At the dawn of the twentieth century, 40 percent of US employment was in agriculture, compared to less than 2 percent today. Nearly half of all workers held blue-collar production and manual-labor jobs in 1960, compared to only 20 percent today. While the occupation structure of the US labor market has changed since 1980, that change has been relatively modest in historical perspective.

In the past, labor market disruption was fueled by breakthrough generalpurpose technologies (GPTs) like steam power and electricity, which enabled the mechanization of agriculture. Mechanization destroyed farming jobs, but it also created factory jobs by increasing labor productivity in manufacturing. Similarly, computer-based manufacturing techniques developed in the 1970s replaced precision production jobs, while also increasing the availability of digital data and the value of analytical and managerial skills.

Although technological breakthroughs often happen suddenly, technology adoption and the pace of labor market change are often gradual. The share of employment in US agriculture fell steadily by 20 percent per decade between 1880 and 1970, and the decline in blue-collar work after 1970 was equally deliberate.

The Malthusian Age In the Malthusian Age, the economy was characterized by a production function of the form Y = A F(T, L) where technology, A, was largely stagnant (or growing very slowly), land, T, was the scarce resource, and humans who provided labor, L, were reproducible. In this era, population growth was constrained by the resources the economy could produce. Given the fixed factor land, an increase in population implied a decreasing marginal product to human labor. Humans reproduced to the point where our marginal product equaled our subsistence level. A larger population could not be supported so any additional population growth led to destitution, hunger, and death. Given this situation, land was the most valuable production factor – far more valuable in relative terms than it is today. Humans earned a subsistence income that was just enough to survive. From a purely economic perspective, the marginal human was dispensable.

The economy of the industrial age has been the main focus of modern economics – it captures the world we are still living in. It is characterized by a production function of the form Y = A F(K, L) where growth in technology, A, is the key driver of economic growth, capital, K, is the reproducible factor and accompanied the growth in A, and labor, L, grows so slowly that is usually taken as exogenous. Since the Industrial Revolution, output, Y, in advanced countries has grown about 20fold, driven by comparatively rapid technological advances.

Given these advances, labor suddenly became the scarce factor of production and experienced large increases in its marginal product in parallel with the 20-fold increase in output, giving rise to the level of wages that we see today. In short, labor transitioned from being dispensable to becoming the bottleneck production factor of the economy.

A key distinction between the 1980 shock and the LLM shock is that the first order effects of the former were on manual occupations such as textile manufacturers, while the latter is expected to primarily affect office jobs such as accounting and back-end administration.

事实上，服务品类如此多样化，以至于服务业行业自身的定义都失去了本来的意义。把国民经济划分为三个行业（第一产业、第二产业和第三产业）是20世纪中期形成的观念，当时这三个产业在经济活动和劳动力中所占的份额大致相当，至少处于可比的级别（见表2.4）。然而当发达国家中70%80%的劳动力都在服务业就业时，这种产业划分就失去了原来的意义，对于贸易的本质和全社会提供的服务提供不了有效的信息。

表2.4 18002012年法国和美国的分行业就业状况（占全部就业的比例）

注：2012年，法国的全部就业中农业占3%，制造业占21%，服务业占76%。建筑业（在2012年约占法国和美国就业人数的7%）包括在制造业中。

资料来源：piketty.pse.ens.fr/capital21c

服务业包含极其广泛的经济活动，其增长是19世纪以来生活水平改善的主要组成部分。为了进行深入考察，我们将服务业划分为几个子行业。首先是医疗和教育服务，在最发达的国家约占全部就业人数的20%以上，与所有工业部门之和相当。各种证据表明，随着医疗业的进步和高等教育的稳步增长，这个子行业还将继续扩张。其次是零售、酒店、餐饮和文化休闲业，就业人数通常占全部就业的20%，同样也在快速增长。再次是对企业的服务（如咨询、会计、设计和数据处理等）、房地产和金融服务（如房地产代理、银行、保险等）以及交通运输业，占全部就业人数的20%左右。另外，政府和安保服务（行政机构、法院、警察和军队等），在大多数国家约占全部就业人数的10%。官方统计中发布的以上服务业在就业中的比重为70%~80%。[17]

需要注意的是，这些服务中有很重要的一部分（特别是医疗和教育）通常靠税收的资金来负担，而且对客户免费。各国的筹资细节各不相同，税收负担的具体比例也不同。例如欧洲国家的比例高于美国和日本。不过在所有发达国家，来自税收的资金比重总体上较高，大致说来，医疗和教育服务有一半的成本靠税收来负担，在某些欧洲国家甚至超过3/4。这给极长期内不同国家生活水平提高的衡量和比较带来了新的困难与不确定性，而且这个问题非常重要，因为医疗和教育行业在最发达国家的国内生产总值和就业中所占的比重超过20%（未来无疑还会提高），同时代表着过去两个世纪以来生活水平改善中最切实可见和最引人关注的部分。在过去的社会，人们的预期寿命只有40岁，绝大多数人是文盲，而今天人们的寿命普遍达到80岁，几乎每个人都有一定文化水平。

It is worth reiterating two critical aspects of the direction of technology during this First, companies continued to automate parts of the production process. In fact, not just in agriculture, but throughout the economy, new machinery substituted for labor in some tasks. The key difference from the first phase of the British industrial revolution was that the reduction in labor requirements driven by automation was offset, sometimes more than one for one, with other aspects of technology that created opportunities for workers, especially for those with some basic schooling, to be employed in manufacturing or services. Second, although some of the benefits for workers from the expansion of various sectors were natural given the productivity improvements and linkages from new factories, others were a result of choices made by companies and the new cadre of engineer-managers. The direction of progress during this era was not an inexorable consequence of the nature of the leading scientific breakthroughs of the age. In fact, electricity, as a general-purpose technology, allowed different applications and distinct development paths. Managers and engineers could have chosen to double down on automation as a method to cut costs in existing industries. Instead, they built on the American path of technology and pushed to build new systems and machinery, increasing efficiency and in the process augmenting the capabilities of both skilled and unskilled labor. These technological choices were foundational to the increase in demand for workers in industry, which more than made up for declining labor intensity in agriculture and in some manufacturing tasks.

This chapter explores the long-run implications of Artificial General Intelligence (AGI) for economic growth and labor markets. AGI makes it feasible to perform all economically valuable work using compute. I distinguish between bottleneck and supplementary work—tasks that are essential versus non-essential for unhindered growth. As computational resources expand: (i) the economy automates all bottleneck work, (ii) some supplementary work may be left exclusively to humans, (iii) output becomes linear in compute and labor and its growth is driven by the expansion of compute, (iv) wages converge to the opportunity cost of computational resources required to reproduce human work, and (v) the share of labor income in GDP converges to zero.

https://www.nber.org/system/files/working_papers/w34423/w34423.pdf

Automation has been accelerating but remains “so-so”

“The density of robots has roughly tripled in the United States and other rich economies since 2010” (p. 3).

“Automation has increased productivity by only about 0.3 percent per year—far below the 2–3 percent contributions many expected” (p. 4, 28–29).

Displacement is real and concentrated

“Each additional industrial robot reduces employment by about 0.4 workers in the average commuting zone” (p. 12).

“Automation accounts for 20 to 30 percent of the decline in the manufacturing employment share since 1980” (p. 13).

“Automation disproportionately reduces employment and wages of non-college workers and middle-skill routine occupations” (p. 14–15).

The missing piece: weak “reinstatement” of labor

“The reinstatement effect—creation of new labor-intensive tasks—has been much weaker since the early 1980s than in earlier periods” (p. 6, 20–22).

“The labor share of national income has declined by more than 10 percentage points in the United States and many other advanced economies since 1980” (p. 22).

Automation vs. productivity paradox explained

“Many of the tasks automated in recent decades were already being performed at low cost with human labor… This explains why automation has delivered only modest productivity gains” (p. 7, 25–26).

“Firms adopt automation technologies primarily to reduce labor costs rather than because they dramatically outperform existing methods” (p. 26).

AI and recent trends follow the same pattern

“Generative AI and other recent software innovations target cognitive and routine tasks previously performed by professionals and clerical workers… early evidence suggests rapid adoption with similar displacement dynamics” (p. 30–31).

“There is little evidence so far of a strong reinstatement effect from AI” (p. 32).

Future outlook is conditional

Pessimistic scenario: “If the reinstatement effect remains weak, the labor share could fall to 40–50 percent and real wages for most workers could stagnate or decline” (p. 34–35).

Optimistic scenario: “With the right policies, automation could contribute an additional 1 to 2 percent annual productivity growth while maintaining or increasing the labor share” (p. 36).

Policy recommendations (explicitly listed on p. 37–39)

Eliminate the tax system’s bias toward capital (e.g., accelerated depreciation).

Redirect R&D subsidies away from pure automation and toward technologies that create new complex tasks for humans.

Invest heavily in education and retraining focused on skills that are hard to automate (creativity, interpersonal interaction, complex problem-solving).

https://www.nber.org/system/files/working_papers/w31910/w31910.pdf

When σ < 1, less labor is allocated to tasks in which labor’s productivity is higher

AI can generate new labor-intensive products or tasks, which corresponds to an increase in N .

In 1965, economic historian Robert Heilbroner confidently stated that “as machines continue to invade society, duplicating greater and greater numbers of social tasks, it is human labor itself—at least, as we now think of ‘labor’—that is gradually rendered redundant” (quoted in Akst 2014, p. 2). Wassily Leontief was equally pessimistic about the implications of new machines. By drawing an analogy with the technologies of the early twentieth century that made horses redundant, in an interview1 he speculated that “Labor will become less and less important… More and more workers will be replaced by machines. I do not see that new industries can employ everybody who wants a job.

While technology-enabled productivity growth led to employment declines in manufacturing, it also created new opportunities. Autor et al. (2024) also show how augmentation innovations in businesses processes and information technology led to the rapid rise in the US of a managerial and professional class of workers. To cite just one example, information technology enabled computer-based manufacturing techniques like the computer numerically controlled machine, allowing workers to program routine production tasks like valve cutting with incredible precision (Bartel, Ichniowski, and Shaw 2007). The automation of routine production tasks enabled more customization and process optimization, thereby complementing high-skilled managerial work (Brynjolfsson and Hitt 2003).

How does the pace of change in the US labor market today compare to that during earlier time periods? The share of employment in farming and agriculture declined steadily by 20 to 25 percent per decade between 1880 and 1980. Employment shares for blue-collar production occupations have declined by an average of 20 percent per decade since 1970. More recently, office and administrative-support occupations declined by 22 percent between 2022 and 2012 after a 12 percent drop the previous decade. Finally, sales occupations have declined by 15 percent in the last ten years, driven by a rapid drop of more than 25 percent in retail sales occupations since 2018. While the decline in office jobs has accelerated in recent years, it does not quite match the magnitude of earlier declines in farming and blue-collar work.

We find a rising share of physical capital in hardware (mechanization) and digital software in software (automation) throughout the period 1968–2019.

Mechanization of production consists in replacing human physical labor with machines within hardware. Large-scale mechanization is observed since the Industrial Revolution (≈1800 CE onwards). Mechanization applies to physical actions but not the instructions defining them.

Automation of production consists in replacing human cognitive work with digital software within software. But for early forerunners, automation is observed since the Digital Revolution (≈1980 CE onwards) when information technologies first came into use as general purpose technologies (Bresnahan and Trajtenberg, 1995). Automation pertains to cases where a task, previously involving human thought and decisions, is autonomously carried out by machines without any human intervention.

Some argue that we need not be concerned about the fate of labor since the comparative advantages of humans and machines will always allow for gainful trade between the two (see, e.g., Smith, 2024). Although it is true that agents with different capabilities – like humans and robots will have comparative advantages, this does not prevent wage declines from technological changes. Comparative advantage just tells us that two agents with different endowments or capabilities benefit from trade – it does not tell us how these benefits will evolve over time. Under AGI, the relative terms-of-trade of humans may deteriorate significantly and may imply significant income losses for the average worker. As an analogy, horses and humans have always had comparative advantages, and they still do, yet the combustion engine has greatly undermined the economic value of horses.

The distinction between mechanization and automation is instrumental in addressing questions like “will humans go the way of horses?” (Brynjolfsson and McAfee, 2014), which is supposed to mean whether human work will be eventually fully replaced by machines. The answer is: as far as physical labor is concerned, we have long gone the way of horses; for cognitive tasks (for which horses are of no use) this has not been the case yet, but it may happen in the future in the scenario of full (AI-driven) automation of production. By the same token, it is false comfort to say that the history of the Industrial Revolution teaches us that when jobs are destroyed, new ones are bound to emerge: that period only tells us that when human physical labor is mechanized, additional workers will be demanded in cognitive occupations, but it tells us nothing about cognitive occupations being automated. The hardware–software framework also helps disentangle the concepts of automation and artificial intelligence (AI). AI algorithms are a special type of software that has the ability to improve its performance based on experience and data. This happens even under a static architecture of AI algorithms, though it is conceivable that in the future AI may also modify its own architecture while heading towards self-improvement. In principle automation does not need AI, and indeed has historically begun prior to AI. However AI can strongly accelerate automation by substituting human cognitive work in non-routine tasks (Brynjolfsson, Rock and Syverson, 2019; Eloundou et al., 2023). According to Agrawal, Gans and Goldfarb (2017), while computers drastically lowered the costs of computing (arithmetic), AI drastically lowers the costs of prediction.

AI could wipe out half of all entry-level white-collar jobs — and spike unemployment to 10-20% in the next one to five years, Amodei told us in an interview from his San Francisco office.

Amodei said AI companies and government need to stop “sugar-coating” what’s coming: the possible mass elimination of jobs across technology, finance, law, consulting and other white-collar professions, especially entry-level gigs.

Few are paying attention. Lawmakers don’t get it or don’t believe it. CEOs are afraid to talk about it. Many workers won’t realize the risks posed by the possible job apocalypse — until after it hits.

“Most of them are unaware that this is about to happen,” Amodei told us. “It sounds crazy, and people just don’t believe it.”

“Cancer is cured, the economy grows at 10% a year, the budget is balanced — and 20% of people don’t have jobs.” That’s one very possible scenario rattling in his mind as AI power expands exponentially.

简单来说，无论是公司账户、国家账户还是全球经济，相应的产出和收入都可以分解为资本收入和劳动收入：

国民收入 = 资本收入 + 劳动收入[7]

但什么是资本？它的局限是什么？它以哪种形式呈现？它的构成是如何随着时间而变化的？这个问题正是本书研究的中心，我将会在随后章节中进行更为详细的阐释。本阶段，主要需要说明以下几点：

首先，在没有特殊说明的情况下，本书中提到的“资本”均不包括经济学家们经常提及的（在我的印象中）“人力资本”。人力资本通常包括个人的劳动力、技术、熟练程度和能力。在本书中，资本指的是能够划分所有权、可在市场中交换的非人力资产的总和，不仅包括所有形式的不动产（含居民住宅），还包括公司和政府机构所使用的金融资本和专业资本（厂房、基础设施、机器、专利等）。

我们之所以在资本的定义中排除了人力资本是有许多原因的，其中最显而易见的原因是人力资本无论何时都不能被另一个人所有，也不能在市场中永久交易。这是人力资本与其他形式的资本最显著的区别。当然，一个人可以在某种劳动合约下雇用另一个人来提供劳务，但是在任何现代法律体系下，类似的合约都需要在时间和范围上加以限定。当然在奴隶社会，情况显然不同：一个奴隶主可以完全彻底地拥有另一个人的人力资本，甚至是那个人的子孙后代。在这样的社会中，奴隶们可以在市场上买卖、通过遗产继承，同时将奴隶算入奴隶主的财富是十分常见的情况。当我在考察1865年以前美国南部的私人资本组成时，将会分析这一情况。除了这些特殊的历史案例之外，尝试将人力资本加入非人力资本是不太合理的。在历史上，两种形式的财富在经济增长和发展中扮演了基础和互补的角色，这也会在21世纪持续下去。但是为了理解增长的过程及其引起的不平等，我们必须仔细区分人力资本与非人力资本，予以区别对待。

非人力资本——在本书中我简单称之为“资本”——包含了私人（或私人团体）可以拥有并且能够在市场上永久交易的所有形式的财富。在现实中，资本能够被私人所有（在这种情况下我们称之为“私人资本”），或者被政府或者政府机构所有（在这种情况下我们称之为“公共资本”）。同时，也存在着中间过渡形式的资产，它们为了追求特殊的目标以集体所有的形式被“法人”所有（例如基金会和教会）。之后我将会继续阐释这个问题。能否被私人所有这一限定，随着时间推移会发生很大改变，在世界的不同地方也会有巨大差异，最极端的案例就像奴隶制这样。空气、海洋、山脉、历史遗迹、知识等的所有权也是类似的情况。某些人出于私人利益希望拥有这些东西，但有时候他们辩解说这么做是为了追求效率而非谋取私利。但是，这不能保证这种追求与大众利益相一致。资本的概念并非是一成不变的，它反映出了每个社会的发展态势及该社会普遍的社会关系。

在19世纪后期，不管是不是重体力劳动，都会因为时间长而不可避免地出现疲劳。在我们的故事开始时的1870年，制造业每周工作时间下降到每天10小时，而1830年开展的一次人口普查显示，超过一半的受访机构认为每天工作11小时以上是常见的做法。马萨诸塞州的纺织工人每天连续工作11.5~12小时，直到1874年法律规定每天工作不应超过10小时。在19世纪后期，一个工人通常每周工作6天，每天工作10小时。到1920年，工作时间减少到每周6天，每天8小时，再到1940年，每天工作同样的时间但每周工作5天。从1940年之前经历了剧变之后，标准的40小时全职工作周在战后保持了惊人的70年稳定状态。

工作时间的快速减少使凯恩斯在1931年做了一个非常有名但事实证明非常错误的预测，即社会生产率会快速提高，使每个工人每周只需工作15小时：

在以后的许多年间，我们的劣根性仍然会根深蒂固，所以，任何人如果想要生活得舒心畅意，他就必须干一点工作。比起现在的富人来，到那个时候我们将为自己多做些事。如果有些什么细小的任务要担当，或者有些日常琐事要料理，我们将感到非常高兴。不过当工作量超过这一限度时，我们将努力减轻每个人的负担，对于到那时仍然必须完成的一些工作，我们将尽可能广泛地进行分配。3小时一轮班或每周15小时的工作，也许会使上述问题在相当长一段时间内得以缓解。

计算机速度和内存的指数级增长显然遥遥领先于机器人复制人类行为的能力。虽然谷歌正在开展形如野生动物的机器人实验，这些机器人能以很快的速度运行，但到目前为止，只是简单站立起来对机器人来说难度就很大。在三年一次的机器人比赛的最近一场总决赛中，研究团队开发出最新机器人，“看到机器人会以各种方法跌倒。他们有时脸冲地，有时仰倒背着地。他们像幼儿一样翻倒，像廉价西装一样折倒，他们倒下就像成吨的砖

本章呼吁关注一个明显的冲突：一方面是技术乐观主义者对人工智能模仿并超越人类活动的新能力提升兴奋不已；一方面是过去10年全要素生产率的缓慢增长。对此的一个解释是，电脑替代人工已经持续了50多年，机器普遍替代人工已经持续了两个多世纪。一些职业如财务顾问、信用分析师、保险代理人和其他职业都处在被替代的过程中，过去20年因互联网而失业的人包括旅行代办员、百科全书推销员、博德斯公司和百视达公司的员工。然而，以前的这些失业并没有阻止美国失业率在2015年下降至接近5％的水平，因为新的就业机会被创造出来，替代了失去的就业岗位。

货物仍然由商店雇员摆放在超市货架上，啤酒、面包和软饮料仍由货车司机运输。零售市场的结账仍由雇员人工操作，而不是机器人完成，自助结账通道少之又少。理发、按摩和美甲服务仍然是完全人工的，就像餐厅需要自己的厨师和服务员一样。酒店仍然需要前台工作人员提供客房服务，而不是由机器人提供。向破坏就业的机器人和人工智能的转变远不是一夜之间发生的，而是如冰川移动一样发生。

As a result of the broadly balanced investments in automation and new tasks in the 1950s and 1960s, worker marginal productivity had increased, and the labor share of income in manufacturing remained broadly constant, hovering close to 70 percent between 1950 and the early 1980s. But by the 1980s, many American managers came to see labor as a cost, not as a resource, and to withstand foreign competition, these costs needed to be cut. This meant reducing the amount of labor used in production through automation. Recall that automation increases output per worker but, by sidelining labor, it limits and may even reduce worker marginal productivity. When this happens on a large enough scale, there is less demand for workers and lower wage growth.

Even more consequential was the elitism of this vision when it came to the nature of work. Most people were not smart enough to even excel at the jobs that were assigned to them, so the use of software designed by technology leaders to reduce corporations’ dependence on these fallible people was fully justified. Hence, automating work became an integral part of this vision, and perhaps its most powerful implication.

As a baseline, it’s useful to understand how labor markets normally respond to advances in technology. When a new technology comes along, it starts by making pieces of a given human job more efficient. For example, early in the Industrial Revolution, machines, such as upgraded plows, enabled human farmers to be more efficient at some aspects of the job. This improved the productivity of farmers, which increased their wages.

In the next step, some parts of the job of farming could be done entirely by machines, for example with the invention of the threshing machine or seed drill. In this phase, humans did a lower and lower fraction of the job, but the work they did complete became more and more leveraged because it is complementary to the work of machines, and their productivity continued to rise. As described by Jevons’ paradox, the wages of farmers and perhaps even the number of farmers continued to increase. Even when 90% of the job is being done by machines, humans can simply do 10x more of the 10% they still do, producing 10x as much output for the same amount of labor.

Eventually, machines do everything or almost everything, as with modern combine harvesters, tractors, and other equipment. At this point farming as a form of human employment really does go into steep decline, and this potentially causes serious disruption in the short term, but because farming is just one of many useful activities that humans are able to do, people eventually switch to other jobs, such as operating factory machines. This is true even though farming accounted for a huge proportion of employment ex ante. 250 years ago, 90% of Americans lived on farms; in Europe, 50–60% of employment was agricultural. Now those percentages are in the low single digits in those places, because workers switched to industrial jobs (and later, knowledge work jobs). The economy can do what previously required most of the labor force with only 1–2% of it, freeing up the rest of the labor force to build an ever more advanced industrial society. There’s no fixed “lump of labor,” just an ever-expanding ability to do more and more with less and less. People’s wages rise in line with the GDP exponential and the economy maintains full employment once disruptions in the short term have passed.

“Many production workers today, including electricians, repair workers, plumbers, nurses, educators, clerical workers, and increasingly many blue-collar workers in factories, are engaged in problem-solving tasks. These tasks require real-time, context-dependent and reliable information. For instance, an electrician dealing with the malfunctioning of advanced equipment or a short-circuit on the electricity grid will be hampered from solving these problems because he or she does not have sufficient expertise and the appropriate information for troubleshooting. Reliable information that can be provided quickly by generative AI tools can lead to significant improvements in productivity.”

AI can predict legal liability from contract language, the likelihood that a medical image indicates a specific pathology, or the next word or phrase in a standard office document, among many other possibilities. In this sense, AI is best understood as a prediction technology (Agrawal, Gans, and Goldfarb 2019). Since most jobs require some prediction and decision-making, AI will augment or automate aspects of nearly all jobs in the US economy (see, for example, Deming 2021; Eloundou et al. 2023). Thus, the impact of AI is likely to be widespread and long-lasting, fitting the mold of past GPTs. However, history teaches us that even if AI disrupts the labor market, its impact will unfold over many decades.

I argue that the emergence of AGI will mark the end of the Industrial Age and usher in a new economic era, just like the Industrial Age ended the Malthusian era. This change in paradigm will involve a significant shift in what factors of production are relevant for the economy, away from labor to reproducible factors such as computational resources and robots, with consequent implications for productivity and output growth.

For the sake of clarity, I will cut through the uncertainty regarding the development of AGI and take it as given that AGI will be reached within the foreseeable future, in line with Hinton’s prediction. Moreover, I will 7 assume that progress in robotics will proceed, and that AGI-powered robots can perform any physical work that humans can perform soon after AGI is reached.

The Age of AGI When AGI is reached, AI systems can, by definition, perform any cognitive work that humans can perform, and, per our earlier assumption, will soon perform any physical work as well, giving rise to a production function of the form Y = A F(K, L + M) in which the new variable, M, captures machines, in the form of AI compute and robots. In this simple model, machines are a perfect substitute for human labor, i.e., for both human compute and human physical capabilities.4 In an AGI-powered world, growth in technology A will also accelerate, as artificial brainpower and robots can drive scientific progress and innovation. Both traditional physical capital K and the new machines will be reproducible resources. The distinction between the two will increasingly blur. They will accumulate without bounds and generate ever more economic capacity, in the spirit of Romer (1986)-style AK-models. An immediate implication is that the growth rate of output will rise. If technology is accelerating and all factors of production are reproducible, there are no more bottlenecks to output growth in the form of fixed factors. The implications for the marginal product of labor are also straightforward. In the Malthusian age, land was the bottleneck factor, and the marginal product of labor was competed down to the human subsistence level by the cruel process of humans multiplying until the available living space and arable land were exhausted. The Industrial Age ushered in a Golden Age for humans: labor was suddenly the bottleneck factor, and the average returns to labor grew about 20-fold above our subsistence needs. In the Age of AGI, there’s no reason for labor to continue to play such a special role. If AI can solve all computational problems that our brains can solve, why should human labor have a special status in our economic system? Labor would be a perfect substitute for machines so the two would be equally scarce. As more and more machines are accumulated, the value of labor and by extension wages will diminish. This will fundamentally challenge our present system of income distribution. However, the divergence of output growth and the marginal product of labor also suggests the contours of a solution. What is needed to create shared prosperity is to take a small sliver of the rapidly rising output from AGI and give it to the workers whose labor market value is diminished. This creates the potential for a Pareto improvement, although such a solution may face significant political economy challenges (see, e.g., Bell and Korinek, 2023).

In a world where AGI and advanced robotics can perform any cognitive and physical task that humans can, labor will lose its special status of being the main irreproducible factor in the economy. The value and role of labor will diminish in such a world.

Production and diffusion lags: Once machines reach human levels of performance and become competitive in cost, there will be significant economic incentives to deploy them throughout the economy. However, the production, distribution, and adoption of large numbers of AI systems and robots throughout the economy will take time, implying that human labor will be employed to perform tasks that can in theory already be done by machines. Part of the reason for such lags is that many consumers and producers as well as employers and employees have long-lasting relationships that involve important social dimensions. However, as AI systems and robots become cheaper and are produced on a greater scale, it will become less and less economical to employ humans when machines can perform the same jobs better and more efficiently. The stronger the competitive forces in a sector, the faster machines will be adopted. For example, consultants or workers in car manufacturing may see their jobs replaced faster than government workers.

Implicit knowledge: AGI together with human-level robots do not imply that machines can immediately perform all human jobs. Even though such machines will be able to rapidly internalize any knowledge that has been written down in explicit format, they will require training and practice to acquire the implicit knowledge necessary for many jobs. In sectors with low stakes and significant volume, training and practice will happen fast. For example, AI powered robots could quickly be trained as super-human cleaners once technically feasible and could be widely deployed once economically viable. On the other hand, it might take longer for robots to acquire human-level performance in smaller markets, say, the restoration of antique clocks (although they may figure it out quickly with super-human cognition). However, even in such niche jobs, it is likely that transferring the implicit knowledge to a machine will be quicker – and more scalable – than transferring it to another human.