Very interesting! I haven’t looked at the model in detail, but I was surprised to see the comparison with the Industrial Revolution and then at the same time growth rates going back to pre-AI levels after a while. In some discussions the Industrial Revolution is thought of as increasing growth rates by an order of magnitude, permanently. What’s the reason the model doesn’t predict that?
Hi Xavi! Great question. The specific way we model the AI-technology boom is:
>In this scenario, over time, AI's usefulness in production is increasing, but only through 2060
>During that period, the US always uses the most advanced technology, and imports lots of capital from around the world to do so. Once 2060 hits though, the opportunity to automate more jobs (relative to baseline) ends. There is still elevated growth beyond that point for the US, because in a higher capital share world, there's still more Solow-style growth to be exploited over the long term. But growth will still asymptote as a function whatever the assumed long-term TFP growth rate is. y
>The rest of the world, which has a lower benefit from automation than the US, takes longer to make the investments to automate lots of jobs. Countries like China, which delay adoption of the most advanced automation, wait for interest rates to come down and for domestic wages to go up. So they see their boost to growth come later in the Century, and into the next one, after the US is already "filled up" with automation
>In the limit you might imagine letting the capital share continue to increase to 100% -- this would be an AK model -- which is something we can explore as an alternate scenario. Recall however that in an AK model the growth rate is (to slightly simplify) just the tech growth rate + the saving rate, so it's easy to reason about if you know those.
So is this exactly analogous to the industrial revolution? It's a bit hard to say. In some ways yes, in some no. In future work, we're eager to plug in other technological scenarios into the model to see how things turn out.
Thanks for the detailed response! Very helpful. If you allow me one more question, I’m very interested in the role of new tasks in the context of AI and automation. I believe new labor intensive tasks are the reason the labor share has remained relatively stable since the Industrial Revolution. If I understand it correctly, your model doesn’t take this force into account. Is this because you don’t think it will be important, or is it because it’s hard to model? Many thanks again!
Hi Xavi, when i say the model automates x% of jobs, i’m speaking loosely (in part because there is no agreed definition of this).
Rather, the model thinks of AI as, like any technology, something that allows us to make more output from a certain amount of input, but also change the relative importance of different inputs. This latter effect comes from some jobs/modes of production being destroyed and some created.
We summarize the net effect of this job creation and destruction on demand for labor by assuming AI changes production share parameters. Following most people’s thinking, on net, this will increase capitals relative importance vs labor.
> While the US benefits the most from the AI-boom, Western Europe and Japan also do well. India actually sees depressed growth due to capital flight. China is a mixed case, in the short term facing capital flight, and lower growth than otherwise, but eventually adopting and benefiting from the frontier AI technology.
Why is this the case? Doesn’t China have a much higher savings rate with a higher inter temporal discount rate? I would expect China to primarily benefit from AI, or is it that China doesn’t have enough money, despite the savings rate and labor costs are still too low to justify automation, similar to your t shirt example?
That's right. China does a great job saving, but in our model some of that investment goes abroad, and China still focuses on relatively labor-intensive production for a time.
To date, I have seen no one address what Ford instinctively saw 100 years ago. Consumer spending is the majority of our economy. Who are the companies all using AI going to sell to when no one has a job because AI is doing their work? They ( corporate America) are eating themselves. What we are most likely to end up with is a few companies in monopolist partnership who make lots of money based on extracting a few pennies from the rest of us who are basically starving. This was a Speilberg movie, I can't remember the name, but it appears to be our future. And for what? How is life going to be better? We already have (or had till Trump started fucking it up) long life spans, lowering poverty worldwide, access to entertainment and learning for the majority of the world. What does AI actually even offer? And you also haven't addressed the environmental issues AND NOT ONE OF YOU, people who are in theory economists, have addressed externality costs at all. If only someone would price what AI actually costs right now, I bet a customer service rep is not much more expensive. So please, give me these analyses.
Seth: "This will get residents of the developing world a higher rate of return, but it is the US residents who will benefit from the investments."
How is this realised as consumption for US residents, despite the shock of labour replacement and falling wage share of output? In the model labour is being substituted by capital, which has a more international composition that is not US resident. Is it via a channel of by increasing dividend payments on invested capital, enough of which flow to US residents to offset the increasing international composition of the capital factor, or by transfers which are paid via international borrowing? How is final demand held up in the US, such that automation does not become a deflationary cycle? Final demand must hold up, or final consumption is displaced by business-to-business transactions, which would increase gross output measures, but not GDP. Alternatively final demand is held by exports of goods+services but then the benefits of investment in the US is not accruing to US final consumption.
1) Even though as a share of output AI needs less labor, if output goes up by a lot (because of foreign investors) this does help support domestic wages
Thanks. Does your model allow you to decompose the changes in the share of final consumption, investment and net exports in GDP, as capital's share of output, and (I think?) net foreign ownership, increase? (Where net foreign assets have to equal final consumption + investment less net exports). Particularly to identify US share of global financial consumption in the counterfactual.
Is wage share approximates consumption share (certainly not, at least to some extent, because of net tax taxes from foreign asset income), then if wage share if 198% counterfactual and wage share 60% of counterfactual, then real consumption levels would be 20% compared to the status quo.
I'm wondering to what extent your simulation predicts an economic structure for the US, in the coutnerfactual, like the capital intensive economies of Ireland and Singapore today. Where their share of global consumption (now, and taking into account smoothed consumption via savings) is necessarily much lower than share of global GDP. (Necessarily because the investment would not happen if it was simply disbursed as wages or taxed away).
Yep our model tracks all this. To be specific: In 2060, in the AI scenario US GDP is 105% higher than baseline. But US national income is only 62% higher. And private consumption is only 46% higher.
Analysis of interest rates has the wrong sign. We'll likely be back at the zero bound in order to stabilise demand. There is no sign that AI capex is utilising significant resources - maybe electricity, but certainly not labour or capital. Collapsing interest rates may well be the dominant impact on fiscal accounts. Of course, productivity growth in healthcare and education also huge.
I must disagree. Almost all leading macroeconomic models of automation predict an increase in interest rates. The exception would be if the scarce factor ends up being some kind of highly skilled or specialized labor. See my paper "Digital Abundance and Scarce Genius".
Which begs the question, what are they missing? All the *empirical* evidence suggests otherwise, *or* other factors dominate interest rate determination and automation is trivial.
From the perspective of the model, interest rates have been declining secularly because of the global savings glut from the rising Asian middle class. So automation has to be "big enough" to overcome that downward pressure, and that's only begun recently.
That’s fair, Seth. My scepticism of higher growth = high rates is formed by cross-sectional evidence. Just look at Taiwan and China. In the ‘near-developed’ world there’s no correlation. Also, if you take Taiwan as an example, you’re talking a vast share of economic activity in tech investment. I’m v sceptical of our models of interest rate determination. Not least, at the moment one needs to specify which interest rate. Our theoretical models don’t distinguish between points on the yield curve, or the cost of equity. These are all lumped into an ‘interest rate’. But the rates are wildly divergent. Btw - I love your posts. Of everything I’m reading, they’re the most relevant and thought-provoking.
Very interesting! I haven’t looked at the model in detail, but I was surprised to see the comparison with the Industrial Revolution and then at the same time growth rates going back to pre-AI levels after a while. In some discussions the Industrial Revolution is thought of as increasing growth rates by an order of magnitude, permanently. What’s the reason the model doesn’t predict that?
Hi Xavi! Great question. The specific way we model the AI-technology boom is:
>In this scenario, over time, AI's usefulness in production is increasing, but only through 2060
>During that period, the US always uses the most advanced technology, and imports lots of capital from around the world to do so. Once 2060 hits though, the opportunity to automate more jobs (relative to baseline) ends. There is still elevated growth beyond that point for the US, because in a higher capital share world, there's still more Solow-style growth to be exploited over the long term. But growth will still asymptote as a function whatever the assumed long-term TFP growth rate is. y
>The rest of the world, which has a lower benefit from automation than the US, takes longer to make the investments to automate lots of jobs. Countries like China, which delay adoption of the most advanced automation, wait for interest rates to come down and for domestic wages to go up. So they see their boost to growth come later in the Century, and into the next one, after the US is already "filled up" with automation
>In the limit you might imagine letting the capital share continue to increase to 100% -- this would be an AK model -- which is something we can explore as an alternate scenario. Recall however that in an AK model the growth rate is (to slightly simplify) just the tech growth rate + the saving rate, so it's easy to reason about if you know those.
So is this exactly analogous to the industrial revolution? It's a bit hard to say. In some ways yes, in some no. In future work, we're eager to plug in other technological scenarios into the model to see how things turn out.
Thanks for the detailed response! Very helpful. If you allow me one more question, I’m very interested in the role of new tasks in the context of AI and automation. I believe new labor intensive tasks are the reason the labor share has remained relatively stable since the Industrial Revolution. If I understand it correctly, your model doesn’t take this force into account. Is this because you don’t think it will be important, or is it because it’s hard to model? Many thanks again!
Hi Xavi, when i say the model automates x% of jobs, i’m speaking loosely (in part because there is no agreed definition of this).
Rather, the model thinks of AI as, like any technology, something that allows us to make more output from a certain amount of input, but also change the relative importance of different inputs. This latter effect comes from some jobs/modes of production being destroyed and some created.
We summarize the net effect of this job creation and destruction on demand for labor by assuming AI changes production share parameters. Following most people’s thinking, on net, this will increase capitals relative importance vs labor.
> While the US benefits the most from the AI-boom, Western Europe and Japan also do well. India actually sees depressed growth due to capital flight. China is a mixed case, in the short term facing capital flight, and lower growth than otherwise, but eventually adopting and benefiting from the frontier AI technology.
Why is this the case? Doesn’t China have a much higher savings rate with a higher inter temporal discount rate? I would expect China to primarily benefit from AI, or is it that China doesn’t have enough money, despite the savings rate and labor costs are still too low to justify automation, similar to your t shirt example?
That's right. China does a great job saving, but in our model some of that investment goes abroad, and China still focuses on relatively labor-intensive production for a time.
To date, I have seen no one address what Ford instinctively saw 100 years ago. Consumer spending is the majority of our economy. Who are the companies all using AI going to sell to when no one has a job because AI is doing their work? They ( corporate America) are eating themselves. What we are most likely to end up with is a few companies in monopolist partnership who make lots of money based on extracting a few pennies from the rest of us who are basically starving. This was a Speilberg movie, I can't remember the name, but it appears to be our future. And for what? How is life going to be better? We already have (or had till Trump started fucking it up) long life spans, lowering poverty worldwide, access to entertainment and learning for the majority of the world. What does AI actually even offer? And you also haven't addressed the environmental issues AND NOT ONE OF YOU, people who are in theory economists, have addressed externality costs at all. If only someone would price what AI actually costs right now, I bet a customer service rep is not much more expensive. So please, give me these analyses.
Seth: "This will get residents of the developing world a higher rate of return, but it is the US residents who will benefit from the investments."
How is this realised as consumption for US residents, despite the shock of labour replacement and falling wage share of output? In the model labour is being substituted by capital, which has a more international composition that is not US resident. Is it via a channel of by increasing dividend payments on invested capital, enough of which flow to US residents to offset the increasing international composition of the capital factor, or by transfers which are paid via international borrowing? How is final demand held up in the US, such that automation does not become a deflationary cycle? Final demand must hold up, or final consumption is displaced by business-to-business transactions, which would increase gross output measures, but not GDP. Alternatively final demand is held by exports of goods+services but then the benefits of investment in the US is not accruing to US final consumption.
It benefits US workers in two ways:
1) Even though as a share of output AI needs less labor, if output goes up by a lot (because of foreign investors) this does help support domestic wages
2) Some of the foreign asset income can be taxed
Thanks. Does your model allow you to decompose the changes in the share of final consumption, investment and net exports in GDP, as capital's share of output, and (I think?) net foreign ownership, increase? (Where net foreign assets have to equal final consumption + investment less net exports). Particularly to identify US share of global financial consumption in the counterfactual.
Is wage share approximates consumption share (certainly not, at least to some extent, because of net tax taxes from foreign asset income), then if wage share if 198% counterfactual and wage share 60% of counterfactual, then real consumption levels would be 20% compared to the status quo.
I'm wondering to what extent your simulation predicts an economic structure for the US, in the coutnerfactual, like the capital intensive economies of Ireland and Singapore today. Where their share of global consumption (now, and taking into account smoothed consumption via savings) is necessarily much lower than share of global GDP. (Necessarily because the investment would not happen if it was simply disbursed as wages or taxed away).
Yep our model tracks all this. To be specific: In 2060, in the AI scenario US GDP is 105% higher than baseline. But US national income is only 62% higher. And private consumption is only 46% higher.
Analysis of interest rates has the wrong sign. We'll likely be back at the zero bound in order to stabilise demand. There is no sign that AI capex is utilising significant resources - maybe electricity, but certainly not labour or capital. Collapsing interest rates may well be the dominant impact on fiscal accounts. Of course, productivity growth in healthcare and education also huge.
I must disagree. Almost all leading macroeconomic models of automation predict an increase in interest rates. The exception would be if the scarce factor ends up being some kind of highly skilled or specialized labor. See my paper "Digital Abundance and Scarce Genius".
Which begs the question, what are they missing? All the *empirical* evidence suggests otherwise, *or* other factors dominate interest rate determination and automation is trivial.
From the perspective of the model, interest rates have been declining secularly because of the global savings glut from the rising Asian middle class. So automation has to be "big enough" to overcome that downward pressure, and that's only begun recently.
That’s fair, Seth. My scepticism of higher growth = high rates is formed by cross-sectional evidence. Just look at Taiwan and China. In the ‘near-developed’ world there’s no correlation. Also, if you take Taiwan as an example, you’re talking a vast share of economic activity in tech investment. I’m v sceptical of our models of interest rate determination. Not least, at the moment one needs to specify which interest rate. Our theoretical models don’t distinguish between points on the yield curve, or the cost of equity. These are all lumped into an ‘interest rate’. But the rates are wildly divergent. Btw - I love your posts. Of everything I’m reading, they’re the most relevant and thought-provoking.