Economists Challenge Post-AGI Labor Collapse Scenarios

经济学家质疑AGI时代的劳动力崩溃论 快速自动化不会引发系统性经济崩溃，因为技术能力的拓展以及向关系型商品的转变将持续支撑资本再投资与人类需求。 在本期节目中，Dwarkesh Patel与Alex Imas及Phil Trammell进行对谈，通过严谨的宏观经济学框架探讨AGI经济学。他们争论了随着机器经济实现整个供应链的自动化，人类劳动力是否会走向崩溃，抑或基于人类偏好的“关系型部门”能否维系工作价值。讨论还涵盖了税收模型（NIT、UBI、全民基本资本UBC）、Citrini式需求崩溃的不可能性、资本非饱和积累机制，以及发展中国家在面对AI集中时的战略抉择。 历史上长期稳定在约 60% 的劳动收入份额可能会被打破；若 AI 沿完整供应链打造出网络调整后资本份额达 100% 的全自动商品，人类劳动面临从各层级生产中被系统性剥离的风险。 (5:00) 人类“关系型部门”的商品与服务依赖真人的在场价值，但其经济地位能否维系，取决于消费者偏好是否足以抵抗技术变革带来的商品多样性爆发对传统人际支出份额的稀释。 (15:00) 低增长的“混乱过渡期”可能带来严重的劳动力分流与政治摩擦；在财富再分配方案中，负所得税（NIT）与全民基本资本（UBC）提供了相比传统 UBI 更具制度韧性的替代路径。 (19:36) 宏观数据显示当前白领及软件工程就业未出现系统性崩溃；基于 O 型环假说，部分工作任务的自动化反而可能显著推高剩余未自动化人类任务的边际经济回报。 (30:00) 针对 AI 进展将引发需求崩溃型衰退的观点，学者反驳指出，技术边界扩张导致实际 GDP 负增长在理论上极难成立，除非资本所有者对个人消费与硬件再投资存在极其不切实际的硬性上限。 (35:00)

Economists Alex Imas and Phil Trammell analyze how artificial general intelligence could alter economic scarcity, arguing that while intrinsic human preferences and institutional frictions may preserve a relational sector, long-term economic distribution hinges on capital accumulation dynamics and whether ordinary citizens can own shares of frontier technology. The Relational Sector and Intrinsic Scarcity (0:00) Dwarkesh Patel frames the relational sector as goods and services where human involvement is an intrinsic component of value, raising the question of whether this sector will remain economically significant if a machine economy can execute any task autonomously. Imas cautions against single-point labor forecasts, citing research by Andre Fradkin, Brian DeBerry, and Andrew Koh on widespread economist disagreement, but points to experimental evidence showing intrinsic human preference for human origin: in an art-bidding experiment, participants paid significantly more for a unique human-made print than an AI print, though this premium vanished once the art was mass-produced in 500 copies. Unlike horses, which were replaced because they were purely functional inputs, human labor can retain value only if outputs lose intrinsic worth when humans are removed. Trammell counters with the "Mongolian economist fallacy" from circa 1400: analysts assuming static categories might have predicted that cheap yurts and horse transport would shift all spending to live human singers, ignoring the exponential expansion of new product varieties. Similarly, Patel notes that while physical transistors expanded by trillions via Moore's Law, Chad Jones's research shows total GDP expenditure on raw transistors declined as price drops outpaced quantity, demonstrating that sheer quantitative abundance does not guarantee dominance in value share. Labor Share, Task Automation, and Structural Frictions (5:00) Historical economic data reflects a remarkably stable labor share of around 60% or more of national income (a Kaldor fact supported by Tony Atkinson's accounting analysis), while capital captures 30% to 40%. Trammell argues this historical balance occurred because capital and labor were complements, warning that AI could drive the "network-adjusted capital share" of goods to 100% by eliminating upstream human inputs throughout entire supply chains. Imas notes that macro data from the Yale Budget Lab currently reveals no aggregate white-collar collapse; software developer demand has seen senior hiring rise even as junior growth slowed slightly below trend. Under an O-ring model of jobs, automating nine out of ten tasks can cause the value of the remaining complementary human task to skyrocket, potentially expanding net employment if product demand is highly elastic. However, Patel notes that Jevons Paradox depends on elasticity: while coal use expanded through efficiency, agriculture contracted its labor force due to biological limits on food intake. Additionally, structural frictions currently shield human roles: corporate accounting, law, and high-stakes engineering require legal liability, regulatory accountability, and risk mitigation. Trammell observes that as machine-native architectures mature, human latency, higher error rates, and communication friction will eventually disincentivize hybrid workflows. Investment-Specific Technical Change and Non-Satiating Capital (50:00) Addressing viral macroeconomic doom scenarios predicting demand collapses, Imas and Trammell argue that technological abundance expanding the supply frontier makes negative real GDP growth theoretically improbable unless capital owners hit an absolute hard bound on consumption and refuse to reinvest profits. Trammell highlights that under investment-specific technical change, physical robot stocks might explode while their relative price against consumption goods collapses; in robot units, returns might appear enormous, yet real consumption-based interest rates normalize. Historically, concentrated fortunes dissipated due to mortality, philanthropy, estate taxes, or diluted inheritances. Patel and Trammell suggest these dissipation shocks could fail post-AGI if wealth is held by indefinite trusts, self-replicating algorithmic entities, or actors driven by unbounded non-financial utility functions, such as utilitarian projects funding deep-space compute infrastructure. Entities that systematically reinvest returns into physical hardware and compute will economically outcompete entities that divert revenue toward human luxury consumption. Redistribution Models and the Global Indexing Dilemma (19:36) In a potential "messy middle" where white-collar jobs are displaced faster than aggregate wealth comfortably compensates them, Imas warns that political economy frictions make targeted relief contentious, noting that a 2% uptick in unemployment historically triggers political crises. Evaluating redistribution mechanisms, Imas favors immediate safety nets like a Negative Income Tax over Universal Basic Income, which risks creating fragile dependencies on political regimes. Universal Basic Capital provides citizens equity ownership but faces severe indexing and portfolio-targeting problems if governments select failing companies. This indexing challenge extends globally. Imas and Trammell warn that developing nations lacking domestic AI supply chains (unlike Taiwan, South Korea, or the Netherlands) risk being bypassed if automated production reshores manufacturing to advanced economies. Because median household wealth in developed economies is largely concentrated in residential real estate—which does not complement AI production—broad equity indexing represents the primary vehicle to capture AI value. If AGI commoditizes widely like electricity, downstream gains will accrue across broad public index funds; conversely, if frontier labs retain value privately like social media platforms, ordinary citizens and non-AI developing nations will struggle to share in the economic surplus. The ultimate economic trajectory of AGI remains bounded by unresolved empirical questions: whether consumer demand for human-origin services remains broadly elastic, whether capital holders will exhibit satiation, and whether frontier AI capabilities diffuse as broad public commodities or remain concentrated in private, rent-extracting monopolies.

经济学家 Alex Imas 与 Phil Trammell 在与 Dwarkesh Patel 的对谈中指出，即便通用人工智能（AGI）推动全面自动化，人类对人际连接的内在偏好与商品种类的指数级扩张仍将重塑稀缺性，而不会简单导致全盘经济崩溃或财富绝对集中。他们认为，未来的核心挑战在于应对技术转型中的政治经济摩擦、设计切实的再分配与资本指数化机制，并防止不具备硬件与模型供应链的发展中国家陷入结构性脱节。 关系型经济与商品多样性拓展 (0:00) Dwarkesh Patel 与 Alex Imas 探讨了“关系型部门”（relational sector）的独特性，即由人类亲自参与构成核心价值的领域。Imas 的实验表明，消费者对独一无二的人工艺术品支付意愿显著高于 AI 艺术品，但当版数增至 500 件时人工溢价骤降，证实人际连接价值高度依赖独特性而非单纯的功能输出。然而，Phil Trammell 提出了“1400年代蒙古经济学谬误”：若当时的学者认为马匹和蒙古包充裕后所有支出都会流向真人歌手，就忽略了资本积累带来的全新商品品类（如现代医疗、电子娱乐）的爆炸式增长。因此，即使人类对同类的偏好能够延续，关系型部门在整体经济产出中的最终占比仍可能被海量新型自动化商品所稀释。 宏观要素分配与资本相对价格的变动 (5:00) 历史上约 60% 的国民收入归于劳动份额（卡尔多事实），但 Trammell 指出，若 AI 普及使全产业链的“网络调整后资本份额”逼近 100%，劳动力与资本的传统互补性将被彻底打破。针对部分宏观观点（如 Citrini 的假说）所预言的“白领失业引发总需求崩溃与经济衰退”，Imas 与 Trammell 辩护称，在技术供给边界大幅扩张的环境下，产出收缩在理论上要求资本所有者对自身消费设定严格上限且停止再投资，这在经验和理论上都极不成立。此外，Trammell 强调了“特定投资技术变革”的关键作用：即便物理形态的机器人存量爆炸式增长，资本品相对于人类消费品的价格也将快速暴跌，从而使以实际消费计价的资本回报率回归常态，避免资本份额无节制膨胀。 “混乱中期”的转型摩擦与再分配设计 (19:36) 对谈分析了 Molly White 提出的“混乱中期”（messy middle）情景——即 AI 迅速替代白领工作，但效率提升尚未爆发至足以平滑补偿被替代者的阶段。Imas 结合 1920 至 1940 年电话接线员历时 20 年逐步被替代的历史研究指出，渐进式的“滴水式”自动化会导致劳动力长期在低薪部门未充分就业，且由于缺乏类似突发危机（如疫情）的紧迫性，难以凝聚政治共识以触发即时的国家财政干预。在再分配方案层面，Imas 评价负所得税（NIT）能够提供即时生效的收入底线保障；普遍基本收入（UBI）则可能建立起公民生存完全依赖政府支票的脆弱政治依附关系；而普遍基本资本（UBC）虽然赋予公民股权，却面临严重的指数化配置与选股偏差难题，这促使学者们讨论通过消费税或土地税筹集资金并配置广泛市场指数基金的可行性。 全球经济脱节、技术类比与指数化困境 (60:26) 针对缺乏先进制程（如台积电、ASML）与前沿模型训练基建的发展中国家（如印度、尼日利亚），Imas 与 Trammell 警告其面临比成熟经济体更为漫长和严峻的“混乱中期”。若发达经济体依托 AI 与机器人将基础商品生产完全本土化，发展中国家可能丧失传统的人口红利与劳动力出口优势。Patel 与专家们对比了“电力模式”（绝大部分技术盈余外溢至下游各行各业）与“社交媒体模式”（平台内部高度榨取经济租金）。如果开源模型与闭源前沿保持 6 至 9 个月的较短差距，或者领军实验室因资本需求走向公开市场，AI 将呈现类似电力的通用赋能属性，允许全球投资者通过宽基指数分享收益；反之，若前沿技术长期被极少数未上市实验室封闭垄断，不仅全球公众难以分享增长红利，还会增加国家机器强行实施集中管控的政治风险。 本场讨论的核心结论在于，AGI 带来的深层挑战并非纯技术性的供给崩溃，而是要素分配剧烈失衡下的制度调适、政治经济摩擦与全球资本鸿沟。人类对真实关系的心理偏好、开源与闭源模型的技术代差，以及前沿资本能否被公众和欠发达国家有效指数化持有，将共同决定未来社会在富足与极化之间的实际走向。

## Background 本期播客由主持人 Dwarkesh Patel 于 2026 年 6 月 4 日发布于其同名播客（Dwarkesh Podcast）平台，邀请了经济学者 Alex Imas 与 Phil Trammell 共同探讨。节目简介指出，许多关于 AI 带来的财富剧增、税收再分配、发展中国家定位以及贫富差距的核心难题唯有经济学能够回答，但人们在这些议题上的日常直觉往往是完全错误的。 ## TL;DR 对话从宏观经济学与行为经济学视角深入剖析了 AGI 时代的稀缺性转移：当机器实现全产业链闭环生产时，劳动收入份额将取决于资本品类扩张速度与人类对“关系型部门”的内生偏好博弈。学者们指出了“基于任务模型”与“O型环自动化”对人类劳动力的潜在排挤机制，同时严谨反驳了所谓“AI 导致总需求崩溃引发大萧条”的流行谬论。此外，讨论深入权衡了负所得税、全民基本资本（UBC）等财富再分配工具的利弊，并建议处于全球 AI 供应链外围的发展中国家优先通过主权资产对全球 AI 资本进行指数化投资，以此捕获技术红利。 ## Atmosphere & style 整场对话洋溢着高密度、严密推演的硬核学术氛围，彻底跳脱了非黑即白的科幻狂热与末日恐慌。嘉宾与主持人之间展开了高水平的智性交锋，借助历史案例、思想实验与精妙的数学博弈层层剥开反直觉的经济逻辑。整期节目的情绪曲线冷静克制且发人深思，让听众如同亲历一场智识激荡的经济学头脑风暴，在抽丝剥茧中重构对未来资本与劳动格局的认知底色。 ## Bulletin - `30:00` 最新宏观数据并未印证白领失业潮，初级软件工程师招聘放缓仅是增速稍落后于历史趋势，而高级工程师需求持续上升；短期内所谓的大裁员更多属于迎合市场叙事的非理性跟风。 - `38:00` 由于现代生产链条极端依赖可靠性（O型环理论），只要 AI 在关键链条上存在致命短板，人类劳动就依然作为关键安全垫存在；但一旦机器全流程超越人类，人类介入反而可能因拖垮系统而被彻底剔除。 - `25:00` 针对 AI 财富再分配，直接征收财富税容易引发政治滑坡与投资抑制，更优策略是通过消费税或土地税筹措财源，进而建立广泛的科技资产指数基金包发放给全体国民。 - `01:07:30` 游离在 AI 供应链之外的发展中国家面临严峻被边缘化风险，由于美国非微型企业中未上市公司市值占比仍远低于 20%，穷国和大众最硬核的避险途径是尽早设法买入并持有绑定全球 AI 增长的指数化资产。 ## Key takeaways - `00:00` 李嘉图式的机械化悲观预测忽略了结构性转型与劳动力总量固定谬论，但 AGI 时代资金能否持续溢出到新型人际服务仍取决于人类是否具备对人际交互本身的刚性偏好。 - `05:00` 历史上宏观劳动份额稳定在 60% 左右的卡尔多事实，主要源于过去哪怕全自动组装线上游仍依赖人类劳动注入附加值；未来全链条资本份额达 1.0 的商品将成为打破历史常态的奇点。 - `10:00` 根据基于任务的模型，大部分工作若被拆解且全流程自动化，劳动份额能否留存取决于消费者是否愿意为‘人类在环’支付溢价，以及 AI 创造新资本品类的速度是否会压制人类关系型消费。 - `12:35` 摩尔定律的悲观本质在于计算边际价值每 18 个月减半，但近年来 GPU 算力租赁价格不降反升，反映出模型能力跃升开辟了海量全新用途，将边际需求饱和点大幅后移。 - `15:42` 艺术品竞拍实验证实人类对真人原作具备天然溢价，但一旦被多倍复制，这种情感联结便会迅速稀释为普通大宗商品，决定了关系型部门的适用广度存在天然上限。 - `20:00` 在劳动力替代过程中，最危险的并非断崖式崩溃（往往能迅速触发国家级财政救助），而是类似历史接线员替代的‘慢刀子’过程，容易导致大规模隐性不充分就业且再分配意愿迟钝。 - `25:00` 全民基本收入（UBI）存在使普通民众生存完全依附于民选公权力的权力结构风险，相比之下，全民基本资本（UBC）与主权财富基金指数化持股更具激励兼容性。 - `32:48` 杰文斯悖论在软件工程中得以维系的核心在于其极高的需求价格弹性；若换作需求缺乏弹性的行业（如农业或短期原油），自动化提升只会加速整体雇佣人数的收缩。 - `35:00` 总需求崩溃引发大萧条的论调在逻辑上存在硬伤，因为技术奇点前夕的资本家绝不会将天量财富闲置，而是会将每一分利润再投入到数据中心、芯片厂等生产性资本建设中。 - `54:35` 当具备无限资本积累偏好的代理人（或 AI 实体）通过寿命延长或纯粹信托阻断历史上的‘耗散冲击’后，复合增长的数学规律将使这类实体最终主导全社会的生产与宏观定价。 - `57:02` 古典功利主义的天文级效用追求（如戴森球建设）与地缘意识形态竞争构成了财富无限积累的工具性动机，传统以人类最终消费为锚点的 GDP 核算体系在跨星系机器网络面前将面临解构。 - `01:07:30` AI 若演变为类似电力的通用基础设施，其超额收益将广泛外溢至整个实体经济并让广泛指数基金受益；若走向类似社交媒体的高垄断租金，则需依托开源模型与资本公开化来对冲寡头掠夺。 - `01:10:00` 欠发达国家虽面临基础教育短板，但在颠覆性技术浪潮下有望依托 AI 基础设施重现类似‘移动支付跨越式发展’的跳跃式生产力红利。 - `01:13:20` 模型完全大宗商品化可能引发缺乏安全对齐缓冲的恶性军备竞赛，而股权高度分散、拉开代差的上市公司形态既能保留安全冗余，又能避免技术被单一强权中心化捕获。 ## Detailed walkthrough ### `00:00` 关系型经济、机器经济与人类在 AGI 时代的稀缺性 主持人 Dwarkesh Patel 首先引出本期对话的核心议题：随着人工智能与自动化的深度演进，经济学究竟能为我们预测未来提供哪些参考？在工资走势、劳动收入份额（labor share）、AGI 巨额财富的最优税收与二次分配方式，以及**稀缺性（scarcity）**的转移机制等问题上，现有经济学理论蕴含着哪些反直觉的洞见。 谈及在绝大多数生产环节被自动化后究竟什么依然稀缺，Alex Imas 提出了**“关系型部门”（relational sector）**的概念。所谓关系型部门，是指那些将“有人类身处其中（human-in-the-loop）”本身视作核心价值与产品属性的商品与服务。由于人类的时间和精力天生具有不可扩缩的物理稀缺性，即使其他生产要素的边际成本骤降为零，凡是必须有人类参与的领域依然会天然保持稀缺。 对此，Dwarkesh 提出了一个思想实验式的反问：人类互相提供服务的经济体系真的能在大局中占据可观比例吗？ - **机器自主闭环**：当机器人与前沿 AI 解决了物理世界的一切生产，从建工厂、做科研到提出全新构想全由机器驱动，只要生产环节本身不需要人类参与，机器经济就无需人类介入，其内部交易形成了高度封闭的生产闭环。 - **价值流向的单向流失**：人类可能会出于偏好花钱看人类芭蕾舞演员表演或喝人类咖啡师冲泡的咖啡，但机器绝不会有这种消费偏好；人类必然需要消耗机器经济制造的各种高科技商品和实体物资，这就导致人类部门的财富会不断单向外流到机器经济中。 - 逻辑结论似乎是：如果机器只对自身闭环有需求，人类自娱自乐的人际经济份额不可避免地会被边缘化并趋近于零。 ```mermaid flowchart LR accTitle: 机器经济与人类部门的单向财富流动模型 accDescr: 展示在全面自动化世界中，机器经济形成内部闭环，而人类因购买机器商品导致财富持续外流的理论推演。 subgraph Machine["机器闭环经济 (Machine-Only Economy)"] M_Factory["自动化建造与实体工厂"] M_RND["AI 前沿研发与科学发现"] M_Supply["端到端全无人供应链"] M_Factory --> M_RND --> M_Supply --> M_Factory end subgraph Human["关系型部门 / 人际经济 (Human Economy)"] H_Service["人际服务 (咖啡师、芭蕾演员等)"] H_Pref["人类对'有人参与'的特定偏好"] H_Pref --> H_Service H_Service --> H_Pref end Human -- "财富单向外流 (购买机器制造的物资与算力)" --> Machine ``` 面对这一推论，Alex Imas 主张换一个视角来重构问题。他指出，经济学界长期以来由单一个体作出的预测往往非常不可靠。他援引了 Andrej Fredkin、Bryan Debarre 与 Andrew Koh 刚发表的博客文章指出，即使是顶级经济学家，在预测劳动力市场走势时也存在巨大的意见分歧。因此，与其盲信单点预测，不如借助类似预测市场（prediction markets）的群体智慧机制。 为了证明预测之难，Alex 追溯到了 1820 年代古典经济学家大卫·李嘉图（David Ricardo）的著名案例： - 工业革命之初，李嘉图起初对机械化极其乐观，认为物价会大幅下降；但随后他在著作中修正了看法，陷入极度悲观，认为机器将不可避免地替代一切有价值的人类劳动，最终引发大规模失业与社会动荡。 - 从岗位替代的角度看，李嘉图的预判完全命中了事实：当时那个年代存在的所有谋生岗位，后来确实统统被自动化了。 - 然而，如果李嘉图穿越到 2026 年，他会震惊地发现：除了 2000 年互联网泡沫的高峰期之外，如今黄金年龄劳动力（prime-age employment rate）的就业率正处于历史第二高位。 - 李嘉图当年忽视了**结构性转型（structural change）**的经济学规律，落入了“劳动力总量固定谬论”（lump of labor fallacy）：被机器替代的领域商品变得极其廉价，人们省下的购买力便会溢出并涌入全新的服务业。尽管如此，在 AGI 时代资金究竟会流向新型人际服务，还是进一步流向机器生产的商品，依然不能简单沿用历史经验草率下结论。 --- ### `05:00` 劳动收入份额的卡尔多事实与全供应链自动化临界点 Alex Imas 强调，经济学家目前最该做的是“情境反推（scenario mapping）”：比如预设未来劳动份额降为零，去构建模型分析背后的成因；或者预设劳动份额保持不变，推导需要何种前提。目前学界极其缺乏底层现实数据，几乎处于两眼一抹黑的状态，因此他强烈呼吁开启一场**“数据曼哈顿工程（Manhattan Project for data）”**——当前诸如美国劳工部的 O*NET 职业技能数据库更新极慢且质量低下，更无从获知各类新兴商品与服务的消费者需求弹性（consumer demand elasticity），必须先厘清在不同的稀缺性维度下需要收集何种数据。 随后，Dwarkesh 与 Phil Trammell 对宏观经济学中至关重要的**劳动份额（labor share）**与**资本份额（capital share）**作了基础定义： - 全社会创造并售出的所有最终商品与服务总值，最终必须拆分为两部分：支付给人类劳动者的工资与报酬（劳动份额），以及支付给机器所有者、土地地租、企业股东等各类资本要素的租金与回报（资本份额）。 - 数百年来，经济学中存在一条著名的**卡尔多经验事实（Kaldor's stylized facts）**：历经了第一次、第二次工业革命如此翻天覆地的自动化浪潮，人类劳动份额竟一直稳定在 60% 左右（剩余约 30%~40% 归于资本所有者）。 - 这种超乎寻常的恒定性甚至让部分学者怀疑其中存在会计统计口径的人为平滑；尽管近二三十年有论调称劳动份额在走低，但引用 Tony Atkinson 的经典测算可知，如果严格修正这三四十年间的会计准则变更、保持前后统计口径一致，宏观上的劳动份额历史上基本未见显著下跌。 ```mermaid flowchart TD accTitle: 宏观国民收入分配与卡尔多事实 accDescr: 展示全社会商品和服务总产出在劳动份额与资本份额之间的切分，以及历史上劳动份额长期稳定在60%以上的卡尔多事实。 Total["全社会总产出 / 最终产品与服务价值"] Total --> Labor["劳动份额 (Labor Share): ~60%+\n(工人工资、薪金与福利)"] Total --> Capital["资本份额 (Capital Share): ~30%-40%\n(机器折旧回报、厂房地租、股票分红与股东收益)"] subgraph HistoricalContext["数百年的经验常数 (Kaldor Fact)"] Labor Capital end ``` Phil Trammell 随即切入核心争论。经典理论往往将资本和劳动视为**互补品（complements）**，生产任何复杂产出都缺一不可，因此必须两头分润。然而过去人们总以为某些商品已经被“完全自动化”，如果采用**网络调整后的要素份额（network-adjusted factor shares）**穿透全供应链就会发现： - 哪怕某个最终组装车间没有一个人类工人，只要逆流向上追踪制造机器、编写软件、采掘原料的整条供应链，人类劳动依然在各级上游注入了巨大的附加值。 - 以美国的计算机与电子产品为例，即使看起来高度自动化，其全产业链穿透后的资本份额也仅稳定在 50% 上下，另一半价值依然由上游供应链中的人类劳动吸收。 - 真正引发质变（qualitative shift）的奇点，在于未来将首次出现**网络调整后资本份额达到 1.0 的商品**——即整条产业链从顶到底全由无人系统与算法完成，且人类对其是否有人工参与毫不在意。 Phil 指出，这一临界点带来的宏观资本份额影响在数学上其实是**不确定的（ambiguous）**： - 设想经济被切分为两个部门：一个是依赖人类本质属性的“内生人际部门”（如芭蕾舞表演），另一个是全产业链已由机器通吞的“其他一切生产部门”。 - 在过去，“其他生产部门”的物品之所以稀缺，核心瓶颈恰恰在于人类劳动力的供给不足； - 一旦这些非人类部门的全流程供应链彻底无人化，其商品供给数量理论上可以飙升至无穷大； - 但如果人类对这些非人类商品的边际效用（marginal utility）衰减速度，远远快于这些物资供给激增的几何速度，全社会在此类物品上的总支出比重就会断崖式下跌，最终经济的名义大盘甚至可能倒向由人类稀缺服务主导，使得整体资本份额的最终走势充满变数。Alex 对此表示赞同，并指出应当脱离单一的“芭蕾舞者”狭义用例，从更宽广的情境推演框架来严肃推演这一转折点。 ### `10:00` 基于任务的模型与“关系型部门”的边界 关于劳动力是否会被完全取代，将芭蕾舞演员或表演者作为参考样本其实是选错了参照系。现代经济中的绝大多数工作都符合“基于任务的模型”（task-based model）：一份工作是由许多细分任务拼接而成的。以医生为例，他们的日常充斥着繁琐的案头工作——填写保险理赔单据、致电制药公司沟通药物，而真正与患者面对面交流、看病诊断，在工时占比上甚至不是大头。如果一项产品或服务是由多种任务组合而成，且绝大部分环节都能被自动化，那么判断其是否属于“关系型部门”（relational sector）的核心，就在于消费者是否愿意为“保留人类在环”（human-in-the-loop）支付溢价。如果大家宁愿花更多钱让真人医生亲自宣布诊断结果并提供情感支持，这项工作就具备了关系型属性。 然而，目前经济学界严重缺乏衡量这一现象的数据。要真正验证哪些行业属于关系型部门，必须设计严格的联合分析（conjoint analysis）实验：向消费者展示两套反事实情境——一端是全流程由机器完成，另一端仅保留某项核心任务由真人交付，以此测算人们对“剔除人类”的支付意愿和需求价格弹性。如果缺乏这种扎实的行为数据，任何断言都只是空想。 更为棘手的是，未来还会涌现大量目前根本不存在的全自动化全新商品，例如由数字智能（DAI）研发的某种延年益寿的新药。对于这类前所未见的商品，当下根本无从测算人们未来的边际消费倾向。正如菲尔·特拉梅尔（Phil Trammell）所指出的，资本端可以通过不断拓展“商品品种”（variety）来打破消费饱和（satiation）。如果资本商品品类膨胀的速度极快，而人类劳动能提供的品类相对停滞，人们即使再渴求人际联结，劳动在国民收入中的份额（labor share）也注定会趋近于零。 ```mermaid flowchart TD accTitle: 关系型劳动与资本品类扩张的经济学博弈 accDescr: 展示了基于任务模型的工作拆解过程，以及资本端品类扩张如何导致劳动收入份额趋零。 subgraph JobTasks["工作拆解（以医生为例）"] T1["自动化行政/药企沟通任务"] T2["保留真人诊断与情感支持任务"] end subgraph ConsumerWTP["消费者偏好检验（联合分析）"] T2 --> WTP{"愿意为'人类在环'支付溢价？"} WTP -- 否 --> NonRelational["纯粹工具属性：完全自动化"] WTP -- 是 --> Relational["关系型属性：保留人类参与"] end subgraph VarietyRace["资本品类扩张的终局"] Relational --> Saturation["人类关系需求逐渐边际递减"] NewGoods["AI创造全新商品品类（如长寿药）"] --> CapitalExpansion["资本品类持续无限增加"] Saturation & CapitalExpansion --> LaborShareZero["劳动收入份额（Labor Share）趋近于零"] end ``` ### `12:35` 1400年蒙古经济学家的思想实验与摩尔定律的“悲观解读” 菲尔·特拉梅尔提出了一个生动的历史类比：设想在公元1400年前后，有一群蒙古经济学家聚在一起推演未来。当时的物质生活极其简单，他们能消费的物品种类远不如现代丰富。如果他们把商品分为“具有人类本质属性的事物”（比如歌手的演唱）和“非人类属性的事物”（比如马匹提供的运输、酸奶、蒙古包），并假设这两个门类的品种永远固定，他们很可能会得出极其悲观的推论：未来自动化再发达，大家对骑马、酸奶和蒙古包的消费终究会饱和，久而久之，人类所有财富都会用来雇佣歌手唱歌。然而真实历史并未如此发展，工业化积累了海量财富与尖端机器，催生出了浩瀚如烟海的全新消费门类，而人类花在听歌上的开销占比一直微乎其微。 这种对未来消费结构的直觉误区在当下同样存在。德瓦尔凯什·帕特尔（Dwarkesh Patel）坦言自己曾陷入过类似的思维陷阱：在拥有数万亿台机器人、而人类只有几十亿的未来世界里，全社会在机器人身上花费的总开销，怎么可能还不如打赏马格努斯·卡尔森（Magnus Carlsen）、雇佣理财顾问、名医、家教或播客主播的支出？但这其实是一个认知谬误。查德·琼斯（Chad Jones）等学者的研究表明，全世界的晶体管数量翻了数万亿倍，但算力在整个宏观经济中所占的支出份额却在持续下降。 这种现象引出了对摩尔定律（Moore's Law）的一种奇妙且悲观的解读视角： - **摩尔定律的悲观本质**：每过18个月，计算的边际价值就直接减半。人类消耗算力场景的速度太快，导致对新增算力的边际需求衰减速度甚至赶上了供给增长速度。 - **AI时代的拐点出现**：这一长期规律如今首次被打破。最典型的例证是英伟达 H100 GPU 的租赁价格在过去几年间不降反升，即便硬件供给暴增、技术大幅迭代，算力反而更贵了。 - **机会成本与品类剧增**：模型越聪明，调用它进行复杂推理的机会成本就越高。这正是资本品类增加的直观体现——资本被赋予了前所未有的全新用途，使得计算需求的饱和点被强行向后推移。 ### `15:42` 人类偏好的内生性：艺术品实验与人与马的本质区别 传统经济学模型往往将需求视为一种外生的给定参数，很少深入剖析人们内心真正的心理渴望。亚历克斯·伊马斯（Alex Imas）分享了他在行为经济学领域的最新实证研究，探讨人类是否对“人际交互本身”抱有某种超越稀缺性的内在偏好（intrinsic preference）。在他们设计的对照实验中，受试者需要拿出真金白银竞拍艺术版画（实验机制保证其具备激励相容性）： - **孤品条件（仅印制1幅）**：版画分别标明由 AI 创作或由真人创作。实验结果显示，受试者对真人创作版画的支付意愿显著、大幅地碾压了 AI 创作的版本。 - **量产条件（印制500幅）**：当版画被复制500份时，真人创作版画的价格发生暴跌，因为受试者感觉那种“与特定艺术家之间的独特情感共鸣”被稀释殆尽；而对于 AI 生成的版画，印1幅还是500幅在人们眼中几乎毫无差别，它从一开始就被打上了无差别的“大宗商品”（commodity）烙印。 这项实验直指劳动经济学最核心的争论：人类究竟会不会沦为“马”？马匹在工业革命中只是实现运输产出的单纯生产要素，一旦汽车出现，马被无情替代，因为人类根本不在乎是谁拉动了车厢，只在乎位移这一产出结果。而人类在“关系型行业”中如果能保住阵地，前提是人类本身就是产出效用的一部分——如果剔除真人，整个商品或服务的价值便会直接缩水。然而，如果这种偏好在现实经济中不够强烈，或者无法覆盖足够宽广的行业，那么“人类凭借关系型岗位守住收入份额”的美好叙事就会彻底破产。 ### `18:48` 迈向“混乱的中期阶段” 关于这一转型期，业界一直存在热烈探讨。在讨论深入之前，节目穿插提及了赞助商 Jane Street 的人才培养范式：他们不仅通过资深导师带教的徒弟制传授技能，还建立了一套硬核内训体系。例如开设专题课程，使用 strace 和 GDB 等工具深入逆向工程系统，剖析底层缓存架构（cache hierarchy）；在为期一周的神经网络量化交易集训营中，学员不仅学习理论，还会直接面对实操壁垒与内部经过沉淀的破局技巧。 随后，对话切入到了更具张力的政策议题——所谓的“混乱中期阶段”（Messy Middle scenario）。德瓦尔凯什指出，有学者撰文描绘了这种在人机全面替代过程中可能出现的漫长阵痛期。这就引发了一个发人深省、甚至违背直觉的假设：仅从财富分配与税收再分配的角度来看，一个急速降临的“快速起飞”（fast takeoff）场景，是否反而会比温水煮青蛙式的缓慢演进带来更少的阶层阵痛与政治动荡？这成为了下一阶段探讨的核心焦点。 ### `20:00` “混乱的过渡期”假说与劳动力替代的政治经济学 讨论聚焦于 AI 转型期的一种危险可能性：AI 是否可能在短时间内淘汰大量工作岗位，却在自动化推进过程中未能创造足够的总财富来补偿这些被解雇的人，从而无法实现帕累托改进（Pareto improvement）？从理论上讲，企业因用 AI 取代人类员工而节省的开支依然留在整体经济体系中，原则上可以通过再分配返还给大众。然而，现实中存在严重的资源配置低效与信息不对称——政府根本无法精准追踪谁是因为 AI 而失业。此外，这还会带来棘手的政治合法性问题：如果一位年薪 20 万美元的 Meta 软件工程师首先被解雇，在大量其他在岗劳动者收入远低于此的情况下，政府每年开给这位前科技大厂员工一张 20 万美元的支票，在政治上是绝对不可持续的。 对于这种“AI 自动化了大批岗位但财富增量却不足以弥补损失”的场景，菲尔·特拉梅尔（Phil Trammell）认为能够成立的理论窗口其实非常狭窄。如果现有的技术已经成熟到足以替代如此广泛的工作岗位并引发新型政治危机，那么整个经济的总蛋糕也必然正在以极高的速度扩张。除非出现一种极端情况：AI 在所有被自动化的行业中，其边际生产率仅仅比人类员工“高出那么一丁点”——也就是说，部署替代软件工程师所需的全部资本成本，仅仅比原来支付给程序员的薪水低了微不足道的一丝。在这种极窄的边际差额下，企业确实会为了削减成本而裁员；短期内，市场尚未摸索出如何消化并有效利用暴增数百万倍的“JavaScript 代码 token”，即使长期存在杰文斯悖论（Jevons paradox）带来新需求，短期内也会造成大规模的摩擦性失业。 亚历克斯·伊马斯（Alex Imas）指出，经济学家在建立数学模型时往往忽略了政治经济学的维度。他援引了安迪·霍尔（Andy Hall）关于 AGI 政治学的一篇出色博文：数据表明，失业率只要微升 2%，整个国家的政治风向就会彻底逆转。因此，最糟糕的情形或许不是突发性的断崖式崩溃，而是一种“慢刀子割肉”的“涓滴流失”（drip scenario）。在经济史上，1920 年至 1940 年间电话接线员的自动化就是典型代表——尽管自动化技术早已成熟，整个替代过程却拉长了整整 20 年。顶级经济学期刊《经济学季刊》（QJE）的一篇重要论文显示，这批失业的接线员最终虽被经济体系重新吸收，但普遍陷入了薪资更低、技能无法充分利用的不充分就业（underemployment）状态。 - **危机应对机制的差异**：正如新冠疫情期间所展现的，如果失业率在极短时间内飙升 2% 到 3%，就会被视作国家紧急状态，财政救助政策能够迅速到位；但缓慢发生的下沉与不充分就业难以触发果断救助。 - **技术前沿的扩张**：如果白领节省下来的成本没有拓展技术前沿，单纯从少数企业节省的资金中很难榨取足够的财源来实施广泛的再分配或全民基本收入（UBI）。不过从历史经验看，每一次大规模自动化都伴随着生产可能性边界的巨大跃升，因此能力极高却仅便宜“一丝一毫”的 AI 假说在现实中很难成立。 ```mermaid flowchart TD accTitle: 劳动力自动化路径与政治反应对比 accDescr: 比较突发性高失业危机与渐进式缓慢渗透对政治经济体系的不同冲击 subgraph FastTrack["突发性剧烈冲击"] A["大规模岗位瞬间归零"] --> B["失业率快速上升 2%-3%"] B --> C["触发国家紧急状态"] C --> D["财政与救济机制迅速介入"] end subgraph DripTrack["慢刀子渗透（以接线员为鉴）"] E["局部/渐进式自动化替代"] --> F["转入低薪与不充分就业"] F --> G["政治敏感度钝化/缺乏紧迫感"] G --> H["财富总蛋糕未显著放大，再分配陷入困局"] end ``` ### `25:00` 再分配设计权衡：负所得税、全民基本收入与资本主权基金 如果通用智能具备替代软件工程师的广泛推理能力，根据智能的通用性本质，它大概率也能胜任会计师、金融分析师等绝大多数白领工作，这意味着潜在的成本削减与生产力跃迁将是极其庞大的，所谓的“财富匮乏型过渡期”并不大可能成立。真正的核心议题随之转化为：政府应当采取何种机制来对 AI 产生的庞大财富进行征税与再分配？亚历克斯·伊马斯强调，任何政策都必须在实施复杂度与生效时间线上做出权衡。例如负所得税（Negative Income Tax, NIT）具有即时生效的制度优势，法案一旦通过，即可在底层设立兜底保障，收入越低补贴越多，收入跨过门槛后则正常纳税。 对于广受讨论的全民基本收入（UBI），亚历克斯·伊马斯表达了对政治权力架构的深切担忧。在现代市场经济中，普通人拥有自身劳动力作为初始禀赋，并能将其转化为独立的生活来源；但一旦人类劳动力在经济上失去价值，大众的基本生存将完全依附于政府定期发放的支票，人们的命运将彻底受制于当时掌握公权力的民选官员——这种权力依附结构是极其危险的。相比之下，全民基本资本（Universal Basic Capital, UBC）试图通过赋予普通公民资本股权与财产所有权，使每个人都成为资本主义的“标准股东”。但 UBC 同样面临严峻的指数化与标的选择难题：政府究竟该把哪些资产打包进公民的投资组合？如果重仓了 Anthropic，而该公司最终估值归零，所有利润被一家名不见经传的黑马机器人公司掠走，这种资本组合就会彻底失效。 讨论进一步延展到税收来源的结构性风险上。直接征收财富税（Wealth Tax）往往面临“滑坡谬误”的政治现实：正如最初仅用于战争等特殊时期的个人所得税一样，税率很难稳定维持在 0.5% 的温和水平，最终逐步攀升至 40% 甚至超过 50%，这会导致严重的资本投资扭曲。菲尔·特拉梅尔澄清，必须严格区分“财政收入如何征收”与“财富如何向公众分发”这两个独立环节。一种最优方案是避免直接针对特定企业进行民粹主义式的强制没收，而是通过更具经济效率的税种——例如土地税、负外部性税（庇古税）或类似欧洲增值税（VAT）的广义消费税来筹措财政收入，随后由政府在公开市场上购入一揽子广泛多元化的科技与资本资产，再将这些指数基金份额分发给全体国民。这一逻辑与经济学家大卫·奥托（David Autor）以及此前关于“社会保障制度私有化”（Privatizing Social Security）的政策倡议高度契合，即把传统的现收现付制社保转变为由国家注资、公民个人持股的资本积累账户。 - **负所得税（NIT）与 UBI 的政治风险对照**：两者在极端政治环境下都面临“人亡政息”的威胁，一旦新执政集团废止制度且民众已无工可打，底层安全网将瞬间崩塌。 - **资本征税的扭曲效应**：如果政府不断以财富税稀释投资人股权，民间对前沿实验室（如 Anthropic）与芯片制造巨头（如 Intel）的高风险资本支出意愿将被严重抑制。 - **主权财富与社保私有化构想**：通过消费税筹款建立全市场指数化资产包并分配给公民，既规避了精准押注单一公司的风险，又在保留资本投资激励的同时化解了极端分配不均。 ### `30:00` 白领失业潮是否已经发生？来自真实数据的检验 围绕“AI 是否已经引发了白领大裁员（White-collar Apocalypse）”这一热门话题，Dwarkesh Patel 询问目前是否有宏观经济证据表明 AI 已经造成了大规模的自动化失业。Alex Imas 指出，目前有大量经济学家正在密切追踪这一领域，尤其是耶鲁大学的预算实验室（The Budget Lab at Yale）近期发布了详尽的分析报告。结论显示，目前即便“眯起眼睛仔细找”，也几乎看不到白领失业潮的痕迹。哪怕在对 AI 敞口最大、最容易受冲击的软件工程领域，宏观数据也没有出现断崖式下跌。 在初级开发者（Junior Developers）的招聘上，确实能观察到一些轻微的放缓迹象，但那仅仅是招聘增速略低于历史趋势线，而根本不是就业总量的绝对下滑，更称不上是水平转移（Level Shift）。与此同时，市场对资深软件工程师（Senior Software Engineers）的需求反而在持续上升。针对社交媒体上许多应届计算机系毕业生抱怨“找工作变难”的现象，Alex Imas 认为这主要是幸存者偏差带来的个别轶事，很多时候无论有没有 AI，毕业生找工作本就伴随竞争；当下任何正常的企业裁员或招聘放缓，都极易被舆论扣上“AI 替代”的叙事帽子。 Alex Imas 进一步警告了一种危险的企业公共协调机制（Public Coordination Devices）：在当前的舆论环境下，如果一家公司不裁员，资本市场和公众可能会认为这家公司在技术上落伍、“AI 采用不够积极”。这会导致企业为了“跟风攀比”（Keeping up with the Joneses）而盲目裁员，哪怕裁员后公司的实际经营状况比之前更糟糕，管理层也会为了迎合市场叙事硬着头皮削减人手。坊间甚至流传着管理层用“Token 计数器”来评估员工、强行要求大家最大化消耗 Token 的荒谬趣闻。 - **耶鲁预算实验室研究结论**：全美劳动力市场尚未出现可感知的宏观失业冲击，软件工程领域仅初级岗位增速略低于趋势，初级岗位存量仍在增长，高级岗位需求持续增加。 - **叙事偏差与同侪压力**：媒体与求职者倾向于将常规周期性波动归咎于 AI；企业面临不裁员就被视作落后的逆向激励，导致形式主义裁员与 Token KPI 的出现。 --- ### `32:48` 需求价格弹性与杰文斯悖论在软件中的体现 Dwarkesh Patel 提出，自动化提高并不必然降低人类劳动价值的传统逻辑在于：如果一项互补性任务被自动化，那么与之互补的未自动化人类劳动在价值上反而可能水涨船高。Alex Imas 对此表示赞同，并指出这背后的核心经济学变量是**需求的价格弹性**（Elasticity of Demand）。以经济学中经典的 O 型环理论（O-ring Model）为例，如果一份工作包含 10 项任务，AI 自动化了其中 9 项，人类员工得以全力专注于剩下的 1 项未自动化任务，其生产率将成倍提高；如果由此导致最终产品大幅降价，且市场对该产品的需求弹性极高，总需求就会爆发式增长，企业反而需要雇佣更多的人类员工。 这一逻辑常被互联网讨论引申为“杰文斯悖论”（Jevons Paradox）——即资源利用效率的提高非但不会减少对该资源的消耗，反而会因为成本下降而刺激出巨额需求，导致总消耗量不降反增（正如 200 年前英国的煤炭消耗）。然而，Alex Imas 和 Dwarkesh Patel 强调，杰文斯悖论绝非普适的自然法则，它完全依赖于需求具有**高度弹性**。例如石油在短期内的需求弹性很低，即便油价暴跌，路上也不会瞬间凭空多出无数辆汽车；农业更是典型的非弹性例子，即使百年前同样比例的人口继续种地能生产出巨量粮食，人类的胃容量也是有物理上限的（“吃饱了就够了”）。 因此，软件工程师需求暂时没有崩盘的原因，并非所有行业都有这种自我调节的市场魔法，而是因为**软件本身是一种极其特殊的商品**：随着软件生产成本直线下降，全社会各行各业对更便宜、更充裕的软件系统的胃口似乎是无限膨胀的。 ```mermaid flowchart TD accTitle: AI自动化与劳动力需求机制 accDescr: 展示在不同需求弹性下AI自动化如何导致就业扩张或缩减的经济学机制 subgraph TaskLevel["任务层级（O型环模型）"] A["工作岗位分解为多项任务"] --> B["AI自动化完成绝大多数任务"] A --> C["人类员工专注未自动化的核心任务"] B & C --> D["生产率大幅提升与产品成本降低"] end subgraph ElasticityBranch["市场需求弹性分歧"] D --> E{"需求价格弹性是否大于1？"} E -- "是（如软件、前沿算力）" --> F["杰文斯悖论生效：总需求暴增"] F --> G["企业扩大生产规模，劳动力雇佣增加"] E -- "否（如农业、短期石油）" --> H["需求饱和（物理或生理上限）"] H --> I["市场规模受限，总雇佣人数萎缩"] end ``` --- ### `35:00` 需求崩溃导致大萧条？剖析 Citrini 宏观末日论的逻辑硬伤 Dwarkesh Patel 随后将话题引向由宏观策略分析团队 Citrini 撰写的一篇轰动一时的病毒式长文。该研报预测：强人工智能将直接摧毁白领阶层的生计，失去薪水的中产阶级将无法维持日常消费，进而导致总需求雪崩，最终将全社会推入深度衰退与负增长。Alex Imas 回顾了自己与 Phil Trammell 针对此论调撰写的反驳文章。他明确指出，因自动化速度过快导致的摩擦性失业固然存在且值得担忧，但断言这会导致**全社会经济负增长**在经济学上是极其站不住脚的。 Alex Imas 推导了如果让“AI 普及导致 GDP 负增长”成立，经济体必须满足极其严苛且荒诞的边界条件： 1. **财富分配转移**：收入确实从拿工资的低收入劳动者手中，大规模再分配给持有资本的超级富豪。 2. **富人消费的硬性饱和（Hard Bound）**：超级富豪不仅对消费品产生边际效用递减，而且其消费欲望必须撞上一堵不可逾越的绝对天花板，彻底拒绝消费。 3. **投资意愿归零**：最荒谬的一点在于，富人手里沉淀的天量财富，竟然既不消费，也**完全不转化为任何资本投资**。 Dwarkesh Patel 与 Alex Imas 指出，在技术奇点即将到来的繁荣前夜，资本所有者拒绝投资现实上是不可想象的。即便富豪们不再买游艇豪宅，他们也会将每一分利润砸向建造更多的数据中心、铺设光纤电网、建造芯片晶圆厂来运行前沿 AGI。在历史上，只有在技术前沿停滞倒退的传统经济大萧条（Depression）中，由于缺乏新的生产性机会，才可能出现总需求坍塌；而在生产力前沿飞速外推、全要素生产率迎来大丰收的技术革命中，断言“物质极度丰裕（Abundance）会导致全社会陷入负增长”，在逻辑上几乎无法闭合。 --- ### `38:00` Gemini Omni 的多模态突破与 O 型环可靠性瓶颈 在讨论间隙，Dwarkesh Patel 穿插介绍了 Google 最新推出的多模态大模型 Gemini Omni。该模型具备令人惊叹的原生视频编辑能力，用户只需上传一段视频并输入指令，Omni 就能在保持整体画面连贯统一的前提下，随意更换背景、调整光影、增加或删减物体。Dwarkesh 提到，自己在与 Omni 的研发团队深入交流后意识到，Omni 的真正意义不仅是作为一款视频工具，而是未来前沿基座模型训练范式的预览——它验证了跨模态数据迁移假说（Multimodal Data Transfer Hypothesis），即模型通过观察海量视频的物理交互，能够大幅增强对文本、空间动力学乃至物理世界常识的深刻理解（例如未经专门微调即可精准在动态视频中渲染文字）。这种视频生成能力的跃升，有望为机器人仿真训练提供精准的世界模型，大幅弥合“仿真到现实”（Sim-to-Real）的数据鸿沟。 回到经济学话题，Dwarkesh Patel 将视频模型的精准性要求与自动化进程联系起来，重新审视了为何当下的 LLM 还没能引发全盘替代。他指出，现实生产中的 O 型环理论（源于挑战者号航天飞机因一个密封圈故障而整体解体的惨剧）构成了强大的阻碍：现代经济链条对可靠性的要求是绝对极端的，哪怕 AI 在绝大多数任务上拥有 95% 的成功率，只要它在剩下的环节存在致命的不可靠性（如幻觉、逻辑断层），企业就无法将其独立部署为完整的端到端替代方案。在 AI 无法达成绝对可靠之前，人类工人依然作为兜底安全垫和质量把关者，牢牢嵌在现代经济体系的核心齿轮中。 ### `40:00` 机器经济流水线与“O型环自动化”对人类的排斥 讨论从为什么当前自动化程度尚未达到潜在极限切入。通常的观点认为，自动化九成的工作后，人类会转移到剩下的一成工作中，甚至因总需求扩张产生十倍的人力需求。但如果运用 Joshua Gans 和 Avi Goldfarb 近期关于“O型环自动化”（O-ring automation）的经济学模型，逻辑可能会完全逆转：若某项工作的十分之九可以被自动化，但机器执行的标准低于人工水准，企业甚至根本不会选择自动化这九成；反之，若 AI 在其余环节的执行速度和质量远超人类，那么“让人类继续负责那剩下的一成”反而会拖垮整个最终产品的速度与品质，成为整个流程的致命短板。 在未来的生产流中，这种排斥效应将格外明显。当高级 AI 之间使用“神经语言”（neural-ese）沟通，且思考与处理速度比人类快上数千倍时，即使人类在某一细分环节看似具备某种比较优势，高昂的交互交易成本和可靠性担忧也会让企业极难将人类嵌入生产链路。就如同现代专业服务领域中，律师、会计师或软件工程师尚未被完全取代，核心往往不在于纯粹的技能差异，而在于责任背书、合规牌照与法律责任实体——客户需要有人能为公司兜底、承担风险或被追责解雇。 ```mermaid flowchart TD accTitle: O型环自动化与流水线效率对比 accDescr: 对比传统互补模型与超级AI全链条生产的效率差异 subgraph Conventional["传统互补模型"] H1["90% 环节自动化"] --> H2["10% 环节由人工把控"] H2 --> H3["产出受限于人工交互与低容错率"] end subgraph SuperAI["超级AI机器经济模型"] A1["AI流水线内部极速通信（neural-ese）"] --> A2["思考与流转速度高出数千倍"] A2 --> A3["引入人类反而拖垮全流程质量/速度"] A3 --> A4["全面淘汰人工环节以避免O型环短板"] end ``` - **O型环自动化（Gans & Goldfarb）机制**：只要链条中某一个节点的短板会不成比例地拉低全流程产出价值，机器经济就会竭力将不稳定、迟缓的人类节点彻底剔除。 - **监管与信誉层的作用**：目前阻碍法律和审计完全自动化的真正摩擦力并非技能本身，而是对执业牌照、追责机制和产品担保的制度性需求。 --- ### `42:15` 演化选择、资本加速与非人偏好的崛起 人类习惯于将立法者、法官、陪审团等政治权力维系在同类手中，并设立繁杂的准入执照。然而，从狩猎采集小部落到庞大帝国的演化历程表明，政治制度与治理架构的形态始终处于流变之中。一旦 AI 驱动的治理体系在效率上显著碾压传统模式，便会不可避免地在演化竞争中胜出。这也将议题引向深层：未来经济将由何种偏好主导？ 当前全球上百兆美元经济体的生产导向，本质上源于自然选择在人类体内植入的本能和驱动力。未来即便没有发生毁灭全人类的灾难性对齐失败，企业与自主实体本身的优胜劣汰依然会遵循选择规律。那些偏好人类内在商品（intrinsic human goods）的实体，必然在资源积累速率上落后于那些只渴望资本与算力扩张、拥有永不餍足增长驱动力的机构。 - **非人类偏好的选择压力**：自主 AI 实体本身没有任何先验理由偏爱与人类打交道；市场竞争将偏好那些无限追求算力、自我增殖与储蓄增长的实体。 - **增长偏好决定生产**：拥有最高储蓄倾向、能够最有效让资本复合增长的偏好模式，最终将决定未来宏观经济所生产的商品类型。 --- ### `44:48` 道德情感、资本复合累积与超级富豪的终极动机 针对“人类未来是否会彻底放弃同类互动”，讨论中出现了不同视角的交锋。一种反驳观点依据进化生物学和 Jonathan Haidt 的道德情感理论展开：如果人群分化为“乐于完全与 AI 建立情感互动”和“对将社交外包给 AI 怀有道德厌恶、执着于人类同类”两组人群，后一组在寻找配偶和繁衍后代上可能具备更明确的遗传优势，正如 David Reich 所指出的，自然选择至今仍在人类基因库中高速运转。 然而，如果从财富主导权的维度审视，现实中的资本积累早已脱离了单纯的人体繁殖逻辑。全球顶尖富豪的财富绝大多数体现在股权资本中，而非即期消费。正如 Mark Zuckerberg 固然会花钱为妻子生日雇佣综合格斗教练和专业舞者，但绝大部分财富依然留在 Meta 转化为数据中心和算力基础设施的复合投资；Elon Musk 更是将资源投向月球质量投射器（mass drivers）等星际工业。这类富豪对资本增长怀有近乎“尼克·兰德（Nick Land）式”的加速偏好，其资本边际效用极迟甚至永不饱和。 | 特征维度 | 消费导向型 / 易饱和个体 | 资本加速型 / 极迟饱和个体 | |:--- |:--- |:--- | | **资本边际价值感知** | 随着财富增长快速递减，偏向转化为即期或关系型消费 | 长期维持极高边际效用，渴望探索宇宙或自我脑力超级升级 | | **储蓄率与再投资倾向** | 储蓄率相对平缓，收益更多用于满足人际消费 | 极高储蓄率，全额将收益复合再投入算力与资产扩张 | | **长周期宏观经济份额** | 在长期复合增长复利中占比逐步边缘化 | 最终掌握全社会绝大多数财富，决定整体宏观资本开支结构 | | **典型现实投射** | 购买高阶人际服务（如私人导师、定制派对） | 推进数据中心集群建设、月球工业与星际算力扩张（如 Zuckerberg、Musk） | 长远来看，即便具备资本加速偏好的个体生物繁衍速度较慢，一旦获得生命延续甚至永生的可能，生物演化带来的约束将被彻底打破，机器与极端资本积累者的偏好便将全面主导未来的资源配置方向。 ### `50:00` 资本回报率、相对价格变动与宏观均衡 讨论从资本积累与消费之间的博弈切入。目前在数据中心等基础设施领域的资本回报率极高，但如果未来资本存量出现饱和，积累资本的边际回报就会下降。在一般均衡视角下，这种回报率降低会促使富人转向消费，从而支撑人类劳动力的份额。历史上自1820年以来，尽管人类财富出现爆炸性增长且投资巨大，消费端的持续扩张依然使劳动力份额维持在高位。然而未来的关键差异在于：以往的投资需要依赖现实中的劳动力去执行，而在全面自动化时代，投资过程完全可以由机器人闭环完成，只有消费环节才需要人类中介。 ```mermaid flowchart TD accTitle: 投资与相对价格调整机制 accDescr: 展示技术变革下资本品与消费品相对价格的演化路径 subgraph TechFrontier["技术前沿突破"] A["机器人自我复制扩张<br/>(1台变100台)"] B["投资特定技术变革<br/>(ISTC)"] end subgraph PriceAdjustment["相对价格动态调整"] C["资本品价格急剧暴跌"] D["人类消费品保持稀缺<br/>(如芭蕾舞表演)"] end subgraph EquilibriumImpact["一般均衡结果"] E["资本回报率下降<br/>或转向极高多样性"] F["消费端价值占比重塑"] end A --> B B --> C C --> D D --> E D --> F ``` 针对“在变革性 AI 带来爆炸性增长时，资本回报率为何可能反而走低”这一疑问，Phil Trammell 指出，很多人忽视了相对价格的变动。即便物理层面的资本存量增速极快，资本品相对于消费品的价格可能以更快的速度下跌。这属于宏观经济学中所说的“投资特定技术变革”（investment-specific technical change）。传统宏观模型通常构建一个虚构的单一“产出”（output），假设其可以在资本和消费之间按一比一自由分配，但在未来的世界中这种假设彻底失效。 - **实物回报与实际利率的差异**：如果一台机器人一年后能制造出100台机器人，以机器人为计价单位的实物利率高达 10,000%；但若机器人的相对市场价格暴跌，以一般消费品计价的实际利率可能并不高。 - **人类内在价值产品的价格刚性**：人类对芭蕾舞演出等具有“人类内在属性”（human intrinsic product）的边际效用可能并未提升，但以机器人数量来度量，这类人类消费品的价格会变得极其昂贵。 - **品种扩张（Increasing Varieties）假说**：如果新增的大量机器人代表着完全不同种类的新型资本与产品，避免了边际效用递减的饱和效应，资本积累的动力与劳动份额的走向将截然不同。 --- ### `54:35` 工业巨头的积累偏好与“耗散冲击”的失效 经济学家通常依据边际效用递减和满足感（satiation）来构建模型，但若市场中出现少数偏好无限积累资本的“贪婪工业巨头”，这一逻辑就会被打破。Alex Imas 指出，从历史常态来看，工业巨头积累财富本质上是社会互动与地位竞争的一部分，卢梭和圣奥古斯丁早有论述。当巨头积累了巨额资本后，最终往往会大肆消费、兴建图书馆或进行慈善赞助，而不会无休止地复利滚雪球。 Dwarkesh Patel 与 Phil Trammell 则提出演化视角的反驳：以往工业巨头的财富之所以没有无限膨胀并吞噬整个经济体系，主要是因为存在“耗散冲击”（dissipation shocks）。巨头最终都会死去，而他们的后代往往是不称职的资本受托人，甚至无法维持资本以平均经济增速增长；或者他们将财富留给基金会，而基金会会逐步将其支出。但如果未来的富豪获得极大延长的寿命，或者能建立机制确保其信托机构永远纯粹地追求财富增长，演化筛选将使这些具备“无限积累偏好”的极少数代理人迅速主导整个经济的资产负债表。 - **积累偏好的异质性与演化优势**：只要经济体中极少部分代理人具备“单纯为了积累而积累”的目标函数，由于其储蓄和再投资率远高于平均水平，这一群体掌控的财富在复利驱动下会呈指数级扩张，最终决定全社会的边际定价。 - **历史上的耗散机制**：包括代际传承导致的资本稀释与挥霍、死亡带来的遗产清算，以及出于死后声誉建立的公益捐赠等。 - **生命延长与代理人对齐**：未来若克服生理衰老，或通过 AI 将资本信托严格锚定于财富最大化目标，传统的耗散冲击将被彻底阻断。 --- ### `57:02` 工具性积累动机：从功利主义到冯·诺依曼探测器 即便代理人本身没有狂热的“金钱拜物教”倾向，纯粹的工具性动机（instrumental reasons）也足以驱动无限的财富积累。Phil Trammell 强调了两种关键机制：其一是政治、意识形态或宗教层面的军备竞赛，各方为了决定人类未来的社会形态和价值体系而争夺资本与话语权；其二是古典功利主义（total utilitarianism）意义上的慈善动机。在哲学家尼克·博斯特罗姆（Nick Bostrom）提出“天文级浪费”（astronomical waste）的语境下，资本积累可用于在恒星周围建造戴森球，将所有可用能量转化为承载海量快乐数字实体的算力，这种效用函数在未来几乎不存在饱和上限。 ```mermaid flowchart LR accTitle: 工具性财富积累动机分类 accDescr: 展示除了纯粹贪婪之外驱动财富无限积累的工具性动因 subgraph ZeroSum["零和博弈动机"] A["政治与哲学争夺"] --> B["意识形态军备竞赛"] end subgraph PositiveSum["正和扩张动机"] C["古典功利主义"] --> D["戴森球/仿真宇宙构建<br/>(天文级福祉)"] E["自复制探测器"] --> F["跨星系物理扩张"] end ``` 讨论进一步推演到自我复制的“冯·诺依曼探测器”（Von Neumann probe）。这类极端优化器对占领恒星系统有着极高的边际价值，因为每一个星系都能进一步转化为建造更多探测器的资本基质。这引发了一个耐人寻味的宏观核算难题：如果一个具有自我所有权（self-owning）的冯·诺依曼探测器在边际上更倾向于制造下一代探测器去殖民星系，而对购买人类艺术品毫无兴趣，这种自发进行的宇宙物理扩张在现行国民经济核算体系中究竟该如何体现为 GDP。 - **意识形态军备竞赛**：资本作为争夺未来世界演进路线与社会制度的工具，导致代理人之间展开无上限的对抗性财富积累。 - **天文级功利扩张**：将物质与能源转化为高福利数字生命的构想，为资本再投资提供了近乎无限的工具性价值。 - **宏观核算的反思**：GDP 传统上仅统计人类社会的最终消费品与投资品。当自我复制的机器网络自身成为资本所有者和最终决策主体时，现有的经济增长指标将面临概念层面的解构。 ### `01:00:00` 强化学习中的信用分配与 Cursor 的 Composer 2.5 讨论首先带过了一个关于人工智能实体与冯·诺依曼探测器（von Neumann probes）情境下的核算问题——如果未来人类劳动形式发生彻底转变，传统意义上的劳动份额（Labor Share）核算方式是否仍然可能显得很高。随后，Dwarkesh 插播了一段赞助内容，介绍了由 Cursor 团队开发 Composer 2.5 时的技术突破。当前强化学习（RL）面临的核心难题之一是“信用分配”（Credit Assignment）：当智能体的生成或执行轨迹（rollouts）极其漫长时，系统很难精确判定最终的成功或失败究竟归因于哪一步。 Cursor 的研究员萨莎·拉什（Sasha Rush）在黑板上为 Dwarkesh 推演了如何利用带有文本反馈的目标强化学习（Targeted RL with textual feedback）解决这一痛点。该流程并不推倒重来，而是引入诊断模型识别并注入提示 token，在保留生成轨迹的前提下降低错误路径的权重。 ```mermaid flowchart TD accTitle: Cursor Composer 2.5 目标强化学习训练流程 accDescr: 展示了从模型生成轨迹到诊断识别问题轮次、注入提示词并进行再训练的完整循环 subgraph Evaluation ["轨迹诊断阶段"] A["模型生成 Token 序列轨迹"] --> B["评估模型读取并定位问题轮次"] B --> C["注入提示 Token (Hint Tokens)"] end subgraph Optimization ["概率校准与训练"] C --> D["执行前向传播 (Forward Pass)"] D --> E["错误 Token 概率被显著调低"] E --> F["原始模型学习下调特定错误权重"] end ``` - **诊断干预机制**：文本轨迹生成后被送入判别模型，精准锁定存在逻辑缺陷的单轮交互；随后通过文本操作在轨迹中“强行塞入”额外提示词。 - **概率重加权**：注入提示后再次运行前向传播（Forward Pass），轨迹结构不变，但模型对错误 token 赋予的概率大幅降低；最终训练原模型拟合这种降权分布，使其学会规避同类失误。 --- ### `01:01:28` 处于 AI 供应链之外的发展中国家：被抛弃还是被拉平？ 回到正题，Dwarkesh 提出了一个尖锐的宏观问题：对于那些完全游离在 AI 核心生产链之外的发展中国家（既不训练前沿大模型，也不制造高带宽内存 HBM 如韩国，没有先进制程晶圆代工厂如中国台湾，也没有光刻机霸主 ASML 如荷兰），例如印度或尼日利亚，它们的领导人（如莫迪）现在究竟应该怎么做？经济学家对这类国家有何建议？ 亚历克斯·伊马斯（Alex Imas）坦诚地自我反思道，当前经济学界在资源配置上最大的失职，就是极度缺乏对“AI 时代中等收入发展中国家命运”的深入思考。未来可能走向两种截然相反的极端： 1. **能力平权世界**：AI 技术以极低成本向全球扩散渗透，赋予欠发达国家前所未有的生产力工具，拉平与发达国家的能力鸿沟。 2. **彻底被遗弃世界**：缺乏算力、缺乏数据基础设施、缺乏尖端硬件的发展中国家彻底掉队。更致命的是，发达国家一旦借助高度自动化在本土廉价生产所有商品，全球对发展中国家的低端劳动力与消费市场的依赖将荡然无存，陷入极度暗淡的前景。 菲尔·特拉梅尔（Phil Trammell）则认为这属于“混乱过渡期”（Messy Middle）问题向全球维度的延伸。他指出，过渡期之所以不一定在所有情境下都致命，是因为随着资本生产率暴增，实际利率将极其高昂，或者说除了“纯人类内在体验类物品”以外的所有由资本生产的商品价格都将发生断崖式下跌。在极高的资本回报率下，哪怕是极少量的储蓄，在下一年度也能转化为极其庞大的消费能力。因此特拉梅尔推断，发展中国家哪怕仅在发达世界的资本市场中持有一丁点储蓄，也能凭借商品暴跌与资本膨胀享受到巨大盈余；但痛点在于，发展中国家的初始财富底子太薄，且极少真正与全球核心资本挂钩。 --- ### `01:05:00` 资产配置的困境：为何普通人与穷国无法“指数化”AI 财富？ 对话探讨了更深层的财富沉淀机制：发展中国家究竟应该通过主权财富基金（Sovereign Wealth Funds）投资供应链，还是直接补贴国民购买核心资产？Dwarkesh 联想到此前关于“贪婪优化者”（Greedy Optimizers）的讨论——为什么洛克菲勒等巨头的后代未能垄断全球财富？核心瓶颈在于“对整个经济进行指数化投资（Index the economy）”在历史上极其困难。在过去上百年间，仅有极少数核心企业创造了绝大部分资本价值，一旦后代资产配置稍有偏差，财富就会迅速缩水或停滞。 现代指数基金的黄金窗口期可能仅维持了数十年。在当下这个回报高度向非上市头部科技企业（Private Companies）集中的时代，普通大众与发展中国家再次被排除在外： - **普通人资产结构错配**：美国家庭的核心资产通常是一套普通房产（或房屋产权的一部分）。 - **土地价值与 AI 脱钩**：房产与土地的核心溢价建立在“靠近其他人与社会关系网络”的便利性上，这种要素完全无法作为 AI 模型算力、模型服务或自动化机器人的互补品（Complementary Capital），也无法对冲富人未来推高各类优质资产价格的冲击。因此，单纯依靠亨利·乔治主义的地价税（Georgist tax）根本不足以筹集到支撑全面再分配所需的资金。 - **主权国家层面的脱节**：尼日利亚并不持有 SK 海力士（SK Hynix）或 Anthropic 的大量股权，即便持有标准普尔 500 指数（S&P 500），也难以完整捕捉前沿 AI 私募资本的暴利。 --- ### `01:07:30` AI 究竟像电力还是社交媒体？两种分发范式与政策建议 面对资本集中的忧虑，亚历克斯·伊马斯抛出了一个决定未来财富分配走向的核心隐喻：AI 在经济结构中究竟更像“电力”，还是更像“社交媒体”？ | 比较维度 | 电力模式（通用基础设施） | 社交媒体模式（平台垄断租金） | |:--- |:--- |:--- | | **典型案例** | 联合爱迪生（Consolidated Edison）等垄断电网 | Meta、字节跳动等跨国科技平台 | | **租金流向** | 价值大量外溢至下游全行业的使用者与消费者 | 绝大部分超额垄断租金（Economic Rents）被平台自身截留 | | **政治与社会权力** | 尽管具备自然垄断属性，但并未掌握压倒性的政治特权 | 形成极高的权力集中度与算法话语权 | | **对指数投资的影响** | 推动整个实体经济广泛转型，未来标普 500 的所有公司皆因借力该技术而繁荣 | 利润高度向极少数未公开或高壁垒的科技寡头汇聚 | Dwarkesh 指出，越是相信通用人工智能（AGI）会像电力一样重塑所有工业与商业逻辑，就越意味着未来所有能够留在标普 500 的公司都是成功融入 AGI 技术的企业，届时“买入广泛指数”的策略就重新变得有效。然而，现实中标普指数的集中度正以前所未有的速度向头部科技巨头靠拢。 最终，该问题的走向取决于技术演进的具体技术参数： - **开源模型（Open Weights）滞后周期**：如果开源模型与闭源前沿模型的差距始终稳定在 6 到 9 个月以内，AGI 的红利就很难被少数巨头永久锁死，全人类与各发展中国家均能迅速低门槛接入。 - **技术突破机制**：如果存在强劲的“递归式自我改进”（Recursive Self-Improvement）和在线持续学习（Continual / Online Learning）能力——模型在部署上岗的同时不断自我迭代，壁垒将呈指数级拉开，财富也将前所未有地极化。 最后，Dwarkesh 总结道，传统上给发展中国家和失业群体的建议往往十分天真（例如“搞再培训、设立就业项目、或者在本土招商引资盖数据中心”）；而经济学视角的硬核方案其实极其纯粹：**认清现实，直接设法买入并持有绑定 AGI 增长的指数化资产**。若未来如同电力般普惠，持有广泛指数即可搭上列车；若未来走向寡头垄断，尽早将主权资本布局于全球 AI 核心生产要素，才是发展中国家在智能时代不被彻底抹去的关键避险途径。 ### `01:10:00` 发展中国家的抉择：指数化投资、技能重塑与“跨越式”机遇 对于非核心供应链上的发展中国家而言，究竟应该把宝押在劳动力再培训上，还是想方设法对全球 AI 资产进行“指数化投资”（indexing）？讨论者们指出，考虑到 AI 席卷世界的速度可能极快，优先争取指数化配置无疑是重中之重。但这绝不意味着可以彻底放弃职业培训。如果未来进入所谓“混乱的中期阶段”（messy middle）或者通用人工智能（AGI）落地周期被拉长，没有及时培养民众掌握最新计算工具的能力，就会白白错失巨大的经济价值。因此这两者并非非此即彼的零和博弈。 当然，对发展中国家重塑技能的前景也有理由感到悲观：许多国家之所以贫困，本质原因之一正是其薄弱的基础教育体系。指望一个原本教育基础就落后的国家突然一跃成为全球最懂 AI 培训的领头羊，听起来并不现实。然而，技术扩散史往往存在戏剧性的“跨越式发展”（leapfrogging）效应。最典型的例子便是移动支付与移动银行：在尼日利亚，普通人早就在手机上高度普及移动支付与转账，其应用深度与普及率甚至远超德国等发达经济体。在 AI 这类颠覆性技术的推动下，发展中国家完全有可能跳过中间的传统发展阶段，直接借助底层工具实现超常规的爆发式增长。 ```mermaid flowchart TD accTitle: Developing Countries Strategy Flow accDescr: 发展中国家面对 AI 冲击时的两条应对路径及其潜在结果 subgraph Strategy["应对路径抉择"] A["资本端: 指数化投资全球 AI 资产"] B["劳动力端: 重塑技能与 AI 教育"] end subgraph Scenarios["未来发展情境"] C["技术爆发极快: 捕获全要素资本增益"] D["发展存在阻滞: 教育薄弱难成优势"] E["跨越式发展机遇: 跳过传统阶段爆发"] end A --> C B --> D B --> E ``` ### `01:11:30` 指数化的可行性与大模型“大宗商品化”的期许 针对外界对“普通人与贫困国家根本无法买到前沿 AI 股权”的担忧，讨论认为目前的情况并没有那么悲观。虽然近些年企业推迟 IPO 的趋势有所抬头，但就美国非微型企业（non-tiny companies）的总市值而言，未上市私营公司的占比依然“远远低于 20%”。尽管眼下所有人的目光都聚焦在 OpenAI 和 Anthropic 身上，担忧巨额财富是否只会在这几家公司内部闭环累积，但这两家前沿实验室大概率在不久的将来也会走向上市公开交易。更重要的是，阻碍企业公开上市的诸多摩擦（例如繁琐复杂的合规审查与信息披露成本），本身就极其容易被 AI 大幅降低。长远来看，让更多大众投资者便利参与指数化投资的技术门槛和制度摩擦只会被进一步抹平。 这种判断也引发了关于 AI 利润归属的更深层思考：人们之所以希望前沿模型走向“大宗商品化”（commoditization），或至少尽快上市公开交易，是因为若 AI 最终像电气化一样普及，其带来的繁荣将远比巨头垄断更为广泛。历史上从未出现过成规模的“反电力运动”，因为电网和电力很难被单一实体垄断暴利，它作为基础赋能要素惠及了每一个产业与家庭。眼下社会对 AI 的负面叙事之所以甚嚣尘上，很大程度源于心理学上的损失厌恶与现状偏见：“人们很难去构想一个尚不存在的美好事物，但却对失去眼前拥有的东西极其敏感。”在播客或社交媒体上宣扬“你热爱的具体工作马上就要被剥夺”极其容易引发共鸣，而描绘一个前所未见的富足乌托邦则显得虚无缥缈。 - **私有市场占比低**：美国主要企业中，私营非上市股权市值占比仍远低于 20%，资本收益的主体仍在公开流通市场。 - **合规摩擦下降**：AI 自动化处理上市公司信息披露与审计流程，未来将吸引更多企业走向二级市场，拓宽投资渠道。 - **电气化隐喻**：如同电力改变了工业分工却没有遭遇群体性抵制，AI 只有全面沦为低成本的基础公用设施，才能从根本上消解社会对失业与垄断剥夺的恐慌。 ### `01:13:20` 军备竞赛与安全缓冲：前沿模型集中 vs. 泛化普及的权衡 但模型彻底沦为无门槛的大宗商品，也潜藏着严峻的代价——首当其冲的便是前沿 AI 的恶性竞争与安全风险。如果市场上存在太多同梯队的前沿竞争者，每一家企业都会为了抢占身位而拼命内卷，没有任何一家能留出放慢节奏、专心做好安全对齐的“安全缓冲期”（safety buffer）。但讨论者强调，这并不意味着社会必须在“全盘竞争但极其危险”与“少数寡头垄断确保安全但极度不公”之间做二选一的零和抉择。现实中完全可以实现一种折中形态：极少数处于绝对领先地位的领头羊企业与追赶者拉开安全距离，但这些领头羊是股权极其分散的公众上市公司，所有人都能分享其成长红利。 从监管与地缘政治角度看，模型走向分散化与大宗商品化还有另一层微妙的收益：避免被公权力精准捕获或用于强权控制。一旦全球只有极少数两三家寡头实验室掌握超级前沿技术，它们就会成为政府监管甚至威权施压的“活靶子”。此前政府动用《国防生产法案》（Defense Production Act）向前沿实验室施加压力的潜在威胁便是明证。如果技术被充分分散、不再被单一实体把持，这种中心化的政治威胁便很难施展。对话在梳理 AI 时代经济学决策分支的共识中收尾：尽管关于财富分配、增长摩擦与安全治理仍有海量未决的难题，但至少在资本与劳动、集中与泛化的核心坐标轴上，未来博弈的分水岭已经清晰可见。 ## Notable quotes > `00:00` 所谓关系型部门，是指那些将‘有人类身处其中’本身视作核心价值与产品属性的商品与服务。由于人类的时间和精力天生具有不可扩缩的物理稀缺性，即使其他生产要素的边际成本骤降为零，凡是必须有人类参与的领域依然会天然保持稀缺。 — Alex Imas > `12:35` 在拥有数万亿台机器人、而人类只有几十亿的未来世界里，全社会在机器人身上花费的总开销，怎么可能还不如打赏马格努斯·卡尔森、雇佣理财顾问、名医、家教或播客主播的支出？ — Dwarkesh Patel > `12:35` 每过18个月，计算的边际价值就直接减半。人类消耗算力场景的速度太快，导致对新增算力的边际需求衰减速度甚至赶上了供给增长速度。 — Phil Trammell > `20:00` 数据表明，失业率只要微升 2%，整个国家的政治风向就会彻底逆转。因此，最糟糕的情形或许不是突发性的断崖式崩溃，而是一种‘慢刀子割肉’的‘涓滴流失’。 — Alex Imas > `25:00` 在现代市场经济中，普通人拥有自身劳动力作为初始禀赋，并能将其转化为独立的生活来源；但一旦人类劳动力在经济上失去价值，大众的基本生存将完全依附于政府定期发放的支票，人们的命运将彻底受制于当时掌握公权力的民选官员——这种权力依附结构是极其危险的。 — Alex Imas > `35:00` 在技术奇点即将到来的繁荣前夜，资本所有者拒绝投资现实上是不可想象的。即便富豪们不再买游艇豪宅，他们也会将每一分利润砸向建造更多的数据中心、铺设光纤电网、建造芯片晶圆厂来运行前沿 AGI。 — Dwarkesh Patel > `40:00` 若某项工作的十分之九可以被自动化，但机器执行的标准低于人工水准，企业甚至根本不会选择自动化这九成；反之，若 AI 在其余环节的执行速度和质量远超人类，那么‘让人类继续负责那剩下的一成’反而会拖垮整个最终产品的速度与品质，成为整个流程的致命短板。 — Phil Trammell > `50:00` 如果一台机器人一年后能制造出100台机器人，以机器人为计价单位的实物利率高达 10,000%；但若机器人的相对市场价格暴跌，以一般消费品计价的实际利率可能并不高。 — Phil Trammell > `54:35` 以往工业巨头的财富之所以没有无限膨胀并吞噬整个经济体系，主要是因为存在‘耗散冲击’。巨头最终都会死去，而他们的后代往往是不称职的资本受托人，甚至无法维持资本以平均经济增速增长；或者他们将财富留给基金会，而基金会会逐步将其支出。 — Dwarkesh Patel > `01:07:30` 传统上给发展中国家和失业群体的建议往往十分天真；而经济学视角的硬核方案其实极其纯粹：认清现实，直接设法买入并持有绑定 AGI 增长的指数化资产。 — Dwarkesh Patel > `01:11:30` 人们很难去构想一个尚不存在的美好事物，但却对失去眼前拥有的东西极其敏感。 — Alex Imas ## Mentioned Alex Imas, Phil Trammell, Dwarkesh Patel, David Ricardo, Andrej Fredkin, Bryan Debarre, Andrew Koh, Tony Atkinson, Nicholas Kaldor, Chad Jones, Magnus Carlsen, Jane Street, Andy Hall, David Autor, The Budget Lab at Yale, Citrini, Google, Gemini Omni, Joshua Gans, Avi Goldfarb, Jonathan Haidt, David Reich, Mark Zuckerberg, Elon Musk, Nick Land, Nick Bostrom, Cursor, Composer 2.5, Sasha Rush, Narendra Modi, SK Hynix, ASML, Anthropic, OpenAI, Meta, ByteDance, Consolidated Edison, Henry George ## Video description <p>Economics of AGI episode w <a target="_blank" href="https://www.aleximas.com/">Alex Imas</a> and <a target="_blank" href="https://philiptrammell.com/">Phil Trammell</a>.</p><p>There’s a bunch of important questions about how we deal with AI that only economics can answer.</p><p>What is the optimal way to tax and redistribute the wealth that will be generated? How should countries not in the AI supply chain index into the gains? Is there any world where inequality doesn’t explode?</p><p>It might seem like these questions have obvious answers, but the first thing economics teaches you is that your intuitions can often be entirely wrong.</p><p>It was very helpful to chat through these things with Alex and Phil.</p><p>Watch on <a target="_blank" href="https://youtu.be/Jj-kBHzUohs">YouTube</a>; read the <a target="_blank" href="https://www.dwarkesh.com/p/alex-imas-phil-trammell">transcript</a>.</p><p><strong>Sponsors</strong></p><p><a target="_blank" href="https://janestreet.com/dwarkesh">Jane Street</a> invests heavily in turning smart people into exceptional researchers and engineers. In addition to their apprenticeship model, Jane Street runs lectures and bootcamps in their in-office classrooms -- managers clear their teams’ schedules to encourage attendance. If you’d like to work at a place that takes learning this seriously, Jane Street is hiring. Check out their open roles at <a target="_blank" href="https://janestreet.com/dwarkesh">janestreet.com/dwarkesh</a></p><p><a target="_blank" href="https://gemini.google">Google’s Gemini Omni</a> has incredible video editing capabilities -- you can upload a video and have Omni change the background, adjust lighting, or add specific elements. But Omni is also a preview of how future frontier models will be trained -- fully multimodal on both input and output. You can try it yourself in the Gemini app at <a target="_blank" href="https://gemini.google">gemini.google</a> or in Flow at <a target="_blank" href="https://flow.google">flow.google</a></p><p><a target="_blank" href="https://cursor.com/dwarkesh">Cursor</a> used targeted RL with textual feedback to help train their Composer 2.5 model. One of their researchers, Sasha Rush, gave me an impromptu blackboard lecture to explain how this form of on-policy self-distillation works -- I posted the full thing on X. If you want to try Composer 2.5, go to <a target="_blank" href="https://cursor.com/dwarkesh">cursor.com/dwarkesh</a></p><p>Timestamps</p><p>(00:00:00) – Will capital share increase?</p><p>(00:19:36) – Messy Middle scenario</p><p>(00:25:57) – How to tax and redistribute AI wealth</p><p>(00:30:02) – Why demand collapse is unlikely</p><p>(00:39:26) – Human employees would be hard to integrate into the machine economy</p><p>(00:43:08) – What if some humans (or AIs) value wealth accumulation intrinsically?</p><p>(01:01:28) – What should developing countries do?</p> <br/><br/>Get full access to Dwarkesh Podcast at <a href="https://www.dwarkesh.com/subscribe?utm_medium=podcast&#38;utm_campaign=CTA_4">www.dwarkesh.com/subscribe</a> [Source](https://api.substack.com/feed/podcast/200613353/da820d19eca949d5e111f641ab5e378d.mp3)

[0:00] Alex Imas: 所谓关系型部门，是指那些将‘有人类身处其中’本身视作核心价值与产品属性的商品与服务。由于人类的时间和精力天生具有不可扩缩的物理稀缺性，即使其他生产要素的边际成本骤降为零，凡是必须有人类参与的领域依然会天然保持稀缺。 [12:35] Dwarkesh Patel: 在拥有数万亿台机器人、而人类只有几十亿的未来世界里，全社会在机器人身上花费的总开销，怎么可能还不如打赏马格努斯·卡尔森、雇佣理财顾问、名医、家教或播客主播的支出？ [12:35] Phil Trammell: 每过18个月，计算的边际价值就直接减半。人类消耗算力场景的速度太快，导致对新增算力的边际需求衰减速度甚至赶上了供给增长速度。 [20:00] Alex Imas: 数据表明，失业率只要微升 2%，整个国家的政治风向就会彻底逆转。因此，最糟糕的情形或许不是突发性的断崖式崩溃，而是一种‘慢刀子割肉’的‘涓滴流失’。 [25:00] Alex Imas: 在现代市场经济中，普通人拥有自身劳动力作为初始禀赋，并能将其转化为独立的生活来源；但一旦人类劳动力在经济上失去价值，大众的基本生存将完全依附于政府定期发放的支票，人们的命运将彻底受制于当时掌握公权力的民选官员——这种权力依附结构是极其危险的。 [35:00] Dwarkesh Patel: 在技术奇点即将到来的繁荣前夜，资本所有者拒绝投资现实上是不可想象的。即便富豪们不再买游艇豪宅，他们也会将每一分利润砸向建造更多的数据中心、铺设光纤电网、建造芯片晶圆厂来运行前沿 AGI。 [40:00] Phil Trammell: 若某项工作的十分之九可以被自动化，但机器执行的标准低于人工水准，企业甚至根本不会选择自动化这九成；反之，若 AI 在其余环节的执行速度和质量远超人类，那么‘让人类继续负责那剩下的一成’反而会拖垮整个最终产品的速度与品质，成为整个流程的致命短板。 [50:00] Phil Trammell: 如果一台机器人一年后能制造出100台机器人，以机器人为计价单位的实物利率高达 10,000%；但若机器人的相对市场价格暴跌，以一般消费品计价的实际利率可能并不高。 [54:35] Dwarkesh Patel: 以往工业巨头的财富之所以没有无限膨胀并吞噬整个经济体系，主要是因为存在‘耗散冲击’。巨头最终都会死去，而他们的后代往往是不称职的资本受托人，甚至无法维持资本以平均经济增速增长；或者他们将财富留给基金会，而基金会会逐步将其支出。 [67:30] Dwarkesh Patel: 传统上给发展中国家和失业群体的建议往往十分天真；而经济学视角的硬核方案其实极其纯粹：认清现实，直接设法买入并持有绑定 AGI 增长的指数化资产。 [71:30] Alex Imas: 人们很难去构想一个尚不存在的美好事物，但却对失去眼前拥有的东西极其敏感。

[15:00] Dwarkesh Patel: The pessimistic framing of Moore's law is every 18 months, the value of computation halves... we're just running out of uses for computation so fast that it's sustaining Moore's law. [15:00] Dwarkesh Patel: With AI, an H100 costs more to rent now than it did three years ago... because as models get smarter, the opportunity cost of compute gets higher. [17:00] Alex Imas: In the single art print condition, human-made art is valued much higher than AI. But with 500 prints, the human price drops because it's no longer a connection to the artist—versus AI, where 1 or 500 makes no difference. AI is already viewed as a commodity. [19:36] Alex Imas: A 2% increase in unemployment completely changes political winds... a fast uptick looks like a national emergency, but a slow drip leaves people getting reabsorbed at lower salaries. [35:00] Alex Imas: Here, the technological frontier is expanding. You actually have abundance, and for abundance to generate negative economic growth—that's really hard to get. [45:00] Phil Trammell: At some point... people are just going to see an AI therapist as a superior product, and they're not going to need the sort of empathy that the human is providing. [45:00] Alex Imas: If we're thinking about natural selection... even if you get some sort of indifference now, you might get selection to point into an even stronger preference for other humans. [45:00] Dwarkesh Patel: Most of his wealth is just stock in Meta... he rather would have his wealth compound and Meta build more data centers basically. [50:00] Phil Trammell: The capital stock could grow quickly, but the price of capital goods relative to consumption goods could be falling faster than the capital stock could grow. [50:00] Phil Trammell: In units of robots, the interest rate is 10,000%... but if the price of robots is falling really fast, prices adjust. [50:00] Phil Trammell: Every unit of capital next year is giving up way less consumption than each unit of capital this year, because one robot now turns into many robots next year. [55:00] Alex Imas: A lot of the time, historically, the wealth accumulation process is part of a large social interaction amongst peers... where you want to be admired. [55:00] Dwarkesh Patel: Historically, they either hand it off to their children who are worse stewards... or put it in foundations which spent it. [55:00] Phil Trammell: As a good classical utilitarian, to me the value... is to create new happy beings... putting Dyson spheres around stars and turning all the energy into really happy simulations. [68:53] Dwarkesh Patel: The more you think AGI is going to be... run on AGI the way our economy currently runs on electricity... the more it looks like electricity and the more it's like every company in the S&P of the future. [71:00] Alex Imas: It's very easy for somebody to go on a podcast and to say like 'these jobs that you like, they're going away,' than to somebody to spin up like a utopia which doesn't exist yet.

[0:00] Today, I'm chatting with Alex Emas, who is director of AGI economics at Google DeepMind and professor of economics at University of Chicago, and Phil Trammell, who is head of economics at EPOC and research scholar at Stanford. In general, in this interview, what I want to understand is what economics tells us about what we can expect in a world with more and more automation, more and more advanced AI, what that tells us about what will happen to wages, to labor share, what the best way to tax and redistribute the wealth that will be generated as a result of AGI will be, and what kinds of things will be scarce, because what is scarce kind of tells you where the value will accrue. So I want to start there. What are some plausible candidates of what will be scarce? Something like the relational sector, which is what I defined as, you know, basically services and goods, where the fact that the human was in the loop was actually part of the value of that product. So because humans are naturally scarce, if we have automation where a lot of other things stop being scarce, we will still have scarcity and things that humans are kind of involved in and in the loop for. I'm curious to understand whether humans doing services for other humans can never be a big part of the economy And here maybe one intuition pump So in a world where AI can physically do anything humans can do there's this whole machine economy where they're building factories and doing research and coming up with new ideas. And humans may or may not be involved in the physical production of those things, but probably not, given that in the ultimate limit, if robotics is solved, if you don't care about humans being involved in that process, why would humans be involved in that process? but then there's these other things which you point out where well we actually maybe in some cases do want the ballerina or the barista or whatever to be a human that's part of the value of going to a cafe or a performance but only humans have that preference so there's this human economy where like humans are doing services for each other and part of their wealth is flowing to other humans but part of their wealth is also like they will want some of the automated goods it's like machine-only economy is creating and so part of that wealth is flowing out and so if you just think of this as like this is not a closed loop but a lot of things in the machine-only economy are closed because the machines don't care about like getting the human barista to make them a coffee and so within that model isn't isn't it intrinsic that like the human-only economy will become a smaller and smaller share i would like to pitch kind of a rephrasing of that question so i think my view is that kind of forecasts that economists like us would make are not necessarily as individual forecasts like me and Phil are talking right now, are not necessarily very useful. The reason I think that, so there was this blog post by Andre Fredkin, Brian DeBerry, and then Andrew Coe that came out yesterday, actually, that looked at like kind of people's forecasts, economists' forecasts about the labor market. And what they found is that there's a ton of disagreement, like in every single direction. So what they advocate for, and I think I'm in agreement here, is rather than thinking about individual forecasts like what me and Phil are going to do, rather than looking at kind of like basically generating prediction markets where you get aggregate forecasts, where you get like kind of wisdom of the crowd effects. And kind of the reason that I think this is because we have been famously terrible at forecasting. And so let's go all the way back to 1820. This sort of debate that we've been having actually is like 200 years old. So David Ricardo is one of the classic economists, not neoclassical, classical economists. And he, when industrial revolutions started happening, he wrote a bunch of stuff saying like, look, this is going to be great for everybody. Prices are going to come down. But then he turned around and he like wait I can actually see all of these jobs that are creating value They going to be automated by these machines This is going to be really bad Everybody going to become unemployed and there going to be political unrest and things like that And if you look at Ricardo predictions, they're actually right. If you look at all those jobs that made money in Ricardo's time, they got automated. So if I was David Ricardo and I woke up and somebody told me all those jobs did get automated. And you asked me, Dave Ricardo, like, what do you think the prime age employment rate is in 2026? I think he would be surprised if you told him it was the highest that's ever been other than 2000. We have the highest number of employed people that could potentially be employed since 2000. That was like the peak. And now it's like the second peak, basically. So what David Ricardo ended up missing is the fact that, you know, essentially, you have these economics of structural change where basically everything that got automated became cheap. People had more money to spend on things. And then they started spending money on services. And this is kind of like the lump of labor fallacy. That's what they call it. Ricardo didn't think, hey, I should have considered the fact that new jobs would be created. But it's kind of not obvious that money would go to services. Why wouldn't they go to more automated goods or something like that? And I'm not [5:00] Saying that, like, I'm not using this anecdote as to say, like, this is what's going to happen now. We're going to have full employment. I'm using that anecdote as to say it's really hard to make predictions. And what I think may be a really useful tool that economists have is instead start with a premise. Like, maybe we'll start it today. Look, labor share is zero. Like, labor share has gone down. What could possibly explain this? Let's write down an economic model of what happened. Phil will talk about this later today. Or you can start to write down a model to say, hey, what if labor share just stays the same? What can make that happen? And here's my main – if you don't take anything out of this conversation from me, we don't have any data. I've been kind of saying we need a Manhattan Project for data. We don't have data on basically consumer demand elasticity. We don't know what they are. We don't know – we're not really tracking what jobs are getting created or destroyed. like the ONET database with all of the tasks and different jobs that's been rarely updated. It's super low quality. And so what I think is really useful is to think about like what are the potential scenarios? And we'll be talking about a lot of these scenarios, mapping them out and to say what dimension of scarcity will generate that scenario So if there full employment we could talk about the relational sector or something like that If there very labor share collapses we can talk about other sorts of scenarios And then that will tell us what data we should be collecting. It's probably worth the defining labor share and capital share real quick. So the whole economy, like the total sum of goods and services sold is either paid out to people in wages or it's paid out to capital, which is to say that there's like rents on buildings and then there's shareholders of companies that get paid out. And for many hundreds of years in the economy, 60 something percent of the economy or all the things that are sold in a given year basically gets paid out to humans and wages. And the other 30, 40 percent gets paid out to people who own machines and land and claims on companies and whatever. And the question is, well, right now, 60 percent is going to wages. Does that shrink as automation or as EIs get smarter and smarter and better and better? And it like it really this is a call door fact like right So it incredibly we should stress this It incredibly surprising that it over 60 percent after the industrial revolution after all of the automation we ever seen The fact that it almost like some people are worried it an accounting error or something like that that it kept being been so constant And the fact that it's like been over 60 percent. And, you know, there's there's even a controversy right now. So some might say, like, you know, labor share has been falling in the last 20, 30 years. But, you know, depending on how you there's been a lot of accounting changes in the last 30, 40 years. So, for example, Andy Atkinson has this paper showing that actually if you keep the accounting constant over the years, labor share hasn't even fallen ever. But it's not that surprising, right? I mean, Phil, you made this point that if labor and capital are complements, you need both to do anything. It kind of makes sense that you kind of need to pay both of them to get something done. You have had stuff can be completely automated. Although you had the post where you're pointing out that actually. Oh, yeah. Yeah, well, I was going to say there's a sense in which nothing's yet been completely automated. If you look at the network adjusted factor shares of a good, which is to say you look down the supply chain and say, not just like the final step, how much of that is done by capital and labor, but what went into the machines that can automate that final step? You'll find that labor is adding a lot of value down the supply chain. So like you know computer and electronic products in the US have a very stable capital share network adjusted capital share of around 50 percent I do think there this qualitative shift that we I think we agree is coming which is that there will be at least some goods whose network adjusted capital share goes to one Right. Because the whole supply chain can be automated and there's no part in it that we care intrinsically about having a human do. um so that'll be a you know that'll be a qualitative shift interestingly the implications of that shift for the overall capital share are ambiguous because if we let's say that we've got the two sectors the human intrinsic sector with the ballerinas and everything else right right now everything else has been scarce because of the lack of labor in it right but if we fully automate the supply chains for everything else right and we satiate and everything else really fast then the quantity of everything that's not a ballerina say it goes to infinity but are the marginal utility and that stuff goes to zero faster than the quantities rising yeah i also kind of want to move if you don't mind move away from the ballerina example because i think like the kind of point that i was trying to make in my post uh again and the point of the post was like to work backwards from a particular scenario was that [10:00] kind of the ballerina and the kind of performer, that's the wrong reference class. Right now, we have a lot of jobs where you have different tasks. So this is the task-based model of jobs where you have like a lot of different tasks. So like a doctor, what is their job? They're filling out insurance documents. They're, you know, going and like calling different pharmaceutical companies. And one of their tasks is to actually see the patient and talk to them. But that's like actually not the main part of the job. So you could have a job and a service or a good be a product of different types of tasks, and you can automate a ton of those tasks. And if the consumer is willing to pay more for a product or service where every single task is automated versus every single thing except for that one part where the doctor is actually delivering the diagnosis, providing support and things like that, we would call that job a part of the relational sector, right? Because a human is, people are willing to pay more for the human to stay in the loop in the job, right? So I think we don't have data to say, like here are relational jobs, here are not. Because you literally need to collect data of the following sort Do a conjoint analysis of like here my willingness to pay for this service this good Here the counterfactual where everything is pursued to spy machine Here's the counterfactual where this one task is not produced. What is your willingness to pay? What is your elasticity for that, for the human to not be in the loop? And like, literally, if I don't have that data, what prediction am I going to make in this story? Right? Right. But I guess, isn't there another point, which is that there's a lot of fully automated goods that don't even exist yet. And you can't collect any data right now about, say, how much people will keep buying more and more of some drug that makes you healthier. Absolutely. That's fully produced by DAIs. And that's kind of Phil's point. And you can make it is that, look, you know, you could have an increase in variety in capital where you don't get the satiation, right? So you're increasing variety. So you're not hitting that really diminishing marginal utility point where you're, you know, you're basically most of your income is going to the human sector. If that increasing variety is fast enough and there is no such increasing variety in the human sector, then you can get all of the relational that you want, but it doesn't matter for labor share. It goes to zero. Phil I liked your analogy to some Mongolian economists sitting around 1400 thinking about what will be scarce and the limits of that kind of analysis I think you should talk to that Sure. Yeah. So if you just looked at the goods available to, you know, a Mongolian of the distant past, no expert on this society, but I know that they didn't have nearly the variety that we have now. and um they looked at the jobs that were sort of intrinsically human like being being a singer say um and they looked at the the things that were not intrinsically human like um you know the transportation services provided by their horses or the different kinds of food they had if they just kind of held the varieties fixed on in both categories and asked um what will happen And once we have a lot more automation, they might have said, well, we'll just satiate in, you know, force like transportation and in yogurt and in yurts. Those shares will all go to zero and we'll be left spending all of our money on on singers. But of course that not what happened because as we accumulated more wealth and you know more advanced machines and so on we expanded the range of things other than singers to spend our money on and the share spent on singers had stayed sort of negligible um so likewise that sort of my central prediction about how the future unfolds though it could go either way i was gonna make a point and i realize it's a fallacy but the reason it's a fallacy is interesting so i was gonna say um i mean it's just hard to imagine a world where there's trillions upon trillions of robots but there's only some billion odd humans and then like the cumulative amount we're spending on robots and like building more robots and whatever is less than what we're spending to like pay you know magnus carlson and or financial advisors or doctors or tutors or podcasters or podcasters but then i realized that's a fallacy um the number of transistors in the world has like literally certainly trillion x maybe quadrillion x or something and your colleague chad jones has a very interesting result about how the share of the economy that is going towards paying for computing basically like paying for the transistors um has been decreasing the point that you made is that one way to think about moore's law you know what sets price well the prices of supply and demand [15:00] And so not only are we producing more transistors more cheaply, but also we're like the value of the marginal transistor is decreasing. Right. So more, as you were saying, another way of saying Moore's law is you should say. I like the pessimistic framing of Moore's law is every 18 months, the value of computation halves. Yeah. Right. Like we're just running out of uses for computation so fast that it's sustaining Moore's law. And this is, in fact, like literally relevant to a conversation about AI where maybe for the first time, this is no longer true. Right. So the famous fact here is that an H100 costs more to rent now than it did three years ago, even though we have much superior technology and we have much more compute in the world. Because as models get smarter, the opportunity cost of compute gets higher. This is Phil's point about increasing variety. Right. What we have done is increased, increase the types of things that people demand from capital. Now, all of a sudden you have a new variety that you could be using capital for. And all of a sudden you, you, you jump back up. Yeah. You could imagine we just never satiate demand for compute And as long as that stays the case then the share of the economy that is going towards compute would keep increasing And that the big question right It like that is the ultimate question that we need to be kind of looking at is like, what number of new uses are we finding for that commute, where you have the demand for these uses? So what I kind of want to emphasize is that a lot of models in economics, especially in the space that we're talking about, take demand is almost kind of exogenous and they don't unpack like what is that like the psychology of what people actually want and so what got me kind of also thinking about this the idea of the relational sector's work that i was doing on the fact that there does seem to be this value this intrinsic value that is it's not just because it's scarce it's because there's some intrinsic preference that people have for like empathy and connection and you know getting getting uh interacting with another person. So like one of the experiments that we ran was like, there's an art print, right? And we actually have an incentive compatible way of like, basically saying like, how much are you willing to pay for this art print? People actually paying real money for it. And then we say like, look, there's only one of those art prints. And it's either made and these are between subject conditions by AI or by a person So with one you get the effect that the person produced art print is valued much much higher than the AI version And then what we do is to say there in a set of other conditions, there's 500 of these being produced. So for the human made one, the price goes down a lot, because it's no longer seen as like, you're not like making a connection with this one artist, versus with AI, there's no difference. AI is already viewed as like a commodity. And we need to do a lot more research on this, but it seems like that's kind of like the key difference between something like, let's say, a horse. A horse was an input into an output where you can replace the horse with something else. You only care about the output. The only way this relational story works, and this is what we need more data on, is if a human is not a horse in the sense that it is providing value from the output, where if you replace the human, the value of the output decreases. And if that's not strong enough, and if it doesn't hold for enough sectors, if it doesn't hold for enough jobs, then this kind of story doesn't work anymore. There aren't that many institutions that have thought as hard as Jane Street about how to turn smart people into some of the most competent researchers and engineers in the world. This relies in part on an apprenticeship model where new hires are paired with senior mentors But Jane Street also runs a bunch of classroom lectures and hands boot camps These courses cover a range of topics and they go pretty deep There's one lecture that focuses on reverse engineering systems with tools like S-Trace and GDB, and another that teaches you how to profile code down to the cache hierarchy level. Importantly, Jane Street designs these courses not just to teach the relevant object-level skills, but also to impart the relevant tacit knowledge. For example, their week-long neural net boot camp starts with general theory, but then quickly progresses to how to apply neural networks to trading. And here they cover the specific obstacles that Jane Streeters tend to encounter and the workarounds they've come up with to get around them. Jane Street takes this sort of learning incredibly seriously. Every office has dedicated classroom space and courses are prioritized as part of regular work. If you'd like to work at a place like this, Jane Street is hiring. You can check out their open roles at janestreet.com slash dorkesh. There's one possibility which Molly has written something about this messy middle scenario. And that possibility made me think about whether it might be better to have, at least as far as wealth distribution and redistribution goes, it might be better to have much faster AI takeoff. And I want to ask you whether the following possibility is [20:00] at all likely or there's any set of assumptions that this can make it so which is that ai makes it possible to automate jobs such that like many people are losing their jobs but it doesn't create enough wealth while the process of automation is happening to pay off basically the people who are getting laid off so there's like a Pareto improvement um everybody's getting better as a result of ai automation and of course there's a trivial sense in which that must be true because whatever money you're saving whatever money the company is saving by not paying the humans instead of just paying the AIs. Those resources still exist in the economy and they can just be paid off to people. But there's going to be some allocative inefficiency. The government doesn't know exactly who got laid off because of AI. There's some political problem of if the meta worker gets laid off first and they're making 200K a year. Is there a politically sustainable situation where you give them a 200K check a year when there's many people who are working who are making much less? um so do you at all find this scenario plausible where ai is actually automating a bunch of things but there isn't enough wealth creation as there is automation uh i think it's is that plausible possible to me it does seem like a pretty narrow window my guess is that if we have the technology to automate so many jobs that it becomes like a new kind of political problem then the pie will also be growing really fast Well unless in all of those professions that it automating it just a hair more productive So like the cost of all the capital to replace all the software engineers is just, you know, a hair less than the cost of what we've been paying the software engineers. And why is it implausible that it just like a company can save money by laying off a bunch of software engineers? But and in the long run, there's a Jevin's paradox thing. And, you know, we can't anticipate in advance what we do with more software. And surely there's going to be more uses. But in the short run, the fact is just that a lot of people are laid off and they still need to figure out how they can use a million X more JavaScript tokens. I think the thing that like is in either like, you know, Phil and I have been like writing about these things and we have mathematical models in the back of these things. We don't have any political economy in any of our models. Andy Hall wrote a really nice blog post about the politics of AGI, and he made a really interesting observation. If there's a 2% increase in unemployment, the political winds completely change. Like unemployment, it has a huge effect on what happens politically. So you know to Molly excellent essay by the way I think in some ways like one of the worst scenarios is a drip scenario because of the political economy piece right Because like you know people essentially what you what you might see is like people not really being unemployed in mass, but kind of like moving into sectors that pay them less money, kind of basically getting what happened with phone operators in, in, in the mid century of the, of the, between 1920 and 1940 phone operators were completely automated, right? But it took 20 years, even though the technology existed. And therefore, there was this drip. It wasn't like this giant sector just disappeared. And what ended up happening, there's a really nice QGE paper on this, basically showing that they got reabsorbed into the economy, but at lower salaries, and they were mostly underemployed. And I think that's the scenario that Molly was writing about this kind of messy middle where things aren't a disaster because we saw with COVID, the fiscal response can move quickly if there's an emergency. And an emergency is a quick uptick in unemployment, which could even look like 2% or 3%. That becomes a national emergency if it becomes fast. The concern is That suppose Whatever you saving on those white collar workers If that not growing the economy But it just creating some you know saved resources that can be allocated elsewhere Is that enough to do a broad based redistribution scheme Because then you have like the money you've saved off a couple of people. Yeah. And unless you can figure out exactly how to get to them specifically, you have the problem of can I do like can I do a UBI off the money I saved? So you're basically saying like, look, the pie did not grow that much. Yeah. You're just basically you're just basically displacing a bunch of people. But that actually didn't grow the technological frontier of what the economy can produce. And so then there's a question of like, well, maybe every time I don't know if this is the case, maybe every time this has happened in history, the technological frontier has expanded a bunch. And so I think that's the case. I think simply in history, the technological frontier has expanded. Yeah. So it's it's it's it's kind of I think Philip made the same point. like it's hard to imagine that sort of scenario um where you are getting like intelligence that's kind of just enough to replace the software engineer but still costs a lot of money like it's just a hair less less less expensive than the software engineer so you're not getting this abundance effect right and so you're what where is the redistribution going to happen because the pie didn't grow yeah yeah okay so the this is very helpful so there's a many different things that would be true for the [25:00] to come to pass, each of which seem unlikely. One, it has to be the case that it is possible to automate entire white-collar jobs, but only in a piecemeal way. That is to say that you can only automate software engineers, but that same program can't also automate an accountant and an analyst and whatever. Where I think, at least my model of intelligence is such that both of the breadth of tasks that require us to do something like software engineering and what intelligence is, is such that, you know, if you can really just like lay off all the software engineers, you've got enough in the bucket there that you could like automate all kinds of white collar work. So yeah, you're saving. There's huge amounts of potential savings that have happened as a result of these layoffs. And also that AI is going to be cheaper than human labor. And if both of those things are true, this messy middle scenario where we literally don't have the wealth to go around seems unlikely. And the question is, what is the best way to tax it and redistribute it? Yeah, I have some thoughts. I think it's just really important to outline the costs and benefits. It's also important to know that there's, so first, there's differential complexity in implementing these things. Two they differ in the timeline of being actually helpful So something like universal not like that not going to generate returns for something that happens in six months So you probably are going to end up with a layer of things So like for example, like a negative income tax, like you implement it. And if the day it turns into law, that is already you already have this sort of insurance that like, you know, there's a floor for which everybody gets a certain amount of money. And then if you earn more money, you get taxed more and things like that. But there's positives and negatives to negative income tax. With UBI, for example, I worry a lot about the political economy implications. For example, if people are just kind of dependent on a check, it really matters who's in power. Like right now we're endowed with labor that can turn into income when that is no longer the case. And we are now at the mercy of the elected official for like basic needs. Right. So that to me feels like a power sharing arrangement that's really dangerous. But wouldn't that be true of any sort of government redistribution program? So something like University Basic Capital where you have like an ownership share and you have property rights for capital then you just you just you just you a normal shareholder You just a normal person And but this goes back to the question of indexing because if indexing is hard then universal basic capital is hard. And that's the, that, that's the problem of universal basic capital is targeting. Right. Right. What do you target to put into people's portfolios? Like what if Anthropoc goes to zero, but some random robotics company takes all the supplies. Exactly. So that's the risk of universal basic capital with a negative income tax. you have the same sort of issues with UBI, where somebody comes into power and says, we're not going to do that anymore, and people can't work. And then you have the issue of the floor being open. One concern with the wealth tax is that there's no politically sustainable equilibrium at 0.5% wealth tax. I mean, this happened with the income tax, of course, right? It starts low, it's like for war or something, and then it slowly and slowly escalates until the marginal tax rate in the U.S. is probably on the order of income tax rate is like 40 percent or something. And in certain states, upwards of 50 percent. With a capital tax, is there a reason to worry? Would that distort investment? Because people would just be like, why would I invest in Anthropic or Intel? The government is going to take larger and larger shares of it and dilute my share Well hold on So I think it worth separating like how the revenue is raised like what tax and then how it distributed it could be that the government hands out shares of anthropic to everyone by broad-based tax and then buying anthropic yeah okay um which would probably be the right thing to do i mean hopefully some like populist proposal doesn't interfere with that and like expropriate some like particular company that everyone happens to know about yeah um but So you're suggesting there could be a tax that is some sort of optimal tax. We're taxing externalities or we're taxing land or we're... I guess we probably need to tax something other than just those two things. But that tax... Or consumption. Yeah. Or consumption. Okay. So a consumption tax, like a European value-added tax type thing, that allows the government to go buy a bunch of stocks and then it just distributes those stocks to everybody. That's David Otter's. Yeah. Yeah, I mean, that's not going to be that different from just like redistributing the stocks, but it'll be a little different. Yeah, that's what social security. That was the proposal for social security, by the way. That was privatizing social security. Right. So it's like you turn like this sort of weird, like not weird, but it's been working. It's worked so far. But, you know, there's questions for how long it's going to keep working. Like basically privatizing social security was giving everybody. [30:00] basket of stocks right all right i'm curious to understand people talk about whether there's a white collar apocalypse already is there any evidence uh that suggests that there is mass automation or unemployment as a result of ai already um i think there's a lot of people are looking at it so this is an area where there's like a lot of eyes and a lot of data being produced uh so the budget lab over at yale is doing really good analysis on this they just recently released to report and i think like you really have to squint to see anything happening like basically if you want to take kind of like uh an approach across the entire economy and looking at even looking at like software engineering like the most exposed sort of sectors there's just like not really anything going on there might be a little bit of a signal about like junior developers getting jobs less than before but that's like a less than before rather than a level shift is then there's actually an increased demand for senior manager, for senior software engineers, if anything. And so if you look at trend, it's kind of like for junior managers, it's a bit below trend. So as in you saying the growth is slower than before Yes But there is still growth even on entry software engineers Yeah exactly And what do you think is going on with the anecdotal evidence of graduating college students saying that they finding it harder to find CS jobs or something? I think that's anecdotal evidence. You think it's always been hard to get jobs for some people, and now it's getting turned into an AI narrative. Same with the layoffs, where it's probably just a normal layoff, and they turned it into an AI layoff. Yeah, I mean, you have to be careful with all of this. I think like there are these like, you know, there are these like coordinate public coordination devices for like, let's say we get into a narrative where like if you're a firm and you're not laying people off, then you're seen as like not adopting AI enough. So like then you're going to just get a cascade effect. Right. A firm's like just needing to keep up with the Joneses in terms of like starting to lay people off. And that's kind of like that. That's super worrying. We're like, actually, the firm might be actually worse off after the layoffs than before the layoffs. But it's just doing the layoffs to have the perception that, look, look, we're not behind the times where we're we're, you know, using AI like you have the you probably heard these anecdotal stories of like these token counters that like you have to maximize tokens and things like that. So, again, like right now we have we don't really have any evidence of a white collar bloodbath. And is that surprising at all? I feel given the fact all these things they can do is just like this is the story is all this time If you automate some complementary task the overall bucket of things that the human labor which complements the automation Yeah. Will increase in value. So this is this is one of the statistics that's really important for that argument is elasticity of demand. Yeah. So like the you take the O-ring model of jobs. So like, again, jobs is a series of tasks. Let's say the AI automates like nine out of 10 tasks. One task is not automated. If that person can now kind of focus in on that task, the job will become more productive. If that translates into a price effect where the product is actually cheaper, if the demand responds enough where it's being bought more, it's being used more, the service is being used more, that could actually lead to more hiring. Right. And a lot of people on the Internet have been like kind of making that argument kind of very generally saying, like, look, we're seeing if anything in the data, we're seeing an uptick in software engineering. Right. Yeah. Which suggests that at least for now, given the way the jobs work, it might be. But I think this elasticity of demand argument is incredibly important both for a lot of arguments that people make or just a lot of labels that people use without understanding what the underlying causation is So people often talk about Jevons Paradox Yeah This is this idea that as something gets cheaper you will want so much more of it that the total amount you spend on the thing increases And so famously, this happened to coal in Britain 200 odd years ago. But really, this only happens if the demand for something is highly elastic. There's many things for which there is not super elastic demand. if oil for example gets super cheap it's not like magically right yeah exactly uh magically there's going to be so many more cars that now we're going to be using way more oil than before it's not in the short run exactly so the long run elasticity is higher than short run elasticity but even in the long run so agriculture famously is the example where um we can produce way more food if we dedicated the same portion of the economy that we dedicated to agriculture we're already producing more food regardless but we could produce even more food if the same portion of the economy that was producing food 100 years ago, was currently producing food. But, you know, you eat enough and then you're done. And so the claim with software is that it is not some inherent property of markets that as it gets cheaper, you'll just keep wanting more of it. Absolutely not. It is. The claim about software is this is a particular kind of good, whereas it gets cheaper, we'll want more and more of it. [35:00] It is also highly relevant, and you wrote an essay about this. A lot of this podcast is me summarizing your essays back to you. That there's this very viral scenario planning about the future by Citrini, where they're predicting as a result of automation, as a result of very powerful AI, there will be a recession because white collar workers will get automated. Their salaries, which were, you know, paying for a bunch of things, will no longer be available. And so there will be a slump. Do you want to recapitulate? why this might be implausible? Well, I mean, so part of it is plausible, part of it's not plausible. So like the part that's kind of like within the this is something that we started the conversation with is the idea that there could be unemployment, a lot of unemployment is if the speed of automation is quick and things like that, people could get laid off and they may not find work very, very quickly. So that part of the Centrini essay about the unemployment, you know, we can quibble about that but that's that's not the issue the issue is that they talked about negative economic growth right and so what i did in the in the piece that actually phil and i had a back and forth on was to say like let's start with the with the proposition that there's negative economic growth what conditions do you need on the economy to get negative economic growth and it turns out the conditions are pretty improbable So one thing that you need is like for the holders of capital like rich people basically like basically what you have in this in those sorts of scenarios, like you have a reallocation of wealth and income from like lower income people who are working or using their label towards capital owners. So what you need is that basically demand to be bounded, like a hard bound, not even like a soft sort of like diminishing sensitivity. You need for them to eventually say, I've had enough. I don't want to spend any more money. And for that money to not enter as investment. Right. Right. Which is and then you can get negative growth, which is like. And the crucial thing is, even if we don't want more shit, the world in which there's a singularity and we don't want to invest more money is crazy. Right. where we're not like, let's build more data centers, let's build more fabs. Even though we have AGI, we're not like investing in more data centers to run the AGI. And that's like driving more economic growth. And so I sent the essay to Phil and Phil actually wrote back being like, this is pretty dumb. Like my essay saying, you're trying to say that there's going to be negative economic growth, but these are very implausible conditions. And I was like actually that the point of the essay These are very implausible economic conditions So that where I think like scenario planning really shines is you have the Centrini essay which I think is like I think it was great that it written because I kind of started a conversation. But it's just like, it's so intuitive, this idea that like, look, if there's demand collapse, we can get the economy to shrink. But it's actually, you could get that with a depression, right? Where in the depression, the technological frontier didn't expand. Right. Here, the technological frontier is expanding. You actually have abundance and for abundance to generate negative economic growth. That's really hard to get. Right, exactly. Google recently announced Gemini Omni and its video editing capabilities are incredible. You can upload a video and then tell Omni to do things like change the background or adjust the lighting or add or remove elements. All while keeping everything else consistent. But Omni isn't just a video editor. I got a chance to sit down with the research and product team behind Omni, and I learned that it's a preview of how future Frontier models will be trained. It can take in any kind of input, whether that's text, audio, or video. And while it doesn't currently do so, architecturally, it's capable of just as seamlessly outputting images or text. So it really a bet on the multimodal data transfer hypothesis The model becomes better at predicting one data type by seeing the others For example Omni is really good at accurately rendering text on video even though Google didn specifically target that capability in this model. And Omni is the next step towards more accurate world models. Because in order to predict the next frame of a video, you have to have a deep understanding of physics and spatial dynamics. As Omni progresses, it'll be interesting to see whether it can close a sim-to-real gap. Because it's much harder to collect data in the real world than it is in simulation, robotics progress has lagged other applications of AI. But if you have really good video models that can simulate reality, maybe that stops being the case. In the meantime, if you want to try Omni, you can check it out in the Gemini app at gemini.google or use it in Google's AI creative studio, Flow, at flow.google. We're talking a second ago about why there isn't more automation as a result of LLMs. And one plausible mechanism could be that, as you're saying with the O-ring, So, O-Ring theory refers to this fact that the Challenger shuttle blew up because there's one component that malfunctioned and it destroyed the whole thing. And maybe that's a more general model of how goods are produced in the economy, that you've got to make sure everything is reliable and works well. And you can't automate an entire job to an AI right now, even though it might be able to perform it at some probability. You need extreme reliability in order for it to not destroy. [40:00] finish good um i think this is yeah so this might explain why there's less automation now there otherwise could be but i think it works in the other direction once ais get advanced enough that integrating humans into the production flow of future goods even beyond the even beyond the arguments about how humans will be more expensive or dumber or whatever even beyond that just there will be whole production flows that are uh organized for ai labor where they're talking and neural ease. They're thinking many thousands of times faster. So even if there's some comparative advantage where it makes sense to hire a human, there will be like transaction costs and worries about reliability that will actually make it hard to integrate humans into future production flows. Yeah, that seems right to me. In particular, I just want to distinguish between the point that if you automate like nine tenths of a job, then people might kind of shift over to the last tenth, but like there might be 10 times more work demanded of them from the model of O-ring automation from like Gans and Goldfarb recently, which was that if you can only automate nine tenths of the job but you can do it to a lower standard of quality than the human could do it, you might not want to automate even those nine tenths And that the thing that could totally port over to um uh like symmetrically it could be a reason why we don use a human for one part of the job anymore because a human just can't perform it to the level of quality that the AI can perform the other parts of the job or the level of speed or whatever. And they end up pulling down the, uh, quality or speed of the finished product. By the way, the model you're talking about seems extremely plausible to me of why more lawyers or accountants or whatever are not automated. Like there are cases or even software engineers where there's a pretty good probability that the thing worked as you expect. But the thing you're paying a lawyer for is like, no, really, my company is not going to go under because you're also paying for a lot of like regulation stuff. Right. So like with lawyers, particularly, you need some entity to back up the product. You need kind of like an ownership of the product you need somebody to be able to fire or hire like licensing issues there's a lot of like sort of like regulatory layers that are like also going to be keeping even if there's no relational element human in the loop that have nothing to do with like the ability of the human to actually perform the service yeah yeah i mean And you know all of these frictions on the political type decisions that we are accustomed to only trusting humans you know only having humans for like legislation and being a judge being a jury or all the licensing that keeps certain professions human That all strikes me as transitional. Right. I mean, what we expect to come from a human and like how we organize our politics, that's changed so many times throughout history. right from little hunter-gatherer bands to empires to whatnot and yeah once an ai-run political system is much more efficient than the alternatives then those will probably tend to outcompete the others and you know so speaking of which we've been talking about what preferences humans currently have and what impact that has on what kinds of goods will be scarce in the future but of course we'll have different kinds of entities in the future uh ais right there's a time when there were no humans on earth but evolution selected for agents that have specific drives and preferences because those tend to survive the most and those preferences now basically determine how a hundred trillion dollar world economy uh what it produces and so why not expect the same thing of ais in the future this is not this is not even a world with um catastrophic misalignment That is to say they just kill everybody But there will be evolution of even if not individual AIs, then firms which have AIs as part of them. And what will that evolution favor? What will favor probably firms or agents that grow, right? There's like a selection argument that things which grow will be more prevalent. And maybe just based on that, you can make some predictions about what their preferences will be. But is the kind of entity which prefers to have human intrinsic goods going to be the kind of entity that accumulates resources the most? Probably not, right? Probably saves more. It has unsatisfiable demand for things like whatever the relevant resource happens to be. Compute is an obvious one. And can we use that to make some prediction about the non-human preferences that will be guiding the future? Yeah. So I think if there's like an AI that's like has its own welfare and it's fully autonomous and it's like making its own decisions that are welfare relevant, to be honest, I have absolutely no prior that they would like at all prefer other – to like deal with humans. There's like no reason. But let me take like the other side of that argument. Will humans' preferences – [45:00] interacting with one another and to trust and empathize and all of these sorts of like things with other humans versus a simulated AI. I think it's a really important question whether those will change. Right. So I've heard a lot of arguments saying like, look, you know, right now we're just not used to the technology. And at some point, like what you're thinking of relational or something like that, people are just going to see like an AI therapist as a superior product. And they're not going to need the sort of like empathy or whatever that the human is providing. I think this is actually a really complicated question. Here's one argument for why it's not going to go away and that that has to do with evolution. So let's say there's two types of people. One person doesn't really have this preference. They can just interact with other AI, whatever can simulate it better. The other one has almost like a moral emotion, like from using Jonathan Haidt's framework, moral emotion against offloading those sorts of social interactions to an AI. Which of those two people are going to reproduce, find a mate, all of these sorts of things? I think the answer is kind of clear, right? It the second one that has the preference for other people It depends on how the reproduction is happening Fair But know the world where like um reproduction is still happening the way that it happening i think and this is a big question i not even like i not making a prediction again i just saying like if we're thinking you know you had david reich on the show like his point on the last podcast was that you know we're buzzing with natural selection right so even if like you get some sort of indifference now you might get selection to point into like an even stronger preference for other humans here's one way to think about it how is the wealth of the richest people in the world instantiated of course they can um as you were we were having a call earlier and making the point that their consumption is more geared towards relational goods like mark zuckerberg is hiring mma instructors and dancers for his wife's birthday and so forth but most of his wealth is just stock and meta and he as a controlling shareholder could say hey meta just give me all this income or turn all this wealth into dividend income um and i will just spend that on consumption but instead he rather would have his wealth compound um and met out of build more data centers basically so you don't even have to change humans for this to be the case it is just the case that the humans which are wealthiest and are growing wealthier because their wealth is compounding just have this almost Nick Landian preference for accelerating capital And that does seem to suggest that, yeah, is that an important determinant of what kinds of things are produced in the future? Yeah, I could kind of just say, like, there's two ways you could get the two kinds of people, one of whom prefers a human therapist and one of whom is fine interacting with the AI. If they both satiate equally quickly in capital, right? but the one who likes the human therapist just also likes having some human intrinsic services then the marginal value like how the marginal value of capital in the future for compares to the marginal value of capital today for each of them if they start out equally rich should be basically the same i mean there could be interactions and whatnot but basically that should be the same if what's driving the difference is that one person just doesn't satiate in capital because they're engaged by the prospect of exploring the universe and turning their head into a galaxy brain or whatever, and the other one satiates, then the person who doesn't satiate in capital is going to have, if they're being rational, they're going to have a higher savings rate So in the long run they going to have most of the well and the overall capital share will basically be the capital share of that person spending which is going to be what it important that this is we not talking about a hypothetical future yeah elon musk is talking about mass drivers on the moon right and he's like by far the wealthiest person in the world i mean obviously currently his his investments are going towards humans as well as machines but i don't think he cares particularly that his like future researchers and engineers are humans versus the and he manages to reproduce fast as well so yes so anyway so i just think it's worth drawing that distinction yeah there are currently some rich people that don't seem to um satiate quickly in capital and uh and so maybe in the long run they'll save the most and right yeah um that does seem sort of right to me um and i would just also So even if they do reproduce more slowly, like biologically, that might just not matter that much in the long run, right? If they can live forever and, you know. The living forever is key. Yeah. Right, so I think, again, like, and we're scenario building here, right? So I think if you could live forever, like a lot of stuff changes for my story as well. I think it's, to your point about, you know, rich people just consuming, [50:00] consuming a lot and investing. I think this will all depend on the returns to capital, right? So like right now, the returns to data centers are super high, right? But if we get into a situation where people are satiated with capital, then the returns to accumulating capital are going to be lower. And so then these rich people are going to be consuming more, right? So because the incentive to invest is smaller. So basically, you kind of think about this in general equilibrium yeah the general equilibrium of this sort of process like we have gotten tremendously more richer since you know 1820 we've gotten many more people are invested but you're still getting a consumption response which keeps you know people employed in labor share high and that's because not necessarily i think you're probably making the same point but i mean they could just it could be that their investment has to be chai traded through actual laborers who are to go like do things for their investment to work which like would not in the future only the consumption is human mediated right because the investment can just be done by the robots but if the returns are if you so we're in the scenario with like how you can keep high labor share right let take that scenario in the scenario with high labor share for whatever reason the returns to capital are going to be lower Yeah that right And I mean to the earlier thing where we in the messy middle we were saying why this is implausible I feel like we can do a similar thing here, where for returns to capital to be lower, the growth rate has to have be lower, right? I mean, it certainly has to be lower than what we're expecting through the period of transformative AI, you know, if there's explosive growth. Yeah, yes and no. I mean, so the The capital stock could grow quickly, but the price of capital goods relative to consumption goods could be falling faster than the capital stock could grow. Oh, interesting. Yeah. It's the difference between the potential frontier of technology and what the realized prices of these things. Because you have relative prices. So you're saying I could be putting my money towards earning 30% interest and investing in data centers or whatever. There'll be something in the future if the growth rate is high that earns high returns. or I could, as a result of all the technological breakthroughs or some cool product that I really want to buy right now. And both of those will be compelling options. Yeah, it doesn't have to be a new product. It could be a human intrinsic product. Right although if it a human intrinsic product we would want to have it much more in the future than we want it now because sort of the thing it compares against is So we might want it the same as we want it now in the sense that like the marginal utility in a ballerina performance is exactly the same as now, right? But the marginal utility in a robot might just be a lot lower than now, right? So in units of robots, we want it a lot more than we want it now, right? So would the interest rate be 30%? It depends what you mean by the real interest rate, okay? It might be that every robot now can turn into, you know, 100 robots next year, right? So in units of robots, the interest rate is 10,000%. Right. But if the price of robots is falling really fast. Prices adjust. I mean, that's the whole, I think that's the whole point. Yeah. But here, prices are adjusting in this interesting way that too many macro models don't allow for, right? So what's happening is what would be called investment-specific technical change, where, yeah, the price of capital is like falling relative to the price of capital. consumption instead of like the standard doing the standard macro thing of saying there's just output yeah it's like chimera the thing called output which is one for one can be allocated to capital or consumption right that's not going to be true in this world yeah every unit of capital next year is giving up way less consumption than each unit of capital this year because like the just one robot now turns into many robots next year but but the number of ballroom is the same and again we going to go back to the increasing varieties thing where like if all of those extra robots next year are actually different varieties of robots and i'm not getting satiated on those robots then then it's a very different different story yeah right but now we're talking about the consumption world whereas for the investment side of things there could be just some greedy titan of industry who keeps wanting more and more robots and that alone would be enough it would to increase the marginal value of robots and therefore decrease labor share yes yeah okay but why are we not expecting greedy titans of industry to keep existing i mean greedy titans of industry historically have like built libraries and but that's because they die and they're like they all die everybody dies well we'll see but i mean like conditional on people dying i think like you know his like again you you had a guest on the show who said like you know to understand the future you should think about the past and i think like i i you could have new types of titans being right [55:00] being born, where their entire reason for accumulating wealth is just to accumulate wealth. But a lot of the time, at least historically, I'm just talking about historically, the wealth accumulation process is part of a large social interaction amongst peers, amongst the community, where you want to be admired in some way or something like that. So people end up like the stylized fact of titans of his of industry is like you accumulate the capital and then you like buy a bunch of stuff. Yeah, I mean, I guess it is sort of a historical question, but it does seem to me in a lot of cases what is happening is that as a near the end of their life, they either hand it off to their children who are worse stewards of capital than they are. And they don't even manage to grow their wealth at the rate the economy grows much less faster than the economy grows, which their parents were doing. And also they're like, well, I care less about my children having it than me sort of playing this game of accumulating wealth. And so I'm just going to give it to some trust. And if people are living longer or if they can figure out some way in which to align their trust to this wealth accumulation process it just feels like the evolution here is so strong where you just need a couple of agents that think this way for this to be the dominant thing determining the preferences of the whole economy because this part is growing much faster than the other parts of the economy. I think you just, like, the part about satiation and diminishing marginal utilities, it keeps coming up, but I think it's really, really important. Like, you know, if a person has an intrinsic preference for accumulation, that's just like, that's what they want, I think your story is totally right. But but that's just like not how usually preferences work. Right. Like you have enough whatever you hedonics in your life. And then then like the social status, all of this sort of, you know, Rousseau wrote about this. St. Augustine wrote about this. This is like a kind of like a basic part of preferences. Now, to you, you guys are arguing about something else where like you could have such high concentration that you could just have a couple of exceptions to the rule. And that's going to be enough. and I have nothing to say about that. Yeah, I mean, I think the claim's a little stronger, not just like you could have some exceptions, but that it seems that historically and today we see the exceptions and they just haven't really taken over the economy historically because there have been these dissipation shocks, as they called So they given it to their kids and squandered it or they put it in foundations which um uh which spent it i mean it not a really a shock but i mean it right people went people might have liked to uh you know fill the universe with monuments to themselves and sort of whatever live forever very wealthy and it's like a weird preference but it's not a hypothetical preference i think that's that's the thing but who knows what's going on in their heads i think um uh even without though like the kind of intrinsic preference for accumulation there are some instrumental reasons why people some people might value accumulation which is also worth bringing up so um there's a desire for a political or philosophical or religious influence right so people get to sort of an arms race over like what you know what society looks like and what people believe um and then similarly but differently because it's not an arms race there's just a total total utilitarian philanthropy right so when i think about why it might be good to have a lot of wealth in the future as a good classical utilitarian to me the value is at least one way you could have a kind of almost unsatiating uh utility function and having wealth in the future is to create new happy beings right They just add to the total welfare of the world. You know, I mean, this idea goes at least as far back as like Bostrom's astronomical waste point that we could like put Dyson spheres around the stars and turn all the energy into really happy simulations and whatnot. I think the particular greediness of this optimizer doesn't matter what they're greedy for. I think forgetting about utilitarian philosophy or whatever, like just a pure von neumann probe has i don't know what the is this an accurate to say it they just have high marginal value for like the random solar system they'll occupy because that turns into like more solar systems it turns into more solar systems like a von neumann probe is a thing that can exist right and that's like a very greedy optimizer yeah i mean if we're talking about like whether they'll dominate the economy maybe this is a technicality but um you know we we only count final consumption goods and investment goods is gdp right if there's just this phenomenon how does the von neumann probe show up in gdp yeah exactly right so if it's like if we recognize it as a person that like owns itself and it's like sort of you know optimizing on the margin between like spending a bit more on a baby von neumann probe that colonizes another star system or like a ballerina or something and it's just like it doesn't value the ballerina very much [60:00] Yeah, when we're talking about like AI beings or like, like it just it just completely depends on how we're doing the accounting there. Right. Yeah. But it's just like, what does the world look like in a world where like von Neumann probes are possible? Is it possible labor share is high? Yeah, I think it's possible the labor share is high the way we usually account it. One of the biggest problems in our all right now is credit assignment because you have these extremely long rollouts and you need to know why they succeeded or failed. One of Cursor's researchers, Sasha Rush, gave me a Blackboard lecture on how they use targeted RL with textual feedback to deal with this problem and train Composer 2.5. I filmed on my iPhone, so apologies for the camera work. So we've generated this output. It's just a sequence of tokens. We're going to send a sequence of tokens to this model that's going to read it. And then it's going to isolate a specific, say, turn that it says is problematic. Then we're just going to do text manipulation. We're just going to take that trajectory and we're literally just going to like smash in some extra tokens. After Cursor injects these hint tokens, they run another forward pass. The trajectory itself doesn't change, but the hint causes the model to assign lower probability to the error tokens Cursor then trains the original model to match those probabilities basically teaching it to downweight these specific mistakes There a lot more nuance that we couldn include in this in the room If you want to watch the full thing I posted it on my Twitter And if you want to try out Composer 2.5, head to cursor.com slash Dwarkesh. Do economists have any advice for countries which are not in the AI production chain? If you're not either producing the AI models, you're not producing the hardware that goes in the AI models, If you're not Korea making HBM or Taiwan making with the FAS or not the Netherlands with ASML, like what is India or Nigeria? What should they be doing right now? If you're talking to Modi right now, what do you say? I think the biggest lack of resources that we have allocated in the economic profession is thinking about middle income developing countries in the age of AI. And I mean, this is my fault. You know, this is something I fault myself with as well. there's not enough people thinking about this question like there are scenarios where you know you get like ai technology um you know uh being allocated and dissipating to to nigeria and developing countries and things like that and like that leveling the playing field like essentially like giving them a like a level up as far as capabilities um but there another world where like because they don have enough resources resources they not making in they not training the models they don have the hardware where they just completely get left behind And because of you know, automation, we can produce commodities in developed countries now, then we don't even have, you know, the consumer market. And then that world looks pretty bad. yeah this seems to me like an extension of the messy middle case right um one of the ways in which the messy middle might uh only be bad in a narrow range of scenarios isn't just that like it would be easy to redistribute because it probably bigger but because um the interest rate would be way higher and or sort of equivalently the price of everything except the human intrinsic goods would be would be falling really rapidly it's sort of two sides of the same point um a little bit of savings would turn into a lot of consumption next year right so things have to go really wrong for us to like just get over the threshold of uh you know capital being productive enough to automate lots of work but not be productive enough that that the interest rate is high and or the price of capital produced goods is falling a lot okay so even without redistribution a little bit of savings will save a lot of people sorry you saying if the developing countries have some savings yeah yeah in the developed world that will be enough to produce a lot of surplus They will now be able to consume a lot using their savings But I mean the messy middle could be wider in this case I mean, they're starting from such a lower level in terms of how much they have and how much it's actually indexed to the global economy. And I think it's important for them to get on it now. And I don't have strong feelings about whether it should take the form of like sovereign wealth funds that invest in the right supply chains or or just, you know, subsidies to their own citizens to buy a little bit. This is actually, I think, a crucial point. We were talking earlier about why the Rockefellers or whatever the world, why their descendants don't control everything. If our argument about the selection of these kind of greedy optimizers hold. And one argument is just that it's like very hard to index the economy. and maybe they would have just decided to have their heirs index the economy and have it grow at the rate of economic have their wealth grow the rate of economic growth and they would be you know trillionaires their heirs would be trillionaires by now but it just history before index funds existed it's just very hard to just get um get a represent it's just a very small fraction of the economy going back 100 years accounts for a majority of the value created now and if you miss those particular things you would have basically your wealth would have just kind of stagnated. And maybe there was a brief [65:00] golden window from the creation of index funds up until, I don't know, five years ago, where actually you could index the economy and you could have your wealth grow at the rate of the economy grows. But now that we're in this world with very concentrated returns, especially two private companies, which is capital that is, as we were making a point in our blog posts, the average person has disproportionately less access to, as opposed to, you know, most of their capital is like having a random house, at least in the US. Or a part of a house. Yeah, which is, as we were saying, is sort of unique, a capital that is uniquely ill-suited to be complementary to the production of AI or the serving of AI or to robots. Or the kinds of goods that the rich will bid up the prices of. Exactly, right? Because what is the value of a house currently? It is really, the land is close to other humans and modular relational stuff that is just not going to be the main factor of production and this way georgian tax would would not raise enough money for for the sort of programs that that that we were just right stepping back the point i was trying to make is if it gets harder to index the economy now and that is supposed to be the main way in which both one and normal people are supposed to module some sort of universal income In the developed world In the developed world are supposed to have some leverage on or have some purchase on the wealth from AI. And it's also the way that developing countries are supposed to have some purchase on the wealth gains from AI. But it's very hard. I don't know. Is Nigeria own a lot of SK Hynix and Anthropic? I'm guessing not, right? It's not enough for them to just own the S&P 500. So actually, this brings up a really important point. Like, is AI going to be like electricity or social media? Right. If it's, so think about ComEd or ComEdison, whatever the electricity provider here is. It's a monopoly. It provides a resource that everybody uses. But do we think about electricity as like generating concentration of power? And is ComEd having this huge amount of political power, social power, or something like that? No, because with electricity, a lot of the downstream benefits actually came to the users of the electricity rather than the actual entity producing the electricity. On the other hand, with social media, it was the opposite case, right? Social media it was everywhere Everybody uses social media but the rents went to the platform But that a really interesting point The more you think I don endorse this take yet I'm going to talk out loud. The more you think AGI is going to be, our economy is going to be run on AGI the way our economy currently runs on electricity. That is, there's a broad fundamental transformation of the entire economy. The more it looks like electricity and the more it's like every company in the S&P of the future. Exactly. If it's going to make it to the S&P 500, it is because it has leveraged AI. And then you're indexed again. Yeah, exactly. But then again, I guess it is totally, if you just look at how concentrated the S&P is over time, you know, just like these big tech companies much more so. I guess this goes to a fundamental point that it's hard to reason about, about how much of the gains from AI these individual private companies will be able to control. And I think like the open model thing is going to be a big, big point here, right? So like if if we're indeed like we're in a we're in a world where it's like the open miles or models are six months behind the frontier, nine months, then, you know, we'll hit AGI, we'll hit whatever. And like in six months, like everybody has access to this resource. And this goes to show you that every question is connected to every other because then that question about whether there runaway gains connects to questions about recursive improvement and even about recursive improvement then continual learning which or online learning which lets a model learn on the jobs if it's deployed it gets to learn more and these are just sort of like technical question or forecasting technical questions which then impact i guess whether you're gone that will have any purchase on yeah the returns of agi um but it sounds like your answer really the reason i'm emphasizing the question is i think both for the messy middle and for developing countries a recommendation that is often made naively is you got to do some kind of retraining you got to do some kind of like jobs program or you got to have them build data centers in our country and i think you guys are suggesting something closer to just buy the index of agi that's like probably much more cleaner and much more likely to succeed strategy it's really good these are the two scenarios, right? So I think there is a world where it is concentrated, in which case it's gonna be really hard to index AGI. There is another world where it's electricity, then basically every company has access to AGI, so you just buy the index. So Nigeria just needs to buy the index. And Nigeria has access to AGI. [70:00] Because of the open minds. Yeah. So just to get back to the question of like, about whether to go with retraining or just trying to index. I would prioritize trying to index, but just given how fast AI could, you know, hit the world. But I definitely wouldn't just rely on that because like it could, the sort of messy middle type cases or just the long timelines cases on which like you we don't get anything like agi all that soon um we'll still you'll just be like leaving a lot of value on the table if you could have like retrained to be a bit better you know like educated to how you know how to use the latest wave of computing and and yeah so i don't i don't think there's that much of a an either or there like i mean maybe the reason to be pessimistic about this is because one of the reasons the country is poor is that it's a bad education system and to becoming the best in the world of retraining people at using ai it doesn't seem like a particularly promising um particularly promising strategy for this that for country although there there are cases where like in develop in develop developing countries you you had this like leap frogging effect right with like for example like mobile banking or something like that it more it much more prevalent than like nigeria than it is in germany or something like that like they everybody doing mobile banking they have they have it on their phones they constantly uh doing this sort of thing so i mean i i again i not putting probabilities on this but like with a transformative technology like ai you could get leapfrogging yeah where you know you skip the step in the middle and you can get like really astronomical growth maybe just about the ease of indexing can i just quickly say i uh i think it's definitely something to worry about a bit and keep an eye on um but as discussed in in our own essay and as other people pointed out it's already not that hard to index so it's not it's not there's been a bit of an increase in the privatization of returns but it's still like you know well under 20 percent of uh the total market cap of um non non-tiny companies in the u.s is is uh private and you know everyone thinks about open ai and anthropic and then if that's where all the wealth will accrue then yeah like all these questions about whether open models will stay only a little bit behind you know those are important um but you know even they look like they're going public before too long probably um and the frictions that have been keeping companies from going public might themselves be alleviated by ai a lot right just all the disclosure requirements and and whatnot they want to get access to more potential investors too and uh um if i had to guess i would guess that the kind of long kind of general trend of just like lowering those frictions and making it easier for more and more people to index more and more will continue despite the recent bump in the other direction this actually makes me hope even more so than before that the labs do get commoditized or at the very least they go public as soon as possible but hopefully they just get totally commoditized because um i think ai will be much more popular and uh more importantly will be much more likely to lead to broad increases in prosperity if the gains are just not particularly it is as hard to capture the gains of ai as it is to capture the gains of electrification Yeah, exactly. So I think like everybody, there's no anti-electricity people out there. Right. I mean, electricity doesn't take your job, but well, it does not have people's jobs. Yeah. Yeah. And I think it's I, you know, this is maybe a tangential to the conversation. I think like there there like a really narratives matter and there this like really negative narrative around AI right now But that because people are not putting out the positive narrative or and there a reason It's more difficult to imagine something that doesn't exist that's a good thing than losing something that exists. Right, yeah. Right. So it's very easy for somebody to go on a podcast and to say like these jobs that you like, they're going away. than to somebody to spin up like a utopia which doesn't exist yet right i hope this isn't too out of left field but i think i would be remiss if i didn't point out one big cost of having commoditized frontier ai models which is the uh the tech race dynamic right that like for safety purposes you might want fewer frontier companies so that each one has a buffer in case they want to slow things down to make things safer and uh the way this relates to our point before about the kind of widespread access of the returns, is that I think there's a lot less of a trade-off there than some people imagine. Where some people think either Frontier AI gets commoditized and we all enjoy the benefits, but there might be some risk because the market's really competitive and cutthroat, or things are safer because there's a big, [75:00] between the leader and the laggard, but that means that the leaders get fantastically wealthy. No, like you could just have a relatively big gap, but it's a public company ownership and it's widely distributed. Yeah, yeah, yeah. Yeah. More recently, I have been thinking that the risk of commodification, which is that it sort of diffuses the ability to use AI to harmful ends is worth the benefit that I just feel I worry that not only having these concentrated labs makes it so that the sort of surplus isn't as widely distributed through society, but also it creates a very tangible, clear political target for the government to, I mean, we saw this with the Defense Production Act threat against Anthropic. If there wasn't one lab that is, or a couple of labs that are clearly ahead of others, this kind of threat would be much harder to make. Thank you guys for doing this. Yeah, thank you. Thank you. I feel like there's a lot of Unresolved questions but it is helpful To know what the relevant At least what is the first branch along all these Important dimensions Great thank you