• 从企业知识库中检索与用户查询相关的内容： 系统首先根据用户查询，将自然语言问题转化为检索指令；通过嵌入向量检索（ embedding search ）、关 键词检索（ keyword search ）或混合检索（ hybrid retrieval ），从企业知识库中迅速筛选出与问题最相关的一组文档片段。这一过程不仅比传统全文检索 更灵活高效，还能处理语义相似但措辞不同的问题。 • 动态构建上下文提示（ prompt ） ：检索回来的内容经过筛选、重排序，并被自动组织成新的提示上下文，与用户原始提问一起拼接，形成 " 携带知识 " 的复合输入。这一动态构建的上下文，是 RAG 系统生成准确回答的基础。 • 模型生成推理式回答 ：最终，大语言模型（ LLM ）在接受到包含检索知识的新提示后，基于外部知识和自身推理能力，生成针对性的自然语言回答。 简单来说， RAG 就像是给大模型装了一个随身小图书馆，每次回答问题前，它可以先去翻翻书，再动笔答题。 这大幅提升了回答的准确性和时效性，使 RAG 被视为现代 AI 应用栈中的重要架构 —— 连接通用大模型与面向企业的高效 AI 应用的的关键纽带。 Z Highli ...

Core Insights - The article discusses the performance decline of large language models (LLMs) as the input context length increases, highlighting that the decline is not uniform but occurs at specific token lengths [10][21][44] - A recent study by the Chroma team tested 18 mainstream LLMs, revealing that models like GPT-4.1 and Claude Sonnet 4 experience significant accuracy drops when processing longer inputs [8][9][19] Group 1: Performance Decline - As input length increases, model performance deteriorates, with a notable drop around 10,000 tokens, where accuracy can fall to approximately 50% [4][21] - Different models exhibit varying thresholds for performance decline, with some models losing accuracy earlier than others [6][7][19] - The study indicates that semantic similarity between the "needle" (target information) and the "problem" significantly affects performance, with lower similarity leading to greater declines [19][21] Group 2: Experimental Findings - Four controlled experiments were conducted to assess the impact of input length on model performance, focusing on factors like semantic similarity, interference information, and text structure [17][35][41] - The first experiment showed that as input length increased, models struggled more with low semantic similarity, leading to a sharper performance drop [19][21] - The second experiment demonstrated that the presence of interference items significantly reduced model accuracy, with multiple interference items causing a 30%-50% drop compared to baseline performance [26][28] Group 3: Structural Impact - The structure of the background text (haystack) also plays a crucial role in model performance, with coherent structures leading to more significant declines in accuracy compared to disordered structures [40][42] - The experiments revealed that most models performed worse with coherent structures as input length increased, while performance decline was less severe with disordered structures [41][44] - The findings suggest that LLMs face challenges in processing complex logical structures in long texts, indicating a need for improved handling of such inputs [41][44] Group 4: Implications and Future Directions - The results highlight the limitations of current LLMs in managing long-context tasks, prompting suggestions for clearer instructions and context management strategies [44] - Chroma, the team behind the research, aims to address these challenges by developing open-source tools to enhance LLM applications in processing long texts [45][48]

近日，云知声联合创始人&董事长&CTO梁家恩博士正式受聘为广西人工智能战略咨询专家委员会委员。这一荣誉不仅是对他在人工智能领域多年深耕所获 成就的认可，更彰显了云知声在AI技术创新与产业落地领域的领先地位，以及企业助力广西区域发展的责任与担当。 2023年3月，云知声与南宁轨道交通合资成立"广西桂云通科技有限公司"，开启了双方合作的全新篇章；同年10月，基于山海大模型，云知声在南宁东站成 功打造一体化换乘引导系统，极大提升了乘客的出行体验；此后，云知声又与自治区自然资源厅展开合作，利用AI技术助力自然资源调查监测，为守护绿 水青山提供了创新方式。 2025年2月，广西壮族自治区党委常委、南宁市委书记农生文会见云知声董事长梁家恩一行，双方就深化合作展开深入交流；同月，《广西日报》特邀梁家 恩博士进行专访并刊发报道，引发社会各界广泛关注。4月，云知声与南宁市人民政府正式签署"云知声东盟总部项目"合作协议，为双方合作注入新动能；6 月，公司与南宁轨道进一步升级合作，对"轨道・慧生活"小程序进行优化迭代；7月，云知声再拓合作领域，与广西卫健委正式签署战略合作协议，将携手 在医疗健康领域开展更深层次的AI应用探索。 此 ...

Group 1: Core Concepts of Machine Emotion - Machine emotion refers to the external manifestation of human-like emotions by AI systems, relying on emotional intelligence and emotional computing as the main technological pathway [3][5][6] - The concept of machine emotion is interdisciplinary, involving cognitive science, emotional philosophy, psychology, computer science, and sociology [5][6] - Emotional intelligence is the foundational capability for machine emotion, encompassing emotional recognition, expression, experience, and control [6][9] Group 2: Construction and Characteristics of Machine Emotion - Machine emotion is characterized by its constructiveness, mimetic nature, embodiment, and computational aspects, emphasizing the generation logic and operational mechanisms of machine emotion [5][12] - The realization of machine emotion depends on emotional computing, which captures human emotional data through various sensors and builds a personalized computational system for emotional understanding and response [7][12] - Machine emotion is limited in its ability to mimic human emotional recognition and expression but lacks genuine emotional experience and control [12][16] Group 3: Human-Machine Emotional Interaction - Human-machine emotion is fundamentally a projection of human emotions onto machines, lacking true intersubjectivity and emotional sharing [13][15] - The construction of human-machine emotion is influenced by psychological mechanisms, such as anthropomorphism and social cultural factors, which shape human emotional responses to machines [15][16] - The emotional interaction between humans and machines can lead to risks such as emotional deception, emotional monitoring, emotional degradation, and emotional manipulation [17][18] Group 4: Ethical Considerations and Development of Machine Emotion - To mitigate risks associated with machine emotion, it is essential to construct machine emotions that serve human needs and enhance interaction [18][19] - The development of machine emotion should adhere to a limited approach, ensuring the appropriateness of emotional capabilities and avoiding exaggerated claims [19][20] - Transparency, authenticity, and rigor should guide the promotion of machine emotions, ensuring users are aware of the simulated nature of emotional responses [20][21]

Core Viewpoint - The article discusses the introduction of the World-Aware Planning Narrative Enhancement (WAP) framework, which significantly improves the performance of large vision-language models (LVLMs) in embodied planning tasks by integrating four-dimensional cognitive narratives and closed-loop observation methods [3][16]. Group 1: Introduction - LVLMs are becoming central in embodied planning, but existing methods often rely on environment-agnostic imitation learning, leading to poor performance in unfamiliar scenarios [3]. - WAP aims to enhance model capabilities by injecting four-dimensional cognitive narratives (visual, spatial, functional, syntactic) into the data layer, allowing models to better understand their environment before reasoning [3][4]. Group 2: Technical Methodology - WAP's main distinction is its explicit binding of instructions to environmental context, relying solely on visual closed-loop feedback without privileged information [6]. - The framework employs a three-stage curriculum learning approach, using only RGB observations and no privileged feedback to train the model [12]. Group 3: Experimental Results - The Qwen2.5-VL model achieved a success rate increase from 2% to 62.7% (+60.7 percentage points) on the EB-ALFRED benchmark, surpassing models like GPT-4o and Claude-3.5 [4][14]. - The model demonstrated a long-range task success rate improvement from 0% to 70%, indicating the effectiveness of the WAP framework in complex planning scenarios [14]. - A case study illustrated WAP's ability to decompose complex instructions into manageable steps, showcasing its superiority over baseline models that failed to consider implicit conditions [15]. Group 4: Conclusion and Future Work - WAP successfully integrates "world knowledge" into data and reasoning chains, allowing small-scale open-source LVLMs to outperform commercial models in pure visual closed-loop settings [16]. - Future work includes enhancing continuous control, expanding to dynamic industrial/outdoor environments, and exploring self-supervised narrative evolution for iterative data-model improvement [17].

2025.07. 17 本文字数：3600，阅读时长大约6分钟 作者 | 一财评论 近年来，美国专利商标局、美国版权局发布了一系列指南或研究报告，明确限制人工智能（AI）辅助创新 的专利和版权保护。例如，美国专利商标局要求人类对AI输出可能形成的发明做出"显著贡献"方能获得专 利保护；美国版权局则指出，当AI技术决定其输出的表达性元素时，生成的材料不受版权保护。这使得商 业秘密法成为企业保护AI相关创新的重要工具，因为它既不要求发明人或作者身份，也无需政府批准或注 册。 商业秘密法凭借其高度包容性和广泛适用范围，为AI技术创新提供了强有力的保护，能够覆盖专利和版权 无法涵盖的敏感信息，例如用于训练AI系统的数据或模型权重。因此，业内有人戏称，用商业秘密保护AI 技术堪称"天作之合"。AI开发者和用户可以借助商业秘密法保护与AI技术相关的核心信息，同时规避传统 知识产权保护所面临的时间限制和程序繁琐等约束。然而，将商业秘密法应用于AI技术的保护也面临诸多 挑战，包括技术透明性不足、法律手段滥用以及公共利益受损等。 AI技术的快速发展带来了"饥饿"与"喂养"的矛盾：一方面，AI对数据、算法和算力有着永无止境的需求 ...

值得注意的是，Ani还拥有"NSFW"模式，即包含裸露、暴力或色情等不适合在工作场合浏览的内容。这也引发了外界对未成年人接触不当内容的担忧。 AI情感陪伴应用可以说是这波AI大模型应用浪潮中最火热的赛道之一。 作者 | 黄昱 编辑 | 王小娟 AI情感陪伴赛道又迎来了一名重量级玩家。 近日，埃隆・马斯克（Elon Musk）旗下人工智能公司xAI开发的人工智能聊天机器人Grok，推出了基于Grok 4大模型的"伴侣"（companions）功能。该功能 旨在提供更具沉浸感和情感参与度的AI互动体验。 马斯克对这一功能十分重视，亲自下场为其摇旗呐喊，并在社交平台X上将这一消息置顶。对于Grok而言，"伴侣"功能的推出是其在AI竞争中寻求差异化、 深化用户关系，并进一步拓展商业模式的重要举措。 Grok 首批上线的两名"伴侣"角色是二次元哥特风女孩形象Ani和卡通风格小熊猫"坏鲁迪"（Bad Rudy），都拥有3D动画形象。用户可以通过语音和文字与角 色进行互动，角色会以其独特的个性和预期作出回应。 目前来看，Ani是主推角色，但Grok这项"伴侣"服务如今仅向每月支付30美元的SuperGrok订阅服务用户开放 ...

Market Overview - On July 16, the Shanghai Composite Index fell by 0.03%, the Shenzhen Component Index decreased by 0.22%, and the ChiNext Index also dropped by 0.22%[3] - The total trading volume of the Shanghai and Shenzhen stock markets exceeded 1.44 trillion, showing a significant decrease compared to the previous period[1] Sector Performance - Leading sectors included chemical pharmaceuticals, automotive parts, oil, automotive services, education, and diversified finance, while insurance, steel, energy metals, banking, electronic components, and non-ferrous metals showed notable adjustments[1] - A total of 3,208 stocks rose, while 1,809 stocks declined, indicating a predominance of gainers in the market[1] Investment Strategy - The current market adjustment is viewed as an opportunity rather than a risk, with expectations for a new round of strong upward movement after technical indicators stabilize[1] - Investors are encouraged to focus on sectors such as digital currency, semiconductors, artificial intelligence, biomedicine, and new energy vehicles[1] Fund Flow - On July 16, net inflows into the Shanghai Stock Exchange were 38.27 billion, while the Shenzhen Stock Exchange saw net inflows of 8.14 billion[4] - The top three sectors for net inflows were automotive parts, chemical pharmaceuticals, and general equipment, while the sectors with the highest outflows included components, securities, and industrial metals[4] Industry Developments - The Ministry of Industry and Information Technology plans to implement stricter technical standards for mobile power sources, which may lead to a reshuffle in the domestic market, benefiting related listed companies[2] - The AI industry is shifting focus from infrastructure to downstream applications, with opportunities emerging in smart manufacturing, smart education, and smart healthcare[2] Fund Dynamics - A total of 136 funds have been liquidated this year, with equity funds making up 65% of the total, indicating a trend towards normalization in fund closures[13] - Fund managers are increasing their stock positions significantly, with some new funds exceeding 90% in stock allocation, reflecting optimism about future market conditions[15]

2022年11月，OpenAI基于GPT 3.5推出了ChatGPT，短时间内便吸引了数千万用户，使大型语言模型 （LLM）正式走进公众视野，也将GPT架构推上了主流AI架构的宝座。随着ChatGPT打响大模型时代的 第一枪，各大厂商纷纷涌入这一赛道。由于ChatGPT无法直接接入国内，一些小作坊看到了套壳牟利的 机会，一时间，各种山寨ChatGPT在互联网上泛滥。 这些套壳行为最初往往不涉及任何二次开发，开发者只是简单地对API进行包装并出售。然而，随着监 管的加强，这种低劣的套壳手段很快就被打击。例如，"ChatGPT在线"公众号因涉嫌仿冒ChatGPT被罚 款6万元，成为首例"ChatGPT套壳"行政处罚案例。 尽管如此，套壳行为并未绝迹。在2023年发布的一些模型中，仍时常出现"GPT味"的回复，引发套壳质 疑。例如，讯飞星火大模型曾因涉嫌套壳ChatGPT而引发关注。对此，一些企业解释称，这可能是由于 训练数据中混入了大量ChatGPT生成的内容，导致模型"身份混淆"。另一种可能是，模型研发团队在微 调训练过程中主动使用了通过ChatGPT等OpenAI旗下模型构造的数据集，即所谓的"数据蒸馏"。 ...

Core Viewpoint - The article discusses the inherent biases present in AI systems, particularly large language models (LLMs), and questions the trustworthiness of their outputs in reflecting a neutral worldview [1][2]. Group 1: AI and Cultural Bias - AI models are found to propagate stereotypes across cultures, reflecting biases such as gender discrimination and cultural prejudices [2][3]. - The SHADES project, led by Hugging Face, identified over 300 global stereotypes and tested various language models, revealing that these models reproduce biases not only in English but also in languages like Arabic, Spanish, and Hindi [2][3]. - Visual biases are evident in image generation models, which often depict stereotypical images based on cultural contexts, reinforcing narrow perceptions of different cultures [2][3]. Group 2: Discrimination Against Low-Resource Languages - AI systems exhibit "invisible discrimination" against low-resource languages, performing poorly compared to high-resource languages [4][5]. - Research indicates that the majority of training data is centered around English and Western cultures, leading to a lack of understanding of non-mainstream languages and cultures [4][5]. - The "curse of multilinguality" phenomenon highlights the challenges AI faces in accurately representing low-resource languages, resulting in biased outputs [4]. Group 3: Addressing AI Bias - Global research institutions and companies are proposing systematic approaches to tackle cultural biases in AI, including investments in low-resource languages and the creation of local language corpora [6]. - The SHADES dataset has become a crucial tool for identifying and correcting cultural biases in AI models, helping to optimize training data and algorithms [6]. - Regulatory frameworks, such as the EU's AI Act, emphasize the need for compliance assessments of high-risk AI systems to ensure non-discrimination and transparency [6]. Group 4: The Nature of AI - AI is described as a "mirror" that reflects the biases and values inputted by humans, suggesting that its worldview is not autonomously generated but rather shaped by human perspectives [7].