Core Viewpoint - Tesla's Full Self-Driving (FSD) is characterized as an end-to-end large model, challenging the notion that it is merely a combination of nearly 200 small scene models [1][11]. Group 1: Model Architecture and Parameters - The B-core neural network parameters significantly exceed those of the A-core, with only 61 shared parameter files, indicating that the redundancy design between A and B cores has become impractical with the rapid expansion of the neural network scale in Tesla V12 [5]. - The discovery of many model parameters being parts of a large model, indicated by naming conventions like FSD E2E FACTORY PART X, suggests a distributed deployment strategy for model parameters across different chips, which is common in the era of large models [6]. - Tesla's HW3 has limited memory bandwidth of 68GB/s, theoretically allowing for a maximum of 1.8GB of model parameters to support a 36Hz output, while HW4, with a bandwidth of 384GB/s, could theoretically support around 10 billion parameters [7][8]. Group 2: Mixture of Experts (MoE) Architecture - The use of a Mixture of Experts (MoE) architecture allows Tesla to run large-scale end-to-end models at high frequencies on relatively older chips by activating only a subset of expert networks, thus optimizing memory bandwidth usage [8][10]. - Elon Musk and Ashok Elluswamy have indicated that the FSD employs MoE architecture, which supports the idea of localized parameters for different regions while maintaining a generalized approach [9][10]. Group 3: Technological Advancement - The assertion that FSD is a backward technology is dismissed, emphasizing that technological advancement is not solely defined by scientific discoveries but also by engineering innovations, as exemplified by Tesla's achievements in rocket technology and engineering [11].

Core Viewpoint - The article discusses the advancements in battery technology by the company, focusing on the development of a battery with a low internal resistance of 0.30 mΩ, which is crucial for achieving high charging speeds and extending battery life [1][4]. Group 1: Battery Development Challenges - During the high-power charging process, battery overheating can significantly affect its long-term lifespan, necessitating effective temperature control [1]. - The two core strategies for temperature control are reducing battery heat generation (lowering internal resistance) and enhancing the cooling capabilities of the battery system [1]. - The company aimed for an internal resistance of 0.30 mΩ, while the industry standard was around 0.5 mΩ, with leading suppliers estimating a minimum of 0.38 mΩ [1]. Group 2: Technical Achievements - The company formed a specialized task force that spent one year overcoming the technical challenges associated with achieving the target internal resistance [1]. - A high-precision model was developed to decouple the internal resistance of the battery cell into three layers and 17 components, which is more detailed than the industry norm of two layers and 5-6 components [1][2]. Group 3: Charging Performance and User Satisfaction - The company set ambitious goals for battery performance, including a range of over 700 km (CLTC) and the ability to charge 400 km in just 10 minutes [4]. - By March 2024, the company aims to achieve a charging time of 12 minutes for 500 km, with subsequent improvements leading to 19 minutes and 15 minutes for 95% charge by mid-2024 [4]. - User satisfaction regarding charging (NPS) has increased from 91.6 to 98.4, indicating a positive reception of the charging performance improvements [4].

TOP2不掌握任何非公开信息，纯粹基于公开信息推理。 根据理想官网，理想正在急招 灵巧手机械设计[1]/算法研发[2]/嵌入式软件工程师[3]，嵌入式软 件工程师-传感器[4]，关节模组嵌入式软件[5]/机械设计[6]工程师，嵌入式硬件工程师[7]，双足 算法研发工程师[8]，全身运控算法研发工程师[9]。 TOP2基于这9份岗位描述推理，理想在研发一款面向工业制造场景的高性能人形机器人(双足+灵巧 手)，谋求量产，不是谋求Demo展示或科研场景，也不是谋求面向普通家庭的消费级产品。 本文侧重分享推理出理想主观诉求的过程，不涉及分析到底这个人形机器人到底能不能做出来。 机械设计工程师-关节模组[6]要求候选人负责解决试产及量产阶段的结构性问题(如装配干涉、密封失 效等)，并要求参与材料选型及供应商技术评估。 在嵌入式硬件工程师[7]要求候选人必须具备大功率 电源产品成功量产的Layout经验，并需要解决DFM(可制造性设计)和DFA(可装配性设计)相关问题。 只有进入正式量产流程的产品，才会对DFM/DFA和供应商评估提出如此具体的要求，Demo不需要考 虑这些。 嵌入式软件工程师-灵巧手[3]机械设计工程师 ...

Core Viewpoint - The company is actively recruiting for various engineering positions related to humanoid robotics, indicating a strategic focus on developing advanced robotic technologies. Group 1: Job Positions and Responsibilities - The company is hiring for roles including dexterous robotic hand design, algorithm development, embedded software engineering, and joint module engineering, suggesting a comprehensive approach to humanoid robot development [1][5] - Specific roles emphasize the need for experience in bipedal walking algorithms, coordination of body parts, and the design of multi-joint systems, indicating a focus on complex motion control [1][8][9] Group 2: Technical Specifications - The robotic hand will utilize a direct drive structure rather than a cable-driven system, highlighting a preference for advanced motor technologies such as hollow cup and brushless DC motors [2] - The core driving mechanism will involve high-power joint modules with hundreds of amperes and a hybrid architecture combining reinforcement learning and model predictive control, emphasizing the importance of advanced algorithms in motion control [3] - The system architecture will incorporate STM32 or TI C2000 series microcontrollers, with real-time operating systems like FreeRTOS and RT-Linux, indicating a focus on high-performance computing and real-time processing [4] Group 3: Design Challenges - The design of the dexterous hand requires independent control of fingers and the ability to handle multimodal data from visual and tactile sensors, which presents significant engineering challenges in terms of bandwidth and latency [4] - The integration of high-torque direct drive motors in compact spaces while managing heat dissipation and wiring issues is a critical design challenge, necessitating collaboration between mechanical and electronic engineering teams [4]

2026年1月19日知乎用户刘延发布《 智驾行业学习笔记|特斯拉FSD模型的非共识 》 最关键的一点是基于X用户green在2025年4月发的推文，如果认可green说的为真，则V13就一定有189个神经网络模型。 节点 A 包含 189 个神经网络，而节点 B 包含 110 个神经网络，61 个神经网络为节点 A 与节点 B 所共享。 HW3 与 HW4 平台之间共享的神经网络数量总计达到 135 个。HW3 平台在当前的 v12.6 版本中，其节点 A 大小为 1.2G， 节点 B 大小为 3.1G；HW4在v13版本中，节点 A 大小为 2.3G，节点B 的大小增至 7.5G。 Got a bit of free time over the weekend and noticed that HW4 size of NNs in v13 ballooned from 2.3G total in v12.x to 7.5G on just node B in v13 (and 2.3g on node A). 由 Google 翻译自英语 周末有点空闲时间,注意到 v13 中 NN 的 HW4 大小从 v1 ...

Core Insights - The article discusses Li Auto's ambitious growth plans for 2026, targeting a 40% increase in sales, which translates to approximately 550,000 units sold [1] - The upgraded L9 model features a battery capacity exceeding 70 kWh, a range of over 400 kilometers, and significant enhancements in size and chassis design [1] - Industry insiders predict that the L9 could achieve sales of over 100,000 units [1] Group 1 - Li Auto is focusing resources on the next-generation L9 model while streamlining its range of extended-range vehicles [1] - A new all-electric SUV is set to launch in 2026, with a product focus returning to extended-range vehicles [1] - The company anticipates that its market share in the 200,000+ price segment will not return to the 40%+ levels seen in 2022 and 2023, but aims to regain a leading position in market share by 2026 [1] Group 2 - The i6 model is expected to have a production capacity of 3,000 units per month [1] - Li Auto is undergoing a systematic review of its offline sales network, with plans to close underperforming retail locations and not renew contracts [1]

Core Viewpoint - The article discusses the expansion plans of Li Auto, highlighting the growth in the number of supercharging stations across various provinces in China, with a target of 1,680 high-speed supercharging stations by the end of 2026, representing a potential increase of approximately 40% from current numbers [1]. Summary by Relevant Sections Expansion Plans - By January 2026, Li Auto aims to have approximately 1,104 service area supercharging stations (excluding nearby highway stations) and plans to build 1,680 high-speed supercharging stations by the end of 2026 [1]. - The company is focusing on increasing its presence in provinces with high highway mileage and service area supercharging stations [1]. Provincial Breakdown - Six provinces have over 90 high-speed service stations: Chongqing (135), Zhejiang (121), Jiangxi (113), Hebei (93), Anhui (92), and Sichuan (92) [1]. - Eight provinces have an average of one Li Auto supercharging station within 120 kilometers in one direction: Chongqing (33), Tianjin (42), Zhejiang (46), Jiangxi (61), Anhui (65), Hebei (94), Tibet (102), and Sichuan (112) [1]. - Fourteen provinces have fewer than 15 service area stations, with the lowest being Inner Mongolia, Heilongjiang, and Liaoning, each having no stations [1]. Highway Mileage and Service Stations - The article provides a detailed table of highway mileage and the number of service area supercharging stations for various provinces, indicating the average distance between stations [2][3][5][6][7][8][9][10]. - For example, Chongqing has 4,500 km of highways with 135 service area stations, averaging 33 km between stations, while Tibet has only 407 km of highways with 4 service area stations, averaging 102 km between stations [2][10].

Core Viewpoint - The company has developed a new AI glasses product called Livis, which aims to extend the smart experience from cars to everyday life, emphasizing comfort, long battery life, and fast response times [1][48]. Group 1: Product Features - Livis glasses weigh only 36 grams, making them one of the lightest in the industry [25]. - The glasses offer a typical usage time of 18.8 hours and standby time of over 70 hours [25]. - The AI response speed is notably fast, achieving 800 milliseconds [25]. - The product has sold out its initial production capacity in just three days, indicating strong market demand [32][33]. Group 2: User Interaction and Experience - The average daily interaction with the glasses is around 40 to 50 voice commands, significantly higher than the 18 to 20 interactions typically seen in cars [36]. - Users engage with the glasses for various tasks, including navigation, media control, and reminders, showcasing their utility beyond just a visual device [37][38]. Group 3: Development and Technology - The glasses utilize a unique single MCU architecture and a custom RTOS, differentiating them from competitors who rely on standard platforms [5][51]. - Collaboration with partners like Hengxuan has led to the development of a specialized system that enhances performance while maintaining low power consumption [5][54]. - The design process involved high standards, such as achieving G3 continuous surfaces, which are typically reserved for luxury products [13][56]. Group 4: Market Position and Strategy - The company aims to create a seamless user experience across multiple devices, positioning the glasses as a step towards a multi-device ecosystem [7][8]. - Despite initial skepticism about diversifying from automotive products, the company believes that the glasses will enhance user engagement and create new value [10][50]. - The product's success is seen as a potential catalyst for future innovations in both the eyewear and automotive sectors [12][53].

Core Viewpoint - The investment decision in Li Auto was significantly influenced by the organizational structure and management approach of the company, particularly the emphasis on aligning the team and fostering communication among employees [2][3]. Group 1: Organizational Structure and Management - The importance of organizational alignment in the electric vehicle industry was highlighted, as it is crucial for integrating diverse talents from automotive, software, and AI backgrounds [2][3]. - Li Auto's CEO, Li Xiang, invested considerable effort in establishing a clear organizational system that ensures all employees understand the company's direction, which instilled confidence in potential investors [3][4]. - Conversations with team members, such as Fan Haoyu, provided deeper insights into the company's internal dynamics, which were deemed more persuasive than discussions with the CEO alone [5][7]. Group 2: Investment Philosophy - The investment strategy focused on identifying companies capable of producing smart electric vehicles rather than traditional electric vehicle manufacturers, emphasizing the need for a strong software development team [9]. - The decision to invest in Li Auto over competitors like Xiaopeng was influenced by Li Xiang's understanding of consumer preferences and his strategic approach to product development [10]. - The company's ability to efficiently utilize funds was recognized as a significant advantage, particularly in a capital-intensive industry like automotive manufacturing [5][11]. Group 3: Funding Challenges and Support - Li Auto faced significant funding challenges, particularly in 2018 and 2019, where external support from investors like Zhang Ying was crucial for the company's survival and growth [10][11]. - The narrative of seeking financial support from influential figures in the industry, such as Wang Xing from Meituan and Zhang Yiming from ByteDance, illustrates the importance of networking and relationships in securing funding [11].

Core Points - The company aims to be among the top three in the domestic large language model sector by December 2024, as stated by Li Xiang during the AI Talk [1][2] - As of December 27, 2024, the potential competitors in the large language model space include Doubao, DeepSeek, Qwen, and Kimi, with a significant challenge to surpass at least two of them by 2025 [1][3] - By January 17, 2026, the competition may also include MiniMax, making it increasingly difficult for the company to outperform three out of these five competitors, particularly DeepSeek [1][3] - In the field of embodied intelligence, the company has a viable opportunity to become a leader, but this is contingent on industry development and Li Xiang's learning and decision-making capabilities [1][4] - Li Xiang envisions achieving capabilities similar to Jarvis from Iron Man, but this is expected to take a considerable amount of time [1][4] Detailed Analysis - Li Xiang emphasized the necessity for the team to ensure that their foundational model for large language models ranks within the top three in China over the next few years, indicating a commitment to invest in the required computational power [2] - The current mobile application of the company is perceived as subpar, and there is no clear standard provided by Li Xiang for evaluating the ranking of large language models [2][3] - The company faces significant challenges in surpassing competitors like DeepSeek, which is noted for its high originality and impactful scientific results [3] - The landscape of embodied intelligence is categorized into four main factions, with the company positioned within the full-stack AI hardware-software integration camp, which could potentially lead to a top position in this domain [4][6] - Li Xiang's past experiences in the automotive media industry highlight his strategic approach to achieving competitive rankings, which may inform his current objectives in AI [6][8] - Observations from industry peers indicate that Li Xiang is rapidly evolving in his understanding of AI, which is crucial for the company's future success [7][8]