加入我们，你可以获得： 编辑部 发自 凹非寺 量子位 | 公众号 QbitAI 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。 岗位面向： 参与核心采访，对话产业专家、技术大牛、撰写AI云落地案例。 任职要求： 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 ...

Core Insights - The release of GLM-5 marks a significant advancement in open-source AI, bringing it into the era of long-task capabilities [2][25] - GLM-5 demonstrates exceptional programming abilities, successfully creating a Game Boy Advance emulator from scratch, showcasing its stability and reliability in complex tasks [3][9][12] - The model has achieved competitive performance, ranking alongside Claude Opus 4.5 in various assessments, indicating its strong programming capabilities and operational stability [15][17] Group 1: Performance and Capabilities - GLM-5 executed over 700 tool calls and 800 context switches while maintaining consistent syntax and accuracy [12] - It has been recognized for its ability to generate complex applications, such as a 3D Monopoly game and an interactive version of Minecraft, demonstrating its versatility [26][35] - The model's performance in the Vending Bench 2 test has positioned it as the leading open-source model in terms of operational capabilities [23] Group 2: Industry Impact - The emergence of GLM-5 signifies a transformative shift in the SaaS industry, as it allows developers to create sophisticated applications without relying on traditional software subscriptions [38][40] - The release has caused market reactions, with significant declines in SaaS-related stocks, reflecting investor concerns about the implications of AI on software sales [39] - GLM-5's capabilities challenge the previous dominance of closed-source models, empowering developers with tools that were once exclusive to major corporations [40] Group 3: Community and Developer Engagement - The open-source nature of GLM-5 has generated significant interest and demand among developers, with many eager to utilize its capabilities [41] - The model's development has become a focal point for the community, with its headquarters attracting attention as a notable location [42] - The ongoing advancements in AI programming, initiated with earlier versions, have positioned GLM-5 as a leading choice for coding tasks in both domestic and international markets [41]

Core Insights - The article introduces MeepleLM, a virtual playtester that simulates diverse player experiences and provides constructive feedback based on dynamic gameplay [1][4] - MeepleLM significantly outperforms general models like GPT-5.1 and Gemini3-Pro in accurately reflecting player reviews and ratings [2] Group 1: MeepleLM Overview - MeepleLM is developed by a collaborative research team from Shanda Tokyo Research Institute, Shanghai Chuangzhi Academy, Nankai University, and Shanghai AI Lab [1] - The model utilizes a dataset of 1,727 structured board game rulebooks and 150,000 real player reviews to create a mapping from objective rules to subjective experiences [1][9] - The MDA (Mechanics-Dynamics-Aesthetics) framework is employed to enhance the model's understanding of gameplay interactions and emotional experiences [12] Group 2: Challenges in Board Game Design - The board game industry is experiencing rapid growth, yet the design process faces significant challenges due to its reliance on social interactions and emergent gameplay [3] - Traditional playtesting methods are time-consuming and often fail to capture the preferences of diverse player types [3] Group 3: Data and Methodology - A high-quality dataset was constructed through a layered sampling strategy, converting unstructured PDF rulebooks into structured documents [9] - The team filtered through 1.8 million reviews to extract approximately 8% of high-quality data that deeply connects game mechanics with dynamic experiences [9] Group 4: Player Personas - Five distinct player personas were identified through clustering analysis, each representing different preferences and reactions to game mechanics [13][14][15][16][17] - MeepleLM can role-play these personas to provide varied feedback based on specific player preferences [18] Group 5: Performance Evaluation - Extensive testing on 207 games demonstrated MeepleLM's superior performance in community alignment, review quality, and utility compared to general models [21][22] - MeepleLM effectively captures the polarized nature of player reviews, identifying both strengths and critical flaws in games [22] Group 6: Practical Applications - MeepleLM's reviews are characterized by factual accuracy and diverse viewpoints, making it a valuable tool for players and designers alike [25][27] - Over 70% of users prefer MeepleLM for purchase decisions, citing its effectiveness in identifying potential design flaws [27] Group 7: Future Implications - MeepleLM establishes a new paradigm for automated virtual testing in interactive systems, paving the way for empathetic human-machine collaboration [28]

Core Viewpoint - The article highlights the significant advancements of iFLYTEK's Starfire Model X2, showcasing a 50% improvement in reasoning performance compared to its predecessor, Starfire X1.5, achieved within just three months, and emphasizes its capabilities based on domestic computing power [2][10][50]. Group 1: Model Performance and Capabilities - Starfire X2 demonstrates outstanding general capabilities, ranking among the top in the industry, and competes closely with international models like GPT-5.2 and Gemini-3-Pro [2][12]. - The model excels particularly in mathematical calculations and logical reasoning, maintaining proficiency in over 130 languages [3][12]. - Starfire X2's reasoning training efficiency has improved by 50%, a notable achievement given the diminishing returns in performance enhancement for large models [8][9]. Group 2: Technical Innovations - The model incorporates several technical innovations, including weight quantization, low-precision KVCache, and a two-stage reasoning sampling method, which enhance training efficiency by 10% [24][27][30]. - A new adaptive calibration algorithm ensures that the model maintains logical consistency during training and inference, addressing common issues in large model architectures [24]. - The recursive high-difficulty data synthesis method allows for the creation of high-quality training data, improving the model's reasoning accuracy [25][26]. Group 3: Industry Applications and Market Position - Starfire X2 is positioned to drive significant advancements in various industries, with a focus on practical applications and user experience [31][54]. - In the healthcare sector, Starfire X2 outperforms competitors in key metrics, establishing a new benchmark for medical AI models [34][35]. - The model's deployment in education enhances personalized learning experiences, allowing for detailed feedback and analysis of student performance [37][41]. Group 4: Strategic Direction and Future Outlook - iFLYTEK's strategy emphasizes a combination of a general-purpose model and specialized industry models, aiming for rapid deployment and practical applications [54][55]. - The company has established itself as a leader in the domestic AI landscape, leveraging its unique approach to overcome challenges related to computing power and model performance [50][56]. - The advancements of Starfire X2 signal a shift towards the application of domestic AI models, marking the beginning of a new phase of growth and opportunity in the industry [56].

上财团队 投稿 量子位 | 公众号 QbitAI 在量化金融的底层，Alpha因子本质上是一段可执行的代码逻辑，它们试图将嘈杂的市场数据映射为精准的交易信号。然而，长期以来，自 动化因子挖掘始终被困在"两难"的夹缝中：传统的遗传规划 （Genetic Programming，GP） 虽然擅长在海量空间中进行进化搜索，但其 本质是"盲目的随机变异"。 它们在回测中过度拟合了历史噪声，却在逻辑上极难解释，如同一个充满巧合的黑盒。而新兴的大语言模型 （LLM） 虽然具备强大的语义 理解能力，能像人类一样读懂金融理论，但在处理高精度的量化逻辑时，却容易陷入"语义漂移"与"幻觉"。生成的代码往往语法完美，但在 经济学逻辑上支离破碎，甚至不仅无法复现，更难以在实盘中生存。 这引发了一个根本性的思考： 我们能否构建一个智能体，既拥有机器般不知疲倦的进化探索能力，又具备人类程序员般严谨的逻辑控制 力？ 正是在这一背景下，QuantaAlpha团队联合上海财经大学AIFin Lab等机构，提出了 QuantaAlpha 框架：一种全新的自进化Agent Alpha 因子挖掘框架。 核心框架解析：基于轨迹的自进化范式 Quant ...

Core Viewpoint - The article emphasizes the ongoing AI boom and invites individuals to join the company "Quantum Bit," which focuses on tracking AI advancements and has established itself as a leading content platform in the industry [1]. Group 1: Job Opportunities - The company is hiring for three main directions: AI Industry, AI Finance, and AI Product, with positions available for both experienced professionals and fresh graduates [2][4]. - Positions are open for various levels, including editors, lead writers, and chief editors, with a focus on matching roles to individual capabilities [6]. Group 2: Job Responsibilities - **AI Industry Direction**: Responsibilities include tracking innovations in infrastructure, such as chips, AI infrastructure, and cloud computing, as well as interpreting technical reports from conferences [6][7]. - **AI Finance Direction**: Focuses on venture capital, financial reports, and capital movements within the AI industry, requiring strong analytical skills and a passion for interviews [11]. - **AI Product Direction**: Involves monitoring AI applications and hardware developments, requiring a keen understanding of product experiences and market trends [11]. Group 3: Benefits and Growth - Employees will have the opportunity to engage with industry leaders, enhance their personal influence through original content creation, and receive professional mentorship from senior editors [6]. - The company offers competitive salaries and comprehensive benefits, including social insurance, meal allowances, and performance bonuses [6]. Group 4: Company Growth Metrics - By 2025, Quantum Bit aims to have over 2.4 million subscribers on WeChat and a total of over 7 million users across platforms, with a daily reading volume exceeding 2 million [12].

Core Insights - The article discusses the capabilities and performance of the GLM-OCR model, highlighting its competitive edge in the OCR technology landscape, particularly in complex scenarios like handwriting and table recognition [1][39]. Performance Comparison - GLM-OCR outperforms several competitors in various OCR tasks, achieving a document parsing accuracy of 94.6% on OmniDocBench V1.5, surpassing PaddleOCR and others [2]. - In text recognition, GLM-OCR achieves 94.0% accuracy, significantly higher than some competitors like Deepseek-OCR2, which only reaches 34.7% [2]. - For formula recognition, GLM-OCR scores 96.5%, indicating strong performance in recognizing mathematical expressions [2]. - The model also excels in table recognition, with an accuracy of 85.2% on PubTabNet, outperforming many alternatives [2]. Practical Applications - GLM-OCR is particularly effective for structured documents such as Word, PPT, and academic papers, as well as for recognizing clear handwriting, receipts, and scanned contracts [3][4]. - The model demonstrates strong capabilities in recognizing handwritten forms, achieving an accuracy of 86.1% [4]. - It can accurately extract information from various documents, including meeting minutes and whiteboard notes, making it suitable for everyday work scenarios [3][4]. User Experience - Users report a generally positive experience with GLM-OCR in standard document parsing tasks, although challenges remain with unclear handwriting and complex layouts [4][12]. - The model's ability to handle low-quality inputs is commendable, with a recognition accuracy of around 96% for mixed content, although some errors were noted in specific cases [13][29]. Structural Extraction - GLM-OCR is capable of structured information extraction, producing outputs in standard JSON format from various documents, which is beneficial for applications like invoicing and identification [36][38]. - The model's performance in structured extraction improves significantly when clear prompts are provided, indicating its adaptability to user requirements [38]. Industry Trends - The OCR technology market is rapidly evolving, with new models like GLM-OCR emerging to meet increasing demands for efficiency and accuracy [39][40]. - The trend towards smaller model parameters (0.07B to 0.9B) is making deployment easier and more cost-effective for users [51]. - Enhanced output quality and reduced processing times are becoming standard expectations in the OCR industry, benefiting users across various sectors [51].

henry 发自 凹非寺 量子位 | 公众号 QbitAI 最强的大模型，已经把scaling卷到了一个新维度： 百万级上下文 。 几天前，Claude Opus 4.6发布，让人第一次真切感受到了百万上下文的涌现能力—— 单次吃进50万字中文内容、实现跨文档法律分析、多轮Agent规划…… 此情此景，用户火速用脚投票，华尔街更是直接给出K线回应。 与此同时，基于SALA注意力架构的模型 MiniCPM-SALA 也将一并开源。 除此之外，面壁还以OpenBMB社区名义，联合SGLang与NVIDIA发起 2026稀疏算子加速大奖赛（SOAR） ，将这套scaling能力直接交到 开发者手中，推动端侧Agent部署的性能突破。 Linear-Sparse混合注意力架构 太长不看，咱直接说重点—— 面壁这次全新的 线性与稀疏注意力混合架构SALA（Sparse Attention-Linear Attention，SALA） ，究竟是怎么个混合法呢？ 简单来说，这套架构将 75%线性注意力（Lightning Attention） 与 25%稀疏注意力（InfLLM v2） 结合，并通过 混合位置编码HyPE ...

梦晨 发自 凹非寺 量子位 | 公众号 QbitAI 24小时内，马斯克的xAI连失两位华人联合创始人。 xAI联合创始人 吴宇怀 （Tony Wu）和 Jimmy Ba 先后在社交平台上宣布离职。 而就在一个月前，另一位华人联合创始人 杨格 （Greg Yang）刚刚因病退出。 三位华人核心科学家，一个月内全部离开。算上此前已出走的三人，xAI成立不到三年，最初12人创始团队已走6人。 同时，一些后加入的核心成员，也纷纷宣布离职。 马斯克的AI团队，发生了什么？ 一月三别：从因病退出到师徒接连告别 这一波离职潮最先离开的是 杨格 。 2026年1月，这位Grok核心架构师宣布，自己被诊断出患有 莱姆病 （Lyme disease） ，不得不退出日常工作，转为公司的非正式顾问。 杨格拥有哈佛大学数学系学位， 师从著名数学家丘成桐 ，曾是微软研究院的研究员。 他在声明中解释自己可能早已感染此病，但在xAI创立期间的 "长期高强度工作"和"把自己逼得太狠" 导致免疫系统受损，最终使病情显现和 恶化。 紧接着是 吴宇怀 。2月10日，他在X平台上发布离职声明： 是时候开启我的下一章了，这是一个充满可能性的时代：一个 ...

Core Viewpoint - The article discusses the limitations of current multimodal large models in fine-grained visual recognition tasks and introduces the Fine-R1 model developed by Professor Peng Yuxin's team at Peking University, which significantly improves recognition accuracy with minimal training data [1][2][5]. Group 1: Fine-Grained Visual Recognition Challenges - Current multimodal large models excel in complex tasks but lag in fine-grained visual recognition compared to their visual encoders like CLIP [1]. - Real-world objects exhibit fine-grained characteristics, with numerous subclasses, such as over 500 types of fixed-wing aircraft, highlighting the importance of fine-grained recognition in practical applications [3]. Group 2: Fine-R1 Model Overview - The Fine-R1 model aims to leverage the rich knowledge of fine-grained subclasses and a generative decoding paradigm to overcome the limitations of traditional recognition methods, enabling fine-grained recognition of any visual object in an open domain [5]. - Fine-R1 enhances the model's ability to reason about unseen subclasses using a small number of training images (only 4 per subclass), outperforming models like OpenAI's CLIP and Google's DeepMind's SigLIP [5][15]. Group 3: Model Development Process - The development of Fine-R1 involves two main steps: 1. Chain-of-thought supervised fine-tuning, which simulates human reasoning to build inference capabilities [7]. 2. Triplet enhancement strategy optimization, which improves robustness to intra-class variations and inter-class distinctions by using positive and negative samples [8][10]. Group 4: Experimental Results - Fine-R1's performance was evaluated on six authoritative fine-grained image classification datasets, demonstrating superior accuracy in both seen and unseen categories compared to other models [15][17]. - The model's ability to utilize fine-grained subclass knowledge effectively was identified as the primary factor for its improved recognition accuracy, rather than enhancements in visual representation or knowledge storage [19]. Group 5: Conclusion and Future Work - The article concludes with the potential of Fine-R1 to excel in fine-grained visual recognition tasks, emphasizing its innovative approach to reasoning and knowledge application [21]. - The research has been accepted for ICLR 2026 and the code is open-sourced for further exploration [2][22].