首页  >>  来自播客: Electrified 更新   反馈  

Electrified - Tesla's Shocking Internal AI Use / Elon & SpaceXAI Have Officially Arrived ⚡️

发布时间:   原节目
以下是播客内容的中文摘要,其中包含所有新闻条目: 主持人 Dylan Lumis 首先承认了人工智能(AI)发展速度惊人,并建议大家不要有“错失恐惧症”(FOMO),强调提高生产力比使用绝对最新的模型更为重要。 **特斯拉的 AI 政策与内部使用:** * **AI Token 支出上限:** 特斯拉已在全公司范围内实施 AI token 每周每位工程师200美元的支出限制。这与此前“无限制”的时代形成了重大转变,过去排行榜还鼓励高使用量(现在可能变成“耻辱榜”)。经高层批准可以获得豁免。 * **Grok 内测:** 特斯拉工程师免费内测 Grok 产品(如 Grok Build),其使用量不计入200美元的限制。XAI 的产品负责人 Andrew Millich 积极与特斯拉工程师合作以获取反馈。 * **工程师偏好:** 据报道,特斯拉工程师更喜欢 Claude Code 和 Codex,而非 Grok。 * **埃隆的鼓励:** 埃隆·马斯克已发送全公司邮件,鼓励使用 AI 代理并推广 Cursor 的 Composer 编码模型。 * **Lars 访谈 - 特斯拉内部 AI 应用:** 1. **经验教训数据库:** 特斯拉将所有错误(例如,电池故障、逆变器问题)整合到一个可供工程师和供应链团队进行 AI 搜索的数据库中。 2. **异常发现:** 新的 AI 工具优化了工厂排行榜(KPI、KPC)的跟踪,以发现设备和加工中细微的偏差。 3. **系统联动与推理:** 各部门系统现在可以相互“推理”,从而使下游问题自动触发上游修正(例如,生产线末端车辆扫描数据立即通知冲压团队)。 4. **服务知识共享:** 服务维修、纠正措施以及客户投诉/解决方案被记录并立即在所有服务中心共享,从而缩短了服务等待时间。 **Grok 4.5 发布与性能:** * **注重现实世界:** 埃隆·马斯克强调 Grok 4.5 旨在解决特斯拉、SpaceX、Neuralink 和 Boring Co. 的“现实世界工程问题”,目标是“现实世界的实用性,而非基准分数”。 * **工程师反馈:** 特斯拉工程师 Yunto Sai 证实 Grok Build 已成为提高生产力的“不可或缺的工具”,以“最少的设置提供深刻的见解”。 * **定价:** Grok 4.5 比 Claude 便宜得多: * 20万 token 以下上下文:输入 $2/M,输出 $6/M。 * 20万 token 以上上下文:输入 $4/M,输出 $12/M。 * Claude (Opus) 对比:输入 $10/M,输出 $50/M。 * **上下文窗口:** Grok 4.5 当前的50万 token 上下文窗口“可能在下周”升级到100万。 * **Token 效率:** Artificial Analysis 的数据显示,Grok 4.5 每任务使用190万 token,而 Opus 4.8 使用920万 token(效率大约高出4.8倍)。 * **每任务成本 (Cursor Bench):** Grok 4.5 每任务成本为1.51美元,显著优于 Fable 5 Max(17.32美元)。 * **重要提示:** Cursor 透露,Grok 4.5 在 Cursor Bench 上拥有“无意中”的优势,因为它在训练中包含了 Cursor 代码库的早期快照,这意味着结果可能受到记忆而非纯粹问题解决能力的影响。 * **模型详情:** Grok 4.5 是一款“Opus 级别的模型,但速度更快,Token 效率更高,成本更低。” 它建立在1.5万亿参数的 v9 基础上(高于4.3版的5000亿参数)。部分 Cursor 数据用于后期的训练阶段。 * **应用:** 主要用于编码(端到端应用构建、Excel、Word/PowerPoint)。 * **法律基准测试:** 它在 Harvey 的法律代理基准测试中出人意料地排名第一。 * **访问方式:** 通过 Cursor 应用(终端/命令行界面)可用。 * **智能指数:** Grok 4.5 在 Artificial Analysis 智能指数上得分为54分(高于 Grok 4.3 的38分),排名第四,仅次于 Fable 5、Opus 4.8 和 GPT 5.5。 **SpaceX AI 与未来战略:** * **Dylan 的预测:** 到2028年,所有其他 AI 公司都将“全力追赶 SpaceX AI”。 * **垂直整合 (Cursor & Origin):** Cursor 现已归属 SpaceX,开发了“Origin”,它是 GitHub 的直接竞争对手,旨在将 AI 代理视为主要的提交代码者。这从真实的开发者工作流程中创造了一个独特的训练数据飞轮。 * **快速发展:** SpaceX 计划每月发布一个新模型。一个2万亿参数的基础模型预计将于本月完成训练,并于下月向客户提供。 * **硬件加速:** SpaceX AI 正在开发将 C/C++ 推理直接映射到 GB300 硬件的技术,可能使速度提高一倍或更多。 * **现实世界影响:** Grok 4.5 在现实世界中获得的积极反响表明 SpaceX AI 是一个重要的参与者,在上下文窗口管理和每 token 成本方面表现出色,这对于95%以上不需要“前沿”原始智能的应用至关重要。 * **埃隆的抱负:** 埃隆相信 SpaceX 最终的价值将超过地球上其他所有事物的总和。 **特斯拉 RoboTaxi 更新:** * **迈阿密上线:** 特斯拉在迈阿密推出 RoboTaxi 所使用的测试司机团队比一年前在奥斯汀使用的300多名司机“小得多”。埃隆希望软件无需广泛的本地测试即可部署。 * **迈阿密详情:** 当前的地理围栏很小(约15平方英里),车辆数量可能少于10辆。 * **FSD 语音指令:** 尽管之前的一些 FSD 功能(如速度控制)因模型复杂性而被移除,Ashok Elluswamy 仍在“努力开发”用户请求的功能,例如告诉汽车驶入哪个车道或识别地标。 * **NHTSA 与方向盘:** 美国国家公路交通安全管理局(NHTSA)负责人表示,该机构将“绝对考虑”取消无人驾驶汽车方向盘的要求,这可能会为自动驾驶汽车公司带来提振,此前他们已对刹车踏板要求进行了审查。 **其他特斯拉新闻:** * **特斯拉 Semi 订单:** QX Logistics,特斯拉和 Forian Mobility 的新客户,订购了20辆特斯拉 Semi,预计2027年交付。这使得公开确认的 Semi 订单总数达到666辆。 * **Cybercab 硬件:** “NotATeslaApp”报道称,Cybercab 生产单元将配备“更强大的双 GPS 配置”,用于精确位置映射,可能有助于远程操作。 * **Model Y 销量:** Model Y 在今年上半年是荷兰所有燃料类型中最畅销的车型。 **市场表现:** * 特斯拉股价收盘报406.55美元,上涨3.17%。 * SpaceX 股价收盘报152.16美元,上涨2.6%。 * NDX(纳斯达克100指数)上涨1.62%。

Here's a summary of the podcast, including every single news item: The host, Dylan Lumis, starts by acknowledging the overwhelming pace of AI development and advises against FOMO, emphasizing that being productive is more important than using the absolute latest model. **Tesla AI Policy & Internal Use:** * **AI Token Spend Cap:** Tesla has implemented a company-wide AI token spend limit of $200 per week per engineer, a significant shift from a previous "free for all" era where leaderboards encouraged high usage (now potentially a "board of shame"). Exemptions are possible with higher-up approval. * **Grok Beta Testing:** Tesla engineers beta-test Grok products (like Grok Build) for free, and this usage is not counted towards the $200 limit. XAI's product lead, Andrew Millich, actively works with Tesla engineers for feedback. * **Engineer Preference:** Tesla engineers reportedly prefer Claude Code and Codex over Grok. * **Elon's Encouragement:** Elon Musk has sent company-wide emails encouraging AI agent use and promoting Cursor's Composer coding model. * **Lars Interview - Tesla's Internal AI Applications:** 1. **Lessons Learned Database:** Tesla consolidated all errors (e.g., battery failures, inverter problems) into an AI-searchable database for engineers and supply chain teams. 2. **Anomaly Spotting:** New AI tools refine tracking on factory leaderboards (KPIs, KPCs) to spot subtle drifts in equipment and machining. 3. **System Linking & Reasoning:** Systems across departments now "reason" with each other, enabling downstream issues to automatically trigger upstream corrections (e.g., end-of-line vehicle scan data immediately informs the stamping team). 4. **Service Knowledge Sharing:** Service fixes, corrections, and customer complaints/solutions are logged and instantly shared across all service centers, leading to reduced service wait times. **Grok 4.5 Launch & Performance:** * **Real-World Focus:** Elon Musk emphasized Grok 4.5 is built for "real-world engineering problems" at Tesla, SpaceX, Neuralink, and Boring Co., aiming for "real-world usefulness, not benchmarks." * **Engineer Feedback:** Tesla engineers like Yunto Sai confirm Grok Build became an "indispensable tool" for productivity, delivering "deep insights with minimal setup." * **Pricing:** Grok 4.5 is significantly cheaper than Claude: * Under 200k context: $2/M input, $6/M output tokens. * Over 200k context: $4/M input, $12/M output tokens. * Claude (Opus) comparison: $10/M input, $50/M output tokens. * **Context Window:** Grok 4.5's current 500,000-token context window will upgrade to 1 million "probably by next week." * **Token Efficiency:** Artificial Analysis showed Grok 4.5 using 1.9 million tokens per task compared to Opus 4.8's 9.2 million (roughly 4.8x more efficient). * **Cost Per Task (Cursor Bench):** Grok 4.5 achieved $1.51 per task, significantly outperforming Fable 5 Max ($17.32). * **Crucial Caveat:** Cursor disclosed that Grok 4.5 had an "unintentional" advantage on Cursor Bench due to an earlier snapshot of the Cursor code base being included in its training, meaning results might be influenced by memorization rather than pure problem-solving. * **Model Details:** Grok 4.5 is an "opus class model but faster, more token efficient, and lower cost." It's built on a 1.5 trillion parameter v9 foundation (up from 4.3's 500 billion). Some Cursor data was used in later training stages. * **Applications:** Primarily built for coding (end-to-end app builds, Excel, Word/PowerPoint). * **Legal Benchmark:** It surprisingly ranks #1 on Harvey's legal agent benchmark. * **Access:** Available via the Cursor app (terminal/CLI). * **Intelligence Index:** Grok 4.5 scored 54 on the Artificial Analysis Intelligence Index (up from Grok 4.3's 38), placing it fourth behind Fable 5, Opus 4.8, and GPT 5.5. **SpaceX AI & Future Strategy:** * **Dylan's Prediction:** By 2028, all other AI companies will be "chasing down SpaceX AI." * **Vertical Integration (Cursor & Origin):** Cursor, now under SpaceX, has developed "Origin," a direct competitor to GitHub, built to treat AI agents as primary code committers. This creates a unique training data flywheel from real developer workflows. * **Rapid Development:** SpaceX plans to launch a new model every month. A 2 trillion parameter foundation model is expected to finish training this month and be available to customers next month. * **Hardware Acceleration:** SpaceX AI is working on C/C++ inference mapped directly to GB300 hardware, potentially doubling speed or more. * **Real-World Impact:** Grok 4.5's positive real-world reception suggests SpaceX AI is a significant player, excelling in context window management and cost per token, crucial for 95%+ of applications that don't need "frontier" raw intelligence. * **Elon's Ambition:** Elon believes SpaceX will eventually be worth more than the rest of Earth. **Tesla RoboTaxi Updates:** * **Miami Launch:** Tesla's RoboTaxi launch in Miami involved a "much smaller" test driver team compared to the 300+ drivers used in Austin a year prior. Elon wants the software to deploy without extensive local testing. * **Miami Details:** The current geofence is small (around 15 sq miles), and the number of vehicles is likely under 10. * **FSD Voice Commands:** Ashok Elluswamy is "working on" user-requested features like telling the car which driveway to pull into or identifying landmarks, despite previous FSD features (like speed control) being removed due to model complexity. * **NHTSA & Steering Wheels:** The head of NHTSA stated the agency will "absolutely consider" ending requirements for steering wheels in driverless cars, a potential boost for autonomous vehicle companies, following current reviews for brake pedal requirements. **Other Tesla News:** * **Tesla Semi Order:** QX Logistics, a new customer of Tesla and Forian Mobility, placed an order for 20 Tesla Semis, with delivery expected in 2027. This brings the total publicly confirmed Semi orders to 666. * **Cybercab Hardware:** "NotATeslaApp" reported that Cybercab production units will feature a "more powerful dual GPS configuration" for pinpoint location mapping, potentially aiding remote operations. * **Model Y Sales:** The Model Y was the best-selling vehicle in the Netherlands in the first half of this year across all fuel types. **Market Performance:** * Tesla stock closed at $406.55, up 3.17%. * SpaceX stock closed at $152.16, up 2.6%. * The NDX (Nasdaq 100) was up 1.62%.