode and and only one way that that that it doesn't. So So we want it to not really not explode but but fly reliably uh on a daily basis like once per hour and obviously you know blows up a lot. It's it's very difficult to maintain that floor cings. Yes.
**Dwarkesh Patel：** 说得对。所以它只需要用很长时间。
**Dwarkesh Patel:** Um and [snorts] then and then I'm going to say like what's the what's the single biggest remaining problem for Starship? It's uh having the heat shield be reusable. Um that such that the no no one has ever made a reusable orbital heat shield. Um so the the sh the heat shield's got to make it through the ascent phase without shocking a bunch of tiles. Um and then it's going to come back in and also not lose a bunch of tiles or or overheat the the main the main uh airframe.
**Elon Musk：** 我们已经让飞船返回并在海面上软着陆了。我们做过好几次。但它掉了很多瓷砖。在不做大量维修的情况下它是不可重复使用的。虽然它确实软着陆了，但没有大量维修的话它不可重复使用。这在那个意义上不算真正可重复使用。这是剩下的最大问题，完全可重复使用的热防护层。你要能着陆，加注推进剂然后再飞。你不能做那种对 40,000 块瓷砖的费力检查。
**Elon Musk:** Isn't that hard? is kind of fundamentally a consumable. Uh well, yes, but your brake pads in your car are also consumable, but they last a very long time.
**Dwarkesh Patel：** 当我读你的传记时，感觉你就是能推动这种紧迫感，推动"这才是能规模化的东西"这种感觉。我好奇为什么你觉得你这样规模的组织——SpaceX 和 Tesla 现在已经是非常大的公司了——你仍然能保持那种文化。其他公司出了什么问题以至于他们做不到？
**Dwarkesh Patel:** Fair. Okay.
**Elon Musk：** 我不知道。
**Elon Musk:** So, it just needs to last a very long time. Um but that's it just you try I mean we have brought the ship back and had it do a soft landing in the ocean. I've done that a few times, but but it lost a lot of tiles, you know, and you know, it was [snorts] not reusable without a lot of work. Yeah.
**Dwarkesh Patel：** 比如今天，你说你有一堆 SpaceX 的会议。你在那里做什么来保持这种状态？是在增加紧迫感？
**Dwarkesh Patel:** So, even though it did land it did come to soft landing, it it was would not have been reusable without a lot of work. Um and and that so it's not really reusable in that sense. So, that's that's the biggest problem that remains is fully reusable heat shield. Um, so so like if you want to be able to land it, uh, refill propellant and fly again, uh, without, you know, you can't go do this laborious inspection of, you know, 40,000 tiles type sort of thing. I I I'm curious how you drive like when when I read biographies of yours, it just uh it seems like you're just able to drive the sense of like urgency and drive the sense of like this is the this is the thing that can scale. Um, and I I'm curious why you think other organizations of your like SpaceX and Tesla are really big companies now and you're still able to keep that culture. What goes wrong with other companies such that they're not able to do that?
**Elon Musk：** 嗯，我不知道。我想紧迫感来自领导公司的人。我有一种疯狂的紧迫感。所以那种疯狂的紧迫感投射到了公司的其他部分。
**Elon Musk:** I don't know. Um,
**Dwarkesh Patel：** 是因为有后果吗？他们觉得"Elon 定了一个疯狂的截止日期，但如果我没做到，我知道会怎样"。是你能够识别瓶颈并消除它们所以人们能快速推进？你怎么看你的公司为什么能快速推进？
**Dwarkesh Patel:** but like today you said you had like a bunch of SpaceX meetings like what what is it that you're doing there that's like keeping that
**Elon Musk：** 我不断地解决限制因素。在截止日期方面，我一般尝试瞄准一个我至少认为在第 50 百分位的截止日期。所以不是一个不可能的截止日期，但它是我能想到的最激进的、有 50% 概率能实现的截止日期。这意味着有一半的时间会迟到。有一种气体膨胀定律适用于日程安排。如果你说我们五年内做某件事——对我来说五年就像是无穷长——它就会膨胀到填满所有可用的日程然后正好用五年。物理学会限制你做某些事情的速度。所以扩展制造，有一个你能移动原子和扩展制造的速率。这就是为什么你不能立刻就做到每年一百万台。你得设计生产线。你得把它立起来。你得走 S 曲线。
我能说什么真正对人有帮助的？一般来说，疯狂的紧迫感是非常重要的。你要有一个激进的时间表，你要弄清楚在任何时间点上限制因素是什么，然后帮团队解决那个限制因素。
我能说什么真正对人有帮助的？一般来说，疯狂的紧迫感是非常重要的。你要有一个激进的时间表，你要弄清楚在任何时间点上限制因素是什么，然后帮团队解决那个限制因素。
**Elon Musk:** that's adding urgency? Yeah. Yeah.
**Dwarkesh Patel：** Starlink 是在酝酿了很多年后慢慢做起来的。我们在公司一开始就讨论过它。然后你在雷德蒙德建了一个团队，然后在某个时候你决定这个团队不行。团队搞了好几年进展缓慢，那你为什么没有更早行动，你为什么在那个时候才行动？为什么那是正确的行动时机？
**Dwarkesh Patel:** Well, I don't know. I guess I guess uh the urgency is going to come from whoever's leading the company. So if my sense of urgency, I have like a maniacal sense of urgency. So
**Elon Musk：** 我有这些非常详细的每周工程评审。那可能是一个不寻常的颗粒度水平。我不知道有谁在运营公司——至少是制造公司——能做到我所做的那种细节程度。这并不是说……我对实际情况有很好的理解因为我们会详细地过一遍事情。我很信任越级会议，就是不是让直接汇报给我的人说事情，而是让所有汇报给他们的人在技术评审中发言。而且不能有事先准备。否则你会被"glazed"——粉饰太平，用现在的话说。
**Elon Musk:** that maniacal sense of urgency projects through the rest of the company. Is it because of consequences? They're like if you know Elon said a crazy deadline, but if I don't get it, I know what happens to me. Is it just um you're able to identify bottlenecks and get rid of them so people can move fast? Like how do you how do you think about why your companies are able to move fast? Yeah, I'm constantly addressing the limiting factor. So, um I mean I mean on the deadlines front I I mean I generally actually try to aim for a deadline that that I at least think is at the 50th percentile. So it's it's not it's not like an impossible deadline, but but it's the most aggressive deadline I can think of that could be achieved with 50% probability. Um which means that it'll be late half the time. Um and um but whatever like there is like a law of gases expansion that applies to schedules like whatever given whatever schedule you you like if if you you said we're going to do this something in like 5 years which to me is like infinity time. Um it it will expand to fully available schedule and it will take 5 years. um you know like there's like there there's a physical limit like that like physics will limit how fast you can do certain things like so like scaling up manufacturing there's like there's a rate which you can move the atoms um and and scale manufacturing that's why you can't like instantly make you know a million of something million a year or something uh you've you've got to design manufacturing line you got to bring it up you got to ride the scurve of production um So yeah, I guess like like what can I say that's that's that's actually helpful to people. Um I I think generally um a maniacal sense of urgency is is a is very big deal. Um so um and and you want to have you want to have you want to have a an an aggressive schedule. Um, and then you and you and you want to figure out what the limiting factor is at any point in time and and help the team address that limiting factor.
**Dwarkesh Patel：** 没错。你很 Z 世代啊。怎么防止事先准备？你随机点名？
**Dwarkesh Patel:** Can you maybe talk about the So, Starlink was slowly in the works for many years. Uh, and
**Elon Musk：** 不，我就绕着房间转。每个人都做一个更新。要在脑子里记住大量信息。如果你每周或每两周开一次会，你就有那个人说过什么的快照。你然后可以把进展点画在一条曲线上，心里画一下说，"我们是不是在收敛到一个解？"我只有在得出结论认为不采取重大行动就没有任何成功可能的时候，才会采取重大行动。所以当我最终得出结论，除非采取重大行动否则我们没有任何成功的机会，那我就必须采取重大行动。2018 年我得出了那个结论，采取了重大行动，解决了问题。
**Elon Musk:** yeah, we talked about it all the way in the beginning of the company. Yeah. And so then there was a team you had built in Redmond and then at one point you decided this team is just not cutting it. But again, how did you like it went for a few years slowly and so why did it why didn't you act earlier and why did you act when you did? Like why was that the right moment at which to act?
**Dwarkesh Patel：** 你有很多很多公司。在每一个公司里听起来你都会做这种对相关瓶颈的深入工程理解以便你能跟人做这些评审。你已经能把它扩展到五个、六个、七个公司了。在一个公司内部，你有很多不同的迷你公司。什么决定了这里的上限？因为你有大概 80 个公司……？
**Dwarkesh Patel:** I mean I had I have these very detailed um engineering reviews weekly. Um that that's that that's maybe a very unusual level of granularity. Um, I don't know anyone who runs a company or at least a manufacturing company that that goes into level of detail that that I go into. Um, so so it's it's not it's not as though like I have a pretty good understanding of what's actually going on.
**Elon Musk：** 80 个？
**Elon Musk:** Mhm. Because we we we we go we go through things in detail. Um, and I'm a big believer in skip level meetings where the individuals in instead of having the person that reports to me say things, it's everyone that reports to them um says something um in in the technical review. Um and um and and there can't be um advanced preparation. So otherwise you you're going to get u you know glazed um as I say these days.
**Dwarkesh Patel：** 不。但你已经有很多了。那已经很了不起了。就这个目前的数量来说。没错。我们连一个公司都勉强维持。
**Dwarkesh Patel:** Yeah, exactly. Very gen Z of view.
**Elon Musk：** 取决于情况。我其实没有跟 The Boring Company 有定期会议，所以 The Boring Company 是某种自己在顺其自然地运行。基本上如果什么东西运行良好而且进展不错，那我花时间在上面没有意义。我实际上是根据限制因素在哪里来分配时间。哪里有问题？哪里在碰壁？什么在拖我们后腿？我聚焦于，冒着说太多遍的风险，限制因素。讽刺的是如果什么东西做得很好，他们见不到我太多。但如果什么东西做得不好，他们会经常见到我。或者不完全是做得不好……如果什么东西是限制因素。
**Elon Musk:** Very generous. You just like call them randomly like No, just go around the room and everyone provides an update.
**Dwarkesh Patel：** 限制因素，没错。不完全是做得不好但它是需要更快推进的东西。当某个东西是 SpaceX 或 Tesla 的限制因素时，你是每周还是每天跟负责的工程师沟通？具体是怎么运作的？
**Dwarkesh Patel:** Okay.
**Elon Musk：** 大多数限制因素是每周的，有些是每周两次。AI5 芯片评审是每周两次。每个周二和周六是芯片评审。
**Elon Musk:** Um so, uh I mean it's it's a lot of information to keep in your head because um you you've got a you you've got then say if you have meetings weekly or twice weekly, you you've got a snapshot of what that person said. Um and and and you can and you can then you know plot the progress points. um you can sort of mentally plot the points on a curve and say are we converging to a solution or not um or or are we you know like I I'll take drastic action uh only when I conclude that um success is not in a set of possible outcomes. Um so when I say okay when when I finally reach the conclusion that okay unless drastic action is done we have no no chance of success then I must take drastic action and so that's that's I came to that conclusion in 2018 took drastic action and and fix the problem. How how many um you know you you've got many many companies and in each of them it sounds like you do this kind of deep engineering understanding of what the relevant bottlenecks are so you can do these um reviews with people. Yeah.
**Dwarkesh Patel：** 时间是开放式的吗？
**Dwarkesh Patel:** Um you've been able to scale it up to five, six, seven companies. Within one of these companies you have many different mini companies within them. What determines the maxim? Could you have like 80 companies?
**Elon Musk：** 技术上是，但通常是两三个小时。有时候更短。取决于我们有多少信息要过。
**Elon Musk:** 80? No. But like you can you have so many already like that's that's already remarkable by this current number. Yeah.
**Dwarkesh Patel：** 那是另一件事。我只是想理清这里的区别因为结果看起来非常不同。我觉得了解投入的不同很有意思。感觉在企业界，第一，正如你说的，CEO 做工程评审并不总是发生，尽管那正是公司在做的事情。但时间通常被切得很细，半小时会议甚至 15 分钟会议。你似乎更多地是"我们讨论直到搞清楚为止"这种开放式的。
**Dwarkesh Patel:** Yeah. Exactly.
**Elon Musk：** 有时候。但大多数似乎或多或少准时结束。今天的 Starship 工程评审长了一点因为有更多话题要讨论。他们在试图搞清楚怎么扩展到每年一百万吨以上的入轨量。相当有挑战性。
**Elon Musk:** Uh no. So um we can barely keep one coming together.
**Dwarkesh Patel：** 我能问个问题吗？你说过 Optimus 和 AI 将会在几年内带来两位数的增长率。
**Dwarkesh Patel:** Um near like it it depends on situation. Um so um I I actually don't don't have regular meetings uh with flooring company. So that boring company's sort of cruising along like look if basically if something is working well and making good progress then there's no point in me spending time on it. So I actually uh allocate time according to where where the where the limiting factor or the problem where are things problematic and um or where are we pushing against uh like what what is holding us back you know I I focus risk of saying the words too many times the limiting factor Um, so, so basically if something's go like the irony is if something's going really well, um, they don't see much of me, but if something is going badly, they'll see a lot of me.
**Elon Musk：** 哦，你是说经济？
**Elon Musk:** Something or not even badly, it's it's like if something's a limiting factor, it's a limiting factor. Exactly. It's not exactly going badly, but it's the thing that's it's it's the thing that we need to make go faster to. And so when something's a limiting factor at SpaceX or Tesla, are you like talking weekly and daily with the engineer that's working on it? How how does that actually work?
**Dwarkesh Patel：** 对。我觉得是的。
**Dwarkesh Patel:** Most things that are learning factor are um weekly and some things are twice weekly. So the the AI5 chip review is twice weekly and and so it's every Tuesday and Saturdays
**Elon Musk：** 我觉得那是对的。
**Elon Musk:** um is is the chip review. Is it open-ended in how long it goes? Uh technically yes, but uh usually it's it's like two or three hours.
**Dwarkesh Patel：** 如果经济会增长那么多的话，DOGE 的削减意义何在？
**Dwarkesh Patel:** So I mean sometimes less. It's it depends on on how much information you've got to go through.
**Elon Musk：** 嗯，我觉得浪费和欺诈不是好事。我实际上相当担心……如果没有 AI 和机器人，我们实际上完全完蛋了，因为国债在疯狂堆积。国债利息支出超过了军事预算，军事预算是一万亿美元。所以我们光利息就超过一万亿美元。我挺担心这个的。也许如果我花一些时间，我们可以减缓美国走向破产的速度，给我们足够的时间让 AI 和机器人来帮助解决国债问题。或者不是帮助解决——它是唯一能解决国债的东西。如果没有 AI 和机器人，我们百分之一千会作为一个国家破产和失败。没有其他东西能解决国债问题。我们只需要足够的时间来建造 AI 和机器人，在那之前不要破产。
**Elon Musk:** Yeah. Well, that's another thing. Again, I'm just trying to tease out the the differences uh here cuz the outcomes seem quite different and so I think it's interesting to note what inputs are different and it feels like the corporate world one like you're saying just the CEO doing engineering reviews does not always happen despite the fact that that is the you know what the company is doing. Um, but then time is often pretty finely sliced into, you know, halfhour meetings or even 15-inute meetings. And it seems like you hold more open-ended we're talking about it until we figure it out type meetings. Yeah. Sometimes, but most of them seem to more or less stay on time. Um so um I mean to today's uh Starship engineering review went a bit longer um because there there were more topics to discuss. Um yeah trying to figure out how to scale to a million plus tons to orbit per year is quite challenging. C
**Dwarkesh Patel：** 我想了解的是，当 DOGE 开始的时候，你有巨大的能力来推动改革。
**Dwarkesh Patel:** can I answer a question? So you you said about um Optimus and AI that they're going to result in doubledigit growth rates within a matter of years.
**Elon Musk：** 没那么巨大。
**Elon Musk:** Oh, like the economy. Yeah.
**Dwarkesh Patel：** 好吧。我完全接受你的观点，AI 和机器人推动生产力提升、推动 GDP 增长很重要。但为什么不直接去解决你指出的那些问题，比如某些零部件的关税，或者审批？
**Dwarkesh Patel:** Um
**Elon Musk：** 我不是总统。而且即使是明显的浪费和欺诈，荒谬的浪费和欺诈，要砍掉都非常难。我发现的是，即使是从政府中砍掉非常明显的浪费和欺诈也极其困难，因为政府必须根据谁在投诉来运作。如果你切断了给欺诈者的付款，他们会立刻想出最让人同情的理由来继续付款。他们不会说"请继续欺诈"。他们会说"你在杀害熊猫宝宝。"但实际上没有熊猫宝宝在死。他们编的。欺诈者有能力编造出极其令人信服的、催人泪下的故事，虽然是假的，但听起来很让人同情。这就是发生的事。也许我应该知道得更清楚。但我想，等等，让我们试试从政府里砍掉一些浪费和冗余。也许社保系统里不应该有 2000 万被标记为活着的人实际上已经确定死了，而且超过 115 岁。美国最年长的人是 114 岁。所以可以肯定地说如果有人 115 岁而且在社保数据库里被标记为活着，要么有打字错误……
**Elon Musk:** yes, I think I think that's right.
**Dwarkesh Patel：** 应该有人打电话给他们说，"你的生日好像搞错了，或者我们需要把你标记为已去世。"二选一。
**Dwarkesh Patel:** What was the point of the Doge cuts if the economy is going to grow so much?
**Elon Musk：** 接到这个电话会很吓人。嗯，这似乎是合理的事。比如说如果他们的生日在未来，而且他们有一笔小企业管理局的贷款，而他们的生日是 2165 年，我们要么有打字错误要么有欺诈。所以我们说"我们似乎搞错了你出生的世纪。"
**Elon Musk:** Well, I think like waste and fraud are not good things to have, you know. Um, I I I was actually pretty worried about I I I guess uh I mean I I think in the absence of AI and robotics, we're actually totally screwed. Uh because the national debt is piling up like crazy. Um now our interest payments, the interest payments to national debt exceed the military budget which is a trillion dollars. So over a trillion dollars just in interest payments. Um, you know, that was like I was like, okay, pretty concerned about that. Maybe if I spend some time, we can slow down the bankruptcy of the United States. Um, and give us enough time for the AI and robots to, you know, help solve the national debt or not help solve, it's the only thing that could solve the national debt. Like, we are 1,000% going to go bankrupt as a country and fail as a country without AI and robots. Nothing else will solve the national debt. Um, and so, so we we'd like to well, we just need we need enough time to get build the AI and robots uh to and not go bankrupt before then. I I I guess the thing I'm curious about is when Doge starts, you have this enormous um ability to enact reform and
**Dwarkesh Patel：** 或者一个很好的电影情节。
**Dwarkesh Patel:** not that enormous.
**Elon Musk：** 对。这就是我说的荒谬欺诈。
**Elon Musk:** Sure. Sure. But to totally by your point that like it's important that AI and robotics drive product improvements, drive GDP growth, but why not just directly go after the things you were pointing out, you know, like the the tariffs on certain components or whether it's like permitting. I'm not the president and and very hard to cut to cut to to even even to cut things that are obvious waste and fraud like like ridiculous waste and fraud. Um what I discovered that is it it's extremely difficult even to cut very obvious ways on fraud um from the government um because the the the government has to operate on a on like who's complaining like if if and if you cut off payments to fraudsters they immediately come up with the most sympathetic sounding uh reasons to continue the payment. They they don't say please keep the fraud going. They say, you know, it's they're like, you're killing baby pandas. Like, meanwhile, there's no baby pandas are dying. They're just making it up. Um, the forces are capable of of coming up with extremely compelling sort of heart-wrenching stories that are false, but nonetheless sound uh sympathetic and that that's what happened. Um, and uh so it's like perhaps I should have known better. Um and uh but I thought wait let's take a let's let's let's try to cut some amount of of waste and core from the government. Maybe there shouldn't be you know 20 million people uh marked as alive in social security who are indefinitely dead and over the age of 115. The oldest American is 114. So, it's safe to say if somebody's 115 and marked as live in the social security database, um something is there's either a typo. So, like somebody should call them and say, "We we seem to have your birthday wrong or or uh or or we need to mark you as dead." [laughter]
**Dwarkesh Patel：** 这些人在领取付款吗？
**Dwarkesh Patel:** One of the two things.
**Elon Musk：** 有些人在从社保领钱。但主要的欺诈途径是把某人在社保里标记为活着，然后用其他所有政府支付系统来做欺诈。因为那些其他政府支付系统做的就是对社保数据库做一个"你还活着吗"的查询。这是一个打擦边球的手法（bank shot）。
**Elon Musk:** Very intimidating call to get. Well, so it seems like a reasonable thing. Um and if if like say their birthday is in the future um [laughter] and they have you know a small business administration loan and their birthday is 2165 um we either again have a typo or we have fraud. Um [laughter] so we say we appear to have gotten the century of your birth incorrect
**Dwarkesh Patel：** 你估计通过这个机制的欺诈总额有多少？顺便说一下，政府问责办公室（GAO）以前做过这些估算。不是只有我一个人。
**Dwarkesh Patel:** or a great plot for a movie.
**Elon Musk：** 事实上我记得 GAO 在拜登政府期间做了一个分析，粗略估计拜登政府期间的欺诈大约有五千亿美元。所以不用信我的。看看拜登政府期间发布的报告好了。怎么样？
**Elon Musk:** Yes. This is this this is when I when I mean about ludicrous fraud that's what I meant ludicrous fraud. Were those people getting payments? Some some were getting payments from social security but but but the main fraud vector uh was to mark somebody as alive in social security and then use every other government payment system uh to uh basically to to do fraud because what those other government payment systems do would do they will simply do an are you alive check to the social security database.
**Dwarkesh Patel：** 来自这个社保机制？
**Dwarkesh Patel:** It's a it's a bank shot.
**Elon Musk：** 这只是许多机制之一。重要的是要认识到政府在制止欺诈方面非常无效。不像一个公司，制止欺诈有动力因为它影响公司的收益。政府只是多印钱。你需要关心和能力。这些在联邦层面严重不足。你去车管所（DMV）的时候，你会觉得"哇，这是一个能力的堡垒"吗？好了，现在想象比车管所更糟因为这是一个能印钱的车管所。至少州级的车管所需要……各州多多少少需要在预算范围内否则就破产。但联邦政府只管印钱。
**Elon Musk:** What would you estimate is like the total uh amount of fraud from this mechanism. Um my guess is and and other by the way the the government accountability office has done these estimates before. I'm not the only one who's coming out of this you know in fact I think they they did the GAO did analysis a rough estimate of fraud during the Biden administration and calculated at roughly half a trillion dollars. So don't take my word for it. Take it a report issued during the Biden administration. How about that
**Dwarkesh Patel：** 如果真的有五千亿的欺诈，为什么不可能把那些都砍掉？
**Dwarkesh Patel:** from this social security mechanism?
**Elon Musk：** 你真的必须后退一步，重新校准你对能力的预期。因为你在一个世界里运作，你得收支平衡。你得付你的账单……找到麦克风。
**Elon Musk:** Uh it's it's one of many. It's important to appreciate that the the government does not is very ineffective at at stopping fraud because uh it's it's not like like if it was a company like like stopping fraud, you've got a motivation because it's affecting the earnings of your company. Uh but the government just just they just print more money. Um, so it's not uh like you you need you need caring and competence and these are in short supply at at the federal level. Um, yeah, sorry. I mean, when you go to the DMV, do you think, wow, this is a bastion of competence. [snorts] Um, well, now imagine it's worse than the DMV because it's a DMV that can print money.
**Dwarkesh Patel：** 没错。
**Dwarkesh Patel:** So, was it not possible? At least the state level DMVs uh need to the states more or less need to stay within their budget to go bankrupt. But the federal government just prints more money.
**Elon Musk：** 不像有一个巨大的、基本上不在乎的怪物官僚机构。它就是一堆过时的电脑在发放付款。DOGE 团队做的一件事听起来很简单，但可能每年能省 1000-2000 亿美元。就是简单地要求从主要的财政部电脑发出的付款——那台电脑叫 PAM，Payment Accounts Master 还是什么，每年有 5 万亿美元的付款——在发出时必须有一个付款拨款代码。把它变成强制的，不是可选的，你必须在备注栏里填写任何东西。你得重新校准事情有多蠢。付款在没有拨款代码的情况下就发出去了，没有回查到任何国会拨款，也没有任何解释。这就是为什么战争部——以前叫国防部——无法通过审计，因为信息根本就不在那里。重新校准你的预期。
**Elon Musk:** Was it not possible to cut that if there's a catchy half a trillion of fraud? Why was it not possible to cut all that? Uh because when when as soon as you we did we we actually no you you you you really have to stand back and reccalibrate your expectations for competence. uh because uh you you're operating in a world where you know you you you've got to sort of make ends meet like you know you got to pay your bills, you got to you know
**Dwarkesh Patel：** 我想更好地理解这个五千亿的数字，因为有一份 2024 年的 IG 报告。
**Dwarkesh Patel:** buy the microphones.
**Elon Musk：** 也许吧，但我们发现七年里社保欺诈他们估计大约 700 亿，所以大约每年 100 亿。所以我很好奇其他的 4900 亿是什么。
**Elon Musk:** Yeah. Yeah. Exactly. Um so so you you you if you don't have it's it's not like there's a giant largely uncaring monster bureaucracy. It's not even and and and a bunch of uh monacistic computers that are just that are just sending payments. Um like one of the things that that that the Doge team did there was and it sounds so simple uh that that probably will save um let's say 100 billion maybe 200 billion a year um is simply requiring that payments from the main treasury computer which is called PM it's like payment accounts master or something like that there's 5 trillion payments a year um requiring that any payment go that goes out have a payment um appropriation code, make it mandatory, not optional, and that you have anything at all in the comment field. Um, because you see you have to recalibrate how dumb things are. You think were being sent out with no appropriation code, not not checking back to any congressional appropriation, and no explanation. And this is why the the Department of War, formerly the Department of Defense, cannot pass an audit because the information is literally not there. Recalibrate your expectations.
**Dwarkesh Patel：** 联邦政府支出每年 7.5 万亿。你觉得政府的能力有多强？
**Dwarkesh Patel:** I want to better understand the South Australian number because there was an IG report in 2024.
**Elon Musk：** 可自由支配的支出大概是……15%？但这不重要。大部分欺诈是非可自由支配的。基本上是欺诈性的 Medicare、Medicaid、社保、残障。有无数的政府付款。其中一堆实际上是对各州的整体转移支付。所以联邦政府在很多情况下甚至没有足够的信息来知道有没有欺诈。让我们考虑一下归谬法。政府是完美的，没有欺诈。你估计这个概率是多少？
**Elon Musk:** How you must like why is it so low? Um maybe. But uh which found that like over seven years this the social security fraud they estimated was like 70 billions over 7 years. So like 10 billion a year. So I'd be curious to see what like the other 490 billion is. Federal government expenditures are 7 and a half trillion a year.
**Dwarkesh Patel：** 零。
**Dwarkesh Patel:** Yeah.
**Elon Musk：** 好的，那你会说政府的欺诈和浪费效率是 90% 吗？那也相当慷慨了。但如果只有 90%，那就意味着每年有 7500 亿美元的浪费和欺诈。而且它不是 90%。它不是 90% 有效的。
**Elon Musk:** Um how what what percentage how competent do you think government is? The the discretionary spending there is like 15%.
**Dwarkesh Patel：** 这似乎是一种奇怪的第一性原理方式来估算政府中的欺诈金额。就像，你觉得有多少？不管怎样，我们不必在这里现场做这个，但我很好奇——
**Dwarkesh Patel:** Yeah. But but it doesn't matter. Most of the ford is non-discretionary. It's it's basically a fraudulent Medicare, Medicaid, uh social security, uh uh uh you know um disability. Uh it's there's there's a zillion government payments. Yeah.
**Elon Musk：** 你了解 Stripe 的欺诈情况吗？人们一直在试图做欺诈。
**Elon Musk:** Um and and a bunch of these payments are in fact u they're uh block transfers to the states. So the federal government doesn't even have the information in a lot of cases to even see know if there's fraud. Let's consider let's like reductio Adam sodom the government the government is perfect and has no fraud. What is your probability estimate of that? I mean zero. Okay. So then would you say that foreign waste that the government uh is has is 90%. That also would be quite generous. But if if it's only 90% that means that there's $750 billion a year of waste and fraud and it's not 90%. It's not 90% effective. This seems like a strange way to first principle is the amount of fraud in the government. just like how much do you think there is and then uh uh I anyways we don't know how to do it live but I'd be curious like see you know a lot about fraud at Stripe people are constantly trying to do fraud.
**Dwarkesh Patel：** 对，但正如你说的，这有点……我们确实磨下来了不少，但这是一个有点不同的问题空间，因为你这里面对的欺诈向量比我们要异质得多。但在 Stripe，你有高能力而且你很努力。
**Dwarkesh Patel:** Yeah but as you say it's like a little bit of a um we've really ground it down but it's a little bit of a different problem space because we're dealing with a much more heterogeneous set of fraud vectors here than we are
**Elon Musk：** 你有高能力和高关注度，但欺诈仍然不是零。现在想象它在一个大得多的规模上，能力差得多，关注度也低得多。在 PayPal 当年，我们试着把欺诈管控到支付量的大约 1%。那非常困难。需要巨大的能力和关注度才能仅仅把欺诈控制到 1%。现在想象你是一个关注度低得多、能力也差得多的组织。那会远超 1%。
**Elon Musk:** Yeah. But I mean I mean at tribe you you you you have high confidence and you try hard. Um you have high competence and high caring but still fraud is non non zero. Um now now now imagine it's at a much bigger scale. Um there's much less competence and much less carrying. You know bank PayPal back in the day we we try to manage forward down to about 1% of of the payment volume. Um and that was very difficult. took a tremendous amount of competence and caring to uh get fraud merely to 1%. Um now imagine that that you have an organization where there's much less caring and much less confidence. It's [snorts] going to be much more than 1%.
**Dwarkesh Patel：** 你现在回头看政治和在那里做事情感觉怎么样？从外面看，两件事产生了相当大的影响：一个是 America PAC，另一个是当时收购 Twitter。但似乎也有很多苦恼。你对整个经历打几分？
**Dwarkesh Patel:** How do you feel now looking back on um kind of politics and and doing stuff there where it feels like from the outside in the two you know two things have been quite impactful. one the America Pack and two um the acquisition of of well Twitter at the time but also it seems like there was a bunch of heartache and so what's your what's your grading of the whole experience
**Elon Musk：** 我觉得那些事情需要做，以最大化未来美好的概率。政治通常是非常部落化的。人们在政治上通常失去客观性。他们通常很难看到对方的好处或自己这方的坏处。那大概是最让我惊讶的事情之一。你经常根本无法跟人讲道理。如果他们在一个部落或另一个部落里。他们就是相信他们部落做的一切都是好的，另一个政治部落做的一切都是坏的。说服他们否则几乎不可能。但我觉得总体上那些行动——收购 Twitter、让 Trump 当选，尽管它让很多人生气——我觉得那些行动对文明是有益的。
**Elon Musk:** well um I think I think those things needed to be done to maximize the probability that the future is good um So um but but politics generally is very tribal. Um and and it's it's very tribal. It's and people lose their objectivity usually with politics like they they generally have trouble seeing the good on the other side or the bad in their own side. That's generally how it goes. Um that that I guess was one of the things that surprised me the most is you you often simply cannot reason with people um if they're in one tribe or the other. they they simply believe that everything their tribe does is good and anything the other political tribe does is bad. Um and persuading them is otherwise is almost impossible. Um, so anyway, but um I think I think overall those actions um acquiring Twitter, getting Trump elected even though, you know, it makes a lot of people angry. Um I think those I think those actions are good for were good for civilization.
**Dwarkesh Patel：** 这如何融入你期待的未来？
**Dwarkesh Patel:** Um yeah, how does it feed into the future you're excited about? Well, um, America needs to contain, America needs to be strong enough to last long enough to, um, extend life to other planets and to get, I guess, AI and robotics to the point where we can ensure that the future is good. Um like on the other hand if if if we were to descend into um say communism or or some situation where where the state was extremely oppressive um that that would mean that we we might not be able to become multilanetary um and we might this the state might um you know stamp out um our progress in AI and robotics. How do you feel about um uh you know Optimus, Grock, etc. are going to be leveraged by and not just yours any revenue maximizing company's products will be leveraged by the government over time? Um how does this concern manifest in what private companies should be willing to give governments? What kinds of guard rails should like should you know the should um AI models be uh um made to do whatever the government that has contracted them out to do asked them to do? Um should like should should Grog get to say like actually even the military wants to do X. No, the Grock will not do that. I I pro probably the biggest danger of AI or maybe the biggest danger of fa for for AI and robotics going wrong wrong is is government interesting you know um I mean the the way like people who are opposed to corporations or or or worried about corporations should um really worry about the most about government cuz government is just a corporation in the limit. It's a government. It is it is it is government is just the biggest corporation with a monopoly on violence. Um so I always find it like a strange dichotomy where where people would think corporations are bad but the government is good when the government is simply the biggest and and and worst corporation. But people have that dichotomy. They somehow think at the same time that government can be good but corporations bad. And this is not true. corporations are have better morality than the government.
**Elon Musk：** 嗯，美国需要足够强大，撑足够久，好把生命延伸到其他星球，好让 AI 和机器人达到我们能确保未来美好的程度。另一方面，如果我们滑入比如说共产主义或某种国家极端压迫的状况，那就意味着我们可能无法成为多行星文明。国家可能扼杀我们在 AI 和机器人方面的进步。
**Elon Musk:** It is.
**Dwarkesh Patel：** Optimus、Grok 等等。不只是你的，任何利润最大化公司的产品都会随着时间被政府利用。这种顾虑在私人公司应该愿意给政府什么方面如何体现？什么样的护栏？AI 模型应该为签约它的政府做任何被要求做的事吗？Grok 应该能说"实际上，即使军方想做 X，不，Grok 不会做那个"吗？
**Dwarkesh Patel:** So I I I actually think it's uh you know that's uh that that is a thing to be worried about. It's like if the you know should should if the government should not like the government could potentially use AI and robotics to suppress the population like that is a serious concern.
**Elon Musk：** 我觉得也许 AI 和机器人出问题的最大危险是政府。反对企业或担心企业的人真的应该最担心政府。因为政府在极端情况下就是一个公司。政府只是最大的公司，对暴力有垄断权。我总是觉得一种奇怪的矛盾——人们认为企业是坏的，但政府是好的，而政府只是最大最差的企业。但人们就是有那种矛盾。他们不知怎的同时认为政府可以是好的但企业是坏的，而这不是真的。企业的道德性比政府更好。我确实觉得这是一件值得担忧的事。政府可能潜在地利用 AI 和机器人来压制人民。这是一个严重的顾虑。
**Elon Musk:** I as a guy building AI and robotics how do you how do you like how do you prevent that? Uh well I think like if you have a limited government um if you limit the powers of government which is like really what the US constitution is intended to do is intended to limit the powers of government then then uh you're probably going to have a better outcome than if you have more government. So will be available to all governments right?
**Dwarkesh Patel：** 作为制造 AI 和机器人的人，你怎么防止这个？
**Dwarkesh Patel:** Yeah about all governments. Um I mean it's difficult to predict the like I said like what what's what's the end end point or like what is what is many years in the future but it's difficult to predict the the sort of path along along that way. Um like if civilization progresses AI will vastly exceed the sum of all human intelligence and and there will be far more robots than humans. um along the way what happens it's very difficult to predict.
**Elon Musk：** 如果你限制政府的权力——这真的是美国宪法的本意，限制政府的权力——那你大概会有一个比政府权力更大的情况更好的结果。
**Elon Musk:** I mean I mean it seems like one thing you could do is just say um uh you are not allowed to whatever government you're not allowed to use Optimus to do XYZ just write out like a policy. I mean you you I think you treated recently that Grock should have a moral constitution. Um and one of those things could be that we we limit what governments are allowed to do with this advanced technology. I mean yeah we we can do what is what I mean technically I mean if if the politicians pass a law uh then and they can enforce that law then it's hard to not do that law. You know the the best thing we can have is is is limited government uh where um you know you have you have the appropriate cross checks between the executive judicial and um legislative branches. I I guess the the reason I'm curious about it is this like at some point it seems like the limits will come from you, right? Like you've got the Optimus, you've got the space GPUs, you've got the
**Dwarkesh Patel：** 机器人将对所有政府都可用，对吧？
**Dwarkesh Patel:** you think I will be the boss of the government
**Elon Musk：** 我不知道是不是所有政府。很难预测。我能说终点是什么，或者很多年后是什么样，但很难预测沿途会发生什么。如果文明继续进步，AI 将远超全人类智能的总和。机器人将远多于人类。沿途会发生什么是很难预测的。
**Elon Musk:** or you will get the you will like the I mean already it's the case with SpaceX that for things that are crucial to the um uh like the government really cares about getting certain satellites up in space whatever like it needs SpaceX. Uh it is the it is the um a necessary contractor and you are in the process of building more and more of the um uh the technological components of the future that that that will have an analogous role in different industries and you could have this ability to like set some policy that um you know suppressing classical liberalism in any way. I my companies will not help in in any way with that or you know some policy like that. Um, I I will do my best to ensure that anything that's within my control maximizes the good outcome for humanity. I think anything else would be shortsighted. Um, because obviously I'm part of humanity. So, um, I like humans. Um,
**Dwarkesh Patel：** 似乎你能做的一件事是直接说，"不管什么政府 X，你不能用 Optimus 做 X、Y、Z。"直接写一个政策。
**Dwarkesh Patel:** pro human pro.
**Elon Musk：** 我觉得你最近发推说 Grok 应该有一个道德宪法。其中一条可以是我们限制政府被允许用这种先进技术做什么。
**Elon Musk:** Um, you you you've mentioned that Dojo 3 will be used for space-based compute. Um, [laughter]
**Dwarkesh Patel：** 技术上如果政治家通过了一项法律而且能执行那项法律，那就很难不遵守那项法律。
**Dwarkesh Patel:** Do you really read my uh what I say?
**Elon Musk：** 我们能做到的最好的是有限政府，有行政、司法和立法部门之间适当的相互制衡。
**Elon Musk:** I don't know if you know Twitter, but I know you lot. [laughter] You have a lot of followers. Big giveaway.
**Dwarkesh Patel：** 我好奇这个的原因是在某个时候，限制似乎会来自你。你有 Optimus，你有太空 GPU……
**Dwarkesh Patel:** Um how do you
**Elon Musk：** 你觉得我会成为政府的老板？
**Elon Musk:** how does you have discern my secrets and I post them. How how do you design this chip for space? What like Yeah. What changes?
**Dwarkesh Patel：** 已经是这种情况了——SpaceX 在关键事项上，政府真的很在意让某些卫星上天什么的，它需要 SpaceX。它是必要的承包商。你正在构建越来越多的未来技术组件，它们会在不同行业扮演类似的角色。你可以有这种能力来设定某种政策——压制经典自由主义的任何方式……"我的公司不会以任何方式帮助这件事"，或者某种这样的政策。
**Dwarkesh Patel:** Well, I guess you want to design it to be um more radiation tolerant and run at a higher temperature. Uh so you can um you know roughly if you increase the um operating temperature by 20% in degrees Kelvin you can cut your radiator mass in half. Um so running at a higher temperature is is helpful in in space. Um there I mean there's various things you can do for shielding the memory and but like neural nets are going to be very resilient to bit flips. Yeah. So like most of what what happens from a radiation is like random bit flips. Um but like if you've got like you know a multi- trillion parameter model and you get a few flips doesn't matter. Um it's it's much like curistic programs are going to be much more sensitive to flips than um some giant parameter file. Um, so I just designed it to run hot and um I think you pretty much do it the same way that you do things on Earth apart from make it run hotter.
**Elon Musk：** 我会尽我所能确保在我控制范围内的任何事情都最大化对人类的好的结果。我觉得任何其他做法都是短视的，因为显然我是人类的一部分，所以我喜欢人类。支持人类。
**Elon Musk:** Um, I mean the solar array is most of the weight on the satellite. Is there a way to make the um the GPUs even more powered ends than what Nvidia and TPUs and etc are planning on doing that would you know be especially privileged in the space-based world? Well, I mean the basic math is like um if you can do about a kilowatt per reticle um and then you'd need um you know 100 million full retical chips uh to do 100 gawatt. Yeah. So yeah, depending what your yield assumptions are, you know, um that that tells you how many trips you need to make. Um, but cool. You need if you want if if if you're going to have 100 gigawatts of power, you need, you know, 100 million chips running that that are running a kilowatt sustained uh quad per reticle.
**Dwarkesh Patel：** 你提到 Dojo 3 将用于太空计算。你真的读我说的东西。
**Dwarkesh Patel:** Um
**Elon Musk：** 我不知道你知不知道，Elon，但你有很多粉丝。
**Elon Musk:** basic math 100 million chips. Uh it depends on Yeah. If if you if you look at the die size of something like black wisps or something and how many you can get out of a wafer, you can get like um on the order of dozens or less uh per wafer. So you're basically you're this is a world where
**Dwarkesh Patel：** 一目了然。你是怎么发现我的秘密的？
**Dwarkesh Patel:** if we're putting that out a every single year you're producing millions millions of wafers a month.
**Elon Musk：** 哦，我把它们发在 X 上了。
**Elon Musk:** Um that's the plan with Terapab millions of wafers a month of advanced process notes. Yeah, it's it's got to be some number north of a million. I think you got to do the memory, too.
**Dwarkesh Patel：** 你怎么为太空设计芯片？有什么变化？
**Dwarkesh Patel:** Yeah. Are you going to make a memory f?
**Elon Musk：** 你要把它设计得更耐辐射，运行温度更高。大致来说，如果你把工作温度按开尔文度提高 20%，你可以把散热器质量减半。所以在太空中运行在更高温度是有帮助的。你可以做各种事情来屏蔽内存。但神经网络对位翻转（bit flips）是非常有韧性的。辐射造成的大部分影响就是随机位翻转。但如果你有一个多万亿参数的模型而且你有几个位翻转，那没关系。启发式程序对位翻转会敏感得多，比某个巨大的参数文件敏感得多。我就是把它设计成运行得热。我觉得你基本上跟在地球上做法一样，除了让它运行得更热。太阳能阵列是卫星上大部分的重量。
**Elon Musk:** I think the teraf's got to do memory. It's got to do logic, memory, and packaging. I'm very curious how somebody like gets started. This is like the most complicated thing man has ever made. And obviously, like if anybody's up to the task, you're up to the task. Like what do you So, you realize it's a bottleneck and you go to your engineers and like what is the next like what what do you tell them to do? [laughter] I want a million papers a month in 2030. What is the next like what do you
**Dwarkesh Patel：** 有没有办法让 GPU 比 Nvidia、TPU 等等计划的更强大，在太空环境中特别有优势？
**Dwarkesh Patel:** That's right.
**Elon Musk：** 基本的数学是，如果你每个 reticle 能做大约一千瓦，那你需要一亿个全 reticle 芯片来做 100 吉瓦。取决于你的良率假设，那告诉你需要造多少芯片。如果你要有 100 吉瓦的功率，你需要一亿个芯片以每个 reticle 持续一千瓦运行。基本数学。一亿个芯片取决于……如果你看 Blackwell GPU 之类的 die 尺寸，以及每片晶圆能切出多少个，每片晶圆大约能出几十个或更少。所以基本上，在我们每年都发射这么多的世界里，你每月要生产几百万片晶圆。
**Elon Musk:** Do you like call ASML? Like what is Ask what I want. [laughter]
**Dwarkesh Patel：** 这就是 TeraFab 的计划？每月几百万片先进制程节点的晶圆？
**Dwarkesh Patel:** What is the next step?
**Elon Musk：** 对，可能超过一百万什么的。你还得做内存。
**Elon Musk:** That's so much to ask. Well, um we make a little fab uh and see what happens. Uh make our mistakes at a small scale and then make a big one.
**Dwarkesh Patel：** 你要建一个内存工厂？
**Dwarkesh Patel:** Is a little fab done or is it
**Elon Musk：** 我觉得 TeraFab 得做内存。它得做逻辑、内存和封装。
**Elon Musk:** No, it's not done. Which I mean people would not keep that cat in the bag. [laughter] that cat's going to come out of the back. It'll be like drones hovering over the bloody thing, you know. You'll be able to like see it construction progress on X, right? You know, in real time. Um, so no, we we I mean, listen, I don't know. We could just flounder in failure to be clear. It's like not uh success is not guaranteed, but um since we want to try to make uh you know something like a 100 million Yeah, we we need we want 100 gigs of power and 100 100 that trips that can take 100 gawatt, right? So call it, you know, but yeah, by by 2030. So then um we'll take as many chips as our suppliers will give us. I've said this to I've actually said this to TSMC and Samsung and Micron. It's like please build your more fabs faster. Um, and we will guarantee to buy the output of those fabs. Um, so so that they're already like moving as fast as they as they can. Like it's it's not like to be clear, it's not like us take, you know, it's not like u either it's it's not like it's us plus them. You know, there's an narrative that the people doing AI want a very large number of, you know, chips as quickly as possible. And then many of the input suppliers, the fabs, but also, you know, the turbine manufacturers are not ramping up production very quickly. And the explan Yeah. And the the explanation you hear is that they're dispositionally conservative. You know, they're Taiwanese or German as the, you know, story may be. And they just like don't believe the like is that really the explanation or is there something else? Well, I mean, it's reasonable to like if somebody's been in say the computer memory business for uh 30 or 40 years
**Dwarkesh Patel：** 我很好奇一个人怎么开始。这是人类有史以来制造的最复杂的东西。显然如果有谁能胜任，那就是你。那你意识到这是一个瓶颈，你去找你的工程师。你跟他们说什么？
**Dwarkesh Patel:** and they've seen cycles,
**Elon Musk：** "我要在 2030 年之前每月一百万片晶圆。"没错。那正是我想要的。
**Elon Musk:** they've seen like boom and bust like 10 times. Yeah.
**Dwarkesh Patel：** 你打电话给 ASML 吗？下一步是什么？
**Dwarkesh Patel:** You know, so so like that's a lot of layers of scar tissue, you know? So, it's like it's like during the boom times looks like everything's going to be great forever and then then then the crash happens and then they're desperately trying to avoid bankruptcy. Um and and then there's another boom and another crash.
**Elon Musk：** 不，没那么多好问的。我们先做一个小的晶圆厂看看会怎样。在小规模上犯错然后再做大的。
**Elon Musk:** Are there other [laughter] are there other ideas you think others should go pursue that you're not for whatever reasons right now? Um [sighs and gasps] I mean there are a few companies that are that are pursuing like uh new ways of doing jobs.
**Dwarkesh Patel：** 小晶圆厂做好了吗？
**Dwarkesh Patel:** Um uh but they're just not scaling fast.
**Elon Musk：** 不，还没好。我们藏不住那只猫的。那只猫会跑出袋子。会有无人机盯着那个该死的东西。你能在 X 上实时看到它的建设进度。听着，我不知道，我们可能就扑腾着失败了，坦白说。成功不是保证的。因为我们想尝试做大约一亿片……我们想要 100 吉瓦的电力和能承受 100 吉瓦的芯片，在 2030 年之前。我们会接受供应商能给我们的所有芯片。我其实已经跟 TSMC 和 Samsung 和 Micron 说了："请更快地建更多的晶圆厂。"我们保证买下那些晶圆厂的产出。所以他们已经在尽可能快了。是我们加他们。
**Elon Musk:** I I don't even mean within AI. I mean just generally I'd say like people should just should do the thing that where they find that they're highly motivated to do that thing. Mhm.
**Dwarkesh Patel：** 有一种说法是，做 AI 的人想要尽可能快地拿到大量芯片。然后很多上游供应商，晶圆厂，还有涡轮机制造商，并没有很快地扩张产能。
**Dwarkesh Patel:** As opposed to, you know, something something some idea that that I suggest, but they should do the thing that they find personally interesting and motivating to do.
**Elon Musk：** 不，他们没有。
**Elon Musk:** Mhm. Um but but you know, going back to the limiting factor, use that phrase about 100 times. Um the the current limiting factor that I see in the time frame you know in the sort of 20 29 20 like in in the in the three 3 to 4 year time frame um it's chips. Um in in the one-year time frame it's it's energy power production electricity. like it's it's not clear to me that there's enough um usable electricity to turn on all the the AI chips that are being made. Um towards the end of this year, I think people are going to have real trouble turning on like the chip output will exceed the the ability to turn chips on.
**Dwarkesh Patel：** 你听到的解释是他们天性保守。他们是台湾人或德国人，看你说的是谁。他们就是不相信……这真的是解释吗还是有别的原因？
**Dwarkesh Patel:** What's your plan to deal with that world?
**Elon Musk：** 嗯，这是合理的……如果一个人在电脑内存行业干了 30 或 40 年……他们见过周期。他们见过繁荣和萧条 10 次。那是很多层伤疤组织了。在繁荣期，看起来一切都会永远美好。然后崩盘来了，他们拼命想避免破产。然后又是繁荣又是崩盘。
**Elon Musk:** Well, we're trying to accelerate electricity production. Um, I I guess that's that's maybe one of the reasons that um XAI will be maybe the leader, hopefully the leader, um, is that we'll be able to turn on more chips than other people can turn on faster. Um, because we're we're we're good at hardware [snorts] and and and and generally the the innovations from the corporations that me that call themselves labs. um the the ideas tend to flow like it's it's rare to see that there's like more than about a six-month difference um between um like the ideas uh travel back and forth um with the people. So, so I think you you sort of hit the hardware wall and um and then whatever whichever company can scale hardware the fastest will be the leader and so I think XCI will be able to scale hardware the fastest and therefore most likely will be the leader. you you you joked or you know um were self-conscious about uh you know using the uh the limiting factor phrase again but I actually think there's something deep here and if you look at a lot of the things we've touched on over the course of it maybe kind of a good note to end on like if you think of a scinesscent lower agency company it would have some bottleneck and not really be doing anything about it. Um, you know, Mark and Dre had the line of, uh, most people are willing to endure any amount of chronic pain to avoid acute pain. Uh, and it feels like a lot of the cases we're talking about are just leaning into the acute pain, whatever it is. It's like, okay, we got to figure out how to, you know, work with steel or we got to figure out how to run the chips in space or like we'll take some near-term acute pain to actually solve the bottleneck. And so that's kind of a unifying theme. I have a high pain threshold. That's helpful.
**Dwarkesh Patel：** 有没有其他你觉得别人应该去追求但你因为某种原因目前没有做的想法？
**Dwarkesh Patel:** Solve the bottlenecks.
**Elon Musk：** 有一些公司在追求芯片的新方式，但它们就是规模化不够快。我甚至不是说在 AI 领域内，我是说总体上。人们应该去做他们发现自己高度有动力去做的事情，而不是某个我建议的想法。他们应该去做他们个人觉得有趣和有动力去做的事。但回到限制因素……我用这个词大概用了 100 次了。我在三到四年时间框架内看到的当前限制因素是芯片。在一年时间框架内是能源、发电、电力。我不确定有足够的可用电力来开启正在制造的所有 AI 芯片。到今年年底，我觉得人们会真的遇到困难……芯片产出将超过开启芯片的能力。
**Elon Musk:** Yes. Um so you know one thing I I can say is like uh I think the future's going to be very interesting. Um and um and as I said the Davos only been I think it was on the ground for like 3 hours or something. Um it's better to be it's better to on the side of optimism and be wrong than on the side of pessimism and be right uh for quality of life. So you know you your your your happiness will be you you'll be happier if you if you are on the side of optimism rather than ingring on the side of pessimism
**Dwarkesh Patel：** 你打算怎么应对那个世界？
**Dwarkesh Patel:** and so I recommend ingring on the side of optimism.
**Elon Musk：** 我们在试图加速电力生产。我想这也许是 xAI 可能会成为领导者的原因之一，希望是领导者。我们能比其他人更快地开启更多芯片，因为我们擅长硬件。总的来说，那些自称实验室的公司的创新，想法倾向于流动……很少见到超过大约六个月的差距。想法随着人一起流动。所以我觉得你在某种程度上撞上了硬件的墙，然后哪家公司能最快地扩展硬件就会成为领导者。所以我觉得 xAI 能最快地扩展硬件，因此最有可能成为领导者。
**Elon Musk:** Let's do that. Cool. Yan, thanks for doing this.
**Dwarkesh Patel：** 你开玩笑或者不好意思又用了"限制因素"这个词。但我其实觉得这里有一些深层的东西。如果你看我们在整个对话中触及的很多事情，这也许是一个很好的结束点。如果你想想一个衰老的、低行动力的公司，它会有某个瓶颈但实际上什么都不做。Marc Andreessen 有一句话，"大多数人宁愿忍受任何程度的慢性痛苦也不愿面对急性痛苦。"感觉我们讨论的很多案例就是迎面去承受急性痛苦，不管它是什么。"好了，我们得搞清楚怎么用钢，或者我们得搞清楚怎么在太空运行芯片。"我们承受一些短期的急性痛苦来真正解决瓶颈。所以这是一种统一的主题。
**Dwarkesh Patel:** Thank you.
**Elon Musk：** 我的痛苦阈值很高。这很有帮助。去解决瓶颈。对。我能说的一件事是，我觉得未来会非常有趣。正如我在达沃斯说的——我觉得我在地面上待了大概三个小时——偏向乐观然后错了比偏向悲观然后对了要好，从生活质量来说。如果你偏向乐观，你会比偏向悲观更快乐。所以我推荐偏向乐观。
**Elon Musk:** All right. place. All right.
**Dwarkesh Patel：** 干杯。好的。Elon，谢谢你做这个。
**Dwarkesh Patel:** Oh, great stamina.
**Elon Musk：** 谢谢。好的，谢谢大家。好的。好耐力啊。希望这在痛苦阈值里不算太痛。
**Elon Musk:** Hopefully this encounter is a pain in the tolerance. [laughter] Hey everybody, I hope you enjoyed that episode. If you did, the most helpful thing you can do is just share it with other people who you think might enjoy it. It's also helpful if you leave a rating or a comment on whatever platform you're listening on. If you're interested in sponsoring the podcast, you [music] can reach out at dwarcash.com/advertise. Otherwise, I'll see you on the next one.
