We Need to Talk About What Elon Just Said About Tesla ⚡️
发布时间 2026-03-19 23:53:06 来源
Finally, after months of waiting, we get an update on 14.3. Elon said it's in testing right now, wide release in a few weeks. The expectations are sky high for this one. I admittedly have played some small role in that. But the understanding is this model should have significantly more reasoning and potentially that 10 times bigger parameter model. And if so, as always, that first release might need some refinement. So it might not go wide until something like 14.3.2. Then from there, it might require another one or two point releases for everything to be smoothed out and to really see where 14.3 will settle in. Just don't forget, there was a time when we thought 14.3 would be coming toward the end of 2025. So right now, the expectation is sometime in April, but do not be surprised if it slips until May.
终于，在经过数月的等待后，我们得到了关于14.3版本的更新。埃隆表示，目前正在进行测试，广泛发布将在几周后进行。大家对这个版本的期待非常高，我承认自己在提升期待上也起了一些小作用。据了解，这个模型应该会有显著的推理能力增强，可能还会增加10倍的参数。如果真是这样的话，首次发布可能需要进行一些调整，所以14.3可能不会立即广泛发布，可能会先到14.3.2才全面推开。从那时起，可能还需要另外一到两个小版本的更新来完善一切，真正看到14.3的最终效果。别忘了，我们曾经以为14.3会在2025年底才问世。而现在预计会在四月份发布，但如果延迟到五月也不要惊讶。
And I am hopeful that this will be the model after a few weeks of refinement out in the wild that we will finally see more unsupervised scaling with. And four things we will all be watching closely, navigation improvements, parking lot behavior, speed profiles, and we'll banish finally arrive with this one, where your car drops you off at the destination and then goes and finds a parking spot on its own. And bonus of fifth, I would love to see a much more reliable, actually smart summon. And I know that's all a big ask, but this is the one that Elon said should be the last big piece of the puzzle.
我希望在经过几周的实际使用和改进后，这将成为我们最终可以看到更多无人监督扩展的模型。我们将密切关注四个方面：导航改进、停车场行为、速度配置，以及最终实现这样的场景：您的车到达目的地时会自动把您放下，然后自己去找停车位。作为额外的惊喜，我希望看到一个更加可靠、真正智能的召唤功能。我知道这些期望很高，但这是埃隆所说的最后一块重要拼图。
If you did not yet think getting into semi-conductors would be important for purposes here, hopefully after the next few minutes you will have changed your mind. Phil Beisel said Tesla's AI5 uses a half-redical design, which is crucial for yield. A redical defines the imaging area of a lithography machine, an EUV machine, fitting two chips per shot effectively doubles yield. If Tesla hits its compute and efficiency targets with AI5 in this half-redical format, it's almost like cutting fab requirements in half. To which, Elon said AI5 will punch far above its weight, because the entire Tesla AI software stack is designed to make maximally effective use of every circuit.
如果你之前还没有考虑进入半导体行业的重要性，希望经过接下来的几分钟后，你会改变主意。Phil Beisel 说，特斯拉的 AI5 使用半掩模设计，这对提升产量至关重要。掩模定义了光刻机（或 EUV 机）的成像区域，每次曝光能容纳两个芯片可以有效地将产量翻倍。如果特斯拉的 AI5 在半掩模格式下达到了计算和效率目标，几乎就像将制造要求减少了一半。对此，Elon 说 AI5 的表现会大大超出预期，因为整个特斯拉的 AI 软件架构都是为了最大限度地利用每一个电路而设计的。
We co-signed our AI software and hardware. Bear in mind, AI5, while it can be used for training in data centers, is primarily optimized for AI edge compute and optimus and robotaxi. There's still significant room for improvement in the same half-redical and the same process node, and we think a single AI6 chip has the potential to match a dual system on chip AI5. Remember, Tesla has decided to remove things like a graphics processor and an ISP or an image signal processor. So removing those elements from the design because Tesla's software no longer needs them, will of course open up more room for AI compute transistors. In short, the chip will be able to do more of the AI math.
我们为我们的人工智能软件和硬件共同签署了合约。请注意，虽然AI5可以用于数据中心的训练，但它主要是为了AI边缘计算、Optimus机器人和Robotaxi（无人驾驶出租车）进行优化的。在相同的半导体架构和工艺节点上，还有相当大的改进空间，我们认为单个AI6芯片有潜力与双系统的AI5芯片相媲美。请记住，特斯拉已经决定移除像图形处理器和图像信号处理器（ISP）这样的组件，因为特斯拉的软件不再需要这些组件。通过从设计中移除这些元素，肯定会为更多的人工智能计算晶体管腾出空间。简而言之，新芯片将能够进行更多的AI计算。
To explain this half-redical, just to make sure everybody is up to speed and actually understands it. The EUV lithography machines that right now only one company, ASML, is making, and they're only making in the neighborhood of 50 to 60 of these giant machines every year. Each machine costs in the neighborhood of $400 million, and there is some new tech being developed, but right now, commercially at scale, all of this industry is reliant on those EUV lithography machines from ASML alone. And just think of those EUV machines as the printers for the actual wafers. So it's the machine that's going to take the design of these chips that companies like Tesla are designing, and then print it or etch it into the actual wafer itself to create the individual dies.
为了确保大家都跟得上并真正理解这一点，解释一下这个半导体行业的重要设备。目前，只有亚斯麦（ASML）这一家公司制造极紫外光刻机（EUV），他们每年只生产大约50到60台这种巨型设备。每台机器的成本大约为4亿美元。虽然有一些新技术正在研发中，但目前在商业规模上，这个行业完全依赖于亚斯麦的这些EUV光刻机。可以把这些EUV机器想象成实际晶圆的“打印机”。它们将像特斯拉这样的公司设计的芯片图案打印或蚀刻到实际的晶圆上，形成单个芯片。
So let's pretend that this is a wafer and this is an EUV machine. The light from the flashlight is the exposure field that the EUV machine creates. And then the radical would be what these companies customized to actually shape the light that's going on to the wafer. And so a full radical design means that a die or a chip itself can actually fit inside of this light field. But it's only one. So for each shot or print from the EUV machine, it only can make one chip or one die. But Tesla's half radical design means that its design for its chips can actually fit two chips inside this same circular light field.
所以，我们假设这是一片晶圆，而这是一个EUV机器。手电筒的光就是EUV机器产生的曝光场。然后，光罩就像是公司定制用来实际塑造照射到晶圆上的光的工具。因此，一个完整的光罩设计意味着一个芯片或一个晶粒可以完整地适应这个光场，但只能包含一个。因此，每次EUV机器的曝光或打印只能制造一个芯片或晶粒。而特斯拉的半光罩设计意味着它的芯片设计可以在同一个圆形光场中同时适应两个芯片。
So in one shot or one print session from the EUV machine, Tesla can actually make two chips instead of one chip for all companies using a full radical design. So all EUV machines create this standard size of the circular light in this flashlight example. But each company will customize one the radical design that would actually shape how this light hits the book. And they would customize the size of their dies to determine how many can actually fit inside the exposure field. So for every one print or shot of the EUV machine, Tesla can create two die instead of one.
在使用EUV（极紫外光刻）机器的一次曝光或打印过程中，特斯拉可以制造出两块芯片，而采用完整设计的其他公司只能制造一块。所有的EUV机器都会产生一个标准尺寸的圆形光，就像手电筒的光圈一样。但每家公司都会定制自己的设计来决定这道光如何投射到硅片上，并调整芯片的大小以确定可以在曝光区域内放置多少芯片。因此，每使用一次EUV机器，特斯拉能够生产出两块芯片，而不是其他公司的一块。
Which means relative to a full radical design, you only need half of the factory space. Or with a comparable footprint, you can create twice as many chips. Another benefit of the half radical design is that since it's smaller, the yields are usually better because you have less surface area that can effectively be damaged from something like dust or any other contaminant. And generally speaking, the less defects and the higher yields you have, the lower the cost per chip will be. And the yield is just the number of good or usable chips per wafer. So think of a wafer as a big round pizza and then each chip is one of the dies that's actually a slice of that pizza. So from one wafer, you might end up with just say 100 actual usable chips.
这段话的大意是：与完全根本性的设计相比，你只需一半的工厂空间。或者在同样的占地面积下，你可以生产出两倍数量的芯片。半根本性设计的另一个好处是，由于其体积较小，产量通常更高，因为受灰尘或其他污染物损坏的表面积较少。总体来说，缺陷越少，产量越高，单个芯片的成本就会越低。产量指的是每片晶圆上可以使用的好芯片数量。你可以把一片晶圆想象成一个大圆形披萨，每个芯片就是披萨上的一块切片。所以从一片晶圆上，你可能最终只得到大约100个可用的芯片。
And of course, like with everything else, there are some trade-offs to going the half radical route. One being you cannot fit as many transistors on a smaller design. So from a raw compute power perspective, you just will not have as much headroom. However, we know Tesla is focused on efficiency and mass scale. And this is why Elon said what he said. AI5 is primarily optimized for AI edge compute, not for training in data centers. Because with that trade-off of the half radical design, if you wanted to connect more of these AI5s together, you would have more packaging challenges. Again, actually getting each of these dies connected, which can result in latency and heat and a host of other challenges.
当然，和其他事情一样，选择半激进路线也有一些权衡。比如，你无法在较小的设计上容纳那么多晶体管。所以从纯计算能力的角度来看，你将没有那么多的余地。然而，我们知道特斯拉专注于效率和大规模生产。这就是为什么埃隆做出那种表述的原因。AI5主要是为AI边缘计算而优化的，而不是用于数据中心的训练。因为采用半激进设计方案，如果你想将更多的AI5连接在一起，会面临更多的封装挑战。实际上连接这些芯片也会导致延迟、发热以及其他一系列问题。
But despite that transistor cap from a half radical design, because Tesla's AI software and hardware are so tightly integrated, it can make up for any of these trade-offs. So with the half radical design, Tesla is clearly prioritizing massive volume, lower costs, better yields and better efficiency. And I've seen questions out there, people saying, well, if AI6 has the potential to be twice as good as AI5, as Elon just said, why wouldn't Tesla just skip to AI6? On the surface, it's a fair question, but think of AI5 as the foundation for this new design. That's where all of this testing, validation, the new equipment, the new design, all of that will be validated with AI5.
尽管半激进设计的晶体管上限存在，但由于特斯拉的AI软件和硬件紧密集成，它可以弥补这些折衷。因此，通过这一半激进设计，特斯拉显然优先考虑的是大规模生产、降低成本、更好的产量和更高的效率。我看到有人提出疑问，既然正如埃隆刚刚所说，AI6有可能比AI5好两倍，为什么特斯拉不直接跳到AI6呢？表面上看，这是个合理的问题，但可以将AI5视为这一新设计的基础。测试、验证、新设备和新设计的所有内容都会在AI5上进行验证。
Tesla will undoubtedly work through a lot of problems and challenges behind the scenes, and all of those learnings are ultimately what's going to enable Tesla to do something like an AI6. And as Elon said, with some luck and acceleration using AI, we might be able to tape out AI6 in December, which likely means the AI6 tape out or final design stage will be sometime in early 2027. And then from there, it would be six months to a year plus before that design can actually make it into production. Phil Beisel also said, for big chips like Tesla's AI5, a half-radical design, roughly 300 to 450 mm squared.
特斯拉无疑会在幕后解决许多问题和挑战，而这些经验最终将使特斯拉有能力完成类似AI6的项目。正如埃隆所说，如果运气好并且能够加速利用AI，我们可能在12月能完成AI6的设计定稿，这很可能意味着AI6的设计定稿或最终设计阶段将在2027年初完成。然后，从那时起，可能需要六个月到一年多的时间，这一设计才能真正投入生产。Phil Beisel还提到，对于像特斯拉的AI5这样的大型芯片，属于半激进的设计，面积大约在300至450平方毫米之间。
Producing 100 million chips a year would require about 100,000 wafer starts per month, roughly the output of a single high volume leading edge fab. Very important distinction, if you see measurements, like we'll say 400 mm squared, that's going to be the surface area of one of these die. And then if you see a measurement for the wafer itself, it's usually going to be just millimeter, so again, that's the diameter of the circle. So it would not be squared, and let's just say the average size of a wafer is 300 mm when it comes to the diameter. And there's a simple equation to go from the diameter of that wafer to the actual surface area of the wafer.
每年生产一亿个芯片需要每月大约10万片晶圆的投片量，这大致相当于一家大型高端晶圆厂的产量。有一个非常重要的区别需要注意：如果你看到的是像400平方毫米这样子的测量单位，那是指一块芯片的表面积。而如果你看到晶圆的测量单位，通常只会是毫米，这指的是晶圆的直径，因此它不是平方毫米。一般来说，一片晶圆的平均直径是300毫米。有一个简单的公式可以将晶圆直径转换成实际的晶圆表面积。
And after doing that math for a standard wafer size, it would put us in the neighborhood of 70,600 mm for the surface area of a wafer. So then we have comparable measurements for the surface area of a wafer, and the surface area of one of the smaller dies that gets cut out of the wafer. And this will hopefully all make sense in a moment, but Elon responded to this saying, probably more like 160,000 waifers per month, factoring in yield. So 160,000 wafer starts per month is 1.92 million per year, just times 12. And since we know to start, Tesla will be aiming for 100 million chips per year.
在对标准晶圆的面积进行计算后，我们得出的晶圆表面积约为70,600平方毫米。这样一来，我们就有了晶圆表面积和从晶圆上切割出来的较小芯片的表面积的可比数据。虽然这些信息可能一时还不太明了，但埃隆对此表示，考虑到良品率，月产晶圆可能会达到16万片。所以每月启动16万片晶圆生产意味着一年是192万片（即16万×12个月）。而根据我们所知道的情况，特斯拉的目标是在开始阶段每年生产1亿颗芯片。
You would simply take that 100 million chip number and divide it by the number of annual wafer starts that 1.92 million, which would mean for Tesla to hit its goals, it would need roughly 52 usable dies from every wafer. And this is where it all ties back to that surface area math I just laid out. If we assume the average wafer is 70,000 mm squared, and we divide that by, let's say, 500 mm squared for the average size of each AI5 die. That would give us 140 potential usable dies from each wafer. Meaning if Tesla is actually getting 52 usable dies from each wafer, divided by that theoretical capacity, that would be a yield of about 37%.
你只需将那1亿芯片的数量除以192万年芯片投片量，这意味着为了达到目标，特斯拉每片晶圆需要大约52个可用芯片。这就与我前面提到的表面积计算相关。如果我们假设每片晶圆的平均面积为70,000平方毫米，并将其除以每个AI5芯片的平均大小500平方毫米，这样每片晶圆理论上可以产生约140个可用芯片。这意味着如果特斯拉每片晶圆实际取得52个可用芯片，相比于理论产能，这样算的话，生产良率大约为37%。
Which for a two nanometer facility that is Frontier level tech, that's typically where things are expected to start. But of course, if long-term, if Tesla can get the yield up into the neighborhood of 60%, with the same wafer in die size, you would take that 140 theoretical capacity of each wafer, times 60%, that would give you 84 chips per wafer. Times that 1.92 million waifers per year, that would be 1.6 million usable chips per year, just jumping up to a 60% yield. And then, in case you forgot, Elon did say that eventually, the tariff lab will be expecting to reach 1 million wafer starts per month. Which is nearly in order of magnitude to jump from where it plans to start at that roughly 160,000 wafer starts per month.
对于一个使用两纳米工艺的前沿技术设施来说，这通常是事情的起点。然而，如果从长期来看，特斯拉能将良品率提高到大约60%，在晶圆和芯片尺寸相同的情况下，那140个理论产能的晶圆乘以60%良品率，就可以每片晶圆生产84个芯片。再乘以每年192万片晶圆，这将意味着每年能生产出160万颗可用芯片，仅仅通过提升到60%的良品率就实现了这样的增长。此外，如果你忘记了，埃隆·马斯克曾表示，晶圆代工实验室最终计划达到每月100万片晶圆的启动量，这几乎是从计划的每月约16万片晶圆启动量开始的一个数量级的增长。
So listen, I understand if you're pretty new to this whole world, you might need to go back and watch that session one or two or more times. Because as we spend more time talking about the semiconductor space, having a base understanding of the naming and the measurements, and the wafers and dies, and all of the fundamentals of the space will be crucial to understand where things are going. And again, to have any chance of actually understanding why Tesla is making the decisions it's making.
所以听我说，我理解如果你对这一领域比较陌生，你可能需要回过头去看一到两遍甚至更多次那个课程。因为随着我们花更多时间讨论半导体领域，对一些基础知识的了解，比如命名、测量，以及晶圆和芯片等基本概念，是理解行业发展方向的关键。同样地，要想弄明白特斯拉为何做出某些决策，也是必不可少的。
And to finish it off, back to what Phil said, at current industry efficiency, that translates to about 1.2 million square feet of clean room space. That being roughly 100,000 wafer starts per month. And since the clean room usually takes up around 20 to 35% of total facility area, the initial tariff lab phase would likely span between 4 and 10 million square feet. For context, Gigatexis is right around 10 million square feet.
最后，回到菲尔所说的，以当前行业效率来看，这相当于大约120万平方英尺的洁净室空间。这样的空间能每月大约生产10万片晶圆。而由于洁净室通常占总设施面积的20%到35%左右，初始阶段的实验室面积可能会在400万到1000万平方英尺之间。作为对比，Gigatexis 的面积大约是1000万平方英尺。
But arguably, the most important thing Phil said, the point of tariff lab's redesign is to break conventional fab assumptions, compressing more output into less space, so that extreme production volume becomes achievable. So we already have four aspects that Tesla will be customizing, the logic, the memory, the packaging, and the clean rooms. And so current industry assumptions and norms are just that. Tesla will be looking to take a step forward from all of those metrics. And so hopefully, this segment made it clear that Tesla going with a half-redical design is not some minor footnote.
可以说，Phil 所说的最重要的一点是，特斯拉实验室重新设计的目的是打破传统制造的假设，将更多的产出压缩到更小的空间中，从而使极高的生产量成为可能。因此，特斯拉已经在四个方面进行定制：逻辑、存储、封装和净化室。当前行业的假设和标准只是现状。特斯拉将努力在所有这些指标上更进一步。希望这一部分能让大家明白，特斯拉选择采用部分激进的设计并不是一个无关紧要的小细节。
It will fundamentally have direct impacts on the efficiency, on the yields, and on the output, and the cost of all of it. And who knows what type of transparency Tesla will give us into the tariff lab developments and in the future when it's actually producing chips? Will it give us a breakdown of what's going on during quarterly or semi-annual reports? Either way, I plan to do everything in my power to keep regular electrified viewers informed and educated on this critical new portion of Tesla's business.
这将从根本上直接影响效率、产量、输出和成本。而且，谁也不知道特斯拉在未来生产芯片时，会给我们提供多大的透明度来了解其在关税实验室的进展。它会在季度或半年度报告中为我们详细介绍相关情况吗？无论如何，我计划竭尽所能，让关心电气化的观众定期了解到这一特斯拉业务新关键领域的最新资讯。
So we've been talking about how Tesla is using AI to do things that most people still think is years away, but the other side of that AI that nobody wants to talk about is how bad actors are using it too. Which brings me to delete the sponsor of this video. They just published their 2026 privacy predictions and one stat was pretty alarming. Scammers now only need a few seconds of your voicemail audio to clone your voice with AI. They can call your parents or your spouse sounding exactly like you, claiming their endanger and need money wired immediately.
我们一直在讨论特斯拉如何使用人工智能（AI）来完成一些大多数人认为还需要几年才能实现的事情，但没有人愿意谈论的另一面是，一些不法之徒也在利用人工智能。接下来我要提到的是本视频的赞助商，他们刚刚发布了2026年的隐私预测，其中有一个数据相当令人警醒。现在，骗子仅需几秒钟的语音邮件音频，就能使用人工智能克隆你的声音。他们可以打电话给你的父母或配偶，声音完全像你一样，然后谎称自己遭遇危险，需要立即汇款。
And here's the thing. These scams work so well because data brokers, specifically, people search sites are linking your family members' profiles to yours. So a bad actor does not just find your name and address, they find your parents' names, your siblings, your kids, all connected, all for sale. Delete me specifically removes your personal information from these people search sites, the ones that show up when someone Googles your name and finds your home address, phone number, your relatives, all of it.
问题就在这里。这些诈骗之所以能如此奏效，是因为数据经纪人，特别是人员搜索网站，把你的家庭成员资料与你的资料关联在一起。因此，坏人不仅能找到你的姓名和地址，还能查到你父母的名字、兄弟姐妹、孩子的信息，这些信息都是相互关联并可被出售的。Delete Me 专门帮助你从这些人员搜索网站上删除你的个人信息，这些网站会在有人通过谷歌搜索你的名字时展示出你的家庭住址、电话号码、亲友信息等一切内容。
I've been using them for years now and what I appreciate most is they don't just remove your information once and then walk away. Every quarter they rescan because these sites constantly repopulate your data. They also offer email and phone masking so when you sign up for new sites, your real contact info never has to touch their database. Plus, delete me is an American company. This is all they do and their trust pilot reviews are excellent.
我已经使用它们好几年了，我最欣赏的是，他们不仅仅是一次性删除你的信息就不管了。每个季度他们都会重新扫描，因为这些网站会不断地重新出现你的数据。他们还提供电子邮件和电话的隐私保护，这样当你注册新网站时，你的真实联系方式就不必被存入他们的数据库。而且，DeleteMe是一家美国公司，他们专门做这些事。在Trustpilot上的评论也非常好。
So if you'd like to support the work I do here and protect your family, you can head to joindeleteme.com slash electrified to get 20% off using my code electrified, linked below or by using the QR code on the screen. Sawyer got some more photos from another follower that Tesla's Robotaxi service is now testing in Orlando, Florida. The same situation we just saw in Las Vegas and Dallas. Multiple model wise with the rear camera washers and Texas manufacturer plates.
如果您想支持我在这里的工作并保护您的家人，可以访问 joindeleteme.com/electrified，使用我的代码 electrified 可享受 20% 的折扣，该链接在下方或通过屏幕上的二维码也可以访问。Sawyer 收到了来自另一位粉丝的照片，显示特斯拉的无人出租车服务现在正在佛罗里达州的奥兰多进行测试。这和我们刚刚在拉斯维加斯和达拉斯看到的情况一样。多个 Model Y 车型配备后置摄像头清洗器和得克萨斯州制造商车牌。
Of course, the big question remains how long from this type of testing until actually launching the service. But it does seem like Tesla launching in Las Vegas, Dallas and Orlando are all imminent. And if you look at Waymo's map, you'll see they are serving riders in Orlando. Now, if you just compare the population of Austin to Orlando, you'll find that Orlando is about one third the size in terms of population compared to Austin. However, I think the more important metric would be how many annual visitors visit each city. And from that perspective, Orlando easily gets twice as many annual visitors as Austin. And so tourism and people wanting experiences and maybe not wanting to rent cars. Orlando should be a great setup from a tourism perspective. And of course, the same thing would apply for Las Vegas.
当然，一个重要的问题仍然是，从这种测试到实际推出服务需要多长时间。但看起来特斯拉在拉斯维加斯、达拉斯和奥兰多的启动都迫在眉睫。如果你查看Waymo的地图，你会看到他们已经在奥兰多为乘客提供服务。现在，如果你比较奥斯汀和奥兰多的人口，你会发现奥兰多的人口大约是奥斯汀的三分之一。不过，我认为更重要的指标是每年有多少游客访问这些城市。从这个角度来看，奥兰多每年的游客数量是奥斯汀的两倍。因此，从旅游和游客希望体验不同、可能不想租车的角度来看，奥兰多应该是一个绝佳的选择。当然，这对于拉斯维加斯也是同样适用的。
And this was the most recent Geofence for Waymo in Orlando that I could find. And some of the reporting from February says it covers around 60 square miles. But in my opinion, by far, the number one question, how long will the safety monitors be in the vehicles for each new city for Tesla's Robotaxi? Will it be seven months like Austin or will Tesla be able to cut that number down significantly? And right now, Waymo's Geofence in Orlando does not cover Disney, the airport, or any parts of downtown Orlando. However, this right here is something to watch. Lyft has actually been named the official ride share of Walt Disney World. Lyft and Disney Parks and Resorts will collaborate in a number of exciting new ways, including an enhanced presence at Disney Resorts.
这段话翻译成中文如下：
“这是我能够找到的关于Waymo在奥兰多的最新地理围栏。根据二月份的一些报告，它覆盖了大约60平方英里。不过我认为，目前最大的一个问题是，特斯拉的自动驾驶出租车在每个新城市中需要配备安全监控人员的时间会有多长？像奥斯汀那样是七个月，还是特斯拉能大幅缩短这个时间？目前，Waymo在奥兰多的地理围栏并未涵盖迪士尼、机场或市中心的任何部分。然而，这里有一个值得关注的事情：Lyft实际上已被命名为沃尔特迪士尼世界的官方打车服务。Lyft与迪士尼公园和度假区将在多个令人激动的新方式上开展合作，其中包括在迪士尼度假区的强势业务存在。”
So that's not an official reason why Waymo is not yet covering parts of Disney. Remember, Waymo just launched its service in Orlando in February this year. But the last point Disney does have plenty of private property or internal roads that any third party provider like Tesla's Robotaxi or Waymo would have to get approval from for Disney. And I'll say what needs to be said, hopefully very soon, Tesla starts shipping customer cars with all of the camera cleaners. Today, Nitsa has escalated one of its probes into Tesla.
这并不是Waymo尚未覆盖迪士尼部分地区的官方原因。请记住，Waymo今年2月才刚刚在奥兰多启动服务。但最后需要指出的是，迪士尼确实有很多私人财产或内部道路，任何第三方提供商，比如特斯拉的机器人出租车或Waymo，都需要获得迪士尼的批准。我希望该说的话能很快实现，希望特斯拉尽快开始为客户汽车配备所有摄像头清洁工具。今天，美国国家公路交通安全管理局（NHTSA）提升了对特斯拉的调查级别。
This on 3.2 million Tesla vehicles with FSD because the system may fail to detect or warn drivers in poor visibility. This preliminary evaluation was opened back October 2024. But now the agency is opening an engineering analysis a required step before it can seek a recall. Nitsa said its investigation raises concerns the Tesla camera-based system did not detect common roadway conditions like glare, dust, or other airborne obstructions that impaired camera visibility or provide alerts when camera performance deteriorated until immediately before a crash. Nitsa said it had reports of nine incidents that may be tied to the issue, including one fatal crash and two injury crashes and was investigating whether six other crashes may be related.
这涉及320万辆配备FSD的特斯拉车辆，因为该系统在能见度差的情况下可能无法检测或警告驾驶员。该初步评估是在2024年10月启动的。但现在，相关机构正在进行工程分析，这是在考虑召回之前必需的步骤。Nitsa表示，其调查引起了对特斯拉基于摄像头的系统不能检测到常见道路状况（如眩光、灰尘或其他悬浮的障碍物）从而导致摄像头能见度受损的担忧，并且只有在即将发生事故前才提供警报。Nitsa称，他们收到九起可能与该问题有关的事件报告，其中包括一起致命事故和两起伤人事故，并正在调查另外六起事故是否有关联。
Nitsa said the data raises concerns the system fails to detect or warn drivers appropriately under degraded visibility conditions, which some of you might find as odd because in my experience anytime there's any bit of rain, I always get that warning, FSD performance may be degraded. But Nitsa said in many of the crashes reviewed, FSD also lost track of or never detected a lead vehicle in its path. Bear in mind, this now preliminary investigation is separate from another one into 2.8 million Tesla vehicles, where there were more than 50 reports of traffic safety violations and a series of crashes. Nitsa said FSD has induced vehicle behavior that violated traffic safety laws.
Nitsa表示，这些数据引起了人们的担忧，因为系统在能见度降低的情况下未能适当检测或警告司机，这可能会让一些人感到奇怪，因为根据我的经验，只要有一点雨，我总能收到FSD性能可能会降低的警告。但Nitsa指出，在他们审查的许多碰撞事件中，FSD也可能丢失或从未检测到其路径上的前方车辆。请注意，这项初步调查与另一项涉及280万辆特斯拉汽车的调查是分开的；在那项调查中，有超过50起交通安全违规报告和一系列碰撞事件。Nitsa表示，FSD引发了一些违反交通安全法律的车辆行为。
Tesla's analysis showed that if an update to the degradation detection system was installed at the time of the crash, it may have affected three of the incidents. Nitsa plans to examine the performance of the updated system, including when it was deployed, how widely it's been rolled out and whether it improves the system's ability to detect visibility issues and alert drivers in time. You guys know I always try to avoid pointing the finger for any daily Tesla stock price performance, but I will say in this case I understand why so many would point the finger at this news.
特斯拉的分析显示，如果在车祸发生时安装了更新的降级检测系统，有可能对其中三个事件产生影响。Nitsa计划研究更新系统的性能，包括它何时部署、推广范围，以及是否能改进系统检测可见性问题并及时提醒司机的能力。大家都知道，我一向避免归咎于任何特斯拉股价的日常表现，但在这个案例中，我可以理解为什么许多人会把矛头指向这个消息。
Now, an engineering analysis upgrade like this is certainly nothing new for Tesla it's happened before. But from here, it might be anywhere from three months to a year plus until Nitsa actually decides what they wanna do. There will be a back and forth, data gathering, communications and testing. And if Nitsa decides it could issue a recall, but as we know, it would most likely just be a software update. And remember, as we saw, Tesla already shipped a software update addressing these concerns that Nitsa will now be reviewing.
现在，像这样的工程分析升级对特斯拉来说并不新鲜，之前也发生过。但从现在开始，国家公路交通安全管理局（NHTSA）可能需要三个月到一年甚至更长时间来决定他们想要做什么。这期间会有来回的沟通、数据收集和测试。如果NHTSA决定采取行动，可能会发出召回通知，但正如我们所知，这很可能只需要进行软件更新。请记住，正如我们所看到的，特斯拉已经发布了一次软件更新来解决这些问题，现在NHTSA将对此进行审查。
So it's something to keep an eye on, not because of the stock move today, but because of all of the federal conversation happening when it comes to sensors, autonomous vehicles, what's required, what's safe and so on. So the sooner Tesla's camera only system can be cleared of any wrongdoing or shortcomings, I think the better. At least when it comes to the regulatory environment being comfortable with Tesla's camera only approach.
因此，这件事值得关注，不是因为今天股价的变动，而是因为涉及传感器、自动驾驶车辆及相关要求和安全等方面的联邦讨论正在进行。因此，我认为，特斯拉的纯摄像头系统越早被证明没有任何问题或不足，就越好。至少在监管环境对特斯拉的纯摄像头方式感到放心方面是如此。
We learned Uber is planning to invest up to $1.25 billion in Rivian through 2031. As Uber is expected to buy 10,000 fully autonomous R2 Robotaxies with the option to buy up to 40,000 more in 2030. Initial commercial deployments are planned for San Francisco and Miami in 2028, scaling to 25 cities through 2031. Now look, this comes at a perfect time. We just got done explaining how everybody and their brother will be announcing partnerships when it comes to autonomy.
我们了解到，Uber 计划在2031年前向Rivian投资最多12.5亿美元。Uber预计将购买1万台全自动驾驶的R2机器人出租车，并有权在2030年之前再购买多达4万台。首次的商业部署计划在2028年于旧金山和迈阿密开始，然后在2031年前扩展到25个城市。现在来看，这个投资时机是非常恰当的。我们刚刚解释过，随着自动化的发展，各种合作伙伴关系将不断被宣布。
Every company wants its shareholders to be excited that they're gonna have a role in our autonomous future. And as we've said, Uber really has no choice but to seek out every single company that says the word autonomy to try to acquire them or create a deal with them. Should all milestones be achieved, the companies will have deployed thousands of unsupervised Rivian R2 Robotaxies across 25 cities in the US, Canada, and Europe by the end of 2031.
每家公司都希望它的股东对自己将在自动化未来中扮演角色感到兴奋。正如我们所说，Uber实际上别无选择，只能去寻找每一个提到“自动化”一词的公司，尝试收购它们或与它们达成交易。如果所有目标都实现，这些公司将在2031年底之前，在美国、加拿大和欧洲的25个城市部署数千辆无监控的Rivian R2无人驾驶出租车。
But initial deployments in two cities in 2028, which effectively gives them two years to make some type of magic happen, but spoiler alert, it's not going to happen. If you're not familiar with where Rivian is at today with its UHF Universal Hands-Free System, I'd highly encourage you to watch this video from Devon Olsen, I will have it linked below. The TLDW, it does not seem very safe. It seems very jerky and uncomfortable. And for somebody like Devon, that's very tech forward and comfortable with autonomous vehicles.
在2028年，最初将在两个城市部署这项技术，这实际上给他们两年的时间来实现某种奇迹，但提示一下，这不太可能实现。如果你不太清楚Rivian目前的UHF全自动免提系统的情况，我强烈建议你观看Devon Olsen的视频，我会在下面附上链接。简而言之，这个系统看起来不太安全，显得非常不稳定和不舒适。即便是像Devon这样对科技非常前沿，并对自动驾驶车辆感到舒适的人，也感到问题重重。
Even he was having some issues with the UI in the Rivian and understanding what level of driver assistance was active. Now in fairness, I will say anybody coming from FSD trying the system from Rivian, the expectations are impossible to meet. And of course, the experience would be very different if someone has never used FSD or really any advanced driver assistance features. And let me be clear, I would love to be wrong about all of this.
即使是他，在使用Rivian的用户界面时也遇到了一些问题，不太明白当前的驾驶辅助系统级别。公平地说，对于那些习惯了FSD（全自动驾驶）的人来说，转而使用Rivian的系统，期望值很难达到。当然，如果有人从来没有用过FSD或者任何高级驾驶辅助功能，体验自然会非常不同。我得澄清一下，我其实很希望我对这些看法是错的。
I desperately want to see autonomous vehicles that are actually safe proliferate around the world. I have just seen way too many promises and way too many failures from legacy auto companies and other tech companies, trying to copy Tesla, thinking they're going to do it, but in the end, ultimately failing. And Rivian's not a legacy auto company, but we have legacy auto that could not figure out how to make EVs profitably, yet now they're all partnering up with Nvidia saying that they're gonna solve for autonomy.
我迫切希望能够真正安全的自动驾驶汽车在全球普及。我已经看到太多的承诺和失败，许多传统汽车公司和其他科技公司试图模仿特斯拉，以为自己能做到，结果最终失败。虽然Rivian不是一家传统汽车公司，但传统车企无法找到盈利制造电动车的方法，而现在他们都和英伟达合作，声称能够解决自动驾驶技术的问题。
I'm sorry, but I'm just not buying it. Tesla has some interesting new job postings, one for a technical program manager for infrastructure semiconductor. AKA Tesla is officially hiring for the TerraFab. This one will be engineering integration, utility planning and factory design and construction from concept through execution. Tesla also has 12 job postings right now for various social media content roles, video content producer for AI, and one is content producer for Robotaxi.
抱歉，我实在难以相信。特斯拉最近发布了一些有趣的新职位招聘，其中一个是基础设施半导体的技术项目经理。这也就是说，特斯拉正式开始为TerraFab招聘了。这个职位将负责从概念到执行的工程整合、公用事业规划以及工厂设计和建设。此外，特斯拉目前还有12个社交媒体内容相关的职位空缺，包括人工智能的视频内容制作人，还有一个是为Robotaxi做内容制作的职位。
This person will conceptualize, capture and edit video projects that support the Robotaxi program. The content will include public-facing videos that tell the story of Robotaxi, like product, factory and industry updates, customer engagement content, and creative content for assembly line and manufacturing scenarios. I do think it's fair to take this as a sign that Tesla is gearing up to get the word out there a bit more about Robotaxi.
这个人将构思、拍摄和编辑支持Robotaxi项目的视频。这些内容将包括面向公众的视频，如产品、工厂和行业更新、客户互动内容，以及用于装配线和制造场景的创意内容。我确实认为这可以被看作是特斯拉准备更广泛宣传Robotaxi的一个信号。
And on that point, Morgan Stanley just said they're more optimistic about Tesla's progress toward an unsupervised Robotaxi rollout after a recent visit to Gigatexis. Particularly its progress in addressing edge cases around pickup and drop off, which ties right back to 14.3 if it can add a significant reasoning, especially when it comes to parking lot behavior, which will be massive for pickup and drop off, that absolutely could be one of the bottlenecks for scaling right now.
关于这一点，摩根士丹利公司在最近访问特斯拉的德州超级工厂后表示，他们对特斯拉在推出无需监督的无人出租车计划方面的进展更加乐观。特别是在解决上下车时出现的极端情况方面取得了进展，这与第14.3版密切相关，尤其是在停车场行为上的显著推理能力，这对于上下车来说至关重要。目前，这可能是其扩大规模的一个瓶颈。
Morgan Stanley said FSD scaling is the most important thing for Tesla this year, no surprise, but after a site visit and likely talking to some engineers, they're feeling more confident. I was waiting for some more clarity on that Fox reporting on the Cybertruck crash that was framed as FSD on the Cybertruck nearly driving a mother and her baby off an overpass.
摩根士丹利表示，FSD（完全自动驾驶）的规模化是特斯拉今年最重要的事情，这并不令人意外。但在实地考察并可能与一些工程师交流后，他们对这点更加有信心。我一直在等待关于福克斯新闻报道的更多细节，该报道称一辆配备FSD的Cybertruck差点将一位母亲和她的婴儿推下立交桥。
But anybody that saw the video and has ever used FSD would know that FSD would not behave like that. And Elon said the log shows the driver disengaged autopilot four seconds before crashing. So for the entirety of that video, FSD was not engaged. And if you missed it, this was the video in question. I'm not going to give this clown any more airtime than he deserves.
任何看过该视频并使用过FSD（全自动驾驶系统）的人都会知道，FSD不会那样表现。埃隆提到，日志显示司机在碰撞前四秒解除了自动驾驶模式。所以在整个视频中，FSD并没有启动。如果你错过了，这就是我们在讨论的视频。我不想给这个小丑更多的关注。
I learned months back that Matt Watson from CarWOWUK is super anti-Tesla, very similar to JerryRig, everything, suffering from a severe case of EDS. They just released a video that was initially titled Tesla full self driving is pointless, but after being called out vehemently in the comments, he has since changed it to why Tesla's enhanced autopilot is pointless. But of course in the video, he still refers to it as FSD. But as we know, FSD is not even available in the UK. So ultimately, he's running tests with EAP on streets that it's not designed to be used on. And thankfully, plenty of people are in the comments saying how embarrassing this video actually is.
几个月前，我了解到CarWOWUK的Matt Watson非常不喜欢特斯拉，这和JerryRig Everything很相似，他们都严重反对特斯拉的FSD（自动驾驶）。他们刚刚发布了一个视频，最初的标题是“特斯拉完全自动驾驶毫无意义”，但在评论中遭到强烈批评后，他将标题改为了“为什么特斯拉增强版自动驾驶毫无意义”。不过在视频中，他依然将其称为FSD。而我们知道，FSD在英国根本没有推出。所以实际上，他是在不适用的街道上测试EAP（增强版自动驾驶）。值得庆幸的是，很多人都在评论中指出这个视频有多么尴尬。
But I bring it up as a reminder, as Tesla gets closer to actually scaling, there will be targeted attacks and smear campaigns against Tesla and FSD, especially across parts of the EU, where largely those people have not experienced true FSD. As we know, legacy automakers over there have zero chance of offering anything close to the full capability of FSD. So once customers over there realize what Tesla has to offer and the regulatory blockers are removed, we say things like it's over a lot, so it becomes tight. But once people across the EU actually see FSD for what it can really do, it really should spread like wildfire, even if it's supervised.
我提到这一点是为了提醒大家，随着特斯拉在规模化发展上越来越接近实现，会有针对特斯拉和全自动驾驶（FSD）的攻击和抹黑活动，尤其是在欧盟的一些地区，那里的多数人还没有真正体验过真正的FSD。正如我们所知，那边的传统汽车制造商完全没有能力提供接近FSD全功能的产品。所以，一旦那里的消费者了解特斯拉的产品优势，并且监管障碍被清除，我们常说"一旦发生变化，市场就紧张"，但真正让欧盟地区的人们见识到FSD的实际能力后，它就应该像野火般迅速传播，即使是在监督下运行。
But again, just prepare yourself for all kinds of frustrating nonsense from all corners of the earth against Tesla and FSD. This one is basically Mark Rober, Luminar and Lidar, part two. Sawyer said Tesla's asking a Colorado federal court to throw out a suit alleging a defect in its autopilot tech led to a fatal car crash. Tesla saying the evidence shows autopilot was not on and that the driver was drunk well beyond the legal limit. Elon replied to that saying common story.
但是，请你做好准备，全球各地都会有各种针对特斯拉和全自动驾驶（FSD）的令人沮丧的无稽之谈。这件事基本上就是马克·罗伯（Mark Rober）、Luminar 和激光雷达的续集。据索耶（Sawyer）所说，特斯拉正在请求科罗拉多州的联邦法院驳回一项诉讼，该诉讼称其自动驾驶技术的缺陷导致了一次致命车祸。特斯拉表示证据显示自动驾驶并未开启，而且司机的酒精含量远超法律限度。埃隆对此回应称这是一个常见的故事。
So not only will there be targeted smear campaigns taking place, but we will have plenty of greedy, opportunistic people that will be looking to point the finger blaming Tesla and its technology instead of taking accountability and responsibility themselves, trying to get a payout. And that won't be the case every time there may be some instances where Tesla is actually at fault to a degree. But by and large, many of the cases we have seen have ended up in Tesla's favor once all of the facts and the data come out.
所以，不仅会有针对性的抹黑活动，还有许多贪婪和机会主义的人会试图把责任推给特斯拉及其技术，而不是承担他们自己的责任，希望能得到赔偿。当然，并不是每次都是这样，有些情况下特斯拉确实存在一定程度的过错。但总体来说，许多我们看到的案例在所有事实和数据公布后，结果对特斯拉是有利的。
I will say sometimes Elon does not make sense or I'm just not comprehending what he's saying. Today he said Google will win the AI race in the West, China on earth and SpaceX in space. This only a few hours after saying that SpaceX in a few years would dominate everybody else combined. But now Google is going to win in the West, China on earth, so does that exclude the West? What does that mean for Google in the West? Because if China wins on earth, then that would mean they out-compete Google. And now Elon's just saying SpaceX will win in space.
有时候我觉得埃隆·马斯克说的话不太合逻辑，要么就是我没理解他的话。今天他提到Google将在西方的AI竞赛中获胜，中国将在地球上取得胜利，而SpaceX将在太空中领先。可就在几小时前，他还说SpaceX将在几年内超过所有其他竞争者的总和。现在又说Google将在西方获胜，而中国在地球上获胜，那这是否意味着西方不包括在内？那对Google在西方的优势意味着什么呢？因为如果中国在地球上获胜，那这岂不是意味着他们会超过Google。而现在，埃隆只是说SpaceX将在太空中取得胜利。
Honestly guys, sometimes all I'm going to have for you is in, I don't know, and this is one of those times. Frankly, I think that most predictions about AI, no matter who it's coming from, will ultimately end up being wrong. But you can make of that what you will. Tesla stock closed the day at $380.30 down 3.18%. While the NDX was down 0.29%. The volume was 14% above the average.
说实话，各位，有时候我真的不知道要怎么说，而这就是其中一次。坦白讲，我认为无论是谁对人工智能作出的预测，最终大多会被证明是错误的。但你可以根据自己的理解来解读。特斯拉的股票今天收盘价为380.30美元，下跌了3.18%。而纳斯达克100指数下跌了0.29%。当天的交易量比平均水平高出14%。
Don't forget, if you'd like to metaphorically slap some of these scammers to sign up for, delete me, links below, grab that 20% off, and thank you for supporting the channel. Hope you all have a wonderful day and a huge thank you to all of my Patreon supporters.
别忘了，如果你想用隐喻手法给这些骗子来点颜色看看，可以注册“删除我”，链接在下方，还能获得八折优惠。感谢你对频道的支持。希望你们都有美好的一天，并特别感谢我所有的Patreon支持者。