2019 年,ARK Invest 的三个人坐在弗里蒙特(Fremont)工厂里,问马斯克一个略带挑战意味的问题。分析师 Tasha Keeney 把他 2013 年说过的话翻了出来:自动驾驶最后 10% 极其困难,必须始终留一个人在环里。她想知道是什么让他改变了主意。
马斯克的回答是,他没有改变主意。
最后 10% 依然极其困难,最后 1% 更是如此。他改变的是「完成」这个词的定义。
这场对话由 ARK Invest 录制,是他们播客 FYI(For Your Innovation)的一期。主持人是 ARK 创始人 Cathie Wood(当时还叫 Kathy Wood)和分析师 Tasha Keeney。整场访谈可以切成三块:产能与指数增长、自动驾驶的技术路线、以及最后突然聊起来的比特币。
信息密度最高的部分集中在自动驾驶。马斯克在里面讲清楚了几件事:他为什么认定激光雷达是死路,为什么纯视觉方案够用,为什么别人的自动驾驶会卡在局部最优,以及他凭什么敢说特斯拉的真实道路数据比同行多出一百倍。
2023 年 2600 万辆电动车,他觉得这个数字八九不离十
访谈从电动车的销量预测开始。ARK 有一个自上而下的模型:如果只看电池成本的下降曲线,加上电池包系统和新化学体系的改进,考虑到需求的价格弹性,全球电动车销量会从 2018 年的 130 万辆涨到 2023 年的 2600 万辆。五年,二十倍。前提是电池产能跟得上。这个数字大概是当年全球汽车总销量的三分之一。
马斯克的评价很直接:
「自上而下这么算,听起来差不多。你可能差个一两年,但差不了太多。」(马斯克)
ARK 自己的自下而上模型反而要保守得多。他们给特斯拉算的是那一年卖 160 万辆,据此得出一个熊市到牛市的目标价区间,每股 700 到 4000 美元。Tasha 当场承认,160 万辆这个假设看起来越来越站不住脚了,因为特斯拉领先其他车厂和科技公司的距离太大。
马斯克接过这个话题,讲了一段关于指数曲线的常识。
2017 年他被骂得很惨,原因是交付量不够。他的解释是,那一年还处在指数爬坡的起点,曲线下方的面积很小。等到曲线真正陡起来,底下的面积会大得吓人。问题在于大部分人是用线性方式思考的。
「我一直在说这件事,但人们不理解指数是什么意思。去年我们基本上把全球保有量翻了一倍,因为累计产量差不多等于公司历史上造过的所有车的总和。你拿线性去外推,那当然是不可能的。」(马斯克)
他顺手举了个更容易理解的例子。2018 年特斯拉冲到每周 5000 辆的峰值产能,比他最初的估计晚了大约六个月。在宏大叙事里,一个全新车型项目晚半年,真的不算什么。但媒体报道是按百分比来算缺口的,于是六个月的时间平移被描述成了一场巨大的失败。
同一个道理反过来也成立:在指数曲线上,日历时间挪动半年到一年,百分比差异会大得离谱。所以他对未来的猜测必须带上误差范围。半年和一年的区别,在指数曲线上可能是五倍的差距。
他自己给出的猜测是:2021 年特斯拉 150 万辆,2023 年如果电池成本下降的判断成立,全行业 2600 万辆不是没可能。他还补了一句关于中国的判断:特斯拉在中国大概能卖到 300 万辆。
顺带他解释了一个容易被忽略的口径问题。所谓的「周产 5000 辆」是峰值产能,不是季度平均。把公共假日和设备检修算进去,一个季度的平均产能大概是峰值的 80% 到 85%。媒体拿峰值当成平均值去倒推季度交付,误差自然就出来了。
访谈开头还有一小段关于资本市场的讨论。Cathie Wood 说 ARK 成立的动机之一,是创新在公开市场被系统性地错误定价,而大型资产管理公司已经开始跑到 pre-IPO 阶段去找创新。她感谢马斯克没有把特斯拉私有化,因为要享受指数级增长,就得留在公开市场里快速扩张。马斯克没在这上面纠缠,很快把话题带回了产品。
「功能完备」和「能睡觉」之间,隔着无数个 9
回到 Tasha 那个问题。马斯克说,他依然认为最后那一点点极其困难,但他对 2019 年有一个明确的判断:
「今年我们会实现功能完备的全自动驾驶。意思是车能在停车场找到你,接上你,一路把你送到目的地,全程不需要任何干预。今年,我对这一点是确定的,这不是一个问号。」(马斯克)
紧接着他做了区分,这段区分值得单独拿出来说。功能完备只是第一层。
第二层是可靠性。功能齐了之后,剩下的工作是沿着「9 的个数」往上爬。两个 9 是 99%,三个 9 是 99.9%,每多一个 9,事故率就下降一个数量级。要几个 9 才算够,是个工程和统计问题,也是一个成本问题。每多一个 9,需要的真实里程数都在翻倍,而这些里程只能靠车队一台一台车跑出来。
第三层是监管。监管机构什么时候认可这个可靠性水平,允许多少人在什么条件下使用,这不在特斯拉的控制范围内。
三层拆开之后,时间表就清楚了。功能完备在 2019 年,这一点他反复说自己能控制,因为他每周都直接盯着 Autopilot 的工程细节。至于什么时候普通人可以「上车睡着,醒来到站」,他的猜测是:
「我猜大概是明年年底。我认为那时候安全性已经足够了。但监管什么时候同意,我不知道。」(马斯克)
2019 年底,那个「功能完备」并没有如期发生。这是后话。但如果你把他自己讲的「9 的个数」套回去,就会发现这个延迟根本不意外,因为功能完备只是爬坡的起点,真正消耗时间的是后面每一层 9。他自己在访谈里就把答案说出来了。
监管这件事,中国比欧洲更懂技术
Cathie Wood 追问了一句,你觉得中国会先同意吗,那些市长和监管者看起来挺懂技术的。
马斯克的评价相当正面。他说中国的监管者非常有想法,也相当聪明,一路上打交道下来感觉很好。美国联邦层面同样不错。加州有时候会过度热情(overzealous)。欧洲偏保守,节奏慢一些。
但他也把话说得很清楚:监管基本没有卡住特斯拉。欧洲对导航式 Autopilot 的批准预计下周就能下来,这类事情在宏观视角下都是微不足道的延迟。全世界只有少数几个辖区格外保守。
监管愿意听什么?马斯克的答案很朴素:数据。你说自己有一台很快的电脑,一切都会好用,这只是个声明。你拿出几十亿英里的行驶数据,摆出事故率和干预率,证明不开 Autopilot 反而更不安全,这才是能谈的东西。他说到这个结论时用了一个词:无可辩驳(unequivocal)。
主持人提到一个细节。在 Autopilot 涉及的第一起致死事故之后,监管机构看完数据,态度反而朝特斯拉这边偏了。马斯克承认存在一种现象:人会过度信任这套系统。所以每开启一次,车都会提示必须看着路、手必须放在方向盘上。手离开太久,车会先报警,然后自动减速。
数据护城河:几百万个你想象不到的怪场景
关于数据,Tasha 问的重点在质量,不在数量。马斯克说,质量的优势来自一个别人很难复制的东西:干预数据。
每一次有人从 Autopilot 手里接管方向盘,这次干预的信息都会被保存并上传。系统不知道是哪台车,也不知道驾驶员是谁,没有任何个体归因,只知道「这里发生了一次接管」。然后工程团队去看要修什么才能消除这次接管。这个设计同时解决了两件事:数据能拿到,隐私不用碰。
他提到,现在已经开始做到不太需要人工标注了。人在路口怎么开,本身就是标注。
「有数以百万计的边缘场景,奇怪到你无法相信。不同国家的道路规则不一样,人们的预期也不一样。雨、雪、雨夹雪、冰雹、飓风、洪水、火灾、烟雾、沙尘,全都有。而我们的车几乎跑遍了所有这些环境。」(马斯克)
还有一类更麻烦的场景:规则本身被破坏了。施工路段的人可能摆错了牌子,锥桶没放,路面上突然出现一个大坑,或者一个从没见过的物体横在路中间。
数量级是这样的:特斯拉的真实道路里程比任何同行都多得多,可能是所有人的一百倍。主持人提到其他公司加起来大概占 5%,马斯克说,更接近 1%。数据量随车队规模增长,而车队规模在指数增长,所以这个差距会越拉越大。
激光雷达为什么是死路:局部最优 vs 全局最优
主持人们抛出了他们在研究 Waymo 和 Cruise 时的困惑。那些公司用的是确定性、规则化的方法,而特斯拉用的是概率性的深度神经网络。区别到底在哪里?
马斯克的回答毫不含糊:
「本质上是启发式的方法,会把你带到一个能力的局部最优,而不是全局最优。你必须用足够复杂的神经网络才能达到全局最优。这就是为什么他们依赖激光雷达是不明智的。它能带你走一段,然后就没了。一堆 if-then-else 语句加上激光雷达,解决不了这个问题。忘了它吧,游戏结束。你必须解决视觉感知,也就是理解。视觉解决了,其他什么都不需要,不需要任何别的传感器。」(马斯克)
他给的参照物是人类自己。人开车靠的是两只成像质量相当一般的眼睛,装在一个转速很慢的云台上。职业司机几乎不出事故。如果几亿年进化出来的生物视觉够用,工程视觉就没有理由不够用。
这里藏着技术路线的分岔。规则化方案在早期看起来进步飞快,因为每一个能想到的场景都可以手写一条规则。但规则的数量会随着长尾场景的暴露而指数膨胀,最终撞上维护成本的天花板。神经网络在早期笨得让人绝望,但它的上限取决于数据和算力,而这两样都在指数增长。起点高度和爬升斜率,是两件完全不同的事。
自研芯片:三年前立项,提升 2000%
访谈里马斯克提了一件他觉得外界重视不够的事:特斯拉的自动驾驶 AI 芯片要量产了。所有订购了全自动驾驶(FSD)的车主会免费拿到这块板子。
性能对比很夸张。相比之前用的英伟达系统,他说提升了一个数量级,实际数字大概是 2000%。主持人补了一个外部佐证:一位在英伟达做了九年深度学习的分析师,看完参数后认为特斯拉至少领先同行三年。
为什么要自己做芯片?马斯克讲了电子游戏的那个老道理:软件和硬件必须配套,再聪明的软件也没法在一台过时的硬件上跑出今天的画面。神经网络也一样,模型规模和硬件吞吐互相锁死。
当时的限制很具体。系统每秒大概能处理 100 帧,为了塞进算力预算,必须对画面做裁剪、变形,不能所有摄像头都跑全分辨率。换成新芯片之后,所有摄像头、全帧、全分辨率,算力还有富余。
用老硬件能不能做到全自动驾驶?他说原则上可以,但软件问题会难得多的多,因为你得不停地做算力取舍、玩各种技巧。硬件的余量,本质上买的是软件的自由度。三年前立项做芯片,赌的就是这个自由度迟早会成为瓶颈。
至于安全性的差距,他给了两个粗略的数字。用旧硬件硬啃,大概只能做到比人类安全一点点;换上新硬件,安全性可能到 1000% 的量级。这台机器还在不断学习,所以数字会继续变。
铺开路径:先高速,再路口,再一个国家一个国家
Waymo 那种打法是一座城市一座城市地开,带着地理围栏。特斯拉走的是另一条路。
顺序是从高速公路开始。原因有两个。第一是收益最大,堵车和走走停停最折磨人,高速上开 Autopilot 的体验差异最明显。第二是危险最大,高速上的事故速度更高,致死风险更大。
马斯克自己就是用户。他会为了能用上 Autopilot 而选择绕高速,哪怕那条路更长。他甚至还说自己已经不用 Waze 了,因为穿一堆弯弯曲曲的小街,精神开销太大,不如在高速上坐着。
下一个战场是路口。识别停止标志和红绿灯在开发层面已经不是问题,难的是歧义:一个复杂的路口有七八个灯,到底该看哪一个,人类司机有时候也说不清。
这一步会先在美国跑通,然后往其他地区移植。欧洲高速的样式差别很大,得一个个处理角落案例。挪威是优先项,因为那边的车主多。接下来是荷兰、德国、瑞士。
要不要开放给别的车厂:可以,但别改需求
Cathie Wood 问了一个开放平台的问题。马斯克的态度是这样的:
传统车厂并不容易合作。首先,没人真的在门口排队。特斯拉开放了专利,但真正用起来的人不多,这和提供一整套集成系统是两码事。如果有车厂愿意原封不动地用特斯拉的硬件加软件,特斯拉很开放。问题在于,对方坐下来谈的第一句话通常是「我们想改这六件事」。
「那样会严重拖慢我们。你想用一模一样的东西,没问题。但你不想用一模一样的东西。用超充网络、用 Autopilot 都可以,前提是别给特斯拉的工程团队增加大量额外负担。」(马斯克)
比特币:结构精妙,但太耗电
访谈最后几分钟,话题突然拐到了加密资产。Tasha 说这事儿最近老上新闻。
马斯克讲了个小插曲。推特上冒充他的以太坊骗子太猖獗,他回应了一下,结果账号被自动规则封了,大概是因为涉及比特币交易的判定。
他对技术本身评价不低。比特币的结构相当精妙,他有些深度参与加密的朋友,以太坊那套设计也有可取之处。但他不确定特斯拉把资源投进加密是不是个好主意,因为公司的使命是加速可持续能源的到来。
「我认为加密的一个缺点是在计算上非常耗能。这本身就会成为创造加密货币的一种约束。就现在这个阶段,挖出增量比特币极其耗电。」(马斯克)
Cathie Wood 从另一个角度补了论点。比特币网络上流转的交易额是万亿美元级别的,只是大家看不见。这些交易里没有披萨,也没有可乐,全是企业对企业。在非洲,把一种本国货币换成另一种的成本高得离谱,必须先换成美元。对他们来说,加密货币就是「跑在 IP 上的钱」,是免费的价值传输,还能绕开货币管制。
马斯克的回应很有意思。他没有反驳,而是把视角拉到更远:
「纸币是会消失的。加密是比纸片好得多的价值转移方式,这一点毫无疑问。当然有利也有弊。」(马斯克)
主持人最后确认了一句:特斯拉不会开始卖比特币吧。答案是,不会。
这场访谈里真正能拿走的东西
把整场对话拆开看,马斯克反复在做同一件事:把「某件事什么时候发生」拆成「哪些部分我能控制,哪些部分我控制不了」。
产能是可以控制的,所以他敢说 5000 辆每周;技术功能完备是可以控制的,所以他敢说 2019 年;用户行为和监管批准是控制不了的,所以他给的是猜测,而且提前把误差说明白。他唯一没料准的,是爬完那些 9 需要多久。
另一个反复出现的判断标准是上限。启发式加激光雷达能快速做出漂亮的演示,但天花板很低;纯视觉加神经网络起手很笨,但天花板取决于数据量和算力,而这两个变量都在指数增长。选路线的时候,起点的高度不重要,斜率才重要。
做技术、做投资、做产品的人,这场 2019 年的对话有几条可以直接拿来用的判断方法:
- 听到任何关于自动驾驶的进展,先问它的数据闭环是怎么建的。是人工标注的规则库,还是真实干预数据的自动回流。前者会随长尾场景膨胀而崩溃,后者随车队规模扩大而变强。
- 判断一家公司有没有被算力卡住,看它的软件里有没有大量裁剪、降采样、分时复用的技巧。有,说明硬件是瓶颈;没有,说明还有余量。
- 看到指数型业务的预测,把日期当成区间读,把数字当成条件句读。凡是没写「如果电池产能跟得上」「如果监管批准」的预测,都可以直接打折。
- 自己给团队定目标时,学他把交付物拆成三层:功能做完、可靠性达标、外部批准。混在一起报一个日期,最后一定是信任受损。
- 评估技术路线别只看谁的演示更惊艳,要看谁的上限受物理定律约束、谁的上限受工程投入约束。受工程约束的那条路,长期一定能走通。
- 用「9 的个数」来提要求。问对方现在是两个 9 还是三个 9,比问「离全自动驾驶还差多久」有用得多。
最后一条值得单独说。他在这一个小时里说了好些现在回看已经不准的时间点,但同时他也把不准的原因讲清楚了。预测错了不稀奇,讲不清为什么可能错,才是真的问题。
0:00welcome back to FYI today we're thrilled to be here in Fremont with Elon Musk I have with me Kathy wood Arkin best og innovator and CEO and I am titled what that was until today and I'm I'm Tasha Keaney I'm an analyst and I cover autonomous cars at arc Elon thanks for joining us today you're welcome get to join you have a great radio voice fYI this is called fYI
0:37for your innovation that's our weekly podcast Tasha there was a competition in the office and Tasha one as the best radio voice okay makes sense I'm aggressive I'll try talking and I also want to thank you Elon we're very excited to do this one of the reasons I started arc is because innovation is being completely mispriced in the public markets as the public markets go passive yeah totally
1:10John Bogle who's awesome rest in peace like when it's Pat last comment was like there was too much the passive index funds are too big and usually got really you know obviously came up with the you know the low fee index fund yes it's and the other thing that's happening again a problem for innovation in the public markets is the big public asset managers they started looking for innovation in the pre IPO space right so you know
1:39where we operate is in efficiently priced like I've never seen in my career most people looking at our portfolios would not agree because our p/e ratios to this year next year look very high but we want companies like Tesla to be investing aggressively to capitalize on the massive opportunities we see ahead so we thank you for not going private and because we do think if you are going to achieve your goals you need to be public to scale that quickly and to enjoy the kind of exponential growth that you will enjoy because you're on the right side of change so we're really excited to be here now what we're not going to talk about too much is electric
2:26vehicle sales and certainly not production right because we believe and just want to get your high-level reaction to this based on and this is based on Sam choruses battery work cost declined some of the amazing things you're doing with battery pack systems new chemistries and so forth the cost declines if we were relying on that analysis alone we would see evey sales given the price elasticity of demand increased from 1.3 million last year to
2:5626 million in 2023 so in five years so that's a 20-fold increase if the battery capacity were there and so for M production capacity what do you think about that that would be the high level assumption top-down sounds about right Wow so that's a about 1/3 of total global sales at that time I mean you might be off by a year or two but not by much that's huge now for our bottom up modeling we're very conservative so the range that we are out there in terms of our bare bull range 700 to 4,000 that is based on one point Tesla selling 1.6 million units in that
3:42year which given our top-down analysis and given how far you are ahead of any other auto manufacturer and any other technology company that 1.6 million is increasingly looking way too conservative if you can scale to take the proper amount of share or the share you should take given how advanced you are so maybe you could give us some of your thoughts there yeah Webb says that if I give estimates these are really there's a lot of guesswork here and especially on an exponential curve a year or two different space is is enormous he says we're gonna lot of criticism for the number of cars we delivered in 2017 because if you say like the area under
4:26the curve of production in 2017 was quite small because it was the beginning of an exponential ramp but then once that got going the area under the curve was enormous and that's why people were so shocked I kept trying to say this but that people don't understand I guess what an exponential means or we don't there are all linear thinkers yes yes exactly so like last year we basically doubled our global fleet because we cumulatively we're made a little bit more knowing that's bad about as many cars as we had made in our entire history right exactly
5:02what the audience out there it was impossible yes and if you do a linear extrapolation it certainly would be but when making estimates against an exponential small changes in the calendar break point have enormous percentage differences the time difference is small but the percentage difference is enormous now just I'm saying things that you guys know but it's this is for the broader audience out there yes no we want this yes we want people to understand this right like getting to 5,000 cars peak production per week was offset by approximately six months from my initial estimate last year so I thought we'd be there and to 2017 it took us six months
5:46longer in the grand scheme of things six months late for a massive new program is not much but this was characterized in the press in terms of the percentage of units and instead of a calendar shift and so it was perceived as a massive shortfall when in fact it was merely a six-month delay and when I give these say my guesses for the future if you move six months or 12 months these numbers are it can be 52 harps that different so but if you say like what's my guess for 2021 is one and a
6:19half million cars for Tesla tell me like this 21 and in 2023 which is that 26 million if we knew only the cost decline associated with the batteries 22:23 number if we're right could be twenty-six million but it it's it's actually dependent on what you do more than almost any and what China does and what you do in China again just guess probably we do three million so
6:45very exciting times ahead that's way above what anyone is thinking out there and even above what we have assumed bottom up just to remain conservative because we already seem a little out there in terms of our expectations so what we wanted to do with this podcast is spend more time on your autonomous strategy we're very happy that retail investors are asking questions about this on your call we wish more institutional institutional investors would understand how important this is not only to your model but how it's going to change your valuation from that of from today to really software-as-a-service almost so yes so maybe sorry yeah no absolutely we sort
7:32of think of of course maybe long time as being effectively carriers for the autonomy that's autonomy software so it's like there were vehicle literally and figuratively for autonomy as the software that rides on them yes yes so Tasha has been doing a lot of our work on autonomous taxi networks we've been working on these models for five years and it's extended into autonomous truck platoons into parcel drones and
8:03passenger drones so we're very focused on this and would love to explore it a little bit more with you in terms of what you're doing Tasha yes so my first question is you know Tesla wasn't always pursuing fully autonomous driving if you look back in 2013 there's a couple quotes from you saying you know that last 10% is incredibly hard and you need to keep a human in the loop what's changed your mind I still think the last 10% is over ton of me is extremely difficult or even the last 1% of autonomy is really difficult I think there's like a couple of things that we helpful to calibrate with doing and if I may go back to to one thing like generally what you're
8:39like for example if I say we will reach 5,000 cars a week what I mean is that's that that's the peak production but then just-just reference 1 if you have P production of X then you're gonna be like some between 80 and 85 percent of that for average production through the quarter when taking public holidays and equipment maintenance into account then also for in defining autonomy or full self-driving I think we will be feature complete full self-driving this year meaning the call will be able to find you in a parking lot pick you up take you all the way to your destination without an intervention this year I would say that I am certain of that that
9:21is not a question mark however people sometimes will extrapolate that to mean now it works with 100 cents certainty we're requiring no observation perfectly this is not the case once it is future complete then you're sort of kind of a march of nines like how many nines a reliability do you do you want it to be and then when do regulators agree that it is that that is that reliable so there's future complete full self-driving this year with certainty this is something that we control and I've managed to autopilot engineering directly every week in detail so I'm uncertain to this then when will regulators allow us even to have these features turned on with human
10:05oversight but that's a bearable which we have limited control over that is when will regulators agree that that these things could be done without human oversight that isn't another level beyond that so these are externalities we don't quite control and the conservatism of regulators varies conserve a lot from one jurisdiction to another my guess as to when we would think it's safe for somebody to essentially fall asleep and wake up the destination probably towards the end of next year that's that's what I wouldn't I would think it's most likely it will be safe enough for that I don't know when regulators will agree do you think they'll agree in China
10:46first could you imagine that happening China really wants to leapfrog the rest of the world and as you've said the the mayors of cities and presumably the regulators are tech savvy because that's what's going to get them there we've generally found the regulator's and China to be very thoughtful quite smart yeah I've really found them to be very good we've also found the u.s. at a federal level to be very good as well California can be a little overzealous but US federal China federal it's good Europe is a little conservative let's regard yeah we've heard sort of similar things about Europe so the environments moving a little bit slower they're interesting so
11:26I mean this is an amazing technological feat what gives you the confidence that this is a solvable problem and then why should Tess look either one to solve this well first of all I think it's helpful to clarify with people think sometimes that I'm like a business person or finance person or something like that I I'm an engineer I do engineering always have so Irena wrote software for like 15 years 20 years and I understand technology and software are quite fundamental level I know what we need to solve to make will self-driving feature complete I think we've got an extremely good technical team I think we really have the best people it's an
12:03honor to work with them I'm certain that we will get this done this year so can I ask you about autopilot so to what degree is consumer use of autopilot important in terms of furthering along the process because it seems like you know not every consumer is going to use it it might vary person a person should you sort of incentivize people to use it maybe auto initiate is that even possible we're here I think we're talking about quality of the data is that what you're talking about the data that you're getting because you can monitor all the data but the quality with autopilot is superior I would imagine the advantage that we have that I think is very difficult to
12:38overcome is that we have just a vast amount of data on interventions so effectively the customers are training the system on how to drive and there are millions of corner cases they're so obscure and weird you wouldn't believe it I mean there's this different road walking is different rules in different countries different expectations you've got rain snow sleet hail you know
13:02hurricanes floods fires smoke dust it's insane I know but we've got cars and almost all really in all those environments and so we every time somebody intervenes takes over from autopilot it saves that information on and uploads it to our system we don't know which car it was or how it what happened you know there's no individual attribution for the car we just know that that intervention took place and then we see what is required to fix that intervention and it were really starting to get quite good at not even requiring human labeling basically the person say drives the intersection we know then and is there by training order pilot what to do right do you
13:46think one of the questions we debate a lot when we look at what wait happening is at what is happening at way Moe and cruise automation is that it seems as though the artificial intelligence are deep learning our deep reinforcement learning that you're using you know is probabilistic they seem to be more deterministic in their approach because we're just looking at what's going wrong there yes and that's what that suggests is does that wait yeah essentially and heuristics approach to this will result in a local maximum of capabilities so not a global maximum I think you really have to apply a sophisticated neural net to achieve a global maximum and that's why their
14:30reliance on lidar is unwise it gets you to a certain point but no further and so basically a series if then else statements and lidar is not gonna solve it forget it game over you have to solve vision perception essentially understanding and with vision and then it's off you don't need anything else no other sensors at all when we drive cars with basically two cameras that aren't very good and on a gimbal that doesn't move very fast and a professional driver will almost not have it almost never have an accident yeah yeah you know a
15:04lot of the time we get a lot of pushback in this idea of data and sort of what data is more important than others I guess maybe highlight for us like a I mean Tesla has more real-world miles than anyone else by far I think we must have more than everyone else maybe a hundred times more than 5% I think on you you said they have altogether 5% of its probably closer to 1% yeah I mean that's amazing so there must be sort of events that are specific to Tesla in there that no one else has I guess could you sort of talk about some of that uniqueness I think it's just really the long tail of weird events the million weird situations you know also it's where the conditions all kinds of
15:45road conditions situations where the road rules aren't even followed like they're not always follow it's like you know say somebody working on the road might make a mistake and then suddenly you've got a situation where there no Road cones or there's a big hole in the road or something really nutty or completely a strange object that's on the road that's not recognized so it's really that the reason tails I think is making rapid progress is because we have vastly more data and this is increasing exponentially with as our fleet is increasing exponentially our data is increasing exponentially then I think what Blake said the best technical team
16:21and although I've brought it up before people don't seem to really be taking note of the fact that the Tesla autopilot AI computer is about to roll into production anyone who's ordered full self-driving will get that real plate we'll get that for free it's really you know said that's like maybe it's like a order of magnitude improvement over the Nvidia system that we had but it's really more like or 2,000 percent yeah analyst on AI
16:48deep learning spent nine years at in Nvidia and concluded that you were at least three years just looking initially at the specs at least three years ahead of anybody else out there any other auto manufacturer yeah and we started the chip program about three years ago because it just seems as though we would want to advance making things if you want to have a complex neural network you need a combination of software hardware and your software needs to be that much better in order to compensate for hardware if the hardware's weaker you know if you look sort of like say have video games and how they progressed is a combination of software and hardware no
17:24amount of clever software could produce a video game on an old hardware that you have today just doesn't matter you know so the same thing with neural nets so right now we can process on the order of you know a hundred frames a second and we really need to do a lot of work in terms of cropping the frames and sort of bending the pixels and not going to full resolution on all camera isn't that kind of thing with the card hardware we're full frame is full resolution with the tip of the Tesla hardware all cameras at full resolution full frame this fool hasn't tapped out yeah so is the amount of compute processing power that you need for a full autonomy fixed or would you
18:04have to upgrade this every two years how do you think about that well I mean I do think that we could we can achieve it principle achieve full autonomy with the end video hardware but it's much harder software a problem and it's like so you know like you so really have to try to budget your compute and do all sorts of tricks to manage how you use your compute so it's a harder software problem with hardware that's you know two thousand percent better you just you don't have to do that that constant budgeting and so the software software problem is much easier so like I think with the current hardware and a lot of effort we could we could get to full
18:38self driving it's maybe maybe being like fifty-two harps and safer than a person but with hardware three I think it's probably like a thousand percent safe or the new prison are you presenting this to the regulators do they understand this do they want to understand this I think they are sent data so if we show you know billions of miles with a given safety level then they will they sort of appreciate that I mean somebody saying like hey we've got this really fast computer and everything's gonna work it's like well you know that's just a statement like but if you've got hard data billions of miles and you can show the accident rate and the intervention
19:10rate and that essentially it's unsafe if you don't have autopilot on right which I think really is unequivocal at this point so no matter how you slice slice the data is unequivocal that's this point that it's safer to have a pilot on the data comment you just made we were very interested that the regulator's after they examined the first fatality I think it was the the data that data around the first fatality that they just swung in your favor I think it was shocking and I'm surprised more people haven't really taken note of that yeah there is a phenomena that we that we noticed that would that would happen essentially it's like if you yeah if you
19:50will just get it overconfident with the system even though we repeatedly won them you must pay attention to the road like literally every time you use autopilot it says you must pay attention to the road you must keep your hands on the wheel every single time you use it it says this and if you take your hands off the wheel for too long it will start beeping at you and then slow down and I can't think so anyway so just but really at this point there's this just flat-out no question let it safer was you know like I would recommend it to anyone it's just getting better so yeah so have you noticed a change in the regulatory environment are you feeling sort of more confident than
20:24previously or is this sort of you know it's still such an open question that I mean Tesla's in a really unique position because you have the data to show them so it almost seems like you know you could be having sort of more advanced conversations about sort of proving that safety level I guess what are your what are your thoughts right now we're with a few exceptions we are not being held back by regulators is that a u.s.
20:46comment or is it global I mean it's that there are a few jurisdictions in the world that are more conservative right now it's it's I would not say that we're held back by regulators in my inner way it's like for example like we expect to get the latest autopilot approval navigator or pilot I think is gonna get approved in Europe next week or something like that but this is these are tiny delays and the grand scheme of things does really matter yeah that's great and so if you think about sort of this development of autonomous driving and the rollout strategy you know Tesla's not taking this other approach that's a way mo and others are doing going city by city
21:21you're gonna do a wider rollout they're still sort of I don't know if you want to call it geofencing but there's you know it's you start with highways and then you go through the Baath highway and then are you going to go country by country or sort of how are you thinking of this evolving in terms of sort of the full steps to I guess to get to full autonomy you know then tackling intersections are you sort of adding feature feature feature and then country by country yeah we start off with with high rate because that tends to be the what actually matters the most if like you if particularly if you're stuck stuck in traffic and stop-and-go traffic that's
21:53very painful and it has the the most benefit to have autonomy on freeways which are usually condition jest that in almost every city in the world so in fact myself and I was doing is even if those are shorter way home you still take the highway because you can use autopilot I stopped using Waze for example just like you know take the highway cuz then I bought a pilot on your going through a bunch of windy streets verses which is kind of like a lot of mental overhead as opposed to just sitting on highway increasing lung is better so we don't like got gone with okay what's gonna add the most valued people but also how you know highway accidents hits me higher velocity and so
22:31more potentially more dangerous like if fatalities are very fortunate to speed blow best fan speed it's very difficult to hurt yourself or die in it in particular in Tesla boson speed it's more dangerous so it's it's like okay where can it be most helpful from a convenience and safety standpoint so it's refocus on highways for that reason then intersections all the next thing and some of them that they have a lot of variants and intersections so that's what you know we're working right now and you know it's working at a development level no problem recognizing stop signs and traffic lights but do you do get ambiguity and some complex intersections
23:07with traffic lights like which ones the light the right light to focus on even if you're a person it's not always clear so that that's what we're working on there so we'll try to make that work in the US and then we'll extend that functionality elsewhere so it and like highways in Europe are different particularly we just deal with corner cases so we're kind of making it work in like Norway as a priority for us because we've got a lot of customers there and I think we're close to having that ready Netherlands Germany Switzerland I guess one general question what do you think of letting other automakers build on Tesla's platform haha I mean generally we found it's like quite it's
23:42not it's not easy to work with traditional automakers it's not like first of all they're not exactly banging down our door to work with us nobody took you up on your patents actually depends they have that's very different from say creating an integrated system so you know if there was a quota maker out there that wanted to implement the same hardware system as as Tesla and and yours our software would be very open to it but we're not going to change it what has happened is they want to work with us but then they'll say over we want you to change the following like six things like no slow us down massively yeah and and so we know it's like if you want to use
24:20exactly out of thing that's fine but then they don't wanna use exactly out thing we're open to other or to make is using a supercharged network or opens MT isn't using our autopilot system I just need to make it work without a ton of overhead on Tesla engineering absolutely the right thing to do I have a little bit of a different question for the it's a little bit off-topic but it's in the news or it's becoming more in the news so it's about crypto crypto assets crypto you know -huh crypto seriously Chris seriously seriously very seriously I think you and Jack Dorsey have chatted a bit recently about this right but coined the theorem scammers were so
24:57rampant on on Twitter that I said it just join in SS had like one point one buy some Bitcoin my account got suspended like some automatic rule that would if you try to sell Bitcoin or something it's a little different from crypto currencies although given your history in the payments ecosystem it would be very interesting to know if you agree with Jack and there is going to be one native crypto currency when it comes to the internet he thinks it's Bitcoin it's interesting have some friends mine that are could really involved in crypto I mean I think like the Brit coin structure was quite brilliant yeah that seems like the Samaritan theorem as well
25:47and there may be some of the others but you know I'm not sure it's like I'm not sure that it would be a good use of tells the resources to get involved in crypto I mean really just trying to accelerate the events of sustainable energy and I mean I think actually one of the downsides of crypto is that it computationally it's like quite energy intensive so like that to be some kind of constraint on the creation of crypto so but it's very energy intensive to creates like the incremental Bitcoin at this point yeah but it's it by the same time there were 1.3 trillion dollars worth of transactions in Bitcoin and we don't see it here because it's not for
26:22pizza or coke it's business-to-business in the correct is business-to-business in Africa where it is prohibitively expensive to convert from one nation's currency to another you have to go through the dollar that I mean it really is very important to its money over IP for them it's free transmission of money and that's really important to opening up the world and you know you know that bypasses currency controls yes yeah yes yeah yes it's it's
26:55well yeah paper money is going away yes and crypto is a far better way to transfer value than pieces of paper that's for sure without a doubt that has pros and cons just to clarify Tesla's not going to start selling Bitcoin so no well we we thank you so much for doing this and thank you for just doing what you're doing to change the world and make it a better place right you know that's what we're doing in terms of it no seriously think of Silicon Valley that you know comedy I'm
27:33sure I didn't buy social mobile different we totally agree with you yeah so yeah okay you're doing the real thing cool so thank you Thanks [Music]
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