为何科学之问永不落幕 科学的使命在于理解世界，而人类迄今所获的认知仍极其有限。世界顶尖科学家协会主席、2006年诺贝 尔化学奖得主罗杰·科恩伯格举例道，"我们对人类生物学的了解甚至不到1%。"正因如此，科学的未来 才显得无限广阔。 历史上，总有人声称科学已近终点。19世纪末，便有不少观点认为人类知识体系已趋完备，今后只需在 工程与应用上优化即可，甚至提出"物理学大厦已建成"。然而，1905年爱因斯坦发表的三篇论文彻底重 塑了物理学的面貌，其影响贯穿整个20世纪并延续至今。 AI for Science（科学智能）的兴起，正在深刻改变科学研究的范式。当AI能够预测蛋白质结构、筛选药 物靶点，并行执行数以万计的实验模拟时，一个根本性问题也随之浮现：人类科学研究的价值何在？ 南方财经 21世纪经济报道记者雷若馨迪拜报道 2月3日，世界顶尖科学家峰会（WLS）在阿联酋迪拜圆满闭幕。在为期三日的峰会中，与会科学家、 政策制定者与青年领袖在自由对话中达成深度共识：在充满不确定性的变局中，基础科学仍是人类能够 依赖的确定性力量。而基础科学的飞跃往往源于认知之外的未知，而这正是算法尚难抵达的疆域。 与此同时，AI与区块链等技 ...

Core Insights - Google DeepMind is at the forefront of AI research, focusing on breakthroughs that impact science, business, and society, particularly in the context of the AGI race [1][3][4] - The company has made significant advancements, including the development of Gemini, which is now competitive with ChatGPT, and has roots in technologies originally developed by Google [3][4][28] - The investment made by Google in DeepMind in 2014, approximately £400 million (around $540 million), has potentially grown to hundreds of billions, highlighting the strategic importance of this acquisition [4][28] Company Overview - Google DeepMind was founded in 2010 in London by Demis Hassabis, Shane Legg, and Mustafa Suleyman, with the latter now working at Microsoft [2][3] - The company has been pivotal in Google's AI advancements, particularly with consumer-facing products like Gemini, which leverage DeepMind's foundational technologies [4][28] Technological Developments - The AI landscape has evolved significantly since the emergence of ChatGPT, with Google facing internal restructuring to adapt to the competitive environment [3][4] - DeepMind's previous breakthroughs, such as AlphaGo and AlphaFold, have set the stage for its current innovations, emphasizing the company's commitment to solving fundamental scientific problems [4][5] AGI and Future Prospects - The pursuit of AGI is a long-term mission for DeepMind, with expectations of achieving significant milestones within the next 5 to 10 years [10][11] - Current AI systems, including LLMs, face limitations in achieving true AGI, particularly in areas like continuous learning and creative hypothesis generation [7][8][10] Energy and Efficiency Challenges - There are physical limitations in AI development, particularly concerning energy consumption and computational power, which need to be addressed as the field progresses [11][12] - Innovations in model efficiency, such as the use of Distillation, are expected to enhance performance significantly, with annual improvements projected at around 10 times [12][13] Competitive Landscape - The AI industry is experiencing intense competition, with many players, including startups and established tech giants, vying for leadership [28][29] - Concerns about potential financial bubbles in the AI sector are acknowledged, with some segments showing signs of unsustainable valuations [32][33] Global AI Dynamics - The competition between the U.S. and China in AI development is intensifying, with Chinese companies like DeepSeek and Alibaba making notable advancements [35][36] - Despite rapid progress, there are questions about whether Chinese firms can achieve significant innovations beyond existing technologies [36][38] Collaboration and Integration - Google DeepMind operates as a central hub for AI research within Google, integrating technologies across various products and ensuring rapid deployment of new capabilities [41][42] - The collaboration between DeepMind and Google is characterized by a close iterative process, allowing for swift adjustments to strategic goals and product development [42][43]

图片来源： Youtube 的。外界曾有一种看法，认为Google让ChatGPT把这项技术抢先用起来了。但在我看来，现在的Gemini已经几乎可以和ChatGPT平起平坐，甚至在某些方面 表现更好。 Arjun： Google DeepMind在这当中起着核心作用。我之前提到，它成立于2010年，而Google在2014年将其收购。当时我刚刚进入科技报道行业，Google 为DeepMind支付了大约4亿英镑，也就是2014年约5.4亿美元。按照现在的估算，这笔投资的价值可能已经达到数百亿，甚至上千亿美元。 Arjun： 实际上，DeepMind对Google的AI发展至关重要。以Gemini这个面向消费者发布的聊天机器人为例，它的背后技术很大程度上都来自DeepMind。 但早在这些之前，DeepMind就已经有过一些重大突破。几年前，他们推出了名为AlphaGo的系统，引起了全球轰动。这是第一个能够击败围棋世界冠军的 计算机程序。围棋是一种非常复杂的棋类游戏，当时被视为AI的重大挑战之一，因为它的变化极其多样，可能的组合数量非常庞大。 Z Highlights： 2026 年 1 月 16 日由 Arj ...

知情人士称，Silver 正在伦敦创办一家名为 Ineffable Intelligence 的新公司。该公司目前正在积极招聘人工智能研究人员，并寻求风险投资。 Google DeepMind 已于本月初向员工宣布了 Silver 的离职消息。Silver 在离职前的几个月里一直处于休假状态，并未正式返回 DeepMind 工作岗位。 Google DeepMind 的一位发言人在电子邮件声明中证实了 Silver 离职的信息，表示：「Dave 的贡献是无价的，我们非常感谢他对 Google DeepMind 工 作所做出的贡献。」 编辑 | 泽南 又一位 AI 大佬决定创业，这位更是重量级。 《财富》等媒体本周五报道说，在 Google DeepMind 众多著名突破性研究中发挥关键作用的知名研究员 David Silver 已离开公司，创办了自己的初创公 司。 根据英国公司注册处 Companies House 的文件显示，Ineffable Intelligence 公司成立于 2025 年 11 月，Silver 于今年 1 月 16 日被任命为该公司董 事。 此外，Silver 的个人网页现在将他的 ...

Core Insights - The article discusses the transition from AI for Science (AI4S) to AGI for Science (AGI4S), emphasizing the need for a specialized generalist model to enhance scientific discovery and reasoning capabilities [1][2][71]. Group 1: Current State of AI in Science - AI for Science, exemplified by AlphaFold, has achieved significant milestones in specific fields like protein folding and weather prediction, but reliance on existing deep learning models may limit the exploration of new knowledge and hinder innovation [1][71]. - A systematic evaluation involving 100 scientists from 10 different scientific fields revealed that cutting-edge models scored 50 out of 100 in general scientific reasoning tasks, but dropped to scores between 15 and 30 in specialized reasoning tasks [1][71]. Group 2: The Need for AGI4S - The transition from AI4S 1.0 to AGI4S 2.0 is necessary to integrate general reasoning with specialized capabilities, addressing the limitations of current models in scientific discovery [2][71]. - The concept of "Specialized Generalist" is proposed as a feasible path to achieve AGI, which combines deep specialization with general capabilities [2][90]. Group 3: Technological Framework - SAGE - The "SAGE" architecture is introduced as a synergistic framework for developing generalizable experts, consisting of three layers: foundational, collaborative, and evolutionary [3][18]. - The foundational layer focuses on decoupling knowledge and reasoning capabilities, while the collaborative layer employs reinforcement learning to balance intuitive and logical reasoning [27][28]. - The evolutionary layer aims to enable self-evolution of models through continuous interaction and feedback, addressing the challenges of adapting to complex tasks [55][56]. Group 4: Innovations in Reinforcement Learning - The article highlights the development of the PRIME algorithm, which provides dense rewards for reinforcement learning without the need for extensive manual labeling, significantly improving model performance [38][39]. - FlowRL is introduced to enhance the diversity of reasoning paths in models, allowing them to explore multiple solutions rather than converging on a single answer [47][50]. Group 5: Applications and Case Studies - The Intern-S1 model is designed to be a deep specialized generalist for scientific applications, demonstrating superior performance in various scientific domains compared to existing models [77][79]. - The Intern-Discovery platform integrates the Intern-S1 model with extensive data and tools, facilitating a closed-loop system for hypothesis generation and experimental validation [80][84]. Group 6: Future Directions - The article calls for collaboration among researchers to fill the gaps in the current framework and advance the development of AGI4S, emphasizing the potential for AI to revolutionize scientific research [89][90].

Core Insights - OpenAI has launched a new AI research tool called Prism, powered by GPT-5.2, aimed at assisting scientists in writing and collaborating on research, now available for free to all ChatGPT personal account users [1] - The company aims to empower scientists with AI capabilities to accelerate research, with a vision to enable scientific advancements by 2030 that would typically be expected by 2050 [1][2] - OpenAI's entry into the scientific field comes after competitors like Google DeepMind have already established their presence with AI-for-science teams and groundbreaking models [2] Group 1: OpenAI's Strategic Goals - OpenAI's goal is to enhance the capabilities of scientists, allowing them to focus on more complex problems rather than previously solved issues, thereby accelerating research [2][3] - The company plans to optimize its models by reducing confidence levels in answers and implementing self-fact-checking mechanisms [3][15] - OpenAI's mission is to develop general artificial intelligence (AGI) that benefits humanity, with a focus on transforming scientific research through new drugs, materials, and instruments [3][4] Group 2: Model Performance and Capabilities - GPT-5 has shown significant improvements, achieving a 92% accuracy rate in the GPQA benchmark, surpassing the performance of 90% of graduate students [5] - The model has been recognized for its ability to assist researchers in finding connections between existing research and generating new insights, although it still makes errors [10][11] - OpenAI acknowledges that while the model can assist in research, it has not yet reached the level of making groundbreaking discoveries [6][8] Group 3: Industry Context and Competition - OpenAI's late entry into the AI-for-science domain is notable, as competitors like Google DeepMind have already made significant advancements [2][16] - The company is aware of the competitive landscape and aims to establish a strong foothold in the scientific research sector [16] - OpenAI's focus on optimizing model features and enhancing collaboration with researchers is part of its strategy to differentiate itself from other AI models in the market [15][16]

Core Viewpoint - OpenAI has launched Prism, a new AI research tool powered by GPT-5.2, aimed at enhancing scientific research collaboration and efficiency, now available for free to all ChatGPT personal account users [2][3]. Group 1: OpenAI's Strategic Move - OpenAI's entry into the scientific research field is seen as a response to the growing importance of AI in academia, with the goal of empowering scientists to conduct advanced research by 2030 [2][3]. - The establishment of the OpenAI for Science team indicates a focused effort to explore how large language models (LLMs) can assist researchers and optimize tools for scientific support [2][3]. Group 2: Model Capabilities and Limitations - Kevin Weil, OpenAI's VP, acknowledges that while current models can accelerate research by preventing time wastage on solved problems, they are not yet capable of making groundbreaking discoveries [4][5]. - The latest version, GPT-5.2, has shown significant improvement, achieving a 92% accuracy rate in the GPQA benchmark, surpassing the performance of 90% of graduate students [7][8]. Group 3: Research Applications and Feedback - Researchers have reported that GPT-5 can assist in brainstorming, summarizing papers, and planning experiments, significantly reducing the time needed for data analysis [13][14]. - Feedback from various scientists indicates that while GPT-5 can provide valuable insights, it still makes basic errors, and its role is more about integrating existing knowledge rather than generating entirely new ideas [14][15]. Group 4: Future Directions and Enhancements - OpenAI is working on two main optimizations for GPT-5: reducing confidence in its answers to promote humility and enabling the model to fact-check its outputs [4][19]. - The goal is to create a collaborative workflow where the model can serve as its own verifier, enhancing the reliability of its contributions to scientific research [19][20].

作者 | Big Technology Podcast 编译 | 王启隆 出品丨AI 科技大本营（ID：rgznai100） 如果说 Sam Altman 是 AI 时代的布道者，善于用宏大的愿景点燃公众的想象力；那么 Demis Hassabis 更像是一位在实验室里盯着显微镜的科学 家，冷静、严谨，对"炒作"有着天然的免疫力。 一年前，当整个硅谷都在因为 ChatGPT 的红利期似乎见顶而焦虑，甚至开始讨论"大语言模型（LLM）是否撞墙"时，Demis 却感到困惑。在他看来， 进步从未停止。他掌舵的 Google DeepMind 刚刚经历了 AlphaFold 3 的高光时刻，正试图将 AI 的触角从简单的聊天机器人延伸到生物学、物理学乃 至材料科学的最深处。 在达沃斯的一间木质会议室里，Demis 近期接受了 Big Technology 播客的专访。这场对话的特别之处在于，他没有回避那些尖锐的问题： 现在的 AI 是不是只有"金鱼记忆"？谷歌会不会为了财报在 Gemini 里塞满广告？所谓的 AGI 究竟是营销话术还是科学定义？ 最令人印象深刻的是他对"智能载体"的断言。在纪录片《The Think ...

Core Insights - The article discusses the launch of AlphaGenome, a new AI tool by DeepMind that predicts the effects of single nucleotide mutations in human DNA sequences, enhancing the understanding of genetic diseases and guiding DNA design [2][3]. Group 1: AlphaGenome Overview - AlphaGenome is a DNA sequence model capable of processing up to 1 million base pairs, accurately predicting a wide range of genomic features and mutation effects [10]. - The model represents a significant advancement in genomic AI, moving from specialized models to a unified approach that can handle multiple tasks simultaneously [11][12]. Group 2: Technical Innovations - AlphaGenome achieves a breakthrough by maintaining single-base resolution while analyzing long sequences, combining the strengths of convolutional neural networks and transformer architectures [11][15]. - It can evaluate the impact of genetic mutations on various molecular characteristics in just one second, facilitating rapid identification of potentially disease-causing genetic variations [13]. Group 3: Performance Metrics - In 24 DNA sequence function prediction tasks, AlphaGenome achieved state-of-the-art performance in 22 tasks, and in 26 genetic variant impact prediction tasks, it excelled in 24 tasks, outperforming many specialized models [19]. Group 4: Practical Applications - AlphaGenome has been utilized to explore the mechanisms of mutations related to cancer, linking non-coding region mutations to the activation of oncogenes [22]. - It also aids in understanding rare genetic diseases caused by RNA splicing errors and can guide the design of synthetic DNA sequences for targeted gene therapy [24]. Group 5: Future Implications - The introduction of AlphaGenome signifies a shift in genomic AI from single-task specialists to comprehensive models, paving the way for predictive science in biology [26]. - It enhances the ability to predict molecular functions and mutation effects from DNA sequences, opening new avenues for biological discoveries and applications in biotechnology [26].

Core Insights - The article discusses the groundbreaking advancements of AlphaGenome, a new AI model by Google DeepMind, which aims to decode the 98% of the human genome that does not code for proteins, previously referred to as "junk DNA" [1][4][22]. Group 1: AlphaGenome's Purpose and Innovations - AlphaGenome is designed to decipher the regulatory mechanisms of non-coding DNA, which plays a crucial role in gene expression and is linked to various diseases [5][6]. - The model utilizes three key innovations: ultra-long context, single-base precision, and multi-modal predictions, allowing it to analyze vast sequences of DNA and predict multiple biological features simultaneously [6][10]. - A specific case study highlighted the model's ability to identify a mutation in a non-coding region that led to a form of leukemia, showcasing its precision in detecting subtle genetic changes [7]. Group 2: Evolution of the Alpha Family - The Alpha family of AI models has evolved from AlphaGo, which focused on game strategy, to AlphaFold, which predicts protein structures, and now to AlphaGenome, which aims to understand the dynamic regulatory processes of life [9][10]. - This progression signifies a shift from static predictions to dynamic understanding of biological systems, moving closer to the core of life processes [10][22]. Group 3: Implications for Drug Development and Healthcare - AlphaGenome is set to revolutionize drug development by enabling faster identification of disease-causing mutations and designing targeted therapies, potentially reducing the development timeline from ten years to just two or three [13]. - The model also paves the way for personalized medicine by analyzing individual genetic variations, allowing for tailored drug dosages and treatment plans [14][15]. - The advancements in synthetic biology facilitated by AlphaGenome will enable precise genetic modifications, significantly enhancing the efficiency of biotechnological applications [16]. Group 4: Limitations and Ethical Considerations - Despite its capabilities, AlphaGenome is described as a "black box" model, meaning it can predict outcomes but lacks the ability to explain the underlying biological mechanisms [18]. - There are concerns regarding the model's training data, which predominantly represents European populations, potentially leading to disparities in healthcare outcomes for other ethnic groups [18]. - Ethical dilemmas arise from the potential for gene editing technologies to create "designer babies," raising questions about regulation and societal implications [18]. Group 5: Recommendations for Stakeholders - For students and professionals in the field, there is a growing demand for expertise in bioinformatics and computational biology, emphasizing the need for interdisciplinary knowledge [20]. - Healthcare professionals are encouraged to familiarize themselves with AI tools, as those who do not adapt may be left behind in the evolving landscape of medicine [20]. - Investors and entrepreneurs should focus on niche areas such as non-coding variant detection services and AI-driven personalized medicine, as these sectors are expected to see significant growth and investment opportunities [20][21].