https://github.com/HW-whistleblower/True-Story-of-Pangu 截至 2025-07-10 为止，作者「HW吹哨人」已追更补充 4 次，该文已收过 10500+ Star。347 个 issues，也是相当热闹。 推荐阅读 点击标题可跳转 1、 腾讯：鸿蒙架构特殊，微信适配必须从零重写代码 2、 华为盘古团队：否认抄袭！ 3、 B 站游戏高管"小姐姐"被逮捕 7 月 5 日， 华为盘古技术团队公开发声明，否认了最近的抄袭传闻 。 7 月 6 日，一篇自称盘古团队内部员工写的文档，在网上引发热议。 该文也发在 GitHub 上了。 - EOF - ...

Core Insights - Musk's xAI has launched its latest AI model, Grok 4, which is touted as the "strongest model in the universe" and claims to outperform competitors in various academic assessments [2][5][6] - Grok 4 achieved a 38.6% accuracy rate in the "Humanity's Last Exam," surpassing Google's Gemini 2.5 Pro and OpenAI's o3 [2][5] - The model's training involved a significant increase in computational resources, utilizing a supercomputing center with 100,000 H100 GPUs, resulting in a training volume 10 times that of Grok 3 and 100 times that of Grok 2 [2][6] Performance Metrics - Grok 4 demonstrated exceptional reasoning capabilities, scoring 88.9% in the Graduate-level Question Answering (GPQA) and achieving a perfect score in the American Mathematics Invitational Exam (AIME25) [6] - In a commercial simulation task, Grok 4 managed an average net asset of $4,684.15, double that of its closest competitor, showcasing its long-term planning and multi-step reasoning abilities [6] Business Strategy - xAI has introduced a premium subscription plan for Grok 4 at $300 per month, which is 50% more expensive than OpenAI's top-tier subscription [7] - The API pricing is also aggressive, charging $3 per million tokens for input and $15 for output, reflecting the high training costs associated with Grok 4 [7] Future Developments - Musk plans to integrate Grok 4 with humanoid robots and aims to create high-precision physical simulators, including black hole simulations, to test AI against physical laws [7][8] - Grok 4 is expected to be embedded in Tesla's latest firmware, potentially serving as the brain for voice assistance and autonomous driving [3][7] Industry Context - The AI arms race is intensifying, with Musk's aggressive pace in advancing AI models and applications, positioning xAI as a formidable competitor in the market [9] - The integration of various technologies, including autonomous driving and commercial space ventures, is creating a closed-loop system that enhances the capabilities of Grok 4 [9]

去年有一个双非的同学找到我们，情况是：老师没有人带，但自己想申请博士，想咨询有没有快速发表论文的 渠道。在分析这位同学的基础和硬件资源后，我们为他快速制定了一个研究方向，并匹配到了相关的老师！经 过近10个月的沟通、实验、写作，最终成功投出到了CVPR25，并被录取。成为学院首个发CVPR的硕士研究 生。 SCI一区~四区； 中科院1区，2区，3区，4区； 谈到这个，归咎于2点。没人指导不可怕，可怕的是自己不行动，主动出击才有胜算。如果当时没有主动找老 师辅导，也许CVPR对他来说只是一个梦。还有就是同学性格很主动、肯吃苦，经常分析到凌晨。遇到问题不 逃避，敢于直面！ EI/中文核心； 毕设论文/申博/比赛等； 如果你缺乏指导、身边没有老师带着科研，欢迎联系具身智能之心！我们提供从idea->实验->写作->投稿一站 式服务。 辅导方向：大模型、VLA、视觉语言导航、端到端、强化学习、Diffusion Policy、sim2real、具身交互、抓取 点预测与位姿估计、机器人决策规划、运动规划、3DGS、SLAM、触觉感知、双足/四足机器人、遥控操作、 零样本学习等方向，如果您有任意论文发表需求，支持带课题/ ...

点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 端到端自动驾驶 - 下一代智能驾驶量产核心算法 端到端自动驾驶（End-to-End Autonomous Driving）作为目前智驾量产的核心算法，可以分为一段式端到端、二段式端到端两个大的技术方向。自UniAD获得 CVPR Best Paper以来，正式拉开了国内新一轮的智驾军备竞赛。 2024年理想汽车更是宣布E2E+VLM的双系统架构量产！ 端到端自动驾驶通过传感器数据输入 （视觉/Lidar等）直接输出自车规划或控制信息，是目前智能驾驶最具代表性的方向。 目前VLM/VLA也是招聘的刚需，3-5年就能冲击百万年薪！ 而随着学术界和工业界的目光投向端到端这个技术领域，我们发现了很多问题。UniAD是端到端的最终解吗？显然不是！一系列算法如雨后春笋般冒出： 技术栈多？入门困难？ 去年我们推出了《首个面向工业级的端到端算法与实战教程》，今年很多小伙伴反馈技术发展太快了，先前的技术方案已经不适合当下的大环境。端到端目前发 展出多个领域技术的方向，需要掌握多模态大模型、BEV感知、强化学习、视觉Trans ...

Core Insights - The article discusses the evolution of artificial intelligence from 1.0 to 2.0, highlighting SenseTime's breakthroughs in multimodal interaction technology and its applications across various sectors [1][2]. Group 1: AI Evolution - SenseTime has transitioned from focusing on computer vision in the AI 1.0 era to promoting multimodal interaction innovations in the AI 2.0 era, driven by the rise of large model technologies in 2023 [1]. - The concept of "seeing is believing" is emphasized, integrating video, images, and voice to enable real-time interaction with humans [1]. Group 2: Applications in Education - In the education sector, SenseTime collaborates with learning device manufacturers to develop interactive devices that utilize real-time algorithms to assist children in solving problems and recognizing errors [2]. - The system supports interactive storytelling for young children by converting images into narratives, and SenseTime has partnered with around 10 schools to create smart campus assistants for managing course schedules and grade inquiries [2]. Group 3: Intelligent Applications - SenseTime's intelligent applications include algorithms that analyze industry data to assist in warehouse leasing scenarios and generate lease management solutions [2]. - In customer service, SenseTime collaborates with well-known operators to create efficient intelligent agents, and in smart home applications, it enhances family interaction through AI technology [2]. - The advantage of multimodal large models lies in enabling smoother interactions beyond text command recognition, utilizing visual and multidimensional information [2].

左手刚刚融资，右手就发大模型，马斯克重金打造的Grok 4，正式面世！ 7月10日，特斯拉创始人兼首席执行官马斯克旗下的人工智能公司xAI正式发布了Grok 4。在将近1小时 的发布会直播中，xAI发布了这个系列的两款模型，分别是Grok 4（单智能体版本）和Grok 4 Heavy （多智能体版本），其中后者支持4个智能体并行思考，在推理过程中横向比对、纵向协同，调用更大 规模的计算资源以完成更复杂、更精密的任务。 作为xAI在2023年推出首代大模型以来的第四次重要更新，Grok 4在"人类的最后考试"（Humanity's Last Exam）取得了25.4%的准确率，超过了谷歌Gemini 2.5 Pro的21.6%和OpenAI o3（高版本）的21%，被称 为"世界上最强AI模型"。 据xAI的研究人员介绍，Humanity's Last Exam测试总共有2500个问题，包括数学、自然科学、工程以及 所有人文学科，问题广泛且都是博士甚至高级研究水平，极具挑战性，但Grok 4在这些问题上都可以得 到很好的分数。 此外，据发布会披露，在GPQA、AIME25、LCB（Jan-May）、HMMT25 ...

7月初，华为盘古大模型陷入一场"抄袭门"风波。 起因是一位开源社区GitHub用户HonestAGI发表报告，称盘古Pro MoE模型注意力参数分布与阿里通义千问Qwen-2.5 14B模型相似度极高，平均相关系数 达0.927（接近完全一致的1.0），而且代码文件中含阿里Qwen的版权声明。该报告作者认为，盘古模型可能在千问模型的基础上进行了增量训练，而非完 全从零训练。 随后，一份自称为"华为盘古大模型团队、华为诺亚方舟实验室员工"的用户HW-whistleblower（华为吹哨人）在GitHub发表博文《盘古之殇》，以"亲历 者"的口吻，讲述了他眼中的"盘古套壳"事件，进而将此事推向舆论漩涡。 于盘古大模型开源代码的讨论。 盘古 Pro MoE开源模型是基于昇腾硬件 平台开发、训练的基础大模型,并非基 于其他厂商模型增量训练而来,在架构 设计、技术特性等方面做了关键创新, 是全球首个面向异腾硬件平台设计的同 规格混合专家模型,创新性地提出了分 组混合专家模型 (MoGE) 架构,有效 解决了大规模分布式训练的负载均衡难 题,提升训练效率。其他技术特性创 新,请参考异腾生态竞争力系列技术报 告披露内容。 ...

Core Insights - The launch of the Ketu 2.1 model by Keling AI significantly enhances image generation capabilities, including improved instruction adherence, stunning portrait aesthetics, and cinematic quality [1][11] - The model is available for free to all member users for a limited time, allowing creators to explore its features [11] Image Generation Capabilities - Ketu 2.1 excels in following complex instructions, accurately capturing multiple elements and details in prompts, resulting in high-quality images that showcase creative imagination [1][3] - The model demonstrates a notable improvement in image quality, including clarity, richness of elements, and realism, particularly in portrait aesthetics [3][5] Artistic and Cinematic Quality - The model can generate images with a cinematic feel, effectively recreating scenes with unique aesthetic tones and advanced composition [6] - It supports over 180 different styles, allowing creators to choose from various artistic expressions, from vintage photography to futuristic digital art [10] Text Generation Features - Ketu 2.1 also enhances text generation, producing clear and design-oriented text in both Chinese and English, facilitating smoother integration of text and images for marketing and creative projects [8] User Engagement and Growth - Keling AI has achieved significant user engagement, with a total of 344 million images and 168 million videos generated since its launch, showcasing its strength in the image generation sector [11]

来源：光明网 中国科学院院士、中国自动化学会理事长、中国空间技术研究院研究员杨孟飞出席。中国自动化学会特 聘顾问、青岛科技大学副校长、上海交通大学教授李少远，华北电力大学教授肖峰，中国自动化学会副 监事长、安徽大学教授孙长银，中国自动化学会理事、武汉大学教授张俊作主旨报告。中国自动化学会 理事、中国科学院自动化研究所研究员魏庆来，英国格拉斯哥大学教授于慧，英国剑桥大学助理教授那 晓翔作专题报告。北京航空航天大学教授王卓主持报告环节。中国科协第十届青年人才托举工程入选 者、中国科学院自动化研究所副研究员王晨主持圆桌讨论环节。 7月5日，第二十七届中国科协年会复杂系统自学习"逆最优"理论与方法专题论坛在北京召开。本次专题 论坛由中国科协主办，中国自动化学会承办，与会专家围绕实际复杂系统最优运行建模等非共识议题， 共同探讨复杂非线性系统自学习"逆最优"发展路径。 张俊教授作题为"基于生成式人工智能和科学智能（AI4S）的复杂电力系统数智化关键技术与应用"的报 告 孙长银教授作题为"试错驱动具身智能学习与进化"的报告 魏庆来研究员作题为"自学习最优控制"的报告 杨孟飞理事长出席论坛 李少远教授作题为"基于'智能'增强 ...

Core Insights - The future of large models and intelligent agents will focus on enhancing capabilities, scene migration, and convenient access, leading to a powerful single-agent and collaborative multi-agent ecosystem [4][8] - Global large models are evolving towards stronger capabilities, more modalities, and higher efficiency, with a significant improvement in intelligent emergence and cross-modal interaction abilities [5] - The application scenarios for large models will become more generalized, supported by distributed deployment, open-source collaboration, and the integration of models, computing power, and data [6] Group 1 - The integration of connectors will facilitate seamless interaction between models and intelligent agents, preventing "model islands" and "intelligent gaps" [7] - A collaborative model system is emerging, characterized by the synergy of "basic-specialized," "large-small," and "central-edge" models, marking a breakthrough year for intelligent agents [8] - Large model-driven innovations will create new opportunities in intelligent industries and significantly transform productivity across sectors, potentially triggering a new wave of AI industrialization [8]