Nvidia's Keynote: Trouble for Tesla? / Jensen Talks Elon and Tesla / Grok Coming to Teslas ⚡️

Welcome to Electrified, it's your host Dylan Loomis. Quick shout out to Michael G. Thank you for using my tesla referral link and congratulations on your new tesla. Earlier today, Elon was running some streaming tests on x and have a listen to what he said. I grok in teslas is coming soon. So you'll just be able to like talk to your tesla and ask for anything. Hey christen, yeah all teslas will be able to use grok. The grock kind of the intelligence to grok kind of lives in the cloud in the data center. So it doesn't matter what's in the car, it can just chat to grok in the car. Definitely the one I've been waiting for. Now yes, there are still plenty of questions and yes, he said the dreaded coming soon. So it could be two weeks, it could be six months plus.

欢迎来到Electrified，这里是你的主持人Dylan Loomis。首先感谢Michael G.，谢谢你使用我的特斯拉推荐链接，恭喜你购入新的特斯拉。今天早些时候，Elon在进行一些流媒体测试，来听听他说了什么。“我在特斯拉上理解（grok）即将上线。”这意味着你将可以直接和你的特斯拉对话，提出任何请求。嘿，克里斯汀，是的，所有特斯拉都会支持这项功能。Grok的智能主要在云端的数据中心，所以车内硬件无所谓，车子可以直接和云端的Grok通信。这绝对是我一直以来期待的功能。当然，目前还有许多疑问，而且他用了“即将上线”这个不确定的词，所以可能是两周，也可能是超过六个月之后。

For now, my main question is how integrated will grok be with the tesla UI and tesla feature. So will this be a major improvement to voice commands for control functions? Or to start, will it just be grok's more traditional functionality just in your tesla vehicle? Either way though, the ability to talk to it, aka audio inputs is exactly what I was hoping for. So a very encouraging update. And to everybody that argued xai would be a distraction for tesla. If and when this feature comes to fruition, that'll be a pretty strong counterpoint. Tesla, full self driving, unsupervised.

目前，我主要的问题是，Grok 将如何与特斯拉的用户界面和功能集成。这是否会大幅改善语音指令的控制功能？还是最初只是在特斯拉汽车中提供 Grok 的一些传统功能？无论如何，能够通过语音输入进行交流正是我所期待的，非常令人鼓舞的更新。对于那些认为 XAI 会分散特斯拉注意力的人来说，如果这个功能实现了，那将是一个有力的反驳。特斯拉，全自动驾驶，无需监督。

Let me mention we're going to be when we actually are doing trials of that for Tesla employees already. And we expect to have that in commercial service sometime this year, which I mentioned at the last their next call. Just to be clear that last clip is not actually news when it comes to unsupervised FSD trials already taking place internally at tesla for employees. We were reminded of that on the Q3 call. NHTSA has opened a new probe into 2.6 million tesla's over smart summon.

让我提一下，我们实际上已经在为特斯拉员工进行试验。我们预计在今年某个时候将把这项服务投入商业使用，这在上次电话会上已经提到过。需要明确的是，关于无人监督的FSD（全自动驾驶）试验已经在特斯拉内部员工中进行，这并不算是新闻。在第三季度的电话会议上，我们就被提醒过。NHTSA（美国国家公路交通安全管理局）已经对260万辆特斯拉汽车启动了关于智能召唤功能的新调查。

Just know some of the reports on this news are conflating actually smart summon with regular smart summon. But this preliminary evaluation is based on reports of four crashes involving tests of vehicles. The vehicles failed to detect posts or parked vehicles when they were operating on actually smart summon. NHTSA said it had reports where users had too little reaction time to avoid a crash either with the available line of sight or releasing the phone app button, which stops the vehicle's movement.

要注意，有些报道将真正的智能召唤功能与普通智能召唤混为一谈。然而，这项初步评估是基于四起涉及车辆测试的事故报告。这些车辆在使用真正的智能召唤功能时未能检测到路标或停泊的车辆。美国国家公路交通安全管理局（NHTSA）表示，有报告称用户由于可视范围有限或需要释放手机应用程序按钮来停止车辆运动，反应时间过短，无法避免碰撞。

Remember the three steps for NHTSA are one preliminary evaluation. Two, it could be upgraded to an engineering analysis and then third would be a recall if it gets that far. NHTSA said it will assess actually smart summons maximum speed, use on public roads and line of sight requirements. The probe will also cover remote control through the phone app, the impact of connectivity delays and the system's performance in unanticipated conditions. One complaint said a model three in Houston in 2023 struck a parked car with a passenger inside.

请记住，美国国家公路交通安全管理局（NHTSA）的三个步骤：第一步是初步评估；第二步，如果需要，升级为工程分析；第三步，如果问题严重，可能会进行召回。NHTSA表示，他们将评估智能召唤功能的最高速度、是否适用于公共道路以及视线要求。调查还将涵盖通过手机应用程序进行远程控制、连接延迟对功能的影响以及系统在意外情况下的表现。有一项投诉称，2023年，在休斯顿，一辆Model 3撞到了一辆停着的车，车里当时有乘客。

But we need to remember that actually smart summon wasn't even released until 2024 in the fall. On that point, NHTSA said that regular smart summon was the subject of 12 separate customer complaints. NHTSA said Tesla had not reported any of the crashes despite rules requiring manufacturers to report crashes involving automated driving systems. This investigation by NHTSA does cover the entire sexy lineup. I'm sure many of you remember years ago, most analysts were saying that Tesla's regulatory credit revenue was about to go to zero. And here we are in 2025. And it may continue to increase already for 2025.

但我们需要记住，实际上智能召唤功能直到2024年秋天才发布。就此话题，美国国家公路交通安全管理局（NHTSA）指出，常规智能召唤功能已经成为12个不同客户投诉的对象。NHTSA表示，尽管法规要求制造商报告涉及自动驾驶系统的事故，但特斯拉未报告任何事故。NHTSA的这项调查涵盖了整个Tesla车型系列。我相信你们中的许多人还记得多年前，大多数分析师都在说特斯拉的监管信用收入将要降为零。而我们现在是在2025年，这一收入可能在2025年继续增加。

Stellantis Toyota Ford Subaru and Mazda are planning to join an EV credit pool with Tesla. The EU's CO2 targets for this year are about 15% lower than 2021 levels and experts say automakers will have to sell at least 20% full EVs while the EV market in Europe is stuck around 14%. Missing the target results in a fine of about $98 per gram of CO2 over the limit per vehicle. According to an estimate by the ACA last year, automakers could face a total of more than $15.5 billion in fines. However, to highlight the uncertainty of these figures, transport and environment said that the fines may be around $1 billion.

Stellantis、丰田、福特、斯巴鲁和马自达计划与特斯拉加入一个电动汽车（EV）积分池。今年欧盟的二氧化碳排放目标比2021年的水平降低了大约15%。专家称，汽车制造商需要至少销售20%的全电动汽车，而欧洲的电动车市场目前停滞在14%左右。如果未能达到目标，每辆车超过限额的每克二氧化碳将面临大约98美元的罚款。根据ACA去年的一项估算，汽车制造商可能面临总额超过155亿美元的罚款。然而，为了强调这些数据的不确定性，运输与环境组织表示，罚款金额可能约为10亿美元。

A European EV analyst said that the rush for automakers to form pools showed automakers were seeking an early backup plan if EV sales don't accelerate. Tesla will be the manager of one of these open pools and the Tesla group is open to new applicants until February 5th. Tesla's regulatory credit sales through the first 3 quarters of 2024 were $2.07 billion. For all 4 quarters of 2023, they came in at $1.79 billion. For all of 2022, they were $1.77 billion. And for all of 2021, they were $1.46 billion. So it's been years since those analysts made those predictions and not only have Tesla's regulatory credits not gone to zero, they're still increasing.

一位欧洲电动车分析师表示，汽车制造商急于组建联盟，显示他们在为电动车销售未能加速做好早期的后备计划。特斯拉将管理其中一个开放联盟，该集团接受新申请的截止日期是2月5日。特斯拉在2024年前三季度的监管信用销售额为20.7亿美元。2023年全年为17.9亿美元，2022年全年为17.7亿美元，而2021年全年则为14.6亿美元。这说明自那些分析师做出预测以来，不仅特斯拉的监管信用销售没有降为零，反而还在增长。

This compared the market company put out the most searched car brands for 2024, just focusing on North America, Tesla took home the number one spot for Canada and the US, whereas the most searched brand in Mexico was Nissan. In 2023, Toyota was the number one brand in 64 countries and that has not changed for 2024. However, Tesla has risen from being the number one brand in 29 countries in 2023 to 34 in 2024. For the past 5 years, Toyota has held a top spot in terms of number of countries where they're the number one search. But Tesla is charging ahead as in 2022, it was not even on the list and as of last year, it's in the number two spot again, with even more countries searching for Tesla the most.

这项比较分析了市场公司发布的2024年搜索最多的汽车品牌，聚焦于北美地区，结果显示特斯拉在加拿大和美国排名第一，而在墨西哥搜索最多的品牌是日产。2023年，丰田在64个国家中位居第一，这一情况在2024年没有改变。然而，特斯拉已经从2023年的29个国家增至2024年的34个国家排名第一。在过去的5年里，丰田一直在多数国家中保持搜索量第一的位置。但特斯拉正在快速追赶，因为在2022年，它甚至没有上榜，而到了去年，它又重新回到了第二名的位置，有更多的国家搜索特斯拉最多。

There was plenty of chatter out there today that Tesla stock was down because of this note from Bank of America, so I wanted to touch on why I don't think that's the case. BOA said that Tesla's latest valuation already reflects long-term growth potential, including Robloxaxes. Tesla is trading at a level that captures much of our base case long-term potential from Corados, Robloxoxes, Optimus, and Energy Generation in storage. In their eyes, they're saying that Robloxes already account for about 50% of Tesla's valuation, and based off today's closing price, that would be about $600 billion.

今天有很多人在讨论特斯拉的股票下跌是因为美国银行的一份报告，所以我想谈谈为什么我不这样认为。美国银行表示，特斯拉当前的估值已经反映了其长期增长潜力，包括Robloxaxes（假设这是某个术语）。特斯拉的交易水平已经涵盖了我们基本情景下的长期潜力，包括Corados、Robloxoxes、Optimus和能源存储。在他们看来，Robloxes已经占到了特斯拉估值的约50%，按照今天的收盘价计算，这大约是6000亿美元。

I'm sure many of you heard gents and say last night at NVIDIA's keynote that autonomous vehicles will be the first multi-trillion-dollar real-world AI industry. Per the usual though, these analysts note, leave us scratching our heads, no different this time around. BOA raised its Tesla price target to $490 from 400. However, they said the execution risk is high, so they downgraded the rating for Tesla to neutral from buy. That implies nearly 25% stock upside in the next 12 months, but they have a neutral rating. And at the end of today's video, I'll explain why I think there's a lot more going on that was a driver of Tesla's stock price today.

我相信很多人昨晚在NVIDIA的主题演讲中听到大家说，自动驾驶汽车将成为第一个万亿美元级的现实世界人工智能产业。但是，和往常一样，分析师们的预测让我们摸不着头脑，这次也不例外。美国银行将特斯拉的目标股价从400美元提高到490美元。然而，他们表示执行风险很高，所以将特斯拉的评级从买入下调到中性。这意味着在接下来的12个月内，特斯拉的股价可能有近25%的上涨空间，但他们的评级却是中性。在今天视频的最后，我会解释为什么我认为导致特斯拉股价波动的背后还有更多因素。

We got the Tesla China weekly number for week 1 of 2025, and it came in at 5,500, which yes, is the best first week ever. Week 1 of quarter 4 came in at 1,800, so quarter over quarter, Tesla China's up over 205%. Week 1 of quarter 1 last year came in at 3,200. Thus, year over year, Tesla China is up 71.8%. The new record quarter for Tesla China was quarter 4 last year, 194,952, meaning Tesla would have to average over 15,788 units per week over the next 12 weeks. And of course, the year to date figures are just the same as the year over year comparison. For this reading, it was about 1100 Model 3s and 4300 Model Wives.

我们得到了2025年第一周特斯拉中国的周数据，达到了5500辆，这是历年来最佳的第一周。第四季度的第一周数据是1800辆，因此环比增长超过205%。去年第一季度的第一周是3200辆，因此同比增长71.8%。去年第四季度是特斯拉中国的新纪录，达到194,952辆，这意味着接下来的12周每周平均需要超过15,788辆。而今年截至目前的数据与去年的同比数据相同。在这次的数据中，大约有1100辆Model 3和4300辆Model Y。

I'm in a data kind of a mood today, so I wanted to quickly share some of AG1's clinical trials. It was shown to enrich the gut microbiome more than doubling the levels of healthy bacteria known to bolster digestion. Specifically, a 2.9x increase in the healthy bacteria in the gut. That relative to the control group. I can tell you right now, if you've taken antibiotics or NSAIDs like ibuprofen or Advil in the past few years, there's a very good chance that your gut biome could absolutely use some help. Last time we talked about how roughly 90% of your serotonin production happens in the gut. Well, in an in vitro study of AG1, they found a 9.8x increase in serotonin production in the gut, taking AG1.

今天我的心情偏向于数据分析，所以我想快速分享一下AG1的一些临床试验结果。研究表明，AG1能显著改善肠道微生物群，将有助于消化的健康细菌水平提高了超过两倍，具体来说是增加了2.9倍，这个结果是相对于对照组而言的。可以说，如果你在过去几年内服用过抗生素或像布洛芬、Advil这样的非甾体抗炎药，你的肠道菌群可能确实需要一些帮助。上次我们提到，大约90%的血清素都是在肠道中产生的。而在AG1的一项体外研究中，发现服用AG1能够使肠道中的血清素产量增加到9.8倍。

We haven't talked about this recently, also in that in vitro study, the powder form of AG1 was seen to digest faster than a multivitamin tablet. Specifically, 4.4x more minerals available for absorption versus a multivitamin tab. And finally, in an observational study with regular use over three months, AG1 was seen to positively impact feelings of calm, energy, and improved digestion. I sit mine throughout the day and it does really seem to help with any afternoon crashing. Plus, it's nice because I can drink it into the evening and not worry about problems falling asleep at night. And as I always say, we're all different, so if you feel like you want to support your energy levels, digestion, focus, etc., I just think it's worth a shot for a few months to see how it works for you. AG1 is still offering that limited edition gift, so if you'd like to support the channel and your own health, you can head to drinkag1.com slash electrified to check it out. Or you can use the QR code on the screen.

我们最近还没聊过这个话题，在某项体外研究中，AG1粉末形式被发现比多种维生素片剂消化更快。具体来说，其矿物质吸收率是多种维生素片的4.4倍。而且，在一项为期三个月的观察性研究中，定期使用AG1对身心平和感、能量和消化都有积极影响。我是一天里慢慢喝的，的确有助于避免下午的疲倦。此外，它的好处是晚上喝也不会影响睡眠。正如我常说的，每个人都不一样，所以如果你想支持自己的能量水平、消化、专注力等，我认为值得尝试几个月，看看效果如何。AG1仍在提供限量版礼品，如果你想支持这个频道和自己的健康，可以前往drinkag1.com/electrified查看。或者可以使用屏幕上的二维码。

I think it'll be valuable to spend a bit of time talking about Nvidia's keynote and what it'll really mean for Tesla's FSD lead going forward. I've seen plenty of commentary out there that I certainly don't agree with, and I want to try to provide some data and facts to back that up. Just know some of my data and facts will be commentary from people that are actually in the machine learning trenches and really know what they're talking about. To start, Gary Black, who you guys know, I have a lot of respect for said Tesla will not be the only automaker with fully autonomous EVs. Tesla is clearly the market leader today with its generalized and scalable approach, but with Nvidia providing synthetic AI generated driving data to any OEM willing to partner with Nvidia, others will get there too. In valuation models, assuming Tesla with winner take all market share are unrealistic. Now it's true, in the long term we're talking decades, Tesla won't be the only automaker with generalized autonomy, but the question is for how long will that be the case? I also agree that valuation models with Tesla taking 100% of the market share are unrealistic. But my main disagreement here is that for any OEM that's just willing to partner with Nvidia, that then means they're also going to have generalized autonomy too. So I just want to try to explain based on everything I've read and learned over the years why I think that is such a gigantic leap.

我认为花一点时间来讨论英伟达（Nvidia）的主题演讲及其对特斯拉FSD（全自动驾驶）未来领先地位的真正意义是非常有价值的。我看到很多评论，其中有不少我并不同意，我打算尝试提供一些数据和事实来支持我的观点。请注意，我所引用的一些数据和事实会来自于那些真正研究机器学习且经验丰富的人士的评论。 首先，大家都知道我很尊敬的Gary Black表示，特斯拉不会是唯一一家拥有全自动驾驶电动车的汽车制造商。特斯拉由于其通用和可扩展的方法，目前显然是市场的领导者，但随着英伟达为任何愿意合作的OEM提供合成AI生成的驾驶数据，其他公司也会达到这个水平。在估值模型中，假设特斯拉能独占市场份额是不现实的。长期来看，可能经过几十年，特斯拉将不再是唯一拥有关联化自主技术的汽车公司，但问题是这种情况会持续多久。我也同意认为特斯拉占据市场100%份额的估值模型不现实。但我主要不同意的是，只要任何OEM愿意与英伟达合作，就意味着他们也能实现通用自动驾驶。 因此，我想基于我多年来阅读和学习到的知识，解释为什么我认为这是一个巨大的飞跃。

First, if you did not see Nvidia's keynote, I really want to focus on one thing they announced and that's their new world foundation models. This offers developers an easy way to generate massive amounts of photoreal physics-based synthetic data to train and evaluate their existing models. Developers can also build custom models by fine-tuning Cosmos WFM's world foundation models. This suite of open models that these legacy OEMs or robotics companies will be licensing from Nvidia means developers can customize the WFM's with data sets like video recordings of AV trips or robots navigating a warehouse. But for our purposes today, just think about Cosmos and how that's going to help these companies generate synthetic data.

首先，如果你没看到英伟达的主题演讲，我特别想强调他们宣布的一件事，那就是他们的新世界基础模型。这个模型为开发者提供了一种简便的方法，可以生成大量基于物理的真实感合成数据，以训练和评估现有模型。开发者还可以通过调整Cosmos WFM的世界基础模型来构建定制模型。这些开放模型套件将由传统OEM或机器人公司向英伟达授权，这意味着开发者可以用视频记录，如自动驾驶汽车的行程或机器人在仓库中的导航过程，来自定义WFM。但对于我们今天的目的，只需想想Cosmos，以及它将如何帮助这些公司生成合成数据。

If you're not that familiar with Nvidia, the easiest way to think about it is like they provide toolkits to these legacy OEMs. So yes, Toyota can go buy these tools specifically built for autonomous driving from Nvidia, but after that purchase, all they have are some tools. They still need a talented team of machine learning engineers to put those tools to use and to actually build the project, which in this case is a model. As Ashok said in this video at CVPR in 2023, Tesla is already using very similar techniques. So that's why we are working on learning a more general world model that can really just represent arbitrary things. So in this case, what we do is we have a neural network that can be conditioned on the past or other things to predict the future. What you're seeing here is purely generated video sequences. Given the past videos, the network predicts some samples from the future, hopefully the most likely sample. And you can see that it is being predicted not just for one camera, but it predicts all the eight cameras around the car jointly. And you can see how the car colors are consistent across the camera as the motion of objects is consistent in 3D. Even though we have not explicitly asked it to do anything in 3D or not, they are baked in any 3D priors. This is just the network understanding depth and motion on its own without us informing it off. So this is super powerful because now you have essentially a neural network simulator that can simulate different futures based on different actions.

如果你对Nvidia不太熟悉，可以简单地把它看作是向传统设备制造商提供工具包的公司。是的，丰田可以从Nvidia那里购买专为自动驾驶设计的工具，但购买后，他们得到的只是一些工具。他们仍然需要一支有才华的机器学习工程师团队来使用这些工具，真正建立项目，在这个案例中就是一个模型。正如Ashok在2023年CVPR的视频中所说，特斯拉已经在使用非常类似的技术。这就是为什么我们正在努力学习一个更通用的世界模型，它可以代表任意事物。所以在这种情况下，我们所做的是让一个神经网络基于过去或其他条件来预测未来。你现在看到的只是生成的视频序列。基于过去的视频，网络会预测未来的一些样本，希望是最有可能的样本。你可以看到它不仅为一个摄像头进行预测，而且是为车周围的八个摄像头一起预测的。你还可以看到，随着物体运动在3D中的一致性，汽车的颜色在摄像头之间也是一致的。即使我们没有明确要求它进行任何3D相关的操作，也没有嵌入任何3D先验知识，这仅仅是网络自己理解深度和运动而已。因此，这非常强大，因为现在你基本上拥有了一个神经网络模拟器，可以根据不同的动作模拟出不同的未来场景。

My point Tesla has been using synthetic data and simulations to augment its library of real world video data for years. Late last year, Andrei Carpathi said in the context of LLMs, I think synthetic data is absolutely the future. We're not going to run out of data, but we do need to be careful. Right. But what about the argument like in that domain that that was easier when we were taking internet data and we're out of internet data. And so the questions are really around synthetic data or more expensive data collection. So I think that's good point. So that's where a lot of the activity is now in LLMs. So the internet data is not the data you want for your transformer. It's like a nearest neighbor that actually gets you really far surprisingly.

我的观点是，特斯拉多年来一直在使用合成数据和模拟技术来扩充其真实世界视频数据库。去年年底，Andrei Carpathi在谈到大型语言模型（LLMs）时表示，他认为合成数据绝对是未来。我们不会耗尽数据，但我们确实需要小心对待。然而，有人提出了一个问题：当我们在使用互联网数据时，这似乎相对容易，但现在我们正在耗尽互联网数据。因此，问题就转向了合成数据或更昂贵的数据收集。我认为这是个好点子。所以，现在很多关于LLM的活动都集中在这方面。互联网数据并不是你想要用在转换器上的数据，但它作为“最近邻”的替代品实际上能带来惊人的进展。

But the internet data is a bunch of internet web pages. It's just like what you want is the inner thought monologue of your brain. The trajectories in your brain. The trajectories in your brain as you're doing problem solving. If we had a billion of that, like AGI is here, roughly speaking, to a very large extent. And we just don't have that. So where a lot of activity is now, I think, is with the internet data that actually gets you really close because it just so happens that internet has enough of reasoning traces in it and a bunch of knowledge. And the transformer just makes it work. Okay. So I think a lot of activity now is around refactoring the data set into these inner monologue formats. And I think there's a ton of synthetic data generation that's helpful for that. How important do you think the synthetic data piece is? It's incredibly important. I think it's the only way we can make progress is we have to make it work. I think with synthetic data, you just have to be careful because these models are silently collapsed as one of the major issues. So if you go to chat, TPT, and you ask it to give you a joke, you'll notice that it only knows like three jokes. It gives you like one joke, I think most of the time, and sometimes it gives you like three jokes. And it's because the models are collapsed and it's silent.

互联网数据是一堆网页，就像你希望能够获得自己内心思考的独白，即大脑处理问题的思路轨迹。如果我们拥有亿万个这样的轨迹，大致来说，强人工智能（AGI）就已经实现了。然而，我们目前还没有这种资源。所以，我认为现在很多研究活动集中在利用互联网数据，因为互联网恰好有足够的推理痕迹和大量知识。而变压器（Transformer）模型正好能够很好地利用这些。现在，很多活动都在重新构建数据集，使其更接近于这些内心独白的格式。我认为合成数据生成在这方面很有帮助。你觉得合成数据有多重要？非常重要。我认为这是我们取得进展的唯一途径，我们必须让它发挥作用。不过在使用合成数据时需要小心，因为这些模型有时会出现静默崩溃的问题。比如，当你与聊天机器人互动，并要求它讲个笑话时，你会发现它只知道三四个笑话。大多数时候，它会给你讲一个笑话，有时会给你三个。这是因为模型崩溃了，而且是默默地崩溃的。

So when you're looking at any single individual output, you're just seeing a single example. But when you actually look at the distribution, you'll notice that it's not a very diverse distribution. It's silently collapsed. When you're doing synthetic data generation, this is a problem because you actually really want that entropy. You want the diversity and richness in your data set. Otherwise, you're getting collapsed data sets. And you can't see it when you look at any individual example, but the distribution is has lost a ton of interest and richness. And so it silently gets worse. And so that's why you have to be very careful and you have to make sure that you maintain your entropy in your data set.

当你观察某个单一输出时，你只能看到一个单独的例子。然而，当你查看整个分布时，你会注意到这个分布并不多样化，它悄然崩溃了。在进行合成数据生成时，这是个问题，因为你实际上非常需要混乱度（熵）——你需要数据集中的多样性和丰富性。否则，你将得到崩溃的数据集。单从某个个别例子来看，你是无法察觉的，但总体分布已经失去了大量趣味性和丰富性，因此它会悄悄地变得更糟。这就是为什么你必须非常小心，并确保保持数据集中的混乱度（熵）。

There's the context of that conversation for LLMs, but many of those principles will still translate when it comes to autonomy. In March of last year, Elon said the vector sum of humans on the team usable compute and unique access to data define AI competitiveness. It's remarkable how quickly we run out of human created data reality itself and synthetic data for the win. In April of last year, Elon said two sources of data scale infinitely synthetic data, which has an is it true problem and a real world video, which does not you heard a car path. He talked about that model collapse problem when it comes to using too much synthetic data.

翻译如下： 在讨论大型语言模型（LLM）的背景下，我们可以看到许多原则在谈到自主性时依然适用。去年三月，埃隆（马斯克）表示，团队中人类的向量总和、可用的计算资源和独特的数据访问定义了人工智能的竞争力。我们很快就会耗尽人类创造的数据，而现实本身和合成数据成为关键。去年四月，埃隆提到有两种数据来源可以无限扩展：合成数据（存在真实性问题）和真实世界的视频（不存在这种问题）。你可能听说过一条路径，他谈到了使用太多合成数据时出现的模型崩溃问题。

And I found this paper exploring synthetic data for AI and autonomous systems. One of the most prominent risks with using synthetic data is called the reality gap. This refers to the subtle differences between the synthetic data and the real world. Sophisticated machine learning models often learn to exploit small discrepancies, making simulated environments difficult to learn from. In other words, if synthetic data is not simulated properly, it can run into issues of not being able to fully replicate the complex and chaotic physics of the real world, and may fail to properly capture the unexpected shifts or one off cases that emerge in real world data.

我发现了一篇论文，讨论了合成数据在人工智能和自动系统中的应用。使用合成数据的一个突出风险被称为“现实差距”。这指的是合成数据与现实世界之间的细微差异。复杂的机器学习模型往往会利用这些小的差别，这使得从模拟环境中学习变得困难。换句话说，如果合成数据模拟得不好，它可能无法完全复制现实世界中复杂而混乱的物理现象，并且可能无法正确捕捉到现实数据中出现的意外变化或偶尔的案例。

Another way to think about this model degradation if you train with too much synthetic data is the telephone game. Ordinarily, every time the message is passed from one person to the next, parts of the actual message are lost. The same thing can happen with AI models. They make mistakes, they hallucinate and over time, if AI is trained on AI and synthetic data, it can actually lose sight of the real world or the base that it's starting from. A 2023 study at Rice and Stanford found that over reliance on synthetic data during training can create models whose quality or diversity progressively decrease. Sampling bias or poor representation of the real world causes a model's diversity to worsen after a few generations of training.

如果你用过多的合成数据训练模型，可以用"传话游戏"来理解这种模型劣化的现象。通常情况下，每次信息从一个人传递到下一个人，原本的信息都会有些丢失。同样的事情也可能发生在人工智能模型上。模型可能会出错、产生幻觉，随着时间的推移，如果人工智能在训练时过于依赖人工智能生成的数据和合成数据，它可能会逐渐偏离现实世界或最初的基础。赖斯大学和斯坦福大学在2023年的一项研究发现，过度依赖合成数据进行训练会导致模型的质量或多样性逐步下降。抽样偏差或对现实世界的代表性不足会导致模型的多样性在多轮训练后恶化。

And guess what, one of the ways to combat that model degradation with synthetic data is by balancing it out with real world data. I hope that's at least a little bit of context to help explain that these legacy automakers having access to Nvidia's synthetic data is by no means a panacea to solving FSD. It absolutely has a place as a tool to help solve the problem, but the core foundation is always going to be real world data because it's reliable. And one piece I don't see many people talking about, what about all of the intervention data that Tesla has that really helps to train the system. In short, Tesla is closing in on a decade being laser focused on solving this problem with real world data, synthetic data, simulation, a lot of compute, the most talented engineers, the best inference in the world, and yet the problem is still not solved, but they're close.

猜猜看，对抗模型退化的一个方法是用真实世界的数据来平衡合成数据。我希望这至少能提供一些背景信息，以帮助解释这些传统汽车制造商能够使用Nvidia的合成数据并不是解决全自动驾驶难题的万能钥匙。合成数据确实是可以帮助解决问题的一种工具，但核心基础总是依赖于真实世界的数据，因为真实数据是可靠的。还有一个我看到不多人讨论的方面，那就是特斯拉拥有的大量干预数据，这些数据对系统的训练非常有帮助。简而言之，特斯拉近十年来一直专注于用真实世界的数据、合成数据、模拟、大量计算资源、最有才华的工程师和世界上最好的推理能力来解决这个问题。尽管问题仍未完全解决，但他们已经很接近了。

So now to think that legacy OEMs that are mostly still at the starting line can just have access to this synthetic data from Nvidia and all of the sudden be on Tesla's level with solving generalized FSD. To me is totally insane, even to think that they're going to figure it out in the next few years just does not make any sense. Elon Musk is a customer of yours. And Tesla, their theory or practice is based on real world data gathered through vision. Does the synthetic data underpinning of cosmos kind of contradict that? It doesn't replace at augments. And so you're going to, you should collect as much world data as you can. Of course collecting world data is very expensive. And Elon has a great advantage because number one, his AI factory for his cars is fantastic. It has a lot of Nvidia gear in it.

所以现在，认为那些大多数仍在起步阶段的传统汽车制造商可以通过英伟达的合成数据就突然达到特斯拉在解决通用完全自动驾驶水平的想法，是完全不现实的。即使认为他们在未来几年就能搞清楚这一点也是毫无意义的。埃隆·马斯克是你的客户。特斯拉的理论或实践是基于通过视觉收集的真实世界数据。那么，宇宙的合成数据基础是否与此相矛盾？它不是替代而是增强。所以你应该尽可能多地收集真实世界的数据。当然，收集真实世界数据是非常昂贵的。埃隆有很大的优势，因为首先，他的汽车人工智能工厂非常出色，里面有很多英伟达的设备。

His AV algorithms is incredible. It's the best in the world. And he has a very large fleet of cars on the road that allows him to collect a lot of data. And so I think he has just a phenomenal position and he's been working on this for a long time. And so he's going to be in a great position to take advantage of it. I know Elon's attitude towards AI. And he's very optimistic about its future. And obviously he's working on some of the most important AI areas. XAI is working on foundation, cognitive intelligence AI, Tesla is working on Thomas vehicles, and optimists are human robotics. These three areas of AI are the three most important areas of AI. And so I think he's working on exactly the right things.

他的自动驾驶算法非常厉害，是世界上最好的。他有一个庞大的车队在路上行驶，这使他能够收集大量数据。所以我认为他在这个领域处于一个非常优秀的位置，而且他已经在这个领域努力了很长时间，因此他有很好的优势。我了解Elon对人工智能的态度，他对其未来非常乐观，而且显然他正在从事一些最重要的人工智能领域。XAI专注于基础和认知智能AI，特斯拉专注于自动驾驶汽车，而Optimus专注于人形机器人。这三个领域是人工智能中最重要的三个领域。所以我认为他正在研究的方向完全正确。

When it comes to this Elon and Jensen takeover of the world, I think Phil Bizell does a great job of explaining it. And you guys know I like highlighting people that deserve more attention. In my opinion, Phil is one of those people. He said Tesla goes hard with hardcore engineering and produces real world AI products, namely FSD for Roblotaxian optimists. Nvidia sells massive compute to Tesla for training its real world AI solutions. Watches what Tesla does with their chips does a copy light version of Tesla's work in the form of SDKs, which is software development kits, convinces OEMs like Toyota that Nvidia SDKs and platforms are the basis for their own AI dreams sells those OEMs tons of chips for training inference and now generative compute in the form of simulation Tesla demands more chips to push faster against the perceived competition and then rinse wash repeat. Plus you guessed it, these legacy OEMs that will buy inference and SDKs and compute from Nvidia will be paying those 50% plus margins to Nvidia. So right out of the gate just when it comes to the hardware and software suite for autonomy, legacy OEMs will be at a much higher price point.

关于埃隆和黄仁勋接管世界的讨论，我觉得菲尔·比泽尔对此的解释非常出色。大家都知道，我喜欢突出那些值得更多关注的人，而在我看来，菲尔就是其中之一。他说，特斯拉通过硬核工程大力发展，推出了真实世界的AI产品，尤其是为乐观的机器人出租车支持者提供的FSD（全自动驾驶）。而英伟达则向特斯拉出售大量计算能力，用于训练其真实世界的AI解决方案。英伟达观察特斯拉如何使用他们的芯片，然后以软件开发工具包（SDK）的形式做出简化版本的仿制，并说服像丰田这样的原始设备制造商（OEMs）相信英伟达的SDK和平台是他们自己AI梦想的基础，向这些OEM出售大量芯片，用于训练推理和现在的生成计算（如仿真）。特斯拉则需要更多芯片以更快地迎接竞争，而这一过程就不断重复。没错，那些将从英伟达购买推理和SDK以及计算的传统OEMs，将为英伟达支付超过50%的利润率。因此，单在自动化的硬件和软件套件方面，传统OEMs就处于更高的价格点。

Then when you layer in the skill of building EVs profitably at scale, the challenge becomes that much harder. And guess what, these legacy OEMs still all need to buy these tools from Nvidia, get them integrated into their vehicles which we know is going to take three to five years and then they need to get that fleet on the road to begin collecting real world data. And at that point they can begin to augment that with some of this synthetic data. The AV revolution has arrived. After so many years with Waymo success and Tesla's success, it is very, very clear autonomous vehicles has finally arrived. Elon shared a post from Sawyer saying correct in which Sawyer said this seems like it's aimed at legacy automakers that have virtually zero real world video self driving data collection. They're years behind Tesla and have no real shot at catching up. He said synthetic driving data is kind of like using chat GBT you might trust what you see is true, but you often can't be entirely certain without further validation. In contrast, real world video driving data is reliable and represents true legitimate scenarios as they occurred. It's the superior method for solving self driving as long as you have enough video data collection.

当你再加上大规模、盈利地生产电动汽车的技能时，挑战就更加艰巨了。你猜怎么着，这些传统的汽车制造商仍然需要从Nvidia购买这些工具，并将它们集成到他们的车辆中。而这过程，众所周知，需要三到五年时间。然后，他们需要把这些车队投放到道路上，以开始收集真实世界的数据。在那时，他们才可以开始用一些合成数据进行补充。自动驾驶革命已经到来。经过多年的努力，Waymo和Tesla的成功让人非常明确地意识到：自动驾驶汽车终于到来了。Elon分享了Sawyer的一篇帖子，Sawyer说，这似乎是针对那些几乎没有真实世界视频数据收集的传统汽车制造商。它们落后于Tesla好几年，而且几乎没有机会赶上。他说，合成驾驶数据有点像使用ChatGPT，你可能相信你看到的是真的，但你常常需要进一步验证才能完全确定。相比之下，真实世界的视频驾驶数据是可靠的，代表了真实发生的合法场景。只要你有足够的视频数据收集，它就是解决自动驾驶问题的更好方法。

When it comes to Jensen saying that self driving vehicles will be the first multi trillion dollar robotics industry, Elon replied saying Jensen is right. May of 2024 Jensen said Tesla is far ahead in self driving cars. The story is still the same even after last night. In theory, yes, Nvidia's advancements should help legacy OEMs to solve generalized autonomy a bit faster. But the way I see it, I don't see any legacy OEM solving generalized autonomy anytime before 2030. The only way that would happen is if they choose to license FSD from Tesla and get to work effectively now. And don't misunderstand me what Nvidia is doing is incredible and impressive and will be very valuable to the industry. The problem is we know too much about that legacy industry we've been covering now for the past few years. They move slow, they don't have software talent, and they're not making EVs profitably. And nothing Nvidia said last night changes any of that

关于詹森说自动驾驶汽车将成为第一个万亿美元级机器人产业，埃隆回应说詹森是对的。2024年5月，詹森表示特斯拉在自动驾驶汽车方面遥遥领先。即便在昨晚过后，这个说法依旧成立。从理论上讲，英伟达的进步应该能帮助传统汽车制造商更快解决通用自动驾驶的问题。但在我看来，至少在2030年之前，我觉得没有哪个传统车企能够解决这个问题。唯一的可能是他们选择从特斯拉获得自动驾驶软件许可，并立即有效开展工作。请不要误解，我认为英伟达所做的事情确实令人惊叹，将对行业非常有价值。问题是，我们对这个传统行业已经了解太多了，在过去几年我们一直在关注。他们发展缓慢，缺乏软件人才，电动车也没能盈利。而昨晚英伟达所说的一切并没有改变这些。

Tesla supplier Panasonic plans to eliminate its supply chain dependence on China for EV batteries made in the US likely in response to the fact that Trump has vowed to impose tariffs of 10% on global imports into the US along with a 60% tariff on Chinese goods. In November, he specifically pledged a 25% tariff on imports from Canada and Mexico. Panasonic said we do have some Chinese supply, but we don't have a lot. The bulk of the raw materials for Panasonic energy's US made batteries come from overseas suppliers. The bulk of the raw materials for Panasonic's US made batteries come from overseas suppliers, including ones from Canada. It's speculation on my part, but this could cause a slight bump in cost of goods sold for Tesla in the near term.

特斯拉的供应商松下计划减少对中国供应链的依赖，以便在美国生产电动汽车电池。这可能是因为特朗普表示将对全球进口征收10%的关税，并对中国商品征收高达60%的关税。而在11月，特朗普特别承诺对加拿大和墨西哥的进口征收25%的关税。松下表示，我们确实有一些来自中国的供应，但并不多。松下在美国生产的电池所需的原材料大多来自海外供应商，其中包括来自加拿大的。虽然这只是我的猜测，但这可能会导致特斯拉在短期内的商品销售成本略有上升。

At CES, Honda gave us a bit more information on their zero series models that are set to enter production in 2026 in Ohio. We don't have the official names, but a midsize SUV is set to come first. They're both supposedly going to have level 3 automated driving. The plan is to launch in the US first and then head to Japan and Europe. We really don't have much information about the vehicles other than the steering wheel, suspension and brakes will be controlled by a steered by wire system. And the word is this sporty saloon is going to follow the SUV in the latter part of 2026. Tesla has asked a court in Sweden to ensure that the country's transport agency provides access to license plates that are currently blocked by postal workers in a wider labor conflict. While the postal blockade makes access to license plates more difficult, Swedish media has reported that Tesla has found ways to circumvent the unions by asking car buyers to order plates themselves. Tesla has already lost similar appeals in other courts, so we'll see how this one goes. Speaking of Honda, they also have their partnership with Sony and they announced that their vehicle starting at over $102,000 that will only be available in California is set to arrive in the middle of 2026. Then in 2027, a less expensive version will be priced at just under $90,000. Right now, you can reserve in a feel of one with a $200 deposit. It'll have a 91kWh battery pack with an EPA estimated range of 300 miles. It will have a NAC support as standard, but the charge rate is only up to 150kWh. For now, the companies have not announced when and if it'll be offered in other states outside of California. So yeah, color me quite skeptical with this one.

在CES上，本田透露了一些关于他们零系列车型的信息，这些车型计划于2026年在俄亥俄州投产。我们还不知道这些车型的正式名称，但首先会推出一款中型SUV。传闻这两款车型都将具备三级自动驾驶能力。计划是先在美国推出，然后再进入日本和欧洲。除了方向盘、悬挂和刹车将由线控系统操控之外，我们对这些车辆的其他信息了解不多。据说这款运动轿车将在2026年后期跟随SUV发布。 特斯拉在瑞典请求法院确保该国的运输机构提供目前被邮政工人封锁的车牌访问权限，这涉及更广泛的劳动冲突。虽然邮政封锁使得车牌获取更加困难，但瑞典媒体报道，特斯拉已经找到绕过工会的方法，要求购车者自己订购车牌。特斯拉在其他法院的类似上诉已经失败，所以我们将看看这次结果如何。 说到本田，他们还和索尼合作，宣布将在2026年中期推出只在加利福尼亚州销售的车辆，起价超过10.2万美元。然后在2027年，将推出一个价格较低的版本，售价不到9万美元。现在，你可以通过支付200美元订金预订一辆，车辆将配备91千瓦时的电池组，EPA预计续航里程为300英里。它将标配NAC支持，但充电速度仅为150千瓦。目前，这两家公司尚未宣布是否会在加利福尼亚州以外的其他州销售。所以对这点我持怀疑态度。

Tesla stock closed the day at $394.36 down 4.06% while the NASDAQ was down 1.89%. As you can see though, it was a red day across the board, so it's not just like Tesla was down because of that BOA downgrade. The volume for Tesla was 14% below the average. As promised, just some quick context for what might also be going on in the markets and with Tesla, you know I've always said that the macro drives things first. Well, since the Fed pivot just over 100 days ago, the 10-year treasury yield has actually gone up over 1%. So the Fed has been cutting rates, but the 10-year is actually up 1% over the past 3 months. You can interpret this in different ways, but it feels like the market is not buying this Fed pivot and is actually still pricing in the fact that inflation may not be done. I know that most people here do not want to hear about the macro space, but I need you guys to know it's crucially important to the direction of stocks. And trends like this do not usually happen and in those cases, the market is telling you something. Don't forget to check out AG1 links below if you're interested and as always, thank you very much for considering supporting the channel in that way. Hope you all have a wonderful day. Please like the video if you did, you can find me on X-linked below and a huge thank you to all of my Patreon supporters.

特斯拉的股价收盘时为394.36美元，下跌4.06%，同时纳斯达克指数下跌1.89%。正如你所看到的，整体市场都呈现下跌趋势，所以这不仅仅是因为美国银行下调评级导致特斯拉股价下跌。特斯拉的交易量比平均水平低了14%。如我承诺的，我会简要说明一下市场和特斯拉可能正在发生的情况。我始终认为宏观经济因素是首要影响。自从美联储在大约100天前转向以来，10年期国债收益率实际上已经上涨了超过1%。虽然美联储在降息，但10年期国债收益率在过去三个月中却上涨了1%。对此可以有不同的解读，但市场似乎并不买账美联储的转向，实际上仍然在考虑通胀可能尚未结束。我知道大多数人不太想听宏观经济方面的分析，但我需要大家知道这是决定股票走向的重要因素。这样的趋势并不常见，而在这些情况下，市场其实在向你传递信息。假如你感兴趣，请查看下面的AG1链接，并且一如既往地感谢大家以这种方式支持本频道。希望你们度过美好的一天。如果喜欢这个视频，请点赞，你可以在下面链接中找到我在X平台的账号，也非常感谢所有Patreon的支持者。