首页  >>  来自播客: Electrified 更新   反馈  

Electrified - Tesla Just Rewrote FSD Entirely - The Major Technical Breakthrough of FSD 14.3 ⚡️

发布时间:   原节目
以下是内容的中文翻译: 迪伦·卢米斯(Dylan Loomis)开篇承认,用户对早期14.3 FSD视频和反馈感到失望,但他认为真正的重点应该放在一次重要的底层重写上。 **FSD 14.3:根本性重写** 核心信息是,特斯拉从零开始对其AI编译器和运行时系统进行了“大规模的、深层次的、根本性的重写”,并利用了MLIR(多级中间表示)。这使得反应时间加快了20%,并提高了模型迭代速度。 * **编译器的作用:** AI编译器将特斯拉训练好的神经网络FSD模型(复杂的矩阵运算)转换为可在AI4和Hardware 3芯片上执行的代码。运行时系统则实时执行这些代码。 * **MLIR的优势:** 与之前的“扁平层”编译器不同,MLIR引入了三层结构(高层、中层、低层)。这使得每一层都能进行专门优化,工程师可以在不破坏其他部分的情况下进行更改,累积改进,并简化调试(可能在几分钟而非几天内发现bug)。 * **影响:** 这次重写为更大的视觉编码器、更多的参数以及增强的推理能力创造了巨大的“发展空间”。它还使特斯拉能够更快地迭代和部署新训练的模型。 * **专家认可:** LLVM(开发了MLIR)的创始人、特斯拉前自动驾驶软件副总裁克里斯·拉特纳(Chris Lattner)称此次采用为潜在的“突破”。 * **战略举措:** 特斯拉在多年来一直守口如瓶之后,决定公布这一技术细节,这被视为其重要性的一个信号。 * **未来展望:** 迪伦认为14.3是一个新的“基础”,将为后续的小版本更新(14.3.x)中的“炫目的面向客户的功能”奠定基础。尽管目前存在实际性能问题,他仍为埃隆·马斯克“拼图的最后一块重要部分”的评论提供合理解释,并表示希望在未来三个月内看到“惊喜”。 * **Hardware 3:** 这次重写带来的额外潜力也让老款Hardware 3单元的车主感到鼓舞。 **FSD 14.3 初期用户反馈与轶事:** * **停车:** 在停车场内,决策速度更快,信心更足,但并非完美;仍然存在问题,车辆有时会陷入循环。新的蓝色“P”停车预测图标通常会指向离入口更近的停车位,且不可调节。 * **标志识别:** 据说有所改进,但仍不完美(例如,驶过“禁止驶入”标志,停在预留车位)。Dirty Tesla报告了一个案例,其中它*确实*识别了“禁止驶入”标志。 * **错误:** 报告了“红手”错误,表明存在并非实际发生的摄像头遮挡,但车辆仍继续行驶。 * **驾驶:** 反馈好坏参半——一些人称其为“有史以来最平稳、最自信的版本”,而另一些人则指出了一些问题,例如在免下车餐厅过早起步、在停车场内陷入无限循环、对警察手势识别不一致、导航问题、“左车道占道”问题重现,以及在小雨中被强制进入“冷静/缓慢模式”且无法超控。 * **反应时间:** 测试中被指出“相当惊人”。 * **速度曲线:** 似乎基本没有变化。 **其他新闻:** * **Cybercab目击:** 在奥斯汀发现了大约60辆基于Cybertruck的“Cybercab”,被识别为工程测试车,带有可见方向盘,表明正在全国范围内进行广泛测试和验证。 * **驾驶员监控系统(DMS):** 由于NHTSA的监管压力以及欧洲等市场对监督式FSD的需求,改进仍在继续。 * **推理:** 尽管14.3的主要推理重点是目的地停车,但推理元素(例如施工期间导航路线变更)在14.2中已经存在,并且正在模型中逐步扩展。 * **TeraFab项目:** 英特尔已加入SpaceX、XAI和特斯拉,共同重构硅晶圆制造技术,目标是每年生产1太瓦的计算能力。此次合作显著降低了项目风险,利用了英特尔的制造专长以及美国政府通过《芯片法案》提供的近90亿美元股权投资,该法案优先发展美国半导体制造业。 * **特斯拉股价:** 在五天内获得了2.56亿美元的散户资金流入,表明尽管其他“MAG-7”股票需求降温,投资者仍对其抱有强烈信念。特斯拉股价收盘下跌0.98%,报343.25美元。 * **Waymo扩张:** Waymo已在田纳西州纳什维尔正式向公众开放,覆盖60平方英里的地理围栏,并计划在机场提供服务。一起事件显示一辆Waymo开到了新铺设的沥青路上并绕过了一名工人。 * **德国销量:** 3月份是特斯拉在德国有史以来销量最好的月份,2024年第一季度的销量约占特斯拉2023年全年总注册量的66%,表明强劲复苏。 * **特斯拉Semi:** 一辆特斯拉Semi被发现拖着一辆Watt EV和亚马逊的拖车。 * **Model S/X库存:** Model S和X的库存只剩下几百辆。

Dylan Loomis opens by acknowledging user disappointment with early 14.3 FSD videos and feedback, but argues the true focus should be on a significant underlying rewrite. **FSD 14.3: The Fundamental Rewrite** The core message is that Tesla undertook a "massive, deep-seated, fundamental rewrite" of its AI compiler and runtime from the ground up, utilizing MLIR (Multi-Level Intermediate Representation). This resulted in a 20% faster reaction time and improved model iteration speed. * **Compiler's Role:** The AI compiler translates Tesla's trained neural network FSD model (complex matrix math) into executable code for AI4 and Hardware 3 chips. The runtime then executes this in real-time. * **MLIR Benefits:** Unlike the previous "flat layer" compiler, MLIR introduces a three-level structure (high, medium, low). This allows for specialized optimizations at each layer, enabling engineers to make changes without breaking other parts, compounding improvements, and simplifying debugging (potentially spotting bugs in minutes instead of days). * **Impact:** This rewrite creates significant "headroom" for larger vision encoders, more parameters, and enhanced reasoning. It also enables Tesla to iterate and deploy newly trained models much faster. * **Expert Endorsement:** Chris Lattner, founder of LLVM (which developed MLIR) and former VP of Autopilot Software at Tesla, called this adoption a potential "breakthrough." * **Strategic Move:** Tesla's decision to call out this technical detail, after years of being tight-lipped, is seen as a signal of its importance. * **Future Expectations:** Dylan believes 14.3 is a new "foundation" that will unlock future "flashy customer-facing features" in subsequent dot releases (14.3.x). He justifies Elon Musk's "last big piece of the puzzle" comment, despite current real-world performance issues, and expresses hope for "fireworks" within the next three months. * **Hardware 3:** The rewrite's potential for extra headroom also encourages owners of older Hardware 3 units. **FSD 14.3 Initial User Feedback & Anecdotes:** * **Parking:** Decisiveness in parking lots is faster and more confident but not perfect; still sees issues and cars getting stuck in loops. The new blue 'P' pin for parking prediction often targets spots closer to entrances and isn't adjustable. * **Sign Reading:** Anecdotally improved, but still imperfect (e.g., driving through "do not enter" signs, parking in reserved spots). Dirty Tesla reported an instance where it *did* read a "do not enter" sign. * **Bugs:** "Red hands" bug reported, indicating a camera occlusion that wasn't present, though the car continued driving. * **Driving:** Mixed reports – some call it the "smoothest and most confident version ever," while others note issues like taking off too early at drive-thrus, getting stuck in infinite loops in parking garages, inconsistent police hand signal recognition, navigation problems, a return of "left lane camping," and being forced into "chill/sloth mode" in light rain without override. * **Reaction Time:** Noted to be "quite incredible" in testing. * **Speed Profiles:** Seem largely unchanged. **Other News:** * **Cybercab Sightings:** Around 60 Cybertruck-based "cybercabs" were seen in Austin, identified as engineering vehicles with visible steering wheels, suggesting extensive testing and validation across the country. * **Driver Monitoring System (DMS):** Improvements are continuing due to regulatory pressure from NHTSA and the need for supervised FSD in markets like Europe. * **Reasoning:** While 14.3's primary reasoning focus is on destination parking, elements of reasoning (e.g., navigation route changes during construction) were already present in 14.2, and it's being gradually expanded across the model. * **TeraFab Project:** Intel has joined SpaceX, XAI, and Tesla to refactor silicon fab technology, aiming to produce 1 terawatt per year of compute. This partnership significantly de-risks the project, leveraging Intel's manufacturing expertise and a nearly $9 billion equity investment from the U.S. government via the CHIPS Act, which prioritizes American semiconductor manufacturing. * **Tesla Stock:** Experienced $256 million in retail inflows over five days, indicating strong conviction among investors despite cooling demand for other "MAG-7" stocks. Tesla stock closed down 0.98% at $343.25. * **Waymo Expansion:** Waymo has officially opened to the public in Nashville, TN, covering a 60-square-mile geofence and planning service at the airport. An incident showed a Waymo driving onto fresh asphalt and around a worker. * **Germany Sales:** March was Tesla's best sales month ever in Germany, with Q1 2024 volumes representing approximately 66% of Tesla's total registrations for all of 2023, suggesting a strong recovery. * **Tesla Semi:** A Tesla Semi was spotted with a Watt EV and Amazon trailer. * **Model S/X Inventory:** Only a few hundred Model S and X vehicles are left in inventory.