近期，一则关于AWS亚马逊云科技上海AI研究院解散的消息在科技圈内引起了广泛关注。据悉，该研 究院的首席应用科学家王敏捷在个人社交平台上透露，AWS亚马逊云科技上海AI研究院，也是AWS在 全球范围内设立的最后一个海外研究院，已接到正式解散的通知。 王敏捷回顾了自己近六年的带队经历，感慨万分。他表示，这段时光恰逢外企研究院的黄金发展期，自 己更是在张峥老师的悉心指导下，有幸成为了AWS亚太地区最年轻的首席应用科学家。在这段期间， 他们从零开始，成功孵化出了全球知名的图神经网络开源框架DGL，为亚马逊电商创造了近10亿美元 的营收。尽管团队规模不大，但他们在机器学习与系统领域取得了显著成就，发表了100余篇顶级会议 论文，实现了该领域顶会全满贯的佳绩。 AWS亚马逊云科技上海AI研究院自2018年秋季成立以来，便隶属于亚马逊网络服务机器学习部门 （Amazon Web Services Machine Learning），致力于推动人工智能技术的发展。上海纽约大学计算机教 授张峥曾担任该研究院的首任院长，为研究院的发展奠定了坚实基础。 面对此次解散的决定，亚马逊云科技方面给出了正式回应。他们表示，经过对公司组织 ...

7月23日，市场有传闻称亚马逊云科技（AWS）上海AI研究院已解散。 网传消息显示，亚马逊AWS上海AI研究院的首席应用科学家王敏捷朋友圈发布内容称："刚收到通知，AWS亚马逊云科技上海AI研究院因中美战略调整 解散。" 值得注意的是，这是亚马逊云科技最后一个海外研究中心。目前尚不清楚此次变动涉及多少名员工，但有消息人士向英国《金融时报》透露称，AWS在 巅峰时期在中国拥有超过1000名员工。 网传图片中，王敏捷感慨道，"近6年带队时光，赶上了外企研究院的黄金周期，更得益于张峥老师的细心指导，有幸成为AWS亚太地区最年轻的首席应 用科学家。值得骄傲的是：我们从零孵化出全球知名的图神经网络开源框架DGL，为亚马逊电商创造了近10亿美元的营收；仅实验室规模的团队，拿下 机器学习与系统领域顶会全满贯，发表100余篇顶会论文。转向AI Agent后，敢说这支队伍在技术深度、科学素养与执行力上，都是最顶尖的Agentic AI 团队之一 —— 从框架到落地经验，全是现成的。" 网传图片 针对这一消息，亚马逊发言人23日向媒体回应称："经过对公司组织、发展重点及未来战略方向的深入评估，我们决定对亚马逊云科技部分团队进行人 ...

公众号记得加星标⭐️，第一时间看推送不会错过。 不断增长的人工智能（AI）需求暴露出一个严峻的"计算危机"，其特点是能源消耗不可持续、训练成本过高以及传统互补式金属氧化物 半导体（CMOS）微缩技术接近极限。「基于物理的专用集成电路（ASIC）」提供了一种变革性的范式，它直接利用固有的物理动力学 进行计算，而不是耗费资源来强制实现理想化的数字抽象。 通过放宽传统ASIC所需的约束，例如强制无状态性、单向性、确定性和同步性，这些设备旨在作为物理过程的精确实现而运行，从而在 能源效率和计算吞吐量方面获得显著提升。这种方法能够实现新颖的协同设计策略，使算法需求与物理系统固有的计算原语相吻合。 基于物理的ASIC可以加速关键的AI应用，例如扩散模型、采样、优化和神经网络推理，以及材料和分子科学模拟等传统计算负载。最 终，这一愿景指向了一个异构、高度专业化的计算平台未来，它能够克服当前的扩展瓶颈，并开启计算能力和效率的新前沿。 一、引言：计算危机 在过去十年中，人工智能（AI）应用的快速扩展显著增加了对计算基础设施的需求，暴露了基础硬件范式中的关键限制。支撑AI模型的 基础设施从未考虑到今天的规模、复杂性或能源需求。因 ...

优胜奖得主： • Dr. Aditya Nair，加州理工学院和斯坦福大学的博士后研究员及 NIH NeuroAI 项目青年学者，专注于 AI 与神经科学的融合研究 • Dr. Alizée Roobaert，比利时佛兰德海洋研究所（Vlaams Instituut voor de Zee）研究员，开发了监测海 洋气候动态的创新 AI 解决方案 天桥脑科学研究院（Tianqiao and Chrissy Chen Institute）与美国科学促进会（American Association for the Advancement of Science, AAAS）今天联合宣布，首届"天桥脑科学研究院与《科学》杂志 AI 驱动科 学大奖"获奖名单正式公布。这一重要年度奖项旨在表彰利用 AI 驱动科学发现的创新性研究。三位获奖 者将分享总计 5 万美元的现金奖励，其获奖研究论文也将在《科学》杂志上发表。 大奖得主： • Dr. Zhuoran Qiao，机器学习科学家，旧金山 Chai Discovery 公司创始科学家，因其在生物化学领域运 用 AI 的突破性工作获得大奖。 奖项设置与后续活动 大奖得主 Dr ...

文 | 追问nextquestion 当以ChatGPT为代表的许多大语言模型，能够实现相对准确地预测大脑对语言任务的反应时，是否可以 认为大语言模型捕捉到了大脑语言认知加工的一些深层机制？换言之，大脑也采用类似大语言模型的预 测编码机制——不断预测并修正错误？ 这种推论是否经得起科学的检验？GPT的预测与人脑语言反应的高度相关，究竟是"认知本质"，还是 只是"统计上的巧合"？ 01 预测编码理论 在20世纪，我们认为大脑从感官中提取知识。21世纪则见证了一场"奇怪的反转"，大脑被视为一个推理 的器官，会主动地为外部世界发生的事情构建解释[1]。在这场转变中，预测编码（Predictive coding） 理论扮演了重要角色。 20世纪90年代，心理学家Karl Friston提出了预测编码理论，提供了一个关于大脑如何加工的高层次描 述。该理论认为，大脑在未来事件发生之前就在不断地尝试对其进行预测，然后将预测与观测进行比 较，当预测与实际的感官输入不匹配时，大脑会对预测进行调整与更新以减少这种预测误差 （prediction error）。作为一种认知理论，预测编码理论为大脑信息加工提供了一种概念简洁、机制合 ...

Core Insights - The Guangdong Province is accelerating the development of a modern industrial system, focusing on emerging industries such as artificial intelligence and robotics [2] - Brain-inspired intelligence (BI) is highlighted as a significant development direction for the next generation of AI, with disruptive potential in technology and industry transformation [4] Group 1: Brain-Inspired Intelligence - Brain-inspired intelligence is a new path to overcome AI technology bottlenecks, drawing inspiration from brain science and neuroscience [4] - The core advantages of brain-inspired intelligence include low power consumption, self-evolution, and small sample learning, which can lead to breakthroughs in computation speed, energy consumption, and reasoning capabilities [4] - The broad scope of brain-inspired intelligence encompasses areas such as brain-like vascular systems and organ engineering, with applications in simulating brain functions and disease mechanisms [4] Group 2: Research and Development Progress - The Guangdong Provincial Institute of Intelligent Science and Technology has made significant research advancements in brain-inspired intelligence, promoting comprehensive innovation from algorithms to chips [5] - The team has developed the "Intuitive Neural Network" (INN), which integrates symbolic computation with data-driven approaches, achieving both interpretability and high energy efficiency [5] - The "Tianqin Chip" BPU processor and the "Tianqin·Hai" wafer chip support brain-like computing systems, achieving breakthroughs in speed and power consumption [5] Group 3: Industrial Ecosystem and Recommendations - The brain-inspired intelligence ecosystem is categorized into four levels: core research institutions, infrastructure companies (chip and computer manufacturers), hardware enterprises in the supply chain, and application-end companies [5] - Brain-like computing technology can drive the miniaturization and high integration of intelligent terminals, activating emerging industries such as chip manufacturing and brain-machine medical applications [5] - Recommendations include accelerating the construction of heterogeneous fusion supercomputing centers, promoting energy-saving technologies, and focusing on the development of brain-like chips, sensors, and distributed intelligent computing networks [5]

Core Insights - Chipone Technology has successfully established itself as a leading player in the semiconductor IP industry, being recognized as "China's first semiconductor IP stock" after its listing on the STAR Market in August 2020 [1][2] - The company has achieved significant milestones, including being included in the STAR 50 index and successfully completing a refinancing issuance under the "light asset, high R&D investment" recognition standard [1][3] - As of the end of Q1 2025, Chipone's order backlog reached a record high of 2.456 billion yuan, maintaining a high level for six consecutive quarters, indicating strong demand and growth potential [1][3] Company Development - Chipone Technology was founded in 2001 and has evolved from providing standard cell libraries to offering comprehensive chip customization services and semiconductor IP licensing [2] - The company initially aimed for a NASDAQ listing but chose to return to the Chinese capital market, benefiting from the reforms and growth prospects of the STAR Market [2] R&D and Market Position - Chipone ranks first in China and eighth globally in semiconductor IP licensing revenue as of 2024, with a strong focus on high R&D investment, maintaining over 30% of revenue allocated to R&D from 2020 to 2024 [3][4] - The company has successfully developed 5nm system-on-chip (SoC) technology and is executing multiple projects in the 4nm/5nm range, showcasing its advanced design capabilities [3] Strategic Growth and Financing - The recent A-share private placement marks a significant step for Chipone, allowing for more flexible fund allocation towards R&D and innovation, which is crucial for maintaining competitive advantages [4][5] - The company is also exploring mergers and acquisitions to enhance its industry position and foster ecosystem development, leveraging its status as a leading semiconductor IP provider [5][6] AI and Chiplet Technology - The rise of artificial intelligence has created substantial demand for high-performance chips, particularly in the AI ASIC sector, where Chipone has made significant advancements [6][7] - Chipone's neural network processor (NPU) IP has been adopted in 142 AI chip models across various sectors, including servers and automotive, with over 100 million units shipped [6][7] - The company has been proactive in developing Chiplet technology, which allows for modular integration of chips, enhancing performance while balancing costs, particularly in cloud AI and high-end automotive applications [8]

中集安瑞科投资控股（深圳）有限公司，成立于2010年，位于深圳市，是一家以从事科技推广和应用服 务业为主的企业。企业注册资本8000万美元。通过天眼查大数据分析，中集安瑞科投资控股（深圳）有 限公司共对外投资了25家企业，参与招投标项目1次，财产线索方面有商标信息103条，专利信息941 条，此外企业还拥有行政许可3个。 本文源自金融界 天眼查资料显示，中集安瑞科工程科技有限公司，成立于2001年，位于南京市，是一家以从事建筑装 饰、装修和其他建筑业为主的企业。企业注册资本11000万人民币。通过天眼查大数据分析，中集安瑞 科工程科技有限公司共对外投资了2家企业，参与招投标项目262次，财产线索方面有商标信息3条，专 利信息111条，此外企业还拥有行政许可16个。 中国国际海运集装箱（集团）股份有限公司，成立于1980年，位于深圳市，是一家以从事金属制品业为 主的企业。企业注册资本539252.0385万人民币。通过天眼查大数据分析，中国国际海运集装箱（集 团）股份有限公司共对外投资了23家企业，参与招投标项目54次，财产线索方面有商标信息155条，专 利信息4903条，此外企业还拥有行政许可25个。 金融界 ...

Core Viewpoint - Tesla's stock opened down 7.56% on July 7, with a closing price of $291.51 per share and a market capitalization of $938.94 billion, reflecting a significant decline in revenue and net profit for the fiscal year ending March 31, 2025 [1][2]. Financial Performance - As of March 31, 2025, Tesla reported total revenue of $19.335 billion, a year-over-year decrease of 9.23% [1]. - The net profit attributable to shareholders was $409 million, representing a substantial year-over-year decline of 70.58% [1]. Analyst Ratings and Future Reports - On July 3, HSBC reaffirmed a "Reduce" rating for Tesla, raising the target price to $120 [2]. - Tesla is scheduled to disclose its fiscal year 2025 mid-term report on July 23, 2023, after market hours [2]. Company Overview - Tesla, founded on July 1, 2003, by Martin Eberhard and Marc Tarpenning, is an American electric vehicle and energy company [2]. - The company designs, develops, manufactures, sells, and leases high-performance all-electric vehicles and energy generation and storage systems, providing related services [2]. - Tesla is recognized as the world's first vertically integrated sustainable energy company, offering end-to-end clean energy products, including generation, storage, and consumption [2]. Product Line and Technological Advancements - Tesla is planning to launch electric vehicles to cater to a broad consumer and commercial vehicle market, including models such as Model 3, Model Y, Model S, Model X, Cybertruck, Tesla Semi, and a new Tesla Roadster [2]. - The electric vehicles feature advanced power systems, autonomous driving capabilities, and Full Self-Driving (FSD) hardware, providing advantages in range, charging flexibility, acceleration, handling, safety, and user-friendly infotainment features [2].

Group 1 - The market cannot be precisely predicted, and the goal is to build a cognitive framework to understand its current state and infer short-term evolution [1] - Traditional technical analysis attempts to reduce the complexity of market processes but often overlooks the high-dimensional latent space that drives price movements [1] Group 2 - Early deep learning models like CNNs capture local spatial patterns but fail to understand the path dependency of time series data [2] - LSTM and its variants address the limitations of CNNs by capturing sequential dependencies, but they assume a linear flow of information, which does not reflect the complex interactions in real markets [3] Group 3 - A paradigm shift is needed from sequential dependency modeling to spatio-temporal structural dependency modeling to better capture market dynamics [5] - The core of the proposed approach is a dynamic temporal knowledge graph that models relationships among entities, which is essential for understanding market interactions [6] Group 4 - The use of heterogeneous Hawkes processes allows for modeling event flows within the knowledge graph, capturing the ripple effects of market events [6] - By maximizing the log-likelihood function, the system can derive embedding vectors for entities and relationships, projecting the knowledge graph into a lower-dimensional latent space [7] Group 5 - The model's output is a posterior probability that combines likelihood from data and prior probability based on human insights, emphasizing the importance of human judgment in the decision-making process [9][10] - The company aims to create a decision framework that optimizes long-term expected value rather than focusing on short-term gains, leveraging the cognitive spread between its insights and market averages [11]