语言模型

Search documents
全球首个宠物翻译器,上线爆火
3 6 Ke· 2025-05-23 00:47
近期,谷歌推出 DolphinGemma 大模型, 称将让人类听懂海豚的语言,实现人与海豚在水下的实时交流。另一个由华人团队研发、面向全球英文用户的 人狗交流应用Traini在去年6月出现,成为全球首个实现人宠语言互译的AI原生应用。AI正步入跨物种交流领域,拓宽着人们对非人类语言理解的边界。 一条联系了Traini的CEO孙邻家,他是80后、中国人,老家在吉林长白山。我们与他聊了聊AI新技术对人宠交流领域的影响,从0到1探索过程中的挑战, 以及他三年来身处行业内部的感受。 除此之外,我们还想知道:当人类暂时离开语言的中心,开始尝试建立起与非人类语言平等对话的可能,在新奇感过后,AI+跨物种交流对我们具有怎样 的意义? 孙邻家,80后,吉林长白山人 知名投行高盛近期的一份报告显示,中国的宠物数量首次超过 4 岁以下婴幼儿总量。 同时 根据艾媒咨询的数据, 2023 年中国宠物经济产业规模 就已经 达到 5928 亿元。 根据《2025宠物品牌网红营销生态报告》,以年轻群体为代表的养宠人多将宠物视作"孩子"与"朋友",呈现出情感消费与拟人化养宠的趋势。 这样的需求也催生了相关产业,比如几年前备受争议的宠物灵媒师 ...
腾讯混元TurboS技术报告首次全公开:560B参数混合Mamba架构,自适应长短链融合
AI前线· 2025-05-22 19:57
随着大型语言模型(LLM)的飞速发展,模型能力与效率的平衡成为了前沿研究的关键议题。 腾讯混 元团队最新推出的混元TurboS模型,是一款新颖的 超大型 Hybrid Transformer-Mamba架构MoE模型 。该模型通过Mamba架构在长序列处理上的卓越效率与Transformer架构在上下文理解上的固有优势的 有机协同,实现了性能与效率的精妙平衡。 混元TurboS引入了创新的自适应长短思维链机制,能够根据问题复杂度动态切换快速响应模式与深度 思考模式,从而优化计算资源分配。更重要的是,其模型激活参数达到了56B(总参数560B),是业 界首个大规模部署的Transformer-Mamba专家混合(MoE)模型。 架构创新以及参数量的保证,让模型效果进步明显,国际最权威的大模型评测榜单LMSYS Chatbot Arena最新排名显示: 混元Turbo S 取得了整体1356的高分,在所有239个参赛模型中位列全球前7名。 | Rank* | Rank | Model | Arena 4 | વેરૂર A | Votes | A Organizatio License | 4 | | --- | ...
领域驱动的 RAG:基于分布式所有权构建精准的企业知识系统
Sou Hu Cai Jing· 2025-05-22 13:37
Core Insights - The company is leveraging Retrieval-Augmented Generation (RAG) technology to enhance the accuracy and efficiency of information retrieval within its extensive product line [2][3][5] - A distributed ownership model is being implemented, assigning domain experts to oversee the integration and fine-tuning of the RAG system in their respective areas [3][4][10] - The company is focusing on metadata strategies to improve the context and relevance of information retrieved by the RAG applications [6][7][29] RAG Technology Implementation - RAG combines intelligent search engines with AI-generated responses to provide accurate answers from vast data sources [2][5] - The system is designed to assist human consultants, who are responsible for validating and modifying AI-generated outputs to ensure accuracy [3][4] - The company has developed a comprehensive RAG application that integrates seamlessly into existing workflows, enhancing user experience and information accuracy [10][21] Knowledge Management - The RAG system utilizes a structured approach to generate metadata, which helps users understand the context of system responses [6][29] - Domain experts are tasked with creating high-quality documentation and training materials to ensure effective use of the RAG system [4][5] - The integration of UML diagrams into the knowledge base enhances the understanding of system architecture and component relationships [16][17] Performance Evaluation - The evaluation framework includes metrics such as classifier accuracy (81.7%) and response accuracy (97.4% for correctly classified questions) [22][24] - Findings indicate that specialized models outperform general queries, highlighting the importance of accurate classification in improving answer quality [24][28] - The company aims to continuously enhance the classification system to further improve response accuracy and overall system performance [28][29]
昇腾杀手锏FlashComm,让模型推理单车道变多车道
雷峰网· 2025-05-22 11:29
" MoE模型推理面临的3大通信难题,被通信尖子生华为逐一突 破,未来将进一步优化。 " 作者丨李希 大语言模型 (Large Language Models, LLMs) 自从其问世以来,便迅速成为全球科技领域乃至整个社会 的焦点。根据 Scaling law ,大语言模型的能力与其参数量的对数正相关,因此大语言模型的参数规模也 在指数级增长。随之而来的,是大语言模型部署形态的变化,从神经网络时代的单卡部署,到稠密模型时 代的多卡 / 单节点部署,再到以最近发布的 DeepSeek V3/R1 模型为代表的混合专家( Mixture of Experts, MoE )模型,它甚至会采用数百卡组成的集群和超节点来部署。 而在这基于集群的大模型推理中,集合通信操作就像是一群工人协作盖房子时传递材料和信息的方式,能 让多个计算节点高效配合完成任务。有一些常用集合通信操作,比如全量规约(A ll Reduce)可以想象 成一群工人各自收集了不同区域的建筑材料数据,全量规约就是把所有工人手里的数据汇总到一个地方, 进行求和、求平均值等计算。 大模型的推理,就只是算力吗? 在大模型里,多个计算节点可能各自计算了一部分参 ...
智能辅助驾驶竞速与暗战:自研派VS合作派,功能水平分化加剧
Bei Ke Cai Jing· 2025-05-22 10:37
Core Insights - The article discusses the advancements and competitive landscape of the assisted driving industry, highlighting various companies' self-developed systems and strategies [1][4]. Group 1: Company Developments - Li Auto has launched its new generation dual-system intelligent driving solution, focusing on upgrading driving capabilities and synchronizing updates for smart electric vehicles [3]. - NIO's intelligent assisted driving system has reportedly avoided over 3.5 million collision risks, accumulating a total driving mileage of approximately 4.94 billion kilometers as of May 15, 2025 [3]. - Chery's Hawk 500 has achieved widespread adoption of assisted driving features, with the Hawk 700 targeting mid-to-high-end models and the Hawk 900 positioned as a flagship [3]. - GAC Group's GSD intelligent driving assistance system has accumulated 5 million user driving scenarios and over 40 million kilometers of high-level autonomous driving data [3]. Group 2: Industry Trends - BYD and XPeng are recognized as leaders in self-developed intelligent driving systems, with BYD's high-end system named "Tianshen Eye" [4]. - Bosch's China president has expressed skepticism about the self-development model, suggesting that mid-level intelligent driving should become standard and that costs could be better managed through supply chain partnerships [4]. - Huawei is positioned as a top player in the intelligent driving system market, with plans for 10 brands from 7 automakers to adopt its solutions, potentially exceeding 500,000 vehicles [4][5]. - Huawei's collaboration models include component supply, Huawei Inside (HI) partnerships, and deep cooperation with automakers, with the latter being the most integrated approach [5]. Group 3: Strategic Partnerships - SAIC Group has publicly stated its intention to maintain control over core technologies while also choosing to collaborate with Huawei [6]. - The partnerships with Huawei have led to increased sales for collaborating automakers, but questions remain about their ability to independently develop high-quality vehicles [6].
鸿蒙折叠电脑官网预约量超10万
第一财经· 2025-05-22 06:08
2025.05. 22 本文字数:2239,阅读时长大约4分钟 作者 | 第一财经 李娜 值得注意的是,在这个战略项目中,一个名为"543-AI"的项目也在被紧张推进,主要为鸿蒙的原生 智能提供系统性产品规划,最终的使命是将AI融入新一代终端设备的核心。 5月22日午间,华为商城数据显示,鸿蒙电脑的预约量接近14万人,其中售价23999元起的鸿蒙折 叠电脑预约人数超过10万。 而在手机的备货量上,一位华为渠道经销商对记者表示,"nova系列的备货量是上一代的两倍,目前 看首销是比较乐观的,后面就要看首批用户的反馈了。" 华为电脑以及nova系列开始搭载鸿蒙系统,这被外界视作鸿蒙5向主流消费市场渗透的标志。在不久 前的一场大学演讲中,华为常务董事、终端BG董事长余承东透露,仅nova14系列的备货量就在千 万级别,而在下个月,搭载鸿蒙的华为旗舰手机也会上市。 "如果有人拧熄了灯塔,我们怎么航行?"这句曾经由华为创始人任正非发出的疑问正在交由华为人 自己解答。在科技生态日益分裂的大时代下,鸿蒙这个曾经作为华为"逃生计划"的技术储备,正在 成为华为手机重生后市场破局的关键。 "将AI融入鸿蒙的一部分" "小艺帮我接 ...
澜起科技:业绩高增,运力芯片前景广阔
He Xun Wang· 2025-05-21 14:39
Core Viewpoint - The company aims to become a leading international interconnected chip design firm over the next five to ten years, focusing on interconnect chips to support cloud computing and AI infrastructure [1] Business Strategy - The company will expand its business layout from three dimensions: - Continuous investment in DDR memory interface product upgrades in the memory interconnect field to lead technological innovation [1] - Strengthening core underlying technology research and development in the PCIe/CXL interconnect field to promote product upgrades and market expansion [1] - Exploring niche markets in the Ethernet and optical interconnect fields through various methods to advance product layout [1] Financial Performance - The company reported a revenue of 3.639 billion yuan for 2024, a year-on-year increase of 59.2%, and a net profit of 1.412 billion yuan, up 213.1% [1] - In Q1 2025, the company achieved a record high in revenue and net profit, with revenue of 1.222 billion yuan, a year-on-year growth of 65.78%, and a net profit of 525 million yuan, an increase of 135.14% [1] - As of April 22, 2025, the company expects over 1.29 billion yuan in orders for interconnect chips to be delivered in Q2 2025, with new orders still being received [1] Industry Insights - The chairman noted that advancements in AI technology are shifting the industry from "computing" to "intelligent computing," with generative AI and large language models reshaping the tech sector and driving growth in the AI server market [1] - The AI infrastructure sector continues to benefit from this transformation, with interconnect capabilities becoming increasingly important [1] - The company has a strong technical foundation in interconnect chips and actively participates in the formulation of related product standards [1]
何恺明等新作大道至简,瞬时速度改为平均速度,一步生成表现提升70%
量子位· 2025-05-21 06:31
白交 发自 凹非寺 量子位 | 公众号 QbitAI 何恺明等团队新作新鲜出炉,再次大道至简—— 他们引入平均速度,实现「一步生成」新SOTA。 CMU博士生耿正阳一作,何恺明的学生邓明扬、白行健参与。 他们提出的模型是从头开始训练的,没有任何预训练、蒸馏或课程学习,最终实现了3.43的FID值,明显优于之前最先进的一步扩散/流模型。 一步生成框架:引入平均速度 一次生成模型,指的是只需一步计算就产生高质量的结果,而无需多次迭代。 团队提出了一个原则性强且有效的单步生成框架MeanFlow。其核心思想是引入平均速度的概念来表征流场,这与流匹配方法所模拟的瞬时速 度截然不同。 △ 流匹配的速度场,瞬时速度 平均速度被定义为位移与时间间隔的比率,位移由瞬时速度的时间积分给出。 根据这一定义,这说明平均速度和瞬时速度之间定义明确的内在联系,这自然成为指导网络训练的原则基础。 我们的方法被称为MeanFlow模型,它自成一体,无需预先训练、提炼或课程学习。 演示1:通过jvp计算只需要一次后向传递,类似于神经网络中的标准反向传播,开销不到总训练时间的20%。 它在从零开始训练的ImageNet 256×256上通过1 ...
大语言模型“吵架水平”超越人类
Huan Qiu Wang Zi Xun· 2025-05-21 02:57
该研究的辩论采取了一种结构性方法,而现实世界辩论的自由度更高,且辩论有时间限制。研究者指 出,研究结果揭示了人工智能驱动的工具影响人类观点的潜力,可能对在线平台的设计具有借鉴意义。 (冯维维) 相关论文信息: https://doi.org/10.1038/s41562-025-02194-6 《中国科学报》 (2025-05-21 第2版 国际) 来源:中国科学报 科学家发现,在线辩论中,GPT-4一类的大语言模型(LLM)如能根据对手的个性化信息调整论据,其 说服力将比人类高64.4%。研究显示,GPT-4具有生成有针对性和说服力论据的能力,并提出应进一步 研究如何降低其用于说服时的风险。相关研究5月19日发表于《自然-人类行为》。 有研究显示,随着人类与LLM的对话日益普遍,LLM可能变得更有说服力,即能改变一个人的信念或 观点。然而,之前并不清楚这些模型能否根据个性化信息进行调整,提出更能针对辩论对手的论点。 瑞士洛桑联邦理工学院的Francesco Salvi和同事分别将900名美国人与另一个人或GPT-4配对,使双方辩 论各种社会政治议题。在有些配对中,辩论对手——无论是人工智能还是人类,均能获得 ...
ICML 2025 Spotlight | 多模态大模型暴露短板?EMMA基准深度揭秘多模态推理能力
机器之心· 2025-05-20 04:58
「三个点电荷 + Q、-2Q 和 + 3Q 等距放置,哪个向量最能描述作用在 + Q 电荷上的净电力方向?」 在解这道题时,我们可以通过绘制受力分析草图轻松解决。但即使是先进的多模态大语言模型,如 GPT-4o,也可能在理解「同性相斥」的基本物理原则时,错误 地判断斥力的方向(例如,错误地将 + 3Q 对 + Q 的斥力方向判断为右下方而非正确的左上方)。 这个看似简单的物理问题,却暴露了多模态大模型一个「致命缺陷」: 当前的 MLLMs 仍然无法进行需要深度视觉与文本融合的复杂多模态推理 !一项最新研究 推出的 EMMA 基准测试,如同一面「照妖镜」,揭示了即使是顶尖 MLLMs 也在这关键能力上显著不足。 目前该研究已被 ICML 2025 接收为 spotlight,代码数据已全部开源 ! 目前已有多个模型 / 方法在 EMMA 上验证其多模态推理能力,研究发现: 即使最先进的模型 ——Gemini-2.5-pro-exp-03-25 ,或者是能够进行视觉工具调用的 o3/o4-mini 模型在 EMMA 上的表现仍然落后人类专家超 20% ! 标题: Can MLLMs Reason in Multi ...