Workflow
中国的“马斯克” 李想分享VLA训练细节
Jing Ji Guan Cha Bao·2025-05-08 04:08

Core Insights - The rapid development of AI technology is acknowledged by Li Xiang, founder of Li Auto, who emphasizes the need for a calm approach to its practical application [2] - Li Auto is focusing on the integration of AI tools into its production processes, moving from simple information and assistance tools to core value creation tools [2][4] - The company has made significant advancements in its VLA (Vision-Language-Action) strategy, which aims to develop a humanoid driving model capable of independent decision-making and action [4][5] AI Development and Strategy - Li Auto has experienced a significant boost from the domestic large model DeepSeek, which has accelerated the completion of its foundational model infrastructure by nine months [2][3] - The VLA strategy is divided into three phases: rule-based algorithms, end-to-end models, and the final VLA stage, which represents the beginning of humanoid intelligence [4][5] - The company is training a 32 billion parameter VL model, incorporating extensive 3D visual data and language inputs related to driving behavior to enhance semantic understanding [5][6] Industry Perspective - Li Xiang compares the current state of AI in the automotive industry to a "Linux moment," with the potential for a significant breakthrough akin to an "Android moment" when AI is effectively deployed in specific fields [7] - The transition from traditional rule-driven approaches to intelligent agent-driven paradigms is seen as a major shift in the industry [7] - Li Auto aims to redefine the relationship between humans and vehicles, positioning AI not just as an efficiency tool but as a systemic and scalable production tool [7]