Have only scanned this but it seems to have flaws I've seen elsewhere. In general. I recommend reading @Charles Dillon 🔸's article on comparative advantage (Charles, I couldn't find it here, but I suggest posting it here):
The following framework explains why horses suffered complete replacement by more advanced technology and why humans are unlikely to face the same fate due to artificial intelligence.
Humans and AIs Aren't Perfect Substitutes but Horses and Engines Were
Technological Growth and Capital Accumulation Will Raise Human Labor Productivity; Horses Can't Use Technology or Capital
Humans Own AIs and Will Spend the Productivity Gains on Goods and Services that Humans Can Produce
Comparative advantage means I'm guaranteed work but not that that work will provide enough for me to eat
I agree. The last section of the post above briefly discusses this.
The argument is plausible and supported by history but it’s not a mathematical deduction. The key elements are relative productivity differences, technological improvements that increase labor productivity, and increased income generating demand for goods and services produced by humans.
[...]
Higher wages are not always and everywhere guaranteed, but humans are not likely to face the same fate as horses. We are far from perfect substitutes for AIs which means we can specialize and trade with them, raising our productivity as the AI labor force multiplies. We can take advantage of technological growth and capital accumulation to raise our productivity further. We'll continue inventing new ways to profitably integrate with automated production processes as we have in the past. And we control the abundant wealth that AI automation will create and will funnel it into human pursuits.
Thanks Vasco, I hadn't seen that. Do you know if anyone has addressed Nathan's "Comparative advantage means I'm guaranteed work but not that that work will provide enough for me to eat" point? (Apart from Maxwell, who I guess concedes the point?)
I think MaxWell conceded Nathan's point, and I do not know about anyone disputing it in a mathematical sense (for all possible parameters of economic models). However, in practice, what matters is how automation will plausibly affect wages, and human welfare more broadly.
The key crux here seems to be: will AGI systems become so cheap to run and scalable that they will make it unviable to instead pay a human to do any work?
I think Tabarrok makes key assumptions supporting his view that the answer is "no" without fully defending them. Tabarrok dismisses the idea that running an AGI will be "costless," but it doesn't need to be costless, it just needs to be highly scalable and cheaper than than the minimum viable cost of supporting a living human. Then human labor would be essentially worthless.
It's possible AGI will never become cheap and scalable enough for this to happen, but Tabarrok doesn't ever really make an argument that this is so.
It's possible that AGI will drive down the cost of supporting human lives, but it would also drive down the cost of producing more AGI systems even faster, and there seem to be harder limits on how cheap it can be to support a human life.
will AGI systems become so cheap to run and scalable that they will make it unviable to instead pay a human to do any work?
It is not enough for AIs to be better than humans at jobs defined in an overly narrow sense. Chess engines are much cheaper to run, and play much better than top chess human players, but these still have jobs.
It's possible AGI will never become cheap and scalable enough for this to happen, but Tabarrok doesn't ever really make an argument that this is so.
I agree Maxwell does not make that argument. On the other hand, humans eventually running out of jobs is not necessarily bad either. Huge automation would increase wealth per capita a lot, and this has been associated with improvements in human welfare per capita throughout history.
If AIs are a perfect substitute for humans with lower absolute costs of production – where "costs" mean the physical resources needed to keep a flesh-and-blood human alive and productive – humans will have a comparative advantage only in theory. In practice, it would make more sense to get rid of the humans and use the inputs that would have sustained them to produce more AI labor.
Thanks, Matt. I agree. However, "If AIs are a perfect substitute for humans" is a very big if. In particular, it is not enough for AIs to be better than humans at jobs defined in an overly narrow sense. Chess engines are much cheaper to run, and play much better than top chess human players, but these still have jobs.
Yes, but this shows your claim here is actually just empirical skepticism about how general and how capable AI systems will be.
It is true that loose talk of AIs being "[merely] better than" all humans at all tasks does not imply doom, but the "merely" part is not what doomers believe.
Comparative advantage is not at all load bearing here. The only thing that is load bearing is skepticism that AI capabilities will advance sufficiently and be sufficiently cheap.
Comparative advantage alone doesn’t mean you get to eat This is Arthur B’s point above, and mine when I pointed out that there are fewer horses these days than in the past.
It is simply a claim that the market clearing price of labour does not have to be sufficient for people to get as much resources as they want. Comparative advantage tells you which actor should do which task in order to maximise output, not what anyone gets paid.
We can see this even today, without needing to imagine a future very different from our present. Around 20% of people in Africa are undernourished, with 78% (!), or one billion people, unable to afford a healthy diet. In Western countries, there is strong empirical evidence and a clear intuitive case for why e.g. some cognitively disabled individuals have market clearing wages substantially below minimum wage thresholds.
If we take sufficiently capable AI as given, we are all cognitively and physically disabled individuals for the purpose of this conversation. If AI labour is sufficiently cheap, and additional workers are sufficiently easy to spin up, the market clearing price of human labour should theoretically approach 0.
The discovery and use of machinery may be … injurious to the labouring class, as some of their number will be thrown out of employment, and population will become redundant.
Fears about human labor getting replaced by machines go back hundreds if not thousands of years. These fears continue today in response to the rapid progress of AI in several previously secure human domains.
People's labour gives them less leverage than ever before.
Achieving outlier success through your labour in most or all areas is now impossible.
Scott Alexander responded to this post with some reasons to doubt the implications of complete labor redundancy, but he accepts the premise that AGI will make human labor valueless without contest.
The basic argument in favor of this position is twofold. First, that AI will continue to rapidly automate tasks that humans used to do. Second, is that as a consequence of this automation, there will be more available labor available for a shrinking number of jobs, thus lowering wages until eventually all tasks are produced by AIs and the only way to make income is to own some GPUs, power plants, or land.
The long history of this idea proves that it is convincing but it probably isn’t true. Even if AGI continues to advance and is eventually able to do any task that humans currently do, human labor will not become worthless.
The Labor Share of GDP Has Been Constant For 200+ Years
The labor share of income - the fraction of total output that is paid to workers versus capital owners - has been constant for so long that it has become one of the fundamental regularities that any model of long run economic growth must explain. For at least200 years, 50-60% of GDP has gone to pay workers with the rest paid to machines or materials. There is lots of hubbub about falling labor share since the 1990s, but this is due to mismeasurement of labor income which has shifted increasingly into non-wage benefits like employer healthcare and retirement plans. Accounting for this, the regularity of labor’s share of income at around 50% remains.
The basic story that motivates fear of AI automation predicts that more automation leads to lower value and leverage for labor, but this story cannot a explain a flat labor share of income since 1800. The past 200+ year period of industrial economic growth has been defined by the rapid growth of labor-replacing automation, but labor’s share of income has been constant. Almost all the tasks that were done by humans in 1800 are now automated, but the labor share of income did not go to zero. This is not to say that a change in labor’s share of GDP is impossible, but the constancy of this measure through all of our past experience with labor-replacing technologies is an important foundation to keep in mind when predicting the effects of the next such technology.
Labor’s constant share of income is not due to luck or coincidence but equilibrating forces in overall demand, relative factor prices, and directed technological change.
A technological innovation that automates tasks previously done by labor does displace people and decrease labor’s share of income. However, there are four countervailing effects which keep labor’s share constant.
Automation increases productivity and output. A task won’t be automated unless it increases productivity to do it with machines instead of labor. Therefore, once something is automated, people can afford more of the automated product at lower cost and their budget constraint is relaxed everywhere else, essentially increasing their total incomes.
When incomes increase, people demand more goods from all sectors, including the sector getting automated and other labor intensive sectors. If demand for the good produced by the automated tasks is very elastic, this can even increase employment in the automated sector. For example, the printing press automated the most labor intensive part of authorship, but the increase in the quantity of books demanded was so great that in addition to each author getting hundreds of times more productive, there was also an explosion in the number of people employed as authors after the printing press automated away many of their tasks.
Alternatively, the productivity increase will have spillovers into demand for other areas of the economy that aren’t as easily automated. For example, as farming was mechanized and automated, people didn’t need to spend as much on food, so their incomes effectively increased. While they didn’t spend enough extra money on food to offset labor displacement there, they spent enough on other industries like textiles and automobiles which employed more labor to meet that demand.
Automation raises the amount of capital available to each worker. Automation makes capital more productive, thus raising the returns to investment and leading people to build more capital machines. This increase in the capital stock complements labor and makes it more productive, and thus better compensated.
For example, the automation of computing displaced many “computers”, but it made certain types of capital much more valuable. As firms invested in computers, the marginal worker became more productive since they now had access to powerful machines that helped them in their work. The rise in labor demand for more productive computer-powered secretaries, stock traders, engineers, and programmers far outweighed the displacement caused by the automation in the first place.
Additionally, since the supply of capital responds elastically to changes in its productivity while labor does not, the productivity gains from automation end up accruing to labor in a type of Baumol’s cost disease.
Automation often improves capital productivity Automation can help capital replace a task that is currently done by labor, but it can also replace old capital and improve productivity in already automated tasks. For example, the transition from horse-drawn plows to tractors doesn't have any displacement effect but does have an income effect, which again increases demand for all goods including from labor intensive sectors or sectors where labor has a comparative advantage, thus increasing wages and employment in those sectors.
Automation creates new tasks that labor can perform. Automation often straightforwardly creates a new task for labor by requiring someone to produce, supervise, and maintain the automating machinery. Even when this is not the case, the displacement caused by automation also incentivizes more investment into technology which creates new tasks for labor rather than capital. Automation means that capital has a growing opportunity cost, while labor has a falling opportunity cost. There are more valuable tasks that capital can do and fewer that labor can. So a new technology which takes advantage of underutilized labor will get higher returns than one that relies on in-demand and constrained capital.
These forces have kept the labor share of income in equilibrium through all of the major technological revolutions of the past 200+ years. Therefore, a convincing and likely true story about how AI will automate most of the tasks currently performed by humans is insufficient evidence to conclude that the labor share of income will go to zero.
Self-Replicating Genius Labor Did Not Make Average Labor Worthless
The obvious response to the above argument is that AGI is different than all the labor replacing technology which came before it. These previous technologies automated particular, narrow tasks rather than emulating the general learning capacity of the human mind.
It may be true that labor can retain it’s value against capital that automates one task at a time, but what would happen if tens or hundreds of millions of fully general human-level general intelligences suddenly entered the labor market and started competing for jobs?
We needn’t speculate because this has already happened. Over the past three centuries, population growth, urbanization, transportation increases, and global communications technology has expanded the labor market of everyone on earth to include tens or hundreds of millions of extra people. This influx of intelligence came out of a recursively self-improving feedback loop between people, the ideas they create, the resource they can produce using those new ideas, and the new people those resources can support.
For below average intelligence laborers, this means competing with an ever-growing population of people who are better than them at any conceivable task. The amount of intelligence competing in the average person’s labor market increased by several orders of magnitude over the past few centuries.
But this did not turn low skilled laborers into a permanently unemployed underclass who could not make any money from their labor. In fact, their incomes increased a lot as they specialized and traded with these much more intelligent workers.
The reason for this is comparative advantage. Low skilled laborers are worse than high skilled ones at all tasks. But high-skilled laborers face constraints: if they work on an assembly line, they can’t also teach at the university. Low-skilled workers don’t face this tradeoff and thus can outcompete high-skilled workers for assembly line jobs even when they are less productive at them. Exactly because of their superior ability at all tasks, high skilled workers give up more when they choose to do something that they could trade for.
This applies just as strongly to human level AGIs. They would face very different constraints than human geniuses, but they would still face constraints. There would still not be an infinite or costless supply of intelligence as some assume. The advanced AIs will face constraints, pushing them to specialize in their comparative advantage and trade with humans for other services even when they could do those tasks better themselves, just like advanced humans do.
This is the third in a sequence of posts taken from my recent report: Why Did Environmentalism Become Partisan?
Summary
Rising partisanship did not make environmentalism more popular or politically effective. Instead, it saw flat or falling overall public opinion, fewer major legislative achievements, and fluctuating executive actions.
Public Opinion...
This post presents the executive summary from Giving What We Can’s impact evaluation for 2025. At the end of this post we share links to more information, including the full report and...
Have only scanned this but it seems to have flaws I've seen elsewhere. In general. I recommend reading @Charles Dillon 🔸's article on comparative advantage (Charles, I couldn't find it here, but I suggest posting it here):
https://open.substack.com/pub/charlesd353/p/on-comparative-advantage-and-agi?utm_campaign=post&utm_medium=web
The quickest summary is:
+1 to this being an important question to ask.
Hi Nathan and Ben.
I liked Maxwell's follow-up post What About The Horses?.
I agree. The last section of the post above briefly discusses this.
Also on comparative advantage, I liked Noah Smith's post Plentiful, high-paying jobs in the age of AI.
Thanks Vasco, I hadn't seen that. Do you know if anyone has addressed Nathan's "Comparative advantage means I'm guaranteed work but not that that work will provide enough for me to eat" point? (Apart from Maxwell, who I guess concedes the point?)
I think MaxWell conceded Nathan's point, and I do not know about anyone disputing it in a mathematical sense (for all possible parameters of economic models). However, in practice, what matters is how automation will plausibly affect wages, and human welfare more broadly.
And let's not gloss over this, right. His concession is a knockdown argument to the overall thesis.
If AI means I can't eat, but can still work, I cannot eat. Game over is much more likely.
I do not think the concession matters much. I ultimately care about expected changes in welfare, not whether something is possible.