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 Viewpoint - The article discusses the application of deep learning models, particularly LSTM, in predicting government bond yields, highlighting its advantages in handling complex financial time series data [4][5]. Group 1: Financial Time Series Prediction - Financial time series prediction has evolved through three main stages: traditional econometric models, traditional machine learning models, and deep learning models [4]. - Deep learning models, especially LSTM, are currently among the mainstream methods for financial time series prediction due to their ability to adapt to non-stationary, non-linear, high-noise, and long-memory characteristics [4]. Group 2: LSTM Model for Bond Yield Prediction - A three-layer stacked LSTM model with Dropout regularization was developed to predict the 10-year government bond yield, exploring the application and effectiveness of deep learning in fixed income quantitative analysis [5]. - The model utilized data from early 2021 to December 12, 2025, with approximately 130,000 adjustable parameters, achieving an average absolute error of 1.43 basis points in predictions [5]. - The model predicts a slight decline in the 10-year government bond yield, with a forecasted value of 1.8330% for December 19, 2025, down from 1.8396% on December 12, 2025 [5]. Group 3: Future Optimization Directions - Future optimizations include adjusting the model design regarding time windows, data processing, network architecture, and training strategies [6]. - Expanding input variables to include macroeconomic, market, and sentiment data will enhance the model's predictive accuracy and economic logic [6]. - Combining LSTM with traditional econometric models or other machine learning models to create hybrid models like ARIMAX-LSTM and CNN-LSTM-ATT can improve prediction precision [7].

Core Insights - Tesla's CEO Elon Musk announced the detailed rollout plan for the highly anticipated FSD V14 version, with the first deployment expected next week, followed by V14.1 and V14.2 in a staggered update approach [1][2] Group 1: FSD V14 Rollout Plan - The initial release of FSD V14 will be an "early wide release," primarily targeting Tesla's internal employees, early testers, and select industry influencers, rather than all vehicle owners [2] - Approximately two weeks after the initial version, Tesla plans to launch FSD V14.1, which is expected to address various errors identified in the initial release and will likely be the first version pushed to the general public [4] Group 2: Technical Enhancements - The FSD V14 update is characterized as a fundamental architectural upgrade, comparable to V12, and is considered the second-largest update in FSD's history, featuring a neural network model with a tenfold increase in parameters [4] - The new version will reduce video data compression rates, enhancing the vehicle's visual clarity and precision in judgment, while also decreasing the frequency of driver monitoring [4] Group 3: Additional Features and Goals - Features previously listed as "coming soon" in the V13.2.9 update, such as audio input recognition for emergency vehicles and further optimization of "phantom braking," are expected to be implemented in the V14 series [5] - However, the highly requested "destination auto-parking" feature was not mentioned by Musk, indicating it may not be included in the initial V14 release [5] - The core objective of the new architecture in V14 is to provide vehicles with driving intuition closer to that of humans, enabling them to handle more complex road scenarios [5]