What Elon Just Said: What It Really Means ⚡️

Welcome to Electrified, it's your host Dylan Loomis. Quick shout out to my newest patron, Eric G. Thank you for choosing to support the channel.

欢迎来到《电子深度解析》，我是主持人迪伦·卢米斯。快速向最新的赞助人埃里克·G.致以问候。感谢您选择支持本频道。

Today's video is going to be a bit different. We're going to rip through some news items and then I want to take a few minutes diving into what Elon just said when it comes to FSD and the last puzzle piece. This is arguably the most important project Tesla is working on right now, so certainly worth our time.

今天的视频会有一些不同。我们将迅速浏览一些新闻，然后我想用几分钟的时间来探讨埃隆刚刚谈到的关于全自动驾驶（FSD）和最后的拼图。这可以说是特斯拉目前最重要的项目，所以绝对值得我们花时间去看。

A few weeks ago, we had talked about Tesla's backup switch, which Elon and the team confirmed was going to replace the old gateway. And now we have Tesla energy sharing a video that shows with the new way, the new power wall with the backup switch, it only takes 20 minutes to install compared to the old version where it took over four hours. There weren't many headlines on this at the time when it first came out, but clearly another example of Tesla continuing to innovate and in this time, that's a huge change on the time for install. During the math, excluding travel time, Tesla can now install 12 power walls with the backup switch. And the same time it took them to install one with the old gateway.

几周前，我们曾讨论过特斯拉的备用开关，埃隆和团队确认将用它来替换旧的网关。现在，特斯拉能源分享了一个视频，展示了新的路线，新的储能墙配备备用开关，只需要20分钟安装，而旧版本则需要超过四小时。虽然一开始并没有太多的头条关注这个消息，但显然这是特斯拉继续创新的又一个例子，对于安装时间来说，这是一个巨大的改变。在计算过程中，不计算行程时间，特斯拉现在能够安装12个备用开关的储能墙，而旧网关则只能完成一个的安装时间。

We have Jared sharing an image showing us some new Tesla supercharger stalls with the magic dock, this one in Fort Worth, Texas. Earlier this year, we saw 10 plus in the New York Buffalo area, maybe a few in California. Now Texas added to the list after a bit of a gap as far as we know. If you're in the area, it's the East Chase Parkway location and it's 45 cents per kilowatt hour for non Teslas without a membership, 33 cents per kilowatt hour for Tesla drivers or non Teslas with a membership. Many people have asked why is Tesla still rolling this out with so many people announcing the next? Well, don't forget there's still going to be plenty of CCS cars on the road for years to come and federal funding could be another reason.

我们有贾里德分享了一张图片，展示了位于得克萨斯州沃思堡的一家新特斯拉超级充电站，其中带有魔术对接设备。今年早些时候，我们看到了纽约布法罗地区有10多个这样的充电站，加利福尼亚可能也有几个。现在，得克萨斯也在这个名单上加入了，在我们所知道的情况下，他们中间有一段时间没有新的充电站了。如果你在这个地区，该充电站位于东车道公园路位置，非特斯拉车主没有会员资格的话，每千瓦时要收取45美分电费，特斯拉车主或者有会员资格的非特斯拉车主则是每千瓦时33美分。很多人问为什么特斯拉在这么多其他公司宣布新产品的时候还在推出新的充电站呢？嗯，不要忘记还会有很多CCS电动车在路上行驶很多年，而且联邦资金可能也是另一个原因。

Lucas shared some images of Tesla doing FSD data collection in Switzerland, of course, preparing for FSD.

Lucas分享了一些特斯拉在瑞士进行FSD数据收集的图像，当然，为的是为FSD做准备。

We have one of Tesla's lithium hydroxide suppliers, the Yohua Group that has revised their battery grade supply agreement with Tesla from 2020. Now it's going to be extended from 2025 to the end of 2030. If this report is correct, the new long term contract could reach up to 81 billion yuan in value or about $11 billion. It looks like originally the deal was going to be 2021 to 2025 and less than a billion dollars of total value. So again, if this is accurate, this is a huge extension and a long term deal, just another example of suppliers kind of itching to work with Tesla and Tesla needing massive supply.

我们拥有特斯拉的锂氢氧化物供应商之一，即永华集团，他们已经修订了与特斯拉的电池级供应协议，从2020年开始延长至2030年底。如果这份报告属实，新的长期合同的价值可能达到810亿元人民币，约合110亿美元。最初的交易计划是2021年至2025年，总价值不到10亿美元。所以，如果这是真实的，这是一项巨大的延期和长期协议，再次证明供应商急于与特斯拉合作，而特斯拉需要大规模的供应。

The talks between Tesla and India are still ongoing but Tesla's India arm motor and energy has actually leased some office space about 6000 square feet in Pune's Vietnam Nagar location. As I've said before, whether Gigat India is next or not, it seems very likely that at some point it's definitely coming.

特斯拉与印度之间的谈判仍在进行中，但特斯拉印度分公司——摩托和能源部门已经在普纳的越南纳加地区租下了约6000平方英尺的办公空间。正如我之前所说，无论印度Gigat项目是不是下一个，似乎有很大可能性在某个时候它终将到来。

To any of you that might be concerned about Elon's work habits or commitment to Tesla, he just said he was buried in Tesla work all day.

对于可能对埃隆的工作习惯或对特斯拉的承诺有所担心的人，他刚刚表示自己整天都在忙于特斯拉工作。

The UAW president just said he sees merit in a 32 hour work week, not 60 to 80 hours a week, we shouldn't have to spend 7 days a week, 12 hours a day living in factories. And he said four Giamens to Lantis can easily afford big changes in the upcoming contract that help workers, saying in the first half of this year, those three combined made $21 billion in North America alone, saying in the first half of this year, those three automakers made $21 billion combined. He also shared a list of demand some of the UAW employees are asking for, eliminating wage tiers, substantial wage increases, restoration of COLA increases, more paid time off and increased benefits to current retirees among others. As we all should know, sure, maybe the big three automakers can afford these changes now, but the question becomes for how long, how long can they keep up these profits and when is the ICE train gravy train coming to an end?

美国汽车工人联合会（UAW）主席刚刚表示，他认为每周工作32小时是有道理的，而不是每周60到80小时，我们不应该每天在工厂里过着7天12小时的生活。他还表示，吉粤曼斯、兰蒂斯等四家汽车制造商在即将到来的合同中可以轻松承担起对工人有利的重大改变，因为今年上半年，这三个公司在北美就赚了210亿美元。他还列举了一些工人的要求清单，包括取消工资分层制度、大幅提高工资、恢复消费者物价指数适应津贴增长、增加带薪休假和提高现有退休人员福利等。我们都知道，当然，也许这三家大型汽车制造商现在能够承担这些改变，但问题是能够维持多久，他们能够保持这些利润又将何时终结呢？

And it looks like those 12 former Tesla employees that left to work at Rivian will not be of avoiding trial for their breach of Tesla's confidentiality agreements. A ruling still needs to be finalized, but Tesla was arguing Rivian had directly hired at least 70 of its former employees, some of whom were caught red-handed stealing core technology for next generation batteries. It looks like this is headed to a trial.

看起来，那12名之前离开特斯拉去在Rivian工作的员工将无法逃避由于违反特斯拉的保密协议而可能面临的审判。虽然仍需要最终裁决，但特斯拉主张Rivian直接雇佣了至少70名前员工，其中一些人当场被抓到窃取下一代电池核心技术。看起来这将进入审判阶段。

On the Tesla configurator in some locations in Europe, the Model 3 delivery dates have been pushed back in some cases for the performance all the way until January 2024, potentially another crumb to the pile of Highland Clues.

在欧洲某些地区的特斯拉配置器上，Model 3的交付日期在某些情况下一直被推迟到2024年1月，这可能是与《高地谜团》有关的另一个线索。

Subaru changed its EV plans and now plans to build EVs in the United States around 2027 and to sell about 400,000 EVs in the US by 2028. Subaru will also expand its EV lineup to eight models from an earlier planned four. The new CEO said the US market is shifting to electrification at a rapid clip and the situation has changed considerably in just the past few months.

斯巴鲁改变了其电动车计划，现在计划在2027年左右在美国建设电动车，并计划到2028年在美国销售约40万辆电动车。斯巴鲁还将扩大其电动车产品线，从原计划的四款车型增加到八款车型。新任首席执行官表示，美国市场正在以惊人的速度转向电动化，而且在过去几个月中情况已经发生了很大变化。

We also got the best look yet at the upcoming Porsche Macon without any camo. This car is supposed to launch sometime next year and yes that includes the US market. We don't yet have official specs but there are rumors of it starting around $80,000 and having over 300 miles of range. And Porsche has come out and said this car is supposed to be highly digital so we'll see what that translates to.

我们还以前所未有的清晰度看到了没有伪装的即将推出的保时捷Macon。这辆车计划在明年推出，并且会包括美国市场。尽管我们还没有官方规格，但有传言称其起价约为8万美元，并具有超过300英里的续航里程。保时捷还表示这辆车将是高度数字化的，所以我们拭目以待，看看它会带来什么样的变化。

And here we have it, this is where we need to spend a few minutes explaining what this means and what the reality might actually be. Elon said vehicle control is the final piece of the Tesla FSD AI puzzle. That'll drop over 300,000 lines of C++ control code by around two orders of magnitude. It's training as I write this. Our progress is currently training compute constrained, not engineer constrained. So what's going on here?

在这里，我们需要花几分钟解释这意味着什么，以及现实可能是什么样子的。埃隆说车辆控制是特斯拉全自动驾驶人工智能拼图的最后一块。这将减少300,000行C++控制代码，大约降低两个数量级。正在我写这篇文章的时候进行训练。我们的进展目前是受计算能力限制，而不是工程师的限制。那么这里到底发生了什么呢？

So I have to temper my expectations here given Elon's track record. He said many times the FSD team will make a breakthrough and things look great but a few months down the line they hit a local maximum and need to adjust the plan. When it comes to solving for autonomy think about this analogy. Your goal is to climb the tallest mountain in the world but you can only see a mile in front of yourself and you have no pre-existing knowledge of the location. All you know is that if you're going uphill it's a good thing and if you're going downhill that's a bad thing.

所以考虑到埃隆过去的纪录，我必须调整自己的期望。他曾多次说全自动驾驶团队会取得突破，形势看起来不错，但几个月后他们达到了一个局部最优点，需要调整计划。当涉及到实现自动驾驶时，可以用以下比喻来思考。你的目标是攀登世界上最高的山，但你只能看到前方一英里的路程，而且对于目的地的位置没有任何预先的知识。你所知道的只有，如果你往上走，那是件好事；如果你往下走，那是件坏事。

When you get to the top of the first hill you can then see that maybe there's a bigger mountain that you can now see what you'll call mountain A. So even though going downhill is bad you have to go down the hill towards the taller mountain and then eventually you start going back uphill. And then maybe once you get to the peak of mountain A you see mountains B and C but one's way to the left and one's way to the right and you have to make a decision which one do you pursue. So maybe you pick going right you end up at the top of mountain C.

当你到达第一个山坡的顶部时，你可能会看到可能还有一个更大的山，你现在称之为山A。所以即使下坡很糟糕，你仍然需要下山，朝着更高的山的方向前进，然后最终你会开始再次上山。然后当你到达山A的山顶时，你会看到山B和山C，但一个在左边，一个在右边，你必须做出决定选择哪个追求。所以也许你选择向右走，最终你会到达山C的顶部。

That might mean things work great in California which we've seen when it comes to self-driving but it may not be so great in another place like Michigan. But of course it's not that simple because you can never search the entire problem space because there's no actual defined goal in self-driving. And the complexity of the problem increases exponentially with the number of variables in the world and when it comes to driving there are plenty.

这可能意味着在加利福尼亚州，我们看到自动驾驶的运行表现非常好，但在密歇根州等其他地方可能不太理想。但当然，情况并非如此简单，因为在自动驾驶中，你永远无法搜索整个问题空间，因为没有实际定义的目标。而且，随着世界上变量的数量增加，问题的复杂性呈指数增长，而在驾驶方面，存在大量变量。

So the best thing Tesla can do is find the fastest computer it can dojo and check as many mountains as it can which is basically training the neural net. And then you hope that your output FSD does a slightly better job than a human driver. Here are just a few examples of Elon talking about local maximums for Tesla's FSD progress in the past 2020 autopilot was trapped in a local maximum labeling single camera images uncorrelated in time now it's not beta 10 or maybe 10.1 going to pure vision sets us back initially going down the hill vision plus course radar had us trapped in a local maximum like a level cap pure vision requires fairly advanced real world AI but that's how our whole road system is designed to work neural networks with vision when it comes to high resolution maps Elon said we barked up that tree for way too long gives a false sense of victory of being close that should sound familiar a tantalizing local maximum but reality is just too messy and weird our new system is capable of driving in locations we've never seen once.

因此，特斯拉能做的最好的事情就是找到最快的计算机，来训练神经网络，尽可能多地检查山脉。然后，希望你的输出全自动驾驶(FSD)比人类驾驶员做得更好一些。以下是一些埃隆关于特斯拉FSD在过去的2020年的局部极值的讨论。在过去，自动驾驶一直陷入了局部极值，标记了不相关的单摄像头图像，现在不是beta10或者10.1。从一开始纯视觉对我们来说是个挑战，视觉加上雷达束缚我们在一个局部极值中，纯视觉需要相当先进的现实世界人工智能，但这就是我们整个道路系统的设计原理，神经网络和视觉一起工作。当谈到高分辨率地图时，埃隆说我们在这个方向上努力了太久，给人一种接近成功的错误感觉。这应该听起来很熟悉，是一个令人心动的局部极值，但现实太杂乱和奇怪。我们的新系统能够在我们从未见过的地点驾驶。

So keeping all of that context in mind what does the rest of this mean basically in this realm you don't know it's solved until well it's solved yes it may be training right now like Elon said making great progress but until it hits another local maximum and then Tesla has to figure out why so you don't really know you're in a local maximum unless you've explored the whole problem space improved you're at a local maximum aka seeing a new higher mountain eventually though one of these times Tesla will find the global maximum the highest peak in the world per se but is that going to be this time around we'll see there are people out there that argue Tesla needs radar or LiDAR to climb the global peak but personally i'm not one of them you may also hear people talking about end to end the simplest explanation is to think of three pillars or phases of full self driving perception the car perceiving the world around it planning deciding what to do and when and control actually changing lanes breaking etc if you're an office fan just remember the ppc party planning committee or in this case perception planning control

在这个背景下，就基本上可以解释剩下的意思了。在这个领域里，直到问题得到解决之前，你并不知道它是否解决了。是的，现在可能正在进行训练，就像埃隆所说的，取得了巨大的进展，但直到它达到另一个局部最大值，并且特斯拉必须弄清楚原因，你才能知道你是否处于局部最大值。除非你已经探索了整个问题空间并提高了你的局部最大值，也就是说，看到了新的更高的山峰，否则你不会真正知道自己是否处于局部最大值。然而，总有一天，特斯拉会找到全球最大值，也就是所谓的世界最高峰，但这次会是这个时候吗，我们拭目以待。有些人认为特斯拉需要雷达或激光雷达来攀登全球最高峰，但是就个人而言，我不是其中之一。你可能还会听到人们谈论端到端，最简单的解释就是将完全自动驾驶分为三个支柱或阶段：感知（车辆感知周围世界）、规划（决定该做什么以及何时做）和控制（实际上改变车道、刹车等）。如果你是《办公室》的粉丝，只需要记住 PCC（党务策划委员会），或者在这种情况下，感知、规划、控制。

right now Tesla is in the process of transitioning from. code written by engineers for the control pillar but they're working on using neural nets alongside of the code to eventually replace the manual code so from perception to control it'll actually be controlled by neural networks that can essentially train themselves as you can imagine though to do the math and explore all of the scenarios that exist hiking to different mountains it takes an absurd amount of compute power what Elon said makes it seem like hiring more engineers would not help expedite the process nearly as much as more compute power would hence project dojo

现在特斯拉正在过渡中，从工程师编写的控制代码转变为使用神经网络与代码结合的方式，并最终替代手动代码。从感知到控制，实际上将由能够自我训练的神经网络来进行控制。不过，可以想象，为了进行数学计算并探索所有可能的情景，比如攀登不同的山，需要大量的计算能力。埃隆说的意思是，雇佣更多工程师并不能像增加计算能力那样显著加快这个过程，因此需要“项目道场”。

the big question remains with vehicle control dropping from 300 000 lines of c++ code down to around 3000 that's a significant drop in manual code but the trade off becomes more training compute power is then needed to process video data that's fed through the neural nets controlling that to control or the third pillar i'm not a machine learning engineer so i can't really say what the remaining 3000 lines of code would be used for but two orders of magnitude fewer lines of code could of course free up some engineers to work on other things and it's a huge step toward tesla getting to end to end with what we believe will be fsd beta 12 when it's no longer a beta then from there with most of the fsd stack running neural nets and dojo continually scaling and doing more work things could most certainly accelerate paving away for huge progress throughout 2024

一个重要的问题仍然存在，那就是车辆控制从30万行的C++代码降低到大约3000行，这是手动代码的显著减少，但在此过程中需要更多的培训和计算能力来处理通过神经网络控制的视频数据。对于我这种非机器学习工程师来说，我无法确定剩下的3000行代码将用于什么，但是由于代码行数减少了两个数量级，当然可以释放一些工程师来从事其他工作，这对于特斯拉实现我们所认为的完整自动驾驶测试版本12是一个巨大的进步，当它不再是测试版时，随着大部分自动驾驶系统中的神经网络和Dojo的不断扩展和工作，事情肯定会加快，为2024年取得巨大进展铺平道路。

this all becomes confusing though because the car doesn't need to be perfect it just needs to be safer than a human not all drivers are max for stapping out there many are more like michael scott so the real question is which mountain peak will be good enough for regulators and that's good enough for tesla to take responsibility and shift to a level four or five sae framework when elon hinted that they figured out fsd a few weeks back perhaps this is what he meant tesla knows what to do now they just need to do it so this is certainly a huge step forward and if you just imagine 300 000 lines of code you can understand why it's great progress but the question remains will they hit another local maximum in another few months or will tesla find a global peak that's good enough and then we're off to the races either way we'll be here watching it all unfold you can find me on x at dylan lumus 22 hope you guys have a wonderful day please like the video if you did and a huge thank you to all of my patreon supporters

尽管一切变得混乱不过这没关系，因为汽车不需要完美，它只需要比人类驾驶员更安全。并非所有驾驶员都是能对待自己的工作严谨如同马克斯，很多人更像是迈克尔·斯科特。所以真正的问题是，哪个山峰足够满足监管机构的要求，并且足够满足特斯拉负责并转向四级或五级SAE框架。当埃隆暗示几周前他们找到了完全自动驾驶（FSD）时，也许就是指的这个。特斯拉现在知道该如何做，他们只需要去做。所以这肯定是一个巨大的进步，如果你想象一下30万行代码的话，你就能理解为什么这是巨大的进步。但问题仍然存在，他们是否会在另外几个月内达到另一个局部最大值，还是特斯拉会找到一个足够好的全局峰值，然后我们就可以起步了。无论如何，我们都会在这里观察一切的发展。你可以在X上找到我，我的用户名是dylanlumus22。希望大家有美好的一天。如果你喜欢这个视频，请给它点赞。再次感谢所有支持我赞助者。