编辑 | 理想TOP2 转自 | 李想: 特斯拉V14也用了VLA相同技术|25年10月18日B站图文版压缩版 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近30个 方向 学习 路线 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 本文只做学术分享，如有侵权，联系删文 压缩版： 视频共计21min24s，花了10min51s介绍对OpenAI定义的5阶段的理解，做了很多类比。认为OpenAI在AI应用/模型/规范定义都做得非常好。 聊天机器人 （Chatbots）：背后是基座模型，功能是压缩人类已知的数字知识。好比人上学到大学毕业，打下知识基础。 推理者 （Reasoners）：具备思维链，能进行连续性思考和任务，主要依赖SFT和RLHF训练。好比人读研或有师傅带，获得经验。 智能体 （Agents）：AI真正开始工作，能使用工具完成长任务。这对AI的专业性和可靠性要求极高（需达到八九十分才合格），好比人胜任一个专业岗位。 创新者 （Innovators）：为解决智能体专业性难题，通过出题和解题来进行强化训练。这需要世界模型和RLAIF（AI反馈强化学习）来模拟真实环境的训练 ...

Core Viewpoint - The article discusses the five stages of artificial intelligence (AI) as defined by OpenAI, emphasizing the importance of each stage in the development and application of AI technologies [10][11]. Group 1: Stages of AI - The first stage is Chatbots, which serve as a foundational model that compresses human knowledge, akin to a person completing their education [2][14]. - The second stage is Reasoners, which utilize supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) to perform continuous reasoning tasks, similar to advanced academic training [3][16]. - The third stage is Agents, where AI begins to perform tasks autonomously, requiring a high level of reliability and professionalism, comparable to a person in a specialized job [4][17]. - The fourth stage is Innovators, focusing on generating and solving problems through reinforcement training, necessitating a world model for effective training [5][19]. - The fifth stage is Organizations, which manage multiple agents and innovations to prevent chaos, similar to corporate management [4][21]. Group 2: Computational Needs - The demand for reasoning computational power is expected to increase by 100 times, while training computational needs may expand by 10 times over the next five years [7][23]. - The article highlights the necessity for both edge and cloud computing to support the various stages of AI development, particularly in the Agent and Innovator phases [6][22]. Group 3: Ideal Self-Developed Technologies - The company is developing its own reasoning models (MindVLA/MindGPT), agents (Driver Agent/Ideal Classmate Agent), and world models to enhance its AI capabilities [8][24]. - By 2026, the company plans to equip its autonomous driving technology with self-developed advanced edge chips for deeper integration with AI [9][26]. Group 4: Training and Skill Development - The article emphasizes the importance of training in three key areas: information processing ability, problem formulation and solving ability, and resource allocation ability [33][36]. - It suggests that effective training requires real-world experience and feedback, akin to the 10,000-hour rule for mastering a profession [29][30].

Core Viewpoint - The article discusses the importance of reinforcement learning (RL) in enhancing the intelligence of large models, emphasizing the need for effective interaction between models and their environments to obtain high-quality feedback [1][2]. Summary by Sections Section 1: Importance of Reinforcement Learning - The article highlights that RL is crucial for the advancement of large model intelligence, with a focus on how to enable models to interact with broader environments to achieve capability generalization [1][8]. - It mentions various RL techniques such as RLHF (Reinforcement Learning from Human Feedback), RLAIF (AI Feedback Reinforcement Learning), and RLVR (Verifiable Reward Reinforcement Learning) as key areas of exploration [1][8]. Section 2: RuscaRL Framework - The RuscaRL framework is introduced as a solution to the exploration bottleneck in RL, utilizing educational psychology's scaffolding theory to enhance the reasoning capabilities of large language models (LLMs) [12][13]. - The framework employs explicit scaffolding and verifiable rewards to guide model training and improve response quality [13][15]. Section 3: Mechanisms of RuscaRL - **Explicit Scaffolding**: This mechanism provides structured guidance through rubrics, helping models generate diverse and high-quality responses while gradually reducing external support as the model's capabilities improve [14]. - **Verifiable Rewards**: RuscaRL designs rewards based on rubrics, allowing for stable and reliable feedback during training, which enhances exploration diversity and ensures knowledge consistency across tasks [15][16]. Section 4: Future Implications - The article suggests that both MindGPT and MindVLA, which target digital and physical worlds respectively, could benefit from the advancements made through RuscaRL, indicating a promising future for self-evolving models [9][10]. - It emphasizes that the current challenges in RL are not just algorithmic but also involve systemic integration of algorithms and infrastructure, highlighting the need for innovative approaches in building capabilities [9].

Core Viewpoint - The article discusses the advanced technologies used in Li Auto's autonomous driving solutions, specifically focusing on the "MoE + Sparse Attention" efficient structure that enhances the performance and efficiency of large models in 3D spatial understanding and reasoning [3][6]. Group 1: Introduction to Technologies - The article introduces a series of posts that delve deeper into the advanced technologies involved in Li Auto's VLM and VLA solutions, which were only briefly discussed in previous articles [3]. - The focus is on the "MoE + Sparse Attention" structure, which is crucial for improving the efficiency and performance of large models [3][6]. Group 2: Sparse Attention - Sparse Attention limits the complexity of the attention mechanism by focusing only on key input parts, rather than computing globally, which is particularly beneficial in 3D scenarios [6][10]. - The structure combines local attention and strided attention to create a sparse yet effective attention mechanism, ensuring that each token can quickly propagate information while maintaining local modeling capabilities [10][11]. Group 3: MoE (Mixture of Experts) - MoE architecture divides computations into multiple expert sub-networks, allowing only a subset of experts to be activated for each input, thus enhancing computational efficiency without significantly increasing inference costs [22][24]. - The article outlines the core components of MoE, including the Gate module for selecting experts, the Experts module as independent networks, and the Dispatcher for optimizing computation [24][25]. Group 4: Implementation and Communication - The article provides insights into the implementation of MoE using DeepSpeed, highlighting its flexibility and efficiency in handling large models [27][29]. - It discusses the communication mechanisms required for efficient data distribution across multiple GPUs, emphasizing the importance of the all-to-all communication strategy in distributed training [34][37].