labor economics i

...sigh

i wasn’t planning to write this post because i believe most economics research is quite partisan and therefore difficult to fully trust, and also to avoid potential inflammatory comments, but after seeing too many awful takes on labor over the last few months i decided it would be useful to share my thoughts somewhere. this post focuses on the historical effect of technology on labor, the elon/vivek vs maga h1b disputes at the end of 2024, and implications for agi. i try to avoid explaining things in too much detail and instead link to references that you can read if interested

this is quite different from my typical post so if you’re not interested in labor economics you can stop reading. that being said i strongly believe this is a subject you should care about as it’s otherwise not really possible to understand the problems of people outside academic + tech bubbles / fears around “X is taking our jobs” / what the policy tradeoffs are (and also just because it’s interesting to understand the nature of work and wealth creation)

thanks to NF and sonnet 3.5 for feedback on drafts

i. technology (historically)

early economists thought that technology was great for laborers. this is in large part because classical economics was founded around the time when machines made cotton spinning far more efficient - this drove down the price of cotton, caused an explosion in cotton demand, and led to the creation of many new weaving jobs. much of the classical economics position on technology - any increase in total productivity creates wealth, which raises wages and generates more opportunity - was derived from this example [1]

of course with hindsight we know that shortly afterwards power looms made weaving much more efficient, causing wage collapse for hand-held weavers. there were no other fast-growing industries at the time so weavers had to choose between moving to factories or simply living on less money (this sounds stupid but was technically feasible because industrialization made clothes cheaper, potatoes made food cheaper, etc). either choice was a severe quality-of-life reduction [2]

i mostly bring this up to illustrate 1) the fallibility of economic theory even when it is data-driven 2) labor impacts depend heavily on the nature of the specific technology in question. one popular framework is to divide technologies into two categories: automation (replacing jobs) and augmentation (creating new jobs). this is an oversimplification - can you really determine if a technology is automating or augmenting without hindsight? - but it’s good enough for many case studies

so a natural thing to ask is where recent technological progress falls on the automation / augmentation spectrum. one popular method for doing this analysis systematically is to jointly examine usa patent data + census data and link new technologies to changes in job demographics [3]. by doing this analysis for data from 1940-2018 economists find that the effects of automation and augmentation were relatively balanced from 1940-1980, and unbalanced (in favor of automation) from 1980 onwards

specifically, they find that from 1940-1980 augmentation created middle-skill jobs in roughly the same places automation replaced them for both college-educated and non-college workers, whereas from 1980 onwards augmentation primarily created high-skill / professional jobs for college-educated workers and more low-skill / service jobs for non-college workers. this is often referred to as job polarization and is one of the primary drivers of middle-class decline

(as far as i know the change around 1980 is not well-understood - it’s not super clear if this kind of change is a property of technological progress itself and therefore inevitable, or if it’s about the relationship between technological progress and society)

so to reiterate, we know job polarization over the last 40 years has displaced middle-skill workers and driven people towards high-paying jobs that benefit from machine and computer skills as well as low-paying jobs primarily in the service industry. this has not led to a reduction in work; instead it has led to a decrease in job quality, with more people trapped in lower-paying jobs with little opportunity for skill specialization or acquisition [4]

technology has also changed the distribution of wealth - because technology often comes in the form of capital (expensive machines and computers, etc) and capital is generally not owned by workers, wealth tends to concentrate and re-invest in capital rather than labor [5]

this concentration effect also helps explain the rise of superstar firms - firms that dominate sales / profits / returns of their markets, such as faang. adopting technology often involves paying a per-period fixed cost, so larger and more productive firms are more likely to automate, which boosts labor productivity and causes further expansion. the end result is that automation raises sales concentration much more than labor concentration [6]

ii. maga disputes

the story here is roughly that in december 2024 elon musk and vivek ramaswamy came out in support of h1b visas and green cards to bring more engineers into the us. the maga response was “why are you bringing in foreign engineers instead of hiring americans”, to which elon / vivek essentially said “we can’t hire you because you’re not good at engineering and (white) americans don’t raise their kids to care about stem”. a large number of people then claimed that they actually have immense engineering potential but can’t get a job because companies are refusing to train them and discriminating in favor of cheaper non-american workers [7]

while many liberals may find it fun to point at trump supporters complaining about job discrimination and say “haha skill issue”, i think doing so is politically counterproductive (maga is largely built around loss of american jobs and won’t pass until that concern is resolved) and also ignores the important policy question of “what should we actually do about these people struggling to get jobs”

one point in the dispute revolves around work culture - the elon camp repeatedly shared anecdotes about asian immigrants being willing to work longer hours and not take vacations, to which the maga response was “well maybe work-life balance and vacations are important american values that we should preserve instead of gradually sliding towards a 996 culture”, which i think is actually a reasonable position [8]

however, there’s a revealing question buried in the work culture dispute: is it actually possible (government welfare aside) for a group of people to do the same task as everyone else while maintaining a higher standard of living and remaining competitive? if you believe markets are relatively efficient then the answer should be no in the long run, which implies any policy addressing job loss due to foreign competition is essentially choosing between the following tradeoffs*:

1. being uncompetitive, ie. sacrificing competitiveness to protect jobs. this seems to be the approach favored by maga, and involves a combination of tariffs and cracking down on immigration

2. a decrease in standard of living (or making up for the difference with welfare programs, though the usa seems allergic to this). this is essentially the default path if you follow markets to their natural conclusion, and largely describes what happened with manufacturing jobs / the rust belt

3. pivoting the industry, ie. avoiding competition by relocating workers to a related but distinct task. this is usually done through vocational training / apprenticeships / upskilling, and requires either industry or government intervention to educate workers

economists tend to agree that 1) is almost never a good idea since being uncompetitive raises domestic prices and causes more problems long-term [9]. obviously 2) sucks for everyone involved and contributes to right-wing populism. 3) is tricky to implement correctly but probably has the best outcomes when done well - many countries other than the usa also faced similar issues with competition to manufacturing jobs** but some were able to pivot to high-tech manufacturing, eg. germany (strong apprenticeship program for advanced manufacturing) [10] and sweden (early adoption of cad / 3d printing / related techniques) [11]

from this perspective a lot of maga arguments (with racist and xenophobic elements stripped away) of “why are we relying on foreign labor instead of training americans to do these jobs” become more reasonable, especially as job polarization makes it increasingly difficult for people to find well-paying jobs. more worker upskilling should’ve happened a long time ago with advanced manufacturing, and it probably should be happening now with engineering

---

*note 1: i did not discuss other relevant factors like exchange rates or access to resources for simplicity. these are all significant but do not change any of the high-level conclusions. it’s also possible there are other ways to be more productive long-term in a way that is hard to replicate, such as better culture, though i don’t think this is well-understood in the context of a fixed task

**note 2: some people may argue that the usa primarily lost manufacturing jobs due to automation and not globalization. this was actually the consensus among economists and politicians for a long time and was used as an excuse to not do anything about the problem, but turns out to be an artifact of misleading aggregate statistics - essentially the aggregate stats indicated large productivity gains in manufacturing, but if you break down manufacturing by sector it turns out almost all the productivity gains are concentrated in computer / electronics manufacturing (largely due to moore’s law) [12]. when excluding computers / electronics and analyzing the rest of the manufacturing industry, most evidence points to job loss from competition rather than automation [13]. this was an extremely stupid mistake that may have significantly influenced the course of american politics, please remember to look at the individual components of your aggregate stats:

iii. ai and agi

okay so now we can move onto everyone’s favorite topic, agi, which in many ways is the direct sequel to sections i and ii. the first thing i’ll mention is that anyone who claims to know how ai will affect jobs, wages, or employment is likely extremely overconfident [14]; goldman sachs interviewed several well-reputed economists who started with similar assumptions about ai and ended up with wildly different estimates for productivity increases in the next few years (anywhere from 1% to 10%), and this is without accounting for higher-variance events like further research breakthroughs [15]

a. things we can be confident about

some things we can be confident about: ai automation is most likely in tasks that are not natural for humans and benefit from scaling raw computational power, such as mass-producing copy and artwork, designing new molecules and materials, programming and complex logical reasoning. non-examples include most manual labor and service jobs, which rely on the physical and social dexterity that evolution has optimized humans for. task-based surveys agree that generative ai will affect higher-skill knowledge work first, with the potential to automate the majority of tasks for between 20-50% of workers in the next decade [16]. many people are excited about using ai to distill knowledge and onboard new employees faster, eg. there was a customer support case study where people trained a model on the best support agents and offered the model’s assistance to new support agents, allowing them to develop skills much faster [17]

another thing we can be confident about: even if agi is achieved, the economy will not transform overnight. this is because transformative growth requires transformative capabilities in every sector of the economy, not just a few sectors or most sectors. for instance, we saw in the last 60 years that automating manufacturing, agriculture, and textiles without automating education and healthcare simply caused education and healthcare to take up a larger share of income and weigh down on growth. this effect is known as baumol cost disease, and it occurs because productivity gains in one sector of the economy result in wage increases in unrelated sectors that have not become more productive, creating bottlenecks to growth [18]. because its capabilities are distributed very unevenly (eg. llms are currently much better at math and coding than everything else, and spikiness seems likely to persist) agi is more likely to have an effect similar to the industrial revolution, albeit perhaps at a more rapid pace, rather than being a direct route to post-scarcity [19]

b. things that are more speculative

at this point i’m going to speculate a bit about the future of how people get paid. the main point i want to emphasize is that because overnight transformation is unlikely, policies must be judged by how well they hold up during the transition to superintelligence rather than how well they hold up after superintelligence is achieved. this is the main problem with ubi in my view - maybe it results in a stable equilibrium after ai can do everything, but what about transition periods where ai can do 90% of tasks and 10% of jobs still require humans? the incentives are not well-aligned if you have 90% of people on ubi but require 10% of the population to work. there are probably variants of ubi which address this problem; i am just pointing out that the mechanism design requires more careful consideration

there has been a lot of discussion about how power will concentrate in capital rather than labor by default [20]. i agree this is the default path, but will also point out that one of the reasons power often concentrates in capital is that we currently tax most capital at ~5% and most labor at ~25% [21], and tax reform can change the balance of power between capital and labor (in fact this seems like a prerequisite for ending up in a reasonable future where a few entities don’t control everything / probably should be much higher on the policy priority list). there are also important questions about the legal status of ai / whether it can own assets that probably should be resolved in favor of ai not having rights, so that wealth remains in human control [22]

more miscellaneous thoughts:

• it seems likely that we’ll want some kind of welfare which expands over time as capabilities increase. the welfare design space is very large and i won’t pretend to understand it, but i will point out that creatively-designed welfare programs can pay for themselves and incentivize specific positive outcomes, eg. brazil’s cct (conditional on children attending school and receiving healthcare) [23] and the usa’s eitc (incentivizing low-income individuals to keep working) [24]. i’m confident someone will come up with something good here

• one idea people often discuss is compensating people for training on their data. this seems good for fairness reasons but probably does not solve any long-term problems because of increases in data efficiency, progress with synthetic data generation, and because many demographics (eg. lower-income, less fluent in english) have significantly less high-quality training data

• i previously mentioned reskilling workers as a good solution to job loss from competition, but am less sure about this approach for job loss due to ai. reskilling programs take years to complete and the average half-life of skills is estimated to be around five years [25], so if you expect human history to continue accelerating then we either need to figure out how to teach people much faster or reskilling becomes infeasible

• social movements can significantly alleviate the labor problem. for instance, in 19th century england, unions fought to eliminate child labor and move from 6-day to 5-day workweeks, which increased productivity while creating more jobs [26]. among other things this required changing cultural norms around the protestant work ethic, and it seems important to have a modern-day version of this (otherwise everyone places their value in work and then feels hopeless when ai is better than them at work)