FYI about terminology before people who don't read the paper comment
1. GPT means general purpose technology or any sort of new technology that has a compounding effect on productivity, not the OpenAI model.
2. Productivity in this case means economic output, not the colloquial definition that means "hard work". If it takes 5 automotive factory workers to assemble a car manually but 2 with industrial automation, then the latter are more productive than the former despite expending equal amounts of effort.
3. The crux of this paper is that existing economic metrics are not able to adequately measure the impact of IP and R&D driven innovations in the larger economy. For example, think about how it took 20-30 years for traditional econometrics to fully gauge the impact of digitization and industrial automation that began in earnest in the 1990s and early 2000s.
> 2. Productivity in this case means economic output, not the colloquial definition that means "hard work".
The colloquial definition doesn't mean “hard work”, it also means you “produce” more (optionally, for the same amount of time). The difference is how you measure what's being “produced”.
With the “economic output” definition a barista in a posh bar in SF is “more productive” than their counterpart in a popular place in rural Minnesota, because it generates more revenue, even if the later serves twice as many patrons while also keeping the restroom clean (which would colloquially make the later “more productive”).
“Economic output productivity” can also decline if consumer spending do so, not because workers are “less productive” (in the colloquial sense), but because unsold goods or services don't count as “economic output”.
(IMHO overloading things that have a well-defined colloquial term is a very bad habit of economists and it makes things needlessly confusing for laypersons)
MS Word can be a surprisingly good and powerful type setting system if you use its fancier features correctly, like named styles. I used to have a Latex resume and switched to LibreOffice several years ago. It gives me more than enough precise control over layout and styling, with less effort, and it still generates good PDFs.
This completely breaks down under the current reality of AI investment, as players large and small are no longer price-takers. The marginal costs of investment are not constant because we have finite supplies of GPUs, TPUs, memory, hard drives, and power. The Hamiltonian in equations 5 and 6 needs to account for this.
It's not that supply was actually infinite, but you didn't realistically have situations where you said "I want to buy GPUs for a data center" only to be told "there's a 3 year waiting list."
You might have two months after NVidia 3090s came out where they were short, but it is nothing like today.
Citation needed. Industries that faced multi year supply constraints in recent memory include: nuclear power, battery manufacturing, flagship commercial aircraft models, late-stage pharmaceutical safety certification, certain luxury cars, and more.
At the same time, a lot of interest in this paper is specific to the AI industry (even if it's not only about that), and outside of high-end labor, the other constraints are new there.
Isn't this true in a lot of situations? Basically anytime maximum production capacity of an input is limited, or scale-up time is very long, firms are large relative to the supply so they are not price takers.
Q: The J-dip is where capital stock is just about to overtake investment growth, why should it lag the hype trough where presumably value overtakes interest ?
I did a write-up on the history of the J-curve and some 2026 macro data that supports it for generative AI. The short version: U.S. productivity is climbing again after a decade of stagnation, and the original j curve economist proposes it might be due to AI hitting the upper part of the j-curve.
FYI about terminology before people who don't read the paper comment
1. GPT means general purpose technology or any sort of new technology that has a compounding effect on productivity, not the OpenAI model.
2. Productivity in this case means economic output, not the colloquial definition that means "hard work". If it takes 5 automotive factory workers to assemble a car manually but 2 with industrial automation, then the latter are more productive than the former despite expending equal amounts of effort.
3. The crux of this paper is that existing economic metrics are not able to adequately measure the impact of IP and R&D driven innovations in the larger economy. For example, think about how it took 20-30 years for traditional econometrics to fully gauge the impact of digitization and industrial automation that began in earnest in the 1990s and early 2000s.
> 2. Productivity in this case means economic output, not the colloquial definition that means "hard work".
The colloquial definition doesn't mean “hard work”, it also means you “produce” more (optionally, for the same amount of time). The difference is how you measure what's being “produced”.
With the “economic output” definition a barista in a posh bar in SF is “more productive” than their counterpart in a popular place in rural Minnesota, because it generates more revenue, even if the later serves twice as many patrons while also keeping the restroom clean (which would colloquially make the later “more productive”).
“Economic output productivity” can also decline if consumer spending do so, not because workers are “less productive” (in the colloquial sense), but because unsold goods or services don't count as “economic output”.
(IMHO overloading things that have a well-defined colloquial term is a very bad habit of economists and it makes things needlessly confusing for laypersons)
This paper clearly shows why you shouldn't use MS Word to typeset research.
MS Word can be a surprisingly good and powerful type setting system if you use its fancier features correctly, like named styles. I used to have a Latex resume and switched to LibreOffice several years ago. It gives me more than enough precise control over layout and styling, with less effort, and it still generates good PDFs.
TBH, if a paper is not written by Latex, I naturally question the research and learning ability of the authors, and I don't want to read it.
So I guess you ignore all of biology and just focus on math and physics ;)
This completely breaks down under the current reality of AI investment, as players large and small are no longer price-takers. The marginal costs of investment are not constant because we have finite supplies of GPUs, TPUs, memory, hard drives, and power. The Hamiltonian in equations 5 and 6 needs to account for this.
are you saying that previous technologies had effectively infinite supply?
It's not that supply was actually infinite, but you didn't realistically have situations where you said "I want to buy GPUs for a data center" only to be told "there's a 3 year waiting list."
You might have two months after NVidia 3090s came out where they were short, but it is nothing like today.
Citation needed. Industries that faced multi year supply constraints in recent memory include: nuclear power, battery manufacturing, flagship commercial aircraft models, late-stage pharmaceutical safety certification, certain luxury cars, and more.
Fair!
At the same time, a lot of interest in this paper is specific to the AI industry (even if it's not only about that), and outside of high-end labor, the other constraints are new there.
No. I'm stating where the paper's assumptions are clearly violated.
AI companies are intentionally trying to monopolize the supply of inputs needed for R&D. This violates homogeneity of degree 1.
Isn't this true in a lot of situations? Basically anytime maximum production capacity of an input is limited, or scale-up time is very long, firms are large relative to the supply so they are not price takers.
PP's point has (mostly?) been addressed in
https://www.aeaweb.org/articles?id=10.1257/mac.20180386
Inputs that can be monopolized are manifestly _not_ the intangibles; price-taking is independent of their effect
I don't think they implied that those inputs are the intangibles, only that the functional form assumption in the 2018 paper is unrealistic.
If you are one of those that are amused by attempts to synthesize paradigms, here's one that superposes J-curve on the hype curve
https://www.financialprofessionals.org/training-resources/re...
Q: The J-dip is where capital stock is just about to overtake investment growth, why should it lag the hype trough where presumably value overtakes interest ?
I did a write-up on the history of the J-curve and some 2026 macro data that supports it for generative AI. The short version: U.S. productivity is climbing again after a decade of stagnation, and the original j curve economist proposes it might be due to AI hitting the upper part of the j-curve.
https://lightsight.ai/blog/j-curve (disclosure: my company’s blog)
[flagged]
[dead]
[dead]