特斯拉FSD將迎來重大更新 馬斯克：V15版本在無人監管下安全性遠超人類水平

🔑 增強重點摘要

• •Tesla has transitioned to an end-to-end neural network architecture for FSD V15, moving away from C++ code-based heuristics to a system trained on massive datasets of human driving behavior.
• •Regulatory scrutiny remains a significant hurdle, as V15's 'unsupervised' claims face pending investigations by the NHTSA regarding the system's performance in low-visibility conditions.
• •The rollout of V15 is tied to the deployment of Tesla's new 'AI5' hardware suite, which provides significantly higher compute capacity compared to the previous Hardware 4.0 platform.

🛠️ 技術深入

• •Architecture: Shift to a unified transformer-based model that processes raw video input directly into control outputs (steering, braking, acceleration).
• •Compute: Utilization of the AI5 inference computer, featuring a custom-designed NPU optimized for high-throughput transformer model execution.
• •Training: Leverages the 'Dojo' supercomputing cluster to process petabytes of fleet-collected video data, focusing on 'edge case' scenarios identified by the shadow-mode fleet.
• •Latency: Reduced end-to-end latency by 40% compared to V12, allowing for faster reaction times in high-speed urban environments.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線