来源：陆向谦 作者：赵老师 过去一周，中国科技圈发生了两件震撼全网的事。 PART.01 两件震撼全网的事件 第一件：腾讯官宣，OpenAI科学家、清华姚班校友姚顺雨正式出任CEO/总裁办公室首席AI科学家，直 接向腾讯总裁刘炽平汇报。 同时兼任AI Infra部、大语言模型部负责人。 第二件：小米官宣，95后"AI才女"罗福莉担任MiMo大模型负责人，首次公开亮相。 据传，雷军曾开出千万年薪挖角她。 两个年轻人，一个27岁，一个95后（不到30岁），站上了中国AI战场的最中央。 家长圈瞬间炸了。 有人震惊："27岁就能直接向总裁汇报？" 也有人焦虑："95后都拿千万年薪了，我家孩子怎么办？" 但定居硅谷多年、在清华、香港科技大学都长期任教过，深耕创新人才培养的教育专家陆向谦教授，听 到这个消息，说了一段让所有人震惊的话： "95后站上AI舞台中央，不算什么。奥特曼28岁创立OpenAI，乔布斯、比尔·盖茨、扎克伯格、马斯克创 业时都才20岁左右，大学都没读完。" "AI人才平均年薪，在硅谷至少100万美金，在中国是100万人民币。这是平均工资，骨干再乘10倍也正 常。" "新事物往往在边缘，由年轻人创造。孩 ...

ReCode 作者团队来自于 Foundation Agents 开源社区。第一作者为 DeepWisdom 研究员于兆洋，同时也是 OpenManus 的发起人之一。共同通讯作者 为 DeepWisdom 创始人兼 CEO 吴承霖，以及蒙特利尔大学与 MILA 实验室的副教授刘邦。 想象你在准备早餐：你不会先写一份详细到「左手抓鸡蛋、右手拿碗、手腕旋转 45 度敲击蛋壳」这样的清单，也不会只有一个笼统的计划叫「做个早 餐」，然后不知所措。 人类大脑会自然地在煎培根、鸡蛋和敲开鸡蛋这样不同 决策粒度 间无缝切换，粗粒度的高层规划和细粒度的具体动作融为一体，但目前的 AI 智能体很难 做到这一点。 最近，来自 DeepWisdom 的研究员在论文中指出，当前主流智能体框架都被固定的决策粒度束缚住了。ReAct 智能体只会一步步执行细粒度动作，缺乏全 局规划；而带规划器（Planner）的智能体虽然能制定高层计划，但规划和执行被硬生生分成两个模块，难以动态调整和优化。 论文标题 ：ReCode: Unify Plan and Action for Universal Granularity Control 论文地 ...

论文链接：https://arxiv.org/pdf/2510.23822 斯坦福和MIT的研究团队推出了一种新的AI智能体推理框架ReCAP，在长上下文任务中全面超越了现有的主流框架ReAct，性能提升显著。 ReCAP通过独特的递归树结构和三大机制，解决了大语言模型在复杂任务中常见的目标漂移、上下文断层和成本爆炸等问题。在多项基准测 试中，ReCAP均取得了大幅领先的成绩，展现出强大的通用性和稳定性。尽管计算成本略有增加，但其在关键任务中的表现使其成为极具潜 力的新一代通用推理架构。 自2022年ReAct框架提出以来，AI智能体推理领域便进入了百家争鸣的时代，各种复杂架构如雨后春笋般涌现。 然而，这些架构大多昙花一现，因其复杂的结构导致在更换评测基准时需要大幅修改示例，表现远不如ReAct稳定泛用，这也使得ReAct在过去三年中， 成为了该领域事实上的主流与标杆。 但是，我们真的不能再做得更好了吗？ 面对大模型在长上下文任务中走几步就忘的短期记忆顽疾，业界是否只能止步于此？ 来自斯坦福大学与MIT的研究团队给出了肯定答案，正式发布的AI Agent推理新框架——ReCAP（递归上下文感知推理与规划），从 ...

Core Insights - The article discusses the evolution of AI, particularly focusing on the transition to the "second half" of AI development, emphasizing the importance of language and reasoning in creating more generalizable AI systems [4][62]. Group 1: AI Evolution and Language - The concept of AI has evolved from rule-based systems to deep reinforcement learning, and now to language models that can reason and generalize across tasks [41][43]. - Language is highlighted as a fundamental tool for generalization, allowing AI to tackle a variety of tasks by leveraging reasoning capabilities [77][79]. Group 2: Agent Systems - The definition of an "Agent" has expanded to include systems that can interact with their environment and make decisions based on reasoning, rather than just following predefined rules [33][36]. - The development of language agents represents a significant shift, as they can perform tasks in more complex environments, such as coding and internet navigation, which were previously challenging for AI [43][54]. Group 3: Task Design and Reward Mechanisms - The article emphasizes the importance of defining effective tasks and environments for AI training, suggesting that the current bottleneck lies in task design rather than model training [62][64]. - A focus on intrinsic rewards, which are based on outcomes rather than processes, is proposed as a key factor for successful reinforcement learning applications [88][66]. Group 4: Future Directions - The future of AI development is seen as a combination of enhancing agent capabilities through better memory systems and intrinsic rewards, as well as exploring multi-agent systems [88][89]. - The potential for AI to generalize across various tasks is highlighted, with coding and mathematical tasks serving as prime examples of areas where AI can excel [80][82].

Core Viewpoint - Recent rumors about former OpenAI researcher Yao Shunyu joining Tencent with a salary exceeding 100 million have been officially denied by Tencent, labeling the reports as false [1][2][3]. Group 1: Yao Shunyu's Background - Yao Shunyu graduated from Tsinghua University and holds a PhD in Computer Science from Princeton University, joining OpenAI in 2024 [4]. - He was recognized as one of the "Innovators Under 35" in the China region by MIT Technology Review, being the youngest at 27 years old [4]. - Yao is a key contributor to OpenAI's early research, particularly in the development of language agents [4]. Group 2: Contributions to AI - Yao Shunyu proposed the ReAct method, which integrates reasoning and action in the development of language agents, establishing a foundational approach for creating versatile and scalable AI systems [5]. - The core idea of ReAct is to enable large language models to perform internal reasoning before making decisions, enhancing model controllability and expanding applicability across various fields [5]. - ReAct has become a mainstream method for constructing language agents, widely adopted in both academia and industry [5].

Group 1 - The news about former OpenAI researcher Yao Shunyu joining Tencent for a salary exceeding 100 million is false, as clarified by Tencent's official account [1] - Yao Shunyu graduated from Tsinghua University and later obtained a PhD from Princeton University, where he developed the Tree of Thoughts framework and CoALA modular cognitive architecture [1] - He joined OpenAI in 2024 and contributed significantly to the development of intelligent agent products, being recognized as a core contributor [1][5] Group 2 - Yao Shunyu's ReAct method introduced a paradigm combining reasoning and action for intelligent agents, enhancing the controllability and applicability of large language models [5] - The AI talent competition is intensifying globally, with companies like Meta offering over $200 million in total compensation to attract top talent, including researchers from OpenAI [6] - In China, major internet companies are expanding their recruitment for AI-related positions, with a reported increase of over 10 times in new AI job postings compared to the previous year [6]

Group 1 - Recent rumors suggest that former OpenAI researcher Yao Shunyu has joined Tencent with a salary exceeding 100 million [1] - Tencent officially refuted the claim regarding Yao Shunyu's salary through its public account, labeling the report as a rumor [1] - Yao Shunyu, a graduate of Tsinghua University's Yao Class and a PhD in Computer Science from Princeton University, joined OpenAI in 2024 [1] Group 2 - Yao Shunyu introduced the ReAct method, which combines reasoning and action in the agent paradigm, laying the foundation for creating generalizable and scalable language agents [2] - The core idea of ReAct is to enable large language models to perform explainable internal reasoning before making decisions and actions, enhancing model controllability and expanding applicability across various fields [2] - ReAct has become the most mainstream method for constructing language agents globally, widely adopted in both academia and industry [2]

Core Viewpoint - The article discusses the career transition of Shunyu Yao, a prominent researcher from OpenAI, as he embarks on a new phase in the AI field, focusing on personal AI and the evolving landscape of AI development, which is now entering its "second half" [2][47]. Group 1: Background and Achievements - Shunyu Yao, a 29-year-old researcher, has an impressive academic background, including graduating from Tsinghua University and obtaining a PhD from Princeton, where he focused on natural language processing and reinforcement learning [4][22]. - His notable contributions to AI include the development of frameworks like Tree of Thoughts, SWE-bench, and ReAct, which enhance the reasoning and decision-making capabilities of language models [6][36]. Group 2: Career Transition - Yao's departure from OpenAI has been confirmed through various channels, and he is rumored to be considering entrepreneurship or joining another tech giant [3][51]. - His recent work emphasizes the shift in AI development from model-centric approaches to defining meaningful tasks and evaluating AI systems' performance in real-world scenarios [47][48]. Group 3: Philosophical Insights - Yao's approach to research is characterized by a cross-disciplinary perspective, drawing inspiration from various fields, which he believes is essential for innovation in AI [9][20]. - He advocates for the importance of language as a medium for reasoning and decision-making in AI, highlighting its role in enabling agents to generalize across different contexts [28][30].