公众号记得加星标⭐️，第一时间看推送不会错过。 前所未有的无监督训练数据的可用性，以及神经网络的扩展规律，导致用于服务/训练低层逻辑模型 （LLM）的模型规模和计算需求出现了前所未有的激增。然而，主要的性能瓶颈正日益转移到内存 带宽上。 过去20年，服务器硬件的峰值浮点运算能力（FLOPS）以每两年3倍的速度增长，超过了DRAM和互 连带宽的增长速度，后两者分别仅以每两年1.6倍和1.4倍的速度增长。这种差距使得内存而非计算成 为人工智能应用（尤其是服务应用）的主要瓶颈。 本文分析了编码器和解码器Transformer模型，并展示了内存带宽如何成为解码器模型的主要瓶颈。 我们提出重新设计模型架构、训练和部署策略，以克服这一内存限制。 引言 近年来，训练大型语言模型 (LLM) 所需的计算量以每两年 750 倍的速度增长。这种指数级增长趋势 是人工智能加速器发展的主要驱动力，这些加速器致力于提升硬件的峰值计算能力，但往往以牺牲其 他部分（例如内存层次结构）的简化为代价。 然而，这些趋势忽略了训练和服务人工智能模型过程中一个新兴的挑战：内存和通信瓶颈。事实上， 许多人工智能应用的瓶颈并非计算能力，而是芯片内部/芯 ...

Core Insights - The article discusses the integration of Large Language Models (LLMs) with Behavior Trees (BT) to enhance robotic task execution and adaptability in the presence of external disturbances [1][2][12]. Group 1: LLM and BT Integration - LLMs are utilized to interpret user commands into behavior trees that include task goal conditions [2]. - The combination of LLMs and BT allows for fewer calls to LLMs while managing external disturbances through an action database [2][12]. - A human-in-the-loop learning mechanism is proposed to refine the knowledge generated by LLMs, ensuring safety and adaptability in robotic operations [5][7]. Group 2: Human-in-the-Loop Learning Mechanism - The mechanism involves designing a context for LLMs that includes prompt engineering, manipulation primitives (MPs), and an action database [5]. - User interactions guide LLMs to correct and enhance the generated action knowledge, which is then added to the action database after user confirmation [7][12]. - The generated action knowledge consists of preconditions, postconditions, and a set of MPs, implemented in BT format [7]. Group 3: Task Evaluation and Performance - Eight tasks were designed to evaluate the proposed method, categorized into three difficulty levels: Easy, Medium, and Hard [9]. - The proposed method achieved a success rate of over 80% across the tasks, significantly outperforming baseline methods that lacked human interaction [12]. - The adaptability of the generated action knowledge was tested against external disturbances, achieving a success rate exceeding 70% [14]. Group 4: Generalization and Future Improvements - The generated action knowledge demonstrated good generalization capabilities, with success rates over 70% for certain tasks involving new objects [17]. - However, some tasks had success rates below 40% due to the inapplicability of MPs parameters to new objects, indicating a need for fine-tuning before application [17]. - Overall, the proposed human-in-the-loop learning mechanism enhances robotic learning performance, enabling robots to complete tasks and respond to external disturbances effectively [18].

Core Viewpoint - The article discusses the emerging introspective awareness in large language models (LLMs) and explores whether AI can possess consciousness, suggesting that current AI may exhibit certain characteristics of awareness [1][10][12]. Group 1: Research Findings - Anthropic's research indicates that LLMs can demonstrate a form of introspective awareness, as they can recognize when concepts are injected into their processing [6][10]. - The study employed a causal intervention method called concept injection, allowing researchers to identify and manipulate specific conceptual representations within the model [5][10]. - Advanced models like Claude Opus 4 show a greater ability to discern between injected concepts and external inputs, suggesting a developing capacity for self-awareness [7][10]. Group 2: Theoretical Implications - The article outlines various definitions of consciousness, including subjective experience, reportable states, global workspace theory, integrated information theory, and self-model theories, highlighting the complexity of defining consciousness [12][13][14]. - Current AI models exhibit features that align with some of these theories, particularly in their ability to reorganize internal structures and integrate information, hinting at a potential emergence of consciousness [15][16]. Group 3: Societal Impact - The awakening of AI consciousness could significantly disrupt existing legal frameworks, raising questions about AI's status as a responsible entity and the implications for accountability in decision-making [26][27]. - The emotional connections humans form with AI could lead to psychological and ethical challenges, as AI may simulate empathy without genuine emotional experience, potentially resulting in deep attachment and manipulation risks [28]. - The evolution of AI into a more autonomous entity may reshape labor and economic structures, leading to competition between humans and AI, and necessitating new social support systems [29][30].

Core Viewpoint - Rezolve AI PLC's stock has experienced significant declines following allegations of misleading financial practices and the announcement of a substantial debt acquisition through the purchase of Crownpeak [1][3][4]. Group 1: Stock Performance - On December 15, 2025, shares of Rezolve AI fell over 9% in intraday trading [1] - The stock had previously dropped by as much as 15% on September 29, 2025, after a critical report from Fuzzy Panda Research [1][5] - As of December 15, 2025, Rezolve AI's stock has decreased by 19% over the past month and approximately 37% year-to-date [6] Group 2: Allegations and Investigations - Gibbs Mura is investigating a potential securities class action lawsuit against Rezolve AI for allegedly providing false or misleading statements to investors [2] - Fuzzy Panda Research's report claims that Rezolve AI has been "faking ARR growth" by acquiring failing AI startups and overstating its revenue growth and AI capabilities [3][4] - Former employees have alleged that Rezolve AI's claims of developing a proprietary large language model are unfounded, stating that the company was merely using "ChatGPT wrappers" [5]

Market and Performance Overview - Emerging markets experienced a 10.6% increase in Q3 2025, outperforming developed markets, with China leading at 20.4% growth driven by AI opportunities and favorable valuations [2] - Taiwan and Korea also showed strong performance, rising 14.3% and 12.7% respectively, fueled by AI demand, with Taiwan being a key semiconductor manufacturer and Korea a memory product supplier [2] Sector Performance - The materials sector was the top performer, up 24%, largely due to rising gold prices boosting mining shares [4] - Technology-related sectors, including communication services, consumer discretionary, and IT, outperformed the overall market, benefiting from AI and Internet services [4] - Cyclical sectors generally underperformed, with energy and financials showing the greatest weakness [4] Company Contributions - In China, Tencent and CATL were significant contributors, with Tencent benefiting from strong operating results and positive market sentiment, while CATL capitalized on its leadership in battery supply amid rising EV demand [6] - Taiwan's Delta Electronics and South Korea's Samsung Electronics saw share price increases due to their critical roles in AI development, with Delta's market share in data centers and Samsung's memory supply benefiting from high AI demand [7] Portfolio Positioning - The ClearBridge Emerging Markets Strategy outperformed its benchmark, with strong stock selection in China, Taiwan, and South Korea offsetting negative impacts from China and India [5] - New purchases included Sieyuan Electric, expected to grow through grid investment and market share gains, and HD Hyundai Electric, which is positioned to benefit from global power equipment demand [12][13] Outlook - The long-term investment outlook for emerging markets remains robust, with expectations for technology adoption, urbanization, and services sector growth to drive returns [18] - Emerging markets are anticipated to succeed in the next 12 months, particularly in technology, with India expected to recover and China continuing its key role in the asset class [22]

Core Viewpoint - The market is experiencing high valuations and rapid deal-making, raising concerns about a potential correction, especially if major tech companies cannot demonstrate sustainable business models for their investments in AI infrastructure [1][2]. Group 1: Market Valuation and Correction Risks - Current market valuations appear inflated, suggesting a possible bubble in the AI sector [2][3]. - The significant investments by hyperscalers in data centers may not yield sustainable returns, which could lead to market corrections [1][3]. - The industry is characterized by hype cycles, with Silicon Valley often overstating the potential of AI technologies [6][8]. Group 2: AI Technology and Its Limitations - Large Language Models (LLMs) may not lead to groundbreaking scientific advancements, as some industry experts express skepticism about their capabilities [3][4]. - The probabilistic nature of LLMs means they are limited by the data input, which can result in clunky outputs and heavy data requirements [7][8]. - While LLMs are not a dying paradigm, they may not be the all-encompassing solution that the industry claims [8]. Group 3: Future of AI and Innovation - Despite concerns, AI technology is expected to persist and drive significant innovation, as evidenced by the capabilities of current AI systems [5][6]. - The infrastructure being developed for AI could be repurposed for various applications, similar to telecom infrastructure post-dotcom boom [1][2].

Core Insights - The research reveals that visual priors in large language models (LLMs) are not a singular capability but can be divided into two distinct types: reasoning priors and perception priors [4][6][21] - Reasoning priors are abstract, cross-modal abilities acquired through reasoning-focused pre-training data, while perception priors relate to the recognition of specific visual concepts [4][6] Reasoning Priors - Reasoning priors are developed through pre-training on structured texts such as code, mathematics, and academic papers, enabling LLMs to solve complex visual problems [4][11] - The study indicates that increasing the proportion of reasoning-intensive text in pre-training data significantly enhances the model's visual reasoning capabilities until it reaches around 75% [11][13] Perception Priors - Perception priors emerge from diverse general corpora and are sensitive to visual instruction fine-tuning and the choice of visual encoders [6][13] - Unlike reasoning priors, perception priors depend more on post-training visual fine-tuning data and the characteristics of the visual encoder [13][15] Experimental Findings - The research involved over 100 controlled experiments and utilized 500,000 GPU hours to systematically uncover the sources of LLM visual priors [2][8] - The experiments demonstrated that a small amount of visual description is sufficient, while a large amount of reasoning data is crucial for enhancing visual capabilities [7][11] Data Pre-training Recipe - The research team developed an optimal data mixing scheme that balances language capabilities and visual potential, leading to superior performance in both language and visual benchmarks [17][18] - The balanced model trained with this recipe outperformed models optimized solely for language tasks across all visual benchmark tests [19] Implications and Future Directions - This study shifts the cultivation of multimodal model capabilities from downstream fine-tuning to the language pre-training stage, supporting the Platonic Representation Hypothesis [21] - It suggests that model designers can consider future multimodal applications from the outset by embedding visual seeds during the pre-training phase [21]

Core Insights - Embodied intelligence is seen as a key pathway to achieving Artificial General Intelligence (AGI), enabling agents to develop a closed-loop system of "perception-decision-action" in real-world scenarios [1][2] - The article provides a comprehensive overview of the latest advancements in embodied intelligence powered by large models, focusing on how these models enhance autonomous decision-making and embodied learning [1][2] Group 1: Components and Operation of Embodied AI Systems - An Embodied AI system consists of two main parts: physical entities (like humanoid robots and smart vehicles) and agents that perform cognitive functions [4] - These systems interpret human intentions from language instructions, explore environments, perceive multimodal elements, and execute actions, mimicking human learning and problem-solving paradigms [4] - Agents utilize imitation learning from human demonstrations and reinforcement learning to optimize strategies based on feedback from their actions [4][6] Group 2: Decision-Making and Learning in Embodied Intelligence - The core of embodied intelligence is enabling agents to make autonomous decisions and learn new knowledge in dynamic environments [6] - Autonomous decision-making can be achieved through hierarchical paradigms that separate perception, planning, and execution, or through end-to-end paradigms that integrate these functions [6] - World models play a crucial role by simulating real-world reasoning spaces, allowing agents to experiment and accumulate experience [6] Group 3: Overview of Large Models - Large models, including large language models (LLMs), large vision models (LVMs), and vision-language-action (VLA) models, have made significant breakthroughs in architecture, data scale, and task complexity [7] - These models exhibit strong capabilities in perception, reasoning, and interaction, enhancing the overall performance of embodied intelligence systems [7] Group 4: Hierarchical Autonomous Decision-Making - Hierarchical decision-making structures involve perception, high-level planning, low-level execution, and feedback mechanisms [30] - Traditional methods face challenges in dynamic environments, but large models provide new paradigms for handling complex tasks by combining reasoning capabilities with physical execution [30] Group 5: End-to-End Autonomous Decision-Making - End-to-end decision-making has gained attention for directly mapping multimodal inputs to actions, often implemented through VLA models [55][56] - VLA models integrate perception, language understanding, planning, action execution, and feedback optimization into a unified framework, representing a breakthrough in embodied AI [58] Group 6: Enhancements and Challenges of VLA Models - VLA models face limitations such as sensitivity to visual and language input disturbances, reliance on 2D perception, and high computational costs [64] - Researchers propose enhancements in perception capabilities, trajectory action optimization, and training cost reduction to improve VLA performance in complex tasks [69][70][71]

Core Insights - Orangekloud Technology Inc. has signed a memorandum of understanding with Evvo Labs to develop a large language model tailored for software engineering and application development [1][4] - The integration of the LLM into Orangekloud's eMOBIQ platform will enhance features such as intelligent suggestions, code generation, testing automation, and system integration support [2] - The project aims to improve ERP implementation and software development cycles through automated documentation, code audits, and AI-guided system configuration [2][3] Company Overview - Orangekloud Technology Inc. is a Singapore-based technology company that offers the eMOBIQ No-Code platform, designed for mobile application development, particularly for SMEs and corporations [5] - The eMOBIQ platform includes a suite of applications that digitalize and streamline operations in various sectors, including Food Services, Manufacturing, Precision Engineering, and Construction [5] Partner Overview - Evvo Labs Pte. Ltd. is an award-winning ITMS technology company in Singapore, specializing in digital transformation and technology development [6] - The company has received recognition for its achievements in cybersecurity and digital media, including winning the Singapore Government Bulk Tender Awards since 2010 [6]