Core Viewpoint - The article discusses the launch of RynnBrain, the first embodied brain model with spatiotemporal memory, developed by Alibaba's Damo Academy, which significantly enhances the capabilities of embodied robots in understanding and interacting with the physical world [7][9][76]. Group 1: RynnBrain Model Features - RynnBrain consists of seven models ranging from 2B to 30B parameters, designed to understand both "time" and "space," allowing it to remember past trajectories and predict future actions [7][9]. - It outperforms leading models like Nvidia's Cosmos-reason2 and Google's Gemini Robotics ER 1.5 across 20 benchmarks, achieving 16 state-of-the-art (SOTA) results [7]. - RynnBrain-30B-A3B, the first MoE architecture in embodied models, demonstrates exceptional efficiency, requiring only 3B active parameters while surpassing the performance of a 72B model [10][11]. Group 2: Training and Data Utilization - The model was trained using over 20 million pairs of high-quality data, incorporating various multimodal training datasets to enhance its understanding of physical space [19][20]. - A unique aspect of the training involved generating 1 million pairs of "self-centered" OCR question-answer data, enabling the robot to interpret labels and numbers in its environment [21][23]. Group 3: Functional Capabilities - RynnBrain exhibits strong flexibility in input and output, capable of processing images and videos of varying resolutions and providing multiple modalities of output, such as trajectories and poses [26][28]. - It possesses spatiotemporal memory, allowing it to maintain awareness of object locations and trajectories even after interruptions, which is crucial for long-term tasks [34][40]. Group 4: System Architecture and Scalability - The model employs a "big brain-small brain" layered architecture, where RynnBrain handles long-term planning and scene understanding, while a smaller execution layer focuses on motor control [54][56]. - This architecture facilitates modular iteration and enhances the model's adaptability to various tasks, such as complex navigation and planning [57][58]. Group 5: Open Source and Industry Impact - Damo Academy has open-sourced RynnBrain along with comprehensive training codes and a new evaluation benchmark, RynnBrain-Bench, which assesses the model's understanding of video sequences and spatial positioning [60][62]. - This initiative aims to lower barriers in the industry by providing a shared infrastructure for understanding physical concepts, improving system efficiency, and fostering healthy competition among teams [66][69].

Core Viewpoint - Tencent Hunyuan has launched a new ultra-small model, HY-1.8B-2Bit, designed for consumer-grade hardware, which is significantly smaller than many common mobile applications, making it suitable for edge deployment [2][13]. Group 1: Model Specifications and Performance - The HY-1.8B-2Bit model has a parameter count of only 0.3 billion and a memory footprint of just 600MB, making it ideal for deployment on edge devices [1][13]. - The model utilizes a unique 2-bit quantization scheme, which reduces the parameter count by six times compared to the original model while maintaining its full cognitive capabilities [2][6]. - Compared to the original precision model, the generation speed of HY-1.8B-2Bit is enhanced by 2-3 times on real edge devices, significantly improving user experience [2][6][13]. Group 2: Quantization Techniques - The model employs Quantization Aware Training (QAT) to mitigate the precision loss typically associated with 2-bit quantization, allowing it to approach the performance of full-precision models [6][11]. - The "Elastic Stretch Quantization" (SEQ) strategy is introduced to address the challenges of low precision, enhancing the model's ability to capture high-dimensional feature distributions [9][11]. - Data optimization strategies have been implemented, increasing the proportion of scientific data and incorporating long-text data to improve the model's overall capabilities [8][7]. Group 3: Training and Deployment - The training process for HY-1.8B-2Bit was optimized to require only 10% of the tokens needed for training the Bitnet-2B model, demonstrating efficiency in achieving low-bit model performance [12][11]. - The model is compatible with Arm computing platforms and has been tested on devices like the MacBook M4 and Dimensity 9500, showing significant acceleration in both latency and generation speed compared to original models [13][14]. - Future developments will focus on reinforcement learning and model distillation to further enhance the capabilities of low-bit quantized models, aiming to bridge the performance gap with full-precision models [15].

AI热潮还在汹涌，但如果你还不知道如何参与……那为什么不来 量子位 呢？ 我们是一家以 追踪AI新进展 为核心的内容平台，经过8年积累，目前拥有顶流影响力，广泛且备受认可的产业资源，以及时代风口的最佳观 测和学习生态位。 目前，我们有 三大方向 岗位招聘，希望你是 （或者能成为） 这三个方向的内容专家： 岗位均为全职，工作地点：北京中关村。 编辑部 发自 凹非寺 量子位 | 公众号 QbitAI 岗位面向： 加入我们，你可以获得： 以下是岗位详情： 所有岗位不同能力层级职位均在开放，欢迎结合个人履历和经验申请。 AI产业方向 岗位职责： AI产业方向 ：关注基建层创新，包含芯片、AI Infra、云计算； AI财经方向 ：关注AI领域创投和财报，跟踪产业链资本动向； AI产品方向 ：关注AI在应用和硬件终端方向的进展。 社招：覆盖编辑、主笔、主编各个层级，按能力匹配岗位； 校招：应届毕业生，接受实习且可转正。 站在AI浪潮之巅 ：第一时间接触和了解AI领域最新技术和产品，构建完整的AI认知体系。 玩转AI新工具 ：将各种AI新技术、新工具应用于工作，提升工作效率和创造力。 打造个人影响力 ：通过撰写独家原创内 ...

Core Viewpoint - Typeless, an AI voice keyboard, has gained significant popularity with over 100,000 users subscribing despite its high monthly fee of over 200 yuan, primarily due to its unique voice input capabilities [2][3]. Group 1: Product Features - Typeless offers a distinct functionality by converting voice directly into structured text, eliminating filler words and redundancies [11][20]. - The AI can automatically organize spoken content into coherent sections, making it easier for users to manage information [22]. - It supports over 100 languages and operates seamlessly on both computer and mobile platforms, with no typing required [12]. Group 2: Performance Evaluation - The transcription accuracy of Typeless is notably high, effectively filtering out filler words and maintaining the core message of the spoken input [14][18]. - In tests involving complex phrases and homophones, the AI demonstrated a strong ability to capture the intended meaning, although occasional issues with rapid speech were noted [17][18]. - The structured output from the AI is comparable to markdown formats commonly used in professional settings, enhancing usability for tasks like note-taking [24]. Group 3: AI Enhancement Capabilities - Typeless includes an AI refinement feature that improves the clarity and flow of the transcribed text, although its utility may vary based on user needs [25][26]. - The AI's ability to refine text is limited to the current input and does not support historical edits, which may restrict its application in certain scenarios [28][29]. - Despite some limitations, the overall user experience with the AI refinement feature is positive, contributing to a more polished final output [30][32]. Group 4: Market Position and User Demographics - The product is positioned as a versatile tool for individuals who require high transcription accuracy, such as those who drive or have limited typing capabilities [34]. - Users have identified various innovative applications for Typeless, including social media content creation and enhancing interactions with AI models like ChatGPT [36][38]. - The subscription price is considered high, suggesting that the product may be more suitable for users with specific needs, such as content creators and professionals who frequently record notes [42].

Core Viewpoint - The article discusses the phenomenon of "AI fatigue," where increased productivity through AI tools leads to greater stress and exhaustion among developers, rather than the anticipated efficiency gains [1][42]. Group 1: AI's Impact on Productivity - AI has the potential to significantly enhance productivity, allowing tasks that previously took a day to be completed in an hour [9]. - However, this efficiency often results in an increased workload, as developers are expected to handle multiple tasks simultaneously, leading to fragmented attention and higher energy consumption [10][9]. - The shift from a creator role to a quality control role means developers spend more time evaluating and correcting AI-generated outputs, which is more mentally taxing than traditional coding [12][14]. Group 2: Psychological and Emotional Effects - The unpredictability of AI outputs creates anxiety, as developers cannot rely on consistent results, leading to a constant state of alertness [18][20]. - The rapid evolution of AI tools requires continuous learning, which can lead to feelings of inadequacy and pressure to keep up with peers, exacerbating stress levels [23][39]. - Over-reliance on AI can result in cognitive decline, as critical thinking skills may diminish when individuals do not engage in independent problem-solving [33]. Group 3: Strategies for Managing AI Fatigue - The author suggests implementing time limits for AI tasks, distinguishing between thinking and execution time, and accepting that AI outputs do not need to be perfect [43][45]. - Developers are encouraged to focus on foundational concepts rather than chasing every new tool, and to document the efficiency of AI usage to determine when to rely on it [43][45]. - Emphasizing the importance of mental breaks and allowing for downtime can help maintain overall well-being and productivity in the AI-driven work environment [47].

HVO-GRPO团队 投稿 量子位 | 公众号 QbitAI 文本摘要作为自然语言处理 （NLP） 的核心任务，其质量评估通常需要兼顾 一 致性 （Consistency） 、连贯性 （Coherence） 、流畅 性 （Fluency） 和相关性 （Relevance） 等多个维度。 然而，在实际优化过程中，开发者往往面临"拆东墙补西墙"的窘境：提升了相关性，一致性可能随之下降。如何让模型在多个目标之间达成完 美的"帕累托最优" （Pareto optimal） ？ 近日，Li Auto团队一项被 ICASSP 2026 接收的研究提出了 HyperVolume Optimization （HVO） 。这是一种全新的多目标强化学习 （MORL） 策略，它基于GRPO框架，无需SFT或冷启动，就能让7B参数的模型在摘要任务上展现出媲美GPT-4的性能，且生成内容更加简 洁。 △ HVO性能对比雷达图 研究背景 核心痛点：多目标优化的"不平衡" 文本摘要生成是自然语言处理 （NLP） 中的一项核心且具有挑战性的任务。为了全面评估生成摘要的质量，研究人员通常会考察多个维度， 例如连贯性、一致性、流畅性和相关性。然 ...

克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 2026以来，还有比 OpenClaw （原Clawdbot/Moltbot） 更火的AI和Agent应用吗？ 别的不说，光是GitHub上的星星就有17.7万之多，而且增长速度就像在坐火箭。 能有这样的热度，是因为它早就脱离了聊天机器人的范畴，是公认能真正指挥电脑干活的"数字员工"。 比如这位网友就用三台电脑、15个Agent部署了自己的"数字军团"，他只需要在一个Discord频道里坐镇指挥，就能让它们完成处理邮件、读 PPT、写代码、发推文，甚至撰写每日汇报等一系列工作。 OpenClaw，一行代码真能搞定？ 说实话，为了搭上OpenClaw的这班车，我属实是没少折腾。 不过，想把这位"大神"请进自己的设备，还是得费点周折。 所以很多人都是看着视频里它干活挺爽，但只能眼巴巴地看，真到了自己想用的时候，却发现连大门都进不去。 现在，这个"看得到吃不到"的尴尬局面，终于有解了。 我发现百度智能云搞了个 "作弊级"的方案 ，直接把这个复杂的部署过程给"降维打击"了。 它不需要你懂代码，也不用你折腾环境，直接把这个顶流AI变成了人人都能上手的生产力工具。 ...

Core Insights - The article discusses the significant advancements in reasoning capabilities of large language models (LLMs) through reinforcement learning fine-tuning, particularly highlighting the high costs associated with inefficient training processes [1][2]. Group 1: Training Efficiency - Traditional training methods like "Uniform Sampling" waste computational resources by randomly selecting questions that do not provide effective learning signals [2]. - The "Dynamic Sampling" approach, while more efficient, still incurs high costs due to the need for extensive self-evaluation by the model [2][6]. - The proposed MoPPS framework aims to dynamically predict question difficulty without the expensive self-evaluation process, thus enhancing training efficiency [3][6]. Group 2: MoPPS Framework - MoPPS utilizes a lightweight Bayesian model to quickly estimate question difficulty, allowing for efficient selection of training data [8][10]. - The framework models each question as a "bandit" problem, using a Beta distribution to estimate success rates based on training feedback [9][10]. - MoPPS introduces a recursive update mechanism that improves difficulty estimation over time, adapting to the model's evolving capabilities [11][13]. Group 3: Performance Improvements - MoPPS has demonstrated a training speed increase of 1.6x to 1.8x while reducing inference costs by up to 78.46% compared to traditional methods [18][21]. - The framework has shown significant advantages across various reasoning tasks, achieving better performance with fewer computational resources [18][21]. - The correlation between predicted and actual question difficulty is high, validating the effectiveness of MoPPS in accurately estimating task challenges [25][29]. Group 4: Versatility and Future Applications - MoPPS is compatible with multiple reinforcement learning algorithms and can adapt to different sampling strategies, enhancing its applicability [26][28]. - The framework's ability to incorporate prior knowledge can further accelerate initial training phases, making it a versatile tool for large-scale model fine-tuning [28][31]. - The research indicates potential for broader applications in the reinforcement learning fine-tuning of larger models in the future [31].

ProjDevBench团队 投稿 量子位 | 公众号 QbitAI AI编程是一项非常有实用价值的能力，但网络上不时也能看到程序员抱怨AI"听不懂人话"、"难以找到根本问题"，更有直接建议"每次生成代码 不要超过5行"的经验分享。 而近期又有很多AI工具声称可以从零快速构建完整代码项目。 所以AI编程智能体真的能从零构建完整软件项目吗？近日一多校联合研究团队针对这一问题进行了探索。 上海交通大学、上海创智学院、加州大学默塞德分校、 北京理工大学（按论文作者顺序） 联合发布 ProjDevBench ——首个通过OJ细粒度 反馈评估AI编程智能体端到端项目开发能力的基准测试，要求智能体仅凭自然语言需求文档，从零开始构建完整、可运行的软件仓库。 当任务从"补全现有代码"变为"从零构建"时，性能出现断崖式下跌。 结果令人深思： 所有智能体总体提交AC率仅27.38% 。 该研究得出的结论摘要： 为什么需要端到端项目开发基准 现有基准测试如HumanEval、MBPP聚焦于函数级代码生成，SWE-bench关注issue修复，但真实软件工程需要的远不止这些。当开发者使 用Cursor或GitHub Copilot进 ...

梦晨 发自 凹非寺 量子位 | 公众号 QbitAI 北大 谢俊逸、袁新意 合作论文，被数学四大顶刊接收！ 还是四大顶刊中年发文量最少的 《Acta Mathematica》 。 | For Issue | Seq. | Title | Author(s) | Date of | | --- | --- | --- | --- | --- | | | | | | Acceptance | | : | | On Kähler Ricci shrinker surfaces | Yu Li, Bing Wang | 21 January 2025 | | -- | | On Stevenhagen's conjecture | Peter Koymans, Carlo Pagano | 23 January 2025 | | -- | | Ray structures on Teichmüller space | Huiping Pan, Michael Wolf | 8 June 2025 | | -- | | Primes of the form p^2 + nq^2 | Ben Green, Mehtaab ...