George Hotz: Hacking the Simulation & Learning to Drive with Neural Nets | Lex Fridman Podcast #132

George Hotz：破解模拟世界与利用神经网络学习驾驶 | Lex Fridman 播客 #132 George Hotz 阐述了 Comma.ai 的使命，即提供面向消费者的端到端半自动驾驶软硬件，并将 openpilot 的开放生态系统与特斯拉 Autopilot 的封闭 iOS 模式类比为 Android 与 iOS 的竞争。他剖析了自主系统的工程权衡，详细说明了为何驾驶员监控摄像头以及类似 MuZero 的行为模仿模型将超越手工制定的规则集和 Waymo 资本密集型的自动驾驶出租车路线。在长达三小时的对话中，Hotz 还探讨了基础编程范式、智能合约、模拟理论，以及让他确信人类认知终将被淘汰的智力里程碑。 Comma.ai 仅凭 810 万美元的融资就已实现消费级硬件公司的盈利，其 Comma Two 套件定价为 1,200 美元（主机 1,000 美元，线束 200 美元），在已售出的约 3,000 台设备中保持着超过 70% 的日活跃使用率。 (105:13) George Hotz 预测端到端深度学习是实现 L5 级自动驾驶的唯一可行途径，并将 DeepMind 的 MuZero 视为开创性架构，因为它是在内部学习环境动力学，而不是依赖脆弱的启发式处理流水线。 (70:16) Hotz 坚持一项 10,000 美元的赌注，即埃隆·马斯克将在实现 L5 级自动驾驶之前部署基于摄像头的活跃驾驶员监控系统；他认为，与转向控制要求的零容错可靠性相比，可靠的驾驶员监控只需要约 92% 的准确率。 (76:09) Hotz 认为像 Waymo 这样的纯 L4 级自动驾驶出租车模式是不盈利的、类似于共享电动滑板车的“逐底竞争”，并指出自动驾驶出租车在商业上很难与人类网约车司机竞争，因为人类司机在拥堵路况下的通行速度大约要快 20%。 (114:28)

In conversation with Lex Fridman, Comma.ai founder George Hotz argues that real-world autonomy and artificial intelligence require end-to-end learning grounded in shippable hardware rather than modular heuristics, corporate hype, or unfalsifiable theories. End-to-End Deep Learning Over Heuristic Driving (59:24) Hotz contends that Level 5 autonomous driving will only be solved through end-to-end machine learning rather than modular, hand-engineered perception pipelines. Drawing an analogy to computer chess, Hotz compares modular autonomous architectures—such as Tesla's multi-task task engineering—to hand-tuned engines like Stockfish, whereas end-to-end learning mirrors AlphaZero. He asserts that across the history of artificial intelligence, feature engineering has systematically lost to end-to-end architectures. To resolve the need for real-world simulation, Hotz points to DeepMind's MuZero algorithm, which learns environmental dynamics models internally to perform rollouts. Hotz identifies MuZero as the conceptual blueprint for autonomous vehicles and forecasts that neural networks will fully solve autonomous driving in the latter half of the 2020s. In the interim, Comma.ai trains driving policies via disengagements collected from real human driving supervision. Comma.ai Strategy, Competitor Flaws, and Silicon Monopolies (105:13) Comma.ai positions its openpilot ecosystem as an open 'Android' counterpart to Tesla's vertically integrated 'iOS.' Hotz notes that Comma.ai raised $8.1 million and operates profitably, selling its flagship Comma Two device for $1,000 alongside a $200 vehicle harness. By late 2020, the company had sold between 2,500 and 3,000 units, maintaining roughly 2,000 daily active users and an active driver engagement rate exceeding 70%. Hotz criticizes competing industry paths, arguing that Waymo's Level 4 robotaxi model is commercially unviable because human ride-hail drivers will consistently outperform autonomous cabs on transit times and unit economics. Furthermore, Hotz highlights a hardware bottleneck in AI development driven by an Nvidia and Google duopoly. He criticizes Nvidia for enterprise markups—such as pricing the A100 at roughly $10,000 despite architectural similarities to the $699 consumer RTX 3080—and notes that Google restricts Cloud TPUs behind anti-competitive terms of service, validating Tesla's move to build its own Dojo silicon. Driver Monitoring as an Essential Intermediate Step (76:09) While Fridman advocates for active driver-monitoring systems (DMS) based on human-machine interaction research, Hotz treats DMS as a pragmatic necessity for Level 2 systems rather than an insurmountable AI hurdle. Hotz reiterates a standing $10,000 bet that Elon Musk will ship active driver monitoring before achieving Level 5 autonomy, contrasting Comma Two’s dedicated infrared cabin sensors with Tesla's reliance on visible-light interior cameras. Unlike vehicle path planning, where ground truth can be derived automatically from human steering behavior, driver attentiveness relies on subjective evaluation. However, Hotz argues that driver monitoring does not demand four-nines reliability; achieving roughly 92% reliability suffices because its operational objective is simply keeping the human attentive in an interactive loop while the human retains ultimate legal liability. Pragmatic Engineering, AI Scaling, and Reality Exploits (27:27) Hotz rejects theoretical models that offer no testable leverage over the physical world, dismissing both unfalsifiable simulation hypotheses and String Theory. To take simulation claims seriously, he looks for computational exploits in physics, such as read leaks or arbitrary state modifications. In foundational AI, Hotz credits Marcus Hutter's AIXI and the Hutter Prize for demonstrating that intelligence is formalizable as lossless data compression. Regarding modern language models, Hotz argues that architectures like GPT-3 are overhyped and incapable of scaling into general intelligence. He points out that minimizing cross-entropy loss over text lacks the agentic structure and long-term memory needed for complex execution. Furthermore, Hotz emphasizes that driving states are largely Markovian—where the immediate prior step contains almost all necessary predictive information—unlike the deep contextual dependencies of human language. Hotz maintains that technological survival requires radical honesty and superior engineering, predicting that marketing fabrications will inevitably collapse. Unresolved questions remain regarding whether end-to-end architectures can fully handle long-tail driving edge cases, how consumer Level 2 systems will cross the regulatory liability chasm into true autonomy, and whether open-source hardware ecosystems can survive deepening silicon monopolies.

黑客及 Comma.ai 创始人 George Hotz（geohot）与 Lex Fridman 展开深入对话，主张完全自动驾驶必须依赖端到端神经网络与学习型动态模型，而非过度工程化的人工规则。Hotz 强调高精度的商业落地应立足于实用硬件、激进的真实度与真实代码实现，反对空谈理论与不可证伪的虚构系统。 端到端自动驾驶与 MuZero 范式 (59:24) Hotz 认为实现 L5 级完全自动驾驶的唯一可行路径是端到端（end-to-end）机器学习，而非特斯拉等团队采用的多任务特征切片与人工分解模式。他以国际象棋作比，指出手工分解规则的工程思维（如分别为“象”或“车”雇用专门工程师）在人工智能历史上始终会被 AlphaZero 这类端到端系统击败。针对模拟环境的真实性难题，Hotz 将 DeepMind 的 MuZero 算法视为解决自动驾驶的关键里程碑，因为 MuZero 能够自主学习环境的动态模型，并在该内部模型中展开前向预测与搜索。 关于落地时间线与责任归属，Hotz 预计神经网络将在 2020 年代后期彻底攻克自动驾驶难题。在此之前，Comma.ai 严格专注于渐进式交付具备商业价值的 L2 辅助驾驶产品。Hotz 直言他无意承担法律责任，而 L2 与更高级别自动驾驶的核心分歧仅在于谁承担事故责任。通过让用户承担责任，企业能够专注于通过 OTA 更新实现渐进式迭代，并通过系统介入收集宝贵的人类接管数据，将强化学习直接应用于现实世界。 硬件博弈、驾驶员监控与商业路径 (76:09) Comma.ai 推出的 Comma Two 定价 1000 美元（加上适配器共 1200 美元），安装简易如宜家家具，已支持超过 90 种车型，拥有超 70% 的高日活率并保持盈利。Hotz 将其开放生态与特斯拉的垂直整合相比，类比为 Android 与苹果 iOS 的关系：特斯拉通过封闭的自研 FSD 芯片、Dojo 超算和整车制造赚取极高单车利润，而 Comma.ai 则以开放软件模式拓展更广泛的设备基数。Hotz 批评英伟达高价收割机器学习生态（如将 A100 标价约 10000 美元），并指出谷歌 TPU 绑定云服务且协议严苛，迫使特斯拉等企业不得不走向自研硬件路线。 在车内安全层面，Hotz 设立了一项 10,000 美元的公开赌注，断言马斯克在交付 L5 之前必将先推出驾驶员监控系统。Fridman 与 Hotz 一致认为特斯拉最终会屈服于现实妥协。Hotz 强调，与极其复杂的外部路况感知不同，驾驶员监控属于特征工程能够有效解决的低门槛任务；该系统即使仅达到 92% 的可靠性，也足以与驾驶员形成有效的交互闭环，因而不需要外部驾驶所需的高精可靠性指标。 商业模式的硬核真伪：否定 Waymo 与警惕吹嘘 针对 Waymo 在亚利桑那州推行的 L4 Robotaxi 业务，Hotz 从商业逻辑上直斥其为根本性缺陷的产品概念，预言其市场终局可能类似于共享滑板车。他指出，自动驾驶出租车在遇到行人或复杂路口时表现保守，在通行效率上必然落后于真实的人类司机；消费者在日常出行中高度看重时间效率，仅有少数人愿意为了折扣牺牲时间。面对数十亿美元的沉没成本，Hotz 戏称若让自己担任 Waymo 的 CEO，唯一破局方案是“把 Anthony Levandowski 从监狱里弄出来让他接管”。 Hotz 强调，企业的长期立足之本在于过硬的技术与彻底的诚实。对于初创团队，他坚决反对硅谷常见的“弄假直到成真”（fake it till you make it）套路，认为谎言终将失败，唯有更好的技术才能胜出。此外，他将网络社区（如 Hacker News）上充斥的冷嘲热讽归结为平庸者的心理防御机制——外界往往因为自身无法达成突破，而倾向于唱衰那些雄心勃勃的创新者。 工程哲学：可证伪性、经验主义与学习之法 在对待物理学前沿理论或模拟假设时，Hotz 表现出彻底的实用主义倾向：他不关心缺乏可证伪性的理论美感，只在乎理论是否能赋予人类掌控自然的力量（例如寻找物理学中的“内存泄漏”或实现超高频处理器）。对于当前火热的 GPT-3 等 Transformer 大语言模型，Hotz 直言其被过度神化；基于字符交叉熵损失训练的模型缺乏长期记忆与行动能力，不能直接缩放为通用人工智能，且驾驶任务具备强烈的马尔可夫性质，与自然语言的远距离依赖结构迥然不同。 在工程实践上，Hotz 摒弃形式化的自学指南，主张“学习编程的唯一途径就是动手去写”。他将计算机体系划分为必须掌握的基础阶梯（从汇编、C 语言到 Python），使开发者理解硬件寄存器与抽象层运作机制。Fridman 则补充道，高水平工程能力来自两项互补技能的交替：独自面对白纸从第一性原理深思，以及通过网络与极客社群高效检索前人经验；而 Linus Torvalds 的名言“少说废话，放码过来”（Talk is cheap, show me the code）正是对这种工程实证主义的最佳注解。 本期对话展现了 George Hotz 极具个人风格的硬核黑客哲学：拒绝不可证伪的抽象概念与脱离实际的商业吹捧，主张用最精炼的代码和最直接的实践解决工程难题。无论是在自动驾驶领域押注端到端强化学习，还是在商业运营中践行小步快跑的硬件交付，其实质都是以绝对的真实面对世界，通过不断交付实体系统来兑现技术力量。

## Background 本期视频由著名计算机科学家与播客主持人 Lex Fridman 于 2020 年 10 月 22 日发布在 YouTube 频道上。节目的核心嘉宾是传奇黑客、Comma.ai 创始人 George Hotz（geohot），访谈围绕破解模拟世界、通过神经网络学习驾驶、自动驾驶路线之争以及人工智能与编程哲学展开深入探讨。 ## TL;DR 在这场长达三个多小时的对谈中，George Hotz 详尽拆解了 Comma.ai 的商业模式、开源生态以及与特斯拉（Tesla）和 Waymo 的路线分歧，坚称 Openpilot 将成为自动驾驶领域的开放式 Android。他深入剖析了端到端模型、3D 视觉感知、硬件与芯片垄断格局，并激烈抨击了传统车企与官僚资本的虚伪。此外，Hotz 分享了重塑其认知的三大思想支柱、四大核心编程范式，并结合对 AGI、数据压缩、药物体验与神经多样性的反思，阐明了他以绝对真实、技术唯尊和高效生产力为核心的人生哲学。 ## Atmosphere & style 整场对话洋溢着高密度的极客智力激荡与不羁的赛博朋克黑客风骨。George Hotz 语速飞快、思维跳跃且言辞极度犀利坦率，毫不掩饰对行业主流与平庸建制的鄙夷，而 Lex Fridman 则以沉稳细腻的哲思与耐心追问适时托底平衡。听众既能感受到纯粹技术推演带来的思维过山车快感，又能体会到一位技术天才在面对自然规律时的敬畏，以及践行“绝不说谎、只写代码”这一信条时的狂热与真实。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=0s) 播客介绍：Comma.ai、驾驶员监控与现实世界测试 Lex Fridman 在开场中介绍了本期嘉宾 George Hotz（网名 geohot），这是 Hotz 第二次做客该播客。Fridman 介绍了 Hotz 创立的自动驾驶与半自动驾驶技术公司 Comma.ai，并将其核心愿景与移动操作系统的竞争格局进行类比：正如 Android 之于 iOS，Comma.ai 旗下的开源系统 openpilot 致力于成为汽车领域的 Android，而 Tesla Autopilot 则扮演封闭生态的 iOS。Comma.ai 推出的 Comma Two 硬件设备售价 1,000 美元，安装在许多兼容车型上后，即使在没有车道标线的路况下也能实现车道居中保持。 设备中还集成了基于摄像头的驾驶员状态监控系统（Driver Sensing），以确保驾驶员的视线始终保持在前方道路上。Fridman 强调自己一直是驾驶员监控系统的坚定拥护者，并指出无论是 Tesla Autopilot 还是其他辅助驾驶方案，都理应将该技术纳入标准传感器套件中。 Fridman 结合自己在 MIT 从事自动驾驶与半自动驾驶车辆的研究背景指出，驾驶涉及人机交互的复杂维度，其广度与深度远超实验室内部环境能够模拟的范畴。来自 Comma.ai、Tesla Autopilot 以及凯迪拉克 Super Cruise 等系统在真实世界收集的行驶数据，为理解如何在现实环境下为普通人设计安全系统提供了关键支撑，这也要求工程师必须回归驾驶任务的第一性原理进行大胆探索与创新。 --- ### [02:31](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=151s) 费米悖论与人类文明的毁灭风险 进入正题后，Fridman 提起上一次谈话中关于“模拟假说（Simulation Theory）”的讨论，并抛出一个根本问题：宇宙中是否存在其他智慧生命？Hotz 表达了自己对费米悖论的直观判断——宇宙中曾经存在过大量智慧生命，但那些文明极有可能在演化过程中自我毁灭了。这直接映射了德雷克方程（Drake Equation）中关于文明平均寿命（$L$）的参数：高度发展的文明往往难以维系太长时间。 在谈到人类自身是否已接近自我毁灭时，Hotz 引用了 Eliezer Yudkowsky 的论断：“毁灭世界所需的智商每年下降一点”（“The IQ required to destroy the world falls by one point every year.”）。技术的发展正在不可逆转地使毁灭性力量走向大众化和平民化。随后两人探讨了文化基因与观念（Memes）对现实的破坏力。Fridman 打趣问道“Meme 什么时候会毁灭世界”，Hotz 直言其实这种影响在当下已经悄然发生，且最恶劣的情况远未到来，整个世界正变得愈发怪诞。Fridman 随即调侃道，或许也存在拯救世界的 Meme，例如代表光明力量的“Meme 之王”Elon Musk，对抗如同达斯·维达（Darth Vader）般的暗黑力量代表 Donald Trump。 针对那种“英雄击败恶龙后永远幸福地生活在一起”的童话结局，Hotz 提出了现实维度的冷峻反思： > “I think in every fairy tale they always end it with 'and they lived happily ever after', and I'm like, please tell me more about this happily ever after; I've heard 50% of marriages end in divorce, why doesn't your marriage end up there? You can't just say happily ever after. So it's the thing about destruction is it's over after the destruction. We have to do everything right in order to avoid it, and one thing wrong...” —— George Hotz Hotz 特别指出了密码学（Cryptography）与核武器在攻防博弈上的本质不对称性： - **密码学是防御占优的体系**：在这个领域中防御方如同巨人歌利亚，而攻击者如同大卫，构建坚不可摧的数学防线比攻破它更容易。 - **核武器是进攻绝对占优的体系**：制造一枚能够分裂出 100 枚分导核弹头的洲际导弹，在技术上远比拦截并摧毁 100 枚分导核弹头的反导系统简单得多；攻击只需要计算轨道轨迹发射即可，这使得全球防御体系极易崩溃。 ```mermaid flowchart TD accTitle: 密码学与核武器的攻防不对称对比 accDescr: 对比密码学以防御占优和核武器以进攻占优的博弈结构特征 subgraph CryptoArea ["密码学体系 (Defense-Dominant)"] direction TB C_Def["防御方 (数学加密体系)"] -->|"天然具备结构优势"| C_Adv["防守成本远低于破解成本"] C_Atk["攻击者 (黑客 / 试探者)"] -.->|"难以攻破算法屏障"| C_Def end subgraph NuclearArea ["核武战略体系 (Offense-Dominant)"] direction TB N_Atk["进攻方 (多弹头弹道导弹)"] -->|"轨道投放技术门槛低"| N_Adv["单次突防即可造成绝对毁灭"] N_Def["防御方 (拦截与反导网络)"] -.->|"拦截百枚弹头极度复杂"| N_Atk end ``` --- ### [05:18](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=318s) 外星生命假说：冯·诺依曼探测器与废热难题 如果地外文明没有走向物理毁灭，Hotz 认为它们可能陷入了另一种终极稳态——“脑部插管/电极刺激（Wireheading）”。也就是说，文明成员集体选择通过刺激自身大脑的快乐中枢，沉溺于人工制造的无尽极乐中，从而停止了对现实宇宙的向外扩张。 Hotz 坚信宇宙中目前没有高阶文明在进行大尺度活动，核心证据依旧是费米悖论所提出的核心疑问：“他们都在哪儿？”银河系的直径仅约 70,000 光年（Hotz 提到该数值可能源自《星际迷航：航海家号》），对于一个拥有充足时间的文明而言，这个尺度其实极为微小： - 甚至不需要依赖虫洞等超光速假设。 - 只需部署具有自我复制能力的冯·诺依曼探测器（Von Neumann probe），配合亚光速（Sublight travel）航行，在短短数百万年内就能将探测器散布至整个星系乃至宇宙每一个角落。 - 现实中既然没有任何探测器造访或接管星系的痕迹，说明扩张过程必然在某个节点遭遇了阻断。 Fridman 提出了另一种可能：高等智慧生命可能并不遵循人类的物理感知范式。如果采用 Stephen Wolfram 的计算宇宙视角，将宇宙的本质视作信息处理过程与熵的流动，外星文明或许以纯计算的形式穿行于现实，甚至地球上未解的自然奥秘本身就是它们留下的智能脚印。 Hotz 对这类推测持怀疑态度。他认为物理学在银河系各处应当具备普适性，文明只要向外扩张、进行物质改造，就无法规避热力学第二定律： - 扩张必定伴随庞大的能量消耗。 - 无论技术多么发达，“废热（Waste Heat）”排放是不可逾越的物理限制；然而人类当前的观测仪器在太空中从未探测到此类异常的大规模能量废热聚集。 为了反驳“人类感知过于低等因而无法察觉高等文明”的观点，Hotz 提出了一个百万年跨度的时间思想实验：假如让一百万年前生活在荒漠中、刚刚学会用火的原初人类，去观测现代人类文明，答案是显而易见的——如果一架波音 747 喷气式客机从他们头顶掠过，原始人即便不懂其工作原理，也必然能确凿无疑地感知到某种超常造物的存在。 Fridman 补充指出，这个实验的前提是 747 客机飞过原始人的头顶；如果客机飞在火星上，原始人受制于感知工具依然一无所知。人类向宇宙广播无线电波的历史至今不过 150 年左右，信号仅仅扩散到了极狭窄的光锥（Light cone）范围内。就在两人继续深入外星文明探测的话题时，Fridman 顺势提到了五角大楼解密公布的著名“Tic Tac”UFO 视频，以及他此前与美军战斗机飞行员 David Fravor 的深度访谈。 ### [10:19](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=619s) UFO 目击事件、心理战与 Bob Lazar Lex Fridman 提到了《纽约时报》关于 UFO 目击事件的报道。他指出，历史上其实有数十万起不明飞行物目击记录，其中绝大多数可以通过自然物理现象或人类心理的“意愿投射”（想要相信超自然现象）来解释。然而，像美军海军飞行员 David Fravor 这样训练有素的专业观察员所目击的“井字糖”（Tic Tac）不明飞行物，其飞行轨迹异常诡异。Lex 罗列了可能的解释：美军未公开的绝密技术、先进的无人机或光学欺骗系统、中俄的尖端武器，亦或是来自外星文明。 George Hotz 对此持强烈的怀疑态度，直接抛出一句定调：“我认为所有新闻都是心理战（psyop）。”（“I think all the news is a psyop.”）他提到自己曾在 Joe Rogan 的播客中听过 Bob Lazar 的访谈。尽管 Bob Lazar 听上去极具说服力，但 Hotz 倾向于认为这整件事不过是一场类似 MKUltra 的认知操纵实验。军方制造出某种奇异装置并冠以“外星飞船”之名，其真实目的可能只是为了刺激年轻物理学家的想象力，测试他们在误以为面对外星科技时能逆向工程出什么成果。 ### [12:33](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=753s) 阴谋论、假新闻与信息泛滥的控制术 对话顺势转向广泛存在的阴谋论。George Hotz 提到一个著名的元阴谋论：“我听过一种阴谋论说，‘阴谋论’这个词本身就是 CIA 在 60 年代发明出来用来给真实真相抹黑的。”（“I've heard a conspiracy theory that conspiracy theories were invented by the CIA in the 60s to discredit true things.”）他认为，像“地平说”或“披萨门”这种荒谬透顶的言论，实际上掩盖了现实中真正发生过的惊人真相——例如早期中国对新冠病毒预警医生的打压，或是美元彻底脱离金本位变成法定纯信用货币，这些事实常被人误当作无稽之谈。 Lex 表达了他对现代“假新闻”泛滥的深层担忧：当一切都被虚无化、任何事物都不再可信时，客观真理的根基就会彻底瓦解，每个人都缩进属于自己的“主观真相”里。Hotz 对此总结了权力机构控制信息手段的演进模式，并讲述了一个民间传说：有人在爱尔兰小矮妖（leprechaun）埋金子的地方系了一根红布条，并让小矮妖发誓不准拆掉布条；结果寻宝人带着铁锹回来时，发现整片田野的每一棵草上都系满了红布条。互联网让中心化审查彻底失效，于是控制者不再封杀信息，而是主动注入海量噪音和虚假叙事。Lex 也指出，这在密码学中是经典的隐写术（steganography）与隐私保护策略：为了掩盖一次真正的对话，系统可以生成数百万个深度伪造（deepfake）副本，从而彻底隐藏原始真相。 ```mermaid flowchart TD accTitle: 信息控制与真理稀释机制 accDescr: 展示从传统中心化审查到现代信息泛滥混淆的控制策略演变。 subgraph Classical["传统控制模式"] A["敏感真相出现"] --> B["中心化信息审查与封锁"] B --> C["阻断信息传播"] end subgraph Modern["现代网络环境模式"] D["互联网彻底打破审查限制"] --> E["主动生成海量虚假叙事与荒谬阴谋论"] E --> F["真伪彻底混淆（红布条掩盖金子）"] F --> G["公众丧失对客观事实的信任"] end C -.->|"网络时代失效"| D ``` ### [14:36](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=876s) 生命的编程语言：DNA 是 CAD 模型而非代码 Lex 将话题引回 Hotz 最擅长的领域，询问地外智慧生命是否必然会发展出计算机、算法与编程语言，以及这些系统的指令集是否会与人类所写代码存在共性。Hotz 认为自然语言和编程语言之间没有本质鸿沟，外星人的表达系统应该与人类大同小异，跨语言翻译不会特别困难。针对 Lex 提出的“自然界的运行逻辑（如 DNA）难以被图灵机代码（如 Python 或 C）编译”的疑问，Hotz 给出了明确的反驳： “DNA 是一个 CAD 模型，它不算是一种编程语言，它没有任何串行执行机制。”（“DNA is a cad model. It's not quite a programming language, it has no sort of serial execution.”） Hotz 解释道，人类并非不懂 DNA 的内在机理，真正的问题在于仿真代价高昂。如果要在分子层面上精确模拟蛋白质折叠等现象，所需的算力是天文数字。生命之所以高效，是因为它“在纯粹的自然物理硬件上直接裸机运行”（runs well on raw nature）。当探讨生物实验与自动化时，Hotz 吐槽了当今生物学实验室工具的极端落后。他直言自己无法忍受手动用移液枪吸取几十毫升溶液的原始工作，并嘲弄血液检测骗局 Theranos 的机器“就像高中科技节拼凑出来的简单 XY 轴导轨”。他坚信未来的生物工程必将走向完全可编程的微流控芯片，实现计算逻辑与生物界面的无缝对接。 - **DNA 的本质属性**：空间拓扑的 CAD 结构蓝图，而非具有串行指令流的执行程序。 - **生物模拟的算力瓶颈**：物理现实以纳秒级的分子力学并行运转，宏观模拟需耗费极其庞大的算力。 - **生物实验工具现状**：自动化与计算集成度严重脱节，仍高度依赖手动移液等低效机械动作。 ### [18:57](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=1137s) 黑客帝国隐喻：人类在上帝计算机中扮演的微不足道 谈及在上一次 SXSW 大会上引发热议的“入侵模拟宇宙（Hacking the Simulation）”话题，Lex 询问 Hotz 是否有了新的突破性见解。Hotz 澄清说，很多人曲解了他当时的意图，那场演讲并不是真的要大家去编写入侵宇宙系统的 Exploit 代码。探寻宇宙漏洞本质上就是理论物理学家的终极目标——构建出大统一理论（Grand Unified Theory），进而寻找底层维度的异常。 Hotz 说明那场演讲的真正内核是给人类敲响一记警钟：“你们在上帝的计算机里不过是几个字节而已。”（“You are bytes in God's computer really.”）他感叹人类总是把毕生心血和充沛情绪浪费在极其微小的事情上。面对 Lex 追问这是否是指人们沉迷卡戴珊家族八卦等娱乐消遣，Hotz 否认了：大众消遣无可厚非，他真正失望的是现代经济体系与社会所构筑的扭曲博弈。整个社会机器都在盲目追求拉抬股票价格与短期利益，将人类的全局目标函数（goal function）异化。虽然 Hotz 明确声明自己是资本主义的忠实信徒，但他冷峻地指出，当前社会运行的根本不是真正的自由市场资本主义，而是一场“规则已被暗中操控的作弊游戏”（a different game where the rules are rigged）。 ### [21:10](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=1270s) 真实游戏与人造游戏的博弈 George Hotz 阐述了人类所玩的各类“游戏”。他指出，人类创造了许多人造的、虚假的游戏，但真正值得投入的是大自然制定规则的真实游戏：“停止玩这些虚假的人类游戏，这里有一个真实的游戏，我们可以玩真实的游戏；真实的游戏就是自然写下的规则。”Lex Fridman 则提出了反思：大自然的游戏并不温情，正如 Instagram 账号“Nature is Metal”展示的那样，自然法则是残酷的掠食——大生物吃掉小生物，以此向另一个大生物炫耀以便交配繁衍。大自然本身没有艺术、音乐，也没有像 Comma.ai、Comma Two 或者 George Hotz 在 SXSW 的演讲那样的人类文明创造。 ```mermaid flowchart TD accTitle: 大自然与人类游戏层级 accDescr: 展示从自然底层到人类虚构层及死循环刷经验的演化路径 subgraph NatureLayer["大自然底层规则"] A["开放世界 MMORPG: 弱肉强食 / 繁殖淘汰"] end subgraph HumanConstructs["人类人造层"] B["艺术、音乐、技术创新 (Comma.ai)"] C["自我设限的子游戏 (如刷经验循环)"] end subgraph Danger["异化风险"] D["牺牲世界探索与真实乐趣<br/>陷入虚妄的多巴胺反馈"] end A --> HumanConstructs C --> D ``` George 反驳了 Lex 对大自然的悲观定义。他认为大自然本质上提供了一个完全开放世界的 MMORPG，具体的玩法完全取决于参与者自己。Lex 用经典游戏《暗黑破坏神》（Diablo）中的“母牛关”（The Secret Cow Level）做比喻： - 玩家发现快速获取经验值（XP）的最优解就是在母牛关中无休止地屠杀母牛。 - 所有人都默认在 RPG 框架下获取更多装备、技能和等级是唯一目标。 - 结果玩家牺牲了去探索广袤世界、享受游戏本身的乐趣，耗费数百上千小时沉溺于单调的杀牛循环中。 - 这种机制带来的是一种怪异、虚幻的多巴胺刺激，本质上是人类自我设限并败给了自然游戏机制的表现。 --- ### [23:51](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=1431s) 永生、终极价值与好奇心的衰竭 在探讨打破自然规则时，George 抛出了自己的终极目标——活下去：“我想永生，这……就是目标。”面对 Lex 关于死亡是否赋予生命意义的质问，George 澄清自己追求的并不是绝对不可逆的永生：“我想要的并不是那种绝对无法死去的真正永生，如果是那样，我宁可选择我们现在的状态；但我想自己决定何时去死。我不想让大自然来决定我何时死亡，我要赢，我要击败它。”他表示，自己会一直活着，直到“知道一切”为止；当真正无所不知时，才做好了死去的准备。 Lex 提出反驳：人类那种渴望知晓一切的驱动力，恰恰是因为生命是有限的；如果时间无限，这种好奇心可能像晨雾一样瞬间蒸发，导致生命陷入彻底的虚无。George 对此回应： - 对认知的渴望是他的终极价值（Terminal Value），不服务于任何其他功利目的。 - 即使好奇心真的依赖于有限感，未来也可以通过修改大脑代码来“自我欺骗”，例如植入让大脑以为“生命将在下周结束”的假象。 - 他设想通过版本备份机制（如类似 WaveNet 结构、按对数时间间隔留存的自我快照）来进行状态回滚，确保自我认知系统的稳定性。 - Lex 指出许多单机游戏一旦输入作弊码（Cheat Code），整个游戏瞬间丧失所有乐趣；George 则认为这表明当前对大脑的干预技术还不够成熟，未来应直接擦除大脑中记住作弊码的那部分神经回路。 --- ### [27:27](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=1647s) 模拟宇宙与物理法则的“内存泄漏” 对话由思想实验转向了物理现实的模拟假设。George 表示自己并不怀疑宇宙是个模拟系统，但他对纯哲学的思辨毫无兴趣：“我只在模拟假说能赋予我凌驾于自然之上的力量时，才对它感兴趣；如果它完全不可证伪，那谁在乎呢？”如果要在现实物理世界中寻找宇宙是虚拟环境的证据，最明确的信号就是能够发现并利用模拟系统的漏洞。 ```mermaid flowchart LR accTitle: 模拟漏洞类型对比 accDescr: 区分只读与可写系统漏洞在物理宇宙中的危险等级 subgraph ReadVulnerability["只读类漏洞 (Read)"] R1["千里眼 / 遥视 (Scrying)"] R2["任意内存读取<br/>类似 Heartbleed 漏洞"] end subgraph WriteVulnerability["可写类漏洞 (Write)"] W1["向物理法则写入随机噪声"] W2["重写基础参数与物质结构<br/>极度危险甚至毁灭宇宙"] end ReadVulnerability -->|"安全边界尚存"| Normal["可控观察"] WriteVulnerability -->|"灾难性后果"| Catastrophe["不可逆毁灭"] ``` George 提出： - 如果存在类似“遥视”（Scrying，即不受物理空间限制观测远方）的现象，这在计算机意义上就等同于宇宙底层存在“任意内存读取”（Arbitrary Memory Read）漏洞，类似互联网的 Heartbleed 漏洞。 - 读取漏洞还不算恐怖，真正毁灭性的是“内存写入”（Arbitrary Memory Write）漏洞——一旦任何人拥有向现实宇宙底层写入哪怕一个随机噪声比特的能力，整个物理现实的稳定性都将被打破。 --- ### [29:58](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=1798s) 万物理论与对现实掌控的力量 谈及物理学与计算理论的探索，Lex 提到了 Eric Weinstein 提出的“几何统一论”（Geometric Unity）以及 Stephen Wolfram 基于超图（Hypergraphs）的计算万物理论。George 坦言，自己对这些理论是否在数学上优美毫无偏好，这也正是他厌恶弦理论的原因。 > “做出一个可检验的预测吧。我对理论内在的美感毫无兴趣，我只对能赋予我掌控宇宙力量的理论感兴趣。” —— George Hotz George 进一步强调了他的实用主义态度：“我想要 100 GHz 的处理器，我想要比原子还小的晶体管，我想要力量。”Lex 对这种极其直白、剥离了伪善学术辞藻的坦率表示认同，但他同时指出，当力量达到这种极限尺度时，必然引出权力归属的问题——究竟该由政府、马克·扎克伯格（Mark Zuckerberg）、分布式黑客群体，还是 George Hotz 本人来掌握？George 随即指出，正因没有任何单一实体值得被托付这种终极权力，去中心化的区块链（Blockchains）技术才展现出了真正的价值，并顺势引出了他早年参与以太坊（Ethereum）项目的往事。 ### [31:49](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=1909s) 纽约餐厅的以太坊扩容偶遇与 Layer 2 原理 Lex Fridman 提起了一段广为流传的传奇轶事：George Hotz 在纽约的一家餐厅里，偶遇了一群正在为以太坊扩容方案赶工的加密团队工程师。他们面临着与 Solidity 到 OVM（Optimistic Virtual Machine）编译器相关的技术硬通牒。对方认出了 Hotz 并将他拉到一旁解释了困境，Hotz 当即出手帮他们解决了这个核心瓶颈，从而名声大噪。Hotz 证实了这件事完全属实，并称对方团队事后写了一篇非常溢美且动人的文章。这家公司名为 Optimism（源自 Plasma 项目的分支），其目标是在以太坊上构建 Layer 2（L2）扩容解决方案。 Hotz 解释了以太坊面临的本质瓶颈：在现有的 Layer 1（L1）基础链上，网络中的每一个节点都必须全量执行每一笔交易。这意味着如果你有 $N$ 台计算机，将节点增加到 $2N$ 台时，网络整体能处理的计算吞吐量其实完全没有增加，整体呈现 $O(1)$ 的计算伸缩性。虽然节点增多能带来更强的区块链安全性，但在基础链已经极其安全的前提下，完全可以通过某种方式用部分多余的安全性来置换计算速度，而这正是 L2 方案的意义所在。以太坊虽然在严格意义上因为资源 Gas Limit（燃料上限，以太坊原生用来给每条指令计价以防范 DDoS 攻击的机制）而不是纯粹的图灵完备机，但其可编程性远非只能处理基础脚本且非图灵完备的比特币可比。 - **Layer 1 与 Layer 2 分工**：L1 是以太坊底层基础结算链；L2 则是运行在链外的独立计算层，运算在外部执行，但其最终状态仍由 L1 的安全机制进行兜底背书。 - **计算瓶颈本质**：传统全节点验证导致链上算力呈现 $O(1)$ 扩展，增加节点仅增强冗余和共识安全性，无法提升计算通量。 - **共识与哈希机制**：以太坊当时仍运行在以太坊 1.0 的工作量证明（Proof of Work）机制下，采用与比特币不同、更具 ASIC 抗性（依赖 RAM）的哈希算法，同时正在规划转向以太坊 2.0 的权益证明（Proof of Stake）。 ### [35:00](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=2100s) 从 3000 行转译器到 300 行编译器：Hotz 的降维解决 谈到 Optimism 团队当时尝试解决的具体技术问题，Hotz 解释了链下计算与欺诈证明（Fraud Proof）的运作逻辑：团队希望把繁重的智能合约计算完全搬到链下执行，仅将最终结果写回链上。由于链下提交可能存在作恶和虚假结果，系统引入了挑战博弈机制。提交者必须质押资金（例如质押 10,000 美元）担保结果无误；任何人如果发现作假，即使在 L1 上执行单步验证需要花费昂贵的 Gas 费（例如 3,000 美元），也能因戳穿谎言而赢得那 10,000 美元的悬赏，从而在经济学意义上保证了安全性。 为了实现这一点，L2 必须是一个受控的“沙箱”（Sandbox），就像 Chrome 浏览器渲染进程不能直接调用 Linux 的 `open` 系统调用读取文件系统一样，任何读取和写入以太坊链上状态的系统调用都必须被拦截并转交给 Hypervisor 处理。当时 Optimism 采用的方法是先将 Solidity 编译为字节码，再用一个自主编写的“转译器”（Transpiler）对字节码进行拆解、篡改并重新组装，把违规的系统调用替换为指向 Hypervisor 的调用。这个转译器已经变成了一个长达 3,000 行且漏洞频出的泥潭。Hotz 听到这个设计后立刻反问他们为什么不在生成字节码的第一站——编译器本身去做这件事。在谈好报酬后，Hotz 直捣黄龙，直接给官方 C++ 编译器提交了一个仅约 300 行的简洁 Diff（补丁）代码，优雅地彻底解决了问题。当被问及这种化繁为简的方法论时，Hotz 坦言对方觉得他“先思考，然后用正确的方法做”的说法很有趣，但实际上根本没有什么深奥捷径：“This is always what I say, I'm like, do you want to be a good programmer? Do it for 20 years.” ```mermaid flowchart TD accTitle: Solidity到L2执行流对比 accDescr: 对比Optimism团队原本繁琐的转译器路线与George Hotz直接修改编译器的简洁路线 subgraph Origin["Optimism 原有方案（3000 行泥潭）"] A["Solidity 源码"] --> B["标准以太坊编译器"] B --> C["原始 EVM 字节码"] C --> D["3000 行字节码转译器\n(拆解、修改、重构)"] D --> E["OVM 安全字节码\n(转交 Hypervisor)"] end subgraph Hotz["George Hotz 方案（300 行编译器 Diff）"] F["Solidity 源码"] --> G["修改后的编译器\n(直接拦截敏感系统调用)"] G --> H["OVM 原生安全字节码"] end ``` - **沙箱隔离模型**：合约运行在 L2 容器中，所有涉及状态读写的底层系统调用（Syscalls）被重定向至宿主监管程序（Hypervisor）。 - **重构复杂度对比**：反向解析并重写已生成的二进制字节码极其繁琐脆弱；而在 AST（抽象语法树）生成底层代码的编译阶段进行注入，代码仅需约 300 行 C++。 - **编程没有捷径**：所谓直觉和快速破局，本质上是超过二十年高强度编码训练所积累的系统化工程洞察。 ### [39:30](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=2370s) 加密货币的宏大构想：中本聪共识与“代码即法律” 从具体的编译器工程跳出来，Lex 询问 Hotz 对整个加密货币领域的技术与哲学看法。Hotz 表示自己长期来看极度看好加密货币（Bullish），虽然他不一定认同某一个具体的币种或代币项目，但他认为加密货币带来了两项人类历史上极其震撼的发明。第一项是中本聪共识算法（Nakamoto Consensus），Hotz 称其为“21 世纪最伟大的创新之一”，它用一个相对简单直白的数学与算法机制，首次让人类群体在去中心化、无需相互信任的条件下达成了确定性的共识。Hotz 坦言自己最崇敬的人就是中本聪，并坚决否认了社区中猜测他是中本聪或 Elon Musk 是中本聪的荒谬假说。 第二项深刻的革命则是智能合约以及“代码即法律”（Code is Law）的思想。Hotz 将传统法律体系与计算机程序做了一个辛辣且幽默的对比：在现实世界中，两方签订合同后一旦出现争议，解释权落在了律师手里。Hotz 评价道：“Lawyers are just really shitty overpaid interpreters.”他用 Python 与律师进行了多维度对照：在 Azure 的优质计算节点上跑 Python，每小时成本只要 10 美分，而顶尖律师每小时要收 1,000 美元，Python 在经济效率上胜出整整一万倍；更致命的是，面对同样一段条款，不同的律师给出的判决和解释经常飘忽不定、自相矛盾，而 Python 运行相同的逻辑却几乎永远能输出确定一致的结果。以一个存放 100 万美元资产、约定需三方中任意两方到场才能解锁的保险箱为例，现实中的银行合约完全依赖于银行和司法体系的执行意愿，而智能合约则以不可篡改的数学执行力消除了人的不确定性，Hotz 极其渴望看到一个全社会普遍认同并接纳“代码即法律”的未来世界。 - **中本聪共识**：打破了拜占庭容错问题在社会学尺度上的困局，提供了一种去中心化、形式化且不可篡改的群体共识范式。 - **法律与解释器对比**：律师被视为昂贵且低效的不确定性“解释器”；智能合约提供了一种毫秒级、极低成本、无主观偏差的状态转移保障。 - **多签信任机制**：以 2/3 门限签名机制（Multisig）替代传统银行的三方保险柜协议，将制度信用彻底锚定在数学确定性之上。 ### [42:28](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=2548s) 智能合约、法律成本与对确定性的追求 乔治·霍兹（George Hotz）对比了传统法律体系与基于密码学的智能合约之间的巨大鸿沟。在现实生活中，如果有人过世，往往会引发漫长而丑陋的法律纠纷，争执遗产究竟是否在遗嘱之内，而这种查明真相的司法成本极其高昂。相比之下，智能合约利用密码学机制即可轻松解决争议，例如只需验证三把密钥中是否出现两把。这种纯粹的数学逻辑能带来绝对的确定性，而这正是商业世界最核心的诉求。霍兹以维亚康姆（Viacom）起诉 YouTube 的经典版权诉讼为例指出，大公司甚至根本不在乎判决倒向哪一边，正如 YouTube 当时的态度：大家只希望尽快拿到结果，以便获得经营上的确定性。 莱克斯·弗里德曼（Lex Fridman）询问这是否仅局限于金融交易，霍兹则直言智能合约的应用边界可以无限延伸——甚至婚前协议都可以部署在以太坊区块链上。霍兹戏谑地表示：“抱歉了离婚律师们，你们要被 Python 替代了。”无论是婚姻契约还是商业协作，代码的确定性正在逐步侵蚀模糊的人为裁量空间。 ```mermaid flowchart TD accTitle: 传统法律纠纷与智能合约执行机制对比 accDescr: 对比传统法律系统中冗长昂贵的司法判定流程与智能合约依靠密码学即时确定的执行路径。 subgraph LegalSystem["传统法律争议"] A["发生纠纷 (如遗嘱/婚前协议)"] --> B["漫长法庭诉讼"] B --> C["高昂判定成本与人为不确定性"] end subgraph SmartContract["智能合约体系"] D["链上触发条件"] --> E["密码学校验 (如 2/3 多签)"] E --> F["即时执行与绝对确定性"] end ``` --- ### [43:32](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=2612s) 加密货币格局与“天才领袖”的系统性缺失 当谈到加密货币领域数十年的长期演变以及究竟谁能胜出时，霍兹坦言自己从不做市场投机，也不在乎哪个项目最终获胜。在他眼中，比特币在技术上显得粗糙落后，本质上更像是一种“模因币”（Meme coin），而历史上那场闹得沸沸扬扬的比特币“区块大小之争”（Block Size Debate）在他看来简直愚蠢到不值一提。相反，他更认可治理模式相对优秀的以太坊，并转变观念接受了权益证明（Proof of Stake）等机制。 两人进而深入探讨了系统治理与领袖的关系。弗里德曼指出，即使在分布式系统中，人类似乎仍需要一个具象化的面孔来凝聚愿景，维塔利克·布特林（Vitalik Buterin）正是以太坊的象征。霍兹指出，维塔利克能成为核心是因为他本身是个无可挑剔的天才，而非传统意义上的政治领袖。两人共同感叹当前现代民主制度的失灵：在现行体制中脱颖而出的往往不是系统中最能干的精英。 霍兹幻想总统大选辩论应该提出硬核常识与科学问题：“想象一下辩论赛上会提问：宪法第六修正案是什么？宇宙的四种基本力是什么？$2^x$ 的积分是多少？”弗里德曼深表赞同，并以著名历史播客主播丹·卡林（Dan Carlin）为例，指出真正的顶尖能力就像卡林那样，虽然没有历史学博士学位，却能凭大脑串联起成吉思汗到二十世纪的整个脉络，洞悉每场战役细节与背后的权力博弈。顶尖程序员也具备这种将海量知识内化为知识树的能力，只可惜这种真正的胜任力很少出现在现实政治领袖身上。 --- ### [47:23](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=2843s) 代码分叉（Fork）作为一种社会与商业实验机制 霍兹认为加密世界最强大的核心思想是“分叉”（Fork）。在现实政治中，公众无法同时检验两条道路，但在区块链世界中，当社区发生根本分歧时，系统可以通过硬分叉裂变——例如以太坊分叉为以太坊（ETH）与以太坊经典（ETC），资本和用户可以直接用资金投票，将流动性从一条链撤出并投入另一条链。霍兹调侃自己在 2012 年开发的一款恶搞应用中就提过这个设想：“巴拉克·奥巴马与米特·罗姆尼，让我们让他们俩都当总统吧！想象一下我们可以分叉美国，让他们同时执政，让两个美国展开竞争。” 霍兹进一步主张，这种机制理应引入到商业实体中： - **英伟达的分叉实验**：霍兹表示自己很想“分叉英伟达”，让不同团队以完全不同的商业战略去市场中竞争验证。 - **Comma.ai 的开源优势**：一家 Comma.ai 可以选择闭源，另一家保持开源；一家可以被通用汽车（GM）收购，另一家作为独立的反叛力量继续运营。霍兹兴奋地表示，正因为 Comma.ai 坚持开源，任何人如果对路线不满，随时都可以直接 Fork 它的代码库另立山头。 --- ### [48:57](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=2937s) 拒绝平庸的成功学：在实战中学习 面对大众普遍好奇的“如何持续跨领域快速学习新知”这一问题，霍兹表现出了对市面上鸡汤式成功学的极大反感。他戏称自己完全可以写一本定价 67 美元的畅销自救书，封面上就大写一句：“书中所有的建议都毫无意义，但你反正也是个冤大头肯定会买。”他坦言自己根本无法给出程式化的方法论，因为他极度抗拒那种过于刻意、按部就班的生活方式。 弗里德曼分析道，霍兹目前正处于激烈的实战前线，就像那些全神贯注于比赛的职业运动员，对停下来写一本“如何掌握这项运动”的教程毫无兴趣。霍兹的学习正是通过高强度的直接实践完成的。不过，弗里德曼也指出，随着一个人在特定领域变得极度专业且公司事务繁忙，走出舒适区去涉足不擅长的新领域会变得痛苦不堪。而霍兹在 Twitch 等平台直播写代码时，展现出了一种不惧暴露无知、在镜头前即兴探索并快速解决问题的罕见特质。 霍兹坦承自己也并非无所不能，并非所有尝试都能迅速见效。例如他在国际象棋上投入了数百个小时，等级分却依然停留在孱弱的 1400 分左右。他自嘲道：“我知道如果用更好的方法我可以下得更好，比如别下 5 分钟快棋，至少下 15 分钟慢棋……但我知道归知道，这些东西就是无法顺畅地长进我的知识树里。” --- ### [52:37](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=3157s) Comma.ai 的核心愿景：交付可落地的阶段性成果 在访谈话题转向霍兹创立的自动驾驶公司 Comma.ai 时，弗里德曼询问这家公司的宏观愿景究竟是什么。与硅谷同行那些空洞宏大的画饼不同，霍兹脱口说出了一句极其精准利落的企业使命陈述： > “解决自动驾驶问题，同时持续交付可出货的阶段性产品。”（“Solve self-driving cars while delivering shippable intermediaries.”） 霍兹阐明，虽然最终目标是攻克完全无人驾驶，但关键路径必须依赖于在每个演进阶段都制造出真实存在、能卖给终端用户并产生实际价值的软硬件产品，而不是躲在实验室里等待数十年后虚无缥缈的突破。 ### [53:00](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=3180s) 什么是“解决自动驾驶”与基于共识的资本主义 关于究竟什么是“解决自动驾驶”（solve self-driving cars），George Hotz 给出了一个极其直接的定义：你并不是要制造一辆全新的汽车，而是要制造一个“人类驾驶员的替代品”（a person replacement）。这个替代品必须能坐在驾驶位上，把人带到任何人类能开车到达的地方，并且在安全性、速度、驾驶质感与舒适度上达到甚至超越人类的水平。 谈及 Comma.ai 的商业模式，Lex Fridman 询问向用户持续交付可发货的中间产品（shippable intermediaries）是否仅仅是为了给公司提供资金。George 回应称，自我造血当然重要，但交付产品更是“保持诚实”（keep us honest）的核心机制。如果没有实际交付的压力，在研发这项技术的过程中，团队极易陷入自我感动的幻觉，误以为自己取得了重大进展。正如 Hacker News 上一句一针见血的评论所言：你可以随意设定任何虚假的目标节点，哪怕全部达成，你也可能距离真正解决自动驾驶还有无穷远的距离。对于 Cruise 或 Waymo 这样没有实际商业营收的公司来说，由于缺乏真实的交付反馈，很难设定出真正有意义的硬性截止日期。 在 George 看来，真正的资本主义建立在自愿与共识的基础之上。如果用户愿意掏出 1,000 美元购买一套 Comma Two 设备，唯一的原因就是这套设备为他们提供了至少价值 1,000 美元的效用，否则交易根本不会发生。George 犀利地总结道：“真正的资本主义是建立在自愿共识之上的。市面上确实存在骗局，但真正的资本主义之所以能获得收入，是因为我们向另一个人提供了至少等值的价值。” ```mermaid flowchart LR accTitle: Comma.ai Honest Progress Flywheel accDescr: A compact diagram illustrating how selling products to real users validates actual technological progress. subgraph Validation["真实反馈机制"] A["交付可发货产品 (Comma Two)"] --> B["用户自愿支付 ($1,000)"] B --> C["创造真实市场价值"] end subgraph Research["避免伪进展"] C --> D["迫使团队保持诚实"] D --> E["杜绝虚假里程碑幻觉"] E --> A end ``` --- ### [54:30](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=3270s) Comma Two 硬件与极简设计哲学 在回顾 Comma.ai 的产品演进时，George 直言早期版本不过是对同一构想的渐进打磨，目前核心在售的硬件仅有 Comma Two。可以将 Comma Two 拟人化为一个人类驾驶员：它拥有双眼（前后摄像头）、耳朵、嘴巴、大脑以及与汽车进行物理交互的接口。面对 Lex 开玩笑询问该设备是否具备“自我意识”时，George 幽默地反驳称自己同样没有意识，两者并无本质区别，唯一的差异不过是人脑的计算能效远高于芯片——人脑消耗约 20 瓦功率即可输出约 20 PFLOPS 的计算能力。 关于设备发布的时间点，Comma Two 于 2020 年初正式面世。Lex 感慨 2020 年漫长得如同过去了十年。当被问及 Comma Two 最性感的功能特征时，George 用官方口号精准概括了其核心目的：“让驾驶变得轻松惬意”（make driving chill）。它绝非简单的定速巡航，而是包含自适应巡航控制（ACC）与车道保持系统（LKA），能够自动跟随前车减速、精准保持车距，并且在机器学习算法的加持下，其居中行驶质感随着迭代越发丝滑。 在硬件物理架构层面，Comma Two 本质上是一台经过深度魔改的智能手机，并加装了专属散热系统与车载通信接口： - **核心计算平台**：高通骁龙 821 处理器（Qualcomm Snapdragon 821），充分榨干手机芯片的每一项能力。 - **无线连接模组**：集成 Wi-Fi 与 LTE 通信模块，用于系统 OTA 升级与行驶数据云端上传。 - **人机交互与传感**：自带显示屏与前视/车内双向摄像头，车内摄像头专用于驾驶员监控（Driver Monitoring）。 - **车载物理接口**：设备通过 USB 端口连接专用的汽车线束（Harness），向外引出 4 路 CAN 总线（Controller Area Network），实现对车辆总线数据的底层读取与控制指令下发。 --- ### [58:07](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=3487s) openpilot 软件生态与端到端横向控制 相比于硬件的精炼，开源自动驾驶系统 openpilot 才是系统的真正核心。外界常常质疑模型权重是一个非开源的代码块（blob of weights），George 对此回应道，权重本身就是一个权重文件，真正的软件工程在于围绕这颗深度神经网络所构建的整套运行时生态。openpilot 扮演着粘合剂的角色：清洗输入喂给神经网络、校准传感数据、执行输出控制、利用传感器融合与定位算法处理道路坡度（例如在倾斜路面上对冲重力转向），同时负责热管理、存储调度以及将驾驶数据切片回传到服务器供云端训练。 谈及算法架构的飞跃，George 指出团队兑现了一年前的承诺：实现了横向控制（Lateral Policy）近乎完全的“端到端”（End-to-End）闭环。现在的 openpilot 彻底摒弃了传统的语义分割、车道线识别（Lane Marking Detection）与显式物体检测流水线，而是完全由人类日常驾驶行为的真实轨迹数据训练而成。甚至在关闭车道线显示模式下，系统依然能仅凭视觉画面顺畅驾驶，尽管在高速公路上的平稳度略微逊色于标线完好的情况。接下来，团队正致力于将纵向控制（Longitudinal Policy，即加减速与跟车逻辑）全面改造为端到端架构。 ```mermaid flowchart TD accTitle: openpilot End-to-End Control Pipeline accDescr: Pipeline showing raw inputs flowing into the neural network and how openpilot glue logic outputs to the CAN bus. subgraph InputProcessing["输入与校准"] CAM["摄像头原始视频流"] --> PRE["输入清洗与校准"] SENS["IMU 与车身传感器"] --> FUS["传感器融合与定位"] end subgraph NeuralCore["神经网络推理"] PRE --> NN["端到端驾驶模型权重 (Blob of Weights)"] FUS --> NN end subgraph Execution["控制执行与车载通信"] NN --> POST["输出清洗与重力坡度对冲"] POST --> CAN["4 路 CAN 总线指令下发"] CAN --> CAR["控制方向盘转向与油门刹车"] end ``` --- ### [01:00:44](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=3644s) 数据筛选哲学与强化学习闭环 面对海量涌入的用户驾驶数据，如何筛选高价值的训练集成为核心挑战。Lex 指出驾驶场景充斥着各种长尾极端边缘案例（Edge Cases），筛选机制决定了模型的智能上限。George 从监督学习与强化学习的数学本质切入，给出了一个极富哲理的辨析：“在监督学习中，模型权重取决于数据；而在强化学习中，数据则取决于模型权重。” 基于这一定义，George 描绘了 Comma.ai 的宏大蓝图——在现实物理世界中运行强化学习闭环（RL on the world）： - 只要车辆在实际道路上出现判断失误，逼迫人类驾驶员接管干预（Disengagement），系统就会立刻把这次接管事件记录为关键样本。 - 云端针对接管场景重新训练模型，生成新权重并推送到整车终端。 - 每一次全新推送的软件模型，实质上就构成了现实世界强化学习中的一个训练轮次（Epoch）。 关于全自动驾驶的时间表，George 明确拒绝像埃隆·马斯克那样做出遥遥无期的激进承诺，坚持将当前系统严格限制在 L2 级辅助驾驶的范畴，并强制执行人工监控。从硬性指标来看，openpilot 的非计划接管里程（Unplanned Disengagements）已经从约一年前的每 10 英里一次大幅优化到目前的每 100 英里一次，在一年内实现了整整 10 倍的可靠性提升。但 George 清醒地坦言，距离商业级全自动驾驶所必需的 100,000 英里接管一次的标准，整个系统在未来仍然需要再跨越连续三个数量级（10x × 10x × 10x = 1000x）的极限工程突破。 - **接管里程跨越**：从 ~10 英里/次 优化至 ~100 英里/次（实现 10 倍跃升）。 - **完全自动驾驶目标壁垒**：必须达到 ~100,000 英里/次接管，尚需三次 10 倍的指数级提升。 - **系统层级定调**：坚决锚定在受人类监督的 L2 级别，杜绝无边界的营销炒作。 ### [01:02:50](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=3770s) 自动驾驶的技术范式之争：Tesla 的多任务架构 vs. Comma.ai 的端到端哲学 讨论从自动驾驶系统的演进速度展开。Lex Fridman 提到人们常在直觉上默认系统性能存在“指数级提升”，但即便存在这种倍增关系，也绝不意味着明天就会凭空出现 100 GHz 的处理器，从最顶尖的机器系统跨越到人类驾驶水平，中间仍隔着巨大的鸿沟。 紧接着，Lex 剖析了特斯拉（Tesla）与 Comma.ai 截然不同、且均与 Waymo 等传统车企大相径庭的技术路径： - **Tesla 方案（多任务学习 + 数据引擎）**：将驾驶任务拆解为数百个高度细分的子任务（车道线检测、停止标志识别、红绿灯分类、可通行区域分割、行人和自行车检测、轨迹预测等），底层使用多头神经网络（Multi-headed Neural Network），上方再由控制策略模块将信息拼接以执行决策。其优势在于可以建立一套主动挖掘边界案例（edge cases）的数据引擎，并配备专职工程师团队将单个任务做到极致。 - **Comma.ai 方案（端到端监督学习）**：将整个驾驶视为统一的机器学习问题，直接基于人类驾驶员的日常操控数据（纵向与横向控制）以及发生人工接管（disengagement）的失败样本进行端到端（End-to-End）训练。 ```mermaid flowchart TD accTitle: Tesla Multi-task vs Comma 端到端架构对比 accDescr: 对比特斯拉细分任务的数据引擎架构与 Comma.ai 的端到端机器学习流程 subgraph Tesla["Tesla: 多任务拆解与特征工程"] T_Data["摄像头/传感器数据"] --> T_Net["多头神经网络 (Multi-task Heads)"] T_Net --> T_Tasks["子任务拆解<br/>• 车道线 / 信号灯<br/>• 障碍物 / 可通行区域<br/>• 轨迹预测"] T_Tasks --> T_Engine["数据引擎持续挖掘 Edge Cases"] T_Tasks --> T_Policy["规则/策略拼接层"] T_Policy --> T_Control["车辆控制决策"] end subgraph Comma["Comma.ai: 端到端学习哲学"] C_Data["人类真实驾驶流数据<br/>+ 接管失败样本"] --> C_Model["端到端大模型<br/>(端到端感知到轨迹控制)"] C_Model --> C_Control["车辆直接控制输出"] end ``` 面对两种路线的优劣之争，George Hotz 用国际象棋打了个辛辣的比方。他反问 Lex：如果今天要创办一家国际象棋引擎公司，难道还要去专门招聘一个“象专家”（Bishop guy）或“车专家”（Rook guy），到处炫耀“我们把那家公司的车专家挖过来了，我们的车会走得无敌好”吗？在 AlphaZero 早已扫平棋坛的今天，这种做法极其荒唐。 Lex 辩解称当年由于特征工程占主导，确实会做这种模块拆解，而且驾驶的复杂度远超国际象棋。但 George 坚信特斯拉迟早会彻底转向端到端： > “End-to-end is going to be the only solution for Level 5.”（George Hotz：“端到端将是实现 L5 级自动驾驶唯一的解决方案。”） Hotz 承认他不确定人类何时能走到那一步，但他直言，特斯拉目前拆解任务并逐个挖掘边界案例的做法，本质上仍是高阶的“特征工程”（Feature Engineering）或“任务工程”： > “And anything about the history of AI has taught us anything, it's that feature engineering approaches will always be replaced and lose to end to end.”（George Hotz：“如果说人工智能的历史教会了我们什么，那就是特征工程方案终究会被取代，并且彻底输给端到端方案。”） 他对比了传统国际象棋软件 Stockfish 与 AlphaZero 的代码量——AlphaZero 的代码不仅短得多、优雅得多，而且耗费的程序员人时要少几个数量级。 - 特斯拉的多任务学习将感知与规划显式解耦，便于针对特定场景定向挖掘训练样本。 - Comma.ai 主张跳过人工拆解子任务的过程，直接用海量人类驾驶与接管样本训练端到端网络。 - 历史经验表明，深度学习领域中所有高度依赖人工定义特征的架构，最终都会被统一的端到端学习模型取代。 --- ### [01:09:12](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=4152s) MuZero、世界模型与终极解决路径 Lex 随即对端到端提出了一个尖锐的工程现实问题：AlphaZero 之所以能够摆脱特征工程，是因为它拥有一个完美的棋盘模拟器，在虚拟世界里“残杀”（murdered）了数以亿计失败的智能体；这与地球生物演化在通往人类的过程中牺牲数以万亿计的有机体如出一辙。但在现实物理世界中，自动驾驶的试错代价极其高昂，不可能在真实马路上任由汽车撞毁来试错。 George 对此给出了明确答案：**MuZero** 就是解开这个死结的关键钥匙。 ```mermaid flowchart LR accTitle: MuZero 在驾驶决策中的推演机制 accDescr: 展示环境输入通过学习到的动力学模型进行隐空间推演与价值评估的闭环 Obs["真实环境输入 / 历史观测"] --> Latent["状态隐编码 (Representation)"] subgraph Dynamics["内部学习的世界模型 (Learned Simulator)"] Latent -->|"虚拟动作"| Pred["动力学前向推演 (Rollout)"] Pred -->|"预测下一步"| Value["价值与奖励预估"] end Dynamics --> Action["输出最优驾驶轨迹动作"] ``` 在 George 看来，DeepMind 提出的 MuZero 论文是整个深度学习时代最被低估的里程碑之作： > “MuZero is the solution to self-driving cars... it does those roll outs and those murdering in in a learned simulator, in a learned dynamics model.”（George Hotz：“MuZero 就是自动驾驶的终解……它在一个自学习的模拟器中、在一个学出来的动力学模型里，去执行前向推演和那些‘残酷淘汰’。”） MuZero 不再依赖人类写好的物理仿真器，而是自己从数据中学会环境的动力学模型，并在内部隐空间中完成推演与决策。Hotz 直言自己创办 Comma.ai 以来一直在静候自动驾驶的终极理论，而看到 MuZero 论文的那一刻，他确信自己找到了。 谈及自动驾驶落地的时间表，Hotz 认为纯神经网络终将完全解决驾驶问题，时间点预计落在 **2020 年代的后半期（late 2020s）**。但在通往终极目标的漫长道路上，企业必须有可盈利的产品落地支撑研发。特斯拉和 Comma.ai 目前都在打造体验不断进化的 L2（Level 2）高级辅助驾驶系统。 对于 SAE 自动驾驶等级中的 L3、L4 和 L5，Hotz 表现出极具极客务实风格的不屑： > “The only difference between L2 and the other levels is who takes liability, and I'm not a liability guy. I want to take liability Level 2 forever.”（George Hotz：“L2 和其他更高等级之间唯一的区别就在于谁来承担法律责任，而我可不是来背锅的。我宁愿永远承担 L2 级别的责任。”） - AlphaZero 依赖已知的完美游戏规则模拟器，而现实世界的物理规律与多智能体交互无法精确写出模拟代码。 - MuZero 能够自主学习内部动力学模型（Learned Simulator），使得算法可以在想象的世界模型中进行推演和自我博弈。 - 自动驾驶商业落地不必死磕法律责任划分，提供持续进化的 L2 级系统是兼顾商业闭环与技术迭代的最优解。 ### [01:12:41](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=4361s) 场景自适应的驾驶员监控与风险评估 Lex Fridman 抛出了一个高级辅助驾驶面临的核心难题：随着辅助驾驶系统性能不断精进，人类驾驶员必然会产生过度信任（over-trust），从而变得越来越懒散、注意力涣散，系统该如何拿捏监控的尺度？George Hotz 表示，Comma.ai 的驾驶员监控系统（Driver Monitoring）已经具备了“场景自适应”（scene-adaptive）能力。车内配备的红外摄像头会时刻追踪驾驶员，但系统执行监控策略的严苛程度完全取决于当前的路况与驾驶场景。 ```mermaid flowchart TD accTitle: 场景自适应驾驶员监控逻辑 accDescr: 展示Comma.ai系统如何根据道路风险动态调节监控策略 subgraph Perception["环境与风险评估 (ML)"] A["车速与车辆静止判断"] B["道路类型 (直线高速 vs 城市道路)"] C["障碍物与信号 (红绿灯/行人人行横道)"] end subgraph Policy["自适应监控策略"] D["车辆完全静止: 不强制监控"] E["空旷笔直公路: 放宽注意力阈值"] F["复杂路况 (45 mph/信号灯/行人): 严格高频警报"] end Perception --> Policy Policy --> G["动态警报与介入机制"] ``` George Hotz 解释道，如果车辆完全处于静止状态，系统根本不会主动施加监控压力；而当车速在 45 英里/小时、穿行于遍布红绿灯、停车标志和潜在行人的复杂街区时，系统就会执行极其严格的驾驶员注意力监控。如果车子正行驶在一条空无一人的笔直高速公路上，策略则会相应放宽。除了车辆静止由手写规则判断外，其余场景几乎完全依赖机器学习对潜在风险进行实时评估。Lex Fridman 对这种风险估计思路赞不绝口，指出人类在高度专注时本身就是极佳的风险评估者，机器走向这一步至关重要。George Hotz 强调，绝不能为了所谓的“安全”而把监控策略全程调到最激进，否则驾驶员不仅会产生严重的警报疲劳（alert fatigue），还会因为在毫无危险的场景被频繁打扰而产生逆反心理，最终甚至会想方设法去欺骗系统（比如在方向盘上挂配重环）。 - Comma.ai 采用红外摄像头（IR camera）进行车内驾驶员面部与视线追踪，能在夜间与复杂光照下工作。 - 场景风险评估主要依靠机器学习模型驱动，根据道路复杂度、周边障碍物与车速动态平衡监控灵敏度。 - 过于频繁且不合时宜的警报会导致用户彻底将其当成耳边风，并在心理上对系统产生反感与抵触。 --- ### [01:15:11](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=4511s) 特斯拉的路线转向与一万美元赌局 谈及驾驶员注意力机制，Lex Fridman 认为这不仅是一个驾驶技术问题，更是一个探究人类行为与认知心理学的绝佳切入点。这也正是他和 George Hotz 此前从不同角度批评 Elon Musk 的地方。Lex 提到，在自己与 Elon 的线上和线下交流中，Elon 始终把解决自动驾驶的终极目标放在“将人类彻底移出闭环”，因此早期对车内视觉监控人类驾驶员并不感冒；然而，在真正实现完全自动驾驶之前，人类依然是无法替代的安全底线。近期业内已有消息证实，特斯拉正开始在“影子模式”（shadow mode）下测试其驾驶员监控模型，尽管其车内配备的只是一颗低分辨率的可见光摄像头。 面对特斯拉这一潜在的路线转向，George Hotz 毫不避讳地指出 Comma.ai 的硬件优势：在整车硬件的其他维度上 Comma 确实远落后于特斯拉，但在车内驾驶员监控摄像头这一项上，Comma 的红外方案绝对领先于特斯拉目前的可见光/鱼眼方案。George Hotz 重申了他曾在播客上立下的著名赌局：“Elon will definitely ship driver monitoring before he ships level five.” Lex Fridman 对此表示赞同，但也为 Elon 的处境辩护，认为作为一家大规模量产车企的掌舵人，公开承认错误并掉头极其艰难。当车里最初没有标配专用红外硬件时，CEO 在公关层面上很难直接向公众承认现有配置的不足。George 则笑着回应称自己不是什么企业领袖，因而非常乐意随时改变想法，他相信 Elon 最终也能找到体面的台阶，在心理上跨过这道坎。 - 特斯拉正在影子模式下静默运行车内可见光摄像头的驾驶员监控模型，硬件上属于低分辨率普通镜头，与 Comma 的早期产品 EON 类似。 - George Hotz 愿意押注 10,000 美元，赌特斯拉必然会先正式推送驾驶员监控功能，然后才可能实现 Level 5 级完全自动驾驶。 - Lex Fridman 指出，人因工程学界（human factors community）虽然普遍倡导必须引入驾驶员监控，但仅仅安装摄像头并不等于解决了注意力问题，设计出兼顾安全与体验的产品极具挑战。 --- ### [01:19:54](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=4794s) 数据收集与 Super Cruise 的缺陷对比 George Hotz 进一步理清了驾驶员监控的设计哲学：在车内安装摄像头与何时向驾驶员发出警报，在工程逻辑上是完全解耦的两件事。摄像头哪怕只是在后台默默观察、从不发出一次误报，车企也能藉此收集关键的注意力数据，用于长期模型迭代和决策优化，这与粗暴地向用户报警打扰完全不同。Comma.ai 早在三年前就启动了这项数据沉淀，George 自豪地表示，Comma.ai 拥有目前全世界最好的驾驶员监控系统，其实际体验远超通用汽车的 Super Cruise。 谈到通用的 Super Cruise，George Hotz 分享了自己的亲身体验与痛点：当阳光穿过车窗斜射进来时，强光会直接导致 Super Cruise 的红外摄像头致盲，车辆随后就会判定驾驶员“注意力不集中”；更糟糕的是，此时驾驶员就算死死盯着正前方、甚至用手紧握方向盘施加扭矩，系统也不允许用户通过物理动作重置注意力状态，而是强行退出辅助驾驶。George Hotz 认为，人机交互本质上是一场持续的对话，系统不能把用户逼入绝境。为了将驾驶员监控打磨到极致，OTA（无线软件更新）是绝对不可或缺的基础设施。Comma.ai 每月都会推送更新，并频繁在 Discord 社区收集极客用户的真实反馈，动态微调监控策略的严格程度，力求在不引起用户反感与抱怨的前提下，将安全约束推到最紧绷的平衡点。 - 通用 Super Cruise 的主要缺陷在于抗强光干扰能力差，且在误判驾驶员分心时，不允许通过触碰方向盘来重置注意力状态，直接发生强行脱离（disengage）。 - Comma.ai 保持高频迭代，通常每个月通过 OTA 推送一套新的驾驶员监控模型与策略。 - 团队在 Discord 用户社群中实时收集体验反馈，以“在用户不反感、不觉得受打扰的前提下尽可能保持最高激进程度”为策略调优目标。 - Lex Fridman 总结指出，这种由数据驱动并通过 OTA 持续重塑产品体验的闭环能力，正是类似特斯拉与 Comma.ai 体系最核心的颠覆性创新。 ### [01:22:00](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=4920s) OTA 更新的本质与智能手机的启示 Lex Fridman 和 George Hotz 探讨了特斯拉如何让整个汽车行业彻底意识到无线软件更新（OTA, Over-The-Air）的核心地位。在 George 看来，这从来不是什么新鲜发明，早期的 iPhone 才是第一个真正普及 OTA 更新并证明其颠覆性价值的产品。两人回顾了过去的技术生态：以前的游戏机根本无法联网更新，后来才逐步跟进。智能手机真正令人惊叹的特质绝不仅仅是触摸屏这一块硬件界面，而在于它能够通过软件迭代不断自我升级、变得越来越强大。George 开玩笑提到自己非常喜欢当时的 iOS 14，强调硬件只是载体，能持续进化的软件才是产品的灵魂所在。 --- ### [01:22:38](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=4958s) 驾驶员监控为何远比外部感知简单 针对驾驶员监控（Driver Monitoring, DM）系统，Lex 提出异议，指出 George 过去常把某些难题说得轻描淡写，尤其是认为车内驾驶员监控远比外部环境感知容易得多。George 随即从底层工程逻辑进行了解释：在驾驶员监控领域，“特征工程（feature engineering）”依然行之有效，但在车外全自动驾驶感知中，传统的特征工程早已彻底失效。更关键的差异在于对系统容错率的要求完全不同。车内监控本质上是一场与人类的持续交互，系统是在“训练并引导”驾驶员保持警觉；即使监测算法只有 92% 的准确率，依然能发挥绝大部分安全效用。而车外环境感知与控制完全不容许闪失，一旦车辆判断失误冲向路边大树，留给系统的容错空间就是零。 Lex 承认在容错率这一维度上车内监控确实更容易，但他从人机交互（HCI）与计算机视觉的角度指出，人脸与微表情的复杂性实际上深不可测，车内本应有空间构建更为丰富、沉浸的产品体验。George 则持完全务实的实用主义态度，认为驾驶员监控没必要搞得那么复杂，系统唯一需要传达给人类的信号就是提醒对方“这里可能会出状况，请集中注意力”。谈到高速公路上的 L3 级自动驾驶，George 认为人类届时只要能低头玩《糖果传奇》（Candy Crush）或刷 Hulu 视频即可，所谓 L3 就是系统能在特定场景下允许人类暂时脱离驾驶任务，并给予驾驶员大约 5 秒的接管缓冲时间；而车辆能够准确评估当前路况的危险系数，正是实现这一缓冲机制的核心前提。 > “Feature engineering works for driver monitoring, feature engineering does not work for the external.” —— George Hotz > “If you have something that's detecting whether the human is paying attention, it only works 92% of the time, you're still getting almost all the benefit of that because the human—like you're training the human.” —— George Hotz --- ### [01:26:03](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=5163s) 传递不确定性与 Comma 的 UI 架构改造 Lex 提到了 Comma 和特斯拉 Autopilot 在控制风格上的微妙差异：当环境变得模糊或不确定时，Comma 系统会通过转向手感与触觉反馈（haptic feel）向驾驶员传递这种不确定性；而特斯拉 Autopilot 则倾向于表现得始终极其自信，哪怕面对险情也显得泰然自若，直到最后一刻可能突然失控。George 透露，自己很早以前与 Elon Musk 交流时就提到过这种“向人类传达不确定性”的概念，但对方当时并不买账。Comma 除了通过物理阻尼传递动态，也会在屏幕上明确标出车道线识别状态和轨迹规划。 为了把系统内部的置信度更直观地展现给用户，George 表示他们正在积极招募 UI 工程师与 Qt 开发者。Comma 之前在前端使用了部分 React 框架，但 George 决定将新版系统的 UI 完全迁移到 C++ 搭配 Qt 的技术栈上。尽管 React 等基于 JavaScript 的方案开发便捷，但作为推崇底层性能的黑客，George 坚定地偏爱 C++ 所带来的确定性与高效率。 ```mermaid flowchart LR accTitle: Comma 与 Autopilot 面对不确定性时的表现差异 accDescr: 对比两者在不确定场景下的信息传递方式，Comma 采用触觉与显式 UI，Autopilot 保持强行自信。 subgraph Uncertainty ["环境不确定性输入"] U["复杂/模糊路况"] end subgraph CommaPath ["Comma.ai 应对策略"] C1["方向盘触觉反馈 (Haptic)"] C2["显式 UI 渲染车道线/路径"] C3["人类驾驶员提前接管"] C1 --> C3 C2 --> C3 end subgraph TeslaPath ["Tesla Autopilot 应对策略"] T1["维持高置信度表面控制"] T2["延迟报警或突发失误"] T1 --> T2 end U --> CommaPath U --> TeslaPath ``` --- ### [01:27:51](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=5271s) 特斯拉 Dojo 芯片与英伟达的垄断困局 话题转向特斯拉自研的超算芯片架构 Dojo。Lex 询问 Dojo 是否类似于专门针对驾驶数据定制的谷歌 TPU，George 指出 Dojo 并非专门针对自动驾驶数据，本质上就是对标 TPU 的全自研机器学习训练硬件，特斯拉这么做完全是走苹果式的垂直整合路线。George 剖析了当前 AI 训练基础设施面临的严峻困局：全球开发者实际上被困在谷歌和英伟达构成的双头格局中，而谷歌根本不是合格的平台供应商。 George 吐槽谷歌云有着苛刻且排他的服务条款，早前甚至明文禁止用户将 Google Cloud ML 用于训练自动驾驶汽车或从事任何与其自身业务竞争的项目，因此他绝不可能去碰谷歌 TPU。这就导致整个市场完全被英伟达垄断。George 回忆道，自己在 2016 年曾是英伟达的狂热粉丝，当时黄仁勋（Jensen Huang）甚至亲自坐进过 Comma 的测试车里，彼时英伟达的股价才 30 美元每股。然而在 2018 年左右，英伟达内部管理层发生明显转变，彻底走向了“竭泽而渔”、榨取生态内每一分钱的商业模式。 George 极为尖锐地批评了英伟达的定价策略：售价高达 10,000 美元的 A100 数据中心芯片在底层架构上与售价 699 美元的 RTX 3080 并没有天壤之别，但英伟达却依仗技术垄断攫取难以想象的暴利。在他看来，英伟达这种追求短期股价最大化、拼命压榨生态价值的行为是在自掘坟墓。那些试图在自动驾驶领域追赶特斯拉的传统大型主机厂由于缺乏底层研发能力，只能病急乱投医般地与英伟达深度绑定，但这类合作注定步履维艰。如果英伟达愿意像全盛时期的英特尔那样，以合理利润将芯片不设限地卖给每一个人，本可以长久地统治计算市场，然而过度商业化和高溢价正在逼迫包括特斯拉在内的巨头以及社区开发者联手将其推翻。 > “Google has absolutely onerous terms of service restrictions... back in Google's terms of service it said explicitly you are not allowed to use Google Cloud ML for training autonomous vehicles or for doing anything that competes with Google without Google's prior written permission.” —— George Hotz > “Let's charge ten thousand dollars for an A100 when it's really not that different from 3080 which is 699.” —— George Hotz - **关键观点与事实要点**： - **训练硬件的现实制约**：市场上可用于大规模神经网络训练的体系实际只有 NVIDIA GPU 与 Google TPU；由于谷歌的排他性条款，绝大多数商业公司只能被动接受英伟达的垄断。 - **技术定价悬殊**：A100 企业级芯片与消费级旗舰 RTX 3080 的成本差异远小于其数十倍的标价差异（10,000 美元 vs. 699 美元），激怒了许多算力需求巨大的 AI 创业公司。 - **商业模式对比**：英特尔依靠开放、合理溢价的通用芯片建立了长期统治；而英伟达在数据中心与车机芯片领域的过度产品化和封锁策略，正直接促使特斯拉等巨头加速自研芯片 Dojo。 ### [01:31:30](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=5490s) 芯片策略：NVIDIA 的定价与特斯拉芯片外售的可能性 谈到半导体行业时，莱克斯·弗里德曼（Lex Fridman）指出，NVIDIA 首席执行官黄仁勋（Jensen Huang）是一位极具远见的企业家，通常会制定长达五十年的远期规划，但他对目前显卡高昂的定价策略感到不解，怀疑黄仁勋可能听取了错误的建议。尽管如此，弗里德曼对 NVIDIA 怀有深切的感激之情，称赞其产品对深度学习领域贡献巨大，并提到自己依然保留着许多 GTX 1080 Ti 显卡。然而，将所有赌注押在 NVIDIA 一家身上充满风险，正如埃隆·马斯克（Elon Musk）带领特斯拉研发 Autopilot 芯片和超级计算机 Dojo 一样，特斯拉采取了类似苹果的封闭生态路线，外界很难直接使用其硬件。 乔治·霍茨（George Hotz）认为特斯拉完全应该转变思路，对外出售硬件，包括车辆内部功耗仅为 30 瓦的神经网络推理芯片。霍茨指出：“Make some money off the chips it doesn't take away from your chip you're going to make some money free money and also the world is going to build an ecosystem of tooling for you right you're not going to have to fix the bug in your tanh layer someone else already did.”（通过卖芯片赚点外快，这不会削弱芯片本身的价值，既能获得额外收入，全世界还会为你构建一套庞大的工具链生态，你就不用自己去修 tanh 层里的 bug 了，别人早就替你修好了。） 弗里德曼提出疑问：特斯拉是否愿意让自己的技术出现在其他品牌的汽车或 iRobot 扫地机器人内部？霍茨分析称，关键在于企业要分清自身不可替代的核心优势。例如，特斯拉不应该向竞争对手出售电池包，因为电池直接关乎整车竞争壁垒；但 AI 加速器并非不可逾越的护城河，只要有人愿意投入约 1 亿美元，就能制造出类似的芯片。当前已有大量芯片初创公司涌入 AI 加速器赛道，迟早会有公司找对方向，甚至中国市场最终也会出现平价替代产品。 - NVIDIA 的 GTX 1080 Ti 至今仍被广泛视作一代经典算力卡。 - 特斯拉车载 FSD 芯片的功耗约为 30 瓦，霍茨建议将其独立商业化出售。 - 霍茨认为专用 AI 加速器的研发门槛并非遥不可及，大约 1 亿美元的资金投入即可造出同类芯片。 --- ### [01:34:19](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=5659s) 特斯拉 Autopilot 大重构与 Comma.ai 的 3D 模型升级 弗里德曼谈到特斯拉 Autopilot 团队正在进行的“全量重构”（Big Rewrite），埃隆·马斯克与安德烈·卡帕斯（Andrej Karpathy）试图将多个传感器的数据在时间维度上统一融合，构建出随时间推移的四维（4D，即 3D 空间加时间维度）世界模型。这种架构不仅让多任务网络（HydraNet）在空间感知与预测上更加系统，还能大幅简化人工标注流程——标注团队只需在统一的空间模型中完成标记，便可反向投影并分发给所有传感器与分支任务。弗里德曼借此询问霍茨，这是否会动摇 Comma.ai 坚持的端到端学习路线。 霍茨透露，Comma.ai 当时也正处于底层模型的重大重构阶段。此前旧版本采用“平坦世界假设”（Flat World Hypothesis），遇到车身纵仰（Pitch）时车道线会在视觉上发生偏移；而全新重构的模型摒弃了这一假设，直接在三维空间中输出目标信息。此外，Comma.ai 的训练架构也全面由 TensorFlow 转向了 PyTorch。谈及特斯拉在时间维度上进行标注的做法，霍茨显得不以为然：“Why are you building an annotator?”（为什么还要费尽心思去造一个标注工具？）在霍茨看来，端到端学习天然包含了人类行为的真实输出，过度依赖精密的人工标注工具链属于走偏了方向。 ```mermaid flowchart TD accTitle: Comma.ai 架构重构对比 accDescr: 对比 Comma.ai 旧版 2D 平面假设与新版 3D 模型的升级变化 subgraph OLD["旧版架构"] A1["TensorFlow 框架"] --> A2["平坦世界假设 (Flat World)"] A2 --> A3["2D 空间输出 (车身纵仰导致车道线变形)"] end subgraph NEW["新版架构"] B1["迁移至 PyTorch 框架"] --> B2["三维空间感知 (引入 Z 轴)"] B2 --> B3["完整 3D 轨迹与无标注重构"] end ``` --- ### [01:36:31](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=5791s) 数据引擎与驾驶员监控的端到端争议 弗里德曼对特斯拉的数据引擎（Data Engine）赞赏有加。特斯拉依托数十万车主的日常行驶自动收集极端边缘案例（Edge Cases），再通过特征工程与监督学习精准筛选难例进行重训，弗里德曼认为这种工程闭环能力极为强大。卡帕斯曾在推特上提到，将验证集按照损失值（Loss）从高到低排序，直接审视高损失的样本能够带来意想不到的洞察。霍茨认同这一思路，并坦言 Comma.ai 也运行着一套类似但更轻量级的数据引擎：每次训练新的分割模型后，在 10 万张图像上进行推理，筛出 Loss 最高的 1000 张，再由人工挑出其中的 100 张打标签后送入重新训练。霍茨承认特斯拉的数据闭环是顶尖的监督学习工程，但他依然坚持，纯粹的自动驾驶并不需要这种人工监督，因为人类驾驶行为本身就是最完美的天然标注。 双方随后将讨论延伸至驾驶员状态监控系统（Driver Monitoring System, DMS）。弗里德曼提出，究竟是采用端到端方式直接处理驾驶员面部视频，还是通过关键点检测识别眼部、瞳孔、眨眼频率和面部朝向，再辅以心理学指标（如方向盘微调逆转、使用手机动作）来做规则推理？霍茨指出，驾驶员监控不同于车辆控制，因为视频本身并不自带“是否走神”的真实客观标准（Ground Truth）：“The only way to get them is with you know other humans labeling it really.”（获取真值的唯一途径，实际上就是靠其他人类去手动标记。）霍茨透露，Comma.ai 当时负责驾驶员监控的团队仅有“半个人”，后续一旦团队扩张至两到三人，也会探索利用行为特征建立自监督或自动标注机制。 - 卡帕斯的方法论：在模型验证集中按 Loss 排序，优先分析 Loss 最高的极端样本。 - Comma.ai 的轻量数据闭环：从 10 万张推理图像中筛选 1000 张高 Loss 样本，人工确认 100 张进行标注重训。 - 驾驶员监控现状：该模块的判定具有极强的主观性，Comma.ai 早期仅由半个全职人力维持维护。 --- ### [01:40:38](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6038s) Comma Two 的安装与车型支持 在播客后半段，弗里德曼将话题引回 Comma.ai 的消费级硬件产品 Comma Two。对于想要尝试该系统的汽车爱好者，弗里德曼询问具体该如何将其安装在常见的量产车上（例如 2020 款丰田卡罗拉），支持的车型有哪些，以及安装过程是否有门槛。 霍茨介绍，系统目前已经支持了大约 91 款车型。通过即插即用的线束套件与原车辅助驾驶接口连接，用户无需对车辆原厂结构进行破坏性改造，即可在几十分钟内完成软硬件的部署并启动体验。 - Comma.ai 当时已原生适配支持约 91 款市售汽车。 - 安装过程主要通过专用线束连接车内现有的后视镜雷达或摄像头总线接口。 ### [01:41:04](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6064s) 车型支持、改装与“宜家家具级”的安装体验 George Hotz 介绍，Comma.ai 目前已支持 91 种不同汽车品牌与型号，并计划在年内将支持车型扩充至 100 款。在推荐车型时，他首推 2020 款丰田卡罗拉（Corolla）和 2020 款现代索纳塔（Sonata）。索纳塔虽然使用的是原厂纵向控制系统（仅采用 Comma 的横向控制），但整体调校非常平顺精致，原厂的跟车加速与刹车体验相当成熟。对于愿意深入钻研、动手能力强的用户，本田思域（Honda Civic）也是极佳的选择，但前提是必须从互联网某些极客圈子里找到并刷入修改过的电动转向助力（EPS）固件，以获取更大的转向扭矩。对此 George 笑着划清界限：Comma 官方绝不背书该改装，团队虽然早已掌握这一技术，但从未公开发布，而是耐心等待外部社区独立发现它。 针对普通消费者的安装门槛，George 强调只要车辆在官方支持列表内，整个过程极其傻瓜化。“如果你会拼装宜家的桌子，你就能在受支持的汽车上安装好一台 Comma Two，而且它插上就能正常工作。”（“If you can set up a table from IKEA, you can install a Comma Two in your supported car and it will just work.”）全套安装仅需约 10 分钟，YouTube 上也有详尽的步骤教程；但如果用户想要折腾高端定制功能或修复细枝末节的 bug，那就意味着进入了开发者社区的领地。 - 支持车型数量：现已达 91 款，目标年内突破 100 款。 - 推荐首选车型：2020 款 Toyota Corolla、2020 款 Hyundai Sonata（使用原厂纵向控制，Comma 提供横向控制）。 - 极客改装：Honda Civic 需刷写非官方 EPS 固件以提高转向扭矩极限，官方 Discord 社区保持高度活跃。 --- ### [01:42:29](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6149s) 开发者社区、SSH 开关与分叉（Forking）生态 Lex Fridman 询问，如果用户希望将 Comma 当作纯硬件平台，运行自己编写的代码或修改过的 Openpilot，具体需要怎么做。George 回答说，系统设置中有一个直接的“Enable SSH”开关，只要打开该开关，用户就可以远程登录到设备根目录，上传并运行任意自定义程序。目前社区的生态分布大约是：60% 的用户运行 Comma 官方原生代码，40% 的用户运行各种社区分支（Forks）。 这些第三方分支由活跃的开发者维护，有的用于适配官方尚未支持的冷门车型，有的用于定制用户界面——比如为喜欢机械仪表感的用户在屏幕上渲染大量跳动的实时参数读数，或者取消踩油门自动解除辅助驾驶的限制。然而，George 对安全边界有着极其严苛的态度。他明确表示，虽然用户拥有自己的车，可以做任何想做的事，但凡是开发用来绕过或关闭驾驶员监控系统（Driver Monitoring）的分支，其作者和传播者都会直接被剔除出官方 Discord 服务器：“如果你试图破坏安全功能，你就会被我们的 Discord 封禁，我不想和你有任何瓜葛。”（“If you're trying to subvert safety features, you're banned from our Discord. I don't want anything to do with you.”） - 用户群体分布：约 60% 运行官方 Stock 系统，约 40% 运行第三方维护的分支（Forks）。 - 开发者入口：系统内置一键开启的 SSH 访问权限，支持完全开源的代码替换。 - 社区红线：严禁篡改或禁用疲劳与注意力监测机制，违者直接封禁。 --- ### [01:44:00](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6240s) 接管逻辑与避免模式混淆（Mode Confusion） 当谈到实际驾驶中的接管逻辑时，George 深入探讨了“模式混淆（Mode Confusion）”带来的安全隐患。在许多传统车企的驾驶辅助系统乃至特斯拉早期 Autopilot 和通用 Super Cruise 上，横向车道保持与纵向自适应巡航（ACC）是割裂的。George 举例说，在驾驶 Autopilot 时常遇到这种尴尬：车道保持因某种原因退出，但自适应巡航依然挂着；当正前方的前车稍微变道让开后，特斯拉突然大油门猛烈加速冲向原本设定的巡航速度，让毫无心理准备的驾驶员措手不及。 为了杜绝这种横纵向状态不一致导致的认知混乱，Openpilot 采用了航空领域中飞行员与副驾驶的交接理念：系统要么完全接管，要么完全交回控制权，绝不允许处于半睡半醒的中间模式。“要么是飞行员在控制，要么是副驾驶在控制。”（“My understanding is either the pilot is in control or the copilot is in control.”）通过定速巡航按键即可开启 Openpilot，而只要驾驶员踩下刹车、踩下油门（在默认逻辑下）或按下取消键，系统就会彻底退出，把车辆的横纵向控制权完整还给人。 ```mermaid flowchart TD accTitle: Openpilot 接管与退出控制逻辑 accDescr: 描述 Openpilot 与驾驶员之间清晰二元的控制状态转换 subgraph UserAction["驾驶员操作"] A["开启：按下定速巡航 (Cruise Control)"] B["退出：踩刹车 / 踩油门 / 按 Cancel"] end subgraph States["二元系统状态（避免模式混淆）"] OFF["未激活 (Disengaged)<br/>人类驾驶员完全掌控横向与纵向"] ON["完全激活 (Engaged)<br/>Openpilot 统一接管（或统一步调）"] end A --> ON ON --> B B --> OFF ``` --- ### [01:45:13](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6313s) 商业模式：高留存率与 Comma 硬件体系 谈及 Comma.ai 的商业模式与盈利现状，George 骄傲地宣布公司目前已经处于盈利状态。在定价上，Comma Two 设备售价为 1,000 美元，加上连接汽车的线束接口套件（200 美元），全套落地价格为 1,200 美元，并提供 30 天无条件全额退款保证。针对未来的硬件规划，George 透露硬件产品未来只可能变得更贵，如果推出下一代硬件（如 Comma Three），售价将远高于 1,200 美元。下一代设备与 Comma Two 之间的性能级差，就如同特斯拉 Autopilot 硬件一代（HW1）与二代（HW2）之间的跨越。 最让 George 自豪的不是销售数量，而是惊人的用户留存率。他指出，许多消费级前沿硬件（例如 VR 头显）平均被用户使用一次后就被扔进抽屉吃灰，但 Comma 完全相反。“Comma Two 的购买者中有 70% 是日活跃用户。”（“70% of Comma Two buyers are daily active users.”）在目前的活跃指标中，每日活跃用户（DAU）约为 2,000 人，周活跃用户（WAU）约 2,500 人，月活跃用户（MAU）已超过 3,000 人。对于一款单价上千美元的汽车后装硬件而言，这种极具黏性的日常使用率证明了产品真正创造了不可替代的日常出行价值。 - 产品定价：Comma Two 售价 1,000 美元，汽车接口 200 美元，整套 1,200 美元。 - 留存数据：70% 的设备购买者为 DAU（日活用户），DAU 约 2,000，MAU 超 3,000。 - 硬件迭代规划：未来高级硬件将与 Comma Two 形成高低搭配，定价显著提升。 --- ### [01:47:49](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6469s) Openpilot 做 Android，Autopilot 做 iOS：超越特斯拉的愿景 面对 Lex 抛出的终极问题——Comma 的目标是否是超越特斯拉，George 毫不犹豫地给出了肯定的答案。他将当前的自动驾驶竞争格局直接类比为智能手机早期的操作系统大战：特斯拉就像垂直封闭整合的 iOS，而 Openpilot 则是开放赋能所有汽车品牌的 Android。正如 Android 最终在整体装机量和用户规模上超越了 iOS 一样，Openpilot 也将在用户量上实现对特斯拉 Autopilot 的历史性反超。 对于开源软件如何变现的质疑，George 指出 Google 靠 Android 驱动搜索和生态变现，而 Comma 本质上是一家消费电子硬件公司（类似 Google 的 Pixel 手机业务）。只要未来哪怕只有 10% 的 Openpilot 用户购买 Comma.ai 官方出品的硬件，产生的利润就足以支撑一家极其庞大且健康的企业。George 算了一笔账：目前已售出约 2,500 到 3,000 台设备；预计明年出货 10,000 台，就能直接带来 1,000 万美元的硬件收入；而一旦规模攀升到 100,000 台，配合未来更高售价的硬件，年收入将轻松突破 2 亿美元。“我认为在某个时间节点，Openpilot 的用户量将会超过 Autopilot，就像 Android 超过 iOS 那样。”（“I think Openpilot at some point will cross over Autopilot in users, just like Android crossed over iOS.”）在绝大多数自动驾驶与辅助驾驶研发企业都在巨额烧钱的当下，Comma.ai 凭借硬件销售与开源软件生态的良性循环，成为了行业中极其罕见的自我造血、稳定盈利的异类。 - 竞争定位：Openpilot = 汽车辅助驾驶领域的 Android；Tesla Autopilot = 汽车领域的 iOS。 - 硬件营收推算： - 10,000 台/年（按 1,000 美元单价计算）= 1,000 万美元年收入。 - 100,000 台/年（单价提升至约 2,000 美元计算）= 2 亿美元年收入。 - 商业哲学：以卓越的消费级硬件品质正面竞争，即便仅占 Openpilot 总体硬件装机量的 10%，也能达成巨大的商业成功并维持盈利。 ### [01:50:28](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6628s) Android 对阵 iOS：Comma 与 Tesla 的终局格局 谈到神经网络模型的迭代，George Hotz 指出在 A/B 测试中，如果某个分支的损失在前 5 个 Epoch 内下降得非常迅速，它最终收敛出的模型往往会显著好于那些收敛缓慢的版本。人们总以为 Waymo 最终会在技术上碾压后发者，但 Hotz 坚决予以否认：“No, they're not.”（不，他们不会。） 当 Lex Fridman 询问未来是否可能出现特斯拉车主只需按下一个按键就能一键切换到 openpilot 的情况时，Hotz 认为这在技术生态上并不合理，并借用 Android 与 iOS 的生态格局进行了精准类比： ```mermaid flowchart LR accTitle: Comma.ai 与 Tesla 的生态模式类比 accDescr: 对比 openpilot 与 Android 的开放碎片化模式，以及 Autopilot 与 iOS 的垂直整合模式。 subgraph OpenEcosystem["开放 / 规模生态"] direction TB Android["Android / Google"] -->|"海量用户 & 多硬件适配"| Comma["Comma.ai / openpilot"] end subgraph ClosedEcosystem["封闭 / 高利润生态"] direction TB iOS["iOS / Apple"] -->|"垂直整合 & 单客高营收"| Tesla["Tesla / Autopilot"] end ``` Hotz 认为，Comma.ai 在最终用户总数上有可能超越 Tesla，但在营收规模上绝不可能战胜对方，因为特斯拉从每个用户身上榨取的单客收入要高得多。就像市场上 Android 设备的保有量远超 iPhone，但市场上卖出的 iPhone 数量却能轻松碾压 Google 官方的 Pixel 手机一样；未来售出的特斯拉汽车数量必然会远远超过 Comma 独立硬件的销量。至于特斯拉车主能否刷入 openpilot，Hotz 反问道：iPhone 理论上能刷 Android 吗？虽然硬要折腾确实办得到，但根本没有人会这么做，因为毫无意义。 - **用户规模 vs. 商业变现**：Comma 追求 Android 式的广泛覆盖与装机量，Tesla 则牢牢把持 iOS 式的高客单价与垂直控制权。 --- ### [01:51:46](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6706s) 巨头收购与“美联储代币”：Hotz 对传统车企的蔑视 Lex Fridman 抛出假想：如果福特、通用、丰田等老牌车企，或者亚马逊、Facebook、谷歌等科技巨头带着海量现金上门收购，George Hotz 是否会动心？Hotz 直言现在完全不予考虑，除非对方愿意满足他提出的各种苛刻条件陪他玩到底。谈及当下的并购（M&A）风气，Hotz 表现出极强的反感与不屑： > “The M&A people are so demeaning to companies. They treat the startup world as their innovation ecosystem, and they think that I'm cool with going along with that so I can have some of their scam fake Fed dollars, you know, Fedcoin.” —— George Hotz Hotz 讥讽并购团队把初创企业当成外部的“创新外包生态”，妄图用虚无缥缈的“美联储代币”（Fedcoin，即通胀美元）来收割核心技术。当被问及是否可能对传统厂商保有尊重时，Hotz 表示自己绝不可能尊重通用汽车（GM）；在传统厂商中，现代汽车（Hyundai）稍微让人觉得接近可敬，而日系（丰田）、德系和美系车企都沉浸在过去的辉煌中过于自大。在科技巨头中，苹果是唯一他还能稍微尊重的公司；至于 Facebook，他早在十年前就因鄙视其商业模式而选择离开，而谷歌在过去五年中也迅速退化。 --- ### [01:54:28](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=6868s) 拆解 Waymo：商业模式缺陷与 Level 4 的“滑板车陷阱” Lex Fridman 为 Waymo 进行了辩护。Fridman 表示自己曾多次拜访 Waymo，并在亚利桑那州钱德勒（Chandler）试乘过其无人驾驶汽车。Waymo 在全栈工程、软硬件自研、真实道路落地以及安全与创新的平衡上给他留下了极深印象，其车辆在设定路线上行驶果断，甚至能达到或超越限速，界面交互也非常出色。Fridman 据此认为 Waymo 绝非没有胜算。 然而 Hotz 一针见血地指出，Waymo 的致命伤根本不在技术或官僚作风，而在于其**产品定义本身就是错的**： | 对比维度 | 优质网约车（Uber） | Waymo 自动驾驶出租车 | |:--- |:--- |:--- | | **行驶效率与通行果断度** | 遇到无保护左转或行人过马路，司机能灵活果断借道或不完全刹停，速度快 20% | 遇到行人持续穿行便彻底卡死在停滞标志（Stop Sign）前，极度保守延误时间 | | **消费者权衡心理** | 仅约 15% 的乘客愿意为了省钱选择更慢的拼车（Uber Pool） | 试图用“无人驾驶成本更低”说服乘客，但乘客根本不愿用时间换微小的差价 | | **商业壁垒与行业本质** | 双边网络效应（双寡头竞合，如 Uber 与 Lyft） | 极易沦为重资产、低毛利的共享滑板车模式（Scooter Market），陷入技术投入与降价内卷 | Hotz 强调，Level 4 共享自动驾驶车要想实现盈利，必须建立在某一家公司形成绝对垄断的前提之上。一旦多家公司入局，这门生意就会彻底退化成类似共享滑板车的恶性价格战（Race to the bottom）。如今滑板车公司已不再宣传其具备严肃的公共交通实用价值，转而强调“好玩有趣”；在 Hotz 看来，从严肃出行的商业逻辑来看，“共享滑板车不可行，Level 4 自动驾驶汽车也同样不可行”。 --- ### [01:57:49](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=7069s) 假如执掌特斯拉：红外驾驶员监控与社交媒体哲学 面对 Fridman 提出的思想实验——“如果 Elon Musk 休假一年去荒岛隐居，董事会让你暂代特斯拉 CEO 一年，你会怎么做？”——Hotz 坦率承认自己其实“不会改动太多”（“I wouldn't change much”）。 > “If I actually were CEO of Tesla, first off, I'd be horrified that I wouldn't be able to do a good job as Elon.” —— George Hotz Hotz 认为特斯拉的大方向完全正确。尽管他曾公开与特斯拉争论“端到端学习与语义分割网络（SegNets）”的优劣，但他承认这纯属软件架构问题，现有的语义分割网络非常易于用来排查问题，未来平滑过渡到全端到端架构也并非难事。如果他真走马上任，唯一的硬件改动就是在 Model 3 和 Model Y 上**立刻加装红外摄像头和红外补光灯**，以实现更稳健的驾驶员面部注意力监测并采集闭环数据。 当话题转向社交媒体时，Fridman 好奇为什么极具黑客精神的 Hotz 在 Twitter 上相对沉寂，而在 Instagram 和直播间却十分活跃。Hotz 分享了他对主流社交网络价值观的深刻体悟： - **Instagram**：推崇具象的美感（“Instagram is a beautiful place. It glorifies beauty. I like Instagram's values as a network”）。 - **Twitter**：本质上是在挑唆矛盾、崇尚攻击与互撕（“Twitter glorifies conflict... taking shots at people”），这也正是 Twitter 与特朗普彼此完美适配的原因。因此他更倾向于远离 Twitter 的争吵旋涡。 ### [02:00:12](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=7212s) 拯救 Waymo 的极端方案：让安东尼·莱万多夫斯基出狱掌舵 当 Lex Fridman 抛出一个思想实验，询问如果让 George Hotz 接管在无人出租车（Robotaxi）和 L4 领域领跑的 Waymo 一年，他会如何大刀阔斧地改革时，George 的回答出人意料却斩钉截铁：他会把 Anthony Levandowski 从监狱里保释出来，直接任命他为 Waymo 的最高负责人。面对 Lex 关于这究竟是想彻底毁掉 Waymo 还是看中其抛弃官僚主义的第一性原理思维的追问，George 毫不讽刺地表示，Levandowski 绝对是一个不折不扣的天才，在处理 Waymo 这种局面上有远比自己更出色的构想与魄力。 Lex 坦言自己过去曾打算邀请 Levandowski 上播客，但他觉得自己过于轻信且天真，总倾向于看到人性最好的一面，后来才发觉有些人戴着虚伪的面具示人，行事充满欺骗。George 则拒绝从道德制高点做价值评判，而是回归商业与工程现实：Waymo 已经砸进去了整整 100 亿美元的无底洞，而 Comma.ai 仅仅融资了 810 万美元就已经实现了自我造血与盈利。作为一家成功的消费级硬件公司，Comma.ai 的逻辑根本无法填补 Waymo 高达百亿的亏损深坑。Waymo 要想活命，必须采取极端冒险的自救路径，因此 George 戏称自己如果当 CEO，唯一要做的事就是充当“过渡期 CEO”，全力帮助 Levandowski 解决牢狱之灾并把权杖交给他。 > “Whatever the idea is that's going to save Waymo, I don't have it, it's going to have to be a big risk idea, and I cannot think of a better person than Anthony Levandowski to do it.” —— George Hotz --- ### [02:03:40](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=7420s) 虚假的温情公关与硅谷创业精神的背叛 谈及对 Waymo 当前运营方式的批评，Lex 和 George 都对该公司过早陷入“做作与煽情模式”（corny mode）深感不满。Lex 认为，Waymo 本质上仍处于尚未交付实质商业闭环的创业阶段，应当像 SpaceX 那样把工程师和硬核极客推到前台，坦然直播火箭试飞甚至是爆炸失败的全过程，这种真诚反而能激发公众的信任与期待。George 对此强烈赞同，并尖锐讽刺了 Waymo 那些自诩为“你友好的汽车伙伴”的广告片，他直言广告里所谓的乘客完全是拿着通告单的受雇演员，这种虚假宣传只会招致反感。 两人认为，如今掌舵自动驾驶巨头的职业经理人缺少了真正的反叛精神。像 John Krafcik 或 Chris Urmson 这类管理者或许在传统车企或特定流程上各有所长，但他们并不是那种敢于打破常规的狂徒。在没有实际营收（pre-revenue）的情况下疯狂融资上百亿美元，却拿不出最小可行产品（MVP）、没有真实的用户增长指标，这完全背离了硅谷赖以立足的初衷。这些巨头似乎在取得商业成功之前，就已经患上了“大到不能倒”（too big to fail）的傲慢病。 > “You lie to me, I want you to fail; you tell me the truth, you'll be honest with me, I want you to succeed.” —— George Hotz - **Waymo 商业困境**：在未产生可规模化商业收入的前提下已烧掉约 100 亿美元，商业模式严重偏离精益创业轨道。 - **公关反差**：SpaceX 依靠坦承技术缺陷和爆炸现场赢得敬意，Waymo 却依靠雇佣演员拍摄温情广告掩饰工程落地的停滞。 - **领导层特质**：缺乏像马斯克或莱万多夫斯基式的“叛逆型（renegade）极客”，充斥着循规蹈矩的职业官僚。 --- ### [02:05:41](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=7541s) 自动驾驶改变社会？保持专注，不要傲慢面对大自然 当 Lex 畅想自动驾驶将如何重塑人类社会，甚至探索无人车是否会演进为能够与人类建立深厚情感联结的通用机器人时，George 给出了极为冷静务实的降温思考。在 George 眼中，自动驾驶是目前最酷的人工智能应用课题，因为它具备直接且清晰的商业变现路径；但从人类文明的历史尺度来看，它并非像互联网那样在前人认知之外凭空诞生的神迹——五十年前科幻作品中就早已充斥着对无人驾驶汽车的幻想，只有极少数真正的远见卓识者才在半个世纪前预见到了互联网。 对于 Lex 提出未来 Comma.ai 或特斯拉是否会借此底座全面进军广义机器人领域的设想，George 坚决主张在当下阶段必须绝对聚焦。他时刻告诫团队，无论 Comma.ai 在技术效率上看起来走得多顺，在没有彻底攻克 L5 级全自动驾驶之前，公司绝对算不上赢家。盲目扩张战线是对工程复杂性的轻视，创业者绝不能产生大而不能倒的幻觉。 > “You want to be arrogant in the face of other people, great; you want to be arrogant in the face of nature, you're an idiot.” —— George Hotz --- ### [02:07:53](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=7673s) 告别旧金山：一线城市的停滞与迁徙圣地亚哥 Lex 提到自己筹备成立新公司，原本在疫情前计划搬往旧金山，但如今也在考虑得克萨斯州的奥斯汀，或是二人当前录制播客所在的圣地亚哥。George 毫不犹豫地劝阻他前往旧金山，并阐述了 Comma.ai 搬迁至圣地亚哥的真实考量。虽然圣地亚哥拥有高通（Qualcomm）总部及其骁龙芯片生态，这对硬件开发确实是极佳的加分项，但这并不是搬迁的核心动力。 George 指出，团队核心成员依然热爱加州的气候、阳光与海岸线，加州的许多顽疾实际上是旧金山这座城市特有的体制病。以旧金山为代表的传统一线城市已经失去了对“增长”的渴望：城市的房产拥有者构成了主要的选民阶层，他们通过投票阻止新建住房，以此维持高房价让自己的资产升值。这种凭借选票自我敛财的政治寻租让整座城市陷入了不可救药的腐败与失衡之中，彻底失去了创新的活力土壤。 > “Once people have figured out how to vote themselves more money, they're going to do it.” —— George Hotz ### [02:09:29](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=7769s) 硅谷的虚伪、出逃潮与新城市的兴起 Lex Fridman 和 George Hotz 聊起了旧金山和硅谷令人窒息的社会环境。Lex 直言旧金山在政党倾向上完全是一党独大的城市，最讽刺的是，整座城市把“多元化”（diversity）整天挂在嘴边，却极度匮乏思想、背景、行事策略和创意上的真正多元性。他感叹道：“喊多元化声音最响的人，恰恰最缺乏多样性。”George 笑着附和：“这就是所谓的投射效应（projection）。” 两人提到，即便是硅谷的高层精英，私下里也普遍承认旧金山已经成了一个糟糕透顶的地方。而在新冠疫情期间，旧金山更是成了全美人口外逃最严重的城市。谈到理想的搬迁目的地： - **圣地亚哥**：George 仍然认为对于 Comma.ai 来说是一个极佳的落脚点。 - **奥斯汀**：Lex 觉得奥斯汀是一个非常有趣的生态系统，唯一的硬伤是夏天极端的高温。 - **科罗拉多**：两人都对科罗拉多这类治理得当的州给予了高度评价，反衬出加州（尤其是旧金山）那副自命不凡却千疮百孔的傲慢姿态。 --- ### [02:11:01](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=7861s) 创业的核心法则：技术必胜，谎言必败 Lex 顺势向 George 讨教创业心得，问他要打造一家成功的初创企业需要具备什么条件。George 一开始开启了自嘲模式，打趣说自己还没真正成功过，Lex 应该去请教那些真正赚大钱的人。两人甚至开玩笑说，将来他们可以把自己的经历写成 67 美元一本的畅销书，名字就叫《我是如何越狱 iPhone 的，你也可以》（*How I Jailbroke the iPhone and You Can Do It*）。 当 Lex 追问，如果两个人深夜喝着啤酒、自己深陷绝境与抑郁时，George 会给出怎样的建议？George 先是戏谑地回答“那就再来点啤酒”，随后神情严肃地给出了他最根本的准则。正如 George Hotz 所言：“更好的技术总是会赢，而撒谎总是会输。做好技术，不要撒谎。”（“The better technology always wins and lying always loses. Build technology and don't lie.”） Lex 指出，这在长期来看当然成立，但正如金融界的名言所说：“市场的非理性维持的时间，往往比你能保持不破产的时间还要长。”在资本涌动的 AI 和自动驾驶赛道上，许多公司靠着毫无实质的幻影和吹嘘（smoke and mirrors）就能轻松融到数亿美元。面对“为了生存是否该‘弄假直到成真’（fake it till you make it）”的疑问，George 表现得极为坚决。他痛批了因虚假宣传而锒铛入狱的 Fyre 音效节创始人 Billy McFarland，强调创业者绝对不能靠欺骗过日子，而是必须以诚实立足，并坚守技术终会兑现的信念。 ```mermaid flowchart TD accTitle: 创业宣传与技术真实性对比 accDescr: 对比正向炒作与虚假欺骗对长期结果的影响 subgraph Legitimate["正向推动（技术支撑）"] A["高调宣传 / 雄心目标"] --> B["背后有真实工程与实质产品支撑"] B --> C["交付时间延期，但最终攻克"] C --> D["长期胜利（如 SpaceX / Tesla / 优秀科技产品）"] end subgraph Fraudulent["欺骗骗局（虚妄空想）"] E["夸大概念 / 烟幕弹"] --> F["根本没有技术积累，靠虚假演示融资"] F --> G["违背物理与工程事实，深陷谎言"] G --> H["彻底崩盘（如 Fyre Festival / 概念骗局）"] end ``` --- ### [02:14:01](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=8041s) 吹嘘与过度承诺的区别：以 Elon Musk 为例 谈及炒作（hype），George 并不完全排斥。他回忆起自己年轻时为了买新款 iPhone 通宵在苹果店外排队的兴奋感。只要狂热的预期背后有能够真正买到的、扎实的产品作为支撑，炒作就是一件极富魅力的好事；只有当预期背后纯粹是一堆谎言时，它才令人作呕。 很多人指责 Elon Musk 是个骗子（fraud），但 George 和 Lex 并不赞同这一评价。他们剖析了程序员常犯的通病——预估工期过于乐观。Lex 指出，马斯克并非撒谎，而是习惯性地给出了一个过于激进、最终无法如期兑现的时间表： > “如果马斯克说‘我们一年后就会让火箭回收到地球’，结果花了四年——可他确实把火箭降落回了地球，而且他全程都在为此努力。”——George Hotz George 认为，只要你真诚地相信自己正在追求的目标，并且倾尽全力去攻克它，哪怕时间严重延期，这也是完全可以接受的；只有当你明明清楚自己绝对无法达成 deadline，却依然对外承诺时，那才是彻头彻尾的撒谎。在自动驾驶领域，虽然 George 常常批评特斯拉在纯视觉端到端路线上的一些误判，但他坦承马斯克在调动海量资源方面的能力远超自己，特斯拉依然有极大的概率最终赢得胜利。 - **炒作与交付的核心界线**： - **良性期望管理**：目标宏大、时间表不断延期，但每一次迭代都在靠近物理实现，本人对目标深信不疑。 - **恶性商业诈骗**：明知不可为而编造演示，靠资金链轮动维持虚假表象（以 Billy McFarland 式的邪教式狂热为典型）。 --- ### [02:16:51](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=8211s) 为什么人们总盼着成功者失败？Hacker News 评论区的阴暗心理 Lex 对当下社会的一种普遍心态感到极度困惑和沮丧：为什么那么多人不去庆祝他人的成功，反而迫切希望看到有雄心的人跌落神坛？不仅是马斯克，甚至在学术界、政界，许多人都在暗中诅咒开拓者失败。Lex 动情地表示，人类文明想要生存，客观目标函数理应是鼓励和庆祝那些取得突破的人，而不是把所有人拉回到泥潭里。 George 以极具洞察力的视角分析了这种阴暗心理，特别是科技论坛（如 Hacker News）上的网络键盘侠。首先，网络上的评论者绝不代表普通大众，他们只是极窄的一小撮人。其次，当一个惊世骇俗、雄心勃勃的初创项目出现时，那些老牌科技大厂员工之所以疯狂唱衰，本质上是为了捍卫他们自己的世界观： > “如果这东西成了，那我这 36 岁、在大科技公司干了一辈子的生活算什么？如果他们成功了，就意味着我本可以过上截然不同的人生；但我认定了我不能，所以他们必须失败。”——George Hotz 这种为了维护自身平庸选择的自尊防御机制，迫使这群人必须站在失败的看台上喝彩。George 奉劝创业者对此彻底充耳不闻。Lex 则透露，自己正考虑从产品和平台机制层面发起一场变革，通过新的技术形态去改变这种只允许少数刻薄之人发声的网络交流生态。 ### [02:19:42](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=8382s) 社交网络的毒性与商业模式困境 Lex Fridman 和 George Hotz 探讨了主流互联网平台的讨论环境恶化。谈及 Reddit 时，George 直言无法理解 Reddit 评论区严重的群体思维（groupthink）。Lex 指出，Reddit 存在子版块（subreddits）结构，某些小型子版块依然是充满善意、积极情绪并能进行深度微妙探讨的避风港；但 George 强调这属于用户自发涌现的现象，平台本身并未提供帮助。Lex 总结道：“某种程度上，如果不花巨大力气去鼓励微妙深刻与积极友善的氛围，互联网自然而然就会堕入混乱。” 针对 Twitter 的戾气，George 引用了认知科学家 Joscha Bach 的观点：Twitter 仅将“点赞”与“转推”设置为正向路径，用户若想表达任何负面看法，唯一的途径就是写评论反驳，这种机制的不对称性极大地放大了平台的毒性；相比之下，YouTube 评论虽然有赞和踩，但不展示踩的统计，体验好得多。Lex 认为这本质上是一个技术与机制设计问题，未来人类回头看今天的 Twitter 时，会觉得不可思议——“我们居然能容忍这种程度的戾气”。 两人进一步将矛头对准了基于广告的商业模式。George 提到纪录片《监视资本主义：智能陷阱》（The Social Dilemma）中的犀利洞察：传统观念认为“如果你不花钱，你就是产品”，但纪录片指出了更深层的本质——“被出售的产品，实际上是对你的影响力”（the product being sold is influence over you）。如果平台的核心逻辑是贩卖对用户的操控权，那么它就不可能消除毒性。Lex 探讨了是否能通过给用户更多透明度和控制权来解决，但 George 怀疑普通用户根本没有自省动力。Lex 承认自省的成本极高，但如果平台能从机制上奖励自省行为，结果或许会令人惊喜；而在当下的网络环境中，充斥着虚伪与冷漠，许多互动纯粹是为了操控他人，用户自身也沦为了一种广告形态。 --- ### [02:24:00](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=8640s) 极简编程环境与工具之争 Lex 将话题转向轻松的编程习惯，询问 George 心目中完美的编程装备。出乎意料的是，George 对硬件装备完全没有物欲崇拜，他表示：“给我一台坐在酒店房间角落里的 MacBook Air，我也能干活。”虽然在办公室他也会配置三块 24 英寸显示器、Happy Hacking Keyboard（HHKB）机械键盘以及 Razer DeathAdder（炼狱蝰蛇）鼠标，但他认为这些外设绝不会自动释放生产力，很多时候他在一台 MacBook Air 上产出反而最高。 在软件方面，George 坦言自己主要使用 70 年代诞生的 Vim，至今没有哪个现代工具真正实现了革命性超越；VS Code 仅适合阅读代码，而在引用跳转（xrefs）等功能上，逆向工程工具 IDA 表现得远比 VS Code 出色。这引发了 Lex 的共鸣与坦白： - **Lex 对 Emacs 的偏爱**：Lex 承认自己从没用过 Vim，一直坚持用 Emacs。他开玩笑说这可能源于某种恐惧——害怕发现自己过去的人生都被浪费了。他甚至将这种感觉与机器学习框架的迁移进行类比：“我感觉这就像我对 TensorFlow 与 PyTorch 的感受一样，刚刚把所有东西都切到了 PyTorch，花了好几个月，我感觉自己在 TensorFlow 上浪费了好几年，PyTorch 实在好太多了。” - **对 Lisp 的情结**：Lex 提到自己对 Lisp 语言情有独钟，或许是因为那是他接触的第一种函数式语言，也可能是因为在深度学习革命之前，很多经典 AI 课程都是以 Lisp 为教学核心。 - **硬件习惯差异**：Lex 表示自己很难脱离多屏环境，而且极度依赖 Kinesis 人体工学键盘，离开这些工具在笔记本上办公会让他感到非常痛苦。 --- ### [02:27:01](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=8821s) 彻底颠覆思维认知的三天 Lex 询问 George 是否有某种充满仪式感的高效工作日程。George 回避了时间表，反而回顾了他生命中宛如顿悟般、彻底改变认知模式的三个特定日子。他感慨道：“我愿意掏十万美元，如果明天能再过上哪怕其中一天的话。”这三天为他提供了理解技术、世界与社会的底层框架： ```mermaid flowchart LR accTitle: 三个颠覆思维认知的核心转折点 accDescr: 展示乔治·霍兹在人生中建立理解世界三大思维框架的关键时刻 subgraph Shift1["技术奇点与理性"] A["阅读 Eliezer Yudkowsky"] --> B["理解技术奇点与机器智能演进"] end subgraph Shift2["智能与压缩理论"] C["领会 Hutter Prize 与 AIXI"] --> D["顿悟：无损压缩等价于智能"] end subgraph Shift3["社会与政治现实"] E["阅读 Curtis Yarvin 博客"] --> F["获得解析现实政治的全新框架"] end ``` - **第一天：初识 Eliezer Yudkowsky** 读到 Yudkowsky 关于奇点理论（Singularity）的论述，思维受到了极大的震撼，建立了对超人类智能演进的全新预期。 - **第二天：洞悉胡特奖（Hutter Prize）与 AIXI** 他在 18、19 岁时就接触过相关理论但未读懂，直到那天真正领会了“无损压缩等价于智能”（lossless compression implies intelligence）以及 Marcus Hutter 提出的通用人工智能数学模型 AIXI，这一理论彻底奠定了他对现代人工智能本质的理解。 - **第三天：发现博客《Unqualified Reservations》** 在大约“玩家门”（Gamergate）和 2016 年美国大选前夕，他读到了 Curtis Yarvin（笔名 Mencius Moldbug）的新反动主义博客《Unqualified Reservations》。尽管 Yarvin 的观点极具争议、在外界看来亦疯亦天才，但该博客为 George 提供了除 AI 和技术之外的全新现实分析坐标系，让他突然觉得原本混乱的世界能够被逻辑性地解释通顺。 ### [02:29:05](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=8945s) 平凡日常、自发顿悟与科罗拉多暴风雪 乔治·霍茨（George Hotz）探讨了改变人生的那些灵感时刻究竟是如何降临的。莱克斯·弗里德曼（Lex Fridman）询问，日常生活中诸如吃早饭等琐碎细节是否存在某种迎接灵感的“魔法仪式”，抑或霍茨只是在特定时刻恰好做好了接收信息的准备。霍茨断然否定了任何日常结构的作用，坦言自己大部分时间只是随波逐流：“我漫无目的地度日，毫无章法，我随波逐流，希望并祈祷能再遇到像过去那些日子一样的一天。”（“I drift through life without structure. I drift through life hoping and praying that I will get another day like those days.”） 他回顾了自己过去几次漫无目的的外出经历：去纽约待了一个月，除了了解以太坊（Ethereum）算是个亮点之外，大部分时间糟透了；还有一次为期两周的公路自驾游，当他驱车行至科罗拉多州的甘尼森（Gunnison），打算翻越莫纳克山口（Monarch Pass）前往丹佛时，暴风雪铺天盖地而来。他亲眼目睹前方的福特 F-150 皮卡滑出路面，逼得他不得不掉头折返甘尼森留宿一宿。霍茨感慨道，自己生活中极少遇到这样真实发生的事件，哪怕只是因为大雪被逼掉头这种琐碎遭遇，那种直面物理世界现实的冲击感也显得格外珍贵而有意义。 --- ### [02:30:42](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=9042s) 尤德科夫斯基、指数增长与“人类过时论” 话题转回到霍茨早年受到的思想冲击，特别是埃利亚泽·尤德科夫斯基（Eliezer Yudkowsky）的早期著述。弗里德曼指出尤德科夫斯基如今主要因对人工通用智能（AGI）威胁论和 AI 安全（AI safety）的警示而闻名，但霍茨澄清，自己当年读到的是早期的作品，那时页面上根本没有任何“AI 安全”的字眼，甚至笑称尤德科夫斯基现在可能会谴责自己当年的文章，怪罪正是那类文字催生了今天他不愿看到的局面。 霍茨记忆中最深刻、最颠覆认知的核心观点，是计算机与人类之间算力增长速率的残酷对比： ```mermaid flowchart LR accTitle: Hotz's View on Exponential Computing vs Static Human Intelligence accDescr: Graph contrasting computer power growing exponentially every 18 months against flat human capability over time, leading directly to human obsolescence. subgraph Trends ["增长趋势对比"] A["计算机算力<br/>每 18 个月翻一番 (对数线性)"] B["人类智能与生理极限<br/>维持水平停滞"] end subgraph Consequence ["终局认知"] C["算力交叉点来临"] D["人类走向过时<br/>(在我有生之年被取代)"] end A --> C B --> C C --> D ``` 这彻底打开了霍茨的指数思考方式，让他意识到技术的非线性飞跃终将改变整个世界的权力结构： - 计算机算力每 18 个月翻倍一次，而人类的能力则维持水平不变。 - 尽管两者的能力曲线在当时尚未交叉，但在对数坐标图上，斜率恒正的直线必然会跨越水平线。 - 这一数学推论让他确信：“在我有生之年，人类注定会被取代。”（“In my lifetime, humans are going to be replaced.”） --- ### [02:32:22](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=9142s) 胡特奖、作为压缩的智能与无损等价性 紧接着尤德科夫斯基的思想，霍茨接触到了马库斯·胡特（Marcus Hutter）设立的胡特奖（Hutter Prize）。胡特的核心假说将通用人工智能的构建直接约化为数据压缩问题，认为宇宙中所有形式的智能本质上都可以视作某种压缩算法。霍茨对此深表赞同，表示自己以往一直在苦苦探寻“智能”的严格定义却不可得，而“无损压缩”（Lossless Compression）提供了一个清晰、毫无歧义的数学框架。 ```mermaid flowchart TD accTitle: Hutter Prize Equivalence Flow accDescr: Conceptual diagram showing how probabilistic next-character prediction connects to arithmetic coding, proving how intelligence equals lossless compression. subgraph Intelligence ["智能模型"] A["预测下一个字符 / 词元"] --> B["输出精准的概率分布"] end subgraph CompressionEngine ["编码与压缩"] B --> C["算术编码器<br/>(Arithmetic Encoder)"] C --> D["以最高效率编码噪声与熵"] end subgraph Result ["等价结论"] D --> E["无损压缩上限 = 通用智能水平"] end ``` 霍茨在了解胡特奖之后，随即投入了整整六个月时间尝试攻破这项竞赛。尽管在日常工程中该定义未必随时有用，但这开启了他的 AI 探索之路：随后他前往 Vicarious 工作，狂读深度学习论文，迅速补齐了对现代 AI 的全部理解。针对弗里德曼提出的疑问——人类智能和神经网络似乎都是模糊、有损的（lossy），为何胡特奖坚持要求无损（lossless）压缩？霍茨解释了信息论的实现机制： - 只要配合算术编码器（Arithmetic Encoder），任何有损预测器都能轻松转换为无损压缩器。 - 模型无需百分之百预知下一个字符究竟是 `e` 还是 `i`，它只需要为候选字符提供极其精准的概率分布。 - 算术编码器能根据概率分布以最大效率编码剩余的不确定性（噪声），压缩的极致即是对系统熵（entropy）的准确捕捉。 --- ### [02:35:06](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=9306s) GPT-3 的局限性与自动驾驶中的马尔可夫性 谈及 2020 年引发轰动的 GPT-3、大型语言模型与 Transformer 架构时，霍茨表现得相当冷静，直言其被严重过度炒作（overhyped）。他坚信仅凭扩大规模（如一路堆叠到 GPT-12）不可能自然涌现出真正的通用人工智能。 > “你的损失函数仅仅是对下一个字符做交叉熵损失，对吧？那根本就不是通用智能的损失函数。”（“Your loss function is literally just cross-entropy loss on the character, right? That's not the loss function of general intelligence.” — George Hotz） 霍茨阐述了 GPT 系列模型的几个硬伤与自动驾驶的本质区别： - **缺乏长期记忆**：自回归 Transformer 依赖硬编码的有限上下文窗口，哪怕做宽做大，也无法拥有真正的终生长期记忆，更不可能凭此造出一艘火箭飞船。 - **自动驾驶的马尔可夫属性（Markovian）**：弗里德曼追问是否能将视频 Transformer 用于彻底解决自动驾驶，霍茨给出了否定答复。他指出自动驾驶环境在本质上比语言具有更强的马尔可夫性质：绝大部分关于 $t$ 时刻的状态信息都已经包含在 $t-1$ 时刻之中，更早历史的信息权重大致呈指数级平滑衰减；而自然语言完全不同，读者往往需要回头检索“第二页第三段”的细节才能理解当前内容。 - **Comma.ai 的架构实践**：Comma.ai 目前使用视觉模型输出一个 1024 维的感知空间，后端序列模型采用 GRU（门控循环单元）。虽然未来完全可以把 GRU 替换为 Transformer 并测试效果，但霍茨强调这纯粹是普通的工程更替（就像过去从 RNN 换到 GRU 一样），根本谈不上任何颠覆性的技术革命。 ### [02:38:26](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=9506s) 通用人工智能（AGI）与技术奇点 Lex Fridman 将话题转向通用人工智能（AGI），询问 George Hotz 对雷·库兹韦尔（Ray Kurzweil）提出的“技术奇点”有何看法。Fridman 提到，Hotz 此前认为自动驾驶是个相对容易解决的问题，但他自己倾向于认为解决自动驾驶将对实现 AGI 产生深远影响。自动驾驶深度植根于人类经验之中，两者在底层是紧密相连的邻居。Fridman 追问 Hotz：未来是否会出现一个标志着断层式跃迁的“奇点”时刻，还是说人类其实已经身处奇点之中？ Hotz 斩钉截铁地回答：“我们绝对已经身处奇点之中了（We're definitely in the singularity now）。”他解释说，衡量奇点的核心指标在于人与人之间的通信带宽——如今整个世界都已高度网络化，网络传输速率不断飙升，全球正在形成紧密互联的单一文化（monoculture）。Fridman 提问这种分布式的信息传递是否代表着群体智能的增强，Hotz 澄清说自己并不是在评判这是好事还是坏事，但这就是奇点的现实定义。 对于那种“单一超级系统爆发式硬着陆（hard takeoff AGI）”的科幻式设想，Hotz 和 Fridman 都持否定态度。两人一致认为，真正的奇点并非源自某个孤独诞生的孤立 AGI 系统，而是源于全人类像蚁群一般通过高带宽网络协作所迸发出的集体智慧。Hotz 进一步指出，现代 AI 在本质上与人类一直在构建的各种技术工具并没有根本性区别。 ```mermaid flowchart TD accTitle: Hotz 对 AGI 与技术奇点的认知路径 accDescr: 展示通信带宽驱动奇点，以及自动驾驶作为 AGI 完整子集的验证逻辑 subgraph S1["奇点的本质"] A["高带宽通信网络"] --> B["全球互联与单一文化"] B --> C["人类蚁群式集体智能"] end subgraph S2["Comma.ai 的工程路径"] D["自动驾驶 (L5)"] -->|"作为完备子集"| E["通用智能验证"] E --> F["构建真实财务与工程价值"] end C -.->|"非单一孤立系统爆发"| E ``` - **关于自动驾驶与 AGI 的关系**：Hotz 认为自动驾驶是 AGI 的一个高度完备的子集。在 Comma.ai 研发的工具未来不仅能解决行车问题，更能直接迁移到 AGI 的探索中。 - **做自动驾驶的真实动机**：Hotz 坦白自己其实根本不在乎无人驾驶汽车本身，它只是一个能让他享受在技术上击败竞争对手快感的绝佳难题。 - **避免自我欺骗的工程信条**：研究 AGI 和 AI 极容易让人陷入自欺欺人的幻觉，误以为自己参透了某些奥秘；但如果能脚踏实地造出一辆 L5 级别的自动驾驶汽车，那就是无可争议的实质成果。Hotz 总结道：“它是不是通用人工智能？我不想争论这个。我只能说我们造出了具有巨大财务价值的东西（Is it general intelligence? I'm not going to debate that. I will say we've built something that provides huge financial value）。” - **后现代主义者的分工**：Fridman 将这种做法类比为埃隆·马斯克执着于“登陆火星”，只要确立一个极端清晰的目标并全力达成，过程中自然会攻克无数意想不到的难题。Hotz 赞同火星任务对人类文明的价值超越自动驾驶，但他笑称自己更适合处理信息流：“我是一个信息流派的人，我不是现代主义者，我是个后现代主义者（I'm an information guy. I'm not a modernist, I'm a postmodernist）。” --- ### [02:42:29](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=9749s) 程序员必学的四种编程范式与核心语言 Fridman 请 Hotz 为想要精进技能的开发者推荐 3 到 5 门必学的编程语言。Hotz 从计算机底层抽象机制出发，梳理出了一套从底层硬件到现代机器学习的完整知识梯队： ```mermaid flowchart LR accTitle: Hotz 推荐的四种编程范式与语言阶梯 accDescr: 从底层汇编到机器学习的层级与范式对照 subgraph P1["指令式抽象阶梯"] L1["汇编语言 (Assembly)"] --> L2["C 语言"] --> L3["Python"] end subgraph P2["其他核心范式"] L4["函数式: Haskell"] L5["硬件级并发: Verilog / HDL"] L6["数据驱动/ML: PyTorch"] end P1 -.->|"范式对比"| P2 ``` 1. **第一梯队：指令式语言的抽象堆叠（汇编、C、Python）** Hotz 强调每个人都必须学习**汇编语言（Assembly）**和 **C 语言**。他解释道：“如果你不会写汇编，你根本就不知道计算机在做什么（If you can't code in assembly, you don't know what the computer's doing）。”写汇编不需要达到大师水准，但必须亲自动手体会——唯有经历过汇编中受限的寄存器分配与图着色（Graph Coloring），才能真正理解编译器的价值；唯有写过 C 语言的内存管理，才能真正体会 Python 带来的强大生产力。第一版 C 编译器是在汇编中写就的，而 Python 则是用 C 编写的，这三者在抽象层级上构成了直观的垂直堆叠。 2. **第二范式：纯函数式编程（Haskell）** 跳出指令式逻辑后，必须接触重度依赖编程语言理论（PL theory）和依赖类型（Dependent Types）的领域。Hotz 推荐将 **Haskell** 从头到尾学透。虽然他自认不是顶级 Haskell 极客，但他曾用 Haskell 独立写过一个编译器，这种严格的形式化思维对拓宽视野至关重要。 3. **第三范式：硬件描述语言的全局并发（Verilog / HDL）** Hotz 指出，汇编语言虽然底层，但仍属于串行编译的思维框架。要想真正理解物理硬件是如何运转的，必须学习 **Verilog** 或 **VHDL**。在硬件描述语言中，一个代码块内部的所有指令是**同时瞬间执行**的，完全颠覆了软件开发者固有的单步顺序执行假设。 4. **第四范式：数据驱动编程（机器学习 / PyTorch）** 结合安德烈·卡帕斯（Andrej Karpathy）提出的“软件 2.0（Software 2.0）”，Hotz 认为这代表了一种崭新的编程范式——**数据驱动编程（Data-driven Programming）**。针对开发框架的选择，Hotz 毫不犹豫地给出明确建议：“不要学 TensorFlow，去学 PyTorch（Don't learn TensorFlow, learn PyTorch）。”PyTorch 凭借优秀的动态计算图设计赢得了这场范式之争。 --- ### [02:46:11](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=9971s) 给编程初学者的建议：在真实项目中死磕 Fridman 提到 Hotz 曾发布过一个名为《什么是编程：新手教程！》（*What is programming, noobs!*）的视频，而该视频最热门的评论却是调侃“警告：这视频根本不是给新手看的”。Fridman 顺势向 Hotz 请教，对于毫无基础的新人，究竟该如何入门编程。 Hotz 尖锐地打破了各种教学视频的虚妄承诺，直言：“你永远不可能通过看一个叫‘学习编程’的视频学会编程（You are never going to learn programming by watching a video called learn programming）。”他坚信，全世界所有顶尖程序员的成长路径完全相同： - **从具体渴望出发**：学习从来不是因为“想学编程”这一空泛概念，而是源于心中有一个明确想做出的东西，或者是想让计算机帮自己自动化实现某项具体任务。 - **在解决问题中推进**：动手尝试，遭遇阻碍，不断查阅摸索，在项目遇到死胡同的过程中一点点撬动知识网络，硬生生把项目推过终点。 - **唯一的有效建议**：学编程的唯一方法就是**立刻去动手写程序（Go program）**。 Hotz 举例反讽了那种脱离实际需求的学习状态：有人曾给他写信咨询循环神经网络（RNN）和时间序列数据，信中声称自己的公司打算在时间序列上引入 RNN，但目前其实根本还没想好具体应用场景，只是想先提前了解技术。在 Hotz 看来，缺乏实际业务与任务痛点的所谓“技术预研”，往往只是空耗精力的自我感动。 ### [02:47:28](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=10048s) 解决真问题与学会善用 Google 话题从如何学习机器学习等宽泛技术展开。George Hotz 指出，学习任何技术的真正秘诀在于先有一个明确想解决的具体问题。每当有人提出那种假大空的问题时，他都感到十分抓狂：“我听说机器学习很重要，你能帮我们把机器学习整合到通心粉和奶酪的生产中去吗？”George 毫不客气地吐槽：“我甚至不知道怎么帮这些人，到底是谁允许这种人去负责任何事情的？” Lex Fridman 试图缓和气氛，认为每个人都有刚入门当新手的时候。但 George 强调，问题不在于对方是新手、不懂机器学习，而在于那种荒谬的思维模式——盲目觉得新技术一定能套用在任何风马牛不相及的事情上。Lex 赞同解决问题需要前置条件：第一，你得真的在通心粉生产中遇到了故障，比如某些盒装产品没装够分量；第二，从传统方案入手，再去看看机器学习是否真的适用于这一场景。 由此 Lex 分享了自己对求学者的普遍建议：学会“在 Google 上搜索”（Google it）。在 Lex 看来，优秀的工程师和研究员需要两种互补的核心能力： 1. 一张白纸、完全断网状态下的深层独立思考能力； 2. 彻底深挖 Google 的信息检索能力。 检索是一门经常被低估的硬本领：顺着线索追踪术语、深入 GitHub 上只有两颗星的冷门仓库去阅读源码、混入 Twitter 上的尖端讨论圈子，甚至跨界寻找灵感——神经科学家应该去翻阅信号处理社区，AI 研究员应该去读心理学文献。Lex 指出，与其闭门造车从零折腾，不如去找到已经花多年时间把该问题解决掉的先驱。George 认为这两者其实殊途同归：“你去尝试解决问题，而在尝试解决问题的过程中，如果你擅长解决问题，你自然会撞上那个已经解决它的人。”当面对“我该如何入门深度学习”这类宽泛提问时，George 坦言：“相信我，Google 上排第一的答案，绝对比我能告诉你的任何话都要好得多。” ```mermaid flowchart TD accTitle: 学习新技术的有效闭环与错误路径 accDescr: 对比先找具体问题并通过深度检索来解决，与拿着流行技术盲目套用现实场景这两种模式 subgraph WrongPath ["常见误区"] W1["追逐热词：'我想把 ML 整合进业务'"] --> W2["无目标的虚假需求<br/>（如空谈通心粉生产）"] end subgraph RightPath ["正确闭环"] R1["发现具体真实问题<br/>（如包装盒未装满）"] --> R2["第一原则深度思考<br/>（断网/白纸演算）"] R2 --> R3["极客式检索能力<br/>（GitHub 冷门代码 / 跨领域社区）"] R3 --> R4["吸收先驱现成成果，完成闭环"] end ``` ### [02:51:34](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=10294s) 塑造认知的书籍推荐：《无尽的玩笑》与《阿特拉斯耸肩》 Lex 询问 George 哪三本技术、虚构或哲学书籍对他的人生产生了重大影响。George 首先推荐了相对最不具争议的一本——大卫·福斯特·华莱士（David Foster Wallace）的《无尽的玩笑》（*Infinite Jest*）。George 评价它是一部关于“神经刺激享乐”（wireheading）的书，不仅文笔极佳、阅读体验极度愉悦，而且“读这本书会让你作为一个人类获得成长”，尽管阅读它需要付出巨大的智力心力。 随后，George 语出惊人地拿出了第二本推荐：“第二本书就是色情读物，它叫《阿特拉斯耸肩》（*Atlas Shrugged*）。”面对 Lex 的惊讶，George 直白表示自己不会为这本书的文学价值辩护，它绝非一本写法精妙的作品，但阅读起来确实极具娱乐性，“就像色情片一样，制作水准不高”。他提到安·兰德（Ayn Rand）的书里有一段长达 60 页的单白独白，当时编辑强烈要求删掉，但兰德甚至自掏腰包也要把这段长篇大论留在书里。 尽管如此，George 强调该书为理解人类社会关系提供了一套坚固的框架：“我知道很多人会说‘对，但那是个糟糕透顶的框架’——可它毕竟是一个框架。”Lex 也提到自己过几天将与在世的核心客观主义哲学家 Yaron Brook 进行四小时以上的长谈，Lex 发现客观主义哲学非常耐人寻味：很大一部分人对其极度反感甚至生理厌恶，但兰德哲学中将资本主义与“道德的生活就是富有生产力的生活”紧密绑定的观点，本身具有独特的吸引力。 George 回忆自己 19 岁阅读该书时的震撼体验。书中反派角色的口号是“各尽所能，按需分配”，George 第一次读到时觉得这简直是“火箭队级别（Team Rocket）的卡通反派脸谱”，世界上怎么可能有这么弱智的反派信条；然而当他后来得知这竟然是真实的共产主义口号时，整个人目瞪口呆。Lex 与 George 一致认为，兰德笔下的世界善恶过于二元对立，正如色情片没有角色成长与心理纵深一样，《阿特拉斯耸肩》也没有，但这丝毫不影响它在 19 岁的 George 心中掀起风暴。 - 大卫·福斯特·华莱士的《无尽的玩笑》（*Infinite Jest*）：深入探讨极致成瘾、wireheading 与现代人精神困境的巨著。 - 安·兰德的《阿特拉斯耸肩》（*Atlas Shrugged*）：客观主义哲学代表作，包含 60 页著名哲学独白，提倡理性利己主义与生产创造的道德价值。 ### [02:55:16](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=10516s) 硬核科幻与奇点推演：《排列城市》与《主智的蜕变》 在第三本书的名额上，George 推荐了两部硬核科幻作品： 1. 澳大利亚科幻作家格雷格·伊根（Greg Egan）的《排列城市》（*Permutation City*）：George 称其为思考意识上传、数字复制品与虚拟现实本体论的必读之作。 2. 罗杰·威廉姆斯（Roger Williams）发表在网络上的小说《主智的蜕变》（*The Metamorphosis of Prime Intellect*）：一部描绘“技术奇点之后”（post-singularity）的世界的震撼之作。 George 深入分析了《主智的蜕变》，指出这部作品带有极其鲜明的“工程师写作”印记，用近乎实用主义的冷峻视角去构建一个绝对乌托邦。在这个世界里，超级人工智能控制了一切物理法则，人类获得了无限的寿命与随心所欲的力量，然而这种终极极乐却引发了深刻的虚无与异化。George 强调该书的结局耐人寻味，至于那是积极还是消极，“这完全取决于读者在阅读过程中的自我判断”。他在 15 岁时第一次读到这部仅约 50 页的短篇巨作，随后一生中反复重读了多次，强烈建议所有人去读。 在随后的科幻闲聊中，Lex 提到自己近期正在补课经典科幻，正在读阿西莫夫的《基地》（*Foundation*）系列，想知道 George 心目中最伟大的科幻小说是什么。George 直言自己并不喜欢《基地》，觉得它“充斥着过时的现代主义说教”。在被问及其他作品时，George 表示自己还没读过《沙丘》（*Dune*），但觉得弗诺·文奇（Vernor Vinge）的《深渊上的火》（*A Fire Upon the Deep*）很有趣。最后 George 总结，在赛博朋克与硬科幻领域，每个人都必须去读一读尼尔·斯蒂芬森（Neal Stephenson）的《雪崩》（*Snow Crash*）以及威廉·吉布森（William Gibson）的《神经漫游者》（*Neuromancer*）。 ### [02:57:12](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=10632s) 书籍推荐与给年轻人的建议 对话首先承接上一部分关于阅读书单的讨论。George Hotz 提到，如果还没读过那些基础经典，应该先从那里读起。Lex Fridman 表示自己还没读过《雪崩》（Snow Crash），Hotz 则力荐道格拉斯·霍夫施塔特（Douglas Hofstadter）的《哥德尔、艾舍尔、巴赫》（Gödel, Escher, Bach: An Eternal Golden Braid），并补充说如果想要看一本具有争议性的书，可以去读《青铜时代心态》（Bronze Age Mindset）。随后，Fridman 顺着书籍话题转入了社交媒体上的征集提问：很多 Twitter 和 Reddit 网友想知道，以 Hotz 如此年轻就完成诸多惊人成就的履历，他会对今天的年轻人给出怎样的生活建议？ 面对这个宏大的问题，Hotz 带着讽刺与幽默的态度表示，如果大学以后重新开学，他很乐意去发表一场毕业演讲，到那时他会投入大量精力去撰写一份充满讽刺意味的讲稿。他随口调侃起那些泛滥的名人演讲套话：“永远涂防晒霜”（涂防晒霜演讲）、“这就是水”（大卫·福斯特·华莱士的经典毕业演讲），或者乔丹·彼得森的“打扫你的房间”。在 Hotz 看来，市面上所有的自助类书籍本质上都是同一套东西在互相抄袭。他尖锐地指出，这类建议根本无法解决实际问题：“自助类书籍不是为了帮助你而设计的，它们是为了让你感觉良好而设计的。”（“Self-help books aren't designed to help you, they're designed to make you feel good.”） Hotz 进一步剖析道，任何别人能给出的生活建议，当事人其实早就心知肚明。“要好好吃饭”、“多读点书”，这些话说出来毫无意义，提问者听了也不会发生任何改变。如果非要说有什么建议，那就是别在脑子里把这些东西想得太复杂，不要作茧自缚。Fridman 评价称，Hotz 本人的整个人生和行动本身就是对这种思维抗拒的活生生例证——他完全不纠结、不过度思考。Hotz 对此回应，自己的逻辑仅仅适用于他自己： > “即使我谈论我的建议，我也是说，我的建议只对我自己有用，对其他人毫无意义。”（“Even when I talk about my advice, I'm like, my advice is only relevant to me. It's not relevant to anybody else.”） 如果一个人本身就习惯于反复过度思考，强行劝对方停下来也毫无用处。每个人必须找到适合自己的运作方式。 --- ### [02:59:57](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=10797s) 爱情的三种定义与道德生产力 Fridman 提起两人上次对话曾讨论过“生命的意义在于赢”，但鲜少探讨“爱”（Love）这个命题，无论是浪漫之爱还是对全人类的博爱，并询问在 Hotz 追逐胜利的人生中，爱究竟扮演了怎样的角色。Hotz 将“爱”拆解为三种截然不同的生理与心理机制： 1. **阿片类药物式的爱（Opiates）**：带来麻醉般的舒适感与被动依赖； 2. **催产素式的爱（Oxytocin）**：建立人际联结与依附关系的生物学驱动； 3. **抽象维度的爱**：比如对数学（Math）纯粹求知、探索的热爱，这完全脱离了前两者的范畴。 对于前两种以生物化学为驱动的情感，Hotz 直言自己并不感冒。他解释称，这与自己不碰阿片类药物和摇头丸（Ecstasy/MDMA）是同一个原因。尽管他曾经尝试过两者（并且相比摇头丸更偏好阿片类），但他拒绝沉湎其中。他给出了一个极其理性主义的伦理法则： > “有道德的生活就是富有生产力的生活。”（“The ethical life is the productive life.”） Hotz 认为，消耗在纯粹感官刺激或沉溺型依恋中的状态无法带来生产力。而谈到第三种——对人类这一物种抽象的爱，Hotz 表示这种情感对自己来说就像底层代码或系统背景音一样，一直自然存在。他甚至很难想象一个人心中如果完全没有对人类的抽象之爱会是什么状态。 --- ### [03:01:46](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=10906s) 迷幻药、认知边界与学习率 话题顺理成章地延展至药物与致幻体验。Fridman 提到自己最近在与几位研究迷幻药（Psychedelics）的学者交流，并且自己尝试吃过几次致幻蘑菇（Shrooms）。他发现迷幻药能将大脑带到一个前所未见的奇异境地，体验异常友好、积极、充满笑意与狂喜，这让他好奇心灵究竟能去往何处，体验归来后是否有益。 Hotz 提出了一个独到的批判视角：社会上唯独赋予了迷幻药“教导人生哲理”的崇高地位，却几乎没有人会说苯丙胺类（Amphetamines）或兴奋剂有东西能教给人类。在 Hotz 看来，只要保持清醒的认知，任何药物都能带来观察自我的样本：阿片类、苯丙胺、大麻或迷幻药，它们都有独特的视角，但关键在于不要搞错认知的客体： > “但我同时也认识到，我认为你并不是在了解这个世界，你是在了解你自己。”（“...recognize that I don't think you're learning things about the world, I think you're learning things about yourself.”） Hotz 引用了类似蒂莫西·利里（Timothy Leary）的名言指出：每个人都应该推开那扇门看一眼门后到底有什么，但一旦你看过了，就完全没有必要一次又一次地反复回去。除日常饮用的咖啡外，Hotz 大约每年只重新接触一次迷幻药，且通常使用极小的微剂量（microdose）。当 Fridman 提出迷幻药或许能“调高大脑的学习率（learning rate）”时，Hotz 眼前一亮，用机器学习的隐喻风趣地赞同道：学习率调大确实能加速收敛，但学习率过大同样伴随着极高发散与系统崩溃的风险（pros and cons）。 --- ### [03:04:07](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=11047s) 疯狂的定义与拒绝被驯化的神经多样性 访谈进入尾声，Fridman 问起 Hotz 曾戏称自己“疯狂”（Crazy），如果要给自己的疯狂程度打分（1 到 10 分），会打几分。Hotz 笑着回应，这取决于对标的对象：“如果你拿我和埃隆·马斯克（Elon Musk）以及安东尼·莱万多夫斯基（Anthony Levandowski）比，我其实没那么疯。”经过在 6 分和 7 分之间的权衡，Hotz 最终将自己定在 6 到 7 分的区间。 ```mermaid flowchart LR accTitle: 神经多样性与精神疾病界定谱系 accDescr: 展示 George Hotz 对人类心理状态从病理极端到正常个性谱系的划分 subgraph Extreme["器质性极端疾病"] A["双相情感障碍<br/>(Bipolar)"] B["精神分裂症<br/>(Schizophrenia)"] C["明确的视觉幻觉"] end subgraph Neurodiversity["广泛的神经多样性"] D["注意力缺陷 (ADD)"] E["对立违抗性障碍 (ODD)"] F["不迎合社会规则的特异个体"] end Extreme ---|"病理分界线"| Neurodiversity ``` 谈及“疯狂”究竟是大脑的缺陷（Bug）还是特性（Feature），Hotz 直言自己并不赞同现代主流心理学对“心理健康”所贴的很多标签： - **器质性精神疾病**：真正的重度双相情感障碍（Bipolar）、精神分裂症（Schizophrenia）或者肉眼可见的幻觉，确实是生物学层面的器质性病变； - **过度医疗化的行为谱系**：像注意力缺陷（ADD）或对立违抗性障碍（ODD），在 Hotz 看来根本不属于疾病，而只是人类丰富神经多样性（Neurodiversity）中的一种正常频谱分布。 Hotz 吐槽说，自己在做编程直播时，弹幕常常刷屏怀疑他磕了药。但他澄清自己除了早上一杯咖啡外什么都没碰，“你们看到的仅仅是我大脑高速运转的原生状态”。Fridman 用圣诞树上的装饰物打比方，形容 Hotz 就是那颗最亮眼、最不愿平庸无奇的闪耀饰品（sparkle）。 Hotz 总结说，这种特立独行在某种程度上是自我构建的选择： > “我是我选择成为的人。我会尽可能真实地说出我所看到的事情，我绝不说谎。”（“I am who I choose to be. I'm going to say things as true as I can see them, I'm not gonna lie.”） 在节目的最后几秒，Hotz 用一句耐人寻味的格言为整场对话收尾：“追求正确是至关重要的，除非以犯错为代价。”（“Being right is super important, except at the expense of being wrong.”）Fridman 赞叹 Hotz 是世界上最独特的人类之一，采访在思辨的余韵中落下帷幕。 ### [03:07:42](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=11262s) 节目尾声与赞助商致谢 长达三个多小时的深度对话接近尾声，Lex Fridman 向 George Hotz 表达了由衷的敬意与感谢，称能与他交谈深感荣幸。George Hotz 随即回应致谢。随后，Lex Fridman 转向对听众的结语，并逐一致谢了本期播客的赞助商：包括制作美味蘑菇咖啡的 Four Sigmatic、他平时常听且喜爱的科技播客 Decoding Digital，以及他多年来一直在使用的虚拟专用网络服务 ExpressVPN。Lex 提醒大家可以通过视频简介中的链接获取专属折扣，以此来支持节目的持续制作。 Lex 接着向观众发出了常规的互动邀请，希望喜欢节目的听众能够在 YouTube 上订阅频道，在 Apple Podcasts 上留下五星好评，在 Spotify 上关注，在 Patreon 上提供赞助支持，或者在 Twitter（@lexfridman）上与他取得联系。 - **赞助商支持**：Four Sigmatic（蘑菇咖啡）、Decoding Digital（科技播客）以及 ExpressVPN（VPN 服务）。 - **互动平台**：YouTube 订阅、Apple Podcasts 评分、Spotify 关注、Patreon 支持以及 Twitter 社交连接。 ### [03:08:24](https://www.youtube.com/watch?v=_L3gNaAVjQ4&t=11304s) 结语：林纳斯·托瓦兹的名言 在正式结束这期围绕黑客精神、自动驾驶与宇宙模拟的马拉松式探讨前，Lex Fridman 选择以开源世界最具标志性的人物之一——Linus Torvalds（林纳斯·托瓦兹）的名言作为全篇的句号。这句话精准契合了 George Hotz 崇尚实干、用代码说话的技术黑客作风。Lex 引用道：“Talk is cheap. Show me the code.” 随后，他感谢所有人的收听，并期待在下一期节目中与听众再次相遇。 ## Video description George Hotz (geohot) is a programmer, hacker, and the founder of Comma.ai. Please support this podcast by checking out our sponsors: - Four Sigmatic: https://foursigmatic.com/lex and use code LexPod to get up to 40% & free shipping - Decoding Digital: https://appdirect.com/decoding-digital - ExpressVPN: https://expressvpn.com/lexpod and use code LexPod to get 3 months free EPISODE LINKS: Comma.ai's Twitter: https://twitter.com/comma_ai Comma.ai's Website: https://comma.ai/ George's Instagram: https://www.instagram.com/georgehotz George's Twitch: https://www.twitch.tv/georgehotz George's Twitter: https://twitter.com/realgeorgehotz Comma.ai YouTube (unofficial): https://www.youtube.com/channel/UCwgKmJM4ZJQRJ-U5NjvR2dg PODCAST INFO: Podcast website: https://lexfridman.com/podcast Apple Podcasts: https://apple.co/2lwqZIr Spotify: https://open.spotify.com/episode/1LUuv5NInzIUf3I4sCzZOk RSS: https://lexfridman.com/feed/podcast/ Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4 Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41 OUTLINE: 0:00 - Introduction 2:31 - Will human civilization destroy itself? 5:18 - Where are the aliens? 10:05 - Tic Tac UFO and Bob Lazar 12:33 - Conspiracy theories 14:36 - The programming language of life 18:57 - The games that humans play 27:27 - Memory leaks in the simulation 29:58 - Theories of everything 31:43 - Ethereum startup story 39:30 - Cryptocurrency 48:57 - Self-help advice 52:37 - Comma.ai 54:30 - Comma two 1:03:19 - Tesla vs Comma.ai 1:12:22 - Driver monitoring 1:26:03 - Communicating uncertainty 1:27:51 - Tesla Dojo 1:34:19 - Tesla Autopilot big rewrite 1:40:37 - How to install the Comma Two 1:45:13 - Openpilot is Android & Autopilot is iOS 1:54:28 - Waymo 2:05:41 - Autonomous driving and society 2:07:53 - Moving 2:10:58 - Advice to Startups 2:24:00 - Programming setup 2:27:01 - Ideas that changed my life 2:35:06 - GPT-3 2:38:26 - AGI 2:42:29 - Programming languages that everyone should learn 2:49:02 - How to learn anything 2:51:34 - Book recommendations 2:59:57 - Love 3:01:46 - Psychedelics 3:04:07 - Crazy CONNECT: - Subscribe to this YouTube channel - Twitter: https://twitter.com/lexfridman - LinkedIn: https://www.linkedin.com/in/lexfridman - Facebook: https://www.facebook.com/LexFridmanPage - Instagram: https://www.instagram.com/lexfridman - Medium: https://medium.com/@lexfridman - Support on Patreon: https://www.patreon.com/lexfridman [Source](https://www.youtube.com/watch?v=_L3gNaAVjQ4)

[130:58] George Hotz: The better technology always wins and lying always loses. Build technology and don't lie. [59:24] George Hotz: In supervised learning the weights depend on the data... in reinforcement learning the data depends on the weights. [66:19] George Hotz: End-to-end is going to be the only solution for level five... and of course Tesla's gonna come around to my way. [39:30] George Hotz: The Nakamoto consensus algorithm is, I think, one of the greatest innovations of the 21st century. [14:36] George Hotz: DNA is a CAD model. It's not quite a programming language it has no sort of serial execution it's not quite a yeah it's a CAD model. [52:45] George Hotz: Solve self-driving cars while delivering shippable intermediaries. [53:00] George Hotz: You can set any arbitrary milestone, meet that milestone, and still be infinitely far away from solving self-driving cars. [162:29] George Hotz: If you can't code in assembly, you don't know what the computer's doing.

[0:00] the following is a conversation with [0:01] george hotz a.k.a [0:03] geohot his second time on the podcast [0:06] he's the founder of comma ai an [0:09] autonomous and semi-autonomous vehicle [0:11] technology company [0:12] that seeks to be to tesla autopilot what [0:16] android is [0:17] to the ios they sell the [0:20] comma two device for one thousand [0:21] dollars that when installed [0:23] in many of their supported cars can keep [0:25] the vehicle centered in the lane [0:28] even when there are no lane markings it [0:30] includes [0:31] driver sensing that ensures that the [0:33] driver's eyes are on the road [0:35] as you may know i'm a big fan of driver [0:37] sensing i do believe tesla autopilot and [0:40] others [0:40] should definitely include it in their [0:42] sensor suite also [0:44] i'm a fan of android and a big fan of [0:46] george for many reasons [0:48] including his non-linear out of the box [0:50] brilliance [0:51] and the fact that he's a superstar [0:54] programmer [0:55] of a very different style than myself [0:57] styles make [0:58] fights and styles make conversations so [1:01] i [1:01] really enjoyed this chat i'm sure we'll [1:03] talk many more times on this podcast [1:06] quick mention of each sponsor followed [1:07] by some thoughts related to the episode [1:10] first is four sigmatic the maker of [1:12] delicious [1:13] mushroom coffee second is the coding [1:16] digital [1:17] a podcast on tech and entrepreneurship [1:19] that i listen to [1:20] and enjoy and finally expressvpn [1:24] the vpn i've used for many years to [1:26] protect my privacy on the internet [1:29] please check out the sponsors in the [1:30] description to get a discount [1:32] and to support this podcast as a side [1:35] note [1:35] let me say that my work at mit on [1:38] autonomous and semiautonomous vehicles [1:40] led me to study the human side of [1:42] autonomy enough [1:43] to understand that it's a beautifully [1:45] complicated and interesting problem [1:47] space [1:48] much richer than what can be studied in [1:50] the lab [1:51] in that sense the data that comma ai [1:54] tesla autopilot [1:55] and perhaps others like cadillac super [1:57] crews are collecting [1:58] gives us a chance to understand how we [2:00] can design safe [2:02] semiautonomous vehicles for real human [2:04] beings [2:05] in real world conditions i think this [2:08] requires bold innovation [2:09] and a serious exploration of the first [2:12] principles of the driving task [2:14] itself if you enjoy this thing subscribe [2:17] on youtube [2:17] review it with five stars and up a [2:19] podcast follow on spotify [2:21] support on patreon or connect with me on [2:23] twitter at lex friedman [2:26] and now here's my conversation with [2:28] george [2:29] hotz so last time we started talking [2:32] about the simulation [2:34] this time let me ask you do you think [2:35] there's intelligent life out there in [2:37] the universe [2:38] i've always maintained my answer to the [2:40] fermi paradox i think there [2:42] has been intelligent life elsewhere in [2:44] the universe [2:45] so the intelligent civilizations existed [2:47] but they've blown themselves up so your [2:49] general intuition is that intelligent [2:52] civilizations quickly like there's that [2:54] parameter [2:55] in in the drake equation your senses [2:58] they don't last very long [2:59] yeah how are we doing on that like have [3:01] we lasted pretty [3:02] pretty good i don't know we do oh yeah [3:06] i mean not quite yet [3:09] well what telly has your caskey the iq [3:11] required to destroy the world falls by [3:13] one point every year [3:15] okay so technology democratizes [3:18] the destruction of the world when can a [3:21] meme destroy the world [3:25] it kind of is already right somewhat [3:28] i don't think i don't think we've seen [3:30] anywhere near the worst of it yet [3:32] world's going to get weird well maybe a [3:34] mu can save the world [3:36] you thought about that the meme lord [3:38] elon musk fighting on the side of good [3:40] versus the uh the meme lord of the [3:43] darkness which is uh not saying anything [3:46] bad about donald trump [3:48] but he is the the lord of the meme on [3:50] the dark side he's a darth vader of [3:52] memes [3:53] i think in every fairy tale [3:57] they always end it with and they lived [3:58] happily ever after and i'm like please [4:00] tell me more about this happily ever [4:01] after i've heard [4:03] 50 percent of marriages end in divorce [4:05] uh why doesn't your marriage end up [4:07] there you can't just say happily ever [4:08] after so [4:10] it's the thing about destruction is it's [4:12] over after the destruction [4:14] um we have to do everything right in [4:16] order to avoid it [4:18] and uh one thing wrong i mean actually [4:21] that's what i really like about [4:22] cryptography [4:22] cryptography it seems like we live in a [4:24] world where the defense wins [4:27] um versus like nuclear weapons the [4:29] opposite is true [4:30] it is much easier to build a warhead [4:32] that splits into 100 little warheads [4:34] than to build something that can [4:35] you know take out 100 little warheads uh [4:38] the offense has the advantage there [4:40] um so maybe our future is in crypto but [4:44] uh so cryptography right the goliath is [4:47] the [4:47] the defense and then all the different [4:50] hackers [4:51] are the uh are the davids and that [4:54] equation is flipped for nuclear war [4:57] because there's so many like one nuclear [4:59] weapon destroys everything [5:01] essentially yeah and it is much easier [5:03] to uh [5:05] attack with a nuclear weapon than it is [5:07] to like the technology required to [5:09] intercept and destroy [5:10] a rocket is much more complicated than [5:12] the technology required to just you know [5:13] orbital trajectory send a rocket to [5:15] somebody [5:17] so okay your intuition that the there [5:19] were intelligent [5:21] civilizations out there but it's very [5:23] possible that they're [5:24] no longer there that's kind of a sad [5:26] picture they [5:27] enter some steady state they all [5:29] wirehead themselves [5:31] what's wirehead um stimulate stimulate [5:34] their pleasure centers [5:36] uh and just you know live forever in [5:37] this kind of stasis [5:39] they become well i mean i think the [5:42] reason i believe this is because where [5:44] are they [5:46] if there's some reason they stopped [5:49] expanding [5:50] because otherwise they would have taken [5:51] over the universe the universe isn't [5:52] that big or at least you know [5:54] let's just talk about the galaxy right [5:56] 70 000 light years across [5:58] uh i took that number from star trek [5:59] voyager i don't know how true it is but [6:01] um [6:02] uh yeah that's not big right 70 [6:05] 000 light years is nothing for some [6:08] possible technology that you can imagine [6:09] that could leverage like wormholes or [6:11] something like that [6:12] you don't even need wormholes just a von [6:13] neumann probe is enough a von [6:15] neumann probe and a million years of [6:17] sublight travel and you'd have taken [6:19] over the whole universe [6:20] that clearly uh didn't happen so [6:22] something stopped it [6:23] so you mean if you right for for like a [6:26] few million years [6:26] if you sent out probes that travel close [6:29] what's sublight [6:30] meaning close to the speed of light [6:32] let's it just spreads [6:34] interesting actually that's an [6:35] interesting calculation [6:38] huh so what makes you think that would [6:40] be able to uh communicate with them [6:42] like uh yeah what what's [6:45] why do you think we would able to be [6:47] able to comprehend [6:48] intelligent lives that are out there [6:51] like [6:52] even if they were among us kind of thing [6:54] like [6:55] or even just flying around well [6:58] i mean that's possible it's possible [7:02] that there is some sort of prime [7:03] directive [7:04] uh that'd be a really cool universe to [7:06] live in um and there's some reason [7:07] they're not making themselves visible to [7:09] us [7:10] but it makes sense that they would use [7:13] the same [7:15] well at least the same entropy well [7:17] you're implying the same laws of physics [7:18] i don't know what you mean by entropy [7:20] in this case oh yeah i mean if entropy [7:22] is the scarce resource in the universe [7:24] so what do you think about like stephen [7:26] wolfram and everything is a computation [7:28] and then what if they are traveling [7:31] through this world of computation so [7:33] if you think of the universe as just [7:34] information processing [7:36] then uh what you're referring to with [7:39] with entropy [7:41] and then these these pockets of [7:42] interesting complex computations [7:44] swimming around [7:45] how do we know they're not already here [7:48] how do we know that this like all the [7:51] different [7:52] amazing things that are full of mystery [7:54] on earth [7:55] are just like little footprints of [7:57] intelligence from [7:59] light years away maybe i mean [8:03] i tend to think that as civilizations [8:05] expand they use more and more energy [8:07] uh and you can never overcome the [8:08] problem of waste heat so where is their [8:10] waste heat [8:11] so we'd be able to with our crude [8:13] methods be able to see like there's a [8:15] whole lot of [8:17] energy here but it could be something [8:19] we're not i mean we don't understand [8:21] dark energy right dark matter [8:23] it could be just stuff we don't [8:24] understand at all or they could have a [8:26] fundamentally different physics [8:28] you know like that that we just don't [8:31] even compromise well i think [8:33] okay i mean it depends how far out you [8:34] want to go i don't think physics is very [8:36] different on the other side of the [8:37] galaxy [8:39] i would suspect that they have i mean if [8:42] they're in our [8:43] universe they have the same physics well [8:45] yeah that's the assumption we have [8:47] but there could be like super trippy [8:49] things like [8:51] like our cognition [8:54] only gets to a slice oh and all the [8:57] possible instruments that we can design [8:59] only get to a particular slice of the [9:00] universe and there's something much [9:02] like weirder maybe we can try a thought [9:05] experiment [9:05] um would people from the past [9:09] be able to detect the remnants of our [9:13] uh we would be able to detect our modern [9:15] civilization i think the answer is [9:17] obviously yes [9:18] you mean past from 100 years ago well [9:20] let's even go back further let's go to a [9:22] million years ago [9:24] right the humans who were lying around [9:25] in the desert probably didn't even have [9:27] maybe they just barely had fire uh [9:31] they would understand if a 747 flew [9:33] overhead [9:35] in in in this vicinity but not um [9:39] if the if a 747 flew on mars [9:43] because they wouldn't be able to see far [9:44] because we're not actually communicating [9:46] that well [9:47] with the rest of the universe we're [9:48] doing okay we're just sending out random [9:51] like [9:51] 50s tracks of music true [9:55] and yeah i mean they'd have to you know [9:57] the we've only been broadcasting radio [9:59] waves for [10:00] um 150 years and well there's your light [10:03] cone [10:04] so yeah okay what do you make about all [10:07] the [10:08] i recently came across this uh having [10:12] talked to [10:12] david fravor i don't know if you caught [10:15] what the [10:16] the videos that pentagon released and [10:19] uh the new york times reporting of the [10:21] ufo sightings [10:23] so i kind of looked into it quote [10:25] unquote and [10:27] there's actually been like [10:30] hundreds of thousands of ufo sightings [10:32] right [10:33] and a lot of it you can explain away in [10:36] different kinds of ways [10:37] so one is it could be interesting [10:38] physical phenomena [10:40] two it could be people wanting to [10:43] believe [10:44] and therefore they conjure up a lot of [10:46] different things that just you know when [10:47] you see different kinds of lights some [10:49] basic physics phenomena [10:50] and then you just conjure up ideas of [10:53] possible [10:54] out there mysterious worlds but you know [10:57] it's also possible like you have a case [10:59] of [11:00] david fravor who is a navy pilot [11:04] who's you know as legit as a guest in [11:07] terms of [11:08] humans who are able to perceive [11:11] things in the environment and make [11:14] conclusions whether those things are a [11:16] threat or not [11:17] and he and several other pilots saw [11:21] a thing i don't know if you followed [11:22] this but they saw a thing that they've [11:25] since then called tick tock that moved [11:27] in all kinds of weird ways [11:29] they don't know what it is it could be [11:32] technology developed by by the united [11:35] states [11:36] and they're just not aware of it and the [11:38] surface level from the navy right it [11:40] could be [11:40] different kind of lighting technology or [11:42] drone technology all that kind of stuff [11:44] it could be the russians and then [11:46] chinese all that kind of stuff [11:47] and of course their mind our mind [11:51] can also venture into the possibility [11:54] that it's from another world [11:56] have you looked into this at all what do [11:58] you think about it [11:59] i think all the news is a psyop [12:03] i think that the most closing is real [12:06] yeah i listened to the uh i think it was [12:08] bob lazar [12:10] um on joe rogan and like i believe [12:14] everything this guy is saying and then i [12:16] think that it's probably just some like [12:17] mk ultra kind of thing you know [12:20] uh what do you mean like they they uh [12:23] you know they made some weird thing and [12:24] they called it an alien spaceship you [12:26] know maybe it was just to like [12:27] stimulate young physicist minds we'll [12:29] tell them it's alien technology and [12:31] we'll see what they come up with right [12:33] do you find any conspiracy theories [12:35] compelling like have you [12:37] pulled at the string of the of the rich [12:40] complex world of conspiracy theories [12:42] that's out there [12:43] i think that uh i've heard a conspiracy [12:46] theory that conspiracy theories were [12:47] invented by the cia [12:48] in the 60s to discredit true things yeah [12:52] um so you know you can go to ridiculous [12:57] conspiracy theories like flat earth and [13:00] pizzagate [13:00] and uh you know these things are almost [13:04] to hide like [13:06] conspiracy theories that like you know [13:08] remember when the chinese like locked up [13:09] the doctors who discovered coronavirus [13:11] like i tell people this and i'm like [13:12] no no that's not a conspiracy theory [13:14] that actually happened [13:15] do you remember the time that the money [13:17] used to be backed by gold and now it's [13:18] backed by nothing this is not a [13:20] conspiracy theory this actually [13:22] happened that's one of my worries [13:25] today with the idea of fake news is that [13:29] when nothing is real [13:32] then like you dilute the possibility of [13:36] anything being true by [13:38] conjuring up all kinds of conspiracy [13:40] theories and then you don't know what to [13:41] believe and then [13:43] like the idea of truth of objectivity is [13:46] lost completely [13:47] everybody has their own truth so you [13:50] used to control information by censoring [13:52] it [13:53] then the internet happened and [13:54] governments are like oh we can't [13:56] censor things anymore [13:58] i know what we'll do you know it's the [14:01] old story [14:02] of uh the story of like tying a flag [14:04] where the leprechaun tells you the gold [14:06] is buried [14:06] and you tie one flag and you make the [14:08] leprechaun swear to not remove the flag [14:09] and you come back to the field later [14:11] with a shovel and there's flags [14:12] everywhere [14:14] that's one way to maintain privacy right [14:16] is like [14:18] in order to protect the contents of this [14:20] conversation for example we could just [14:22] generate like millions of deep fake [14:25] conversations where you and i talk [14:27] and say random things yeah so this is [14:29] just one of them and nobody knows which [14:31] one was the real one [14:32] this this could be fake right now [14:34] classic steganography technique [14:37] okay another absurd question about [14:38] intelligent life because uh [14:41] you know you're you're an incredible [14:43] programmer outside of everything else [14:44] we'll talk about just as a programmer [14:48] uh do you think intelligent beings out [14:52] there the civilizations that were out [14:54] there [14:54] had computers and programming [14:58] did they do we naturally have to develop [15:00] something where we engineer machines [15:03] and are able to encode both [15:06] knowledge into those machines and [15:09] instructions that [15:10] process that knowledge process that [15:12] information to to make decisions and [15:14] actions and so on [15:15] and with those programming languages if [15:18] you think they exist [15:20] be at all similar to anything we've [15:22] developed [15:24] so i don't see that much of a difference [15:26] between [15:27] quote-unquote natural languages and [15:29] programming languages [15:32] um yeah i think there's so many [15:35] similarities so [15:37] when asked the question what do alien [15:40] languages look like [15:42] i imagine they're not all that [15:44] dissimilar from ours [15:46] and i think translating in and out of [15:48] them [15:50] uh wouldn't be that crazy well it's [15:53] difficult to compile [15:54] like dna to python and then to see [15:58] i mean there is a little bit of a gap in [16:00] in the kind of languages we use for [16:03] for uh touring machines and the kind of [16:07] languages [16:08] nature seems to use a little bit maybe [16:10] that's just [16:11] we just haven't cr we haven't understood [16:13] the kind of language that nature uses [16:15] well [16:15] yet dna is a cad model [16:18] it's not quite a programming language it [16:21] has no sort of [16:22] serial execution it's not quite a [16:27] yeah it's a cad model um so i think in [16:30] that sense [16:30] we actually completely understand it the [16:32] problem is um you know [16:35] well simulating on these cad models i [16:37] played with it a bit this year [16:38] is super uh computationally intensive if [16:41] you want to go down to like the [16:42] molecular level [16:43] um where you need to go to see a lot of [16:45] these phenomena like protein folding [16:47] um so yeah it's not that it's it's not [16:50] it's not that we don't understand it it [16:52] just requires a whole lot of compute to [16:53] kind of [16:53] compile it for our human minds it's [16:56] inefficient both for the pro [16:58] for the data representation and for the [16:59] programming yeah it runs well on raw [17:01] nature [17:02] it runs well in raw nature and when we [17:04] try to build uh emulators or simulators [17:06] for that [17:06] uh well then manslaughter and i've tried [17:09] it [17:10] it runs in yeah you've commented [17:13] elsewhere [17:14] i don't remember where that uh [17:17] one of the problems is simulating nature [17:19] is tough [17:21] and if you want to sort of deploy a [17:23] prototype [17:25] i forgot how you you put it but it made [17:27] me laugh but animals or humans would [17:29] need to be involved [17:31] in order to in order to try to run some [17:34] prototype code on um [17:38] like if we're talking about covid and [17:39] viruses and so on yeah [17:41] if you were trying to engineer some kind [17:43] of defense mechanisms [17:44] like a vaccine uh against coven or [17:48] all that kind of stuff that doing any [17:50] kind of experimentation like you can [17:52] with like autonomous vehicles would be [17:54] very [17:55] technically cost technically and [17:58] ethically costly i'm not sure about that [18:00] i think you can do tons of [18:02] uh crazy biology and in test tubes i [18:05] think [18:05] my bigger complaint is more all the [18:08] tools are so bad [18:11] like literally you mean like like i'm [18:13] not libraries and i'm not pipetting [18:15] like you're handing me a i gotta no no [18:19] no no there has to be some [18:22] like automating stuff and like the [18:25] yeah but human biology is messy like it [18:28] seems like [18:29] look at those toronto's videos they were [18:31] a joke it's like it's like a little [18:32] gantry it's like a little xy gantry high [18:34] school science project with the pipette [18:36] i'm like [18:37] really gotta be something better you [18:39] can't build like nice microfluidics and [18:41] i can program the you know [18:43] computation to bio interface i mean this [18:45] is going to happen [18:47] but like right now if you are asking me [18:49] to pipette [18:50] 50 milliliters of solution amount [18:54] this is so crude yeah okay [18:57] let's get all the crazy out of the way [18:59] uh so a bunch of people ask me [19:02] since we talked about the simulation [19:03] last time we talked about hacking the [19:06] simulation [19:06] do you have any updates any insights [19:09] about how we might be able to go about [19:12] hacking simulation if we indeed do live [19:15] in a simulation [19:17] i think a lot of people misinterpreted [19:19] the point of that south by talk [19:22] the point of the south by talk was not [19:23] literally to hack the simulation [19:26] uh i think that this [19:30] we this is this is an idea is literally [19:33] just i think theoretical physics i think [19:34] that's the whole [19:36] you know the whole goal [19:39] right you want your grand unified theory [19:41] but then okay build a brand new five [19:42] theory search for exploits [19:44] right i think we're nowhere near [19:46] actually there yet [19:47] my hope with that was just more to like [19:50] like are you people kidding me with the [19:52] things you spend time thinking about [19:54] do you understand like kind of how small [19:56] you are [19:57] you are you are bites and god's computer [20:01] really and the things that people get [20:03] worked up about and [20:05] you know so basically it was more [20:09] a message of uh we should humble [20:11] ourselves [20:12] that we we get to uh like what [20:16] what are we humans in this bite code [20:19] yeah [20:20] and not just just humble ourselves but [20:22] like like i'm not trying to like make [20:24] you feel guilty or anything like that [20:25] i'm trying to say like literally [20:27] look at what you are spending time on [20:29] right what are you referring to you're [20:31] referring to the kardashians what are we [20:32] talking about [20:33] from twitter to no the kardashians see [20:36] everyone knows that's kind of [20:37] fun i'm referring more to like the [20:40] economy [20:42] you know this idea that [20:46] we gotta up our stock price like [20:50] or or what is what is the goal function [20:53] of humanity [20:55] you don't like the game of capitalism [20:57] like you don't like the games we've [20:58] constructed for ourselves as humans [21:00] i'm a big fan of of capitalism i don't [21:03] think that's really the game we're [21:04] playing right now i think we're playing [21:05] a [21:06] a different game where the rules are [21:08] rigged [21:10] look at which games are interesting to [21:12] you that we humans have constructed and [21:13] which aren't [21:14] which are productive and which [21:17] are not actually maybe that's the real [21:20] point of the [21:21] of the talk it's like stop playing these [21:24] fake human games there's a real game [21:26] here [21:26] we can play the real game the real game [21:29] is you know nature wrote the rules [21:31] this is a real game there still is a [21:33] game to play [21:35] but if you look at sergeant drop i don't [21:37] know if you've seen the instagram [21:38] account nature is metal [21:40] the game that nature seems to be playing [21:42] is a lot [21:43] a lot more cruel than we humans want to [21:46] put up with or [21:47] at least we see it as cool it's like the [21:50] bigger [21:50] thing eats the smaller thing and uh [21:54] does it to impress another big thing [21:58] so it can mate with that thing and [22:00] that's it that [22:01] seems to be the entirety of it well [22:05] there's no art there's no music there's [22:07] no [22:08] comma ai there's no comma one no comma [22:11] two no george [22:13] hotz with his brilliant talks at south [22:15] by southwest [22:16] see i disagree though i disagree that [22:18] this is what nature is i think nature [22:20] just provided basically a uh [22:24] open world and mmorpg and um [22:28] you know here it's open world i mean if [22:30] that's the game you want to play you can [22:31] play that game but [22:32] isn't that isn't that beautiful i know [22:33] if you play diablo they used to have uh [22:36] i think cow level [22:38] where it's [22:41] so everybody will go just they figured [22:43] out this like the best way to gain [22:46] like experience points is to just [22:48] slaughter cows over and over and over [22:52] and uh so they figured out this little [22:55] sub game within the bigger game that [22:56] this is the most efficient way [22:58] to get experience points and everybody [23:01] somehow agreed that getting experience [23:02] points in [23:03] rpg context where you always want to be [23:05] getting more stuff [23:06] more skills more levels keep advancing [23:09] that seems to be good [23:10] so might as well spend sacrifice [23:13] actual enjoyment of playing a game [23:16] exploring a world [23:17] and spending like hundreds of hours of [23:20] your time in cow level i mean [23:22] the number of hours i spent in cow level [23:26] i'm not like the most impressive person [23:28] because people have probably thousands [23:29] of hours there but it's ridiculous [23:31] so that's a little absurd game that [23:34] brought me joy in some weird dopamine [23:36] drug kind of way [23:37] yeah so you you don't like those games [23:39] you don't [23:40] you don't think that's us humans failing [23:43] the the yeah nature i think so [23:47] and that was the point of the talk yeah [23:50] so how do we hack it then [23:51] well i want to live forever and wait [23:54] i want to live forever and this is like [23:56] the goal well that's a game against [23:57] nature [23:59] yeah immortality is the good objective [24:01] function to you [24:03] i mean start there and then you can do [24:04] whatever else you want cause you got a [24:05] long time [24:07] what if mortality makes the game just [24:09] totally not fun [24:10] i mean like why do you assume [24:12] immortality [24:13] is uh somehow uh it's not [24:16] a good objective function it's not [24:18] immortality that i want a true [24:20] immortality where i could not die [24:22] i would prefer what we have right now um [24:24] but i want to choose my own death of [24:26] course [24:27] i don't want nature to decide when i die [24:29] i'm going to win i'm going to be you [24:32] and then at some point if you choose [24:35] commit suicide [24:36] like how long you think you'd live [24:41] until i get bored see i don't think [24:43] people [24:45] like in like brilliant people like you [24:48] that really [24:48] ponder living a long time [24:52] are really considering how [24:55] how meaningless life becomes well i want [24:58] to know everything and then i'm ready [25:00] to die [25:03] as long as why do you want isn't it [25:05] possible that you want to know [25:06] everything because [25:08] it's finite like the reason you want to [25:10] know quote unquote everything is because [25:12] you don't have enough time to know [25:15] everything [25:16] and once you have unlimited time then [25:18] you realize like [25:20] why do anything like why learn anything [25:24] i want to know everything and i'm ready [25:26] to die so you have yeah [25:28] well it's not it's not a like it's a [25:30] terminal value it's not [25:32] it's not in service of anything else i'm [25:34] conscious of the possibility this is not [25:36] a certainty [25:37] but the possibility is of that engine of [25:40] curiosity that you're speaking to [25:43] is actually a [25:46] a symptom of uh the finiteness of life [25:49] like [25:50] without that finiteness [25:53] your curiosity would vanish like like a [25:55] like a morning fog [25:56] all right cool then you talked about [25:58] love like that then um [26:00] let me solve immortality let me change [26:01] the thing in my brain that reminds me of [26:03] the fact that i'm immortal tells me that [26:05] life is finite maybe i'll have it [26:06] tell me that life ends next week [26:08] right i'm okay with some self [26:11] manipulation like that i'm okay with [26:13] with deceiving oh change [26:14] oh rico changing the code if that's the [26:17] problem right if the problem is that i [26:18] will no longer have that [26:20] that curiosity i'd like to have backup [26:22] copies of myself [26:24] uh which yeah well which i check in with [26:26] occasionally to make sure they're okay [26:28] with the trajectory and they can kind of [26:29] override it [26:30] maybe a nice like i think of like those [26:32] wavenets those like logarithmic go back [26:34] to the copies yeah but sometimes it's [26:35] not reversible like uh [26:37] sure i've done this with video games [26:39] when once you figure out the cheat code [26:41] or like you look up [26:42] how to cheat old school like single [26:44] player [26:45] it ruins the game for you absolutely i [26:47] know that feeling but again [26:50] that just means our brain manipulation [26:51] technology is not good enough yet remove [26:53] that cheat code from your brain [26:54] what if we all so it's also possible [26:56] that if we figure out immortality [27:00] that all of us will kill ourselves [27:03] before we advance [27:04] far enough to uh to be able to revert to [27:08] change [27:08] i'm not killing myself till i know [27:09] everything so [27:11] that's what you say now because your [27:13] life is finite [27:15] you know i think yes self-modifying [27:17] systems gets [27:18] comes up with all these hairy [27:20] complexities and can i promise that i'll [27:22] do it perfectly no but i think i can put [27:24] good safety structures in place [27:27] so that talk in your thinking here is [27:29] not literally [27:33] referring to uh a simulation in that our [27:37] our universe is a kind of computer [27:40] program running in a computer [27:42] that's more of a thought experiment um [27:45] do you also think of the potential of [27:48] the sort of uh [27:50] bostrom elon musk [27:53] and others that talk about an actual [27:56] program [27:57] that simulates our universe oh i don't [28:00] doubt that we're in a simulation i just [28:02] think that it's not [28:03] quite that important i mean i'm [28:05] interested only in simulation theory as [28:06] far as like it gives me power [28:08] over nature uh if it's totally [28:10] unfalsifiable then [28:12] who cares i mean what do you think that [28:13] experiment would look like like somebody [28:15] uh [28:16] on twitter asked ask george what signs [28:19] we would look for to know whether or not [28:21] we're in the simulation [28:22] which is exactly what you're asking is [28:24] like [28:26] the step that precedes the step of [28:29] knowing how to get more power [28:30] from this knowledge is to get an [28:32] indication that there is some power to [28:34] be gained so [28:35] get an indication that there you can [28:38] discover [28:39] and exploit cracks in the simulation or [28:42] it doesn't [28:43] you know in the physics of the universe [28:45] yeah [28:46] show me i mean like a memory leak would [28:49] be cool [28:51] like some scrying technology you know [28:53] what what kind of technology [28:55] scrying what's that oh that's a weird uh [28:58] this crying is the [28:59] is the uh paranormal ability to uh [29:02] like like remote viewing like being able [29:05] to see somewhere where you're not [29:07] um so you know i don't think you can do [29:09] it by chanting in a room but um [29:12] if we could find as a memory leak [29:14] basically [29:17] yeah you're able to access parts you're [29:19] not supposed to yeah yeah yeah and [29:20] thereby discover shortcut [29:22] yeah maybe memory leak means the other [29:24] thing as well but i mean like yeah like [29:25] an ability to read arbitrary memory [29:27] yeah right and that one's not that [29:28] horrifying right the the right ones [29:30] start to be horrifying [29:31] read right so the the reading is not the [29:34] problem yeah [29:35] it's like heartbleed for the universe oh [29:37] boy the writing is a big big problem [29:40] it's a big problem it's the moment you [29:43] can write anything even if it's just [29:45] random noise that's terrifying [29:49] i mean even without even without that [29:51] like even some of the you know the [29:52] nanotech stuff that's coming i think is [29:56] i don't know if you're paying attention [29:58] but actually eric weinstein came out [30:00] with the theory of everything [30:02] i mean that came out he's been working [30:04] on a theory of everything in the physics [30:05] world [30:06] called geometric community and then for [30:08] me from computer science person [30:10] like you stephen wolfram's theory of [30:13] everything of like hypographs is super [30:15] interesting and beautiful [30:17] but not from a physics perspective but [30:19] from a computational perspective i don't [30:20] know have you paid attention [30:22] to any of that so again like what would [30:25] make me pay attention and like why like [30:27] i hate string theory is okay make a [30:30] testable prediction [30:31] right i'm only interested in i'm not [30:33] interested in theories for their [30:34] intrinsic beauty i'm interested in [30:36] theories that give me power over the [30:37] universe [30:39] so if these theories do i'm very [30:41] interested um can i just say how [30:43] beautiful that [30:44] is because a lot of physicists say i'm [30:47] interested in experimental validation [30:49] and they skip out the part [30:51] where they say to give me more power in [30:54] the universe [30:55] i just love the um yo i want i want i [30:58] want the clarity of that [30:59] i want 100 gigahertz processors i want [31:02] transistors that are smaller than atoms [31:04] i want like [31:04] power [31:08] that's uh that's true and that's where [31:10] people [31:11] from aliens to this kind of technology [31:13] where people are worried that [31:15] governments like who owns that power [31:19] is it george hearts is it thousands of [31:22] distributed hackers across the world [31:24] is it governments you know is it mark [31:27] zuckerberg [31:28] there's a lot of people that uh [31:32] i don't know if anyone trusts any one [31:34] individual with power so they're always [31:36] worried [31:37] it's the beauty of blockchains that's [31:39] the beauty of blockchains [31:41] which we'll talk about on twitter [31:43] somebody pointed me to a story uh [31:46] a bunch of people pointed me to a story [31:48] a few months ago [31:49] where you went into a restaurant in new [31:50] york and you can correct me fame this is [31:52] wrong [31:53] and ran into a bunch of folks from a [31:56] company in a [31:57] crypto company who are trying to scale [31:59] up ethereum [32:01] and they had a technical deadline [32:03] related to a solidity to ovm compiler [32:07] so these are all ethereum technologies [32:09] so [32:10] you stepped in they recognized you uh [32:14] pulled you aside explained their problem [32:16] and you stepped in and helped them solve [32:17] the problem [32:19] uh thereby creating legend status [32:22] story so uh can you uh [32:26] tell me the story the little more detail [32:28] it seems kind of incredible [32:31] this did this happen yeah yeah it's a [32:32] true story it's a true story i mean they [32:34] wrote a very flattering [32:35] account of it um they [32:39] so optimism is the spin the company's [32:42] called optimism [32:43] spin-off of plasma they're trying to [32:45] build l2 solutions on ethereum [32:47] so right now uh [32:50] every ethereum node has to run every [32:53] transaction [32:54] on the ethereum network um and this kind [32:57] of doesn't scale right because if you [32:58] have n computers [32:59] well you know if that becomes two n [33:01] computers you actually still get the [33:03] same amount of compute [33:04] right this is this is like like o of one [33:07] scaling [33:08] um because they all have to run it okay [33:10] fine you get more [33:11] blockchain security but like the [33:13] blockchain's already so secure [33:15] can we trade some of that off for speed [33:17] uh so that's kind of what these l2 [33:18] solutions are [33:20] they built this thing which kind of um [33:22] kind of sandbox [33:24] uh for ethereum contracts so they can [33:26] run it in this l2 world and it can't do [33:28] certain things in l world [33:30] in l1 i can ask you for some definitions [33:32] what's l2 [33:33] oh l2 is layer 2. so l1 is like the base [33:36] ethereum chain [33:37] and then layer two is like a [33:39] computational layer that runs [33:42] um elsewhere but still is kind of [33:45] secured by layer one [33:47] and i'm sure a lot of people know but [33:49] ethereum is a cryptocurrency probably [33:52] one of the most popular cryptocurrencies [33:53] second to bitcoin [33:55] and a lot of interesting technological [33:57] innovations there [33:58] maybe you can also slip in [34:01] whenever you talk about this any things [34:03] that are exciting to you in the [34:05] ethereum space and why ethereum well i [34:08] mean [34:08] bitcoin uh is not turn complete [34:12] well ethereum is not technically a [34:13] terrain complete with a gas limit but [34:14] close enough [34:15] well the gas limit what's the gas limit [34:18] resources [34:18] yeah i mean no computers actually turn [34:20] complete right right [34:23] you're fine at ram you know what if i [34:24] can actually solve this gas limit you [34:26] just have so many brilliant [34:28] words i'm not even gonna ask but that's [34:29] what that's no that's not my word that's [34:30] ethereum's word [34:32] gasoline ethereum you have to spend gas [34:34] per instruction [34:35] so like different op codes use different [34:36] amounts of gas and you [34:38] buy gas with ether to prevent people [34:40] from basically ddosing the network [34:42] so uh bitcoin is proof of work and then [34:46] what's ethereum it's also proof of work [34:48] uh they're working on some [34:49] proof-of-stake ethereum 2.0 stuff but [34:51] right now it's [34:51] it's proof of work usually a different [34:53] hash function from bitcoin that's more [34:55] asic resistance because you need ram [34:57] so we're all talking about ethereum 1.0 [34:59] yeah so what uh [35:00] what were they trying to do to scale [35:02] this whole process so they were like [35:04] well [35:04] if we could run contracts elsewhere um [35:07] and then only save the results of that [35:09] computation [35:11] uh you know well we don't actually have [35:13] to do the computer on the chain we can [35:14] do the compute off chain and just post [35:16] what the results are [35:17] now the problem with that is well [35:18] somebody could lie about what the [35:20] results are [35:21] so you need a resolution mechanism and [35:23] the resolution mechanism can be really [35:24] expensive [35:26] uh because you know you just have to [35:28] make sure that like [35:29] the person who is saying look i swear [35:32] that this is the real computation [35:33] i'm staking ten thousand dollars on that [35:35] fact and [35:37] if you prove it wrong yeah it might cost [35:39] you three thousand dollars in gas fees [35:41] to prove wrong [35:42] but you'll get the ten thousand dollar [35:43] bounty so you can secure [35:45] using those kind of systems um [35:48] so it's effectively a sandbox which runs [35:51] contracts [35:52] uh and like just like any kind of normal [35:55] sandbox you have to like replace [35:56] syscalls with um you know calls into the [35:59] hypervisor [36:01] uh sandbox this calls hypervisor what do [36:05] these things mean [36:06] uh as long as it's interesting to talk [36:08] about yeah i mean you can take like the [36:10] chrome [36:11] is maybe the one to think about right so [36:13] the chrome process that's doing a [36:14] rendering [36:15] uh can't for example read a file from [36:17] the file system yeah [36:18] it has if it tries to make an open [36:20] syscall in linux the open system [36:22] you can't make it open says call no no [36:24] no uh you have to request [36:26] from the kind of uh hypervisor process [36:29] or [36:29] like i don't know what's called in [36:31] chrome but um [36:33] the canoe hey could you open this file [36:35] for me and then it does all these checks [36:37] and then it passes the file handle back [36:38] in if it's approved [36:40] um so that's yeah uh so what's the in [36:43] the context of ethereum [36:45] what are the boundaries of the sandbox [36:47] that we're talking about um well like [36:48] one of the calls that you [36:50] actually reading and uh writing any [36:53] state to the ethereum contract [36:55] to the ethereum blockchain um [36:58] writing state is one of those calls that [37:01] you're going to have to sandbox [37:03] in layer two because if you let layer [37:06] two just [37:06] arbitrarily right to the ethereum [37:08] blockchain um [37:10] so layer two is except is really sitting [37:12] on top of [37:14] layer one so you're gonna have a lot of [37:16] different kinds of ideas that you can [37:17] play with [37:18] yeah and they're all they're not [37:19] fundamentally changing [37:21] the source code level of ethereum [37:25] well you have to replace a bunch of [37:28] calls [37:28] with calls into the hypervisor so [37:31] instead of doing the syscall directly [37:33] you you replace it with a call to the [37:35] hypervisor [37:37] so originally they were doing this by [37:40] first running the so solidity is the [37:42] language that most ethereum contracts [37:44] are written in [37:45] it compiles to a byte code and then they [37:47] wrote this thing they called the [37:48] transpiler and the transpiler took the [37:51] byte code [37:52] and it transpiled it into ovm safe [37:55] bytecode [37:55] basically bytecode that didn't make any [37:57] of those restricted syscalls and added [37:59] the calls to the hypervisor [38:01] this transpiler was a 3000 line mess [38:05] and it's hard to do it's hard to do if [38:07] you're trying to do it like that because [38:09] you have to kind of like deconstruct the [38:10] byte code [38:12] change things about it and then [38:13] reconstruct it [38:15] and i mean as soon as i hear this i'm [38:16] like why don't you just change the [38:18] compiler [38:19] right why not the first place you build [38:21] the bytecode just do it in the compiler [38:24] uh so yeah you know i asked them how [38:27] much they wanted it [38:28] uh of course measured in dollars and i'm [38:31] like well okay [38:32] um and yeah and you wrote the compiler [38:35] yeah i modified i wrote a 300 line diff [38:38] to the compiler [38:38] uh it's open source you can look at it [38:40] yeah it's yeah i looked at the code last [38:42] night [38:43] [Laughter] [38:45] yeah exactly cute good is a good word [38:48] for it [38:48] uh and it's um c plus plus [38:51] see if it's lost yeah so when [38:55] asked how you were able to do it you [38:57] said [38:58] you just gotta think and then do it [39:01] right [39:02] so can you break that apart a little bit [39:04] what's what's your process of uh [39:06] one thinking and two doing it right [39:09] you know they they the people i was [39:11] working for are amused that i said that [39:13] it doesn't really mean anything [39:14] okay i mean is there some [39:17] deep profound insights to draw from like [39:21] how you problem solve from that because [39:23] this is always what i say i'm like do [39:24] you want to be a good programmer do it [39:26] for 20 years [39:27] yeah there's no shortcuts yeah [39:31] what are your thoughts on crypto in [39:32] general so would [39:34] what what parts technically or [39:37] philosophically do you find especially [39:38] beautiful maybe [39:39] oh i'm extremely bullish on crypto long [39:42] term not any specific crypto project [39:44] but this idea of [39:48] well two ideas one um [39:51] the nakamoto consensus algorithm is [39:54] i think one of the greatest innovations [39:57] of the 21st century [39:58] this idea that people can reach [40:00] consensus you can reach a group [40:02] consensus [40:03] using a relatively straightforward [40:05] algorithm [40:06] um is wild and [40:10] like you know satoshi nakamoto people [40:14] always ask me who i look up to it's like [40:16] whoever that is who do you think it is i [40:19] mean [40:19] elon musk is it you it is definitely not [40:23] me and i do not think it's elon musk [40:26] but yeah this idea of uh [40:29] groups reaching consensus in a [40:32] decentralized yet formulaic way [40:34] is one extremely powerful idea from [40:36] crypto [40:40] maybe the second idea is this idea of [40:43] smart contracts [40:45] when you write a contract between two [40:48] parties [40:49] any contract um this contract [40:52] if there are disputes it's interpreted [40:54] by lawyers [40:56] lawyers are just really shitty overpaid [40:58] interpreters [40:59] imagine you had let's talk about them in [41:01] terms of a in terms of like let's [41:02] compare a lawyer to python [41:04] right so lawyer well okay that's really [41:08] oh i never thought of it that way it's [41:10] hilarious [41:11] so python i'm paying i'm paying um [41:14] you know even 10 cents an hour i'll use [41:16] the nice azure machine i can run python [41:17] for 10 cents an hour [41:19] lawyers cost a thousand dollars an hour [41:21] so python is is is [41:23] 10 000 x uh better on that axis um [41:26] lawyers don't always return the same [41:28] answer [41:30] um python almost always does [41:34] uh cost [41:37] yeah i mean just just cost reliability [41:40] everything about python is so much [41:42] better than lawyers um so if you can [41:44] make smart contracts [41:45] this whole concept of code is law [41:50] i i love and i would love to live in a [41:52] world where everybody accepted that fact [41:55] so so maybe uh you can talk about what [41:59] smart contracts [42:00] are so let's say um [42:03] let's say you know we have a uh [42:07] even something as simple as a safety [42:10] deposit box [42:11] right safety deposit box that holds a [42:13] million dollars [42:14] i have a contract with the bank that [42:16] says two out of these three parties [42:19] uh must uh be present to open the safety [42:23] deposit box and get the money out [42:25] so that's a contract for the bank and [42:26] it's only as good as the bank and the [42:27] lawyers right [42:28] let's say you know somebody dies and now [42:32] oh we're going to go through a big legal [42:34] dispute about whether oh well was it in [42:35] the will was it not in the well [42:37] what like it's just so messy and the [42:40] cost [42:40] to determine truth is so expensive [42:44] versus a smart contract which just uses [42:46] cryptography to check if two out of [42:48] three keys are present [42:50] well i can look at that and i can have [42:52] certainty [42:53] in the answer that it's going to return [42:55] that's what all businesses want [42:57] certainty you know they say businesses [42:58] don't care viacom youtube [43:00] youtube's like look we don't care which [43:02] way this lawsuit goes [43:04] just please tell us so we can have [43:06] certainty yeah i wonder how many [43:08] agreements in this world because we're [43:09] talking about financial transactions [43:11] only in this case correct the smart the [43:14] smart contracts oh you can go to you can [43:16] go to anything [43:17] you can go you could put a prenup in the [43:18] theorem blockchain [43:21] a married smart contract sorry divorce [43:23] lawyers sorry you're going gonna be [43:25] replaced by [43:26] python uh [43:29] okay so that's uh [43:32] so that's that's another beautiful idea [43:34] do you think [43:36] there's something that's appealing to [43:37] you about any one specific [43:38] implementation [43:40] so if you look 10 20 50 years down the [43:44] line [43:44] do you see any like bitcoin ethereum [43:48] any of the other hundreds of [43:49] cryptocurrencies winning out is there [43:51] like what's your intuition about the [43:52] space are you just sitting back and [43:54] watching the chaos and look who cares [43:56] what emerges [43:57] oh i don't i don't speculate i don't [43:59] really care i don't really care which [44:00] one of these projects wins [44:02] i'm kind of in the bitcoin as a meme [44:04] coin camp i mean why does bitcoin have [44:06] value [44:06] it's technically kind of you know [44:10] what yeah not great like the block size [44:13] debate or [44:14] when i found out what the block size [44:15] debate was i'm like are you guys [44:17] kidding what's the block size debate [44:21] you know what it's really it's too [44:22] stupid to even talk about people [44:24] people people can look it up but i'm [44:26] like wow you know ethereum seems the [44:28] governance of ethereum seems much better [44:30] um i've come around i've been on proof [44:32] of stake ideas [44:34] uh you know very smart people thinking [44:36] about some things yeah [44:38] you know governance is interesting it [44:40] does feel like uh [44:42] vitalik it could just feel like an [44:45] open in even in these distributed [44:47] systems leaders and [44:49] are helpful because they kind of [44:52] help you drive the mission and the [44:54] vision [44:55] and they put a face to a project it's a [44:58] weird thing about us humans [45:00] geniuses are helpful like mattel right [45:02] yeah [45:03] brilliant [45:06] leaders are not necessarily yeah [45:10] so you think the reason he's uh he's the [45:13] face [45:14] of a theorem is because he's a genius [45:16] that's interesting [45:18] i mean that was um [45:21] it's interesting to think about that we [45:23] need to create systems [45:25] in which uh the quote unquote leaders [45:29] that emerge [45:30] are the geniuses in the system i mean [45:33] that's [45:33] arguably why the current state of [45:35] democracy is broken is the people who [45:37] are [45:38] emerging as the leaders are not the most [45:40] competent are not the superstars of the [45:42] system [45:43] and it seems like at least for now in [45:45] the crypto world oftentimes [45:47] the leaders are the superstars imagine [45:49] at the debate they asked [45:51] what's the sixth amendment what are the [45:53] four fundamental forces in the universe [45:55] right what's the integral of two to the [45:57] x [45:58] yeah i i'd love to see those questions [46:01] asked and that's what i want as our [46:02] leader [46:03] it's it's a little bit of a bayes rule [46:07] yeah i mean even oh wow you're hurting [46:10] my brain [46:11] it's that my standard was even lower but [46:15] i would have loved to see just this [46:18] basic [46:19] brilliance like i've talked to [46:21] historians there's just these [46:23] they're not even like they don't have a [46:24] phd or even education history they just [46:27] like a dan carlin type character who [46:30] just like [46:31] holy how did all this information [46:34] get into your head [46:35] they're able to just connect uh genghis [46:38] khan to [46:39] the entirety of the history of the 20th [46:40] century they [46:42] they know everything about every single [46:44] battle that happened [46:46] and they know the the the like [46:49] the game of thrones of the of the [46:52] different power plays and all that [46:54] happened there [46:55] and they know like the individuals they [46:56] know all the documents involved [46:58] and it's and that they integrate that [47:00] into their regular life it's not like [47:02] they're ultra history nerds they're just [47:04] they know this information that's what [47:06] competence looks like yeah [47:08] because i've seen that with programmers [47:10] too right that's what great programmers [47:11] do [47:12] but yeah it would be uh it's really [47:15] unfortunate that those kinds of people [47:16] aren't emerging as as our leaders but [47:19] for [47:19] now at least in the crypto world that [47:21] seems to be the case [47:23] i don't know if that always uh you could [47:26] imagine that in a hundred years that's [47:27] not the case right the crypto world has [47:29] one very powerful idea going for it and [47:31] that's the idea of forks [47:33] right i mean [47:37] you know imagine uh [47:40] we'll use a less controversial example [47:42] um this was actually in my joke [47:44] uh app in 2012 i was like barack obama [47:48] mitt romney let's let him both be [47:50] president right like imagine we could [47:51] fork america [47:52] and just let them both be president and [47:54] then the americas could compete [47:56] and you know people could invest in one [47:58] pull their liquidity out of one put it [47:59] in the other [48:00] you have this in the crypto world [48:02] ethereum forks into ethereum and [48:03] ethereum classic [48:05] and you can pull your liquidity out of [48:06] one and put it in another [48:08] and people vote with their dollars um [48:11] which forks companies should be able to [48:15] fork [48:16] i'd love to fork nvidia you know [48:20] yeah like different business strategies [48:22] and yeah and then try them out and see [48:24] see what works like even take uh [48:30] uh yeah take comedy i that closes its [48:33] source [48:34] and then take one that's open source and [48:36] see what works [48:38] take one that's purchased by gm and one [48:41] that remains [48:42] android renegade and all these different [48:44] versions and see the beauty of comma ai [48:46] is someone could actually do that [48:47] yeah please take come ai and fork it [48:50] that's right [48:51] that's the beauty of open source so [48:53] you're i mean [48:54] we'll talk about autonomous vehicle [48:56] space but it does seem that [48:59] you're really knowledgeable about [49:02] a lot of different topics so the natural [49:04] question a bunch of people ask this [49:06] which is [49:07] uh how do you keep learning new things [49:09] do you have [49:10] like practical advice if you were to [49:12] introspect [49:14] like taking notes allocate time [49:17] or do you just mess around and just [49:19] allow your curiosity to drive i'll write [49:21] these people a self-help book and i'll [49:23] charge 67 [49:24] for it and i will i will write i will [49:27] write chapter one i will write on the [49:29] cover of the self-help [49:30] book all of this advice is completely [49:31] meaningless you're gonna be a sucker and [49:33] buy this book anyway [49:34] yeah and the one lesson that i hope they [49:36] take away from the book [49:38] is that i can't give you a meaningful [49:40] answer to that [49:42] that's interesting let me translate that [49:46] is you haven't really thought about what [49:49] it is you do [49:51] systematically because you could reduce [49:53] it and there's some people i mean i've [49:54] met [49:55] brilliant people that this is really [49:58] clear with athletes [50:00] some are just you know the best in the [50:03] world that's something [50:05] and they they have zero interest in [50:06] writing like a self-help book [50:08] but or how to master this game and then [50:11] there's [50:12] some athletes who become great coaches [50:15] and they love the analysis [50:17] perhaps the over analysis and you right [50:19] now at least at your age [50:20] which isn't interesting you're in the [50:22] middle of the battle you're like the [50:23] warriors that have zero interest in [50:25] writing books [50:26] uh so you're in the middle of the battle [50:28] so you have yeah [50:30] this is this is a fair point i do think [50:32] i have a certain aversion [50:33] to um this kind of deliberate [50:37] intentional way of living life [50:40] here eventually the hilarity of this [50:43] especially [50:43] since this is recorded [50:47] it will reveal beautifully the absurdity [50:49] when you finally do publish this book [50:51] and i guarantee you you will the story [50:54] of comma ai [50:55] would be maybe it'll be a biography [50:57] written about you [50:58] they'll be they'll be better i guess and [51:00] you might be able to learn some cute [51:01] lessons if you're starting a company [51:03] like comma ai from that book [51:05] but if you're asking generic questions [51:07] like how do i be good at things [51:10] dude i don't know well learn i mean the [51:12] interesting do them a lot [51:14] i do them a lot but the interesting [51:16] thing here is [51:17] learning things outside [51:20] of your current trajectory which is what [51:22] it feels like from an outsider's [51:23] perspective i mean that [51:26] uh you know that i don't know if there's [51:28] an advice on that [51:30] but it is an interesting curiosity when [51:33] you become really busy [51:34] you're running a company [51:37] part time yeah [51:41] but like there's a natural inclination [51:43] and [51:45] trend like just the the the momentum of [51:47] life [51:48] carries you into a particular direction [51:49] of wanting to focus and this kind of [51:52] dispersion that curiosity can lead to [51:55] gets harder and harder with time because [51:58] you're [51:58] you get really good at certain things [52:00] and it sucks [52:02] trying things that you're not good at [52:03] like trying to figure them out [52:05] you do this with your live streams [52:07] you're on the [52:08] fly figuring stuff out you don't mind [52:10] looking dumb [52:13] you just figured out figure it out [52:15] pretty quickly sometimes i try things [52:17] and i don't figure them out [52:19] my chest rating is like a 1400 despite [52:21] putting like [52:22] a couple hundred hours in it's pathetic [52:24] i mean to be fair i know that i could do [52:26] it better if i did it better like don't [52:27] play you know don't play five-minute [52:29] games play 15-minute games at least like [52:31] i know these things but it just doesn't [52:34] it doesn't stick nicely in my knowledge [52:35] tree [52:37] all right let's talk about comma ai [52:39] what's the mission of the company [52:42] let's like look at the biggest picture [52:44] oh i have an exact statement [52:46] solve self-driving cars while delivering [52:49] shippable intermediaries [52:51] so long-term vision is have [52:54] fully autonomous vehicles and make sure [52:56] you're making money along the way [52:58] i think it doesn't really speak to money [53:00] but i can talk i can talk about what [53:01] solve self-driving cars means [53:03] solve self-driving cars of course means [53:05] um [53:06] you're not building a new car you're [53:08] building a person replacement [53:10] uh that person can sit in the driver's [53:11] seat and drive you anywhere a person can [53:13] drive [53:14] with a human or better level of safety [53:17] speed quality comfort [53:21] and what's the second part of that [53:23] delivering shippable intermediaries [53:25] um is well it's a way to fund the [53:27] company that's true but it's also a way [53:28] to keep us honest [53:30] uh if you don't have that it is very [53:34] easy [53:34] with this technology to think you're [53:38] making progress when you're not [53:40] i've heard it best described on hacker [53:41] news as [53:43] you can set any arbitrary milestone [53:46] meet that milestone and still be [53:48] infinitely far away from solving [53:49] self-driving cars [53:51] so it's hard to have like real deadlines [53:53] when you're like [53:55] cruz or waymo [53:58] when uh you don't have uh revenue [54:02] is that i mean is revenue essentially [54:05] the thing we're talking about here [54:07] revenue is [54:08] is capitalism is based around consent [54:11] capitalism the way that you get revenue [54:12] is kind of [54:13] real capitalism commas in the real [54:15] capital is in camp there's definitely [54:16] scams out there but [54:17] real capitalism is based around consent [54:19] it's based around this idea that like if [54:20] we're getting revenue [54:21] it's because we're providing at least [54:23] that much value another person when [54:24] someone buys a thousand dollar comment [54:26] two from us [54:27] we're providing them at least a thousand [54:28] dollars of value where they wouldn't buy [54:29] it brilliant [54:30] so can you give a whirlwind overview of [54:32] the products that come i [54:34] provides like uh throughout its history [54:37] and today [54:38] i mean yeah the past ones aren't really [54:40] that interesting it's kind of just been [54:41] refinement of the same idea uh [54:45] the real only product we sell today is [54:47] the comma two which is a piece of [54:48] hardware [54:49] with cameras um so the comet to [54:53] i mean you can think about it kind of [54:54] like a person uh you know when future [54:57] hardware will probably be even more and [54:58] more person-like [54:59] um so it has uh you know eyes [55:03] ears a mouth a brain [55:06] uh and a way to interface with the car [55:09] does it have consciousness [55:10] just kidding that was a trick question [55:12] because i don't have consciousness [55:14] either [55:15] me and the common two are the same [55:16] they're the same i have a little more [55:17] compute than it [55:18] it only has like the same computer [55:20] interesting b [55:21] uh you know you're more efficient energy [55:24] wise [55:25] for the compute you're doing far more [55:26] efficient energy-wise [55:28] huh 20 paid flaps 20 watts crazy you [55:30] lack consciousness [55:32] sure do you fear death you do you want [55:34] immortality [55:35] does comey i fear death i don't think so [55:39] of course it does it very much fears [55:41] while it fears negative loss [55:42] oh yeah okay [55:46] so come a comma two when did that come [55:48] out that that was a year ago [55:50] no two uh early this year [55:53] wow time it feels like yeah [55:56] 2020 feels like uh it's taken 10 years [56:00] to get to the end it's a long year [56:01] it's a long year so um [56:04] what uh what's the sexiest thing about [56:07] comma [56:08] too feature-wise so i mean maybe you can [56:11] also link on like [56:12] what is it like what's its purpose [56:15] because there's a hardware there's a [56:16] software component [56:18] you've mentioned the sensors but also [56:20] like what is of its features and [56:22] capabilities [56:22] i think our slogan summarizes it well uh [56:25] comma slogan is make driving chill [56:28] i love it okay yeah i mean it is [56:32] you know if you like cruise control [56:34] imagine cruise control but much much [56:36] more [56:37] so it can uh do adaptive cruise control [56:40] things which is like slow down for cars [56:42] in front of it maintain a certain speed [56:44] and you can also do lane keeping so [56:46] staying in the lane [56:47] and doing it better and better and [56:48] better over time that's very much [56:50] machine learning based [56:53] so this camera is there's a driver [56:55] facing camera [56:56] too that's [56:59] um what else is there what am i thinking [57:02] so the hardware versus software so [57:04] open pilot versus the actual hardware of [57:07] the device [57:08] what's can you draw that distinction [57:10] what's one what's the other i mean the [57:11] hardware is pretty much [57:13] a cell phone with a few additions a cell [57:14] phone with a cooling system [57:16] and with a car interface connecting to [57:20] it [57:20] and as by cell phone you mean uh like [57:23] qualcomm snapdragon [57:25] uh yeah the current hardware is a [57:26] snapdragon 821 [57:28] uh it has wi-fi radio it has an lte [57:30] radio it has a screen [57:32] uh we use every part of the cell phone [57:35] and then the interface of the car is [57:37] specific to the car so you keep [57:38] supporting more and more cars [57:41] um yeah so the interface to the car i [57:42] mean the device itself just has four can [57:44] buses [57:45] has four cam interfaces on it that are [57:46] connected through the usb port to the [57:48] phone [57:49] um and then yeah on those four can buses [57:52] uh you connect it to the car and there's [57:54] a little harness to do this [57:56] cars are actually surprisingly similar [57:58] so can is the the protocol by which cars [58:00] communicate [58:01] and then you're able to read stuff and [58:03] write stuff to be able to control the [58:05] car [58:05] depending on the car so what's the [58:07] software side what's open pilot [58:09] um so i mean open pilot is the hardware [58:12] is pretty simple compared to open pilot [58:13] open pilot is [58:15] uh well [58:19] so you have a machine learning model [58:20] which it's an open pilot it's a [58:22] you know it's a blob it's just a blob of [58:25] weights it's not like people are like oh [58:26] it's closed source i'm like it's a blob [58:28] of weights what do you expect [58:29] um you know primarily neural network [58:32] based [58:33] ul open pilot is all the software kind [58:36] of around that neural network that if [58:37] you have a neural network that says [58:38] here's where you want to send the car [58:40] openpilot actually goes and executes all [58:43] of that [58:44] it cleans up the input to the neural [58:46] network it cleans up the output and [58:48] executes on it so it connects [58:50] it's the glue that connects everything [58:51] together runs the sensors does a bunch [58:53] of calibration for the neural network [58:55] does you know deals with like you know [58:57] if the car is on a banked road [58:59] uh you have to counter steer against [59:01] that and the neural network can't [59:02] necessarily know that by looking at the [59:04] picture [59:06] so you do that with with other sensors [59:08] infusion and localizer [59:09] openpilot also is responsible for [59:12] sending the data up to our servers so we [59:14] can learn from it [59:16] logging it recording it running the [59:18] cameras thermally managing the device [59:21] managing the disk space on the device [59:23] managing all the resources of the device [59:24] so what um since we last spoke i don't [59:26] remember when maybe a year ago maybe a [59:28] little bit longer [59:30] how uh has open pilot improved we did [59:33] exactly what i promised you [59:34] i promised you that by the end of the [59:36] year you'd be able to remove the lanes [59:39] um the lateral policy is now [59:43] uh almost completely end to end you can [59:46] turn the lanes off and it will drive [59:48] drive slightly worse on the highway if [59:49] you turn the lanes off but you can turn [59:51] the lanes off and it will drive [59:53] well trained completely end to end on [59:56] user data [59:57] um and this year we hope to do the same [59:58] for the longitudinal policy so that's [60:00] the interesting thing [60:01] is you're not doing you don't appear to [60:04] be you can correct me you don't appear [60:05] to be doing lean [60:07] detection or lane marking detection or [60:10] kind of the segmentation task or [60:13] any kind of object detection task you're [60:15] doing what's [60:16] traditionally more called like [60:17] end-to-end learning [60:19] so and trained on actual behavior of [60:23] drivers [60:24] when they're driving the car manually [60:27] and this is hard to do you know it's not [60:30] supervised learning [60:32] yeah but uh so the nice thing is there's [60:33] a lot of data so it's hard and easy [60:36] right it's uh we have a lot of high [60:38] quality data yeah [60:39] like more than you need in the senate [60:41] well we way more than we do we have way [60:43] more data than we need [60:44] i mean it's it's an interesting question [60:46] actually because [60:48] in terms of amount you have more than [60:49] you need [60:51] but the you know driving is full of edge [60:53] cases so how do you [60:55] select the data you train on [60:58] i i think this is an interesting open [61:00] question like what's [61:01] what's the cleverest way to select data [61:04] that's the question tesla is probably [61:05] working on [61:07] uh that's i mean the entirety of machine [61:09] learning can be they don't seem to [61:10] really care they just kind of select [61:11] data but i feel like that [61:13] if you want to solve if you want to [61:15] create intelligent systems you have to [61:16] pick data well [61:18] right and so would you have any hints [61:21] ideas of how to do it well [61:22] so in some ways that is the definition i [61:26] like of reinforcement learning versus [61:27] supervised learning [61:29] in supervised learning the weights [61:31] depend on the data [61:32] right um and this is obviously true but [61:35] the [61:36] uh in reinforcement learning the data [61:38] depends on the weights [61:40] yeah right and actually both ways that's [61:42] that's poetry [61:43] so it's brilliant how does it know what [61:45] data to turn on well let it pick [61:47] we're not there yet but that's the [61:48] eventual so you're thinking this [61:50] almost like a reinforcement learning [61:52] framework we're going to do rl on the [61:54] world [61:55] every time a car makes a mistake user [61:57] disengages we [61:58] train on that and do our all in the [61:59] world ship out a new model that's an [62:01] epoch right [62:03] and uh for now you're not [62:06] doing the elon style promising that it's [62:08] going to be fully autonomous you really [62:10] are sticking to level two [62:12] and like it's supposed to be supervised [62:14] oh it is definitely supposed to be [62:16] supervising reinforced the fact that [62:17] it's supervised [62:18] um we look at our rate of improvement [62:22] in disengagements um open pilot now has [62:24] an unplanned engagement about every 100 [62:26] miles [62:27] this is up from 10 miles like [62:30] maybe maybe uh [62:35] maybe a year ago yeah so maybe we've [62:37] seen 10x improvement in a year but [62:39] a hundred miles is still a far cry from [62:41] the hundred thousand you're going to [62:42] need [62:43] so you're going to somehow need to get [62:45] um [62:46] three more 10xs in there and your [62:50] what's your intuition uh you're [62:52] basically hoping that there's [62:53] exponential improvement built into the [62:55] baked into the cake somewhere well [62:56] that's even like i mean 10x improvement [62:58] that's already assuming exponential [63:00] right there's definitely exponential [63:01] improvement and i think when elon talks [63:03] about exponential like [63:04] these things these systems are going to [63:06] exponentially improve [63:07] just exponential doesn't mean you're [63:09] getting 100 gigahertz processors [63:11] tomorrow [63:12] right like it's going to still take a [63:14] while because the gap [63:16] between even our best system and humans [63:18] is still large [63:20] so that's an interesting distinction to [63:21] draw so if you look at the way tesla's [63:23] approaching the problem [63:26] and the way you're approaching the [63:27] problem which is very different than the [63:29] rest of the [63:30] self-driving car world so let's put them [63:33] aside [63:33] is you're treating most the driving task [63:36] is a machine learning problem [63:37] and the way tesla is approaching it is [63:39] with the multi-task learning [63:40] where you break the task of driving into [63:43] hundreds of different desks [63:45] and you have this multi-headed neural [63:47] network that's very good at [63:49] performing each task and there there's [63:52] presumably something on top that's [63:54] stitching stuff together in order to uh [63:58] make controlled decisions policy [64:00] decisions about how you move the car [64:02] but what that allows you there's a [64:03] brilliance to this because it allows you [64:05] to [64:05] um master each task like [64:08] lane detection uh stop sign detection [64:13] traffic light detection uh drivable area [64:16] segmentation [64:18] uh you know vehicle bicycle pedestrian [64:20] detection [64:22] uh there's some localization tasks in [64:24] there [64:25] also predicting of [64:29] like yeah predicting how the the [64:32] entities in the scene are going to move [64:33] like everything is basically a machine [64:35] learning task well there's a [64:36] classification [64:37] segmentation prediction and [64:41] it's nice because you can have this [64:43] entire engine data engine that's mining [64:45] for edge cases [64:48] for each one of these tasks and you can [64:49] have people like engineers that are [64:51] basically masters of that task [64:53] like becoming the best person in the [64:55] world that uh as you talk about the cone [64:57] guy [64:57] for uh for for waymo the good old phone [65:00] guy [65:01] the the becoming the best person in the [65:02] world at uh at uh [65:05] at cone detection i i so that's a [65:07] compelling notion from a supervised [65:09] learning perspective [65:12] automating much of the process of [65:14] educates discovery and retraining neural [65:16] network for each of the individual [65:18] perception tasks and then you're looking [65:21] at the machine learning [65:22] in a more holistic way basically doing [65:25] end-to-end learning [65:26] on the driving task supervised trained [65:30] on the data of the actual [65:32] driving of people that use comma ai [65:36] like actual human drivers do manual [65:37] control plus the [65:40] moments of disengagement that uh [65:43] maybe with some labeling could indicate [65:46] the failure of the system so you have [65:47] the [65:48] you have a huge amount of data for [65:50] positive [65:51] control of the vehicle like successful [65:53] control of the vehicle [65:55] both maintaining the lane as as i think [65:58] you're also working on [65:59] longitudinal control of the vehicle and [66:02] then failure cases where the vehicle [66:04] does something [66:05] wrong that needs disengagement [66:08] so like what why do you think you're [66:10] right and tesla is wrong [66:13] on this and do you think do you think [66:15] you'll come around the tesla way do you [66:17] think tesla will come around [66:19] to your way if you were to start a chess [66:22] engine company [66:23] would you hire a bishop guy see we have [66:27] uh this is monday morning quarterbacking [66:30] as uh yes probably [66:33] [Laughter] [66:35] so oh oh our rook guy oh we stole the [66:38] rook guy from that company oh we're [66:39] gonna have real good [66:40] rooks well there's not many pieces right [66:43] you can uh yeah there's not many [66:47] guys and gals to hire you just have a [66:49] few that work on the bishop a few that [66:51] work in the rook [66:52] but is that not ludicrous today to think [66:54] about in in the world of alpha zero [66:57] but alpha zero is jessica so the the [66:59] fundamental question [67:00] is how hard is driving compared to chess [67:04] because so long term end to end [67:08] will be the right solution the question [67:11] is how many years away is that [67:13] end-to-end is going to be the only [67:14] solution for level five for the only way [67:16] we can [67:17] of course and of course tesla's gonna [67:18] come around to my way and if you're a [67:20] rook guy out there i'm sorry [67:22] the cone guy i don't know we're gonna [67:25] specialize each task we're gonna really [67:27] understand [67:28] rook placement yeah i understand the [67:31] intuition you have [67:32] i mean that [67:35] that uh is very compelling notion that [67:36] we can learn the task and to end [67:39] like the same compelling notion you [67:40] might have for natural language [67:41] conversation [67:42] i'm not sure [67:46] because one thing you sneaked in there [67:48] is the assertion that [67:50] it's impossible to get to level five [67:53] without this kind of approach [67:55] i don't know if that's obvious i don't [67:57] know if that's obvious either i don't [67:59] actually uh mean that i think that it is [68:02] much easier [68:03] to get to level five with an end-to-end [68:04] approach i think that the other approach [68:06] is [68:07] doable but the magnitude of the [68:10] engineering challenge may exceed what [68:11] humanity is capable of [68:13] so but what do you think of the tesla [68:17] data engine approach which to me is an [68:20] active learning task is kind of [68:21] fascinating [68:23] is breaking it down into these multiple [68:24] tasks and [68:26] mining their data constantly for like [68:28] edge cases [68:29] for these different tasks but the tasks [68:31] themselves are not being learned this is [68:32] feature engineering [68:35] yeah i mean it's it's a it's a higher [68:38] abstraction level of feature engineering [68:41] for the different tasks [68:43] it's task engineering in a sense it's [68:45] slightly better feature engineering but [68:46] it still fundamentally is feature [68:48] engineering [68:49] and anything about the history of ai has [68:51] taught us anything [68:52] it's that feature engineering approaches [68:54] will always be replaced and lose [68:56] to end to end now to be fair i cannot [68:59] really make promises on timelines [69:02] but i can say that when you look at the [69:04] code for stockfish and the code for [69:06] alpha zero [69:06] one is a lot shorter than the other a [69:09] lot more elegant required a lot less [69:10] programmer hours to write [69:12] yeah but there was a lot more murder [69:16] of bad uh [69:20] agents on the uh alpha zero side [69:24] by murder i mean uh agents that played a [69:28] game and failed miserably [69:30] yeah oh in simulation that failure is [69:34] less costly yeah in in real world it's [69:37] wait do you mean in practice like alpha [69:38] zero has lost games miserably [69:40] no well i haven't seen that no but i [69:43] know but the the [69:44] the the requirement for alpha zero is a [69:47] simulator [69:48] to be able to like evolution human [69:50] evolution [69:51] not human evolution biological evolution [69:53] of life on earth from the origin of life [69:56] has murdered trillions upon trillions of [69:59] organisms [70:00] on the path to us humans yeah so the [70:03] question is can [70:04] can we uh stitch together a human-like [70:06] object without having to go through the [70:08] entirety process of evolution [70:09] well no but do the evolution in [70:11] simulation yeah that's the question can [70:12] we simulate so do you have a sense [70:14] that's possible to simulate some [70:16] mu zero is exactly this mu zero is [70:19] is the solution to this mu zero i think [70:22] is going to look be looked back [70:23] as the canonical paper and i don't think [70:25] deep learning is everything i think that [70:27] there's still a bunch of things missing [70:28] to get there [70:29] but mu zero i think is going to be [70:30] looked back as the kind of cornerstone [70:32] paper [70:33] um of this whole deep learning era [70:36] and mu zero is the solution to [70:38] self-driving cars you have to make a few [70:40] tweaks to it [70:41] but mu0 does effectively that it does [70:42] those roll outs [70:44] and those murdering in in a learned [70:47] simulator in a learned dynamics model [70:50] it's interesting it doesn't get enough [70:51] love i was blown away when i i was blown [70:53] away when i read that paper i'm like [70:54] you know okay i've always had a comma [70:56] i'm gonna sit and i'm gonna wait for the [70:58] solution to self-driving cars to come [70:59] along [71:00] this year i saw it it's me zero yeah [71:05] so uh sit back and let the winning roll [71:08] in [71:09] so your sense just to elaborate a little [71:11] bit [71:12] to link on the topic your senses in your [71:14] networks will solve driving [71:16] yes like we don't need anything else i [71:18] think the same way chess was maybe the [71:20] chess and maybe google are the pinnacle [71:22] of like [71:24] search algorithms and things that look [71:25] kind of like a star [71:27] um the pinnacle of [71:30] this era is going to be self-driving [71:32] cars [71:34] but on the path of that you have to [71:36] deliver products [71:38] and it's possible that the path to full [71:41] self-driving cars [71:42] will take decades i doubt it so how long [71:45] would you put on it [71:47] like what what are we you're chasing it [71:50] tesla's chasing it what are we talking [71:54] about five years 10 years [71:55] 50 years in the 2020s in the 2020s [71:59] the later part of the 2020s [72:03] with the neural network well that would [72:05] be nice to see and on the path to that [72:07] you're delivering products [72:09] which is a nice l2 system that's what [72:10] tesla's doing a nice l2 system i'm just [72:13] going to do better every time [72:14] l2 the only difference between l2 and [72:16] the other levels is who takes liability [72:17] and [72:17] i'm not a liability guy i want to take [72:19] liability level 2 forever [72:22] now on that little transition i mean [72:27] how do you make the transition work is [72:29] uh is this where driver sensing [72:31] comes in like how do you make the [72:34] because you said [72:34] 100 miles like is is there some [72:39] sort of human factor psychology thing [72:41] where people start to over trust the [72:42] system all those kinds of [72:44] effects once it gets better and better [72:46] and better and better they get lazier [72:47] and lazier and lazier [72:49] is that like how do you get that [72:50] transition right first off our [72:52] monitoring is already adaptive [72:54] our monitoring is already seen adaptive [72:56] driver monitoring [72:58] is this the camera that's looking at the [72:59] driver you have an infrared camera [73:01] in the our policy for [73:04] how we enforce the driver monitoring is [73:06] scene adaptive what's that mean [73:08] well for example in one of the extreme [73:10] cases um [73:12] if you uh if the car is not moving we do [73:15] not [73:15] uh actively enforce driver monitor right [73:18] um if you are going through a [73:21] uh like a 45 mile an hour road with [73:24] lights [73:25] um and stop signs and potentially [73:27] pedestrians we enforce a very tight [73:29] driver monitoring policy [73:30] if you are alone on a perfectly straight [73:32] highway [73:33] um and this is it's all machine learning [73:35] none of that is hand coded [73:36] actually the stop is hand coded but so [73:39] there's some kind of machine learning [73:41] estimation of risk yes yeah i mean i've [73:44] always been a huge fan of that that [73:46] that's uh because it's difficult to do [73:50] every step into that direction [73:53] is a worthwhile step to take it might be [73:55] difficult to do really well like us [73:56] humans are able to [73:57] estimate risk pretty damn well whatever [74:00] the hell that is [74:01] that feels like one of the nice features [74:04] of us humans [74:05] uh because like we humans are really [74:08] good drivers when we're really like [74:10] tuned in and we're good at estimating [74:12] risk like when are we supposed to be [74:14] tuned in [74:14] yeah and you know people are like oh [74:17] well you know why would you ever make [74:18] the driver monitoring policy [74:20] less aggressive why would you always not [74:22] keep it at its most aggressive [74:23] because then people are just going to [74:24] get fatigued from it yes when they get [74:26] annoyed [74:27] you want them yeah you they want you [74:28] want the experience to be pleasant [74:30] obviously i want the experience to be [74:32] pleasant but even just from a straight [74:34] up safety perspective [74:35] if you alert people when they look [74:38] around and they're like [74:39] why is this thing alerting me there's [74:41] nothing i could possibly hit right now [74:42] people will just learn to tune it out [74:45] people will just learn to tune it out to [74:46] put weights on the steering wheel to do [74:48] whatever to overcome it [74:49] and remember that you're always part of [74:52] this adaptive system so [74:53] all i can really say about you know how [74:55] this scale is going forward is yeah [74:57] something we have to monitor for [74:59] we don't know this is a great psychology [75:01] experiment at scale like we'll see [75:03] yeah it's fascinating track it and [75:04] making sure you have a good [75:06] understanding of attention is a very key [75:09] part of that psychology problem [75:11] yeah i think i mean you and i probably [75:13] have a different come to it differently [75:15] but to me [75:16] it's an it's a fascinating psychology [75:18] problem to explore something much deeper [75:20] than just driving [75:22] it's a it's such a nice way to explore [75:24] human [75:25] attention and human behavior [75:28] which is why again we've probably both [75:31] criticized [75:32] mr elon musk on this one topic from [75:35] different [75:36] avenues uh so both offline and online i [75:39] had little chats with elon and [75:44] like i love human beings as a as a [75:47] as a computer vision problem as an ai [75:49] problem it's fascinating [75:51] he wasn't so much interested in that [75:52] problem [75:54] it's like in order to solve driving the [75:56] whole point is you want to remove the [75:58] human from the picture [76:01] and it seems like you can't do that [76:03] quite yet eventually yes but [76:05] you can't quite do that yet so [76:09] this is the moment where and you can't [76:11] yet say [76:12] i told you so uh to tesla [76:16] but it's getting there because i don't [76:18] know if you've seen this there's some [76:19] reporting that they're in fact starting [76:21] to do [76:22] drive them off yeah they ship the model [76:23] in shadow mode [76:25] uh without uh i believe only a visible [76:28] light camera it might even be fisheye [76:30] uh it's like a low resolution low [76:33] resolution visible light [76:34] i mean to be fair that's what we have in [76:36] the eon as well our last generation [76:38] product [76:38] this is the one area where i can say our [76:41] hardware's ahead of tesla the rest of [76:42] our hardware way way behind but our [76:44] driver monitoring camera [76:45] do you think uh i think on the [76:48] third row tesla podcast or somewhere [76:51] else i've heard you say that [76:53] obviously eventually they're gonna have [76:55] driver monitoring [76:57] i think what i've said is elon will [76:58] definitely ship driver monitoring before [77:00] he ships level five the beautiful level [77:02] and i'm willing to about 10 grand on [77:03] that and you better ground on that [77:06] uh i mean now i want to take the bet but [77:08] before maybe someone would have i should [77:09] have got my money [77:10] yeah that's an interesting bet i think i [77:13] i think you're right i'm actually [77:17] on a human level because he's been [77:21] he's made the decision like he said that [77:25] driver monitoring is the wrong way to go [77:27] but like you have to think of as a human [77:30] as a ceo [77:31] i think that's the right thing to say [77:34] when like [77:37] sometimes you have to say things [77:39] publicly they're different than when you [77:40] actually believe [77:41] because when you're producing a large [77:43] number of vehicles [77:45] and the decision was made not to include [77:47] the camera like what are you supposed to [77:48] say [77:48] yeah like our cars don't have the thing [77:51] that i think is right to have [77:53] uh it's an interesting thing but like on [77:56] the other side as a ceo i mean something [77:58] you could probably speak to [78:00] as a leader i think about me as a human [78:04] to publicly change your mind on [78:06] something how hard is that [78:08] well especially when like [78:09] george haas say i told you so [78:12] all i will say is i am not a leader and [78:14] i am happy to change my mind [78:16] um and i think elon will [78:20] yeah i do i think he'll come up with a [78:23] good way to [78:24] make it psychologically okay for him [78:27] well [78:27] it's such an important thing man [78:29] especially for a first principles [78:30] thinker because he made a decision [78:33] that uh driver monitoring is not the [78:35] right way to go and i could see that [78:36] decision [78:37] and i i could even make that decision [78:39] like i was on the fence [78:41] too like i'm not driving monitoring [78:45] is such an obvious simple solution to [78:48] the problem of attention [78:49] it's not obvious to me that just by [78:51] putting a camera there you solve [78:53] things you have to create an [78:56] incredible compelling experience just [78:59] like you're you're talking about [79:00] and i don't know if it's easy to do that [79:03] it's not at all easy to do that in fact [79:05] i think [79:05] so as a creator of a car that's trying [79:09] to [79:10] create a product that people love which [79:12] is what tesla tries to do right [79:14] it's not obvious to me that uh you know [79:17] as a design decision whether adding a [79:19] camera [79:19] is a good idea from a safety perspective [79:22] either like [79:23] in the human factors community everybody [79:25] says [79:26] that like you should obviously have [79:27] driver sensing driving monitoring [79:30] but like that [79:34] that's like saying it's obvious as [79:36] parents you shouldn't let your kids go [79:38] out at night [79:39] but okay but like [79:43] they're still gonna find ways to do [79:45] drugs [79:47] yeah you have to also be good parents so [79:50] like it's it's much more complicated [79:51] than just like [79:52] you need to have drive and monitoring i [79:54] totally disagree on [79:56] okay if you have a camera there and the [79:59] camera's watching the person [80:00] but never throws an alert they'll never [80:02] think about it [80:03] right the the driver monitoring policy [80:07] that you choose to how you choose to [80:09] communicate with the user is [80:10] entirely separate from the data [80:13] collection perspective right [80:15] right so you know like [80:19] there's one thing to say like you know [80:22] tell your teenager they can't do [80:23] something [80:24] there's another thing to like you know [80:26] gather the data so you can [80:27] make informed decisions that's really [80:29] interesting but you have to make that [80:31] that's the interesting thing about cars [80:33] uh but even true with com ai [80:35] like you don't have to manufacture the [80:37] thing into the car [80:38] is you have to make a decision that [80:40] anticipates [80:41] the right strategy long term so like you [80:44] have to start collecting the data [80:46] and start making decisions starter date [80:48] started it three years ago [80:50] i believe that we have the best driver [80:52] monitoring solution in the world [80:54] um i think that when you compare it to [80:56] well supercruise is the only other one [80:58] that i really know that shipped [80:59] and ours is better what uh [81:03] what do you like and not like about [81:04] super coos um i mean i had a few [81:07] super crews uh the sun would be shining [81:11] through the window [81:11] would blind the camera and it would say [81:13] i wasn't paying attention when i was [81:14] looking completely straight [81:16] i couldn't reset the attention with a [81:17] steering wheel touch and supercrews [81:19] would disengage [81:20] like i was communicating to the car i'm [81:22] like look i'm here i'm paying attention [81:24] why are you really gonna force me to [81:25] disengage and [81:26] it did um so it's it's a constant [81:30] conversation with the user and yeah [81:32] there's no way to ship a system like [81:33] this if you can ota [81:35] right we're shipping a new one every [81:36] month sometimes we we [81:38] balance it with our users on discord [81:40] like when sometimes we make the driver [81:41] monitoring a little more aggressive and [81:42] people complain [81:43] sometimes they don't you know we want it [81:45] to be as aggressive as possible where [81:47] people [81:47] don't complain it doesn't feel intrusive [81:49] so being able to update the system over [81:50] the air is an essential component [81:52] i mean that's probably to me you [81:54] mentioned uh [81:56] i mean to me that is the biggest [81:59] innovation of tesla [82:00] that it it made it people realize that [82:03] over-the-air updates [82:04] is essential yeah i mean [82:07] yeah was that not obvious from the [82:09] iphone the iphone was the first real [82:11] product that ota'ed i think [82:12] was it actually that's that's brilliant [82:14] you're right i mean the game consoles [82:15] used to not right [82:16] the game consoles are maybe the second [82:18] thing that did well i didn't really [82:19] think about [82:20] one of the amazing features of a [82:22] smartphone [82:24] isn't just like the touchscreen isn't [82:26] the thing [82:27] it's the ability to constantly update [82:30] yeah [82:30] get better it gets better [82:34] i love my ios 14. uh [82:38] one thing that i probably disagree with [82:41] you on on driving monitoring [82:43] is you've said that it's easy like i [82:46] mean you you tend to say stuff is easy [82:50] the uh i guess you said it's easy [82:52] relative to the [82:53] external perception problem [82:58] can you elaborate why you think it's [82:59] easy feature engineering works [83:02] for driver monitoring feature [83:03] engineering does not work for the [83:04] external [83:05] so human faces are not [83:08] human faces and the movement of human [83:11] faces and head [83:12] and body is not as variable as the [83:14] external environment [83:16] is your intuition um yes and there's [83:17] another big difference as well um your [83:20] reliability of a driver monitoring [83:21] system [83:22] doesn't actually need to be that high [83:24] the uncertainty [83:25] if you have something that's detecting [83:27] whether the human is paying attention it [83:28] only works 92 [83:29] of the time you're still getting almost [83:31] all the benefit of that because [83:33] the human like you're training the human [83:34] yeah right you're you're dealing with a [83:36] system that's [83:37] really helping you out it's a [83:39] conversation it's not [83:40] like the external thing where guess what [83:43] if you swerve into a tree you swerve [83:44] into a tree [83:45] right like you get no margin for error [83:48] yeah i think that's really well put i [83:49] i i think that's the right exactly the [83:53] place [83:54] where um where comparing to the external [83:58] perception the control problem [84:00] driver monitoring is easier because you [84:01] know the bar for success is much lower [84:05] yeah but i i still think like the human [84:07] face [84:09] is more complicated actually than the [84:11] external environment but for driving you [84:12] don't give a damn [84:14] i don't need you i don't need something [84:15] i don't need something that complicated [84:18] to to to have to communicate the idea [84:21] to the human that i want to communicate [84:23] which is yo [84:24] system might mess up here you got to pay [84:26] attention [84:27] yeah see that's that's my love and [84:30] fascination is the [84:31] the human face and it feels like this is [84:34] a nice place to um create products that [84:38] create an experience in the car [84:39] so like it feels like there should be [84:42] more richer experiences [84:45] in the car you know like that's an [84:48] opportunity for [84:49] like something like my eye or just any [84:52] kind of system like a tesla or any of [84:54] the autonomous vehicle companies [84:56] is because software's and there's much [84:58] more sensors [84:59] and so much is running on software and [85:00] you're doing machine learning anyway [85:02] there's an opportunity to create totally [85:04] new experiences [85:06] that we're not even anticipating you [85:08] don't think so [85:10] no you think it's a box that gets you [85:12] from a to b and you want to do it [85:13] chill yeah i mean i think as soon as we [85:16] get to level three on highways okay [85:17] enjoy your candy crush [85:19] enjoy your hulu enjoy your [85:22] you know whatever whatever sure you get [85:24] this you can look at screens basically [85:26] versus right now what do you have music [85:27] and audio books so level three is where [85:29] you can kind of disengage in [85:30] in stretches of time [85:33] [Music] [85:33] [Applause] [85:34] well you think level three is possible [85:37] like on the highway going for 100 miles [85:39] and you can just go to sleep oh yeah [85:41] uh sleep so again [85:44] i think it's really all on a spectrum i [85:47] think that [85:47] being able to use your phone while [85:50] you're on the highway [85:51] and like this all being okay and being [85:53] aware that the car might alert you and [85:55] you have five seconds to basically [85:56] so the five second thing that you think [85:58] is possible yeah i think it is oh yeah [86:00] not not in all scenarios right some [86:02] scenarios it's not [86:03] it's the whole risk thing that you [86:05] mentioned is nice is to be able to [86:06] estimate like [86:08] how risky is this situation exactly [86:10] that's really important to understand [86:12] one other thing you mentioned comparing [86:14] comma and [86:16] autopilot is that um [86:19] something about the haptic feel of the [86:22] way combo controls the car [86:24] when things are uncertain like it [86:26] behaves a little bit more uncertain when [86:27] things are uncertain [86:29] that's kind of an interesting point and [86:31] then autopilot is much more confident [86:33] always even when it's uncertain until it [86:36] runs into trouble [86:38] yeah that's that's a funny thing i [86:40] actually mentioned that to elon i think [86:42] and then the first time we talked he [86:45] wasn't biting [86:46] is like communicating uncertainty i [86:48] guess comet doesn't really communicate [86:51] uncertainty explicitly communicates it [86:53] through haptic feel [86:55] like what what's the role of [86:56] communicating uncertainty do you think [86:57] oh we do some stuff explicitly like we [87:00] do detect the lanes when you're on the [87:01] highway and we'll show you how many [87:02] lanes we're using to drive with [87:04] you can look at where it thinks the [87:05] lanes are you can look at um [87:07] the path and yeah we want to be better [87:09] about this we're actually hiring [87:11] i want to hire some new ui people ui [87:13] people you mentioned this because it's [87:14] such an [87:15] it's a ui problem too right it's we're [87:17] we have a great designer now [87:19] but you know we need people who are just [87:20] gonna like build this and debug these [87:21] uis [87:22] qt people and okay is that what the ui [87:25] has done with this key [87:26] we're moving the new ui is in qt c plus [87:29] plus qt [87:32] uses it yeah we had some react stuff in [87:35] there [87:37] react js would just react react is his [87:40] own language right [87:41] react native reaction react to the [87:43] javascript framework [87:44] yeah it's all it's all based on [87:46] javascript but it's [87:47] you know i like c plus plus [87:51] what do you think about uh dojo with [87:54] tesla [87:55] and their foray into what appears to be [88:00] specialized hardware for uh training [88:02] neonets [88:05] i guess it's something maybe you can [88:06] correct me from my [88:08] shallow looking at it it seems like [88:10] something like google did with tpus but [88:12] specialized [88:13] for uh driving data i don't think it's [88:16] specialized for driving data [88:18] it's just legit just tpu they want to [88:20] inter go the apple way basically [88:22] everything required in the chain is done [88:24] in-house well so you have a problem [88:27] right now and [88:28] this is one of my one of my concerns i [88:31] really would like to see [88:32] somebody deal with this if anyone out [88:34] there is doing it i'd like to help them [88:36] if i can [88:37] um you basically have two options right [88:39] now to train [88:41] your options are nvidia or google [88:45] so google is not even an option [88:50] their tpus are only available in google [88:51] cloud [88:53] google has absolutely onerous terms of [88:55] service restrictions [88:57] uh they may have changed it but back in [88:59] google's terms of service it said [89:00] explicitly you are not allowed to use [89:02] google cloud ml [89:03] for training autonomous vehicles or for [89:05] doing anything that competes with google [89:07] without google's prior written [89:08] permission [89:09] well okay i mean google is not a [89:11] platform company [89:12] uh i wouldn't i wouldn't touch tpus with [89:15] a 10-foot pole [89:16] so that leaves you with the monopoly uh [89:19] nvidia and video [89:20] so i mean that you're not a fan of well [89:24] look i was a huge fan of in 2016 nvidia [89:28] jensen came sat in the car um [89:31] cool guy when the stock was 30 a share [89:35] uh nvidia stock has skyrocketed i [89:38] witnessed a real [89:39] change in who was in management over [89:41] there in like 2018 [89:43] and now they are let's exploit let's [89:46] take every dollar we possibly can out of [89:48] this ecosystem [89:49] let's charge ten thousand dollars for [89:50] a100s because we know we got the best [89:52] in the game [89:54] and let's charge ten thousand dollars [89:56] for an a100 [89:57] when it's really not that different from [89:59] 3080 which is 699. [90:02] um the margins that they are making off [90:05] of those high-end chips are [90:06] so high that i mean i think they're [90:09] shooting themselves in the foot just [90:10] from a business perspective [90:12] because there's a lot of people talking [90:13] like me now who are like somebody's got [90:16] taken video down [90:18] yeah where they could dominate it nvidia [90:21] could be the new intel [90:22] yeah to be in inside everything [90:25] essentially [90:26] and and yet the winners in in certain [90:30] spaces like [90:31] autonomous driving the winners only the [90:34] people who are like [90:35] desperately falling back and trying to [90:36] catch up and have a ton of money [90:38] like the big automakers are the ones [90:40] interested in partnering with nvidia [90:42] oh and then i think a lot of those [90:44] things are going to fall through if i [90:46] were nvidia [90:47] sell chips sell chips at a reasonable [90:50] markup [90:52] to everybody to everybody without any [90:54] restrictions without any restrictions [90:56] intel did this look at intel they had a [90:58] great long run [90:59] nvidia is trying to turn their they're [91:01] like trying to productize their chips [91:04] way too much they're trying to extract [91:06] way more value than they can sustainably [91:09] sure you can do it tomorrow is it going [91:10] to up your share price sure if you're [91:12] one of those ceos [91:13] like how much can i strip mine this [91:14] company and you know and that's what's [91:16] weird about it too [91:17] like the ceo is the founder it's the [91:19] same guy yeah i mean i still think [91:21] jensen's a great guy that's great why do [91:24] this [91:25] you have a choice you have a choice [91:27] right now are you trying to cash out are [91:28] you trying to buy a yacht [91:30] if you are fine but if you're trying to [91:33] be [91:34] the next huge semiconductor company sell [91:36] chips [91:37] well the the interesting thing about [91:38] jensen is he is [91:40] a big vision guy so he has [91:44] a plan like for 50 years down the road [91:48] so it makes me wonder like how does [91:50] price gouging fit into it yeah how does [91:52] that [91:53] like it's it doesn't seem to make sense [91:55] to plan [91:56] i worry that he's listening to the wrong [91:58] people yeah that that's the sense i have [92:01] too sometimes because i despite [92:04] everything [92:05] i think nvidia is an incredible company [92:08] well one [92:09] sort of i'm deeply grateful to nvidia [92:12] for the products they've created me to [92:13] pass right [92:14] and so the 1080 ti was a great gpu still [92:18] have a lot of them [92:18] it was yeah but [92:22] at the same time it just feels like [92:26] feels like you don't want to put all [92:28] your stock in nvidia and so like elon [92:30] is doing um what tesla is doing with [92:32] autopilot and dojo [92:34] is the apple way is because they're not [92:37] going to share dojo with uh george hotz [92:40] uh [92:40] i i know they should sell that chip [92:43] oh they should sell their even their [92:45] their accelerator the accelerator that's [92:47] in all the cars the 30 watt one [92:49] sell it why not so open it up [92:52] make me why does this have to be a car [92:54] company [92:55] well if you sell the chip here's what [92:57] you get yeah [92:58] make some money off the chips it doesn't [93:00] take away from your chip [93:01] you're going to make some money free [93:02] money and also [93:05] the world is going to build an ecosystem [93:07] of tooling for you [93:08] right you're not going to have to fix [93:10] the bug in your tanh layer [93:12] someone else already did well the [93:15] question that's an interesting question [93:16] i mean that's the [93:17] question steve jobs asked that's the [93:18] question elon musk is [93:21] uh perhaps asking is uh [93:25] do you want tesla stuff inside other [93:26] vehicles in [93:28] inside potentially inside like uh irobot [93:31] vacuum cleaner [93:32] yeah i think you should decide where [93:35] your advantages are [93:37] i'm not saying tesla should start [93:38] selling battery packs to automakers [93:40] because battery packs to automakers they [93:41] are straight up in competition with you [93:43] if i were tesla i'd keep the battery [93:44] technology totally yeah ssr's we make [93:47] batteries [93:47] but the thing about the tesla tpu um [93:51] is anybody can build that it's just a [93:53] question of you know [93:54] are you willing to spend the you know [93:56] the money it could be a huge source of [93:58] revenue potentially [94:00] are you willing to spend 100 million [94:01] dollars right anyone can build it [94:03] and someone will and a bunch of [94:05] companies now are starting trying to [94:06] build ai accelerators somebody's going [94:08] to get the idea right [94:10] and yeah hopefully they don't get greedy [94:13] because they'll just lose to the next [94:14] guy who finally and then eventually the [94:16] chinese are going to make knock off and [94:17] video chips and that's [94:19] from your perspective i don't know if [94:20] you're also paying attention to stan [94:22] tesla for a moment [94:23] all dave elon musk has talked about a [94:26] complete rewrite [94:28] of uh the neural net that they're using [94:31] that seems to [94:32] again i'm half paying attention but it [94:35] seems to involve [94:36] basically a kind of integration of all [94:39] the sensors to where [94:42] it's a four dimensional view you know [94:44] you have a 3d [94:45] model of the world over time and then [94:47] you can i think it's done [94:49] both for the for actually you know [94:53] so the neural network is able to in a [94:55] more holistic way deal with the world [94:57] and make predictions and so on [94:59] but also to make the annotation task [95:02] more uh [95:03] you know easier like you can annotate [95:06] the world [95:07] in one place and then kind of distribute [95:09] itself across the sensors and [95:10] across a different like the hundreds of [95:13] tasks that are involved [95:15] in the hydronet what are your thoughts [95:18] about this rewrite is it just like some [95:19] details [95:21] that are kind of obvious there are steps [95:22] that should be taken or is there [95:24] something fundamental [95:26] that could challenge your idea that end [95:27] to end is the right solution [95:30] uh we're in the middle of a bakery right [95:32] now as well we haven't shipped a new [95:33] model in a bit [95:34] of what kind uh we're going from 2d to [95:37] 3d [95:38] right now all our stuff like for example [95:39] when the car pitches back the lane lines [95:41] also pitch back [95:42] uh because we're assuming the flat ro [95:45] flat world hypothesis [95:46] uh the new models do not do this the new [95:48] models output everything in 3d [95:50] um so this but there's still no [95:52] annotation so the 3d is [95:54] it's more about the opposite yeah uh [95:57] we have we have disease and everything [95:59] uh we've disease [96:01] yeah we had a disease we had a disease [96:03] um we unified [96:04] a lot of stuff as well uh we switched [96:06] from tensorflow to pi torch [96:08] yes uh my understanding [96:11] of what tesla's thing is is that their [96:14] annotator now annotates across the time [96:16] dimension [96:17] uh i mean cute [96:21] why are you building an annotator i i [96:24] find their entire pipeline [96:28] i find your vision i mean the vision [96:30] event to end [96:31] very compelling but i also like the [96:33] engineering of the data engine that [96:35] they've created [96:37] in terms of supervised learning [96:40] pipelines [96:41] that thing is damn impressive you're [96:44] basically the [96:44] the idea is that you have hundreds of [96:48] thousands of people that are doing data [96:50] collection for you by doing their [96:51] experience so that's kind of similar to [96:53] the common ai model [96:55] and you're able to [96:58] mine that data based on the kind of edge [97:00] cases you need [97:02] i i think it's harder to do in the end [97:05] to end learning [97:07] the mining of the right edge cases like [97:09] that's where feature engineering [97:11] is actually really powerful because [97:15] like us humans are able to do this kind [97:17] of mining a little better but but [97:20] yeah there's obvious as we as we know [97:21] there's obvious constraints and [97:23] limitations to that [97:24] idea uh carpathi just tweeted he's like [97:27] um [97:28] you get really interesting insights if [97:29] you saw if you sort your validation set [97:32] by loss [97:33] and look at the highest loss examples [97:36] yeah [97:37] uh so yeah i mean you can do we we have [97:40] we have a little data engine like thing [97:41] we're training a segment uh [97:43] it's not fancy it's just like okay train [97:46] the new segment [97:48] run it on 100 000 images and now take [97:50] the thousand with highest loss [97:52] select 100 of those by human put those [97:54] get those ones labeled [97:56] retrain do it again right so it's a much [97:59] less well-written data engine and yeah [98:01] you can you can take these things really [98:03] far and it is [98:04] impressive engineering and if you truly [98:07] need supervised data for a problem [98:09] yeah things like data engine are the [98:11] high end of the [98:14] what is attention is a human paying [98:15] attention i mean we're going to probably [98:17] build something that looks like data [98:18] engine to push our driver monitoring [98:19] further [98:20] um but for driving itself you have it [98:23] all annotated beautifully [98:24] by what the human does so yeah that's [98:26] interesting i mean that applies to [98:27] driver [98:28] attention as well do you want to detect [98:30] the eyes do you want to detect blinking [98:32] and pupil movement [98:33] do you want to detect all the like face [98:36] alignments the landmark detection and so [98:37] on [98:38] and then doing kind of reasoning based [98:41] on that or do you want to take the [98:42] entirety of the face over time and [98:44] do and i mean it's obvious that over [98:47] eventually you have to do end to end [98:48] with some calibration with some fixes [98:50] and so on [98:51] but it's uh like i don't know when [98:54] that's the right move [98:55] even if it's end to end there actually [98:58] is there is no [98:59] kind of um you have to supervise that [99:02] with humans [99:03] whether a human is paying attention or [99:05] not is a completely subjective judgment [99:07] um like you can try to like [99:10] automatically do it with some stuff but [99:11] you don't [99:12] have if i record a video of a human i [99:15] don't have [99:16] true annotations anywhere in that video [99:18] the only way to get them [99:19] is with you know other humans labeling [99:22] it really [99:22] well i don't know uh you you so [99:26] if you think deeply about it you could [99:29] you might be able to just depending on [99:30] the task maybe you'll discover [99:32] self-annotating things like [99:34] you know you can look at like steering [99:36] wheel reversal or something like that [99:37] you can discover little moments of lapse [99:39] of attention yeah [99:41] i mean that's that's where psychology [99:44] comes in [99:44] is there indicate because you have so so [99:46] much data to look at [99:47] so you might be able to find moments [99:50] when there's like just [99:52] inattention that even with smartphone if [99:54] you want to detect smartphone use [99:56] yeah you can start to zoom in i mean [99:58] that's the gold mine [99:59] sort of the comma ai i mean tesla's [100:02] doing this too right is there they're [100:03] doing [100:05] annotation based on it's like uh [100:08] self-supervised [100:09] learning too it's just a small part of [100:11] the entire picture [100:13] it's that's kind of the challenge [100:16] of solving a problem in machine learning [100:18] if you can discover [100:20] self-annotating parts of the problem [100:23] right our driver monitoring team is half [100:26] a person right now i have a problem [100:28] you know once we have skill to full once [100:30] we have two people once we have two [100:31] three people on that team [100:33] i definitely want to look at [100:34] self-annotating stuff or yeah for [100:35] attention [100:38] let's go back for a sec to uh to a comma [100:41] and how what you know for people who are [100:44] curious to try it out [100:46] how do you install a comma in say a 2020 [100:49] toyota corolla [100:50] or like what are the cars that are [100:52] supported what are the cars that you [100:54] recommend [100:55] and what does it take you have a few [100:58] videos out but maybe through words can [101:00] you explain [101:00] what's it take to actually install a [101:02] thing so we support uh i think it's 91 [101:04] cars [101:05] 91 makes models um we get to 100 this [101:08] year [101:10] nice the yeah the [101:13] 2020 corolla great choice [101:16] the um 2020 sonata uh it's using the [101:20] stock longitudinal it's using just our [101:22] lateral control [101:23] but it's a very refined car their [101:25] longitudinal control is not [101:26] bad at all um so yeah corolla [101:30] sonata or if you're willing to get your [101:33] hands a little dirty and look in the [101:34] right places on the internet the honda [101:36] civic is great [101:37] uh but you're going to have to install a [101:38] modified eps firmware [101:40] in order to get a little bit more torque [101:42] and i can't help you with that comma [101:43] does not officially endorse that [101:45] um but we have been doing it we didn't [101:47] ever release it [101:49] uh we waited for someone else to [101:50] discover it and then you know [101:52] and you have a discord server where [101:54] people there's a very active [101:57] developer community yeah as opposed to [102:00] so [102:01] depending on the level of [102:02] experimentation you're willing to [102:04] to do that's the community if you [102:08] if you just want to buy it and you have [102:09] a supported car yeah it's [102:12] 10 minutes to install there's youtube [102:14] videos it's [102:15] ikea furniture level if you can set up a [102:17] table from ikea [102:18] you can install a comma 2 in your [102:20] supported car and it will just work [102:22] now you're like oh but i want this [102:23] high-end feature or i want to fix this [102:25] bug [102:26] okay well welcome to the developer [102:27] community uh [102:29] so what if i wanted to this is something [102:31] i asked you offline [102:33] or like a few months ago if i wanted to [102:35] run my own code [102:38] to um so use comma as a [102:42] platform and try to run something like [102:44] open pilot [102:45] what does it take to do that um so [102:48] there's a toggle in the settings called [102:50] enable ssh [102:51] and if you toggle that you can ssh into [102:54] your device you can modify the code you [102:55] can upload whatever code you want to it [102:57] um there's a whole lot of people so [102:59] about 60 of people are running stock [103:02] comma [103:03] about 40 percent of people are running [103:04] forks and there's a community [103:06] of there's a bunch of people who [103:08] maintain these forks and these forks [103:10] support [103:11] different cars or they have you know uh [103:14] different toggles [103:15] we try to keep away from the toggles [103:17] that are like disabled driver monitoring [103:18] but [103:19] you know there's some people might want [103:20] that kind of thing and like you know [103:22] yeah you can it's your car it's your [103:24] i'm not here to tell you you know [103:27] uh we we have some you know we ban [103:30] if you're trying to subvert safety [103:32] features you're banned from our discord [103:33] i don't want anything to do with you [103:35] but there's some forks doing that you [103:37] got it [103:39] so you encourage responsible uh forking [103:42] yeah yeah some people you know yeah some [103:44] people uh like like there's forks that [103:47] will do [103:47] um some people just like having a lot of [103:49] uh readouts on the ui like a lot of like [103:52] flashing numbers so there's forks that [103:54] do that [103:54] uh some people don't like the fact that [103:56] it disengages when you press the gas [103:57] pedal there's forks that disable that [104:00] got it now the the stock experience is [104:03] is what [104:04] like so it does both lane keeping and [104:06] longitudinal control [104:08] all together so it's not separate like [104:09] it is an autopilot [104:11] no so okay um some cars we use the stock [104:14] longitudinal control we don't do the [104:15] longitudinal control on all the cars [104:17] uh some cars the acc's are pretty good [104:18] in the cars it's the lane keep that's [104:20] atrocious in anything except for [104:21] autopilot super cruise [104:23] but you know you just turn it on and it [104:27] it works what does this engagement look [104:28] like yeah so we have i mean i'm very [104:31] concerned about mode confusion [104:32] i've experienced it on supercruise and [104:36] and autopilot where [104:37] like autopilot like autopilot disengages [104:39] i don't realize that the acc is still on [104:42] the lead car moves slightly over and [104:44] then the tesla accelerates to like [104:46] whatever my set speed is super fast and [104:48] like [104:48] what's going on here um we have [104:51] engaged and disengaged and this is [104:54] similar to my understanding i'm not a [104:55] pilot but my understanding is either the [104:57] pilot [104:58] is in control or the co-pilot is in [105:01] control [105:02] and we have the same kind of transition [105:04] system either open pilot is engaged or [105:06] open pilot is disengaged [105:08] engage with cruise control disengage [105:10] with either gas break or cancel [105:13] let's talk about money what's uh the [105:15] business strategy for comma [105:17] profitable well it's you're good [105:21] so congratulations yeah uh what uh [105:24] so basically selling so we should say [105:26] combo cost uh [105:27] a thousand bucks comment 200 for the [105:30] interface to the car as well it's 1200 [105:32] all said done [105:34] nobody's usually up front like this yeah [105:36] you got it you got to have the tac [105:37] on right yeah i love it this side i'm [105:40] not going to lie to you [105:41] trust me it will add 1200 value to your [105:43] life yes [105:44] it's still super cheap 30 days no [105:46] questions asked money back guarantee and [105:48] prices are only going up you know if [105:50] there ever is future hardware [105:52] it could cost a lot more than twelve [105:53] hundred dollars so comma three is in the [105:54] works [105:56] so it could be all i will say is future [105:59] hardware is going to cost a lot more [106:00] than the current hardware [106:02] yeah like i mean the people that use uh [106:05] the people i've spoken with that use [106:06] comma [106:07] use open pilot they first of all they [106:10] use it a lot [106:12] so people that use it they they fall in [106:13] love with oh our retention rate is [106:15] insane [106:16] this is a good sign yeah it's a really [106:18] good sign um [106:19] 70 of comma 2 buyers are daily active [106:23] users [106:23] yeah it's amazing [106:26] um oh also we don't plan on stopping [106:29] selling the comma too [106:30] like like it's you know so whatever you [106:33] create that's beyond comma two [106:36] it would be uh it would be potentially a [106:39] phase shift [106:40] like it's you it's so much better that [106:42] like you could use comma two and you can [106:44] use [106:45] comma depends what you want it's three [106:46] point four one kind of 42 [106:48] yeah you know autopilot hardware one [106:50] versus hardware two yeah the comma two [106:52] is kind of like hardware one got it got [106:54] it [106:55] got it got it i think i heard you talk [106:57] about retention rate with the [106:58] vr headsets that the average is just [107:00] once yeah which is fast i mean it's such [107:02] a fascinating way to think about [107:04] technology [107:05] and this is a really really good sign [107:07] and the other thing that people say [107:08] about comm is like they can't believe [107:09] they're getting this for a thousand [107:11] bucks [107:11] right it's it seems it seems like a some [107:14] kind of steal [107:16] so but in terms of like long-term [107:19] business strategies it basically to [107:21] put so it's currently in like a thousand [107:24] plus cars [107:27] 1200 more uh [107:30] so yeah dailies is about uh [107:34] dailies is about 2 000 weekly is about 2 [107:36] 500. monthlies is over 3 000. [107:38] wow we've grown a lot since we've [107:41] stopped [107:42] is the goal like can we talk crazy for a [107:44] second i mean what's the [107:46] goal to overtake tesla let's talk [107:49] okay so i mean android did overtake ios [107:51] yeah that's exactly it right so [107:54] yeah they did it i actually don't know [107:55] the timeline of that one [107:57] they but let let's talk uh [108:00] because everything is in alpha now the [108:02] autopilot you could argue is in alpha in [108:04] terms of [108:04] towards the big mission of autonomous [108:06] driving right and so [108:09] what yes your goal to overtake into [108:11] millions of cars [108:13] essentially of course where would it [108:15] stop [108:16] like it's open source software it might [108:18] not be millions of cars with a piece of [108:19] comma hardware but yeah [108:21] i think open pilot at some point [108:24] will cross over autopilot in in users [108:26] just like android crossed over ios [108:29] how does google make money from android [108:32] uh it's it's complicated their own [108:35] devices make money [108:37] google google makes money by just kind [108:39] of having you on the internet [108:41] uh yes google search is built in gmail [108:44] is built in [108:45] android is just a shill for the rest of [108:46] google's ecosystem yeah but the problem [108:48] is [108:49] android is not is a brilliant thing i [108:52] mean [108:53] android arguably changed the world so [108:55] there you go that's [108:56] you can you can feel good ethically [109:00] speaking but as a business strategy it's [109:03] questionable i'll sell hardware so [109:05] hardware i mean it took google a long [109:07] time to come around to it but they are [109:08] now making money on the pixel [109:10] you're not about money you're more about [109:12] winning yeah of course [109:14] but if only if only 10 percent of open [109:16] pilot devices [109:18] come from comma ai we still make a lot [109:20] that is still yes that is a ton of money [109:22] for our company but can't somebody [109:23] create [109:24] a better comma using open pilot or [109:27] you're basically saying well i'll [109:28] compete well i'll compete [109:29] is can you create a better android phone [109:31] than the google pixel right [109:32] i mean you can but like i love that so [109:35] you're confident [109:36] like you know what the hell you're doing [109:38] yeah it's [109:39] it's uh confidence in merit i mean our [109:43] money yeah our money comes from we're a [109:44] consumer electronics company [109:46] yeah and put it this way so we sold we [109:48] sold like three thousand company's [109:50] um i'm 2500 right now uh [109:54] and like [109:58] okay we're probably going to sell 10 000 [110:00] units next year right 10 000 units [110:02] even just a thousand dollars a unit okay [110:04] we're 10 million in uh [110:06] in in in in revenue um get that up to a [110:09] hundred thousand [110:10] maybe double the price of the unit now [110:12] we're talking like 200 million in [110:13] revenue yeah [110:13] actually making money uh one of the rare [110:17] semi-autonomous or autonomous vehicle [110:18] companies that are actually making [110:20] money yeah yeah you know [110:23] if you have if you look at a model when [110:25] we were just talking about this [110:26] yesterday if you look at a model [110:27] and like you're testing like you're a b [110:28] testing your model and if you're you're [110:30] you're [110:31] one branch of the a b test the losses go [110:32] down very fast in the first five epochs [110:34] yeah [110:35] that model is probably going to converge [110:37] to something considerably better than [110:38] the one where the losses are going down [110:40] slower [110:41] why do people think this is going to [110:42] stop why do people think one day there's [110:44] going to be a great like [110:45] well waymo's eventually going to surpass [110:47] you guys [110:49] no they're not you see like a world [110:52] where like [110:53] a tesla or a car like a tesla would be [110:56] able to [110:57] basically press a button and you like [110:59] switch to open pilot [111:01] you know you know they load in i don't [111:04] know so [111:05] i think so first off i think that we may [111:07] surpass tesla [111:08] in terms of users uh i do not think [111:11] we're gonna surpass tesla ever in terms [111:12] of revenue [111:13] i think tesla can capture a lot more [111:14] revenue per user than we can [111:16] um but this mimics the android ios model [111:19] exactly [111:20] there may be more android devices but [111:22] you know there's a lot more iphones than [111:23] google pixels [111:24] so i think there'll be a lot more tesla [111:25] cars sold than pieces of comma hardware [111:28] um and then as far as a [111:32] tesla owner being able to switch to open [111:35] pilot [111:36] uh does ios does iphones run android [111:40] no but you can if you really want to do [111:43] it but it doesn't really make sense like [111:44] it's not [111:45] it doesn't make sense who cares what [111:46] about if uh a large company like [111:50] automakers 4g m toyota came to george [111:53] hotz or [111:54] on the tech space amazon facebook google [111:57] came with a large pile of cash [112:01] would would you consider being [112:04] purchased [112:07] do you see that as a one possible not [112:10] seriously now [112:11] um i would probably uh [112:15] see how much uh they'll entertain [112:18] for me [112:19] um and if they're willing to like jump [112:20] through a bunch of my hoops then maybe [112:22] but like no not the way that m a works [112:24] today i mean we've been approached [112:26] and i laugh in these people's faces i'm [112:28] like are you kidding me [112:30] yeah you know because you're so it's so [112:32] it's so demeaning the m a people are [112:34] so demeaning to companies they treat the [112:37] startup world [112:38] as their innovation ecosystem and they [112:41] think that i'm cool with going along [112:42] with that [112:43] so i can have some of their scam fake [112:44] fed dollars you know fedcoin i don't [112:47] what am i gonna do with more fedcoin you [112:49] know i had coin fat coin man [112:51] i love that so that's the cool thing [112:53] about podcasting actually is uh [112:55] people criticize i don't know if you're [112:56] familiar with the spotify [112:58] uh giving joe rogan a hundred million [113:01] and something about that [113:03] and you know they respect [113:07] despite all the that people are [113:08] talking about spotify [113:11] people understand that podcasters like [113:14] joe rogan [113:15] know what the hell they're doing yeah so [113:17] they give them money [113:18] and say just do what you do and [113:21] like the equivalent for you would be [113:24] like [113:25] george do what the hell you do because [113:27] you're good at it try not to murder too [113:29] many people [113:30] like try like there's some kind of [113:32] common sense things like [113:33] just don't go on a weird rampage of [113:37] yeah it comes down to what companies i [113:39] could respect [113:40] right um you know could i respect gm [113:44] never [113:45] um no i couldn't i mean could i respect [113:48] like [113:49] a hyundai more so right that's that's a [113:52] lot closer toyota [113:54] what's your nah nah it's korean is the [113:58] way [113:59] i think i think that you know the [114:00] japanese the germans the us they're all [114:02] too [114:02] they're all too you know they all think [114:04] they're too great what about the tech [114:06] companies [114:07] apple apple is of the tech companies [114:10] that i could respect apple's the closest [114:12] yeah i mean i could never subscribe it [114:14] would be ironic [114:15] oh if uh if common ai is acquired by [114:19] apple [114:19] i mean facebook look i quit facebook 10 [114:21] years ago because i didn't respect the [114:22] business model [114:23] um google has declined so fast in the [114:26] last five years [114:28] what are your thoughts about waymo as [114:31] present and future so let me let me see [114:34] let me start by saying something uh nice [114:37] which is uh [114:38] i've visited them a few times and i've [114:41] [Music] [114:43] have ridden in their cars and [114:46] the engineering that they're doing [114:49] both the research and the actual [114:51] development and the engineering they're [114:52] doing [114:53] and the scale they're actually achieving [114:55] by doing it all themselves [114:56] is really impressive and the the balance [114:59] of safety and innovation [115:01] and like the cars work [115:04] really well for the routes they drive [115:07] like they drive [115:08] fast which was very surprising to me [115:10] like it drives like [115:12] the speed limit or faster the speed [115:14] limit it goes [115:16] and it works really damn well and the [115:17] interface is nice and chandler arizona [115:19] yeah yeah yeah in challengers in a very [115:21] specific environment so [115:22] it i you know it gives me enough [115:25] material [115:26] in my mind to push back against the [115:28] madman of the world like [115:30] george hotz to be like [115:34] like because you kind of imply there's [115:36] zero probability they're going to win [115:38] yeah and and after i've used [115:41] after i've written in it to me it's not [115:43] zero oh it's not for technology reasons [115:46] bureaucracy no it's worse than that it's [115:49] actually for product reasons i think [115:51] oh you think they're just not capable of [115:53] creating an amazing product uh [115:55] no i think that the product that they're [115:57] building doesn't make sense [116:00] um so a few things uh you say the weimos [116:04] are fast [116:05] um benchmark away mo against a competent [116:07] uber driver [116:09] right right the uber driver is faster [116:11] it's not even about speed it's the thing [116:12] you said [116:13] it's about the experience of being stuck [116:15] at a stop sign [116:16] because pedestrians are crossing [116:18] non-stop [116:20] i like when my uber driver doesn't come [116:21] to a full stop at the stop sign yeah [116:23] you know and so [116:27] let's say the waymo's are 20 slower than [116:30] than an uber [116:31] right um you can argue they're going to [116:33] be cheaper [116:34] and i argue that users already have the [116:36] choice to trade off money for speed [116:39] it's called uberpool um i think it's [116:42] like 15 [116:43] of rides or uber pools right users are [116:46] not willing to trade off [116:47] money for speed so the whole product [116:50] that they're building [116:52] is not going to be competitive with [116:55] traditional ride sharing networks [116:56] right um like and also [117:02] whether there's profit to be made [117:04] depends entirely on one company having a [117:06] monopoly [117:07] i think that the level for autonomous [117:10] ride sharing vehicles [117:12] market is going to look a lot like the [117:13] scooter market if even the technology [117:16] does come to exist which i question [117:18] who's doing well in that market yeah [117:20] it's a race to the bottom [117:21] you know well they could be it could be [117:23] closer like an uber and a lyft [117:25] where it's just a one or two players [117:27] well [117:28] the scooter people have given up trying [117:31] to market scooters [117:32] as a practical means of transportation [117:35] and they're just like [117:36] they're super fun to ride look at wheels [117:38] i love those things and they're great on [117:39] that front [117:40] yeah but from an actual transportation [117:43] product perspective [117:44] i do not think scooters are viable and i [117:46] do not think level 4 autonomous cars are [117:47] viable [117:49] if you uh let's play a fun experiment if [117:51] you [117:52] ran let's do a tesla and let's do waymo [117:56] if uh elon musk took a vacation for [118:00] a year he just said screw it i'm gonna [118:02] go live on an island [118:03] no electronics and the board decides [118:06] that we need to find somebody to run the [118:07] company [118:09] and they they decide that you should run [118:10] the company for a year [118:12] how do you run tesla differently i [118:14] wouldn't change much [118:16] do you think they're on the right track [118:17] i wouldn't change i mean i'd have some [118:20] minor changes but even even my debate [118:23] with tesla [118:24] about you know end to end versus segnets [118:27] like that's just software who cares [118:29] right [118:30] like it's not gonna it's not like you're [118:32] doing something terrible with segnats [118:34] you're probably building something [118:35] that's at least going to help you debug [118:36] the end-to-end system a lot [118:38] right it's very easy to transition from [118:41] what they have [118:42] to like an end-to-end kind of thing [118:45] and then i presume you would uh [118:49] in the model y or maybe in the model 3 [118:51] start adding driver sensing with [118:52] infrared yes i would add i would [118:54] i would add infrared camera infrared [118:57] lights [118:58] right away to those cars um [119:02] and start collecting that data and do [119:03] all that kind of stuff yeah [119:05] very much i think they're already kind [119:06] of doing it it's it's an incredibly [119:08] minor change [119:09] if i actually were ceo of tesla first [119:10] off i'd be horrified that i wouldn't be [119:12] able to do a good job as elon [119:14] and then i would try to you know [119:15] understand the way he's done things [119:17] before he would also have to take over [119:18] his twitter [119:20] so god i don't tweet yeah what's your [119:22] twitter situation why [119:23] why why are you so quiet on twitter [119:26] comma is [119:27] like what what's your social network [119:29] presence like because you you [119:31] on instagram you're you're you uh you do [119:33] live streams you're [119:35] you're you're um you understand the [119:37] music of the internet [119:39] but you don't always fully engage into [119:41] it you're part-time i used to have a [119:43] twitter [119:44] yeah i mean it's the pr instagram is a [119:46] pretty place [119:47] instagram is a beautiful place it [119:49] glorifies beauty i like i like [119:50] instagram's values as a network [119:53] um twitter glorifies conflict glorifies [119:55] you know [119:57] like like like like like you know just [119:59] shots [120:00] taking shots at people and it's like you [120:01] know well [120:03] you know twitter and donald trump are [120:05] perfectly [120:07] they're perfect for each other so [120:08] tesla's on uh tesla's on the right track [120:11] in your view yeah [120:12] okay so let's try let's like really try [120:15] this experiment if you ran [120:17] way more let's say they're i don't know [120:19] if you agree but [120:20] they seem to be at the head of the pack [120:22] of the kind of [120:24] uh what would you call that approach [120:26] like it's not necessarily lighter based [120:28] because it's not about lighter but [120:30] before robot taxi [120:31] level four robot taxi all in before any [120:34] before making your revenue [120:36] uh so they're probably at the head of [120:38] the pack if you were [120:41] said hey george can you please run this [120:43] company for a year how would you [120:45] change it uh i would go i would get [120:48] anthony levandowski out of jail [120:49] and i would put him in charge of the [120:52] company [120:55] um let's try to break that apart [120:58] why do you do you want to make do you [120:59] want to destroy the company by doing [121:01] that or [121:01] do you mean or do you mean uh you like [121:05] renegade style thinking that uh pushes [121:09] that that like throws away bureaucracy [121:11] and goes to first principle thinking [121:12] what [121:12] what do you mean by that um i think [121:14] anthony lewandowski is a genius [121:16] and i think he would come up with a much [121:18] better idea of what to do with waymo [121:20] than me [121:22] so you mean that unironically he is a [121:24] genius oh yes oh absolutely [121:26] without a doubt i mean i'm not saying [121:29] there's no shortcomings [121:30] but in the interactions i've had with [121:32] him yeah [121:34] what um he's also willing to take like [121:37] who knows what he would do with waymo [121:39] i mean he's also out there like far more [121:41] out there than i am yeah there's big [121:42] risks [121:43] yeah what do you make of him i was i was [121:45] going to talk to him in this podcast and [121:47] i was going back and forth [121:48] i'm i'm such a gullible naive human like [121:51] i see the best in people [121:53] and i slowly started to realize that [121:56] there might be some people out there [121:59] that like have multiple faces to the [122:03] world [122:05] they're like deceiving and dishonest i [122:08] still [122:09] refuse to like i i just i trust people [122:12] and i don't care if i get hurt by it but [122:14] like [122:15] you know sometimes you have to be a [122:17] little bit careful especially platform [122:18] wise and [122:19] podcast wise what do you what am i [122:21] supposed to think [122:23] so you think you think he's a good [122:24] person [122:26] oh i don't know i don't really make [122:28] moral judgments [122:29] and it's difficult to oh oh i mean this [122:31] about the waymo actually i mean that [122:33] whole idea very non-ironically about [122:35] what i would do [122:36] the problem with putting me in charge of [122:37] waymo is waymo is already [122:39] 10 billion dollars in the hull right [122:41] whatever idea waymo does look [122:43] com is profitable comes raised 8.1 [122:45] million dollars that's small [122:47] you know that's small money like i can [122:48] build a reasonable consumer electronics [122:50] company [122:50] and succeed wildly at that and still [122:53] never be able to pay back weight most 10 [122:54] billion [122:55] so i i think the basic idea with women [122:58] well forget the 10 billion because they [122:59] have some backing but [123:01] your basic thing is like what can we do [123:03] to stop making some money [123:05] well no i mean my bigger idea is like [123:07] whatever the idea is that's going to [123:09] save waymo [123:10] i don't have it it's going to have to be [123:12] a big risk idea and i cannot think of a [123:14] better person than anthony lewandowski [123:16] to do it [123:17] so that is completely what i would do as [123:19] ceo of waymo i call myself a [123:21] transitionary ceo [123:22] do everything i can to fix that [123:24] situation yeah [123:26] uh yeah cause i can't i can't do it [123:29] right like i can't i can't [123:30] oh i mean i can talk about how what i [123:33] really want to do is just apologize for [123:35] all those corny [123:36] uh you know ad campaigns and be like [123:38] here's the real state of the technology [123:40] yeah um like i have several criticism [123:42] i'm a little bit more bullish on [123:43] waymo than than you seem to be but [123:46] one criticism i have is it went into [123:49] corny mode too early [123:50] like it's still a startup it hasn't [123:52] delivered on anything so [123:53] it should be like more renegade and show [123:56] off the engineering that they're doing [123:59] which just can be impressive as opposed [124:00] to doing these weird commercials of like [124:03] your friendly yeah your friendly car [124:06] company [124:06] i mean that's my biggest my biggest [124:08] snipe at waymo was always that guy's a [124:10] paid actor [124:11] that guy's not a waymo user he's a paid [124:13] actor look here i found his call sheet [124:15] do kind of like what spacex is doing [124:17] with uh the rocket launchers just [124:19] get put the nerds up front put the [124:21] engineers up front and just like [124:23] show failures too just i love i love [124:26] spacex's [124:27] yeah yeah the thing they're doing is [124:29] right and it just [124:30] feels like the right but we're all so [124:32] excited to see them succeed [124:34] yeah i can't wait to see when it won't [124:36] fail you know like you lie to me i want [124:38] you to fail [124:39] you tell me the truth you'll be honest [124:40] with me i want you to succeed yeah [124:44] uh yeah and that requires the uh the [124:46] renegade ceo [124:48] right i'm with you i'm with you i still [124:51] have a little bit of faith in waymo to [124:53] for for the renegade ceo to step forward [124:56] but [124:57] it's not it's not john krafter yeah [125:01] it's uh you can't it's not chris [125:03] holmstone [125:04] and those people may be very good at [125:06] certain things yeah [125:08] but they're not renegades because these [125:10] companies are fundamentally even though [125:12] we're talking about billion dollars [125:14] all these crazy numbers they're still [125:16] like early stage startups [125:19] i mean i i just i if you are pre-revenue [125:21] and you've raised 10 billion dollars i [125:23] have no idea like [125:24] like this just doesn't work no it's [125:26] against everything silicon valley [125:27] where's your minimum viable product [125:29] you know where's your users what's your [125:31] growth numbers [125:33] this is traditional silicon valley why [125:36] do you not apply it to what you think [125:37] you're too big to fail already [125:39] like how do you think autonomous [125:43] driving will change society so the [125:46] mission is [125:46] for comma to uh solve [125:50] self-driving do you have like a vision [125:53] of the world of how [125:54] it'll be different [125:57] is it as simple as a to b transportation [125:59] or is there like [126:01] because these are robots it's not about [126:04] autonomous driving in and of itself it's [126:05] what the technology enables [126:09] it's i think it's the coolest applied ai [126:11] problem i like it because it has a clear [126:14] path to monetary value um [126:17] but as far as that being the thing that [126:19] changes the world [126:21] i mean no like like there's cute things [126:24] we're doing in common like who thought [126:25] you could stick a [126:26] phone on the windshield middle drive um [126:28] but like really the product that you're [126:30] building is not [126:31] something that people were not capable [126:33] of imagining 50 years ago [126:35] so no it doesn't change the world in [126:36] that front could people imagine the [126:38] internet 50 years ago only true [126:40] junior genius visionaries yeah everyone [126:42] could have imagined autonomous cars 50 [126:44] years ago [126:44] it's like a core but i don't drive it [126:46] see i i have this sense [126:48] and i told you like i'm my long-term [126:51] dream [126:52] is robots with which you have deep [126:55] with whom you have deep connections [126:57] right [126:59] and there's different trajectories [127:01] towards that [127:03] and i've been thinking so i've been [127:05] thinking of launching a startup [127:08] i see autonomous vehicles as a potential [127:10] trajectory to that that [127:13] i'm that's not where the direction i [127:15] would like to go but [127:16] i also see tesla or even kamehameha like [127:19] pivoting into [127:21] into robotics broadly defined [127:24] that's at some stage in the way like [127:26] you're mentioning the internet didn't [127:27] expect [127:29] let's solve you know when i say a comma [127:31] about this we could talk about this but [127:33] let's solve self-driving cars first [127:35] got to stay focused on the mission don't [127:37] don't don't you're not too big to fail [127:39] for however much i think kama's winning [127:41] like no no [127:42] no you're winning when you solve level [127:44] five self-driving cars and until then [127:45] you haven't win and [127:46] one and you know again you want to be [127:48] arrogant in the face of other people [127:50] great you want to be arrogant in the [127:51] face of nature you're an idiot right [127:53] stay mission focused brilliantly put uh [127:56] like i mentioned [127:57] thinking of launching a startup i've [127:58] been considering actually before [128:00] cove i've been thinking of moving to san [128:02] francisco oh [128:03] oh i wouldn't go there so why is uh [128:07] okay well and now i'm thinking about [128:09] potentially austin [128:10] um and we're in san diego now san diego [128:14] come here [128:14] so why what um [128:18] i mean you're such an interesting human [128:20] you've launched so many successful [128:21] things [128:23] what uh why san diego what do you [128:26] recommend why not san francisco [128:29] have you thought so for in your case san [128:31] diego with qualcomm and snapdragon i [128:33] mean [128:34] that's an amazing combination but that [128:37] wasn't really why [128:38] that wasn't the why no i mean qualcomm [128:40] was an afterthought qualcomm was it was [128:42] a nice thing to think about it's like [128:42] you can have [128:43] a tech company here and a good one i [128:45] mean you know i like qualcomm but [128:48] no um what's the west san diego better [128:50] than stephanie why does san francisco [128:51] suck [128:51] well so okay so first off we all kind of [128:53] said like we want to stay in california [128:55] people like the ocean you know [128:57] california [128:58] for for its flaws it's like a lot of the [129:01] flaws of california are not necessarily [129:03] california as a whole and they're much [129:04] more san francisco specific [129:05] yeah um san francisco so i think first [129:08] year cities in general [129:09] have stopped wanting growth uh [129:13] well you have like in san francisco you [129:15] know the voting class always votes to [129:17] not build more houses because they own [129:19] all the houses and they're like well [129:20] you know once people have figured out [129:22] how to vote themselves more money [129:24] they're going to do it [129:25] it is so insanely corrupt um it is not [129:28] balanced at all [129:29] like political party-wise you know it's [129:32] it's a one-party city and [129:34] for all the discussion of diversity it's [129:38] has it's stops lacking real diversity of [129:41] thought of background of uh [129:43] approaches to strategies of yeah ideas [129:48] it's it's kind of a strange place that [129:51] it's the loudest [129:52] people about diversity and the biggest [129:54] lack of diversity [129:56] well i mean that's that's what they say [129:58] right it's the projection [130:00] projection yeah yeah it's interesting [130:02] and even people in silicon valley tell [130:04] me that's uh [130:06] like high up people but everybody is [130:08] like this is a terrible place it doesn't [130:10] make i mean and coronavirus is really [130:12] what killed it yeah [130:13] san francisco was the number one uh [130:16] exodus [130:17] during coronavirus we still think san [130:19] diego is a good place to be [130:21] yeah yeah i mean we'll see [130:24] we'll see what happens with california [130:28] a bit longer term yeah i like austrians [130:30] and austin's an interesting choice i [130:32] wouldn't i wouldn't i don't have really [130:33] anything bad to say about austin [130:35] either except for the extreme heat in [130:36] the summer um which you know but that's [130:38] like very on the surface right i think [130:40] as far as like an ecosystem [130:42] goes it's it's cool i personally love [130:44] colorado [130:45] colorado uh yeah i mean you have these [130:47] states that are [130:48] you know like just way better run [130:51] um california is you know it's [130:54] especially san francisco so it's high [130:55] horse and like [130:57] yeah can i ask you for advice [131:01] to me and to others about [131:04] what's it take to build a successful [131:06] startup oh i don't know i haven't done [131:08] that [131:09] talk to someone who did that well you [131:11] know [131:14] uh this is like another book of years [131:16] that i'll buy for 67 [131:18] i suppose uh so there's um [131:22] one of these days will sell out yeah [131:24] that's right jail breaks are going to be [131:25] a dollar and books are going to be 67. [131:27] how i uh how i joe broke the iphone by [131:31] george cotts that's right [131:32] how i jailbroke the iphone and you can [131:35] do it [131:38] that's right that's right oh god okay i [131:41] can't wait but [131:42] quite so you haven't introspected you [131:45] have built [131:46] a very unique company i mean [131:49] not not you but you and others [131:53] but i don't know um there's no [131:56] there's nothing you have an interest but [131:57] you haven't really sat down and thought [131:59] about like [132:01] well like if you and i we're having a [132:03] bunch of we're having some beers [132:06] and you're seeing that i'm depressed and [132:08] whatever i'm struggling [132:09] there's no advice you can give oh i mean [132:13] more beer more beer [132:18] yeah i think it's all very like [132:20] situation dependent [132:22] um here's okay if i can give a generic [132:25] piece of advice [132:25] it's the technology always wins the [132:28] better technology [132:29] always wins and lying always [132:32] loses build technology [132:36] and don't lie i'm with you i agree very [132:39] much [132:40] the long run long run sure it's the long [132:42] run you know what the market can remain [132:43] irrational longer than you can remain [132:45] solvent [132:46] true fact well this is this is an [132:48] interesting point because i [132:50] ethically and just as a human believe [132:52] that um [132:54] like sm like hype and smoke and mirrors [132:57] is not at any stage of the company is a [133:01] good strategy [133:02] i mean there's some like you know pr [133:05] magic kind of like [133:06] you know you want a new product yeah if [133:08] there's a call to action [133:09] if there's like a call to action like [133:11] buy my new gpu look at it it takes up [133:13] three slots and it's this big it's huge [133:15] buy my g for you yeah [133:16] that's great if you look at you know [133:18] especially in that in the ai space [133:20] broadly but autonomous vehicles [133:22] like you can raise a huge amount of [133:24] money on nothing [133:26] and the question to me is like i'm [133:28] against that i'll never be [133:30] part of that i don't think i hope not [133:34] willingly not but like is there [133:37] something to be said [133:40] to uh essentially lying to raise [133:44] money like fake it till you make it kind [133:46] of thing [133:47] i mean this is billy mcfarland the fire [133:49] festival like we all we all [133:51] experienced uh you know what happens [133:53] with that no [133:55] no don't fake it till you make it be [133:57] honest [133:58] and hope you make it the whole way the [134:00] technology wins [134:01] right the technology wins and like there [134:04] is i'm not [134:05] i use like the anti-hype you know that's [134:07] that's a slava kpss reference but [134:09] um hype isn't necessarily bad [134:13] i loved camping out for the iphones [134:16] um you know and as long as the hype is [134:19] backed by [134:20] like substance as long as it's backed by [134:21] something i can actually [134:23] buy and like it's real then hype is [134:26] great and it's a great [134:27] feeling it's when the hype is backed by [134:30] lies that it's a bad feeling [134:31] i mean a lot of people call elon musk a [134:33] fraud how could he be a fraud [134:35] i've noticed this this kind of [134:36] interesting effect which is [134:38] he does tend to over promise [134:42] and deliver what's what's the better way [134:44] to phrase it [134:45] promise a timeline that he doesn't [134:48] deliver on he delivers much [134:50] later on what do you think about that [134:52] because i do that i think that's a [134:54] programmer thing [134:55] yeah i do that as well you think that's [134:58] a really bad thing to do or [135:00] is that okay i think that's again as [135:02] long as like [135:03] you're working toward it and you're [135:05] gonna deliver on it [135:06] it's not too far off right [135:10] right like like you know the whole the [135:13] whole [135:13] autonomous vehicle thing it's like i [135:16] mean i still think [135:17] tesla's on track to beat us i still [135:19] think even with their even with their [135:21] missteps they [135:22] have advantages we don't have um [135:25] you know illness is better than me at [135:28] at like marshaling massive amounts of [135:31] resources [135:33] so you know i still think given the fact [135:36] they're maybe making some wrong [135:37] decisions [135:38] they'll end up winning and like [135:41] it's fine to hype it if you're actually [135:43] gonna win [135:44] right if elon says look we're gonna be [135:46] landing rockets back on earth in a year [135:48] and it takes four [135:49] like you know he landed a rocket back on [135:52] earth [135:53] and he was working toward it the whole [135:55] time i think there's some amount of like [135:56] i think when it becomes wrong is if you [135:58] know you're not gonna meet that deadline [136:00] if you're lying yeah that's brilliantly [136:02] put like [136:03] this is what people don't understand i [136:06] think [136:06] like elon believes everything he says he [136:09] does [136:10] as far as i can tell he does and i i [136:12] detected that [136:13] in myself too like if i it's only [136:16] if [136:18] you're like conscious of yourself lying [136:21] yeah i think so yeah you know you can't [136:23] take that to such an extreme [136:25] right like in a way i think maybe billy [136:26] mcfarland believed everything he said [136:28] too [136:29] right that's how you started cult and [136:31] everybody uh kills themselves [136:33] yeah yeah like it's you need you need if [136:36] there's like some [136:37] factor on it it's fine and you need some [136:39] people to like [136:41] you know keep you in check but like [136:44] if you deliver on most of the things you [136:46] say and just the timelines are off [136:48] yeah it does piss people off though i [136:50] wonder [136:51] but who cares in the long arc of history [136:54] the people [136:55] everybody gets pissed off at the people [136:57] who succeed which is [136:58] one of the things that frustrates me [137:00] about this world [137:01] is uh they don't celebrate [137:05] the success of others like [137:09] there's so many people that want elon to [137:11] fail [137:12] it's so fascinating to me like what [137:15] is wrong with you like so [137:19] elon musk talks about like people short [137:21] like they talk about financial [137:23] yeah but i think it's much bigger than [137:24] the financials i've seen like the human [137:26] factors community they [137:28] want they want other people to fail [137:31] why why why like even people the [137:34] harshest thing is like [137:36] you know even people that like seem to [137:38] really hate donald trump [137:39] they want him to fail yeah or like the [137:42] other president or [137:43] they want barack obama to fail it's like [137:47] we're almost involved it's weird but i [137:50] i want that i would love to inspire that [137:52] part of the world to change [137:54] because well damn it if the human [137:57] species is going to survive we should [137:58] celebrate success [138:00] like it seems like the efficient thing [138:02] to do in this objective function that [138:04] like we're all striving for [138:06] is to celebrate the ones that like [138:08] figure out how to like do better [138:10] at that objective function as opposed to [138:12] like dragging them down [138:13] back into them into the mud i think [138:17] there is [138:17] this is the speech i always give about [138:19] the commenters on hacker news [138:21] um so first off something to remember [138:23] about the internet in general [138:24] is commenters are not representative of [138:27] the population [138:28] yeah i don't comment on anything i don't [138:31] you know [138:32] commenters are are representative of a [138:34] certain sliver of the population [138:36] and on hacker news a common thing i'll [138:39] say is when you'll see something that's [138:40] like [138:42] you know promises to be wild out there [138:45] and [138:45] in innovative there is some amount of [138:48] you know checking them back to earth but [138:50] there's also some amount of [138:52] if this thing succeeds well [138:56] i'm 36 and i've worked at large tech [138:58] companies my whole life [139:02] they can't succeed because if they [139:04] succeed that would mean that i could [139:06] have done something different with my [139:07] life [139:07] but we know that i couldn't have we know [139:09] that i couldn't have and and that's why [139:10] they're going to fail [139:11] and they have to root for them to fail [139:13] to kind of maintain their world image [139:15] so tune it out and they comment well [139:19] it's hard i uh so one of the things one [139:22] of the things i'm considering [139:24] startup wise is to change that [139:27] because i think the i think it's also [139:30] a technology problem it's a platform [139:32] problem i agree it's like [139:34] because the thing you said most people [139:36] don't comment [139:39] i think most people want to comment [139:42] they just don't because it's all the [139:44] for commenting exactly i don't [139:46] want to be grouped in with that or not [139:47] you don't want to be in a at a party [139:49] where everyone is an [139:50] yeah so they but that's a platform [139:53] problem that's [139:54] i can't believe what reddit's become i [139:56] can't believe the group think in reddit [139:58] comments [140:00] there's a red is an interesting one [140:02] because they're subreddits [140:05] and so you can still see especially [140:07] small subreddits [140:09] that like that are little like havens of [140:12] like [140:12] joy and positivity and like deep [140:16] even disagreement but like nuanced [140:17] discussion but it's only like small [140:20] little pockets [140:21] but that's uh that's emergent the [140:23] platform's not helping that or [140:25] hurting that so i guess naturally [140:29] something about the internet uh if you [140:31] don't put in a lot of effort to [140:32] encourage [140:34] nuance and positive good vibes it's [140:37] naturally going to decline into chaos [140:41] i would love to see someone do this well [140:42] yeah um i think it's [140:44] yeah very doable this is uh i think [140:46] actually so [140:47] i i feel like twitter could be [140:50] overthrown joshua bach talked about how [140:53] like uh if you have like and retweet [140:57] like [140:58] that's only positive wiring right [141:02] the only way to do anything like [141:04] negative there [141:05] is um with a comment and [141:08] that's like that asymmetry is what gives [141:12] you know twitter its particular [141:14] toxicness [141:15] whereas i find youtube comments to be [141:16] much better because youtube comments [141:18] have a have a [141:19] of an up and a down and they don't show [141:22] the downloads [141:23] without getting into depth of this [141:25] particular discussion [141:26] the point is to explore possibilities [141:29] and get a lot of data on it because uh [141:32] i mean i could disagree with what you [141:33] just said it's it's uh the point is it's [141:35] unclear [141:35] it's a it hasn't been explored in a [141:37] really rich way [141:39] like the these questions of how to [141:42] create [141:44] platforms that encourage positivity yeah [141:47] i think it's a it's a technology problem [141:49] and i think we'll look back at twitter [141:51] as it is now maybe it'll happen within [141:52] twitter [141:53] but most likely somebody overthrows them [141:56] is [141:57] we'll look back at twitter and say we [142:00] can't believe we put up with this level [142:02] of toxicity you need a different [142:03] business model too [142:04] any any social network that [142:06] fundamentally has advertising as a [142:07] business model [142:08] this was in the social dilemma which i [142:10] didn't watch but i liked it it's like [142:11] you know there's always the you know [142:12] you're the product you're not the uh but [142:15] they had a nuanced take on it that i [142:17] really liked and it said [142:18] the product being sold is influence over [142:22] you [142:24] the product being sold is literally your [142:26] you know [142:27] influence on you like [142:30] that can't be if that's your idea okay [142:33] well [142:34] you know guess what it cannot be toxic [142:36] yeah [142:37] maybe there's ways to spin it like with [142:40] with uh giving [142:40] a lot more control to the user and [142:42] transparency to see what is happening to [142:44] them as opposed to [142:45] in the shadows as possible but that [142:47] can't be the primary source of [142:49] but the users aren't no one's going to [142:50] use that it depends [142:52] it depends it depends i think i think [142:54] that the you're [142:55] you're not going to you can't depend on [142:58] self-awareness of the users [143:00] it's a it's another it's a longer [143:01] discussion because uh [143:03] you can't depend on it but [143:06] you can reward self-awareness [143:09] like if for the ones who are willing to [143:11] put in the work of self-awareness [143:13] you can reward them and incentivize and [143:16] perhaps be pleasantly surprised how many [143:18] people [143:20] are are willing to be self-aware on the [143:22] internet [143:23] like we are in real life like i'm [143:24] putting a lot of effort with you right [143:26] now [143:26] being self-aware about if i say [143:28] something stupid or mean [143:29] sure i'll like look at your like body [143:31] language like i'm putting in that effort [143:33] it's costly for an introvert it's very [143:35] costly [143:36] but on the internet it [143:39] like most people are like i don't care [143:41] if if this hurts somebody i don't care [143:43] if this uh [143:45] is not interesting or if this is yeah [143:47] the mean or whatever [143:48] i think so much of the engagement today [143:50] on the internet is so disingenuous too [143:52] yeah you're not doing this out of a [143:54] genuine this is what you think you're [143:55] doing this just straight up to [143:56] manipulate others whether you're in you [143:58] just became an ad okay [144:00] okay let's talk about a fun topic which [144:02] is programming [144:04] here's another book idea for you let me [144:06] pitch uh what's your uh perfect [144:08] programming setup [144:09] so like this by george hotz so [144:12] uh like what listen you're [144:17] giving me give me a macbook air sitting [144:18] in a corner of a hotel room and you know [144:20] i'll still have so you really don't care [144:21] you don't [144:22] fetishize like multiple monitors [144:25] keyboard [144:26] uh those things are nice and i'm not [144:28] going to say no to them [144:30] but do they automatically unlock tons of [144:32] productivity no not at all i have [144:34] definitely been more productive on a [144:35] macbook air in a corner of a hotel room [144:38] what about ide [144:41] so uh which operating system [144:44] do you love what uh text editor do you [144:47] use [144:48] ide what um is there is there something [144:50] that [144:52] is like the perfect if you could just [144:53] say the perfect productivity [144:56] set up for george hawks doesn't matter [144:58] it doesn't doesn't matter [144:59] it really doesn't matter you know i [145:00] guess i code most of the time in vim [145:02] like literally i'm using an editor from [145:04] the 70s you know you didn't [145:06] make anything better okay vs code is [145:08] nice for reading code there's a few [145:09] things that are nice about [145:10] it uh i think that they're you can build [145:12] much better tools [145:13] how like ida's xrefs work way better [145:15] than vx vs codes why [145:18] yeah actually that's a good question [145:20] like why i i still use [145:21] sorry emacs eat for most uh [145:25] i've actually know i have to confess [145:26] something dark [145:28] cause i've never used bim yeah [145:31] it's i think maybe i'm just afraid [145:36] that my life has been a like a waste [145:39] [Laughter] [145:40] i'm so i'm not i'm not evangelical about [145:42] emacs i [145:43] i think this this is how i feel about [145:45] tender flow versus pie torch [145:47] yeah having just like we've switched [145:49] everything to pie torch now put months [145:50] into the switch [145:51] i have felt like i've wasted years on [145:53] tensorflow i can't believe it [145:56] i can't believe how much better pie [145:57] torch is yeah [145:59] i've used emacs and them doesn't matter [146:01] yeah still just my heart [146:02] somehow i fell in love with lisp i don't [146:04] know why you can't [146:06] the heart wants what the heart wants i [146:08] don't i don't understand it but it just [146:09] connected with me maybe it's the [146:10] functional language at first i connected [146:12] with [146:13] maybe it's because so many of the ai [146:14] courses before the deep learning [146:16] revolution were taught [146:17] with lisp in mind i don't know i don't [146:20] know what it is but i'm i'm [146:21] stuck with it but at the same time like [146:23] why am i not using a modern id for some [146:25] of these programming like i don't know [146:27] they're not that much better i've used [146:28] modernity to use them but at the same [146:30] time so [146:31] to just not to disagree with you but [146:33] like i like multiple monitors [146:35] like i've i have to do work on a laptop [146:38] and [146:39] it's a it's a pain in the ass and also [146:41] i'm addicted to the [146:43] kinesis weird keyboard that you could [146:45] you could see [146:47] uh yeah so you don't have any of that [146:50] you can just be in a macbook [146:51] i mean look at work i have three 24-inch [146:54] monitors i have a happy hacking keyboard [146:56] i have a [146:57] razer death header mouse like but it's [147:00] not essential for you no [147:01] let's go to a day in the life of george [147:04] hotz what is the [147:05] perfect day productivity-wise so we're [147:08] not talking about like [147:10] hunter s thompson uh drugs yeah and uh [147:13] let's let's look at productivity like [147:16] what [147:16] what's the day look like on like hour by [147:19] hour is there any [147:21] irregularities that create a magical [147:24] george hawks experience [147:25] i can remember three days in my life and [147:28] i remember these days vividly [147:30] when i've gone through kind of [147:34] radical transformations to the way i [147:37] think [147:37] and what i would give i would pay a [147:39] hundred thousand dollars if i could have [147:40] one of these days tomorrow [147:42] um the days have been so impactful and [147:44] one was first discovering eliezer [147:46] yukowski on the singularity [147:49] and reading that stuff and like you know [147:51] my mind was blown [147:53] um the next was discovering [147:56] uh the hutter price and then ai is just [147:58] compression [147:59] like finally understanding aix i and [148:02] what all that was [148:03] you know i like read about it when i was [148:04] 18 19 i didn't understand it and then [148:06] the fact that like lossless compression [148:08] implies intelligence [148:09] the day that i was shown that and then [148:12] the third one is controversial [148:14] the day i found a blog called [148:15] unqualified reservations [148:17] and uh read that and i was like [148:20] wait which one is that that's uh what's [148:22] the guy's name curtis garvin [148:24] yeah so many people tell me i'm supposed [148:26] to talk to him [148:27] yeah but he looks he sounds insane or [148:30] brilliant [148:31] but insane or both i don't know the day [148:33] i found that blog was another like this [148:35] was during like like [148:36] gamergate and kind of the run-up to the [148:37] 2016 election and i'm like [148:40] wow okay the world makes sense now this [148:42] this like i had a framework now to [148:44] interpret this just like i got the [148:46] framework for ai and a framework to [148:47] interpret technological progress like [148:49] those days when i discovered these new [148:51] frameworks were oh interesting it's just [148:53] not about [148:54] but what was special about those days [148:57] how did those days come to be [148:58] is it just you got lucky like sure i [149:01] like [149:02] well you just encounter hutter prize on [149:05] uh on hack news or something like that [149:07] um like what but you see i don't think [149:11] it's just [149:11] see i don't think it's just that like i [149:13] could have gotten lucky at any point i [149:14] think that in a way [149:16] you were ready at that moment yeah [149:18] exactly to receive the information [149:21] but is there some magic to the day today [149:24] of like [149:25] like eating breakfast and it's the [149:27] mundane things [149:28] nah nothing no i drift i drift through [149:31] life [149:32] without structure i drift through life [149:35] hoping and praying that i will get [149:36] another day like those days [149:38] and there's nothing in particular you do [149:40] to uh to be a receptacle [149:42] for another for day number four [149:46] no i didn't do anything to get the other [149:47] ones so i don't think i have to really [149:50] do anything now i took a month-long trip [149:52] to new york and [149:54] i mean the ethereum thing was the [149:55] highlight of it but the rest of it was [149:56] pretty terrible [149:57] i did a two-week road trip and i got i [150:00] had to turn around i had to turn around [150:02] i'm driving in uh [150:05] in gunnison colorado i passed through [150:06] gunnison and uh [150:09] the snow starts coming down this path up [150:11] there called monarch pass in order to [150:12] get through to denver you gotta get over [150:13] the rockies [150:14] and i had to turn my car around i [150:16] couldn't i watched [150:18] i watched a f-150 go off the road i'm [150:20] like i gotta go back and [150:23] like that day was meaningful because [150:24] like like it was real like i actually [150:26] had to turn my car around [150:28] um it's rare that anything even real [150:30] happens in my life even as [150:31] you know mundane is the fact that yeah [150:34] there was snow [150:35] i had to turn around stay in gunnison [150:36] and leave the next day something about [150:38] that moment for [150:38] real okay so actually it's interesting [150:42] to break apart the three moments you [150:43] mentioned [150:44] if it's okay so uh i always have trouble [150:47] pronouncing his name but [150:49] allows yakowski yeah [150:53] so what how did your [150:56] world view change in starting to [150:59] consider [151:00] the the exponential growth of ai and agi [151:03] that [151:03] he thinks about and the the the threats [151:06] of artificial intelligence and all that [151:07] kind of ideas like [151:09] can you is it j like can you maybe uh [151:11] break apart like what [151:13] exactly was so magical to use a [151:15] transformational experience [151:17] today everyone knows him for threats and [151:18] ai safety um this was pre [151:21] that stuff there was i don't think a [151:22] mention of ai safety on the page [151:25] um this is this is old yukowski stuff [151:27] he'd probably denounce it all now he'd [151:29] probably be like that's exactly what i [151:30] didn't want to happen [151:31] well sorry man [151:35] is there something specific you can take [151:36] from his work that you can remember [151:38] yeah uh it was this realization that [151:42] uh computers double in power every 18 [151:45] months [151:45] and humans do not and they haven't [151:49] crossed yet [151:50] but if you have one thing that's [151:51] doubling every 18 months and one thing [151:53] that's staying like this [151:54] you know here's your log graph here's [151:57] your line [151:58] you know you calculate that [152:01] and that did that open the door to the [152:03] exponential thinking [152:05] like thinking that like you know what [152:06] with technology we can actually [152:09] transformed the world [152:11] it opened the door to human obsolescence [152:13] it opened the door to realize that in my [152:15] lifetime [152:17] humans are going to be replaced [152:20] and then the matching idea to that of [152:22] artificial intelligence with the hutter [152:24] prize [152:26] you know i'm torn i go back and forth on [152:29] what i think about it [152:30] yeah but the the the basic thesis [152:33] is it's nice to com it's a nice [152:35] compelling notion that [152:36] we can reduce the task of creating an [152:38] intelligent system a general [152:40] intelligence system [152:41] into the task of compression so you can [152:44] think of all of intelligence in the [152:45] universe in fact [152:46] as a kind of compression [152:50] do you find that was that just at the [152:52] time you found that as a compelling idea [152:53] do you still find that [152:54] a compelling idea i still find that [152:57] compelling idea [152:58] um i think that it's not that useful day [153:01] to day [153:02] but actually um one of maybe my quests [153:05] before that was a search for the [153:07] definition of the word intelligence [153:09] and i never had one and i definitely [153:12] have a definition of the word [153:13] compression [153:14] it's a very uh simple uh straightforward [153:17] one [153:18] and uh you know what confession is you [153:19] know what lossless is lossless [153:20] compassion not lossy lossless [153:21] compression [153:22] and that that is equivalent to [153:24] intelligence which i believe [153:26] i'm not sure how useful that definition [153:27] is day to day but like i now have a [153:29] framework to understand what it is [153:32] and he just 10x 10xed the uh the prize [153:35] for that competition like recently a few [153:36] months ago [153:37] you ever thought of taking a crack at [153:39] that oh i did [153:40] oh i did i spent i spent the next after [153:43] i found the prize i spent the next [153:45] six months of my life trying it and uh [153:48] well that's when i started learning [153:49] everything about [153:50] ai and then i worked vicarious for a bit [153:53] and then i learned [153:54] read all the deep learning stuff and i'm [153:55] like okay now i like i'm called up to [153:57] modern ai [153:58] wow and i had i had a really good [153:59] framework to put it all in from the [154:01] compression stuff [154:04] right like some of the first uh some of [154:05] the first deep learning models i played [154:07] with were [154:07] uh like gpt basically but before [154:11] transformers before it was still [154:13] uh rnn's to to do uh [154:16] character prediction but by the way on [154:18] the compression side [154:19] i mean the especially neural networks [154:22] what do you make of the lossless [154:24] requirement with the hudder prize so [154:27] you know human intelligence and neural [154:30] networks can probably compress [154:32] stuff pretty well but it would be lossy [154:35] it's imperfect [154:36] uh you can turn a lossy compressor into [154:37] a lossless compressor pretty easily [154:39] using an arithmetic encoder right you [154:41] can take an arithmetic encoder [154:42] and you can just encode the noise with [154:44] maximum efficiency [154:45] right so even if you can't predict [154:48] exactly what the next character is [154:50] the better a probability distribution [154:52] you can put over the next character [154:54] you can then use an arithmetic encoder [154:56] to uh right you don't have to know [154:57] whether it's an e or an i [154:59] you just have to put good probabilities [155:00] on them and then you know code those [155:02] and if you have it's a bits of entropy [155:04] thing right [155:06] so let me on that topic could be [155:07] interesting as a little side [155:09] tour what are your thoughts in this year [155:11] about gpt3 [155:13] and these language models and these [155:14] transformers is there something [155:16] interesting [155:17] to you as an ai researcher or is there [155:21] something interesting to you [155:22] as an autonomous vehicle developer nah i [155:25] think uh [155:26] i think it's overhyped i mean it's not [155:27] like it's cool it's cool for what it is [155:30] but no [155:30] we're not just going to be able to scale [155:32] up to gpg 12 and get [155:33] general-purpose intelligence like your [155:36] loss function [155:37] is literally just you know you know [155:39] cross-entropy loss on the character [155:41] right like that's not the loss function [155:42] of general intelligence [155:44] is that obvious to you yes can you [155:47] imagine [155:49] that like to play devil's advocate on [155:52] yourself is it [155:53] possible that you can the gpt-12 will [155:57] achieve general intelligence [155:58] with something as dumb as this kind of [156:00] loss function [156:01] i guess it depends what you mean by [156:03] general intelligence [156:04] so there's another problem with the gpts [156:07] and that's that they don't have [156:09] a uh they don't have long-term memory [156:12] right right so like just [156:16] gpt 12 a scaled up version of gpt two or [156:20] three [156:22] i find it hard to believe [156:25] well you can scale it in it's yeah so [156:28] it's a hardcore [156:30] hard-coded length but you can make it [156:33] wider and wider and wider yeah [156:36] you're gonna get you're gonna get cool [156:38] things from those systems [156:40] but i i don't think you're ever gonna [156:44] get something that can like [156:45] you know build me a rocket ship what [156:47] about solve driving [156:49] so you know you can use transformer with [156:52] video for example [156:54] you think is there something in there no [156:57] because [156:58] hey look we use we as a group we use a [157:01] group we could change that group out to [157:02] a transformer [157:04] um i think driving is much more [157:07] markovian than language [157:09] so markov you mean like the memory which [157:11] which aspect of uh [157:13] i mean that like most of the information [157:16] in the state at t [157:17] minus one is also in the in is in state [157:19] t yeah right and it kind of like drops [157:21] off [157:21] nicely like this where sometime with [157:23] language you have to refer back to the [157:25] third paragraph on the second page [157:27] i feel like there's not many like like [157:29] you can say like speed limit signs but [157:30] there's really not many things in [157:32] autonomous driving that look like that [157:33] but if you look at uh to play devil's [157:36] advocate [157:37] is uh the risk estimation thing that [157:38] you've talked about it's kind of [157:40] interesting [157:41] is uh it feels like there might be some [157:43] longer term [157:45] uh aggregation of context necessary to [157:48] be able to figure out like [157:50] the context yeah i'm not even sure i'm [157:53] i'm believing my my own devil's we have [157:56] a nice we have a nice like vision model [157:58] which outputs like a a [157:59] one two four dimensional perception [158:01] space um [158:03] can i try transformers on it sure i [158:05] probably will [158:06] at some point we'll try transformers and [158:08] then we'll just see do they do better [158:09] sure [158:10] i'm well it might not be a game changer [158:12] no well i'm not like like [158:13] might transformers work better than [158:14] grooves for autonomous driving sure [158:16] might we switch sure is this some [158:18] radical change no [158:19] okay we use a slightly different you [158:21] know we switch from rnns to grooves [158:23] like okay maybe it's greased to [158:24] transformers but no it's not [158:26] yeah i well on the on the topic of [158:29] general intelligence i don't know [158:30] how much i've talked to you about it [158:32] like what um [158:34] do you think will actually build an agi [158:37] like if if you look at ray kurzweil with [158:39] a singularity do [158:41] you have like an intuition about you're [158:43] kind of saying driving is easy [158:45] yeah and i [158:48] i tend to personally believe that [158:50] solving driving [158:52] will have really deep important [158:55] impacts on our ability to solve general [158:58] intelligence [158:59] like i i think driving doesn't require [159:01] general intelligence [159:03] but i think they're going to be [159:04] neighbors in a way that it's like deeply [159:07] tied [159:08] because it's so like driving is so [159:10] deeply connected to the human experience [159:13] that i think solving one will help solve [159:16] the other [159:17] but but so i don't see i don't see [159:19] driving is like [159:20] easy and almost like separate than [159:22] general intelligence but like what's [159:24] your [159:24] vision of a future with a singular do [159:26] you see there'll be a single moment like [159:28] a singularity where it'll be a phase [159:29] shift [159:30] are we in the singularity now like what [159:32] do you have crazy ideas about the future [159:34] in terms of [159:35] agi we're definitely in the singularity [159:36] now um we are [159:38] coolers of course look at the bandwidth [159:40] between people the bandwidth between [159:42] people goes up [159:43] all right um the singularity is just you [159:45] know when the bandwidth but [159:47] what do you mean by the bandwidth of the [159:48] people communications tools the whole [159:50] world is networked [159:51] the whole world is networked and we [159:52] raise the speed of that network right [159:54] oh so you think the communication of [159:56] information in a [159:57] distributed way is a empowering thing [160:00] for collective intelligence [160:02] oh i didn't say it's necessarily a good [160:03] thing but i think that's like when i [160:04] think of the definition of the [160:05] singularity yeah it seems kind of right [160:08] i see like it's a change in the world [160:12] beyond which [160:13] like the world be transformed in ways [160:15] that we can't possibly imagine no i mean [160:16] i think we're in the singularity now in [160:18] the sense that there's like you know one [160:19] world and a monoculture and it's also [160:21] linked [160:22] yeah i mean i i kind of shared the [160:24] intuition that [160:25] the the singularity will originate from [160:27] the collective intelligence [160:30] of us ants versus the like some single [160:33] system agi [160:34] type thing oh i totally agree with that [160:36] yeah i don't i don't really believe in [160:38] like like a hard take off agi kind of [160:40] thing [160:43] um yeah i don't think i don't even think [160:46] ai is all that different [160:48] in kind from what we've already been [160:50] building [160:51] um with respect to driving i think [160:54] driving is a subset of general [160:55] intelligence and i think it's a pretty [160:56] complete subset [160:58] i think the tools we develop at comma [161:00] will also be extremely helpful [161:02] to solving general intelligence and [161:04] that's i think the real reason why i'm [161:05] doing it [161:06] i don't care about self-driving cars [161:08] it's a cool problem to beat people at [161:10] but yeah i mean yeah you're kind of [161:12] you're of two minds [161:14] so one you do have to have a mission and [161:16] you want to focus and make sure you [161:17] get you get there you can't forget that [161:20] but [161:21] at the same time there is a thread [161:23] that's much [161:25] bigger than uh the connects the entirety [161:27] of your effort that's much bigger than [161:29] just driving [161:31] with ai and with general intelligence it [161:33] is so easy to delude yourself [161:35] into thinking you've figured something [161:36] out when you haven't if we build a level [161:38] 5 self-driving car [161:39] we have indisputably built something [161:42] yeah [161:43] is it general intelligence i'm not going [161:45] to debate that i will say we've built [161:46] something [161:47] that provides huge financial value yeah [161:49] beautifully put that's the engineering [161:51] credo like just just build the thing [161:53] it's like that's why i'm with uh with [161:55] the with elon on uh [161:57] go to mars yeah that's a great one you [161:59] can argue like [162:00] who the hell cares about going to mars [162:03] but the reality is [162:05] set that as a mission get it done yeah [162:07] and then you're going to crack some pro [162:09] problem that you've never even expected [162:11] in the process of doing that yeah [162:13] yeah i mean no i think if i had a choice [162:16] between humanity going to mars and [162:17] solving self-driving cars i think going [162:18] to mars is uh [162:20] better but i don't know i'm more suited [162:22] for self-driving cars i'm an information [162:24] guy i'm not a modernist i'm a [162:25] postmodernist [162:26] post modernist all right beautifully put [162:29] let me let me drag you back to [162:30] programming for a sec what [162:32] three maybe three to five programming [162:34] languages should people learn [162:36] do you think like if you look at [162:37] yourself what did you get the most out [162:39] of [162:40] from learning uh well so everybody [162:44] should learn [162:44] c and assembly we'll start with those [162:46] two right assembly [162:48] yeah if you can't code in assembly you [162:50] don't know what the computer's doing [162:51] you don't understand like you don't have [162:53] to be great in assembly but you have to [162:55] code in it and then like you have to [162:57] appreciate assembly in order to [162:59] appreciate all the great things c [163:00] gets you and then you have to code and [163:03] see in order to appreciate all the great [163:04] things python gets you [163:06] so i'll just say assembly c and python [163:07] we'll start with those three [163:09] the memory allocation of of c and the [163:12] the the fact that so assemblies give you [163:15] a sense of just how many levels of [163:17] abstraction [163:18] you get to work on in modern day [163:20] programs yeah yeah graph coloring for [163:22] assignment register assignment and [163:23] compilers yeah like you know you got to [163:25] do you know the compiler your computer [163:26] only has a certain number of registers [163:28] you can have all the variables you want [163:29] a c [163:30] function you know so you get to start [163:31] your build intuition about [163:33] compilation like what a compiler gets [163:35] you [163:37] what else um well then there's then [163:39] there's kind of uh [163:41] so those are all very imperative [163:42] programming languages um [163:45] then there's two other paradigms for [163:47] programming that everybody should be [163:48] familiar with [163:49] i'm one of them is functional uh you [163:51] should learn haskell and take that all [163:53] the way through [163:54] learn a language with dependent types [163:55] like [163:57] learn that whole space like the very pl [164:00] theory [164:00] heavy languages and haskell is your [164:03] favorite functional [164:04] what is that the go-to you would say [164:06] yeah i'm not a great haskell programmer [164:08] i [164:08] wrote a compiler in haskell once there's [164:10] another paradigm and actually there's [164:11] one more paradigm that i'll even talk [164:12] about [164:13] after that that i never used to talk [164:14] about when i would think about this but [164:16] the next paradigm is learn verilog of [164:17] hdl [164:19] um understand this idea of all of the [164:21] instructions execute at once [164:25] if i have a block in verilog and i write [164:27] stuff in it [164:28] it's not sequential they all execute it [164:30] once [164:33] and then like think like that that's how [164:34] hardware works [164:36] to be so i guess assembly doesn't quite [164:39] get you that [164:39] assembly's more about compilation and [164:42] verilog is more about the hardware [164:44] like giving a sense of what actually is [164:46] the hardware is doing [164:48] assembly c python are straight like they [164:50] sit right on top of each other in fact c [164:53] is well let's see it's kind of coded in [164:55] c but you could imagine the first c was [164:56] coded in assembly and python is actually [164:58] coded in c [164:59] um so you know you can straight up go on [165:02] that [165:03] got it and then verilog gives you that's [165:05] brilliant [165:06] okay and then i think there's another [165:08] one now everyone should carpathi calls [165:10] it programming 2.0 [165:12] which is learn a i'm not even gonna [165:16] don't learn tensorflow learn pi torch so [165:18] machine learning [165:20] we've got to come up with a better term [165:21] than programming 2.0 or um [165:25] but yeah it's a programming language [165:29] i wonder if it could be formalized a [165:31] little bit better which we feels like [165:33] we're in the early days [165:34] of what that actually entails [165:36] data-driven programming [165:39] data-driven programming yeah but it's so [165:42] fundamentally different as a paradigm [165:43] than the others uh like it almost [165:46] requires a different skill set [165:50] but you think it's still yeah [165:53] and ply torch versus tensorflow pytorch [165:55] wins it's the fourth paradigm it's the [165:57] fourth paradigm that i've kind of seen [165:59] there's like this you know imperative [166:02] functional hardware i don't know a [166:05] better word for it [166:06] and then ml do you have advice [166:11] for people uh that want to you know get [166:14] into programming want to learn [166:15] programming you have a [166:17] a video uh what is programming new [166:21] blessings exclamation point [166:22] and i think the top comment is like [166:24] warning this is not for noobs [166:27] uh do you have a noob like uh [166:30] tldw for that video but also uh [166:34] a new but friendly advice on [166:38] how to get into programming you are [166:39] never going to learn programming by [166:41] watching a video [166:42] called learn programming the only way to [166:45] learn programming i think and the only [166:46] one is the only way [166:48] everyone i've ever met who can program [166:49] well learned it all in the same way they [166:52] had something they wanted to do [166:54] and then they tried to do it and then [166:56] they were like [166:58] oh well okay this is kind of you know be [167:01] nice if the computer could kind of do [167:02] this thing and then [167:03] you know that's how you learn you just [167:04] keep pushing on a project [167:08] um so the only advice i have for [167:10] learning programming is go program [167:12] somebody wrote to me a question like we [167:14] don't really [167:15] they're looking to learn about recurring [167:17] neural networks [167:18] he's saying like my company is thinking [167:20] of doing recruit using recurring neural [167:22] networks for time series data [167:24] but we don't really have an idea of [167:25] where to use it yet [167:27] we just want to like do you have any [167:28] advice on how to learn about [167:30] these are these kind of general machine [167:32] learning questions and i think [167:34] the answer is like actually [167:37] have a problem that you're trying to [167:38] solve and and just i see that stuff oh [167:41] my god when people talk like that [167:42] they're like [167:43] i heard machine learning's important [167:45] could you help us integrate machine [167:47] learning with macaroni and cheese [167:48] production [167:51] you just i don't even you can't help [167:53] these people like who lets you run [167:55] anything who lets that kind of person [167:57] run anything [167:58] i think we're we're all um we're all [168:01] beginners at some point [168:02] so it's not like they're a beginner it's [168:05] it's like [168:06] my problem is not that they don't know [168:07] about machine learning my problem is [168:09] that they think that [168:09] machine learning has something to say [168:11] about macaroni and cheese production [168:14] or like i heard about this new [168:16] technology how can i use it for why [168:19] like i don't know what it is but how can [168:21] i use it for why [168:23] that's true you have to build up an [168:24] intuition of how because you might be [168:26] able to figure out a way but [168:28] like the prerequisites you should have a [168:29] macaroni and cheese problem to solve [168:31] first [168:32] exactly and then two you should [168:35] have more traditional like in the [168:37] learning process should involve [168:39] more traditionally applicable problems [168:41] in the space of whatever that is of [168:43] machine learning [168:44] and then see if it could be applied to [168:46] background at least start with tell me [168:48] about a problem like if you have a [168:49] problem you're like [168:50] you know some of my boxes aren't getting [168:52] enough macaroni in them [168:53] um can we use machine learning to solve [168:55] this problem that's [168:57] much much better than how do i apply [168:59] machine learning to macaroni and cheese [169:01] one big thing maybe this is me uh [169:04] talking to the audience a little bit [169:06] because i get these days [169:08] so many messages a device on how to like [169:12] learn stuff okay my [169:15] this this this is not me being mean i [169:18] think this is quite a profound actually [169:20] is you should google it oh yeah [169:24] like one of the uh [169:27] like skills that you should really [169:28] acquire as an engineer [169:31] as a researcher as a thinker like one [169:34] there's two [169:34] two complementary skills like one is [169:37] with a blank sheet of paper with no [169:39] internet to think deeply [169:41] and then the other is to google the crap [169:44] out of the questions you have like [169:46] that's actually a skill i [169:47] don't people often talk about but like [169:49] doing research like pulling at the [169:51] thread like looking up different words [169:53] going into like [169:55] github repositories with two stars and [169:58] like looking how they did stuff like [169:59] looking at the code or going on twitter [170:03] seeing like there's little pockets of [170:04] brilliant people that are like having [170:06] discussions [170:07] like if you're a neuroscientist go into [170:10] signal processing community if you're an [170:12] ai person going into the [170:14] psychology community like like switch [170:17] communities that keep searching [170:18] searching searching [170:19] because it's so much better [170:22] to invest in like finding somebody else [170:25] who already solved your problem [170:28] than than this to try to solve the [170:29] problem and because they've [170:32] often invested years of their life like [170:34] entire communities [170:36] are probably already out there who have [170:37] tried to solve your problem [170:39] i think they're the same thing i think [170:41] you go try to solve the problem [170:44] and then in trying to solve the problem [170:46] if you're good at solving problems [170:47] you'll stumble upon the person who [170:48] solved it already yeah [170:50] but the stumbling is really important i [170:52] think that's a skill that people should [170:53] really approach especially in undergrad [170:56] like search if you ask me a question how [170:58] should i get started in deep learning [171:00] like especially [171:04] like that is just so google like [171:08] the whole point is you google that and [171:10] you get a million pages [171:11] and just start looking at them yeah [171:14] start pulling at the thread [171:15] start exploring start taking notes start [171:17] getting it [171:18] advice from a million people that [171:20] already like spent their [171:22] life answering that question actually oh [171:25] well yeah i mean that's definitely also [171:26] yeah when people like ask me things like [171:27] that i'm like trust me the top answer on [171:29] google is much much better than anything [171:30] i'm going to tell you [171:32] right yeah people ask [171:35] it's an interesting question let me know [171:38] if you have any recommendations what [171:40] three books [171:41] technical or fiction or philosophical [171:43] had an impact [171:44] on your life or you would recommend [171:46] perhaps uh [171:49] maybe we'll start with the least [171:50] controversial uh infinite jest [171:52] um infinite jest is a [171:57] david foster wallace yeah it's a book [171:59] about wireheading really [172:03] very enjoyable to read very uh [172:06] well-written [172:07] you know you will you will you will grow [172:08] as a person reading this book [172:10] uh it's effort um and i'll set that up [172:13] for the second book which is pornography [172:16] it's called atlas shrugged um [172:19] uh which um atlas drug is pornography [172:22] i mean it is i will not i will not [172:24] defend the i will not say atlas shrugged [172:26] is a well-written book [172:28] it is entertaining to read certainly [172:30] just like pornography [172:31] the production value isn't great um you [172:33] know there's a 60-page [172:35] monologue in there that ann rand's [172:36] editor really wanted to take out [172:38] and she uh paid she paid out of her [172:42] pocket to keep that 60 page monologue in [172:44] the book [172:45] um but it is a great book for a kind of [172:50] framework um of human relations [172:54] and i know a lot of people are like yeah [172:56] but it's a terrible framework [172:58] yeah but it's a framework just for [173:00] context [173:01] in a couple days i'm speaking with for [173:04] uh [173:04] probably four plus hours with euron [173:06] brook who's [173:08] the main living remaining objectivists [173:12] objectivist interesting uh so [173:15] i've always found this philosophy quite [173:18] interesting on many levels one of how [173:21] repulsive [173:23] some percent of large percent of the [173:24] population find it [173:26] which is always uh always funny to me [173:29] when people [173:29] are like unable to even read a [173:32] philosophy [173:32] because uh of some i think that says [173:36] more about their psychological [173:39] perspective on it [173:40] yeah but but there is something about [173:44] objectivism and iran's philosophy that's [173:47] very deeply connected to this idea of [173:49] capitalism [173:50] of uh the ethical life is the productive [173:54] life [173:55] um that was always um [173:59] compelling to me it didn't seem as like [174:01] i didn't seem to interpret it in the [174:03] negative sense that some people do [174:05] to be fair i read that book when i was [174:06] 19. so you had an impact at that point [174:09] yeah [174:09] yeah and the the bad guys in the book [174:12] have this slogan from each according to [174:14] their ability to each according to their [174:16] need [174:17] and i'm looking at this and i'm like [174:19] these are the most cards this is team [174:20] rocket level cartoonishness [174:22] right no bad guy and then when i [174:24] realized that was actually the slogan of [174:25] the communist party [174:27] i'm like wait a second wait no no no no [174:30] no just you're telling me this really [174:32] happened yeah it's interesting i mean [174:35] one of the criticisms of her work is she [174:37] has a cartoonish view of good and evil [174:39] like [174:40] that there's like the the reality isn't [174:43] jordan peterson says this [174:44] is that each of us have the capacity for [174:46] good and evil in us as opposed to like [174:48] there's some characters who are purely [174:50] evil and some characters are purely good [174:52] and that's in a way why it's [174:53] pornographic [174:55] the production value i love it well evil [174:57] is punished and there's very clearly [174:59] you know there's no there's no you know [175:02] uh just like porn doesn't have uh you [175:05] know like character growth well you know [175:07] neither does alex shrugged like [175:09] brilliant well put but as a 19 year old [175:12] george cotts it was [175:13] it was good enough yeah yeah what uh [175:15] what's the third [175:16] you have something um i i could give [175:19] these these two i'll just throw out uh [175:21] there's sci-fi uh permutation city [175:23] um great things to start thinking about [175:25] copies of yourself and then um [175:28] that is uh greg egan [175:31] uh he's uh that might not be his real [175:33] name some australian guy might not be [175:34] australian [175:35] i don't know um and then this one's [175:38] online it's called the metamorphosis of [175:39] prime intellect [175:42] um it's a story set in a [175:44] post-singularity world it's interesting [175:46] is there uh can you if either of the [175:48] worlds do you find something uh [175:50] philosophy interesting in them that you [175:51] can comment on [175:53] i mean it is clear to me that uh [175:57] metamorphosis prime intellect is like [175:59] written by [176:00] an engineer uh which is [176:05] it's very it's very almost a pragmatic [176:08] take on a utopia in a way [176:12] positive or negative well that's up to [176:15] you to decide reading the book [176:17] and the ending of it is very interesting [176:20] as well and i didn't realize what it was [176:23] i first read that when i was 15. i've [176:25] reread that book several times in my [176:26] life [176:27] and it's sure it's 50 pages everyone [176:29] should go read it [176:30] what's uh sorry this is a little tangent [176:33] i've been working through the foundation [176:34] i've been i've haven't read much sci-fi [176:36] my whole life and i'm trying to fix that [176:38] the last few months that's been a little [176:40] side project [176:42] what's uh to use the greatest sci-fi [176:44] novel [176:45] uh that uh people should read or is that [176:48] or [176:49] i mean i would yeah i would i would say [176:50] like yeah permutation city metamorphosis [176:52] environmental i got it [176:53] um i don't know i i didn't like [176:54] foundation uh i thought it was way too [176:57] modernist [176:58] i feel like dune i've never read dune [177:01] i've never read dune i have to read it [177:04] uh fire upon the deep is [177:06] interesting uh [177:08] okay i mean look everyone should read [177:10] everyone's reading romance everyone [177:11] should read snow crash [177:12] if you haven't read those like start [177:14] there um yeah i haven't read snow [177:16] questions [177:17] yeah no it means very entertaining go to [177:20] lecture bach and if you want the [177:21] controversial one [177:22] bronze age mindset [177:25] all right i'll look into that one those [177:27] aren't sci-fi [177:28] but just to round out books so [177:32] a bunch of people asked me on twitter [177:34] and read it and so on for advice [177:36] so what advice would you give a young [177:38] person today about life [177:41] another way [177:44] what uh yeah i mean looking back [177:47] especially when you're [177:48] young younger you did and you continued [177:51] it [177:51] you've accomplished a lot of interesting [177:54] things [177:54] is there some advice from those [177:59] i'm that life of yours that you can pass [178:01] on if college ever opens again i would [178:04] love to give a graduation speech [178:07] um at that point i will put a lot of [178:09] somewhat satirical effort into this [178:11] question [178:12] yeah at this you haven't written [178:14] anything at this point [178:15] oh you know what always wear sunscreen [178:18] this is water [178:19] like you're plagiarizing i mean you know [178:22] but that's the that's the like clean [178:25] your room [178:26] you know yeah you can play drugs from [178:27] from all this stuff and it's it's [178:33] there is no self-help books aren't [178:36] designed to help you they're designed to [178:38] make you feel good [178:39] like whatever advice i could give you [178:42] already know [178:44] everyone already knows sorry it doesn't [178:46] feel good [178:50] right like you know you know what what [178:52] if [178:53] if i tell you that you should you know [178:56] eat well and and and read more and [179:00] it's not gonna do anything i think the [179:02] whole like genre of those kind of [179:04] questions is [179:05] is is meaningless i don't know if [179:08] anything [179:08] it's don't worry so much about that [179:10] stuff don't be so caught up in your head [179:12] right i mean you're yeah in the sense [179:14] that your whole life [179:16] is your whole existence is like moving [179:18] version of that advice [179:20] i don't know yeah [179:23] there's there's something i mean there's [179:25] something in you that resists that kind [179:27] of thinking and that in itself [179:28] is it's just illustrative of uh [179:33] who you are and there's something to [179:35] learn from that [179:36] i think you're you're clearly not [179:38] overthinking stuff [179:41] yeah and you know it's a gut thing i [179:43] even when i talk about my advice i'm [179:44] like my advice is only relevant to me [179:47] it's not relevant to anybody else i'm [179:48] not saying you should go out if you're [179:50] the kind of person who overthinks things [179:51] to stop overthinking things it's not bad [179:53] it doesn't work for me maybe it works [179:55] for you i you know i don't know [179:57] let me ask you about love yeah [180:01] uh so i think last time we talked about [180:03] the meaning of life [180:05] and it was it was kind of about winning [180:08] of course uh i don't think i've talked [180:11] to you about [180:12] love much whether romantic or just love [180:14] for [180:15] the common humanity amongst us all what [180:18] role has love played in your life in [180:21] this [180:22] in this quest for winning where does [180:24] love fit in [180:26] well the word love i think means uh [180:28] several different things there's [180:30] uh love in the sense of maybe i could [180:32] just say there's like love in the sense [180:33] of opiates and love in the sense of [180:35] uh oxytocin and then love in the sense [180:39] of [180:42] maybe like a love for math i don't think [180:44] fits into either those first two [180:46] paradigms [180:47] uh so each of those have they uh [180:51] have they have they given something to [180:53] you [180:55] in your life i'm not that big of a fan [180:57] of the first two [180:59] um what [181:03] the same reason i'm not a fan of you [181:06] know the same reason i [181:07] don't do opiates and don't take ecstasy [181:10] right [181:11] and there were times look i've tried [181:12] both um [181:15] i like opiates way more than i liked [181:16] ecstasy uh but they're not [181:21] the ethical life is the productive life [181:24] so maybe that's [181:25] my problem with with those and then like [181:27] yeah a sense of [181:28] i don't know like abstract love for [181:30] humanity [181:32] i mean the abstract love for humanity [181:34] i'm like yeah i've always [181:35] felt that and i guess it's hard for me [181:39] to imagine not feeling it and maybe [181:40] there's people who don't and [181:42] i don't know but yeah that's just like a [181:44] background thing that's there [181:46] i mean since we brought up uh drugs let [181:48] me ask you [181:51] this is becoming more and more part of [181:53] my life because i'm talking a few [181:54] researchers that are working on [181:56] psychedelics [181:57] i've eaten shrooms a couple times and it [182:00] was [182:00] fascinating to me that like the mind can [182:03] go [182:04] like it's fascinating the mine can go to [182:07] places i didn't imagine it could go and [182:09] it was [182:09] very friendly and and positive and [182:12] exciting and [182:13] everything was kind of hilarious in the [182:15] in the place wherever my mind went [182:17] that's where i went [182:18] is uh what do you think about [182:19] psychedelics do you think they have [182:22] where do you think the mind goes have [182:24] you done psychedelics [182:26] where do you think the mind goes uh is [182:28] there something useful to learn about [182:30] the places it goes [182:32] once you come back you know i find it [182:34] interesting [182:35] that this idea that psychedelics have [182:39] something to teach [182:40] is almost unique to psychedelics right [182:43] people don't argue this about [182:44] amphetamines and [182:47] that's true and i'm not really sure why [182:49] yeah i think [182:51] all of the drugs have lessons to teach i [182:53] think there's things to learn from [182:54] opiates i think there's things to learn [182:55] from amphetamines i think there's things [182:57] to learn from psychedelics things to [182:58] learn from marijuana [183:00] um but also at the same time [183:04] recognize that i don't think you're [183:06] learning things about the world i think [183:07] you're learning things about yourself [183:09] yes um and you know what's the [183:12] even though it might have even been uh [183:15] might have been a timothy leary quote i [183:17] don't want to miss about him but the [183:18] idea is basically like [183:19] you know everybody should look behind [183:21] the door but then once you've seen [183:22] behind the door you don't need to keep [183:23] going back [183:24] um so i mean and that's my thoughts on [183:27] on all real drug use too [183:29] except maybe for caffeine it's a it's a [183:33] little experience [183:34] that uh it's good to have but oh yeah no [183:37] i mean [183:38] yeah i guess yeah psychedelics are [183:39] definitely um [183:41] so you're a fan of new experiences i [183:43] suppose yes because they all [183:44] contain a little especially the first [183:46] few times it contains some lessons that [183:47] could be picked up [183:49] yeah and i'll i'll revisit psychedelics [183:53] maybe once a year [183:55] um usually small smaller doses [183:58] maybe they turn up the learning rate of [183:59] your brain i've heard that [184:02] i like that yeah that's cool big [184:04] learning rates have pros and cons [184:07] last question this is a little weird one [184:09] but you've called yourself [184:10] crazy in the past uh [184:14] first of all on a scale of one to ten [184:16] how crazy would you say [184:17] are you oh i mean it depends how you you [184:19] know when you compare me to elon musk [184:20] and anthony lewandowski not so crazy [184:23] so like like a seven let's go with six [184:27] six yes six what uh [184:31] well like seven seven's a good number [184:33] seven sorry well yeah i'm sure [184:35] day by day changes right so but you're [184:37] in that [184:38] in that area what uh [184:42] in thinking about that what do you think [184:43] is the role of madness [184:45] is that a feature or a bug if you were [184:48] to uh dissect your [184:50] brain so okay from like a [184:53] like mental health lens on crazy [184:56] i'm not sure i really believe in that [184:58] i'm not sure i really believe in like [185:00] a lot of that stuff right this concept [185:03] of [185:04] okay you know when you get over to like [185:06] like like like [185:07] hardcore bipolar and schizophrenia these [185:09] things are clearly [185:11] real somewhat biological and then over [185:13] here on the spectrum you have like [185:15] a dd and oppositional defiance disorder [185:18] and [185:18] these things that are like wait this is [185:21] normal spectrum human behavior like this [185:23] isn't [185:25] you know where's the the line here and [185:28] why is this like a [185:30] problem so there's this whole this you [185:32] know the neurodiversity of humanity is [185:35] huge like people think i'm always on [185:37] drugs people are saying this to me on my [185:38] streams and like guys you know like i'm [185:40] real open with my drug use i'd tell you [185:41] if i was on drugs [185:43] yeah i had like a cup of coffee this [185:45] morning but other than that this is just [185:46] me [185:47] you're witnessing my brain and action [185:51] so so the word madness doesn't even uh [185:54] make sense and then you're in the rich [185:56] neurodiversity of [185:58] humans i think it makes sense but only [186:02] for [186:04] like some insane extremes like if you [186:07] are [186:07] actually like visibly hallucinating [186:12] um you know that's okay but there is the [186:15] kind of spectrum on which you stand out [186:17] like that that's uh like if i were to [186:20] look you know at decorations on a [186:23] christmas tree or something like that [186:24] like if you were a decoration out that [186:27] would catch my eye like [186:29] that thing is sparkly [186:32] whatever the hell that thing is uh [186:35] there's something to that [186:37] just like refusing to be um [186:41] boring or maybe boring is the wrong word [186:43] but to um [186:46] yeah i mean be willing to sparkle [186:51] you know it's it's like somewhat [186:53] constructed i mean [186:54] i am who i choose to be uh [186:58] i'm gonna say things as true as i can [187:00] see them i'm not gonna i'm not gonna [187:03] lie and but that's a really important [187:05] feature in itself so like whatever the [187:07] neurodiversity of your [187:08] whatever your brain is not putting um [187:12] constraints on it that force it to to [187:15] fit into the mold of what society is [187:17] like [187:18] defines what you're supposed to be so [187:20] you're one of the specimens that [187:24] that doesn't mind being yourself [187:27] being right is super important [187:31] except at the expense of being wrong [187:37] without breaking that apart i think it's [187:38] a beautiful way to end it and george [187:40] you're one of the most special humans i [187:42] know it's truly an honor to talk to you [187:44] thanks so much for doing it thank you [187:45] for having me [187:47] thanks for listening to this [187:48] conversation with george hotz and thank [187:50] you to our sponsors [187:52] for sigmatic which is the maker of [187:54] delicious [187:55] mushroom coffee decoding digital which [187:58] is a tech podcast that i listen to [188:00] and enjoy and expressvpn [188:04] which is the vpn i've used for many [188:06] years please check out these sponsors in [188:08] the description to get a discount [188:10] and to support this podcast if you enjoy [188:13] this thing [188:14] subscribe on youtube review it with five [188:16] stars in apple podcast [188:17] follow on spotify support on patreon or [188:20] connect with me on twitter [188:22] at lex friedman and now let me leave you [188:25] with some words [188:26] from the great and powerful linus [188:29] torvald [188:30] talk is cheap show me the code [188:34] thank you for listening and hope to see [188:35] you next time