Group 1: Company Overview - The company specializes in the research, production, and sales of IoT intelligent products, focusing on communication, positioning, and AI technologies [1] - In 2024, the company achieved a revenue of CNY 1.402 billion, a year-on-year increase of 24.85%, and a net profit of CNY 176 million, an increase of 0.81% [1] - In Q1 2025, the company reported a revenue of CNY 348 million, a 40.28% increase year-on-year, and a net profit of CNY 40 million, up 14.42% [2] Group 2: Product Development and Technology - The company continuously launches new products based on core technologies such as communication, positioning, and AI, which serve as the foundation for expanding into various IoT application scenarios [2][3] - The company has developed a modular and standardized cloud management platform to meet diverse industry needs, enhancing product performance and reducing production costs [3] - In 2024, revenue from other smart hardware reached CNY 142 million, a growth of 21.70% compared to 2023 [3] Group 3: Product Applications - The company offers over twenty types of IoT products, including electronic student ID cards, smart wearable watches, portable mistake printers, and smart security cameras, currently in market development and incubation stages [4] - The electronic student ID card focuses on "safe campus" applications, providing features like student tracking and SOS alerts [5] - The smart wearable watch targets "elderly care" and "safe campus" scenarios, boasting a battery life of over 12 days on a single charge [5] Group 4: Market Impact and Risks - The company’s smart vehicle terminal products are primarily sold in Africa, Southeast Asia, and West Asia, while smart payment hardware is mainly distributed in Southeast Asia [5] - Changes in U.S. tariff policies have minimal impact on the company, as the customer bears the costs associated with tariffs under the EXW delivery model [5] - The company advises investors to make rational decisions and be aware of investment risks related to industry forecasts and strategic planning [5]

编者荐语： 日经中文网与36氪展开内容交换合作，精选36氪的精彩独家财经、科技、企业资讯，与读者分享。 以下文章来源于36氪Pro ，作者李安琪 李勤 36氪Pro . 36氪旗下官方账号。深度、前瞻，为1%的人捕捉商业先机。 车企的AI辅助驾驶人才饥渴症。 文 ｜ 李安琪 编辑 ｜ 李勤 封面来源 ｜ 日经中文网 入职新公司第一天，张杨（化名）被要求"吐露"上家公司的辅助驾驶算法与代码。因没有积极配合，张杨没在新公司待多久就离 开了。 张杨的前东家是理想汽车，近年因迅速落地辅助驾驶而被行业关注，成为同行重点"探秘"的对象。 辅助驾驶的技术演化在持续喷发。从传统的基于规则的方案转向"端到端"模型路线后，车企的人才画像需求发生了极大变化，中 国车企像互联网大厂与AI公司一样渴求AI人才。 行业竞争激烈而持续。车企内部，团队赛马、立军令状、集体封闭式开发、"做不出来就换人"等，已经成为辅助驾驶部门的常 态。在高压的交付压力下，挖角高端人才、解密头部公司的技术，成为企业的一些"水下动作"。 尤其今年以来，辅助驾驶第一梯队公司的人才遭到了哄抢。有猎头人士告诉36氪，在端到端、AI大模型这波浪潮中，华为、理 想、Mom ...

Core Insights - The article discusses the shift from traditional Search Engine Optimization (SEO) to Generative Engine Optimization (GEO) as a new strategy for enhancing brand marketing effectiveness in the age of AI-driven information retrieval [1][2] - A16z emphasizes that the focus of brand competition will transition from manipulating search rankings to being actively referenced by AI models, indicating that brand success will hinge on being "remembered" by AI rather than just being found through search engines [1][2] Industry Overview - For over two decades, SEO has been the gold standard for online exposure, leading to the emergence of various tools and services aimed at optimizing digital marketing [2] - By 2025, the landscape of search is expected to change dramatically, with traditional search engines being replaced by large language model (LLM) platforms, challenging Google's dominance in the search market [2] - The SEO market, valued at over $80 billion, is beginning to wane as a new paradigm driven by language models emerges, marking the onset of the GEO era [2] Transition from SEO to GEO - Traditional search relied on "links," while GEO relies on "language," shifting the definition of visibility from high rankings in search results to being integrated into AI-generated answers [3][6] - The format of search answers is evolving, with AI-native searches becoming more decentralized across platforms like Instagram, Amazon, and Siri, leading to longer queries and extended session durations [3][5] Differences Between SEO and GEO - GEO differs fundamentally from traditional SEO in content optimization logic, requiring content to have clear structure and semantic depth for effective extraction by generative language models [6][11] - The business models and incentives of traditional search engines and language models differ significantly, impacting how content is referenced and monetized [7][11] New Metrics for Brand Visibility - The core metrics for brand communication are shifting from click-through rates (CTR) to citation rates, which measure how often brand content is referenced in AI-generated answers [11][12] - Emerging platforms like Profound, Goodie, and Daydream are utilizing AI analysis to help brands track their presence in generative AI responses, focusing on frequency and sentiment of mentions [11][12] Tools and Strategies in GEO - Companies are developing tools to monitor brand mentions in AI outputs, with platforms like Ahrefs and Semrush adapting to the GEO landscape [12][15] - GEO represents a paradigm shift in brand marketing strategies, emphasizing how brands are "written into" AI knowledge layers as a competitive advantage [12][15] Future of GEO - The future of GEO platforms will involve not only brand perception analysis but also the ability to generate AI-friendly marketing content and respond to changes in model behavior [17][18] - The rapid migration of budgets towards LLMs and GEO platforms indicates a significant shift in marketing strategies, with brands needing to ensure they are remembered by AI before user searches occur [18]

如果您希望可以时常见面，欢迎标星收藏哦~ 来源：内容 编译自 ieee 。 对于那些喜欢支持弱者的人来说，最新的MLPerf基准测试结果可能会令人失望：Nvidia 的GPU仍 然占据主导地位再次。这包括在最新、最苛刻的基准测试中，对Llama 3.1 403B 大型语言模型进 行预训练时，所展现出的顶级性能。即便如此，基于最新AMD GPU MI325X 构建的计算机在最 流行的 LLM 微调基准测试中，其性能与Blackwell 的前代产品 Nvidia H200 相当。这表明 AMD 落后了Nvidia一代。 MLPerf训练是MLCommons联盟举办的机器学习竞赛之一。"AI 性能有时就像狂野西部。MLPerf 致力于打破这种混乱，" Nvidia 加速计算产品总监Dave Salvator表示。"这并非易事。" 本次比赛包含六个基准测试，每个基准测试分别针对一个与行业相关的机器学习任务。这些基准测 试包括内容推荐、大型语言模型预训练、大型语言模型微调、机器视觉应用的目标检测、图像生成 以及用于欺诈检测和药物研发 等应用的图节点分类。 大型语言模型预训练任务是最耗费资源的，而本轮更新后更是资源密集。 ...

《硅谷101》创始人泓君邀请了Energent.ai联合创始人Kimi Kong和HeyRevia创始人Shaun Wei，一起和两 位前Google的技术专家聊聊Gemini模型登顶背后的底层逻辑。 以下是这次对话内容的精选: 一、Gemini2.5崛起背后的底层逻辑 泓君：谷歌此次发布的Gemini 2.5 Pro，在当前各项评测中的数据都是所有大模型中最好的，Kimi你可 以分析一下它是如何做到的吗？ 从去年在大会前夜被OpenAI的4o模型"精准狙击"，到今年Gemini 2.5 Pro全面霸榜。短短一年时间， Gemini是如何完成从追赶者到领跑者的逆转？ Kimi：我已经离开DeepMind快一年时间了，也不太清楚我的前同事们在这一年中又做了哪些新的创 新。但大语言模型训练根本的步骤是不变的，包括以下三点：Pre-training（预训练）、SFT（Supervised Fine-tuning，监督微调）和利用RLHF（基于人类反馈的强化学习）技术做的Alignment（对齐）。 大概在去年的NeurIPS（神经信息处理系统大会）上，业内已经普遍承认，公开网络数据基本都已经抓 完了，就像化石燃料已 ...

一位眼科医生通过用人工晶体置换混浊的晶状体（白内障），能在半小时内改变一个人的生活。在许多医疗领 域，从业者可以用明确的指标（如血液检测、骨骼扫描和其他生理指标）来评估干预的效果。 在心理健康护理领域，用于诊断和提供治疗方法的数据通常主要由逐步积累的文本构成。虽然标准化问卷和评分 量表提供了一些可量化的指标，但它们仍然依赖于自我报告或临床医生的判断。这使得叙述中可能出现许多漏 洞、误解和认知偏差。一位患者可能因为各种原因无法准确地在日志中记录自己的情绪、活动或行为，而临床医 生在有限的环境和较短的时间内与患者互动时，可能会对患者的病情做出错误的判断。 数字技术可以通过提供更客观和持续的数据收集方法来帮助缓解这些问题。现在可以利用智能手机和可穿戴设备 实现对行为的被动监测。能够主动提示用户情绪和行为状态的心理健康应用程序可以帮助人们实现更一致的自我 监测。利用AI分析地理定位数据、短信发送频率和通话时长，可以预测抑郁症或双相情感障碍的发作。 大语言模型还可以分析大量的治疗会话记录，以更好地了解在不同情境下哪种干预措施效果最好，以及哪些咨询 师行为可以带来不错的治疗效果。例如，2024年1月，隶属于质量保证和临床 ...

Core Viewpoint - The article discusses the considerations for choosing between local and cloud-based AI models, emphasizing the importance of computational requirements and privacy needs when selecting hardware for AI applications [5][6][17]. Group 1: AI Model Deployment - Local deployment of AI models is suitable for applications requiring high computational power and privacy, particularly when handling sensitive data [16][17]. - The article outlines the parameters of AI models, indicating that a model with 1 billion parameters requires approximately 4GB of memory for full precision, while half-precision models can reduce this requirement significantly [11][14]. - For local deployment, models with fewer than 14 billion parameters are generally manageable, while larger models may necessitate high-end GPUs like the RTX 4090 or 5090 [14][19]. Group 2: Hardware Recommendations - The article provides recommendations for laptops suitable for AI applications across different price ranges, highlighting models with specific GPU configurations [26][29][31]. - For a budget of around 5000 yuan, the Mechrevo Aurora X with a 5060 GPU is suggested as a high-value option [26]. - In the 6000 yuan range, the HP Shadow Elf 11 with a 5060 GPU is recommended, while the 7000 yuan range includes upgraded versions of the same model [29][31]. Group 3: Privacy and Security - Local deployment is emphasized as a necessity for applications involving sensitive data, such as business secrets or medical information, to prevent data leaks [17][18]. - The article argues that using local models ensures that all computations are performed on the user's hardware, eliminating the risk of data exposure to third-party services [16][17].

本文作者分别来自新加坡国立大学、北京大学与清华大学。第一作者王宇豪与共同第一作者屈文杰来自新加坡国立大学，研究方向聚焦于大语言模型中的安 全与隐私风险。共同通讯作者为北京大学翟胜方博士，指导教师为新加坡国立大学张嘉恒助理教授。 本研究聚焦于当前广泛应用的 RAG (Retrieval-Augmented Generation) 系统，提出了一种全新的黑盒攻击方法： 隐式知识提取攻击 (IKEA) 。不同于以 往依赖提示注入 (Prompt Injection) 或越狱操作 (Jailbreak) 的 RAG 提取攻击手段， IKEA 不依赖任何异常指令，完全通过自然、常规的查询，即可高效 引导系统暴露其知识库中的私有信息。 在基于多个真实数据集与真实防御场景下的评估中，IKEA 展现出超过 91% 的提取效率与 96% 的攻击成功率，远超现有攻击基线；此外，本文通过多项 实验证实了隐式提取的 RAG 数据的有效性。本研究揭示了 RAG 系统在表面「无异常」交互下潜在的严重隐私风险。 论文题目：Silent Leaks: Implicit Knowledge Extraction Attack on RAG S ...

GPT 系列模型的记忆容量约为每个参数 3.6 比特。 语言模型到底能记住多少信息？Meta、DeepMind、康奈尔大学和英伟达的一项测量结果显示： 每个 参数大 约 3.6 比特 。一旦达到这个极限，它们就会停止记忆 并开始泛化。 长期以来，记忆与泛化之间的模糊性一直困扰着对模型能力和风险的评估，即区分其输出究竟源于对训练数据的「记忆」 （对其训练数据分布的编码程度） ，还 是对潜在模式的「泛化」理解 （将理解扩展到未见过的新输入）。 这种不确定性阻碍了在模型训练、安全、可靠性和关键应用部署方面的针对性改进。 机器之心报道 编辑：+0、张倩 这就好比我们想知道一个学生考试得了高分，是因为他真的理解了知识点（泛化），能够举一反三，还是仅仅因为他把教科书上的例题和答案都背下来了（记 忆）。 基于此，研究团队提出了一种新方法，用于估计一个模型对某个数据点的「了解」程度，并利用该方法来衡量现代语言模型的容量。 研究团队从形式上将记忆分为两个组成部分： 通过消除泛化部分，可以计算出给定模型的总记忆量，从而估计出模型容量：测量结果估计， GPT 系列模型的容量约为每个参数 3.6 比特 。 研究团队训练了数百个参数量 ...

强化学习（RL）到底是语言模型能力进化的「发动机」，还是只是更努力地背题、换个方式答题？这个问题，学界争论已久：RL 真能让模型学会新的推理 技能吗，还是只是提高了已有知识的调用效率？ 过去的研究多数持悲观态度：认为 RL 带来的收益非常有限，有时甚至会让模型「同质化」加重，失去多样性。然而，来自英伟达的这项研究指出，造成这 一现象的根本原因在于：数学、编程等任务在 base model 的训练数据中被过度呈现，以及 RL 训练步数不足。 论文题目：ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models 链接：https://arxiv.org/pdf/2505.24864 ProRL 来了！长期训练 = 推理能力质变！ 由 NVIDIA 团队提出的 ProRL（Prolonged Reinforcement Learning）框架，将 RL 训练步数从传统的几百步大幅提升至 2000 步以上，释放了小模型潜 藏的巨大潜力。结果令人震惊： KL 正则化 + 周期性策略重置 这一突 ...