A16z Partners Say AI Scaling Shifts Software Bottleneck to Capital

A16z合伙人：AI规模扩展将软件瓶颈转移至资本 生成式模型正在改变技术经济学，初创企业得以通过直接部署算力而非招募人才，来解决以往难以攻克的工程瓶颈。 本次对话探讨了AI如何通过将历史上受限于工程能力的问题转化为资本和算力问题，从而深刻改变数学与软件领域。对话者梳理了计算工具的演进历程——从算盘、Curta计算器、早期IBM大型机到图形计算器，以此展示更高维度的抽象层如何持续重塑人类的能力基准。他们探讨了AI时代的“创新者窘境”，认为资本极为雄厚的初创公司如今能够绕过传统的分发壁垒，直接挑战老牌科技巨头。最终他们总结道，尽管当前的AI模型在统计学上仍是受限的贝叶斯系统，但规模达100亿至200亿美元的算力产物所催生的能力，依然让专家难以预测。 现代AI进步的根本驱动力，在于从受工程限制（受制于人才和带宽）的范式转向资本与算力问题，资本能够直接扩展模型能力。 (0:00) 专业数学家对AI的热情接纳给怀疑论者带来了悖论，这表明领域专家十分欢迎能够提升其抽象水平的工具。 (2:38) 纯数学是一个规则严格的公理化领域，AI在此擅长将不同分支联系起来；而物理现实包含混乱的变量与经验约束，纯数学无法绕过这些限制。 (2:38) 从Curta计算器到TI-85图形计算器，历史上的计算工具表明，新技术不断改变人类解决问题的基准，且每一次都会引发对自动化“作弊”的担忧。 (14:45) 历史上受经典IBM 1953年抽象概念（输入、存储、运算、控制、输出）支配的计算机架构，最初是为了解决弹道计算和导航等直接的军事与经济需求而发展演进的。 (19:43)

Steven Sinofsky and Martin Casado argue that artificial intelligence has returned computing to a capital-bound paradigm, allowing funding to translate directly into compute and model capabilities. However, they caution that while capital-rich models excel within axiomatic and abstract domains, their economic impact remains bounded by physical and empirical bottlenecks. From Engineering Bottlenecks to Capital-Driven Scaling (38:16) Historically, computing transitioned from being capital-bound in the mainframe era (1950s–1980s) to being engineering-bound from the 1990s through the 2010s, when expanding engineering teams ran into Fred Brooks's "Mythical Man-Month" management limits. Casado and Sinofsky contend that computing has inverted back to a capital-bound discipline: a 20-person startup can now absorb $1 billion productively by converting funds directly into compute and model training rather than ballooning human overhead. Sinofsky notes that this shift alters the basic dynamics of technological progress. Previously open-ended or intractable engineering challenges—such as exhaustively exploring vast combinations of protein structures—are transformed from open-ended engineering bottlenecks into solvable, finite capital problems. Shifting Computational Abstractions from Deterministic to Stochastic (29:28) Technological advancements have continually reset the baseline of intellectual labor, moving from mechanical devices like the 600-part Curta calculator to early electronic tools like the TI-35, the TI-85 graphing calculator, and early symbolic math engines like Maxima. Historically, computing architectures—such as the five-part framework of input, storage, arithmetic, control, and output outlined in IBM's 1953 brochure—provided deterministic infrastructure while human operators retained direct control over program logic. Casado emphasizes that AI breaks this deterministic continuity, moving software through three core paradigms: imperative (step-by-step recipes), declarative (specifying target end states like SQL), and stochastic computing. In this model-driven paradigm, users supply natural language prompts without defining an exact end state or execution path, probabilisticly delegating the reasoning process itself to the underlying model. Axiomatic Problem-Solving Versus Empirical Physical Constraints (12:05) AI displays remarkable capability in pure mathematics and strategy games because these are axiomatic systems operating under strict, self-contained rules. This compute-intensive dynamic mirrors the 1976 proof of the Four-Color Theorem, which abandoned the pursuit of an elegant equation in favor of brute-force computer verification across 1,936 planar configurations. AI models similarly synthesize mathematical connections across obscure domains that human specialists lack the time to span. However, Casado notes that proficiency in formal mathematical logic does not naturally extend to empirical reality. Modeling complex physical systems—such as fluid dynamics over an airplane wing or exploding stars—relies on empirical measurements and differential equations of state that cannot be solved by pure mathematical shortcuts. Similarly, while AI synthesizes medical research patterns and identifies candidate molecules, real-world drug development remains bottlenecked by clinical trials and safety testing in human biology. Incumbent Vulnerabilities and Unprecedented Model Scale (46:28) Unlike the cloud computing wave where startups built on top of an established hyperscaler oligopoly, massive capital availability allows AI startups to challenge incumbents directly. Sinofsky and Casado point out that major tech incumbents face severe institutional frictions, including balance-sheet pressures that require bond deals, token rationing that starves internal engineering teams of GPUs, and cultural rigidities reminiscent of Intel initially dismissing ARM chips as low-end processors for printers. Addressing model architecture, Casado notes that current models mechanically operate as Bayesian statistical engines bounded by their training distributions rather than autonomous systems undergoing recursive self-improvement. Nonetheless, he argues that deep experts cannot reliably forecast the limits of single digital artifacts built with $10 billion to $20 billion worth of compute and FLOP density, as humanity has never before concentrated that volume of resources into a single system. While AI scaling converts complex engineering challenges into brute-force capital allocation problems, the speakers emphasize that translation into the broader economy remains bounded by physical testing, clinical validation, and organizational inertia. The ultimate trajectory of multi-billion-dollar artifacts will depend not just on raw computational scale, but on where axiomatic reasoning meets irreducible real-world constraints.

Steven Sinofsky 与 Martin Casado 认为，当前人工智能正在重塑创新的经济逻辑：技术突破的决定性门槛已从依赖人类组织与人力规模的“工程受限型”，转变为能够将巨额资金直接兑换为算力与模型能力的“资本受限型”。 计算范式重构：从工程瓶颈转向资本驱动 (38:16) Martin Casado 指出，在过去的软件时代，技术创新严格受制于布鲁克斯法则（《人月神话》），向团队盲目注入资金往往因沟通与管理损耗导致系统瘫痪。如今，仅由约20人组成的精简团队便能有效吸收10亿美元资本，并将其直接转化为原始算力与模型能力。Steven Sinofsky 补充称，计算历史经历了一场螺旋上升：20世纪50至80年代的大型机时代属于资本受限型，90年代至2010年代演变为工程师人力受限型，而2020年代的人工智能则让计算全面回归资本受限型。 这种机制从根本上颠覆了传统“精益创业”模式。Casado 反驳了风险投资界所谓“过多资金追逐过少标的”的零和观点，强调资本涌入私有市场支持了高风险的长期研发，实际扩大了整体潜在市场空间（TAM）。Sinofsky 认为，这种将复杂工程转化为资本计算的机制改变了创新的底层规则：以往诸如全量遍历蛋白质结构等需要耗费数十年工程巧思的无限难题，如今正在转变成能够通过集中调配算力在有限时间内攻克的确定性任务。 在位者困境重演：资本反转瓦解分发壁垒 (46:28) 在传统商业竞争中，科技巨头通常凭借庞大的分发渠道与充沛现金流压制初创企业。然而 Casado 与 Sinofsky 观察到，AI 正在颠覆克里斯坦森经典的“颠覆者困境”。一方面，市场对算力与 Token 的旺盛需求消除了传统客户获取与分发阻碍，使 OpenAI、Anthropic 和 Cursor 等初创公司能直接将融得的巨资转化为顶层增长；另一方面，前沿模型能力主要由资本与算力驱动而非数十年积累的工程壁垒，导致敏捷的初创团队在基座模型表现上反超拥有海量数据和工程师队伍的在位巨头。 与此相对，科技在位者内部正面临严重的财务与组织摩擦。Sinofsky 指出，巨头往往受困于现有销售激励、客户承诺与考核体系，甚至承受沉重的资产负债表压力——例如谷歌通过发债将资本支出转移出表，部分云厂商因优先向外部企业客户配售 GPU 而导致内部团队陷入“算力饥渴”。Sinofsky 援引历史指出，当年英特尔因深陷 x86 摩尔定律惯性，将微软早期的 ARM 架构 Surface 贬低为“打印机芯片”；如今科技巨头同样极易被自身的传统云架构与商业模式捆绑，从而错失平台跃迁的机会。 抽象层级跃迁：从确定性系统到概率驱动 (29:28) Sinofsky 与 Casado 通过计算历史考察了技术抽象层的升级演进——从拥有600多个精细零件的 Curta 机械计算器，到引发教育界作弊恐慌的 TI-85 图形计算器，再到80年代麻省理工学院用于自动化微积分的 Maxima 符号数学工具。Casado 强调，当前 AI 突破了以往所有抽象工具的本质：在传统命令式（Imperative）与声明式（Declarative）编程中，机器仅提供算力与存储，人类依然完全掌控确定性的逻辑输入、输出与终态定义；而现代 AI 属于随机/模型驱动（Stochastic computing），人类将核心推理过程交由概率模型完成，只需提供自然语言输入。 这种抽象跃升使得各行各业的专家无需经过漫长的软件工程培训即可直接构建复杂系统。Sinofsky 以早年一位懂编程的医生耗费数年用 DOS 系统编写门诊排班工具为例（该系统需动态权衡患者病情并协调血液采集、X光与心电图等并行设备资源），指出过去创业的瓶颈在于领域专家缺乏技术实现手段。如今 AI 接管了底层代码编写与工程招聘，使得医疗、法律和商业地产等长期未被传统软件充分覆盖的专业领域迎来了自动化契机。 公理断层与规模涌现：物理现实的约束 (12:05) 尽管 AI 在黎曼猜想探讨以及四色定理验证（1976年计算机通过穷举1936种构型完成证明）等公理化领域进展迅速，但 Casado 提醒，绝不能将纯数学与形式逻辑的突破盲目等同于对物理现实的预测能力。AI 擅长在规则封闭、公理完备的合成数学或博弈环境中建立跨领域关联，但复杂物理现象（如超新星爆发或机翼气流模拟）依赖于基于真实测量的经验状态方程。同样，在生物医药领域，AI 虽能迅速生成候选分子，但药物研发的真实核心瓶颈依然是耗时漫长的人体临床安全性与有效性试验，纯数字计算无法绕过现实世界的物理与经验阻碍。 关于当前架构的技术极限，Casado 承认大语言模型在机制上属于受制于训练数据分布的贝叶斯统计系统，并未出现自主递归自我改进的“奇点”。但他强调，人类历史上从未创造过包含100亿至200亿美元 FLOPs 算力和超大规模数据的单一数字实体。面对这种前所未有的超大规模产物，即便模型仍受限于数据流形，其在复杂模式识别中所展现的实际问题解决能力依然超出了专家的直觉预判范围。 该讨论总结指出，AI 创新的核心本质在于将过去看似无解的工程瓶颈转化为可通过资本与算力精确兑现的经济问题。然而该趋势留下的核心未解难题在于：随着单次训练成本向千亿美元级迈进，单纯依赖资本与数据规模构建的概率模型，究竟能在多大程度上跨越经验科学与物理现实的客观鸿沟；以及当极大规模的计算资源高度集中于极少数实体时，创新红利与潜在的系统性安全风险将如何达到平衡。

## Background 该视频由风险投资机构 a16z 于其官方 YouTube 频道发布，由 a16z 普通合伙人 Martin Casado、Erik Torenberg 与董事合伙人（前微软高管）Steven Sinofsky 共同对谈。视频探讨了近期 AI 在数学与推理领域的突破，反思了数十年来支配计算机科学的底层假设，并重点剖析了创新制约因素如何从“工程人才瓶颈”彻底转移至“资本与算力规模”，进而重塑初创企业与科技巨头的竞争格局。 ## TL;DR 对话深入剖析了生成式 AI 引领的核心范式转移：软件研发与科技创新已从依赖稀缺工程人才的线性突破，演变为纯粹由资本与算力规模吞吐量决定的全新战场。嘉宾们借由四色定理、算盘到图形计算器等技术史，阐明大模型在数学形式系统中的飞跃是计算抽象层的跃迁，但其真实物理与商业效用仍有待杀手级应用检验。同时，初创团队凭借高度精简的精英架构与海量资本杠杆得以正面迎击巨头，而科技巨头则受困于既有的财务、销售体制与文化惯性，印证了“创新者的窘境”在 AI 时代的再度上演。最终，千亿美元级单一数字造物的涌现，正在将原本看似无限的科学探索转化为可计算的资本问题，彻底重构了创新的底层物理法则。 ## Atmosphere & style 整场对谈兼具硬核系统工程的严谨纵深与顶级风投的高维商业思辨，洋溢着一种既沉着务实又充满历史厚重感的极客氛围。Steven Sinofsky 带来算盘、Curta 机械计算器和 1953 年 IBM 宣传手册等珍贵实物穿插历史轶事，与 Martin Casado 针砭学术效用与“模型祈祷”现象的辛辣冷峻形成了绝佳的智识互补。从对老派工程奇迹的复盘怀旧，到面对前所未有的百亿美元级黑盒模型时的敬畏与审慎，对话情绪层层收紧，带给听众一种俯瞰半个世纪计算文明演进、目睹底层游戏规则正在被资本与算力重写的巨大震撼。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=0s) 技术演进瓶颈发生根本性迁移：软件创新已从过去数十年受限于工程团队复杂度的模式，转变为纯粹由资本规模与算力驱动的模式，20 人的初创团队如今能够高效消化并转化 10 亿美元以上的算力投入。 - [02:32](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=152s) 大语言模型在数学与符号逻辑上的突破，本质上是在封闭公理系统中的高效拼图，与其能否直接映射并预测复杂物理现实之间存在巨大的逻辑鸿沟与实证缺失。 - [46:28](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2788s) 初创企业正面击败科技巨头的机制正在被重构：凭借极高的资本吸纳能力和无限的市场需求，初创公司跨越了巨头的分发与工程壁垒，而巨头则深陷公开市场财报约束与组织惯性中难以激进自我革命。 - [55:07](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3307s) 人类首次面对耗资数百亿乃至千亿美元打造的单一数字造物（Digital Artifact），即使其底层机械原理已知，在极致规模与数据压缩下产生的能力边界和潜在风险仍完全未知。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=0s) 软件工程创新的底层驱动力已从稀缺的顶尖工程人才转向巨额资本与算力投入，彻底重构了传统科技商业的“物理法则”。 - [02:32](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=152s) AI 在纯数学难题上的求解更多得益于跨领域的知识广度与严密公理系统，但其突破是否具备明确的市场经济效用仍面临严峻拷问。 - [08:54](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=534s) 数学与计算机科学的演进历史表明，算力暴力穷举能够推动学科跃迁至更高抽象层，使后来的开发者无需再为底层基础数据结构重复造轮子。 - [12:05](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=725s) 物理现实动力学往往具有计算不可约性（Computational Irreducibility），脱离经验测量状态方程的纯数学大模型无法直接跳步预测现实世界演变。 - [14:07](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=847s) 技术工具的迭代周期性引发社会恐慌；每一代人往往将自己入行时的工具视为“真理基线”，并将随后的自动化抽象误判为学科的堕落或“作弊”。 - [17:33](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1053s) 历史上计算机软硬件的重大突破均源于航海潮汐预测、军工弹道计算等紧迫的现实经济与军事刚需，AI 基础设施最终同样必须靠解决具体问题的杀手级应用来证明商业闭环。 - [35:34](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2134s) 编程范式正在从确定性的指令式与声明式编程，彻底转向基于自然语言交互、概率统计驱动的“向模型之神祈祷”的新范式。 - [41:20](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2480s) 传统“精益创业”神话在算力超级周期中被打破，早期风险投资应坚持正和思维，通过巨额资金支持优秀企业在私募阶段最大化做大潜在市场（TAM）。 - [43:55](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2635s) 深厚行业专业知识如今可以直接通过资本与大模型转化为垂直软件，终结了行业专家耗时数年物色并培训软件工程团队的低效模式。 - [46:28](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2788s) 现有科技巨头因组织架构包袱、现有客户群体与财务报表考核，形成了难以打破的体制物理学，使巨头在面对激进技术范式转移时屡屡陷入“创新者的窘境”。 - [52:31](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3151s) 由于财报利润压力与合规顾虑，巨头甚至在内部实行算力与 Token 配给制，反而将激进扩大资本支出以冲刺模型上限的先机让渡给了风投资本支持的初创团队。 - [55:07](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3307s) 传统对模型分布外泛化（OOD）能力的机械怀疑，在面对千亿美元级前所未有的超大规模数字造物时，无法解释极致算力压缩所可能涌现出的能力未知性。 - [58:03](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3483s) 在生物医疗等关键领域，AI 的核心杠杆在于海量文献与复杂数据的跨学科模式识别，但新药落地的根本路障依然是漫长严苛的人体临床试验。 - [01:01:48](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3708s) 原本在科学探索中被视为不可逾越的“无限工程难题”（如全量蛋白质构象搜索），正被 AI 与算力系统化地压缩为经济上可衡量的“资本有限问题”。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=0s) 范式转移：从“工程瓶颈”到“资本与算力驱动” 本期节目的开篇直接点出了当前人工智能浪潮最根本的范式转变：技术演进的瓶颈已经彻底迁移。过去，软件与科技创业面临的是极其严苛的工程人才瓶颈；而如今，“如果你给 20 个人 10 亿美元，他们真的能把这笔钱有效地花掉”。整个行业已经从受限于工程能力的阶段，演变为纯粹由资本与算力规模驱动的问题。这一转变颠覆了过往几十年的商业逻辑。 与此同时，巨头与初创公司的竞争格局也发生了意料之外的扭转。传统硅谷大公司普遍有一种自信，认为凭借庞大的资源定能“碾压所有小团队”，但在现实中，初创公司往往不直接对抗现有巨头，而巨头真正的注意力完全被同级别的对手牵制——微软最忌惮的是亚马逊和谷歌，根本无暇顾及创业生态。正是这种盲区，让像 OpenAI、Anthropic 以及 Cursor 这样的公司迎来了流星般的崛起。即便资本在宏观上依然稀缺、获取门槛极高，但一旦资金到位，它便能以惊人的杠杆直接换取技术突破。 ```mermaid flowchart LR accTitle: AI创新驱动范式的转移 accDescr: 对比传统软件工程瓶颈时代与当前资本与算力驱动时代的发展路径。 subgraph OLD["传统软件时代 (人才驱动)"] A["依赖稀缺高级工程师"] --> B["算法与工程瓶颈"] B --> C["线性产出与渐进创新"] end subgraph NEW["现代 AI 时代 (资本/算力驱动)"] D["精简团队 (如20人) + 巨额资本"] --> E["大规模算力集群 (Compute)"] E --> F["突变式模型能力提升"] end ``` --- ### [00:56](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=56s) 黎曼猜想时刻：数学突破与社会的认知分化 主持人 Erik Torenberg 引出话题：几天前 Jared Friedman 在社交平台发文称，他让 Claude 尝试证明“黎曼猜想（Riemann Hypothesis）”并要求其“更加努力地思考”。虽然尚不清楚具体取得了何种实质进展，但这引发了全网关于“AI 在数学领域取得重大突破”的广泛热议。Erik 借此向 Steven Sinofsky 和 Martin Casado 提问：我们究竟该如何理解这一现象？这对于数学和实际应用到底意味着什么？ Steven Sinofsky 率先回应，他坦言自己绝非数学家，但他认为这个节点极具象征意义，因为它清晰地将社会割裂为两种态度截然不同的人群： 1. **盲目兴奋派**：哪怕并不理解背后的具体数学原理（实际上全天下能看懂这类前沿命题的人屈指可数），单纯因为“AI 攻克了千古难题”这一事实而欢欣鼓舞； 2. **悲观怀疑派**：认定这要么是虚假噱头，要么是人类工作的终结，甚至哀叹人类将全面进入“蠢人时代（Idiocracy）”，因为所有深奥的智力工作都被机器代劳。 令人玩味的是，最感到兴奋的反而是真正身处前沿的数学家群体。这一现象让许多坚信“技术将带来失业与退化”的外行感到无比困惑：为什么这群最直接面临“被替代风险”的专业学者，反而表现得最为雀跃？Steven 指出，这一反差精准地映射了当下技术变革期的独特张力。 --- ### [02:32](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=152s) 公理系统、经济效用与“游戏玩家”的局限 Martin Casado 明确表示，他和 Steven 都是“系统架构与产品出身（Systems and Product Guys）”，并非纯粹的数学专家。从系统工程与市场经济的视角审视，Martin 提出了两条尖锐的观察： 首先，AI 解决长期未决的数学难题，是否真的能释放出巨大的**经济效用（Economic Utility）**？Martin 指出，如果把多年来全球所有研究特定数学难题的博士后薪水加在一起，总额可能少得可怜。社会之所以长久没有攻克某些数学问题，很可能只是因为**缺乏足够的市场经济激励去攻坚**，而非它在工程应用上具有不可逾越的关键瓶颈。 其次，大型语言模型在数学上的惊人表现**并不令人意外**。数学本质上是一个纯粹的公理化系统（Axiomatic System）。AI 拥有远超人类生物极限的广度，能够自如地从极其分散、孤立的数学子领域中抽取理论拼图，并迅速拼凑出严谨的解答。许多数学突破在被 AI 整理出来后，人类会发现逻辑步骤其实十分直接，“无非是借鉴了一门我碰巧不知道的偏僻数学分支”。 ```mermaid flowchart TD accTitle: AI解决数学难题的逻辑与现实脱节 accDescr: 展示AI在公理系统中的优势，及其与真实物理世界和商业价值之间的鸿沟。 subgraph CAPABILITY["AI 核心优势"] M1["海量知识广度 (跨领域交叉)"] M2["严密公理化系统 (Axiomatic Domain)"] M1 & M2 --> M3["极擅长定理证明与符号规则推理"] end subgraph DISCONNECT["现实与市场脱节"] M3 --> G["类似顶尖星际争霸玩家 (闭环游戏逻辑)"] G -. 无法直接映射.-> R["复杂的物理与现实世界"] G -. 经济激励长期缺失.-> E["缺乏明确的经济闭环与应用场景"] end ``` 针对行业内常有的“数学是通往 AGI 与通用推理的基石，掌握数学公理就能理解宇宙万物”的融资叙事，Martin 表现得尤为清醒甚至冷淡。他认为这在本质上依然类似于“训练出了一个史上最强的《星际争霸（StarCraft）》玩家”——它在既定的符号与规则游戏内所向披靡，但这种封闭域的能力究竟如何映射到混乱的物理现实和商业世界，依然存在巨大的问号： - “我连 12 维空间是什么概念、生活在 12 维空间的人是不是特别瘦小都完全搞不明白。” - 单纯在形式化逻辑中取得高分，并不等同于解决了现实世界中制约经济生产力的关键路障。 --- ### [06:54](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=414s) 抽象层跃迁 vs 职业生存危机 针对 Martin 的实用主义视角，Steven Sinofsky 提出了互补的观点：数学是一门通过层层叠加**抽象层（Abstraction）**向前演进的学科。这与历史学完全不同——历史学缺乏这种形式化的抽象，面对战争或饥荒，人们只能建立类似受力图（Force Diagrams）的经验模型来解释事实。而数学的演进逻辑，正是在积累到一定阶段后，将低阶的繁琐推演打包成更高阶的抽象概念。对于那些感到兴奋的数学家而言，AI 实际上承担了他们平日里最厌倦的 20% 的符号推演繁杂工作，从而让他们能够站上更高的抽象维度去拓展前沿边界。 Martin 随即指出了其中的微妙与讽刺：数学界内部的情绪其实是分裂的，一部分学者迎来了生产力解放，另一部分人则陷入了“存在主义危机（Existential Crisis）”。他做了一个耐人寻味的对比： - 如果 AI 彻底攻克了癌症，研究癌症的医学家绝不会感到抑郁，全人类只会欢呼雀跃； - 但如果 AI 攻克了某个数学难题，研究该领域的学者可能会陷入抑郁。 这促使 Martin 提出反思：某些纯数学研究问题存在的唯一现实效用，是否仅仅是“给研究者提供一份工作、让他们能够撰写论文并在结尾写下‘另一次尝试与错因分析’”？一旦问题被机器瞬间破解，掩盖在学术活动之下的经济与现实价值空白就被无情地戳破了。虽然 Martin 自嘲这番言论显得“过于愤世嫉俗”，但它切中了一个核心命题：在技术跨入新的抽象层时，旧有的智力劳动范式正在经受剧烈冲击。 ### [08:54](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=534s) 算法复杂度与四色定理：算力推动的抽象层跃迁 Martin Casado 与 Steven Sinofsky 回顾起当年的大学计算机教育。过去，离散数学和算法复杂度理论是计算机系逃不掉的必修课，而如今翻看各大高校的课程大纲，许多学校已经不再硬性要求。Steven 调侃起计算机科学泰斗高德纳（Donald Knuth）的名著《具体数学》（*Concrete Mathematics*），并拿自己母校康奈尔与斯坦福开起玩笑，提到当年在康奈尔由算法领域的传奇宗师约翰·霍普克洛夫特（John Hopcroft，他在斯坦福博士期间发明了 2-3 树）亲自教授 P vs NP 等理论。当时学界之所以对这类问题极度痴迷，是因为理论家设想，若能在多项式时间（而非指数级时间）内解决此类问题，像旅行商问题（TSP）这样的大规模组合优化就能迎刃而解——这在当年极其关键，因为整个计算机行业都严重受制于算力瓶颈（compute-bound）。以 AT&T 为例，想要在拥有 6000 个交换机节点的网络中计算最优路由，单是运行模拟就可能要耗费整整两年。 正是在这种背景下，四色定理（Four-Color Theorem，即任何平面地图只需四种颜色即可确保相邻区域不同色）的计算机证明成为了历史转折点。它并非像经典微积分那样给出优美简洁的人类解析证明，而是数学家先证明了可能的图结构组态是有限的，随后写出长达 200 页的组合情况，依靠计算机暴力穷举了所有可能解，硬生生砸出了“确实只需要四种颜色”的结论。这种“打擦边球式”的证明完全是算力加持下的产物，但对于拓扑学界和实际工程设定强边界却极具价值。这一历史案例给两人的核心启示在于：**计算抽象层的跃迁会重塑整个学科的工作范式**。一旦证明某一整类问题在新的抽象层级上能够被系统化解决，业界就能直接在该层级上构建全新工具，后来的开发者再也无需为底层的 2-3 树这类基础数据结构重复造轮子。 ```mermaid flowchart TD accTitle: 四色定理计算证明与抽象层跃迁流程 accDescr: 展示四色定理如何从数学理论化简为有限组态，再通过算力暴力穷举证明，最终促成高层抽象工具的发展。 subgraph Theory["理论化简阶段"] A["图论问题：四色猜想"] --> B["数学证明：潜在解形态为有限集"] end subgraph Compute["算力穷举阶段"] B --> C["计算机暴力遍历数千种构型"] C --> D["机器证明：四色定理成立"] end subgraph Abstraction["抽象层跃迁"] D --> E["确立全新数学边界与理论基准"] E --> F["构建更高抽象层工具\n(无需从底层2-3树从头实现)"] end ``` --- ### [12:05](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=725s) 数学能否表征物理现实：模拟的计算不可约性与 AI 的逻辑跳跃 话题随即转向更深层的科学哲学讨论：数学究竟能否完全精准地表征物理现象？如果 AI 真的攻克了大量高阶纯数学问题，这是否意味着人类能顺理成章地预测物理现实？Martin 坦言，自己在参与大型物理模拟系统研发时（如恒星爆炸坍缩模拟、飞机风洞模拟等），系统内部虽然在解复杂的微分方程，但这些方程全部建立在经验状态方程（empirical equations of state）和实际物理测量的基础之上。纯粹的数学推导脱离了实证数据，根本无法凭空模拟真实的物理演进。 这引发了对于“模拟是否在计算上不可约（computationally irreducible）”的质疑。如果一个物理过程必须一步一步完整计算才能得出结果，那么单纯在数学定理证明或算法上取得突破，并不等于能用 AI 绕过物理模拟本身。Martin 指出，AI 或许在建模优化、物流调度等运筹学领域大放异彩，但在“这颗恒星会不会爆炸”或“这座大楼会不会倒塌”等真实物理世界动力学问题上，纯数学求解与物理现实之间存在一条明显的鸿沟。目前舆论中充斥着一种“只要 AI 能解开所有数学题，就能预测世间万物”的宏大叙事，而这在逻辑上存在极其巨大的跳跃，缺乏任何扎实的实证支撑。 - **经验物理 vs 纯数学**：大型物理模拟软件的核心在于测量温度、压力等经验数据拟合出的状态方程，而非脱离实证的纯数学孤岛。 - **计算不可约性悬念**：复杂的真实物理过程若必须完整模拟时间步长，则无法直接用高阶数学捷径或大模型一步跳到终局。 - **过度外推的风险**：从“AI 擅长解数学奥数题或符号证明”推论到“AI 即将预测全物理宇宙”，是典型的概念偷换。 --- ### [14:07](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=847s) 从算盘、Curta 机械计算器到图形计算器：工具演变与教师的技术恐慌 Steven 自称本质上是一个“工具论者（tools person）”。为了具象化工具抽象的演变，他特意带来了几个实物道具：首先是一个他在北京潘家园地摊上砍价到仅花 7 美分淘来的中国传统算盘；接着，他拿出了一个极为精巧罕见的圆柱形机械装置——20 世纪中叶奥地利发明的柯塔计算器（Curta calculator）。这个形状酷似咖啡研磨机或胡椒磨的金属圆筒，内部塞满了 600 多个微型精密机械加工零件，在当年相当于纯机械版的算术计算机，如今复刻一台的成本可能高达 5 万美元。使用者通过拨动侧面的数位滑轨，正摇把手做加法，提起并倒摇做减法。Steven 感慨，每一次算力载体的迭代，都让整整一个阶梯的复杂计算化为平淡无奇的基础操作。 这种演进随即映照出教育与学术界对新技术的周期性恐慌。Martin 和 Steven 回忆起当年德州仪器（Texas Instruments）计算器进入校园的剧烈震荡：从只能算阶乘（最大只能算到 59!）的 TI-35，到后来具备方程求解和函数绘图功能的 TI-85 图形计算器。当年 TI-85 问世时，全美国的数学老师陷入了集体恐慌，痛呼“过去我们发一张坐标纸让学生手动画 XY 轴点，现在按个键就能自动画出曲线、解出方程，数学这门学科彻底完蛋了！”为了防范学生依赖机器，大学考试强制要求使用蓝色答题本（Blue Book）详尽写出每一步手算推导。Steven 一针见血地指出其背后的社会心理机制：**“人们对‘变化’的反应远比对‘基线’强烈得多。”** 当年微积分诞生时颠覆了传统算术，但后来的一代人却理所当然地将其视为起点；每一代人都会将自己入行时的工具视为不可逾越的“真理基线”，而将随后的每一次自动化抽象视为学科的堕落与危机。 ### [17:33](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1053s) 从潮汐预测到弹道计算：驱动数学与计算创新的经济与军事刚需 回顾数学与计算机的发展史，技术的每一次阶跃几乎都源于迫切的现实需求。从代数、线性代数、微积分，再到傅里叶变换（Fourier transforms）和流体力学，数学工具的诞生始终受到实际应用场景的驱动。计算机的雏形差分机最初是为了计算积分，但追溯其技术源头，最早的机械积分器是为了预测潮汐——在航海贸易和港口物流中，精准预测潮汐蕴藏着巨大的商业与经济价值。随后，这类架构在战争时期被军方迅速征用，转向计算火炮与导弹的弹道抛物线表格。与人工计算相比，机器的计算速度提升了约 5,000 倍，而且彻底消除了人为失误，展现出无可替代的经济与军事效用。 在冷战背景下，这种对基础微积分的攻坚和弹道计算的需求，直接催生了太空竞赛、喷气式发动机研发以及现代工厂自动化等一系列技术奇迹。当时的社会文化对计算与数学突破报以狂热支持，不仅全民欢呼雀跃，家长们也纷纷敦促孩子钻研数理化，争取在西屋科学天才奖（Westinghouse Science Talent Search）或通用电气（GE）数学竞赛中崭露头角。对比当下，马丁·卡萨多（Martin Casado）提出了一个核心疑问：如今的大型语言模型在某些数学领域确实表现优异，但这种数学究竟是在攻克阻碍经济发展的现实瓶颈，还是仅仅在抽象的数字迷宫中自娱自乐？目前尚无定论。 - **微积分的工程化落地**：最初为战争研制导弹弹道表和火炮射击表的计算技术，后续全面外溢，支撑了空间探索、喷气推进技术与工业流水线自动化。 - **冷战时期的科学狂热**：全社会将数学能力视为面向未来的核心竞争力，催生了西屋科学奖等标志性青少年竞赛的繁荣。 - **AI 数学的效用拷问**：现代 AI 在数学形式系统上的突破，是否具备当年潮汐预测或弹道计算那样清晰的经济杠杆与实际效用，仍待检验。 --- ### [19:43](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1183s) 1953 年 IBM 宣传手册与计算机科学的五大经典抽象架构 史蒂文·西诺夫斯基（Steven Sinofsky）展示了一份他在家中偶然淘到的珍贵藏品——一份由 IBM 在 1953 年印制的宣传手册，题为《IBM：照亮未来》（*IBM: Light on the Future*）。这份手册诞生于 ENIAC 之后、IBM 704 与 System/360 系列问世之前的计算黎明期，封面印着浓郁冷战科技风的火箭、试管、示波器波形，以及一个头顶环绕着原子轨道的未来人形象。手册开篇写道：“人类耗费了数百万年才发明轮子并认识到其效用。”当时的公众甚至连二进制、十进制与十六进制的概念都不清楚，手册不得不拿罗马数字作类比，并在附录词汇表中向人们艰难地解释什么是“算术单元”、“二进制位（bit）”、“阴极射线管”以及“静电存储管”。 ```mermaid flowchart TD accTitle: 1953年IBM提出的数字计算机经典五大抽象架构 accDescr: 展示输入、存储、运算、控制和输出五大经典模块及其向现代计算机科学分支的演进 subgraph Core["1953年经典计算机架构五大支柱"] Input["输入 (Input)"] Storage["存储 (Storage)"] Arithmetic["运算 (Arithmetic)"] Control["控制 (Control)"] Output["输出 (Output)"] end subgraph Evolution["工程与学科演进"] Storage --> S_Evol["存储介质迭代<br/>电子管 → 磁鼓 → 磁盘 → 磁带"] Output --> O_Evol["终端形态进化<br/>电传打字机 → 黑白终端 → 彩色矢量显示"] Control --> Net["网络技术分支<br/>(后独立演进为系统核心)"] end ``` 在这份历史档案中，真正触及计算本质的是它对数字计算机组织的定义：**输入（Input）、存储（Storage）、运算（Arithmetic）、控制（Control）与输出（Output）**。这五大模块奠定了过去 75 年计算机科学的抽象框架。高校的计算机系（最早大都因导弹计算需求脱胎于数学系）课程设置完全映射了这五大领域，工程师往往穷尽整个职业生涯去攻坚其中某一模块——例如在存储领域见证从静电管、磁鼓、旋转磁盘到磁带的变迁，或在输出领域经历从行式电传打字机到黑白显示终端再到彩色矢量图形的迭代。卡萨多补充道，人们在五大抽象之外往往容易忽略“网络（Networking）”，网络本质上属于控制的延伸，但在数据包离开主机之后，网络学科在 20 世纪 90 年代后期逐渐成熟，如同晶体管物理机制从计算机核心课程中被层层抽象掉一样，最终沉淀为了理所当然的基础设施。 - **1953 年 IBM 手册的历史意义**：在普通大众对计算机毫无概念的时代，IBM 试图向商业社会描绘可编程机器如何不仅能算数，还能处理文本。 - **五大经典抽象**：输入、存储、运算、控制、输出构成了现代系统架构的底层共识，随后整合演进为当代的“系统工程（Systems）”大类。 - **抽象层的上移**：早期计算机专业学生必须搞懂晶体管物理特性，随着层次抽象，晶体管变成了电路图上的一个三角符号，网络协议与底层硬件也逐渐成为开箱即用的透明层。 --- ### [23:33](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1413s) 实用主义 vs. 纯粹解题：从布莱切利园、AlphaGo 到杀手级应用 面对技术突破，技术人员与社会大众的反馈往往取决于该技术究竟是在解决真正的经济难题，还是仅仅完成了一场智力炫技。在二战期间，布莱切利园（Bletchley Park）图灵等人的计算创新是为了破译恩尼格玛密码以挽救国家命运；ENIAC 的建立是为了核武器研发与战时弹道测算。这些突破被全社会奉为“拯救人类”的壮举，物理学家和工程界立刻全盘接纳并投入使用。然而反观深蓝（Deep Blue）战胜国际象棋特级大师或 AlphaGo 击败李世石的时刻，虽然技术上震撼，但行业却不得不费尽口舌向公众解释“这为什么是一件好事”，因为其背后的直接经济效益并不直观。 西诺夫斯基指出，如果人们把“解决问题本身”当成了目的，而忽略了问题所服务的目标，那么这种突破就如同僧侣绘制“沙画曼陀罗”——或许能带来智识上的圆满，却无法推动实体经济前进。对于当下的 AI 技术也是如此，卡萨多感慨道：“如果你为某个问题被机器攻克而感到难过或失落，或许说明那个问题一开始就不值得花费精力去解决。”从系统底层走向应用层是所有技术浪潮的必经之路。正如 1995 至 1996 年互联网兴起之初，大众也曾茫然困惑于“这东西到底对我有什么用”。西诺夫斯基特别推荐了关于史蒂夫·乔布斯离开苹果创业时期的传记《流亡中的史蒂夫·乔布斯》（*Steve Jobs in Exile*），书中记述了他与埃德·卡特姆（Ed Catmull）、丹·列文（Dan'l Lewin）在 NeXT 和皮克斯的探索历程：伟大的硬件与系统唯有催生出切中用户痛点的杀手级应用，才算真正完成了商业闭环——就连蒂姆·伯纳斯-李（Tim Berners-Lee）发明万维网时，使用的也是一台 NeXT 工作站。 - **历史使命与社会认同**：伴随军工、密码破译和核能工程崛起的早期计算机技术，由于具备立竿见影的生死存亡价值，享有极高的社会支持度。 - **棋类 AI 的效用鸿沟**：从深蓝到 AlphaGo，纯粹追求在封闭规则游戏中战胜人类，容易陷入“为了解题而解题”的象征意义，难以直接转换为生产力。 - **应用层（Apps）决定技术胜负**：如同 NeXT 工作站最终成为万维网的发源地，AI 基础设施当前同样面临 1995 年互联网初期的拷问，其真正价值需要由解决具体问题的新一代应用来证明。 ### [26:33](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1593s) 新平台的早期迷茫：从 NeXT 到“外卖披萨” 每当全新的计算平台问世时，人们往往对其真实用途一无所知。蒂姆·伯纳斯-李（Tim Berners-Lee）当年正是用一台 NeXT 工作站编写出了最初的 HTTP 协议与万维网原型。然而，即便他做出了这个划时代的成果，外界依然茫然不解——伯纳斯-李当时给出的解释只是“这东西可以用来查找其他研究人员的电话号码和共享论文”。 西雅图曾有一家名为 CyberSlice（节目中提及的 Cyber Pizza）的早期互联网公司，堪称那个时代的披萨版 DoorDash。在 NeXTSTEP / OPENSTEP 的发布会舞台上，创始人做了现场点餐演示：系统根据用户位置自动寻找附近的披萨店并下单送达。为了防止现场演示翻车，团队甚至在后台偷偷备好了现成的热披萨。尽管演示很震撼，台下观众当时的真实反应却是：“哇，这太酷了，但你难道没听说过电话吗？”这深刻说明了新计算平台在诞生初期，往往要面对公众对其必要性的普遍质疑。 ```mermaid flowchart TD accTitle: 技术演化中的早期认知断层 accDescr: 展示新技术从概念诞生到公众理解之间经历的认知落差。 subgraph 平台早期探索 A["新平台问世<br/>(NeXT 工作站 / 早期 PC)"] --> B["原生杀手级应用诞生<br/>(万维网原型 / 早期在线订餐)"] end subgraph 公众与传统认知 B --> C{"公众反应"} C -->|"怀疑现有价值"| D["'我们不是已经有电话了吗？'"] C -->|"抵制与恐慌"| E["视作作弊 / 校园全面禁用"] end subgraph 范式迁移 D --> F["更高层级的抽象被完全接受"] E --> F end ``` ### [27:50](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1670s) 25 磅的 Osborne 与哈佛法学院的“作弊”禁令 到了 1982 年左右，字处理软件开始在 Apple II 以及运行 CP/M（DOS 的前身）的电脑上兴起，彻底颠覆了传统的打字机打字模式。史蒂文·西诺夫斯基（Steven Sinofsky）拿出了自己大学时代的珍藏——重达 25 磅、外形堪比缝纫机的 Osborne 1 电脑。当被问及它的电池续航时，埃里克·托伦伯格（Erik Torenberg）猜测“应该不长吧”，西诺夫斯基揭晓了答案：它根本没有电池，必须插着电源线使用，体积甚至超过了民航合法的随身行李限制。 这种设备在诞生之初遭遇了极大的文化阻力。在哈佛法学院，原本学生们都是带着打字机进考场，以便教授能看清卷面。但当两名学生分别带着一台 Apple II 和一台 Osborne 1 进考场时，校方立刻下达了禁令。当时的《时代》周刊和《纽约时报》对此大加报道，其抵触论调与后来人们反对图形计算器、反对说唱乐、反对爵士乐、反对《龙与地下城》，乃至今日大学抵制 AI 如出一辙。西诺夫斯基回忆，他在 1983 年秋季刚入读康奈尔大学时，是整栋 90 人宿舍楼里唯一拥有电脑的学生，当时他甚至必须获得系主任的特批，才被准许用电脑完成大学一年级英语课的论文写作；而几年前他向康奈尔建议在写作课中引入 AI 时，校方却直接回避了沟通。 ### [30:26](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1826s) 抽象层级的演进，还是人类逻辑的彻底让渡？ 马丁·卡萨多（Martin Casado）对“AI 只是又一次抽象层级提升”的观点提出了深刻质疑。在以往的计算机科学历史中，计算、网络与存储始终只是人类调用的资源，人类负责定义问题规范、决定程序逻辑并确保正确性，计算机仅用来执行算力密集型的求解任务；即便是使用第三方函数库或使用谷歌搜索，其边界和输入输出也是明确可控的。但当前的 AI 却截然不同——人类似乎正在“放弃思考与逻辑本身”（abdicating logic），演变成直接向机器索要最终答案，甚至很多时候人类连问题本身该如何定义都不清楚。 西诺夫斯基用自身的经历反驳称，每一代技术突破都会给旧观念带来“作弊”的冲击感。当图形计算器出现时，考题恰恰就是“绘制函数图像”，这让工具显得像在作弊；1980 年代初麻省理工学院（MIT）符号数学计算包 Maxima（以及运行在 IBM PC 上的 muMATH）横空出世，学生可以直接把微积分积分题敲进电脑得出答案，学生们跑到工程图书馆借一张 PC 软盘就能搞定微积分作业，当时的人们同样感到这不可思议且形同作弊。 ### [32:47](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1967s) 专家系统与现代 AI：确定性算法与端到端决策的本质分野 两人的论战进一步延伸至 1980 年代斯坦福大学的“AI 寒冬”与专家系统（Expert Systems）时代。西诺夫斯基指出，当时斯坦福曾大力推动 AI 与医学院结合，研发用于医疗诊断、化疗方案制定及有机化学合成的系统，试图让机器接管人类的决策过程。卡萨多反驳道，那个时代的专家系统与 Prolog 逻辑编程语言与今天的生成式模型有着本质区别： - **专家系统的局限性**：过去的 Prolog 规则系统本质上仍然是确定性的算法实现（algorithmic），程序的目标终态（end state）完全由人类程序员预先设定，计算机仅是在明确定义的逻辑规则图谱中搜索达成路径，而且“当年那套根本跑不通”。 - **生成式 AI 的范式颠覆**：当下的 AI 则是真正能跑通的实用系统，但人类不仅让机器去寻找路径，甚至连“终态应该是什么”都直接丢给了模型，逻辑与正确性的控制权被彻底转移。 西诺夫斯基总结道，人们面对新技术时产生的心理不安与焦虑（"the willies"），往往源于当下所处的特定语境与社会背景。正如公众对数据中心建设的态度一样，两年前各地州长还在争先恐后招商引资，十年前大家还在争夺重污染的汽车制造厂，如今围绕算力建设和 AI 逻辑让渡的社会恐慌，本质上也是技术迭代与旧有认知范式激烈碰撞所引发的周期性反应。 ### [35:34](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2134s) 跨越抽象层：从确定性指令到概率性“祈求模型之神” Martin Casado 与 Steven Sinofsky 展开了一场关于计算堆栈（computing stack）本质的深刻辩论。Martin 回顾自己的整个职业生涯，计算技术的演进一直是在技术栈上不断向更高抽象层跃升——从底层的晶体管逻辑（transistor logic），到硬件、操作系统（OS）、应用程序、再到平台层。以往每一次向上迁移，底层依然是完全确定性（deterministic）的计算机逻辑，任何高层抽象最终都能精准映射回底层指令。但当前的生成式 AI 第一次让人感觉踏入了一个截然不同的层级：它不再像传统计算抽象那样能够向下严格映射，而是一种更加契合人类心智的“人类级抽象”（human-level abstraction）。 Steven Sinofsky 提炼了编程范式的三次跃迁：最初行业是从纯粹的算术计算跨越到指令式编程（imperative programming），由程序员编写精确的“步骤与食谱”；之后曾短暂进入以数据驱动程序的阶段，即推理系统以及类似 Prolog、Datalog、SQL 或 Makefile 的声明式编程（declarative programming）——人们只需定义最终状态（end state），把执行路径交给计算机，尽管消耗的计算时间难以严格界定；而现在，AI 带来了全新的随机（stochastic）、统计驱动的范式。Martin 形象地调侃道，现在的开发者甚至无法预先界定具体的目标终态，“你只是用恰当的词句向‘模型之神’祈祷，然后它就会吐出一个恰好有用的答案”。由于大模型的输入和输出直接交织于人类的自然语言与现实世界的非结构化表达，这标志着计算科学正前所未有地与自然界和人类认知现象深度交融。 ```mermaid flowchart LR accTitle: 计算与编程范式的演进历程 accDescr: 展示从早期的指令式编程，到声明式编程，再到当前统计与概率驱动的 AI 范式的演进路径。 subgraph S1["传统确定性范式"] A["指令式编程<br/>(Imperative)"] -->|"明确执行步骤<br/>写明食谱 (Recipe)"| B["确定性结果"] C["声明式编程<br/>(Declarative: SQL / Prolog)"] -->|"明确终态 (End State)<br/>机器自寻求解路径"| B end subgraph S2["AI 新范式"] D["统计与生成式模型<br/>(Stochastic / Statistical)"] -->|"提示词交互<br/>向模型祈祷 (Pray to Model God)"| E["人类级语言输出<br/>概率性且有用的解答"] end S1 -.->|"打破确定性映射"| S2 ``` --- ### [38:16](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2296s) 从“工程瓶颈”到“资本瓶颈”：重写软件商业的底层物理法则 面对这种颠覆性范式，讨论转向了对科技商业基本假设的彻底反思。Martin 指出，像他们这样在行业摸爬滚打四五十年的技术专家和投资人，对于软件系统运作、价值分配（价值流向基础模型还是上层应用）、资本效率及生产力提升早已建立了极深的直觉与前置假设（priors）。然而如今，支撑这些直觉的“物理法则”（laws of physics）正在瓦解。 Martin 提出了一个极具震撼力的思想实验来对比资本在不同时代的吸收能力： - **20 年前**：如果给一个 10 人的初创团队 10 亿美元，团队根本无所适从。除了盲目采购大量惠普（HP）服务器、疯狂招人，最终只会导致公司因管理混乱而迅速炸毁。 - **10 年前**：如果给一个 10 人软件团队 10 亿美元，同样难以花掉。软件工程受到《人月神话》（*The Mythical Man-Month*）与布鲁克斯法则（Brooks's Law）的死死约束，软件规模受制于沟通与工程复杂度，“堆人”非但不能加快进度，反而会拖垮系统。 - **现如今**：如果把 10 亿美元给一个仅有 20 人的前沿 AI 团队，他们能够极其高效、合理地将这笔巨资全部转化为算力与模型训练资源。 这一质的飞跃意味着，软件创新从一个长达数十年受限于“工程能力”（engineering-bound）的难题，彻底转变为了“资本密集型”（capital-bound）问题。Steven Sinofsky 补充指出，其实计算产业最初的 30 到 40 年也是高度受制于资本的——就像美剧《广告狂人》（*Mad Men*）里的经典场景，广告公司斥巨资买回一台庞大笨重的大型机，全员跑来围观却根本不知道它能干什么，拥有机器本身就是一种拥有雄厚资本的实力象征。过去几十年，行业从资本受限走向工程人才受限；而今，AI 竟然让整个行业“倒退回 40 年前”，再度演变为极度依赖资本的战场。 --- ### [41:20](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2480s) 打破精益创业神话：VC 零和思维与超级周期的资本逻辑 随着资本再度成为核心驱动力，关于创业方法论与风险投资的争论也迎来了全新视角。Erik Torenberg 提到 5 年前 Patrick Collison（Stripe 创始人）采访 Sam Altman 时的提问：在硅谷长期奉行“精益创业”（Lean Startup）的时代，Sam Altman 无论是做 OpenAI、核能项目还是长寿研究，为何从一开始就能募集天文数字般的资金？这种模式是否一直被市场低估？ Martin 深入剖析了硅谷历史上埃里克·莱斯（Eric Ries）的“精益创业”与本·霍洛维茨（Ben Horowitz）倡导的“肥胖创业”（Fat Startup，即储备巨额资本果断全力出击）之争。在过去，尽管“肥胖创业”逻辑上可行，但工程复杂度的极限始终是一道客观天花板。Martin 感慨自己在早期 VC 投资工作中，最痛苦的莫过于给团队投了钱，帮他们招募工程师，然后坐等漫长的两年看着工程进度原地爬坡——因为工程团队规模往往是无法线性扩张的。而如今，前沿 AI 探索建立了一套可以用极小的精英团队高效消化巨额资本的新规律。 最后，Martin 尖锐抨击了当下风险投资界弥漫的“零和思维”（zero-sum thinking）。很多投资人频繁抱怨“市场资金泛滥，过多资本在追逐过少的优质标的”；Martin 认为这种悲观论调对 VC 而言极其荒谬与短视： - 优秀的早期风险投资人本应是坚定的“正和思维”（positive-sum）信徒。 - 更多资本注入一级市场，不仅不会摊薄收益，反而从根本上做大了整体潜在市场空间（TAM）。 - 一方面，AI 等底层技术浪潮拥有前所未有的庞大吞吐资本的能力；另一方面，充沛的一级市场资金允许优秀公司在私有阶段停留更长时间，让绝大部分爆发性价值在未上市之前就沉淀在私募生态中。这种资本正循环，正为即将到来的全新一波 AI 杀手级应用浪潮铺平道路。 ### [43:55](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2635s) 领域知识与资本的直接结合：告别十年工程师招聘 Martin Casado 与 Steven Sinofsky 讨论了技术抽象层提升所带来的颠覆性变化：过去，若想在尚未被软件覆盖的领域创业，创始人往往需要花费数年时间四处物色并招聘顶尖工程师；而现在，资本可以直接转化为产出。如今世界上绝大多数行业其实依然处于未被软件充分服务的状态——无论是医疗记录管理、诊所预约挂号，还是法律服务。谈及法律行业时，Sinofsky 幽默地指出，在当今世界上几乎每个人都反对各种繁文缛节，但唯一能让大家同声欢呼的事情就是“终于能用 AI 把律师给自动化了”。过去风险投资中最纠结的情景莫过于：某人在商业地产领域深耕了二十年，非常懂行业痛点，但他不懂写软件；传统的解决办法是找一个懂技术的联合创始人，再花几年时间把二十年的商业地产经验灌输给对方，这种方式极其低效且痛苦。但如今，将专业领域构想转化为现实软件，已经彻底变成了一个可以用资本与算力直接解决的高阶抽象问题。 Sinofsky 分享了自己作为专业产品开发者时的第一次客户拜访经历。当时年仅 20 岁的他去拜访一位医生，那位医生恰好在读医学院之前主修了早期计算机科学，并且自己纯手写了一个 DOS 程序用来管理诊所预约。当时年轻的 Sinofsky 以为诊所挂号不过就是带时间的日历，但那位医生向他解释：“不，你根本不懂。”当病人打电话进来时，分诊人员是在根据关键词进行判断：这个问诊是需要 5 分钟还是 20 分钟？病人是否需要做 X 光检查或心电图（EKG）？因此软件在后台其实是在并行调度抽血、检查设备和医生那宝贵的 10 分钟。那位医生花了好几年才敲出这套系统。而在今天，拥有深厚行业洞察的人不再需要用几年时间死磕底层代码，“无代码/低代码”的真正抽象层正在降临。甚至整个技术底座的抽象层都在上移：开发者无需再去手动构建一个文本输入框或 UI 控件，在手机和现代平台上开发，最繁重的基础构件早已成为现成设施。 ```mermaid flowchart TD accTitle: 软件开发范式与护城河的演进 accDescr: 对比传统软件开发由工程团队驱动与现代AI应用由领域专业知识与资本驱动的不同路径 subgraph Traditional ["传统软件开发范式"] T1["深厚行业洞察 (如医疗、地产、法律)"] --> T2["招聘稀缺软件工程团队 (耗时数年)"] T2 --> T3["手写基础构件 (操作系统/UI控件/分布式系统)"] T3 --> T4["漫长的业务逻辑实现与交付"] end subgraph Modern ["现代 AI 驱动范式"] M1["深厚行业专业知识"] --> M2["高层抽象工具 & 算力资本注入"] M2 --> M3["大模型 & 现代化平台现成底座"] M3 --> M4["快速生成端到端垂直行业解决方案"] end ``` --- ### [46:28](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2788s) 颠覆的物理定律：大公司为何扼杀不了初创企业 话题转向经典的管理学命题——现有巨头（Incumbents）与初创企业（Startups）之间的博弈与“创新者的窘境”（Innovator's Dilemma）。Casado 坦言，如果六个月前有人问大公司有什么优势，答案依然是老一套：巨头拥有无尽的资本、充沛的现金流和牢不可破的分发渠道（Distribution）。然而在当前的 AI 浪潮中，这种逻辑正在被打破。一方面，AI 产品面对的市场需求是近乎无限的——业界对 Token 和 GPU 算力的渴望永无止境，初创公司可以直接通过算力和资金投入来撬动漏斗顶端的用户增长，分发难题在极度饥渴的需求面前迎刃而解；另一方面，像 OpenAI、Anthropic、Cursor 这样的初创企业具备惊人的吸金能力，能够融到数百亿美元，从而直接在资本规模上与微软、Meta 等科技巨头平起平坐。 Sinofsky 回顾了自己在微软的职业生涯，当时他的核心工作就是管理数千名顶级工程师，去攻克除微软外全世界任何地方都造不出来的工程奇迹。比如为了打造 Windows NT 操作系统，必须依赖像 Dave Cutler 这样的传奇计算机科学家，打造操作系统的工程壁垒宛如不可逾越的天堑；即便当年史蒂夫·乔布斯在 NeXT 创业，也是直接沿用了卡耐基梅隆大学（CMU）的 Mach 微内核代码，根本无法从零凭空手搓。Sinofsky 引用了哈佛商学院的一个经典典故：在克莱顿·克里斯坦森（Clayton Christensen）撰写《创新者的窘境》时，哈佛商学院流传着一句玩笑——“把颠覆理论放在商学院当管理理论来讲真是奇怪，它根本就应该作为一条不可抗拒的定律被编入物理系课程”。 Sinofsky 进一步剖析了大公司无法轻易碾压小团队的深层机制。在大公司工作过的人往往有一种迷思，总觉得自己庞大的战舰开过去就能把所有小公司碾得粉碎，但现实中这种情况极少发生： - 初创企业通常不会直奔巨头的腹地开火，而是切入巨头忽视的死角。 - 巨头的注意力永远只聚焦在同体量的对手身上——微软真正在意的是亚马逊和谷歌在做什么，而不是雷达边缘的小初创。 - 真正决定颠覆的是大公司内部无法更改的“文化与体制物理学”：销售激励机制（Scorecards）、现场销售模式、薪酬结构、组织架构包袱以及庞大的存量客户。当你服务着 50 万家企业客户时，你的行为模式就被彻底锁死了，根本无法掉头做出激进的自我革命。 --- ### [50:38](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3038s) 从工程奇迹到算力资本：初创企业如何正面迎击巨头 Sinofsky 指出，当前的 AI 创业生态不仅继承了云计算时代的基础设施抽象优势——初创团队不需要自己建数据中心、拉网络专线、给 AT&T 打电话，在创业的第一天甚至一顿晚饭的工夫就能部署上线。但 Casado 提出了一个更本质的认知升级：在云计算时代，没有任何人妄想过能把 AWS 赶下台，大家默认接受了云巨头的寡头垄断，只是小心翼翼地在巨头的生态夹缝里寻找生存空间，整天提心吊胆巨头是否会推出一个免费功能把自己的小工具顺手捏死。 然而在今天的 AI 时代，初创企业不再是在边缘苟延残喘，而是在直接正面硬刚甚至击败巨头。Casado 借用普利策奖名著《新机器的灵魂》（The Soul of a New Machine）中的典故总结道：过去的计算机工业史完全是由极其庞大、复杂的工程奇迹所定义的——无论是当年 Data General 研发小型机、微软打造操作系统，还是早期 Jeff Dean 等传奇工程师在谷歌从零搭建分布式集群，那些壁垒都是初创企业无论如何都无法逾越的“人类智力与团队协作极限”。 但如今的基础大模型已经打破了这一规律：问题的本质已经从“不可复制的超级工程实现”转变为“对资本与算力的获取能力”。规则的物理定律彻底变了——只要初创公司能汇聚足够的资本和算力，就能直接在模型能力上完成突围。正因如此，才会出现即便谷歌拥有全球最海量的数据、顶尖的研究天才和无与伦比的基础设施，其核心大模型依然会被 OpenAI 和 Anthropic 屡屡击败的惊人景象，而这种落后的根源，恰恰在于大公司根深蒂固的文化与组织惯性。 ### [52:31](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3151s) 科技巨头的资本紧缩与传统工程文化的桎梏 局外人往往低估了大企业内部因文化惯性而遭遇的转型阻力。像 Google Cloud（GCP）这样的部门，其工程执行力毋庸置疑，但这种成功反而将其固化在既有的工程解决路径上。巨头们不仅在文化上面临转型难题，在调配巨额资本时也显得捉襟见肘。观察各大科技巨头的财报电话会议便可发现，它们对当下狂飙突进的资本支出充满了焦虑与不安。例如，Google 甚至通过发行债券等复杂的财技手段将开支移出资产负债表；更有传闻指出，某些巨头为了优先保障企业级客户的算力需求，不得不严格配给 Token，导致自家的内部产品长期处于 AI 算力饥渴的状态，而它们的外部竞争对手却不受任何算力限制。 ```mermaid flowchart TD accTitle: 巨头与初创公司的资本配置对比 accDescr: 展示现有巨头在资本与算力上的受限与初创公司的激进扩张策略 subgraph Incumbent["现有科技巨头"] A["财报业绩与资本焦虑"] --> B["限制内部算力支出 / 发债出表"] B --> C["Token 配给优先给企业客户"] C --> D["内部产品遭遇 AI 算力饥荒"] end subgraph Startup["AI 初创公司"] E["顺应资本反转逻辑"] --> F["风投机构无上限资金支持"] F --> G["全部资本直接转化为算力与模型训练"] end ``` 这种文化与战略的脱节在计算历史上屡见不鲜。Steven Sinofsky 分享了当年在微软主导初代 Surface 研发时的亲身经历：他带着 Surface 原型机拜访英特尔（Intel）高管团队，对方一开始表现得极为兴奋，但一听说机身内部搭载的是 ARM 架构芯片，整个会议室的氛围瞬间降至冰点。在当时的英特尔文化认知中，ARM 只配用在打印机这种边缘设备上，英特尔坚信摩尔定律主导的 X86 架构才是正统，傲慢地声称自己懂全部的架构许可。他们完全忽略了低功耗、优秀图形处理能力以及“始终连接”（Always Connected）对新计算形态的重要性。同样，今天的 Google 骨子里是做超大规模数据中心（Hyperscale）的，一旦 AI 趋势转向端侧（On-device），这就完全脱离了他们的文化舒适区；而微软此刻也在被两面夹击。相比之下，初创公司迎来了绝佳的机遇——在当前的“资本反转”（Capital Inversion）时代，只要创业团队敢于融资买算力，投资人绝不会投以怀疑的眼光，巨额融资也因而屡见不鲜。 - 英特尔早期对移动架构嗤之以鼻，认为 ARM 芯片不过是“打印机里用的芯片”，无法撼动 PC 计算格局。 - 巨头受制于公开市场的财务约束，导致内部业务反而出现算力配给制，失去了与初创公司全力赛跑的先发优势。 --- ### [55:07](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3307s) 超大规模数字造物的能力未知性与贝叶斯黑盒 关于大语言模型架构的终极局限性，科技界的讨论正在发生深刻转向。过去，包括许多知名技术专家在内的主流声音普遍对大模型能够带来的真正科学突破持悲观态度。从底层机械原理来看，大模型的运行逻辑非常清晰：输入海量数据，参数被锚定在数据流形（Manifold）之中，以严格的贝叶斯分布方式移动。模型既无法做出真正的分布外推理（Out-of-Distribution），也不具备所谓的跨领域无缝迁移学习，距离超级智能的“快速爆发”（Fast Takeoff）或奇点降临仍有本质鸿沟。 然而，Martin Casado 指出，人类的理性在一种全新的现象面前失效了：人类历史上从未打造过一个耗资 50 亿美元甚至 200 亿美元的单一数字造物（Digital Artifact）。即便其机械机理是已知的，但当将前所未有的海量数据和算力（FLOPs）全部压缩进这个模型时，没有人能够准确预言其能力的边界在哪里。这种由 Anthropic 等前沿机构所代表的“元经济机器”（Meta-economic Machinery）——即能够无上限吸纳数百亿美元资本并将其一滴不漏地倾注于模型训练的能力——彻底重塑了讨论的语境。现在的核心问题不再是“模型机械上是如何运转的”，而是“耗费 200 亿美元锻造出的系统到底能干什么？它是否能靠压缩一切已知而实质上攻克癌症？”这种对极致规模的未知，让哪怕最资深的行业观察家也必须承认无法做出精准预测。 - 传统的争议主要聚焦在模型能否跨越分布限制，而现在的共识转向探讨极限算力灌注下的能力涌现。 - 历史上从未存在过能吸收单次百亿美元级训练成本的单一软件或算法对象，这已经超出了经典计算机科学的评估框架。 --- ### [58:03](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3483s) 算力杠杆下的生物医药与百亿美元训练的真正风险 谈及跨学科的实际应用，Steven Sinofsky 提到了医疗科研的真实案例。他的妻子是一名使用 AI 进行脑外科与神经科学研究的医学博士。家中配有专用的英伟达（NVIDIA）工作站（现场两人还打趣澄清绝非 80 年代 Scott McNeely 执掌的 Sun SPARC 小型机，而是带 K 的现代化硬件）。AI 在该领域的惊人之处在于海量模式匹配的能力：如果某特定病理课题有上万篇学术论文，人类医生单凭经验绝不可能通读并建立关联，但 AI 可以迅速跨越这些论文捕捉到此前隐匿的模式，为研究指明全新路径。不过，Sinofsky 也强调，AI 并非解决新药研发的万灵药，因为药物研发的真正瓶颈从来不是发现候选分子（自 20 世纪 80 年代起合成化合物的能力就已远超测试能力），而是临床人体试验中漫长且高风险的安全性与有效性验证。 针对模型演进的本质，Martin 坦陈了自己此前的认知盲区：他曾坚决驳斥尼克·博斯特罗姆（Nick Bostrom）关于递归自我改进（Recursive Self-Improvement）和 AI 瞬间统治世界的论调，因为所谓快速爆发并未发生；但他此前未能预料到的是，Scaling Law（缩放定律）竟然能一直成立，并且金融资本机器能够支撑起高达 1000 亿美元的单次模型训练。这种将全人类庞大资源以前所未有的密度集中于单一技术方向的能力，才是真正意义上的未知与潜在危险。如果有人把千亿美元级的单一数字机器用于研发武器而非攻克癌症，其后果将无人能预估。这种以资本为驱动的指数级暴涨，再次印证了人类思维对指数级增长认知的极度匮乏——正如当年英特尔不断推高芯片主频从几个兆赫兹到 5GHz 时，工程师们完全不知道未来究竟该拿这些晶体管做什么一样。 - 现代 AI 在生物医药中的主要价值是跨越海量文献与复杂数据进行模式识别，而非凭空越过严苛的临床安全与有效性检验。 - 当前真正的行业挑战不再是科幻式的自我意识觉醒，而是面对千亿美元规模训练出的极高浓度资源造物，人类缺乏应对其潜在用途（乃至恶意武器化）的评估与治理工具。 ### [01:01:48](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3708s) 将无限工程难题转化为资本问题：创新的物理法则重构 讨论临近尾声时，嘉宾们提炼出本期对话中最核心的范式转移：AI 正在根本性地重塑创新的经济学逻辑。以生命科学领域为例，如果现在的目标是“穷尽式探索每一种蛋白质组合”，在过去这几乎是一个不可逾越的无限问题，受制于极其复杂的工程瓶颈与实验局限；但在如今的计算与 AI 架构下，这可以直接被转化成一个“金钱问题（money problem）”。资本与算力的注入，使得原本在传统视角下近乎无限的探索空间，迅速收敛为在经济上可量化、可计算的有限问题。 这一转变深刻改变了技术研发的本质。以往依赖极少数顶尖工程师艰难攻关、反复试错的工程壁垒，正被直接替换为纯粹的资本规模竞赛——正如对话中所总结的，这让难题“变成了资本问题，而非工程问题”，其运作机理遵循着一套与传统软件工程截然不同的“物理法则”。在确立了这一极具启发性的总结之后，主持人 Erik Torenberg 提议就此收尾，感谢了 Martin Casado 与 Steven Sinofsky 的深入对谈，为整期节目画上了圆满句号。 - **无限到有限的资本化转化**：以往被视为计算或搜索空间近乎无限的难题（如全量蛋白质构象筛选），如今只要投入足够的资本与算力即可求解。 - **创新范式重塑**：研发创新的制约要素正从“顶尖工程人力的稀缺性”转变为“资本分配与计算资源的吞吐量”，打破了传统软件研发的成本与规模假设。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=0s) 如果你给 20 个人 10 亿美元，他们真的能把这笔钱有效地花掉 — Martin Casado > [02:32](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=152s) 我连 12 维空间是什么概念、生活在 12 维空间的人是不是特别瘦小都完全搞不明白。 — Martin Casado > [02:32](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=152s) 无非是借鉴了一门我碰巧不知道的偏僻数学分支 — Martin Casado > [14:07](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=847s) 人们对‘变化’的反应远比对‘基线’强烈得多。 — Steven Sinofsky > [14:07](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=847s) 过去我们发一张坐标纸让学生手动画 XY 轴点，现在按个键就能自动画出曲线、解出方程，数学这门学科彻底完蛋了！ — Steven Sinofsky > [19:43](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1183s) 人类耗费了数百万年才发明轮子并认识到其效用。 — Steven Sinofsky > [23:33](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1413s) 如果你为某个问题被机器攻克而感到难过或失落，或许说明那个问题一开始就不值得花费精力去解决。 — Martin Casado > [26:33](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=1593s) 哇，这太酷了，但你难道没听说过电话吗？ — Steven Sinofsky > [35:34](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2134s) 你只是用恰当的词句向‘模型之神’祈祷，然后它就会吐出一个恰好有用的答案 — Martin Casado > [43:55](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2635s) 终于能用 AI 把律师给自动化了 — Steven Sinofsky > [46:28](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=2788s) 把颠覆理论放在商学院当管理理论来讲真是奇怪，它根本就应该作为一条不可抗拒的定律被编入物理系课程 — Steven Sinofsky > [01:01:48](https://www.youtube.com/watch?v=GHPB1MwlKU0&t=3708s) 变成了资本问题，而非工程问题 — Martin Casado ## Mentioned Martin Casado, Steven Sinofsky, Erik Torenberg, Jared Friedman, Donald Knuth, John Hopcroft, Steve Jobs, Ed Catmull, Dan'l Lewin, Tim Berners-Lee, Sam Altman, Patrick Collison, Eric Ries, Ben Horowitz, Clayton Christensen, Dave Cutler, Jeff Dean, Nick Bostrom, Scott McNeely, a16z, OpenAI, Anthropic, Cursor, Microsoft, Google, Amazon, Meta, Apple, Intel, NVIDIA, IBM, AT&T, NeXT, Texas Instruments, CyberSlice, Pixar, Sun Microsystems, Claude, Osborne 1, Surface, TI-85, TI-35, Curta 计算器, AlphaGo, Deep Blue, ENIAC, 《人月神话》, 《创新者的窘境》, 《新机器的灵魂》, 《IBM：照亮未来》, 《流亡中的史蒂夫·乔布斯》, 《具体数学》, 黎曼猜想, 四色定理, P vs NP, 旅行商问题 (TSP), Prolog, 康奈尔大学, 斯坦福大学, 哈佛法学院, 麻省理工学院 (MIT), 卡耐基梅隆大学 (CMU) ## Video description a16z General Partners Martin Casado and Erik Torenberg are joined by Board Partner Steven Sinofsky to explore what recent breakthroughs in AI and mathematics tell us about where the technology is headed, and whether some of the basic assumptions that have governed computing for decades are starting to break. Martin and Steven debate whether AI's progress in mathematics represents a genuine leap in reasoning or simply a new tool for solving problems at a higher level of abstraction. From the four-color theorem and early computers to graphing calculators and today's models, they trace how new technologies have repeatedly changed which problems humans need to solve themselves, and ask what makes this moment different. The conversation then turns to one of the biggest shifts in AI: problems that were once constrained by engineering talent can increasingly be attacked with capital and compute. They discuss what that means for startups versus incumbents, venture capital, the coming wave of AI applications, and why pouring billions into increasingly capable models may force us to rethink what these systems can ultimately accomplish. Timestamps: 00:00 - Intro 00:56 - Making Sense of AI & Math: The Riemann Hypothesis Moment 12:05 - Will AI Math Ever Map Onto Physical Reality? 19:43 - The Cold War, IBM 1953 & the Cultural Roots of Computing 38:16 - Rethinking Fundamental Assumptions About Software 46:28 - Incumbents vs Startups: Why the Innovator's Dilemma Still Wins 55:05 - The Limits of Current AI Architecture & What Comes Next Resources: Follow Martin Casado on X: https://x.com/martin_casado Follow Steven Sinofsky on X: https://x.com/stevesi Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=GHPB1MwlKU0)

[0:00] Martin Casado: 如果你给 20 个人 10 亿美元，他们真的能把这笔钱有效地花掉 [2:32] Martin Casado: 我连 12 维空间是什么概念、生活在 12 维空间的人是不是特别瘦小都完全搞不明白。 [2:32] Martin Casado: 无非是借鉴了一门我碰巧不知道的偏僻数学分支 [14:07] Steven Sinofsky: 人们对‘变化’的反应远比对‘基线’强烈得多。 [14:07] Steven Sinofsky: 过去我们发一张坐标纸让学生手动画 XY 轴点，现在按个键就能自动画出曲线、解出方程，数学这门学科彻底完蛋了！ [19:43] Steven Sinofsky: 人类耗费了数百万年才发明轮子并认识到其效用。 [23:33] Martin Casado: 如果你为某个问题被机器攻克而感到难过或失落，或许说明那个问题一开始就不值得花费精力去解决。 [26:33] Steven Sinofsky: 哇，这太酷了，但你难道没听说过电话吗？ [35:34] Martin Casado: 你只是用恰当的词句向‘模型之神’祈祷，然后它就会吐出一个恰好有用的答案 [43:55] Steven Sinofsky: 终于能用 AI 把律师给自动化了 [46:28] Steven Sinofsky: 把颠覆理论放在商学院当管理理论来讲真是奇怪，它根本就应该作为一条不可抗拒的定律被编入物理系课程 [61:48] Martin Casado: 变成了资本问题，而非工程问题

[2:38] Martin Casado: The solution was pretty straightforward... It just borrowed from a bit of math that I didn't know. The metalearning here is there is a set of problems that require you being too broad for most humans or most education, and it's going to solve those. [18:02] Steven Sinofsky: Those people didn't complain about when calculus came out because calculus was a baseline to them... people react to change more than they react to the baseline of where they all started. [23:33] Martin Casado: If you're really like sad about something being solved, maybe it wasn't worth working on to begin with. [33:34] Martin Casado: In imperative programming, you know all of the steps... In declarative programming, you know the end state... And this is like this new thing where you don't really know what the end state is specifically, and you just prey to the model God in like the right words and then it produces the answer that just ends up being useful. [43:55] Martin Casado: Everybody is against everything except having more lawyers. [43:55] Steven Sinofsky: No, you don't understand... they're actually scheduling like a blood draw and all of this stuff in parallel, not just the 10 minutes you need with the doctor. [46:28] Steven Sinofsky: Disruption... really it should just be a fact in the physics department. [46:28] Steven Sinofsky: Here's our new computer... 'But what's in here?' I said, 'Well, it's an ARM chip.' [55:05] Martin Casado: In the history of humanity we've never created a single digital artifact that had that many FLOPs and that much data in it. [55:05] Martin Casado: I have decided that I cannot predict what an artifact... that you used $20 billion to create is capable of. [55:05] Steven Sinofsky: I wrote 58 memos on what the internet was going to be, and I was wrong a lot of them by far. [61:02] Steven Sinofsky: We can take previously infinite problems and apply capital and it becomes finite.

[0:00] Right now, if I give 20 people a billion [0:03] dollars, they can actually use it [0:04] usefully. We've kind of moved the [0:06] industry from like this engineering [0:07] bound problem to a capital problem. [0:09] That's fundamentally very different. [0:11] >> Math is very much a leading edge [0:13] indicator of what the market might be [0:15] interested in. Why? Some people will [0:17] walk in and say the foundations to AGI [0:20] and to reasoning is going to be math, [0:22] but like that doesn't tell you anything [0:23] about reality. For me, it's still in the [0:24] domain of like it's really good at [0:26] playing a game. The startups don't aim [0:28] straight at the incumbents and the [0:30] incumbents just don't pay attention. [0:31] Microsoft is worried way more about what [0:33] Amazon and Google are doing than anyone [0:36] in a startup space. Everybody who's from [0:38] a big company in Silicon Valley, you [0:39] always think, "Oh my god, we're just [0:40] going to crush all of these little [0:41] companies." And then you realize they [0:43] never get crushed. And I think this is [0:45] why we're seeing such meteoric roasts of [0:47] the cursors, the anthropics, and the [0:49] open AIs. Although capital is scarce and [0:52] it's hard to get and all of these other [0:53] things once you get it. [0:56] >> First off, thanks for both of you making [0:58] making time [0:59] >> to be here. That's great. [1:00] >> Um Jared Ser tweeted a few days ago [1:03] something along lines of uh how he told [1:05] Claude to try to solve the reman [1:07] hypothesis and to try harder and uh I [1:10] don't know if there was actually any any [1:11] progress made but it's part of the [1:13] larger conversation around hey it seems [1:15] like there's some uh accomplishments [1:17] that that are that are being made. H how [1:19] do we make sense of this in terms of [1:21] what is actually happening and what does [1:23] it mean for for math? [1:24] >> I'm gonna let uh Steve go. [1:26] >> Oh, well, I'm no mathematician at all, [1:28] but I am I mean I think it's it's just [1:32] it's an important moment because it it [1:34] sort of divides the world into into two [1:37] groups like the groups that are just [1:38] very very excited that like oh my god [1:41] these these things are being solved. It [1:43] doesn't matter if you understand them. [1:44] actually nobody unders the number the [1:46] universe of people who understand what [1:47] these things are is very small and then [1:50] there are the people who are just like [1:51] oh it's fake it's going to put people [1:53] out of jobs that no one's going to know [1:55] the future of where these fields go and [1:57] the most interesting thing about it is [1:59] the group that's most excited are mostly [2:02] the mathematicians [2:04] and and they're the ones and so that [2:06] actually confuses everybody because if [2:08] you're of this school of the people who [2:10] are like it's going to put people out of [2:12] work and it's going to we're going to [2:13] all get dumber and it's, you know, the [2:14] dawn of idiocracy because computers are [2:16] doing all of our work. You're confused [2:18] that the people who are impacted most by [2:22] what this level of AI did are the most [2:25] excited. [2:25] >> Yeah. Yeah. [2:26] >> And I think that's just I I I think that [2:28] that is itself shining a light on this [2:30] moment that we're in right now. [2:32] >> You know, you're you're talking to two [2:34] systems guys, two product guys. You're [2:37] going to get like we're we're like to [2:38] have the same caveat. I feel there's [2:40] some things like we're actually both [2:41] very expert on. This is not one of them. [2:43] So I'm going to kind of from the peanut [2:45] gallery I I've got a two comments. So [2:47] one of them is like okay so I I view [2:49] like economic utility to be a very [2:51] important um uh measure when you're [2:55] talking about AI right so I was trying [2:57] to think like there's a lot of hours [2:59] been trying to solve some math thing [3:01] right but like if you sum up the entire [3:04] posttock salaries of all the people that [3:06] have been working over the years on [3:08] these problems is probably not very much [3:11] and so part of me is saying like it's [3:13] great that there's these capabilities [3:14] I'm not sure that the fact have been [3:17] longstanding as that much of an [3:19] indication because there hasn't been a [3:21] huge economic incentive um in order to [3:24] solve. Now that doesn't mean that it's [3:25] not hard or whatever. It's just like I [3:27] just don't think we have like that like [3:30] that that validation of this unlock some [3:32] deluge of like economic value. And the [3:36] second one is [3:38] it's kind of not surprising to me that [3:40] AI is very good at solving [3:43] a almost purely axiomatic [3:46] domain that you know requires knowing a [3:51] whole bunch of different things and and [3:52] and putting you know putting the [3:55] solutions together from very disperate [3:57] spaces because often really when I read [4:00] so I've I've been reading all these like [4:01] everybody else has been obsessively [4:02] they're like oh like I came up with the [4:04] solution it's like Yeah, the solution [4:05] was pretty straightforward. It just like [4:06] it borrowed from a bit of math that I [4:08] didn't know. And so I I think if there's [4:10] like a metalarning here, the metalarning [4:12] is is there is a set of problems that [4:14] probably, you know, require you being [4:17] too broad for most humans or most [4:18] education and it's going to solve those. [4:21] It's clearly very good at solving aatic [4:23] systems, but it's it I don't think it [4:24] provides a strong indication of is this [4:28] solving things that the market hasn't [4:29] been able to solve because there really [4:31] hasn't been a market around these. And [4:32] so I think that's the best questions for [4:33] us to answer. So very exciting. Seems [4:36] kind of reasonable and understandable. [4:38] Not sure what the longer term [4:39] implications are. [4:40] >> I I do think that there's something [4:42] interesting that that math is very much [4:45] a a leading edge indicator of what the [4:49] market might be interested in. Why? I [4:51] mean, if you like I remember when I was [4:52] in school, like there was some big thing [4:54] that someone at AT&T invented a new um [4:58] algorithmic, a new program for doing um [5:01] linear algebra, like a new way to solve [5:03] linear, which is super important right [5:04] now in the AI world. But his big thing [5:07] was, well, now we can just calculate [5:09] like the United Airlines flight map in [5:11] like 3 hours less time than we could [5:15] last week, [5:16] >> right? But but let's dig into So, it's [5:17] just not clear to me that the problems [5:19] being solved are those that are [5:20] roadblocks to like existingly [5:23] economically useful tasks, [5:24] >> right? And if they were, it's not clear [5:26] to me that they wouldn't have been [5:27] solved. Like posttock that's been [5:30] ruminating on a problem getting paid 30k [5:32] a year for 5 years. Like it's very [5:35] different than like the market has [5:37] decided that this is like the one thing [5:39] to unlock. And and maybe they're there. [5:41] maybe these problems that are being [5:42] solved are like the key problems to [5:44] unlocking some big economically [5:46] productive use case. I just haven't seen [5:48] that yet. So that for me is like the [5:50] next thing I'm kind of looking for. [5:51] >> I don't even know what 12dimensional [5:53] spaces or what that means. And so like I [5:55] I'm completely with you on like I don't [5:58] even know what problems are in 12 [6:00] dimensional space. Like are you very [6:01] skinny? Are you very tiny? I'm really [6:03] confused by that. [6:04] >> And and maybe I'm wrong here, but for me [6:05] it's still in the domain of like it's [6:07] really good at playing a game. like this [6:09] is the best Starcraft player ever, which [6:11] is cool and it's very powerful, but like [6:13] I have a hard time connecting that with [6:17] a like maybe the reason we didn't have [6:19] them before is because there just was an [6:21] economic need and b like how does that [6:23] actually map? And so listen, I there's a [6:25] huge range of these things. We get [6:26] pitches all the time. Some people will [6:27] walk in and say, you know, the the the [6:30] foundations to AGI and to reasoning is [6:33] going to be math. And once you do that, [6:34] you'll be able to answer every question [6:36] because the universe is based on some, [6:38] you know, fundamental mathematical [6:39] principles. And once you understand [6:40] that, you understand everything. And [6:42] then, you know, there's other people [6:43] candidly that walk in the door and [6:44] they're just like, listen, that's great. [6:46] Um, but like that doesn't tell you [6:47] anything about reality. And so, you [6:50] know, I think that there's more work to [6:51] do and this isn't just about like [6:53] getting better at math. [6:54] >> Yeah. I do think what's interesting is [6:55] that the part of the reason that the [6:57] mathematicians are very excited about it [6:59] though is because they they work a [7:01] certain way. Like if you work in in [7:03] history, there's basically no [7:06] abstraction in history. Like there's [7:08] just a bunch of facts and then people [7:10] develop like sort of these models that [7:12] you can think of almost as force [7:13] diagrams that explain war or famine or [7:16] or whatever. Whereas mathematicians and [7:18] mathematics has this super long historic [7:21] arc of layering on abstractions after [7:24] abstra and we'll get don't worry we'll [7:25] get to OSI in a minute. But but but but [7:28] like this idea that that where what why [7:31] they're so excited is like a bunch of [7:33] math all of a sudden becomes a new level [7:35] of of abstraction. [7:36] >> Is it true that they're so I I've found [7:38] that it's a mixed. I found that some are [7:40] very excited and some are in like an [7:42] existential crisis. The ones who are [7:45] excited they basically say listen it it [7:47] it solves 20% of my job is the 20% I [7:49] didn't like anyways. So like this allows [7:51] me to explore a new frontier that's very [7:53] important or whatever. And what I've [7:55] always wondered is like is that a [7:58] function of the type of problem being [8:00] solved? Like I just can't imagine if AI [8:02] came and solved cancer [8:04] like whatever someone that works on [8:06] cancer would be like oh I'm so [8:07] existentially depressed. [8:10] This is amazing. But let me I where on [8:13] the other hand of like oh we saw this [8:14] math problem oh I'm so depressed I saw [8:16] the math problem like maybe like [8:18] literally the entire utility of that [8:20] problem was keeping somebody employed to [8:22] solve the problem [8:23] >> or or just writing articles in the back [8:26] like another attempt at and here's where [8:28] I went wrong. So let me offer it this [8:30] way. [8:30] >> There's there's nothing on the other [8:31] side of the solution and so like we're [8:33] depressed because like now like whatever [8:36] like this useless activity is gone. Let [8:38] let let me let me stop. No, I mean too [8:42] cynical on this thing. I love math. [8:43] >> I think I think we caught you but like [8:46] being a little cynical but not really [8:47] but more that it's just it's it's [8:51] let me take a side of it this way [8:53] looking at the history of computer [8:54] science [8:55] >> because I had to take this class which I [8:57] looked at all the course cataloges for a [8:58] bunch of schools. You don't have to take [8:59] it anymore. That was like discrete math [9:01] bas. [9:02] >> Yeah. Yeah. and and like or and then or [9:04] algorithmic complexity theory which was [9:06] a required class for a very long time [9:08] and now [9:08] >> do you remember concrete mathematics [9:10] from Donald can like [9:12] >> I didn't I mean you're a Stanford guy [9:13] I'm not but like my state a school my [9:16] state a school we didn't have that but [9:18] uh that's a Cornell joke for us [9:19] Cornellians but but um you know my class [9:22] I got taught by one of the luminaries in [9:24] the field of algorithms ironically a [9:26] Stanford PhD uh um John Hopcraftoft of [9:29] course [9:29] >> who who invented for for the people who [9:31] are pragmatic IC invented like two three [9:33] trees and a bunch of stuff as his thesis [9:35] at Stanford. That's a legend. [9:36] >> But but John was our our professor in [9:39] all this crap and we had to learn like [9:41] all this P equals NP stuff and I [9:43] remember like the this is the four color [9:45] this is the four color [9:46] >> proven by computers. [9:47] >> No ex but that's where I'm going. You [9:49] just buried the lead. [9:51] >> Yeah, but but like so those of you that [9:53] don't know we had to take a whole course [9:54] in college on that basically boiled down [9:56] to this problem. And the interesting [9:57] thing is why and it was because the [10:01] theoreticians had postulated that if you [10:03] can solve this problem in in algorithmic [10:06] in in exponential in non-exponential [10:08] time in polomial time then you could [10:10] solve all these other problems like the [10:12] traveling salesperson problem and all [10:14] these other problems much much faster [10:17] which mattered because all of our [10:18] computers were just so computebound. So, [10:21] if you were the AT&T people that gave, [10:24] you know, like here's our node of like [10:25] 6,000 switches, like how do you route [10:28] optimally? You'd be like, well, we don't [10:30] have enough. That's like two years of [10:31] running the the simulation to solve [10:33] this. [10:34] >> Yeah. Yeah. [10:34] >> And and so it turns out that one of the [10:36] interesting things was they proved the [10:39] fourcolor theorem. [10:40] >> Yeah. [10:40] >> But they did it which by the way, I mean [10:43] just the four color theorem says for any [10:45] 2D planer map, you can color it. You can [10:48] use only four colors. So such that no [10:50] two adjacent areas have the same color. [10:52] Right. Exactly. [10:53] >> And um uh and you only need four colors. [10:57] You'll never need five colors. [10:58] >> And and we learned it just so people [11:00] like you kids know that's literally how [11:01] we learned it. And we could all repeat [11:03] it like that. It's this very weird [11:05] imprint over this problem. [11:07] >> And so what what sort of happened was no [11:09] one ever arrived at a at a basically [11:11] what you could think of as like a proof [11:13] that looked like calculus. Instead, what [11:15] they did is they actually proved that [11:17] the number of potential solutions was [11:19] finite. [11:20] >> Have you actually seen the proof? [11:21] >> Yeah. Yeah. 200 pages of combinations, [11:23] >> but they basically proved that you [11:25] there's a finite number of them and then [11:27] they just computed all of them and said, [11:29] "Look, it's only four colors." And so [11:31] it's this sort of bankshot proof, but it [11:34] was only possible because of compute. [11:35] >> And to your point, that was actually [11:37] very very useful in the the the the [11:39] practical applications. [11:40] >> Right. Right. and certainly as a [11:41] topology person [11:42] >> like setting setting strong bounds and [11:43] and things like that. So actually I see [11:45] >> and I think it I think that that to me [11:48] was just a really good lesson in in when [11:51] you have like a new level of abstraction [11:53] that says this is a whole class of [11:54] problems that can be solved. [11:56] >> Yeah. [11:56] >> You can then build tools working at that [11:58] level of abstraction and everybody [12:00] doesn't have to start from like okay [12:02] what's the two three tree representation [12:04] of what we're doing. [12:05] >> So listen I listen it's hard not to get [12:07] philosophical when you're talking about [12:08] AI. So, I'm gonna get philosophical and [12:10] you can tell me to shut up, but I I just [12:12] can't like you you kind of do. So, so [12:14] the the math this math thing seems to me [12:17] a little different because like it kind [12:18] of begs the following question, which is [12:21] will math ever be represented of [12:24] physical phenomenon, right? Like has [12:26] anybody ever like taken a bunch of [12:28] equations and actually predicted [12:29] something like physical? And I don't [12:32] know the answer to that. Like so I [12:33] worked in these large simulation codes [12:35] and these large simulation codes are [12:37] actually um trying to compute physical [12:40] phenomenon like the explosion of a star [12:42] or like you know what would happen to [12:44] like whatever an airplane in like an uh [12:47] an air simulator or a wind simulator. Um [12:51] but all of those and even though they're [12:52] just calculating these like large you [12:55] know differential equations they were [12:57] all based on empirical results. [12:59] >> Yeah. like literally the equations of [13:01] state for the [13:01] >> well they were model they were just they [13:03] were like we could measure temperature [13:04] in these places [13:05] >> that's exactly right so it was all it [13:06] was all based on empirical equations of [13:08] state and so I've always wondered like [13:11] >> like is simulation computationally [13:14] irreducible and so you actually have to [13:16] actually run the simulation in that case [13:18] it's not clear to me to what extent AI [13:21] helps like I know people are trying to [13:22] solve this problem with AI but like I [13:25] don't know if these math an these math [13:28] answers [13:30] have any impact on that type of stuff, [13:31] right? Right. So maybe there's some [13:33] separate algorithmics domain to to your [13:35] point where they do or you know maybe [13:37] like like modeling or logistics but when [13:39] it comes to like you know will you know [13:42] will this star explode will this [13:44] building stand up like the actual [13:46] simulation I think these things things [13:47] are pretty disjoint and and then I read [13:50] a lot of these discourses on the the [13:52] math solutions and there's kind of these [13:54] claims where if it can solve all math [13:56] you can predict anything and I just [13:57] think that that's a huge huge logical [13:59] leap which is not clear to me that is is [14:02] is is obviously true. [14:04] >> Yeah. [14:04] >> Or or there's any indication it's true [14:06] at all. like it. [14:07] >> So the way one way to that I think I [14:09] might [14:11] >> um talk about that you know again like [14:13] this is so out of my league on the [14:14] actual math [14:15] >> two and two systems people [14:17] >> I'm good but but I'm [14:18] >> compelled I'm I'm inherently a tool I'm [14:20] inherently a tools person and so I kind [14:23] of get this part of it which is [14:25] >> that what's happened is is that that AI [14:29] might not be the next tool to solve math [14:32] problems [14:33] >> at at some scale that matters but it [14:36] might lead to the development of a new [14:37] kind a new level of model and so I [14:39] brought like props to shows this off. [14:41] So, of course, this is the original [14:43] >> math tool. And so, before something like [14:45] this, this is a, you know, one of these [14:46] real ones from like Beijing market. [14:50] >> Well, I you know, it's the ones they [14:51] tell tourists in French. But, but I'm [14:54] very proud of that because I negotiated [14:55] it down to like seven cents. But, um, [14:58] >> but but you know, that became a level of [15:01] abstraction and all of a sudden like you [15:03] just had this basic math thing and then [15:05] you just fast forward a whole bunch. I [15:06] brought this because it's just so [15:07] freaking cool. It is. So this everybody [15:09] knows what slide rules are. You know, [15:11] nobody knows how to use them. Yeah. This [15:12] is called a a kerta and which is a [15:15] Austrian uh basically it's a round uh [15:18] slide rule. [15:19] >> Yeah. [15:19] >> And so it it's like a coffee grinder or [15:21] a pepper mill and you you you have the [15:23] all these ways you set the numbers on [15:25] the side and then you turn it one way to [15:27] add another way to subtract. [15:28] >> Whoa. [15:29] >> And it's this thing is it's [15:31] >> Wait, is that used for like multi-umber [15:33] arithmetic or is it used for stuff like [15:35] like logarithm? [15:36] >> No, it's only arithmetic. Okay. [15:38] >> Well, I think but of course it depends [15:40] on how you use it. But but um [15:43] >> it it it's from the the 20th mid 20th [15:47] century, I think. [15:48] >> And uh my uncle brought this back from [15:50] the war. And uh but it's what's [15:51] incredible is this is like 600 pieces of [15:53] machined metal. [15:54] >> Wow. [15:55] >> Inside this. It would cost like $50,000 [15:58] to make one. Now, [15:58] >> do you know how to use it? [15:59] >> I I actually did, but I'm not going to [16:01] try to do it. I I actually for prepping [16:03] for this so I wouldn't be a complete [16:04] Just go look what I have. I [16:06] actually went through the trouble of [16:08] learning how to use it, although it's [16:08] been sitting on my shelf for years. [16:10] >> But it's um but the interesting thing [16:12] is, you know, then all of a sudden a [16:13] whole new level of problems get solved. [16:15] And [16:16] >> wait, so you're saying that the the the [16:17] new model is the new calculator or the [16:19] new graphing calculator, the new I [16:21] actually remember when like remember the [16:23] TI85. [16:24] >> Oh, of course. Yeah. Yeah. [16:25] >> I remember that came out. They're like, [16:27] "All of the math teachers had this [16:28] crisis." I'm like, "You know, we used to [16:30] give you a piece of paper. We plot the [16:32] XY equation. Now they could do it on the [16:34] calculator and they can solve equations [16:36] and our field is dead." [16:37] >> But but but what's interesting is this [16:39] is why it's so important to AI today. [16:41] Those people didn't complain about when [16:43] calculus came out because calculus was a [16:45] baseline to them. And and what it is is [16:49] there's this notion this this people [16:50] react to change more than they react to [16:53] the baseline of where they all started. [16:56] And so so much of like the concerns in a [16:59] in a in a like I lived I literally got [17:01] like a uh the TI35 were the first [17:04] calculators in schools. The only [17:06] advanced math they did they had a [17:07] percent key and factorial which we [17:09] didn't even know what it was and you [17:10] could do like 59 factorial and that was [17:12] the max that you could display. And you [17:14] know I went to college and the classes [17:16] were no calculators allowed [17:18] >> the whole I was on that I my whole life [17:20] I've been on the cus of allowed and not [17:22] allowed for everybody. I mean, I was I [17:24] was there for the graphing calculator. [17:26] Like, you literally have a blue book. [17:27] >> Yeah. [17:28] >> Just to show all of your work, just to [17:30] show that you weren't plugging it into [17:31] the graph. [17:32] >> See, I missed the graph, which most of [17:33] us were like actually writing video [17:35] games in the back and could care less [17:36] about its ability to write math. But [17:38] >> Absolutely. Absolutely. And but but you [17:40] just play that backwards and you realize [17:42] that after you know after these guys you [17:45] went through this march of of algebra [17:47] and then linear algebra and then [17:49] calculus and you know all of and then [17:52] you know and then with calculus then you [17:53] ended up with fier transforms and fluid [17:55] dynamics and all of that was first to [17:57] your earlier point were all based on [17:59] need. I mean so much of this math [18:01] >> well all of computer computers are [18:02] basically from difference engines which [18:04] are just trying to calculate integrals. [18:05] But and and but and of course but to be [18:08] really clear to calculate integrals so [18:10] that we could shoot missiles and at and [18:12] cannons at each other [18:13] >> that okay so that's what [18:16] yes [18:16] >> which I'm not judging it I'm just saying [18:18] >> well well yes I don't mean to be [18:20] pedantic about this but like it's one of [18:21] my favorite parts of history. It [18:23] actually started with with tides which [18:24] also had massive economic value which [18:26] you're trying to calculate the tides and [18:28] this is where you kind of had like the [18:30] the old you know um and then that those [18:34] architectures got co-opted into of [18:37] course the ward effort for the [18:38] logorithms for that that's what I came [18:40] from it actually it was very interesting [18:42] uh was about 5,000 times faster than a [18:45] human being when it came to like you [18:48] know doing this and then of course [18:49] >> and it didn't make mistakes which was [18:51] sort of the [18:51] >> but it was but it was very specifically [18:53] math and very specific for economic [18:55] utility. And the interesting question to [18:56] me is is like these models are clearly [18:58] good at a type of math. It is is it one [19:01] that has somehow blocked some sort of [19:02] economic [19:03] >> Yeah. [19:04] >> And I don't know of the answer to that. [19:05] >> Oh yeah. I I you know, but I think it's [19:07] super interesting to keep going with [19:08] that because to me that it's so what's [19:10] so cool is that that that doing that [19:14] that basic calculus for the war and [19:16] making those missile tables and and [19:18] things like that, then it unlocked the [19:21] space race basically and jet engines and [19:24] factory automation and all of these [19:26] things. And you know, people were [19:27] cheering that on like that to me [19:30] culturally is the most interesting [19:31] thing. Like there was just this not only [19:34] were they cheering it on, they were [19:35] every parent was looking at their kids [19:36] saying, "Go learn that in school. Go win [19:39] the Westinghouse competition. Go win the [19:41] GE math competition." [19:42] >> And was that because of the Cold War? [19:44] Was it because [19:45] >> Well, obviously the Cold War was a big [19:47] cultural part of it for sure, but it was [19:49] just a general the the future. I like I [19:52] found this uh incredibly cool brochure [19:55] from uh IBM from it's from 1953. So 193. [20:00] >> Like do you just have this stuff in your [20:01] house? I just stumbled across it. Like [20:02] this this one I just got. I can't even [20:04] believe this exists. But this is like [20:07] this is a brochure about the future is [20:09] computing. [20:09] >> Wait, I want to see it. [20:11] >> But but but like first you got to look. [20:13] It's it's got like nuclear like the [20:15] whole thing. The future of computing is [20:17] like a guy with like Adams racing around [20:19] his head. Oh wait, we're zooming in and [20:21] doing the Carol Merrell thing. So but [20:23] the the fascinating thing is it's from [20:25] 1953. So you're Aniac and and that point [20:28] like that's it. That's the computer at [20:30] the time. This is pre74, pre 370. And so [20:34] it's a brochure from IBM explaining what [20:38] a computer might be. Not even is. And [20:40] it's like it took millions of years to [20:42] invent and recognize the usefulness of [20:44] the wheel. That's the opening sentence [20:46] of of of the and and but like you people [20:49] were eating the stuff up. But here's the [20:51] part that I want to get to. It talks [20:52] about computers and it's the two [20:54] families of computers. Yeah. And so of [20:56] course you get the slide rule and that's [20:58] explaining the history. And what this is [20:59] really leading up to is we could do this [21:02] for text too. [21:04] >> Yeah. [21:04] >> And so the idea and I mean like imagine [21:06] who was reading this in 1953 that it has [21:08] to explain hex and decimal and binary [21:11] and compare it to Roman. [21:12] >> That's amazing. [21:12] >> And because like nobody like nobody knew [21:15] >> is the name of that thing. [21:16] >> It's just called IBM light on the future [21:18] with like a rocket test tubeish [21:20] spotlight looking thing. And it's [21:22] incredible. There's like oscilloscope [21:24] waves in the back. It is the most [21:26] incredible thing. It has this dictionary [21:27] in the back. Imagine the first time [21:29] someone explains a computer and the [21:31] diction the dictionary isn't, you know, [21:33] arithmetic unit binary digit bit you [21:37] know like cathode ray tube electrostatic [21:40] storage tube and you know but but the [21:43] thing is is the reason I open this [21:44] because there's one cool page that [21:45] really matters. What is the organization [21:48] of digital computers? And so this is the [21:50] thing that that I gets to this point [21:52] about abstraction for us and AI. This is [21:55] these have been for for 75 years how we [21:59] thought computers are organized. [22:00] >> Yeah. [22:01] >> Input, storage, arithmetic, control, and [22:04] output. [22:04] >> Yeah. [22:05] >> And that's all that's what we learned in [22:07] school. You took courses basically in [22:09] each one of those. Last night we were [22:10] going back and forth on the abstractions [22:13] that will remain in computer science. [22:14] And you tossed in networking, which is [22:16] sort of control. [22:16] >> Yeah. Everybody forgets networking by [22:18] the way. Of course that was [22:19] >> well because most people stop worrying [22:22] about networking [22:24] >> stop worrying about as soon as the [22:25] packet leaves the computer I would say [22:28] you know the late 90s was the end of [22:30] basically a mandatory networking class [22:33] >> because because like but it was solved [22:35] like like there was no you know for me [22:38] it was it was the transistor I was like [22:40] the last time that computer science [22:42] majors had to know what a transistor was [22:44] and trust me I actually don't get what [22:46] one is now it's like a triangle symbol [22:48] But but the interesting thing is those [22:49] abstractions led to okay so now we have [22:52] those abstractions they were basically [22:54] fields that did each one of them like [22:55] you spent 20 years of your career on [22:57] storage and you watched the march from [23:00] from tubes to drums to spinning discs to [23:03] tapes and so on. And you know if you did [23:05] output you watch the invention of going [23:08] from a teletype to a lineoriented [23:10] teletype to a terminal black and white [23:13] to color to vector and the whole deal. [23:16] And all of those were the fields and [23:18] they all rose in parallel. Any CS [23:20] department which came out of the math [23:21] department because of the missiles [23:23] >> ended up being like departments made up [23:26] of those things and then it all [23:27] collapsed and produced us to the systems [23:30] group. [23:31] >> Sure. Yeah. Yeah. Yeah. So let let me [23:33] just push on one angle of this. Sure. [23:34] Sure. Because I I listen I I clearly [23:36] love the framing and we move up an [23:38] abstraction and every abstraction [23:39] there's still a set of problems. It's [23:40] just a higher level of abstraction. But [23:42] I still think this kind of notion of [23:43] economic meat is very important. [23:44] >> Oh yeah. Yeah. Right. So for example um [23:48] we we we you know Bletzley Park was [23:50] about cracking a code for a war and so [23:53] like there's this effort that created [23:56] innovation that the outcome was you know [23:58] winning World War II um any act we were [24:02] trying to do nuclear not just research [24:05] but like innovation you know in terms of [24:07] a war effort and so we needed to like [24:10] calculate integrals and we were doing it [24:12] by hand and so so at that point. These [24:16] things were lauded as like saving [24:18] humanity. Everybody was super excited. [24:20] All the physicists loved computers and [24:22] used computers. And for me, the thing [24:25] about the current solving math is I [24:28] don't know what that thing on the other [24:29] >> Oh, yeah. No, but on the other side is [24:32] we I I do think we've had that in the [24:33] past. And so, you know, [24:35] >> well, we did we had the Alph Go moment. [24:36] It was the same thing. Like we I we did [24:39] a podcast not in this room, but [24:41] >> but even before Alpha Go, we had like [24:42] remember when chess got bitten in chess? [24:44] Yes, we had the IBM chess thing when it [24:45] was and and Frank Chen and I we did this [24:47] podcast at Alpha Go and we had to try to [24:49] make people understand like why it was a [24:52] good idea and [24:52] >> and I think it's actually pretty [24:54] reasonable for us to ask the question [24:55] which is there's things that these [24:56] things solve and you know there's a lot [24:58] of utility and value and that and like [25:00] that's going to move things forward and [25:01] and when that tends to happen people [25:03] tend to be excited and get behind it and [25:05] there's these things you solve where I [25:07] think people [25:08] >> I think don't have like as positive a [25:11] view and I I would submit that's because [25:14] it's almost like solving the problem had [25:16] become the the end as opposed to the [25:19] actual end. But as we should maybe all [25:21] step back and be like if you're really [25:23] like sad about something being solved, [25:26] maybe it wasn't worth working on to [25:28] begin with, [25:28] >> right? And so this course, you know, and [25:31] you're like doing the San Mandala and [25:32] like your inner piece or something like [25:34] but that's not moving the economy [25:35] forward, [25:36] >> right? Well, we're both look, we're [25:37] systems people, but I'm actually an apps [25:38] person. I know you're not I don't [25:40] system. Yeah, we're both people but like [25:43] system [25:43] >> I of course absolutely think that that [25:46] the wave that matters are apps and of [25:48] course the internet [25:50] >> also this same problem happened in 1995 [25:52] and 96 with the internet which was it [25:56] was very exciting but most people just [25:58] sat around saying I don't know what that [26:00] does for me. Look, there's a great book [26:02] out now um called Steve Jobs in Exile, [26:04] which I absolutely think is required [26:06] reading if you're listening to this [26:08] podcast. Um so, uh Kane wrote the book, [26:11] but it's with Catm who at Pixar and with [26:13] Dan Leuen, who was at at Steve's super [26:16] good friend and was also at Microsoft. [26:18] They all they they this book is just [26:20] fantastic because it explains all of it [26:23] encapsulates all of this notion of like [26:26] building things that people actually [26:27] need and solve problems. But it pointed [26:29] out very clearly, you know, the the next [26:32] was actually the machine that Tim [26:33] Berners Lee used to write the HTTP [26:35] protocol, [26:36] >> right? So he actually [26:37] >> a next machine [26:38] >> and he used the next machine. That's an [26:40] interesting [26:40] >> and it's super interesting because [26:43] nobody knew what this machine was for or [26:45] what it did. But then he built that and [26:47] still nobody knew what the machine was [26:49] for or what it did because he's like [26:50] well it's to find the phone numbers and [26:52] other researchers and to share papers [26:54] and I'm like h like and and and I think [26:57] there was a great example of a company a [26:58] Seattle based company that was called uh [27:00] cyber pizza and this was like a dot [27:03] thing that didn't even make it to the [27:04] 2000 I think but the idea was it was [27:07] basically um uh uh in or Door Dash for [27:10] pizza only pizza and they would [27:12] basically you would order and then they [27:14] would figure out a pizza place near and [27:16] send the pizza. That was the launch demo [27:18] for the next on stage. They did that and [27:20] they had actually pizzas in the back in [27:22] case it didn't work. And and I should [27:25] say for next step or open step. But um [27:28] but the idea was that that that that was [27:32] showing what you could do with it. And [27:34] literally the reaction was like, "Wow, [27:36] that's really cool, but have you heard [27:38] of the telephone?" [27:38] >> Yeah. Yeah. Your point is not everything [27:40] we've known how to use. So I I'm a [27:42] little focused on like there was a [27:44] solution on the other side that people [27:45] are going for. You're making a point [27:47] that there's a lot of platforms that get [27:48] built where that's not clear, but [27:49] clearly they [27:50] >> well the the spreadsheet was like I I [27:52] will I will show here's my probably one [27:53] of my last visual aids for today. But [27:55] like the word processor came out and and [27:57] and this is in 1982 and people are using [28:00] it on Apple 2 computers and this new [28:02] kind of computer called CPM which is the [28:04] origin of DOSs and people were like I [28:07] don't understand why you just type and [28:09] the people once you used a computer the [28:11] idea of typing really really just didn't [28:14] work anymore and and so some people at [28:17] law school you have to show it now. [28:18] >> Yeah, I will. I'm just I like building [28:20] up. But these people in at Harvard Law [28:22] School, they brought in the first um [28:24] laptop. So that's the first laptop. [28:27] >> Weren't those called Lugables? [28:28] >> Well, no. They This was just called This [28:30] was literally just called an Osborne, [28:31] and it was the only one. So So as a [28:33] guess, how Eric, you're you're a kid. [28:36] How How much How was the battery life in [28:39] this? [28:40] >> Um [28:42] not long. [28:43] >> There was no battery. This giant case [28:46] that weighed 25 lbs, there's no battery [28:48] in it. It just plugged in. [28:50] >> But that was a trick question because [28:51] every time I've ever plugged mine out, I [28:53] have mine from college. Like people are [28:55] like, "Well, how long does the battery [28:56] last and and so it's literally the size [28:58] of a sewing machine. It's bigger than a [29:00] legal carry-on ever was." And that was [29:03] my college computer. But in in my senior [29:05] year of high school, it got banned from [29:07] Harvard Law School. So some someone [29:09] showed up to do their exams. So at [29:11] Harvard, they used to bring your [29:12] typewriter to exams because that way the [29:14] professor could read it. And it and two [29:16] kids brought brought computers in. One [29:18] brought an Apple 2 and one brought the [29:19] Osborne and then the school banned them. [29:21] >> Wow. [29:22] >> They just said this is and for all the [29:24] every reason you could read and I have [29:25] the Time magazine articles and the New [29:28] York Times. Every article you could read [29:30] reads like don't use the graphing [29:32] calculator, [29:33] >> don't listen to rap music or don't read [29:37] don't know jazz [29:38] >> or don't play Dungeons and Dragons [29:40] >> or the the arguments that are going on [29:42] now. Three years ago, I I tried to get [29:44] Cornell to use AI in in freshman writing [29:48] when the first and and they just stopped [29:51] talking to me. [29:52] >> Wow. [29:52] >> And but here's the irony of that. My [29:54] freshman year when I had this computer, [29:56] I was of course the only person in my 90 [29:58] person dorm with a computer and and I [30:01] had to get permission from the dean to [30:02] use it to write my papers for freshman [30:05] English. This is the fall of 1983. And [30:08] so and so that's exactly where we are [30:11] now on all of this stuff and this this [30:13] whole and it but the thing is it's a you [30:15] could also think of it as a level of [30:16] abstraction because like no one's going [30:18] to college now without a computer like [30:21] can I can I can I can I just put no so [30:24] so I I agree with you but let me [30:26] >> Right. Right. [30:26] >> Every once in a while I'm like well [30:28] maybe it's a little different. So here' [30:29] be the argument. Um [30:32] I don't think in the history of computer [30:33] science that I can recall have we ever [30:36] um [30:38] abdicated [30:39] actual reasoning or logic. It's always [30:42] been a resource, right? It's been like [30:44] comput network and storage and like [30:46] that's what you're providing and then [30:49] the human is like putting in the highle [30:52] thing and then it's using the compute [30:53] network and storage to to like calculate [30:56] the answer. But like all of the kind of [30:57] initial setup we're providing wherein [31:00] and I guess maybe it's not true for the [31:02] internet but now I feel like you're [31:04] actually abdicating thinking in a way [31:07] where you're like tell me the answer [31:09] like I'm not even really sure what the [31:11] question is. Right. And again, like I [31:12] think maybe you could say, well, Google [31:13] was kind of like that, too. But it was [31:14] still a very much a social thing and not [31:16] very much. And so it does feel like [31:17] that's a little different than just [31:18] going up in abstractions because going [31:21] up in abstractions, you still tend to [31:22] have like a deterministic system that's [31:24] a higher level of distraction that like [31:26] you have a computer and like it's the [31:27] human being that's kind of defining [31:29] everything about the problem statement. [31:30] It feels a little different. [31:31] >> Well, it definitely feels different. [31:34] Here's I I also think for me, graphing [31:37] calculators felt different. To me, [31:39] graphing calculators felt like cheating. [31:41] and and because you know the test [31:43] question was make a graph [31:46] >> and so that's what's going on right now [31:47] is that the capabilities match the test [31:50] question. Now getting us full circle to [31:52] what we were talking about about [31:54] computers and mathematicians. My [31:55] freshman year also a new product a new [31:57] thing came out and it was Maxima which [32:00] was the MIT symbolic math package and so [32:04] this was a way you could literally type [32:06] in like an integral [32:09] >> into a computer. I remember the first [32:10] time I saw Mathematica, I'm like this [32:12] stuff is black. [32:13] >> Maxima is you know uh machine aided what [32:17] was it? Machine [32:19] >> aided computation symbolic math I think [32:21] was the and that was the lab at MIT [32:23] started in the late 60s early 70s and [32:26] that had started to sweep through. So my [32:28] freshman engineering class we had a [32:30] version of it that ran on IBM PC. It it [32:33] was called MUMath and like you could [32:36] like we got our calculus homework you [32:38] marched over to the engineering library [32:40] checked out a PC disc and then just [32:41] typed in the answers to [32:44] >> so you don't think that so [32:45] >> and that was cheating let me just push [32:47] on just a little bit because I tend to [32:49] agree with every once in a while I have [32:50] like moments of doubt so so uh I don't [32:53] remember writing programs where you [32:55] actually abdicate logic like like if I'm [32:58] writing a program I'll like whatever [33:01] I'll use a cloud database [33:03] I'll use storage. I'll use networking. [33:06] You know, whatever it is, but like [33:08] correctness and logic for the program is [33:10] under the programmer's control. Maybe [33:12] I'll use a third party library. But [33:14] again, like I'm choosing the library. I [33:16] know the inputs. I know the outputs. And [33:18] I feel like we're entering this realm [33:19] where you're actually abdicating logic [33:21] to a third party. You're like, tell me [33:23] the answer. So maybe that's just a [33:26] higher level of traction. It feels a [33:27] little different to me. [33:28] >> No, like that's the debate. I'm like, [33:30] I'm all in on the debate. Like here's a [33:32] example of that a Stanford example. So [33:34] in the in the during the AI winter that [33:37] was the 80s [33:38] >> Stanford the biggest [33:39] >> one of the one of the AI winters [33:42] >> Stanford was and we have a podcast on [33:44] that from 15 years ago. One of the AI [33:47] one of the biggest things at Stanford [33:48] was to combine new AI with the medical [33:51] school. And so there were all of these [33:53] projects to do like medical diagnosis, [33:56] chemotherapy [33:57] kind of stuff. I worked on on one that [33:59] was doing organic synthesis with a team [34:01] at Harvard and all of those were sort of [34:04] the earliest like let me turn over the [34:07] decision-making. In fact, that's the [34:08] whole era of of um the 80s in computers [34:12] were the dawn of what they used to call [34:14] expert systems. [34:14] >> I remember [34:15] >> and so expert systems were the first [34:16] time we got a taste of this debate. [34:18] >> I I remember very well. I just [34:20] >> your classes were all this [34:21] >> it just didn't work. [34:23] >> Your classes were mostly about like you [34:26] had a bunch of classes on this stuff. [34:27] tons of expert systems. I've had to [34:29] build expert systems, right? I I've I've [34:32] written a lot of prologue. Exactly. So, [34:35] I very much understand it. I just [34:36] thought like that never really worked [34:38] and [34:39] >> Right. So, the big difference is that [34:40] stuff was working, but now [34:41] >> does work and we're abdicating logic [34:44] using these [34:45] >> but so it's interesting because to [34:47] compare and contrast [34:48] >> and even in the case of prologue, you're [34:49] kind of like for these to [34:50] >> you're coding it. It's algorithmic. [34:51] you're still providing the end state and [34:54] it's just like finding a way to get to [34:56] the end state where here like you're [34:57] almost asking it what the end state [34:58] should be. So it just feels a little [35:00] different [35:00] >> and it's especially so I agree like I I [35:03] love having this debate because I I [35:05] think [35:07] >> so much of it boils down to the concern [35:09] and the willies that you get thinking [35:10] about it. It's actually because of of [35:13] the context we're in [35:14] >> and and like because you know think [35:16] about like we we have all this stuff [35:17] going on where people don't want to [35:18] build data centers but like two years [35:20] ago people were like beating each other [35:22] to please governors were racing to have [35:24] data centers built or you know 10 years [35:26] ago like build a car factory in our [35:29] state the one that billows smoke and and [35:31] and is really hard labor and and so the [35:34] context really matters to to these [35:37] discussions. You can't separate them [35:39] from [35:39] >> right but I just want to go back this [35:41] layer and I don't mean to I just I just [35:43] think so um [35:45] >> your my entire career has been moving up [35:47] layers of stack but like there's always [35:49] a computer layer of stack. [35:50] >> Yeah. Yeah. [35:50] >> You could always map it down to like the [35:53] next layer in basically a deterministic [35:55] way. Higher levels of like compute [35:57] abstractions. [35:58] >> This is the first time it feels like a [36:00] different layer of the stack. like maybe [36:02] this is like really is the next [36:04] abstraction which is more of a human [36:06] level abstraction which doesn't map [36:08] directly and so is actually different. [36:09] So it may like I think you know whatever [36:12] it is starting with like [36:14] >> you know transistor logic and then going [36:17] to compute and then going to like [36:18] hardware and then going to oss and then [36:20] going to like applications and then [36:23] going to platforms like you've been [36:24] moving up the stack that way. It could [36:26] be the case that like we're at a layer [36:28] where like we have to rethink [36:29] fundamentals [36:30] >> because it feels a lot different to me [36:32] than just like this is the next layer. [36:33] >> The the big difference is and we can [36:35] argue you can argue this or debate it or [36:37] label it either side which is we [36:40] actually are making the leap from [36:42] calculating to imperative programming. [36:45] >> Yeah. which is where we've been and [36:46] which where everybody is that to now and [36:50] then we were in this for a brief time we [36:52] were in this mode where basically the [36:53] data really determined the program and [36:55] that was the first recognition all of [36:57] the inference and everything and now [36:58] we're at this where where it's it's [37:01] arbitrary it's random and it's [37:02] statistical [37:03] >> right so so the way that I think about [37:04] it is the following so imperative [37:07] programming you know all of the steps so [37:09] you write the recipe and it follows the [37:11] steps okay then there's declarative [37:13] programming declarative programming is [37:15] you know the end state [37:16] >> which is this prology kind of thing for [37:18] people [37:18] >> or or data log you know the or SQL you [37:21] know the end state [37:22] >> but then the computer does all the stuff [37:24] to get to that end state and you can't [37:25] really bound the computer time so you're [37:27] like this is like a make files it's like [37:28] here's what the end state looks like and [37:30] it does it [37:31] >> and this is like this new thing where [37:33] >> it's almost like um you don't really [37:36] know what the end state is specifically [37:39] and you just kind of like you know you [37:42] kind of like uh you know prey to the [37:44] model God in like the right words and [37:46] then it produces the answer that just [37:48] ends up being useful. [37:49] >> Yeah. Yeah. [37:50] >> And [37:50] >> but I I look and it is stoastic. [37:52] >> That's a factual statement like that. [37:54] But it's also interesting to think about [37:56] it going forward in terms of is that [37:59] itself the next layer of abstraction in [38:02] how we think of computing. [38:03] >> Yeah. And it may be like computing a [38:04] baby like this is where compute and like [38:06] or like natural like phenomenon actually [38:09] intersects pretty heavily because the [38:11] answer is produced from like human [38:13] output which is language which is kind [38:14] of different than like [38:16] >> to if we do need to rethink some [38:18] fundamental assumptions what may that [38:20] look like? [38:23] Well, I just think that like um people [38:26] like you know, Stephen and myself have [38:29] built these deep intuitions on how [38:32] systems function and how they hit the [38:35] industry based on 40 50 years of like [38:39] watching this stuff and I just don't [38:42] know like things like will value go to [38:44] the model or to the app? How much [38:47] capital can you apply to this stuff? [38:48] What classes of problems can you solve [38:50] versus not solve? um what guarantees [38:53] that can you provide uh how does this [38:56] impact productivity? There's a lot of [38:57] things that we've got intuitions on and [38:59] for me the big question is do we have to [39:01] like reshape those assumptions or not [39:03] and to what extent do we have to because [39:05] the laws of physics feel a little bit [39:06] different I'll just give you one example [39:08] I mean I've said this many times I think [39:10] it's so important 20 years ago and if [39:12] you're a startup of 10 people and I gave [39:14] you a billion dollars what would you do [39:15] with it [39:16] >> you would have end up spending a a ton [39:18] of money on building buying your own [39:19] computers and things if that's where [39:20] you're going if that's what hire people, [39:22] you buy computers, you blow up. Like you [39:23] wouldn't know what to do with a billion [39:24] dollar. [39:25] >> Oh. Oh, I see what you're saying. Yeah. [39:26] Yeah. [39:26] >> Yeah. 10 years ago. I give you a [39:27] billion. You hire you you hire engineers [39:29] and you'd be Right. Right. [39:31] Right. Like what do you do? Like right [39:33] code. You've got to you know you've got [39:35] >> you know the billion. The important part [39:37] of that is it's a billion. [39:39] >> It's not that you got money. It's that [39:40] it's a it's a huge billion. It's a ton [39:42] of money. [39:42] >> If I give you a billion dollars two [39:44] years ago [39:44] >> because 10 million you'd buy a bunch of [39:46] stuff from Hila Packard and the money [39:47] would be gone and for sure. For sure. [39:49] This one is a billion dollars. I mean [39:51] like in software you hire people and [39:53] then it's all about the FK scale. The [39:55] mythical man is very real. [39:57] >> Yep. [39:57] >> And right now if I give 20 people a [40:00] billion dollars they can actually use it [40:02] usefully. It's very so it's like it's [40:04] like we've kind of moved the industry [40:05] from like this engineering bound problem [40:07] to a capital problem that's [40:09] fundamentally very different. We've [40:10] never been like that before. And so like [40:12] this is like a law of physics where like [40:14] our early intuition which is like all [40:16] problems are engineering problems starts [40:18] to change. So I think there's this very [40:19] open question we should be especially [40:21] people like us should be asking which is [40:22] like to what extent do we have to [40:24] re-evaluate our priors on this stuff and [40:26] it's not just one level of abstraction [40:27] it actually changes the nature of [40:29] capital versus innovation versus [40:31] competition versus defensibility etc. [40:33] >> I that's a great a great way to to think [40:35] about it because it forces you to think [40:37] about a a new model. It's also [40:39] interesting that computing was capital [40:42] bound for the the the first no 30 or 40 [40:45] years. Like if you wanted to do [40:47] something [40:47] >> such an important point if you wanted to [40:49] do something with a computer like your [40:51] first step was we have to get one and [40:52] then you couldn't [40:53] >> you were capital bound and then you were [40:54] engineering bound and now we're capital [40:56] again which is crazy. So it's almost [40:58] like you have to like hop back 40 years. [41:00] >> Yeah. Madman goes through the scenario [41:02] where the computer shows up at the [41:03] advertising agency and and they run [41:06] around trying to figure out explain what [41:07] it does for people which they also got a [41:09] copy machine the same thing they did but [41:11] but it was interesting because they they [41:13] they couldn't figure out what to do but [41:15] they were excited that they had the [41:16] capital to acquire one and it made them [41:18] look like they knew what they were [41:19] doing. Five years ago, Patrick Olson [41:21] interviewed Sam Alman um in in a podcast [41:23] and Patrick was saying, "Hey, you know, [41:25] we've been in this era of lean startup, [41:27] but for your projects, you know, OpenAI, [41:29] um this sort of energy, you know, [41:31] project was involved with a few other [41:33] aging thing. You've raised colossal [41:35] amounts of money right out the gate, is [41:37] that underrated?" Um and and it's sort [41:39] of just speaking to what you're saying. [41:40] >> Yeah. Yeah. You know, it's interesting. [41:42] So, so prior to AI, there was always [41:44] this battle between Eric Rise and Ben [41:45] Horowits, right? Yeah. Yeah. See lean [41:47] startup and then and then you know Mark [41:48] and Ben wrote like the art of the fat [41:50] startup [41:51] >> which basically argued raise the money [41:53] and go for it. But there's always been [41:55] this natural limiter actually which is [41:58] engineering. [41:58] >> Yeah. [41:59] >> Complexity is that's actually been the [42:00] reality. And so Patrick Collison is [42:03] right is like we now have a discipline [42:06] for taking a lot of money with small [42:08] teams and using it productively. That's [42:10] a very very big change. I don't think [42:11] we've uh internalized it [42:13] >> which also it's incredibly [42:16] that is why there can be so much [42:19] optimism now because although capital is [42:22] scarce and it's hard to get and all of [42:24] these other things once you get it you [42:26] the as we know the building based on [42:28] people was also hard like just scaling [42:30] that and doing more and then nine people [42:32] can't do anything faster and [42:34] >> I'm telling you yeah my my my my [42:37] 10year job you know [42:38] >> was recruited and giving yeah literally [42:41] giving these early teams money and then [42:44] helping them recruit and then watching [42:46] then waiting for two years while the [42:47] engine and I think engineering just [42:49] doesn't scale. [42:50] >> It also has implications for venture [42:51] capital because for the last decade [42:52] people have been saying hey there's way [42:53] too much capital way too much capital. [42:55] >> I just think this is such a crazy view. [42:57] So there's been this view in venture [42:58] this zero sum thinking which is funny [42:59] from the people that shouldn't be zero [43:00] sum thinking you know like and they'll [43:02] go up on [43:03] >> too much capital is chasing too few [43:04] deals and all this is like this like [43:06] you're a venture capitalist don't you [43:08] believe in positive sum stuff right and [43:10] and but if you look at the numbers the [43:12] more capital that flows into private [43:14] markets the larger the market gets and [43:16] there's a there's a there's a there's a [43:17] you know there a couple reasons one of [43:18] them is the one we've talked about like [43:20] technical waves that actually are able [43:22] to consume capital like AI but there's [43:24] another one is if there's more capital [43:25] available on the private markets, [43:26] companies will stay private longer, so [43:28] more value acrru on the private side. [43:29] And so I think capital going to private [43:32] markets grows the TAM. It's not a [43:34] limited TAM. And I think the people that [43:35] should be, that's so funny, early stage [43:37] venture investors who should think of [43:39] like, you know, positive sum outcomes [43:41] need to stop thinking about zero sum. [43:42] Well, one way to think about that is is [43:44] I'll bring it back to what I I think the [43:47] your foundation enables what I think is [43:50] the most exciting thing, which is we're [43:51] really on the cusp of a wave of of apps [43:55] >> and and like the fact that now you can [43:57] apply capital without also being a [43:59] recruiter for 10 years and and have [44:01] output now all of the world that's [44:04] unserved by software, which is literally [44:06] all of it. Like everybody who complains [44:08] about whether it's medical records or [44:10] scheduling it to go to to go to a a a [44:14] doctor or my favorite are lawyers. Like [44:16] nobody has cheered more that oh my god [44:18] we're finally going to be able to [44:19] automate lawyers with AI which is the [44:21] weirdest thing in a world where [44:22] everybody is against everything except [44:24] having more lawyers and but but like all [44:27] of all of this this means that the [44:30] person who has the domain experience [44:32] like we used to love like venture [44:34] capital thing like oh you know it turns [44:36] out it's like really really hard to like [44:38] build commercial real estate. Wouldn't [44:40] it be great if somebody who understands [44:41] commercial real estate built a software [44:43] company but then they don't know how to [44:44] build software. Well, they should get a [44:46] co-founder who knows how to build [44:47] software and teach them about commercial [44:48] 20 years of commercial reality. It's [44:50] really hard. But now the path from that [44:53] kind of idea is a capital problem and [44:56] that's a new level of abstraction. And I [44:58] mean I remember my very very first [44:59] customer visit as a professional product [45:01] developer was to visit a doctor who [45:04] happened to have gone to medical school [45:05] after majoring in the earliest computer [45:07] science. Oh wow. [45:08] >> And he wrote like a DOSS program to [45:11] schedule a doctor's office. That's [45:12] amazing. [45:13] >> Which you think it's just scheduling. [45:14] It's a calendar with hours. But it turns [45:16] out this was me, 20-year-old me, hearing [45:19] this guy explain, "No, you don't [45:20] understand." You call the doctor and you [45:23] know, you're talking to a scheduler. So, [45:24] they're listening for keywords to [45:26] decide, is this 5 minutes, 20 minutes, [45:29] do they need the X-ray machines, do they [45:31] need the EKG? And so, they're actually [45:33] scheduling like a blood draw and all of [45:35] this stuff in parallel, not just the 10 [45:37] minutes you need with the doctor. And [45:39] so, that's what his software did. It [45:42] took him years to bang that out himself. [45:44] >> Yeah. [45:45] >> And and that's what you just [45:47] >> but that's the kind of thing that's [45:49] gonna be able to h like now that problem [45:52] can get solved by the person who knows [45:54] >> code no code is finally here. [45:55] >> Well, it could really be and and and it [45:58] and it might actually be that that [46:00] you're not just building this throwaway [46:02] code that's hard to you but also [46:03] everybody else's abstraction layer is [46:05] rising. So the you know you don't need [46:08] to design that piece of code like you [46:09] you know if you're doing it for a phone [46:11] well the phone's abstraction level has [46:12] risen so you're not building a text [46:14] control you're not building UI controls [46:16] whereas 20 years ago step one of [46:18] building a company was building all of [46:20] those things and so there's there's a [46:23] lot to how important this is in terms of [46:26] what you're able to do [46:28] >> I want to talk about any other [46:29] fundamental assumptions that might be [46:30] interesting to revisit is is it sort of [46:32] how about incumbents versus startups you [46:34] know the we've talked a lot about [46:35] innovators dilemma. Does it does that [46:37] you know now that these startups are [46:39] earth is incumbent or you know have the [46:41] capital advantage are they able to to do [46:43] more but at the same time we're seeing [46:44] startups that you would think incumbents [46:46] would just destroy. The crazy thing if [46:47] you would have told me six months ago [46:50] you would have asked this question say [46:51] like what like what advantages do [46:52] incumbents have? They have the same [46:53] advantage incumbents always have they [46:55] have the capital and they have the cash [46:57] flow and they have like whatever [46:59] >> distribution distribution and like [47:01] what's crazy is AI a solves the [47:04] distribution problem. it just solves the [47:06] demand problem and B these companies are [47:08] able to raise so much money that they're [47:10] actually on competitive footing with [47:12] like the Microsofts and the medicine and [47:14] the Microsofts and so I think we're in a [47:16] very new territory when it comes to [47:18] these new challenges versus the [47:20] incumbents specifically for these two [47:22] reasons you know I think that the the [47:25] the um the distribution point is is is [47:29] often misunderstood how impactful it is [47:31] in the past if you had a company and you [47:33] wanted to get people to use your stuff. [47:35] It was hard. You'd hire marketing. You [47:36] have no idea how much like to invest and [47:39] where and like you didn't know what [47:40] you're getting out of return on [47:41] investment. But the demand is so [47:43] unlimited for tokens and for GPUs. Like [47:45] literally, you can just decide how much [47:47] money you're putting into it in order to [47:48] drive top of funnel growth. And so the [47:51] things have typically been very very [47:52] hard for startups, you know, are much [47:54] easier now. And I think this is why [47:56] we're seeing um such meteoric growth of [47:59] the cursors, the anthropics and the open [48:01] AIs that results in capital access and [48:05] that has put them on on uneven footing. [48:08] So very [48:10] >> I and I think it's to your point about [48:12] how hard look my whole life was managing [48:14] thousands of people of engineers to [48:16] build things that couldn't be built [48:17] anywhere else. Like it was the moat to [48:20] build an operating system. It was [48:22] infinite and [48:23] >> and I think [48:24] >> you have to have one cutler. Is that the [48:26] >> Well, it's but it's it it really it read [48:29] read [48:30] >> that was mostly [48:30] >> Yeah, I know. No, I get it. No, but but [48:32] he's brilliant. [48:33] >> But read the um read the exile Steve [48:36] Jobs and exile book because it you can [48:39] you really get a sense for like building [48:40] up. In fact, you know, of course, next [48:42] was famously just it took the code from [48:44] mock at at Carnegie Melon and started [48:46] from there. We couldn't have done it [48:47] from from scratch completely. But this [48:49] this whole idea [48:52] of of just um how how important it is to [48:56] to think through the the domain specific [48:58] and the and how you disrupt people [49:00] because there was an old joke at Harvard [49:03] Business School when Clay was was still [49:04] with us which was they really it's weird [49:06] that they teach disruption as a theory [49:08] in the business school when really it [49:10] should just be a fact in the physics [49:12] department [49:13] >> and and I love that I was there in 98 [49:15] when he was writing the book and the [49:16] paper and everything. That's when I was [49:18] teaching. [49:18] >> That's nice. Great. and and and I really [49:21] I used to be of course there's a lore [49:23] with everybody who's from a big company [49:25] in Silicon Valley or when you arrive [49:26] like I did the theory is always like you [49:29] always think oh my god we're just going [49:30] to crush all of these little companies [49:32] you always think that when you're at the [49:33] big company and then you realize they [49:35] never get crushed [49:37] >> like and that Ben always makes this [49:38] point like they just and Mark does in [49:40] his um [49:42] >> his movie it was AWS actually put out of [49:44] business [49:44] >> right right exactly and because and the [49:47] and because you know the startups don't [49:48] aim aim straight at the incumbents and [49:51] the incumbents just don't pay attention. [49:52] They the incumbents are only interested [49:53] in what the other incumbents are doing. [49:55] Microsoft is worried way more about what [49:57] Amazon and Google are doing than anyone [50:00] in a startup space. But that that this [50:03] what the elements of disruption that [50:05] matter are the the cultural ones of [50:07] being a big company and those are [50:09] constant. Those are the laws of physics [50:11] and so you can't you just can't change [50:13] those. You can't change scorecards. who [50:15] can't change field sales and go to [50:18] market and compensation and org [50:20] structures and and legacy and customers [50:23] because you know the way you behave if [50:25] you have you know 500,000 customers [50:27] you're serving [50:28] >> you there's a bunch of stuff you just [50:30] can't do like that you just you're just [50:32] stuck [50:33] >> like you and and that is really the [50:36] essence of of disruption and that's why [50:38] we're at a magic moment where it's not [50:41] just that that the culture is there like [50:43] it always is but the startup ecosystem. [50:46] It is very it has it has it's a [50:48] reflection of what happened during cloud [50:50] which was a whole bunch of stuff that [50:52] you needed. Again, it's this abstraction [50:54] layer. You don't you know if you're a [50:56] startup like you were, you don't have to [50:58] go build a data center and build your [51:00] own egress and call AT&T and do all of [51:03] that stuff. You now you're up and [51:05] running in the first hours of your first [51:08] dinner. [51:08] >> But the but the thing with cloud I [51:10] actually think it's Yeah, you actually [51:12] articulated it very well. And I actually [51:14] use this because this is great. Like the [51:15] thing with cloud is like nobody thought [51:17] they could put AWS out of business. [51:18] >> Right. Right. [51:19] >> Like you just kind of accepted the [51:21] oligopy and you built on top of it. And [51:23] the question is is like will they kill [51:25] us in our little kind of pipsqueak [51:27] corner? The answer was [51:28] >> will they just add you for free or for [51:29] some price could match or whatever. [51:31] >> I actually think you know and that's [51:32] always been the question. Will Microsoft [51:34] the app? Right. But but now these [51:36] companies are actually taking on the [51:38] incumbents. And you make this great [51:40] point which I actually had thought about [51:41] it this way but like coffee has been [51:43] defined by these very complex large [51:45] engineering efforts building a chip [51:48] building a system. What's that the the [51:49] age of the new machine or [51:50] >> Yeah. Yeah. Yeah. The solar well [51:51] >> the solution [51:52] >> solar machine. Yeah. [51:53] >> Beautiful book right talked about how [51:54] hard it was to build like you know these [51:56] jack systems building an operating [51:58] system even in the cloud like I mean [52:00] like Jeff Dean in that era of people [52:03] were building these distributed clusters [52:04] and they're the first time that people [52:06] could figure out how to do that. Once [52:07] you had that was a massive advantage. So [52:09] these were these massive engineering [52:10] efforts that no startup could do and now [52:14] for these models it really is just [52:15] capital access and so it's a very [52:17] different laws of physics where like if [52:19] you can amass the capital you can do [52:20] something like I mean you know I mean [52:22] Google is Google they have all the data [52:24] they have all the intelligence and like [52:25] their models are getting trounced yeah [52:27] by open AI and by anthropic and it just [52:29] goes [52:29] >> because the cultural element [52:31] >> I I think that people people on the [52:33] outside underestimate it until you've [52:35] lived the cultural element of trying to [52:37] to do I'll bet it's I'll bet it's [52:39] cultural. It's not like a typical, you [52:41] know, engineering problem like they're [52:43] very good at because they've out [52:44] executed like like GCP is fantastic. [52:46] That's engineering after and then also I [52:49] bet it's probably hard to free up that [52:50] much capital for one of these companies [52:52] honestly. [52:52] >> Well, all the all the big all the big [52:54] companies you can tell from their [52:56] earnings calls how constrnated they've [52:59] been over the capital. You know, you [53:00] have Google doing their bond deal to [53:02] move it off balance sheet basically in [53:03] some weird way. You know, you had the [53:05] rumors of of I don't remember which [53:07] company you know, the rumors of like, [53:09] well, they're rationing the tokens so [53:11] that they they go to the enterprise [53:13] customers and not to the the internal [53:16] products and so the internal products [53:17] are are AI starved. [53:20] >> And of course, none of their competitors [53:21] to those products are are starved. And I [53:24] I [53:25] >> I just I I've learned to really [53:26] appreciate the the the I mean, look, I I [53:31] fought and fought and fought to not be [53:33] disrupted by by the mobile platforms [53:36] like by ARM basically. And and Intel [53:40] just didn't care. [53:41] >> Yeah. [53:42] >> You know, I I came down here, I sat [53:44] across the table from all the Intel [53:45] leadership and I pulled out the first [53:47] Surface and I said, "Here's our new [53:48] computer." And they got very excited and [53:50] then they were like, "But what's in [53:52] here?" I said, "Well, it's an ARM chip." [53:55] >> Oh, wow. [53:56] >> And and you know, like the fact I even [53:57] brought one into the building, you know, [54:00] and it was very very tough. and and and [54:04] they just never felt that that was going [54:06] to that that that was like a chip used [54:08] in a in a printer and also and they [54:10] looked at me like we're in we're we're [54:12] ARM lences we we knew all and I'm like [54:15] but it's the power it's the graphics [54:17] it's the you know always connected all [54:20] of this stuff and the culture was they [54:22] they do moors law at Intel and just like [54:27] with Google they do hypers scale [54:29] >> so like if if AI moves on device. [54:32] >> Yeah. Yeah. Yeah. Of course. [54:33] >> Like that's not what they do. [54:34] >> Yeah. Sure. [54:35] >> And and you know, and if it with with [54:38] Microsoft, they were squeezed. They're [54:40] squeezed now, you know. And I I think [54:42] you raised super interesting points [54:44] about the opportunity though for [54:46] >> for start [54:46] >> with this capital inversion kind of [54:48] thing. [54:48] >> Go raise go raise capital and go after [54:49] the [54:50] >> Well, and it's not just it's like you're [54:52] also saying like we're actually not [54:54] going to question you if you're trying [54:56] to raise that capital. like we're not [54:57] going to look at you like you're crazy [54:59] >> and you just look at the raises that are [55:00] happening right now and like you know [55:02] these companies have been quite [55:03] successful as a result. The um last [55:07] thing when Vichel came in uh and we had [55:09] him on the podcast, he was sort of um he [55:12] thought elements were a great [55:13] achievement, but he was bearish on their [55:14] ability to invent new discoveries, [55:18] particularly like scientific [55:19] breakthroughs or things like that. And [55:21] I'm curious if you think the the sort of [55:24] math progress um is consistent with that [55:27] or or or what is your latest thinking on [55:30] sort of the limitations of the of the [55:32] current sort of uh you know model [55:34] architecture versus like well we need [55:36] more um so here's my here's kind of my [55:39] new view which is I think we know [55:41] exactly how these things work. You put a [55:43] bunch of data in them. They're stuck to [55:44] that data. They can only do distribution [55:46] stuff and they can move along that [55:49] manifold. um in a perfectly Beijian way. [55:52] So we okay so we can say these words and [55:55] then then the question is is okay but [55:57] what are the implications of that like [55:58] what problems can it solve right? I [56:01] think it's just so hard for a human [56:03] being to reason to reason about a [56:05] digital artifact in this case the model [56:08] that was built with $5 billion. [56:12] So like in the history of humanity we've [56:14] never created a single digital artifact [56:15] that had that many flops and that much [56:17] data in it. So on one hand we know [56:19] exactly how it works from a mechanic [56:22] standpoint. On the other hand that is so [56:24] much data and that is so much compute [56:27] maybe all of that stuff's already in [56:29] there and it can solve anything that you [56:31] want and so you know the conversation [56:33] has moved from but how do these things [56:36] work? We know can it do out of [56:38] distribution stuff? No. Um uh does you [56:43] know is there transfer learning? [56:45] Probably not. like if IRL one thing it [56:47] doesn't tweet something else like are [56:48] these is the singularity here probably [56:50] not I think everybody kind of most many [56:52] people kind of agree on like we're not [56:54] in fast takeoff you know we're stuck to [56:57] it being in distribution we haven't [56:58] closed we all agree about that but what [57:02] I don't think anybody knows is okay but [57:04] you're still putting 10 billions of [57:05] dollars in that thing what's it capable [57:07] of now and if you if you consider this [57:09] metaeconomic machinery which means the [57:12] ability from anthropic to raise lots of [57:13] money then then pour all of that money [57:15] into this thing to create this super [57:17] powerful thing. I don't think any of us [57:19] can predict what that means and where [57:20] that goes. And so it's a different [57:22] conversation, but the question is the [57:24] same. It's like will that be able to [57:25] cure cancer? Maybe. But if you put $20 [57:26] billion into something, maybe it can [57:28] cure cancer effectively. And that's [57:30] where I think the discourse has evolved [57:31] and where it is now. And I I honestly [57:33] have decided that I cannot predict what [57:36] an artifact worth that was, you know, [57:38] that like you used $20 billion to create [57:40] is capable of. I look, I I think it's [57:43] just so important. [57:45] It's important for people who are deep [57:47] in watching everything that's new to [57:49] admit that they can't predict. And I [57:51] think that that's great because it turns [57:52] out like I wrote 58 memos on what the [57:55] internet was going to be. And I was [57:57] wrong a lot of them by far. But I I do [58:02] think on on the and and I [58:03] >> but but even this one is a little [58:04] different. This is like I I I take $20 [58:07] billion and I put it into a model, [58:08] right? and and then you and I look at [58:10] that model and we can do whatever we [58:12] want. I don't think we can comprehend [58:13] the like what that even means. It's so [58:16] many flops and so much data like I don't [58:18] know what that's capable of. [58:19] >> I I and I think we are I think that [58:21] that's really true and I and I think but [58:23] I will say on on biio medicine in [58:26] particular look the other half of my [58:27] household is a research doctor who uses [58:30] AI. We have a spark at home and she's [58:32] loaded [58:34] >> like a a ton [58:35] >> like a sun spark. No, no. Uh uh Nvidia [58:39] Spark. Oh yeah. No, no, no. Not not with [58:41] a C or with a K. Oh yeah. Wow. We were [58:44] in old times there for I was like not a [58:46] Scott McNeely Spark. No. No. Um and like [58:48] a relic [58:49] >> and No. And and um and like it's all AI [58:53] like she does brain stuff and and [58:56] surgical brain stuff. All AI. And it's [58:59] so interesting to see because what it [59:01] what it really can do is it it just it [59:05] it sees the patterns that you can't that [59:08] only experience could tell somebody. But [59:10] if there's 10,000 papers on a topic [59:13] that's part of her model, then like it's [59:16] just finding the patterns that you just [59:18] that no one has. And that's a pretty [59:20] basic AI capability at this point, but [59:23] it's actually opening up solutions or [59:26] problems or research directions and [59:28] things like that. I will say just for [59:30] the like this is not a magic to discover [59:32] drugs because the hard part of drugs has [59:35] always been candidates not candidate. [59:38] It's always been efficacy and safety. [59:40] The candidates have since the 80s have [59:42] been able to develop more than we could [59:44] test. It's human patients and it's very [59:46] very very hard. Can I can I just tell [59:49] you something that I got wrong on this? [59:50] So um I I love the question that you [59:52] asked which is how's our thinking [59:53] evolved on like you know whether these [59:55] things you know like their capabilities [59:57] in generality which is um I was [60:00] responding to this boastrum notion of [60:03] recursive self-improvement fast takeoff [60:06] you create one of these things you step [60:08] back and it takes over the world right [60:09] and so I kind of poo pooed that because [60:11] that's clearly not what's happening and [60:12] I think most a lot of people agree that [60:14] that's the case right but here's what I [60:16] got wrong what I got wrong is I did not [60:19] know that But we could effectively just [60:22] continue to pour money in this like the [60:24] scaling laws are holding and I I don't [60:27] you know I don't know what it means to [60:29] just let's say we do a hundred billion [60:30] dollar training run to like have this [60:32] thing that you're putting a hundred [60:33] billion dollars in and then that that [60:36] money comes from this meta economic [60:38] machinery that may be able want to solve [60:40] whatever they may want to solve cancer [60:42] but they may also want to create a [60:43] weapon like who knows and so this [60:47] concentration of this many resources in [60:49] a useful way I think is very new. I [60:52] don't think we understand the [60:52] implications. I think you could [60:54] reasonably argue that that's very [60:55] dangerous if you kind of apply that $100 [60:57] billion in the wrong way. So I think [60:59] that's kind of where this conversation [61:00] needs to evolve [61:02] to. So less the [61:05] fume, you know, and more the what does [61:09] it mean to be able to concentrate [61:10] resources, [61:11] >> which also was in all I mean this is [61:14] you're you're basically talking about [61:15] exponential growth and and this is just [61:18] exponential in dollars and and we all [61:21] know none of us can model exponential [61:23] very well. [61:24] >> Yeah, we we've never been able to do [61:25] that like like complex engineering [61:27] project. You were not like tackling one [61:29] problem with a lot of money or just kind [61:31] of building this machine. Well, we we [61:33] you know it's you're right. You're 100% [61:35] right and I completely agree, but just I [61:38] I remember just sitting in meeting after [61:40] meeting Intel saying we have 5 [61:42] gigahertz, we have this many gigahertz, [61:44] this many transistors and literally [61:46] nobody knows what we're going to do with [61:47] them all. [61:48] >> No, no, you're building the machinery. [61:50] I'm saying in this case, if you're like, [61:51] I want to exhaustively explore every [61:53] protein combination, right? We can just [61:55] turn that into a money problem. Yes. Got [61:57] a very strange, [61:58] >> which is a great way to say it that we [61:59] can take previously [62:02] um infinite problems and apply capital [62:05] and it becomes finite. [62:06] >> Just makes it a capital problem and not [62:08] an engineering problem. Yeah. Yeah. [62:09] Which is just a very different laws of [62:10] physics. [62:10] >> Yeah. Yeah. Let's wrap on that cuz it's [62:11] 2:30. This has been a great episode. [62:13] Thank you guys for [62:15] >> Thank you.