Economist Korinek Warns Reproducible AI Labor May Cut Real Wages

When Intelligence Becomes Cheap: The Economic Consequences of Transformative AI — Ai4

Ai4

Because artificial intelligence makes cognitive and physical labor economically reproducible, the resulting collapse in labor's income share threatens to overpower explosive output growth and drive real wages down.

AI-assisted summary of When Intelligence Becomes Cheap: The Economic Consequences of Transformative AI — original source. Check the original for full context.

Source published: 2026-08-18

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 18, 2026, this talk features Anton Korinek, Professor of Economics at the University of Virginia and Head of Transformative AI Economic Studies at Anthropic, speaking in his independent academic capacity. The session explores the macroeconomic consequences of cheap and abundant machine intelligence, presenting empirical measurements of the surging 'AI economy' and detailing the feedback loops that could reshape labor markets, wages, and economic growth.

## TL;DR Anton Korinek examines the economic transformation driven by compounding compute scaling, noting that effective frontier AI compute has expanded by roughly 10x per year over the past 15 years. Unlike the Industrial Revolution—which made non-reproducible human labor scarce and drove a 20x surge in real wages—the AI era introduces reproducible cognitive and physical labor, potentially depressing labor's share of income even as overall GDP growth accelerates dramatically. By modeling a distinct 'AI Economy' that is expanding at roughly 2,000% annually in quality-adjusted terms, Korinek illustrates how AI labor feedback loops can trigger explosive growth across R&D and production. Consequently, pure cognitive execution will face market devaluation, shifting human comparative advantage toward interpersonal connection and governance while requiring deliberate policy to steer post-AGI abundance toward broadly shared prosperity.

## Atmosphere & style The presentation carries the measured, analytical rigor of an academic lecture while laying out profound, civilization-scale economic shifts. Korinek delivers his data-driven forecasts with calm intellectual clarity, balancing dry, self-deprecating humor about human biological limits with rigorous macroeconomic modeling. The talk feels sober yet intellectually electrifying, moving systematically from historical baselines to exponential scenarios without lapsing into hyperbole.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=zzGvFhsayVc&t=4s) Modern large language models continue the deep evolutionary trajectory of knowledge transmission, governed by the core principle that generated intelligence scales directly as a function of invested compute. - [02:07](https://www.youtube.com/watch?v=zzGvFhsayVc&t=127s) Effective compute dedicated to frontier AI training has compounded at roughly 10x per year over the past 15 years, propelled by a 4x annual increase in raw compute combined with a 2.5x annual gain in algorithmic efficiency. - [03:37](https://www.youtube.com/watch?v=zzGvFhsayVc&t=217s) Autonomous recursive self-improvement—AI systems capable of engineering and training superior successor models—represents the decisive tipping point toward AGI and transformative economic acceleration. - [05:12](https://www.youtube.com/watch?v=zzGvFhsayVc&t=312s) The Industrial Revolution broke the Malthusian trap by making physical capital reproducible, which turned non-reproducible human labor into the primary economic bottleneck and drove a 20-fold rise in real wages. - [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) In the Age of AI, cognitive and physical labor itself becomes economically reproducible through software agents and robotics, stripping labor of its historical scarcity. - [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) While total economic output is projected to accelerate by an order of magnitude under AGI, the trajectory of real wages is deeply uncertain because a collapsing labor share of income could outpace overall GDP gains. - [12:57](https://www.youtube.com/watch?v=zzGvFhsayVc&t=777s) Modeling macroeconomic activity as 'two separate islands' reveals an emerging AI economy growing at rates exceeding 150% annual capital spending and 200% annual compute purchasing power. - [14:28](https://www.youtube.com/watch?v=zzGvFhsayVc&t=868s) When quality-adjusted for rapid capability gains, the frontier AI economy expanded at approximately 2,000% (20x) annually across 2024 and 2025—a 400x cumulative increase driven predominantly by algorithmic progress. - [15:31](https://www.youtube.com/watch?v=zzGvFhsayVc&t=931s) Deploying AI as economic labor generates powerful recursive feedback loops, as artificial workers simultaneously perform direct production, advance scientific R&D, and optimize the hardware and software architectures powering future AI. - [17:13](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1033s) As raw cognitive computation and manual labor are commoditized, human comparative advantage will pivot toward interpersonal empathy, high-level governance, and ethical oversight. - [19:25](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1165s) The paramount institutional challenge of a post-AGI transition is not restricting technological progress, but deliberately steering the resulting unprecedented economic abundance toward broadly shared human prosperity.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=zzGvFhsayVc&t=4s) Introduction and the Long Horizon of Human Language to LLMs

Anton Korinek opens his talk by clarifying his roles. While introduced as a Professor of Economics at the University of Virginia, he notes that he also recently joined Anthropic. However, because his appointment at Anthropic is brand new, he emphasizes that he is speaking solely in his academic capacity as a UVA professor based on his independent research, and his remarks do not represent the official views of Anthropic or any other affiliated institutions.

To contextualize the trajectory of artificial intelligence, Korinek traces the origins of intelligence and communication back to human evolutionary history:

```mermaid flowchart LR accTitle: Evolution of Language and Machine Intelligence accDescr: A timeline tracing human language 300k years ago, writing 5.5k years ago, and modern LLM scaling. subgraph HumanHistory["Human Foundations"] A["Spoken Language ~300,000 Years Ago (Abstract knowledge sharing)"] --> B["Written Language ~5,500 Years Ago (Recorded knowledge)"] end subgraph ModernAI["Machine Intelligence"] B --> C["Large Language Models ~9 Years Ago (Machine learning on text)"] C --> D["Exponential Compute Scaling Intelligence = f(Compute)"] end ```

Spoken language emerged roughly 300,000 years ago, distinguishing humans from all other animals by allowing knowledge to be transferred across generations through abstract learning. For hundreds of millennia, this transmission occurred strictly between human minds. Written language emerged around 5,500 years ago, marking another foundational milestone. Modern Large Language Models (LLMs), which arrived around nine years ago, represent the continuation of this lineage. Korinek summarizes the fundamental principle governing current frontier systems with an intentionally simplified equation: the amount of intelligence produced by an AI system is fundamentally a function of the compute invested into training it.

---

[02:07](https://www.youtube.com/watch?v=zzGvFhsayVc&t=127s) The 10x Annual Compounding of Effective Frontier Compute

Looking at empirical trends since the onset of the deep learning era around 2010, the raw physical compute dedicated to training frontier AI models has grown by roughly 4x per year. On a logarithmic scale, this manifests as a straight upward line representing rapid exponential expansion—a pace far outstripping the historic pace of Moore's Law. This growth is driven both by improvements in underlying semiconductor hardware and by surging financial capital investments into compute infrastructure.

```mermaid flowchart TD accTitle: Effective Compute Growth Drivers accDescr: How raw compute growth and algorithmic efficiency combine into 10x annual effective compute. subgraph Drivers["Annual Growth Multipliers"] A["Raw Compute Scaling 4.0x / year (Hardware & Capital)"] B["Algorithmic Efficiency 2.5x / year (Software & Architecture)"] end A & B --> C["Effective Training Compute 10.0x / year (Compounded annually over 15 years)"] ```

In addition to expanding hardware capacity, algorithmic efficiency—how effectively algorithms utilize that compute—has grown at an estimated 2.5x per year. Multiplying these factors ($4\times \times 2.5\times$) reveals that the *effective* compute applied to training frontier systems has expanded by an astonishing 10x per year, compounded continuously over the past 15 years. Korinek jokes about the contrast with biological limits, remarking, "And yeah, my poor brain is if anything slowly depreciating, right? But that tells you something about the future that we are heading into."

---

[03:37](https://www.youtube.com/watch?v=zzGvFhsayVc&t=217s) Recursive Self-Improvement and the Tipping Point to AGI

The central concept currently commanding Silicon Valley's focus is recursive self-improvement. This occurs when an AI system reaches sufficient competence in software engineering and computer science to autonomously design, test, and train superior descendant models.

While the field has not yet crossed this threshold, Korinek underscores that reaching it would constitute an economic and technological tipping point. Once models begin autonomously improving subsequent iterations, the rate of progress could accelerate dramatically: - **Artificial General Intelligence (AGI):** Systems capable of matching or outperforming human workers across virtually all cognitive capabilities. - **Transformative AI (TAI):** AI capabilities advanced enough to fundamentally restructure macroeconomic production, labor markets, and broader social institutions. - **The Technological Singularity:** An extreme scenario where compounding recursive self-improvement leads to runaway, unprecedented technological transformation.

---

[05:12](https://www.youtube.com/watch?v=zzGvFhsayVc&t=312s) Historical Paradigm Shifts: The Malthusian Age vs. The Industrial Revolution

To understand the magnitude of the economic transition ahead, Korinek compares it to the previous macroeconomic paradigm shift: the Industrial Revolution.

During the pre-industrial Malthusian age, technology was essentially stagnant, and land was the fixed, non-reproducible bottleneck. Output was primarily agricultural. Because land could not be manufactured, any small technological improvement caused population growth that quickly brought average living standards back down to the brink of starvation. Labor was economically dispensable, earning only enough to survive.

The Industrial Revolution broke this trap through the scientific method and the Enlightenment. Machines and physical capital replaced land as the core driver of production. Crucially, machines were *reproducible*—economies could build more factories and equipment to increase capacity. In this new regime, non-reproducible human labor became the scarce economic bottleneck. Because human labor was indispensable to operate machines, its market value skyrocketed: - Real wages and salaries surged by approximately 20x across advanced economies over 250 years. - Labor scarcity served as the structural mechanism that distributed the massive wealth of industrialization across broad segments of society, producing unprecedented shared prosperity.

---

[09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) The Age of AI: Reproducible Labor and Uncertain Wage Trajectories

The transition into the Age of AI presents a structural parallel to the Industrial Revolution, but with a profound difference: labor itself becomes economically reproducible. - **Cognitive Labor:** When a new cognitive worker is needed, an organization can spin up a software agent at the press of a button. - **Physical Labor:** When physical tasks are required, factories can manufacture and deploy robotic workers.

```mermaid flowchart TD accTitle: AI Macroeconomic Impact Flowchart accDescr: How reproducible labor drives both massive GDP growth and downward pressure on the labor income share. A["Reproducible Labor (Software Agents + Robotics)"] --> B["Turbocharged GDP Growth (Order of magnitude higher)"] A --> C["Collapse of Labor Scarcity (Declining labor share of income)"] B & C --> D{"Net Wage Impact? Total Output × Labor Share"} D --> E["Scenario 1: Moderate Growth (Output growth outweighs share loss)"] D --> F["Scenario 2: Absolute Wage Decline (Share collapse outweighs GDP gains)"] ```

This dynamic has two profound macroeconomic consequences: 1. **Accelerated Growth:** Economic output expands by another order of magnitude, continuing the step-change observed between the Malthusian and Industrial eras. 2. **Reduced Labor Scarcity:** Because artificial systems can replicate human work, labor loses its status as the scarce bottleneck, leading to a shrinking labor share of aggregate income.

Drawing from an IMF research paper he authored, Korinek discusses simulated macroeconomic scenarios for output and wages under AGI: - **Output:** Across traditional, baseline AGI, and aggressive AGI models, total economic output accelerates dramatically. - **Wages:** Because average wages equal total output multiplied by labor's share of income, the trajectory of real wages is highly ambiguous. In moderate scenarios, wages could rise slowly alongside a booming economy. However, in aggressive AGI simulations, the sharp drop in labor's income share overpowers overall GDP growth after a few years, causing real wages to decline outright.

Korinek emphasizes that wage trajectory is the single most uncertain variable in his modeling, though a declining labor share of national income remains a near certainty.

---

[12:57](https://www.youtube.com/watch?v=zzGvFhsayVc&t=777s) Economic Symptoms of AI Takeoff: The "Two Islands" Model

Korinek details the concrete market indicators that signal the onset of an AI takeoff: - Substantial reallocation of foundational resources—financial capital, semiconductor chips, technical talent, and energy infrastructure—into the AI ecosystem. - Major price adjustments, quantity shifts, and supply-chain bottlenecks across input and output markets. - Rapid, continuous leaps in frontier AI capabilities.

To rigorously analyze these trends, Korinek models the macroeconomy as "two separate islands": the traditional **Human Economy** and the expanding **AI Economy**. Tracking the AI island directly reveals growth rates without precedent in modern economic history: - **AI Sector Spending:** Growing at rates exceeding +150% year-over-year for the past two years. - **Compute Purchasing Power:** The volume of compute acquired by that capital is expanding at approximately 200% year-over-year.

[14:28](https://www.youtube.com/watch?v=zzGvFhsayVc&t=868s) Measuring Quality-Adjusted Growth in the AI Economy

When adjusting for the sheer speed of quality improvements across frontier AI models over the past two years, the real scale of expansion becomes apparent. If output is scaled by capability and model quality, the AI economy has achieved a quality-adjusted growth rate of roughly 2,000% annually across 2024 and 2025. Compounding at approximately 20x per year translates to a cumulative 400x expansion in quality-adjusted terms over a two-year window.

```mermaid flowchart LR accTitle: Growth Drivers of the AI Economy subgraph Drivers["Components of Growth"] A["Increased Capital Spending"] B["Hardware Efficiency Gains"] C["Algorithmic Progress"] end subgraph Outcome["Total Expansion"] D["20x Annual Growth (400x over 2 Years)"] end A --> D B --> D C --> D ```

This exponential trajectory is driven by three distinct factors: direct financial expenditure, hardware efficiency, and algorithmic progress. While rising capital investment represents a substantial driver and hardware improvements contribute a steady middle sliver, the overwhelming majority of this surge stems from software and algorithmic efficiency gains that allow systems to perform dramatically more computation and reasoning per unit of resource.

- **Annual Quality-Adjusted Growth**: ~2,000% (or 20x) year-over-year in 2024 and 2025. - **Two-Year Cumulative Scale**: Approximately 400x increase in quality-adjusted output. - **Primary Driver**: Algorithmic progress dominates over raw compute spend and hardware improvements.

---

[15:31](https://www.youtube.com/watch?v=zzGvFhsayVc&t=931s) Feedback Loops and Explosive Growth from AI Labor

Once artificial intelligence can effectively perform useful economic labor, it initiates powerful dynamic feedback loops. In traditional economic models, macroeconomic expansion is driven straightforwardly by total factor productivity (TFP) growth paired with capital accumulation, which jointly generate economic output. Introducing capable AI labor fundamentally alters this dynamic by establishing self-reinforcing cycles across the entire productive system.

```mermaid flowchart TD accTitle: AI Labor Feedback Loops in the Macroeconomy subgraph Engine["AI Labor Engine"] A["AI Labor"] B["Software Quality Advances"] C["Hardware Quality Advances"] end subgraph Economy["Macroeconomic Output"] D["Scientific Research & TFP"] E["Direct Goods & Services Production"] F["Capital Accumulation"] G["Turbocharged Economic Output"] end A -->|"Conducts R&D"| B A -->|"Optimizes Design"| C B --> A C --> A A -->|"Drives Discovery"| D A -->|"Produces Directly"| E D --> G E --> G F --> G ```

Drawing on recent research examining when automating AI research triggers explosive economic growth, AI labor acts as a direct input on multiple fronts: it conducts scientific research to drive TFP, and it directly assists in producing goods and services. Crucially, AI labor also applies itself recursively toward improving both software architectures and hardware systems. By directly accelerating its own efficiency and development, the total pool of available AI labor expands rapidly, turbocharging macroeconomic growth well beyond historical baselines.

---

[17:13](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1033s) The Transformation of Work and the Dual Economy

Operating in an economy where these feedback loops run at full steam will reshape labor markets and scientific discovery. High-leverage intelligence will spark rapid scientific breakthroughs in domains such as healthcare and broader human living standards. However, as pure cognitive problem-solving and intelligence become abundant, pure cognitive labor will face significant relative devaluation. Similarly, the eventual arrival of self-replicating robotics will devalue routine physical labor, leaving AI and automated systems as the primary engines of economic value creation.

This transition will split economic activity into a rapidly expanding AI economy and a distinct human economy. While raw cognitive and physical execution will be handled by machines, human labor will shift fundamentally toward interpersonal skills, AI oversight, and high-level governance—domains where human judgment and connection remain essential.

- **Scientific Acceleration**: Massive advances across healthcare, biotechnology, and quality of life. - **Factor Devaluation**: Relative market devaluation of pure cognitive tasks and physical labor as robotics matures. - **Human Comparative Advantage**: Rising importance of interpersonal connection, ethical oversight, and strategic governance.

---

[19:25](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1165s) Steering Toward Human Prosperity in a Post-AGI World

Even in a post-AGI regime where the autonomous AI economy expands far faster than traditional sectors, human labor and societal purpose remain central. The critical policy and societal objective is not to constrain technological capability, but to ensure that the human sphere expands alongside it.

The central imperative of this economic transition is to ensure that the wealth and productivity unlocked by transformative AI directly enrich human life. By channeling these unprecedented technological capabilities toward shared prosperity, future generations can inherit an economy that is vastly larger, more fulfilling, and more equitable than today's baseline.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=zzGvFhsayVc&t=4s) the amount of intelligence produced by an AI system is fundamentally a function of the compute invested into training it. — Anton Korinek

> [02:07](https://www.youtube.com/watch?v=zzGvFhsayVc&t=127s) And yeah, my poor brain is if anything slowly depreciating, right? But that tells you something about the future that we are heading into. — Anton Korinek

> [03:37](https://www.youtube.com/watch?v=zzGvFhsayVc&t=217s) Once models begin autonomously improving subsequent iterations, the rate of progress could accelerate dramatically — Anton Korinek

> [05:12](https://www.youtube.com/watch?v=zzGvFhsayVc&t=312s) Because human labor was indispensable to operate machines, its market value skyrocketed — Anton Korinek

> [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) Because artificial systems can replicate human work, labor loses its status as the scarce bottleneck, leading to a shrinking labor share of aggregate income. — Anton Korinek

> [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) in aggressive AGI simulations, the sharp drop in labor's income share overpowers overall GDP growth after a few years, causing real wages to decline outright. — Anton Korinek

> [12:57](https://www.youtube.com/watch?v=zzGvFhsayVc&t=777s) Tracking the AI island directly reveals growth rates without precedent in modern economic history — Anton Korinek

> [14:28](https://www.youtube.com/watch?v=zzGvFhsayVc&t=868s) If output is scaled by capability and model quality, the AI economy has achieved a quality-adjusted growth rate of roughly 2,000% annually across 2024 and 2025. — Anton Korinek

> [15:31](https://www.youtube.com/watch?v=zzGvFhsayVc&t=931s) By directly accelerating its own efficiency and development, the total pool of available AI labor expands rapidly, turbocharging macroeconomic growth well beyond historical baselines. — Anton Korinek

> [17:13](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1033s) human labor will shift fundamentally toward interpersonal skills, AI oversight, and high-level governance—domains where human judgment and connection remain essential. — Anton Korinek

> [19:25](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1165s) The critical policy and societal objective is not to constrain technological capability, but to ensure that the human sphere expands alongside it. — Anton Korinek

## Mentioned Anton Korinek, Anthropic, University of Virginia, Ai4, International Monetary Fund (IMF), Moore's Law, Large Language Models (LLMs), Artificial General Intelligence (AGI), Transformative AI (TAI), Technological Singularity, Industrial Revolution, Enlightenment, Malthusian Age

## Video description Anton Korinek - Head, Transformative AI Economic Studies, Anthropic; Professor of Economics, University of Virginia, Anthropic | University of Virginia

What happens to the economy when intelligence itself becomes cheap and abundant? The compute powering frontier AI is growing more than fourfold per year, and a distinct "AI economy" is already emerging and expanding at rates that dwarf anything in economic history. Drawing on my recent research, I will present new measurements of this AI economy and explain the feedback loops that could drive an economic take-off. The talk will then explore what this means for the questions every leader should be asking: What happens to jobs and wages when AI can perform most cognitive tasks? Join me for a data-driven tour of the most consequential economic transformation of our lifetimes — and what we can do to steer it toward broadly shared prosperity.

[Source](https://www.youtube.com/watch?v=zzGvFhsayVc)

中文简报

经济学家 Korinek 警告：可复制的 AI 劳动力或将导致实际工资下降由于人工智能使认知和体力劳动在经济上可被复制，由此导致的劳动收入份额暴跌可能压倒爆炸性的产出增长，并拉低实际工资。 Anton Korinek 探讨了由复合式算力扩展所驱动的经济变革，指出在过去 15 年中，前沿 AI 的有效算力大约以每年 10 倍的速度扩张。与使不可复制的人类劳动变得稀缺并推动实际工资激增 20 倍的工业革命不同，AI 时代引入了可复制的认知与体力劳动，即使整体 GDP 增长急剧加速，也可能压低劳动收入份额。通过对一个按质量调整后年增长率约为 2,000% 的独立“AI 经济”进行建模，Korinek 阐明了 AI 劳动力反馈回路如何在研发与生产中触发爆炸性增长。因此，纯粹的认知执行力将面临市场贬值，从而促使人类的比较优势转向人际交往和治理，同时也需要审慎的政策来引导后 AGI 时代的富足走向广泛共享的繁荣。过去15年间，前沿AI训练的有效算力以每年约10倍的速度持续复合增长，这源于硬件与资本驱动的物理算力年增4倍，以及算法效率年均2.5倍提升的叠加效应。 (2:07) 工业革命使机器资本可复制并将人类劳动推为经济核心瓶颈，带动实际工资提升20倍；但在AI时代，软件智能体与机器人使认知与体力劳动本身变得可复制，打破了劳动的历史稀缺性根基。 (5:12) 尽管AGI有望推动宏观经济总产出实现数量级跃升，但劳动所得份额下滑已基本成为定局，若其萎缩速度超过整体GDP扩张，实际工资将面临绝对下降的风险。 (9:45) 在按模型能力进行质量调整后，新兴的前沿AI经济在2024至2025年间实现了约2,000%（20倍）的年均增长，两年累计扩张400倍，其中算法效率突破构成了最主要的驱动力量。 (14:28)

Reading summary

University of Virginia economics professor Anton Korinek argues that transformative artificial intelligence is turning human labor into an economically reproducible factor, setting off exponential growth while threatening the traditional scarcity mechanism that historically raised wages. Compounding Compute and Recursive Self-Improvement (2:07) Since the onset of modern deep learning around 2010, the raw physical compute dedicated to training frontier AI models has expanded by roughly 4x per year. Combined with algorithmic efficiency gains estimated at 2.5x per year, effective training compute has compounded at roughly 10x annually over the past 15 years. Korinek frames intelligence in contemporary frontier systems as fundamentally driven by this compute investment, far outstripping the historical cadence of Moore's Law. Looking ahead, Korinek identifies recursive self-improvement as the primary threshold commanding industry focus. If models achieve sufficient proficiency in computer science and software engineering to autonomously design, test, and refine subsequent iterations, the transition toward Artificial General Intelligence (AGI) and Transformative AI (TAI) could trigger self-reinforcing, runaway acceleration beyond biological human limits. Reproducible Labor and the Breakdown of Historical Wage Dynamics (5:12) To contextualize the scale of this transition, Korinek compares it to the shift from the Malthusian era to the Industrial Revolution. In the pre-industrial economy, land was the fixed bottleneck, leaving labor economically dispensable at subsistence levels. The Industrial Revolution introduced reproducible physical capital, which made non-reproducible human labor the scarce bottleneck. This scarcity drove a roughly 20-fold surge in real wages across advanced economies over 250 years and broadly distributed industrial wealth. The AI era upends this paradigm by making cognitive and physical labor itself reproducible through software agents and robotics. Drawing on his research for the International Monetary Fund, Korinek models how this structural change could accelerate aggregate economic output while simultaneously eroding labor's share of national income. While moderate transition scenarios project modest wage gains driven by surging GDP, aggressive AGI simulations show the collapse in labor's income share outpacing economic growth, leading to outright declines in real wages. The Dual Economy and Measured Takeoff Indicators (12:57) To track empirical symptoms of an AI takeoff, Korinek conceptualizes the macroeconomy as two interconnected sectors: a traditional Human Economy and an AI Economy. While the traditional economy faces standard physical constraints, the AI sector exhibits growth rates without precedent, including sector spending expanding at over 150% year-over-year and compute purchasing power growing at roughly 200% annually. When adjusting for frontier capability and model quality across 2024 and 2025, Korinek calculates that the AI economy is expanding at roughly 2,000% annually, or 20x per year, yielding a 400-fold cumulative increase over two years. While surging financial capital and hardware improvements contribute to this trajectory, Korinek attributes the overwhelming majority of this quality-adjusted expansion to algorithmic progress. Macroeconomic Feedback Loops and the Specialization of Work (15:31) As AI systems take on economic tasks, they introduce self-reinforcing feedback loops that depart from traditional growth models. Rather than relying solely on conventional capital accumulation and exogenous productivity shifts, AI labor directly accelerates scientific research, software development, and semiconductor optimization. This recursive involvement expands the effective pool of AI labor, driving output further beyond historical baselines. In this environment, Korinek anticipates that pure cognitive execution and routine physical tasks will experience significant relative economic devaluation. Human economic comparative advantage will consequently pivot toward areas where interpersonal connection, ethical oversight, and strategic governance remain uniquely valuable. Korinek concludes that while a post-AGI world will decouple autonomous machine expansion from traditional inputs, the core policy imperative is ensuring that the wealth and productivity unlocked by transformative systems expand human living standards and broadly shared prosperity.

精炼阅读摘要

弗吉尼亚大学经济学教授安东·科里内克（Anton Korinek，以学术学者身份发言）指出，随着前沿算力以每年十倍的有效速度复合增长，人工智能正在将人类劳动转变为可无限复制的要素，这可能在推动宏观经济产出出现数量级跃升的同时，对劳动收入份额造成严重冲击。算力复合扩张与递归自我改进的奇点 (2:07) 科里内克回顾了从30万年前人类口语、5500年前书面语到9年前大型语言模型的演变路径，指出当前前沿系统的智能产出本质上是训练算力的函数。自2010年进入深度学习时代以来，前沿物理训练算力每年增长约4倍，算法效率每年增长约2.5倍，两者结合推动前沿系统的“有效训练算力”在过去15年中以每年10倍的速度持续复合增长，远超历史上的摩尔定律节奏。硅谷当前聚焦的核心门槛是“递归自我改进”，即AI系统具备足够的软件工程与计算机科学能力，能够自主设计、测试和训练更先进的后代模型。科里内克强调，虽然该临界点尚未被完全突破，但一旦跨越，系统自主改良将触发极速的技术加速，推动体系迈向在各项认知能力上匹敌人类的通用人工智能（AGI）、重塑宏观生产的变革性AI（TAI），乃至技术奇点。从马尔萨斯陷阱到劳动力可复制的新时代 (5:12) 为了界定即将到来的范式转移，科里内克将其与历史上的宏观转型进行了对比。在工业革命前的马尔萨斯时代，不可复制的土地是主要瓶颈，劳动力在经济上可被替代且仅能维持生存，技术微幅提升带来的收益会被人口增长吞噬；工业革命引入了由机器构成的可复制资本，使不可复制的人类劳动成为稀缺瓶颈，推动先进经济体在过去250年中实际工资增长约20倍，实现了广泛共享的繁荣。然而，AI时代带来了一个颠覆性的结构变化：劳动本身变成了可复制的。认知劳动可以通过点击按钮生成软件代理，物理劳动未来则可通过工厂制造人形机器人实现。虽然这将使生产摆脱人类劳动力供给的硬约束，推动GDP再次出现数量级的加速，但劳动力也将彻底丧失其作为稀缺生产要素的地位。宏观产出分化与实际工资走势的极端不确定性 (9:45) 劳动力可复制性确立了两个宏观层面的必然与未知。根据科里内克参与撰写的国际货币基金组织（IMF）研究论文模拟，在传统、基线AGI及激进AGI三种情景下，经济总产出都将呈现爆发式增长，而国民收入中流向劳动者的份额下降几乎是定局。然而，平均工资取决于“总产出”与“劳动收入份额”的乘积，这导致实际工资的最终轨迹充满不确定性。在温和情景下，巨大的GDP扩张效应可能压过份额的流失，带动实际工资缓慢上升；但在激进的AGI情景中，劳动份额的断崖式下滑在数年内就会彻底压倒GDP的增幅，最终导致实际工资绝对值出现萎缩。科里内克强调，工资走向是其宏观模型中最难预测的核心变量。双岛经济模型与质量调整后20倍的年增长 (12:57) 科里内克观察到，资本、芯片、技术人才和能源基础设施正发生巨量转移，标志着AI起飞阶段的开启。为了量化这些趋势，他将宏观经济抽象为“传统人类经济”和“AI经济”两座相互独立的岛屿。数据显示，过去两年间AI领域的资本支出年增幅超过150%，其所采购的原始算力年增幅达200%。若进一步对前沿模型的能力跃升进行质量调整，AI经济体在2024至2025年间实现了每年约2000%（即20倍）的惊人增长，两年累计达400倍。这一扩张由资本支出、硬件效率和算法进步共同支撑，但其中绝大部分增长动力源于软件和算法层面的突破，它们使得单位资源所能支持的计算与推理能力成倍释放。反馈回路与劳动力市场的人类比较优势 (15:31) 当AI能够胜任实际经济工作时，传统的“全要素生产率（TFP）与资本积累带动增长”的线性链条将被打破。AI不仅能参与科学研究以提升TFP，直接生产商品和服务，还能将其认知能力递归应用于优化自身的软件架构与芯片硬件设计，从而形成自我强化的正反馈循环，大幅提升总产出。这一动态将对就业格局产生深远影响。高杠杆智能有望在医疗健康等领域带来颠覆性突破，但纯粹的认知解决能力将变得极其廉价，常规体力劳动也将随机器人的普及而被贬值。科里内克预测，未来经济活动将走向双轨制，机器主导纯粹的认知与物理执行，而人类的比较优势将全面退守至人际交流、伦理监督以及高层级战略治理等核心领域。科里内克总结道，即使在自主AI经济增速远超传统部门的后AGI时代，人类社会的根本目标也不是限制技术演进，而是构建适配的机制，确保变革性AI所创造的巨大财富和生产力能够真正反哺人类生活，避免分配失衡并实现广泛共享的繁荣。

Detailed notes — zh

## Background 本视频由 YouTube 频道 Ai4 于 2026 年 8 月 18 日发布。演讲者为弗吉尼亚大学经济学教授兼 Anthropic 变革性 AI 经济研究主管安东·科里内克（Anton Korinek），演讲基于其前沿学术研究，深入探讨当“智能”变得廉价充裕时，前沿算力爆发驱动的“AI 经济”与递归自我改进闭环将如何重构人类社会的就业、工资结构与宏观经济增长路径。 ## TL;DR 科里内克通过梳理从马尔萨斯时代、工业革命到 AI 时代的要素演变，指出当前前沿 AI 的“有效算力”正以每年 10 倍的速度复利增长，劳动力自身正逐步沦为可无限复制的生产要素。他提出划分“人类经济”与“AI 经济”的双岛模型，揭示出经质量调整后的 AI 经济正以每年 20 倍的超几何级数扩张，并通过 AI 参与软硬件研发的自举机制涡轮增压全要素生产率。随着纯粹认知劳动与标准物理劳动的相对贬值，劳动在总产出中的分配份额将大幅萎缩，而人际情感沟通、顶层制度治理与 AI 监督将成为人类独特的核心价值支柱。他呼吁全球决策者主动设计分配与监管机制，引导这场空前的生产力爆炸外溢为人类整体福祉的切实增长，走向广泛共享的代际繁荣。 ## Atmosphere & style 整场演讲兼具顶尖经济学者的严谨实证底色与宏大的文明历史纵深，科里内克以沉稳克制的学术语言和数据图表层层拆解前沿颠覆性概念，偶以自嘲大脑折旧调节气氛。叙事从数十万年的语言演进无缝切入至令人眩晕的超指数级增长模型，在冷峻推演传统工资与劳动力市场可能遭受的断崖式震荡时保持着高度的客观与清醒。听众在被未来技术奇点与自动化浪潮的巨大冲击力所震撼的同时，也能在结尾对扩大人类福祉圆圈的深情呼吁中感受到温和而坚定的建设性力量。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=zzGvFhsayVc&t=4s) 前沿 AI 的缩放定律表明，AI 输出的智能水平本质上是其输入计算量（Compute）的函数。 - [02:07](https://www.youtube.com/watch?v=zzGvFhsayVc&t=127s) 物理算力（每年约 4 倍）与算法效率（每年约 2.5 倍）的提升相互叠加，使过去 15 年用于训练顶尖模型的有效算力保持每年 10 倍的复利增长。 - [03:37](https://www.youtube.com/watch?v=zzGvFhsayVc&t=217s) 一旦 AI 跨越自主进行软件工程与科研的临界点，递归式自我改进将开启正反馈循环，推动通用人工智能（AGI）及变革性 AI 的迅速到来。 - [05:12](https://www.youtube.com/watch?v=zzGvFhsayVc&t=312s) 工业革命打破了马尔萨斯陷阱，可复制的机器使不可复制的人类劳动成为稀缺瓶颈，从而推动劳动价值在过去 250 年暴涨 20 倍并带来共享繁荣。 - [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) AI 时代使劳动力自身（认知与体力）沦为可无限复制的要素，这虽将大幅拉高经济总产出，但也打破了人类劳动的稀缺性，导致劳动收入份额急剧萎缩。 - [12:57](https://www.youtube.com/watch?v=zzGvFhsayVc&t=777s) 科里内克提出“双岛模型”刻画人类经济与 AI 经济的异质化脱节，测算显示 AI 领域支出年增 150%，可采购算力规模年增 200%。 - [14:28](https://www.youtube.com/watch?v=zzGvFhsayVc&t=868s) 经问题解决质量与效能调整后，2024 与 2025 年 AI 经济的年增长率高达约 2000%，两年复合增长超 400 倍，且算法进步贡献了绝大部分增量。 - [15:31](https://www.youtube.com/watch?v=zzGvFhsayVc&t=931s) 当 AI 劳动力直接投入下一代软硬件研发与宏观生产时，将形成自举式正反馈闭环，彻底打破人口与生理限制并涡轮增压宏观经济。 - [17:13](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1033s) 劳动力市场将面临价值重塑：纯认知分析与标准化体力劳动大幅贬值，而人际交往能力、同理心、AI 监督审查（AI oversight）与顶层治理将迎来价值重估。 - [19:25](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1165s) 领导者与政策制定者的核心使命并非单纯竞逐 AI 规模，而是建立合理的分配与治理路径，将 AI 释放的超级生产力转化为整个人类社会的普惠繁荣。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=zzGvFhsayVc&t=4s) 语言的演进与大语言模型的缩放定律安东·科里内克（Anton Korinek）在开场时阐明了自己的双重身份：他既是弗吉尼亚大学（University of Virginia）的经济学教授，也是近期加入 Anthropic 的研究员。不过由于入职 Anthropic 时间尚短，他特别声明当天的演讲完全以大学教授的学术身份展开，所分享的内容均基于其在弗吉尼亚大学的研究成果，仅代表个人观点。为了深入探讨人类当前所处的经济轨迹，科里内克将视野拉回到极为宏大的历史尺度——追溯至语言的起源。人类语言诞生于大约 30 万年前，这项使人类彻底区别于其他动物的创举，让人类得以通过“抽象学习”在个体之间传递知识，进而开启了整个人类文明与文化的演进；约 5,500 年前，书面文字的出现构成了知识积累的另一个关键里程碑；而大语言模型（LLM）的兴起距今不过短短 9 年左右。大语言模型问世以来，技术展现出极其惊人的缩放（Scaling）轨迹。科里内克将前沿 AI 的缩放定律（Scaling Laws）提炼为一个简明扼要的核心公式：**AI 系统输出的“智能水平”是其输入的“计算量”（Compute）的函数**。这一关系成为了驱动前沿模型能力爆发的底层逻辑。 ```mermaid flowchart LR accTitle: 知识传播与智能载体的演进历程 subgraph S1["生物与文字阶段"] A["口头语言 (~30万年前)"] --> B["书面文字 (~5500年前)"] end subgraph S2["机器智能阶段"] B --> C["大语言模型 (~9年前)"] C --> D["前沿模型缩放 (智能 = 算力的函数)"] end ``` --- ### [02:07](https://www.youtube.com/watch?v=zzGvFhsayVc&t=127s) 算力爆炸与 10 倍有效算力增长从 2010 年左右深度学习时代开启算起，用于训练最顶尖 AI 模型的物理算力正以每年约 4 倍（4x/year）的指数级速度狂飙。在对数坐标图上，这种增长表现为一条笔直向上的斜线，其扩张斜率远远甩开了传统的摩尔定律。这背后既得益于芯片硬件效率的提升，更源于全行业向算力基础设施砸下的巨额投资。更关键的是，除了纯物理算力投入的 4 倍增长，顶尖系统利用与转化算力的算法与架构效率（Algorithmic Efficiency）也在以每年约 2.5 倍的速度迭代。当两者叠加（$4 \times 2.5$），意味着过去 15 年来，注入前沿系统训练的“有效算力”（Effective Compute）每年都在以整整 **10 倍（10x/year）** 的速度复利膨胀。科里内克幽默地自嘲道：“反观我可怜的大脑，就算没有退化，顶多也就是在慢慢折旧。”这一鲜明反差预示着人类正无可阻挡地步入一个全新的智能时代。 - 物理算力增长：每年约 4 倍（4x/year），主要由资本开支狂飙与硬件进步共同驱动。 - 算法使用效率：每年约提升 2.5 倍（2.5x/year）。 - 综合有效算力：训练最前沿系统所用的实际有效算力每年增长 10 倍，并已持续复合增长 15 年。 --- ### [03:37](https://www.youtube.com/watch?v=zzGvFhsayVc&t=217s) 递归式自我改进与变革性 AI 当前硅谷的核心议题聚焦于“递归式自我改进”（Recursive Self-Improvement）。这一构想的核心逻辑在于：一旦 AI 系统的认知水平达到能够胜任软件工程和计算机科学研究的门槛，它们就能自主训练、优化出能力更强大的后代模型，进而拉开自主进化的正反馈循环。科里内克指出，尽管人类目前尚未真正跨越这一临界点，但一旦迈过，技术进化的齿轮将急剧加速，AI 可能会在极短时间内于所有认知领域全面碾压人类。届时，一系列核心概念将变为现实： 1. **通用人工智能（AGI）**：当递归自我改进的系统能独立完成几乎所有人类任务； 2. **变革性 AI（Transformative AI）**：AI 彻底颠覆现有经济结构与人类社会运行秩序； 3. **技术奇点（Singularity）**：在极限推演下，技术爆发将带来人类认知无法企及的文明质变。 ```mermaid flowchart TD accTitle: 递归式自我改进与变革性 AI 演进闭环 A["前沿 AI 模型具备软件与科研能力"] --> B["自主研发并训练更强的后代模型"] B --> C["突破递归临界点 (Tipping Point)"] C --> D["全面超越人类认知能力 (AGI)"] D --> E["经济与社会彻底重塑 (变革性 AI / 奇点)"] E --> B ``` --- ### [05:12](https://www.youtube.com/watch?v=zzGvFhsayVc&t=312s) 历史范式变迁：从马尔萨斯陷阱到工业革命为了推演 AI 带来的经济剧变，科里内克将其置于人类经济史的长期范式转移中进行考察。距今最近的一次同等量级的范式跃迁正是“工业革命”。在工业革命之前，人类长期受困于漫长且停滞的“马尔萨斯时代”（Malthusian Age，如中世纪及更早时期）：在马尔萨斯时代，技术进步极其缓慢，整个经济体系的核心瓶颈是**固定不变的土地**（农业和粮食是主导产出）。当时附着在土地上的劳动力在经济学意义上近乎“可有可无、极易替代”——边际劳动者获得的报酬仅仅勉强维持生存，全社会始终在饥饿线边缘挣扎。一旦出现微小的技术改良带来微弱增产，人口便迅速膨胀并将人均生活水平重新拉回糊口红线，陷入无休止的贫困循环。然而，随着科学方法与启蒙运动的爆发，人类迎来了工业革命的奇迹，彻底改写了经济底层逻辑： 1. **技术进步常态化**：技术从停滞变为驱动经济持续增长的核心动力； 2. **要素结构重组**：经济增长的支柱从固定的土地转移为“机器与资本”。机器与土地最大的区别在于**可复制性**——土地总量固定，只能通过零和战争争夺；但机器和工厂可以源源不断地快速制造和复刻。 3. **人类劳动力成为终极稀缺瓶颈**：随着资本无限制膨胀，唯一无法快速复制的要素变成了人类自身所能提供的劳动。在经济学规律中，成为生产瓶颈意味着高度稀缺与极高的市场需求，这直接拉升了劳动力的市场价值。在过去的 250 年里，劳动力的市场价值（工资与薪酬）暴涨了约 20 倍。工业革命之所以能带来覆盖广泛社会阶层的“共享繁荣”（Shared Prosperity），本质原因正是劳动力的稀缺性确保了新增的巨大财富能够分配到广大普通劳动者手中。 ```mermaid flowchart TD accTitle: 历史三大经济范式对比 subgraph Malthusian["马尔萨斯时代"] M1["核心瓶颈：土地 (总量固定)"] M2["劳动力报酬：仅够糊口 (饥饿边缘)"] M3["增长特征：技术停滞 / 零和博弈"] end subgraph Industrial["工业革命时代"] I1["核心支撑：资本与机器 (可复制)"] I2["核心瓶颈：人类劳动 (高度稀缺)"] I3["分配结果：劳动价值暴涨20倍 / 共享繁荣"] end subgraph AIAge["AI 时代"] A1["资本与劳动力双重可复制化"] A2["认知与体力工作者按需批量生成"] A3["产出涡轮加速 / 劳动所得份额缩水"] end Malthusian --> Industrial --> AIAge ``` --- ### [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) AI 时代的范式重塑：可复制的劳动力与工资走向站在迈向 AI 时代的门槛上，科里内克预测技术进步将进一步提速，资本将维持高速积累与无限复制。但最颠覆性的变革在于——**劳动力自身也将沦为“可复制品”**： - **认知劳动力**：每当系统需要一名脑力工作者，只需按下一个按钮，就能在云端启动一个 AI Agent； - **体力劳动力**：每当系统需要一名体力工作者，工厂流水线就能下线一台人形机器人执行任务。当劳动力能够被无限复制，经济学将迎来双重震荡： 1. **经济增速再度跃升一个数量级**：如同马尔萨斯时代跃迁至工业时代增长提升了一个数量级，AI 时代的经济总产出将迎来真正的“涡轮增压式增长”（Turbocharged Growth）。 2. **劳动力稀缺性彻底破灭**：当机器与算法能全面替代人工，劳动力在经济中的稀缺性将不复存在，劳动在国民收入中的分配份额（Labor Share）注定大幅萎缩。关于未来劳动者实际工资（Wages）的走向，科里内克展示了其三年前为国际货币基金组织（IMF）撰写论文时的仿真模型。总产出的爆炸式膨胀乘以急剧萎缩的劳动所得份额，导致工资的最终走向充满未知： - **传统/温和情景**：整体经济蛋糕急剧做大，即便分配比例下降，绝对工资水平短期内仍可能微幅上升； - **基准与激进 AGI 情景（Baseline / Aggressive AGI）**：在全面体现 AGI 替代效应的模型推演中，人工工资将在几年内见顶并出现断崖式逆转下降。科里内克坦言，未来工资到底是涨是跌是其研究中最不确定的变量，但有一点毋庸置疑：**劳动在全社会总经济产出中所占的比重将不可逆转地大幅走低**。 --- ### [12:57](https://www.youtube.com/watch?v=zzGvFhsayVc&t=777s) 经济腾飞的表征与“双岛模型”（AI 经济 vs 人类经济）经济若全面迈向 AI 腾飞（Takeoff），市场上必然显现出一系列前置经济表征：海量资本、先进制程芯片、顶尖人才和巨量能源等核心资源将大规模向 AI 部门集聚，并引发全要素市场的价格剧烈重构。随着经济重心全面向 AI 倾斜，各种前所未见的宏观经济异象也将接踵而至。为了精准刻画这一异质化脱节现象，科里内克在近期的学术论文中提出了著名的**“双岛模型”**，将整体宏观经济抽象为两座相互独立的岛屿：一座是由传统人力驱动的**“人类经济岛”**，另一座则是完全由算力与自动化运转的**“AI 经济岛”**。通过对 AI 经济岛的量化测算，多项颠覆传统经济学常识的数据浮出水面： - **资金投入增速**：AI 领域的总体支出在过去两年中以每年近 150%（+150%/year）的惊人速度增长； - **硬件采购规模**：所能购买到的有效算力规模年均增幅高达 200%（+200%/year），即每年翻三倍。这种以超几何级数扩张的“AI 经济”，其内生增长速率正在将传统人类经济远远甩在身后，成为人类经济史上前所未有的膨胀奇观。 ### [14:28](https://www.youtube.com/watch?v=zzGvFhsayVc&t=868s) 经质量调整后的 AI 经济爆炸式增长若将过去几年人工智能领域的质量跃升纳入计量考量，整个“AI 经济”的扩张速度将呈现出令传统经济学难以置信的图景。过去两年中，前沿大模型的性能实现了跨越式的飞跃；如果根据模型解决问题的实际质量与效能对产出进行调整，那么在 2024 年和 2025 年，AI 经济的“经质量调整增长率”每年都高达惊人的 2000%（即每年扩大 20 倍）。两年复合累计之下，经质量调整后的 AI 经济体量实际上暴增了约 400 倍。这一惊人增长的动力来源可以清晰地拆解为三项支柱：首先是资本支出的激增，即行业向算力与基础设施投入了海量资金；其次是硬件能效的稳步提升，但在整体增长图表中，硬件进步只占中间很薄的一小块比例；驱动增长最核心的绝对主力，是“算法层面的突破与软件系统的效率演进”。正是由于基础模型与算法架构变得越来越高效，三者叠加才最终形成了这种每年超过 20 倍的指数级扩张。 - **经质量调整年增长率**：2024 年与 2025 年每年均达到约 2000%（年环比扩张 20 倍）。 - **两年复合增长倍数**：质量调整后的 AI 经济规模整体暴涨达 400 倍（400x）。 - **增长归因结构**：硬件效率提升仅占小部分份额，海量资金投入构成基础，而“算法进步”（algorithmic progress）贡献了绝大部分的系统效率飞跃。 --- ### [15:31](https://www.youtube.com/watch?v=zzGvFhsayVc&t=931s) AI 劳动力驱动的正反馈闭环与经济起飞当人工智能能够实质性地承担具有经济价值的劳动时，传统宏观经济学的传导链条将被彻底重塑。在 Anton Korinek 最新的学术研究《自动化 AI 研究何时会带来爆炸式增长？》（*When Does Automating AI Research Produce Explosive Growth?*）中，他对比了传统经济增长模型与引入 AI 劳动力后的全新循环。在传统经济中，产出的增长主要依赖于右侧的两大支柱：全要素生产率（TFP）的自然增长以及资本的持续积累。 ```mermaid flowchart LR accTitle: AI Labor and Economic Feedback Loops subgraph S_AILabor["AI 劳动力自我强化系统"] AILabor["AI 劳动力总量"] Improvement["软件与硬件质跃提升"] AILabor -->|"参与软硬件前沿研发"| Improvement Improvement -->|"加速自身迭代与供给"| AILabor end subgraph S_Macro["宏观经济增长系统"] TFP["全要素生产率 (TFP)"] DirectOutput["经济总产出 (Output)"] Capital["资本积累 (Capital)"] TFP --> DirectOutput Capital --> DirectOutput end AILabor -->|"推动多领域科学研究"| TFP AILabor -->|"直接参与产品与服务生产"| DirectOutput ``` 然而，一旦 AI 劳动力具备了执行实际劳动的能力，系统左侧便会触发一连串前所未有的正反馈闭环。一方面，AI 劳动力能够直接从事前沿科学研究，大幅拉升全要素生产率；另一方面，它能直接作为劳动力投入到各类生产制造与服务流程中，直接拔高产出。更为关键的是，AI 自身的供给量受制于软件架构与硬件算力的提升，而 AI 劳动力反过来会直接投身于改进下一代 AI 软件算法与算力芯片的研发工作。这种“自己改进自己”的自举机制，意味着经济中可用的 AI 劳动力存量将以数十倍甚至数百倍的速度自我繁衍，进而全面“涡轮增压”（turbocharge）整体宏观经济的总产出增长。 - **传统经济增长动力**：仅依靠全要素生产率（TFP）缓慢增长与实物/金融资本积累的相互作用拉动产出。 - **AI 赋能的双重产出路径**：AI 劳动力一方面直接参与终端生产，另一方面加速科学研发、跃升全要素生产率。 - **自举式研发循环（Self-improving loop）**：AI 介入算法与硬件的迭代研发，导致可用 AI 劳动力的增长脱离人类生理与人口学限制，进入超指数级爆发区间。 --- ### [17:13](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1033s) 劳动力市场的重塑：认知贬值、机器繁衍与人类新定位当这种由 AI 自主研发驱动的正反馈回路“全速开动”（working on full steam）时，人类社会将迎来全方位的深刻巨变。首先，智能的充裕将外溢到 AI 之外的广泛科学领域，在医疗健康、生命科学以及所有依赖智力探索的领域实现突破性进展，极大改善人类的生活生存条件。但在劳动力市场内部，一场不可避免的相对价值大洗牌即将展开：由于智能本身的供应变得极其廉价，纯粹的“认知劳动”（cognitive labor）将遭遇相对贬值；与此同时，真正体现人类特质的“人际沟通与交往能力”（interpersonal skills）将迎来价值的重新发现与重估。在不久的将来，具备自我复制与装配能力的自主机器人（self-replicating robots）也将落地，这将进一步导致传统“物理/体力劳动”（physical labor）的价值大幅缩水。伴随这种空前生产力释放的是极其迅猛的经济整体增长，但这也意味着 AI 经济体系的膨胀速度将远超传统人类经济，在一些更为激进的宏观情景下甚至会呈现悬殊分化。在这样一个 AGI 之后的后奇点经济（post-AGI economy）中，人类劳动并不会彻底消失，但其角色与今天的定义有着根本性差异。纯粹的智力推演与物理执行均可交由机器与机器人完成，这蕴含着极大的劳动力市场颠覆风险；AI 与自动化系统势必会成为财富与价值创造的主要核心驱动力。对于人类劳动者而言，除了依靠人际情感连接的纽带之外，诸如“AI 监督审查”（AI oversight）与“顶层治理”（governance）等决定系统方向与责任归属的角色，将上升到比当下重要得多的核心地位。 - **行业外溢效应**：不仅加速计算科学，医疗、长寿与健康科技等一切依赖智能分析的领域均将迎来跨越式进步。 - **劳动力价值两极分化**： - *贬值资产*：纯粹的认知劳动（分析、编程、文案等）与标准化的物理/体力操作。 - *升值资产*：真实的人际交往能力、同理心连接、AI 行为的监督与审查机制（AI oversight）以及宏观治理决策（governance）。 - **经济结构二元化**：由算力和自主智能主导的 AI 经济，其扩展曲线将彻底脱离传统人类消费与生产的自然节奏。 --- ### [19:25](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1165s) 走向普惠繁荣的人类经济愿景在演讲的尾声，Anton Korinek 引导全场审视一个宏观经济结构对比图：左侧代表传统的“人类经济”（human economy），右侧则是体量庞大、飞速膨胀的“AI 经济”（AI economy）。他指出，虽然 AI 经济在相对比例上可能会以压倒性的规模增长，但只要治理与分配路径选择得当，左侧代表人类自身福祉与生活品质的经济圈层，其绝对体量完全可以被推向远比今天更宏大、更富足的维度。 ```mermaid flowchart TD accTitle: Dual Economy Structure subgraph System["宏观未来经济形态"] direction LR subgraph HumanEco["人类经济 (Human Economy)"] H1["高质量人际服务"] H2["顶层制度与安全治理"] H3["人类福祉、健康与终极繁荣"] end subgraph AIEco["AI 经济 (AI Economy)"] A1["超高速自我繁殖认知劳动力"] A2["自主机器人与物理生产"] A3["科学发现与算力自循环"] end AIEco ==>|"外溢庞大生产力与财富红利"| HumanEco end ``` Korinek 饱含深情地呼吁，在全人类共同经历这场波澜壮阔的 AI 历史大转型的当下，社会领导者、决策层与研究者的终极使命，绝非坐视技术对人的排挤，而是必须全力确保“代表人类福祉的这个圆圈尽可能扩大”。核心目标是让 AI 释放出的无限生产力真正惠及每一个人，不仅让当下世代享受到科技的红利，更为子孙后代缔造一个物质生活更繁荣、精神追求更充实、生命体验更幸福的经济社会形态。 - **核心治理诉求**：关注点不应仅仅停留在 AI 算力与经济规模的单边竞逐，而必须聚焦于把 AI 创造的增量转化为整个人类社会的实质福祉。 - **代际繁荣承诺**：通过合理的制度设计、监管重塑与经济转向，确保技术带来的超级丰裕能够普惠大众，为后代创造一个“比今天更加富裕、更有意义且更加快乐”的可持续世界。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=zzGvFhsayVc&t=4s) AI 系统输出的“智能水平”是其输入的“计算量”（Compute）的函数。 — Anton Korinek > [02:07](https://www.youtube.com/watch?v=zzGvFhsayVc&t=127s) 反观我可怜的大脑，就算没有退化，顶多也就是在慢慢折旧。 — Anton Korinek > [05:12](https://www.youtube.com/watch?v=zzGvFhsayVc&t=312s) 机器与土地最大的区别在于可复制性——土地总量固定，只能通过零和战争争夺；但机器和工厂可以源源不断地快速制造和复刻。 — Anton Korinek > [05:12](https://www.youtube.com/watch?v=zzGvFhsayVc&t=312s) 工业革命之所以能带来覆盖广泛社会阶层的“共享繁荣”（Shared Prosperity），本质原因正是劳动力的稀缺性确保了新增的巨大财富能够分配到广大普通劳动者手中。 — Anton Korinek > [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) 每当系统需要一名脑力工作者，只需按下一个按钮，就能在云端启动一个 AI Agent；每当系统需要一名体力工作者，工厂流水线就能下线一台人形机器人执行任务。 — Anton Korinek > [09:45](https://www.youtube.com/watch?v=zzGvFhsayVc&t=585s) 劳动在全社会总经济产出中所占的比重将不可逆转地大幅走低。 — Anton Korinek > [15:31](https://www.youtube.com/watch?v=zzGvFhsayVc&t=931s) 这种“自己改进自己”的自举机制，意味着经济中可用的 AI 劳动力存量将以数十倍甚至数百倍的速度自我繁衍，进而全面“涡轮增压”整体宏观经济的总产出增长。 — Anton Korinek > [17:13](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1033s) 由于智能本身的供应变得极其廉价，纯粹的“认知劳动”将遭遇相对贬值；与此同时，真正体现人类特质的“人际沟通与交往能力”将迎来价值的重新发现与重估。 — Anton Korinek > [19:25](https://www.youtube.com/watch?v=zzGvFhsayVc&t=1165s) 绝非坐视技术对人的排挤，而是必须全力确保“代表人类福祉的这个圆圈尽可能扩大”。核心目标是让 AI 释放出的无限生产力真正惠及每一个人。 — Anton Korinek ## Mentioned Anton Korinek（安东·科里内克）, Anthropic, University of Virginia（弗吉尼亚大学）, Ai4, IMF（国际货币基金组织）, 大语言模型（LLM）, 通用人工智能（AGI）, 变革性 AI（Transformative AI）, 马尔萨斯陷阱 / 马尔萨斯时代, 工业革命, 摩尔定律, 《When Does Automating AI Research Produce Explosive Growth?》 ## Video description Anton Korinek - Head, Transformative AI Economic Studies, Anthropic; Professor of Economics, University of Virginia, Anthropic | University of Virginia What happens to the economy when intelligence itself becomes cheap and abundant? The compute powering frontier AI is growing more than fourfold per year, and a distinct "AI economy" is already emerging and expanding at rates that dwarf anything in economic history. Drawing on my recent research, I will present new measurements of this AI economy and explain the feedback loops that could drive an economic take-off. The talk will then explore what this means for the questions every leader should be asking: What happens to jobs and wages when AI can perform most cognitive tasks? Join me for a data-driven tour of the most consequential economic transformation of our lifetimes — and what we can do to steer it toward broadly shared prosperity. [Source](https://www.youtube.com/watch?v=zzGvFhsayVc)

Source quotes

[0:04] Anton Korinek: AI 系统输出的“智能水平”是其输入的“计算量”（Compute）的函数。 [2:07] Anton Korinek: 反观我可怜的大脑，就算没有退化，顶多也就是在慢慢折旧。 [5:12] Anton Korinek: 机器与土地最大的区别在于可复制性——土地总量固定，只能通过零和战争争夺；但机器和工厂可以源源不断地快速制造和复刻。 [5:12] Anton Korinek: 工业革命之所以能带来覆盖广泛社会阶层的“共享繁荣”（Shared Prosperity），本质原因正是劳动力的稀缺性确保了新增的巨大财富能够分配到广大普通劳动者手中。 [9:45] Anton Korinek: 每当系统需要一名脑力工作者，只需按下一个按钮，就能在云端启动一个 AI Agent；每当系统需要一名体力工作者，工厂流水线就能下线一台人形机器人执行任务。 [9:45] Anton Korinek: 劳动在全社会总经济产出中所占的比重将不可逆转地大幅走低。 [15:31] Anton Korinek: 这种“自己改进自己”的自举机制，意味着经济中可用的 AI 劳动力存量将以数十倍甚至数百倍的速度自我繁衍，进而全面“涡轮增压”整体宏观经济的总产出增长。 [17:13] Anton Korinek: 由于智能本身的供应变得极其廉价，纯粹的“认知劳动”将遭遇相对贬值；与此同时，真正体现人类特质的“人际沟通与交往能力”将迎来价值的重新发现与重估。 [19:25] Anton Korinek: 绝非坐视技术对人的排挤，而是必须全力确保“代表人类福祉的这个圆圈尽可能扩大”。核心目标是让 AI 释放出的无限生产力真正惠及每一个人。

Source quotes

[0:04] Anton Korinek: the amount of intelligence produced by an AI system is fundamentally a function of the compute invested into training it. [2:07] Anton Korinek: And yeah, my poor brain is if anything slowly depreciating, right? But that tells you something about the future that we are heading into. [3:37] Anton Korinek: Once models begin autonomously improving subsequent iterations, the rate of progress could accelerate dramatically [5:12] Anton Korinek: Because human labor was indispensable to operate machines, its market value skyrocketed [9:45] Anton Korinek: Because artificial systems can replicate human work, labor loses its status as the scarce bottleneck, leading to a shrinking labor share of aggregate income. [9:45] Anton Korinek: in aggressive AGI simulations, the sharp drop in labor's income share overpowers overall GDP growth after a few years, causing real wages to decline outright. [12:57] Anton Korinek: Tracking the AI island directly reveals growth rates without precedent in modern economic history [14:28] Anton Korinek: If output is scaled by capability and model quality, the AI economy has achieved a quality-adjusted growth rate of roughly 2,000% annually across 2024 and 2025. [15:31] Anton Korinek: By directly accelerating its own efficiency and development, the total pool of available AI labor expands rapidly, turbocharging macroeconomic growth well beyond historical baselines. [17:13] Anton Korinek: human labor will shift fundamentally toward interpersonal skills, AI oversight, and high-level governance—domains where human judgment and connection remain essential. [19:25] Anton Korinek: The critical policy and societal objective is not to constrain technological capability, but to ensure that the human sphere expands alongside it.

Original transcript

[0:04] Hello everybody. So as uh I I was just introduced as professor of economics at the University of [0:11] Virginia, I also recently joined anthropic. Uh but uh that's been so recent that I haven't really [0:19] been able to do very much work there yet. So I am speaking in my capacity as a professor today [0:26] based on research that I have conducted at the University of Virginia and of course the views [0:31] presented are my own and don't represent any of the organizations that I'm affiliated with. [0:38] So I want to start by talking about the rise of AI because I think it's useful when we want to [0:46] talk about the economic trajectory that we're on that we start way back in some sense you can say [0:51] the current generation of language models uh are the frontier of AI and they're of course [0:58] based on language and language is something that emerged some 300,000 years ago in our species [1:06] and that's what distinguished us from all other animals. It allowed us to pass on knowledge [1:11] via abstract learning. And originally, it allowed us to pass on knowledge only between humans. You [1:18] can already see where I'm going. And that set in motion the development of human culture. [1:25] Now, written language emerged something like 5,500 years ago. Again, an important milestone [1:32] in what will ultimately feed us into LLMs. And then large language models came about [1:38] about nine years ago, right? And since then the trajectory has been one of tremendous scaling. [1:47] Uh and if we want to summarize uh the scaling laws in one simple [1:53] admittedly perhaps uh oversimplified equation, it would be that the amount of intelligence [2:00] that we get of out of AI systems is a function of the amount of compute that we put into them. [2:07] So how much has the compute going into our leading AI systems grown in recent times? [2:15] Well, if you start all the way uh from the beginning of the deep learning era around 2010, [2:22] the amount of compute going into the training of the most cutting edge models has risen [2:27] at something like a factor of 4x per year. So this is exponential growth. Here on this chart, [2:34] you actually see a logarithmic chart. That means a straight line corresponds to an exponential. [2:41] So this is much much faster than Moore's law and it's driven by growing investments into compute in [2:50] addition to uh growing efficiency coming from Mo's law. But on top of these 4x increases in compute [3:00] the efficiency at which the leading systems have used the compute has also risen at a rate that's [3:07] estimated to be 2.5x per year. So let's multiply those two. 4x * 2.5x means the effective amount [3:15] of compute going into training the most frontier systems has risen at 10x every year year after [3:23] year for the past 15 years. And yeah, my poor brain is if anything slowly depreciating, right? [3:32] But that tells you something about the future that we are heading into [3:37] now. uh at present uh all the talk in Silicon Valley is about recursive self-improvement. [3:45] So what does that mean? The basic idea is that sufficiently advanced AI models may be capable [3:54] of autonomously training even better descendant models. So at some point when they are smart [4:00] enough to do uh software engineering and computer science, one of the things that they will be able [4:07] to do is to create even better models and then set in motion this recursive process. [4:14] Uh and if that happens and to be sure we are not there yet, that would represent a tipping point [4:22] beyond beyond which progress may accelerate and after which AI could soon outperform humans at all [4:31] cognitive capabilities and of course that would have pretty transformative implications across the [4:37] economy but also across our society at large. Now there are some concepts that are closely related [4:45] and I'm sure all of you have heard about like hi. uh once we have recursively self-improving models, [4:52] they could probably soon do pretty much everything and then that would be hi. Transformative AI [4:58] captures the notion that such systems would lead to a transformation of the economy and of society [5:05] and in the extreme it could even give rise to something like a singularity. [5:12] So what does it mean for the economy side of the equation? Well, what I want to propose is that it [5:20] will lead to an economic paradigm change and again I want to kind of anchor the change in paradigm [5:29] in the longer term trajectory that we have seen in our economy. So that means I want to go back [5:35] to the last time that we had a paradigm change of similar magnitude and that was essentially [5:42] the industrial revolution. Before the industrial revolution, we lived in a Malthusian age. [5:50] During Malfusian times, you know, basically think of it as like the middle ages and so on, [5:56] the level of technology in our economy was pretty stagnant. There wasn't really very much progress. [6:04] The big bottleneck in the economy which determined how much can be produced was land [6:10] because the main product was agricultural output was food. And the labor that worked on the land [6:18] was unfortunately almost dispensable. And when I say that I mean this in an economic sense. [6:25] I mean in the sense that the marginal worker earned just enough to feed themselves. [6:32] So during those malfusian times people lived on the brink of starvation. [6:38] the population essentially expanded to the point that the available land could just feed [6:45] the population uh that lived on the land and by many measures living standards were very poor. uh [6:55] people had barely enough to survive and whenever uh they made small uh forms of technological [7:04] progress, the population just expanded again and again everybody was on the brink of starvation. [7:11] So yeah, those were dismal times and then something miraculous happened and that was [7:19] the industrial revolution. So what changed from an economic perspective in the industrial age? [7:25] Well, first of all, you know, driven by the scientific method and enlightenment and so on, [7:31] we suddenly started to make sustained technological progress. Uh technology [7:38] suddenly became a driving force of the economy. And the other really critical change that occurred [7:47] was that instead of relying primarily on land, the economy started to rely on machines [7:55] in addition to labor. And what's the amazing thing about machines compared to land? Well, [8:01] the amount of land that you have available is fixed. Uh you can perhaps engage in zero sum wars [8:08] over it, but you can't just produce more land. However, you can easily produce more machines. You [8:15] can produce more factories, be more productive, and that's exactly what happened during the [8:21] industrial age. We produced more and more and more capital, and we could reproduce that rapidly. [8:28] And what became the bottleneck was human labor. All the things that only humans could do. Now, [8:36] economically speaking, what happens when you are a bottleneck? Well, it means that you are scarce, [8:44] that you're in very high demand. If you're in high demand, your value rises. So, [8:50] the market value of labor has risen by something like a factor of 20x over the past 250 years. [8:59] And in some sense you could see that market value of labor which we also call our salaries and wages [9:08] is what lies behind the broadly based increases of living standards that we have seen across [9:16] advanced economies for the past two centuries. So we saw uh basically shared prosperity over [9:26] the past 250 years because labor was so scarce and that's what made sure that all this extra [9:35] wealth created by the economy uh created by the industrial revolution was also spread around. [9:45] Now what happens if we fast forward to the age of AI and I would say in many ways we are not quite [9:52] there yet but I think we can speculate that the pace of technological progress is almost likely to [10:00] accelerate. The amount of capital in the economy will uh almost certainly continue to go up and [10:08] capital as it has been throughout the industrial revolution will be reproducible. But the other [10:15] thing that will change is that labor will suddenly also become reproducible in the following sense. [10:21] Every time that you need a new cognitive worker, you can press a button and spin up an agent. [10:29] Every time that you need a new physical worker, you press another button and a factory puts out [10:36] another robot that can perform work for you. And so if labor suddenly becomes reproducible, [10:44] it means two things. First of all, economic growth is going to accelerate even more. So we saw a step [10:52] change from the Maltusian age to the industrial age where growth grew by an order of magnitude. [11:00] In the age of AI, it's very likely to grow another order of magnitude. And then uh we have this [11:07] rapidly growing output. And here from a paper that I wrote for the IMF three years ago uh I'm showing [11:15] you a simulation for what may happen to wages. Well, so we will have this turbocharged growth. [11:23] But if you have artificial machines that can perform labor, in some sense the scarcity of labor [11:29] is reduced. And if labor is less scarce, it means it will earn a smaller share of overall output. [11:40] Now, if you multiply a rapidly growing economy times a declining share, we can't quite see for [11:49] sure where wages will go. But I'm showing you three scenarios here in the graph on [11:54] the right hand side. So um in the center of the slide you see scenarios for output. Output will [12:02] uh depending on if you look at the traditional the baseline hi or an aggressive hi scenario [12:08] output will be turbocharged. Wages we can be less sure about which direction they'll move in. [12:16] They could continue to rise perhaps at a smaller rate than the economy as a whole, but they could [12:24] also reverse. And so in the two scenarios that capture the full effects of AGI here, we see that [12:34] in this simulation, wages can actually go down after a couple of years. And frankly speaking, [12:43] which direction they will go is the thing that I feel the most uncertain about. But I think we [12:49] can be pretty confident that the share of labor in the economy as a whole is likely to decline. [12:57] Now what would be the economic symptoms if this kind of AI takeoff starts to happen? [13:05] We would see significant resources flowing into the AI sector. And when I say resources, [13:11] I mean funding, chips, tile, and energy that would have financial implications uh across [13:17] all of those markets that would have price and quantity implications for the inputs and outputs. [13:24] Uh lots of weird things may happen as the economy pivots more and more towards AI. [13:30] And of course, it would also imply rapid advances in the AI capabilities. [13:36] And then I think it becomes useful to think about the economy as if it was two separate islands. [13:43] And in fact in a recent research paper I have done that I have kind of separated [13:48] all the economic activity uh that is created in the AI economy versus in the human economy. [13:56] And if you follow that and you try to measure progress in the AI economy, what you see [14:03] is first of all the amount of spending in the economy is growing rapidly at rates like plus 150% [14:11] a year almost in each of the past uh two years. Uh the amount of compute that you can buy with that [14:19] it's been rising at like 200% year after year. uh but the most uh spectacular increase is once [14:28] you adjust for the quality improvements that we have seen in the AI economy. So as we all know [14:35] models have gotten rapidly better over the past two years. If you adjust the output by how good [14:43] it was, uh, then you get a quality adjusted growth in the AI economy of 2,000% each year in 2024 and [14:52] 2025. So in total, this is like 20x each year. In total, it's like 400x which this AI economy grew [15:02] in quality adjusted terms. And this breaks out how much of the growth comes from spending more money [15:11] versus the hardware getting more efficient. That's like the sliver in the middle. [15:15] But the vast majority is because systems have become more efficient because of essentially [15:22] algorithmic progress. And those three combine to like growth rates of more than 20x each year. [15:31] Now once the AI can effectively perform labor there will also be a whole range of feedback [15:39] effects triggered in the economy. So this is a recent research paper on when does automating [15:45] AI research produce explosive growth and what we show in this research paper traditionally [15:52] the way that an economy operates was just what you see in the very right column here on this [15:58] slide. You have total factor productivity or generally speaking we just call it productivity [16:05] that grows and you have capital accumulation in the bottom right and the growing productivity [16:10] and the accumulation of capital together increase output. Now if you have AI that can do actually [16:19] useful labor in the economy then you have all these things that you see on the left hand side. [16:23] So the AI labor is going to do scientific research is going to increase total factor productivity. [16:31] It's going to directly help us produce more things. Those are the two arrows from AI labor [16:37] to the top right and middle right. And how will the AI labor itself uh improve? Well, [16:44] it's going to improve due to progress in software quality and progress in hardware quality. [16:50] And of course it's going to itself work on making those advances. And that means [16:56] the total amount of AI labor available to the economy could rapidly grow at the kinds of rates [17:04] that I showed on the previous slides and that in turn is going to turbocharge output growth. [17:13] Now over time what would we expect to see in the economy where all these feedback loops [17:20] are working on full steam? We would expect rapid scientific progress in fields other than AI [17:28] perhaps health care or things uh that essentially improve human living conditions uh in anything [17:37] that benefits from intelligence. We would see a relative devaluation of cognitive labor. [17:45] Perhaps we'll see a re-evaluation of the value of interpersonal skills. [17:53] Soon we should also expect self-replicating robots that will devalue physical labor. [18:00] But all of that will go hand in hand with rapid economic growth. And then the AI economy may grow [18:10] in comparison to the human economy and it's going to grow faster than the human economy. [18:16] It could, you know, in the more extreme scenarios grow very fast indeed. Um I think there will still [18:24] be a role for labor in this kind of postgi economy but it will play a different role today. So [18:33] in some sense if we look at the pure cognitive and physical tasks uh in the economy uh if it's just [18:41] about intelligence and about robotic execution machines will be able to do all of that and [18:49] that carries the potential for significant labor market disruption. So it will also imply that uh [18:57] AI and robots uh will at some point uh turn into the primary uh economic value creation drivers. [19:07] However, the interpersonal skills of humans will still be valuable and roles like AI oversight, [19:18] uh roles like governance, uh will probably take on a much much larger importance [19:25] than they have today. And then so let's look again at this economy. We have the human economy to the [19:32] left, the AI economy to the right. And let's zoom into this human economy. If things go well, this [19:40] circle that represents the human economy could actually be much much larger than it is today [19:45] and could make for much greater human well-being uh and could give us much more prosperity. [19:54] And I think our goal uh as we are kind of going through this AI transition has to be to make sure [20:01] that this human circle uh is as big as possible uh that humans can benefit from the AI advances [20:10] and live in an economy uh for our children and children's children in which people are more [20:17] prosperous, more fulfilled uh and happier than they are today. Thank you very much.