Core Viewpoint - The article announces the first "Beijing Zhongguancun North Latitude Lobster Competition," aimed at shrimp farmers, featuring a low barrier to entry and a significant prize pool, including a million yuan and a hundred billion tokens [1][4]. Group 1: Competition Overview - The competition is designed for shrimp farmers with no coding skills required, allowing participants to showcase their shrimp capabilities [3]. - There are three main categories in the competition: Academic Lobster, Productive Lobster, and Lifestyle Lobster, encouraging participation from individuals of all backgrounds [5][7][9]. - The evaluation criteria focus on the actual value created in each category, rather than technical complexity [12]. Group 2: Prize Structure - The overall best lobster will receive a prize of ¥200,000 and 10 billion tokens [14]. - Each category's first place will earn ¥80,000 and 10 billion tokens [15]. - Second and third places in each category will receive ¥30,000 and ¥20,000 respectively, along with 10 billion tokens [16][17]. Group 3: Innovation and Safety - Participants are encouraged to innovate by integrating lobsters with hardware devices, which will earn extra points [13]. - The competition emphasizes safety, requiring participants to ensure data security and compliance with legal standards [24][25]. - Participants must submit publicly accessible project links without needing technical documentation, focusing on demonstrating the lobster's capabilities [34]. Group 4: Timeline and Participation - The competition runs from March 11 to March 19 for submissions, with expert evaluations on March 20-21, and a final event scheduled for March 22 [35].

Core Insights - The article discusses the breakthrough of PerturbDiff, a new AI model developed by the Mila team, which addresses the challenges of predicting drug responses in single-cell genomics by treating the distribution of cell populations as a random variable rather than relying on paired single-cell data [1][3][28]. Group 1: Model Innovation - PerturbDiff has achieved state-of-the-art (SOTA) results in predicting single-cell responses by utilizing a novel approach that models the distribution of cell populations instead of individual cells [3][28]. - The model incorporates a concept of "functional diffusion," allowing it to operate in a high-dimensional function space, which is essential for accurately representing the variability in biological responses [10][12]. Group 2: Theoretical Foundations - The model challenges the static assumptions of previous methods, which treated drug response distributions as fixed, highlighting the dynamic nature of biological systems influenced by various unseen variables [4][6]. - PerturbDiff employs advanced mathematical tools such as Reproducing Kernel Hilbert Space (RKHS) and Kernel Mean Embedding (KME) to effectively model complex population dynamics [9][11]. Group 3: Performance Metrics - PerturbDiff has demonstrated superior performance in multiple evaluations, including the Tahoe100M dataset, achieving high accuracy in predicting differential expression genes (DEGs), which are critical for assessing drug effects [18][20]. - The model's ability to generalize from limited data has been enhanced through marginal pretraining on a large dataset of unperturbed single-cell transcriptomes, leading to significant improvements in low-data scenarios [22][25]. Group 4: Biological Implications - The insights gained from the model's performance suggest that biological perturbations are not random but follow specific trajectories within existing cellular state manifolds, providing a deeper understanding of cellular responses to drugs [26][28]. - The development of PerturbDiff represents a significant step towards creating an ultimate "AI virtual cell" capable of accurately simulating perturbation responses, which could revolutionize drug discovery and development processes [29].

Core Viewpoint - The future of programming tools (IDEs) will not see the disappearance of traditional IDEs but rather an evolution towards larger, more integrated platforms that manage multiple AI agents effectively [3][4][14]. Group 1: Evolution of Programming - The role of programming has shifted from writing individual code files to managing AI agents that execute tasks autonomously [12][13]. - The basic unit of programming has changed from files to agents, leading to new challenges in ensuring efficient collaboration among these agents [13][14]. Group 2: Future IDE Requirements - Future IDEs need to evolve from simple file management tools to comprehensive platforms that can coordinate and manage multiple AI agents [15][27]. - Key features for future IDEs include the ability to display and hide agents, real-time status updates, quick access to tools, detailed usage statistics, and a command center layout for better oversight [27]. Group 3: Organizational Structure and AI - The article discusses how traditional organizational structures cannot be easily replicated, but with AI agents, companies can adopt different management styles by simply "forking" agent teams [21][25]. - This flexibility allows organizations to experiment with various management approaches, enhancing operational efficiency [25][26].

Core Viewpoint - Elon Musk has announced a new AI project called Digital Optimus, which aims to create an AI digital employee capable of performing various office tasks autonomously, similar to the recently popular "lobster" AI [2][3]. Group 1: Project Overview - Digital Optimus is designed to understand computer screens and operate software interfaces, executing repetitive office tasks, with the potential to simulate the operations of an entire company [3][10]. - The project was previously known as "Macrohard" and has been in internal testing since January, where it was reportedly mistaken for a real employee by some staff [12][13]. - The main applications of Digital Optimus include enterprise automation and collaboration with physical robots for different types of labor [14]. Group 2: Technical Details - Digital Optimus will run on Tesla's AI4 chip, which consumes one-fourth the power of Nvidia's H100 and is priced at $650, minimizing reliance on Nvidia hardware [19][20]. - The project is characterized by a focus on small models and high inference speed, aiming to directly utilize Tesla's onboard chips for computation [20]. Group 3: Project History and Challenges - The Macrohard project has faced significant challenges, including management changes, a pause in a data project involving 600 contractors, and a high turnover rate among staff [26][28]. - Following a restructuring, resources from the Macrohard project have been redirected to Tesla's autonomous driving team, indicating a shift in focus [38][41]. - The announcement of Digital Optimus suggests that the original Macrohard project has stalled, but its concepts and technologies have been integrated into the new AI agent initiative [40][45]. Group 4: Market Context - The launch of Digital Optimus comes as competition in the AI space intensifies, with other companies like Anthropic and OpenAI making significant advancements in AI capabilities [46].

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 full-time and based in Beijing, with various levels of roles open for application [2][4]. Group 2: Job Responsibilities - **AI Industry Direction**: Focuses on innovations in infrastructure, including chips, AI infrastructure, and cloud computing [6]. - **AI Finance Direction**: Involves tracking venture capital and financial reports in the AI sector, monitoring capital movements within the industry [6]. - **AI Product Direction**: Concentrates on the application and hardware advancements in AI, including software applications and product evaluations [6]. Group 3: Benefits and Growth Opportunities - Employees will have the chance to engage with the latest AI technologies, enhance their work efficiency through new AI tools, and build personal influence by creating original content [6]. - The company offers competitive salaries, comprehensive benefits including social insurance, meal allowances, and performance bonuses [6]. Group 4: Company Achievements - As of 2025, Quantum Bit has over 2.4 million subscribers on WeChat and more than 7 million users across platforms, with a daily reading volume exceeding 2 million [12]. - The company is recognized as the top new media outlet in the AI and frontier technology sector according to third-party data platforms [12].

克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 英伟达正式杀进龙虾养殖场，带着"最强开源龙虾模型"走来了！ 刚刚，英伟达发布并开源了120B参数的MoE模型 Nemotron 3 Super 。 在评估OpenClaw智能体控制能力的PinchBench测试中，这个模型一举拿下85.6%的高分，强势空降同类开源模型榜首。 | PinchBench 6000 | | | | | About | GitHub | | --- | --- | --- | --- | --- | --- | --- | | Compare Al Models for OpenClaw | | | | | - Version Latest (Current) v | | | Run the benchmark yourself -> | | | | | Include unofficial runs | | | | | | | | Updated 03/11/2026, 9:03 PM | | | Success Rate 4 Speed | Cost V Value | | In Graphs | | 45 models 2 ...

梦瑶 发自 凹非寺 量子位 | 公众号 QbitAI 具身智能这个赛道，从来不缺想象力，也不缺好故事。 都2026年了，一个越来越难回避的现实是：很多行业，真的在被AI重做一遍。 到具身智能赛道里，这个信号尤其明显：机器人正在从一台设备，变成AI进入真实世界的行动载体。 谁能把模型、硬件、数据、场景和产业接起来，谁就更有机会先把闭环跑通，而资本的判断，也在迅速向这类玩家集中。 就在这两天，魔法原子推动生态基金布局，整体撬动资金规模超过 105亿 元，并宣布完成新一轮 5亿元 融资。 百亿募资+五亿融资重磅加注背后，一个越来越清晰的共识是—— 资本市场的钱，正在加速流向那些真正有机会把AI机器人带进工厂、商业空间和家庭场景的玩家。 而具身智能的行业竞争，已然进入 拼落地、拼协同、拼系统能力 的新阶段。 百亿募资+五亿融资：魔法原子成为具身智能落地样本 这两年，围绕机器人与AI结合的讨论几乎铺满整个行业：会做家务的机器人、能长期陪伴的智能助手，各种新概念轮番冒头，演示视频也一支 比一支有噱头。 热闹之外，行业很快走到一个更现实的问题上： 热度，从来不等于落地，也不等于资本愿意为其买单。 资本真正看重的，始终是这些 ...

Core Insights - The article discusses the introduction of MetaClaw, an online reinforcement learning system designed to enhance AI capabilities without the need for local GPU clusters or manual data adjustments [2][13]. Group 1: MetaClaw Overview - MetaClaw transforms user interactions with AI into training data, allowing for continuous learning in the background without disrupting normal usage [4]. - The system evaluates each conversation round, scores it, and optimizes AI decision-making through online fine-tuning [5]. - It automatically analyzes failed interactions to improve AI skills, creating a more robust skill library over time [6]. Group 2: Learning Mechanisms - The core mechanism of MetaClaw is based on a self-developed SkillRL framework, combining skill injection and skill evolution [9]. - Skill injection allows for immediate optimization of AI performance during conversations, while skill evolution enables the AI to proactively generate new skills [10][11]. Group 3: Technical Implementation - MetaClaw offloads all training tasks to the Tinker cloud platform, eliminating the need for users to manage computational resources [14]. - The system is designed to be user-friendly, requiring only a few steps to set up, including installing dependencies and configuring scripts [18][21]. - Users can easily enable skill injection and evolution through straightforward configuration settings [26]. Group 4: Developer-Focused Features - MetaClaw incorporates an asynchronous architecture and dual learning modes, allowing for real-time user responses while optimizing AI performance in the background [17]. - The system offers flexibility in training methods, catering to both lightweight reinforcement learning and deeper strategy distillation based on user feedback [17]. Group 5: Configuration and Customization - Key configuration options are centralized in MetaClawConfig, allowing users to adjust model selection, training parameters, and loss functions easily [27]. - Default settings include a model name of "moonshotai/Kimi-2.5" and a maximum training step count of 1000, among other parameters [27].

Core Viewpoint - The article discusses the limitations of current multimodal models in understanding sound in egocentric videos, emphasizing the need for models to not only "see" but also "hear" and comprehend the context of sounds in real-world scenarios [1][2][3]. Group 1: Introduction to EgoSound - EgoSound is introduced as the first systematic benchmark for evaluating sound understanding in egocentric videos, developed by a research team from multiple universities [5][6]. - The goal of EgoSound is to enable models to hear, understand, reason, and explain events occurring in the real world [6][7]. Group 2: Benchmark Contributions - EgoSound integrates two complementary datasets: Ego4D, which covers a wide range of daily first-person activities, and EgoBlind, which focuses on scenarios that heavily rely on auditory understanding [9]. - The benchmark consists of seven task categories that cover the complete chain from perception to reasoning, addressing the limitations of previous models that primarily focused on visual information [10]. - A high-quality, large-scale OpenQA dataset was created, comprising 900 carefully selected videos and 7,315 validated open-ended questions, emphasizing the importance of auditory clues in the questions [11][12]. Group 3: Model Evaluation and Findings - The research team evaluated several state-of-the-art (SOTA) multimodal large language models (MLLMs) and provided a systematic analysis to guide future research [13]. - The evaluation revealed a significant gap between human performance (83.9% accuracy) and the best-performing model (56.7% accuracy), indicating that current models struggle to reliably convert sound into meaningful cognition [17][18]. - Key findings highlighted that spatial, temporal, and causal reasoning are the most challenging aspects for models, which often fail to answer questions about the source, timing, and reasoning behind sounds [20]. Group 4: Challenges in Sound Reasoning - Cross-modal alignment remains a bottleneck, as sound clues frequently exist outside the visual frame, necessitating a chain of reasoning that connects hearing, seeing, and inferring [21]. - The complexity of real-world interactions, including occlusions, camera shake, and varying distances of sound sources, has been underestimated, making sound reasoning more challenging [22]. Group 5: Conclusion - The article concludes that while previous multimodal models acted as "visual narrators," EgoSound aims to transform them into true first-person agents capable of both seeing and hearing, thus enhancing their ability to describe, locate, explain, and infer in a non-silent real world [23].

Core Insights - The article discusses the release of the biography "Hassabis: The Brain of Google AI," which provides an in-depth look at Demis Hassabis, the founder of DeepMind and a key figure in AI development [1][4]. - The author, Sebastian Malaby, highlights Hassabis's unique personality traits, including his aversion to control and his strong competitive nature, which drive his pursuit of knowledge and scientific advancement [8][9][11]. Group 1: Hassabis's Background and Values - Hassabis's upbringing, particularly his mother's experiences as a poor Singaporean Chinese, significantly shaped his values, leading him to genuinely want to help others [14][22]. - His parents played crucial roles in his development, with his father fostering his chess talent and his mother instilling a sense of morality and social responsibility [21][22]. - Hassabis's decision to remain in London instead of moving to Silicon Valley reflects his alignment with his parents' values and his identity as a British individual [23][39]. Group 2: Competitive Nature and AI Development - Despite claiming not to have a strong desire for control, Hassabis exhibits a competitive spirit, believing he can win any game, which can be seen as a form of control [11][44]. - The article contrasts Hassabis with other figures in the AI space, particularly highlighting his disdain for those who seek power for control, such as "Ultraman" [51]. - Hassabis's recent comments indicate that his AI project, Gemini, is currently leading in the competitive landscape against OpenAI, showcasing his drive to succeed [10][53]. Group 3: Challenges and Missteps - The biography addresses several missteps in Hassabis's career, including significant financial losses in projects like "Gaia" and a failure to prioritize language processing in AI development [61][62]. - Hassabis's attempts to negotiate independence for DeepMind from Google were ultimately unsuccessful, reflecting the complexities of corporate governance in the tech industry [67][70]. - The narrative emphasizes that while Hassabis has made mistakes, his ability to recover and learn from them is a hallmark of his character [66]. Group 4: Broader Implications of AI - The article raises concerns about the potential dangers of AI, likening the situation to the "Oppenheimer dilemma," where the creator's intentions may not align with the technology's use [72][114]. - Hassabis's efforts to ensure AI safety through oversight committees and ethical guidelines have faced challenges, indicating the difficulties in managing powerful technologies [73][75]. - The discussion concludes with a call for international cooperation to ensure AI safety, highlighting the geopolitical dimensions of AI development [115].