鹭羽 发自 凹非寺 量子位 | 公众号 QbitAI 开年，DeepSeek论文火遍全网，内容聚焦 大模型 记忆 。 无独有偶，谷歌近期也发布了一篇被誉为"Attention is all you need"V2 （Nested Learning: The Illusion of Deep Learning Architectures） 的重磅论文，核心同样指向记忆瓶颈。 就连最近这只彻底破圈的AI大龙虾——OpenClaw （原名Clawdbot） ，其亮点之一依旧是记忆。 也就是说， 记忆 ≈今年全球AI圈集体押注的技术风口≈皇冠明珠。 几乎所有你能想到的大模型团队，都在加班加点往自家模型里塞记忆功能…… 但这一次，让我们把视线从这些科技巨头身上稍稍挪开，就会发现有一支后起之秀同样不容小觑。 他们就是陈天桥和邓亚峰带队的 EverMind 。 为什么这样说呢？ 且看产品，最新发布世界级长期记忆系统—— EverMemOS ，发布即SOTA。 一举打破多项记忆基准测试的同时，还能远超此前所有的基线方法。 其次，它是真正能用的。 不是只会跑测试的"花架子"，实际部署后效果照样能打。而且团队有底气有信心，技术代 ...

Core Insights - Shanda Group's Shanda AI Research Tokyo made its debut at SIGGRAPH Asia 2025, focusing on "Interactive Intelligence" and "Spatiotemporal Intelligence" in digital human research, reflecting the long-term vision of founder Chen Tianqiao [1][10] - The article discusses the systemic challenges leading to the "soul" deficiency in current digital human interactions, which is a significant barrier to user engagement despite substantial investments in visual effects [2][3] Systemic Challenges - **Long-term Memory and Personality Consistency**: Current large language models (LLMs) struggle with maintaining a stable personality over extended conversations, leading to "persona drift" and inconsistent narrative logic [3] - **Lack of Multimodal Emotional Expression**: Digital humans often exhibit "zombie-face" phenomena, lacking natural micro-expressions and emotional responses, which diminishes immersive experiences [3] - **Absence of Self-evolution Capability**: Most digital humans operate as passive systems, unable to learn from interactions or adapt to user preferences, hindering their evolution into truly intelligent entities [3] Industry Consensus - Experts at the SIGGRAPH Asia conference reached a consensus that the bottleneck in digital human development has shifted from visual fidelity to cognitive and interaction logic, emphasizing the need for long-term memory, multimodal emotional expression, and self-evolution as core competencies [13][10] Introduction of Mio - Shanda AI Tokyo Research introduced Mio (Multimodal Interactive Omni-Avatar), a framework designed to transform digital humans from passive entities into intelligent partners capable of autonomous thought and interaction [16][22] - Mio's architecture includes five core modules: Thinker (cognitive core), Talker (voice engine), Facial Animator, Body Animator, and Renderer, which work together to create a seamless interaction loop [20][21] Performance Metrics - Mio achieved an overall Interactive Intelligence Score (IIS) of 76.0, representing an 8.4 point improvement over previous technologies, setting a new performance benchmark in the industry [25][22] Future Outlook - The development of Mio signifies a paradigm shift in digital human technology, moving focus from static visual realism to dynamic, meaningful interactive intelligence, with potential applications in virtual companionship, interactive storytelling, and immersive gaming [22][25] - Shanda AI Tokyo Research has made the complete technical report, pre-trained models, and evaluation benchmarks of the Mio project publicly available to foster collaboration in advancing this field [28]

Core Insights - The article highlights significant developments in the AI industry, including new startups, funding rounds, and technological advancements, reflecting the rapid growth and investment in AI technologies globally. Group 1: AI Startups and Innovations - Jeff Bezos has founded a new AI startup named "Project Prometheus," which has raised $6.2 billion in funding and employs nearly 100 staff from major tech companies [3] - The humanoid robot company, UBTECH, has responded to allegations that its delivery video for the Walker S2 robot was CGI, asserting that the video was filmed live during the delivery of hundreds of robots [3] - Tencent has open-sourced its latest video generation model, HunyuanVideo 1.5, which is a lightweight model capable of generating high-definition videos [8] Group 2: Funding and Valuations - Genspark has completed a $275 million Series B funding round, bringing its valuation to $1.25 billion [7] - The "Dark Side of the Moon" project is planning an IPO in the second half of next year, with a post-funding valuation expected to reach approximately $4 billion [6] - Databricks is in talks for a new funding round that could exceed a valuation of $130 billion, reflecting strong interest in AI infrastructure despite concerns about a potential bubble [15][29] Group 3: Market Trends and Predictions - The global 3D printing market is projected to reach $24.61 billion in 2024, with a compound annual growth rate of 18.5% from 2024 to 2034, indicating a robust growth phase for the industry [7] - AI mobile phones are expected to see shipments challenge 600 million units by 2026, driven by the expansion of 5G networks and the replacement of feature phones in emerging markets [32] - The AI sector is witnessing a significant shift, with predictions that generative AI could eliminate up to 50% of entry-level white-collar jobs, potentially raising unemployment rates to 10%-20% [12] Group 4: AI Financing Overview - A total of 18 AI financing events were disclosed globally this week, with a total financing scale of 66.75 billion RMB, averaging 3.93 billion RMB per event [36] - In the domestic market, the disclosed financing total reached 3.125 billion RMB, with the highest amount coming from the humanoid robot developer, Star Motion Era, which completed a 900 million RMB A+ round [44] - Internationally, the disclosed financing total was 63.629 billion RMB, with the highest being the $6.2 billion seed round for Project Prometheus [54]

Core Insights - EverMind has launched its flagship product EverMemOS, a world-class long-term memory operating system for AI agents, which has been released as an open-source version on GitHub for developers and AI teams to deploy and test [1] - The cloud service version is expected to be released within the year, providing enhanced technical support, data persistence, and scalability for enterprise users [1] - EverMemOS has surpassed previous works in mainstream long-term memory evaluation sets, becoming the new state-of-the-art (SOTA) [1][4] Group 1: Product Features and Innovations - EverMemOS is designed based on a brain-like architecture, allowing AI to possess continuity over time, addressing the limitations of large language models (LLMs) that often "forget" during long-term tasks [3][4] - The system features a four-layer architecture inspired by human memory mechanisms, including an agent layer for task understanding, a memory layer for long-term memory management, an indexing layer for efficient memory retrieval, and an interface layer for seamless integration with enterprise applications [6][7] - Key innovations include a modular memory framework that allows for dynamic organization and retrieval of memories, ensuring that AI interactions are coherent and personalized based on long-term user understanding [7] Group 2: Performance Metrics - EverMemOS achieved scores of 92.3% and 82% on the LoCoMo and LongMemEval-S long-term memory evaluation sets, respectively, significantly exceeding the previous SOTA levels [4][6] - The system is the first to support both one-on-one conversations and complex multi-party collaborations, marking a significant advancement in memory systems for AI applications [4]

Core Viewpoint - The EverMind team, founded by Chen Tianqiao, has launched EverMemOS, a world-class open-source long-term memory system for AI agents, aiming to establish a data infrastructure for future intelligent systems [3][4]. Group 1: Long-term Memory in AI - Long-term memory is seen as a strategic feature for AI applications, with Chen Tianqiao asserting that memory will be a core competitive advantage for AI, transitioning from tools to intelligent agents [3][4]. - The design of EverMemOS is inspired by human memory mechanisms, allowing AI to think, remember, and grow like humans [3][5]. Group 2: Discovering Intelligence - Chen Tianqiao emphasizes the concept of "discovery intelligence," which includes five core capabilities, with long-term memory being a crucial aspect [4][5]. - He argues that true discovery involves asking questions and understanding principles rather than merely predicting outcomes [4][5]. Group 3: Paths to Discovery Intelligence - Two paths to achieving discovery intelligence are proposed: the "scale path," which focuses on large models and data, and the "structure path," which emphasizes dynamic and continuous understanding akin to cognitive anatomy [5][6]. - The limitations of the scale path are acknowledged, with a call for new theories and interdisciplinary approaches to advance AI understanding [5][6]. Group 4: Investment in Research - The Tianqiao Brain Science Research Institute plans to invest over $1 billion in dedicated computing clusters to support young scientists in exploring memory mechanisms and new causal architectures [6][7]. - An incubator has been established to provide independent research opportunities for young scientists, allowing them to lead their own projects without waiting for graduation [7].

Core Insights - The concept of "memory" is viewed as a critical competitive advantage for future AI applications, marking a transition from tools to intelligent agents and from passive responses to proactive evolution [1][3] - EverMind, a team under Chen Tianqiao's Shanda Group, has launched EverMemOS, a world-class open-source long-term memory system aimed at becoming the data infrastructure for future intelligent agents [1][3] - Chen Tianqiao emphasizes the importance of "discovery-based intelligence," advocating for a research paradigm that endows AI with a persistent, coherent, and evolvable "soul" [1][3] AI Memory Systems - Current AI applications, including Claude and ChatGPT, have integrated long-term memory as a strategic feature, reflecting a broader trend in the industry [3] - EverMemOS is inspired by human memory mechanisms, mimicking the encoding, indexing, and long-term storage processes of the human brain [3][4] - The EverMind team aims to tackle one of the most profound challenges in AI: enabling machines to possess memory, which is seen as a gateway to higher levels of general intelligence [3][4] Discovery-Based Intelligence - Chen Tianqiao argues that true "discovery" in AI should involve posing questions and understanding principles rather than merely predicting outcomes [4] - He outlines two pathways to achieving "discovery-based intelligence": the "scale pathway," which focuses on the size of models and data, and the "structure pathway," which emphasizes dynamic and continuous cognitive processes [4][5] - The "scale pathway" has led to significant applications in AI, such as protein prediction and compound generation, but is viewed as limited in its ability to foster true understanding [5] Research and Development Initiatives - The Tianqiao Brain Science Research Institute plans to invest over $1 billion in dedicated computing clusters to support young scientists in exploring structural intelligence [6] - The institute has established a PI incubator, allowing young researchers to have independent budgets and lead their own experiments without waiting for graduation [6] - Chen Tianqiao stresses the need for interdisciplinary collaboration among neuroscience, information theory, physics, and cognitive psychology to drive innovation in AI [5][6]

Core Insights - EverMind has launched its flagship product EverMemOS, a world-class long-term memory operating system designed for artificial intelligence agents, enabling real-time and proactive memory influence on AI responses [1] - The model has achieved significant breakthroughs in both technical performance and scenario coverage, surpassing state-of-the-art levels in memory evaluation [1] Technical Performance - EverMemOS utilizes innovative biological "Engram" heuristic memory extraction and application technology, scoring 92.3% and 82% on the leading long-term memory evaluation sets LoCoMo and LongMemEval-S, respectively, exceeding state-of-the-art benchmarks [1] Scenario Coverage - The system supports both one-on-one conversations and complex multi-party collaboration, with its first application in the AI Native product Tanka [1] Open Source and Future Plans - EverMind has released an open-source version of EverMemOS on GitHub for developers and AI teams to deploy and test [1] - A cloud service version is expected to be launched later this year, providing enhanced technical support, data persistence, and scalable experiences for enterprise users [1]

Group 1: Apple CEO Transition - Tim Cook, the CEO of Apple, may step down as early as next year, with the company actively seeking a successor [1] - Hardware engineering senior vice president, John Ternus, is widely viewed as the most likely successor, aligning with Apple's shift towards hardware technology [1] - Cook has served as CEO since 2011, marking 14 years in the role, while Ternus is 50 years old, similar to Cook's age when he became CEO [1] Group 2: Xiaomi's CEO Response - Lei Jun, CEO of Xiaomi, has issued multiple clarifications regarding his comments on car design, emphasizing that aesthetics do not compromise safety [1] - His statements were made prior to the launch of the Xiaomi SU7, addressing concerns about prioritizing design over safety standards [1] Group 3: Flying Cars and Market Growth - The world's first "ten-thousand-unit" flying car factory has begun trial production in Guangzhou, capable of producing one flying car every 30 minutes [7] - The low-altitude economy is rapidly emerging, with over 2,000 manufacturing companies in Guangdong, accounting for more than 30% of the national total [7] - The market for electric vertical takeoff and landing (eVTOL) vehicles in China is projected to reach $41 billion by 2040 [7] Group 4: Microsoft Stake Reduction - The Bill and Melinda Gates Foundation Trust sold 17 million shares of Microsoft, reducing its stake by nearly 65%, amounting to approximately $8.7 billion [11] - This sale is considered a portfolio adjustment rather than a reflection of Microsoft's fundamental performance [11] Group 5: Investment in Innovation - Hyundai Motor Group plans to invest 125.2 trillion KRW (approximately 611.1 billion RMB) in South Korea over the next five years, focusing on AI, software-defined vehicles, and hydrogen energy [11] - GAC High Domain has completed a 200 million RMB Pre-A round of financing, aiming to advance its flying car development [12] Group 6: AI and Robotics - The next decade is expected to see robots evolve from industry tools to "life partners," with advancements in AI enabling them to understand the world better [4] - Elon Musk predicts that within 20 years, human consciousness may be uploaded to robots, potentially achieving a form of digital immortality [4] Group 7: Fashion Industry Growth - The global apparel market is projected to reach $1.84 trillion by 2025, representing 1.6% of global GDP [14]

Group 1 - Huawei is set to release a significant AI technology aimed at improving GPU utilization efficiency, potentially increasing the utilization rate from the industry average of 30%-40% to 70% [2] - The EverMind team from Shanda Group has launched EverMemOS, a world-class long-term memory operating system for AI agents, inspired by human memory mechanisms and featuring a modular memory framework [2] - A robot sign language interpreter equipped with China's first sign language interaction system has been showcased, allowing robots with dexterous hands to communicate using sign language [2][3] Group 2 - The sign language library developed by Harbin Institute of Technology (Shenzhen) includes over 5,000 manually recorded sign language entries, combining it with dexterous robotic hands to enhance interaction capabilities [3]

Group 1 - Apple CEO Tim Cook may resign as early as next year, with hardware engineering senior vice president John Ternus seen as a likely successor [1] - Xiaomi's CEO Lei Jun responded to criticism regarding his comments on car aesthetics versus safety, clarifying his stance on product design and safety standards [2][5] - The Shaolin Temple's former abbot, Shi Yongxin, has been arrested for embezzlement and other charges, highlighting issues of financial misconduct within religious institutions [5] Group 2 - The global first "ten-thousand-unit" flying car factory has begun trial production in Guangzhou, with a production rate of one flying car every 30 minutes, indicating rapid advancements in low-altitude economy [9] - The Bill and Melinda Gates Foundation has reduced its stake in Microsoft by 65%, cashing out approximately $8.7 billion, which is seen as a portfolio adjustment rather than a negative outlook on Microsoft's fundamentals [13][14] - Hyundai Motor Group plans to invest over 610 billion yuan in domestic innovation over the next five years, focusing on AI, software-defined vehicles, and hydrogen energy [14] Group 3 - The global apparel market is projected to reach $1.84 trillion by 2025, accounting for 1.6% of global GDP, emphasizing the industry's strategic importance [18] - Cogent Biosciences has completed a $475 million financing round to support clinical trials and early pipeline development, indicating strong investor interest in biotech [15] - GAC's flying car brand, Gaoyu, has completed a 200 million yuan Pre-A round of financing, reflecting growing investment in the flying vehicle sector [15]