以下文章来源于语言即世界language is world ，作者张小珺 语言即世界language is world . 这是原「张小珺」公众号，是我新发起的内容工作室。和我们一起，从这里，探索新世界。 2025 年 4 月， OpenAI 研究员姚顺雨发布了一篇有名的博文《 The Second Half 》，宣告 AI 主线程的游戏已进入下半场。这之 后，我们与他进行了一场播客对谈。 姚顺雨毕业于清华和普林斯顿大学，博士期间意识到语言是人类发明的最重要的工具，也是最有可能构建通用系统的，于是转向 Language Agent 研究，至今已 6 年。 这场对谈有两位主持人，分别是我和李广密。姚顺雨表达了许多此前从未分享过的观点。比如： 我们的谈话从个体出发，共同探索由人、组织、AI 、人与机器的交互，所抵达的这个世界智能的边界以及人类与机器的全景。 此前， 我们关于 Manus 肖宏、 Youware 明超平、 Lovart 陈冕的访谈，记录了华人 Agent 创业者在应用上的探索。而姚顺雨的访 谈，描绘的则是另一面：他在硅谷最前沿的 AI 实验室做 Agent 研究，他如何看待这波浪潮、模型与应用的边界 ...

Core Viewpoint - The article discusses a significant research paper that explores the effectiveness of reinforcement learning (RL) compared to supervised fine-tuning (SFT) in training AI models, particularly focusing on the concept of generalization and transferability of knowledge across different tasks [1][5][14]. Group 1: Training Methods - There are two primary methods for training AI models: imitation (SFT) and exploration (RL) [2][3]. - Imitation learning involves training models to replicate data, while exploration allows models to discover solutions independently, assuming they have a non-random chance of solving problems [3][6]. Group 2: Generalization and Transferability - The core of the research is the concept of generalization, where SFT may hinder the ability to adapt known knowledge to unknown domains, while RL promotes better transferability [5][7]. - A Transferability Index (TI) was introduced to measure the ability to transfer skills across tasks, revealing that RL-trained models showed positive transfer in various reasoning tasks, while SFT models often exhibited negative transfer in non-reasoning tasks [7][8]. Group 3: Experimental Findings - The study conducted rigorous experiments comparing RL and SFT models, finding that RL models improved performance in unrelated fields, while SFT models declined in non-mathematical areas despite performing well in mathematical tasks [10][14]. - The results indicated that RL models maintained a more stable internal knowledge structure, allowing them to adapt better to new domains without losing foundational knowledge [10][14]. Group 4: Implications for AI Development - The findings suggest that while imitation learning has been a preferred method, reinforcement learning offers a promising approach for developing intelligent systems capable of generalizing knowledge across various fields [14][15]. - The research emphasizes that true intelligence in AI involves the ability to apply learned concepts to new situations, akin to human learning processes [14][15].

Core Viewpoint - Mech-Mind Robotics, founded by CEO Shao Tianlan, has focused on developing standardized robotic products that can adapt to various hardware forms, aiming to cover a wide range of industries. The company has achieved significant market penetration, becoming the largest unicorn in the "AI + robotics" sector globally, with a leading market share for four consecutive years [2][3]. Group 1: Company Development and Market Position - Mech-Mind Robotics has been likened to "puzzle-solving" over its eight years of operation, emphasizing the high barriers and challenges in the robotics industry [2]. - The company has successfully implemented its products across multiple sectors, including automotive, logistics, and heavy industry, achieving a leading market share [2]. - The founder, Shao Tianlan, noted that the current robotics industry resembles the state of the autonomous driving sector in 2015, with both opportunities and challenges in scaling technology [3][12]. Group 2: Industry Trends and Comparisons - The robotics industry has seen a shift towards a focus on intelligence, with computer scientists increasingly influencing the field, contrasting with the earlier emphasis on hardware and control [7][8]. - The current landscape is marked by heightened interest and investment in robotics, leading to both opportunities for startups and challenges due to increased competition and unrealistic expectations [11][12]. - Shao Tianlan draws parallels between the current state of robotics and the early days of autonomous driving, highlighting the potential for significant technological advancements alongside the risk of overpromising timelines [12][43]. Group 3: Product Applications and Future Outlook - Mech-Mind Robotics specializes in high-precision industrial 3D cameras and AI software, which have been widely adopted in logistics and manufacturing scenarios [5][20]. - The company aims to enhance robotic intelligence to enable self-perception, planning, and decision-making capabilities, similar to advancements seen in autonomous vehicles [5][6]. - The founder believes that while the timeline for widespread adoption of robots in households may be longer, significant advancements in industrial applications are expected within the next decade [17][48]. Group 4: Global Market Strategy - Mech-Mind Robotics began exploring international markets in 2019, with overseas business now accounting for half of its revenue, driven by the need to meet high standards set by developed countries [28][29]. - The company emphasizes the importance of high standards and quality in its products to compete effectively in the global market, particularly against established players in industrial automation [33][34]. - The founder notes that the robotics market is still in its early stages, with significant room for growth as automation continues to evolve in manufacturing and logistics [36][37].