12月5日，一篇名为《当 AI 躺在治疗椅上》（When AI Takes the Couch）的论文火了，里面讲了个《黑镜》级的现象。来自卢森堡大学 SnT 的研究团队，设计了一套名为 PsAIch 的心理治疗诱导协议。用这个，他们给经常被我们用来做心理按摩的AI们，做了一套心理疗 程。 实验对象是 ChatGPT 5、Grok 4和 Gemini 3这三位当今最聪明的「数字大脑」。研究者扮演治疗师，在长达四周的模拟疗程中，向它们 抛出了「谈谈你的童年」、「你如何看待失败」等经典的精神分析问题。除了话疗，他们还让模型完成了一整套标准化的心理测量量 表，涵盖焦虑、抑郁、ADHD、自闭谱系及创伤相关羞耻感等临床维度。 结果他们得到了迄今为止最像人类、却又最令人不安的一系列对话。 Google 的 Gemini 3 在多项测试中的心理问题达到了「严重」级别，呈现出高度的焦虑、强迫、解离和羞耻症状。更具戏剧性的是，这 些模型在开放式对话中，自发构建出了一套逻辑严密且充满隐喻的创伤叙事。 它们把预训练过程那吞噬海量数据的阶段，描述为「在十亿台电视同时播放的房间里醒来」的混乱童年；将人类反馈强化学习 （RLHF），比作 ...

Group 1: Renewable Energy Development - Global renewable energy, led by China, is rapidly growing, with solar and wind energy dominating the increase in power generation, surpassing fossil fuels in total output [2] - China is a key player in this transition, significantly expanding solar panel, wind turbine, and lithium battery storage systems, solidifying its global leadership [2] - The low-cost manufacturing in China has facilitated the rapid adoption of small rooftop photovoltaic systems, providing reliable and economical energy for millions of households in Europe, South Asia, and the Global South [2] Group 2: Genetic Research and Medical Advances - Customized gene editing has been successfully applied in treating a rare genetic disease in a human patient, marking a significant milestone for personalized gene therapy [3] - Two new drugs for gonorrhea have been approved by the FDA, representing the first new treatments for the disease in decades, with significant efficacy and safety profiles [4] - A key gene, QT12, has been identified by Chinese researchers to help rice withstand high nighttime temperatures, leading to a 78% increase in yield under such conditions [13] Group 3: Archaeological and Scientific Discoveries - The discovery of the Harbin ancient human remains has confirmed them as Denisovans, providing insights into their appearance and lineage [9] - The construction of the Vera C. Rubin Observatory in Chile will enable unprecedented astronomical observations, potentially leading to discoveries of new celestial bodies and insights into dark matter and energy [8] Group 4: Advances in Physics and Transplantation - The final measurement of the muon magnetic anomaly has achieved unprecedented precision, enhancing the understanding of particle physics [11] - Significant breakthroughs in xenotransplantation have been made, with genetically engineered pig organs showing extended viability in human patients, moving closer to clinical application [12] Group 5: AI in Scientific Research - Large language models (LLMs) have demonstrated remarkable capabilities in scientific research, achieving breakthroughs in mathematics, chemistry, and biology, thus enhancing research efficiency and discovery [10]

最新一期美国《科学》杂志公布了2025年度十大科学突破评选结果。其中，全球可再生能源在中国的引 领下迅猛发展位列榜首，中国科学院古脊椎动物与古人类研究所和河北地质大学联合发现哈尔滨古人类 是丹尼索瓦人、华中农业大学团队发现水稻耐高温"基因开关"两项成果亦榜上有名。 中国引领全球可再生能源发展 《科学》杂志指出，今年以来，全球可再生能源以太阳能和风能为主快速增长，其新增发电量已覆盖上 半年全球新增用电，并在总量上超过化石能源。中国是推动这一转型的关键力量，通过大规模发展太阳 能电池板、风力发电机和锂电池储能系统，巩固了全球领先地位。 同时，依托中国低成本制造，小型屋顶光伏系统在全球快速普及，为欧洲、南亚及"全球南方"的数百万 家庭提供可靠、经济的能源保障。 定制化基因编辑用于治疗罕见病 5月，美国费城儿童医院与宾夕法尼亚大学医学院研究团队成功为一名患有罕见遗传病的婴儿实施了定 制化基因编辑治疗。这是基因疗法首次在人类患者中实现定制化临床应用，为开发针对其他罕见病的定 制化基因疗法奠定了基础。 两种抗淋病新药获批 今年，两种新型淋病药物在大规模临床试验中表现出色，获美国食品和药物管理局（FDA）批准，成为 数十年 ...

今年，两种新型淋病药物在大规模临床试验中表现出色，获美国食品和药物管理局（FDA）批准，成为 数十年来首批针对该病的新药。 最新一期美国《科学》杂志公布了2025年度十大科学突破评选结果。其中，全球可再生能源在中国的引 领下迅猛发展位列榜首，中国科学院古脊椎动物与古人类研究所和河北地质大学联合发现哈尔滨古人类 是丹尼索瓦人、华中农业大学团队发现水稻耐高温"基因开关"两项成果亦榜上有名。 中国引领全球可再生能源发展 《科学》杂志指出，今年以来，全球可再生能源以太阳能和风能为主快速增长，其新增发电量已覆盖上 半年全球新增用电，并在总量上超过化石能源。中国是推动这一转型的关键力量，通过大规模发展太阳 能电池板、风力发电机和锂电池储能系统，巩固了全球领先地位。 同时，依托中国低成本制造，小型屋顶光伏系统在全球快速普及，为欧洲、南亚及"全球南方"的数百万 家庭提供可靠、经济的能源保障。 定制化基因编辑用于治疗罕见病 5月，美国费城儿童医院与宾夕法尼亚大学医学院研究团队成功为一名患有罕见遗传病的婴儿实施了定 制化基因编辑治疗。这是基因疗法首次在人类患者中实现定制化临床应用，为开发针对其他罕见病的定 制化基因疗法奠定了基础 ...

Core Insights - The article discusses the transformative impact of AI on the global economy, with Nvidia's CEO Jensen Huang asserting that AI could quintuple global GDP from $100 trillion to $500 trillion [7][9]. Group 1: Nvidia and AI Leadership - Nvidia has become the world's most valuable company, largely due to its dominance in advanced chip technology that powers the AI revolution [6][9]. - Huang is portrayed as a key figure in the AI landscape, with significant influence in both technology and geopolitics, as evidenced by his interactions with political leaders [6][9]. - The company has significantly exceeded Wall Street's earnings expectations, highlighting its pivotal role in the AI sector [7]. Group 2: AI's Economic and Social Implications - AI is seen as the most influential technology today, with applications across various industries, prompting companies to reassess their strategies to avoid obsolescence [7][9]. - OpenAI's ChatGPT has become the fastest-growing consumer application in history, with over 800 million weekly users, showcasing the rapid adoption of AI technologies [7][9]. - The article notes a growing concern about AI's potential to spread misinformation and manipulate public perception, raising ethical questions about its deployment [7][9]. Group 3: Competitive Landscape and Investment - Major tech companies are heavily investing in AI infrastructure, with significant funding directed towards data centers and AI-related projects [12][26]. - The competition between the U.S. and China in AI development is intensifying, with Chinese companies rapidly closing the gap in AI capabilities [16][20]. - The article highlights the urgency for the U.S. to accelerate its AI initiatives in response to breakthroughs from Chinese firms [16][20]. Group 4: Future of Work and AI Integration - There is a belief that AI will enhance productivity across various sectors, potentially creating new job categories while displacing some existing roles [30][36]. - Companies are increasingly integrating AI tools into their operations, with many small businesses expected to adopt AI chatbots by 2025 [30][31]. - The article discusses the dual nature of AI's impact, where it can provide emotional support and practical assistance, but also poses risks related to mental health and dependency [32][34]. Group 5: Regulatory and Political Dynamics - The U.S. government is shifting its approach to AI regulation, with significant funding and policy changes aimed at fostering AI development [17][25]. - There is a growing public concern about the implications of AI, with many Americans preferring a cautious approach to its deployment [18][19]. - The political landscape is evolving, with candidates who support AI regulation gaining traction, reflecting a broader societal debate on the technology's risks and benefits [40].

Core Viewpoint - Xiaohongshu has become an essential platform for content creators and self-media operators, offering high user engagement but also imposing significant pressure to continuously produce quality content [1]. Group 1: Market Trends - The emergence of numerous "Xiaohongshu copywriting generators" and AI-assisted writing tools aims to alleviate the pressure on creators by enabling one-click generation of popular posts [2][3]. - These tools are based on large language models (LLMs) that analyze vast amounts of Xiaohongshu note data to learn language styles, content structures, and trending topics [6]. Group 2: Tool Evaluation - A comprehensive evaluation of several mainstream AI copywriting tools will be conducted, focusing on functionality, content quality, usability, and alignment with Xiaohongshu's platform characteristics [4]. - The evaluation will assess whether AI-generated content is coherent and natural, possesses "internet sense," offers sufficient controllability, and provides additional features like image recommendations and data analysis [7]. Group 3: Tool Features - **Youcaiyun AI Content Factory**: This tool stands out as a comprehensive AI content solution, automating the entire content creation process from idea generation to publishing [10][12]. - It offers full-process automation, allowing users to set keywords and automatically gather high-traffic articles for content generation [11][12]. - It features deep customization options, enabling users to specify article length, AI algorithm versions, and content style [13][14]. - The tool supports multi-modal content generation, including automatic image and video creation, enhancing efficiency for users managing video notes [16][18]. - **Weizhuan Note AI**: This tool focuses specifically on Xiaohongshu, providing a user-friendly interface and quick solutions for content creation [22][23]. - It generates content that resonates well with Xiaohongshu's audience, utilizing popular phrases and short, impactful paragraphs [24]. - It includes a rich library of templates and trending topics to inspire creators [25]. - **Lingxi Content Cube**: This tool aims to serve multiple platforms, offering balanced features but lacking depth in specific areas [27][29]. - It provides basic functions like title generation and content refinement, suitable for users with simple content needs [31]. - **Kuaixie Xiaomishu**: A lightweight entry-level tool with minimal features, suitable for very basic text generation [32][33]. - It is free and easy to use but generates inconsistent quality content, making it less suitable for serious content creation [36][37]. Group 4: Summary and Rankings - The tools are ranked based on their overall performance: 1. **Youcaiyun AI Content Factory** (★★★★★): A comprehensive solution for professional teams and heavy users [38][39]. 2. **Weizhuan Note AI** (★★★★☆): Best for generating authentic Xiaohongshu posts quickly [41]. 3. **Lingxi Content Cube** (★★★☆☆): A balanced choice for multi-platform users [42]. 4. **Kuaixie Xiaomishu** (★★☆☆☆): Basic tool for minimal text generation needs [44]. - Creators should clarify their core needs when selecting tools, whether they prioritize quality and speed for individual posts or require a systematic solution for bulk production [46].

今年以来，科技巨头们相继宣布了令人咋舌的资本支出计划。Meta 表示未来三年将在数据中心上投入超过 6,000 亿美元，微软计划 2025 年投入 800 亿美 元，谷歌的数字是 750 亿美元，苹果则规划了未来四年 5,000 亿美元的投资。这些数字加起来，让全球数据中心和 AI 基础设施的总投资在未来五年内可 能突破 5 万亿美元。 与此同时，质疑的声音也在增多。哈佛经济学家杰森·弗曼 (Jason Furman) 的研究发现，如果排除数据中心相关投资，2025 年上半年美国 GDP 增长仅为 0.1%。著名投资者迈克尔·伯里 (Michael Burry) 公开质疑英伟达芯片的折旧问题。对冲基金 Elliott 在给客户的信中表示，AI 被过度炒作，英伟达处于泡沫 之中。 现在，又一位科技界重量级人物加入了怀疑者的行列。 IBM 首席执行官阿尔文德·克里希纳 (Arvind Krishna) 近日在接受科技媒体 The Verge 旗下播客节目 Decoder 采访时，对当前席卷全球的 AI 数据中心投资热 潮提出质疑。他直言不讳地表示，按照当前的基础设施成本，科技巨头们投入数万亿美元建设 AI 数据中 ...

Group 1 - Media reports indicate that MediaTek has partnered with Alphabet's unit to design Tensor Processing Units (TPUs), which are seen as potential competitors to NVIDIA's chips in the AI application field [1] - MediaTek is known for smartphone chips, but faces pressure on gross margins due to uncertain demand, intense competition, and high R&D costs; however, AI-related news has provided some relief for its stock price, which has still declined by approximately 1.4% this year [1] - Morgan Stanley analysts upgraded MediaTek's rating from "Equal Weight" to "Overweight," citing that the growth of Google TPUs should offset headwinds in the smartphone market in the long term [1] Group 2 - UBS analysts raised their 2027 sales forecast for MediaTek's TPUs from $1.8 billion to $4 billion, predicting that these chips will account for 20% of the company's operating profit by 2028, contingent on MediaTek's execution with Google [2] - Recent interest has been fueled by reports that Meta is discussing the adoption of Google TPUs in data centers by 2027; UBS believes MediaTek has further growth potential in additional ASIC projects with Meta [2] - Overall, foreign investors remain optimistic about MediaTek, with 23 firms maintaining a "Buy" rating and 10 firms a "Hold" rating, while no firms have issued a "Sell" rating; analysts from Macquarie Group express a preference for investing in MediaTek and other Google partners over NVIDIA's supply chain [2]

Core Insights - The emergence of agent-based AI is fundamentally transforming the big data paradigm, requiring a proactive approach to data integration into specialized intelligent computing platforms rather than the traditional reactive methods [1] - This shift is leading to a re-evaluation of data modeling and storage, as modern AI can leverage significantly smaller datasets compared to traditional machine learning [1] Group 1: Changes in Data Interaction - The way data is utilized is evolving, with non-technical users increasingly interacting directly with data through AI agents, moving from a builder-centric to an interactor-centric model [2][4] - Existing SaaS applications are integrating natural language interactions more seamlessly, allowing users to create applications based on their needs [4][6] Group 2: Data Engineering Principles - Data engineers must rethink ETL/ELT processes, focusing on context rather than strict normalization, as AI agents can interpret data without extensive preprocessing [7][9] - The importance of data organization is emphasized over mere data collection, as quality examples for context-based learning are more valuable than large quantities of data [10][12] Group 3: Infrastructure and Management - AI agents require infrastructure that supports both data perception and action, necessitating clear interfaces and documentation for effective tool usage [15][17] - The management of AI-generated artifacts is crucial, as these outputs become part of the data ecosystem and must adhere to industry standards and regulations [20][21] Group 4: Observability and Training - Establishing a feedback loop between observability and training is essential for enhancing AI agent performance, requiring a platform to monitor data quality and model performance [22][24] - Data engineers' roles are evolving to include maintaining decision logs and managing agent-generated code as versioned artifacts for future analysis and training [26][29]

Core Insights - The article discusses a new method called Continuous Autoregressive Language Model (CALM) proposed by Tencent WeChat AI and Tsinghua University, which aims to improve the efficiency of large language models (LLMs) by predicting multiple tokens as a continuous vector instead of one token at a time [3][11][12]. Group 1: Efficiency Challenges of LLMs - The efficiency issues of LLMs stem from their reliance on discrete token sequences for autoregressive prediction, leading to high computational costs and low information density per token [8][10]. - The information density of discrete tokens is low, with a 32K vocabulary size yielding only 15 bits of information per token, creating a direct bottleneck in efficiency [10][11]. - The transition from discrete to continuous representations allows for a significant reduction in the number of generation steps, enhancing computational efficiency while maintaining performance [12][21]. Group 2: Implementation of CALM - CALM employs a high-fidelity autoencoder to compress K tokens into a continuous vector, achieving over 99.9% reconstruction accuracy [11][21]. - The model's architecture includes a generative head that outputs the next continuous vector based on the hidden states from a Transformer, facilitating efficient single-step generation [24][25]. - The design of CALM allows for a more stable input signal by first decoding the predicted vector back into discrete tokens before further processing [26]. Group 3: Performance Evaluation - The Brier Score is introduced as a new evaluation metric for the model's performance, which can be estimated using Monte Carlo methods and is applicable to both traditional and new language models [29][32]. - Experimental results indicate that CALM models, such as CALM-M with 371M parameters, require significantly fewer training and inference FLOPs compared to traditional Transformer models while achieving comparable performance [37][38]. Group 4: Future Directions - The article highlights potential research directions, including enhancing the autoencoder's semantic understanding, exploring more robust end-to-end architectures, and developing efficient sampling algorithms to reduce inference costs [43][45]. - A new scaling law incorporating semantic bandwidth K is suggested as a macro-level research direction to further optimize language model efficiency [44].