播客·中文精读

马斯克 2019 年 4 月 12 日谈 AI:对话莱克斯·弗里德曼

Elon Musk interview From April 12, 2019 With Lex Fridman On AI

中文精读由 AI 基于原始视频与文稿改写,非官方翻译,可能存在偏差或错误;表述以原视频为准。说明

2019年4月12日,马斯克在特斯拉工厂里和MIT研究员莱克斯·弗里德曼(Lex Fridman)聊了两个小时。弗里德曼研究自动驾驶里的人机交互,他的团队跟了特斯拉车主三年,标注了上万次方向盘接管的数据。

这场对话里,马斯克抛出一个违背常识的判断:今天买一辆特斯拉,买的是升值资产,并非贬值资产。

车是消费品,落地打八折,这是所有人的默认常识。马斯克凭什么反着说?理由很硬:硬件已经具备全自动驾驶能力,剩下的只是软件升级。软件没有边际成本。你买的那台车,会在某个深夜通过OTA更新,突然多出一项能力。硬件是固定的,软件是流动的,整台车的价值曲线就被改写了。

这个判断能不能成立,要看后面一整套逻辑。这篇精读把整场对话拆开,从感知系统、数据规模、自研芯片,一路聊到监管悖论、对抗攻击、通用人工智能和模拟假说。

未来人们会觉得手动开车这件事很疯狂

马斯克先给了汽车行业一个二分法。未来只有两场革命:电动化和自动驾驶。前者已经在发生,后者是迟早的事。任何一台不具备自动驾驶能力的车,长期看会像今天的马。

他讲得很直白。今天还有人养马,图的是爱好,交通属性早就没有了。马变成了罕见而古怪的东西。同样,一台不能自己开的车,未来在交通系统里的地位就是马。

按他的估值,一台自动驾驶汽车的价值是非自动驾驶汽车的五到十倍。至少未来五到十年内,这个差距会存在。特斯拉如果不参与这场革命,车的竞争力会被直接抹平。

弗里德曼追问的角度很有意思。他说现在路上跑的Waymo们,严格讲只是L2系统加了个安全员,跟特斯拉的设计哲学不同。马斯克把话头一转,扔出一个更狠的判断:让人手动驾驶一台两吨重的死亡机器,这件事本身就挺疯狂。

他的类比是电梯。几十年前电梯里有专职操作员,你得等那个人扳动拉杆,才能在楼层之间移动。今天没人想要电梯操作员。自动电梯停得稳、判断准,比人可靠。更关键的一点,让一个随机的人拿着一根拉杆控制电梯,本身就是风险源。

他说未来的人回头看今天,会觉得不可思议:人类居然被允许随时随地把两吨重的金属盒子开到任何地方,还能停在两层楼中间。这个画面在将来会像“电梯里站着个操作员”一样荒诞。

仪表盘的作用是健康检查

自动驾驶最早的一个设计选择,是把传感器看到的世界画在屏幕上。Model 3把这块放在中控,Model S和X放在仪表盘。弗里德曼问,当初为什么这么设计,团队里有没有争论。

马斯克的回答很干脆:整个屏幕的唯一目的,是给车辆的感知做一个健康检查。

逻辑是这样。车靠摄像头、雷达、超声波、GPS收集信息,把这些原始数据转成向量空间里的对象:车道线、红绿灯、其他车辆,每个对象带属性。然后系统再把向量空间渲染回屏幕。你坐在车里看一眼屏幕,就能确认车是不是知道周围发生了什么。

这套设计对普通人友好。你不需要懂计算机视觉,也能看出前面有没有车、车道线对不对。马斯克说这能让人“和系统合一”,理解系统能干什么、不能干什么。

弗里德曼接着问了个技术问题:计算机视觉在边缘有不确定性,车道分割、目标检测都可能有置信度问题,为什么不把这种不确定性显示出来,让用户建立对AI的直觉?

马斯克说他自己天天开debug视图。debug视图有两套。一套叫增强视觉(augmented vision),在识别到的物体上画框加标签,就是大家常见的那种;另一套叫可视化器(visualizer),纯向量空间表示,把所有传感器的输入汇总成俯视图,没有画面,只有对象。

这两套东西对普通用户就是天书。马斯克说,把不确定性直接糊到屏幕上,大部分人根本不知道自己在看什么。所以现在展示的东西,是为公众理解做的取舍,本质上是一道人机界面(HMI)难题。开发工程师用开发版能看到全部debug信息,普通用户看的是精简版。

全球99%的自动驾驶数据在特斯拉手里

弗里德曼把自动驾驶拆成三根支柱:算法、数据、硬件。钱和时间都有限,资源该怎么分?马斯克没有回避,他先讲数据。

特斯拉每台车有8个外部摄像头、1个雷达、12个超声波传感器,还有GPS和IMU。这套配置产生的数据流是海量的。当时路上跑着大约40万台具备完整传感器套件的车,正在逼近50万台。

马斯克补了一个对比。他不太确定其他厂商有多少台车装了同级传感器,但他猜全世界加起来可能不超过5000台。也就是说,特斯拉握有全球99%的自动驾驶数据。

这个数字的实际意义在训练深度学习的原料规模上。数据规模是复利,别人要追,先得把车队铺到同等量级。而车队规模又是制造和销售问题,算法再强也绕不过去。

数据进来之后,还需要能处理它的算力。马斯克说他们花了大约三年,做出了全自动驾驶芯片(FSD computer)。这块芯片的处理能力是当时车上英伟达系统的一个数量级。更换方式极简:拔掉英伟达的盒子,插上特斯拉自己的,完事。

更狠的是性能余量。他们能让摄像头跑满帧率、满分辨率,连图像都不裁,芯片还有大量headroom(余量)。这块硬件还是两套完全冗余的系统,两颗系统级芯片(SoC)。马斯克用双引擎客机打比方:两台发动机都在的时候表现最好,只靠一台也能安全飞。

这里的冗余属于纯冗余,跟“两颗芯片分别做决策再投票”的架构完全不同。两颗芯片跑同样的东西,一颗坏了另一颗顶上。弗里德曼问你们是不是还没摸到性能天花板,马斯克说完全没有,甚至连把功能拆分到两颗芯片上的必要都没有,直接每颗跑一份完整副本就行。

所有输入都是错误

聊到数据,就绕不开最有价值的那部分:边缘案例。弗里德曼提到他听过马斯克说,自动驾驶的接管(disengagement)是关键信号。

马斯克把接管分成两类。

第一类是导航性接管。比如你想下高速,但当时的Autopilot不负责导航,你只能自己接管。这类接管跟系统出错无关,只是功能没覆盖到。刚发布的Navigate on Autopilot就是来解决这个的:自动变道、下匝道、过高速立交,大部分这类接管会消失。马斯克说这一步的跨度大到外界没意识到,因为不开车的人感受不到。

第二类才是真正有价值的:系统本该处理,结果没处理好,人被迫接管。这一类是训练素材的金矿。

马斯克在这里给了一句非常锋利的话,弗里德曼当场重复了一遍。

所有输入都是错误。(马斯克)

意思是,只要用户不得不动手,就默认系统有缺陷。这个原则极端,正是这种极端逼着团队去追每一个接管瞬间。当然要区分动机,用户接管可能是图方便,也可能是系统真的不行。把默认值设成“错”,能最大化收集问题样本。

另一类数据来自没有接管的场景。一个复杂路口,系统顺利通过,这段轨迹就是“最优样条”(optimal spline)。大量同类路口的正样本堆起来,模型就学会了在这种路况下该怎么走。负样本找错误,正样本固化正确,两条线同时推进。

交通灯是下一个台阶。开发版上,车已经能在红绿灯前完全停住再起步。早期只是警告提示,后来变成完整控制。这些是一级一级往上垒的台阶。

软件没有边际成本,硬件一次付清

弗里德曼问,全自动驾驶最大的技术路障是什么。马斯克的回答是:路障已经不在硬件了。

FSD芯片已经量产。你订任何一台Model S、Model X,或者带全自动驾驶套件的Model 3,拿到的就是这块芯片。算力够了,剩下是两件事:打磨神经网络、打磨控制软件。这两件事都能通过OTA推送完成。

然后他说了那句最重的话:今天买一台特斯拉,你买的是一项升值资产。

逻辑很干净。硬件已经具备全自动驾驶能力,“具备”这个词是关键。随着软件迭代,能力会指数级上升,可靠性会指数级上升,然后拿到监管批准。整个过程中车主不需要换车,车自己在变强。

这解释了他为什么在投资者日反复强调这一点。传统汽车的生命周期是贬值曲线,出厂即巅峰。特斯拉把它改成了增值曲线,出厂只是起点。硬件的成本一次性付清,软件的边际成本趋近于零,价值增量全部归车主。

弗里德曼追问软件侧还剩多难。马斯克说高速上的体验已经是质变,接下来是把这套能力延伸到城市街道:交通灯识别、复杂路口、复杂的停车场。他描述了一个场景:车在迷宫一样的停车场里自己找车位,你下车走人,它自己停好;你要走的时候,它从车位开出来找到你。

停车这件事本身很烦,自动化带来的收益很直接。

电梯操作员与监管悖论

话题转到人。弗里德曼问,特斯拉的全自动驾驶在一段时间内是不是仍然需要人类监督,就像Waymo车里的安全员。

马斯克估计,检测方向盘上有没有手,这个要求至少还得持续六个月左右。之后是监管问题:自动驾驶要比人类安全多少,才允许人类不盯着?

他说这需要大量数据做统计证明。不光要证明车比人安全,还要证明加上人类监督不会实质性提升安全性。这个门槛可能是人类安全水平的200%到300%。

怎么证明?靠碰撞率。致死事故数量太少,统计上不够显著,碰撞事件足够多。评估要分层:碰撞概率、受伤概率、重伤概率、死亡概率,每一层都得显著优于人类。

弗里德曼提到自己团队在MIT的研究。他们从Autopilot公开起就跟踪车主,采集面部和身体视频,标注了18000次接管,看司机能不能及时接管。结论是:司机保持了功能性警觉,数据不支持“自动化让人松懈”的经典担忧。

马斯克说他在推特上转发过论文摘要。接着他给了个更大的判断:这个问题很快会变得无关紧要。

因为一旦系统比人类安全得多,人类介入这件事本身就开始降低安全性。他的原话大意是,最快今年、最晚明年,人类干预会变成安全负资产。弗里德曼追问,那意味着即使有一部分人出现警觉性下降,也不影响整体安全?马斯克说对,因为基数变了。

他又搬出电梯。自动电梯比操作员安全,让一个拿拉杆的人控制电梯,本身就是风险源。

弗里德曼顺势问驾驶员监控摄像头。业界很多公司认为必须装,用计算机视觉检测司机看哪儿、认知负荷多大、身体姿态如何。马斯克的回答分两层:如果系统可靠性低于人类,监控有意义;如果系统远高于人类,监控帮不上什么忙,甚至有害。他还是那句电梯比喻,你不会想给电梯配个拿拉杆的随机路人。

弗里德曼说,从用户体验和安全角度,基于摄像头的司机状态感知是他个人很在意的研究方向。马斯克承认这问题有意思,立场没变:前提是系统可靠性还没超过人类。

一千倍的媒体关注度

监管之外还有舆论。马斯克说了一件他称为“客观事实”的事:监管部门对能上新闻的事件投入了不成比例的注意力,而特斯拉特别能上新闻。

他给了组数据。美国每年大约4万起汽车致死事故,媒体关注度平平。特斯拉只要出4起,报道量可能是别人的一千倍。

他描述的是一个心理机制。公众对新生事物的风险感知被放大,对存量风险习以为常。4万条人命是背景噪音,4条就是头条。这个不对称会直接影响监管节奏,因为监管要对公众情绪负责。

弗里德曼没在这个话题上纠缠太久,他更关心技术选择背后的哲学。他拿凯迪拉克的Super Cruise做对比:那套系统的运行设计域(ODD)很窄,只在特定高速、高精地图覆盖的路段开放。特斯拉的ODD宽得多,几乎哪都能开。

马斯克承认宽ODD有代价。好处是车主能更早探索系统边界,配合仪表盘的可视化,慢慢建立对能力的认知。坏处是范围太宽,出问题的场景也更多。他反问:把一台两吨重的车交给人类随便开,这件事难道不更疯?

如何防住对抗攻击

弗里德曼提到有黑客用对抗样本(adversarial examples)骗过了Autopilot。神经网络对输入的微小扰动很敏感,贴个特定图案在路牌上,模型就可能认错。他问,这种攻击能防住吗?

马斯克的回答只有一个词:能。

他解释得很技术。要构造对抗样本,攻击者得深入理解神经网络的矩阵运算,反向工程出能让矩阵计算轻微出错的特定扰动。门槛很高,需要真实的专业能力。但这些都不重要,重要的是怎么防。

他的方案叫负识别。训练模型时,不光教它“什么是车”,还要教它“什么绝对不是车”。这是两个不同的任务。你把对抗样本当作负样本喂进去,模型就学会了识别“看起来像车但并非车”的东西。

他说这在工程上很简单。如果系统看到某种像矩阵攻击的图案,直接排除。前提是训练集里同时包含正例和负例,让模型学会两边。

弗里德曼感叹大部分人对神经网络的理解停留在“它是个黑箱”。马斯克这点说得很朴素:模型要知道什么是什么,也要知道什么不是什么,两件事一起学,防御才成立。

烤面包机和电脑的区别

聊完特斯拉,话题升到通用人工智能。弗里德曼问,现在的深度学习方法能走到通用智能吗,还是需要全新的想法。

马斯克说,还差几个关键概念,但通用人工智能(AGI)会来得非常快。快到我们可能还没想好该怎么办,就得面对“要不要有选择权”这个问题。

他补了一句很多人会忽略的区分:窄AI和通用AI是两回事。让车识别车道线、在街道上导航,这是窄AI。通用智能是另一种东西。他用了个很接地气的比喻:烤面包机和电脑都是机器,复杂度差了几个量级。

弗里德曼开玩笑,问他对特斯拉做出世界最好的烤面包机有没有信心。马斯克接梗说,最好的烤面包机和最好的自动驾驶,他都有信心。

然后弗里德曼抛出一个观察:眼下看起来,特斯拉在自动驾驶上遥遥领先其他人。马斯克说,这可能是他自满,但就目前看,确实是这样。他留了余地,说自己也可能判断错。这种表达方式在这类对话里很少见,多数企业家会把“可能”两个字删掉。

如果无法证伪,那就没有区别

弗里德曼问了个偏哲学的问题:我们会不会造出一个能爱我们、也被我们爱的AI,像电影《她》里那样。

马斯克的回答分两层。第一层是技术判断:AI完全有能力让你相信你爱上了它,而且做得很自然。这一层没什么争议。

第二层就深了。他说人类的情感、意识是不是存在于物理之外的某个领域,这是个形而上学问题,他也不知道。他受的物理训练让他倾向一种判断:如果你没法设计出任何测试,能区分“它真的爱你”和“它表现得像爱你”,那这两者在物理上就没有区别。

弗里德曼接了一句:这就是物理学视角的爱。

马斯克顺着推到模拟假说。如果真实世界和模拟世界之间不存在可执行的区分测试,那从物理角度看,两者就是一回事。他补了个细节:模拟也可能自我修正。一旦模拟里的实体找到检测模拟的方法,模拟可以重启、暂停、开新局,把那个“错误”抹掉。即使你找到了测试,也未必能留下证据。

这段话容易被读成玄学,其实逻辑很硬。它问的是:一个命题如果原则上不可证伪,它在物理上还有意义吗?马斯克的答案是,没有可区分性就没有区别,这正是物理学处理不可观测量的标准做法。

再往下推一层,这其实是整场对话方法论的内核。他判断自动驾驶能不能实现,用的是同一套标准:不看宣传,不看愿景,看数据能不能支撑统计显著性。一个说法如果无法被数据检验,他就不会拿它当依据。

你会问AGI什么问题

弗里德曼在结尾问了个收束性的问题:如果有一天你造出AGI,只能问它一个问题,你问什么。

模拟之外是什么?(马斯克)

这个问题把整场对话串起来了。从车道线识别到通用智能,从对抗样本到模拟假说,所有技术细节最后都指向同一个位置:我们所在的这个系统,边界在哪,外面是什么。

回看这两个小时,马斯克的论证方式很一致。整场对话里愿景占的比重很小,大部分时间在算约束条件。数据够不够,算力够不够,监管门槛多高,舆论会怎么扭曲。每个判断后面都跟着一个可验证的数字:99%的数据、一个数量级的算力提升、200%到300%的安全冗余、4万对4的媒体比例。

这也是这场对话值得细读的原因。它更像一次工程思维的演示:一个人怎么把一个还没解决的问题拆成硬件问题、软件问题、统计问题、心理问题,然后逐个给出量化门槛。推销自动驾驶的成分很少。

至于“升值资产”那个判断,成不成立要看两件事:FSD软件能不能按节奏迭代到位,监管能不能接受人类监督被移除。这两件事在2019年都还没有答案。但马斯克把赌注押得很清楚。

对读者来说,这场对话里有几条能直接用的判断方式:

  • 评估任何自动化系统,先问它比人类基线安全多少倍,再问人类介入是加分还是减分。很多产品的争论卡在第一步没做量化。
  • 看一家公司的技术护城河,别只看算法,看数据规模。99%和1%之间是量级差距,靠效率改进追不上。
  • 接管、报错、用户手动纠正,这些负反馈是最贵的训练素材。把默认值设成“所有输入都是错误”,能逼出系统真正的问题。
  • 对抗攻击的防御思路可以迁移:训练时同时教模型“是什么”和“绝对不是什么”,正负样本一起学。
  • 遇到不可证伪的命题,用物理学的标准处理:设计不出区分测试,就当它没有区别,别在玄学上耗时间。
  • 判断一个新技术的社会阻力,别只看技术成熟度,看它出错时会被放大多少倍。媒体关注度的不对称,往往比技术本身更早决定落地节奏。
YouTube 平台字幕(自动或人工),可能有识别错误,人名与数字请以视频为准。已按语句合并成段落,并去掉 um、uh 等口头填充词。 文稿来源:YouTube(自动/人工字幕)。

0:00the dream of autopilot when in the beginning the big-picture system level when it was first conceived and started being installed in 2014 the hardware in the cars always division the dream I would characterize the division or dream something that there are obviously two massive revolutions and in the

0:19automobile industry one is the transition to electrification and then the other is autonomy and yeah became obvious to me that in the future any a car that did not have autonomy I 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 and if we did not

0:51participate 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 ten times more than I not colored which is not autonomous in the long term as what you doing about long term but let's say at least for the next five years perhaps ten years so there a lot of very interesting design choices with autopilot early on first is showing on the instrument cluster or in the model three on the center stack display what the combined sensor suite sees what was the thinking behind that choice was there debate what was the process the whole point of the display is to provide

1:37a health check on the vehicles perception of reality so the vehicles taking an information for a motion sensor is 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 product with properties like lane lines and traffic lights and other cars and then in vector space that is re-rendered on to your display so you can confirm whether the car knows what's going on or not by looking at the window right 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

2:19of 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 have you considered revealing the parts that the the uncertainty in the system the Sodapop movies associated with will say yeah right now it shows like the vehicles in the vicinity a very clean crisp image and people do confirm there's a car in front of me and the system sees there's a car in front of me but they help people build an intuition of what computer vision is by showing some of the uncertainty well I think it's my car I always look look at this

3:01with the debug view and it's this to debug views one is augmented vision where which I'm sure you've seen where it's basically we draw boxes and labels around objects that are recognized and then there's we're called visualizer which is basically a vector space representation summing up the input from all sensors that doesn't does not show any pictures but it shows all of the it's basically shows the cause view of of the world in vector space but I think this is very cool for people to know know people to understand they would not know what think they're looking at so it's almost an HMI challenge to the current things that are being displayed is optimized

3:45for the general public understanding of what the system is it's like if you no idea what computer vision works or anything you can still look at the screen and see if the car knows what's going on and then if you're you know if you're a 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 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

4:20that this train on and then there's a hardware development there may be others but so look algorithm data hardware you don't 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 do you see it as pretty evenly distributed between those three we all have to get a fast amounts of data because all of our cars have eight external facing cameras and radar and usually twelve ultrasonic sensors GPS obviously and I am you and so we

4:59basically have a fleet that has I we've got about four hundred thousand cars on the road that have that level of data I think you keep quite close track of it actually yes yeah so we're we're approaching half a million cars on the road that have the full sensor suite yeah so this is I'm I'm not sure how many other cars on the road have the sensor suite but every surprised it was more than five thousand which means that we were we have ninety nine percent of all the data so there's this huge inflow of data absolutely massive inflow of data and then we it's taken us about three years but now we finally felt that full self-driving computer which can process

5:40and an order magnitude as much and as the Nvidia system that we currently have in the in the cars and it's really just a to use it you've unplugged the Nvidia computer and plug the tells the computer in and that's it and it's it's a in fact we're not we're still exploring the boundaries of its capabilities we were able to run the cameras at full frame rate full resolution not even crop the images and it's still quite Headroom even on one of the the systems the hard but also private computer is really two computers two systems on a chip that are fully redundant so you could put a boat through basically any part of that system and it still works the redundancy

6:16are 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 you think more like its if you ever say a twin-engine aircraft commercial aircraft this system will operate best if both systems are operating but it's

6:36it's capable of operating safely on one so but as is right now we can just run we haven't even hit the edge of performance so with us no need to actually distribute the functionality across both SOC s we can actually just run a full duplicate on but on on each one you haven't really explored or hit the limit oh this is not yet at the limiter so the magic with deep learning is that 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 so how do you I mean I've heard

7:19you talk somewhere about autopilot disengagement being an important moment of time yes to use is there other edge cases or perhaps can you speak to those edge cases what aspects of that might be valuable or if you have other ideas how to discover more and more and more edge cases and driving well there's a lot of things that I learned though certainly edge cases where I say somebody's on order putt and they take over and then okay that that that's a trigger that goes at justice and I says okay so the trip they take over for convenience or do they take over because the autopilot wasn't working properly there's also like let's say we're trying

7:57to figure out what is the optimal spline for traversing an intersection then even the ones where there are no interventions and we are the right ones see then say okay when it looks like this do the following and then the end and then you get the optimal spline for a complex know getting a complex intersection so that's for this kind of

8:21the common case you're trying to capture a huge amount of samples of a particular intersection how one thing is going right and then there's the edge case where as you said not for convenience but something didn't go all right so he took over somebody sort of manual control from autopilot and really liked the way to look at this is view all input is error if the user had to do input if there's something all input is error that's a powerful line to think of it that way because it may very well be error but if you want to exit the highway or if you want to it's a navigation decision that all autopilot is not currently designed to do then the driver takes over how do

8:57you know that that's gonna change with navigate an autopilot which we're just released and with a stoker phone so the navigation like lane change based it likes a certain control in order to change the lane change or exit freeway or or doing a highway interchange the vast majority that will go away with the release that just went out yeah that that I don't think people quite understand how big of a step that is yeah they don't see if you drive the car than you do so you still have to keep your hands on the steering wheel currently when it does the automatic complaint change what our so there's these these big leaps through the development of autopilot through its

9:34history and what stands out to you as the big leap so I would say this one navigate an autopilot without confirm without having to confirm is a huge leap it is a huge leap but it also automatically overtakes low cars so it's it's both navigation and seeking the fastest lane so it'll it'll tilt it you know overtake a slow cause and exit the freeway and take highway interchanges and and then we have of traffic like

10:06where traffic lights a recognition which is introduced put initially as a as a warning I mean on the development version that I'm driving the car fully fully stops and goes at traffic lights so those are the steps right you've just mentioned some things that an inkling of a step towards full autonomy what would you say are the biggest technological roadblocks to full self-driving actually I don't think I think we're just the full self-driving computer that we just with that but it has the Oracle the FST computer that that's now in production

10:40so if your order any Model S or X or any Model 3 that has the full soft driving package you'll get the FST computer that that was that's important to have enough bass computation then refining the neural net and the control software which but all of that can just providers know their update the thing that's really profound and I'll be emphasizing at the auto sort of what that investor day they were having focused on autonomy is that the cars currently being produced but the hardware currently being produced is capable of full self-driving but capable is an interesting word because the hardware is yeah as we refine the software the capabilities will increase dramatically

11:25and 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 your bike the most profound thing is that if you buy a Tesla today I believe you are buying an appreciating asset not a depreciating asset so that's a really important statement there because if hardware is capable enough that's the hard thing to upgrade yes usually exact so then the rest is a software problem yes I have software has no marginal cost really but what's your intuition on the software side how hard are the remaining steps to get it to where you know the

12:10the experience not just the safety but the full experience is something that people would enjoy I think we will enjoy it very much the on the highways it's it's a total game changer for quality of life for using you know Tesla order piled on the highways is so it's really just extending that functionality to city streets adding in the traffic like traffic light recognition navigating complex intersections and and then being able to

12:42navigate complicated parking lots so the car can exit a parking space and come and find you even if it's in a complete maze of a parking lot and and then if and then you can just pick just drop you off and find a parking spot by itself yeah in terms of enjoy ability and something that people would would actually find a lot of use from the parking lot 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 so let me start injecting the human into this discussion a little bit so let's talk about full autonomy if you look at the current level for vehicles being test on row

13:21like way mo so on they're only technically autonomous they're really level two systems with just the different design philosophy because there's always a safety driver in almost all cases and they're monitoring the system right do you see Tesla's full self-driving as still for a time to come the requiring supervision of 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

13:54fully autonomous vehicles I think it will require detecting hands on wheel for at least six months or something like that from here it really is question of like from a regulatory standpoint what how much safer than a person just autopilot need to be for it took to be okay to not monitor the car you know and and this is a debate that what can happen and then if you need even a large sample a large amount of data so you can prove with high confidence statistically speaking that the car is dramatically safer than a person and that adding in the person monitoring does not materially affect the safety so it might need to be like two or three hundred percent safe in a

14:43person and how do you prove that it's Lisbon well instance morale they'll crashes and fatalities yeah fatality will be a factor but is there they're just not enough fatalities to be statistically significant at scale but there are enough crashes you know there are four more crashes and there are fatalities the Sooners assessed where's the probability of of crash that then there's another step which probability of injury and probability of owned injury the probability of death and all of those need to be much better than a person by at least perhaps 200 percent

15:23and you think there's a the ability to have a healthy discourse with the regulatory bodies on this topic I mean there's no question that but regulators paid a 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 this

15:46panic almost forty thousand automotive deaths per year but if there are four in Tesla they'll probably receive a thousand times more press than anyone else so the psychology of that is actually fascinating I don't think we'll have enough time to talk about that but I have to talk to you about the human side of things so myself and our team at MIT recently released the paper on functional vigilance of drivers while using autopilot this is work we've been doing since autopilot was first released publicly over three years ago collecting video on driver faces and drive her body so I saw that you tweeted a quote from the abstract so I can at least guess

16:26that you've glanced at it yeah all right can I talk you through what we found sure 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 disengagement from autopilot in two thousand nine hundred and annotating were they able to take over control in a timely manner so they were they're present looking at the road to take over control okay so this goes against what

16:55what 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 without apologies this is all really gonna be swept I mean that the systems are proving so much so fast that this is gonna be a moot point very soon where vigilance is like if some

17:33things many times safer than a person then adding a person does if the effect on safety is is limited and in fact it could be negative that's really interesting so the the fact that a human may some percent of the population may exhibit a vigilance decrement will not affect overall sense of safety no in fact I think it will become very very quickly maybe in towards in this year but I say every shark it's not next year at the latest that having the post having human intervene will decrease safety decrease it's a command if you're in an elevator I used to be the throw elevator operators and and you couldn't go in an elevator by yourself and work the the

18:22lever to move between floors and now it nobody wants an elevator operator because the automated elevator that stops the force is much safer than the elevator operator and 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 how guerilla CLE is camera based detection of algae sensing the human but detecting what the driver is looking at cognitive load body pose on the computer vision side that's a fascinating problem

19:01but do you and there's many in industry believe you have to have camera based driver monitoring do you think there could be benefit gained from driver monitoring if you have system that's that's I don't that's outer below human level reliability then driver monitoring makes sense but if your system is dramatically better more level than then a human then drive montane monitoring is not just not help much and like said you

19:28just like as a you wouldn't want someone interviewed like you want someone in the elevator sure an elevator do you really want someone with a big lever so some random person operating the elevator between floors I think they could I wouldn't trust that but I'd rather have the buttons okay you're optimistic about

19:44the pace of improvement of the system that from what you've seen with a full self-driving car computer the rate of improvement is exponential so 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 the so contrast another vehicle system that we're studying is the Cadillac super core system that's in terms of ODG very constraint is particular kinds of highways well mapped tested but it's much narrower than the OD of Tesla vehicles what's theirs there's a TD is good this is good what was a design

20:29decision what's in that different philosophy of thinking where there's pros and cons what we see with a wide o Deedee's 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 so that's a benefit the con is you go you're letting drivers use it basically anywhere anywhere tech lines with continents it was their philosophy design decisions they were challenging they were being made their order from the very beginning was that done on purpose with intent

21:10well I mean I think it's frankly it's pretty crazy giving it 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 about to drive for 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 to can stop it halfway between floors if you want it's pretty crazy so it's gonna seem

21:41like a mad thing in the future that people were driving cars so that I have a bunch of questions about the human psychology about behavior and so on and that would be coming that remember told ya because you have faith in the AI system not faith but the both on the hardware side and the deep learning approach of learning from data will make it just far safer than humans yeah exactly recently there are few hackers who are tricked autopilot act and not unexpected ways of the adversarial examples 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

22:21against something like this wrong for the street-door yeah can you elaborate on the on the confidence behind that answer well the you know when your own there's just like basic bunch of make matrix math oh you have to be like a very sophisticated it's already really understands neural nets and like basically reverse-engineer how the matrix is being built and then create a little thing that's just exactly causes the matrix math to be slightly off but it's very easy then block it block that by having what basically ant here a negative recognition it's like if you if the system sees something that looks like a matrix hack excluded this is such

23:05an easy thing to do so learn both on the validator 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 if you trained for this is a car and this is definitely not a car those are two different things people have no idea neural nets really there's probably thing your lesson is both like you know fishing net only so

23:31as you know so taking a step beyond just Tesla and autopilot current deep learning approaches still seem in some ways to be far from general intelligent systems do 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 gonna be upon us very quickly and then we'll need to figure out what shall we do if we even have that choice would be but it's amazing how people can differentiate between say the narrow AI that you know allows a car to figure out where the lane line is and and and you

24:22know and navigate streets versus general intelligence like these are just very different things like your toaster and your computer or both machines but once much more sophisticated than another you're confident with Tesla you can create the world's best toaster those best toaster yes the world's best self-driving I'm yes I do to me right

24:47now this seems games that match I don't mean that's how much be complacent or confident but if that's what it appears 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 vastly ahead of everyone do you think we'll ever create an AI system that we can love and loves us back and I did 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 start getting into a metaphysical question of like do my emotions and thoughts exist in a different realm in the physical and maybe they do maybe they don't

25:27I don't know but but from a physics standpoint I think I tend to think of things you know like physics was my main sort of training and and from a physics standpoint essentially if it loves you in a way there's that you can to tell whether it's real or not it is real it's a physics view of love yeah if there's no if you if you can't just if you can't prove that it does not if there's no test that you can apply that would make it

26:00may allow you to tell the difference then there is no difference right and it's similar to seeing our world of simulation there may not be a test to tell the difference between what the real world 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 there might be I'm not saying there aren't but you could certainly imagine that a simulation could correct that once an entity in the simulation found a way to detect the simulation it could either restart that before you know pause the written simulation or start a new simulation or do one of many other

26:33things that then corrects for that error so when maybe you or somebody else creates an AGI system and you get to ask her one question or would that question be what's outside the simulation Ilan thank you so much for talking today as a pleasure all right thank you [Music]

资源与出处

相关视频

同频道、同主题或相邻年份