Original Video: Lex Fridman Podcast #475 原始视频: Lex Fridman Podcast #475
Interview with Demis Hassabis, CEO of Google DeepMind. 与Google DeepMind CEO Demis Hassabis的访谈。
The conversation revolves around Demis Hassabis's profound conjecture: "any pattern generatable or discoverable in nature, can be efficiently discovered and modeled by classical learning algorithms." This idea is underpinned by breakthroughs at Google DeepMind with AlphaFold and AlphaGo, and the remarkable ability of the VEO model to simulate physical phenomena like liquids. 本次对话围绕Demis Hassabis提出的一个引人深思的猜想展开，即“自然界中任何可生成或被发现的模式，都能被经典学习算法高效发现和建模”。他通过Google DeepMind在AlphaFold和AlphaGo上取得的突破，以及VEO模型在模拟液体等物理现象方面的惊人能力，来支撑这一观点。
Nature is not random; its complex structures and evolutionary processes imply the existence of learnable low-dimensional manifolds, enabling AI to efficiently "reverse-engineer" physical laws. 核心在于，自然界并非随机，其复杂的结构和演化过程意味着存在可学习的低维流形，使得AI能够高效地“逆向学习”物理规律。
This pie chart illustrates the three key pillars supporting the conjecture: Breakthroughs in AI Games (AlphaGo), Molecular Biology (AlphaFold), and Realistic Physical Simulation (VEO). 本饼图展示了支撑该猜想的三大关键支柱：AI游戏突破 (AlphaGo)、分子生物学 (AlphaFold) 和 真实物理模拟 (VEO)。
Lex Fridman notes the human difficulty in predicting highly non-linear dynamic systems. Demis Hassabis counters by highlighting the surprising ability of classical learning systems, specifically VEO (Video Generation Model), to handle fluid dynamics. Lex Fridman指出人类难以对高度非线性的动态系统做出清晰预测。Demis Hassabis则强调了经典学习系统，特别是VEO（视频生成模型）在处理流体动力学方面的惊人能力。
Fluid Dynamics流体动力学
Realistic liquid flow真实液体流动
Material Behavior材料行为
Deformation & interaction形变与交互
Specular Lighting镜面光照
Realistic reflections真实反射
VEO's capacity to generate coherent 8-second videos, indiscernible from reality by the naked eye, suggests a deeper "understanding" than mere pattern matching. VEO能够生成肉眼几乎无法区分的连贯8秒视频，这表明它不仅仅是简单的模式匹配，而是具备了更深层次的“理解”。
Demis Hassabis's "super interesting conjecture" from his Nobel Prize lecture posits that patterns in nature are efficiently discoverable by classical learning algorithms. This stems from DeepMind's work on AlphaX projects (AlphaGo, AlphaFold), which tackle problems with astronomically vast solution spaces. Demis Hassabis在诺贝尔奖演讲中提出的“超级有趣的猜想”认为，自然界中的模式可以被经典学习算法高效发现。这源于DeepMind在AlphaX项目（AlphaGo，AlphaFold）上的工作，这些项目解决的都是拥有天文数字般巨大解空间的问题。
These numbers far exceed the number of atoms in the universe. Yet, AI finds solutions efficiently. 这些数字都远超宇宙中的原子数量，然而AI却能高效找到解决方案。
The core reason: Natural systems possess inherent structure, shaped by evolutionary processes (e.g., protein folding, geological formation, cosmic evolution). This structure allows AI to learn a "manifold" that guides efficient search for solutions, making seemingly intractable problems tractable. 核心原因在于：自然系统具有内在的结构，这些结构是由演化过程塑造的（例如蛋白质折叠、地质形成、宇宙演化）。这种结构使得AI能够学习到一个“流形”，从而高效地搜索到解决方案，将看似无法处理的问题变得“可处理”。
Demis Hassabis believes "information is the most fundamental unit of the universe," suggesting the universe is an information system. This frames the P=NP problem as a physics question related to what "Learnable Natural Systems" (LNS) can be efficiently modeled by neural networks running on classical Turing machines. Demis Hassabis认为“信息是宇宙最基本单位”，将宇宙看作信息系统。这使得P=NP问题成为一个物理学问题，与经典图灵机上运行的神经网络能高效建模的“可学习自然系统”（LNS）相关。
While most emergent systems (like Cellular Automata) can be modeled by classical systems, Demis acknowledges challenges with chaotic systems (highly sensitive to initial conditions). 虽然大多数涌现系统（如细胞自动机）可以被经典系统建模，但Demis承认混沌系统（初始条件高度敏感）可能难以建模。
Artificial or abstract things without underlying structure (e.g., factoring large numbers unless patterns exist) may not be amenable to this efficient modeling. 人工创造物或抽象事物（如大数分解，除非数字空间中存在模式）可能不适用这种高效建模。
The ability of VEO to render physics and lighting hints at fundamental characteristics of the universe's structure, motivating the quest for AGI to answer questions like P=NP. VEO在视频生成中渲染物理和光照的能力，暗示了宇宙结构的一些基本特点，这激励着人们构建AGI以回答诸如P=NP的问题。
Demis Hassabis's original love for games fuels his vision of AI's transformative impact. He dreams of "shocking games" built with today's AI, particularly focusing on open-world games where simulations, AI characters, and player interaction dynamically adjust the game experience. Demis Hassabis最初对游戏的热爱驱动着他对AI变革性影响的愿景。他梦想着用今天的AI构建“令人震惊的游戏”，特别是专注于开放世界游戏，其中模拟、AI角色和玩家互动能够动态调整游戏体验。
Key Advantage: 关键优势：AI can instantly generate content, overcoming high asset creation costs of traditional AAA games.AI可以即时生成内容，克服了传统AAA级游戏高昂的资产创建成本。
Golden Age: 80s-Early 2000s 黄金时代：80年代-2000年代初
New entertainment media, co-creating stories, social activity. 创造新娱乐媒介，玩家共同创造故事，成为社交活动。
Games as Micro-Simulations 游戏作为微观模拟
Safe, repeatable environments to practice decision-making, winning, and losing (e.g., Jiu-Jitsu). Essential for self-improvement and self-awareness. 安全、可重复的环境，用于练习决策、输赢（如柔术）。有助于自我提升和自我认识。
Demis Hassabis's early work on AI in games like Black and White (AI learning AI via Reinforcement Learning) is a direct lineage to his current pursuit of "general learning systems." He hopes to return to game development as a "post-AGI project." Demis Hassabis早期在《黑与白》等游戏中对AI（通过强化学习的AI学习AI）的研究，与他现在追求的“通用学习系统”一脉相承。他希望在“AGI成功后”能重返游戏开发。
Demis Hassabis predicts a 50% chance of achieving AGI by 2030, based on a stringent definition: Demis Hassabis预测到2030年有50%的几率实现AGI，基于一个严苛的定义：
This implies a general-purpose Turing machine capability, similar to the human brain's role in creating modern civilization. Current systems exhibit "unbalanced intelligence" and lack "true inventiveness and creativity." 这意味着一种通用图灵机能力，类似于人脑在创造现代文明中的作用。目前的系统表现出“不均衡的智能”，并且缺乏“真正的发明能力和创造力”。
Distinguishing Great Scientists 区分伟大科学家
"Research taste" or "judgment" is the most challenging quality to model in AI. It's the ability to "sniff out the right direction, the right experiment, the right problem." “研究品味”或“判断力”是AI最难模仿或建模的品质之一。它指的是“嗅探正确的方向、正确的实验、正确的问题”的能力。
Demis Hassabis's 25-year-old dream is to simulate an entire cell ("Virtual Cell"), enabling "in-silico" experiments that are 100 times faster than wet lab. Demis Hassabis长达25年的梦想是模拟一个完整的细胞（“虚拟细胞”），从而实现“硅中”实验，比湿实验室实验速度提高100倍。
Predicts static protein 3D structures.解决蛋白质静态三维结构。
Models dynamic molecular interactions.迈向建模动态交互。
Simulating Origin of Life 模拟生命起源
One of the deepest questions. AI can aid the combinatorial search. Aiming for a "move 37" in life's origins, seeing it as a continuum from physics to biology. AI is the "ultimate tool." 最深刻的问题之一。AI可以辅助组合空间搜索。期望生命起源中出现一个“move 37”，将其视为从物理学到生物学的连续体。AI是“终极工具”。
World's Best Weather Prediction 世界最佳天气预测
Google DeepMind's WeatherNet outperforms traditional fluid dynamics systems, which take days on supercomputers. It shows neural networks can model complex, chaotic systems like hurricane paths effectively. Google DeepMind的WeatherNet系统优于传统的流体动力学系统，后者在大型超级计算机上需要数天才能计算出来。它证明了神经网络可以有效地建模复杂、混沌的系统，例如飓风路径预测。
Programming and math are "hard skills" easier for AI. Demis predicts that in 5 to 10 years, those embracing AI will become "superhumanly efficient," 10 times better than today. 编程和数学等“硬技能”对AI来说反而更容易。Demis预测在未来5到10年内，那些拥抱AI技术的人将变得“超人般地高效”，甚至比今天好10倍。
"This transformation could be 10 times the Industrial Revolution, but 10 times faster, combining to a 100 times overall effect." This requires profound societal response (e.g., Universal Basic Provision). “这次变革的影响可能是工业革命的10倍，但速度会快10倍，结合起来达到100倍的综合影响。”这需要社会深思如何应对（例如全民基本供给）。
AI is a "multi-purpose technology" capable of solving global challenges like disease, energy, and scarcity, echoing John von Neumann's prediction that computers would have a greater impact than the atomic bomb. The priority is to create "radical abundance." AI是一种“多用途技术”，能够解决疾病、能源和稀缺性等全球挑战，呼应了约翰·冯·诺伊曼关于计算机影响将大于原子弹的预测。当务之急是创造“激进的丰裕”。
Beyond technology, a "spiritual dimension" or "humanistic dimension" is crucial. AI is an enabler, a tool for human flourishing and understanding the world. 除了技术进步，还需要关注“精神维度”或“人文维度”。AI是一种使能者，是帮助人类繁荣和了解世界的工具。
AGI will eventually be used by nations for war. Demis hopes for more cooperative development and deployment, akin to CERN projects, to mitigate the risk of "bad actors abusing universal technology." AGI最终也会被国家用于战争等目的。Demis希望未来能以更合作的方式（如CERN项目）来开发和部署AGI，以减少“坏人滥用通用技术”的风险。
Regarding p_doom (probability of human civilization self-destruction), Demis maintains it's "non-zero" and "possibly non-negligible," but emphasizes "cautious optimism." He is more concerned about bad actors than the alignment problem itself. 关于p_doom（人类文明自我毁灭的概率），Demis认为它“肯定是非零的，可能也不可忽略”，但他强调“谨慎乐观”。他个人更担心“坏人滥用技术”，而不是对齐问题本身。
Google's shift from "behind" to "leading" in LLM products is attributed to three key factors by Demis Hassabis: Google在LLM产品方面从“落后”到“领先”的转变，Demis Hassabis归因于三个关键因素：
Exceptional leaders & members.杰出的领导者与成员。
Essential for large-scale training.大规模训练的必要条件。
Blended Google Brain & old DeepMind.融合Google Brain与DeepMind。
DeepMind operates like a "startup," enabling high-level decision-making and agility. They strive to combine the benefits of a large platform with billions of users and the incredible power of AI research, while "constantly chopping bureaucracy." DeepMind像一家“创业公司”一样运作，具有高层的决策力和小型组织的活力。他们试图兼顾拥有数十亿用户的平台优势，以及AI研究的强大赋能，同时“不断削减官僚主义”。
Demis Hassabis's product leadership stems from his game design experience. He advocates for applying cutting-edge research to create new user experiences, a skill requiring "intuitive imagination" and "good taste." Demis Hassabis的产品领导力源于他早年为数百万玩家设计游戏的经验。他倡导将前沿研究应用于产品，并创造新的用户体验，这项技能需要“直观的想象力”和“良好的品味”。
Demis rejects treating AI as a "win/lose" race due to its significance. He fosters strong relationships with other lab heads, viewing research as collaborative, especially on safety. Science acts as a "beautiful connector." Demis反对将AI视为一场“赢或输”的比赛，因为它事关重大。他与所有主要实验室的负责人保持良好关系，认为研究是“协作性工作”，尤其是在安全等问题上。科学是一种“美丽的连接器”。
David Foster Wallace's graduation speech: "The most obvious, important realities are often the ones that are hardest to see and talk about." David Foster Wallace的毕业演讲：“最明显、最重要的现实，往往是最难看到和谈论的。”
"Everything depends on perspective," and humility broadens wisdom (bar story). “一切都取决于视角”，谦逊地拓宽视角才能获得智慧（酒吧故事）。