基本消息是，所有驶出工厂的 Tesla 车辆都配备了实现 5 级自动驾驶所需的硬件。也就是摄像头和计算能力方面的硬件。

我们生产的每一辆车，大约每周 2000 辆，现在交付时都具备 5 级，也就是硬件有能力

实现全自动驾驶或无人驾驶。

所以我们还需要一些时间

在未来完成软件验证，并且显然还要获得必要的监管批准。但重要的是，

基础已经打好，使这些汽车能够

以我们认为至少是人类 2 倍、也许更高的安全水平实现完全自动驾驶。

所以我想，对大多数人而言，这现在就正在发生，可能出乎意料。

所以，是的，我们对此相当兴奋。这基本上也是 Model 3 发布的第 2 部分，也就是 Model 3 也将配备实现完全自动驾驶所需的硬件。事实上，Tesla 生产的所有汽车

从现在起都将配备

实现完全自动驾驶或无人驾驶所需的硬件。

目前的情况就是这样。而且这全部采用 Tesla 视觉软件。所以在、在视觉处理方面，我们没有使用任何第三方软件之类的东西。这是 Tesla 开发的神经网络。

而且，是的，尽管它在某种程度上独立于硬件，

它可以，我们实际上可以让它运行在

Nvidia、AMD 或 Intel 上。

我们的确选择了 Nvidia Titan GPU

作为神经网络的主芯片，但这是一次相当艰难的

抉择，尤其是在 AMD 和 Nvidia 之间。但最终我们认为 Nvidia 的硬件更好。

所以这大概就是我认为相当重大的消息。

现在说完这些，我们可以进入提问环节。

好的，第 1 个问题来自 Transport Evolve LLC 的 Nikki Gordon Bloomfield。您的麦克风现在已开启。

晚上好，埃隆。我有一个，我有 2 个问题。我不知道你是否方便回答 2 个，所以请选 1 个。如果不方便的话，Autopilot 迄今已证明非常可靠，其安全性方面的问题很少。但我想问的是，当你们车辆上的 5 级自动驾驶启用时，Tesla 是否会像沃尔沃承诺的那样，为涉及完全自动驾驶 Autopilot 模式的撞车事故提供赔偿保障？

不，我认为那应该由个人的保险来负责。我的意思是，就像，你知道，如果是我们设计中固有的问题。那么我们当然会对此负责。但我认为，一旦你使用自动驾驶汽车，就很像建筑物里的电梯。奥的斯会为世界各地的所有电梯负责吗？不，他们不会。所以。

归根结底，这里真正重要的是绝对安全性究竟如何？

有一件事我应该提一下，坦率地说，这让我相当不安，那就是媒体对 Autopilot 撞车事故的报道程度。相较于媒体对每年在人工驾驶撞车事故中死亡的 120 万人的报道之匮乏，Autopilot 撞车事故基本上几乎没有，我认为这件事无法给媒体带来好印象。真的不能。因为。而且真的，你需要认真思考这一点，因为如果你写了一篇负面文章，实际上劝阻人们使用自动驾驶车辆，那你就是在杀人。

下一个问题。

下一个问题来自《连线》的 Jack Stewart。您的线路现在已接通。嘿，晚上好。

我想知道，埃隆，你能否具体说明你如何看待这个

一旦这些车辆开始启用相关功能，自动驾驶功能将如何推出？新闻稿称，新功能将一点一点地开放。您对此有何设想？我们会立刻看到完整的城市驾驶功能，还是会先看到一套较为有限的功能？

是的，由于这是一个新平台，尽管我们一直在投入，我们已经测试了1年多，但最初会禁用这套功能，至少在最初几个月如此。配备Hardware 2的车辆，也就是我们所称的Hardware 2，这可以说是完整的自主驾驶套件，实际上功能会比配备Hardware 1的车辆更少。因此，我们预计在Hardware 2通过实际道路验证后实现功能对等，大概是在12月，也就是可能从现在起2或3个月后。

所以如果是 2 或 3 个月，实际上 Hardware 1 车辆会比 Hardware 2 车辆更好。然后

每隔，大约每隔2或

3个月，此后就是我们预计会发布自动驾驶能力重大改进的时候。

而我们的目标是，我对这个目标感觉相当有信心，我们将能够进行一次全自主驾驶演示，从洛杉矶一路开到纽约。

所以基本上是从洛杉矶的一处住宅

出发，比方说，在纽约时代广场让你下车，然后车辆自行前去停车；到明年年底，整个过程无需触碰一次，包括充电器。

采用新硬件的车辆，其Autopilot会怎么样？您能具体谈谈吗？

采用新硬件的车辆，其Autopilot会怎么样？嗯，这个

新硬件将实现自动驾驶，而这与Autopilot不同。

Autopilot这个术语已经使用了半个多世纪，指的是一种辅助功能，也就是飞机上为飞行员提供的飞行辅助。

这就是我们选择使用它的原因，它并不代表自动驾驶，正如

飞机的Autopilot不会让飞机实现自主飞行一样。

所以我只是想弄明白。也就是说，Autopilot迄今一直与贵公司的高级驾驶辅助技术相关联，那么这套系统会怎么样？如果可以这么说，随着你们推进这款新硬件，这个名称会怎么样？新闻稿里没有说清楚，我只是想了解你们是如何处理这件事的。

是的，会有

基本上会有2个，我

是说，购买购物车时会有2种选择。

一个是我们所称的增强版Autopilot，它在某种程度上与

我们一直提供的Autopilot非常相似，不同之处在于它配备了冗余的前置摄像头，并且有一个左后方

摄像头和一个右后方摄像头，以及经过显著改进的超声波声呐

和强大得多的计算能力。增强版Autopilot的最终效果是，你应该能够从，从高速公路入口匝道行驶到出口，以及

在多条高速公路之间切换，并且

超车和绕过其他车辆，而无需触碰任何东西。

在增强版Autopilot之外，还有某种

完全自动驾驶能力，它将处理城市环境中复杂得多的情况。

所以人们购车时会有2种选择。要么基本上是Autopilot的改进版，要么是完全自动驾驶。所以一个大概有4个摄像头，另一个有8个摄像头。

是的，这大致就是现在会发生的情况。配备Hardware 1 Autopilot的车辆将随着软件改进而继续改进，随着车队学习，随着我们积累更多里程，它会继续变得越来越好，但它会受到硬件上所配备的基础硬件的限制。

1的车辆在这方面只有1个福特摄像头，它们有超声波传感器，但它们有

一半的分辨率范围，配有雷达。

所以在那个背景下

传感器套件，硬件磨损的汽车将随着时间推移继续变得更好。

谢谢。我们的下一个问题来自《旧金山纪事报》的大卫·贝克。您的线路现在已接通。

感谢您参加这次电话会议。实际上，我想就您刚才谈到的内容继续追问。您能否向我们介绍一下，目前已经上市的现有Autopilot汽车所配备的硬件套件，与这套新套件之间有哪些差异？换句话说，您刚才提到了更多摄像头、更大的传感器探测范围。您能否详细说明其中的每一项变化？

可以，当然。

所以我们从1个摄像头增加到

8个摄像头，其中3个是前向摄像头。所以我们在前向

摄像头方面有冗余，算是向前看，而且我们

拥有360度覆盖，也就是车辆周围的视觉覆盖。

所以那是一个相当大的升级。

计算能力提升了40倍，就像是计算能力提升了4.0倍。所以计算能力有了巨大的提升。

事实上，这台计算机将能够

每秒执行12万亿次运算，基本上

一台装在汽车里的超级计算机。

然后，超声波传感器是下一

代超声波声呐，其大约有2倍

当前声呐在分辨率方面的范围。而且它也是360度覆盖。我的意思是，还有一些其他的小改进，比如GPS更准确，更新也更频繁。IMU，也就是初始测量单元，以及其他几个方面的传感器也有一些小改进。

你好，埃隆，我只是想知道你对这方面的法规有什么看法？它会不会因为不同地区而受到限制，比如说，你知道，哦，在得克萨斯州可以这样做，在加利福尼亚州可以那样做，各地会有所不同？还是说你会直接宣布，任何地方的每个人都可以使用？在监管方面你是怎么考虑的？

嗯，这不取决于我们，而是取决于监管机构。希望美国的情况不会

变得四分五裂，以至于在

每个州都不一样，因为我认为那将会

损害美国消费者的利益，而且

我认为在欧盟，我相当有信心它会采用统一标准。

所以真正的问题是，什么，

你知道，公众认为什么是适当的，监管机构认为什么是

适当的，以及收集足够的数据，因为这个

系统将始终以影子

模式运行，这样我们就能收集大量统计数据，以展示

假阳性和假阴性。

计算机本会在什么时候采取行动，而那是否会避免一起事故？或者，如果计算机本会采取行动，而那会造成一起事故。但那是在影子模式下运行，所以我们可以判断它什么时候本会错误地采取行动或没有采取行动。

将其与现实世界中本应采取的做法进行比较。而当它

得到具有统计显著性的结果，表明

一个

相较于人工驾驶汽车事故率的实质性改善。

我认为到那时，监管机构很可能会放心地批准它。但这一审批流程在世界不同地区确实可能截然不同。这不是我们能够控制的事情。

最后再追问一下，你知道，你谈到每2到3个月推出新功能。美国或欧盟或其他任何地方的监管环境会阻碍你们推出这些改进吗？还是你们会先推出，然后，你知道，先斩后奏？如果你们认为它、它更安全的话。

嗯，我的意思是，我们一直都在各个国家的监管框架内推出自动驾驶功能。

所以这不是一种，你知道，请求原谅或请求许可。

而是我们，你知道，我们开展工作，

我们

仔细研究法规，并确保我们的所作所为符合这些法规。

是的，我们不能采取除此之外的任何做法

因为那会违反法律。

你好，埃隆，谢谢你接受提问。

我想知道，既然你们现在正转向硬件2，同时还会继续支持硬件1，是否会有一个时间点，硬件1的潜力将达到上限，然后，你知道，车主需要升级到下一辆、更新的汽车，才能获得你们所能提供的完全自动驾驶能力。

自动驾驶肯定也需要硬件，也就是我们从本周起开始交付的硬件。因为你不能，你就是无法用1个摄像头做到这一点。要实现自动驾驶，你需要8个摄像头。你还需要强大得多的计算能力来运行视觉系统。基本上就是视觉 AI。

所以升级汽车上的硬件并不现实。

这就像给汽车做

脊髓移植，从这方面来说。

所以即使可行，也最好避免。但我应该说，拥有硬件1汽车的人应当记住，他们的汽车实际上会比硬件2汽车拥有更多功能。至少到12月之前如此，也可能更晚，而且大概要到明年某个时候，硬件2的功能才会开始超过硬件1。而且自动驾驶功能的成本也更高。所以这并不是，你知道，像完全自动驾驶这样的功能确实需要支付更高的价格。

要激活全自动驾驶硬件套件，价格是8,000美元。相比之下，硬件1的 Autopilot 是3,000美元。所以价格差异很大。而且至少在接下来的几个月里，搭载硬件1的车实际上功能更多。我希望能有其他办法。这不是，只是我不知道该怎么进入旧车内部，再安装7个摄像头和一套新的线束。

所以没有，只是。

我希望我们能用其他办法做到。

但没有别的。没有。

你好，我想知道你们是否计划把它做成一套可供其他汽车制造商使用的套件。

我认为这很难做成一套套件，因为它需要

软件、传感器和

计算能力的紧密集成，以及进行大规模无线更新的能力。

它不只是某种可以加装到汽车上的套件。我的意思是，我们的系统还做了一件事，就是如果你。而且我们希望会在今晚晚些时候和明天早上发布一些视频，展示车辆运行的情况。

但实际上你无法，除非仔细看，否则你甚至分辨不出一辆

车搭载的是硬件1还是硬件2，因为我们非常谨慎地让

每个

摄像头都成为车身框架的一部分，从而不会有任何东西凸出来。

这绝不会让车变丑。

所以没有奇怪的凸起物。

这一切都极其不显眼。

投入了大量精力，以免

影响汽车的美感。搭载 Hardware 2 的汽车和搭载 Hardware 1 时一样漂亮。

而且根本没办法把

它做成一套套件，再装进其他汽车里。

这不现实。

是的，我对你的看法很感兴趣，如果可以的话，请再详细谈谈你对安全问题的看法。统计数据并不会收集汽车有多安全，只会收集它有多危险。那么，你们如何传达你们的半自动驾驶汽车事实上更加安全这一观点？

这并没有那么困难。我的意思是，只需说明，你知道，有多少

英里的路程已经行驶，而你

知道，然后有多少人死亡、严重受伤，也就是诸如无法康复的伤害，或者轻微事故。

而且显然，还有更多某种程度上的轻微事故。

而严重事故也比死亡人数多。因此，这就为比较相对安全性提供了一个丰富得多的统计样本集

，也就是自动驾驶与非自动驾驶的相对安全性。我们持续看到明显更好的结果

：采用自动驾驶时比不采用时更好。

而且随着系统得到进一步完善，这种情况只会随时间推移而改善。你好，埃隆，我有个简短的问题

，关于恶劣天气，以及硬件

2 相比硬件 1 有何改进？嗯，恶劣天气，你是指它能否在雪天之类的情况下工作，还是说你所说的恶劣天气是什么意思？抱歉，任何时候，比如说，你知道，可能会有

摄像头上出现冷凝水或降水的情况

往往会成为一个问题。

你们是怎么处理这个问题的？

哦，对。所以所有摄像头表面周围都有加热元件。所以它可以，你知道，就是。我们的第一代硬件上也有这个。所以它会自动加热，以防止积雪和结冰。而且摄像头所在的位置非常不容易积灰或结冰。

所以这组 3 个摄像头是。所有 3 个福特摄像头都位于后视镜前方，并处于车辆雨刷的刮扫区域内。然后，朝侧面看的摄像头位于 B 柱内。它们在 B 柱中的位置比较靠上。

就是位于……之间的那种立柱

前排和后排乘客之间。然后是后部。还有左后摄像头和右后摄像头，它们嵌在侧面转向灯中。

所以它们避开了雨雪之类的流动路径。然后后置摄像头位于尾门上，在车牌正上方，不过有点藏在下面。所以你其实得仔细看才能看到这些摄像头，但它们都位于很难积聚泥土、冰或雪的位置。

如果其中一个被遮挡，

系统实际上可以判断其透明度有所降低，并提醒车内乘员清理一下摄像头。

你好。你提到这与 Model 3 和第 2 部分的发布有关，我只是想再稍微多关注一下这方面，以及它究竟如何应用于 Model 3。这些都会成为 Model 3 的标配吗？是否会像现在一样有升级方式，也就是软件标配中包含哪些、哪些需要付费？

是的，从现在起，Tesla 制造的所有车辆都将标配全自动驾驶硬件套件。它还会继续改进，而这些改进将影响……事故发生的概率。但我们相当有信心，目前这套硬件的表现平均至少是人类的 2 倍。

那么，从长远来看，我们希望努力实现 10 倍的

提升，因此这将需要持续

改进硬件。

但即便如此，像 2 倍这样的提升

也会相当不可思议。

我是说，如果这全面应用于汽车，应用于全世界所有车辆，死亡人数就会从 120 万降至 60 万。这将挽救许多生命，更不用说所有严重伤害以及其他那些未被纳入死亡统计的情况了。

所以我认为，努力将安全性从 2 倍提高到 10 倍，仍然非常有价值。

但我应该明确说明，从现在起，每一个进行生产的卡特尔都将具备全自动驾驶硬件能力，包括 Model 3。

嘿，埃文，这里有一个问题总是浮现在脑海中：你们怎么知道，也就是你们在每个阶段进行什么样的测试，才能确信系统是安全的，或者运行得足够好，从而有信心把它装进去？因为这很难想出来。我是说，也许并不难。你们在进行什么样的测试，以确保这些系统运行良好？

是的，其实。这并没有那么难。我是说，很显然，我们一开始是从……

通过 Tesla 进行测试，你知道，由我们的 Tesla 质量保证团队进行测试。所以我们最初会让测试工程师在封闭赛道上进行测试。然后我们会将范围扩大到一组非常有限的

算是 Alpha 用户，包括我。所以我通常会在我们的质量保证之后大约 1 天拿到最新的软件更新

团队拿到它，

而且我本人正在测试

车辆的每一个要素。

然后它会进入我们所说的抢先体验计划，该计划大约有

分布在世界各地的1,000名客户，他们精通技术，而且不介意

或希望使用这种早期软件，并且了解潜在问题

因为整个地球上有太多不同的环境，

质量保证团队根本不可能覆盖所有环境。

如果情况看起来不错，那么我们会

首先以影子模式向整个车队推送。所谓影子模式，是指车辆实际上不会采取任何行动，但会记录它何时会采取行动、何时不会采取行动。然后你可以把它与某些情况进行比较，比如说有人发生了事故。但如果你查看车辆日志，并说，嗯，如果车辆

当时处于自动驾驶模式，那起事故本可以避免，那么，好吧，这显然是一个正面因素。

你也可以说，好吧，车辆会。车辆计算机会做出某种会导致事故的操作。在这种情况下，它当时处于影子模式，那么显然这是一个需要纠正的问题。然后是统计数据，我是说，这些被称为假阳性和假阴性。

随着时间推移，你会收集足够多的这类数据，直到显然形成一个具有统计显著性的样本集。而当毫无疑问地证明启用某项自动驾驶功能能够提高安全性时，我们才允许它实际采取行动。

在此之前，我们不允许

计算机采取行动。

你好，感谢你抽出时间。只是想确认几个问题。你说的是新硬件2达到4级还是5级？另外请明确一下，关于硬件1，目前会不会停用该技术的任何部分，还是会继续照常运行？谢谢。

硬件2将具备5级自动驾驶能力。换句话说，该硬件能够实现最高级别的自动驾驶，而硬件1，我认为，随着我们改进用于操控车辆的软件，它也会继续改进。我的意思是，早在7.0版本时，它就

毫无疑问比

人工驾驶的车辆更安全。而到了800，这一点又有了进一步改善。

所以，显然，

关闭某种正在防止事故发生的东西会很疯狂。

嘿，埃隆，快速问一下，你描述的40倍超级计算机是独立的吗？它

目前是否必须与所有其他车辆系统技术分开，还是与其他车辆技术集成在一起？

不，这是个好问题。它与车辆的其余部分相隔离。所以信息娱乐系统，也就是娱乐仪表板、中控区域，任何与娱乐或网页浏览有关的部分，都与车辆控制计算机相隔离。

好的，很好。

就这些。

谢谢。

是的，肯定不想因为访问了错误的网站而成为那场车祸。

谢谢。

感谢各位参加 Tesla 新闻发布会。

The basic news is that all Tesla vehicles exiting the factory have the hardware necessary for level five autonomy. So that's in terms of the cameras and compute power.

Every car we make on the order of 2000 cars a week are shipping now with level 5 meaning hardware capable

of full self driving or driverless capability.

So it'll take us some time

into the future to complete validation of the software and obviously get the required regulatory approval. But the important thing is that the

foundation is laid for the cars to

be fully autonomous at a safety level we believe to be at least twice that of a person may be better.

So I think that's probably unexpected by most that it's happening right now.

So yeah, so we're pretty excited about that. And that's also essentially part two of the Model 3 announcement, which is that Model 3 will also have the hardware necessary for full autonomy. In fact, all cars that Tesla makes

from here on out will have the

hardware needed to be fully autonomous or driverless.

That's where things are. And this is all Tesla vision software. So we're not using any third party software or anything for, for the vision processing. This is a Tesla developed neural net.

And yeah, although it's somewhat hardware independent,

it could be, we could actually run this on

Nvidia AMD or Intel.

We did pick the Nvidia Titan GPU

as the main chip for the neural net, but it was a pretty tight

call between, particularly between AMD and Nvidia. But ultimately we thought Nvidia had the better hardware.

So that's kind of where I think pretty huge news.

Now with that we can go into questions.

Okay, our first question is from Nikki Gordon Bloomfield from Transport Evolve llc. Your microphone is now open.

Good evening, Elon. I have got a, I've got two questions. I don't know if you're in the place to answer two, so please pick one. If not, Autopilot's proven very reliable thus far with minimal problems with its safety. But I'm asking if Tesla will be offering indemnity against crashes involving fully autonomous autopilot mode in much the same way as Volvo is promising to do when level 5 autonomy is activated on your vehicle?

No, I think that would be up to the individual's insurance. I mean, just like it's, you know, if it's something endemic to our design. So certainly we would take responsibility for that. But I think once you use autonomous cars, much like an elevator in a building. Does Otis take responsibility for all elevators around the world? No, they don't. So.

What really matters at the End of the day here is what is the absolute safety?

One of the things I should mention that frankly has been quite disturbing to me is the degree of media coverage of autopilot crashes, which are basically almost none relative to the paucity of media coverage of the 1.2 million people that die every year in manual crashes, is something that I think does not reflect well upon the media. It really doesn't. Because. And really, you need to think carefully about this, because if in writing some article that's negative, you effectively dissuade people from using an autonomous vehicle, you're killing people.

Next question.

Our next question comes from Jack Stewart from Wired. Your line is now open. Hey, good evening.

I was wondering, Elon, if you could just spell out how you see this

rollout of autonomous features happening once these cars start to become enabled? The press release says that new features will become available bit by bit. How do you see that? Are we going to see full city driving straight away or a more limited feature set?

Yeah, since this is a new platform, although we've been spending, we've spent more than a year in testing, the feature set initially will be disabled, at least for the first few months. The hardware, the cars with Hardware 2, what we call Hardware 2, which is sort of the full autonomy suite, will actually have fewer features than cars with hardware one. So we expect to reach feature parity following field validation of Hardware 2, probably in December, so maybe two or three months from now.

So if there's two or three months, actually a Hardware 1 car will be better than a Hardware 2 car. And then

every approximately every two or

three months thereafter is when we expect to release significant improvements in autonomous capability.

And our goal is, I feel pretty good about this goal, is that we'll be able to do a demonstration drive of full autonomy all the way from LA to New York.

So basically from a home in LA

to, let's say, dropping you off in Times Square in New York and having the car go and park itself by the end of next year without the need for a single touch, including the charger.

What is happening to Autopilot in the vehicles with the new hardware? Could you speak specifically to that?

What is happening to Autopilot in the vehicles with the new hardware? Well, the

new hardware is what will enable self driving, which is different from autopilot.

Autopilot is a term that's been in use for more than half a century as an assistance, a flying assistant in aircraft for pilots.

That's why we chose to use does not represent self driving any more than

autopilot aircraft makes an aircraft self flying.

So I'm just trying to understand. So the Autopilot which has been associated with your advanced driver assistance technology here up until now, is that what happens to that system? What happens to that name, if you will, as you go forward with this new hardware? It's just not clear from the release and I just want to understand how you're handling that.

Yeah, there's

basically there'll be two, I

mean there'll be two options for buying a cart.

One is what we call Enhanced Autopilot, which is kind of quite similar to

what the Autopilot that we have been offering, except that it's got redundant forward cameras and it has a left rear

camera and a right rear camera as well as significantly improved ultrasonic sonar

and much more computing power. The net effect of Enhanced Autopilot is you should be able to go from, from freeway on ramp to exit, as

well as transitioning between multiple freeways and

passing and maneuvering around other cars without touching anything.

With Enhanced Autopilot then there's sort of

the full self driving capability which will take care of the much more complex situations in urban environments.

And so there'll be two options that people can pick in buying the car. Either basically an improved version of Autopilot or full self driving. So one's kind of four cameras, the other's eight cameras.

Yeah, that's sort of what will happen now. The hardware one Autopilot vehicles will continue to improve with improved software as the fleet learning, as we accumulate more miles, it will continue to get better and better, but it is limited by the fundamental hardware that's on the hardware.

One cars in that there's only one Ford camera, they're ultrasonics, but they have

half the range in resolution, has the radar.

And so within the context of that

sensor suite, the hardware worn cars will continue to get better over time.

Thank you. Our next question is from David Baker from the San Francisco Chronicle. Your line is now open.

Thanks for doing the call. I actually wanted to follow up on exactly what you were talking there. Could you walk us through the differences between the hardware suite on the existing Autopilot cars that are out there and this new suite? In other words, you were talking about more cameras, better sensor range. Could you just detail each of those changes?

Yeah, absolutely.

So we go from one camera to

eight cameras and three of which are forward cameras. So we have redundancy in the forward

camera sort of looking forward and we

have 360 degree coverage, visual coverage around the car.

So that's A pretty big upgrade.

The compute power increases by a factor of 40, like so 4.0x increase in compute power. So it's a gigantic increase in computing power.

In fact, the computer will be capable

of 12 trillion operations per second, basically

a supercomputer in a car.

And then the ultrasonic sensors are next

generation ultrasonic sonar, which have about twice

the range in resolution of the current sonar. And that's also 360 degrees. I mean there are also some other minor things like the GPS is more accurate and provides more frequent updates. And there's some other minor sensor improvements on the imu, the initial measurement unit and a few other areas.

Hey Elon, I'm just wondering what are your thoughts on the regulations around this? Is it going to be limited based on, you know, oh, in Texas you can do this and in California you can do that and it's going to be different? Or are you just going to say everyone everywhere can do it? What are you thinking on the regulatory side?

Well, it's not up to us, it's up to the regulators. Hopefully in the US things do not

become balkanized so that it's different in

every state because I think that would

be detrimental to the US consumer and

I think in the EU it's like fairly confident it will be a uniform standard.

So it's really a question of what,

you know, what does the public think is appropriate, what do regulators think is

appropriate and gathering enough data because the

system will always be operating in shadow

mode so we can gather a large volume of statistical data to show the

false positives and false negatives.

When would the computer have acted and would that have prevented an accident? Or if the computer would have acted and that would have produced an accident. But that's operating in shadow mode, so we can say when it would have incorrectly acted or not acted.

Compare that to what should have been done in the real world. And at the point at which it

is statistically significant result that shows

a

material improvement over the accident rate for manually driven cars.

I think at that point regulators are likely to be comfortable approving it. But that approval process really could be radically different from one part of the world to another. It's not something within our control.

One last just follow up to that is will, you know, you talk about rolling out new features every two to three months. Will the regulatory environment in the US or the EU or wherever hold you back from rolling out those improvements? Or will you roll them out and then, you know, ask forgiveness instead of permission? If you think that it's that it's safer.

Well, I mean we've always rolled out our autonomous functionality within the regulatory framework of any given country.

So this is not a, you know, pleading forgiveness or permission.

It's we, you know, we work,

we

look carefully at the regulations and make sure that what we do is in line with those.

Yeah, we can't do anything other than

that because it would be against the law.

Hi Elon, thanks for taking this.

I'm wondering is there now that you're moving on to hardware two and you're going to keep working with hardware one, is there a point at which you can, that you'll reach a maximum of potential for hardware one and then you need to, you know, an owner will need to upgrade to the next to a newer car to get the full self driving capability you can offer.

Self driving will definitely need hardware too, which is what we're currently, we're shipping as of this week. Just because you don't, you just can't do this with one camera. You need eight cameras to be self driving. You also need a lot more computing power to run the vision. Basically the vision AI.

So upgrading hardware on cars is not, it's not realistic.

It would be like giving the car

a spinal cord transplant wise.

So even if possible, best to be avoided. But I should say that people that have Hardware one cars should bear in mind that their car will actually have more functionality than a hardware 2 car. At least until December, maybe later and it's probably sometime next year before the hardware 2 functionality starts to exceed hardware 1. And there's also a higher cost for the autonomous functionality. So it's not, you know, there is a higher price to pay like for the full self driving.

To activate full self driving hardware suite, it's $8,000. So that's as compared to the hardware one autopilot which is $3,000. So meaningful price difference. And the car, the hardware one actually has more functionality at least for the next several months. I wish there was some other way to do it. This is not, it's just I don't know how to go into old cars and install another seven cameras and a new wiring harness.

So there's no, it's just.

I wish we could do it some other way.

But there isn't some other. There isn't.

Hi, I was wondering if you plan on making this a kit that other carmakers could use.

I think this is very hard to turn into a kit because it requires

a tight integration of software, sensors and

computing power as well as the ability to do large scale over the Air updates.

It's not just some kit that you can add onto a car. I mean, one thing that we've also done with our system is that if you. And we'll be releasing some videos hopefully later tonight and tomorrow morning demonstrating the car in operation.

But you actually can't, unless you look closely, you can't even tell that a

car is hardware one or hardware two because we've been so careful about

each

of the cameras being part of the frame of the car, such that nothing's sticking out.

This in no way makes the car ugly.

So there are no weird protuberances.

It's all incredibly subtle.

Put a lot of effort into not

affecting the beauty of the car. The car is as beautiful with Hardware 2 as it is with Hardware 1.

And there's just no way to turn

that into a kit and put it in some other car.

It's not realistic.

Yeah, I was interested in your thoughts, if you could expand your thoughts a little bit about the safety issue. How statistics aren't collected on how safe a car is, only how dangerous it is. So how do you communicate the idea that your semi autonomous cars are in fact safer?

It's not that difficult. I mean, simply say, you know, how

many miles have been driven and you

know, and then how many deaths, serious injuries, meaning like an unrecoverable injury or minor accidents.

And obviously there are many more sort of minor accidents.

And serious accidents than there are fatalities. So that provides a much richer statistical sample set for comparing the relative safety

of autonomy versus not autonomy. And we see consistently significantly better results

with autonomy than without.

And that just gets better over time as the system is further refined. Hi Elon, I have a quick question

about inclement weather and how has hardware

2 improved over hardware 1? Well, inclement weather, you mean does it work in snow and stuff like that or what do you mean by inclement weather? Sorry, Anytime say you know there's potential

for condensation or precipitation on the cameras

tends to be a problem.

How have you dealt with that?

Oh yeah. So there are heater elements around on all of the camera surfaces. So it can, you know, just. We have that on hardware one as well. So just automatically heats it up to prevent snow and ice. And then the positioning of the cameras is in a location where there's very unlikely to be dust or ice accumulation.

So the triple cameras are. All three Ford cameras are in front of the rear view mirror and in the wipe zone of the car. And then the sideways looking cameras are in the B pillar. They're sort of high up in the B pillar.

That's the sort of pillar in between

the front and rear passengers. And then there's the rear. There's left rear and the right rear camera, and those are nested in the side repeaters.

And so they're out of the flow of rain and snow and whatnot. And then the rear camera is on liftgate, just above the license plate, but sort of tucked under. So you actually have to look carefully to see any of the cameras, but they're all in places where it's quite difficult for dirt or ice or snow to accumulate.

If one of them is blocked, the

system can actually tell that it has some opacity reduction and can alert the occupants in the car to just clean up the camera.

Hi. You mentioned that this is related to the Model 3 and the Part 2 unveiling, and I just kind of wanted to focus a little bit on more of that and how it really applies to Model 3. And is this all going to come standard on Model 3? Will there be a way to upgrade like there is now, from what is included as far as the software standard and what's paid?

Yeah, the full Autonomy hardware suite will be standard on all vehicles that Tesla makes from here on out. It will continue to make improvements, and those improvements will affect the probability of. Of an accident. But we feel pretty confident that the current hardware suite will be at least twice as good as people, on average.

Now, in the long term, we want to try to get to a 10x

improvement, and so that will require ongoing

refinement of the hardware.

But still, something like a 2x improvement

would be pretty incredible.

I mean, if that was applied across the board to cars, to all vehicles in the world, you go from 1.2 million deaths to 600,000 deaths. That would be a lot of lives saved, and not to mention all the serious injuries and other things that happen that aren't in the fatality statistics.

So I think there's still a lot of merit in trying to go from twice as safe to 10 times as safe.

But I should be clear that every cartel that produces from now here on out will have the full Autonomy hardware capability, including Model three.

Hey, Avon, one of the questions that kind of keeps coming to mind here is how do you know what kind of testing are you doing to be confident at each stage that the system is safe or is functioning enough to be confident to put it in there? Because it's difficult to come up with. I mean, maybe it's not difficult. What sort of testing are you doing to Ensure that these systems are working well.

Yeah, it's actually. It's not all that difficult. I mean, we start off, obviously with.

With testing with Tesla, you know, with our Tesla QA team. So we're testing initially on a closed track with our test engineers. Then we'll broaden it to a very limited set of

kind of alpha users, including me. So I usually have the latest software update about a day after our QA

team has it,

and I'm personally testing

every element of the car.

Then it goes to what we call our early access program, which is about

1,000 customers distributed around the world who are technically savvy and don't mind

or want to use sort of the early software and are cognizant of potential issues

because there's so many different environments throughout

Earth that it's just not possible for a QA team to cover them all.

If that's looking good, then we will

roll it out initially in shadow mode to the whole fleet. By shadow mode, it means that the car is not actually taking any action, but it is registering when it would take an action and when it would not take an action. And then you can compare that to cases where, say, that somebody has an accident. But if you look at the vehicle logs and say, well, if the car

had been in autonomous mode, that accident would have been avoided, then, okay, that's obviously a plus.

You can also say, okay, the car would. The car computer would have done something that would have resulted in an accident. In that case, it was in shadow mode, then obviously that's an issue that needs to be corrected. And then statistics, I mean, these are called false positives and false negatives.

You gather enough of those over time until there's clearly a statistically significant sample set. And at the point at which it's unequivocal that turning on an autonomy feature would improve safety, that is the point at which we allow to actually take action.

Before that we do not allow the

computer to take action.

Hi, thanks for making time. Just a couple clarifications on some things. Are you talking about the new hardware 2 being level 4 or level 5? And just to be clear, on hardware 1, is there going to be any disabling of that technology at this point, or does this continue to run as is? Thank you.

It will be the hardware 2 is capable of level 5 autonomy. In other words, the hardware is capable of the highest level of Autonomy, and Hardware1, I think, will continue to improve as we improve the software that operates the car. I mean, already with 7.0, it was

unequivocally safer than

manually driven cars. And with 800, that has improved even more.

So it would obviously be crazy to

turn off something that is preventing accidents.

Hey, Elon, just quickly, the 40x supercomputer you described, is it separate, does it

have to be separate from all other vehicle system technology at this point, or is it integrated with other vehicle technologies?

No, it's a good question. It is isolated from the rest of the vehicle. So the kind of infotainment, sort of entertainment instrument panel, the center cluster, anything to do with entertainment or web browsing is isolated from the vehicle control computer.

Okay, great.

That's all.

Thanks.

Yeah, definitely don't want to be the car crash as a result of going to the wrong website.

Thanks.

Thank you everyone for attending the Tesla press conference.