访谈·中文精读

英伟达CEO黄仁勋在GTC 2015对话Tesla马斯克

NVIDIA CEO Jen-Hsun Huang Chats With Tesla's Elon Musk @ GTC 2015

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

2015年3月,圣何塞,英伟达GTC大会。黄仁勋坐在台上,对面是马斯克。

他抛出的第一个问题就带着火药味:你去年说过人工智能比核武器更危险,说搞AI就像"召唤恶魔"。可你现在拼命做的自动驾驶,靠的正是AI。这两件事你怎么放在一起?

马斯克的回答没绕弯子:自动驾驶属于窄人工智能,跟那个让人睡不着觉的东西不是一回事。

这个区分在2015年听起来像技术人的自我安慰。十年后回头看,它是理解马斯克所有AI言论的一把钥匙。他所有的乐观和所有的恐惧,都建立在这条分界线上。

电梯操作员消失了,司机大概率也会

马斯克用的类比是电梯。

一百年前,每部电梯里都坐着一个人。你要去几层,告诉他,他拉杆、关门、停靠。后来出现了简单的电路,按个按钮,电梯自己就上来了。今天没人觉得电梯需要专职司机,电梯甚至学会了提前把轿厢挪到人多的楼层等着。

他说汽车会走同一条路。你上车说一句"送我回家",或者"去公司""去那儿",车就开了。安全性比人高一个数量级。

再往后推一步,他的预测带着点黑色幽默:将来人类开车可能会被立法禁止。理由是数据太难看。让人来操作一台两吨重的金属机器,每年死上百万人,这种"杀人机器"迟早会被当成公共健康问题处理掉。

我不认为我们需要担心自动驾驶,它属于一种窄形式的人工智能。做到比人安全得多,其实比大家想象的容易。(马斯克)

注意他用的是"十倍安全"这个量级,而不是"跟人差不多"。十倍是他心里的及格线。这个标准决定了后面所有的技术选择和监管讨论。

为什么他敢说容易?这里有个容易被忽略的角度。人类开车靠的是一套效率极低的感知系统:注意力有限,视野有盲区,会疲劳,会分心,会因为一条微信把前车距离忘掉。机器不会。360度同时看,永不疲倦,反应时间以毫秒计。人类开车"大体上还行",靠的是路上大部分时间什么都没发生。

这句话放到今天依然成立:自动驾驶真正的难点从来不是把车开好,是把出错率压到人类的一百分之一以下。

二十亿辆车,每年造一亿辆,这道算术题决定了节奏

黄仁勋问技术路线,马斯克却先给了一个产业层面的约束。这一段是整场对话里信息密度最高的部分。

全球在路上的轿车和卡车,大约20亿辆。全球汽车和卡车的年产能,大约1亿辆。

假设明天所有新车都变成自动驾驶,要把现有车队全部换掉,需要20年。就算自动驾驶让车队总规模缩小,也得15年左右。而且不可能一夜之间全切,过渡期会拖得很长。

电动化是同一道算术。假设明天所有产能都变成电动车,替换整个车队同样要20年。而当时电动车的渗透率不到1%。

把这两组数字放在一起,一个让人清醒的结论就出来了:汽车行业的任何技术变革,都要用"十年"为单位计量。软件行业三年一代的节奏,在这里不成立。

马斯克讲这段话的用意,不是给自动驾驶泼冷水,是让人看清体量。一家创业公司做出自动驾驶原型,跟20亿辆车全部变成自动驾驶,中间隔着整个工业体系的惯性和产能上限。

这个提醒对2015年的听众尤其重要。那几年硅谷流行一种叙事:某个技术突破会像开关一样,"啪"地一下把旧世界关掉。汽车不是这样。哪怕技术明天就成熟,物理世界的替换速度也只有每年一亿辆。这条约束比任何算法突破都硬。

现有硬件能走多远:P85D 和它的传感器

黄仁勋自己是Model S P85D的车主。他描述了自己的体验:车道识别,识别路牌,靠计算机视觉判断车道线。他想知道,特斯拉的路线图跟别人有什么不一样。

马斯克给的答案是硬件打底,软件持续升级。因为特斯拉的车一直联网,新软件可以一直推。这台车上的功能会一点点变强,车主不需要换车。

当时那套传感器的清单,他说得很具体:

  • 360度超声波传感器,覆盖车身周围几米的范围
  • 一个前视摄像头
  • 一个前向雷达

就这三样。他的判断是,仅凭这套硬件,就能在自动驾驶上取得巨大进展:高速上自己转向、自己变道,都能做到。

他举了个例子说明"持续升级"意味着什么。当时刚推出的主动巡航控制,用雷达和摄像头融合来追踪前车。软件更新之后,它还会看前车的刹车灯,在对方踩下刹车之前就预判到要减速。同一套传感器,靠软件迭代,能力往上走了一大截。

但他同时划了一条线。同一套硬件理论上能让车完全自动驾驶,可靠性却撑不住复杂城市场景:时速30英里,到处是乱穿马路的人,小孩可能在路边玩,随时有东西从侧面冲过来。要解决这些,需要更大的传感器组合和更强的算力。

说到这里他转向黄仁勋,说英伟达正在做的Tegra芯片会是一个重要的推动力。

马斯克对自动驾驶的定义很有意思。他不谈能力,只谈可靠性等级。同一个系统,在高速公路上叫自动驾驶,在小学门口就叫危险品。区别不在算法会不会开车,在出错概率够不够低。

这也是为什么他敢说"我们知道该做什么,几年之内就能到"。这句话的技术含量不高,时间承诺才是真正的赌注。

如果车不撞,就不用造那么结实

黄仁勋提了个工程师式的联想:现在的车越来越重,因为要装更多结构件、更多气囊,扛住各种碰撞。如果车根本不太会撞,能不能放松这些法规,让车更轻、更省油、更好开?

如果你能指望不出事故,那就可以拿掉一大块碰撞结构和气囊。(马斯克)

这个想法在物理上完全成立。今天的车身设计有一大半是为了应对那0.001%的时刻。如果碰撞概率再降两个数量级,这些重量全都可以省掉,换成续航、空间或者成本。

但他马上补了一句现实:这事离得还很远。路上会长期存在大量老车,你没法假设周围的车都不会撞你。一辆全自动驾驶的车,夹在一堆二十年前的手动挡中间,照样得按最坏情况设计。

顺带说一句,这个数字对比很能说明问题。一辆车95%的时间停在车位上,剩下5%的时间里,绝大多数动作是重复的、低速的、可预测的。人类驾驶的整个安全体系建立在这种冗余之上,而自动驾驶要做的,是在没有冗余的情况下也保证不出事。

监管这道关:比人安全只是起点,还要再等两三年

黄仁勋问了一个很实际的问题:我想在开车上班的路上处理邮件,现在这么做是违法的。法律什么时候能跟上?

马斯克给出了一个清晰的时间差:

从一辆车确定比人安全,到监管机构允许它上路,中间还要两到三年。监管方不满足于"跟人一样安全",他们要看到大量统计数据,证明它明显更安全。

他提出的过桥办法叫影子模式(shadow mode)。车在路上跑,自动驾驶系统在后台计算"如果是我,我会怎么办",但不执行。然后对比每一条决策:这次会不会撞?哪些是误报,哪些是漏报?

跑出足够大的样本之后,拿一份清晰的统计论证去跟监管机构谈。监管方消化、观察一段时间,再判断认不认这个结论。马斯克认为证据会压倒性地清楚,他们最终会同意。

他顺带抱怨了一句当时的舆论:别对这项还不成熟的技术过度反应,别提前监管。

但他也承认,涉及公共安全,谨慎一点有道理,得先确认没问题再改规矩。他特别指出,当时并不存在"某个完全自动驾驶系统被监管卡住不让上"的情况,所以监管暂时还不构成瓶颈。这句话里藏着一层意思:真正的阻力不在审批,在技术本身还没到这个水平。

影子模式这个思路值得多说两句。它的巧妙之处在于绕开了"先批准还是先验证"的死循环。你不需要先拿到许可才能跑数据,车照常上路,系统在后台默默记录。等数据攒够了,讨论的基础就从"你觉得行不行"变成了"这份统计你怎么看"。

真正该怕的是"任意一辆车"被黑

话题转到安全,马斯克先把风险分了级。

如果有人能直接接触到你的车,就像能直接接触一台电脑,甚至一辆传统汽车,那他能做手脚不奇怪。更值得担心的是另一件事:有人能不能远程黑掉任意一辆车,或者一大批车。

特斯拉的精力主要花在防这一类上。思路跟做手机、笔记本一样:让系统性、成规模的入侵变得极其困难。他们花钱请第三方来攻击,主动找漏洞。

车内的架构做了隔离。信息娱乐系统被攻破,最多显示一条奇怪的提示信息,攻不进转向和电机。驱动单元控制器和转向控制器有额外一层安全设计,跟娱乐系统不在同一个信任层级上。攻击面被切成好几块,突破一块不等于全线失守。

他还有一句让人安心的补充:

你永远可以用脚压过刹车的力度,用手压过方向盘的转向。(马斯克)

这句话有前提,前提是车里还有刹车踏板和方向盘。如果几十年后这些都被拿掉,那么一旦出问题,后果会严重得多。他把这个隐患说得很直白,没有回避。

回到开头那个矛盾

黄仁勋一开始提的问题,其实没有被完全回答。

马斯克说自动驾驶是窄AI,所以不用担心。可"召唤恶魔"的担忧指向的是通用人工智能,是能自我改进、目标跟人类不一致的系统。这两个东西在技术上确实隔得很远,在时间上却没有那么远。

2015年,深度学习刚刚证明自己在图像识别上能超过人类。黄仁勋在台上推销GPU和CUDA,马斯克在台上讲车队、传感器和监管。两个人都清楚,脚下踩的是同一块地基。

十年后回看这场对话,最有信息量的地方是两个人对同一件事的判断速度不同。马斯克对自动驾驶的时间表乐观得离谱,对AI风险的时间表却悲观得离谱。这种不对称在旁人眼里是矛盾,在他那里是一套自洽的思维方式:具体的工程问题看得比谁都清楚,抽象的生存风险想得比谁都远。

而这场对话里真正经得起时间检验的,是那几组数字:20亿辆车,1亿辆年产能,20年替换周期,不到1%的电动化渗透率。这些数字到今天依然在决定每一家车企的命运,无论它的算法有多先进。

  • 评估一家公司的自动驾驶能力,先看传感器清单和算力配置,再看它敢不敢公开可靠性数据。只讲功能的,基本可以打折听。
  • 再遇到"某技术将颠覆某行业"的论断,先算一道题:这个行业一年能生产多少台,存量有多少台。两个数字相除,就是最短的替换周期。
  • 影子模式这套方法可以迁移到任何高风险决策场景:先让新系统在后台算、不执行,攒够样本再拿统计说话。比开会吵架有用得多。
  • 看AI监管新闻时,先分清楚讨论的是窄AI还是通用AI。这两类东西的安全标准、监管路径、时间表完全不同,混在一起谈一定谈不出结果。
  • 评估车联网安全,问的问题是"能不能批量攻击",不是"单辆车会不会被黑"。前者才是系统性风险,后者是个人倒霉。
  • 别用软件行业的节奏去预测硬件行业。三年一代的直觉放到车上,会让你错过所有真正重要的时间窗口。
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)。

0:04[Music] that that artificial intelligence is more dangerous than nuclear weapons and that's it potentially and and oh it goes on it goes on you say you say that it's like summoning the demon could be how do you consolidate rationalize the the that conflict between artificial intelligence of course deep learning that that obviously is going to be very important to self how you think that well I I don't think we have to worry about autonomous cars cuz that's sort of like a narrow form of AI and in not something that I think is very difficult actually good to to do autonomous driving to a degree that's much safer than a person is much easier than people think I think it's

0:54going to just become normal like it be like an elevator like no they used to have elevator operators and then we you know developed some simple circuitry to have elevators just automatically come to the floor that you you're at and you can just press the button nobody needs to operate the elevator the car is just going to be like that and the elevators these days are even smart I mean it knows it knows where to position an elevator so so that if you were to need an elevator it's pretty close to you cars in the future will be pretty smart about that too yeah you'll be able to tell your car like take me home go here go there anything and it'll just do it yeah at

1:34an order of magnitude safer than a person and in fact in the in the distant future I think it's probably going to be if people may Outlaw driving cars because it's too dangerous like you can't have a person driving a two-time death machine now if we if we have the right type of intelligence in a car we we also don't have to make the cars that heavy I would think you know cars are getting heavier and heavier and it's got more and more stuff in it because it needs to survive all these incredible collisions and things like that I wonder if if we were to to to design cars that that just simply don't Collide as much I wonder if we could we could relax on some of

2:16those laws and that cause more fuel efficient and lighter and better you could definitely do that if you could count on not not having an accident then you can get rid of a huge amount of the crash structure and the airbags and it we're a long way from that because there's always going to be some for a very long time there'll be some amount of Legacy cars on the road and and I think it is important to just appreciate the size of the automotive industrial base like it's not as though like when somebody makes an autonomous car that suddenly all the cars will be autonomous it's like there's 2 billion of them okay so the the total total number of cars and

2:54trucks on the road is is 2 billion in climbing the capacity of of car and truck production is about 100 million a year so if tomorrow all cars were autonomous it would take 20 years to replace the fleet assuming the fleet stay at the same size arguably it could get smaller things autonomous but still it's it's still you know maybe 15 years or something and it's not all going to transition immediately it'll take quite a while so and it's the same for electrification of cars changing that industrial base to electric I mean if if all cars were suddenly if all cars produce were electric tomorrow it would still take 20 years to replace the fleet and right

3:38now it's less than 1% so now you you you're you mentioned just now about about self-driving hardly easier than people think now you you have a your vision of of how to go from where we are today now my model my p85d has Lane detection and so it gets a little you know when I get close to a to a lane detects the the the speech signs and uses uses computer vision technology to do that and but and that's today's ad what is your what is your road map you know how is that different than other people's road map how do you think about how to get to self part yeah well you kind of need

4:19the the hardware Foundation sort of sensor and Computing foundation and then you can keep uploading new software at least you can with a Tesla cuz it's it's always connected so the the car that you have you noticed like it it's the features are steadily improving we now you know have active cruise control so it'll it'll use radar camera Fusion to track the car in front of you it's also looking at with the some of the things that are coming out it's it looks at the brake lights so it anticipates that the cars the brake lights are active it's going to get basically smer and smarter even with the current Hardware Suite so the current Hardware Suite is 360° ultrasonic

4:57sensors that go up to about just over M it's a Ford camera and a Ford radar so we'll we'll make even with just that sensor Suite we can actually make huge progress in autonomy we can certainly make the car steer itself on on the freeway you do Lane changes it it's really autonomy is about what level of reliability and safety do you want

5:26even with current sensus we we could make the carg go fully autonomous but only to but but not to a level of reliability that would be safe in say a complex Urban environment at 30 mph where the lane walking is out there and children could be playing and things could be coming at you from the side so in order to solve that you need a u a bigger sensor Suite and you need more Computing car and I think what you're doing actually with the the ters in the futures is super interesting and will really be a big enabler for Aon ofous driving we know exactly what to do and we'll be there in a few years right just just like mar mar is not quite as that's kind of

6:16the the the the spirit of of innovators I mean in a lot of ways in your mind you kind of you kind of see things solvable or arguably arguably solved and and a lot of it is is really about getting there yeah we'll take autonomous cost for granted in quite a short period of time it's amazing how comfortable you get and how quickly you comfortable so now what about government government policies like one of the things that I would like to do is I would I would just like to keep working on my email as I'm driving to work sure you know there's there's a some people some people do that already like I said I would like to do it without without without breaking

6:58the law so so where do you where do you think government intervention Falls in in some of this stuff because you know obviously if car drives by itself and it does it even better than people you would like it to drive by itself but largely the laws don't allow you to do that today right absolutely so how do we cross that bridge and and and how do you think about government intervention regulations right so I think it it'll

7:25from the point at which a car is definitely safer than a person there's probably at least another 2 or 3 years after that before Regulators will allow that to be the case cuz they will want to see a large amount of statistical proof that it's not merely as safe as a person but much safer so I think what you can do is you can run run it in Shadow mode and essentially say okay this is this is what the computer would have done in all these circumstances and was there a crash or was there not like what are the false positives false false negatives and then achieve a large population group and then and then make a really clear statistical argument of with the

8:07regulators and then they're going to digest that observe it for a while see if they agree with it and and then I think they will because the evidence will be overwhelming yeah and the evidence is actually already quite overwhelming that if you if you if you would would have noticed a brake light in front of you in the highway and you didn't you didn't crash into it we were in Collision a lot of s LI or you know IDE ideally hopefully people don't don't overreact with this with this unknown technology and and

8:38prematurely regulate no premature regulations well I mean reg I think when it comes to Public Safety I think there's there's an argument for being you know quite cautious and and making sure that things are okay before before there's a change and I mean I don't think it's the case that right now there's a fully autonomous system and Regulators are not approving it that that could really be a substitute for people but there will be in a few years have a mechanical failure of some kind or or a logic failure a fundamental logic failure so you can always overwhelm the the the braking of the car with your foot and you can overwhelm the steering wheel with your

9:24hands so but but when when there isn't a steering wheel there isn't you know brake pedal or something like you know many years from now then it's really really dangerous you know but even as it is right now what we spend most of our time on is making sure that it's it's very difficult to do a multi-car hack like if you have direct access to a car just like if you got direct access to a computer or or even a conventional car but that that's less of a concern than somebody being able to hack an arbitrary car or multiple cars so that's what we we focus on on is making sure that that that way it's a lot like a like a cell phone or or a laptop you know you focus on making

10:08sure that they they can't that it's very difficult for that to be any kind of systemwide hack so take over all put a lot of evid into that we have third parties try to attack it and and then certain parts of the the car at at the very fundamental level like the Drive Unit controller or the steering controller have an additional level of security M MH so somebody mayble to you know hack something that's cosmetic but it's much harder to hack something that's that's actually physically dangerous there's multiple levels of security so this way if you if you were't able to penetrate maybe the infotainment system it doesn't allow you quickly as a result of that right May

10:52display a funny message or something but it would not you would not be able to then control the steering or the the motor yeah

资源与出处

相关视频

同频道、同主题或相邻年份