This Would Be Huge for Tesla / Elon's Tesla Posting Spree / Tesla's 5 New Patents ⚡️
发布时间 2026-01-20 00:53:49 来源
以下是视频内容的中文摘要，包括提到的所有新闻项目：
* **开场与赞助人：** Dylan Lumis 欢迎听众，并向新赞助人 ScooterR 致谢。
* **加拿大与中国之间的电动汽车关税协议：**
* 特斯拉有望从加拿大与中国之间新的电动汽车关税协议中显著受益。
* 拥有北美认证和分销能力的公司（特斯拉、沃尔沃、极星）将率先受益。
* 特斯拉在2023年从中国向加拿大进口了4.4万辆汽车。
* 这项新协议包含了100%的关税税率，但加拿大交通部将在八周内认证新的中国电动汽车。
* 这为特斯拉带来了供应链灵活性，允许从上海向加拿大出口更便宜的汽车。
* 特斯拉很可能占据4.9万辆配额中的大部分，预计该配额将增长到7万辆。
* **XAI 人类模拟器与特斯拉计算机：**
* **Suleiman Gory（前 XAI 技术人员）采访：** 讨论了与 MacroHard 合作构建人类模拟器，以及部署100万个模拟器需要100万台计算机的挑战。
* **特斯拉汽车作为计算平台：** 特斯拉车载计算机（特别是 HW4）在运行模型和人类模拟器方面“资本效率高”，可能比云虚拟机（AWS、Oracle）或专用硬件便宜得多。
* **分布式计算网络：** XAI 旨在利用现有特斯拉汽车网络启动（北美有400万辆，其中2/3到一半配备 HW4，且70-80%处于闲置/充电状态）。
* **车主报酬：** XAI 可以向车主支付费用，以租用其车载计算机的时间，可能覆盖租约付款，从而运行这些“数字擎天柱”人类模拟器。
* **数字擎天柱（Digital Optimus）：** 模仿人类执行的所有数字任务（鼠标、键盘输入），类似于实体擎天柱处理物理任务的方式。
* **Suleiman Gory 的离职：** Suleiman 发帖称已离开 XAI；猜测他因过早泄露这些计划而被解雇。
* **内部测试：** XAI 已经在内部测试这些模拟器，“就像它们是人类员工一样”。
* **路线图：** 该概念被视为一个合法的路线图，而非遥不可及的设想，特斯拉车辆可能提供大规模部署能力。
* **应用领域：** 可能彻底改变数据录入、客户支持、研究、编码、SaaS、网页浏览等领域。
* **CyberCab 和 Robotaxi 最新动态：**
* **CyberCab 行踪：**
* **芝加哥：** 被发现在测试带有后置摄像头自清洁器（喷水器）的功能。
* **奥斯汀（Josh West）：** 被发现没有侧后视镜；希望有新的联邦标准来避免各州法规的差异。特点是新的尾灯（扰流板上有两个独立灯，下方也有灯）。
* **马萨诸塞州：** 被发现，使其成为第4或第5个州（连同纽约州、德克萨斯州、加利福尼亚州、伊利诺伊州）。
* **奥斯汀 Robotaxi 车队扩张：** 奥斯汀的一个大型停车场被发现“停满了”准备就绪的 Robotaxi 车辆，预示着即将到来的“涌入”。
* **高需求：** Joe Tett Meyer 报告称，由于需求旺盛，在奥斯汀无法打到 Robotaxi，再次强调了增加车辆的需求。
* **Robotaxi 追踪器：** 非官方追踪器显示奥斯汀车队有46辆车。
* **美国国家公路交通安全管理局（NHTSA）事故报告（2023年第四季度）：**
* 旧金山和奥斯汀合计数据：10月2起事故，11月4起，12月3起。
* 大多数事故涉及骑自行车的人，但绝大多数未造成人员受伤（仅一例轻伤）。
* 常见细节：车辆速度为0英里/小时，气囊未弹出，车辆未被拖走。
* 报告未说明过错方，但发言人推断，鉴于车辆速度为0英里/小时，骑自行车的人可能经常是过错方。
* Ethan 的数据包括所有 ADS/ADAS，但已筛选出商业用途数据。
* **特斯拉芯片和人工智能芯片战略：**
* **AI5 和 Dojo 3：** 埃隆宣布 AI5 芯片设计进展顺利，特斯拉将重启 Dojo 3 的工作。
* **招聘：** 埃隆正在积极招聘对开发“全球产量最高的芯片”感兴趣的人才。
* **AI5 的重要性：** 埃隆表示解决 AI5 对特斯拉来说是“存亡攸关”的问题，几个月来一直占用他的周六时间。它被描述为“非常强大”，性能大致相当于 Hopper 级别（单片系统芯片），其双核性能则可比肩 Blackwell，但成本更低、能效更高。
* **Dojo 3 重启：** Dojo 3 不是一款新芯片，而是一种“架构或理念”，使用大量 AI6 片上系统进行训练和推理，模糊了两者之间的界限。
* **过往声明：** 埃隆此前（5个月前）曾表示 Dojo 2 是一个“进化的死胡同”，所有路径都汇聚到 AI6。这表明特斯拉的芯片战略正在快速变化。
* **未来人工智能芯片：** AI5 接近完成，AI6 处于早期阶段，AI7 和 AI9 的目标是9个月的*设计*周期（而非车载部署周期）。
* **AI 芯片能力：**
* **AI4：** 预计将实现“远超人类”的自动驾驶安全水平。（CyberCab 将搭载 AI4 推出）。
* **AI5：** 将使汽车“近乎完美”，并“大大增强擎天柱”。（预计最早在2027年末）。
* **AI6：** 用于擎天柱和数据中心。
* **AI7 和 Dojo 3：** 将是“基于空间的AI计算”。
* **HW3 车主：** 特斯拉在搭载 AI4 的无人监督驾驶广泛部署且其能力被充分理解之前，不会就 HW3 升级做出决定。
* **AI4 的充足性：** 埃隆认为 AI4“完全有能力解决无人监督驾驶”，并确认根据目前的表现，“无人监督驾驶不需要更好的硬件”。
* **特斯拉的“数学作弊码”专利：**
* **RopE（旋转位置编码）：** 2024年1月15日公布的五项专利中的关键组成部分。
* **功能：** 允许8位低功耗芯片通过“更粗糙、更简单的数字”实现高精度操作（如记住顺序/位置），从而运行优秀的32位AI模型。它涉及对数压缩（例如，log(1M) = 6）。
* **优势：** 汽车/机器人可以用更低的功耗更长时间地记住世界（例如30秒），从而在边缘（车内）实现AI，而无需大型电池或高端芯片，并使擎天柱能在低功耗下工作数小时。
* **战略意义：** 这些专利确保特斯拉在战略上独立于英伟达的 CUDA 生态系统，实现双代工策略（三星和台积电），并刻意造成计算能力的过剩，可能将其闲置车队转变为一个分布式推理云，在效率上与 AWS 竞争。
* **软件突破：** 发言人强调，现在软件突破比硬件本身更能推动效率，解决特定瓶颈，以便在现有8位芯片上运行 Transformer 级别AI。
* **Dojo 的演变：** 前 Dojo 负责人 Clive Chan 指出 Dojo 1 的雄心壮志但执行面临挑战。埃隆确认，如果 Dojo 3 符合旧有理念，将是“疯狂的”，并且是“太空AI”所必需的。
* **精简的芯片计划：** 特斯拉正转向一种单一芯片，可同时处理推理和训练，并实现大规模生产，效率由软件突破驱动。这使他们相对于从英伟达购买通用芯片的公司具有优势。
* **特斯拉晶圆厂/芯片制造：**
* **“特斯拉超级晶圆厂（Terra Fab）”：** 埃隆对建造特斯拉超级晶圆厂的想法以🎯表情符号回应，表明这是“实现超高产量的唯一选择”。
* **需求与供应：** 台积电/三星目前的产能可能足以应对保守的增长，但埃隆每年2000万辆汽车和数十亿机器人的愿景，产生的需求“远超当前的全球供应”。
* **内部生产：** 到2029-2030年，内部生产被认为是不可避免的，以实现控制、降低成本、确保安全和规模化。
* **即将动工：** 如果埃隆的言论并非只是谈判策略，那么超级晶圆厂的奠基仪式应在今年举行，因为一座晶圆厂从建设到投产需要3-5年时间。
* **台积电亚利桑那州扩建：**
* 台积电在亚利桑那州的扩建“超出预期”。
* 到2035年，美国在全球芯片生产中的份额可能从0%增至25%。
* 台积电第二座晶圆厂已竣工（预计2028年投产），第三座晶圆厂已于去年4月开工。
* 台积电最近又收购了900英亩土地，并暗示将进一步增加产能。
* 台积电将在美国建造至少五座新的半导体工厂，作为贸易协议的一部分，美国将台湾的基线关税从20%降至15%，而台湾承诺在美国科技领域投资2500亿美元。
* 尽管如此，埃隆认为目前的扩张仍不足够。
* **其他新闻：**
* **Grok 的巨像超级计算机：** 现已投入运行，是“世界上第一个千兆瓦级训练集群”，计划在4月升级至1.5千兆瓦。
* **Model Y 加拿大交付：** 加拿大许多 Model Y Performance 预订客户的交付时间被推迟，从2月推迟到最晚7月，没有给出官方原因。
* **特斯拉餐厅超级充电站：** 全球使用率最高的超级充电站（每月4.3万次充电会话或1.5吉瓦时），80个充电桩平均每天每个桩18次充电会话。从中获得的经验将应用于未来的超级充电站设施。
* **爱达荷州 Pi Safe 超级充电站：** 一家私营公司正在爱达荷州以白标签形式运营特斯拉超级充电站，设计融入了当地树种，旨在振兴小镇。
* **埃隆的股票活动：** 埃隆证实他大约三年没有出售特斯拉股票，去年购买了价值10亿美元的股票。
* **市场休市：** 由于市场休市，特斯拉股价没有变化。
Here's a summary of the video transcription, including every single news item mentioned:
* **Intro & Patrons:** Dylan Lumis welcomes listeners, giving a shout-out to new patron ScooterR.
* **Canada EV Tariff Deal with China:**
* Tesla is expected to benefit significantly from Canada's new EV tariff deal with China.
* Companies with North American certification and distribution (Tesla, Volvo, Polestar) will benefit initially.
* Tesla imported 44,000 cars from China into Canada in 2023.
* The new deal includes a 100% tariff rate, but Transport Canada will certify new Chinese EVs within eight weeks.
* This gives Tesla supply chain flexibility, allowing cheaper exports from Shanghai to Canada.
* Tesla will likely take a large portion of the 49,000-unit quota, which is expected to grow to 70,000 units.
* **XAI Human Emulators & Tesla Computers:**
* **Suleiman Gory (ex-XAI Technical Staff) Interview:** Discussed building a human emulator with MacroHard and the challenge of deploying 1 million emulators needing 1 million computers.
* **Tesla Cars as Compute:** Tesla car computers (especially HW4) are "capital efficient" for running models and human emulators, potentially much cheaper than cloud VMs (AWS, Oracle) or dedicated hardware.
* **Distributed Compute Network:** XAI aims to bootstrap off the existing Tesla car network (4 million in North America, with 2/3 to half having HW4, and 70-80% idle/charging).
* **Payment for Owners:** XAI could pay owners to lease time on their car's computer, potentially covering lease payments, to run these "digital Optimus" human emulators.
* **Digital Optimus:** Mimics all digital tasks a human does (mouse, keyboard inputs), similar to how physical Optimus handles physical tasks.
* **Suleiman Gory's Departure:** Suleiman posted he left XAI; speculation suggests he was fired for prematurely revealing these plans.
* **Internal Testing:** XAI has already been testing these emulators internally "as if they were human employees."
* **Roadmap:** The concept was presented as a legitimate roadmap, not a far-off idea, with Tesla vehicles potentially providing deployment at scale.
* **Applications:** Could revolutionize data entry, customer support, research, coding, SaaS, web browsing, etc.
* **CyberCab & Robotaxi Updates:**
* **CyberCab Sightings:**
* **Chicago:** Spotted testing with a rear camera self-cleaner (sprayer).
* **Austin (Josh West):** Seen without side mirrors; hope for new federal standards to avoid state-by-state rules. Features new taillights (two individual lights in the spoiler, and lights below).
* **Massachusetts:** Spotted, making it the 4th/5th state (along with NY, TX, CA, IL).
* **Austin Robotaxi Fleet Expansion:** A large lot in Austin was seen "filled" with Robotaxi-ready vehicles, indicating an imminent "influx."
* **High Demand:** Joe Tett Meyer reported being unable to get a Robotaxi in Austin due to high demand, reinforcing the need for more vehicles.
* **Robotaxi Tracker:** Unofficial tracker shows Austin's fleet at 46 vehicles.
* **NHTSA Incident Reports (Q4 2023):**
* Figures for San Francisco and Austin combined: 2 incidents in October, 4 in November, 3 in December.
* Most incidents involved cyclists, but the majority resulted in no injuries (only one minor injury).
* Common details: cars going 0 mph, airbags not deployed, car not towed.
* Reports don't state fault, but the speaker infers that cyclists may often be at fault given the 0 mph speed of the cars.
* Ethan's data includes all ADS/ADAS but is filtered for commercial use.
* **Tesla Silicon & AI Chip Strategy:**
* **AI5 & Dojo 3:** Elon announced AI5 chip design is in good shape, so Tesla will restart work on Dojo 3.
* **Recruiting:** Elon is actively recruiting for those interested in working on "the highest volume chips in the world."
* **AI5 Importance:** Elon stated solving AI5 was "existential" for Tesla, consuming his Saturdays for months. It's described as "very capable," "roughly hopper class" (single SoC) and "blackwell as dual," but significantly cheaper and more power-efficient.
* **Dojo 3 Revival:** Dojo 3 is not a new chip but an "architecture or a philosophy" using a large number of AI6 system-on-chips for training and inference, blurring the lines between the two.
* **Past Statements:** Elon had previously stated (5 months ago) that Dojo 2 was an "evolutionary dead end" and all paths converged to AI6. This indicates rapid changes in Tesla's chip strategy.
* **Future AI Chips:** AI5 is almost done, AI6 is in early stages, with AI7 and AI9 aiming for a 9-month *design* cycle (not in-car deployment cycle).
* **AI Chip Capabilities:**
* **AI4:** Expected to achieve self-driving safety "very far above human." (CyberCab will launch with AI4).
* **AI5:** Will make cars "almost perfect" and "greatly enhance Optimus." (Expected late 2027 earliest).
* **AI6:** For Optimus and data centers.
* **AI7 & Dojo 3:** Will be "space-based AI compute."
* **Hardware 3 Owners:** Tesla will not make a decision on HW3 upgrades until unsupervised driving with AI4 is widely deployed and its capabilities are fully understood.
* **AI4 Sufficiency:** Elon believes AI4 is "more than capable of solving unsupervised" driving, affirming that "better hardware is not needed for unsupervised driving" based on current performance.
* **Tesla's "Mathematical Cheat Code" Patents:**
* **RopE (Rotary Positional Encoding):** A key component in five patents published Jan 15th, 2024.
* **Functionality:** Allows 8-bit, low-power chips to run elite 32-bit AI models by enabling high-precision operations (like remembering order/position) with "rougher, simpler numbers." It involves logarithmic compression (e.g., log(1M) = 6).
* **Benefits:** Cars/robots can remember the world longer (e.g., 30 seconds) with less power, enabling AI on the edge (in-vehicle) without massive batteries or fancy chips, and Optimus to work for hours on low power.
* **Strategic Implications:** These patents secure Tesla's strategic independence from Nvidia's CUDA ecosystem, enable a dual-foundry strategy (Samsung & TSMC), and create a deliberate oversupply of compute, potentially turning its idle fleet into a distributed inference cloud rivaling AWS in efficiency.
* **Software Breakthroughs:** The speaker emphasizes that software breakthroughs are now driving efficiency more than hardware alone, solving specific bottlenecks to run transformer-class AI on existing 8-bit silicon.
* **Dojo's Evolution:** Clive Chan (former Dojo lead) noted Dojo 1's ambition but execution challenges. Elon confirmed Dojo 3, if it aligns with older ideas, would be "nuts," and is necessary for "space AI."
* **Streamlined Chip Program:** Tesla is moving towards a single chip that handles both inference and training, mass-produced, with efficiency driven by software breakthroughs. This gives them an advantage over companies buying general-purpose chips from Nvidia.
* **Tesla Fab / Chip Manufacturing:**
* **"Tesla Terra Fab":** Elon responded with a target emoji to the idea of a Tesla Terra Fab, indicating it's the "only option for super high volume."
* **Demand vs. Supply:** Current TSMC/Samsung capacity might handle conservative growth, but Elon's vision of 20 million vehicles/year and billions of robots creates demand that "dwarfs current global supply."
* **In-house Production:** In-house production is seen as inevitable by 2029-2030 for control, cost, security, and scale.
* **Groundbreaking Imminent:** If Elon's comments aren't just negotiating tactics, a groundbreaking for a Terra Fab should happen this year, as it takes 3-5 years to bring a fab online.
* **TSMC Arizona Expansion:**
* TSMC's Arizona desert expansion is "exceeding expectations."
* The US could go from 0% to 25% of global chip production by 2035.
* Construction on TSMC's second fab is complete (production expected 2028), and work on Fab 3 started last April.
* TSMC recently acquired 900 more acres and hinted at more capacity.
* TSMC will build at least five new semiconductor facilities in the US as part of a trade deal where the US cut Taiwan's baseline rate from 20% to 15%, and Taiwan pledged $250 billion in US tech investments.
* Despite this, Elon believes current expansion won't be enough.
* **Other News:**
* **Grok's Colossus Supercomputer:** Now operational, it's the "first gigawatt training cluster in the world," with upgrades to 1.5 gigawatts planned for April.
* **Model Y Canada Deliveries:** Many Model Y Performance reservation holders in Canada are seeing delivery timelines pushed back from February to as late as July, with no official reason given.
* **Tesla Diner Supercharger:** The highest usage Supercharger globally (43,000 sessions or 1.5 GWh/month), averaging 18 sessions per stall per day across 80 stalls. Learnings will apply to future Supercharger amenities.
* **Pi Safe Supercharger (Idaho):** A private company is white-labeling Tesla Superchargers in Idaho, with a design featuring native tree species, aiming to revitalize the town.
* **Elon's Stock Activity:** Elon confirmed he hasn't sold Tesla stock for about three years and bought $1 billion worth last year.
* **Market Closed:** Tesla stock had no change due to markets being closed.
摘要
My Tesla referral link: https://ts.la/dillon81150 My fitness channel: https://www.youtube.com/@BuiltxBasicz Bandz Fitness IG: ...
GPT-4正在为你翻译摘要中......
中英文字稿 
Welcome to Electrified, it's your host, Dylan Lumis, quick shout out to my newest patrons, ScooterR. Thank you for choosing to support the channel. Looking back on Friday's episode, I definitely should have been more explicit about how Tesla will benefit from Canada's new EV tariff deal with China. And as we know, at least initially, the companies that stand to benefit will be manufacturers that already have North American certification. And not only that, but North American distribution. Which means companies like Tesla, Volvo, and Polestar.
欢迎来到Electrified，我是你的主持人Dylan Lumis。首先，特别感谢我最新的赞助者ScooterR，感谢你选择支持这个频道。回顾周五的节目，我显然应该更详细地说明特斯拉将如何受益于加拿大与中国的新电动车关税协议。正如我们所知，至少在初期，能够受益的公司将是那些已经获得北美认证的制造商。不仅如此，他们还需要在北美有分销渠道。这意味着像特斯拉、沃尔沃和Polestar这样的公司将受益。
For context, Tesla imported 44,000 cars into Canada from China in 2023, the last full year before the 100% tariff rate. As part of this new deal, Transport Canada will certify new Chinese EVs within just eight weeks. And yes, this will give Tesla more supply chain flexibility, allowing it to export more cars from Shanghai to Canada that presumably have lower costs. So we'll see what Tesla does with pricing in Canada, but it's likely that Tesla makes up a large portion of this 49,000 unit quota that remember is expected to grow to 70,000 in the coming years.
为了提供背景信息，特斯拉在2023年，也就是100%关税率政策实施前的最后一个完整年度，从中国向加拿大进口了44,000辆汽车。根据这项新协议，加拿大交通部将在八周内完成对中国新能源汽车的认证程序。是的，这将为特斯拉提供更多的供应链灵活度，使其能够以更低的成本从上海向加拿大出口更多车辆。至于特斯拉将在加拿大进行何种价格策略，我们将拭目以待，但可以预见到，特斯拉很可能会占据这49,000辆配额中的很大一部分，而这个配额预计在未来几年还会增长到70,000辆。
So actually, this one's public. So one thing that we're thinking about is, okay, like we're building this human emulator with MacroHard. How do we deploy it? Because you actually need, like, if we wanted to deploy one million human emulators, we need one million computers. How do we do that? And the answer showed up two days later in the form of a Tesla computer, because those things are actually the very capital efficient as it turns out.
这其实是公开的信息。我们正在考虑的问题之一是，我们与MacroHard一起打造了一个人类模拟器，如何进行部署？因为如果我们想部署一百万个模拟器，就需要一百万台计算机。我们该怎么做呢？两天后答案出现了：特斯拉计算机。这些计算机实际上非常高效，资金投入也非常合适。
And we can run potentially like our model and the like full computer that a human would otherwise work at on the Tesla computer for much cheaper than you would on a VM on AWS or Oracle or whatever, or even just buying hardware from a video that car computer is actually much more capital efficient. And so it enables us to assume that we can deploy much, much faster at a much higher scale. And so we've adjusted our expectations for that basically.
我们可以在特斯拉计算机上运行我们的模型和完整的计算任务，而这些任务如果在人类使用的普通计算机上进行的话，成本会比在AWS或Oracle的虚拟机上低得多，甚至比购买NVIDIA硬件的成本效率更高。因此，这使我们可以假设我们能够以更高的速度和更大的规模进行部署。因此，我们基本上调整了对这一点的预期。
We're basically able to just bootstrap off of the like car network. So that's one of the one of the potential solutions basically guys like, okay, well, we want one million VMs. There's like four million Tesla cars in North America alone. And like, let's say two thirds or half of them have harbory four. And like somewhere between 70 80 percent of the time, they're sitting there, Idle probably charging. We can just potentially pay, and they have, you know, networking, they have cooling, they have power. We can probably just pay owners to lease time off their car and let us run like a human emulator digital optimist on right on it.
我们基本上能够利用现有的汽车网络。 这就是其中一个潜在的解决方案。比如说，我们想要一百万个虚拟机，但在北美就有大约四百万辆特斯拉汽车。假设其中三分之二或一半配备了硬件四，并且在70%到80%的时间里，它们都处于闲置状态，可能正在充电。我们可以尝试支付费用给车主，让他们将车的时间出租给我们，以便我们在车上运行类似人类模拟器这样的数字运算工具。
And they get, you know, they're at least paid for and we get a full human emulator we can put to work. That was Suleiman Gory, a member of the technical staff at XAI. We know the idea of Optimus is to take physical tasks that humans can do. And then the robot will do those same tasks for a fraction of the cost. So this digital optimist that Suleiman was talking about is doing the same thing. But with everything that humans do digitally.
他们至少得到了报酬，然后我们获得了一个可以投入工作的完整人类模拟器。这是XAI技术团队成员苏莱曼·戈里说的。我们知道Optimus的想法是完成人类可以做的体力工作，然后机器人以低得多的成本完成这些工作。而苏莱曼提到的这个数字化的Optimus也是在做同样的事情，但对象是人类在数字世界中所做的一切。
So think of a human sitting at a computer and all of the different mouse and keyboard inputs. XAI is currently working to emulate all of that so they can deploy in basically any situation that a human is in working at a computer. So this distributed compute idea that Elon's been talking about now for a few years may be a lot closer than some of us were expecting. And yes, I think this is potentially big news and we have some indication that maybe that was not supposed to be public as Suleiman just posted this morning he's left XAI.
想象一个人坐在电脑前，使用各种鼠标和键盘的输入。XAI目前正在努力模拟这些操作，以便在任何人使用电脑的情境下进行部署。埃隆讨论了好几年的这个分布式计算的概念，可能比我们一些人预期的要更接近实现。而且，是的，我认为这可能是个重大新闻，我们有一些迹象表明这本不该公开，因为Suleiman今天早上刚刚发帖说他已经离开XAI了。
Nothing but loved my former team and coworkers. But I'll just say after that interview there's no way he would leave XAI. He very clearly loved working there and loved working for Elon and knew that the XAI team was on to greatness. So it's just speculation on my part, but I'd be pretty confident saying that he was likely fired by XAI for some things he said in that interview. One of those being this new human emulator path to deployment on millions of Tesla vehicles.
我非常欣赏我之前的团队和同事。不过，我只能说经过那次采访后，他绝对不可能离开XAI。他显然很喜欢在那里工作，也很喜欢为Elon工作，并且知道XAI团队将会取得伟大成就。所以，这只是我的猜测，但我很有信心地说，他很可能是因为在那次采访中说的一些话被XAI解雇了。其中之一就是这个关于在人类模拟器上部署到数百万辆特斯拉汽车的新路径。
And he said that XAI has already been testing these emulators internally as if they were human employees. So there's just one more synergy between XAI and Tesla that we may see in the coming years. I'd guess Tesla owners would have the option to opt in or out to allow this use for your car's computer, but it would come with some form of payment from XAI to use the computer to run these human emulators since all of the infrastructure is already there.
他说，XAI 已经在内部测试这些模拟器，就像它们是人类员工一样。因此，这是我们在未来几年可能会看到的 XAI 和特斯拉之间的又一种合作。我猜测特斯拉车主会有选择权，可以决定是否允许 XAI 使用他们车辆的电脑来运行这些人类模拟器，不过 XAI 将会支付一定的费用来使用这些电脑，因为基础设施已经到位。
And the way Suleiman talked about it wasn't some far off idea, it sounded like a legit road map that XAI was planning on. So no guarantees here as XAI is still testing these emulators, but if they figure it out, Tesla vehicles may be the path to deployment at scale. Meaning Tesla owners could have a revenue stream for allowing XAI to use their cars when they're not in use. And the truth is I could do an entire video just on these emulators alone. So this is a part of what macro heart is working on which could revolutionize things like data entry, customer support, research, coding, software as a service, browsing the web and so on. So Optimus to replace physical tasks and digital Optimus to replace digital tasks. So yes, they're both still in very early stages, but I would argue that the path forward is actually pretty clear at this point.
苏莱曼谈论这件事情时，并不像是在说一个遥远的想法，而是听起来像XAI正在筹划的一个实际路线图。所以这里没有任何保证，因为XAI仍然在测试这些模拟器，但如果他们成功了，特斯拉车辆可能会成为大规模部署的途径。这意味着特斯拉车主可能会通过允许XAI在车辆闲置时使用它们来获得一条收入来源。其实，仅仅关于这些模拟器，我就可以做一个完整的视频。这是宏观心脏正在致力于的一部分项目，有可能革新数据输入、客户支持、研究、编码、软件即服务、网络浏览等领域。Optimus可以取代物理任务，而数字Optimus可以取代数字任务。虽然目前这两个项目都处于非常早期的阶段，但我认为现在前进的道路实际上已经相当清晰。
And of course I have no idea what type of payments Tesla car owners would have, but Suleiman did say that least payments could be paid for. So I don't know about you, but I think that's a pretty attractive proposition to become a Tesla owner in the future. This CyberCab was spotted in Chicago out doing some testing, but you'll see that the rear camera does indeed have a self cleaner. You can see Liquid, presumably from a camera sprayer, dripping down the back of the trunk. Which yes, I think would be very nice to have as standard across all Tesla vehicles. We also had Art Gujardo sharing this video, which was one of the first times we've seen a CyberCab out in the wild without SideView mirrors.
当然，我不知道特斯拉车主需要支付什么类型的费用，不过Suleiman提到至少某些费用是可以支付的。所以我不知道你怎么看，但我认为成为特斯拉车主是个相当有吸引力的提议。这个CyberCab在芝加哥被发现正在进行一些测试，你会看到它后置摄像头确实有一个自清洁装置。你可以看到液体（很可能是摄像头喷雾器喷出的）从后备箱滴下来。我认为，如果所有特斯拉车型都标配这个功能，那将会非常好。我们还看到Art Gujardo分享了这段视频，这是我们第一次在野外看到CyberCab没有侧视镜的样子。
And this is one category where hopefully we'll get some new federal standards so each individual state can't have their own rules for registration, operation and inspections of these SideView mirrors. Josh West was able to get some great footage of a CyberCab driving around in Austin also without SideView mirrors. Check out those new tail lights. So the brake lights are two individual lights in the spoiler. There we go. Nice and close. Beautiful. Looks good. Looks good. We got some kind of lights down in the bottom here. Brake lights below. He's probably checking me out in the camera all in the lot where Robotaxi operates out of I came through here like several weeks ago.
这是一个类别，我们希望能够出台新的联邦标准，以避免各州制定自己关于侧视镜的注册、运行和检查的规则。Josh West 在奥斯汀拍到了一个没有侧视镜的 CyberCab 行驶视频。来看一下这些新的尾灯。刹车灯被设计成在扰流板中的两个独立灯。看，我们靠近一些来看。漂亮，看起来不错。我们在底部还看到了一些灯，刹车灯在下方。他可能在用摄像头查看我，我再次来到这个 Robotaxi 的运营场所，几周前我也来过这里。
This lot was empty. And now it is filled filled with an absolute ton of Robotaxi ready vehicles. So check this out guys. Look at all of these Robotaxis. Like as far as you can see, these guys are getting ready to drop the hammer on Austin. I can tell you lately when I've been calling Robotaxi, it's all been, you know, you got to wait, you got to wait, hang tight. But not for long by the looks of it. And hopefully we do get that influx of new vehicle soon because Joe Tett Meyer on Sunday said he was in the heart of Austin, but could not get a Robotaxi due to high demand.
这个场地以前是空的，现在已经被大量准备好的自动驾驶出租车填满了。大家快来看看吧，这些自动驾驶出租车一眼望不到边。他们准备在奥斯汀这边大展拳脚。我最近打电话叫自动驾驶出租车的时候，总是被告知需要等待、再等等。不过看起来这种情况不会持续太久。而且希望我们能很快迎来新车的涌入，因为乔·特特迈尔在周日说，他虽然身在奥斯汀的市中心，但由于需求量大，还是没能叫到一辆自动驾驶出租车。
So we don't know for sure if those vehicles will all be deployed as Robotaxis, but I think there's at least a chance. According to the Robotaxi tracker, which is not official, Austin is up to 46 vehicles in the Robotaxi fleet. We also have the official NITSA incident reports now added to the site. These figures are for both San Francisco and Austin. So for Q4, October, there was two incidents November had four and December had three. And you'll see most of those were actually involving cyclists, but the majority of them resulted in no injuries, only one resulted in a minor injury.
所以我们无法确定这些车辆是否都会被部署为自动驾驶出租车，但我认为至少有这种可能性。根据非官方的自动驾驶出租车追踪器，奥斯汀的自动驾驶出租车队伍已经增加到46辆。我们的网站现在也包含了官方NITSA事故报告，这些数据涵盖旧金山和奥斯汀。在第四季度，10月有两起事故，11月有四起，12月有三起。你会发现大多数事故实际上涉及骑自行车的人，不过大多数都没有造成人员伤害，只有一起导致了轻伤。
But unfortunately, with most of these minor incidents, a lot of the information is redacted. So a lot of times all we get is that a cyclist was involved, no one was injured, the car was going a zero miles an hour, the airbags were not deployed and the car was not towed. And further complicating these reports is that it does not tell us who was actually at fault. But I think it's fair to say if most of the time the cars are going zero miles an hour, then it might actually be the cyclist who's at fault. And Ethan did say he includes all ADS and ADAS filtered by the ones indicated to be commercial.
不幸的是，对于大多数这样的轻微事故，很多信息都被隐藏了。因此，我们通常只知道一个骑自行车的人涉事，没有人受伤，汽车速度为零，安全气囊没有弹出，汽车也没有被拖走。这些报告的问题更在于，它们没有说明到底是谁的责任。不过，我认为，如果大多数时间汽车的速度都是零，那么可能责任在骑自行车的人。而 Ethan 确实说过，他包含所有自动驾驶系统（ADS）和高级驾驶辅助系统（ADAS）的数据，这些数据经过过滤，只显示为商业用途的部分。
But I just wanted to make it clear, these incidents do not necessarily mean that Tesla was at fault. And further, based on the information we get, we have good reason to believe that Tesla may not be at fault in a majority of these instances. And the cyber cab was just spotted in Massachusetts, which now makes four five states, in addition to New York, Texas, California, and Illinois. Devon shared two very impressive clips of FSD so watch and enjoy. I got delayed all that Seattle stuff and then they just go Seattle to Victoria. Oh, that's going to get a thing. That was impressive.
但我只是想说明一点，这些事件并不一定意味着特斯拉有过错。此外，根据我们获得的信息，我们有充分理由相信在大多数情况下特斯拉可能并没有过错。而且，网络出租车刚刚在马萨诸塞被发现，这样一来，加上纽约、德克萨斯、加利福尼亚和伊利诺伊，现在总共有五个州。德文分享了两个非常令人印象深刻的全自动驾驶视频，大家可以观看和欣赏。我因西雅图的事情耽误了，他们就从西雅图去维多利亚。哦，那会引起关注的。真是令人印象深刻。
I was. How did it know the perfect amount of swing? I have no idea. Elon had much to say about silicon over the weekend. He said now that the AI5 chip design is in good shape. Tesla will restart work on Dojo 3. Then he put out a line for recruiting, saying if you're interested in working on what will be the highest volume chips in the world, get in touch. Elon said solving AI5 was existential to Tesla, which is why I had to focus both teams on that chip, and I've personally spent every Saturday for several months working on it. This will be a very capable chip, roughly hopper class as single system on chip and blackwell as dual, but it costs peanuts and uses much less power. Now that AI5 is looking good, we have some headroom to work on Dojo 3.
我在想，它是怎么知道正好合适的摆动幅度的？我完全不知道。埃隆·马斯克在周末对硅技术发表了许多看法。他说，现在AI5芯片设计已经很好了，特斯拉将重启Dojo 3的工作。然后他发布了招聘信息，说如果你有兴趣参与世界上最多产的芯片的研发，请联系我们。埃隆表示，解决AI5对特斯拉至关重要，所以他必须将两个团队的精力都集中在这个芯片上，并且他本人已经连续几个月每个星期六都在研究这个问题。这将是一个非常强大的芯片，单个系统芯片大约相当于Hopper级别，双系统芯片则相当于Blackwell，但成本很低而且耗电量更少。现在AI5进展良好，我们有了一些余地来开展Dojo 3的工作。
At this point, I'll just say this is all changing very quickly for Tesla. It was only five months ago when Elon said this. Once it became clear, all paths converged to AI6. I had to shut down Dojo and make some tough personal choices, as Dojo 2 was an evolutionary dead end. Dojo 3 arguably lives on in the form of a large number of AI6 system on chips on a single board. So to me, it seems like this resurrection of Dojo 3 really is just Tesla having time to focus on AI6 since AI5 is in such a good place, and working on integrating a large number of these AI6 inference chips that will be able to be used for training as well.
目前，我只能说特斯拉的局势变化得非常迅速。就在五个月前，埃隆还说过这话。一旦事情明朗，所有的道路都汇聚到AI6。我不得不关闭Dojo，并做出一些艰难的个人决策，因为Dojo 2是一个进化的死胡同。可以说，Dojo 3以单块板上大量的AI6系统芯片的形式得以延续。在我看来，Dojo 3的“复活”实际上是因为AI5的发展已经很稳固，使得特斯拉有时间专注于AI6，以及致力于整合大量AI6推理芯片，这些芯片也能用于训练。
Meaning Dojo 3 is not some new chip that Tesla is designing on its own from scratch or even reviving from the prior Dojo program. Rather, Dojo 3 is simply an architecture or a philosophy using AI6. And it's likely AI6 where these lines between inference compute and training compute start to be blurred for Tesla. Back to this weekend, Elon said our AI5 chip design is almost done and AI6 is in early stages, but there will be AI7, AI9 aiming for a 9-month design cycle. Now, a lot of people are taking that to mean every 9 months there's going to be a new chip in the vehicles, but personally I do not think that's the case.
"Meaning Dojo 3" 并不是特斯拉自己从头设计的新芯片，也不是从之前的 Dojo 项目中复活的东西。相反，Dojo 3 更像是一种使用AI6的架构或理念。在AI6中，推理计算和训练计算之间的界限可能会开始模糊化。回到这个周末，马斯克说我们的AI5芯片设计几乎完成，AI6还处于早期阶段，但将会有AI7、AI9，目标是一个9个月的设计周期。现在，很多人都认为，这意味着每9个月就会有一款新芯片用于车辆，但我个人认为情况并非如此。
One because Elon said design cycle, meaning every 9 months a new chip is designed. So first of all, that's the goal that doesn't mean it's exactly how it's going to play out, and once you have the chip designed, there's a lot of other steps that need to happen to actually get that into production. And depending on that chip design, the time it takes to get into production will vary as well. And I think further confirming my thesis would be this, Elon just said AI4 by itself will achieve self-driving safety levels very far above human. AI5 will make the cars almost perfect and greatly enhance optimists. AI6 will be for optimists and data centers.
翻译成中文并简化：
埃隆提到设计周期问题，也就是说，每9个月设计一个新芯片。首先，这是目标，但不表示一定会如此。芯片设计完成后，还有许多步骤需要完成才能投入生产。根据芯片设计的不同，进入生产的时间也会有所变化。我认为进一步证明我的观点的是，埃隆刚刚说AI4将单独实现比人类更高的自动驾驶安全水平。AI5将使汽车接近完美，并极大地提升Optimus机器人的能力。AI6则将用于Optimus机器人和数据中心。
AI7 and Dojo3 will be space-based AI compute. A few things, I would say AI4 achieving unsupervised is not a guarantee this is going to boil down to the regulatory environment, but to me it would make sense that if the data shows AI4 is capable of being multiple times safer than a human, then the regulator should be on board allowing that to carry out unsupervised. And let's not forget, the cyber cab is going to launch this year with AI4. We're not going to see AI5 until late next year, late 2027, likely at the earliest.
AI7和Dojo3将是基于太空的AI计算。我想说几件事情。首先，AI4实现无人监督并不是板上钉钉的事，这将取决于监管环境。但在我看来，如果数据显示AI4在安全性方面远超人类，那么监管机构应该支持其在无人监督的情况下运行。此外，不要忘记，搭载AI4的网络出租车将于今年推出。至于AI5，我们要等到明年年底或者最早要到2027年年底才能见到。
So yes, I'm confident saying AI4 will eventually unlock unsupervised. But if AI5 is what's going to make the cars almost perfect, then why would Tesla replace that nine months later with AI6? It wouldn't be required. So I think what's likely to happen is when AI5 starts shipping late 2027, that will be the chip in the fleet for years to come. And this would seem to confirm that AI6 is going to be focused on optimists and data centers, which again is in line with those lines being blurred between inference and compute.
所以，是的，我有信心地说，AI4最终会实现无监督学习。但如果AI5是使汽车几乎完美的关键，那么特斯拉为什么要在九个月后用AI6取而代之呢？这似乎没有必要。因此，我认为更有可能的情况是，当AI5在2027年底开始出货时，这款芯片将在车队中使用很多年。而这似乎也证实了AI6将主要面向优化和数据中心，它将进一步模糊推理和计算之间的界限。
And AI7 and Dojo3 will be space-based AI compute. And I still think Dojo3 from a very oversimplified standpoint will be an architecture or an ethos using AI6 and AI7 in tandem for training compute. And for hardware three owners, I would say what I said a few months ago. It's still too early for Tesla to make a decision about this. Until Tesla solves unsupervised with AI4, it's not going to know what it needs to do for hardware three cars.
AI7和Dojo3将是基于太空的AI计算。我仍然认为，从非常简化的角度来看，Dojo3将是一种架构或理念，利用AI6和AI7协同进行训练计算。对于硬件三代的车主来说，我想重申几个月前说过的话。特斯拉目前仍然无法对此做出决定。在特斯拉用AI4解决无监督学习之前，它不会知道如何处理硬件三代的汽车。
Meaning, until unsupervised actually ships in the wild, Tesla is not going to be able to answer your questions about what hardware three will be replaced with. On that point, Elon said necessity is the mother of invention. The Tesla AI team is epically hardcore. No one can match Tesla's real world AI. That was sharing this post from Ming saying Tesla has patented a mathematical cheat code that forces cheap 8-bit chips to run elite 32-bit AI models and rewrites the rules of silicon.
这段话在表达，直到真正无监督自主驾驶技术在现实生活中应用之前，特斯拉还无法回答关于第三代硬件将被替换成什么的问题。关于这一点，埃隆·马斯克说过“需要是发明之母”。特斯拉的人工智能团队极其拼搏，没人能匹敌特斯拉在实际应用中的人工智能。文中提到Ming分享了一则消息，说特斯拉已经申请了一项专利，通过数学上的巧妙算法，使得便宜的8位芯片能够运行高级的32位AI模型，并重新定义了硅芯片的规则。
And rather than me trying to explain how this works, Cernbacher beat me to it. The fancy math part of this patent, rope, rotary, positional, and coding is AI remembering the order and position of things. It's like giving each item a special tag based on angles and math. But doing this very precisely usually needs a lot of compute power, 32-bit precision, which is like super detailed numbers. The hack? Tesla figured out how to do this high precision stuff on cheaper low power hardware, which is 8-bit, which is like rougher, simpler numbers. Another way to think of it is rounding these numbers to a shorter decimal point. Part of that is done by squeezing big numbers into smaller ones logarithmically, like turning a million into six because the log of a million is six, it saves space. So in short, this means the car or the robot can see and remember the world for longer, like 30 seconds without losing track, without needing a massive battery or fancy chips.
与其让我来解释这是如何运作的，不如说Cernbacher抢先一步为我们讲解了。这项专利中的复杂数学部分——绳索、旋转、定位和编码，是关于AI记住事物的顺序和位置的。就好像给每个物品一个特别的标签，基于角度和数学进行标记。但通常要做到这一点非常精确，需要大量的计算能力，32位精度，就像超级详细的数字。而Tesla的妙招在于，他们找到了一种方法，在便宜的低功耗硬件上实现这种高精度计算，这种硬件是8位的，就像较为粗糙、简单的数字。另一种理解方法是，把这些数字四舍五入到一个更短的小数位。部分处理方法是通过对大的数字进行对数压缩，比如把一百万变成六，因为一百万的对数是六，这样可以节省空间。简而言之，这意味着汽车或机器人可以在不需要一个庞大的电池或昂贵芯片的情况下，更长时间地“看见”和记住周围的世界，比如30秒，而不丢失轨迹。
It's more efficient so optimists can work for hours on low power and cars can run AI on the edge right in the vehicle without sending data to a big server. And guess what? It wasn't just one patent, but rather Tesla had five patents published on the same day, January 15th this year. They all share overlapping inventors and basically identical architectural diagrams, showing things like cars and trucks connected via networks to cloud servers and databases for AI processing and training. And this is a perfect data point to back up what I just said on X. I said over the next few years, we'll likely see more breakthroughs on the software side than we will on the hardware side, specifically when it comes to neural net training. So yes, AI five, six and seven, that will all push things forward significantly, but we cannot overlook the fact that Tesla owns the entire software stack and is pushing forward all of these patents on the software side that are custom built for Tesla's specific applications.
这项技术更高效，因此乐观者可以在低电量模式下工作数小时，车辆可以在边缘计算中运行AI，而无需将数据发送到大型服务器。而且你猜怎么着？这不是一项专利，而是特斯拉在今年1月15日当天发布了五项专利。这些专利有共同的发明者，几乎相同的架构图，展示了通过网络连接到云服务器和数据库进行AI处理和训练的汽车和卡车。这正好支持了我在X上刚刚说过的观点：未来几年中，我们很可能在软件方面看到比硬件方面更多的突破，特别是在神经网络训练方面。所以，AI五、六、七都将显著推动进步，但我们不能忽视特斯拉掌握了整个软件系统，并在软件方面推动了这些专利的发展，这些都是为特斯拉特定应用量身定制的。
And as SETI Park said right here, each patent solves a specific bottleneck. Together, they enable something that shouldn't be possible, running a transformer class AI on 8-bit production silicon. These patents ensure that evolution is possible. They describe techniques for running algorithms that don't exist yet on hardware that already exists. Meaning the silicon can stay the same, but the brain keeps improving. And there was one phrase in all four patents, product life cycles for automobiles and robotics can span decades across several generations of algorithmic development. And this was from Ming in the post that Elon shared. Finally, baking this math into the silicon secures Tesla's strategic independence. It decouples the company from Nvidia's CUDA ecosystem and enables a dual-foundry strategy with both Samsung and TSMC to mitigate supply chain risks.
正如SETI Park在这里所说，每项专利都解决了一个特定的瓶颈。它们共同实现了一项看似不可能的任务，即在8位生产硅片上运行变压器类人工智能。这些专利确保了演进的可能性。它们描述了如何在现有硬件上运行尚不存在的算法的技术。也就是说，硅片可以保持不变，但“头脑”不断提升。四项专利中都有一句话：汽车和机器人产品生命周期可以跨越几十年，涵盖数代算法开发。而这句话来自Elon分享的Ming的帖子。最终，将这些数学原理嵌入到硅片中，确保了特斯拉的战略独立性。这使公司不再依赖Nvidia的CUDA生态系统，并通过与三星和台积电的双代工策略来降低供应链风险。
This creates a deliberate oversupply of compute potentially turning its idle fleet and unsold chips into a distributed inference cloud that rivals AWS in efficiency, which ties right back to the Suleiman XAI interview. And no, that was not just some AI slop. Elon replied to that post saying pretty much. But again, I want to pause real quick and just remind everybody what Elon said six months ago. Tesla Dojo AI training computer is making progress. We start bringing Dojo 2 online later this year. It takes three major iterations for new tech to be great. Dojo 2 is good, but Dojo 3 will be great. That was June of last year. Then about a month after that, Elon said that Dojo 2 was an evolutionary dead end and fast forward another few months.
这导致了计算资源的过剩，可能会将闲置的计算设施和未售出的芯片转变成一个分布式推理云，其效率可以与AWS媲美，这也和Suleiman关于可解释人工智能（XAI）的采访紧密相关。而且，这并不仅仅是某种AI的随性之举。埃隆对此回复说大致如此。但我想稍微停一下，提醒大家埃隆在六个月前说过的话。特斯拉的Dojo AI训练计算机正在取得进展。我们计划在今年晚些时候上线Dojo 2。新技术通常需要三次重大迭代才能变得优秀。Dojo 2还不错，但Dojo 3将会非常出色。这是去年六月份的事情。然后大约一个月后，埃隆又说Dojo 2是一个演化的死角，再过几个月……
And now Dojo 3 is coming back in a different form. So my point is right now everything Elon is saying about all of this are just the plans right now. But clearly this space is changing and evolving incredibly rapidly. So just know a lot of this is subject to change and anytime Tesla has breakthroughs on the technical side on the software side, that of course might change what their plans are on the hardware side as well. Back to AI4, somebody said I'm waiting for the AI5 chip to buy a Tesla. That's the only version that'll be legally unsupervised FSD. Elon replied to that saying AI4 will far exceed human safety.
现在，Dojo 3 将以不同的形式回归。我的观点是，目前埃隆谈论的所有这些内容都只是计划。但显然，这个领域正在迅速变化和发展。所以请注意，很多东西都是可能会改变的，每当特斯拉在技术或软件方面有突破时，这当然也可能会改变他们在硬件方面的计划。 回到 AI4，有人说，他们在等 AI5 芯片才买特斯拉，因为那是唯一能合法实现无人监督全自动驾驶的版本。对此，埃隆回复说，AI4 的安全性将远超人类。
Which again, I think it would be fair to assume that Elon believes AI4 will unlock unsupervised when it gets there and when regulations allow it across the country. And one more time, once all of this happens, then we will get answers on what's going to happen with hardware 3. I like Tesla said anyone using AI4 right now knows that better hardware is not needed for unsupervised driving. It's incredibly good right now. To which Elon said yep. So it could not be any more clear in Elon's mind right now, AI4 will be more than capable of solving unsupervised. Nick Patain shared what Elon said saying a Tesla Terra Fab sound like a good idea. To which Elon said seems to be the only option for super high volume.
我认为，可以合理地假设，Elon相信AI4在技术成熟和法规允许下将实现无人监督的能力。而且，一旦这些条件都满足，我们就能知道硬件3的发展情况。我很欣赏特斯拉所说的，目前正在使用AI4的人都知道，无人驾驶并不需要更好的硬件。当前的性能已经相当出色。对此，Elon表示赞同。所以，在Elon看来，AI4完全有能力解决无人监督的驾驶问题。Nick Patain分享了Elon所说的话，提到建立一个特斯拉超级工厂听起来是个好主意。Elon对此表示，这似乎是实现超大规模生产的唯一选择。
And Phil Beisel said what we've been arguing the past few months. TSMC and Samsung can handle conservative growth through 2028 to 2029. But Elon's vision explodes demand to 20 million vehicles a year, billions of robots long term, and compute needs dwarfing current global supply. So the expectation AI5 Ramps in 2027 covered by TSMC and Samsung, AI6 in 2028, at Samsung's dedicated factory in Texas. But then by 2029-2030 the shortfall hits hard, and that's when in-house production becomes inevitable for control, cost, security, and insane scale. Probably no way around it. To which Elon replied with the target emoji.
菲尔·贝塞尔提到了我们过去几个月一直在讨论的问题。台积电和三星可以应对直到2028至2029年的保守增长。但埃隆的愿景会导致需求爆发，每年达到2,000万辆汽车，长期来看需要数十亿个机器人，对计算的需求将远远超过目前的全球供应。所以预计AI5将在2027年由台积电和三星加速推广，AI6将在2028年由三星位于德克萨斯的专用工厂负责。但到2029至2030年，供不应求的情况会非常严重，这时候，为了控制成本、安全性以及庞大的规模，内部生产将成为必然选择，几乎无法避免。对此，埃隆用目标表情符号回应。
Which is exactly why I'm saying, if Elon's not just making these comments to pressure TSMC and Samsung, we should hear about a groundbreaking for this Terra Fab this year. That's because it's going to take Tesla anywhere from three to five years to actually get that fab into production. Meaning if Elon's not saying all of this as negotiating tactics, we should get a groundbreaking this year. Clive Chan who actually used to be the Dojo lead at Tesla said Dojo 1 was ultra ambitious and really pushed a lot of technologies, packaging, power delivery, system design, clock distribution, even number formats. But execution and focus could have been better.
这正是我所说的原因，如果埃隆不仅仅是在对台积电和三星施加压力，我们今年应该会听到有关这座"Terra Fab"晶圆厂的奠基消息。这是因为特斯拉需要三到五年的时间才能让这个工厂投产。也就是说，如果埃隆所说的这些并不是作为谈判策略，我们今年就应该看到奠基仪式。实际上，曾是特斯拉Dojo项目负责人的Clive Chan表示，Dojo 1非常有野心，推进了许多技术，如封装、供电、系统设计、时钟分配，甚至是数字格式。但执行和专注度方面还可以更好。
If Dojo 3 is anything, like the ideas floated around the office years ago, it'll be nuts. It'll be interesting to see if things move faster when Elon's focus is on the chips. Go go go and good luck. And Elon replied to that saying yes on all counts has to be done to scale space AI. So if for whatever reason, you still needed more confirmation that Tesla would be heavily involved in the data centers in space, there you go. It won't just be Tesla providing solar panels but providing the compute as well.
如果道场3像几年前在办公室里讨论的那些想法一样，那就会非常疯狂。看看当埃隆把注意力集中在芯片上时事情会不会有更快的进展，这将会很有趣。加油加油，祝好运。而埃隆对此回应说，是的，现在所有事情都必须完成，以扩展太空人工智能。所以，如果你出于某种原因，还需要更多证据来确认特斯拉将大力参与太空数据中心的建设，现在已经很明确了。特斯拉不仅会提供太阳能电池板，还会提供计算能力。
So in short, Tesla is continuing to streamline its entire chip program. Both on the hardware and the software side. So rather than Tesla having one chip for inference and one chip for training, Tesla is just going to have one chip that does both and then they're going to make a bunch of those. And a lot of the breakthroughs and pushing forward more efficiency will happen on the software side. Companies like Ford GM and Toyota aren't designing their own chips, they're buying from Nvidia, paying all of those margins, buying these general purpose chips.
简单来说，特斯拉正在精简其整个芯片项目，涵盖硬件和软件两方面。不再为推理（inference）和训练分别使用不同的芯片，而是将二者合二为一，使用一款兼具训练和推理功能的芯片，并大量生产。同时，很多在推动更高效率方面的突破将主要体现在软件上。像福特、通用和丰田这样的公司没有自行设计芯片，而是从Nvidia购买通用芯片，并支付相关的利润。
So from a first principle standpoint, that's always going to be an inferior approach. But Tesla has been working on things like chip design now for nearly a decade. And further as we always say, they actually have the AI talent to do it. And today we just got news, TSMC's bid to turn the Arizona desert into its number two production nexus is exceeding expectations. Chris Miller, the author of the book, The Chip War said the US could go from 0 to 25% of global chip production by 2035.
从基本原理的角度来看，这始终是一种较差的方法。但是，特斯拉在芯片设计方面已经持续研究了近十年。另外，正如我们常说，他们确实拥有实现这一目标的AI人才。今天，我们刚刚得到消息称，台积电在亚利桑那沙漠建设其第二大生产中心的计划超出预期。《芯片战争》一书的作者克里斯·米勒表示，到2035年，美国的全球芯片产量可能从0增长到25%。
All of this has profound implications for energy production, grid reliability, national defense, employment, GDP, and economic output supply chain resiliency, trade wars and the list goes on. Construction on TSMC's second fab is also complete with production expected in 2028 while work on fab 3 started April last year. And just last week it acquired another 900 acres of land and hinted at plans to build even more capacity. This to further expand its Arizona supercluster.
所有这些对能源生产、电网可靠性、国家防御、就业、GDP、经济产出、供应链韧性、贸易战等方面都有深远的影响。台积电的第二座工厂建设已经完成，预计在2028年投产，而第三座工厂的建设已于去年四月开始。就在上周，台积电又购入了900英亩的土地，并暗示有计划进一步扩大产能，这是为了扩建其在亚利桑那州的超级集群。
And TSMC just said it'll now build at least five new semiconductor facilities in America as part of this newly announced trade deal. The deal announced by the US Department of Commerce, we'll see America cut Taiwan's baseline rate from 20% to 15. In return, Taiwan has pledged that its tech companies will invest $250 billion in the US. So it's not like the company's Tesla is relying on for chips, Samsung and TSMC are just standing still and not increasing their output.
台积电刚刚宣布，将根据新公布的贸易协议，在美国新建至少五座半导体工厂。根据美国商务部公布的该协议，美国将把对台湾的基本税率从20%降至15%。作为回应，台湾承诺其科技公司将在美国投资2500亿美元。因此，特斯拉依赖的芯片供应商三星和台积电并不是在原地踏步，而是在积极扩大生产。
They are very clearly both still expanding, but Elon still thinks it will not be enough. And to top it all off, Elon said the colossus super computer for Grok is now operational. The first gigawatt training cluster in the world upgrades to 1.5 gigawatt in April. Many Model Y performance reservation holders in Canada are having their delivery timelines pushed back by a few months. Some were expecting delivery as soon as February, but they're having their timelines pushed back as late as July. As of now, Tesla has not shared any official reasoning for why.
他们两人显然还在不断扩展，但埃隆仍然认为这还不够。更令人惊讶的是，埃隆说Grok的巨型超级计算机现在已经投入使用。世界上第一个千兆瓦级的训练集群将在四月份升级到1.5千兆瓦。在加拿大，许多预订Model Y高性能版的客户发现他们的交付时间线被推迟了几个月。有些人原本期望在二月份就能收到车，但现在被推迟到了七月。到目前为止，特斯拉还没有公布任何官方原因。
Tesla charging said the Tesla diner is our highest usage supercharger in the world already delivering 43,000 sessions or 1.5 gigawatt hours each month. Max from Tesla said that's already 18 sessions per stall per day across 80 stalls. The food and operations are only getting better, we're always learning and we'll apply these learnings to future supercharger amenities. We have yet another company the Pi safe supercharger white labeling Tesla superchargers. This company's in Idaho.
特斯拉充电站表示，特斯拉餐厅已经是我们全球使用频率最高的超级充电站，每月提供43,000次充电服务，或总计1.5千兆瓦时的电力。特斯拉的Max表示，这意味着在80个充电桩中，每个充电桩每天平均提供18次充电服务。我们的餐饮和运营只会越来越好，我们不断学习，并将这些经验应用于未来的超级充电站设施。我们还有另一家公司Pi safe supercharger在特斯拉充电站上贴牌运营，这家公司位于爱达荷州。
And so far, I'd have to say they're the ones with my favorite design on the superchargers. Isaac French is the man behind this privately owned supercharger, saying it opens up a whole new lane of traffic in our effort to revitalize the town, which is two hours from an interstate or major city or airport.
到目前为止，我不得不说，他们的超级充电站设计是我最喜欢的。艾萨克·弗伦奇是这个私人超级充电站的设计者，他表示，这为我们振兴小镇的努力开辟了一条全新的交通渠道。这个小镇距离高速公路、大城市或机场有两个小时的路程。
And he said the design is the native tree species found in our little town. Elon said I have not sold stock for about three years and bought a billion dollars of Tesla stock last year.
他说，设计是我们小镇本地树种的图案。埃隆说，我大约三年没有卖过股票，去年还购买了价值十亿美元的特斯拉股票。
And no change in Tesla stock today as the markets were closed. Hope you all have a wonderful day and a huge thank you to all of my Patreon supporters.
由于市场休市，特斯拉股票今天没有变化。希望你们都度过一个愉快的日子，并非常感谢我所有的Patreon支持者。