Cruise Cofounder Kyle Vogt Targets Household Robotics With New Lean Venture

Cruise联合创始人Kyle Vogt创立精益新企业，发力家庭机器人领域 Vogt预测，通过利用神经网络规避传统的软件臃肿和监管瓶颈，下一家千亿美元级公司的员工规模将不足100人。 在本次对话中，连续创业者Kyle Vogt阐述了他对新创企业The Bot Company在家庭机器人领域的愿景，解释了现代LLM和神经网络如何最终使通用任务成为可能。结合领导Cruise十年的经验，他将家庭机器人所需的较低可靠性门槛，与自动驾驶汽车所需的“五个九”（99.999%）极高安全标准进行了对比。Vogt还批评了传统硅谷扩张模式下的企业臃肿，指出了Tesla和Waymo等竞争对手的技术局限，并最终表明他致力于打造一支精简、高度专注的工程团队，且誓言绝不出售这家公司。 The Bot Company的核心使命是免除人们每周花费在无薪、低技能家务上的5到10个小时。 (0:00) 得益于大语言模型（LLM）以及用于机器人控制的端到端神经网络的最新进展，通用家庭机器人终于成为可能。 (0:00) 与洗碗机或烤面包机等单一功能电器相比，将多项家务整合到一台通用机器中能为消费者创造巨大的价值。 (2:05) The Bot Company在初期将避开洗碗和洗衣等需求最高的家务，因为消费者对这些任务的要求极高，且现有的家用电器已经高度优化。 (2:05) 当机器人能够根据一条简单的高级指令在家庭中执行通用的、未预先编程的任务时，机器人领域真正的“图灵测试”里程碑就将达成。 (4:27)

Serial entrepreneur Kyle Vogt argues that recent advances in large language models and neural networks have finally made general-purpose home robotics viable, but success requires avoiding both overhyped capability claims and corporate bloat. Modern AI and Household Robotics Strategy (0:00) Vogt contends that household robotics was fundamentally blocked until the emergence of large language models and end-to-end neural networks for physical control. Unlike single-function appliances such as dishwashers or toasters, which consumers evaluate strictly on narrow cost-to-value metrics, a multitask home robot becomes compelling by aggregating the five to ten hours of unpaid chores that individuals perform each week. Rather than programming rigid geometric motions with fragile algorithms and expensive laser scanners, modern systems navigate chaotic, non-standardized home environments through dynamic neural networks that can self-correct and learn from demonstration. Counterintuitively, The Bot Company is not launching with the top requested consumer tasks—washing dishes and doing laundry. Vogt notes that consumer tolerances for mistakes in those categories are unforgiving, and robotic solutions face entrenched competition from mature, dedicated appliances. Instead, the company focuses on the multitude of minor daily tasks, like picking up clutter. For Vogt, the robotics equivalent of the Turing test will arrive when a consumer can issue an open-ended instruction—such as gathering, washing, folding, and putting away bedroom laundry—without explicit, step-by-step programming. Reliability Thresholds and Practical Commercialization (7:37) A critical challenge in robotics is crossing the utility threshold where a machine produces more value than it consumes in maintenance. Vogt contrasts home robotics with autonomous vehicles: while self-driving cars require five or six '9s' of reliability (99.999% to 99.9999%) to operate safely on public roads, consumer interest in home assistance ignites at roughly 95% reliability. In one prototype test, a robot retrieved 48 out of 49 scattered toys over 30 minutes; observers immediately sought to buy it despite the single failure. Because each additional '9' requires ten times more engineering effort, Vogt warns that social media hype around polished video demos risks triggering consumer disappointment if edge cases are overlooked. To manage development risk, Vogt embraces a commercial approach modeled on Elon Musk's early strategies at Tesla and SpaceX. Rather than remaining locked in pre-revenue research and development for a decade, companies must commercialize products using technology that is already robust and affordable. The Bot Company also reduces hardware iteration lag by seating software developers within ten feet of physical robots in an entirely in-person office, enabling real-time code testing directly on the machines. Autonomous Vehicle Lessons: Regulation and Sensors (12:59) Drawing on his decade leading Cruise, Vogt describes autonomous vehicle development as perhaps the most capital-intensive pre-revenue effort in history, reliant on deep corporate backing to fund massive upfront investments. Today, the primary operational bottleneck in the United States is regulatory fragmentation. Because there is no federal preemption, autonomous vehicle fleets must negotiate safety approvals city by city and state by state across dozens of local entities, slowing the deployment of safety-enhancing systems. On the technical debate between sensor suites, Vogt views Tesla's end-to-end vision neural network as the correct long-term trajectory, though he argues Tesla imposed excessive engineering constraints by refusing to use radar or lidar to lower costs and preserve vehicle aesthetics. Conversely, Waymo built a validated, highly reliable system on classical computer vision and motion planning, but faces the logistically intractable burden of maintaining real-time 3D maps of every operating area alongside expensive roof-mounted sensor hardware. Vogt notes that transitioning Waymo's production fleet toward an end-to-end neural approach carries significant operational risk. Lean Teams and Retaining Entrepreneurial Control (17:55) Vogt argues that standard Silicon Valley growth practices produce bloated corporate hierarchies that destroy execution speed. In his estimation, when a startup expands from 80 to 400 employees, individual productivity can drop by 90%, with recovery taking years. Consequently, Vogt intends to keep The Bot Company under 100 people across the next three years, with roughly 95% of staff working directly as engineers. He forecasts that the next $100 billion company founded in 2025 or 2026 will maintain a headcount below 100. Vogt also maintains that he will never sell a company again, framing acquisitions as the surrender of the building process rather than a victory. Reflecting on his sale of Cruise to General Motors, Vogt admits he naively viewed the automaker's scale as a shortcut to deploy self-driving vehicles quickly. He discovered that large corporations function like aircraft carriers that cannot be steered rapidly, concluding that selling ultimately takes a founder off the problem they set out to solve. Vogt's thesis rests on the premise that end-to-end AI models can finally deliver viable consumer robotics, provided startups avoid overpromising on complex household chores, commercialize incrementally, and refuse the bloat of traditional corporate scaling.

连续创业者 Kyle Vogt 结合其创立 Twitch、Cruise 及 The Bot Company 的经验指出，大语言模型与端到端神经网络正推动通用家用机器人走出学术瓶颈；同时他反思了自动驾驶的发展得失，主张深科技企业应采取务实的早期商业化路线，并通过极度精简的研发团队规避大企业管理陷阱。 家用机器人务实路径与范式革新 (0:00) Vogt 认为，现代人每周在无偿家务上耗费 5 至 10 小时，而现有家电自 20 世纪五六十年代技术井喷后基本处于单功能停滞状态，消费者不会为单功能机器支付高昂溢价，但能够整合整理玩具、清空水槽、搬运包裹等杂务的通用多任务机器则极具价值。针对市场最迫切的洗碗和洗衣需求，The Bot Company 反其道而行之，选择先不予涉足。Vogt 解释称，用户对洗烘衣物与洗碗的标准极其严苛，机器人犯错成本过高且必须直面成熟家电的竞争；因此企业应从家庭日常的“一千件琐事”切入，而非押注目前技术尚难实现的替代级全能人形管家。 家庭环境对机器而言堪比“忍者闯关”般混乱，但机器人底层技术范式已与五年前截然不同。行业已从依赖昂贵激光扫描与脆弱的人工编码路径规划，全面转向借助类 ChatGPT 技术与基于演示模仿的自适应神经网络，机器人面对失误已具备动态纠错能力。在 Vogt 看来，机器人的“图灵测试”不在于实验室内的折叠衣物玩具场景，而是当消费者把设备买回家后，给出未预设程序的自然语言指令——如要求其将卧室衣物送洗、折叠并归位，机器人能够自主流畅完成全套泛化操作。 克服95%可靠性假象与硬件敏捷迭代 (7:37) 家用机器人极易被社交媒体上的光鲜演示推入炒作周期，其真实难度在于解决长尾边界情况。Vogt 分享了早期原型测试的案例：机器人在 30 分钟内收拾了散落一地的 49 件儿童玩具中的 48 件，这种约 95% 的可靠性足以让现场观众迫切询问购买渠道；然而在自动驾驶领域，系统需要达到五个或六个“9”（99.999% 以上）的可靠性才具备商业化资格，且每增加一个“9”都需要十倍的工程投入。Vogt 警告，若仅凭精巧的视频剪辑过度拔高消费者预期，行业必将迎来失望周期。 为了加速硬件落地，The Bot Company 借鉴了 Elon Musk 擅长的商业打法：在长期研发期间尽早通过具备商业可行性的产品获取收入，正如特斯拉早期推出低产量的 Roadster、SpaceX 靠 Falcon 1 产生现金流。在团队管理上，Vogt 极力抹平硬件与软件的迭代鸿沟，要求全员线下办公，工程师办公桌距离实体机器人不足 10 英尺，实现代码编写后几秒内即可在物理样机上即时验证，彻底摒弃以周或月为周期的迟缓反馈机制。 自动驾驶技术路线之争与监管碎片化 (15:11) 回顾领导 Cruise 的十年，Vogt 指出自动驾驶技术是商业史上吸纳资金最多的预收入产品，必须依赖通用汽车等具备雄厚财力的大型主机厂承担巨大的初始活化能；相比之下，前沿大语言模型虽资本开支巨大，却享有清晰且即时的销售变现模型。然而，美国自动驾驶的全面普及正受困于联邦立法的缺位。由于缺乏联邦层面的优先管辖权与统一步骤，企业在每个城市落地前必须逐一游说数十个地方工会、市议会与监管实体，地方利益群体的阻挠严重拖慢了这项经数据证实能减少伤亡的技术的推广。 在底层技术路线上，行业呈现出鲜明对立。Vogt 评价特斯拉主导的端到端神经网络代表了正确的长期技术方向，但特斯拉为其工程师套上了过于苛刻的枷锁——既要控制成本至竞品的十分之一，又因美学设计完全摒弃激光雷达与毫米波雷达。相反，Waymo 延续了 DARPA 挑战赛时期的经典机器视觉与路径规划体系，虽在公开道路上安全有效，但长远来看，全球实时维护高精 3D 地图且在车顶安装庞大机械扫描结构的方案在商业上不可持续；Waymo 当前试图向神经网络转型，又必须承受动摇既有安全验证体系的巨大风险。 警惕组织规模膨胀与绝不再出售公司 (17:55) 谈及组织架构时，Vogt 严厉批判了硅谷迷信复杂管理层级的企业文化。他指出，随着团队从 80 人急剧扩充至 400 人，人均生产效率往往会断崖式下跌 90%，大多数创始人在盲目追逐人员扩张时从未理性权衡这种生产力折损。此外，远程办公往往是企业无法在单一地点吸纳足够人才时的无奈妥协，但线下面对面的无缝协作对硬科技攻关依然至关重要。 基于 Cruise 被收购后的经历，Vogt 明确声明自己未来绝不会再出售任何一手创立的公司。他反思当初将 Cruise 出售给通用汽车的想法过于天真，误以为背靠大型车企的体量能加速自动驾驶普及；事实上，巨型企业如同难以转向的“航空母舰”，出售公司实质上宣告创始人丧失对愿景的主导权并被迫退出一线。展望 The Bot Company 的未来三年，Vogt 规划将团队严格控制在 100 人以内（工程师占比 95%），并大胆预测在 2025 至 2026 年间诞生的下一家千亿美元级企业，其员工总数很可能不会超过 100 人。 Kyle Vogt 展现出从重资本驱动向极度精简、强执行力工程师文化回归的创业转向，强调以端到端神经网络与渐进式商业变现破除长周期研发的资本陷阱。然而，避开洗衣等核心高频家务能否建立起足够深厚的用户粘性，以及不足百人的微型团队能否独自支撑起复杂的硬件供应链与大规模量产交付，仍是其愿景中尚待验证的关键环节。

[0:00] Kyle Vogt: holy grail of robotics [0:00] Kyle Vogt: thousand small things [0:00] Kyle Vogt: reaching, perhaps [4:25] Kyle Vogt: with no celebration or fanfare whatsoever [4:25] Kyle Vogt: My clothes are in my bedroom, please put them in the laundry machine and fold them and put them away. [9:47] Kyle Vogt: deceptively simple-looking from the outside [9:47] Kyle Vogt: Where do I buy this? I need it now. [15:11] Waymo/Kyle Vogt: giant spinning KFC buckets on the roof [17:55] Kyle Vogt: dogma of building a Silicon Valley tech company [17:55] Kyle Vogt: productivity per person is gonna drop by 90% [19:21] Kyle Vogt: why would you stop or give up control of that thing? [19:21] Kyle Vogt: aircraft carrier [19:21] Kyle Vogt: I'm done working on the problem [21:39] Kyle Vogt: the next $100 billion company that's created in 2025, 2026 will be under a hundred people

[0:00] Homes are just horribly difficult environment for robots. [0:03] I mean, when you put it that way, [0:04] it sounds like the "Ninja Warrior" obstacle course. [0:06] I think it will be strange to move into a home [0:08] or apartment in five years that doesn't have a home robot. [0:11] So you're very serious about the small team. [0:14] I think the next $100 billion company that's created, [0:17] you know, in 2025, 2026, will be under a hundred people. [0:23] Kyle is one of the only entrepreneurs [0:24] to have started three separate billion dollar companies. [0:26] He started Twitch, Cruise, [0:29] and now he just started The Bot Company. [0:30] He's trying to make household robots finally happen. [0:34] Cheers. [0:37] Okay, so let's back up. [0:39] What is the pitch for The Bot Company? [0:44] Well, I don't like doing chores [0:46] and I think five to 10 hours a week people spend doing [0:50] essentially unpaid, unskilled labor in their own home. [0:54] Yet we all take that for granted, and do it every day. [0:56] I think it's been the holy grail of robotics, [0:59] you know, since I was a kid doing robotics, [1:00] to have the home robot that does everything. [1:02] And it was very clearly not possible [1:05] up until very recently with LLMs [1:08] and then end-to-end neural networks to control robots. [1:11] I think just maybe, maybe this time is the right time [1:14] to build this company and deliver this home robot [1:16] that does all the things that you don't wanna do. [1:18] Okay, so this is vacuuming the floor, ironing the clothes, [1:23] cleaning up after the pets, that kind of stuff. [1:25] We're gonna start very small, [1:26] but it will continue to evolve as the state of AI evolves [1:30] and be able to do more and more things in your house. [1:32] I think it will be strange to move into a home [1:34] or apartment in five years that doesn't have a home robot. [1:37] Or you won't want to without one. [1:40] In the same way that it would feel weird [1:42] to not have like plumbing in a home, [1:44] or a dishwasher if you can afford one, or laundry machines. [1:47] These are all machines that had a huge impact on our lives. [1:50] And these are all from the era of, I dunno, [1:52] the '50s and '60s and we haven't really had that next. [1:54] Yeah, we had a big spurt for a while. [1:56] And people were excited about it. [1:57] There's all these like great ads of like, you know, [1:59] the person in the kitchen being like, [2:00] "Look at how much time I've saved with the microwave." [2:02] And you know, we've gone stagnant since then. [2:05] You could be bearish on home robotics. [2:06] And an argument I think you could construct is [2:08] we are currently so underperforming what we could have [2:14] in terms of household appliances. [2:16] Dishwashers take forever to run [2:19] and don't clean the dishes that well, [2:21] and commercial dishwashers exist [2:23] that are super fast and much more effective [2:25] but we don't do those in our homes. [2:26] You know, the toaster can't tell [2:28] when it's burning the toast, [2:29] even though that seems fairly trivial to detect. [2:32] Should this make one worried about future household robots? [2:35] All of the things you described- [2:37] the toaster, the dishwasher- [2:39] these are like single function machines. [2:42] And I think everyone, [2:43] when they buy a single-function machine [2:45] or any machine is thinking about like the cost and value. [2:47] How much is it worth to me to have toasted bread? [2:50] Maybe $30 for a toaster, [2:52] maybe not $2,000 for a multi-toaster [2:55] that can do it perfectly. [2:56] But the question becomes, if you have a machine [2:58] that is a multitask machine [3:00] and can do lots of small things, [3:01] that you would maybe even pay $0 for [3:03] if it was a standalone machine, [3:04] but when it's bundled into this general-purpose machine [3:07] that can pick up all the kids' toys, [3:09] and clean the dishes off the table, and put stuff in the sink, [3:13] and pick up your packages from the front door, [3:15] and bring them to the kitchen... [3:16] I would never pay any money for a machine [3:18] that just does one of those things. [3:19] But when assembled, I think it becomes extremely valuable, [3:23] or at least that is our theory. [3:24] When you ask people, "If you had a home robot, [3:26] what would you want it to do," [3:28] one of the top three things is: [3:30] "Do my dishes" or "Do my laundry." [3:32] I think those are great tasks to automate. [3:35] I think they are very poor tasks [3:36] to start with, counterintuitively. [3:38] And the reason for that is laundry and dishes are things [3:41] which people are very particular about, [3:43] and the cost of making a mistake is very high. [3:45] You don't wanna like ruin- [3:46] So you don't wanna start with those? [3:47] We will not start with those, [3:49] and that's because we have existing machines [3:51] that do these things. So you're competing [3:52] with the dishwasher, competing with the laundry machine. [3:55] But in between those tasks are a thousand small things [3:57] that we spend our time doing every day around the home, [4:00] and it's our hope that solving those things [4:02] really moves the needle for people. [4:03] And then, of course, over time, [4:06] as the technology improves and I can confidently say to you, [4:08] "We can do the dishes in the exact way that you want," [4:10] then we'll deliver that experience, but not before. [4:12] There are people doing the other thing right now [4:13] with like humanoids, promising they'll be a drop-in [4:16] replacement for human labor and a drop-in replacement [4:18] for a housekeeper on day one. [4:20] Seems really hard. [4:21] I think that's reaching, perhaps. [4:23] Yes. Yes, [4:24] We'll see. [4:25] Okay. People in AI talk about the Turing test, [4:30] where basically, can you have a five-minute conversation [4:32] with the AI over text [4:35] and be able to tell that it's an AI is [4:38] kinda the loose meaning of the term. [4:41] I was gonna ask, when did we cross that threshold? [4:42] It's hard to say. [4:42] I mean, I feel like we've clearly crossed it. [4:45] With no celebration or fanfare whatsoever. [4:47] Exactly with no celebration, no fanfare, in the past few years. [4:50] So, what is the Turing test for robotics? [4:54] Well, where my head is drawn to is some [4:58] of the toy problems in academia, like T-shirt folding. [5:01] There are robots and neural networks now [5:03] that can fold T-shirts. [5:04] So perhaps in the same way [5:05] that the Turing test was crossed [5:06] at some point in time, we're not sure exactly when, [5:08] we just know it's behind us, already for robotics- [5:10] Yes. If you have a set of robot arms that can fold T-shirts, [5:12] it feels like we have crossed some level of- [5:13] That used to be the holy grail of manipulation, [5:15] because classically robots are designed [5:19] for repeatability and precision [5:21] and picking up the same thing in the same place every time. [5:23] And for that to work, [5:24] the thing you pick up also has to be rigid. [5:26] Clothes are hard because you pick them up [5:28] and they collapse and wrinkle and fold over on themselves. [5:30] So to get a machine that thinks in this very rigid world [5:33] to work with such malleable items has been [5:37] this tough research problem for a long time. [5:39] For anyone to be able to go buy a thing, [5:41] put it in their home, and without any other instruction [5:43] than, "My clothes are in my bedroom, [5:46] , please put them in the laundry machine [5:47] and fold them and put them away..." [5:49] That, to me, would signify I think we've made it. [5:51] Yeah. You've talked in the past [5:53] about how homes are just a horribly difficult environment [5:57] for robots. [5:58] If you compare to, say, a warehouse, [6:00] where there's a decent amount of robotics already, [6:02] a warehouse is very standardized, [6:04] and standardized to be easy for the robots, [6:06] whereas homes are an extremely difficult [6:08] and non-standardized environment for the robots. [6:10] They have stairs, they have rugs, [6:12] they have kids and pets running around. [6:15] How do you solve for this? [6:17] I mean, when you put it that way, [6:18] it sounds like the "Ninja Warrior" obstacle course. [6:20] Yes, exactly. ...of robotics, basically, [6:21] and robots have to get through all this stuff. [6:23] Well, I think what's different today, [6:24] and actually, [6:26] if there's one takeaway for today in our conversation, [6:28] I think it would be that robotics today is [6:31] a completely different field and a different industry [6:32] than it was five years ago. [6:34] All of the things we thought we knew, [6:35] all of the businesses that were tried and failed, [6:38] all of the tools that have become best practices [6:40] and standard are now either like worthless [6:42] or completely different. [6:44] Today you don't need a robot that's repeatable, [6:46] you need a robot that's adaptable, [6:48] powered by neural networks. [6:49] And if it makes a mistake [6:50] or doesn't approach this object exactly at the right angle, [6:53] it doesn't matter. It can correct that. [6:55] Like, how a robot sees the world: [6:57] typically it would have expensive laser scanners [6:59] and try to perfectly reconstruct everything that we see, [7:02] and then use very complicated and hard-to-tune algorithms [7:06] to plan how this arm should move through space and time [7:09] to accomplish a task. [7:10] Those systems are very fragile and easy to break. [7:14] If you're willing to completely let go of that [7:16] and embrace today's tools- [7:18] so, you don't have to program anything, [7:19] you have to show a robot how to do a task [7:21] and it can mimic that, and let [7:24] essentially ChatGPT-like technologies instruct these things [7:27] at a high level- [7:28] with these tools, going into a home environment is no longer [7:32] as much of a crazy obstacle course. [7:34] And that, I think, makes it much more tractable [7:36] than it was five years ago. [7:37] For a device to save you work, [7:40] it probably needs to be able to charge itself, [7:42] clean itself, manipulate stairs. [7:45] You probably need some minimum set of functionality. [7:47] The early versions of Roomba is, I think [7:49] part of what frustrated people is [7:50] they probably spent more time cleaning [7:52] the dog poop they spread, [7:53] or it would've been faster for me to just vacuum it myself. [7:56] I think it is dangerous to fall below the threshold of [7:58] creating more work than you, or more value. [8:00] I think that's probably the difference [8:01] between a cool product and a delightful product [8:03] that everyone loves. [8:04] I think some of the things you mentioned are [8:06] what make this problem, as with many AI-powered problems, [8:09] deceptively simple-looking from the outside. [8:11] Yeah, as in it's easy to have a cool demo, [8:13] and it's hard to have something [8:14] that actually saves people time in their home? [8:15] Yeah. And I mean, [8:16] I worked on self-driving cars for a long time. [8:18] We saw this too. It's like, "Well, how hard could it be [8:20] to keep the car between the two yellow lines on the road?" [8:23] And then you think about all the things [8:24] that could go wrong or could happen while you're driving, [8:26] and you start making a list, [8:28] and then you have like three pages of stuff. [8:29] Each one of those is a big technical problem to solve. [8:32] I think we're already seeing this, [8:33] but a similar thing is true for a home robot [8:35] that truly creates more value [8:37] and frees you of work [8:39] rather than consuming all your time [8:40] or asking for help every five minutes. [8:41] Yeah. Is there a risk [8:42] that home robotics have a similar character, [8:46] where that final 1% actually turns out [8:49] to take a pretty long time? [8:50] It's possible, I think. [8:51] What is different to me are a few things. [8:54] I think first of all, just to build a car [8:58] in a highly regulated environment, [8:59] it's a very capital-intensive thing, [9:00] is very different than to build a small consumer product. [9:03] But the other thing that I noticed, several months ago, [9:06] one of our early prototypes, [9:07] we do this thing where we just like dump a basket [9:10] full of kids' toys in a room and say, [9:12] "Hey, robot, clean this up." [9:14] There's 49 toys on the ground. [9:16] And over the course of 30 minutes- [9:18] it took it a long time, this is a prototype- [9:20] it cleaned up all the toys but one. [9:22] And my thought in that moment was, [9:25] "What percentage of success is that? [9:26] That's like 95%, one 9 of reliability." [9:29] Yet everyone who was watching that was just like, [9:31] "Where do I buy this? I need it now." [9:33] And so the takeaway for me is: The bar [9:35] for commercial success for self-driving was [9:37] five, six 9s of reliability. [9:39] And understand that each extra 9 of reliability you add, [9:43] so 10 times better, [9:44] takes probably 10 times more engineering work. [9:46] Robotics is perfect for getting over hyped on social media, [9:52] because it's very easy to have a compelling demo [9:56] that does the numbers on a tweet, [9:58] and all the work is in getting from that demo [10:01] to actually working reliably enough to sell as a product. [10:04] So it feels like we're almost inevitably in [10:06] for a hype cycle in robotics. [10:08] Well, look, I love the demos. [10:09] I think they're inspirational. They get people excited, [10:11] they get more people coming into the industry, [10:12] they get investment dollars. [10:14] So I think they have a purpose. [10:16] I think the problem is when you align customer expectations [10:19] too squarely on like what they see in a demo. [10:21] Or even as an industry, [10:23] if the robotics industry sets the expectations too high [10:26] on a whole for what the next generation of robots will do, [10:29] everyone's gonna be disappointed. [10:30] And I think it without a doubt will happen in robotics, [10:33] and not because of any one bad player, [10:35] more just the natural way that these things go. [10:38] What is your iteration loop for working on robotics? [10:43] Well, I mean, [10:44] if you have a weekly release schedule [10:46] or monthly release schedule, [10:47] what you're really doing is just withholding [10:50] all that useful feedback an arbitrary number of days [10:53] or weeks. [10:54] Doing hardware often requires a lot more upfront thought [10:58] and planning. [10:59] There's lead times, there's manufacturing times, [11:01] all that kind of stuff. [11:02] And so you have to use one process for that [11:04] and it's more schedule driven, [11:06] and then another process for software, [11:07] which is much more iterative, [11:09] because you can make changes on the fly. [11:11] Bringing those together can be tricky, [11:14] but if you set it up right, [11:15] you can actually have hardware development feel [11:18] like software development. Simple hacks: [11:20] You have everyone work in person, in an office, [11:23] you have lots of robots available for developers [11:25] to push code to in real time, [11:27] and you make it like as frictionless as possible [11:29] for people to like try out new stuff on a real machine. [11:31] I don't think you have to walk more than 10 feet [11:34] in our office to go from your desk [11:36] to running code on a robot. [11:39] I feel like one of the underappreciated aspects [11:42] of Elon's playbook for building companies is how much [11:47] of a commercial thinker he is. [11:49] Elon's companies have actually always been [11:51] surprisingly scrappy. [11:52] I mean, famously with the Tesla master plan, [11:54] they started with the Roadster, [11:55] which was deliberately a low-volume car, [11:57] and then kind worked their way up to higher-volume cars. [12:00] With SpaceX, they were selling launches to orbit [12:03] from a very early stage and then progressively [12:05] moved from Falcon to Falcon Heavy to Starship. [12:09] How do you pull the revenue forward [12:12] as early as possible in a robotics company [12:15] so that you're not doing 10 years of R&D [12:18] and then eventually selling a product? [12:20] I think the way that you do that is by understanding [12:24] where the absolute frontier is for technology, [12:27] and then understanding what is commercializable [12:29] in the near term. [12:29] There's usually a gap. [12:31] It can be a small gap or a large gap. [12:33] So if a technology has gone through enough cycles [12:35] of investment by enough companies, or you've done it in house [12:39] and it's at the point where now it's [12:41] affordable, robust, and can work, [12:44] then I think you can build a business [12:45] and get to revenue quickly. [12:47] The problem comes in when you have a business [12:49] that is premised on, or conditioned [12:51] upon, commercializing today's frontier of technology, [12:55] because that will just take time, [12:56] and we don't know if that's like one year or 10 years. [12:59] Let's talk about self-driving. [13:00] You co-founded Cruise, which was acquired by General Motors. [13:06] Is self-driving [13:08] the most capital-intensive pre-revenue product ever? [13:15] It's hard to think of a counterexample. [13:19] I don't have a good one either. [13:20] I think it's insanely capital-intensive, [13:22] and notably, the companies [13:24] who were making these investments were not startups [13:27] that were just doing this by raising venture capital rounds. [13:29] They were large corporations with R&D budgets, [13:33] or basically the pockets that were deep enough [13:36] to make strategic long-term bets [13:39] that could significantly move the needle for the company [13:40] knowing that [13:42] there's a significant activation energy [13:43] to unlock that future value. [13:45] Yeah. Previously it was probably just [13:47] governmental entities, [13:49] and only as of recently do we have companies [13:51] that are going to spend that much money pre-revenue. [13:54] The numbers coming out around large language models [13:56] on the frontier, though, are are getting up [13:58] into that territory. [13:59] They are getting up into that territory. [14:00] But I think with pretty clear user economics, [14:03] where they actually sell a lot of AI these days- [14:06] Healthier business for sure. [14:07] Exactly. Self-driving, it is interesting [14:09] because it was so unproven when all the CapEx was required. [14:13] Self-driving is having a real moment right now [14:18] as we finally see a lot of deployment [14:22] on the streets in volume. [14:25] You worked on this for 10 years. [14:27] How do your industry views differ? [14:30] One is, I think, on the regulatory side, [14:36] and what it will take to truly reach large scale [14:39] for these businesses. [14:40] Right now there's a handful of players [14:42] who are actually doing robotaxis or driverless trucking. [14:45] And then the other is [14:46] like these seemingly diametrically opposed strategies [14:49] of Tesla and Waymo, which everyone likes to talk about. [14:53] So, the less interesting regulatory one first, [14:56] and get it outta the way. [14:57] In the US, [14:59] it is still very much a patchwork of legislation. [15:02] What probably most people don't see, [15:06] Waymo or someone doing is all the groundwork [15:08] in each new city they do. [15:09] The groundwork they're doing is [15:11] because they don't know which small special interest group, [15:14] or union, or local government, or city council, [15:18] or state, you know, whatever it is... [15:20] There's probably two dozen lists of organizations [15:23] that could meaningfully bring the thing [15:24] to a halt in that community [15:26] because there is no federal preemption, [15:28] there's no real federal safety standards [15:30] for autonomous vehicles. [15:31] And so they have to win that battle [15:33] with every single stakeholder in every single location. [15:37] So I hope... And there's maybe some signs of this, [15:39] that the federal government will get ahead of this [15:43] and establish that [15:44] it's pretty clear at this point, that the data shows [15:48] that these cars are saving lives, and reducing crashes. [15:51] So if we think that's important as a government, [15:53] maybe there should be federal preemption, [15:55] and we should ensure that this is open [15:56] for everyone in the US. [15:58] If that happens, I think we'll see more self-driving cars. [16:01] Absent that, I think it's gonna continue [16:02] this really slow sort of city-by-city thing. [16:05] And in the interim, [16:07] a lot of people are going to get hurt [16:08] because these aren't rolling out faster. [16:09] The other big, perhaps false dichotomy [16:12] that people create is LIDAR versus cameras. [16:15] What I see is really Tesla, as a company [16:19] who pioneered the end-to-end neural network approach [16:23] to self-driving, [16:25] which I think is the right technical bet long-term, [16:28] but they put some constraints on it. [16:30] They said, "Hey, engineers, [16:31] you can't have the best sensors," like LIDARs and radars, [16:35] "and the sensors have to look good [16:37] when we put them on the car. [16:38] Oh, and by the way, they have to cost one-tenth as much as [16:40] the guys down the street who are doing this." [16:43] They put some crazy constraints on that. [16:45] So the right technical vector, [16:47] but like really being held back [16:50] by the weight of all these constraints [16:53] that were put on the system. [16:55] But all of their technical approach [16:56] from day one seems to have been pointed [16:58] in the right long-term direction. So that's good. [17:00] With Waymo, they started off [17:02] in the DARPA Grand Challenge era of self-driving, [17:05] which is old-school, classical computer vision, [17:08] classical motion planning. [17:10] And they built this highly-validated, robust system [17:13] that's now on public roads, and it's great, [17:16] but they know that it's the wrong technical approach, [17:18] and they need to move more in the direction of Tesla, [17:20] of more neural networks. [17:22] It's the wrong technical approach [17:23] because it's too expensive? [17:25] Because it is just intractable to maintain a 3D map [17:28] of every square inch of the planet [17:29] and update it in real time, [17:30] and then expect that every time you go somewhere [17:32] the map is still accurate, on one hand. [17:34] And it's also probably unrealistic [17:36] to assume that every car built in the future is going to have [17:38] these giant spinning KFC buckets on the roof. [17:40] To Waymo's credit, I think they know this, [17:42] and they've started moving towards a Tesla-like approach. [17:46] The challenge is, [17:46] they've got a validated safety-critical system on the road, [17:49] and the last thing you want to do [17:50] to a system like that is start changing stuff in it [17:53] because that introduces risk. [17:55] Now that you have a little bit of distance [17:56] from the Cruise experience, what are your reflections? [17:59] Oh, well, many. [18:00] I think a lot of people over-rotate [18:02] on things they would change even the next time around. [18:05] So, The Bot Company is a small company. [18:07] I feel like I, like many of my peers, [18:10] got swept into the dogma [18:11] of building a Silicon Valley tech company, [18:13] which is lots of people, and you have the manager, [18:17] senior manager, director, senior director, VP hierarchy- [18:20] all these like structures that are designed [18:22] to get a lot of people to work well together, [18:26] and they become horribly inefficient, [18:27] and it's very easy for them to become bloated. [18:30] What else? [18:31] I believe in the in-person environment. [18:33] I think everyone ran various experiments of remote work [18:36] during Covid, and has ended up, [18:38] depending on the company, [18:38] in terms of full return to work, or remaining some- [18:41] That pairs also with the small company thing, right, [18:43] where I think anyone would say [18:44] that if you're hiring a small team, it's very attractive... [18:47] Like, the reason companies tend to go remote, [18:49] or certainly go to multiple offices, is [18:51] just ultimately, you need to hire so many people [18:53] that the diminishing marginal returns to being together- [18:55] Well, not to harp on this one thing, [18:57] but there's so many dimensions of it. [18:58] I don't think most people building companies today have [19:01] a conscious decision and say, [19:03] "Well, when we go from 80 people to 400 people, [19:07] our productivity per person is gonna drop by 90%, [19:10] and are we gonna sign up for that [19:11] and understand that it won't get better [19:13] until we're past 400 people?" [19:15] I mean, that's the reality of the situation. [19:17] Not many people like talking about it. [19:18] I think that's a big one. [19:21] You said you're never gonna sell a company again? [19:22] Yeah. [19:24] Why? [19:26] Well, it's- [19:27] Like why is it not, [19:29] "I'm going to be very careful to sell a company [19:32] to the right acquirer with the right vision," [19:34] or a more nuanced statement? [19:36] Let's flip this around. [19:37] If you go through all the pain of starting a company, [19:42] and you do so knowing that you're going to spend 10 plus years [19:45] of your life on something, and it's that important to you, [19:49] and you've told everyone you know about this thing, [19:51] and you've recruited all the best, [19:53] the smartest people in the world that you know [19:54] to work with you on this thing, why would you stop [20:00] or give up control of that thing? [20:02] I think that [20:03] maybe part of the dogma of Silicon Valley is: [20:06] you start a company, [20:07] if you're lucky enough and it's growing fast enough [20:09] someone will make an offer to buy it, and you sell it, [20:11] and that's victory. [20:13] I think if [20:14] financial outcomes are your reward function, [20:18] or fame, or whatever it is, then that's great. [20:21] But I've talked to a lot of people [20:22] who have gone through that, [20:23] and they miss building the company. [20:25] They would've preferred to keep going. [20:26] The robots don't want you to sell. [20:29] Maybe if they're sentient. [20:30] But yeah, I convinced myself when selling Cruise [20:33] that, at the time, [20:36] it was North America's largest automaker, [20:38] and if our vision is to get self-driving cars everywhere, [20:41] isn't that true to the vision? [20:42] I think my heart was in the right place, [20:47] but I was naive about the ability [20:49] to get like a large corporation, which is like [20:52] an aircraft carrier- [20:52] you can't steer it, you can't get it to change its focus, [20:55] it's gonna do what it wants to do, [20:56] or what it's already doing- [20:58] it was naive of me to think [20:59] that I could kind of hitch a ride on that scale [21:01] and make this thing happen. [21:03] Experience has told me now [21:07] that that is not the path to make the thing happen. [21:09] How generalizable is this view? [21:11] Do you think fewer founders should sell their companies, [21:14] or is this a Kyle-specific thing? [21:17] I think selling a company basically means, [21:18] "I'm done working on the problem." [21:20] There probably are cases [21:22] where a founder is tired of it, [21:25] their personal relationships are falling apart, whatever. [21:28] There's a external reason to stop going forward. [21:32] Absent that, and if the intrinsic pull is still there, [21:35] then I think it's a bad idea. [21:39] In three years time, [21:41] how many people is The Bot Company, [21:43] and what percentage are engineers? [21:46] Less than 100. [21:51] 95. [21:52] Oh my God. So you're very serious [21:54] about the small team, all engineers? [21:57] I think the next $100 billion company [21:59] that's created in 2025, 2026 [22:02] will be under a hundred people. [22:03] That's quite provocative. How many people are there [22:05] that have created three billion-dollar companies? [22:09] Not that many. [22:11] I have been very lucky. [22:12] Like I said, good people, good timing. [22:15] Yeah. [22:16] But still, if your view is [22:18] that it's a much smaller headcount, [22:21] we might be in for a new way of building companies. [22:24] I hope so. Yeah. Okay. Thank you. [22:27] Yeah, thanks for having me.