ChatGPT在2022年底把生成式AI推到大众视野后，投资端的变化更快：企业在AI硬件、数据中心上的投入力度，已经接近美国历史上几次最大的 投资浪潮。市场随之抛出一堆漂亮的收入曲线，但问题也变得尖锐——这些预测到底有多大概率能实现，值不值得为此付出资本和时间成本？ 据追风交易台消息，摩根士丹利投资管理旗下Counterpoint Global的Michael J. Mauboussin在10日的报告中直截了当地给出方法论：评估这类前瞻 判断，应该"starting with an initial belief and updating that belief as new results appear"，也就是"贝叶斯公式"："新结论 = 初始判断（先验概率） × 新 证据带来的调整系数（似然比）"。 | New and improved belief | Recent objective data + Initial belief | ll | | | --- | --- | --- | --- | | Recent | New and | improved | objective Initial | | ...

从全年走势来看，去年下半年投融资的"回暖"更加明显。去年12月，张江以29起投融资创下全年新高。从平均来看，每月发生20起投融资，如以工作日 计，每天都有投融资在张江发生。 PART 01 2025张江科创成果多点爆发 据张通社联合加冕研究院发布的《2025年张江企业融资榜单》显示，上海去年每四起科创融资，就有一起发生在张江。过去一年张江共有210家张江企业 完成237起融资，披露融资金额278亿元，投融资事件数量和融资金额同比增长30%和18%。张江融资规模约占浦东65%，约占上海全市的25%。 2025年张江生物医药融资较2024年增长13%，主要融资方向是原创能力强、临床路径清晰，能够真正产品化的企业。具体来看，创新药领域有58起融资， 占整个生物医药赛道融资的58%，医疗器械22起，占比26%。比如，光声制药研发的新一代光敏化合物"华卟啉钠"，成为全球首个在食管癌光动力治疗领 域取得III期临床成功的光敏剂，距离新药上市申报，仅一步之遥。 2025年张江人工智能融资相较2024年增长146%，融资主要集中在AI应用和机器人与具身智能两大方向。"去年获得融资的傅利叶、开普勒和木蚁机器人都 面向真实场景推 ...

机器之心编辑部 马斯克于 2023 年与另外 11 位联合创始人共同创办的 xAI，如今已有 6 人离开。 最新消息，xAI 联合创始人 Jimmy Ba 周二表示，他已经离开了这家 AI 初创公司。 Jimmy 写道：这是我在 xAI 的最后一天。xAI 的使命是推动人类提升卡尔达舍夫等级（Kardashev tech tree）。我非常荣幸能在公司创立之初共同参与这一历程。由 衷感谢 @elonmusk 将我们聚集在一起，开启了这段不可思议的旅程。我为 xAI 团队所取得的成就感到无比自豪，未来也将作为团队的朋友继续保持紧密联系。感 谢大家一起并肩奋斗的岁月，和同事情谊才是我真正的财富。2026 年注定将是疯狂的一年，也极有可能是人类物种未来最繁忙（且最具影响力）的一年。 据知情人士透露，Jimmy 直接向马斯克汇报工作，并在去年年底之前一直负责公司相当大的一部分业务。直到去年晚些时候，他的多项职责被拆分，分别交由另 外两位联合创始人吴宇怀（Tony Wu，已经官宣离职）和 Guodong Zhang 负责。 Jimmy 是在不到 48 小时内第二位离开 xAI 的联合创始人。吴宇怀已于周一晚间宣布从这家 ...

或许 AI 的事还是用 AI 来解决，今年春节的安排，必须得上「深度研究」了。 说来也巧，我们发现美团大模型 LongCat 刚刚上线了这个功能。 编辑 | 泽南、杨文 春节还没到，「过年的气氛」已经渗入科技圈每个人的毛孔。单说 AI 大模型这一块，刚刚发布的有 kimi 2.5 和 Step 3.5 Flash，即将发布的据说还有 DeepSeek V4，GPT-5.3、 Claude Sonnet 5、 Qwen 3.5，GLM-5，说不定一觉醒来，现有的技术就要被颠覆。 再看看千问和元宝发的红包，组团上春晚的机器人，所有厂商在春节期间都摆出一副志在必得的架势。 正因为如此，我们在这个临近长假的时间段，又遭遇了很多的变数。很多人的行程调了又调，不到最后一刻不知道哪几天有空，难不成这场 AI 的革命，反 而会让我们忙到忘了过节？ 这一次的 「深度研究」（DeepResearch）上，LongCat 又把实用化推向新的高度。 LongCat（龙猫）系列大模型，最近才火起来的，在 AI 圈里它的口碑主要是 速度快，工具调用出色，走「实用主义」路线 。 正好，我们就用它来安排一下春节的事儿。 值得注意的是，该功能 ...

Core Viewpoint - The article discusses the rapid rise of Seedance 2.0 and the ensuing copyright concerns, particularly highlighted by the backlash from copyright holders like Stephen Chow's representatives, indicating a growing anxiety within the content industry regarding the unchecked growth of AI technologies [3][6][10]. Group 1: Seedance 2.0 and Its Features - Seedance 2.0 allows users to easily create videos featuring iconic characters from popular IPs, such as those from Stephen Chow's films, by simply inputting basic materials [4][11]. - The generated videos are noted for their high quality, fluid movements, and expressive features, effectively capturing the essence of the original characters [5][11]. - The platform's appeal lies in its "unrestricted" creative freedom, enabling users to generate content that heavily relies on well-known IPs [11][12]. Group 2: Copyright Issues and Industry Response - The surge in AI-generated content has led to significant copyright concerns, with Stephen Chow's management publicly questioning the legality of the widespread use of his characters [6][10]. - The article emphasizes that Seedance 2.0 lacks adequate copyright protection measures, allowing users to generate videos featuring characters from major franchises like Disney and Dragon Ball without authorization [17][18]. - The content generated by Seedance 2.0 poses a higher risk of copyright infringement compared to static images, as the videos can closely replicate the original characters' actions and expressions [26][28]. Group 3: Legal Precedents and Implications - Disney has been proactive in protecting its IP, having filed lawsuits against various AI companies for unauthorized use of its characters, highlighting a broader concern for the film industry regarding AI's impact on copyright [21][24]. - The legal strategies employed by Disney include seeking economic compensation and establishing industry standards for copyright protection in AI technologies [24][25]. - The article suggests that Seedance 2.0 may face similar legal challenges as other AI platforms, given its ability to generate content that closely resembles copyrighted material [26][30]. Group 4: Potential Paths Forward for ByteDance - The article outlines several potential strategies for ByteDance, including litigation and settlement, proactive licensing agreements, or implementing technical measures to avoid copyright infringement [47][49]. - A historical perspective is provided, noting that ByteDance has previously navigated copyright disputes, but the current landscape may present more complex challenges due to the nature of AI-generated content [40][42]. - The company is advised to consider proactive engagement with copyright holders rather than waiting for legal action, as the stakes in the current environment are significantly higher [60][61].

Core Insights - ByteDance's AI video generation model Seedance 2.0 has entered internal testing, being hailed as the strongest video generation model currently available, marking the end of the "childhood era" of AIGC [1] - Following the announcement, the A-share media sector experienced a significant surge, with the cultural media sector rising by 4.79% on February 9, 2026, and multiple stocks hitting their daily limit [1] - Seedance 2.0 features a dual-branch diffusion transformer architecture, capable of generating videos and audio from a single image or detailed prompt within 60 seconds, with unique multi-shot narrative capabilities [1] Technology and Market Impact - Despite the technological advancements, concerns arose regarding the model's ability to replicate personal audio and visuals without authorization, leading to a temporary suspension of certain features [2] - Analysts remain optimistic about Seedance 2.0's potential to accelerate the development of the AI multimodal industry chain, with predictions of a shift from manual to automated content production [2] - The model is expected to significantly reduce the cost of AI video generation, with estimates suggesting a 50% decrease in "抽卡" frequency and a 37% reduction in per-second generation costs compared to competitors [3] Industry Outlook - The capabilities of Seedance 2.0 are seen as a transformative leap in video generation, allowing ordinary users to achieve a "director-level" creative experience while lowering production barriers and costs [3] - The release of Seedance 2.0 is viewed as a sign that the AI application industry cycle is entering a growth phase, with major technological advancements acting as key catalysts for sectors like AI animation, film IP, and data elements [3]

Group 1 - The 2026 Zhangjiang Sci-Tech Financial Salon was held in Shanghai Zhangjiang Science City, focusing on enhancing Zhangjiang's role as a core engine for global innovation and capital aggregation [1] - Zhangjiang Science City is concentrating on three key sectors: artificial intelligence, biomedicine, and integrated circuits, with significant financing achievements in 2025 [1] - In 2025, 210 companies in Zhangjiang completed 237 financing events, totaling 27.8 billion yuan, with a 30% increase in the number of financing events and an 18% increase in disclosed amounts year-on-year [2] Group 2 - Zhangjiang accounted for 65% of the total financing events in the Pudong district and contributed 25% of the total corporate financing in Shanghai [2] - A total of 512 investment institutions participated in Zhangjiang's financing activities, indicating a more diversified capital structure [2] - The financing for the leading sectors (biomedicine, artificial intelligence, and integrated circuits) reached 91% of the total, up from 82% in 2024 [2] Group 3 - The financing stages in Zhangjiang showed a clear characteristic of "early, mid, and late-stage collaboration," with early-stage financing accounting for 64% of total financing [3] - 273 companies were awarded for their financing impact in 2025, covering the core sectors of artificial intelligence, biomedicine, and integrated circuits [3] - Zhangjiang aims to align with national strategies and enhance its innovation ecosystem to contribute to China's technological self-reliance and strength [3]

此外，上海还将调整优化基础研究组织模式，依托尚思研究院，进一步深化基础研究先行区建设，完善选题为基础的选人机制，健全长周期稳定资 助和长周期成果评价机制，促进多元投入，加强非共识项目识别遴选，优化"微环境"，营造有利于基础研究的科研文化和创新生态。 加快形成"产业出题、科技答题"的协同研发模式。 随着上海国际科创中心迈入"强功能"阶段，未来五年上海将继续增强国际科技创新中心策源功能。 在2月11日举行的上海市"十五五"规划主题系列市政府新闻发布会上，上海市科委主任骆大进介绍，到2030年，上海将力争实现基础研究经费支出占 全社会研发经费支出比重达15%左右，产出一批标志性原创性成果，基础研究和原始创新能力得到进一步增强。 刚过去的2025年，上海全社会研发经费支出相当于全市生产总值的比例达到4.5%左右；基础研究投入强度预计12%左右。 关于增强基础研究和原始创新能力，骆大进介绍称，主要从强化基础研究系统布局、发挥战略科技力量支撑作用、畅通基础研究成果转化链条、深 化基础研究体制机制改革几个方面着手。 比如，前瞻性、战略性需求导向，聚焦集成电路、生物医药、人工智能三大先导产业科学问题，聚焦前沿科技和未来产业" ...

国泰海通发布研报称，近期，字节旗下的即梦团队发布Seedance2.0全新视频生成大模型，真正实现了 从"能生成"到"能商用"的跨越式突破，该模型首次实现了文字理解与字幕动效生成，可自动解析参考图 中的文字并添加合理动态效果。此进展标志着AI正从"单模态理解"向"全双工连续感知"与"跨模态深度 创作"跃迁。Seedance2.0的自主创作能力不仅重塑了内容生产与交互模式，更催生了涵盖视频生成、实 时交互、设计工具、端侧智能等环节的全新产业链投资机会。该行看好AI应用加速落地的长期发展空 间。 国泰海通主要观点如下： 根据极客公园测算，Seedance2.0生成15秒视频的可用率或达到90%，相较此前行业内平均值大概20%， 提升幅度较大。当生成的视频可用率提升后，可降低实际成本，以做90分钟的视频项目为例，成本有望 从1万多降低到2000元左右，将提升行业的使用量，其巨大的成本压缩，或可改变整个行业的底层逻 辑。 Seedance2.0标志着AI视频生成从"技术可行"向"商业可用"的关键跨越 过去一年，视频生成领域已完成从512像素静态图像到10秒电影级短片的代际跃升。此次升级的多镜头 叙事与角色一致性保障能 ...

Core Insights - The research team led by Professor Peng Yuxin from Peking University has made significant advancements in fine-grained visual recognition using multi-modal large models, with their latest paper accepted at ICLR 2026 and made open-source [1][19]. Group 1: Fine-Grained Visual Recognition - The real world exhibits fine-grained characteristics, with objects often containing a rich hierarchy of categories, such as the classification of aircraft into specific models like Boeing 707, 717, and 727, with over 500 types of fixed-wing aircraft recorded globally [2]. - The Fine-R1 model aims to leverage the extensive knowledge of fine-grained subcategories contained within multi-modal large models to achieve fine-grained recognition of visual objects in open domains, overcoming the limitations of traditional methods that focus on a closed set of categories [4]. Group 2: Model Development and Methodology - The Fine-R1 model employs a two-phase approach: 1. Chain-of-thought supervised fine-tuning, which simulates human reasoning to enhance the model's inference capabilities [7]. 2. Triplet enhancement strategy optimization, which improves the model's robustness to intra-class variations and its ability to distinguish between different classes [8]. - The model demonstrates superior performance, achieving higher accuracy in recognizing both seen and unseen subcategories with only four training images per class, surpassing models like OpenAI's CLIP and Google's DeepMind's SigLIP [13][14]. Group 3: Experimental Results - Experimental results indicate that Fine-R1 outperforms various models in both closed-set and open-set recognition tasks, showcasing its effectiveness in fine-grained visual recognition [14][16]. - The model's enhancements are attributed primarily to its improved ability to utilize fine-grained subcategory knowledge rather than merely optimizing visual representations or increasing knowledge reserves [16].