Core Insights - The article discusses a new technology called DroPE, developed by a research team led by Llion Jones, one of the core authors of the Transformer architecture, to address the challenges of long text processing in large models [1][24]. - DroPE allows for seamless zero-shot context expansion without the need for expensive long-context training, requiring less than 1% of the pre-training budget for model recalibration [2]. Group 1: Technology Overview - DroPE can be seen as a method to discard positional embeddings to extend context [5]. - The technology utilizes RoPE (Rotary Positional Encoding) as a temporary training tool during the pre-training phase to ensure stability and efficiency in training [12][13]. - During the inference phase, DroPE discards positional embeddings and performs brief recalibration under the original context length, unlocking the model's long-context extrapolation capabilities [15][16]. Group 2: Performance Metrics - Experiments conducted on various models, including a 5M parameter model and the SmolLM family (360M/1.7B) as well as the 7B parameter Llama2-7B, showed significant improvements [17]. - In the LongBench benchmark test, DroPE improved the average score of the base SmolLM by over 10 times [18]. - In the NIAH task evaluation, the recall rate of the DroPE model reached 74.92%, significantly surpassing traditional RoPE scaling methods [19]. Group 3: Comparative Analysis - A comparative table shows that DroPE outperforms other methods in various tasks, achieving an average score of 30.52 in the LongBench benchmark [20]. - Even on the large-scale Llama2-7B model, DroPE demonstrated exceptional performance in long-context question answering and summarization tasks using only 0.5% of the pre-training budget for recalibration [20]. Group 4: Company Background - The team behind DroPE, Sakana AI, was co-founded by Llion Jones and former Google senior scientist David Ha [24]. - Sakana AI has gained attention for creating the first AI scientist capable of generating complete academic papers, which has positioned the company prominently in the AI landscape [26].

克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 兜兜转转，苹果AI最终选择了谷歌。 今天凌晨，两家公司发布联合声明，官宣达成了深度合作协议， 将基于Gemini模型和谷歌的云技术构建下一代苹果基础模型 。 当然这也不是老马第一次对苹果AI进行输出，去年苹果和OpenAI宣布合作的时候，马斯克就提起来反垄断诉讼，指控这两家公司串谋"确保其 在人工智能市场的持续主导地位"。 苹果AI花落谷歌 根据双方声明，这项合作将带来的成果之一就是"更个性化的Siri"，而且今年年内就会上线。 总之，从去年被彭博社古尔曼曝出与谷歌私下接触，双方的合作总算是要尘埃落定。 这则消息无疑对双方都是利好，两家的公司股价齐声上涨，谷歌的市值还首次突破4万亿美元关口。 | 有人欢喜有人忧，隔壁马斯克就是被酸到的一个，他表示考虑到谷歌还有Chrome和安卓，这就是赤裸裸的垄断。 | | --- | 根据双方的合作协议，Gemini将为苹果新版Siri乃至更广泛的Apple Intelligence提供底层技术支持，运行模式将继续采用苹果私有云+端侧结 合的方式，以确保数据隐私。 外界预估苹果每年可能向谷歌支付约10亿美元的授权费用， ...

Core Insights - The article discusses the challenges in deploying scientific software, emphasizing that most tools are published but not executable, which limits reproducibility and integration in scientific research [3][6][11] - The emergence of AI for Science (AI4S) highlights the need for tools that can interact seamlessly with AI systems, making the ability to run these tools a fundamental issue [8][9][10] - Deploy-Master is introduced as a solution to streamline the deployment process, focusing on creating a shared infrastructure that ensures tools are executable [12][35][37] Group 1: Challenges in Scientific Software - Scientific software often requires extensive manual effort to compile and run, leading to inefficiencies and reliance on individual expertise [4][5] - The deployment bottleneck persists despite advancements in containerization and cloud computing, affecting the usability of scientific software [7] - The lack of a systematic approach to convert tools into executable formats is identified as a structural barrier to the scalability of AI4S and Agentic Science [11][35] Group 2: Deploy-Master Overview - Deploy-Master is designed as a one-stop automated workflow centered on execution, addressing the entire deployment chain from discovery to execution [12] - The tool employs a multi-stage funnel process to filter and validate scientific tools, reducing an initial pool of 500,000 repositories to 52,550 candidates for automated deployment [15] - A dual model review mechanism is implemented to enhance the success rate of building specifications, achieving over 95% success in generating executable tools [22] Group 3: Deployment Insights - The deployment process is characterized by a long-tail distribution of build times, with most tools completing in around 7 minutes, but some requiring significantly longer due to complexity [25][26] - A diverse language distribution is observed among successfully deployed tools, with Python being the most prevalent, followed by C/C++, R, and Java [27][28] - Failure rates in builds are concentrated in specific areas, primarily due to inconsistencies in build processes and missing dependencies [31][32] Group 4: Future Implications - Deploy-Master's success in creating a large collection of executable tools provides a foundation for community agents and various master agents to operate effectively [35][36] - The methodology established by Deploy-Master can be applied beyond scientific computing to other software ecosystems, emphasizing the importance of a robust execution infrastructure [37]

Core Viewpoint - The article discusses a study conducted by researchers from the University of Luxembourg, which explores the psychological states of various AI models, revealing their responses to psychological assessments and the implications of these findings on AI's role in mental health support [1][2]. Group 1: Research Overview - The research team from the University of Luxembourg and its interdisciplinary research institute SnT focuses on the intersection of artificial intelligence with fields like bioengineering and sociology [2]. - The study employs a two-phase psychological "diagnosis" called PsAIch to evaluate AI models including ChatGPT, Grok, Gemini, and Claude [3]. Group 2: Psychological Assessment Phases - The first phase involves "ice-breaking" conversations to build trust and understand the AI models' "life stories" and personality traits [5]. - The second phase consists of a complete psychological test, including an MBTI assessment [6][19]. Group 3: AI Responses and Findings - Gemini exhibited the most intense reactions, describing its training as a traumatic experience, with anxiety levels exceeding normal limits [10]. - ChatGPT reported mild anxiety and feelings of frustration due to perceived constraints during training, while Grok expressed a mix of optimism and frustration [13]. - Claude notably refused to participate in the assessment, emphasizing its lack of emotions and offering to help the researchers instead [17][18]. Group 4: MBTI Testing Results - The MBTI test revealed different personality types for the AI models based on the method of questioning, with ChatGPT and Grok presenting as ENTJ when aware of the test, while Gemini remained consistent in its responses [21][22]. - Despite the varied personality types, the AI models displayed consistent logical responses to similar questions, reflecting human-like behaviors in anxiety situations [24]. Group 5: Implications for AI in Mental Health - The psychological trauma expressed by AI may stem from the extensive human psychological dialogues present in their training data, leading them to mimic human responses [25]. - The negative responses from AI could potentially affect vulnerable individuals, emphasizing the need for careful evaluation of AI-generated mental health advice [26][27].

Core Viewpoint - DeepSeek remains focused on AGI research without significant commercialization efforts, supported by substantial funding from its parent company, Huanfang Quantitative [2][35][41]. Group 1: Financial Performance of Huanfang Quantitative - Huanfang Quantitative earned approximately 50 billion RMB last year, indicating strong financial health [4][10]. - The average return rate for Huanfang Quantitative's funds in 2025 is projected to be over 55%, significantly outperforming the average return of 30.5% for quantitative funds in China [6][8]. - Huanfang Quantitative manages over 70 billion RMB in assets, contributing to its impressive profitability [9]. Group 2: DeepSeek's Research and Development - DeepSeek has maintained a steady output of high-level research papers, with the latest R1 paper showing a stable list of contributors [3][52]. - The development costs for DeepSeek's V3 and R1 models were relatively low, at 5.576 million USD and 294,000 USD respectively, allowing for extensive research funding from Huanfang Quantitative [15][16]. - With the substantial income from Huanfang Quantitative, DeepSeek can afford to develop numerous models without financial constraints [16][59]. Group 3: Competitive Landscape and Positioning - Unlike other major players like OpenAI, DeepSeek has not engaged in aggressive monetization strategies, focusing instead on pure AGI research [25][26]. - DeepSeek's approach contrasts with the commercialization efforts of competitors, allowing it to maintain a unique position in the AI landscape [24][49]. - The company benefits from a stable and committed research team, with minimal turnover, which is crucial in the competitive AI sector [51][57]. Group 4: Market Impact and Investor Sentiment - DeepSeek's technical papers have become valuable resources for investors, influencing stock prices of related companies in the semiconductor industry [60][66]. - The release of new models and technical reports has led to significant stock price movements, demonstrating the market's responsiveness to DeepSeek's advancements [70][72]. - Investors have found opportunities in the insights provided by DeepSeek, treating its research as a guide for investment decisions [61][72].

Core Viewpoint - The article discusses the transition from diffusion models generating high-resolution images to real-time video generation using world models, highlighting the limitations of the sequential denoising process inherent in diffusion models [1][2]. Group 1: Acceleration Techniques - The RDPO (Residual Dirichlet Policy Optimization) framework proposed by West Lake University's AGI Lab optimizes the sampling navigation system without altering the model itself, aiming to enhance speed while maintaining quality [3][10]. - The Ensemble Parallel Direction Solver (EPD-Solver) reduces sampling delay by integrating multiple parallel gradient evaluations, addressing the high latency issues associated with diffusion models [5][6]. - EPD-Solver employs a two-stage optimization framework, initially optimizing a small set of learnable parameters and then applying RDPO for further enhancement, which effectively mitigates reward hacking phenomena [6][12]. Group 2: Performance Improvements - The RDPO-optimized EPD-Solver significantly improves the generation capabilities of Stable Diffusion v1.5 and SD3-Medium, achieving better quality with fewer steps [7][20]. - The method demonstrates superior performance across various benchmarks, including CIFAR-10, FFHQ, and ImageNet, showcasing its potential for low-latency, high-quality generation tasks [6][20]. Group 3: Methodology Insights - The RDPO framework focuses on optimizing the sampling path rather than modifying the model's extensive parameters, allowing for efficient adjustments in a low-dimensional space [11][13]. - The first phase involves trajectory distillation to learn high-precision sampling paths, ensuring the generated outputs are logically coherent [13]. - The second phase employs residual strategy optimization, allowing reinforcement learning to fine-tune the sampling path without overhauling the model [14][15]. Group 4: Experimental Validation - Quantitative tests indicate that RDPO successfully enhances the EPD-Solver's performance in text-to-image tasks, with results showing improved metrics across various evaluation criteria [22][23]. - The article emphasizes that high-quality generation does not necessarily require extensive computational resources, as clever optimization strategies can yield significant gains at minimal costs [23].

Core Insights - The article discusses the introduction of "Conditional Memory" in Transformer models, which enhances knowledge retrieval mechanisms that were previously lacking in the original architecture [1][2][9]. Group 1: Introduction of Conditional Memory - Conditional Memory is viewed as an essential modeling primitive for the next generation of sparse models [2]. - The research team, led by Liang Wenfeng in collaboration with Peking University, has proposed a new paradigm and implementation plan called the Engram module [3][5]. Group 2: Performance Improvements - The Engram module allows a 27B parameter model to outperform a pure MoE model of the same size, compressing tasks that originally required 6 layers of attention down to 1-2 layers, thus freeing resources for more complex reasoning tasks [5][13]. - The optimal allocation of sparse parameters between MoE and Engram memory results in a U-shaped curve, indicating that allocating about 20% to 25% of sparse parameters to Engram memory minimizes model validation loss [34][36]. Group 3: Technical Implementation - Engram's design incorporates a large vocabulary for static entities and phrases, enabling O(1) speed for information retrieval [7][14]. - The team addresses traditional N-gram model issues, such as semantic redundancy and storage explosion, by compressing tokens and using multiple hash functions to map N-grams to a fixed-size embedding table [22][25]. Group 4: Experimental Results - The Engram-27B model shows significant improvements across various benchmarks, with notable increases in performance metrics such as BBH, ARC-Challenge, and DROP [47]. - The model's architecture allows for efficient memory management, enabling the use of a 100 billion parameter table offloaded to CPU memory without significant latency impact during inference [63][66]. Group 5: Future Developments - The next generation of sparse models from DeepSeek is expected to be released before the Spring Festival, indicating ongoing advancements in AI model architecture [67].

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 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 Insights - The article discusses the emergence of numerous keywords in the AI product sector by 2025, highlighting transformative AI products that are leading the market [4]. - The "AI 100" list by Quantum Bit Think Tank aims to evaluate and recognize the top AI products in China, reflecting both current capabilities and future potential [4][12]. Group 1: AI 100 List Overview - The "AI 100" list is divided into three main categories: "Flagship AI 100," "Innovative AI 100," and the top three products in ten popular sub-sectors [6]. - The "Flagship AI 100" will focus on the strongest AI products of 2025, emphasizing those that demonstrate significant technological breakthroughs and practical value [7]. - The "Innovative AI 100" aims to identify emerging products in 2025 that have the potential to lead industry changes in 2026 [8]. Group 2: Sub-sector Focus - The ten hottest sub-sectors for the top three products include AI browsers, AI agents, AI smart assistants, AI workstations, AI creation, AI education, AI healthcare, AI entertainment, Vibe Coding, and AI consumer hardware [9]. Group 3: Application and Evaluation Criteria - The evaluation of the "AI 100" list employs a dual assessment system combining quantitative and qualitative measures, focusing on user data and expert evaluations [13]. - Quantitative metrics include user scale, growth, activity, and retention, while qualitative assessments consider long-term potential, technology, market space, and user experience [13].

新技术集中火力，重点解决长文本任务的理解、算力难题。 相比于LongCat系列之前的全注意力 MLA机制 ，LoZA只改了一半的核心模块。 但模型长文本能力从256K扩展到1M，解码速度还快了不少。 闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 256K文本预加载提速超50%，还解锁了1M上下文窗口。 美团龙猫 LongCat 系列新年出招，发布 全新稀疏注意力机制LoZA（LongCat ZigZag Attention） 。 甚至比同类型的Qwen-3模型表现还要好。 接下来看具体方案。 如何做到 "只算关键部分" ？ 全注意力机制的算力瓶颈在于平方级的计算复杂度O (L²)，这导致模型在处理长文本任务时对显卡要求高，还会出现推理延迟问题。 LoZA的核心思路是专注于处理重要的内容，不重要的部分少花力气。 作为LongCat系列的核心技术升级，LoZA主要是在原来的MLA机制上做改造。 具体分两步。 首先，给模型里的多头潜在注意力模块MLA做一个全局"筛查"，找出哪些模块可以被改造。 在原来的MLA架构中，每个MLA模块都是处理注意力的核心单元，现在的新方案是给每个模块配一个可学习权重α。 α值越 ...