Waymo Driver Reaches 500k Rides Weekly as Autonomous AI Scales Up

自动驾驶AI加速规模化，Waymo Driver每周行程达50万次 Waymo通过软硬件深度整合与监管合作而非软件授权拓展至15座城市，同时将严重伤害事故减少至原先的十七分之一。 在Tim O'Reilly的主持下，Waymo领袖Sebastian Thrun与Dmitri Dolgov回顾了公司的发展历程：从早期的Google计时赛与DARPA挑战赛渊源，到如今在全美15座城市每周提供超过50万次出行服务。Dolgov详细阐述了物理AI面临的巨大技术挑战，强调了安全阈值、延迟极限以及依托Waymo基础模型（Waymo Foundation Model）进行海量数据扩展的重要性。讨论内容涵盖Waymo的多支柱架构、业务扩张战略，以及Thrun对空中出租车和人形机器人的见解。圆桌讨论最后向物理AI的开发者发出强烈呼吁，并主张将监管机构视为合作伙伴而非绊脚石。 Waymo 自 2009 年 Project Chauffeur 项目起步，目前已在美国 15 座城市实现常态化运营，每周提供约 50 万次行程，累计完全无人驾驶里程超 2 亿英里，严重伤害事故率相比人类驾驶降低约 17 倍。 (3:22) 联合首席执行官 Dmitri Dolgov 总结了自动驾驶 AI 研发的三大核心启示：利用规模化数据的“苦涩的教训”持续有效、系统可靠性每增加一个“9”难度提升十倍，以及严格的评测（Eval）体系是避免盲点的关键。 (6:39) Waymo 打造了多模态可控生成式世界模型 Waymo Foundation Model，融合物理轨迹预测与通用语义理解，并通过车载 Driver、仿真器 Simulator 和验证评估体系 Critic 三大支柱保障技术落地。 (10:12) 创始人 Sebastian Thrun 表示物理 AI 正步入指数级加速发展阶段，呼吁技术开发者将监管机构视为追求安全的合作伙伴与“最好的朋友”，通过积极协作推动前沿技术在现实世界落地。 (16:31)

Waymo has transitioned autonomous driving from early research moonshots into scaled commercial deployment by pairing data-driven physical AI with rigorous safety validation and institutional collaboration. Drawing on two decades of development, Sebastian Thrun and Dmitri Dolgov argue that real-world physical automation requires tight hardware-software integration and aligned regulatory partnerships rather than ungrounded digital heuristics or adversarial posturing. From Research Milestones to Commercial Ride Scale (3:22) The foundation for autonomous vehicles emerged from the DARPA Grand Challenges in the early 2000s, where military interest in reducing logistics casualties prompted competitive desert trials. After a 2004 trial saw the top vehicle travel only seven miles before catching fire, Sebastian Thrun's Stanford team completed a 140-mile desert course in 2005, convincing technology leaders of commercial viability. This spurred Project Chauffeur in 2009, early executive track demonstrations, and challenging field tests like the Carmel-to-Cambria coastal route. Waymo marked its zero-to-one milestone in 2015 with the driverless, pedal-free 'Firefly' vehicle on public roads, followed by the formal launch of fully autonomous commercial operations in 2020. As of the session, Waymo operates across 15 US cities, serving roughly 500,000 trips across 5 million miles per week. Over its cumulative record of more than 200 million autonomous miles driven, Waymo reports an approximate 17-fold reduction in serious injuries compared to human drivers in its operational zones. Technical Realities and Core AI Lessons (6:39) Dmitri Dolgov outlines autonomous driving as one of artificial intelligence's most demanding physical challenges due to severe safety risks, unpredictable physical environments with dynamic edge cases, and strict latency bounds—noting that at freeway speeds, vehicles travel approximately 100 feet per second. To navigate these constraints, Waymo utilizes the Waymo Foundation Model, combining vision-language models, large language models, and generative world models to predict physical dynamics and interpret nuanced human social interactions. For practitioners in robotics and AI, Dolgov highlights three guiding principles: the Bitter Lesson holds, meaning scalable data-driven architectures consistently surpass hand-engineered heuristics; solving the 'nines' of reliability is nonlinear, where a working prototype is merely the initial 1% and each successive decimal point requires ten times more effort; and evaluation infrastructure ('Eval is king') remains the critical anchor to prevent hidden regressions as underlying models shift. Ecosystem Architecture Versus Simple Software Licensing (13:43) Rather than manufacturing custom vehicles in-house or relying on broad software licensing, Waymo positions itself as building the 'world's most trusted driver,' partnering with automotive OEMs for hardware integration and operational execution. Dolgov explains that simple open-sourcing or detached software licensing fails in safety-critical driving because safety demands a unified three-pillar structure: the Driver (onboard compute, sensors, and execution code), the Simulator (large-scale off-board scenario generation), and the Critic (rigorous safety evaluation and validation frameworks). This operating model is designed to support an addressable global transportation market projected at 20 trillion miles driven per decade. Waymo plans to deploy its core system across multiple transportation vectors beyond urban ride-hailing, including local delivery logistics, long-haul trucking, and personally owned consumer vehicles. Regulatory Alignment and Expanding Physical AI (16:31) Addressing industry tensions surrounding government oversight, Thrun urges builders to view regulators as partners rather than adversaries, noting that Waymo established early progress by aligning with agencies on shared safety benchmarks. Although building safety-critical ground systems required two decades, Thrun asserts that cross-disciplinary physical AI is now entering an exponential compounding curve, dramatically reducing the timeline needed for future breakthroughs. However, expanding physical AI into new form factors involves distinct hurdles. In aviation autonomy, such as electric vertical takeoff and landing (eVTOL) aircraft developed by ventures like Wisk, regulatory certification requirements are substantially stricter due to zero public tolerance for air accidents. Similarly, both speakers express caution regarding humanoid robotics; Thrun notes unresolved reliability issues on everyday obstacles like staircases, while Dolgov questions whether replicating human physical control schemes is an optimal path for practical physical automation. While Waymo has proven that autonomous fleets can surpass human safety benchmarks in commercial city transit, the broader frontier of physical AI remains gated by extreme safety thresholds, regulatory certification challenges in aviation, and the technical complexity of achieving high-order reliability beyond structured road environments.

Waymo从实验室研究演进为全美商业化自动驾驶车队的核心经验表明，物理人工智能的规模化落地依托于海量数据驱动、严苛的可靠性评测与三位一体的系统生态，并需将监管机构视为安全协同创新的伙伴。 规模化验证与安全落地里程碑 (3:22) 自动驾驶技术的商业化历经了二十年的工程沉淀。早期技术突破始于美国国防高级研究计划局（DARPA）大挑战赛，2004年顶尖赛车仅行驶约7英里便起火，而2005年Sebastian Thrun带领的斯坦福大学团队完成了沙漠中约140英里的全自动行驶，促使Larry Page等决策者确认其商业可行性，并在2009年启动Google Chauffeur项目。经过加利福尼亚海岸线等极端路线的无干预测试，Waymo于2015年借助定制的Firefly原型车完成了公开道路上首次无方向盘和踏板的全无人驾驶测试，并于2020年正式推向商业化运营。 截至讨论举行时，Waymo已在全美15个城市开通运营，每周完成约50万次行程，周行驶里程达500万英里。其累计全自动驾驶里程已超过2亿英里。根据运营区域的数据统计，该系统的严重人身伤害事故发生率相较于人类驾驶员降低了约17倍，为自动驾驶的大规模常态化运营提供了安全实证。 物理环境约束与AI研发核心法则 (6:39) Dmitri Dolgov指出，自动驾驶面临极其严苛的安全门槛、连续物理环境中的长尾分布外边缘案例，以及毫秒级的实时低延迟限制——车辆在高速公路行驶时每秒行进约100英尺。针对这些物理限制，Dolgov总结了AI开发者应当遵循的三项法则：首先是理查德·萨顿的“痛苦教训”（The Bitter Lesson）依然成立，依托大规模计算与海量数据训练的方法始终胜过人工编写的启发式规则；其次，原型车研发仅占攻坚任务的1%，工程难度主要集中在对可靠性“9的递增”（从99%提升至99.9%等），每提升一个“9”，难度呈十倍增长；最后，鉴于模型迭代迅速，稳固严谨的评估与评测体系（Eval）才是防止研发盲区的核心。 在环境理解与复杂交互方面，Waymo研发了多模态可控生成式世界模型“Waymo Foundation Model”，该模型融合了大语言模型与视觉语言模型的语义认知，结合物理规律预测，用以解析车辆与人类驾驶员及行人间复杂的社会互动。 系统三支柱与多元商业化路径 (13:43) 在商业定位上，Waymo将自身定位为打造“全球最受信任驾驶员”的技术供应商，其商业路径侧重于与整车代工制造商（OEM）合作开展车辆制造与硬件集成，而非独立承担车辆平台生产。面向未来预计每十年全球行驶里程达20万亿英里的庞大交通市场，Waymo规划将Waymo Driver系统拓展至无人出租车、本地配送、长途货运物流以及个人拥有的私家车等多重场景。 Dolgov同时阐明，由于自动驾驶对生命安全负有不可推卸的责任，简单的软件开源或软件授权模式并不适用。系统必须依靠三项紧密协同的支柱构成闭环生态：其一是部署于车端的传感器、车载计算芯片与实时执行软件“驾驶员”（The Driver）；其二是用于测试边缘复杂场景的云端离线“仿真器”（The Simulator）；其三是负责在实际路测前进行系统性校验与安全把关的“评审系统”（The Critic）。 物理AI前沿载具形态的辨析与审视 (11:04) 针对物理智能向其他硬件形态的延伸，Thrun回顾了他自2010年起参与探索的电动垂直起降飞行器（eVTOL，后发展为波音控股的Wisk）研发历程。Thrun指出，公众对航空事故的容忍度远低于地面交通事故，因此空中自主飞行面临更高的适航适度与监管审定壁垒。不过，他提及了中国亿航智能（EHang）的载人无人机商业化运行，并预测未来几十年内个人空中交通在拉斯维加斯等特定城市有望逐步普及。 对于当前受资本热捧的人形机器人赛道（Thrun提及约64%的人形机器人企业位于中国），Thrun保持了审慎态度，指出目前尚未见到人形机器人能够在楼梯等复杂场景中达到与儿童安全共处的高可靠性水准；Dolgov亦补充认为，人形外观虽有其独特性，但通过复刻人类身体控制方式来解决驾驶或一般实体自动化任务，在工程路径上并非最优选择。 监管协同范式与物理智能的指数跃迁 (16:31) 面对技术开发者常将监管视为创新阻力的心态，Thrun倡导将监管机构重塑为研发者的“盟友”。他回顾项目经验指出，技术团队与监管方的共同底线均是保障公共安全，通过主动透明的沟通，并协助监管部门建立可量化的安全评估标准，才是突破阻力并推动无人驾驶合法驶入公共道路的行之有效的方式。 针对AI快速扩展下为何自动驾驶耗时二十年才走入规模化部署的疑问，Thrun解释称，涉及生命安全的系统对缺陷的零容忍拉长了攻坚周期。然而，他预测物理AI目前正迈入多领域交叉协同的指数级加速拐点，纯数字AI在过去几年的飞跃正在向实体世界渗透，此前需要二十年探索的工程难关，未来有望以更短的时间周期被快速突破。 Waymo的二十年发展验证了依托规模化数据、严苛评测机制与监管主动协同来突破物理AI极限的路径。然而，从特定受控城市的车队运营，进一步向全天候工况、多元车型架构以及长途货运与航空等严监管跨领域场景的大规模落地拓展，如何在保持近乎零故障率的同时控制系统集成的单位经济成本，依然是全行业尚未彻底解决的核心挑战。

## Background 该视频由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台。根据页面简介，这是一场由奥莱利媒体创始人兼 CEO Tim O'Reilly 主持的深度对话，特邀 Waymo 联合 CEO Dmitri Dolgov 与自动驾驶先驱 Sebastian Thrun，共同探讨 Waymo 如何从 2009 年 Google 内部的先锋“登月”实验，演进为如今全球领先的全无人自动驾驶商业服务体系。 ## TL;DR 对话系统回顾了 Waymo 从早期受 DARPA 挑战赛启发、启动 Project Chauffeur，到推出无方向盘 Firefly 原型车并最终实现每周 50 万次纯无人商业出行的演进历程。针对公认难度极高的物理 AI（Physical AI）落地，Dmitri 总结了严苛安全门槛、长尾复杂环境与超低时延三大天堑，并提出了坚信算力规模扩展（The Bitter Lesson）、攻坚可靠性指标中的“9”以及“评测即王道（Eval is King）”的核心工程法则。此外，嘉宾们深入剖析了融合视觉语言模型与动力学规律的 Waymo 基础世界模型，审视了人形机器人与飞行汽车的技术边界，并强调将车载驱动、仿真与严苛裁判紧密结合，以及主动与监管机构构建互信伙伴关系，是推动物理 AI 迎来指数级爆发的关键支柱。 ## Atmosphere & style 全场对话务实沉稳且极具工程深度，洋溢着一种摒弃浮躁炒作、专注技术长跑的硬核极客气质。两位开创者在重温早期沿海公路现场改代码、赛道圈速对决人类高管等趣事时充满默契与情怀，而在探讨安全容错率与物理 AI 规律时则严谨敬畏，整场访谈兼具历史复盘的厚重感与启迪未来的前瞻洞察力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=4s) Waymo 是 AI 领域的典范成功案例，摒弃硅谷常见的炒作营销模式，历经十余年工程深耕，成功将实验室研究转化为数百万次规模的全无人商业出行服务。 - [02:09](https://www.youtube.com/watch?v=wBm_shEsADs&t=129s) 2009 年初创期团队保持了极高的迭代敏捷度，曾通过自动驾驶赛道圈速击败众多 Google 高管赢得内部资源，并在沿海公路测试遇挫后于车内现场重写代码通关。 - [04:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=244s) 从 2015 年无方向盘踏板的 Firefly 原型车到 2020 年商业化落地，Waymo 目前每周运送 50 万单，相比人类驾驶员将严重伤害事故率降低了约 17 倍。 - [05:08](https://www.youtube.com/watch?v=wBm_shEsADs&t=308s) DARPA 挑战赛是自动驾驶产业化的重要催化剂，2005 年斯坦福 Stanley 穿越荒漠的壮举直接坚定了 Larry Page 等科技领袖对该技术商业前景的信心。 - [06:39](https://www.youtube.com/watch?v=wBm_shEsADs&t=399s) 物理 AI 落地面临近乎零容错的安全门槛、混乱物理长尾场景与毫秒级低延迟要求，解决这些难题必须依赖可扩展的算力与数据（苦涩的教训）。 - [06:39](https://www.youtube.com/watch?v=wBm_shEsADs&t=399s) 做出演示原型仅解决了前 1% 的问题，可靠性每增加一个“9”（如 99% 到 99.9%）攻关难度陡增 10 倍，必须构建科学严密的评测体系（Eval is King）来指导迭代。 - [09:40](https://www.youtube.com/watch?v=wBm_shEsADs&t=580s) Waymo 研发了多模态生成式世界模型（Waymo Foundation Model），将前沿视觉语言模型（VLM）掌握的社会语义常识与底层物理动力学规律深度融合。 - [11:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=664s) 物理 AI 具有多样化形态，人形机器人并非解决所有场景的最优解，在特定场景下专用形态（如直接电子线控驱动的无人车、低空飞行器）更具工程效率与安全性。 - [13:44](https://www.youtube.com/watch?v=wBm_shEsADs&t=824s) Waymo 自身定位为提供驾驶大脑的科技公司而非整车制造厂，其全栈能力由车载驱动（Driver）、云端超大规模仿真（Simulator）与严苛裁判评估（Critic）三大支柱紧密咬合构成。 - [16:31](https://www.youtube.com/watch?v=wBm_shEsADs&t=991s) 自动驾驶因对错误极端零容忍而经历了长达 20 年的蓄力，但随着物理 AI 步入指数级增长轨道，过去数十年的技术演进在未来可能在数月内被压缩重演。 - [16:31](https://www.youtube.com/watch?v=wBm_shEsADs&t=991s) AI 创业者应当转变思维，将监管机构视作推进安全落地的合作伙伴与盟友，而非阻碍创新的对立面。 - [19:07](https://www.youtube.com/watch?v=wBm_shEsADs&t=1147s) 继前几年数字 AI 迎来爆发式突破后，融合多模态与实体系统的物理 AI（Physical AI）将成为下一个核心竞技场与重大历史机遇。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=4s) 拒绝炒作的 AI 典范与无人驾驶先驱登场 主持人 Tim O'Reilly（奥莱利媒体创始人兼 CEO）在开场时询问现场有多少人体验过 Waymo。看到举手的人寥寥无几，他直言“这还远远不够”，并称乘车体验具有真正的“变革性颠覆意义”。他透露 Waymo 预计将于次年全面落地拉斯维加斯。在 Tim 看来，Waymo 是整个人工智能领域最伟大的成功案例之一：20 年前它还只是一个看似不可思议的先锋研究项目，如今已承载了数百万次全无人出行服务。更令人钦佩的是，与硅谷惯常充斥的“炒作机器”不同，Waymo 几乎没有过度营销，而是十几年如一日默默执行，最终演变成一种现象级技术。 随后，Tim 隆重请出了两位开创该事业的核心元老：Google X 创始人、Google 自动驾驶项目早期奠基人、如今持续探索具身智能与物理 AI 的行业老将 Sebastian Thrun，以及从项目第一天起就参与其中、现任 Waymo 联合 CEO 的 Dmitri Dolgov。Dmitri 一直坚守在自动驾驶一线超过 15 年，主导研发了每一台 Waymo 车辆的核心大脑——Waymo Driver 全无人自动驾驶 AI 系统。 ### [02:09](https://www.youtube.com/watch?v=wBm_shEsADs&t=129s) 创业初期：从“大战 Google 高管”到沿海公路的代码逆袭 Sebastian Thrun 回顾了 2009 年项目起步（内部代号 Project Chauffeur）时的趣事。为了在 Google 内部赢得管理层的信任与资源支持，Dmitri 亲自打造了一辆具备赛车性能的自动驾驶汽车，并邀请了 Google 高管层中所有好胜心强的“霸气高管（Alpha Males）”前来赛道进行圈速对决。结果极具戏剧性：经过一整天的比拼，最终排行榜前两名全被自动驾驶机器人占据，随后的 42 个名次才是人类高管。 Dmitri 笑称那段日子安全始终是第一位的，回想起来宛如初创团队最纯粹的“蜜月期”，是其职业生涯中最快乐的时光。小团队高度聚焦，每天的迭代速度惊人。当时团队设立了一个极具挑战的目标：在旧金山湾区跑完 10 条各长达 100 英里的严苛路线，且全程零人为接管。首次挑战加州海岸线从卡梅尔（Carmel）到坎布里亚（Cambria）的路线时，系统表现欠佳，中途出现了数次人工干预脱离。团队没有气馁，Dmitri 坐在返程的车内膝上放着电脑现场排错、修改算法并重写代码。当他们掉头回到起点再次挑战时，车辆便一气呵成、完美通关。 ### [04:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=244s) 关键里程碑：从 Firefly 原型到千万里程的超级安全记录 Dmitri 梳理了 Waymo 演进史中的几个关键转折点。2015 年迎来了真正的“从 0 到 1”时刻：Waymo 首次在公开道路上运行了完全没有人类司机的自动驾驶车辆——专门定制研发的 Firefly（萤火虫）原型车，该车甚至直接取消了方向盘和刹车油门踏板。此后团队又攻坚了五年，于 2020 年正式开启全无人商业化运营。 如今，Waymo 已在美国 15 座城市部署了全自动驾驶运营网络，其商业化规模与安全表现已达到坚实的“超人级别（Superhuman）”水准： - **服务规模**：每周提供约 50 万次商业出行，每周行驶里程达 500 万英里。 - **运行总量**：累计已完成超过 2 亿英里的全无人纯商业自动驾驶里程。 - **安全表现**：在其实际运营区域内，相比人类驾驶员，导致严重人身伤害的事故率降低了约 17 倍（17x reduction）。 ```mermaid flowchart LR accTitle: Waymo 发展里程碑演进图 subgraph S1["实验室起步 (2009)"] A["Project Chauffeur<br/>启动研发与赛道测试"] end subgraph S2["从零到一 (2015)"] B["Firefly 原型车<br/>公开道路无方向盘试乘"] end subgraph S3["商业落地 (2020)"] C["全无人商业运营<br/>Phoenix 率先落地"] end subgraph S4["规模化扩展 (当前)"] D["覆盖美 15 城<br/>周运送 50 万单 / 严重事故降 17 倍"] end A --> B --> C --> D ``` ### [05:08](https://www.youtube.com/watch?v=wBm_shEsADs&t=308s) 历史追溯：DARPA 挑战赛与政府催化剂的角色 Tim O'Reilly 强调，在自动驾驶的成功史中，政府的前瞻引导往往被外界忽略。Sebastian Thrun 对此深表赞同，并回顾了 2000 年代初期的伊拉克和阿富汗战争背景。当时简易爆炸装置（IED）对美军士兵造成了巨大杀伤，美国国防高级研究计划局（DARPA）因此设立大奖赛，号召全球顶尖高校探索荒漠环境下的自主地面机动技术。 2004 年在加州巴斯托（Barstow）举行的首届比赛极其惨烈，表现最好的一辆车也仅仅开了约 7 英里便起火瘫痪。但仅过了一年，在 2005 年第二届大奖赛中，Sebastian 率领的斯坦福车队（Stanley）就成功在无任何人员干预的情况下穿行了约 140 英里的荒漠。Sebastian 指出，正是这个转折时刻让来到现场的 Larry Page 等具备远见与资本的科技领袖意识到，无人驾驶技术完全有望成为现实的商业技术。 ### [06:39](https://www.youtube.com/watch?v=wBm_shEsADs&t=399s) 实体 AI 的三重天堑与给 AI 开发者的核心法则 针对多数从事通用 AI 研发的从业者，Tim 提问：自动驾驶作为公认难度极高的 AI 实体落地项目，究竟能给当下 AI 浪潮带来哪些启示？Dmitri 指出，自动驾驶技术之所以极其艰难，除了法规和供应链等非技术因素外，主要在于三个核心技术瓶颈： 1. **极其苛刻的安全门槛**：容错率几乎为零。 2. **嘈杂混乱的物理世界**：连续变量极多，现实世界中的长尾场景极易导致分布漂移（Out of Distribution）并引发恶性反馈回路。 3. **极低的时延要求**：一切必须实时响应。在高速公路行驶时，车辆每秒飞驰约 100 英尺（约 30 米），几毫秒的运算延迟都会直接关乎生死。 结合十几年的工程实践，Dmitri 总结了三条核心经验法则： - **坚信“苦涩的教训”（The Bitter Lesson）**：在每一轮技术浪潮中，能够支持大规模扩展计算与海量数据的架构和方法，最终都会战胜依赖人工定制特征的方案。 - **关注指标中的“9”（Think About the Nines）**：做出演示原型或 Demo 仅解决了前 1% 的问题；之后在性能和可靠性上每增加一个“9”（如从 99% 到 99.9%），攻关难度就会陡增 10 倍。必须据此构建团队组织、流程与架构，做好打持久战的准备。 - **评测即王道（Eval is King）**：行业热衷于追逐最新的模型架构，但“架构会过时，数据虽重要，没有科学严密的评估体系（Eval）和度量指标（Metrics），开发就如同盲人摸象”。 ### [09:40](https://www.youtube.com/watch?v=wBm_shEsADs&t=580s) 长期智能体（Agent）与 Waymo 世界模型的融合 在探讨多智能体交互与行为预测时，Tim 提到自动驾驶车辆本质上是一个长期运行（Long-running）的实体智能体，不仅要应对乘客，还需预测路上所有行人和车辆的行为。 Dmitri 透露，当前最具前景的前沿方向正是**语言模型/视觉语言模型（VLM）与世界模型（World Models）的深度融合**。在物理驾驶中，系统既需要理解底层的物理与动力学规律，又必须掌握语义、社会规则及互动博弈逻辑（即“我采取动作后，周围环境与其他动作者会产生何种反应”）。Waymo 为此研发了 **Waymo 基础模型（Waymo Foundation Model）**——一个多模态、高度拟真且完全可控的生成式世界模型，它将前沿 VLM 获取的通用常识语义，与物理世界的运动预测深度统一在了一起。 ```mermaid flowchart TD accTitle: Waymo 基础模型与闭环生态 subgraph DataWorld["物理世界与常识融合"] A["物理动力学规律<br/>(摩擦力/时延/运动轨迹)"] B["高阶社会语义常识<br/>(行人意图/手势交互/交通社会博弈)"] end subgraph CoreModel["Waymo 基础模型 (Foundation Model)"] C["多模态生成式可控世界模型"] end subgraph ThreePillars["Waymo 三大核心支柱"] D["车载驱动 (Driver)"] E["仿真世界 (Simulator)"] F["严苛裁判评估 (Critic)"] end A --> C B --> C C --> D C --> E E --> F F -->|验证反馈回路| D ``` ### [11:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=664s) 物理 AI 的不同形态：飞行汽车与人形机器人的现实审视 Sebastian 随后分享了他在航空出行领域的探索。2010 年他与 Larry Page 创立了探索空中交通的公司（后大部分业务并入波音旗下合资公司 Wisk），其逻辑在于升至 500 英尺高空后便不再有地面障碍物阻隔，能实现更快捷高效的点对点直线飞行。尽管适航审定与航空级安全监管的标准比地面严苛数倍，但目前中美两国均展现出巨大动力（如中国亿航 EHang 的商业化运营）。他预测未来数十年内，到访拉斯维加斯的旅客可能直接乘坐个人飞行器飞抵酒店。 当被问及人形机器人与自动驾驶车辆这类特定形态机器人的优劣时，Sebastian 幽默地表示：“我们从未想过要在 Waymo 驾驶座上放一个人形机器人来开车，因为直接剪断动力转向与刹车线、用电子线控去驱动车辆显然合理得多。”他提到目前全球人形机器人企业有约 64% 在中国，其核心立论在于人类世界（从门把手到楼梯水龙头）皆是按人体工学打造的；但从实用角度看，目前尚无任何一款人形机器人能以绝对可靠的安全性上下楼梯并让人放心把孩子托付其身旁。Dmitri 亦认为物理 AI 会有多种共存形态，人形绝非解决所有场景的最优解。 ### [13:44](https://www.youtube.com/watch?v=wBm_shEsADs&t=824s) 商业模式：构建生态系统与“三大核心支柱” 在商业模式探讨中，Tim 对比了特斯拉依靠量产私家车铺路、Uber 依赖技术授权等模式，询问 Waymo 选择自营车队的考量。Dmitri 澄清 Waymo 并不想成为重资产的整车制造厂，其定位是一家打造“全球最值得信赖的驾驶大脑”的科技公司，通过开放合作构建 AV 生态： - **生态合作**：与主机厂伙伴合作造车，在制造、供应链和运营维护端深度协同。 - **应用延展**：现阶段聚焦于网约车出行（Ride-hailing），未来将平滑扩展至货运卡车（Trucking）、本地配送（Deliveries）及个人私家车（Personally Owned Vehicles）。面对全球每年 13 万亿英里（预计到本年代末增至 20 万亿英里）的人类驾驶大盘，各种细分模式都将共存。 针对能否将 Waymo Driver 像通用大模型 API 一样开放授权的问题，Dmitri 认为短期内并不现实。自动驾驶是一套高度集成的全栈系统，由紧密咬合的**三大核心支柱**构成： 1. **车载驱动系统（The Driver）**：包含传感器套件、车载计算芯片以及终端算法。 2. **仿真系统（The Simulator）**：在云端构建高度拟真的超大规模并行仿真环境。 3. **评估裁判（The Critic）**：对智能体的各项安全及驾驶行为做极高精度的评测与校验。 脱离了硬件、仿真与严苛评估的任何单一模型组件，都无法直接在物理世界中保障安全落地。 ### [16:31](https://www.youtube.com/watch?v=wBm_shEsADs&t=991s) 苦涩教训下的指数级跃迁与监管者协作之道 针对 Rich Sutton 在 2019 年提出的“苦涩的教训”（算力与数据规模最终压倒人工专家规则），Tim 抛出疑问：既然规模法则如此强大，为何自动驾驶经历了 20 年长跑才得以成型？ Sebastian 解释，这完全是由对错误的极端零容忍决定的——LLM 出现幻觉，人们笑笑就过去了；但如果自动驾驶把红灯幻觉成绿灯，就会危及生命。然而，过去 20 年的沉淀并不意味着未来依然缓慢，随着物理 AI 步入指数级增长轨道，过去需要 20 年的技术积累在未来可能在 2 至 3 年乃至数月内被压缩重演。 最后，Sebastian 给在场的 AI 创业者分享了一条至关重要的心智法则：**将监管机构视作同盟而非敌人**。他谈到，在很多 AI 团队抱怨政策与监管时，Waymo 自初期起就主动与监管部门建立建设性伙伴关系。“监管者要安全，我们也要安全；监管者希望促进行业创新，我们也渴望创新。”通过主动沟通展示技术演进与安全边界，监管者能够成为自动驾驶与前沿 AI 落地过程中最坚定的盟友。 ### [19:07](https://www.youtube.com/watch?v=wBm_shEsADs&t=1147s) 与监管机构合作及物理 AI 的未来前景 在讨论的尾声，发言者探讨了如何正确看待监管机构的作用。大家常常会觉得监管机构总是站在对立面，甚至阻碍创新，但发言者建议创业者和技术人员应当改变这种思维定式，将监管机构视为合作伙伴与朋友，因为在推进安全和合规落地方面，监管方的要求本质上是在为技术的健康普及护航。 谈及当前的时代背景，发言者情绪高涨地表示，当下正处于一个极其不可思议且独特的历史节点。数个领域的指数级技术进步正在相互叠加与放大，创造出前所未有的创新浪潮。虽然过去几年人们见证了数字 AI（Digital AI）领域的突破性爆发，但在接下来的几年里，以自动驾驶为代表的物理 AI（Physical AI）将成为下一个核心战场。他鼓励所有相关领域的从业者抓住这个时代机遇，去创造和构建出更多令人惊叹的产品。 - **核心观点**：监管机构并非创新的阻碍，而应被视为推动技术安全落地的合作伙伴。 - **时代趋势**：继过去两年数字 AI 迎来突破后，未来几年将是物理 AI（Physical AI）快速爆发与应用的关键时期。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=4s) 这还远远不够，乘车体验具有真正的变革性颠覆意义。 — Tim O'Reilly > [02:09](https://www.youtube.com/watch?v=wBm_shEsADs&t=129s) 经过一整天的比拼，最终排行榜前两名全被自动驾驶机器人占据，随后的 42 个名次才是人类高管。 — Sebastian Thrun > [02:09](https://www.youtube.com/watch?v=wBm_shEsADs&t=129s) 安全始终是第一位的，回想起来宛如初创团队最纯粹的‘蜜月期’，是我职业生涯中最快乐的时光。 — Dmitri Dolgov > [06:39](https://www.youtube.com/watch?v=wBm_shEsADs&t=399s) 在每一轮技术浪潮中，能够支持大规模扩展计算与海量数据的架构和方法，最终都会战胜依赖人工定制特征的方案。 — Dmitri Dolgov > [06:39](https://www.youtube.com/watch?v=wBm_shEsADs&t=399s) 做出演示原型或 Demo 仅解决了前 1% 的问题；之后在性能和可靠性上每增加一个‘9’，攻关难度就会陡增 10 倍。 — Dmitri Dolgov > [06:39](https://www.youtube.com/watch?v=wBm_shEsADs&t=399s) 架构会过时，数据虽重要，没有科学严密的评估体系和度量指标，开发就如同盲人摸象。 — Dmitri Dolgov > [11:04](https://www.youtube.com/watch?v=wBm_shEsADs&t=664s) 我们从未想过要在 Waymo 驾驶座上放一个人形机器人来开车，因为直接剪断动力转向与刹车线、用电子线控去驱动车辆显然合理得多。 — Sebastian Thrun > [13:44](https://www.youtube.com/watch?v=wBm_shEsADs&t=824s) Waymo 并不想成为重资产的整车制造厂，其定位是一家打造‘全球最值得信赖的驾驶大脑’的科技公司。 — Dmitri Dolgov > [16:31](https://www.youtube.com/watch?v=wBm_shEsADs&t=991s) LLM 出现幻觉，人们笑笑就过去了；但如果自动驾驶把红灯幻觉成绿灯，就会危及生命。 — Sebastian Thrun > [16:31](https://www.youtube.com/watch?v=wBm_shEsADs&t=991s) 监管者要安全，我们也要安全；监管者希望促进行业创新，我们也渴望创新。 — Sebastian Thrun > [19:07](https://www.youtube.com/watch?v=wBm_shEsADs&t=1147s) 继过去两年数字 AI 迎来突破后，未来几年将是物理 AI 快速爆发与应用的关键时期。 — Sebastian Thrun ## Mentioned Waymo, Dmitri Dolgov, Sebastian Thrun, Tim O'Reilly, O'Reilly Media, Google, Google X, Project Chauffeur, Firefly (萤火虫原型车), Waymo Driver, Waymo Foundation Model, DARPA, Stanley (斯坦福无人驾驶赛车), Larry Page, Rich Sutton, 亿航 (EHang), Wisk, 波音 (Boeing), 特斯拉 (Tesla), Uber, Udacity ## Video description Dmitri Dolgov - Co-CEO, Waymo Sebastian Thrun - Founder, Waymo, Udacity Tim O'Reilly - Founder and CEO, O'Reilly Media In 2009, a small team set out to prove that cars could drive themselves. Today, that "moonshot" is a commercial reality. Join Waymo co-CEO Dmitri Dolgov and the industry veteran Sebastian Thrun—two of the program’s original founders—for a rare conversation on the evolution of autonomous driving. This session, moderated by Tim O'Reilly, will explore the audacious vision that started it all and the strategic pivots required to turn a lab experiment into a global service. While Thrun has since moved on to new frontiers, Dolgov has remained at the helm for over 15 years, guiding Waymo’s current trajectory as the world leader in autonomous mobility. Gain a dual perspective on the breakthroughs of the past and Waymo’s current expansion. [Source](https://www.youtube.com/watch?v=wBm_shEsADs)

[0:04] Tim O'Reilly: 这还远远不够，乘车体验具有真正的变革性颠覆意义。 [2:09] Sebastian Thrun: 经过一整天的比拼，最终排行榜前两名全被自动驾驶机器人占据，随后的 42 个名次才是人类高管。 [2:09] Dmitri Dolgov: 安全始终是第一位的，回想起来宛如初创团队最纯粹的‘蜜月期’，是我职业生涯中最快乐的时光。 [6:39] Dmitri Dolgov: 在每一轮技术浪潮中，能够支持大规模扩展计算与海量数据的架构和方法，最终都会战胜依赖人工定制特征的方案。 [6:39] Dmitri Dolgov: 做出演示原型或 Demo 仅解决了前 1% 的问题；之后在性能和可靠性上每增加一个‘9’，攻关难度就会陡增 10 倍。 [6:39] Dmitri Dolgov: 架构会过时，数据虽重要，没有科学严密的评估体系和度量指标，开发就如同盲人摸象。 [11:04] Sebastian Thrun: 我们从未想过要在 Waymo 驾驶座上放一个人形机器人来开车，因为直接剪断动力转向与刹车线、用电子线控去驱动车辆显然合理得多。 [13:44] Dmitri Dolgov: Waymo 并不想成为重资产的整车制造厂，其定位是一家打造‘全球最值得信赖的驾驶大脑’的科技公司。 [16:31] Sebastian Thrun: LLM 出现幻觉，人们笑笑就过去了；但如果自动驾驶把红灯幻觉成绿灯，就会危及生命。 [16:31] Sebastian Thrun: 监管者要安全，我们也要安全；监管者希望促进行业创新，我们也渴望创新。 [19:07] Sebastian Thrun: 继过去两年数字 AI 迎来突破后，未来几年将是物理 AI 快速爆发与应用的关键时期。

[2:09] Dmitri Dolgov: totally safe [3:22] Dmitri Dolgov: 0 to 1 [6:39] Dmitri Dolgov: The Bitter Lesson Holds [6:39] Dmitri Dolgov: Focus on the "Nines" [6:39] Dmitri Dolgov: Evaluation (Eval) is King [13:43] Dmitri Dolgov: world's most trusted driver [16:31] Sebastian Thrun: Very early on in the project we recognized that the regulators are actually our best friends and we treated them as such... You want safety, we want safety... How can we innovate on that basis? [19:07] Sebastian Thrun: think about regulators as your friend because they actually are [19:07] Sebastian Thrun: absolutely incredible and unique [19:07] Sebastian Thrun: build some amazing stuff

[0:04] How many of you out here have ridden in a Whimo?  Not enough of you. It's a really transformative [0:12] experience. I believe they'll be fully  rolled out in Las Vegas next year. So, [0:17] if you're back here next year, be sure to try  it because Whimo represents really one of the [0:22] greatest success stories in artificial  intelligence. You know, 20 years ago, [0:27] it was an audacious research project, but it's  already now grown into millions of autonomous [0:32] rides. And one of the things that's most striking  to me as a longtime observer of technology is [0:39] just how much they've done it without the hype.  Silicon Valley has been an enormous hype machine. [0:47] uh today, you know, Whimo rolled out just quietly  executing on and on and on until today they're [0:56] just really a phenomenon. So, but today we have  the opportunity to hear that journey from two of [1:01] the people who helped make it possible. Sebastian  Throne is best known for founding Google X, [1:06] but he als he led the original Google self-driving  car project, which laid the foundation for what [1:11] would become Whimo. and he's uh since continued to  advance the frontiers of robotics and physical AI [1:18] which I think is a really important frontier as  an entrepreneur researcher and CEO. Now Dimmitri [1:25] Dolgov is the co-CEO of Whimo but he's been part  of the journey from the very beginning because [1:29] he was also one of the founders of the Google  self-driving car project and today he leads [1:34] the development of the Whimo driver the fully  autonomous AI system that's at the heart of [1:39] every Whimo vehicle. Please join me in welcoming  Sebastian Throne and Dmitri Dolov. Let's start [1:45] with a little bit of history. You know, let's  take us back to the early days. Uh, you know, [1:49] maybe we'll start with you, Sebastian. Demetri,  you can pile on. And then Demetri, I'd love you [1:54] to kind of walk through the journey since, you  know, what were the the breakthroughs? What were [1:59] the turning points? What were the hardest things?  And then after we go through some history, we'll [2:04] start talking about what's what we can learn,  what people here can learn from your journey. [2:09] So over to you Sebastian. So history 2009 Dimmitri  and I started Chaffur. Early on Dimmitri built [2:19] a race car version of a self-driving car.  Yeah. And to get credibility inside Google, [2:25] we invited all these alpha males in the sea suite  of Google to race this car, do a time trial. And [2:33] then we showed them how the car would race itself.  And when the day was over, the top two spots were [2:39] taken by robots and the next 42 spots were taken  by people. Yeah, that's that's that's a vivid [2:47] illustration of where we're we're at. It  brings back memories. Thanks, Sebastian. Yeah, [2:52] this was uh totally safe, by the way. Absolutely  safe. Yeah, safety first. Uh since day one, [2:58] uh yeah, those early days, we can write a book  about all the all the fun stuff and all the crazy [3:03] stuff. This was the first couple of years when we  started. It was like probably the most fun I had [3:10] in my professional life. It's like the honeymoon  period of a startup. Just make small team focused, [3:17] learning a ton, making incredible progress.  How much you can build in a day. It was just [3:22] mind-blowing. I remember one of the goals we had  in the early days uh that we set for the team was [3:28] to drive 10 routes. Each one was 100 miles long  all over the Bay Area. And we had to do each one [3:33] from beginning to end with no human intervention.  Uh so the very first one we attempted was I think [3:39] it was from Carmel to Cambria going down the coast  California and you know we tried it for the first [3:45] time. I was in the car. It didn't do a great job.  We had a few disengages. We drove all the way to [3:50] the end. We figured we try again. So we started  heading home and I was there in the car with a [3:55] computer. I changed a few things. I wrote some  code. We got back to the top. We tried it again. [3:59] We succeeded. Just mind-blowing progress. So  this was uh early days. Then we started building [4:04] towards a product. Uh we had our zero to1 moment.  I would say in maybe 2015. Uh this was our very [4:10] first fully autonomous ride where we put a car on  public roads with no human behind the wheel. In [4:17] fact, the car didn't even have a steering wheel  and pedals. It was this custom design vehicle [4:21] that we called the Firefly. Uh then it took us  another five years to get to the start of a fully [4:27] autonomous commercial operations. That was 2020.  And then today, Whimo uh has fully autonomous cars [4:33] in 15 cities across the US. We're driving about  half a million or serving about half a million [4:39] trips per week, driving about 5 million miles. And  the Whimo driver is doing that uh with a strongly [4:47] superhuman safety record. We've driven well over  200 million fully autonomous miles. and we're [4:52] seeing about a 17x reduction in serious injuries  compared to human drivers in the places where we [4:59] operate. That's an amazing testament to the power  of uh AI. Thank you. Yeah. Uh I want to just pick [5:08] up two elements on that history really quickly.  Um you know before that 2009 race car incident, [5:15] there was the DARPA Grand Challenge. I just would  like to highlight the role of government which so [5:20] often gets uh written out of the story when it's  actually plays a pretty enormous role. Do you have [5:26] any reflections on that? Yes. So this is now going  way back in the early 2000s. The US government [5:33] wanted to see autonomy in the battlefield. We were  heavily engaged in Iraq, Afghanistan and IEDs, [5:40] improvised explosive devices became big troop  killers of our people. uh and they came up with [5:47] this ingenious idea to to make a competition  and invited the best universities in the world [5:52] to showcase if they could drive through a desert  environment. Uh started in Basau, California. The [6:00] the very first race was in 2004. Tells you how  old I am. Um and the best functioning car drove [6:07] about 10 miles or seven miles before it went  up in flames. The second one was 2005 and we [6:13] managed to get roughly 140 miles. Um, completely  autonomous, no person inside through a desert. And [6:20] that was the moment I think that was kind of the  birth moment of the modern self-driving car when [6:24] people of means and vision like Larry Page who's  there started to recognize that this could one day [6:30] be a viable commercial technology and thanks to  Dimmitri it is. Yeah. So the people here may how [6:39] many of you here are involved in robotics? Any of  you? So no, not really. So here's a question then. [6:47] What can people here who are working in a you  know AI today learn from your journey? You know [6:56] it seems to me that you may have had the hardest  deployment in AI. Safety is absolutely critical. [7:03] Uh but this also just you know you've had to  deal with regulation you've had to deal with [7:09] uh you know uh supply chain you've had all of the  you you have any issue that any of these people [7:16] have you've had probably 10x or 100x so what can  they learn from your past so I think you're right [7:27] uh autonomous vehicles I tend to see as one  of the most difficult AI problems and you know [7:35] setting aside the non-technical aspects on the  technical side there's three main things that [7:42] make it so one you mentioned safety Tim uh the bar  on performance is just incredibly high secondly of [7:50] course you're operating in the physical world the  physical world is messy it's noisy there's many [7:56] continuous variables that kind of you pull you  out of distribution or create you know unpleasant [8:00] feedback cycles and the third thing is that  latency really matters. Everything has to happen [8:05] in real time and when you're driving at freeway  speeds we're covering about 100 ft in a second. [8:11] So milliseconds do matter. Uh lessons you know  one thing that we see over over again maybe I'll [8:19] give three. One is that the bitter lesson holds  continues to hold with every wave of technology. [8:24] uh the lesson of building on architectures and  methods that support scale and support data is [8:32] super important. Uh secondly, that's kind of maybe  more of a general lesson is you know whenever [8:39] you're building something hard uh you know then  think about the nines. Think about how many nines [8:45] you get to get to the real product. when you  do a prototype or do a demo that's you know [8:50] the first 1% of the problem and then every other  nine whether we're talking about performance or [8:55] we're talking about reliability uh gets 10 times  harder. So think about you know where you want to [9:01] end up what your product requires and build your  processes build your technology build your team [9:07] especially the leadership team so that you can go  the distance and then maybe the last thing I'll [9:11] mention briefly if you're building AI whether it's  physical AI other eval is king right architectures [9:18] you know everybody likes to talk about the latest  ideas in the field and uh you know architectural [9:23] breakthroughs architectures come and go data is  hugely important but without eval without metrics, [9:30] you're kind of flying blind. So, invest in eval in  invest in metrics and that's the foundation of the [9:35] whatever AI we're going to be building. Strikes  me that might be one or two other things that [9:40] might be relevant. One is a you actually have a  longunning agent in some sense. You know, the cars [9:48] are actually sitting around uh you know, they've  got a basic mission, but then they're also inter [9:54] waiting and interacting you with with with new  humans. But they've also then got to anticipate [10:00] what other humans are going to do. Do you see any  lessons there for the future of agentic software [10:07] agents uh where you've learned something about  interaction absolutely prediction? Yeah. And in [10:12] fact for the physical world and then there you can  expand on that. One of the most interesting things [10:17] that I'm seeing happening is the convergence of  the work on language models or visual language [10:24] models and world models. And you gave a really  good motivating case for that, right? Driving [10:30] you have to understand how the world works. That's  the physics and dynamics of it. But you also have [10:35] to understand this semantic, social interactive  aspects of it, right? You take an action in the [10:41] world and the world reacts. So this is what we've  been building at Whimo called the Whimo foundation [10:46] model and it's really you know a multimodal highly  realistic uh controllable generative world model [10:53] that pulls in the semantics from the general  world knowledge that all of the digital BLMs [10:59] have been acquiring over the years and have  been so successful at. Yeah. So, Sebastian, [11:04] uh, I wanted to ask you a little bit about your  journey since Whimo and since the well, or rather [11:10] since the Google self-driving car project. You've  worked on flying taxis, uh, you know, flying cars, [11:15] you worked on robotics in various ways. So, you're  deep into this physical AI world. Um, can you [11:21] tell us a little bit about what's happening at  the frontier there now? Yeah. So roughly 2010, [11:27] Larry Page and I started a alternative company  where the idea was to effectively build a flying [11:32] car. And it's it's not a flying car with wheels.  It's more like a drone, but the idea was if you go [11:39] 500 ft up near um then there's just no obstacles  anymore. So you can go in a straight line faster, [11:44] more economical. That we sold large parts to  Boeing called Whisk now. And there's in this [11:50] country about four or five very credible companies  in that space and they're all uh working on the [11:56] regulatory framework on safety on certification  which for aviation is a higher hurdle than on the [12:01] ground because general tolerance on the general  public for aviation accidents is lower than for [12:06] car accidents. There's also a lot of motion  in China now. There's a company called E-Hang [12:11] that's commercially operating already um in China.  I would not be surprised if some decades from now [12:18] when you come to Vegas, you would not be on the  ground anymore, but you take your personal drone [12:23] and it brings you all the way to a hotel. What  do you guys think about uh humanoid robots versus [12:30] the kind of robots that you guys are actually  actively building? Well, I I can tell you we never [12:36] tried to put a humanoid robot into a way more and  self-driving car. it felt a little bit wrong given [12:40] that you can just snip the wires to the power  steering and to the power brake and and actuate [12:46] the car that way. Um there's as everybody knows  a massive interest especially in China about 64% [12:52] of human companies are Chinese today in this new  field and they have a very good argument uh that [12:58] there's 100 years we instrumented the world for  people like all the way from the steering wheels [13:02] to faucets to door handles they're all dimensioned  for us so why not just utilize that um environment [13:10] infrastructure investments for robotics um it's  still a bit of a research topic I have not yet to [13:17] a humanoid that can reliably go up and down  staircases reliably. So that I would be happy [13:22] to have my children next to it. I haven't seen  that yet. I think this is going to be one of [13:26] the manifestations of AI, one of the form factors  of AI in the physical world. I don't believe it's [13:31] the right give a good example of that's not the  right way to solve the autonomous vehicle problem. [13:35] I don't believe it's going to be the right form  factor to tackle the mole. I think there's going [13:38] to be multiple ones, but that one is I think uh  has its place. Yeah. So let's talk about business [13:44] models a bit. Whimo, you own your own fleet. Uh,  you know, someone like Tesla is kind of betting [13:52] on the fact that they're putting it into the cars  that individuals drive. And then presumably, I'm [13:57] not quite sure what their plan would be if they  were going to roll out a, you know, a competitor, [14:02] but it could obviously be a version, you know,  Uber could be licensing. You know, there's all [14:07] these possible business models. How do you think  about why the one you've chosen is the best one? [14:14] Well, we think of Whimo as you know a technology  company and with a goal of building the world's [14:22] most trusted driver and how we go about deploying  that driver. We're we're being practical. So right [14:30] now we are doing you know we're very focused on  right hailing uh there the strategy is to build [14:37] an AV ecosystem right we're very focused and we  want to play to our strength which is you know [14:43] AI and the way driver uh and we partner we partner  with com car companies we don't want to build our [14:50] own car we partner with other companies on the  manufacturing and the supply chain as well as [14:54] the operational side of the house and then beyond  that we want to have the way driver deployed in [14:58] other applications which is not just right hailing  right there's 13 trillion miles driven worldwide [15:03] with a ter horable terrific safety record it's  increasing to 20 trillion by end of the decade [15:09] so we want to be right healing deliveries trucking  personally own vehicles and I think the business [15:16] model and the ecosystems and the partnerships that  we weave together in all of those will be a little [15:20] bit different so you could conceive of a future  in which you really in the same way that you know [15:26] LLMs today are uh made available to developers.  You could you could see the Whimo driver becoming [15:33] a platform for other kinds of businesses. I  think some parts of what we're building being [15:40] a platform, I think there's some future there. I  don't want to be too prescriptive. I think it's [15:45] too early. The analogy of you build a thing and  you license it or you open source that I don't [15:51] think quite works because it's the ent driver.  It's not just the models or the software you put [15:55] in the car. is the hardware, sensors, computers,  software that runs in the car, software that runs [16:02] in the cloud, the whole machinery for building the  driver, uh, evaluating it. So, we kind of often [16:09] talk about our AI ecosystem bu being formed around  three key pillars. The driver, the simulator, [16:16] and the critic. And all of that has to be, you  know, validated to an extremely high degree of [16:22] safety before you actually put that agent out in  the physical world. So I think it's a much more [16:26] complicated story because you can't just take one  component and without the rest of it. You know you [16:31] mentioned the bitter lesson earlier. Uh how many  people here are familiar with the concept of the [16:36] bitter lesson? Another may it's really important.  It's it's this fundamental idea from a paper in [16:43] 2019 that basically if you look at the progress in  AI raw scale empowered big data has outperformed [16:52] uh you know specialized human expertise and uh you  know I think a lot about this because my company [16:59] is a company that's devoted to capturing and  sharing human expertise and as it gets absorbed [17:07] into AI models Yet your story also shows that  despite the bitter lesson, you know, there's no [17:16] fast takeoff of Whimo, right? It's been a 20-year  journey. And so this notion that AI will some the [17:22] bitter lesson is just going to lead to some kind  of fast takeoff seems uh to be at odds with the [17:30] reality of the world. Do you have any reflections  on that either of you? It is a 20-year journey [17:35] because of the extreme low tolerance to errors.  Like if a large language model hallucinates, [17:41] we shrug and move on. If a self-driving car  were hallucinate a red light for a green light, [17:47] we would not like that. It might actually hurt a  person. Um, I would not necessarily look at the [17:53] past as a predictor for the future. We live in  an exponential era and the progress has become [17:59] exponential. And as a result, what might have  taken 20 years in the past will soon take three, [18:07] four, five years, two years, one year. And we're  seeing this in many areas of technology today. Um, [18:13] and it's going to embrace safety related  physical AI as much as it does virtual [18:17] AI. Yeah. All right. Um, do you have any other  reflections that would be of interest to this [18:25] audience of people? You asked about lessons.  I have one lesson I want to get rid of here [18:30] which is about regulators. Very early on in the  project we recognized that the regulators actually [18:35] our best friends and we treated them as such.  We would meet with them and say look you guys [18:38] are amazing. You and I want the same thing. Uh  you want safety we want safety. Obviously it's [18:44] in everyone's interest to be safe and then you  want innovation as much as we want innovation. [18:48] So how how can we innovate in that basis at  least when I was on the helmet of the company [18:53] um we had an incredibly positive relationship  and reception with regulators who constructively [18:58] thought about how can we safely evolve this  technology and I saying this because many [19:02] of my friends in AI have a different attitude  they basically say we hate Washington we hate [19:07] regulators they're against us and why it might  feel like this I would recommend everybody to [19:12] think about regulators as your friend because  they actually are well I would say yeah respond [19:19] to that I say the times right now are absolutely  incredible and unique progress on so many fronts [19:28] is exponential and all of those exponential curves  when combined together just amplify the progress [19:36] that can be made and we've seen incredible  progress in the digital AI in the last two [19:41] years I think the next few years are going to  be also in addition to that about physical AI [19:46] So if you're in the space, go build some amazing  stuff on the app. All right. Thank you very much.