首页  >>  来自播客: Electrified 更新   反馈  

Electrified - Tesla's "Final Boss" You've Never Heard Of / Starbase, LA by the Numbers ⚡️

发布时间:   原节目
以下是《Electrified》节目的摘要,涵盖了节目中提到的所有新闻内容: **网站更新与追踪器:** * dloomis.com 上的所有仪表盘均已更新并移植到 Cursor。 * **RoboTaxi 追踪器 (德克萨斯州):** 特斯拉目前已注册270辆车。8月份车队新增了89辆车。目前均为Model Y车型,预计未来几周将迎来Cybercab。 * **特斯拉Semi 卡车追踪器:** 已确认的特斯拉Semi卡车订单为1,267辆。 * **交付势头:** Einride 首席执行官表示,其500份订单中约有75份预计在今年交付,其余部分的目标是在明年年底前完成(原先预计是2028年)。 * **埃隆预测追踪器:** 已移植,将添加新的预测。 **赞助商:** * **Bolt.new:** 一个专为非技术用户设计的平台,可以从一个想法以最小的摩擦力构建并发布应用程序到可用的产品。功能包括自定义域名、GitHub推送和语音生成应用。使用代码“electrified”可获得30天免费专业账户。 **路易斯安那州星际基地与星链分析:** * **路易斯安那州星际基地:** 最终将拥有十几座发射塔,每天可支持超过30次星舰发射,使其成为地球上最大的发射场。 * **地理位置优势:** 胡桃岛 (Pecan Island) 距离最近的城镇约30英里,距离主要城市约50英里,因此比博卡奇卡 (Boca Chica) 的干扰小(博卡奇卡距离南帕德雷岛5英里,距离布朗斯维尔15英里)。它提供了广泛的最佳方位角(发射方向),特别是约190度(正南)用于极地/太阳同步轨道,得益于墨西哥湾的水域。 * **星链发射转型:** 首次从佛罗里达州进行的星舰飞行对星链来说是一个重要的信号。SpaceX 从40号发射台进行的任务是猎鹰9号在佛罗里达州的最后一次星链发射;未来从佛罗里达州进行的星链任务将由星舰执行。佛罗里达州的Gigabay和星舰发射台即将完工。 * **CyberCab集成:** 每辆CyberCab都将内置星链单元。 * **当前星链指标:** * 运行卫星: 约11,000颗。 * 下行容量: 约850太比特/秒。 * 订阅用户: 约1200万。 * 假设均匀分布的容量: 每个用户约71兆比特/秒。 * **星链V3与星舰容量:** * 每次搭载V3卫星的星舰发射将携带60颗V3卫星。 * 每颗V3卫星具有1太比特/秒的下行容量。 * 每次搭载V3卫星的星舰满载发射将部署60太比特/秒(相当于目前整个星链星座容量的7%)。 * 猎鹰9号搭载V2 mini卫星的发射可提供约3太比特/秒的容量。星舰搭载V3卫星的容量是其约20倍。 * SpaceX 需要约1,000颗V3卫星入轨才能显著改变性能,预计在明年第二季度实现。 * **星舰发射组合与未来预测:** * 主持人假设未来1-3年内50%的星舰发射将用于星链(其余用于NASA/HLS、油船发射、星链移动版和Starmind)。 * SpaceX 计划通过星舰将其大部分发射能力分配给其AI部门(Starmind/轨道数据中心)。 * 预计到本十年末,大部分星舰发射将用于Starmind。 * **星舰带来的星链容量假设增长:** * **每年100次星舰飞行(50次星链载荷):** 每年增加3,000颗V3卫星 / 3,000太比特/秒(是当前星座容量的3.5倍)。 * **每年1,000次星舰飞行(500次星链载荷):** 每年增加30,000颗卫星 / 30,000太比特/秒(是当前星座容量的35倍)。 * **每年5,000次星舰飞行(2,500次星链载荷):** 每年增加150,000颗V3卫星 / 150,000太比特/秒(是当前星座容量的176倍)——*将超过当前的授权限制*。 * 维持一个饱和星座每年需要约500次星舰飞行以更换卫星。 * **星链授权限制:** * SpaceX 目前根据第二代(Gen 2)批准(今年获批,是2020年申请的30,000颗卫星的50%)获准最多发射15,000颗卫星。第二代涵盖V2 mini和V3卫星。 * 官方授权总计约19,000颗卫星(第一代+第二代)。 * 当前运行中:约3,200颗(第一代)+ 7,900颗(第二代)。 * 当前第二代上限下的余量:7,100颗卫星。第一代批准正在逐步失效。 * SpaceX 已申请第三代(Gen 3)批准,最多100,000颗卫星(主要用于星链),以及100万颗卫星(主要用于Starmind)。这些目前都只是申请。 * 新的《Part 100》规则旨在加快卫星许可证的审批,SpaceX 已为第三代提交了申请。 * **路易斯安那州星际基地燃料与物流:** 计划在现场生产甲烷和液氧。亨利枢纽(天然气管道)位于北部几十英里处,这对于高频次发射所需的巨大燃料量至关重要(例如,每年1,000次星舰飞行将需要约100万吨甲烷和350万吨液氧,通过卡车/驳船运输不可行)。 **特斯拉FSD/视觉挑战(细水平线):** * **问题:** FSD 在处理线性水平物体时遇到困难,例如铁路道口、细链条、警示带以及I-95公路上的水平栏杆。 * **解释 (万事通博士):** 1. **卷积神经网络 (CNN) 池化:** 卷积神经网络通过“池化”(例如2x2池化)像素来减少大的输入空间。垂直线往往会“变粗”并在池化后得以保留,但如果细水平线与池化滤波器对齐不佳,则可能会消失,这意味着网络没有接收到初始信号。 2. **视差 (立体摄像头):** 前置摄像头为垂直物体提供视差,因为它们相对于背景的位移不同,从而产生强烈的信号。水平线则不会产生同样强烈的视差效应。 3. **视差 (运动):** 当汽车移动时,垂直的柱子会相对于背景显示出显著的视差,提供清晰的信号。随汽车移动的细水平线则不会表现出这种强烈的视差。 4. **体素大小:** 系统使用的“体素”(3D像素)可能为1-5厘米或更大。比这更薄的物体(例如小于1厘米的链条)可能会落在体素之间,或者被平均为空白空间。 5. **数据问题:** 保险杠高度的障碍物(链条、胶带、电缆、HOV车道分隔栏)在训练数据中很少见,并且可能没有像常见的垂直物体那样得到很好的标注。网络也可能将它们忽略为道路上的“水平杂物”,以避免误报。 * **结论:** 这是一个开放的研究问题;对于CNN来说,目前还没有很好的解决方案能够持续检测细水平线。这可能是特斯拉实现真正无监督驾驶的“最终大boss”之一。 **股市收盘情况:** * **特斯拉 (TSLA):** 348.75美元(下跌1.71%) * **SpaceX (私人公司):** 141.50美元(上涨0.45%) * **纳斯达克100指数 (NDX):** 下跌0.7%

Here's a summary of the Electrified episode, including every piece of news mentioned: **Website Updates & Trackers:** * All dashboards on dloomis.com have been updated and ported to Cursor. * **RoboTaxi Tracker (Texas):** Tesla now has 270 vehicles registered. An additional 89 vehicles were added to the fleet in August. All are currently Model Y, with cybercabs expected in the next few weeks. * **Tesla Semi Tracker:** 1,267 verified Tesla Semi Orders. * **Delivery Momentum:** Einride CEO stated ~75 of their 500 order are expected this year, with the remainder targeted by the end of next year (originally thought to be 2028). * **Elon Prediction Tracker:** Now ported, new predictions will be added. **Sponsor:** * **Bolt.new:** A platform designed for non-technical users to build and publish apps from an idea to a working product with minimal friction. Features include custom domains, GitHub push, and voice-to-app creation. A 30-day free pro account is available using the code "electrified". **Starbase Louisiana & Starlink Analysis:** * **Starbase Louisiana:** Will ultimately have over a dozen launch towers, enabling more than 30 Starship flights per day, making it the biggest launch site on Earth. * **Location Advantage:** Pecan Island is ~30 miles from the nearest town and ~50 miles from a major city, making it less disruptive than Boca Chica (5 miles from South Padre Island, 15 miles from Brownsville). It offers a wide range of optimal azimuths (launch directions), particularly ~190 degrees (south) for polar/SSO orbits, benefiting from the Gulf water. * **Starlink Launch Transition:** The first Starship flight out of Florida is a significant signal for Starlink. SpaceX's mission from Pad 40 was the last planned Falcon 9 Starlink launch from Florida; future Starlink missions from Florida will fly on Starship. Gigabay Florida and Starship launch pads are nearing completion. * **CyberCab Integration:** Every CyberCab will have a Starlink unit built in. * **Current Starlink Metrics:** * Operational satellites: ~11,000. * Downlink capacity: ~850 terabits per second. * Subscribers: ~12 million. * Hypothetical evenly distributed capacity: ~71 megabits per second per user. * **Starlink V3 & Starship Capacity:** * A Starship launch with V3 satellites will carry 60 V3 sats. * Each V3 satellite has a 1 terabit/second downlink capacity. * Each full Starship flight with V3s will deploy 60 terabits/second (equivalent to 7% of the entire current Starlink constellation). * Falcon 9 launches with V2 minis send up ~3 terabits/second. Starship with V3s is ~20 times that capacity. * SpaceX needs ~1,000 V3 satellites in orbit to meaningfully change performance, expected by Q2 next year. * **Starship Launch Mix & Future Projections:** * Host assumes 50% of Starship launches for Starlink over the next 1-3 years (the rest for NASA/HLS, tanker launches, Starlink mobile, and Starmind). * SpaceX plans to allocate significant launch capacity to its AI segment (Starmind/Orbital Data Centers) with Starship. * Towards the end of the decade, a majority of Starship launches are expected to be for Starmind. * **Hypothetical Starlink Capacity Growth with Starship:** * **100 Starship flights/year (50 Starlink payloads):** Adds 3,000 V3 satellites / 3,000 terabits per second annually (3.5 times the current constellation capacity). * **1,000 Starship flights/year (500 Starlink payloads):** Adds 30,000 satellites / 30,000 terabits per second annually (35 times the current constellation capacity). * **5,000 Starship flights/year (2,500 Starlink payloads):** Adds 150,000 V3 satellites / 150,000 terabits per second annually (176 times the current constellation capacity) – *would exceed current authorization limits*. * Maintaining a saturated constellation would require ~500 Starship flights annually for replacement satellites. * **Starlink Authorization Limits:** * SpaceX is currently authorized for up to 15,000 satellites under Gen 2 approval (granted this year, 50% of the 30,000 requested in 2020). Gen 2 covers V2 minis and V3 sats. * Officially authorized for ~19,000 satellites total (Gen 1 + Gen 2). * Current operational: ~3,200 (Gen 1) + 7,900 (Gen 2). * Headroom under current Gen 2 cap: 7,100 satellites. Gen 1 is a sunsetting approval. * SpaceX has requested approval for up to 100,000 satellites under a Gen 3 approval (largely for Starlink) and 1 million satellites (predominantly for Starmind). These are currently just requests. * New Part 100 rules aim to expedite satellite licenses, which SpaceX has requested for Gen 3. * **Starbase Louisiana Fuel & Logistics:** The plan is for on-site production of methane and liquid oxygen. The Henry Hub (natural gas pipeline) is a few dozen miles north, crucial for the massive fuel needs of high-cadence launches (e.g., 1,000 Starship flights/year would require ~1 million tons of methane and 3.5 million tons of liquid oxygen, not feasible with trucking/barging). **Tesla FSD/Vision Challenge (Thin Horizontal Lines):** * **Problem:** FSD struggles with linear horizontal objects like railroad crossings, thin chains, caution tape, and horizontal bars on I-95. * **Explanation (Dr. Know-It-All):** 1. **Convolutional Neural Network (CNN) Pooling:** CNNs reduce large input spaces by "pooling" pixels (e.g., 2x2 pooling). Vertical lines tend to "thicken" and survive pooling, but thin horizontal lines can disappear if they align unfavorably with the pooling filters, meaning the network doesn't receive the initial signal. 2. **Parallax (Stereo Cameras):** Front cameras provide parallax for vertical objects, creating a strong signal as they appear to shift differently against the background. Horizontal lines do not produce the same strong parallax effect. 3. **Parallax (Motion):** As a car moves, a vertical post shows significant parallax relative to the background, providing a clear signal. A thin horizontal line moving with the car doesn't exhibit this strong parallax. 4. **Voxel Size:** The "voxels" (3D pixels) used by the system may be 1-5cm or larger. Objects thinner than this (e.g., a chain <1cm) can fall between voxels or be averaged out as empty space. 5. **Data Problem:** Bumper-height obstacles (chains, tape, cables, HOV bars) are rare in training data and likely not as well-labeled as common vertical objects. The network might also ignore them as "horizontal clutter" on roads to avoid false positives. * **Conclusion:** This is an open research problem; no good solution exists for CNNs to consistently detect thin horizontal lines. This might be one of Tesla's "final bosses" for true unsupervised driving. **Stock Market Closings:** * **Tesla (TSLA):** $348.75 (down 1.71%) * **SpaceX (private):** $141.50 (up 0.45%) * **NDX:** Down 0.7%