2019 年,特斯拉路上跑着大约 40 万辆装了完整传感器套件的车。马斯克在播客里说了一句话:全世界其他厂商加起来,有这个套件的车不会超过 5000 辆。
换句话说,全球 99% 的自动驾驶真实路测数据,在特斯拉一家公司手里。
这场对话发生在麻省理工学院(MIT)一个研究团队发表论文之后。论文的主题是特斯拉 Autopilot 使用过程中驾驶员的“功能性警觉”(functional vigilance)。特斯拉团队主动联系了播客主持人莱克斯·弗里德曼(Lex Fridman),提出做一次访谈。弗里德曼答应了,条件是自己控制问题、自己决定放什么出来,最后一刀没剪。
他也把话说在前头:自己没开过特斯拉,没买过特斯拉股票,特斯拉也没资助过他的研究。在这场对话里,他明确表示和马斯克在一个问题上有分歧,就是摄像头驾驶员监控到底有没有用、能管多久。
这场分歧贯穿整场对话,也恰好是整场对话最有意思的部分。一个是造车的人,相信系统足够好之后人类就是多余的;一个是研究人机交互的人,相信只要人还坐在驾驶位上,就得管住这个人。
仪表盘上那块屏幕,是在让你做健康检查
弗里德曼的第一个问题很朴素:Autopilot 最早的愿景是什么?
马斯克说他不用“愿景”或“梦想”这种词。汽车行业正在发生两场革命,一场是电动化,一场是自动驾驶。他判断的逻辑非常直接:未来一辆不能自己开的车,大概跟一匹马差不多有用。并非完全没用,只是很少见、有点古怪,就像今天谁家里养一匹马。
所以他当时的结论是,如果特斯拉不参与自动驾驶革命,那它的车相对别的车就没有竞争力。一辆能自动驾驶的车,价值可能是一辆不能自动驾驶的车的五到十倍。这个时间窗口,他给的判断是至少五年,也许十年。
接下来弗里德曼问了一个很具体的设计问题:为什么要把传感器看到的东西显示在仪表盘上?
马斯克的回答里有个说法很值得记:那块屏幕的全部意义,是给车辆的“现实感知”做一次健康检查。车从摄像头、雷达、超声波、GPS 里拿数据,把这些数据渲染成向量空间,里面有车道线、红绿灯、其他车辆。然后向量空间再渲染到屏幕上,你扭头看一眼窗外,就能确认车是不是搞清楚了状况。
弗里德曼追问,有没有考虑过把系统的不确定性也显示出来。计算机视觉在边缘地带是有概率的,车道分割、车辆检测,都有置信度。让用户看到这些,能不能帮人建立对机器视觉的直觉?
马斯克说他自己车上一直开着调试视图。调试视图有两种:一种叫增强视觉(augmented vision),在识别到的物体上画框、打标签;一种叫可视化器(visualizer),纯粹是向量空间,没有图像,只有车眼中的世界。但他认为普通人看不懂这些,那对他们就是一堆乱码。
“显示内容是为公众理解系统能力而优化的。”(马斯克)
这里藏着一个产品哲学。屏幕上那块干净的、卡通的、你能一眼对照现实的小车模型,是这个系统面向普通人的说明书。它不追求信息完备,追求的是一眼可验证。你能验证,你才敢用。
算法、数据、硬件,钱该往哪投
弗里德曼把 Autopilot 拆成三块:底层算法(神经网络架构)、训练数据、硬件开发。钱和时间都有限,怎么分配?
马斯克的回答是从数据开始讲的。每辆特斯拉有 8 个朝外的摄像头、1 个雷达、通常 12 个超声波传感器,加上 GPS 和惯性测量单元(IMU)。这套配置的车队规模已经到了 40 万辆,正在逼近 50 万辆。
他估计全世界其他有这个级别传感器配置的车,加起来不会超过 5000 辆。这个数量对比决定了数据流向:特斯拉手上有 99% 的自动驾驶数据。
然后是硬件。他们花了大约三年,做出自己的全自动驾驶计算机(FSD computer)。算力是现在车上那颗英伟达(Nvidia)系统的十倍量级。装法很简单,拔掉英伟达的,插上特斯拉的,完事。
这颗芯片有意思的地方在冗余设计。它其实是两台完整的计算机,两块系统级芯片(SOC),完全冗余。你拿电钻在它任何位置打一个洞,它还能工作。
弗里德曼问,这两块芯片是互为完美复制,还是像“争论机”那样各自做决策?马斯克说纯粹是冗余,不是两个系统各自判断然后投票。他打了个比方:双引擎客机,两个引擎都工作时性能最佳,但一个引擎也能安全飞。
更关键的是,他们还没摸到这颗芯片的性能边界。现在能满帧率、满分辨率跑所有摄像头,图像都不用裁,单颗芯片都还有余量。所以暂时没必要把功能拆到两块芯片上跑,两块都跑完整副本就行。
“所有输入都是错误”
深度学习的老规矩:数据越多越好。但开车这件事,真正值钱的是边缘案例(edge case)。
弗里德曼提到马斯克以前说过,Autopilot 被接管的那一刻是重要信号。他问还有哪些边缘案例值得挖。
马斯克的回答分两层。一层是接管:驾驶员从 Autopilot 手里接过控制权,系统会收到一个触发信号,然后去判断,这次接管是因为方便,还是因为 Autopilot 没干好。另一层是无人接管的样本:比如过一个复杂路口,系统自己走完了,没有干预,那就是一条正样本,对应“路况长这样的时候,应该这么走”。积累大量同一个路口的成功轨迹,就能拟合出最优的过弯曲线。
他给出的原则是这样的:
“把所有输入都看成错误。只要用户动手了,那就是系统出错了。”(马斯克)
这句话是个极强的设计约束。它把“人类接管”这件事的默认解释从“人在帮忙”翻转成“系统失职”。开车的人都知道,很多时候你动方向盘只是习惯、只是想变道、只是不放心。但在这套逻辑里,这些都被归到待排查的案子里。
弗里德曼马上抓住了这个模糊地带:如果驾驶员只是想下高速,这是导航决策,Autopilot 当时还不负责这个,那怎么区分?
马斯克说这个问题会随着 Navigate on Autopilot 的发布消失。变道、下匝道、走高速互通,绝大多数这类操作以后不需要人动手。他补了一句:大家还没意识到这一步有多大。
被低估的一步:不用确认的自动变道
弗里德曼请他梳理 Autopilot 历史上的几次大跨越。
马斯克说这次 Navigate on Autopilot 去掉“确认”环节是巨大的一步。它不只是导航,还会自动超越慢车,会自己找最快的那条车道,会下高速、走互通。另外还有红绿灯识别,公开版本先做成警告,他自己开的开发版已经能完整地在红绿灯前停车起步。
这些是通向完全自动驾驶的台阶。那最大的技术路障是什么?
马斯克的答案是:路障已经搬开了。FSD 计算机已经量产,你订任何一台选装全自动驾驶套件的 Model S、X 或 3,拿到的就是这颗芯片。底座算力够了,剩下的就是打磨神经网络和控制软件,而这些都可以通过空中升级(over-the-air update)推送。
他说这就是投资人日要强调的重点:现在生产的车,硬件已经具备完全自动驾驶能力。但“具备”是个微妙的词。硬件具备,软件要迭代,能力会大幅提升,可靠性会大幅提升,然后拿到监管批准。
于是有了那句在股东信里也会出现的表述:
“如果你今天买一台特斯拉,我相信你买的是一项升值资产,而不是贬值资产。”(马斯克)
这话在汽车行业是反常识的。车是消费品,落地打八折,天经地义。马斯克的逻辑是:硬件是最难升级的部分,现在硬件到位了,剩下的是软件,而软件几乎没有边际成本。所以车会随着软件更新变强,变强就意味着值钱。
弗里德曼顺着问,软件这最后几步有多难。马斯克给的路线图是:高速公路上已经很好用了,接下来把它扩展到城市街道,加上红绿灯识别、复杂路口通行,再然后是复杂的停车场。车能自己从车位出来找你,哪怕停车场像个迷宫;车能把你放下,自己去找车位。
他特别提了停车场,理由很实在:自己停车这件事,烦人程度极高,自动化的收益很大。
和 MIT 研究者的正面分歧
聊到这里,弗里德曼把话题转向人。他说,现在路上测试的 Level 4 车辆,比如 Waymo,严格讲只是技术意义上的自动驾驶,实际上大部分时候车里还坐着安全员盯着系统,本质上是另一种设计哲学的 Level 2。
他问,特斯拉的完全自动驾驶,是不是在一段时间内仍然需要人类监督?
马斯克估计,检测方向盘上有没有手,这个要求大概还会持续六个月左右。真正的问题是监管层面的:Autopilot 要比人类安全多少倍,才允许不监控?他认为需要海量数据,才能在统计上高置信度地证明车比人安全得多,并且证明加上人类监控并不会实质提升安全。
那要安全多少?也许两三倍。怎么证明?看每英里的事故率。事故和死亡。死亡样本量太小,统计上不显著,但事故多得多,可以先看事故概率。然后是受伤概率、永久伤害概率、死亡概率,每一层都要比人类好,好至少 200%。
接下来是这场对话的核心冲突。
弗里德曼介绍了他团队的研究。这项工作从 Autopilot 公开上线那年就开始了,三年多,采集驾驶员面部和身体的视频。他们分析了 18000 多次 Autopilot 脱离,标注驾驶员能不能及时接管,结果是:在他们采集的数据里,驾驶员维持了功能性警觉,人是清醒的、看着路的、能接手的。
这个结果和大量关于自动化警觉的文献预测相反。弗里德曼自己也承认样本有限,问马斯克怎么看这个结论能不能推广到更大的人群。
马斯克的回答是不打算在这上面纠缠:
“我认为这一切很快就会被扫进历史。系统进步太快了,这个问题很快就会变得没有意义。”(马斯克)
他的推理是:如果系统比人安全很多倍,那么加一个人进来,对安全的影响是有限的,甚至是负的。
弗里德曼追问:你的意思是,就算一部分人确实出现了警觉下降,也不会影响整体安全统计?
马斯克说会很快,也许是今年年底,最晚明年,人类介入反而会降低安全性。
然后他讲了电梯。
“以前电梯里有操作员,你没法自己进电梯、自己扳杆子上下楼。现在没人想要电梯操作员,因为自动停靠的电梯比人安全得多。事实上,让一个拿着杆子的人控制电梯上下,是相当危险的。”(马斯克)
弗里德曼并不买账。他说从算法角度,摄像头感知驾驶员在看哪、认知负荷多大、身体姿态如何,是个很迷人的问题。行业里很多人相信摄像头驾驶员监控是必须的。他直接问:驾驶员监控到底有没有价值?
马斯克的回答划了一条清晰的线:如果系统可靠性在人类水平或以下,驾驶员监控有意义;如果系统明显比人可靠得多,驾驶员监控帮不上什么忙。他又回到电梯:你在电梯里,真想要一个拿着大杆子的陌生人帮你操作吗?他更信按钮。
弗里德曼在这个点上保留了意见。他在节目开头就写明了:他相信摄像头驾驶员监控在短期和长期都可能有益,而马斯克和特斯拉的焦点是让 Autopilot 的统计安全收益大到足以覆盖人类行为和心理的所有问题。两人没谈拢,但都把自己的道理说清楚了。
四万条人命和四起特斯拉事故
弗里德曼问,和监管机构能不能有一场健康的对话。
马斯克的回答罕见地直白:监管机构对能上新闻的事情,关注度是严重不成比例的,这是客观事实。特斯拉本身就制造大量新闻。美国每年大约有 4 万起汽车死亡事故,但如果特斯拉出了 4 起,拿到的报道可能是别人的一千倍。
他说这个心理机制其实很有意思,可惜没时间展开。
接着弗里德曼问了个技术问题。有黑客用对抗样本(adversarial example)骗过 Autopilot,让它做出异常反应。神经网络对输入的微小扰动很敏感,这个行业能不能防住?
马斯克说能,而且说得轻描淡写。神经网络本质就是一堆矩阵运算,要构造对抗样本,得是非常懂行的人反向工程出矩阵是怎么搭的,再造一个东西让矩阵算偏一点。而防住它很容易:做一种“反识别”,系统看到像矩阵攻击的东西,直接排除掉。
弗里德曼确认:就是在有效数据和无效数据上都训练,用对抗样本训练模型去排除它们?
马斯克的表述更简洁:你要同时知道什么是车,和什么绝对不是车。这是两件不同的事,得分开训练。
他顺带吐槽了一句,普通人对神经网络没有概念,可能以为是个渔网。
从车道线到通用智能,中间隔着好几个关键想法
话题从特斯拉抬到整个人工智能。弗里德曼问,现在的深度学习方法看起来离通用智能还很远,这条路能走到吗,还是需要全新的想法?
马斯克说还缺几个关键想法,但通用人工智能(AGI)会来得很快,快到我们得赶紧想清楚该怎么办,如果我们还有得选的话。
他觉得人们对窄人工智能和通用智能的区别缺乏感知。让车认出车道线、在街上导航,这是一类东西;通用智能是另一类。烤面包机和电脑都是机器,但复杂程度天差地别。
弗里德曼接了个玩笑:所以你很有信心特斯拉能造出世界上最好的烤面包机?马斯克说对。那世界上最好的自动驾驶呢?
马斯克说,现在看起来是“游戏设定,胜负已分”。他马上补了一句,不想显得自满或过度自信,但这就是他看到的局面,他可能错,可目前看特斯拉大幅领先所有人。
弗里德曼问了那个电影里的问题:我们会不会造出一个我们能爱上、它也真心爱我们的 AI,就像《她》(Her)那样?
马斯克的回答绕到了物理学。他说 AI 会有能力让你相信你爱上了它,而且说服力会非常强。这跟人类有什么区别?他说这就要进入形而上的领域了,情感和思想是不是存在于物理之外的某个维度,他不知道。但从物理学的角度看:
“如果它爱你的方式让你分辨不出真假,那它就是真的。如果没有测试能区分,那就不存在区别。”(马斯克)
他说这和模拟假说是同一个逻辑。现实世界和模拟世界之间可能根本没有测试能区分,那从物理学角度,两者就是一回事。
不过他也承认,可能存在办法测试出自己在不在模拟里。但你可以设想,一旦模拟里的实体发现了检测方法,模拟本身可以打补丁:暂停、重启、或者干脆开一个新的模拟,把这个错误修正掉。
弗里德曼最后问:如果有一天你或者别人造出了 AGI,你只能问它一个问题,你问什么?
马斯克说:模拟外面是什么?
这场对话真正的分歧在哪
把整场对话摊开看,两个人其实在回答两个不同的问题。
弗里德曼问的是:在人还坐在驾驶位上的这段过渡期,怎么让人保持状态、怎么让系统理解人。他的研究出发点是,自动驾驶不是纯技术问题,是一个人和机器共享控制权的问题。只要共享还在,人的状态就是安全变量的一部分。
马斯克问的是另一个问题:怎么让系统好到人坐在那里也无所谓。他的路线是把人的作用从安全等式里彻底删掉,而不是优化人的表现。在他的框架里,驾驶员监控是对一个即将消失的过渡期的补丁。
这两个问题不冲突,但它们对应的时间尺度不同。弗里德曼看的是接下来几年,马斯克看的是终局。谁对,取决于系统进步的速度,而这恰恰是两个人分歧最大的地方。弗里德曼在节目开头写得很克制:他不认同马斯克对摄像头监控相关性的判断,但尊重他领导的工程和创新能力。
马斯克那句“所有输入都是错误”,当成产品原则,它是苛刻的;当成工程原则,它是对的。因为一个系统如果默认人类介入是帮忙,它就永远不会去修那些被人默默接管的地方。把每一次动手都当成 bug 上报,才能把长尾一段一段啃下来。
再看那个数据比例。40 万辆车对 5000 辆车,99% 对 1%。这个差距不是算法差距,是车队规模差距。深度学习吃数据,数据来自路上跑的车,路上跑的车来自卖出去的车。这条链条一旦转起来,后来者要追的不是一行代码,是几百万英里的真实里程。这是马斯克敢说“胜负已分”的底气所在。
但硬币的另一面是,数据多不等于覆盖全。真正的长尾,比如某个路口奇怪的施工、某种罕见的光照、某类从未见过的障碍物,出现频率极低。数据量大能提高撞上的概率,但不能保证撞上。这也是为什么“接管即错误”这条原则重要,它把稀有事件从海量数据里筛了出来。
至于那颗双芯片的 FSD 计算机,它透露的是另一种思路。不赌算法什么时候成熟,先把算力和冗余堆到位,让硬件不再是瓶颈。剩下的是软件问题,而软件可以远程推送。这就是“升值资产”这个说法的技术底座。它成不成立,取决于一件事:硬件真的够用到 L4 甚至 L5 吗?马斯克赌的是够。
给读者的行动清单
- 如果你在评估任何“AI 辅助”类产品,先找它的健康检查机制。特斯拉那块屏幕的价值不在于好看,而在于让你一眼验证系统有没有搞懂状况。没有可验证性的自动化,用户只能盲信。
- 做数据驱动的系统,把“用户手动介入”默认标记为失败样本,而不是帮助。这条原则会把大量被忽略的长尾问题翻出来。便利性接管和纠错性接管可以后续再分流,但入口必须严。
- 评估一家公司的技术壁垒时,别只看算法论文。看它的数据飞轮转没转起来:产品卖出去、数据收回来、模型变强、产品更好卖。这个循环一旦成立,追赶者的差距会按年扩大。
- 做硬件产品,优先解决“最难升级的那部分”。特斯拉的逻辑是先让传感器和算力到位,把迭代压力全部转移到可远程更新的软件上。反过来做,硬件会变成负债。
- 面对“人类监督还有没有用”这类争论,先分清时间尺度。过渡期需要人类监控,终局不需要,两个判断可以同时为真。争的其实是曲线斜率,不是终点。
- 如果你在做安全相关的 AI,把对抗样本当成正常训练数据的一部分。既训练“这是什么”,也训练“这绝对不是”,两个方向分开做,比事后打补丁可靠。
- 读技术访谈时,注意说话人回答问题用的是什么框架。马斯克几乎每个回答都指向终局状态,这会让他低估过渡期的成本;弗里德曼从人的行为出发,这会让他高估监控的长期价值。两种偏差都真实存在,合起来看才完整。
0:00The following is a conversation with Elon Musk. He's the CEO of Tesla, SpaceX, Neuralink, and a co-founder of several other companies. This conversation is part of the Artificial Intelligence podcast. The series includes leading researchers in academia and industry, including CEOs and CTOs of automotive, robotics, AI, and technology companies.
0:24This conversation happened after the release of the paper from our group at MIT on driver functional vigilance during use of Tesla's autopilot. The Tesla team reached out to me offering a podcast conversation with Mr. Musk. I accepted with full control of questions I could ask and the choice of what is released publicly. I ended up editing out nothing of substance.
0:46I've never spoken with Elon before this conversation, publicly or privately. Neither he nor his companies have any influence on my opinion, nor on the rigor and integrity of the scientific method that I practice in my position at MIT. Tesla has never financially supported my research, and I've never owned a Tesla vehicle, and I've never owned Tesla stock.
1:10This podcast is not a scientific paper. It is a conversation. I respect Elon as I do all other leaders and engineers I've spoken with. We agree on some things and disagree on others. My goal is always with these conversations is to understand the way the guest sees the world. One particular point of disagreement in this conversation was the extent to which camera-based driver monitoring will improve outcomes and for how long it will remain relevant for AI-assisted driving.
1:39As someone who works on and is fascinated by human-centered artificial intelligence, I believe that if implemented and integrated effectively, camera-based driver monitoring is likely to be of benefit in both the short-term and the long-term. In contrast, Elon and Tesla's focus is on the improvement of autopilot such that its statistical safety benefits override any concern of human behavior and psychology.
2:08Elon and I may not agree on everything, but I deeply respect the engineering and innovation behind the efforts that he leads. My goal here is to catalyze a rigorous, nuanced, and objective discussion in industry and academia on AI-assisted driving. One that ultimately makes for a safer and better world.
2:30And now, here's my conversation with Elon Musk. What was the vision, the dream of autopilot when in the beginning, the big picture system level when it was first conceived and started being installed in 2014 in the hardware in the cars? What was the vision, the dream? I wouldn't characterize it as a vision or dream. It's simply that there are obviously two massive revolutions in in the automobile industry.
3:00One is the transition to electric electrification, and then the other is autonomy. And yeah, it became obvious to me that in the future any any car that did not have autonomy it would be about as useful as a horse. Which is not to say that there's no use. It's just rare. And somewhat idiosyncratic if somebody has a horse at this point. It's just obvious that cars will drive themselves completely. It's just a question of time.
3:29And if we did not participate in the autonomy revolution, then our cars would not be useful to people relative to cars that are autonomous. I mean, an autonomous car is arguably worth five to 10 times more than a car which is not autonomous. In the long term? Depends what you mean by long term, but let's say at least for the next 5 years, perhaps 10 years.
4:01So, there a lot of very interesting design choices with autopilot early on. First is showing on the instrument cluster or in the Model 3 on the center stack display what the combined sensor suite sees. What was the thinking behind that choice? Was there a debate? What was the process?
4:20The whole point of the display is to provide a health check on the the vehicle's perception of reality. So, the vehicle's taking information from a bunch of sensors, primarily cameras, but also radar and ultrasonics, GPS and so forth. And then that that information is then rendered into vector space, and that, you know, with a bunch of objects with with properties like lane lines and traffic lights and other cars. and then in vector space, that is re-rendered onto a display, so you can confirm whether the car knows what's going on or not by looking out the window.
5:02Right. I think that's a extremely powerful thing for people to get an understanding to become one with the system and understanding what the system is capable of. Mhm. Now, have you considered showing more? So, if we look at the computer vision, you know, like road segmentation, lane detection, vehicle detection, object detection underlying the system, there is at the edges some uncertainty.
5:25Have you considered revealing the parts that the the uncertainty in the system, the the sort of probabilities associated with with say image recognition or something like that? Yeah, so right now it shows like the vehicles in the vicinity a very clean, crisp image, and people do confirm that there's a car in front of me and the system sees there's a car in front of me, but did help people build an intuition of what computer vision is by showing some of the uncertainty.
5:53Well, I think it's in my car I always look look at this sort of the the debug view. And there's there's two debug views. one is augmented vision, where which I'm sure you've seen, where it basically we we draw boxes and labels around objects that are recognized. And then there's what we call the visualizer, which is basically a vector space representation summing up the input from all sensors.
6:22That that doesn't does not show any pictures, but it shows all of the it basically shows the car's view of of of the world in vector space. but I think this is very difficult for people to no normal people to understand. They would not know what they're they're looking at. So, it's almost a nature of my challenge to the current things that are being displayed is optimized for the public understanding of what the system is capable of.
6:48If like if you no idea what how computer vision works or anything, you can still look at the screen and and see if the car knows what's going on. And then if you're you know, if you're development engineer or if you're you know, if you're if you have the development build like I do, then you can see you know, all the debug information. But those would just be like total gibberish to most people.
7:10Right. What's your view on how to best distribute effort? So, there's three, I would say, technical aspects of autopilot that are really important. So, it's the underlying algorithms, like the neural network architecture, there's the data so that it's trained on, and then there's the hardware development. There may be others, but so, look, algorithm, data, hardware. You you only have so much money, only have so much time. What do you think is the most important thing to to allocate resources to? Or do you see it as pretty evenly distributed?
7:42Between those three. We automatically get a fast amounts of data because all of our cars have eight external facing cameras and radar and usually 12 ultrasonic sensors. GPS obviously and IMU. And so we we basically have a fleet that has I would got about 400,000 cars on the road that have that level of data. I I actually I think you keep quite close track of it actually.
8:15Yes. Yeah, so we're we're approaching half a million cars on the road that have the full sensor suite. Yeah. but so this is I I'm not I'm not sure how many other cars on the road have the sensor suite, but I'd be surprised if it's more than 5,000. Which means that we have we have 99% of all the data.
8:36So there's this huge inflow of data. Absolutely. Massive inflow of data. And then we it's it's taken us about 3 years, but now we've finally developed our full self-driving computer which can process an an order of magnitude as much as the Nvidia system that we currently have in the in the cars and it's really just a to use it you unplug the Nvidia computer and plug the Tesla computer in. That's it.
9:03And it's it's a In fact, we're not even we're still exploring the boundaries of its capabilities. but we're able to run the cameras at full frame rate, full resolution, not even crop the images. And it's still got headroom even on one of the the systems. The hardware full self-driving computer is really two computers, two systems on a chip that are fully redundant. So you could put a bolt through basically any part of that system and it still works.
9:30The redundancy, are they perfect copies of each other or Yeah. Also, it's purely for redundancy as opposed to an arguing machine kind of architecture where they're both making decisions. This is purely for redundancy. I think of it more like it's if you have, let's say, a twin-engine aircraft, commercial aircraft, the system will operate best if both systems are operating, but it's it's capable of operating safely on one.
9:56So, but as it is right now, we can just run we're we haven't even hit the the the edge the edge of performance, so but there's no need to actually distribute functionality across both SOCs. We we can actually just run a full duplicate on both on on each one. So, you haven't really explored or hit the limit of the We've not yet hit the limit, no.
10:22So, the magic of deep learning is that it gets better with data. You said there's a huge inflow of data, but Yeah. the thing about driving, the really valuable data to learn from is the edge cases. Mhm. So, how do you I mean, I've I've heard you talk somewhere about autopilot disengagements being an important moment of time Yes.
10:46to use. Is there other edge cases, or perhaps can you speak to those edge cases, what aspects of them might be valuable, or if you have other ideas how to discover more and more and more edge cases in driving? Well, there's a lot of things that are learned. There's certainly edge cases where let's say somebody's on autopilot, and they they take over. and then okay, that that will that that's a trigger that goes to our system that says, "Okay, did they Did they take over for convenience, or did they take over because the autopilot wasn't working properly?" There's also, like let's say we're we're trying to figure out what is the optimal spline for traversing an intersection.
11:27then then the the ones where there are no interventions and and we are the right ones. So you then say, "Okay, when it looks like this, do the following." And then and then and and then you get the optimal spline for a complex navigating a a complex intersection. So that's for This So there's kind of like the common case you're trying to capture a huge amount of samples of a particular intersection how when things went right and then there's the edge case where as you said not for convenience but something didn't go exactly right.
12:02over Somebody sort of mangled control from autopilot. And really like the way to look at this is view all input as error. If the user had to do input, it does something. All input is error. That's a powerful line to think of it that way. Cuz it may very well be error but if you want to exit the highway or if you want to is a navigation decision that all autopilot is not currently designed to do, then the driver takes over. How do you know the difference?
12:28that's going to change with navigate on autopilot which we're we're just released. And and we don't still confirm. So the navigation like lane change based like asserting control in order to change do it do a lane change or exit a freeway or or doing a highway interchange, the vast majority of that will go away with the release that just went out.
12:48Yeah, so that that I don't think people quite understand how big of a step that is. Yeah, they don't. If If you drive the car, then you do. So you still have to keep your hands on the steering wheel currently when it does the automatic lane change. What are So there's there's these big leaps through the development of autopilot through its history and what stands out to you as the big leaps?
13:13I would say this one navigate on autopilot without confirm without having to confirm is a huge leap. It is a huge leap. It also automatically overtakes slow cars. So it's it's both navigation and seeking the fastest lane. So it'll it'll it'll take you know overtake a slow cars and exit the freeway and take highway interchanges and and then we have a traffic light traffic lights recognition which is introduced initially as a as a warning. I mean on the development version that I'm driving the car fully fully stops and and goes at traffic lights.
13:56So those are the steps, right? You just mentioned something sort of inkling of a step towards full autonomy. Mhm. What would you say are the biggest technological roadblocks to full self-driving? Actually I think we I think we just the full self-driving computer that we just that the the Tesla what we call the FSD computer that's now in production.
14:21So if you order a any Model S or X or any Model 3 that has the full self-driving package you'll get the FSD computer. That that was that's important have enough base computation. Then refining the neural net and the control software. But but all of that can just be provided as an over-the-air update.
14:45The thing that's really profound and what I'll be emphasizing at the sort of what that investor day that we're having focused on autonomy is that the cars currently being produced with the hardware currently being produced is capable of full self-driving. But capable is a interesting word because Like the hardware is.
15:05Yeah, the hardware And as we refine the software the capabilities will increase dramatically and then the reliability will increase dramatically and then it will receive regulatory approval. So essentially buying a car today is an investment in the future. You're essentially buying You're buying the I think the most profound thing is that if you buy a Tesla today, I believe you are buying an appreciating asset, not a depreciating asset.
15:33So, that's a really important statement there because if hardware is capable enough, that's the hard thing to upgrade Yes. usually. Exactly. So, then the rest is a software problem. I have So, software has like no marginal cost, really. But, what's your intuition on the software side? How hard are the remaining steps
15:55to to get it to where you know, the the experience not just the safety, but the full experience is something that people would enjoy. I think people would enjoy it very much so on any on the highways. It's it's a total game changer for quality of life for using, you know, Tesla autopilot on the highways.
16:20so, it's really just extending that functionality to city streets. Adding in the the traffic light traffic light recognition, navigating complex intersections, and and and then being able to navigate complicated parking lots so the car can exit a parking space and come and find you even if it's in a a complete maze of a parking lot.
16:46And I and and then if and then you can just it just drop you off and find a parking spot by itself. Yeah, in terms of enjoyability and something that people would would actually find a lot of use from. The parking lot is a is a really, you know, it's it's rich of annoyance when you have to do it manually. So, there's a lot of benefit to be gained from automation there.
17:08So, let me start injecting the human into this discussion a little bit. so, let's talk talk about full autonomy. If you look at the current level four vehicles being tested on road like Waymo and so on, they're only technically autonomous. They're really level two systems with just a different design philosophy because there's always a safety driver in almost all cases and they're monitoring the system.
17:33Right. Do you see Tesla's full self-driving as still for a time to come requiring supervision of the the human being? So, its capabilities are powerful enough to drive, but nevertheless requires a human to still be supervising just like a safety driver is in a other fully autonomous vehicles.
17:57I think it'll it'll require detecting hands on wheel for at least 6 months or something like that from here. Really it's a question of like from regulatory standpoint, what what how much safer than a person does autopilot need to be for it to to be okay to not monitor the car? You know, and and this is a debate that one can have it. And then if you but you need need a large samples large amount of data so you can prove with high confidence statistically speaking that the car is dramatically safer than a person.
18:40and that adding in the person monitoring does not materially affect the safety. So, might it need to be like two or 300% safer than a person. And how do you prove that? Incidents per mile. Incidents per mile. So, crashes and fatalities. So, yeah. F- fatalities would be a factor, but there there there are just not enough fatalities to be statistically significant at scale, but there are enough crashes you know, there are much far more crashes than there are fatalities.
19:10So, you can just assess what is the probability of of a crash. That then there's another step which probability of injury. Then probability of permanent injury, then probability of death. And all of those need to be much better than a person by at least perhaps 200%. And you think there's a the ability to have a healthy discourse with the regulatory bodies on this topic?
19:40I mean there's no question that regulators pay just disproportionate amount of attention to that which generates press. This is just an objective fact. and Tesla generates a lot of press. So, the you know, in the United States, there's I think almost 40,000 automotive deaths per year.
20:01but if there are four in Tesla, they'll probably receive a thousand times more press than anyone else. So, the the psychology of that is actually fascinating. I don't think we'll have enough time to talk about that, but I I have to talk to you about the human side of things. So, myself and our team at MIT recently released a paper on functional vigilance of drivers while using autopilot.
20:24This is work we've been doing since autopilot was first released publicly over 3 years ago. Collecting video of driver faces and driver body. So, I saw that you tweeted a quote from the abstract, so I can at least guess that you've glanced at it. Yeah, I read it. Can I talk you through what we found?
20:46Sure. Okay. So, it appears that in the data that we've collected, that drivers are maintaining functional vigilance such that we're looking at 18,000 disengagements from autopilot, 18,900, and annotating were they able to take over control in a timely manner. So, they were there, present, looking at the road to take over control. Okay.
21:11So, this goes against what what many would predict from the body of literature on vigilance with automation. Now, the question is do you think these results hold across the broader population? So, ours is just a small subset. Do you think one of the criticism is that, you know, there's a small minority of drivers that may be highly responsible where their vigilance decrement would increase with autopilot use?
21:40I I I think this is all really going to be swept I mean the system is proving so much so fast that this is going to be a moot point very soon. Where vigilance is I guess if something's many times safer than a person, then adding a person does the the the the effect on safety is is limited.
22:05and in fact, it could be negative. That's really interesting. So, the the So, the fact that a a human may some percent of the population may exhibit a vigilance decrement will not affect the overall statistics numbers of safety. No, in fact, I think it it will become very very quickly maybe even towards the end of this year, but I'd say I'd be shocked if it's not next year at the latest, that having the person having a human intervene will decrease safety.
22:39Decrease. It's I can imagine if you're in an elevator. It used to be that there were elevator operators. and and you you couldn't go in an elevator by yourself and and work the the lever to move between floors. and now nobody wants an elevator operator because the automated elevator that stops the floors is much safer than the elevator operator.
23:03And in fact, it would be quite dangerous to have someone with a lever that can move the elevator between floors. So, that's a that's a really powerful statement and really interesting one. But I also have to ask from a user experience and from a safety perspective, one of the passions for me algorithmically is camera-based detection of of sensing the human, but detecting what the driver's looking at, cognitive load, body pose. On the computer vision side, that's a fascinating problem, but do you think and there's many in industry who believe you have to have camera-based driver monitoring. Do you think there is could be benefit gained from driver monitoring?
23:41If you have a system that's that's at or that's at or below human level reliability, then driver monitoring makes sense. But if if your system is dramatically better more reliable than than a human, then driver monitoring monitoring is not does not help much. And like I said, you you just like as an you wouldn't want someone interview like you wouldn't want someone in an elevator if you're if you're in an elevator, do you really want someone with a big lever? Some some random person operating an elevator between floors? They they could I wouldn't trust that.
24:14I would rather have the buttons. Okay, you're optimistic about the pace of improvement of the system. That from what you've seen with the full self-driving car computer. The rate of improvement is exponential. So, one one of the other very interesting design choices early on that connects to this is the operational design domain of autopilot. So, where autopilot is able to be turned on.
24:42The So, contrast another vehicle system that we are studying is the Cadillac Super Cruise system. That's in terms of ODD, very constrained to particular kinds of highways, well-mapped, tested, but it's much narrower than the ODD of Tesla vehicles. Mhm. What's There's There's pros and ADD. Huh.
25:02Yeah. That's good. That's a That's a good line. What was the design decision what in that different philosophy of thinking where There's pros and cons. What we see with a wide ODD is drive Tesla drivers are able to explore more the limitations of the system, at least early on, and they understand together with the instrument cluster display, they start to understand what are the capabilities.
25:30So, that's a benefit. The con is you're go you're letting drivers use it basically anywhere. Well, anywhere that it has could detect lanes with confidence. Was there a philosophy design decisions that were challenging that were being made there? Or from the very beginning was that done on purpose with intent?
25:54Well, I mean, I think it's Frankly, it's pretty crazy giving letting people drive it a two-ton death machine manually. that's crazy. Like Like in the future, people be like, "I can't believe anyone was just allowed to drive one of these two-ton death machines and they just drive wherever they wanted. Just like elevators. You just like move the elevator with the lever wherever you want and stop it halfway between floors if you want." It's pretty crazy.
26:24So, it it's going to seem like a mad thing in the future that people were driving cars. So, I have a bunch of questions about the human psychology about behavior and so on I don't know. that would become that Yeah, moot. Totally moot. because you have faith in the AI system. not faith, but the both on the hardware a and the deep learning approach of learning from data will make it just far safer than humans.
26:55Yeah, exactly. Recently, there are a few hackers who tricked autopilot to act in unexpected ways with adversarial examples. Mhm. So, we all know that neural network systems are very sensitive to minor disturbances to these adversarial examples on input. Do you think it's possible to defend against something like this for the broader for the industry?
27:16So, Yeah. Can you elaborate on the on the confidence behind that answer? well, the you know, a neural net is just like a basic bunch of matrix math. I mean, you have to be like a very sophisticated somebody who really understands neural nets and like basically reverse engineer how the matrix is being built and then create a a little thing that's just exactly causes the matrix math to be slightly off. But, it's very easy to then block it block that by by having what basically anti-negative recognition. It's like if you if the system sees something that looks like a matrix hack, exclude it.
27:58The so such a easy thing to do. So, learn both on the the valid data and the invalid data. So, basically learn on the adversarial examples to be able to exclude them. Yeah, you like you basically want to both know what is what is a car and what is definitely not a car. You you train for this is a car and this is definitely not a car. Those are two different things.
28:20The People People have no idea of neural nets really. They probably think neural nets involves like, you know, fishing net or something. So, as you know, jer So, taking it a step beyond just Tesla and autopilot, current deep learning approaches still seem in some ways to be far from general intelligence systems.
28:44Do you think the current approaches will take us to general intelligence or do totally new ideas need to be invented? I think we're missing a few key ideas for general intelligence, general artificial general intelligence. But it's going to be upon us very quickly. And then we'll need to figure out what shall we do if we even have that choice?
29:15But it's it's amazing how people can't differentiate between say the narrow AI that you know, allows a car to figure out what a lane line is and and and you know, and navigate streets versus general intelligence. Like these are just very different things. Like your toaster and and your computer are both machines, but one's much more sophisticated than another.
29:38You're confident with Tesla you can create the world's best toaster. The world's best toaster, yes. But the world's best self-driving I'm yes. Yeah, it's to me right now this seems game set match. I don't I mean that's how I don't want to be complacent or over confident, but that's what it that is just literally what it how it appears right now. I could be wrong, but it appears to be the case that Tesla is
30:08vastly ahead of everyone. Do you think we will ever create an AI system that we can love and loves us back in a deep meaningful way like in the movie Her? I think AI will be capable of convincing you to fall in love with it very well. And that's different than us humans? You know, we're starting to get into a metaphysical question of like do emotions and thoughts exist in a different realm than the physical? And maybe they do, maybe they don't, I don't know. But but from a physics standpoint, I don't to I tend to think of things you know, like physics was my main sort of training and and from a physics standpoint essentially if if it loves you in a way
30:52that is that you can't tell whether it's real or not, it is real. That's a physics view of love. Yeah. If there's no if you if you if you cannot just if you cannot prove that it does not if there's no test that you can apply that would make it make it allow you to tell the difference then there is no difference.
31:17Right. And it's similar to seeing our world as a simulation. There may not be a test to tell the difference between what the real world and the simulation and therefore from a physics perspective it might as well be the same thing. Yes. And there may be ways to test whether it's a simulation.
31:34There might be. I'm not saying there aren't. But you could certainly imagine that a simulation could could correct that once an entity in the simulation found a way to detect the simulation it could either restart the you know, pause the simulation start a new simulation or do one of any other things that then corrects for that error.
31:53So when maybe you or somebody else creates an AGI system and you get to ask her one question what would that question be?
32:16What's outside the simulation? Elon, thank you so much for talking today. It was a pleasure. All right. Thank you. Mhm.
资源与出处
- 视频在 YouTube 收看
- 档案silicon-valley.tech 档案页本页视频与文稿数据来自 Elon Musk Archive
- 精读本篇精读 Markdown 版