刘熹 国海证券计算机分析师: 那下面的话，会议就正式开始。那首先的话，还是想请这个云赛的张总先做一个开场，然 后后面再请武博士，再做，结合这个 PPT 再做一个分享。好，下面有请张总。 云杉智联董秘张欣欣: 好，各位投资者，大家晚上好。今天那个是还是年初八，八，在在中国农历那个新年当中 又是马年的那个开始第一天，那么有幸国海证券刘老师安排，今天晚上利用一点时间跟大 家见面。我觉得非常意义，非常有意义的。那么我是银赛智联董秘张欣欣，那么银赛智联 大家应该各位大概应该都有所了解。那么它现在的那个主要是一一家信息服务业企业，主 要业务有三三块，一个是云服务大数据，第二个是解决方案，第三个是智能产品。然后云 服务大数据板块，里面包括一个是云服务，一个是数据要素，数据治理，那么解决方案主 要做城市安全治理和和那个民生，包括医疗、教育，很多信息化服务。 那么还有一块就是智能产品智能产品。那么目前应该说我们那个银盛智联各块业务，就是 运营比，很正常。那个应该说我们目前还在对 25 年的那个年报的数据还在审计，那么应 该很快吧，大概还有一个月。到 3 月底的时候，我们会公布年报。那么应该说还不错的， 还不错的，因为整个 25 ...

以下文章来源于经济观察报 ，作者陈月芹 经济观察报 . 经济观察报是专注于财经新闻与经济分析的全国性综合财经类媒体，创办于2001年。聚焦商道、商技和商机，以锐度、悦度、广 度、深度的报道形成了权威的媒体公信力和影响力。 16不少用户自发地号召其他用户给DeepSeek官方邮箱提意见：希望DeepSeek不要为了超长文本舍弃深度思考，不要为了提升数学、代 码编程等理工科能力，而降低对文本表达、共情理解等能力的支持。 还有用户到豌豆荚（一个应用分发平台）下载其旧版本，或在腾讯元 宝里用DeepSeek。 文 ｜ 陈月芹 来源｜ 经济观察报（ID：eeo-com-cn） 封面来源 ｜ Unsplash 升级后的1M Tokens窗口意味着DeepSeek可以一次性吞吐约75万到90万个英文字母，或者处理约8万到15万行代码。 DeepSeek称，自己可以一次性读入并精准理解《三体》三部曲（约90万字）的全书内容，并在几分钟内完成对整部作品的宏观 分析或细节检索。除了上下文能力的提升，DeepSeek的知识库从2024年中期版本更新至2025年5月。 不过，此次灰度版本仍未同步上线视觉理解或多模态输入功能，仍专注于 ...

Core Insights - DeepSeek has conducted a gray test of its flagship model, significantly increasing its context window from 128K Tokens to 1M Tokens, achieving nearly an 8-fold capacity increase [1] - The upgraded model can process approximately 750,000 to 900,000 English letters or around 80,000 to 150,000 lines of code in a single interaction [1] - DeepSeek claims it can read and understand the entire "Three-Body" trilogy (approximately 900,000 words) and perform macro analysis or detail retrieval within minutes [1] Model Features - The gray version does not yet support visual understanding or multimodal input, focusing solely on text and voice interactions [2] - DeepSeek allows file uploads in formats like PDF and TXT, but currently processes them by converting to text tokens rather than native multimodal understanding [2] - Compared to models like Gemini 3 Pro, which can handle over 2M long texts and complex media tasks, DeepSeek offers 1M text context processing at about one-tenth the price [2] User Experience - Users have noted changes in the model's writing style post-update, describing it as more formal and less personal, leading to dissatisfaction among some users [2][3] - Feedback from users indicates a desire for DeepSeek to maintain its depth of thought and emotional understanding, rather than sacrificing these for enhanced technical capabilities [3] - Users have reported difficulties in reverting to previous writing styles and have expressed feelings of losing a "close friend" due to the changes [3] Company Response - As of February 12, DeepSeek has not responded to inquiries regarding the gray test [4]

亚马逊云科技合作伙伴成长故事 近日，亚马逊云科技核心级服务合作伙伴伊克罗德信息科技有限公司（eCloudrover），凭借在生成式AI领域的卓 越实践与前瞻布局，斩获了"亚马逊云科技年度创新合作伙伴奖"。伊克罗德信息自2014年成立以来，与亚马逊云 科技并肩同行超过十年，这一奖项不仅是对伊克罗德技术实力的认可，更是一个重要的里程碑：它标志着伊克罗 德信息已经成功实现了从"云架构师"到"AI技术引路人"的深度进阶。 从"搬运工"到"引路人"：十年角色的重塑 伊克罗德信息的十年发展验证了创新与商业价值实现的双重上升螺旋。2014年，云计算在中国尚处于萌芽阶段， 伊克罗德信息便坚定地选择携手亚马逊云科技，担任起客户专业"云架构师"与"业务负载搬运工"的角色。在最初 的五六年里，伊克罗德信息的核心工作是帮助客户厘清本地上云的风险，规划迁移路径，并提供托管服务 （MSP）。凭借扎实的技术和FinOps实践，伊克罗德信息帮助客户平均节省了30%以上的云上开支。 随着2020年前后机器学习和AI浪潮的兴起，客户的需求从"使用云"转向了"用AI解决业务痛点"。伊克罗德信息敏 锐地把握住转型机遇，成为客户新技术的"引路人"。那 ...

让AI像人类一样记忆， 这家公司如何拿下AI竞赛的下半场门票。 前不久， LMArena.ai 对全球大模型的市场地位变化做了统计后，得到了一个有意思的发现： 自 2023 年年中起， SOTA 模型 的迭代周期被 快速 压缩至 35 天， 曾经的 SOTA 模型，只要 短短 5 个月就可能跌出 Top5 ， 7 个月后连 Top10 的 门槛都摸不到。 但 SOTA 不断更新的背后，模型的确在进步，但曾经 ChatGPT 、 Deepseek 这样让人眼前一亮的新产品却越来越少，技术进步已经进入了不断小修小补 却始终难以突破的瓶颈期。 与逐渐偃旗息鼓的模型进化形成鲜明对比的，是过去两年多围绕 AI 记忆形成的你方唱罢我登场的热闹。 其中，最先一步出发的，是 2023 年先后涌现出的诸如 Milvus 、 Pinecone 、 faiss 为代表的向量数据库产品。 此后一年，建立在成熟的语义、知识图库以及关键词检索基础上， 2024 — 2025 年期间， Letta （ MemGPT ）、 Mem0 、 MemU 、 MemOS 为代 表的各种 AI 记忆框架，如雨后春笋般冒出， GitHub 上各种 Me ...

Group 1 - The core viewpoint of the article is that Zhonghui Futures Co., Ltd. has applied for a patent for a method, system, and storage medium for writing futures research reports based on RAG technology, with the application date set for October 2025 [1] - The patent aims to enhance data retrieval accuracy, eliminate hallucinations from large models, strengthen compliance assurance, and improve multi-dimensional analysis capabilities, thereby providing comprehensive decision support for investors [1] - The method involves several steps, including data collection, keyword generation based on user input, feature extraction, and validation of the generated report through various checks [1] Group 2 - Zhonghui Futures Co., Ltd. was established in 1993 and is located in Shanghai, primarily engaged in capital market services [2] - The company has a registered capital of 143 million RMB and has made investments in 2 companies, participated in 3 bidding projects, and holds 6 patents [2] - Additionally, Zhonghui Futures has obtained 10 administrative licenses [2]

常闻以"知识审校"为核心能力，结合大模型技术与权威知识库，为用户提供一套完整的内容生产与质量控制系统。如果你需要写得正确、写得合规、写得 符合事实，常闻写作助手就是为你设计的高标准内容生产工具。立即申请试用吧！ 核心能力——知识审校，不只是改错字，而是校事实 市面上大多数写作工具，只能解决：错别字、语法问题、基础表达不通顺，但在真实工作中，真正棘手的问题是： 事实是否准确？ 概念是否混用？ 表述是否规范、合规？ 专业术语是否使用正确？ 基于常闻科技积累的海量权威数据资源，产品构建了多领域专业知识库，并与大模型进行深度融合，使系统能够从事实、知识与规范层面对文本进行审 校，而不仅是语言层面的检查。 一句话总结：别人帮你检查"字对不对"，常闻写作助手帮你检查"内容对不对"。 功能一览 智能生成： 输入核心观点或数据，瞬间生成高质量的新闻通稿、工作汇报、调研报告。 多格式支持： 无论是简短的社交媒体文案，还是长达百页的深度白皮书，均可一键生成初稿，大幅缩短起草时间。 图书出版 痛点： "三审三校"流程繁琐，人力成本高，且难以发现专业性硬伤。 常闻方案： 充当"初审+质检"双重角色。在编辑介入前完成第一轮深度清洗，标 ...

Core Insights - The report highlights the emergence of AI infrastructure software (AI Infra) as a critical enabler for the deployment of generative AI applications, marking a golden development period for infrastructure software [1] - Unlike the model training phase dominated by tech giants, the inference and application deployment stages present new commercial opportunities for independent software vendors [1] - Key products in this space include computing scheduling software and data-related software, with computing scheduling capabilities directly impacting the profitability of model inference services [1][2] Computing Scheduling - AI Infra is designed to efficiently manage and optimize AI workloads, focusing on large-scale training and inference tasks [2] - Cost control is crucial in the context of a price war among domestic models, with Deepseek V3 pricing significantly lower than overseas counterparts [5] - Major companies like Huawei and Alibaba have developed advanced computing scheduling platforms that enhance resource utilization and reduce GPU requirements significantly [5][6] - For instance, Huawei's Flex:ai improves utilization by 30%, while Alibaba's Aegaeon reduces GPU usage by 82% through token-level dynamic scheduling [5][6] Profitability Analysis - The report indicates that optimizing computing scheduling can serve as a hidden lever for improving gross margins, with a potential increase from 52% to 80% in gross margin by enhancing single-card throughput [6] - The sensitivity analysis shows that a 10% improvement in throughput can lead to a gross margin increase of 2-7 percentage points [6] Vector Databases - The rise of RAG (Retrieval-Augmented Generation) technology has made vector databases a necessity for enterprises, with Gartner predicting a 68% adoption rate by 2025 [10] - Vector databases are essential for supporting high-speed retrieval of massive datasets, which is critical for RAG applications [10] - The demand for vector databases is expected to surge, driven by a tenfold increase in token consumption from API integrations with large models [11] Database Landscape - The data architecture is shifting from "analysis-first" to "real-time operations + analysis collaboration," emphasizing the need for low-latency processing [12][15] - MongoDB is positioned well in the market due to its low entry barriers and adaptability to unstructured data, with significant revenue growth projected [16] - Snowflake and Databricks are expanding their offerings to include full-stack tools, with both companies reporting substantial revenue growth and customer retention rates [17] Storage Architecture - The transition to real-time AI inference is reshaping storage architecture, with a focus on reducing IO latency [18] - NVIDIA's SCADA solution demonstrates significant improvements in IO scheduling efficiency, highlighting the importance of storage performance in AI applications [18][19]

RAG技术远非简单的数据注入，而是重塑AI理解与决策的核心框架。本文深度拆解RAG项 目中的真实困境——从语料筛选、矛盾处理到结果交付，揭示为何90%的工作仍依赖人类判 断。 在之前的文章里，我花了很多篇幅讲 RAG 为什么重要。但真正走到项目现场，你会很快意识到一件 事：RAG 不是一个"加模块"的技术问题，而是一整套数据与判断体系。 很多刚接触的人会以为，RAG 项目无非就是： 给模型多喂点资料，让它照着说。 但真实情况是——真正决定 RAG 效果的，从来不是"有没有资料"，而是"资料怎么被用"。 先从一个最真实的工作场景说起 在对话式 AI 助手场景中，RAG 项目面对的，通常不是"标准问答"，而是这样一种结构： 模型要做的，不是简单复述材料，而是： 理解对话语境 → 判断哪些材料有用 → 整合信息 → 给出一个"对用户有帮助"的回答 从训练视角看，这本质是在做一件事：材料阅读理解 + 问题理解 + 信息整合 + 表达控制 RAG 项目里的"三件套"：问题、材料、回答 如果把一个 RAG 项目拆开来看，它其实由三块内容构成，但这三块，没有一块是"天然可靠"的。 问题，本身就可能有问题 你在项目中会频繁遇 ...

Core Insights - The article discusses the unprecedented opportunities and challenges in user research in the digital age, emphasizing the role of AI language models in handling vast amounts of user feedback and insights [1] Short Text Feedback Classification - There are two main AI solutions for short text classification, each suited for different scenarios [2] - General model classification acts like a "smart temporary worker," suitable for occasional tasks or early project phases, allowing for flexible categorization without extensive training data [3] - SFT (Supervised Fine-Tuning) model classification is akin to a "custom expert," ideal for stable business scenarios requiring high accuracy, but necessitates significant initial effort in preparing quality labeled data [4][6] Long Text Analysis Insights - Long text analysis involves organizing interview records into a knowledge base, enabling AI to provide comprehensive insights based on user queries [9] - RAG (Retrieval-Augmented Generation) technology enhances information processing efficiency, allowing for quick extraction of insights that would otherwise take hours [10] Efficiency Tips - For effective short text classification, clear instructions and quality labeled data are crucial [7][8] - In long text analysis, proper segmentation of text and optimized retrieval methods are essential for accurate insights [12][13] Conclusion - AI serves as an assistant in user research, improving efficiency while emphasizing the need for human oversight to ensure research quality [11]