在AI技术狂飙突进的时代，一场关于"可信AI"的商业竞赛正在悄然展开。2月13日，以"AI除幻"为核心业务的海致科技集团登陆港交所，上市首日股价暴涨 242%，市值突破370亿港元，成为今年港股市场最耀眼的新股。这家由百度前高管任旭阳创立的企业，凭借其独特的技术路线，在AI行业激烈竞争中开辟出 一条新赛道。 AI除幻技术瞄准的是当前行业最棘手的痛点——大模型生成的虚假信息问题。随着ChatGPT等生成式AI的普及，内容真实性危机日益凸显。海致科技通过构 建知识图谱与大模型深度融合的推理平台，有效降低AI输出中的错误率。这项技术已获得金融机构、能源企业等客户的认可，其"海致Atlas LLM图模联合推 理平台"成为国内首个解决该问题的商业化产品。 创始人任旭阳的创业履历堪称互联网行业的教科书。这位百度早期元老在2001年加入公司后，先后主导战略合作、投资并购等核心业务。2009年创立爱奇艺 并引入龚宇担任CEO，2011年又与郑朝晖共创一点资讯。2013年，他携手百度前同事史有才、胡嵩成立海致科技，最初聚焦大数据分析领域，直到2021年与 清华团队研发出高性能分布式图数据库，才正式转型AI赛道。 公司业务呈现明 ...

中化新网讯日前，中国信息通信研究院(以下简称中国信通院)正式启动首批可信AI工业智能体评估工 作。 该评估依据中国信通院人工智能研究所依托中国人工智能产业发展联盟智能体创新与应用工作组，联合 上海移动、南方电网、广西电网、中国石化、中国石油等多家企业共同编制完成的《智能体技术要求与 评估方法行业应用工业》技术规范，立足工业行业对复杂性、可靠性的要求，结合智能体的能力特性， 从技术到应用展开全面评估，覆盖基础能力、业务场景、服务应用3大能力域，共计20余个能力项。 其中，基础能力部分主要评估工业智能体在感知、认知、决策、执行等方面的基本技术能力，包含工业 数据采集、工业数据加工、机理融合、生产规划、协同控制等能力项；业务场景主要评估工业智能体场 景应用的丰富程度，包含研发设计、工艺仿真等产品研发场景，生产优化、运行维护、质量控制等生产 管理场景，供应链管理、经营管理等运营管理场景；服务应用主要评估工业智能体服务应用的成熟度， 包含业务效果、智能交互、混合部署、系统兼容、安全保障、运维监控等能力项。(吴班) ...

29日，95后顶尖AI科学家庞天宇通过小红书官宣个人已经加入腾讯混元团队，并担任首席研究科学家 （Principal Scientist，或称首席 / 主任研究员），并写明是"Tech Lead@Multimodal RL Team"，即多模态 强化学习技术负责人。 庞天宇还在帖子中为团队招聘校招、社招、实习岗位。这意味着庞天宇成为腾讯继姚顺雨之后，近期招 揽的又一位95后顶尖AI人才。 从发布招聘前后的IP变化看，这两天他刚刚从新加坡回国。 庞天宇表示，他在加入腾讯混元团队后，主要研究方向为多模态模型的强化学习（Multimodal RL）， 包括生成模型（e.g., diffusion models）和理解模型（e.g., VLMs）。 据公开信息，庞天宇是清华大学计算机系2017级直博生，师从朱军教授。主要研究方向为机器学习，特 别是深度学习及其鲁棒性的研究，并取得了一系列的研究成果。他的研究方向也与当前大模型安全性和 可靠性的发展趋势高度契合。他近期的研究工作涉及大语言模型的安全性问题，包 括"jailbreaking/cheating LLMs"（突破大语言模型限制）和多模态大语言模型(MLLMs) ...

1月25日，第三届全国人工智能应用场景创新挑战赛总决赛暨全国人工智能+应用场景创新大会(简称CICAS)在苏州市会议中心圆满落幕。 明略科技(2718.HK)联合北京大学带来的参赛项目"基于多模态大模型的品牌出海创意生成与情感链接智能平台"，继专项晋级赛特等奖之后，在全国总决赛 再获"特等奖"荣誉。 本届大赛由中国人工智能学会、苏州市人民政府、苏州大学联合主办，全国人工智能应用场景创新挑战赛组委会、姑苏区人民政府承办。以"场景驱动·数智 强国"为主题，旨在全面检验人工智能前沿创新成果，破解AI技术落地难题，释放场景创新潜能，发挥以赛促创、以赛聚智、以赛兴业的创新带动作用。 据CICAS组委会公开介绍，本届大赛共吸引超3250支团队参与角逐，全国总决赛入围项目代表了我国人工智能场景创新领域的先进水平。 全国总决赛现场集结国内外113支顶尖创新团队、350余位参赛选手。50余位两院院士、行业领军企业及知名投资机构代表亲临现场，见证AI+创新成果。 大赛组委会评审团依据严格标准，对入围项目解决方案、数据样本、核心算法及工程化产品等进行多维度实证验证，重点考察真实场景中的问题解决能力、 创新性与成熟度。此外，本届总决 ...

Core Insights - The article emphasizes the transition from merely adopting AI to effectively utilizing it, predicting that by 2026, 5 billion people will use AI daily, highlighting its evolution into a core productivity driver [1][12]. Group 1: AI Adoption Challenges - Companies face significant challenges in AI implementation, including doubts about AI's value, with 37% of enterprises expressing skepticism despite projected spending growth [2]. - Key bottlenecks hindering AI's transition from experimentation to large-scale application include uncontrollable model outputs, unreliable data sources, and inadequate security mechanisms [2][3]. Group 2: Trustworthy AI Framework - Trustworthy AI is defined as the ability to meet stakeholder expectations in a verifiable manner, with a formula proposed: Trustworthy Productivity = Trustworthy Models + Trustworthy Data [4]. - Trustworthy models require not only technical capabilities but also the ability to systematically solve complex problems through trustworthy task planning [4][5]. Group 3: Importance of Trustworthy Data - Trustworthy data is crucial for achieving trustworthy AI, with its reliability ensured through identifying credible data sources and efficiently extracting necessary information [6][7]. - The authority of data sources and the reliability of data acquisition methods are often overlooked factors that significantly impact decision quality [8]. Group 4: Data Source System - The company has established a multi-tiered, high-standard trustworthy data source system, including access to over 1,000 authoritative institutions for macroeconomic and industry data [9]. - It also integrates professional third-party data and enterprise-specific data to provide a comprehensive view of business operations [10]. Group 5: Security and Collaboration Mechanisms - The company prioritizes architecture design over functional promises, ensuring that AI systems can be deployed in a controlled environment to maintain data security [11]. - In critical business decision scenarios, human experts retain final decision-making authority, with AI serving as an efficient execution assistant [11]. Group 6: Practical Applications and Future Outlook - Successful applications of AI have been demonstrated, such as a marketing agency increasing creative material effectiveness from 30% to 70% through predictive testing [12]. - As AI technology continues to permeate industries, the competition will shift from merely having AI to possessing superior AI capabilities, making trustworthy AI a critical component of digital transformation [12].

Core Insights - The banking industry is undergoing a significant digital transformation, entering a phase characterized by data-driven decision-making and operational efficiency [3][10] - Shanghai Bank has successfully completed its "Zhixin Project," marking a new stage in its digital infrastructure with a fully autonomous core system [2][10] - The bank emphasizes the importance of integrating technology with business operations to enhance agility and responsiveness [4][11] Digital Transformation Progress - The digital transformation in the banking sector has reached a "deep water zone" and "critical period," moving beyond initial online service implementations to focus on data-driven and intelligent decision-making [3] - Leading institutions have advanced to a new cycle characterized by data asset operations and AI capabilities, while many smaller banks still face significant challenges [3][4] Organizational Structure and Mechanisms - Successful digital transformation requires a restructuring of production relationships, emphasizing strategic leadership and integration of technology with business [4][5] - Shanghai Bank has adopted a "strong middle platform empowerment and agile tribe combat" principle to facilitate this transformation [4] Evaluation and Decision-Making - The bank has established a three-dimensional evaluation system focusing on value, experience, and efficiency to assess digital initiatives [5] - Decision-making is guided by strategic orientation and value quantification, ensuring that technology investments align with measurable outcomes [5] Challenges in Digitalization - The banking sector faces systemic challenges, including a shortage of skilled talent who understand both finance and technology, and issues related to data ownership and privacy [6] - There is a call for collaborative efforts between regulators and the industry to address these challenges and promote digital transformation [6] Attitude Towards AI and Large Models - Shanghai Bank views AI, particularly large models, as a core strategic element for future competitiveness, transitioning from cost reduction to value creation [7][10] - The bank is cautious in its tactical approach, implementing AI in areas with lower risk while ensuring strict oversight in critical financial operations [7][9] Future Directions in Digitalization - The bank anticipates breakthroughs in AI-native financial products, real-time risk management networks, and the integration of financial services into industrial processes [12] - There is a focus on developing privacy computing technologies and exploring advanced computing solutions to address future challenges in data security [12]

Core Insights - The article highlights the successful IPOs of two companies, Minglue Technology and 51WORLD, in late 2025, marking significant milestones in the fields of Agentic AI and Physical AI respectively [1][5][9] - Both companies faced substantial challenges prior to their IPOs, which they overcame through participation in the Chaos Black Innovation Enterprise Alliance, emphasizing the importance of strategic support and collaboration in entrepreneurship [9][45] Group 1: Minglue Technology - Minglue Technology became the first publicly listed company in the Agentic AI sector, achieving a market capitalization exceeding HKD 40 billion on its listing day [3] - The company has established itself as a leader in data intelligence, providing marketing data support to over 135 Fortune 500 companies and holding more than 2,300 patents [3] - Founder Wu Minghui's journey reflects a transition from a focus on mathematical logic to a broader understanding of trust and human connection in business, which he articulated as a key to his company's future direction [19][28] Group 2: 51WORLD - 51WORLD, recognized as the first Physical AI company to go public, achieved a market valuation of over HKD 15 billion, focusing on creating a digital twin of Earth to address real-world challenges [5][7] - The company has developed capabilities to replicate urban, transportation, and energy systems, serving over 1,000 clients across 19 countries [36] - Founder Li Yi's vision of "cloning the Earth" began with a personal mission to protect the planet, showcasing the blend of ambition and technological innovation in his entrepreneurial approach [34][35] Group 3: Challenges and Support - Both founders faced significant crises in early 2024, which led them to seek support from the Chaos Black Innovation Enterprise Alliance, highlighting the value of mentorship and strategic collaboration in overcoming adversity [9][45] - The article emphasizes the importance of community and shared experiences among entrepreneurs, as both Wu and Li found solace and guidance in their interactions with like-minded peers [40][41] - The strategic co-creation process facilitated by the alliance helped both companies align their visions with actionable steps, reinforcing the idea that collaboration can bridge gaps in understanding and execution [38][45]

近日，冀北电科院发布"智序"电力智能体测评体系，为人工智能在电力行业的有序落地筑牢根基，推动 新型电力系统与智能电网建设迈向新高度。 随着国家"人工智能+"行动持续推进，新型电力系统加快建设，智能体作为人工智能技术的重要应用形 态，正逐步从实验探索走向电网核心业务场景。电力行业的AI应用具有高安全性、高可靠性特点，智 能体一旦参与运行和生产，其行为是否可控、决策是否稳定、结果是否可信，成为必须正面回答的现实 问题。 围绕智能体应用落地前的这一重要关口，冀北电科院立足电力行业实际，打造"智序"电力智能体测评体 系，面向智能体全生命周期构建系统化、工程化的测评方法，致力于以"可度量、可解释、可复现"的专 业评估手段，为电力智能体实现"可用、好用、放心用"提供支撑。 从行业视角看，"智序"电力智能体测评体系为智能体应用提供了一套可复用、可推广的测评范式，有助 于统一能力认知、降低应用风险、提升人工智能应用的可控性和规范性，为后续开展规模化应用和行业 协同奠定基础。 面向未来，冀北电科院将持续深化"智序"测评体系建设，推动其在更多电力业务场景中的实践应用，并 加强与国家人工智能测评与标准体系的协同衔接，不断提升电力智 ...

Core Insights - The article discusses the breakthrough of Robust-R1, a new approach to multi-modal large language models (MLLMs) that addresses the critical issue of visual degradation in real-world applications, such as autonomous driving and medical imaging [1][2][23]. Group 1: Problem Identification - Visual degradation, including blurriness, noise, and occlusion, poses a significant challenge for advanced models like GPT-4V and Qwen-VL, hindering their deployment in key sectors [2][4]. - Existing methods rely on "implicit adaptation" strategies, which attempt to make models resistant to interference but fail to provide a comprehensive understanding of the degradation itself [2][3]. Group 2: Robust-R1 Solution - Robust-R1 introduces a paradigm shift by transforming the perception of visual degradation into an explicit structured reasoning task, allowing models to not only resist but also diagnose interference [2][3][24]. - The core idea of Robust-R1 is to construct a "degradation perception reasoning system" that follows a three-step diagnostic process: degradation diagnosis, semantic impact analysis, and robust conclusion generation [3][5]. Group 3: Technical Implementation - The first phase involves supervised fine-tuning with a structured reasoning chain, enabling the model to learn a "diagnose first, reason later" approach [9]. - The second phase introduces a degradation perception reward function to optimize the model's accuracy in identifying degradation types and intensities [10]. - The third phase employs a dynamic reasoning depth adjustment mechanism, allowing the model to adapt its reasoning based on the severity of degradation [10][11]. Group 4: Performance Validation - Robust-R1 has been tested against various benchmarks, achieving superior performance in understanding real-world degradation compared to existing models, with a comprehensive performance score of 0.5017 on the R-Bench benchmark [14][15]. - In stress tests with varying levels of synthetic degradation, Robust-R1 demonstrated significantly better robustness, maintaining usable accuracy even under extreme conditions [18]. Group 5: Implications and Future Directions - The development of Robust-R1 marks a significant transition in multi-modal models from striving for perfection in clear environments to making reliable decisions in complex realities [23][24]. - This innovation not only enhances the transparency and trustworthiness of AI models but also sets a new direction for robust MLLM research [24].

以下文章来源于快鲤鱼 ，作者杨婧雪 快鲤鱼 . 创业邦旗下AGI矩阵号，寻找海内外创新性的AGI高成长公司，记录AGI商业领袖的成长轨迹。 作者丨杨婧雪 编辑丨刘恒涛 | | | 北京彩智科技有限公司-融资历程 | | | --- | --- | --- | --- | | 融资轮次 | 事件时间 | 融资金额 | 投资方 | | 天使轮 | 2024年11月 | 数千万人民币 | 智谱Al领投 | | | | | 盛景嘉成 | | A轮 | 2025年12月 | 数千万人民币 | 致远互联独家领投 | | | | | 数据来源： 睿兽分析 | 随着 AI 的不断发展，大模型在一些垂直的严肃工作场景落地越来越普遍。比如政务服务大模型、企 业客服大模型，这些严肃的场景，需要大模型对制度、章程的严谨、准确输出，对幻觉"零容忍"。但 目前市面上常见的大模型都是概率模型，尤其遇上章程知识， AI 幻觉更加严重。 但对政企办公来说，规章制度是基本守则。 AI 如果不能解决对规章制度的幻觉，就很难真正进入严 肃办公场景。 彩智科技 正在推进的深知可信知识模型， 针对的就是这一市场痛点。 瞄准规章领域痛点 打造零幻觉 AI ...