以下文章来源于晚点LatePost ，作者晚点团队 晚点LatePost . 晚一点，好一点 先有一个移动能力的 "通才"，才有更强的智驾系统。 文 丨 张家豪 编辑 丨 程曼祺 全无人驾驶，始终被视作自动驾驶行业皇冠上的明珠，就像登顶珠穆朗玛峰有 19 条路线一样，不同的公司选 择了不同的路线通往无人驾驶的最终目标。 Waymo、小马们选择了基于高精地图的 RoboTaxi 路线，在特定的路线已经实现了 RoboTaxi，为市民提供没有 司机的出行服务；以特斯拉为代表的车企与供应商，则是通过渐进式路线，卖车搭配辅助驾驶方案，收集数据 一步步迭代方案，试图逼近技术极限。 没有人能笃定哪条路线一定能成功登顶，也还有不同的公司，在尝试不同的登顶路线。 在今年的英伟达 GTC（ GPU Technology Conference）上，元戎启行周光提出了一套新的解法，他说，大语言 模型的发展，经历了从弱专家模型（初代 Siri）、到通才（ChatGPT）、再到强专家模型（垂直模型）的过 程。智驾也可以复制这样的路线，一个移动能力的通才，能开好汽车、能骑好摩托车、能让配送小车随时找到 你，之后就可能进化到强专家模型—— ...

Core Points - Circle to Search is a new interaction method introduced by Google, first showcased at the Galaxy Unpack event, and is available on Samsung Galaxy S24 and Google Pixel 8 series devices [2][6] - The feature allows users to select screen areas with simple gestures and provides results based on the selection, enhancing the overall user experience on Android devices [3][7] Group 1: Functionality and User Experience - Circle to Search enables users to quickly access search results by selecting text or images on the screen, with additional tools for translation and music identification [3][8] - The design emphasizes rapid response, low error rates, and intuitive screen segmentation, making it more user-friendly compared to Google Lens [8][9] - The development team utilized a streamlined approach, focusing on prototype creation rather than extensive documentation, which led to a more agile development process [9][10] Group 2: Comparison with Competitors - Circle to Search significantly simplifies the search process compared to similar features in other operating systems like HyperOS and Flyme, which require waiting for screen recognition before initiating actions [15][17] - The integration of Google Search provides a robust backend for Circle to Search, allowing for immediate and relevant results, unlike some competitors that rely on slower AI models [17][20] Group 3: Technical Implementation - The feature leverages the new Android Ink API for smooth and responsive gesture recognition, enhancing the user experience with minimal latency [25][27] - Circle to Search supports various Android devices, including Google Pixel and select Samsung models, with a straightforward activation process [29][31] Group 4: Practical Applications - Users can utilize Circle to Search for various tasks, such as quickly sharing screenshots, translating text, and navigating to locations based on selected text [49][51] - The feature's versatility allows for seamless integration into daily activities, making it a valuable tool for information retrieval and sharing [52][53] Group 5: Future Prospects - Google aims to enhance Circle to Search by integrating AI capabilities, potentially improving its functionality while maintaining a smooth user experience [56][57] - The ongoing development reflects a commitment to creating a comprehensive and user-friendly search tool that adapts to the evolving needs of smartphone users [58][59]

西风 一水 发自 凹非寺 量子位 | 公众号 QbitAI 24小时不间断学习且不遗忘，一辈子也只有 4GB 的"知识储量"？ 科学家们最新研究，计算出了人类学习积累上限，就这么多~~ （甚至还不如一块U盘能装） 。 这是来自Cell旗下神经科学顶刊Neuron上的一项工作，它提出了一个发人深省的悖论： 人类信息处理速度仅为每秒10bit，而我们的感官系统却能以每秒10亿bit的速率收集数据。 大语言模型 每个参数就能存储2bit 知识，一个70亿参数的模型就能存储140亿bit的知识。 △ 结论来自华人学者朱泽园"Physics of Language Models"系列论文 难怪研究人员还提出了一项推论： 随着算力的不断提升，机器在各类任务中的表现超越人类只是时间问题。 另外，按照这项研究的结论，马斯克目前的脑机接口研究也有问题了。 研究人员表示： $\mathbf{S}i=\text{Sifting Number}=\dfrac{\text{Sensory information rate}}{\text{Bewiavel throughput}}=\dfrac{1\text{Gbit/s}}{10\t ...

先有一个移动能力的 "通才"，才有更强的智驾系统。 文 丨 张家豪 编辑 丨 程曼祺 全无人驾驶，始终被视作自动驾驶行业皇冠上的明珠，就像登顶珠穆朗玛峰有 19 条路线一样，不同的公司选 择了不同的路线通往无人驾驶的最终目标。 Waymo、小马们选择了基于高精地图的 RoboTaxi 路线，在特定的路线已经实现了 RoboTaxi，为市民提供没有 司机的出行服务；以特斯拉为代表的车企与供应商，则是通过渐进式路线，卖车搭配辅助驾驶方案，收集数据 一步步迭代方案，试图逼近技术极限。 没有人能笃定哪条路线一定能成功登顶，也还有不同的公司，在尝试不同的登顶路线。 在今年的英伟达 GTC（ GPU Technology Conference）上，元戎启行周光提出了一套新的解法，他说，大语言 模型的发展，经历了从弱专家模型（初代 Siri）、到通才（ChatGPT）、再到强专家模型（垂直模型）的过 程。智驾也可以复制这样的路线，一个移动能力的通才，能开好汽车、能骑好摩托车、能让配送小车随时找到 你，之后就可能进化到强专家模型——L5 级别自动驾驶，这套系统被元戎启行称为 RoadAGI，移动能力的通 才系统。 在演示 D ...

Core Insights - The current wave of artificial intelligence, particularly large language models, presents significant opportunities for the banking industry, enhancing service delivery and operational efficiency [1] Group 1: Changes in Service Models - The banking industry has traditionally focused on serving a limited number of key clients, adhering to the "80/20 rule." With the advent of large language models, banks can transition to a "one customer at a time" approach, enabling personalized services for all clients [2] Group 2: Changes in Work Models - The banking sector, being labor-intensive, can leverage large language models to assist or even replace human employees in repetitive tasks, fostering a new collaborative environment between human and intelligent agents [2] Group 3: Changes in Interaction Models - Interaction methods are evolving from purely graphical user interfaces (GUI) to a combination of GUI and chat functionalities, allowing for dynamic interface organization based on customer intent, thus enhancing user experience [2] Group 4: Changes in Data Analysis - Large models significantly improve data analysis efficiency and lower the barriers to entry, enabling more individuals to engage in data analysis activities [3]

广发证券资深金工分析师 李豪 lhao@gf.com.cn 广发证券首席金工分析师 安宁宁 anningning@gf.com.cn 广发金工安宁宁陈原文团队 摘要 大语言模型在金融领域的应用： 近年来，人工智能技术的快速发展推动了大语言模型(LLMs)的革新。作为最前沿的技术之一，大语言 模型正在广泛应用于各行各业。金融行业作为一个高度依赖数据分析和信息处理的领域，对先进的人工 智能技术有着极大的需求。而LLMs凭借其强大的文本理解能力、信息提取能力以及推理和预测能力， 正在逐步改变传统的金融分析和决策方式，为投资管理、市场分析、风险控制等多个领域带来了新的机 遇。 DeepSeek定量解析基金季报行业观点及行业轮动策略构建： 本文中，我们尝试通过DeepSeekV3模型，对于基金季报观点文本中的行业观点进行定量解析，并以此 出发构建行业轮动策略。具体来看，首先我们筛选存续时间较长的主动型权益基金样本，并提取样本基 金不同季度报告期季报中的观点部分文本；而后我们将观点文本输入至DeepSeek模型，加入特定提示 词控制输出的格式，并基于输出结果构建基金季报行业观点指标；最后我们基于基金季报行业观点指标 及观 ...

Group 1 - The core viewpoint of the article emphasizes the transformative potential of Large Language Models (LLMs) in the financial sector, particularly in investment management, market analysis, and risk control [1][7][8]. - LLMs can process vast amounts of unstructured data, such as news articles, social media, and financial reports, enabling faster access to critical information for investors [7][8]. - The DeepSeek model, a representative of advanced LLMs, showcases strong reasoning capabilities and cost-effectiveness, making high-performance AI technology more accessible [13][19]. Group 2 - The article discusses the quantitative analysis of fund quarterly reports using the DeepSeek V3 model to extract industry viewpoints and construct industry rotation strategies [2][22]. - Approximately 18,000 quarterly report texts were analyzed, focusing on active equity funds with a significant equity position over the past five years [26][31]. - The analysis revealed that the proportion of bullish and bearish viewpoints on various industries varies significantly, with certain sectors like electronics and pharmaceuticals receiving more attention [41][42]. Group 3 - The construction of industry viewpoint indicators is based on the quantitative analysis results, leading to the development of 14 indicators to capture the sentiment towards different industries [56][60]. - The article outlines various strategies for industry rotation based on the constructed indicators, highlighting the performance of different combinations during market conditions [62][66]. - The findings suggest that industries with high attention and bullish sentiment tend to perform better, while those with low attention and bearish sentiment may underperform [75][76].

KG-SFT团队 投稿 量子位 | 公众号 QbitAI 让大语言模型更懂特定领域知识，有新招了！ 来自中国科学技术大学MIRA实验室的王杰教授团队提出了提出了一个创新的框架—— 知识图谱驱动的监督微调（KG-SFT） ，该框架通过 引入知识图谱（KG）来提升大语言模型（LLMs）在特定领域的知识理解和处理能力。 实验结果表明，其在多个领域和多种语言的数据集上取得了显著的效果， 成功入选ICLR 2025 。 截至目前，LLMs在常识问答方面表现越来越出色，但它们对领域知识的理解和推理能力仍然有限。 由于难以深入理解专业领域问答背后所蕴含的复杂知识和逻辑关系，因此在面对这类问题时，往往无法准确地给出正确的答案和详细的推理过 程，这极大地限制了其在专业领域的应用价值。 尤其是在数据稀少和知识密集型的场景中， 如何让LLMs更好地理解和操纵知识，成为了研究的关键 。 而中科大MIRA实验室的这项工作即围绕此展开。 KG-SFT是如何工作的 KG-SFT针对LLMs难以理解领域问答背后的知识和逻辑，导致 推理能力弱 的问题，提出 基于知识图谱增强的大语言模型监督微调 技术。 KG-SFT首先通过解析领域知识图谱中的 ...

Core Insights - Google is replacing the head of its Gemini chatbot in an effort to compete with OpenAI's ChatGPT for market share [1] - The current head of Google's product incubation lab, Josh Woodward, will lead the Gemini chatbot team, while the previous head, Xia Qianqi, is being reassigned [1] - According to Similarweb data, Google Gemini's web traffic is only about one-tenth of ChatGPT's, indicating a significant gap in usage [1] Summary by Sections - **Leadership Changes**: Josh Woodward is appointed to lead the Gemini chatbot team, taking over from Xia Qianqi, who has been with Google for a long time [1] - **Market Competition**: The shift in leadership is part of Google's strategy to regain market share from ChatGPT, which has seen a surge in usage [1] - **Product Development**: Woodward's lab has developed several AI products, including AI Studio for developers, the upcoming Project Mariner, and NotebookLM, which gained attention for its AI podcast feature [1]

大模型若要成为可信的研发工具，还要实现科学的ranking技术、可解释性提升等问题。 本文为IPO早知道原创 大模型 加速分子设计、优化，但可解释性仍然不够 科辉 AI 平台包含数据分析软件、分子设计软件，并且公司于 2024 年 11 月上线了内部大语言模 型。朱 振东 表示，大语言模型作为 AI 工具的一种，在药物研发的智能化升级、生物医学文献与知 识管理、基因组学与蛋白质组学的研究解读和临床诊疗与健康管理的革新等领域都带来了显著的效率 提升。 以药物研发的关键环节为例来看，首先在 疾病机制与靶点发现 方面，大语言模型可通过分析海量基 因组、转录组、蛋白质组、文献、专利和数据库，挖掘疾病相关基因的调控网络和信号通路，快速识 别潜在药物靶点（如蛋白质、基因），并预测其与疾病的关联性。 作者｜罗宾 微信公众号｜ipozaozhidao DeepSeek 等大语言模型（ LLMs ）在很多行业已经引领了智能化的变革，它们 在新药研发行业 是否已接近技术的拐点？人工智能正以多大程度推动药物研发的进步？ IPO 早知道对话了科辉智药创始人、董事长 朱 振东博士 ，他分享了 AI 运用于制药行业中的成功经 验与 未来 ...