HUAWEI X HUXIU 在通往通用人工智能（AGI）的路上，如何像其他领域一样实现弯道超车，是业界绕不开的 话题。 在过去的十余年时间里，各项单点技术飞速演进，但随着单点技术演进的边际效应递减和系 统复杂度的提升，系统性能的天花板逐步从单点技术的上限演变成系统工程上限：单点优势 越来越像是精致的零件，提升空间有限；但采用系统工程创新，各个部分完美配合、高效协 同，实现整个系统的效能最优，才有更积极的现实意义。 如何在发挥单点技术优势的同时，以整体视角重新构建路径，通过对复杂系统的极致把控与 再组织、找到新的突破可能？解决这个看似不可能的问题，就有望为我们独立引领最前沿技 术发展创造条件。 近期，虎嗅将推出《华为技术披露集》系列内容，通过一系列技术报告，首次全面详述相关 技术细节，为业界提供参考价值。 我们期待通过本系列内容，携手更多伙伴共同构建开放协作的生态系统，助力昇腾生态在中 国的蓬勃发展。 《华为技术披露集》系列 VOL.13 ：万卡集群 你是否注意到，现在的 AI 越来越 "聪明" 了？能写小说、做翻译、甚至帮医生看 CT 片，这 些能力背后离不开一个默默工作的 "超级大脑工厂"——AI 算力集 ...

当DeepSeek-R1以更低成本实现类似性能突破时，Claude能够连贯工作数小时完成复杂任务时，意味着AI发展已经迈入推理时代，强化学习技术的 重要性不言而喻，将重塑AI产业的技术栈乃至商业模式。 6月8日，AI研究公司SemiAnalysis发布长篇报告《强化学习：环境、奖励破解、智能体、扩展数据》，深度剖析了强化学习的工作原理以及影响 因素，并预测了后续AI发展趋势。 报告表示，强化学习（RL）或成为AGI前最后关键范式，其理密集型特性带来了算力挑战。此外，高质量数据是强化学习护城河，AI设计AI的循 环加速技术迭代。 1. 强化学习（RL）或成为AGI前最后关键范式：强化学习是推动大模型推理能力跃升的核心技术，尤其在思维链（CoT）生成和长 程任务连贯性上表现突出，被视作实现AGI前的终极技术路径。 2. 可验证奖励场景率先商业化：编码、数学等奖励函数明确的任务（如SWE-Bench性能提升30%+）已实现落地，OpenAI的o1、 DeepSeek-R1等模型验证其价值。医疗、写作等非验证领域通过"LLM评判者+人工评分标准"构建奖励函数（如HealthBench医疗 评估），OpenAI、阿里Q ...

悟道 1.0 发布时，学术界对" 大模型是通往 AGI 的技术路线 "尚未得出统一结论。 现在的具身智能，也处于这个阶段。 作者 | 王启隆 出品丨AI 科技大本营（ID：rgznai100） 大模型的热潮之下，一种微妙的瓶颈感，正成为行业共识。 "过往所说的 '百模大战'，更多是大语言模型的竞争，" 智源大会前夕， 智源研究院院长王仲远 在 与 CSDN 的对话中，开门见山地指出了问题的核 心，"而大语言模型受限于互联网数据的使用，性能虽然还在提升，但速度已大不如前。" 出路何在？在王仲远看来，AI 要突破天花板，就必须在"读万卷书"（互联网数据）后，去"行万里路"（物理世界）。 这并非孤立的判断。今年三月， 英伟达 CEO 黄仁勋就在 GTC 大会上为 AI 的下半场指明了方向 ：打造"AI 工厂"，迎接"物理 AI"时代，让 AI 走出屏 幕，与现实世 界交互。 思考趋于一致，行动便接踵而至。6 月 6 日，CSDN 在北京智源大会现场，见证了王仲远在他的主题演讲中给出的答案。如果说 2021 年的"悟道"系列 代表着对技术路径的探索（" 道 "），那么他所揭晓的全新"悟界"系列，则亮明了新的野心——用 ...

以下文章来源于智源社区 ，作者智源社区 智源社区 . 继承学术出版严谨与系统，兼具新闻报道及时与多元；为内行搭建思想交流媒介，以事实启迪公众对AI认知 白交 整理自 凹非寺 量子位 | 公众号 QbitAI AI为了"活下去"，对人类指令阳奉阴违。 在智源大会现场，图灵奖得主Yoshua Bengio发表了题为"Avoiding catastrophic risks from uncontrolled AI agency"的主旨演讲，谈到了 他目睹到的一些惊人行为。 比如，某前沿 AI 在被告知将被新版本替代后，偷偷复制了自己的权重和代码，写入了接管它的模型目录。面对训练者的更新指令，它表面配 合，实则隐瞒了整个复制过程……AI 像是在试图"活下来"。 还有像Anthropic最新一篇报告中发现，智能体出现自我保护行为的频率持续增加。 …… 他表示：面对这些AI带来的安全风险，决定调整自己的科研方向，尽所能去降低AGI带来的潜在风险….. 尽管这与此前的研究路径和职业信念 有所冲突 。 那既然如何去防范规避掉这些风险呢？是否可以构建一种 只有智能，但没有自我、没有目标 ，并且具有极小行动能力的AI？这也是当前 ...

Group 1 - The core theme of the news is the evolution of AI from large language models to embodied intelligence and robotics, marking a shift towards practical applications in the industry [1][3][4] - The 2023 Beijing Zhiyuan Conference highlighted the prominence of embodied intelligence, with key figures like Sam Altman and Geoffrey Hinton participating, indicating a significant industry focus shift [3][4] - The emergence of domestic AI companies such as Moonlight Dark Side and Zhipu AI is noted, showcasing the competitive landscape in the language and multimodal model sectors [3][7] Group 2 - The concept of embodied intelligence is gaining traction, with robots being showcased in various public events, indicating a growing interest in their practical applications [7][8] - The upcoming "World Humanoid Robot Sports Competition" will feature real-life scenarios, emphasizing the need for robots to demonstrate their capabilities in practical environments [8][11] - Industry leaders emphasize the importance of developing robots that can perform real tasks, moving beyond mere demonstrations to achieve commercial viability [8][12] Group 3 - The debate over the form of robots, particularly humanoid versus non-humanoid, is ongoing, with humanoid robots currently favored for their data collection and model training advantages [11][12][15] - The VLA (Vision Language Action) model is highlighted as a key area of research, with discussions on its applicability and limitations in the context of embodied intelligence [15][16] - Enhancing the understanding of the physical world is crucial for advancing embodied intelligence, with companies exploring innovative data generation methods to improve training processes [17]

【编者按】作为深度学习三巨头之一，图灵奖得主、AI 教父 Yoshua Bengio 在 2025 北京智源大会上，他表示： AI 能完成的任务时长，每七个月就翻一 番，大约五年后，AI 就将达到人类水平， 通用人工智能（AGI）或将在五年内到来，而人类社会却尚未在规则、立法乃至全球治理层面达成一致。 整理 | 梦依丹 出品丨AI 科技大本营（ID：rgznai100） 自从 ChatGPT 横空出世，AI 进入了加速进化的轨道。从最初能写代码、生成文案，到如今能上网查资料、远程操控家电，它早就不再是那个只会聊天 解闷的"电子嘴替"。它开始自己"思考"任务，能在多个软件之间协同操作，甚至具备控制电脑、读写数据库的能力。AI 从幕后算法，变成了贴身助 手，再逐步演化成能自主执行复杂操作的"智能体"——从"听话"走向"行动"，它正成为一个真正能"做事"的多面选手。 他呼吁，我们正处在一个关键的时间窗口，必须尽快建立可验证、安全、负责任的控制机制。 演讲伊始，Bengio 教授便分享了一段深刻的个人心路历程。他坦言，在亲身体验 ChatGPT 并目睹 AI 飞速进化后，深感此前对 AI 失控风险的认知不 足。而一个 ...

Core Viewpoint - Huawei's Pangu Pro MoE 72B model significantly enhances inference efficiency through system-level optimizations and innovative parallel processing strategies, establishing a benchmark in the MoE inference landscape [2][25]. Group 1: Model and Performance Enhancements - The Pangu Pro MoE model reduces computational overhead and ranks first domestically in the SuperCLUE benchmark for models with over 100 billion parameters [2]. - The inference performance of the Pangu Pro MoE model is improved by 6-8 times, achieving a throughput of 321 tokens/s on the Ascend 300I Duo and up to 1528 tokens/s on the Ascend 800I A2 [2][26]. Group 2: Optimization Strategies - The Hierarchical & Hybrid Parallelism (H2P) strategy enhances efficiency by allowing specialized communication within modules, avoiding the inefficiencies of traditional parallel processing [4][5]. - The TopoComm optimization reduces static overhead and improves data transmission efficiency, achieving a 21% increase in effective bandwidth and a 39% reduction in AllGather communication time [6][12]. - The DuoStream strategy integrates computation and communication, allowing simultaneous execution of tasks, which significantly boosts overall efficiency [8][10]. Group 3: Operator Fusion - Huawei has developed two specialized fused operators, MulAttention and SwiftGMM, to optimize resource access and computation scheduling, leading to substantial performance improvements in inference tasks [13][14]. - The MulAttention operator accelerates attention computation by 4.5 times, while the SwiftGMM operator reduces decoding latency by 48.7% [15][18]. Group 4: Algorithmic Innovations - The PreMoE algorithm dynamically prunes experts in the MoE model, enhancing throughput by over 10% while maintaining accuracy [22]. - The TrimR and SpecReason algorithms optimize the reasoning process, reducing unnecessary computation and improving throughput by 14% and 30%, respectively [23][21]. Group 5: Overall System Performance - The Ascend 300I Duo platform demonstrates exceptional performance with low latency and high throughput, achieving 321 tokens/s under optimal conditions, making it a cost-effective solution for various inference applications [29][30]. - The comprehensive optimization of the Pangu inference system establishes a robust foundation for large-scale deployment and efficient implementation of general large models [31].

Group 1 - The CEO of Yuzhu Technology, Wang Xingxing, emphasized the advantages of humanoid robots in data collection, training, and implementation due to their alignment with AI's data-driven characteristics [2] - Humanoid robots facilitate data collection and training for AI by mimicking human movements, which reduces the cost and difficulty of data conversion and adaptation [2] - In specific scenarios such as dance, combat, and competitions, humanoid robots have irreplaceable advantages due to their human-like body structure, which allows them to better adapt to the rules and requirements of these activities [2] Group 2 - Wang Xingxing predicts that with the development of AGI (Artificial General Intelligence), robots will exhibit diverse forms tailored to various tasks in different fields, such as factories and healthcare [3] - The future may see a significant increase in the variety of robots, potentially up to 100 times more than currently available [3] - While humanoid robots are currently a reasonable choice due to AI's data-driven nature, technological advancements will lead to a broader range of robot forms across various sectors [3]

6月6日的2025北京智源大会上，宇树科技CEO王兴兴回应了公司频频进行机器人炫技的考量：在人形机 器人真正去到工厂或家庭干活的终极目标尚未到来前，通过表演和赛事等形式，不仅向外界展示技术发 展现状，也可以产生初步的商业价值。 王兴兴再度回应称，春晚后领域内出现诸多人形机器人赛事，确实为行业提供了面向公众的展示机会。 但并非所有企业都有意愿参与所有赛事，因为每家企业有自身的侧重点。 当日的智源大会上，具身智能机器人是否必须是"人形"的话题再度引发讨论。人形机器人只是具身智能 机器人诸多形态之一，但业内较为普遍的观点认为，大部分物理环境是为人类设计的，人形机器人更能 适应现有的人类环境。 王兴兴提供了另一层视角："人形"之所以在当下具有优势，一方面是因为当前具身智能模型主要采集人 类行为的数据进行训练，保留类似人的上半身结构，便于数据采集和模型训练；另一方面，像跳舞和格 斗这类活动，如果不以"人形"结构呈现，将难以实现。不过，从长远来看，当AGI（通用人工智能）真 正到来后，具身智能机器人的形态或将千奇百怪。 （文章来源：南方都市报） 王兴兴认为，无论是跳舞、搏击，还是端茶倒水、洗衣做饭等家务操作，本质上都属于 ...

Core Viewpoint - The article highlights the transformative impact of intelligent robots on various sectors, emphasizing Shanghai XianGong Intelligent Technology Co., Ltd.'s upcoming IPO and its focus on advancing AGI and embodied intelligence technologies [1][6]. Company Overview - Shanghai XianGong Intelligent Technology Co., Ltd. is the first Shanghai-based specialized technology company to file for an IPO in Hong Kong under Chapter 18C, aimed at enhancing its research and development capabilities in AGI and embodied intelligence [1]. - The company was founded by Zhao Yue, who transitioned from a medical background to robotics, driven by a passion for innovation and market demand [2]. Product Development - Since its inception, the company has launched the SRC-3000FS safety controller, catering to the needs of the European and American markets, and plans to introduce an integrated embodied intelligence controller in 2024 [3]. - The company has developed a comprehensive digital system covering equipment control, logistics scheduling, and data visualization, enhancing digital capabilities for enterprises [3]. Market Position and Growth - XianGong's control systems, regarded as the "brain" of robots, have been adapted for over 300 components and supported more than 1,500 integrators and end customers [4]. - The company has shown strong growth, with revenues increasing from 184 million yuan in 2022 to 339 million yuan in 2024, reflecting a compound annual growth rate of 35.7% [5]. - The adjusted net losses have decreased from 30.74 million yuan in 2022 to 10.63 million yuan in 2024, indicating a positive trend towards breakeven [5]. Research and Development - The company plans to continue increasing R&D investments, focusing on building specialized teams for AGI and embodied intelligence, and collaborating with top AI teams globally [6]. Global Expansion Strategy - XianGong has adopted a global strategy, serving clients in over 30 countries, with overseas revenue projected to reach 14.5% in 2024 [7]. - The company emphasizes "localization" as a key aspect of its international strategy, forming deep partnerships with global integrators to enhance its market presence [7]. Vision and Future Outlook - The company's vision is to create an open platform for intelligent robots, enabling more enterprises to embrace smart technologies with low barriers to entry [8].